Use of explicit inf-sup operators

to solve indefinite problems

Patrick Ciarlet
online access to recent Refs: http:/www.ensta.fr/˜ciarlet

POEMS, UMR 7231 CNRS-ENSTA-INRIA

Nancy, April 2012 – p.1/24

 Outline

Well-posedness with the help of explicit inf-sup operators: T-coercivity.

Numerical approximation and convergence via T-coercivity.

Helmholtz equation in acoustics.

Time-harmonic problems in electromagnetics.

Transmission problems with sign changing coefficients.

Conclusion.

Nancy, April 2012 – p.2/24

 Abstract setting
Let
V and W be two Hilbert spaces ;
a(·, ·) be a continuous sesquilinear form over V × W ;
f be an element of W ′ , the dual space of W .
Aim: solve the Variational Formulation

(V F ) Find u ∈ V s.t. ∀w ∈ W, a(u, w) = hf, wi.

Nancy, April 2012 – p.3/24

 Abstract setting
Let
V and W be two Hilbert spaces ;
a(·, ·) be a continuous sesquilinear form over V × W ;
f be an element of W ′ , the dual space of W .
Aim: solve the Variational Formulation

(V F ) Find u ∈ V s.t. ∀w ∈ W, a(u, w) = hf, wi.

[Hadamard] The Problem (V F ) is well-posed if, and only if, for all f , it has one and
only one solution u, with continuous dependence:

∃C > 0, ∀f ∈ W ′ , kukV ≤ C kf kW ′ .

Nancy, April 2012 – p.3/24

 Abstract setting
Let
V and W be two Hilbert spaces ;
a(·, ·) be a continuous sesquilinear form over V × W ;
f be an element of W ′ , the dual space of W .
Aim: solve the Variational Formulation

(V F ) Find u ∈ V s.t. ∀w ∈ W, a(u, w) = hf, wi.

[Hadamard] The Problem (V F ) is well-posed if, and only if, for all f , it has one and
only one solution u, with continuous dependence:

∃C > 0, ∀f ∈ W ′ , kukV ≤ C kf kW ′ .

How can one prove well-posedness?

Nancy, April 2012 – p.3/24

 Abstract setting
Let
V and W be two Hilbert spaces ;
a(·, ·) be a continuous sesquilinear form over V × W ;
f be an element of W ′ , the dual space of W .
Aim: solve the Variational Formulation

(V F ) Find u ∈ V s.t. ∀w ∈ W, a(u, w) = hf, wi.

[Hadamard] The Problem (V F ) is well-posed if, and only if, for all f , it has one and
only one solution u, with continuous dependence:

∃C > 0, ∀f ∈ W ′ , kukV ≤ C kf kW ′ .

How can one prove well-posedness?

[Lax-Milgram] OK provided that a(·, ·) is coercive!

Nancy, April 2012 – p.3/24

 Abstract setting-2
[Banach-Necas-Babuska] Introduce the two conditions

|a(v, w)|
(BN B1 ) ∃α′ > 0, ∀v ∈ V, sup ≥ α′ kvkV .
w∈W \{0} kwkW

(BN B2 ) ∀w ∈ W : {∀v ∈ V, a(v, w) = 0} =⇒ {w = 0}.

NB. Condition (BN B1 ) is called an inf-sup condition, or a stability condition.

Nancy, April 2012 – p.4/24

 Abstract setting-2
[Banach-Necas-Babuska] Introduce the two conditions

|a(v, w)|
(BN B1 ) ∃α′ > 0, ∀v ∈ V, sup ≥ α′ kvkV .
w∈W \{0} kwkW

(BN B2 ) ∀w ∈ W : {∀v ∈ V, a(v, w) = 0} =⇒ {w = 0}.

NB. Condition (BN B1 ) is called an inf-sup condition, or a stability condition.

Theorem (Well-posedness) The two assertions below are equivalent:
(i) the Problem (V F ) is well-posed ;
(ii) the form a(·, ·) satisfies conditions (BN B1 ) and (BN B2 ).

Nancy, April 2012 – p.4/24

 Abstract setting-2
[Banach-Necas-Babuska] Introduce the two conditions

|a(v, w)|
(BN B1 ) ∃α′ > 0, ∀v ∈ V, sup ≥ α′ kvkV .
w∈W \{0} kwkW

(BN B2 ) ∀w ∈ W : {∀v ∈ V, a(v, w) = 0} =⇒ {w = 0}.

NB. Condition (BN B1 ) is called an inf-sup condition, or a stability condition.

Definition (T-coercivity) The form a(·, ·) is T-coercive if

∃T ∈ L(V, W ), bijective, ∃α > 0, ∀v ∈ V, |a(v, Tv)| ≥ α kvk2V .

NB. In other words, the form (v, v ′ ) 7→ a(v, Tv ′ ) is coercive on V × V .

Nancy, April 2012 – p.4/24

 Abstract setting-2
[Banach-Necas-Babuska] Introduce the two conditions

|a(v, w)|
(BN B1 ) ∃α′ > 0, ∀v ∈ V, sup ≥ α′ kvkV .
w∈W \{0} kwkW

(BN B2 ) ∀w ∈ W : {∀v ∈ V, a(v, w) = 0} =⇒ {w = 0}.

NB. Condition (BN B1 ) is called an inf-sup condition, or a stability condition.

Definition (T-coercivity) The form a(·, ·) is T-coercive if

∃T ∈ L(V, W ), bijective, ∃α > 0, ∀v ∈ V, |a(v, Tv)| ≥ α kvk2V .

Theorem (Well-posedness) The three assertions below are equivalent:

(i) the Problem (V F ) is well-posed ;
(ii) the form a(·, ·) satisfies conditions (BN B1 ) and (BN B2 ).
(iii) the form a(·, ·) is T-coercive.

Nancy, April 2012 – p.4/24

 Abstract setting-2
[Banach-Necas-Babuska] Introduce the two conditions

|a(v, w)|
(BN B1 ) ∃α′ > 0, ∀v ∈ V, sup ≥ α′ kvkV .
w∈W \{0} kwkW

(BN B2 ) ∀w ∈ W : {∀v ∈ V, a(v, w) = 0} =⇒ {w = 0}.

NB. Condition (BN B1 ) is called an inf-sup condition, or a stability condition.

Definition (T-coercivity) The form a(·, ·) is T-coercive if

∃T ∈ L(V, W ), bijective, ∃α > 0, ∀v ∈ V, |a(v, Tv)| ≥ α kvk2V .

Theorem (Well-posedness) The three assertions below are equivalent:

(i) the Problem (V F ) is well-posed ;
(ii) the form a(·, ·) satisfies conditions (BN B1 ) and (BN B2 ).
(iii) the form a(·, ·) is T-coercive.

The operator T realizes conditions (BN B1 ) and (BN B2 ) explicitly.

Nancy, April 2012 – p.4/24

 Abstract setting-3
V = W , case of a hermitian form a

The previous definition and theorem can be simplified...

Nancy, April 2012 – p.5/24

 Abstract setting-3
V = W , case of a hermitian form a

The previous definition and theorem can be simplified...

Definition (T-coercivity) The hermitian form a(·, ·) is T-coercive if

∃T ∈ L(V ), ∃α > 0, ∀v ∈ V, |a(v, Tv)| ≥ α kvk2V .

Nancy, April 2012 – p.5/24

 Abstract setting-3
V = W , case of a hermitian form a

The previous definition and theorem can be simplified...

Definition (T-coercivity) The hermitian form a(·, ·) is T-coercive if

∃T ∈ L(V ), ∃α > 0, ∀v ∈ V, |a(v, Tv)| ≥ α kvk2V .

Theorem (Well-posedness) The three assertions below are equivalent:

(i) the Problem (V F ) with hermitian form is well-posed ;
(ii) the hermitian form a(·, ·) satisfies condition (BN B1 ).
(iii) the hermitian form a(·, ·) is T-coercive.

Nancy, April 2012 – p.5/24

 Numerical approximation
Conforming discretization:
let (Vh )h be finite dimensional vector subspaces of V (limh→0 dim(Vh ) = +∞) ;
let (Wh )h be finite dimensional vector subspaces of W (limh→0 dim(Wh ) = +∞).
Aim: solve the Discrete Variational Formulation

(DV F ) Find uh ∈ Vh s.t. ∀wh ∈ Wh , a(uh , wh ) = hf, wh i.

