Parametric shape optimization using the Support

Function

Beniamin Bogosel

CMAP, École Polytechnique

in collaboration with Pedro Antunes

Beniamin Bogosel Parametric shape optimization using the Support Function 1/37
Context

min J(ω)
ω∈A

Questions: existence, regularity, find the shape, qualitative results

The behavior of minimization problems under various constraints is not always

obvious: numerical simulations may guide and indicate new paths for
theoretical study

Problems of interest: considering constraints like: convexity, diameter,

inclusion, constant width, minimal width (or any combination...)
Previous works:
intersection of half-spaces [Lachand-Robert, Oudet, 05]
projection on the convexity constraint [Merigot, Oudet, 14]
[Oudet et al. 07,13] width constraints
[Bartels, Wachsmuth 18] convexity constraint: mesh deformation with
research of admissible perturbations

Beniamin Bogosel Parametric shape optimization using the Support Function 2/37
Numerical optimization in this context

Purpose
what the optimal shape looks like?
give new directions for theoretical study
Questions and objectives
how to discretize shapes?
level-set, phase field, explicit parametrization, polygonal
approximation
compute shape functional efficiently
handle the constraints
evolution of the shape? gradient?
starting point for an optimization algorithm?

Beniamin Bogosel Parametric shape optimization using the Support Function 3/37
Possible numerical difficulties

? not all small perturbations are admissible: imagine a flat part of a

convex boundary

? some constraints are non-local: diameter or constant width

[Lachand-Robert, Oudet], [Lamboley et al.]

Beniamin Bogosel Parametric shape optimization using the Support Function 4/37
Possible numerical difficulties

? not all small perturbations are admissible: imagine a flat part of a

convex boundary
? some constraints are non-local: diameter or constant width

[Lachand-Robert, Oudet], [Lamboley et al.]

Beniamin Bogosel Parametric shape optimization using the Support Function 4/37
Support function parametrization

ω ⊂ Rn

p(θ) = max[x · θ]
x∈ω

? geometric interpretation: distance to the tangent bounding ω in the

direction θ
? good for width constraints
? functional characterization of convexity

p(θ2 )

p(θ1 )
p(θ3 )

Beniamin Bogosel Parametric shape optimization using the Support Function 5/37
Explicit parametrization: support function

Advantages
exact representation of the shape
precise information on geometric elements
can use higher accuracy PDE methods
can handle efficiently the constraints of interest
convexity
constant width
inclusion
diameter inequalities
Weak point - works only for convex sets

? Work with P. Antunes: discretize the support function using a spectral

basis (Fourier, spherical harmonics)
? Recent work: parametrize the support function directly using a
discretization of [0, 2π].

Beniamin Bogosel Parametric shape optimization using the Support Function 6/37
Dimension two
Going from the support function to an actual parametrization:

ω ⊂ R2
p(θ) = max[x · (cos θ, sin θ)]
x∈Ω

If ω is strictly convex and ∂ω is Lipschitz then

(
x(θ) = p(θ) cos θ − p 0 (θ) sin θ
y (θ) = p(θ) sin θ + p 0 (θ) cos θ

p(θ2 )

p(θ1 )
p(θ3 )

Beniamin Bogosel Parametric shape optimization using the Support Function 7/37
Convexity constraint: Fourier coeffs.

? equivalent to the functional inequality: p + p 00 ≥ 0

? discretize the
P support functionPusing Fourier coefficients:
p(θ) = a0 + Nk=1 ak cos(kθ) + N k=1 bk sin(kθ)
? analytic characterization for convexity in terms of (ak ), (bk ): needs special
numerical tools - Semi-definite Programming [Bayen, Henrion, 12]
? a simpler approach: [Antunes 17]
−→ impose p(θn ) + p 00 (θn ) ≥ 0 on a discretization of [0, 2π]

a0
 
 a2 
 . 
··· ···  .. 
  0
1 α1,2 α1,N β1,2 β1,N 
 .. .. .. .. .. .. ..  a 

 .
. . . . . . .   N  ≥  .. 
 b2 
1 αM,2 ··· αM,N βM,2 ··· βM,N   0
 . 
 .. 
bN
2 2
where αm,n = (1 − n ) cos(nθm ) and βm,n = (1 − n ) sin(nθm ).

