A unified analysis of algebraic flux correction schemes for convection–diffusion equations
Abstract
Recent results on the numerical analysis of algebraic flux correction (AFC) finite element schemes for scalar convection–diffusion equations are reviewed and presented in a unified way. A general form of the method is presented using a link between AFC schemes and nonlinear edgebased diffusion schemes. Then, specific versions of the method, that is, different definitions for the flux limiters, are reviewed and their main results stated. Numerical studies compare the different versions of the scheme.
Keywords
Steadystate convectiondiffusion equation Algebraic flux correction Edge diffusion Discrete maximum principle Comparison of limitersMathematics Subject Classification
65N30 65N12 65N151 Introduction
Scalar convection–diffusion equations model the convective and molecular transport of a quantity like temperature or concentration. In applications, the convective transport is usually dominant, which is the case of interest in this paper.
A characteristic feature of solutions of (1) is the appearance of layers, i.e., of narrow regions where the solution has a large gradient. These regions are usually so narrow that the layers cannot be resolved by affordable grids. It is well known that standard discretizations cannot cope with this situation and they lead to meaningless numerical solutions that are globally polluted with huge spurious oscillations. The remedy consists in using stabilized discretizations. In the context of finite element methods, the proposal of the streamlineupwind Petrov–Galerkin (SUPG) method in [15, 22] was the first milestone in this direction. Solutions computed with this method usually have sharp layers at the correct position, but there are still nonnegligible spurious oscillations in a vicinity of layers. Since the publication of [15, 22] the development and analysis of stabilized discretizations for convectiondominated equations has been an active field of research.
In this research, one can distinguish two directions. The first one is the development of stabilized methods with a provable order of convergence in appropriate norms. Examples of this direction are the continuous interior penalty (CIP) method (see, e.g. [16]) and the local projection stabilization (LPS) method (see [12] for the first application of this method to a convectiondominated equation). The second direction consists in finding stabilized methods that compute solutions without spurious oscillations and still with sharp layers. The property of being free of spurious oscillations can be expressed mathematically with the satisfaction of the discrete maximum principle (DMP). Usually, the satisfaction of the DMP is proved by the sufficient condition that the matrix of a linear discretization^{1} is an Mmatrix. However, it is well known that, in the limit case \(\varepsilon = 0\), there is a barrier of the order of the local discretization error for linear discretizations with Mmatrices: these discretizations are at most of first order, e.g., see [42, Thm. 4.2.2].
Since the property of being free of spurious oscillations might be of utmost importance for a method to be applicable in practice, a significant amount of work has been devoted to the development of such methods. Due to the order barrier for linear discretizations, nonlinear discretizations became of interest. One further argument in favor of using nonlinear discretizations for a convectiondominated problem stems from the fact that most of the applications in which convection dominates are modeled by nonlinear partial differential equations. Then, the use of a nonlinear discretization does not constitute a significant overhead. Since the late 1980s, there have been a number of proposals to remove the spurious oscillations of the SUPG method by adding appropriate nonlinear terms. This class of methods is called spurious oscillations at layers diminishing (SOLD) methods, or shock capturing methods. A comprehensive review was carried out in the companion papers [25, 26], and the main conclusion of it was that none of the proposed SOLD methods reduced the spurious oscillations sufficiently well.
Algebraic stabilizations, socalled algebraic flux correction (AFC) schemes, became of interest to us as a result of numerical assessments of stabilized discretizations in [5, 25, 28, 29]. The main motivation for the design of AFC methods is the satisfaction of the DMP. In addition, they provide reasonably sharp approximations of the layers. In contrast to SOLD methods, which are based on variational formulations, the main idea of AFC schemes consists in modifying the algebraic system corresponding to a discrete problem, typically the Galerkin discretization, by means of solutiondependent flux corrections. Consequently, AFC schemes are nonlinear. The basic philosophy of flux correction schemes was formulated already in [14, 43]. Later, the idea was extended to the finite element context, e.g., in [4, 39]. In the last fifteen years, there has been an intensive development of these methods, e.g., see [33, 34, 35, 36, 37].
None of the above references deals with the mathematical analysis of the AFC methods. In fact, the first contributions to the numerical analysis of AFC schemes were presented only recently in [8, 9, 11]. The first paper [8] focuses on the solvability of the nonlinear scheme, while [9] presents the first error analysis of the AFC schemes. Interestingly, the paper [9] also presented negative results, in the sense that it was shown that unless some restrictions are imposed on the mesh, the numerical scheme may not converge. Finally, in the recent paper [11] the role of the linearity preservation was studied. This study is also complemented by the work [10], where a link between the AFC schemes and a nonlinear edgebased diffusion scheme is presented, and the linearity preservation of the scheme is also studied in detail. This latter reformulation offers the applicability of different tools than those used so far for the analysis of AFC schemes. In particular, it facilitated the a posteriori error analysis of the AFC method, presented in [2]. Thus, the present paper aims at providing a review of these works, and performing the analysis in a unified framework.
