Você está na página 1de 41

Arbitrary-order nodal mimetic discretizations of

elliptic problems on polygonal meshes


Gianmarco Manzini
Los Alamos National Laboratory
New Mexico, USA
Istituto di Matematica Applicata e Tecnologie Informatiche
IMATI-CNR, Pavia, Italy

in collaboration with L. Beirao da Veiga and K. Lipnikov

Workshop Bicocca September 2012


Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

1 / 29

Outline

1 Low- and high-order formulation of the method:


-

meshes;
degrees of freedom;
approximation of the bilinear form;
approximation of the loading term.

2. The stability condition.


3. Convergence results and numerical experiments.
4. Conclusions.

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

2 / 29

The linear diffusion problem

Differential formulation:
div(Ku) = f

in ,

u = g on ,
Variational formulation:
Find u Hg1 () such that:
Z

Z
Ku v dV =

Manzini, G. (LANL & IMATI-CNR)

fv dV

High order Nodal MFD Methods

v H01 (),

Bicocca 2012

3 / 29

Scheme construction in five steps


Steps 1 and 2

1. We decompose into a mesh h of polygons (2-D) or polyhedrons


(3-D);
- admissible meshes may contain non-convex cells and singular
(AMR) cells;
- we need some regularity assumptions to avoid pathological cases
and perform the convergence analysis;
2. degrees of freedom: Vh
u, v Hg1 () C ()

uh , vh Vh ,

with 0;

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

4 / 29

Scheme construction in five steps


Steps 3 and 4


3. bilinear form: Ah , : Vh Vh R
Z

Ku v dV ,
Ah uh , vh

it is built by mimicking a fundamental relation of calculus


(integration by parts);

4. linear functional: f , h : Vh R
f , vh

Manzini, G. (LANL & IMATI-CNR)


h

fv dV .

High order Nodal MFD Methods

Bicocca 2012

5 / 29

MFD construction in five steps


Step 5

5. The variational formulation


Find u Hg1 () such that:
Z

Z
Ku v dV =

fv dV

v H01 (),

becomes the mimetic variational formulation:


Find uh Vh,g such that:


Ah uh , vh = f , vh h

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

vh Vh,0 .

Bicocca 2012

6 / 29

Notation and References


0: regularity index, C -regular approximations;
m + 1: consistency index, degree of the polynomials of the
consistency condition.
low-order method, = 0, m = 1, 2-D, 3-D,
Brezzi, Buffa, Lipnikov (M2AN, 2009);
high-order methods, m 1, 2-D
= 0, MFD, Beirao da Veiga, Lipnikov, M. (SINUM, 2011);
and VEM reformulation, Brezzi et. al., (M3AS, 2013);
0, MFD & VEM for highly regular approximations, Beirao da
Veiga, M., (submitted).

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

7 / 29

1. Meshes: basic assumptions


There exists a conforming sub-decomposition Sh of h in triangles (2-D)
such that
1. for any polygonal cell P h , the restriction Sh,P is formed by a
uniformly bounded number of triangles T;
2. the triangles T Sh are shape-regular (Ciarlet);
3. Convergence analysis of the mimetic high-order setting is more
restrictive: 1. and 2. must hold but
- any polygonal cell P h is star-shaped with respect to an
internal point xP ;
- Sh,P is given by connecting xP to the cell vertices.
We do not need to build Sh in practice!
Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

8 / 29

Meshes: 2-D examples


Examples: convex and non-convex polygonal cells

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

9 / 29

Meshes: 2-D examples


Examples: randomized quads and Adaptive Mesh Refinements (AMR)

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

10 / 29

2. Degrees of freedom
C 0 -regular approximations

A discrete scalar field vh in Vh is given by:

m=1

low-order setting: = 0, m = 1, 2-D, 3-D:


- one real number per mesh vertex v;
=0, m=2

high-order setting: = 0, m > 1, 2-D:


- one real number per mesh vertex v;
- (m 1) real numbers per mesh edge e;
- m(m 1)/2 real numbers per mesh cell P;

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

11 / 29

Degrees of freedom
Approximations with high regularity

A discrete scalar field vh in Vh is given by:


high-order setting: = 2, m = 4:

=2, m=4

- vertex dofs solution and derivatives at


each vertex;
- cell dofs solution moments inside the
cells;
- edge dofs solution and normal
derivatives along the edges;

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

12 / 29

3. Construction of Ah uh , vh


Ah uh , vh must be
symmetric, bounded and semi-positive;
locally defined through an assembly process (like FEM):
 X

Ah,P uh,P , vh,P
Ah uh , vh =
P

where uh,P = uh|P , vh,P = vh|P ;



Any Ah,P uh,P , vh,P must be a local approximation:
Z

P h : Ah,P uh,P , vh,P
Ku v dV .
P

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

13 / 29

Towards a local consistency condition


The low-order setting, = 0, m = 1

Let K be constant on P.
1. We integrate by parts on every mesh cell P.
If u is a linear polynomial on P:
Z
Z
Z
Ku v dV = div(Ku) v dV +
Ku nP vds;
P
P
|P
{z
}
equal to zero!

