You are on page 1of 301

Lecture

Notes by Professor B. Cockburn

1. Discontinuous Galerkin Methods (11/07/2017)

2. Discontinuous Galerkin Methods for Computational Fluid Dynamics

3. Static condensation, hybridization, and the devising of the HDG methods

4. HDG methods for hyperbolic problems


Discontinuous Galerkin Methods

Bernardo Cockburn

School of Mathematics
University of Minnesota

U. Politecnica de Catalunya
Summer School
July 11-14, 2017

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 1 / 120
Outline I
1 DG methods for first-order PDEs
Super-short historical overview
Why use DG methods?
The original DG method: Linear transport
DG methods for linear symmetric hyperbolic systems
The RKDG methods for nonlinear hyperbolic conservation laws
2 The HDG methods for diffusion
Static condensation and hybridization
Static condensation of the continuous Galerkin method
Static condensation of mixed methods
Definition of the HDG methods
Formulations of the HDG methods
Existence and Uniqueness
Brief overview of the evolution of the HDG methods
Ongoing work
References
Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 2 / 120
A short (and biased) historical overview of the DG methods

First DG method introduced in 1973 by Reed and Hill for linear transport.
First studied in 1974 by Lesaint and Raviart.
Extended to nonlinear hyperbolic conservation laws in the 90’s by B.C. and
C.-W. Shu.
Extended to compressible flow in 1997 first by F. Bassi and S. Rebay.
New DG methods for diffusion appear and some old ones (the IP methods of
the late 70’s) are resuscitated. A unified analysis is proposed in 2002 by D.
Arnold, F. Brezzi, B.C. and D. Marini.
Explosive extension to a wide variety of equations.
They clash with the well-established mixed and continuous Galerkin
methods. In response, the HDG methods are introduced in 2009 by B.C., J.
Gopalakrishnan and R. Lazarov. The HDG methods are strongly related to
the hybrid methods and to the hybridization techniques of the mid 60’s
introduced as implementation techniques for mixed methods.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 3 / 120
A short (and biased) historical overview of the DG methods

B.C., G. Karniadakis, C.-W. Shu, The development of Discontinuous


Galerkin methods, in Discontinuous Galerkin methods. Theory, computation
and applications, Lecture Notes in Computational Sicence and Engineering,
Volume 11, Springer, 2000.
B.C. and C.-W. Shu, Runge-Kutta Discontinuous Galerkin methods for
convection-dominated problems, J. Sci. Comput. 16 (2001), pp. 173–261.
D. Arnold, F. Brezzi, B.C. and D. Marini, Unified analysis of discontinuous
Galerkin methods for elliptic problems, SINUM 39 (2002), pp. 1749–1779.
B.C., Discontinuous Galerkin methods, ZAMM Z. Angew. Math. Mech. 83
(2003), pp. 731–754.
B.C., Discontinuous Galerkin methods for Computational Fluid Dynamics,
Encyclopedia of Computational Mechanics, Volume 3: Fluids, E. Stein, R.
de Borst and T.J.R. Hughes, Eds., Wiley, 2004, pp. 91—123.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 4 / 120
Motivation
Why use DG methods? Good approximation of smooth solutions.

Z Z

Y Y

X X

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

1 1
0 0
0.8 0.8
0 0
0.6 0.6
0.2 0.2
0.4 0.4 0.4 0.4
0.6 0.6
0.2 0.2
0.8 0.8

Approximate solution at T = 0.0 (left), T = 83 π (right) with quadratic


polynomials.

(B.C. and C.-W. Shu, 1990.)

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 5 / 120
Motivation
Why use DG methods? Good approximation of smooth solutions.

0.25 1

0.8

0.6
Y

0.4

0.2

1
0
0.8
0 0
0.6
0.2
0.4 0.4
0.6
0.25 0.5 0.8
0.2

Approximate solution at T = 43 π with quadratic polynomials.

(B.C. and C.-W. Shu, 1990.)

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 6 / 120
Motivation
Why use DG methods? Local postprocessing enhances the accuracy.

2 2
P , before post-processing P , after post-processing
10-2 10-2
|error|

|error|
N=10

10-4 N=20 10-4


N=10
N=40

-6 N=80 -6
10 10 N=20

N=160

-8 -8 N=40
10 10

N=80
-10 -10
10 10

N=160

-12 -12
10 10
1 2 3 4 5 6 1 2 3 4 5 6
x x

The absolute value of the errors for P 2 with N=10, 20, 40, 40, 80 and 160
elements. Before post-processing (left) and after post-processing (right).

(B.C., M. Luskin, C.-W. Shu and E. Suli, 2003).


Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 7 / 120
Motivation
Why use DG methods? Good approximation of discontinuities.

-0.22
0.75

-0.23
0.5

0.25 -0.24

0 -0.25

-0.25
-0.26
0 0.25 0.5 0.75 1

Inviscid Burgers’ equation.

(B.C. and C.-W. Shu, 1990).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 8 / 120
Motivation
Why use DG methods? Good approximation of contacts and shocks.

Rectangles P2, ∆ x = ∆ y = 1/480

0.4

0.3

0.2

0.1

0.0

2.0 2.2 2.4 2.6 2.8

Isolines of the density for the Double Mach reflection problem.

(B.C. and C.-W. Shu, 1998).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 9 / 120
Motivation
Why use DG methods? Ideally suited for adaptivity.

0.8
7 (1%)
0.6

0.4

0.2 6 (17%)

−0.2
5 (59%)
−0.4

−0.6

−0.8
4 (23%)
−1
−1 −0.5 0 0.5 1

Subsonic flow around a NACA0012 airfoil. The hp–mesh has 325


elements, 45008 degrees of freedom, and produces an error
|J(u) − J(uh )| = 3.756 × 10−7 .

(Houston and Suli, 2002).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 10 / 120
Motivation
Why use DG methods? Ideally suited for adaptivity.

1
6 (1%)
0.8

0.6
5 (12%)
0.4

0.2

0 4 (42%)

−0.2

−0.4 3 (44%)
−0.6

−0.8
2 (1%)
−1
−1 −0.5 0 0.5 1

Supersonic flow around a NACA0012 airfoil: The hp–mesh has 783


elements, 69956 degrees of freedom, and produces an error of
|J(u) − J(uDG )| = 1.311 × 10−4 .

(Houston and Suli, 2002).


Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 11 / 120
The original DG method.
Transport of neutrons.

The original DG method was devised for numerically solving equations


modeling the transport of neutrons. A simplified version of that model is
the following:

σ u + ∇ · (a u) = f in Ω,
u = uD on ∂Ω− ,

where σ > 0 , a is a constant vector and ∂Ω− the inflow boundary of


Ω ⊂ Rd , that is, ∂Ω− = {x ∈ ∂Ω : a · n (x ) < 0}.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 12 / 120
The original DG method
Transport of neutrons.

Triangulation Ωh = {K } of Ω and boundary data uD on ∂Ω− .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 13 / 120
The original DG method.
Rewriting the equations.

Set ub := uD on ∂Ω− .
Given ub on ∂K− , compute u by solving

σ u + ∇ · (a u) = f in K ,
u = ub on ∂K − .

Given u in K , set ub := u on ∂K \ ∂K − .

Here ∂K − := {x ∈ ∂K : a · n(x ) < 0}.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 14 / 120
The original DG method
Solving the equations.

Given ub on ∂K − (left), compute u on K (right).

a a

K K

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 15 / 120
The original DG method
Solving the equations.

Set ub := u on ∂K \ ∂K − (left). The computation on other elements can


now proceed (right).

a a

K K

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 16 / 120
The original DG method
Solving the equations.

Given ub := uD on ∂Ω− , compute u (left) and then obtain ub (right).

a a

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 17 / 120
The original DG method
Solving the equations.

Given ub, compute u (left) and then obtain ub (right).

a a

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 18 / 120
The original DG method
The weak formulation on each element.

Given ub on ∂K − , we have that u satisfies the weak formulation

σ (u, w )K − (u, a · ∇w )K + ha · n u, w i∂K \∂K −


= (f , w )K − ha · n ub, w i∂K − ,

for all w ∈ W (K ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 19 / 120
The original DG method
The Galerkin method on each element.

The Galerkin method on the element K ∈ Ωh is defined as follows. We


take uh in the space W (K ) and determine it by requiring that

σ (uh , w )K − (uh , a · ∇w )K + ha · n uh , w i∂K \∂K −


= (f , w )K − ha · n ubh , w i∂K − ,

for all w ∈ W (K ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 20 / 120
The original DG method
Implementation.

Set ubh := uD on ∂Ω− .


Given ubh on ∂K− , compute uh in K as the element of W (K ) such
that

σ (uh , w )K − (a uh , ∇w )K + ha · n uh , w i∂K \∂K −


=(f , w )K − ha · n ubh , w i∂K − ,

for all w ∈ W (K ).
Given uh in K , set ubh := uh on ∂K \ ∂K− .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 21 / 120
The original DG method.
The stabilization mechanism. The jumps uh − ubh stabilize the method.

The energy identity for the exact solution is

σku − f /2σk2L2 (Ω) + h|a · n |u, ui∂Ω+ = Ψ(f , uD ),


1
2
and for the approximate solution,

σkuh − f /2σk2L2 (Ω) + h|a · n |uh , uh i∂Ω+ + Θh (uh − ubh ) = Ψ(f , uD ),


1
2
where Θh (uh − ubh ) := 12 K ∈Ωh h|a · n |(uh − ubh ), uh − ubh i∂K − .
P

The method is stabilized by the term Θh (uh − ubh ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 22 / 120
The original DG method.
The stabilization mechanism. The jumps uh − ubh control the residuals.

The Galerkin formulation on the element K reads

σ (uh , w )K − (uh , a · ∇w )K + ha · n ubh , w i∂K = (f , w )K ∀ w ∈ W (K ),

or, equivalently,

(RK , w )K = hR∂K , w i∂K ∀ w ∈ W (K ),

where RK := σ uh + ∇ · (a uh ) − f and R∂K := a · n (uh − ubh ).

Thus, the L2 -projection of RK into W (K ) is controlled by the jumps


R∂K = a · n (uh − ubh ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 23 / 120
The original DG method.
The stabilization mechanism. The case of non-smooth solutions.

The exact solution u in the element K is not smooth.


The residual RK is big.
The jump R∂K = |a · n |(uh − ubh ) is big.
The dissipation produced by Θh (uh − ubh ) damps the spurious
oscillations.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 24 / 120
The original DG method.
Convergence properties. The spaces and the triangulations.

special triangulations Ωh made of shape-regular simplexes K ,


W (K ) := Pk (K ),

A special triangulation for a = (1, 0).


Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 25 / 120
The original DG method.
Convergence properties. Another triangulation.

Another special triangulation for a = (1, 0).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 26 / 120
The original DG method.
Convergence properties. The auxiliary projection.

We can find projections such that the projection of the errors


Π : H 1 (K ) → W (K ), εu := Π(u − uh ),
P∂ : L2 (F ) → M(F ), εub := P∂ (u − ubh ),
satisfy
a · n εub = a · n εbu ,
for all w ∈ W (K ),

σ (εu , w )K − (εu , a · ∇w )K + ha · n εbu , w i∂K = σ (Πu − u, w )K ,

εbu = 0 on ∂Ω− .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 27 / 120
The original DG method.
Convergence properties. The jumps uh − ubh are controlled by the projection.

From the energy identity


1 σ
σkεu − (Πu − u)k2L2 (Ω) + Θh (εu − εbu ) = kΠu − uk2L2 (Ω) ,
2 4
we deduce that
1/2
ku − uh kL2 (Ω) + σ −1/2 Θh (εu − εbu ) ≤ C kΠu − ukL2 (Ω)
≤ C | u |H k+1 (Ωh ) hk+1 .

Thus, optimal convergence orders are obtained for smooth solutions.


(B.C., B. Dong and J. Guzmán, 2008.)

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 28 / 120
The original DG method.
Conclusion.

We have seen that the original DG method:


Uses discontinuous approximations for both the solution inside each
element and its trace on the element boundary.
Uses a Galerkin method to weakly enforce the equations on each
element.
Is devised so that they can be efficiently implemented.
Has a stabilization mechanism that allows it to damp away spurious
oscillations and reach optimal orders of convergence at the same time.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 29 / 120
DG methods for linear symmetric hyperbolic systems

u t + ∇ · F(u ) + Bu = f in Ω × (0, T ),
u=g at t = 0,
F(u ) n − Nu = 0 on ∂Ω × [0, T ].

Here (F(u ))ij := m


P
ℓ=1 (Aj )iℓ uℓ , and Aj , j = 1, . . . , N, are constant,
symmetric matrices.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 30 / 120
DG methods for linear symmetric hyperbolic systems
Friedrichs’ result.

In 1958, Friedrichs showed that the above problem has a unique solution if

N + N⋆ ≥ 0,
B + B ⋆ ≥ σ Id, σ ≥ 0,
ker (An − N) + ker (An + N) = Rm .

nj . Note that F(u ) n = An u .


PN
Here An := i=1 Aj

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 31 / 120
DG methods for symmetric hyperbolic systems
Acoustics: The first-order system

∂2u
ρ − ∇ · (A∇u) = f in Ω × (0, T ).
∂t 2

∂q
c − ∇v = 0 in Ω × (0, T ),
∂t
−∇·q = f
∂v
ρ in Ω × (0, T ).
∂t

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 32 / 120
DG methods for linear symmetric hyperbolic systems
Elastodynamics: The first-order system

∂2u
ρ − ∇ · [µ∇u + (µ + λ)(∇ · u )I] = b in Ω × (0, T ).
∂t 2

− ∇v
∂H
= 0 in Ω × (0, T ),
∂t
∂v
ρ − ∇ · (µH + pI) = b in Ω × (0, T ),
∂t
−∇·v = 0
∂p
ǫ in Ω × (0, T ).
∂t

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 33 / 120
DG methods for linear symmetric hyperbolic systems
Maxwell’s equations

∂H
µ + ∇ × E = 0,
∂t
∂E
ǫ − ∇ × H = 0,
∂t

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 34 / 120
Space discretization of linear symmetric hyperbolic systems
The Galerkin method and the numerical flux.

We take u h (t) in the space W (K ) and determine it by requiring that


D E
((u h )t , w )K − (F(u h ), ∇w )K + Fdh n , w = (f , w )K ,
∂K

for all w ∈ W (K ), where


h n := An+ ( (u h + u h )) + Nn± (u h − u h )
d 1 + − 1 + −
F
2 2
The matrix Nn± is called the dissipation matrix.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 35 / 120
Space discretization of linear symmetric hyperbolic systems
Examples of numerical fluxes.

• In the scalar case, F(u ) = a u. For the original DG method: F


d h n is the
upwinding numerical flux An+ := a · n and Nn± = |a · n |.
+

h n in the general case is obtained as


d
• The upwinding numerical flux F
follows:
Diagonalize An = P −1 Λ P,
Set Nn± := P −1 | Λ | P.

h n is obtained as follows:
d
• The Lax-Friedrichs numerical flux F
Diagonalize An = P −1 Λ P,
Set Nn± := λmax Id.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 36 / 120
Space discretization of linear symmetric hyperbolic systems
Main properties of the DG method.

The jumps u h + − u h − stabilize the method when Nn ± is positive


definite.
The jumps control the residuals.
Spurious oscillations are damped in the presence of discontinuities.
The method converges with order k+1/2.
After a local postprocessing, with order 2k + 1 for locally uniform
grids.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 37 / 120
Time discretization of linear symmetric hyperbolic systems
Main strategies.

Explicit Runge-Kutta methods: SSP methods. (C.-W. Shu 88, S.


Gottlieb, C.-W. Shu and E. Tadmor, 2001.)
Space-time methods: Locally implicit. (R. Haber; J. van der Vegt; R.
Falk and G. Richter, 1999; see also, Gopalakrishnan, Schöberl and
Winterstiger, 1917)
Globally implicit methods: Efficient multigrid techniques. (J. van der
Vegt; P. Persson and J. Peraire.)

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 38 / 120
The original DG method.
Dispersion and dissipation properties.

For the semidiscrete transport equation:

Order of dispersion: 2 k + 3.
Order of dissipation: 2 k + 2.

(M. Ainsworth, 2004).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 39 / 120
The original DG method.
Dispersion and dissipation properties.

1 1 1

0 0 0

0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

1 1 1

0 0 0

0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

Effect of the polynomial degree on the approximation of discontinuities.


(B.C. and C.-W. Shu, 2001.)
Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 40 / 120
DG methods for linear symmetric hyperbolic problems.
Conclusion.

We have devised DG methods that:


Use discontinuous approximations for both the solution inside each
element and its trace on the element boundary.
Use a Galerkin method to weakly enforce the equations on each
element.
Have a stabilization mechanism that allows it to damp away spurious
oscillations and reach almost optimal orders of convergence at the
same time.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 41 / 120
The RKDG methods.
Non-linear hyperbolic problems.

ut + ∇ · f(u) = 0.
Pd ∂fi
Hyperbolic: i=1 ∂u (u) ni is diagonalizable and has real eigenvalues.

Example 1: The Euler equations of gas dynamics:

ρt + (ρ vj ),j = 0,
(ρ vi )t + (ρ vi vj − σij ),j = fi ,
(ρ e)t + (ρ e vj − σij vi ),j = fi vi ,
p
where σij = −p δij and e = (γ−1) ρ + 12 | v |2 .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 42 / 120
The RKDG methods.
Non-linear hyperbolic problems.

Example 2: Isentropic gas dynamics in Lagrangian coordinates,

τt − ux = 0,
ut + (p(τ ))x = 0,

and in Eulerian coordinates

ρt + (v ρ)x = 0,
(ρ v )t + (ρ v 2 + p(ρ−1 ))x = 0,

where p(τ ) = Aτ −γ for a polytropic ideal gas.


Example 3: Scalar hyperbolic conservation law:

ut + ∇ · f(u) = 0.

Inviscid Burgers equation: 1D and f(u) = u2 /2.


Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 43 / 120
The RKDG methods.
Non-linear hyperbolic problems.

Main difficulties:
Convergence to the physically relevant solution must be ensured.
An additional mechanism to properly capture discontinuities is needed.
Implicit methods are very inefficient in the presence of discontinuities.

Solution:
DG-space discretization with suitable numerical traces
(approximate Riemann solvers).
SSP, explicit time-marching algorithms.
Slope limiters (part of an artificial viscosity hidden term!).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 44 / 120
The RKDG methods.
Non-linear hyperbolic problems.

Examples of numerical fluxes:


The Godunov flux:
fb(a, b) = min f (u), if a ≤ b,
a≤u≤b

fb(a, b) = max f (u), otherwise.


b≤u≤a

The Engquist-Osher flux:


Z b
fb(a, b) = min(f ′ (s), 0) ds
0
Z a
+ max(f ′ (s), 0) ds + f (0).
0

The Lax-Friedrichs flux:


1
fb(a, b) = [f (a) + f (b) − C (b − a)],
2
C= 0
max 0 |f ′ (s)|.
inf u (x)≤s≤sup u (x)

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 45 / 120
The RKDG methods.
Non-linear hyperbolic problems.

Development of the RKDG method:


1982: G.Chavent and G.Salzano: Use the DG-space discretization
with Godunov flux.
1989: G.Chavent and B.C.: Incorporate the slope limiter.
1991: B.C. and C.-W.Shu: Incorporate an SSP time-marching
method: First RKDG method.
89-98: B.C. and C.-W.Shu (+S.Hou+S.Lin) : RKDG methods.
A parallel development:
1987: Allmaras and Giles: Euler equations.
1989: Allmaras: P 1 and 3-stage second-order RK.
1991: Halt and Agarwall
1992: Halt: high polynomial degree.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 46 / 120
The RKDG method

We construct the RKDG methods for the non-linear hyperbolic model


problem

ut + f (u)x = 0, in (0, 1) × (0, T ),


u(·, 0) = u0 (·) on (0, 1),
u(0+, ·) = u(1−, ·) on (0, T ),

The main components of the RKDG methods are:


A DG space discretization,
A strongly-stable RK time-marching discretization,
A generalized slope limiter,

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 47 / 120
• Discontinuous Galerkin discretization in space

The approximate solution uh restricted to the interval Ij belongs to the


space P(Ij ).

The non-linear conservation law element-by-element by requiring that for


every function vh in the space P(Ij )
xj+1/2

b
((uh )t , v )Ij − (f (uh ), (v )x )Ij + f (uh ) v = 0,
xj−1/2

where fb(uh ) is the so-called numerical flux has the following general form:

fb(uh )(xj+1/2 ) = fb(uh (xj+1/2


− +
), uh (xj+1/2 )).

Monotone schemes are obtained with k = 0.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 48 / 120
• Strong-Stability-Preserving RK methods
d
Each time step for dt uh = L(uh ) is of the form

(0)
1 set uh = uhn ;
2 for i = 1, ..., K compute the intermediate
functions:
i−1
X
(i)
uh = αil whl ,
l=0
(l) βil (l)
whl = uh + ∆t n Lh (uh );
αil

3 set uhn+1 = uhK .

Note that αil ∈ [0, 1], and that, if αil = 0, then βil = 0.
Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 49 / 120
Set
wh = uh + δLh (uh ) ≡ EULER(uh ; δ),
and assume that
| wh | ≤ | u h | ∀| δ | ≤ δ0 .
Then
i−1
X i−1
X
(i) (l)
| uh | ≤ αil | whl |≤ αil | uh |,
l=0 l=0

provided that
βil
∆t n ≤ δ0 .
αil
This implies that

| uhn | ≤ | uh0 | ∀n = 0, . . . , N.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 50 / 120
The Euler step is non-increasing in | · | if:

We take the semi-norm | · | to be


X
| uh | ≡ | u j+1 − u j |,
j

1
R
where u j = ∆j Ij uh (x) dx.
We take !
| fb(a, ·) |Lip | fb(·, b) |Lip
δ0−1 = 2 + .
∆j+1 ∆j
We assume that the following sign conditions are satisfied:
+ +
sign (uj+1/2 − uj−1/2 ) = sign (u j+1 − u j ),
− −
sign (uj+1/2 − uj−1/2 ) = sign (u j − u j−1 ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 51 / 120
•The generalized slope limiter

Since the sign conditions are not automatically satisfied, we enforce them
by means of a simple projection called the generalized slope limiter, ΛΠh .

It is indeed possible to construct generalized slope limiters that enforce the


sign conditions which, moreover, have the following properties:

Is a projection into the finite element space.


Leaves the averages unchanged.
Leaves a linear function unchanged.
Can be efficiently parallelized.

The slope limiter of the MUSCL scheme is the prototypical example.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 52 / 120
•The RKDG method

Set uh0 = ΛΠh Ph u0 .


For n = 0 until N − 1 do:

(0)
1 set uh = uhn ;
2 for i = 1, ..., K compute:
i−1
!
(i)
X
uh = ΛΠh αil whl ,
l=0
(l) βil
whl = EULER(uh ; ∆t n );
αij

3 set uhn+1 = uhK .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 53 / 120
We have the following boundedness result.
Theorem
Assume that
βil
∆t n ≤ δ0 .
αij
Then, we have that
| uhn | ≤ | u0 |TV (0,1) ,
where uhn is given by an RKDG scheme.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 54 / 120
The RKDG methods.
Non-linear hyperbolic problems.

Positivity-preserving RKDG methods (Shu et al.).


How to avoid the use of slope limiters?
Rigorous error analysis for shocks?

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 55 / 120
The HDG methods for diffusion
Static condensation of the exact solution.

We provide a ”static condensation” characterization of the solution of the


following second-order elliptic model problem:

c q + ∇u = 0 in Ω,
∇·q =f in Ω,
ub = uD on ∂Ω.

Here c is a matrix-valued function which is symmetric and uniformly


positive definite on Ω.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 56 / 120
The HDG methods for diffusion
Static condensation of the exact solution: Local problems and transmission conditions.

We have that the exact solution satisfies the local problems

c q + ∇u = 0 in K ,
∇·q =f in K ,

the transmission conditions

[[b
u ]] = 0 if F ∈ Eoh ,
q ]] = 0
[[b if F ∈ Eoh ,

and the Dirichlet boundary condition

ub = uD if F ∈ E∂h .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 57 / 120
The HDG methods for diffusion
Static condensation of the exact solution: Rewriting the equations.

We can obtain (q , u) in K in terms of ub on ∂K and f by solving

c q + ∇u = 0 in K ,
∇·q =f in K ,
u = ub on ∂K .

The function ub can now be determined as the solution, on each F ∈ Eh , of


the equations

q ]] = 0
[[b if F ∈ Eoh ,
ub = uD if F ∈ E∂h ,

where q
b is the trace of q = q (bu, f ) on ∂K .
Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 58 / 120
The HDG methods for diffusion
Static condensation of the exact solution: A characterization of the solution.

We have that (q , u) = (Qub, Uub) + (Qf , Uf ), where

c Qub + ∇Uub = 0 in K , c Qf + ∇Uf = 0 in K ,


∇ · Qub = 0 in K , ∇ · Qf =f in K ,
Uub = ub on ∂K , Uf =0 on ∂K .

The function ub can now be determined as the solution, on each F ∈ Eh , of


the equations
b ub]] = [[Q
− [[Q b f ]] if F ∈ Eoh ,
ub = uD if F ∈ E∂h .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 59 / 120
The HDG methods for diffusion
Static condensation of the exact solution. The one-dimensional case K = (xi−1 , xi ) for
i = 1, . . . , I , with c = 1.

We have that (q , u) = (Qub, Uub) + (Qf , Uf ), where

d d
Qub + Ub = 0 in (xi−1 , xi ), Qf + Uf = 0 in (xi−1 , xi ),
dx u dx
d d
Qb = 0 in (xi−1 , xi ), Qf =f in (xi−1 , xi ),
dx u dx
Uub = ub on {xi−1 , xi }, Uf =0 on {xi−1 , xi }.

The function ub is the solution of


b ub(x + ) − Q
Q b ub(x − ) = −Q
b f (x + ) + Q
b f (x − ) for i = 1, . . . , I − 1,
i i i i
ub(xi ) = uD (xi ) for i = 0, I .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 60 / 120
The HDG methods for diffusion
Static condensation of the exact solution. The one-dimensional case K = (xi−1 , xi ) for
i = 1, . . . , I , with c = 1.

We have that (q , u) = (Qub, Uub) + (Qf , Uf ), where, for x ∈ (xi−1 , xi ),


Z xi
1
u (x) = − (b
Qb ui − ubi−1 ), Qf (x) = − Gxi (x, s)f (s) ds,
h xi−1
Z xi
Ub
u (x) = ϕi (x) u
bi + ϕi−1 (x) ubi−1 Uf (x) = G i (x, s)f (s) ds.
xi−1

The function ub is the solution of


Z xi+1
1
− (bui−1 − 2 ubi + ubi+1 ) = ϕi (s) f (s) ds for i = 1, . . . , I − 1,
h xi−1
ub(xi ) = uD (xi ) for i = 0, I .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 61 / 120
The HDG methods for diffusion
Static condensation of the continuous Galerkin method. (Guyan 65)

The continuous Galerkin method provides an approximation to u,


uh ∈ Wh (uD ), determined by

(a ∇uh , ∇w )Ω = (f , w )Ω ∀w ∈ Wh (0).

where

Wh = {w ∈ C0 (Ω) : w |K ∈ W (K ) ∀K ∈ Ωh },
Wh (g ) = {w ∈ Wh : w = Ih (g ) on ∂Ω}.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 62 / 120
The HDG methods for diffusion
Static condensation of the continuous Galerkin method. Splitting the degrees of freedom.

For each element K ∈ Ωh ,


W (K ) = W0 (K ) ⊕ W∂ (K ),
W0 (K ) := {w ∈ W (K ) : w |∂K = 0},
W∂ (K ) := {w ∈ W (K ) : w |∂K = 0 =⇒ w |K = 0}.
This implies
Wh = W0,h ⊕ WEh
W0,h := {w ∈ Wh : w |K ∈ W0 (K ) ∀K ∈ Ωh },
WEh := {w ∈ Wh : w |K ∈ W∂ (K ) ∀K ∈ Ωh },
and
Mh := {w |Eh : w ∈ Wh },
Mh (g ) := {µ ∈ Mh : µ|∂Ω = Ih (g )}.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 63 / 120
The HDG methods for diffusion
Static condensation of the continuous Galerkin method. Local problems and transmission
condition.

We obtain U ∈ W (K ) in terms of ubh and f by solving

(a ∇U, ∇w )K = (f , w )K ∀w ∈ W0 (K ),
U = ubh on ∂K .

The function ubh ∈ Mh is determined as the solution of

(a ∇U, ∇w )Ω = (f , w )Ω ∀w ∈ WEh (0),


ubh = Ih (uD ) on ∂Ω.

Note that we have a transmission condition:

0 =ha ∇U · n , w
b i∂Ωh − (∇ · (a ∇U) + f , w )Ωh = ha ∇U · n + r∂ , w
b i∂Ωh

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 64 / 120
The HDG methods for diffusion
Static condensation of the CG method: A characterization of the approximate solution.

We have that uh = Uubh + Uf , where

(a ∇Uubh , ∇w )K = 0 ∀w ∈ W0 (K ),
Uubh = ubh on ∂K ,
(a ∇Uf , ∇w )K = (f , w )K ∀w ∈ W0 (K ),
Uf = 0 on ∂K ,

and ubh is the element of Mh (uD ) that solves the global problem

(a ∇Uubh , ∇Uµ )Ω = (f , Uµ )Ω ∀ µ ∈ Mh (0).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 65 / 120
The HDG methods for diffusion
Static condensation of the CG method: The original one (Guyan 65)!

The system of equations is

K [uh ] = [f ],

and, after splitting the degrees of freedom, it is


    
K00 K0∂ [U] f
= 0 .
K∂0 K∂∂ [b uh ] f∂

The solution of the local problems is


−1 −1
[U] = −K00 K0∂ [b
uh ] + K00 [f 0 ].

and the transmission condition


−1 −1
(−K∂0 K00 uh ] = −K∂0 K00
K0∂ + K∂∂ )[b [f 0 ] + [f ∂ ].

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 66 / 120
The HDG methods for diffusion
Static condensation of the CG method: The 1D case.

For W (K ) := Pk (K ), the solution of the local problems are


Z xi
Uub(x) = ϕi (x) ubi + ϕi−1 (x) ubi−1 Uf (x) = Ghi (x, s)f (s) ds,
xi−1

and where the global problem for the values {b ui }N


i=0 is
Z xi+1
1
− (bui−1 − 2 ubi + ubi+1 ) = ϕi (s) f (s) ds for i = 1, . . . , N − 1,
h xi−1
ubj = uD (xj ) for j = 0, N.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 67 / 120
The HDG methods for diffusion
Static condensation of mixed methods (deVeubeke 65).

The function (q h , uh ) is the only element of Vh × Wh satisfying the


equations

(c q h , v )Ω − (uh , ∇ · v )Ω = −huD , v · n i∂Ω ∀v ∈ Vh ,


(∇ · q h , w )Ω = (f , w )Ω ∀w ∈ Wh .

where

Vh = {v ∈ H (div , Ω) : v |K ∈ V (K ) ∀K ∈ Ωh }.
Wh = {w ∈ L2 (Ω) : w |K ∈ W (K ) ∀K ∈ Ωh }.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 68 / 120
The HDG methods for diffusion
Static condensation of mixed methods: Local problems and transmission conditions.

We define (Q, U) ∈ V (K ) × W (K ) in terms of ubh and f as the solution of


the local problem

(c Q, v )K − (U, ∇ · v )K = hb
uh , v · n i∂K ∀v ∈ V (K ),
(∇ · Q, w )K = (f , w )K ∀w ∈ W (K ).

The function ubh in the space Mh is such that

[[Q]] = 0 on Eoh ,
ubh = uD on ∂Ω.

The weak form of the transmission condition is

hQ · n , µi∂Ωh = hQ, µi∂Ωh \∂Ω = h [[Q]], µiEoh = 0 ∀µ ∈ Mh (0).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 69 / 120
The HDG methods for diffusion
Static condensation of mixed methods: A characterization of the approximate solution.

We have that (q h , uh ) = (Qubh , Uubh ) + (Qf , Uf ), where, ∀K ∈ Ωh ,

(c Qµ , v )K − (Uµ , ∇ · v )K = −hµ, v · n i∂K ∀ v ∈ V (K ),


(∇ · Qµ , w )K = 0 ∀ w ∈ W (K ),
(c Qf , v )K − (Uf , ∇ · v )K = 0 ∀ v ∈ V (K ),
(∇ · Qf , w )K = (f , w )K ∀ w ∈ W (K ),

and the function ubh is the element of Mh (uD ) which solves the global
problem

(c Qubh , Qµ )Ωh = (f , Uµ )Ωh ∀ µ ∈ Mh (0).

Note that
0 = hQ · n , µi∂Ωh = hQb
uh · n , µi∂Ωh + hQf · n , µi∂Ωh = −(c Qb
uh , Qµ )Ωh + (Uµ , f )∂Ωh .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 70 / 120
The HDG methods for diffusion
Static condensation of mixed methods: The original hybridization (deVeubeqe 65)!

The system of equations is


    
A B [q h ] [uD ]
= .
Bt 0 [uh ] [f ]
which, after hybridization, becomes
    
A B C [Q] −C∂ [uD ]
B t 0 0   [U]  =  [f ]  .
Ct 0 0 uh ]
[b 0
The solution of the local problems is
   −1  
[Q] A B uh ] − C∂ [uD ]
−C [b
= .
[U] Bt 0 [f ]
and the transmission condition is H[ ubh ] = H∂ [ uD ] + J [ f ],
H := C t (A−1 − A−1 B (B t A−1 B)−1 B t A−1 ) C .
Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 71 / 120
The HDG methods for diffusion
Static condensation of mixed methods: The 1D case.

For V (K ) × W (K ) := Pk+1 (K ) × Pk (K ), the solution of the local


problems is
Z xi
ubi − ubi−1
Qub(x) = − , Qf (x) = Hhi (x, s)f (s) ds,
h xi−1
Z xi
b b
Uub(x) = ϕi (x) ui + ϕi−1 (x) ui−1 , Uf (x) = Ghi (x, s)f (s) ds,
xi−1

and the global problem for the values {b ui }Ni=0 is


Z xi+1
1
− (b b b
ui−1 − 2 ui + ui+1 ) = ϕi (s) f (s) ds for i = 1, . . . , N − 1,
h xi−1
ubi = uD (xj ) for i = 0, N.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 72 / 120
The HDG methods for diffusion
Devising HDG methods: The main idea

The HDG methods are obtained by constructing discrete versions


(based on discontinuous Galerkin methods) of the above
characterization of the exact solution.

In this way, the globally coupled degrees of freedom will be those of


the corresponding global formulations.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 73 / 120
The HDG methods for diffusion
Devising HDG methods. (B.C., J.Gopalakrishnan and R.Lazarov, SINUM, 2009.) The local problems: A weak
formulation on each element.

On the element K ∈ Ωh , we define (q h , uh ) terms of (b


uh , f ) as the element
of V (K ) × W (K ) such that

(c q h , v )K − (uh , ∇ · v )K + hb
uh , v · n i∂K = 0,
−(q h , ∇w )K + hb
q h · n, w i∂K = (f , w )K ,
for all (v , w ) ∈ V (K ) × W (K ), where

qbh · n = q h · n + τ (uh − ubh ) on ∂K .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 74 / 120
The HDG methods for diffusion
Devising HDG methods. The global problem: The weak formulation for ubh .

For each face F ∈ Eoh , we take ubh |F in the space M(F ). We determine ubh
by requiring that,

q h ]]iF = 0 ∀ µ ∈ M(F )
hµ, [[b if F ∈ Eoh ,
ubh = uD if F ∈ E∂h .

All the HDG methods are generated by choosing the local spaces
V (K ), W (K ), M(F ) and the stabilization function τ .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 75 / 120
Formulation for (q h , q
b h , uh , ubh )
Characterization of the approximate solution (B.C., J.Gopalakrishnan and R.Lazarov, SINUM, 2009.).

The approximate solution (q h , uh , ubh ) is the element of the space


V h × Wh × Mh (uD ) satisfying the equations
(c q h , v )Ωh − (uh , ∇ · v )Ωh + hb
uh , v · n i∂Ωh = 0,
−(q h , ∇w )Ωh + hb
q h · n, w i∂Ωh = (f , w )Ωh ,
hµ, qb h · n i∂Ωh = 0,

for all (v , w , µ) ∈ V h × Wh × Mh (0), where

qbh · n = q h · n + τ (uh − ubh ) on ∂Ωh .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 76 / 120
The HDG methods.
The transmission condition.

q h ]] = 0 on a face
Suppose that the transmission condition implies that [[b
F ∈ Eoh . Then, on that face, we have that

[[q h ]] + τ + (uh + − ubh ) + τ − (uh − − ubh ) = 0,

which holds if
τ + uh + + τ − uh −
[[q ]],
1
ubh = + +
τ+ + τ− τ + τ− h
qbh = τ qτh + ++ ττ −q h + τ +τ +τ τ − [[uh ]]
− + + − + −

provided τ + + τ − > 0.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 77 / 120
Formulation for (uh , ubh )
Characterization of the approximate solution (D.Arnold and F.Brezzi, RAIRO, 1985; ABCD, SINUM, 02; B.C.
and K.Shi, C&F, 2014.)

For any (w , µ) ∈ Wh × Mh , define q w ,µ ∈ V h as the solution of


(c q w ,µ , v )Ωh − (w , ∇ · v )Ω + hµ, v · n i∂Ω = 0,
h h

for all v ∈ Vh .
The approximate solution is (q uh ,buh , uh , ubh ) where (uh , ubh ) is the element
of Wh × Mh (uD ) satisfying the equations

(∇ · q uh ,buh , w )Ωh + hτ (uh − ubh ), w i∂Ωh = (f , w )Ωh ,


hµ, q uh ,buh · n + τ (uh − ubh )i∂Ωh = 0,

for all (w , µ) ∈ Wh × Mh (0).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 78 / 120
Formulation for (uh , ubh )
Characterization of the approximate solution (D.Arnold and F.Brezzi, RAIRO, 1985; ABCD, SINUM, 02; B.C.
and K.Shi, C&F, 2014.)

For any (w , µ) ∈ Wh × Mh , define q w ,µ ∈ V h as the solution of


(c q w ,µ , v )Ωh − (w , ∇ · v )Ω + hµ, v · n i∂Ω = 0,
h h

for all v ∈ Vh .
The approximate solution is (q uh ,buh , uh , ubh ) where (uh , ubh ) is the element
of Wh × Mh (uD ) satisfying the equations

(c q uh ,buh , q w ,µ )Ωh + hµ, q uh ,buh · n i∂Ωh + hτ (uh − ubh ), w i∂Ωh = (f , w )Ωh ,


hµ, q uh ,buh · n + τ (uh − ubh )i∂Ωh = 0,

for all (w , µ) ∈ Wh × Mh (0).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 79 / 120
Formulation for (uh , ubh )
Characterization of the approximate solution (D.Arnold and F.Brezzi, RAIRO, 1985; ABCD, SINUM, 02; B.C.
and K.Shi, C&F, 2014.)

For any (w , µ) ∈ Wh × Mh , define q w ,µ ∈ V h as the solution of


(c q w ,µ , v )Ωh − (w , ∇ · v )Ω + hµ, v · n i∂Ω = 0,
h h

for all v ∈ Vh .
The approximate solution is (q uh ,buh , uh , ubh ) where (uh , ubh ) is the element
of Wh × Mh (uD ) satisfying the equations

(c q uh ,buh , q w ,µ )Ωh + hτ (uh − ubh ), w − µi∂Ωh = (f , w )Ωh ,

for all (w , µ) ∈ Wh × Mh (0).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 80 / 120
Formulation for (uh , ubh )
The associated minimization property. (H. Kabbaria, A. Lew, and B.C., 14; B.C. and K.Shi, 14; B.C. and
J.Shen, 15)

The function (uh , ubh ) minimizes the quadratic functional

Jh (w , µ) := (c q w ,µ , q w ,µ )Ωh + hτ (w − µ), (w − µ)i∂Ωh − (f , w )Ωh ,


1 1
2 2
over the functions (w , µ) ∈ Wh × Mh (uD ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 81 / 120
Formulation for ubh
Characterization of the approximate solution (B.C. and J.Gopalakrishnan, SINUM, 2005; B.C. and
J.Gopalakrishnan and R.Lazarov, SINUM, 2009.)

We have that (q h , uh ) = (Qubh , Uubh ) + (Qf , Uf ) where

uh , 0), U(b
(Qubh , Uubh ) := (Q(b uh , 0)), (Qf , Uf ) := (Q(0, f ), U(0, f )).

where (Q(b uh , f ), U(b


uh , f )) is the linear mapping that associates (b
uh , f ) to
(q h , uh ), and where the numerical trace ubh is the element of the space

Mh (uD ) := {µ ∈ L2 (Eh ) : µ|F ∈ M(F ) ∀ F ∈ Eh , uh |∂Ω := P∂ uD },

satisfying the equations

uh , µ) = ℓh (µ)
ah (b ∀ µ ∈ Mh (0),

b λ · n i∂Ω , and ℓh (µ) := hµ, Q


where ah (µ, λ) := −hµ, Q b f · n i∂Ω .
h h

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 82 / 120
Formulation for ubh
The associated minimization problem (B.C. and K.Shi, C&F, 14; B.C. and J.Shen, 15)

Theorem
We have that

ah (µ, λ) = (cQµ , Qλ )∂Ωh + hτ (Uµ − µ), (Uλ − λ)i∂Ωh ,


ℓh (µ) = (f , Uµ )∂Ωh .

Moreover, ah (·, ·) is positive definite on Mh (0) × Mh (0).

The numerical trace ubh minimizes the quadratic functional


1
Jh (η) := ah (η, η) − ℓh (η),
2
over the functions η in Mh (uD ).
Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 83 / 120
Formulation for ubh
Condition number of the stiffness matrix.

Theorem
If V (K ) = Pk (K ), W (K ) = Pk (K ) and M(F ) = Pk (K ), k ≥ 0,the
condition number of ah (·, ·) (on Mh,0 × Mh,0 ) is of order

(1 + (τ ∗ h)2 )h−2 .

Here τ ∗ := maxK ∈Ωh τ |∂K \FK∗ , where FK∗ is an arbitrary face of the simplex
K.

Note that the matrix is invertible even if τ ≡ 0!

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 84 / 120
Existence and uniqueness.
The local problems are well defined.

Theorem
The local solver on K is well defined if
τ > 0 on ∂K ,
∇W (K ) ⊂ V (K ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 85 / 120
Existence and uniqueness.
Proof.

The system is square. Set ubh = 0 and f = 0.


For (v , w ) := (q h , uh ), the equations read

(c q h , q h )K − (uh , ∇ · q h )K = 0,
−(q h , ∇uh )K + hb
q h · n, uh i∂K = 0.
Hence
(c q h , q h )K + h(b
q h − q h ) · n, uh i∂K = 0,
and since q
b h · n = q h · n + τ (uh ), we get

(c q h , q h )K + hτ (uh ), uh i∂K = 0.

This implies that q h = 0 on K , and that uh = 0 on ∂K .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 86 / 120
Existence and uniqueness.
Proof.

Now, the first equation defining the local problems reads

−(uh , ∇ · v )K = 0,

for all v ∈ V (K ). Hence


(∇uh , v )K = 0,
and so ∇uh = 0. This proves the result.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 87 / 120
Existence and uniqueness.
The numerical trace ubh is well defined.

Theorem
The numerical trace ubh is well defined if, for each K ∈ ∂Ωh ,
τ > 0 on ∂K ,
∇W (K ) ⊂ V (K ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 88 / 120
Existence and uniqueness.
Proof.

The system is square. Set uD = 0 and f = 0. For µ := ubh , the equation


reads
X X
0= hb q h ]]iF = hbuh , qbh · ni∂K =: hbuh , qbh · ni∂Ωh .
uh , [[b
F ∈Eoh K ∈Ωh

Note that

uh , q
−hb b h · n i∂Ωh = − hb
uh , q h · n + τ (uh − ubh )i∂Ωh
uh , q h · n i∂Ωh − huh , τ (uh − ubh )i∂Ωh
= − hb
+ h(uh − ubh ), τ (uh − ubh )i∂Ωh
uh , q h · n i∂Ωh − huh , q
= − hb b h · n i∂Ωh + huh , q h · n i∂Ωh
+ h(uh − ubh ), τ (uh − ubh )i∂Ωh

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 89 / 120
Existence and uniqueness.
Proof.

For (v , w ) := (q h , uh ), the equations of the local problems read

(c q h , q h )K − (uh , ∇ · q h )K + hb
uh , q h · n i∂K = 0,
−(q h , ∇uh )K + hb
q h · n, uh i∂K = 0.
Then

uh , q
−hb b h · n i∂Ωh =(c q h , q h )Ωh + h(uh − ubh ), τ (uh − ubh )i∂Ωh .

uh , q
As a consequence, hb b h · n i∂Ωh = 0 implies q h = 0 on Ωh and uh = ubh
on ∂Ωh .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 90 / 120
Existence and uniqueness.
Proof.

Now, the first equation defining the local problems reads

−(uh , ∇ · v )K + huh , v · n i∂K = 0,

for all v ∈ V (K ). Hence


(∇uh , v )K = 0,
and so ∇uh = 0.

This shows that uh is a constant and, since uh = ubh = 0 on ∂Ω, we can


conclude that uh = 0 on Ωh . We now have that ubh = uh = 0 on ∂Ωh .
This proves the result.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 91 / 120
Devising superconvergent methods.
Superconvergence and postprocessing.

We seek HDG methods for which the local averages of the error u − uh ,
converge faster than the errors u − uh and q − q h .

If this property holds, we introduce a new approximation uh⋆ . On each


element K it lies in the space W ∗ (K ) and defined by

(∇uh ⋆ , ∇w )K = − (cq h , ∇w )K for all w ∈ W ∗ (K ),


(uh ⋆ , 1)K =(uh , 1)K ,

Then u − uh⋆ will converge faster than u − uh . This does happen for mixed
methods!

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 92 / 120
Illustration of the postprocessing.
An HDG method for linear elasticity.(S.-C. Soon, B.C. and H. Stolarski, 2008.)

k=1 k=1

Z Z

X Y X Y

Comparison between the approximate solution (left) and the


post-processed solution (right) for linear polynomial approximations.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 93 / 120
Illustration of the postprocessing.
An HDG method for linear elasticity.(S.-C. Soon, B.C. and H. Stolarski, 2008.)

k=2 k=2

Z Z

X Y X Y

Comparison between the approximate solution (left) and the


post-processed solution (right) for quadratic polynomial approximations.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 94 / 120
Illustration of the postprocessing.
An HDG method for linear elasticity.(S.-C. Soon, B.C. and H. Stolarski, 2008.)

k=3 k=3

Z Z

X Y X Y

Comparison between the approximate solution (left) and the


post-processed solution (right) for cubic polynomial approximations.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 95 / 120
First superconvergent HDG methods.(B.C, B.Dong and J.Guzman, 08; B.C.,
J.Gopalakrishnan and F.-J. Sayas, 10)

The first superconvergent HDG method: the SFH method

Method τ qh uh uh k

RT 0 k +1 k +1 k +2 ≥0
SFH >0 k +1 k +1 k +2 ≥1
LDG-H O(1) k +1 k +1 k +2 ≥1
BDM 0 k +1 k k +2 ≥2

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 96 / 120
Sufficient conditions for superconvergence
The conditions on the local spaces.(B.C., W.Qiu and K.Shi, Math. Comp.,2012 + SINUM, 2012.)

Theorem
Suppose that the local spaces are such that

V (K ) · n + W (K ) ⊂ M(∂K ),
P0 (K ) × P0 (K ) ⊂ ∇W (K ) × ∇ · V (K ) ⊂ V
e (K ) × W
f(K ),

Ve · n ⊕ W

f⊥ = M(∂K ).

Then there is a stabilization function τ such that the HDG method


superconverges.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 97 / 120
Sufficient conditions for superconvergence.
Methods for which M(F ) = Q k (F ), k ≥ 1, and K is a square. (B.C., W.Qiu and K.Shi, Math.
Comp.,2012 + SINUM, 2012.)

method V (K ) W (K )

RT[k] P k+1,k (K ) Q k (K )
×P k,k+1 (K )
TNT[k] Q k (K ) ⊕ H k3 (K ) Q k (K )

HDGQ
[k] Q k (K ) ⊕ H k2 (K ) Q k (K )

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 98 / 120
Sufficient conditions for superconvergence.
Methods for which M(F ) = Q k (F ), k ≥ 1, and K is a cube. (B.C., W.Qiu and K.Shi, Math.
Comp.,2012 + SINUM, 2012.)

method V (K ) W (K )

RT[k] P k+1,k,k (K ) Q k (K )
×P k,k+1,k (K )
×P k,k,k+1 (K )
TNT[k] Q k (K ) ⊕ H k7 (K ) Q k (K )
HDGQ
[k] Q k (K ) ⊕ H k6 (K ) Q k (K )

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 99 / 120
Sufficient conditions for superconvergence.
Methods for which M(F ) = Q k (F ), k ≥ 1, and K is a square or a cube. (B.C., W.Qiu and K.Shi,
Math. Comp.,2012 + SINUM, 2012.)

method τ kq − q h kΩ kΠW u − uh kΩ ku − uh⋆ kΩ

RT[k+1] 0 k +1 k +2 k +2
TNT[k] 0 k +1 k +2 k +2
HDGQ[k] O(1) > 0 k +1 k +2 k +2

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 100 / 120
Sufficient conditions for superconvergence.
TNT in 3D: The space H k7 (K ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 101 / 120
Sufficient conditions for superconvergence.
TNT in 3D: The space H k7 (K ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 102 / 120
Sufficient conditions for superconvergence.
TNT in 3D: The space H k7 (K ).

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 103 / 120
The theory of M-decompositions.
(B.C., G.Fu, F.-J. Sayas, Math. Comp., to appear; B.C. and G.Fu, 2D+3D, M2 AN, to appear)

Definition (The M-decomposition)


We say that V ×W admits an M-decomposition when
(a) tr(V × W ) ⊂ M,
and there exists a subspace V f of
e ×W V ×W satisfying
(b) ∇W × ∇ · V ⊂ V e ×W f,
(c) tr : V
e⊥ × W
f⊥ → M is an isomorphism.
Here V f⊥ are the L2 (K )-orthogonal complements of V
e ⊥ and W e in V , and
f in W , respectively.
of W

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 104 / 120
The theory of M-decompositions.
A characterization of M-decompositions. (B.C., G.Fu, F.-J. Sayas, Math. Comp., to appear)

IM (V × W ) := dim M − dim{v · n |∂K : v ∈ V , ∇ · v = 0}


− dim{w |∂K : w ∈ W , ∇w = 0}.

Theorem
For a given space of traces M, the space V ×W admits an
M-decomposition if and only if
(a) tr(V × W ) ⊂ M,
(b) ∇W × ∇ · V ⊂ V × W ,
(c) IM (V × W ) = 0.
In this case, we have

M = {v · n |∂K : v ∈ V , ∇ · v = 0}⊕{w |∂K : w ∈ W , ∇w = 0},


where the sum is orthogonal.
Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 105 / 120
The theory of M-decompositions.
Construction of M-decompositions. (B.C., G.Fu, F.-J. Sayas, Math. Comp., to appear)

Table: Construction of spaces V × W admitting an M-decomposition, where the


space of traces M(∂K ) includes the constants. The given space Vg × Wg satisfies
the inclusion properties (a) and (b).
V W ∇·V
Vg ⊕ δVfillM ⊕ δVfillW Wg (if ⊃ P0 (K )) = Wg
Vg ⊕ δVfillM Wg (if ⊃ P0 (K )) ⊂ Wg
Vg ⊕ δVfillM ∇ · Vg (if ⊃ P0 (K )) = ∇ · Vg

δV ∇ · δV γδ V dim δ V

δ V fillM {0} ⊂ M, ∩γ Vgs = {0} IM (Vg × Wg )


δ V fillW ⊂ Wg , ∩∇ · Vg = {0} ⊂M IS (Vg × Wg )

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 106 / 120
Construction of M-decompositions

Theorem
Let Vg × Wg satisfy properties (a) and (b) of an M-decomposition.
Assume that δ V fillM satisfies the following hypotheses:
(a) ∇ · δ V fillM = {0},
(b) δ V fillM · n |∂K ⊂ M,
(c) δ V fillM · n |∂K and {v · n |∂K : v ∈ V , ∇· v = 0} are linearly
independent,
(d) dim δ V fillM = dim δ V fillM · n |∂K = IM (Vg × Wg )
Then, (Vg ⊕ δ V fillM ) × Wg admits an M-decomposition.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 107 / 120
A construction of M-decompositions
A three-step procedure to construct the filling space δ VfillM

(1) Characterize the trace space {v · n |∂K : v ∈ V , ∇· v = 0}


(2) Find a trace space CM ⊂ M(∂K ) such that

CM ⊕ {v · n |∂K : v ∈ V , ∇· v = 0} = {µ ∈ M : hµ, 1i∂K = 0}

note that the dimension of the space CM is equal to IM (V × W )

(3) Set δ V fillM := {v µ : µ ∈ CM }, where v µ is divergence-free function


such that v µ · n |∂K = µ

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 108 / 120
A construction of M-decompositions
The M-indexes for different elements

V × W × M := Pk (K ) × Pk (K ) × Pk (∂K )
2D element I M (V × W ) 3D element I M (V × W )
triangle 0 tetrahedron 0
(k≥0) (k≥0)
quadrilateral 1 2 pyramid 1 3
(k=0) (k≥1) (k=0) (k≥1)
pentagon 2 4 5 prism1 1 3
(k=0) (k=1) (k≥2) (k=0) (k≥1)

hexagon 3 6 8 9 hexahedron2 2 6 9
(k=0) (k=1) (k=2) (k≥3) (k=0) (k=1) (k≥2)

1
no parallel faces
Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 109 / 120
A construction of M-decompositions
An example of δ VfillM on a quadrilateral

V × W × M := Pk (K ) × Pk (K ) × Pk (∂K ),
δ V fillM := span{∇×(ξ4 λk4 ), ∇×(ξ4 λk3 )}.

λi is a linear function that vanishes on edge ei .


ξ4 ∈ H 1 (K ) is a function such that its trace on each edge is linear
and vanishes at the vertices v1 , v2 , and v3 .

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 110 / 120
A construction of M-decompositions
An example of δ VfillM on the reference pyramid

K := {(x, y , z) : 0 < x, 0 < y , 0 < z, x + z < 1, y + z < 1}

V × W × M := Pk (K ) × Pk (K ) × Pk (∂K )
 xy
span{∇×( 1−z ∇z)} if k = 0
δ V fillM :=
 k+1 k+1
span{∇×( x1−z
y
∇z), ∇×( y1−z
x xy
∇z), ∇×( 1−z ∇x)} if k ≥ 1

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 111 / 120
A construction of M-decompositions.
From M-decompositions to hybridized mixed methods

Theorem
Let the space V × W admit an M-decomposition and assume that
∇ · Vg ( W . Then,
V × ∇ · V admits an M-decomposition.
Moreover, let δ V fillW satisfy the following hypotheses:
(a) δ V fillW · n |∂K ⊂ M,
(b) ∇ · δ V fillW ⊕ ∇ · V = Wg ,
(c) dim δ V fillW = dim ∇ · δ V fillW ,
Then (V ⊕ δ V fillW ) × W admits an M-decomposition.

For the above choices of spaces, we can set stabilization operator τ = 0 in


and obtain hybridized mixed methods.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 112 / 120
A construction of M-decompositions
Spaces for hybridized mixed methods on a quadrilateral

V hdg × W hdg × M := Pk (K ) ⊕ δVfillM × Pk (K ) × Pk (∂K ),


δ V fillM := span{∇×(ξ4 λk4 ), ∇×(ξ4 λk3 )}.
δ V fillW := x PkK .

V W M τ
UMX V hdg ⊕ δVfillW Pk (K ) Pk (∂K ) 0
HDG V hdg Pk (K ) Pk (∂K ) >0
LMX V hdg Pk−1 (K ) Pk (∂K ) 0

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 113 / 120
A construction of M-decompositions
Spaces for hybridized mixed method on a pyramid

V hdg × W hdg × M := Pk (K ) ⊕ δVfillM × Pk (K ) × Pk (∂K ), k ≥ 1


x y k+1 y x k+1
δ V fillM := span{∇×(
xy
∇z), ∇×( ∇z), ∇×( ∇x)}.
1−z 1−z 1−z
δ V fillW := x PlkK .

V W M τ
UMX V hdg
⊕ δ V fillW Pk (K ) Pk (∂K ) 0
HDG V hdg Pk (K ) Pk (∂K ) >0
LMX V hdg Pk−1 (K ) Pk (∂K ) 0

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 114 / 120
The theory of M-decompositions.
Numerical experiments.

History of convergence of LDG-H with k = 1

h ku − uh⋆ kΩh rate ku − uh⋆ kΩh rate ku − uh⋆ kΩh rate


τ =1
0.1 0.15E-2 - 0.83E-2 - 0.52E-2 -
0.05 0.18E-3 3.06 0.16E-2 2.36 0.10E-2 2.34
0.025 0.23E-4 3.03 0.28E-3 2.52 0.19E-3 2.43
0.0125 0.28E-5 3.02 0.44E-4 2.68 0.35E-4 2.46

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 115 / 120
The theory of M-decompositions.
Numerical experiments.

History of convergence of M-decompositions with k = 1

h ku − uh⋆ kΩh rate ku − uh⋆ kΩh rate ku − uh⋆ kΩh rate


τ =1
0.1 0.15E-2 - 0.26E-2 - 0.17E-2 -
0.05 0.18E-3 3.06 0.31E-3 3.06 0.21E-3 3.02
0.025 0.23E-4 3.03 0.38E-4 3.03 0.27E-4 2.95
0.0125 0.28E-5 3.02 0.47E-5 3.02 0.35E-5 2.96

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 115 / 120
The theory of M-decompositions

Provides:

1 A systematic way of constructing superconvergent HDG and


hybridized mixed methods for elements of arbitrary shapes.
2 A systematic approach to satisfying elementwise inf-sup conditions,
stabilized (HDG) or not (mixed methods).
3 A systematic way of constructing finite element commuting diagrams.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 116 / 120
The evolution of HDG methods.
Steady-state diffusion

Relation with old DG methods. (C. Gopalakrishnan, Lazarov, 09; C.,


Guzman, Wang, 09).
Relation with mixed methods:
The SFH method + relation with SDG method (C., Dong, Guzman,
09; SDG Chung, C., Fu, 12).
Necessary conditions for superconvergence (C., Qiu, Shi, 12, 13, 14).
Theory of M-decompositions + new mixed methods (C., Fu, Qiu,
Sayas, 16, 17).
New stabilization functions (Lehrenfeld, Schöberl, 10; Oikawa, 14;
HHO Di Pietro, Ern, Lemaire, 14).
Different formulations of the same method (C. 16).
Different characterizations leading to the same scheme (C., 16).
Applications to a wide variety of PDEs.