The rest of the manuscript is organized as follows. After having introduced AFC methods in Sect. 2, a rewriting as an edge diffusion scheme is presented. A unified analysis is given in Sect. 3, covering the existence of a solution, minimal conditions for the validity of the DMP, and finite element error estimates. Three definitions of limiters are provided in Sect. 4. Strategies for the solution of the nonlinear problems are discussed in Sect. 5. Numerical studies for different limiters used in AFC schemes and the edge diffusion scheme proposed in [10] are presented in Sect. 6. Finally, Sect. 7 states the most important open problems in the field of AFC schemes.
2 The model problem and a unified presentation of AFC schemes

\(\{\mathscr {T}_h\}_{h>0}\) denotes a family of shape regular simplicial triangulations of \(\overline{\varOmega }\).

For a given triangulation \(\mathscr {T}_h\), \(\mathscr {E}_h^{}\) denotes the set of its internal edges.

For every edge \(E\in \mathscr {E}_h^{}\), we denote by \(h_E\) the length of E and by \(\varvec{x}_{E,1}\), \(\varvec{x}_{E,2}\) the endpoints of E. Furthermore, for every \(E\in \mathscr {E}_h^{}\), we choose one unit tangent vector \(\varvec{t}_E^{}\). Its orientation is of no importance.

For every edge \(E\in \mathscr {E}_h^{}\), we define the neighborhood \(\omega _E^{}:=\cup \{T\in \mathscr {T}_h:T\cap E\ne \emptyset \}\).

For a given triangulation \(\mathscr {T}_h\), \(\{\varvec{x}_1^{},\ldots ,\varvec{x}_N^{}\}\) is the set of its nodes. We will assume that the nodes \(\varvec{x}_1^{},\ldots ,\varvec{x}_M^{}\) are the internal nodes, and \(\varvec{x}_{M+1}^{},\ldots ,\varvec{x}_N^{}\) are boundary nodes, i.e., the nodes where the Dirichlet boundary condition is imposed.
 For a node \(\varvec{x}_i^{}\), \(i=1,\ldots ,N\), we define$$\begin{aligned} \mathscr {E}_i^{}:=\{E\in \mathscr {E}_h^{}: \varvec{x}_i^{}\;\text {is an endpoint of}\; E\}. \end{aligned}$$

For a node \(\varvec{x}_i^{}\), \(i=1,\ldots ,N\), we define \(\varDelta _i:=\{T\in \mathscr {T}_h:\varvec{x}_i\in T\}\).
 For an interior node \(\varvec{x}_i^{}, i=1,\ldots ,M\), we define the index set of its neighbors$$\begin{aligned}&S_i:=\{j\in \{1,\dots ,N\}\setminus \{i\}\,:\,\, \varvec{x}_i\text { and }\varvec{x}_j\text { are endpoints of the same}\nonumber \\&\quad \text {internal edge}\; E\in \mathscr {E}_h^{}\}. \end{aligned}$$
 The finite element spaces used in this work are given by$$\begin{aligned} V_h:=\{v_h\in C^0(\overline{\varOmega }): v_h_T^{}\in \mathbb {P}_1(T)\;\;\forall \, T\in \mathscr {T}_h\},\quad V_{h,0}:=V_h\cap H^1_0(\varOmega ). \end{aligned}$$

These spaces have standard nodal basis functions denoted by \(\{\varphi _1^{},\ldots ,\varphi _N^{}\}\), uniquely determined by the conditions \(\varphi _i^{}(\varvec{x}_j)=\delta _{ij}^{}\) for all \(i,j=1,\ldots ,N\). We further notice that \(\text {supp}\,\varphi _i^{}=\varDelta _i\).
 The Lagrange interpolation operator \(i_h:C^0(\overline{\varOmega })\rightarrow V_h^{}\) is given byIn addition, we set$$\begin{aligned} i_h v=\sum _{i=1}^N v(\varvec{x}_i^{})\,\varphi _i. \end{aligned}$$$$\begin{aligned} i_h u_D=\sum _{i=M+1}^N u_D(\varvec{x}_i^{})\,\varphi _i_{\partial \varOmega }^{}. \end{aligned}$$
2.1 A variational formulation and a rewriting as an edge diffusion scheme
The solutiondependent limiters \(\beta _E\) are still assumed to satisfy \(\beta _E\in [0,1]\) and to assure the solvability of (14) (see the next section), we further make the following continuity assumption:
Assumption (A1)
For any \(E\in \mathscr {E}_h\), the function \(\beta _E(u_h)(\nabla u_h)_E^{}\cdot \varvec{t}_E\) is a continuous function of \(u_h\in V_h\).
It will be shown in Sect. 4 that the limiters defined in [9, 10, 11] satisfy Assumption (A1).
