Numer. Math.

Numerische
DOI 10.1007/s00211-016-0808-z Mathematik

Edge-based nonlinear diffusion for finite element

approximations of convection–diffusion equations
and its relation to algebraic flux-correction schemes

Gabriel R. Barrenechea1 · Erik Burman2 ·

Fotini Karakatsani3

Received: 23 September 2015 / Revised: 7 March 2016

© The Author(s) 2016. This article is published with open access at Springerlink.com

Abstract For the case of approximation of convection–diffusion equations using

piecewise affine continuous finite elements a new edge-based nonlinear diffusion
operator is proposed that makes the scheme satisfy a discrete maximum principle.
The diffusion operator is shown to be Lipschitz continuous and linearity preserving.
Using these properties we provide a full stability and error analysis, which, in the diffu-
sion dominated regime, shows existence, uniqueness and optimal convergence. Then
the algebraic flux correction method is recalled and we show that the present method
can be interpreted as an algebraic flux correction method for a particular definition of
the flux limiters. The performance of the method is illustrated on some numerical test
cases in two space dimensions.

Mathematics Subject Classification 65N30 · 65N12

1 Introduction

For an open bounded polygonal (polyhedral) domain  ⊆ Rd , d = 2, 3, with

Lipschitz boundary, we consider in this work the steady-state convection–diffusion–
reaction equation

B Gabriel R. Barrenechea
[email protected]

1 Department of Mathematics and Statistics, University of Strathclyde, 26 Richmond Street,

Glasgow G1 1XH, UK
2 Department of Mathematics, University College London, Gower Street, London WC1E 6BY, UK
3 Department of Mathematics, University of Chester, Thornton Science Park,
Chester CH2 4NU, UK

123
 G. R. Barrenechea et al.


−ε u + b · ∇u + σ u = f in ,
(1.1)
u=g on ∂,
where ε > 0 is the diffusion coefficient, b ∈ L ∞ ()2 is a solenoidal convective field,
1
σ > 0 is a real constant, and f ∈ L 2 (), g ∈ H 2 (∂), are given data. In this work
we adopt the standard notation for Sobolev spaces. In particular, for D ⊂ Rd we
denote (·, ·) D the L 2 (D) [or L 2 (D)d ] inner product, and by  · l,D (| · |l,D ) the norm
(seminorm) in H l (D) [with the usual convention that H 0 (D) = L 2 (D)].
The weak form of problem (1.1) is: Find u ∈ H 1 () such that u = g on ∂ and

a(u, v) = ( f, v) ∀v ∈ H01 (), (1.2)

where the bilinear form a is given by

a(u, v) := ε (∇u, ∇v) + (b · ∇u, v) + σ (u, v) .

The weak problem (1.2) has a unique solution u ∈ H 1 () and its solution satisfies
the following maximum principle (see [10]).
Definition 1 (Maximum principle) Assume that f ≥ 0, g ≥ 0 (resp. ≤ 0) and the
solution u of (1.2) is smooth enough. Then, if σ = 0 and u attains a strict minimum
(resp. maximum) at an interior point x̃ ∈ , then u is constant in . If σ > 0, then
the same conclusion remains valid if we suppose in addition that u(x̃) < 0 [resp.
u(x̃) > 0].
This work deals with the development of a method that satisfies the discrete analogue
of the last definition. The quest for such a method has been a constant for the last
couple of decades. Several methods have been proposed over the years, both in the
finite element and finite volume contexts (see [21] for a review). Overall, the common
point of all discretisations that satisfy a discrete maximum principle (DMP) is that they
add some diffusion to the equations. This extra diffusion can lead to a linear method,
but it is a well-known fact that such a method will provide very diffused numerical
solutions, which will converge suboptimally. Due to the previous fact, several methods
that add nonlinear diffusion have been proposed.
One approach has been to add a so-called shock-capturing term to the finite element
formulation. This typically amounts to a nonlinear diffusion term where the diffusion
coefficient depends nonlinearly on the finite element residual, making it large in the
zones where the solution is underresolved, but vanish in smooth regions. An analy-
sis showing that nonlinear shock capturing methods may lead to a DMP was first
proposed in [5], and then developed further for the Laplace operator in [6], and for
the convection–diffusion equation in [7]. For a review of shock capturing methods,
designed to reduce spurious oscillations, without necessarily satisfying a DMP, see
[14]. More recent nonlinear discretisations, these ones based on the idea of blending
in order to satisfy the DMP, are the works [1,9], where the emphasis has been given to
prove the convergence to an entropy solution. Most shock capturing techniques suf-
fer from the strong nonlinearity introduced when the diffusion coefficient is made to
depend on the finite element residual (and therefore the gradient of the approximation

123
Nonlinear diffusion and AFC schemes

function). Because of this the analysis of such methods is incomplete even when linear
model problems with constant coefficients are considered. In particular, in most cases
uniqueness of solutions can not be proved, and the convergence theory is incomplete.
On the other hand, driven initially by the design of explicit time stepping schemes
for compressible flows, so called flux corrected transport (FCT) schemes and the
related algebraic flux correction (AFC) schemes were introduced [15,19,20]. These
schemes act on the algebraic level by first modifying the system matrix so that it has
suitable properties to make the system monotonous, while perturbing the method as
little as possible. In the most elementary case the system matrix is simply perturbed
to make it an M-matrix, resulting in a linear method. This crude strategy, however,
necessarily results in a first order scheme. Then, AFC schemes introduce a nonlinear
switch, or flux limiter, thus making the low order monotone scheme active only in the
zones where the DMP may be violated. These schemes have also resisted mathematical
analysis for a long time, but a number of results have been proved recently in [2,3].
Indeed, in these references, existence of solutions and positivity have been proved, and
a first error analysis has been performed. Nevertheless, it was shown that the DMP,
and even the convergence of the discrete solution to the continuous one, depend on
the geometry of the mesh.
Another approach to combine monotone (low order) finite element methods with
linear diffusion and high order FEM using flux-limiters was proposed very recently
in [13]. It then appears that a cross pollination between the idea of AFC and shock-
capturing could be fruitful.
The objective of the present paper is to further bridge the gap between the shock
capturing approach and the algebraic flux correction. Indeed we will consider a gen-
eralisation of the shock-capturing term first introduced in [4] to several dimensions,
using an anisotropic diffusion operator along element edges similar to that introduced
in [7]. We show that the resulting scheme satisfies the DMP and give an analysis of the
method. In particular we show that the new shock capturing term is Lipschitz continu-
ous, and, if the mesh is sufficiently regular, linearity preserving (see Sect. 2.1), which
allows us to improve greatly on previous results. In Sect. 2.2 we prove existence of
solutions, the discrete maximum principle, and noticeably, uniqueness in the diffusion
dominated regime. We then show error estimates, which, thanks to the combined use of
linearity preservation and Lipschitz continuity, turn out to be optimal in the diffusion
dominated regime, for a special class of meshes (see Sect. 3). In Sect. 4, we revisit
the design principles of AFC and show that the proposed shock-capturing term can
be interpreted as an AFC scheme using a special flux, allowing both for a DMP and
Lipschitz continuity. Some numerical results are finally shown in Sect. 5.