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 117 / 120
Ongoing work and open problems

A posteriori error estimates: Only in terms of uh − ubh and τ ?


Efficient solvers: Domain decomposition methods?
Stokes flow: Superconvergence with other formulations?
Solid mechanics: Optimal convergence for all variables?
Are there HDG methods which conserve energy?
Linear transport: Which unknowns superconverge?
HDG methods for KdV equations: Superconvergence?
Nonlinear hyperbolic conservation laws: New ways to deal with
shocks?

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 118 / 120
References

Static condensation, hybridization and the devising of the HDG


methods. (48p.)
The Discontinuous Galerkin methods for fluid dynamics. (111 pp.)
HDG methods for hyperbolic problems. (20 pp., with N.C.Nguyen
and J. Peraire.)

Bernardo Cockburn (U. of Minnesota, USA) Discontinuous Galerkin Methods Barcelona, 2014 119 / 120
Discontinuous Galerkin Methods
for Computational Fluid Dynamics

B. Cockburn

University of Minnesota, MN, USA

ABSTRACT

The discontinuous Galerkin methods are locally conservative, high-order accurate, robust methods
which can easily handle elements of arbitrary shapes, irregular meshes with hanging nodes, and
polynomial approximations of different degrees in different elements. These properties, which render
them ideal for hp-adaptivity in domains of complex geometry, have brought them to the main stream of
computational fluid dynamics. We study the properties of the DG methods as applied to a wide variety
of problems arising in fluid dynamics with a special emphasis on linear, symmetric positive hyperbolic
systems, the Euler equations of gas dynamics, purely elliptic problems, and the incompressible and
compressible Navier-Stokes equations. In each instance, we discuss the main properties of the methods,
display the mechanisms that make them work so well, and present numerical experiments showing
their performance.

key words: computational fluid dynamics, discontinuous finite elements

1. Introduction

This is a short, introductory essay to the study of the so-called Discontinuous Galerkin (DG)
methods for fluid dynamics. The DG methods provide discontinuous approximations defined
by using a Galerkin method element by element, the connection between the values of the
approximation in different elements being established by the so-called numerical traces. Since
the methods use discontinuous approximations, they can easily handle elements of arbitrary
shapes, irregular meshes with hanging nodes, and polynomial approximations of different
degrees in different elements. The methods are thus ideally suited for hp-adaptivity in domains
of complex geometry. Moreover, since they use a Galerkin method on each element, they can
easily achieve high-order accuracy when the exact solution is smooth and high resolution when

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

2 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

it is not. Finally, when their numerical traces are properly chosen, they achieve a high degree of
locality (and hence a high degree of parallelizability for time-dependent hyperbolic problems),
they become locally conservative (a highly valued property in computational fluid dynamics),
easy to solve and, last but not least, very stable even in the presence of discontinuities or
strong gradients.

The first DG method was introduced by Reed and Hill (1973 for numerically solving the
neutron transport equation, a linear hyperbolic equation for a scalar-valued unknown. Lesaint
and Raviart (1974 recognized the relevance of the method and carried out its first theoretical
analysis. Since then, the method has been slowly evolving as it was applied to different
problems. In the 1990s, the method was successfully extended to nonlinear time-dependent
hyperbolic systems by Cockburn and Shu (see the review by Cockburn and Shu (2001) and
since then the method has known a remarkably vigorous development, as suggested in Figure
1, where we display the number of papers (in the American Mathematical Society database
MathSciNet) whose title contains the words discontinuous Galerkin. In this paper, we describe
the method, discuss its main properties, uncover the mechanisms that make it work so well,
and show its performance in a variety of problems in fluid dynamics.

Figure 1. Cumulative number of papers, in MathSciNet, containing the words discontinuous Galerkin
in their title.

The paper is organized as follows. Since the main properties of all DG methods are already
displayed in the very first DG method, we begin, in Section 2, by considering the original
DG method for the neutron transport equation. In Section 3, we extend the DG method
to linear, symmetric positive hyperbolic systems, and, In Section 4, to nonlinear hyperbolic
conservation laws. This key extension, which takes advantage of the relation between the so-
called finite-volume monotone schemes and the DG methods, is what made these methods
relevant in computational fluid dynamics. In Section 5, we consider DG methods for steady-
state diffusion problems. We argue that the DG methods can be considered to lie in between
the well known continuous Galerkin and the mixed methods for second-order elliptic problems.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 3

In Section 6, we extend theses methods to the Stokes equations of incompressible flow. In


Section 7, we consider convection-dominated flows including convection-diffusion, the Oseen
and the compressible Navier-Stokes equations. In these last three sections, special attention
will be devoted to the so-called hybridizable DG (HDG) methods, which constitute a subclass
of DG methods amenable to static condensation. Finally, we end in Section 8 with some
concluding remarks and bibliographical notes.

2. The neutron transport equation

In this section, we consider the original DG method for numerically solving the neutron
transport equation. It is in this framework that the idea of combining a Galerkin method on
each element together with a suitably defined numerical trace linking the different elements
was introduced. We discuss the properties of local conservativity and local solvability of the
method and show that the jumps of the approximate solution across interelement boundaries
enhance the stability of the method. We also show that these jumps are related to the local
residuals in a linear fashion. This establishes that the original DG method is what we nowadays
call a residual-stabilized method.

2.1. The original DG method

We begin by considering the original DG method of Reed and Hill (1973 which was devised
to numerically solve the neutron transport equation,
σ u + ∇ · (a u) = f in Ω
u = uD on ∂Ω−
where σ is a positive number, a a constant vector and ∂Ω− the inflow boundary of Ω, that is,
∂Ω− = {x ∈ ∂Ω : a · n(x) < 0}
Here n(x) is the outward unit normal at x.

2.2. Definition

To define the method, we proceed as follows. First, we find the weak formulation that the
Galerkin procedure will be based upon. For each element K of the mesh Th of the domain Ω,
we multiply the neutron transport equation by a test function v and integrate over K to get
σ (u, v)K − (u, a · ∇v)K + ha · nK u, vi∂K = (f, v)K (1)
where nK is the outward unit normal to K,
Z Z
(u, v)K = u vdx, and hw, vi∂K = w vds
K ∂K

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

4 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

This is the weak formulation with which we define the approximation to u, uh . Thus, for each
element K ∈ Th , we take uh |K in the space of polynomials of degree k, Pk (K), and define it
by requiring that
σ (uh , v)K − (uh , a · ∇v)K + ha · nK u
bh , vi∂K = (f, v)K (2)
for all v ∈ Pk (K). Here, the numerical trace u bh is given by

uD (x) for x ∈ ∂Ω−
u
bh (x) = (3)
limǫ↓0 uh (x − ǫa) otherwise.
Note that if the vector a is perpendicular to the normal nK , the above numerical trace is not
well defined. However, the numerical trace of the flux a · nK u
bh , usually called the numerical
flux , is actually well defined and is called the upwinding numerical flux. This completes the
definition of the DG method.

2.3. The numerical trace of the flux

Now, let us discuss some of the properties of this method. Let us begin by noting that the
numerical trace of the flux, a · nK u
bh , is a linear function of the trace of uh which is consistent
and single valued. The numerical trace is easy of evaluate and its form ensures a great degree
of locality of the method. The fact that it is consistent, namely, that
a · nK u
b = a · nK u,
where u is the exact solution, ensures that we are approximating the correct exact solution.
The fact that it is single valued implies that the DG method is a locally conservative method.
Indeed, since on any face F := ∂K1 ∩ ∂K2 , we have
bh + a · nK1 u
a · nK2 u bh = 0,
for any set S which is the union of elements K ∈ Th , we have
Z Z Z
σ uh dx + a · nS u
bh ds = f dx
S ∂S S

This equation is obtained by simply taking v := 1 in the weak formulation (2) for each K in
S and then adding the equations.

2.4. A local energy identity and the elementwise solvability of the method

Next, note that, thanks for the definition of the numerical trace of the flux, the method satisfies
a local energy identity we deduce next. Setting v := uh in the weak formulation (2), we get
that
σ kuh k2K + 12 h1, a · nK u
bh2 i∂K + ΘK (uh ) = (f, uh )K
1/2
where kζkK := (ζ, ζ)K and
uh − uh )2 i∂K = 21 h1, |a · nK | (b
ΘK (uh ) := − 12 h1, a · nK (b uh − uh )2 i∂K−

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 5

by the definition of the numerical trace of the flux. Completing squares, we get that
2
1
2 σ kuh kK + 12 σkuh − f /σk2K + 21 h1, a · nK u
bh2 i∂K + ΘK (uh ) = 1 2
2σ kf kK ,

and finally,
2
1
2 σ kuh kK + 21 σkuh − f /σk2K + 12 h1, |a · nK | u
bh2 i∂K+ + ΘK (uh ) = 1 2
2σ kf kK bh2 i∂K− ,
+ 12 h1, |a · nK | u

where ∂K+ = ∂K\∂K− .

We claim that an immediate consequence of this local energy identity is the fact that the
approximate solution can be efficiently computed in an element-by-element fashion. Indeed,
from the weak formulation (2) and the definition of the numerical trace (3), we have

bh , vi∂K− ,
(σ uh , v)K − (uh , a · ∇v)K + ha · nK uh , vi∂K+ = (f, v)K − ha · nK u

for all v ∈ Pk (K), since, on ∂K+ , a · nK u


bh = a · nK uh |K . Since the above formulation defines
a square system, the existence and uniqueness of the approximation uh on K holds if and only
if, when we set the data f |K and ubh |∂K− to zero, the only solution is uh = 0. But, in this case,
the above energy identity reads

σkuh k2K + 21 h1, |a · nK | uh2 i∂K = 0,

which immediately implies that uh = 0.This proves the claim.

In Figure 2.4, the approximate solution uh on the elements of number i can only be computed
after the approximate solution on the neighboring elements of number j < i were obtained. The
approximate solution uh on the elements with equal number can be computed simultaneously.

2 3 7 12
1

4 11
6
2
3
10
1
5
a 4
2 9

1 5
6 8
2
7
10
3 5
4

Ω 7 8
5 6

Figure 2. Solving the neutron transport equation with the DG method.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

6 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

2.5. A global energy identity and stabilization by the jumps

Another consequence of the above local energy identity is a global energy identity obtained by
simply adding the identities for all the elements K ∈ Th , that is,
2
1
2 σ kuh kΩ + 1
2σ kσ uh − f k2Ω + 21 h1, |a · n| uh2 i∂Ω+ + Θh (uh ) = 1 2
2σ kf kΩ
2
+ 12 h1, |a · n| uD i∂Ω− ,
where
1 X 1 X
Θh (uh ) := ΘK (uh ) = uh − uh )2 i∂K− .
h1, | a · nK | (b
2 2
K∈Th K∈Th
P
Of course, kζk2Ω := 2
K∈Th kζkK .

Note that the left-hand side of the above identity can be considered to be a discrete energy.
As a consequence, the term Θh (uh ) can be interpreted as the energy associated to the inter-
element jumps. This is why we say that the method is stabilized by the jumps. Moreover, we
see that the energy of the inter-element jumps is uniformly bounded regardless of the mesh
Th and of the approximating spaces Pk (K), K ∈ Th . Thus, the method exerts an automatic
control on the size of the jumps.

2.6. Control of the residuals by the inter-element jumps

Next, we argue that the size of the jumps is related to the ability of the method to solve the
partial differential equation inside the element. To see this, let us first note that the size of
the local residuals,
RK := σ uh + ∇ · ( a uh ) − f and r∂K := a · nK (b
uh − uh ),
control the quality of the approximation since only when they are zero, we have that uh
coincides with the exact solution u. Now, let us show that these two residuals are related by
the very definition of the method. Indeed, a simple integration by parts in the weak formulation
(2), reveals that
(RK , v)K = hr∂K , vi∂K ,
for all v ∈ Pk (K). Taking v := Pk RK , where Pk denotes the L2 (K)−projection into Pk (K),
we immediately obtain that
−1/2
kPk RK kK ≤ C hK kr∂K k∂K
This implies that only the size of the jumps a·nK (buh −uh ), and the size of f −Pk f , control the
quality of the approximation. Thus, if the size of the jumps is very small, then the projection
of the residual Pk RK is also very small. If f |K is very smooth, then the residual RK is also
very small. As a consequence, the quality of the approximation on K is very good.

On the other hand, if the quality of the approximation in the element K is very poor, as we
expect it would be, for example, in the presence of discontinuities, then the projection of the
residual Pk RK , and hence the jumps a · nK (b
uh − uh ) across the inflow boundary of K, would

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 7

be huge. Fortunately, this faux pas of the method is automatically compensated by an increase
in the dissipative term ΘK (uh ), see the above energy identities, which in practice results in
the damping of the typical spurious oscillations that appear around the discontinuities.

2.7. Convergence properties

For very smooth solutions and general meshes, we have the following simple a priori error
estimate.
Theorem 1. (Error estimates for the original DG method) Consider the original DG method
for the neutron transport equation given by (2) and (3). Then, for regular meshes Th made of
arbitrarily shaped elements, we have that
1/2
σ 1/2 k u − uh kΩ + h1, |a · n|(u − uh )2 i∂Ω+ + Θh (uh ) ≤ C| u |H k+1 (Th ) hk+1/2
where C is independent of u and h is the maximum of the diameters hK of the elements
K ∈ Th .

This elementary result was improved by Johnson and Pitkäranta (1986 who proved, among
other things, the same order of convergence of k + 1/2 for k u − uh kΩ independently of σ. This
order of convergence was shown to be sharp by Peterson (1991 and then by Richter (2008.
However, in some instances, the order of convergence of k + 1 can be obtained. This happens
for Cartesian meshes and tensor- product polynomials of degree k (see Lesaint and Raviart
(1974), for some structured meshes of triangles and spaces of polynomials of degree k (see
Richter (1988), and for some unstructured meshes which are, roughly speaking, aligned with
the transport velocity a (see Cockburn et al. (2008a; Cockburn et al. (2010a.

An early a posteriori error estimate and superconvergence results can be found in Adjerid and
Massey (2002. For finer results, see the work by Adjerid and Mechai (2014 and the references
therein.

For early work on adaptivity on linear, steady state hyperbolic problems, see the paper by
Bey (1994, Bey and Oden (1996, and then to the papers Houston et al. (2000; Houston et al.
(2001; Houston et al. (2002, Houston and Süli (2001, and Süli and Houston (2002.

2.8. Devising DG methods for general partial differential equations

Let us end by pointing out that, by rewriting a partial differential equation as first-order
system, the same approach used to define the original DG method can be readily applied.
As we are going to see, all the resulting DG methods share with the original DG method
several important, distinctive properties. First is the use of approximations of the variables
inside the elements as well as on their boundary. Since no interelement continuity is required
for the approximations, the DG methods can handle arbitrarily-shaped elements and general
basis functions. Because of this, the DG methods are ideally suited for hp-adaptivity and for

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

8 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

capturing special features of the exact solution by using special basis functions. Second is the
enforcement of the equations by means of an elementwise Galerkin method. This results in
locally conservative numerical methods and automatically implies the control of the residuals
inside the elements by the residuals at the boundaries (which depend on the jumps of the
variables). Finally, by suitably defining their numerical traces, a stabilization mechanism is
introduced which depends on the interelement jumps of the variables. It renders the methods
robust and can even improve their convergence properties.

3. Linear, symmetric positive hyperbolic systems

In this section, we exploit the hyperbolic nature of the neutron transport problem to extend the
original DG method to linear, symmetric positive hyperbolic systems like the wave equation
or the Maxwell equations. To show how to do that, we consider the model problem
N
X
ut + Ai uxi + Bu = f in Ω × (0, T )
i=1
(An − M ) (u − uD ) = 0 on ∂Ω × [0, T ]
u(t = 0) = u0 on Ω

where u is an Rm -valued
PN function and Ai is a symmetric matrix for i = 1, . . . , N . Here An
i
denotes the matrix i=1 ni A where n = (n1 , . . . , nN ) is the unit outward normal at the
boundary of Ω. Friedrichs (1958 has shown that this problem has a unique solution under
some smoothness conditions on the data, under the positivity property
N
X
B + B∗ − Aixi ≥ βI, β>0
i=1

and under the following properties on the boundary condition matrix M :

M + M∗ ≥ 0
ker(An − M ) + ker(An + M ) = Rm on ∂Ω × [0, T ].

We can easily verify that the neutron transport equation is a particular case of the above
problem. Indeed, in that case, we have that m = 1 and so the matrices Ai are real numbers;
moreover, we have that a = (A1 , . . . , AN ) and that An = a · n. The boundary condition matrix
M is simply | a · n |, so that the boundary condition reads

(a · n − | a · n |)(u − uD ) = 0 on ∂Ω × [0, T ]

that is,
u = uD on ∂Ω− × [0, T ]

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 9

3.1. Definition

To define a DG method, we proceed as follows. First, we obtain a mesh Th of the space-time


domain Ω×(0, T ). Then, for each element K ∈ Th , we take uh |K to be in the finite dimensional
space V (K) and define it by requiring that
N
X N
X
−(uh , vt )K − (Ai uh , vxi )K + hA\
nK uh , vi∂K + ((B − Aixi )uh , v)K = (f, v)K
i=1 i=1

for all v ∈ V (K). Here we have taken AnK = An + nt Id, where nK = (n, nt ). Next, let us
\
define the numerical flux AnK u.

On the boundary of the space-time domain, we take




 u0 on Ω × {t = 0}

u h on Ω × {t = T }
A\nK u h = 1

 A n (u h + u D )
2
+ 12 M (uh − uD ) on ∂Ω × (0, T )

and on the interelement boundaries,

A\
nK uh = AnK {uh } +
1
2 M [[uh ]]nK

where {uh } = 12 (u− + − + ±


h + uh ), [[uh ]]nK = uh − uh and uh (x) = limǫ↓0 uh (x ± ǫ nK ).

It only remains to define the matrices M. The two main choices are M = | AnK |, which gives
rise to the so-called upwinding numerical flux, and M = ̺(AnK ) Id, where ̺(E) is the spectral
radius of the matrix E, which gives rise to the so-called Lax-Friedrichs numerical flux. This
completes the definition of the DG method.

3.2. The numerical trace of the flux

Note how the numerical trace of the flux, A\nK uh , is a linear function of the traces of uh which
is consistent and single valued. As in the case of neutron transport, the form the numerical
race of the flux is easy to evaluate and ensures a high degree of locality of the method. The
property of consistency ensures that we are approximating the correct exact solution. It is
satisfied if
\
A nK u = AnK u

where u is the exact solution. Since [[u]]nK = 0, we do have that the numerical flux is consistent
in the interelement boundaries. It is trivial to see it is consistent on Ω × {t = 0} and on
Ω × {t = T }. It remains to see what happens on ∂Ω × (0, T ). But there we have

\
AnK u = 1
An (u + uD ) + 12 M (u − uD )
2
= An u + 12 (M − An )(u − uD ) = An u

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

10 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

since the exact solution satisfies the boundary condition (M −An )(u−uD ) = 0. The numerical
trace of the flux is thus consistent.

Finally, the fact that it is single valued, that is , that on the face F := ∂K1 ∩ ∂K2 ,

A\ \
nK1 uh + AnK2 uh = 0

ensures the highly valued property of local conservatively which states that for any set S,
which is the union of elements K ∈ Th , we have

Z N
X Z Z
(B − Aixi )uh dx + A\
nS uh ds = f dx.
S i=1 ∂S S

t n +1

tn

t n −1

(a)

t n +1

tn
11 12
10 6
10 9 5 2
9 10
8 7 8 4
5 3
5 6 1
4 3 4 2
1 2 1
1
t n −1

(b)

Figure 3. Space-time meshes for the DG method. On each time slab Ω × (tn , tn+1 ), the resolution can
be globally implicit (a) or implicit only on each element (b).

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 11

3.3. A local energy identity and the elementwise solvability of the method

To obtain the local energy identity, we consider the case in which M = M = |AnK | in order to
emphasize the closeness of this case with that of the neutron transport problem. In this case,
the resulting numerical flux is called the unwinding flux:

A\
nK uh = AnK {uh } +
1
2 |AnK | [[uh ]]nK = A+ − − +
nK uh + AnK uh ,

where A− +
nK := (AnK − |AnK |)/2 and AnK := (AnK + |AnK |)/2. We can now proceed exactly as
in the neutron transport problem to obtain the following local energy identity:

2 2 − − + +
1 1
2 kuh kK,Σ + 2 kΣuh − fkK,Σ−1 + 21 hA+ 1 2 1 −
nK uh , uh i∂K + ΘK (uh ) = 2 kf kK,Σ−1 − 2 hAnK uh , uh i∂K ,

where

ΘK (uh ) := − 21 hA−
nK [[uh ]], [[uh ]]i∂K .

PN
Here, we are using the notation kzk2K,C := (Cz, z)K and Σ := 21 (B + B ∗ − i=1 Aixi ).

This identity implies that the approximate solution on each element is well defined provided
appropriate data is provided. Indeed, on the element K ∈ Th , we have that uh is the element
of V (K) that solves

N
X N
X
− +
−(uh , vt )K − (Ai uh , vxi )K +hA+
nK uh , vi∂K +((B− Aixi )uh , v)K = (f, v)K −hA−
nK uh , vi∂K
i=1 i=1

for all v ∈ V (K). Since this defines a square system, to prove the existence and uniqueness of
+
uh , it is enough to set the data A−nK uh and f to zero and show that the only possible solution
is the trivial one. But in this case, the above local energy identity gives

kuh k2K,Σ + 12 h|AnK |u− −


h , uh i∂K = 0,

which immediately implies that uh = 0 in K.

This means that approximate solution uh |K can be obtained on the element K once the trace
u+ −
h |∂K such that AnK = 0 is available. In spite of the possibility of solving in an elementwise
manner, sometimes it is possible to solve only on the time slab Ω × [tn , tn+1 ], as indicated
by Johnson et al. (1984. This gives rise to a globally implicit method; see Figure 3(a). This
difficulty can be avoided if suitably defined meshes are used together with the upwinding
numerical flux as shown by Lowrie et al. (1995, Lowrie (1996, and Lowrie et al. (1998 in
the frame of nonlinear hyperbolic systems, and later by Yin et al. (2000 in the framework of
elastodynamics, and by Falk and Richter (1999 in the framework of linear symmetric positive
hyperbolic systems. See Figure 3(b).

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

12 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

3.4. A global energy identity and stabilization by the jumps

As in the case of the neutron transport problem, we can easily get the following global energy
identity:
2 − −
1
2 kuh kΩT ,Σ + 21 kΣuh − fk2ΩT ,Σ−1 + 21 hA+
n uh , uh i∂ΩT + Θh (uh )

= 21 kf k2ΩT ,Σ−1 − 12 hA− 1


n uD , uD i∂Ω×[0,T ] + 2 hu0 , u0 iΩ ,

where ΩT := Ω × (0, T ) and


1 X 1 1 X
Θh (uh ) := ΘK (uh ) = − h A−
n [[uh ]], [[uh ]]i∂ΩT + h| AnF | [[uh ]], [[uh ]]iF .
2 2 2
K∈Th i F ∈F

Here, nF is any unit vector normal to the face F and Fhi


is the collection of all interior
faces of the mesh Th . Again, we see that the method is stabilized by the jumps and that the
energy of the inter-element jumps is uniformly bounded regardless of the mesh Th and of the
approximating spaces V (K), K ∈ Th .

3.5. Control of the residuals by the inter-element jumps

Now, consider the residuals


N
X
RK := (uh )t + Ai (uh )xi + Buh − f and r∂K := A\ −
nK uh − AnK uh = −AnK [[uh ]].
i=1

A simple integration by parts in the weak formulation defining the scheme gives
(RK , v)K = hr∂K , vi∂K ,
for all v ∈ V (K), and this implies that
−1/2
kPV RK kK ≤ C hK kr∂K k∂K
where PV is the L2 (K)-projection into the space V (K). If we assume, for simplicity, that the
matrices Ai , i = 1, · · · , N and B are constant, this implies that only the size of the jumps
−A−nK [[uh ]], and the size of f − PV f control the quality of the approximation.

3.6. Convergence properties

For very smooth solutions and general meshes, we have the following simple a priori error
estimate.
Theorem 2. (Error estimates for the DG method) Consider the DG method for which each
component of the space V (K) contains the space of polynomials Pk (K). Then, for regular
meshes Th made of arbitrarily shaped elements, we have that
1/2 1/2
k u − uh kΩT ,Σ + hA+
n (u − uh ), (u − uh )i∂Ω + Θh (uh ) ≤ C| u |H k+1 (Th ) h
k+1/2

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 13

where C is independent of u and h is the maximum of the diameters hK of the elements


K ∈ Th .

See the a priori error estimates in Falk and Richter (1999. Asymptotically exact
discontinuous Galerkin error estimates for linear symmetric hyperbolic systems have been
obtained by Adjerid and Weinhart (2014; see also Adjerid and Weinhart (2011.

3.7. Enhanced accuracy by local postprocessing

The method of lines for these linear systems has been studied by Cockburn et al. (2003 where it
was shown that, if uniform meshes are used, a local postprocessing of the approximate solution
of the DG method is of order 2 k + 1 when polynomials of degree k are used.

This remarkable technique, which can be used for all other DG methods, was introduced
by Bramble and Schatz (1977 for finite element solutions to elliptic equations. It was explored
from a Fourier perspective, and for derivative filtering, by Thomée (1977. Its application to
discontinuous Galerkin solutions to linear hyperbolic equations was then done by Cockburn
et al. (2003. Applications to the approximation of derivatives were carried by Ryan and
Cockburn (2009 and more recently by Li et al. (2016. For applications to convection-diffusion
equations see Ji et al. (2012, to variable-coefficient equations see Mirzaee et al. (2011, and
to nonlinear hyperbolic equations with smooth solutions see Ji et al. (2013. This filtering
technique, later called Smoothness-Increasing Accuracy-Conserving (SIAC) filtering, requires
translation-invariant meshes, used originally a symmetric convolution kernel, and cannot be
applied up to the boundary. Extensions to smoothly varying mesh sizes-uniform cartesian
meshes were carried out by Curtis et al. (0708, and to unstructured triangular meshes by
Mirzaee et al. (2013, (applications to structured triangular meshes were done by Mirzaee et al.
(2011 and to structured tetrahedral meshes by Mirzaee et al. (2014. The development of one-
sided convolution kernels which can be applied up to the boundary, was carried out in the
series of papers by Ryan and Shu (2003; van Slingerland et al. (2011; Ryan et al. (2015; see
also the L∞ −error estimates by Ji et al. (2014. The application of SIAC filtering to streamline
visualization and isosurface extraction was done by Walfisch et al. (2009.

Let us illustrate how powerful is this filtering on the model problem,

ut + ux = 0, in (0, 2π) × (0, T )


u(x, 0) = sin(x) x ∈ (0, 2π) (4)

with periodic boundary conditions. In Table 1, we see that the order of convergence of both
the L2 and L∞ errors for P k elements is of (k + 1) before postprocessing and of at least
(2k + 1) after postprocessing, for k = 1, 2, 3, 4. In Figure 4, we see the absolute errors before
and after postprocessing for P 2 . The postprocessing of the approximate solution is obtained
by convolution with a kernel whose support is the union of a number of elements, which only
depends on k; for details, see Cockburn et al. (2003.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

14 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

P 2, before postprocessing
10−2
N = 10

10−4 N = 20
N = 40

10−6 N = 80
|Error|

N = 160

10−8

10−10

10−12
1 2 3 4 5 6
(a) x

P 2, after postprocessing
10−2

10−4
N = 10

10−6
N = 20
|Error|

10−8 N = 40

N = 80
10−10
N = 160

10−12
1 2 3 4 5 6
(b) x

Figure 4. The absolute value of the errors for P 2 with N = 10, 20, 40, 40, 80 and 160 elements.
Before postprocessing (a) and after postprocessing (b). (From Cockburn B, Luskin M, Shu C-W and
Süli E. Enhanced accuracy by post-processing for finite element methods for hyperbolic equations.
Math. Comput. 2003; 72:577-606.)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 15

4. Nonlinear Hyperbolic Problems

In this section, we consider DG methods for nonlinear hyperbolic problems. We begin by


showing that to ensure convergence towards the physically relevant solution, usually called
the entropy solution, the DG methods need to use a numerical flux based on a suitable
approximate Riemann solver and that they must use either a shock-capturing term or a slope
limiter . We show that the shock-capturing DG methods are strongly related to stabilized
methods like the streamline diffusion method, and that slope-limiter DG methods can be
considered to be an extension of finite volume methods. We then show computational results
for some shock-capturing DG methods and describe and analyze the so-called Runge-Kutta
DG (RKDG) method, a slope-limiter DG method. We show how the different ingredients of the
method, namely, the DG space discretization, a special Runge-Kutta time discretization, and a
generalized slope limiter, are put together to ensure its stability. Several numerical experiments
showing the performance of the RKDG methods are given. Particular attention is devoted to
the Euler equations of gas dynamics. Finally, we discuss, and illustrate, a remarkable technique
for enforcing a range-invariance property of the RKDG approximations.

4.1. The main difficulty: the loss of well-posedness

Devising numerical methods for nonlinear hyperbolic problems is dramatically different from
devising methods for linear symmetric hyperbolic problems. This is due to the fact that whereas
linear, symmetric hyperbolic problems are well posed, nonlinear hyperbolic problems are not.

This difficulty was uncovered first in the framework of the Euler equations of gas dynamics;
indeed, this equation has several nonphysical weak solutions. This happens because the
Euler equations of gas dynamics are obtained from the compressible Navier-Stokes by simply
dropping from the equations, the terms modeling viscosity and heat transfer effects. As a
consequence, the information concerning the second law of thermodynamics is completely lost
and discontinuous solutions, which violate such a law suddenly appear. To devise numerical
schemes that are guaranteed to converge to the entropy solution and not to any other weak
solution constitutes the main difficulty of devising numerical methods for nonlinear hyperbolic
problems.

This difficulty is present even in the simplest hyperbolic problem, namely, the scalar
hyperbolic conservation law

ut + (f (u))x = 0 in (0, 1) × (0, T )


u(t = 0) = u0 on (0, 1)

with periodic boundary conditions. To illustrate this phenomenon, consider the well-known
Engquist-Osher and Lax-Wendroff schemes and let us apply them to the above equation for
f (u) = u2 /2 and u0 (x) = 1 on (0.4, 0.6) and u0 (x) = 0 otherwise. In the Figure 5, we see that
the approximation given by the Engquist-Osher scheme converges to the entropy solution,
whereas that given by the Lax-Wendroff scheme does not. The Lax-Wendroff scheme lacks a

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

16 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Table 1. The effect of post-processing the approximate solution. (From Cockburn B, Luskin M,
Shu C-W and Süli E. Enhanced accuracy by postprocessing for finite element methods for hyperbolic
equations. Math. Comput. 2003; 72:577-606.)

Mesh Before postprocessing After postprocessing


L2 error Order L∞ error Order L2 error Order L∞ error Order

P1
10 3.29E-02 - 5.81E-02 - 3.01E-02 - 4.22E-02 -
20 5.63E-03 2.55 1.06E-02 2.45 3.84E-03 2.97 5.44E-03 2.96
40 1.16E-03 2.28 2.89E-03 1.88 4.79E-04 3.00 6.78E-04 3.01
80 2.72E-04 2.09 8.08E-04 1.84 5.97E-05 3.00 8.45E-05 3.00
160 6.68E-05 2.03 2.13E-04 1.93 7.45E-06 3.00 1.05E-05 3.00
320 1.66E-05 2.01 5.45E-05 1.96 9.30E-07 3.00 1.32E-06 3.00
P2
10 8.63E-04 - 2.86E-03 - 2.52E-04 - 3.57E-04 -
20 1.07E-04 3.01 3.69E-04 2.95 5.96E-06 5.40 8.41E-06 5.41
40 1.34E-05 3.00 4.63E-05 3.00 1.53E-07 5.29 2.16E-07 5.28
80 1.67E-06 3.00 5.78E-06 3.00 4.22E-09 5.18 5.97E-09 5.18
160 2.09E-07 3.00 7.23E-07 3.00 1.27E-10 5.06 1.80E-10 5.06
P3
10 3.30E-05 - 9.59E-05 - 1.64E-05 - 2.31E-05 -
20 2.06E-06 4.00 6.07E-06 3.98 7.07E-08 7.85 1.00E-07 7.85
40 1.29E-07 4.00 3.80E-07 4.00 2.91E-10 7.92 4.15E-10 7.91
50 5.29E-08 4.00 1.56E-07 4.00 5.03E-11 7.87 7.24E-11 7.83
P4
10 1.02E-06 - 2.30E-06 - 1.98E-06 - 2.81E-06 -
20 3.21E-08 5.00 7.30E-08 4.98 2.20E-09 9.82 3.11E-09 9.82
30 4.23E-09 5.00 9.66E-09 4.99 4.34E-11 9.68 6.66E-11 9.48

mechanism that ensures its convergence towards the entropy solution and, as a consequence,
can converge to a weak solution, which is not the entropy solution.

4.2. Tools for capturing the entropy solution: heuristics

The DG methods try to ensure convergence towards the entropy solution by using the so-
called Riemann solvers and either a shock-capturing term or a slope limiter . To describe the
heuristics behind their construction, we consider the parabolic problem

ut + (f (u))x = ν uxx in (0, 1) × (0, T )


u(t = 0) = u0 on (0, 1)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 17

1
u
0.8

0.6

0.4

0.2 uh
0

−0.2

−0.4
x
−0.6
(a) 0 0.25 0.5 0.75 1

0.8

0.6
uh
0.4

0.2
u
0

−0.2

−0.4
x
−0.6
(b) 0 0.25 0.5 0.75 1

Figure 5. The entropy solution, u, and its approximation uh at time T = 1/2: Engquist-Osher
scheme (a) and Lax-Wendroff scheme (b). (From Cockburn B. Continuous dependence and error
estimation for viscosity methods. Acta Numer. 2003a; 12:127-180.)

since it is known that as the viscosity coefficient ν goes to zero, the solution of the above
problem converges to the entropy solution of our scalar hyperbolic conservation law.

Given this property, our strategy is to use the weak formulation of the parabolic problem to
see what the tools are that should be used to devise DG methods that converge to the entropy
solution. Thus, multiplying the parabolic equation by a test function ϕ, and integrating over
the space-time element K, we get
Z
(f (u) − ν ux , u) · (nx , nt ) ϕds
∂K
Z
− (f (u) − ν ux , u) · (ϕx , ϕt )dxdt = 0
K

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

18 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Now, if we set FnK (u) = (f (u) − ν ux , u) · (nx , nt ), on ∂K, we end up with


Z Z
FnK (u) ϕds − (f (u), u) · (ϕx , ϕt )
∂K K
Z
+ ν ux ϕx dxdt = 0
K
Finally, noting that we have
Z x
ν ux (x, t) = R(u)(y, t)dy
x(t)

where R(u) = ut + (f (u))x and x(t) is such that


ux (x(t), t) = 0 (such a point always exists because we
have periodic boundary conditions), this suggests the so-called shock-capturing DG methods:
Z Z
FbnK (uh ) ϕds − (f (uh ), uh ) · (ϕx , ϕt )
∂K K
Z
+ νb (uh )x ϕx dxdt = 0
K

where FbnK (uh ) is the approximate Riemann solver and the last term is the shock-capturing
term.

The approximate Riemann solver is nothing but a numerical trace for the function FnK (u);
it only depends on the two traces of the function u, that is, FbnK (uh ) = gb(u− +
h , uh ). The main
examples are the following:

(i) The Godunov flux:



mina≤u≤b g(u), if a ≤ b
gbG (a, b) =
maxb≤u≤a g(u), otherwise
(ii) The Engquist-Osher flux:
Z b
gbEO (a, b) = min(g ′ (s), 0)ds
0
Z a
+ max(g ′ (s), 0)ds + g(0)
0
(iii) The Lax-Friedrichs flux:
gbLF (a, b) = 1
2 [g(a) + g(b) − C(b − a)]
C = max |g ′ (s)|
inf u0 (x)≤s≤sup u0 (x)

The shock-capturing term has the same structure as the corresponding term for the parabolic
equation and typically, has a viscosity coefficient νb that depends on the residual as follows:
| R(uh ) |
νb = δ α
| (uh )x | + ǫ

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 19

where the auxiliary parameter δ is usually taken to be of the order of the diameter of K and
α is a parameter usually bigger than one and smaller than two. The purpose of the small
number ǫ is to prevent a division by zero when (uh )x = 0. The shock-capturing DG methods
considered by Jaffré et al. (1995 and by Cockburn and Gremaud (1996 for the scalar hyperbolic
conservation law in several space dimensions are of this form.

The DG methods that do not have a shock-capturing term must have a slope limiter in order
to ensure that the information about the entropy solution is incorporated into the scheme. In
fact, as we argue next, the slope limiters and the shock-capturing terms have exactly the same
origin. The DG methods with a slope limiter are obtained as follows. Instead of keeping the
shock-capturing term in a single equation, that term is split-off in a way typical of operator
splitting techniques. Take K = I × (tn , tn+1 ). To march from time tn to tn+1 , we first compute
n+1/2
uh from unh by using the scheme
Z
(fbh , u
bh ) · (nx , nt ) ϕds
∂K
Z
− (f (uh ), uh ) · (ϕx , ϕt )dxdt = 0
K
n+1/2
for some numerical flux (fbh , u
bh ) · (nx , nt ), and then, compute un+1 h from uh by using
Z
n+1/2
(un+1
h − uh ) ϕdx
I
Z
n+1/2
− (tn+1 − tn ) ν (uh )x ϕx dx = 0
I

We thus see that the function un+1


h captures the information contained in the shock-capturing
term. The link between this second step and the so-called slope limiters can be easily
established if we realize that, if we write,
n+1/2
un+1
h = ΛΠh uh
then, the operator ΛΠh is actually a (generalized) slope limiter. For details, see the work by
Cockburn (2001.

Let us illustrate this fact on a simple case. Consider the piecewise linear function vh and set
uh = ΛΠ(vh ), that is, uh is the piecewise linear function defined by
Z Z Z
uh ϕdx = vh ϕdx − slc (vh )x ϕx dx
I I I

where slc := (tn+1 − tn ) ν, for all linear functions ϕ. If we write


vh (x) = v j + (x − xj ) vx,j
on each interval Ij , and take
(
0 if vx,j = 0
slc = h2j h 
v j −v j−1 2 v j+1 −v j
i
12 1 − m 1, h2j vx,j , hj vx,j otherwise

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

20 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Figure 6. The ΛΠOh limiter: an example. Displayed are the local means of uh ( ), the linear function
uh in the element of the middle before limiting (- - - -) and the resulting function after limiting ( ).

where the minmod function m is defined by


(
s min1≤n≤3 | an | if s = sign(a1 )
m (a1 , a2 , a3 ) = = sign(a2 ) = sign(a3 )
0 otherwise
we obtain that
uj = v j , and ux,j = m(vx,j , v j − v j−1 , v j+1 − v j )
We thus see that the mean of uh coincides with that of vh . Moreover, since, by definition of
the function m, we have
| ux,j | ≤ | vx,j |
it is reasonable to call the operator ΛΠ a slope limiter. This slope limiter, which we denote
by ΛΠO h , is due to Osher (1984; see Figure 6. It is less restrictive than the limiters originally
considered by van Leer (1974 and by van Leer (1979.

Next, we show computational results for some shock-capturing DG methods and then study
the main example of DG methods using slope limiters, namely, the Runge-Kutta discontinuous
Galerkin (RKDG) methods.

4.3. Shock-capturing DG methods

There are only two theoretical results concerning shock-capturing methods, and those concern
the scalar hyperbolic conservation law. The first is by Jaffré et al. (1995, who proved
convergence to the entropy solution. The second is by Cockburn and Gremaud (1996, who

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 21

obtained a priori error estimates showing that this convergence takes place at rate of at least
h1/4 in the L∞ (0, T ; L1 (RN ))-norm, as well as a posteriori error estimates which can be used
for adaptivity purposes.

Space-time DG methods for nonlinear hyperbolic conservation laws were considered by


Lowrie et al. (1995 (Lowrie et al. (1995, Lowrie et al. (1998) and Lowrie (1996. DG shock-
capturing methods have been considered by Hartmann and Houston (2002a (Hartmann and
Houston (2002a,b) for adaptively solving for values of linear functionals of solutions of steady
state nonlinear hyperbolic conservation laws with remarkable success; see also Süli and Houston
(2002. More recently, shock-capturing DG methods with PDE-based artificial viscosity were
considered by Barter and Darmofal (2010. See also the shock-capturing DG method by Huerta
et al. (2012.

To give an example, let us consider the Burger’s equation



ut + 21 u2 x = 0, in Ω × (0, T )
where Ω = (0, 3) and T = 2, subject to the initial condition

 2 sin2 (πx), 0 ≤ x ≤ 1
u(x, 0) = sin2 (πx), 1≤x≤2

0, 2≤x≤3
and boundary condition u(0, t) = 0, for t ∈ [0, T ]; see Figure 7. The exact solution develops
two shocks, which eventually merge. The functional of interest J(·) is the value of the solution
before these two shocks collapse into each other. We thus take,
J(u) = u(2.3, 1.5) = 0.664442403975254670
see Figure 7.

In Figure 8, we compare the performance of the h- and hp-mesh refinement algorithms


for this problem. Again, we observe exponential convergence of the error in the computed
functional using hp-refinement; on the linear-log scale, the convergence line is straight. On
the final mesh the true error between J(u) and J(uDG ) using hp-refinement is almost five
orders of magnitude smaller than the corresponding quantity when h-refinement is employed
alone. Furthermore, in Figure 8, we observe that the hp-refinement algorithm also outperforms
the h-refinement strategy, when comparing the error in the computed target functional with
respect to the computational cost. Indeed, Figure 9 clearly shows that for the hp-DGFEM
the cost per degree of freedom when hp-refinement is employed is comparable to that of using
h-refinement.

Finally, in Figure 10, we show the primal mesh after 11 adaptive hp-mesh refinements.
Here, we see that the h-mesh has been refined in the region upstream of the point of interest,
thereby isolating the smooth region of u from the two interacting shock waves; this renders
the subsequent p-refinement in this region much more effective.

Now, let us consider the problem of computing the drag coefficient, J(u), of the NACA0012
airfoil for two flows. The first is subsonic and is obtained by imposing on the outer boundary

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

22 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

1.8

1.6

1.4

1.2

0.8

0.6

0.4

0.2

0
0 0.5 1 1.5 2 2.5 3
(a) xi

Point of
interest

(b)

Figure 7. Burgers’ problem initial condition (a) and isolines for the exact solution (b). (From Süli E
and Houston P. Adaptive finite element approximation of hyperbolic problems. In Error Estimation
and Adaptive Discretization Methods in Computational Fluid Dynamics, Volume 25 of Lecture Notes
in Computational Science and Engineering, Barth T and Deconink H (eds). Springer-Verlag: Berlin,
2002; 269-344.)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 23

100
hp-refinement
h-refinement

10−5
|J(u) − J(uDG)|

10−10

10−15
0 50 100 150 200 250 300 350 400
(a) sqrt (degrees of freedom)

100

hp-refinement
h-refinement

10−5
|J(u) − J(uDG)|

10−10

10−15 0 20 40 60 80 100
(b) sqrt (time (seconds))

Figure 8. Burgers’ equation. Comparison between h- and hp-adaptive mesh refinement: |J(u)−J(uDG )|
versus number of degrees of freedom (a); |J(u) − J(uDG )| versus computational time (b). (From
Süli E and Houston P. Error Estimation and Adaptive Discretization Methods in Computational
Fluid Dynamics, Volume 25 of Lecture Notes in Computational Science and Engineering, Adaptive
finite element approximation of hyperbolic problems, Barth T and Deconink H (eds). 269-344, 2002,
Copyright Springer-Verlag Gmbh & Co.KG, Berlin.)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

24 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

104 hp-refinement
h-refinement

103
Time (seconds)

102

101

100

102 103 104 105


Degrees of freedom

Figure 9. Burgers’ equation. Computational time versus number of degrees of freedom. (From
Süli E and Houston P. Error Estimation and Adaptive Discretization Methods in Computational
Fluid Dynamics, Volume 25 of Lecture Notes in Computational Science and Engineering, Adaptive
finite element approximation of hyperbolic problems, Barth T and Deconink H (eds). 269-344, 2002,
Copyright Springer-Verlag Gmbh & Co.KG, Berlin.)

a Mach 0.5 flow at a zero angle of attack, and a far-field density ̺ = 1 and pressure p = 1. In
this case, no shock-capturing term is used since the solution is very smooth. The second flow is
obtained by imposing this time a Mach 0.8 flow and an angle of attack α = 1.25◦ . Since in this
case, the solution presents a shock, the shock-capturing term is turned on. In Figure 11, we see
how easily the DG method handles meshes with hanging nodes and with different polynomials
degrees in different elements. In Figure 12, we also see that hp-adaptivity is more efficient than
h-adaptivity even in the presence of shocks.

In van der Vegt and van der Ven (2002b (see also the paper by van der Ven and van der
Vegt (2002) have considered shock-capturing DG methods for the time-dependent compressible
Euler equations of gas dynamics. Accordingly, they have used space-time elements, which
allow them to easily deal with moving bodies. Their shock-capturing term uses both the local
residuals as well as the jumps, which, as we have seen, are also related to the local residuals; for
details, see van der Vegt and van der Ven (2002b. They have shown that this method can be
efficiently used with mesh adaptation. As an example, we show in Figure 13, the approximation
of the method with mesh adaptation on a time-dependent Mach 0.8 flow around an oscillating
NACA 0012 airfoil. The pitching angle is between −0.5◦ and 4.5◦ , and the circular frequency
is ω = π/10. A more spectacular example is shown in Figures 14, 15, and 16, where their DG
method is applied to helicopter flight.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 25

Figure 10. Burgers’ equation. h- and hp-meshes after 11 refinements, with 999 elements and 26 020
degrees of freedom; here, |J(u) − J(uDG )| = 1.010 × 10−7 . (From Süli E and Houston P. Error
Estimation and Adaptive Discretization Methods in Computational Fluid Dynamics, Volume 25 of
Lecture Notes in Computational Science and Engineering, Adaptive finite element approximation of
hyperbolic problems, Barth T and Deconink H (eds). 269-344, 2002, Copyright Springer-Verlag Gmbh
& Co.KG, Berlin.)

1 1

0.8 0.8 6(1%)


7(1%)
0.6 0.6

0.4 0.4 5(12%)

0.2 6(17%) 0.2

0 0 4(42%)

−0.2 5(59%) −0.2

−0.4 −0.4 3(44%)

−0.6 −0.6
4(23%)
−0.8 −0.8 2(1%)

−1 −1
(a) −1 −0.5 0 0.5 1 (b) −1 −0.5 0 0.5 1

Figure 11. Flow around a NACA0012 airfoil sub-sonic (a) and supersonic (b). The actual hp-meshes
after 10 refinements. For the subsonic flow, the hp-mesh has 325 elements, 45 008 degrees of freedom,
and produces an error |J(u) − J(uh )| = 3.756 × 10−7 . For the transonic flow, it has 783 elements,
69 956 degrees of freedom, and produces an error of |J(u) − J(uDG )| = 1.311 × 10−4 . (From Süli E and
Houston P. Error Estimation and Adaptive Discretization Methods in Computational Fluid Dynamics,
Volume 25 of Lecture Notes in Computational Science and Engineering, Adaptive finite element
approximation of hyperbolic problems, Barth T and Deconink H (eds). 269-344, 2002, Copyright
Springer-Verlag Gmbh & Co.KG, Berlin.)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

26 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

10−2 hp-refinement hp-refinement


h-refinement 10−2 h-refinement
10−3

Enorm
Enorm

10−4
10−3

10−5

10−6 10−4

50 100 150 200 250 300 103 104 105


(a) Degrees (b) Degrees

Figure 12. Flow around a NACA0012 airfoil sub-sonic (a) and supersonic (b). Comparison
between h- and hp-adaptive mesh refinement. (From Süli E and Houston P. Error Estimation and
Adaptive Discretization Methods in Computational Fluid Dynamics, Volume 25 of Lecture Notes in
Computational Science and Engineering, Adaptive finite element approximation of hyperbolic problems,
Barth T and Deconink H (eds). 269-344, 2002, Copyright Springer-Verlag Gmbh & Co.KG, Berlin.)

4.4. The RKDG methods

There are two main differences between the RKDG methods and the shock-capturing DG
methods. The first is that the RKDG methods use an explicit Runge-Kutta scheme to evolve
the approximate solution in time; this renders them very easy to implement and much more
parallelizable. The second is that whereas the shock-capturing DG methods converge to the
entropy solution, thanks to the inclusion in their weak formulation of the shock-capturing
terms, the RKDG achieve this by using a slope limiters. Although these two techniques have
the very same origin, as we showed in the previous section, the use of the slope limiters results
in sharper approximations to the shocks and contact discontinuities.

In this section, we consider the Runge-Kutta discontinuous Galerkin (RKDG) methods for
nonlinear hyperbolic systems in divergence form,
N
X
ut + (fi (u))xi = 0
i=1

To define the RKDG methods, we proceed in three steps. In the first, the conservation law is
discretized in space by using a discontinuous Galerkin (DG) method. After discretization,
the system of ordinary differential equations (d/dt)uh = L(uh ) is obtained. Since the
approximation is discontinuous, the so-called mass matrix is block diagonal and hence, easily
invertible. In the second step, an explicit strong stability preserving (SSP) Runge-Kutta
method is used to march in time. The distinctive feature of the strong stability preserving
Runge-Kutta (SSP-RK) methods is that their stability follows from the stability of the forward
Euler step. Finally, in the third step, a generalized slope limiter ΛΠh is introduced in order to
enforce the above-mentioned stability property of the Euler forward step.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 27

1
1

0.5
0.5
–Cp

–Cp

0
0
T = 37.5 T = 47.0
α = 0.23° α = 4.0°
−0.5 −0.5
Medium grid with adaptation Medium grid with adaptation
Fine grid Fine grid
−1 −1

0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1


X X

Figure 13. Adapted mesh around oscillating NACA 0012 airfoil, contours of density, and pressure
coefficient Cp on the airfoil surface for α = 0.23◦ (pitching upward) and α = 4.0◦ (pitching downward)
(M∞ = 0.8, ω = π/10, α = 2◦ ± 2.5◦ ). (From van der Vegt JJW and van der Ven H. Space-time
discontinuous Galerkin finite element method with dynamic mesh motion for inviscid compressible
flows: I. General formulation. J. Comput. Phys. 2002b; 182:546-585.)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

28 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

(a)

(b)

Figure 14. Four-dimensional simulation of the Operational Loads Survey rotor in forward flight. Grid
cross-section at z = 0 (a) and vorticity levels (b) on adapted mesh at azimuth ψ = 140◦ . The tip
vortex of the blade in the upper corner lies above the z = 0 plane. (Mtip = 0.664, advance ratio 0.164,
and thrust 0.0054, flow is coming from the left). (Reproduced from van der Ven H and Boelens OJ.
(2003). Towards affordable CFD simulations of rotors in forward flight, A feasibility study with future
application to vibrational analysis. In 59th American Helicopter Society Forum, Phoenix, Arizona,
NLR-TP-2003-100, 6-8 May, 2003, by permission of American Helicopter Society.)

In what follows, we give a detailed construction of the RKDG method for the model problem
of the scalar conservation law in one space dimension. Then, we briefly discuss the extension
of the method to hyperbolic systems in several space dimensions and present numerical results
showing the performance of the method.

4.5. RKDG methods for scalar hyperbolic nonlinear conservation laws

Let us define the RKDG method for the Cauchy problem for the scalar hyperbolic nonlinear
conservation law
ut + f (u)x = 0, in (0, 1) × (0, T )
u(x, 0) = u0 (x), ∀ x ∈ (0, 1)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 29

X
Y

Vortex blade 2

Vortex blade 1
Blade 1

Vortex blade 2

(a)

(b)

Figure 15. Adapted Caradonna-Tung rotor mesh (135.280 elements) with periodic plane at z = 0 and
horizontal plane at x = −3.6, showing the refined regions at the vortex locations (a). Vorticity contours
(|ω| = 0.175) for the Caradonna-Tung rotor in hover, collective pitch 12◦ , and Mtip = 0.61 (b). (From
Boelens OJ, van der Ven H, Oskam B and Hassan AA. The boundary conforming discontinuous
Galerkin finite element approach for rotorcraft simulations. J. Aircraft 2002; 39:776-785.)

with periodic boundary conditions.

4.5.1. The DG space discretization. Let us triangulate the domain [0, 1) with the partition
Th = { Ij }Ni=1 where Ij = (xj−1/2 , xj+1/2 ). The initial data uh (·, 0)|Ij is simply the L -
2

projection of u0 |Ij on the space Pk (Ij ), that is, it is the only element of Pk (Ij ) such that

(uh (·, 0), v)Ij = (u0 , v)Ij (5)

for all v ∈ Pk (Ij ). For t > 0, we take the approximate solution uh (·, t)Ij to be the element of
Pk (Ij ) such that

(uh (·, t))t , v)Ij − (f (uh (·, t)), vx )Ij


D E
+ fb(uh (·, t)) nIj , v =0 (6)
∂Ij

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

30 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Figure 16. Four-dimensional simulation of the Operational Loads Survey rotor in forward flight. Full
space mesh of adapted mesh in intermediate time level at azimuth ψ = 151.25◦ . This intermediate
time level has completely been generated by local mesh refinement in time. (M = 0.664, advance
ratio 0.164, and thrust 0.0054). (Reproduced from van der Ven H and Boelens OJ. Towards affordable
CFD simulations of rotors in forward flight, A feasibility study with future application to vibrational
analysis. In 59th American Helicopter Society Forum, Phoenix, Arizona, NLR-TP-2003-100, 6-8 May,
2003, by permission of American Helicopter Society.)

for all v ∈ Pk (Ij ), where fb(uh ) is the numerical flux, which can be taken as indicated in the
previous section. This completes the definition of the DG space discretization.

Note that, thanks to the fact that the approximations are discontinuous, the mass matrix
is block diagonal, each block being of order (k + 1). Moreover, this matrix can be rendered
diagonal if we use (properly mapped) Legendre polynomials. Indeed, if, for x ∈ Ij , we write

k
X  
2 (x − xj )
uh (x, t) = uℓj ϕjℓ (x), ϕjℓ (x) = Pℓ
∆xj
ℓ=0
∆xj = xj+1/2 − xj−1/2

then, the initial condition (5) becomes


Z
(2ℓ + 1)
uℓj (0) = u0 (x) ϕjℓ (x)dx, ℓ = 0, . . . , k
∆xj Ij

and the weak formulation (6) takes the following simple form:

d ℓ (2ℓ + 1)
uj (t) + (−(f (uh (·, t)) (ϕjℓ )x )Ij
dt ∆xj
+ hfb(uh (·, t)), ϕj i∂I ) = 0
ℓ j

for ℓ = 0, . . . , k.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 31

Note that when f (u) = u, the system of equations for the degrees of freedom are:

d ℓ (2ℓ+1) Pℓ−1 ℓ+m m
u
dt j (t) + ∆xj m=0 (−1) uj
Pk Pk 
+ m=ℓ um j −
ℓ m
m=0 (−1) uj−1 = 0

for ℓ = 0, . . . , k. The dissipation of this method was shown to be of order 2k + 2 and the
dispersion of order 2k + 3 by Hu and Atkins (2002; Ainsworth (2004; see also Sherwin (2000).
A sharp stability analysis of the method has been carried out by Krivodonova and Qin (2013a;
Krivodonova and Qin (2013b.

Let us verify that the approximate solution remains bounded in the L2 -norm. It is easy to
see that the exact solution satisfies
d
k u(·, t) k2L2 (0,1) = 0
dt
The approximate solution satisfies, instead
d
k uh (·, t) k2L2 (0,1) + Θh (uh (·, t)) = 0
dt
where !
N
X Z u+
h
Θh (v) = (f (s) − fb(u− +
h , uh ))ds (xi+1/2 ) ≥ 0
i=1 u−
h

For details, see Jiang and Shu (1994; see also Cockburn and Gremaud (1996.

4.5.2. The SSP-RK time discretization. We discretize in time by using the following K-stage
SSP-RK method:

(0)
1. Set uh = unh ;
2. For i = 1, . . . , K compute the intermediate functions:
i−1
X
(i) (l) βil n (l)
uh = αil whil , whil = uh + ∆t Lh (uh );
αil
l=0

3. Set un+1
h = uK
h.

The method is called SSP if

(i) If βil 6= 0 then αil 6= 0,


(ii) αil ≥ 0,
Pi−1
(iii) l=0 αil = 1.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

32 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Table 2. TVD-RK time discretization parameters.

Order αil βil max{βil /αil }

2 1 1 1
1 1 1
2 2 0 2
1 1
3 1 1
3 4 4 0 4 1
1 2 2
3 0 3 00 3

These methods were originally called TVD-RK methods as they preserved the TVD property
of numerical schemes for nonlinear conservation laws. They were introduced by Shu (1988 and
by Shu and Osher (1988. Examples are displayed in Table 2; more can be found in the paper
by Gottlieb and Shu (1998. See also the recent review by Gottlieb et al. (2000.

The main property of these methods is that their stability follows from the stability of the
(l)
forward Euler steps whil = uh + (βil /αil )∆tn . Indeed, assume that each of the Euler steps
satisfy the following stability property
(l)
| whil | ≤ | uh |
for some seminorm | · |. Then
i−1
X
(i) il
| uh | = αil wh

l=0
i−1
X
≤ αil | whil | by the positivity property (ii)
l=0
i−1
X (l)
≤ αil | uh | by the stability assumption
l=0
(l)
≤ max | uh | by the consistency property (iii)
0≤l≤i−1

It is clear now that the inequality |unh | ≤ |Ph u0 |, ∀ n ≥ 0, follows from the above inequality by
a simple induction argument.