Remark 1
The fact that the restriction of the functions v and w to the internal edges is a linear function is what makes it possible to obtain the expression (16) for \(D_h\). This property also holds for unmapped \(\mathbb {Q}_1\) finite elements, and for mapped \(\mathbb {Q}_1\) finite elements on parallelepipeds, although in that case this methodology would lead to a completely different method, as the crossterms would not be included in the method. The implications of this remark are the topic of current investigations and are to be reported elsewhere. On a related note to the previous point, AFCrelated schemes using higher order elements combined with Bernstein basis functions have been developed recently in the work [38], but the full stability and error analysis of the methods is lacking.
3 General properties of the nonlinear scheme
In this section we present the main results associated to the nonlinear scheme (14). More precisely, we present results on its solvability, minimal conditions for the validity of the discrete maximum principle, and a first error estimate for the method. In the following section the conditions imposed herein will be checked for different definitions of the limiters \(\beta _E\).
3.1 Existence of solutions
Lemma 1
(Consequence of Brouwer’s fixedpoint theorem) Let X be a finitedimensional Hilbert space with inner product \((\cdot ,\cdot )_X\) and norm \(\Vert \cdot \Vert _X\). Let \(T:X\rightarrow X\) be a continuous mapping and \(K>0\) a real number such that \((Tx,x)_X>0\) for any \(x\in X\) with \(\Vert x\Vert _X=K\). Then there exists \(x\in X\) such that \(\Vert x\Vert _X < K\) and \(Tx=0\).
A proof of Lemma 1 can be found in [40, p. 164, Lemma 1.4]. Now, the existence of solutions for the nonlinear scheme (14) can be proved.
Theorem 1
(Existence of a solution of (14)) If Assumption (A1) holds, then there exists a solution \(u_h^{}\) of (14).
Proof
3.2 The discrete maximum principle
In this section we shall formulate general properties of the limiters \(\beta _E\) under which the AFC scheme (14) satisfies the local and global DMP. The local DMP will be formulated on the patches \(\varDelta _i\) defined in Sect. 2.
To prove the DMP, we make the following general assumption, which is a reformulation of an analogous assumption introduced in [32].
Assumption (A2)
Theorem 2
Proof
For proving (17), it suffices to consider the case \(A_i>0\). Let us assume that \(\max _{\varDelta _i}u_h>\max _{\partial \varDelta _i}u_h^+\). Then again \(\max _{\varDelta _i}u_h>\max _{\partial \varDelta _i}u_h\) and also \(u_i>0\). Like before, the sum in (24) is nonnegative and since \(A_i\,u_i>0\), the lefthand side of (24) is positive, which is again a contradiction proving the assertion.
The implications (18) and (20) follow in an analogous way. \(\square \)
Theorem 3
Proof
Now we want to prove that the relations (21), (23), (29), and (30) imply (25) and (27). If \(c=0\) in \(\varOmega \) and hence \(\sum _{j=1}^N\widetilde{a}_{ij}=0\) for \(i=1,\dots ,M\) (see (23)), then (21) still holds if one adds a constant to all components of the vector \((u_1,\dots ,u_N)\) so that one can assume that \(s>0\). If \(\sum _{j=1}^N\widetilde{a}_{ij}>0\), then \(s>0\) can be also assumed since otherwise (25) trivially holds.
The relations (26) and (28) can be proved analogously. \(\square \)
Remark 2
3.3 An a priori error estimate
Theorem 4
Proof
A simple estimate of the consistency error \(D_h^{}(u_h;i_h u,i_h u)^{\frac{1}{2}}\) is given in the following lemma.
Lemma 2
Proof
Lemma 2 shows that if \(d_E\) is defined by (8), then the convergence order of \(\Vert uu_h\Vert _h^{}\) is reduced to 1/2 in the convectiondominated case and no convergence follows in the diffusiondominated case. It was demonstrated in [9] that these results are sharp. On the other hand, the results of [10, 11] indicate that a better convergence behaviour in the diffusiondominated case may be expected if the AFC scheme is linearity preserving, i.e., if the stabilization originating from the AFC vanishes in regions where the approximate solution is a polynomial of degree 1. This property can be formulated in terms of the limiters \(\beta _E\) in the following way.
Assumption (A3)
The linearity preservation leads to an improved bound of the consistency error provided that the limiters satisfy the following Lipschitzcontinuity assumption.
Assumption (A4)
Lemma 3
Proof
4 Various definitions of the limiters
4.1 The Kuzmin limiter
In this section we review the results obtained when implementing the method with the definition of the limiters proposed in [34]. In that work, the algorithm, originally proposed by Zalesak in [43] was adapted to the steadystate case, and exploited further. We refer then to this limiter as the Kuzmin limiter. This limiter has been used in numerous works, for example [5, 9], where a detailed study of its performance for the convection–diffusion equation is carried out. The numbers \(d_E\) in the definition of \(D_h^{}\) are given by (8) in this case.
