a lagrangian scheme 'à la brenier' for the incompressible ... · a lagrangian scheme...
TRANSCRIPT
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
A Lagrangian scheme ”a la Brenier” for theincompressible Euler equations
T. O. Gallouet1 , Q. Merigot 2,E. Shwindt 3, B. Levy 3,
F.X. Vialard 1
Inria, Paris 1
Universite Paris Sud-Orsay 2
Inria, Nancy 3
Paris 6, LJLL, 16 march 2018
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Incompressible Euler equations
Domain: Ω ⊆ Rd with Leb measure. Eulerian formulation:∂tu(t, x) + (u(t, x) · ∇) u(t, x) = −∇p(t, x),
div(u(t, x)) = 0
u(t, x) · n = 0
u(0, x) = u0.
Lagrangian formulation:ddtφ(t, x) = u (t, φ(t, x)) for t ∈ [0,T ], x ∈ Ω ,
φ(0, ·) = id,
∂tφ(0, ·) = u0.
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Incompressible Euler equations II
Measure preserving maps:
S =s ∈ L2(Ω,Rd) | s# Leb = Leb
,
The incompressibility constraint reads
div u = 0 ⇐⇒ det∇φ(t) = 1 ⇐⇒ φ(t, ·) ∈ S
The evolution equation
d2
dt2φ(t) = −∇p(t, φ(t, x)) ∈ (TφS)⊥
Conclusion: solutions of the incompressible Euler equations aregeodesics of the measure preserving maps S for the L2 metric.
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Approached geodesics a la Brenier
Simple example: we consider R× 0 ⊂ R2.
geodesic: γ: [0,T ]→ R2 withγ(t) = (t, 0), t ∈ [0,T ],
γ(0) = (0, 0),
γ(0) = (1, 0).
Approached geodesic z with initial error δ.z(0) = (0, δ),
z(0) = (1, 0).
ε-evolution :
z(t) =1
ε2(PR(z)− z)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Approached geodesics a la Brenier II
Exact solution:
z(t) =(t, δ cos
t
ε
).
Solution for the Hamiltonian system defined with
H(z , v) =1
2||v ||2 +
1
2ε2d2R×0(z)
Quantifying the convergence: the modulated energy
Eγ(t) =1
2||z(t)− γ(t)||2 +
1
2ε2d2R×0(z(t)).
Eγ(t) =δ2
ε2
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
naive idea
Compute the solution of the Hamiltonian system associated to
H(f , v) =1
2‖v‖2
M +d2S (f )
2ε2.
How to compute d2S (f ) ?
→ Optimal Transport (Brenier)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Settings
• two probability measures µ, ν on Rd .
• cost c : X × Y → R, in particular c(x , y) = |x − y |2.
• t#µ = ν: for any measurable set B we haveν(B) = µ
(t−1(B)
).
Problem (Monge 1781)
Find T such that
T = argmint#µ=ν
(∫c (x , t(x)) dµ(x)
).
MK22(µ, ν) = min
t#µ=ν
(∫|t(x)− x |2 dµ(x)
)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Existence
Theorem (Existence, Brenier)
Let µ ac. w.r.t. Leb. Then there exists a unique ∇ϕ with ϕconvex such that T = ∇ϕ.
Theorem (Polar decomposition, Brenier)
Let f ∈ L2(Ω,Rd) and ν = f# Leb. Then there exists ϕ convexand σ ∈ S such that
• f = ∇ϕ σ,
• d2S (f ) = MK2
2(Leb, ν) = MK22(Leb, f# Leb),
• σ ∈ PS(f ).
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Kantorovich potentials
Owner point of view:
Factory points: Xα, α ∈ 1, ...,N, producing 1N .
Distribution points: Yβ, β ∈ 1, ...,N, distributing 1N .
Transport cost: c(Xα,Yβ).
Shipper point of view: At a factory Xα, I buy −ϕ(Xα). At a
distribution point Yβ, I sell ψ(Yβ). I must ship everything and I
guarantee for all α, β
(∗∗) ϕ(Xα) + ψ(Yβ) ≤ c(Xα,Yβ).
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Kantorovich potentials II
The (∗∗) deal means∫ϕ(x)dµ(x) +
∫ψ(y)dν(y) ≤ MK2
2(µ, ν).
Kantorovich duality
supϕ⊕ψ≤c
∫ϕ(x)dµ(x) +
∫ψ(y)dν(y) = MK2
2(µ, ν).
c-transform
ψc(x) = supy
(c(x , y)− φ(y)) .
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Semi-Discrete optimal transport (Q. Merigot)
We want to compute MK22
(Leb, 1
N
∑Nβ=1 δYβ
).
Laguerre Cells for points Yβ and prices ψβ.
