Pith. sign in

REVIEW 3 major objections 4 minor 16 references

Stabilization of isogeometric finite element method with optimal test functions computed from $L_2$ norm residual minimization

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read By combining standard Galerkin testing with L2-optimal test functions, this paper derives a stabilization for isogeometric B-spline discretizations that needs no tuning parameters and matches SUPG on advection-dominated diffusion.

desk verdict A broken first model problem invalidates the headline numerical comparison; the stability proof is a plausible separate contribution. read the letter →

arxiv 2411.15565 v1 pith:TV2GD2UU submitted 2024-11-23 math.NA cs.NA

Marcin {L}oś , Tomasz S{l}użalec , Maciej Paszyński , Eirik Valseth This is my paper · ORCID
classification math.NAcs.NA MSC 65N3065N1265D07
keywords isogeometricanalysisB-splinebasisfunctionsadvection-diffusionequationPetrov-Galerkinoptimaltestleast-squaresfiniteelementmethodGalerkin/least-squaresstabilizationSUPGcoercivity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Advection-dominated diffusion is a standard test case where plain Galerkin finite element methods produce spurious oscillations unless the mesh is carefully adapted. This paper argues that a stabilization scheme for isogeometric (B-spline) discretizations can be built directly from residual minimization: using the $L_2$ norm, the optimal test function for each trial basis function is just the advection-diffusion operator applied to that basis function, which turns the method into a least-squares formulation. Combining that with an ordinary Galerkin term yields a Galerkin/least-squares method whose coercivity is proved for B-spline spaces under a mesh-size-diffusion relation, with no tuning parameter. The paper's numerical comparison on two benchmark problems with coarse uniform meshes indicates this combined scheme is as accurate as SUPG and sometimes better, especially for small diffusion. If the claim is right, stabilized simulations on B-spline meshes can be set up without hand-tuning stabilization constants.

What carries the argument

The central object is the $L_2$-optimal test function. For each trial basis function $e_i$, minimizing the residual in $L_2$ yields the test function $\hat{e}_i = \partial e_i/\partial x - \varepsilon(\partial^2 e_i/\partial x^2 + \partial^2 e_i/\partial y^2)$, the image of the trial function under the advection-diffusion operator; this explicit formula turns the Petrov-Galerkin method into a least-squares method without solving auxiliary problems. The stabilized Galerkin/least-squares variant adds a $1/h$-weighted Galerkin term to this optimal testing, and the coercivity proof rests on an inverse inequality for B-splines, $\|\Delta v_h\|_{L_2} \le C h^{-1}\|\nabla v_h\|_{L_2}$, whose constant $C$ enters the stability condition $\varepsilon \le h/(2C^2)$.

What would settle it

Substitute the exact solution in Eq. (2) into the PDE in Eq. (1) with $\beta=(1,1)$, including the y-advection term, and evaluate the residual; if the residual is not zero, the benchmark and all error tables do not measure the model problem the paper intends to solve.

Watch

Extended reading notes

Core claim

The paper's central claim is that $L_2$-norm residual minimization provides an explicit, analytic family of optimal test functions for advection-diffusion problems, and that combining these with standard Galerkin test functions produces a stabilized isogeometric method with proven stability. In the $L_2$ setting, the optimal test function for a B-spline basis function $e_i$ is the image of $e_i$ under the differential operator, $\partial e_i/\partial x - \varepsilon(\partial^2 e_i/\partial x^2 + \partial^2 e_i/\partial y^2)$; the resulting Petrov-Galerkin method is exactly a least-squares formulation. Adding a mesh-weighted Galerkin term with a $1/h$ factor gives the Galerkin/least-squares scheme, and the paper proves a coercivity estimate $b(v_h, v_h) \ge \frac{1}{2}((\varepsilon/h)\|\nabla v_h\|^2 + \|\beta\cdot\nabla v_h\|^2)$ for $\varepsilon \le h/(2C^2)$, where $C$ comes from a B-spline inverse inequality. The numerical experiments on two advection-diffusion benchmarks with ten uniform elements show the combined scheme matches or beats SUPG in $L_2$ accuracy for the smallest diffusion coefficients tested, and it requires no stabilization parameter to tune.

Load-bearing premise

The load-bearing premise is that the benchmark exact solutions stated in the paper really do solve the model problems as written, so the reported relative errors compare each method against the correct target.