It is well known that the L2 -stability of the method (in the linear case) is necessary in order
to prevent the growth of the round-off errors. Such a stability property is usually achieved
under a condition of the form
∆t
| c| ≤ CFLL2
∆x
In Table 3, we display the numbers CFLL2 for a wide variety of time and space discretizations;
they have been obtained by numerically. The symbol ‘⋆’ indicates that the method is unstable

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 33

Table 3. The CFLL2 numbers for polynomials of degree k and RK methods of order ν.

k 0 1 2 3 4 5 6 7 8

ν =1 1.000 ⋆ ⋆ ⋆ ⋆ ⋆ ⋆ ⋆ ⋆
ν =2 1.000 0.333 ⋆ ⋆ ⋆ ⋆ ⋆ ⋆ ⋆
ν =3 1.256 0.409 0.209 0.130 0.089 0.066 0.051 0.040 0.033
ν =4 1.392 0.464 0.235 0.145 0.100 0.073 0.056 0.045 0.037
ν =5 1.608 0.534 0.271 0.167 0.115 0.085 0.065 0.052 0.042
ν =6 1.776 0.592 0.300 0.185 0.127 0.093 0.072 0.057 0.047
ν =7 1.977 0.659 0.333 0.206 0.142 0.104 0.080 0.064 0.052
ν =8 2.156 0.718 0.364 0.225 0.154 0.114 0.087 0.070 0.057
ν =9 2.350 0.783 0.396 0.245 0.168 0.124 0.095 0.076 0.062
ν = 10 2.534 0.844 0.428 0.264 0.182 0.134 0.103 0.082 0.067
ν = 11 2.725 0.908 0.460 0.284 0.195 0.144 0.111 0.088 0.072
ν = 12 2.911 0.970 0.491 0.303 0.209 0.153 0.118 0.094 0.077

when the ratio ∆t/∆x is held constant. For DG discretizations using polynomials of degree k
and a k + 1 stage RK method of order k + 1 (which give rise to an (k + 1)-th order accurate
method), we can take
1
CFLL2 =
2k + 1

The issue of the stability of the Euler forward step wh = uh + δ ∆tn L(uh ), where δ is a
positive parameter, is by far more delicate. Indeed, from Table 3, we see that this step is always
unstable in L2 . On the other hand, when the method uses piecewise-constant approximations,
then the forward Euler step is nothing but a monotone scheme, which is total variation
diminishing (TVD), that is,
| wh |TV(0,1) ≤ | uh |TV(0,1)
where X
| uh |TV(0,1) ≡ | uj+1 − uj |
1≤j≤N

is the total variation of uh . Hence, if we use piecewise polynomial approximations, it is


reasonable to try to see if the Euler forward step under consideration is stable for the following
seminorm X
| uh |TVM(0,1) ≡ | uj+1 − uj |
1≤j≤N

where uj is the mean of uh in the interval Ij . Thus, this seminorm is the total variation of the
local means of uh . The following result gives the conditions for the Euler forward step to be
nonexpansive with respect to this seminorm.
Proposition 1. (The sign conditions) We have
| wh |TVM(0,1) ≤ | uh |TVM(0,1)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

34 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

provided that
sign (u+ +
j+1/2 − uj−1/2 ) = sign (u0j+1 − u0j )
sign (u− −
j+1/2 − uj−1/2 ) = sign (u0j − u0j−1 )
and provided that !
| fb(a, ·) |Lip | fb(·, b) |Lip
|δ| + ≤1
∆j+1 ∆xj

This result states that a discretization in space by the DG method and an SSP-RK time
discretization of the resulting system of ordinary differential equations does not guarantee
a nonexpansive total variation in the local means. Fortunately, the sign conditions can be
enforced by a generalized slope limiter, ΛΠh .

4.5.3. The generalized slope limiter. Next, we construct the operator ΛΠh . To do that, let
us denote by vh1 the L2 -projection of vh into the space of piecewise-linear functions. We then
define uh = ΛΠh (vh ) on the interval Ij , as follows:

(i) Compute
u−
j+1/2 = −
v j + m ( vj+1/2 − v j , v j − v j−1 , v j+1 − v j )
u+
j−1/2 = +
v j − m ( v j − vj−1/2 , v j − v j−1 , v j+1 − v j )

(ii) If u− − + +
j+1/2 = vj+1/2 and uj−1/2 = vj−1/2 , set uh |Ij = vh |Ij ,
O 1
(iii) If not, take uh |Ij equal to ΛΠh (vh ).

This generalized slope limiter does not degrade the accuracy of the scheme, except at critical
points. In order to avoid that, we replace the minmod function m by the corrected minmod
function mj defined by

a1 if |a1 | ≤ M ∆x2j
mj (a1 , a2 , a3 ) =
m (a1 , a2 , a3 ) otherwise
where M is an upper bound of the absolute value of the second-order derivative of the solution
at local extrema.

We have the following result.


Proposition 2. (The TVBM property) Suppose that for j = 1, . . . , N
!
| fb(a, ·) |Lip | fb(·, b) |Lip 1
|δ| + ≤
∆j+1 ∆xj 2
Then, if uh = ΛΠh,M vh , then
| wh |TVM(0,1) ≤ | uh |TVM(0,1) + CM ∆x

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 35

Note that the condition on δ is independent of the form that the approximate solution has in
space.

4.5.4. The nonlinear boundedness of the RKDG method. For this method, we have the
following boundedness result.
Theorem 3. (TVBM-stability of the RKDG method) Let each time step ∆tn satisfy the
following CFL condition:
!
βil n | fb(a, ·) |Lip | fb(·, b) |Lip 1
max ∆t + ≤ (7)
il αil ∆j+1 ∆xj 2

Then we have
| unh |TVM(0,1) ≤ | u0 |TV(0,1) + CM Q ∀ n = 0, . . . , L
where L ∆x ≤ Q.

Let us emphasize that, as we have seen, the DG space discretization, the RK time
discretization, and the generalized slope limiter are intertwined just in the right way to achieve
the above nonlinear stability result. Thus, although the DG space discretization of this method
is an essential distinctive feature, the other two ingredients are of no less relevance.

Note that the above result holds for any polynomial degree and for any order of accuracy in
time. This shows that this stability result does not impose an accuracy barrier to the method,
as happens with many other methods. The RKDG method can actually achieve high-order
accuracy when the exact solution is smooth because the generalized slope limiter does not
degrade the high-order accuracy of the space and time discretizations. Although there are
no theoretical error estimates that justify this above statement, it is actually supported by
overwhelming practical evidence.

Note also that for the linear case f (u) = c u, the CFL condition (7) becomes
∆t 1
|c| ≤ CFLTV ≡
∆x 2 max αβilil

In general, the restriction of the time step imposed by the TVBM property is much weaker than
that required to achieve L2 -stability. However, it is the condition for L2 stability that needs to
be respected; otherwise, the round-off errors would get amplified and the high-order accuracy
of the method would degenerate even though the RKDG method remains TVBM-stable.

4.5.5. Generalized slope limiters, discontinuity detection and artificial viscosity techniques.
Although the generalized slope limiter just discussed is fairly simple to implement, how to
estimate the constant M is not easy. The generalized slope limiters by Biswas et al. (1994;
Burbeau et al. (2001, see also Krivodonova (2007, bypass this difficulty but they can still modify
the approximation and degrade the original accuracy of the method. A new approach which

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

36 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

overcomes this problem, based on seeing the slope limiter as an artificial diffusion operator
(see Cockburn (2001) was proposed in Casoni et al. (2013.

Another approach is the essentially non-oscillatory and weighted essentially non-oscillatory


techniques, see Zhu et al. (2013 and the references therein, which maintain nonlinear
boundedness and maintains the formal order of accuracy of the method. However, they are
very involved, especially for high-degree polynomial approximations and require structured
meshes. On the other hand, the limiter proposed in Zhu et al. (2013 works for unstructured
triangular meshes. To render the generalized slope limiters more efficient, a popular approach
is to apply them only at the elements at which the approximate solution might be capturing a
discontinuity of the exact solution. In this way, a sophisticated generalized slope limiter could
be applied there. See the comparison carried out by Qiu and Shu (2005.

4.5.6. Convergence properties. It is not difficult to use Theorem 3 to conclude, by using a


discrete version of the Ascoli-Arzelá theorem, that from the sequence { uh }∆x>0 , it is possible
to extract a subsequence strongly converging in L∞ (0, T ; L1 (0, 1)) to a limit u⋆ . That this limit
is a weak solution of the nonlinear conservation law can be easily shown. However, while there
is ample numerical evidence that suggests that u⋆ is actually the entropy solution, this fact is
still a challenging theoretical open problem.

In the one-dimensional case of the transport equation with an initial condition displaying a
discontinuity, it was shown by Cockburn and Guzmán (2008 that the RKDG with k = 1 and a
second-order time-marching RK method that, at time T , the L2 -error is second order in the size
of the mesh, h, outside a region of size O(T 1/2 h1/2 log 1/h) to the right of the discontinuity
and of size O(T 1/3 h2/3 log 1/h) to the left. For any polynomial degree k ≥ 1 and a third-order
time-marching RK method , it was shown by Zhang and Shu (2014 that the L2 -error is of
order min{k + 1, 3} in the size of the mesh, outside a region of size O(T 1/2 h1/2 log 1/h) to
both the right and the left of the location of the discontinuity. In both cases, the standard
CFL condition is assumed.

Error estimates for k = 1 elements and a second-order time-marching RK method are


obtained in the L∞ (0, T ; L2 (Rd ))-norm for nonlinear conservation laws in one-space dimension
and for linear conservation laws in multiple space dimensions by Shu and Zhang (2004. The
order of convergence of min{k + 1/2, 2} is obtained for general monotone numerical fluxes, and
that of min{k + 1, 2} for upwind numerical fluxes; the minimal CFL condition for stability of
the method. is assumed. The extension to symmetrizable hyperbolic systems was carried out
by Zhang and Shu (2006. Results for the case in which the time-marching RK is of order three
have been done by Zhang and Shu (2010a; Luo et al. (2015; Meng et al. (2016.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 37

4.6. RKDG methods for multidimensional hyperbolic systems

The extension of the RKDG methods to the model multidimensional hyperbolic system
N
X
ut + (fi (u))xi = 0
i=1

deserves comments on a few key points.

4.6.1. The basis functions. Just as in the one dimensional case, the mass matrix is block-
diagonal; the block associated with the element K is a square matrix of order equal to the
dimension of the local space and hence, can be easily inverted. Moreover, for a variety of
elements and spaces, a basis can be found, which is orthonormal in L2 . This is the case, for
example, of rectangles and tensor product polynomials, in which case, the orthonormal basis
is a properly scaled tensor product of Legendre polynomials. For simplices and polynomials
of a given total degree, there is also an orthonormal basis; see the work by Dubiner (1991,
Karniadakis and Sherwin (1999, and Warburton (1998, and the recent implementations by
Aizinger et al. (2000 and Hesthaven and Warburton (2002.

4.6.2. Quadrature rules. In practice, the integrals appearing in the weak formulation need
to be approximated by quadrature rules. It was proven by Cockburn et al. (1990 that

kLh (u) + ∇ · f (u)kL∞ (K) ≤ C hk+1 |f (u)|W k+2,∞ (K)

if the quadrature rules over each of the faces of the border of the element K are exact for
polynomials of degree 2k + 1, and if the one over the element is exact for polynomials of degree
2k. The fact that these requirements are also necessary, can be easily numerically verified;
moreover, the method is more sensitive to the quality of the quadrature rules used on the
boundary of the elements than to that used in their interior.

Finally, let us point out that a quadrature-free version of the method was devised by Atkins
and Shu (1998, which results in a very efficient method for linear problems and certain nonlinear
problems such as Euler equations of gas dynamics. A very efficient quadrature rule was obtained
by van der Ven and van der Vegt (2002 for the Euler equations of gas dynamics by suitably
exploiting the structure of the equations.

4.6.3. Numerical fluxes. When dealing with multidimensional hyperbolic systems, the so-
called local Lax-Friedrichs numerical flux is a particularly convenient choice of numerical flux.
Indeed, it can be easily applied to any nonlinear hyperbolic system, it is simplePto compute,
N
and yields good results. This numerical flux is defined as follows. If we set fnK = i=1 ni fi (u),
we define the local Lax-Friedrichs numerical flux as

bf LLF (uh ) = {f (uh )} − C [[uh ]]n


nK nK K
2
Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

38 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

where C = C(K ± ) is the larger one of the largest eigenvalue (in absolute value) of
∂/(∂u± )fnK ± (u± ), or, in practice, of ∂/(∂u± )fnK ± (uK ± , where uK ± are the means of the
approximate solution uh in the elements K ± .

For symmetric hyperbolic systems, it is possible to devise numerical fluxes that render the
method of lines (or the space-time methods) L2 -stable; see Barth (2000.

4.6.4. The slope limiter ΛΠh . When we dealt with the scalar one-dimensional conservation
law, the role of the generalized slope limiter ΛΠh was to enforce the TVBM property of a
typical Euler forward time step. In the case of multidimensional scalar conservation laws, we
cannot rely anymore on the TVBM property of the Euler forward step because such a property
does not hold for monotone schemes on general meshes; it has been proven only for monotone
schemes in nonuniform Cartesian meshes by Sanders (1983. We can, instead, rely on a local
maximum principle; see the paper by Cockburn et al. (1990.

A practical and effective generalized slope limiter ΛΠh,M was later developed by Cockburn
and Shu (1998b. To apply it to the function vh , we proceed on the element K as follows:

(i) Compute the L2 -projection of vh into the linear functions on K, vh1 |K ,


(ii) Compute rh |K = ΛΠ1h,M vh1 |K ,
(iii) If rh |K = vh1 |K , set uh |K = vh |K ,
(iv) If not, set uh |K = rh |K .

Note that in order to use this generalized slope limiter, one only needs to know how to slope
limit piecewise linear functions; for the details of the definition of ΛΠ1h,M , we refer the reader
to the paper by Cockburn and Shu (1998b.

An interesting limiter has been proposed by Wierse (1997. Kershaw et al. (1998 introduced
a limiter based on quadratic programming. Biswas et al. (1994 devised a limiter based on local
moments, rather than on slopes, and used it for adaptivity purposes. Burbeau et al. (2001
proposed what they call a problem-independent slope limiter.

4.6.5. Characteristic variables. For systems, limiting in the local characteristic variables gives
remarkably superior results than doing it component-by-component.

To limit the vector ṽh (mi , K0 ) in the element K0 (see Figure 4.6.5), we proceed as follows:

• Find the matrix R and its inverse R−1 , which diagonalizes the Jacobian
∂ mi − b 0
J= f (v K0 ) ·
∂u |mi − b0 |

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 39

b3
K3
b1
b0 m1
K1

K0

b2

K2

Figure 17. Illustration of limiting.

that is, R−1 JR = Λ, where Λ is a diagonal matrix containing the eigenvalues of J. Notice
that the columns of R are the right eigenvectors of J and the rows of R−1 are the left
eigenvectors.
• Transform veh (mi , K0 ) and ∆v(mi , K0 ) to the characteristic fields. This is achieved by
left multiplying these vectors by R−1 .
• Apply the scalar limiter to each of the components of the transformed vectors.
• Multiply by R on the left to transform the result back to the original space.

4.7. Computational results

In this section, we display computational results that show that the RKDG method can achieve
exponential convergence when the solution is very smooth and that it can perform as well as
the high-resolution methods when discontinuities are present. We also show results showing
its excellent handling of boundary conditions and its remarkable parallelization properties.
Finally, we also show that the use of higher-degree polynomials results in a more efficient
method, even in the presence of discontinuities.

4.7.1. Exponential convergence. To show that exponential convergence can be achieved and
that it is always more efficient to use higher-degree polynomials when the exact solution is
very smooth, we consider
ut + ∇ · (v u) = 0
where v = 2π (−y, x) and the initial condition is a Gaussian hill. In Figure 4.7.1, we see the
L2 -error at time T = 1 versus the CPU time for the four different successively refined meshes
described below and for polynomials of degree up to six. The refinement of the mesh is obtained
by dividing the triangles in four congruent triangles. Each line corresponds to a different mesh,

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

40 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Fixed mesh performance


0
Mesh1
−2
Mesh2
−4 Mesh3
−6 Mesh4

−8
Log (|error|)

−10
−12
−14
−16
−18
−20
−22
−5 −4 −3 −2 −1 0 1
Log (CPU time)

Figure 18. Spectral convergence and comparison of L2 -error versus CPU time for 4 successively refined
meshes and polynomials of degree 1 to 6. (From Aizinger V, Dawson CN, Cockburn B and Castillo P.
Local discontinuous Galerkin method for contaminant transport. Adv. Water Res. 2000; 24:73-87.)

with the symbols on each line representing the error for the six different approximating spaces.
We easily observe that exponential convergence is achieved and that it is always more efficient
to use a coarser mesh with a higher-order polynomial approximation.

4.7.2. Treatment of the boundary conditions. To show the ease with which the method deals
with the boundary conditions, we consider a variation of the above problem

ut + ∇ · (v u) = 0

where v = (−y + 1/2, x − 3/4) and the initial data is

u(x, t = 0)
(    2
exp 8 − 8
1−8|x−(3/4,1)|2 , if 8 x− 34 ,1 < 1
=
0, otherwise

An RKDG method using quadratic polynomial approximations and a SSP-RK method of order
three is used. Note that, unlike the previous example, only part of the initial data is in the
computational domain. The boundary conditions are taken by using the Lax-Friedrichs flux
and by giving the exact solution as the exterior trace. In Table 4, we see that the full order
three has been achieved, as expected. In Figure 19, we also see that the boundary conditions
have been captured very well by the RKDG method.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 41

Z Z
Y Y

X X

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2
1 1
0 0
0.8 0.8
0 0.6 0 0.6
0.2 0.2
0.4 0.4 0.4 0.4
0.6 0.2 0.6 0.2
(a) 0.8 (b) 0.8

Z
Y

X
0.25
1

0.8
Y

0.6

0.4

0.2
0 1
0
0.8
0 0.6
0.2
0.25 0.5 0.4 0.4
0.6 0.2
(c) X 0.8

Figure 19. Approximate solution at T = 0.0 (a), T = (3/8)π (b) and T = (3/4)π (c). The mesh is a
uniform 64 × 64 of triangles.

Table 4. Errors at T = (3/4)π.

Mesh k eu (T ) kL∞ (Ω) Order k eu (T ) kL1 (Ω) Order

16 × 16 0.21E-01 2.01 0.42E-03 3.32


32 × 32 0.25E-02 3.07 0.42E-04 3.31
64 × 64 0.32E-03 2.96 0.49E-05 3.11
128 × 128 0.52E-04 2.64 0.60E-06 3.01

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

42 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

1 1 1

0 0 0

(a) 0 0.25 0.5 0.75 1 (b) 0 0.25 0.5 0.75 1 (c) 0 0.25 0.5 0.75 1

1 1 1

0 0 0

(d) 0 0.25 0.5 0.75 1 (e) 0 0.25 0.5 0.75 1 (f) 0 0.25 0.5 0.75 1

Figure 20. Effect of the order of the RKDG method on the approximation of discontinuities. The
exact solution at time T = 100, u (- - - -), is contrasted against the approximate solution uh ( )
obtained with the RKDG method of order k + 1 on a mesh of 40 elements for the values k = 0 (a),
k = 1 (b), k = 2 (c), k = 3 (d), k = 4 (e), and k = 5 (f). No limiter was used.

4.7.3. Approximation of contact discontinuities. Let us now show how the contact
discontinuities are approximates by the RKDG methods. To do that, we consider the problem
ut + ux = 0, in [0, 1) × (0, T )
with periodic boundary conditions and initial condition

1, if x ∈ (0.25, 0.75)
u(x, 0) =
0, otherwise
In Figures 20 and 21, we show the results given by RKDG methods using polynomials of degree
k and a (k +1)-stage, (k +1)th-order accurate SSP-RK method. We see that as the polynomials
degree increases, so does the quality of the approximation of the contact discontinuity, except,
perhaps for the unwanted oscillations near them.

4.7.4. Approximation of shocks. First, let us show in a simple example that the RKDG
methods can capture shocks as well as any high-resolution finite difference or finite volume
scheme. Consider the approximation of the entropy solution of the inviscid Burgers equation
 2
u
ut + =0
2 x
on the domain (0, 1)×(0, T ) with initial condition 1/4+sin(π(2 x−1))/2 and periodic boundary
conditions. In Figure 22, we display the RKDG solution using piecewise linear and piecewise

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 43

P1 P6

1.2 1.2
1 1
0.8 0.8
0.6 0.6
n

n
0.4 0.4
0.2 0.2
0 0
0 0
0 1 0 1
1 2 1 2
2 3 2 3
3 4 x 3 4 x
y 4 5 y 4 5
5 6 5 6
(a) 6 (b) 6

Figure 21. Effect of the order of the RKDG method on the approximation of discontinuities.
Comparison of the exact and the RKDG solutions at T = 100π with k = 1 (a) and k = 6 (b).
Two dimensional results with 40 × 40 squares. No limiter was used. (Reproduced from Cockburn B
and Shu C-W. Runge-Kutta discontinuous Galerkin methods for convection-dominated problems. J.
Sci. Comput. 2001; 16:173-261, by permission of Kluwer Academic/Plenum Publishers.)

quadratic approximations; note how, in both cases, the shock has been captured within three
elements as would be expected of any high-resolution scheme.

4.7.5. Parallelizability. Let us address the parallelizability of the RKDG method. In Table 5
below, we display the results obtained by Biswas et al. (1994; we see the solution time and
total execution time for the two-dimensional problem
ut + ux + uy = 0
2
on the domain (−π, π) × (0, T ) with initial condition u(x, y, 0) = sin(πx) sin(πy) and periodic
boundary conditions. Biswas et al. (1994 used 256 elements per processor and ran the RKDG
method with polynomials of degree two and eight time steps; the work per processor was kept
constant. Note how the solution time increases only slightly with the number of processors
and the remarkable parallel efficiency of the method.

4.7.6. Approximation of complex solutions. Let us show that the RKDG method can handle
solutions with very complicated structure. Consider the classical double-Mach reflection
problem for the Euler equations of gas dynamics. In Figure 23, details of the approximation
of the density are shown. Note that the strong shocks are very well resolved by the RKDG
solution using piecewise linear and piecewise quadratic polynomials defined on squares. Also,

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

44 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

0.75

0.5

0.25

−0.25

0 0.25 0.5 0.75 1


(a)

−0.22

−0.23

−0.24

−0.25

−0.26

(b)

Figure 22. Burgers equation: Comparison of the exact and the RKDG solutions obtained with
∆x = 1/40 at T = 0.40. Full domain (a) and zoom on three elements (b) the first of which contains the
exact shock. Exact solution ( ), piecewise linear approximation (. . . . . . . . . . . . . .), and piecewise
quadratic approximation (- - - -). (From Cockburn B. High-Order Methods for Computational Physics,
Volume 9 of Lecture Notes in Computational Science and Engineering, Discontinuous Galerkin
methods for convection-dominated problems, Barth T and Deconink H (eds), 69-224, 1999, Copyright
Springer-Verlag Gmbh & Co.KG, Berlin.)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 45

Rectangles P 2, ∆x = ∆y = 1/240
0.5

0.4

0.3

0.2

0.1

0.0
2.0 2.2 2.4 2.6 2.8
(a)

Rectangles P 1, ∆x = ∆y = 1/480

0.4

0.3

0.2

0.1

0.0
2.0 2.2 2.4 2.6 2.8
(b)

Rectangles P 2, ∆x = ∆y = 1/480

0.4

0.3

0.2

0.1

0.0
2.0 2.2 2.4 2.6 2.8
(c)

Figure 23. Euler equations of gas dynamics double Mach reflection problem. Isolines of the density
around the double Mach stems. Quadratic polynomials on squares ∆x = ∆y = (1/240) (a); linear
polynomials on squares ∆x = ∆y = (1/480) (b); and quadratic polynomials on squares∆x = ∆y =
(1/480) (c). (From Cockburn B and Shu C-W. The Runge-Kutta discontinuous Galerkin finite element
method for conservation laws V: multidimensional systems. J. Comput. Phys. 1998b; 141:199-224.)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

46 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Rectangles P1, ∆x = ∆y = 1/320


1.0
0.8
0.6
0.4
0.2
0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
(a)

Rectangles P 2, ∆x = ∆y = 1/320
1.0
0.8
0.6
0.4
0.2
0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
(b)

Figure 24. Forward facing step problem. Approximation of the density ̺. 30 equally spaced contour
lines from ̺ = 0.090338 to ̺ = 6.2365. (From Cockburn B and Shu C-W. The Runge-Kutta
discontinuous Galerkin finite element method for conservation laws V: multidimensional systems.
J. Comput. Phys. 1998b; 141:199-224.)

note that there is a remarkable improvement in the approximation of the density near the
contacts when going from linear to quadratic polynomials.

A similar conclusion can be drawn in the case of the flow of a gas past a forward facing step
(see Figure 24); see, also, the study by Woodward and Colella (1984.

4.7.7. Problems with curved boundaries. Bassi and Rebay (1997b showed the importance
of approximating as accurately as possible the boundaries of the physical domain and the
ease with which this is achieved by using the RKDG methods. Indeed, for the classical
two-dimensional isentropic flow around a circle, they showed that approximating the circle
by a polygon results in nonphysical entropy production at each of the kinks, which is then
carried downstream and accumulates into a nonphysical wake, which does not disappear by
further refining the mesh. However, by simply taking into account the exact shape of the
boundary, a remarkably improved approximation is obtained; see Figure 25.

On the other hand, van der Vegt and van der Ven (2002a have shown that the high-order
accurate representation of the curved boundary can be avoided by using local mesh refinement.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 47

Table 5. Scaled parallel efficiency. Solution times (without I/O) and total execution times measured on
the nCUBE/2. (From Biswas R, Devine KD and Flaherty J. Parallel, adaptive finite element methods
for conservation laws. Appl. Numer. Math. 1994; 14:255-283.)

Number of Solution parallel Total parallel


processors Work (W ) Solution time (secs.) efficiency (%) Total time (secs.) efficiency (%)

1 18 432 926.92 - 927.16 -


2 36 864 927.06 99.98 927.31 99.98
4 73 728 927.13 99.97 927.45 99.96
8 147 456 927.17 99.97 927.58 99.95
16 294 912 927.38 99.95 928.13 99.89
32 589 824 927.89 99.89 929.90 99.70
64 1 179 648 928.63 99.81 931.28 99.55
128 2 359 296 930.14 99.65 937.67 98.88
256 4 718 592 933.97 99.24 950.25 97.57

(a) (b)

Figure 25. Mach isolines of the DG approximation with P1 elements the circle is approximated by a
polygonal (a) and rendered exactly (b). (From Bassi F and Rebay S. High-order accurate discontinuous
finite element solution of the 2-D Euler equations. J. Comput. Phys. 1997b; 138:251-285.)

4.7.8. Adaptivity for the Euler equations of gas dynamics. Next, we give examples of
adaptivity using the RKDG method with anisotropic mesh refinement. The first two examples
illustrate the use of conforming mesh refinement. For the first example, two Sedov-type
explosions in an open square domain develop and interact while bouncing on square obstacles
and interacting with each other; see Figure 26. In the second example, the blast of a cannon
is simulated in order to understand the shape of the blast waves around the muzzle break; see
Figure 27.

Finally, we present an example of a steady state computation on an ONERA M6 wing for


which nonconforming refinement has been employed; see Figure 28.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

48 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

(a)

(b)

Figure 26. Two explosions in a square domain with obstacle density (a), and
the corresponding mesh (b) after 1 second. (From Remacle J-F, Li X, Chevau-
geon N, Shephard MS and Flaherty JE. (2003). Anisotropic Adaptive Simulation
of Transient Flows Using Discontinuous Galerkin Methods, 2003, Submitted.)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 49

(a) (b)

Figure 27. Cannon-blast simulation density and the corresponding mesh. Close up on the muzzle break
at an early stage (a), and after the main blast wave left the muzzle (b). (From Remacle J-F, Li X,
Chevaugeon N, Shephard MS and Flaherty JE. (2003). Anisotropic Adaptive Simulation of Transient
Flows Using Discontinuous Galerkin Methods. 2003, Submitted.)

4.7.9. Simulation of inertial confinement fusion. Our final example is the simulation of the
implosion of a NIF capsule which consists of a nearly vacuous inner region enclosed by two
spherical shells. For details, see Shestakov et al. (2001. This is a complicated and very difficult
problem which involves the simulation of hydrodynamics, heat conduction and radiation
transport phenomena. Only the hydrodynamics part of the problem is simulated by using
a one-step ALE, RKDG method proposed by Kershaw et al. (1998 and implemented in the
ICF3D code by Shestakov et al. (2000. In Figure 31, we see the mesh and several physical
quantities after 8 nanoseconds of having deposited energy on the outer surface of the capsule.
Note the near spherical symmetry of the implosion.

4.8. Enforcing range invariance of the DG approximations

To end this section, we describe a technique for enforcing a range-invariance property of the
RKDG methods was introduced by Zhang and Shu (2010b in the framework of nonlinear scalar
hyperbolic conservation laws.

For the sake of clarity, we describe this technique for RKDG methods for the initial-value
problem:
ut + f (u)x = 0 in (0, T ) × R u = u0 on {T } × R.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

50 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Figure 28. Final adapted mesh on ONERA M6 wing, clearly showing the lambda shock (flow conditions
M∞ = 0.84, α = 3.06◦ ). (From van der Vegt JJW and van der Ven H. Discontinuous Galerkin finite
element method with anisotropic local mesh refinement for inviscid compressible flows. J. Comput.
Phys. 1998; 182:46-77.)

In this case, it is well known that the exact solution satisfies the range-invariance (or maximum
principle) property
u(t, x) ∈ [m, M ] for all (t, x) ∈ (0, T ) × R,
where m := inf x∈R u0 (x) and M := supx∈R u0 (x).

Suppose that we are given the RKDG approximation at the i-th intermediate stage after the
application of the generalized slope limiter ΛΠh,M . Then, on the element Ij := (xj−1/2 , xj+1/2 ),
we modify the function as follows:
(i) (i) (i) (i)
uh := θ(uh − uh ) + uh ,

where ( )
M − u(i) m − u(i)
h h
θ := , ,1 ,
Mj − u(i) mj − u(i)
h h
(i) (i)
and mj := inf x∈Sj uh (x) and Mj := supx∈Sj uh (x). Here Sj denotes a set of suitably chosen
points lying on [xj−1/2 , xj+1/2 ]. Typically, they are chosen as Gauss-Lobatto quadrature points.
This modification does not alter the high-order accuracy of the method and enforces the range-
invariance property

uh (tn , x) ∈ [m, M ] for all (t, x) ∈ (0, T ) × Sj ∀j ∈ Z.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 51

under a suitable CFL condition. An illustration is provided in Figures 29 and 30. Therein, the
slope limiter ΛΠh,M has not been applied since it is not needed to enforce convergence to the
physically relevant solution.

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

-0.2 -0.2
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
X X

1.2 1.2

+++++ +++++++++++++++++++++++++++++++++++++++++ ++++++


1 +++ +++ + 1 ++++++++++++++++++++++++++++++++++++++++++++++++++++
+
+ ++ ++
+ ++
+
0.8 + 0.8 + +
+
+
+
0.6 + 0.6 +
+ + +
+ +
+ + + +
0.4 + 0.4 +
+ +
+
+
0.2 + 0.2 + +
+ ++
+
+ ++ +
+ ++ ++
0 +++++++++++++++++++++++++++++++++++++++++++++++++++
0 +++++++++++++++++++++++++++++++++++++++++++++++++++
++++ ++

-0.2 -0.2
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
X X

Figure 29. Approximations provided by the RKDG method at time T = 100 of the solution of the
transport equation ut + ux = 0 with an initial condition u0 (x) = 0 for x ≤ 0 and u0 (x) = 1 for x > 0
1
defined on the interval (−1, 1) with periodic boundary conditions. Here, k = 5, ∆t = 16 ∆x, ∆x = 1/16
and the time-marching method is the SSP RK method or order 3. Without (left) and with (right)
bound-preserving limiter. Cell averages (top) and Gauss-Lobatto points (bottom) are plotted. Note
how the limiter enforces strict maximum principle for all Gauss-Lobatto point values and cell averages
of the numerical polynomial solution. Courtesy of Chi-Wang Shu and Xiongxiang Zhang.

Applications and extensions of this technique were carried out by Zhang and Shu (2010c;
Zhang and Shu (2011; Hu et al. (2013 to the compressible Euler equations, by Zhang
et al. (2012; Zhang et al. (2013 for triangular meshes, by Qin et al. (2016 to relativistic
hydrodynamics, and by Vilar et al. (2016b; Vilar et al. (2016a to multimaterial compressible

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

52 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Reference Reference
TVD Limiter TVD Limiter
1.2 Bound-Preserving Limiter Bound-Preserving Limiter
+ No Limiter + No Limiter
0.004 +
1 +
++ +
++
+ + + + +
+ + + + 0.003
0.8 + + +
+ +
+ + + +
0.002
0.6
+ + + +

0.4 + + + + 0.001

+ + + +
0.2 + + + + +
0 + +
+ + + + +
+ + + + + +
++ ++
+++
0 +++ ++
+++ +++
+++ ++
+++
-0.001
-1 -0.5 0 0.5 1 -0.1 0 0.1
X X

Figure 30. Approximations provided by the RKDG method at time T = 2 of the solution of the
transport equation ut + ux = 0 with an initial condition u0 (x) = sin4 (πx) defined on the interval
(−1, 1) with periodic boundary conditions. Here, k = 2, ∆t = 51 ∆x, ∆x = 1/16 and the time-marching
method is the SSP RK method or order 3. Three Gauss-Lobatto point values are ploted for each
element. Note how the limiter does not ”clip” a smooth extremum, unlike the TVD limiter. Courtesy
of Chi-Wang Shu and Xiongxiang Zhang.

flows.

For4.9. A posteriori
pioneering workerror estimationfor
on adaptivity andnonlinear
adaptivity
problems , see the papers by Biswas et al.
(1994, Devine et al. (1995, Devine and Flaherty (1996, Flaherty et al. (1997; Flaherty et al.
(1998; Flaherty et al. (1999, Flaherty et al. (2000.

The only rigorous a posteriori error estimate in the L∞ (0, T ; L1 (RN ))-norm was obtained by
Cockburn and Gremaud (1996. It holds for the approximation given by an space-time, shock-
capturing DG method for nonlinear hyperbolic conservation laws in multiple-space dimensions.
The entropy solution can display discontinuities. For more recent work on the subject see the
review by Ohlberger (2009. See also the paper by Giesselmann et al. (2015 on a posteriori
error estimates for the semidiscrete version of DG methods for smooth solutions of nonlinear
systems of conservation laws.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 53

5. Steady-state diffusion

In this section, we consider DG methods for a second-order elliptic model problem, steady-
state diffusion. We do this not only because this also has direct applications to fluid flow (in
porous media, for example) but mainly as a much needed transition towards the devising of DG
methods for convection-dominated flows. The emphasis here is on the fact that the numerical
fluxes used in the framework of hyperbolic conservation laws are not associated to approximate
Riemann solvers anymore. Instead, they are better understood if they are considered to be
numerical traces that must be chosen in order to render the DG method both stable and
accurate. We show that, also in this context, the stability of the DG methods is enhanced by
the jumps of the approximate solution and that they are a particular case of stabilized mixed
methods. We also show that the inter-element jumps control the quality of the approximate
solution and discuss the convergence properties of the methods for various choices of numerical
traces.

We present these ideas for the second-order elliptic model problem:


−∆u = f in Ω, u = uD on ∂Ω,
where Ω is a bounded domain of RN and n is the outward unit normal to its boundary.

5.1. General form of the DG methods

5.1.1. Definition. A way to define a DG method consists in rewriting the elliptic model
problem as a system of first-order equations, namely,
q = ∇u, −∇ · q = f in Ω, u = uD on ∂Ω
and then applying to it a DG discretization. Thus, the approximate solution (qh , uh ) on
the element K is taken in the space V(K) × W (K) and is defined as the solution, for all
(v, w) ∈ V(K) × W (K), of the equations
(qh , v)K + (uh , ∇ · v)K − hb
uh , nK · vi∂K = 0,
(qh , ∇v)K − hbqh · nK , vi∂K = (f, v)K .
The Dirichlet boundary condition is enforced by requiring that ubh = uD on ∂K ∩ ∂Ω. All the
DG methods are generated by choosing the local spaces V(K) × W (K) and the numerical
traces b
qh · nK and u
bh . This completes the definition of the DG methods.

Next, we discuss some simple properties that hold for all of them.

5.1.2. The numerical traces. Just as for DG methods for hyperbolic problems, the definition
of the numerical traces b
qh and u
bh strongly influences the properties of the corresponding DG
method. In this context, we also require that the numerical traces be linear functions of the
traces of qh · nK and uh (on the boundary of the element ∂K) which are consistent and

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

54 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Cells Z
PE Cells 8.002194e-09
63 D
56.7 3.9
50.4 3.51
44.1 3.12
37.8 2.73
2.34
31.5
1.95
25.2 1.56
18.9 1.17
12.6 0.784
Y X
6.3 0.394
0 0.00503
(a) (b)

Cells 8.002194e-09 Cells 8.002194e-09


Te Tr
0.715 0.158
0.646 0.144
0.576 0.13
0.506 0.116
0.436 0.102
0.367 0.0881
0.297 0.0741
0.227 0.0601
0.158 0.0461
0.0879 0.0322
0.0182 0.0182
(c) (d)

Figure 31. ICF capsule implosion mesh (a) with colors indicating to which processor the element
belongs. Density (b), matter temperature (c), and radiation temperature (d) at 8 nanoseconds. (From
Shestakov AI, Milovich JL and Prasad MK. Combining cell- and point-centered methods in 3-D,
unstructured-mesh hydrodynamics codes. J. Comput. Phys. 2001; 170:81-111.)

single valued. As for the original DG method, the first property renders the numerical traces
easy to evaluate and ensures a high degree of locality of the method. The very form of the
numerical traces also influences the way the DG methods can be implemented, as we will see.
The consistency of the traces ensures the convergence of the method to the correct solution.
The single-valuedness property, which is highly valued in computational fluid dynamics, is
also very important in this context. If violated, the method produces a stiffness matrix for the
primal variable, which is not symmetric; see the paper by Arnold et al. (2002 for a complete
discussion. More importantly, it induces a loss in the rate of convergence in uh as well as a
significant degradation in the quality of the approximation of linear functionals as shown by
Harriman et al. (2003.

5.1.3. Energy identities. If we take (v, w) := (qh , uh ) and add the resulting equations, we
obtain what we could call the local energy identity, namely,
(qh , qh )K + ΘK = hb
qh · nK , u
bh i∂K + (f, uh )K ,

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 55

where
ΘK := h(qh − b
qh ) · n K , u h − u
bh i∂K .
By adding on all the elements these equations, we obtain the global energy identity

(qh , qh )Ω + Θh = hb
qh · n, u
bh i∂Ω + (f, uh )Ω ,
P
where Θh := K∈Th ΘK . We see that the jumps (qh − b qh ) · nK and uh − u
bh can stabilize the
method it they are chosen in such a way that Θh ≥ 0.

5.1.4. Residuals and jumps. Let us next show that the method establishes linear relations
between the residuals in the interior of the element K

Rr := qh − ∇uh and Rv := −∇ · qh − f,

and the residuals on its boundary ∂K,

rr := u
bh − uh and rv := (b
q h − qh ) · n K

Indeed, a simple integration by parts gives that for all (v, w) ∈ V(K) × W (K),

(Rv , v)K = hrr , nK · vi∂K ,


(Rw , w)K = hrw , wi∂K .

Taking v := PV Rv and w := PW , where PN denotes the L2 (K)-projection into the finite


dimensional space N (K), we obtain that
−1/2
kPV Rv kK ≤ C hK krr k∂K ,
−1/2
kPW Rw kK ≤ C hK krw k∂K .

If we assume that ∇W (K) ⊂ V(K) so that PV Rr = Rr , these inequalities say that the jumps
(b
qh −qh )·nK and ubh −uh , and f −PW f , control the size of the residuals, and hence the quality
of the approximation. This implies that any a posteriori error estimate for DG methods should
only depend on those quantities.

5.1.5. An example. Let us consider the numerical traces proposed in Cockburn and Shu
(1998a and Castillo et al. (2000. With the notation

[[uh ]] = u+ + − −
h n + uh n
and [[qh ]] = q+ + −
h · n + qh · n
+

±
where ωh (x) = lim ωh (x − ǫ n± )
ǫ↓0

the numerical traces are defined as follows. Inside the domain Ω, we take

b
qh := {qh } − Cqq [[qh ]] − Cqu [[uh ]]
u
bh := {uh } − Cuu ·[[uh ]] − Cuq [[qh ]]

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

56 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

and on its boundary, we take


b
qh := qh − Cqu (uh − uD ) n
u
bh := uD .
In this case, assuming that Cqq + Cuu = 0, a simple computation gives us that the global
energy identity is
(qh , qh )Ω + Θh = hb
qh · n, uD i∂Ω + (f, uh )Ω ,
where
Θh (qh , uh ) := hCuq [[qh ]], [[qh ]]iFhi + hCqu [[uh ]], [[uh ]]iFhi + hCqu (uh − uD ), uh − uD i∂Ω .
We see that, if we assume that Cuq and Cqu are nonnegative, the quantity Θh can be
interpreted as the term capturing the energy of the inter-element jumps [[qh ]] and [[uh ]], and the
jumps uh − uD on ∂Ω. This suggests that the role of Cuq and Cqu is to stabilize the method.
The role of the other coefficient Cqq = −Cuu , can be that of maximizing the sparsity of the
matrices, see Cockburn et al. (2001, or even that of enforcing the stability and convergence
properties of the DG method resulting when Cuq = Cqu = 0 on Fhi , see Cockburn and Dong
(2007.

We can use the above global energy identity to obtain the existence and uniqueness of the
approximate solution in the simple but important case in which
(i) Cuq ≥ 0 on Fhi ,
(ii) Cqu > 0 on Fhi ∪ ∂Ω,
(iii) ∇W (K) ⊂ V(K) ∀K ∈ Th .
Again, we only have to show that, when we set the data f and uD to zero, the only solution
is the trivial one. But in this case, the above energy identity gives that
(qh , qh )Ω + hCuq [[qh ]], [[qh ]]iFhi + hCqu [[uh ]], [[uh ]]iFhi + hCqu uh , uh i∂Ω = 0,

which implies, by (i) and (ii), that qh = 0 on Ω, that [[uh ]] = 0 on Fhi , and that Cqu uh = 0 on
∂Ω. As a consequence, we have that u bh = uh on Fhi ∪ ∂Ω, and the first equation defining the
DG method reads
−(∇uh , r)K = 0
for all r ∈ V(K). By (iii) we can take r := ∇uh and conclude that uh is a constant on the
element K. Since uh = ubh on the interelement boundaries, uh is a constant on Ω, and since
uh = 0 on ∂Ω, we get that uh = 0 on Ω. This completes the proof.

Let us end by pointing out that this DG method is in fact a stabilized mixed finite element
method whenever we require that Cuu + Cqq = 0 and Cqu and Cuq are non-negative. To see
this, we only have to notice that (qh , uh ) is the element of Vh × Wh , where
Vh = {v ∈ L2 (Ω) : v ∈ V(K) ∀ K ∈ Th }
Wh = {w ∈ L2 (Ω) : w ∈ W (K) ∀ K ∈ Th }

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 57

that solves the equations

ah (qh , v) + bh (uh , v) = huD , v · ni∂Ω


−bh (w, qw ) + ch (uh , w) = (f, v)Ω + hCqu uD , vi∂Ω

for all (v, w) ∈ Qh × Uh , where

ah (q, r) := (q, r)Ω + hCuq [[qh ]], [[r]]iFhi


bh (u, r) := (u, ∇ · r)Th − h{u} − Cuu · [[u]], [[r]]iFhi
ch (u, v) :=
hCqu [[u]], [[v]]iFhi + hCqu u, vi∂Ω ,
P P
where (·, ·)Th := K∈Th (·, ·)K and (·, ·)Fhi := F ∈F i (·, ·)F , where Fhi is the set of faces F of
h
the triangulation Th . Note that Cqq does not appear in the definition of these bilinear forms
because we are assuming that Cuu + Cqq = 0. Note that the method has two stabilization
terms, namely, the one associated to the jumps in the normal component of the flux, [[qh ]], and
the other asociated to the jumps of the scalar vairable, [[uh ]]. Both have important effects on
the accuracy of the method.

Indeed, some convergence properties of these methods in terms of the spaces and numerical
traces are given in the Table 6; for a proof, see Castillo et al. (2000. The so-called minimal
dissipation DG method takes Cqu = 0 and Cuq = 0 on the interior faces, and Cqu = 1/h and
Cuq = 0 on the boundary faces. For simplexes, it has the same order of convergence that the
third method in Table 6. This can be achieved by taking a special choice of Cuu and Cqq ; see
the work by Cockburn and Dong (2007.

Table 6. Orders of convergence of the DG methods in terms of the local spaces and the stabilization
parameters. We assume that Cuu and Cqq are uniformly bounded and that Cuu + Cqq = 0. The meshes
are made of shape-regular elements of arbitrary shape. The space of vector-valued functions P ℓ (K)
can be replaced by the space ∇Pℓ+1 (K) without altering the orders of covergence.

V(K) W (K) Cqu Cuq qh uh


P k (K) Pk (K) h 1/h k+1 k
P k (K) Pk (K) 1 1 k + 1/2 k+1
P k (K) Pk (K) 1/h h, 0 k k+1
P k−1 (K) Pk (K) 1/h h, 0 k k+1

5.2. DG methods allowing for an easy elimination of qh

Let us consider DG methods having a numerical trace u bh independent of qh . This allows for the
easy elimination of the variable qh , which can now be expressed in terms of uh in an elementwise
manner, and results in the so-called primal formulation of the method. The main DG methods

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

58 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

of this type are displayed in Table 7. There, αj ([[uh ]]) := Cqu [[uh ]] and αr ([[uh ]]) denotes the
stabilization introduced by Bassi et al. (1997 and later studied by Brezzi et al. (2000. The effect
of both stabilizations is essentially the same; however, that latter produces a method with a
more sparse stiffness matrix for uh . Note that the methods proposed by Cockburn and Shu
(1998a are called the local DG (LDG) methods, whereas the methods by Baker (1977; Arnold
(1982 are called the IP (IP) methods. Let us note that the so-called compact DG method
proposed by Peraire and Persson (2008 increases the sparsity of the matrix for uh of the LDG
method while maintaining the same orders of convergence. In Castillo (2010, an algorithm to
increase the sparsity of the LDG method wss introduced.

Not all DG methods were originally proposed in the form we have used to present them.
Many of them were proposed directly in the primal form

Bh (uh , v) = (f v)Ω − huD , ∇v · n + Cqu vi∂Ω

The two main examples are the IP method, proposed by Arnold (1982, see also Baker (1977
where the biharmonic problem was considered, for which we have

Bh (uh , v) = (∇uh , ∇v)Th


−h[[uh ]], {∇v}iFhi − huh , ∇v · ni∂Ω
−h{∇uh }, [[v]]iFhi − h∇uh · n, vi∂Ω
+hCqu [[uh ]], [[v]]iFhi + hCqu uh , vi∂Ω ,

and the method proposed by Baumann and Oden (1999, for which we have

Bh (uh , v) = (∇uh , ∇v)Th


−h[[uh ]], {∇v}iFhi − huh , ∇v · ni∂Ω
+h{∇uh }, [[v]]iFhi + h∇uh · n, vi∂Ω
+hCqu [[uh ]], [[v]]iFhi + hCqu uh , vi∂Ω ,

where Cqu = 0.

A theoretical study of the main DG methods introduced up to the end of last century for
the elliptic model problem is carried out in a single, unified approach by Arnold et al. (2002.
Here, the mixed formulation is not used to carry out the analysis. Instead, the variable qh
is eliminated from the equations. A primal formulation is thus obtained, which is used to
analyze the method. In Table 8, we have summarized the properties of various DG methods.
We display the properties of consistency and conservativity of the numerical fluxes, of stability
of the method, of the type of stability (the symbol αj is used to denote the stabilization
associated with the terms Cqu [[uh ]]2 ), the condition on Cqu for achieving stability, and the
corresponding rates of convergence.

Error estimates in the L∞ -norm for the IP methods have been obtained in Kanschat and
Rannacher (2002 and Chen and Chen (2004. For the other DG methods displaying consistent
and single-valued numerical traces, see Guzmán (2006. The sub-optimal convergence, for odd
polynomial approximations, in the L2 -norm of the error of the scalar approximation of the

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 59

Table 7. Some DG methods and their numerical traces. The elements K are simplexes, the spaces are
V(K) = [Pk (K)]d , W (K) = Pk (K). We also have that Cuu + Cqq = 0.

Method b
qh,K u
bh,K

Bassi and Rebay (1997a {qh } {uh }


Cockburn and Shu (1998a {qh } − Cqq [[qh ]] − Cqu [[uh ]] {uh } − Cuu · [[uh ]]
Brezzi et al. (2000 {qh } − αr ([[uh ]]) {uh }
Baker (1977,Arnold (1982 {∇uh } − Cqu [[uh ]] {uh }
Bassi et al. (1997 {∇uh } − αr ([[uh ]]) {uh }
Baumann and Oden (1999 {∇uh } {uh } − nK · [[uh ]]
Rivière et al. (1999 {∇uh } − Cqu [[uh ]] {uh } − nK · [[uh ]]
Babuška and Zlámal (1973 −Cqu [[uh ]] u h |K
Brezzi et al. (2000 −αr ([[uh ]]) u h |K
(Modified from Arnold DN, Brezzi F, Cockburn B and Marini D. SIAM Journal on
Numerical Analysis, Unified analysis of discontinuous Galerkin methods for elliptic problems,
39:1749-1779, 2001, Copyright Society for Industrial and Applied Mathematics,
Philadelphia.)

Table 8. Properties of the DG methods. The elements K are simplexes, the spaces are V(K) =
[Pk (K)]d , W (K) = Pk (K). We also have Cuu + Cqq = 0.

Method Consistency Conservativity Stability Type η0 = h Cqu H 1 L2


√ √ √
Brezzi et al. (1999 √ √ √ αr η0 > 0 hp hp+1
Cockburn and Shu (1998a √ √ √ αj η0 > 0 hp hp+1
Baker (1977,Arnold (1982 √ √ √ αj η0 > η ∗ hp hp+1
Bassi et al. (1997 √ √ αr η0 > 3 hp hp+1
Rivière et al. (1999 ×
√ αj η0 > 0 hp hp
Babuška and Zlámal (1973 × ×
√ αj η0 ≈ h−2p hp hp+1
Brezzi et al. (2000 ×
√ × αr η0 ≈ h−2p hp hp+1
Baumann and Oden (1999 (p = 1) √ × × - - × ×
Baumann and Oden (1999 (p ≥ 2) √ ×
√ × - - hp hp
Bassi and Rebay (1997a × - - [h ] [hp+1 ]
p

(Modified from Arnold DN, Brezzi F, Cockburn B and Marini D. SIAM Journal on Numerical
Analysis, Unified analysis of discontinuous Galerkin methods for elliptic problems, 39:1749-
1779, 2001, Copyright Society for Industrial and Applied Mathematics, Philadelphia.)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

60 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

DG methods proposed in Baumann and Oden (1999 and Rivière et al. (1999 was proven in
Guzmán and Rivière (2009.

Superconvergence results for the approximation inside the elements have been obtained by
Adjerid and Baccouch (2012 and for the numerical traces by Celiker and Cockburn (2007.

A numerical study of various of these methods was carried out in Castillo (2002. He found,
in particular, that qh is a much better approximation to q than ∇uh ; that the penalization
parameter Cqu must be taken as η0 /h, where h is a measure of the diameters of the elements,
to obtain a condition number of the matrix for uh of order h−2 ; and that when η0 is taken to be
too big, all the DG methods give similar approximations. The hp-version of the IP method on
general computational meshes consisting of polygonal/polyhedral elements of arbitrary shapes
was considered by Cangiani et al. (2014; the elements are also allowed to have degenerating
faces.

A posteriori error estimates for the IP method were obtained by Karakashian and Pascal
(2003; Houston et al. (2007; Ainsworth (2007; Cochez-Dhondt and Nicaise (2008; R. Lazarov
and Tomar (2009; Creusé and Nicaise (2010 for the symmetric IP methods, in Becker et al.
(2003; A. Ern and Vohralı́k (2007; Ainsworth and Rankin (2010 for the symmetric and non-
symmetric IP methods, in Bustinza et al. (2005 for the local DG methods, and in Juntunen
and Stenberg (2008 for the Bassi-Rebay DG methods. For the L2 -norm of the error in the
scalar variable, they have been obtained in Rivière and Wheeler (2003 for the non-symmetric
IP method and in Castillo (2005 for the local DG method. A unified a posteriori error analysis
of all the DG methods for second-order elliptic problems considered in Arnold et al. (2002 was
carried out in Carstensen et al. (2009. Another unified a posteriori error analysis, based on the
point of view proposed in Brezzi et al. (2006, was carried out in Lovadina and Marini (2009.

The convergence of an adaptive algorithm for the IP method was proved by Karakashian
and Pascal (2007; Hoppe et al. (2008; Bonito and Nochetto (2010.

5.3. The HDG methods

The DG methods just considered have numerical traces which allow for the elimination of the
variable qh in order to obtain a formulation of the method in terms of uh only. The HDG
methods have numerical traces which allow for the elimination of both qh and uh in order to
obtain a formulation of the method in terms of u
bh only. The size of the stiffness matrix of the
HDG methods are thus significantly smaller than those of the previously mentioned methods.
The HDG methods are also more accurate, as we are going to see. We follow closely the work
done by Cockburn (2015.

5.3.1. Definition. The HDG methods are obtained as a discrete version of the following
characterization of the exact solution. On the element K, for any given f |K and Dirichlet

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 61

boundary data u
b |∂K , we take (q, u) to be the solution of the local problem
q = ∇u, −∇ · q = f in K, u=u
b in ∂K.
Then the single-valued function ub is determined by solving the global problem obtained by
imposing the following transmission and boundary conditions:
[[q ]] = 0 on Fhi and u
b = uD in ∂Ω.

To obtain the HDG methods, we solve the local problem on each element K ∈ Th by
using a DG method. We solve the global problem by imposing the transmission and boundary
conditions weakly. So, on the element K ∈ Th , for any given f |K and Dirichlet boundary data
u
bh |∂K ,we define (qh , uh ) ∈ V(K) × W (K) as the solution of the local problem
(qh , v)K + (uh , ∇ · v)K − hb
uh , v · ni∂K = 0 ∀v ∈ V(K),
(qh , ∇w)K − hb qh · n, wi∂K = (f, w)K ∀w ∈ W (K),
where
b
q · n := qh · n − τ (uh − u
bh ) on ∂K,
where the function τ is linear. Note that here we consider u
bh |∂K to be data of this problem.
To determine it, we take the numerical trace u
bh in the space
Mh := {µ ∈ L2 (Fh ) : µF ∈ M (F ) ∀F ∈ Fh },
where M (F ) is a suitably chosen finite dimensional space, and require that it be determined
as the solution of the following global problem consisting in weakly imposed transmission and
boundary conditions:
qh ]]iFhi = hµ, b
hµ, [[b qh · ni∂Th \∂Ω = 0,
hµ, u
bh i∂Ω = hµ, uD i∂Ω ,
for all µ ∈ Mh . This completes the definition of the HDG methods.

Notice that u
bh is the data of the local problems but is the unknown of the global problem. So,
the only globally-coupled degrees of freedom are those of u bh . By solving the local problems, we
express qh , uh and bqh in terms of ubh and f . With these expressions, we construct the matrix
equation associated to the global problem. After solving it, we can insert the actual values
bh in the expressions we had obtained for qh , uh and b
of u qh . Next, we describe this procedure
more precisely.

5.3.2. The problem for u bh . We begin by introducing some notation associated to the local
problems. On the element K ∈ Th , for any µ ∈ L2 (∂K), the function (Qµ , Uµ ) ∈ V(K)×W (K)
is the solution of the local problem
(Qµ , v)K − (Uµ , ∇ · v)K + hµ, v · ni∂K = 0 ∀v ∈ V(K),
−(Qµ , ∇w)K + hQ b µ · n, wi∂K = 0 ∀w ∈ W (K),
Qb µ · n := Qµ · n + τ (Uµ − µ) on ∂K,

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

62 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

and, for any f ∈ L2 (K), the function (Qf , Uf ) ∈ V(K) × W (K) is the solution of the local
problem
(Qf , v)K − (Uf , ∇ · v)K = 0 ∀v ∈ V(K),
−(Qf , ∇w)K + hQb f · n, wi∂K = (f, w)K ∀w ∈ W (K),
b f · n := Qf · n + τ (Uf )
Q on ∂K.
With this notation, we can write that
(qh , uh ) = (Qµ , Uµ ) + (Qf , Uf ),
where ubh is the solution of the global problem. Next, we give a characterization of the problem
for the globally-coupled unknown u bh by suitably rewriting the transmission condition. We are
now ready to state the result. Therein, we are going to use the following notation:
Mh (g) := {µ ∈ Mh : hλ, µi∂Ω = hλ, gi∂Ω ∀λ ∈ Mh },
Theorem 4 (Characterization of u
bh ) Assume that
(i) τ |F is a strictly positive constant ∀ F ∈ Fh ,
(ii) ∇W (K) ⊂ V(K) ∀ K ∈ Th .
Then, the function u
bh is the element of Mh (uD ) such that
ah (b
uh , µ) = ℓh (µ) ∀ µ ∈ Mh (0),
b · ni∂T and ℓh (µ) := hµ, Q
where ah (µ, λ) := −hµ, Q b · ni∂T . Moreover,
λ h f h

ah (µ, λ) = (Qµ , Qλ )Th + hUµ − µ, τ (Uλ − λ)i∂Th , ℓh (µ)= (f, Uµ ),


and ah (·, ·) is symmetric and positive definite on Mh (0) × Mh (0). Thus, u
bh minimizes the total
energy functional Jh (µ) := 21 ah (µ, µ) − ℓh (µ) over Mh (uD ).

This result shows that the method can be implemented in a way typical of finite element
methods. Note that the equation satisfied by u bh is nothing but the transmission condition.
Indeed, since b b ub · n + Q
qh · n = Q b f · n, the transmission condition hµ, b
qh · ni∂Th \∂Ω = 0 ∀µ ∈ Mh ,
becomes
−hµ, Q b ub · ni∂T = hµ, Qb f · ni∂T ∀µ ∈ Mh (0).
h h

Note also that to obtain the matrix equation for the degrees of freedom of u b, we only need to
compute the mapping µ 7→ (Qµ , Uµ ). The computation of the mapping is not f 7→ (Qf , Uf )
required. Finally, note that the fact that the bilinear for ah (·, ·) is symmetric and positive
definite on Mh (0) × Mh (0) is a reflection that it approximates the solution u of a strangle
elliptic, self adjoint problem.