1.1 Notations

We now introduce some notation that will be needed for the discrete setting. We
consider a family {Th }h>0 of shape-regular triangulations of  consisting of disjoint
d-simplices K . We define h K := diam(K ), and h = max{h K : K ∈ Th }. We associate
with the triangulation Th the finite element spaces

123
 G. R. Barrenechea et al.

Fig. 1 In two dimensions, meshes (a–c) are examples of symmetric meshes. Mesh (d) is a non-symmetric,
non-Delaunay mesh

Vh := {χ ∈ H 1 () : χ | K ∈ P1 (K ) ∀K ∈ Th }, and Vh0 := Vh ∩ H01 (), (1.3)

where P (D) is the space of polynomials of degree at most  on D. The nodes of Th

are denoted by {xi }i=1
N , and the usual associated basis functions of V are denoted by
h
{ψi }i=1 .
N

We let Eh be the set of the interior edges of Th . For every edge E ∈ Eh , we define
h E := |E| and ω E := {K ∈ Th : K ∩ E = ∅}, and fix one unit tangent vector, denoted
by t.
For an interior node xi , we define the associated edges Ei := {E ∈ Eh : xi ∈ E}
and the subset of Rd defined by the union of all elements K sharing the node xi ,
i := {x ∈  ¯ : ∃K ∈ Th : x ∈ K and xi ∈ K }, and the set

Si := { j ∈ {1, . . . , N }\{i} : x j shares an internal edge with xi }. (1.4)

Finally, we will say that the triangulation Th is symmetric with respect to its internal
nodes if for every internal node xi the following holds: for all j ∈ Si there exists k ∈ Si
such that x j − xi = −(xk − xi ) (see Fig. 1 for examples in two space dimensions).

2 The nonlinear discretisation

The standard finite element method for the problem (1.2) takes the form: find u h ∈ Vh
such that u h − u bh ∈ Vh0 and

a(u h , vh ) = ( f, vh ) ∀vh ∈ Vh0 . (2.1)

Here, u bh ∈ Vh is introduced to approximate the boundary condition g. Then, we

propose the following stabilised method to discretise (1.2): find u h ∈ Vh such that
u h − u bh ∈ Vh0 and

ah (u h ; vh ) := a(u h , vh ) + dh (u h ; u h , vh ) = ( f, vh ) ∀vh ∈ Vh0 . (2.2)

The stabilisation term dh (· ; ·, ·) is defined by


dh (wh ; u h , vh ) = γ0 h dE α E (wh )(∂ t u h , ∂ t vh ) E . (2.3)
E∈Eh

123
Nonlinear diffusion and AFC schemes

Here, γ0 > 0, and α E : Vh → [0, 1] is defined as follows. First, for wh ∈ Vh , we

define ξwh as the unique element in Vh0 whose nodal values are given by
⎧  
⎨  j∈Si wh (xi )−wh (x j ) 
ξwh (xi ) :=

|wh (xi )−wh (x j )|
, if j∈Si |wh (xi ) − wh (x j )| = 0, (2.4)
⎩ j∈Si
0, otherwise.

Then, on each E, α E is defined by

p
α E (wh ) := max ξwh (x) , p ∈ [1, +∞). (2.5)
x∈E

The value for p will determine the rate of decay of the numerical diffusion with the
distance to the critical points. A value closer to 1 will add more diffusion in the far
field, while a larger value will make the diffusion vanish faster, but on the other hand,
increasing p may make the nonlinear system more difficult to solve. In principle, as p
goes to infinity the method will add the perturbations only in points with local extrema.
In our calculations we have tested several different values for p, and have presented
those for p = 1, 4, 8, and 10. The higher values provide better numerical results, while
keeping the nonlinear solver converging within a reasonable number of iterations. In
Sect. 5 below we present a more detailed study of the behavior of the nonlinear solver
with respect to the value of p.We finally stress the fact that, for any value of p, the
function α E (wh ) is equal to 1 if wh has a local extremum in one of the end points of
the edge E. This property is of fundamental importance for the proof of the discrete
maximum principle below.

2.1 Properties of dh (·; ·, ·)

We start noticing that


|wh (xi ) − wh (x j )| = 0 ⇒ wh |i = c ∈ R.
j∈Si

This prevents the method from adding artificial diffusion to the equations in regions
in which the solution is constant. Moreover, the method is as well linearity preserving
if the mesh is symmetric with respect to its interior nodes. In fact, if E ∈ Eh has
endpoints xi and x j , and vh ∈ P1 (ω E ), then
 
vh (xi ) − vh (xl ) = 0 and vh (x j ) − vh (xl ) = 0, (2.6)
l∈Si l∈S j

which gives α E (vh ) = 0. Then, the method does not add extra diffusion in smooth
regions, whenever the mesh is sufficiently structured. We now state this in a more pre-
cise way. Let us decompose the stabilisation term dh as the sum of edge contributions
as follows:

123
 G. R. Barrenechea et al.


dh (u h ; vh , z h ) = d E (u h ; vh , z h )
E∈Eh

with d E (u h ; vh , z h ) := γ0 h dE α E (u h )(∂ t vh , ∂ t z h ) E .

Then, if the mesh is symmetric with respect to its internal nodes and E ∈ Eh , whenever
vh ∈ P1 (ω E ), the edge diffusion vanishes, this is

d E (vh ; wh , z h ) = 0 ∀wh , z h ∈ Vh .

As a consequence, if, for a given node xi , with associated basis function ψi , we denote
the extended macro element  ˜ i := ∪ E∈E ω E , then
i

˜ i ).
dh (vh ; wh , ψi ) = 0, ∀wh ∈ Vh and ∀vh : vh |˜ i ∈ P1 (

The next step is to show that dh (·; ·, ·) is continuous. More precisely, it is Lipschitz
continuous, and the next result is the first step towards this.

Lemma 1 For any vh , wh ∈ Vh , and any given internal node xi , the following holds

E∈Ei h E |∂ t (vh − wh )|
|ξvh (xi ) − ξwh (xi )| ≤ 4  . (2.7)
E∈Ei h E (|∂ t vh | + |∂ t wh |)
 
Proof It is enough to suppose that j∈Si |vh (xi ) − vh (x j )| > 0 and j∈Si |wh (xi ) −
wh (x j )| > 0, otherwise the claim is obvious. A quick calculation gives

|ξvh (xi ) − ξwh (xi )|

   
 


 v (x ) − v (x )   w (x ) − w (x ) 
 j∈Si h i h j j∈Si h i h j
=   −  

 E∈Ei h E |∂ t vh | E∈Ei h E |∂ t wh | 
   
 



 j∈Si vh (x i ) − vh (x j ) −  j∈Si wh (x i ) − wh (x j ) 
≤   

 E∈Ei h E |∂ t vh | 
  
  
  1 1 
+  wh (xi ) − wh (x j )   − 
 j∈Si   E∈Ei h E |∂ t vh | E∈Ei h E |∂ t wh | 

E∈Ei h E |∂ t (vh − wh )|
≤ 
E∈Ei h E |∂ t vh |
   
  
 j∈Si wh (xi ) − wh (x j )  E∈Ei h E (|∂ t wh | − |∂ t vh |)
+  
E∈Ei h E |∂ t vh | E∈Ei h E |∂ t wh |

E∈Ei h E |∂ t (vh − wh )|
≤2  .
E∈Ei h E |∂ t vh |

123
Nonlinear diffusion and AFC schemes

The following estimate can be proved in an analogous way


E∈Ei h E |∂ t (vh − wh )|
|ξvh (xi ) − ξwh (xi )| ≤ 2  .
E∈Ei h E |∂ t wh |

Then,

|ξvh (xi ) − ξwh (xi )|

1 1 
≤ 2 min  , h E |∂ t (vh − wh )|, (2.8)
E∈Ei h E |∂ t v h | E∈Ei E |∂ t wh |
h
E∈Ei

which gives the desired result upon applying the estimate min{a −1 , b−1 } ≤ 2
a+b , for
two positive numbers a and b. 