Editorial extensions

If this is right

  • On a uniform mesh of ten elements, the combined Galerkin/least-squares formulation gives the lowest $L_2$ errors among all tested methods for the smallest diffusion parameters in both benchmark problems.
  • The pure $L_2$-optimal-test (least-squares) method is unconditionally stable but overly diffusive on coarse meshes for very small $\varepsilon$, so the combination, not the pure form, is the practically useful scheme.
  • The stability theorem guarantees coercivity whenever $\varepsilon \le h/(2C^2)$, so the mesh-to-diffusion ratio, not a user-tuned constant, determines stability.
  • The paper's comparisons indicate the combined method requires no tuning of stabilization parameters, unlike SUPG, which needs the parameter $\tau$ to be chosen.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the coercivity proof transfers to other B-spline spaces satisfying the same inverse inequality, the no-tune stabilization could extend to locally refined, trimmed, or higher-dimensional isogeometric meshes without new parameter choices.
  • Because the $L_2$-optimal test function is known in closed form, the same operator-action identity could be reused as a cheap residual or error indicator for adaptive mesh refinement; the paper does not explore this.
  • The mesh-dependent $1/h$ coupling suggests the stabilization strength is encoded in the discretization, which may make the method easier to port to time-dependent or nonlinear advection-diffusion problems where SUPG constants are usually re-tuned.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper studies stabilization of isogeometric finite element methods for advection-diffusion problems. It derives a least-squares method from Petrov-Galerkin optimal testing in the L2 norm, combines it with the standard Galerkin term to obtain a Galerkin/least-squares method, proves a coercivity bound for B-spline spaces, and compares the resulting schemes with SUPG on two model problems. The theoretical stability analysis in Section 4 is a self-contained argument based on an inverse inequality for B-splines. However, the numerical part is seriously compromised: the first benchmark problem is mis-specified, with an exact solution that does not satisfy the stated PDE, weak forms that omit the y-advection term for beta=(1,1), and an abstract and conclusions that contradict the reported error tables.

Significance. If the stability analysis and the numerical comparisons were correct, the paper would offer a parameter-free stabilized isogeometric variant with a rigorous coercivity bound. The coercivity proof for the Galerkin/least-squares form in Section 4 is a useful and largely self-contained contribution. However, the paper's central comparative claim—that the proposed Galerkin/least-squares method is competitive with or better than SUPG without tuning—rests on Tables 11 and 12, and for the first model problem those tables are obtained from an incorrectly posed benchmark. The second model problem is correctly posed, but it alone does not establish the headline conclusions, which are also in direct contradiction with the table values.

major comments (3)
  1. [Section 2.1.1, Eq. (2)] The exact solution in Eq. (2) is not a solution of the stated homogeneous PDE (1) with beta=(1,1). For f(t)=t+(e^{t/epsilon}-1)/(1-e^{1/epsilon}) one has -epsilon f''+f'=1, so for u=f(x)f(y) the left-hand side of (1) equals f(x)+f(y), which is not zero. With zero Dirichlet data, the unique solution of the homogeneous equation (1) is u identically zero. Consequently, the relative errors reported in Tables 1, 3, 5, 7, 9, and 11 for the first model problem are computed against a function that is not the exact solution of the problem being solved, and the comparison for that benchmark is meaningless.
  2. [Section 2.2, Eqs. (4), (5), (12), and (14)-(16)] The weak forms used throughout the paper contain only the x-advection term and the two diffusion terms; the y-advection term is absent. For the first model problem with beta=(1,1), this means the implemented method solves a different boundary value problem, namely u_x - epsilon(u_xx+u_yy)=0 with zero Dirichlet data, whose only solution is zero. The non-zero computed solutions reported in the tables indicate that the actual right-hand side or boundary data used in the implementation differ from the equations stated in the paper, so the experiments are not reproducible from the written formulation.
  3. [Abstract and Section 5 vs. Tables 11-12] The abstract states that least-squares (Petrov-Galerkin with L2-optimal test functions) outperforms the other stabilization methods for small Peclet numbers, and the conclusion states that the optimal test functions behave better than SUPG for epsilon=0.1 and 0.01. Tables 11 and 12 show the opposite: for example, Table 11 at epsilon=0.01 gives L2 errors 86.70 (PG-optimal), 28.20 (GLS), and 33.71 (SUPG); Table 12 at epsilon=0.01 gives 52.87, 17.10, and 20.88, respectively. The headline claims are thus not supported by the paper's own data.
minor comments (4)
  1. [Section 2.1, Eq. (1)] Equation (1) writes 'beta_y d u/d x' for the second advection term; presumably this should be 'beta_y d u/d y'.
  2. [Figure 7 caption] The caption reads 'epsilon in {0.1, 0.001, 0.0003}' while the text and Table 3 use {0.1, 0.01, 0.003}.
  3. [Section 2.4, Table 8] Table 8 includes a row for epsilon=0.1 for the second problem, although Section 2.1.2 defines the second problem for epsilon in {0.01, 0.001, 0.0001}.
  4. [Section 4, Lemma 2] The continuity bound for the term h^{-1}(beta·grad u_h, v_h) is K|beta|/epsilon, which grows as epsilon tends to zero; the phrase 'provided sufficiently small epsilon > 0' is therefore misleading, since the inequality itself holds for any fixed epsilon with a constant that depends on epsilon.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central derivation is self-contained and the stability proof is independent.

