PaperlayerView exact revision on arXiv

Source-linked AI summary

Fortin operators for DPG advection discretizations

Pablo Cortés Castillo, Jay Gopalakrishnan

arXiv:2608.30043v1math.NA

TL;DR

Practical DPG advection discretizations need finite-dimensional test spaces that preserve stability while avoiding global infinite-dimensional dual-norm solves. The paper constructs minimal local Fortin-compatible polynomial spaces and proves mesh-uniform bounds, yielding quasioptimality and reliable error estimation, with improved lowest-order behavior after augmentation. The analysis assumes piecewise constant advection and requires non-characteristic facets for k ≥2.

  • Problem

    Practical DPG replaces an infinite-dimensional global dual-norm computation with local finite-dimensional test solves, creating the need for a mesh-uniform Fortin operator for advection.

  • Method

    The paper constructs minimal elementwise polynomial test spaces from advection-dependent facet and interior bubbles and analyzes their Fortin operators in β-weighted broken graph norms.

  • Results

    The practical method is quasioptimal in the DPG energy norm and has a globally reliable and efficient computable residual estimator for every 1 < p < ∞.

  • Takeaways & Limitations

    At k = 1, characteristic facets are allowed; augmentation and a suitable scale recover optimal first-order L2 convergence and estimator control up to data oscillation.

  • Takeaways & Limitations

    The analysis assumes piecewise constant β, and for k ≥2 requires a uniform non-characteristic facet condition; extending it to variable fields is deferred.

Abstract

from arXiv · show

We construct Fortin operators for discontinuous Petrov-Galerkin (DPG) discretizations of the advection equation with a piecewise constant divergence-free advection vector $β$, on simplicial meshes of any spatial dimension $N$ and for any polynomial degree $k \ge 1$ of the trial space. A minimal test space is built on each element from facet and interior bubbles and later augmented. The Fortin operator is shown to be bounded, uniformly over shape-regular mesh families, in the natural $β$-weighted broken test graph norm built on $L_q$ for every $1 < q < \infty$, where $q$ is the exponent conjugate to the trial exponent $p$. A non-characteristic facet condition is assumed when $k \ge 2$, while the lowest-order case requires no such condition and admits characteristic facets. As applications we prove that the practical fully discrete residual minimization method is quasioptimal in the DPG energy norm for every $1 < p < \infty$, with a quasioptimality constant governed solely by the Fortin operator, that its computable residual is a globally reliable and efficient a posteriori error estimator, and that augmenting the test space and changing the test norm improves the results in the lowest-order case.

1. Introduction

The paper develops minimal polynomial Fortin operators for practical DPG advection discretizations in arbitrary dimension and polynomial degree. It establishes mesh-uniform bounds, quasioptimality, reliable error estimation, and improved lowest-order behavior under weighted norms and augmentation.

  • Motivation: DPG replaces global infinite-dimensional dual-norm solves with local finite-dimensional test-space residual minimization, whose stability is analyzed here for advection.The Fortin operator controls the replacement of the full test space by a discrete test space.
  • Problem and discretization: The model uses piecewise constant divergence-free advection and discontinuous polynomial trial functions of degree k −1 with degree-k interface variables for arbitrary k ≥1.The discretization uses broken test spaces rather than globally conforming ones.
  • Contribution: The minimal local test space combines signed facet bubbles with interior bubbles, has dimension matching active trial degrees of freedom, and yields a mesh-uniform Fortin operator.The construction is explicit on each simplex and uses polynomial degree k+N.
  • Main results: Fortin boundedness holds in a β-weighted broken graph norm for every 1 < p < ∞, with no directional restriction at k = 1 and a uniform non-characteristic facet condition for k ≥2.The lowest-order case admits characteristic facets, whereas higher orders require the stated geometric condition.
  • Applications: The practical method is quasioptimal in the DPG energy norm, converges at the expected rate, and has a globally reliable and efficient residual estimator.These results extend the Fortin analysis to the fully discrete residual minimization method.
  • Lowest-order improvement: At lowest order, augmenting the test space by at most two functions per element and introducing a scale σ recovers optimal first-order L2 convergence and direct estimator control up to vanishing oscillation for piecewise constant data.The result is obtained for fixed σ above the intrinsic element-wise scale.