This result shows that the HDG methods are amenable to static condensation, see Guyan
(1965, thanks to the fact that the method can be hybridized, see Fraejis de Veubeke (1977.
The relation between static condensation, hybridization and the way the HDG methods are
devised is explored by Cockburn (2015. Also there, one can find different ways of rewriting the
HDG methods and how other characterizations of the exact solution can give rise to the same
DHG methods.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 63

5.3.3. The numerical traces of the HDG methods. Suppose that the stabilization function τ is
a constant on each face and that the local spaces are such that the numerical trace b qh · n|F lies
in the space M (F ) for all interior faces F ∈ Fhi . Then, the transmission condition is equivalent
to
qh ]] = 0 on Fhi .
[[b
A simple calculation shows us that the above equation can take place if and only if
τ + uh + + τ − uh − 1
u
bh = − + [[q ]],
τ+ + τ− τ + τ− h
τ − q h + + τ + qh − τ +τ −
b
qh = − [[uh ]],
τ+ + τ− τ+ + τ−
provided, of course, that τ + + τ − 6= 0. In other words, the DG methods proposed in Cockburn
and Shu (1998a and Castillo et al. (2000 with
1
Cuq = ,
τ+ + τ−
τ +τ −
Cqu = + ,
τ + τ−
τ + n+ + τ − n−
−Cqq = Cuu = .
2(τ + + τ − )
are also HDG methods. These are called LDG-H methods since the numerical method used to
define the local problems is the local DG (LDG) method.

5.3.4. Mixed methods and superconvergent HDG methods. As pointed out in Cockburn et al.
(2009b, when the stabilization function τ can be set identically to zero, we obtain nothing but
the well known hybridized version of the well known mixed methods. In this case, the above
expressions for the numerical traces are not valid anymore and, instead, we simply have that

qh · n± = q−
b ± − ±
h · n = qh · n ,

on all interelement boundaries; the unknown u


bh can be characterized exactly as in Theorem
4. This suggests that the HDG methods might share with the mixed methods some of its
convergence properties.

This was proven to be true by Cockburn et al. (2008b for a special LDG-H method obtained
by setting τ = 0 on all the faces of the simplex K except one. For this reason, it was called
the single face-hybridizable (SFH) DG method. Moreover, it was shown that the bilinear
forms ah (·, ·) of the Raviart-Thomas (RT), Brezzi-Douglas-Marini (BDM) and SFH methods
are the same, and that these three methods share the same superconvergence property (we
describe below). A similar result was later obtained by Chung et al. (2014 who proved that
the staggered discontinuous Galerkin (SDG) method, originally introduced in the framework
of wave propagation in Chung and Engquist (2009, can be obtained as the limit when the
non-zero values of the stabilization function, which must be defined in suitable manner, of the
SFH method goes to infinity.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

64 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

By the above-mentioned superconvergence property we mean that the elemenwise averages


of the error u − uh , converge faster than the errors u − uh and q − qh . As a consequence, we
can define, on the each element K, the new approximation u⋆h ∈ W ∗ (K) := Pk+1 (K) as the
solution of

(∇u⋆h , ∇w)K = − (c qh , ∇w)K for all w ∈ W ∗ (K),


(u⋆h , 1)K =(uh , 1)K ,

Then u − u⋆h will converge faster than u − uh . Any HDG method with this property will be
called a superconvergent method.

The orders of convergence, for conforming the meshes made of simplexes, are displayed in
Tables 9 and 10. The symbol ⋆ indicates that the non-zero values of the stabilization function
τ only need to be uniformly bounded by below.

Table 9. Examples of mixed and HDG methods defined on simplexes.

Method V(K) W (K) M (F )

RT Raviart and Thomas (1977 ek (K)


P k (K) + x P Pk (K) Pk (F )
SFH Cockburn et al. (2008b P k (K) Pk (K) Pk (F )
SDG Chung and Engquist (2009 P k (K) Pk (K) Pk (F )
LDG-H Cockburn et al. (2009b P k (K) Pk (K) Pk (F )
BDM Brezzi et al. (1985 P k (K) Pk−1 (K) Pk (F )

Table 10. Orders of convergence (in the corresponding L2 -norms) for simplicial, conforming meshes.

Method τ qh uh uh k
RT Arnold and Brezzi (1985 0 k+1 k+1 k+2 ≥0
SFH Cockburn et al. (2008b ⋆ k+1 k+1 k+2 ≥1
SDG Chung et al. (2014 0, ∞ k+1 k+1 k+2 ≥1
LDG-H Cockburn et al. (2010d O(1) k+1 k+1 k+2 ≥1
BDM Brezzi et al. (1985 0 k+1 k k+2 ≥2
LDG-H Castillo et al. (2000 O(1/h) k k+1 k+1 ≥1

The presence of hanging nodes in the meshes does not alter the superconvergence of the
HDG methods, as shown by Chen and Cockburn (2012; Chen and Cockburn (2014. A posteriori
error estimates were obtained by Cockburn and Zhang (2012; Cockburn and Zhang (2013. A
proof of the convergence of an adaptive algorithm for the LDG-H method was obtained by
Cockburn et al. (2016c. The performance of the LDG-H method (with τ of order one) for the
p-Laplacian was numerically explored by Cockburn and Shen (2016. The superconvergence

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 65

Figure 32. HDG approximations, on Ω := (0, 1)2 , of the p-Laplacian for p = 1.05 (top) and p = 15
with f equal to ten times the characteristic function of the square (1/4, 3/4)2 . The interelement jumps
in the piecewise linear approximation uh (left column) indicate the need of smaller meshes or higher-
degree polynomials. The interelement jumps of the piecewise quadratic postprocessing u∗h are smaller
than those of uh as the former function is usually a better approximation than the latter. Note the
ability of the method to capture at the same time very strong and very weak gradients (p = 1.05), as
well as functions displaying kinks (p = 15). Courtesy of Jiguang Shen.

properties for the case p = 2 were recovered. In Figure 5.3.4, examples of approximations uh
and the corresponding postprocessing u∗h are provided. Notice how the size of the interelement
jumps increases whenever the solution has steeper gradients, as expected.

The systematic construction of superconvergent HDG methods was undertaken in Cockburn

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

66 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

et al. (2012a where the following sufficient conditions were found. The space V(K) × W (K)
e
must have a subspace V(K) ×Wf (K) satisfying inclusions

e
P 0 (K) ⊂ ∇W (K) ⊂ V(K),
f (K),
P0 (K) ⊂ ∇ · V (K) ⊂ W
V (K) · n|∂K + W (K)|∂K ⊂ M (∂K).
and whose orthogonal complement satisfies the identity

Ve · n|∂K ⊕ W
f ⊥ |∂K = M (∂K).

Many new superconvergence HDG and mixed methods were found for simplexes, squares, cubes
and prisms; see also the related new commuting diagrams for the so-called TNT elements on
cubes obtained by Cockburn and Qiu (2014 for the DeRham complex. For curved elements,
see Cockburn et al. (2012b. This work was then further refined by Cockburn et al. (2016b
who introduced the so-called M-decompositions as a tool for the systematic construction of
superconvergent HDG and mixed methods. The actual construction for general polygonal
elements was carried out in Cockburn and Fu (2016a and the actual construction for arbitrary
pyramids, prisms and hexahedral was carried out in Cockburn and Fu (2016b.

5.3.5. Other stabilization functions. So far, we have only considered stabilization function
τ which are simple mutiplication operators. A more sophisticated stabilization function was
introduced in Lehrenfeld (2010, Remark 1.2.4 by Lehrenfeld and Schöberl, see also Lehrenfeld
and Schöberl (2015, and independently by Oikawa (2015. The stabilization function is simply
τ LS (uh − u
bh ) := h−1 · (PM (uh ) − u
bh ),
and was introduced to deal with the case in which W (K)|∂K is not included in M (∂K).
Thanks to this choice of stabilization function optimal orders of convergence for both qh
and uh for regular-shaped, general polyhedral elements can be obtained whenever we take
V(K) := P k (K), W (K) := Pk+1 (K) and M (F ) := Pk (F ); see the proof by Oikawa (2015.

The very same orders of convergence can actually be obtained with the much
smaller local spaces V(K) := ∇Pk+1 (K), W (K) := Pk (K) and M (F ) := Pk (F )
if yet another more sophisticated stabilization function is properly constructed: See
the so-called hybrid high-order (HHO) methods introduced by Di-Pietro and Ern
(2015DiPietroErnLemaire14DiPietroErnCRAS14. Although these methods were originally
introduced in a primal form, their relation to the HDG methods was recently established
in Cockburn et al. (2015.

5.4. The embedded DG methods

An embedded DG (EDG) method is obtained by simply taking an already existing HDG


method using the space Mh and requiring that its approximation u bh lie in a space embedded in
Mh . In this way, the computational complexity of the global problem can be greatly reduced.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 67

An important particular case is the case in which the above-mentioned subspace is Mh ∩ C 0 (Fh ).
Note that the EDG has the very same local problems than the associated HDG method and a
smaller matrix for the global system of the form E t AE, where A is the matrix for the global
system of the original HDG method and E is the matrix associated to the embedding.

Next consider the HDG method for which V(K) := ∇Pk (K), W (K) := Pk (K) and
M (F ) := Pk (F ) with τ = 1, where the elements are simplexes. It was shown by Cockburn et al.
(2009c that the corresponding EDG method with Mh ∩C 0 (Fh ) as the embedded subspace looses
the superconverence properties of the original HDG method. So, the reduction of computational
complexity comes at a heavy price.

The EDG methods were introduced in the framework of linear shells by Güzey et al. (2007,
see also Cockburn et al. (2009b. The method seems to be identical to the second version of
the so-called multiscale DG method introduced by Hughes et al. (2006; the first version was
introduced by Bochev et al. (2006.

How
5.5.toExtensions
couple DG methods with the classical conforming methods was shown by Alotto et al.
(2001 and Perugia and Schötzau (2001. Moreover, Perugia and Schötzau (2001 combined the
theoretical framework developed by Arnold et al. (2002 with the techniques of analysis of
nonconforming methods to obtain optimal error estimates for the resulting coupling. How to
couple DG methods with mixed methods was shown by Cockburn and Dawson (2002. The
coupling at a distance of HDG methods with BEM was carried out by Cockburn et al. (2012c.

DG methods for multiscale problems have been considered by Wang et al. (2011; Efendiev
et al. (2015.

Extensions to the approximation of eigenvalues and eigenfunctions by using the LDG-H


methods was explored by Gopalakrishnan et al. (2015; see also the corresponding work for the
RT method by Cockburn et al. (2010c.

Extensions to the heat equation are straightforward; see, in particular, the HDG methods
developed by Chabaud and Cockburn (2012. For the heat equation, the so-called direct DG
(DDG) method has been developed by Liu and Yan (2010 which has the distinctive feature
of bypassing the definition of an approximation for the flux and letting its numerical flux to
depend of the interelement jumps of second and higher-order derivatives of the approximation
to the scalar variable. The methods display an optimal order of convergence in an energy-like
norm.

A comparison between the HDG and the continuous Galerkin method has been carried out
by Kirby et al. (2012. See also the study by Huerta et al. (2013.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

68 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

5.6. Solvers

A domain decomposition preconditioner for DG approximations for purely elliptic problems


was proposed by Feng and Karakashian (2001. The condition number of their nonoverlapping
preconditioner grows linearly with the number of degrees of freedom in each subdomain. Later,
Lasser and Toselli (2000 found an overlapping domain decomposition method for DG methods
for linear advection-diffusion problems whose condition number is independent of the number
of degrees of freedom and the number of subdomains.

Another significant result has been obtained by Gopalakrishnan and Kanschat (2003b who
devised a multigrid method for solving the matrix equation of the IP method for elliptic
problems. They proved that it convergences in a fixed number of iterations; they have also
devised a method for the steady state convection-diffusion problem, which converges with a
fixed number of iterations independently of the size of the convection coefficients. These solvers
were generalized to the LDG method in primal form, the method by Bassi and Rebay (1997a,
and the method by Brezzi et al. (1999 by Gopalakrishnan and Kanschat (2003a. On the basis
of these solvers, preconditioners for the LDG saddle point systems arising from the mixed
discretization of Poisson and Stokes equations were introduced by Kanschat (2003.

A a semi-algebraic multilevel preconditioner for the local discontinuous Galerkin method


was proposed and studied in P. E. Castillo and Velázquez (2008; Castillo and Sequeira (2013.
A nonnested multigrid V-cycle, with one smoothing per level for the HDG method was
introduced in Cockburn et al. (2014 and proven to converge at a mesh-independent rate.
Domain decomposition methods on complicated domains have been investigated by Antonietta
et al. (2014.

6. The Stokes equations of incompressible fluid flow

In this section, we study the application of DG methods to the Stokes equations of


incompressible fluid flow
L − ∇u = 0 in Ω,
−ν∇ · L + ∇p = f in Ω,
∇·u = 0 in Ω,
(p, 1)Ω = 0,
u = uD , on ∂Ω,
d
where Ω is a bounded domain in R with Lipschitz boundary ∂Ω, ν is a viscosity. We
assume that ν is a constant function on Ω and that uD satisfies the compatibility condition
(uD · n, 1)∂Ω = 0.

Since the Laplace operator is applied to each of the components of the velocity, to devise
a DG method, we can simply use any DG method for the steady-state diffusion equation by
applying it to each component of the velocity. This does not present any major difficulty and

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 69

so, here the novelty resides on how to carry out the DG-discretization of the incompressibility
condition. In particular, we show that, even though the approximation of the velocity is
discontinuous, it is possible to obtain an H(div, Ω)-conforming, divergence-free approximate
velocities by means of an elementwise postprocessing. We also show how to devise DG methods
with H(div, Ω)-conforming velocity spaces. Finally, we show how to use finite dimensional
subspaces of divergence-free, H(div, Ω)-conforming without having to actually carry out the
almost impossible construction of their bases. This can only be done in a few cases, see
Thomasset (1981; Hecht (1981; Scott and Vogelius (1985.

6.1. The general form of the DG methods

6.1.1. Definition. The approximation (Lh , uh , ph ) on the element K ∈ Th is taken in the space
G(K) × V(K) × Q(K) and is defined as the solution, for all (G, v, q) ∈ G(K) × V(K) × Q(K),
of the equations
(Lh , G)K + (uh , ∇ · G)K − hbuLh , GnK i∂K = 0,
b h nK − pbh nK , vi∂K = (f , v)K ,
ν (Lh , ∇v)K − (ph , ∇ · v)K − hν L
uph · nK , qi∂K = 0.
−(uh , ∇q)K + hb
The average condition on the pressure is (ph , 1)Ω = 0. Finally, the Dirichlet boundary condition
b Lh = u
is imposed by setting u b ph = uD on ∂Ω. To complete the definition of the DG method,
we only need to define the numerical traces ν L b Lh and u
b h n − pbh n, u b ph .

6.1.2. The numerical traces. Taking advantage that the Laplacian operator acts on each
component of the velocity, the numerical traces L b Lh can be chosen by using, also in
b h n and u
a componentwise manner, any of the numerical traces for the DG method for steady-state
b ph can be picked independently.
diffusion. The numerical traces pbh n and u

6.1.3. Energy identities. Taking (G, v, q) := (νLh , uh , ph ) and adding the resulting equations,
we obtain the following local energy identity:
ν(Lh , Lh )K + ΘK = hν L b Lh i∂K − hb
b h nK , u b p i∂K + (f, uh )K ,
ph n K , u
h
where
ΘK := ΘLK + ΘpK ,
ΘL := hν(Lh − L
K b L i∂K ,
b h )nK , uh − u
h
b ph i∂K .
ΘpK := −h(ph − pbh )nK , uh − u
By adding on all the elements, we obtain the global energy identity
b h n − pbh n, u
ν(Lh , Lh )Ω + Θh = hν L b D i∂Ω + (f, uh )Ω ,
P
where Θh := K∈Th ΘK . We now see that, if the numerical traces have to be defined so that
the term Θh is positive, it can be interpreted as the energy associated to the interelement
jumps.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

70 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

6.1.4. Residuals and jumps. Let us show that the method establishes a linear relation between
the residuals in the interior or the element K,
RG := Lh − ∇uh , Rv := −ν∇ · Lh + ∇ph − f, Rq := ∇ · uh ,
and the residuals on its boundaries ∂K,
b Lh − uh ,
rG := u b h − Lh )nK − (b
rv := ν(L ph − ph )nK , uph − uh ) · nK ,
rq := (b
Simple integration by parts in the three equations defining the DG methods give us that, for
all (G, v, q) ∈ G(K) × V(K) × Q(K), we have that
(RG , G)K = hrL , GnK i∂K ,
(Rv , v)K = hrv , vi∂K ,
(rq , q)K = hrq , qi∂K .
This immediately implies that
−1/2
kPG RG kK ≤ C hK krG k∂K ,
−1/2
kPV Rv kK ≤ C hK krv k∂K ,
−1/2
kPQ Rq kK ≤ C hK krq k∂K ,
where PG , PV and PQ denote the L2 (K)-projections into the spaces G(K), V(K) and Q(K),
respectively. We thus see that, whenever
∇V(K) ⊂ G(K), ∇ · G(K) ⊂ V(K) and ∇ · V(K) ⊂ Q(K),
the quality of the approximation only depends on the residuals on the boundaries and on
f − PV f.

6.1.5. An example. Let us consider the numerical traces proposed in Cockburn et al. (2002;
they use the numerical traces used for discretizing the Laplacian in Cockburn and Shu (1998a
and Castillo et al. (2000. Le us begin with the traces associated to the discretization of the
Laplacian. Inside the domain Ω, we take
bh
L := {Lh } − [[Lh ]] ⊗ CLL − CLu [[uh ]],
bh
u := {uh } − [[uh ]]Cuu − CuL [[Lh ]],
where [[Lh ]] := L+ + − − + + − −
h n + Lh n and [[uh ]] := uh ⊗ n + uh ⊗ n , and on its boundary, we take

bh
L := Lh − CLu (uh − uD ) ⊗ n
bh
u := uD .
Now, let us consider the traces associated to the pressure and the discretization of the
divergence-free condition. On the interior faces, we take
pbh := {ph } + Dpp · [[ph ]] + Dpu [[uh,n ]],
b ph := {uh } + Duu [[uh,n ]] + Dup · [[ph ]],
u

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 71

where [[uh,n ]] := u+ + − − + + − −
h · n + uh · n and [[ph ]] := ph n + ph n , and on the boundary,

pbh := ph + Dpu (uh − uD ) · n,


b ph := uD .
u
In this case, a simple computation gives us that the global energy identity is
b h n − pbh n, uD i∂Ω + (f, uh )Ω ,
ν(Lh , Lh )Ω + Θh = hν L
where
Θh = hCuL [[Lh ]], [[Lh ]]iFhi + hCLu [[uh ]], [[uh ]]iFhi + hCLu (uh − uD ), uh − uD i∂Ω
+ hDup [[ph ]], [[ph ]]iFhi + hDpu [[uh,n ]], [[uh,n ]]iFhi + hDpu (uh − uD ), uh − uD i∂Ω ,
whenever CLL + Cuu = 0 and Dpp + Duu = 0. We see that CuL , CLu , Cup and Cpu stabilize
the method.

Let us use the above global energy identity to obtain the existence and uniqueness of the
approximate solution in the case in which
(i) CuL , Dpu ≥ 0 on Fhi ,
(ii) CLu , Dup > 0 on Fhi ∪ ∂Ω,
(iii) ∇V(K) ⊂ G(K) ∀K ∈ Th ,
(iv) ∇Q(K) ⊂ V(K) ∀K ∈ Th .
Again, we only have to show that, when we set the data f and uD to zero, the only solution
is the trivial one. But in this case, the above energy identity gives that, by (i) and (ii), that
Lh = 0 on Ω, that [[uh ]] = [[ph ]] = 0 on Fhi , and that uh = 0 on ∂Ω. As a consequence, we have
bL
that u i
h = uh on Fh ∪ ∂Ω, and the first equation defining the DG method reads

−(∇uh , G)K = 0
for all G ∈ G(K). By (iii) we can take G := ∇uh and conclude that uh is a constant vector
on the element K. Since uh = u b h on the interelement boundaries, uh is a vector constant on
Ω, and since uh = 0 on ∂Ω, we get that uh = 0 on Ω.

We still need to show that the approximate pressure is zero. Taking into account that the
interelement jumps of ph are zero, the second equation defining the DG method reads
(∇ph , v)K = 0,
for all v ∈ V(K). By property (iv), we can take v := ∇ph to conclude that ph is a constant on
the element K. Since ph is continuous on Ω and has average equal to zero, we conclude that
ph is also zero. This completes the proof.

6.2. DG methods allowing for an easy elimination of Lh

If we discretize each of the components of the Laplacian by using DG methods for the Laplacian
allowing for an easy elimination of the variable qh , we immediately obtain a DG method for

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

72 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

which the velocity gradient Lh can be easily eliminated. As a consequence, we obtain methods
expressed in terms of the velocity uh and the pressure ph only. Examples of these methods
are the very first DG method proposed by Baker et al. (1990, which uses the IP method to
approximate the Laplacian, elementwise solenoidal velocities and a continuous approximation
of the pressure, the methods proposed by Cockburn et al. (2002 (for arbitrarily -shaped
elements) and Schötzau et al. (2003a (for square and cubic elements), which uses the LDG
method, and the method proposed by Girault et al. (2005 (for triangular elements), which
considers both the IP and the Baumann-Oden methods.

The convergence properties of these methods are as follows. When the velocities contain
the piecewise polynomials of degree k and the pressure (and velocity gradient) are piecewise
polynomials of degree k − 1, both the velocity and the pressure converge optimally (in
particular, with order k + 1 and k in L2 -norms) for any k ≥ 1; for Girault et al. (2005 the
estimates hold for p = 1, 2, 3. The p-version of the methods is treated in Schötzau et al. (2003a
where optimal convergence properties are proven.

6.3. The HDG methods

The DG methods we just considered allow for an easy elimination of the approximation Lh
which gives rise to formulations in terms of uh and ph . Here, we consider DG methods which
b h and the
allow for the elimination of Lh , uh and ph which results in formulations in terms of u
average of the pressure on each element ph only. This new formulation maintains the original
saddle- point structure of the problem for uh and ph , but has remarkably less globally-coupled
degrees of freedom, which provides an efficient implementation of the method. We follow
Cockburn and Shi (2014.

6.3.1. Definition. The HDG methods are obtained as a discrete version of the following
characterization of the exact solution. On the element k ∈ Th , given f |K , the constant p and
the Dirichlet boundary data u b |∂K , we take (L, u, p) to be the solution of the following local
problem:
1
L = ∇u, −ν∇ · L + ∇p = f, ∇ · u = hb
u · nK , 1i∂K in K,
|K|
1
(p, 1)K = p, u b = uD on ∂K,
|K|
Then, the piecewise-constant function p and the single-valued velocity u b is determined as
the solution of the global problem consisting the following transmission, divergence-free and
boundary and average conditions:
[[νL − pId]] = 0 on Fhi , hb
u · nK , 1i∂K = 0 ∀K ∈ Th , b=u
u bD on ∂Ω, (p, 1)Ω = 0.
+ + − − + + − −
Here [[νL − pId]] := ν(L n − L n ) − (p n + p n ), with the obvious notation.

Let us note that the velocity u solving the local problems is not necessarily divergence-free
because we do not want to assume that the Dirichlet data u b |∂K is such that hbu · n, 1iK = 0.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 73

We want these local problems to be solvable for any values of u b |∂K . For this reason, we need
to introduce the second equation in the the definition of the global problem.

To obtain the HDG methods, we solve the local problem on each element K ∈ Th by
using a DG method. We solve the global problem by imposing the transmission, divergence-
free, boundary and beverage conditions weakly. So, for any given arbitrary function u b h |∂K
and constant ph , we define (Lh , uh , ph ) ∈ G(K) × V(K) × Q(K) as the solution of the local
problem for all (G, v, q) ∈ G(K) × V(K) × Q(K),
(Lh , G)K + (uh , ∇ · G)K − hbuh , Gni∂K = 0,
b
ν (Lh , ∇v)K − (ph , ∇ · v)K − hν Lh n − pbh n, vi∂K = (f , v)K ,
−(uh , ∇q)K + hb
uh · n, qi∂K = hbuh · n, qi∂K ,
(ph , 1)K = (ph , 1)K ,
b h n − pbh n := νLh n − ph n − S(uh − u
νL bh) on ∂K.
uh , ph ) in the space M h × Q0h , where,
We take the function (b
M h := {µ ∈ L2 (Fh ) : µ|F ∈ M (F ) ∀ F ∈ Fh },
Q0h 2
:= {q ∈ L (Th ) : q|K is a constant ∀ K ∈ Th },
where the local space M (F ) is a general finite dimensional space, and determine it by requiring
that, for all (µ, q) ∈ M h × Q0h ,
b h n + pbh n, µi∂T \∂Ω = 0,
h−ν L h

uh · n, qi∂Th = 0,
hb
hb
uh , µi∂Ω = hg, µi∂Ω ,
(ph , 1)Ω = 0.

uh , ph ) is data of the local problem but the unknown of the global problem.
Note, again, that (b
We need to solve the global problem in order to get the actual values of (b uh , ph ). However,
we first need to solve the local problems to be able to construct the matrix equations of the
global problem. We show how to do that next.

uh , ph ). We start by introducing notation related to the local


6.3.2. The problem for (b
problems. On the element K ∈ Th , for any µ ∈ L2 (∂K), we define (Lµ , Uµ , Pµ ) ∈
G(K) × V(K) × Q(K) as the solution, for all (G, v, q) ∈ G(K) × V(K) × Q(K), of the
equations
(Lµ , G)K + (Uµ , ∇ · G)K − hµ, Gni∂K = 0,
b µ n, vi∂K = 0,
b µn − P
ν (Lµ , ∇v)K − (Pµ , ∇ · v)K − hν L
−(Uµ , ∇q)K + hµ · n, qi∂K = hµ · n, qi∂K ,
(Pµ , 1)K = 0,
νL b µ n := νLµ n − Pµ n − S(Uµ − µ)
b µn − P on ∂K.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

74 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Similarly, for any f ∈ L2 (K), we define (Lf , Uf , Pf ) ∈ G(K) × V(K) × Q(K) as the solution,
for all (G, v, q) ∈ G(K) × V(K) × Q(K), the equations
(Lf , G)K + (Uf , ∇ · G)K = 0,
ν (Lf , ∇v)K − (Pf , ∇ · v)K − hν L b f n, vi∂K = (f , v)K ,
b fn − P
−(Uf , ∇q)K = 0,
(Pf , 1)K = 0,
νL b f n := νLf n − Pf n − S(Uf )
bfn − P on ∂K.

With this notation, we can write that


(Lh , uh , ph ) = (Lbuh , Ubuh , Pbuh ) + (Lf , Uf , Pf ) + (0, 0, ph ),
where (buh , ph ) is the solution of the global problem. The next result, by Nguyen et al. (2010b,
gives a characterization of this function. Therein, we use the following notation:
M h (g) := {µ ∈ M h : hµ, λi∂Ω = hg, λi∂Ω ∀λ ∈ M h }.
Theorem 5 (Characterization of (b
uh , ph )) Assume that
(i) S|F is symmetric and uniformly positive definite ∀F ∈ Fh ,
(ii) ∇V(K) ⊂ G(K) ∀K ∈ Th ,
(iii) ∇Q(K) ⊂ V(K) ∀K ∈ Th .
uh , ph ) is the element in Mh (uD ) × Q0h satisfying
The function (b
uh , µ) + bh (µ, ph ) =(f, uµ
ah (b h )T h ∀µ ∈ Mh (0),
−bh (b
uh , q) =0 ∀q ∈ Q0h ,
(ph , 1)Ω =0,
b µ n − pbµ n, λi∂T and bh (λ, q) := −hq, λ · ni∂T . Moreover,
where ah (λ, µ) := hν L h h h h

µ µ
ah (λ, µ) = ν (Lλ λ
h , Lh )Th + hS(uh − λ), (uh − µ)i∂Th ,

for all λ, µ in M h , and ah (·, ·) is symmetric and positive definite in M h (0) × M h (0).

This implies that, as claimed, the HDG methods can be easily implemented. Indeed, we see
that the only globally coupled degrees of freedom are those of the approximation of the velocity
b h , and those of the elementwise average of the pressure ph . This global problem can
on Fh , u
be solved by using, for example, the augmented Lagrangian method, see Nguyen et al. (2010b.
Once this global problem is solved, the approximate solution (Lh , uh , ph ) can be easily obtained
in an element-by-element fashion by using the very first identity of this Subsection.

b h as the only minimum of the functional


This also implies that we can see u
1
Jh (λ) := ah (λ, λ) − (f, uλ
h )Th ,
2
Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 75

on the space D := {µ ∈ M h (g) : bh (λ, q) = 0 ∀ q ∈ Qh }. Note that a related method


can be obtained by dropping the restriction that λ lie on the space D, by requiring that the
approximate velocities be divergence-free on each element and by penalizing their interelement
normal jump. This was originally proposed by Hansbo and Larson (2008 for a DG method
defined by means of an IP discretization. This approach was then applied to a DG method
proposed in Montlaur et al. (2008 which uses globally divergence-free velocities. The resulting
method turned out to be identical to the DG method proposed by Hansbo and Larson (2008.

6.3.3. The numerical traces. Let us give an idea of the explicit form of the numerical traces
of the HDG methods in the simple case in which the stabilization tensor is just a constant
time the identity, that is, S := τ Id where τ is taken to be constant on each face. Then, when
the local spaces are taken in such a way that the transmission condition implies
b h n + pbh n]] = 0
[[−ν L on Fhi ,
a simple computation gives us that the numerical traces must given by the following formulas:
τ + uh + + τ − uh − 1
bh =
u + + [[−νLh + ph Id]],
τ+ + τ− τ + τ−
b h + pbh Id = τ−
−ν L (−νL+ +
h + ph Id)
τ+ + τ−
τ+ τ +τ −
+ + −
(−νL− −
h + ph Id) + + [[uh ]].
τ +τ τ + τ−
Note that for this HDG method, we have that u b Lh = u
b ph = u
b h and so, it is impossible to relate it
to any of the DG methods previously considered. We can, however, compare the stabilization
term which in this case is
X
Θh = b h − pbh Id))nK , uh − u
h((νLh − ph Id) − (ν L b h i∂K
K∈Th
X
= b h ), uh − u
hτ (uh − u b h i∂K
K∈Th
1
=h [[νLh − ph Id]], [[νLh − ph Id]]iFhi
τ++ τ−
τ +τ −
+h + [[uh ]], [[uh ]]iFhi + hτ (uh − uD ), (uh − uD )i∂Ω .
τ + τ−
We thus see that the stabilization of the HDG methods is through the interelement jumps of
uh and those of the normal component of νLh − ph Id.

6.3.4. Convergence properties. Consider meshes Th is a regular-shaped elements polygonal


or polyhedral element K. For the HDG methods with the local spaces
G(K) = Pk (K), V (K) = P k (K), Q(K) = Pk (K), M (F ) = P k (F ).
and the stabilization function S = τ Id, it was proven in Cockburn and Shi (2014 that, when τ
is a positive constant, the order of convergence for the velocity is k+1 but those for the pressure

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

76 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

and the velocity gradient are only k + 1/2 (The same orders hold if we take G(K) = ∇P k (K)).
Whenever τ = 1/h, then the order of convergence for the velocity is k + 1 but those for the
pressure and the velocity gradient are only k. The same orders hold if we take G(K) = Pk−1 (K)
and Q(K) = Pk−1 (K). In this case, the orders are optimal.

6.3.5. Superconvergent HDG methods. It is possible to devise superconvergent HDG


methods for the Stokes flow in terms superconvergent HDG methods for the model diffusion
problem previously considered. Suppose that the local spaces of the latter methods are
V D (K), W D (K) and M D (F ), and that their stabilization function is τ D . We construct a
superconvergent HDG method as follows.

Denote by Gi (K) the space of all the i-th rows of functions in G(K), and by V i (K) and
M i (F ) the space of the i-th component of functions in V (K) and M (F ), respectively, for
i = 1, . . . , d. Then, we take the local spaces as

Gi (K) := V D (K), V i (K) := W D (K), M i (F ) := M D (F ), (8a)

for i = 1, . . . , d, and the stabilization function as

S := τ D Id. (8b)

The choice of the space for the pressure Q(K) has to be done in such a way that

d
X
∂j W D (K) ⊂ Q(K) ⊂ ∩dj=1 {vj : v ∈ V D (K) : vi = 0 for i 6= j}. (8c)
j=1

Examples of such methods are displayed in Table 11, taken from Cockburn et al. (2012a. We
display the orders of convergence for mixed methods (τ D = 0) and HDG methods (τ D = 1)
using different elements K. We only show the space for the pressure Q(K); the the other spaces
are explicitly given in Cockburn et al. (2012a. The new approximation of the velocity u∗h on
the element K ∈ Th is defined, see Gastaldi and Nochetto (1989; Stenberg (1988; Stenberg
(1991, as the element of finite dimensional space V ∗ (K) such that

(∇u∗h , ∇v)K = (Lh , ∇v)K ∀v ∈ V ∗ (K), (9a)


(u∗h , v)K = (uh , v)K ∀ ∈ P 0 (K). (9b)

In the examples in Table 11, it is enough to take V ∗h (K) ⊃ P k+1 (K) for all elements K ∈ Th .

The application of the theory of M-decompositions to devise superconvergent HDG and


mixed methods has been developed in Cockburn et al. (2016a. Examples of other methods are
the staggered DG method obtained by Kim et al. (2013, which is related to the SFH method,
as pointed out by Chung et al. (2016, and the method proposed by Oikawa (2016, which is
related to the reduced-order HDG method analyzed in Oikawa (2015.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 77

Table 11. Order of convergence for superconvergent HDG methods (k ≥ 1)


Q(K) kL − Lh kL2 (Ω) kp − ph kL2 (Ω) ku − u⋆h kL2 (Ω)

K simplex and M (F ) = P k (F )
BDFMk+1 Pk (K) k+1 k+1 k+2
RTk Pk (K) k+1 k+1 k+2
HDGk Pk (K) k+1 k+1 k+2
BDMk Pk (K) k+1 k+1 k+2
k≥2

K square or cube and M (F ) = P k (F )


BDFM[k+1] Pk (K) k+1 k+1 k+2
HDGP[k] Pk (K) k+1 k+1 k+2
BDM[k] Pk (K) k+1 k+1 k+2
k≥2

K square or cube and M (F ) = Qk (F )


RT[k] Qk (K) k+1 k+1 k+2
TNT[k] Qk (K) k+1 k+1 k+2
HDGQ [k]
Qk (K) k+1 k+1 k+2

6.4. Obtaining divergence-free approximate velocities

6.4.1. Using H(div, Ω)-conforming velocity spaces. A very simple way to obtain divergence-
free approximate velocities, see Cockburn et al. (2005; Cockburn et al. (2007, is to pick the
velocity space in such a way that

(i) ∇ · V(K) ⊂ Q(K) ∀ K ∈ Th ,


(ii) V(K) ⊂ H(div, Ω),

and set
b h · nK := uh · nK
u on ∂K ∀ K ∈ Th .

Indeed, in the case, the third equation defining the DG method reads, after a simple integration
by parts,
(∇ · uh , q)K = 0

for all q ∈ Q(K). By property (i) and the choice of the normal component of the numerical
trace, this implies that we can take q := ∇ · uh to conclude that ∇ · uh = 0 on the element K.
By property (ii), this implies that uh is a divergence-free function in H(div, Ω).

In Cockburn and Sayas (2014, a new approach was proposed to devising HDG methods with
H(div)-conforming velocity spaces which are also superconvergent. It consists in considering
the superconvergent HDGk method in Table 11 with the following stabilization function

S := ν τn n ⊗ n + ν τt (Id − n ⊗ n),

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

78 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

where n is the normal to the faces on ∂Th , and let τn go to infinity. In Cockburn and Sayas
(2014, it was proven that, when τn goes to infinity, we obtain a well defined method with the
very same convergence and superconvegence properties of the original method HDGk .

6.4.2. Divergence-free velocities by local postprocessing. An alternative to the previous


approach is to postprocess the approximation to obtain the wanted H(div, Ω)-conforming,
divergence-free approximate velocity; see Bastian and Rivière (2003; Cockburn et al. (2005;
A. Ern and Vohralı́k (2007. For example, let us assume that all the elements K are tetrahedra.
Then, on the tetrahedron K ∈ Th , the postprocessed velocity u∗h is defined as the element of
P k (K) such that
h(u∗h − u
b h ) · n, µiF = 0 ∀ µ ∈ Pk−1 (F ),

for all faces F of K, and such that

(u∗h − uh , ∇w)K = 0 ∀ w ∈ Pk−1 (K),


(∇ × u∗h − wh , (∇ × v) BK )K = 0 ∀ v ∈ S k−1 (K)
where BK is the so-called symmetric bubble matrix introduced in Cockburn et al. (2010b,
namely,
X3
BK := λℓ−3 λℓ−2 λℓ−1 ∇λℓ ⊗ ∇λℓ ,
ℓ=0

where λi are the barycentric coordinates associated


Pk−1 with the tetrahedron K, the subindices
being counted modulo 4. Finally, S k−1 (K) := ℓ=1 S ℓ (K) where S ℓ is the space of vector-
valued homogeneous polynomials v of degree ℓ such that v · x = 0, see Nédélec (1980; Nédélec
(1986.

The fact that u∗h is well defined follows from the projection of the so-called BDM method. The
fact that u∗h lies in H(div, Ω) is a consequence of the first equation defining the postprocessing
and from the fact that u b h is single valued. Finally, the fact that the divergence of u∗h is zero
follows from the third equation defining the approximation. Indeed, we have, for all Q ∈ Q(K),
(∇ · u∗h , q)K = − (u∗h , ∇q)K + hu∗h · nK , qi∂K
= − (uh , ∇q)K + hbuh · nK , qi∂K
= 0.
It is not difficult to show that this new approximation to the velocity has the same convergence
properties than the original approximation, that is,
ku − u∗h kL2 (Ω) ≤ C hk+1 ,
when the exact solution is smooth enough and whenever k ≥ 1.

6.4.3. Divergence-free superconvergent velocities by local postprocessing. The following


postprocessing was introduced in Cockburn et al. (2011 and, in three-space dimensions, is

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 79

defined as follows. On the tetrahedron K ∈ Th , u∗h is defined as the element of P k+1 (K) such
that
h(u∗h − u
b h ) · n, µiF = 0 ∀ µ ∈ Pk (F ), (10a)
h(n × ∇)(u∗h · n) − n × ({Lth }n), (n × ∇)µiF = 0 ∀ µ ∈ Pk+1 (F ) ,⊥
(10b)
for all faces F of K, and such that
(u∗h − uh , ∇w)K = 0 ∀ w ∈ Pk (K), (10c)
(∇ × u∗h − wh , (∇ × v) BK )K = 0 ∀ v ∈ S k (K). (10d)
Here
Pk+1 (F )⊥ := {µ ∈ Pk+1 (F ) : hµ, µ
eiF = 0, ∀e
µ ∈ Pk (F )},
the operator n×∇ is the tangential gradient and the function {Lth } is the single-valued function
on Eh equal to ((Lth )+ + (Lth )− )/2 on the set Eh \ ∂Ω and equal to Lth on ∂Ω. Moreover,
wh := (L32 h − L23 h , L13 h − L31 h , L21 h − L12 h )
is the approximation to the vorticity. This elementwise post-processing has many properties
we gather in the following result.

The fact that u∗h is well defined and that it is a divergence-free function in H(div, Ω) can be
proven exactly as in the previous case. In Cockburn et al. (2011, it has been shown that we
also have
ku − u∗h kL2 (Ω) ≤ C hk+2 ,
when the exact solution is smooth enough and whenever k ≥ 1.

6.4.4. Using H(div, Ω)-conforming divergence-free velocity spaces. It is very well known that
the construction of finite dimensional spaces of divergence-free velocities functions is, in
practice, impossible due to the many inter-element continuity constraints that need to be
imposed; see Thomasset (1981; Hecht (1981; Scott and Vogelius (1985. This difficulty can
completely be bypassed by a simple technique called hybridization. Roughly speaking, it
consists two steps. In the first step, we remove the continuity constraints of the interelement
normal component of the space of velocities. In this way, the new space of velocities is
completely discontinuous across elements. In the second step, we restore the above interelement
constraints, but only for the approximation of the velocity. In this manner, only locally
divergence-free approximations need to be constructed. The technique was introduced in
Carrero et al. (2006 for DG methods and then in Cockburn and Gopalakrishnan (2005a;
Cockburn and Gopalakrishnan (2005b for a classical mixed method. See also the reviews
Cockburn and Gopalakrishnan (2005c; Cockburn (2009. This hybridization technique coincides
with the way in which the HDG methods are devised. Thus, HDG methods using divergence-
free, globally divergence-free velocities are very easy to define.

6.5. Extensions

Although we have used a velocity gradient-velocity-pressure formulation of the Stokes system,


we could have used others using, for example, the symmetric gradient of the velocity or the

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

80 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

vorticity instead of the velocity gradient. Numerical experiments carried out in Nguyen et al.
(2010a for the HDG method show that the formulation using the velocity gradient is superior
to that of the symmetric gradient which in turn is better than that of the vorticity; see
also Cockburn and Cui (2012a; Cockburn and Cui (2012b. The use of the velocity gradient
formulation does not allow in a natural way the imposition of the normal stress as a boundary
condition. This problem has been partially addressed in Nguyen et al. (2011, but although the
optimal convergence of all the variables was retained, the superconvergence of the velocity is
lost.

In defining the HDG methods, we have used as data of the local problems the velocity at the
boundary and the average on the pressure in the element. However, this is certainly not the
only way to define the local problems. For example, an HDG method based on the vorticty
formulation was studied in Cockburn and Gopalakrishnan (2009. The very same method can
be obtained in four different ways according to what are the data of the local problems. Each
of these ways can be thought as a different way of implementing the method.

7. Convection-dominated Problems

In this section, we consider the application of the DG method to various problems in fluid
dynamics in which convection plays a dominant role. We start with the convection-diffusion
and the shallow water equations. We then consider the equations of incompressible and
compressible fluid flow. To obtain the DG methods, we simply have to combine the DG
discretization techniques for the equations for the corresponding purely hyperbolic problems
with those of the purely elliptic equations. Since the discretizations are straightforward, we
only discuss important features of the discretization not previously considered.

7.1. Convection-diffusion problems

In this section, we consider the LDG methods for the following convection-diffusion model
problem
ut + ∇ · ( f(u) − a(u)∇u) = 0 in Ω × (0, T )
u(x, 0) = u0 (x) ∀x ∈ Ω
To define a DG method, we first notice that, since the matrix a(u) is assumed to be symmetric
and semipositive definite, there exists a symmetric matrix b(u) such that a = b2 . This allows
us to introduce the auxiliary variable q = b ∇u, and rewrite the model problem as follows:
ut + ∇ · f (u) − ∇ · (b(u)q) = 0 in Ω × (0, T )
qi = ∇ · gi (u) in Ω × (0, T ), 1 ≤ i ≤ N
u(x, 0) = u0 (x) ∀ x ∈ Ω
where
R u qi is the i−th component of the vector q, and gi (u) is the vector whose jth component
is bji (s)ds. A DG method is now obtained in a most straightforward way. For details, see

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 81

Cockburn and Shu (1998a and Cockburn and Dawson (2000.

Let us give a computational result of the application of the LDG method to the two-
dimensional flow and transport in shallow water from the paper by Aizinger and Dawson
(2003; see also Dawson and Proft (2002; Dawson and Proft (2003; Dawson and Proft (2004.
The system of shallow water equations can be written as

ct + ∇ · (A + (D∇)c) = h(c)

where ct = (ξ, u, v); here, ξ is the deflection of the air-water interface from the mean sea level,
and (u, v) is the depth-averaged horizontal velocity. For details about the remaining terms, see
Aizinger and Dawson (2003. What is relevant for our purposes is that the above is a nonlinear
convection-diffusion-reaction equation which can be easily discretized by the LDG method.
In Figure 33, we see a mesh (top) of 14 269 triangles, highly graded towards the coast, and
the function ξ (bottom) computed for a high inflow of 35 000 m3 s−1 for the Mississippi River;
an open sea boundary condition is assumed. In comparison with the no-inflow situation (not
shown here), the elevation increases about half a foot near the lower Louisiana coast.

A posteriori error estimates have been obtained, for example, by Baccouch and Adjerid
(2015. Error estimates of LDG methods using implicit-explicit time-marching schemes was
carried out by Wang et al. (2015; Wang et al. (2016.

The definition of HDG methods for steady-state convection-diffusion problems is


straightforward. For details, see Cockburn et al. (2009a and Nguyen et al. (2009a, where
linear convection was considered, and Nguyen et al. (2009b, where nonlinear convection
was treated. When the diffusion dominates, the convergence properties of the methods are
those of the purely diffusive case whereas when the convection dominates. In particular, the
superconvergence of scalar variable can be recovered. When convection dominates, the method
behaves like the original DG method as shown by Fu et al. (2015. See also Egger and Schöberl
(2010, where a mixed method is used to approximate the second-order elliptic term. An space-
time HDG method for the advection-diffusion equation on moving and deforming meshes was
proposed by Rhebergen and Cockburn (2013.

7.2. Oseen flow

Next, let us consider the Oseen equations of incompressible fluid flow, namely,

−ν∆u + (w · ∇)u + γ u + ∇p = f in Ω
∇·u = 0 in Ω
u = g on Γ

where u is the velocity, p the pressure, f ∈ L2 (Ω)2 a prescribed external body force, ν > 0
the kinematic viscosity, w a convective velocity field and γ a given scalar function. As usual,
we take Ω to be a bounded domain of R2 with boundary Γ = ∂Ω, and the Dirichlet datum

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

82 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

3.5 × 10+06 Louisiana


Mississippi

Land
Y 3.0 × 10+06

Open sea
2.5 × 10+06 Land
Open sea
Land

2.0 × 10+06
0.0 × 10+00 5.0 × 10+05 1.0 × 10+06
(a) X

ETA
3.5 × 10+06 0.25
0.225
0.2
0.175
0.15
0.125
0.1
0.075
0.05
3.0 × 10+06 0.025
0
−0.025
−0.05
Y

2.5 × 10+06

2.0 × 10+06
0.0 × 10+00 5.0 × 10+05 1.0 × 10+06
(b) X

Figure 33. Gulf of Mexico mesh (a) and surface elevation for high inflow of the Mississippi river (b).
(From Aizinger V and Dawson CN. A Discontinuous Galerkin Method for Two-Dimensional Flow and
Transport in Shallow Water. Technical Report 03-16, TICAM, 2003.)

R
g ∈ H 1/2 (Γ)2 to satisfy the compatibility condition Γ
g · nds = 0, where n denotes the unit
outward normal vector to Γ. We also assume that
γ(x) − 12 ∇ · w(x) =: γ0 (x) ≥ 0, x∈Ω (11)
1 2 2
This condition guarantees the existence and uniqueness of a solution (u,
R p) ∈ Hg (Ω) × L0 (Ω)
1 2 1 2 2 2
where Hg (Ω) := {u ∈ H (Ω) : u|Γ = g} and L0 (Ω) := {p ∈ L (Ω) : Ω pdx = 0}.

In Figure 34, we display the norms of the error in the velocity and the pressure for the
LDG method as a function of the mesh size for several Reynolds numbers for the so-called
Kovasznay flow. Bi-quadratic approximations on squares are used. The norms are scaled with

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 83

1e+ 01

1e+ 00

1e− 01

ν1/2||eu||L2 1e− 02

1e− 03
Re = 1
Re = 2
1e− 04 Re = 5
Re = 10
Re = 20
1e− 05 Re = 50
Re = 100
1e− 06 Re = 200
Re = 500
Re = 1000
1e− 07
2 3 4 5 6 7
Refinement level

1e+ 02

1e+ 01

1e+ 00

1e− 01

1e− 02
ν−1/2||ep||L2

1e− 03 Re = 1
Re = 2
Re = 5
1e− 04 Re = 10
Re = 20
1e− 05 Re = 50
Re = 100
1e− 06 Re = 200
Re = 500
Re = 1000
1e− 07
2 3 4 5 6 7
Refinement level

Figure 34. Scaled L2 -errors in u and p with bilinear approximations for different Reynolds numbers.
(From Cockburn B, Kanschat G and Schötzau D. Local discontinuous Galerkin methods for the Oseen
equations. Math. Comput. 2004; 73:569-593.)

the appropriate powers of ν so as the make all the quantities dimensionally equivalent - see
Cockburn et al. (2004 for details. We can see that the convergence of the above errors is not
altered as the Reynolds number varies from 1 to 1000 which confirms the expected robustness
of the LDG method with respect to an increase in the strength of the convection.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

84 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

7.3. The Navier-Stokes of equations incompressible flow

Here, we consider DG methods for the stationary incompressible Navier-Stokes equations


L − ∇u = 0 in Ω,
−ν∇ · L + ∇ · (u ⊗ u) + ∇p = f in Ω,
∇·u=0 in Ω,
(p, 1)Ω = 0,
u = uD on ∂Ω,
where huD · n, 1i∂Ω = 0. Since the only difference with the Oseen equations is the nonlinear
convective term, we restrict our discussion to the DG discretization of such a term. The
emphasis will be placed on how to obtain (provable) H1 -bounded DG methods which are also
locally conservative.

7.3.1. H1 -boundedness, local conservativity and incompressibility. Témam (1966; Témam


(1968, see also Témam (1979, proposed a way to discretize the nonlinear convective term
of the Navier-Stokes equations, ∇ · (u ⊗ u). It consists in replacing the convective nonlinear
term by the term
1
∇ · (u ⊗ u) − (∇ · u) u.
2
This approach became the approach of choice in the finite element literature because it allowed
to obtain the H 1 -boundedness of the approximate velocity without requiring that it be exactly
divergence free. The first DG method for the Navier-Stokes equations proposed by Karakashian
and Jureidini (1998, which also uses divergence-free polynomial approximations of the velocity
inside each element, and the method proposed by Girault et al. (2005 use this classic approach.
The only drawback is that, unlike all other DG methods, these methods cannot achieve local
conservativity because the above nonlinear term does not have divergence form.

As pointed out by Cockburn et al. (2005, this difficulty can be overcome by simply replacing
the pressure p by the new unknown P := p − 12 |u|2 . Indeed, in this case the nonlinear term,
namely,
1
∇ · (u ⊗ u) + ∇|u|2
2
is in divergence form. Hence, locally conservative (and H1 -bounded) DG methods can be
obtained. The drawback is, however, that the new unknown P does not have a clear physical
interpretation.

Yet another possibility to achieve H1 -boundedness and local conservativity is to rewrite the
Navier-Stokes equations as the Oseen problem
−ν∆u + (w · ∇) u + ∇p = f in Ω,
∇·u=0 in Ω,
u=0 on Γ,

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 85

where w = u. It is known that, for f small enough, the sequence sequence {un }n∈N , where
un+1 is the solution of the above Oseen problem with w = un . The limit is thus a solution of
the Navier-Stokes equations. Thus, at the discrete level, the idea is to use an approximation
for u and another for w which should be globally divergence free. This can be achieved by an
elementwise postprocessing as pointed out in the previous section. The resulting method is
both H1 -bounded and local conservative.

7.3.2. Examples. DG methods for the Navier-Stokes can be obtained by combining a DG


method for the Stokes system with an approximation of the nonlinear convective terms. The
corresponding convergence properties are usually identical. Recall that to define a DG method
for the Stokes system, it is enough to choose a DG method for a scalar second-order elliptic
problem (to discretize the terms associated to the viscosity) and to choose to weakly or strongly
impose the divergence-free condition. Let us give some examples:

• The DG method by Karakashian and Jureidini (1998; Karakashian and Katsaounis (2000
extend the IP method DG method by Baker et al. (1990 using locally divergence-free velocities.
It approximates the nonlinear terms by using Témam’s approach. Similarly, the main DG
method proposed by Girault et al. (2005 extends the IP method and also approximates the
nonlinear terms by using Témam’s approach.

• The method by Crivellini et al. (2013 is an extension of the DG method by Bassi et al.
(2006.

• The hybrid method by Montlaur et al. (2010 extends the hybrid method by Montlaur et al.
(2008; since it uses exactly divergence-free velocities, it can discretize directly the nonlinear
term ∇ · (u ⊗ u).

• The LDG methods by Cockburn et al. (2005; Cockburn et al. (2007; Cockburn et al. (2009d
extend the LDG methods by Cockburn et al. (2002 and use the last of the above-mentioned
approaches to discretize the nonlinearity. Similarly, the LDG method by Schötzau et al. (2003a
extends the LDG method in Schötzau et al. (2003b.

• The so-called Galerkin interface stabilisation (GIS) method Labeur and Wells (2007;
Labeur and Wells (2012 does not quite fit the general form of the HDG methods considered
here. It is is, in part, an extension of the multiscale DG method by Bochev et al. (2006;
Hughes et al. (2006; Buffa et al. (2006. It approximates the nonlinear convective term by
directly discretizing the term
∇ · (u ⊗ u) − (1 − χ)(∇ · u) u,
where χ ∈ [0, 1]. Interestingly enough, the method is actually H1 -bounded even though its
approximate velocities are not exactly divergence free. If the trace spaces are continuous, the
numerical traces of the method are no longer single-valued, as is typical of the multiscale DG
method.

• The superconvergent HDG method by Nguyen et al. (2010c; Nguyen et al. (2011 extends

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

86 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

the superconvergent HDG method for Stokes by Nguyen et al. (2010b also by using last of the
above-mentioned approaches to discretize the nonlinearity. (An HDG method was proposed
and analyzed by Cesmelioglu et al. (2013.) Space-time HDG methods were considered in
Rhebergen and Cockburn (2012; Rhebergen et al. (2013.

• The HDG method by Lehrenfeld and Schöberl (2015 extends the method using the
Lehrenfeld-Schöberl stabilization function introduced by Lehrenfeld (2010; since it uses exactly
divergence-free velocities, it can discretize directly the nonlinear term ∇ · (u ⊗ u).

• The staggered DG method by Cheung et al. (2015 extends the corresponding method for
the Stokes system by Kim et al. (2013. It employs an exactly divergence-free approximation
of the velocity but its discretization of the nonlinear convection does not quite fit our general
framework.

7.4. The compressible Navier-Stokes equations of fluid flow

To end this section, we consider the compressible Navier-Stokes equations:


̺t + (̺ vj ),j =0
(̺ vi )t + (̺ vi vj − σij ),j = fi
(̺ e)t + (̺ e vj − σij vi + qi ),j = fi v i
where ̺ is the density, v the velocity, e the internal energy, and f the external body forces.
The viscous stress σ and the heat flux q are given by
σij = (−p + λ vi,i ) δij + µ (vi,j + vj,i )
qi = −κ T,i
where p is the pressure and T the temperature.

To obtain a DG method, we first rewrite the equations as


q − ∇u = 0 in Ω × (0, T ),
∂t u + ∇ · (F(u) + G(u, q)) = 0 in Ω × (0, T ),
where G(u, q) are the viscous fluxes. We then can use a space-time, or a semidiscrete DG
discretization of these equations by properly choosing the convective fluxes and the viscous
fluxes as sketched in the previous sections. A suitable time-marching scheme has then to be
applied to the semidiscretization. For example, see the high-order accurate implicit Runge-
Kutta methods in Montlaur et al. (2011.

7.4.1. Examples of DG methods. There are many DG methods for the compressible Navier-
Stokes equations. Let us mention a few to give an idea.

The first DG method for the compressible Navier-Stokes equations was proposed by Bassi
and Rebay (1997a. Based on the work by Bassi and Rebay (1997a and by Cockburn and

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 87

Shu (1998a, a DG method for the compressible Navier-Stokes equations was proposed by
Lomtev and Karniadakis (1999. A DG method based on the IP discretization of the second-
order operators was proposed by Hartmann and Houston (2002b. A space-Time Discontinuous
Galerkin method for the compressible Navier-Stokes were proposed by Klaij et al. (2006. The
compact discontinuous Galerkin method Peraire and Persson (2008 has also been applied to
compressible viscous flows Persson and Peraire (2008 and turbulent flows Nguyen et al. (2007.
More recently, an HDG methods was developed by Peraire et al. (2010. Its EDG version
was explored in Nguyen et al. (2015. A hybrid mixed method for the compressible Navier-
Stokes equations was proposed by Schütz and May (2013. Adjoint-based error estimation and
mesh adaptation using HDG methods were explored by Woopen et al. (2014. A high-order
discontinuous Galerkin discretization with multiwavelet-based grid adaptation for compressible
flows was explored by Gerhard et al. (2015.

For time-marching methods for DG methods for compressible flow, see Nigro et al. (2014a;
Nigro et al. (2014b; Bassi et al. (2015. For implicit high-order DG methods DNS and implicit
LES of turbulent flows see Bassi et al. (2016.

Domain decomposition preconditioners have been explored by Giani and Houston (2014.
And p-multigrid discontinuous Galerkin solver in Ghidoni et al. (2014.

7.4.2. Some numerical examples. We present some numerical experiments obtained by using
the LDG method developed by Lomtev and Karniadakis (1999. We present some numerical
results for the compressible Navier-Stokes equations, In Figure 35, we show a steady state
calculation of the laminar Mach 0.8 flow around a NACA 0012 airfoil with Reynolds number
73. No limiter was applied.

Another example is a time-dependent computation of the flow around a cylinder in two space
dimensions. The Reynolds number is 10 000 and the Mach number 0.2. In Figure 36, we see
the detail of a mesh of 680 triangles (with curved sides fitting the cylinder) and polynomials
whose degree could vary from element to element; the maximum degree was 5. Note how the
method is able to capture the shear layer instability observed experimentally.

We end by presenting the application of the HDG method by Peraire et al. (2010 to the
Reynolds-averaged Navier-Stokes equations, see Spalart and Allmaras (1994, in Figure 7.4.2.
We also show its application to the Euler equations in Figure 7.4.2.

