Professional Documents
Culture Documents
Communicated by R. Glowinski
1. Introduction
1This work was supported in part by the Fonds zur Forderungder wissenschaftlichen Forschung
under SFB 03 Optimization and Control.
2Professor, Institut fur Mathematik, Karl-Franzens-Universitat Oraz, Graz, Austria.
3Assistant Professor, Institut fur Mathematik, Karl-Franzenf-Universitat Graz, Graz, Austria.
345
0022-3239/99/0800-0345$ 16.00/0 © 1999 Plenum Publishing Corporation
346 JOTA: VOL. 102, NO. 2, AUGUST 1999
system evaluated at the locations xi. Consequently, (1/m) £mj=1 Yi,j can be
viewed as the time-averaged mean of the trajectory at location xi.
Singular-value decomposition guarantees (Ref. 11) the existence of real
numbers c1 > c2> ... >c k >0 and unitary matrices
such that
where
where (.,.)c n denotes the canonical inner product in C". In terms of the
columns yj of Y, we express the last equality as
Here, || .||F denotes the Frobenius norm, Ul denotes the first l columns of
U, Bl denotes the first l rows of B, and similarly for Ul and Cl. Note that
the jth column of UlBl represents the Fourier expansion of order l of the
jth column yj of Y in the orthonormed basis {ui)li=1 . Utilizing the fact that
UCl has rank l and recalling that
the estimate (6) follows directly from singular-value analysis (Ref. 11). We
refer to Ul as the POD-basis of rank l. Note that
Since
Inequalities (7) and (8) establish that, for every 1 <l<k, the POD basis of
rank l is optimal in the sense of representing in the mean the columns of Y
as a linear combination by a basis of rank l. Adopting the interpretation of
the Yi,j as the velocity of a fluid at location xi and time tj, inequality (8)
expresses the fact that the first l POD basis functions capture more energy
on average than the first l functions of any other basis.
The POD expansion Yl of rank l is given by
350 JOTA: VOL. 102, NO. 2, AUGUST 1999
This property is referred to as the fact that the POD coefficients are uncorre-
lated. For further properties of the POD approximations concerning for
instance symmetry properties, we refer to the literature (Ref. 12).
Concerning the practical computation of a POD basis of rank l, let us
note that, if m<n, then one can choose to determine M eigenvectors v,
corresponding to the largest eigenvalues of YH Ye Cm x m and by (2) determine
the POD basis from
Note that the square matrix YHY has the dimension of number of time
instances tj. For historical reasons (Ref. 13), this method of determining the
POD basis is sometimes called the method of snapshots.
For the application of POD to concrete problems, the choice of lis
certainly of central importance for applying POD. It appears that no general
a priori rules are available. Rather, the choice of l is based on heuristic
considerations combined with observing the ratio of the modeled energy to
the total energy contained in the system Y, which is expressed by
where (•, • )L2 denotes the canonical inner product in L2. Instead of (5), we
obtain
JOTA: VOL. 102, NO. 2, AUGUST 1999 351
where
where again Uk and Vk contain the first k columns of the matrices U and
V, respectively. Using (11), we determine the coefficient matrix
Since the vectors c.,j, l<j<k, are the eigenvectors of problem (9) with
corresponding eigenvalues a2j, the functions cj solve (P) with
where
we infer that
Note that the elements of the matrix YTSY are given by the integrals
which is a Hilbert space endowed with the common inner product. For a
control ueL2(Q), the state ye W is given by the weak solution of the unsteady
Burgers equation,
where A denotes the Laplace operator from H01(O) to H -1 (O), the optimal
control problem can be written in the form
where the prime denotes the Frechet derivative with respect to the variable
(y, u). In particular, A* is the weak solution of the backward differential
equation
Thus, we set
Recall that the POD basis functions are orthonormal in L2(O). Hence, the
mass matrix
where
and
JOTA: VOL. 102, NO. 2, AUGUST 1999 357
Furthermore, we set
A choice has to be made concerning the dynamics from which the POD
basis is computed. In this example, we constructed the POD basis from the
dynamics of the uncontrolled Burgers equation, [i.e., Eq. (16b) with u = 0],
which was solved by the Newton method. Alternatively, we could have
determined the POD basis from the dynamics of the Burgers equation using
an expected control. The numerical solution is shown in Fig. 1. One can
observe that the dynamics go from the right to the left as t increases. The
decay of the eigenvalues ci2 of the matrix YTSY is presented in Fig. 2. The
ratio of the modeled energy to the total energy contained in the system Y
is given in Table 1 for different values of l. We chose l=5 POD basis
functions, which are presented in Fig. 3.
requires 80 sec. On the other hand, the CPU time is less than 2 sec for the
reduced-order model utilizing 5 POD basis functions (see Table 2), The
values of the cost functional are presented in Table 3. In Fig. 4, the optimal
state computed by the POD basis is shown. To compare the POD solution
to the finite-elements (FE) solution, the function t|->||y(t)||L2 is plotted for
the FE solution (—) and POD solution (+) of the uncontrolled Burgers
where a>0, v>0, and peL2(O) are given. Discretizing (17) with the tech-
niques developed in Section 4 we arrive at
subject to
Taking the inner product of Eq. (18b) with y(t), integrating on [0, i), and
using the fact that
with tf chosen sufficiently large. Associated to (19), the minimal value func-
tion Vtf: [0, tf] x R l ->R is defined by
where
JOTA: VOL. 102, NO. 2, AUGUST 1999 363
Let (y tf , utf) denote the solution to (19). Then, due to the Bellman optimality
principle, the solution to
The tensor
is given by
Since (20)-(22) hold for every te[0, tf], we have a state feedback law
of the form
where
and hence,
and thus the mapping y |->Htf(0, y) can be evaluated by solving the coupled
system consisting of the primal equation and (21), (22) on [0, tf], except for
364 JOTA: VOL. 102, NO. 2, AUGUST 1999
the difficulty that remains due to the fact that, for the terminal condition at
t = tf, the value of Vtf is unknown. In computations, we replaced the terminal
condition by
where g>0 and yeRl is a steady state of the dynamical system, in our case
y = 0. Then, assuming that y |—> Htf(0, y) is available at least for y-values on
an appropriate grid in £ C Rl, a last approximation step is necessary before
we have an implementable feedback heuristic. It consists in replacing (23)
by
As grid, we choose
As in Example 4.1, we compute the POD basis from the dynamics of the
uncontrolled Burgers equation. The numerical solution of the uncontrolled
Burgers equation via FE is shown in Fig. 6. The eigenvalues a] of the
matrix YTSY decrease very rapidly; see Fig. 7. The ratio of the modeled
energy to the total energy contained in the system Y is given in Table 4 for
different values of l. We took l=4 POD basis functions and computed the
adjoints Aj(0) for initial values
Fig. 9. L2-norms.
368 JOTA: VOL. 102, NO. 2, AUGUST 1999
i.e., for m = 54 = 625 initial data. Here, the set G was chosen in such a way that
it contains the trajectory of the POD solution to the uncontrolled Burgers
equation. Using the augmented Lagrangian-SQP method, with c=1, we
needed only 2 sec on average to compute A j (0). Finally, we solved with
piecewise-linear finite elements the closed-loop equation
(o), and the closed-loop state (x ). In Table 5, the values of the cost func-
tional for the uncontrolled solution, for the optimal state via finite elements,
for the optimal state via POD, and for the closed-loop solution are given.
For a = 0.1, similar results were obtained. Since the suboptimal feedback
strategy is based on some heuristic step, it is a priori obvious that the
feedback H constructed for a specific initial condition is also effective for
other initial conditions. In the numerical tests, the feedback operator was
successful for several other initial conditions. In Figs. 10 and 11, numerical
results are presented for
with a = 0.1.
370 JOTA: VOL. 102, NO. 2, AUGUST 1999
References
1. CHOI, H., HINZE, M., and KUNISCH, K., Suboptimal Control of Backward-
Facing-Step-Flow, Preprint 571/97, Technical University of Technology, Berlin,
Germany, 1997.
2. CHOI, H., TEMAM, R., MOIN, P., and KIM, J., Feedback Control for Unsteady
Flow and Its Application to the Stochastic Burgers Equation, Journal of Fluid
Mechanics, Vol. 253, pp. 509-543, 1993.
3. ITO, K., and RAVINDRAN, S. S., A Reduced-Basis Method for Control Problems
Governed by PDEs, Control and Estimation of Distributed Parameter Systems,
International Series of Numerical Mathematics, Vol. 126, pp. 153-168, 1998.
4. BROOMHEAD, D. S., and KING, G. P., Extracting Qualitative Dynamics from
Experimental Data, Physica, Vol. 20D, pp. 217-236, 1986.
5. BERKOOZ, G., HOLMES, P., and LUMLEY, J. L., Turbulence, Coherent Structures,
Dynamical Systems, and Symmetry, Cambridge Monographs on Mechanics,
Cambridge University Press, Cambridge, England, 1996.
6. BURGERS, J. M., Application of a Model System to Illustrate Some Points of the
Statistical Theory of Free Turbulence, Proc. Acad. Sci. Amsterdam, Vol. 43,
pp. 2-12, 1940.
7. LIGHTHILL, M. J., Viscosity Effects in Sound Waves of Finite Amplitude, Surveys
in Mechanics, pp. 250-351, 1956.
8. CHAMBERS, D. H., ADRIAN, R. J., MOIN, P., STEWART, D. S., and SUNG, H.
J., Karhunen-Loeve Expansion of the Burgers Model of Turbulence, Physics of
Fluids, Vol. 31, pp. 2573-2582, 1988.
9. TANG, K. Y., GRAHAM, W. R., and PERAIRE, J., Active Flow Control Using a
Reduced-Order Model and Optimum Control, Technical Report, Computational
Aerospace Sciences Laboratory, Department of Aeronautics and Astronautics,
MIT, 1996.
10. LY, H. V., and TRAN, H. T., Proper Orthogonal Decomposition for Flow Calcula-
tions and Optimal Control in a Horizontal CVD Reactor, Preprint CRSC-TR98-
12, Center for Research in Scientific Computation, North Carolina State Univer-
sity, 1998.
11. NOBLE, B., Applied Linear Algebra, Prentice-Hall, Englewood Cliffs, New Jersey,
1969.
12. AUBRY, N., LIAN, W. Y., and TITI, E. S., Preserving Symmetries in the Proper
Orthogonal Decomposition, SI AM Journal on Scientific Computing, Vol. 14,
pp. 483-505, 1993.
13. SIROVICH, L., Turbulence and the Dynamics of Coherent Structures, Parts 1-3,
Quarterly of Applied Mathematics, Vol. 45, pp. 561-590, 1987.
14. GOLUB, G. H., and VAN LOAN, C. F., Matrix Computations, Johns Hopkins
University Press, Baltimore, Maryland, 1989.
15. VOLKWEIN, S., Mesh Independence of an Augmented Lagrangian-SQP Method
in Hilbert Spaces and Control Problems for the Burgers Equation, PhD Thesis,
Department of Mathematics, University of Technology, Berlin, Germany, 1997.
16. ITO, K., and KUNISCH, K., Augmented Lagrangian-SQP Methods in Hilbert
Spaces and Application to Control in the Coefficient Problems, SIAM Journal on
Optimization, Vol. 6, pp. 96-125, 1996.
JOTA: VOL. 102, NO. 2, AUGUST 1999 371
17. ITO, K., and KUNISCH, K., Optimal Control, Encyclopedia of Electrical and
Electronics Engineering, John Wiley, New York, New York, Vol. 15, pp. 364-
379, 1999.
18. PRAGER, W., Numerical Computation of the Optimal Feedback Law for Nonlinear
Infinite-Time Horizon Control Problems, Technical Report, Karl-Franzens Univ-
ersitat, Graz, Austria, 1996.