arxiv: 2604.16975 · v1 · submitted 2026-04-18 · 🧮 math.NA · cs.LG· cs.NA· stat.ML

Recognition: unknown

Convergence theory for Hermite approximations under adaptive coordinate transformations

Yahya Saleh

Authors on Pith no claims yet

Pith reviewed 2026-05-10 06:47 UTC · model grok-4.3

classification 🧮 math.NA cs.LGcs.NAstat.ML

keywords Hermite expansionsadaptive coordinate transformationsnormalizing flowsspectral convergenceerror estimatespullback functionsmonotone transport mapsspectral methods

0 comments

The pith

Approximating a function with a transformed Hermite basis equals approximating its pullback with ordinary Hermite functions.

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

The paper gives the first error estimates for Hermite spectral methods that incorporate adaptive coordinate transformations. It proves an equivalence: the problem of approximating a target function f in the span of the transformed basis is identical to the standard problem of approximating the pullback of f in the ordinary Hermite basis. This reduction lets the authors import classical Hermite approximation rates that depend only on the smoothness and decay of the pullback. The authors construct an explicit monotone transport map for smooth functions with exponential decay that restores spectral accuracy. The result supplies the missing theory behind recent numerical work that uses normalizing flows to adapt coordinates in spectral calculations.

Core claim

Approximating a function f in the span of the transformed basis is equivalent to approximating the pullback of f in the span of Hermite functions. This allows us to leverage the classical approximation theory of Hermite expansions to derive error estimates in transformed coordinates in terms of the regularity of the pullback.

What carries the argument

The equivalence principle that reduces approximation in the transformed Hermite span to standard Hermite approximation of the pullback function.

If this is right

Error bounds for the adaptive approximation follow directly from the regularity and decay of the pullback under classical Hermite theory.
Spectral convergence is recovered once the monotone map aligns the function's decay with the weight of the Hermite basis.
The estimates hold for any monotone transport map that produces a sufficiently regular pullback, not only those produced by normalizing flows.
Convergence rates in the original variables are controlled by how smooth and rapidly decaying the pullback turns out to be.

Where Pith is reading between the lines

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

Optimizing the coordinate map is equivalent to maximizing the regularity of the pullback, which could be turned into a practical training objective.
The same equivalence argument could be carried over to other orthogonal bases, such as Laguerre functions on the half-line, with suitable transport maps.
In higher dimensions the construction of a transport map that keeps the pullback regular becomes the main remaining obstacle to fast convergence.

Load-bearing premise

The coordinate transformation is a monotone transport map that aligns the target function's decay with the Hermite basis while keeping the pullback sufficiently regular for classical Hermite theory to apply.

What would settle it

Choose a concrete smooth exponentially decaying function, apply a nonlinear monotone map, expand the pullback in the standard Hermite series, and check whether the approximation error measured in the original coordinates matches the rate predicted by the pullback's Hermite coefficients.

Figures

Figures reproduced from arXiv: 2604.16975 by Yahya Saleh.

**Figure 1.** Figure 1: Shown is the relative error |E˜m−E ref m | |Eref m | in the computation of 23 eigenvalues of the Morse oscillator (1.3) using Hermite approximations (blue circles), Hermite approximations employing a linear coordinate transformation g (red triangles), and Hermite approximations employing a nonlinear coordinate transformation g parameterized by a normalizing flow (green squares) as a function of the trunca… view at source ↗

