betaFeedback

WorksMathematics

Signed QoI Gaps for Goal-Oriented Error Estimation: Reliability and Sharpness from the Same First-Order System Residuals

Abstract

There are two common ways to estimate the error in a quantity of interest (QoI). The product of primal and adjoint error estimators gives a reliable but pessimistic bound. The dual-weighted residual method gives a sharp signed estimate from a more accurate approximation of the adjoint solution. We write the primal and adjoint problems as first-order systems and obtain both a reliable bound and a sharp estimate from the same residuals. The first paper of this series (arXiv:2607.23850) uses two consistent evaluators of a quantity of interest, one built from the potentials and one in which the gradients are expressed through the fluxes. We show that they differ exactly by the A−1A^{-1} inner product of the primal and adjoint constitutive residuals, and we call this difference the signed QoI gap. The identity holds for arbitrary primal and adjoint approximations. In our numerical experiments, the signed QoI gap is of the size of the larger of the errors of the two evaluators, while the product of the estimators exceeds this error by factors of up to several hundred. We therefore use the least-squares estimators to refine the mesh and the signed QoI gap to decide when to stop. We also introduce a third evaluator, built from the fluxes alone, which restores the reaction term. The resulting reaction-restored QoI gap estimates the error when reaction dominates. With the non-intrusive least-squares flux recovery, we can apply the method to any discretization that produces a conforming potential, including finite difference and finite volume schemes.

Introduction

In many applications the quantity of interest (QoI) is a scalar functional of the solution of a partial differential equation, such as a drag coefficient, an average over a subdomain, or a flux through an interface. We want to know how large the error of the computed value is. In the standard Galerkin setting, with the Galerkin approximation uh∈Xhu_h \in X_h, γ=m(u)\gamma= m(u) and the adjoint solution zz satisfying a(s,z)=m(s)a(s,z)=m(s) for all admissible ss,

γ−m(uh)=a(u−uh,z−ψh)∀ψh∈Xh.\gamma-m(u_h)=a(u-u_h,z-\psi_h)\qquad\forall\psi_h\in X_h.

The dual-weighted residual method [4, 2] replaces zz by a more accurate approximation of the adjoint solution, computed in an enriched space or reconstructed from the Galerkin approximation zh∈Xhz_h\in X_h of zz. The resulting signed estimate is accurate in practice but not guaranteed, and its elementwise terms serve as refinement indicators. The other common approach takes ψh=zh\psi_h=z_h and bounds the two factors by estimators of the primal and the adjoint error [33, 29], ∣γ−m(uh)∣≤Cηpηd|\gamma-m(u_h)|\leq C\eta_p\eta_d. This bound is reliable and is the basis of the convergence theory of goal-oriented adaptive methods [28, 3, 18, 20], but it can be very pessimistic. In this paper we obtain both a reliable bound and a signed estimate from the same first-order system residuals.

The first paper of this series [35] considers the general second-order elliptic equation with diffusion coefficient AA, convection coefficient bb and reaction coefficient cc, which is also our model problem. That paper writes the primal problem and a physically meaningful adjoint problem as first-order systems for a flux and a potential. The norms of their residuals, the least-squares estimators ηpLS\eta_p^{\mathrm{LS}} and ηdLS\eta_d^{\mathrm{LS}}, are equivalent to the primal and adjoint errors for arbitrary approximations [10, 11, 6, 7, 37, 14]. The same paper uses two consistent evaluators of γ\gamma. The first, which we call the potential evaluator, is the corrected functional of Giles and Süli [19], γ0(w,v):=m(w)+ℓ(v)−a(w,v)\gamma_0(w,v) := m(w) + \ell(v) - a(w,v), where ℓ\ell is the right-hand side of the primal problem. It satisfies γ−γ0(w,v)=a(u−w,z−v)\gamma- \gamma_0(w,v) = a(u-w,z-v) for all admissible ww and vv, so we do not need Galerkin orthogonality. The second, γ1\gamma_1, also uses the flux approximations that a least-squares method computes along with the potentials. It expresses the gradients in γ0\gamma_0 through these fluxes and adds two residual corrections. Another way to combine least squares with a quantity of interest is to build the quantity into the least-squares functional itself [17].