2. we introduce a numerical integration rule for each edge e = (v0 , v00 ) that
is exact for linear polynomials:
Z
X
X
v (xv0 ) + v (xv00 )
vds
Ku nP,e |e|
.
Ku nP,e
| {z } e
|
{z2
}
eP
eP
constant
trapezoidal rule

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

14 / 29

The local consistency condition


The low-order setting, = 0, m = 1

HENCE,
Z
Ku v dV

P h :
P

Ku nP,e |e|

eP

v (xv0 ) + v (xv00 )
2

for any linear polynomial u and v enough regular.


THUS, WE REQUIRE THAT:
the degrees of freedom approximate the vertex values:
vh|v := vv v (xv0 );
for every linear polynomial u and every discrete field vh Vh the
bilinear form satisfies
X


|e|
vv0 + vv00 .
Ah,P vh , u I :=
Ku nP,e
2
eP

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

15 / 29

Towards a local consistency condition


General case: 0, m + 1

Let K be constant on P and integrate by parts on cell P:


Z
Z
XZ
Ku v dV = div(Ku) v dV +
Ku nP,e v ds.
{z }
| {z }
P|
P
ee e
not zero!
not constant!
If u is a polynomial of degree m on P:
div(Ku) is a polynomial of degree m 2;
Ku nP,e is a polynomial of degree m 1;

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

16 / 29

Divergence term
Internal degrees of freedom, 0, m + 1

1. We use the moments of v to express the integral over P:


if

div(Ku) = a0 1 + a1 x + a2 y + . . . Pm2 (P)

then
Z

div(Ku) v dV = a0
P

Z
Z
1v dV +a1 xv dV +a2 y v dV + . . .
| P {z }
| P {z }
| P {z }
vP,0

vP,1,x

vP,1,y

= a0 v P,0 + a1 v P,1,x + a2 v P,1,y + . . .


This choice suggests us to define
- m(m 1)/2 internal degrees of freedom v P,0 , v P,1,x , v P,1,y , . . .
Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

17 / 29

Edge integrals
Nodal degrees of freedom, = 0, m > 1

2. Let K|P constant, u Pm (P). Then,




we take v (xi ) , i = 0, . . . , m, along e = (x0 , xm );
we interpolate these values ve Pm (e);
we integrate the polynomial nP,e Ku ve along the edge using the
quadrature rule {(xe,q , we,q )}
Z
nP,e Ku(s) v (s)ds
e

m+1
X

we,q Ku(xe,q ) nP,e ve (xe,q ).

q=0

This choice suggests us to define:


- one degree of freedom per vertex,
ve,0 = vv0 v (xv0 ), ve,m+1 = vv00 v (xv00 );
- (m 1) nodal degrees of freedom per edge of P,
ve,q v (xe,q ) for q = 1, 2, . . . m 1.
Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

18 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 0 regularity

= 0, m = 1

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 0 regularity

= 0, m = 2

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 0 regularity

= 0, m = 3

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 0 regularity

= 0, m = 4

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 1 regularity

= 1, m = 2

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 1 regularity

= 1, m = 3

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 1 regularity

= 1, m = 4

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 1 regularity

= 1, m = 5

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 2 regularity

= 2, m = 3

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 2 regularity

= 2, m = 4

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 2 regularity

= 2, m = 5

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

Edge integrals
General case: 0, m + 1

2. We interpolate v along e using all the available degrees of freedom


(Hermite-like interpolation) and integrate using ve
Z
Ku nP,e v ds Ie (u; ve ; nP,e )
e

Example: C 2 regularity

= 2, m = 6

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

19 / 29

The local consistency condition


General case: 0, m + 1

RECALL THAT:
 X

Ah,P uh,P , vh,P .
Ah uh , vh =
P

WE REQUIRE:
1. local consistency: for every u Pm (P) (m 1) and every discrete field
vh,P Vh there holds:
m(m1)/2
X
X

Ah,P vh,P , u I :=
aj vP,j +
Ie (u; ve ; nP,e ).
j=0

|
{z
}
divergence

Manzini, G. (LANL & IMATI-CNR)

eP

|
{z
}
boundary, ( 0)

High order Nodal MFD Methods

Bicocca 2012

20 / 29

The stability condition


For both low and high-order methods we require:
2. stability, there exist two positive constants and such that for every
vh,P Vh,P there holds:

||vh,P ||21,h,P Ah,P vh,P , vh,P ||vh,P ||21,h,P ;
where the mesh-dependent norm || ||1,h,P mimics | |1,P .
We modify the bilinear form Ah,P by adding a stabilization term Sh,P
that depends on a set of parameters family of schemes!
The stabilization term Sh,P has the same form of that used by the
VEM.