Nancy, April 2012 – p.6/24

 Numerical approximation
Conforming discretization:
let (Vh )h be finite dimensional vector subspaces of V (limh→0 dim(Vh ) = +∞) ;
let (Wh )h be finite dimensional vector subspaces of W (limh→0 dim(Wh ) = +∞).
Aim: solve the Discrete Variational Formulation

(DV F ) Find uh ∈ Vh s.t. ∀wh ∈ Wh , a(uh , wh ) = hf, wh i.

NB. For simplicity, the discrete forms are assumed to be exact.

Nancy, April 2012 – p.6/24

 Numerical approximation
Conforming discretization:
let (Vh )h be finite dimensional vector subspaces of V (limh→0 dim(Vh ) = +∞) ;
let (Wh )h be finite dimensional vector subspaces of W (limh→0 dim(Wh ) = +∞).
Aim: solve the Discrete Variational Formulation

(DV F ) Find uh ∈ Vh s.t. ∀wh ∈ Wh , a(uh , wh ) = hf, wh i.

[Babuska-Brezzi] Introduce the uniform discrete inf-sup condition

|a(vh , wh )|
(U DISC) ∃α† > 0, ∀h > 0, ∀vh ∈ Vh , sup ≥ α† kvh kV .
wh ∈Wh \{0} kwh kW

Nancy, April 2012 – p.6/24

 Numerical approximation
Conforming discretization:
let (Vh )h be finite dimensional vector subspaces of V (limh→0 dim(Vh ) = +∞) ;
let (Wh )h be finite dimensional vector subspaces of W (limh→0 dim(Wh ) = +∞).
Aim: solve the Discrete Variational Formulation

(DV F ) Find uh ∈ Vh s.t. ∀wh ∈ Wh , a(uh , wh ) = hf, wh i.

[Babuska-Brezzi] Introduce the uniform discrete inf-sup condition

|a(vh , wh )|
(U DISC) ∃α† > 0, ∀h > 0, ∀vh ∈ Vh , sup ≥ α† kvh kV .
wh ∈Wh \{0} kwh kW

Definition (Th -coercivity) The form a(·, ·) is uniformly Th -coercive if

∃α⋆ , β ⋆ > 0, ∀h > 0, ∃Th ∈ L(Vh , Wh ), ∀vh ∈ Vh ,

|a(vh , Th vh )| ≥ α⋆ kvh k2V and |||Th ||| ≤ β ⋆ .

Nancy, April 2012 – p.6/24

 Numerical approximation-2
Theorem (approximation error) The three assertions below are equivalent:
(i) Problems (DV F ) are well-posed with uniform continuous dependence ;
(ii) the form a(·, ·) satisfies the uniform discrete inf-sup condition (U DISC) ;
(iii) the form a(·, ·) is uniformly Th -coercive.
If one of these conditions is satisfied, the error ku − uh kV is bounded by

(Strang) ku − uh kV ≤ C inf ku − vh kV ,
vh ∈Vh

with C independent of f and h.

Nancy, April 2012 – p.7/24

 Numerical approximation-2
Theorem (approximation error) The three assertions below are equivalent:
(i) Problems (DV F ) are well-posed with uniform continuous dependence ;
(ii) the form a(·, ·) satisfies the uniform discrete inf-sup condition (U DISC) ;
(iii) the form a(·, ·) is uniformly Th -coercive.
If one of these conditions is satisfied, the error ku − uh kV is bounded by

(Strang) ku − uh kV ≤ C inf ku − vh kV ,
vh ∈Vh

with C independent of f and h.

Proposition (Th -coercivity)
Assume
∃T ∈ L(V, W ), bijective, such that (v, v ′ ) 7→ a(v, Tv ′ ) is coercive on V × V ;
“ ”
||(Th −T)(vh )||W
∃(Th )h , Th ∈ L(Vh , Wh ) s.t. limh→0 supvh ∈Vh \{0} ||v ||
= 0.
h V
Then, the form a(·, ·) is uniformly Th -coercive for h small enough.

Nancy, April 2012 – p.7/24

 Numerical approximation-2
Theorem (approximation error) The three assertions below are equivalent:
(i) Problems (DV F ) are well-posed with uniform continuous dependence ;
(ii) the form a(·, ·) satisfies the uniform discrete inf-sup condition (U DISC) ;
(iii) the form a(·, ·) is uniformly Th -coercive.
If one of these conditions is satisfied, the error ku − uh kV is bounded by

(Strang) ku − uh kV ≤ C inf ku − vh kV ,
vh ∈Vh

with C independent of f and h.

Proposition (Th -coercivity)
Assume
∃T ∈ L(V, W ), bijective, such that (v, v ′ ) 7→ a(v, Tv ′ ) is coercive on V × V ;
“ ”
||(Th −T)(vh )||W
∃(Th )h , Th ∈ L(Vh , Wh ) s.t. limh→0 supvh ∈Vh \{0} ||v ||
= 0.
h V
Then, the form a(·, ·) is uniformly Th -coercive for h small enough.
Similar approach, see [Buffa-Costabel-Schwab’02] for BEM.
Non-conforming discretization, see [Chung-Jr’1x] for DG.

Nancy, April 2012 – p.7/24

Helmholtz equation in acoustics
Consider a bounded domain Ω of Rd , with d = 1, 2, 3.
We study the classical problem
8
1
< Find u ∈ H (Ω) such that
>
>
div (σ∇u) + ω 2 ηu = f in Ω
>
>
u = 0 on ∂Ω.
:

Above, f is a source, ω > 0 is the given pulsation.

σ, η ∈ L∞ (Ω), and ∃σ− , η− > 0 such that σ > σ− and η > η− a.e. in Ω.
NB. Other boundary conditions are possible...

Nancy, April 2012 – p.8/24

Helmholtz equation in acoustics
Consider a bounded domain Ω of Rd , with d = 1, 2, 3.
We study the classical problem
8
< Find u ∈ H01 (Ω) such that
Z Z
: σ∇u · ∇v dΩ − ω 2 ηuv dΩ = −hf, vi, ∀v ∈ H01 (Ω).
Ω Ω

Above, f ∈ H −1 (Ω).

Nancy, April 2012 – p.8/24

Helmholtz equation in acoustics
Consider a bounded domain Ω of Rd , with d = 1, 2, 3.
We study the classical problem
8
< Find u ∈ H01 (Ω) such that
Z Z
: σ∇u · ∇v dΩ − ω 2 ηuv dΩ = −hf, vi, ∀v ∈ H01 (Ω).
Ω Ω

Within our framework:

V = W = H01 (Ω).
Z
aac (v, w) = (σ∇v · ∇w − ω 2 ηvw) dΩ.
Ω

How can one achieve T-coercivity of the form aac (·, ·)?

Nancy, April 2012 – p.8/24

Helmholtz equation in acoustics
Consider a bounded domain Ω of Rd , with d = 1, 2, 3.
We study the classical problem
8
< Find u ∈ H01 (Ω) such that
Z Z
: σ∇u · ∇v dΩ − ω 2 ηuv dΩ = −hf, vi, ∀v ∈ H01 (Ω).
Ω Ω

Within our framework:

V = W = H01 (Ω).
Z
aac (v, w) = (σ∇v · ∇w − ω 2 ηvw) dΩ.
Ω

How can one achieve T-coercivity of the form aac (·, ·)?

Choose the norms:

„Z «1/2
v 7→ kvk0 := ηv 2 dΩ in L2 (Ω).
Ω
„Z Z «1/2
v 7→ kvk1 := ηv 2 dΩ + σ|∇v|2 dΩ in H 1 (Ω).
Ω Ω

Nancy, April 2012 – p.8/24

Helmholtz equation in acoustics-2
Spectral Theorem: ∃(vℓ )ℓ≥0 , a Hilbert basis of H01 (Ω) made up of eigenfunctions

8
< Find (vℓ , λℓ ) ∈ H01 (Ω) × R such that vℓ 6= 0 and
Z Z
: σ∇vℓ · ∇w dΩ = λℓ ηvℓ w dΩ, ∀w ∈ H01 (Ω).
Ω Ω

In addition
(vℓ )ℓ≥0 is also an orthogonal basis of L2 (Ω) ;
all eigenvalues are of finite multiplicity ;
λ0 > 0, and limℓ→∞ λℓ = +∞.
NB. The eigenpairs are ordered by increasing values of the eigenvalues.

