Potential Fredholm Neural Network (PFNN)

• Potential Fredholm Neural Network (PFNN) is a neural architecture that leverages boundary integral formulations and potential theory to solve 2D elliptic PDEs.
• It unrolls fixed-point iterations as network layers, combining a Fredholm NN block for boundary density with an additional potential layer for interior evaluation.
• The design offers explainability by linking each network component to discretized integral operations, with error bounds tied to iteration and quadrature accuracy.

to=arxiv_search.search 天天中彩票app 彩神争霸怎么json {"query":"\"Potential Fredholm Neural Network\" OR \"Fredholm Neural Networks\" elliptic PDEs boundary integral", "max_results": 10, "sort_by": "relevance"} to=arxiv_search.search 大发快三大小单双 լինգjson {"query":"PFNN penalty-free neural network boundary value problems complex geometries", "max_results": 10, "sort_by": "relevance"} to=arxiv_search.search ส่งเงินบาทไทยjson {"query":"Fredholm integral equations training neural networks tensor train Ritz-Galerkin", "max_results": 10, "sort_by": "relevance"} Potential Fredholm Neural Network (PFNN) denotes a class of deep neural architectures for elliptic partial differential equations in which the network is derived directly from a boundary integral formulation and a fixed-point scheme from potential theory. In the formulation developed for two-dimensional linear and semi-linear elliptic problems, a PFNN combines a Fredholm Neural Network (FNN) that solves the boundary integral equation for an unknown boundary density with a further layer that evaluates the corresponding double-layer potential inside the domain. Its layers, widths, weights, biases, and hyperparameters are determined by the fundamental solution, quadrature rules, and the fixed-point iteration, rather than by generic end-to-end training of internal parameters (Georgiou et al., 8 Jul 2025).

1. Terminology and scope

Within recent arXiv literature, the acronym PFNN is not unique. It refers both to Potential Fredholm Neural Network in the boundary-integral literature and to Penalty-Free Neural Network in variational neural PDE solvers on complex geometries. The distinction is substantive rather than stylistic: the former is rooted in potential theory and Fredholm integral equations, while the latter is rooted in weak formulations and exact architectural enforcement of essential boundary conditions (Georgiou et al., 8 Jul 2025, Sheng et al., 2020).

Acronym usage	Meaning	Representative source
PFNN	Potential Fredholm Neural Network	(Georgiou et al., 8 Jul 2025)
PFNN	Penalty-Free Neural Network	(Sheng et al., 2020)
PFNN-2	Domain decomposed penalty-free variant	(Sheng et al., 2022)

In the Potential Fredholm sense, the term “potential” refers to the boundary-integral method and layer-potential representations of elliptic PDEs. This should not be conflated with the “potential” or energy interpretation used in variational neural solvers, nor with function-space Fredholm formulations in which a parameter function acts as a “potential” over parameter space (Gelß et al., 2023). A recurring misconception is therefore to treat all PFNN papers as belonging to one method family; the arXiv record instead shows two distinct lines of development that share an acronym but differ in mathematical formulation and computational workflow.

2. Boundary-integral and potential-theoretic formulation

The PFNN developed for elliptic PDEs addresses two-dimensional Poisson, modified Helmholtz, and semi-linear elliptic equations on a bounded domain $\Omega \subset \mathbb{R}^2$, with detailed treatment of Dirichlet boundary conditions. For Poisson,

$$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$

the solution is represented by a double-layer potential plus a volume potential,

$$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$

where $\Phi$ is the fundamental solution and $\beta$ is an unknown boundary density. Imposing the boundary condition through the jump relation produces a Fredholm integral equation of the second kind for $\beta$ on $\partial\Omega$ (Georgiou et al., 8 Jul 2025).

For Poisson, the fundamental solution is

$\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$

while for $\mathcal L=\Delta-\lambda$ the modified Helmholtz fundamental solution is

$$\Phi(x,y)=-\frac{1}{2\pi}K_0(\sqrt{\lambda}\,|x-y|),$$

with $$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$0 the modified Bessel function of the second kind. In the Poisson case, the boundary density satisfies

$$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$1

This Fredholm equation is the object solved by the FNN block inside the PFNN (Georgiou et al., 8 Jul 2025).

A technical difficulty is the jump behavior of the double-layer potential near the boundary. The PFNN construction therefore uses a boundary-regular representation in which a boundary point $$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$2 with the same angular coordinate as $$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$3 is introduced, and the potential is rewritten in terms of $$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$4. For Poisson,

$$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$5

For Helmholtz, an analogous identity uses $$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$6 and modified coefficients involving $$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$7 (Georgiou et al., 8 Jul 2025). This boundary-regular reformulation is central: it removes the singular jump term from the integral part and transfers it into explicit additive terms that are directly encoded in the network.

