betaFeedback

WorksMathematics

Quantitative approximation for Fourier neural operators in controlled ordinary differential equations

Abstract

Fourier neural operators (FNOs) avoid the curse of parametric complexity when approximating the solution operator of certain partial differential equations (PDEs). While recent studies primarily focus on PDEs, in this article we derive upper error bounds for approximating the solution operator of controlled ordinary differential equations (CODEs) y˙=f(y,u)\dot{y}=f(y,u) using a FNO. Under a Carathéodory setting and suitable regularity assumptions, including a smallness condition, we prove that the size of the FNO grows at most algebraically in 1/ε1/\varepsilon, namely like ε−(1+d/(2(s−1)))\varepsilon^{-(1+d/(2(s-1)))} up to logarithmic factors for GELU networks, and is thus sub-quadratic for sufficiently smooth ff, where d=1+ny+nud=1+n_y+n_u and ss denotes the Sobolev regularity of ff. Our proof is based on emulating the Picard iteration through the layers of the FNO while exploiting the differentiation formula of the discrete Fourier transform. This result contributes to the quantitative approximation theory of FNOs for physically motivated problems and provides a theoretical foundation for leveraging a FNO as efficient forward surrogate for the solution operator of CODEs, especially in inverse parameter estimation where repeated evaluations of the forward operator are needed.

Introduction

Neural operators, acting between infinite-dimensional function spaces, constitute a natural extension of classical finite neural networks. In recent years, they have attracted considerable attention in the field of machine learning due to their ability to enabling surrogate modeling of high-dimensional and complex physical systems. Such systems can be modeled by partial differential equations (PDEs) or controlled ordinary differential equations (CODEs), whose numerical solving may become time consuming, particularly when repeated evaluations are required. Using neural operators to replace the solution operator of a PDE or CODE, where the latter also acts as an operator between function spaces due to the presence of control functions, may work out this issue. Once a neural operator is learned, it allows for a faster computational inference time while also maintaining an architecture that enables the processing of batched data. This also makes them an encouraging tool in solving (Bayesian) inverse problems via the Ensemble Kalman Inversion, especially in cases where a large data set of multiple experiments is present.

Various different architectures to realize a neural operator were introduced over the last years. One of the most promising ones is the so-called Fourier neural operator (FNO) which employs the (inverse) Fourier transform inside the network’s layered structure. While Universal Approximation Theorems of the FNO are well studied and suggest that any continuous operator can be approximated, quantitative approximation studies show that FNO faces exponential growth in model complexity, the so-called curse of parametric complexity [34], even when approximating CrC^r- or Lipschitz-regular operators. The curse of parametric complexity is an infinite-dimensional analogue to the well-known curse of dimensionality common in high-dimensional approximation problems.

Recent research suggests that this curse of parametric complexity can be overcome when approximating the solution operator of certain PDEs by emulating numerical methods like the Picard iteration. While current studies therefore expand the class of PDEs which allow for breaking the curse using a FNO, CODEs have not been studied in this context, although they play an important role in modeling physics while also being solvable through the Picard iteration. Therefore in this article, we focus on CODEs of form y˙=f(t,y,u)\dot{y}=f(t,y,u), where ff is a non-linear function and uu are the Lebesgue-integrable control functions.

This article starts with introducing a system of CODEs under a Carathéodory setting, that is that the controls uu do not need to be continuous, to determine appropriate mathematical assumptions to show the well-posedness of such a system while also deriving a continuous solution operator. This results in reformulating the CODE as a fixed point integral formula and thus, we follow the idea of emulating the Picard iteration by relying on the differentiation formula of the (discrete) Fourier transform to realize the integration of the controlled ordinary differential equation. Thus, the FNO constitutes a natural candidate for this approach. By sharpening our assumptions, which requires to assume more regularity on uu and ff, while for the latter also a smallness condition is needed, we break the curse by showing that the size of the FNO scales at most algebraically, and under a sufficiently smooth ff even sub-quadratically up to logarithmic factors.

Our article is organized as follows:

  • In § 2 we introduce the controlled ordinary differential equation y˙=f(t,y,u)\dot{y}=f(t,y,u) under suitable assumptions, such that we obtain the existence of a continuous control-to-solution operator. As we are interested in deriving error bounds for approximating this solution operator with a Fourier neural operator, we start to sharpening the assumptions on the CODE such that we can apply tools from Fourier analysis.

  • The Fourier neural operator is being formally introduced and defined in § 3. One essential key-point of the FNO is the usage of the (inverse) Fourier transform inside the network’s architecture. Since deriving error bounds requires to quantify the network’s size, we define the Ψ\Psi-FNO which relates to the numerical implementable version of a FNO. Afterwards, we state known quantitative approximation results regarding finite neural networks, as we need to approximate ff by such networks. At the end of § 3, we propose a Picard type fixed-point iteration scheme which produces an approximate output to the true solution of the CODE while also allowing to be emulated through the layers of Ψ\Psi-FNO.