full rationale

The L2-optimal test functions are derived in Sections 2.2 and 2.3 directly from Riesz representation and the Fourier lemma (Eqs. (7)-(11)), with no fitted parameters or renamed predictions. The Galerkin/least-squares form combines this optimal test with a boundary-vanishing stabilization term, and Section 4 proves coercivity using an external inverse inequality [3], Young's inequality, and Poincare's inequality, with the condition epsilon <= h/(2C^2) stated explicitly. Self-citations [11,14] are used for background and inspiration but are not load-bearing in the derivation. Separate concerns about the correctness of the first benchmark's exact solution and omitted y-advection terms are noted as benchmark-correctness issues, not circularity of the method derivation.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The proposed GLS method introduces no new physical or mathematical entities and has no fitted stabilization parameter, which is a strength. The only tuned parameter appears in the SUPG baseline. The stability theorem relies on an inverse inequality from lecture notes that is not machine-checked, and on standard PDE inequalities. The benchmark inconsistency is not a matter of free parameters but of an incorrect statement of the model problem.

free parameters (1)
  • tau_SUPG = tau^{-1} = beta_x/h_x + beta_y/h_y + 3*epsilon/(h_x^2 + h_y^2)
    The SUPG stabilization parameter used in the baseline comparison is taken from the literature and includes a hand-set factor of 3. It is not fitted to the benchmarks, but it affects the SUPG results that the proposed method is compared against.
assumptions (5)
  • standard math Riesz representation theorem guarantees existence and uniqueness of optimal test functions in L2
    Used in Section 2.3 to define the optimal test function for each trial basis function via Eq. (7).
  • standard math Fourier lemma identifies the optimal test function with the residual operator applied to the trial basis
    Used in Section 2.3, Eqs. (8)-(11), to infer the analytic form of the optimal test function.
  • domain assumption Inverse inequality for B-splines: |v_h|_{H1} <= 2*sqrt(3d)*p^2*h^{-1}*||v_h||
    Invoked in Lemma 1 and cited to unpublished lecture notes [3]. The constant C = 2*sqrt(3d)*(p-1)^2 is load-bearing for the coercivity theorem and the stability condition epsilon <= h/(2*C^2).
  • standard math Poincare inequality ||v|| <= K*||grad v|| for v in H1_0
    Used in Lemma 2 to bound the advective term in the continuity estimate.
  • standard math Skew-symmetry of the advection form: (beta*grad v, v) = 0 for v vanishing on the boundary
    Used in Theorem 1 to eliminate the (1/h)(beta*grad v, v) term in the coercivity proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stabilization of isogeometric finite element method with optimal test functions computed from $L_2$ norm residual minimization." pith.science (2026). https://pith.science/paper/TV2GD2UU

@misc{pith2026241115565,
  author       = {Pith},
  title        = {Pith review of: Stabilization of isogeometric finite element method with optimal test functions computed from $L_2$ norm residual minimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TV2GD2UU}},
  note         = {Machine review of arXiv:2411.15565}
}
abstract

We compare several stabilization methods in the context of isogeometric analysis and B-spline basis functions, using an advection-dominated advection\revision{-}diffusion as a model problem. We derive (1) the least-squares finite element method formulation using the framework of Petrov-Galerkin method with optimal test functions in the $L_2$ norm, which guarantee automatic preservation of the \emph{inf-sup} condition of the continuous formulation. We also combine it with the standard Galerkin method to recover (2) the Galerkin/least-squares formulation, and derive coercivity constant bounds valid for B-spline basis functions. The resulting stabilization method are compared with the least-squares and (3) the Streamline-Upwind Petrov-Galerkin (SUPG)method using again the Eriksson-Johnson model problem. The results indicate that least-squares (equivalent to Petrov-Galerkin with $L_2$-optimal test functions) outperforms the other stabilization methods for small P\'eclet numbers, while strongly advection-dominated problems are better handled with SUPG or Galerkin/least-squares.