The Lipschitz continuity of dh (·; ·, ·) appears then as a consequence of the previous

result.

Lemma 2 The nonlinear form dh (·; ·, ·) is Lipschitz continuous. More precisely, there
exists Clip > 0, independent of h, such that, for all vh , wh , z h ∈ Vh , the following
holds

|dh (vh ; vh , z h ) − dh (wh ; wh , z h )| ≤ Clip γ0 h |vh − wh |1, |z h |1, . (2.9)

Proof We have

dh (vh ; vh , z h ) − dh (wh ; wh , z h )
 
= γ0 h dE α E (vh )∂ t vh − α E (wh )∂ t wh , ∂ t z h E
E∈Eh

= γ0 h dE α E (vh )(∂ t vh − ∂ t wh , ∂ t z h ) E
E∈Eh

+ γ0 h dE (α E (vh ) − α E (wh ))(∂ t wh , ∂ t z h ) E . (2.10)

The first term in the above estimate is bounded using the fact that |α E (vh )| ≤ 1, the
Cauchy–Schwarz inequality, a local trace inequality, and the shape regularity of the
mesh sequence, to give

γ0 h dE α E (vh )(∂ t vh − ∂ t wh , ∂ t z h ) E ≤ Cγ0 h |vh − wh |1, |z h |1, . (2.11)
E∈Eh

The second term is bounded next. For this, a general edge E ∈ Eh will be considered as
having xi and x j as endpoints, where xi is chosen to be the vertex such that α E (vh ) =
p
ξvh (xi ). We then divide Eh = E 1 ∪ E 2 , where

E 1 := {E ∈ Eh : α E (vh ) = ξvph (xi ), α E (wh ) = ξwph (xi )},

123
 G. R. Barrenechea et al.

E 2 := {E ∈ Eh : α E (vh ) = ξvph (xi ), α E (wh ) = ξwph (x j )},

and the second term in (2.10) reduces to

 
γ0 h dE (ξvph (xi ) − ξwph (xi ))∂ t wh , ∂ t z h E
E∈E 1
 
+ γ0 h dE (ξvph (xi ) − ξwph (x j ))∂ t wh , ∂ t z h E
.
E∈E 2

We now remark that for two numbers a, b ∈ [0, 1] we have


p−1
|a p − b p | = |a − b| a l b p−1−l ≤ p |a − b|,
l=0

and the term in E 1 is bounded using Lemma 1. In fact, from the shape regularity of the
mesh sequence there exists C > 0, independent of h, such that for all E, F ∈ Ei , h F ≤
Ch E . Moreover, the number of edges in Ei is uniformly bounded, independently of h.
Then, using Cauchy–Schwarz’s inequality and a local trace inequality we arrive at
 
γ0 h dE (ξvph (xi ) − ξwph (xi ))∂ t wh , ∂ t z h E
E∈E 1
 
≤p γ0 h dE |ξvh (xi ) − ξwh (xi )|∂ t wh , ∂ t z h E
E∈E 1
  
 F∈Ei h F |∂ t (vh − wh )| F |
≤p γ0 h dE 4 |∂ t wh |, |∂ t z h |
h F (|∂ t vh | F | + |∂ t wh | F |)
F∈Ei
E∈E 1 E
⎛ ⎞
  
≤ 4 p γ0 h dE ⎝ ∂ t (vh − wh )| , |∂ t z h |⎠
F
E∈E 1 F∈Ei E
≤ Cγ0 h |vh − wh |1, |z h |1, . (2.12)

The sum over E 2 is bounded next. First, using (2.12) we get

 
γ0 h dE (ξvph (xi ) − ξwph (x j ))∂ t wh , ∂ t z h E
E∈E 2
 
= γ0 h dE (ξvph (xi ) − ξwph (xi ))∂ t wh , ∂ t z h E
E∈E 2
 
+ γ0 h dE (ξwph (xi ) − ξwph (x j ))∂ t wh , ∂ t z h E
E∈E 2
 
≤ Cγ0 h |vh − wh |1, |z h |1, + γ0 h dE (ξwph (xi ) − ξwph (x j ))∂ t wh , ∂ t z h E
.
E∈E 2

123
Nonlinear diffusion and AFC schemes

In an analogous way we obtain

 
γ0 h dE (ξvph (xi ) − ξwph (x j ))∂ t wh , ∂ t z h E
≤ Cγ0 h |vh − wh |1, |z h |1,
E∈E 2
 
+ γ0 h dE (ξvph (xi ) − ξvph (x j ))∂ t wh , ∂ t z h E
.
E∈E 2

Hence

γ0 h dE ((ξvph (xi ) − ξwph (x j ))∂ t wh , ∂ t z h ) E
E∈E 2
≤ Cγ0 h |vh − wh |1, |z h |1,

+ γ0 h dE min{(ξvph (xi ) − ξvph (x j ))(∂ t wh , ∂ t z h ) E ,
E∈E 2
(ξwph (xi ) − ξwph (x j ))(∂ t wh , ∂ t z h ) E }
≤ Cγ0 h |vh − wh |1, |z h |1, , (2.13)

since the last term in the middle inequality is always non-positive, since by construc-
p p p p
tion, for E ∈ E 2 , ξvh (xi ) − ξvh (x j ) ≥ 0 and ξwh (xi ) − ξwh (x j ) ≤ 0. The result then
follows collecting (2.10)–(2.13). 


Remark 1 It is worth remarking that a modification of the method can be introduced

in such a way that the method becomes linearity preserving on general meshes. This
modification is based on the introduction of appropriate weights in the definition of ξwh .
More precisely, instead of its original definition (2.4), we can introduce the following
modified one: for wh ∈ Vh and any internal node xi
⎧ 
⎨ j∈Si βi j (wh (x j )−wh (xi )) 
ξwh (xi ) :=

βi j |wh (xi )−wh (x j )|
if j∈Si βi j |wh (x j ) − wh (xi )| = 0,
⎩ j∈Si
0 otherwise.

