A posteriori error estimates of finite element methods by preconditioning

02/16/2020
by   Yuwen Li, et al.
0

We present a framework that relates preconditioning with a posteriori error estimates in finite element methods. In particular, we use standard tools in subspace correction methods to obtain reliable and efficient error estimators. As a simple example, we recover the classical residual error estimators for the second order elliptic equations.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

11/20/2019

Residual-based a posteriori error estimates of mixed methods in Biot's consolidation model

We present residual-based a posteriori error estimates of mixed finite e...
10/14/2020

Nodal auxiliary a posteriori error estimates

We introduce and explain key relations between a posteriori error estima...
09/14/2020

A Posteriori Error Estimates for Elliptic Eigenvalue Problems Using Auxiliary Subspace Techniques

We propose an a posteriori error estimator for high-order p- or hp-finit...
08/09/2019

A posteriori error estimates for the mortar staggered DG method

Two residual-type error estimators for the mortar staggered discontinuou...
03/06/2020

A posteriori Error Estimation for the Spectral Deferred Correction Method

The spectral deferred correction method is a variant of the deferred cor...
08/12/2019

A new anisotropic mesh adaptation method based upon hierarchical a posteriori error estimates

A new anisotropic mesh adaptation strategy for finite element solution o...
11/26/2021

A-posteriori error estimation and adaptivity for multiple-network poroelasticity

The multiple-network poroelasticity (MPET) equations describe deformatio...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1. Introduction

Adaptive finite element methods (AFEMs) have been an active research area since the pioneering work [2]. In contrast to finite elements based on quasi-uniform meshes, AFEMs produce a sequence of locally refined grids that is able to resolve the singularity arising from irregular data in the underlying boundary value problems. Readers are referred to e.g., [3, 31, 36] for a thorough introduction. Among the key concepts in AFEMs, a posteriori error estimates are the building block for comparing errors on different elements and marking elements with large errors for refinement. For details on various AFEM error estimation techniques we refer to works on: explicit residual estimators [36]; implicit estimators based on local problems [2, 4, 14, 29]; recovery-based estimators; [42, 13, 7, 8, 41]; hierarchical basis estimators [5, 6, 24, 23]; functional estimators [32]; and equilibrated estimators [1, 26, 10, 19].

On the other hand, parallel with the development of AFEMs, there are also substantial research efforts in studying efficient preconditioning, which is a technique for approximating the inverse of a differential operator. Usually, such approximations are aimed at accelerating Krylov subspace iterative methods for solving linear systems resulting from discretized partial differential equations. Popular techniques used for preconditioning include e.g., multigrid 

[35, 11, 21, 22, 40, 39] and domain decomposition/subspace correction methods [18, 34, 37]. In practice, subspace correction methods provide an efficient way of reducing the condition number of a large-scale but finite-dimensional linear system. However, the analysis of uniform convergence rate of those methods often benefits from the general setting of infinite-dimensional Hilbert spaces (see, for example, [30, 20, 38]).

In this paper we present a general framework relating abstract operator preconditioning [37, 20, 38, 25, 27] to a posteriori error estimates. In particular, we shall show that such standard techniques for developing preconditioners also yield reliable and efficient error estimators. Here, for clarity of presentation, we focus on the symmetric and positive-definite problems although extensions to more general cases are definitely within reach. As a simple example, with this framework, we are able to recover the classical residual error estimators for elliptic equations in primal form.

The rest of this paper is organized as follows. In section 2, we set up the model variational problem and define the operator notation which is convenient when constructing preconditioners. In section 3, we develop the main theory on posteriori error estimates via preconditioning. Section 4 is devoted to the example of second order elliptic equation that illustrates the aforementioned abstract theory. Concluding remarks are found in Section 5.

2. Preliminaries

Let be a Hilbert space and denote the dual space of Let be a continuous bilinear form and . We consider the following variational problem: Find such that for all

(2.1)

Here is the duality pairing between and . Let denote the norm on and the dual norm of For simplicity, we assume that the bilinear form is symmetric and positive-definite (SPD). The continuity and positive-definiteness of imply

(2.2a)
(2.2b)

for all where are absolute constants. Such a bilinear form naturally defines a bounded isomophism for which we have

Hence (2.1) is equivalent to the operator equation

(2.3)

(2.2a) and (2.2b) imply that induces the inner product on For all the -norm on is defined as , which is equivalent to the -norm.