Figures

Figures reproduced from arXiv: 2411.15565 by the authors.

Figure 1
Figure 1. The exact solutions for the first problem for ϵ ∈ {0.1, 0.01, 0.003} [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The exact solutions for the second problem for ϵ ∈ {0.01, 0.001, 0.0001} The problem is driven by the inflow Dirichlet boundary condition and develops a boundary layer of width ϵ at the outflow x = 1. The exact solution is given by u(x1, x2) = exp(r1(x1 − 1)) − exp(r2(x1 − 1)) exp(−r1) − exp(−r2) sin(πx2) (3) where r1 = 1+√ 1+4ϵ 2π2 2ϵ , r2 = 1− √ 1+4ϵ 2π2 2ϵ , see [2]. The exact solutions for ϵ ∈ {0.01, 0.001, 0.00… view at source ↗
Figure 3
Figure 3. B-spline basis functions for non-uniform grid for the Eriksson￾Johnsons problem [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: The first model problem, for ϵ ∈ {0.1, 0.01, 0.003}. The solution of the Bubnov-Galerkin problem on a manually refined grid. for the order p defined as the number of repetitions of the first knot x[1] minus one, assuming that the subsequent knots inserted into the deno…
Figure 5
Figure 5. Figure 5: The second model problem, Eriksson-Johnson for ϵ ∈ {0.01, 0.001, 0.0001}. The solution of the Bubnov-Galerkin problem on a man￾ually refined grid. ϵ 100 ∥u−uexact∥L2 ∥uexact∥L2 100 ∥u−uexact∥H1 ∥uexact∥H1 ϵ = 0.1 0.49 2.60 ϵ = 0.01 0.12 2.30 ϵ = 0.003 0.07 2.32 [PITH_…
Figure 6
Figure 6. Figure 6: B-spline basis functions for uniform grid for the Eriksson-Johnson problem [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: The first model problem for ϵ ∈ {0.1, 0.001, 0.0003}. The solution of the Bubnov-Galerkin problem on a uniform grid. knot y = [0 0 0 1 2 3 4 4 4]; points y = [0 0.25 0.5 0.75 1]; for the second problem. They define the B-splines on the uniform mesh illustrated in [PIT…
Figure 8
Figure 8. Figure 8: The second model problem, Eriksson-Johnson for ϵ ∈ {0.01, 0.001, 0.0001}. The solution of the Bubnov-Galerkin problem on a uni￾form grid. ϵ 100 ∥u−uexact∥L2 ∥uexact∥L2 100 ∥u−uexact∥H1 ∥uexact∥H1 ϵ = 0.1 0.60 4.69 ϵ = 0.01 46.11 60.31 ϵ = 0.003 87.17 189.62 [PITH_FULL…
Figure 9
Figure 9. Figure 9: The first model problem for ϵ ∈ {0.1, 0.01, 0.003}. The solution of the Petrov-Galerkin with optimal test functions in L2 setup [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Eriksson-Johnson for ϵ ∈ {0.01, 0.001, 0.0001}. The solution of the Petrov-Galerkin with optimal test functions in L2 setup. 2.3.3 Petrov-Galerkin method with optimal test functions in L2 setup on a uniform grid Next, we apply the Petrov-Galerkin formulation with opti…
Figure 11
Figure 11. Figure 11: The first model problem for ϵ = {0.1, 0.01, 0.003}. The solution with the modified stabilization on a uniform mesh of ten elements [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: The second model problem, Eriksson-Johnson for ϵ ∈ {0.01, 0.001, 0.0001}. The solution with the modified stabilization on a uni￾form mesh of ten elements. problem or ϵ = 0.0001 for the second model problem, the solution has a good shape on the mesh with only ten unifo…
Figure 13
Figure 13. Figure 13: The first model problem for ϵ ∈ {0.1, 0.001, 0.003}. The solution with the SUPG stabilization on a uniform mesh of ten elements [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]
Figure 14
Figure 14. Figure 14: The second model problem, Eriksson-Johnson for ϵ ∈ {0.01, 0.001, 0.0001}. The solution with the SUPG stabilization on a uniform mesh of ten elements. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

  1. [1]

    John Austin Cottrell, Thomas J. R. Hughes, Yuri, Bazilevs, Isogeometric analysis: toward integration of Computer Aided Design and Finite Element Analysis, John Wiley & Sons, 2009

    work page 2009

  2. [2]

    Eriksson, C

    K. Eriksson, C. Johnson, Adaptive finite element methods for parabolic problems I: A linear model problem, SIAM Journal on Numerical Analysis, 28 (1991), 43–77

    work page 1991

  3. [3]