After having presented the Kuzmin limiter, we will show that it satisfies Assumption (A1) and, under an additional assumption on the matrix \(\mathbb {A}\), also Assumption (A2). Consequently, the nonlinear problem (14) possesses a solution and satisfies the discrete maximum principle. It was demonstrated in [11, Ex. 7.2] with the help of a numerical example that the AFC scheme with the Kuzmin limiter is not linearity preserving in general.
There is an obvious ambiguity in the definition of \(\beta _E\) if \(a_{ij}=a_{ji}\). This ambiguity does not influence the resulting method if \(\min \{a_{ij},a_{ji}\}\le 0\) since then \(d_E=0\) and the respective term with \(\beta _E\) does not occur in (15). To fulfill the condition \(\min \{a_{ij},a_{ji}\}\le 0\), which also assures the DMP (cf. Lemma 5), it may help to replace the matrix corresponding to the reaction term by a lumped diagonal matrix, see [9].
Lemma 4
The Kuzmin limiter satisfies Assumption (A1).
Proof
Let E be an internal edge that connects the nodes \(\varvec{x}_i\) and \(\varvec{x}_j\). Then it suffices to show that \(\alpha _{ij}(u_h)(u_ju_i)\) is a continuous function of \(u_h\in V_h\). Because of \(\beta _{E}(u) = \beta _{ij} = \beta _{ji}\), \(\alpha _{ij}=\alpha _{ji}\), we can restrict these considerations to the situation that \(a_{ji}\le a_{ij}\). Moreover, it suffices to consider \(d_{ij}<0\) since otherwise \(\alpha _{ij}\equiv 1\).
The last case is \(f_{ij}{(\bar{u}_h)}=0\) which leads to \(\alpha _{ij}(\bar{u}_h)(\bar{u}_j\bar{u}_i)=0\). Since \(\alpha _{ij}\) is bounded by definition, \(\alpha _{ij}(u_h)(u_ju_i) \rightarrow 0\) as \(u_j\rightarrow u_i\). Consequently, \(\alpha _{ij}(u_h)(u_ju_i)\) is continuous at \(\bar{u}_h\). \(\square \)
Remark 3
In [9] it was shown that the terms \(\alpha _{ij}(u_h)(u_ju_i)\) are even Lipschitzcontinuous. The proof of this property is based on the representations (40) and (41) of the coefficients \(\alpha _{ij}\). The sums in these representations are Lipschitzcontinuous and then one can show that the function which is obtained by multiplying these representations with \((u_j  u_i)\) is Lipschitzcontinuous, too.
Lemma 5
Proof
4.2 A limiter leading to linearity preservation and DMP on general meshes (BJK limiter)
Here we present a limiter recently proposed in [11] using some ideas of [35]. This limiter is designed in such a way that the AFC scheme satisfies the discrete maximum principle and linearitypreservation property on arbitrary meshes, which is a substantial improvement in comparison with the Kuzmin limiter. Like in the previous section, the numbers \(d_E\) used in (15) are given by (8).
Lemma 6
The above limiter satisfies Assumptions (A1) and (A2).
Proof
Lemma 7
The above limiter satisfies Assumption (A3).
Proof
Remark 4
Note that large values of the constants \(\gamma _i\) cause that more limiters \(\alpha _{ij}\) will be equal to 1 and hence less artificial diffusion is added, which makes it possible to obtain sharp approximations of layers. On the other hand, however, large values of \(\gamma _i\)’s also cause that the numerical solution of the nonlinear algebraic problem becomes more involved.
4.3 A limiter based on the variation of the discrete solution (BBK limiter)
In this section we review briefly the limiter presented in [10] and its main results. This limiter, also referred to as smoothnessbased viscosity, has its origin in the finite volume literature (see, e.g., [24] and [23]), and has also been used (although in a slightly modified way) in the recent work [19].
Remark 5
In [10, Remark 1] a process to generate a method which is linearity preserving on general meshes is described. It involves a minimization process per node to determine a set of weights. The same results that hold for the method presented in this work hold for that variant.
The next result states that the limiter defined in (46) satisfies Assumptions (A1), (A2), and (A4).
Lemma 8
Proof
4.4 Related recent work
We finish this section by mentioning that, to the research reviewed in this paper, work has been done in parallel, e.g., in [6, 7, 19, 20]. In those references, a stabilizing term similar to the one defined in (15) is added to the formulation, and referred to as the Graph Laplacian. The stabilizing mechanism of the methods presented in those works and the ones reviewed in this manuscript are very similar. There are, nevertheless, significant differences in the limiters \(\beta _E^{}\), some of the results obtained in terms of the satisfaction of the discrete maximum principle, and the linearity preservation of the final schemes.