2.1. Approximation from a subspace

Let us consider a general case where we approximate the solution to (2.1) by restricting it to a subspace , namely: Find such that

(2.4)

Note that the subspace does not even have to be finite dimensional, although it usually is in applications. It follows from (2.2a), (2.2b) and the well-known Lax–Milgram theorem that (2.4) admits a unique solution.

For such a subspace , we consider the natural inclusion and its adjoint defined as

We introduce the operator which approximates on . In this way, the discrete problem (2.4) reads

3. A posteriori error estimates by preconditioning

A posteriori error estimates are of the form

where are absolute positive constants and is computed from . In AFEMs, is the sum of error indicators on all elements. The local error indicators can be used to compare errors on different elements and those elements with large errors will be refined. In this way, the errors estimated by are equidistributed over all elements in the mesh. The optimal computational complexity of AFEMs is often attributed to the aforementioned equidistribution of errors. Rigorous analysis of convergence and optimality of AFEMs can be found in e.g., [17, 28, 9, 33, 15].

3.1. Links with operator preconditioning

Let

Clearly, from our discusion above, it follows that constructing a posteriori error estimators is equivalent to estimating a norm of the error by computable bounds. We note, however, that a direct computation of the norm of will be, in general, impossible or too expensive, since one needs to compute the action of on . As we pointed out in the introduction, approximating such action has been also studied for several decades and is known as as preconditioning. Following this simple observation we now borrow some simple ideas from this field and apply them in constructing a posteriori error estimators.

First, we need a bounded isomorphism (the preconditioner) , whose particular form will be given later. For the time being we only assume that is bounded and SPD, i.e., is an inner product on . Let be a SPD operator, which we will refer to as “the smoother” and is such that its range approximates well the high frequency part of the range of , i.e., the result of the action provides a good approximation to the high frequency components of the error. Now, a simple choice for is

which is known as additive Schwarz preconditioner. Just to simplify the presentation, we will not consider the multiplicative preconditioner in this paper although following the abstract framework developed in [20, 38] similar results can also be obtained in the multiplicative case as well. Let be two positive absolute constants. We say that is a preconditioner for provided there exist constants and , such that

(3.1)

The inequality (3.1) is known as spectral equivalence, or norm equivalence, and is a common ingredient in the analysis of convergence of iterative methods for large-scale linear systems.

3.2. Estimating the residual

We now show that the norm (spectral) equivalence (3.1) naturally yields a two-sided estimate on . This is the central result in this paper.

Theorem 3.1.

Let (3.1) hold. Then we have the following two sided bound

Proof.

Since is SPD, we use the Cauchy–Schwarz inequality to obtain

(3.2)

The inequality (3.1) implies

(3.3)

Combining (3.2) and (3.3) yields

where we used in the first equality. The upper bound

can be shown in a similar fashion. In summary, we have

(3.4)

On the other hand, for any , (2.4) implies

i.e., Hence

(3.5)

Combining (3.4) and (3.5) completes the proof. ∎

Throughout the rest, will serve as a (nearly) computable a posteriori error estimator that is proved to be both an upper and lower bound of the error . In order to derive an error estimator within our framework, the key step is to suitably select the smoother such that the spectral equivalence (3.1) holds.

3.3. Additive Schwarz smoother

In this subsection, we construct a particular using the additive Schwarz method. For such a smoother, we present a lemma that serves as a criterion for verifying (3.1).

For , , let be subspaces providing a decomposition of , namely,

(3.6)

Let be the natural inclusion and denote its adjoint. We further set . Next, let be spectrally equivalent to . More precisely, for and , we assume that

(3.7)

where are positive absolute constants. The smoother (additive Schwarz method) is then defined to be

By the definition of , we obtain

The norm of can be estimated using the following lemma, which can be found in e.g., [20, 38, 34, 12].

Lemma 3.2.

We have the following identity

where the infimum is taken over and for

The proof that is a good preconditioner for is standard. We include it here for completeness and we follow the proof in [38].

Lemma 3.3.

For each , let

and . In addition, assume that for all , there exist with satisfying

(3.8)

Then (3.1) holds with constants ,

Proof.