  • In § 4 we establish the main result. Under the sharpened assumptions from § 2, which include a smallness condition on ff, we show that the size of the Ψ\Psi-FNO scales at most algebraically when approximating the control-to-solution operator. Thus, the size of the Ψ\Psi-FNO avoids the curse of parametric complexity.

  • To conclude the article, we present in § 5 the application of a learned FNO in an inverse parameter estimation problem motivated by a physical example. Due to the nature of the given example, it is the case that a large data set of multiple physical experiments is present. We therefore propose to utilize the learned FNO in an Ensemble Kalman Inversion which is extended to a Kaczmarz-like structure.

Related works

Neural operators [28] were first introduced in [3] by realizing them as graph kernel networks. By doing so, they offer a new approach to operator learning compared to other methods such as DeepONet [42] and PCA-Net [31]. Numerous neural operator architectures, such as the Fourier neural operator [39], the convolutional neural operator [49], the Laplace neural operator [11], the orthogonal polynomial neural operator [41] and the Wavelet neural operators [63], have been developed in recent years. Owing to their strong empirical approximation capabilities and fast inference times, they have established themselves as state-of-the-art tools for surrogate modeling physical simulations [64], which also makes them interesting objects to study in the context of inverse problems [68, 5, 69, 10, 56, 65, 45], due to the reliance on repeatedly evaluating the forward operator to numerically solve inverse problems. We refer the reader to [4, 29] for an overview of neural operators and to [47] for neural operators in the context of inverse problems.

One of the most important theoretical foundations in the field of machine learning are the Universal Approximation Theorems, which state that for a given continuous operator and any accuracy level, there exists a neural operator capable of approximating the target operator over a compact set to any desired degree of accuracy. These theorems further justify the application of neural operators. Multiple Universal Approximation Theorems have been established for various kinds of neural operators [28], including the FNO [27, 32], as well as for the DeepONet [33] and PCA-Net [31].

Analogous to the case of finite-dimensional neural networks, neural operators suffer from the curse of parametric complexity, the infinite-dimensional analogue to the curse of dimensionality. This curse is an exponential growth in size with respect to an increasing accuracy, which is also present when approximating Cr−C^{r-} or Lipschitz-regular operators as discussed in [34]. In [27], they overcome this curse when approximating the solution operator of the Darcy-type elliptic equation and the incompressible Navier–Stokes equations. The idea of the proof relies on recasting the underlying PDEs as fixed point problems which then allows for utilizing the Picard iteration. The latter is then realized by the layers of the investigated neural operator. The authors of [17, 16, 30] follow the same idea, leading to multiple neural operators that approximate solution operators of various PDEs that do not suffer from the curse of parametric complexity.

Although much encouraging work has been done to couple neural operators with ODEs [42, 24, 5, 40, 71], no effort has yet been made to derive quantitative approximation rates for approximating the solution operator of a controlled ODE with a Fourier neural operator. Since neural operators usually act on function spaces, the object of interest in the case of ODEs is the setting when so-called control functions are present, which then act as the input functions for the neural operator. We therefore refer to such ODEs as controlled ordinary differential equations (CODEs) and highlight [59, 12] for further reading on this topic. In this article, we derive error bounds for the Fourier neural operator when approximating the solution operator of CODEs. This is motivated by the fact that the differentiation formula of the Fourier transform suggests that the emulation of a CODE, when recast as a fixed point integral formula, is realizable through the layers of a FNO.

Such controlled systems can arise in practice, e.g., when controlling and modeling the process of electric arc furnaces [6, 44]. To obtain meaningful predictions, simulations of these furnaces require inverse parameter estimation [50, 43, 2]. As electric arc furnaces are run under different control settings, their operation generates large data sets that can be used for parameter estimation. However, the long inference times of complex physical simulations can hinder a successful inversion. We therefore propose to apply the Ensemble Kalman Inversion (EKI) where one additionally replaces the simulation by a FNO surrogate. Interestingly the authors of [30, 23] share the same idea of deploying the EKI with a learned neural operator. However, we extend this approach to a Kaczmarz structure, which is commonly encountered in deterministic inverse problems [36, 51, 70, 37]. For further reading on EKI we refer the reader to [22, 21, 20] and for inverse problems we suggest [14, 61, 52, 54, 53, 26].

