FEM3
FEM3
∂u
+ Lu = f,
∂t
∂u ∂u
+a + bu = f,
∂t ∂x
∂2u
+ Lu = f,
∂t2
∂2u ∂u
+γ + Lu = f,
∂t2 ∂t
to which initial and boundary conditions will be added, to uniquely dene
their solutions. The term Lu contains partial derivatives (of order up to 2)
with respect to the space variable, and represents the elliptic part of the
equation. It can take, for example, any of the forms we have considered in
Chapters 2 and 3.
The rst equation represents a parabolic equation, while the following
three are of hyperbolic type. Here are some examples of second order time
dependent PDE we will consider.
1. Some diusion equations (parabolic)
(i) Heat equation:
∂u ∂2u
1D : − ν 2 = 0, 0 < x < 1, 0 < t ≤ T
∂t ∂x
∂u
2D, 3D : − ν∇2 u = 0, x ∈ Ω, 0 < t < T
∂t
where ν is a real positive constant (thermal diusion constant).
1
(ii) More general diusion equations:
∂u ∂ ∂u
1D : − α + βu = 0, 0 < x < 1, 0 < t ≤ T
∂t ∂x ∂x
∂u
2D, 3D : − div (µ∇u) + b · ∇u + σu = f, x ∈ Ω, 0 < t < T
∂t
where α, µ, σ, f are scalar functions, β is a real number, and b is a (real)
vector eld. In particular, α and µ are diusion coecients, β and σ are
reaction coecients, while b is a velocity eld.
3. (Non relativistic) Schrödinger equation (parabolic):
∂u h̄2
2D, 3D : ıh̄ = − ∇2 u + V u, x ∈ Ω, 0 < t ≤ T
∂t 2µ
where h̄ is the reduced Plank constant, µ is a real positive constant dening
the particle reduced mass, and V = V (x, t) is a given potential energy. The
solution u is called position-space wave function.
4. Wave equation (hyperbolic):
∂2u 2
2∂ u
1D : − c = 0, 0 < x < 1, 0 < t ≤ T
∂t2 ∂x2
∂2u
2D, 3D : − c2 ∇2 u = 0, x ∈ Ω, 0 < t ≤ T
∂t2
Here and in the following two equations, the (positive) constat c denotes the
wave propagation velocity.
5. Telegrapher's equation (1D transmission line equation; hyperbolic):
∂2u ∂u 2
2∂ u
+ a − c + b u = 0, 0 < x < 1, 0 < t ≤ T
∂t2 ∂t ∂x2
where a and b denotes the dissipation and dispersion coecients, respectively.
6. Klein-Gordon equation (relativistic wave equation; hyperbolic):
1 ∂2u 2 m2 c2
2D, 3D : − ∇ u + u = 0, x ∈ Ω, 0 < t ≤ T
c2 ∂t2 h̄2
where m is the mass parameter.
The solutions of all the above equations are time dependent, i.e., they are
of the form u = u(x, t), where the space variable x belongs to some bounded
physical domain Ω, as in the elliptic case, while 0 < t ≤ T .
We will derive the weak formulations of our time depended problems,
hence construct the associated FEM discretizations, only in the space do-
mains. In this rst main stage, the time variable t will be considered as an
additional parameter, that only later will be discretized. Thus, it is con-
venient to interpret the unknown solution u(x, t), x ∈ Ω, t ∈ (0, T ), as
follows:
2
for some given integer k ≥ 0. That is, for each value of t ∈ (0, T ], u(t) is the
associated symbol denoting a function of the solely variable x ∈ Ω, which we
assume to belong to the space H k (Ω).
The notation u(t) is introduced to interpret u(x, t) as a family of functions
of the variable x, depending on the parameter t, where for each value of
t ∈ (0, T ] the corresponding function u(t), whose values are denoted by
u(t)(x), belongs to H k (Ω). In this case u(t) denotes the name of a function,
whose independent variable is x, with u(t)(x) = u(x, t). Using this notation,
0
we can write u(t) =
d ∂u
dx u(t) instead of ∂x .