2. The DPG method for advection

This section formulates practical DPG residual minimization for advection using broken graph test spaces and discrete interface variables. The finite-dimensional test space must contain an advection-adapted minimal subspace, while its norm and implementation enable local computation.

  • Mesh setting: Shape-regular simplicial meshes and affine element mappings provide the geometric framework for transferring reference-element estimates to physical elements.The mesh family is controlled by a fixed shape-regularity bound.
  • Test spaces: The graph space consists of Lr functions whose directional derivative β · ∇v also lies in Lr, with a broken product space defined element by element.This accommodates transport-oriented test norms beyond the Hilbert setting.
  • Practical DPG method: The practical DPG method minimizes the residual over a discontinuous polynomial trial space while measuring it in the dual norm of a finite-dimensional broken test space Yh.For p = 2 the formulation admits elementwise computable linear-system or saddle-point representations; for p ≠ 2 it yields a nonlinear mixed system.
  • Test-space requirement: The discrete test space Yh may be any computationally convenient space containing the advection-adapted minimal space Vh; a simple choice is piecewise polynomials of degree r ≥ k + N.The minimal space is constructed locally and then assembled globally.

3. The discrete local test space

The discrete local test space is built from advection-dependent facet bubbles, orthogonal polynomial components, and interior bubbles on each simplex. Its dimension is minimal, with characteristic facets allowed only in the lowest-order case.

  • Advection geometry: The advection sign vector records the signs of β · ni on the facets, and every simplex has at least two non-characteristic facets.A facet is characteristic when its corresponding sign is zero.
  • Local construction: The local space combines a signed facet-bubble term bs P ⊥_k(K) with an interior-bubble term b Pk−2(K), where bs tracks inflow and outflow orientation.The orthogonal polynomial component excludes degree-k interior bubbles.
  • Global assembly: The resulting advection-adapted space is assembled elementwise into a global subspace of the broken graph space and can be embedded in piecewise polynomials of degree k+N.Any larger computationally convenient test space containing it is admissible.
  • Geometric assumption: For k ≥2, a uniformly non-characteristic facet condition is imposed across the mesh family, while k = 1 requires no such condition and permits characteristic facets.The higher-order assumption excludes facets parallel to the elementwise advection direction uniformly over the mesh family.
  • Minimality: The space has minimal dimension because its dimension equals the number of independent Fortin conditions associated with active trace and directional-derivative trial degrees of freedom.The relevant trial quantities are γ(Pk(K)) and β · ∇Pk−1(K) = Pk−2(K).

4. The Fortin operator

The Fortin operator is defined locally by a square system of moment conditions and then assembled elementwise. Unisolvency and the Fortin property follow from the structure of the minimal test space, but exact characteristic facets can destroy uniqueness at higher order.

  • Operator definition: The local Fortin operator ΠK maps each graph-space test function to the minimal space V(K) through moment conditions involving boundary operators and interior moments.For k = 1, the interior condition is vacuous.
  • Unisolvency: The defining system is square because its independent conditions equal dim P ⊥_k(K) + dim Pk−2(K) = dim V(K).Uniqueness is proved by decomposing a candidate function into facet and interior bubble components.
  • Failure case: At N = 2 and k = 2, an exactly characteristic facet can make the defining conditions non-unique because a nonzero bubble function satisfies both boundary and mean constraints.Thus the higher-order non-characteristic condition is structurally needed for this construction.
  • Fortin property: The global operator is assembled elementwise, and its correction is invisible to the discrete trial space, establishing the Fortin property.The proof uses β · ∇w ∈ Pk−2(K) and the defining moment conditions.

5. Scaling estimates

The section derives element estimates by mapping simplices to a reference element, exploiting homogeneity and compactness, and tracking affine scaling. The resulting bounds are uniform over sign patterns and shape-regular meshes, with non-characteristicity required for k ≥2 but not k = 1.

