Discontinuous Galerkin methods Lecture 3 - Brown University
Discontinuous Galerkin methods Lecture 3 - Brown University
Discontinuous Galerkin methods Lecture 3 - Brown University
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
RMMC 2008<br />
<strong>Discontinuous</strong> <strong>Galerkin</strong> <strong>methods</strong><br />
<strong>Lecture</strong> 3<br />
Jan S Hesthaven<br />
<strong>Brown</strong> <strong>University</strong><br />
-0.75<br />
Jan.Hesthaven@<strong>Brown</strong>.edu x<br />
y<br />
y<br />
1<br />
0.75<br />
0.5<br />
0.25<br />
0<br />
-0.25<br />
-0.5<br />
-1<br />
-1 -0.5 0 0.5 1<br />
1<br />
0.75<br />
0.5<br />
0.25<br />
0<br />
-0.25<br />
-0.5<br />
-0.75<br />
1.203<br />
1.142<br />
1.081<br />
1.019<br />
0.958<br />
0.896<br />
0.835<br />
0.774<br />
0.712<br />
0.651<br />
0.590<br />
0.528<br />
0.467<br />
0.405<br />
0.344<br />
0.283<br />
0.221<br />
0.160<br />
0.099<br />
0.037<br />
y<br />
1<br />
0.75<br />
0.5<br />
0.25<br />
0<br />
-0.25<br />
-0.5<br />
-0.75<br />
8.4 Scattering about a vertical cylinder in a finite-width channel -0.0881 147<br />
10<br />
0.75<br />
9<br />
test the consistency of the imposed bound-<br />
8<br />
7<br />
ary conditions and investigate 0.5<br />
6how<br />
the geo-<br />
0<br />
cal model based on the linear Padé (2,2) ro-<br />
0.02<br />
tational velocity version we set up a test-0.25 for<br />
0.01<br />
open-channel flow. We seek to model the scat-<br />
0<br />
-0.5<br />
tering of an incident wave field which is prop-<br />
-0.01<br />
y<br />
-0.02<br />
der positioned -0.03 in the middle of a finite-width<br />
-0.03 -0.02 -0.01 0 0.01 0.02 0.03<br />
x<br />
-1<br />
-1 -0.5 0 0.5 1<br />
x<br />
y<br />
-1<br />
-1 -0.5 0 0.5 1<br />
x<br />
8.4 Scattering about a vertical cylinder in a finitewidth<br />
channel<br />
A numerical study is carried out to both<br />
metric representation of the the domain may 0.25<br />
affect the computed solution. In a numeri-<br />
agating toward a bottom-mounted rigid cylin-<br />
channel.<br />
1<br />
-0.75<br />
Due to the symmetry, the solution is mathe-<br />
y<br />
0.02<br />
0.01<br />
0<br />
-0.01<br />
-0.02<br />
-0.03<br />
-0.0028<br />
-0.0072<br />
-0.0117<br />
-0.0162<br />
-0.0207<br />
-0.0252<br />
-0.0297<br />
-0.0342<br />
-0.0387<br />
-0.0432<br />
-0.0477<br />
-0.0522<br />
-0.0567<br />
-0.0612<br />
-0.0657<br />
-0.0702<br />
-0.0747<br />
-0.0791<br />
-0.0836<br />
1.95E-06<br />
1.70E-06<br />
1.45E-06<br />
1.20E-06<br />
9.45E-07<br />
6.94E-07<br />
4.43E-07<br />
1.92E-07<br />
-5.90E-08<br />
-3.10E-07<br />
-5.61E-07<br />
-8.12E-07<br />
-1.06E-06<br />
-1.31E-06<br />
-1.57E-06<br />
-0.03 -0.02 -0.01 0 0.01 0.02 0.03<br />
-1<br />
x<br />
-1 -0.5 0 0.5 1<br />
Figure 8.12: Scattering x of waves about a<br />
Darmstadt International Workshop, October 2004 – p.79<br />
cylinder in a finite-width channel.
A brief overview of what’s to come<br />
• <strong>Lecture</strong> 1: Introduction, Motivation, History<br />
• <strong>Lecture</strong> 2: Basic elements of DG-FEM<br />
• <strong>Lecture</strong> 3: Linear systems and some theory<br />
• <strong>Lecture</strong> 4: A bit more theory and discrete stability<br />
• <strong>Lecture</strong> 5: Attention to implementations<br />
• <strong>Lecture</strong> 6: Nonlinear problems and properties<br />
• <strong>Lecture</strong> 7: Problems with discontinuities and shocks<br />
• <strong>Lecture</strong> 8: Higher order/Global problems
<strong>Lecture</strong> 3<br />
• Let’s briefly recall what we know<br />
• Constructing fluxes for linear systems<br />
• The local basis and Legendre polynomials<br />
• Approximation theory on the interval
Let us recall<br />
We already know a lot about the basic DG-FEM<br />
• Stability is provided by carefully choosing the<br />
numerical flux.<br />
• Accuracy appear to be given by the local solution<br />
representation.<br />
• We can utilize major advances on monotone<br />
schemes to design fluxes.<br />
• The scheme generalizes with very few changes to<br />
very general problems -- multidimensional systems<br />
of conservation laws.
Let us recall<br />
We already know a lot about the basic DG-FEM<br />
• Stability is provided by carefully choosing the<br />
numerical flux.<br />
• Accuracy appear to be given by the local solution<br />
representation.<br />
• We can utilize major advances on monotone<br />
schemes to design fluxes.<br />
• The scheme generalizes with very few changes to<br />
very general problems -- multidimensional systems<br />
of conservation laws.<br />
At least in principle -- but what can we actually prove ?
2.4 Interlude on linear hyperbolic problems<br />
For But linear systems, first a the bit construction more of on thefluxes upwind numerical flux is particularly<br />
simple and we will discuss this in a bit more detail. Important application<br />
areas include Maxwell’s equations and the equations of acoustics and<br />
elasticity. Let us briefly look a little more carefully at linear<br />
To systems illustrate the basic approach, let us consider the two-dimensional<br />
system<br />
Q(x) ∂u<br />
+ ∇ ·F = Q(x)∂u<br />
∂t ∂t + ∂F 1<br />
∂x + ∂F linear systems, the construction of the upwind numerical flux is particrly<br />
simple and we will discuss this in a bit more detail. Important appliion<br />
areas include Maxwell’s equations and the equations of acoustics and<br />
sticity.<br />
To illustrate the basic approach, let us consider the two-dimensional<br />
tem<br />
Q(x)<br />
2<br />
=0, (2.19)<br />
∂y<br />
where the flux is assumed to be given as<br />
F =[F1, F 2] =[A1(x)u, A2(x)u] .<br />
Furthermore, we will make the natural assumption that Q(x) is invertible and<br />
symmetric for all x ∈ Ω. To formulate the numerical flux, we will need an<br />
∂u<br />
+ ∇ ·F = Q(x)∂u<br />
∂t ∂t + ∂F 1<br />
∂x + ∂F 2<br />
=0, (2.19)<br />
∂y<br />
ere the flux is assumed to be given as<br />
F =[F1, F 2] =[A1(x)u, A2(x)u] .<br />
thermore, Prominent we will make examples the natural areassumption<br />
that Q(x) is invertible and<br />
metric • Acoustics<br />
for all x ∈ Ω. To formulate the numerical flux, we will need an<br />
• Electromagnetics<br />
• Elasticity<br />
In such cases we can derive exact upwind fluxes
∂t ∂x ∂y<br />
assume approximation that Q(x) of ˆn ·F andutilizing Ai(x) information vary smoothly from both throughout sides of the Ω. interface. In<br />
anExactly rewrite Linear how Eq. this systems (2.19) information as and is combined fluxes should follow the dynamics of the<br />
equations.<br />
Let us first assume that Q(x) and Ai(x) vary smoothly throughout Ω. In<br />
Q(x) thisAssume case, we can first rewrite that Eq. all (2.19) coefficients as vary smoothly<br />
∂u<br />
rested in the formulation Q Q(x) of a flux+ along A1(x) the normal, + A2(x) ˆn, we+<br />
B(x)u<br />
operator<br />
∂t ∂x ∂y<br />
ll low-order terms; for example, it vanishes if Ai is constant.<br />
rested in Π<br />
where the = (ˆnxA1(x) formulation B collects all low-order terms; for example, it vanishe<br />
∂u + ˆnyA2(x)) of a flux<br />
∂u . along the normal, ˆn, we<br />
operator+<br />
Since A1(x) we are + interested A2(x) + inB(x)u the formulation =0, of a flux along<br />
∂t ∂x ∂y<br />
r that will consider the operator<br />
−1 Π = SΛS −1 ,<br />
diagonal matrix, Λ, has purely real entries; that is, Eq. (2.19) is a<br />
yperbolic system [142]. We express this as<br />
Q(x) ∂u ∂u ∂u<br />
s all low-order Π = (ˆnxA1(x) terms; ˆn ·F = for + A1(x) Πu. + example, ˆnyA2(x)) + A2(x) it vanishes . + B(x)u if Ai is =0, constant.<br />
∂t ∂x ∂y<br />
terested in the formulation of a flux Π along = (ˆnxA1(x) the normal, + ˆnyA2(x)) ˆn, we .<br />
the linear system can be understood by considering Q−1 ing to the elements of Λ that have positive and negative signs, re-<br />
Π.<br />
r that<br />
Λ = Λ + + Λ − ,<br />
The flux along a normal ˆn is then<br />
ewhere operator B collects all low-order terms; for example, it vanishes if Ai is constant.<br />
Since we Note are interested inˆn particular ·F in= the Πu. formulation that of a flux along the normal, ˆn, we<br />
will consider Π = the(ˆnxA1(x) operator + ˆnyA2(x)) . ˆn ·F = Πu.<br />
the linear system can be understood by considering Q<br />
Π = (ˆnxA1(x) + ˆnyA2(x)) .<br />
−1 at Q<br />
Π.<br />
−1Π can be diagonalized as<br />
Q −1 Π = SΛS −1 Thus, the nonzero elements of Λ<br />
,<br />
− correspond to those elements<br />
racteristic vector S−1u where the direction of propagation is ophe<br />
normal (i.e., they are incoming components). In contrast, those<br />
orresponding to Λ + 36 2 The key ideas<br />
reflect components propagating along ˆn (i.e.,<br />
Note in particular that<br />
ˆn ·F = Πu.<br />
The dynamics of the linear system can be understood by considering Q−1Π. Let us assume that Q−1Π can be diagonalized as<br />
Q −1 Π = SΛS −1 lar that<br />
ˆn ·F = Πu.<br />
of the linear system can be understood by considering Q<br />
,<br />
where the diagonal matrix, Λ, has purely real entries; that is, Eq. (2.19) is a<br />
−1Π. that Q−1Π can be diagonalized as<br />
Q −1 Π = SΛS −1 The dynamics of the linear system can be understood by c<br />
Let us assume that Q<br />
,<br />
onal matrix, Λ, has purely real entries; that is, Eq. (2.19) is a<br />
−1Π can be diagonalized as<br />
Q −1 Π = SΛS −1 at Q<br />
,<br />
where the diagonal matrix, Λ, has purely real entries; that<br />
strongly hyperbolic system [142]. We express this as<br />
−1Π can be diagonalized as<br />
Q −1 Π = SΛS −1 al matrix, Λ, has purely real entries; that is, Eq. (2.19) is a<br />
lic system [142]. We express this , as<br />
al matrix, Λ, Λ has = purely Λ real entries; that is, Eq. (2.19) is a<br />
ic system [142]. We express this as<br />
+ + Λ − ,<br />
the elements of Λ that have positive and negative signs, rethe<br />
nonzero elements of Λ− eaving through the boundary). The basic picture is illustrated in<br />
here we, for completeness, also illustrate a λ2 = 0 eigenvalue (i.e.,<br />
agating mode).<br />
his basic understanding, it is clear that a numerical upwind flux can<br />
d as<br />
(ˆn ·F) correspond to those elements<br />
∗ = QS Λ + S −1 u − + Λ − S −1 u + Now diagonalize this as<br />
u<br />
,<br />
–<br />
u * u **<br />
u +<br />
λ2 λ1 λ3 and we obtain<br />
Fig. 2.3. Sketch of the characteristic wave speeds of a t<br />
Λ = Λ + −<br />
+ + Λ − −1 boundary between , two states, u − and u + . The two intermed
know that<br />
with λ > 0; that is, the wave corresponding to λ1 is entering the domain, the<br />
λ1 = −λ, λ2 =0, λ3 = λ,<br />
Linear systems and fluxes<br />
∀i : −λiQ[u − − u + ] + [(Πu) − − (Πu) + ]=0<br />
wave corresponding to λ3 is leaving, and λ2 corresponds to a stationary wave<br />
must hold across each wave. This is also known as the Ra<br />
condition and is a simple consequence of conservation of u ac<br />
discontinuity. To appreciate this, consider the scalar wave eq<br />
