Convex discretization of functionals involving the Monge-Amp`ere operator
Quentin M´erigot
CNRS / Universit´e Paris-Dauphine
Joint work with J.D. Benamou, G. Carlier and ´E. Oudet
Workshop on Optimal Transport in the Applied Sciences
1. Motivation: Gradient flows in Wasserstein space
Background: Optimal transport
I Wasserstein distance between µ, ν ∈ P2(Rd), P2(Rd) = {µ ∈ P(Rd); R
kxk2 d µ(x) < +∞}
Γ(µ, ν) := {π ∈ P(Rd × Rd); p1#π = µ, p2#π = ν} P2ac(Rd) = P2(Rd) ∩ L1(Rd)
Rd
Rd π
µ ν
p1 p2
Definition: W22(µ, ν) := minπ∈Γ(µ,ν) R
kx − yk2 d π(x, y).
Background: Optimal transport
I Wasserstein distance between µ, ν ∈ P2(Rd), P2(Rd) = {µ ∈ P(Rd); R
kxk2 d µ(x) < +∞}
Γ(µ, ν) := {π ∈ P(Rd × Rd); p1#π = µ, p2#π = ν} P2ac(Rd) = P2(Rd) ∩ L1(Rd)
Rd
Rd π
µ ν
p1 p2
I Relation to convex functions: Def: K := finite convex functions on Rd Definition: W22(µ, ν) := minπ∈Γ(µ,ν) R
kx − yk2 d π(x, y).
Background: Optimal transport
Theorem (Brenier): Given µ ∈ P2ac(Rd) the map φ ∈ K 7→ ∇φ#µ ∈ P2(Rd) is surjective and moreover,
I Wasserstein distance between µ, ν ∈ P2(Rd), P2(Rd) = {µ ∈ P(Rd); R
kxk2 d µ(x) < +∞}
Γ(µ, ν) := {π ∈ P(Rd × Rd); p1#π = µ, p2#π = ν} P2ac(Rd) = P2(Rd) ∩ L1(Rd)
Rd
Rd π
µ ν
p1 p2
[Brenier ’91]
I Relation to convex functions: Def: K := finite convex functions on Rd Definition: W22(µ, ν) := minπ∈Γ(µ,ν) R
kx − yk2 d π(x, y).
W22(µ, ∇φ#µ) = R
Rd kx − ∇φ(x)k2 d µ(x)
Background: Optimal transport
Theorem (Brenier): Given µ ∈ P2ac(Rd) the map φ ∈ K 7→ ∇φ#µ ∈ P2(Rd) is surjective and moreover,
I Wasserstein distance between µ, ν ∈ P2(Rd), P2(Rd) = {µ ∈ P(Rd); R
kxk2 d µ(x) < +∞}
Γ(µ, ν) := {π ∈ P(Rd × Rd); p1#π = µ, p2#π = ν} P2ac(Rd) = P2(Rd) ∩ L1(Rd)
Rd
Rd π
µ ν
p1 p2
[Brenier ’91]
I Relation to convex functions: Def: K := finite convex functions on Rd Definition: W22(µ, ν) := minπ∈Γ(µ,ν) R
kx − yk2 d π(x, y).
Given any µ ∈ P2ac(Rd), we get a ”parameterization” of P2(Rd), as ”seen” from µ.
W22(µ, ∇φ#µ) = R
Rd kx − ∇φ(x)k2 d µ(x)
Motivation 1: Crowd Motion Under Congestion
[Maury-Roudneff-Chupin-Santambrogio 10]
ρτk+1 = minσ∈P2(X) 2τ1 W22(ρτk, σ) + E(σ) + U(σ) I JKO scheme for crowd motion with hard congestion:
U(ν) :=
( 0 if d ν = f d Hd, f ≤ 1
+ ∞ if not congestion potential energy
E(ν) := R
Rd V (x) d ν(x) (∗)
where X ⊆ Rd is convex and bounded, and
Motivation 1: Crowd Motion Under Congestion
[Maury-Roudneff-Chupin-Santambrogio 10]
minφ 2τ1 R
Rd kx − ∇φ(x)k2ρτk(x) d x + E(∇φ#ρτk) + U(∇φ#ρτk) ρτk+1 = minσ∈P2(X) 2τ1 W22(ρτk, σ) + E(σ) + U(σ)
I JKO scheme for crowd motion with hard congestion:
U(ν) :=
( 0 if d ν = f d Hd, f ≤ 1
+ ∞ if not congestion potential energy
E(ν) := R
Rd V (x) d ν(x)
I Assuming σ = ∇φ#ρτk with φ convex, the Wasserstein term becomes explicit:
On the other hand, the constraint becomes strongly nonlinear:
(∗)
(∗) ⇐⇒
U(∇φ ρτ) < +∞ ⇐⇒ det D2φ(x) ≥ ρ (x) where X ⊆ Rd is convex and bounded, and
Motivation 2: Nonlinear Diffusion
∂ρ
∂t = div [ρ∇(U0(ρ) + V + W ∗ ρ)] ρ(0, .) = ρ0
(∗) ρ(t, .) ∈ Pac(Rd)
