Download as pdf or txt
Download as pdf or txt
You are on page 1of 153

DISCONTINUOUS PETROV-GALERKIN (DPG)

METHOD
WITH OPTIMAL TEST FUNCTIONS
Tutorial and Perspectives

Leszek Demkowicz
ICES, The University of Texas at Austin

Advanced Numerical Methods in the Mathematical Sciences


Texas A&M University
College Station, May 4 - May 7, 2015
Collaboration:

Portland State: J. Gopalakrishnan


Humboldt U: C. Carstensen
ICES: T. Bui-Thanh, O. Ghattas, B. Moser. T. Ellis, F. Fuentes, B.
Keith, S. Nagaraj, S. Petrides
Argonne: N. Roberts
Basque U: D. Pardo
C.U. of Hong-Kong: W. Qiu
KAUST: V. Calo
Rice: J. Chan
C.U. Chile: I. Muga, N. Heuer
U. Helsinki: A. Niemi
U. Tel-Aviv I. Harari

Texas A&M, May 2015 DPG Method 2 / 131


Act 1: The Big (Functional Analysis) Picture
Act 2: Various Variational Formulations and the Paradigm of Breaking
Test Functions
Act 3: A Tutorial on Coding DPG
Act 4: 3D div-grad Examples
Act 5: 3D Maxwell

Texas A&M, May 2015 DPG Method 3 / 131


Act One

The Big (Functional Analysis) Picture

Texas A&M, May 2015 DPG Method 4 / 131


Three Interpretations of DPG

Texas A&M, May 2015 DPG Method 5 / 131


Abstract variational problem

U, V - Hilbert spaces,
b(u, v) - bilinear (sesquilinear) continuous form on U × V ,

|b(u, v)| ≤ kbk kukU kvkV ,


|{z}
=:M

l(v) - linear (antilinear) continuous functional on V ,

|l(v)k ≤ klkV 0 kvk

The abstract variational problem:

Bu = l B : U → V 0

u∈U

b(u, v) = l(v) ∀v ∈ V < Bu, v >= b(u, v) v ∈ V

Texas A&M, May 2015 DPG Method 6 / 131


Banach - Babuška - Nečas Theorem

If b satisfies the inf-sup condition (⇔ B is bounded below),

|b(u, v)|
inf sup |b(u, v)| =: γ > 0 ⇔ sup ≥ γkukU
kukU =1 kvkV =1 v∈V kvkV

and l satisfies the compatibility condition:

l(v) = 0 ∀v ∈ V0

where
V0 := N (B 0 ) = {v ∈ V : b(u, v) = 0 ∀u ∈ U }
then the variational problem has a unique solution u that satisfies the stability
estimate:
1
kuk ≤ klkV 0 .
γ
Proof: Direct interpretation of Banach Closed Range Theorem∗ .

∗ see e.g. Oden, D, Functional Analysis, Chapman & Hall, 2nd ed., 2010, p.518
Texas A&M, May 2015 DPG Method 7 / 131
Petrov-Galerkin Method and Babuška Theorem

Uh ⊂ U, Vh ⊂ V, dim Uh = dim Vh - finite-dimensional trial and test (sub)spaces



uh ∈ Uh
b(uh , vh ) = l(vh ), ∀vh ∈ Vh

Theorem (Babuška† ).
The discrete inf-sup condition

|b(uh , vh )|
sup ≥ γh kuh kU
vh ∈Vh kvh kV

implies existence, uniqueness and discrete stability

kuh kU ≤ γh−1 klkVh0

† I. Babuska, “Error-bounds for Finite Element Method.”, Numer. Math, 16, 1970/1971.
Texas A&M, May 2015 DPG Method 8 / 131
Petrov-Galerkin Method and Babuška Theorem
Uh ⊂ U, Vh ⊂ V, dim Uh = dim Vh - finite-dimensional trial and test (sub)spaces

uh ∈ Uh
b(uh , vh ) = l(vh ), ∀vh ∈ Vh

Theorem (Babuška† ).
The discrete inf-sup condition
|b(uh , vh )|
sup ≥ γh kuh kU
vh ∈Vh kvh kV
implies existence, uniqueness and discrete stability
kuh kU ≤ γh−1 klkVh0
and convergence
M
ku − uh kU ≤ inf ku − wh kU
γh wh ∈Uh

(Uniform) discrete stability and approximability imply convergence.


† I. Babuska, “Error-bounds for Finite Element Method.”, Numer. Math, 16, 1970/1971.
Texas A&M, May 2015 DPG Method 8 / 131
Optimal test functions

The main trouble:

continuous inf-sup condtion =⇒


/ discrete inf-sup condition

Texas A&M, May 2015 DPG Method 9 / 131


Optimal test functions

The main trouble:

continuous inf-sup condtion =⇒


/ discrete inf-sup condition

unless

‡ L.D., J. Gopalakrishnan. “A Class of Discontinuous Petrov-Galerkin Methods. Part II:

Optimal Test Functions.”Numer. Meth. Part. D. E., 27, 70-105, 2011.


Texas A&M, May 2015 DPG Method 9 / 131
Optimal test functions

The main trouble:

continuous inf-sup condtion =⇒


/ discrete inf-sup condition

unless ‡ we employ special test functions that realize the supremum in the inf-sup
condition:
|b(uh , v)|
vh = arg maxv∈V
kvk

‡ L.D., J. Gopalakrishnan. “A Class of Discontinuous Petrov-Galerkin Methods. Part II:

Optimal Test Functions.”Numer. Meth. Part. D. E., 27, 70-105, 2011.


Texas A&M, May 2015 DPG Method 9 / 131
Optimal test functions

The main trouble:

continuous inf-sup condtion =⇒


/ discrete inf-sup condition

unless ‡ we employ special test functions that realize the supremum in the inf-sup
condition:
|b(uh , v)|
vh = arg maxv∈V
kvk
Recall that the Riesz operator RV : V → V 0 is an isometry. Then:
−1
|b(uh ,v)| (RV Buh ,vh )V
supv kvh k
= kBuh kV 0 = k RV−1 Buh kV = kvh kV
| {z }
=vh
hBuh ,vh i b(uh ,vh )
= kvkV
= kvkV

vh ∈ V
Variational definition of vh :
(v, δv)V = b(uh , δv) ∀δv ∈ V .
The operator T := RV−1 B : Uh → V will be called the trial to test operator.
‡ L.D., J. Gopalakrishnan. “A Class of Discontinuous Petrov-Galerkin Methods. Part II:

Optimal Test Functions.”Numer. Meth. Part. D. E., 27, 70-105, 2011.


