DG = FEM + flat elements

Václav Kučera, Jonathan Lambrechts, Nicolas Moës,

Jean-François Remacle, Jiřı́ Szotkowski, Antoine Quiriny
January 19, 2025

1 Nonlinear convection in 1D
We consider the following nonlinear first order hyperbolic equation on an interval
(a, b) ⊂ R:
ut + f (u)x = 0, on (a, b) × (0, T ). (1)
For simplicity, we consider zero boundary conditions or a compactly supported
initial condition. In the standard DG scheme we take a partition (mesh) of [a, b]
into intervals determined by the mesh nodes

a = x0 < x1 < . . . < xN +1 = b (2)

which determine the intervals Ii = [xi , xi+1 ], i = 0, . . . , N . We define the

discontinuous piecewise linear space

Sh = {v : v|Ii ∈ P 1 (Ii ), i = 0, . . . , N }, (3)

where P 1 (Ii ) is the space of linear functions on Ii . A function vh ∈ Sh has

two values at every interior node xi : vh (x± −
i ) and the jump [vh ]xi = vh (xi ) −
+ + −
vh (xi ). For consistency, we define [vh ]x0 = −vh (x0 ) and [vh ]xN +1 = vh (xN +1 ).
Moreover, when applying integration by parts on an interval Ii , we will use the
x
standard notation [vh ]xi+1
i = vh (x− +
i+1 ) − vh (xi ).
The standard DG formulation of (1) reads: Find uh : (0, T ) → Sh such that
Z b N Z N +1
fˆ(uh (x−
X X
(uh )t vh dx − f (uh )vh0 dx + +
i ), uh (xi ))[vh ]xi = 0, (4)
a i=0 Ii i=0

for all vh ∈ Sh . Here fˆ is a consistent and monotone numerical flux, e.g. the
Lax-Friedrichs numerical flux
1 1
fˆLF (u− , u+ ) = f (u− ) + f (u+ ) − α(u+ − u− ), max |f 0 (u)|.


α= (5)
2 2 u

1
 b
uh (xi ) b
uh (x+
i )
uh (yi ) b b
b
uh (x−
i ) b b
b

b b b b b b

yi xi yi+1 xi+1 xi xi+1

Figure 1: FEM solution with inserted auxiliary nodes yi in the mesh (left) and
limiting DG solution (right).

1.1 From FEM to DG

Our goal is to derive the DG scheme (4) from standard finite elements. To this
end, we enrich the partition (2) with nodes yi :

a = y0 < x0 < y1 < x1 < . . . < yN +1 < xN +1 = b, (6)

where we consider two types of intervals: Ii = [xi , yi+1 ] and Ii0 = [yi , xi ]. The
idea is that we let the intervals Ii0 eventually degenerate to the single point xi .
Specifically, we let h̄ → 0, where h̄ = |Ii0 | for all i = 0, . . . , N + 1.
On the mesh (6) we consider the continuous piecewise linear space

Vh = {v ∈ C[a, b] : v|Ii ∈ P 1 (Ii ), i = 0, . . . , N, v|Ii0 ∈ P 1 (Ii0 ), i = 0, . . . , N +1, v(a) = v(b) = 0}.

(7)
Our starting point is a modified version of (1):

ut + f (u)x − (d(x)ux )x = 0, (8)

where d(x) is an artificial diffusion parameter of the form

(
0, x ∈ Ii , i = 0, . . . , N,
d(x) = (9)
h̄D, x ∈ Ii0 , i = 0, . . . , N + 1,

with a constant D which will be determined appropriately later. We note that

the added diffusion is ‘active’ only on the degenerating elements Ii0 .
The FEM formulation of (8) reads: Find uh : (0, T ) → Vh such that
Z b Z b Z b
(uh )t vh dx − f (uh )vh0 dx + d(x)(uh )x vh0 dx = 0 (10)
a a a

for all vh ∈ Vh .
We now let the size h̄ of the intervals Ii0 go to zero. In the FEM formulation,
the piecewise linear solution is determined by the its values uh (yi ), uh (xi ), while
the DG solution is determined by its values uh (x− +
i ), uh (xi ). As we degenerate
0
the auxiliary elements Ii to the single point xi (e.g. by letting yi → xi ), we
identify the values

uh (yi ) = uh (x− +
i ), uh (xi ) = uh (xi ) (11)

