SlideShare a Scribd company logo
1 of 71
Download to read offline
International Research Training Group IGDK 1754
Optimal Control for Linear Second-Order Hyperbolic
Equations by BV-Functions in Time
Sebastian Engel
Rigorosum
01.03.2019
Sebastian Engel Optimal Control and BV-Functions Rigorosum 1
International Research Training Group IGDK 1754
Optimal Control of the Wave Equation with BV-Functions
We focus on optimal control of hyperbolic partial dierential equations:
Figure: [Hor16]
Sebastian Engel Optimal Control and BV-Functions Rigorosum 2
International Research Training Group IGDK 1754
Background - Related Work
§ Optimization with 2nd order linear hyperbolic equations:
§ Lions (Hilbert control).
and Kröner et al. (Distributed, Dirichlet, Neumann L2
-control).
Kunisch et al., and Trautmann et al.(Mp
; L2
p0;Tqq,
L2
w˚ p0;T; Mp
qq-control).
§ Thesis: 2nd order linear hyperbolic equations with BV(0,T)-controls.
§ FE with hyperbolic equations in optimal control:
§ Kröner et al. lin. cont. FE for state and control discretization.
Trautmann et al. lin. cont. state, variational discretization control.
§ Thesis: Variational discretization of u, lin. cont. FE for Dtu and state.
§ Hyperbolic equations and sparse controls
§ Kunisch et al., and Trautmann et al..
§ Thesis: Sparse BV-controls.
§ Optimization with BV-controls and other PDE constraints
§ Casas, Kunisch (semilinear elliptic, BVp
q).
Casas, Kruse, Kunisch (semilinear parabolic, BVp0;Tq-controls).
§ Thesis: 2nd order Hyperbolic equations with BVp0;Tq-controls
Sebastian Engel Optimal Control and BV-Functions Rigorosum 3
International Research Training Group IGDK 1754
Optimal Control of the Wave Equation with BV-Functions
pPq
$
’’’’’’’’’’
’’’’’’’’’’%
min
uPBVp0YTqm
1
2
ż
ˆr0YTs
pyu ´ ydq2
dxdt `
mÿ
j“1
j
ż
r0YTs
d|Dtuj|ptq “: Jpuq
s.t.
$
’’’
’’’%
Btty ´ 4y “
mÿ
j“1
ujgj in p0YTq ˆ 
y “ 0 on p0YTq ˆ B
pyYBtyq “ py0Yy1q in t0u ˆ 
§  Ă Rn (n=1,2,3) open bounded, B Lipschitz, T P p0Y8q
§ yd P W1Y1
pr0YTs; L2
pqq, py0Yy1q P H1
0
pq ˆ L2
pq
§ pgjqm
j Ă L8
pqzt0u pairwise disjoint supports wj
This strictly convex problem has a unique solution.
Clason, Kunisch [CK11]
§ Penalization by the } ¨ }BV -norm corresponds to settings where the cost is
proportional to changes in the control.
- On/O structure, less blurring (H1
´control).
Sebastian Engel Optimal Control and BV-Functions Rigorosum 4
International Research Training Group IGDK 1754
Equivalent Problem p˜Pq
Consider the following equivalent optimal control problem w.r.t. pPq:
p˜Pq
#
min
pvYcqPMp0YTqmˆRm
1
2
}SpvYcq ´ yd}2
L2pT q `
mÿ
j“1
j
ż T
0
|vj|dx “: JpvYcq
with uptq “
ˆtş
0
dvjpsq ` cj
˙m
i“1
resp.
ˆ
Dtu
up0q
˙
“
ˆ
v
c
˙ ˆ
Fundamental
theorem of calculus
˙
.
§ S is the ane control-to-state operator.
§ BVp0YTqm – Mp0YTqm ˆ Rm only possible in one dim.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 5
International Research Training Group IGDK 1754
Optimality Conditions
Adjoint Wave Equation
We dene by p˚
phq the solution of the adjoint wave equation:
$

%
Bttp˚
´ 4p˚
“ h in p0YTq ˆ 
p˚
“ 0 on p0YTq ˆ B
pp˚
YBtp˚
q “ p0Y0q in tTu ˆ 
Consider the following integrated adjoint functions for j=1,...,m:
p1YjpvYcqptq “
Tż
t
ż
wj
p˚
pSpvYcq ´ ydq gjdxdsY pvYcq P Mp0YTqm ˆ Rm
Sebastian Engel Optimal Control and BV-Functions Rigorosum 6
International Research Training Group IGDK 1754
Optimality Conditions
Theorem (Necessary and Sucient Condition)
pÝÑv YÝÑc q P Mp0YTqm ˆ Rm is the solution of p˜Pq if, for all j “ 1YXXXYm holds
´
ˆ
p1pÝÑv YÝÑc q
p1pÝÑv YÝÑc qp0q
˙
P
ˆ`
iB}ÝÑv i}Mp0YTq
˘m
i“1
0Rm
˙
Y
with p1pÝÑv YÝÑc q P C0pr0YTsq X C2
pr0YTsq.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 7
International Research Training Group IGDK 1754
Regularized Problem
Regularized Problem
Sebastian Engel Optimal Control and BV-Functions Rigorosum 8
International Research Training Group IGDK 1754
Regularized Problem
For the numerical realization, we would now like to consider a problem that is
numerically easier to solve (Pieper [Pi15]):
p˜Pq
#
min
pvYcqPL2p0YTqmˆRm
JpvYcq ` 
2
˜
mÿ
j“1
}vj}2
L2p0YTq ` }c}2
Rm
¸
“: J1
pvYcq
Theorem
Denote by pÝÑv YÝÑc q the unique solutions of p˜Pq and by u their BVp0YTqm
representation. Then we have:
§ 0 ď J1
pÝÑv YÝÑc q ´ Jpuq “ Opq
§ u
Ñ0
ÝÝÝÑ u strictly in BVp0YTqm (Fundamental theorem of calculus)
§ p
1
H2
p0YTqm
ÝÝÝÝÝÝÑ p1 with p
1
ptq :“
Tş
t
ş
wj
p˚
pSpÝÑv YÝÑc q ´ ydqgjdxds
Sebastian Engel Optimal Control and BV-Functions Rigorosum 9
International Research Training Group IGDK 1754
Regularized Problem - Optimality Conditions
Due to the additional L2
´regularization, we can use a Prox-Operator approach:
Theorem
ˆÝÑv 
ÝÑc 
˙
is optimal i
$
’
’%
ÝÑv  “ Proxř
i
i }¨}L1p0;Tq
p´1
 p
1
q
p
1
p0q ` ÝÑc  “ 0Rm
,
/.
/-
with Proxř
i
i }¨}L1p0;Tq
p´1
 p
1
q :“
¨
˝
max
´
0Y´1
 p
1Yipsq ´ i

¯
`
` min
´
0Y´1
 p
1Yipsq ` i

¯
˛
‚
m
i“1
§ Proxř
i
i }¨}L1p0;Tq
p´1
 p
1
q is semismooth.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 10
International Research Training Group IGDK 1754
BV Path-Following Algorithm
Dene the following function: F
pÝÑv ;ÝÑc q :“
˜ÝÑv ´ Prox
ř
i
i }¨}L1p0;Tq
p´ 1

 p
1 q
ÝÑc ` 1

 p1p0q
¸
Since we approximate p˜Pq by p˜Pq we consider the Path-Following algorithm:
BV Path-Following Algorithm
Input: u0 P L2
p0YTqm ˆ Rm, 0 ą 0, TOL ą 0, TOLN ą 0, k “ 0 and
# P p0Y1q
while k ą TOL do
Set i “ 0, ui
k “ uk
while }F
k pui
kq}L2p0;TqmˆRm ą TOLN do
Solve DF
k pukqpuq “ ´F
k pukq, set ui`1
k`1 “ ui
k ` u; i “ i ` 1.
end
Dene uk`1 “ ui
k, and 
k`1 “ 
k; set k “ k ` 1.
end
A similar approach was used for a fully discretized control problem with
semi-linear parabolic constraints in [CKK17].
Sebastian Engel Optimal Control and BV-Functions Rigorosum 11
International Research Training Group IGDK 1754
BV Path-Following Algorithm - Super Linearity
Theorem (Super Linearity of the Newton Method)
For each  ą 0 there exists a  ą 0 s.t. for all pÝÑv YÝÑc q P L2
p0YTqm ˆ Rm with
}pÝÑv YÝÑc q ´ pÝÑv YÝÑc q}L2p0YTqmˆRm ă Y
the semismooth Newton method algorithm converging superlinearly.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 12
International Research Training Group IGDK 1754
Numerical Example
Example:  “ r´1Y1s2
, T “ 2,  “ 0X005, and patch gpxq “ 1r´0X5Y0X5s2 pxq, i.e.
Desired state: yd :“ Spuq ´ pBtt ´ 4q9ptYxq with py0Yy1q “ p0Y0q for S and
9ptYxq :“ sinp3%tq sinp3%
2
tq
dź
i“1
cosp
%
2
xiq with  “ 3%l
4
´
2
?
2
%
¯´2
Sebastian Engel Optimal Control and BV-Functions Rigorosum 13
International Research Training Group IGDK 1754
Numerical Example
Dirac-example until  “ 10´8
(Full Discretization: v P S( and y P S( b Sh).
Sebastian Engel Optimal Control and BV-Functions Rigorosum 14
International Research Training Group IGDK 1754
Additional Control Problems
All presented results are extended for other linear dierential equations, i.e.
§ Hyperbolic constraints: Btty ` Ay “
mř
i“1
ui ¨ gi,
§ A as elliptic operator,
§ with Dirichlet, Neumann, or Robin B.C.
§ Linear parabolic constraints, which extends the results of [CKK17].
Sebastian Engel Optimal Control and BV-Functions Rigorosum 15
International Research Training Group IGDK 1754
Error Rates for Variational Discretization
Error Rates for
Variational Discretization
Sebastian Engel Optimal Control and BV-Functions Rigorosum 16
International Research Training Group IGDK 1754
Error Rates BV-Control Problems
§ i[Bar12]: Image reconstruction by TV-based optimal control
(no PDE constraints).
§ linear continuous uh : }u ´ uh}L2p
q ď ch
1
6 .
§ i[CKK17]: Optimal control of a semi-linear parabolic equation.
§ One dimensional BV-controls, u;h cellwise constant.
}y ´ y;h}L2p
q ` |Jpuq ´ J;hpu;hq| ď cp
?
 ` hq
§ i[HMNV19]: Optimal control of one dimensional elliptic equations.
§ Variational discretization:
Optimal convergence results for state, adjoint, and control
(piecewise constant assumption).
Sebastian Engel Optimal Control and BV-Functions Rigorosum 17
International Research Training Group IGDK 1754
Standard Approach - Error Estimates
§ In case of smooth cost functions Jpuq, the standard approach uses
coercivity properties, e.g.
§ appropriate testing of the 1st
order optimality conditions, to derive error
estimates for the control.
§ Error rates for the optimal states, costs and TV-semi-norm of the optimal
controls can be obtained in a direct manner, which are sub-optimal.
§ For optimal error rates of the controls in the strict BV-topology, state, and
costs, we need several assumptions on the adjoint function p1.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 18
International Research Training Group IGDK 1754
Variational Discretization of p˜Pq
In the following we discretize the state equation by linear continuous FE in time
and space (S( b Sh), where controls will not be changed:
p˜Psemi
(Yh q
$
’
’%
min
v P Mp0YTqm
c P Rm
1
2
}S(YhpvYcq ´ yd}2
L2pT q `
mÿ
j“1
j
ż T
0
d|vj| “: J(YhpvYcq
Optimality Conditions
pÝÑv (YhYÝÑc (Yhq P Mp0YTqm ˆ Rm is the solution of p˜Psemi
(Yh q, if
´
ˆ
p1Yp(Yhq
p1Yp(Yhqp0q
˙
:“ ´
ˆ
p1Yp(YhqpÝÑv (YhYÝÑc (Yhq
p1Yp(YhqpÝÑv (YhYÝÑc (Yhqp0q
˙
P
ˆ`
iB}ÝÑv iY(Yh}Mp0YTq
˘m
i“1
0Rm
˙
§ }p1Y(Yh ´ p1}L8p0YTqm Ñ 0X
Sebastian Engel Optimal Control and BV-Functions Rigorosum 19
International Research Training Group IGDK 1754
Convergence Results - Standard Approach
 “ 2
3
: yd P C1
pI; L2
pqq and py0Yy1q P H1
0
pq ˆ L2
pq.
 “ 1: yd P C1
pI; H1
0
pqq, g P pHp2q
qm, and py0Yy1q P Hp3q
ˆ Hp2q
.
State Error Rates
}SpÝÑv YÝÑc q ´ S(YhpÝÑv (YhYÝÑc (Yhq}L2pT q P Op( ` hq
Cost Error Rates
|Jpuq ´ J(Yhpu(Yhq| P Op( ` hq Y
with  “  for  “ 2
3
and  “ 2 elseX
ˇ
ˇ
ˇ
ˇ}Dtu}Mp0YTqm ´ }Dtu(Yh}Mp0YTqm
ˇ
ˇ
ˇ
ˇ P O`
( ` h
˘
Y
with u(Yhptq “
ş
r0Yts dÝÑv (Yh ` ÝÑc (YhX
Sebastian Engel Optimal Control and BV-Functions Rigorosum 20
International Research Training Group IGDK 1754
Optimal Control with PDE constraints and Sparse Controls
§ What are sparse controls?
Sparsity
[Cas17]: BV-controls are sparse, if their distributional derivative is singular with
respect to the Lebesgue measure.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 21
International Research Training Group IGDK 1754
Optimality Conditions - Consequences
For the optimal control pÝÑv YÝÑc q of p˜Pq holds:
$

