PaperlayerView exact revision on arXiv

Source-linked AI summary

One-Sided $p$-Explicit Stability for High-Order Cut Finite Element Nitsche Methods

Erik Burman, Peter Hansbo, Mats G. Larson, Karl Larsson

arXiv:2610.34853v1math.NAcs.NA

TL;DR

Cut elements can prevent the direct use of fitted-element inverse-trace arguments needed for symmetric Nitsche stability. The paper uses a patch projection to control a polynomial on the physical side and stabilize only its residual on complete elements. This yields an explicit p^2/h flux and penalty scale, alongside approximation estimates and bounded normalized flux constants in the reported disk experiments.

  • Problem

    Arbitrarily small physical intersections in cut elements prevent directly applying the fitted-element inverse-trace argument used to control the normal flux for symmetric Nitsche stability.

    Source: §1 · p. 1

  • Method

    A patch H^1-projection separates a polynomial controlled on inward physical segments from a residual controlled by gradient stabilization on complete elements.

    Source: §1 · p. 2

  • Results

    The flux estimate gives a cut- and degree-independent sufficient Nitsche penalty at scale p^2/h, while translated-disk experiments report unit-stabilization flux constants of 0.560–0.892 for p ≤ 8 and accurate degree decay through p = 12.

    Source: §1 · p. 2, §7 · p. 23

  • Takeaways & Limitations

    The stability proof avoids the exponential cost of full-element polynomial continuation and yields hp-error estimates plus exponential degree convergence for analytic extensions on sufficiently fine meshes.

    Source: §3 · p. 7, §7 · p. 23

    The coercivity result does not imply a cut-independent stiffness-matrix condition number because its norm is not uniformly equivalent to the Euclidean norm of arbitrary nodal coefficients.

    Source: §4.4 · p. 12

Abstract

from arXiv · show

We give a one-sided stability argument for high-order cut finite elements with symmetric Nitsche boundary conditions. On each boundary patch, an $H^1$-projection splits the discrete function into a single polynomial and a stabilized residual. Inward segments control the polynomial inside the physical domain; complete-element estimates control only the residual. For smooth exterior boundaries and under a geometry-dependent mesh restriction, this gives an explicit normal-flux estimate with factor $p^2/h$ and a sufficient Nitsche penalty with the same scaling. The constants are independent of the cut position and polynomial degree. A neighboring-interval estimate quantifies the exponential cost of polynomial continuation and motivates avoiding it in the stability proof. We obtain $hp$-error estimates, with half a power of degree lost in energy and a full power in $L^2$ relative to best approximation, and exponential degree convergence for analytic extensions on sufficiently fine meshes. Translated-disk experiments show bounded normalized flux constants and penalty thresholds over the tested degrees and cuts, and resolved convergence curves follow the predicted orders. Physical and extension-dependent stabilization errors are reported separately.

1 Introduction

Cut intersections obstruct the fitted-element inverse-trace argument behind symmetric Nitsche stability. The paper proves a one-sided flux estimate by controlling a patch polynomial inside the physical domain and stabilizing only its residual on complete elements.

Main Results

  • Full-element continuation can grow exponentially with degree, so that intermediate estimate may yield an exponentially large sufficient penalty without implying the method requires one.The paper motivates its stability proof by avoiding continuation of the projected polynomial beyond the physical domain.Source: §1 · pp. 1–2
  • A patch H^1-projection separates one polynomial, controlled on inward physical segments, from a residual controlled by gradient stabilization on complete elements.The Nitsche form remains symmetric; “one-sided” describes the polynomial’s physical-side control.Source: §1 · p. 2
  • Physical-domain coercivity does not guarantee a cut- or p-independent stiffness-matrix condition number; conditioning is a separate issue addressed by preconditioning approaches.The translated-disk experiments test flux and convergence scaling, while their meshes do not meet the sufficient restriction in the explicit disk calculation.Source: §1 · p. 2
  • The explicit normal-flux bound yields symmetric Nitsche coercivity with penalty scale p^2/h and a sufficient coefficient independent of cut position and degree.The result requires h ≤ h0, where h0 is determined by boundary geometry and mesh regularity.Source: §1 · p. 2
  • The analysis gives hp-energy and L2-error estimates and exponential degree convergence for analytically extendable solutions on sufficiently fine meshes.Relative to optimal approximation, the energy and L2 bounds lose half a power and a full power of p, respectively.Source: §1 · p. 2