The product of the two estimators gives a reliable upper bound for the errors of both evaluators, and its elementwise contributions [35] are good refinement indicators. In our experiments, however, the product exceeds the larger of these errors by factors of up to several hundred. If we stop when the product falls below a tolerance, we refine far more than needed. This overestimate does not come from the estimators. For the Poisson model problem each of them is asymptotically exact at the least-squares solution [16], so there the whole overestimate comes from bounding the error a(u−w,z−v)a(u-w,z-v) of γ0(w,v)\gamma_0(w,v) by the product of the two error norms. To decide when to stop, we need to know how large the error actually is.

For this we compare two approximations of the same quantity. We compute γ\gamma twice from the same primal and adjoint approximations (τ,w)(\tau,w) and (ρ,v)(\rho,v), once as γ0(w,v)\gamma_0(w,v) from the potentials and once as γ1(w,τ,v,ρ)\gamma_1(w,\tau,v,\rho) with the gradients expressed through the fluxes. We do not compute a second, more accurate approximation. We show that the difference of the two values, which we call the signed QoI gap, is exactly the A−1A^{-1} inner product of the primal and adjoint constitutive residuals,

γ1(w,τ,v,ρ)−γ0(w,v)=(A−1Rpcst(τ,w),Rdcst(ρ,v)),\gamma_1(w,\tau,v,\rho)-\gamma_0(w,v) = \left(A^{-1}R_p^{\mathrm{cst}}(\tau,w),R_d^{\mathrm{cst}}(\rho,v)\right),

with Rpcst(τ,w)=f2+τ+A∇wR_p^{\mathrm{cst}}(\tau,w) = f_2 + \tau+ A\nabla w and Rdcst(ρ,v)=g2+ρ+A∇v+bvR_d^{\mathrm{cst}}(\rho,v) = g_2 + \rho+ A\nabla v + bv, where f2f_2 and g2g_2 are the vector-valued data of the primal problem and of the quantity of interest. The identity is algebraic and holds for arbitrary primal and adjoint approximations. When γ\gamma lies between the two values, the QoI gap is between one and two times the larger of the two errors. Otherwise the QoI gap is smaller than the larger error, and much smaller only when the two values have nearly the same error.

The dual-weighted residual estimate also computes γ\gamma twice, but it evaluates γ0\gamma_0 the second time at a more accurate adjoint approximation. It is a good estimate when the second value is much more accurate than the first. The signed QoI gap is a generalized dual-weighted residual estimate with flux-based weights. We do not assume that either of its two values is the more accurate, and we need no more accurate approximation. The fluxes need not be more accurate than the gradients of the potentials either. The signed QoI gap and the product bound come from the same residuals. For the QoI gap we take the inner product of the primal and adjoint constitutive residuals, and for the bound we take the product of the norms of the primal and adjoint residuals.

Since −A−1Rpcst(τ,w)=∇(u−w)+A−1(σ−τ)-A^{-1}R_p^{\mathrm{cst}}(\tau,w) = \nabla(u-w) + A^{-1}(\sigma-\tau) and −Rdcst(ρ,v)=A∇(z−v)+b(z−v)+(r−ρ)-R_d^{\mathrm{cst}}(\rho,v) = A\nabla(z-v) + b(z-v) + (r-\rho), the QoI gap contains the error a(u−w,z−v)a(u-w,z-v) of γ0(w,v)\gamma_0(w,v) without its reaction term, plus terms in the flux errors. Where reaction dominates, the reaction term can dominate the errors of both evaluators, and the signed QoI gap does not contain it. For this case we introduce the reaction-restored evaluator γ2(τ,ρ)\gamma_2(\tau,\rho), built from the fluxes alone. The reaction-restored QoI gap γ2(τ,ρ)−γ0(w,v)\gamma_2(\tau,\rho)-\gamma_0(w,v) then estimates the error of γ0(w,v)\gamma_0(w,v). We refine the mesh with the least-squares estimators and stop when the magnitude of the signed QoI gap, or of the reaction-restored QoI gap where reaction dominates, is below a tolerance. In our experiments the QoI gaps follow the errors closely, even where the product bound is very pessimistic.

