Summation of Gaussian shifts as Jacobi's third Theta function

Shengxin Zhu

Abstract

A proper choice of parameters of the Jacobi modular identity (Jacobi Imaginary transformation) implies that the summation of Gaussian shifts on infinity periodic grids can be represented as the Jacobi's third Theta function. As such, connection between summation of Gaussian shifts and the solution to a Schrödinger equation is explicitly shown. A concise and controllable upper bound of the saturation error for approximating constant functions with summation of Gaussian shifts can be immediately obtained in terms of the underlying shape parameter of the Gaussian. This shed light on how to choose a shape parameter and provides further understanding on using Gaussians with increasingly flatness.

Introduction: The Challenge of Saturation Error

In numerical analysis and machine learning, Gaussian Radial Basis Functions (RBFs) are essential tools for interpolation and solving partial differential equations. However, they are subject to a "saturation error"—a fundamental limit on approximation accuracy. Unlike many numerical methods where increasing grid density leads to arbitrary precision, Gaussian interpolation often hits a floor where the error ceases to decrease.

Historically, it was known that making Gaussians "flatter" by adjusting their shape parameter could mitigate this error, but quantifying this effect required complex harmonic analysis. This paper introduces an elegant mathematical bridge to simplify this problem, providing a direct way to calculate and control this error.

The Jacobi Theta Identity

The core contribution of this work is the realization that the sum of shifted Gaussians on an infinite periodic grid is exactly equivalent to Jacobi’s third Theta function ($\vartheta_3$). By leveraging the Jacobi modular identity (or Jacobi imaginary transformation), the author proves that: $$\frac{1}{\sqrt{\pi d}} \sum_{n=-\infty}^{\infty} e^{-(x-n)^2/d} = \vartheta_3(\pi x, i\pi d)$$ This identity, established in Theorem 2.1, allows researchers to bypass traditional integral transforms and use the well-established properties of elliptic functions to analyze RBFs.

Visualizing the shape parameter $d$: as $d$ increases, the Gaussian $e^{-x^2/d}$ becomes increasingly flat, which is key to reducing approximation error.

Furthermore, because Theta functions satisfy specific differential equations, this identity explicitly links the summation of Gaussian shifts to solutions of the Schrödinger equation, connecting approximation theory to fundamental physics. This approach avoids the need for sophisticated harmonic analysis, making the underlying theory more accessible to practitioners.

Quantifying Accuracy and Flatness

The paper provides a rigorous mathematical basis for the heuristic that "flatter is better." Theorem 2.2 derives a remarkably simple and controllable upper bound for the saturation error in terms of the shape parameter $d$: $$\left| \text{Error} \right| < \text{csch}(\pi^2 d)$$ This hyperbolic cosecant bound decays extremely rapidly as $d$ increases. Numerical verification confirms that as $d$ scales from 1 to 4, the saturation error plummets from $10^{-4}$ to $10^{-16}$, effectively reaching machine precision.

Numerical verification of saturation error across different values of $d$. The error magnitude decreases from $10^{-4}$ at $d=1$ to machine epsilon ($10^{-16}$) at $d=4$.

Conclusion

By identifying Gaussian sums as Jacobi Theta functions, this research provides a practical guide for practitioners: to achieve a specific level of accuracy (such as machine epsilon), one can now calculate the exact required flatness of the Gaussian kernel. While currently limited to one-dimensional periodic grids, this framework opens new avenues for studying "learnability" in machine learning and extending these concise bounds to higher-dimensional spaces.

Summary updated

Abstract

A proper choice of parameters of the Jacobi modular identity (Jacobi Imaginary transformation) implies that the summation of Gaussian shifts on infinity periodic grids can be represented as the Jacobi's third Theta function. As such, connection between summation of Gaussian shifts and the solution to a Schrödinger equation is explicitly shown. A concise and controllable upper bound of the saturation error for approximating constant functions with summation of Gaussian shifts can be immediately obtained in terms of the underlying shape parameter of the Gaussian. This shed light on how to choose a shape parameter and provides further understanding on using Gaussians with increasingly flatness.
BibTeX
@inproceedings{zhu2020summation,
  author = {Shengxin Zhu},
  title = {{Summation of Gaussian shifts as Jacobi's third Theta function}},
  booktitle = {Math. Found. Comput},
  year = {2020},
  url = {https://openreview.net/forum?id=TqKkutWIN8r},
}
Figure