In the following sections we will consider only problems whose
solutions u(x, t) are suciently smooth (at least C [0, T ] continuous 1
respect to the variable t, and, for each value of t, u(t) ∈ H (Ω) for k
∂2u
∂u
∂t − ν ∂x2 = f
0 < x < 1, 0 < t ≤ T
u(x, 0) = u0 (x), 0≤x≤1 (5.1)
u(0, t) = u(1, t) = 0, 0<t≤T
Z1 Z1 Z1
∂u ∂2u
vdx − ν vdx = f vdx, ∀v ∈ V
∂t ∂x2
0 0 0
3
from which it follows:
1 R1 R1
d R
u(t)0 v 0 dx =
dt u(t)vdx + ν f (t)vdx, ∀v ∈ V
0 0 0
u(0)(x) = u (x), 0 ≤ x ≤ 1
0
Note that the test function v = v(x) depends only on the space
variable x .
The weak formulation for the above 1D heat equation then takes the
following form:
nd u(t) ∈ V such that:
(
d
dt (u(t), v) + a(u(t), v) = (f (t), v) ∀v ∈ V
(5.2)
u(0) = u0
R1 0 0
where (·, ·) is the L2 (Ω) scalar product, and a(u, v) = ν 0 u v dx is the
bilinear form associated with the elliptic operator of our equation. Recall
that u(0) and u0 are the symbols which identify the functions u(x, 0) and
u0 (x), x ∈ Ω, respectively. As in the elliptic case, (5.2) has a unique solution,
which coincides with that of (5.1).
We recall that the statement u(t) ∈ H01 means that for each t, u(x, t)
∂u
and
∂x (x, t) are both square integrable in Ω = (0, 1). Furthermore, since
Ω is a one dimensional domain, the property u(t) ∈ H01 allows us to claim
that u(x, t) is, with respect to the variable x, a continuous function in [0, 1],
vanishing at the endpoints x = 0, 1.
d
Note that in (5.2) the derivative
dt and the equation itself are dened in
the classical sense, since, under the smoothness assumption we have made
on u(x, t), the scalar product (u(t), v) is at least a C1 continuous function
of t.
To apply the FEM to formulation (5.2), we partition the space domain
[0, 1] into N subintervals of equal length h = 1/N , and we choose an approx-
imate solution having separated variables, of the form:
Nh
X
uh (x, t) = uh (t)(x) = cj (t)ϕj (x) ≈ u(x, t), 0≤t≤T
j=1
The symbol cj (t) does not have to be interpreted, as we do for uh (t), as the
name of a function of x; it denotes a function of t. For each value of t we
4
have a dierent set of coecients {cj }, i.e., a dierent approximant uh (t) of
the unknown u(t). We also remark that these coecients depend also on the
chosen space domain partition, although, for notational simplicity, we have
not indicated this dependence explicitly.
Since we have assumed that u(x, t) is at least C 1 continuous with respect
du(t) ∂ dcj (t)
to the variable t, the derivatives
dt (x) = u̇(t)(x) = ∂t u(x, t) and dt =
ċj (t) are dened in the classical sense and are continuous functions (with
respect to the variable t) in [0, T ]. Thus the variational equation in (5.2) can
be rewritten as follows:
We recall that since u0 , u(t), uh (t) ∈ H01 (0, 1), they are continuous functions
of x ∈ [0, 1]).
Thus from (5.2) we obtain:
for 0<t≤T the following relationship
fi (t)
z }| {
Nh
X Z1 Nh
X Z1
ċj (t) ϕj ϕi dx + cj (t) a(ϕj , ϕi ) = f (t)ϕi dx, i = 1 : Nh
j=1 j=1
| {z }
0 aij 0
| {z }
mij
uh (xj , 0) = u0 (xj ), j = 1 : Nh ,
cj (0) = u0 (xj ), j = 1 : Nh .
Thus, if we dene:
5
symmetric (besides being positive denite), while the stiness matrix A is
symmetric only if the bilinear form is such.
Z1 Z1 Z1
∂u ∂ ∂u
vdx − α vdx + β uvdx = 0, ∀v ∈ V
∂t ∂x ∂x
0 0 0
After applying the integration by part rule to the second integral, hence
taking into account of the Robin condition at x = 1, we obtain:
Z1 Z1 Z1
d ∂u(t) 0
u(t)vdx + α v dx + β u(t)vdx + δ u(t)(1)v(1) = 0, ∀v ∈ V
dt ∂x
0 0 0
(5.6)
The weak formulation for the above problem then takes the following
form:
nd u(t) ∈ W = H 1 (0, 1) such that:
d
dt (u(t), v) + a(u(t), v) = 0
∀v ∈ V
u(t)(0) = γ 0<t≤T (5.7)
0≤x≤1
u(0)(x) = u (x),
0
where
Z1 Z1
0 0
a(u(t), v) = α u(t) v dx + β u(t)vdx + δu(t)(1)v(1)
0 0
is the (symmetric) bilinear form associated with the elliptic operator of our
equation. Recall that u(0) and u0 are the symbols which identify the func-
tions u(x, 0) and u0 (x), x ∈ [0, 1], respectively.
