Journal of Computational Physics 272 (2014) 360385

Contents lists available at ScienceDirect

Journal of Computational Physics

www.elsevier.com/locate/jcp

A high-order mimetic method on unstructured polyhedral

meshes for the diffusion equation
K. Lipnikov a , G. Manzini a,b,c,
a

Los Alamos National Laboratory, Theoretical Division, Group T-5, MS B284, Los Alamos, NM 87545, USA
Istituto di Matematica Applicata e Tecnologie Informatiche, Consiglio Nazionale delle Ricerche (IMATI-CNR), via Ferrata 1,
I-27100 Pavia, Italy
c
Centro di Simulazione Numerica Avanzata (CeSNA) IUSS Pavia, v.le Lungo Ticino Sforza 56, I-27100 Pavia, Italy
b

a r t i c l e

i n f o

Article history:
Received 7 December 2013
Received in revised form 5 March 2014
Accepted 4 April 2014
Available online 16 April 2014
Keywords:
High-order method
Unstructured polyhedral mesh
Mimetic nite difference method
Diffusion problem

a b s t r a c t
We present a new family of mimetic nite difference schemes for solving elliptic partial
differential equations in the primal form on unstructured polyhedral meshes. These
mimetic discretizations are built to satisfy local consistency and stability conditions. The
consistency condition is an exactness property, i.e., the mimetic schemes are exact when
the solution is a polynomial of an assigned degree. The stability condition ensures the
well-posedness of the method. The degrees of freedom are the solution moments on mesh
faces and inside mesh cells. Higher order schemes are built using higher order moments.
The developed schemes are veried numerically on diffusion problems with constant and
spatially variable (possibly, discontinuous) tensorial coecients.
Published by Elsevier Inc.

1. Introduction
The mimetic discretization framework has been developed to solve PDEs on arbitrary polygonal and polyhedral
meshes [11]. In contrast to nite volume methods [22] that can also handle general meshes, it has a solid mathematical foundation based on a discrete vector and tensor calculus [28]. Thus, the resulting discrete schemes preserve or mimic
important properties of continuum PDEs such as symmetry and positivity of discrete operators, exact discrete identities, and
discrete Helmholtz space decompositions. In this paper, we exploit exibility of the mimetic framework to mix and match
degrees of freedom of various nature to develop a new family of mimetic nite difference (MFD) schemes for the diffusion
equation.
The development of the MFD method has a long history. It has been applied successfully to a wide range of scientic and
engineering problems, such as continuum mechanics [33], discretization of differential forms [14,34], electromagnetics [16,
25,27], gas dynamics [18], linear diffusion equation [2,12,17,20,26,29,30], convectiondiffusion equation [6,21], steady Stokes
equations [710], elasticity [3], elliptic obstacle [1], ReissnerMindlin plates [13], eigenvalues [19] and two-phase ows in
porous media [31].
Due to the importance of discrete conservations laws for engineering simulations, the MFD schemes were originally
developed for systems of PDEs formulated using rst-order differential operators. The discrete operators are constructed in
pairs: rst principles are used to discretize one operator (e.g. primary gradient, divergence, and curl), and discrete duality
relationships are used to derive the second discrete operator (respectively, derived divergence, gradient, and curl).

Corresponding author.

http://dx.doi.org/10.1016/j.jcp.2014.04.021
0021-9991/Published by Elsevier Inc.

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

361

In [15], it was realized that two underlying discretization principles guiding construction of the derived mimetic operators, named consistency and stability, can be applied to primal formulations of PDEs. More precisely, the resulting MFD
method builds a discrete bilinear form Ah that approximates the continuum form A and preserves its kernel. This construction is done cell-by-cell like in the nite element method (FEM). In contrast to the FEM, there is no analog of a unisolvency
condition that often makes construction of high-order elements a non-trivial task. The mimetic framework allows us to use
variable number of degrees of freedom per cell that can be associated with different geometric objects and have different
meaning (e.g. pointwise values, moments, or projection of vector functions on face normals). Usage of additional degrees of
freedom to simplify the construction violates the unisolvency condition but does not break the convergence and stability
properties. It simply leads to a parametric family of mimetic schemes that can be further analyzed for existence of schemes
with additional properties such as the discrete maximum principle [29].
In [9], the consistency and stability conditions are developed for arbitrary-order mimetic schemes for elliptic equations
in two dimensions. The consistency condition is formulated as an exactness property for polynomials of a given degree.
On its turn, the stability condition enforces the coercivity of the discrete bilinear form and, eventually, the well-posedness
of the resulting mimetic scheme. Extension of these schemes to three dimensions requires the construction of high-order
quadrature rules for polygonal faces of polyhedral cells. Such quadrature rules are not available for an arbitrary polygon and
their numerical construction will make the method too expensive. The polygonal FEMs also suffers of this issue [32,3539],
which is overcome only by the Galerkin reformulation of the mimetic methods, recently proposed as the virtual element
method in [4].
In this paper, we resolve this issue using a special choice of the degrees of freedom. Instead of using nodal degrees of
freedom, which may be associated with either the mesh vertices and other special nodes on the cell interfaces, we use
solution moments on faces and inside cells. The construction requires to calculate moments of only polynomial functions
which is a problem with a well-known solution. The new mimetic schemes are suitable to the numerical approximation
of two- and three-dimensional elliptic problems at any order of accuracy on an arbitrary polygonal or polyhedral mesh.
It is worth mentioning that this kind of approximation shares many characteristics with the non-conforming nite element
method, and, for this reason, we might refer to it as the non-conforming mimetic method.
The paper is organized as follows. In Section 2, we introduce the model problem, we briey review the basic concepts
of mimetic discretizations, and we describe the construction of the new non-conforming mimetic method. In Section 3
we prove the consistency and stability of the proposed mimetic method. In Section 4 we discuss two special cases where
the general formulation is signicantly simpler: the low-order approximation and the non-conforming mimetic method
(of any order of accuracy) for problems with constant diffusion tensors. In Section 5, we study the performance of the
proposed method for the numerical resolution of problems with smooth and discontinuous coecients. Final conclusions
are in Section 6.
2. The mimetic nite difference method
In this section, we introduce the model problem, we give a brief overview of the basic concepts of the MFD method, and
we describe the construction of the new non-conforming mimetic method.
2.1. Model problem
The steady diffusion problem for the scalar solution eld u is governed by the following equation:

div(K u ) = f
u=g

in ,

(1)

on ,

(2)

where R is a polyhedral domain with Lipschitz boundary , K is a diffusion tensor describing material properties,
3

f L 2 () is a given loading term, and g H 2 ( ) is a given function. We assume that K is a bounded, measurable, and
symmetric tensor in ( W 1, ())33 . We also assume that K is strongly elliptic, i.e., there exist two positive constants
and such that

v2  v K(x)v  v2 v R3 , x ,

where v is the Euclidean norm of vector v. The strong ellipticity implies that matrix K(x) is strictly positive denite and
thus non-singular for every x .
We consider the ane subspace of H 1 (),




V g = v H 1 ()  v | = g ,
and the linear subspace V0 for g = 0. Let us introduce the bilinear form

A( u , v ) =

K u v dV

(3)

362

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

and the linear functional

f , v
=

f v dV .

(4)

The variational form of problem (1)(2) reads as:

nd u V g such that:

A(u , v ) = f , v
v V0 ().

(5)

The existence and uniqueness of the weak solution follows from the continuity and coercivity of the bilinear form, see, for
instance, [23].
2.2. Mesh notations
Let {h }h be a sequence of conforming mesh partitions of , where h is the mesh size parameter. The meshes h are
regular in
the sense specied in Section 3. Each mesh h is a collection of closed non-overlapping polyhedral cells P such
that = Ph P. We denote the volume of P by |P|, its center of gravity by xP = (xP , y P , zP ), its boundary by P, and its
diameter by hP . As usual, we set h = maxPh hP . We denote a mesh face by f, its barycenter by xf , its area by |f|, and its
diameter by hf . Let P be the number of faces in cell P. Moreover, we will use a local numbering of the cell faces, e.g., fi for
i = 1, . . . , P . Finally, let nP,f be the unit normal vector to f pointing out of P.
2.3. Basic concepts of the MFD method
The formulation of the MFD method includes three basic steps. In the rst step, we introduce the degrees of freedom
which allow us to represent a scalar function v H 1 () by a discrete eld v h (see Section 2.5). Each v h is an ordered set
of real numbers associated with mesh faces and possibly with mesh cells and can be treated as an algebraic vector. Later,
we will refer to v h as either the discrete eld or the algebraic vector of degrees of freedom. The set of all vectors v h forms
a linear space V h . Given v h V h and a cell P, we consider a discrete eld v hP := v h |P , the restriction of v h to P, which is
formed by the degrees of freedom of v h associated with cell P. The algebraic vectors v hP form a linear space VPh , which can
be considered as the restriction of V h to P.
To formalize the construction of discrete elds, we introduce the reduction operator h : H 1 () V h , e.g., v h = h ( v ).
Similarly, for each cell P, the discrete representation of v H 1 (P) is returned by the local reduction operator Ph : H 1 (P)
VPh , i.e., v hP = Ph ( v ). The local and global reductions operators are dened consistently so that Ph ( v ) is the restriction of
h ( v ) to cell P.
In the second step, we construct a discrete bilinear form Ah : V h V h R and a discrete functional f ,
h : V h R that
approximate the left- and right-hand sides of (5), respectively:



Ah h (u ), h ( v )

K u v dV

and

f , h (v )

f v dV .

