A coercive space-time variational approach to fractional diffusion problems

Herbert Egger; Marvin Fritz; Barbara Wohlmuth

Source-linked AI summary

A coercive space-time variational approach to fractional diffusion problems

Herbert Egger, Marvin Fritz, Barbara Wohlmuth

arXiv:2609.00091v1Aug 31, 2026math.NA

TL;DR

The paper studies a time-fractional diffusion equation and develops a coercive space–time variational framework in an appropriate fractional Sobolev setting. It establishes well-posedness and regularity, derives quasi-optimal Galerkin error estimates, and preserves causal time-stepping with memory while retaining efficient computational structure.

Problem
The paper considers a time-fractional diffusion equation as its model problem.
Method
The paper extends a space–time variational approach using a natural coercivity structure, then develops a conforming tensor-product Galerkin discretization and fractional-integral analysis.
Results
The formulation establishes continuity, coercivity, well-posedness, regularity, quasi-optimal energy-norm error estimates, and improved L2-estimates through duality.
Takeaways & Limitations
The tensor-product framework combines the analytical advantages of space–time methods with causal time-stepping and efficient realization for evolution problems with memory.
Takeaways & Limitations
Inhomogeneous initial conditions introduce singularities at t = 0 that require careful treatment, and extensions to incompatible initial data may produce stronger initial-time singularities.

Abstract

from arXiv · show

We consider a fractional diffusion problem with temporal nonlocality acting on the diffusive flux. A coercive space--time variational formulation in Bochner-valued fractional Sobolev spaces is derived and the existence, uniqueness, and regularity of solutions are established. We further develop a conforming tensor-product Galerkin discretization and prove quasi-optimal error estimates in the anisotropic energy norm and improved convergence rates in weaker norms using duality arguments. In contrast to some space-time formulations for classical diffusion, the method preserves the causal structure of the evolution problem and leads to a time-stepping procedure with memory terms. On uniform time grids, the discrete history operator has a lower-triangular Toeplitz structure which enables an efficient implementation using fast recursive convolution techniques.

1. Introduction.

The paper formulates a time-fractional diffusion problem whose temporal nonlocality creates analytical and computational challenges, then develops a coercive space–time framework and causal discretization.

Motivation: Fractional diffusion models anomalous transport in heterogeneous and multiscale media, while temporal nonlocality couples solution values across different times.These features motivate both the analytical treatment and the need to handle memory computationally.
Contributions: The work extends a space–time variational approach to the time-fractional diffusion equation in its natural functional-analytic setting.The formulation incorporates initial conditions in strong form but requires additional fractional-calculus arguments.
Contributions: The authors establish continuity, coercivity, well-posedness, and regularity for the resulting variational formulation.These properties provide the analytical foundation for the subsequent discretization and error analysis.
Discretization: A conforming tensor-product Galerkin method yields quasi-optimal energy-norm estimates and improved L2-estimates obtained through duality.The discretization and a priori estimates are developed in the later sections of the paper.
Implementation: The discretization preserves causality and can be realized as time stepping for an evolution problem with memory.On uniform grids, the history operator produces lower-triangular Toeplitz systems suitable for fast convolution techniques.

2. Notation and fractional calculus.

This section establishes the fractional Sobolev and Bochner-space notation and recalls fractional-integral identities, adjointness, and coercivity tools used later.

Fractional Sobolev spaces: Fractional Sobolev spaces are defined through Fourier-based norms and restriction to the interval (0, T), with equivalent Sobolev-Slobodeckij descriptions.The recalled embeddings also provide continuity and endpoint traces in the applicable regularity range.
Fractional integrals: Left- and right-sided Riemann–Liouville fractional integrals satisfy a semigroup property and are adjoint under the L2(0, T) inner product.These relations support the distributional extensions and subsequent variational identities.
Coercivity: For 0 < β < 1/2, the recalled fractional-integral estimate establishes coercivity, with a constant involving cos(πβ).The corresponding coercivity constant degenerates as β approaches 1/2.
Fractional derivatives: For 1/2 < β < 1, fractional derivatives and endpoint conditions characterize relevant fractional Sobolev spaces in the distributional setting.The natural norms induced by the alternative definitions are equivalent in the stated spaces.
Bochner spaces: Bochner spaces Hs(0, T; X) describe fractional time regularity for functions taking values in a spatial space X.The notation also identifies product-domain Lebesgue spaces with iterated Bochner spaces through Fubini’s theorem.

3. Space-Time Variational Formulation.

The paper formulates fractional diffusion in a natural coercive space–time setting, proves well-posedness and regularity, and derives a conforming variational framework with established solution properties.

3. Space-Time Variational Formulation.: The model is a fractional diffusion equation on a bounded Lipschitz domain with 0 < α < 1 and homogeneous boundary and initial conditions.
3.1. Weak formulation.: The weak formulation is derived by testing the differential equation with a time derivative, integrating spatially by parts, and rewriting the fractional term through fractional-integral identities.
3.1. Weak formulation.: The trial space V is identified as the natural energy space, and the variational problem seeks a weak solution satisfying A(u, v) = ℓ(v).
3.2. Well-posedness.: The space V has an equivalent Hilbert-space norm, while the bilinear form A is well-defined, continuous, and coercive with constants depending only on α and T.
3.2. Well-posedness.: The coercivity constant degenerates as α → 1, coinciding with loss of ellipticity of the space–time bilinear form in the standard diffusion limit.
3.2. Well-posedness.: For every f ∈ L2(Q), the initial-boundary value problem admits a unique weak solution by applying the Lax–Milgram lemma.
3.3. Regularity.: For f ∈ L2(Q), regularity arguments and elliptic regularity yield additional spatial-temporal regularity, with solution norms bounded by the data norms.
3.3. Regularity.: Higher temporal regularity follows by differentiating the fractional diffusion equation, with convex domains and suitable temporal data assumptions supporting the resulting estimates.