Motivation 2: Nonlinear Diffusion
∂ρ
∂t = div [ρ∇(U0(ρ) + V + W ∗ ρ)] ρ(0, .) = ρ0 (∗)
I Formally, (∗) can be seen as the W2-gradient flow of U + E, with
ρ(t, .) ∈ Pac(Rd)
U(ν) :=
Z
Rd
U(f(x)) d x if d ν = f d Hd + ∞ if not
E(ν) := R
Rd V (x) d ν(x) + R
Rd W (x − y) d[ν ⊗ ν](x, y)
internal energy, ex: U(r) = r log r =⇒ entropy
potential energy interaction energy
Motivation 2: Nonlinear Diffusion
∂ρ
∂t = div [ρ∇(U0(ρ) + V + W ∗ ρ)]
I JKO time discrete scheme: for τ > 0,
ρ(0, .) = ρ0 (∗)
ρτk+1 = arg minσ∈P(Rd) 2τ1 W2(ρτk, σ)2 + U(σ) + E(σ) I Formally, (∗) can be seen as the W2-gradient flow of U + E, with
ρ(t, .) ∈ Pac(Rd)
[Jordan-Kinderlehrer-Otto ’98]
U(ν) :=
Z
Rd
U(f(x)) d x if d ν = f d Hd + ∞ if not
E(ν) := R
Rd V (x) d ν(x) + R
Rd W (x − y) d[ν ⊗ ν](x, y)
internal energy, ex: U(r) = r log r =⇒ entropy
potential energy interaction energy
Motivation 2: Nonlinear Diffusion
∂ρ
∂t = div [ρ∇(U0(ρ) + V + W ∗ ρ)]
I JKO time discrete scheme: for τ > 0,
ρ(0, .) = ρ0 (∗)
ρτk+1 = arg minσ∈P(Rd) 2τ1 W2(ρτk, σ)2 + U(σ) + E(σ)
−→ Many applications: porous medium equation, cell movement via chemotaxis, I Formally, (∗) can be seen as the W2-gradient flow of U + E, with
ρ(t, .) ∈ Pac(Rd)
[Jordan-Kinderlehrer-Otto ’98]
U(ν) :=
Z
Rd
U(f(x)) d x if d ν = f d Hd + ∞ if not
E(ν) := R
Rd V (x) d ν(x) + R
Rd W (x − y) d[ν ⊗ ν](x, y)
internal energy, ex: U(r) = r log r =⇒ entropy
potential energy interaction energy
Displacement Convex Setting
minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν) For X convex bounded and µ ∈ Pac(X),
spt(ν) ⊆ X
Displacement Convex Setting
minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν) For X convex bounded and µ ∈ Pac(X),
(∗X)
minφ∈KX 2τ1 W22(µ, ∇φ#µ) + U(∇φ#µ) + E(∇φ#µ)
⇐⇒
KX := {φ convex ; ∇φ ∈ X}
Displacement Convex Setting
When is the minimization problem (∗X) convex ?
E(ν) := R
Rd(V + W ∗ ν) d ν U(ν) := R
Rd U ddHνd
d x minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν)
For X convex bounded and µ ∈ Pac(X),
(∗X)
minφ∈KX 2τ1 W22(µ, ∇φ#µ) + U(∇φ#µ) + E(∇φ#µ)
⇐⇒
KX := {φ convex ; ∇φ ∈ X}
Displacement Convex Setting
When is the minimization problem (∗X) convex ?
E(ν) := R
Rd(V + W ∗ ν) d ν U(ν) := R
Rd U ddHνd
d x minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν)
For X convex bounded and µ ∈ Pac(X),
(∗X)
minφ∈KX 2τ1 W22(µ, ∇φ#µ) + U(∇φ#µ) + E(∇φ#µ)
⇐⇒
NB: U(∇φ#ρ) = R
U
ρ(x) MA[φ](x)
MA[φ](x) d x MA[φ](x) := det(D2φ(x)) KX := {φ convex ; ∇φ ∈ X}
(H2) rdU(r−d) is convex non-increasing, U(0) = 0.
Theorem: (∗X) is convex if (H1) V, W : Rd → R are convex functions,
[McCann ’94]
Displacement Convex Setting
When is the minimization problem (∗X) convex ?
E(ν) := R
Rd(V + W ∗ ν) d ν U(ν) := R
Rd U ddHνd
d x minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν)
For X convex bounded and µ ∈ Pac(X),
(∗X)
minφ∈KX 2τ1 W22(µ, ∇φ#µ) + U(∇φ#µ) + E(∇φ#µ)
⇐⇒
NB: U(∇φ#ρ) = R
U
ρ(x) MA[φ](x)
MA[φ](x) d x MA[φ](x) := det(D2φ(x)) KX := {φ convex ; ∇φ ∈ X}
(H2) rdU(r−d) is convex non-increasing, U(0) = 0.