Nancy, April 2012 – p.9/24

Helmholtz equation in acoustics-2
Spectral Theorem: ∃(vℓ )ℓ≥0 , a Hilbert basis of H01 (Ω) made up of eigenfunctions

8
< Find (vℓ , λℓ ) ∈ H01 (Ω) × R such that vℓ 6= 0 and
Z Z
: σ∇vℓ · ∇w dΩ = λℓ ηvℓ w dΩ, ∀w ∈ H01 (Ω).
Ω Ω

Choice of Tac :
Let ℓmax denote the largest index ℓ ≥ 0 such that λℓ < ω 2 . Introduce:
V − := span0≤ℓ≤ℓmax (vℓ ), a finite dimensional vector subspace of H01 (Ω) ;
the orthogonal projection operator P− from H01 (Ω) to V − .
NB. When ω 2 is smaller than λ0 , ℓmax = −1, V − = {0} and P− = 0...

Nancy, April 2012 – p.9/24

Helmholtz equation in acoustics-2
Spectral Theorem: ∃(vℓ )ℓ≥0 , a Hilbert basis of H01 (Ω) made up of eigenfunctions

8
< Find (vℓ , λℓ ) ∈ H01 (Ω) × R such that vℓ 6= 0 and
Z Z
: σ∇vℓ · ∇w dΩ = λℓ ηvℓ w dΩ, ∀w ∈ H01 (Ω).
Ω Ω

Choice of Tac :
Let ℓmax denote the largest index ℓ ≥ 0 such that λℓ < ω 2 . Introduce:
V − := span0≤ℓ≤ℓmax (vℓ ), a finite dimensional vector subspace of H01 (Ω) ;
the orthogonal projection operator P− from H01 (Ω) to V − .
NB. When ω 2 is smaller than λ0 , ℓmax = −1, V − = {0} and P− = 0...
Define Tac := IH 1 (Ω) − 2P− :
0

8
< −v if 0 ≤ ℓ ≤ ℓ
ac ℓ max
T vℓ :=
: +vℓ if ℓ > ℓmax .

Nancy, April 2012 – p.9/24

Helmholtz equation in acoustics-2
Spectral Theorem: ∃(vℓ )ℓ≥0 , a Hilbert basis of H01 (Ω) made up of eigenfunctions

8
< Find (vℓ , λℓ ) ∈ H01 (Ω) × R such that vℓ 6= 0 and
Z Z
: σ∇vℓ · ∇w dΩ = λℓ ηvℓ w dΩ, ∀w ∈ H01 (Ω).
Ω Ω

Choice of Tac :
Let ℓmax denote the largest index ℓ ≥ 0 such that λℓ < ω 2 . Introduce:
V − := span0≤ℓ≤ℓmax (vℓ ), a finite dimensional vector subspace of H01 (Ω) ;
the orthogonal projection operator P− from H01 (Ω) to V − .
NB. When ω 2 is smaller than λ0 , ℓmax = −1, V − = {0} and P− = 0...
Define Tac := IH 1 (Ω) − 2P− .
0
Z
Proposition aac : (v, w) 7→ (σ∇v · ∇w − ω 2 ηvw) dΩ is T-coercive:
Ω

˛ λℓ − ω 2 ˛
˛ ˛
∀v ∈ H01 (Ω), |aac (v, Tac v)| ≥ α kvk2V , with α := min ˛˛ ˛.
ℓ≥0 1 + λℓ ˛

Nancy, April 2012 – p.9/24

Helmholtz equation in acoustics-3
Conforming discretization: Lagrange finite elements =⇒ (Vh )h ...
The Discrete Variational Formulation writes:

Find uh ∈ Vh s.t. aac (uh , vh ) = −hf, vh i, ∀vh ∈ Vh .

How can one achieve the uniform Th -coercivity of the form aac (·, ·)?

Nancy, April 2012 – p.10/24

Helmholtz equation in acoustics-3
Conforming discretization: Lagrange finite elements =⇒ (Vh )h ...
The Discrete Variational Formulation writes:

Find uh ∈ Vh s.t. aac (uh , vh ) = −hf, vh i, ∀vh ∈ Vh .

How can one achieve the uniform Th -coercivity of the form aac (·, ·)?

Idea (simple!): Build a suitable approximation of V − in Vh .

Choose approximations (vℓ,h )0≤ℓ≤ℓmax of the basis vectors (vℓ )0≤ℓ≤ℓmax , and set

Vh− := span0≤ℓ≤ℓmax (vℓ,h ).

Nancy, April 2012 – p.10/24

Helmholtz equation in acoustics-3
Conforming discretization: Lagrange finite elements =⇒ (Vh )h ...
The Discrete Variational Formulation writes:

Find uh ∈ Vh s.t. aac (uh , vh ) = −hf, vh i, ∀vh ∈ Vh .

How can one achieve the uniform Th -coercivity of the form aac (·, ·)?

Idea (simple!): Build a suitable approximation of V − in Vh .

Because V − is finite dimensional, one can find, for h small enough, a sequence of
orthonormal families (vℓ,h )0≤ℓ≤ℓmax ,h and a uniform bound δ (limh→0 δ(h) = 0) s.t.

kvℓ − vℓ,h k1 ≤ δ(h), 0 ≤ ℓ ≤ ℓmax , for h small enough.

Nancy, April 2012 – p.10/24

Helmholtz equation in acoustics-3
Conforming discretization: Lagrange finite elements =⇒ (Vh )h ...
The Discrete Variational Formulation writes:

Find uh ∈ Vh s.t. aac (uh , vh ) = −hf, vh i, ∀vh ∈ Vh .

How can one achieve the uniform Th -coercivity of the form aac (·, ·)?

Idea (simple!): Build a suitable approximation of V − in Vh .

Because V − is finite dimensional, one can find, for h small enough, a sequence of
orthonormal families (vℓ,h )0≤ℓ≤ℓmax ,h and a uniform bound δ (limh→0 δ(h) = 0) s.t.

kvℓ − vℓ,h k1 ≤ δ(h), 0 ≤ ℓ ≤ ℓmax , for h small enough.

Introduce:
the orthogonal projection operator P− −
h from Vh to Vh = span0≤ℓ≤ℓmax (vℓ,h ) ;
−
the operator Tac
h := IVh − 2Ph of L(Vh ).

Nancy, April 2012 – p.10/24

Helmholtz equation in acoustics-3
Conforming discretization: Lagrange finite elements =⇒ (Vh )h ...
The Discrete Variational Formulation writes:

Find uh ∈ Vh s.t. aac (uh , vh ) = −hf, vh i, ∀vh ∈ Vh .

How can one achieve the uniform Th -coercivity of the form aac (·, ·)?

Idea (simple!): Build a suitable approximation of V − in Vh .

Because V − is finite dimensional, one can find, for h small enough, a sequence of
orthonormal families (vℓ,h )0≤ℓ≤ℓmax ,h and a uniform bound δ (limh→0 δ(h) = 0) s.t.

kvℓ − vℓ,h k1 ≤ δ(h), 0 ≤ ℓ ≤ ℓmax , for h small enough.

Introduce:
the orthogonal projection operator P− −
h from Vh to Vh = span0≤ℓ≤ℓmax (vℓ,h ) ;
−
the operator Tac
h := IVh − 2Ph of L(Vh ).
||(Tac ac )(v )||
“ ”
h −T h 1
Proposition There holds limh→0 supvh ∈Vh \{0} ||vh ||1
= 0.
Hence, the discrete solution uh converges to u, with a rate governed by (Strang).

Nancy, April 2012 – p.10/24

Time-harmonic problem in EM-ics
Consider a bounded domain Ω of R3 .
We study the classical problem
8
< Find e ∈ H(curl; Ω) such that
>
>
−ω 2 εe + curl(µ−1 curl e) = f in Ω
>
>
e × n = 0 on ∂Ω.
:

Above, f is a source, ω > 0 is the given pulsation.

ε, µ ∈ L∞ (Ω), and ∃ε− , µ− > 0 such that ε > ε− and µ > µ− a.e. in Ω.
NB. Other boundary conditions are possible...

Nancy, April 2012 – p.11/24

Time-harmonic problem in EM-ics
Consider a bounded domain Ω of R3 .
We study the classical problem
8
< Find e ∈ H0 (curl; Ω) such that
Z Z Z
: µ−1 curl e · curl v dΩ − ω 2 εe · v dΩ = f · v dΩ, ∀v ∈ H0 (curl; Ω).
Ω Ω Ω

Above, f ∈ L2 (Ω).

Nancy, April 2012 – p.11/24

Time-harmonic problem in EM-ics
Consider a bounded domain Ω of R3 .
We study the classical problem
8
< Find e ∈ H0 (curl; Ω) such that
Z Z Z
: µ−1 curl e · curl v dΩ − ω 2 εe · v dΩ = f · v dΩ, ∀v ∈ H0 (curl; Ω).
Ω Ω Ω

Within our framework:

V = W = H0 (curl; Ω).
Z
aEM (v, w) = (µ−1 curl v · curl w − ω 2 εv · w) dΩ.
Ω

How can one achieve T-coercivity of the form aEM (·, ·)?