The construction does not depend on the discretization. We can take the potentials from any method that produces conforming potentials, and the fluxes from least-squares flux recovery [14, 24] or from a mixed method. An augmented mixed method [27, 25] gives both, and so does a least-squares finite element method. Our experiments include potentials from finite difference and finite volume schemes, equilibrated mixed fluxes for Kellogg's problem, a comparison with a dual-weighted residual computation, and computations with reaction coefficients from c=1c=1 to c=108c=10^8.

Section 2 recalls the setting and the evaluators of [35]. Section 3 derives the QoI gap identity, shows what the signed QoI gap estimates, and introduces γ2\gamma_2 and the reaction-restored QoI gap. Section 4 relates the QoI gaps to the product bound, the dual-weighted residual method, flux recovery and, for symmetric problems, the primal–dual bounds of [36]. Section 5 describes how we compute the fluxes and how we refine the meshes, and Section 6 reports the numerical experiments.

Primal and adjoint first-order systems and two evaluators

The primal problem, the quantity of interest and the adjoint problem

Let Ω⊂Rd\Omega\subset\mathbb{R}^{d}, d=2,3d=2,3, be a bounded Lipschitz polyhedral domain, whose boundary is split into relatively open, disjoint parts ΓD\Gamma_{D} and ΓN\Gamma_{N} with ∂Ω=Γ‾D∪Γ‾N\partial\Omega=\overline{\Gamma}_{D}\cup\overline{\Gamma}_{N}, and let n\mathbf{n} be the outward unit normal. The coefficients are A∈L∞(Ω)d×dA\in L^{\infty}(\Omega)^{d\times d}, symmetric and uniformly positive definite, b∈L∞(Ω)d\mathbf{b}\in L^{\infty}(\Omega)^{d} and c∈L∞(Ω)c\in L^{\infty}(\Omega). The data of the primal problem are f1∈L2(Ω)f_{1}\in L^{2}(\Omega), f2∈L2(Ω)d\mathbf{f}_{2}\in L^{2}(\Omega)^{d}, ϕp∈H1/2(ΓD)\phi_{p}\in H^{1/2}(\Gamma_{D}) and ψp∈L2(ΓN)\psi_{p}\in L^{2}(\Gamma_{N}), and those of the quantity of interest are g1∈L2(Ω)g_{1}\in L^{2}(\Omega), g2∈L2(Ω)d\mathbf{g}_{2}\in L^{2}(\Omega)^{d}, ϕd∈H1/2(ΓD)\phi_{d}\in H^{1/2}(\Gamma_{D}) and ψd∈L2(ΓN)\psi_{d}\in L^{2}(\Gamma_{N}). We write (⋅,⋅)(\cdot,\cdot) and ∥⋅∥0\|\cdot\|_{0} for the inner product and norm of L2(Ω)L^{2}(\Omega), and ⟨⋅,⋅⟩Γ\langle\cdot,\cdot\rangle_{\Gamma} for the L2(Γ)L^{2}(\Gamma) inner product on Γ⊂∂Ω\Gamma\subset\partial\Omega, or the duality pairing of H−1/2H^{-1/2} and H1/2H^{1/2} there. The spaces are