∂u<br />
Consider the problem + λ∂u =0, x ∈ [a, b].<br />
∂t ∂x<br />
Integrating u over the interval, we have<br />
−<br />
with λ > 0; that is, the wave corresponding to λ1 is entering the domain, the<br />
wave corresponding to λ3 is leaving, and λ2 corresponds to a stationary wave<br />
as illustrated in Fig. 2.3.<br />
Following the well-developed theory or Riemann solvers [218, 303], we<br />
know that<br />
∀i : −λiQ[u<br />
λ<br />
− − u + ] + [(Πu) − − (Πu) + as illustrated in Fig. 2.3.<br />
Following the well-developed theory or Riemann solvers [218, 303], we<br />
know that<br />
∀i : −λiQ[u<br />
]=0, (2.20)<br />
must hold across each wave. This is also known as the Rankine-Hugoniot<br />
condition and is a simple consequence of conservation of u across the point of<br />
discontinuity. To appreciate this, consider the scalar wave equation<br />
− − u + ] + [(Πu) − − (Πu) + ]=0, (2.20)<br />
must hold across each wave. This is also known as the Rankine-Hugoniot<br />
condition and is a simple consequence of conservation of u across the point of<br />
discontinuity. To appreciate this, consider the scalar wave equation<br />
∂u<br />
+ λ∂u =0, x ∈ [a, b].<br />
∂t ∂x<br />
For non-smooth coefficients, it is a little more complex<br />
d<br />
dt<br />
Then we clearly have<br />
b<br />
Integrating over the interval, ∂u we have<br />
∂t a<br />
∂x<br />
b<br />
d<br />
Integrating over the interval, we have<br />
u dx = −λ (u(b, t) − u(a, t)) = f(a, t) − f(b<br />
a + λ∂u =0, x ∈ [a, b]. b<br />
since f u= dx λu. = −λ On(u(b, thet) other − u(a, hand, t)) = f(a, since t) −the f(b, wave t), is propagati<br />
dt a<br />
speed, λ, we also have<br />
b<br />
d<br />
dt<br />
u dx = d<br />
dt<br />
u +<br />
since f = λu. d On the u dx other = −λ hand,<br />
dt (u(b, t) since − u(a, thet)) wave = f(a, is propagating t) − f(b, t), at a constant<br />
speed, λ, we also b<br />
a have<br />
b<br />
(λt − a)u − +(b − λt)u + = λ(u −<br />
since f = λu. d On the other ahand,<br />
since the wave is propagating at a constant<br />
speed, λ, we alsou dt<br />
have dx = d − +<br />
(λt − a)u +(b− λt)u<br />
dt<br />
= λ(u − − u + ).<br />
Taking a → x− and b → x + a<br />
, we recover the jump conditions<br />
d b<br />
d − + Taking a → x − +<br />
− and b → x + , we recover the jump conditions
d<br />
dt<br />
b<br />
u dx = d<br />
(λt − a)u − +(b − λt)u + = λ(u − − u + ).<br />
Linear systems and fluxes<br />
a dt<br />
a →Hence, x by simple mass conservation, we achieve<br />
− and b → x + erive the proper numerical upwind flux for such cases, we need to<br />
careful. One should keep in mind that the much simpler local Laxhs<br />
flux also works in this case, albeit most likely leading to more<br />
ion than if the upwind flux , weis recover used. the jump conditions<br />
simplicity, we assume that we have only three entries in Λ, given as<br />
−λ(u − − u + )+(f − − f + )=0.<br />
for a → x− ,b→ x +<br />
λ1 = −λ, λ2 =0, λ3 = λ,<br />
0; that is, the wave corresponding to λ1 is entering the domain, the<br />
eralization to Eq. (2.20) is now straightforward.<br />
These are the Rankine-Hugoniot conditions<br />
rresponding to λ3 is leaving, and λ2 corresponds to a stationary wave<br />
rated in Fig. 2.3.<br />
wingFor the the well-developed general system, theory orthese Riemann are solvers [218, 303], we<br />
at<br />
36 2 The key ideas<br />
∀i : −λiQ[u − − u + ] + [(Πu) − − (Πu) + ]=0, (2.20)<br />
u * u **<br />
ld across each wave. This is also known as the Rankine-Hugoniot<br />
λ1 n and is a simple consequence of conservation of u across the point of<br />
nuity. They To appreciate must hold this, consider across the each scalar wave equation<br />
wave and can be used to<br />
connect across the interface<br />
∂u<br />
∂t<br />
+ λ∂u<br />
∂x<br />
=0, x ∈ [a, b].<br />
ing over the interval, we have<br />
Fig. 2.3. Sketch of the characteristic wave speeds of<br />
u –<br />
λ 2<br />
u +<br />
λ 3
Linear systems and fluxes<br />
2.4 Interlude on linear hyperbolic problems 37<br />
36 2 The key ideas<br />
So for the 3-wave problem we have<br />
turning to the problem in Fig. 2.3, we have the system of equations<br />
λQ − (u ∗ − u − )+ (Πu) ∗ − (Πu) − =0,<br />
[(Πu) ∗ − (Πu) ∗∗ ]=0,<br />
−λQ + (u ∗∗ − u + )+ (Πu) ∗∗ − (Πu) + =0,<br />
(u<br />
and the numerical flux is given as<br />
∗ , u∗∗ ) represents the intermediate states.<br />
e numerical flux can then be obtained by realizing that<br />
2.4 Interlude on linear hyperbolic problems 3<br />
eturning to the problem in Fig. 2.3, we have the system of equations<br />
λQ − (u ∗ − u − )+ (Πu) ∗ − (Πu) − =0,<br />
[(Πu) ∗ − (Πu) ∗∗ ]=0,<br />
−λQ + (u ∗∗ − u + )+ (Πu) ∗∗ − (Πu) + =0,<br />
e (u∗ , u∗∗ ) represents the intermediate states.<br />
he numerical flux can then be obtained by realizing that<br />
(ˆn ·F) ∗ =(Πu) ∗ =(Πu) ∗∗ ,<br />
λ 1<br />
u –<br />
λ 2<br />
u * u **<br />
Fig. 2.3. Sketch of the characteristic wave speeds of a<br />
boundary between two states, u − and u + . The two interme<br />
are used to derive the upwind flux. In this case, λ1 < 0, λ2<br />
(ˆn ·F) ∗ =(Πu) ∗ =(Πu) ∗∗ ,<br />
one can attempt to express using (u− , u + ) through the jump conditions<br />
. This leads to an upwind flux for the general discontinuous case.<br />
appreciate this approach, we consider a few examples.<br />
ple 2.5. Consider first the linear hyperbolic problem<br />
<br />
∂q<br />
∂ u a(x) 0 ∂ u<br />
+ A∂q = +<br />
=0,<br />
∂t ∂x ∂t v 0 −a(x) ∂x v<br />
To derive the proper numerical upwind flux for s<br />
be more careful. One should keep in mind that the m<br />
Friedrichs flux also works in this case, albeit most<br />
dissipation than if the upwind flux is used.<br />
For simplicity, we assume that we have only three<br />
one can attempt to express using (u− , u + ) through the jump condition<br />
e. This This leads approach to an upwind is general flux for and theyields general the discontinuous exact case.<br />
o appreciate this approach, we consider a few examples.<br />
upwind fluxes -- but requires that the system<br />
can be solved !<br />
ple 2.5. Consider first the linear hyperbolic problem<br />
∂q ∂q ∂ <br />
∂<br />
<br />
u +<br />
λ 3<br />
λ1 = −λ, λ2 =0, λ3 = λ,<br />
with λ > 0; that is, the wave corresponding to λ1 is en<br />
wave corresponding to λ3 is leaving, and λ2 correspond
which one can attempt to express using (u<br />
Linear systems and fluxes -- an example<br />
Consider<br />
− , u + ) through the jump conditions<br />
above. This leads to an upwind flux for the general discontinuous case.<br />
To appreciate this approach, we consider a few examples.<br />
Example 2.5. Consider first the linear hyperbolic problem<br />
<br />
∂q<br />
∂ u a(x) 0 ∂ u<br />
+ A∂q = +<br />
=0,<br />
∂t ∂x ∂t v 0 −a(x) ∂x v<br />
with a(x) being piecewise constant. For this simple equation, it is clear that<br />
u(x, t) propagates right while v(x, t) propagates left and we could use this to<br />
form a simple upwind flux.<br />
However, let us proceed using the Riemann jump conditions. If we introduce<br />
a ± as the values of a(x) on two sides of the interface, we recover the<br />
conditions<br />
a − (q ∗ − q − )+(Πq) ∗ − (Πq) − =0,<br />
−a + (q ∗ − q + )+(Πq) ∗ − (Πq) + =0,<br />
where q∗ refers to the intermediate state, (Πq) ∗ is the numerical flux along<br />
ˆn, and<br />
(Πq) ± = ˆn · (Aq) ± <br />
± a 0<br />
= ˆn ·<br />
0 −a ±<br />
<br />
± u<br />
v ±<br />
<br />
± ± a u<br />
= ˆn ·<br />
−a ± v ±<br />
Example 2.5. Consider first the linear hyperbolic problem<br />
<br />
∂q<br />
∂ u a(x) 0 ∂ u<br />
+ A∂q = +<br />
=0,<br />
∂t ∂x ∂t v 0 −a(x) ∂x v<br />
with a(x) being piecewise constant. For this simple equation, it is clear that<br />
u(x, t) propagates right while v(x, t) propagates left and we could use this to<br />
form a simple upwind flux.<br />
However, let us proceed using the Riemann jump conditions. If we introduce<br />
a<br />
<br />
.<br />
± as the values of a(x) on two sides of the interface, we recover the<br />
conditions<br />
a − (q ∗ − q − )+(Πq) ∗ − (Πq) − =0,<br />
−a + (q ∗ − q + )+(Πq) ∗ − (Πq) + =0,<br />
where q∗ refers to the intermediate state, (Πq) ∗ is the numerical flux along<br />
ˆn, and<br />
(Πq) ± = ˆn · (Aq) ± <br />
± a 0<br />
= ˆn ·<br />
0 −a ±<br />
<br />
± u<br />
v ±<br />
<br />
± ± a u<br />
= ˆn ·<br />
−a ± v ±<br />
<br />
.<br />
A bit of manipulation yields<br />
(Πq) ∗ 1<br />
=<br />
a + + a− ∂q<br />
∂ u a(x) 0 ∂ u<br />
+ A∂q = +<br />
∂t ∂x ∂t v 0 −a(x) ∂x v<br />
+ − − + + − − +<br />
a (Πq) + a (Πq) + a a (q − q ) ,<br />
which simplifies as<br />
=0,<br />
with a(x) being piecewise constant. For this simple equation, it is clear that<br />
u(x, t) propagates right while v(x, t) propagates left and we could use this to<br />
form a simple upwind flux.<br />
However, let us proceed using the Riemann jump conditions. If we introduce<br />
a ± as the values of a(x) on two sides of the interface, we recover the<br />
conditions<br />
a − (q ∗ − q − )+(Πq) ∗ − (Πq) − =0,<br />
−a + (q ∗ − q + )+(Πq) ∗ − (Πq) + =0,<br />
where q∗ refers to the intermediate state, (Πq) ∗ is the numerical flux along<br />
ˆn, and<br />
(Πq) ± = ˆn · (Aq) ± <br />
± a 0<br />
= ˆn ·<br />
0 −a ±<br />
<br />
± u<br />
v ±<br />
<br />
± ± a u<br />
= ˆn ·<br />
−a ± v ±<br />
<br />
.<br />
A bit of manipulation yields<br />
(Πq) ∗ 1<br />
=<br />
a + + a− + − − + + − − +<br />
a (Πq) + a (Πq) + a a (q − q ) ,<br />
which simplifies as<br />
(Πq) ∗ = 2a+ a− a + <br />
{{u}}<br />
ˆn ·<br />
+<br />
+ a− −{{v}}<br />
1<br />
propagates right while v(x, t) propagates left and we could use this to<br />
simple upwind flux.<br />
ever, let us proceed using the Riemann jump conditions. If we introas<br />
the values of a(x) on two sides of the interface, we recover the<br />
ns<br />
a<br />
<br />
[[u]]<br />
,<br />
2 [[v]]<br />
− (q ∗ − q − )+(Πq) ∗ − (Πq) − =0,<br />
−a + (q ∗ − q + )+(Πq) ∗ − (Πq) + =0,<br />
∗ ∗ refers to the intermediate state, (Πq) is the numerical flux along<br />
(Πq) ± = ˆn · (Aq) ± <br />
± a 0<br />
= ˆn ·<br />
0 −a ±<br />
<br />
± u<br />
v ±<br />
<br />
± ± a u<br />
= ˆn ·<br />
−a ± v ±<br />
<br />
.<br />
f manipulation yields<br />
(Πq) ∗ 1<br />
=<br />
a + + a− + − − + + − − +<br />
a (Πq) + a (Πq) + a a (q − q ) ,<br />
implifies as<br />
(Πq) ∗ = 2a+ a− a + <br />
{{u}}<br />
ˆn ·<br />
+<br />
+ a− −{{v}}<br />
1<br />
<br />
[[u]]<br />
,<br />
2 [[v]]<br />
numerical flux (Aq) ∗ follows directly from the definition of (Πq) ∗<br />
erve that if a(x) is smooth (i.e., a− = a + 38 2 The key ideas<br />
simply upwinding. Furthermore, for the general case, t<br />
to defining an intermediate wave speed, a<br />
) then the numerical flux is<br />
∗ , as<br />
a ∗ = 2a−a +<br />
a + Following the general approach, we have<br />
with<br />
Solving this yields Intermediate<br />
velocity<br />
,<br />
+ a− which is the harmonic average.