Nancy, April 2012 – p.11/24

Time-harmonic problem in EM-ics
Consider a bounded domain Ω of R3 .
We study the classical problem
8
< Find e ∈ H0 (curl; Ω) such that
Z Z Z
: µ−1 curl e · curl v dΩ − ω 2 εe · v dΩ = f · v dΩ, ∀v ∈ H0 (curl; Ω).
Ω Ω Ω

Within our framework:

V = W = H0 (curl; Ω).
Z
aEM (v, w) = (µ−1 curl v · curl w − ω 2 εv · w) dΩ.
Ω

How can one achieve T-coercivity of the form aEM (·, ·)?

Choose the norms:

„Z «1/2
v 7→ kvk0 := ε|v|2 dΩ in L2 (Ω).
Ω
„Z Z «1/2
v 7→ kvkcurl := ε|v|2 dΩ + µ−1 | curl v|2 dΩ in H(curl; Ω).
Ω Ω

Nancy, April 2012 – p.11/24

Time-harmonic problem in EM-ics-2
DIFFICULTY: the embedding of H0 (curl; Ω) into L2 (Ω) is not compact!
Hence, the Spectral Theorem can not be applied “as is”...

Nancy, April 2012 – p.12/24

Time-harmonic problem in EM-ics-2
DIFFICULTY: the embedding of H0 (curl; Ω) into L2 (Ω) is not compact!
Hence, the Spectral Theorem can not be applied “as is”...
Proposition There holds the decomposition

⊥curl
H0 (curl; Ω) = G ⊕ Wε
where G := ∇H01 (Ω), W ε := {w ∈ H0 (curl; Ω) : div (εw) = 0}.

Nancy, April 2012 – p.12/24

Time-harmonic problem in EM-ics-2
DIFFICULTY: the embedding of H0 (curl; Ω) into L2 (Ω) is not compact!
Hence, the Spectral Theorem can not be applied “as is”...
Proposition There holds the decomposition

⊥curl
H0 (curl; Ω) = G ⊕ Wε
where G := ∇H01 (Ω), W ε := {w ∈ H0 (curl; Ω) : div (εw) = 0}.

Idea: one can try and build two Hilbert bases:

one for G (cf. acoustics section): (eℓ )ℓ<0 , with eℓ := ∇v−(1+ℓ) for ℓ < 0 ;
one for W ε .

Nancy, April 2012 – p.12/24

Time-harmonic problem in EM-ics-2
DIFFICULTY: the embedding of H0 (curl; Ω) into L2 (Ω) is not compact!
Hence, the Spectral Theorem can not be applied “as is”...
Proposition There holds the decomposition

⊥curl
H0 (curl; Ω) = G ⊕ Wε
where G := ∇H01 (Ω), W ε := {w ∈ H0 (curl; Ω) : div (εw) = 0}.

(eℓ )ℓ<0 Hilbert basis of G, with eℓ := ∇v−(1+ℓ) for ℓ < 0.

Theorem [Weber’80] W ε is compactly embedded into L2 (Ω).
DIFFICULTY: W ε is not dense in L2 (Ω).

Nancy, April 2012 – p.12/24

Time-harmonic problem in EM-ics-2
DIFFICULTY: the embedding of H0 (curl; Ω) into L2 (Ω) is not compact!
Hence, the Spectral Theorem can not be applied “as is”...
Proposition There holds the decomposition

⊥curl
H0 (curl; Ω) = G ⊕ Wε
where G := ∇H01 (Ω), W ε := {w ∈ H0 (curl; Ω) : div (εw) = 0}.

(eℓ )ℓ<0 Hilbert basis of G, with eℓ := ∇v−(1+ℓ) for ℓ < 0.

Theorem [Weber’80] W ε is compactly embedded into L2 (Ω).
DIFFICULTY: W ε is not dense in L2 (Ω).
New pivot space: H(div ε0; Ω) := {w ∈ H(div ε; Ω) : div (εw) = 0}.
(+) W ε is compactly embedded into H(div ε0; Ω) ;
(+) one can prove that W ε is dense in H(div ε0; Ω).

Nancy, April 2012 – p.12/24

Time-harmonic problem in EM-ics-2
DIFFICULTY: the embedding of H0 (curl; Ω) into L2 (Ω) is not compact!
Hence, the Spectral Theorem can not be applied “as is”...
Proposition There holds the decomposition

⊥curl
H0 (curl; Ω) = G ⊕ Wε
where G := ∇H01 (Ω), W ε := {w ∈ H0 (curl; Ω) : div (εw) = 0}.

(eℓ )ℓ<0 Hilbert basis of G, with eℓ := ∇v−(1+ℓ) for ℓ < 0.

Spectral Theorem: ∃(eℓ )ℓ≥0 a Hilbert basis of W ε made up of eigenfunctions

8
< Find (eℓ , µℓ ) ∈ W ε × R such that eℓ 6= 0 and
Z Z
: (εeℓ · w + µ−1 curl eℓ · curl w) dΩ = (1 + µℓ ) εeℓ · w dΩ, ∀w ∈ W ε .
Ω Ω

all eigenvalues are of finite multiplicity ;

µℓ = 0 occurs K times, with K + 1 number of c.c. of ∂Ω, and limℓ→∞ µℓ = +∞.
NB. The eigenpairs are ordered by increasing values of the eigenvalues.

Nancy, April 2012 – p.12/24

Time-harmonic problem in EM-ics-3
Conclusion: (eℓ )ℓ is a Hilbert basis of H0 (curl; Ω) such that
Z
∀ℓ, ∃µℓ ≥ 0, (eℓ , w)curl = (1 + µℓ ) εeℓ · w dΩ, ∀w ∈ H0 (curl; Ω).
Ω

For ℓ < 0: eℓ ∈ G and µℓ = 0 ;

For ℓ ≥ 0: eℓ ∈ W ε and µℓ are eigenpairs, and
all eigenvalues are of finite multiplicity ;
µℓ = 0 occurs K times, and limℓ→∞ µℓ = +∞.

Nancy, April 2012 – p.13/24

Time-harmonic problem in EM-ics-3
Conclusion: (eℓ )ℓ is a Hilbert basis of H0 (curl; Ω) such that
Z
∀ℓ, ∃µℓ ≥ 0, (eℓ , w)curl = (1 + µℓ ) εeℓ · w dΩ, ∀w ∈ H0 (curl; Ω).
Ω

NB. Given any ω > 0, there is an infinite number of ℓ s.t. µℓ < ω 2 .

Nancy, April 2012 – p.13/24

Time-harmonic problem in EM-ics-3
Conclusion: (eℓ )ℓ is a Hilbert basis of H0 (curl; Ω) such that
Z
∀ℓ, ∃µℓ ≥ 0, (eℓ , w)curl = (1 + µℓ ) εeℓ · w dΩ, ∀w ∈ H0 (curl; Ω).
Ω

NB. Given any ω > 0, there is an infinite number of ℓ s.t. µℓ < ω 2 .

Choice of TEM :
Let ℓmax denote the largest index ℓ such that µℓ < ω 2 . Introduce:
V − := span0≤ℓ≤ℓmax (eℓ ), a finite dimensional vector subspace of W ε ;
the orthogonal projection operator P− from H0 (curl; Ω) to V − .

Nancy, April 2012 – p.13/24

Time-harmonic problem in EM-ics-3
Conclusion: (eℓ )ℓ is a Hilbert basis of H0 (curl; Ω) such that
Z
∀ℓ, ∃µℓ ≥ 0, (eℓ , w)curl = (1 + µℓ ) εeℓ · w dΩ, ∀w ∈ H0 (curl; Ω).
Ω

NB. Given any ω > 0, there is an infinite number of ℓ s.t. µℓ < ω 2 .

Choice of TEM :
Let ℓmax denote the largest index ℓ such that µℓ < ω 2 . Introduce:
V − := span0≤ℓ≤ℓmax (eℓ ), a finite dimensional vector subspace of W ε ;
the orthogonal projection operator P− from H0 (curl; Ω) to V − .
Define TEM := −iG + iW ε − 2P− :
8
< −e if ℓ ≤ ℓ
EM ℓ max
T eℓ :=
: +eℓ if ℓ > ℓmax .