  • Affine scaling: Reference-element estimates transfer to physical simplices through affine maps while preserving the relevant polynomial and bubble structures.Pullback preserves L2 orthogonality and maps facet and interior bubbles to their reference-element counterparts.
  • Reference estimates: The reference estimates bound polynomial norms and streamline derivatives uniformly over admissible advection vectors and facet sign patterns.The derivative estimate has an upper bound proportional to |β||c|, while polynomial norms are equivalent to the coefficient norm.
  • Reference estimates: Compactness and homogeneity yield constants depending only on N, k, p, and, when needed, the non-characteristicity parameter θ.For k = 1, the lower-bound constant is independent of θ because two non-characteristic facets always suffice.
  • Affine scaling: Scaling identities relate physical-element quantities to reference-element quantities, including the streamline derivatives of facet and interior bubble components.Integration by parts connects the derivative pairings used in the element estimates.
  • Interior bubbles: Interior bubble estimates control paired Lr and Lr′ norms with constants independent of the simplex geometry.The proof uses equivalence of finite-dimensional norms on the reference simplex before transferring the estimate by change of variables.
  • Element estimates: Physical-element estimates inherit uniform dependence on N, k, p, shape regularity, and non-characteristicity, with θ-independent constants when k = 1.The physical advection scaling is controlled through the affine map and the relation between the element and reference-element vectors.

6. Continuity of Fortin and augmented Fortin operators

The Fortin operator is bounded in a β-weighted broken graph norm, and its augmented variants preserve constants or means while retaining the Fortin conditions. The construction is minimal, supports k = 1 without non-characteristic facets, and gains stronger estimates under additional moment conditions.

  • Fortin boundedness: The weighted test norm uses aK = hK/|βK| so Fortin bounds remain uniform over shape-regular meshes and invariant under rescaling β.On each fixed element this norm is topologically equivalent to the canonical broken graph norm, but with mesh- and advection-dependent equivalence constants.
  • Fortin boundedness: Theorem 6.1 establishes a mesh-uniform Fortin bound for every 1 < p < ∞, requiring non-characteristic facets only when k ≥2.The constant depends on N, k, q, θ0, and κ0, but is independent of θ0 for k = 1.
  • Refined estimates: Under additional orthogonality conditions, the operator satisfies ∥ΠKv∥Lq(K) ≤ C aK∥β · ∇v∥Lq(K) and preserves the streamline-derivative bound.These conditions remove the elementwise Lq contribution from the estimate.
  • Minimal test space: The minimal local test space is parameterized by facet and interior bubble components, with the Fortin image represented as φ = φ⊥ + ˚φ.The decomposition separates the facet-driven component from the interior bubble correction.
  • Augmented operators: Augmenting by one constant per element preserves the Fortin conditions and improves control of the mean component.The correction is designed to reproduce constants and, in the further augmentation, preserve element means.
  • Augmented operators: For k = 1 and N ≥2, the augmented space is a direct sum of the original space, constants, and one additional bubble direction.The directness argument uses a non-characteristic facet and then boundary vanishing, while the one-dimensional case is exceptional.

7. Quasioptimality of the practical DPG method

Under flow-based wellposedness assumptions, the practical DPG method is uniquely solvable and quasioptimal for 1 < p < ∞ because a uniformly bounded Fortin operator transfers stability from the continuous problem.

  • 7.1. The undiscretized problem: Flow conditions provide a curved Poincaré–Friedrichs inequality and unique solvability for transport data in Lr(Ω).The inequality controls the Lr norm by the directional derivative for functions vanishing on inflow or outflow, while inflow data determine a unique solution.
  • 7.1. The undiscretized problem: The practical method uses a β-weighted broken test norm and realizes interface variables through the operator Dh.The norm is obtained by elementwise rescaling with aK = hK/|βK|, and the discrete test spaces form a chain inside the broken graph space.
  • 7.2. Stability of the practical DPG method: The continuous bilinear form is injective, so the energy norm is a norm on X and its discrete counterpart is a norm on Xh through the Fortin property.Injectivity follows from the transport equation, trace conditions, and uniqueness under the flow assumptions.
  • 7.2. Stability of the practical DPG method: The Fortin operator is uniformly bounded in the rescaled test norm, with constant CΠ depending only on N, k, q, and mesh and flow-shape parameters, independent of θ0 when k = 1.The same type of bound holds for the mean-preserving augmented operator.
  • 7.2. Stability of the practical DPG method: The discrete solution exists uniquely and satisfies quasioptimality in the DPG energy norm, with error bounded by (1 + 2CΠ) times the best approximation error.The result applies for 1 < p < ∞ under the flow, shape-regularity, test-space, and either non-characteristic-facet or k = 1 assumptions.
  • 7.3. Convergence rates: The method achieves convergence rates in the energy norm and corresponding interior-variable Lp estimates under the stated regularity hypotheses.These conclusions follow from the quasioptimality theorem and the convergence-rate and interior-variable corollaries.