Texas A&M, May 2015 DPG Method 9 / 131
DPG is a Minimum Residual Method
With the optimal test functions in place, γh ≥ γ, and the Galerkin method is
automatically stable. Trade now the original norm in U for an energy norm§ :

|b(u, v)|
kukE := kRV−1 BukV = kBukV 0 = sup
v∈V kvkV

Two points:
I With respect to the new, energy norm, both continuity constant M and inf-sup
constant γ are unity.
I The use of optimal test functions (their construction is independent of the choice of
trial norm) implies that γh ≥ γ = 1.
Thus, by the Babuška Theorem,
M
ku − uh kE ≤ inf ku − wh kE .
γh wh ∈Uh
|{z}
=1

In other words, FE solution uh is the best approximation of the exact solution u in the
energy norm. We have arrived through a back door at a Minimum Residual Method.
§ Residual norm really...
Texas A&M, May 2015 DPG Method 10 / 131
Moral of the story

The minimum residual method,


with the residual measured in the dual test norm,
is the most stable Petrov-Galerkin method
you can have.

Texas A&M, May 2015 DPG Method 11 / 131



DPG is a minimum residual method

Bu = l B : U → V 0

u∈U

b(u, v) = l(v) v ∈ V hBu, vi = b(u, v)


I J.H. Bramble, R.D. Lazarov, J.E. Pasciak, “A Least-squares Approach Based on a Discrete Minus One Inner Product for First Order
Systems”Math. Comp, 66, 935-955, 1997.
I L.D., J. Gopalakrishnan. “A Class of Discontinuous Petrov-Galerkin Methods. Part II: Optimal Test Functions.”Numer. Meth. Part. D. E., 27,
70-105, 2011.

Texas A&M, May 2015 DPG Method 12 / 131



DPG is a minimum residual method

Bu = l B : U → V 0

u∈U

b(u, v) = l(v) v ∈ V hBu, vi = b(u, v)

I Minimum residual method: Uh ⊂ U ,


1
2 kBuh − lk2V 0 → min
uh ∈Uh


I J.H. Bramble, R.D. Lazarov, J.E. Pasciak, “A Least-squares Approach Based on a Discrete Minus One Inner Product for First Order
Systems”Math. Comp, 66, 935-955, 1997.
I L.D., J. Gopalakrishnan. “A Class of Discontinuous Petrov-Galerkin Methods. Part II: Optimal Test Functions.”Numer. Meth. Part. D. E., 27,
70-105, 2011.

Texas A&M, May 2015 DPG Method 12 / 131



DPG is a minimum residual method

Bu = l B : U → V 0

u∈U

b(u, v) = l(v) v ∈ V hBu, vi = b(u, v)

I Minimum residual method: Uh ⊂ U ,


1
2 kBuh − lk2V 0 → min
uh ∈Uh

I Riesz operator:
RV : V → V 0 , hRV v, δvi = (v, δv)V
is an isometry, kRV vkV 0 = kvkV .


I J.H. Bramble, R.D. Lazarov, J.E. Pasciak, “A Least-squares Approach Based on a Discrete Minus One Inner Product for First Order
Systems”Math. Comp, 66, 935-955, 1997.
I L.D., J. Gopalakrishnan. “A Class of Discontinuous Petrov-Galerkin Methods. Part II: Optimal Test Functions.”Numer. Meth. Part. D. E., 27,
70-105, 2011.

Texas A&M, May 2015 DPG Method 12 / 131



DPG is a minimum residual method

Bu = l B : U → V 0

u∈U

b(u, v) = l(v) v ∈ V hBu, vi = b(u, v)

I Minimum residual method: Uh ⊂ U ,


1
2 kBuh − lk2V 0 → min
uh ∈Uh

I Riesz operator:
RV : V → V 0 , hRV v, δvi = (v, δv)V
is an isometry, kRV vkV 0 = kvkV .
I Minimum residual method reformulated:
1
2 kBuh − lk2V 0 = 21 kRV−1 (Buh − l)k2V → min
uh ∈Uh


I J.H. Bramble, R.D. Lazarov, J.E. Pasciak, “A Least-squares Approach Based on a Discrete Minus One Inner Product for First Order
Systems”Math. Comp, 66, 935-955, 1997.
I L.D., J. Gopalakrishnan. “A Class of Discontinuous Petrov-Galerkin Methods. Part II: Optimal Test Functions.”Numer. Meth. Part. D. E., 27,
70-105, 2011.

Texas A&M, May 2015 DPG Method 12 / 131


DPG is a minimum residual method

Taking Gâteaux derivative,

(RV−1 (Buh − l), RV−1 Bδuh )V = 0 δuh ∈ Uh

Texas A&M, May 2015 DPG Method 13 / 131


DPG is a minimum residual method

Taking Gâteaux derivative,

(RV−1 (Buh − l), RV−1 Bδuh )V = 0 δuh ∈ Uh

or
hBuh − l, RV−1 Bδuh i = 0 δuh ∈ Uh

Texas A&M, May 2015 DPG Method 13 / 131


DPG is a minimum residual method

Taking Gâteaux derivative,

(RV−1 (Buh − l), RV−1 Bδuh )V = 0 δuh ∈ Uh

or
hBuh − l, RV−1 Bδuh i = 0 δuh ∈ Uh
| {z }
vh

Texas A&M, May 2015 DPG Method 13 / 131


DPG is a minimum residual method

Taking Gâteaux derivative,

(RV−1 (Buh − l), RV−1 Bδuh )V = 0 δuh ∈ Uh

or
hBuh − l, vh i = 0 vh = RV−1 Bδuh

Texas A&M, May 2015 DPG Method 13 / 131


DPG is a minimum residual method

Taking Gâteaux derivative,

(RV−1 (Buh − l), RV−1 Bδuh )V = 0 δuh ∈ Uh

or
hBuh , vh i = hl, vh i vh = RV−1 Bδuh

Texas A&M, May 2015 DPG Method 13 / 131


DPG is a minimum residual method

Taking Gâteaux derivative,

(RV−1 (Buh − l), RV−1 Bδuh )V = 0 δuh ∈ Uh

or
b(uh , vh ) = l(vh )
where 
vh ∈ V
(vh , δv)V = b(δuh , δv) δv ∈ V

Texas A&M, May 2015 DPG Method 13 / 131


DPG is a mixed method

An alternate route k ,

( RV−1 (Buh − l) , RV−1 Bδuh )V = 0 δuh ∈ Uh


| {z }
=:ψ(error representation function)