For example, in [6] the emphasis is in the regularization of the limiter proposed there (related to the BBK one) in order to make the limiter differentiable, to allow the use of Newton’s method to solve the nonlinear system. The regularization proposed in there used regularization parameters that had an impact on the performance of the method. In addition, although the results concerning the discrete maximum principle were not too different from the ones reviewed in this work, the linearity preservation was not guaranteed for the regularized limiters. In [19] the definition of the limiter (nondifferentiable this time) is modified using generalized barycentric coordinates in order to make it differentiable on meshes for which the support of basis functions is convex.
A final important difference between the works reviewed in this paper and the abovementioned references consists in the emphasis. While the papers just quoted deal with first order hyperbolic systems, the results reviewed in this paper deal with the convection–diffusion equation. The presence of the Laplacian in the partial differential equation makes the method satisfy very different properties, especially on nonDelaunay meshes, as it can be seen in [9] where an example of nonconvergence was given for the Kuzmin limiter.
5 Iterative schemes for solving the nonlinear problem
Consider the weak formulation (13) and the equivalent formulation (11), (12) in matrixvector notation. For simplicity, we will restrict the discussion to the case of homogeneous boundary conditions. These formulations represent a nonlinear problem since the coefficients \(\beta _{ij}\) depend on the finite element solution \(u_h\). Applying an iterative scheme for solving the nonlinear problem, our experience is that usually damping is necessary to achieve convergence. Let \(u_h^{(m)}\), \(m \ge 0\), be a given approximation of \(u_h\).
6 Numerical studies
6.1 The Hemker example
We will consider the socalled Hemker example, which was proposed in [21]. It models the convection of temperature from a hot circle (2d cylinder) in a channel. The convection field is constant. There are exponential layers at the circle and interior layers downstream from the circle. The Hemker problem can be considered as a standard benchmark problem for convection–diffusion equations. It was used in [5] for comparing a number of stabilized discretizations. Here, the same setup as in this paper will be considered.
The simulations were performed with \(\mathbb {P}_1\) finite elements on two types of grids, see Fig. 2. Grid 1 is aligned downstream to the convection and it has edges at the position where the interior layer is expected. The stopping criterion for the nonlinear iteration was based on the Euclidean norm of the residual vector, which should be smaller or equal than \(10^{13}\ (\#\mathrm {dof})^{\frac{1}{2}}\), where \(\#\mathrm {dof}\) is the number of degrees of freedom (including Dirichlet nodes) on the respective grid. As initial iterate, a function that vanishes on all degrees of freedom was used. The linear systems were solved with the sparse direct solver UMFPACK, [17]. The simulations were performed with the code MooNMD [27].
By construction of the Hemker example, the solution takes values in [0, 1]. The first quantity of interest from [5] considers the violation of this range by the numerical solutions. Since the AFC methods satisfy the DMP, it is expected that there are no violations if the nonlinear problems are solved exactly. In fact, we could observe in the numerical results only negligible violations of the order of the stopping criterion for the iteration of the nonlinear problem.
Finally, the costs for solving the nonlinear problems is studied. In the used code, only the fixed point iteration (49) is implemented. Either, the selection of the damping parameter as described in [26] can be used or the Anderson acceleration with \(l>0\) vectors and a fixed damping parameter \(\omega \). Results are presented for \(\omega = 0.5\) and \( l =5,10,25\) vectors in the Anderson acceleration. The numbers of iterations that were necessary for solving the nonlinear problems are illustrated in Fig. 6. It can be seen that generally fewer iterations were needed for the Kuzmin limiter. On the structured grid, the variant with 25 vectors in the Anderson acceleration needed often the smallest number of iterations and the fixed point iteration with an adaptive selection of the damping parameter needed most iterations. But on the unstructured grid, there is no clear picture. Using many vectors in the Anderson iteration did even result in failing to reach the stopping criterion on certain levels. For example, results have been obtained for other values of the damping parameter \(\omega \) (not reported here due to space restrictions). More precisely, the use of \(\omega = 0.25\) gave similar results to the ones shown in Fig. 6, needing a few more iterations in most cases. Using \(\omega = 0.7\), we observed that the stopping criterion was not reached in many more cases than for \(\omega =0.5\).
6.2 Illustration of the smearing of layers
A motivation for studying convection–diffusion equations in channel geometries comes from the simulation of population balance systems in chemical engineering. For experiments, chemical engineers often use long and thin pipes. That means, the diameter of the pipes is of the order of a few millimeters or centimeters and the length of the order of several meters. There are several specific properties when considering convection–diffusion equations in pipes or channels. First, a preferred flow direction exists. Second, the grids are eventually aligned with the flow direction and third, the mesh cells might be anisotropic. For convectiondominated problems there is the experience that it is of advantage to align the grid with the convection. In the literature, one finds already observations that report notable smearing of layers for algebraic stabilizations in examples where the grid is aligned to the convection, e.g., in [13, 28].