(6)

This construction will be carried out independently on each mesh cell using the consistency and stability conditions (see
Section 2.6 for the general scheme construction and Section 4.1 for the low-order case).
In the third step, we dene the mimetic approximation of problem (5) using Ah (, ) and f ,
h . To this end, we need the
ane subspace V gh of V h , whose elements use the function g to dene the values of the degrees of freedom on , and V0h ,

the linear subspace of V h , that corresponds to g = 0. The mimetic approximation of the variational problem (5) reads:

nd u h V gh such that:




h
Ah uh , v h = f , v h

v h V0h .

(7)

The coercivity and continuity of Ah (see Section 3) imply the well-posedness of this approximation.
2.4. The mimetic stiffness matrix
We construct the bilinear form Ah as the sum of local discrete bilinear forms AhP : VPh VPh R acting on the local
discrete elds:


  h h h
Ah uh , v h =
AP u P , v P u h , v h V h .

(8)

Ph

Each local bilinear form AhP is associated with a polyhedral cell P and must be an approximation of the corresponding
continuous bilinear form (compare with the rst relationship in (6)):



AhP Ph (u ), Ph ( v )

K u v dV AP (u , v ).
P

(9)

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

363

Thus, AhP must be symmetric and positive semi-denite and its action on the vectors u hP and v hP can be represented by a
symmetric and positive semi-denite matrix MP :


  T
AhP uhP , v hP = v hP MP uhP .

(10)

Due to the locality of our construction, we will drop the subscript P and write M instead of MP . In contrast to the FEM,
we derive the matrix M as a solution of a set of algebraic equations. More precisely, we require that AhP satises two
fundamental properties called the consistency and stability conditions.
2.4.1. The consistency condition
In this subsection, we present the consistency condition for the general case of a spatially variable tensor K. Let us
introduce the orthogonal projector m1 : L 2 (P) Pm1 (P) for m  1. By denition, for any scalar function w L 2 (P) it
holds that

w m1 ( w ) q dV = 0 q Pm1 (P).

(11)

We say that the bilinear form AhP satises the consistency condition if

AhP

Ph (q), Ph ( v )

m1 (K q) v dV q Pm (P), v H 1 (P).

(12)

Remark 2.1. In the MFD methodology we do not use the basis functions as in the nite element method. Thus, we cannot
approximate directly the local integrals like that in (9) by using high-order accurate quadrature rules. The projector m1 is
introduced to overcome this problem. It allows us to compute the bilinear form AhP exactly using various moments of v as
the degrees of freedom. In addition, the orthogonality property of this projector plays a critical role in showing symmetry
of the resulting discretization.
Remark 2.2. When m = 1 or K is constant, the formulation of the consistency condition can be simplied as we do not need
the projection operator m1 anymore, see Section 4.2.
2.4.2. The stability condition
The stability condition guarantees the well-posedness of the discrete problem. To formalize it, let us introduce the local
semi-norm on VPh ,

 h 
 v 
P

VPh



= min v hP c Ph (1) ,

(13)

c R

and the global semi-norm

  2
 h 2
 v  h =
 v h  h ,
P V
V
P

Ph

which is also a norm on V0h . We say that the bilinear form AhP satises the stability condition if there exist two positive
constants C and C independent of h and P such that:

2

2

C hP  v hP V h  AhP v hP , v hP  C hP  v hP V h
P

v hP VPh .

(14)

The stability of the bilinear form AhP will be analyzed in Section 3.

2.5. The degrees of freedom
The high-order formulation requires two different types of degrees of freedom for a discrete eld v h , which are associated with the mesh cells and faces. For convenience of exposition, we will refer to them as the cell moments and the face
moments to highlight their connection to moments of a continuous function v.
2.5.1. The cell moments
Let us consider the polynomial space Pm (P) for any integer m  1. It has dimension nm = (m + 1)(m + 2)(m + 3)/6 and
nm
with q1 = 1.
is generated by the set of basis functions {q j } j =
1
Let k = (x,k , y ,k , z,k ) denote a triple of integer indices x,k , y ,k , z,k  0 with the usual multi-index notation
|k | = x,k + y,k + z,k . We enumerate all possible triples k with |k |  m using the subscript k running from 1 to nm .

364

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

Fig. 1. Degrees of freedom for m = 1, 2, 3 on three visible faces of a cubic cell (circles) and internal degrees of freedom (diamonds).
n

m
The polynomial space Pm (P) is generated by the hierarchical basis {qk (x)}k=
where
1

qk (x) =

x xP

k

hP

x xP

x,k

y yP

hP

 y,k

hP

z zP

z,k

hP

(15)

For m  2 and 1  k  nm2 (where nm2 = (m 1)m(m + 1)/6), we denote the k-th cell moment of a function v by

v P,k =

|P |

qk v dV .

(16)

For m = 1 there are no cell moments in the formulation of the MFD method.
2.5.2. The face moments
To dene the moments of the function v on a two-dimensional face f we consider the space of two-dimensional poly
nomials Pm1 (f), which has dimension nm
1 = (m + 1)m/2. Let us introduce a two-dimensional local coordinate system
= (, ) on f. In this coordinate system, we denote the barycenter xf of f by f = (f , f ).
Let k = (x,k , y ,k ) denote a pair of integer indices x,k , y ,k  0 with the usual multi-index notation |k | = x,k + y ,k .

We enumerate all the possible pairs k with |k |  m 1 using the subscript k running from 1 to nm
1 . The polynomial
n

m1
where
space Pm1 (f) is generated by the hierarchical basis {rk ( )}k=
1

rk ( ) =

k

hf

f
hf

,k

,k

hf

(17)


For m  1 and 1  k  nm
1 , we denote the k-th face moments of function v on face f by

v f,k =

|f|

rf,k v dS .

(18)

The middle and right plots in Fig. 1 illustrate the degrees of freedom for m = 2, 3 for a cubic cell.
2.5.3. The linear spaces V h and VPh
The degrees of freedom of a discrete function v h are its face and cell moments

 }f , { v P,1 , v P,2 , . . . , v P,n }P .

v h = { v f,1 , v f,2 , . . . , v f,nm
m
h
h

The linear space V h of such discrete elds has dimension


h = nm
1 [number of mesh faces] + nm2 [number of mesh cells].
h

The restriction of v to a polyhedral cell P, i.e.

moments related to P:

v hP = { v f,1 , v f,2 , . . . , v f,n

m 1

v hP ,

(19)

contains the face moments related to the faces f of P and the cell


}f P , v P,1 , v P,2 , . . . , v P,nm2 .


The local linear spaces VPh of discrete elds v hP has dimension dim(VPh ) = nm
1 P + nm2 .

The local reduction operator Ph : H 1 (P) VPh , when applied to a function v, returns the discrete eld v hP whose cell
and face moments are dened by (16) and (18), respectively.

Remark 2.3. The cell and face moments are dened in (16) and (18) by using the monomials as basis functions for the
polynomial space on P and on every f P. It is worth noting that this choice implicitly gives a hierarchical structure to the
formulation of the method in the sense that to increase the order from m to m + 1 we only need to add a few monomials
to the implementation for m. In practice, we just need to add a few rows (moments) and columns (polynomials) to the

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

365

matrices N and R; hence, the hierarchical structure is helpful in the implementation. However, the formulation of the
method is independent of the choice of the basis for the various polynomials spaces, and, in principle, any polynomial basis
can be used to compute the moments and dene the degrees of freedom.
Remark 2.4. The moments of the polynomials can be computed by recursively reducing the calculation to the polyhedron
faces and, then, to the edges through the divergence theorem.
2.6. General scheme construction
The construction of the method is based on the consistency condition (12), which is an exactness property that holds
when one of the arguments of the discrete bilinear form is a polynomial. As already noted in (2.1), if K(x) is a general
function of x, the term K q is not a polynomial vector function and the use of the orthogonal projection m1 introduced
in (11) is necessary. Using m1 we expand all the integrands in the right-hand side of (21) in a polynomial basis by
projecting the components of K q onto Pm1 (P). The notation m1 (K q) means that the orthogonal projector is applied
independently to each component of K q. In such a case, m1 : ( L 2 (P))3 (Pm1 (P))3 so that

nm1

m1 (K q) =

gs q s ,

(20)

s =1

where gs are the projection vector coecients. We insert this projection operator into the denition of the bilinear form
AP and integrate the result by parts:

m1 (K q) v dV =
P

v div

m1 (K q) dV +

vnP,f m1 (K q) dS

(21)

f P f

(compare with (61)).

We reformulate (21) in three steps by considering separately the cell and face integrals. First, we introduce the expansion coecients c P,k associated with cell P and the divergence term. Second, we introduce the expansion coecients c f,k
associated with each face f of P and the gradient term. Finally, we combine these expansions to obtain the formula for the
right-hand side of (21).
Step (i): divergence term. The divergence term div(m1 (K q)) is a polynomial of degree at most m 2 on P and can be
expanded using the basis functions (15). First, we take the divergence of (20) to obtain:
nm1
 
div m1 (K q) =
gs q s .

(22)

s =2

The summation starts from 2 because q1 = 0. Then, we expand each term gs q s in the basis functions (15):

nm2

gs q s =

g Ps ,k qk

s  2,

(23)

k =1

where g Ps ,k are the expansion coecients and the superscript s indicates the dependence on q s . By substituting (23) into (22)
and rearranging the summations we obtain
nm2
 
div m1 (K q) =

k =1

 nm1



g Ps ,k

nm2

qk =

s =2

nm1

c P,k qk

where c P,k =

g Ps ,k

s =2

k =1

where c P,k are the expansion coecients. Using this formula, we reformulate the divergence term in (21) as


nm2
nm2



v div m1 (K q) dV =
c P,k qk v dV =
c P,k |P| v P,k .


k =1

(24)

k =1

Step (ii ): face terms. The trace of the projected gradient, nP,f m1 (K q), on face f is a polynomial of degree at most
m 1 and can be expanded using the basis functions (17). First, we expand the trace of q s on f in the functions rk :

nm
1

q s |f =


k =1

g fs,k rk .

