pith. sign in

arxiv: 2605.04126 · v1 · submitted 2026-05-05 · 💻 cs.LG · cs.NA· math.NA

Simultaneous CNN Approximation on Manifolds with Applications to Boundary Value Problems

Hanfei Zhou , Lei Shi This is my paper

Pith reviewed 2026-05-08 18:17 UTC · model grok-4.3

classification 💻 cs.LG cs.NAmath.NA
keywords CNN approximationRiemannian manifoldsSobolev approximationphysics-informed neural networksboundary value problemsspectral boundary lossintrinsic dimension
0
0 comments X

The pith

CNNs approximate manifold functions and their derivatives at rates set by intrinsic dimension alone.

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

The paper establishes that single- and multichannel convolutional neural networks can approximate functions on compact Riemannian manifolds together with their derivatives in Sobolev norms. Approximation rates depend on the manifold's intrinsic dimension and the smoothness gap rather than the ambient embedding dimension. This removes the usual curse of dimensionality for data supported on lower-dimensional structures. The authors then build a physics-informed CNN solver for elliptic boundary-value problems that replaces standard boundary penalties with a spectral loss derived from the boundary Laplace-Beltrami operator.

Core claim

We prove simultaneous Sobolev approximation theorems for CNNs on compact Riemannian manifolds, with error rates controlled by intrinsic dimension and smoothness. We then introduce a physics-informed CNN framework whose boundary loss is expressed as weighted frequency energies via the boundary Laplace-Beltrami operator, yielding stable Sobolev-trace control without auxiliary extensions or singular integrals.

What carries the argument

Simultaneous Sobolev approximation for CNNs transferred to Riemannian manifolds, together with the spectral boundary loss based on the boundary Laplace-Beltrami operator.

If this is right

  • High-dimensional manifold data can be processed by CNNs without exponential growth in network size.
  • Simultaneous derivative approximation enables direct use inside differential operators for PDE solvers.
  • The spectral boundary loss supports FFT-based or precomputed-basis implementations on structured boundaries.
  • Numerical experiments show improved accuracy and stability compared with standard physics-informed networks.

Where Pith is reading between the lines

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

  • The same rate statements may apply to other architectures once they are lifted to manifold-valued inputs.
  • The spectral loss construction could be reused for time-dependent or nonlinear problems on manifolds.
  • The approach suggests a route to parameter-free error bounds when the manifold is sampled rather than given explicitly.

Load-bearing premise

The CNN architecture can be adapted to the Riemannian manifold so that Euclidean approximation rates transfer without extra manifold-specific error terms that would dominate the claimed rates.

What would settle it

A concrete manifold and target function where measured Sobolev approximation error grows with ambient dimension or where the spectral boundary loss fails to produce the predicted trace-norm decay.

Figures

Figures reproduced from arXiv: 2605.04126 by Hanfei Zhou, Lei Shi.

Figure 1
Figure 1. Figure 1: Empirical convergence curves (log–log) and fitted slopes on two manifolds view at source ↗
Figure 2
Figure 2. Figure 2: We compare the function error |uθ − u| and Laplacian error |∆Muθ − ∆Mu| across all four settings. To study the optimization behavior, we track the test RelL 2 and Rel H2 errors over epochs on both MS and MT. With all other hyperparameters fixed, we compare the Sobolev penalty with the standard L 2 penalty. Empirically, the Sobolev penalty L H3/2 bnd yields faster 13 view at source ↗
Figure 2
Figure 2. Figure 2: Pointwise error visualization at N = 4096: left/right subplots correspond to the spatial distributions of |uθ−u| and |∆Muθ−∆Mu|, comparing four combinations of boundary penalty and manifold type. error decay in the early stage and smoother convergence trajectories view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of test L 2 error versus epoch under different interior sample sizes. (a) RelL 2 , N = 16384 (b) RelL 2 , N = 32768 (c) Rel H2s , N = 16384 (d) Rel H2s , N = 32768 view at source ↗
Figure 4
Figure 4. Figure 4: Channel-width sweep results on the upper-torus manifold view at source ↗
read the original abstract