Beniamin Bogosel Parametric shape optimization using the Support Function 8/37
Compute the gradient

Z
F: a functional with shape derivative F 0 (ω)(V ) = f V .ndσ
∂ω
F(ω) = F(a0 , a1 , ..., aN , b1 , ..., bN )
p 7→ p + t cos(kθ)
x→
7 x + t[cos(kθ) cos θ + k sin(kθ) sin θ]
y→7 y + t[cos(kθ) sin θ − k sin(kθ) cos θ]

n(θ) = (cos θ, sin θ) so V .n = cos(kθ)

Jacobian: p + p 00
Formula
Z 2π
dF
= f · (p(θ) + p 00 (θ)) cos(kθ)dθ
dak 0

Beniamin Bogosel Parametric shape optimization using the Support Function 9/37
Dimension three

use Spherical Harmonics decomposition

N X
X l
p(φ, ψ) = al,m Ylm (φ, ψ).
l=0 m=−l

General parametrization

Sd−1 3 u 7→ p(u).u + ∇τ p(u) ∈ Rd ,

? explicit formulas for finding a basis of the tangent space

n = n(φ, ψ) 7→ (sin φ sin ψ, cos φ sin ψ, cos ψ), φ ∈ [−π, π), ψ ∈ [0, π).

pφ (φ, ψ)
xp (φ, ψ) = p(φ, ψ)n + nφ + pψ (φ, ψ)nψ
sin2 ψ

Beniamin Bogosel Parametric shape optimization using the Support Function 10/37
Convexity constraint in 3D

? non-negative principal curvatures

? Hack: it is enough to look at the Gaussian curvature

In dimension three, if the Gaussian curvature of a closed surface is everywhere

non-negative then the surface bounds a convex body

? complicated but computable expression.

   2
pφφ 1 pφ cos ψ
p sin ψ + + pψ cos ψ (p + pψψ ) + − pψφ ≥0
sin ψ sin ψ sin ψ

Beniamin Bogosel Parametric shape optimization using the Support Function 11/37
Constant width constraint

? Dimension two: p(θ) + p(θ + π) = w for θ ∈ [0, 2π]

? Dimension three: p(θ) + p(−θ) = w for θ ∈ S2

The even part of the support function is equal to zero:

a0 = w /2, a2k = b2k = 0 in 2D
a0 = w /2, coeffs of even spherical harmonics are zero in 3D

? other inequalities on the diameter can be introduced imposing them

again at a finite but large enough number of points

Beniamin Bogosel Parametric shape optimization using the Support Function 12/37
Other advantages - numerical side

Computing basic quantities

in 2D area and perimeter can be computed explicitly:
N
X
P(p) = 2πa0 , A(p) = πa02 + π/2 (1 − k 2 )(ak2 + bk2 )
i=1

in 3D things are more complicated: for constant width bodies,

however, there are explicit formulas

h = p − S2 p/H2 (S2 )
R

Volume: V = π6 w 3 − w2 E(h)
Area: A = πw 2 − R E(h)
Energy: E(h) = S2 12 |∇τ h|2 − h2 dA 



PN Pl λ 2
In terms of coeffs: E(h) = l=1 m=−l l,m 2 − 1 al,m

Beniamin Bogosel Parametric shape optimization using the Support Function 13/37
Theoretical questions: convergence?

? FN denotes the class of convex sets for which the support function p
has only the first N Fourier or spherical harmonics coefficients not equal
to zero
? if ωN is the solution of the problem in FN , then as N → ∞ does ωN
converge to a solution of the original problem?
Answer: Yes, but not straightforward.