(25)

366

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

We use (25) in (20) to compute the trace of m1 (K q). Then, we substitute the resulting expression into the face integrals
of (21) and rearrange the summations to obtain


nm
1 nm1


vnP,f m1 (K q) dS =

k =1

s =1


nP,f gs g fs,k


nm
1

rk v dS =

c f,k |f| v f,k

(26)

k =1

where the expansion coecients c f,k are given by:

nm1

c f,k =

nP,f gs g fs,k .

s =1

Step (iii ): reformulation of (21). Now, let us insert (24) and (26) in (21):


nm
1

nm2

m1 (K q) v dV =

c P,k |P| v P,k +

k =1

c f,k |f| v f,k ,

(27)

k =1

which holds for every q Pm (P) and every function v H 1 (P). Taking q = q j and v = qi for i , j = 1, . . . , nm , we obtain the
entries of the stiffness matrix with respect to the polynomial basis of Pm (P):

nm2

m1 (K q j ) qi dV =


nm
1

j
c P,k |P|qi P,k

k =1

c f,k |f|qi f,k ,

(28)

k =1
j

where we added the superscript j to the expansion coecients c P,k and c f,k to stress their dependence on q j . Formula (28)
denes a part of the stiffness matrix M.
2.6.1. The stiffness matrix formula
Let us introduce the two auxiliary matrices N and R that will be used for the construction of matrix M. To ease the
notation, in this subsection and in Section 3 we will omit the subindex P in the matrix symbols M, R and N, although the
denition of these matrices is strictly local and related to a given cell P. The j-th column vector N j of matrix N contains
the degrees of freedom of the j-th polynomial function q j :

N j = Ph (q j )

j = 1, 2, . . . , nm .

(29)

Using the local numbering of the faces of P, i.e., f1 , f2 , . . . , fP , the j-th column of matrix N is given by:

N Tj = q j f ,1 , q j f ,2 , . . . , q j f ,n ,
1
1
1 m 1



q j f ,1 , q j f ,2 , . . . , q j f ,n ,
P
P
P m1

...,



moments of face f1

q j P,1 , q j P,2 , . . . , q j P,n

m 2



(30)

cell moments

moments of face fP

where the face coecients q j f ,t and the cell coecients q j P,s are dened through the developments of Section 2.6. Matrix
s
R has the same size as matrix N. Its j-th column vector R j is given implicitly by

R Tj v hP = AhP Ph (q j ), v hP

v hP VPh .

(31)

According to (31), the columns of R collect the coecients of the various polynomial expansions of the divergence and
gradient terms in (67) and (28):

R Tj = c f ,1 , c f ,1 , . . . , c f ,n ,
1
1
1 m 1



...,

expansion of face f1

c f ,1 , c f ,1 , . . . , c f ,n ,
P m1
P
P



expansion of face fP

qj

c P,1 , c P,2 , . . . , c P,nm2 ,



(32)

expansion of cell P

qj

where the face coecients c fs ,t and the cell coecients c P,s are dened through the developments of Section 2.6.

Let q = q j dened as in (15) and v hP = Ph ( v ) for an enough smooth function v. Using the matrices N and R, Eq. (27)
can be rewritten as

T

m1 (K qi ) v dV = v hP R j .

(33)

Combining (10), the consistency condition (12), relation (33) and denitions (29) and (31), we obtain

v hP

T

R j = AhP v hP , Ph (q j ) = v hP

T

MPh (q j ) = v hP

T

MN j .

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

367

As v hP is arbitrary, we can identify the algebraic relationship MN j = R j for any column index j. Thus, matrix M must be a
solution of the matrix equation

MN = R.

(34)

We will show in Section 3 that the solution to the matrix equation (34) is not unique, i.e. the consistency and stability
conditions do not determine a single scheme but a family of schemes, all sharing the same properties of accuracy and
well-posedness. This feature is common to all MFD formulations that are available in the literature and, according to [24],
it is possible to identify member of the family with additional (superior) properties compared to the other schemes in the
family. This feature originated a new adaptation methodology, called M-adaptation [24], that analyzes how to the enforce
the discrete maximum principles for the diffusion equation in the primal and dual forms, how to the reduce numerical
dispersion and anisotropy for the acoustic wave equation, and how to optimize the performance of multi-grid solvers.
M-adaptivity for the non-conforming mimetic method is out of the scope of the present paper and will be the topic of
future work. Thus, for a practical implementation of the MFD method we will refer to a very specic choice of the mimetic
stiffness matrix, without addressing anymore the issue of its non-uniqueness.
To derive the formula of the mimetic stiffness matrix, let us observe that the reduction operators Ph satisfy Ph (q1 ) = 1,
where 1 = (1, 1, . . . , 1) T . Thus, we can consider the following block column partitioning N = [1, 
N], where 
N collects the
remaining nm 1 columns. Moreover, for j = 1 the consistency condition (12) requires that AhP (Ph (q1 ), v hP ) = 0 for any
v hP = Ph ( v ) since q1 = 0. As v hP is an arbitrary vector in (31), we obtain R1 = 0, where 0 = (0, 0, . . . , 0) T . Thus, we can
consider the following block column partitioning R = [0, 
R]. Let h = dim(V h ) be the dimension of the space V h , i.e., the
total number of degrees of freedom, cf. (19). The most general form for matrix M1

M1 = CUC T ,
where C is a h (h 3)-sized matrix such that N T C = 0 and the columns of N and C span Rh , i.e., span{N, C} = Rh ; U is
a (h 3) (h 3)-sized symmetric and positive denite matrix of parameters. An effective choice that leads to a single
parameter family of schemes is given by U = I, where is a real positive parameter and I is the (properly sized) identity
matrix. Alternatively, we can set U = (C T C)1 , where is a real positive parameter, and we obtain M1 = C(C T C)1 C T .
Now, we observe that

 1

C CT C

 1

CT + N NT N

N T = I,

since span{N, C} = Rh by construction, and C is orthogonal to the columns of N. Thus, we can reformulate this single
parameter family by taking

 1

M1 = I N N T N

NT .

Experimental evidence shows that the MFD method has good approximation properties for a wide range of values (several
orders of magnitude) for ; therefore, we can set = P to x a unique choice for this matrix. Finally, the formula for
matrix M is given by

 T  1


 1 T 

M=
R
R 
N
R + P I N N T N
N ,

 

 1

P = trace 
R
RT 
N



RT .

(35)

In Section 3, we will prove that (35) satises the consistency and stability conditions.
2.7. Discretization of the loading term
We project the loading term f on the polynomial space Pm2 (P) using the basis function qm dened in (15) with m
such that |m |  m 2:

nm2

m2 ( f ) =

f P,k qm ,

m =1

where f P,m are the expansion coecients. We consider the following approximation for the right-hand side of (5):


f v dV

m2 ( f ) v dV =
P

nm2
k =1

qk v dV =

f P,k
P

nm2

f P,k |P| v P,k =: f , v h P ,

(36)

k =1

where v h = Ph ( v ).
The third-order method (m = 2) needs a special treatment. Note that nm2 = 1 and (36) yields

f v dV
P

0 ( f ) v dV = |P| f P,1 v P,1 .

P

(37)

368

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

By denition, f P,1 and v P,1 are the cell averages of f and v, respectively. As pointed out in [9], this approximation is not
accurate enough to provide the O (h3 ) convergence rate in the L 2 -norm. In accordance with [5], this issue can be tackled by
using a better approximation of f and v. Let

f = f P,1 + ( f )P,1

x xP

v = v P,1 + ( v )P,1

and

hP

x xP

hP

where ( f )P,1 and ( v )P,1 are the cell averages of f and v, respectively. Using this and a few algebraic manipulations,
we approximate the integral of the loading term as follows:


f v dV

(x xP )(x xP )T

f v dV = f P,1 v P,1 |P| + ( f )P,1

dV

h2P

( v )P,1 .

(38)

Thus, this strategy adds a correction term to (37). We can further develop the correction term by expressing the cell average
of v through the face moments of v:

( v )P,1 =

|P |

v dV =
P

|P |

v dS =

nP,f

f P

|P |

nP,f |f| v f,1 .

f P

Inserting this into (38) yields the nal expression:


f v dV |P| f P,1 v P,1 +
P

|P |

f P





h

(x xP )(x xP )T
( f )P,1
dV nP,f |f| v f,1 =: f , v h P .
2
hP

(39)

3. Consistency and stability analysis

Let us recall the key matrix equation, MN = R, see formula (34). This equation is the matrix form of the consistency
condition (12). In the next subsection, we will show that formula (35) is a solution to (34).
3.1. Derivation of matrix M and consistency of AhP
Let us rst note that R T 1 = 0. Indeed, using rst (31), then the consistency condition, and nally q1 = 0, we obtain

R Tj 1 = AhP Ph (q j ), 1 = AhP Ph (q j ), Ph (q1 ) =

m1 (K q j ) q1 dV = 0.
P

Matrix N R plays a crucial role in the derivation of formula (35). Repeating the above argument, we obtain a useful formula:

R Tj Ni = AhP Ph (q j ), Ph (qi ) =

m1 (K q j ) qi dV .

