Source-linked AI summary
A discrete Grönwall inequality with application to numerical schemes for subdiffusion problems
Hong-lin Liao, William McLean, Jiwei Zhang
TL;DR
Numerical approximations for Caputo derivatives must accommodate solutions with limited regularity and singular behavior at t=0. This paper generalizes a discrete fractional Grönwall inequality to varied discretizations and nonuniform time grids, supporting stability, convergence, and later higher-order and fast-scheme analyses.
Problem
Solutions of fractional reaction-subdiffusion problems are typically less regular than solutions of classical parabolic PDEs, with singular behavior at t=0 limiting numerical convergence.
Method
The paper generalizes the discrete Grönwall inequality beyond the L1 scheme using general discrete kernels and assumptions involving kernel monotonicity, lower bounds, and local step-size ratios.
Results
Alikhanov’s L2-1σ formula can achieve second-order accuracy on certain nonuniform time grids, while a linearized fast algorithm is unconditionally convergent for nonlinear subdiffusion equations.
Takeaways & Limitations
The generalized inequality provides a basis for stability and convergence estimates across discretizations of the Caputo derivative and facilitates analysis of higher-order and linearized fast schemes.
Takeaways & Limitations
It remains unknown whether variable-weights recombination works on nonuniform grids or which mesh constraints would ensure assumptions A1–A3.
Abstract
from arXiv · showhide
We consider a class of numerical approximations to the Caputo fractional derivative. Our assumptions permit the use of nonuniform time steps, such as is appropriate for accurately resolving the behavior of a solution whose derivatives are singular at~$t=0$. The main result is a type of fractional Grönwall inequality and we illustrate its use by outlining some stability and convergence estimates of schemes for fractional reaction-subdiffusion problems. This approach extends earlier work that used the familiar L1 approximation to the Caputo fractional derivative, and will facilitate the analysis of higher order and linearized fast schemes.
1. Introduction.
Fractional reaction-subdiffusion solutions can be weakly regular near t=0, limiting time-discretization accuracy and motivating graded or adaptive meshes. The paper generalizes the discrete fractional Grönwall inequality beyond L1 schemes to broad nonuniform meshes, supporting analyses of higher-order and fast schemes.
- Motivation: Fractional reaction-subdiffusion solutions are typically less regular than classical parabolic solutions, with derivatives that can be singular near t=0.For an eigenfunction initial condition, the paper gives ∂u/∂t = O(t^(α−1)) as t → 0.
- Motivation: Polynomial-interpolation time approximations are limited by solution smoothness, so graded meshes can restore optimal convergence for fixed initial singularities.Graded meshes concentrate points near t=0 to resolve this behavior; the grading parameter γ controls that concentration.
- Limitations of earlier analysis: Nonuniform meshes complicate stability and consistency analysis, while the basic discrete Grönwall inequality is restricted to L1 kernels and excludes some adaptive meshes.The earlier proof relies substantially on the simple form of the L1 kernels and rough truncation-error estimates.
- Contributions: The paper generalizes the discrete Grönwall inequality to varied Caputo discretizations using assumptions on kernel monotonicity, lower bounds, and local step-size ratios.It defines general discrete kernels and complementary kernels, then proves an estimate involving the Mittag–Leffler function.
- Contributions: The framework permits general nonuniform or adaptive time grids rather than only graded meshes designed for initial singularities.Its intended applications include higher-order time discretizations and linearized fast algorithms with nonuniform step sizes.
- Implications: The generalized results are intended to support second-order accuracy for Alikhanov’s formula on certain nonuniform grids and unconditional convergence of a linearized fast algorithm.These outcomes are stated for companion papers applying the present Grönwall theory.
2. Discrete fractional derivative.
The section defines discrete Caputo derivatives through convolution-like sums on possibly nonuniform time levels, imposes kernel and step-ratio assumptions, and introduces complementary discrete kernels for analysis.
- General discrete derivative: Discrete Caputo derivatives are represented by convolution-like sums using general kernels A^(n)_{n-k} over possibly nonuniform time levels.The time levels satisfy 0 = t_0 < t_1 < ··· < t_N = T, with step sizes τ_n and offset parameter θ.
- Kernel assumptions: The theory requires positive, monotone kernels, a kernel lower bound, and a restriction on adjacent step-size ratios.These are assumptions A1, A2, and A3, respectively.
- Admissible schemes: The framework covers several discrete Caputo formulas, including L1, fast L1, Alikhanov, and multi-term or distributed-order extensions.The listed schemes satisfy the kernel assumptions when the local mesh condition is appropriately enforced.