Our results and contributions

After introducing the CODE system in a Carathéodory setting, we give appropriate assumptions that lead to a well-defined operator acting between the spaces of Lebesgue integrable and continuous functions. The corresponding result is summarized in Theorem 2.1. We then strengthen these assumptions to establish the setting for deriving error bounds when deploying the Fourier neural operator as a surrogate of the solution operator. Our main result Theorem 4.1 states that the size of the FNO scales at most algebraically and thus, exponential growth in the size of the FNO with respect to an increasing accuracy can be avoided. Under a sufficiently smooth ff the size in fact grows at most sub-quadratically up to logarithmic factors. Proving this relies on a reformulation of the CODE as a fixed point integral formula, while utilizing numerous results from Fourier analysis. The non-linearity ff in the CODE system can be efficiently approximated by a finite neural network. Consequently, the idea of our proof is similar to that of [17], implying that other problems with a comparable structure can be resolved in the same manner. At the end of the paper, we introduce an inverse parameter estimation problem in which a set of measurement data from multiple experiments is present. We propose solving this problem through surrogate modeling and applying the Ensemble Kalman Inversion in a Kaczmarz-type fashion.

Notation

Let us briefly introduce the notation used throughout this paper.

  • We set N={1,2,…}\mathbb{N}=\{1,2,\ldots\} and N0=N∪{0}\mathbb{N}_0=\mathbb{N}\cup\{0\}. Further we set R+=(0,∞)\mathbb{R}^{+}=(0,\infty).

  • Let Td=[0,2π]d\mathbb{T}^{d}=[0,2\pi]^{d}, d∈Nd\in\mathbb{N}, be the periodic domain which we identify with the standard torus.

  • By ∥⋅∥Rn\|\cdot\|_{\mathbb{R}^{n}}, n∈Nn\in\mathbb{N}, we denote the euclidean norm.

  • For q∈[1,∞]q\in[1,\infty] the Lebesgue space Lq(I;Rn)L^{q}(I;\mathbb{R}^{n}) consists of all measurable functions u:I→Rnu:I\to\mathbb{R}^{n} such that

∥u∥Lq(I;Rn):={(∫I∥u(x)∥Rnq dx)1/q,if q<∞,ess sup⁡x∈I∥u(x)∥Rn,if q=∞.\|u\|_{L^{q}(I;\mathbb{R}^{n})}:= \begin{cases} \left(\int_{I}\|u(x)\|_{\mathbb{R}^{n}}^{q}\,\mathrm{d}x\right)^{1/q}, & \text{if }q<\infty,\\ \underset{x\in I}{\operatorname{ess\,sup}}\|u(x)\|_{\mathbb{R}^{n}}, & \text{if }q=\infty. \end{cases}
  • By Cabs(I;Rn)C_{\mathrm{abs}}(I;\mathbb{R}^{n}) we denote the space of absolutely continuous functions endowed with the sup-norm. For 0≤s≤∞0 \le s \le\infty we denote by Cs(I;Rn)C^{s}(I;\mathbb{R}^{n}) the space of ss-times continuously differentiable functions.

  • For s∈N0s \in\mathbb{N}_{0}, p∈[1,∞]p \in[1,\infty] the Sobolev space Ws,p(I;Rn)W^{s,p}(I;\mathbb{R}^{n}) is defined by

Ws,p(I;Rn):={u∈Lp(I;Rn)∣Dαu∈Lp(I:Rn) for all α∈N0d with ∣α∣≤s},W^{s,p}(I;\mathbb{R}^{n}) := \{u \in L^{p}(I;\mathbb{R}^{n}) \mid D^{\alpha}u \in L^{p}(I:\mathbb{R}^{n}) \text{ for all } \alpha\in\mathbb{N}_{0}^{d} \text{ with } |\alpha| \le s\},

where DαuD^{\alpha}u denotes the weak derivative of order ss and α\alpha is a multi-index such that

α=(α1,…,αd),Dαu=(∂α1∂α1x1⋯∂αd∂αdxd)u,∣α∣=∑i=1dαi.\alpha= (\alpha_{1},\ldots,\alpha_{d}),\quad D^{\alpha}u = \left(\frac{\partial^{\alpha_{1}}}{\partial^{\alpha_{1}}x_{1}}\cdots\frac{\partial^{\alpha_{d}}}{\partial^{\alpha_{d}}x_{d}}\right)u,\quad|\alpha| = \sum_{i=1}^{d}\alpha_{i}.