k W. Dahmen, Ch. Huang, Ch. Schwab, and G. Welper. “Adaptive Petrov Galerkin methods

for first order transport equations”, SIAM J. Num. Anal. 50(5): 242-2445, 2012
Texas A&M, May 2015 DPG Method 14 / 131
DPG is a mixed method

An alternate route k ,

( RV−1 (Buh − l) , RV−1 Bδuh )V = 0 δuh ∈ Uh


| {z }
=:ψ(error representation function)

or
ψ = RV−1 (Buh − l)
(

(ψ, RV−1 Bδuh )V = 0 δuh ∈ Uh

k W. Dahmen, Ch. Huang, Ch. Schwab, and G. Welper. “Adaptive Petrov Galerkin methods

for first order transport equations”, SIAM J. Num. Anal. 50(5): 242-2445, 2012
Texas A&M, May 2015 DPG Method 14 / 131
DPG is a mixed method

An alternate route k ,

( RV−1 (Buh − l) , RV−1 Bδuh )V = 0 δuh ∈ Uh


| {z }
=:ψ(error representation function)

or
(ψ, δv)V − b(uh , δv) = −l(δv) ∀δv ∈ V
(

b(δuh , ψ) =0 ∀δuh ∈ Uh

k W. Dahmen, Ch. Huang, Ch. Schwab, and G. Welper. “Adaptive Petrov Galerkin methods

for first order transport equations”, SIAM J. Num. Anal. 50(5): 242-2445, 2012
Texas A&M, May 2015 DPG Method 14 / 131
DPG method, a summary so far
I Stiffness matrix is always hermitian and positive-definite (it is a
generalization of the least squares method...).

∗∗ J.T.Oden, L.D., T.Strouboulis and Ph. Devloo, “Adaptive Methods for Problems in Solid and Fluid Mechanics”, in Accuracy Estimates and
Adaptive Refinements in Finite Element Computations, Wiley & Sons, London 1986
Texas A&M, May 2015 DPG Method 15 / 131
DPG method, a summary so far
I Stiffness matrix is always hermitian and positive-definite (it is a
generalization of the least squares method...).
I The method delivers the best approximation error (BAE) in the “energy
norm”:
|b(u, v)|
kukE := kBukV 0 = sup
v∈V kvkV

∗∗ J.T.Oden, L.D., T.Strouboulis and Ph. Devloo, “Adaptive Methods for Problems in Solid and Fluid Mechanics”, in Accuracy Estimates and
Adaptive Refinements in Finite Element Computations, Wiley & Sons, London 1986
Texas A&M, May 2015 DPG Method 15 / 131
DPG method, a summary so far
I Stiffness matrix is always hermitian and positive-definite (it is a
generalization of the least squares method...).
I The method delivers the best approximation error (BAE) in the “energy
norm”:
|b(u, v)|
kukE := kBukV 0 = sup
v∈V kvkV
I The energy norm of the FE error u − uh equals the residual and can be
computed,

ku − uh kE = kBu − Buh kV 0 = kl − Buh kV 0 = kRV−1 (l − Buh )kV = kψkV

where the error representation function ψ comes from



ψ∈V
(ψ, δv)V = hl − Buh , δvi = l(δv) − b(uh , δv), δv ∈ V

No need for a-posteriori error estimation, note the connection with implicit
a-posteriori error estimation techniques ∗∗
∗∗ J.T.Oden, L.D., T.Strouboulis and Ph. Devloo, “Adaptive Methods for Problems in Solid and Fluid Mechanics”, in Accuracy Estimates and
Adaptive Refinements in Finite Element Computations, Wiley & Sons, London 1986
Texas A&M, May 2015 DPG Method 15 / 131
DPG method, a summary

I A lot depends upon the choice of the test norm k · kV ; for different test
norms, we get get different methods.

Texas A&M, May 2015 DPG Method 16 / 131


DPG method, a summary

I A lot depends upon the choice of the test norm k · kV ; for different test
norms, we get get different methods.
I How to choose the test norm in a systematic way ?

Texas A&M, May 2015 DPG Method 16 / 131


DPG method, a summary

I A lot depends upon the choice of the test norm k · kV ; for different test
norms, we get get different methods.
I How to choose the test norm in a systematic way ?
I Is the inversion of Riesz operator (computation of the optimal test functions,
energy error) feasible ?

Texas A&M, May 2015 DPG Method 16 / 131


DPG method, a summary

I A lot depends upon the choice of the test norm k · kV ; for different test
norms, we get get different methods.
I How to choose the test norm in a systematic way ?
I Is the inversion of Riesz operator (computation of the optimal test functions,
energy error) feasible ?
I Being a Ritz method, DPG does not experience any preasymptotic
limitations.

Texas A&M, May 2015 DPG Method 16 / 131


Wait a minute!

You cannot compute the optimal test functions!

Texas A&M, May 2015 DPG Method 17 / 131


Approximate optimal test functions

Take a finite-dimensional enriched test space: Ṽ ⊂ V , dim Ṽ >> dim Uh , and


invert the Riesz operator approximately,

ṽh ∈ Ṽ
(ṽh , δv) f ∀δv
f V = b(uh , δv) f ∈ Ṽ .

This leads to an approximate trial to test operator:

T̃ : Uh → Ṽ T̃ uh := ṽh

and approximate optimal test space:

Ṽh := T̃ Uh .