The coefficients βi j are designed in such a way that they satisfy the linearirty preser-
vation property. Denoting τ i j = x j − xi , this condition reads

⎛ ⎞
   
∀v ∈ P1 (i ) βi j v(x j ) − v(xi ) = βi j ∇v · τ i j = ∇v · ⎝ βi j τ i j ⎠ = 0,
j∈Si j∈Si j∈Si

which is equivalent to imposing


βi j τ i j = 0. (2.14)
j∈Si

123
 G. R. Barrenechea et al.

The Eq. (2.14) is a first restriction that the coefficients have to satisfy. A further
restriction on βi j is their strict positivity. Then, we impose

βi j ≥ C0 > 0, (2.15)

where the value of C0 is of no great importance. Finally, in case the mesh is symetric
with respect to its interior nodes, then βi j = 1 for all i, j should be an acceptable (and
preferred) solution. Then, we find βi j as the solution of the following problem: for all
internal node xi , find
⎧ ⎫
 ⎨ ⎬
βi j = argmin |δi j − 1|2 : {δi j } satisfies the restrictions (2.14), (2.15) .
j∈Si ⎩ ⎭
j∈Si
(2.16)
The same results that are presented for the original definition of ξ in (2.4) can be
obtained for the present modification. For simplicity of the presentation, and also to
avoid the computational complexity of solving the constrained optimisation problem
(2.16), we have preferred to use in the rest of the paper the original definition (2.4).

2.2 Solvability of the discrete problem

This section is devoted to analyse the existence of solutions for (2.2). It is interesting to
remark that, thanks to the Lipschitz continuity of dh (·; ·, ·), the solution can be proved
to be unique in the diffusion-dominated regime.

Lemma 3 Let Th : Vh0 → [Vh0 ] be the operator defined by

[Th z h , vh ] = ah (z h + u bh ; vh ) − ( f, vh ) , z h , vh ∈ Vh0 , (2.17)

where [·, ·] denotes the duality pairing between Vh0 and its dual. Then,

[Th z h , z h ] ≥ c1 |z h |21, − c2 (u bh 21, +  f 20, ), (2.18)

where c1 , c2 are positive constants independent of z h , f , and g.

Proof For this proof only, we will consider constants C > 0 that may depend on the
physical coefficients. From the definition of a it follows that

a(z h , z h ) = ε |z h |21, + (σ z h , z h ) ≥ ε |z h |21, . (2.19)

Moreover, the definition of dh (·; ·, ·) and the fact that 0 ≤ α E (z h + u bh ) give


dh (z h + u bh ; z h , z h ) = γ0 h dE α E (z h + u bh )∂ t z h 20,E ≥ 0. (2.20)
E∈Eh

123
Nonlinear diffusion and AFC schemes

Then, the definition of the operator Th gives

[Th z h , z h ] ≥ ε|z h |21, + a(u bh , z h ) + dh (z h + u bh ; u bh , z h ) − ( f, z h ) . (2.21)

Next, the Cauchy–Schwarz and Poincaré inequalities lead to the following bound

a(u bh , z h )| = |ε(∇u bh , ∇z h ) + (b · ∇u bh , z h ) + (σ u bh , z h )|
≤ ε |u bh |1, |z h |1, + b∞, u bh 1, z h 0, + Cσ u bh 0, z h 0,
≤ Cu bh 1, |z h |1, . (2.22)

In addition, using the shape regularity of the mesh sequence, α E (·) ≤ 1, and the local
trace inequality, we arrive at

|dh (z h + u bh ; u bh , z h )| = γ0 h dE α E (z h + u bh )(∂ t u bh , ∂ t z h ) E
E∈Eh

≤ γ0 h dE ∂ t u bh 0,E ∂ t z h 0,E
E∈Eh
≤ C h |u bh |1, |z h |1, . (2.23)

We can thus conclude that

[Th z h , z h ] ≥ ε|z h |21, − C u bh 1, |z h |1, −  f 0, z h 0, .

The claimed result arises by applying the Poincaré and Young inequalities to the last
relation. 


The solvability of the nonlinear problem (2.2) appears as a consequence of the above
result and Brower’s fixed point theorem.

Theorem 1 The discrete problem (2.2) has at least one solution. Moreover, if
Clip γ0 h < ε, where Clip is the constant from Lemma 2, then the solution is unique.

Proof First, since the bilinear form a(·, ·) is continuous, and dh (·; ·, ·) is Lipschitz
continuous, then the operator Th is Lipschitz continuous. Next, in view of (2.18), for
any z h ∈ Vh0 such that

c2 (u bh 21, +  f 20, )

|z h |21, = 2 ,
c1

Lemma 3 gives
[Th z h , z h ] = c2 (u bh 20, +  f 20, ) > 0. (2.24)
Then, using a consequence of Brower’s fixed point Theorem (see [11, Corollary 1.1,
Ch. IV]), there exists ṽh ∈ Vh0 such that Th (ṽh ) = 0. Hence, u h := ṽh + u bh solves
(2.2).

123
 G. R. Barrenechea et al.

In order to prove uniqueness, let u 1h , u 2h be two solutions of (2.2). Then, using (2.2)
for both solutions, denoting ẽh := u 1h − u 2h , and using the Lipschitz continuity of
dh (·; ·, ·), we obtain

ε |ẽh |21, ≤ a(ẽh , ẽh ) = −dh (u 1h ; u 1h , ẽh ) + dh (u 2h ; u 2h , ẽh ) ≤ Clip γ0 h |ẽh |21, .
(2.25)

This leads to
(ε − Clip γ0 h) |ẽh |21, ≤ 0, (2.26)

which, using that ẽh ∈ H01 (), finishes the proof. 



2.3 The discrete maximum principle

This section is devoted to prove that method (2.2) preserves positivity. For this, we
will impose the following geometric hypothesis on the mesh. This hypothesis can be
tracked back to [22], and in two space dimensions it reduces to impose that the mesh
is Delaunay.

Assumption 1 (Hypothesis of Xu and Zikatanov, cf. [22]) For every internal edge
E ∈ Eh with end points xi and x j the following inequality holds

1 
|ωiKj | cot(θiKj ) ≥ 0, (2.27)
d(d − 1)
K ∈ω E

where θiKj is the angle between the two facets in K opposite to xi and x j (denoted by
Fi,K and F j,K , respectively), and ωiKj is the (d − 2)-dimensional simplex Fi,K ∩ F j,K
opposite to the edge E.

We now introduce the discrete analogue of the maximum principle. This definition
is related to the one from [7], and it leads to results which are, essentially, identical to
those from that reference.

Definition 2 (DMP) The semilinear form ah (·; ·) is said to satisfy the strong DMP
property if the following holds: for all u h ∈ Vh and for all interior vertices xi , if u h
is locally minimal (resp. maximal) on the vertex xi over the macro-element i , then
there exist negative quantites (c E ) E∈Ei such that
  
ah (u h ; ψi ) ≤ c E ∂ t u h | E , (2.28)
E∈Ei

  
[resp. ah (u h ; ψi ) ≥ − E∈Ei c E ∂ t u h | E ]. Furthermore, we will say that the semi-
linear form satisfies the weak DMP property, related to local minima, if (2.28) holds
only under the additional assumption that the local minimum above is supposed to be
negative.

123
Nonlinear diffusion and AFC schemes

A direct consequence of this definition is the following result analoguous to that of

[7, Proposition 2.5]. We reproduce the proof here for the reader’s convenience.