The Ws,p(I;Rn)W^{s,p}(I;\mathbb{R}^{n})-norm is given by

∥u∥Ws,p(I;Rn):={(∑0≤∣α∣≤s∥Dαu∥Lp(I;Rn)p)1/p,if p<∞,max⁡0≤∣α∣≤s∥Dαu∥L∞(I;Rn),if p=∞.\lVert u\rVert_{W^{s,p}(I;\mathbb{R}^{n})}:= \begin{cases} \left(\sum_{0\le|\alpha|\le s}\lVert D^{\alpha}u\rVert_{L^{p}(I;\mathbb{R}^{n})}^{p}\right)^{1/p}, & \text{if }p<\infty,\\ \max_{0\le|\alpha|\le s}\lVert D^{\alpha}u\rVert_{L^{\infty}(I;\mathbb{R}^{n})}, & \text{if }p=\infty. \end{cases}

We define the notation Hs(I;Rn):=Ws,2(I;Rn)H^{s}(I;\mathbb{R}^{n}) := W^{s,2}(I;\mathbb{R}^{n}).

  • In general we write ∥f∥X(I)\lVert f\rVert_{\mathcal{X}(I)} for functions f∈X(I;R)f \in\mathcal{X}(I;\mathbb{R}). Thus, we also just write X(I)\mathcal{X}(I) instead of X(I;R)\mathcal{X}(I;\mathbb{R}).

All proofs of our results are found in the appendix of the respective chapters.

Well-posedness for controlled ordinary differential equations (CODE)

We are interested in systems of controlled ordinary differential equations, that is

y˙(t)=f(t,y(t),u(t)),t∈I:=(t1,t2)⊂R,−∞<t1<t2<∞,\dot{y}(t) = f(t,y(t),u(t)),\quad t \in I := (t_{1},t_{2}) \subset\mathbb{R},\quad-\infty< t_{1} < t_{2} < \infty,
y(t1)=y0,(1)y(t_{1}) = y^{0}, \tag*{(1)}

where f:I×Rny×Rnu→Rnyf : I \times\mathbb{R}^{n_{y}} \times\mathbb{R}^{n_{u}} \to\mathbb{R}^{n_{y}}, y:I→Rnyy : I \to\mathbb{R}^{n_{y}}, u:I→Rnuu : I \to\mathbb{R}^{n_{u}}, ny,nu∈Nn_{y},n_{u} \in\mathbb{N}. Further we assume that the control function uu does not necessarily need to be a continuous but a Lebesgue integrable function. Such setting is usually referred to as Carathéodory setting in literature. See [59], [12], [55], [48] and [7] for the study of controlled ODEs. We introduce standard assumptions, under which the system is well posed. By well-posedness we mean that for every control function uu and initial value y0y^{0} there exists a unique solution yy of Equation (1). By slightly extending the assumptions we obtain continuity of the solution operator, i.e. the mapping

(u,y0)↦y(u,y^{0}) \mapsto y

is continuous with respect to appropriately chosen function spaces and their respective norms. Note that the derivative and integration are understood component-wise throughout this paper.