(40)

To prove the symmetry of (40) with respect to swapping indices i and j, we note that q i (Pm1 (P))3 . Hence, we can use
the denition of the orthogonal projector (see (11)) for each component to obtain:

m1 (K q j ) qi dV =

K q j qi dV

[use the symmetry of K ]

P

q j K qi dV

use again (11)

P
q j m1 (K qi ) dV .

=
P

Thus, R T N is a symmetric and semi-positive denite matrix. Moreover, 1 is its sole eigenvector corresponding to the zero
eigenvalue. A straightforward calculation shows that


RT N =

0T

RT

1 
N =

0T 1 0T 
N

RT 1 
RT 
N


=

0 0T
.
0 
RT 
N

The sub-matrix 
RT 
N is symmetric and strictly positive denite because it represents the bilinear form AP (see (9)) restricted
to the polynomial space 
Pm (P) = span{q j , j > 1}.

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

369

Let us, now, consider the matrix

M0 = R R T N R T ,

(41)

where (R T N) denotes the pseudo-inverse of the singular matrix R T N, see the next formula. A straightforward calculation
shows that

R
M0 = R R T N R T = 0 


 0

0T
T  1

(R N)



0

RT

 T  1 T

=
R
R 
N
R .

(42)

In view of denition (42), matrix M0 satises (34) because

 T  1 T
 T  1 T 



M0 N = 
R
R 
N
R [1, 
N] = 0, 
R
R 
N
R 
N = [0, 
R] = R.
However, M0 may not satisfy the stability condition. In fact, the kernel of M0 equals the kernel of 
R T , which is generally
larger than span{1} on polyhedral cells. In such a case, we can nd a non-constant function v with a non-zero energy,
AP ( v , v ) > 0, such that v hP = Ph ( v ) ker(M0 ). Since AhP ( v hP , v hP ) = ( v hP )T M0 v hP = 0, the discrete eld v hP does not satisfy the
left inequality in (14).
To x this issue, we add a correction term to M0 , a matrix M1 , which we require to be symmetric, positive semi-denite
and such that

ker M1 = img(N).

(43)

The mimetic stiffness matrix is now given by

M = M0 + M1 .

(44)

Matrix M is obviously symmetric, positive semi-denite and satises (34) since M1 N = 0 and MN = M0 N = R. Moreover, the kernel of M is exactly given by span{1}, so that AhP ( v hP , v hP ) = 0 if and only if v hP is a multiple of a constant vector. In other words, the discrete bilinear form preserves the kernel of the continuous bilinear form. To prove
the kernel characterization, ker(M) = span{1}, let us consider a discrete eld v hP such that M v hP = 0. Then, we obtain
0 = ( v hP ) T M v hP = ( v hP ) T M0 v hP + ( v hP ) T M1 v hP , which implies that both terms must be zero. The second term and formula (43)
zh ) T ) in accordance with
imply the existence of a vector zh such that v hP = N zh . Let us partition such vector as ( zh ) T = ( zh0 , (
the column partitioning of matrix N. Then, the rst term implies

0 = v hP

T

 T

M0 v hP = zh

 T  1 T h  h
 T T  T  1 T  h



NT 
RT 
R 
N
R N z = z0 1 + 
N
zh 
R 
R 
N
R z0 1 + 
N
zh =
zT 
RT 
N
z

from which we conclude that 

zh = 0 as 
RT 
N is a strictly positive denite matrix. Therefore, v hP = zh0 1.
As the choice of M1 is not unique, we have a family of mimetic schemes, all sharing the same accuracy properties.
A possible choice of M1 is given by

 1

M 1 = P I N N T N

NT ,

P = trace M0 .

(45)

Substituting (41) and (45) in (44), we obtain formula (35).

3.2. Stability condition and well-posedness of the method
Stability analysis is limited to shape-regular meshes {h }h . We modify the mesh shape-regularity conditions proposed
in [15].
(MR) A mesh h is called shape-regular if there exist two positive real numbers N and such that
(MR1) every polyhedral cell P h admits a conformal tetrahedral decomposition TP,h that is made of less than N
tetrahedra and includes all vertices of P;
(MR2) each tetrahedron T TP,h is shape-regular, i.e. the radius rT of the inscribed sphere is uniformly bounded from
below:

rT  hP ;

(46)

(MR3) every cell P h is star-shaped with respect to its centroid xP which is the node of TP,h .
The shape-regularity conditions impose some restrictions on the shape of cell P in order to avoid pathological situations
as hP 0. Nonetheless, the meshes of {h }h may contain very general cells including non-convex or degenerate cells as the
ones used in the adaptive mesh renement technology. An example is given in the numerical section, where we consider
a sequence of cubic meshes that are locally rened towards a corner. Such renement leads to cubic-shaped degenerate
polyhedra that have faces subdivided into four co-planar sub-faces as shown in Fig. 2. In a conformal mesh, these cells
are no longer cubic cells, they are actually general polyhedra with 9 or more distinct faces. It is worth noting that not all
methods available in the literature may work on these kind of meshes, e.g. the polyhedral nite element method.

370

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

Fig. 2. Convergence analysis: the rst two random hexahedral meshes (top row) and the rst two cubic meshes with local renement at a corner (bottom
row). In the two plots on top, part of the mesh around the vertex (1, 1, 1) is removed to show the internal structure.

Lemma 3.1. There exists a positive constant depending only on the shape-regularity constants an N such that

vh

T

M 0 v h  P  v h 

v h img(
N),

(47)

for all cells P.

The proof of this lemma is rather technical and is reported in Appendix B.
On its turn, the matrix M1 given by (45) satises the following bounds:

s P  v h   v h


h T

T

M1 v h

 
v h ker NT ,

 2
M v  s P  v h 
v h VPh ,
1 h

(48)
(49)

for some positive constant s , s independent of P. The rst inequality follows from the fact that matrix M1 is a projector
onto (img(N)) = ker(N T ) times P ; thus, it holds

 h 2  h  T h  h  T 

 1 T  h


v  = v
v = v
I N NT N
N v
v h img(N) .
The second inequality uses the fact that the above matrix is the orthogonal projector:

vh

T 

 1

I N NT N

NT v h   v h 

v h VPh .

h of vectors v h V h that are orthogonal to a constant vector, i.e. 1 T v h = 0. The following

Consider a linear space V
P
P
P
P
h .
stability results holds true with respect to the Cartesian norm of vectors in V
P
Theorem 1. Let h be a shape-regular polyhedral mesh that satises conditions (MR1)(MR2). Let matrices M0 and M1 be given
by (42) and (45), respectively. Then, the local bilinear form AhP , represented by the matrix M = M0 + M1 , satises the stability inequalities

C P  v h   v h

T

M v h  C P  v h 

vh 
VPh ,

(50)

for a pair of constants C and C that are independent of h and P.

h and decompose it as v h = 
v h + v h,c , where 
v h img(
N)
Proof. We rst prove the left inequality in (50). Consider v h V
P
h,c

1
1h
and v (img(N)) . From the denition of matrix M , we have M v = 0 (see (43)). Then,

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385


Let

vh

T

Mvh = vh

T

M0 v h + v h,c

T

M1 v h,c = T1 + T2 .

371

(51)

< 1 be a positive number that will be specied later. We bound T1 as follows:

 h T 0  h,c 
 T
 h T 0  h   h,c T 0  h,c 
+2 
T1 = v h M0 v h = 
v
M 
v + v
M v
v
M v


 h T 0  h 
1  h,c  T 0  h,c 
= T11 + T12 .
v
M 
v + 1
v
M v
 (1 ) 

We control T11 by using (47):

T11  (1 ) P 
vh .
We control T12 by noting that

v h,c

T

2
2
2
M0 v h,c  max M0  v h,c   trace M0  v h,c  = P  v h,c 

Since (1 1/ ) is a negative factor whenever

T12  1

(52)

< 1, it holds that

P  v h,c  .

Since v h,c (img(N)) = ker(N T ), term T2 is bounded by (48):

T2 = v h,c

T

M1 v h,c  s P  v h,c  .

(53)

Now, we combine all inequalities in (51):



 h 2

2
 h 2  h,c 2 
1


v M v  (1 ) P 
v
+ s + 1
P  v h,c   C P 
v  + v  ,
h

where C = min{(1 ) , (s + 1 1/ )} is strictly positive when 1/(1 + s ) < < 1.

To prove the right inequality in (50), we use again the inequality ( v h ) T M0 v h  P  v h 2 (see (52)). Using the upper
bound (49) yields:

vh

T

Mvh = vh

T 

M0 + M1 v h  max 1, s

P  v h  .

The right inequality and the theorem are proved by setting C = max{1, s }.