The deeper understanding of the reasons for the smearing effect and the finding of remedies are open problems. So far, the probably best explanation is given in [28]. Algebraic stabilizations are by construction multidimensional schemes, i.e., there is no dimensional splitting in the construction of the limiters. Such a splitting would be of advantage in this example since it is basically onedimensional. However, the limiters see the layers of the solution that are vertical to the convection and they do not recognize that it is not necessary to introduce notable diffusion for preventing spurious oscillations.
6.3 A threedimensional example
Let \(\varOmega =\varOmega _1\setminus \overline{\varOmega _2}\) with \(\varOmega _1=(0,5)\times (0,2)\times (0,2)\) and \(\varOmega _2=(0.5,0.8)\times (0.8,1.2)\times (0.8,1.2)\). We consider problem (1) with \(\varepsilon =10^{5}\), \(\varvec{b}=(1,\ell (x),\ell (x))^T\) where \(\ell (x)=(0.19x^31.42x^2+2.38x)/4\), \(u_D=1\) on \(\partial \varOmega _1\), \(c=g=0\), and \(u_D=0\) on \(\partial \varOmega _2\). An initial mesh containing 842 elements was generated using gmsh and adaptively refined to a mesh containing 1,308,237 elements by using an SUPG method combined with the a posteriori error estimator from [1]. This adaptively refined mesh was then used to obtain approximations using various AFC methods. The nonlinear problems were solved using the damped fixed point algorithm from [26, Figure 12], and the initial guess was obtained using a standard unstabilized Galerkin approximation.
There is a slight violation of the DMP for the method with the Kuzmin limiter. This violation is due to the fact that the mesh does not respect the hypotheses under which the DMP can be shown, cf. Lemma 5. For this mesh we have found violations of this condition, which explains the numerical results and confirms the sharpness of the analytical results. The boundary and inner layers are significantly sharper for the method with the BJK limiter, although this comes at the price of having to perform significantly more fixed point iterations than with the other methods. In fact, for this example the method using the Kuzmin limiter took 70 iterations to reach convergence, while the method using the BBK limiter took 166 iterations, and the use of the BJK limiter took 1117 iterations to reach convergence. As was mentioned earlier, the BJK and Kuzmin limiters provide sharper profiles than the BBK one. This has been observed not only in this example. This behavior seems to be related to the equal weight given to all fluxes by the construction of the BBK limiter, different from the Kuzmin one (essentially, an upwind limiter), and the BJK limiter, which has the flux associated to the local extremum as main ingredient, and also includes some explicit mesh information to make it linearity preserving.
7 Open problems
The improvement of AFC schemes and the further development of their analysis have been listed in [30] among the most important open problems for \(H^1\)conforming finite elements for convection–diffusion equations. Some concrete issues are the following. It was shown by means of a numerical example that the general a priori estimate given in [9] is sharp. However, one can observe for the Kuzmin limiter and the BJK limiter higher orders of convergence than proved in [9], at least on special grids. So far, there is no concrete characterization of the necessary properties of such grids and no corresponding analysis. A priori analysis of AFC schemes for anisotropic grids remains an open problem. In addition, numerical analysis of AFC schemes for timedependent equations is not available. Last but not least, efficient numerical methods for solving the nonlinear problems have to be developed.
Footnotes
References
 1.Ainsworth, M., Allendes, A., Barrenechea, G.R., Rankin, R.: Fully computable a posteriori error bounds for stabilised FEM approximations of convection–reaction–diffusion problems in three dimensions. Int. J. Numer. Methods Fluids 73(9), 765–790 (2013)MathSciNetGoogle Scholar
 2.Allendes, A., Barrenechea, G.R., Rankin, R.: Fully computable error estimation of a nonlinear, positivitypreserving discretization of the convection–diffusion–reaction equation. SIAM J. Sci. Comput. 39(5), A1903–A1927 (2017)MathSciNetCrossRefGoogle Scholar