    Stefan Takacs, Lecture notes on Numerical Analysis of Isogeometric Methods, Institute of Numerical Mathematics (NuMa), Johannes Ke- pler University Linz https://numa.jku.at/media/filer_public/3c/2a/ 3c2a6222-fd4d-4405-b866-e3757698fe8c/lecture_notes_iga.pdf

    work page

  4. [4]

    Hughes, Leopoldo P

    Thomas J.R. Hughes, Leopoldo P. Franca, Michel Mallet. A new finite element formulation for computational fluid dynamics: VI. Convergence analysis of the generalized SUPG formulation for linear time-dependent multidimensional advective-diffusive systems, Computer Methods in Ap- plied Mechanics and Engineering 63(1) (1987), 97–112

    work page 1987

  5. [5]

    Franca, Sergio L

    Leopoldo P. Franca, Sergio L. Frey 1, Thomas J.R. Hughes. Stabilized finite element methods: I. Application to the advective-diffusive model, Computer Methods in Applied Mechanics and Engineering 95(2) (1992), 253-276

    work page 1992

  6. [6]

    Evans, A minimum-residual finite element method for the convection-diffusion equation

    Jesse Chan, John A. Evans, A minimum-residual finite element method for the convection-diffusion equation. Technical Reports, The University of Texas at Austin (2013)

    work page 2013

  7. [7]

    Evans, and Weifeng Qiu, A dual Petrov–Galerkin finite element method for the convection–diffusion equation, Computers & Mathematics with Applications 68(11) (2014) 1513–1529

    Jesse Chan, John A. Evans, and Weifeng Qiu, A dual Petrov–Galerkin finite element method for the convection–diffusion equation, Computers & Mathematics with Applications 68(11) (2014) 1513–1529. 21

    work page 2014

  8. [8]

    Minimal residual space-time discretiza- tions of parabolic equations: Asymmetric spatial operators, Computers & Mathematics with Applications 101 (2021) 107–118

    Rob Stevenson, Jan Westerdiep. Minimal residual space-time discretiza- tions of parabolic equations: Asymmetric spatial operators, Computers & Mathematics with Applications 101 (2021) 107–118

    work page 2021

Show all 16 references
  1. [9]

    A class of discontinuous Petrov- Galerkin methods

    Leszek Demkowicz, Jay Gopalakrishnan. A class of discontinuous Petrov- Galerkin methods. Part I: The transport equation, Computer Methods in Applied Mechanics and Engineering, 199 (2010) 1558-1572

    2010

  2. [10]

    Pavel Bochev, Max Gunzburger (2009) Least-squares finite element meth- ods (Vol. 166). Springer Science & Business Media

    2009

  3. [11]

    Tomasz S lu˙ zalec, Mateusz Dobija, Anna Paszy´ nska, Ignacio Muga, Marcin Lo´ s and Maciej Paszy´ nski, Automatic stabilization of finite-element simula- tions using neural networks and hierarchical matrices, Computer Methods in Applied Mechanics and Engineering 411 (2023) 116073

    2023

  4. [12]

    Victor Manuel Calo, Marcin Lo´ s, Quanling Deng, Ignacio Muga, Maciej Paszy´ nski, Isogeometric Residual Minimization Method (iGRM) with di- rection splitting preconditioner for stationary advection-dominated diffu- sion problems, Computer Methods in Applied Mechanics and Engi...

    2021

  5. [13]

    Chennakesava Kadapa, Wulf G. Dettmer, Djordje Peri´ c, NURBS based least-squares finite element methods for fluid and solid mechanics, Interna- tional Journal for Numerical Methods in Engineering, 101(7) (2015) 521- 539

    2015

  6. [14]

    Boundary and Interior Layers, Computational and Asymptotic Methods BAIL 2018

    Calo, Victor M., Albert Romkes, and Eirik Valseth, Automatic varia- tionally stable analysis for FE computations: an introduction. Boundary and Interior Layers, Computational and Asymptotic Methods BAIL 2018. Springer International Publishing, (2020)

    2020

  7. [15]

    J., Franca, L

    Hughes, T. J., Franca, L. P., and Hulbert, G. M. A new finite element formulation for computational fluid dynamics: VIII. The Galerkin/least- squares method for advective-diffusive equations. Computer methods in applied mechanics and engineering, (1989) 73(2), 173-189

    1989

  8. [16]

    Error bounds for finite element method

    Babuˇ ska, I. Error bounds for finite element method. Numerische Mathe- matik (1971) 16, 322—333. 22

    1971

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.