• Aucun résultat trouvé

Asymptotically constant-free, p-robust and guaranteed a posteriori error estimates for the Helmholtz equation

N/A
N/A
Protected

Academic year: 2021

Partager "Asymptotically constant-free, p-robust and guaranteed a posteriori error estimates for the Helmholtz equation"

Copied!
81
0
0

Texte intégral

(1)

HAL Id: hal-02321140

https://hal.inria.fr/hal-02321140

Submitted on 21 Oct 2019

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of

sci-entific research documents, whether they are

pub-lished or not. The documents may come from

teaching and research institutions in France or

abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est

destinée au dépôt et à la diffusion de documents

scientifiques de niveau recherche, publiés ou non,

émanant des établissements d’enseignement et de

recherche français ou étrangers, des laboratoires

publics ou privés.

Asymptotically constant-free, p-robust and guaranteed a

posteriori error estimates for the Helmholtz equation

Théophile Chaumont-Frelet, Alexandre Ern, Martin Vohralík

To cite this version:

Théophile Chaumont-Frelet, Alexandre Ern, Martin Vohralík. Asymptotically constant-free, p-robust

and guaranteed a posteriori error estimates for the Helmholtz equation. EnuMath 2019 - European

Numerical Mathematics and Advanced Applications Conference, Sep 2019, Egmond aan Zee,

Nether-lands. �hal-02321140�

(2)

Asymptotically constant-free, p-robust and guaranteed

a posteriori error estimates for the Helmholtz equation

T. Chaumont-Frelet (Inria project-team Nachos, CNRS, UCA LJAD) A. Ern (CERMICS, Inria project-team Serena)

M. Vohral´ık (Inria project-team Serena, CERMICS)

Numerical methods for wave propagation with applications in electromagnetics and geophysics

(3)

Asymptotically constant-free, p-robust and guaranteed

a posteriori error estimates for the Helmholtz equation

T. Chaumont-Frelet (Inria project-team Nachos, CNRS, UCA LJAD) A. Ern (CERMICS, Inria project-team Serena)

M. Vohral´ık (Inria project-team Serena, CERMICS)

Numerical methods for wave propagation with applications in electromagnetics and geophysics

(4)

Asymptotically constant-free

,

p-robust

and

guaranteed

a posteriori error estimates for the Helmholtz equation

T. Chaumont-Frelet (Inria project-team Nachos, CNRS, UCA LJAD) A. Ern (CERMICS, Inria project-team Serena)

M. Vohral´ık (Inria project-team Serena, CERMICS)

Numerical methods for wave propagation with applications in electromagnetics and geophysics

(5)

Model problem: the Helmholtz equation

Given f , we seek u such that    −k2 u − ∆u = f in Ω, u = 0 on ΓD, ∇u · n − iku = 0 on ΓA, where: Ω ⊂ Rd is a polytopal domain, ∂Ω = ΓD∪ ΓA, k ≥ 0 is the wavenumber. ΓA D ΓD Ω

(6)

Model problem: variational formulation

After integration by parts, we obtain a weak formulation: Find u ∈ HΓ1D(Ω) such that

b(u, v ) = (f , v ) ∀v ∈ HΓ1D(Ω),

where

(7)

Model problem: FEM discretization

The discrete problem consists in finding uh∈ Vhsuch that

b(uh, vh) = (f , vh) ∀vh∈ Vh,

where Vhis the usual Lagrange FE discretization subspace of degree p ≥ 1

Vh:=

n

vh∈ HΓ1D(Ω) | vh|K ∈ Pp(K ) ∀K ∈ Th

o ,

and where This a “shape-regular” mesh of Ω.

(8)

Main motivation: a posteriori error estimation

We would like to estimate the discretization error

eh:= u − uh,

in some suitable norm k · k?,Ω.

We associate with each K ∈ Th a computable quantity

ηK:= ηK(uh, f )

that we call a local “estimator”. We gather the contributions

η :=   X K ∈Th η2K   1/2 ,

(9)

Main motivation: a posteriori error estimation

A “good” a posteriori estimator is reliable

kehk?,Ω≤Cη,

and locally efficient

ηK≤Ckehk?,ωK,

where:

C is a “nice” constant,

C only depends on the “shape-regularity” and p, ωK is the patch of elements sharing a vertex with K .

(10)

Some vocabulary

An error estimate is guaranteed, if

kehk?,Ω≤Cη,

with a fully computable constantC.

An error estimate is constant-free ifC= 1, i.e. kehk?,Ω≤ η.