%
supppÝÑv ˘
i q Ă tt P r0YTs | p1Yiptq “ ¯i u
}p1Yi}C0pIq ď i
We have analogous results for pÝÑv (YhYh(Yhq.
§ If D :“ tp1Yi “ ˘iu is a nite set, we nd that u is piecewise constant, i.e.
uiptq “
ÿ
aPD
'a ¨ 1raYTsptq ` ÝÑc
§ In practice, we often observe piecewise constant controls.
§ In general, we cannot expect
piecewise constant controls u.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 22
International Research Training Group IGDK 1754
Optimal Convergence Results - Preliminaries
Assumptions
A1: t t P p0YTq | |p1Yiptq| “  u “ tt1YiY¨ ¨ ¨ Ytmi Yiu
for mi P Ną0, with i “ 1Y¨ ¨ ¨ Ym and mi P N.
A2: Bttp1ptjYiq ‰ 0, for i “ 1Y¨ ¨ ¨ Ym and j “ 1Y¨ ¨ ¨ Ymi.
§ Assumption A1 implies:
ui “
miř
–“1
ci
– 1rt`;i YTs ` ciY
whereby ci
– can be 0, if ui has no jump in t–Yi.
[HiDe] for an elliptic problem without L2
-regularization: |tp˚
pxq “ 0u| “ 0
Bang-Bang
control ´ Variational Discretization ´ }u´uh}L1pq ď ch2
|lnphq|
Sebastian Engel Optimal Control and BV-Functions Rigorosum 23
International Research Training Group IGDK 1754
Optimal Convergence Results - Preliminaries
Due to the structural assumptions A1, A2, we obtain:
Explicit Form
There exists a p(0Yh0q such that @p(Yhq “ 5 ď p(0Yh0q it holds
uptq “
¨
˚
˝
c1
...
cm
˛
‹
‚`
¨
˚
˚
˚
˚
˚
˝
m1ř
j“1
c1
j 1ptj;1YTsptq
...
mmř
j“1
cm
j 1ptj;mYTsptq
˛
‹
‹
‹
‹
‹
‚
Y u5ptq “
¨
˚
˝
c1Y5
...
cmY5
˛
‹
‚`
¨
˚
˚
˚
˚
˚
˝
m1ř
j“1
c1
jY51pt1
j;#YTsptq
...
mmř
j“1
cm
jY51ptm
j;#YTsptq
˛
‹
‹
‹
‹
‹
‚
Sebastian Engel Optimal Control and BV-Functions Rigorosum 24
International Research Training Group IGDK 1754
Optimal Convergence Results - Preliminaries
This implies the following estimate for the optimal controls:
L1
´ Estimate
For all p(Yhq “ 5 ď p(0Yh0q holds
}u ´ u5}L1p0YTqm ď rc
ˆ mř
i“1
|ci ´ ciY5| `
miř
j“1
|ci
j | ¨ |tjYi ´ ti
jY5| ` |ci
j ´ ci
jY5|
˙
Sebastian Engel Optimal Control and BV-Functions Rigorosum 25
International Research Training Group IGDK 1754
Optimal Convergence Results
We dene  “ 2
3
for pydYÝÑg Yy0Yy1q P C1
pI; L2
pqq ˆ L2
pqm ˆ H1
0
pq ˆ L2
pq
We eand  “ 1 for pydYÝÑg Yy0Yy1q P C1
pI; H1
0
pqq ˆ pHp2q
qm ˆ Hp3q
ˆ Hp2q
.
Amplitude |ci
j ´ ci
jY5|
Jump |tjYi ´ ti
jY5|
Constant |ci ´ ciY5|
,
////.
////-
“ Op( ` hq if  “ 2
3
Y
ď c
`
(2
` h2
` }Spuq ´ S5pu5q}L2pT q
˘
, if  “ 1X
Control ´ Error Rate:
This implies:
}u ´ u5}L1p0YTqm P Op( ` hq (suboptimal)X
Sebastian Engel Optimal Control and BV-Functions Rigorosum 26
International Research Training Group IGDK 1754
Optimal Convergence Results
In case of pydYÝÑg Yy0Yy1q P C1
pI; H1
0
pqq ˆ pHp2q
qm ˆ Hp3q
ˆ Hp2q
we obtain
(by Young inequality and control rates for  “ 1):
Optimal Control Error Rates
}u ´ u5}L1pIqm Y |ci ´ ciY5|Y
|tjYi ´ ti
jY5|Y |ci
j ´ ci
jY5|
,
.
-
“ Op(2
` h2
q
with i “ 1Y¨ ¨ ¨ Ym, j “ 1Y¨ ¨ ¨ Ymi,
Optimal State and Total Variation Error Rates
}Spuq ´ S5pu5q}L2pT q “ Op(2
` h2
q
ˇ
ˇ
ˇ
ˇ}Dtu}MpIq ´ }Dtu5}MpIq
ˇ
ˇ
ˇ
ˇ “ Op(2
` h2
q pBV-Strict Convergence!qX
Sebastian Engel Optimal Control and BV-Functions Rigorosum 27
International Research Training Group IGDK 1754
Optimal Control of Hyperbolic Equations
Additional Results
and Outlook
Sebastian Engel Optimal Control and BV-Functions Rigorosum 28
International Research Training Group IGDK 1754
Sound Propagation in an Urban Environment(rHhsrTtsrTtsrPpsrSss?|rFf srTtsrPpsq:TrueWave:swf (rHhsrTtsrTtsrPpsrSss?|rFf srTtsrPpsq:
TrueWave:swf
§ Hyperbolic Equation:
Btty ´ divpbpxqOyq “
řm
i“1
uiptq ¨ gipxq
§ Sound is emitted by acoustic speakers,
placed in gipxq
Sebastian Engel Optimal Control and BV-Functions Rigorosum 29
International Research Training Group IGDK 1754
Sound Propagation in an Urban Environment
Data of the microphone recordings can be used to obtain a possible position
using optimal control theory:
Optimal Control - Example of Penalization Costs
§ Examples of penalization terms, used to nd an optimal control, e.g.
Moving Microphone:
Tş
0
´
´
ş
Br pptqq yptqdx ´ ydataptq
¯2
dt
§ The solution of the following minimization problem leads to a possible
sound source position:
min
uPX
1
2
ř
iYj
´
´
ş
Br pxj q yptiqdx ´ ydata
¯2
` 
2
}u}X
where X “ BVp0YTqm or X “ H1
p0YTqm with
}u}BVp0YTqm “
mř
i“1
}Btui}Mp0YTq or }u}H1p0YTqm “
mř
i“1
}Btui}2
L2p0YTq ` }up0q}2
Rm .
Sebastian Engel Optimal Control and BV-Functions Rigorosum 30
International Research Training Group IGDK 1754
Sound Propagation in an Urban Environment(rHhsrTtsrTtsrPpsrSss?|rFf srTtsrPpsq:Comparision:swf (rHhsrTtsrTtsrPpsrSss?|rFf srTtsrPpsq:
Comparision:swf
i “ 0X1 for BV and H1
,
bpxq “
$

%
10´10
Y inside
0X35Y outside
r10´10
Y0X35sY else
Sebastian Engel Optimal Control and BV-Functions Rigorosum 31
International Research Training Group IGDK 1754
Bayesian Inversion with BV-Prior
Based on [EHMS19], we are able to consider a stochastic view on the sound
source identication problem before:
ydata “
´
´
ş
Br pxj q yuptiqdx
¯
i“1M
` 
with noise  „ Np0Y'q and stochastic control u “
kř
–“0
–1pt`YTs ` c „ 0
Bayesian Inversion
§ Well-posedness of the posterior solution
dypuq “ expp´1
2
}yu ´ ydata}2
¦qd0puq.
§ Convergence results for the SMC-Algorithm and FE-Discretization
§ Expectation, condence region, MAP-Estimator, ...
A similar Bayesian inversion problem is considered by [Yao16]
(TV-Gaussian Prior, Splitting pCN).
Sebastian Engel Optimal Control and BV-Functions Rigorosum 32
International Research Training Group IGDK 1754
Summary
We analyzed control problems with second-order hyperbolic equations, with
various types of cost functionals, which are regularized by a non-smooth
semi-norm of BVp0YTq.
§ Optimality conditions + sparsity + explicit examples.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
International Research Training Group IGDK 1754
Summary
We analyzed control problems with second-order hyperbolic equations, with
various types of cost functionals, which are regularized by a non-smooth
semi-norm of BVp0YTq.
§ Optimality conditions + sparsity + explicit examples.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
International Research Training Group IGDK 1754
Summary
We analyzed control problems with second-order hyperbolic equations, with
various types of cost functionals, which are regularized by a non-smooth
semi-norm of BVp0YTq.
§ Optimality conditions + sparsity + explicit examples.
§ Regularized problem pPq (H1
´norm)
+ convergence results (control, costs, adjoint)
Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
International Research Training Group IGDK 1754
Summary
We analyzed control problems with second-order hyperbolic equations, with
various types of cost functionals, which are regularized by a non-smooth
semi-norm of BVp0YTq.
§ Optimality conditions + sparsity + explicit examples.
§ Regularized problem pPq (H1
´norm)
+ convergence results (control, costs, adjoint)
Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
International Research Training Group IGDK 1754
Summary
We analyzed control problems with second-order hyperbolic equations, with
various types of cost functionals, which are regularized by a non-smooth
semi-norm of BVp0YTq.
§ Optimality conditions + sparsity + explicit examples.
§ Regularized problem pPq (H1
´norm)
+ convergence results (control, costs, adjoint)
§ Numerical analysis of a semismooth Newton algorithm for pPq
+ super-linear convergence.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
International Research Training Group IGDK 1754
Summary
We analyzed control problems with second-order hyperbolic equations, with
various types of cost functionals, which are regularized by a non-smooth
semi-norm of BVp0YTq.
§ Optimality conditions + sparsity + explicit examples.
§ Regularized problem pPq (H1
´norm)
+ convergence results (control, costs, adjoint)
§ Numerical analysis of a semismooth Newton algorithm for pPq
+ super-linear convergence.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
International Research Training Group IGDK 1754
Summary
We analyzed control problems with second-order hyperbolic equations, with
various types of cost functionals, which are regularized by a non-smooth
semi-norm of BVp0YTq.
§ Optimality conditions + sparsity + explicit examples.
§ Regularized problem pPq (H1
´norm)
+ convergence results (control, costs, adjoint)
§ Numerical analysis of a semismooth Newton algorithm for pPq
+ super-linear convergence.
§ Error estimates for a variational discretization
+ optimal rates under specic assumptions (sparsity in pPq).
Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
International Research Training Group IGDK 1754
Summary
We analyzed control problems with second-order hyperbolic equations, with
various types of cost functionals, which are regularized by a non-smooth
semi-norm of BVp0YTq.
§ Optimality conditions + sparsity + explicit examples.
§ Regularized problem pPq (H1
´norm)
+ convergence results (control, costs, adjoint)
§ Numerical analysis of a semismooth Newton algorithm for pPq
+ super-linear convergence.
§ Error estimates for a variational discretization
+ optimal rates under specic assumptions (sparsity in pPq).
Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
International Research Training Group IGDK 1754
Summary
We analyzed control problems with second-order hyperbolic equations, with
various types of cost functionals, which are regularized by a non-smooth
semi-norm of BVp0YTq.
§ Optimality conditions + sparsity + explicit examples.
§ Regularized problem pPq (H1
´norm)
+ convergence results (control, costs, adjoint)
§ Numerical analysis of a semismooth Newton algorithm for pPq
+ super-linear convergence.
§ Error estimates for a variational discretization
+ optimal rates under specic assumptions (sparsity in pPq).
§ A Bayesian perspective on pPq.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
International Research Training Group IGDK 1754
Literature
[Bar12], S. Bartels, Total variation minimization with nite elements:
convergence and iterative solution. SIAM J. Numer. Anal.,
50(3):1162-1180, 2012.
[Bog98], V. I. Bogachev, Gaussian measures, vol.62, Mathematical Surveys
and Monographs, American Mathematical Society, 1998.
[Cas17], E. Casas, A review on sparse solutions in optimal control of partial
dierential equations. SEMA Journal, 74, pp. 319-344, 2017.
[CK17], E. Casas, K. Kunisch, Analysis of optimal control problems of
semilinear elliptic equations by bv-functions. Set-Valued and Variational
Analysis, 1-25, 2017.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 34
International Research Training Group IGDK 1754
Literature
[CKK17], E. Casas, K. Kunisch, and F. Kruse. Optimal control of
semilinear parabolic equations by bv-functions. SIAM Journal on Control
and Optimization, 55:1752-1788, 2017.
[CK11], C. Clason and K. Kunisch, A duality based approach to elliptic
control problems in non refelxive banach spaces. ESIAM Control Optim.
Calc. Var., 17, pp. 243-266, 2011.
[HiDe], K. Deckelnick, M. Hinze, A note on the approximation of ellliptic
control problems with bang-bang controls. Comput. Optim. Appl.,
51:931-939, 2010.
[EHMS19], S. Engel, D. Hafemeyer, C. Münch, and D. Schaden. An
application of sparse measure valued Bayesian inversion to acoustic sound
source identication. Inverse Problems, accepted.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 35
International Research Training Group IGDK 1754
Literature
[HMNV19], D. Hafemeyer, F. Mannel, I. Neitzel, B. Vexler, Finite element
error estimates for elliptic optimal control by BV functions, arxiv, 2019.
[Hor16], M. Hornikx, Ten questions concerning computational urban
acoustics. Building and Environment 106 : 409-421, 2016.
[KTV16], K. Kunisch, P. Trautmann, B. Vexler, Optimal control of the
undamped linear wave equation with measure valued controls. SIAM
Journal on Control and Optimization, 54(3), 1212-1244, 2016.
[KKV10], A. Kröner, K. Kunisch, B. Vexler, Semismooth Newton Methods
for an Optimal Boundary Control Problem of Wave Equations, in: Recent
Advances in Optimization and its Applications in Engineering, M. Diehl, F.
Glineur, E. Jarlebring, W. Michiels (eds.), Springer, 389-398, 2010.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 36
International Research Training Group IGDK 1754
Literature
[KKV11], A. Kröner, K. Kunisch, B. Vexler, Semismooth Newton Methods
for Optimal Control of the Wave Equation with Control Constraints, SIAM
J.Control Optim. 49: 830-858, 2011.
[Lion], J. L. Lions: Optimal Control of Systems Governed by Partial
Dierential Equations. Springer-Verlag, Berlin, 1971.
[MR04], B. S. Mordukhovich, J. P. Raymond, Neumann boundary control
of hyperbolic equations with pointwise state constraints. SIAM journal on
control and optimization, 43(4), 1354-1372, 2004.
[Pi15], K. Pieper, Finite element discretization and ecient numerical
solution of elliptic and parabolic sparse control problems. PhD thesis,
Technical University Munich, 2015.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 37
International Research Training Group IGDK 1754
Literature
[Sta09], G. Stadler, Elliptic optimal control problems with L1
-control cost
and applications for the placement of control devices. Comput. Optim.
Appl., 44(2):159-181, 2009.
[TVZ17], P. Trautmann, B. Vexler, A. Zlotnik, Finite element error analysis
for measure-valued optimal control problems governed by a 1D wave
equation with variable coecients, arXiv , 2017.
[Yao16], Z. Yao, Z. Hu, and J. Li. A tv-gaussian prior for
innite-dimensional bayesian inverse problems and its numerical
implementations. IOPscience, 2016.
[Zlo94], A. A. Zlotnik, Convergence rate estimates of nite-element
methods for second-order hyperbolic equations. numerical methods and
applications, p.153 et.seq. Guri I. Marchuk, CRC Press, 1994.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 38
International Research Training Group IGDK 1754
Appendix
Variational Discretization
Sebastian Engel Optimal Control and BV-Functions Rigorosum 39
International Research Training Group IGDK 1754
Optimal Control of the Wave Equation with BV-Functions
Consider the following semi-discretized optimal control problem:
pPsemi
(Yh q
$
’’’’
’’’’%
min
uPBVp0YTqm
1
2
}yuY(Yh ´ yd}2
L2pT q `
mÿ
j“1
j
ż
r0YTs
d|Dtuj|ptq “: J(Yhpuq
with }y ´ y(Yh}Cp0YT;L2pqq P Op(2
3 ` h
2
3 qY for pFEM Zlotnikq
 f “forcing y0 “displacement y1 “velocity
3
C1
pr0YTs; H1
0
pqq or
BVp0YT; H2
q
H3
H2
2 C1
pr0YTs; L2
pqq H2
H1
0
pq
1 L2
pTq H1
0
pq L2
pq
§ H “

w P L2
pq
ˇ
ˇ
ˇ
ˇ
ř
kě1
!
k xwYky2
L2pq ă 8
*
with eigenvalues and
eigenfunctions p!kYkqkě1 of ´4 with homogeneous Dirichlet boundary
conditions.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 40
International Research Training Group IGDK 1754
Optimal Convergence Results - Preliminaries
Due to the structural assumptions A1 and A2, we obtain:
Lemma
There exists a  ą 0, and (0Yh0 ą 0 such that
i
@p(Yhq “ 5 ď p(0Yh0q holds:
ÝÑv 5Yi “
miř
l“1
ci
lY5ti
l;#
with ti
jY5 P BptjYiq
where pBptjYiqqmi
j“1
are pairwise disjoint for a xed i “ 1Y¨ ¨ ¨ Ym with 0 ă 5 Ñ 0
Sebastian Engel Optimal Control and BV-Functions Rigorosum 41
International Research Training Group IGDK 1754
Optimal Convergence Result
Optimal Rates:
1st-Step, State Dependence:
}u ´ u5}L1pIqm ď ˆcp(2
` h2
` }Spuq ´ S5pu5q}L2pT qq
2nd-Step, Scaled Young Inequality:
}Spuq ´ S5pu5q}L2pT q
ďljhn
FOOC
#
c}Spuq ´ S5puq}L2pT q
`cpÝÑg q}u5 ´ u}
1
2
L1pIqm }L˚
pSpuq ´ ydq ´ L˚
5pSpuq ´ ydq}
1
2
L8p0YT;L2pqq
+
ďljhn
Zlotnik
Young Ineq.Y ą 0
#
cp(2
` h2
q ` ˆc}Spuq ´ S5puq}L2pT q
`cp
ÝÑg q
4 }L˚
pSpuq ´ ydq ´ L˚
5pSpuq ´ ydq}L8p0YT;L2pqq
+
Sebastian Engel Optimal Control and BV-Functions Rigorosum 42
International Research Training Group IGDK 1754
Optimal Convergence Results - Consequences
Let the assumptions A1 and A2 hold:
Properties of u5:
Consider an amplitude ci
jY5 of an optimal control of pPsemi
5 q, with i “ 1Y¨ ¨ ¨ Ym,
and j “ 1Y¨ ¨ ¨ Ymi:
a) Assume that |ci
j | ą 0. The optimal control of pPsemi
5 q has a jump
(|ci
jY5| ą 0) inside BptjYiq for all 0 ă 5 ă 50 and 50 small enough.
b) Assume that ci
j “ 0. The optimal control of pPsemi
5 q can have a jump in
BptjYiq for all 0 ă 5 ă 50, with 50 small enough, but the jump height has
to decrease with some specic rate pci
jY5 P Op( ` hqq.
c) Let tt P I||p1Yiptq| “ u “ H for i “ 1Y¨ ¨ ¨ Ym, and m ě 1. Then the
optimal control sub function u5Yi of pPsemi
5 q has no jumps pu5Yi “ const.q for
all 0 ă 5 ď 50.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 43
International Research Training Group IGDK 1754
Appendix
More General Costs
Sebastian Engel Optimal Control and BV-Functions Rigorosum 44
International Research Training Group IGDK 1754
Control Problem
pP¥q
$
’’’’’’’’’
’’’’’’’’’%
min
uPBVp0YTqm
3ř
j“1
j
2
} r¥jpyuq ´ zj}2
Oj
`
mř
–“1
–}Dtu–}Mp0YTq “: J¥puq
s.t.
$
’’’
’’’%
lyu :“ Bttyu ´ 4yu “
mÿ
j“1
ujgj in p0YTq ˆ 
y “ 0 on p0YTq ˆ B
pyp0qYBtyp0qq “ py0Yy1q P H1
0
pq ˆ L2
pq
§ zj P Oj sep. Hilbert spaces and pg–qm
– Ă L8
pqzt0u with pairwise disjoint
supports.
r¥2pyuq :“ ¥2pyupt–qqr
–“1
Y
r¥3pyuq :“ ¥3pBtyupt–qqr
–“1
Y
0 “ t0 ă t1 ă ¨ ¨ ¨ ă tr “ TY
Lin. indep. p r¥j pyg`qqm
`“1
for one j ą0
with
$

%
r¥1 P Lp L2
pTq Y O1qY
¥2 P Lp L2
pqr Y O2qY
¥3 P Lp H´1
pqr Y O3qX
This convex problem has a solution.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 45
International Research Training Group IGDK 1754
Equivalent Problem and Optimality Conditions
prP¥q
$
’’’
’’’%
min
pvYcqPMp0YTqmˆRm
J¥puq :“ rJ¥pvYcq
with uptq “
ˆtş
0
dvjpsq ` cj
˙m
i“1
resp. pDtuYup0qq “ pvYcqX
The rst-order optimality condition has the following form:
ˆ
p¥
1
ptq
p¥
1
p0q
˙
“
¨
˝´
Tş
t
ş

ˆppÝÑv YÝÑc qÝÑg
´p¥
1
p0q
˛
‚P
ˆ
piB}pÝÑv i}Mqqi
0
˙
text
with the jumping
wave function
ˆppÝÑv YÝÑc q :“
řr
i“1
pi1Ii
with Ii :“ pti´1Ytiq
l pi “ gi in Ii ˆ Y piptiq “ pi`1ptiq ` hiY Btpiptiq “ Btpi`1ptiq ` ki
gi “ 1¥˚
1
p¥1pyuq ´ z1q|Ii ˆ P L2
pIi ˆ qY
hi “ 3p´4q´1
”
¥˚
3
p r¥3pyuq ´ z3q
ı
i
P H1
0
pqY
ki “ ´2
”
¥˚
2
p r¥2pyuq ´ z2q
ı
i
P L2
pqX
Sebastian Engel Optimal Control and BV-Functions Rigorosum 46
International Research Training Group IGDK 1754
Regularized Control Problem
prP¥
 q
#
min
pvYcqPL2p0YTqmˆRm
rJ¥pvYcq ` 
2
˜
mř
j“1
}vj}2
L2p0YTq ` }c}Rm
¸
For the BV-representation u of the optimal control pvYcq holds u
strict
ÝÝÝÑ
BV
u.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 47
International Research Training Group IGDK 1754
Regularized Control Problem - Jumping Wave Example
Let 1 “ 3 “ 0, and consider –, – “ 1Y¨ ¨ ¨ Yk, as pairwise disj. balls inside :
r¥2pyuq :“
ˆ´
1
|`|
ş
`
yuptiqdx
¯k
–“1
˙r
i“1
P Rk¨r p„ Microphones)
One experimental result with ÝÑ
t “ p0X098Y0X199Y0X297Y0X347Y0X398Y1q:
Sebastian Engel Optimal Control and BV-Functions Rigorosum 48
International Research Training Group IGDK 1754
Appendix
Bayesian Inversion
Sebastian Engel Optimal Control and BV-Functions Rigorosum 49
International Research Training Group IGDK 1754
Bayesian Inversion and pP¥q - Specic Example
Consider the following stochastic problem:
pI¥q z2 “ r¥2pyuq ` Y  „ Np0Y 1
2
¨ idRk¨r qY
where r¥2 is dened as in (Microphones) and u K .
Sebastian Engel Optimal Control and BV-Functions Rigorosum 50
International Research Training Group IGDK 1754
Bayesian Inversion and pP¥q - Specic Example
Consider the following stochastic problem:
pI¥q z2 “ r¥2pyuq ` Y  „ Np0Y 1
2
¨ idRk¨r qY
where r¥2 is dened as in (Microphones) and u K .
§ Without any knowledge of z2, we assume that u „ 0 (Prior)
Ñ How should 0 be dened? Gaussian?
Sebastian Engel Optimal Control and BV-Functions Rigorosum 50
International Research Training Group IGDK 1754
Bayesian Inversion and pP¥q - Specic Example
Consider the following stochastic problem:
pI¥q z2 “ r¥2pyuq ` Y  „ Np0Y 1
2
¨ idRk¨r qY
where r¥2 is dened as in (Microphones) and u K .
§ Without any knowledge of z2, we assume that u „ 0 (Prior)
Ñ How should 0 be dened? Gaussian?
§ We consider ˜u „ Np0Y 1
 Iq “  with:
$

%
I P LpL2
pIqq trace class
rgpI1{2
q “ H1
p0YTqY
(see [Bog98,EX 2.3.4]).
Sebastian Engel Optimal Control and BV-Functions Rigorosum 50
International Research Training Group IGDK 1754
Bayesian Inversion and pP¥q - Specic Example
Consider the following stochastic problem:
pI¥q z2 “ r¥2pyuq ` Y  „ Np0Y 1
2
¨ idRk¨r qY
where r¥2 is dened as in (Microphones) and u K .
§ Without any knowledge of z2, we assume that u „ 0 (Prior)
Ñ How should 0 be dened? Gaussian?
§ We consider ˜u „ Np0Y 1
 Iq “  with:
$

%
I P LpL2
pIqq trace class
rgpI1{2
q “ H1
p0YTqY
(see [Bog98,EX 2.3.4]).
§ We weight  by f pwq “ 1
£0
expp´}Btw}Mq
Ñ concentration to functions with small total variation, i.e.
0p¨q “
ş
¨
f pwqdpwq (see [Yao18]).
Sebastian Engel Optimal Control and BV-Functions Rigorosum 50
International Research Training Group IGDK 1754
Bayesian Inversion and pP¥q - Specic Example
§ We have u|z2 „ z, with zp¨q “ 1
£z
ş
¨
expp´2
2
}Ă¥2pyuq ´ z2}2
Rk¨r qd0puq
(Posterior).
Sebastian Engel Optimal Control and BV-Functions Rigorosum 51
International Research Training Group IGDK 1754
Bayesian Inversion and pP¥q - Specic Example
§ We have u|z2 „ z, with zp¨q “ 1
£z
ş
¨
expp´2
2
}Ă¥2pyuq ´ z2}2
Rk¨r qd0puq
(Posterior).
§ Maximum a Posteriori Estimator:
min
uPH1p0YTq
2
2
} r¥2pyuq ´ z2}2
Rk¨m ` }Btu}M ` p}Btu}2
L2 ` }u}2
L2 q
Sebastian Engel Optimal Control and BV-Functions Rigorosum 51
International Research Training Group IGDK 1754
Bayesian Inversion and pP¥q - Specic Example
§ We have u|z2 „ z, with zp¨q “ 1
£z
ş
¨
expp´2
2
}Ă¥2pyuq ´ z2}2
Rk¨r qd0puq
(Posterior).
§ Maximum a Posteriori Estimator:
min
uPH1p0YTq
2
2
} r¥2pyuq ´ z2}2
Rk¨m ` }Btu}M ` p}Btu}2
L2 ` }u}2
L2 q
§ Sampling and Empirical MAP (Splitting pCN with 100,000 sample steps):
We used ˜uptq “
8ř
–“0
Np0Y!2
– q ¨ 9–ptq „ Np0Yp´1
 40q´1
q,
with MAP:
min
uPH1
0
p0YTq
2
2
} r¥2pyuq ´ z2}2
Rk¨m
`}Btu}M ` }Btu}2
L2
Sebastian Engel Optimal Control and BV-Functions Rigorosum 51
International Research Training Group IGDK 1754
Bayesian Inversion and pP¥q - Specic Example
§ We have u|z2 „ z, with zp¨q “ 1
£z
ş
¨
expp´2
2
}Ă¥2pyuq ´ z2}2
Rk¨r qd0puq
(Posterior).
§ Maximum a Posteriori Estimator:
min
uPH1p0YTq
2
2
} r¥2pyuq ´ z2}2
Rk¨m ` }Btu}M ` p}Btu}2
L2 ` }u}2
L2 q
§ Sampling and Empirical MAP (Splitting pCN with 100,000 sample steps):
We used ˜uptq “
8ř
–“0
Np0Y!2
– q ¨ 9–ptq „ Np0Yp´1
 40q´1
q,
with MAP:
min
uPH1
0
p0YTq
2
2
} r¥2pyuq ´ z2}2
Rk¨m
`}Btu}M ` }Btu}2
L2
§ Disadvantages (true function is sparse): Computationally expensive (KLE),
high number of samples needed (lack of sparsity), good initial value needed.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 51
International Research Training Group IGDK 1754
BV-Prior
§ We now assume that u „
kř
–“0
–1pt`YTsptq ` c („ 0 BV-Prior).
Sebastian Engel Optimal Control and BV-Functions Rigorosum 52
International Research Training Group IGDK 1754
BV-Prior
§ We now assume that u „
kř
–“0
–1pt`YTsptq ` c („ 0 BV-Prior).
§ The set of all realizations of u is dense in BVp0YTq, in the strict BV-topo.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 52
International Research Training Group IGDK 1754
BV-Prior
§ We now assume that u „
kř
–“0
–1pt`YTsptq ` c („ 0 BV-Prior).
§ The set of all realizations of u is dense in BVp0YTq, in the strict BV-topo.
§
Sampling + Empirical MAP +
Optimal Control pP¥q
(SMC 5,000 samples):
Sebastian Engel Optimal Control and BV-Functions Rigorosum 52
International Research Training Group IGDK 1754
BV-Prior
§ We now assume that u „
kř
–“0
–1pt`YTsptq ` c („ 0 BV-Prior).
§ The set of all realizations of u is dense in BVp0YTq, in the strict BV-topo.
§
Sampling + Empirical MAP +
Optimal Control pP¥q
(SMC 5,000 samples):
§ Advantage: Faster algorithms by sparsity.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 52
International Research Training Group IGDK 1754
Gaussian Measure
[Bog98, Theorem 3.6.1]
Let  be a Radon Gaussian measure on a locally convex space X. Then the
topological support of  coincides with the ane subspace Epq ` H
X
, where
H
X
stands for the closure in X.
Sebastian Engel Optimal Control and BV-Functions Rigorosum 53
International Research Training Group IGDK 1754
Thank you for your attention
Supported by the DFG through the International Research Training Group IGDK
1754 Optimization and Numerical Analysis for Partial Dierential Equations
with Nonsmooth Structures
Sebastian Engel Optimal Control and BV-Functions Rigorosum 54

More Related Content

Similar to Optimal Control for Linear Second-Order Hyperbolic Equations by BV-Functions in Time, KFU Graz 2019

POSTER [Compatibility Mode]2
POSTER [Compatibility Mode]2POSTER [Compatibility Mode]2
POSTER [Compatibility Mode]2
AMIT PATHAK
 
Constraint optimal control
Constraint optimal controlConstraint optimal control
Constraint optimal control
Rodrigue Tchamna
 
Presentation FOPID Boost DC-DC Converter.pptx
Presentation FOPID Boost DC-DC Converter.pptxPresentation FOPID Boost DC-DC Converter.pptx
Presentation FOPID Boost DC-DC Converter.pptx
SherAli260123
 
buoyantBousinessqSimpleFoam
buoyantBousinessqSimpleFoambuoyantBousinessqSimpleFoam
buoyantBousinessqSimpleFoam
Milad Sm
 

Similar to Optimal Control for Linear Second-Order Hyperbolic Equations by BV-Functions in Time, KFU Graz 2019 (20)

first law of thermodynamics and second law
first law of thermodynamics and second lawfirst law of thermodynamics and second law
first law of thermodynamics and second law
 
1413570.ppt
1413570.ppt1413570.ppt
1413570.ppt
 
Optimum Solution of Quadratic Programming Problem: By Wolfe’s Modified Simple...
Optimum Solution of Quadratic Programming Problem: By Wolfe’s Modified Simple...Optimum Solution of Quadratic Programming Problem: By Wolfe’s Modified Simple...
Optimum Solution of Quadratic Programming Problem: By Wolfe’s Modified Simple...
 
Tuning of PID, SVFB and LQ Controllers Using Genetic Algorithms
Tuning of PID, SVFB and LQ Controllers Using Genetic AlgorithmsTuning of PID, SVFB and LQ Controllers Using Genetic Algorithms
Tuning of PID, SVFB and LQ Controllers Using Genetic Algorithms
 
Chebyshev Inequality
Chebyshev InequalityChebyshev Inequality
Chebyshev Inequality
 
Unit 1.2 thm
Unit 1.2 thmUnit 1.2 thm
Unit 1.2 thm
 
Db36619623
Db36619623Db36619623
Db36619623
 
POSTER [Compatibility Mode]2
POSTER [Compatibility Mode]2POSTER [Compatibility Mode]2
POSTER [Compatibility Mode]2
 
Min Max Model Predictive Control for Polysolenoid Linear Motor
Min Max Model Predictive Control for Polysolenoid Linear MotorMin Max Model Predictive Control for Polysolenoid Linear Motor
Min Max Model Predictive Control for Polysolenoid Linear Motor
 
A New Adaptive PID Controller
A New Adaptive PID ControllerA New Adaptive PID Controller
A New Adaptive PID Controller
 
G010525868
G010525868G010525868
G010525868
 
Constraint optimal control
Constraint optimal controlConstraint optimal control
Constraint optimal control
 
TOCbw I&ECPDD Oct67
TOCbw I&ECPDD Oct67TOCbw I&ECPDD Oct67
TOCbw I&ECPDD Oct67
 
Presentation FOPID Boost DC-DC Converter.pptx
Presentation FOPID Boost DC-DC Converter.pptxPresentation FOPID Boost DC-DC Converter.pptx
Presentation FOPID Boost DC-DC Converter.pptx
 
Ch16_1_27_05.ppt
Ch16_1_27_05.pptCh16_1_27_05.ppt
Ch16_1_27_05.ppt
 
buoyantBousinessqSimpleFoam
buoyantBousinessqSimpleFoambuoyantBousinessqSimpleFoam
buoyantBousinessqSimpleFoam
 
5977026.ppt
5977026.ppt5977026.ppt
5977026.ppt
 
fluid mechanics
fluid mechanicsfluid mechanics
fluid mechanics
 
70
7070
70
 
Controller design of inverted pendulum using pole placement and lqr
Controller design of inverted pendulum using pole placement and lqrController design of inverted pendulum using pole placement and lqr
Controller design of inverted pendulum using pole placement and lqr
 

Recently uploaded

Electricity and Circuits for Grade 9 students
Electricity and Circuits for Grade 9 studentsElectricity and Circuits for Grade 9 students
Electricity and Circuits for Grade 9 students
levieagacer
 
Nanoparticles for the Treatment of Alzheimer’s Disease_102718.pptx
Nanoparticles for the Treatment of Alzheimer’s Disease_102718.pptxNanoparticles for the Treatment of Alzheimer’s Disease_102718.pptx
Nanoparticles for the Treatment of Alzheimer’s Disease_102718.pptx
ssusera4ec7b
 
Warming the earth and the atmosphere.pptx
Warming the earth and the atmosphere.pptxWarming the earth and the atmosphere.pptx
Warming the earth and the atmosphere.pptx
GlendelCaroz
 

Recently uploaded (20)

NuGOweek 2024 programme final FLYER short.pdf
NuGOweek 2024 programme final FLYER short.pdfNuGOweek 2024 programme final FLYER short.pdf
NuGOweek 2024 programme final FLYER short.pdf
 
Heads-Up Multitasker: CHI 2024 Presentation.pdf
Heads-Up Multitasker: CHI 2024 Presentation.pdfHeads-Up Multitasker: CHI 2024 Presentation.pdf
Heads-Up Multitasker: CHI 2024 Presentation.pdf
 
GBSN - Microbiology (Unit 4) Concept of Asepsis
GBSN - Microbiology (Unit 4) Concept of AsepsisGBSN - Microbiology (Unit 4) Concept of Asepsis
GBSN - Microbiology (Unit 4) Concept of Asepsis
 
GBSN - Biochemistry (Unit 8) Enzymology
GBSN - Biochemistry (Unit 8) EnzymologyGBSN - Biochemistry (Unit 8) Enzymology
GBSN - Biochemistry (Unit 8) Enzymology
 
Electricity and Circuits for Grade 9 students
Electricity and Circuits for Grade 9 studentsElectricity and Circuits for Grade 9 students
Electricity and Circuits for Grade 9 students
 
Taphonomy and Quality of the Fossil Record
Taphonomy and Quality of the  Fossil RecordTaphonomy and Quality of the  Fossil Record
Taphonomy and Quality of the Fossil Record
 
ABHISHEK ANTIBIOTICS PPT MICROBIOLOGY // USES OF ANTIOBIOTICS TYPES OF ANTIB...
ABHISHEK ANTIBIOTICS PPT MICROBIOLOGY  // USES OF ANTIOBIOTICS TYPES OF ANTIB...ABHISHEK ANTIBIOTICS PPT MICROBIOLOGY  // USES OF ANTIOBIOTICS TYPES OF ANTIB...
ABHISHEK ANTIBIOTICS PPT MICROBIOLOGY // USES OF ANTIOBIOTICS TYPES OF ANTIB...
 
SaffronCrocusGenomicsThessalonikiOnlineMay2024TalkOnline.pptx
SaffronCrocusGenomicsThessalonikiOnlineMay2024TalkOnline.pptxSaffronCrocusGenomicsThessalonikiOnlineMay2024TalkOnline.pptx
SaffronCrocusGenomicsThessalonikiOnlineMay2024TalkOnline.pptx
 
VILLAGE ATTACHMENT For rural agriculture PPT.pptx
VILLAGE ATTACHMENT For rural agriculture  PPT.pptxVILLAGE ATTACHMENT For rural agriculture  PPT.pptx
VILLAGE ATTACHMENT For rural agriculture PPT.pptx
 
A Scientific PowerPoint on Albert Einstein
A Scientific PowerPoint on Albert EinsteinA Scientific PowerPoint on Albert Einstein
A Scientific PowerPoint on Albert Einstein
 
Factor Causing low production and physiology of mamary Gland
Factor Causing low production and physiology of mamary GlandFactor Causing low production and physiology of mamary Gland
Factor Causing low production and physiology of mamary Gland
 
Manganese‐RichSandstonesasanIndicatorofAncientOxic LakeWaterConditionsinGale...
Manganese‐RichSandstonesasanIndicatorofAncientOxic  LakeWaterConditionsinGale...Manganese‐RichSandstonesasanIndicatorofAncientOxic  LakeWaterConditionsinGale...
Manganese‐RichSandstonesasanIndicatorofAncientOxic LakeWaterConditionsinGale...
 
Fun for mover student's book- English book for teaching.pdf
Fun for mover student's book- English book for teaching.pdfFun for mover student's book- English book for teaching.pdf
Fun for mover student's book- English book for teaching.pdf
 
Adaptive Restore algorithm & importance Monte Carlo
Adaptive Restore algorithm & importance Monte CarloAdaptive Restore algorithm & importance Monte Carlo
Adaptive Restore algorithm & importance Monte Carlo
 
X-rays from a Central “Exhaust Vent” of the Galactic Center Chimney
X-rays from a Central “Exhaust Vent” of the Galactic Center ChimneyX-rays from a Central “Exhaust Vent” of the Galactic Center Chimney
X-rays from a Central “Exhaust Vent” of the Galactic Center Chimney
 
TEST BANK for Organic Chemistry 6th Edition.pdf
TEST BANK for Organic Chemistry 6th Edition.pdfTEST BANK for Organic Chemistry 6th Edition.pdf
TEST BANK for Organic Chemistry 6th Edition.pdf
 
Nanoparticles for the Treatment of Alzheimer’s Disease_102718.pptx
Nanoparticles for the Treatment of Alzheimer’s Disease_102718.pptxNanoparticles for the Treatment of Alzheimer’s Disease_102718.pptx
Nanoparticles for the Treatment of Alzheimer’s Disease_102718.pptx
 
GBSN - Microbiology (Unit 5) Concept of isolation
GBSN - Microbiology (Unit 5) Concept of isolationGBSN - Microbiology (Unit 5) Concept of isolation
GBSN - Microbiology (Unit 5) Concept of isolation
 
EU START PROJECT. START-Newsletter_Issue_4.pdf
EU START PROJECT. START-Newsletter_Issue_4.pdfEU START PROJECT. START-Newsletter_Issue_4.pdf
EU START PROJECT. START-Newsletter_Issue_4.pdf
 
Warming the earth and the atmosphere.pptx
Warming the earth and the atmosphere.pptxWarming the earth and the atmosphere.pptx
Warming the earth and the atmosphere.pptx
 

Optimal Control for Linear Second-Order Hyperbolic Equations by BV-Functions in Time, KFU Graz 2019

  • 1. International Research Training Group IGDK 1754 Optimal Control for Linear Second-Order Hyperbolic Equations by BV-Functions in Time Sebastian Engel Rigorosum 01.03.2019 Sebastian Engel Optimal Control and BV-Functions Rigorosum 1
  • 2. International Research Training Group IGDK 1754 Optimal Control of the Wave Equation with BV-Functions We focus on optimal control of hyperbolic partial dierential equations: Figure: [Hor16] Sebastian Engel Optimal Control and BV-Functions Rigorosum 2
  • 3. International Research Training Group IGDK 1754 Background - Related Work § Optimization with 2nd order linear hyperbolic equations: § Lions (Hilbert control). and Kröner et al. (Distributed, Dirichlet, Neumann L2 -control). Kunisch et al., and Trautmann et al.(Mp ; L2 p0;Tqq, L2 w˚ p0;T; Mp qq-control). § Thesis: 2nd order linear hyperbolic equations with BV(0,T)-controls. § FE with hyperbolic equations in optimal control: § Kröner et al. lin. cont. FE for state and control discretization. Trautmann et al. lin. cont. state, variational discretization control. § Thesis: Variational discretization of u, lin. cont. FE for Dtu and state. § Hyperbolic equations and sparse controls § Kunisch et al., and Trautmann et al.. § Thesis: Sparse BV-controls. § Optimization with BV-controls and other PDE constraints § Casas, Kunisch (semilinear elliptic, BVp q). Casas, Kruse, Kunisch (semilinear parabolic, BVp0;Tq-controls). § Thesis: 2nd order Hyperbolic equations with BVp0;Tq-controls Sebastian Engel Optimal Control and BV-Functions Rigorosum 3
  • 4. International Research Training Group IGDK 1754 Optimal Control of the Wave Equation with BV-Functions pPq $ ’’’’’’’’’’ ’’’’’’’’’’% min uPBVp0YTqm 1 2 ż ˆr0YTs pyu ´ ydq2 dxdt ` mÿ j“1 j ż r0YTs d|Dtuj|ptq “: Jpuq s.t. $ ’’’ ’’’% Btty ´ 4y “ mÿ j“1 ujgj in p0YTq ˆ y “ 0 on p0YTq ˆ B pyYBtyq “ py0Yy1q in t0u ˆ § Ă Rn (n=1,2,3) open bounded, B Lipschitz, T P p0Y8q § yd P W1Y1 pr0YTs; L2 pqq, py0Yy1q P H1 0 pq ˆ L2 pq § pgjqm j Ă L8 pqzt0u pairwise disjoint supports wj This strictly convex problem has a unique solution. Clason, Kunisch [CK11] § Penalization by the } ¨ }BV -norm corresponds to settings where the cost is proportional to changes in the control. - On/O structure, less blurring (H1 ´control). Sebastian Engel Optimal Control and BV-Functions Rigorosum 4
  • 5. International Research Training Group IGDK 1754 Equivalent Problem p˜Pq Consider the following equivalent optimal control problem w.r.t. pPq: p˜Pq # min pvYcqPMp0YTqmˆRm 1 2 }SpvYcq ´ yd}2 L2pT q ` mÿ j“1 j ż T 0 |vj|dx “: JpvYcq with uptq “ ˆtş 0 dvjpsq ` cj ˙m i“1 resp. ˆ Dtu up0q ˙ “ ˆ v c ˙ ˆ Fundamental theorem of calculus ˙ . § S is the ane control-to-state operator. § BVp0YTqm – Mp0YTqm ˆ Rm only possible in one dim. Sebastian Engel Optimal Control and BV-Functions Rigorosum 5
  • 6. International Research Training Group IGDK 1754 Optimality Conditions Adjoint Wave Equation We dene by p˚ phq the solution of the adjoint wave equation: $ % Bttp˚ ´ 4p˚ “ h in p0YTq ˆ p˚ “ 0 on p0YTq ˆ B pp˚ YBtp˚ q “ p0Y0q in tTu ˆ Consider the following integrated adjoint functions for j=1,...,m: p1YjpvYcqptq “ Tż t ż wj p˚ pSpvYcq ´ ydq gjdxdsY pvYcq P Mp0YTqm ˆ Rm Sebastian Engel Optimal Control and BV-Functions Rigorosum 6
  • 7. International Research Training Group IGDK 1754 Optimality Conditions Theorem (Necessary and Sucient Condition) pÝÑv YÝÑc q P Mp0YTqm ˆ Rm is the solution of p˜Pq if, for all j “ 1YXXXYm holds ´ ˆ p1pÝÑv YÝÑc q p1pÝÑv YÝÑc qp0q ˙ P ˆ` iB}ÝÑv i}Mp0YTq ˘m i“1 0Rm ˙ Y with p1pÝÑv YÝÑc q P C0pr0YTsq X C2 pr0YTsq. Sebastian Engel Optimal Control and BV-Functions Rigorosum 7
  • 8. International Research Training Group IGDK 1754 Regularized Problem Regularized Problem Sebastian Engel Optimal Control and BV-Functions Rigorosum 8
  • 9. International Research Training Group IGDK 1754 Regularized Problem For the numerical realization, we would now like to consider a problem that is numerically easier to solve (Pieper [Pi15]): p˜Pq # min pvYcqPL2p0YTqmˆRm JpvYcq ` 2 ˜ mÿ j“1 }vj}2 L2p0YTq ` }c}2 Rm ¸ “: J1 pvYcq Theorem Denote by pÝÑv YÝÑc q the unique solutions of p˜Pq and by u their BVp0YTqm representation. Then we have: § 0 ď J1 pÝÑv YÝÑc q ´ Jpuq “ Opq § u Ñ0 ÝÝÝÑ u strictly in BVp0YTqm (Fundamental theorem of calculus) § p 1 H2 p0YTqm ÝÝÝÝÝÝÑ p1 with p 1 ptq :“ Tş t ş wj p˚ pSpÝÑv YÝÑc q ´ ydqgjdxds Sebastian Engel Optimal Control and BV-Functions Rigorosum 9
  • 10. International Research Training Group IGDK 1754 Regularized Problem - Optimality Conditions Due to the additional L2 ´regularization, we can use a Prox-Operator approach: Theorem ˆÝÑv ÝÑc ˙ is optimal i $ ’ ’% ÝÑv “ Proxř i i }¨}L1p0;Tq p´1 p 1 q p 1 p0q ` ÝÑc “ 0Rm , /. /- with Proxř i i }¨}L1p0;Tq p´1 p 1 q :“ ¨ ˝ max ´ 0Y´1 p 1Yipsq ´ i ¯ ` ` min ´ 0Y´1 p 1Yipsq ` i ¯ ˛ ‚ m i“1 § Proxř i i }¨}L1p0;Tq p´1 p 1 q is semismooth. Sebastian Engel Optimal Control and BV-Functions Rigorosum 10
  • 11. International Research Training Group IGDK 1754 BV Path-Following Algorithm Dene the following function: F pÝÑv ;ÝÑc q :“ ˜ÝÑv ´ Prox ř i i }¨}L1p0;Tq p´ 1 p 1 q ÝÑc ` 1 p1p0q ¸ Since we approximate p˜Pq by p˜Pq we consider the Path-Following algorithm: BV Path-Following Algorithm Input: u0 P L2 p0YTqm ˆ Rm, 0 ą 0, TOL ą 0, TOLN ą 0, k “ 0 and # P p0Y1q while k ą TOL do Set i “ 0, ui k “ uk while }F k pui kq}L2p0;TqmˆRm ą TOLN do Solve DF k pukqpuq “ ´F k pukq, set ui`1 k`1 “ ui k ` u; i “ i ` 1. end Dene uk`1 “ ui k, and k`1 “ k; set k “ k ` 1. end A similar approach was used for a fully discretized control problem with semi-linear parabolic constraints in [CKK17]. Sebastian Engel Optimal Control and BV-Functions Rigorosum 11
  • 12. International Research Training Group IGDK 1754 BV Path-Following Algorithm - Super Linearity Theorem (Super Linearity of the Newton Method) For each ą 0 there exists a ą 0 s.t. for all pÝÑv YÝÑc q P L2 p0YTqm ˆ Rm with }pÝÑv YÝÑc q ´ pÝÑv YÝÑc q}L2p0YTqmˆRm ă Y the semismooth Newton method algorithm converging superlinearly. Sebastian Engel Optimal Control and BV-Functions Rigorosum 12
  • 13. International Research Training Group IGDK 1754 Numerical Example Example: “ r´1Y1s2 , T “ 2, “ 0X005, and patch gpxq “ 1r´0X5Y0X5s2 pxq, i.e. Desired state: yd :“ Spuq ´ pBtt ´ 4q9ptYxq with py0Yy1q “ p0Y0q for S and 9ptYxq :“ sinp3%tq sinp3% 2 tq dź i“1 cosp % 2 xiq with “ 3%l 4 ´ 2 ? 2 % ¯´2 Sebastian Engel Optimal Control and BV-Functions Rigorosum 13
  • 14. International Research Training Group IGDK 1754 Numerical Example Dirac-example until “ 10´8 (Full Discretization: v P S( and y P S( b Sh). Sebastian Engel Optimal Control and BV-Functions Rigorosum 14
  • 15. International Research Training Group IGDK 1754 Additional Control Problems All presented results are extended for other linear dierential equations, i.e. § Hyperbolic constraints: Btty ` Ay “ mř i“1 ui ¨ gi, § A as elliptic operator, § with Dirichlet, Neumann, or Robin B.C. § Linear parabolic constraints, which extends the results of [CKK17]. Sebastian Engel Optimal Control and BV-Functions Rigorosum 15
  • 16. International Research Training Group IGDK 1754 Error Rates for Variational Discretization Error Rates for Variational Discretization Sebastian Engel Optimal Control and BV-Functions Rigorosum 16
  • 17. International Research Training Group IGDK 1754 Error Rates BV-Control Problems § i[Bar12]: Image reconstruction by TV-based optimal control (no PDE constraints). § linear continuous uh : }u ´ uh}L2p q ď ch 1 6 . § i[CKK17]: Optimal control of a semi-linear parabolic equation. § One dimensional BV-controls, u;h cellwise constant. }y ´ y;h}L2p q ` |Jpuq ´ J;hpu;hq| ď cp ? ` hq § i[HMNV19]: Optimal control of one dimensional elliptic equations. § Variational discretization: Optimal convergence results for state, adjoint, and control (piecewise constant assumption). Sebastian Engel Optimal Control and BV-Functions Rigorosum 17
  • 18. International Research Training Group IGDK 1754 Standard Approach - Error Estimates § In case of smooth cost functions Jpuq, the standard approach uses coercivity properties, e.g. § appropriate testing of the 1st order optimality conditions, to derive error estimates for the control. § Error rates for the optimal states, costs and TV-semi-norm of the optimal controls can be obtained in a direct manner, which are sub-optimal. § For optimal error rates of the controls in the strict BV-topology, state, and costs, we need several assumptions on the adjoint function p1. Sebastian Engel Optimal Control and BV-Functions Rigorosum 18
  • 19. International Research Training Group IGDK 1754 Variational Discretization of p˜Pq In the following we discretize the state equation by linear continuous FE in time and space (S( b Sh), where controls will not be changed: p˜Psemi (Yh q $ ’ ’% min v P Mp0YTqm c P Rm 1 2 }S(YhpvYcq ´ yd}2 L2pT q ` mÿ j“1 j ż T 0 d|vj| “: J(YhpvYcq Optimality Conditions pÝÑv (YhYÝÑc (Yhq P Mp0YTqm ˆ Rm is the solution of p˜Psemi (Yh q, if ´ ˆ p1Yp(Yhq p1Yp(Yhqp0q ˙ :“ ´ ˆ p1Yp(YhqpÝÑv (YhYÝÑc (Yhq p1Yp(YhqpÝÑv (YhYÝÑc (Yhqp0q ˙ P ˆ` iB}ÝÑv iY(Yh}Mp0YTq ˘m i“1 0Rm ˙ § }p1Y(Yh ´ p1}L8p0YTqm Ñ 0X Sebastian Engel Optimal Control and BV-Functions Rigorosum 19
  • 20. International Research Training Group IGDK 1754 Convergence Results - Standard Approach “ 2 3 : yd P C1 pI; L2 pqq and py0Yy1q P H1 0 pq ˆ L2 pq. “ 1: yd P C1 pI; H1 0 pqq, g P pHp2q qm, and py0Yy1q P Hp3q ˆ Hp2q . State Error Rates }SpÝÑv YÝÑc q ´ S(YhpÝÑv (YhYÝÑc (Yhq}L2pT q P Op( ` hq Cost Error Rates |Jpuq ´ J(Yhpu(Yhq| P Op( ` hq Y with “ for “ 2 3 and “ 2 elseX ˇ ˇ ˇ ˇ}Dtu}Mp0YTqm ´ }Dtu(Yh}Mp0YTqm ˇ ˇ ˇ ˇ P O` ( ` h ˘ Y with u(Yhptq “ ş r0Yts dÝÑv (Yh ` ÝÑc (YhX Sebastian Engel Optimal Control and BV-Functions Rigorosum 20
  • 21. International Research Training Group IGDK 1754 Optimal Control with PDE constraints and Sparse Controls § What are sparse controls? Sparsity [Cas17]: BV-controls are sparse, if their distributional derivative is singular with respect to the Lebesgue measure. Sebastian Engel Optimal Control and BV-Functions Rigorosum 21
  • 22. International Research Training Group IGDK 1754 Optimality Conditions - Consequences For the optimal control pÝÑv YÝÑc q of p˜Pq holds: $ % supppÝÑv ˘ i q Ă tt P r0YTs | p1Yiptq “ ¯i u }p1Yi}C0pIq ď i We have analogous results for pÝÑv (YhYh(Yhq. § If D :“ tp1Yi “ ˘iu is a nite set, we nd that u is piecewise constant, i.e. uiptq “ ÿ aPD 'a ¨ 1raYTsptq ` ÝÑc § In practice, we often observe piecewise constant controls. § In general, we cannot expect piecewise constant controls u. Sebastian Engel Optimal Control and BV-Functions Rigorosum 22
  • 23. International Research Training Group IGDK 1754 Optimal Convergence Results - Preliminaries Assumptions A1: t t P p0YTq | |p1Yiptq| “ u “ tt1YiY¨ ¨ ¨ Ytmi Yiu for mi P Ną0, with i “ 1Y¨ ¨ ¨ Ym and mi P N. A2: Bttp1ptjYiq ‰ 0, for i “ 1Y¨ ¨ ¨ Ym and j “ 1Y¨ ¨ ¨ Ymi. § Assumption A1 implies: ui “ miř –“1 ci – 1rt`;i YTs ` ciY whereby ci – can be 0, if ui has no jump in t–Yi. [HiDe] for an elliptic problem without L2 -regularization: |tp˚ pxq “ 0u| “ 0 Bang-Bang control ´ Variational Discretization ´ }u´uh}L1pq ď ch2 |lnphq| Sebastian Engel Optimal Control and BV-Functions Rigorosum 23
  • 24. International Research Training Group IGDK 1754 Optimal Convergence Results - Preliminaries Due to the structural assumptions A1, A2, we obtain: Explicit Form There exists a p(0Yh0q such that @p(Yhq “ 5 ď p(0Yh0q it holds uptq “ ¨ ˚ ˝ c1 ... cm ˛ ‹ ‚` ¨ ˚ ˚ ˚ ˚ ˚ ˝ m1ř j“1 c1 j 1ptj;1YTsptq ... mmř j“1 cm j 1ptj;mYTsptq ˛ ‹ ‹ ‹ ‹ ‹ ‚ Y u5ptq “ ¨ ˚ ˝ c1Y5 ... cmY5 ˛ ‹ ‚` ¨ ˚ ˚ ˚ ˚ ˚ ˝ m1ř j“1 c1 jY51pt1 j;#YTsptq ... mmř j“1 cm jY51ptm j;#YTsptq ˛ ‹ ‹ ‹ ‹ ‹ ‚ Sebastian Engel Optimal Control and BV-Functions Rigorosum 24
  • 25. International Research Training Group IGDK 1754 Optimal Convergence Results - Preliminaries This implies the following estimate for the optimal controls: L1 ´ Estimate For all p(Yhq “ 5 ď p(0Yh0q holds }u ´ u5}L1p0YTqm ď rc ˆ mř i“1 |ci ´ ciY5| ` miř j“1 |ci j | ¨ |tjYi ´ ti jY5| ` |ci j ´ ci jY5| ˙ Sebastian Engel Optimal Control and BV-Functions Rigorosum 25
  • 26. International Research Training Group IGDK 1754 Optimal Convergence Results We dene “ 2 3 for pydYÝÑg Yy0Yy1q P C1 pI; L2 pqq ˆ L2 pqm ˆ H1 0 pq ˆ L2 pq We eand “ 1 for pydYÝÑg Yy0Yy1q P C1 pI; H1 0 pqq ˆ pHp2q qm ˆ Hp3q ˆ Hp2q . Amplitude |ci j ´ ci jY5| Jump |tjYi ´ ti jY5| Constant |ci ´ ciY5| , ////. ////- “ Op( ` hq if “ 2 3 Y ď c ` (2 ` h2 ` }Spuq ´ S5pu5q}L2pT q ˘ , if “ 1X Control ´ Error Rate: This implies: }u ´ u5}L1p0YTqm P Op( ` hq (suboptimal)X Sebastian Engel Optimal Control and BV-Functions Rigorosum 26
  • 27. International Research Training Group IGDK 1754 Optimal Convergence Results In case of pydYÝÑg Yy0Yy1q P C1 pI; H1 0 pqq ˆ pHp2q qm ˆ Hp3q ˆ Hp2q we obtain (by Young inequality and control rates for “ 1): Optimal Control Error Rates }u ´ u5}L1pIqm Y |ci ´ ciY5|Y |tjYi ´ ti jY5|Y |ci j ´ ci jY5| , . - “ Op(2 ` h2 q with i “ 1Y¨ ¨ ¨ Ym, j “ 1Y¨ ¨ ¨ Ymi, Optimal State and Total Variation Error Rates }Spuq ´ S5pu5q}L2pT q “ Op(2 ` h2 q ˇ ˇ ˇ ˇ}Dtu}MpIq ´ }Dtu5}MpIq ˇ ˇ ˇ ˇ “ Op(2 ` h2 q pBV-Strict Convergence!qX Sebastian Engel Optimal Control and BV-Functions Rigorosum 27
  • 28. International Research Training Group IGDK 1754 Optimal Control of Hyperbolic Equations Additional Results and Outlook Sebastian Engel Optimal Control and BV-Functions Rigorosum 28
  • 29. International Research Training Group IGDK 1754 Sound Propagation in an Urban Environment(rHhsrTtsrTtsrPpsrSss?|rFf srTtsrPpsq:TrueWave:swf (rHhsrTtsrTtsrPpsrSss?|rFf srTtsrPpsq: TrueWave:swf § Hyperbolic Equation: Btty ´ divpbpxqOyq “ řm i“1 uiptq ¨ gipxq § Sound is emitted by acoustic speakers, placed in gipxq Sebastian Engel Optimal Control and BV-Functions Rigorosum 29
  • 30. International Research Training Group IGDK 1754 Sound Propagation in an Urban Environment Data of the microphone recordings can be used to obtain a possible position using optimal control theory: Optimal Control - Example of Penalization Costs § Examples of penalization terms, used to nd an optimal control, e.g. Moving Microphone: Tş 0 ´ ´ ş Br pptqq yptqdx ´ ydataptq ¯2 dt § The solution of the following minimization problem leads to a possible sound source position: min uPX 1 2 ř iYj ´ ´ ş Br pxj q yptiqdx ´ ydata ¯2 ` 2 }u}X where X “ BVp0YTqm or X “ H1 p0YTqm with }u}BVp0YTqm “ mř i“1 }Btui}Mp0YTq or }u}H1p0YTqm “ mř i“1 }Btui}2 L2p0YTq ` }up0q}2 Rm . Sebastian Engel Optimal Control and BV-Functions Rigorosum 30
  • 31. International Research Training Group IGDK 1754 Sound Propagation in an Urban Environment(rHhsrTtsrTtsrPpsrSss?|rFf srTtsrPpsq:Comparision:swf (rHhsrTtsrTtsrPpsrSss?|rFf srTtsrPpsq: Comparision:swf i “ 0X1 for BV and H1 , bpxq “ $ % 10´10 Y inside 0X35Y outside r10´10 Y0X35sY else Sebastian Engel Optimal Control and BV-Functions Rigorosum 31
  • 32. International Research Training Group IGDK 1754 Bayesian Inversion with BV-Prior Based on [EHMS19], we are able to consider a stochastic view on the sound source identication problem before: ydata “ ´ ´ ş Br pxj q yuptiqdx ¯ i“1M ` with noise „ Np0Y'q and stochastic control u “ kř –“0 –1pt`YTs ` c „ 0 Bayesian Inversion § Well-posedness of the posterior solution dypuq “ expp´1 2 }yu ´ ydata}2 ¦qd0puq. § Convergence results for the SMC-Algorithm and FE-Discretization § Expectation, condence region, MAP-Estimator, ... A similar Bayesian inversion problem is considered by [Yao16] (TV-Gaussian Prior, Splitting pCN). Sebastian Engel Optimal Control and BV-Functions Rigorosum 32
  • 33. International Research Training Group IGDK 1754 Summary We analyzed control problems with second-order hyperbolic equations, with various types of cost functionals, which are regularized by a non-smooth semi-norm of BVp0YTq. § Optimality conditions + sparsity + explicit examples. Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
  • 34. International Research Training Group IGDK 1754 Summary We analyzed control problems with second-order hyperbolic equations, with various types of cost functionals, which are regularized by a non-smooth semi-norm of BVp0YTq. § Optimality conditions + sparsity + explicit examples. Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
  • 35. International Research Training Group IGDK 1754 Summary We analyzed control problems with second-order hyperbolic equations, with various types of cost functionals, which are regularized by a non-smooth semi-norm of BVp0YTq. § Optimality conditions + sparsity + explicit examples. § Regularized problem pPq (H1 ´norm) + convergence results (control, costs, adjoint) Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
  • 36. International Research Training Group IGDK 1754 Summary We analyzed control problems with second-order hyperbolic equations, with various types of cost functionals, which are regularized by a non-smooth semi-norm of BVp0YTq. § Optimality conditions + sparsity + explicit examples. § Regularized problem pPq (H1 ´norm) + convergence results (control, costs, adjoint) Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
  • 37. International Research Training Group IGDK 1754 Summary We analyzed control problems with second-order hyperbolic equations, with various types of cost functionals, which are regularized by a non-smooth semi-norm of BVp0YTq. § Optimality conditions + sparsity + explicit examples. § Regularized problem pPq (H1 ´norm) + convergence results (control, costs, adjoint) § Numerical analysis of a semismooth Newton algorithm for pPq + super-linear convergence. Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
  • 38. International Research Training Group IGDK 1754 Summary We analyzed control problems with second-order hyperbolic equations, with various types of cost functionals, which are regularized by a non-smooth semi-norm of BVp0YTq. § Optimality conditions + sparsity + explicit examples. § Regularized problem pPq (H1 ´norm) + convergence results (control, costs, adjoint) § Numerical analysis of a semismooth Newton algorithm for pPq + super-linear convergence. Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
  • 39. International Research Training Group IGDK 1754 Summary We analyzed control problems with second-order hyperbolic equations, with various types of cost functionals, which are regularized by a non-smooth semi-norm of BVp0YTq. § Optimality conditions + sparsity + explicit examples. § Regularized problem pPq (H1 ´norm) + convergence results (control, costs, adjoint) § Numerical analysis of a semismooth Newton algorithm for pPq + super-linear convergence. § Error estimates for a variational discretization + optimal rates under specic assumptions (sparsity in pPq). Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
  • 40. International Research Training Group IGDK 1754 Summary We analyzed control problems with second-order hyperbolic equations, with various types of cost functionals, which are regularized by a non-smooth semi-norm of BVp0YTq. § Optimality conditions + sparsity + explicit examples. § Regularized problem pPq (H1 ´norm) + convergence results (control, costs, adjoint) § Numerical analysis of a semismooth Newton algorithm for pPq + super-linear convergence. § Error estimates for a variational discretization + optimal rates under specic assumptions (sparsity in pPq). Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
  • 41. International Research Training Group IGDK 1754 Summary We analyzed control problems with second-order hyperbolic equations, with various types of cost functionals, which are regularized by a non-smooth semi-norm of BVp0YTq. § Optimality conditions + sparsity + explicit examples. § Regularized problem pPq (H1 ´norm) + convergence results (control, costs, adjoint) § Numerical analysis of a semismooth Newton algorithm for pPq + super-linear convergence. § Error estimates for a variational discretization + optimal rates under specic assumptions (sparsity in pPq). § A Bayesian perspective on pPq. Sebastian Engel Optimal Control and BV-Functions Rigorosum 33
  • 42. International Research Training Group IGDK 1754 Literature [Bar12], S. Bartels, Total variation minimization with nite elements: convergence and iterative solution. SIAM J. Numer. Anal., 50(3):1162-1180, 2012. [Bog98], V. I. Bogachev, Gaussian measures, vol.62, Mathematical Surveys and Monographs, American Mathematical Society, 1998. [Cas17], E. Casas, A review on sparse solutions in optimal control of partial dierential equations. SEMA Journal, 74, pp. 319-344, 2017. [CK17], E. Casas, K. Kunisch, Analysis of optimal control problems of semilinear elliptic equations by bv-functions. Set-Valued and Variational Analysis, 1-25, 2017. Sebastian Engel Optimal Control and BV-Functions Rigorosum 34
  • 43. International Research Training Group IGDK 1754 Literature [CKK17], E. Casas, K. Kunisch, and F. Kruse. Optimal control of semilinear parabolic equations by bv-functions. SIAM Journal on Control and Optimization, 55:1752-1788, 2017. [CK11], C. Clason and K. Kunisch, A duality based approach to elliptic control problems in non refelxive banach spaces. ESIAM Control Optim. Calc. Var., 17, pp. 243-266, 2011. [HiDe], K. Deckelnick, M. Hinze, A note on the approximation of ellliptic control problems with bang-bang controls. Comput. Optim. Appl., 51:931-939, 2010. [EHMS19], S. Engel, D. Hafemeyer, C. Münch, and D. Schaden. An application of sparse measure valued Bayesian inversion to acoustic sound source identication. Inverse Problems, accepted. Sebastian Engel Optimal Control and BV-Functions Rigorosum 35
  • 44. International Research Training Group IGDK 1754 Literature [HMNV19], D. Hafemeyer, F. Mannel, I. Neitzel, B. Vexler, Finite element error estimates for elliptic optimal control by BV functions, arxiv, 2019. [Hor16], M. Hornikx, Ten questions concerning computational urban acoustics. Building and Environment 106 : 409-421, 2016. [KTV16], K. Kunisch, P. Trautmann, B. Vexler, Optimal control of the undamped linear wave equation with measure valued controls. SIAM Journal on Control and Optimization, 54(3), 1212-1244, 2016. [KKV10], A. Kröner, K. Kunisch, B. Vexler, Semismooth Newton Methods for an Optimal Boundary Control Problem of Wave Equations, in: Recent Advances in Optimization and its Applications in Engineering, M. Diehl, F. Glineur, E. Jarlebring, W. Michiels (eds.), Springer, 389-398, 2010. Sebastian Engel Optimal Control and BV-Functions Rigorosum 36
  • 45. International Research Training Group IGDK 1754 Literature [KKV11], A. Kröner, K. Kunisch, B. Vexler, Semismooth Newton Methods for Optimal Control of the Wave Equation with Control Constraints, SIAM J.Control Optim. 49: 830-858, 2011. [Lion], J. L. Lions: Optimal Control of Systems Governed by Partial Dierential Equations. Springer-Verlag, Berlin, 1971. [MR04], B. S. Mordukhovich, J. P. Raymond, Neumann boundary control of hyperbolic equations with pointwise state constraints. SIAM journal on control and optimization, 43(4), 1354-1372, 2004. [Pi15], K. Pieper, Finite element discretization and ecient numerical solution of elliptic and parabolic sparse control problems. PhD thesis, Technical University Munich, 2015. Sebastian Engel Optimal Control and BV-Functions Rigorosum 37
  • 46. International Research Training Group IGDK 1754 Literature [Sta09], G. Stadler, Elliptic optimal control problems with L1 -control cost and applications for the placement of control devices. Comput. Optim. Appl., 44(2):159-181, 2009. [TVZ17], P. Trautmann, B. Vexler, A. Zlotnik, Finite element error analysis for measure-valued optimal control problems governed by a 1D wave equation with variable coecients, arXiv , 2017. [Yao16], Z. Yao, Z. Hu, and J. Li. A tv-gaussian prior for innite-dimensional bayesian inverse problems and its numerical implementations. IOPscience, 2016. [Zlo94], A. A. Zlotnik, Convergence rate estimates of nite-element methods for second-order hyperbolic equations. numerical methods and applications, p.153 et.seq. Guri I. Marchuk, CRC Press, 1994. Sebastian Engel Optimal Control and BV-Functions Rigorosum 38
  • 47. International Research Training Group IGDK 1754 Appendix Variational Discretization Sebastian Engel Optimal Control and BV-Functions Rigorosum 39
  • 48. International Research Training Group IGDK 1754 Optimal Control of the Wave Equation with BV-Functions Consider the following semi-discretized optimal control problem: pPsemi (Yh q $ ’’’’ ’’’’% min uPBVp0YTqm 1 2 }yuY(Yh ´ yd}2 L2pT q ` mÿ j“1 j ż r0YTs d|Dtuj|ptq “: J(Yhpuq with }y ´ y(Yh}Cp0YT;L2pqq P Op(2 3 ` h 2 3 qY for pFEM Zlotnikq f “forcing y0 “displacement y1 “velocity 3 C1 pr0YTs; H1 0 pqq or BVp0YT; H2 q H3 H2 2 C1 pr0YTs; L2 pqq H2 H1 0 pq 1 L2 pTq H1 0 pq L2 pq § H “ w P L2 pq ˇ ˇ ˇ ˇ ř kě1 ! k xwYky2 L2pq ă 8 * with eigenvalues and eigenfunctions p!kYkqkě1 of ´4 with homogeneous Dirichlet boundary conditions. Sebastian Engel Optimal Control and BV-Functions Rigorosum 40
  • 49. International Research Training Group IGDK 1754 Optimal Convergence Results - Preliminaries Due to the structural assumptions A1 and A2, we obtain: Lemma There exists a ą 0, and (0Yh0 ą 0 such that i @p(Yhq “ 5 ď p(0Yh0q holds: ÝÑv 5Yi “ miř l“1 ci lY5ti l;# with ti jY5 P BptjYiq where pBptjYiqqmi j“1 are pairwise disjoint for a xed i “ 1Y¨ ¨ ¨ Ym with 0 ă 5 Ñ 0 Sebastian Engel Optimal Control and BV-Functions Rigorosum 41
  • 50. International Research Training Group IGDK 1754 Optimal Convergence Result Optimal Rates: 1st-Step, State Dependence: }u ´ u5}L1pIqm ď ˆcp(2 ` h2 ` }Spuq ´ S5pu5q}L2pT qq 2nd-Step, Scaled Young Inequality: }Spuq ´ S5pu5q}L2pT q ďljhn FOOC # c}Spuq ´ S5puq}L2pT q `cpÝÑg q}u5 ´ u} 1 2 L1pIqm }L˚ pSpuq ´ ydq ´ L˚ 5pSpuq ´ ydq} 1 2 L8p0YT;L2pqq + ďljhn Zlotnik Young Ineq.Y ą 0 # cp(2 ` h2 q ` ˆc}Spuq ´ S5puq}L2pT q `cp ÝÑg q 4 }L˚ pSpuq ´ ydq ´ L˚ 5pSpuq ´ ydq}L8p0YT;L2pqq + Sebastian Engel Optimal Control and BV-Functions Rigorosum 42
  • 51. International Research Training Group IGDK 1754 Optimal Convergence Results - Consequences Let the assumptions A1 and A2 hold: Properties of u5: Consider an amplitude ci jY5 of an optimal control of pPsemi 5 q, with i “ 1Y¨ ¨ ¨ Ym, and j “ 1Y¨ ¨ ¨ Ymi: a) Assume that |ci j | ą 0. The optimal control of pPsemi 5 q has a jump (|ci jY5| ą 0) inside BptjYiq for all 0 ă 5 ă 50 and 50 small enough. b) Assume that ci j “ 0. The optimal control of pPsemi 5 q can have a jump in BptjYiq for all 0 ă 5 ă 50, with 50 small enough, but the jump height has to decrease with some specic rate pci jY5 P Op( ` hqq. c) Let tt P I||p1Yiptq| “ u “ H for i “ 1Y¨ ¨ ¨ Ym, and m ě 1. Then the optimal control sub function u5Yi of pPsemi 5 q has no jumps pu5Yi “ const.q for all 0 ă 5 ď 50. Sebastian Engel Optimal Control and BV-Functions Rigorosum 43
  • 52. International Research Training Group IGDK 1754 Appendix More General Costs Sebastian Engel Optimal Control and BV-Functions Rigorosum 44
  • 53. International Research Training Group IGDK 1754 Control Problem pP¥q $ ’’’’’’’’’ ’’’’’’’’’% min uPBVp0YTqm 3ř j“1 j 2 } r¥jpyuq ´ zj}2 Oj ` mř –“1 –}Dtu–}Mp0YTq “: J¥puq s.t. $ ’’’ ’’’% lyu :“ Bttyu ´ 4yu “ mÿ j“1 ujgj in p0YTq ˆ y “ 0 on p0YTq ˆ B pyp0qYBtyp0qq “ py0Yy1q P H1 0 pq ˆ L2 pq § zj P Oj sep. Hilbert spaces and pg–qm – Ă L8 pqzt0u with pairwise disjoint supports. r¥2pyuq :“ ¥2pyupt–qqr –“1 Y r¥3pyuq :“ ¥3pBtyupt–qqr –“1 Y 0 “ t0 ă t1 ă ¨ ¨ ¨ ă tr “ TY Lin. indep. p r¥j pyg`qqm `“1 for one j ą0 with $ % r¥1 P Lp L2 pTq Y O1qY ¥2 P Lp L2 pqr Y O2qY ¥3 P Lp H´1 pqr Y O3qX This convex problem has a solution. Sebastian Engel Optimal Control and BV-Functions Rigorosum 45
  • 54. International Research Training Group IGDK 1754 Equivalent Problem and Optimality Conditions prP¥q $ ’’’ ’’’% min pvYcqPMp0YTqmˆRm J¥puq :“ rJ¥pvYcq with uptq “ ˆtş 0 dvjpsq ` cj ˙m i“1 resp. pDtuYup0qq “ pvYcqX The rst-order optimality condition has the following form: ˆ p¥ 1 ptq p¥ 1 p0q ˙ “ ¨ ˝´ Tş t ş ˆppÝÑv YÝÑc qÝÑg ´p¥ 1 p0q ˛ ‚P ˆ piB}pÝÑv i}Mqqi 0 ˙ text with the jumping wave function ˆppÝÑv YÝÑc q :“ řr i“1 pi1Ii with Ii :“ pti´1Ytiq l pi “ gi in Ii ˆ Y piptiq “ pi`1ptiq ` hiY Btpiptiq “ Btpi`1ptiq ` ki gi “ 1¥˚ 1 p¥1pyuq ´ z1q|Ii ˆ P L2 pIi ˆ qY hi “ 3p´4q´1 ” ¥˚ 3 p r¥3pyuq ´ z3q ı i P H1 0 pqY ki “ ´2 ” ¥˚ 2 p r¥2pyuq ´ z2q ı i P L2 pqX Sebastian Engel Optimal Control and BV-Functions Rigorosum 46
  • 55. International Research Training Group IGDK 1754 Regularized Control Problem prP¥ q # min pvYcqPL2p0YTqmˆRm rJ¥pvYcq ` 2 ˜ mř j“1 }vj}2 L2p0YTq ` }c}Rm ¸ For the BV-representation u of the optimal control pvYcq holds u strict ÝÝÝÑ BV u. Sebastian Engel Optimal Control and BV-Functions Rigorosum 47
  • 56. International Research Training Group IGDK 1754 Regularized Control Problem - Jumping Wave Example Let 1 “ 3 “ 0, and consider –, – “ 1Y¨ ¨ ¨ Yk, as pairwise disj. balls inside : r¥2pyuq :“ ˆ´ 1 |`| ş ` yuptiqdx ¯k –“1 ˙r i“1 P Rk¨r p„ Microphones) One experimental result with ÝÑ t “ p0X098Y0X199Y0X297Y0X347Y0X398Y1q: Sebastian Engel Optimal Control and BV-Functions Rigorosum 48
  • 57. International Research Training Group IGDK 1754 Appendix Bayesian Inversion Sebastian Engel Optimal Control and BV-Functions Rigorosum 49
  • 58. International Research Training Group IGDK 1754 Bayesian Inversion and pP¥q - Specic Example Consider the following stochastic problem: pI¥q z2 “ r¥2pyuq ` Y „ Np0Y 1 2 ¨ idRk¨r qY where r¥2 is dened as in (Microphones) and u K . Sebastian Engel Optimal Control and BV-Functions Rigorosum 50
  • 59. International Research Training Group IGDK 1754 Bayesian Inversion and pP¥q - Specic Example Consider the following stochastic problem: pI¥q z2 “ r¥2pyuq ` Y „ Np0Y 1 2 ¨ idRk¨r qY where r¥2 is dened as in (Microphones) and u K . § Without any knowledge of z2, we assume that u „ 0 (Prior) Ñ How should 0 be dened? Gaussian? Sebastian Engel Optimal Control and BV-Functions Rigorosum 50
  • 60. International Research Training Group IGDK 1754 Bayesian Inversion and pP¥q - Specic Example Consider the following stochastic problem: pI¥q z2 “ r¥2pyuq ` Y „ Np0Y 1 2 ¨ idRk¨r qY where r¥2 is dened as in (Microphones) and u K . § Without any knowledge of z2, we assume that u „ 0 (Prior) Ñ How should 0 be dened? Gaussian? § We consider ˜u „ Np0Y 1 Iq “ with: $ % I P LpL2 pIqq trace class rgpI1{2 q “ H1 p0YTqY (see [Bog98,EX 2.3.4]). Sebastian Engel Optimal Control and BV-Functions Rigorosum 50
  • 61. International Research Training Group IGDK 1754 Bayesian Inversion and pP¥q - Specic Example Consider the following stochastic problem: pI¥q z2 “ r¥2pyuq ` Y „ Np0Y 1 2 ¨ idRk¨r qY where r¥2 is dened as in (Microphones) and u K . § Without any knowledge of z2, we assume that u „ 0 (Prior) Ñ How should 0 be dened? Gaussian? § We consider ˜u „ Np0Y 1 Iq “ with: $ % I P LpL2 pIqq trace class rgpI1{2 q “ H1 p0YTqY (see [Bog98,EX 2.3.4]). § We weight by f pwq “ 1 £0 expp´}Btw}Mq Ñ concentration to functions with small total variation, i.e. 0p¨q “ ş ¨ f pwqdpwq (see [Yao18]). Sebastian Engel Optimal Control and BV-Functions Rigorosum 50
  • 62. International Research Training Group IGDK 1754 Bayesian Inversion and pP¥q - Specic Example § We have u|z2 „ z, with zp¨q “ 1 £z ş ¨ expp´2 2 }Ă¥2pyuq ´ z2}2 Rk¨r qd0puq (Posterior). Sebastian Engel Optimal Control and BV-Functions Rigorosum 51
  • 63. International Research Training Group IGDK 1754 Bayesian Inversion and pP¥q - Specic Example § We have u|z2 „ z, with zp¨q “ 1 £z ş ¨ expp´2 2 }Ă¥2pyuq ´ z2}2 Rk¨r qd0puq (Posterior). § Maximum a Posteriori Estimator: min uPH1p0YTq 2 2 } r¥2pyuq ´ z2}2 Rk¨m ` }Btu}M ` p}Btu}2 L2 ` }u}2 L2 q Sebastian Engel Optimal Control and BV-Functions Rigorosum 51
  • 64. International Research Training Group IGDK 1754 Bayesian Inversion and pP¥q - Specic Example § We have u|z2 „ z, with zp¨q “ 1 £z ş ¨ expp´2 2 }Ă¥2pyuq ´ z2}2 Rk¨r qd0puq (Posterior). § Maximum a Posteriori Estimator: min uPH1p0YTq 2 2 } r¥2pyuq ´ z2}2 Rk¨m ` }Btu}M ` p}Btu}2 L2 ` }u}2 L2 q § Sampling and Empirical MAP (Splitting pCN with 100,000 sample steps): We used ˜uptq “ 8ř –“0 Np0Y!2 – q ¨ 9–ptq „ Np0Yp´1 40q´1 q, with MAP: min uPH1 0 p0YTq 2 2 } r¥2pyuq ´ z2}2 Rk¨m `}Btu}M ` }Btu}2 L2 Sebastian Engel Optimal Control and BV-Functions Rigorosum 51
  • 65. International Research Training Group IGDK 1754 Bayesian Inversion and pP¥q - Specic Example § We have u|z2 „ z, with zp¨q “ 1 £z ş ¨ expp´2 2 }Ă¥2pyuq ´ z2}2 Rk¨r qd0puq (Posterior). § Maximum a Posteriori Estimator: min uPH1p0YTq 2 2 } r¥2pyuq ´ z2}2 Rk¨m ` }Btu}M ` p}Btu}2 L2 ` }u}2 L2 q § Sampling and Empirical MAP (Splitting pCN with 100,000 sample steps): We used ˜uptq “ 8ř –“0 Np0Y!2 – q ¨ 9–ptq „ Np0Yp´1 40q´1 q, with MAP: min uPH1 0 p0YTq 2 2 } r¥2pyuq ´ z2}2 Rk¨m `}Btu}M ` }Btu}2 L2 § Disadvantages (true function is sparse): Computationally expensive (KLE), high number of samples needed (lack of sparsity), good initial value needed. Sebastian Engel Optimal Control and BV-Functions Rigorosum 51
  • 66. International Research Training Group IGDK 1754 BV-Prior § We now assume that u „ kř –“0 –1pt`YTsptq ` c („ 0 BV-Prior). Sebastian Engel Optimal Control and BV-Functions Rigorosum 52
  • 67. International Research Training Group IGDK 1754 BV-Prior § We now assume that u „ kř –“0 –1pt`YTsptq ` c („ 0 BV-Prior). § The set of all realizations of u is dense in BVp0YTq, in the strict BV-topo. Sebastian Engel Optimal Control and BV-Functions Rigorosum 52
  • 68. International Research Training Group IGDK 1754 BV-Prior § We now assume that u „ kř –“0 –1pt`YTsptq ` c („ 0 BV-Prior). § The set of all realizations of u is dense in BVp0YTq, in the strict BV-topo. § Sampling + Empirical MAP + Optimal Control pP¥q (SMC 5,000 samples): Sebastian Engel Optimal Control and BV-Functions Rigorosum 52
  • 69. International Research Training Group IGDK 1754 BV-Prior § We now assume that u „ kř –“0 –1pt`YTsptq ` c („ 0 BV-Prior). § The set of all realizations of u is dense in BVp0YTq, in the strict BV-topo. § Sampling + Empirical MAP + Optimal Control pP¥q (SMC 5,000 samples): § Advantage: Faster algorithms by sparsity. Sebastian Engel Optimal Control and BV-Functions Rigorosum 52
  • 70. International Research Training Group IGDK 1754 Gaussian Measure [Bog98, Theorem 3.6.1] Let be a Radon Gaussian measure on a locally convex space X. Then the topological support of coincides with the ane subspace Epq ` H X , where H X stands for the closure in X. Sebastian Engel Optimal Control and BV-Functions Rigorosum 53
  • 71. International Research Training Group IGDK 1754 Thank you for your attention Supported by the DFG through the International Research Training Group IGDK 1754 Optimization and Numerical Analysis for Partial Dierential Equations with Nonsmooth Structures Sebastian Engel Optimal Control and BV-Functions Rigorosum 54