Some stability must be lost. How much ?

Texas A&M, May 2015 DPG Method 18 / 131


Approximate mixed problem


 ψ̃ ∈ Ṽ , ũh ∈ Uh
f V − b(ũh , δψ)
(ψ̃, δψ) f = −l(δψ)
f δψf ∈ Ṽ

b(δuh , ψ̃) =0 δuh ∈ Uh
The (discrete) inf sup condition must be satisfied:

˜
|b(uh , δψ)|
sup ≥ γh kuh k
˜ Ṽ
δψ∈
˜
kδψk

Back to square one ??

Texas A&M, May 2015 DPG Method 19 / 131


Fortin operator

Coming up with a Fortin operator

Π̃ : V → Ṽ

such that
kΠ̃vkV ≤ CkvkV
and
b(uh , Π̃v − v) = 0 ∀uh ∈ Uh
††
solves the problem

†† J. Gopalakrishnan and W. Qiu. “An analysis of the practical DPG method.”, Math. Comp.,

2013 (posted May 31, 2013).


See also the poster of S. Nagaraj and S. Petrides
Texas A&M, May 2015 DPG Method 20 / 131
Act Two

Various Variational Formulations and


the Paradigm of Breaking Test Functions
‡‡

‡‡ L.D., Various Variational Formulations and Closed Range Theorem, ICES Report 15/03.
Texas A&M, May 2015 DPG Method 21 / 131
Notation

I N Lipschitz domain,
Standard assumptions: Ω ⊂ R

Elements:K
Edges:e S
Skeleton:Γh = K ∂K
Internal skeleton:Γ0h = Γh − ∂Ω

Texas A&M, May 2015 DPG Method 22 / 131


Diffusion-Convection-Reaction Problem


 u = u0 on Γ1
u = u0 on Γ1


σn = σ0 on Γ2
 
∂u
 ∂n − βn u = σ0 on Γ2 ⇔
σ = ∇u − βu /τ (1)
−∆u + ∇ · (βu) + cu =f in Ω
 

−div σ + cu =f /v (2)

To relax or not to relax ?

(1) (2) name energy setting


1 no no trivial (strong) u ∈ H 1 , σ ∈ H(div), v ∈ L2 , τ ∈ (L2 )N
2 no yes classical u, v ∈ H 1 , σ, τ ∈ (L2 )N
3 yes no mixed u, v ∈ L2 , σ, τ ∈ H(div)
4 yes yes ultraweak u ∈ L2 , σ ∈ (L2 )N , v ∈ H 1 , τ ∈ H(div)

The inf-sup constants for different variational formulations are equal or O(1)-equivalent
(Closed Range Theorem at work).

Texas A&M, May 2015 DPG Method 23 / 131


Classical Formulation



 u = u0 on Γ1
σn = σ0 on Γ2


 σ = ∇u − βu /τ (1)
−div σ + cu =f /v (2)

Relax (2):

(σ, ∇v) + (cu, v) = (f, v) + hσ0 , vi, v ∈ H 1 (Ω) : v = 0 on Γ1

Use the strong form of (1) to eliminate σ,

u ∈ H 1 (Ω) : u = u0 on Γ1
(

(∇u − βu, ∇v) + (cu, v) = (f, v) + hσ0 , vi, v ∈ H 1 (Ω) : v = 0 on Γ1

Texas A&M, May 2015 DPG Method 24 / 131


Mixed Formulation


 u = u0 on Γ1
σn = σ0 on Γ2

1
(σ + βu) = ∇u /τ (1)
 


−div σ + cu =f /v (2)
Relax (1):
1
(σ + βu, τ ) = −(u, divτ ) + hu0 , τn i, τ ∈ H(div, Ω) : τn = 0 on Γ2

Use the strong form of (2) to eliminate u,

 σ ∈ H(div, Ω), σn = σ0 on Γ2
1
(σ + β 1c (f + div σ), τ ) = −( 1c (f + div σ), divτ ) + hu0 , τn i
 
τ ∈ H(div, Ω) : τn = 0 on Γ2
or leave it alone in the strong form (a must if c = 0),

 u ∈ L2 (Ω), σ ∈ H(div, Ω), σn = σ0 on Γ2
1
 (σ + βu, τ ) + (u, divτ ) = hu0 , τn i, τ ∈ H(div, Ω) : τn = 0 on Γ2
−(div σ + cu, v) = (f, v) v ∈ L2 (Ω)

Texas A&M, May 2015 DPG Method 25 / 131


Ultraweak Formulation



 u = u0 on Γ1
σn = σ0 on Γ2

1
(σ + βu) = ∇u /τ (1)
 


−div σ + cu =f /v (2)
Relax (1) and (2):

 u ∈ L2 (Ω), σ ∈ (L2 (Ω))N
1
(σ + βu, τ ) = −(u, divτ ) + hu0 , τn i, τ ∈ H(div, Ω) : τn = 0 on Γ2
 
(σ, ∇v) + (cu, v) = (f, v) + hσ0 , vi, v ∈ H 1 (Ω) : v = 0 on Γ1

Texas A&M, May 2015 DPG Method 26 / 131


Breaking Test Functions

In each of the possible variational formulations we can replace the conforming test
functions with broken test functions. The resulting inf-sup constants remain of the
same order and, in particular, they are mesh independent .

C. Carstensen, L.D., and J. Gopalakrishnan, The Paradigm of Broken Test Functions in DPG
Discretizations of Elliptic Second–Order PDEs, in preparation.
Texas A&M, May 2015 DPG Method 27 / 131
Classical Formulation with Broken Test Functions

Remove the BC from the test functions


u ∈ H 1 (Ω) : u = u0 on Γ1




t̂ ∈ H −1/2 (Γ) : t̂ = σ0 on Γ2


v ∈ H 1 (Ω)

(∇u − βu, ∇v) + (cu, v)−ht̂, viΓ = (f, v)

Break test functions:


u ∈ H 1 (Ω) : u = u0 on Γ1




t̂ ∈ H −1/2 (Γh ) : t̂ = σ0 on Γ2


v ∈ H 1 (Ωh )

(∇u − βu, ∇h v) + (cu, v)−ht̂, viΓh = (f, v)

where Γh is now the mesh skeleton and

H −1/2 (Γh ) := tr H(div, Ω) on Γh

Primal DPG Method


Texas A&M, May 2015 DPG Method 28 / 131
Act Three

A Tutorial on Coding the DPG Method

Texas A&M, May 2015 DPG Method 29 / 131


The main point

The test norm must be localizable, e.g.,


X
kvk2H 1 (Ωh ) = kv|K k2H 1 (K) .
| {z } K | {z }
global norm local norm

The (approximate) inversion of the Riesz operator is done elementwise.