Lagα ((Yβ)1...N , (ψβ)1...N) =x ∈ Ω
∣∣∣−ψα + |x − Yα|2 ≤ −ψβ + |x − Yβ|2 , for all β
→ Partition of the space Ω.
To be sure that the price gives a transport maps, it must holds
|Lagα (Yβ, ψβ)| = 1/N.
Numerically: Newton on this condition.
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Discretization space
Ω partitioned in N cells Pα with mass 1N .
MN = piecewise constant functions on Pα ⊂ L2(Ω,Rd).
For f ∈MN :
d2S (f ) = MK2
2(Leb, f# Leb) = MK22(Leb,
1
N
N∑α=1
δfα).
Projection set of f on S:
PS(f ) = σ ∈ S | ∀α, σ(Pα) = Lagα(fβ, ψopt).
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Approximate Geodesics
d2S as a function : MN → R:
H(f , v) =1
2‖v‖2
M +d2S (f )
2ε2
=1
2‖v‖2
M +1
2ε2MK 2
2 (Leb,1
N
N∑α=1
δfα).
Scheme: f (t) +∇d2
S (f (t))
2ε2 = 0, for t ∈ [0,T ] ,
(f (0), f (0)) ∈M2N
With1
2∇d2
S (f (t)) = f − PM PS(f ) ∈MN
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
Approximate Geodesics II
PM PS(f ) is the application
Pα → Bar(Lagα(fβ, ψ
opt))
Scheme rewritesf (t) = PMPS(f )−fε2 = 0, for t ∈ [0,T ] ,
(f (0), f (0)) ∈M2N
Full discretization (Euler semi-symplectic), step time τ .V n+1 = V n − τ f n−PMN
PS(f n)
ε2
f n+1 = f n + τV n+1,
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
ConvergenceLet u be a smooth (eulerian) solution of the incompressible Eulerequations.
The modulated energy:
Eu(t) =1
2‖f (t)− u(t, f (t))‖2
M +d2S (f )
2ε2.
Gronwall estimates, Semi-Discrete:
Eu(t) ≤ C
[ε2 + hN +
h2N
ε2
]Full discrete:
Eu(t) ≤ C
[ε2 + hN +
h2N
ε2+ κ+
τ
ε
]Where
κ = maxn∈N∩[0,T/τ ]
(Hn − H0
)≤ τ
ε2.
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
Space: (S,L2) ⊂ L2(Ω)
Projection: d2S (f (t)) = MK2
2(Leb, f# Leb), polar decomposition.
Geodesics: Incompressible Euler equations.
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
Space: (S,L2) ⊂ L2(Ω)
Projection: d2S (f (t)) = MK2
2(Leb, f# Leb), polar decomposition.
Geodesics: Incompressible Euler equations.
Behind the structure: the Otto Riemannian submersion
π : (Diff(Ω),L2)→ (Dens(Ω), MK2)
φ 7→ φ#(Leb).
DefinitionThe map π is a Riemannian submersion if π is a submersion andfor any x ∈ M, the map dfx : Ker(dπx)⊥ 7→ Tπ(x)N is an isometry.
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
Space: (S,L2) ⊂ L2(Ω)
Projection: d2S (f (t)) = MK2
2(Leb, f# Leb), polar decomposition.
Geodesics: Incompressible Euler equations.
Behind the structure: the Otto Riemannian submersion
π : (Diff(Ω),L2)→ (Dens(Ω), MK2)
φ 7→ φ#(Leb).
DefinitionThe map π is a Riemannian submersion if π is a submersion andfor any x ∈ M, the map dfx : Ker(dπx)⊥ 7→ Tπ(x)N is an isometry.
S = π−1(Leb), ∂tρ = div(ρv), v = φ(t) φ−1(t)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersionSpace: (S,L2) ⊂ L2(Ω)
Projection: d2S (f (t)) = MK2
2(Leb, f# Leb), polar decomposition.
Geodesics: Incompressible Euler equations.
Behind the structure: the Otto Riemannian submersion
π : (Diff(Ω),L2)→ (Dens(Ω), MK2)
φ 7→ φ#(Leb).
DefinitionThe map π is a Riemannian submersion if π is a submersion andfor any x ∈ M, the map dfx : Ker(dπx)⊥ 7→ Tπ(x)N is an isometry.
S = π−1(Leb), ∂tρ = div(ρv), v = φ(t) φ−1(t)
v = u+∇p, div(u) = 0, W22(∂tρ) = inf
∂tρ=div(ρv)‖u‖L2(ρ) = ‖∇p‖L2(ρ)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
What if
Projection: given by the WMKHFR distance on M+(Ω)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
What if
Projection: given by the HK distance on M+(Ω)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
What if
Projection: given by the d distance on M+(Ω)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
What if
Projection: given by the WFR distance on M+(Ω)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
What if
Projection: given by the HK-WFR distance on M+(Ω)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
What if
Projection: given by the HK-WFR distance on M+(Ω)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
Yes
Projection: given by the HK-WFR distance on M+(Ω)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
Projection: given by the HK-WFR distance on M+(Ω)
Submersion
π :(Diff(Ω) n C∞(Ω,R+), dc
)→ (M+(Ω), WFR)
(φ, λ) 7→ π(φ, λ) = φ#(λ2 Leb).