This paper develops convolutional neural network (CNN) methods for simultaneous approximation and elliptic boundary value problems on compact Riemannian manifolds. We establish simultaneous Sobolev approximation results for single- and multichannel CNNs, showing that manifold functions and their derivatives can be approximated with rates governed by the intrinsic dimension and the smoothness gap, rather than by the ambient dimension, thereby mitigating the curse of dimensionality. Building on this approximation theory, we propose a physics-informed CNN (PICNN) framework specially designed for boundary value problems. The main numerical issue is a boundary-norm mismatch: standard PINNs usually impose boundary data through low-order, often L2-type, penalties, whereas elliptic stability requires Sobolev trace control. We address this by introducing a spectral boundary loss based on the boundary Laplace-Beltrami operator, which represents trace errors as weighted frequency energies and relates truncation error to boundary eigenvalue decay. This avoids smooth auxiliary constructions required by exact boundary enforcement and singular double integrals arising in Sobolev-Slobodeckij penalties, while enabling implementations based on Fast Fourier Transforms (FFTs) or precomputed spectral bases on structured boundaries. Numerical experiments demonstrate improved accuracy, convergence, and stability over standard PINNs.

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

2 major / 2 minor

Summary. The manuscript develops simultaneous Sobolev approximation results for single- and multichannel CNNs on compact Riemannian manifolds, showing that manifold functions and their derivatives can be approximated with rates governed by the intrinsic dimension and the smoothness gap rather than the ambient dimension. Building on this, it proposes a physics-informed CNN (PICNN) framework for elliptic boundary value problems that introduces a spectral boundary loss based on the boundary Laplace-Beltrami operator to enforce trace control and avoid issues with standard L2 penalties or Sobolev-Slobodeckij terms.

Significance. If the approximation rates hold as stated, the work would advance manifold learning and physics-informed neural networks by mitigating the curse of dimensionality in high-codimension embeddings and providing a stable, FFT-friendly boundary enforcement method that improves convergence over standard PINNs. The combination of theory and numerics on structured boundaries adds practical utility for PDEs on manifolds.

major comments (2)
  1. [Simultaneous Sobolev approximation results] The transfer of Euclidean CNN Sobolev rates to the manifold setting (via charts, exponential maps, or pullbacks) is load-bearing for the central claim that rates depend only on intrinsic dimension. Explicit bounds are needed on manifold-specific contributions such as metric distortion, Christoffel symbols in derivative approximations, and partition-of-unity overlap errors to confirm they are absorbed into the stated rates without introducing ambient-dimension or curvature-dependent factors that would dominate.
  2. [PICNN framework and spectral boundary loss] The spectral boundary loss is claimed to relate truncation error to boundary eigenvalue decay and to provide Sobolev trace control. A precise stability estimate or comparison theorem showing that this loss yields the required elliptic regularity without auxiliary smooth extensions would be necessary to support the improved accuracy and stability reported in the numerical experiments.
minor comments (2)
  1. [CNN architecture adaptation] Clarify the precise definition of multichannel CNNs in the approximation theorems and how the multichannel structure interacts with the manifold tangent bundle.
  2. [Numerical experiments] In the numerical section, report both intrinsic and embedding dimensions for each manifold example to allow direct assessment of the claimed dimension-independent rates.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments. We address the two major points below by clarifying the proof structure and committing to added estimates and lemmas in the revision.

read point-by-point responses

  1. Referee: [Simultaneous Sobolev approximation results] The transfer of Euclidean CNN Sobolev rates to the manifold setting (via charts, exponential maps, or pullbacks) is load-bearing for the central claim that rates depend only on intrinsic dimension. Explicit bounds are needed on manifold-specific contributions such as metric distortion, Christoffel symbols in derivative approximations, and partition-of-unity overlap errors to confirm they are absorbed into the stated rates without introducing ambient-dimension or curvature-dependent factors that would dominate.