Position Relative to the Closest Work

  • The result addresses a specific gap: an exterior boundary has no complementary physical subcell for the one-sided interface-flux strategy of Wu and Xiao.The patch projection and inward-segment argument replace that missing side while preserving p^2/h scaling.Source: §1 · p. 2

2 Problem Setting and Discrete Method

The method uses a continuous piecewise-polynomial space on a background mesh and stabilizes boundary patches built from complete active elements. Each patch projection reproduces one polynomial, separating it from a residual for stabilization in the symmetric Nitsche formulation.

2.1 Model Problem

  • The model assumes a bounded C2 domain in two or three dimensions and exact-domain integrals, excluding geometry-approximation and quadrature errors.The background meshes are shape-regular, quasi-uniform affine simplicial meshes, and the mesh size is restricted by a geometry-dependent threshold.Source: §2 · p. 3

2.2 Active Mesh and Finite Element Space

  • The active finite element space is the restriction to the active mesh of continuous functions that are degree-p polynomials on each background element.The background mesh is a shape-regular, quasi-uniform family of affine simplicial meshes.Source: §2.2 · p. 3

2.3 Constructed Boundary Patches

  • Boundary patches are face-connected unions of complete active elements covering inward tubes, with uniformly bounded element counts and cut-independent geometric constants.The construction partitions the boundary into cells, enlarges them along the boundary, and sweeps the cell and enlargement inward to form the tubes.Source: §2.3 · p. 4, Figure 1 · p. 4

2.4 Patch Projection and Stabilization

  • On each patch, the H1-projection maps onto restrictions of single global degree-p polynomials and reproduces those polynomials.The stabilization seminorm vanishes on global polynomials; mean preservation and a uniform Poincaré inequality make a separate L2-residual penalty unnecessary.Source: §2.4 · p. 6

2.5 Nitsche Formulation

  • The patch projection and gradient stabilization define a weakly consistent symmetric Nitsche method whose stabilization contribution is retained in the Strang estimate.That contribution is bounded using polynomial reproduction and local approximation.Source: §2.5 · p. 6

3 The Full-Element Continuation Barrier

Standard full-element transfer can be exponentially costly for high-degree polynomials, complicating stability proofs that extend control beyond the physical domain. The section quantifies neighboring-interval continuation and explains why avoiding that intermediate estimate matters for the penalty bound.

Sharp Order of Neighboring-Interval Continuation

  • Neighboring-interval L2 continuation has a sharp exponential order with an algebraic prefactor, with constants in the stated comparison independent of p.The lower-bound construction uses a Chebyshev polynomial; the passage does not state a full formula for the order.Source: §3 · p. 7
  • The Chebyshev trial polynomial establishes the sharp exponential base and algebraic prefactor but does not establish an asymptotic equivalent for the optimal continuation constant.A simpler trial polynomial provides a lower bound but loses the sharp exponential base.Source: §3 · p. 8

Implication for Stability

  • Combining full-element continuation with a complete-element inverse trace inequality yields an exponentially large sufficient Nitsche penalty.This is a limitation of that proof mechanism, not evidence that the discrete method itself requires an exponential penalty.Source: §3 · p. 8
  • The continuation barrier motivates replacing full-element transfer with a one-sided estimate that controls the projected polynomial only inside the physical domain.The standard approach first extends control to complete active elements before applying an inverse trace inequality.Source: §3 · pp. 7–8

4 One-Sided Stability Analysis

The stability analysis controls the projected polynomial on inward segments inside the physical domain and the stabilized residual on complete elements. These controls yield a cut-independent normal-flux estimate and sufficient Nitsche coercivity under a geometry-dependent mesh restriction.