? consider the case of the area: if pn → p in L2 then Area(pn ) does not

necessarily converge to Area(p). We have upper-semicontinuity, not
lower semicontinuity as usual. We need convergence pf pn → p in H 1 !
? Exploit the convexity using the Blaschke selection theorem and the link
between convergence in Hausdorff metric and convergence of support
functions!
? proofs in the revised version of [Antunes, B., 18]

Beniamin Bogosel Parametric shape optimization using the Support Function 14/37
Numerical Optimization

start from randomized coefficients: avoid local minima

compute gradient with respect to each coefficient
compute gradients of the constraint (if it is non-affine)
use a constrained optimization algorithm:
fmincon in Matlab or IPOPT in FreeFEM

Beniamin Bogosel Parametric shape optimization using the Support Function 15/37
Objective functions studied

? various functionals related to volume, perimeter, Dirichlet-Laplace

eigenvalues λk (ω)

−∆u = λk (ω)u in ω
u = 0 on ∂ω

? constraints: convexity, constant width, minimal width, diameter

inequalities, inclusion
Goal
Show that the method works for a large variety of problems with basic
generally available software.

Beniamin Bogosel Parametric shape optimization using the Support Function 16/37
Example: minimize λk among convex sets

?: why? - λk is non-linear, non-quadratic

?: this computation goes out of the scope of [Bayen, Henrion 12]

λ2 = 38.00 λ4 = 65.28 λ5 = 79.70

λ6 = 88.54 λ7 = 109.44 λ8 = 120.58

λ9 = 137.38 λ10 = 143.15

Beniamin Bogosel Parametric shape optimization using the Support Function 17/37
Example: minimize λk among convex sets

?: why? - λk is non-linear, non-quadratic

?: this computation goes out of the scope of [Bayen, Henrion 12]

λ2 = 38.00 λ4 = 65.28 λ5 = 79.70

λ6 = 88.54 λ7 = 109.44 λ8 = 120.58

λ9 = 137.38 λ10 = 143.15

Remark: Possible problems capturing segments!?

Beniamin Bogosel Parametric shape optimization using the Support Function 17/37
3D case

λ2 = 43.07 λ3 = 49.37 λ5 = 73.81 λ6 = 75.64

λ7 = 80.74 λ8 = 85.29 λ10 = 93.22

Beniamin Bogosel Parametric shape optimization using the Support Function 18/37
Blaschke-Lebesgue

Constraints: convexity+constant width

The Reuleaux triangle minimizes the area at constant width

Beniamin Bogosel Parametric shape optimization using the Support Function 19/37
Constant Width constraint in 3D

the convexity constraint is quadratic (non-linear)

exact Hessian available for the volume
confirmation of Meissner’s conjecture: bodies of constant width and
minimal volume in 3D
−→ starting from random initial coefficients

Beniamin Bogosel Parametric shape optimization using the Support Function 20/37
Constant width: eigenvalues

Minimizers Maximizers

Also confirmed using FreeFEM!

Nice theoretical challenge, even for λ1 !
[B., Henrot, Lucardesi]

Beniamin Bogosel Parametric shape optimization using the Support Function 21/37
Constant width: DL eigenvalues in 3D

? for small indices the ball is often optimal

? below are some examples where the ball is not optimal

λ10 = 39.41 λ46 = 88.80 λ99 = 156.14

λ10 ( ) = 39.48 λ46 ( ) = 88.83 λ99 ( ) = 157.92

Beniamin Bogosel Parametric shape optimization using the Support Function 22/37
Minimal area Rotors

Rotor constraint on the coefficients:

- for n-gon: use only coeffs in (nZ − 1) ∪ (nZ + 1)
- first coeff s.t. 4a0 sin2 (π/n) = 1 gives rotor of width 1
- Applications: drill (almost) polygonal holes

Beniamin Bogosel Parametric shape optimization using the Support Function 23/37
Rotors of minimal volume in 3D

? such rotors only exist for the regular tetrahedron, the regular
octahedron and the cube (bodies of constant width)
? characterization: only spherical harmonics of certain indices are
non-zero

Using spectral decomposition is particularly useful here!

Beniamin Bogosel Parametric shape optimization using the Support Function 24/37
Diameter constraint

[Henrot, Lamboley, Privat]

Jγ (ω) = γ|ω| − H1 (∂ω)

? minimize Jγ when diam(ω) ≤ 1
? results below for γ = 0.5, 0.9, 1.

? polygons are captured quite well for this problem although looking
closely you can see that the segments are not really straight...

