Abstract
The Double-Exponential (DE) transformation, developed primarily in Japan over the last 50 years, represents a significant advancement in numerical integration and function approximation. By utilizing a specific change of variables that maps finite intervals to infinite ones with double-exponential decay, this method achieves near-optimal convergence rates and exceptional robustness against endpoint singularities. This paper reviews the core principles of the DE transformation, its extension to Sinc methods, and its comparative performance against traditional numerical techniques.
Introduction
Numerical integration (quadrature) and function approximation are fundamental tools in science and engineering. However, standard methods often struggle when dealing with singularities—points where a function or its derivative becomes infinite. Traditional rules, such as Gaussian quadrature, often require specific adjustments for different types of singularities. The Double-Exponential (DE) transformation offers a "universal" and highly efficient framework that handles a wide variety of analytic functions, even those with endpoint singularities, without requiring prior knowledge of the singularity's nature.
The Double-Exponential Transformation
The core of the DE tradition is a change of variable $x = \phi(t)$, specifically the tanh-sinh transformation: $\phi(t) = \tanh(\frac{\pi}{2} \sinh t)$. This mapping transforms a finite interval $[-1, 1]$ into an infinite one $(-\infty, \infty)$. The derivative $\phi'(t)$ decays at a "double exponential" rate (e.g., proportional to $\exp(-\exp|t|)$).
When the simple trapezoidal rule is applied in this transformed space, the errors vanish remarkably fast. This phenomenon is often referred to as the "magic" of the trapezoidal rule, which becomes the most efficient method possible for integrating certain analytic functions on an infinite interval. Research by Masaaki Sugihara has proven that the DE formula is optimal within specific function classes, such as Hardy spaces, meaning no other quadrature formula can converge significantly faster for these functions.
DE-Sinc Methods and Applications
Beyond integration, the DE principle has been adapted for function approximation and solving equations through DE-Sinc methods. While traditional Sinc methods use "single exponential" transformations, DE-Sinc methods apply the double-exponential mapping to approximate functions or solve differential and integral equations.
As shown in comparative studies, the DE-Sinc method demonstrates superior efficiency. For instance, it can reach an error of $10^{-14}$ (near the limits of double-precision arithmetic) with approximately 70 sample points. In contrast, Ordinary-Sinc and Chebyshev interpolation require significantly more points to achieve similar accuracy or may fail to reach that level entirely for certain test functions.
Key Results and Robustness
The primary advantage of the DE method is its convergence rate. While the "tanh-rule" (Single Exponential) converges at a rate of $\exp(-C\sqrt{N})$, the DE formula converges at the much faster rate of $\exp(-CN/\log N)$. Furthermore, DE formulas are remarkably robust; they are essentially "blind" to endpoint singularities. Because the transformation pushes these points toward infinity where the integration weights become negligible, the same code can handle integrands that go to infinity at boundaries without modification.
Limitations and Conclusion
Despite its power, the DE transformation is not a universal "silver bullet." Standard tanh-sinh quadrature is less effective for highly oscillatory integrals, although specialized Fourier-type transformations (developed by Ooura and Mori) have been created to address this. Additionally, the rapid convergence often necessitates arbitrary-precision arithmetic to fully realize the method's benefits, as it quickly reaches the limits of standard 64-bit floating-point representations. Nevertheless, the DE transformation remains a cornerstone of high-precision numerical analysis, offering a robust and efficient solution for a wide range of practical engineering and mathematical problems.
Summary updated