Let us also consider a slightly more complicated problem, originating<br />
Linear systems and fluxes -- an example<br />
electromagnetics.<br />
2.5 Exercises 39<br />
Example Consider 2.6. Consider Maxwell’s the equations<br />
one-dimensional Maxwell’s equations<br />
<br />
1<br />
ε(x) 0 ∂ E 01 ∂ E<br />
{{ZH}} + +<br />
=0. (2.2<br />
{{Z}}<br />
0 µ(x) ∂t H 10 ∂x H<br />
1<br />
2 [[E]]<br />
<br />
, E ∗ = 1<br />
<br />
{{YE}} +<br />
{{Y }}<br />
1<br />
2 [[H]]<br />
or<br />
H<br />
<br />
,<br />
∗ = 1<br />
<br />
{{ZH}} +<br />
{{Z}}<br />
1<br />
2 [[E]]<br />
<br />
, E ∗ = 1<br />
<br />
{{Y }}<br />
where<br />
Z ± <br />
µ ±<br />
=<br />
ε ± =(Y ± ) −1 ,<br />
2.5 Exercises 39<br />
Here (E, H) = (E(x, t),H(x, t)) represent the electric and magnetic fiel<br />
respectively, while ε and µ are the electric and magnetic material propertie<br />
known as permittivity and permeability, respectively.<br />
To simplify the notation, let us write this as<br />
Q ∂q<br />
+ A∂q<br />
∂t ∂x =0,<br />
or<br />
H<br />
where<br />
<br />
ε(x) 0<br />
01 E<br />
Q =<br />
, A = , q = ,<br />
0 µ(x) 10 H<br />
reflect the spatially varying material coefficients, the one-dimensional rotatio<br />
operator, and the vector of state variables, respectively.<br />
∗ = 1<br />
<br />
{{ZH}} +<br />
{{Z}}<br />
1<br />
2 [[E]]<br />
<br />
, E ∗ = 1<br />
<br />
{{YE}} +<br />
{{Y }}<br />
1<br />
2 [[H]]<br />
<br />
,<br />
where<br />
Z ± <br />
µ ±<br />
=<br />
ε ± =(Y ± ) −1 ,<br />
represents the impedance of the medium.<br />
If we again consider the simplest case of a continuous medium, things<br />
simplify considerably as<br />
H ∗ = {{H}} + Y<br />
2 [[E]], E∗ = {{E}} + Z<br />
2 [[H]],<br />
or<br />
H<br />
which we recognize as the Lax-Friedrichs flux since<br />
∗ = 1<br />
<br />
{{ZH}} +<br />
{{Z}}<br />
1<br />
2 [[E]]<br />
<br />
, E ∗ = 1<br />
<br />
{{Y }}<br />
where<br />
Z ± <br />
µ ±<br />
=<br />
ε ± =(Y ± ) −1 ,<br />
represents the impedance of the medium.<br />
If we again consider the simplest case of a cont<br />
simplify considerably as<br />
H ∗ = {{H}} + Y<br />
2 [[E]], E∗ Z<br />
= {{E}} +<br />
± <br />
µ ±<br />
=<br />
ε ± =(Y ± ) −1 ,<br />
the impedance of the medium.<br />
again consider the simplest case of a continuous medium, things<br />
onsiderably as<br />
H ∗ = {{H}} + Y<br />
2 [[E]], E∗ = {{E}} + Z<br />
2 [[H]],<br />
recognize as the Lax-Friedrichs flux since<br />
Y Z<br />
=<br />
ε µ = 1<br />
represents the impedance of the medium.<br />
If we again consider the simplest case of a cont<br />
simplify considerably as<br />
H<br />
√ = c<br />
εµ ∗ = {{H}} + Y<br />
2 [[E]], E∗ = {{E}} +<br />
which we recognize as the Lax-Friedrichs flux since<br />
Y Z<br />
=<br />
ε µ = 1<br />
The exact same approach leads to<br />
Now assume smooth materials:<br />
√ = c<br />
εµ<br />
We have recovered isthe the LF speed flux!<br />
of light (i.e., the fastest wave speed in th<br />
−1 −1/2
x ∈ D<br />
Lets move on<br />
At this point we have a good understanding of<br />
stability for linear problems -- through the flux.<br />
Lets now look at accuracy in more detail.<br />
k : u k Np <br />
h(x, t) = û<br />
n=1<br />
k Np <br />
n(t)ψn(x) = u<br />
i=1<br />
k h(x k i ,t)ℓ k i (x).<br />
Here, we have introduced two complementary expressions for the local solution<br />
known as the modal form, we use a local polynomial basis, ψn(x). A simple exa<br />
be ψn(x) =xn−1 . In the alternative form, known as the nodal representation, w<br />
N + 1 local grid points, xk i ∈ Dk , and express the polynomial through the associ<br />
Lagrange polynomial, ℓk i (x). The connection between these two forms is through<br />
the expansion coefficients, ûk n. We return to a discussion of these choices in much<br />
for now it suffices to assume that we have chosen one of these representations.<br />
The global solution u(x, t) is then assumed to be approximated by the piec<br />
polynomial approximation uh(x, t),<br />
K<br />
u(x, t) uh(x, t) = u<br />
k=1<br />
k y are d-dimensional simplexes.<br />
e local inner product and L<br />
h(x, t),<br />
2 (D k ) norm<br />
<br />
(u, v) k<br />
D =<br />
D k<br />
uv dx, u 2<br />
D k =(u, u) k<br />
D ,<br />
lobal broken inner product and norm<br />
K<br />
(u, v) Ω,h = (u, v) k<br />
D , u<br />
k=1<br />
2 Ω,h =(u, u) Ω,h .<br />
ects that Ω is only approximated by the union of D k , that is<br />
K<br />
Ω Ωh = D<br />
k=1<br />
k this slightly more general approach will pay off when we begin to consider<br />
more complex nonlinear and/or higher-dimensional problems.<br />
3.1 Legendre polynomials and nodal elements<br />
In Chapter 2 we started out by assuming that one can represent the global<br />
solution as a the direct sum of local piecewise polynomial solution as<br />
K<br />
u(x, t) uh(x, t) = u<br />
k=1<br />
Recall<br />
,<br />
ll not distinguish the two domains unless needed.<br />
we assume the local solution to be<br />
ne has local information as well as information from the neighalong<br />
an intersection between two elements. Often we will refer<br />
these intersections in an element as the trace of the element.<br />
s we discuss here, we will have two or more solutions or boundt<br />
the same physical location along the trace of the element.<br />
k h(x k ,t).<br />
The careful reader will note that this notation is a bit careless, as we do not<br />
address what exactly happens at the overlapping interfaces. However, a more<br />
careful definition does not add anything essential at this point and we will use<br />
this notation to reflect that the global solution is obtained by combining the<br />
K local solutions as defined by the scheme.<br />
The local solutions are assumed to be of the form<br />
x ∈ D k =[x k l ,x k r]: u k Np <br />
h(x, t) = û<br />
n=1<br />
k Np <br />
n(t)ψn(x) = u<br />
i=1<br />
k h(x k i ,t)ℓ k i (x).<br />
modal basis nodal basis
choosing this.<br />
rtant from a theoretical point of view. The results in Example 2.4 clearly<br />
rate, Local however, approximation<br />
that the accuracy of the method is closely linked to the<br />
r of the local polynomial representation and some care is warranted when<br />
sing this.<br />
et us To beginsimplify by introducing matters, the affine introduce mapping local affine mapping<br />
x ∈ D k : x(r) =x k l + 1+r<br />
2 hk , h k = x k r − x k l , (3.1)<br />
the reference variable r ∈ I = [−1, 1]. We consider local polynomial<br />
sentations of the form<br />
x ∈ D k : u k Np <br />
h(x(r),t)= û<br />
n=1<br />
k Np <br />
n(t)ψn(r) = u<br />
i=1<br />
k h(x k i ,t)ℓ k i (r).<br />
s first discuss the local modal expansion,<br />
Np <br />
uh(r) = ûnψn(r).<br />
n=1<br />
e we have dropped the superscripts for element k and the explicit time<br />
ndence, t, for clarity of notation.<br />
s a first choice, one could consider ψn(r) =rn−1 Let us begin by introducing the affine mapping<br />
x ∈ D<br />
r ∈ [−1, 1]<br />
Consider first the modal expansion<br />
(i.e., the simple monobasis).<br />
This leaves only the question of how to recover ûn. A natural way<br />
k : x(r) =x k l + 1+r<br />
2 hk , h k = x k r − x k l , (3.1)<br />
with the reference variable r ∈ I = [−1, 1]. We consider local polynomial<br />
representations of the form<br />
x ∈ D k : u k Np <br />
h(x(r),t)= û<br />
n=1<br />
k Np <br />
n(t)ψn(r) = u<br />
i=1<br />
k h(x k i ,t)ℓ k i (r).<br />
Let us first discuss the local modal expansion,<br />
Np <br />
uh(r) = ûnψn(r).<br />
n=1<br />
where we have dropped the superscripts for element k and the explicit time<br />
dependence, t, for clarity of notation.<br />
As a first choice, one could consider ψn(r) =rn−1 (i.e., the simple monomial<br />
basis). This leaves only the question of how to recover ûn. A natural way<br />
is by an L2 x ∈ D<br />
-projection; that is by requiring that<br />
Np <br />
(u(r), ψm(r)) I = ûn (ψn(r), ψm(r)) I ,<br />
n=1<br />
for each of the Np basis functions ψn. We have introduced the inner product<br />
on the interval I as 1<br />
k : u k Np <br />
h(x(r),t)= û<br />
n=1<br />
k Np <br />
n(t)ψn(r) = u<br />
i=1<br />
k h(x k i ,t)ℓ k i (r).<br />
et us first discuss the local modal expansion,<br />
Np <br />
uh(r) = ûnψn(r).<br />
n=1<br />
here we have dropped the superscripts for element k and the explicit time<br />
ependence, t, for clarity of notation.<br />
As a first choice, one could consider ψn(r) =rn−1 (i.e., the simple monoial<br />
basis). This leaves only the question of how to recover ûn. A natural way<br />
by an L2 Let us first discuss the local modal expansion,<br />
Np <br />
uh(r) = ûnψn(r).<br />
n=1<br />
where we have dropped the superscripts for element k<br />
dependence, t, for clarity of notation.<br />
As a first choice, one could consider ψn(r) =r<br />
-projection; that is by requiring that<br />
Np <br />
(u(r), ψm(r)) I = ûn (ψn(r), ψm(r)) I ,<br />
n=1<br />
r each of the Np basis functions ψn. We have introduced the inner product<br />
the interval I as<br />
1<br />
(u, v) I = uv dx.<br />
−1<br />
his yields<br />
n−1 (<br />
mial basis). This leaves only the question of how to reco<br />
is by an L2 Np <br />
uh(r) = ûnψn(r).<br />
n=1<br />
where we have dropped the superscripts for element k and the explicit time<br />
dependence, t, for clarity of notation.<br />
As a first choice, one could consider ψn(r) =r<br />
-projection; that is by requiring that<br />
Np <br />
(u(r), ψm(r)) I = ûn (ψn(r), ψm(r<br />
n=1<br />
for each of the Np basis functions ψn. We have introdu<br />
on the interval I as<br />
1<br />
(u, v) I = uv dx.<br />
−1<br />
This yields<br />
Mû = u,<br />
where<br />
leading to Np equations for the Np unknown expansion<br />
n−1 (i.e., the simple monomial<br />
basis). This leaves only the question of how to recover ûn. A natural way<br />
is by an L2-projection; that is by requiring that<br />
Np <br />