4. Discretization and error analysis.

The paper uses a conforming tensor-product Galerkin discretization with continuous piecewise linear functions in space and time, establishing quasi-optimal energy-norm error estimates and improved L2(Q) rates under regularity assumptions.

Discretization: The approximation space uses continuous piecewise linear functions in both spatial and temporal variables on a tensor-product discretization.Uniform temporal meshes are analyzed, while the interpolation framework also covers suitable locally refined meshes.
Discretization: The discrete problem has a unique solution because the finite-dimensional bilinear form inherits ellipticity and continuity.Existence and uniqueness follow by applying the Lax–Milgram lemma to the discrete space.
Extensions: The interpolation and error estimates extend to a broad class of adaptively refined tensor-product meshes and can also be generalized to higher-order polynomial approximations.The paper reports corresponding adaptive numerical results and notes that the extension to higher-order approximations is straightforward.
Energy-norm error estimates: For 0 ≤ r,s ≤ 1, the energy-norm error satisfies ∥u − uhτ∥V = O(τ^r + h^s), reaching O(τ + h) under stronger regularity.The optimal-order bound requires u ∈ H2(0,T;L2(Ω)) ∩ H2−α(0,T;H2(Ω)).
Energy-norm error estimates: The O(h + τ) energy-norm rate is optimal under the stated regularity assumptions and transfers to equivalent standard Bochner-Sobolev norms.The assumptions are supported, for example, when f ∈ H1(0,T;L2(Ω)), f(0)=0, and Ω is convex.
Improved L2-estimates: Under the same stronger regularity and convexity of Ω, the L2(Q)-error improves to O(τ^2 + h^2) through a duality argument.The dual problem supplies the regularity needed for the improved estimate.

5. Efficient implementation.

The tensor-product structure is reformulated in temporal derivative variables, exposing a lower-triangular causal system that supports time stepping and fast history updates.

Algebraic formulation: Because temporal derivatives are the only time-dependent operators, the discretization can use discontinuous piecewise constant basis functions for p = ∂tuhτ and q = ∂tvhτ.The resulting representation uses spatial mass and stiffness matrices together with temporal matrices.
Causal structure: Causality makes the temporal matrix Cτ lower triangular, while its entries can be computed analytically on general time grids.The diagonal matrix Dτ and lower-triangular Cτ produce the history structure of the discrete system.
Time stepping: The discrete problem reduces to a lower-triangular block system and can be realized as time stepping with one elliptic finite element solve per time step.The spatial system matrix remains the same across time steps and can therefore be factorized once.
Fast history updates: On uniform time grids, the system matrix has a lower-triangular block-Toeplitz structure with sparse blocks.This structure comes from the temporal history operator and is the basis for fast convolution updates.
Fast history updates: Recursive fast Toeplitz multiplications reduce the residual-computation cost, while approximate circular embedding also permits parallel-in-time solution.The reported complexity expressions depend on the spatial and temporal degrees of freedom and are given in the implementation analysis.

6. Numerical results.

Numerical experiments confirm the predicted convergence rates for smooth and temporally nonsmooth solutions, while graded time meshes recover optimal orders and Toeplitz-based history updates improve computational performance.

6.1. Convergence for smooth solutions.: For the smooth case β = 2 with h ≃τ, the energy error converges as O(h + τ), while the space–time L2-error converges as O(h2 + τ 2).The manufactured solution satisfies the regularity assumptions used for the theoretical estimates.
6.1. Convergence for smooth solutions.: The fractional-gradient component dominates the energy error, although both energy-error components converge with first order under h ≃τ.The space–time L2-error and final-time error both converge with second order in the smooth experiment.
6.2. Reduced temporal regularity.: For non-integer 1/2 < β < 3/2, the predicted rates are O(τ β−1/2 + h) in the energy norm and O(τ β+1/2 + h2) in L2(Q).Uniform meshes with h ≃τ show reduced asymptotic rates, with a pre-asymptotic phase caused by contributions of different orders.
6.3. Graded temporal meshes.: Grading the temporal mesh near t = 0 regains full convergence orders, up to logarithmic powers in τ, without increasing the number of time steps.The chosen grading parameters are γ = 4 for β = 0.75 and γ = 4/3 for β = 1.25.
6.4. Efficient solvers.: Exploiting the lower-triangular Toeplitz structure and fast convolution substantially improves computational performance compared with basic history updates.The comparison examines total solver times and history-update times for methods (M1) and (M2).

7. Discussion.

The work establishes a coercive space-time framework for fractional diffusion and shows that its tensor-product discretization can be realized efficiently as time stepping. The authors identify extensions addressing broader regularity, adaptivity, and fractional evolution settings.

The formulation yields well-posedness of the continuous problem and quasi-optimality of the corresponding Galerkin approximations.
A systematic tensor-product finite element discretization achieves optimal convergence rates under regularity assumptions reflecting reduced temporal regularity.
The tensor-product structure permits efficient realization as a time-stepping procedure while retaining the analytical advantages of the space-time framework.
Extensions include incompatible initial data, higher-order and adaptive discretizations, graded temporal meshes, nonuniform-grid implementations, and a posteriori error estimates.

Declaration of AI use.

ChatGPT was used during manuscript preparation for proof-checking, language revision, and typesetting assistance. The authors independently verified the mathematical content and accept full responsibility.

ChatGPT assisted with proof-checking during manuscript preparation.
ChatGPT assisted with language revision and typesetting.
The authors independently verified the mathematical content and take full responsibility for the manuscript.