Beniamin Bogosel Parametric shape optimization using the Support Function 25/37
Minimal width in 3D

? minimize the total surface area

? minimal width constraint: p(θ) + p(−θ) ≥ 1

Result of [Oudet 13]: 2.9249

Our result: 2.9154
? we wish to do the same thing for the volume: conjecture −→ Heil’s body is optimal
? either Heil’s body is too singular, or the algorithm stops too soon

Beniamin Bogosel Parametric shape optimization using the Support Function 26/37
Inclusion constraint: Cheeger sets

B1 ⊂ B2 ⇔ pB1 (θ) ≤ pB2 (θ) for every θ ∈ Sn−1

Hn−1 (∂ω)
min
ω⊂D, ω convex |ω|
? for polyhedra only one constraint per face is necessary

? presence of flat parts, difficult to capture using smooth support

functions generated by spherical harmonics!

Beniamin Bogosel Parametric shape optimization using the Support Function 27/37
Summary

Constraints
convexity constant-width diameter
2D linear ineq. coeffs. linear ineq.
3D quadratic coeffs. linear ineq.

Functionals
Volume Eigenvalues
2D quadratic non-linear
3D quadratic non-linear

? the method can handle arbitrarily complex functionals

? affine, quadratic (nonlinear) constraints

? Problems: capture segments or singularities which often appear when

imposing convexity!

Beniamin Bogosel Parametric shape optimization using the Support Function 28/37
Ongoing work

? Deal with segments in the boundary: the support function is not

smooth there −→ using Fourier coeffs not a good idea!
? how about working directly with p(θ0 ), ..., p(θN ) avec
θi = 2πi
N , i = 0, ..., N − 1.??
? approximate p 0 , p 00 using finite differences
pi+1 − pi−1 pi+1 + pi−1 − 2pi
p 0 (θi ) = , p 00 (θi ) = .
2h h2
? Shape derivative −→ functional derivative

Z Z 2π
J 0 (Ω)(V ) = f V .n ←→ J 0 (p)(ϕ) = f ϕ(θ)(p(θ) + p 00 (θ))dθ
∂Ω 0

? Discrete variant:
∂J 2π
= f (θi )(p(θi ) + p 00 (θi )).
∂pi N

Beniamin Bogosel Parametric shape optimization using the Support Function 29/37
Ongoing work

? Deal with segments in the boundary: the support function is not

smooth there −→ using Fourier coeffs not a good idea!
? how about working directly with p(θ0 ), ..., p(θN ) avec
θi = 2πi
N , i = 0, ..., N − 1.??
? approximate p 0 , p 00 using finite differences
pi+1 − pi−1 pi+1 + pi−1 − 2pi
p 0 (θi ) = , p 00 (θi ) = .
2h h2
? Shape derivative −→ functional derivative

Z Z 2π
0 0
J (Ω)(V ) = f V .n ←→ J (p)(ϕ) = f ϕ(θ)(p(θ) + p 00 (θ))dθ
∂Ω 0

? Discrete variant:
∂J 2π
= f (θi )(p(θi ) + p 00 (θi )). WRONG!
∂pi N

Beniamin Bogosel Parametric shape optimization using the Support Function 29/37
Ongoing work

? Deal with segments in the boundary: the support function is not

smooth there −→ using Fourier coeffs not a good idea!
? how about working directly with p(θ0 ), ..., p(θN ) avec
θi = 2πi
N , i = 0, ..., N − 1.??
? approximate p 0 , p 00 using finite differences
pi+1 − pi−1 pi+1 + pi−1 − 2pi
p 0 (θi ) = , p 00 (θi ) = .
2h h2
? Shape derivative −→ functional derivative

Z Z 2π
0 0
J (Ω)(V ) = f V .n ←→ J (p)(ϕ) = f ϕ(θ)(p(θ) + p 00 (θ))dθ
∂Ω 0

? Discrete variant:
Z
∂J
= χ[θi −h/2,θi +h/2] f . Is this ok?
∂pi ∂Ω

Beniamin Bogosel Parametric shape optimization using the Support Function 29/37
Does it work?
? yes (tested in 2D), imposing convexity as p(θi ) + p 00 (θi ) ≥ 0.
? solvers: fmincon in Matlab or IPOPT in FreeFEM
? no problems with segments in the boundary: one segment corresponds
to one θi
Example 1: min Area(ω) for ω with width ≥ 1: equilateral triangle