8. Concluding remarks and bibliographic notes

Most of the material in this article has been taken from the monograph by Cockburn (1999,
from the reviews by Cockburn and Shu (2001; Cockburn (2003; Cockburn and Shi (2014 and
from the short essay by Cockburn (2015. The references do not pretend to be exhaustive.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

88 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

1
Cp

0 0.2 0.4 0.6 0.8 1


(a) X /C

0.5
Cf

−0.5
0 0.2 0.4 0.6 0.8 1
(b) X /C

Figure 35. Pressure (a) and drag (b) coefficient distributions. The squares were obtained by using
polynomials of degree 3 by Bassi F and Rebay S. A high-order accurate discontinuous finite element
method for the numerical solution of the compressible Navier-Stokes equations. J. Comput. Phys.
1997a; 131:267-279; and the crosses by using polynomials of degree 6 by Lomtev I and Karniadakis GE.
A discontinuous Galerkin method for the Navier-Stokes equations. Int. J. Numer. Methods Fluids 1999;
29:587-603.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 89

0.8

0.6

0.4

0.2

0
y

−0.2

−0.4

−0.6

−0.8
−0.5 0 0.5 1
(a) x

0.8

0.6

0.4

0.2
y

−0.2

−0.4

−0.6

−0.8
−0.5 0 0.5 1
(b) x

Figure 36. Flow around a cylinder with Reynolds number 10 000 and Mach number 0.2. Detail of the
mesh (a) and density (b) around the cylinder. (From Lomtev I and Karniadakis GE. A discontinuous
Galerkin method for the Navier-Stokes equations. Int. J. Numer. Methods Fluids 1999; 29:587-603.)

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

90 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Figure 37. Reynolds-averaged Navier-Stokes equations over the NACA 0012 airfoil at Mach number
0.3, Reynolds number 1.85 millions, and 0 angle of attack. Mach number contour plot provided by the
HDG method with k = 4. Courtesy of Ngoc-Cuong Nguyen.

Figure 38. Euler equations over the Trefftz airfoil at Mach number 0.2 and 0 angle of attack. Pressure
contour plot provided by the EDG method with k = 4. Courtesy of Ngoc-Cuong Nguyen.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 91

For a history of the development of DG methods up to 1999, see Cockburn et al. (2000.
For a theory of DG methods for Friedrichs’ systems, see Ern and Guermond (2006a; Ern
and Guermond (2006b; Ern and Guermond (2008; see also a systematic way of picking the
corresponding numerical traces by Bui-Thanh (2015. See also the reviews by Cheng and Shu
(2013; Shu (2014 and the monograph by Shu (2009. For books on DG methods, see Di-Pietro
and Ern (2012, Hesthaven and Warburton (2008, Kanschat (2008, Li (2006, and Rivière (2008.

9. Related Chapters

(See also Finite Element Methods, Finite Volume Methods: Foundation and Analysis;
Multiscale and Stabilized Methods)

Acknowledgments

The author would like to thank Thomas Hughes for the kind invitation to write this paper.
Concerning the first edition, the author would like to thank the many colleagues that provided
most of the figures in this paper: Francesco Bassi, Paul Castillo, Nicolas Chauvegeon, Clint
Dawson, Paul Houston, Igor Lomtev, Guido Kanschat, George Karniadakis, Manoj Prasad,
Stefano Rebay, Chi-Wang Shu, Jaap van der Vegt and Harmen van der Ven. Thanks are also
due to Paul Houston for a careful reading of the paper. Finally, the author would like to thank
Clint Dawson, Guido Kanschat, Paul Houston, Dominik Schötzau and Harmen van der Ven
for useful feedback on the first version of the manuscript. Concerning the second edition, the
author would like to thank Ngoc-Cuong Nguyen, Jiguang Shen, Chi-Wang Shu and Xiongxiang
Zhang for providing new figures, and to Mauricio Flores, Ilaria Perugia and Jennifer Ryan for
providing bibliographic material.

REFERENCES

S. Adjerid and M. Baccouch (2012). A superconvergent local discontinuous Galerkin method


for elliptic problems. J. Sci. Comput., 52(1):113–152.
S. Adjerid and T.C. Massey (2002). A posteriori discontinuous finite element error
estimation for two-dimensional hyperbolic problems. Comput. Methods Appl. Mech.
Engrg., 191:5877–5897.
S. Adjerid and I. Mechai (2014). A superconvergent discontinuous Galerkin method for
hyperbolic problems on tetrahedral meshes. J. Sci. Comput., 58(1):203–248.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

92 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

S. Adjerid and T. Weinhart (2011). Discontinuous Galerkin error estimation for linear
symmetrizable hyperbolic systems. Math. Comp., 80(275):1335–1367.
S. Adjerid and T. Weinhart (2014). Asymptotically exact discontinuous Galerkin error
estimates for linear symmetric hyperbolic systems. Appl. Numer. Math., 76:101–131.
M. Ainsworth (2004). Dispersive and dissipative behaviour of high order discontinuous
Galerkin finite element methods. J. Comput. Phys., 198(1):106–130.
M. Ainsworth (2007). A posteriori error estimation for discontinuous Galerkin finite element
approximation. SIAM J. Numer. Anal., 44:1777–1798.
M. Ainsworth and R. Rankin (2010). Fully computable error bounds for discontinuous Galerkin
finite element approximations on meshes with an arbitrary number of levels of hanging
nodes. SIAM J. Numer. Anal., 46:4112–4141.
V. Aizinger and C.N. Dawson (2003). A Discontinuous Galerkin Method for two-dimensional
flow and transport in shallow water. Technical Report 03-16, TICAM.
V. Aizinger, C.N. Dawson, B. Cockburn and P. Castillo (2000). Local discontinuous Galerkin
method for contaminant transport. Advances in Water Resources, 24:73–87.
P. Alotto, A. Bertoni, I. Perugia and D. Schötzau (2001). Discontinuous Finite Element
Methods for the Simulation of Rotating Electrical Machines. COMPEL, 20:448–462.
P. F. Antonietta, S. Giani and P. Houston (2014). Domain decomposition preconditioners for
discontinuous Galerkin methods for elliptic problems on complicated domains. J. Sci.
Comput., 60(1):203–227.
D. N. Arnold (1982). An interior penalty finite element method with discontinuous elements.
SIAM J. Numer. Anal., 19:742–760.
D. N. Arnold and F. Brezzi (1985). Mixed and nonconforming finite element methods:
implementation, postprocessing and error estimates. RAIRO Modél. Math. Anal. Numér.,
19:7–32.
D. N. Arnold, F. Brezzi, B. Cockburn and L. D. Marini (2002). Unified analysis of discontinuous
Galerkin methods for elliptic problems. SIAM J. Numer. Anal., 39:1749–1779.
H.L. Atkins and C.-W. Shu (1998). Quadrature-free implementation of discontinuous Galerkin
methods for hyperbolic equations. AIAA Journal, 36.
I. Babuška and M. Zlámal (1973). Nonconforming elements in the finite element method with
penalty. SIAM J. Numer. Anal., 10:863–875.
M. Baccouch and S. Adjerid (2015). A posteriori local discontinuous Galerkin error estimation
for two-dimensional convection-diffusion problems. J. Sci. Comput., 62(2):399–430.
G. A. Baker (1977). Finite element methods for elliptic equations using nonconforming
elements. Math. Comp., 31:45–59.
G. A. Baker, W.N. Jureidini and O. A. Karakashian (1990). Piecewise solenoidal vector fields
and the Stokes problem. SIAM J. Numer. Anal., 27:1466–1485.
G. E. Barter and D. L. Darmofal (2010). Shock capturing with PDE-based artificial viscosity
for DGFEM. I. Formulation. J. Comput. Phys., 229(5):1810–1827.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 93

T. Barth (2000). Simplified DG methods for systems of conservation laws with convex
extension. In Cockburn, B., Karniadakis, G., and Shu, C.-W., editors, Discontinuous
Galerkin Methods. Theory, Computation and Applications, volume 11 of Lect. Notes
Comput. Sci. Engrg., pages 63–75, Berlin. Springer Verlag.
F. Bassi, L. Botti, A. Colombo, A. Crivellini, A. Ghidoni and F. Massa (2016). On the
development of an implicit high-order Discontinuous Galerkin method for DNS and
implicit LES of turbulent flows. Eur. J. Mech. B Fluids, 55(part 2):367–379.
F. Bassi, L. Botti and A. Colombo, A. Ghidoni and F. Massa (2015). Linearly implicit
Rosenbrock-type Runge-Kutta schemes applied to the discontinuous Galerkin solution
of compressible and incompressible unsteady flows. Comput. & Fluids, 118:305–320.
F. Bassi, A. Crivellini, D. A. Di-Pietro and S. Rebay (2006). An artificial compressibility flux
for the discontinuous Galerkin solution of the incompressible Navier-Stokes equation. J.
Comput. Phys., 218:794–815.
F. Bassi and S. Rebay (1997a). A high-order accurate discontinuous finite element method for
the numerical solution of the compressible Navier-Stokes equations. J. Comput. Phys.,
131:267–279.
F. Bassi and S. Rebay (1997b). High-order accurate discontinuous finite element solution of
the 2D Euler equations. J. Comput. Phys., 138:251–285.
F. Bassi, S. Rebay, G. Mariotti, S. Pedinotti and M. Savini (1997). A High-Order Accurate
Discontinuous Finite Element Method for Inviscid and Viscous Turbomachinery Flows.
In Decuypere, R. and Dibelius, G., editors, 2nd European Conference on Turbomachinery
Fluid Dynamics and Thermodynamics, pages 99–108, Antwerpen, Belgium. Technologisch
Instituut.
P. Bastian and B. Rivière (2003). Superconvergence and H(div) projection for discontinuous
Galerkin methods. Internat. J. Numer. Methods Fluids, 42:1043–1057.
C. E. Baumann and J. T. Oden (1999). A discontinuous hp-finite element method for
convection-diffusion problems. Comput. Methods Appl. Mech. Engrg., 175:311–341.
R. Becker, P. Hansbo and M. G. Larson (2003). Energy norm a posteriori error estimation for
discontinuous Galerkin methods. Comput. Methods Appl. Mech. Engrg., 192:723–733.
K.S. Bey (1994). An hp-adaptive discontinuous Galerkin method for hyperbolic conservation
laws. PhD thesis, The University of Texas at Austin.
K.S. Bey and J. T. Oden (1996). hp-version discontinuous Galerkin methods for hyperbolic
conservation laws. Comput. Methods Appl. Mech. Engrg., 133:259–286.
R. Biswas, K.D. Devine and J.E. Flaherty (1994). Parallel, adaptive finite element methods
for conservation laws. Appl. Numer. Math., 14:255–283.
P. B. Bochev, J. T. R. Hughes and G. Scovazzi (2006). A multiscale discontinuous Galerkin
method. In I. Lirkov, S. M. and Waśniewski, J., editors, Large-scale scientific computing,
volume 3743 of Lecture Notes in Comput. Sci., pages 84–93. Springer-Verlag Berlin
Heidelberg.
A. Bonito and R. H. Nochetto (2010). Quasi-optimal convergence rate of an adaptive
discontinuous Galerkin method. SIAM J. Numer. Anal., 48(2):734–771.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

94 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

J. H. Bramble and A. H. Schatz (1977). Higher order local accuracy by averaging in the finite
element method. Math. Comp., 31:94–111.
F. Brezzi, B. Cockburn, L. D. Marini and E. Süli (2006). Stabilization mechanisms in
discontinuous Galerkin finite element methods. Comput. Methods Appl. Mech. Engrg.,
195:3293–3310. C. Dawson, Ed.
F. Brezzi, J. Douglas, Jr. and L. D. Marini (1985). Two families of mixed finite elements for
second order elliptic problems. Numer. Math., 47:217–235.
F. Brezzi, G. Manzini, L. D. Marini, P. Pietra and A. Russo (1999). Discontinuous finite
elements for diffusion problems. In in Atti Convegno in onore di F. Brioschi (Milan
1997), pages 197–217. Istituto Lombardo, Accademia di Scienze e Lettere.
F. Brezzi, G. Manzini, L. D. Marini, P. Pietra and A. Russo (2000). Discontinuous Galerkin
Approximations for elliptic Problems. Numer. Methods Partial Differential Equations.,
16:365–378.
A. Buffa, T. J. R. Hughes and G. Sangalli (2006). Analysis of a multiscale discontinuous
Galerkin method for convection-diffusion problems. SIAM J. Numer. Anal., 44:1420–
1440 (electronic).
T. Bui-Thanh (2015). From Godunov to a unified hybridized discontinuous Galerkin framework
for partial differential equations. J. Comput. Phys., 295:114–146.
A. Burbeau, P. Sagaut and Ch.-H. Bruneau (2001). A problem-independent limiter for high-
order Runge-Kutta discontinuous Galerkin methods. J. Comput. Phys., 169:111–150.
R. Bustinza, G.N. Gatica and B. Cockburn (2005). An a posteriori error estimate for the local
discontinuous Galerkin method applied to linear and nonlinear diffusion problems. J. Sci.
Comput., 22/23:147–185.
A. Cangiani, E. H. Georgoulis and P. Houston (2014). hp-version discontinuous Galerkin
methods on polygonal and polyhedral meshes. Math. Models Methods Appl. Sci.,
24(10):2009–2041.
J. Carrero, B. Cockburn and D. Schötzau (2006). Hybridized, globally divergence-free LDG
methods. Part I: The Stokes problem. Math. Comp., 75:533–563.
C. Carstensen, T. Gudi and M. Jensen (2009). A unifying theory of a posteriori error control
for discontinuous Galerkin FEM. Numer. Math., 112:363–379.
E. Casoni, J. Peraire and A. Huerta (2013). One-dimensional shock-capturing for high-order
discontinuous Galerkin methods. Internat. J. Numer. Methods Fluids, 71(6):737–755.
P. Castillo (2002). Performance of discontinuous Galerkin methods for elliptic PDE’s. SIAM
J. Sci. Comput., 24:524–547.
P. Castillo (2005). An a posteriori error estimate for the local discontinuous Galerkin method.
J. Sci. Comput., 22/23:187–204.
P. Castillo, B. Cockburn, I. Perugia and D. Schötzau (2000). An a priori error analysis of
the local discontinuous Galerkin method for elliptic problems. SIAM J. Numer. Anal.,
38:1676–1706.
P. E. Castillo (2010). Stencil reduction algorithms for the local discontinuous Galerkin method.
Internat. J. Numer. Methods Engrg., 81(12):1475–1491.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 95

P. E. Castillo and F. A. Sequeira (2013). Computational aspects of the local discontinuous


Galerkin method on unstructured grids in three dimensions. Math. Comput. Modelling,
57(9-10):2279–2288.
F. Celiker and B. Cockburn (2007). Superconvergence of the numerical traces of discontinuous
Galerkin and hybridized mixed methods for convection-diffusion problems in one space
dimension. Math. Comp., 76:67–96.
A. Cesmelioglu, B. Cockburn, N.-C. Nguyen and J. Peraire (2013). Analysis of HDG methods
for Oseen equations. J. Sci. Comput., 55(2):392–431.
B. Chabaud and B. Cockburn (2012). Uniform-in-time superconvergence of HDG methods for
the heat equation. Math. Comp., 81:107–129.
Y. Chen and B. Cockburn (2012). Analysis of variable-degree HDG methods for Convection-
Diffusion equations. Part I: General nonconforming meshes. IMA J. Num. Anal., 32:1267–
1293.
Y. Chen and B. Cockburn (2014). Analysis of variable-degree HDG methods for Convection-
Diffusion equations. Part II: Semimatching nonconforming meshes. Math. Comp., 83:87–
111.
Z. Chen and H. Chen (2004). Pointwise error estimates of discontinuous Galerkin methods
with penalty for second-order elliptic problems. SIAM J. Numer. Anal., 42(3):1146–1166
(electronic).
J. Cheng and C.-W. Shu (2013). Recent developments on the Lagrangian and remapping
methods for compressible fluid flows. In Numerical methods for hyperbolic equations:
theory and applications, pages 9–19. CRC Press/Balkema, Leiden.
S. W. Cheung, E. T. Chung, H. H. Kim and Y. Qian (2015). Staggered discontinuous Galerkin
methods for the incompressible Navier-Stokes equations. J. Comput. Phys., 302:251–266.
E. T. Chung, B. Cockburn and G. Fu (2014). The staggered DG method is the limit of a
hybridizable DG method. SIAM J. Numer. Anal., 52:915–932.
E. T. Chung, B. Cockburn and G. Fu (2016). The staggered DG method is the limit of a
hybridizable DG method. Part II: The Stokes flow. J. Sci. Comput., 66(2):870–887.
E. T. Chung and B. Engquist (2009). Optimal discontinuous Galerkin methods for the acoustic
wave equation in higher dimensions. SIAM J. Numer. Anal., 47(5):3820–3848.
S. Cochez-Dhondt and S. Nicaise (2008). Equilibrated error estimators for discontinuous
Galerkin methods. Numer. Methods Partial Differential Equations., 24:1236–1252.
B. Cockburn (1999). Discontinuous Galerkin Methods for Convection-Dominated Problems.
In Barth, T. and Deconink, H., editors, High-Order Methods for Computational Physics,
volume 9 of Lect. Notes Comput. Sci. Engrg., pages 69–224. Springer Verlag, Berlin.
B. Cockburn (2001). Devising discontinuous Galerkin methods for non-linear hyperbolic
conservation laws. Journal of Computational and Applied Mathematics, 128:187–204.
B. Cockburn (2003). Discontinuous Galerkin Methods. ZAMM Z. Angew. Math. Mech.,
83:731–754.
B. Cockburn (2009). Two new techniques for generating exactly incompressible approximate
velocities. In Deconinck, H. and Dick, E., editors, Computational Fluid Dynamics 2006.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

96 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

Proceedings of the Fourth International Conference in Fluid Dynamics, ICCDF4, Ghent,


Belgium, 10-14 July 2006, pages 1–11. Springer Verlag.
B. Cockburn (2015). Static condensation, hybridization, and the devising of the HDG methods.
In G.R. Barrenechea, F. Brezzi, A. C. and Georgulis, E., editors, Building Bridges:
Connections and Challenges in Modern Approaches to Numerical Partial Differential
Equations, volume 114 of Lect. Notes Comput. Sci. Engrg., pages 1–45. Springer Verlag,
Berlin. LMS Durham Symposia funded by the London Mathematical Society. Durham,
U.K., on July 8–16, 2014.
B. Cockburn and J. Cui (2012a). An analysis of HDG methods for the vorticity-velocity-
pressure formulation of the Stokes problem in three dimensions. Math. Comp., 81:1355–
1368.
B. Cockburn and J. Cui (2012b). Divergence-Free HDG Methods for the Vorticity-Velocity
Formulation of the Stokes Problem. J. Sci. Comput., 52:256–270.
B. Cockburn and C. Dawson (2000). Some extensions of the local discontinuous Galerkin
method for convection-diffusion equations in multidimensions. In Whiteman, J.,
editor, The Proceedings of the Conference on the Mathematics of Finite Elements and
Applications: MAFELAP X, pages 225–238. Elsevier.
B. Cockburn and C. Dawson (2002). Approximation of the velocity by coupling discontinuous
Galerkin and mixed finite element methods for flow problems. Comput. Geosci. (Special
issue: Locally Conservative Numerical Methods for Flow in Porous Media), 6:502–522.
B. Cockburn, D. A. Di-Pietro and A. Ern (2015). Bridging the Hybrid High-Order and
Hybridizable Discontinuous Galerkin Methods. ESAIM Math. Model. Numer. Anal. to
appear.
B. Cockburn and B. Dong (2007). An analysis of the minimal dissipation local discontinuous
Galerkin method for convection-diffusion problems. J. Sci. Comput., 32:233–262.
B. Cockburn and B. Dong and J. Guzmán (2008a). Optimal convergence of the original DG
method for the transport-reaction equation on special meshes. SIAM J. Numer. Anal.,
46:1250–1265.
B. Cockburn and B. Dong and J. Guzmán (2008b). A superconvergent LDG-hybridizable
Galerkin method for second-order elliptic problems. Math. Comp., 77:1887–1916.
B. Cockburn, B. Dong, J. Guzmán and J. Qian (2010a). Optimal convergence of the original
DG method on special meshes for variable convective velocity. SIAM J. Numer. Anal.,
48(1):133–146.
B. Cockburn, B. Dong, J. Guzmán, M. Restelli and R. Sacco (2009a). Superconvergent and
optimally convergent LDG-hybridizable discontinuous Galerkin methods for convection-
diffusion-reaction problems. SIAM J. Sci. Comput., 31:3827–3846.
B. Cockburn, O. Dubois, J. Gopalakrishnan and S. Tan (2014). Multigrid for an HDG method.
IMA J. Numer. Anal., 34(4):1386–1425.
B. Cockburn and G. Fu (2016a). Superconvergence by M-decompositions. Part II: Construction
of two-dimensional finite elements. Modél. Math. Anal. Numér. To appear.
B. Cockburn and G. Fu (2016b). Superconvergence by M-decompositions. Part III:
Construction of three-dimensional finite elements. Modél. Math. Anal. Numér. To appear.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 97

B. Cockburn, G. Fu and W. Qiu (2016a). A note on the devising of superconvergent HDG


methods for Stokes flow by M-decompositions. IMA J. Num. Anal. To appear.
B. Cockburn, G. Fu and F.-J. Sayas (2016b). Superconvergence by M-decompositions. Part I:
General theory for HDG methods for diffusion. Math. Comp. to appear.
B. Cockburn and J. Gopalakrishnan (2005a). Incompressible finite elements via hybridization.
Part I: The Stokes system in two space dimensions. SIAM J. Numer. Anal., 43:1627–1650.
B. Cockburn and J. Gopalakrishnan (2005b). Incompressible finite elements via hybridization.
Part II: The Stokes system in three space dimensions. SIAM J. Numer. Anal., 43:1651–
1672.
B. Cockburn and J. Gopalakrishnan (2005c). New hybridization techniques. GAMM Mitt.
Ges. Angew. Math. Mech., 2:154–183.
B. Cockburn and J. Gopalakrishnan (2009). The derivation of hybridizable discontinuous
Galerkin methods for Stokes flow. SIAM J. Numer. Anal., 47:1092–1125.
B. Cockburn, J. Gopalakrishnan and J. Guzmán (2010b). A new elasticity element made for
enforcing weak stress symmetry. Math. Comp., 79:1331–1349.
B. Cockburn, J. Gopalakrishnan and R. Lazarov (2009b). Unified Hybridization of
discontinuous Galerkin, mixed and continuous Galerkin methods for second order elliptic
problems. SIAM J. Numer. Anal., 47:1319–1365.
B. Cockburn, J. Gopalakrishnan, F. Li, N.-C. Nguyen and J. Peraire (2010c). Hybridization
and postprocessing techniques for mixed eigenfunctions. SIAM J. Numer. Anal., 48:857–
881.
B. Cockburn, J. Gopalakrishnan, N. C. Nguyen, J. Peraire and F.-J. Sayas (2011). Analysis
of an HDG method for Stokes flow. Math. Comp., 80:723–760.
B. Cockburn, J. Gopalakrishnan and F.-J. Sayas (2010d). A projection-based error analysis
of HDG methods. Math. Comp., 79:1351–1367.
B. Cockburn and P.-A. Gremaud (1996). Error estimates for finite element methods for
nonlinear conservation laws. SIAM J. Numer. Anal., 33:522–554.
B. Cockburn and J. Guzmán (2008). Error estimates for the Runge-Kutta discontinuous
Galerkin method for the transport equation with discontinuous initial data. SIAM J.
Numer. Anal., 46:1364–1398.
B. Cockburn, J. Guzmán, S.-C. Soon and H.K. Stolarski (2009c). An analysis of the embedded
discontinuous Galerkin method for second-order elliptic problems. SIAM J. Numer. Anal.,
47:2686–2707.
B. Cockburn, S. Hou and C.-W. Shu (1990). TVB Runge-Kutta Local Projection Discontinuous
Galerkin Finite Element Method for Conservation Laws IV: The Multidimensional Case.
Math. Comp., 54:545–581.
B. Cockburn, G. Kanschat, I. Perugia and D. Schötzau (2001). Superconvergence of the local
discontinuous Galerkin method for elliptic problems on Cartesian grids. SIAM J. Numer.
Anal., 39:264–285.
B. Cockburn, G. Kanschat and D. Schötzau (2004). Local discontinuous Galerkin methods for
the Oseen equations. Math. Comp., 73:569–593.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

98 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

B. Cockburn, G. Kanschat and D. Schötzau (2005). A locally conservative LDG method for
the incompressible Navier-Stokes equations. Math. Comp., 74:1067–1095.
B. Cockburn, G. Kanschat and D. Schötzau (2007). A note on discontinuous Galerkin
divergence-free solutions of the Navier-Stokes equations. J. Sci. Comput., 31:61–73.
B. Cockburn, G. Kanschat and D. Schötzau (2009d). An equal-order DG method for the
incompressible Navier-Stokes equations. J. Sci. Comput., 40:141–187.
B. Cockburn, G. Kanschat, D. Schötzau and C. Schwab (2002). Local discontinuous Galerkin
methods for the Stokes system. SIAM J. Numer. Anal., 40:319–343.
B. Cockburn, G.E. Karniadakis and C.-W. Shu (2000). The development of discontinuous
Galerkin methods. In Cockburn, B., Karniadakis, G., and Shu, C.-W., editors,
Discontinuous Galerkin Methods. Theory, Computation and Applications, volume 11 of
Lect. Notes Comput. Sci. Engrg., pages 3–50, Berlin. Springer Verlag.
B. Cockburn, M. Luskin, C.-W. Shu and E. Süli (2003). Enhanced accuracy by post-processing
for finite element methods for hyperbolic equations. Math. Comp., 72:577–606 (electronic).
B. Cockburn and R. H. Nochetto and W. Zhang (2016c). Contraction property of adaptive
hybridizable discontinuous Galerkin methods. Math. Comp., 85(299):1113–1141.
B. Cockburn and W. Qiu (2014). Commuting diagrams for the TNT elements on cubes. Math.
Comp., 83:603–633.
B. Cockburn, W. Qiu and K. Shi (2012a). Conditions for superconvergence of HDG methods
for second-order eliptic problems. Math. Comp., 81:1327–1353.
B. Cockburn, W. Qui and K. Shi (2012b). Conditions for superconvergence of HDG methods on
curvilinear elements for second-order eliptic problems. SIAM J. Numer. Anal., 50:1417–
1432.
B. Cockburn and F.-J. Sayas (2014). Divergence–conforming HDG methods for Stokes flow.
Math. Comp., 83:1571–1598.
B. Cockburn, F.-J. Sayas and M. Solano (2012c). Coupling at a distance HDG and BEM.
SIAM J. Sci. Comput., 34:A28–A47.
B. Cockburn and Shen, Jiguang (2016). A Hybridizable Discontinuous Galerkin Method for
the p-Laplacian. SIAM J. Sci. Comput., 38(1):A545–A566.
B. Cockburn and K. Shi (2014). Devising HDG methods for Stokes flow: An overview.
Computers & Fluids, 98:221–229.
B. Cockburn and C.-W. Shu (1998a). The local Discontinuous Galerkin Method for time-
dependent convection-diffusion systems. SIAM J. Numer. Anal., 35:2440–2463.
B. Cockburn and C.-W. Shu (1998b). The Runge-Kutta Discontinuous Galerkin Finite Element
Method for Conservation Laws V: Multidimensional Systems. J. Comput. Phys., 141:199–
224.
B. Cockburn and C.-W. Shu (2001). Runge-Kutta Discontinuous Galerkin Methods for
convection-dominated problems. J. Sci. Comput., 16:173–261.
B. Cockburn and W. Zhang (2012). A posteriori error estimates for HDG methods. J. Sci.
Comput., 51(3):582–607.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 99

B. Cockburn and W. Zhang (2013). A posteriori error analysis for hybridizable discontinuous
Galerkin methods for second order elliptic problems. SIAM J. Numer. Anal., 51:676–693.
E. Creusé and S. Nicaise (2010). A posteriori error estimator based on gradient recovery by
averaging for discontinuous Galerkin methods. J. Comput. Appl. Math., 234:2903–2915.
A. Crivellini, V. D’Alessandro and F. Bassi (2013). Assessment of a high-order
discontinuous Galerkin method for incompressible three-dimensional Navier-Stokes
equations: benchmark results for the flow past a sphere up to Re = 500. Comput. &
Fluids, 86:442–458.
S. Curtis, R. M. Kirby, J. K. Ryan and C.-W. Shu (2007/08). Postprocessing for the
discontinuous Galerkin method over nonuniform meshes. SIAM J. Sci. Comput.,
30(1):272–289.
C. Dawson and J. Proft (2002). Discontinuous and coupled continuous/discontinuous Galerkin
methods for the shallow water equations. Comput. Methods Appl. Mech. Engrg., 191(41-
42):4721–4746.
C. Dawson and J. Proft (2003). Discontinuous/continuous Galerkin methods for coupling the
primitive and wave continuity equations of shallow water. Comput. Methods Appl. Mech.
Engrg., 192(47-48):5123–5145.
C. Dawson and J. Proft (2004). Coupled discontinuous and continuous Galerkin finite element
methods for the depth-integrated shallow water equations. Comput. Methods Appl. Mech.
Engrg., 193(3-5):289–318.
K.D. Devine and J.E. Flaherty (1996). Parallel adaptive hp-refinement techniques for
conservation laws. Appl. Numer. Math., 20:367–386.
K.D. Devine, J.E. Flaherty, R.M. Loy and S.R. Wheat (1995). Parallel partitioning strategies
for the adaptive solution of conservation laws. In Babuška, I., Flaherty, J., Henshaw,
W., Hopcroft, J., Oliger, J., and Tezduyar, T., editors, Modeling, mesh generation, and
adaptive numerical methods for partial differential equations, volume 75 of The IMA
Volumes in Mathematics and its Applications, pages 215–242, New York, New York.
Sringer-Verlag.
D. A. Di-Pietro and A. Ern (2012). Mathematical aspects of discontinuous Galerkin methods,
volume 69 of Mathématiques & Applications (Berlin) [Mathematics & Applications].
Springer, Heidelberg.
D. A. Di-Pietro and A. Ern (2015). A hybrid high-order locking-free method for linear elasticity
on general meshes. Comput. Meth. Appl. Mech. Engrg., 283:1–21.
M. Dubiner (1991). Spectral Methods on Triangles and Other Domains. J. Sci. Comp., 6:345–
390.
Y. Efendiev, R. Lazarov and K. Shi (2015). A multiscale HDG method for second order
elliptic equations. Part I. Polynomial and homogenization-based multiscale spaces. SIAM
J. Numer. Anal., 53(1):342–369.
H. Egger and J. Schöberl (2010). A hybrid mixed discontinuous Galerkin finite-element method
for convection-diffusion problems. IMA J. Numer. Anal., 30(4):1206–1234.
A. Ern and J. L. Guermond (2006a). Discontinuous Galerkin Methods for Friedrichs’ Systems.
I. General theory. SIAM J. Numer. Anal., 44:753–778.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

100 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

A. Ern and J.-L. Guermond (2006b). Discontinuous Galerkin methods for Friedrichs’ systems.
II. Second-order elliptic PDEs. SIAM J. Numer. Anal., 44(6):2363–2388.
A. Ern and J.-L. Guermond (2008). Discontinuous Galerkin methods for Friedrichs’ systems.
III. Multifield theories with partial coercivity. SIAM J. Numer. Anal., 46(2):776–804.
A. Ern, S. Nicaise and M. Vohralı́k (2007). An accurate H(div) flux reconstruction for
discontinuous Galerkin approximations of elliptic problems. C. R. Math. Acad. Sci. Paris,
345:709–712.
R.S. Falk and G.R. Richter (1999). Explicit finite element methods for symmetric hyperbolic
equations. SIAM J. Numer. Anal., 36:935–952.
X. Feng and O. A. Karakashian (2001). Two-level non-overlapping Schwarz methods for a
discontinuous Galerkin method. SIAM J. Numer. Anal., 39:1343–1365 (electronic).
J.E. Flaherty, R.M. Loy, C. Özturan, M.S. Shephard, B.K. Szymanski, J.D. Teresco and L.H.
Ziantz (1998). Parallel Structures and Dynamic Load Balancing for Adaptive Finite
Element Computation. Appl. Numer. Math., 26:241–265.
J.E. Flaherty, R.M. Loy, M.S. Shephard, M.L. Simone, B.K. Szymanski , J.D. Teresco and
L.H. Ziantz (1999). Distributed Octree Data Structures and Local Refinement Method
for the Parallel Solution of Three-Dimensional Conservation Laws. In Bern, M., Flaherty,
J., and Luskin, M., editors, Grid Generation and Adaptive Algorithms, volume 113 of The
IMA Volumes in Mathematics and its Applications, pages 113–134, Minneapolis. Institute
for Mathematics and its Applications, Springer.
J.E. Flaherty, R.M. Loy, M.S. Shephard, B.K. Szymanski, J.D. Teresco and L.H. Ziantz (1997).
Adaptive local refinement with octree load-balancing for the parallel solution of three-
dimensional conservation laws. J. Parallel and Dist. Comput., 47:139–152.
J.E. Flaherty, R. Loy, M.S. Shephard and J. Teresco (2000). Software for the parallel
adaptive solution of conservation laws by a discontinuous Galerkin method. In Cockburn,
B., Karniadakis, G., and Shu, C.-W., editors, Discontinuous Galerkin Methods. Theory,
Computation and Applications, volume 11 of Lect. Notes Comput. Sci. Engrg., pages 113–
123, Berlin. Springer Verlag.
B. M. Fraejis de Veubeke (1977). Displacement and equilibrium models in the finite element
method. In Zienkiewicz, O. and Holister, G., editors, Stress Analysis, pages 145–197.
Wiley, New York.
K.O. Friedrichs (1958). Symmetric positive linear differential equations. Comm. Pure and
Appl. Math., 11:333–418.
G. Fu, W. Qiu and W. Zhang (2015). An analysis of HDG methods for convection-dominated
diffusion problems. ESAIM Math. Model. Numer. Anal., 49(1):225–256.
L. Gastaldi and R. H. Nochetto (1989). Sharp maximum norm error estimates for general
mixed finite element approximations to second order elliptic equations. RAIRO Modél.
Math. Anal. Numér., 23:103–128.
N. Gerhard, F. Iacono, G. May, S. Müller and R. Schäfer (2015). A high-order discontinuous
Galerkin discretization with multiwavelet-based grid adaptation for compressible flows.
J. Sci. Comput., 62(1):25–52.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 101

A. Ghidoni, A. Colombo, F. Bassi and S. Rebay (2014). Efficient p-multigrid discontinuous


Galerkin solver for complex viscous flows on stretched grids. Internat. J. Numer. Methods
Fluids, 75(2):134–154.
S. Giani and P. Houston (2014). Domain decomposition preconditioners for discontinuous
Galerkin discretizations of compressible fluid flows. Numer. Math. Theory Methods Appl.,
7(2):123–148.
J. Giesselmann, Ch. Makridakis and T. Pryer (2015). A posteriori analysis of discontinuous
Galerkin schemes for systems of hyperbolic conservation laws. SIAM J. Numer. Anal.,
53(3):1280–1303.
V. Girault, B. Rivière and M. F. Wheeler (2005). A discontinuous Galerkin method with
non-overlapping domain decomposition for the Stokes and Navier-Stokes problems. Math.
Comp., 74:53–84.
J. Gopalakrishnan and G. Kanschat (2003a). Application of unified DG analysis to
preconditioning DG methods. In Bathe, K., editor, Proceedings of the Second M.I.T.
Conference on Computational Fluid and Solid Mechanics, Cambridge, MA, USA, pages
1943–1945. Elsevier.
J. Gopalakrishnan and G. Kanschat (2003b). A multilevel discontinuous Galerkin method.
Numer. Math., 95:527–550.
J. Gopalakrishnan, F. Li, N.-C. Nguyen and J. Peraire (2015). Spectral approximations by the
HDG method. Math. Comp., 84:1037–1059.
S. Gottlieb and C.-W. Shu (1998). Total variation diminishing Runge-Kutta schemes. Math.
Comp., 67:73–85.
S. Gottlieb, C.-W. Shu and E. Tadmor (2000). Strong stability preserving high order time
discretization methods. SIAM Rev., 43:89–112.
R. J. Guyan (1965). Reduction of Stiffness and Mass Matrices. J. Am. Inst. Aron. and Astro.,
3:380.
S. Güzey, B. Cockburn and H.K. Stolarski (2007). The embedded discontinuous Galerkin
methods: Application to linear shells problems. Internat. J. Numer. Methods Engrg.,
70:757–790.
J. Guzmán (2006). Pointwise error estimates for discontinuous Galerkin methods with lifting
operators for elliptic problems. Math. Comp., 75(255):1067–1085 (electronic).
J. Guzmán and B. Rivière (2009). Sub-optimal convergence of non-symmetric discontinuous
Galerkin methods for odd polynomial approximations. J. Sci. Comput., 40(1-3):273–280.
P. Hansbo and M.G. Larson (2008). Piecewise divergence-free discontinuous Galerkin methods
for Stokes flow. Comm. Numer. Methods Engrg., 24:355–366.
K. Harriman, P. Houston, B. Senior and E. Süli (2003). hp-version discontinuous Galerkin
methods with interior penalty for partial differential equations with nonnegative
characteristic form. In Shu, C.-W., Tang, T., and Cheng, S.-Y., editors, Recent advances
in scientific computing and partial differential equations (Hong Kong, 2002), volume 330
of Contemp. Math., pages 89–119. Amer. Math. Soc., Providence, RI.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

102 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

R. Hartmann and P. Houston (2002a). Adaptive discontinuous Galerkin finite element


methods for nonlinear hyperbolic conservation laws. SIAM J. Sci. Comput., 24:979–1004
(electronic).
R. Hartmann and P. Houston (2002b). Adaptive discontinuous Galerkin finite element methods
for the compressible Euler equations. J. Comput. Phys., 183:508–532.
F. Hecht (1981). Construction d’une base P1 non conforme à divergence nulle. RAIRO Modél.
Math. Anal. Numér., 15:119–150.
J. S. Hesthaven and T. Warburton (2002). Nodal high-order methods on unstructured grids.
I. Time-domain solution of Maxwell’s equations. J. Comput. Phys., 181:186–221.
J. S. Hesthaven and T. Warburton (2008). Nodal discontinuous Galerkin methods, volume 54
of Texts in Applied Mathematics. Springer, New York. Algorithms, analysis, and
applications.
R. H. W. Hoppe , G. Kanschat and T. Warburton (2008). Convergence analysis of an adaptive
interior penalty discontinuous Galerkin method. SIAM J. Numer. Anal., 47(1).
P. Houston, D. Schötzau and T. P. Wihler (2007). Energy norm a posteriori error estimation of
hp-adaptive discontinuous Galerkin methods for elliptic problems. Math. Models Methods
Appl. Sci., 17:33–62.
P. Houston, C. Schwab and E. Süli (2000). Stabilized hp-finite element methods for hyperbolic
problems. SIAM J. Numer. Anal., 37:1618–1643.
P. Houston, C. Schwab and E. Süli (2001). hp-adaptive discontinuous Galerkin finite element
methods for hyperbolic problems. SIAM J. Sci. Comput., 23:1226–1252.
P. Houston, B. Senior and E. Süli (2002). hp-discontinuous Galerkin finite element methods for
hyperbolic problems: error analysis and adaptivity. Internat. J. Numer. Methods Fluids,
40:153–169. ICFD Conference on Numerical Methods for Fluid Dynamics (Oxford, 2001).
P. Houston and E. Süli (2001). hp-adaptive discontinuous Galerkin finite element methods for
first-order hyperbolic problems. SIAM J. Sci. Comput., 23:1226–1252 (electronic).
X. Y. Hu, N. A. Adams and C.-W. Shu (2013). Positivity-preserving method for high-order
conservative schemes solving compressible Euler equations. J. Comput. Phys., 242:169–
180.
F. Q. Hu and H. L. Atkins (2002). Eigensolution analysis of the discontinuous Galerkin method
with nonuniform grids. I. One space dimension. J. Comput. Phys., 182(2):516–545.
A. Huerta, A. Angeloski, X. Roca and J. Peraire (2013). Efficiency of high-order elements for
continuous and discontinuous Galerkin methods. Internat. J. Numer. Methods Engrg.,
96(9):529–560.
A. Huerta, E. Casoni and J. Peraire (2012). A simple shock-capturing technique for high-order
discontinuous Galerkin methods. Internat. J. Numer. Methods Fluids, 69(10):1614–1632.
J. T. R. Hughes, G. Scovazzi, P. B. Bochev and A. Buffa (2006). A multiscale discontinuous
Galerkin method with the computational structure of a continuous Galerkin method.
Comput. Methods Appl. Mech. Engrg., 195:2761–2787.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 103

J. Jaffré, C. Johnson and A. Szepessy (1995). Convergence of the discontinuous Galerkin finite
element method for hyperbolic conservation laws. Mathematical Models & Methods in
Applied Sciences, 5:367–386.
L. Ji, P. van Slingerland, J. K. Ryan and K. Vuik (2014). Superconvergent error estimates for
position-dependent smoothness-increasing accuracy-conserving (SIAC) post-processing of
discontinuous Galerkin solutions. Math. Comp., 83(289):2239–2262.
L. Ji, Y. Xu and J. K. Ryan (2012). Accuracy-enhancement of discontinuous Galerkin solutions
for convection-diffusion equations in multiple-dimensions. Math. Comp., 81(280):1929–
1950.
L. Ji, Y. Xu and J. K. Ryan (2013). Negative-order norm estimates for nonlinear hyperbolic
conservation laws. J. Sci. Comput., 54(2-3):531–548.
G.-S. Jiang and C.-W. Shu (1994). On a cell entropy inequality for discontinuous Galerkin
methods. Math. Comp., 62:531–538.
C. Johnson, U. Nävert and J. Pitkäranta (1984). Finite element methods for linear hyperbolic
problems. Comput. Methods Appl. Mech. Engrg., 45:285–312.
C. Johnson and J. Pitkäranta (1986). An analysis of the discontinuous Galerkin method for a
scalar hyperbolic equation. Math. Comp., 46:1–26.
M. Juntunen and R. Stenberg (2008). On a mixed discontinuous Galerkin method. Electron.
Trans. Numer. Anal., 32:17–32.
G. Kanschat (2003). Preconditioning Methods for local discontinuous Galerkin Discretizations.
SIAM J. Sci. Comp., 25:815–831.
G. Kanschat (2008). Discontinuous Galerkin methods for viscous incompressible flows.
Teubner, Deutscher Universits-Verlag, GWV Fachverlage GmbH, Wiesbaden.
G. Kanschat and R. Rannacher (2002). Local error analysis of the interior penalty
discontinuous Galerkin method for second order elliptic problems. J. Numer. Math.,
10(4):249–274.
O. A. Karakashian and W.N. Jureidini (1998). A nonconforming finite element method for
the stationary Navier-Stokes equations. SIAM J. Numer. Anal., 35:93–120.
O. A. Karakashian and T. Katsaounis (2000). A discontinuous Galerkin method for the
incompressible Navier-Stokes equations. In Cockburn, B., Karniadakis, G., and Shu,
C.-W., editors, Discontinuous Galerkin Methods. Theory, Computation and Applications,
volume 11 of Lect. Notes Comput. Sci. Engrg., pages 157–166, Berlin. Springer Verlag.
O. A. Karakashian and F. Pascal (2003). A posteriori error estimates for a discontinuous
Galerkin approximation of a second order elliptic problems. SIAM J. Numer. Anal.,
41:2374–2399.
O. A. Karakashian and F. Pascal (2007). Convergence of adaptive discontinuous Galerkin
approximations of second-order elliptic problems. SIAM J. Numer. Anal., 45(2):641–665
(electronic).
G.E. Karniadakis and S.J. Sherwin (1999). Spectral/hp Element Methods in CFD. Oxford
University Press.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

104 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

D.S. Kershaw, M.K. Prasad, M.J. Shaw and J.L. Milovich (1998). 3D unstructured mesh ALE
hydrodynamics with the upwind discontinuous Galerkin method. Comput. Methods Appl.
Mech. Engrg., 158:81–116.
H. H. Kim, E. T. Chung and C. S. Lee (2013). A staggered discontinuous Galerkin method
for the Stokes system. SIAM J. Numer. Anal., 51(6):3327–3350.
R.M. Kirby, S.J. Sherwin and B. Cockburn (2012). To HDG or to CG: A comparative study.
J. Sci. Comput., 51(1):183–212.
C.M. Klaij, J. J. W. van der Vegt and H. van der Ven (2006). Space-Time Discontinuous
Galerkin method for the compressible Navier-Stokes equations. J. Comput. Phys.,
217:589–611.
L. Krivodonova (2007). Limiters for high-order discontinuous Galerkin methods. J. Comput.
Phys., 226(1):879–896.
L. Krivodonova and R. Qin (2013a). An analysis of the spectrum of the discontinuous Galerkin
method. Appl. Numer. Math., 64:1–18.
L. Krivodonova and R. Qin (2013b). An analysis of the spectrum of the discontinuous Galerkin
method II: Nonuniform grids. Appl. Numer. Math., 71:41–62.
R. J. Labeur and G. N. Wells (2007). A Galerkin interface stabilisation method for the
advection-diffusion and incompressible Navier-Stokes equations. Comput. Methods Appl.
Mech. Engrg., 196(49-52):4985–5000.
R. J. Labeur and G. N. Wells (2012). Energy stable and momentum conserving hybrid finite
element method for the incompressible Navier-Stokes equations. SIAM J. Sci. Comput.,
34(2):A889–A913.
C. Lasser and A. Toselli (2000). An overlapping domain decomposition preconditioner for a
class of discontinuous Galerkin approximations of advection-diffusion problems. Technical
Report 2000-12, Seminar für Angewandte Mathematik, ETH Zürich.
C. Lehrenfeld (2010). Hybrid discontinuous Galerkin methods for solving incompressible flow
problems. PhD thesis, Diplomigenieur Rheinisch-Westfalishen Technischen Hochchule
Aachen.
C. Lehrenfeld and J. Schöberl (2015). High order exactly divergence-free hybrid discontinuous
Galerkin methods for unsteady incompressible flows. ASC Report No. 27.
P. Lesaint and P. A. Raviart (1974). On a finite element method for solving the neutron
transport equation. In de Boor, C., editor, Mathematical aspects of finite elements in
partial differential equations, pages 89–145. Academic Press.
B. Q. Li (2006). Discontinuous finite elements in fluid dynamics and heat transfer.
Computational Fluid and Solid Mechanics. Springer-Verlag London, Ltd., London.
X. Li, J. K. Ryan, R. M. Kirby and C. Vuik (2016). Smoothness-increasing accuracy-conserving
(SIAC) filters for derivative approximations of discontinuous Galerkin (DG) solutions over
nonuniform meshes and near boundaries. J. Comput. Appl. Math., 294:275–296.
H. Liu and J. Yan (2010). The direct discontinuous Galerkin (DDG) method for diffusion with
interface corrections. Commun. Comput. Phys., 8(3):541–564.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 105

I. Lomtev and G.E. Karniadakis (1999). A discontinuous Galerkin method for the Navier-
Stokes equations. Int. J. Numer. Meth. Fluids, 29:587–603.
C. Lovadina and L. D. Marini (2009). A-posteriori error estimates for discontinuous Galerkin
approximations of second order elliptic problems. J. Sci. Comput., 40(1-3):340–359.
R.B. Lowrie (1996). Compact higher-order numerical methods for hyperbolic conservation laws.
PhD thesis, University of Michigan.
R. B. Lowrie, P. L. Roe and B. van Leer (1995). A Space-Time Discontinuous Galerkin
Method for the Time-Accurate Numerical Solution of Hyperbolic Conservation Laws. In
Proceedings of the 12th AIAA Computational Fluid Dynamics Conference. Paper 95-1658.
R. B. Lowrie, P. L. Roe and B. van Leer (1998). Space-Time Methods for Hyperbolic
Conservation Laws. In Barriers and Challenges in Computational Fluid Dynamics,
volume 6 of ICASE/LaRC Interdisciplinary Series in Science and Engineering, pages
79–98. Kluwer.
J. Luo, C.-W. Shu and Q. Zhang (2015). A priori error estimates to smooth solutions of the
third order Runge-Kutta discontinuous Galerkin method for symmetrizable systems of
conservation laws. ESAIM Math. Model. Numer. Anal., 49(4):991–1018.
X. Meng, C.-W. Shu and B. Wu (2016). Optimal error estimates for discontinuous Galerkin
methods based on upwind-biased fluxes for linear hyperbolic equations. Math. Comp.,
85(299):1225–1261.
H. Mirzaee, L. Ji, J. K. Ryan and R. M. Kirby (2011). Smoothness-increasing accuracy-
conserving (SIAC) postprocessing for discontinuous Galerkin solutions over structured
triangular meshes. SIAM J. Numer. Anal., 49(5):1899–1920.
H. Mirzaee, J. King, J. K. Ryan and R. M. Kirby (2013). Smoothness-increasing accuracy-
conserving filters for discontinuous Galerkin solutions over unstructured triangular
meshes. SIAM J. Sci. Comput., 35(1):A212–A230.
H. Mirzaee, J. K. Ryan and R. M. Kirby (2014). Smoothness-increasing accuracy-conserving
(SIAC) filters for discontinuous Galerkin solutions: application to structured tetrahedral
meshes. J. Sci. Comput., 58(3):690–704.
A. Montlaur, S. Fernández-Méndez and A. Huerta (2008). Discontinuous Galerkin methods for
the Stokes equations using divergence-free approximations. Internat. J. Numer. Methods
Fluids, 57(9):1071–1092.
A. Montlaur, S. Fernández-Méndez and A. Huerta (2011). Higher-order implicit Runge-Kutta
methods for unsteady incompressible flows. Rev. Internac. Métod. Numér. Cálc. Diseñ.
Ingr., 27(1):77–91.
A. Montlaur, S. Fernández-Méndez, J. Peraire and A. Huerta (2010). Discontinuous Galerkin
methods for the Navier-Stokes equations using solenoidal approximations. Internat. J.
Numer. Methods Fluids, 64(5):549–564.
J.-C. Nédélec (1980). Mixed finite elements in R3 . Numer. Math., 35:315–341.
J.-C. Nédélec (1986). A new family of mixed finite elements in R3 . Numer. Math., 50:57–81.
N. C. Nguyen, J. Peraire and B. Cockburn (2009a). An implicit high-order hybridizable
discontinuous Galerkin method for linear convection-diffusion equations. J. Comput.
Phys., 228:3232–3254.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

106 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

N. C. Nguyen, J. Peraire and B. Cockburn (2009b). An implicit high-order hybridizable


discontinuous Galerkin method for nonlinear convection-diffusion equations. J. Comput.
Phys., 228:8841–8855.
N. C. Nguyen, J. Peraire and B. Cockburn (2010a). A comparison of HDG methods for Stokes
flow. J. Sci. Comput., 45:215–237.
N. C. Nguyen, J. Peraire and B. Cockburn (2010b). A hybridizable discontinuous Galerkin
method for Stokes flow. Comput. Methods Appl. Mech. Engrg., 199:582–597.
N. C. Nguyen, J. Peraire and B. Cockburn (2010c). A Hybridizable Discontinuous Galerkin
Method for the Incompressible Navier-Stokes Equations (AIAA Paper 2010-362). In
Proceedings of the 48th AIAA Aerospace Sciences Meeting, Exhibit, Orlando, Florida.
N. C. Nguyen, J. Peraire and B. Cockburn (2011). An implicit high-order hybridizable
discontinuous Galerkin method for the incompressible Navier-Stokes equations. J.
Comput. Phys., 230:1147–1170.
N. C. Nguyen, J. Peraire and B. Cockburn (2015). A class of embedded discontinuous Galerkin
methods for computational fluid dynamics. J. Comput. Phys., 302:674–692.
N. C. Nguyen, P.-O. Persson and J. Peraire (2007). RANS solutions using high order discontin-
uous Galerkin methods. In 45th. AAIA Aerospace Sciences Meeting and Exhibit, Reno,
Nevada,.
A. Nigro, C. De Bartolo, F. Bassi and A. Ghidoni (2014a). Up to sixth-order accurate A-
stable implicit schemes applied to the discontinuous Galerkin discretized Navier-Stokes
equations. J. Comput. Phys., 276:136–162.
A. Nigro, A. Ghidoni, S. Rebay and F. Bassi (2014b). Modified extended BDF scheme for
the discontinuous Galerkin solution of unsteady compressible flows. Internat. J. Numer.
Methods Fluids, 76(9):549–574.
M. Ohlberger (2009). A review of a posteriori error control and adaptivity for approximations
of non-linear conservation laws. Internat. J. Numer. Methods Fluids, 59(3):333–354.
I. Oikawa (2015). A hybridized discontinuous Galerkin method with reduced stabilization. J.
Sci. Comput., 65:327–340.
I. Oikawa (2016). Analysis of a Reduced-Order HDG Method for the Stokes Equations. J.
Sci. Comput., 67(2):475–492.
S. Osher (1984). Convergence of Generalized MUSCL Schemes. SIAM J. Numer. Anal.,
22:947–961.
P. E. Castillo, Paul and E. S. Velázquez (2008). A numerical study of a semi-algebraic multilevel
preconditioner for the local discontinuous Galerkin method. Internat. J. Numer. Methods
Engrg., 74(2):255–268.
J. Peraire, N. C. Nguyen and B. Cockburn (2010). A Hybridizable Discontinuous Galerkin
Method for the compressible Euler and Navier-Stokes Equations (AIAA Paper 2010-
363). In Proceedings of the 48th AIAA Aerospace Sciences Meeting and Exhibit, Orlando,
Florida.
J. Peraire and P.-O. Persson (2008). The compact discontinuous Galerkin (CDG) method for
elliptic problems. SIAM J. Sci. Comput., 30(4):1806–1824.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 107

P.-O. Persson and J. Peraire (2008). Newton-GMRES preconditioning for discontinuous


Galerkin discretizations of the Navier-Stokes equations. SIAM J. Sci. Comput.,
30(6):2709–2733.
I. Perugia and D. Schötzau (2001). On the coupling of local discontinuous Galerkin and
conforming finite element methods. J. Sci. Comput., 16:411–433.
T. Peterson (1991). A note on the convergence of the Discontinuous Galerkin method for a
scalar hyperbolic equation. SIAM J. Numer. Anal., 28:133–140.
T. Qin, C.-W. Shu and Y. Yang (2016). Bound-preserving discontinuous Galerkin methods
for relativistic hydrodynamics. J. Comput. Phys., 315:323–347.
J. Qiu and C.-W. Shu (2005). A comparison of troubled-cell indicators for Runge-Kutta
discontinuous Galerkin methods using weighted essentially nonoscillatory limiters. SIAM
J. Sci. Comput., 27(3):995–1013.
R. Lazarov, S. Repin and S. K. Tomar (2009). Functional a posteriori error estimates for
discontinuous Galerkin approximations of elliptic problems. Numer. Methods Partial
Differential Equations., 25:952–971.
P. A. Raviart and J. M. Thomas (1977). A mixed finite element method for second order
elliptic problems. In Galligani, I. and Magenes, E., editors, Mathematical Aspects of Finite
Element Method, Lecture Notes in Math. 606, pages 292–315. Springer-Verlag, New York.
W.H. Reed and T.R. Hill (1973). Triangular mesh methods for the neutron transport equation.
Technical Report LA-UR-73-479, Los Alamos Scientific Laboratory.
S. Rhebergen and B. Cockburn (2012). A space-time hybridizable discontinuous Galerkin
method for incompressible flows on deforming domains. J. Comput. Phys., 231(11):4185–
4204.
S. Rhebergen and B. Cockburn (2013). Space-time hybridizable discontinuous Galerkin method
for the advection-diffusion equation on moving and deforming meshes. In The Courant-
Friedrichs-Lewy (CFL) condition, pages 45–63. Birkhäuser/Springer, New York.
S. Rhebergen and B. Cockburn and van der Vegt, Jaap J. W. (2013). A space-time
discontinuous Galerkin method for the incompressible Navier-Stokes equations. J.
Comput. Phys., 233:339–358.
G.R. Richter (1988). An optimal-order error estimate for the Discontinuous Galerkin method.
Math. Comp., 50:75–88.
G. R. Richter (2008). On the order of convergence of the discontinuous Galerkin method for
hyperbolic equations. Math. Comp., 77:1871–1885.
B. Rivière (2008). Discontinuous Galerkin methods for solving elliptic and parabolic equations,
volume 35 of Frontiers in Applied Mathematics. Society for Industrial and Applied
Mathematics (SIAM), Philadelphia, PA. Theory and implementation.
B. Rivière and M. F. Wheeler (2003). A posteriori error estimates for a discontinuous Galerkin
method applied to elliptic problems. Log number: R74. Comput. Math. Appl., 46(1):141–
163. p-FEM2000: p and hp finite element methods—mathematics and engineering practice
(St. Louis, MO).

