Professional Documents
Culture Documents
Shape Optimization Among Convex Sets
Shape Optimization Among Convex Sets
Beniamin Bogosel
14/06/2022
min J(ω)
ω∈A
Purpose
given a functional and a set of constraints that guarantee existence of an optimal
set: what the optimal shape looks like?
confirm/disprove conjectures, give new directions for theoretical study
Questions
how to discretize shapes?
level-set, phase field, explicit parametrization, polygonal approximation
handle all constraints of interest
evolve of the shape? shape derivative? gradient?
p(θ) = max(x · θ)
x∈ω
? geometric interpretation: distance to the tangent bounding ω in the direction θ
? can naturally handle width constraints: look at p(θ) + p(−θ)!
? Convexity constraint (in 2D): functional inequality p + p00 ≥ 0
p(θ2 )
p(θ1 )
p(θ3 )
Advantages
exact representation of the shape
precise information on geometric elements
can handle efficiently the constraints of interest
convexity
constant width
inclusion
diameter inequalities
Objective
Given a general shape functional, its shape derivative and a set of constraints, find an
approximation of the optimal shape!
Options
The choice of p and p0 gives a parametrization of ω:
spectral decomposition: direct access to p and p0 - limited to strictly convex sets!
direct choice of values for some angle discretization: how to choose p0 rigorously?
[Antunes, B.]
[B., Henrot, Lucardesi]
wi ≤ pi + pi+N/2 ≤ Wi .
? diameter constraint: upper bound D for all directions, lower bound D for one
direction.
Beniamin Bogosel Numerical shape optimization among convex sets 14/25
Example
[Al Sayed, B., Henrot, Nacry, Maximization of the Steklov Eigenvalues with a Diameter
Constraint]
? Maximize the Steklov eigenvalues under diameter constraint:
−∆u = 0 in Ω
∂n u = σk (Ω)u on ∂Ω
? Example: with and without convexity constraint
? applying the same method to other functionals works very well. Why?
Beniamin Bogosel Numerical shape optimization among convex sets 15/25
Discrete configuration
1
Area(∆A1 A2 A3 ) = [6%1 %2 + 12%1 %3 + 6%2 %3 + (p1 − p3 )(%1 − %3 )] h3 + O(h5 )
48
? p(θ) 7→ p(θ) + a cos θ + b sin θ corresponds to a translation with vector (a, b).
? using first order finite differences:
translation does not commute with the discretization process
discrete curvature radii are not preserved under translations
? Choose the simplest option which fixes this:
pi+1 − pi−1 pi+1 + pi−1 − 2pi
p0 (θi ) ≈ , p(θi ) + p00 (θi ) ≈ %i = pi + .
2 sin h 2 − 2 cos h
? observe that the formulas are still consistent as h → 0.
? discrete shapes obtained are convex provided %i ≥ 0!
pi+1
Ai
pi+1 −pi−1
pi 2 sin h
O
pi+1 +pi−1 X(pi , θi )
2 cos h
pi−1
X(pi−1 , θi−1 )
? pi+1 + pi−1 − 2pi cos h ≥ 0: Ai is on the correct side of the tangents at Ai−1 , Ai+1 .
−pi−1
? pi+1
2 sin h is the only tangential coordinate that always produces a point on the correct
side of the tangents at Ai−1 , Ai+1
? all convex shapes can be approximated arbitrarily well with the method proposed!
λ2 (Ω∗ ) = 37.9855
Obtain best known shape
without any effort!
? Minimize λ1 (Ω)/h(Ω)2
? compute the Cheeger constant of Ω at every iteration
? compute the Cheeger set
? compute the associated shape derivatives
? the optimal shape found numerically is the square.