**Figure 2.** Figure 2: Empirical convergence of function approximations using Hermite [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Figure 3: displays the computed moments as a function of s. The results indicate that Ψ0 exhibits super-Gaussian decay, while Wg −1 l Ψ0 and Wg −1 r Ψ0 exhibit a slower decay. Interestingly, Wg −1 r Ψ0 aligns more closely with the Gaussian behavior than Wg −1 l Ψ0, which is consistent with the superior performance of the iResNet-based Hermite basis compared to the linear-based basis. These findings support our hypot… view at source ↗

read the original abstract

Recent work has shown that parameterizing and optimizing coordinate transformations using normalizing flows, i.e., invertible neural networks, can significantly accelerate the convergence of spectral approximations. We present the first error estimates for approximating functions using Hermite expansions composed with adaptive coordinate transformations. Our analysis establishes an equivalence principle: approximating a function $f$ in the span of the transformed basis is equivalent to approximating the pullback of $f$ in the span of Hermite functions. This allows us to leverage the classical approximation theory of Hermite expansions to derive error estimates in transformed coordinates in terms of the regularity of the pullback. We present an example demonstrating how a nonlinear coordinate transformation can enhance the convergence of Hermite expansions. Focusing on smooth functions decaying along the real axis, we construct a monotone transport map that aligns the decay of the target function with the Hermite basis. This guarantees spectral convergence rates for the corresponding Hermite expansion. Our analysis provides theoretical insight into the convergence behavior of adaptive Hermite approximations based on normalizing flows, as recently explored in the computational quantum physics literature.

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.

Desk Editor's Note private letter to a colleague

This paper gives the first error estimates for Hermite approximations after adaptive coordinate changes by reducing the problem to classical theory via pullback equivalence and constructing a monotone map that enforces spectral rates.

read the letter

The paper supplies the first rigorous error estimates for Hermite spectral approximations composed with adaptive coordinate transformations. It does this by establishing an equivalence between the transformed approximation error and the standard Hermite error on the pullback function, then leveraging known rates once the pullback has enough smoothness and decay. The central new element is the explicit statement of this equivalence for error analysis plus the construction of a monotone transport map that aligns decay for smooth functions on the real line. The example shows a nonlinear map improving convergence in a concrete case. The work stays close to classical approximation theory and supplies theoretical grounding for normalizing-flow techniques already used in computational quantum physics without overclaiming. The soft spots are limited and mostly about verification. The argument assumes the pullback preserves the needed regularity under the map, and while the construction aims to guarantee this, the full proofs would need checking for boundary terms or edge cases where the learned flow might violate the assumptions. No load-bearing circularity or invented entities appear. This is for numerical analysts and computational physicists working on spectral methods for decaying functions on unbounded domains. A reader who already knows Hermite expansions and change-of-variables arguments will get immediate value and can use the map construction directly. It deserves a serious referee because the reduction is clean, the map is explicit, and the field needs this justification for methods that have been running heuristically. I recommend sending it to peer review.

Referee Report

2 major / 3 minor

Summary. The paper claims to provide the first error estimates for Hermite spectral approximations composed with adaptive coordinate transformations parameterized by normalizing flows. It establishes an equivalence principle: the best approximation error of a function f from the span of the transformed Hermite basis equals the best approximation error of its pullback f ∘ φ from the standard Hermite span in the weighted L2 space. Classical Hermite approximation theory then yields rates in terms of the pullback's regularity. For smooth functions with Gaussian decay, the paper constructs an explicit monotone transport map that aligns the decay to guarantee spectral convergence, offering insight into empirically successful adaptive methods in quantum physics.

Significance. If the equivalence and regularity preservation hold, the work supplies a rigorous theoretical foundation for adaptive Hermite methods that have shown practical acceleration in computational physics. By reducing the transformed problem to a classical, parameter-free result on Hermite expansions, it explains observed convergence improvements through decay alignment and provides a template for analyzing other ML-parameterized bases. The explicit transport map construction demonstrates how nonlinear changes of variables can restore spectral rates without introducing fitted parameters.

major comments (2)

[Equivalence principle and transport map construction] The equivalence principle (stated in the abstract and developed in the main argument) reduces the transformed approximation to classical Hermite theory, but the manuscript must explicitly verify that the pullback under the constructed monotone map remains in the weighted Sobolev space required for the classical rates (e.g., sufficient smoothness plus Gaussian decay). This assumption is load-bearing for all derived error estimates; without the verification, the claimed spectral rates do not necessarily follow.
[Discussion of normalizing flows] The analysis supplies insight into normalizing-flow-based adaptations, yet the error estimates are derived for an explicitly constructed ideal map rather than for a learned flow. The manuscript should address how the approximation error of the flow itself affects the overall rate, or clarify that the theory applies only in the limit of perfect alignment; otherwise the connection to the motivating computational literature remains incomplete.