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

108 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

B. Rivière, M. F. Wheeler and V. Girault (1999). Improved energy estimates for interior
penalty, constrained and discontinuous Galerkin methods for elliptic problems. Part I.
Comp. Geosci., 3:337–360.
J. Ryan and C.-W. Shu (2003). On a one-sided post-processing technique for the discontinuous
Galerkin methods. Methods and Applications of Analysis, 10:295–307.
J. K. Ryan and B. Cockburn (2009). Local derivative post-processing for the discontinuous
Galerkin method. J. Comput. Phys., 228(23):8642–8664.
J. K. Ryan, X. Li, R. M. Kirby and K. Vuik (2015). One-sided position-dependent smoothness-
increasing accuracy-conserving (SIAC) filtering over uniform and non-uniform meshes. J.
Sci. Comput., 64(3):773–817.
R. Sanders (1983). On convergence of monotone finite difference schemes with variable spacing
differencing. Math. Comp., 40:91–106.
D. Schötzau, C. Schwab and A. Toselli (2003a). hp-DGFEM for incompressible flows. SIAM
J. Numer. Anal., 40:2171–2194.
D. Schötzau, C. Schwab and A. Toselli (2003b). Stabilized hp-DGFEM for incompressible flow.
Math. Models Methods Appl. Sci., 13:1413–1436.
J. Schütz and G. May (2013). A hybrid mixed method for the compressible Navier-Stokes
equations. J. Comput. Phys., 240:58–75.
L. R. Scott and M. Vogelius (1985). Norm estimates for a maximal right inverse of the
divergence operator in spaces of piecewise polynomials. RAIRO Modél. Math. Anal.
Numér., 19:111–143.
S.J. Sherwin (2000). Dispersion analysis of the continuous and discontinuous Galerkin
formulations. In Cockburn, B., Karniadakis, G., and Shu, C.-W., editors, Discontinuous
Galerkin Methods. Theory, Computation and Applications, volume 11 of Lect. Notes
Comput. Sci. Engrg., pages 425–431, Berlin. Springer Verlag.
A.I. Shestakov, J.L. Milovich and M.K. Prasad (2001). Combining cell- and point-centered
methods in 3D, unstructured-grid hydrodynamics codes. J. Comput. Phys., 170:81–111.
A.I. Shestakov, M.K. Prasad, J.L. Milovich, N.A. Gentile, J.F. Painter and G. Furnish (2000).
The radition-hydrodynamic ICF3D code. Comput. Methods Appl. Mech. Engrg., 187:181–.
C.-W. Shu (1988). TVD time discretizations. SIAM J. Sci. Stat. Comput., 9:1073–1084.
C.-W. Shu (2009). Discontinuous Galerkin methods: general approach and stability. In
Numerical solutions of partial differential equations, Adv. Courses Math. CRM Barcelona,
pages 149–201. Birkhäuser, Basel.
C.-W. Shu (2014). Discontinuous Galerkin method for time-dependent problems: survey and
recent developments. In Recent developments in discontinuous Galerkin finite element
methods for partial differential equations, volume 157 of IMA Vol. Math. Appl., pages
25–62. Springer, Cham.
C.-W. Shu and S. Osher (1988). Efficient implementation of essentially non-oscillatory shock-
capturing methods. J. Comput. Phys., 77:439–471.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 109

C.-W. Shu and Q. Zhang (2004). Error estimates to smooth solutions of Runge-Kutta
discontinuous Galerkin methods for scalar conservation laws. SIAM J. Numer. Anal.,
42:641–666.
P. R. Spalart and S. R. Allmaras (1994). A one-equation turbulence model for aerodynamic
flows. La Recherche Aerospatiale, 1:5–21.
R. Stenberg (1988). A Family of Mixed Finite Elements for the Elasticity Problem. Numer.
Math., 53:513–538.
R. Stenberg (1991). Postprocessing schemes for some mixed finite elements. RAIRO Modél.
Math. Anal. Numér., 25:151–167.
E. Süli and P. Houston (2002). Adaptive Finite Element Approximation of Hyperbolic
Problems. In Barth, T. and Deconink, H., editors, Error Estimation and Adaptive
Discretization Methods in Computational Fluid Dynamics, volume 25 of Lect. Notes
Comput. Sci. Engrg., pages 269–344. Springer Verlag, Berlin.
R. Témam (1966). Sur l’approximation des solutions des équations de Navier-Stokes. C. R.
Acad. Sci. Paris Sér. A, 216:219–221.
R. Témam (1968). Une méthode d’approximation de la solutions des équations de Navier-
Stokes. Bull. Soc. Math. France, 98:115–152.
R. Témam (1979). Navier-Stokes equations, volume 2 of Studies in Mathematics and its
Applications. North-Holland Publishing Co., Amsterdam-New York, revised edition.
Theory and numerical analysis, With an appendix by F. Thomasset.
F. Thomasset (1981). Implementation of finite element methods for Navier-Stokes equations.
Springer Series in Computational Physics. Springer Verlag, New York.
V. Thomée (1977). High order local approximations to derivatives in the finite element method.
Math. Comp., 31:652–660.
J. J. W. van der Vegt and H. van der Ven (2002a). Slip flow boundary conditions in
discontinuous Galerkin discretizations of the Euler equations of gas dynamics. In Fifth
World Congress on Computational Mechanics, July 7-12, Vienna, Austria.
J. J. W. van der Vegt and H. van der Ven (2002b). Space-Time Discontinuous Galerkin Finite
Element Method with Dynamic Grid Motion for Inviscid Compressible Flows: I. General
Formulation. J. Comput. Phys., 182:546–585.
H. van der Ven and J. J. W. van der Vegt (2002). Space-Time Discontinuous Galerkin Finite
Element Method with Dynamic Grid Motion for Inviscid Compressible Flows: II. Efficient
flux quadrature. Comput. Methods Appl. Mech. Engrg., 191:4747–4780.
B. van Leer (1974). Towards the ultimate conservation difference scheme, II. J. Comput.
Phys., 14:361–376.
B. van Leer (1979). Towards the ultimate conservation difference scheme, V. J. Comput.
Phys., 32:1–136.
P. van Slingerland, J. K. Ryan and C. Vuik (2011). Position-dependent smoothness-increasing
accuracy-conserving (SIAC) filtering for improving discontinuous Galerkin solutions.
SIAM J. Sci. Comput., 33(2):802–825.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

110 ENCYCLOPEDIA OF COMPUTATIONAL MECHANICS

F. Vilar, C.-W. Shu and P.-H. Maire (2016a). Positivity-preserving cell-centered Lagrangian
schemes for multi-material compressible flows: From first-order to high-orders. Part I: The
one-dimensional case. J. Comput. Phys., 312:385–415.
F. Vilar, C.-W. Shu and P.-H. Maire (2016b). Positivity-preserving cell-centered Lagrangian
schemes for multi-material compressible flows: From first-order to high-orders. Part II:
The two-dimensional case. J. Comput. Phys., 312:416–442.
D. Walfisch, J. K. Ryan, R. M. Kirby and R. Haimes (2009). One-sided smoothness-increasing
accuracy-conserving filtering for enhanced streamline integration through discontinuous
fields. J. Sci. Comput., 38(2):164–184.
H. Wang, C.-W. Shu and Q. Zhang (2015). Stability and error estimates of local discontinuous
Galerkin methods with implicit-explicit time-marching for advection-diffusion problems.
SIAM J. Numer. Anal., 53(1):206–227.
H. Wang, C.-W. Shu and Q. Zhang (2016). Stability analysis and error estimates of
local discontinuous Galerkin methods with implicit-explicit time-marching for nonlinear
convection-diffusion problems. Appl. Math. Comput., 272(part 2):237–258.
W. Wang, J. Guzmán and C.-W. Shu (2011). The multiscale discontinuous Galerkin method
for solving a class of second order elliptic problems with rough coefficients. Int. J. Numer.
Anal. Model., 8(1):28–47.
T.C. Warburton (1998). Spectral/hp methods on polymorphic multi-domains: Algorithms and
Applications. PhD thesis, Brown University.
M. Wierse (1997). A new theoretically motivated higher order upwind scheme on unstructured
grids of simplices. Adv. Comput. Math., 7:303–335.
P. Woodward and P. Colella (1984). The numerical simulation of two-dimensional fluid flow
with strong shocks. J. Comput. Phys., 54:115–173.
M. Woopen, G. May and J. Schütz (2014). Adjoint-based error estimation and mesh adaptation
for hybridized discontinuous Galerkin methods. Internat. J. Numer. Methods Fluids,
76(11):811–834.
L. Yin, A. Acharya, N. Sobh, R.B.Haber and D.A.Tortorelli (2000). A space-time discontinuous
Galerkin method for elastodynamic analysis. In Cockburn, B., Karniadakis, G., and Shu,
C.-W., editors, Discontinuous Galerkin Methods. Theory, Computation and Applications,
volume 11 of Lect. Notes Comput. Sci. Engrg., pages 459–464, Berlin. Springer Verlag.
Q. Zhang and C.-W. Shu (2006). Error estimates to smooth solutions of Runge-Kutta
discontinuous Galerkin method for symmetrizable systems of conservation laws. SIAM J.
Numer. Anal., 44(4):1703–1720 (electronic).
Q. Zhang and C.-W. Shu (2010a). Stability analysis and a priori error estimates of the third
order explicit Runge-Kutta discontinuous Galerkin method for scalar conservation laws.
SIAM J. Numer. Anal., 48.
Q. Zhang and C.-W. Shu (2014). Error estimates for the third order explicit Runge-Kutta
discontinuous Galerkin method for a linear hyperbolic equation in one-dimension with
discontinuous initial data. Numer. Math., 126(4):703–740.
X. Zhang and C.-W. Shu (2010b). On maximum-principle-satisfying high order schemes for
scalar conservation laws. J. Comput. Phys., 229(9):3091–3120.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

DISCONTINUOUS GALERKIN METHODS FOR COMPUTATIONAL FLUID DYNAMICS 111

X. Zhang and C.-W. Shu (2010c). On positivity-preserving high order discontinuous Galerkin
schemes for compressible Euler equations on rectangular meshes. J. Comput. Phys.,
229(23):8918–8934.
X. Zhang and C.-W. Shu (2011). Positivity-preserving high order discontinuous Galerkin
schemes for compressible Euler equations with source terms. J. Comput. Phys.,
230(4):1238–1248.
X. Zhang, Y. Xia and C.-W. Shu (2012). Maximum-principle-satisfying and positivity-
preserving high order discontinuous Galerkin schemes for conservation laws on triangular
meshes. J. Sci. Comput., 50.
Y. Zhang, X. Zhang and C.-W. Shu (2013). Maximum-principle-satisfying second order
discontinuous Galerkin schemes for convection-diffusion equations on triangular meshes.
J. Comput. Phys., 234:295–316.
J. Zhu, X. Zhong, C.-W. Shu and J. Qiu (2013). Runge-Kutta discontinuous Galerkin method
using a new type of WENO limiters on unstructured meshes. J. Comput. Phys., 248:200–
220.

Encyclopedia of Computational Mechanics. Edited by Erwin Stein, René de Borst and Thomas J.R. Hughes.
c John Wiley & Sons, Ltd. ISBN: 0-470-84699-2

Static condensation, hybridization, and the
devising of the HDG methods

Bernardo Cockburn

Abstract In this paper, we review and refine the main ideas for devising the so-called
hybridizable discontinuous Galerkin (HDG) methods; we do that in the framework
of steady-state diffusion problems. We begin by revisiting the classic techniques of
static condensation of continuous finite element methods and that of hybridization
of mixed methods, and show that they can be reinterpreted as discrete versions of
a characterization of the associated exact solution in terms of solutions of Dirich-
let boundary-value problems on each element of the mesh which are then patched
together by transmission conditions across interelement boundaries. We then define
the HDG methods associated to this characterization as those using discontinuous
Galerkin (DG) methods to approximate the local Dirichlet boundary-value prob-
lems, and using weak impositions of the transmission conditions. We give simple
conditions guaranteeing the existence and uniqueness of their approximate solu-
tions, and show that, by their very construction, the HDG methods are amenable to
static condensation. We do this assuming that the diffusivity tensor can be inverted;
we also briefly discuss the case in which it cannot. We then show how a differ-
ent characterization of the exact solution, gives rise to a different way of statically
condensing an already known HDG method. We devote the rest of the paper to es-
tablishing bridges between the HDG methods and other methods (the old DG meth-
ods, the mixed methods, the staggered DG method and the so-called Weak Galerkin
method) and to describing recent efforts for the construction of HDG methods (one
for systematically obtaining superconvergent methods and another, quite different,
which gives rise to optimally convergent methods). We end by providing a few bib-
liographical notes and by briefly describing ongoing work.

Bernardo Cockburn
School of Mathematics, University of Minnesota, e-mail: cockburn@math.umn.edu
Submitted on June 30, 2015. To appear in Building Bridges: Connections and Challenges in
Modern Approaches to Numerical Partial Differential Equations, G.R.Barrenechea, F.Brezzi,
A.Cagniani and E.H.Georgulis, eds., Lecture Notes in Computational Science and Engineer-
ing 114, pages 129-177.
The LMS Durham Symposia (of the same title) was funded by the London Mathematical Society
and took place at Durham, U.K., on July 7–16, 2014.

1
2 Bernardo Cockburn

1 Introduction

In this paper, we give a short introduction to the devising of the hybridizable dis-
continuous Galerkin (HDG) in the framework of the following steady-state diffusion
model problem:

c q + ∇u = 0 in Ω ⊂ Rd , (1a)
∇·q = f in Ω , (1b)
u = uD on ∂ Ω . (1c)

We assume that the data c , f and uD are smooth functions such that the solution itself
is smooth. Here c is a matrix-valued function which is symmetric and uniformly
positive definite on Ω . We are going to closely follow [24], where the HDG methods
were introduced.
Since the HDG methods are discontinuous Galerkin (DG) methods, [40], we
begin by defining the DG methods for the above boundary-value problem; we follow
[3]. Let us first discretize the domain Ω . We denote a triangulation of the domain
Ω by Ωh := {K} and set ∂ Ωh := {∂ K : K ∈ Ωh }. The outward unit normal to the
element K is denoted by n. The set of faces of the element K is denoted by F(K).
An interior face F of the triangulation Ωh is any set of the form ∂ K + ∩ ∂ K − , where
K ± are elements of Ωh ; we assume that the (d − 1)-Lebesgue measure of F is not
zero. The set of all interior faces is denoted by Fhi . Similarly, a boundary face F of
the triangulation Ωh is any set of the form ∂ K ∩ ∂ Ω , where K are elements of Ωh ;
again, we assume that the (d − 1) Lebesgue measure of F is not zero. The set of
all boundary faces is denoted by Fh∂ . The set of interior and boundary faces of the
triangulation is denoted by Fh .
The notation associated to the weak formulation of the method is the following.
We set
(·, ·)Ωh := ∑ (·, ·)K and h·, ·i∂ Ωh := ∑ h·, ·i∂ K ,
K∈Ωh K∈Ωh

where (·, ·)K denotes the standard L2 (K)-inner product, and h·, ·i∂ K denotes the stan-
dard L2 (∂ K)-inner product.
We can now introduce the general form of a DG method. The approximate solu-
tion (qh , uh ) given by a DG method is the element of the space Vh ×Wh , where

Vh := {v ∈ L2 (Ω ) :v|K ∈ V(K) ∀K ∈ Ωh },
Wh := {w ∈ L2 (Ω ) :w|K ∈ W (K) ∀K ∈ Ωh }.

satisfying the equations

(c qh , v)Ωh − (uh , ∇ · v)Ωh + hb


uh , v · ni∂ Ωh = 0,
−(qh , ∇w)Ωh + hb
qh · n, wi∂ Ωh = ( f , w)Ωh ,
Static condensation, hybridization, and the devising of the HDG methods 3

for all (v, w) ∈ Vh ×Wh , where the numerical traces ubh and q bh · n are approximations
to u|∂ Ωh and q · n|∂ Ωh , respectively. The finite dimensional space Vh ×Wh is chosen
so that all the integrals in the above weak formulation are well defined.
It remains to discuss how to choose the numerical traces. To do that, let us begin
by introducing some useful notation. The traces of the functions ζ and z defined
on K ± ∈ Ωh on the boundary ∂ K ± are denoted by ζ ± and z± , respectively. We use
the same notation if the functions ζ and z are defined on ∂ Ωh . Thus, we define the
jumps of ζ and z across the interior face F = ∂ K + ∩ ∂ K − by

[[ζ ]] := ζ + n+ + ζ − n− and [[z]] := z + · n+ + z − · n− ,

respectively, where n± is the outward unit normal to K ± . On boundary faces F, we


simply write

[[ζ ]] := ζ n and [[z]] := z · n,

with the obvious notation. We say that the numerical traces are single-valued if, on
Fhi , [[b
uh ]] = 0 and [[b
qh ]] = 0.
Slightly extending what was done in [3], the numerical traces ubh and (the nor-
mal component of) q bh are linear mappings ubh : H1 (Ωh ) × H 1 (Ωh ) → L2 (∂ Ωh )
bh : H (Ωh ) × H (Ωh ) → L2 (∂ Ωh ) which approximate the traces of u and (the nor-
q 1 1

mal component of) q on ∂ Ωh , respectively. We take these numerical traces to be


consistent. We say that they are consistent if

ubh (−a∇v, v) = v|∂ Ωh , bh (−a∇v, v) · n = −(a∇v) · n|∂ Ωh ,


q

whenever [[a∇v]] = 0 and [[v]] = 0 on the interior faces Fhi . Here a := c −1 . This
completes the description of the DG methods.
The HDG methods are the DG methods just described which are amenable to
static condensation. They are thus efficiently implementable and turn out to be more
accurate than its predecessors in many instances. None of them fit in the unify-
ing framework developed in [3], since the numerical trace ubh of the HDG methods
depends on the approximate flux too. The family of DG methods analyzed in [4]
includes some HDG methods.
The paper is organized as follows. In section 2, we show that the classic tech-
niques of static condensation of continuous finite element methods and that of hy-
bridization of mixed methods, introduced back in 1965 in [56] and [53], respec-
tively, can be reinterpreted as discrete versions of a characterization of the asso-
ciated exact solution expressed in terms of solutions of Dirichlet boundary-value
problems on each element of the mesh patched together by transmission conditions
across interelement boundaries. In section 3, we use this reinterpretation to define
the HDG methods associated to this characterization as those using discontinuous
Galerkin (DG) methods to approximate the local Dirichlet boundary-value prob-
lems, and using weak impositions of the transmission conditions. We show that, by
construction, the global problem of these HDG methods only involves the approx-
imation to the trace of the scalar variable on the faces of the triangulation. We do
4 Bernardo Cockburn

this assuming that the diffusivity tensor a is invertible; in section 4, we show that it
is trivial to treat the case in which it is not. In section 5, we show that a new char-
acterization of the exact solution, based on the elementwise solution of Neumann
boundary-value problems, can be used to produce a different type of static conden-
sation of already known HDG methods. In section 6, we establish bridges between
the HDG and several other methods and comment on two promising ways of de-
vising new HDG methods. We end by providing a few bibliographical notes and by
briefly describing ongoing work.

Note to the reader

Engineering and Mathematics Graduate Students interested in numerical methods


for partial differential equations should be able to read this paper. An elementary
background in finite element methods should be enough since here we focus on the
ideas guiding the devising of the methods rather than in their rigorous error analyses.
The material of these notes is strongly related to the one presented at the Durham
Symposium entitled ”Building bridges: Connections and challenges in modern ap-
proaches to numerical partial differential equations” at Durham, U.K., July 8-16,
2014, sponsored by the London Mathematical Society, and EPSRC. I would like
to express my gratitude to the organizers, especially to G.R. Barrenechea and E.
Georgoulis, for the invitation to talk about HDG methods at that meeting.
These notes have evolved from several short courses the author has given: at
the Basque Center of Applied Mathematics, Bilbao, Spain, July 9–17, 2009; at the
University of Pavia, May 28–June 1, 2012; at the Department of Mathematics &
Statistics of the King Fahad University of Petroleum and Minerals, Dec. 2012; at
the International Center for Numerical Methods in Engineering, and Universidad
Polytecnica de Catalunya, Barcelona, Spain, July 11–15, 2012; at the US National
Conference on Computational Mechanics 12, Raleigh, North Carolina, July 22–25,
2013; and at the Department of Mathematics of the Chinese University of Hong
Kong, March 19–21, 2014.

2 Static condensation and hybridization

Here we argue that the static condensation of the continuous Galerkin method, an
implementation technique introduced by R.J. Guyan 1965 in [56], can be reinter-
preted as a discrete version of a characterization of the exact solution. We also ar-
gue that a similar interpretation can be given to the static condensation of a mixed
method as proposed by Fraejis de Veubeque also in 1965 [53], who showed that this
can be achieved provided the mixed method is hybridized first. Although the above-
mentioned procedures were carried out in the setting of linear elasticity, we present
them for our simpler model problem of steady-state diffusion (1).
Static condensation, hybridization, and the devising of the HDG methods 5

We proceed as follows. First, we present a characterization of the exact solution


in terms of solutions of local problems patched together by means of transmission
and boundary conditions. We then show how the original static condensation of
the continuous Galerkin method and that of a mixed method can be thought of as
discrete versions of such characterization.

2.1 Static condensation of the exact solution

2.1.1 A characterization of the exact solution

Here, for any given triangulation Ωh := {K} of Ω , we give a characterization of the


exact solution in terms of solutions on each of the elements K ∈ Ωh , and a single
global problem expressed in terms of transmission and boundary conditions.
Suppose that, for each element K ∈ Ωh , we define (Q, U) as the solution of the
local problem

c Q + ∇U = 0 in K,
∇·Q = f in K,
U = ub on ∂ K,

where we want the single-valued function ub to be such that (Q, U) = (q, u) on each
element K ∈ Ωh . We know that this happens if and only if ub enforces the following
transmission and boundary conditions:

[[Q]] = 0 on F ∈ Fhi ,
ub = uD on F ∈ Fh∂ .

If we now separate the influence of ub form that of f , we can easily see that we
obtained the following result.

Theorem 1 (Characterization of the exact solution). We have that

(q, u) = (Q, U) = (Qub, Uub) + (Q f , U f ),

where, on the element K ∈ Ωh , (Qub, Uub) and (Q f , U f ) are the solutions of

c Qub + ∇Uub = 0 in K, c Q f + ∇U f = 0 in K,
∇ · Qub = 0 in K, ∇·Qf = f in K,
Uub = ub on ∂ K, Uf = 0 on ∂ K,

and where ub is the single-valued function solution of


6 Bernardo Cockburn

− [[Qub]] = [[Q f ]] if F ∈ Fhi ,


ub = uD if F ∈ Fh∂ .

2.1.2 An example

Let us illustrate this result with a simple but revealing case. Take Ω := (0, 1) with
K = (xi−1 , xi ) for i = 1, . . . , N where x0 = 0 and xN = 1. For simplicity, we take c to
be a constant. We then have that

(q, u) = (Qub, Uub) + (Q f , U f ),

where, for i = 1, . . . , N, the functions (Qub, Uub) and (Q f , U f ) are the solutions of the
local problem

d d
c Qub + U =0 in (xi−1 , xi ), cQf + Uf = 0 in (xi−1 , xi ),
dx ub dx
d d
Q =0 in (xi−1 , xi ), Q =f in (xi−1 , xi ),
dx ub dx f
Uub = ub on {xi−1 , xi }, Uf = 0 on {xi−1 , xi }.

Note that we still do not know the actual values of the function ub : {xi }Ni=0 7→ R,
but once we obtain them, we can readily get the exact solution (q, u). To find those
values, we only have to solve the global problem

−Qub(xi− ) + Qub(xi+ ) = Q f (xi− ) − Q f (xi+ ) for i = 1, . . . , N − 1,


ub(xi ) = uD (xi ) for i = 0, N.

Now, let us solve the local problems and then find the global problem. A simple
computation gives that the solutions of the local problems are
Z xi
ubi − ubi−1
Qub(x) = − , Q f (x) = −c −1 Gix (x, s) f (s) ds,
c hi xi−1
Z xi
Uub(x) = ϕi (x) ubi + ϕi−1 (x) ubi−1 , U f (x) = Gi (x, s) f (s) ds,
xi−1

where hi := xi − xi−1 and Gi is the Green’s function of the second local problem,
namely, (
i c hi ϕi (s) ϕi−1 (x) if xi−1 ≤ s ≤ x,
G (x, s) :=
c hi ϕi (x) ϕi−1 (s) if x ≤ s ≤ xi .
where (
(s − xi−1 )/hi if xi−1 ≤ s ≤ xi ,
ϕi (s) :=
(xi+1 − s)/hi+1 if xi ≤ s ≤ xi+1 .
Static condensation, hybridization, and the devising of the HDG methods 7

ui }Ni=0 is
As a consequence, the global problem for the values {b
Z xi+1
ubi − ubi−1 ubi+1 − ubi
− = ϕi (s) f (s) ds for i = 1, . . . , N − 1,
c hi c hi+1 xi−1
ubj = uD (x j ) for j = 0, N.

In other words, the values of the exact solution at the nodes of the triangulation,
ui }Ni=0 , can be obtained by inverting a (symmetric positive definite) tridiagonal
{b
matrix of order N + 1.

2.2 Static condensation of the continuous Galerkin method

Now, we show that a characterization of the continuous Galerkin method similar to


that one just obtained for the exact solution can be interpreted as the original static
condensation of the method [56].

2.2.1 A characterization of the approximate solution

The continuous Galerkin method provides an approximation to u, uh , in the space

Wh = {w ∈ C0 (Ω ) : w|K ∈ W (K) ∀K ∈ Ωh }.

It determines it by requiring that it be the only solution in Wh (uD ) of the equation

(a ∇uh , ∇w)Ω = ( f , w)Ω ∀w ∈ Wh (0).

where Wh (g) = {w ∈ Wh : w = Ih (g) on ∂ Ω }, and Ih is a suitably defined interpola-


tion operator.
Now, to obtain our characterization of the approximate solution, we need to split
the spaces in a suitable manner. Thus, for each element K ∈ Ωh , we define the space
associated to the interior degrees of freedom,

W0 (K) := {w ∈ W (K) : w|∂ K = 0},

and the space associated to the degrees of freedom on the boundary,

W∂ (K) := {w ∈ W (K) : w|∂ K = 0 =⇒ w|K = 0}.

Clearly, W (K) = W0 (K) +W∂ (K) for all K ∈ Ωh , and so Wh = W0,h +WFh where

W0,h := {w ∈ Wh : w|K ∈ W0 (K) ∀K ∈ Ωh },


WFh := {w ∈ Wh : w|K ∈ W∂ (K) ∀K ∈ Ωh }.
8 Bernardo Cockburn

We also need to introduce the following sets of traces on Fh :

Mh := {w|Fh : w ∈ Wh },
Mh (g) := {µ ∈ Mh : µ|∂ Ω = Ih (g)}.

Note that the trace into Fh is an isomorphism between WFh and Mh .


Suppose that, for each element K ∈ Ωh , we define U ∈ W (K) as the solution of
the local problem

(a ∇U, ∇w)K = ( f , w)K ∀w ∈ W0 (K),


U = ubh on ∂ K,

where we want to chose the function ubh ∈ Mh in such a way that U = uh on each
element K ∈ Ωh . This happens if and only ubh is such that

(a ∇U, ∇w)Ω = ( f , w)Ω ∀w ∈ WFh ,


ubh = Ih (uD ) on ∂ Ω .

If we separate the influence of ubh from that of f in the definition of the local
problems, and rework the formulation of the global problem, we get the following
result.

Theorem 2 (Characterization of the continuous Galerkin method). The approx-


imation given by the continuous Galerkin method can be written as

uh = U = Uubh + U f ,

where, on the element K ∈ Ωh , Uubh and U f are the elements of W (K) that solve the
local problems

(a ∇Uubh , ∇w)K = 0 ∀w ∈ W0 (K) (a ∇U f , ∇w)K = ( f , w)K ∀w ∈ W0 (K),


Uubh = ubh on ∂ K Uf = 0 on ∂ K,

and ubh is the element of Mh (uD ) that solves the global problem

(a ∇Uubh , ∇Uµ )Ω = ( f , Uµ )Ω ∀ µ ∈ Mh (0).

Note that, although the static condensation [56] is carried out directly on the
stiffness matrix of the method, this result shows how to use (local and global) weak
formulations to achieve exactly the same thing.

Proof. By the linearity of the problem, we only have to justify the characterization
of the function ubh . Let us start from the fact that ubh is the element of Mh (uD ) which
solves the global problem

(a ∇Uubh , ∇w)Ω + (a ∇U f , ∇w)Ω = ( f , w)Ω ∀ w ∈ WFh .


Static condensation, hybridization, and the devising of the HDG methods 9

Now, note that, for any w ∈ Wh , we can define the function w0 by the equation

w = Uµ + w0 ,

where µ := w|Fh ; this readily implies that w0 ∈ W0,h . If we now insert this expression
in the equation and take into consideration the definition of the solution of the local
problems, that is, that

(a ∇Uubh , ∇w0 )Ω = 0,
(a ∇U f , ∇Uµ )Ω = 0,
(a ∇U f , ∇w0 )Ω = ( f , w0 )Ω ,

we finally get the wanted formulation. This completes the proof. 

2.2.2 The numerical trace of the flux

A quick comparison of the above result with the one for the exact solution, suggests
that the global problem for the continuos Galerkin method is a transmission con-
dition on a discrete version of the normal component of the flux. This little known
fact will allow us to identify the numerical trace of the approximate flux for the
continuous Galerkin method.
To do this, we first write the global problem in its original form, that is,

(a ∇uh , ∇w)Ω = ( f , w)Ω ∀ w ∈ WFh ,

and perform a simple integration by parts to get

−(∇ · (a ∇uh ), w)Ωh + h(a ∇uh ) · n, wi∂ Ωh = ( f , w)Ω ∀ w ∈ WFh ,

Let us now define, for each element K ∈ Ωh , the function Rh ∈ W∂ (K) satisfying the
equation
hRh , wi∂ K = (∇ · (a ∇uh ) + f , w)K ∀w ∈ W∂ (K).
Thus, the function Rh is a projection of the residual ∇ · (a ∇uh ) + f . With this defi-
nition, we get that

h(−a ∇uh ) · n + Rh , wi∂ Ωh = 0 ∀ w ∈ WFh ,

which can be interpreted as a transmission condition forcing the normal component


of numerical trace of the flux

bh · n := (−a ∇uh ) · n + Rh
q on ∂ Ωh ,

to be weakly continuous across interelement boundaries.


10 Bernardo Cockburn

2.2.3 Relation with static condensation

Let us now show that this characterization is nothing but an application of the well-
known technique of static condensation [56]. Static condensation was conceived as
a way to reducing the size of the stiffness matrix. Indeed, if [uh ] is the vector of de-
grees of freedom of the approximation uh , and the matrix equation of the continuous
Galerkin method is
K [uh ] = [ f ],
the static condensation consists in partitioning the vector of degrees of freedom [uh ]
into two smaller vectors, namely, the degrees of freedom interior to the elements,
[U], and the degrees of freedom associated to the boundaries of the elements, [b uh ],
and then eliminating [U] from the equations. Indeed, taking into account this parti-
tion, the above equation reads
    
K00 K0∂ [U] f
= 0 .
K∂ 0 K∂ ∂ [b uh ] f∂

By our choice of the degrees of freedom, the matrix K00 is easy to invert since it is
block diagonal, each block being associated to a local problem. We thus get
−1 −1
[U] = −K00 K0∂ [b
uh ] + K00 [ f0 ].

We can now eliminate [U] from the original matrix equation to obtain
−1 −1
(−K∂ 0 K00 uh ] = −K∂ 0 K00
K0∂ + K∂ ∂ )[b [ f0 ] + [ f∂ ].

The matrix in the left-hand side, nowadays called the Schur complement of the ma-
trix K00 , is clearly smaller than the original matrix K and is also easier to numerically
invert. We have thus shown that our characterization of the approximate solution of
the continuous Galerkin method is nothing but another way of carrying out the good,
old static condensation. The former expresses in terms of weak formulations what
the latter does directly on the matrix equations itself.

2.2.4 An example

Let us now illustrate this procedure in our simple one-dimensional example. We


take
W (K) := Pk (K),
where Pk (K) denotes the space of polynomials of degree at most k defined on the
set K. We begin by solving the local problems. If we use the notation ubi = ubh (xi )
for i = 0, . . . , N, a few manipulations (and the proper choice of the basis functions)
allow us to see that the solutions of the local problems are
Static condensation, hybridization, and the devising of the HDG methods 11
Z xi
Uub(x) = ϕi (x) ubi + ϕi−1 (x) ubi−1 U f (x) = Gih (x, s) f (s) ds,
xi−1

where hi := xi − xi−1 and Gih is the discrete Green’s function of the second local
problem, namely,

hi k−1 1
Gih (x, s) := i
∑ 2` + 1 (P`+1 i
− P`−1 i
)(x) (P`+1 i
− P`−1 )(s)
4a `=1

where Pni (x) := Pn (T i (x)), T i (ζ ) := (ζ − (xi + xi−1 )/2)/(hi /2) and Pn is the Leg-
endre polynomial of degree n. As a consequence, the global problem for the values
ui }Ni=0 is
{b
Z xi+1
ubi − ubi−1 ubi+1 − ubi
a −a = ϕi (s) f (s) ds for i = 1, . . . , N − 1,
hi hi+1 xi−1
ubj = uD (x j ) for j = 0, N.

Note that the size of the matrix equation of the global problem is independent of
the value of the polynomial degree k, a reflection of the effectiveness of the static
condensation technique. Note also that the values of the approximate solution at the
ui }Ni=0 , are actually exact, as expected.
nodes of the triangulation, {b

2.3 Static condensation of mixed methods by hybridization

Next, we show how to extend what was done for the continuous Galerkin method
to mixed methods. A particular important point we want to emphasize here is that
hybridization of a mixed method is what allows it to be statically condensed, as first
realized in [53].

2.3.1 A characterization of the approximate solution

A mixed method seeks approximations to the flux q := −a∇u, qh , and the scalar u,
uh , in the finite dimensional spaces

Vh = {v ∈ H(div, Ω ) : v|K ∈ V(K) ∀K ∈ Ωh }.


Wh = {w ∈ L2 (Ω ) : w|K ∈ W (K) ∀K ∈ Ωh },

respectively. It determines the function (qh , uh ) as the only element of Vh ×Wh sat-
isfying the equations

(c qh , v)Ω − (uh , ∇ · v)Ω = −huD , v · ni∂ Ω ∀v ∈ Vh ,


(∇ · qh , w)Ω = ( f , w)Ω ∀w ∈ Wh .
12 Bernardo Cockburn

For mixed methods, the choice of the finite dimensional space Vh ×Wh is not simple,
but here we assume that it has been properly chosen as to define a unique approxi-
mate solution.
Now, suppose that, for each element K ∈ Ωh , we define (Q, U) ∈ V(K) ×W (K)
as the solution of the local problem

(c Q, v)K − (U, ∇ · v)K = hb


uh , v · ni∂ K ∀v ∈ V(K),
(∇ · Q, w)K = ( f , w)K ∀w ∈ W (K).

This problem is well defined since it is nothing but the application of the mixed
method, which we assume to be well defined, to the single element K ∈ Ωh . As
before, we want to choose the function ubh in some finite dimensional space Mh in
such a way that (Q, U) = (qh , uh ) on each element K ∈ Ωh . For this to hold, we only
need to guarantee that

Q ∈ Vh ,
hb
uh , v · ni∂ Ω = huD , v · ni∂ Ω ∀ v ∈ Vh .

The first property is a transmission condition since it holds if and only if the normal
component of Q is continuous across interelement boundaries. The second condition
is nothing but a weak form of the Dirichlet boundary condition.
As for the case of the continuous Galerkin method, the choice of the space Mh has
to be made in such a way that the above two conditions do determine the numerical
trace ubh . Typically, we take

Mh := {µ ∈ L2 (Fh ) : ∃ v ∈ Vh : µ = [[v]] on Fh }.

Thus, if we set Mh (g) := {µ ∈ Mh : hµ, ηi∂ Ω = hg, ηi∂ Ω ∀η ∈ Mh }, the global


problem can be expressed as follows:

hQ · n, µi∂ Ωh = 0 ∀ µ ∈ Mh (0),
ubh ∈ Mh (uD ).

Indeed, note that, for any µ ∈ Mh (0),

hQ · n, µi∂ Ωh = hQ, µi∂ Ωh \∂ Ω = h [[Q]], µiFi ,


h

and if this quantity is zero, we certainly have that Q ∈ Vh , as wanted. So, let us
assume then that the above global problem for ub ∈ Mh is well defined.
So, we have obtained the following result.

Theorem 3 (Characterization of the mixed method). The solution of the mixed


method can be written as

(qh , uh ) = (Q, U) = (Qubh , Uubh ) + (Q f , U f ),


Static condensation, hybridization, and the devising of the HDG methods 13

where, on each element K ∈ Ωh , for any µ ∈ L2 (∂ K) and f ∈ L2 (K), the functions


(Qµ , Uµ ) and (Q f , U f ) are the elements of V(K) × W (K) which solve the local
problems

(c Qµ , v)K − (Uµ , ∇ · v)K = −hµ, v · ni∂ K , (c Q f , v)K − (U f , ∇ · v)K = 0,


(∇ · Qµ , w)K = 0, (∇ · Q f , w)K = ( f , w)K ,

for all (v, w) ∈ V(K) × W (K), and the function ubh is the element of Mh (uD ) which
solves the global problem

(c Qubh , Qµ )Ωh = ( f , Uµ )Ωh ∀ µ ∈ Mh (0),

Proof. We only have to prove that ubh ∈ Mh (uD ) satisfies the equation

−hQubh · n, µi∂ Ωh = hQ f · n, µi∂ Ωh ∀ µ ∈ Mh (0).

But, by the definition of the local problems, we have

−hQubh · n, µi∂ Ωh = (c Qµ , Qubh )Ωh ,


hQ f · n, µi∂ Ωh = −(c Qµ , Q f )Ωh + (Uµ , ∇ · Q f )Ωh
= −(U f , ∇ · Qubh )Ωh + (Uµ , ∇ · Q f )Ωh
= (Uµ , ∇ · Q f )Ωh
= ( f , Uµ )Ωh ,

and the identity follows. This completes the proof. 

2.3.2 Relation with static condensation and hybridization

Let us now show that what we have done is nothing but the static condensation of
the hybridized version of the mixed method as done by Fraejis de Veubeke in [53].
Suppose that the matrix equation of the mixed method reads

A B [qh ]
    
[uD ]
= .
Bt 0 [uh ] [f]

It is not easy to eliminate [qh ] from this equation since the matrix A is not block
diagonal because, since qh ∈ Vh , its normal component is continuous across in-
ter element boundaries. To overcome this unwanted feature, Fraejis de Veubeque
relaxed the continuity condition on qh and worked with a totally discontinuous ap-
proximation Q instead. Because of this, he had to introduce the hybrid unknown ubh ,
an approximation to the trace of u on each element; this is why this procedure re-
ceives the name of hybridization of the mixed method. Finally, in order to guarantee
that Q be identical to the original function qh , he then forced it to have a continu-
ous normal component at the interelement boundaries. This operation resulted the
14 Bernardo Cockburn

following matrix equation:


    
A BC [Q] −C∂ [uD ]
Bt 0 0   [U]  =  [ f ]  .
Ct 0 0 [buh ] 0

Here, [b
uh ] denotes the digressive freedom of the function ubh restricted to the interior
faces. On the boundary faces, the relation of ubh to uD is already captured by the
right-hand side of the first equation. Note that, since the first two equations define
the local problems, we can easily solve them to obtain
   −1  
[Q] A B −C[b
uh ] −C∂ [uD ]
= t .
[U] B 0 [f]

The third equation, C [Q] = 0 enforces the continuity of the normal component of
Q across inter element boundaries; it is this equation that determines the hybrid
unknown in the interior faces, [b
uh ]. A few computations show that the resulting
matrix equation is of the form

H[ ub] = H∂ [ uD ]+J [ f ], H := Ct E C, E := A−1 −A−1 B (Bt A−1 B)−1 Bt A−1 ,

and we see that, as expected, the matrix H is symmetric. Moreover, H is positive


definite and E is block-diagonal.

2.3.3 An example

Next, let us illustrate this procedure in our simple one-dimensional example. We


take
V(K) ×W (K) := Pk+1 (K) × Pk (K).
We begin by solving the local problems. A little computation gives that the solutions
of the local problems are
Z xi
ubi − ubi−1
Qub(x) = − , Q f (x) = Hhi (x, s) f (s) ds,
c hi xi−1
Z xi
Uub(x) = ϕi (x) ubi + ϕi−1 (x) ubi−1 , U f (x) = Gih (x, s) f (s) ds,
xi−1

where hi := xi − xi−1 and

1 k
Hhi (x, s) := ϕi (x) ϕi (s) − ϕi−1 (x) ϕi−1 (s) + i
∑ (P`+1 i
− P`−1 )(x) P`i (s),
2 `=1
c hi k−1 1
Gih (x, s) := i
∑ 2` + 1 (P`+1 i
− P`−1 i
)(x) (P`+1 i
− P`−1 )(s).
4 `=1
Static condensation, hybridization, and the devising of the HDG methods 15

Let us recall that Pni (x) := Pn (T i (x)), T i (ζ ) := (ζ − (xi + xi−1 )/2)/(hi /2) and Pn
is the Legendre polynomial of degree n. Note that the function Gih approximates
the Green function Gi whereas −c Hhi approximates its partial derivative Gix . As a
consequence, the global problem for the values {b ui }Ni=0 is
Z xi+1
ubi − ubi−1 ubi+1 − ubi
− = ϕi (s) f (s) ds for i = 1, . . . , N − 1,
c hi c hi+1 xi−1
ubj = uD (x j ) for j = 0, N.

We thus see that the values of the approximate solution at the nodes of the triangu-
ui }Ni=0 , are actually exact, as expected.
lation, {b

3 HDG methods

In this section, we show how to use a discrete version of the characterization of


the exact solution obtained in the previous section to devise HDG methods for our
model problem (1). The local problems are solved by using a DG method and the
transmission conditions by a simple weak formulation. As a consequence, the re-
sulting HDG methods are DG methods whose distinctive feature is that they are
amenable to hybridization and hence to static condensation. Let us emphasize that
this does not happen by accident, but because they are constructed by using a dis-
crete version of the characterization of the exact solution worked out in the previous
section.
After defining the HDG methods, we establish a simple result about the existence
and uniqueness of their approximate solution and display some examples. We end by
showing several different ways of presenting them which will be useful for relating
them to other numerical methods.
We follow closely the work done in 2009 [24] for the original HDG methods,
as well as the work done in the 2014 review paper [38] for HDG methods for the
Stokes system of incompressible fluid flow.

3.1 Definition

We take the approximate solution of the HDG methods to be the function

(qh , uh ) = (Q, U),

where, on the element K ∈ Ωh , (Q, U) ∈ V(K) × W (K) is the solution of the local
problem
16 Bernardo Cockburn

(c Q, v)K − (U, ∇ · v)K + hb


uh , v · ni∂ K = 0 ∀v ∈ V(K),
−(Q, ∇w)K + hQ b · n, wi∂ K = ( f , w)K ∀w ∈ W (K),

where the numerical trace Qb has to be suitably chosen. Ideally, the numerical trace
of the flux Q
b should be chosen so that it

(i) is consistent,
(ii) only depends (linearly) on Q|K , U|K and ubh |∂ K ,
(iii) renders the local problem solvable.
Our favorite choice is
b · n := Q · n + τ(U − ubh ) on ∂ K,
Q

where the function τ is linear. We are also going to require that τ be symmetric, that
is, that, for all K ∈ Ωh ,

hτ(w), ωi∂ K = hw, τ(ω)i∂ K ∀ w, ω ∈ W (K) + Mh (∂ K).

Although there are many other choices, we are going to use this one from now on;
not only it is very natural but it actually covers all the known HDG methods.
To complete the definition of the HDG methods, we take the function ubh in the
space
Mh := {µ ∈ L2 (Fh ) : µF ∈ M(F) ∀F ∈ Fh },
where M(F) is a suitably chosen finite dimensional space, and require that it be de-
termined as the solution of the following weakly imposed transmission and bound-
ary conditions:

hµ, [[Q]]i
b
F i = hµ, Q · ni∂ Ωh \∂ Ω = 0,
b
h

hµ, ubh i∂ Ω = hµ, uD i∂ Ω ,

for all µ ∈ Mh . This completes the definition of the HDG methods.

The HDG methods are obtained by choosing different functions τ and dif-
ferent local spaces V(K), W (K) and M(F).

3.2 Existence and uniqueness

We now provide simple conditions on the local spaces and the function τ ensuring,
not only that the local problems are solvable, but that the global problem is also
well posed. To do that, we use an energy identity we obtain next which will also
shed light on the role to the function τ.
Proposition 1 (The local energy identity). For any element K ∈ Ωh , we have
Static condensation, hybridization, and the devising of the HDG methods 17

(c Q, Q)K + h(U − ubh ), τ(U − ubh )i∂ K = ( f , U)K − hb b · ni∂ K .


uh , Q

Note that the exact solution satisfies the following energy identity:

(c q, q)K = ( f , u)K − hu, q · ni∂ K .

Typically, the terms (c q, q)K and (c Q, Q)K are interpreted as the energy stored in-
side the element K. It is thus reasonable to interpret the term h(U− ubh ), τ(U− ubh )i∂ K
as an energy associated with the jumps U − ubh at the boundary of the element ∂ K.
Since all energies are nonnegative, we assume that the function τ is such that

hτ(w − µ), w − µi∂ K ≥ 0 ∀(w, µ) ∈ W (K) × M(∂ K), (2a)

where

M(∂ K) := {µ ∈ L2 (∂ K) : µ|F ∈ M(F), for any face F ∈ Fh lying on ∂ K}. (2b)

We now see that the role of τ is to transform the discrepancy between U and ubh
on ∂ K into an energy. Since an increase of energy is typically associated with an
enhancement of the stability properties of the numerical method, τ is called the
stabilization function.
Let us now prove Proposition 1.

Proof. If we take (v, w) := (Q, U) in the equations of the local problems, we get

(c Q, Q)K − (U, ∇ · Q)K + hb


uh , Q · ni∂ K = 0,
−(Q, ∇U)K + hQ b · n, Ui∂ K = ( f , U)K ,

and adding the two equations, we obtain

(c Q, Q)K + h(Q
b − Q) · n, U − ubh i∂ K = ( f , U)K − hQ
b · n, ubh i∂ K .

The energy identity now follows by simply inserting the definition of the numerical
trace Q.
b This completes the proof. 

We are now ready to present our main result. It is a variation of a similar result
in [24].

Theorem 4. Assume that the stabilization function τ satisfies the nonnegativity con-
dition (2). Assume also that, for each element K ∈ Ωh , we have that if (w, µ) ∈
W (K) × M(∂ K) is such that

(i) hτ(w − µ), w − µi∂ K = 0,


(ii) (∇w, v)K + hµ − w, v · ni∂ K = 0 ∀ v ∈ V(K),

then w is a constant on K and w = µ on ∂ K. Then the approximate solution


(qh , uh , ubh ) ∈ Vh ×Wh × Mh of the HDG method is well defined.
18 Bernardo Cockburn

Note that condition (ii) establishes a relation between the local spaces V(K),
W (K) and M(∂ K) and the stabilization function τ guaranteeing that the local prob-
lems as well as the global problem have a unique solution. Note also that if condition
(i) were not necessary to obtain that w is a constant on K and w = µ on ∂ K, we can
simply take τ ≡ 0. However, for most cases, without condition (i), the method might
simply fail to be well defined. The role of τ, is thus to prevent this failure.
Let us now prove Theorem 4.

Proof. Since the HDG method defines a finite dimensional square system for the
unknowns (Q, U, ubh ) ∈ Vh × Wh × Mh , we only have to show that, when we set the
data f and uD to zero, the only solution is the trivial one.
Thus, setting µ := ubh in the transmission condition, and recalling that, by the
boundary condition, ubh = 0 on ∂ Ω , we get

0 = − hb b · ni∂ Ω = (c Q, Q)Ω + h(U − ubh ), τ(U − ubh )i∂ Ω ,


uh , Q h h h

by the energy identity of the previous proposition. By assumption (i), we get that
Q = 0 on Ω and that h(U − ubh ), τ(U − ubh )i∂ K = 0 for any K ∈ Ωh . Moreover, the
first equation defining the local problem now reads

(∇U, v)K + hb
uh − U, v · ni∂ K = 0 ∀ v ∈ V(K).

By assumption (ii) with (w, µ) := (U, ubh ), we have that, on each element K ∈ Ωh ,
U is a constant on K and that U = ubh on ∂ K. As a consequence, U is a constant on
Ω and U = ubh on Fh . Since ubh = 0 on ∂ Ω we finally get that U = 0 on Ω and that
ubh = 0 on Fh . This completes the proof. 

Let us now present an almost direct consequence of the previous result in a case
in which the stabilization function τ is very strong.
Corollary 1 ([24]). Assume that the stabilization function τ satisfies the nonnega-
tivity condition (2). Assume also that, for every element K ∈ Ωh ,
(a) (w, µ) ∈ W (K) × M(∂ K) : hτ(w − µ), w − µi∂ K = 0 =⇒ w = µ on ∂ K,
(b) ∇W (K) ⊂ V(K).
Then the approximate solution (qh , uh , ubh ) ∈ Vh × Wh × Mh of the HDG method is
well defined.
A remarkable feature of this result is that the method is well defined completely in-
dependently of the choice of the space Mh . This is a direct consequence of condition
(a), which is clearly stronger than condition (i) of Theorem 4 on the stabilization
function τ. Thanks to condition (a) , we can replace condition (ii) of Theorem 4 by
the simpler condition (b), as we see next.

Proof. We only have to show that the assumptions the previous result are satisfied.
Since τ is a linear mapping, assumption (a) implies condition (i) of Theorem 4.
Now, by assumption (a), if hτ(w − µ), w − µi∂ K = 0, we have that w = µ on ∂ K and
we get that condition (ii) of Theorem 4 reads
Static condensation, hybridization, and the devising of the HDG methods 19

(∇w, v)K = 0 ∀v ∈ V(K).

By assumption (b), we can take v := ∇w and conclude that w is a constant on K.


This implies that the second assumption of Theorem 4 holds. This completes the
proof. 

3.3 Characterizations of the HDG methods

Here, we provide two characterizations of the approximate solution provided by the


HDG methods just introduced. We are going to use the set

Mh (g) := {µ ∈ Mh : hµ, ηi∂ Ω := hg, ηi∂ Ω ∀η ∈ Mh }.

3.3.1 Formulation in terms of (qh , uh , ubh )

Static condensation formulation

The following result reflects the way in which the HDG methods were devised and
renders evident the way in which their implementation by static condensation can
be achieved.

Theorem 5 (First characterization of HDG methods). The approximate solution


of the HDG method is given by

(qh , uh ) = (Q, U) = (Qubh , Uubh ) + (Q f , U f ),

where, on the element K ∈ Ωh , for any µ ∈ L2 (∂ K), the function (Qµ , Uµ ) ∈ V(K)×
W (K) is the solution of the local problem

(c Qµ , v)K − (Uµ , ∇ · v)K + hµ, v · ni∂ K = 0 ∀v ∈ V(K),


−(Qµ , ∇w)K + hQ b µ · n, wi∂ K = 0 ∀w ∈ W (K),

Qb µ · n := Qµ · n + τ(Uµ − µ) on ∂ K,

and, for any f ∈ L2 (K), the function (Q f , U f ) ∈ V(K) ×W (K) is the solution of the
local problem

(c Q f , v)K − (U f , ∇ · v)K = 0 ∀v ∈ V(K),


−(Q f , ∇w)K + hQ b f · n, wi∂ K = ( f , w)K ∀w ∈ W (K),
b f · n := Q f · n + τ(U f )
Q on ∂ K.

The function ubh is the element of Mh (uD ) such that


20 Bernardo Cockburn

uh , µ) = `h (µ) ∀ µ ∈ Mh (0),
ah (b

where ah (µ, λ ) := −hµ, Q


b λ · ni∂ Ω , and `h (µ) := hµ, Q
h
b f · ni∂ Ω . Moreover,
h

ah (µ, λ ) = (c Qµ , Qλ )Ωh + hUµ − µ, τ(Uλ − λ )i∂ Ωh , `h (µ)= ( f , Uµ ),

and ah (·, ·) is symmetric and positive definite on Mh (0) × Mh (0). Thus, ubh minimizes
the total energy functional Jh (µ) := 12 ah (µ, µ) − `h (µ) over Mh (uD ).

Proof. We only need to prove the last two identities and the property of positive
definiteness of the bilinear form ah (·, ·).
Let us prove the first identity. If we take v := Qλ in the first equation defining
the first local problem, replace µ by λ in the second equation defining the first local
problem and set w := Uµ , we get

(c Qµ , Qλ )K − (Uµ , ∇ · Qλ )K + hµ, Qλ · ni∂ K = 0,


−(Qλ , ∇Uµ )K + hQ b λ · n, Uµ i∂ K = 0.

Adding the two equations, we obtain

(c Qµ , Qλ )K + h(Q
b λ − Qλ ) · n, Uµ − µi∂ K = −hQ
b λ · n, µi∂ K .

The first identity follows by inserting the definition of the numerical trace Q b λ and
adding over the elements K ∈ Ωh .
Let us prove the second identity. If we take v := Q f in the first equation defining
the first local problem and w := Uµ in the second equation defining the second local
problem, we get

(c Qµ , Q f )K − (Uµ , ∇ · Q f )K + hµ, Q f · ni∂ K = 0,


−(Q f , ∇Uµ )K + hQ b f · n, Uµ i∂ K = ( f , Uµ )k

and if we add the two equations and insert the definition of Q


b f , we obtain

(c Qµ , Q f )K + hτ(U f ), Uµ − µi∂ K = ( f , Uµ )K − hQ
b f · n, µi∂ K .

If we now take v := Qµ in the first equation defining the second local problem and
w := U f in the second equation defining the first local problem with ubh := µ, we get

(c Q f , Qµ )K − (U f , ∇ · Qµ )K = 0,
−(Qµ , ∇U f )K + hQ b µ · n, U f i∂ K = 0,

and if we proceed as before, we get

(c Q f , Qµ )K + hτ(Uµ − µ), U f i∂ K = 0.

This implies that


Static condensation, hybridization, and the devising of the HDG methods 21

−hτ(Uµ − µ), U f i∂ K + hτ(U f ), Uµ − µi∂ K = ( f , Uµ )K − hQ


b f · n, µi∂ K ,

and the result follows by the fact that τ is symmetric.


The fact that ah (·, ·) is symmetric follows from the previous identities and the fact
that τ is also symmetric. Finally the fact that it is positive definite on Mh (0) × Mh (0)
follows exactly as in the proof of Theorem 4. This completes the proof. 

Two compact formulations

Let us now show how to rewrite the HDG methods in a more compact manner. It
does not suggest a way to statically condense the methods but it is our favorite way
of presenting them concisely. It emphasizes the role of the numerical traces of the
methods and is suitable for carrying out their analysis. It is the following.
The approximate solution given by the HDG method is the function (qh , uh , ubh ) ∈
Vh ×Wh × Mh (uD ) satisfying the equations

(c qh , v)Ωh − (uh , ∇ · v)Ωh + hb


uh , v · ni∂ Ωh = 0 ∀v ∈ Vh , (3a)
−(qh , ∇w)Ωh + hb
qh · n, wi∂ Ωh = ( f , w)Ωh ∀w ∈ Wh , (3b)
bh · n := qh · n + τ(uh − ubh )
q on ∂ Ωh , (3c)
hµ, q
bh · ni∂ Ωh = 0 ∀µ ∈ Mh (0). (3d)

Indeed, note that the first, second and third equations correspond to the definition of
the local problems and that the weakly imposed boundary conditions are enforced
by requesting that ubh be an element of Mh (uD ).
We can also eliminate the numerical trace q bh to obtain yet another rewriting of the
methods. Once again, it hides the numerical trace of the flux, but emphasizes what
we could call the stabilized mixed method structure of the methods. The formulation
is the following. The approximate solution given by the HDG method is the function
(qh , uh , ubh ) ∈ Vh ×Wh × Mh (uD ) satisfying the equations

Ah (qh , v) + Bh (uh , ubh ; v) = 0 ∀v ∈ Vh ,


−Bh (w, µ; qh ) + Sh (uh , ubh ; w, µ) = ( f , w)Ωh ∀(w, µ) ∈ Wh × Mh (0),

where

Ah (p, v) := (c p, v)Ωh ∀p, v ∈ Vh ,


Bh (w, µ; v) := −(w, ∇ · v)Ωh + hµ, v · ni∂ Ωh ∀(v, w, µ) ∈ Vh ×Wh × Mh ,
Sh (ω, λ ; w, µ) := hτ(ω − λ ), w − µi∂ Ωh ∀(ω, λ ), (w, µ) ∈ Wh × Mh .

Indeed, the first equation follows from the definition of the bilinear forms Ah (·, ·)
and Bh (·, ·). It remains to prove that

Bh (w, µ; qh ) + Sh (uh , ubh ; w, µ) = −(qh , ∇w)Ωh + hb


qh · n, wi∂ Ωh − hb
qh · n, µi∂ Ωh .
22 Bernardo Cockburn

But, we have, by the definition of the bilinear forms Bh (·, ·) and Sh (·, ·), that

Θ := − Bh (w, µ; qh ) + Sh (uh , ubh ; w, µ)


= (w, ∇ · q)Ωh − hµ, q · ni∂ Ωh + hτ(uh − ubh ), w − µi∂ Ωh
= − (qh , ∇w)Ωh + hq · n + τ(uh − ubh ), w − µi∂ Ωh ,

by integration by parts. The identity we want follows now by using the definition of
the numerical trace of the flux.
To end, we note that, thanks to the structure of the methods, it is easy to see that
the solution (qh , uh , ubh ) ∈ Vh ×Wh × Mh (uD ) minimizes the functional

1
Jh (v, w, µ) := {Ah (v, v) + Sh (w, µ; w, µ)} − ( f , w)Ωh (4a)
2
over all functions (v, w, µ) in Vh ×Wh × Mh (uD ) such that

Ah (v, p) + Bh (w, µ; p) = 0 ∀p ∈ Vh . (4b)

Note that the last equation can be interpreted as the elimination of qh from the
equations. The minimization problem would then be one on the affine space Wh ×
Mh (uD ) and would correspond to a problem formulated solely in terms of uh and ubh .
Next, we explore the static condensation of such reformulation.

3.3.2 Formulation in terms of (uh , ubh )

So, here we eliminate the approximate flux qh from the equations defining the HDG
method in order to formulate it solely in terms of (uh , ubh ). To achieve this, we simply
rewrite qh as a linear mapping applied to (uh , ubh ). This mapping is defined by using
the first equation defining the HDG methods. Thus, for any (w, µ) ∈ Wh × Mh , we
define Qw,µ ∈ Vh as the solution of

(c Qw,µ , v)Ωh − (w, ∇ · v)Ωh + hµ, v · ni∂ Ωh = 0 ∀ v ∈ Vh ,

In this way, we are going to have that qh = Quh ,buh . Note that the above equation is
nothing but a rewriting of the equation (4b).

Static condensation formulation

Using this approach, we obtain the following characterization of the HDG methods.
It is useful for their implementation and involves less unknows than the previous
characterization since the unknown for the approximate flux has been eliminated.
(Of course, the price to pay for this is that we now we have to work with the mapping
(w, µ) 7→ Qw,µ .) This characterization better shows the role of τ as a stabilization
Static condensation, hybridization, and the devising of the HDG methods 23

function but it hides its relation with the numerical trace of the flux and does not
clearly indicate the associated transmission condition.