Finally, we say that an error estimate is p-robust, if

ηK≤Ckehk?,ωK

with a constantC that is independent of p.

The goal of this talk is to (try to) achieve these properties using an “equilibrated” estimator.

(11)

Outline

1 The low wavenumber regime: the Poisson equation

2 The Helmholtz equation in the high-wavenumber regime

3 Numerical examples

(12)

Outline

1 The low wavenumber regime: the Poisson equation

Error upper bounds via equilibrated fluxes Practical construction of the equilibrated flux Reliability and p-robustness

Summary

2 The Helmholtz equation in the high-wavenumber regime

3 Numerical examples

(13)

Outline

1 The low wavenumber regime: the Poisson equation

Error upper bounds via equilibrated fluxes

Practical construction of the equilibrated flux Reliability and p-robustness

Summary

2 The Helmholtz equation in the high-wavenumber regime

3 Numerical examples

(14)

Low-wavenumber regime: the Poisson equation

Let us first set k = 0. We obtain a “Poisson” problem:

   −∆u = f in Ω, u = 0 on ΓD, ∇u · n = 0 on ΓA.

Equivalently, we seek u ∈ HΓ1D(Ω) such that

b(u, v ) = (f , v ) ∀v ∈ HΓ1D(Ω),

where now

b(u, v ) = (∇u, ∇v ).

The natural norm for the analysis is

(15)

Equilibrated fluxes

We call “equilibrated flux” an element σ ∈ H(div, Ω) such that

 ∇ · σ = f in Ω, σ · n = 0 on ΓA. Recalling that  ∇ · (−∇u) = f in Ω, (−∇u) · n = 0 on ΓA,

this definition mimics the properties of −∇u.

We have

|eh|1,Ω= k(−∇u) + ∇uhk0,Ω≤ kσ + ∇uhk0,Ω

for any equilibrated flux σ ∈ H(div, Ω).

(16)

Equilibrated fluxes

We call “equilibrated flux” an element σ ∈ H(div, Ω) such that

 ∇ · σ = f in Ω, σ · n = 0 on ΓA. Recalling that  ∇ · (−∇u) = f in Ω, (−∇u) · n = 0 on ΓA,

this definition mimics the properties of −∇u.

We have

|eh|1,Ω= k(−∇u) + ∇uhk0,Ω≤ kσ + ∇uhk0,Ω

(17)

Equilibrated fluxes

We call “equilibrated flux” an element σ ∈ H(div, Ω) such that

 ∇ · σ = f in Ω, σ · n = 0 on ΓA. Recalling that  ∇ · (−∇u) = f in Ω, (−∇u) · n = 0 on ΓA,

this definition mimics the properties of −∇u.

We have

|eh|1,Ω= k(−∇u) + ∇uhk0,Ω≤ kσ + ∇uhk0,Ω

for any equilibrated flux σ ∈ H(div, Ω).

(18)

The Prager-Synge theorem

This is known as the Prager-Synge theorem. How do we establish this result?

We have the following expression for the error:

|eh| 2

1,Ω= b(eh, eh)

= b(u, eh) − b(uh, eh)

= (f , eh) − (∇uh, ∇eh)

We can rewrite the first term

(f , eh) = (∇ · σ, eh) = −(σ, ∇eh).

This leads to

|eh| 2

1,Ω= −(∇uh+ σ, ∇eh) ≤ kσ + ∇uhk0,Ω|eh|1,Ω,

(19)

The Prager-Synge theorem

This is known as the Prager-Synge theorem. How do we establish this result?

We have the following expression for the error:

|eh| 2

1,Ω= b(eh, eh)

= b(u, eh) − b(uh, eh)

= (f , eh) − (∇uh, ∇eh)

We can rewrite the first term

(f , eh) = (∇ · σ, eh) = −(σ, ∇eh).

This leads to

|eh| 2

1,Ω= −(∇uh+ σ, ∇eh) ≤ kσ + ∇uhk0,Ω|eh|1,Ω,

and the result follows.

(20)

The Prager-Synge theorem

This is known as the Prager-Synge theorem. How do we establish this result?

We have the following expression for the error:

|eh| 2

1,Ω= b(eh, eh)

= b(u, eh) − b(uh, eh)

= (f , eh) − (∇uh, ∇eh)

We can rewrite the first term

(f , eh) = (∇ · σ, eh) = −(σ, ∇eh).

This leads to

|eh| 2

1,Ω= −(∇uh+ σ, ∇eh) ≤ kσ + ∇uhk0,Ω|eh|1,Ω,

(21)

The Prager-Synge theorem

This is known as the Prager-Synge theorem. How do we establish this result?