References

  1. [1]H.W. Alt and R. Nürn­berg. Linear Functional Analysis: An Application-Oriented Introduction. Universitext. Springer-Verlag Berlin Heidelberg, 2016. ISBN: 9781447172802. URL: https://books.google.de/books?id=mzilDAAAQBAJ.
  2. [2]Mauricio Álvarez, Carlos Henao Baena, and Jesser Marulanda-Durango. “Calibration of parameters for electric arc furnace model using simulation and neural networks”. In: Revista EIA 11.22 (Dec. 2014), pp. 35–44.
  3. [3]Anima Anandkumar et al. “Neural Operator: Graph Kernel Network for Partial Differential Equations”. In: ICLR 2020 Workshop on Integration of Deep Neural Models and Differential Equations. 2019. URL: https://openreview.net/forum?id=fG2ZFmXF03.arxiv.org/abs/2003.03485
  4. [4]Kamyar Azizadehnesheli et al. “Neural operators for accelerating scientific simulations and design”. In: Nature Reviews Physics 6.5 (2024), pp. 320–328.
  5. [5]Abdolmehdi Behroozi, Chaopeng Shen, and Daniel Kifer. “Sensitivity-Constrained Fourier Neural Operators for Forward and Inverse Problems in Parametric Differential Equations”. In: The Thirteenth International Conference on Learning Representations. 2025. URL: https://openreview.net/forum?id=DPzQ5n3mNm.arxiv.org/abs/2505.08740
  6. [6]Johannes Gerhardt Bekker, Jan Karel Craig, and Petrus Christiaan Pistorius. “Modeling and Simulation of an Electric Arc Furnace Process”. In: ISIJ International 39.1 (1999), pp. 23–32. DOI: 10.2355/isijinternational.39.23.DOI
  7. [7]Michał Bełdziński, Marek Galewski, and Igor Kossowski. “Dependence on parameters for nonlinear equations—Abstract principles and applications”. In: Mathematical Methods in the Applied Sciences 45.3 (2022), pp. 1668–1686. DOI: 10.1002/mma.7882. URL: https://onlinelibrary.wiley.com/doi/abs/10.1002/mma.7882.
  8. [8]H. Brezis. Functional Analysis, Sobolev Spaces and Partial Differential Equations. Universitext. Springer New York, 2010. ISBN: 9780387097147. URL: https://books.google.de/books?id=1bFCAAAQBAJ.DOI
  9. [9]Jené Brynjarsdóttir and Anthony O’Hagan. “Learning about physical parameters: the importance of model discrepancy”. In: Inverse Problems 30 (Oct. 2014). DOI: 10.1088/0266-5611/30/11/114007.DOI
  10. [10]Lianghao Cao et al. “Residual-based error correction for neural operator accelerated infinite-dimensional Bayesian inverse problems”. In: Journal of Computational Physics 486 (Aug. 2023), p. 112104. ISSN: 0021-9991. DOI: 10.1016/j.jcp.2023.112104. URL: http://dx.doi.org/10.1016/j.jcp.2023.112104.
  11. [11]Qianying Cao, Somdatta Goswami, and George Em Karniadakis. “Laplace neural operator for solving differential equations”. In: Nature Machine Intelligence 6.6 (2024), pp. 631–640.DOI
  12. [12]Paulo M. de Carvalho-Neto, Cícero L. Frota, and Pedro G. P. Torelli. A General Version of Carathéodory’s Existence and Uniqueness Theorem. 2025. arXiv: 2505.24516 [math.CA]. URL: https://arxiv.org/abs/2505.24516.
  13. [13]M.E. Causon et al. “Real-time Bayesian inversion in resin transfer moulding using neural surrogates”. In: Composites Part A: Applied Science and Manufacturing 185 (2024), p. 108355. ISSN: 1359-835X. DOI: https://doi.org/10.1016/j.compositesa.2024.108355. URL: https://www.sciencedirect.com/science/article/pii/S1359835X2400352X.
  14. [14]Christian Clason. Regularization of Inverse Problems. 2021. arXiv: 2001.00617 [math.FA]. URL: https://arxiv.org/abs/2001.00617.
  15. [15]Alexandre Ern and Jean-Luc Guermond. “Mollification in Strongly Lipschitz Domains with Application to Continuous and Discrete De Rham Complexes”. In: Computational Methods in Applied Mathematics 16.1 (2016), pp. 51–75. DOI: 10.1515/cmam-2015-0034. URL: https://doi.org/10.1515/cmam-2015-0034.
  16. [16]Takashi Furuya, Ryo Ozawa, and Jenn-Nan Wang. Approximation Theory of Laplacian-Based Neural Operators for Reaction-Diffusion System. 2026. arXiv: 2605.12025 [cs.LG]. URL: https://arxiv.org/abs/2605.12025.
  17. [17]Takashi Furuya, Koichi Taniguchi, and Satoshi Okuda. “Quantitative approximation for neural operators in nonlinear parabolic equations”. In: International Conference on Learning Representations. 2025, pp. 85528–85556.arxiv.org/abs/2410.02151