minor comments (3)

The abstract asserts that the work presents 'the first error estimates' but should cite the specific classical Hermite results (e.g., the theorem on spectral convergence for functions in weighted Sobolev spaces) being invoked, to make the reduction fully traceable.
[Example demonstrating nonlinear coordinate transformation] In the example section demonstrating the nonlinear transformation, include a quantitative comparison (e.g., tabulated or plotted convergence rates) between the transformed and untransformed Hermite expansions for the chosen test function, to illustrate the improvement concretely.
Notation for the transformed basis functions and the pullback operator should be introduced with a short table or diagram early in the paper to aid readability, especially when switching between original and transformed coordinates.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive review and the recommendation for minor revision. The comments identify key points that strengthen the connection between the equivalence principle, the explicit construction, and the motivating applications. We address each major comment below and will incorporate the suggested clarifications and verifications into the revised manuscript.

read point-by-point responses

Referee: [Equivalence principle and transport map construction] The equivalence principle (stated in the abstract and developed in the main argument) reduces the transformed approximation to classical Hermite theory, but the manuscript must explicitly verify that the pullback under the constructed monotone map remains in the weighted Sobolev space required for the classical rates (e.g., sufficient smoothness plus Gaussian decay). This assumption is load-bearing for all derived error estimates; without the verification, the claimed spectral rates do not necessarily follow.

Authors: We agree that an explicit verification of the pullback's membership in the requisite weighted Sobolev space is necessary to rigorously justify the spectral rates. The construction in Section 4 is designed so that the monotone transport map aligns the Gaussian decay while preserving smoothness, but we will add a new lemma (or proposition) that directly confirms the pullback satisfies the weighted Sobolev regularity conditions used in the classical Hermite theory. This addition will make the load-bearing assumption fully explicit and support all subsequent error bounds. revision: yes
Referee: [Discussion of normalizing flows] The analysis supplies insight into normalizing-flow-based adaptations, yet the error estimates are derived for an explicitly constructed ideal map rather than for a learned flow. The manuscript should address how the approximation error of the flow itself affects the overall rate, or clarify that the theory applies only in the limit of perfect alignment; otherwise the connection to the motivating computational literature remains incomplete.

Authors: We accept that the present error estimates apply to the ideal (perfectly aligned) transport map. In the revised discussion, we will explicitly state that the theory characterizes the limiting case of perfect alignment and that, for a learned normalizing flow, the total approximation error decomposes into the pullback approximation error plus a term controlled by the flow's approximation quality to the ideal map. We will note that a quantitative bound on the flow error lies beyond the scope of the current work but represents a natural extension; this clarification will better connect the results to the computational literature while preserving the focus on the mechanism of decay alignment. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper establishes a direct equivalence via change of variables: the approximation error of f in the transformed Hermite basis equals the error of the pullback f ∘ φ in the standard Hermite span (in the weighted L2 space). Classical Hermite approximation theory then supplies the rates once the pullback's regularity and decay are verified. The monotone transport map is constructed explicitly from the target function's smoothness and decay properties, without fitted parameters, self-definitions, or load-bearing self-citations. The derivation is self-contained and reduces to externally known results on Hermite expansions.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The analysis rests on the classical approximation theory of Hermite expansions (standard in the literature) and the assumption that the learned coordinate transformation is a diffeomorphism preserving the necessary decay and regularity properties. No new free parameters or invented entities are introduced in the abstract.

axioms (2)

standard math Classical Hermite approximation theory applies directly to the pullback function once the coordinate change is performed.
Invoked to transfer error bounds from the transformed problem to the untransformed Hermite setting.
domain assumption The adaptive coordinate transformation is a monotone transport map that aligns decay rates without destroying smoothness.
Required for the concrete example that guarantees spectral convergence.