Nancy, April 2012 – p.13/24

Time-harmonic problem in EM-ics-3
Conclusion: (eℓ )ℓ is a Hilbert basis of H0 (curl; Ω) such that
Z
∀ℓ, ∃µℓ ≥ 0, (eℓ , w)curl = (1 + µℓ ) εeℓ · w dΩ, ∀w ∈ H0 (curl; Ω).
Ω

NB. Given any ω > 0, there is an infinite number of ℓ s.t. µℓ < ω 2 .

Choice of TEM :
Let ℓmax denote the largest index ℓ such that µℓ < ω 2 . Introduce:
V − := span0≤ℓ≤ℓmax (eℓ ), a finite dimensional vector subspace of W ε ;
the orthogonal projection operator P− from H0 (curl; Ω) to V − .
Define TEM := −iG + iW ε − 2P− .
Z
Proposition aEM : (v, w) 7→ (µ−1 curl v · curl w − ω 2 εv · w) dΩ is T-coercive.
Ω

Nancy, April 2012 – p.13/24

Time-harmonic problem in EM-ics-4
Conforming discretization: Nédélec’s first family finite elements =⇒ (V h )h ...
The Discrete Variational Formulation writes:
Z
Find eh ∈ V h s.t. aEM (eh , vh ) = f · vh dΩ, ∀vh ∈ V h .
Ω

How can one achieve the uniform Th -coercivity of the form aEM (·, ·)?

Nancy, April 2012 – p.14/24

Time-harmonic problem in EM-ics-4
Conforming discretization: Nédélec’s first family finite elements =⇒ (V h )h ...
The Discrete Variational Formulation writes:
Z
Find eh ∈ V h s.t. aEM (eh , vh ) = f · vh dΩ, ∀vh ∈ V h .
Ω

How can one achieve the uniform Th -coercivity of the form aEM (·, ·)?

DIFFICULTY: Given any ω > 0, there is an infinite number of ℓ s.t. µℓ < ω 2 .

Nancy, April 2012 – p.14/24

Time-harmonic problem in EM-ics-4
Conforming discretization: Nédélec’s first family finite elements =⇒ (V h )h ...
The Discrete Variational Formulation writes:
Z
Find eh ∈ V h s.t. aEM (eh , vh ) = f · vh dΩ, ∀vh ∈ V h .
Ω

How can one achieve the uniform Th -coercivity of the form aEM (·, ·)?

DIFFICULTY: Given any ω > 0, there is an infinite number of ℓ s.t. µℓ < ω 2 .

Idea:
split elements of V h (≈ exact decomposition H0 (curl; Ω) = G ⊕ W ε ) ;
take the opposite of the gradient part ;
use the orthogonal projection on the other part (cf. acoustics section).

Nancy, April 2012 – p.14/24

Time-harmonic problem in EM-ics-4
Conforming discretization: Nédélec’s first family finite elements =⇒ (V h )h ...
The Discrete Variational Formulation writes:
Z
Find eh ∈ V h s.t. aEM (eh , vh ) = f · vh dΩ, ∀vh ∈ V h .
Ω

How can one achieve the uniform Th -coercivity of the form aEM (·, ·)?

DIFFICULTY: Given any ω > 0, there is an infinite number of ℓ s.t. µℓ < ω 2 .

Idea:
split elements of V h (≈ exact decomposition H0 (curl; Ω) = G ⊕ W ε ) ;
take the opposite of the gradient part ;
use the orthogonal projection on the other part (cf. acoustics section).
DIFFICULTY: The discrete splitting needs to be uniformly close to the exact splitting.

Nancy, April 2012 – p.14/24

Time-harmonic problem in EM-ics-5
Given vh ∈ V h :
the exact splitting is ∃!(ϕ, w) ∈ H01 (Ω) × W ε , vh = ∇ϕ + w.
a discrete splitting is (ϕh , wh ) ∈ Vh × V h , vh = ∇ϕh + wh .
NB. Provided the orders of FE are appropriately chosen, there holds ∇Vh ⊂ V h .

Nancy, April 2012 – p.15/24

Time-harmonic problem in EM-ics-5
Given vh ∈ V h :
the exact splitting is ∃!(ϕ, w) ∈ H01 (Ω) × W ε , vh = ∇ϕ + w.
a discrete splitting is (ϕh , wh ) ∈ Vh × V h , vh = ∇ϕh + wh .
NB. Provided the orders of FE are appropriately chosen, there holds ∇Vh ⊂ V h .
Proposition (Uniform discrete splittings)
Assume that ε is piecewise-constant: there exists a discrete splitting such that

k∇(ϕ − ϕh )kcurl = kw − wh kcurl ≤ Cr hs kvh kcurl ,

with s := s(Ω, ε) > 0, Cr > 0 independent of vh .

Nancy, April 2012 – p.15/24

Time-harmonic problem in EM-ics-5
Given vh ∈ V h :
the exact splitting is ∃!(ϕ, w) ∈ H01 (Ω) × W ε , vh = ∇ϕ + w.
a discrete splitting is (ϕh , wh ) ∈ Vh × V h , vh = ∇ϕh + wh .
NB. Provided the orders of FE are appropriately chosen, there holds ∇Vh ⊂ V h .
Proposition (Uniform discrete splittings)
Assume that ε is piecewise-constant: there exists a discrete splitting such that

k∇(ϕ − ϕh )kcurl = kw − wh kcurl ≤ Cr hs kvh kcurl ,

with s := s(Ω, ε) > 0, Cr > 0 independent of vh .

Proof (main ingredients!)
regular-singular splitting of elements of W ε , cf. [Costabel-Dauge-Nicaise’99] ;
edge element approximability of piecewise-smooth fields, cf. [Monk’03] ;
edge element interpolation of gradients, cf. [Nédélec’80].

Nancy, April 2012 – p.15/24

Time-harmonic problem in EM-ics-5
Given vh ∈ V h :
the exact splitting is ∃!(ϕ, w) ∈ H01 (Ω) × W ε , vh = ∇ϕ + w.
a discrete splitting is (ϕh , wh ) ∈ Vh × V h , vh = ∇ϕh + wh .
NB. Provided the orders of FE are appropriately chosen, there holds ∇Vh ⊂ V h .
Proposition (Uniform discrete splittings)
Assume that ε is piecewise-constant: there exists a discrete splitting such that

k∇(ϕ − ϕh )kcurl = kw − wh kcurl ≤ Cr hs kvh kcurl ,

with s := s(Ω, ε) > 0, Cr > 0 independent of vh .

Approximate V − in V h , cf. acoustics section: V −
h := span0≤ℓ≤ℓmax (eℓ,h ), with

keℓ − eℓ,h kcurl ≤ δ(h), 0 ≤ ℓ ≤ ℓmax , for h small enough ( lim δ(h) = 0).
h→0

Nancy, April 2012 – p.15/24

Time-harmonic problem in EM-ics-5
Given vh ∈ V h :
the exact splitting is ∃!(ϕ, w) ∈ H01 (Ω) × W ε , vh = ∇ϕ + w.
a discrete splitting is (ϕh , wh ) ∈ Vh × V h , vh = ∇ϕh + wh .
NB. Provided the orders of FE are appropriately chosen, there holds ∇Vh ⊂ V h .
Proposition (Uniform discrete splittings)
Assume that ε is piecewise-constant: there exists a discrete splitting such that

k∇(ϕ − ϕh )kcurl = kw − wh kcurl ≤ Cr hs kvh kcurl ,

with s := s(Ω, ε) > 0, Cr > 0 independent of vh .

Introduce:
the orthogonal projection operator P−
h from V h to V −
h ;
the operator TEM
h of L(V h ) defined by TEM
h (vh ) := −∇ϕh + (IV h − 2P−
h )(wh ).

Nancy, April 2012 – p.15/24

Time-harmonic problem in EM-ics-5
Given vh ∈ V h :
the exact splitting is ∃!(ϕ, w) ∈ H01 (Ω) × W ε , vh = ∇ϕ + w.
a discrete splitting is (ϕh , wh ) ∈ Vh × V h , vh = ∇ϕh + wh .
NB. Provided the orders of FE are appropriately chosen, there holds ∇Vh ⊂ V h .
Proposition (Uniform discrete splittings)
Assume that ε is piecewise-constant: there exists a discrete splitting such that

k∇(ϕ − ϕh )kcurl = kw − wh kcurl ≤ Cr hs kvh kcurl ,

with s := s(Ω, ε) > 0, Cr > 0 independent of vh .