  18. [18]Ingo Gü​hring, Gitta Kutyniok, and Philipp Petersen. “Error bounds for approximations with deep ReLU neural networks in Ws,p norms”. In: Analysis and Applications 18.05 (2020), pp. 803–859. DOI: 10.1142/S0219530519410021. eprint: https://doi.org/10.1142/S0219530519410021. URL: https://doi.org/10.1142/S0219530519410021.
  19. [19]Dan Hendrycks and Kevin Gimpel. Gaussian Error Linear Units (GELUs). 2023. arXiv: 1606.08415 [cs.LG]. URL: https://arxiv.org/abs/1606.08415.
  20. [20]Marco Iglesias and Yüchen Yang. Adaptive regularisation for ensemble Kalman inversion. 2020. arXiv: 2006.14980 [math.NA]. URL: https://arxiv.org/abs/2006.14980.
  21. [21]Marco A Iglesias. “A regularizing iterative ensemble Kalman method for PDE-constrained inverse problems”. In: Inverse Problems 32.2 (Jan. 2016), p. 025002. ISSN: 1361-6420. DOI: 10.1088/0266-5611/32/2/025002. URL: http://dx.doi.org/10.1088/0266-5611/32/2/025002.
  22. [22]Marco A Iglesias, Kody J H Law, and Andrew M Stuart. “Ensemble Kalman methods for inverse problems”. In: Inverse Problems 29.4 (Mar. 2013), p. 045001. ISSN: 1361-6420. DOI: 10.1088/0266-5611/29/4/045001. URL: http://dx.doi.org/10.1088/0266-5611/29/4/045001.
  23. [23]Marco A Iglesias et al. “DeepONet-accelerated Bayesian inversion for moving boundary problems”. In: Computer Methods in Applied Mechanics and Engineering 457 (2026), p. 118988.
  24. [24]Pengzhan Jin, Shuai Meng, and Lu Lu. “MIONet: Learning multiple-input operators via tensor product”. In: SIAM Journal on Scientific Computing 44.6 (2022), A3490–A3514.arxiv.org/abs/2202.06137
  25. [25]Marc C. Kennedy and Anthony O’Hagan. “Bayesian calibration of computer models”. In: Journal of the Royal Statistical Society: Series B (Statistical Methodology) 63 (2001). URL: https://api.semanticscholar.org/CorpusID:119562136.DOI
  26. [26]Rebecca Klein, Thomas Schuster, and Anne Wald. “Sequential subspace optimization for recovering stored energy functions in hyperelastic materials from time-dependent data”. In: Time-dependent problems in imaging and parameter identification. Springer, 2021, pp. 165–190.
  27. [27]Nikola Kovachki, Samuel Lanthaler, and Siddhartha Mishra. “On universal approximation and error bounds for Fourier neural operators”. In: Journal of Machine Learning Research 22.290 (2021), pp. 1–76.arxiv.org/abs/2107.07562
  28. [28]Nikola Kovachki et al. “Neural operator: learning maps between function spaces with applications to PDEs”. In: J. Mach. Learn. Res. 24.1 (Jan. 2023). ISSN: 1532-4435.DOI
  29. [29]Nikola B Kovachki, Samuel Lanthaler, and Andrew M Stuart. “Operator learning: Algorithms and analysis”. In: Handbook of Numerical Analysis 25 (2024), pp. 419–467.arxiv.org/abs/2402.15715
  30. [30]Pu-Yun Kwok et al. Quantitative approximation theorem of neural operators for the Dirichlet-to-Neumann map. 2026. arXiv: 2610.35404 [math.AP]. URL: https://arxiv.org/abs/2610.35404.
  31. [31]Samuel Lanthaler. “Operator learning with PCA-Net: upper and lower complexity bounds”. In: Journal of Machine Learning Research 24.318 (2023), pp. 1–67.arxiv.org/abs/2303.16317
  32. [32]Samuel Lanthaler, Zongyi Li, and Andrew M Stuart. “Nonlocality and nonlinearity implies universality in operator learning”. In: Constructive Approximation 62.2 (2025), pp. 261–303.arxiv.org/abs/2304.13221
  33. [33]Samuel Lanthaler, Siddhartha Mishra, and George E Karniadakis. “Error estimates for deeponets: A deep learning framework in infinite dimensions”. In: Transactions of Mathematics and its Applications 6.1 (2022), tnac001.arxiv.org/abs/2102.09618
  34. [34]Samuel Lanthaler and Andrew M Stuart. “The parametric complexity of operator learning”. In: IMA Journal of Numerical Analysis 46.2 (2026), pp. 647–712.arxiv.org/abs/2306.15924
  35. [35]B. Laurent and P. Massart. “Adaptive estimation of a quadratic functional by model selection”. In: The Annals of Statistics 28.5 (2000), pp. 1302–1338. DOI: 10.1214/aos/1015957395. URL: https://doi.org/10.1214/aos/1015957395.
  36. [36]A. Leitão and B F Svaiter. “On projective Landweber–Kaczmarz methods for solving systems of nonlinear ill-posed equations”. In: Inverse Problems 32.2 (Jan. 2016), p. 025004. ISSN: 1361-6420. DOI: 10.1088/0266-5611/32/2/025004. URL: http://dx.doi.org/10.1088/0266-5611/32/2/025004.
  37. [37]Housen Li and Markus Haltmeier. “The Averaged Kaczmarz Iteration for Solving Inverse Problems”. In: SIAM Journal on Imaging Sciences 11.1 (Jan. 2018), pp. 618–642. ISSN: 1936-4954. DOI: 10.1137/17m1146178. URL: http://dx.doi.org/10.1137/17m1146178.