Texas A&M, May 2015 DPG Method 30 / 131


DPG element stiffness matrix and load vector

N
X M
X
uh = ui ei , vh ≈ vj gj , M >> N
i=1 j=1

Computation of (approximate) optimal test function v = T ei ,


X
(gj , gl ) vji = b(ei , gl ) , l = 1, . . . , M
| {z } | {z }
j
Gram matrix G expanded stiffness matrix B

or
v = G−1 Bδu
The DPG stiffness matrix and load vector:

v T Bu = (G−1 Bδu)T Bu = (δu)T B T G−1 Bu

v T b = (G−1 Bδu)T b = (δu)T B T G−1 b

Texas A&M, May 2015 DPG Method 31 / 131


Same result with the mixed method interpretation

    
G −B ψ −b
=
BT u 0
Condensing out error indication function ψ,

ψ = G−1 (Bu − b)

we get again,
B T G−1 B u = G−1 Bb

Texas A&M, May 2015 DPG Method 32 / 131


Primal DPG Formulation

Group unknown (watch for the overloaded symbol):

uh := ( uh , t̂h )
|{z} |{z}
field flux

Mixed system:
    
G −B 1 −B 2 ψ −b
 B T1 0 0  u = 0 
B T2 0 0 t̂ 0

where B 1 , B 2 correspond to (∇uh , ∇h ṽ) and −ht̂h , ṽi, resp.


Eliminate ψ to get the DPG system:
 T −1
B 1 G B 1 B T1 G−1 B 2
   T −1 
u B1 G b
=
B T2 G−1 B 1 B T2 G−1 B 2 t̂ B T2 G−1 b

Texas A&M, May 2015 DPG Method 33 / 131


Primal DPG method

Neglecting the error steming from the approximation of optimal test function
(computation of residual), we have,
 1/2
ku − uh k2H 1 (Ω) +kt̂ − t̂h k2H −1/2 (Γ
h)
 1/2
≤ 1
inf
γ w ,r
ku − wh k2H 1 (Ω) + kt̂ − rh k2H −1/2 (Γh )
h h
| {z }
best approximation error

Additionally,
 1/2
ku − uh k2H 1 (Ω) +kt̂ − t̂h k2H −1/2 (Γ
h)

1 |(∇uh , ∇h v) − ht̂h , viΓh |


≤ γ
sup
v∈H 1 (Ωh ) kvkH 1 (Ωh )
| {z }
computable residual
P 1/2
1 2
= γ K kψK kH 1 (K)

Texas A&M, May 2015 DPG Method 34 / 131


2D convergence rates

Texas A&M, May 2015 DPG Method 35 / 131


hp3d and shape functions

All reported 3D examples were coded within hp3d - a three-dimensional hp code


supporting:
I hybrid meshes with elements of all shapes: hexas, tets, prisms and pyramids,
I first Nedèlèc family of H 1 -,H(curl)-, H(div)-, and L2 -conforming elements,
I 1-irregular meshes and anisotropic hp-refinements.
The computations used a recently developed suite of orientation embedded shape
functions for elements of all shapes and the whole exact sequence that can be
downloaded from:

https://github.com/libESEAS/ESEAS

Developed with Paolo Gatto and Kyungjoo Kim.


F. Fuentes, B. Keith, L. D. and S. Nagaraj, Orientation Embedded High Order Shape
Functions for the Exact Sequence Elements of All Shapes, ICES Report 2015/07
Texas A&M, May 2015 DPG Method 36 / 131
Act Four

3D div-grad Examples:
I Poisson problem
I Reaction-dominated diffusion
I Convection-dominated diffusion

Texas A&M, May 2015 DPG Method 37 / 131


FE discretization for div-grad problems

Hexahedral meshes
H 1 element for field uh :
P p ⊗ P p ⊗ P p,
Trace of H(div) element:

(P p ⊗ P p−1 ⊗ P p−1 ) × (P p−1 ⊗ P p ⊗ P p−1 ) × (P p−1 ⊗ P p−1 ⊗ P p )

for flux t̂h ,


and the enriched element:

P p+∆p ⊗ P p+∆p ⊗ P p+∆p ,

for test function vh .


In reported experiments: p = 1, 2, 3, ∆p = 2.

Texas A&M, May 2015 DPG Method 38 / 131


Poisson problem, smooth solution, uniform refinements
Rectangular domain Ω = (0, 1) × (0, 2) × (0, 1),
Smooth solution: u = sin πx sin πy sin πz
Boundary condition: u = 0.

Residual versus H 1 error.


Texas A&M, May 2015 DPG Method 39 / 131
Poisson problem, manufactured shock solution
BC: u = u0 .

Texas A&M, May 2015 DPG Method 40 / 131


Shock solution, uniform and h-adaptive refinements, p = 1

Convergence history for the residual and H 1 error

Texas A&M, May 2015 DPG Method 41 / 131


Shock solution, uniform and h-adaptive refinements, p = 2

Convergence history for the residual and H 1 error

Texas A&M, May 2015 DPG Method 42 / 131


Shock solution, uniform and h-adaptive refinements, p = 3

Convergence history for the residual and H 1 error

Texas A&M, May 2015 DPG Method 43 / 131


Shock solution, p = 3, Mixed BC
Mixed BC: trace: bottom, top, flux: sides.

Convergence history for the residual and H 1 error


Texas A&M, May 2015 DPG Method 44 / 131
Reaction-dominated diffusion, p = 2.


u =0 on Γ
−2 ∆u + u = 1 in Ω

 = 0.01, left: solution after 7 iterations, right: convergence history

Texas A&M, May 2015 DPG Method 45 / 131


Convection-dominated diffusion, p = 2.