We have the following expression for the error:

|eh| 2

1,Ω= b(eh, eh)

= b(u, eh) − b(uh, eh)

= (f , eh) − (∇uh, ∇eh)

We can rewrite the first term

(f , eh) = (∇ · σ, eh) = −(σ, ∇eh).

This leads to

|eh|21,Ω= −(∇uh+ σ, ∇eh) ≤ kσ + ∇uhk0,Ω|eh|1,Ω,

and the result follows.

(22)

Constant-free error estimates

Given an “equilibrated flux” σ ∈ H(div, Ω), we set

ηK:= kσ + ∇uhk0,K ∀K ∈ Th and η2:= X K ∈Th η2K= kσ + ∇uhk20,Ω. Then, we have |eh|1,Ω≤ η = kσ + ∇uhk0,Ω.

(23)

Outline

1 The low wavenumber regime: the Poisson equation

Error upper bounds via equilibrated fluxes

Practical construction of the equilibrated flux

Reliability and p-robustness Summary

2 The Helmholtz equation in the high-wavenumber regime

3 Numerical examples

(24)

Practical construction of the equilibrated flux

For the sake of simplicity, we assume that f |K∈ Pp−1(K ) for all K ∈ Th.

All the presented result naturally extends by adding “oscillation terms”

oscK(f ) :=

hK

π kf − π

p−1 K f k0,K.

Let Wh⊂ H(div, Ω) denote the Raviart-Thomas space of degree p.

There exists σh∈ Whsuch that

∇ · σh= f ,

(25)

“Ideal” discrete flux

We recall that we have the upper bound

|eh|1,Ω≤ η := kσh+ ∇uhk0,Ω.

Thus the “ideal” discrete flux would be

σh:= argmin τh∈Wh

∇·τh=f in ωa

τh·n=0 on ∂ωa\ΓD

kτh+ ∇uhk0,Ω.

It is computable, through resolution of a discrete global saddle point problem.

Fortunaly, it is possible to localize the computations, vertex by vertex.

(26)

Construction through “patch” problems (hidden magic)

For each vertex a ∈ Vh, consider the following vertex-patch problem

σah:= argmin τah∈Wh(ωa) ∇·τa h=fain ωa τa h·n=0 on ∂ωa\ΓD kτah+ ψa∇uhk0,Ω,

where ωa is the “vertex patch” around a

fa= ψaf − ∇ψa· ∇uh,

and ψa is the “hat function” associated with a.

We obtain an equilibrated flux by setting

σh:=

X

a∈Vh

σah.

(27)

Outline

1 The low wavenumber regime: the Poisson equation

Error upper bounds via equilibrated fluxes Practical construction of the equilibrated flux

Reliability and p-robustness

Summary

2 The Helmholtz equation in the high-wavenumber regime

3 Numerical examples

(28)

Reliability and p-robustness (hidden magic)

Our particular construction of σhleads to p robust estmiates: we have

ηK:= kσh+ ∇uhk0,K ≤C|eh|1,ωK

whereC is independent of p.

D. Braess, V. Pillwein, J. Sch¨oberl, CMAME, 2009:

Equilibrated residual error estimates are p-robust. A. Ern and M. Vohral´ık, Math. Comp., 2019:

(29)

Outline

1 The low wavenumber regime: the Poisson equation

Error upper bounds via equilibrated fluxes Practical construction of the equilibrated flux Reliability and p-robustness

Summary

2 The Helmholtz equation in the high-wavenumber regime

3 Numerical examples

(30)

Summary

We construct an equilibrated flux σh based on local finite element problems.

This flux provides a constant-free error estimate

|eh|1,Ω≤ kσh+ ∇uhk0,Ω=: η.

This estimator is locally efficient and p-robust since

ηK:= kσh+ ∇uhk0,K ≤C|eh|1,ωK,

(31)

Outline

1 The low wavenumber regime: the Poisson equation

2 The Helmholtz equation in the high-wavenumber regime

How do we deal with the absence of coercivity? Error upper bound via equilibrated fluxes

Practical construction, reliability and p-robustness Summary

3 Numerical examples

(32)

Outline

1 The low wavenumber regime: the Poisson equation

2 The Helmholtz equation in the high-wavenumber regime

How do we deal with the absence of coercivity?

Error upper bound via equilibrated fluxes

Practical construction, reliability and p-robustness Summary

3 Numerical examples

(33)

The Helmholtz problem

Let us now assume that k is large. We recall that

b(u, v ) := −k2(u, v ) − ik(u, v )ΓA+ (∇u, ∇v ) ∀u, v ∈ H

1 ΓD(Ω).