  38. [38]Jiyong Li. “Uniform error bounds of a nested Picard iterative integrator for the Klein-Gordon-Zakharov system in the subsonic limit regime”. In: Advances in Computational Mathematics 51 (July 2025). DOI: 10.1007/s10444-025-02051-x.DOI
  39. [39]Zongyi Li et al. “Fourier Neural Operator for Parametric Partial Differential Equations”. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. URL: https://openreview.net/forum?id=c8P9NQVtmn0.
  40. [40]Zhi-Song Li et al. Inverse Neural Operator for PDE Parameter Optimization. 2026. arXiv: 2603.11854 [cs.LG]. URL: https://arxiv.org/abs/2603.11854.
  41. [41]Ziyuan Liu et al. “Render unto numerics: Orthogonal polynomial neural operator for PDEs with nonperiodic boundary conditions”. In: SIAM Journal on Scientific Computing 46.4 (2024), pp. C323–C348.arxiv.org/abs/2206.12698
  42. [42]Lu Lu et al. “Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators”. In: Nature Machine Intelligence 3.3 (Mar. 2021), pp. 218–229. ISSN: 2522-5839. DOI: 10.1038/s42256-021-00302-5. URL: http://dx.doi.org/10.1038/s42256-021-00302-5.
  43. [43]Jesser Marulanda-Durango and C.D. Zuluaga-Ríos. “A Meta-heuristic Optimization-based Method for Parameter Estimation of an Electric Arc Furnace Model”. In: Results in Engineering 17 (Dec. 2022), p. 100850. DOI: 10.1016/j.rineng.2022.100850.DOI
  44. [44]Thomas Meier et al. “Process Modeling and Simulation of the Radiation in the Electric Arc Furnace”. In: steel research international 89.4 (2018), p. 1700487. DOI: 10.1002/srin.201700487. URL: https://onlinelibrary.wiley.com/doi/abs/10.1002/srin.201700487.
  45. [45]Roberto Molinaro et al. “Neural Inverse Operators for Solving PDE Inverse Problems”. In: International Conference on Machine Learning. 2023. URL: https://api.semanticscholar.org/CorpusID:256274656.arxiv.org/abs/2301.11167
  46. [46]François Murat and Cristina Trombetti. “A chain rule formula for the composition of a vector-valued function by a piecewise smooth function”. In: Bollettino della Unione Matematica Italiana. Serie VIII. Sezione B. Articoli di Ricerca Matematica 3 (Oct. 2003).
  47. [47]Derick Nganyu Tanyu et al. “Deep learning methods for partial differential equations and related parameter identification problems”. In: Inverse Problems 39.10 (Aug. 2023), p. 103001. ISSN: 1361-6420. DOI: 10.1088/1361-6420/ace9d4. URL: http://dx.doi.org/10.1088/1361-6420/ace9d4.
  48. [48]D. O’Regan. Existence Theory for Nonlinear Ordinary Differential Equations. Mathematics and Its Applications. Springer Netherlands, 2013. ISBN: 9789401715171. URL: https://books.google.de/books?id=ijjpCAAAQBAJ.
  49. [49]Bogdan Raonic et al. “Convolutional neural operators for robust and accurate learning of pdes”. In: Advances in Neural Information Processing Systems 36 (2023), pp. 77187–77200.arxiv.org/abs/2302.01178
  50. [50]P.L. Rathaba et al. “Identifiability of an Electric Arc Furnace Model”. In: IFAC Proceedings Volumes 36.24 (2003). IFAC Workshop on New Technologies for Automation of the Metallurgical Industry 2003, Shanghai, PR China, 11-13 October 2003, pp. 89–94. ISSN: 1474-6670. DOI: 10.1016/S1474-6670(17)37609-7. URL: https://www.sciencedirect.com/science/article/pii/S1474667017376097.DOI
  51. [51]Kowar Richard and O. Scherzer. “Convergence analysis of a Landweber-Kaczmarz method for solving nonlinear ill-posed problems”. In: Ill-Posed and Inverse Problems 23 (Jan. 2002).DOI
  52. [52]Andreas Rieder and Thomas Schuster. “The Approximate Inverse in Action II: Convergence and Stability”. In: Mathematics of Computation 72.243 (2003), pp. 1399–1415. ISSN: 00255718, 10886842. URL: http://www.jstor.org/stable/409980 (visited on 11/09/2026).DOI
  53. [53]Dimitri Rothermel and Thomas Schuster. “Solving an inverse heat convection problem with an implicit forward operator by using a projected quasi-Newton method”. In: Inverse Problems 37 (2020). URL: https://api.semanticscholar.org/CorpusID:222133406.
  54. [54]Dimitri Rothermel et al. Parameter estimation of temperature dependent material parameters in the cooling process of TMCP steel plates. 2019. arXiv: 1912.12166 [math.NA]. URL: https://arxiv.org/abs/1912.12166.