(u(r), ψm(r)) I = ûn (ψn(r), ψm(r)) I ,<br />
n=1<br />
for each of the Np basis functions ψn. We have introduced the inner product<br />
on the interval I as<br />
1<br />
(u, v) I = uv dx.<br />
−1<br />
This yields<br />
Mû = u,<br />
where<br />
Mij =(ψi, ψj) I , û = [û1, . . . , ûNp ]T We immediately have<br />
or<br />
, ui =(u, ψi) I ,<br />
leading to Np equations for the Np unknown expansion coefficients, ûi. How-<br />
Mij =(ψi, ψj) I , û = [û1, . . . , ûNp ]T , ui =
mial basis). This leaves only the question of how to recover ûn. A natural<br />
is by an L2-projection; that is by requiring that<br />
Local approximation<br />
(u(r), ψm(r)) I =<br />
Np <br />
Consider the most straightforward choice<br />
... and look at the condition number of M<br />
Table 3.1. Condition number of M based on the monomial basis for increasing<br />
value of N.<br />
(u, v) I = uv dx.<br />
N 2 4 8 16<br />
κ(M) 1.4e+1 3.6e+2 3.1e+5 3.0e+11<br />
This does not look like a robust choice for high N !<br />
(i + j − 1) −1 implies an increasing close to linear dependence of the basis as<br />
(i, j) increases, resulting in the ill-conditioning. The impact of this is that,<br />
even at moderately high order (e.g., N>10), one cannot accurately compute<br />
û and therefore not a good polynomial representation, uh.<br />
A natural solution to this problem is to seek an orthonormal basis as<br />
a more suitable and computationally stable approach. We can simply take<br />
the monomial basis, rn , and recover an orthonormal basis from this through<br />
an L2 But ever, why note? that - note that<br />
1 i+j<br />
Mij = 1+(−1)<br />
i + j − 1<br />
-based Gram-Schmidt orthogonalization approach. This results in the<br />
orthonormal basis [296]<br />
,<br />
Linear dependence for high N (i,j)<br />
n=1<br />
ûn (ψn(r), ψm(r)) I ,<br />
ψn(r) =rn−1 for each of the Np basis 3.1 Legendre functions polynomials ψn. Weand have nodal introduced elements the45 inner pro<br />
on the interval I as<br />
1<br />
This yields<br />
where<br />
−1<br />
Mû = u,<br />
Mij =(ψi, ψj) I , û = [û1, . . . , ûNp ]T , ui =(u, ψi) I ,<br />
leading to Np equations for the Np unknown expansion coefficients, ûi. H<br />
which resembles a Hilbert matrix, known to be very poorly conditioned.<br />
compute the condition number, κ(M), for M for increasing order of app<br />
imation, N, we observe in Table 3.1 the very rapidly deteriorating condi
− 1) Local approximation<br />
With this understanding, look for a better basis<br />
Idea - make M diagonal - an orthonormal basis<br />
−1 implies an increasing close to linear dependence of the basis as<br />
increases, resulting in the ill-conditioning. The impact of this is that,<br />
t moderately high order (e.g., N>10), one cannot accurately compute<br />
therefore not a good polynomial representation, uh.<br />
natural solution to this problem is to seek an orthonormal basis as<br />
e suitable and computationally stable approach. We can simply take<br />
onomial basis, rn , and recover an orthonormal basis from this through<br />
-based Gram-Schmidt orthogonalization approach. This results in the<br />
ormal basis [296]<br />
ψn(r) = ˜ Pn−1(r) = Pn−1(r)<br />
√ ,<br />
γn−1<br />
Pn(r) are the classic Legendre polynomials of order n and<br />
2<br />
γn =<br />
2n +1<br />
normalization. An easy way to compute this new basis is through the<br />
ence<br />
r ˜ Pn(r) =an ˜ Pn−1(r)+an+1 ˜ Pn+1(r),<br />
<br />
n<br />
an =<br />
2<br />
even at moderately high order (e.g., N>10), one cannot accurately<br />
û and therefore not a good polynomial representation, uh.<br />
A natural solution to this problem is to seek an orthonorma<br />
a more suitable and computationally stable approach. We can si<br />
the monomial basis, r<br />
,<br />
n , and recover an orthonormal basis from th<br />
an L2-based Gram-Schmidt orthogonalization approach. This resu<br />
orthonormal basis [296]<br />
ψn(r) = ˜ Pn−1(r) = Pn−1(r)<br />
√ ,<br />
γn−1<br />
where Pn(r) are the classic Legendre polynomials of order n and<br />
2<br />
γn =<br />
2n +1<br />
is the normalization. An easy way to compute this new basis is th<br />
recurrence<br />
r ˜ Pn(r) =an ˜ Pn−1(r)+an+1 ˜ Pn+1(r),<br />
<br />
n<br />
an =<br />
2<br />
(2n + 1)(2n − 1) ,<br />
with<br />
ψ1(r) = ˜ P0(r) = 1<br />
√ , ψ2(r) =<br />
2 ˜ <br />
3<br />
P1(r) =<br />
2 r.<br />
h<br />
A natural solution to this problem is to seek an orthonormal basis<br />
a more suitable and computationally stable approach. We can simply tak<br />
the monomial basis, r<br />
Appendix A discusses how to implement these recurrences and a<br />
n , and recover an orthonormal basis from this throug<br />
an L2-based Gram-Schmidt orthogonalization approach. This results in th<br />
orthonormal basis [296]<br />
ψn(r) = ˜ Pn−1(r) = Pn−1(r)<br />
√ ,<br />
γn−1<br />
where Pn(r) are the classic Legendre polynomials of order n and<br />
2<br />
γn =<br />
2n +1<br />
is the normalization. An easy way to compute this new basis is through th<br />
recurrence<br />
r ˜ Pn(r) =an ˜ Pn−1(r)+an+1 ˜ Pn+1(r),<br />
<br />
n<br />
an =<br />
2<br />
(2n + 1)(2n − 1) ,<br />
with<br />
ψ1(r) = ˜ P0(r) = 1<br />
√ , ψ2(r) =<br />
2 ˜ <br />
3<br />
P1(r) =<br />
2 r.<br />
even at moderately high order (e.g., N>10), one cannot accurately compute<br />
û and therefore not a good polynomial representation, uh.<br />
A natural solution to this problem is to seek an orthonormal basis as<br />
a more suitable and computationally stable approach. We can simply take<br />
the monomial basis, r<br />
Appendix A discusses how to implement these recurrences and a number<br />
n , and recover an orthonormal basis from this through<br />
an L2-based Gram-Schmidt orthogonalization approach. This results in the<br />
orthonormal basis [296]<br />
ψn(r) = ˜ Pn−1(r) = Pn−1(r)<br />
√ ,<br />
γn−1<br />
where Pn(r) are the classic Legendre polynomials of order n and<br />
2<br />
γn =<br />
2n +1<br />
is the normalization. An easy way to compute this new basis is through the<br />
recurrence<br />
r ˜ Pn(r) =an ˜ Pn−1(r)+an+1 ˜ Pn+1(r),<br />
<br />
n<br />
an =<br />
2<br />
(2n + 1)(2n − 1) ,<br />
with<br />
ψ1(r) = ˜ P0(r) = 1<br />
√ , ψ2(r) =<br />
2 ˜ <br />
3<br />
P1(r) =<br />
2 r.<br />
The Legendre Polynomials<br />
Evaluated as<br />
Appendix A discusses how to implement these recurrences and a number of<br />
˜
erty quadrature ˜Pn(r) of wesuch use can the quadratures library be computed routine is that<br />
46<br />
JacobiP.m using theyJacobiGQ.m, are<br />
3 Making<br />
asexact,<br />
provided as discussed u is in Appendix<br />
it work in one dimension<br />
fTo order compute 2Np −the 1 [90]. Np = The N +1 weights nodes and andnodes weights for (r, thew) Gaussian for an (2N +1)-th-or<br />
Local approximation<br />
obiP(r,0,0,n);<br />
s, the mass matrix, M, is the identity and the problem of<br />
resolved. So we It leaves could openevaluate the question these of how asto<br />
compute ûn,<br />
s<br />
ûn =(u, ψn) I =(u, ˜ For a general function u, the evaluation of this inner produ<br />
only practical way is to approximate the integral with a<br />
one can use a Gaussian quadrature of the form<br />
Np <br />
Pn−1) I .<br />
ûn u(ri) ˜ accurate be computed quadrature, using JacobiGQ.m, one executes as discussed the command in Appendix A.<br />
ature, >> one [r,w] executes =JacobiGQ(0,0,N);<br />
the command<br />
For JacobiGQ(0,0,N);<br />
the purpose of later generalizations, we consider a slightly different<br />
Pn−1(ri)wi,<br />
e Np = N +1 nodes and weights (r, w) for an (2N +1)-th-order<br />
proach and define ûn such that the approximation is interpolatory; tha<br />
e of later generalizations, we consider a slightly different i=1 ap-<br />
fine ûn such that the approximation<br />
<br />
is interpolatory; that is<br />
u(ξi) = ûn ˜ We instead require that<br />
Pn−1(ξi),<br />
u(ξi) =<br />
Np <br />
Np<br />
where ri are the quadrature points and wi are the quad<br />
important property of such quadratures is that they are e<br />
a polynomialn=1 of order 2Np − 1 [90]. The weights and nod<br />
quadrature canfor be computed some using ξi JacobiGQ.m, as discus<br />
To compute the Np = N +1 nodes and weights (r, w) for a<br />
accurate quadrature, one executes the command<br />
ûn<br />
n=1<br />
˜ Pn−1(ξi),<br />
where ξi represents a set of Np distinct grid points. Note that these need<br />
ents<br />
be associated<br />
a set of<br />
with<br />
Np distinct<br />
any<br />
grid<br />
quadrature<br />
points.<br />
points.<br />
Note that<br />
We<br />
these<br />
can now<br />
need<br />
write<br />
not<br />
ith any This quadrature yields points. We >> can [r,w] Vû now =JacobiGQ(0,0,N);<br />
= write u,<br />
where<br />
Vû = u,<br />
For the purpose of later generalizations, we consider a sl<br />
proach and define ûn such that the approximation is int<br />
Vij = ˜ Pj−1(ξi), ûi =ûi, ui = u(ξi).<br />
This matrix, V, is recognized as a generalized Vandermonde matrix and<br />
play a pivotal role in all of the subsequent developments of the scheme. T<br />
matrix establishes the connection between the modes, û, and the nodal val<br />
u(ξi) = ûn<br />
n=1<br />
˜ Vij =<br />
Pn−1(ξi),<br />
˜ Pj−1(ξi), ûi =ûi, ui = u(ξi).<br />
V is a Vandermonde matrix<br />
, is recognized as a generalized Vandermonde matrix and will<br />
or<br />
Np
imating polynomial of order N. Hence, the Lebesque constant indicates h<br />
o appreciate how this relates to the conditioning of V, recognize t<br />
Since we have already determined that<br />
Local approximation<br />
˜ Pn is the optimal<br />
Np with the need to choose the grid points ξi to define the Vand<br />
u(r) uh(r) = ûn There is significant freedom in this choice, so let us try to<br />
n=1<br />
intuition for a reasonable criteria.<br />
We first recognize that if<br />
˜ re · ∞ is the usual maximum norm and u<br />
Pn−1(r),<br />