Nh
X
uh (x, t) = uh (t)(x) = γϕ0 (x) + cj (t)ϕj (x) ≈ u(t)(x), t≥0
j=1
6
where, adopting the same notation we have used in the previous example,
{ϕj (x), j = 1, . . . , Nh } is the Lagrangian basis of the space
1
Xh,Γ of the
D
continuous piecewise linear functions, associated with the chosen mesh, van-
ishing at the endpoint x = 0. Note that uh (t)(0) = γ , so that the Dirichlet
boundary condition is automatically satised for any choice of the coecients
cj (t) and we only need to impose the initial condition uh (0)(x) = u0 (x).
This latter, in general, can only be approximated. As in the previous
case, we impose the interpolation conditions
uh (xj , 0) = u0 (xj ), j = 1 : Nh ,
cj (0) = u0 (xj ), j = 1 : Nh .
we obtain:
fi
z Z 1 }| {
Nh
X Z 1 Nh
X
ċj (t) ϕj ϕi dx + cj (t) a(ϕj , ϕi ) = −γ a(ϕ0 , ϕi )dx, i = 1 : Nh
j=1 |0 j=1 0
| {z }
{z } aij
mij
that is,
(
Mċ(t) + Ac(t) = f , 0<t≤T
(5.9)
c(0) = u0
Nh
X
u(x, t) ≈ uh (x, t) = γϕ0 (x) + cj (t)ϕj (x), 0 < t ≤ T.
j=1
∂u
∂t − div (µ∇u) + b · ∇u + σu = f, in Ω, 0 < t < T
u(0) =u 0 in Ω
(5.10)
u(x, t) = 0, on ΓD , 0 < t ≤ T
∂u
µ ∂n = φ on ΓN , 0 < t ≤ T
where the data are in general functions satisfying the following assumptions
with respect to the x variable: f, σ ∈ L2 (Ω), σ ≥ 0, µ ∈ L∞ (Ω), µ ≥ µ0 > 0,
7
µ0 being a real number, b ∈ [L∞ (Ω)]2 , divb, φ ∈ L2 (Ω), u0 ∈ C(Ω). Those
which depend also on t are assumed to be at least continuous with respect
to this variable.
To derive the associated weak formulation, we proceed exactly as we did
in the previous case, except that now we have to apply Green's formula
for the divergence. Thus, taking into account of the particular boundary
conditions, rst we choose W = V = HΓ1D (Ω); then we multiply the equation
by v∈V and integrate all terms over the domain Ω. We obtain:
R R R R
Ω u̇(t)v dx − Ω div (µ∇u(t)) v dx + ΩR(b · ∇u(t)) v dx + Ω σu(t)v dx
= Ω f v dx, ∀v ∈ V 0 < t ≤ T
u(0) = u 0 in Ω (5.11)
u(t) = 0, on ΓD , 0 < t ≤ T
∂u
µ ∂n = φ on ΓN , 0 < t ≤ T
Finally, we apply the Green formula for the divergence and insert the Neu-
mann boundary condition. The problem weak formulation takes the follow-
ing form: for any given t, 0 < t ≤ T , nd u(t) ∈ V such that
R R R R
Ω u̇(t)v dx + Ω µ (∇u(t) · ∇v) dx
R
+ Ω (b · R∇u(t)) v dx + Ω σu(t)v dx
= Ω f v dx + ΓN φv dΓN , ∀v ∈ V (5.12)
u(0) = u0 in Ω
Note that this formulation can also be written in the following (standard)
more concise form:
(
(u̇(t), v) + a(u(t), v) = (f (t), v) + bN (φ, v), ∀v ∈ V
(5.13)
u(0) = u0 in Ω
where
Z Z Z
a(u, v) = µ (∇u(t) · ∇v) dx + (b · ∇u(t)) vdx + σu(t)vdx
Ω Ω Ω
and Z
bN (φ, v) = φv dΓN .
ΓN
For each 0 < t ≤ T, the above weak formulation has a unique solution
u(t) ∈ HΓ1D (Ω). If we consider also t as a variable, then u(t)(x) = u(x, t) is
also the (unique) solution of the original problem (5.10).