4.1 Endpoint Inverse Control on Inward Segments

  • The projected polynomial’s normal flux is controlled using its gradient on inward segments inside the physical domain, while the residual is controlled on the complete patch.The endpoint inverse argument gives a sharp square-root norm constant of m + 1.Source: §4.1 · p. 9, Figure 2 · p. 9

4.2 Control of the Projection Residual

  • Complete-element inverse trace estimates control the projection residual with a p^2/h scaling and a constant independent of cut position, h, and p.The estimate requires h ≤ h0, set by boundary geometry and mesh regularity.Source: §4.2 · p. 10

4.3 The One-Sided Normal-Flux Estimate

  • The one-sided normal-flux estimate uses the full-element trace estimate only for the stabilized residual, keeping polynomial continuation out of the bound.The resulting constant depends on fixed geometric parameters and is independent of the cut configuration, h, and p.Source: §4.3 · pp. 10–11

4.4 Coercivity and Continuity

  • For h ≤ h0, a sufficient Nitsche penalty gives cut-independent coercivity, with constants independent of h and p.The estimate does not imply a cut-independent stiffness-matrix condition number because the coercivity norm is not uniformly equivalent to the Euclidean norm of arbitrary nodal coefficients.Source: §4.4 · p. 12

5 hp-Error Analysis

The analysis builds conforming and elementwise polynomial approximants to bound approximation errors in the augmented and stabilization norms. These bounds feed Strang-based energy estimates and a duality argument for the L2 estimate, with a separate construction for analytic extensions.

5.1 Approximation Tools

  • Conforming and possibly discontinuous piecewise polynomial approximants are used separately to support the hp approximation bounds.The conforming approximant belongs to the active-mesh space; patchwise polynomial approximation supplies the local estimates.Source: §5.1 · p. 13
  • The normal-trace contribution is higher order in p than the boundary-value contribution after multiplication by h/p^2.The proof combines the trace estimate with approximation bounds to establish the augmented-norm estimate.Source: §5.1 · p. 14

5.2 Strang Estimate and Energy Error

  • The Strang estimate and approximation lemmas yield an hp-energy estimate for extensions in H^s with s ≥ 2.Its constant may depend on s, but is independent of the cut configuration, h, and p.Source: §5.2 · pp. 14–15
  • For analytic extensions, the error decays exponentially with p on each fixed mesh satisfying θ = CaAeh ≤ θ0 < 1.A direct approximation construction avoids substituting s = p + 1 into an estimate whose constant depends on s.Source: §5.2 · p. 15

5.3 L2-Error Estimate

  • The L2-error estimate uses dual regularity and bounds the error under the assumptions of the hp-energy theorem.The proof uses an H2 extension of the dual solution and interpolation for noninteger regularity.Source: §5.3 · p. 16
  • Relative to optimal approximation, the energy estimate loses half a power of p and the L2 estimate loses a full power.These are upper-bound losses; faster observed rates do not establish that either loss is necessary.Source: §5.3 · p. 16

6 Numerical Results

The experiments examine continuation, flux control, penalty thresholds, convergence, and conditioning on translated disk cuts. They report bounded flux constants and penalty thresholds over tested cuts and degrees, convergence behavior consistent with predicted orders in resolved ranges, and strong cut-related growth in coefficient conditioning.