References

  1. [1]M. Ainsworth and T. Vejchodský, Robust error bounds for finite element approximation of reaction–diffusion problems with non-constant reaction coefficient in arbitrary space dimension, Comput. Methods Appl. Mech. Engrg., 281 (2014), pp. 184–199.DOI
  2. [2]W. Bangerth and R. Rannacher, Adaptive finite element methods for differential equations, Birkhäuser, Basel, 2003.DOI
  3. [3]R. Becker, E. Estecahandy, and D. Trujillo, Weighted marking for goal-oriented adaptive finite element methods, SIAM J. Numer. Anal., 49 (2011), pp. 2451–2469.DOI
  4. [4]R. Becker and R. Rannacher, An optimal control approach to a posteriori error estimation in finite element methods, Acta Numerica, 10 (2001), pp. 1–102.
  5. [5]M. Berndt, T. A. Manteuffel, and S. F. McCormick, Local error estimates and adaptive refinement for first-order system least squares (FOSLS), Electron. Trans. Numer. Anal., 6 (1997), pp. 35–43.
  6. [6]P. B. Bochev and M. D. Gunzburger, Finite element methods of least-squares type, SIAM Rev., 40 (1998), pp. 789–837.DOI
  7. [7]—, Least-Squares Finite Element Methods, Applied Mathematical Sciences, 166, Springer, 2009.
  8. [8]D. Cai, Z. Cai, and S. Zhang, Robust equilibrated error estimator for diffusion problems: mixed finite elements in two dimensions, J. Sci. Comput., 83 (2020), 22.DOI
  9. [9]Z. Cai, R. Falgout, and S. Zhang, Div first-order system LL* (FOSLL*) for second-order elliptic partial differential equations, SIAM J. Numer. Anal., 53 (2015), pp. 405–420.arxiv.org/abs/1407.4558
  10. [10]Z. Cai, R. D. Lazarov, T. A. Manteuffel, and S. F. McCormick, First-order system least-squares for second order partial differential equations: Part I, SIAM J. Numer. Anal., 31 (1994), pp. 1785–1799.DOI
  11. [11]Z. Cai, T. Manteuffel, and S. F. McCormick, First-order system least squares for second-order partial differential equations: Part II, SIAM J. Numer. Anal., 34 (1997), pp. 425–454.DOI
  12. [12]Z. Cai and S. Zhang, Recovery-based error estimator for interface problems: conforming linear elements, SIAM J. Numer. Anal., 47 (2009), pp. 2132–2156.DOI
  13. [13]Z. Cai and S. Zhang, Recovery-based error estimators for interface problems: mixed and nonconforming finite elements, SIAM J. Numer. Anal., 48 (2010), pp. 30–52.DOI
  14. [14]Z. Cai and S. Zhang, Flux recovery and a posteriori error estimators: conforming elements for scalar elliptic equations, SIAM J. Numer. Anal., 48 (2010), pp. 578–602.DOI
  15. [15]Z. Cai and S. Zhang, Robust equilibrated residual error estimator for diffusion problems: conforming elements, SIAM J. Numer. Anal., 50 (2012), pp. 151–170.DOI
  16. [16]C. Carstensen and J. Storn, Asymptotic exactness of the least-squares finite element residual, SIAM J. Numer. Anal., 56 (2018), pp. 2008–2028.DOI
  17. [17]J. H. Chaudhry, E. C. Cyr, K. Liu, T. A. Manteuffel, L. N. Olson, and L. Tang, Enhancing least-squares finite element methods through a quantity-of-interest, SIAM J. Numer. Anal., 52 (2014), pp. 3085–3105.DOI
  18. [18]M. Feischl, D. Praetorius, and K. G. van der Zee, An abstract analysis of optimal goal-oriented adaptivity, SIAM J. Numer. Anal., 54 (2016), pp. 1423–1448.DOI
  19. [19]M. B. Giles and E. Süli, Adjoint methods for PDEs: a posteriori error analysis and postprocessing by duality, Acta Numerica, 11 (2002), pp. 145–236.
  20. [20]M. Inneberger and D. Praetorius, Instance-optimal goal-oriented adaptivity, Comput. Methods Appl. Math., 21 (2021), pp. 109–126.
  21. [21]D. C. Jespersen, A least squares decomposition method for solving elliptic equations, Math. Comp., 31 (1977), pp. 873–880.DOI
  22. [22]B.-N. Jiang, The Least-Squares Finite Element Method: Theory and Applications in Computational Fluid Dynamics and Electromagnetics, Springer, Berlin, 1998.DOI