We will consider the “natural” norm

kuk21,k,Ω:= k 2 kuk20,Ω+ kkuk 2 0,ΓA+ k∇uk 2 0,Ω. In particular, we have |b(u, v )| ≤ kuk1,k,Ωkv k1,k,Ω,

which is interesting for goal-oriented error estimation and adaptivity.

(34)

ardling-like inequality

The sesquilinear form is not coercive

Re b(φ, φ) = |φ|21,Ω− k 2kφk2

0,Ω, φ ∈ H 1 ΓD(Ω).

Instead, it satisfies a G˚arding-like inequality:

Re b(φ, φ) = kφk21,k,Ω−  2k2kφk20,Ω+ kkφk 2 0,ΓA  , φ ∈ HΓ1D(Ω).

(35)

Recovering coercivity

For φ ∈ HΓ1D(Ω), there exists a uniqueS

? φ∈ H 1 ΓD(Ω) such that b(w ,Sφ?) = 2k 2 (w , φ) + k(w , φ)ΓA. Importantly, we have b(φ,Sφ?) = 2k 2 kφk20,Ω+ kkφk 2 0,ΓA, for all φ ∈ HΓ1D(Ω). Recalling that Re b(φ, φ) = kφk21,k,Ω−  2k2kφk20,Ω+ kkφk 2 0,ΓA  , we have Re b(φ, φ +Sφ?) = kφk 2 1,k,Ω for all φ ∈ H1 ΓD(Ω).

(36)

Recovering coercivity

For φ ∈ HΓ1D(Ω), there exists a uniqueS

? φ ∈ H 1 ΓD(Ω) such that b(w ,Sφ?) = 2k 2 (w , φ) + k(w , φ)ΓA. Importantly, we have b(φ,Sφ?) = 2k 2 kφk20,Ω+ kkφk 2 0,ΓA, for all φ ∈ HΓ1D(Ω). Recalling that Re b(φ, φ) = kφk21,k,Ω−  2k2kφk20,Ω+ kkφk 2 0,ΓA  , we have Re b(φ, φ +Sφ?) = kφk 2 1,k,Ω for all φ ∈ H1 ΓD(Ω).

(37)

Recovering coercivity

For φ ∈ HΓ1D(Ω), there exists a uniqueS

? φ ∈ H 1 ΓD(Ω) such that b(w ,Sφ?) = 2k 2 (w , φ) + k(w , φ)ΓA. Importantly, we have b(φ,Sφ?) = 2k 2 kφk20,Ω+ kkφk 2 0,ΓA, for all φ ∈ HΓ1D(Ω). Recalling that Re b(φ, φ) = kφk21,k,Ω−  2k2kφk20,Ω+ kkφk 2 0,ΓA  , we have Re b(φ, φ +Sφ?) = kφk 2 1,k,Ω for all φ ∈ H1 ΓD(Ω).

(38)

Recovering coercivity

For φ ∈ HΓ1D(Ω), there exists a uniqueS

? φ ∈ H 1 ΓD(Ω) such that b(w ,Sφ?) = 2k 2 (w , φ) + k(w , φ)ΓA. Importantly, we have b(φ,Sφ?) = 2k 2 kφk20,Ω+ kkφk 2 0,ΓA, for all φ ∈ HΓ1D(Ω). Recalling that Re b(φ, φ) = kφk21,k,Ω−  2k2kφk20,Ω+ kkφk 2 0,ΓA  , we have Re b(φ, φ +Sφ?) = kφk 2 1,k,Ω for all φ ∈ H1 ΓD(Ω).

(39)

Recovering coercivity

For φ ∈ HΓ1D(Ω), there exists a uniqueS

? φ ∈ H 1 ΓD(Ω) such that b(w ,Sφ?) = 2k 2 (w , φ) + k(w , φ)ΓA. Importantly, we have b(φ,Sφ?) = 2k 2 kφk20,Ω+ kkφk 2 0,ΓA, for all φ ∈ HΓ1D(Ω). Recalling that Re b(φ, φ) = kφk21,k,Ω−  2k2kφk20,Ω+ kkφk 2 0,ΓA  , we have Re b(φ, φ +Sφ?) = kφk 2 1,k,Ω for all φ ∈ H1 ΓD(Ω).

(40)

Approximation factor

We then introduce γba:= sup φ∈H1 ΓD(Ω)\{0} min vh∈Vh kS? φ− vhk1,k,Ω kφk1,k,Ω .

It is the best constant such that

min

vh∈VhkS ?

