Source-linked AI summary
Stable Generalized Finite Element Method (SGFEM)
I. Babuska, U. Banerjee
TL;DR
GFEM enriches finite-element spaces to approximate solutions with complicated local features, but its stiffness matrices can be badly conditioned and cause severe accuracy loss in linear-system solutions. This paper proposes SGFEM, analyzes its conditioning, and reports that it is no worse than standard FEM and robust to enrichment parameters.
Problem
GFEM can provide excellent convergence but may produce stiffness matrices much more ill-conditioned than standard FEM, causing severe loss of accuracy in computed linear-system solutions.
Method
The paper proposes SGFEM, analyzes its stiffness-matrix conditioning using the scaled condition number, and evaluates it through examples and validation problems.
Results
SGFEM stiffness-matrix conditioning is not worse than standard FEM, and the method is reported to be robust with respect to enrichment parameters.
Takeaways & Limitations
The authors prefer SGFEM over GFEM because its linear system is less prone to accuracy loss when solved using an elimination method.
Abstract
from arXiv · showhide
The Generalized Finite Element Method (GFEM) is a Partition of Unity Method (PUM), where the trial space of standard Finite Element Method (FEM) is augmented with non-polynomial shape functions with compact support. These shape functions, which are also known as the enrichments, mimic the local behavior of the unknown solution of the underlying variational problem. GFEM has been successfully used to solve a variety of problems with complicated features and microstructure. However, the stiffness matrix of GFEM is badly conditioned (much worse compared to the standard FEM) and there could be a severe loss of accuracy in the computed solution of the associated linear system. In this paper, we address this issue and propose a modification of the GFEM, referred to as the Stable GFEM (SGFEM). We show that the conditioning of the stiffness matrix of SGFEM is not worse than that of the standard FEM. Moreover, SGFEM is very robust with respect to the parameters of the enrichments. We show these features of SGFEM on several examples.
I. Babuˇska ∗ U. Banerjee †
The paper concerns GFEM, PUM, XFEM, approximation, conditioning, accuracy loss, and validation and verification.
- GFEM, PUM, and XFEM are the paper's central finite-element methods.
- The paper addresses approximation, condition numbers, and loss of accuracy in linear systems.
- Validation and verification are included among the paper's concerns.
1 Introduction
The introduction motivates SGFEM as a stable modification of GFEM/XFEM that preserves strong approximation while addressing severe conditioning and accuracy-loss problems. It also presents the paper’s one-dimensional exposition, indicator, analysis plan, and validation strategy.
- Background: GFEM and XFEM are closely related Partition of Unity approaches used for problems including cracks, interfaces, heterogeneous materials, and microstructure.
- Background: GFEM enriches the standard finite-element space with compactly supported non-polynomial functions that mimic local solution features.
- GFEM conditioning: Hat-function partitions of unity can produce almost linearly dependent GFEM shape functions and stiffness matrices much more ill-conditioned than standard FEM.
- SGFEM motivation: The paper proposes SGFEM to retain GFEM’s excellent convergence while making computed linear-system accuracy loss comparable in order to standard FEM.
- Accuracy indicator: The paper uses the scaled condition number K(A)=κ2(H) as an indicator for comparing loss of accuracy across GFEM variants.
- Scope and organization: The main analysis is developed through a one-dimensional model, with higher-dimensional generalization described as future work.
2 Model problem
The paper formulates a one-dimensional variational problem on Ω=(0,1), allowing bounded smooth or rough coefficients and a solution defined up to an additive constant.
- Model setting: The model domain is Ω=(0,1), with standard Sobolev spaces H^k(Ω) used to describe the variational setting.
- Model assumptions: The coefficient a(x) is assumed bounded between positive constants, while it may be smooth or rough.
- Solution properties: The variational problem has a unique solution up to an additive constant, and its energy norm is defined on subdomains.
- Strong formulation: Under differentiability of au′, the variational solution is also characterized by a strong-form boundary value problem.
3 Generalized Finite Element Method (GFEM):
GFEM constructs a Partition of Unity approximation by augmenting a standard finite-element space with locally selected enrichment functions. Its local approximation flexibility supports strong accuracy, but the resulting stiffness matrix can be severely ill-conditioned and cause accuracy loss.
- Galerkin framework: The Ritz-Galerkin method approximates the model solution in any chosen finite-dimensional subspace of H1(Ω).
- Partition of Unity Method: A PUM combines a partition of unity with local approximating spaces, whose local accuracy controls the global energy-norm accuracy.
- GFEM construction: GFEM uses finite-element stars and piecewise-linear hat functions as its partition of unity, with the standard FEM space as its basic part.
- Enrichments: Enrichment functions are selected to mimic local solution behavior, including polynomial, interface-adapted, singular, or discontinuous functions.
- Flexibility: The framework permits nonuniform local enrichment degrees and can be formulated on locally quasi-uniform meshes.
- Conditioning: The scaled condition number K(A) is used because the GFEM stiffness matrix can be much worse conditioned than standard FEM, whose condition number is O(h^-2).
- Conditioning example: For a singular enrichment example, the GFEM reproduces the exact solution with no approximation error while still exhibiting severe conditioning concerns.
4 Stable Generalized Finite Element Method (SGFEM):
SGFEM modifies GFEM enrichment spaces to preserve approximation quality while controlling stiffness-matrix conditioning. The construction approximates u−I_hu locally, achieves optimal convergence in the example, and yields conditioning of the same order as standard FEM under stated assumptions.