Lemma 4 Assume that the semilinear form ah (·; ·) satisfies the DMP property.
Assume that u h ∈ Vh solves (2.2) and that f ≥ 0. Then u h reaches its minimum
on the boundary ∂ and for the weak DMP-property, if g ≥ 0, then u h ≥ 0 in .

Proof Assume that the DMP is satisfied and u h reaches its minimum in an interior
vertex xi . Since ah (·; ·) satisfies (2.28), u h is constant over i , implying that the
minimum is taken in all vertices x j ∈ i . Repeating the argument we eventually
deduce that the minimum is reached on the boundary. 


The following result states the DMP for (2.2).

Theorem 2 Let us suppose that the mesh Th satisfies Assumption 1, and that the
parameter γ0 is large enough. Then, the semilinear form ah (·; ·) satisfies the weak
DMP property for σ > 0 and the strong DMP-property for σ = 0.

Proof Let us suppose that u h has a negative local minimum at an interior node xi .
Then, α E (u h ) = 1 for all E ∈ Ei , which gives

ah (u h ; ψi ) = (σ u h , ψi ) + ε(∇u h , ∇ψi ) + (b · ∇u h , ψi )

+ γ0 h dE (∂ t u h , ∂ t ψi ) E . (2.29)
E∈Ei

We will analyse the expression above term-by-term. First, if u h ≤ 0 in the support of

ψi , then (σ u h , ψi ) ≤ 0. Let us suppose now that u h changes sign in the support of
ψi , and let K ∈ i be an element in which u h changes sign. Let xk be a node in K
such that u h (xk ) ≥ 0, and let E ik be the edge connecting these two nodes. Then, using
the Cauchy–Schwarz inequality, a Poincaré inequality in K , and the shape regularity
of the mesh sequence, we arrive at

(σ u h , ψi ) K ≤ σ u h 0,K ψi 0,K

d
≤ Cσ h K2 u h 0,K
 
≤ Cσ h dK h Eik ∂ t u h | Eik .

Then, adding up over all K ∈ i and using the shape regularity of the mesh sequence
we obtain 
(σ u h , ψi ) ≤ C0 σ h d+1
E |∂t u h | E |. (2.30)
E∈Ei

Also, as in [7] (see also [21]), Assumption 1 on the mesh leads to

ε(∇u h , ∇ψi ) ≤ 0. (2.31)

123
 G. R. Barrenechea et al.

N 
Moreover j=1 ψ j = 1 gives j∈Si (b · ∇ψ j , ψi ) = 0, and then

(b · ∇u h , ψi ) = (b · ∇ψ j , ψi ) u h (x j ) + (b · ∇ψi , ψi ) u h (xi )
j∈Si
 
= (b · ∇ψ j , ψi ) u h (x j ) − u h (xi )
j∈Si
  
= (b · ∇ψ j , ψi ) h E ∂ t u h | E , (2.32)
E∈Ei

which, using the shape regularity of the mesh sequence gives


(b · ∇u h , ψi ) ≤ C1 b∞,E h dE |∂ t u h | E |. (2.33)
E∈Ei

Finally, since u h (xi ) is a local minimum, then in every E ∈ Ei , ∂ t u h and ∂ t ψi have

different signs (independently of the orientation of the tangential vector in E), which
gives    
γ0 h dE (∂ t u h , ∂ t ψi ) E = − γ0 h dE ∂ t u h | E . (2.34)
E∈Ei E∈Ei
Hence, gathering all the above computations, we arrive at
  
ah (u h ; ψi ) ≤ − (γ0 − C0 σ h E − C1 b∞,E )h dE ∂ t u h | E , (2.35)
E∈Ei

and the result follows assuming that γ0 > C0 σ h E + C1 b∞,E . Finally, we notice
that if σ = 0 then the sign of the strict minimum is irrelevant, which proves the strong
DMP property. 

Remark 2 It is interesting to remark that the hypothesis on the meshes of the triangu-
lation can be avoided if the problem is supposed to be strongly convection-dominated.
In fact, following analogous steps to those used to prove (2.32) we can arrive at
 
ε(∇u h , ∇ψi ) = ε (∇ψ j , ∇ψi ) h E |∂ t u h | ≤ C2 εh d−1
E |∂ t u h |. (2.36)
E∈Ei E∈Ei

Replacing this into the steps leading to (2.35) gives


ah (u h ; ψi ) ≤ − (γ0 − C0 σ h E − C1 b∞,E − C2 εh −1
E )h E |∂ t u h |, (2.37)
d

E∈Ei

and the proof follows by assuming that γ0 > C0 σ h E + C1 b∞,E + C2 εh −1 E .

−1
The last result is only interesting if εh E stays bounded, which means this is applica-
ble only in the case the problem is highly convection-dominated. In this sense, the
method proposed in this work can be applied to scalar conservation laws, regardless
of the geometrical impositions on the mesh. Similar results have been obtained recently
in [12,13].

123
Nonlinear diffusion and AFC schemes

3 Convergence

The error will be analysed using the following norm:

vh 2h := σ vh 20, + ε |vh |21, + dh (u h ; vh , vh ). (3.1)

This norm is not only mesh-dependent, but also depends on the discrete solution. The
inclusion of the last term in it is made mostly for convenience, but the fact that it
controls the usual H 1 ()-norm (weighted by physical coefficients) guarantees that
the convergence of the method is valid with respect to the standard norm as well. As
usual, the error e := u − u h is split as follows

e = u − u h = (u − i h u) + (i h u − u h ) := ρh + eh , (3.2)

where i h : C 0 () ∩ H01 () → Vh0 stands for the Clément interpolation operator.
Using standard interpolation estimates (see [8]), the fact that α E (·) ≤ 1, and the shape
regularity of the mesh sequence, the following bound for ρh follows:

1 1 1
ρh h ≤ C(ε 2 + σ 2 h + γ0 h 2 ) h u2, . (3.3)

The next result states a bound for eh .

Lemma 5 Let us suppose u ∈ H 2 ()∩ H01 (). Then, there exists C > 0, independent
of h and ε, such that

1 1
eh h ≤ C ε + σ −1 {b2∞, + σ 2 } 2 hu2, + Ch 2 u1, . (3.4)

Proof First, from the definition of a and dh we get

eh 2h = a(eh , eh ) + dh (u h ; eh , eh )

= a(i h u, eh ) − {a(u h , eh ) + dh (u h ; u h , eh )} + dh (u h ; i h u, eh )
= − a(ρh , eh ) + dh (u h ; i h u, eh ). (3.5)

Next, the continuity of a gives

1
a(ρh , eh ) ≤ (σ ρh 20, + [ε + σ −1 b2∞, ] |ρh |21, ) 2 eh h
1 1
≤ C(ε 2 + σ −1/2 b∞, + σ 2 h) h u2, eh h . (3.6)

Moreover, since dh (u h ; ·, ·) is a symmetric positive semi-definite bilinear form it sat-

isfies Cauchy–Schwarz’s inequality, which gives

1 1 1
dh (u h ; i h u, eh ) ≤ dh (u h ; i h u, i h u) 2 dh (u h ; eh , eh ) 2 ≤ dh (u h ; i h u, i h u) 2 eh h .
(3.7)

123
 G. R. Barrenechea et al.

Then, inserting (3.6) and (3.7) into (3.5), and using Young’s inequality, we arrive at
1 1
eh 2h ≤ C(ε 2 + σ −1/2 b∞, + σ 2 h)2 h 2 u22, + C dh (u h ; i h u, i h u). (3.8)

It only remains to bound the consistency error dh (u h ; i h u, i h u) in (3.8). The definition

of dh (·; ·, ·), α E (u h ) ≤ 1, a local trace inequality, the shape regularity of the mesh
sequence, and the H 1 ()-stability of i h , give

dh (u h ; i h u, i h u) = γ0 h dE α E (u h )∂ t i h u20,E
E∈Eh

≤ γ0 h h d−1
E ∂ t i h u0,E ≤ Ch u1, .
2 2
(3.9)
E∈Eh

Then, the result arises inserting (3.9) into (3.8). 