Introduce:
the orthogonal projection operator P−
h from V h to V −
h ;
the operator TEM
h of L(V h ) defined by TEM
h (vh ) := −∇ϕh + (IV h − 2P− h )(wh ).
„ «
||(TEM
h −T EM
)(v h )||curl
Proposition There holds limh→0 supv h ∈V h \{0} ||v ||
= 0.
h curl

Hence, the discrete solution eh converges to e, with a rate governed by (Strang).

Nancy, April 2012 – p.15/24

 Sign-changing coefficients

Consider a scalar transmission problem, set in a bounded domain Ω of Rd , d = 1, 2, 3.

8
< Find u ∈ H 1 (Ω) such that
0
: div (σ∇u) = f in Ω.

8
< σ > 0 in Ω , with meas(Ω ) > 0 ;
1 1
σ ∈ L∞ (Ω) is a sign-changing coefficient:
: σ < 0 in Ω2 , with meas(Ω2 ) > 0.

σ −1 ∈ L∞ (Ω).

Nancy, April 2012 – p.16/24

 Sign-changing coefficients

Consider a scalar transmission problem, set in a bounded domain Ω of Rd , d = 1, 2, 3.

8
< Find u ∈ H 1 (Ω) such that
0
: div (σ∇u) = f in Ω.

σ ∈ L∞ (Ω), is a sign-changing coefficient.

σ −1 ∈ L∞ (Ω).
NB. The “generalized” Helmholtz equation div (σ∇u) + ω 2 ηu = f with η ∈ L∞ (Ω)
can be analyzed similarly, cf. [BonnetBenDhia-Jr-Zwölf’10].
When σ < 0, this models a metamaterial.
One can also consider a Neumann b.c., cf. [BonnetBenDhia-Chesnel-Jr’12].

Nancy, April 2012 – p.16/24

 Sign-changing coefficients

Consider a scalar transmission problem, set in a bounded domain Ω of Rd , d = 1, 2, 3.

8
< Find u ∈ H 1 (Ω) such that
0
: div (σ∇u) = f in Ω.

σ ∈ L∞ (Ω), is a sign-changing coefficient.

σ −1 ∈ L∞ (Ω).
Z
The main dificulty is that (v, w) 7→ σ ∇v · ∇w dΩ is not coercive in H01 (Ω).
Ω

Nancy, April 2012 – p.16/24

 Sign-changing coefficients

Consider a scalar transmission problem, set in a bounded domain Ω of Rd , d = 1, 2, 3.

8
< Find u ∈ H 1 (Ω) such that
0
: div (σ∇u) = f in Ω.

σ ∈ L∞ (Ω), is a sign-changing coefficient.

σ −1 ∈ L∞ (Ω).
Z
The main dificulty is that (v, w) 7→ σ ∇v · ∇w dΩ is not coercive in H01 (Ω).
Ω
Structure of spectrum? Use of the Spectral Theorem?
=⇒ New approach to achieve T-coercivity!

Nancy, April 2012 – p.16/24

 Sign-changing coefficients

Consider a scalar transmission problem, set in a bounded domain Ω of Rd , d = 1, 2, 3.

8
< Find u ∈ H 1 (Ω) such that
0
: div (σ∇u) = f in Ω.

σ ∈ L∞ (Ω), is a sign-changing coefficient.

σ −1 ∈ L∞ (Ω).
Z
The main dificulty is that (v, w) 7→ σ ∇v · ∇w dΩ is not coercive in H01 (Ω).
Ω
Structure of spectrum? Use of the Spectral Theorem?
=⇒ New approach to achieve T-coercivity!

We follow [BonnetBenDhia-Jr-Zwölf’10]:
Ω1 and Ω2 are domains of Rd ;
Σ := Ω1 ∩ Ω2 is the interface ;
Γk := ∂Ω ∩ ∂Ωk , k = 1, 2 are the boundaries.

Nancy, April 2012 – p.16/24

 Sign-changing coefficients-2

For the transmission problem with sign-changing coefficient:

V = H01 (Ω) ;
Z
the sesquilinear form is atr (v, w) = σ ∇v · ∇w dΩ.
Ω
NB. Complex-valued forms, to enable the introduction of dissipation...

Nancy, April 2012 – p.17/24

 Sign-changing coefficients-2

For the transmission problem with sign-changing coefficient:

V = H01 (Ω) ;
Z
the sesquilinear form is atr (v, w) = σ ∇v · ∇w dΩ.
Ω

Introduce Vk := {vk ∈ H 1 (Ωk ) | vk |Γk = 0}, k = 1, 2:

V = {v | v|Ωk ∈ Vk , k = 1, 2, MatchingΣ (v|Ω1 , v|Ω2 ) = 0} ,

with MatchingΣ (v1 , v2 ) := v1 |Σ − v2 |Σ .

Nancy, April 2012 – p.17/24

 Sign-changing coefficients-2

For the transmission problem with sign-changing coefficient:

V = H01 (Ω) ;
Z
the sesquilinear form is atr (v, w) = σ ∇v · ∇w dΩ.
Ω
Z
Introduce atr
k (vk , wk ) := σ ∇vk · ∇wk dΩ, k = 1, 2:
Ωk

∀v, w ∈ V, atr (v, w) = atr tr

1 (v|Ω1 , w|Ω1 ) + a2 (v|Ω2 , w|Ω2 ) ;

Nancy, April 2012 – p.17/24

 Sign-changing coefficients-2

For the transmission problem with sign-changing coefficient:

V = H01 (Ω) ;
Z
the sesquilinear form is atr (v, w) = σ ∇v · ∇w dΩ.
Ω
Z
Introduce atr
k (vk , wk ) := σ ∇vk · ∇wk dΩ, k = 1, 2:
Ωk

∀v, w ∈ V, atr (v, w) = atr tr

1 (v|Ω1 , w|Ω1 ) + a2 (v|Ω2 , w|Ω2 ) ;

∀v1 ∈ V1 , σ1− k∇v1 k2L 2 (Ω ≤ +atr

1 (v1 , v1 ) ≤ σ + 2
1 k∇v1 kL 2 (Ω ;
1) 1)

∀v2 ∈ V2 , σ2− k∇v2 k2L 2 (Ω ≤ −atr + 2

2 (v2 , v2 ) ≤ σ2 k∇v2 kL 2 (Ω ) .
2) 2

NB. We have 0 < σk− ≤ σk+ < ∞, k = 1, 2.

Nancy, April 2012 – p.17/24

 Sign-changing coefficients-3
First try:
8
< v in Ω1
1
∀v ∈ H01 (Ω), T− v := .
: −v2 in Ω2

NB. Given v ∈ H01 (Ω), we set vk := v|Ωk , k = 1, 2.

Nancy, April 2012 – p.18/24

 Sign-changing coefficients-3
First try:
8
< v in Ω1
1
∀v ∈ H01 (Ω), T− v := .
: −v2 in Ω2

(+) Obviously, (T− )2 = IH 1 (Ω) .

0

(–) But T− 6∈ L(H01 (Ω)), because the matching condition is not enforced.

Nancy, April 2012 – p.18/24

 Sign-changing coefficients-3
First try:
8
< v in Ω1
1
∀v ∈ H01 (Ω), T− v := .
: −v2 in Ω2

(+) Obviously, (T− )2 = IH 1 (Ω) .

0

(–) But T− 6∈ L(H01 (Ω)), because the matching condition is not enforced.

Second try: let R1 ∈ L(V1 , V2 ) s.t. for all v1 ∈ V1 , MatchingΣ (v1 , R1 v1 ) = 0.

8
< v in Ω1
1
∀v ∈ H01 (Ω), T v := .
: −v2 +2R1 v1 in Ω2

Nancy, April 2012 – p.18/24

 Sign-changing coefficients-3
First try:
8
< v in Ω1
1
∀v ∈ H01 (Ω), T− v := .
: −v2 in Ω2

(+) Obviously, (T− )2 = IH 1 (Ω) .

0

(–) But T− 6∈ L(H01 (Ω)), because the matching condition is not enforced.

Second try: let R1 ∈ L(V1 , V2 ) s.t. for all v1 ∈ V1 , MatchingΣ (v1 , R1 v1 ) = 0.

8
< v in Ω1
1
∀v ∈ H01 (Ω), T v := .
: −v2 +2R1 v1 in Ω2

(+) T ∈ L(H01 (Ω)).

(+) One checks easily that T2 = IH 1 (Ω) !
0

Nancy, April 2012 – p.18/24

 Sign-changing coefficients-3
First try:
8
< v in Ω1
1
∀v ∈ H01 (Ω), T− v := .
: −v2 in Ω2

(+) Obviously, (T− )2 = IH 1 (Ω) .