Theorem: (∗X) is convex if (H1) V, W : Rd → R are convex functions,
[McCann ’94]
Numerical applications of the JKO scheme
I Numerical applications of the (variational) JKO scheme are still limited:
Numerical applications of the JKO scheme
1D: monotone rearrangement e.g. [Kinderleherer-Walkington 99]
I Numerical applications of the (variational) JKO scheme are still limited:
[Blanchet-Calvez-Carrillo 08] [Agueh-Bowles 09]
Numerical applications of the JKO scheme
1D: monotone rearrangement e.g. [Kinderleherer-Walkington 99]
2D: optimal transport plans → diffeomorphisms [Carrillo-Moll 09]
U = hard congestion term [Maury-Roudneff-Chupin-Santambrogio 10]
I Numerical applications of the (variational) JKO scheme are still limited:
[Blanchet-Calvez-Carrillo 08] [Agueh-Bowles 09]
Numerical applications of the JKO scheme
I Our goal is to approximate a JKO step numerically, in dimension ≥ 2:
minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν)
(∗X)
1D: monotone rearrangement e.g. [Kinderleherer-Walkington 99]
2D: optimal transport plans → diffeomorphisms [Carrillo-Moll 09]
U = hard congestion term [Maury-Roudneff-Chupin-Santambrogio 10]
I Numerical applications of the (variational) JKO scheme are still limited:
For X convex bounded and µ ∈ Pac(X),
[Blanchet-Calvez-Carrillo 08] [Agueh-Bowles 09]
Numerical applications of the JKO scheme
I Our goal is to approximate a JKO step numerically, in dimension ≥ 2:
Part 2: Convex discretization of the problem (∗X) under McCann’s hypotheses.
Part 3: Γ-convergence results from the discrete problem to the continuous one.
Part 4: Numerical simulations: non-linear diffusion, crowd motion.
minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν)
(∗X)
1D: monotone rearrangement e.g. [Kinderleherer-Walkington 99]
2D: optimal transport plans → diffeomorphisms [Carrillo-Moll 09]
U = hard congestion term [Maury-Roudneff-Chupin-Santambrogio 10]
I Numerical applications of the (variational) JKO scheme are still limited:
For X convex bounded and µ ∈ Pac(X),
I Plan of the talk:
[Blanchet-Calvez-Carrillo 08] [Agueh-Bowles 09]
2. Convex discretization of a JKO step
Prior work: Discretizing the Space of Convex Functions
minφ∈K R
X F (φ(x), ∇φ(x)) d µ(x)
I Finite elements: piecewise-linear convex functions K := {φ convex}
−→ number of constraints is ∼ N := card(P )
P ⊆ X
over a fixed mesh
Prior work: Discretizing the Space of Convex Functions
minφ∈K R
X F (φ(x), ∇φ(x)) d µ(x)
[Chon´e-Le Meur ’99]
I Finite elements: piecewise-linear convex functions K := {φ convex}
−→ non-convergence result
−→ number of constraints is ∼ N := card(P )
P ⊆ X
over a fixed mesh
Prior work: Discretizing the Space of Convex Functions
−→ convergent, but number of constraints is ' N2
[Ekeland—Moreno-Bromberg ’10]
minφ∈K R
X F (φ(x), ∇φ(x)) d µ(x)
[Chon´e-Le Meur ’99]
I Finite elements: piecewise-linear convex functions
I Finite differences, using convex interpolates
K := {φ convex}
[Carlier-Lachand-Robert-Maury ’01]
−→ non-convergence result
−→ number of constraints is ∼ N := card(P )
P ⊆ X
over a fixed mesh
Prior work: Discretizing the Space of Convex Functions
−→ convergent, but number of constraints is ' N2
[Ekeland—Moreno-Bromberg ’10]
[Mirebeau ’14]
minφ∈K R
X F (φ(x), ∇φ(x)) d µ(x)
[Chon´e-Le Meur ’99]
I Finite elements: piecewise-linear convex functions
I Finite differences, using convex interpolates