Beniamin Bogosel Parametric shape optimization using the Support Function 30/37
Blaschke-Lebesgue
? minimize area under constant width condition

p(θi ) + p(θi + π) = D, ∀i
p(θi ) + p 0 (θi ) ≥ 0

? Note that the Reuleaux triangle has smooth support function despite
the corners!
Beniamin Bogosel Parametric shape optimization using the Support Function 31/37
Another classical problem

min λ2 (ω) · Area(ω)

ω convex
[Oudet 03],[Antunes, Henrot, 11]
? the minimizer seems to have two parallel segments in its boundary
? Antunes and Henrot did a parametric search for the min with
fundamental solutions and found 37.987
? using Fourier coeffs: 38.00 - segments cannot be captured here since
the support function is smooth!

? using direct discretization with finite elements in FreeFEM: 37.987 -

Same result as [Antunes, Henrot] without any assumption on the shape
apart from convexity

Beniamin Bogosel Parametric shape optimization using the Support Function 32/37
Test cases in [Bartels, Wachsmuth, 18]

? min ω u, where −∆u = f , u ∈ H01 (ω), with ω convex

R

? Singular optimal shapes: optimal values are -6.120356e-03,

-2.762311e-02

? Results obtained with supp. func.: optimal values: -6.12125e-03,

-2.76101e-02

Beniamin Bogosel Parametric shape optimization using the Support Function 33/37
Ongoing work

? question of A. Henrot: maximize the Steklov eigenvalues under

diameter constraint and convexity

max σk (ω)
diam(ω)=c

−∆u = 0 in ω
∂n u = σk (ω)u on ∂ω
? diameter constraint can be imposed in the following way:

p(θi ) + p(θi + π) ≤ D, ∀i

p(θ1 ) + p(θ1 + π) = D

? the question remains valid when removing the convexity assumption:

non-degenerate shapes exist. How to deal numerically with this case?
The support function cannot be used here.

Beniamin Bogosel Parametric shape optimization using the Support Function 34/37
Ongoing work (2) - 3D case

? Choose a uniform distribution of points (φi , ψi ) on the sphere so that

finite differences can be used
? Discretizing the convexity constraint!
   2
pφφ 1 pφ cos ψ
p sin ψ + + pψ cos ψ (p + pψψ ) + − pψφ ≥0
sin ψ sin ψ sin ψ

? Try and test Heil’s conjecture [Oudet, 13] concerning bodies of given
minimal width and minimal volume.

Beniamin Bogosel Parametric shape optimization using the Support Function 35/37
Conclusions

support function parametrization can handle complex non-local

constraints
discrete version of the constraints is enough
use Fourier coeffs for smooth support functions
use direct discretization to capture non-smooth behavior: segments
in the boundary
complex problems can be tackled using fmincon in Matlab and
IPOPT in FreeFEM

Beniamin Bogosel Parametric shape optimization using the Support Function 36/37
Questions
1. Parametric derivative:
Z Z 2π
J 0 (Ω)(V ) = f V .n ←→ J 0 (p)(ϕ) = f ϕ(θ)(p(θ) + p 00 (θ))dθ
∂Ω 0
Z
∂J
= χ[θi −h/2,θi +h/2] f . Is this ok?
∂pi ∂Ω
2. How to optimize among sets which have a given diameter, but are not
necessarily convex?
Possible Idea: Optimize the couple (ω, chull(ω)), using the support
function for chull(ω) and something else for ω?
3. Extension to 3D:
? Choose a uniform distribution of points (φi , ψi ) on the sphere so that
finite differences can be used
? Discretizing the convexity constraint!
   2
pφφ 1 pφ cos ψ
p sin ψ + + pψ cos ψ (p + pψψ ) + − pψφ ≥0
sin ψ sin ψ sin ψ

Beniamin Bogosel Parametric shape optimization using the Support Function 37/37