6.1 Continuation and the One-Sided Flux Constant

  • At p = 14, the computed continuation constant is 9.57615 × 10^9, about 18% above the Chebyshev lower bound of 8.14801 × 10^9.Continuation has an exponential asymptotic slope, whereas endpoint control grows linearly.Source: §6.1 · p. 17, Figure 3 · p. 17
  • With stabilization, normalized flux maxima remain bounded over tested degrees; without it, individual-basis quotients grow strongly as cells become small.The unit and computational stabilization measurements lie in [0.560, 0.892] and [0.735, 3.741], respectively; the sliver witness scales as ηmin^-1/2.Source: §6.1 · p. 18, Figure 4 · p. 18

6.2 Penalty Thresholds Across Cuts and Degrees

  • Thresholds remain bounded down to ηmin = 7.50 × 10^-10; the largest axis-path threshold falls from 1.4263 at p = 2 to 0.9974 at p = 6.At matching degrees, axis and diagonal cuts need no larger penalty than the fitted control, though the largest oblique p = 2 endpoint is higher.Source: §6.2 · p. 19, Figure 5 · p. 19, Table 1 · p. 19

6.3 Mesh and Degree Convergence

  • Resolved mesh slopes on disk A agree with the predicted p + 1 order in L2 and p order in gradient and energy.Slopes above the reference orders are treated as pre-asymptotic observations, not stronger asymptotic rates.Source: §6.3 · p. 20, Figure 6 · p. 20, Table 2 · p. 20
  • At N = 6, exponential degree-convergence factors are approximately 0.06 in L2 and 0.09 in energy over the resolved ranges.The two disks have nearly coincident curves; the factors are measured fits, not sharp predictions from the Taylor-based bound.Source: §6.3 · p. 20, Figure 7 · p. 21
  • At p = 8, the extension seminorm is 3.7% larger on disk A and 4.2% larger on disk B than the stabilization-scale error seminorm.The reported energy errors separate physical and full energy contributions; this comparison does not make full energy independent of the chosen extension.Source: §6.3 · pp. 20, 22, Figure 8 · p. 21
  • For finite-regularity degree convergence, L2 slopes are −3.33 and −3.36 on disks A and B; gradient slopes are −2.56 and −2.58, and energy slopes are −2.59 and −2.61.These finite-window fits do not establish asymptotic sharpness of the upper bounds.Source: §6.3 · p. 22, Figure 9 · p. 22

6.4 Practical Accuracy and Conditioning

  • At p = 6, the smallest axis cut has Jacobi condition number 4.65 × 10^6 versus 132.6 for the fitted control, despite stabilization and patch-layer variations.The fitted geometry differs, and the comparison does not establish an asymptotic law.Source: §6.4 · p. 23, Table 4 · p. 23

7 Conclusions

The paper avoids exponential continuation costs in symmetric Nitsche stability through a patchwise polynomial-residual split, yielding explicit flux and penalty bounds under a geometry-dependent mesh restriction. Translated-disk experiments report bounded normalized flux constants and resolved convergence, while identifying limits on what the experiments certify.

7 Conclusions

  • The patchwise projection controls the polynomial on inward physical segments and the stabilized residual on complete elements, producing a p^2/h flux estimate and sufficient Nitsche penalty under a geometry-dependent mesh restriction.The analysis also gives a Strang estimate, hp-energy and L2-error estimates, and exponential degree convergence for analytic extensions on sufficiently fine meshes.Source: §7 · p. 23
  • Unit-stabilization flux constants lie in [0.560, 0.892] for p ≤8, and accurate assembly resolves degree decay through p = 12.Matched penalty thresholds remain below measured computational flux constants; physical and extension-dependent stabilization errors are reported separately.Source: §7 · p. 23
  • The explicit analytic constant is not certified on the tested meshes, and the fitted-polygon comparison addresses stability and conditioning rather than disk accuracy.Same-domain accuracy comparisons and smooth nonconvex tests are outside the study's scope, and dense computations do not establish a scalable-solver claim.Source: §7 · p. 23

A Verification Protocol and Patch Geometry

The verification material documents reproducible numerical procedures and certifies tube coverage on the saved geometries. Its explicit disk trace and penalty calculation illustrates the analytic bounds but does not certify the constants on the experimental meshes.