[Oudet-M. ’14]
[Oberman ’14]
−→ exterior parameterization
K := {φ convex}
[Carlier-Lachand-Robert-Maury ’01]
−→ non-convergence result
−→ adaptive method
−→ number of constraints is ∼ N := card(P )
P ⊆ X
over a fixed mesh
Prior work: Discretizing the Space of Convex Functions
−→ convergent, but number of constraints is ' N2
[Ekeland—Moreno-Bromberg ’10]
[Mirebeau ’14]
minφ∈K R
X F (φ(x), ∇φ(x)) d µ(x)
[Chon´e-Le Meur ’99]
I Finite elements: piecewise-linear convex functions
I Finite differences, using convex interpolates
[Oudet-M. ’14]
[Oberman ’14]
−→ exterior parameterization
K := {φ convex}
[Carlier-Lachand-Robert-Maury ’01]
−→ non-convergence result
−→ adaptive method
−→ number of constraints is ∼ N := card(P )
P ⊆ X
over a fixed mesh
Discretizing the Space of Convex Functions
minφ∈KX 2τ1 W22(µ, ∇φ#µ) + E(∇φ#µ) + U(∇φ#µ) KX := {φ cvx ; ∇φ ∈ X}
Definition: Given P ⊆ X finite, KX(P ) = {ψ|P ; ψ ∈ KX}
= φ : P → R
P ⊆ X
Discretizing the Space of Convex Functions
minφ∈KX 2τ1 W22(µ, ∇φ#µ) + E(∇φ#µ) + U(∇φ#µ) KX := {φ cvx ; ∇φ ∈ X}
Definition: Given P ⊆ X finite, KX(P ) = {ψ|P ; ψ ∈ KX}
−→ finite-dimensional convex set
= φ : P → R
P ⊆ X
Discretizing the Space of Convex Functions
minφ∈KX 2τ1 W22(µ, ∇φ#µ) + E(∇φ#µ) + U(∇φ#µ) KX := {φ cvx ; ∇φ ∈ X}
Definition: Given P ⊆ X finite, KX(P ) = {ψ|P ; ψ ∈ KX}
−→ finite-dimensional convex set
−→ extension of φ ∈ KX(P ): = φ : P → R
φˆ
P ⊆ X
φˆ := max{ψ; ψ ∈ KX and ψ|P ≤ φ} ∈ KX
φ : P → R φˆ : Rd → R ∈ KX
X = [−1, 1]
Discretizing the Space of Convex Functions
minφ∈KX 2τ1 W22(µ, ∇φ#µ) + E(∇φ#µ) + U(∇φ#µ) KX := {φ cvx ; ∇φ ∈ X}
Definition: Given P ⊆ X finite, KX(P ) = {ψ|P ; ψ ∈ KX}
−→ finite-dimensional convex set
−→ extension of φ ∈ KX(P ): = φ : P → R
φˆ
P ⊆ X
φˆ := max{ψ; ψ ∈ KX and ψ|P ≤ φ} ∈ KX
Discrete push-forward of µP = P
p∈P µpδp ∈ P(P)
φ : P → R φˆ : Rd → R ∈ KX
X = [−1, 1]
Discretizing the Space of Convex Functions
minφ∈KX 2τ1 W22(µ, ∇φ#µ) + E(∇φ#µ) + U(∇φ#µ) KX := {φ cvx ; ∇φ ∈ X}
Definition: Given P ⊆ X finite, KX(P ) = {ψ|P ; ψ ∈ KX}
−→ finite-dimensional convex set
−→ extension of φ ∈ KX(P ): = φ : P → R
φˆ
P ⊆ X
φˆ := max{ψ; ψ ∈ KX and ψ|P ≤ φ} ∈ KX
p
Discrete push-forward of µP = P
p∈P µpδp ∈ P(P) Definition: For φ ∈ KX(P ), let Vp := ∂φ(p)ˆ and
φ : P → R φˆ : Rd → R ∈ KX
X = [−1, 1]
Discretizing the Space of Convex Functions
minφ∈KX 2τ1 W22(µ, ∇φ#µ) + E(∇φ#µ) + U(∇φ#µ) KX := {φ cvx ; ∇φ ∈ X}
Definition: Given P ⊆ X finite, KX(P ) = {ψ|P ; ψ ∈ KX}
−→ finite-dimensional convex set
−→ extension of φ ∈ KX(P ): = φ : P → R
φˆ
P ⊆ X
φˆ := max{ψ; ψ ∈ KX and ψ|P ≤ φ} ∈ KX
X Gφ#µP := P
p∈P
µp
Hd(Vp) Hd
Vp ∈ Pac(X) p
Vp
Discrete push-forward of µP = P
p∈P µpδp ∈ P(P) Definition: For φ ∈ KX(P ), let Vp := ∂φ(p)ˆ and
φ : P → R φˆ : Rd → R ∈ KX
X = [−1, 1]
Convex Space-Discretization of One JKO Step
Theorem: Under McCann’s hypothesis (H1) and (H2), the problem minφ∈KG
X(P) 1
2τ W22(µP , Hφ#µP ) + E(Hφ#µP ) + U(Gφ#µP ) is convex, and the minimum is unique if rdU(r−d) is strictly convex.
[Benamou, Carlier, M., Oudet ’14]
Convex Space-Discretization of One JKO Step
Theorem: Under McCann’s hypothesis (H1) and (H2), the problem minφ∈KG
X(P) 1
2τ W22(µP , Hφ#µP ) + E(Hφ#µP ) + U(Gφ#µP ) is convex, and the minimum is unique if rdU(r−d) is strictly convex.
[Benamou, Carlier, M., Oudet ’14]
I Unfortunately, two different definitions of push-forward seem necessary.