  23. [23]R. B. Kellogg, On the Poisson equation with intersecting interfaces, Applicable Anal., 4 (1974), pp. 101–129.DOI
  24. [24]Z. Li and S. Zhang, Non-intrusive least-squares functional a posteriori error estimator: Linear and nonlinear problems with plain convergence, Comput. Math. Appl., 191 (2025), pp. 275–295.arxiv.org/abs/2407.13253
  25. [25]Y. Liang and S. Zhang, Least-squares versus partial least-squares finite element methods: Robust a priori and a posteriori error estimates of augmented mixed finite element methods, Comput. Math. Appl., 184 (2025), pp. 45–64.arxiv.org/abs/2312.15459
  26. [26]Q. Liu and S. Zhang, Adaptive flux-only least-squares finite element methods for linear transport equations, J. Sci. Comput., 84 (2020), 26.arxiv.org/abs/1907.06991
  27. [27]A. Masud and T. J. R. Hughes, A stabilized mixed finite element method for Darcy flow, Comput. Methods Appl. Mech. Engrg., 191 (2002), pp. 4341–4370.
  28. [28]M. S. Mommer and R. Stevenson, A goal-oriented adaptive finite element method with convergence rates, SIAM J. Numer. Anal., 47 (2009), pp. 861–886.DOI
  29. [29]J. T. Oden and S. Prudhomme, Goal-oriented error estimation and adaptivity for the finite element method, Comput. Math. Appl., 41 (2001), pp. 735–756.DOI
  30. [30]M. Paraschivoiu, J. Peraire, and A. T. Patera, A posteriori finite element bounds for linear-functional outputs of elliptic partial differential equations, Comput. Methods Appl. Mech. Engrg., 150 (1997), pp. 289–312.DOI
  31. [31]N. Parés, N. C. Nguyen, P. Díez, and J. Peraire, A posteriori goal-oriented bounds for the Poisson problem using potential and equilibrated flux reconstructions: application to the hybridizable discontinuous Galerkin method, Comput. Methods Appl. Mech. Engrg., 386 (2021), 114088.DOI
  32. [32]A. I. Pehlivanov, G. F. Carey, and R. D. Lazarov, Least-squares mixed finite elements for second-order elliptic problems, SIAM J. Numer. Anal., 31 (1994), pp. 1368–1377.DOI
  33. [33]S. Prudhomme and J. T. Oden, On goal-oriented error estimation for elliptic problems: application to the control of pointwise error, Comput. Methods Appl. Mech. Engrg., 176 (1999), pp. 313–331.DOI
  34. [34]A. M. Sauer-Budge, J. Bonet, A. Huerta, and J. Peraire, Computing bounds for linear functionals of exact weak solutions to Poisson’s equation, SIAM J. Numer. Anal., 42 (2004), pp. 1610–1630.
  35. [35]Y. Wu and S. Zhang, Goal-oriented error estimation for least-squares finite element methods via physically meaningful adjoint PDEs, arXiv:2607.23850, 2026.arxiv.org/abs/2607.23850
  36. [36]S. Zhang, Primal-dual reduced basis methods for convex minimization variational problems: robust true solution a posteriori error certification and adaptive greedy algorithms, SIAM J. Sci. Comput., 42 (2020), pp. A3638–A3676.arxiv.org/abs/1810.04073
  37. [37]S. Zhang, A simple proof of coerciveness of first-order system least-squares methods for general second-order elliptic PDEs, Comput. Math. Appl., 130 (2023), pp. 98–104.DOI
  38. [38]S. Zhang, Why the Kellogg mesh is radial: a mathematical explanation of a classical computational benchmark, arXiv:2608.22991, 2026.arxiv.org/abs/2608.22991
  39. [39]O. C. Zienkiewicz and J. Z. Zhu, A simple error estimator and adaptive procedure for practical engineering analysis, Internat. J. Numer. Methods Engrg., 24 (1987), pp. 337–357.DOI
  40. [40]O. C. Zienkiewicz and J. Z. Zhu, The superconvergent patch recovery and a posteriori error estimates. Part 1: The recovery technique, Internat. J. Numer. Methods Engrg., 33 (1992), pp. 1331–1364.DOI

Paper details

vlnc:W_wjrkhhsz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents