Differential Inclusions, Control and Optimization 27 (2007 ) 7–22
ERROR ESTIMATES FOR FINITE ELEMENT APPROXIMATIONS OF ELLIPTIC
CONTROL PROBLEMS
Walter Alt
Friedrich-Schiller-Universit¨ at Jena Institute for Applied Mathematics
D–07740 Jena, Germany Nils Br¨ autigam
Friedrich-Alexander-Universit¨ at Erlangen-N¨ urnberg Institute for Applied Mathematics
D–91058 Erlangen, Germany and
Arnd R¨ osch
Johann Radon Institute for Computational and Applied Mathematics (RICAM), A–4040 Linz, Austria
Abstract
We investigate finite element approximations of one-dimensional elliptic control problems. For semidiscretizations and full discretiza- tions with piecewise constant controls we derive error estimates in the maximum norm.
Keywords: Linear quadratic optimal control problems, elliptic equa- tions, finite element approximations, error estimates.
2000 Mathematics Subject Classification: 49K20, 49M25, 65N30.
1. Introduction
The paper discusses the one-dimensional elliptic optimal control problem (CP1) min J(z, u) = 1
2 Z
T0
|z(t) − z
d(t)|
2+ ν |u(t)|
2dt s.t.
−¨z(t) + Az(t) = Bu(t) + e(t) for a.a. t ∈ [0, T ] , z(0) = z(T ) = 0 ,
a ≤ u(t) ≤ b for a.a. t ∈ [0, T ] ,
where u ∈ L
2(0, T ; R
m), z, z
d∈ W
22(0, T ; R
n), e ∈ L
∞(0, T ; R
m), A ∈ R
n×nis positive semidefinite, B ∈ R
n×mand a, b ∈ R
m, a < b.
Hinze [6] and Meyer/R¨osch [8] consider multidimensional elliptic control problems and derive error estimates in the L
2-norm for finite element dis- cretizations. We consider only one-dimensional problems, however, we deal with vector-valued controls and states. Based on the ideas of Sendov/Popov [10] we derive error estimates in the L
∞-norm, and we prove a convergence result according to Meyer/R¨osch [8].
The following notations are used. By X(0, T ; R
n) we denote a space of functions on [0, T ] with values in R
n. By L
2(0, T ; R
n) we denote the Hilbert space of square-integrable functions with the usual scalar product (·, ·) and the corresponding norm k · k
2. By L
∞(0, T ; R
n) we denote the space of essentially bounded functions with norm k · k
∞, by W
pk(0, T ; R
n) the space of absolutely continuous functions which are k times differentiable a.e. on [0, T ] and whose k-th derivative belongs to L
p(0, T ; R
n). The norm of a function z ∈ W
21(0, T ; R
n) is defined by kzk
1,2= kzk
22+ k ˙zk
22 12. By W
p,0k(0, T ; R
n) we denote the space of functions z ∈ W
pk(0, T ; R
n) satisfying the boundary conditions z(0) = z(T ) = 0. For a matrix A ∈ R
n×nwe use the matrix norm kAk = sup
|x|=1|Ax|, where | · | denotes the Euclidian norm of a vector in R
n.
2. Optimality conditions
In the case of one-dimensional domains considered here the state equation
(1) −¨z(t) + Az(t) = y(t) for a.a. t ∈ [0, T ] ,
z(0) = 0 , z(T ) = 0 ,
is equivalent to its variational formulation
(2) ( ˙z, ˙v) + (Az, v) = (y, v) ∀v ∈ V
0= W
2,01(0, T ; R
n)
(see e.g., Br¨autigam [2]). By S : L
2(0, T ; R
n) → W
2,01(0, T ; R
n) we denote the continuous linear operator which assigns to each y ∈ L
2(0, T ; R
n) the unique solution z = S(y) of (2). With the aid of the operator S, Problem (CP1) can be equivalently written in the form
(CP2) min F (u) = J(S(Bu + e), u) s.t. u ∈ U
ad,
where U
ad= {u ∈ L
2(0, T ; R
m) | a ≤ u(t) ≤ b for a.a.t ∈ [0, T ]}.
Problem (CP2) has a unique solution ¯ u which is characterized by the pointwise variational inequality
(3)
B
Tp(t) + ν ¯ ¯ u(t)
T(u − ¯u(t)) ≥ 0 ∀u ∈ U
for a.a. t ∈ [0, T ], where U = {u ∈ R
m| a ≤ u ≤ b}, ¯p = S
∗(¯ z − z
d) = S(¯ z − z
d) is the solution of the adjoint equation, and ¯ z = S(B ¯ u + e) is the solution of (2). Equation (3) implies that ¯ u(t) is the projection of −
1νB
Tp(t) ¯ onto [a, b], i.e.,
(4) u(t) = P r ¯
[a,b]− 1
ν B
Tp(t) ¯
(compare e.g. Malanowski [7]). This further implies ¯ u ∈ W
∞1(0, T ; R
m).
3. Discretization of the state equation
For discretizations of Problem (CP2) we use a finite element discretization of the state equation. To this end we define a mesh size h = T /N for N ≥ 2 and a uniform grid
(5) G = {t
i= ih | i = 0, . . . , N}
on [0, T ]. In the variational equation (2) we replace the space V
0by the
finite-dimensional subspace V
h,0of continuous, piecewise linear functions v
hon the grid G satisfying the boundary conditions v
h(0) = v
h(T ) = 0. In this way we obtain the finite element discretization
(6) ( ˙z, ˙v
h) + (Az, v
h) = (y, v
h) ∀v
h∈ V
h,0of the state equation.
The functions v
h(j), j = 1, . . . , N − 1, defined by
(7) v
h(j)(t) =
t − t
j−1h , t ∈ [t
j−1, t
j] t
j+1− t
h , t ∈ [t
j, t
j+1]
0, else,
form a basis of V
h,0.
By S
h: L
2(0, T ; R
n) → V
h,0we denote the continuous linear operator which assigns to each y ∈ L
2(0, T ; R
n) the unique solution z
h= S
h(y) of (6).
The operators S
hare stable in the following sense:
Theorem 1. There are nonnegative constants c
1and c
2independent of h such that
kS
h(y)k
2≤ c
1kyk
2∀y ∈ L
2(0, T ; R
n) , and
kS
h(y)k
∞≤ c
2kyk
∞∀y ∈ L
∞(0, T ; R
n) . P roof. The inequalities directly follow from (6) with v
h= z
h.
Estimates for the discretization error S(y) − S
h(y) play a crucial role in the proof of error estimates for discretizations of Problem (CP2). For es- timates in the L
2-norm we use the Aubin-Nitsche-Lemma (compare Aubin [1], Nitsche [9], Ciarlet [4]).
Lemma 2. For an arbitrary y ∈ L
2(0, T ; R
n) the error estimate kS(y) − S
h(y)k
2≤ c kyk
2h
2holds with a constant c ≥ 0 independent of y and h.
For error estimates in the L
∞-norm Sendov/Popov [10] state a result for
scalar equations (see [10], Theorem 7.8). The following extension to systems
of equations has been proved in Br¨autigam [2].
Theorem 3. Let y ∈ L
∞(0, T ; R
n). Then the error estimate kS(y) − S
h(y)k
∞≤ c kyk
∞h
2holds with a constant c ≥ 0 independent of y and h.
4. Semi-discretization of the control problem
We consider the following discretization of Problem (CP2), where only the state variables are discretized:
(CP3) min F
h(u) = J(S
h(Bu + e), u) s.t. u ∈ U
ad.
As in the case of Problem (CP2) it is well-known that Problem (CP3) has a unique solution ¯ u
hwhich is characterized by the pointwise variational inequality
(8)
B
Tp ¯
h(t) + ν ¯ u
h(t)
T(u − ¯u
h(t)) ≥ 0 ∀u ∈ U
for a.a. t ∈ [0, T ], where U = {u ∈ R
m| a ≤ u ≤ b}, ¯ p
h= S
h∗(¯ z
h− z
d) = S
h(¯ z
h− z
d) is the solution of the discretized adjoint equation and ¯ z
h= S
h(B ¯ u
h+ e) is the solution of the discretized state equation (compare e.g., Malanowski [7], Tr¨oltzsch [11]).
Since the optimal control ¯ u is feasible for (CP3) we can use (8) to prove the following estimate in the L
∞-norm (see Hinze [6], Theorem 3.6):
k¯u − ¯u
hk
∞≤ C h
k(S
∗− S
h∗)(S(B ¯ u + e) − z
d)k
∞+ h
2kB¯uk
2i .
Combining this result with Theorem 3 and using the fact that S
∗h= S
hand S
∗= S we obtain
Theorem 4. For the solution ¯ u to Problem (CP2) and the solution ¯ u
hto Problem (CP3) the error estimate
k¯u − ¯u
hk
∞≤ c h
2holds with a constant c ≥ 0 independent of h.
5. Full discretization of the control problem
For a full discretization we use the grid G defined by (5) and we denote by U
hthe finite-dimensional space of piecewise constant polynomials on the grid G. Further we define
s
i= 1
2 (t
i+ t
i+1) , i = 0, . . . , N − 1 . In this way we obtain the fully discretized problem (CP4) min F
h(u) = J(S
h(Bu + e), u)
s.t. u ∈ U
ad∩ U
h.
As in the case of Problems (CP2) and (CP3) it is well-known that Prob- lem (CP4) has a unique solution ¯ u
hwhich is characterized by the pointwise variational inequality
(9)
B
Tp ¯
h(s
i) + ν ¯ u
h(s
i)
T(u − ¯u
h(s
i)) ≥ 0 ∀u ∈ U
for i = 0, . . . , N − 1, where U = {u ∈ R
m| a ≤ u ≤ b}, ¯p
h= S
h∗(¯ z
h− z
d) = S
h(¯ z
h− z
d) is the solution of the discretized adjoint equation and ¯ z
h= S
h(B ¯ u
h+ e) is the solution of the discretized state equation.
For y ∈ L
2(0, T ; R
n) we define Y (t) =
Z
t 0y(s) ds , t ∈ [0, T ] .
In view of estimates for the discretization error P
hu − ¯u ¯
hwe need the fol- lowing auxiliary result.
Lemma 5. Let y ∈ L
2(0, T ; R
n) and let z
h= S
h(y) be the solution of (6).
Then we have
(10) kz
hk
∞≤ c kY k
1with some constant c ≥ 0 independent of y.
P roof. We denote by z
0h∈ V
h,0the solution of (6) for A = 0. Then Z
T0
( ˙z
h0(t) + Y (t))
T˙v
h(t) dt = 0 ∀v
h∈ V
h,0.
Choosing the basis functions v
h= v
(j)h, j = 1, . . . , N − 1, defined by (7), we obtain
− z
h0(t
j+1) − 2z
0h(t
j) + z
0h(t
j−1)
h + 1
h Y
j−1− 1
h Y
j= 0 , for j = 1, . . . , N − 1, where
Y
j= 1 h
Z
tj+1tj
Y (t) dt , j = 0, . . . , N − 1 .
This shows that the derivative of the continuous, piecewise linear function w
hdefined by
w
h(t
j) = −z
h0(t
j+1) + z
h0(t
j) − Y
j, j = 0, . . . , N − 1 ,
vanishes. Hence, with some constant q we have w
h≡ q on [t
0, t
N−1], i.e., z
h0(t
j+1) − z
h0(t
j) + Y
j= q , j = 0, . . . , N − 1 ,
From z
0h(0) = z
h0(T ) = 0 we obtain
0 = z
h0(T ) = h
N−1
X
j=0
z
0h(t
j+1) − z
h0(t
j)
h = N q −
N−1
X
j=0
Y
j,
and therefore
q = 1 N
N−1
X
j=0
Y
j= 1 N
Z
T 0Y (t) dt .