Theorem 6 (Second characterization of HDG methods). The approximate solu-


tion of the HDG method is given by

(qh , uh ) = (Q, U) = (QUub uh , Uubh ) + (QU f ,0 , U f ),


,b
h

where, on the element K ∈ Ωh , for any µ ∈ L2 (∂ K) and any f ∈ L2 (K), the functions
Uµ , U f ∈ W (K) are the solutions of the local problems

(c QUµ ,µ , Qw,0 )K + hτ(Uµ − µ), wi∂ K = 0 ∀w ∈ W (K),


(c QU f ,0 , Qw,0 )K + hτ(U f ), wi∂ K = ( f , w)K ∀w ∈ W (K),

respectively. The function ubh is the element of Mh (uD ) such that

uh , µ) = `h (µ) ∀ µ ∈ Mh (0),
ah (b

where ah (µ, λ ) := −hµ, Q


b U ,λ · ni∂ Ω and `h (µ) := hµ, Q
λ h
b U ,0 · ni∂ Ω . Moreover,
f h

ah (µ, λ ) = (c QUµ ,µ , QUλ ,λ )∂ Ωh + hUµ − µ, τ(Uλ − λ )i∂ Ωh , `h (µ)= ( f , Uµ ),

and ah (·, ·) is symmetric and positive definite on Mh (0) × Mh (0). Thus, ubh minimizes
the total energy functional Jh (µ) := 12 ah (µ, µ) − `h (µ) over Mh (uD ).

Proof. This results follows easily from the first characterization of the HDG meth-
ods given in Theorem 5. We only have to show that the solutions of the local prob-
lems coincide, that is, that (QUµ ,µ , Uµ ) ∈ V(K) ×W (K) is the solution of

(c QUµ ,µ , v)K − (Uµ , ∇ · v)K + hµ, v · ni∂ K = 0 ∀v ∈ V(K),


−(QUµ ,µ , ∇w)K + hQ
b U ,µ · n, wi∂ K = 0
µ ∀w ∈ W (K),
b µ · n := Qµ · n + τ(Uµ − µ)
Q on ∂ K,

and that (QU f ,0 , U f ) ∈ V(K) ×W (K) is the solution of

(c QU f ,0 , v)K − (U f , ∇ · v)K = 0 ∀v ∈ V(K),


−(QU f ,0 , ∇w)K + hQ
b U ,0 · n, wi∂ K = ( f , w)K
f
∀w ∈ W (K),
b f · n := Q f · n + τ(U f )
Q on ∂ K.

Since the first equation of these problems is nothing but the definition of the operator
Qw,µ , we only have to show that

(∇ · QUµ ,µ , w)K + hτ(Uµ − µ), wi∂ K = 0 ∀w ∈ W (K),


(∇ · QU f ,0 , w)K + hτ(U f ), wi∂ K = ( f , w)K ∀w ∈ W (K).
24 Bernardo Cockburn

But, by the definition of Qw,0 , we have

(∇ · QUµ ,µ , w)K = (c Qw,0 , QUµ ,µ )K ,


(∇ · QU f ,0 , w)K = (c Qw,0 , QU f ,0 )K ,

and the result follows. This completes the proof. 

A compact formulation

As we did for the first characterization of the HDG methods, we can rewrite the
above result in a compact manner as follows. The approximate flux provided by the
HDG method is qh = Quh ,buh and (uh , ubh ) ∈ Wh × Mh (uD ) is the solution of

(c Quh ,buh , Qw,µ )Ωh + hτ(uh − ubh ), w − µi∂ Ωh = ( f , w)Ωh ∀(w, µ) ∈ Wh × Mh (0).

We immediately see that (uh , ubh ) is the only minimum over Wh × Mh (0) of the total
energy functional
1
Jh (w, µ) := {(c Qw,µ , Qw,µ )Ωh + hτ(w − µ), w − µi∂ Ωh } − ( f , w)Ωh .
2
This minimization problem is identical to the minimization (with restrictions) prob-
lem (4).

4 HDG methods using only the tensor a := c −1

4.1 Motivation

Note that the the first three equations of the weak formulation of the DG methods
we have been considering can also be expressed as

−(gh , v)Ωh − (uh , ∇ · v)Ωh + hb


uh , v · ni∂ Ωh = 0,
(c qh , v)Ωh = −(gh , v)Ωh ,
−(qh , ∇w)Ωh + hb
qh · n, wi∂ Ωh = ( f , w)Ωh ,

where the approximate gradient gh is taken in Vh If one prefers to work with the
tensor a := c −1 , we can simply use the equations

−(gh , v)Ωh − (uh , ∇ · v)Ωh + hb


uh , v · ni∂ Ωh = 0,
(qh , v)Ωh = −(a gh , v)Ωh ,
−(qh , ∇w)Ωh + hb
qh · n, wi∂ Ωh = ( f , w)Ωh ,
Static condensation, hybridization, and the devising of the HDG methods 25

for all (v, w) ∈ Vh ×Wh , where the numerical traces ubh and qbh · n are approximations
to u|∂ Ωh and q · n|∂ Ωh , respectively. The difference between these two DG methods
is certainly not abysmal since it consists in picking one of the two ways of relating
the approximate gradient gh to the approximate flux qh , namely,

(c qh , v)Ωh = −(gh , v)Ωh or (qh , v)Ωh = −(a gh , v)Ωh .

As a consequence, there is a one-to-one correspondence between these two weak


formulations, provided both a and c are well defined. Moreover, both formulations
coincide whenever a and c are constant on each element K ∈ Ωh which gives rise to
super-closeness of their approximations, as noted in [42].
However, if a degenerates and is not invertible at every point, the second for-
mulation might be preferable. This is also what motivated the so-called “extended”
form of the mixed methods introduced in [1, 62, 11].
Finally, let us note that in elasticity, g corresponds to the strain, q to the stress, a to
the so-called constitutive tensor and c to the so-called compliance tensor. Thus, the
HDG methods obtained for linear and nonlinear elasticity, see the HDG methods for
elasticity considered in 2008 [86], 2009 [87] and 2014 [54] and in 2015 [60], can be
immediately reduced to our simpler case; see also the 2006 DG method proposed in
[51]. It is well known that to work with the constitutive tensor is usually preferred in
the case of nonlinear elasticity. Next, we briefly show how to define and characterize
the HDG methods associated with using the tensor a := c −1 .

4.2 Definition, existence and uniqueness

We take the approximate solution of the HDG methods to be the function

(qh , gh , uh ) = (Q, G, U),

where, on the element K ∈ Ωh , (Q, G, U) ∈ V(K) × V(K) ×W (K) is the solution of


the local problem

−(G, v)K − (U, ∇ · v)K + hb


uh , v · ni∂ K = 0 ∀v ∈ V(K),
(Q, v)K = −(a G, v)K ∀v ∈ V(K),
−(Q, ∇w)K + hQ
b · n, wi∂ K = ( f , w)K ∀w ∈ W (K),

b is suitably chosen, and ubh ∈ Mh is the solution of the


where the numerical trace Q
following weakly imposed transmission and boundary conditions:

hµ, Q
b · ni∂ Ω \∂ Ω = 0,
h

hµ, ubh i∂ Ω = hµ, uD i∂ Ω ,

for all µ ∈ Mh . This completes the definition of the HDG methods.


26 Bernardo Cockburn

It is not difficult to see that the existence and uniqueness in Theorem 4 and its
Corollary 1 do hold unchanged.

4.3 Characterizations of the HDG methods

4.3.1 Formulation in terms of (qh , gh , uh , ubh )

Static condensation formulation

We have the following result which is analogous to Theorem 5.


Theorem 7 (First characterization of HDG methods). The approximate solution
of the HDG method is given by

(qh , gh , uh ) = (Q, G, U) = (Qubh , Gubh , Uubh ) + (Q f , G f , U f ),

where, on the element K ∈ Ωh , for any µ ∈ L2 (∂ K), the function (Qµ , Gµ , Uµ ) ∈


V(K) × V(K) ×W (K) is the solution of the local problem

−(Gµ , v)K − (Uµ , ∇ · v)K + hµ, v · ni∂ K = 0 ∀v ∈ V(K),


(Qµ , v)K = −(aGµ , v)K ∀v ∈ V(K),
−(Qµ , ∇w)K + hQ b µ · n, wi∂ K = 0 ∀w ∈ W (K),
b µ · n := Qµ · n + τ(Uµ − µ)
Q on ∂ K,

and, for any f ∈ L2 (K), the function (Q f , G f , U f ) ∈ V(K) × V(K) × W (K) is the
solution of the local problem

−(G f , v)K − (U f , ∇ · v)K = 0 ∀v ∈ V(K),


(Q f , v)K = −(aG f , v)K ∀v ∈ V(K),
−(Q f , ∇w)K + hQ f · n, wi∂ K = ( f , w)K
b ∀w ∈ W (K),
b f · n := Q f · n + τ(U f )
Q on ∂ K.

The function ubh is the element of Mh (uD ) such that

uh , µ) = `h (µ) ∀ µ ∈ Mh (0),
ah (b

where ah (µ, λ ) := −hµ, Q


b λ · ni∂ Ω , and `h (µ) := hµ, Q
h
b f · ni∂ Ω . Moreover,
h

ah (µ, λ ) = (aGµ , Gλ )∂ Ωh + hUµ − µ, τ(Uλ − λ )i∂ Ωh , `h (µ)= ( f , Uµ ),

and ah (·, ·) is symmetric and positive definite on Mh (0) × Mh (0). Thus, ubh minimizes
the functional Jh (µ) := 21 ah (µ, µ) − `h (µ) over Mh (uD ).
Static condensation, hybridization, and the devising of the HDG methods 27

Two compact formulations

Proceeding as for the first family of HDG methods, we obtain the following two
compact formulations. The first emphazised the role of the numerical traces. It reads
as follows. The approximate solution given by the HDG method is the function
(qh , gh , uh , ubh ) ∈ Vh × Vh ×Wh × Mh (uD ) satisfying the equations

−(gh , v)Ωh − (uh , ∇ · v)Ωh + hb


uh , v · ni∂ Ωh = 0 ∀v ∈ Vh ,
(qh , v)Ωh = −(a gh , v)Ωh ∀v ∈ Vh ,
−(qh , ∇w)Ωh + hb
qh · n, wi∂ Ωh = ( f , w)Ωh ∀w ∈ Wh ,
bh · n := qh · n + τ(uh − ubh )
q on ∂ Ωh ,
hµ, q
bh · ni∂ Ωh = 0 ∀µ ∈ Mh (0).

The second emphasizes the stabilized mixed structure of the method. It is the
following. The approximate solution given by the HDG method is the function
(qh , uh , ubh ) ∈ Vh ×Wh × Mh (uD ) satisfying the equations

Ah (gh , v) + Bh (qh , v) = 0 ∀v ∈ Vh ,
−Bh (v, gh ) + Bh (uh , ubh ; v) = 0 ∀v ∈ Vh ,
−Bh (w, µ; qh ) + Sh (uh , ubh ; w, µ) = ( f , w)Ωh ∀(w, µ) ∈ Wh × Mh (0),

where

Ah (p, v) := (a p, v)Ωh , ∀p, v ∈ Vh ,


Bh (p, v) := (p, v)Ωh , ∀p, v ∈ Vh ,
Bh (w, µ; v) := −(w, ∇ · v)Ωh + hµ, v · ni∂ Ωh ∀(v, w, µ) ∈ Vh ×Wh × Mh ,
Sh (ω, λ ; w, µ) := hτ(ω − λ ), w − µi∂ Ωh ∀(ω, λ ), (w, µ) ∈ Wh × Mh .

Thanks to the structure of the method, it is easy to see that the solution (gh , uh , ubh ) ∈
Vh ×Wh × Mh (uD ) minimizes the functional

1
Jh (v, w, µ) := {Ah (v, v) + Sh (w, µ; w, µ)} − ( f , w)Ωh (5a)
2
over the functions (v, w, µ) in the space Vh × Wh × Mh (uD ) such that there exist
qh = qh (v, w, µ) ∈ Vh such that

Ah (v, p) + Bh (qh , p) = 0 ∀p ∈ Vh , (5b)


−Bh (p, v) + Bh (w, µ; p) = 0 ∀p ∈ Vh . (5c)

Once again, Note that the last two equations can be interpreted as the elimination
of (qh , gh ) from the equations. The minimization problem would then be one on the
affine space Wh × Mh (uD ) and would correspond to a problem formulated solely in
terms of uh and ubh . Next, we explore such reformulation.
28 Bernardo Cockburn

4.3.2 Formulation in terms of (uh , ubh )

We eliminate the approximate gradient gh and the approximate flux qh from the
equations defining the HDG method in order to formulate it solely in terms of
(uh , ubh ). To achieve that, we simply rewrite gh and qh as a linear mappings applied
to (uh , ubh ). These mappings are defined by using the first equation defining the HDG
methods. Thus, for any (w, µ) ∈ Wh × Mh , we define (Gw,µ , Qw,µ ) ∈ Vh × Vh as the
solution of

−(Gw,µ , v)Ωh − (w, ∇ · v)Ωh + hµ, v · ni∂ Ωh = 0 ∀ v ∈ Vh ,


(Qw,µ , v)Ωh = −(a Gw,µ , v)Ωh ∀ v ∈ Vh .

In this way, we are going to have that (qh , gh ) = (Quh ,buh , Guh ,buh ). Note that these two
equations are nothing but a rewriting of equations (5b) and (5c).

Static condensation formulation

We have the following result.

Theorem 8 (Second characterization of HDG methods). The approximate solu-


tion of the HDG method is given by

(qh , gh , uh ) = (Q, G, U) = (QUub ,bu , GUub uh , Uubh ) + (QU f ,0 , GU f ,0 , U f ),


,b
h h

where, on the element K ∈ Ωh , for any µ ∈ L2 (∂ K) and f ∈ L2 (K), the functions


Uµ , U f ∈ W (K) are the solutions of the local problems

(a GUµ ,µ , Gw,0 )K + hτ(Uµ − µ), wi∂ K = 0 ∀w ∈ W (K),


(a GU f ,0 , Gw,0 )K + hτ(U f ), wi∂ K = ( f , w)K ∀w ∈ W (K),

respectively. The function ubh is the element of Mh (uD ) such that

uh , µ) = `h (µ) ∀ µ ∈ Mh (0),
ah (b

where ah (µ, λ ) := −hµ, Q


b U ,λ · ni∂ Ω , and `h (µ) := hµ, Q
λ h
b U ,0 · ni∂ Ω . Moreover,
f h

ah (µ, λ ) = (aGUµ ,µ , GUλ ,λ )∂ Ωh + hUµ − µ, τ(Uλ − λ )i∂ Ωh , `h (µ)= ( f , Uµ ),

and ah (·, ·) is symmetric and positive definite on Mh (0) × Mh (0). Thus, ubh minimizes
the functional Jh (µ) := 21 ah (µ, µ) − `h (µ) over Mh (uD ).
Static condensation, hybridization, and the devising of the HDG methods 29

Compact formulation

Finally, we display the compact form of this formulation of the HDG method. We
have that (qh , gh ) = (Quh ,buh , Guh ,buh ) where (uh , ubh ) ∈ Wh × Mh (uD ) is the solution of

(a Guh ,buh , Gw,µ )Ωh +hτ(uh − ubh ), w − µi∂ Ωh = ( f , w)Ωh ∀(w, µ) ∈ Wh × Mh (0). (6)

In other words, (uh , ubh ) is the only minimum over Wh × Mh (0) of the functional

1
Jh (w, µ) := {(a Gw,µ , Gu,µ )Ωh + hτ(w − µ), w − µi∂ Ωh } − ( f , w)Ωh .
2
This is exactly the minimization problem (5).

5 Using Neumann instead of Dirichlet boundary conditions

In the previous two sections, we have shown how a characterization of the exact
solution can be used to generate HDG methods. Here we show how a different
characterization of the exact solution can be used to produce a different static con-
densation, that is , a different way of implementing, an already known HDG method.
We proceed as follows. First, we present a characterization of the exact solution
which uses Neumann boundary-value problems instead of the Dirichlet boundary-
value problems to define the local problems. Then, we consider some HDG methods
devised in the previous sections and show how a discrete version of the new charac-
terization of the exact solution is nothing but a new way of implementing them. The
resulting form of the HDG method has already been used in the work on multiscale
methods in [50]. Recently, two different ways of statically condensing the very same
method were proposed in [49].
The idea of using different characterizations of the exact solution to devise HDG
methods was introduced back in 2009 in [23] where four different ways were pre-
sented to devise HDG methods for the vorticity-velocity-pressure formulation of the
Stokes system, as the exact solution could be characterized in terms of four different
local problems and transmissions conditions. Just as it happens with the exact solu-
tion, the very same HDG method could be obtained by using any of the four ways.
In other words, the HDG method could be hybridized and then statically condensed
in each of the above-mentioned four different manners.

5.1 A second characterization of the exact solution

Let us then show how to use local Neumann boundary-value problems to obtain a
characterization of the exact solution.
30 Bernardo Cockburn

Suppose that, for every element K ∈ Ωh , we define (Q, U) as the solution of the
local problem

c Q + ∇U = 0 in K,
∇ · Q = f + {hb
q · n, 1i∂ K − ( f , 1)K }/|K| in K,
Q·n = q b·n on ∂ K,
(U, 1)K = (u, 1)K ,

where we want the function q b, which has a single-valued normal component, and
the constant u, to be such (q, u) = (Q, U) on K. This happens if and only if q
b and u
satisfy the equations

[[U]] = 0 for F ∈ Fhi ,


hb
q · n, 1i∂ K = ( f , 1)K for K ∈ Th ,
U = uD for F ∈ Fh∂ .

Note that we have to provide the average to U on the element, u, otherwise the solu-
tion U is not uniquely determined. Note also that, we have had to add an additional
term to the right-hand side of the second equation in order to ensure that the local
problem has a solution for any boundary data q b · n. As a consequence, we have to
make sure that such term is zero. This explains why the global problem consists not
only of transmission and boundary conditions, as in the case of Dirichlet boundary-
value local problems.
If we now separate the influence of q
b, u and f , we readily get the following char-
acterization of the exact solution.

Theorem 9 (Characterization of the exact solution). We have that

(q, u) = (Q, U) = (Qqb , Uqb ) + (0, u) + (Q f , U f ),

where (Qqb , Uqb ) and (Q f , U f ) are the solution of the local problems

c Qqb + ∇Uqb = 0 in K, c Q f + ∇U f = 0 in K,
∇ · Qqb = hb
q · n, 1i∂ K /|K| in K, ∇·Qf = f − ( f , 1)K /|K| in K,
Qqb · n = q
b·n on ∂ K, Qf ·n =0 on ∂ K,
(Uqb , 1)K = 0, (U f , 1)K = 0.

b · n and u are determined as the solution of the equations


where the functions q

− [[Uqb ]] − [[u]] = [[U f ]] on Fhi ,


hb
q · n, 1i∂ K = ( f , 1)K for K ∈ Th ,
Uqb + u = −U f + uD on Fh∂ .
Static condensation, hybridization, and the devising of the HDG methods 31

5.2 An example

In the case of our one-dimensional example, this result reads as follows. We have
that
(q, u) = (Qqb , Uqb ) + (0, u) + (Q f , U f ),
where
d d
c Qqb + U =0 in (xi−1 , xi ), cQf + Uf = 0 in (xi−1 , xi ),
dx qb dx
d 1 d 1 xi
Z
Q = (b q −q
bi−1 ) in (xi−1 , xi ), Qf = f − f in (xi−1 , xi ),
dx qb hi i dx hi xi−1
Qqb · n = q
b·n on {xi−1 , xi }, Q f · n = 0, on {xi−1 , xi },
Z xi Z xi
Uqb = 0, U f = 0,
xi−1 xi−1

and where the functions q


b and u are the solution of
Uqb (xi+ ) − Uqb (xi− ) + ui+1/2 − ui−1/2 = −U f (xi+ ) + U f (xi− ) for i = 1, . . . , N − 1,
Z xi
bi − q
q bi−1 = f for i = 1, . . . , N − 1,
xi−1

Uqb (x0+ ) + u1/2 = −U f (x0+ ) + uD (x0 ),


Uqb (xN− ) + uN−1/2 = −U f (x0+ ) + uD (xN ).

Since the solution of the local problems are


Z xi
Qqb (x) = ϕi (x)b
qi + ϕi−1 (x)b
qi−1 , Q f (x) = −c −1 Gix (x, s) f (s) ds,
xi−1
Z xi
c hi
Uqb (x) = {ψi (x)b
qi − ψi−1 (x)b
qi−1 } U f (x) = Gi (x, s) f (s) ds.
6 xi−1

where Gi is the Green’s function of the second local problem, namely,


(
c hi
i [1 − 3ϕi2 (s) − 3ϕi−1
2 (x)] if xi−1 ≤ s ≤ x,
G (x, s) := c6hi 2 2
6 [1 − 3ϕi (x) − 3ϕi−1 (s)] if x ≤ s ≤ xi .

and ψi := 1 − 3ϕi2 , and where the functions q


b and u are the solution of
c hi ch
(b bi ) + i+1 (2 q
qi−1 + 2 q bi+1 ) + ui+1/2 − ui−1/2 = −U f (xi+ ) + U f (xi− )
bi + q for i = 1, . . . , N − 1,
6 6
Z xi
bi − q
q bi−1 = f for i = 1, . . . , N − 1,
xi−1
c h1
(2b b1 ) + u1/2 = −U f (x0+ ) + uD (x0 ),
q0 + q
6
c hN
qN ) − uN−1/2 = U f (xN− ) − uD (xN ).
qN−1 − 2b
(b
6
32 Bernardo Cockburn

5.3 Another static condensation of known HDG methods

Let us consider the HDG methods introduced in Section 3. Next, we show that those
methods can be statically condensed in the way suggested by our new characteriza-
tion of the exact solution.

5.3.1 Rewriting the compact formulation based on the numerical traces

First, we rewrite them in such a way that the numerical trace q bh , and not ubh , is an
independent unknown. We can do that very easily if we use the compact formulation
of those methods based on the numerical traces, (3). It states that the approximate
solution given by the HDG method is the function (qh , uh , ubh ) ∈ Vh ×Wh × Mh (uD )
satisfying the equations

(c qh , v)Ωh − (uh , ∇ · v)Ωh + hb


uh , v · ni∂ Ωh = 0 ∀v ∈ Vh ,
−(qh , ∇w)Ωh + hb
qh · n, wi∂ Ωh = ( f , w)Ωh ∀w ∈ Wh ,
bh · n := qh · n + τ(uh − ubh )
q on ∂ Ωh ,
hµ, q
bh · ni∂ Ωh = 0 ∀µ ∈ Mh (0).

Now, if we take the stabilization function τ(·) to be the simple multiplication by the
scalar function τ, we have that

ubh = uh + τ −1 (qh · n − q
bh · n) on ∂ Ωh .

If the local space V(K) ×W (K) is such that, for each face F of the element K,

V(K) · n|F ⊂ M(F),


W (K)|F ⊂ M(F),

and take τ to be constant on each face of the triangulation, we have that q


bh belongs
to the space

Nh := {νν ∈ L2 (Fh ) : ν · n|∂ K ∈ M(∂ K), [[νν ]] = 0 on Fhi }.

We can thus rewrite the HDG method as follows. The approximate solution given
bh ) ∈ Vh ×Wh × Nh satisfying the equa-
by the HDG method is the function (qh , uh , q
tions

(c qh , v)Ωh − (uh , ∇ · v)Ωh + hb


uh , v · ni∂ Ωh = 0 ∀v ∈ Vh ,
−(qh , ∇w)Ωh + hb
qh · n, wi∂ Ωh = ( f , w)Ωh ∀w ∈ Wh ,
−1
ubh = uh + τ (qh · n − q
bh · n) on ∂ Ωh ,
hb
uh , · ni∂ Ωh = huD , ν · ni∂ Ω
ν ∀νν ∈ Nh .
Static condensation, hybridization, and the devising of the HDG methods 33

Note that the last equation enforces both the single-valuedness of ubh as well as the
Dirichlet boundary conditions of the model problem (1).

5.3.2 The new static condensation

So, suppose that, for every element K ∈ Ωh , we define (Q, U) ∈ V(K) ×W (K) to be
the solution of the local problem

(c Q, v)K − (U, ∇ · v)K + hU,b v · ni∂ K = 0 ∀v ∈ V(K),


−(Q, ∇w)K + hb qh · n, w − wi∂ K = ( f , w − w)∂ K ∀w ∈ W (K),
b := U + τ −1 (Q − q
U b )·n on ∂ K,
h
(U, 1)K = (uh , 1)K ,

where w|K := (w, 1)K /|K|, and where we want to take q bh ∈ Nh and the piecewise
constant function uh such that (qh , uh ) = (Q, U). Clearly, this happens if we have
that (b
qh , uh ) is the solution of the global problem

hνν · n, Ui
b ∂ Ω = hνν · n, uD i∂ Ω
h
∀νν ∈ Nh ,
hb
qh · n, 1i∂ K = ( f , 1)∂ K ∀ K ∈ Ωh .

Separating the influence of q


bh from that of uh and f , we obtain the following,
new static condensation of the HDG method. In what follows, W h denotes the space
of real-valued functions which are constant on each element K ∈ Ωh .

Theorem 10 (New static condensation of HDG methods). The approximate solu-


tion of the HDG method is

(qh , uh ) = (Q, U) = (Qqbh , Uqbh ) + (0, uh ) + (Q f , U f ),

where, for each element K ∈ Ωh , for any η ∈ L 2 (∂ K), the function (Qη , Uη ) ∈
V(K) ×W (K) is the solution of the local problem

(c Qη , v)K − (Uη , ∇ · v)K + hUb η , v · ni∂ K = 0 ∀v ∈ V(K),


−(Qη , ∇w)K + hη η · n, w − wi∂ K = 0 ∀w ∈ W (K),
b η := Uη + τ −1 (Qη − η ) · n
U on ∂ K,
(Uη , 1)K = 0,

and, for any f ∈ L2 (K), the function (Q f , U f ) ∈ V(K) ×W (K) is the solution of the
local problem
34 Bernardo Cockburn

(c Q f , v)K − (U f , ∇ · v)K + hU
b f , v · ni∂ K = 0 ∀v ∈ V(K),
−(Q f , ∇w)K = ( f , w − w)∂ K ∀w ∈ W (K),
U f := U f + τ −1 Q f · n
b on ∂ K,
(U f , 1)K = 0,

qh , uh ) ∈ Nh ×W h is the solution of the global problem


and where (b

qh , ν ) + bh (uh , ν ) = `h (νν ) − huD , ν · ni∂ Ω


ah (b ∀νν ∈ Nh ,
bh (ω, q
bh ) = ( f , ω)Ωh ∀ω ∈ W h ,

where

η , ν ) := −hνν ·n, U
ah (η b η i∂ Ω ,
h
bh (ω, ν ) := −hνν ·n, ωi∂ Ωh , `h (νν ) := hνν ·n, U
b f i∂ Ω .
h

Moreover,

η , ν ) = (c Qη , Qν )∂ Ωh + h(Qη − η ) · n, τ −1 (Qν − ν ) · ni∂ Ωh ,


ah (η `h (νν )= ( f , Uν )Ωh ,

and q
bh minimizes the complementary energy functional

1
Jh (νν ) := ah (νν , ν ) − `h (νν ) + huD , ν · ni∂ Ω ,
2
over the functions ν ∈ Nh such that bh (ω, ν ) = ( f , ω)Ωh ∀ω ∈ W h .

The proof of this result goes along the very same lines of the proof of the character-
ization Theorem 5.

5.3.3 The stabilized mixed compact formulation

Let us end this section by displaying the compact formulation of the method ob-
tained when we eliminate the numerical trace ubh . Proceeding as for the first charac-
terization, we can obtain that the approximate solution given by the HDG method is
bh ) ∈ Vh ×Wh × Nh satisfying the equations
the function (qh , uh , q

bh ; v, ν ) + Bh (uh ; v, ν ) = −huD , ν · ni∂ Ωh ,


Ah (qh , v) + Sh (qh , q (7a)
−Bh (w; qh , q
bh ) = ( f , w)Ωh , (7b)

for all (v, w, ν ) ∈ Vh ×Wh × Nh , where

Ah (p, v) := (c p, v)Ωh ∀p, v ∈ Vh , (7c)


Bh (w; v, ν ) := (∇w, v)Ωh − hw, ν · ni∂ Ωh ∀(v, w, ν ) ∈ Vh ×Wh × Nh , (7d)
Sh (p, η ; v, ν ) := h(p − η ) · n, τ −1 (v − ν ) · ni∂ Ωh ∀(p, η ), (v, ν ) ∈ Vh × Nh . (7e)
Static condensation, hybridization, and the devising of the HDG methods 35

bh ) ∈ Vh × Nh minimizes the complementary


As a consequence, the solution (qh , q
energy functional
1
Jh (v, ν ) := {Ah (v, v) + Sh (v, ν ; v, ν )} + huD , ν · ni∂ Ωh
2
over all functions (v, µ) in Vh × Mh (uD ) such that Bh (w; v, ν ) = ( f , w) ∀w ∈ Wh .

6 Building bridges and constructing methods

Here, we briefly discuss the evolution of the HDG methods. We being by show-
ing that (some of the earliest) HDG methods can be seen as a particular case of
the DG methods introduced in 1998 [39] and analyzed in 2000 [4]. We then recall
the strong relation between the HDG and the mixed methods, already pointed out
in 2009 [24], and show how this relation drove (and is still driving) the develop-
ment of superconvergent HDG methods. The bridge built in 2014 [13] between the
HDG and the so-called staggered discontinuous Galerkin (SDG), a DG method in-
troduced in 2009 [14] and apparently unrelated to the HDG methods, can be seen
as part of this development. We discuss the stabilization introduced by Lehrenfeld
(and Schöberl) in 2010 [63]. We end by showing that the so-called Weak-Galerkin
methods proposed in 2014 [89] and in 2015 [66, 67], are nothing but rewritings of
the HDG methods.

6.1 Relating HDG to old DG methods

Here, we consider HDG methods whose numerical method defining the local prob-
lems is the so-called local discontinuous Galerkin (LDG) method introduced in [39].
The resulting HDG methods are then called the LDG-H methods. For all of them,
the stabilization function τ on any face F ∈ Fh is a simple multiplication by a con-
stant which we also denote by τ, that is,

bh · n := qh · n + τ · (uh − ubh )
q on ∂ Ωh .

Examples of local spaces, taken from [24], are shown in the table below.
Method V(K) W (K) M(F)

LDG-H Pk−1 (K) Pk (K) Pk (F)


LDG-H Pk (K) Pk (K) Pk (F)
LDG-H Pk (K) Pk−1 (K) Pk (F)

In all these cases, we have that the local spaces V(K) ×W (K) are such that, for each
face F of the element K,
36 Bernardo Cockburn

V(K) · n|F ⊂ M(F),


W (K)|F ⊂ M(F).

This implies that [[qh ]] ∈ Mh and the transmission condition becomes [[b
qh ]] = 0 on
Fhi . This can only hold if and only if, on Fhi ,

τ + uh + + τ − uh − 1
ubh = + + [[qh ]],
τ+ + τ− τ + τ−
τ − qh + + τ + qh − τ +τ −
bh =
q + [[uh ]].
τ+ + τ− τ+ + τ−
This implies that the DG methods introduced in [39] and analyzed in [4] that have
the above choice of numerical traces can be hybridized and then statically con-
densed. This is why we call these methods the hybridizable DG methods.
Note, that none of these LDG-H methods is an LDG method if we take τ ± ∈
(0, ∞) since for the method to be an LDG method, we must have that 1/(τ + + τ − ) =
0. This shows that none of the DG methods considered in [3] is an LDG-H method
with finite values of the stabilization function. In fact, these methods can converge
faster than any of the DG methods considered therein. For example, in the case in
which c = Id, V(K) × W (K) = Pk (K) × Pk (K) and M(F) = Pk (F) this LDG-H
method was analyzed in [4], where is was proven that, for arbitrary shape-regular,
polyhedral elements, qh converges with order k + 1/2 and uh with order k + 1, for
any k ≥ 0, provided τ is of order one. The convergence is in the L2 (Ω )-norm. On the
other hand, other LDG-H methods do have the same order of convergence than those
considered in [3]. Indeed, by using the same approach in [4], one can easily show
that in the case in which V(K) × W (K) = Pk−1 (K) × Pk (K) and M(F) = Pk (F),
qh converges with order k and uh with order k + 1, for any k ≥ 1, provided τ is of
order 1/h. This result holds for meshes made of general shape-regular, polyhedral
meshes.

6.2 Relating HDG to mixed methods

As pointed out in [24], if the stabilization function τ is taken to be identically zero


so that qbh · n = qh · n on Fh , and the transmission condition implies that [[b
qh ]] = 0
on Fhi , we recover the so-called (hybridized version of the) mixed methods if the
mixed method is used to define the local problems; see also [2]. In the table below,
we display the main examples of mixed methods with this property when K is a
simplex and we compare it with one of the first HDG methods, the LDG-H method.
Static condensation, hybridization, and the devising of the HDG methods 37

Method V(K) W (K) M(F)

RT Pk (K) + x P
e k (K) Pk (K) Pk (F)
LDG-H Pk (K) Pk (K) Pk (F)
BDM Pk (K) Pk−1 (K) Pk (F)

The strong relation between the mixed method and the HDG methods suggested
that the HDG methods might share with the mixed methods some of its convergence
properties. This was proven to be true for a special LDG-H method obtained by
setting τ = 0 on all the faces of the simplex K except one. This method, called
the single face-hybridizable (SFH) method, was introduced and analyzed in [18].
Therein, it was shown that the SFH method is strongly related to the RT and BDM
mixed methods. Indeed, the bilinear forms ah (·, ·) of the RT, BDM and SFH methods
are the same, and the SFH shares with the RT and BDM the same superconvergence
properties.
Next, we briefly describe this superconvergence property. For all of the above
methods, the local averages of the error u − uh , converge faster than the errors u −
uh and q − qh . As a consequence, we can define, on the each element K, the new
approximation u?h ∈ W ∗ (K) := Pk+1 (K) as the solution of

(∇u?h , ∇w)K = − (c qh , ∇w)K for all w ∈ W ∗ (K),


(u?h , 1)K =(uh , 1)K ,

Then u − u?h will converge faster than u − uh . The orders of convergence are dis-
played in the table below; see [18] for the results on the SFH method and [26] for
those on the general LGD-H method. The symbol ? indicates that the non-zero val-
ues of the stabilization function τ only need to be uniformly bounded by below.

Method τ qh uh uh k

RT 0 k+1 k+1 k+2 ≥ 0


SFH ? k+1 k+1 k+2 ≥ 1
LDG-H O(1) k + 1 k + 1 k + 2 ≥ 1
BDM 0 k+1 k k+2 ≥ 2
LDG-H O(1/h) k k + 1 k + 1 ≥ 1

6.3 The SDG method as a limit of SFH methods

In [13], it was proved that the staggered discontinuous Galerkin (SDG) method,
originally introduced in the framework of wave propagation in [14], can be obtained
38 Bernardo Cockburn

as the limit when the non-zero values of the stabilization function of a special SFH
method goes to infinity. The special SFH method is obtained as follows. The mesh
consists of triangles or tetrahedra subdivided into three triangles or four tetrahedra.
On the faces of the bigger simplexes, the stabilization function is not zero; it is equal
to zero on all the remaining faces.
By building this bridge between the SDG and the SFH methods, the SDG can
now be implemented by hybridization and can share the superconvergence prop-
erties of the SFH method. Similarly, the SFH method now share the (related but
different) superconvergence property of the SDG method.

6.4 Constructing superconvergent HDG methods

The first superconvergent HDG method was the SFH method. A systematic ap-
proach to uncover superconvergent HDG methods was undertaken in [31] where
the following sufficient conditions were found. The space V(K) ×W (K) must have
a subspace V(K)
e ×W e (K) satisfying inclusions

P0 (K) ⊂ ∇W (K) ⊂ V(K),


e
P0 (K) ⊂ ∇ ·V
V (K) ⊂ W
e (K),
V (K) · n +W (K) ⊂ M(∂ K).

and whose orthogonal complement satisfies the identity

⊥ ⊥
Ve · n ⊕ W
e = M(∂ K).

Let us present examples taken from [31] in the case in which K is a cube; the first
corresponds to the choice M(F) = Qk (F) and the second to the choice M(F) =
Pk (F).

M(F) = Qk (F), k ≥ 1
method V (K) W (K)

RT[k] Pk+1,k,k (K) Qk (K)


×Pk,k+1,k (K)
×Pk,k,k+1 (K)
TNT[k] Q k (K) ⊕ H k7 (K) Qk (K)
HDGQ[k]
Q k (K) ⊕ H k6 (K) Qk (K)
Static condensation, hybridization, and the devising of the HDG methods 39

M(F) = Pk (F), k ≥ 1
method V (K) W (K)

BDFM[k+1] Pk+1 (K)\P


e k+1 (y, z) Pk (K)
×P (K)\P
k+1 e k+1 (x, z)
×Pk+1 (K)\P
e k+1 (x, y)
HDGP[k] P k (K) Pk (K)
⊕∇ × (yz Pe k (K), 0, 0)
⊕∇ × (0, zx Pe k (K), 0)
BDM[k] P k (K) Pk−1 (K)
k≥2 ⊕∇ × (0, 0, xy P (y, z))
e k

⊕∇ × (0, xz P
e k (x, y), 0)
⊕∇ × (yz Pe k (x, z), 0, 0)

In the first example, the HDG method denoted by HDGQ


[k]
and the mixed method
denoted by TNT[k] are new. The 7-dimensional space H k7 (K) is obtained by adding
a basis function to the space H k6 (K). The precise description of these spaces can be
found in [31] or, better, in [41], where commuting diagrams for the TNT elements
on cubes were obtained for the DeRham complex.
In the second example, the HDG method denoted by HDGP[k] is new. In the
corresponding table, we abuse the notation slightly to keep it simple. Thus, by
Pk+1 (K)\P e k+1 (y, z) we mean the span of {xα yβ zγ : α + β + γ ≤ k + 1, α > 0}.
In [31], many new superconvergence HDG methods were found for simplexes,
squares, cubes and prisms. For curved elements, see [32].

6.5 The Lehrenfeld-Schöberl stabilization function

Let us recall that the case in which M(F) := Pk (K) and V (K)×W (K) := Pk−1 (K)×
Pk (K), and the stabilization function τ is the multiplicative stabilization function,
namely,
τ(uh − ubh ) := τ · (uh − ubh ),
the resulting method is an LDG-H method. Moreover, for arbitrary shape-regular,
polyhedral elements, we have that qh converges with order k and uh with order k + 1,
for any k ≥ 0, provided τ is of order 1/h. Since the size of the stiffness matrix of
the local problem is proportional to the number of faces of the triangulation times
the dimension of the space M(F), a reduction of the space M(F) would result in a
smaller global problem. The question is if this is possible to achieve without loosing
the convergence properties of the method.
In 2010, Ch. Lehrenfeld (and J. Schöberl) [63, Remark 1.2.4] noted that the an-
swer is affirmative, see also [64], if we modify the above stabilization function by
40 Bernardo Cockburn

simply projecting uh into Mh :

τ LS (uh − ubh ) := h−1 · (PM (uh ) − ubh ).

The error analysis of this HDG method was carried out in 2014 by I. Oikawa [77]
who proved optimal orders of convergence for both qh and uh for regular-shaped,
general polyhedral elements.
For the sake of fairness in the attribution of this simple but remarkable projection,
I would like to emphasize that it was announced in 2009 by J. Schöberl in his plenary
talk at the ICOSAHOM in Trondheim, Norway; at the 2010 Finite Element Circus
in Minneapolis, USA; and then again at Oberwolfach, Germany, February 10-12,
2012; see [85]. I personally knew about it through Ch. Lehrenfeld, who told me
about it during a Ph.D. Course in Pavia, May 28- June 1, 2010. At that time, the error
estimates obtained later by I. Oikawa [77] were already known to Ch. Lehrenfeld
even though he did not include them in [63].

6.6 Relating HDG with the Weak Galerkin method

So far, no effort has been made to render clear the relation between the HDG and the
so-called Weak Galerkin methods. The first Weak Galerkin method was proposed
in 2013 [88] in the framework of convection-diffusion-reaction equations. Therein,
it is pointed out that the Weak Galerkin is identical to some mixed and HDG meth-
ods but only in the purely diffusion case and whenever the diffusivity tensor is a
constant. This is not an accurate statement which will be discussed elsewhere since
it requires addressing issues related to the convective and reaction terms. Instead,
here we restrict ourselves to discussing other versions of the Weak Galerkin method
devised specifically for steady-state diffusion in [66, 67, 89]. We show that all these
Weak Galerkin methods are rewritings of the HDG methods.
The Weak Galerkin method proposed in 2015 [66] (deposited in the archives
in 2012), was described therein as identical to the HDG methods for the Poisson
equation. Here we show that it is also identical for the model problem under consid-
eration. Indeed, it is nothing but the compact form of the HDG methods (6) in sub-
section 4.3.2 using the multiplication stabilization function τ(w− µ) := h−1 .(w− µ)
and the tensor a := c −1 . Let us point out that, although the HDG methods were intro-
duced in 2009 [24] (submitted in 2007) by using the formulation with the tensor c ,
the extension to the formulation with a := c −1 is straightforward. In fact, as argued
in subsection 4.1, these HDG methods can be obtained by reducing to the model
problem under consideration the HDG methods for the more difficult problem of
linear and nonlinear elasticity. Specifically, the HDG methods for elasticity were
obtained in 2008 in [86] and in 2009 [87] (submitted in 2008). The Weak Galerkin
method in [66] is thus a simple rewriting of HDG methods.
The Weak Galerkin method proposed in 2015 [67] (deposited in the archives in
2013) is nothing but the compact form of the HDG methods (6) in subsection 4.3.2
Static condensation, hybridization, and the devising of the HDG methods 41

using the Lehrenfeld-Schöberl stabilization function and the tensor a := c −1 . Thus,


the Weak Galerkin method in [67] is also a simple rewriting of HDG methods.
Finally, the Weak Galerkin method proposed in 2014 [89] (submitted in 2012)
is nothing but the HDG method (7) in subsection 5.3.3 corresponding to the
Lehrenfeld-Schöberl stabilization function and the tensor c . Although the results of
subsection 5.3 have been obtained when the stabilization function is a simple mul-
tiplication, the extension to the Lehrenfeld-Schöberl function are straightforward.
Indeed, the numerical trace for the HDG method with the Lehrenfeld-Schöberl sta-
bilization is
1
bh · n = qh · n + (PM (uh ) − ubh )
q
h
which implies that
ubh = PM (uh ) + h(qh − q
bh ) · n.
All the results of subsection 5.3 now follow from this simple identity and from the
fact that V(K) · n|F ⊂ M(F) for each face F of the triangulation. In other words, the
Weak Galerkin method in [89] is also a simple rewriting of HDG methods.
Let us end by pointing out that, by the previous argument, the Weak Galerkin in
[89] is identical to the Weak Galerkin method in [67] when the tensors c and a are
piecewise constant.

7 Bibliographical notes and ongoing work

After the introduction of the HDG methods in 2009 [24], we have extended the
methods to a variety of partial differential equations and introduced a variation of
the methods called the embedded discontinuous Galerkn (EDG) methods. The EDG
methods were introduced in 2007 [57] in the framework of linear shells, and then
analyzed in 2009 [28] for steady-state diffusion. (The HDG and EDG methods were
devised almost at the same time but the publication of the HDG methods [24] took
much more time than the publication of the EDG methods [28]).
The HDG methods for diffusion were devised and analyzed in [18, 29, 26, 34,
27, 36, 8, 31, 32, 59, 9, 10], multigrid methods for them in [21], a posteriori error
estimation for HDG methods in [43, 44, 45], and the convergence of adaptive HDG
methods in [30]. The implementation of the HDG methods in 2D was considered
in [61] and in 3D in [55]. The methods have been extended to convection-diffusion
in [20, 70, 71, 82], to the Stokes flow of incompressible fluids in [73, 72, 25, 15,
16, 81, 37, 38, 35], to the Oseen equations in [7], to the incompressible Navier-
Stokes equations in [69, 83], to the compressible Euler and Navier-Stokes Equations
[76, 78], to several problems in continuum mechanics in [57, 86, 87, 5, 6, 68, 37,
54, 60], to wave propagation in [74, 75, 33], to the biharmonic in [19] and to scalar
conservation laws in [65].
The current search for more efficient, superconvergent or optimally convergent,
HDG methods seems to be going in three main directions: (1) The refinement of
the sufficient conditions guaranteeing the superconvergence of the HDG methods
42 Bernardo Cockburn

through the so-called technique of M-decompositions [22], (2) the exploration of


the properties of the Lehrenfeld-Schöberl stabilization function [77, 79, 80, 64], and
(3) the exploration of the new, remarkable technique for devising numerical traces
for the hybrid high-order (HHO) methods [46, 48, 47].
In fact, a bridge between the HHO and HDG methods was recently established
in [17]. It would also be interesting to establish bridges with other numerical meth-
ods like, for example, the SUSHI methods [52], the elements constructed by Chris-
tiansen and Gillette [12], the BEM-based methods proposed in [84, 90], and the
methods introduced in [58] for multiscale problems.

Acknowledgements The author would like to thank Prof. Henryk K. Stolarski for providing the
earliest reference on static condensation. He would also thank Yanlai Chen, Mauricio Flores, Gu-
osheng Fu, Matthias Maier and an anonymous referee for feedback leading to a better presentation
of the material in this paper.

References

1. T. Arbogast, M.F. Wheeler, and I. Yotov, Mixed finite elements for elliptic problems with tensor
coefficients as cell-centered finite differences, SIAM J. Numer. Anal. 34 (1997), no. 2, 828–
852. MR 1442940 (98g:65105)
2. D.N. Arnold and F. Brezzi, Mixed and nonconforming finite element methods: implementation,
postprocessing and error estimates, RAIRO 19 (1985), 7–32.
3. D. N. Arnold, F. Brezzi, B. Cockburn, and L. D. Marini, Unified analysis of discontinuous
Galerkin methods for elliptic problems, SIAM J. Numer. Anal. 39 (2002), 1749–1779.
4. P. Castillo, B. Cockburn, I. Perugia, and D. Schötzau, An a priori error analysis of the local
discontinuous Galerkin method for elliptic problems, SIAM J. Numer. Anal. 38 (2000), 1676–
1706.
5. F. Celiker, B. Cockburn, and K. Shi, Hybridizable discontinuous Galerkin methods for Timo-
shenko beams, J. Sci. Comput. 44 (2010), 1–37.
6. , A projection-based error analysis of HDG methods for Timoshenko beams, Math.
Comp. 81 (2012), 131–151.
7. A. Cesmelioglu and B. Cockburn, Analysis of HDG methods for the Oseen equations, J. Sci.
Comput. 55 (2013), no. 2, 392–431.
8. B. Chabaud and B. Cockburn, Uniform-in-time superconvergence of HDG methods for the
heat equation, Math. Comp. 81 (2012), 107–129.
9. Y. Chen and B. Cockburn, Analysis of variable-degree HDG methods for convection-diffusion
equations. Part I: General nonconforming meshes, IMA J. Num. Anal. 32 (2012), 1267–1293.
10. , Analysis of variable-degree HDG methods for convection-diffusion equations. Part
II: Semimatching nonconforming meshes, Math. Comp. 83 (2014), 87–111.
11. Z. Chen, BDM mixed methods for a nonlinear elliptic problem, J. Comput. Appl. Math. 53
(1994), no. 2, 207–223. MR 1306126 (95i:65153)
12. S.H. Christiansen and A. Gillette, Constructions of some minimal finite element systems,
arXiv:1504.04670v1 (18 Apr., 2015).
13. E. Chung, B. Cockburn, and G. Fu, The staggered DG method is the limit of a hybridizable
DG method, SIAM J. Numer. Anal. 52 (2014), 915–932.
14. E. Chung and B. Engquist, Optimal discontinuous Galerkin methods for the acoustic wave
equation in higher dimensions, SIAM J. Numer. Anal. 47 (2009), no. 5, 3820–3848.
15. B. Cockburn and J. Cui, An analysis of HDG methods for the vorticity-velocity-pressure for-
mulation of the Stokes problem in three dimensions, Math. Comp. 81 (2012), 1355–1368.
Static condensation, hybridization, and the devising of the HDG methods 43

16. , Divergence-free HDG methods for the vorticity-velocity formulation of the Stokes
problem, J. Sci. Comput. 52 (2012), 256–270.
17. B. Cockburn, D. A. Di Pietro, and A. Ern, Bridging the hybrid high-order and hybridizable
discontinuous Galerkin methods, ESAIM Math. Model. Numer. Anal., to appear.
18. B. Cockburn, B. Dong, and J. Guzmán, A superconvergent LDG-hybridizable Galerkin method
for second-order elliptic problems, Math. Comp. 77 (2008), 1887–1916.
19. , A hybridizable and superconvergent discontinuous Galerkin method for biharmonic
problems, J. Sci. Comput. 40 (2009), 141–187.
20. B. Cockburn, B. Dong, J. Guzmán, M. Restelli, and R. Sacco, Superconvergent and opti-
mally convergent LDG-hybridizable discontinuous Galerkin methods for convection-diffusion-
reaction problems, SIAM J. Sci. Comput. 31 (2009), 3827–3846.
21. B. Cockburn, O. Dubois, J. Gopalakrishnan, and S. Tan, Multigrid for an HDG method, IMA
J. Numer. Anal. 34 (2014), no. 4, 1386–1425.
22. B. Cockburn, G. Fu and F.J. Sayas, Superconvergence by M-decompositions: General theory
for diffusion problems, submitted.
23. B. Cockburn and J. Gopalakrishnan, The derivation of hybridizable discontinuous Galerkin
methods for Stokes flow, SIAM J. Numer. Anal. 47 (2009), 1092–1125.
24. B. Cockburn, J. Gopalakrishnan, and R. Lazarov, Unified hybridization of discontinuous
Galerkin, mixed and continuous Galerkin methods for second order elliptic problems, SIAM
J. Numer. Anal. 47 (2009), 1319–1365.
25. B. Cockburn, J. Gopalakrishnan, N.C. Nguyen, J. Peraire, and F.-J. Sayas, Analysis of an HDG
method for Stokes flow, Math. Comp. 80 (2011), 723–760.
26. B. Cockburn, J. Gopalakrishnan, and F.-J. Sayas, A projection-based error analysis of HDG
methods, Math. Comp. 79 (2010), 1351–1367.
27. B. Cockburn, J. Guzmán, and F.-J. Sayas, Coupling of Raviart-Thomas and hybridizable dis-
continuous Galerkin methods with BEM, SIAM J. Numer. Anal. 50 (2012), 2778–2801.
28. B. Cockburn, J. Guzmán, S.-C. Soon, and H.K. Stolarski, An analysis of the embedded discon-
tinuous Galerkin method for second-order elliptic problems, SIAM J. Numer. Anal. 47 (2009),
2686–2707.
29. B. Cockburn, J. Guzmán, and H. Wang, Superconvergent discontinuous Galerkin methods for
second-order elliptic problems, Math. Comp. 78 (2009), 1–24.
30. B. Cockburn, R. Nochetto, and W. Zhang, Contraction property of adaptive hybridizable dis-
continuous Galerkin methods, Math. Comp., To appear.
31. B. Cockburn, W. Qiu, and K. Shi, Conditions for superconvergence of HDG methods for
second-order eliptic problems, Math. Comp. 81 (2012), 1327–1353.
32. , Conditions for superconvergence of HDG methods on curvilinear elements for
second-order eliptic problems, SIAM J. Numer. Anal. 50 (2012), 1417–1432.
33. B. Cockburn and V. Quenneville-Bélair, Uniform-in-time superconvergence of HDG methods
for the acoustic wave equation, Math. Comp. 83 (2014), 65–85.
34. B. Cockburn and F.-J. Sayas, The devising of symmetric couplings of boundary element and
discontinuous Galerkin methods, IMA J. Numer. Anal. 32 (2012), 765–794.
35. , Divergence–conforming HDG methods for Stokes flow, Math. Comp. 83 (2014),
1571–1598.
36. B. Cockburn, F.-J. Sayas, and M. Solano, Coupling at a distance HDG and BEM, SIAM J.
Sci. Comput. 34 (2012), A28–A47.
37. B. Cockburn and K. Shi, Conditions for superconvergence of HDG methods for Stokes flow,
Math. Comp. 82 (2013), 651–671.
38. , Devising HDG methods for Stokes flow: An overview, Computers & Fluids 98 (2014),
221–229.
39. B. Cockburn and C.-W. Shu, The local discontinuous Galerkin method for time-dependent
convection-diffusion systems, SIAM J. Numer. Anal. 35 (1998), 2440–2463.
40. , Runge-Kutta discontinuous Galerkin methods for convection-dominated problems, J.
Sci. Comput. 16 (2001), 173–261.
41. B. Cockburn and W. Qiu, Commuting diagrams for the TNT elements on cubes, Math. Comp.
83 (2014), 603–633.
44 Bernardo Cockburn

42. B. Cockburn, G. Fu and C. Xiong, Supercloseness of two formulations of DG, HDG and mixed
methods, in preparation.
43. B. Cockburn and W. Zhang, A posteriori error estimates for HDG methods, J. Sci. Comput.
51 (2012), no. 3, 582–607.
44. , A posteriori error analysis for hybridizable discontinuous Galerkin methods for sec-
ond order elliptic problems, SIAM J. Numer. Anal. 51 (2013), 676–693.
45. , An a posteriori error estimate for the variable-degree Raviart-Thomas method, Math.
Comp. 83 (2014), 1063–1082.
46. D. A. Di Pietro and A. Ern, A hybrid high-order locking-free method for linear elasticity on
general meshes, Comput. Meth. Appl. Mech. Engrg. 283 (2015), 1–21.
47. , Hybrid high-order methods for variable-diffusion problems on general meshes, C. R.
Acad. Sci Paris, Ser. I 353 (2015), 31–34.
48. D. A. Di Pietro, A. Ern, and S. Lemaire, An arbitrary-order and compact-stencil discretization
of diffusion on general meshes based on local reconstruction operators, Comput. Meth. Appl.
Math. 14 (2014), no. 4, 461–472.
49. , Arbitrary-order mixed method for heterogeneous anisotropic diffusion on genela
meshes, HAL Id: hal-00918482 (Aug. 1, 2015).
50. Y. Efendiev, R, Lazarov, and K. Shi, A multiscale HDG method for second order elliptic equa-
tions. part i. polynomial and homogenization-based multiscale spaces, SINUM 53 (2015),
no. 1, 342–369.
51. A. Ten Eyck and A. Lew, Discontinuous Galerkin methods for non-linear elasticity, Internat.
J. Numer. Methods Engrg. 67 (2006), 1204–1243.
52. R. Eymard, T. Gallouët, and R. Herbin, Discretization of heterogeneous and anisotropic dif-
fusion problems on general nonconforming meshes. SUSHI: a scheme using stabilization and
hybrid interfaces, IMA J. Numer. Anal. 30 (2010), no. 4, 1009–1043.
53. B. M. Fraejis de Veubeke, Displacement and equilibrium models in the finite element method,
Stress Analysis (O.C. Zienkiewicz and G. Holister, eds.), Wiley, New York, 1977, pp. 145–
197.
54. G. Fu, B. Cockburn, and H.K. Stolarski, Analysis of an HDG method for linear elasticity,
Internat. J. Numer. Methods Engrg., to appear.
55. Z. Fu, L.F. Gatica, and F.-J. Sayas, Algorithm 494: MATLAB tools for HDG in three dimen-
sions, ACM. Trans. Math. Softw. 41 Art. No. 20 (May 2015), 21 pages.
56. R.J. Guyan, Reduction of stiffness and mass matrices, J. Am. Inst. Aron. and Astro. 3 (1965),
380.
57. S. Güzey, B. Cockburn, and H.K. Stolarski, The embedded discontinuous Galerkin methods:
Application to linear shells problems, Internat. J. Numer. Methods Engrg. 70 (2007), 757–790.
58. Ch. Harder, D. Paredes, and F. Valentin, On a multi scale hybrid-mixed method for advective-
reactive dominated problems with heterogenous coefficients, Multiscale Model. Simul. 13
(2015), no. 2, 491–518.
59. L. N. T. Huynh, N. C. Nguyen, J. Peraire, and B. C. Khoo, A high-order hybridizable discon-
tinuous Galerkin method for elliptic interface problems, Internat. J. Numer. Methods Engrg.
93 (2013), 183–200.
60. H. Kabaria, A. Lew, and B. Cockburn, A hybridizable discontinuous Galerkin formulation for
nonlinear elasicity, Comput. Methods Appl. Mech. Engrg. 283 (2015), 303–329.
61. R.M. Kirby, S.J. Sherwin, and B. Cockburn, To HDG or to CG: A comparative study, J. Sci.
Comput. 51 (2012), no. 1, 183–212.
62. J. Koebbe, A computationally efficient modification of mixed finite element methods for flow
problems with full transmissivity tensors, Numer. Methods Partial Differential Equations 9
(1993), no. 4, 339–355. MR 1223251 (94c:76054)
63. C. Lehrenfeld, Hybrid discontinuous Galerkin methods for solving incompressible flow prob-
lems, Diplomigenieur Rheinisch-Westfalishen Technischen Hochchule Aachen (2010).
64. C. Lehrenfeld and J. Schöberl, High order exactly divergence-free hybrid discontinuous
Galerkin methods for unsteady incompressible flows, ASC Report No. 27 (2015).
65. D. Moro, N. C. Nguyen, and J. Peraire, A hybridized discontinuous Petrov-Galerkin scheme
for scalar conservation laws, Internat. J. Numer. Methods Engrg. 91 (2012), 950–970.
Static condensation, hybridization, and the devising of the HDG methods 45

66. L. Mu, J. Wang, and X. Ye, Weak Galerkin finite element methods on polytopal meshes, Int. J.
Numer. Anal. Model. 12 (2015), 31–53.
67. , A weak Galerkin finite element method with polynomial reduction, J. Comput. Appl.
Math. 285 (2015), 45–58.
68. N.C. Nguyen and J. Peraire, Hybridizable discontinuous Galerkin methods for partial differ-
ential equations in continuum mechanics, J. Comput. Phys. 231 (2012), 5955–5988.
69. N.C. Nguyen, J. Peraire, and B. Cockburn, A hybridizable discontinuous Galerkin method for
the incompressible Navier-Stokes equations (AIAA Paper 2010-362), Proceedings of the 48th
AIAA Aerospace Sciences Meeting and Exhibit (Orlando, Florida), January 2010-362.
70. N.C. Nguyen, J. Peraire, and B. Cockburn, An implicit high-order hybridizable discontinu-
ous Galerkin method for linear convection-diffusion equations, J. Comput. Phys. 228 (2009),
3232–3254.
71. , An implicit high-order hybridizable discontinuous Galerkin method for nonlinear
convection-diffusion equations, J. Comput. Phys. 228 (2009), 8841–8855.
72. , A comparison of HDG methods for Stokes flow, J. Sci. Comput. 45 (2010), 215–237.
73. , A hybridizable discontinuous Galerkin method for Stokes flow, Comput. Methods
Appl. Mech. Engrg. 199 (2010), 582–597.
74. , High-order implicit hybridizable discontinuous Galerkin methods for acoustics and
elastodynamics, J. Comput. Phys. 230 (2011), 3695–3718.
75. , Hybridizable discontinuous Galerkin methods for the time-harmonic Maxwell’s
equations, J. Comput. Phys. 230 (2011), 7151–7175.
76. , An implicit high-order hybridizable discontinuous Galerkin method for the incom-
pressible Navier-Stokes equations, J. Comput. Phys. 230 (2011), 1147–1170.
77. I. Oikawa, A hybridized discontinuous Galerkin method with reduced stabilization, J. Sci.
Comput. (published online on December 2014).
78. J. Peraire, N. C. Nguyen, and B. Cockburn, A hybridizable discontinuous Galerkin method for
the compressible Euler and Navier-Stokes equations (AIAA Paper 2010-363), Proceedings of
the 48th AIAA Aerospace Sciences Meeting and Exhibit (Orlando, Florida), January 2010-
363.
79. W. Qiu and K. Shi, An HDG method for convection-diffusion equations, J. Sci. Comput.,to
appear.
80. , An HDG method for linear elasticity with strongly symmetric stresses, submitted.
81. S. Rhebergen and B. Cockburn, A space-time hybridizable discontinuous Galerkin method for
incompressible flows on deforming domains, J. Comput. Phys. 231 (2012), 4185–4204.
82. , Space-time hybridizable discontinuous Galerkin method for the advection-diffusion
equation on moving and deforming meshes, The Courant-Friedrichs-Lewy (CFL) condition,
Birkhäuser/Springer, New York, 2013, pp. 45–63.
83. S. Rhebergen, B. Cockburn, and J.J.W. van der Vegt, A space-time discontinuous Galerkin
method for the incompressible Navier-Stokes equations, J. Comput. Phys. 233 (2013), 339–
358.
84. S. Rjasanow and S. Weißer, Higher order BEM-based FEM on polygonal meshes, SIAM J.
Numer. Anal. 50 (2012), 2357–2378.
85. J. Schöberl, private communication, 2015. The 2012 pdf file of J. Schöberl’s Oberwolfach talk
can be found in http://www.asc.tuwien.ac.at/ schoeberl/wiki/index.php/Talks.
86. S.-C. Soon, Hybridizable discontinuous Galerkin methods for solid mechanics, Ph.D. thesis,
University of Minnesota, Minneapolis, 2008.
87. S.-C. Soon, B. Cockburn, and H.K. Stolarski, A hybridizable discontinuous Galerkin method
for linear elasticity, Internat. J. Numer. Methods Engrg. 80 (2009), no. 8, 1058–1092.
88. J. Wang and X. Ye, A weak Galerkin finite element method for second order elliptic problems,
J. Comput. Appl. Math. 241 (2013), 103–115.
89. , A weak Galerkin mixed finite element method for second order elliptic problems,
Math. Comp. 289 (2014), 2101–2126.
90. S. Weißer, Arbitrary order Trefftz-like basis functions on polygonal meshes and realization in
BEM-based FEM, Comput. Math. Appl. 67 (2014), 1390–1406.
HDG methods for hyperbolic problems
∗ †
Bernardo Cockburn Ngoc Cuong Nguyen

Jaime Peraire
August 3, 2016

Abstract
We give a short overview of the development of the so-called hybridiz-
able discontinuous Galerkin methods for hyperbolic problems. We de-
scribe the methods, discuss their main features and display numerical
results which illustrate their performance. We do this in the framework
of wave propagation problems. In particular, we show that these methods
are amenable to static condensation, and hence to efficient implementa-
tion, both for time-dependent (with implicit time-marching schemes) as
well as for time-harmonic problems; we also show that they can be used
with explicit time-marching schemes. We discuss an unexpected, recently
uncovered superconvergence property and introduce a new postprocessing
for time-harmonic Maxwell’s equations. We end by providing bibliograph-
ical notes.

