PaperlayerView exact revision on arXiv

Source-linked AI summary

A Double-Adaptivity Solver for Parabolic PDEs

Gregor Gantner, Robin Smeets, Rob Stevenson

arXiv:2609.01748v1math.NA

TL;DR

General space-time finite-element partitions may lack the uniform inf-sup stability needed for quasi-optimal minimal-residual discretizations of parabolic problems. The paper replaces this missing uniform guarantee with a data-dependent a posteriori condition, estimators for both variables, and a double-adaptive loop; experiments show refinement toward singularities and improved rates.

  • Problem

    General locally refined space-time partitions do not permit a time-slab decomposition, so uniform inf-sup stability cannot be expected for them.

  • Method

    The method introduces the residual’s Riesz lift, estimates the secondary and principal variables, and alternates test-space and trial-space enrichment in a double-adaptive loop.

  • Results

    0.5 is obtained for the non-matching initial datum in 1 + 1 dimensions across tested polynomial degrees, compared with about 0.12 under uniform refinement.

  • Takeaways & Limitations

    The experiments show refinement toward expected singularities and systematic gains from the two adaptive mechanisms, with more modest gains in 2 + 1 dimensions.

  • Takeaways & Limitations

    The certified secondary-variable estimator requires an auxiliary refined partition whose cardinality cannot generally be bounded linearly by the original partition.

Abstract

from arXiv · show

We study minimal residual space-time finite element discretizations of linear parabolic initial value problems in canonical space-time variational form. To deal with the arising dual norm, we introduce the Riesz lift of the residual as an additional variable. Quasi-optimality of the primal variable of the mixed system follows from a uniform inf-sup condition. This condition is known to be satisfied for finite element spaces w.r.t. prismatic partitions of the space-time cylinder that allow for a decomposition into time-slabs. We prove that this condition cannot be expected to hold otherwise. To recover stability for general partitions and the data at hand, we derive an a posteriori condition on the error between the exact Riesz lift of the residual and its Galerkin approximation -- being the secondary variable of our system -- under which the primal variable is quasi-optimal. We derive a posteriori error estimators for both variables, and use them in a double-adaptive loop that alternates test-space with trial-space enrichment. We illustrate our findings with numerical experiments in $1+1$ and $2+1$ dimensions.

1. INTRODUCTION

The paper develops adaptive monolithic space-time discretizations for linear parabolic problems, using minimal residuals and a Riesz-lift saddle-point formulation. Because uniform inf-sup stability fails on general space-time partitions, the authors introduce data-dependent certification, estimators, and a double-adaptive algorithm validated numerically.

  • Motivation: Monolithic space-time methods support parallel implementations, quasi-best trial-space approximations, and simultaneous local refinement in time and space.Their higher memory cost is less problematic for coupled forward-adjoint or goal-oriented problems.
  • Minimal residual formulation: The discretization minimizes residuals over finite-dimensional trial spaces using a discretized dual norm and introduces the residual’s Riesz lift as an auxiliary variable.The resulting formulation is an equivalent saddle-point system whose primal approximation is quasi-best under uniform inf-sup stability.
  • Stability: Uniform inf-sup stability is known for tensor-product and time-slab spaces but cannot be expected for finite-element partitions that lack a time-slab decomposition.This limitation conflicts with general local space-time refinement, which motivates a data-dependent condition for the data at hand.
  • Adaptivity: The proposed double-adaptive loop enriches the test space until the quasi-optimality condition is met, then enriches the trial space using a posteriori error indicators.The method estimates both the principal-variable error and the Riesz-lift approximation error.
  • Estimator comparison: The one-degree-enriched surrogate is not guaranteed to upper-bound the certified estimator but tracks it within the same order of magnitude in experiments.The certified estimator is efficient and reliable modulo higher-order oscillation terms, though its refined auxiliary partition can prevent linear-complexity computation.
  • Numerical findings: 0.5 is the adaptive rate for the non-matching initial datum in 1 + 1 dimensions across tested polynomial degrees, versus about 0.12 under uniform refinement.For reduced boundary regularity, adaptive rates reach approximately 0.88 for p = 2 and 0.93 for p = 3; gains in 2 + 1 dimensions remain systematic but more modest.

2. PARABOLIC INITIAL VALUE PROBLEM

The paper formulates a linear parabolic initial value problem on a Gelfand triple in space-time variational form. Boundedness and uniform coercivity of the spatial operator provide the well-posedness framework for the evolution equation.