    Authors: We agree that explicit control of these terms strengthens the presentation. The proof localizes via a finite smooth atlas, applies the Euclidean CNN result in each chart (with rates depending only on intrinsic dimension), and controls transition errors by the C^infty regularity of the charts and metric. Metric distortion and Christoffel symbols are bounded by constants depending solely on intrinsic invariants (injectivity radius, sectional curvature bounds, atlas overlap multiplicity) and are absorbed into the overall constant prefactor; they introduce no ambient-dimension dependence. We will add a new lemma (Lemma 3.4) that states these bounds explicitly in terms of manifold geometry only. revision: yes

  2. Referee: [PICNN framework and spectral boundary loss] The spectral boundary loss is claimed to relate truncation error to boundary eigenvalue decay and to provide Sobolev trace control. A precise stability estimate or comparison theorem showing that this loss yields the required elliptic regularity without auxiliary smooth extensions would be necessary to support the improved accuracy and stability reported in the numerical experiments.

    Authors: The spectral loss is defined via the eigen-expansion of the boundary Laplace-Beltrami operator and is equivalent (up to constants) to the H^{1/2} trace norm by the spectral theorem; truncation after M modes is controlled by the eigenvalue decay rate. This directly supplies the Sobolev trace control needed for elliptic regularity without requiring smooth extensions or singular integrals. We will add an appendix proposition establishing the norm equivalence and a remark explaining how the combined loss (interior residual plus spectral boundary term) inherits the required stability. A complete a-priori optimizer error bound is beyond the present scope but is not needed for the stated claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained against external benchmarks

full rationale

The paper's central claims rest on establishing new simultaneous Sobolev approximation rates for CNNs on manifolds (governed by intrinsic dimension) and introducing a spectral boundary loss via the Laplace-Beltrami operator. No load-bearing step reduces by construction to a fitted parameter renamed as prediction, a self-citation chain, or an ansatz smuggled from prior work by the same authors. The approximation results are presented as extensions of Euclidean CNN theory with manifold adaptation, and the spectral loss is constructed explicitly to address boundary-norm mismatch without re-expressing existing quantities. The derivation chain is independent of the target results and externally falsifiable via numerical experiments on BVP accuracy.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The claims rest on standard Sobolev space theory on Riemannian manifolds, elliptic regularity for boundary value problems, and existing CNN approximation results in Euclidean space extended to manifolds; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (2)
  • standard math Compact Riemannian manifolds admit well-defined Sobolev spaces and trace operators with standard embedding and stability properties.
    Invoked implicitly when stating Sobolev approximation rates and boundary trace control for elliptic problems.
  • domain assumption Convolutional neural networks can be defined or pulled back onto the manifold while preserving their approximation capabilities from the ambient Euclidean setting.
    Required for the intrinsic-dimension rates to hold without additional manifold curvature penalties.

pith-pipeline@v0.9.0 · 5506 in / 1390 out tokens · 38448 ms · 2026-05-08T18:17:53.271338+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages

  1. [1]

    Convergence of dif- fusion models under the manifold hypothesis in high-dimensions.arXiv preprint arXiv:2409.18804,

    Iskander Azangulov, George Deligiannidis, and Judith Rousseau. Convergence of diffusion models under the manifold hypothesis in high-dimensions.arXiv preprint arXiv:2409.18804, 2024

    work page arXiv 2024

  2. [2]

    Representation learning: A review and new perspectives.IEEE transactions on pattern analysis and machine intelligence, 35(8):1798–1828, 2013

    Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives.IEEE transactions on pattern analysis and machine intelligence, 35(8):1798–1828, 2013

    work page 2013

  3. [3]

    Enforc- ing dirichlet boundary conditions in physics-informed neural networks and variational physics-informed neural networks.Heliyon, 9(8), 2023