0

(–) But T− 6∈ L(H01 (Ω)), because the matching condition is not enforced.

Second try: let R1 ∈ L(V1 , V2 ) s.t. for all v1 ∈ V1 , MatchingΣ (v1 , R1 v1 ) = 0.

8
< v in Ω1
1
∀v ∈ H01 (Ω), T v := .
: −v2 +2R1 v1 in Ω2

Can one achieve T-coercivity?

Nancy, April 2012 – p.18/24

 Sign-changing coefficients-4
Some elementary computations:

|atr (v, Tv)| = |atr tr tr

1 (v1 , v1 ) − a2 (v2 , v2 ) + 2a2 (v2 , R1 v1 )|

≥ |atr tr tr
1 (v1 , v1 ) − a2 (v2 , v2 )| − 2|a2 (v2 , R1 v1 )|

≥ σ1− kv1 k2V1 − atr tr

2 (v2 , v2 ) − 2|a2 (v2 , R1 v1 )|

Nancy, April 2012 – p.19/24

 Sign-changing coefficients-4
Some elementary computations: let δ > 0, apply Young’s inequality

|atr (v, Tv)| = |atr tr tr

1 (v1 , v1 ) − a2 (v2 , v2 ) + 2a2 (v2 , R1 v1 )|

≥ |atr tr tr
1 (v1 , v1 ) − a2 (v2 , v2 )| − 2|a2 (v2 , R1 v1 )|

≥ σ1− kv1 k2V1 − atr tr

2 (v2 , v2 ) − 2|a2 (v2 , R1 v1 )|

≥ σ1− kv1 k2V1 − atr tr

2 (v2 , v2 ) + δa2 (v2 , v2 ) + δ
−1 tr
a2 (R1 v1 , R1 v1 )
≥ (σ1− − δ −1 σ2+ |||R1 |||2 )kv1 k2V1 − (1 − δ)atr
2 (v2 , v2 ).

Nancy, April 2012 – p.19/24

 Sign-changing coefficients-4
Some elementary computations: let δ > 0, apply Young’s inequality

|atr (v, Tv)| = |atr tr tr

1 (v1 , v1 ) − a2 (v2 , v2 ) + 2a2 (v2 , R1 v1 )|

≥ |atr tr tr
1 (v1 , v1 ) − a2 (v2 , v2 )| − 2|a2 (v2 , R1 v1 )|

≥ σ1− kv1 k2V1 − atr tr

2 (v2 , v2 ) − 2|a2 (v2 , R1 v1 )|

≥ σ1− kv1 k2V1 − atr tr

2 (v2 , v2 ) + δa2 (v2 , v2 ) + δ
−1 tr
a2 (R1 v1 , R1 v1 )
≥ (σ1− − δ −1 σ2+ |||R1 |||2 )kv1 k2V1 − (1 − δ)atr
2 (v2 , v2 ).

Hence, to obtain |atr (v, Tv)| ≥ α kvk2V with α > 0, it is sufficient that

σ1−
> |||R1 |||2 .
σ2+

Nancy, April 2012 – p.19/24

 Sign-changing coefficients-5
Third try: let R2 ∈ L(V2 , V1 ) s.t. for all v2 ∈ V2 , MatchingΣ (R2 v2 , v2 ) = 0.
8
< v −2R v in Ω1
1 2 2
∀v ∈ H01 (Ω), T v := .
: −v2 in Ω2

Nancy, April 2012 – p.20/24

 Sign-changing coefficients-5
Third try: let R2 ∈ L(V2 , V1 ) s.t. for all v2 ∈ V2 , MatchingΣ (R2 v2 , v2 ) = 0.
8
< v −2R v in Ω1
1 2 2
∀v ∈ H01 (Ω), T v := .
: −v2 in Ω2

(+) T ∈ L(H01 (Ω)).

(+) One checks easily that T2 = IH 1 (Ω) !
0

Nancy, April 2012 – p.20/24

 Sign-changing coefficients-5
Third try: let R2 ∈ L(V2 , V1 ) s.t. for all v2 ∈ V2 , MatchingΣ (R2 v2 , v2 ) = 0.
8
< v −2R v in Ω1
1 2 2
∀v ∈ H01 (Ω), T v := .
: −v2 in Ω2

(+) T ∈ L(H01 (Ω)).

(+) One checks easily that T2 = IH 1 (Ω) !
0

To obtain |atr (v, Tv)| ≥ α kvk2V with α > 0, it is sufficient that

σ2−
> |||R2 |||2 .
σ1+

Nancy, April 2012 – p.20/24

 Sign-changing coefficients-5
Third try: let R2 ∈ L(V2 , V1 ) s.t. for all v2 ∈ V2 , MatchingΣ (R2 v2 , v2 ) = 0.
8
< v −2R v in Ω1
1 2 2
∀v ∈ H01 (Ω), T v := .
: −v2 in Ω2

(+) T ∈ L(H01 (Ω)).

(+) One checks easily that T2 = IH 1 (Ω) !
0

To obtain |atr (v, Tv)| ≥ α kvk2V with α > 0, it is sufficient that

σ2−
> |||R2 |||2 .
σ1+

Conclusion: to achieve T-coercivity , one needs

σ1− σ2−
„ «2 „ «2
> inf |||R1 ||| or > inf |||R2 ||| .
σ2+ R1 σ1+ R2

Nancy, April 2012 – p.20/24

 Sign-changing coefficients-6

How to choose the operators R1 or R2 ?

using traces on Σ, liftings, cf. [BonnetBenDhia-Jr-Zwölf’10], [Nicaise-Venel’11] ;
using geometrical transformations, cf. [BonnetBenDhia-Chesnel-Jr’12].

Nancy, April 2012 – p.21/24

 Sign-changing coefficients-6

How to choose the operators R1 or R2 ?

using traces on Σ, liftings, cf. [BonnetBenDhia-Jr-Zwölf’10], [Nicaise-Venel’11] ;
using geometrical transformations, cf. [BonnetBenDhia-Chesnel-Jr’12].
Numerical studies in (Vh )h :
in general,
“ one cannot build discrete operators
” (Th )h s.t.
||(Th −T)(vh )||V
limh→0 supvh ∈Vh \{0} ||vh ||V
= 0;
one can only prove that (Rk,h )h is bounded wrt |||Rk |||, k = 1, 2.

Nancy, April 2012 – p.21/24

 Sign-changing coefficients-6

How to choose the operators R1 or R2 ?

using traces on Σ, liftings, cf. [BonnetBenDhia-Jr-Zwölf’10], [Nicaise-Venel’11] ;
using geometrical transformations, cf. [BonnetBenDhia-Chesnel-Jr’12].
Numerical studies in (Vh )h :
in general,
“ one cannot build discrete operators
” (Th )h s.t.
||(Th −T)(vh )||V
limh→0 supvh ∈Vh \{0} ||vh ||V
= 0;
one can only prove that (Rk,h )h is bounded wrt |||Rk |||, k = 1, 2.
Safety net: choose σ s.t. σ1− /σ2+ or σ2− /σ1+ are sufficiently large to ensure

σ1− 2 σ2−
> |||R1,h ||| or > |||R2,h |||2 , for h small enough.
σ2+ σ1+

Nancy, April 2012 – p.21/24

 Sign-changing coefficients-6

How to choose the operators R1 or R2 ?

using traces on Σ, liftings, cf. [BonnetBenDhia-Jr-Zwölf’10], [Nicaise-Venel’11] ;
using geometrical transformations, cf. [BonnetBenDhia-Chesnel-Jr’12].
Numerical studies in (Vh )h :
in general,
“ one cannot build discrete operators
” (Th )h s.t.
||(Th −T)(vh )||V
limh→0 supvh ∈Vh \{0} ||vh ||V
= 0;
one can only prove that (Rk,h )h is bounded wrt |||Rk |||, k = 1, 2.
Safety net: choose σ s.t. σ1− /σ2+ or σ2− /σ1+ are sufficiently large to ensure

σ1− 2 σ2−
> |||R1,h ||| or > |||R2,h |||2 , for h small enough.
σ2+ σ1+

Under this last assumption, convergence follows.