Collecting (3.3) and Lemma 5 we then obtain the following error estimate for (2.2).
Theorem 3 Let us suppose u ∈ H 2 () ∩ H01 (). Then, there exists C > 0, indepen-
dent of h and ε, such that
1 1
eh ≤ C ε + σ −1 {b2∞, + σ 2 } 2 hu2, + Ch 2 u1, . (3.10)

The following result states that for meshes which are symmetric with respect to their
interior nodes, the method converges with a higher order. This result’s main interest
1
lies in the diffusion dominated regime, due to the factor ε− 2 present in the estimate.
The combination of Lipschitz continuity and linearity preservation seems to be novel,
and that is why we do detail it now.
Theorem 4 Let us suppose u ∈ H 2 () ∩ H01 () and that the mesh is symmetric with
respect to its internal nodes. Then, there exists C > 0, independent of h and ε, such
that
1 h
eh ≤ C ε + σ −1 {b2∞, + σ 2 } 2 hu2, + C √ u1, . (3.11)
ε
Proof It is enough to bound the consistency error d(u h ; i h u, i h u). We have

dh (u h ; i h u, i h u) = {dh (u h ; i h u, i h u) − dh (i h u; i h u, i h u)} + dh (i h u; i h u, i h u)
=: I + II. (3.12)

The first term is bounded as in the proof of Lemma 2. In fact, in that proof, the bound
for the second term in (2.10) leads to the following

I= (α E (u h ) − α E (i h u))γ0 h dE (∂ t i h u, ∂ t i h u) E
E∈Eh
≤ Ch|u h − i h u|1, |i h u|1,
ε h2
≤ |u h − i h u|21, + C u21, , (3.13)
2 ε

123
Nonlinear diffusion and AFC schemes

where we have also used the H 1 ()-stability of i h . To bound II we use the linearity
preservation and the Lipschitz continuity of dh (·; ·, ·). More precisely, for a given
E ∈ Eh we introduce the function i E u ∈ P1 (ω E ) as the unique solution of the
problem

(∇i E u, ∇ψ)ω E = (∇u, ∇ψ)ω E ∀ψ ∈ P1 (ω E ), (3.14)

(i E u, 1)ω E = (u, 1)ω E .

Using standard finite element approximation results (see [8]), i E u satisfies

|u − i E u|1,ω ≤ Ch E |u|2,ω . (3.15)

E E

Since the mesh is symmetric with respect to its internal nodes, α E (i E u) = 0. Then,
proceeding as in the bound for I we obtain

II = (α E (i h u) − α E (i E u))γ0 h dE (∂ t i h u, ∂ t i h u) E
E∈Eh
⎧ ⎫1
⎨ ⎬2
≤ Ch |i h u − i E u|21,ω |i h u|1,
⎩ E ⎭
E∈Eh

≤ Ch 2 |u|2, u1, . (3.16)

Then, inserting (3.13) and (3.16) into (3.12) we obtain

ε h2
dh (u h ; i h u, i h u) ≤ |u h − i h u|21, + C u21, + Ch 2 |u|2, u1, , (3.17)
2 ε

and the result follows by rearranging terms. 



4 A link to algebraic flux correction schemes

Method (2.2) has been presented having as motivation the study of the effect of
adding edge-based diffusion into the equations to impose the discrete maximum
principle. Another family of methods that are built with the same purpose is the
AFC schemes. This section is devoted to study the relationship between the two
approaches, and that is why we now summarise the main building principles of AFC
schemes.
The starting point of an algebraic flux-correction scheme is a discretisation of the
convection–diffusion–reaction equation which leads to the linear system

AU = G, (4.1)

123
 G. R. Barrenechea et al.

where A = (ai j )i,N j=1 , U = {u h (xi )}i=1

N and G = {g } N . The first step of these
i i=1
schemes is to identify which parts of the system matrix A are responsible for the
violation of the discrete maximum principle. To achieve this, the diffusion matrix
D = (di j )i,N j=1 is built, where

di j = d ji = − max{ai j , 0, a ji } ∀i = j dii = − di j .
j =i

Adding DU both sides of (4.1) we obtain

ÃU = G + DU, (4.2)

where à := A + D. Since the matrix à fullfils the hypothesis to guarantee the discrete
maximum principle, then the oscillations that appear in a non-stabilised discretisation
(4.1) are due to the right-hand side. This is why the right-hand side is now rewritten.
Using that the row-sums of D are zero, then

(DU)i = f i j where f i j = di j (u h (x j ) − u h (xi )).
j =i

The quantities f i j are called fluxes. Then, the AFC schemes are based on introducing
limiters αi j (u h ) such that αi j ∈ [0, 1], αi j = α ji , and αi j = 1 if xi and x j are both
Dirichlet nodes. Then, after introducing these limiters, the method reads as follows:


N
AUi + (1 − αi j (u h ))di j (u h (x j ) − u h (xi )) = gi . (4.3)
i, j=1

The most popular limiters in practice are Zalesak’s limiters (see, Refs. [15–17,23],
and the recent review [18] for examples). The analysis of these methods for a class
of limiters that includes the Zalesak one has been carried out recently in [2,3]. In
1
particular, in [2] an O(h 2 ) convergence rate was proved for the case in which the mesh
used satisfies Assumption 1. In the case of meshes that do not satisfy this assumption,
then no convergence can be proved, unless some appropriate modifications are done
to the algorithm. This result is optimal, as the numerical results in [2] show.
Following [2], Eq. (4.3) can be written as the following weak problem: find u h ∈ Vh
such that u h − u bh ∈ Vh0 , and

a(u h , vh ) + d̃h (u h ; u h , vh ) = ( f, vh ) ∀vh ∈ Vh0 , (4.4)

where the nonlinear form d̃h (·; ·, ·) is given by


N
d̃h (u h ; u h , vh ) = (1 − αi j (u h ))di j (u h (x j ) − u h (xi ))vh (xi ). (4.5)
i, j=1

123
Nonlinear diffusion and AFC schemes

Next, to link this to the method analysed in the last sections, we use the symmetry of
D, and of the limiters αi j = α ji , and a simple calculation gives:

d̃h (u h ; u h , vh ) = (1 − αi j (u h ))di j (u h (x j ) − u h (xi ))vh (xi )
i> j