olant (i.e., u(ξi) =uh(ξi) at the grid points, ξi), then we can write<br />
∗ far away the interpolation may be from the best<br />
represents<br />
possible<br />
the<br />
polynomial<br />
best appr<br />
rep<br />
tingsentation polynomial u of order N. Hence, the Lebesque constant indicates h<br />
away the interpolation may be from the best possible polynomial rep<br />
∗ nsequence of uniqueness . Note that Λofisthe determined polynomial solelyinterpolation, by the grid points, we have ξi. To<br />
an optimal approximation, we should therefore aim to identify those poin<br />
Recall first that we have<br />
Np <br />
u(r) uh(r) = ûn<br />
n=1<br />
˜ ation u<br />
Np <br />
u(r) uh(r) = u(ξi)ℓi(r).<br />
Pn−1(r),<br />
i=1<br />
∗ ξi, that . Note minimize thatthe Λ is Lebesque determined constant. solely by the grid points, ξi. To<br />
optimalTo approximation, appreciate how we thisshould relates to therefore the conditioning aim to identify of V, recognize those that poi<br />
hat aminimize consequence theof Lebesque uniqueness constant. of the polynomial interpolation, we have<br />
To appreciate how this relates to the conditioning of V, recognize that<br />
V<br />
nsequence of uniqueness of the polynomial interpolation, we have<br />
T ℓ(r) = ˜ V<br />
P (r),<br />
T ℓ(r) = ˜ P (r),<br />
re ℓ =[ℓ1(r),...,ℓNp (r)]T and ˜ P (r) =[ ˜ P0(r),..., ˜ PN (r)] T and<br />
. We are<br />
d in the particular solution, ℓ, which minimizes the Lebesque const<br />
ecall Cramer’s rule for solving linear systems of equations<br />
This immediately implies<br />
is an interpolant (i.e., u(ξi) =uh(ξi) at the grid points, ξi),<br />
it as<br />
Np <br />
u(r) uh(r) = u(ξi)ℓi(r).<br />
where ℓ =[ℓ1(r),...,ℓNp (r)]T and ˜ P (r) =[ ˜ P0(r),..., ˜ PN (r)] T . We are int<br />
ested in the particular solution, ℓ, which minimizes the Lebesque constant<br />
V i=1<br />
T ℓ(r) = ˜ P (r),<br />
ℓi(r) = Det[VT (:, 1), VT (:, 2), . . . , ˜ P (r), VT (:,i+ 1),...,VT (:,Np)]<br />
or<br />
re ℓ =[ℓ1(r),...,ℓNp (r)]T and ˜ P (r) =[ ˜ P0(r),..., ˜ PN (r)] T . We are in<br />
d in the particular solution, ℓ, which minimizes the Lebesque constant<br />
recall Cramer’s rule for solving linear systems of equations<br />
ℓi(r) = Det[VT (:, 1), VT (:, 2), . . . , ˜ P (r), VT (:,i+ 1),...,VT (:,Np)]<br />
Det(VT we recall Cramer’s rule for solving linear systems of equations<br />
ℓi(r) =<br />
.<br />
)<br />
Det[VT (:, 1), VT (:, 2), . . . , ˜ P (r), VT (:,i+ 1),...,VT (:,Np)]<br />
Det(VT Det(V<br />
.<br />
)<br />
It suggests that it is reasonable to seek ξi such that the denominator (i.e.,<br />
determinant of V), is maximized.<br />
T )<br />
ggests that it is reasonable to seek ξi such that the denominator (i.<br />
rminant of V), is maximized.<br />
or this<br />
so we<br />
one<br />
should<br />
dimensional<br />
choose<br />
case,<br />
the<br />
the<br />
point<br />
solution<br />
to maximize<br />
to this problem<br />
the<br />
is know<br />
tively simple form as the Np zeros of [151, 159]<br />
determinant -- the solution is<br />
For this one dimensional case, the solution to this problem is known i<br />
relatively simple form as the Np zeros of [151, 159]<br />
zeros of<br />
f(r) = (1 − r 2 ) ˜ P ′ N (r).<br />
uggests that it is reasonable to seek ξi such that the denominator (i.e.,<br />
2 ˜′
It suggests that it is reasonable to seek ξi such that the denomin<br />
determinant Local approximation<br />
of V), is maximized.<br />
For this one dimensional case, the solution to this problem<br />
relatively simple form as the Np zeros of [151, 159]<br />
This problem has an exact solution<br />
zeros of<br />
f(r) = (1 − r 2 ) ˜ P ′ N (r).<br />
also known as the Gauss-Lobatto quadrature points<br />
These are closely related to the normalized Legendre polynom<br />
46 3 Making it work in one dimension<br />
known as the Legendre-Gauss-Lobatto (LGL) quadrature poin<br />
library For a general routine function JacobiGL.m u, the evaluation in Appendix of this inner A, product theseisnodes nontrivial. canThe be c<br />
Note: These happen to also be quadrature points<br />
which we could use to evaluate<br />
only practical way is to approximate the integral with a sum. In particular,<br />
one can use a Gaussian quadrature of the form<br />
>> [r] = JacobiGL(0,0,N);<br />
ûn <br />
Np <br />
i=1<br />
u(ri) ˜ Pn−1(ri)wi,<br />
However, they were only chosen to be good for<br />
interpolation -- this is essential in 2D/3D<br />
where ri are the quadrature points and wi are the quadrature weights. An<br />
important property of such quadratures is that they are exact, provided u is<br />
a polynomial of order 2Np − 1 [90]. The weights and nodes for the Gaussian<br />
quadrature can be computed using JacobiGQ.m, as discussed in Appendix A.<br />
To compute the Np = N +1 nodes and weights (r, w) for an (2N +1)-th-order
determinant continues to increase in size with N, reflecting a well-behaved<br />
interpolation where ℓi(r) takes small values in between the nodes. For the<br />
Local approximation - an example<br />
equidistant nodes, the situation is entirely different.<br />
While the determinant has comparable values for small values of N, the<br />
determinant for the equidistant nodes begins to decay exponentially fast once<br />
N exceeds So does 16, caused it really by the matter? close to linear -- dependence consider of Vthe<br />
last columns<br />
a)<br />
c)<br />
1.00 −1.00 1.00 −1.00 1.00 −1.00 1.00<br />
1.00 −0.67 0.44 −0.30 0.20 −0.13 0.09<br />
1.00 −0.33 0.11 −0.04 0.01 −0.00 0.00<br />
1.00 0.00 0.00 0.00 0.00 0.00 0.00<br />
1.00 0.33 0.11 0.04 0.01 0.00 0.00<br />
1.00 0.67 0.44 0.30 0.20 0.13 0.09<br />
1.00 1.00 1.00 1.00 1.00 1.00 1.00<br />
0.71 −1.22 1.58 −1.87 2.12 −2.35 2.55<br />
0.71 −0.82 0.26 0.49 −0.91 0.72 −0.04<br />
0.71 −0.41 −0.53 0.76 0.03 −0.78 0.49<br />
0.71 0.00 −0.79 0.00 0.80 0.00 −0.80<br />
0.71 0.41 −0.53 −0.76 0.03 0.78 0.49<br />
0.71 0.82 0.26 −0.49 −0.91 −0.72 −0.04<br />
0.71 1.22 1.58 1.87 2.12 2.35 2.55<br />
b)<br />
d)<br />
1.00 −1.00 1.00 −1.00 1.00 −1.00 1.00<br />
1.00 −0.83 0.69 −0.57 0.48 −0.39 0.33<br />
1.00 −0.47 0.22 −0.10 0.05 −0.02 0.01<br />
1.00 0.00 0.00 0.00 0.00 0.00 0.00<br />
1.00 0.47 0.22 0.10 0.05 0.02 0.01<br />
1.00 0.83 0.69 0.57 0.48 0.39 0.33<br />
1.00 1.00 1.00 1.00 1.00 1.00 1.00<br />
0.71 −1.22 1.58 −1.87 2.12 −2.35 2.55<br />
0.71 −1.02 0.84 −0.35 −0.28 0.81 −1.06<br />
0.71 −0.57 −0.27 0.83 −0.50 −0.37 0.85<br />
0.71 0.00 −0.79 0.00 0.80 0.00 −0.80<br />
0.71 0.57 −0.27 −0.83 −0.50 0.37 0.85<br />
0.71 1.02 0.84 0.35 −0.29 −0.81 −1.06<br />
0.71 1.22 1.58 1.87 2.12 2.35 2.55<br />
Bad points Good points<br />
Fig. 3.1. Entries of V for N = 6 and different choices of the basis, ψn(r), and<br />
evaluation points, ξi. For (a) and (c) we use equidistant points and (b) and (d) are<br />
based on LGL points. Furthermore, (a) and (b) are based on V computed using a<br />
simple monomial basis, ψn(r) =r n−1 , while (c) and (d) are based on the orthonormal<br />
Bad<br />
basis<br />
Good<br />
basis
Local approximation - an example<br />
Determinant of V<br />
10 25<br />
10 20<br />
10 15<br />
10 10<br />
10 5<br />
10 0<br />
3.1 Legendre polynomials and nodal elements 49<br />
Equidistant nodes<br />
Legendre Gauss<br />
Lobatto nodes<br />
2 6 10 14 18<br />
N<br />
22 26 30 34<br />
50 3 Making it work in one dimension<br />
Fig. 3.2. Growth of the determinant of the Vandermonde matrix, V, when computed<br />
based on LGL nodes and the simple equidistant nodes.<br />
It is the basic<br />
in V. For the interpolation, this manifests itself as the well-known Runge<br />
phenomenon [90] for interpolation at equidistant grids. The rapid decay of<br />
the determinant enables an exponentially fast growth of ℓi(r) between the<br />
grid points, resulting in a very poorly behaved interpolation. This is also<br />
reflected in the corresponding Lebesque constant, which grows like 2Np 7<br />
6<br />
5<br />
4<br />
structure that 3<br />
for<br />
the equidistant nodes while it grows like2 log Np in the optimal case of the<br />
LGLreally nodes (see [151, matters<br />
307] and references 1therein).<br />
The main lesson here is that it is generally 0 not a good idea to use equidistant<br />
grid points when representing the local −1 interpolating polynomial solution.<br />
This point is particularly important when the interest is in the development<br />
of <strong>methods</strong> performing robustly at higher values of N.<br />
ξi One can ask, however, whether there is anything special about the LGL<br />
α<br />
10<br />
9<br />
8<br />
−1.25 −1 −0.75 −0.5 −0.25 0 0.25<br />
It matters a great deal<br />
Determinant of V<br />
10 25<br />
10 20<br />
10 15<br />
10 10<br />
10 5<br />
10 0<br />
N=4<br />
N=8<br />
N=16<br />
N=24<br />
N=32<br />
0 2 4 6 8 10<br />
α
where To summarize ℓ =[ℓ1(r),...,ℓNp matters, we have local approximations of the form<br />
(r)]T and ˜ P (r) =[ ˜ P0(r),..., ˜ PN (r)] T . W<br />
ested in the particular solution, ℓ, which minimizes the Lebesque<br />
This highlights that it is the overall structure of the nodes rather than<br />
the details of the individual node position that is important; for example, one<br />
could optimize these nodal sets Npfor<br />
various applications. Np<br />
Lets summarize this<br />
<br />
u(r) uh(r) = ûn ˜ we recall Cramer’s rule for solving linear systems of equations<br />
Pn−1(r) = u(ri)ℓi(r), (3.3<br />
So we have the local approximations<br />
n=1<br />
To summarize matters, we have local approximations i=1 of the form<br />
ℓi(r) = Det[VT (:, 1), VT (:, 2), . . . , ˜ P (r), VT (:,i+ 1),...,VT (:<br />
Det(VT )<br />
where ξi = ri are the Legendre-Gauss-Lobatto quadrature points. A centr<br />
Np <br />
u(r) uh(r) = ûn ˜ Np <br />
component of this construction is thePn−1(r) Vandermonde = u(ri)ℓi(r), matrix, V, which (3.3) estab<br />
lishes the connections<br />
n=1<br />
i=1<br />
