Silvia Bertoluzza
FEF2017, Rome
Model problem:
u = f , in , u = 0 on
Variational form
Z
u H01 (), a(u, v ) := u v = hf , v i v H01 ()
Polygonal tessellation
Assumptions
Quasi-uniformity
h: for any two vertices A, B of an element K , |A B| ' h
VEM discretization of order k
Au = f
For VEM proof requires the construction of one stable lifting, which need the construction of a
Scott-Zhang type operator [B., Pennacchio 15]
Interface problem
w Wh = Vh | , s(w , z) = hg , zi, z Wh
with X
sh (w , z) = ah` (L`h w , L`h z), hg , zi = hf , L`h zi
`
Proposition
There exist constant , such that
X X
|w |2H 1/2 (` ) sh (w , w ) |w |2H 1/2 (` )
` `
u W : s(u, v ) = hf , v i v W .
Linear operators
S : W W 0 , hSu, v i = s(u, v ) v W
S : W W 0 , hSu, v i = s(u, v ) v W
P : W W , P 2 = P, range P = W
B : W , null B = W , range B = = W /W
BDT : W , BBDT = I
R : W W natural injection
Assumption
B, BDT and P satisfy
BDT B + P = IW
S. Bertoluzza (IMATI-CNR) FETI-DP and BDDC for VEM FEF2017 - Rome 8 / 24
BDDC
Solve
Su = f
MBDDC = P S 1 P T
Su + B T = P T f , Bu = 0
Eliminate u = S 1 (P T f B T ) equation in
B S 1 B T = B S 1 P T f
MFETI DP = BD SBDT
BDDC = FETI DP
Theorem
The condition numbers FETI DP and BDDC for, resp., FETI-DP and
BDDC applied to VEM satisfy
Proof relies on series of technical lemmata holding for FEM carried over to
VEM
Example
E macroedge of ` , w Wh` , w 0 Wh` defined as
Then
kw 0 kH 1/2 (E ) . (1 + log(Hk 2 /h))2 kw kH 1/2 (E ) .
00
Test problem
u = 1 in , u = 0 on , =]0, 1[2
u = 1
= 10 , random
in SD
Operator E built by
weighted averaging
+ wi+ + wi
wi =
+ + n
Figure: Left. Largest eigenvalue for increasing polynomial degree, Right. Largest
eigenvalue for increasing H/h and k = 4, 8, N = 256, 64.
1 uniform grid;
2 random grid
#(d.o.f .) N + 1,
h 1/N;
3 random grid
#(d.o.f .) N + 1,
h 1/(2N);
4 random grid
#(d.o.f .) N + 1,
h 1010 .
W in 3D:
continuity at cross points is Rnot enough
continuity of edge average ( E [u] = 0) - easy to impose since on edge
we have pw polynomials R
continuity of face average ( F [u] = 0) - easy to impose for k > 1 since
average on elements are d.o.f
For k > 1 Rieszs basis property will depend on the choice of degrees
of freedom