Keywords: discontinuous Galerkin methods, hybridization, hyperbolic problems.

1 Introduction
We give a short overview of the development of the so-called hybridizable dis-
continuous (HDG) methods for hyperbolic problems. The HDG methods are
discontinuous Galerkin methods which were originally devised for numerically
approximating steady-state problems and implicit time-marching schemes for
time-dependent problems. Their distinctive feature is that they are amenable
to static condensation and hence to efficient implementation. They turned out
to be more accurate that other DG methods, as will be shown below.
∗ School of Mathematics, University of Minnesota, Minneapolis, MN 55455, USA, email:

cockburn@math.umn.edu. Supported in part by the National Science Foundation (Grant DMS-


1522657) and by the University of Minnesota Supercomputing Institute.
† Department of Aeronautics and Astronautics, Massachusetts Institute of Technology, MA

02139, USA, email: cuongng@mit.edu. Supported in part by the Singapore-MIT Alliance.


‡ Department of Aeronautics and Astronautics, Massachusetts Institute of Technology, MA

02139, USA, email: peraire@mit.edu. Supported in part by the Singapore-MIT Alliance.

1
The HDG methods were introduced in [14] in the framework of steady-state
diffusion as part of the effort that started at the end of last century to devise
efficient DG methods for second-order elliptic problems. The development of
the HDG methods was then spearheaded by the authors who extended them
to a variety of problems in computational fluid dynamics including convection-
diffusion [33, 34], the incompressible Navier-Stokes equations [36, 40], and the
compressible Euler and Navier-Stokes equations [43, 41]; to partial differential
equations in continuum mechanics, see [32] and the references therein; and to
wave propagation problems in the time-domain [37, 46] as well as to the fre-
quency domain [39, 42].
In this paper, we describe the HDG methods, highlight some of their main
features and provide numerical experiments displaying their performance. In
particular, we show that they can be efficiently implemented, that they can
be used with either implicit or explicit time-marching schemes, and that they
do possess recently uncovered superconvergence properties. We do this for the
acoustic wave equation in Section 2, for the elastic wave equation in Section 3,
and for the time-harmonic Maxwell’s equation in Section 4. In Section 5, we
end with a few bibliographic notes.

2 The Acoustics Wave Equation


In this section we describe HDG methods for the numerical solution of the
acoustic wave equation
∂2u
ρ − ∇ · (κ∇u) = f, in Ω × (0, T ]. (1)
∂t2
By introducing the velocity v = ut and the flux q = −κ∇u, we can write (1) as
the following system of first-order equations:
∂q
κ−1 + ∇v = 0, in Ω × (0, T ],
∂t (2a)
∂v
ρ +∇·q = f, in Ω × (0, T ].
∂t
The exact solution (v, q) satisfies the following initial conditions
v(x, t = 0) = v0 (x),
(2b)
q(x, t = 0) = q0 (x),
and a Robin boundary condition
−q · n + αv = g, on ∂Ω × (0, T ]. (2c)
The coefficient α varies on the boundary ∂Ω and represents different types
of boundary conditions. Specifically, the Neumann boundary condition corre-
sponds to α = 0, the Dirichlet boundary condition to 1/α = 0, and the first-

order absorbing boundary condition to α = κρ. We assume that κ(x), ρ(x),
and α(x) are scalar positive functions.

2
We begin with the spatial discretization of the wave equation (2) and the
temporal integration of the semi-discrete form using both explicit and implicit
time-stepping methods. We end by presenting numerical experiments to demon-
strate their performance.

2.1 Spatial discretization


Let Th be a collection of disjoint elements that partition Ω. We denote by ∂Th
the set {∂K : K ∈ Th }. For an element K of the collection Th , F = ∂K ∩ ∂Ω
is the boundary face if the d − 1 Lebesgue measure of F is nonzero. For two
elements K + and K − of the collection Th , F = ∂K + ∩ ∂K − is the interior face
between K + and K − if the d − 1 Lebesgue measure of F is nonzero. Let Eho
and Eh∂ denote the set of interior and boundary faces, respectively. We denote
by Eh the union of Eho and Eh∂ .
Let Pk (D) denote the set of polynomials of degree at most k on a domain
D. We are going to use the following discontinuous finite element spaces:
Wh = {w ∈ L2 (Ω) : w|K ∈ W (K), ∀K ∈ Th },
2 d
Vh = {p ∈ (L (Ω)) : p|K ∈ V (K), ∀K ∈ Th }.
Some appropriate choices for the local space W (K) × V (K) on K include

 Pk (K) × (Pk (K))d ,
W (K) × V (K) ≡ Pk−1 (K) × (Pk (K))d , 

Pk (K) × (Pk (K))d + xPk (K) .
These spaces correspond to the equal-order elements, the BDM elements [1],
and the RT elements [44], respectively. In addition, we introduce a traced finite
element space
Mh = {µ ∈ L2 (Eh ) : µ|F ∈ Pk (F ), ∀F ∈ Eh }.
R
For functions w and v in (L2 (D))d , we denote
R (w, v)D = D w ·v. For functions
w and v inR L2 (D), we denote (w, v)D = D wv if D is a domain in Rd and
hw, viD = D wv if D is a domain in Rd−1 . We then introduce
X X
(w, v)Th = (w, v)K , hµ, ηi∂Th = hµ, ηi∂K ,
K∈Th K∈Th

for w, v defined on Th and µ, η defined on ∂Th .


The HDG methods for the wave equation (2) seek to define (qh , vh , vbh )(t) ∈
Vh × Wh × Mh , for t ∈ [0, T ], as a solution of the following system
 ∂qh 
κ−1 ,r − (vh , ∇ · r)Th + hb
vh , r · ni∂Th = 0, (3a)
∂t Th
 ∂v 
h
ρ ,w − (qh , ∇w)Th + hqbh · n, wi∂Th = (f, w)Th , (3b)
∂t Th

− hqbh · n, µi∂Th \∂Ω + h−qbh · n + αb


vh , µi∂Ω = hg, µi∂Ω , (3c)

3
for all (r, w, µ) ∈ Vh × Wh × Mh and all t ∈ (0, T ], where the numerical flux is
defined as

qbh · n = qh · n + τ (vh − vbh ), on ∂Th . (3d)



If the stabilization function is taken as τ = κρ, we obtain the well known
upwinding flux.
Note that the equations (2a) require v and the normal component of q to be
continuous across the set of interior faces Eho × (0, T ). The HDG method takes
into account these transmission conditions by requiring that the corresponding
numerical traces vbh and the normal component of qbh be single valued on that
set. The first condition is satisfied by taking vbh (t) in Mh and the second by
impossing equation (3c) for any t ∈ [0, T ]. For other ways of defining HDG
methods, see [8] and the references therein.
This semidiscretization gives rise to a system of ODEs to be solved by using
some time-marching methods. As we are going to see in the next subsection,
the form presented here is useful when using implicit time-marching methods
because it takes advantage of the fact that the HDG methods are amenable to
static condensation. When using explicit time-marching methods, a better way
of presenting the method is the following: Find (qh , vh ) ∈ Vh × Wh such that
for all K ∈ Th ,
 ∂qh 
κ−1 ,r − (vh , ∇ · r)K + hb
vh , r · ni∂K = 0, ∀ r ∈ V (K), (4a)
∂t K
 ∂v 
h
ρ ,w − (qh , ∇w)K + hqbh · n, wi∂K =(f, w)K , ∀ w ∈ W (K), (4b)
∂t K

where, for any given face F ∈ ∂K,


 + + − −
 τ vh + τ vh +
 1
(q + · n+ + qh− · n− ), if F ∈ Eho ,
vbh =
+
τ +τ − τ + τ− h
+
(4c)

 τ 1
vh + (Pg + qh · n), if F ∈ ∂Ω,
τ +α τ +α
and
qbh · n = qh · n + τ (vh − vbh ) on ∂K. (4d)
2
Here Pg denotes the L projection of g onto the space Mh , and

vh± |F = vh |F ∈∂K ± , and qh± |F = qh |F ∈∂K ± ,

where K + and K − are two elements sharing the face F . Hence, vh− and qh−
(respectively, vh+ and qh+ ) are nothing but the value of vh and qh on the face
F from the element K − (respectively, K + ). We can easily show that when the
stabilization function is taken to be a constant on each face lying on ∂Th , the
system (4) is equivalent to the original formulation (3) [37, 46].

4
2.2 Temporal discretization
We now show how to obtain a fully discrete scheme by discretizing the above
system of ODEs by several different time-marching methods, two being implicit
and the other two explicit.

BDF methods
We will only discuss the backward-Euler method since higher-order BDF meth-
ods follow a similar way. Using the backward-Euler scheme for the discretization
of the time derivative in (3), we find that the approximate solution (qhn , vhn , vbhn ) ∈
Vh × Wh × Mh at time step n satisfies the following equations
 qn   q n−1 
h
,r − (vhn , ∇ · r)Th + hb
vhn , r · ni∂Th = h
,r , (5a)
κ∆t Th κ∆t Th
 vn   v n−1 
ρ h ,w − (qhn , ∇w)Th + hqbhn · n, wi∂Th = f n + ρ h , w , (5b)
∆t Th ∆t Th

− hqbhn · n, µi∂Th \∂Ω + h−qbhn · n + αb


vhn , µi∂Ω = hg n , µi∂Ω , (5c)
for all (r, w, µ) ∈ Vh × Wh × Mh , where
qbhn · n = qhn · n + τ (vhn − vbhn ), on ∂Th . (5d)
Here (qhn , vhn , vbhn ) represents the numerical approximation to the exact solution
(q(tn ), u(tn ), vb(tn )) at time tn . We then find unh ∈ Wh such that
1 n 1 n−1
(uh , w)Th = (vhn , w)Th + (u , w)Th , ∀w ∈ Wh . (6)
∆t ∆t h
The fully discrete system (5) can be efficiently solved by locally eliminating
(qh , uh ) to obtain a linear system in terms of the globally coupled degrees of
freedom of vbh . We refer to [37] for a detailed discussion.

DIRK methods
Next, we apply the DIRK formula represented by the coefficients (aij , bi , ci ),
1 ≤ i ≤ q, 1 ≤ j ≤ i, to the semidiscrete system (3). We denote by (qhn.i , vhn,i , vbhn,i )
the numerical approximation to the exact solution (q(tn,i )|Th , v(tn,i )|Th , v(tn,i )|Eh ),
where tn,i = tn−1 + ci ∆t, 1 ≤ i ≤ q. Given the approximate solution at time
tn−1 , (qhn−1 , vhn−1 , vbhn−1 ), we find the intermediate solutions (qhn.i , vhn,i , vbhn,i ) ∈
Vh × Wh × Mh satisfying
 q n,i  D E  pn,i 
h
,v − (vhn,i , ∇ · v)Th + vbhn,i , v · n = h
,v , (7a)
κ∆t Th ∂Th κ Th
 ρv n,i  D E
h
,w − (qhn,i , ∇w)Th + qbhn,i · n, w = (f n,i + ρsn,i
h , w)Th , (7b)
aii ∆t Th ∂Th
D E D E

− qbhn,i · n, µ + −qbhn,i · n + αb vhn,i , µ = g n,i , µ ∂Ω , (7c)
∂Th \∂Ω ∂Ω

5
for all (v, w, µ) ∈ Vh × Wh × Mh , where

qbhn,i · n = qhn,i · n + τ (vhn,i − vbhn,i ), on ∂Th . (7d)

The terms sn,i n,i


h and ph on the right-hand side of (7) are given by

vhn−1 X aij  v n,j


i−1 
sn,i
h = + h
− sn,j
h , i = 1, . . . , q,
aii ∆t j=1 aii ajj ∆t

qhn−1 X aij  qhn,j 


i−1
pn,i
h = + − pn,j
h , i = 1, . . . , q.
aii ∆t j=1 aii ajj ∆t

The discrete systems (7) must be solved sequentially from i = 1, 2, . . . , q. Hence,


a q-stage DIRK scheme requires us to solve q discrete systems which are very
similar to the system (5) resulting from the backward-Euler method.
Once the intermediate solutions have been computed, the numerical solution
(qhn , vhn ) at time tn is given by
q
X
(qhn , vhn ) = (qhn−1 , vhn−1 ) + ∆t bi (yhn,i , zhn,i ), (8)
i=1

where
i−1
qhn,i − qhn−1 X aij n,j
yhn,i = − y , i = 1, . . . , q,
aii ∆t a h
j=1 ii
i−1
vhn,i − vhn−1 X aij n,j
zhn,i = − z , i = 1, . . . , q.
aii ∆t a h
j=1 ii

Finally, we compute unh by using the same DIRK scheme to discretize the ODE
∂uh /∂t = vh .

Adams–Bashforth methods
The Adams–Bashforth (AB) methods are linear multistep explicit methods. The
forward-Euler method is a first-order AB method. Here we discuss the forward-
Euler method since higher-order AB methods can be constructed in a similar
way. Given the solution at the previous time step (qhn , vhn , unh ), we first compute
the approximate traces as
 + +n − −n
 τ vh + τ vh −
 1
(q +n · n+ + qh−n · n− ), if F ∈ Eh \∂Ω,
n
vbh =
+
τ +τ − τ + τ− h
+

 τ 1
vn + (Pg n + αqhn · n), if F ∈ ∂Ω,
τ +α h τ +α
(9)

6
and qbhn · n = qhn · n + τ (vhn − vbhn ) for all faces F of Eh . We then determine the
numerical solution (qhn+1 , vhn+1 , un+1h ) ∈ V (K) × W (K) × W (K) at the next
time step by solving
 1 q n+1 − q n 
h h
− (vhn , ∇ · r)K + hb
,r vhn , r · ni∂K = 0,
κ ∆tn K
 v n+1 − v n 
ρ h h
,w − (qhn , ∇w)K + hqbhn · n, wi∂K = (f n , w)K , (10)
∆tn K
 un+1 − un 
h h
,z − (vhn , z)K = 0,
∆tn K

for all (r, w, z) ∈ V (K) × W (K) × W (K) and for all elements K ∈ Th .
It is clear that we compute the numerical solution at any time step in an
element-by-element fashion. Therefore, explicit HDG methods have the same
computational complexity as other explicit DG methods. Higher-order AB
methods can be used as well, provided that the numerical solutions at the earlier
time steps are available.

2.3 SSP-RK methods


Lastly, we describe the SSP-RK(q, q) scheme [4, 27] to integrate the semidiscrete
system (4) in time. For i = 0, . . . , q − 1, we compute
 + +n,i−1

 τ vh + τ − vh−n,i−1


 τ + τ−
+
n,i−1
vbh = + τ + +τ − (qh+n,i−1 · n+ + qh−n,i−1 · n− ),
1
if F ∈ Eh \∂Ω,



 τ 1
 v n,i−1 + (Pg n,i−1 + αqhn,i−1 · n), if F ∈ ∂Ω,
τ +α h τ +α
(11)
and qbhn,i−1 · n = qhn,i−1 · n + τ (vhn,i−1 − vbhn,i−1 ) for all faces F of Eh ; we then
determine (qhn,i , vhn,i , vhn,i ) ∈ V (K) × W (K) × W (K) as the solution of
 1 q n,i − q n,i−1    D E
h h
,r − vhn,i−1 , ∇ · r + vbhn,i−1 , r · n = 0,
κ ∆t K K ∂K
 v n,i − v n,i−1    D E
h h n,i−1 n,i−1
ρ ,w − qh , ∇w + qbh · n, w = (f n,i−1 , w)K ,
∆t K K ∂K
 v n,i − v n,i−1 
h h
,z − (vhn,i−1 , z)K = 0,
∆t K
(12)
for all (r, w, z) ∈ V (K) × W (K) × W (K) and for all elements K ∈ Th . We
finally set
s
X
(qhn , vhn , vhn ) = αq,i (qhn,i , vhn,i , vhn,i ), (13)
i=0

7
where the coefficients αq,i are precisely those corresponding to the SSPRK
scheme (q, q) [4, 27], namely
1
α1,0 = 1, αq,i = αq−1,i−1 , i = 1, . . . , q − 2,
i
q−1
X (14)
1
αq,q = , αq,q−1 = 0, αq,0 = 1 − αq,i .
q! i=1

The SSP-RK(q, q) scheme has q stages and q orders of accuracy. Each stage of
the SSP-RK(q, q) scheme is exactly the forward-Euler method described earlier.

2.4 Postprocessing
The numerical results we present in the next subsection are going to involve two
elementwise postprocessings defined as follows. The first is a new approximation
to u: On every simplex K ∈ Th , we take unh ∗ ∈ Pk+1 (K), such that

(∇unh ∗ , ∇w)K = (qhn , ∇w)K , ∀ w ∈ Pk+1 (K),


(15)
(unh ∗ , 1)K = (unh , 1)K .

The second is a new approximation to ut : On every simplex K ∈ Th , we take


vhn ∗ ∈ Pk+1 (K), such that

(∇vhn ∗ , ∇w)K = − (vhn , ∆w)K + hb


vhn , ∇w · ni∂K , ∀ w ∈ Pk+1 (K),
(16)
(vhn ∗ , 1)K = (vhn , 1)K .

As we are going to see, it turns out that both postprocessings u∗h and vh∗ have
better orders of convergence that the original approximations uh and vh , respec-
tively. Note that this local postprocessing can be performed at suitable time
steps, where these more accurate approximations are needed.

2.5 Numerical results


We consider the wave equation on a unit square Ω = (0, 1)×(0, 1) with boundary
condition v = 0 on ∂Ω and initial condition u0 = 0 and v0 = sin(πx) sin(πy).
For ρ = κ = 1 and f = 0, the problem has the following exact solution
1 √ √
u = √ sin(πx) sin(πy) sin( 2πt), v = sin(πx) sin(πy) cos( 2πt).

We use triangular meshes obtained by splitting a regular n × n Cartesian grid
into a total of 2n2 triangles, giving uniform element sizes of h = 1/n.
We present the L2 -errors and orders of convergence for the numerical ap-
proximations in Table 1 for the DIRK schemes and Table 2 for the SSP-RK
schemes. We observe that the approximate field variables converge with the op-
timal order k + 1, while the postprocessed displacement and velocity converge

8
with order k + 2. The HDG methods yield optimal convergence for the approx-
imate gradient, while many other DG methods provide suboptimal convergence
of order k. Furthermore, the postprocessed displacement and velocity converge
one order higher than the original approximations.
These convergence properties were first reported in [37] and later proven (for
the semidiscrete case) in [18]. A priori error estimates for v − vh∗ remain an open
problem though.

ku − uh kTh kv − vh kTh kq − qh kTh ku − u∗ kTh kv − v ∗ kTh


k 1/h error order error order error order error h order error h order
2 7.29e-3 −− 1.72e-2 −− 3.01e-2 −− 6.16e-3 −− 1.71e-2 −−
4 4.80e-4 3.92 2.16e-3 2.99 2.00e-3 3.91 2.77e-4 4.48 1.99e-3 3.11
2 8 4.47e-5 3.42 1.86e-4 3.54 1.84e-4 3.44 7.02e-6 5.30 1.40e-4 3.83
16 5.24e-6 3.09 1.81e-5 3.36 2.15e-5 3.10 2.54e-7 4.79 8.73e-6 4.00
32 6.36e-7 3.04 2.08e-6 3.12 2.61e-6 3.04 1.44e-8 4.14 5.36e-7 4.03
2 5.80e-4 −− 1.60e-3 −− 2.67e-3 −− 1.97e-4 −− 1.59e-3 −−
4 3.12e-5 4.22 8.22e-5 4.29 1.38e-4 4.27 4.92e-6 5.33 8.05e-5 4.30
3 8 1.78e-6 4.13 5.20e-6 3.98 7.74e-6 4.16 1.37e-7 5.17 3.78e-6 4.41
16 1.06e-7 4.07 3.32e-7 3.97 4.56e-7 4.08 4.05e-9 5.08 1.14e-7 5.05
32 6.46e-9 4.04 2.09e-8 3.99 2.77e-8 4.04 1.24e-10 5.03 1.50e-9 6.24

Table 1: History of convergence results using DIRK(k + 1, k + 2) schemes.

ku − uh kTh kv − vh kTh kq − qh kTh ku − u∗ kTh kv − v ∗ kTh


k 1/h error order error order error order error h order error h order
2 4.13e-3 −− 9.84e-3 −− 1.65e-2 −− 2.13e-3 −− 8.64e-3 −−
4 4.01e-4 3.37 1.06e-3 3.22 1.65e-3 3.32 1.02e-4 4.38 5.19e-4 4.06
2 8 4.44e-5 3.17 1.27e-4 3.06 1.83e-4 3.18 4.82e-6 4.40 2.80e-5 4.21
16 5.24e-6 3.08 1.60e-5 2.99 2.15e-5 3.09 2.59e-7 4.22 1.61e-6 4.12
32 6.36e-7 3.04 2.02e-6 2.99 2.61e-6 3.04 1.53e-8 4.08 9.81e-8 4.04
2 5.75e-4 −− 1.62e-3 −− 2.66e-3 −− 1.82e-4 −− 1.33e-3 −−
4 3.12e-5 4.21 8.22e-5 4.30 1.38e-4 4.27 4.63e-6 5.29 3.59e-5 5.21
3 8 1.78e-6 4.13 5.21e-6 3.98 7.74e-6 4.15 1.31e-7 5.15 1.03e-6 5.13
16 1.06e-7 4.07 3.32e-7 3.97 4.56e-7 4.08 3.88e-9 5.07 3.05e-8 5.07
32 6.46e-9 4.04 2.09e-8 3.99 2.77e-8 4.04 1.19e-10 5.03 8.97e-10 5.09

Table 2: History of convergence results using SSP-RK(k + 2, k + 2) schemes.

3 The Elastic Wave Equations


The elastic wave equations are different from the scalar acoustic wave equation
in that they are vectorial and have two different wave speeds, namely, pressure
(primary) wave speed and shear (secondary) wave speed. Although there are
several different formulations of the elastic wave equations, we will focus on
HDG methods for the displacement gradient-velocity-pressure formulation.
Let u represent the displacement field, λ and µ the Lamé moduli, ρ the
density of the elastic isotropic material, and b a time-dependent body force.
Let Ω be an open bounded domain in Rd and T a fixed final time. The motion

9
of the elastic isotropic body Ω is governed by:

∂2u
ρ − ∇ · [µ∇u + (µ + λ)(∇ · u)I] = b, in Ω × (0, T ]. (17)
∂t2
We introduce the velocity field v = ∂u/∂t, the displacement gradient tensor
H = ∇u, and the hydrostatic pressure p = (µ + λ)(∇ · u). We then rewrite (17)
as the first-order system
∂H
− ∇v = 0, in Ω × (0, T ],
∂t
∂v
ρ − ∇ · (µH + pI) = b, in Ω × (0, T ], (18)
∂t
∂p
ǫ −∇·v = 0, in Ω × (0, T ],
∂t
where ǫ = 1/(µ + λ), and I is the second-order identity tensor. Associated with
this system is the boundary condition

(µH + pI) · n + αv = g, on ∂Ω × (0, T ],

and initial condition

v = v0 , H = H0 , p = p0 , on Ω × {t = 0}.

For simplicity of exposition, we assume that ǫ > 0, which in essence means that
the elastic solid is either compressible or nearly incompressible. The incom-
pressible limit ǫ = 0 requires the average pressure condition and can be treated
by the augmented Lagrangian method [38, 35].

3.1 Spatial discretization


In addition to the finite element spaces defined in Section 2.2, we introduce the
following new finite element spaces:

Gh = {N ∈ (L2 (Th ))d×d : N|K ∈ (W (K))d×d , ∀K ∈ Th },


Mh = {µ ∈ (L2 (Eh ))d : µ|F ∈ (Pk (F ))d , ∀F ∈ Eh }.

We then define volume and boundary inner products associated with Gh as


X X
(N, L)Th = (N, L)K , hN, Li∂Th = hN, Li∂K ,
K∈Th K∈Th

for N, L ∈ (L2 (Th ))d×d . Note that (N, L)D denotes the integral of tr(NT L)
over D, where tr is the trace operator.

10
The HDG methods then find an approximation (Hh , vh , ph , v bh ) ∈ Gh ×Vh ×
Wh × Mh at time t such that
 
∂Hh
,N + (vh , ∇ · N)Th − hb vh , N · ni∂Th = 0, (19a)
∂t Th
 
∂vh
ρ ,w + (µHh + ph I, ∇w)Th
∂t Th
D E
− (µH b h + pbh I) · n, w = (b, w)Th , (19b)
∂Th
 
∂ph
ǫ ,q + (vh , ∇q)Th − hb vh · n, qi∂Th = 0, (19c)
∂t Th
D E
(µHb h + pbh I) · n, µ + hαb vh , µi∂Ω = hg, µi∂Ω , (19d)
∂Th

for all (N, w, q, µ) ∈ Gh × Vh × Wh × Mh , where


b h + pbh I) · n = (µHh + ph I) · n − S(vh − v
(µH bh ). (19e)
Here S is p a second-order tensor consisting of stabilization parameters which can
be set to (µ + λ)ρI.
The semidiscrete form (19) can be equivalently reformulated into finding
(Hh , vh , ph ) such that for all K ∈ Th ,
 
∂Hh
,N + (vh , ∇ · N)K − hb vh , N · ni∂K = 0, (20a)
∂t
K 
∂vh
ρ ,w + (µHh + ph I, ∇w)K
∂t K
D E
− (µH b h + pbh I) · n, w = (b, w)K , (20b)
∂K
 
∂ph
ǫ ,q + (vh , ∇q)K − hb vh · n, qi∂K = 0, (20c)
∂t K
where, for any given face F ∈ ∂K,
 + +

 τ vh + τ − vh−


 τ+ + τ−
1 + + + − − −
bh =
v − τ + +τ − ((µHh + ph I) · n + (µHh + ph I) · n ), if F ∈ Eho ,



 τ 1
 vh + (Pg − (µHh + ph I) · n), if F ∈ ∂Ω,
τ +α τ +α
(20d)
and
b h + pbh I) · n = (µHh + ph I) · n − S(vh − v
(µH bh ) on ∂K. (20e)
In this formulation, both v b
bh and (µHh + pbh I) · n are explicitly determined from
the numerical solution (Hh , vh , ph ).
While the first formulation (19) is useful for implicit time-stepping methods,
the second formulation (20) is convenient for explicit time-stepping methods.
Since the temporal discretization in this case is very similar to that in the scalar
wave equation, it will not be discussed here.

11
3.2 Local Postprocessing
As with the acoustic wave equation, we can define two new approximations
which will converge faster than the corresponding original approximations. The
postprocessed displacement field unh ∗ ∈ (Pk+1 (K))d satisfies, on every simplex
K ∈ Th ,

(∇unh ∗ , ∇w)K = (Hnh , ∇w)K , ∀ w ∈ (Pk+1 (K))d ,


(21)
(unh ∗ , 1)K = (unh , 1)K .

The postprocessed velocity field vhn ∗ ∈ (Pk+1 (K))d is obtained by locally solving

(∇vhn ∗ , ∇w)K = −(vhn , ∆w)K + hb


vhn , ∇w · ni∂K ∀ w ∈ (Pk+1 (K))d ,
(22)
(vhn ∗ , 1)K = (vhn , 1)K .

Since the local postprocessing can be carried out at any particular timestep and
performed at the element level, the postprocesssed displacement and velocity
are very inexpensive to compute. Note that the postprocessing is effective only
if the temporal accuracy is of order k + 2.

3.3 Numerical Results


We consider the elastic wave equations on a unit square Ω = (0, 1) × (0, 1) with
µ = 1 and ρ = 1. The exact solution is given by

u1 = −x2 y(2y − 1)(x − 1)2 (y − 1) sin(πt),


u2 = xy 2 (2x − 1)(x − 1)(y − 1)2 sin(πt).

The source term b is determined from the above solution. The Dirichlet bound-
ary data are determined as the restriction of the exact solution on the boundary.
Likewise the initial data is taken as an instantiation of the exact solution at time
t = 0. Our triangular meshes are constructed upon regular n×n Cartesian grids
(h = 1/n). The stabilization parameter is set to τ = 1.

ku − uh kTh kv − vh kTh kσ − σh kTh ku − u∗ kTh kv − v ∗ kTh


k 1/h error order error order error order error h order error horder
4 3.79e-4 −− 1.94e-3 −− 2.08e-3 −− 1.74e-4 −− 1.28e-3 −−
8 1.12e-4 1.76 4.51e-4 2.11 5.07e-4 2.04 2.53e-5 2.78 1.74e-4 2.88
1 16 3.04e-5 1.88 1.06e-4 2.09 1.26e-4 2.01 3.27e-6 2.95 2.18e-5 2.99
32 7.90e-6 1.94 2.60e-5 2.03 3.16e-5 2.00 4.12e-7 2.99 2.96e-6 2.89
64 2.01e-6 1.97 6.45e-6 2.01 7.93e-6 2.00 5.16e-8 3.00 3.99e-7 2.89
4 5.14e-5 −− 2.26e-4 −− 3.27e-4 −− 1.78e-5 −− 2.41e-4 −−
8 8.01e-6 2.68 2.90e-5 2.96 4.21e-5 2.96 1.20e-6 3.89 7.10e-6 5.08
2 16 1.10e-6 2.87 3.67e-6 2.98 5.25e-6 3.00 7.39e-8 4.02 4.53e-7 3.97
32 1.43e-7 2.94 4.60e-7 3.00 6.54e-7 3.01 4.52e-9 4.03 2.70e-8 4.07
64 1.82e-8 2.97 5.75e-8 3.00 8.14e-8 3.00 2.78e-10 4.02 1.68e-9 4.01

Table 3: History of convergence results for a compressible material (λ = 1).

12
ku − uh kTh kv − vh kTh kσ − σh kTh ku − u∗ kTh kv − v ∗ kTh
k 1/h error order error order error order error h order error horder
4 3.75e-4 −− 1.94e-3 −− 2.2e-3 −− 1.72e-4 −− 1.26e-3 −−
8 1.12e-4 1.75 4.49e-4 2.11 5.41e-4 2.02 2.57e-5 2.74 1.71e-4 2.89
1 16 3.04e-5 1.88 1.06e-4 2.08 1.33e-4 2.02 3.37e-6 2.93 2.13e-5 3.00
32 7.90e-6 1.94 2.60e-5 2.03 3.33e-5 2.00 4.26e-7 2.98 2.87e-6 2.89
64 2.01e-6 1.97 6.45e-6 2.01 8.33e-6 2.00 5.34e-8 2.99 3.85e-7 2.90
4 5.11e-5 −− 2.24e-4 −− 3.67e-4 −− 1.80e-5 −− 2.40e-4 −−
8 7.98e-6 2.68 2.88e-5 2.96 4.82e-5 2.93 1.22e-6 3.89 6.91e-6 5.12
2 16 1.09e-6 2.87 3.66e-6 2.98 6.12e-6 2.98 7.44e-8 4.03 4.20e-7 4.04
32 1.43e-7 2.94 4.59e-7 2.99 7.89e-7 2.96 4.52e-9 4.04 2.48e-8 4.08
64 1.82e-8 2.97 5.75e-8 3.00 9.95e-8 2.99 2.78e-10 4.02 1.48e-9 4.07

Table 4: History of convergence results for a nearly incompressible material


(λ = 1000).

We present the L2 -errors and orders of convergence of the numerical ap-


proximations at the final time T = 0.5 in Table 3 for λ = 1 (compressible case)
and in Table 4 for λ = 1000 (nearly incompressible case). These results are
obtained using the DIRK(2,3) scheme for k = 1 and the DIRK(3,4) scheme for
k = 2, and a fixed ratio h/∆t = 4. We observe that the approximate field vari-
ables converge with the optimal order k + 1 even for the nearly incompressible
case. Furthermore, we observe that both the postprocessed displacement and
velocity converge with order k + 2, which are one order higher than the orig-
inal approximations. Since the local postprocessing is inexpensive, the HDG
methods provide better convergence and accuracy than other DG methods.
These convergence properties were first reported in [37]. For the semidiscrete
case, they can be obtained by an analysis similar to that for the acoustic wave
equation in [18]. Again, a priori error estimates for v − vh∗ remain an open
problem.

4 The Electromagnetic Wave Equations


In this section, we restrict our attention to devising HDG methods for the
Maxwell’s equations in frequency domain. Numerical treatment of the Maxwell’s
equations in time domain follows from the discussion in this section and the
second section.
Let us consider the time-harmonic Maxwell’s equations in a connected and
bounded domain Ω ∈ R3 with zero charge density and zero conductivity:

∇ × E + iµωH = 0, ∇ × H − iǫωE = J , in Ω ⊂ R3 , (23)

where E, H, and J are the electric field, magnetic field, and current source,
respectively. In addition, ω is the frequency, ǫ the permittivity, and µ the
permeability of the medium. We assume that the electromagnetic field satisfies
the following impedance condition

−n × H + αn × E × n = g, on ∂Ω, (24)

13
for some given scalar function α and vectorial function g.

4.1 Numerical discretization


To define the numerical approximation of the Maxwell’s equations (23), we
introduce the following approximation spaces
Vh := {v ∈ [L2 (Th )]3 : v|K ∈ [Ck (K)]3 , ∀ K ∈ Th },
Mht := {η ∈ [L2 (Eh )]3 η|F ∈ [Ck (F )]3 , (η · n)|F = 0, ∀ F ∈ Eh }.
:
(25)
Heret Ck (D) denote the space of complex-valued polynomials of degree at most
k on D. We then define the inner products for our approximation spaces as
X Z 3
X
(w, v)Th := wv, (w, v)Th := (wj , vj )Th ,
K∈Th K j=1
(26)
X Z X3
hw, vi∂Th := wv, hw, vi∂Th := hwj , vj i∂Th .
K∈Th ∂K j=1

Here the bar denotes a complex conjugate which is applied only to the second
argument of the inner products.
Note that Mht consists of vector-valued functions whose normal component
is zero on any face F ∈ Eh . Hence, a member of Mht can be characterized by
two tangential vectors on the faces: if tF F
1 and t2 denote independent tangent
vectors on F , we can write the restriction of η ∈ Mht on F as

η|F = η1F tF F F
1 + η 2 t2 , (27)

where η1F ∈ Ck (F ) and η2F ∈ Ck (F ) are complex-valued polynomials of degree


at most k on F . Hence, the vector-valued function η ∈ Mht is characterized by
two scalar functions η1 and η2 .
The HDG method seeks (E h , Hh , E b t ) ∈ Vh × Vh × M t such that
h h
D E
(iωµH h , R)Th + (Eh , ∇ × R)Th + E bt , R × n = 0, (28a)
h
∂Th
D E
(H h , ∇ × W )Th + H ch , W × n − (iǫωEh , W )Th = (J , W )Th , (28b)
∂Th
D E D E
− n×H ch , η + αE bt ,η = hg, ηi∂Ω , (28c)
h
∂Th ∂Ω

for all (R, W , η) ∈ Vh × Vh × Mht , where


c h = H h + τ (Eh − E
H b t ) × n. (28d)
h
q
ǫω 2
Here the stabilization parameter is chosen as τ = µ . This completes the
HDG method for solving the time-harmonic Maxwell’s equations.
The structure of the HDG method makes itself amenable to an efficient
implementation. Note that the first two equations in (28) can be thought as

14
bh . The equation (28c) is then the equation
characterizing (Eh , Hh ) in terms of E
that determines the actual values of the unknown E bh . In this manner, the only
globally-coupled degrees of freedom are those of E bh . As a result, the HDG
method can provide more accurate solutions at much lower computational cost
than standard frequency-domain DG method.

4.2 Local postprocessing


We propose a new local postprocessing to obtain new approximations of the
electric and magnetic fields, which converges with an additional order in the
Hcurl (Th )-norm. A remarkable feature of this new local postprocessing is that it
is effective even in three dimensions, whereas the local postprocessing introduced
in our previous work [39] is only applicable in two dimensions.
We find the new approximate electric field Eh∗ as the element of [Ck+1 (K)]3
such that for all K ∈ Th ,

(∇ × Eh∗ , W )K = −(iµωHh , W )K , ∀ W ∈ ∇ × [Ck+1 (K)]3 , (29a)


(Eh∗ , Y )K = (Eh , Y )K , ∀ Y ∈ ∇Ck+2 (K). (29b)

Similarly, we find the new approximate magnetic field Hh∗ as the element of
[Ck+1 (K)]3 such that for all K ∈ Th ,

(∇ × Hh∗ , W )K = (iǫωEh + J , W )K , ∀ W ∈ ∇ × [Ck+1 (K)]3 , (30a)


(Hh∗ , Y )K = (Hh , Y )K , ∀ Y ∈ ∇Ck+2 (K). (30b)

It is obvious that ∇ × Eh∗ and ∇ × Hh∗ are nothing but the projection of
iµωHh and iǫωEh + J , respectively, onto the space of divergence-free functions
in [Pk+1 (K)]3 . Therefore, we expect that both Eh∗ and Hh∗ converge with order
k + 1 in the Hcurl (Th )-norm, whereas Eh and Hh converge with order k in the
Hcurl (Th )-norm.

4.3 Numerical results


We consider the time-harmonic Maxwell’s equations on a unit cube Ω = (0, 1)3
with µ = 1, ǫ = 2, α = 0, and ω = 1. For J = 0 the problem has the exact
solution
Ex = sin(ωy) sin(ωz), Hx = i sin(ωx)(cos(ωy) − cos(ωz)),
Ey = sin(ωx) sin(ωz), Hy = i sin(ωy)(cos(ωz) − cos(ωx)),
Ez = sin(ωy) sin(ωx), Hz = i sin(ωz)(cos(ωx) − cos(ωy)),

The boundary data g is determined from the exact solution. The tetrahedral
meshes are constructed upon regular n × n × n Cartesian grids (h = 1/n) by
splitting each cube into 6 tetrahedral.
We present the L2 -errors and orders of convergence of the numerical approxi-
mations in Table 5 and the postprocessed quantities in Table 6. We observe that
the approximate electric and magnetic fields converge with order k + 1 in the

15
L2 -norm, but only order k in the Hcurl (Th )-norm. Furthermore, we observe that
the postprocessed electric and magnetic fields converge with order k + 1 in the
Hcurl (Th )-norm, which are one order higher than the original approximations.
The theoretical justification of these results is still an open problem.

kE − Eh kTh kE − Eh kHcurl (Th ) kH − Hh kTh kH − Hh kHcurl (Th )


k 1/h error order error order error order error order
2 2.94e-2 −− 9.90e-2 −− 8.41e-3 −− 2.20e-1 −−
4 7.77e-3 1.92 4.46e-2 1.15 2.18e-3 1.95 1.10e-1 1.00
1 6 1.94e-3 2.00 2.14e-2 1.06 5.85e-4 1.90 5.52e-2 1.00
8 4.81e-4 2.01 1.05e-2 1.02 1.54e-4 1.93 2.76e-2 1.00
2 9.49e-4 −− 1.32e-2 −− 6.56e-4 −− 3.28e-2 −−
4 1.33e-4 2.84 3.37e-3 1.97 8.74e-5 2.91 8.15e-3 2.01
2 6 1.90e-5 2.81 8.47e-4 1.99 1.12e-5 2.96 2.03e-3 2.00
8 2.87e-6 2.73 2.12e-4 2.00 1.42e-6 2.98 5.09e-4 2.00
2 8.72e-5 −− 1.40e-3 −− 5.51e-5 −− 1.74e-3 −−
4 5.59e-6 3.96 1.73e-4 3.02 3.51e-6 3.97 2.28e-4 2.93
3 6 3.53e-7 3.99 2.15e-5 3.01 2.23e-7 3.98 2.92e-5 2.97
8 2.22e-8 3.99 2.67e-6 3.00 1.41e-8 3.99 3.69e-6 2.98

Table 5: History of convergence results for the approximate solution.

kE − Eh∗ kTh kE − Eh∗ kHcurl (Th ) kH − Hh∗ kTh kH − Hh∗ kHcurl (Th )
k 1/h error order error order error order error order
2 3.19e-2 −− 3.44e-2 −− 1.05e-2 −− 6.26e-2 −−
4 8.42e-3 1.92 9.05e-3 1.93 2.69e-3 1.97 1.67e-2 1.90
1 6 2.10e-3 2.00 2.27e-3 1.99 7.05e-4 1.93 4.21e-3 1.99
8 5.23e-4 2.01 5.68e-4 2.00 1.83e-4 1.95 1.05e-3 2.00
2 9.56e-4 −− 1.58e-3 −− 8.34e-4 −− 2.06e-3 −−
4 1.34e-4 2.84 2.07e-4 2.93 1.08e-4 2.95 2.82e-4 2.87
2 6 1.91e-5 2.81 2.76e-5 2.91 1.38e-5 2.97 3.85e-5 2.87
8 2.88e-6 2.73 3.81e-6 2.86 1.74e-6 2.99 5.46e-6 2.82
2 8.36e-5 −− 1.03e-4 −− 4.88e-5 −− 1.75e-4 −−
4 5.43e-6 3.95 6.71e-6 3.95 3.20e-6 3.93 1.13e-5 3.94
3 6 3.44e-7 3.98 4.26e-7 3.98 2.05e-7 3.97 7.20e-7 3.98
8 2.17e-8 3.99 2.69e-8 3.99 1.29e-8 3.98 4.54e-8 3.99

Table 6: History of convergence results for the postprocessed solution.

5 Bibliographic notes
Time-dependent wave propagation
The devising of HDG methods for the acoustic wave equation was carried out
as an almost immediate consequence of the introduction of HDG methods for
steady-state diffusion. After all, both equations share the same second-order
strongly elliptic operator. However, not all convergence properties which hold

16
for HDG methods for steady-state diffusion problems [9, 16, 15, 5, 6, 17, 19] can
be immediately obtained for time-dependent wave equations. In particular, the
wave equation does not have a smoothing effect which could generate supercon-
vergence results, as happens for the heat equation, see [3]. However, in [18], it
was shown how to obtain the superconvergence results we have illustrated in
Section 2; a comparison with other mixed and DG methods can also be found
there. Although therein we only used simplexes and spaces of polynomials of
degree k, similar convergence and superconvergence results do hold for meshes
made of elements of arbitrary shape. This can be obtained by using the so-
called theory of M-decompositions developed in [12, 10, 11]. In a similar way,
HDG methods for the elastic wave equation can be easily obtained once HDG
methods for linear elasticity [45, 32, 20, 26] are obtained.
The first HDG methods for wave propagation were proposed in [37], where
implicit time-marching methods were used, and in [46], where explicit time-
marching methods were used. In both papers, the superconvergence properties
of the semidiscrete method uncovered in [18] were shown to hold for the corre-
sponding implicit and explicit time-marching schemes, respectively.
The HDG methods we have presented here can be also used with PML
absorbing boundary conditions, as shown in [37]. HDG methods which are not
dissipative, and have similar superconvergence properties, have been developed
in [13].

Time-harmonic wave propagation


HDG methods for time-harmonic hyperbolic equations are also strongly related
to the HDG methods for steady-state diffusion problems. The first HDG method
for the Helmholtz equation was introduced in [28]. The same optimal conver-
gence and superconvergence properties of the associated steady-state diffusion
were proven. In [25], a wide family of discontinuous Galerkin methods, which
included the HDG methods, were proven to be stable regardless of the wave
number. The methods used piecewise linear approximations. In [21], an analy-
sis of the HDG methods for the Helmholtz equations was carried which shows
that the method is stable for any wave number, mesh and polynomial degree
and which recovers the orders of convergence and superconvergence obtained
previously in [28]. A method for arbitrarily large wave numbers is proposed in
[42].
The first HDG for the time-harmonic Maxwell’s equations was proposed in
[39] in two-space dimensions. The extension of the method to three-dimensions
was done in [30]. A variation was introduced in [31]. HDG method for the
time-harmonic equations of elastodynamics can be found in [29].

Further reading material


A systematic way of defining HDG methods for Friedrichs’ systems has been
developed in [2]. A general construction of DG methods for these problems
can be found in [22, 23, 24]. An overview of the development of DG (and in

17
particular, HDG) methods for fluid dynamics can be found in [7]. An overview
of the HDG methods for steady-state diffusion can be found in [8]. Therein, the
relation between static condensation, hybridization and the devising of HDG
methods is carefully explored.

Acknowledgements. The authors would like to thank Rémi Abgrall and


Chi-Wang Shu for the invitation to write this paper.

References
[1] D. N. Arnold and F. Brezzi. Mixed and nonconforming finite element
methods: implementation, postprocessing and error estimates. RAIRO
Modél. Math. Anal. Numér., 19:7–32, 1985.

[2] T. Bui-Thanh. From Godunov to a unified hybridized discontinuous


Galerkin framework for partial differential equations. J. Comput. Phys.,
295:114–146, 2015.

[3] B. Chabaud and B. Cockburn. Uniform-in-time superconvergence of HDG


methods for the heat equation. Math. Comp., 81:107–129, 2012.

[4] M.-H. Chen, B. Cockburn, and F. Reitich. High-order RKDG methods for
computational electromagnetics. J. Sci. Comput., 22/23:205–226, 2005.

[5] Y. Chen and B. Cockburn. Analysis of variable-degree HDG methods


for convection-diffusion equations. Part I: General nonconforming meshes.
IMA J. Num. Anal., 32:1267–1293, 2012.

[6] Y. Chen and B. Cockburn. Analysis of variable-degree HDG methods


for convection-diffusion equations. Part II: Semimatching nonconforming
meshes. Math. Comp., 83:87–111, 2014.

[7] B. Cockburn. Discontinuous Galerkin methods for computational fluid dy-


namics. In E. Stein, R.de Borst, and T.J.R. Hughes, editors, Encyclopedia
of Computational Mechanics, Second Edition, volume 3. John Wiley &
Sons, Ltd., England, 2016. 111 pages. To appear.

[8] B. Cockburn. Static condensation, hybridization, and the devising of the


HDG methods. In G.R. Barrenechea, F. Brezzi, A. Cagniani, and E.H.
Georgoulis, editors, Building Bridges: Connections and Challenges in Mod-
ern Approaches to Numerical Partial Differential Equations, volume 114
of Lect. Notes Comput. Sci. Engrg. Springer Verlag, Berlin, 2016. LMS
Durham Symposia funded by the London Mathematical Society. Durham,
U.K., on July 8–16, 2014. 51 pages.

[9] B. Cockburn, B. Dong, and J. Guzmán. A superconvergent LDG-


hybridizable Galerkin method for second-order elliptic problems. Math.
Comp., 77:1887–1916, 2008.

18
[10] B. Cockburn and G. Fu. Superconvergence by M-decompositions. Part
II: Construction of two-dimensional finite elements. Modél. Math. Anal.
Numér., 2016. To appear.
[11] B. Cockburn and G. Fu. Superconvergence by M-decompositions. Part
III: Construction of three-dimensional finite elements. Modél. Math. Anal.
Numér., 2016. To appear.

[12] B. Cockburn, G. Fu, and F.-J. Sayas. Superconvergence by M-


decompositions. Part I: General theory for HDG methods for diffusion.
Math. Comp., 2016. To appear.

[13] B. Cockburn, X. Fu, A. Hungria, L. Ji, and F.-J. Sayas. Stormer-Numerov


HDG methods for the acoustic wave equation. Submitted.

[14] B. Cockburn, J. Gopalakrishnan, and R. Lazarov. Unified hybridization of


discontinuous Galerkin, mixed and continuous Galerkin methods for second
order elliptic problems. SIAM J. Numer. Anal., 47:1319–1365, 2009.

[15] B. Cockburn, J. Gopalakrishnan, and F.-J. Sayas. A projection-based error


analysis of HDG methods. Math. Comp., 79:1351–1367, 2010.

[16] B. Cockburn, J. Guzmán, and H. Wang. Superconvergent discontinuous


Galerkin methods for second-order elliptic problems. Math. Comp., 78:1–
24, 2009.

[17] B. Cockburn, W. Qiu, and K. Shi. Conditions for superconvergence of


HDG methods for second-order eliptic problems. Math. Comp., 81:1327–
1353, 2012.

[18] B. Cockburn and V. Quenneville-Bélair. Uniform-in-time superconvergence


of HDG methods for the acoustic wave equation. Math. Comp., 83:65–85,
2014.

[19] B. Cockburn, W. Qui, and K. Shi. Conditions for superconvergence of HDG


methods on curvilinear elements for second-order eliptic problems. SIAM
J. Numer. Anal., 50:1417–1432, 2012.

[20] B. Cockburn and K. Shi. Superconvergent HDG methods for linear elas-
ticity with weakly symmetric stresses. IMA J. Numer. Anal., 33:747–770,
2013.

[21] J. Cui and W. Zhang. An analysis of HDG methods for the Helmholtz
equation. IMA J. Numer. Anal., 34(1):279–295, 2014.

[22] A. Ern and J. L. Guermond. Discontinuous Galerkin methods for


Friedrichs’ systems. i. general theory. SIAM J. Numer. Anal., 44:753–778,
2006.

19
[23] A. Ern and J.-L. Guermond. Discontinuous Galerkin methods for
Friedrichs’ systems. II. Second-order elliptic PDEs. SIAM J. Numer. Anal.,
44(6):2363–2388, 2006.
[24] A. Ern and J.-L. Guermond. Discontinuous Galerkin methods for
Friedrichs’ systems. III. Multifield theories with partial coercivity. SIAM
J. Numer. Anal., 46(2):776–804, 2008.

[25] X. Feng and Y. Xing. Absolutely stable local discontinuous Galerkin meth-
ods for the Helmholtz equation with large wave number. Math. Comp.,
82:1269–1296, 2013.

[26] G. Fu, B. Cockburn, and H. Stolarski. Analysis of an HDG method for


linear elasticity. Internat. J. Numer. Methods Engrg., 102(3-4):551–575,
2015.

[27] S. Gottlieb, C.-W. Shu, and E. Tadmor. Strong stability preserving high
order time discretization methods. SIAM Rev., 43:89–112, 2000.

[28] R. Griesmaier and P. Monk. Error analysis for a hybridizable discontinuous


Galerkin method for the Helmholtz equation. J. Sci. Comput., 49(3):291–
310, 2011.

[29] A. Hungria, D. Prada, and F.-J. Sayas. HDG methods for elastodynamics.
Submitted.

[30] L. Li, S. Lanteri, and R. Perrussel. A hybridizable discontinuous Galerkin


method for solving 3D time-harmonic Maxwell’s equations. In Numeri-
cal mathematics and advanced applications 2011, pages 119–128. Springer,
Heidelberg, 2013.
[31] L. Li, S. Lanteri, and R. Perrussel. A class of locally well-posed hybridiz-
able discontinuous Galerkin methods for the solution of time-harmonic
Maxwell’s equations. Comput. Phys. Commun., 192:23–31, 2015.

[32] N. C. Nguyen and J. Peraire. Hybridizable discontinuous Galerkin methods


for partial differential equations in continuum mechanics. J. Comput. Phys.,
231:5955–5988, 2012.

[33] N. C. Nguyen, J. Peraire, and B. Cockburn. An implicit high-order hy-


bridizable discontinuous Galerkin method for linear convection-diffusion
equations. J. Comput. Phys., 228:3232–3254, 2009.

[34] N. C. Nguyen, J. Peraire, and B. Cockburn. An implicit high-order


hybridizable discontinuous Galerkin method for nonlinear convection-
diffusion equations. J. Comput. Phys., 228:8841–8855, 2009.

[35] N. C. Nguyen, J. Peraire, and B. Cockburn. A hybridizable discontinuous


Galerkin method for Stokes flow. Comput. Methods Appl. Mech. Engrg.,
199:582–597, 2010.

20
[36] N. C. Nguyen, J. Peraire, and B. Cockburn. A hybridizable discontinuous
Galerkin method for the incompressible Navier-Stokes equations (AIAA
Paper 2010-362). In Proceedings of the 48th AIAA Aerospace Sciences
Meeting and Exhibit, Orlando, Florida, January 2010.

[37] N. C. Nguyen, J. Peraire, and B. Cockburn. High-order implicit hybridiz-


able discontinuous Galerkin methods for acoustics and elastodynamics. J.
Comput. Phys., 230:3695–3718, 2011.
[38] N. C. Nguyen, J. Peraire, and B. Cockburn. Hybridizable discontinuous
Galerkin methods. In J. Hesthaven and E. Ronquist, editors, Spectral and
High Order Methods for Partial Differential Equations, volume 76 of Lect.
Notes Comput. Sci. Engrg., pages 63–84, Berlin Heidelberg, 2011. Springer
Verlag.

[39] N. C. Nguyen, J. Peraire, and B. Cockburn. Hybridizable discontinuous


Galerkin methods for the time-harmonic Maxwell’s equations. J. Comput.
Phys., 230:7151–7175, 2011.

[40] N. C. Nguyen, J. Peraire, and B. Cockburn. An implicit high-order hy-


bridizable discontinuous Galerkin method for the incompressible Navier-
Stokes equations. J. Comput. Phys., 230:1147–1170, 2011.

[41] N. C. Nguyen, J. Peraire, and B. Cockburn. A class of embedded discon-


tinuous Galerkin methods for computational fluid dynamics. J. Comput.
Phys., 302:674–692, 2015.

[42] N. C. Nguyen, J. Peraire, F. Reitich, and B. Cockburn. A phase-based


hybridizable discontinuous Galerkin method for the numerical solution of
the Helmholtz equation. J. Comput. Phys., 290:318–335, 2015.

[43] J. Peraire, N. C. Nguyen, and B. Cockburn. A hybridizable discontinuous


Galerkin method for the compressible Euler and Navier-Stokes equations
(AIAA Paper 2010-363). In Proceedings of the 48th AIAA Aerospace Sci-
ences Meeting and Exhibit, Orlando, Florida, January 2010.

[44] P. A. Raviart and J. M. Thomas. A mixed finite element method for


second order elliptic problems. In I. Galligani and E. Magenes, editors,
Mathematical Aspects of Finite Element Method, Lecture Notes in Math.
606, pages 292–315. Springer-Verlag, New York, 1977.

[45] S.-C. Soon, B. Cockburn, and H.K. Stolarski. A hybridizable discontinuous


Galerkin method for linear elasticity. Internat. J. Numer. Methods Engrg.,
80(8):1058–1092, 2009.

[46] M. Stanglmeier, N.-C. Nguyen, J. Peraire, and B. Cockburn. An explicit


hybridizable discontinuous Galerkin method for the acoustic wave equation.
Comput. Methods Appl. Mech. Engrg., 300:748–769, 2016.

21