spectral optimization problems - univpm.it · spectral optimization problems giuseppe buttazzo...
TRANSCRIPT
Spectral Optimization Problems
Giuseppe Buttazzo
Dipartimento di Matematica
Universita di Pisa
http://cvgmt.sns.it
Workshop on “Calculus of Variations”Ancona, June 6–8, 2011
We deal with the optimization problem
minF (Ω) : Ω ∈ A(D)
where D is a given bounded domain of Rd
and A(D) is a class of admissible choicesmade of subsets of D.
Typical examples are:
F (Ω) = Φ(λ(Ω)
)where λ(Ω) is the spec-
trum of the Dirichlet Laplacian in Ω andΦ : RN → [0,+∞] is a given function; forinstance
F (Ω) = λk(Ω).
2
F (Ω) =∫Dj(x, uΩ) dx where j is a given in-
tegrand and uΩ is the solution in H10(Ω) of
−∆u = f ; for instance
j(x, u) = −a(x)up.
In both cases a very natural choice for ad-
missible domains is
A(D) =
Ω ⊂ D, Ω quasi-open, |Ω| ≤ m
.
Some good sources for results and problems...
3
Dorin B
ucur and Giuseppe B
uttazzoV
ariational Methods in Shape O
petimization P
roblems
PNLDE65
Dorin BucurGiuseppe Buttazzo
Variational Methodsin Shape OptimizationProblems
Birkhäuser
Progress in Nonlinear Differential Equationsand Their Applications
BirkhäuserISBN 0-8176-4359-1www.birkhauser.com
The study of shape optimization problems encompasses a wide spectrum ofacademic research with numerous applications to the real world. In thiswork these problems are treated from both the classical and modernperspectives and target a broad audience of graduate students in pure andapplied mathematics, as well as engineers requiring a solid mathematicalbasis for the solution of practical problems.
Key topics and features:
• Presents foundational introduction to shape optimization theory
• Studies certain classical problems: the isoperimetric problem and the Newton problem involving the best aerodynamical shape, and optimization problems over classes of convex domains
• Treats optimal control problems under a general scheme, giving a topological framework, a survey of γ-convergence, and problems governed by ODE
• Examines shape optimization problems with Dirichlet and Neumann conditions on the free boundary, along with the existence of classical solutions
• Studies optimization problems for obstacles and eigenvalues of elliptic operators
• Poses several open problems for further research
• Substantial bibliography and index
Driven by good examples and illustrations and requiring only a standardknowledge in the calculus of variations, differential equations, andfunctional analysis, the book can serve as a text for a graduate course incomputational methods of optimal design and optimization, as well as anexcellent reference for applied mathematicians addressing functional shapeoptimization problems.
Dorin Bucur and Giuseppe ButtazzoVariational Methods in Shape Optimization Problems
4
2005. XII, 334 p. (Mathématiques et
Applications, Vol. 48) Softcover
59,72 €
$129.00
SFr. 99.50
£47.99
ISBN 978-3-540-26211-4
A. Henrot, Université Henri Poincaré, Vandoeuvre-les-Nancy, France; M. Pierre, Antenne de Bre-
tagne de I'ENS Cachan, Bruz, France
Variation et optimisation de formes
Une analyse géométrique
Ce livre est une initiation aux approches modernes de l’optimisation mathématique de
formes. On y développe la méthodologie ainsi que les outils d’analyse mathématique et
de géométrie nécessaires à l’étude des variations de domaines. On y trouve une étude
systématique des questions géométriques associées à l’opérateur de Laplace, de la capa-
cité classique, de la dérivation par rapport à une forme, ainsi qu’un FAQ sur les topolo-
gies usuelles sur les domaines et sur les propriétés géométriques des formes optimales
avec ce qui se passe quand elles n’existent pas, le tout avec une importante bibliogra-
phie.... more on http://springer.com/978-3-540-26211-4
Yes, please send me copies
7 Call: + 49 (0) 6221-345-4301 7 Fax: +49 (0) 6221-345-4229 7 Web: springer.com7 Email: [email protected]
SpringerOrder DepartmentPO Box 2485Secaucus, NJ 07096-2485USA
Name
Address
Street Address
(Sorry, we cannot deliver to P.O. boxes)
City / State / ZIP-Code
Date Signature
Country
Telephone / Email
Methods of Payment Check/Money Order enclosed AmEx MasterCard VISA
Card No. Exp. Date
Order Now !