Convex Space-Discretization of One JKO Step
Theorem: Under McCann’s hypothesis (H1) and (H2), the problem minφ∈KG
X(P) 1
2τ W22(µP , Hφ#µP ) + E(Hφ#µP ) + U(Gφ#µP ) is convex, and the minimum is unique if rdU(r−d) is strictly convex.
[Benamou, Carlier, M., Oudet ’14]
I Unfortunately, two different definitions of push-forward seem necessary.
of the discrete Monge-Amp`ere operator:
I The convexity of φ ∈ KX(P ) 7→ U(Gφ#µP) follows from the log-concavity MA[φ](p) := Hd(∂φ(p)).ˆ
Convex Space-Discretization of One JKO Step
Theorem: Under McCann’s hypothesis (H1) and (H2), the problem minφ∈KG
X(P) 1
2τ W22(µP , Hφ#µP ) + E(Hφ#µP ) + U(Gφ#µP ) is convex, and the minimum is unique if rdU(r−d) is strictly convex.
U(Gφ#µP) < +∞ =⇒ ∀p ∈ P, MA[φ](p) = Hd(∂φ(p))ˆ > 0
I If limr→∞ U(r)/r = +∞, the internal energy is a barrier for convexity, i.e.
=⇒ φ is in the interior of KX(P ).
2
[Benamou, Carlier, M., Oudet ’14]
I Unfortunately, two different definitions of push-forward seem necessary.
of the discrete Monge-Amp`ere operator:
I The convexity of φ ∈ KX(P ) 7→ U(Gφ#µP) follows from the log-concavity MA[φ](p) := Hd(∂φ(p)).ˆ
Convex Space-Discretization of the Internal Energy
Proposition: Under McCann’s assumption, φ ∈ KX(P ) 7→ U(Gφ#µP ) is convex.
Convex Space-Discretization of the Internal Energy
Proposition: Under McCann’s assumption, φ ∈ KX(P ) 7→ U(Gφ#µP ) is convex.
Discrete Monge-Amp`ere operator: MA[φ](p) := Hd(∂φ(p)).ˆ
Convex Space-Discretization of the Internal Energy
U(Gacφ#µP ) = P
p∈P U (µp/MA[φ](p)) MA[φ](p)
Proposition: Under McCann’s assumption, φ ∈ KX(P ) 7→ U(Gφ#µP ) is convex.
Proof: Recall Gφ#µP = P
p∈P [µp/MA[φ](p)]1∂φ(p)ˆ . With U(σ) = R
U(σ(x)) d x, Discrete Monge-Amp`ere operator: MA[φ](p) := Hd(∂φ(p)).ˆ
NB: similarity with U(∇φ#ρ) = R
U
ρ(x) MA[φ](x)
MA[φ](x) d x
Convex Space-Discretization of the Internal Energy
U(Gacφ#µP ) = P
p∈P U (µp/MA[φ](p)) MA[φ](p)
Proposition: Under McCann’s assumption, φ ∈ KX(P ) 7→ U(Gφ#µP ) is convex.
Proof: Recall Gφ#µP = P
p∈P [µp/MA[φ](p)]1∂φ(p)ˆ . With U(σ) = R
U(σ(x)) d x,
= − P
p∈P µp log(MA[φ](p)) + P
p∈P µp log(µp) case U(r) = r log r
Discrete Monge-Amp`ere operator: MA[φ](p) := Hd(∂φ(p)).ˆ
Convex Space-Discretization of the Internal Energy
U(Gacφ#µP ) = P
p∈P U (µp/MA[φ](p)) MA[φ](p)
Proposition: Under McCann’s assumption, φ ∈ KX(P ) 7→ U(Gφ#µP ) is convex.
Proof: Recall Gφ#µP = P
p∈P [µp/MA[φ](p)]1∂φ(p)ˆ . With U(σ) = R
U(σ(x)) d x,
= − P
p∈P µp log(MA[φ](p)) + P
p∈P µp log(µp)
Lemma: Given φ0, φ1 ∈ KX(P ), φt := (1 − t)φ0 + tφ1 and
∂φˆt(p) ⊇ (1 − t)∂φˆ0(p) ⊕ t∂φˆ1(p).
case U(r) = r log r
⊕ = Minkowski sum Discrete Monge-Amp`ere operator: MA[φ](p) := Hd(∂φ(p)).ˆ
Convex Space-Discretization of the Internal Energy
U(Gacφ#µP ) = P
p∈P U (µp/MA[φ](p)) MA[φ](p)
Proposition: Under McCann’s assumption, φ ∈ KX(P ) 7→ U(Gφ#µP ) is convex.
Proof: Recall Gφ#µP = P
p∈P [µp/MA[φ](p)]1∂φ(p)ˆ . With U(σ) = R
U(σ(x)) d x,
= − P
p∈P µp log(MA[φ](p)) + P
p∈P µp log(µp)
log Hd(∂φˆt) ≥ log Hd((1 − t)∂φˆ0(p) ⊕ t∂φˆ1(p)).
≥ (1 − t) log Hd(∂φˆ0(p)) + t log Hd( ˆφ1(p))).
Brunn-Minkowski
Lemma: Given φ0, φ1 ∈ KX(P ), φt := (1 − t)φ0 + tφ1 and
∂φˆt(p) ⊇ (1 − t)∂φˆ0(p) ⊕ t∂φˆ1(p).
case U(r) = r log r
⊕ = Minkowski sum
=⇒
Discrete Monge-Amp`ere operator: MA[φ](p) := Hd(∂φ(p)).ˆ
Convex Space-Discretization of the Internal Energy
U(Gacφ#µP ) = P
p∈P U (µp/MA[φ](p)) MA[φ](p)
Proposition: Under McCann’s assumption, φ ∈ KX(P ) 7→ U(Gφ#µP ) is convex.
Proof: Recall Gφ#µP = P
p∈P [µp/MA[φ](p)]1∂φ(p)ˆ . With U(σ) = R
U(σ(x)) d x,
= − P
p∈P µp log(MA[φ](p)) + P
p∈P µp log(µp)
log Hd(∂φˆt) ≥ log Hd((1 − t)∂φˆ0(p) ⊕ t∂φˆ1(p)).
≥ (1 − t) log Hd(∂φˆ0(p)) + t log Hd( ˆφ1(p))).
Brunn-Minkowski
Lemma: Given φ0, φ1 ∈ KX(P ), φt := (1 − t)φ0 + tφ1 and
∂φˆt(p) ⊇ (1 − t)∂φˆ0(p) ⊕ t∂φˆ1(p).
case U(r) = r log r
⊕ = Minkowski sum
=⇒
Discrete Monge-Amp`ere operator: MA[φ](p) := Hd(∂φ(p)).ˆ
3. A Γ-convergence result
Convergence of the Space-Discretization
Setting: X bounded and convex, µ ∈ Pac(X) with density c−1 ≤ ρ ≤ c minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν) (∗)
Convergence of the Space-Discretization
Setting: X bounded and convex, µ ∈ Pac(X) with density c−1 ≤ ρ ≤ c
(C1) E continuous, U l.s.c on (P(X), W2)
minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν) (∗)
Convergence of the Space-Discretization
Setting: X bounded and convex, µ ∈ Pac(X) with density c−1 ≤ ρ ≤ c
(C1) E continuous, U l.s.c on (P(X), W2)
minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν) (∗)
(C2) U(ρ) = R
U(ρ(x)) d x, where U ≥ M is convex.
Convergence of the Space-Discretization
Setting: X bounded and convex, µ ∈ Pac(X) with density c−1 ≤ ρ ≤ c
(C1) E continuous, U l.s.c on (P(X), W2)
minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν) (∗)
(C2) U(ρ) = R
U(ρ(x)) d x, where U ≥ M is convex.
6= McCann’s condition for displacement-convexity
Convergence of the Space-Discretization
Setting: X bounded and convex, µ ∈ Pac(X) with density c−1 ≤ ρ ≤ c
(C1) E continuous, U l.s.c on (P(X), W2)
minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν) (∗)
Theorem: Let Pn ⊆ X finite, µn ∈ P(Pn) with lim W2(µn, µ) = 0, and:
minφ∈KG
X(Pn) 1
2τ W2(µn, Hφ#µn) + E(Hφ#µn) + U(Gφ#µn) (∗)n If φn minimizes (∗)n, then (Gφn#µn) is a minimizing sequence for (∗).
(C2) U(ρ) = R
U(ρ(x)) d x, where U ≥ M is convex.
6= McCann’s condition for displacement-convexity
[Benamou, Carlier, M., Oudet ’14]
Convergence of the Space-Discretization
Setting: X bounded and convex, µ ∈ Pac(X) with density c−1 ≤ ρ ≤ c
(C1) E continuous, U l.s.c on (P(X), W2)
minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν) (∗)
Theorem: Let Pn ⊆ X finite, µn ∈ P(Pn) with lim W2(µn, µ) = 0, and:
minφ∈KG
X(Pn) 1
2τ W2(µn, Hφ#µn) + E(Hφ#µn) + U(Gφ#µn) (∗)n If φn minimizes (∗)n, then (Gφn#µn) is a minimizing sequence for (∗).
(C2) U(ρ) = R
U(ρ(x)) d x, where U ≥ M is convex.
6= McCann’s condition for displacement-convexity
I Proof relies on Caffarelli’s regularity theorem.
[Benamou, Carlier, M., Oudet ’14]
Convergence of the Space-Discretization
Setting: X bounded and convex, µ ∈ Pac(X) with density c−1 ≤ ρ ≤ c
(C1) E continuous, U l.s.c on (P(X), W2)
minν∈P(X) 2τ1 W22(µ, ν) + E(ν) + U(ν) (∗)
Theorem: Let Pn ⊆ X finite, µn ∈ P(Pn) with lim W2(µn, µ) = 0, and:
minφ∈KG
X(Pn) 1
2τ W2(µn, Hφ#µn) + E(Hφ#µn) + U(Gφ#µn) (∗)n If φn minimizes (∗)n, then (Gφn#µn) is a minimizing sequence for (∗).
(C2) U(ρ) = R
U(ρ(x)) d x, where U ≥ M is convex.
6= McCann’s condition for displacement-convexity
I Proof relies on Caffarelli’s regularity theorem.
[Benamou, Carlier, M., Oudet ’14]
Proof of the convergence theorem: Lower bound
Let m := minσ∈P(X) F(σ)
mn := minφn∈KX(Pn) Fn(Gφn#µn) I Step 1: lim inf mn ≥ m
JKO step: X bounded and convex, µ ∈ Pac(X) with density c−1 ≤ ρ ≤ c
F(σ) := W22(µ, σ) + E(σ) + U(σ) Fn(σ) := W22(µn, σ) + E(σ) + U(σ)
Proof of the convergence theorem: Lower bound
Let m := minσ∈P(X) F(σ)
mn := minφn∈KX(Pn) Fn(Gφn#µn) I Step 1: lim inf mn ≥ m
JKO step: X bounded and convex, µ ∈ Pac(X) with density c−1 ≤ ρ ≤ c
I Step 2: lim sup mn ≤ m = minσ∈P(X) F(σ)
F(σ) := W22(µ, σ) + E(σ) + U(σ) Fn(σ) := W22(µn, σ) + E(σ) + U(σ)
Proof of the convergence theorem: Lower bound
Let m := minσ∈P(X) F(σ)
mn := minφn∈KX(Pn) Fn(Gφn#µn) I Step 1: lim inf mn ≥ m
JKO step: X bounded and convex, µ ∈ Pac(X) with density c−1 ≤ ρ ≤ c
I Step 2: lim sup mn ≤ m = minσ∈P(X) F(σ)
s.t. limn→∞ kσ − σnkL∞(X) = 0, where σn is the density of Gφn#µn.
Using (C2) and a convolution argument
F(σ) := W22(µ, σ) + E(σ) + U(σ) Fn(σ) := W22(µn, σ) + E(σ) + U(σ)
⇐= Given any probability density σ ∈ C0(X) with ε < σ < ε−1, ∃φn ∈ KX(Pn)
Proof of the convergence theorem: Lower bound
X X
∇φ
µ σ = ∇φ#µ
I Step 2’:
s.t. limn→∞ kσ − σnkL∞(X) = 0, where σn := density of Gφn#µn. Given any prob. density σ ∈ C0(X) s.t. ε < σ < ε−1, ∃φn ∈ KX(Pn)
Proof of the convergence theorem: Lower bound
X X X
spt(µ ) = P ⊆ X
∇φ
µ σ = ∇φ#µ
I Step 2’:
∀p ∈ Pn, σ(Vpn) = µp, where Vpn := ∂φˆn(p)
∂φˆn
(a) Given µn = P
p∈Pn µpδp, take φn ∈ KX(Pn) O.T. potential between µn and σ:
p Vpn
s.t. limn→∞ kσ − σnkL∞(X) = 0, where σn := density of Gφn#µn. Given any prob. density σ ∈ C0(X) s.t. ε < σ < ε−1, ∃φn ∈ KX(Pn)
Proof of the convergence theorem: Lower bound
X X X
∇φ
µ σ = ∇φ#µ
I Step 2’:
∀p ∈ Pn, σ(Vpn) = µp, where Vpn := ∂φˆn(p) (b) σn := Gφn#µn = P
p∈Pn
σ(Vpn)
Hd(Vpn)1V n
p .
∂φˆn
(a) Given µn = P
p∈Pn µpδp, take φn ∈ KX(Pn) O.T. potential between µn and σ:
p Vpn
s.t. limn→∞ kσ − σnkL∞(X) = 0, where σn := density of Gφn#µn. Given any prob. density σ ∈ C0(X) s.t. ε < σ < ε−1, ∃φn ∈ KX(Pn)
Proof of the convergence theorem: Lower bound
X X X
spt(µ ) = P ⊆ X
∇φ
µ σ = ∇φ#µ
I Step 2’:
∀p ∈ Pn, σ(Vpn) = µp, where Vpn := ∂φˆn(p) (b) σn := Gφn#µn = P
p∈Pn
σ(Vpn)
Hd(Vpn)1V n
p . (c) If maxp∈Pn diam(Vpn) −−−→n→∞ 0 , then
∂φˆn
(a) Given µn = P
p∈Pn µpδp, take φn ∈ KX(Pn) O.T. potential between µn and σ:
p Vpn
s.t. limn→∞ kσ − σnkL∞(X) = 0, where σn := density of Gφn#µn. Given any prob. density σ ∈ C0(X) s.t. ε < σ < ε−1, ∃φn ∈ KX(Pn)
limn→∞ kσ − σnkL∞(X) = 0.