To apply to it the FEM method, for notational convenience we number
the interior nodes and the Neumann boundary ones from 1 to Nh . Then we
choose Vh = span{ϕj , j = 1 : Nh } ⊂ V , replace V by Vh , hence u(t) by
uh (t) and v ∈ V by vh ∈ Vh . Note that ∀vh ∈ Vh is equivalent to: for all
ϕi , i = 1 : Nh . This gives rise to the following problem discretization:
8
nd uh (t) ∈ Vh such that :
(
(u̇h (t), ϕi ) + a(uh (t), ϕi ) = (f (t), ϕi ) + bN (φ, ϕi ), i = 1 : Nh
(5.14)
uh (0) = u0h
Nh
X
uh (t)(x) = uh (x, t) = cj (t)ϕj (x), 0≤t≤T
j=1
where {cj (t) = uh (xj , t)} represents the set of unknown coecients to be
determined.
For t = 0, the interpolation conditions
implies
cj (0) = u0 (xj ), j = 1 : Nh .
Thus, for t > 0, the coecients {cj (t), j = 1 : Nh } are the (unique)
solutions of the following system of initial value ODEs:
N Z N Z Z
Ph Ph
ċj (t) ϕ j ϕ i dx + cj (t) a(ϕj , ϕi ) = f (t)ϕi dx + φϕi dΓN ,
j=1 | Ω {z } j=1 Ω ΓN
| {z }
aij | {z }
mij di (5.15)
i = 1 : Nh
cj (0) = u0 (xj ), j = 1 : Nh
(
Mċ(t) + Ac(t) = d(t)
c(0) = u0
The mass matrix M is always symmetric and positive denite, while the
mass matrix A is positive denite but not symmetric. Both of them are
highly sparse.
∂u
∂t + Lu = f x ∈ Ω, 0 < t ≤ T
u(0) = u
0 in Ω
(5.16)
u(x, t) = g(x, t), x ∈ ΓD , 0 < t ≤ T
∂u (x, t) = 0,
x ∈ ΓN , 0 < t ≤ T
∂nL
9
∂u
where L is the elliptic operator dened at the end of Chapter 2 and
∂nL
is the conormal derivative associated with it, the PDE is rewritten in the
following weak form:
where a(·, ·) is the bilinear form associated with the elliptic operator L.
Thus, the above problem is represented by the following (equivalent)
canonical weak formulation:
for any t ∈ (0, T ] nd u(t) ∈ W = H 1 (Ω) such that
(u̇(t), v) + a(u(t), v) = (f (t), v),
∀v ∈ V
u(t) = g on ΓD (5.18)
u(0) = u
0 in Ω
Note that, although not explicitly stated, the initial conditions we have im-
posed in all the previous examples could have been written in the concise
form: uh (0) = u0h in Ω.
The approximant uh (t) is dened by the expression
Nh
X Nt
X
uh (t)(x) = uh (x, t) = cj (t)ϕj (x) + cj (t)ϕj (x), 0 ≤ t ≤ T
j=1 j=Nh +1
where {cj (t) = uh (xj , t)} represents the set of unknown coecients to be
determined.
To this end, rst we impose the Dirichlet boundary condition at the
Dirichlet boundary mesh points:
10
which, recalling the properties of the basis functions, gives:
Note that only the coecients {cj (t), j = 1 : Nh , t > 0, } still need to be
determined. Then, we impose the given (approximated) initial condition
uh (0) = u0h , from which it follows:
There are many numerical methods for solving initial value problems for
ordinary dierential equations. Here we recall the very basic ones, which are
particular cases of a family of methods, called the θ−method.
To construct these methods, we rewrite the ODE system in the canonical
form:
(
ċ(t) = M−1 [d(t) − Ac(t)] , 0<t≤T
c(0) = u0
11
For simplicity, in the following its order will be denoted by N , instead of Nh .
To construct a simple numerical method for its solution, we consider rst
a sequence of temporal instants, for simplicity equidistant in the interval
[0, T ]: tk = k∆t, k = 0, 1, . . . , NT , with ∆t = T /NT . Then, we consider the
dierential equation at the time instant t = tk + θ∆t, for a chosen value
of 0 ≤ θ ≤ 1. Since the numerical method we are going to construct must
involve only quantities evaluated at tk , tk+1 , not at intermediate values t, we
need to discretize both members of the above ODE in such a way that only
the wanted quantities appear.
To this end, we rst discretize the temporal derivative by a simple incre-
mental ratio:
c(tk+1 ) − c(tk )
ċ(t) = + err1
∆t
Correspondingly, we approximate the right hand side, here denoted by F (t, c(t)),
where t = tk +θ∆t = θtk+1 +(1−θ)tk , by the same convex linear combination:
1 1
erri = O(∆t ), if θ 6= , and erri = O(∆2t ), if θ =
2 2
At this point we neglect these error terms (called local truncation errors),
hence replace the exact values c(tk ) by the associated approximate ones, that
we denote by ck . We obtain the so-called θ -method for solving our ODE
system:
ck+1 − ck
M + A[θck+1 + (1 − θ)ck ] = θf k+1 + (1 − θ)f k , 0≤θ≤1
∆t
where we have set
1 1
Mck+1 = M − A ck + f k , k = 0, 1, . . .
∆t ∆t
which is a rst-order method (with respect to ∆t.)
2. When θ=1 we have the implicit Euler method
12
ck+1 − ck
M + Ack+1 = f k+1
∆t
that is,
1 1
M + A ck+1 = Mck + f k+1 , k = 0, 1, . . .
∆t ∆t
of order 1 (with respect to ∆t).
3. When θ= 1
2 we have the Crank-Nicolson (or trapezoidal) method
ck+1 − ck 1 1
M + A(ck+1 + ck ) = (f k+1 + f k )
∆t 2 2
i.e.,
1 1 1 1 1
M + A ck+1 = M − A ck + (f k+1 + f k ), k = 0, 1, . . .
∆t 2 ∆t 2 2
All three methods lead to a linear system for the unknown ck+1 , since at
k
this stage c is known. In particular, for θ=0 its matrix is:
1
M,
∆t
while for θ=1 its matrix is:
1
M+A
∆t
In general, for any given 0≤θ≤1 the system will have the following form:
1
Kck+1 = gk , K= M + θA
∆t
where the matrix K does not depend on the time t, but only on the time
integration stepsize ∆t. Moreover, since both matrices M, A are symmetric
and positive denite, then also K is symmetric and positive denite. Indeed
we have:
1 1
KT = MT + θAT = M + θA = K
∆t ∆t
and
1 T
x Mx + θxT Ax > 0, ∀x 6= o
xT Kx =
∆t
We also remark that in general the matrix K is a banded matrix, with a very
small bandwidth. Therefore if the space domain mesh does not change with
t, this matrix can be factorized once for all at the beginning of the process.
In particular we can construct the Cholesky factorization K = HHT , which
13
in this case will give rise to a banded lower triangular matrix H, once for
all, and then, for each k , nd the unknown ck+1 by solving the following two
triangular systems:
(
Hy = gk ⇒ y
HT ck+1 = y ⇒ ck+1
After having solved the ODE system, the nal approximant of u(x, t), 0 <
t ≤ T, that we obtain, which is discrete with respect to the t variable, is:
Nh
X Nt
X
ukh (x) = ckj ϕj (x)+ g(xj , tk )ϕj (x) ≈ uh (x, tk ) ≈ u(x, tk ), k = 1 : NT .
j=1 j=Nh +1
Stability
1
(i) if θ≥ 2 , the θ−method is unconditionally stable, i.e. it is stable for each
∆t (no matter what is the value of the space discretization parameter h)
1
(ii) if θ< 2 the θ−method is stable only for
2
∆t ≤
(1 − 2θ)λN
h
∆t ≤ C(θ)h2 ,
Convergence
Under the assumption that the data f, u0 and the solution u are su-
ciently smooth, the θ-FEM method we have dened in the previous section
is, in the L2 (Ω)-norm, unconditionally convergent in case (i), and condition-
ally convergent in case (ii). Moreover, for the associated errors we have the
following estimates:
1
max1≤k≤NT ku(tk ) − ukh kL2 (Ω) ≤ C1 h2 + ∆t
if p 6= 2
1
max1≤k≤NT ku(tk ) − ukh kL2 (Ω) ≤ C2 h2 + (∆t)2
if p = 2
where the constants C1 , C2 do not depend on h and ∆t.
14
condition and a Dirichlet boundary condition:
∂u ∂u
∂t + a ∂x + bu = f,
0 < x < 1, 0 < t ≤ T
u(x, 0) = u0 (x), 0≤x≤1 (5.21)
0<t≤T
u(0, t) = α(t),
where a = a(x) ≥ a0 > 0, b = b(x) ≥ 0, f = f (x, t), u0 (x), α(t) are given
functions. We remark that since this equation is of rst order with respect
to the space variable x, the boundary condition must be imposed only at
one of the two space interval endpoints 0, 1. To uniquely dene the problem
solution, this endpoint must be x = 0 when a is positive, and x = 1 when a
is negative; thus in our case it is x = 0. We assume that the data functions
are suciently smooth and guarantee the existence (and uniqueness) of a
solution which is at least C 1 [0, T ] with respect to the t variable, and, for
each t∈ [0, 1], H 1 with respect to the x variable.
where u = u(x, t), f = f (x, t), u0 = u0 (x) are vector functions with p com-
ponents each, and A, B are p × p real matrices, is dened hyperbolic if the
eigenvalues of A are all real and the associated eigenvectors are linearly in-
dependent. This is true, for example, when the matrix A is symmetric, or
when its eigenvalues are all real and distinct. Note that the PDE in (5.21)
is of this type, where p = 1 and the 1 × 1 matrix A coincides with the real
number a.
Note also that the classical wave equation, which is a second order PDE,
is hyperbolic also according to this denition. Indeed, if we consider, for
simplicity, the 1D case, the equation
∂2u 2
2∂ u
− ν = f, ν>0
∂t2 ∂x2
with initial conditions u(x, 0) = u0 (x) and ∂u
∂t (x, 0) = v0 (x), can always be
rewritten as an equivalent system of two rst order PDE:
(
∂z1
∂t − ∂z
∂x = 0
2
∂z2 (5.23)
∂t − ν 2 ∂z
∂x = f
1
with initial conditions z1 (x, 0) = u00 (x), z2 (x, 0) = v0 (x). This system is
easily obtained by setting
∂u ∂u
z1 = , z2 =
∂x ∂t
15
and recalling that
∂2u ∂2u
= .
∂t∂x ∂x∂t
The latter equality holds since when we apply the FEM we always assume
that the associated problem data are such that u ∈ H 2 with respect to x, and
u ∈ C 2 with respect to t.
The associated matrix A is:
!
0 −1
A= ,
−ν 2 0
It is not symmetric, but its eigenvalues are −ν, ν , which besides being real
are also distinct.
Let us go back to our problem (5.21). Since we have a Dirichlet condition
atx = 0, the test function space is V = {v ∈ H 1 (0, 1) : v(0) = 0}. Moreover,
1
when α(t) is not the zero constant, we take W = H (0, 1). Thus, setting as
usual u(x, t) = u(t)(x), we easily obtain the following weak formulation:
for every given 0 < t ≤ T , nd u(t) ∈ W such that
R
1 R1 0
R1 R1
0 u̇(t)vdx + 0 au(t) vdx + 0 bu(t)vdx = 0 f vdx,
∀v ∈ V
u(0)(x) = u0 (x), 0≤x≤1 (5.24)
u(t)(0) = α(t), t>0
Note that since we only have a rst order (partial) derivative with respect
to the x variable, no integration by parts is performed.
To apply our nite element method, as usual we subdivide the interval
[0, 1] into N sub-intervals of maximum length h, and associate with it the
standard piecewise linear Lagrange basis. For the above problem, using the
same notation we have introduced in the previous section, we set Nh = N and
replace the innite dimensional spaces W, V by Wh = span{ϕ0 , ϕ1 , . . . , ϕNh }
and Vh = span{ϕ1 , . . . , ϕNh } , respectively. The corresponding discretized
weak formulation then takes the following form:
for every given 0 < t ≤ T , nd uh (t) ∈ Wh such that
R
1 R1 0
R1 R1
0 u̇h (t)vh dx + 0 auh (t) vh dx + 0 buh (t)vh dx = 0 f vh dx,
∀vh ∈ Vh
uh (0) = u0h , in 0 ≤ x ≤ 1 (5.25)
u (t)(0) = α(t),
h t>0
where u0h is the element of Wh which interpolates the datum u0 at the space
domain mesh points points, that is:
Nh
X
u(t)(x) ≈ uh (t)(x) = cj (t)ϕj (x), 0 < t ≤ T.
j=0
16
Before inserting this expression into the above weak formulation, we im-
pose the boundary condition:
Nh
X
uh (t)(0) = cj (t)ϕj (0) = c0 (t) = α(t), t>0
j=0
Nh
X
uh (t)(x) = α(t)ϕ0 (x) + cj (t)ϕj (x), t > 0.