φ− vhk1,k,Ω≤γbakφk1,k,Ω ∀φ ∈ HΓ1D(Ω).

(41)

Approximation factor

In the strong sense,Sφ?is defined by

   −k2S? φ− ∆Sφ? = 2k2φ in Ω, S? φ = 0 on ΓD, ∇S? φ· n + ikSφ? = kφ on ΓA. In particularSφ?∈ H 1+ε

(Ω) for some ε > 0, so that

γba≤C(Ω, ΓD, k)

 h p

→ 0. For non-trapping domains, we can show that

γba≤C(Ω, ΓD)  kh p + k  kh p p .

J.M. Melenk and S. Sauter, SIAM J. Numer. Anal., 2011:

Wavenumber explicit convergence analysis for Galerkin discretizations of the Helmholtz equation. T. Chaumont-Frelet and S. Nicaise, IMA J. Numer. Anal., 2019:

Wavenumber explicit convergence analysis for finite element discretizations of general wave propagation problems.

(42)

Outline

1 The low wavenumber regime: the Poisson equation

2 The Helmholtz equation in the high-wavenumber regime

How do we deal with the absence of coercivity?

Error upper bound via equilibrated fluxes

Practical construction, reliability and p-robustness Summary

3 Numerical examples

(43)

Equilibrated flux

We observe that −∇u ∈ H(div, Ω) and



∇ · (−∇u) = f + k2u in Ω, (−∇u) · n = −iku on ΓA.

This motivation the following definition:

we say that σ ∈ H(div, Ω) is an equilibrated flux if



∇ · σ = f + k2u

h in Ω,

σ · n = −ikuh on ΓA.

Can we relate kehk1,k,Ωand kσ + ∇uhk0,Ω?

(44)

A type of Prager-Synge theorem

As a comparison, the starting point was

|eh|21,Ω= b(eh, eh)

in the coercive case. Here, the situation is more complex.

Using the definition ofS?

eh and Galerkin orthogonality, we have

kehk21,k,Ω= b(eh, eh+Se?h) = b(eh, eh+S ? eh− vh) = b(eh, ψ) where ψ := eh+Se?h− vh vh:= argmin vh∈Vh kS? eh− vhk1,k,Ω.

We can show that

kψk1,k,Ω≤ kehk1,k,Ω+ min vh∈Vh

kS?

eh− vhk1,k,Ω

(45)

A type of Prager-Synge theorem

As a comparison, the starting point was

|eh|21,Ω= b(eh, eh)

in the coercive case. Here, the situation is more complex. Using the definition ofS?

eh and Galerkin orthogonality, we have

kehk21,k,Ω= b(eh, eh+Se?h) = b(eh, eh+S ? eh− vh) = b(eh, ψ) where ψ := eh+Se?h− vh vh:= argmin vh∈Vh kS? eh− vhk1,k,Ω.

We can show that

kψk1,k,Ω≤ kehk1,k,Ω+ min vh∈Vh

kS?

eh− vhk1,k,Ω

≤ (1 +γba)kehk1,k,Ω.

(46)

A type of Prager-Synge theorem

As a comparison, the starting point was

|eh|21,Ω= b(eh, eh)

in the coercive case. Here, the situation is more complex. Using the definition ofS?

eh and Galerkin orthogonality, we have

kehk21,k,Ω= b(eh, eh+Se?h) = b(eh, eh+S ? eh− vh) = b(eh, ψ) where ψ := eh+Se?h− vh vh:= argmin vh∈Vh kS? eh− vhk1,k,Ω.

We can show that

kψk1,k,Ω≤ kehk1,k,Ω+ min vh∈Vh

kS?

eh− vhk1,k,Ω

(47)

A type of Prager-Synge theorem

We have

kehk21,k,Ω= b(eh, ψ) = b(u, ψ) − b(eh, ψ)

= (f , ψ) + k2(uh, ψ) + ik(uh, ψ)ΓA− (∇uh, ∇ψ)

= (f + k2uh, ψ) + ik(uh, ψ)ΓA− (∇uh, ∇ψ).

For any equilibrated flux σ ∈ H(div, Ω), we have

(f + k2uh, ψ) + ik(uh, ψ)ΓA = (∇ · σ, ψ) − (σ · n, ψ)∂Ω= −(σ, ∇ψ).

It follows that

kehk 2

1,k,Ω= −(σ + ∇uh, ∇ψ) ≤ kσ + ∇uhk0,Ω|ψ|1,Ω,

and we can conclude since

|ψ|1,Ω≤ kψk1,k,Ω≤ (1 +γba)kehk1,k,Ω.