(1)
Proof of the convergence theorem: Lower bound
X X X
∇φ
µ σ = ∇φ#µ
I Step 2’:
∀p ∈ Pn, σ(Vpn) = µp, where Vpn := ∂φˆn(p) (b) σn := Gφn#µn = P
p∈Pn
σ(Vpn)
Hd(Vpn)1V n
p . (c) If maxp∈Pn diam(Vpn) −−−→n→∞ 0 , then
∂φˆn
(a) Given µn = P
p∈Pn µpδp, take φn ∈ KX(Pn) O.T. potential between µn and σ:
p Vpn
s.t. limn→∞ kσ − σnkL∞(X) = 0, where σn := density of Gφn#µn. Given any prob. density σ ∈ C0(X) s.t. ε < σ < ε−1, ∃φn ∈ KX(Pn)
(d) Assume not (1): ∃pn ∈ Pn s.t. diam(Vpn) ≥ ε.
diam(∂φ(x)) ≥ ε where x = limn pn ∈ X. limn→∞ kσ − σnkL∞(X) = 0.
(1)
Up to extraction, φˆn −−−→k.k∞ φ, so that (2)
Proof of the convergence theorem: Lower bound
X X X
spt(µ ) = P ⊆ X
∇φ
µ σ = ∇φ#µ
I Step 2’:
∀p ∈ Pn, σ(Vpn) = µp, where Vpn := ∂φˆn(p) (b) σn := Gφn#µn = P
p∈Pn
σ(Vpn)
Hd(Vpn)1V n
p . (c) If maxp∈Pn diam(Vpn) −−−→n→∞ 0 , then
∂φˆn
(a) Given µn = P
p∈Pn µpδp, take φn ∈ KX(Pn) O.T. potential between µn and σ:
p Vpn
s.t. limn→∞ kσ − σnkL∞(X) = 0, where σn := density of Gφn#µn. Given any prob. density σ ∈ C0(X) s.t. ε < σ < ε−1, ∃φn ∈ KX(Pn)
(d) Assume not (1): ∃pn ∈ Pn s.t. diam(Vpn) ≥ ε.
diam(∂φ(x)) ≥ ε where x = limn pn ∈ X. (e) Moreover, ∇φ#ρ = σ. By Caffarelli’s regularity
limn→∞ kσ − σnkL∞(X) = 0.
(1)
Up to extraction, φˆn −−−→k.k∞ φ, so that (2)
4. Numerical results
Computing the discrete Monge-Amp`ere operator
U(Gφ#µP ) = P
p∈P U (µp/MA[φ](p)) MA[φ](p)
with MA[φ](p) := Hd(∂φ(p)).ˆ
I Goal: fast computation of MA[φ](p) and its 1st/2nd derivatives w.r.t φ
Computing the discrete Monge-Amp`ere operator
X
Definition: Given a function φ : P → R,
I We rely on the notion of Laguerre (or power) cell in computational geometry
LagφP (p) := {y ∈ Rd; ∀q ∈ P, φ(q) ≥ φ(p) + hq − p|yi}
U(Gφ#µP ) = P
p∈P U (µp/MA[φ](p)) MA[φ](p)
with MA[φ](p) := Hd(∂φ(p)).ˆ
LagφP(p)
p
I Goal: fast computation of MA[φ](p) and its 1st/2nd derivatives w.r.t φ
Computing the discrete Monge-Amp`ere operator
X
Definition: Given a function φ : P → R,
I We rely on the notion of Laguerre (or power) cell in computational geometry
LagφP (p) := {y ∈ Rd; ∀q ∈ P, φ(q) ≥ φ(p) + hq − p|yi}
U(Gφ#µP ) = P
p∈P U (µp/MA[φ](p)) MA[φ](p)
with MA[φ](p) := Hd(∂φ(p)).ˆ
LagφP(p)
p
I Goal: fast computation of MA[φ](p) and its 1st/2nd derivatives w.r.t φ
Computing the discrete Monge-Amp`ere operator
LagφP(p) := {y; ∀q ∈ P, kq − yk2 ≥ kp − yk2} X
Definition: Given a function φ : P → R,
I We rely on the notion of Laguerre (or power) cell in computational geometry
LagφP (p) := {y ∈ Rd; ∀q ∈ P, φ(q) ≥ φ(p) + hq − p|yi}
−→ For φ(p) = kpk2/2, one gets the Voronoi cell:
U(Gφ#µP ) = P
p∈P U (µp/MA[φ](p)) MA[φ](p)
with MA[φ](p) := Hd(∂φ(p)).ˆ
LagφP(p)
p
I Goal: fast computation of MA[φ](p) and its 1st/2nd derivatives w.r.t φ
−→ Computation in time O(|P | log |P|) in 2D