The result of the previous theorem implies the following corollary, which states the stability condition (14) for the MFD
method.
Corollary 3.1. Under the assumptions of Theorem 1, the stability condition (14) holds with constants C and C independent of h
and P.
Proof. The proof of Lemma 3.1 (in particular, see Eq. (B.3)) shows that the positive eigenvalues of M0 are uniformly bounded
from above and from below by P , i.e., by hP . Since the semi-norm ||V h in (14) is equivalent to the Euclidean norm in (50)

h , the corollary follows immediately.

on V
P

3.3. Complexity of the proposed scheme

The cost of a matrixvector product is one of the fair metrics for comparing different methods with the same asymptotic
convergence rate. In fact, modern iterative solvers for linear algebra problems are based on Krylov methods, e.g., GMRES,
BiCG-STAB, Conjugate Gradient, and their implementation essentially requires matrixvector products. To have an idea of
the complexity of the new scheme, we estimate the cost of a matrixvector product for m = 1, 2, 3 on a cubic mesh and
compare it with the cost of the same operation for a Lagrange nite element method.
The cost (per mesh cell) can be calculated approximately as the number of non-zeroes in the matrix divided by the
number of mesh cells, see Table 1. We also present the number of degrees of freedom per mesh cell, labeled as dofs. The
Lagrange FE method is only 20% more cost-ecient for m = 1 and less ecient for m = 2 and m = 3. Note that the proposed
method uses less degrees of freedom per cell when m = 3.
Remark 3.1. A static condensation can be employed to eliminate internal degrees of freedom in both methods making the
Lagrange FE method more competitive for large values of m. However, a detailed cost/benet analysis of the proposed
method is beyond the scope of this work.

372

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

Table 1
Cost per cubic cell of a matrixvector product operation for the Lagrange FEM (left columns) and the new MFD method (right columns).
Lagrange FE method
m=1
m=2
m=3

New MFD method

dofs

cost

dofs

cost

1
8
27

27
476
3231

3
10
22

33
334
1492

4. Special cases
In this section we discuss two important special cases: the low-order formulation in Section 4.1 and the case of constant
diffusion coecients in Section 4.2. In both cases, the construction of the method is signicantly simpler than that in
Section 2 because we do not need to use the orthogonal projection m1 . The stiffness matrix M in (10) is still given by the
matrix formula (35), and requires the modied construction of matrices N and R that is described below.
4.1. The low-order formulation
In this section, we derive the low-order MFD method corresponding to m = 1, i.e. q P1 (P) in the subsequent formulas.
Since q is a constant vector function, it holds that 0 (K q) = 0 (K) q and we can approximate the diffusion coecient
K on cell P by its cell-average value KP = 0 (K). Moreover, the method will still work even if we take the value of the
diffusion tensor at the centroid of P. An integration by parts formula (21) yields:


KP q v dV =


div(KP q) v dV +

(nP KP q) v dS .
P

Since KP q is a constant vector, its divergence is zero. Thus, splitting the surface integral into integrals over the polyhedral
faces, the relation above becomes:


KP q v dV =


nP,f KP q

f P

v dS .

(54)

At this point, it is quite natural to consider the face averages of function v as the degrees of freedom, i.e. dene the local
reduction operator as

Ph ( v ) = v hP { v f }f P where v f =


v dS .

|f|

(55)

Thus, a discrete function v V becomes an ordered set of all face-based values, v h = { v f }fh . The dimension of V h equals
the number of mesh faces; the dimension of VPh equals the number of faces in P, see Fig. 1. Summarizing, we dene the
local bilinear form for every polynomial q and every discrete eld v hP through the formula:
h




AhP Ph (q), v hP :=
|f| v f nP,f KP q.

(56)

f P

This is the low order formulation of the consistency condition (12). Indeed, when v hP = Ph ( v ), the right-hand side of (56)
is equal to the left-hand side of (54), and we recover the original formula (12) for m = 1.
N and 
R required by formula (35), we evaluate (56) on the following polynomials:
To derive matrices 

q 1 = 1,

q2 =

x xP
hP

q3 =

y yP
hP

and q4 =

z zP
hP

which form a basis for P1 (P). Since the face averages of q1 are all equal to 1, it holds that N1 = Ph (q1 ) = 1. Furthermore,
the centroid rule yields

Ph (q2 )|f

|f|

x xP
hP

dS =

xf xP
hP

f P,

and similar relations holds for q3 and q4 . Using the local numbering of faces, fi for i = 1, 2, . . . , P , and the compact notation
(xfi xP )T = (xfi xP , y fi y P , zfi zP ) for the rows of matrix 
N, denition (29) yields

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

(xf1 xP ) T

hP
(xf2 xP ) T
hP


N = [1, N] =
..
.

..
.

(xf xP ) T

373

(57)

hP

As we noted in the previous section, the rst column of matrix R corresponding to q1 = 1 equals 0. To nd the other
columns, we evaluate (56) on polynomials q2 , q3 , and q4 . For q2 , we obtain

AhP

Ph (q2 ), v hP

 

1 
hP

|f| v f nP,f KP

f P

1
0
0

(58)

as (x xP ) = (1, 0, 0) T . Two similar relations are obtained for q3 and q4 using ( y y P ) = (0, 1, 0) T and ( z zP ) =
(0, 0, 1)T , respectively. Thus, matrix R is given by

|f1 |nPT ,f1 KP

T

1
0 |f2 |nP,f2 KP
R = [0, 
R] =
.
.
.
..

hP ..
T
0 |fP |nP,f KP

(59)

Remark 4.1. An alternative choice for the degrees of freedoms is to consider the values of the function v at the barycenters xf . In such a case, we substitute (55) with

Ph ( v ) = v hP { v f }f P where v f = v (xf ).
It is easy to observe that the resulting matrices R and N are not changed. Hence, this choice of the degrees of freedom
leads to the same scheme.
To complete the formulation of the low-order MFD method, we need to derive an approximation of the loading term.
We assume that f ,
h can be split into the sum of local terms f ,
hP such that

f , vh


Ph

v hP

where

Ph ( v ).

f , v
P =

f , v hP P ,

f , v hP

f v dV ,
P

Let f P be the cell average of f . Then, we consider the following approximation:

v dV f P IP v hP ,

f v dV f P

(60)

where IP ( v hP ) is a second-order accurate quadrature rule based on the degrees of freedom of P, see Appendix A. Using this,
we dene the local discrete loading term as

f , v hP

h
P

 
:= f P IP v hP v hP VPh .

Remark 4.2. Consider a tetrahedron P. By design, the discrete bilinear form is exact when both its arguments are linear
functions. The moment of a linear function q on face f of P equals its values at the barycenter xf and the latter are the
degrees of freedom in the CrouzierRaviart nite element method. Thus, both methods produce identical stiffness matrices
for a tetrahedral cell.
4.2. Constant diffusion coecient
Let K be constant in each cell P. In this case, the projection operator
the bilinear form AhP satises the consistency condition if



AhP Ph (q), Ph ( v ) =

m1 is the identity operator. Thus, we say that

K q v dV

q Pm (P), v H 1 (P),

where Pm (P) denotes the space of the polynomials of degree at most m on P. Now, let us integrate by parts the right-hand
side of this formula:

374

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385


K q v dV =

v div(K q) dV +

vnP,f K q dS .

(61)

f P f

We reformulate (61) by following the same steps of the general construction in Section 2.6.
Step (i ): divergence term. Since the diffusion tensor K is constant and q Pm (P), the divergence term div(K q) is a
polynomial of degree at most m 2 and can be expanded in the basis functions (15) with k such that |k |  m 2:

nm2

div(K q) =

c P,k qk ,

(62)

k =1

where c P,k are the expansion coecients. Using (16) and (62), we reformulate the volume integral in (61) as

nm2

v div(K q) dV =

k =1

qk v dV =

c P,k

nm2

c P,k |P| v P,k .

(63)

k =1

Step (ii ): face terms. Similarly, the trace of nP,f K q on f is a polynomial of degree at most m 1 and can be expanded
using the basis functions (17):

nm
1

(nP,f K q)|f =

c f,k rk ,

(64)

k =1

where c f,k are the expansion coecients. Using (18) and (64), we reformulate the face integrals in (61) as

nm
1

vnP,f K q dV =

k =1


nm
1


rf,k v dS =

c f,k

c f,k |f| v f,k .

(65)

k =1

Step (iii ): reformulation of (61). Now, let us insert (63) and (65) in (61):


nm
1

nm2

K q v dV =

c P,k |P| v P,k +

k =1

c f,k |f| v f,k ,

(66)

k =1

which holds for every q Pm (P) and every function v H 1 (P). Taking q = q j and v = qi for i , j = 1, . . . , nm , we obtain the
entries of the stiffness matrix with respect to the polynomial basis of Pm (P):

nm2

K q j qi dV =
P

k =1

j
c P,k |P|qi P,k


nm
1

c f,k |f|qi f,k ,

(67)

k =1
j

where we added the superscript j to the expansion coecients c P,k and c f,k to stress their dependence on q j . In the FEM,
this step often completes the derivation of the stiffness matrix while, in the MFD method, formula (67) gives only a part
of the stiffness matrix M. According to formula (35), the stiffness matrix M is given in terms of matrices N and R, see
formulas (30) and (32), respectively.
Remark 4.3. The denition of matrix N does not depend on the diffusion tensor K; therefore, its expression given by (30) is
the same for constant and variable K.
j

Remark 4.4. Matrix R does depend on K through the coecients c P,k and c f,k . Despite the different denitions of these
coecients, the nal expansions (66) and (27) are identical, which explains a single formula (32).
5. Numerical experiments
5.1. Convergence analysis
Let = ]0, 1[3 in problem (1)(2). We choose the boundary data g and the loading term f such that the exact solution is

u (x, y , z) = x3 y 2 z + x sin(2 xy ) sin(2 yz) sin(2 z).

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

375

We consider the constant permeability tensor K1 = I and the variable permeability tensor


K2 =

1 + y 2 + z2
xy
xz

xy
xz
1 + x2 + z2
yz
yz
1 + x2 + y 2


(68)

The Dirichlet boundary conditions are easily implemented in the code by directly imposing the value of all the degrees of
freedom that are associated with the boundary faces of the mesh.
h,
Let u h denote the discrete solution. For any cell P, let u P be the unique polynomial in Pm (P) such that





AP uhP, , q = AhP uhP , Ph (q) q Pm (P),
!"