    Stefano Berrone, Claudio Canuto, Moreno Pintore, and Natarajan Sukumar. Enforc- ing dirichlet boundary conditions in physics-informed neural networks and variational physics-informed neural networks.Heliyon, 9(8), 2023

    work page 2023

  4. [4]

    Convergence and error control of consistent PINNs for elliptic PDEs.IMA journal of numerical analysis, 46(1):90–148, 2026

    Andrea Bonito, Ronald DeVore, Guergana Petrova, and Jonathan W Siegel. Convergence and error control of consistent PINNs for elliptic PDEs.IMA journal of numerical analysis, 46(1):90–148, 2026

    work page 2026

  5. [5]

    Reproducing kernel Hilbert spaces on manifolds: Sobolev and diffusion spaces.Analysis and Applications, 19(03):363–396, 2021

    Ernesto De Vito, Nicole M¨ ucke, and Lorenzo Rosasco. Reproducing kernel Hilbert spaces on manifolds: Sobolev and diffusion spaces.Analysis and Applications, 19(03):363–396, 2021

    work page 2021

  6. [6]

    Nonlinear approximation.Acta Numerica, 7:51–150, 1998

    Ronald A DeVore. Nonlinear approximation.Acta Numerica, 7:51–150, 1998

    work page 1998

  7. [7]

    Theory of deep convolu- tional neural networks II: Spherical analysis.Neural Networks, 131:154–162, 2020

    Zhiying Fang, Han Feng, Shuo Huang, and Ding-Xuan Zhou. Theory of deep convolu- tional neural networks II: Spherical analysis.Neural Networks, 131:154–162, 2020

    work page 2020

  8. [8]

    Scattered data interpolation on embedded sub- manifolds with restricted positive definite kernels: Sobolev error estimates.SIAM Journal on Numerical Analysis, 50(3):1753–1776, 2012

    Edward Fuselier and Grady B Wright. Scattered data interpolation on embedded sub- manifolds with restricted positive definite kernels: Sobolev error estimates.SIAM Journal on Numerical Analysis, 50(3):1753–1776, 2012

    work page 2012

  9. [9]

    Evarist M. Gin´ e. The addition formula for the eigenfunctions of the Laplacian.Advances in Mathematics, 18(1):102–107, 1975

    work page 1975

  10. [10]

    Gaussian Error Linear Units (GELUs)

    D Hendrycks. Gaussian error linear units (GELUs).arXiv preprint arXiv:1606.08415, 2016

    work page Pith review arXiv 2016

  11. [11]

    Trace regularity PINNs: EnforcingH 1/2(∂Ω) for bound- ary data.arXiv preprint arXiv:2510.16817, 2025

    Doyoon Kim and Junbin Song. Trace regularity PINNs: EnforcingH 1/2(∂Ω) for bound- ary data.arXiv preprint arXiv:2510.16817, 2025

    work page arXiv 2025

  12. [12]

    Imagenet classification with deep convolutional neural networks.Advances in neural information processing systems, 25, 2012

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks.Advances in neural information processing systems, 25, 2012

    work page 2012

  13. [13]

    Solving PDEs on spheres with physics-informed convolutional neural networks.Applied and Computational Harmonic Analysis, 74:101714, 2025

    Guanhang Lei, Zhen Lei, Lei Shi, Chenyu Zeng, and Ding-Xuan Zhou. Solving PDEs on spheres with physics-informed convolutional neural networks.Applied and Computational Harmonic Analysis, 74:101714, 2025

    work page 2025

  14. [14]

    Understanding deep convolutional networks.Philosophical Trans- actions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 374 (2065):20150203, 2016

    St´ ephane Mallat. Understanding deep convolutional networks.Philosophical Trans- actions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 374 (2065):20150203, 2016

    work page 2065

  15. [15]

    Consistent PINNs for higher-order elliptic PDEs.Inter- national Journal for Numerical Methods in Engineering, 127(7):e70320, 2026

    Shiv Mishra and Arbaz Khan. Consistent PINNs for higher-order elliptic PDEs.Inter- national Journal for Numerical Methods in Engineering, 127(7):e70320, 2026. 23

    work page 2026

  16. [16]