It suggests that it is reasonable to seek ξi such that the denomina<br />
where ξi = ri are the Legendre-Gauss-Lobatto quadrature points. A central<br />
component of this u = construction Vû, V is the Vandermonde matrix, V, which establishes<br />
the connections<br />
T ℓ(r) = ˜ P (r), Vij = ˜ determinant of V), is maximized.<br />
Pj(ri).<br />
For this one dimensional case, the solution to this problem is<br />
relatively simple form as the Np zeros of [151, 159]<br />
and are the Legendre Gauss Lobatto points:<br />
u = Vû, V T ℓ(r) = ˜ P (r), Vij = ˜ By carefully choosing the orthonormal Legendre basis,<br />
Pj(ri).<br />
˜ ri<br />
Pn(r), and the nod<br />
points, ri, we have ensured that V is a well-conditioned object and that th<br />
resulting interpolation zeros of is well behaved. A script for initializing V is given i<br />
Vandermonde1D.m.<br />
By carefully choosing the orthonormal Legendre basis, ˜ f(r) = (1 − r Pn(r), and the nodal<br />
points, ri, we have ensured that V is a well-conditioned object and that the<br />
resulting interpolation is well behaved. A script for initializing V is given in<br />
Vandermonde1D.m.<br />
2 ) ˜ P ′ N (r).<br />
These This areleads closely to a related robust toway theof normalized computing/evaluating<br />
Legendre polynomi<br />
known a high-order as the Legendre-Gauss-Lobatto polynomial approximation. (LGL) quadrature points<br />
library routine JacobiGL.m in Appendix A, these nodes can be co<br />
but is it accurate ?<br />
>> [r] = JacobiGL(0,0,N);
e notation A second look at approximation<br />
k on a more rigorous discussion, we need to introduce some<br />
ion. InWe particular, will need both global a little norms, more defined notation on Ω, and broed<br />
over Ωh as sums of K elements D<br />
Regular energy norms<br />
k , need to be introduced.<br />
ion, u(x,t), we define the global L2-norm over the domain<br />
u 2 <br />
Ω = u<br />
Ω<br />
2 dx<br />
oken norms<br />
u 2 K<br />
Ω,h = u<br />
k=1<br />
2<br />
<br />
k<br />
D , u2 k<br />
D =<br />
D k<br />
u 2 dx.<br />
we define the associated Sobolev norms<br />
u (α) 2 Ω, u 2 K<br />
Ω,q,h = u 2<br />
D k , u2<br />
,q D k ,q =<br />
q<br />
u (α) 2<br />
additional notation. In particular, both global norms, defined on Ω, a<br />
ken norms, defined over Ωh as sums of K elements D k , need to be intr<br />
For the solution, u(x,t), we define the global L2-norm over the<br />
Ω ∈ R d as<br />
u 2 <br />
Ω = u<br />
Ω<br />
2 dx<br />
as well as the broken norms<br />
u 2 K<br />
Ω,h = u<br />
k=1<br />
2<br />
<br />
k<br />
D , u2 k<br />
D =<br />
D k<br />
u 2 dx.<br />
In a similar way, we define the associated Sobolev norms<br />
u 2 q<br />
Ω,q = u<br />
|α|=0<br />
(α) 2 Ω, u 2 K<br />
Ω,q,h = u<br />
k=1<br />
2<br />
D k , u2<br />
,q D k ,q =<br />
q<br />
u<br />
|α|=0<br />
where α is a multi-index of length d. For the one-dimensional cas<br />
often considered in this chapter, this norm is simply the L2 Before we embark on a more rigorous discussion, we need to introduce some<br />
additional notation. In particular, both global norms, defined on Ω, and broken<br />
norms, defined over Ωh as sums of K elements D<br />
-norm of<br />
k , need to be introduced.<br />
For the solution, u(x,t), we define the global L2-norm over the domain<br />
Ω ∈ R d as<br />
u 2 <br />
Ω = u<br />
Ω<br />
2 dx<br />
as well as the broken norms<br />
u 2 K<br />
Ω,h = u<br />
k=1<br />
2<br />
<br />
k<br />
D , u2 k<br />
D =<br />
D k<br />
u 2 dx.<br />
In a similar way, we define the associated Sobolev norms<br />
u 2 q<br />
Ω,q = u<br />
|α|=0<br />
(α) 2 Ω, u 2 K<br />
Ω,q,h = u<br />
k=1<br />
2<br />
D k , u2<br />
,q D k ,q =<br />
q<br />
u<br />
|α|=0<br />
(α) 2<br />
D k,<br />
where α is a multi-index of length d. For the one-dimensional case, most<br />
often considered in this chapter, this norm is simply the L2 Sobolev norms<br />
76 4 Insight through theory<br />
derivative. We will also define the space of functions, u ∈ H<br />
Semi-norms<br />
-norm of the q-th<br />
q (Ω), as those<br />
functions for which uΩ,q or uΩ,q,h is bounded.<br />
Finally, we will need the semi-norms<br />
k=1<br />
|u| 2 Ω,q,h =<br />
|α|=0<br />
D k ,q<br />
D k,<br />
u (α) 2<br />
D k.<br />
ulti-index of length d. For the one-dimensional case, most<br />
in this chapter, this norm is simply the L2-norm of the q-th<br />
4.2 Briefly on convergence<br />
K<br />
k=1<br />
|u| 2<br />
D k ,q<br />
, |u|2<br />
= <br />
|α|=q
Lagrange polynomial, ℓ<br />
Approximation theory<br />
k i (x). The connection between these two forms is through<br />
the expansion coefficients, ûk n. We return to a discussion of these choices in much m<br />
for now it suffices to assume that we have chosen one of these representations.<br />
The global solution u(x, t) is then assumed to be approximated by the piecew<br />
polynomial approximation uh(x, t),<br />
K<br />
u(x, t) uh(x, t) = u<br />
k=1<br />
k K<br />
, v) Ω,h = (u, v) k<br />
D , u<br />
k=1<br />
h(x, t),<br />
2 Ω,h =(u, u) Ω,h .<br />
cts that Ω is only approximated by the union of D k , that is<br />
K<br />
Ω Ωh = D<br />
k=1<br />
k K<br />
u(x, t) uh(x, t) = u<br />
k=1<br />
Recall<br />
,<br />
not distinguish we assume the twothe domains local unless solution needed. to be<br />
has local information as well as information from the neighong<br />
an intersection between two elements. Often we will refer<br />
ese intersections in an element as the trace of the element.<br />
e discuss here, we will have two or more solutions or boundthe<br />
same physical location along the trace of the element.<br />
k h(x k ,t).<br />
The careful reader will note that this notation is a bit careless, as we do not<br />
address what exactly happens at the overlapping interfaces. However, a more<br />
careful definition does not add anything essential at this point and we will use<br />
this notation to reflect that the global solution is obtained by combining the<br />
K local solutions as defined by the scheme.<br />
The local solutions are assumed to be of the form<br />
x ∈ D k =[x k l ,x k r]: u k Np <br />
h(x, t) = û<br />
n=1<br />
k Np <br />
n(t)ψn(x) = u<br />
i=1<br />
k h(x k i ,t)ℓ k Here, we have introduced two complementary expressions<br />
known as the modal form, we use a local polynomial basis<br />
be ψn(x) =x<br />
i (x).<br />
n−1 . In the alternative form, known as the n<br />
N + 1 local grid points, xk i ∈ Dk , and express the polynom<br />
Lagrange polynomial, ℓk i (x). The connection between thes<br />
the expansion coefficients, ûk n. We return to a discussion of<br />
for now it suffices to assume that we have chosen one of th<br />
The global solution u(x, t) is then assumed to be app<br />
polynomial approximation uh(x, t),<br />
K<br />
The question is in what sense is<br />
u(x, t) uh(x, t) =<br />
We have observed improved accuracy in two ways<br />
• Increase K/decrease h<br />
• Increase N<br />
k=1<br />
u
simplicity, we assume that all elements have length h (i.e., D k = x k + r<br />
˜Pn(r) = Pn(r)<br />
√ , γn =<br />
γn<br />
simplicity, we assume that all elements have length h (i.e., D<br />
Approximation theory<br />
k = xk +<br />
where xk = 1<br />
2 (xkr +xk where x<br />
l ) represents the cell center and r ∈ [−1, 1] is the refere<br />
k = 1<br />
2 (xk r +xk l ) represents the cell center and r ∈ vh(r) [−1, 1] = is the reference ˆvn<br />
coordinate).<br />
n=0<br />
˜ Pn(r),<br />
coordinate). We begin by considering the standard interval and introduce the new vari-<br />
We able Let begin us assume by considering all elements the standard have interval size h and introduce consider the new va<br />
able<br />
v(r) =u(hr) =u(x);<br />
that is, v is defined on thev(r) standard =u(hr) interval, =u(x); I =[−1, 1], and xk represents v ∈ L<br />
= 0, x ∈<br />
2 (I). Note, that for simplicity of the n<br />
with standard notation, we now have the sum runn<br />
than from 1 to N + 1 = Np, as used previously.<br />
d<br />
r<br />
[−h, h]. We discussed in Chapter 3 the advantage of using a local orthonormal<br />
(1 − r2 ) d<br />
dr ˜ Pn + n(n + 1) ˜ Pn =0. (4.2)<br />
Prior to discussing interpolation, used throughou<br />
sic question of how well the properties of the projection where we utilize the<br />
that is, v is defined on the standard interval, I =[−1, 1], and xk basis – in this case, the normalized Legendre polynomials<br />
= 0, x<br />
[−h, h]. We discussed in Chapter 3 the advantage of using a local orthonorm<br />
basis – in this case, the normalized ˜Pn(r) = Legendre polynomials<br />
Pn(r)<br />
to find ˜vn as<br />
N<br />
<br />
2<br />
√ γn<br />
, γn =<br />
˜Pn(r) = Pn(r)<br />
√ , γn =<br />
γn<br />
2<br />
2n +1 .<br />
Here, Pn(r) are the classic Legendre polynomials of order n. A key prope<br />
dr<br />
of these polynomials is that they satisfy a singular Sturm-Liouville proble<br />
˜ Pn + n(n + 1) ˜ Pn interpolation, used throughout this text, we consider =0. (4.2)<br />
d<br />
dr (1 − r2 ) d<br />
dr ˜ Pn + n(n + 1) ˜ N Pn =0. (4<br />
n=0<br />
Let us consider the basic question of how well<br />
2n +1 .<br />
Here, Pn(r) are the classic Legendre polynomials of order n. A key property<br />
of these polynomials is that they satisfy a singular Sturm-Liouville problem<br />
d<br />
dr (1 − r2 ) d<br />
Let us consider the basic question of how well<br />
vh(r) = ˆvn ˜ ote, that for simplicity of the notation and to conform<br />
n, we now have the sum running from 0 to N rather<br />
= Np, as used previously.<br />
projection where we utilize the orthonomality of<br />
Pn(r),<br />
˜ Pn(r)<br />
<br />
˜vn = v(r)<br />
I<br />
˜ Pn(r) dr.<br />
2<br />
2<br />
2n +1 .<br />
assic Legendre polynomials of order n. A key property<br />
s that they satisfy a singular Sturm-Liouville problem<br />
We consider expansions as<br />
vh(r) =<br />
n=0<br />
ˆvn ˜ Pn(r),<br />
˜vn =<br />
I<br />
N<br />
2 h<br />
v(r) ˜ Pn(r) dr.