2
in the limit, cf. Figure 1. Using this convention, we can identify any vh ∈ Vh
with a limiting vh ∈ Sh . It will also be important to note that for any vh ∈ Vh
−
[vh ]xyii = vh (xi ) − vh (yi ) −→ vh (x+
i ) − vh (xi ) = −[vh ]xi , (12)

as h̄ → 0. Using this convention, we can identify the limiting FEM scheme (10)
for h̄ → 0 with the following DG form.
Theorem 1. The finite element scheme (10) for h̄ = 0 is equivalent to the
formulation: Find uh : (0, T ) → Sh such that
N Z N Z N +1
fˆI (uh (x−
X X X
(uh )t vh dx− f (uh )vh0 dx+ +
i ), uh (xi ))[vh ]xi = 0, (13)
i=0 Ii i=0 Ii i=0

for all vh ∈ Sh , where fˆI is the numerical flux defined by

Z 1
ˆ − +
f (1 − τ )u− + τ u+ dτ − D(u+ − u− ).


fI (u , u ) = (14)
0

Remark 1. We note that the intervals Ii0 have ‘disappeared’ from the limiting
formulation, their only remnant being the emergence of a numerical flux fˆI .
Remark 2. The numerical flux fˆI , which naturally arises in the limiting scheme,
consists of an integral average of f over the values between u− and u+ along
with a diffusion term. If we approximate the integral average by trapezoidal
quadrature, we get the Lax-Friedrichs flux fˆLF from (5). We call the flux from
(14) the Integral Lax-Friedrichs numerical flux.
Remark 3. If we consider linear advection, i.e. f (u) = au for some a ∈ R, and
take D = |a|/2, then fˆI reduces to the standard upwind numerical flux :
Z 1  a |a|   a |a| 
ˆ − +
fI (u , u ) = a (1−τ )u− +τ u+ dτ −D(u+ −u− ) = + u− + − u+ ,
0 2 2 2 2
hence (
au− , a ≥ 0,
fˆI (u− , u+ ) = (15)
au+ , a < 0.

An essential property of numerical fluxes is consistency (i.e. fˆ(u, u) = f (u))

and monotonicity (fˆ is non-decreasing in its first argument and non-increasing
in its second argument), symbolically fˆ(↑, ↓).

Theorem 2. The integral Lax-Friedrichs numerical flux fˆI is consistent. More-

over let D ≥ 21 maxu |f 0 (u)|, then fˆI is monotone.
Proof. For consistency we have
Z 1 Z 1
ˆ



fI (u, u) = f (1 − τ )u + τ u dτ − D̃(u − u) = f u dτ = f (u).
0 0

3
Monotonicity follows from

∂ fˆI (u− , u+ )
Z 1
f 0 (1 − τ )u− + τ u+ (1 − τ ) dτ + D̃


−
=
∂u 0
Z 1
1
≥ D̃ − max |f 0 (u)| (1 − τ ) dτ = D̃ − max |f 0 (u)|,
u 0 2 u
1
which is nonnegative iff D̃ ≥ 2 maxu |f 0 (u)|. Similarly,

∂ fˆI (u− , u+ ) 1
Z
f 0 (1 − τ )u− + τ u+ τ dτ − D̃


=
∂u+ 0
1
1
Z
≤ max |f 0 (u)| τ dτ − D̃ = max |f 0 (u)| − D̃,
u 0 2 u
1
which is nonpositive iff D̃ ≥ 2 maxu |f 0 (u)|.