pith-pipeline@v0.9.0 · 5480 in / 1411 out tokens · 29889 ms · 2026-05-10T06:47:05.701554+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

32 extracted references · 15 canonical work pages · 1 internal anchor

[1]

Lubich: From quantum to classical molecular dynamics: reduced models and numerical analysis

C. Lubich: From quantum to classical molecular dynamics: reduced models and numerical analysis. European Mathematical Society (2008)

2008
[2]

Yurchenko: Computational Spectroscopy of Polyatomic Molecules

S. Yurchenko: Computational Spectroscopy of Polyatomic Molecules. CRC Press (2023)

2023
[3]

Boyd: Dynamics of the Equatorial Ocean

J. Boyd: Dynamics of the Equatorial Ocean. Springer (2018)

2018
[4]

Saleh, Á

Y. Saleh, Á. Fernández Corral, E. Vogt, A. Iske, J. Küpper, and A. Yachmenev: Computing Excited States of Molecules Using Normalizing Flows. J. Chem. Theory Comput.21, 5221–5229 (2025).https://doi.org/10.1021/acs.jctc.5c00590

work page doi:10.1021/acs.jctc.5c00590 2025
[5]

E. Vogt, Á. Fernández Corral, Y. Saleh, and A. Yachmenev: Transferability and interpretability of vibrational normalizing-flow coordinates. J. Chem. Phys.163, 154106 (2025).https://doi. org/10.1063/5.0285954

work page doi:10.1063/5.0285954 2025
[6]

Zhang, R.-S

Q. Zhang, R.-S. Wang, and L. Wang: Neural canonical transformations for vibrational spectra of molecules. J. Chem. Phys.161, 024103 (2024).https://doi.org/https://doi.org/10.1063/ 5.0209255

2024
[7]

Z. Li, H. Xie, X. Dong, and L. Wang: Deep Variational Free Energy Calculation of Hydrogen Hugoniot. Phys. Rev. Lett.136, 076504 (2026).https://doi.org/10.1103/8zn5-6dnt

work page doi:10.1103/8zn5-6dnt 2026
[8]

Zhang, X

Q. Zhang, X. Wang, R. Shi, X. Ren, H. Wang, and L. Wang: Neural Canonical Transformations for Quantum Anharmonic Solids of Lithium. Phys. Rev. Lett.134, 246101 (2025). https : //doi.org/10.1103/p3th-25bc

work page doi:10.1103/p3th-25bc 2025
[9]

H. Xie, L. Zhang, and L. Wang: Ab-initio study of interacting fermions at finite temperature with neural canonical transformation. arXiv preprint arXiv:2105.08644 (2021)

work page arXiv 2021
[10]

and Brubaker, Marcus A

I. Kobyzev, S. J. Prince, and M. A. Brubaker: Normalizing flows: An introduction and review of current methods. IEEE Trans. Pattern Anal. Mach. Intell.43, 3964–3979 (2020).https: //doi.org/10.1109/TPAMI.2020.2992934

work page doi:10.1109/tpami.2020.2992934 2020
[11]

J. P. Boyd: Chebyshev and Fourier Spectral Methods. Springer (2000). 15

2000
[12]

Canuto, M

C. Canuto, M. Y. Hussaini, A. Quarteroni, and T. A. Zang: Spectral methods: fundamentals in single domains. Springer Science & Business Media (2007)

2007
[13]

Guo: Spectral methods and their applications

B. Guo: Spectral methods and their applications. World Scientific (1998)

1998
[14]

J. Shen, T. Tang, and L.-L. Wang: Spectral methods: algorithms, analysis and applications. Springer Science & Business Media (2011)

2011
[15]

Shen and L.-L

J. Shen and L.-L. Wang: Some recent advances on spectral methods for unbounded domains. Commun. Comput. Phys.5, 195–241 (2009)

2009
[16]