simplicity, we assume that all elements have length h (i.e., D k = x k + r<br />
˜Pn(r) = Pn(r)<br />
√ , γn =<br />
γn<br />
2<br />
2n +1 .<br />
simplicity, we assume that all elements have length h (i.e., D<br />
Approximation theory<br />
k = xk +<br />
where xk = 1<br />
2 (xkr +xk where x<br />
l ) represents the cell center and r ∈ [−1, 1] is the refere<br />
k = 1<br />
2 (xk r +xk l ) represents the cell center and r ∈ vh(r) [−1, 1] = is the reference ˆvn<br />
coordinate).<br />
n=0<br />
˜ Pn(r),<br />
assic Legendre polynomials of order n. A key property<br />
s that they satisfy a singular Sturm-Liouville problem<br />
coordinate). We begin by considering the standard interval and introduce the new vari-<br />
We able Let begin us assume by considering all elements the standard have interval size h and introduce consider the new va<br />
able<br />
v(r) =u(hr) =u(x);<br />
that is, v is defined on thev(r) standard =u(hr) interval, =u(x); I =[−1, 1], and xk represents v ∈ L<br />
4.3 Approximations by orthogonal polynomials and consistency 79<br />
= 0, x ∈<br />
2 (I). Note, that for simplicity of the n<br />
with standard notation, we now have the sum runn<br />
than from 1 to N + 1 = Np, as used previously.<br />
d<br />
r<br />
[−h, h]. We discussed in Chapter 3 the advantage of using a local orthonormal<br />
(1 − r2 ) d<br />
dr ˜ Pn + n(n + 1) ˜ Pn =0. (4.2)<br />
An immediate consequence of<br />
Prior<br />
the orthonormality<br />
to discussing<br />
of<br />
interpolation,<br />
the basis is that<br />
used throughou<br />
sic question of how well the properties∞ of the projection where we utilize the<br />
We consider expansions as<br />
that is, v is defined on the standard interval, I =[−1, 1], and xk basis – in this case, the normalized Legendre polynomials<br />
= 0, x<br />
[−h, h]. We discussed in Chapter 3 the advantage of using a local orthonorm<br />
basis – in this case, the normalized ˜Pn(r) = Legendre polynomials<br />
Pn(r)<br />
v − vh<br />
N<br />
2<br />
2<br />
I = |˜vn|<br />
n=N+1<br />
2 to find ,<br />
˜vn as<br />
<br />
˜Pn(r) = Pn(r) 2<br />
√ , γn =<br />
γn 2n +1 .<br />
√ , γn =<br />
γn 2n +1 .<br />
Here, Pn(r) are the classic Legendre polynomials of order n. A key property<br />
of these polynomials is that they satisfy a singular Sturm-Liouville problem<br />
d<br />
dr (1 − r2 ) d<br />
Let us consider the basic question of how well<br />
vh(r) = ˆvn ˜ vh(r) = ˆvn<br />
n=0<br />
Pn(r),<br />
˜ Pn(r),<br />
ote, that for simplicity of the notation and to conform<br />
n, we now have the sum running from 0 to N rather<br />
= Np, as used previously.<br />
projection where we utilize the orthonomality of ˜ Pn(r)<br />
<br />
˜vn = v(r)<br />
I<br />
˜ recognized as Parseval’s identity. A basic result for this approximation error<br />
follows directly.<br />
Theorem 4.1. Assume that v ∈ H<br />
Pn(r) dr.<br />
p ˜vn = v(r)<br />
I<br />
(I) and that vh represents a polynomial<br />
projection of order N. Then<br />
where<br />
<br />
3q,<br />
0 ≤ q ≤ 1<br />
and 0 ≤ q ≤ p.<br />
˜ Pn(r) dr.<br />
Here, Pn(r) are the classic Legendre polynomials of order n. A key prope<br />
dr<br />
of these polynomials is that they satisfy a singular Sturm-Liouville proble<br />
˜ Pn + n(n + 1) ˜ Pn interpolation, used throughout v −this vhtext, I,q ≤ we N consider =0. (4.2)<br />
ρ−p |v| I,p ,<br />
d<br />
dr (1 − r2 ) d<br />
dr ˜ Pn + n(n + 1) ˜ ρ = 2<br />
2q − N Pn =0. (4<br />
1<br />
2 , q ≥ 1<br />
n=0<br />
Let us consider the basic question of how well<br />
Proof. We will just sketch the proof; the details can be found in [43]. Com-<br />
bining the definition 2 of ˜vn and Eq. (4.2), we recover<br />
N<br />
2 h
(N +1+σ)! n=N+1<br />
v − vh<br />
Approximation theory<br />
2<br />
|˜vn| I,0 (n − σ)!<br />
n=N+1<br />
,<br />
hich, combined with Lemma 4.2 and σ = min(N +1,p), gives the result.<br />
(N +1− σ)!<br />
≤<br />
(N +1+σ)!<br />
A sharper result can be obtained by using<br />
∞<br />
2 (n + σ)!<br />
which, combined with Lemma 4.2 and σ = min(N +1,p), gives the result.<br />
A generalization of this is the following result.<br />
A generalization of this is the following result.<br />
emma 4.4. If v ∈ H p (I), p ≥ 1 then<br />
Lemma 4.4. If v ∈ H p (I), p ≥ 1 then<br />
v (q) − v (q)<br />
h I,0 ≤<br />
v (q) − v (q)<br />
h I,0 ≤<br />
where σ = min(N +1,p) and q ≤ p.<br />
here σ = min(N +1,p) and q ≤ p.<br />
<br />
1/2 (N +1− σ)!<br />
|v| I,σ |v| , I,σ ,<br />
(N + 1 + σ − 4q)!<br />
The proof follows the one above and is omitted. Note that in the limit<br />
N ≫ p, the Stirling formula gives<br />
in agreement with Theorem 4.1.<br />
<br />
(N +1− σ)!<br />
(N + 1 + σ − 4q)!<br />
1/2 v (q) − v (q)<br />
h I,0 ≤ N 2q−p |v| I,p ,<br />
(n − σ)!<br />
The proof follows the one above and is omitted. Note that in the limit<br />
≫ p, Note the Stirling that in formula the limit gives of N>>p we recover<br />
agreement with Theorem 4.1.<br />
v (q) − v (q)<br />
h I,0 ≤ N 2q−p |v| I,p ,<br />
A minor issues arises -- these results are based on<br />
projections and we are using interpolations ?