Remark 4. In analogy with the local Lax-Friedrichs flux, in the condition on D

in Theorem 1 we can take the maximum over the interval between u− and u+ .
Proof of Theorem 1. Let ϕi , i = 1, . . . , 2N , be the standard ‘tent’ basis func-
tions of Vh corresponding to the individual interior nodes xi , yi . We note that
there are no basis functions corresponding to y0 = a and xN +1 = b due to the
zero Dirichlet boundary
P2N conditions at these points. We write the solution of
(10) as uh (x, t) = j=1 Uj (t)ϕj (x) and set vh := ϕk for some k. Then (10)
becomes
2N
X Z b Z b 2N
X Z b
Uj0 ϕj ϕk dx − f (uh )ϕ0k dx + Uj d(x)ϕ0j ϕ0k dx = 0. (16)
j=1 a a j=1 a

Evolutionary term: Since |ϕj | ≤ 1 for all j, in the first term we have
Z
ϕj ϕk dx ≤ h̄ → 0 as h̄ → 0, (17)
Ii0

therefore, as h̄ → 0, we get
Z b N Z
X N
X +1 Z N Z
X
ϕj ϕk dx = ϕj ϕk dx + ϕj ϕk dx → ϕj ϕk dx. (18)
a i=0 Ii i=0 Ii0 i=0 Ii

Convective term: On an element Ii0 , we have ϕ0k = [ϕk ]xyii /h̄, hence
Z xi
1 1 xi
Z Z
− f (uh )ϕ0k dx = − f (uh ) [ϕk ]xyii dx = − f (uh ) dx [ϕk ]xyii
Ii0 yi h̄ h̄ yi
Z 1 (19)
f (1 − τ )uh (yi ) + τ uh (xi ) dτ [ϕk ]xyii ,


=−
0

4
where we have used the substitution x = (1 − τ )yi + τ xi in the integral, and the
fact that uh is a linear function on [yi , xi ]. Therefore, we get
Z b N Z
X N
X +1 Z
− f (uh )ϕ0k dx = − f (uh )ϕ0k dx + f (uh )ϕ0k dx
a i=0 Ii i=0 Ii0
(20)
N Z
X N
X +1 Z 1
f (uh )ϕ0k dx − f (1 − τ )uh (yi ) + τ uh (xi ) dτ [ϕk ]xyii .


=−
i=0 Ii i=0 0

Diffusion term: We have d(x) = h̄D on elements Ii0 and zero elsewhere. More-
over on these elements ϕ0j = [ϕj ]xyii /h̄ for all j. Therefore

1 1
Z Z
d(x)ϕ0j ϕ0k dx = h̄D [ϕj ]xyii [ϕk ]xyii dx = D[ϕj ]xyii [ϕk ]xyii , (21)
Ii0 Ii0 h̄ h̄

since |Ii0 | = h̄ and all the integrands are constant. Hence

Z b N
X +1 Z N
X +1
d(x)ϕ0j ϕ0k dx = h̄Dϕ0j ϕ0k dx = D̃[ϕj ]xyii [ϕk ]xyii . (22)
a i=0 Ii0 i=0

Altogether the FEM scheme (16) becomes

2N
X N Z
X 2N
X N
X +1 Z N Z
X
Uj0 ϕj ϕk dx + Uj0 ϕj ϕk dx − f (uh )ϕ0k dx
j=1 i=0 Ii j=1 i=0 Ii0 i=0 Ii
(23)
N
X +1  Z 1 
f (1 − τ )uh (yi ) + τ uh (xi ) dτ − D̃[uh ]xyii [ϕk ]xyii = 0.


−
i=0 0

In the limit h̄ → 0 the second term in (23) ‘disappears’ from the formulation
due to (18). Moreover, we use the aforementioned convention (11) in the limit.
Finally, if we use (12) in the limit h̄ → 0 the scheme (23) becomes (13).

References