(48)

A type of Prager-Synge theorem

We have

kehk21,k,Ω= b(eh, ψ) = b(u, ψ) − b(eh, ψ)

= (f , ψ) + k2(uh, ψ) + ik(uh, ψ)ΓA− (∇uh, ∇ψ)

= (f + k2uh, ψ) + ik(uh, ψ)ΓA− (∇uh, ∇ψ).

For any equilibrated flux σ ∈ H(div, Ω), we have

(f + k2uh, ψ) + ik(uh, ψ)ΓA = (∇ · σ, ψ) − (σ · n, ψ)∂Ω= −(σ, ∇ψ).

It follows that

kehk 2

1,k,Ω= −(σ + ∇uh, ∇ψ) ≤ kσ + ∇uhk0,Ω|ψ|1,Ω,

and we can conclude since

(49)

A type of Prager-Synge theorem

We have

kehk21,k,Ω= b(eh, ψ) = b(u, ψ) − b(eh, ψ)

= (f , ψ) + k2(uh, ψ) + ik(uh, ψ)ΓA− (∇uh, ∇ψ)

= (f + k2uh, ψ) + ik(uh, ψ)ΓA− (∇uh, ∇ψ).

For any equilibrated flux σ ∈ H(div, Ω), we have

(f + k2uh, ψ) + ik(uh, ψ)ΓA = (∇ · σ, ψ) − (σ · n, ψ)∂Ω= −(σ, ∇ψ).

It follows that

kehk 2

1,k,Ω= −(σ + ∇uh, ∇ψ) ≤ kσ + ∇uhk0,Ω|ψ|1,Ω,

and we can conclude since

|ψ|1,Ω≤ kψk1,k,Ω≤ (1 +γba)kehk1,k,Ω.

(50)

A type of Prager-Synge theorem

We have

kehk21,k,Ω= b(eh, ψ) = b(u, ψ) − b(eh, ψ)

= (f , ψ) + k2(uh, ψ) + ik(uh, ψ)ΓA− (∇uh, ∇ψ)

= (f + k2uh, ψ) + ik(uh, ψ)ΓA− (∇uh, ∇ψ).

For any equilibrated flux σ ∈ H(div, Ω), we have

(f + k2uh, ψ) + ik(uh, ψ)ΓA = (∇ · σ, ψ) − (σ · n, ψ)∂Ω= −(σ, ∇ψ).

It follows that

kehk21,k,Ω= −(σ + ∇uh, ∇ψ) ≤ kσ + ∇uhk0,Ω|ψ|1,Ω,

and we can conclude since

(51)

A type of Prager-Synge theorem

We have

kehk21,k,Ω= b(eh, ψ) = b(u, ψ) − b(eh, ψ)

= (f , ψ) + k2(uh, ψ) + ik(uh, ψ)ΓA− (∇uh, ∇ψ)

= (f + k2uh, ψ) + ik(uh, ψ)ΓA− (∇uh, ∇ψ).

For any equilibrated flux σ ∈ H(div, Ω), we have

(f + k2uh, ψ) + ik(uh, ψ)ΓA = (∇ · σ, ψ) − (σ · n, ψ)∂Ω= −(σ, ∇ψ).

It follows that

kehk21,k,Ω= −(σ + ∇uh, ∇ψ) ≤ kσ + ∇uhk0,Ω|ψ|1,Ω,

and we can conclude since

|ψ|1,Ω≤ kψk1,k,Ω≤ (1 +γba)kehk1,k,Ω.

(52)

Error estimate using equilibrated fluxes

We thus obtain that

kehk1,k,Ω≤ (1 +γba)kσ + ∇uhk0,Ω

for any equilibrated flux σ.

This estimate is “asymptotically constant-free”, asγba→ 0 as (h/p) → 0.

Guaranteed estimates can be obtained if computabable upper-bounds

γba≤γeba

are provided, which is possible for non-trapping domains.

T. Chaumont-Frelet, A. Ern, M. Vohral´ık, submittted, 2019:

(53)

Outline

1 The low wavenumber regime: the Poisson equation

2 The Helmholtz equation in the high-wavenumber regime

How do we deal with the absence of coercivity? Error upper bound via equilibrated fluxes

Practical construction, reliability and p-robustness

Summary

3 Numerical examples

(54)

Practical construction of the equilibrated

We follow the techniques used for the Poisson problem.

For a ∈ Vh, we introduce vertex-patch minimization problems

σah:= argmin τah∈Wh(ωa) ∇·τah=dain ωa τah·n=baon ∂ω a\ΓD kτah− ∇uhk0,ωa. where da:= ψa(f + k2uh) − ∇ψa· ∇uh ba:= ikψauh1ΓA.