- Examples: For the nonuniform L1 formula, the discrete kernel assumptions hold with π_A = 1.The formula uses linear interpolation of v′ over each time interval.
- Examples: For the fast L1 formula, sufficiently small kernel-approximation tolerance gives A1–A2 with π_A = 3/2.The fast approximation uses a sum of exponentials for the weakly singular kernel.
- Examples: For the nonuniform Alikhanov formula, A1–A2 hold with π_A = 11/4 when the maximum step-size ratio is ρ = 7/4.The construction combines linear and quadratic interpolation with θ = α/2.
3. Discrete fractional Gr¨onwall inequality.
The paper establishes discrete fractional Grönwall inequalities under kernel monotonicity, lower-bound, and mesh-ratio assumptions, covering broad nonuniform meshes and multiple Caputo discretizations. The results support stability analysis for subdiffusion schemes while identifying limitations for Caputo BDF2-like formulas.
- Theorem 3.1: Theorem 3.1 provides a discrete fractional Grönwall inequality under assumptions A1–A3, including a restriction involving the nonuniform mesh parameter ρ.The proof uses induction and incorporates ρ into the estimate, unlike classical parabolic discrete Grönwall inequalities.
- Alternative estimates: If the relevant constant Λ is non-positive, a simpler inequality holds using only A1–A2 and no time-step restriction.For the general case, the maximum-step restriction is described as not stringent in practical applications.
- Mesh dependence: A Mittag–Leffler factor enters the estimates, and the mesh parameter ρ indicates that sudden, drastic reductions of the time step should be avoided.The inequality nevertheless permits heterogeneous general nonuniform meshes.
- Scope of applicability: The inequalities apply on very general nonuniform time meshes and cover L1, fast L1, Alikhanov, multi-term, and distributed-order Caputo discretizations when the assumptions hold.The fractional exponent used in the estimates is determined by the lower-bound kernel rather than necessarily by a continuous Caputo counterpart.
- Assumptions and kernel framework: The theory starts from a discrete convolution form with general kernels, requiring kernel monotonicity, a lower bound, and a mild local step-size-ratio restriction.Complementary discrete kernels are defined to support the resulting convolution estimates.
- Caputo BDF2-like formulas: The standard assumptions may fail for Caputo BDF2-like formulas, although weighted recombination can make the Grönwall inequalities applicable to a transformed scheme.Whether this recombination works on nonuniform meshes, and which mesh constraints are required, remains open.
4. Stability and consistency.
The section applies the general discrete fractional framework to Galerkin time-stepping schemes for reaction-subdiffusion problems, establishing stability and outlining consistency analysis.
- Stability: Coercivity of the bilinear form induced by the strongly elliptic operator provides the positivity needed in the stability argument.The bilinear form is assumed coercive after increasing κ if necessary.
- Spatial and fully discrete schemes: The spatial discretization uses a finite-dimensional subspace X_h of H^1, with Galerkin approximation u_h and fully discrete solution u_h^n.Finite element spaces based on triangulations are given as an example, and finite differences can be treated similarly.
- Spatial and fully discrete schemes: The fully discrete scheme applies the general discrete fractional derivative approximation to the fractional term at each time level on a nonuniform mesh.The formulation is imposed for all test functions in X_h and for 1 ≤ n ≤ N.
- Stability: Under A1 and 0 ≤ θ ≤ θ(n), Theorem 4.2 yields stability of the fully discrete solution in L2(Ω).The proof combines the auxiliary discrete inequality, positivity of the bilinear form, and the discrete fractional Grönwall inequality.
Appendix A. Two technical inequalities.
The appendix proves two technical inequalities for sequences under the kernel monotonicity assumption A1, preparing the stability analysis used with the discrete fractional Grönwall inequality.
- Lemma A.1: Lemma A.1 begins by analyzing a sequence difference and using an identity that decomposes the relevant terms into increments.The proof introduces auxiliary quantities to organize the sequence differences.
- Lemma A.1: Assumption A1 implies the auxiliary coefficients satisfy Q1 ≥ Q2 ≥ ··· ≥ Qn, which controls the sequence increments in the proof.The relation vj − vj−1 = Qj(wj − wj−1) is used for 2 ≤ j ≤ n.
- Role in the analysis: The resulting inequality supplies the technical estimate needed before applying the discrete fractional Grönwall inequality.The appendix completes the proof by relating v_n−1 to v_n and the final sequence increment.