The above estimates is based on areN all + n=0 related 1 points. to The projections. difference However, mayn=0 beasminor, we disbut<br />
it is essential to<br />
cussedisinbased Chapter appreciate on N 3, + we1 are points. it. Consider oftenThe concerned difference withmay the be interpolations minor, butofitvis essential to<br />
where<br />
Approximation theory<br />
4.3 Approximations by orthogonal polynomials a<br />
v = V ˆv,<br />
The above estimates are all related to projections.<br />
We consider cussed in Chapter 3, we are often concerned with th<br />
where N<br />
vh(r) = ˆvn v = V ˆv,<br />
˜ N<br />
Pn(r), ˜vh(r) = ˜vn ˜ appreciate it. Consider<br />
N<br />
vh(r) = ˆvn<br />
n=0<br />
Pn(r),<br />
˜ N<br />
Pn(r), ˜vh(r) = ˜vn<br />
n=0<br />
˜ N<br />
vh(r) = ˆvn<br />
n=0<br />
Pn(r),<br />
where vh(x) is the usual approximation based on interpolation and ˜vh(r) refers<br />
˜ N<br />
Pn(r), ˜vh(r) = ˜vn<br />
n=0<br />
˜ ere vh(x) is the usual approximation based on interpolation and ˜vh(r) ref<br />
the approximation based on projection.<br />
Pn(r),<br />
By the interpolation property we have<br />
where vh(x) is the usual approximation based on interpolation and ˜vh(r) refers<br />
∞<br />
to the approximation based on projection.<br />
(V ˆv)i = vh(ri) = ˜vn ˜ N<br />
Pn(ri) = ˜vn ˜ ∞<br />
Pn(ri)+ ˜vn ˜ interpolation projection<br />
Pn(ri),<br />
is based on N + 1 points. The difference may be minor, but it is essential to<br />
appreciate it. Consider<br />
to the approximation By the interpolation based on projection. property we have<br />
n=0<br />
n=0<br />
By the interpolation is based property on N ∞ + we1 have points. The difference may be minor,<br />
appreciate (V ˆv)i = vh(ri)<br />
∞ it. = ˜vn Consider N<br />
∞<br />
˜ N<br />
Pn(ri) = ˜vn ˜ ∞<br />
Pn(ri)+ ˜vn ˜ n=0<br />
n=0<br />
n=N+1<br />
Compare the two<br />
Pn(ri),<br />
m which we recover<br />
where vh(x) is the usual approximation based on interpolation and ˜vh(r) refers<br />
to the approximation based on projection.<br />
By the interpolation property we haven=0<br />
n=0<br />
N<br />
vh(r) = ˆvn ˜ ∞<br />
˜vn Pn(r), ˜vh(r) =<br />
˜ n=0<br />
n=N+1<br />
˜vn Pn(ri),<br />
˜ Pn(r), r =(r0,...,rN ) T .<br />
∞<br />
(V ˆv)i = vh(ri) = ˜vn<br />
n=0<br />
n=0<br />
˜ N<br />
Pn(ri) = ˜vn<br />
n=0<br />
˜ from which we recover n=0<br />
Pn(ri)+<br />
from which we recover<br />
∞<br />
n=N+1<br />
V ˆv = V ˜v + ˜vn<br />
∞<br />
from which we recover<br />
n=N+1<br />
˜ Pn(r), r =(r0,...,rN ) T V ˆv = V ˜v +<br />
n=N+1<br />
.<br />
is implies<br />
This implies<br />
This implies<br />
(V ˆv)i = vh(ri) =<br />
V ˆv = V ˜v +<br />
˜vn ˜ Pn(ri) =<br />
∞<br />
N<br />
n=0<br />
˜vn ˜ Pn<br />
where∞ vh(x) is the usual approximation based on interpol<br />
˜vn to the approximation based on projection.<br />
n=N+1<br />
By the interpolation property we have<br />
˜ Pn(r), r =(r0,...,rN ) T n=N+1<br />
.<br />
vh(r) = ˜vh(r)+ ˜ P T (r)V −1<br />
∞<br />
˜vn ˜ ∞<br />
˜vn Pn(r). ˜ Pn(r).<br />
This implies<br />
V ˆv = V ˜v +<br />
vh(r) = ˜vh(r)+ ˜ P T (r)V −1<br />
vh(r) = ˜vh(r)+ ∞<br />
˜ P T (r)V −1<br />
˜vn ˜ n=N+1<br />
Now, consider the additional term on the right-hand<br />
Pn(r).<br />
side,<br />
∞<br />
(V ˆv)i = vh(ri) = ˜vn ˜ N<br />
Pn(ri) = ˜vn ˜ vh(r) = ˜vh(r)+<br />
Pn(ri)+<br />
˜ P T (r)V −1<br />
˜vn<br />
n=N+1<br />
˜ Pn(r). n=N+1<br />
Now, consider the˜P additional term on the right-hand side,<br />
T (r)V −1<br />
∞<br />
˜vn ˜ w, consider the additional term on the right-hand side,<br />
∞ <br />
T<br />
Pn(r) = ˜vn<br />
˜P −1<br />
(r)V Pn(r) ˜<br />
∞ ∞<br />
,<br />
<br />
Now, consider the additional term on the right-hand side,<br />
˜vn ˜ Pn(ri)+<br />
˜vn ˜ Pn(r), r =(r0,...,rN ) T .<br />
∞<br />
n=N+1<br />
n=N+1<br />
˜vn ˜ Pn(ri),
h h<br />
This implies<br />
vh(r) = ˜vh(r)+ ˜ P T (r)V −1<br />
∞<br />
˜vn ˜ n=N+1<br />
Now, consider the additional term on the right-handPn(r). side,<br />
Approximation theory<br />
n=N+1<br />
Now, Consider consider the this additional term term on the right-hand side,<br />
˜P T (r)V −1<br />
∞<br />
˜vn ˜ ˜P<br />
∞ <br />
T<br />
Pn(r) = ˜vn<br />
˜P −1<br />
(r)V Pn(r) ˜ ,<br />
T (r)V −1<br />
∞<br />
˜vn<br />
n=N+1<br />
˜ ∞ <br />
T<br />
Pn(r) = ˜vn<br />
˜P −1<br />
(r)V Pn(r) ˜ ,<br />
n=N+1<br />
which is allowed provided v ∈ Hp P1: FQF/GQE (I), p>1/2 P2: FQF/GQE[280]. QC: FQF/GQE Since T1: FQF<br />
n=N+1<br />
n=N+1<br />
which is allowed provided v ∈ Hp (I), p>1/2 [280]. Since<br />
˜P T (r)V −1 N<br />
Pn(r) ˜ = ˜pl<br />
l=0<br />
˜ Pl(r), V ˜p = ˜ ˜P<br />
Pn(r),<br />
we can interpret the additional term as those high-order modes (recall n>N)<br />
that look like lower order modes on the grid. It is exactly this phenomenon<br />
that is known as aliasing and which is the fundamental difference between an<br />
interpolation and a projection.<br />
The final statement, the proof of which is technical and given in [27], yields<br />
the following<br />
T (r)V −1 N<br />
Pn(r) ˜ = ˜pl<br />
l=0<br />
˜ Pl(r), V ˜p = ˜ Pn(r),<br />
we Caused can interpret by interpolation the additional termof ashigh those high-order modes (recall n>N)<br />
that look like lower order modes on the grid. It is exactly this phenomenon<br />
frequency unresolved modes<br />
n = 6<br />
that is known as aliasing and which is the fundamental difference between an<br />
interpolation and a projection.<br />
The final statement, Aliasing the proof of which is technical and given in [27], yields<br />
n = −2<br />
the following<br />
n = 10<br />
Caused by the grid<br />
n n<br />
CUUK702-Hesthaven June 6, 2006 18:54<br />
2.2 Discrete trigonometric polynomials 3
Approximation theory<br />
This 82 has 4 Insight a some through impact theory on the accuracy<br />
Theorem 4.5. Assume that v ∈ Hp (I), p > 1<br />
Theorem 4.5. Assume that v ∈ H<br />
polynomial interpolation of order N. Then<br />
p (I), p > 1<br />
polynomial interpolation of order N. Then<br />
where 0 ≤ q ≤ p.<br />
v − vhI,q ≤ N 2q−p+1/2 v − vhI,q ≤ N |v| I,p ,<br />
2q−p+1/2 |v| I,p ,<br />
2 , and that vh 2 represents a<br />
, and that vh represents a<br />
Note in particular that the order of convergence can be up to one order<br />
lower due to the aliasing errors, but not worse than that.<br />
Example 4.6. Let us revisit Example 3.2 where we considered the error when<br />
computing the discrete derivative. Using Theorem 4.5, we get a crude estimate<br />
as<br />
v ′ − DrvhI,0 ≤v− vhI,1 ≤ N 5/2−p Example 4.6. Let us revisit Example 3.2 where we considered the error when<br />
computing the discrete derivative. Using Theorem 4.5, we get a crude estimate<br />
as<br />
v |v| I,p .<br />
′ − DrvhI,0 ≤v− vhI,1 ≤ N 5/2−p |v| I,p .<br />
For the first example, that is,<br />
v(r) = exp(sin(πr)),
the situation is different. Note that v (i) ∈ Hi (I). In this case, the computations<br />
indicate<br />
Approximation theory<br />
This 82 has 4 Insight a some through impact theory on the accuracy<br />
Theorem 4.5. Assume that v ∈ Hp (I), p > 1<br />
which is considerably better than indicated by the general result. An improved<br />
estimate, valid only for interpolation of Gauss-Lobatto grid points, is [27]<br />
polynomial interpolation v of order N. Then<br />
′ − DrvhI,0 ≤ N 1−p Theorem 4.5. Assume that v ∈ H<br />
|v| I,p ,<br />
p (I), p > 1<br />
polynomial interpolation of order N. Then<br />
2 , and that vh 2 represents a<br />
, and that vh represents a<br />
v − vhI,q ≤ N 2q−p+1/2 and is much closer to the observed v − vh behavior but suboptimal. |v| I,p , This is, however,<br />
I,q ≤ N<br />
a sharp result for the general case, reflecting that special cases may show<br />
better whereresults. 0 ≤ q ≤ p.<br />
2q−p+1/2 |v| I,p ,<br />
where 0 ≤ q ≤ p.<br />
Note in particular that the order of convergence can be up to one order<br />
We return to the behavior for the general element of length h to recover<br />
lower due to the aliasing errors, but not worse than that.<br />
the estimates for u(x) rather than v(r). We have the following bound:<br />
To also account for the cell size we have<br />
Example 4.6. Let us revisit Example 3.2 where we considered the error when<br />
computing the discrete derivative. Using Theorem 4.5, we get a crude estimate<br />
as<br />
v ′ − DrvhI,0 ≤v− vhI,1 ≤ N 5/2−p Theorem 4.7. Assume that u ∈ H<br />
|v| I,p .<br />
p (D k ) and that uh represents a piecewise<br />
polynomial approximation of order N. Then<br />
u − uhΩ,q,h ≤ Ch σ−q Example 4.6. Let us revisit Example 3.2 where we considered the error when<br />
computing the discrete derivative. Using Theorem 4.5, we get a crude estimate<br />
as<br />
v |u|Ω,σ,h,<br />
′ − DrvhI,0 ≤v− vhI,1 ≤ N 5/2−p |v| I,p .<br />
for For 0 ≤ the q ≤ first σ, example, and σ = min(N that is, +1,p).<br />
(v (i) )r − Drv (i)<br />
h I,0 ≤ CN1/2−i ,<br />
v(r) = exp(sin(πr)),
u − uh 2<br />
D k ,q ≤ h1−2qv − vh 2<br />
Similarly fashion, we have<br />
Approximation theory<br />
I,q ≤ h1−2q |v| 2<br />
I,σ = h2σ−2q |u| 2<br />
D k ,σ ,<br />
where we have used Lemma 4.4 and defined σ = min(N +1,p). Summing over<br />
all elements yields the result. <br />
For a more general grid where the element length is variable, it is natural to<br />
use h = maxk hk u<br />
(i.e., the maximum interval length). Combining this with<br />
Theorem 4.7 yields the main approximation result:<br />
2<br />
D k ,q =<br />
q<br />
|u| 2<br />
D k ,p =<br />
q<br />
h 1−2p |v| 2<br />
p=0<br />
p=0<br />
Combining everything, we have the general result<br />
We combine these estimates to obtain<br />
Theorem 4.8. Assume that u ∈ Hp (D k ), p>1/2, and that uh represents a<br />
piecewise polynomial interpolation of order N. Then<br />
h<br />
u − uhΩ,q,h ≤ C<br />
σ−q<br />
u − uh<br />
|u|Ω,σ,h,<br />
N p−2q−1/2 2<br />
D k ,q ≤ h1−2qv − vh 2<br />
I,q ≤ h1−2q |v| 2<br />
I,σ<br />
for 0 ≤ q ≤ σ, and σ = min(N +1,p).<br />
This result gives the essence of the approximation properties; that is, it shows<br />
clearly under which conditions on u we can expect consistency and what<br />
convergence rate to expect depending on the regularity of u and the norm in<br />
which Theorem the error4.7 is measured. yields the main approximation result:<br />
4.4 Stability<br />
I,p ≤ h1−2q <br />
= h2σ−<br />
where we have used Lemma 4.4 and defined σ = min(N +1,p<br />
all elements yields the result.<br />
For a more general grid where the element length is variable<br />
use h = maxk hk (i.e., the maximum interval length). Com<br />
with<br />
Theorem 4.8. Assume that u ∈ Hp (D k ), p>1/2, and tha<br />
piecewise polynomial interpolation of order N. Then
Lets summarize<br />
Fluxes:<br />
• For linear systems, we can derive exact upwind fluxes<br />
using Rankine-Hugonoit conditions.<br />
Accuracy:<br />
• Legendre polynomials are the right basis<br />
• We must use good interpolation points<br />
• Local accuracy depends on elementwise smoothness<br />
• Aliasing appears due to the grid but is under control<br />
• For smooth problems, we have a spectral method<br />
• Convergence can be recovered in two ways<br />
• Increase N<br />
• Decrease h