For , assume the decomposition with , . Direct calculation shows that

(3.9)

The definition of and implies

Combining the previous estimate with (3.9) and (3.7) gives

(3.10)

Taking the infimum with respect to all decompositions and using Lemma 3.2, we obtain the upper bound

For the lower bound in (3.1), let be the decomposition that satisfies (3.8). It then follows from (3.7) and (3.8) that

Using the previous estimate and Lemma 3.2, we obtain

The proof is complete. ∎

4. Examples

In this section, we consider the typical example of a scalar elliptic equation. Let where is a Lipschitz polytope. For a given and , the bilinear and linear forms in equation (2.1) are:

We assume is uniformly elliptic, i.e.,

Hence (2.2a) and (2.2b) holds.

Let be a conforming and shape-regular simplicial partition of . Let denote the set of polynomials of degree at most . The subspace is

where is an integer.

Let denote the set of interior vertices in For each , let denote the continuous piecewise linear function that takes the value 1 at and 0 at other vertices. We denote for . Obviously we have

(4.1)
(4.2)

4.1. A posteriori error estimates for Lagrange elements

Now, let which is a subspace of by zero extension. The partition of unity (4.1) implies

We note that the framework also works for other local patches. For instance, can be chosen as the union of the two elements around each interior face.

For a fixed , the set defined in Lemma 3.3 translates into

In this case, is an absolute constant by the shape-regularity of . Throughout the rest, we adopt the notation provided with being a generic constant dependent only on and We say provided and .

We set and thus in (3.7). The corresponding smoother yields an error estimator. In order to show the reliability and efficiency, we need to verify (3.8) in Lemma 3.3.

Corollary 4.1.

We have the following estimate

Proof.

To verify (3.8), we take , where is a

-stable interpolation which also enjoys standard approximation properties:

(4.3)

A simple choice for is the Clément interpolation  [16]. We now set . Hence is a decomposition. It follows from (4.2) and (4.3) that

Hence (3.8) are verified. Finally, we conclude Corollary 4.1 from Theorem 3.1 and Lemma 3.3. ∎

For we have

Hence computing amounts to solving the variational problem:

(4.4)

Taking in (4.4) implies that

It then follows from the previous identity and Corollary 4.1 that

(4.5)

4.2. Computable error estimator

Unfortunately, is not available in practice because (4.4) is local but still not fully computable. To implement the estimator in Corollary 4.1, we consider the approximate problem: Find such that

(4.6)

where is a subspace of polynomials. Generally speaking, the approximate estimator is expected to be an accurate upper and lower bound of provided the degree of polynomials in is sufficiently high.

Taking in (4.6) and (4.4), we obtain

Combining the previous estimate with (4.5) provides the following computable lower bound for the error

To derive a computable upper bound for , let us first write the action of the residual on . We denote the set of all -dimensional faces in the triangulation by . Clearly, where denotes the set of all boundary faces and the interior faces. We further denote by and by , respectively. Note that does not include the edges on . For each , let

Further, for each , let be the two elements sharing , (resp. ) the outward unit normal to (resp. ), and

It then follows from (4.4) and integration by parts that

(4.7)

Given a Lipschitz domain , let denote the diameter of The shape-regularity of implies that and we will use these notions interchangeably. Now, let us introduce the computable quantity

which is the standard explicit residual error estimator. We take and use (4.7) and the Cauchy–Schwarz inequality to obtain that

Finally, combining the previous inequality with the trace inequality and the Poincaré inequality yields

Hence we obtain the following computable upper bound for the error

5. Concluding remarks

For SPD problems, we have shown how preconditioning can be used to derive a posteriori error estimates. Extensions of this abstract theoretical framework and its application to derive estimators for indefinite, nonconforming, and discontinuous Galerkin methods are ongoing. A close inspection of the arguments shows that not only preconditioning can give a unified way to derive a posteriori error estimators. This is a two-way street: the a posteriori error estimators may provide efficient smoothers for multilevel methods. For example, the operator we have introduced in our framework is a clear analogue of smoothing (relaxation) operator. We hope that some of the error indicators and estimators may give efficient smoothers in case of non-symmetric and or indefinite problems which are, in general, hard to precondition.