 −2 ∆u − u = sin πy sin πz at x = 0
u =0 on the rest of Γ
−2 ∆u + ∂u =0 in Ω

∂x

 = 0.01, left: solution after 5 iterations, right: convergence history

Texas A&M, May 2015 DPG Method 46 / 131


Act Five

3D Maxwell

Texas A&M, May 2015 DPG Method 47 / 131


Time-harmonic Maxwell equations



 ∇ × E + iωµH =0 Faraday Law



= J imp




 ∇ × H − iωE − σE Ampère Law


 ∇ · (µH) =0 Gauss Magnetic Law



= ρimp + ρ




 −∇ · (()E) Gauss Electric Law


−iωρ + ∇(σE) =0 conservation of charge

where: 
E electric field 
H magnetic field the unknowns
ρ free charge 

µ permeability 
 permittivity material constants
σ conductivity

J imp

impressed current
load data
ρimp impressed charge

Texas A&M, May 2015 DPG Method 48 / 131


A bit of history

Electrostatics:
Coulomb’s Law (1775) electric field
→ →
polarization, dielectrics,  Gauss Electric Law (1835)

Magnetostatics:

Ampère Force Law (1820) magnetic field Ampère Law


→ →
(steady currents) magnetic polarization, µ Gauss Magnetic Law

Faraday Law (1831)


Maxwell equations (1856) including correction to Ampère Law.
Current formalism due to Heaviside (1884)
Inspired Einstein on his way to Special Relativity.

Texas A&M, May 2015 DPG Method 49 / 131


Time-harmonic Maxwell equations with ρ eliminated
Eliminating the free charge, we get:
∇ × E + iωµH =0 Faraday Law






= J imp

 ∇ × H − iωE − σE Ampère Law

∇ · (µH) =0 Gauss Magnetic Law










= −iωρimp

−∇ · ((σ + iω)E) continuity equation
The equations are linearly dependent:
Take curl of the Faraday Law to obtain the Gauss Magnetic Law.
Take curl of the Ampère Law to obtain the continuity equation. Notice that J imp , ρimp
must satisfy the compatibility condition:
∇ · J imp = −iωρimp .
Boundary Conditions (BC):
n×E = n × E imp on Γ1

n×H = n × H imp =: JSimp on Γ2


|{z}
impressed surface current

Texas A&M, May 2015 DPG Method 50 / 131


Different variational formulations

= n × E imp


 n×E on Γ1
n×H = n × H imp on Γ2


 ∇ × E + iωµH =0 /φ (1)
∇ × H − (iω + σ)E = J imp /ψ (2)

To relax or not to relax ?

(1) (2) name energy setting


1 no no trivial (strong) E, H ∈ H(curl), φ, ψ ∈ L2
2 no yes standard E, ψ ∈ H(curl), H, φ ∈ L2
3 yes no standard H, φ ∈ H(curl), E, ψ ∈ L2
4 yes yes ultraweak E, H ∈ L2 , φ, ψ ∈ H(curl)

The inf-sup constants for different variational formulations are equal or O(1)-equivalent
(Closed Range Theorem at work).
Only standard variational formulations are eligible for the Bubnov-Galerkin method
(symmetric functional setting)

L.D. “Various Variational Formulations and Closed Range Theorem”, ICES Report
15/03
Texas A&M, May 2015 DPG Method 51 / 131
Standard variational formulation
Relax the Ampère equation:
I multiply with factor −iω,

∇ × (−iωH) + (−ω 2  + iωσ)E = −iωJ imp ,

I multiply with a test function ψ and integrate by parts the curl term,

(−iωH, ∇ × ψ) + hn × (−iωH), ψi + ((−ω 2  + iωσ)E, ψ) = −iω(J imp , ψ) ,

I build the second boundary condition into the formulation and eliminate the rest of
the boundary term by not testing on Γ1 , i.e. assuming n × ψ = 0 on Γ1 ,

(−iωH, ∇×ψ)+((−ω 2 +iωσ)E, ψ) = −iω(J imp , ψ)+iωhJSimp , ψiΓ2 n×ψ = 0 on Γ1 .

Use the strong form of the Faraday equation to eliminate H:


E ∈ H(curl, Ω), n × E = n × E imp on Γ1 ,





( µ1 ∇ × E, ∇ × ψ) + ((−ω 2  + iωσ)E, ψ) = −iω(J imp , ψ) + iωhJSimp , ψiΓ2



ψ ∈ H(curl, Ω) : n × ψ = 0 on Γ1 .

Texas A&M, May 2015 DPG Method 52 / 131


Ultraweak variational formulation

Relax both equations:


E, H ∈ L2 (Ω)





= −hn × E inc , φi

(E, ∇ × φ) + (iωµH, φ)



φ ∈ H(curl, Ω), n × φ = 0 on Γ2



(H, ∇ × ψ) − ((iω + σ)E, ψ) = (J imp , ψ) − hn × H inc , ψi




ψ ∈ H(curl, Ω), n × ψ = 0 on Γ1

You may test on the whole boundary:


−1/2


 E ∈ L2 (Ω), Ê ∈ Ht (curl, Γ), n × Ê = n × E inc on Γ1



−1/2
H ∈ L2 (Ω), Ĥ ∈ Ht = n × H inc


 (curl, Γ), n × Ĥ on Γ2

(E, ∇ × φ) + (iωµH, φ) + hn × Ê, φi =0 φ ∈ H(curl, Ω)









(H, ∇ × ψ) − ((iω + σ)E, ψ) + hn × Ĥ, ψi = (J imp , ψ) ψ ∈ H(curl, Ω)

−1/2
where Ht (curl, Γ) := trΓ H(curl, Ω).