7 Call toll-free 1-800-SPRINGER
8:30 am – 5:30 pm ET7 Fax your order to (201) 348-45057 Web springer.com7 Email [email protected]
CA, MA, NJ, NY, and PA residents, please add sales tax. Canadian residents, please add 5% GST. Please add $5.00 for shipping one book and $1.00 for each additional book. Outside the US and Canada add $10.00 for first book, $5.00 for each additional book. All orders are processed upon receipt. If an order cannot be fulfilled within 90 days, payment will be refunded upon request. Prices are payable in US currency or its equivalent. Remember, your 30-day return privilege is always guaranteed. Pre-publication pricing: Unless otherwise stated, pre-pub prices are valid through the end of the third month following publication, and therefore are subject to change.
Springer Customer Service Center GmbH Haberstrasse 7 69126 Heidelberg Germany
All € and £ prices are net prices subject to local VAT, e.g. in Germany 7% VAT for books and 19% VAT for electronic products. Pre-publication pricing: Unless otherwise stated, pre-pub prices are valid through the end of the third month following publication, and therefore are subject to change. All prices exclusive of carriage charges. Prices and other details are subject to change without notice. All errors and omissions excepted.Please consult springer.com for information on postage.
Please send orders to: Outside the Americas:
"Variation et optimisation de formes"ISBN 978-3-540-26211-4
15/10/09 12:2241NCA1K2F5L.jpg 354×500 pixels
Page 1 of 1http://ec1.images-amazon.com/images/I/41NCA1K2F5L.jpg
One issue we intend to develop is the studyof gradient flow evolutions for shape opti-mization problems, via minimizing movements.
The framework is a metric space (space ofshapes) and a functional F : X → [0,+∞];via the Euler scheme of time step ε andinitial condition u0 ∈ X we may constructa step function uε(t) = w([t/ε]) by settingw(0) = u0 and
w(n+ 1) ∈ argminF(v) +
d2(v, w(n)
)2ε
.
The gradient flow u(t) is then a limit of asequence uεn with εn → 0.
6
In spectral optimization we take a suitabledistance on the space of shapes and
F(Ω) = Φ(λ(Ω)
).
The goal is to study existence and propertiesof spectral flows Ω(t). At this stage the onlyavailable results are obtained for debondingproblems, where mushy regions may appearduring the evolution.
D.Bucur, G.Buttazzo: J. Convex Anal. 2008D.Bucur, G.Buttazzo, A.Lux: Arch. Ratio-nal Mech. Anal. 2008.
From now on we restrict to static shape op-timization problems.
7
In general one cannot expect the existence
of a solution; for instance, the problem
min ∫
D|u− c|2 dx : −∆u = 1 in H1
0(Ω)
has no solution if c is small, and one has
to deal with relaxed solutions that in this
case are capacitary mesures µ, i.e. Borel
countably additive set functions with values
in [0,+∞] and such that
µ(E) = 0 whenever cap(E) = 0.
8
On the other hand, adding some geomet-rical constraint to the admissible domains,gives extra compactness that provides theexistence of a solution in a very large num-ber of situation. For instance:• The class Aconvex of convex sets containedin D.• The class Aunif cone of domains satisfyinga uniform exterior cone property.• The class Aunif flat cone of domains satis-fying a uniform flat cone condition, i.e., asabove, but with the weaker requirement thatthe cone may be flat, that is of dimensiond− 1.
9
• The class Acap density of domains satisfying
a uniform capacity density condition.
• The class Aunif Wiener of domains satisfy-
ing a uniform Wiener condition.
Aconvex ⊂ Aunif cone ⊂ Aunif flat cone⊂ Acap density ⊂ Aunif Wiener
• Another interesting class, which is only
of topological type and is not contained in
any of the previous ones, is (for d = 2) the
class of domains for which the number of
connected components of D \Ω is uniformly
bounded (Sverak 1993).
10
A powerful result which applies in the sit-
uations when no geometric constraints are
imposed(Buttazzo-Dal Maso 1993):
Theorem Assume that
1. F is decreasing for the set inclusion;
2. F is l.s.c. for the γ-convergence.
Then the minimum problem
minF (Ω) : |Ω| ≤ m, Ω ⊂ D quasi open
admits at least a solution Ωopt.
11
Example 1. Take F (Ω) = Φ(λ(Ω)
)with Φ
l.s.c. and increasing in each of its variables;
then the assumptions are verified. For in-
stance we may take F (Ω) = λk(Ω).