  • Functional setting: The formulation uses the Gelfand triple V ,→ H ,→ V′ and the space-time setting associated with Y = L2(Ξ; V).The trace operator γt from the solution space X into H is bounded uniformly in time.
  • Problem statement: The initial value problem seeks u on Ξ = (0, T) satisfying dt u(t) + A(t)u(t) = ℓ(t) almost everywhere and u(0) = u0.The data consist of a forcing term ℓ and initial value u0.
  • Variational form: The evolution equation is expressed through a space-time operator B and the initial trace condition γ0u = u0.This yields the operator form of the initial value problem for prescribed data (ℓ, u0).
  • Assumptions: The operator A(t) is assumed bounded uniformly in time and uniformly coercive from V to V′.A uniform Gårding inequality would also suffice for the stated well-posedness result.

3. MINIMAL RESIDUAL DISCRETIZATION

The minimal residual method discretizes the space-time operator over trial and test subspaces and rewrites the residual minimization as a saddle-point system with a Riesz-lift variable. Its quasi-optimality depends on an inf-sup constant that may require data-dependent certification.

  • Discrete spaces: Finite-dimensional trial spaces Xδ and test spaces Yδ are selected as subspaces of the continuous spaces X and Y.The discretized dual norm replaces the full Y′ norm, producing mesh-dependent norms on the discrete trial space.
  • Saddle-point system: The minimal residual approximation is characterized by an Euler–Lagrange saddle-point system whose solvability is equivalent to a positive discrete inf-sup constant.This constant also controls uniqueness of the discrete solution.
  • Stability characterization: The inf-sup constant measures how adequately the test space represents the optimal test space, with alternative characterizations through approximability or Fortin projectors.Because of the parabolic operator’s structure, these characterizations do not provide easy access to the constant.
  • Quasi-optimality: For suitable tensor-product and time-slab settings, the inf-sup constant can be bounded uniformly away from zero, yielding quasi-optimality.The associated theory also supports a posteriori error estimation under suitable stability conditions.
  • A posteriori control: The standard estimator is efficient, but reliability is not automatic because of a data-oscillation term and the lack of a uniformly valid saturation assumption.A data-dependent a posteriori condition is therefore used to restore efficient and reliable error control.

4. INF-SUP STABILITY CONDITION (3.6)

Uniform inf-sup stability is established for tensor-product and time-slab finite-element settings, including cases with different spatial meshes on different time intervals. These settings do not cover arbitrary local space-time refinements.

  • Projector condition: The required spatial projector is uniformly bounded for families of conforming newest-vertex-bisection meshes in dimensions d < 7.This result applies to Lagrange finite-element spaces of arbitrary degree generated from an initial mesh.
  • Tensor-product setting: Tensor-product space-time partitions combine a temporal partition of Ξ with a spatial partition of Ω and realize the stability condition.The construction relies on compatible finite-element spaces and suitable spatial projector properties.
  • Time-slab setting: Time-slab spaces generalize tensor products by allowing different spatial finite-element spaces on different time intervals.This retains a decomposition into time slabs while permitting temporal variation in the spatial discretization.
  • Extensions: Sparse tensor-product constructions extend the setting beyond standard finite elements by allowing localized refinements in time and space.These refinements are described as enrichments rather than conventional finite-element mesh refinements.

5. A DATA-DEPENDENT CONDITION FOR QUASI-OPTIMALITY

The paper replaces unavailable uniform inf-sup stability on general space-time partitions with a data-dependent a posteriori condition controlling the Riesz-lift approximation. Under this condition, the primal approximation is quasi-optimal and its estimator is efficient and reliable.

  • Motivation: Uniform inf-sup stability cannot generally be expected for locally space-time-refined partitions that do not decompose into time-slabs.The time-slab condition is tied to uniform approximability of the optimal test space.
  • Data-dependent condition: The proposed condition controls the Y-norm distance between the exact optimal test function A_s^-1(B_a u_δ) and its Galerkin approximation in Y_δ.Unlike uniform inf-sup stability, it is required only for the data at hand.
  • Relation to prior work: The condition is motivated by prior data-dependent stability ideas, while this theorem permits any constant ϱ ≥ 0.The cited predecessor restricted its corresponding constant to [0, 2).
  • Consequences: The condition yields quasi-optimality of u_δ in the X-norm and permits an efficient and reliable a posteriori estimator for the primal error.The resulting estimator includes the residual-lift contribution and a data-oscillation term.
  • Test-space enrichment: For any fixed finite-dimensional trial space X_δ, enlarging Y_δ sufficiently can enforce the data-dependent condition.The required enlargement depends on the supplied data and a chosen tolerance parameter.
  • Verification: Verification requires an a posteriori estimator for the difference between the exact and Galerkin residual lifts in the Y-norm.Constructing that estimator is the subject of the following section.