 3.Anderson, D.G.: Iterative procedures for nonlinear integral equations. J. Assoc. Comput. Mach. 12, 547–560 (1965)MathSciNetCrossRefGoogle Scholar
 4.Arminjon, P., Dervieux, A.: Construction of TVDlike artificial viscosities on twodimensional arbitrary FEM grids. J. Comput. Phys. 106(1), 176–198 (1993)MathSciNetCrossRefGoogle Scholar
 5.Augustin, M., Caiazzo, A., Fiebach, A., Fuhrmann, J., John, V., Linke, A., Umla, R.: An assessment of discretizations for convectiondominated convection–diffusion equations. Comput. Methods Appl. Mech. Eng. 200(47–48), 3395–3409 (2011)MathSciNetCrossRefGoogle Scholar
 6.Badia, S., Bonilla, J.: Monotonicitypreserving finite element schemes based on differentiable nonlinear stabilization. Comput. Methods Appl. Mech. Eng. 313, 133–158 (2017)MathSciNetCrossRefGoogle Scholar
 7.Badia, S., Hierro, A.: On monotonicitypreserving stabilized finite element approximations of transport problems. SIAM J. Sci. Comput. 36(6), A2673–A2697 (2014)MathSciNetCrossRefGoogle Scholar
 8.Barrenechea, G.R., John, V., Knobloch, P.: Some analytical results for an algebraic flux correction scheme for a steady convection–diffusion equation in one dimension. IMA J. Numer. Anal. 35(4), 1729–1756 (2015)MathSciNetCrossRefGoogle Scholar
 9.Barrenechea, G.R., John, V., Knobloch, P.: Analysis of algebraic flux correction schemes. SIAM J. Numer. Anal. 54(4), 2427–2451 (2016)MathSciNetCrossRefGoogle Scholar
 10.Barrenechea, G.R., Burman, E., Karakatsani, F.: Edgebased nonlinear diffusion for finite element approximations of convection–diffusion equations and its relation to algebraic fluxcorrection schemes. Numer. Math. 135(2), 521–545 (2017)MathSciNetCrossRefGoogle Scholar
 11.Barrenechea, G.R., John, V., Knobloch, P.: An algebraic flux correction scheme satisfying the discrete maximum principle and linearity preservation on general meshes. Math. Models Methods Appl. Sci. 27(3), 525–548 (2017)MathSciNetCrossRefGoogle Scholar
 12.Becker, R., Braack, M.: A twolevel stabilization scheme for the Navier–Stokes equations. In: Feistauer, M., Dolejší, V., Knobloch, P., Najzar, K. (eds.) Numerical Mathematics and Advanced Applications, pp. 123–130. Springer, Berlin (2004)CrossRefGoogle Scholar
 13.Bordás, R., John, V., Schmeyer, E., Thévenin, D.: Numerical methods for the simulation of a coalescencedriven droplet size distribution. Theor. Comput. Fluid Dyn. 27(3–4), 253–271 (2013)CrossRefGoogle Scholar
 14.Boris, J.P., Book, D.L.: Fluxcorrected transport. I. SHASTA, a fluid transport algorithm that works. J. Comput. Phys. 11(1), 38–69 (1973)CrossRefGoogle Scholar
 15.Brooks, A.N., Hughes, T.J.R.: Streamline upwind/Petrov–Galerkin formulations for convection dominated flows with particular emphasis on the incompressible Navier–Stokes equations. Comput. Methods Appl. Mech. Eng. 32(1–3), 199–259 (1982). FENOMECH ’81, Part I (Stuttgart, 1981)MathSciNetCrossRefGoogle Scholar
 16.Burman, E., Hansbo, P.: Edge stabilization for Galerkin approximations of convection–diffusion–reaction problems. Comput. Methods Appl. Mech. Eng. 193(15–16), 1437–1453 (2004)MathSciNetCrossRefGoogle Scholar
 17.Davis, T.A.: Algorithm 832: UMFPACK V4.3—an unsymmetricpattern multifrontal method. ACM Trans. Math. Softw. 30(2), 196–199 (2004)MathSciNetCrossRefGoogle Scholar
 18.Ern, A., Guermond, J.L.: Theory and Practice of Finite Elements, Applied Mathematical Sciences, vol. 159. Springer, New York (2004)CrossRefGoogle Scholar
 19.Guermond, J.L., Popov, B.: Invariant domains and secondorder continuous finite element approximation for scalar conservation equations. SIAM J. Numer. Anal. 55(6), 3120–3146 (2017)MathSciNetCrossRefGoogle Scholar
 20.Guermond, J.L., Nazarov, M., Popov, B., Yang, Y.: A secondorder maximum principle preserving Lagrange finite element technique for nonlinear scalar conservation equations. SIAM J. Numer. Anal. 52(4), 2163–2182 (2014)MathSciNetCrossRefGoogle Scholar
 21.Hemker, P.W.: A singularly perturbed model problem for numerical computation. J. Comput. Appl. Math. 76(1–2), 277–285 (1996)MathSciNetCrossRefGoogle Scholar
 22.Hughes, T.J.R., Brooks, A.: A multidimensional upwind scheme with no crosswind diffusion. In: Finite Element Methods for Convection Dominated Flows (Papers, Winter Ann. Meeting Amer. Soc. Mech. Engrs., New York, 1979), AMD, vol. 34, pp. 19–35. Amer. Soc. Mech. Engrs. (ASME), New York (1979)Google Scholar
 23.Jameson, A.: Origins and further development of the Jameson–Schmidt–Turkel scheme. AIAA J. 55, 1487–1510 (2017)CrossRefGoogle Scholar