1
h,
f P P,f u f for m = 1,
u P dV =
u P,1
for m > 1,
|P |

(69)
(70)

where

P,f are the weights built in Appendix A. The rst condition implies that the gradient of uhP, is determined by the
h,

degrees of freedom of u hP . The second condition implies that for m > 1 the cell average of u P on P coincides with the
approximate zero-th order moment of u hP . Then, we dene the post-processed solution as the piecewise polynomial funch,

tion u h, such that u h, |P = u P . Note that the post-processed solution is discontinuous across the cell interfaces. Once u h,
has been determined, we dene the relative error by calculating the L 2 norm or the H 1 semi-norm of u u h, and dividing
it by the corresponding norm of u.
5.1.1. Comparison with other low-order methods
The plots in Figs. 34 compare the new MFD method (m = 1) with the mixed MFD method [17], the nodal MFD
method [15] and the polyhedral nite element method (PFEM) [39] based on the three-dimensional barycentric coordinates
of [40]. These methods are applied on a family of random hexahedral meshes, whose cells are obtained by partitioning
tetrahedral cells. Mesh data are reported in Table 2; the rst two meshes in the family are displayed in the top row in
Fig. 2. A portion of each mesh close to the vertex (1, 1, 1) has been removed to show the interior structure.
The approximation errors for the PFEM and the nodal MFD method are dened by taking the L 2 -norm and the
H 1 -seminorm of the difference between the exact solution and the interpolation that is built using the three-dimensional
barycentric coordinates of [40]. The mixed MFD method is based on a different choice of degrees of freedom, i.e., uxes and
cell-averages of the scalar variable, which requires a different post-processing treatment. Therefore, the approximation error
for such a method is computed by using the discrete norms induced by the mimetic inner products for discrete uxes and
scalars, cf. [17].
In each gure, the two plots in the top row refer to calculations with K = K1 = I (constant permeability), the two
plots in the bottom row refer to calculations with K = K2 . We present the relative error curves with respect to the mesh
size parameter h and the number of non-zeroes of the stiffness matrix. The curves for PFEM, the mixed MFD, the nodal
MFD method, and the new MFD method are marked by circles, triangles, squares, and diamonds, respectively. For both
constant and variable diffusion tensors, these error curves indicate the second-order accuracy of such approximations, i.e.,
the quadratic convergence in the L 2 -norm and the linear convergence in the H 1 -seminorm. It is worth noting that for
variable K the approximation errors of the new MFD method are the smallest one on almost all the meshes in the L 2 norm
and comparable to those of the other methods in the energy norm.
5.1.2. Accuracy of the higher-order MFD schemes
We consider three MFD schemes corresponding to m = 1, 2 and 3. For m = 2, we use formula (39) to approximate the
loading term. The convergence analysis is performed on two different families of meshes. The rst family is the family of the
random hexahedral meshes described in the previous subsection. The second family is formed by a regular cubic partition
of with a local renement around the corner vertex (1, 1, 1). Mesh data are reported in Table 3; the rst two meshes are
displayed in the bottom row in Fig. 2.
In Figs. 58 we present the curves of the relative errors. In each gure, the two plots in the top row refer to calculations
with K = K1 = I, the two plots in the bottom row refer to calculations with K = K2 . In each plot, the low-order scheme
(m = 1) is marked by circles; the scheme based on quadratic polynomials (m = 2) is marked by squares; the scheme based
on cubic polynomials (m = 3) is marked by diamonds.
Let Nm denote the number of degrees of freedom in the scheme based on polynomials of degree m. Assuming that
the degrees of freedom are distributed uniformly over the computational domain, we have Nm h3 . Therefore, the
(m+1)/3
) when Nm . The H 1 -error is expected to decrease as
L 2 -error is expected to decrease as O (hm+1 ) O (Nm

m
/
3
O(hm ) O(Nh
) when Nm . The total number of non-zeroes in the stiffness matrix grows almost linearly with the
number of the degrees of freedom, so that we expect to observe a similar dependence of the errors. The numerical results
conrm these theoretical expectations, see Figs. 56 for the rst mesh family and Figs. 78 for the second mesh family.

376

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

Fig. 3. The family of random hexahedral meshes: relative L 2 errors for the linear PFEM (circles), the mixed MFD method (triangles), the nodal MFD method
(squares) and the new MFD method (diamonds) with m = 1. The top row corresponds to K = K1 = I; the bottom row corresponds to K = K2 . The expected
slopes are indicated by triangles.

5.2. Discontinuous permeabilities.

We investigate the behavior of the MFD schemes for m = 1, 2, 3 when the diffusion tensor is discontinuous. We solve
problem (1)(2) on the domain = 1 2 , where





1 = ]0, 1[ 0, 12 ]0, 1[ and 2 = ]0, 1[ 12 , 1 ]0, 1[,
with K1 = 1 I and K2 = 2 I. We set the Dirichlet boundary data g and the loading term f such that the exact solution is

n
on 1 ,
ax + by
+ cz + d

2 1 1 n
u (x, y , z) =
+
y + cz + d on 2 .
ax + b
n

2 2

Note that this solution is continuous and ux continuous across the interface separating 1 and 2 . The calculations are
carried out by taking a = b = c = d = 1, n = 4, 1 = 1 and 2 = 104 . We consider a sequence of random hexahedral
partitionings of that match the discontinuity of the diffusion tensor. Fig. 9 shows three different plots of the rst mesh of
the family. In these plots, part of the mesh has been removed to show the internal structure, and, in particular, the middle
panel shows the mesh lling exactly subdomain 1 . Mesh data are reported in Table 4.

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

377

Fig. 4. The family of random hexahedral meshes: relative H 1 errors for the linear PFEM (circles), the mixed MFD method (triangles), the nodal MFD method
(squares) and the new MFD method (diamonds) with m = 1. The top row corresponds to K = K1 = I; the bottom row corresponds to K = K2 . The expected
slopes are indicated by triangles.
Table 2
The family of random hexahedral meshes: NP , Ne , Nf , and Nv are the numbers of cells, faces, edges, and vertices, respectively; h is the mesh size
parameter; N1 , N2 , N3 are the numbers of degrees of freedom for m = 1, m = 2, and m = 3, respectively.
n

NP

Nf

Ne

Nv

N1

N2

0
1
2
3

176
888
11 444
94 320

600
2865
35 451
287 607

698
3153
37 495
299 613

275
1177
13 489
106 327

1.78E01
1.04E01
4.44E02
2.20E02

600
2865
35 451
287 607

1976
9483
117 797
957 141

N3
4304
20 742
258 482
2 102 922

Table 3
The family of locally rened meshes: NP , Ne , Nf , and Nv are the numbers of cells, faces, edges, and vertices, respectively; h is the mesh size parameter;
N1 , N2 , N3 are the numbers of degrees of freedom for m = 1, m = 2, and m = 3, respectively.
n

NP

Nf

Ne

Nv

N1

N2

0
1
2
3
4

15
120
960
7680
61 440

66
444
3216
24 384
189 696

96
546
3588
25 800
195 216

46
223
1333
9097
66 961

4.05E01
2.03E01
1.01E01
5.07E02
2.53E02

66
444
3216
24 384
189 696

213
1452
10 608
80 832
630 528

N3
456
3144
23 136
177 024
1 383 936

378

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

Fig. 5. The family of random hexahedral meshes: relative L 2 errors in the MFD schemes with m = 1 (circles), m = 2 (squares), and m = 3 (diamonds). The
top row corresponds to K = K1 = I; the bottom row corresponds to K = K2 . The expected slopes are indicated by triangles.

The plots of Fig. 10 show the relative errors for m = 1, m = 2 and m = 3 in the L 2 norm (top panels) and the
H -seminorm (bottom panels). Errors are computed using the post-processed solution u h, dened by (70), and are displayed versus the mesh size parameter h (left panels) and the number of non-zeroes of the stiffness matrix (right panels).
In each plot, the low-order scheme (m = 1) is marked by circles; the scheme based on quadratic polynomials (m = 2) is
marked by squares; the scheme based on cubic polynomials (m = 3) is marked by diamonds. For m = 2, we use formula (39)
to approximate the loading term. Despite the global low regularity of the exact solution, the MFD method for m = 1, 2, 3
shows the optimal convergence rates for both L 2 norm and the H 1 -seminorm.
1

6. Conclusions
In this work, we presented a new family of mimetic nite difference schemes for solving elliptic partial differential
equations in the primal form on unstructured polyhedral meshes. The degrees of freedom are the solution moments on
mesh faces and inside mesh cells. The schemes are built from the local consistency and stability conditions. Higher-order
mimetic schemes use the consistency condition that holds for polynomials of degree m > 1. Such schemes show an (m + 1)
convergence rate in the L 2 norm on unstructured polyhedral meshes. The accuracy of new discretizations is investigated
numerically for m = 1, m = 2, and m = 3 on diffusion problems with constant, spatially variable, and discontinuous permeability tensors.

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

379

Fig. 6. The family of random hexahedral meshes: relative errors in the H 1 semi-norm in the MFD schemes with m = 1 (circles), m = 2 (squares), and m = 3
(diamonds). The top row corresponds to K = K1 = I; the bottom row corresponds to K = K2 . The expected slopes are indicated by triangles.