For k = 1 (and large D) the best domain is
a ball of volume m (Faber-Krahn 1923);
For k = 2 (and large D) the best domain is
the union of two disjoint balls of volume m/2
(Polya-Szego 1955).
12
Here is a numerical computation of optimalshapes for k = 3− 10 (Oudet)
k best array of balls best shape
12/03/09 12:20Eigenvalues
Page 1 of 1http://www.lama.univ-savoie.fr/~oudet/index.php?page=Galerie/../Eigenvalues/eigenvalues&lang=fr
Minimisation des modes propres
d'une membrane par rapport au
domaine
Le tableau ci-dessous présente les formes obtenues numériquement par uneméthode mixte de type level set/ relaxation lors de la minimisation des premiersmodes propres (k=3 à 10) d'une membrane à bord fixe (condition au bord de typeDirichlet) d'aire fixée (pour plus de détails voir Numerical minimization ofeigenmodes of a membrane with respect to the domain ainsi que Minimizing thesecond eigenvalue of the Laplace operator with Dirichlet boundary conditions).On compare ces résultats aux valeurs obtenues par Wolf et Keller en serestreignant à des réunions de disques dans l'article :
Wolf, Sven Andreas(1-STF); Keller, Joseph B.(1-STF) Range of the first two eigenvalues of the Laplacian. (English. English summary) Proc. Roy. Soc. London Ser. A 447 (1994), no. 1930, 397--412. 35P15 (58G25)
ACCUEIL
TRAVAUX RÉCENTS
PUBLICATIONS
UNE TOOLBOX POUR LA
GÉOMETRIE CONVEXE
GALERIE
La conjecture de Kelvin
Partitionnement optimal
Transport avec effet decongestion
Minimisation souscontraintes de volume
Ensembles de largeurconstante
Corps convexes
Un problème de transport
Valeur propres duLaplacien
Les arbres de Steiner
Simulation localisée d'unfaisceau de particules
CV, RAPPORT ACTIVITÉS
RESUME DE THÈSE
13
Very little is known on regularity of optimaldomains; apart the case of λ1 the proof thatthey are actually open sets is still missing.
For spectral optimization problems that donot fulfill the decreasing property of theoremabove, the case of
min
Φ(λ1, λ2) : |Ω| ≤ m, Ω ⊂ D
is interesting. Indeed, if D is large enough, itis possible to prove (Bucur-Buttazzo-Figueiredo1999) that the existence of an optimal do-main always occur, i.e. for every (continu-ous) function Φ.
14
Here is the plot of the set E of points of R2
that are of the form(λ1(Ω), λ2(Ω)
)for some
Ω ⊂ D with |Ω| ≤ m (Keller-Wolf 1994).6.4 The first two eigenvalues 155
Figure 6.1. The set E for N = 2 and c = 1.
Theorem 6.4.1 The set E is closed in R2.
The proof of the theorem above is based on the following lemma.
Lemma 6.4.2 If the set E is convex on the vertical and horizontal directions, thenE is closed in R2.
Proof Consider (x, y) ! E . There exists a sequence of sets (An)n!N " Ac(B)
such that s(An) # (x, y). From the weak ! -compactness of the set Ac(B), for asubsequence still denoted by the same indices we can write An # A in the weak! -sense. Then A ! Ac(B) and since the eigenvalues of the Laplacian are weakly! -lower semicontinuous we get
"1(A) $ liminfn#%
"1(An) = x and "2(A) $ liminfn#%
"2(An) = y.
From the vertical convexity of E , the vertical segment joining s(A) with the halfline ts(B1) : t & 1 is contained in E . If y < "2(B1) we can find the point("1(A), y) on this segment and using now the horizontal convexity, the segmentjoining ("1(A), y) to ts(B2) : t & 1 is in E . But this segment contains the point(x, y) since "1(A) $ x .
If y & "2(B1), then the horizontal convexity gives directly (x, y) ! E .
Following Lemma 6.4.2 it suffices to prove the convexity of E on vertical andhorizontal directions. For this purpose, we split the proof in two steps:
15
Optimal partitions
We look for a partition Ω1, . . . ,ΩN of Ω whichminimizes a cost of the form
F (Ω1, . . . ,ΩN)
among all partitions of Ω (volume constraints|Ωi| = mi can be added). The cost F couldbe very general, for instance
F (Ω1, . . . ,ΩN) = Φ(λk1
(Ω1), . . . , λkN(ΩN)).