3. Network architecture and fixed-point interpretation

The PFNN consists of two coupled components. The first is a Fredholm NN that solves the boundary integral equation for $$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$8 by unrolling a fixed-point iteration; the second is a potential block that evaluates the PDE solution $$\begin{cases} \Delta u(x) = \psi(x), & x \in \Omega,\ u(x) = f(x), & x \in \partial\Omega, \end{cases}$$9 from the approximate boundary density. In the linear boundary-integral setting, the fixed-point step is written abstractly as

$$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$0

and approximated by Krasnosel’skii–Mann iteration,

$$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$1

The PFNN depth therefore has a direct numerical interpretation: the number of hidden layers equals the number of fixed-point iterations, and the layer width equals the number of quadrature nodes on the boundary (Georgiou et al., 8 Jul 2025).

On a boundary grid $$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$2, the first hidden layer encodes the source term through

$$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$3

For inner layers $$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$4, the weight matrix has entries

$$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$5

and the bias is

$$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$6

The activation in this block is linear. This is a literal network encoding of the discretized integral operator and its relaxation parameter, not a metaphorical analogy to iteration (Georgiou et al., 8 Jul 2025).

The additional “potential” part of the PFNN is what distinguishes it from a generic Fredholm NN. After the boundary-density approximation $$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$7 has been obtained, an extra hidden layer computes the vector $$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$8 via

$$u(x)=\int_{\partial\Omega}\beta(y)\,\frac{\partial\Phi(x,y)}{\partial n_y}\,d\sigma_y +\int_\Omega \Phi(x,y)\,\psi(y)\,dy,$$9

The output layer then evaluates the regularized double-layer potential. In the Poisson case, the output weights are

$\Phi$0

where

$\Phi$1

and the output bias contains the explicit jump and volume terms,

$\Phi$2

The modified Helmholtz case uses the corresponding formula with $\Phi$3 and $\Phi$4-dependent coefficients (Georgiou et al., 8 Jul 2025).

This architecture makes PFNN explainable in a strong numerical-analysis sense. Each layer corresponds to a known operation: application of the boundary kernel, fixed-point relaxation, subtraction of the boundary trace, and potential evaluation. Hyperparameters are likewise numerical parameters: depth is iteration count, width is quadrature resolution, and the weights are discretized kernel evaluations. The closely related “Fredholm Neural Networks” framework formalized this iteration-as-network construction for linear and nonlinear Fredholm equations of the second kind and later served as the basis of the PFNN extension to elliptic PDEs (Georgiou et al., 2024).

4. Theoretical properties and error decomposition

The PFNN theory is built on two error layers: the error in approximating the boundary density by the FNN block, and the discretization error in evaluating the boundary and volume potentials. For a nonexpansive Fredholm operator, the underlying FNN approximation $\Phi$5 satisfies an explicit estimate of the form

$\Phi$6

so the boundary-density error is controlled by the iteration count $\Phi$7, the quadrature resolution $\Phi$8, and the operator’s nonexpansive constant. In the constant-$\Phi$9 case, $\beta$0, making the dependence of accuracy on depth explicit (Georgiou et al., 8 Jul 2025).

The PFNN error bound for the PDE solution $\beta$1 is then expressed in terms of the boundary-density approximation error $\beta$2 and four discretization terms $\beta$3 associated with the potential and boundary integrals. For interior points,

$\beta$4

while on the boundary,

$\beta$5

The omitted coefficients are explicit integrals of $\beta$6, $\beta$7, and $\beta$8. The structural point is that the total PFNN error is decomposed into iteration error from the boundary Fredholm solve and quadrature error from the potential evaluation; both contributions correspond directly to identifiable architectural components (Georgiou et al., 8 Jul 2025).

For semi-linear elliptic equations, the PFNN is embedded in an outer fixed-point loop, yielding a Recurrent PFNN (RPFNN). If $\beta$9 denotes the outer contraction map, then after $\beta$0 outer iterations

$\beta$1

where the $\beta$2 terms are precisely the inner PFNN discretization errors. This gives the semi-linear architecture a two-scale convergence structure: geometric decay from the outer contraction and explicit discretization control from the inner boundary-integral solver (Georgiou et al., 8 Jul 2025).

A distinctive claim of this framework is boundary fidelity. Because the PFNN is derived from a representation that already encodes the jump behavior of the double-layer potential, boundary conditions are respected by construction rather than through penalty terms. The earlier Fredholm NN paper showed a related Laplace-unit-disc example with maximum error about $\beta$3 for the potential-theoretic architecture, again emphasizing that the boundary treatment is inherited from the integral representation rather than imposed by an auxiliary loss term (Georgiou et al., 2024).

5. Forward and inverse formulations

For forward problems, PFNN is a deterministic solver. Given boundary data $\beta$4 and source information $\beta$5 or $\beta$6, the architecture is assembled from the PDE data, the boundary grid, quadrature rules, and the chosen number of iterations. Internal PFNN weights are not trained by generic backpropagation; they are precomputed from the Green’s function, quadrature, and fixed-point coefficients. This differentiates PFNN from PINN-style residual minimization, even though both are neural-network-based PDE solvers (Georgiou et al., 8 Jul 2025).

