pith. sign in

arxiv: 2606.08685 · v1 · pith:GPPFELCQnew · submitted 2026-06-07 · 🧮 math.NA · cs.NA

ND-TNN: Tensor-Neural-Network Approximation for High-Dimensional Nonlocal Diffusion Models

Ziyue Cai , Zuoqiang Shi This is my paper

Pith reviewed 2026-06-27 17:51 UTC · model grok-4.3

classification 🧮 math.NA cs.NA
keywords tensor neural networknonlocal diffusionhigh-dimensional approximationasymptotically compatible schemeGaussian kernelerror estimatesDirichlet boundary conditionNeumann boundary condition
0
0 comments X

The pith

Tensor neural network reduces high-dimensional nonlocal diffusion integrals to low-dimensional products while preserving asymptotically compatible error bounds.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a tensor neural network method for nonlocal diffusion models in spaces up to dimension twenty. The tensor-product form of the network together with the separable Gaussian kernel converts the high-dimensional integrals that appear in the nonlocal energy into products of low-dimensional integrals that are evaluated by Gauss-Legendre quadrature. Error analysis establishes asymptotically compatible L2 bounds for both Dirichlet and Neumann problems, with an additional smoothed H1 estimate available after post-processing for the Neumann case. Numerical tests on tensor-product domains and on L-shaped domains confirm that the observed errors follow the predicted rates.

Core claim

The TNN ansatz, when combined with the separable Gaussian kernel, reduces the nonlocal energy integrals exactly to products of low-dimensional integrals that are computed by quadrature; the resulting approximation satisfies the asymptotically compatible L2 estimate ||u_loc - u_δ,p||_L2(Ω) ≤ C(ε_f/√δ + ε_g/δ + ε_u/√δ + η_opt) + C√δ for Dirichlet conditions and an improved O(ε_f + ε_g/√δ + ε_u + η_opt + δ) bound together with a smoothed H1 estimate for Neumann conditions.

What carries the argument

The tensor neural network (TNN) ansatz paired with the separable Gaussian kernel, which factors the high-dimensional nonlocal integrals into products of one-dimensional integrals evaluated by Gauss-Legendre quadrature.

If this is right

  • The L2 error for Dirichlet conditions remains controlled by data approximation errors, optimization residual, and a term of order √δ as the nonlocal horizon δ approaches zero.
  • Neumann conditions yield an improved L2 rate of order δ and permit an H1 estimate after a smoothing post-processing step.
  • The method scales to tensor-product domains of dimension up to twenty without the integrals becoming prohibitive.
  • A TNN-based preconditioning step extends the scheme to non-separable source and boundary data.
  • The approach remains practically accurate on L-shaped domains even though the analysis assumes smooth tensor-product domains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same separability reduction could be applied to other nonlocal operators whose kernels admit a product structure.
  • Quadrature accuracy on the reduced integrals may become the dominant error source once network optimization reaches machine precision.
  • The smoothing step used for the H1 estimate suggests a general post-processing route for recovering derivatives from L2 approximations in nonlocal settings.
  • Adaptive choice of quadrature nodes per dimension could further improve performance in dimensions beyond twenty.

Load-bearing premise

The Gaussian kernel is separable and the domain admits a tensor-product structure that lets the high-dimensional integrals factor exactly into products of low-dimensional integrals.

What would settle it

A computation on a five-dimensional tensor-product domain that, after driving the optimization residual below 10^-8, produces an L2 error larger than twice the sum of the predicted approximation terms plus √δ would falsify the error bound.

Figures

Figures reproduced from arXiv: 2606.08685 by Ziyue Cai, Zuoqiang Shi.