A Verification Protocol and Patch Geometry

  • The companion guide maps studies to verification sections and provides case parameters, reference results, data files, and commands for reproduction and export.Four-layer patches need no reassignment.Source: A Verification Protocol and Patch… · p. 23, Table 5 · p. 24, Table 6 · p. 25
  • The baseline accuracy protocol compares two assembly rules within 1% of the smaller discretization error and independently regenerated error rules within 10^-4 relatively.The accurate Figure 7 points use analytic circle moments, high-precision patch projections, iterative refinement, and coefficient storage as sums of two binary64 numbers.Source: A Verification Protocol and Patch… · p. 24
  • An exact full-arc certificate establishes tube depth h/16 on all 30 saved geometries in the standard PDE study, after reassignment of 97 of 5324 open angular pieces resolves the two-layer seed failures.The seed assignment failed on 18 of 24 two-layer geometries; the reassignment uses existing neighboring patches without changing the method.Source: A Verification Protocol and Patch… · pp. 24–25
  • For the disk, h ≤ ϱR/6 gives c* = ϱ/2; at R = 0.62 this requires h ≤ approximately 0.03027, finer than the smallest sampled spacing of 1/12.The sufficient restriction follows from the crude Taylor bound and is not a restriction of the finite element method; the analytic constant is not certified on the experimental meshes.Source: A Verification Protocol and Patch… · p. 25
  • The unit-stabilization analytic bound gives β* ≃ 311.1, whereas the computational penalty was chosen as β0 = 16; the gap reflects conservative bounds and the polynomial-residual split.For the disk tube, the exact lower Jacobian bound is 1 − h/(16R), compared with the conservative bound cJ = 1/2.Source: A Verification Protocol and Patch… · p. 25

B Explicit Constant in the Complete-Element Trace Estimate

The trace argument constructs inward segments with a uniform transverse direction, then uses them to control polynomial gradients inside the physical domain. A weighted polynomial estimate supplies the degree dependence in dimensions two and three.

B Explicit Constant in the Complete-Element Trace Estimate

  • Shape regularity supplies a point P in each cut element at distance at least ϱh from the tangent plane, and the C2 boundary condition preserves a common transverse sign when h is sufficiently small.The resulting sign is common across the boundary intersection, including distinct connected components.Source: B Explicit Constant in the… · pp. 25–26
  • The map Ψ(y,t) = P + t(y − P) sends inward segments into the element, and convexity plus the common transverse sign makes it injective for applying the area formula.A ray from P cannot meet the boundary twice: consecutive transverse crossings would require opposite outward-normal components along the ray.Source: B Explicit Constant in the… · p. 26
  • Applying the weighted estimate to polynomial components of degree at most p − 1 yields the factor p(p + d − 1) ≤ dp^2 for p ≥ 1 in dimensions two and three.The argument uses the weighted estimate with m = p − 1 and α = d − 1.Source: B Explicit Constant in the… · p. 26

C A Physical Cone Alternative

The physical-cone alternative replaces a single inward normal tube with finitely many inward cones, subject to a bounded-multiplicity condition. The extension is conditional and does not replace the main results' C2 assumptions.

C A Physical Cone Alternative

  • The alternative permits corners, but the main theorems retain C2 assumptions and the argument does not supply dual H2 regularity at reentrant corners.Uniform cone height, multiplicity, and patch geometry remain construction requirements.Source: C A Physical Cone Alternative · pp. 26–27
  • Each boundary piece must admit an inward apex whose segments remain in the patch and physical domain, with the cone maps having bounded total volume multiplicity.The boundary pieces are Lipschitz, with outward normals defined almost everywhere.Source: C A Physical Cone Alternative · p. 26
  • The cone estimate controls polynomial gradients by applying the weighted polynomial bound along each segment and summing over pieces under the multiplicity bound.A cone-based flux proof can use Ccone = dMc/cc alongside complete-element trace control and residual control.Source: C A Physical Cone Alternative · p. 27
Loading 2610.34853v1…