Again with z
0h(0) = 0 this implies
z
0h(t
i) = h
i−1
X
j=0
z
h0(t
j+1) − z
h0(t
j)
h = iq −
i−1
X
j=0
Y
j= i
N − 1 X
i−1j=0
Y
j+ i N
N−1
X
j=i
Y
j,
for i = 1, . . . , N − 1. This further implies
|z
h0(t
i)| ≤
N−1
X
j=0
|Y
j| , i = 1, . . . , N − 1 .
Using the boundary conditions z
h0(0) = z
h0(T ) = 0 again we obtain
kz
h0k
∞≤
N−1
X
j=0
|Y
j| ≤ Z
T0
|Y (t)| dt = kY k
1.
Observing that z
h− z
0hsolves the discretized variational equation ( ˙z, ˙v
h) + (Az, v
h) = (Az
h0, v
h) ∀v
h∈ V
h,0it finally follows from Theorem 1 that
kz
hk
∞≤ kz
h− z
h0k
∞+ kz
h0k
∞≤ c
2kAk kz
h0k
∞+ kz
h0k
∞≤ (c
2kAk + 1) kY k
1, which proves the assertion.
In the following, we use the notation BV (0, T ; R
m) for the linear space of functions u : [0, T ] → R
mof bounded variation and V
τσ(u) for the variation of the function u on the interval [τ, σ]. Further, we define z
h(u) = S
h(Bu+e) and we approximate a control function u by the piecewise constant function P
hu defined by
(P
hu)(t) = u(s
i) ∀t ∈ [t
i, t
i+1[ , i = 0, . . . , N − 1 .
Lemma 6. Let ¯ u be the solution to Problem (CP2), and suppose that ˙¯ u ∈ BV (0, T ; R
m). Then
(11) kz
h(¯ u) − z
h(P
hu)k ¯
∞≤ c (V
0T( ˙¯ u) + k ˙¯uk
∞) h
2, with some constant c ≥ 0 independent of ¯u and h.
P roof. The piecewise linear function z
h(¯ u) − z
h(P
hu) ∈ V ¯
h,0solves the discretized variational equation (6) for y = B(¯ u − P
hu). Therefore, with ¯
Y (t) = B Z
t0
(¯ u − P
hu)(s) ds ¯
it follows from Lemma 5 that
kz
h(¯ u) − z
h(P
hu)k ¯
∞≤ c
1kY k
1= c
1N−1
X
i=0
Z
ti+1ti
|Y (t)| dt .
By the definition of Y we have Y (t) = Y (t
i) + B
Z
t ti(¯ u − P
hu)(s) ds ¯ ∀t ∈ [t
i, t
i+1] . and therefore
(12)
N−1
X
i=0
Z
ti+1ti
|Y (t)| dt =
N−1
X
i=0
Z
ti+1ti
Y (t
i) + Z
tti
B(¯ u − P
hu)(s) ds ¯ dt.
Because of Y (0) = 0 the first term on the right hand side of (12) can be estimated by
Z
ti+1ti
|Y (t
i)| ds = h
i−1
X
j=0
Y (t
j+1) − Y (t
j) = h
i−1
X
j=0
Z
tj+1tj
Y (t) dt ˙
≤ h kBk
i−1
X
j=0
Z
tj+1tj
u(t) − ¯u(s ¯
j) dt
≤ h kBk
i−1
X
j=0
Z
h20
(¯ u(s
j+ t) − ¯u(s
j)) − (¯u(s
j) − ¯u(s
j− t) dt
≤ h kBk
N−1
X
j=0
Z
h20
Z
t0
| ˙¯u(s
j+ s) − ˙¯u(s
j+ s − t)| ds dt .
Assuming ˙¯ u ∈ BV (0, T ; R
m) we obtain Z
ti+1ti
|Y (t
i)| ds ≤ h
34 kBk
N−1
X
j=0
V
ttjj+1( ˙¯ u) = h
34 kBk V
0T( ˙¯ u) , and therefore
(13)
N−1
X
i=0
Z
ti+1ti
|Y (t
i)| dt ≤ T
4 kBk V
0T( ˙¯ u) h
2.