General existence theoremBucur-B.-Henrot Adv. Math. Sci. Appl. ’98
16
RelaxationB., Timofte Adv. Math. Sci. Appl. ’02
Regularity for λ1(Ω1) + · · ·+ λ1(ΩN)Caffarelli, Lin J. Sci. Comp. ’06
Other variants (manifolds, nonlinear, . . . )Conti, Terracini, Verzini JFA ’03, Calc.Var ’05Helffer, Hoffmann-Ostenhof Preprint
Conjecture: dim=2, optimal partitions forλ1(Ω1) + · · ·+ λ1(ΩN) approach as N → ∞a regular exagonal tiling.
17
Replacing the volume constraint |Ω| ≤ m by
a perimeter constraint Per(Ω) ≤ L provides
interesting variants to the spectral optimiza-
tion problem (Bucur-B.-Henrot ’09, Van den
Berg-Iversen ’09). Note that spectral func-
tional are in general not l.s.c. for the L1
convergence.
Theorem Assume that F is as above (de-
creasing and γ-l.s.c.). Then the minimum
problem
minF (Ω) : Per(Ω) ≤ L, Ω ⊂ D
.
admits at least a solution Ωopt.
18
For F (Ω) = λ2(Ω) and D = Rd we have:
• in the case d = 2 Ωopt is convex, of class
C∞, its boundary does not contain any seg-
ment, does not contain any arc of circle,
contains exactly two points where the cur-
vature vanishes.
• In the case d ≥ 3 Ωopt is not convex, is
connected, regularity of Ωopt is not known
(even not if it is an open set).
19
• A similar problem has been studied by Athana-
sopoulos, Caffarelli, Kenig, Salsa (CPAM 2001):
minE(Ω) + Per(Ω) : Ω ⊂ D
where E(Ω) is the Dirichlet energy
E(Ω) = min ∫
D|Dv|2 dx :
v = 0 on D \Ωv = g on ∂D
.
• numerical plot of Ωopt for d = 2 (courtesy
of E. Oudet).
20
We consider now Cheeger-like rescaled shape
optimization problems
minM(Ω)J(Ω) : Ω ⊂ D
where M(Ω) is a scaling factor and J(Ω) a
shape functional. The coercivity condition
limM(Ω)→0
M(Ω)J(Ω) = +∞
is assumed, which means that the scaling is
“above” the scaling invariance, as it happens
in the Cheeger problem where
M(Ω) = Per(Ω), J(Ω) = |Ω|−α, α > 1−1/d.
22
Theorem (B.-Wagner) If M and J are non-negative, fulfill the coercivity condition aboveand• J is γ-l.s.c. and decreasing for inclusion;• M is wγ-l.s.c..the minimum problem above has a solution.
Examples• |Ω|αλk(Ω), α < 2/d
• |Ω|α(C(Ω)
)−1, α < 1 + 2
dwhere C(Ω) is the compliance functional
∫Ω fuΩ dx
•(
Per(Ω))αλk(Ω), α < 2/(d− 1);
•(
Per(Ω))α(
C(Ω))−1
, α < 1 + 2d−1.
23
Note that Per(Ω) is not a wγ-l.s.c. and the
existence proof requires some extra devices.
Necessary conditions of optimality can be
found; for instance in the case of |Ω|α(C(Ω)
)−1
if Ω is an optimal domain in D and u is the
corresponding solution:
|∇u|2 = αC(Ω)|Ω| on ∂Ω ∩D (free boundary);
|∇u|2 ≥ αC(Ω)|Ω| on ∂Ω∩∂D (common bound.).
For instance if D has a corner the optimal
domain Ω cannot fill the entire D.
24
To finish, a challenging open question. Con-sider the problem
minC(Ω)λαk(Ω) : Ω ⊂ D
where α is above the scaling invariance 1 +d/2. Thanks to a result of Kohler-Jobin the
minimum of C(Ω)λ1+d/21 (Ω) is reached by a
ball, and so when α > 1 + d/2 the coercivitycondition
limM(Ω)→0
M(Ω)J(Ω) = +∞
is fulfilled, taking
M(Ω) = C(Ω) and J(Ω) = λαk(Ω).
25
However, the other conditions are not ful-
filled; in particular C(Ω) is not wγ-l.s.c. and,
even if strongly expected, the existence of an
optimal domain is still missing.
Once the existence of an optimal domain is
established, the subsequent step would be
the regularity of the free boundary and the
necessary conditions of optimality.
26