References

  • [1] Mark Ainsworth and J. Tinsley Oden. A posteriori error estimation in finite element analysis. Pure and Applied Mathematics (New York). Wiley-Interscience [John Wiley & Sons], New York, 2000.
  • [2] I. Babuška and W. C. Rheinboldt. Error estimates for adaptive finite element computations. SIAM J. Numer. Anal., 15(4):736–754, 1978.
  • [3] Ivo Babuška and Theofanis Strouboulis. The finite element method and its reliability. Numerical Mathematics and Scientific Computation. The Clarendon Press, Oxford University Press, New York, 2001.
  • [4] R. E. Bank and A. Weiser. Some a posteriori error estimators for elliptic partial differential equations. Math. Comp., 44(170):283–301, 1985.
  • [5] Randolph E. Bank. Hierarchical bases and the finite element method. In Acta numerica, 1996, volume 5 of Acta Numer., pages 1–43. Cambridge Univ. Press, Cambridge, 1996.
  • [6] Randolph E. Bank and R. Kent Smith. A posteriori error estimates based on hierarchical bases. SIAM J. Numer. Anal., 30(4):921–935, 1993.
  • [7] Randolph E. Bank and Jinchao Xu. Asymptotically exact a posteriori error estimators. I. Grids with superconvergence. SIAM J. Numer. Anal., 41(6):2294–2312, 2003.
  • [8] Randolph E. Bank and Jinchao Xu. Asymptotically exact a posteriori error estimators. II. General unstructured grids. SIAM J. Numer. Anal., 41(6):2313–2332, 2003.
  • [9] Peter Binev, Wolfgang Dahmen, and Ron DeVore. Adaptive finite element methods with convergence rates. Numer. Math., 97(2):219–268, 2004.
  • [10] Dietrich Braess and Joachim Schöberl. Equilibrated residual error estimator for edge elements. Math. Comp., 77(262):651–672, 2008.
  • [11] Achi Brandt. Multi-level adaptive technique (MLAT) for fast numerical solution to boundary value problems. In Proc. 3rd Internat. Conf. on Numerical Methods in Fluid Mechanics (Paris, 1972), volume 18 of Lecture Notes in Physics, pages 82–89, Berlin, New York, 1973. Springer-Verlag.
  • [12] Susanne C. Brenner and L. Ridgway Scott. The mathematical theory of finite element methods, volume 35 of Texts in Applied Mathematics, 15. Springer, New York, 3 edition, 2008.
  • [13] Carsten Carstensen and Sören Bartels. Each averaging technique yields reliable a posteriori error control in FEM on unstructured grids. I. Low order conforming, nonconforming, and mixed FEM. Math. Comp., 71(239):945–969, 2002.
  • [14] Carsten Carstensen and Stefan A. Funken. Fully reliable localized error control in the FEM. SIAM J. Sci. Comput., 21(4):1465–1484, 1999/00.
  • [15] J. Manuel Cascon, Christian Kreuzer, Ricardo H. Nochetto, and Kunibert G. Siebert. Quasi-optimal convergence rate for an adaptive finite element method. SIAM J. Numer. Anal., 46(5):2524–2550, 2008.
  • [16] Ph. Clément. Approximation by finite element functions using local regularization. Rev. Française Automat. Informat. Recherche Opérationnelle Sér., 9(R-2):77–84, 1975.
  • [17] Willy Dörfler. A convergent adaptive algorithm for Poisson’s equation. SIAM J. Numer. Anal., 33(3):1106–1124, 1996.
  • [18] Maksymilian Dryja and Olof B. Widlund. An additive variant of the Schwarz alternating method for the case of many subregions. Technical Report 339, also Ultracomputer Note 131, Department of Computer Science, Courant Institute, 1987.
  • [19] Alexandre Ern and Martin Vohralík. Polynomial-degree-robust a posteriori estimates in a unified setting for conforming, nonconforming, discontinuous galerkin, and mixed discretizations. SIAM J. Numer. Anal., 53(2):1058–1081, 2015.
  • [20] M. Griebel and P. Oswald. On the abstract theory of additive and multiplicative Schwarz algorithms. Numer. Math., 70(2):163–180, 1995.