We obtain the global equilibrated flux by summation:

σh:=

X

a∈Vh

(55)

Reliability and p-robustness

Thanks to our particular construction, we additionally have

ηK:= kσh+ ∇uhk0,K ≤C 1 +  kh p 1/2 +kh p ! kehk1,k,ωK,

where the constantC only depends on the shape-regularity of the mesh.

The estimator is reliable and p-robust as long as kh

p ≤C which means that Ndofs/λ≥C.

(56)

Outline

1 The low wavenumber regime: the Poisson equation

2 The Helmholtz equation in the high-wavenumber regime

How do we deal with the absence of coercivity? Error upper bound via equilibrated fluxes

Practical construction, reliability and p-robustness

Summary

3 Numerical examples

(57)

Summary

We can construct an equilibrated flux σhusing local finite element problems.

The construction is very similar to the coercive case (different rhs).

(58)

Summary

The error estimates

kehk1,k,Ω≤ (1 +γba)η

is asymptotically constant-free sinceγba→ 0 as h/p → 0.

For non-trapping domains,γba 1 means

Ndofs/λC(Ω, ΓD)(1 + k1/p),

i.e. many dofs per wavelength or large p.

Guaranteed upper bounds can be obtain if computable upper bounds

γba≤γeba,

(59)

Summary

The p-robust local lower-bound

ηK≤C 1 +  kh p 1/2 +kh p ! kehk1,k,ωK holds.

Thus, the estimator is efficient and p-robust when we have sufficiently many dofs per wavelength:

kh

p ≤C or Ndofs/λ≥C.

(60)

Outline

1 The low wavenumber regime: the Poisson equation

2 The Helmholtz equation in the high-wavenumber regime

3 Numerical examples

A validation experiment A more realistic example

(61)

Outline

1 The low wavenumber regime: the Poisson equation

2 The Helmholtz equation in the high-wavenumber regime

3 Numerical examples

A validation experiment

A more realistic example

(62)

Propagation of a plane wave

We consider the propagation of a plane wave in Ω = (−1, 1)2

   −k2 u − ∆u = 0 in Ω, u = 0 on ΓD, ∇u·n − iku = g on ΓA, where g := ∇ξθ· n − ikξθ ξθ:= eikd ·x

with d := (cos θ, sin θ) and θ = π/12. The solution is u = ξθ.

(63)

Plane wave experiment p = 1 and k = π

10−3 10−2 10−1 100 10−1 100 101 102 103 Efem Eest e Eest h (k = π) Relative erro r (%) Efem:= kehk1,k,Ω Eest:= η e Eest:= (1 +eγba)η

(64)

Plane wave experiment p = 1 and k = 4π

10−3 10−2 10−1 100 10−1 100 101 102 103 Efem Eest e Eest h (k = 4π) Relative erro r (%) Efem:= kehk1,k,Ω Eest:= η e Eest:= (1 +γeba)η

(65)

Plane wave experiment p = 1 and k = 10π

10−3 10−2 10−1 100 101 102 103 Efem Eest e Eest h (k = 10π) Relative erro r (%) Efem:= kehk1,k,Ω Eest:= η e Eest:= (1 +γeba)η

(66)

Plane wave experiment p = 1 and k = 20π

10−3 10−2 10−1 100 101 102 103 Efem Eest e Eest h (k = 20π) Relative erro r (%) Efem:= kehk1,k,Ω Eest:= η e Eest:= (1 +γeba)η

(67)

Plane wave experiment p = 4 and k = 10π

10−2.5 10−2 10−1.5 10−5 10−3 10−1 101 Efem Eest e Eest h (k = 10π) Relative erro r (%) Efem:= kehk1,k,Ω Eest:= η e Eest:= (1 +γeba)η

(68)

Plane wave experiment p = 4 and k = 20π

10−2.5 10−2 10−1.5 10−5 10−3 10−1 101 Efem Eest e Eest h (k = 20π) Relative erro r (%) Efem:= kehk1,k,Ω Eest:= η e Eest:= (1 +γeba)η

(69)

Plane wave experiment p = 4 and k = 40π

10−2.5 10−2 10−1.5 10−2 100 102 Efem Eest e Eest h (k = 40π) Relative erro r (%) Efem:= kehk1,k,Ω Eest:= η e Eest:= (1 +γeba)η

(70)

Plane wave experiment p = 4 and k = 60π

10−2.5 10−2 10−1.5 10−2 100 102 Efem Eest e Eest h (k = 60π) Relative erro r (%) Efem:= kehk1,k,Ω Eest:= η e Eest:= (1 +γeba)η

(71)

Outline

1 The low wavenumber regime: the Poisson equation

2 The Helmholtz equation in the high-wavenumber regime

3 Numerical examples

A validation experiment

A more realistic example

(72)

Scattering by an non-trapping obstacle

We now consider a scattering problem    −k2u − ∆u = 0 in Ω, u = 0 on ΓD, ∇u·n − iku = g on ΓA,

where again g = ∇ξθ· n − ikξθ.