Texas A&M, May 2015 DPG Method 53 / 131


Breaking test functions
Primal DPG formulation:


 E ∈ H(curl, Ω), n × E = n × E imp on Γ1 ,



−1/2
= n × H inc


 Ĥ ∈ Ht (curl, Γh ), n × Ĥ on Γ2

( µ1 ∇ × E, ∇h × ψ) + ((−ω 2  + iωσ)E, ψ) + iωhĤ, ψi = −iω(J imp , ψ)









ψ ∈ H(curl, Ωh )

Ultraweak DPG formulation:


−1/2


 E ∈ L2 (Ω), Ê ∈ Ht (curl, Γh ), n × Ê = n × E inc on Γ1



−1/2
H ∈ L2 (Ω), Ĥ ∈ Ht = n × H inc


 (curl, Γh ), n × Ĥ on Γ2

(E, ∇h × φ) + (iωµH, φ) + hn × Ê, φi =0 φ ∈ H(curl, Ωh )









(H, ∇h × ψ) − ((iω + σ)E, ψ) + hn × Ĥ, ψi = (J imp , ψ) ψ ∈ H(curl, Ωh )

Both formulations retain stability properties of the original variational formulations with
mesh independent inf-sup constants of the same order of magnitude.
C. Carstensen, L.D., and J. Gopalakrishnan, The Paradigm of Broken Test Functions in DPG
Discretizations of Elliptic Second–Order PDEs, in preparation.
Texas A&M, May 2015 DPG Method 54 / 131
FE discretization for curl problems

Hexahedral meshes
H(curl) element for electric field E:

(P p−1 ⊗ P p ⊗ P p ) × (P p ⊗ P p−1 ⊗ P p ) × (P p ⊗ P p ⊗ P p−1 )

and trace of the same element for flux (surface current) Ĥ.
Same element for the enriched space but with order p + ∆p.
In reported experiments: p = 2, ∆p = 2.

Texas A&M, May 2015 DPG Method 55 / 131


Fichera corner

Divide it into eight smaller cubes and remove one:

Texas A&M, May 2015 DPG Method 56 / 131


Fichera corner microwave
Attach a waveguide:

 = µ = 1, σ = 0
ω = 5(1.6 wavelengths in the cube)

Cut the waveguide and use the lowest propagating mode for BC along the cut.
Texas A&M, May 2015 DPG Method 57 / 131
Fichera corner microwave

Standard variational formulation


(Primal DPG method)

Standard test norm:

kψk2V := kψk2H(curl,Ω) = kψk2 + k∇ × ψk2

Texas A&M, May 2015 DPG Method 58 / 131


Initial mesh and real part of E1

Texas A&M, May 2015 DPG Method 59 / 131


Mesh and real part of E1 after 1st refinement

Texas A&M, May 2015 DPG Method 60 / 131


Mesh and real part of E1 after 2nd refinement

Texas A&M, May 2015 DPG Method 61 / 131


Mesh and real part of E1 after 3rd refinement

Texas A&M, May 2015 DPG Method 62 / 131


Mesh and real part of E1 after 4th refinement

Texas A&M, May 2015 DPG Method 63 / 131


Mesh and real part of E1 after 5th refinement

Texas A&M, May 2015 DPG Method 64 / 131


Mesh and real part of E1 after 6th refinement

Texas A&M, May 2015 DPG Method 65 / 131


Mesh and real part of E1 after 7th refinement

Texas A&M, May 2015 DPG Method 66 / 131


Mesh and real part of E1 after 8th refinement

Texas A&M, May 2015 DPG Method 67 / 131


Mesh and real part of E1 after 9th refinement

Texas A&M, May 2015 DPG Method 68 / 131


Mesh and real part of E1 after 10th refinement

Texas A&M, May 2015 DPG Method 69 / 131


Residual history

Residual decreases monotonically.

Texas A&M, May 2015 DPG Method 70 / 131


Fichera corner microwave

Ultraweak variational formulation


(Original DPG method)

Adjoint graph norm:

k(φ, ψ)k2V := kφk2 + kψk2 + k∇ × φ + (iω − σ)ψk2 + k∇ × ψ − iωµφk2

Texas A&M, May 2015 DPG Method 71 / 131


Initial mesh and real part of E1

Texas A&M, May 2015 DPG Method 72 / 131


Mesh and real part of E1 after 1st refinement

Texas A&M, May 2015 DPG Method 73 / 131


Mesh and real part of E1 after 2nd refinement

Texas A&M, May 2015 DPG Method 74 / 131


Mesh and real part of E1 after 3rd refinement

Texas A&M, May 2015 DPG Method 75 / 131


Mesh and real part of E1 after 4th refinement

Texas A&M, May 2015 DPG Method 76 / 131


Mesh and real part of E1 after 5th refinement

Texas A&M, May 2015 DPG Method 77 / 131


Mesh and real part of E1 after 6th refinement

Texas A&M, May 2015 DPG Method 78 / 131


Mesh and real part of E1 after 7th refinement

Texas A&M, May 2015 DPG Method 79 / 131


Mesh and real part of E1 after 8th refinement

Texas A&M, May 2015 DPG Method 80 / 131


Residual history

Residual decreases monotonically.

Texas A&M, May 2015 DPG Method 81 / 131


2D Linear acoustics (equivalent to 2D Maxwell)