6. A POSTERIORI ERROR ESTIMATOR FOR ∥λδ −λδ∥Y

The estimator construction reformulates the residual-lift problem on a coarsest time-slab-compatible refinement and uses equilibrated fluxes for the resulting spatially elliptic problems. This introduces computational overhead because the auxiliary refinement may be much larger and is not uniformly shape-regular.

  • Finite element spaces: The finite element construction uses prismatic tensor-product spaces, newest-vertex bisection in higher dimensions, and grading constraints to preserve spatial conformity.An additional coefficient assumption ensures data-oscillation terms depend only on ℓ rather than u_δ.
  • Estimator setting: The estimator targets the Y-norm error between the exact residual lift and its Galerkin approximation for parabolic problems with piecewise diffusion coefficients.On each time-slab, the spatial operator is elliptic, enabling independent slabwise treatment.
  • Auxiliary refinement: The auxiliary mesh qP_δ is the coarsest prismatic refinement of P_δ that permits decomposition into time-slabs.It is formed using temporal cuts and removes prisms with hanging lateral-edge midpoints.
  • Mesh properties: The auxiliary refinement is not uniformly shape-regular, even though its spatial simplices are uniformly shape-regular.This distinction is central to the construction of the estimator on qP_δ.
  • Computational cost: Computing the auxiliary lift requires independent elliptic solves on each time slab, but the cardinality ratio #qP_δ/#P_δ has no uniform bound.Optimal preconditioning can reduce the work by exploiting slabwise independence, yet the auxiliary mesh may still be substantially larger.
  • Flux estimator: Equilibrated flux estimators provide an upper bound that is efficient and reliable up to higher-order data-oscillation terms.The leading term can avoid a generic unknown constant, and smooth-data oscillation is higher order.

7. THE DOUBLE-ADAPTIVE LOOP

The double-adaptive algorithm alternates test-space and trial-space enrichment, using estimators and an a posteriori quasi-optimality condition to guide refinement. A higher-order surrogate can avoid expensive certified estimator components but is heuristic.

  • Estimator and stability: The estimator ϑ(δ) bounds the Riesz-lift error up to a higher-order data-oscillation term, which was negligible in the experiments.The observed oscillation term was 8–12 orders of magnitude smaller than ϑ(δ), depending on the example.
  • Adaptive loop: When the test-space condition fails, marked elements are refined until the condition is addressed, after which the algorithm returns to the mixed solve.The marked set is selected by a bulk criterion involving the local ϑP(δ) contributions.
  • Estimator and stability: The resulting condition ensures that the primal solution is quasi-best, linking test-space adequacy to quasi-optimality in the trial variable.The construction is motivated by the theorem connecting the estimator-controlled test-space error with validity of (5.2).
  • Adaptive loop: The method solves the mixed system on trial and test spaces, then uses estimator values to decide whether further test-space enrichment is needed.The loop begins by solving for (λδ, uδ) and evaluating the relevant estimators before refinement decisions.
  • Practical limitations: Fixing the number of inner test-space enrichments in advance can simplify the loop, but quasi-optimality is then not guaranteed a priori.The standard strategy instead adapts the number of inner enrichments according to condition (7.1).
  • Practical limitations: The certified estimator requires an auxiliary time-slab partition whose size cannot generally be bounded by the original partition size.This makes the certified computation potentially more expensive than the original space-time discretization.
  • Practical limitations: The higher-order space Zδ provides a computable surrogate that avoids auxiliary partitions and local flux equilibration, but it supplies no guaranteed upper bound.Its use is therefore empirical, and Zδ may contain substantially more degrees of freedom than Yδ.

8. NUMERICAL RESULTS

The experiments compare uniform and double-adaptive refinement for heat-equation problems with limited regularity in 1+1 and 2+1 dimensions. Adaptivity refines near singularities and improves convergence, while the test-space mechanism is often inexpensive in the reported runs.