 24.Jameson, A., Schmidt, W., Turkel, E.: Numerical solution of the Euler equations by finite volume methods using Runge–Kutta timestepping schemes. AIAA J. In: 14th AIAA Fluid and Plasma Dynamics Conference, AIAA paper 19811259 (1981)Google Scholar
 25.John, V., Knobloch, P.: On spurious oscillations at layers diminishing (SOLD) methods for convection–diffusion equations. I. A review. Comput. Methods Appl. Mech. Eng. 196(17–20), 2197–2215 (2007)MathSciNetCrossRefGoogle Scholar
 26.John, V., Knobloch, P.: On spurious oscillations at layers diminishing (SOLD) methods for convection–diffusion equations. II. Analysis for \(P_1\) and \(Q_1\) finite elements. Comput. Methods Appl. Mech. Eng. 197(21–24), 1997–2014 (2008)CrossRefGoogle Scholar
 27.John, V., Matthies, G.: MooNMD—a program package based on mapped finite element methods. Comput. Vis. Sci. 6(2–3), 163–169 (2004)MathSciNetCrossRefGoogle Scholar
 28.John, V., Novo, J.: On (essentially) nonoscillatory discretizations of evolutionary convection–diffusion equations. J. Comput. Phys. 231(4), 1570–1586 (2012)MathSciNetCrossRefGoogle Scholar
 29.John, V., Schmeyer, E.: Finite element methods for timedependent convection–diffusion–reaction equations with small diffusion. Comput. Methods Appl. Mech. Eng. 198(3–4), 475–494 (2008)MathSciNetCrossRefGoogle Scholar
 30.John, V., Knobloch, P., Novo, J.: Finite elements for scalar convectiondominated equations and incompressible flow problems—a never ending story? Comput. Vis. Sci. (2018). https://doi.org/10.1007/s0079101802905
 31.Knobloch, P.: Numerical solution of convection–diffusion equations using a nonlinear method of upwind type. J. Sci. Comput. 43(3), 454–470 (2010)MathSciNetCrossRefGoogle Scholar
 32.Knobloch, P.: On the discrete maximum principle for algebraic flux correction schemes with limiters of upwind type. In: Zhang, Z., Huang, Z., Stynes, M. (eds.) Boundary and Interior Layers, Computational and Asymptotic Methods BAIL 2016, Lecture Notes in Computational Science and Engineering, vol. 120, pp. 129–139. Springer, New York (2017)Google Scholar
 33.Kuzmin, D.: On the design of generalpurpose flux limiters for finite element schemes. I. Scalar convection. J. Comput. Phys. 219(2), 513–531 (2006)MathSciNetCrossRefGoogle Scholar
 34.Kuzmin, D.: Algebraic flux correction for finite element discretizations of coupled systems. In: Papadrakakis, M., Oñate, E., Schrefler, B. (eds.) Proceedings of the International Conference on Computational Methods for Coupled Problems in Science and Engineering, pp. 1–5. CIMNE, Barcelona (2007)Google Scholar
 35.Kuzmin, D.: Linearitypreserving flux correction and convergence acceleration for constrained Galerkin schemes. J. Comput. Appl. Math. 236(9), 2317–2337 (2012)MathSciNetCrossRefGoogle Scholar
 36.Kuzmin, D., Möller, M.: Algebraic flux correction I. Scalar conservation laws. In: Kuzmin, D., Löhner, R., Turek, S. (eds.) FluxCorrected Transport. Principles, Algorithms, and Applications, pp. 155–206. Springer, Berlin (2005)Google Scholar
 37.Kuzmin, D., Turek, S.: Highresolution FEMTVD schemes based on a fully multidimensional flux limiter. J. Comput. Phys. 198(1), 131–158 (2004)MathSciNetCrossRefGoogle Scholar
 38.Lohmann, C., Kuzmin, D., Shadid, J.N., Mabuza, S.: Fluxcorrected transport algorithms for continuous Galerkin methods based on high order Bernstein finite elements. J. Comput. Phys. 344, 151–186 (2017)MathSciNetCrossRefGoogle Scholar
 39.Löhner, R., Morgan, K., Peraire, J., Vahdati, M.: Finite element fluxcorrected transport (FEMFCT) for the Euler and Navier–Stokes equations. Int. J. Numer. Methods Fluids 7(10), 1093–1109 (1987)CrossRefGoogle Scholar
 40.Temam, R.: Navier–Stokes Equations. Theory and Numerical Analysis. Studies in Mathematics and its Applications, vol. 2. NorthHolland, Amsterdam, (1977)Google Scholar
 41.Walker, H.F., Ni, P.: Anderson acceleration for fixedpoint iterations. SIAM J. Numer. Anal. 49(4), 1715–1735 (2011)MathSciNetCrossRefGoogle Scholar
 42.Wesseling, P.: Principles of Computational Fluid Dynamics, Springer Series in Computational Mathematics, vol. 29. Springer, Berlin (2001)CrossRefGoogle Scholar
 43.Zalesak, S.T.: Fully multidimensional fluxcorrected transport algorithms for fluids. J. Comput. Phys. 31(3), 335–362 (1979)MathSciNetCrossRefGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.