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
Projection: given by the HK-WFR distance on M+(Ω)
Submersion
π :(Diff(Ω) n C∞(Ω,R+), dc
)→ (M+(Ω), WFR)
(φ, λ) 7→ π(φ, λ) = φ#(λ2 Leb).
S = π−1(Leb) =(φ,√
Jac(φ))
, ∂tρ = div(ρv) + ρr ,
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
Projection: given by the HK-WFR distance on M+(Ω)
Submersion
π :(Diff(Ω) n C∞(Ω,R+), dc
)→ (M+(Ω), WFR)
(φ, λ) 7→ π(φ, λ) = φ#(λ2 Leb).
S = π−1(Leb) =(φ,√
Jac(φ))
, ∂tρ = div(ρv) + ρr ,
WFR22(∂tρ) = inf
∂tρ=div(ρv)+rρ‖v‖2
L2(ρ) + ‖r‖2L2(ρ) = gc(∇r , r)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersion
Space: S ⊂ Diff(Ω) n C∞(Ω,R+)
Projection: given by the HK-WFR distance on M+(Ω)
Submersion
π :(Diff(Ω) n C∞(Ω,R+), dc
)→ (M+(Ω), WFR)
(φ, λ) 7→ π(φ, λ) = φ#(λ2 Leb).
S = π−1(Leb) =(φ,√
Jac(φ))
, ∂tρ = div(ρv) + ρr ,
WFR22(∂tρ) = inf
∂tρ=div(ρv)+rρ‖v‖2
L2(ρ) + ‖r‖2L2(ρ) = gc(∇r , r)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersionSpace: S ⊂ Diff(Ω) n C∞(Ω,R+)
Projection: given by the HK-WFR distance on M+(Ω) New polardecomposition
Geodesics:
Submersion
π :(Diff(Ω) n C∞(Ω,R+), dc
)→ (M+(Ω), WFR)
(φ, λ) 7→ π(φ, λ) = φ#(λ2 Leb).
S = π−1(Leb) =(φ,√
Jac(φ))
, ∂tρ = div(ρv) + ρr ,
WFR22(∂tρ) = inf
∂tρ=div(ρv)+rρ‖v‖2
L2(ρ) + ‖r‖2L2(ρ) = gc(∇r , r)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersionSpace: S ⊂ Diff(Ω) n C∞(Ω,R+)
Projection: given by the HK-WFR distance on M+(Ω) New polardecomposition
Geodesics:
Submersion
π :(Diff(Ω) n C∞(Ω,R+), dc
)→ (M+(Ω), WFR)
(φ, λ) 7→ π(φ, λ) = φ#(λ2 Leb).
S = π−1(Leb) =(φ,√
Jac(φ))
, ∂tρ = div(ρv) + ρr ,
WFR22(∂tρ) = inf
∂tρ=div(ρv)+rρ‖v‖2
L2(ρ) + ‖r‖2L2(ρ) = gc(∇r , r)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
The Riemannian submersionSpace: S ⊂ Diff(Ω) n C∞(Ω,R+)
Projection: given by the HK-WFR distance on M+(Ω) New polardecomposition
Geodesics: Camassa-Holm equations.
Submersion
π :(Diff(Ω) n C∞(Ω,R+), dc
)→ (M+(Ω), WFR)
(φ, λ) 7→ π(φ, λ) = φ#(λ2 Leb).
S = π−1(Leb) =(φ,√
Jac(φ))
, ∂tρ = div(ρv) + ρr ,
WFR22(∂tρ) = inf
∂tρ=div(ρv)+rρ‖v‖2
L2(ρ) + ‖r‖2L2(ρ) = gc(∇r , r)
Arnold interpretation Approached geodesics scheme Optimal Transport Semi-Discrete OT Numerical Scheme Bonus
ConclusionRevisit Brenier’s work for the incompressible Euler equations in thiscase
• Generalized solutions
• Regularity of the pression
• Numerical scheme
New polar decompostion:
(φ, λ) = expC(M)
(−1
2∇pz0 ,−pz0
) (s,
√Jac(s))
or equivalently
(φ, λ) =
(ϕ, e−z0
√1 + ‖∇z0‖2
)· (s,
√Jac(s)) ,
where pz0 = ez0 − 1 and
ϕ(x) = expMx
(− arctan
(1
2‖∇z0(x)‖
)∇z0(x)
‖∇z0(x)‖
).