The second term on the right hand side of (12) can be estimated by
N−1
X
i=0
Z
ti+1ti
Z
t ti(¯ u − P
hu)(s) ds ¯ dt ≤
N−1
X
i=0
Z
ti+1ti
Z
ti+1ti
|(¯u − P
hu)(s)| ds dt ¯
= h
N−1
X
i=0
Z
ti+1ti
|¯u(s) − ¯u(s
i)| ds ≤ h k ˙¯uk
∞ N−1X
i=0
Z
ti+1ti
|s − s
i| ds
≤ h k ˙¯uk
∞ N−1X
i=0
h
24 = T
4 k ˙¯uk
∞h
2. Together with (12) and (13) we obtain
Z
T0
|Y (t)| dt ≤ T
4 kBk (V
0T( ˙¯ u) + k ˙¯uk
∞) h
2. Choosing c =
T4kBk we get (11).
Defining p
h(u) = S
∗h(S
h(Bu + e) − z
d) we obtain from Lemma 6:
Corollary 7. Let ¯ u be the solution to Problem (CP2), and suppose ˙¯ u ∈ BV (0, T ; R
m). Then
(14) kp
h(¯ u) − p
h(P
hu)k ¯
∞≤ c
1(V
0T( ˙¯ u) + k ˙¯uk
∞) h
2and
(15) k¯ p − p
h(P
hu)k ¯
∞≤ c
2(kB¯u + ek
∞+ kz
dk
∞+ V
0T( ˙¯ u) + k ˙¯uk
∞) h
2with nonnegative constants c
1and c
2independent of ¯ u and h.
P roof. The estimate (14) is an immediate consequence of Theorem 1 and Lemma 6. Since
k¯ p − p
h(P
hu)k ¯
∞≤ k¯ p − p
h(¯ u)k
∞+ kp
h(¯ u) − p
h(P
hu)k ¯
∞the second estimate (15) follows from Theorem 3 and (14).
Now we prove a result on discrete convergence of the optimal controls, which
is similar to that of Theorem 2.3 in Meyer/R¨osch [8]. However, Assump-
tion (A3) used in [8] is replaced by a weaker assumption that the derivative
of the optimal control ˙¯ u has bounded variation.
Theorem 8. Let ¯ u be the solution to Problem (CP2), ¯ u
hthe solution to the discretized Problem (CP4) and suppose ˙¯ u ∈ BV (0, T ; R
m). Then
(16) k¯u
h− P
huk ¯
2≤ c (kB¯u + ek
∞+ kz
dk
∞+ V
0T( ˙¯ u) + k ˙¯uk
∞) h
2with a constant c ≥ 0 independent of ¯u and h.
P roof. Let p
h(u) = S
h∗(z
h(u)−z
d) be the solution of the discretized adjoint equation and z
h(u) = S
h(Bu + e) the solution of the discretized state equa- tion. Based on the optimality conditions for Problems (CP2) and (CP4) we obtain with the help of the same arguments as in Meyer/R¨osch [8]
ν k¯u
h− P
huk ¯
22≤ (P
hp − p ¯
h(¯ u
h), B(¯ u
h− P
hu)) . ¯ For the right hand side we find
(P
hp − p ¯
h(¯ u
h), B(¯ u
h− P
hu)) = (p ¯
h(P
hu) − p ¯
h(¯ u
h), B(¯ u
h− P
hu)) ¯ +(¯ p − p
h(P
hu), B(¯ ¯ u
h− P
hu)) ¯ +(P
hp − ¯p, B(¯u ¯
h− P
hu) . ¯
The first and third term on the right hand side can be estimated according to Meyer/R¨osch [8], Section 4. For the second term we obtain from (15)
(¯ p − p
h(P
hu), B(¯ ¯ u
h− P
hu)) ¯
≤ kBk k¯ p − p
h(P
hu)k ¯
2k¯u
h− P
huk ¯
2≤ ˜ck¯u
h− P
huk ¯
2(kB¯u + ek
∞+ kz
dk
∞+ V
0T( ˙¯ u) + k ˙¯uk
∞) h
2with a constant ˜ c ≥ 0 independent of h. Now we can proceed in the same way as in Meyer/R¨osch [8], Section 4.