Figure 1
Figure 1. Figure 1: Illustration of the tensor neural network architecture. Proposition 2.7 (Approximation property of TNNs). Let m ≥ 0 and f ∈ Hm(Ω). Then for any ε > 0, there exist a positive integer p and a TNN output Ψ(x; Θ) ∈ Vd p of the form (2.10) such that ∥f − Ψ(·; Θ)∥Hm(Ω) < ε. (2.11) Remark 2.8. Proposition 2.7 states that S p≥1 V d p is dense in Hm(Ω), but does not provide an explicit relationship between the rank… view at source ↗
Figure 2
Figure 2. Figure 2: Mean pointwise residual rmean versus training iteration for the Dirichlet case with tensor-product data (δ = 0.05, d = 3, 5, 10, 20). As in the Dirichlet case, we define the residual of (2.7) by rN (x; uδ,p,N ) := 1 δ 2 Z Ω Rδ(x, y) [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: L 2 and H1 errors between the local solution uloc and the TNN output uδ,p as functions of δ for the Dirichlet case with tensor-product data (d = 3, 5, 10, 20). εu are nonzero in Theorems 3.1–3.3. We work on Ω = [0, 1]d with the exact solution uloc(x) = uloc,N (x) = exp 1 d X d i=1 xixi+1! , xd+1 := x1. For convenience, we write S(x) := 1 d Pd i=1 xixi+1, so that uloc = uloc,N = e S. We consider the Neumann… view at source ↗
Figure 4
Figure 4. Figure 4: Mean pointwise residual rmean,N versus training iter￾ation for the Neumann case with tensor-product data (δ = 0.05, d = 3, 5, 10, 20). indicators to measure the approximation quality: RMSE := vuut 1 Napprox Napprox X j=1 [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: L 2 and H1 errors between the local solution uloc,N and the TNN output uδ,p,N as functions of δ for the Neumann case with tensor-product data (d = 3, 5, 10, 20) [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Mean pointwise residual rmean,N versus training iter￾ation for the Neumann case with non-tensor-product data (d = 3, 5, 10, 20). 5.2.3. Dirichlet case. We also test the same TNN-based method on the Dirichlet problem with the same exact solution. Here the source term is f(x) = −∆uloc(x) = −|∇S| 2 uloc, where we again used ∆S = 0, and the boundary data g is obtained by restricting uloc to each face of ∂Ω. Th… view at source ↗
Figure 7
Figure 7. Figure 7: L 2 and H1 errors between the local solution uloc,N and the TNN output uδ,p,N as functions of δ for the Neumann case with non-tensor-product data (d = 3, 5, 10, 20). 5.3.1. Two-dimensional case. We first consider the two-dimensional setting. The domain Ω is Ω = [0, 1] × [0, 0.5] ∪ [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Mean pointwise residual rmean versus training iter￾ation for the Dirichlet case with non-tensor-product data (d = 3, 5, 10, 20). Following this procedure, the method can be applied as in Section 5.2. The numeri￾cal results are displayed in Figures 11a and 11b, supporting the empirical robustness of the method on the two-dimensional L-shaped domain. 5.3.2. Three-dimensional case. We now consider a three-dim… view at source ↗
Figure 9
Figure 9. Figure 9: L 2 and H1 errors between the local solution uloc and the TNN output uδ,p as functions of δ for the Dirichlet case with non-tensor-product data (d = 3, 5, 10, 20). (a) Two-dimensional L￾shaped region. (b) Three-dimensional L￾shaped region [PITH_FULL_IMAGE:figures/full_fig_p029_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Illustration of the L-shaped computational regions. As in the two-dimensional case, this experiment uses the same approximation mech￾anism but lies outside the smooth-domain assumptions used in the proof of Theo￾rems 3.2–3.3. The numerical results are reported in Figures 11c and 11d, confirming the practical robustness of the method on the three-dimensional L-shaped domain [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 11
Figure 11. Figure 11: Mean pointwise residual rmean,N and the L 2 , H1 er￾rors between the local solution uloc,N and the TNN output uδ,p,N on L-shaped regions under Neumann boundary conditions. 6. Conclusions In this paper we have constructed and analyzed a variational solver for nonlo￾cal diffusion models with Dirichlet and Neumann boundary conditions using the existing tensor neural network architecture as the trial class. T… view at source ↗
read the original abstract