+ (1 − αi j (u h ))di j (u h (x j ) − u h (xi ))vh (xi )
i< j

= (1 − αi j (u h ))di j (u h (x j ) − u h (xi ))vh (xi )
i> j

+ (1 − α ji (u h ))d ji (u h (xi ) − u h (x j ))vh (x j )
i> j

= (1 − αi j (u h ))di j (u h (x j ) − u h (xi ))(vh (xi ) − vh (x j )). (4.6)
i> j

Then, since di j = 0 for j ∈

/ Si , d̃h (·; ·, ·) can be rewritten as

d̃h (u h ; u h , vh ) = (1 − αi j (u h ))|di j |h E (∂ t u h , ∂ t vh ) E , (4.7)
E∈Eh

where we have adopted the convention that an edge E ∈ Eh has endpoints xi and x j ,
and used that αi j = 1 for edges included in the Dirichlet boundary.
Method (2.2) then appears as an algebraic flux-correction scheme, with a different
definition of the limiters. Indeed comparing (2.2) with (4.7) we get the equivalent AFC
scheme if we choose αi j (u h ) such that

(1 − αi j (u h ))|di j |h E = γ0 h dE α E (u h ).

The new definition of the limiters made it possible to write some convergence and
existence results, also present in [2], in a more precise way, and improve in some of
them. In particular, the new limiters make it possible to prove convergence for general
meshes, as well as to prove uniqueness of solutions and optimal convergence in the
diffusion dominated regime.

5 Numerical results

In this section we present three sets of numerical results for bi-dimensional problems.
All three cases are set in  = (0, 1)2 . The nonlinear system (2.2) has been solved
using the following fixed-point algorithm with damping: starting with the Galerkin
solution u 0h , then compute a sequence {u kh } defined by

u k+1
h = u kh + ω (ũ k+1
h − u kh ) k = 0, 1, 2, . . . , (5.1)

123
 G. R. Barrenechea et al.

where ω ∈ (0, 1) is a damping parameter, and ũ k+1

h solves: ũ k+1
h − u bh ∈ Vh0 , and

a(ũ k+1 k k+1

h , vh ) + dh (u h ; ũ h , vh ) = ( f, vh ) ∀vh ∈ Vh .
0
(5.2)

In all our calculations we have used ω = 0.1, and stopped the iterations when the
residual Rk := (ah (u k+1
h ; ψi )−( f, ψi ) )i=1,...,dim(V 0 ) has an euclidean norm smaller
h
than, or equal to, 10−8 .

5.1 Convergence for a smooth solution

We take b = (2, 1), σ = 1, and different values for ε. We have selected the right-hand-
side and boundary conditions in such a way that the solution is given by u(x, y) =
sin(2π x) sin(2π y). The meshes used were the three-directional mesh (c) and the
non-Delaunay mesh (d) in Fig. 1. In these calculations we have used γ0 = 3 and
p = 4.
The results in Tables 1, 2, 3 and 4 match the theoretical results. In particular we
observe a first order convergence in the diffusion-dominated regime for the mesh (c),
as predicted by Theorem 4, and a second order convergence in the L 2 norm of the error
for both the convection and diffusion-dominated regimes. The latter is in accordance
with the empirical observations that linearity preservation implies such a convergence.
For mesh (d), which is non-symmetric, and hence the method is no longer linearity
preserving, we can observe a first order convergence in both regimes. This convergence
is not affected by the non-Delaunay character of the mesh.
We finish this example by a deeper study of the behavior of the nonlinear fixed-point
iteration with respect to the value of p. The results are reported in Table 5. For these
results, we have used the three-directional mesh (c), with l = 5. We can observe that,
for the values of p ranging from 1 to 10 the iterations needed to reach convergence are
essentially independent of the value of p. This behavior is kept until a value around 20,
and then some non-convergence is observed in the scheme. Here, by non-convergence
we mean that the desired residual reduction has not been achieved after 5000 iterations.
The same qualitative behavior has been observed for other meshes, and the two other
settings presented later. In those cases, non-convergence has been observed starting at

Table 1 ε = 10−6 , numerical results for grid (c)

l u − u h 0, Ord. |u − u h |1, Ord. u − u h h Ord.

3 0.49391 – 4.38896 – 3.62380 –

4 0.47965 0.04 4.26871 0.04 3.08479 0.23
5 0.19110 1.33 2.71665 0.65 1.08371 1.51
6 0.04080 2.23 1.55469 0.81 0.22671 2.26
7 0.00683 2.58 0.64692 1.27 0.03904 2.54
8 0.00119 2.52 0.27480 1.24 0.00689 2.50

123
Nonlinear diffusion and AFC schemes

Table 2 ε = 1, numerical results for grid (c)

l u − u h 0, Ord. |u − u h |1, Ord. u − u h h Ord.

3 0.38594 – 3.48242 – 5.44504 –

4 0.16557 1.22 1.90920 0.87 2.26966 1.26
5 0.03268 2.34 0.89029 1.10 0.92785 1.29
6 0.00612 2.42 0.43637 1.03 0.43912 1.08
7 0.00141 2.12 0.21800 1.00 0.21818 1.01
8 0.00035 2.02 0.10903 1.00 0.10904 1.00

Table 3 ε = 10−6 , numerical results for grid (d)

l u − u h 0, Ord. |u − u h |1, Ord. u − u h h Ord.

3 0.48754 – 4.33607 – 5.06989 –

4 0.45680 0.09 4.11426 0.08 2.93242 0.79
5 0.17080 1.42 3.15455 0.38 1.05213 1.48
6 0.04330 1.98 2.23948 0.49 0.26065 2.01
7 0.01165 1.89 1.72410 0.38 0.05482 2.25
8 0.00474 1.30 1.63424 0.08 0.02087 1.39

Table 4 ε = 1, numerical results for grid (d)

l u − u h 0, Ord. |u − u h |1, Ord. u − u h h Ord.

3 0.38351 – 3.52996 – 5.57464 –

4 0.16616 1.21 2.00539 0.82 2.41681 1.21
5 0.04513 1.88 0.98086 1.03 1.03172 1.23
6 0.01277 1.82 0.48118 1.03 0.48720 1.08
7 0.00423 1.59 0.23973 1.01 0.24059 1.02
8 0.00163 1.38 0.11982 1.00 0.11998 1.00

Table 5 Iterations needed to reach convergence

p 1 2 3 4 5 6 7 8 9 10 15 20

Iter. 224 218 261 262 278 286 211 227 197 197 218 206

values of about 10 or 15, depending on the case. Then, we believe that it is safe to use
this scheme for values of p not much higher than 10. Of course, further work could
be used to find the right damping parameters for each case, but this would come at the
price of having to perform much more iterations.

123
 G. R. Barrenechea et al.

5.2 A problem with one inner layer, and a rotating convective field

We use ε = 10−5 , f = 0, σ = 0, b = (−y, x), homogeneous Neumann boundary

conditions on exit, and


1 if x ≤ 0.5,
g(x, y) =
0 else,

as Dirichlet condition at entry. We have solved this problem on a uniform refinement

of the three-directional from mesh (c) in Fig. 1. The parameter γ0 has been set to 1,
and the results show no violation of the DMP. The results for this case are depicted in
Fig. 2. We can observe that the increase in the value of p provides a solution whose
inner layer is much sharper than the choice p = 1. For both higher values for p, a
similar behaviour to the one in Table 5 was observed in terms of number of iterations
needed for convergence.