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

21 / 29

4. The linear functional (f , vh )h


General case: 0, m + 1

WE REQUIRE THAT
X


f , vh h :=
f , vh h,P where

f , vh


h,P

fv dV .
P

For m = 1 we use a numerical integration formula over P based on the


vertex values.
For m > 1 we consider the orthogonal projection of f onto the
polynomials of degree m 2:
f = b0 1 + b1 x + b2 y + . . . Pm2 (P)
and use the moments of v to express the r.h.s. integral:
Z
Z
Z
Z
f v dV b0 1v dV +b1 xv dV +b2 y v dV + . . .
P
| P {z }
| P {z }
| P {z }
vP,0

vP,1,x

vP,1,y

= b0 v P,0 + b1 v P,1,x + b2 v P,1,y + . . .


Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

22 / 29

A mesh-dependent norm
We consider the mesh-dependent norm
X
||vh ||21,h =
||vh ||21,h,P
Ph

that mimics the | |1, semi-norm;


for the low-order method (m = 1, d = 2, 3), e = (v0 , v00 ) being an edge,
X
2
|vv00 vv0 | ;
||vh ||21,h,P = ||GRADh (vh )||2h,P = hP
eP

for the high-order method (m > 1, d = 2), e = (v0 , v00 ) being an edge,
||vh ||21,h,P = hP

h
i
X vh,f 2

2 + moments
s L (e)

eP

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

23 / 29

Convergence results
The consistency and the stability conditions allow us to determine a family
of mimetic schemes:
for the low-order method = 0, m = 1:

||u I uh ||1,h < Ch |f |0, + |u|1, + |u|2, ;
(Brezzi, Buffa, Lipnikov, M2AN (2009)),

for the high-order method = 0, m > 1:


||u I uh ||1,h < Chm ||u||m+1, ;
(Beirao da Veiga, Lipnikov, M., SINUM (2011); VEM, Brezzi et. al. M3AS (to
appear)

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

24 / 29

Numerical experiments
Meshes with non-convex polygons

Meshes:

Exact solution: u(x, y) = e2y sin(2x)


Diffusion tensor


1 0
K=
0 1
Manzini, G. (LANL & IMATI-CNR)


and

K(x, y) =

(x + 1)2 + y 2
xy

High order Nodal MFD Methods

xy
(x + 1)2

Bicocca 2012

25 / 29

Results
= 0, non-convex polygons, || ||1,h errors, constant K

n
0
1
2
3

h
1.458 101
7.289 102
3.644 102
1.822 102

m=2
Error
2.858
7.867 101
2.049 101
5.289 102

Rate

1.86
1.94
1.95

m=3
Error
1.007
2.819 101
5.597 102
8.897 103

Rate

1.84
2.33
2.65

n
0
1
2
3

h
1.458 101
7.289 102
3.644 102
1.822 102

m=4
Error
1.943 101
1.276 102
7.075 104
3.950 105

Rate

3.93
4.17
4.16

m=5
Error
2.282 102
1.128 103
4.406 105

Rate

4.34
4.68

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

26 / 29

Results
= 0, non-convex polygons, || ||1,h errors, non-constant K

n
0
1
2
3

h
1.458 101
7.289 102
3.644 102
1.822 102

m=2
Error
3.007
8.081 101
2.071 101
5.303 102

Rate

1.89
1.96
1.97

m=3
Error
9.873 101
2.760 101
5.621 102
9.083 103

Rate

1.84
2.29
2.63

n
0
1
2
3

h
1.458 101
7.289 102
3.644 102
1.822 102

m=4
Error
2.059 101
1.367 102
7.562 104
4.210 105

Rate

3.92
4.18
4.17

m=5
Error
1.988 102
1.016 103
3.924 105

Rate

4.29
4.69

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

27 / 29

Results
= 1, 2; non-convex polygons, || ||1,h errors, constant K

n
0
1
2
3

h
1.458 101
7.289 102
3.644 102
1.822 102

Manzini, G. (LANL & IMATI-CNR)

= 1, m = 2
Error
Rate
8.901 102

1.983 102 2.26


4.815 103 2.08
1.198 103
2.03

High order Nodal MFD Methods

= 2, m = 3
Error
Rate
1.054 102

4.543 104 4.72


4.663 105 3.36
5.528 106
3.11

Bicocca 2012

28 / 29

Conclusions
We are developing a new family of nodal mimetic finite difference
methods such that:
(i) the low-order formulation, using vertex values to represent linear
polynomials, works in 2-D (and in 3-D);
(ii) an high-order formulation, using nodal values, normal derivatives
along edges and moments to represent m-degree polynomials,
works in 2-D.

Possible developments:
(i)
(ii)
(iii)
(iv)

extension to 3-D;
new formulations using moments also for face and edge terms;
p-refinements;
...

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

29 / 29

Conclusions
We are developing a new family of nodal mimetic finite difference
methods such that:
(i) the low-order formulation, using vertex values to represent linear
polynomials, works in 2-D (and in 3-D);
(ii) an high-order formulation, using nodal values, normal derivatives
along edges and moments to represent m-degree polynomials,
works in 2-D.

Possible developments:
(i)
(ii)
(iii)
(iv)

extension to 3-D;
new formulations using moments also for face and edge terms;
p-refinements;
...

Manzini, G. (LANL & IMATI-CNR)

High order Nodal MFD Methods

Bicocca 2012

29 / 29

Você também pode gostar