Acknowledgements
The authors gratefully thank the anonymous reviewers for many useful suggestions on how to improve the presentation
of this paper. This work was partially supported by the National Nuclear Security Administration of the U.S. Department of
Energy at Los Alamos National Laboratory under Contract No. DE-AC52-06NA25396 and the DOE Oce of Science Advanced
Scientic Computing Research (ASCR) Program in Applied Mathematics.
Appendix A. A second-order accurate quadrature rule based on face averages
Let the mesh satises assumptions (MR1)(MR3). We will rst show that there exists a set of uniformly bounded coecients {P,f }f P such that

xP =

f P

P,f xf ,

P,f = 1, P,f > 0.

(A.1)

f P

Then, we will prove that (A.1) is enough to ensure the second order of accuracy to the quadrature rule IP ( v hP ) in (60) and
built using the pairs of nodes and weights {(xf , P,f )}f P . The weights P,f can also be used to evaluate the cell average of
the solution in post-processing (70).

380

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

Fig. 7. The family of cubic meshes with local renement: relative L 2 errors in the MFD schemes with m = 1 (circles), m = 2 (squares), and m = 3 (diamonds).
The top row corresponds to K = K1 = I; the bottom row corresponds to K = K2 . The expected slopes are indicated by triangles.

Let Pf be the pyramid with apex xP , the barycenter of P, and base f. Furthermore, let xPf denote the barycenter of Pf
and |Pf | its volume. The barycenter of Pf is a convex combination of xP and xf , i.e. xPf = x
f + (1 )xP where
" = 1/4.
Due to the mesh assumptions (MR), the set {Pf }f P is a conformal partition of cell P, i.e. P = f P Pf and |P| = f P |Pf |.
Thus, xP is the weighted average of the barycenters xPf :

|P |x P =

|Pf |xPf =

f P




|Pf | xf + (1 )xP =
|Pf |xf + (1 )|P|xP ,

f P

f P

from which it follows that

xP =

 |P f |
xf .
|P |

f P

By comparison with (A.1), we deduce that we can take P,f = |Pf |/|P| and (A.1) is a trivial consequence of such denition.
Moreover, assumptions (MR) imply that |Pf |  C |P| where C depends only on the mesh regularity constants N and ,
which implies the uniform boundedness of the weights P,f .
Now, let v be a function dened on cell P. We assume that v is suciently smooth for all the developments that follow.
Let us consider the Taylor expansion of v around xf , the barycenter of face f:

v (x) = v (xf ) + v (xf ) (x xf ) + O h2P .

(A.2)

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

381

Fig. 8. The family of cubic meshes with local renement: relative H 1 errors in the MFD schemes with m = 1 (circles), m = 2 (squares), and m = 3 (diamonds). The top row corresponds to K = K1 = I; the bottom row corresponds to K = K2 . The expected slopes are indicated by triangles.

Fig. 9. Discontinuous permeability problem: the rst random hexahedral mesh. Part of the mesh is removed in the three panels to show the internal
structure. The panel in the middle shows that the mesh lls subdomain 1 .

Its cell-average on P is given by

|P |

v dV = v (xf ) + v (xf ) (xP xf ) + O h2P .

P

Let us expand v (xf ) in (A.3) around xP :

(A.3)

382

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

Table 4
The family of meshes used in the discontinuous permeability test case: NP , Ne , Nf , and Nv are the numbers of cells, faces, edges, and vertices, respectively;
h is the mesh size parameter; N1 , N2 , N3 are the numbers of degrees of freedom for m = 1, m = 2, and m = 3, respectively.
n

NP

0
1
2
3
4
5

384
3072
24 576
82 944
196 608
384 000

Nf
1224
9504
74 880
251 424
594 432
1 159 200

Ne
1340
10 072
78 128
261 000
615 520
1 198 520

Nv

501
3641
27 825
92 521
217 697
423 321

0.13758
0.0687902
0.0343951
0.0229301
0.0171975
0.013758

N1
1224
9504
74 880
251 424
594 432
1 159 200

N2
4056
31 584
249 216
837 216
1 979 904
3 861 600

N3
8880
69 312
547 584
1 840 320
4 353 024

Fig. 10. Double permeability test: the top row corresponds to relative L 2 errors, the bottom row corresponds to relative H1 errors. The MFD schemes
corresponds to m = 1 (circles), m = 2 (squares), and m = 3 (diamonds). The slopes of the optimal convergence rates are indicated by triangles.

 
v (xf ) = v (xP ) + H( v ; xf ) (xf xP ) + O h2P ,

(A.4)

where H( v ; xf ) is the Hessian of v at xf . Inserting this expansion in (A.3) and keeping only the linear terms, we obtain:

|P |

v dV = v (xf ) + v (xP ) (xP xf ) + O h2P .

P

(A.5)

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

Note that v (xf ) = v f + O (h2P ). Then, we multiply both sides of (A.5) by

properties of the weights listed in (A.1):

|P |


v dV =

P,f v (xf ) + v (xP ) xP

f P

Eq. (A.6) implies that we can take IP ( v hP ) =

"

f P

P,f , sum over all faces f of P, and apply the



P,f xf + O h2P =

f P

383

P,f v f + O h2P .

(A.6)

f P

P,f v f as second-order accurate quadrature rule.

Appendix B. Proof of Lemma 3.1

Since 
RT 
N is a symmetric positive denite matrix, for any v h = 
N w h img(
N) we have

vh

T

M0 v h = w h

T

 T T

N T M0 
Nwh = wh 
N 
R w h  0,

and the last equality is achieved only when w h = 0, i.e. v h = 0. Thus, the Rayleighs quotient of M0 satises the inequalities

0 < +
M0 
min

 
( v h ) T M0 v h
 max M0 v h img(
N)\{0},
h
2
v 

(B.1)

where +
(M0 ) and max (M0 ) denote the minimum positive eigenvalue and the maximum eigenvalue of M0 , respectively.
min
Let us introduce the auxiliary quantity

 =

+
1 min (M0 )

N max (M0 )

Now, (MR1) implies that

 P 

(B.2)

P = trace(M0 )  N max (M0 ). A straightforward calculation and the left inequality in (B.1) yield:

+
1 min (M0 )

N max

(M0 )

N max M0 = +
M0 
min

( v h ) T M0 v h
 v h 2

v h img(
N)\{0}.

(B.3)

 satises lemmas inequality (47), but cannot be chosen as as it strongly depends on P through the
The quantity
eigenvalues of M0 . Nonetheless, (B.3) suggests that we can prove the lemma if we remove such dependence by taking the
 on all admissible P in accordance with the shape-regularity conditions (MR). Of course, we have also to
minimum of
 , which
show that such minimum exists and is strictly positive. This strategy will lead us to determine a quantity 
only depends on the shape-regularity constants an N .
To this end, let us"rst note that each vector w h corresponds to a polynomial function w of degree m through the
nm
correspondence w =
q w hj 
Pm (P), where 
Pm (P) = Pm (P)\span{q1 }. According to (40) and (42), we have:
j =2 j

vh

T

M0 v h = w h

T


RT 
N w h = AhP ( w , w )  0

(B.4)

and the equality is achieved only when w = 0.

To remove the dependence of P on its size, we remap the polyhedron P of diameter hP to a polyhedron 
P of diameter
1
h

h
=
1 through the homothetic transformation x := h
P
P x. Under this transformation, the bilinear form AP gains a factor h P .
Instead, the triple-bar norm | w |P 
N w h  =  v h  remains unchanged, since the corresponding basis functions q j and qj
h
dened on P and 
P are both properly scaled by the size of such cells. The following relation holds between AhP and A
and
P

between the triple-bar norms on P and 

P:
h
AhP ( w , w ) = hP A
(
w, 
w ),
P

| w |P = |
w |
P.

(B.5)

Let SP = { w 
Pm (P): | w |P = 1}. Relations (B.5) implies that

 0
 0
h
 = hP min A
+
M = min AhP ( w , w ) = hP
M
(
w, 
w ),
+
min
min
P
w SP


w S
P

where all the quantities with the hat accent refer to the transformed polyhedron 
P.
The transformed polyhedron 
P is uniquely dened by the position of its vertices and how these vertices are connected,
i.e., by its connectivity map M. The concept of connectivity maps somehow generalizes the concept of the reference cell
that is used in the nite element analysis. In fact, each connectivity map can be considered as a family of reference unitsize polyhedra 
P, where each member of the family may differ only for the position of the vertices forming 
P under the

constraint that h
=
1,
but
not
for
the
number
of
the
vertices
forming
P
and
the
way
they
are
connected.
The
crucial
point
P
of this argument is that the shape regularity assumption (MR1) implies that the number of vertices forming an admissible polyhedron 
P is uniformly bounded. Therefore, there exists only a nite set SM of connectivity maps, i.e., of possible
families of such polyhedra.