- Approximation and convergence: ∥u−u_h∥_E(Ω) = O(h^2) for the example, matching the stated optimal convergence order.The result follows from the local approximation construction and Proposition 4.1.
- Conditioning analysis: Orthogonality of S1 and S2 makes the coupling blocks A12 and A21 zero in the example.This structure is used in the conditioning analysis, but the paper notes that exact orthogonality does not hold generally.
- Conditioning analysis: K(A) is of the same order as K(A_11), the scaled condition number of the standard-FEM basic part.The analysis establishes matching orders for the enriched GFEM matrix and its basic part.
- SGFEM construction: SGFEM enrichments approximate u−I_hu locally, rather than approximating u directly as in standard GFEM.This modified local target is identified as the main idea of SGFEM.
- Approximation and convergence: The construction attains O(h) convergence without a ramp-function in the cited crack-propagation setting.The paper contrasts this with Corrected XFEM, where a ramp-function is required for that rate in crack propagation problems.
- General result: Under Assumptions 1, 2, and 3, the GFEM with S = S1 + S2 is an SGFEM.The general result extends the example’s conclusion to specified local approximation and enrichment spaces.
5 Applications:
The applications analyze SGFEM for one-dimensional problems with discontinuous coefficients and singular solutions. Across these cases, the method retains convergence while achieving stiffness-matrix conditioning of order O(h^-2), including enrichment parameters near 0 or 1.
- Applications: The applications use a uniform one-dimensional mesh and analyze the scaled stiffness-matrix condition number for several problem classes.The examples include piecewise constant coefficients and singular solutions.
- Discontinuous coefficients: For discontinuous-coefficient cases, the enrichment parameters satisfy the required assumptions uniformly for 0 < β < 1.The constants in the assumptions are independent of β, including when β approaches 0 or 1.
- Discontinuous coefficients: K(A) = O(h^-2) for the analyzed discontinuous-coefficient problems, matching the standard finite-element conditioning order.This conclusion holds even when the discontinuity point is close to mesh vertices.
- Singular solutions: For the singular-solution example, the SGFEM solution satisfies ||u - uh||_E(Ω) = O(h) when the singularity neighborhood is fixed.The estimate is obtained using local approximation and standard interpolation results.
- Singular solutions: If the enriched singularity neighborhood has size h^γ with γ < 1, the error becomes O(h^(1−γ)), so enriching only finitely many nearby patches loses optimal convergence.The neighborhood is otherwise assumed independent of h.
- SGFEM construction: The SGFEM construction combines the standard space S1 with an enrichment space S2 and satisfies the framework's assumptions in the presented examples.The resulting method is identified as SGFEM in the discontinuous-coefficient analyses.
6 Conclusion
The conclusion presents SGFEM as a modification that preserves GFEM's convergence advantages while avoiding severe accuracy loss from ill-conditioning. The paper reports robustness to enrichment parameters, while noting that the analysis is demonstrated in one dimension.
- Conclusion: GFEM enrichments can provide excellent convergence but may produce ill-conditioned stiffness matrices and severe loss of accuracy in the computed linear system solution.The loss can exceed that experienced with standard FEM.
- Conclusion: SGFEM is proposed as a modification that retains GFEM's advantages without the severe loss-of-accuracy problem.The paper characterizes loss of accuracy through the scaled condition number and validates Hypothesis H on examples.
- Conclusion: SGFEM is reported to be robust with respect to enrichment parameters such as β.This robustness is stated alongside the retained convergence properties.
- Scope: The abstract framework is applied to a one-dimensional problem for clarity, with higher-dimensional applications deferred to future work.The authors state that the framework and analysis could be generalized to higher dimensions.
7 Appendix
The appendix validates Hypothesis H for selected FEM, GFEM, and SGFEM problems, relating relative computed-solution error to a scaled condition number. The results also show that validation depends on the tested meshes, tolerances, software, and arithmetic setting.
- Hypothesis H: Hypothesis H models relative error as η = Cn^βK(A)ε, with bounded C and small |β| over the range where accuracy is not almost entirely lost.Validation uses tolerances on the coefficient ratio and exponent, with the exponent tolerance identified as primary.
- Conditioning indicator: The scaled condition number K(A) is used because κ2(D∗AD∗) more accurately reflects η than the unscaled matrix condition number κ2(A).The practical indicator uses diagonal scaling with Dii = Aii^-1/2.
- Validation results: For Problem 3, the observed GFEM error satisfies C̄1N^4 ≤ η ≤ C̄2N^4 with C̄2/C̄1 ≤ 340, but all accuracy digits are lost for N ≥ 9000.Beyond that range, η is of order 1 and oscillates around 1, so Hypothesis H does not address those meshes.
- Interpretation and scope: Validation outcomes depend on the chosen tolerance and mesh values, while software changes can also alter computed errors, as observed for MUMPS beyond N > 5 × 10^3.The authors accept SGFEM over GFEM under a tolerance τ1 = 500 because GFEM error is much larger for large N.
- Conclusion: The authors report confidence in Hypothesis H for Problems 1–4 and prefer SGFEM because its elimination-method linear systems are less prone to accuracy loss than GFEM systems.They also note that more substantial validation, including other two- and three-dimensional problems, is future work.