Example 9. In order to illustrate the result of Theorem 8, we consider the same problem as in Hinze [6], Section 4.2, for which the optimal solution is known. In this problem, we choose T = 1, z
d≡ 2, ν = 0.1, a = −∞, b = 2.5( √
2 − 1), and e(t) = −2 + t
2− t − min(−
1ν(t
2− t), b). The optimal control is
u(t) = min(− ¯
1ν(t
2− t), b) .
Figure 1 shows the optimal control ¯ u and the discrete optimal control ¯ u
h.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0 0.5 1 1.5 2 2.5
t
Figure 1. Controls ¯ u (thin line) and ¯ u
h(thick line).
Based on the projection formula (4) we use the solution ¯ u
hof the discretized Problem (CP4) to define a new control ˜ u ∈ U
adby
(17) u = P r ˜
[a,b]− 1
ν B
TS
h∗(S
h(B ¯ u
h+ e) − z
d)
.
Meyer/R¨osch [8] have shown a superconvergence estimate for this control in the L
2-norm. We prove a corresponding result using the L
∞-norm.
Theorem 10. Let ¯ u be the solution to Problem (CP2), ¯ u
hthe solution to the discretized Problem (CP4), and suppose ˙¯ u ∈ BV (0, T ; R
m). If ˜ u is defined by (17), then we have
(18) k¯u − ˜uk
∞≤ c (kB¯u + ek
∞+ kz
dk
∞+ V
0T( ˙¯ u) + k ˙¯uk
∞) h
2with a constant c ≥ 0 independent of ¯u and h.
P roof. With the notations of Theorem 8 we have
k¯ p − p
h(¯ u
h)k
∞≤ k¯ p − p
h(P
hu)k ¯
∞+ kp
h(P
hu) − p ¯
h(¯ u
h)k
∞.
For the first term on the right hand side we obtain from Corollary 7 k¯ p − p
h(P
hu)k ¯
∞≤ c
1(kB¯u + ek
∞+ kz
dk
∞+ V
0T( ˙¯ u) + k ˙¯uk
∞) h
2. For the last term we have by Theorem 1, Lemma 2, and Theorem 8
kp
h(P
hu) − p ¯
h(¯ u
h)k
∞≤ ˜c
2kP
hu − ¯u ¯
hk
2≤ c
2(kBu + ek
∞+ kz
dk
∞+ V
0T( ˙¯ u) + k ˙¯uk
∞) h
2. Combining these estimates and using the fact that the operator P r
[a,b]is Lipschitz continuous, we get the desired result.
Example 11. In order to illustrate the result of Theorem 10 we consider the same problem as in Example 9. Figure 2 shows the function
w = − 1
ν B
TS
h∗(S
h(B ¯ u
h+ e) − z
d)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0 0.5 1 1.5 2 2.5
t
Figure 2. Control ˜ u (thick line) and function w (thin line).
and the control ˜ u = P r
[a,b](w) defined by (17). Column 3 of Table 1 and Figure 3 show the linear order of convergence w.r.t. the mesh size h of the error k¯u − ¯u
hk
∞, while columns 4 and 5 of Table 1 and Figure 4 show quadratic convergence w.r.t. the mesh size h of the error k¯u − ˜uk
∞.
h F
h(¯ u
h) k¯u − ¯u
hk
∞k¯u − ˜uk
∞k¯u − ˜uk
∞/h
21/3 2.6535 1.0355 0.2532 2.2787
1/4 2.5621 0.9382 0.1550 2.4797
1/6 2.4654 0.6952 0.0686 2.4712
1/10 2.4017 0.4502 0.0248 2.4843 1/20 2.3710 0.2375 0.0062 2.4953 1/50 2.3618 0.0980 9.9840 · 10
−42.4963 1/100 2.3604 0.0495 2.4985 · 10
−42.4985
Table 1. Error as a function of mesh size h.
0 0.05 0.1 0.15 0.2 0.25
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
h
Figure 3. Error k¯u − ¯u
hk
∞as a function of mesh size h.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0
0.05 0.1 0.15 0.2 0.25 0.3 0.35
h