8. A posteriori error control

The computable residual representation yields a localized estimator whose global reliability and efficiency follow from the Fortin property, with efficiency constant one.

  • 8. A posteriori error control: The minimized residual is computable and serves as a natural a posteriori error estimator.Its discrete residual representative is available from the mixed formulation and can be localized into elementwise indicators.
  • 8. A posteriori error control: For all 1 < p < ∞, the estimator is globally reliable and efficient, with efficiency constant one and reliability controlled mesh-independently by the augmented Fortin constant.The reliability bound includes data oscillation, which vanishes when f is piecewise constant.
  • 8. A posteriori error control: The residual representation space contains at least one independent dimension per mesh element when N ≥ 2.This guarantees the estimator has nontrivial residual degrees of freedom rather than being identically zero.
  • 8. A posteriori error control: The bilinear map from Xh is injective, and rank-nullity yields lower bounds on the residual representation space dimension.The argument uses the global Fortin property and continuous injectivity to establish the dimension estimates.

9. Lowest order special case

For the lowest-order case, augmenting the test space and introducing a fixed test-norm scale improves L2 error behavior while retaining Fortin control and estimator reliability.

  • 9. Lowest order special case: A test-norm scale σ yields an optimal first-order L2(Ω) rate for k = 1, improving the unscaled lowest-order estimate.This is analyzed through an augmented Fortin operator for the practical DPG method.
  • 9. Lowest order special case: The scaled analysis is restricted to the Hilbert case p = q = 2 because the energy-norm characterization used is available only there.The restriction is methodological rather than a claim about the general Fortin construction.
  • 9. Lowest order special case: The scaled test norm defines corresponding scaled dual and energy norms for the residual minimization method and estimator.The practical discrete space is considered with the new norm ∥·∥Y,σ.
  • 9. Lowest order special case: Theorem 9.1 applies for any fixed σ > 0 on meshes satisfying maxK aK ≤ σ, with constants independent of β, σ, and the mesh.No non-characteristic assumption is needed in this lowest-order theorem.
  • 9. Lowest order special case: For piecewise-constant f, the scaled data oscillation vanishes, and the resulting estimator remains a reliable indicator of interior-variable error.The scaled estimator inherits efficiency and reliability from the scaled energy analysis.
  • 9. Lowest order special case: The augmented Fortin operator preserves the Fortin property and is bounded in the scaled norm by a constant depending only on N and κ0.The bound holds when σK = σ ≥ aK.
  • 9. Lowest order special case: The optimal rate requires decoupling σ from the mesh scale: σ fixed as h → 0, rather than choosing σ = maxK aK.The latter choice makes the bound diverge like h^-1, whereas the product governing the estimate has a finite optimum determined by flow constants.

10. Closing remarks

The closing remarks emphasize the minimal, same-mesh Fortin construction and its lowest-order benefits, while identifying unresolved extensions to higher-order facets and variable advection fields.

  • Construction: The facet and interior bubble ingredients make the Fortin system square and unisolvent, yielding a minimal test space without mesh refinement.The quantitative bounds rely on scaling estimates uniform over admissible flow directions and invariant under rescaling of β.
  • Lowest-order case: At k = 1, adding one constant per element and choosing σ ≥ maxK aK recovers the optimal first-order L2(Ω) rate, while mean-preserving augmentation improves the a posteriori bound.The unaugmented minimal space alone gives the vacuous mesh-independent-norm rate hmin(s,k)−1 = 1.
  • Lowest-order case: For k = 1, the analysis permits characteristic facets and imposes no directional restrictions, unlike the qualified higher-order result.For k ≥2, the result requires Assumption 3.6.
  • Open problems: Removing the higher-order qualification remains open because a uniformly bounded correction map RK has not been constructed over admissible flow directions.The required annihilation property holds for the mean projection at k = 1 but generally fails for k ≥3.
  • Open problems: The analysis assumes piecewise constant β, while extension to variable advection fields is postponed to future work.The authors also report no further numerical experiments because related numerical work is ongoing and prior DPG studies already exist.
Loading 2608.30043v1…