  55. [55]Tomáš Roubíček. “Preliminary general material”. In: Nonlinear Partial Differential Equations with Applications. Basel: Springer Basel, 2013, pp. 1–27. ISBN: 978-3-0348-0513-1. DOI: 10.1007/978-3-0348-0513-1_1. URL: https://doi.org/10.1007/978-3-0348-0513-1_1.
  56. [56]Otmar Scherzer, Thi Lan Nhi Vu, and Jikai Yan. “Neural operators for solving nonlinear inverse problems”. In: Numerical Functional Analysis and Optimization 47.11-14 (2026), pp. 315–356.arxiv.org/abs/2508.19347
  57. [57]Claudia Schillings and Andrew M. Stuart. “Analysis of the ensemble Kalman filter for inverse problems”. In: SIAM Journal on Numerical Analysis 55.3 (2017), pp. 1264–1290. DOI: 10.1137/16M105959X.DOI
  58. [58]Jie Shen, Tao Tang, and Li-Lian Wang. Spectral Methods: Algorithms, Analysis and Applications. Vol. 41. Springer Series in Computational Mathematics. Springer, Jan. 2011. ISBN: 978-3-540-71040-0. DOI: 10.1007/978-3-540-71041-7.DOI
  59. [59]Eduardo D. Sontag. “Systems”. In: Mathematical Control Theory: Deterministic Finite Dimensional Systems. New York, NY: Springer New York, 1998, pp. 25–80. ISBN: 978-1-4612-0577-7. DOI: 10.1007/978-1-4612-0577-7_2. URL: https://doi.org/10.1007/978-1-4612-0577-7_2.
  60. [60]Elias M. Stein. Singular Integrals and Differentiability Properties of Functions (PMS-30). Princeton University Press, 1970. ISBN: 9780691080796. URL: http://www.jstor.org/stable/j.ctt1bpmb07 (visited on 10/31/2026).
  61. [61]Albert Tarantola. Inverse Problem Theory and Methods for Model Parameter Estimation. Society for Industrial and Applied Mathematics, 2005. DOI: 10.1137/1.9780898717921. URL: https://epubs.siam.org/doi/abs/10.1137/1.9780898717921.DOI
  62. [62]H. Triebel. Interpolation Theory, Function Spaces, Differential Operators. Huethig & Wepf, 1995. ISBN: 9783335004202. URL: https://books.google.de/books?id=ZzvvAAAAMAAJ.
  63. [63]Tapas Tripura and Souvik Chakraborty. “Wavelet Neural Operator for solving parametric partial differential equations in computational mechanics problems”. In: Computer Methods in Applied Mechanics and Engineering 404 (Feb. 2023), p. 115783. DOI: 10.1016/j.cma.2022.115783.DOI
  64. [64]Jakob Elias Wagner et al. “Neural Operators as Fast Surrogate Models for the Transmission Loss of Parameterized Sonic Crystals”. In: NeurIPS 2024 Workshop on Data-Driven and Differentiable Simulations, Surrogates, and Solvers. 2024. URL: https://openreview.net/forum?id=1tXlQvjhjt.
  65. [65]Tian Wang and Chuang Wang. “Latent neural operator for solving forward and inverse pde problems”. In: Advances in Neural Information Processing Systems 37 (2024), pp. 33085–33107.arxiv.org/abs/2406.03923
  66. [66]Simon Weissmann. “Gradient flow structure and convergence analysis of the ensemble Kalman inversion for nonlinear forward models”. In: Inverse Problems 38.10 (2022), p. 105011.DOI
  67. [67]Yahong Yang and Juncai He. Deep Neural Networks with General Activations: Super-Convergence in Sobolev Norms. 2026. arXiv: 2508.05141 [cs.LG]. URL: https://arxiv.org/abs/2508.05141.
  68. [68]Ziyi Yin et al. “Learned coupled inversion for carbon sequestration monitoring and forecasting with Fourier neural operators”. In: Second International Meeting for Applied Geoscience and Energy. SEG Technical Program Expanded Abstracts. Aug. 2022, pp. 467–472. DOI: 10.1190/image2022-3722848.1. URL: https://doi.org/10.1190/image2022-3722848.1.
  69. [69]Ziyi Yin et al. “Solving multiphysics-based inverse problems with learned surrogates and constraints”. In: Advanced Modeling and Simulation in Engineering Sciences 10.1 (Oct. 2023). ISSN: 2213-7467. DOI: 10.1186/s40323-023-00252-0. URL: http://dx.doi.org/10.1186/s40323-023-00252-0.
  70. [70]Nan Yu et al. “Landweber–Kaczmarz method for inverse problems using multiple repeated measurements”. In: Inverse Problems and Imaging 20.0 (2026), pp. 1–24. ISSN: 1930-8337. DOI: 10.3934/ipi.2025016. URL: https://www.aimsciences.org/article/id/67f8db1e45e64536d7529a5f.DOI
  71. [71]Mingyuan Zhou et al. “Parameter estimation of structural dynamics with neural operators enabled surrogate modeling”. In: Mechanical Systems and Signal Processing 237 (2025), p. 112914.arxiv.org/abs/2410.11712

Paper details

vlnc:W_6wut3gva
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents