Final Exam
January 14th, 2014
INSTRUCTIONS
Solve three out of four problems.
Read carefully the questions and answer in a pertinent, clear and concise way.
Write in a clear and readable way: illegible texts will not be corrected and will be graded 0
points.
You may write your answers in English, Spanish or Catalan.
Books and/or personal notes are not allowed during the exam.
The use of mobile phones, tablets, notebooks and calculators is strictly forbidden.
Time allowed: 2 hours.
NAME:
FAMILY NAME:
CHOSEN PROBLEMS:
1
2
Problem 1 [50 points]
Let Rd (d = 2, 3) be an open bounded domain and consider the boundary value problem:
u + u = f in
(1)
u n = g on
3
Problem 2 [50 points]
Let Rd (d = 2, 3) be an open bounded domain with boundary = D N , int(D )
int(N ) = . Consider the steady Stokes problem:
u + p = f in
div u = 0 in
(2)
u = 0 on D
u n pn = 0 on N ,
where f [L2 ()]d is an assigned forcing term and the viscosity > 0 is assumed to be constant.
1. [10 points] For the finite elements approximation, we consider P1 elements for the velocity
and P0 elements for the pressure. Write the definition of the finite dimensional spaces V h
and Qh . Will the discrete pressure be continuous or discontinuous? Why?
2. [20 points] Suppose that = (0, 1)2 R2 with D and N as shown in the following figure:
N D
where
X Z
h (vh ) = h2K (uh + ph f ) (vh )
KTh K
X Z
h (qh ) = h2K (uh + ph f ) qh
KTh K
4
Problem 3 [50 points]
Let Rd (d = 2, 3) be an open bounded domain. Consider the following parabolic problem:
find u = u(x, t) such that
u
u + b u = f in , t > 0
t (4)
u = 0 on , t > 0
u = u0 in , t = 0,
1. [20 points] Consider a discretization in space using the finite elements method and a dis-
cretization in time using a finite difference scheme. Write the semi-discrete problem in space
and the associated system of ordinary differential equations.
2. [15 points] Use the -method to perform the time discretization of the obtained system of
ordinary differential equations. For which values of the parameter do we obtain an explicit
scheme? And when an implicit method?
3. [15 points] If we use the forward Euler scheme or the backward Euler scheme, which restric-
tions do we have on the choice of the time step t? Which of these two methods is more
attractive from the computational point of view?
5
Problem 4 [50 points]
Let Rd (d = 2, 3) be an open bounded domain and consider the diffusion-transport problem:
find u such that
u + b u = f in = (0, 1) (0, 1)
(5)
u = 0 on .
1. [10 points] To compute the finite element approximation of problem (5) we use Lagrangian
elements of degree r 0. Will the stiffness matrix be symmetric? Why?
2. [20 points] A finite element code has been implemented to compute the numerical approx-
imation of problem (5). For = 1, b = 0 and a suitable f , the following error estimates
have been obtained with respect to a smooth exact solution u C ():
6
NUMERICAL METHODS FOR PARTIAL DIFFERENTIAL EQUATIONS
Master in Applied Mathematics and Mathematical Engineering 2013
Dr. M. Discacciati
Problem 1
Problem 2
The pressure will be discontinuous since it will approximated by a constant value in each
element K of the triangulation and there are no continuity conditions between such values
in adjacent elements K.
2. Let N = dim(V h ) and M = dim(Qh ).
The dimensions N and M correspond to the number of degrees of freedom for velocity and
pressure, respectively.
1
M coincides with the number of triangles of the triangulation:
Lets compute the number of degrees of freedom for each component of the velocity. In total,
we have k 2 nodes. We notice that on three edges of the domain we impose a Dirichlet
boundary condition, thus we have to remove these nodes from the degrees of freedom. The
nodes in which we impose a Dirichlet boundary condition are k + k + (k 2). Thus, for each
component of the velocity we have
k 2 (k + k + (k 2)) = k 2 3k + 2 = (k 1)(k 2)
degrees of freedom.
The problem is two-dimensional, so that we have two components of the velocity and N =
2(k 1)(k 2).
We can easily see that M > N and we can conclude that the inf-sup condition cannot be
satisfied.
3. If we consider the proposed finite element spaces, we can see that
where
Z
Aij = i : j i, j = 1, . . . , N
Z
Bji = j div i i = 1, . . . , N, j = 1, . . . , M
Z
Fi = f i i = 1, . . . , N
and {i }i=1,...,N and {j }j=1,...,M are suitable basis functions for the spaces V h and Qh ,
respectively.
The linear system that we have obtained coincides with the one that we would obtain without
stabilization. Thus, the proposed stabilization method gives no contribution for the finite
elements spaces that we are using and it will not be effective. Other stabilization techniques
should be considered.
2
Problem 3
1. Let V = H01 () and Vh V a suitable finite dimensional approximation. The semi-discrete
problem in space becomes: for each t > 0, find uh (t) Vh such that
Z Z Z Z
uh (t)
vh + uh (t) vh + (b uh (t))vh = f vh vh Vh
t
with uh (0) = u0h . The corresponding system of ordinary differential equations reads:
du(t)
M + Au(t) = f (t).
dt
Here, M is the mass matrix, A is the stiffness matrix associated to the bilinear form involving
the diffusion and convection terms, while u(t) = (u1 (t), . . . , uNh (t))T is the vector of the
unknown nodal values.
2. Using the -method for the time discretization we finally obtain
un+1 un
M + A(un+1 + (1 )un ) = f n+1 + (1 )f n .
t
If = 0 the scheme is explicit (and it corresponds to the forward Euler method), otherwise
it is implicit.
3. The backward Euler scheme ( = 1) is unconditionally stable, while for the forward Euler
method the following restriction holds:
t Ch2
for a suitable positive constant C > 0. At each time step, the forward Euler method
is cheaper than the backward Euler one, since one has to solve a linear system with mass
matrix (that, possibly, can be diagonalized). If the stability condition of this scheme does not
become too restrictive for the problem at hand, the forward Euler method is then generally
cheaper than the backward Euler one.
Problem 4
1. The bilinear form associated to the problem is
Z Z
a(u, v) = u v + (b u)v u, v H 1 ()
which is clearly non-symmetric due to the transport term. Thus, also the corresponding
stiffness matrix will not be symmetric.
2. We must refer to the error estimates:
ku uh kL2 () C0 hs+1 |u|H s+1 ()
ku uh kH 1 () C1 hs |u|H s+1 ()
where C0 and C1 are suitable positive constants, independent of h, and s = min{p, r}
supposing that u H p+1 () for some p 1. In this case u C () so that s = r. Thus,
using elements of degree r, we expect convergence r + 1 in L2 -norm and r in H 1 -norm.
For P1 elements, the L2 error is divided roughly by 4 when halving h denoting quadratic
convergence, while the H 1 error is divided by 2 indicating a linear convergence.
For P2 elements, the L2 error is reduced by a factor 8, while the H 1 error by 4.
The numerical convergence rates fit the theoretical estimates. Thus, we may assume that
the implementation is correct.
3
3. We have a transport-diffusion problem. Unphysical oscillations may be due to a Peclet
number larger than 1. In this case,
kbkh 1 101
Pe = = = 5 > 1.
2 2 102
Thus, the oscillations are due to a dominating transport term. A suitable stabilization
technique should be applied.