  • [21] W. Hackbusch. A fast iterative method for solving Poisson’s equation in a general region. In Numerical treatment of differential equations (Proc. Conf., Math. Forschungsinst., Oberwolfach, 1976), pages 51–62. Lecture Notes in Math., Vol. 631. Springer, Berlin, 1978. Longer version: Ein Iteratives Verfahren zur Schnellen Auflösung Elliptischer Randwertprobleme, Math. Inst., Universität zu Köln, Report 76-12 (November 1976).
  • [22] Wolfgang Hackbusch. Multigrid methods and applications, volume 4 of Springer Series in Computational Mathematics. Springer-Verlag, Berlin, 1985.
  • [23] Harri Hakula, Michael Neilan, and Jeffrey S. Ovall. A posteriori estimates using auxiliary subspace techniques. J. Sci. Comput., 72(1):97–127, 2017.
  • [24] Hengguang Li and Jeffrey S. Ovall. A posteriori error estimation of hierarchical type for the Schrödinger operator with inverse square potential. Numer. Math., 128(4):707–740, 2014.
  • [25] D. Loghin and A. J. Wathen. Analysis of preconditioners for saddle-point problems. SIAM J. Sci. Comput., 25(6):2029–2049 (electronic), 2004.
  • [26] R. Luce and B. I. Wohlmuth. A local a posteriori error estimator based on equilibrated fluxes. SIAM J. Numer. Anal., 42(4):1394–1414, 2004.
  • [27] Kent-Andre Mardal and Ragnar Winther. Preconditioning discretizations of systems of partial differential equations. Numer. Linear Algebra Appl., 18(1):1–40, 2011.
  • [28] Pedro Morin, Ricardo H. Nochetto, and Kunibert G. Siebert. Data oscillation and convergence of adaptive FEM. SIAM J. Numer. Anal., 38(2):466–488, 2000.
  • [29] Pedro Morin, Ricardo H. Nochetto, and Kunibert G. Siebert. Local problems on stars: a posteriori error estimators, convergence, and performance. Math. Comp., 72(243):1067–1097, 2003.
  • [30] S. V. Nepomnyaschikh. Decomposition and fictitious domains methods for elliptic boundary value problems. In Fifth International Symposium on Domain Decomposition Methods for Partial Differential Equations (Norfolk, VA, 1991), pages 62–72. SIAM, Philadelphia, PA, 1992.
  • [31] Ricardo H. Nochetto, Kunibert G. Siebert, and Andreas Veeser. Theory of adaptive finite element methods: an introduction, pages 409–542. Springer, Berlin, 2009.
  • [32] Sergey Repin. A posteriori estimates for partial differential equations, volume 4 of Radon Series on Computational and Applied Mathematics. Walter de Gruyter GmbH & Co. KG, Berlin, 2008.
  • [33] Rob Stevenson. Optimality of a standard adaptive finite element method. Found. Comput. Math., 7(2):245–269, 2007.
  • [34] Andrea Toselli and Olof Widlund. Domain decomposition methods—algorithms and theory, volume 34 of Springer Series in Computational Mathematics. Springer-Verlag, Berlin, 2005.
  • [35] U. Trottenberg, C. W. Oosterlee, and A. Schüller. Multigrid. Academic Press, Inc., San Diego, CA, 2001. With contributions by A. Brandt, P. Oswald and K. Stüben.
  • [36] Rüdiger Verfürth. A posteriori error estimation techniques for finite element methods. Numerical Mathematics and Scientific Computation. Oxford University Press, Oxford, 2013.
  • [37] Jinchao Xu. Iterative methods by space decomposition and subspace correction. SIAM Rev., 34(4):581–613, 1992.
  • [38] Jinchao Xu and Ludmil Zikatanov. The method of alternating projections and the method of subspace corrections in Hilbert space. J. Amer. Math. Soc., 15(3):573–597, 2002.
  • [39] Jinchao Xu and Ludmil Zikatanov. Algebraic multigrid methods. Acta Numer., 26:591–721, 2017.
  • [40] Harry Yserentant. Old and new convergence proofs for multigrid methods. In Acta numerica, 1993, Acta Numer., pages 285–326. Cambridge Univ. Press, Cambridge, 1993.
  • [41] Zhimin Zhang and Ahmed Naga. A new finite element gradient recovery method: superconvergence property. SIAM J. Sci. Comput., 26(4):1192–1213, 2005.
  • [42] O. C. Zienkiewicz and J. Zhu. The superconvergent patch recovery and a posteriori error estimates. II. error estimates and adaptivity. Internat. J. Numer. Methods Engrg., 33(7):1365–1382, 1992.