  • Experimental setup: The experiments use heat equations on Ξ=(0,1) and Ω=(0,1)^d for d∈{1,2}, comparing uniform refinement with the double-adaptive strategy.The adaptive runs use ϱ=1 and ϖ=0.4, with newest-vertex bisection in d>1 and an additional grading condition in d=1.
  • Test-space diagnostics: The no-inner-enrichment variant closely matched standard adaptive results in the reported experiments and was omitted from the figures.For the 1+1-dimensional reduced-regularity case, however, its average ϱeff values were 1.26, 0.91, and 1.02 for p=1,2,3.
  • Experimental setup: Smooth manufactured-solution tests attained the expected optimal convergence rates under both uniform and adaptive refinement.These tests were used as implementation checks and are not included in the detailed figures.
  • 1+1 dimensions: For the 1+1-dimensional non-matching initial datum, adaptive refinement reached approximately 0.5 versus 0.12 for uniform refinement, independently of p.Meshes refined toward the singular corners at (t,x)=(0,0) and (0,1).
  • 2+1 dimensions: For the 2+1-dimensional non-smooth datum, adaptive rates were 0.33 for p=1 and 0.46 for p=2,3, versus uniform rates of 0.27 and approximately 0.33.Refinement concentrated toward the singularity at {0}×{x=0}.

9. CONCLUSION

The paper establishes that uniform inf-sup stability fails generally without time-slab partitions, then restores data-dependent quasi-optimality through estimators and double adaptivity. Numerical experiments show refinement toward singularities and improved rates over uniform refinement, while nonlinear operators and more efficient inner-loop estimators remain future work.

  • Uniform inf-sup stability cannot generally be expected for non-time-slab partitions, unlike for partitions decomposable into time-slabs.
  • A data-dependent a posteriori test-space condition guarantees quasi-optimality, under which the principal-variable estimator is efficient and reliable with explicit constants.
  • A computable secondary-variable estimator enables a double-adaptive loop that enriches the test space until stability holds and then enriches the trial space using principal-error indicators.
  • Numerical experiments show refinement toward expected singularities and improved convergence rates compared with uniform refinement.
  • Extending the approach to nonlinear spatial operators and avoiding auxiliary time-slab partitions for inner-loop estimation remain open issues.

APPENDIX A. THE NECESSITY OF TIME-SLABS FOR CONDITIONS (3.6) AND (3.3)

The appendix tests whether the inf-sup conditions remain valid on general prismatic finite element partitions. It frames the analysis through the canonical spaces, Riesz lifts, and optimal test functions used to construct counterexamples.

  • The appendix uses the canonical domain and codomain together with the Riesz map to analyze residual lifts associated with finite element trial functions.
  • For joint shape-regular isotropic prismatic partitions that cannot be subdivided into time-slabs, conditions (3.6) and (3.3) cannot generally be expected to hold.The result also extends when the test-space partition is a fixed-depth refinement of the trial-space partition.
  • The inverse Riesz lift is identified with the optimal test function for a trial function and the parabolic operator.
  • The counterexample analysis begins with a nonzero trial function whose operator image is nonzero and examines the resulting test-space behavior.

A.1. The case d = 1.

For d = 1, the appendix constructs square partitions with separated coarse and fine scales and uses localized functions to show that both inf-sup conditions can fail without time-slab structure.

  • The one-dimensional construction uses square partitions with edge lengths between h_δ and H_δ, subject to p + 2 ≤ H_δ.The relevant partition contains coarse and fine square regions, as illustrated in Figure 8.
  • A piecewise-constant function g_δ is constructed on p + 2 intervals so that it has zero sum and is orthogonal to every polynomial in P_p(0,(p + 2)h_δ).
  • The function g_δ is extended by zero and integrated into a continuous piecewise-linear function y_δ for the counterexample construction.
  • Local support and orthogonality on the coarse region produce the estimates needed to show that condition (3.6) is generally invalid without the time-slab restriction.
  • A second consequence of the construction shows that condition (3.3) is also generally invalid.
  • The argument relies on coarse and fine squares remaining within spatial distance O(H_δ).

A.2. The case d ≥2.

For spatial dimension d ≥ 2, the appendix transfers the one-dimensional obstruction to shape-regular isotropic prismatic partitions using localized spatial hat functions and scale-separated supports.

  • The higher-dimensional construction uses shape-regular isotropic prisms on Σ = (0,1) × Ω and spatially localized hat functions.
  • The selected supports have diameter comparable to H_δ and are separated by distance at most proportional to H_δ.
  • A Poincaré estimate on a ball containing both supports supplies the key bound for the localized construction.
  • For d = 2, a lower bound for the H^-1(Ω)-norm of one hat function is obtained by placing its support near the origin and using a radial construction.
Loading 2609.01748v1…