We study a numerical method, built on the tensor neural network (TNN) architecture introduced in \cite{wang2022tensor}, for solving nonlocal diffusion models in high-dimensional spaces. The tensor-product structure of the TNN ansatz, combined with the separability of the Gaussian kernel, reduces the high-dimensional integrals in the nonlocal energy to products of low-dimensional integrals, which are evaluated by Gauss--Legendre quadrature; nonseparable source and boundary data are handled by a TNN-based preconditioning step. For the Dirichlet boundary condition, we establish the asymptotically compatible $L^2$ error estimate \[ \|u_{\mathrm{loc}}-u_{\delta,p}\|_{L^2(\Omega)} \le C\!\left(\frac{\varepsilon_f}{\sqrt\delta} +\frac{\varepsilon_g}{\delta} +\frac{\varepsilon_u}{\sqrt\delta} +\eta_{\mathrm{opt}}\right) +C\sqrt\delta, \] where $\varepsilon_f$, $\varepsilon_g$ and $\varepsilon_u$ are the data and trial-class approximation errors and $\eta_{\mathrm{opt}}$ is the optimization residual. For the Neumann boundary condition, the $L^2$ estimate is improved to $O(\varepsilon_f+\varepsilon_g/\sqrt\delta+\varepsilon_u +\eta_{\mathrm{opt}}+\delta)$, and an $H^1$ gradient estimate is further obtained through a smoothing post-processing step. Numerical experiments on tensor-product domains up to $d=20$ support the theoretical results, and additional tests on two- and three-dimensional $L$-shaped domains demonstrate the practical robustness of the method beyond the smooth-domain setting covered by the analysis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The manuscript develops ND-TNN, a tensor-neural-network method for high-dimensional nonlocal diffusion problems. It uses the tensor-product structure of the TNN ansatz together with separability of the Gaussian kernel to reduce the nonlocal energy integrals to products of low-dimensional quadratures; non-separable data are handled via a TNN preconditioner. Asymptotically compatible L² error bounds are stated that explicitly incorporate data-approximation errors (ε_f, ε_g), trial-space error (ε_u), optimization residual (η_opt), and the standard O(√δ) consistency term; improved L² and post-processed H¹ bounds are given for Neumann conditions. Numerical tests on tensor-product domains up to dimension 20 and on L-shaped domains are reported in support of the theory.

Significance. If the stated error estimates are valid, the work supplies a scalable, theoretically supported route to high-dimensional nonlocal models whose cost grows only linearly with dimension under the tensor-product assumption. The explicit inclusion of optimization residuals and data-approximation terms inside the bounds, together with the numerical demonstration in d=20, constitutes a concrete advance over purely heuristic neural-network approaches for nonlocal problems.

minor comments (3)
  1. [Section 2] §2 (or wherever the TNN construction is recalled): the precise definition of the trial-class approximation error ε_u should be stated explicitly, including its dependence on the TNN rank and quadrature parameters, so that the error bound can be read without reference to the cited prior work.
  2. [Numerical experiments] The numerical section should report the observed scaling of the optimization residual η_opt with respect to the number of training iterations or network size, so that readers can judge whether the term remains negligible in the high-dimensional regime.
  3. [Numerical experiments] Figure captions and the text around the L-shaped-domain tests should clarify whether the reported errors are measured against a reference solution obtained on a fine mesh or against the local limit; this affects interpretation of the O(√δ) term.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary of our work on ND-TNN and for recommending minor revision. No specific major comments appear in the provided report, so we have no individual points to address. We remain available to incorporate any editorial suggestions.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The derivation of the asymptotically compatible L2 error bound proceeds from standard approximation properties of the TNN trial space, quadrature accuracy for the separable Gaussian kernel, and the known nonlocal-to-local consistency term O(√δ). The TNN ansatz is imported from an external citation (wang2022tensor) with no author overlap indicated, and the error terms ε_f, ε_g, ε_u, η_opt are defined as independent inputs measuring data and optimization residuals rather than being fitted or redefined to match the target bound. No self-definitional steps, fitted predictions, or load-bearing self-citations appear in the analysis or numerical method.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the separability of the Gaussian kernel, the tensor-product structure of the domain and ansatz, and standard properties of Gauss-Legendre quadrature. No new entities are postulated. The TNN architecture is imported from the cited reference.

axioms (2)
  • domain assumption The Gaussian kernel admits a separable product representation that allows the nonlocal integral to factor into products of one-dimensional integrals.
    Invoked in the description of how high-dimensional integrals reduce to low-dimensional ones.
  • standard math Gauss-Legendre quadrature accurately approximates the resulting low-dimensional integrals to sufficient precision.
    Used to evaluate the factored integrals; standard assumption for quadrature-based methods.