384

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

Let us x a connectivity map M, which in particular means that we consider a subset of polyhedra 
P with a xed
number of vertices N. We dene a vector X = X (
P) R3N formed by the coordinates of the vertices of 
P. Let SM,X
indicate the set of all admissible vectors X (
P), each vector corresponding to an admissible unit-size polyhedron 
P with
connectivity map M. Clearly, SM,X is a nite-dimensional set. It is also bounded as the entries of each X (
P) must be
such that h
P = 1. Moreover, the shape-regularity condition (MR2) indicates that any Cauchy sequence of vectors in SM,X
converges to a vector in this set; hence, SM,X is closed. A proof of the completeness property is by contradiction; more
details can be found in [8]. We conclude that SM,X is a compact set.
h
0
on the polynomial functions 
w S
The action of the bilinear form A
P can be represented by the matrix M acting
P

0 are a continuous function of the

w h corresponding to such polynomials. The coecients of matrix M
on the vectors 
h
+
coordinates of the vertices forming 
P. Thus, A
is continuous on SM,X , and so is its eigenvalue 
, which must achieve
min
P
its minimum on this set. Let us call


= min 
+
+
min,M
min
SM,X

h
+
such minimum. The quantity 
is actually achieved by A
(
w, 
w ) for some polynomial function 
w with |
w |
P =1
min,M
P
that is dened on some polyhedral cell 
P corresponding to an element of SM,X . Inequality (B.4) (see also the following
+
+
comment) implies that 
is strictly positive. The quantity 
still depends on the choice of the connectivity
min,M
min,M
map M, as indicated by its subscript. Let us now consider the nite set SM of all possible connectivity maps M and
+
+
the set of corresponding strictly positive 
. The minimum of all corresponding 
must be a positive number
min,M
min,M

+
that depends only on the shape regularity constants of the transformed cell but not on its geometric shape. Let 
=
min
+
min,M denote such minimum.
minMSM 

1
0
0 ) = h
max that provides an upper bound for 
max (M
Using the same strategy, we can dene a quantity 
P max (M ).
Eventually, we set

+
1 
min

N 
max

 , from which the result of the lemma follows.

and note that 0 < 

References
[1] P. Antonietti, L. Beiro da Veiga, N. Bigoni, M. Verani, Mimetic nite differences for nonlinear and control problems, Math. Models Methods Appl. Sci.
(2014), http://dx.doi.org/10.1142/S0218202514400016.
[2] L. Beiro da Veiga, A residual based error estimator for the mimetic nite difference method, Numer. Math. 108 (3) (2008) 387406.
[3] L. Beiro da Veiga, A mimetic discretization method for linear elasticity, Math. Model. Numer. Anal. 44 (2) (2010) 231250.
[4] L. Beiro da Veiga, F. Brezzi, A. Cangiani, G. Manzini, L.D. Marini, A. Russo, Basic principles of virtual element methods, Math. Models Methods Appl.
Sci. 23 (2013) 119214.
[5] L. Beiro da Veiga, F. Brezzi, L.D. Marini, Virtual elements for linear elasticity problems, SIAM J. Numer. Anal. 51 (2013) 794812.
[6] L. Beiro da Veiga, J. Droniou, G. Manzini, A unied approach to handle convection term in nite volumes and mimetic discretization methods for
elliptic problems, IMA J. Numer. Anal. 31 (4) (2011) 13571401.
[7] L. Beiro da Veiga, V. Gyrya, K. Lipnikov, G. Manzini, Mimetic nite difference method for the Stokes problem on polygonal meshes, J. Comput. Phys.
228 (19) (2009) 72157232.
[8] L. Beiro da Veiga, K. Lipnikov, A mimetic discretization of the Stokes problem with selected edge bubbles, SIAM J. Sci. Comput. 32 (2) (2010) 875893.
[9] L. Beiro da Veiga, K. Lipnikov, G. Manzini, Arbitrary-order nodal mimetic discretizations of elliptic problems on polygonal meshes, SIAM J. Numer.
Anal. 49 (5) (2011) 17371760.
[10] L. Beiro da Veiga, K. Lipnikov, G. Manzini, Error analysis for a mimetic discretization of the steady Stokes problem on polyhedral meshes, SIAM J.
Numer. Anal. 48 (2011) 14191443.
[11] L. Beiro da Veiga, K. Lipnikov, G. Manzini, The Mimetic Finite Difference Method, 1st edition, Modeling, Simulation and Applications, vol. 11, SpringerVerlag, New York, 2013.
[12] L. Beiro da Veiga, G. Manzini, An a posteriori error estimator for the mimetic nite difference approximation of elliptic problems, Int. J. Numer.
Methods Eng. 76 (11) (2008) 16961723.
[13] L. Beiro da Veiga, D. Mora, A mimetic discretization of the ReissnerMindlin plate bending problem, Numer. Math. 117 (3) (2011) 425462.
[14] P. Bochev, J.M. Hyman, Principle of mimetic discretizations of differential operators, in: D. Arnold, P. Bochev, R. Lehoucq, R. Nicolaides, M. Shashkov
(Eds.), Compatible Discretizations. Proceedings of IMA Hot Topics Workshop on Compatible Discretizations, in: IMA Vol. Math. Appl., vol. 142, SpringerVerlag, 2006, pp. 89120.
[15] F. Brezzi, A. Buffa, K. Lipnikov, Mimetic nite differences for elliptic problems, M2AN Math. Model. Numer. Anal. 43 (2) (2009) 277295.
[16] F. Brezzi, A. Buffa, G. Manzini, Mimetic inner products for discrete differential forms, J. Comput. Phys. 257 (2014) 12281259, Part B.
[17] F. Brezzi, K. Lipnikov, M. Shashkov, Convergence of the mimetic nite difference method for diffusion problems on polyhedral meshes, SIAM J. Numer.
Anal. 43 (5) (2005) 18721896.
[18] J. Campbell, M. Shashkov, A tensor articial viscosity using a mimetic nite difference algorithm, J. Comput. Phys. 172 (2001) 739765.
[19] A. Cangiani, F. Gardini, G. Manzini, Convergence of the mimetic nite difference method for eigenvalue problems in mixed form, Comput. Methods
Appl. Mech. Eng. 200 (912) (2011) 11501160.
[20] A. Cangiani, G. Manzini, Flux reconstruction and pressure post-processing in mimetic nite difference methods, Comput. Methods Appl. Mech. Eng.
197 (912) (2008) 933945.
[21] A. Cangiani, G. Manzini, A. Russo, Convergence analysis of the mimetic nite difference method for elliptic problems, SIAM J. Numer. Anal. 47 (4)
(2009) 26122637.

K. Lipnikov, G. Manzini / Journal of Computational Physics 272 (2014) 360385

385

[22] J. Droniou, Finite volume schemes for diffusion equations: introduction to and review of modern methods, Math. Models Methods Appl. Sci. (2014).
[23] P. Grisvard, Elliptic Problems in Nonsmooth Domains, Monographs and Studies in Mathematics, vol. 24, Pitman, Boston, 1985.
[24] V. Gyrya, K. Lipnikov, G. Manzini, D. Svyatskiy, M-adaptation in mimetic nite difference method, Math. Models Methods Appl. Sci. (2014), http://
dx.doi.org/10.1142/S0218202514400053.
[25] J. Hyman, M. Shashkov, Mimetic discretizations for Maxwells equations and the equations of magnetic diffusion, Prog. Electromagn. Res. 32 (2001)
89121.
[26] J.M. Hyman, M.J. Shashkov, S. Steinberg, The numerical solution of diffusion problems in strongly heterogeneous non-isotropic materials, J. Comput.
Phys. 132 (1) (1997) 130148.
[27] K. Lipnikov, G. Manzini, F. Brezzi, A. Buffa, The mimetic nite difference method for 3D magnetostatics elds problems, J. Comput. Phys. 230 (2) (2011)
305328.
[28] K. Lipnikov, G. Manzini, M. Shashkov, Mimetic nite difference method, J. Comput. Phys. 257 (Part B) (2014) 11631227.
[29] K. Lipnikov, G. Manzini, D. Svyatskiy, Analysis of the monotonicity conditions in the mimetic nite difference method for elliptic problems, J. Comput.
Phys. 230 (7) (2011) 26202642.
[30] K. Lipnikov, J. Morel, M. Shashkov, Mimetic nite difference methods for diffusion equations on non-orthogonal non-conformal meshes, J. Comput.
Phys. 199 (2) (2004) 589597.
[31] K. Lipnikov, J.D. Moulton, D. Svyatskiy, A Multilevel Multiscale Mimetic (M3 ) method for two-phase ows in porous media, J. Comput. Phys. 227 (2008)
67276753.
[32] G. Manzini, A. Russo, N. Sukumar, New perspectives on polygonal and polyhedral nite element methods, Math. Models Methods Appl. Sci. (2014),
http://dx.doi.org/10.1142/S0218202514400065.
[33] L. Margolin, M. Shashkov, P. Smolarkiewicz, A discrete operator calculus for nite difference approximations, Comput. Methods Appl. Mech. Eng. 187
(2000) 365383.
[34] A. Palha, P.P. Rebelo, R. Hiemstra, J. Kreeft, M. Gerritsma, Physics-compatible discretization techniques on single and dual grids, with application to the
Poisson equation of volume forms, J. Comput. Phys. 257 (Part B) (2014) 13941422.
[35] N. Sukumar, E. Malsch, Recent advances in the construction of polygonal nite element interpolants, Arch. Comput. Methods Eng. 13 (1) (2006)
129163, http://dx.doi.org/10.1007/BF02905933.
[36] N. Sukumar, A. Tabarraei, Conforming polygonal nite elements, Int. J. Numer. Methods Eng. 61 (2004) 20452066.
[37] C. Talischi, G.H. Paulino, Addressing integration error for polygonal nite element through polynomial projections: a patch test connection, Math.
Models Methods Appl. Sci. (2014), http://dx.doi.org/10.1142/S0218202514400077.
[38] C. Talischi, G.H. Paulino, C.H. Le, Honeycomb Wachspress nite elements for structural topology optimization, Struct. Multidiscip. Optim. 37 (6) (2009)
569583.
[39] E. Wachspress, A Rational Finite Element Basis, Academic Press, 1975.
[40] J.D. Warren, S. Schaefer, A.N. Hirani, M. Desbrun, Barycentric coordinates for convex sets, Adv. Comput. Math. 27 (3) (2007) 319338.