Nancy, April 2012 – p.21/24

 Sign-changing coefficients-6

How to choose the operators R1 or R2 ?

using traces on Σ, liftings, cf. [BonnetBenDhia-Jr-Zwölf’10], [Nicaise-Venel’11] ;
using geometrical transformations, cf. [BonnetBenDhia-Chesnel-Jr’12].
Numerical studies in (Vh )h :
in general,
“ one cannot build discrete operators
” (Th )h s.t.
||(Th −T)(vh )||V
limh→0 supvh ∈Vh \{0} ||vh ||V
= 0;
one can only prove that (Rk,h )h is bounded wrt |||Rk |||, k = 1, 2.
Safety net: choose σ s.t. σ1− /σ2+ or σ2− /σ1+ are sufficiently large to ensure

σ1− 2 σ2−
> |||R1,h ||| or > |||R2,h |||2 , for h small enough.
σ2+ σ1+

Under this last assumption, convergence follows.

NB. One can also add dissipation, cf. [Chesnel-Jr’1x]:
(+) convergence follows without safety net ;
(–) convergence rate is reduced.

Nancy, April 2012 – p.21/24

 Sign-changing coefficients-6

How to choose the operators R1 or R2 ?

using traces on Σ, liftings, cf. [BonnetBenDhia-Jr-Zwölf’10], [Nicaise-Venel’11] ;
using geometrical transformations, cf. [BonnetBenDhia-Chesnel-Jr’12].
Numerical studies in (Vh )h :
in general,
“ one cannot build discrete operators
” (Th )h s.t.
||(Th −T)(vh )||V
limh→0 supvh ∈Vh \{0} ||vh ||V
= 0;
one can only prove that (Rk,h )h is bounded wrt |||Rk |||, k = 1, 2.
Safety net: choose σ s.t. σ1− /σ2+ or σ2− /σ1+ are sufficiently large to ensure

σ1− 2 σ2−
> |||R1,h ||| or > |||R2,h |||2 , for h small enough.
σ2+ σ1+

Under this last assumption, convergence follows.

Numerical results:
conforming discretization, cf. [Chesnel-Jr’1x].
non-conforming discretization, cf. [Chung-Jr’1x] ;

Nancy, April 2012 – p.21/24

 Sign-changing coefficients-7
In a symmetric domain. Here, Ω =] − 1, 1[×]0, 1[, Ω1 and Ω2 are unit squares.
σk := σ|Ωk , k = 1, 2, are constant numbers, and σ2 /σ1 = −1.001 ; ω = 0.
An exact piecewise smooth solution of the transmission problem is available.
Conforming discretization using P1 Lagrange FE.
We study below the influence of the meshes (errors in L2 -norm ; O(h2 ) is expected).

Nancy, April 2012 – p.22/24

 Sign-changing coefficients-7
In a symmetric domain.
σk := σ|Ωk , k = 1, 2, are constant numbers, and σ2 /σ1 = −1.001 ; ω = 0.
An exact piecewise smooth solution of the transmission problem is available.
Conforming discretization using P1 Lagrange FE.
We study below the influence of the meshes (errors in L2 -norm ; O(h2 ) is expected).
Contrast =−1.001
1
Non symmetric mesh
Locally symmetric mesh
0 Symmetric mesh

−1
Relative errors

−2 a =−1.2433

−3

a =−1.2539
−4

−5

a =−1.9993
−6
0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
log(1/h)

Nancy, April 2012 – p.22/24

 Sign-changing coefficients-8
In a rectangle. Here, Ω =]0, 5[×]0, 2[, Ω2 =]1, 3[×]0, 2[, and Ω1 = Ω \ Ω2 .

metamaterial
Ω1 Ω2 Ω1

(σk )k=1,2 are constant numbers, and σ2 /σ1 = −1/3 ; ω = 1.6 and η = σ −1 .
An exact piecewise smooth solution of the transmission problem is available.
Non-conforming discretization using staggered DG1 FE, cf. [Chung-Engquist’06/’09].
Errors in L2 -norm ; O(h2 ) is expected.

Nancy, April 2012 – p.23/24

 Sign-changing coefficients-8
In a rectangle.

metamaterial
Ω1 Ω2 Ω1

(σk )k=1,2 are constant numbers, and σ2 /σ1 = −1/3 ; ω = 1.6 and η = σ −1 .
An exact piecewise smooth solution of the transmission problem is available.
Non-conforming discretization using staggered DG1 FE, cf. [Chung-Engquist’06/’09].
Errors in L2 -norm ; O(h2 ) is expected. Numerically, the order is 1.9999.

Nancy, April 2012 – p.23/24

 Sign-changing coefficients-8
In a rectangle.

metamaterial
Ω1 Ω2 Ω1

(σk )k=1,2 are constant numbers, and σ2 /σ1 = −1/3 ; ω = 1.6 and η = σ −1 .
An exact piecewise smooth solution of the transmission problem is available.
Non-conforming discretization using staggered DG1 FE, cf. [Chung-Engquist’06/’09].
Errors in L2 -norm ; O(h2 ) is expected. Numerically, the order is 1.9999.
Exact solution Numerical solution Numerical and exact solutions at y=0.98
0.05
0 0

0
−0.05 −0.05
−0.05
−0.1 −0.1
−0.1

−0.15 −0.15
−0.15

−0.2 −0.2
−0.2

−0.25 −0.25 −0.25

Numerical
−0.3 −0.3 −0.3 Exact

−0.35
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Nancy, April 2012 – p.23/24

 Conclusion/Perspectives
T-coercivity is versatile!
BEM for the classical Maxwell problem (cf. [Buffa-Costabel-Schwab’02]) ;
FEM for the classical scalar or Maxwell problems (cf. [Jr’12]) ;
Vol. Int. Eq. Methods for scattering from gratings (cf. [Lechleiter-Nguyen’1x]) ;
study of Interior Transmission Eigenvalue Problems:
scalar case (cf. [BonnetBenDhia-Chesnel-Haddar’11]) ;
Maxwell problem (cf. [Chesnel’1x]) ;
etc.

Nancy, April 2012 – p.24/24

 Conclusion/Perspectives
T-coercivity is versatile!

Scalar problems with sign-shifting coefficients:

introduction of T-coercivity during WAVES’07 (cf. [BonnetBenDhia-Jr-Zwölf’10]) ;
numerical analysis when T-coercivity applies (cf. [BonnetBenDhia-Jr-Zwölf’10],
[Nicaise-Venel’11], [Chesnel-Jr’1x], DG-approach [Chung-Jr’1x], etc.) ;
theoretical study of well-posedness (cf. [BonnetBenDhia-Chesnel-Jr’12]) ;
theoretical study of the critical cases (with [BonnetBenDhia-Chesnel-Claeys’1x]) ;
discretization and numerical analysis of the critical cases.

Nancy, April 2012 – p.24/24

 Conclusion/Perspectives
T-coercivity is versatile!

Scalar problems with sign-shifting coefficients:

introduction of T-coercivity during WAVES’07 (cf. [BonnetBenDhia-Jr-Zwölf’10]) ;
numerical analysis when T-coercivity applies (cf. [BonnetBenDhia-Jr-Zwölf’10],
[Nicaise-Venel’11], [Chesnel-Jr’1x], DG-approach [Chung-Jr’1x], etc.) ;
theoretical study of well-posedness (cf. [BonnetBenDhia-Chesnel-Jr’12]) ;
theoretical study of the critical cases (with [BonnetBenDhia-Chesnel-Claeys’1x]) ;
discretization and numerical analysis of the critical cases.
Maxwell problem(s) with sign-shifting coefficients:
T-coercivity + side results during NELIA’11 (cf. [BonnetBenDhia-Chesnel-Jr’??]) ;
numerical analysis when T-coercivity applies.

Nancy, April 2012 – p.24/24

 Conclusion/Perspectives
T-coercivity is versatile!

Scalar problems with sign-shifting coefficients:

introduction of T-coercivity during WAVES’07 (cf. [BonnetBenDhia-Jr-Zwölf’10]) ;
numerical analysis when T-coercivity applies (cf. [BonnetBenDhia-Jr-Zwölf’10],
[Nicaise-Venel’11], [Chesnel-Jr’1x], DG-approach [Chung-Jr’1x], etc.) ;
theoretical study of well-posedness (cf. [BonnetBenDhia-Chesnel-Jr’12]) ;
theoretical study of the critical cases (with [BonnetBenDhia-Chesnel-Claeys’1x]) ;
discretization and numerical analysis of the critical cases.
Maxwell problem(s) with sign-shifting coefficients:
T-coercivity + side results during NELIA’11 (cf. [BonnetBenDhia-Chesnel-Jr’??]) ;
numerical analysis when T-coercivity applies.
In the critical cases: are models derived from physics still relevant?
re-visit models (homogenization, multi-scale numerics, etc.).
(A.N.R. METAMATH Project ; coordinator S. Fliss (POEMS)).

Nancy, April 2012 – p.24/24