j=1
After inserting this expression into (5.25), and imposing the correspond-
ing initial conditions
uh (0)(xk ) = u0 (xk ), k = 1, . . . , Nh
from which it follows ck (0) = u0 (xk ), we obtain the nal linear system for
the unknown coecients {cj (t), j = 1, . . . , Nh }:
for every given 0 < t ≤ T , nd {cj (t), j = 1, . . . , Nh } such that
PNh R1 PNh R1 0
j=1 ċj (t) 0 ϕj ϕi dx + j=1 cj (t) 0 (aϕj + bϕj )ϕi dx =
R
1 R 01 R 1
0 f ϕi dx − α(t) 0 aϕ0 ϕi dx − α̇(t) 0 ϕ0 ϕi dx, i = 1, . . . , Nh
(5.26)
cj (0) = u0 (xj ), j = 1, . . . , Nh
(
Mċ(t) + Ac(t) = f (t), 0<t≤T
c(0) = u0
This latter can be solved by using the θ-method we have described in the
previous section.
17
where a(·, ·) is the bilinear form associated with the elliptic operator L of
the given PDE.
To apply to it the FEM method we repeat the same steps we have
performed for the corresponding parabolic case. Thus, as in the parabolic
case, we number the interior nodes and, if they are present, the Neumann
boundary ones, from 1 to Nh . Those belonging to a Dirichlet boundary,
if this is present, are numbered from Nh + 1 up to Nt . Then we choose
Wh = span{ϕj , j = 1 : Nt } ⊂ W and Vh = span{ϕj , j = 1 : Nh } ⊂ V ,
replace W by Wh , V by Vh , hence u(t) by uh (t) and v by vh . Finally, we ap-
ply the Galerkin method to (5.27). This gives rise to the following problem
semi-discretization:
nd uh (t) ∈ Wh such that :
(üh (t), vh ) + a(uh (t), vh ) = (f (t), vh ), ∀vh ∈ Vh
Discrete Dirichlet boundary condition (if required)
uh (0) = u0h in Ω
u̇h (0) = v0h in Ω
where u0h , v0h are the elements of Wh which interpolate the data u0 , v0 at
all mesh points (internal and boundary ones).
In the more general case of mixed type boundary conditions, the approx-
imant uh (t), t > 0, is dened by the expression
Nh
X Nt
X
uh (t)(x) = uh (x, t) = cj (t)ϕj (x) + g(xj , t)ϕj (x)
j=1 j=Nh +1
where g(x, t) is the given Dirichlet datum and {cj (t) = uh (xj , t)} represents
the set of unknown coecients to be determined. Note that the last sum
above is empty when the problem does not have a Dirichlet datum.
Then we choose
Nt
X
u0h (x) = u0 (xj )ϕj (x)
j=1
and
Nt
X
v0h (x) = v0 (xj )ϕj (x).
j=1
Therefore we have:
and
v0h (xi ) = v0 (xi ), i = 1 : Nh
i.e.,
uh (0)(xj ) = cj (0) = u0 (xj ), j = 1 : Nh
18
and
u̇h (0)(xj ) = ċj (0) = v0 (xj ), j = 1 : Nh .
The coecients {cj (t), j = 1 : Nh are uniquely dened by the following
initial value ODE system:
NPh
Z N
Ph
c̈ (t) ϕj ϕi dx + cj (t) a(ϕj ϕi )
j
j=1 Ω j=1 | {z }
| {z } aij
mij
fi
z }| {
Nt
(5.28)
Z X Z Z
=
f (t)ϕi dx − ġ(xj , t) ϕj ϕi dx − g(xj , t) a(ϕj , ϕi )dx ,
Ω Ω Ω
j=Nh +1
i = 1 : Nh
cj (0) = u0 (xj ), i = 1 : Nh
ċj (0) = v0 (xj ), i = 1 : Nh
We recall that the matrices M, A are both positive denite and (highly)
sparse. Moreover, the mass matrix M is always symmetric, while A is sym-
metric only if the bilinear form a(·, ·) is symmetric. For notational simplicity,
in the following their order will be denoted by N .
As in the previous cases, we have still obtained a second order linear
ODE system, with initial conditions.
Formulation (5.28) denes a semi-discretization of (5.18), since the time
variable t has not yet been discretized. This variable will then be descritized,
when we will solve the ODE system by applying to it a numerical method.