    A novel Galerkin method for solving PDEs on the sphere using highly localized kernel bases.Mathematics of Computation, 86(303):197–231, 2017

    Francis Narcowich, Stephen Rowe, and Joseph Ward. A novel Galerkin method for solving PDEs on the sphere using highly localized kernel bases.Mathematics of Computation, 86(303):197–231, 2017

    work page 2017

  17. [17]

    Diffusion models are minimax opti- mal distribution estimators

    Kazusato Oko, Shunta Akiyama, and Taiji Suzuki. Diffusion models are minimax opti- mal distribution estimators. InICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2023

    work page 2023

  18. [18]

    Springer Science & Business Media, 2012

    Allan Pinkus.N-widths in Approximation Theory, volume 7. Springer Science & Business Media, 2012

    work page 2012

  19. [19]

    Sharp lower bounds on the manifold widths of Sobolev and Besov spaces.Journal of Complexity, 85:101884, 2024

    Jonathan W Siegel. Sharp lower bounds on the manifold widths of Sobolev and Besov spaces.Journal of Complexity, 85:101884, 2024

    work page 2024

  20. [20]

    Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks.Computer Methods in Applied Mechanics and Engineering, 389:114333, 2022

    Natarajan Sukumar and Ankit Srivastava. Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks.Computer Methods in Applied Mechanics and Engineering, 389:114333, 2022

    work page 2022

  21. [21]

    Two-point homogeneous spaces.Annals of Mathematics, 55(1): 177–191, 1952

    Hsien-Chung Wang. Two-point homogeneous spaces.Annals of Mathematics, 55(1): 177–191, 1952

    work page 1952

  22. [22]

    Cambridge University Press, 2004

    Holger Wendland.Scattered Data Approximation, volume 17. Cambridge University Press, 2004

    work page 2004

  23. [23]

    On the rates of convergence for learning with convolutional neural networks.SIAM Journal on Mathematics of Data Science, 7 (4):1755–1772, 2025

    Yunfei Yang, Han Feng, and Ding-Xuan Zhou. On the rates of convergence for learning with convolutional neural networks.SIAM Journal on Mathematics of Data Science, 7 (4):1755–1772, 2025

    work page 2025

  24. [24]

    Deep distributed convolutional neural networks: Universality.Analysis and Applications, 16(06):895–919, 2018

    Ding-Xuan Zhou. Deep distributed convolutional neural networks: Universality.Analysis and Applications, 16(06):895–919, 2018

    work page 2018

  25. [25]

    Theory of deep convolutional neural networks: Downsampling.Neural Networks, 124:319–327, 2020

    Ding-Xuan Zhou. Theory of deep convolutional neural networks: Downsampling.Neural Networks, 124:319–327, 2020

    work page 2020

  26. [26]

    Universality of deep convolutional neural networks.Applied and Com- putational Harmonic Analysis, 48(2):787–794, 2020

    Ding-Xuan Zhou. Universality of deep convolutional neural networks.Applied and Com- putational Harmonic Analysis, 48(2):787–794, 2020

    work page 2020

  27. [27]

    Expressive power of deep networks on manifolds: Simultaneous approximation.arXiv preprint arXiv:2509.09362, 2025

    Hanfei Zhou and Lei Shi. Expressive power of deep networks on manifolds: Simultaneous approximation.arXiv preprint arXiv:2509.09362, 2025

    work page arXiv 2025

  28. [28]

    SSBE-PINN: A sobolev boundary scheme boosting stability and accuracy in elliptic/parabolic PDE learning.arXiv preprint arXiv:2508.10322, 2025

    Qixuan Zhou, Chuqi Chen, Tao Luo, and Yang Xiang. SSBE-PINN: A sobolev boundary scheme boosting stability and accuracy in elliptic/parabolic PDE learning.arXiv preprint arXiv:2508.10322, 2025. 24

    work page arXiv 2025