ΓA

D ΓD

We fix the wavenumber k = 10π and employ P3elements.

(73)

Solution of the scattering problem

-2.15 0.0 2.15

Real (left) and imaginary (right) parts of the solution

(74)

Estimated error in mesh #1

0.00 0.23 0.46

(75)

Estimated error in mesh #2

0.00 0.04 0.08

Estimator ηK (left) and elementwise error kehk1,k,K (right)

(76)

Estimated error in mesh #3

0.00 0.01 0.02

(77)

Behavior of the estimator through the adaptive procedure

5 10 10−2 101 104 Efem Eest e Eest Iterations Relative erro r (%) Efem:= kehk1,k,Ω Eest:= η e Eest:= (1 +eγba)η

Behaviors of the estimated and analytical errors in the adaptive procedure

(78)

Outline

1 The low wavenumber regime: the Poisson equation

2 The Helmholtz equation in the high-wavenumber regime

3 Numerical examples

(79)

Upper bounds suffer from the lack of coercivity...

The upper bound

kehk1,k,Ω≤ (1 +γba)η

is “asymptotically constant-free” sinceγba→ 0 as h/p → 0.

We “qualitatively” know the behaviour ofγbafor non-trapping domains

γba≤C(Ω, ΓD)  kh p + k  kh p p .

It is possible to obtain (coarse) guaranteed upper bounds, by deriving computable estimatesγba≤γeba.

(80)

But the lower bounds are okay!

We obtain lower bounds of the form

ηK≤C 1 +  kh p 1/2 +kh p ! kehk1,k,Ω,

where the constantC only depends on the shape-regularity parameter of Th.

(81)

Some references on a posteriori estimation for the Helmholtz equation

Early work in 1D, with first-order FE:

I. Babuˇska, F. Ihlenburg, T. Strouboulis and S.K. Gangaraj, Int. J. Numer. Meth. Engrg., 1997: A posteriori error estimation for finite element solutions of Helmholtz equation. Part I & II.

Residual-based estimators in 3D with high-order FE and DG:

W. D¨orfler and S. Sauter, Comput. Meth. Appl. Math., 2013: A posteriori error estimation for highly indefinite Helmholtz problems. S. Sauter and J. Zech, SIAM J. Numer. Anal., 2015:

A posteriori error estimation of hp − dG finite element methods for highly indefinite Helmholtz problems.

Recent works using equilibrated fluxes:

S. Congreve, J. Gedicke, I. Perugia, SIAM J. Sci. Comp., 2019:

Robust adaptive hp-discontinuous Galerkin finite element methods for the Helmholtz equation. T. Chaumont-Frelet, A. Ern, M. Vohral´ık, submittted, 2019:

On the derivation of guaranteed and p-robust a posteriori error estimates for the Helmholtz equation.

Références

Documents relatifs

We consider the a posteriori error analysis of fully discrete approximations of parabolic problems based on conforming hp-finite element methods in space and an arbitrary

2) Pliez à nouveau en deux dans la seconde diagonale du carré. Le motif doit rester visible. 3) Pliez un tiers du triangle qui vous reste, en rabattant la partie qui ne contient pas

Since the dE(bc)s are an unrelaxed cluster population, and since even in their centers, young stars make up only a few percent of the stellar mass (Lisker et al. 2006), we may be

Grosman, An equilibrated residual method with a computable error approximation for a singularly perturbed reaction-diffusion problem on anisotropic finite element meshes , M2AN

We finally in Section 5 present our second main result, the local efficiency and robustness, with respect to reaction (and also diffusion) dominance and also with respect to the

Lately, there is an increasing interest on the use of goal-oriented error estimates which help to measure and control the local error on a linear or

A posteriori error estimates for finite element discretization of (1.1a)–(1.1b) have been a pop- ular research subject starting from the Babuˇska and Rheinboldt work [8]. One

The pragmatics of reference in children language: some issues in developmental theory, in Social and Functional Approaches to Language and Thought (pp.. The development