Approximate Solution Approximate Solution

1
1
0.9
0.9
0.8
0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0
0 0 0 0.2
1 0.2 1 0.4
0.8 0.4 0.8 0.6
0.6 0.6 0.6
0.4 0.4 0.8
0.2 0.8 0.2 1
0 1 0

(a) (b)

Approximate Solution

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1 0
0.2
0 0.4
1 0.6
0.8
0.6 0.8
0.4
0.2 1
0

(c)
Fig. 2 Discrete solution for p = 1 (top left) and p = 4 (top right), and p = 8 (bottom)

123
Nonlinear diffusion and AFC schemes
Approximate Solution Approximate Solution

1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
1
0.1 1 0.1
0.5 0.5
0 0
0 0.1 0 0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

(a) (b)
Approximate Solution

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
1
0.1
0 0.5
0 0.2 0.4 0
0.6 0.8 1

(c)
Fig. 3 Discrete solution for p = 1 (top left) and p = 4 (top right), and p = 10 (bottom)

5.3 Advection skew to the mesh

π
 π

We use ε = 10−5 , f = 0, σ = 0 b = cos 3 , sin 3 , and


1 if x = 0 or y = 1,
g(x, y) =
0 else,

as Dirichlet condition. We have solved this problem on a criss-cross mesh as shown in

mesh (a) in Fig. 1. We have used the parameter γ0 = 0.75, and, again, no violations of
the DMP have been observed. The results are depicted in Fig. 3, where we can observe
much sharper layers (especially the internal one) when higher values for p have been
used. Again, for both higher values for p, a similar behaviour to the one in Table 5
was observed in terms of number of iterations needed for convergence.

Acknowledgments The work of GRB and FK has been partially funded by the Leverhulme Trust via the
Research Project Grant No. RPG-2012-483. The authors would like to thank Volker John and Petr Knobloch
for very helpful discussions.

Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 Interna-
tional License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution,

123
 G. R. Barrenechea et al.

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.

References
1. Badia, S., Hierro, A.: On monotonicity-preserving stabilized finite element approximations of transport
problems. SIAM J. Sci. Comput. 36(6), A2673–A2697 (2014). doi:10.1137/130927206
2. Barrenechea, G.R., John, V., Knobloch, P.: Analysis of algebraic flux correction schemes. SIAM J.
Numer. Anal. (2015, in press)
3. 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). doi:10.1093/imanum/dru041
4. Burman, E.: On nonlinear artificial viscosity, discrete maximum principle and hyperbolic conservation
laws. BIT 47(4), 715–733 (2007). doi:10.1007/s10543-007-0147-7
5. Burman, E., Ern, A.: Nonlinear diffusion and discrete maximum principle for stabilized Galerkin
approximations of the convection–diffusion–reaction equation. Comput. Methods Appl. Mech. Eng.
191(35), 3833–3855 (2002). doi:10.1016/S0045-7825(02)00318-3
6. Burman, E., Ern, A.: Discrete maximum principle for Galerkin approximations of the Laplace operator
on arbitrary meshes. C. R. Math. Acad. Sci. Paris 338(8), 641–646 (2004). doi:10.1016/j.crma.2004.
02.010
7. Burman, E., Ern, A.: Stabilized Galerkin approximation of convection–diffusion–reaction equations:
discrete maximum principle and convergence. Math. Comput. 74, 1637–1652 (2005)
8. Ern, A., Guermond, J.L.: Theory and Practice of Finite Elements. Springer, New York (2004)
9. Ern, A., Guermond, J.L.: Weighting the edge stabilization. SIAM J. Numer. Anal. 51(3), 1655–1677
(2013). doi:10.1137/120867482
10. Gilbarg, D., Trudinger, N.S.: Elliptic partial differential equations of second order. In: Grundlehren der
Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 224, 2nd
edn. Springer, Berlin (1983). doi:10.1007/978-3-642-61798-0
11. Girault, V., Raviart, P.A.: Finite element methods for Navier–Stokes equations. In: Springer Series in
Computational Mathematics, vol. 5. Springer, Berlin (1986). doi:10.1007/978-3-642-61623-5
12. Guermond, J.L., Nazarov, M.: A maximum-principle preserving C 0 finite element method for scalar
conservation equations. Comput. Methods Appl. Mech. Eng. 272, 198–213 (2014). doi:10.1016/j.cma.
2013.12.015
13. Guermond, J.L., Nazarov, M., Popov, B., Yang, Y.: A second-order maximum principle preserving
Lagrange finite element technique for nonlinear scalar conservation equations. SIAM J. Numer. Anal.
52(4), 2163–2182 (2014). doi:10.1137/130950240
14. John, V., Knobloch, P.: On spurious oscillations at layers diminishing (SOLD) methods for convection–
diffusion equations: part I—a review. Comput. Methods Appl. Mech. Eng. 196, 2197–2215 (2007)
15. Kuzmin, D.: On the design of general-purpose flux limiters for finite element schemes. I. Scalar
convection. J. Comput. Phys. 219, 513–531 (2006)
16. Kuzmin, D.: Algebraic flux correction for finite element discretizations of coupled systems. In:
Papadrakakis, M., Oñate, E., Schrefler, B. (eds.) Proceedings of the Int. Conf. on Computational
Methods for Coupled Problems in Science and Engineering, pp. 1–5. CIMNE, Barcelona (2007)
17. Kuzmin, D.: Linearity-preserving flux correction and convergence acceleration for constrained
Galerkin schemes. J. Comput. Appl. Math. 236, 2317–2337 (2012)
18. Kuzmin, D., Hämäläinen, J.: Finite element methods for computational fluid dynamics. In: Compu-
tational Science and Engineering, vol. 14. Society for Industrial and Applied Mathematics (SIAM),
Philadelphia (2015). (A practical guide)
19. Kuzmin, D., Turek, S.: Flux correction tools for finite elements. J. Comput. Phys. 175(2), 525–558
(2002). doi:10.1006/jcph.2001.6955
20. Löhner, R., Morgan, K., Vahdati, M., Boris, J.P., Book, D.L.: FEM-FCT: combining unstructured
grids with high resolution. Commun. Appl. Numer. Methods 4(6), 717–729 (1988). doi:10.1002/cnm.
1630040605
21. Roos, H.G., Stynes, M., Tobiska, L.: Robust Numerical Methods for Singularly Perturbed Differential
Equations. Convection–Diffusion–Reaction and Flow Problems, 2nd edn. Springer, Berlin (2008)

123
Nonlinear diffusion and AFC schemes

22. Xu, J., Zikatanov, L.: A monotone finite element scheme for convection–diffusion equations. Math.
Comput. 68(228), 1429–1446 (1999). doi:10.1090/S0025-5718-99-01148-5
23. Zalesak, S.T.: Fully multidimensional flux-corrected transport algorithms for fluids. J. Comput. Phys.
31, 335–362 (1979)

123