The reported forward experiments concern the unit disk. For Poisson,

$\beta$7

with exact solution

$\beta$8

a boundary grid $\beta$9 and radial grid size $\partial\Omega$0 were used, and both $\partial\Omega$1 and MAE errors in the interior and on the boundary decreased monotonically with the number of layers $\partial\Omega$2. For the modified Helmholtz example,

$\partial\Omega$3

with exact solution $\partial\Omega$4, the PFNN again produced accurate solutions with very small boundary error. For the semi-linear equation

$\partial\Omega$5

with exact solution $\partial\Omega$6, an RPFNN with $\partial\Omega$7 outer iterations, $\partial\Omega$8 PFNN layers per iteration, a $\partial\Omega$9 polar grid, and $\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$0 was used; the boundary error was again reported as near machine precision (Georgiou et al., 8 Jul 2025).

Inverse problems are handled differently. In the source-identification example for Poisson,

$\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$1

the PFNN is kept fixed as the PDE solver, while a separate shallow neural model $\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$2 for the unknown source is trained from coarse interior measurements. The loss is

$\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$3

with Tikhonov regularization $\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$4. In the reported implementation, the source model had one hidden layer with $\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$5 nodes and $\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$6 activation, the optimizer was Levenberg–Marquardt with $\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$7 iterations, and $\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$8. On a test grid of size $\Phi(x,y)=\frac{1}{2\pi}\ln|x-y|,$9, the mean training MSE was $\mathcal L=\Delta-\lambda$0, the mean test MSE was $\mathcal L=\Delta-\lambda$1, the mean test $\mathcal L=\Delta-\lambda$2 error was $\mathcal L=\Delta-\lambda$3, and the boundary test MAE and $\mathcal L=\Delta-\lambda$4 were $\mathcal L=\Delta-\lambda$5 and $\mathcal L=\Delta-\lambda$6, respectively (Georgiou et al., 8 Jul 2025).

6. Relation to adjacent Fredholm and neural PDE frameworks

PFNN did not emerge in isolation. The immediate precursor is the Fredholm NN framework for linear and nonlinear Fredholm integral equations of the second kind, where a fully connected DNN with analytically specified weights reproduces fixed-point iterations and can be extended by a final integral layer to evaluate potential representations for elliptic PDEs (Georgiou et al., 2024). A later generalization, Fredholm Integral Neural Operators (FREDINOs), moved from fixed known kernels to learned contractive integral operators in arbitrary dimensions and explicitly incorporated PFNNs for a nonlinear elliptic PDE through a boundary integral equation formulation (Georgiou et al., 3 Apr 2026).

Two other nearby strands are conceptually related but mathematically distinct. First, the paper on “Fredholm integral equations for function approximation and the training of neural networks” recasts shallow-network training as the approximate least-squares solution of a Fredholm integral equation of the first kind, with a parameter function over parameter space learned by Ritz–Galerkin, Tikhonov regularization, and tensor-train methods (Gelß et al., 2023). Second, the LDNN work on nonlinear Volterra–Fredholm–Hammerstein integral equations uses a Legendre-based network and Gaussian quadrature collocation to minimize integral-equation residuals; it is not a PFNN in the potential-theoretic sense, but it shows another way of hard-wiring Fredholm structure into a neural architecture (Hajimohammadi et al., 2021).

The most frequent confusion concerns the acronym collision with the Penalty-Free Neural Network literature. That PFNN uses two neural networks, a length factor function, and weak/energy formulations to enforce Dirichlet conditions without penalty terms for second-order boundary-value problems on complex geometries; PFNN-2 extends that line to non-self-adjoint and time-dependent PDEs with overlapping domain decomposition (Sheng et al., 2020, Sheng et al., 2022). The shared acronym should therefore not be taken as evidence of a shared architecture. Potential Fredholm PFNNs are boundary-integral fixed-point networks; Penalty-Free PFNNs are weak-form trial-space constructions.

7. Limitations and open directions

The current Potential Fredholm formulation is developed for two-dimensional elliptic PDEs with smooth boundaries and Dirichlet conditions, with examples on the unit disk. Extension to three dimensions is described as conceptually possible but computationally more expensive, and Neumann or Robin conditions would require modified integral formulations and jump relations. The semi-linear treatment is likewise restricted to monotone-iteration settings in which a contraction can be ensured (Georgiou et al., 8 Jul 2025).

The framework’s numerical strengths depend on boundary discretization and quadrature accuracy. For very complex geometries, both the construction of the boundary grid and the evaluation of the singular or weakly singular kernels become more delicate. This suggests a practical boundary between current PFNN implementations and more general geometric PDE solvers: PFNNs are strongest where boundary integral methods are already numerically advantageous, and less mature where geometry or coefficient structure makes boundary-element-style preprocessing expensive (Georgiou et al., 8 Jul 2025).

A plausible implication is that the later FREDINO program points toward a hybrid future for PFNNs. FREDINOs learn contractive linear and nonlinear integral operators while preserving fixed-point convergence, and in the PDE setting they learn a corrected potential kernel $\mathcal L=\Delta-\lambda$7 within a PFNN-style boundary-integral architecture (Georgiou et al., 3 Apr 2026). This suggests a broader interpretation of PFNN: not only as a deterministic encoding of a known boundary integral method, but also as an explainable scaffold into which learned kernels, operator corrections, or uncertainty-aware components can be inserted without abandoning the Fredholm and potential-theoretic structure.