Source-linked AI summary
Revisiting Diffusive Representations for Enhanced Numerical Approximation of Fractional Integrals
Renu Chaudhary, Kai Diethelm
TL;DR
Numerical evaluation of fractional integrals is resource-intensive because of their non-local character. This paper develops a diffusive representation and quadrature-based approximations, reporting improved algorithms with memory and computational benefits. The stated truncation-error theorem is formally limited to t ≤ 1, with rescaling offered to overcome that limit.
Problem
The non-local character of fractional integrals makes their numerical evaluation resource-intensive in runtime and memory, particularly for complex systems or evaluations at multiple points.
Source: §1 · p. 1Method
Results
Takeaways & Limitations
The proposed approximation approach harmonizes with common techniques for integer-order operators.
Source: §4 · p. 6Theorem 6 is formally limited to t ≤ 1, but the authors state that simple rescaling overcomes this limitation.
Source: §2.3 · p. 4
Abstract
from arXiv · showhide
This study reexamines diffusive representations for fractional integrals with the goal of pioneering new variants of such representations. These variants aim to offer highly efficient numerical algorithms for the approximate computation of fractional integrals. The approach seamlessly aligns with established techniques used in addressing problems involving integer-order operators, contributing to a unified framework for numerical solutions.
1. INTRODUCTION AND PRELIMINARIES
Fractional integrals are costly to evaluate because their non-local character raises runtime and memory demands, especially for complex systems or repeated evaluations. The paper develops a diffusive representation for Riemann–Liouville integrals of order α ∈ (0, 1), using an auxiliary function and quadrature to support numerical approximation.
1. INTRODUCTION AND PRELIMINARIES
- The method targets the Riemann–Liouville integral of order α ∈ (0, 1) for functions f ∈ C[a, b], commencing from a.Its diffusive representation uses an auxiliary bivariate function, which can be found by solving a first-order differential equation, avoiding the need to account for memory effects in the computation.Source: §1 · p. 1
- The proposed approach improves diffusive representations for numerical approximation of fractional integrals, with the stated aim of better algorithms.The conclusion describes benefits in memory efficiency and computational simplicity.Source: §1 · p. 1
- An admissible transformation is a strictly increasing C1 function from a non-empty open interval Ω to (0, ∞), with limits 0 and +∞ at the respective ends of Ω.These conditions define the transformation used in the representation.Source: §1 · p. 1
- Earlier approaches include quadrature of a kernel integral, exponential-sum approximations (Beylkin and Monzón, 2010; McLean, 2018), and Gauss–Jacobi quadrature (Baffet, 2019).The paper situates its new diffusive representation among these methods for approximating fractional integrals.Source: §1 · p. 2
- A specially selected admissible transformation makes the auxiliary function decay exponentially near the boundaries of its domain, supporting trapezoidal or Gauss–Laguerre quadrature.The auxiliary function is studied by numerically solving an initial-value problem with backward Euler or the trapezoidal method.Source: §1 · pp. 1–2
2. VARIANTS OF DIFFUSIVE REPRESENTATION ENABLING EFFICIENT NUMERICAL TREATMENT
The paper develops a diffusive-representation variant whose integrand decays exponentially at the domain boundaries, enabling more efficient numerical integration. It then constructs an exponential-sum kernel approximation and estimates the error from truncating that sum.
2. VARIANTS OF DIFFUSIVE REPRESENTATION ENABLING EFFICIENT NUMERICAL TREATMENT
- The variant’s integrand decays exponentially at the domain boundaries, and its smoothness supports more efficient numerical integration.The section presents this decay-and-smoothness combination as the basis for efficient numerical treatment.Source: §2 · p. 2
2.1 The Fundamental Approach
- For fixed t, the integrand decays asymmetrically: at least as fast as e^-αr as r → ∞, and much faster as r → −∞.The stated faster positive-direction rate is possible only for rare and exceptional choices of f.Source: §2.1 · p. 3
2.2 Kernel Function Approximation by Exponential Sums
- The kernel integral over an unbounded domain is discretized by the trapezoidal rule to produce an exponential sum; a finite approximation is considered for δ ≤ t ≤ b − a.The authors note that Prony’s method can reduce excess terms without compromising accuracy, as discussed by Beylkin and Monzón (2010).Source: §2.2 · pp. 3–4
2.3 Truncation Error Estimation of the Kernel Function
- The truncation analysis gives Λ = M + 1 + N terms for relative accuracy ϵ on δ ≤ t ≤ 1, with order (log ϵ^-1) log(δ^-1 log ϵ^-1).The stated theorem formally covers t ≤ 1; the authors note that rescaling can overcome this restriction.Source: §2.3 · p. 4
3. EFFICIENT NUMERICAL ALGORITHMS FOR APPROXIMATE EVALUATION OF FRACTIONAL INTEGRAL
The section presents two algorithms for evaluating the fractional integral: exponential-sum approximation of the kernel and Gauss-Laguerre quadrature. It describes their computational procedures, including a backward Euler choice for the differential equation.
3.1 Numerical Algorithm for Approximating Fractional Integral using Exponential Sum Approximation of the Kernel Function
- The exponential-sum algorithm approximates the kernel using a moderate number of terms Λ, with δ chosen below every time step, and evaluates the integral using a recursion formula.The approach discretizes the kernel integral over an unbounded domain by the trapezoidal rule, yielding a sum of exponential terms.Source: §3.1 · pp. 4–5
- O(Λ·P) operations and storage proportional to Λ replace conventional O(P^2) computation and storage proportional to P in the exponential-sum approach.The algorithm can overwrite stored values; the stated savings are substantial particularly when Λ ≪ P.Source: §3.1 · pp. 4–5
3.2 Numerical Algorithm for Approximating Fractional Integral using Gauss-Laguerre Quadrature Formula
- The Gauss-Laguerre algorithm applies quadrature to the integral, using nodes and weights associated with the Laguerre polynomial.The section notes that this method is natural for the [0, ∞) component; the other component decays faster and may use a truncated sum or another Gauss-type quadrature.Source: §3.2 · p. 5
- Under Theorem 4’s assumptions, the Gauss-Laguerre method has a convergence result for all t ∈ [a, b].The proof invokes multiple differentiability of φ(t, ·) and its decay properties.Source: §3.2 · p. 5
3.3 Methodology and Computational Procedures
- The computational procedure uses backward Euler for the differential equation because it is the simplest A-stable method; the trapezoidal method is an alternative requiring formula modifications.The procedure takes the initial point, order, grid points, and number of quadrature nodes as inputs, then computes Gauss-Laguerre nodes and weights.Source: §3.3 · pp. 5–6
4. CONCLUSION AND FUTURE DIRECTIONS
The paper presents a diffusive representation using an admissible transformation chosen to give the integrand exponential decay, enabling trapezoidal or Gauss-Laguerre quadrature. It identifies broader applicability, algorithm optimization, and rigorous error analysis as future directions.
4. CONCLUSION AND FUTURE DIRECTIONS
- The proposed diffusive representation uses an admissible transformation that gives the integrand exponential decay, supporting trapezoidal discretization on unbounded domains or Gauss-Laguerre quadrature.The approach is described as harmonizing with techniques for integer-order operators, with benefits in memory efficiency and computational simplicity.Source: §4 · p. 6
- Future work includes broadening the admissible transformations and domains, optimizing algorithms for intricate functions, and evaluating practical applications in environmental modeling, finance, and control systems.The paper also calls for rigorous error analysis to validate accuracy and convergence across applications.Source: §4 · p. 6