pith-pipeline@v0.9.1-grok · 5835 in / 1439 out tokens · 23511 ms · 2026-06-27T17:51:52.611199+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    Abdulle, Y

    A. Abdulle, Y. Bai, and G. Vilmart. Reduced basis finite element heterogeneous multiscale method for quasilinear elliptic homogenization problems. Discrete and Continuous Dynamical Systems-Series S , 8(1):91–118, 2015

    2015

  2. [2]

    Ainsworth and C

    M. Ainsworth and C. Glusa. Towards an efficient finite element method for the integral fractional laplacian on polygonal domains. In Contemporary Computational Mathematics-A Celebration of the 80th Birthday of Ian Sloan , pages 17–57. Springer, 2018

    2018

  3. [3]

    Askari, F

    E. Askari, F. Bobaru, R. Lehoucq, M. Parks, S. A. Silling, and O. Weckner. Peridynamics for multiscale materials modeling. In Journal of Physics: Conference Series , volume 125, page 012078, 2008

    2008

  4. [4]

    B. Bahr, M. Faustmann, and J. M. Melenk. An implementation of hp-fem for the fractional laplacian. Computers & Mathematics with Applications , 176:324–348, 2024

    2024

  5. [5]

    Bonito, J

    A. Bonito, J. P. Borthagaray, R. H. Nochetto, E. Otárola, and A. J. Salgado. Numerical methods for fractional diffusion. Computing and Visualization in Science , 19(5):19–46, 2018

    2018

  6. [6]

    Bourgain, H

    J. Bourgain, H. Brezis, and P. Mironescu. Another look at sobolev spaces. 2001

    2001

  7. [7]

    Buades, B

    A. Buades, B. Coll, and J.-M. Morel. Image denoising methods. a new nonlocal principle. SIAM review , 52(1):113–147, 2010

    2010

  8. [8]

    Chen and M

    X. Chen and M. Gunzburger. Continuous and discontinuous finite element methods for a peridynamics model of mechanics. Computer Methods in Applied Mechanics and Engineering , 200(9-12):1237–1250, 2011

    2011

  9. [9]

    Q. Du. Nonlocal Modeling, Analysis, and Computation: Nonlocal Modeling, Analysis, and Computation. SIAM, 2019

    2019

  10. [10]

    Q. Du, B. Engquist, and X. Tian. Multiscale modeling, homogenization and nonlocal effects: Mathematical and computational issues. arXiv preprint arXiv:1909.00708 , 2019

    work page arXiv 1909

  11. [11]

    Q. Du, L. Ju, L. Tian, and K. Zhou. A posteriori error analysis of finite element method for linear nonlocal diffusion and peridynamic models. Mathematics of computation, 82(284):1889– 1922, 2013

    1922

  12. [12]

    Du and J

    Q. Du and J. Yang. Asymptotically compatible fourier spectral approximations of nonlocal allen–cahn equations. SIAM Journal on Numerical Analysis , 54(3):1899–1919, 2016

    1919

  13. [13]

    L. C. Evans. Partial differential equations , volume 19. American mathematical society, 2022

    2022

  14. [14]

    Gilboa and S

    G. Gilboa and S. Osher. Nonlocal linear image regularization and supervised segmentation. Multiscale Modeling & Simulation , 6(2):595–630, 2007

    2007

  15. [15]

    Gilboa and S

    G. Gilboa and S. Osher. Nonlocal operators with applications to image processing. Multiscale Modeling & Simulation , 7(3):1005–1028, 2009

    2009

  16. [16]

    Griebel and S

    M. Griebel and S. Knapek. Optimized tensor-product approximation spaces. Constructive Approximation, 16(4):525–540, 2000

    2000

  17. [17]

    Y. D. Ha and F. Bobaru. Characteristics of dynamic brittle fracture captured with peridy- namics. Engineering Fracture Mechanics, 78(6):1156–1168, 2011

    2011

  18. [18]

    Y. Li, Z. Lin, Y. Wang, and H. Xie. Tensor neural network interpolation and its applications. arXiv preprint arXiv:2404.07805 , 2024

    work page arXiv 2024

  19. [19]