Texas A&M, May 2015 DPG Method 82 / 131


Example: Gaussian beam

Texas A&M, May 2015 DPG Method 83 / 131


Mesh 1 and real part of pressure

Texas A&M, May 2015 DPG Method 84 / 131


Mesh 2 and real part of pressure

Texas A&M, May 2015 DPG Method 85 / 131


Mesh 3 and real part of pressure

Texas A&M, May 2015 DPG Method 86 / 131


Mesh 4 and real part of pressure

Texas A&M, May 2015 DPG Method 87 / 131


Mesh 5 and real part of pressure

Texas A&M, May 2015 DPG Method 88 / 131


Mesh 6 and real part of pressure

Texas A&M, May 2015 DPG Method 89 / 131


Mesh 7 and real part of pressure

Texas A&M, May 2015 DPG Method 90 / 131


Mesh 8 and real part of pressure

Texas A&M, May 2015 DPG Method 91 / 131


Mesh 9 and real part of pressure

Texas A&M, May 2015 DPG Method 92 / 131


Mesh 10 and real part of pressure

Texas A&M, May 2015 DPG Method 93 / 131


Mesh 11 and real part of pressure

Texas A&M, May 2015 DPG Method 94 / 131


Mesh 12 and real part of pressure

Texas A&M, May 2015 DPG Method 95 / 131


Mesh 13 and real part of pressure

Texas A&M, May 2015 DPG Method 96 / 131


Mesh 14 and real part of pressure

Texas A&M, May 2015 DPG Method 97 / 131


Mesh 15 and real part of pressure

Texas A&M, May 2015 DPG Method 98 / 131


Mesh 16 and real part of pressure

Texas A&M, May 2015 DPG Method 99 / 131


Mesh 17 and real part of pressure

Texas A&M, May 2015 DPG Method 100 / 131


Mesh 18 and real part of pressure

Texas A&M, May 2015 DPG Method 101 / 131


Mesh 19 and real part of pressure

Texas A&M, May 2015 DPG Method 102 / 131


Mesh 20 and real part of pressure

Texas A&M, May 2015 DPG Method 103 / 131


Mesh 21 and real part of pressure

Texas A&M, May 2015 DPG Method 104 / 131


Mesh 22 and real part of pressure

Texas A&M, May 2015 DPG Method 105 / 131


Mesh 23 and real part of pressure

Texas A&M, May 2015 DPG Method 106 / 131


Mesh 24 and real part of pressure

Texas A&M, May 2015 DPG Method 107 / 131


Mesh 25 and real part of pressure

Texas A&M, May 2015 DPG Method 108 / 131


Mesh 26 and real part of pressure

Texas A&M, May 2015 DPG Method 109 / 131


Mesh 27 and real part of pressure

Texas A&M, May 2015 DPG Method 110 / 131


Mesh 28 and real part of pressure

Texas A&M, May 2015 DPG Method 111 / 131


Mesh 29 and real part of pressure

Texas A&M, May 2015 DPG Method 112 / 131


Mesh 30 and real part of pressure

Texas A&M, May 2015 DPG Method 113 / 131


Mesh 31 and real part of pressure

Texas A&M, May 2015 DPG Method 114 / 131


Mesh 32 and real part of pressure

Texas A&M, May 2015 DPG Method 115 / 131


Mesh 33 and real part of pressure

Texas A&M, May 2015 DPG Method 116 / 131


Mesh 34 and real part of pressure

Texas A&M, May 2015 DPG Method 117 / 131


Mesh 35 and real part of pressure

Texas A&M, May 2015 DPG Method 118 / 131


Mesh 36 and real part of pressure

Texas A&M, May 2015 DPG Method 119 / 131


Mesh 37 and real part of pressure

Texas A&M, May 2015 DPG Method 120 / 131


Mesh 38 and real part of pressure

Texas A&M, May 2015 DPG Method 121 / 131


Mesh 39 and real part of pressure

Texas A&M, May 2015 DPG Method 122 / 131


Mesh 40 and real part of pressure

Texas A&M, May 2015 DPG Method 123 / 131


Mesh 41 and real part of pressure

Texas A&M, May 2015 DPG Method 124 / 131


Mesh 42 and real part of pressure

Texas A&M, May 2015 DPG Method 125 / 131


Mesh 44 and real part of pressure

Texas A&M, May 2015 DPG Method 126 / 131


Mesh 46 and real part of pressure

Texas A&M, May 2015 DPG Method 127 / 131


Residual and L2 error convergence history

Residual (dotted line) decreases always while error (solid line) decreases only in
the asymptotic regime.
Texas A&M, May 2015 DPG Method 128 / 131
A non-trivial question

Texas A&M, May 2015 DPG Method 129 / 131


A non-trivial question

I Does the satisfaction of the weak form of the Ampère equation imply the
satisfaction of the continuity equation?

Texas A&M, May 2015 DPG Method 129 / 131


A non-trivial question

I Does the satisfaction of the weak form of the Ampère equation imply the
satisfaction of the continuity equation?
I More precisely, does the minimization of the residual corresponding to the
Ampère eqn imply the control of the residual corresponding to the continuity
equation ? (Not the case for an explicit a-posteriori error estimation for the
standard Galerkin method.)

Texas A&M, May 2015 DPG Method 129 / 131


A non-trivial question

I Does the satisfaction of the weak form of the Ampère equation imply the
satisfaction of the continuity equation?
I More precisely, does the minimization of the residual corresponding to the
Ampère eqn imply the control of the residual corresponding to the continuity
equation ? (Not the case for an explicit a-posteriori error estimation for the
standard Galerkin method.)
I The answer to the last two questions seems to be positive. For the primal
DPG method,

|b((E, Ĥ), ∇h q)| |b((E, Ĥ), ∇h q)| |b((E, Ĥ), ψ)|


sup ≤ sup ≤ sup
q kqkH 1 (Ωh ) q k∇h qk ψ kψkH(curl,Ωh )
| {z } | {z }
controlled minimized

since ∇h H 1 (Ωh ) ⊂ H(curl, Ωh ).

Texas A&M, May 2015 DPG Method 129 / 131


Take home messages
I Different variational formulations are possible. One can implement DPG
within all of them. Different functional setting leads to (optimal)
convergence in different norms.
I Do not confuse functional setting with selection of concrete test norm.
Different test norms lead to different versions of DPG. Selecting a right test
norm is the whole ball game...
I DPG reproduces stability properties from the continuous level.
I Coding DPG method with a code supporting the exact sequence is
straightforward: DPG traces (fluxes) are discretized with traces of
H 1 , H(curl), and H(div) elements.
I Being a minimum residual method, DPG comes with a-posteriori error
estimate built in. One can exercise adaptivity from day one.
I Being a Ritz method, DPG does not suffer from any preasymptotic behavior.
One can start adaptivity from very coarse meshes.
In progress:
I Speeding up element computations (fast integration, GPU implementation,
precomputing).
I Integration of solvers with adaptivity (iterative “multigrid” solvers).
Texas A&M, May 2015 DPG Method 130 / 131
Acknowledgements

AFOSR grant FA9550-12-1-0484


NSF grant DMS-1418822
SANDIA grant 1536119

Thank you

Texas A&M, May 2015 DPG Method 131 / 131

You might also like