Tang: The Hermite spectral method for Gaussian-type functions

T. Tang: The Hermite spectral method for Gaussian-type functions. SIAM J. Sci. Comput.14, 594–606 (1993).https://doi.org/10.1137/0914038

work page doi:10.1137/0914038 1993
[17]

J. P. Boyd: The rate of convergence of Hermite function series. Math. Comput.35, 1309–1316 (1980)

1980
[18]

J. P. Boyd: Asymptotic coefficients of Hermite function series. J. Comput. Phys.54, 382–410 (1984)

1984
[19]

Hu and H

H. Hu and H. Yu: Scaling optimized Hermite approximation methods. SIAM J. Numer. Anal. 64, 125–147 (2026).https://doi.org/10.1137/25M1737146

work page doi:10.1137/25m1737146 2026
[20]

Jaming, A

P. Jaming, A. Karoui, R. Kerman, and S. Spektor: Approximation of almost time and band limited functions I: Hermite expansions. arXiv preprint arXiv:1407.1293 (2014)

work page arXiv 2014
[21]

M. Xia, S. Shao, and T. Chou: Efficient Scaling and Moving Techniques for Spectral Methods in Unbounded Domains. SIAM J. Sci. Comput.43, A3244–A3268 (2021).https://doi.org/10. 1137/20M1347711

2021
[22]

Takagi and K

H. Takagi and K. Yokouchi: Multiplication and Composition Operators. In: Function Spaces: Proceedings of the Third Conference on Function Spaces, May 19-23, 1998, Southern Illinois University at Edwardsville, 321 (1999)

1998
[23]

Saleh, A

Y. Saleh, A. Iske, A. Yachmenev, and J. Küpper: Augmenting basis sets by normalizing flows. Proc. Appl. Math. Mech.23, e202200239 (2023).https://doi.org/10.1002/pamm.202200239

work page doi:10.1002/pamm.202200239 2023
[24]

Saleh and A

Y. Saleh and A. Iske: Inducing Riesz Bases inL2 via Composition Operators. Complex Anal. Oper. Theory20, 21 (2025).https://doi.org/10.1007/s11785-025-01874-5

work page doi:10.1007/s11785-025-01874-5 2025
[25]

Behrmann, W

J. Behrmann, W. Grathwohl, R. T. Q. Chen, D. Duvenaud, and J.-H. Jacobsen: Invertible Residual Networks. In: Proceedings of the 36th International Conference on Machine Learning, 573–582. PMLR (2019)

2019
[26]

Hille: Contributions to the theory of Hermitian series

E. Hille: Contributions to the theory of Hermitian series. Duke Math. J.5(1939)

1939
[27]

Hille: Contributions to the theory of Hermitian series

E. Hille: Contributions to the theory of Hermitian series. II. The representation problem. Trans. Amer. Math. Soc.47, 80–94 (1940)

1940
[28]

Wang and L

H. Wang and L. Zhang: Convergence analysis of Hermite approximations for analytic functions. arXiv: eprint:2312.07940(math.NA) (2025)

work page arXiv 2025
[29]

Scaling Optimized Hermite Approximation Methods

H. Hu and H. Yu: Scaling Optimized Hermite Approximation Methods. arXiv: eprint:2412.08044 (cs) (2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024
[30]

Villani: Topics in Optimal Transportation

C. Villani: Topics in Optimal Transportation. American Mathematical Society, Providence, RI (2003)

2003
[31]

Á. F. Corral and Y. Saleh: Enhancing polynomial approximation of continuous functions by composition with homeomorphisms. arXiv: eprint:2512.13740(cs.NA) (2025)

work page arXiv 2025
[32]

R. K. Singh and J. S. Manhas: Composition operators on function spaces. Elsevier (1993). 16 A Appendix A.1 Equivalence principle for Riesz bases Analogously to Theorem 4.1 we establish a principle that allows us to transfer convergence results of Hermite expansions to approximations by Riesz bases of the form (2.2). We begin by establishing basic properti...

work page arXiv 1993