    Z. Li, Z. Shi, and J. Sun. Point integral method for solving poisson-type equations on man- ifolds from point clouds with convergence guarantees. Communications in Computational Physics, 22(1):228–258, 2017

    2017

  20. [20]

    Z. Lin, Q. Ma, H. Xie, and X. Yin. Solving time-fractional partial integro-differential equations using tensor neural network. SIAM Journal on Scientific Computing , 48(1):C164–C189, 2026. 36 ND-TNN

    2026

  21. [21]

    D. J. Littlewood. Simulation of dynamic fracture using peridynamics, finite element modeling, and contact. In ASME International mechanical engineering congress and exposition , volume 44465, pages 209–217, 2010

    2010

  22. [22]

    Meng and Z

    Y. Meng and Z. Shi. Maximum principle preserving nonlocal diffusion model with dirichlet boundary condition. arXiv preprint arXiv:2310.01221 , 2023

    work page arXiv 2023

  23. [23]

    Meng and Z

    Y. Meng and Z. Shi. Asymptotically compatible error bound of finite element method for nonlocal diffusion model with an efficient implementation. Journal of Scientific Computing , 107(2):71, 2026

    2026

  24. [24]

    Z. Shi. Enforce the dirichlet boundary condition by volume constraint in point integral method. arXiv preprint arXiv:1506.02343 , 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  25. [25]

    Z. Shi, S. Osher, and W. Zhu. Weighted nonlocal laplacian on interpolation from sparse data. Journal of Scientific Computing , 73(2):1164–1177, 2017

    2017

  26. [26]

    Shi and J

    Z. Shi and J. Sun. Convergence of the point integral method for laplace–beltrami equation on point cloud. Research in the Mathematical Sciences , 4(1):22, 2017

    2017

  27. [27]

    S. A. Silling. Reformulation of elasticity theory for discontinuities and long-range forces. Journal of the Mechanics and Physics of Solids , 48(1):175–209, 2000

    2000

  28. [28]

    Y. Tao, Q. Sun, Q. Du, and W. Liu. Nonlocal neural networks, nonlocal diffusion and nonlocal modeling. Advances in Neural Information Processing Systems , 31, 2018

    2018

  29. [29]

    Tian and Q

    X. Tian and Q. Du. Analysis and comparison of different approximations to nonlocal diffusion and linear peridynamic equations. SIAM Journal on Numerical Analysis , 51(6):3458–3482, 2013

    2013

  30. [30]

    X. Wang, R. Girshick, A. Gupta, and K. He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 7794–7803, 2018

    2018

  31. [31]

    Y. Wang, P. Jin, and H. Xie. Tensor neural network and its numerical integration. arXiv preprint arXiv:2207.02754 , 2022

    work page arXiv 2022

  32. [32]

    Wang and H

    Y. Wang and H. Xie. Computing multi-eigenpairs of high-dimensional eigenvalue problems using tensor neural networks. Journal of Computational Physics , 506:112928, 2024

    2024

  33. [33]

    Zhang, M

    X. Zhang, M. Gunzburger, and L. Ju. Nodal-type collocation methods for hypersingular integral equations and nonlocal diffusion problems. Computer Methods in Applied Mechanics and Engineering , 299:401–420, 2016

    2016

  34. [34]

    Zhou and Q

    K. Zhou and Q. Du. Mathematical and numerical analysis of linear peridynamic models with nonlocal boundary conditions. SIAM Journal on Numerical Analysis , 48(5):1759–1780, 2010

    2010

  35. [35]

    Q. Zhou, T. Wu, J. Liu, Q. Sun, H. Xie, and Z. Xu. Sum-of-gaussians tensor neural networks for high-dimensional Schrödinger equation. arXiv preprint arXiv:2508.10454 , 2025

    work page arXiv 2025

  36. [36]

    X. Zhu, Z. Ghahramani, and J. D. Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In Proceedings of the 20th International conference on Machine learning (ICML-03), pages 912–919, 2003. Ziyue Cai: Qiuzhen College, Tsinghua University, Beijing, China, 100084. Email address : cai-zy22@mails.tsinghua.edu.cn Zuoqiang Shi: Yau M...

    2003