To construct a numerical method for the solution of the above initial
value ODE problems, the ODE system must be rewritten in the following
form:
−1
c̈(t) = M [f (t) − Ac(t)] ,
0<t≤T
c(0) = u
0
ċ(0) = v
0
19
a dissipation term γ ∂u
∂t , as it happens in the telegrapher's equation, after
applying the FE method we obtain a nal ODE system of the following
form:
−1
c̈(t) = M [f (t) − γMċ(t) − Ac(t)] ,
0<t≤T
c(0) = u 0 (5.29)
ċ(0) = v
0
Since the method commonly used to solve this second order initial value
problem is the Newmark method , we briey describe the construction of
this method for a general second order problem of the form:
ÿ(t) = F (t, y(t), ẏ(t)),
t>0
y(0) = y0 (5.30)
ẏ(0) = z
0
∆2t 00
y(tk+1 ) = y(tk ) + ∆t y 0 (tk ) + y (tk ) + O(∆3t )
2
y 0 (tk+1 ) = y 0 (tk ) + ∆t y 00 (tk ) + O(∆2t )
Since
y 00 (tk+1 ) − y 00 (tk ) = O(∆t )
for arbitrary reals β and γ, from the previous expansions we derive the
equivalent ones:
∆2t 00
y(tk+1 ) = y(tk ) + ∆t y 0 (tk ) + y (tk ) + ∆2t β y 00 (tk+1 − y 00 (tk ) + O(∆3t )
2
y 0 (tk+1 ) = y 0 (tk ) + ∆t y 00 (tk ) + ∆t γ y 00 (tk+1 − y 00 (tk ) + O(∆2t )
i.e.
1
0 00
y(tk+1 ) = y(tk ) + ∆t y (tk ) + ∆2t βy (tk+1 ) + − β y 00 (tk ) + O(∆3t )
2
20
These relationships are satised by any function y ∈ C 4 [0, T ]. Now we take
as y(t) the solution of our initial value problem (5.30); we obtain:
h i
0 2 1
y(tk+1 ) = y(tk ) + ∆t y (tk ) + ∆t βF (tk+1 , y(tk+1 )) + 2 − β F (tk , y(tk ))
+O(∆t ) 3
0
y (tk+1 ) = y 0 (tk ) + ∆t [γF (tk+1 , y(tk+1 )) + (1 − γ)F (tk , y(tk ))] + O(∆2t ).
Finally, by setting z(t) = y 0 (t), and neglecting the O(∆2t ), O(∆3t ) terms, from
the latter we obtain the following recursive relationships, which dene the
Newmark method:
F or k = 0, 1, . . . , NT : h
i
yk+1 = yk + ∆t zk + ∆2t βFk+1 + 21 − β Fk ,
zk+1 = zk + ∆t [γFk+1 + (1 − γ)Fk ] ,
∆2t
The matrix M+ 4 A is always positive denite, and it is also symmetric
whenever the bilinear form a(u, v) is symmetric. Thus, the associated system
is always non singular. Recall that also the mass matrix M is symmetric
21
and positive denite. So, rst we determine ck+1 by solving the rst linear
system in (5.31); then, after inserting this vector in the right hand side of
the second equation, we solve this latter and determine dk+1 .
The approximation we obtain, after having coupled the above Newmark
method with the FEM, is:
Nh
X Nt
X
ukh (x) = ckj ϕj (x) + g(xj , t)ϕj (x), k = 1, . . . , NT
j=1 j=Nh +1
It can be shown that if the data are suciently smooth, and satisfy the
required compatibility conditions, then the overall method (Newmark+FEM)
is unconditionally convergent in the L2 -norm, as h, ∆t → 0, and the following
error estimate holds:
2 ∆t ∆t h k
i
M dk+1 = 1−γ M dk + b − A ck + ck+1
2 + γ∆t 2 2
from the rst equation we obtain:
" #
∆2
M + t (1 − δγ )A ck+1 =
4
" #
∆2 ∆t ∆t ∆t ∆2
M − t (1 − δγ )A ck + 2−γ − 1−γ δγ M dk + t (1 − δγ )bk
4 2 2 2 4
γ∆t
where we have set δγ = 2+γ∆t . Note that 0 ≤ δ γ < 1; thus the matrix
∆2t
M+ (1 − δγ )A,
4
besides being symmetric, is also positive denite, hence non singular. Once
this system is solved, by inserting its solution ck+1 into (5.32), hence solving
this latter system, which is symmetric and positive denite, we obtain dk+1 .
This denes an iterative process similar to that we have presented above in
(5.31) for the case γ=0 (which implies δγ = 0)
22