Spaces of measures with nonnegative cross-curvature
Flavien Léger
joint work with Gabriele Todeschi and François-Xavier Vialard

∀(ξ,η),
Introduction: the MTW tensor
Sc(x,y)(ξ,η)=(cikmˉcmˉncnˉℓˉ−ciˉkℓˉ)ξiηˉξkηℓˉ
X,Y, d-dimensional manifolds, c∈C4(X×Y), ∇xy2c(x,y) nonsingular
MTW conditions
MTW tensor
(Ma–Trudinger–Wang ’05)
Sc(x,y)(ξ,η)≥κ∣ξ∣x2∣η∣x,y2
κ>0
κ=0
(Ma–Trudinger–Wang ’05)
(Trudinger–Wang ’09)
Sc(x,y)(ξ,η)≥0
(Kim–McCann ’10)
Sc(x,y)(ξ,η)≥κ∣ξ∣x2∣η∣x,y2−C∣⟨ξ,η⟩∣∣ξ∣∣η∣
(Loeper–Villani ’10)
⏵ ∀(ξ,η),ξ⊥η
Nonnegative cross-curvature (NNCC)
⏵ ∀(ξ,η),
⏵
Motivation 1: principal–agents problems
Ψc= set of c-concave functions
ϕc(x):=y∈Yminc(x,y)+ϕ(y)
Ψc is a convex set (of functions)
T H E O R E M (FIGALLI–KIM–MCCANN '11)
If c has NNCC + other assumptions then
Tϕ(x):=y∈Yargmin_________
Motivations 2: Evolution variational inequalities
(X,d) metric space
Convergence rate
EVI
(X,d) non-positively curved, Mayer/Jost
Ambrosio–Gigli–Savaré
E convex on geodesics
E convex on curves x(t) such that d2(x,xn−1) is 1-convex, i.e. t↦d2(x(t),xn−1)−t2d2(x(1),x(0)) is convex
t↦(x(t),yˉ) is a c-segment if ∇yc(x(t),yˉ)=(1−t)∇yc(x(0),yˉ)+t∇yc(x(1),yˉ).
D E F I N I T I O N
Background
X,Y two d-dimensional manifolds,
c∈C4(X×Y), nondegenerate, twist,
X×Y c-convex
T H E O R E M (KIM–MCCANN '10)
c has nonnegative cross-curvature iff for any c-segment t↦(x(t),yˉ), ∀y∈Y,t↦c(x(t),yˉ)−c(x(t),y) is convex.
Key observation
c∈C1(X×Y. If x(t) is any curve such that t↦c(x(t),yˉ)−c(x(t),y) is convex for every y∈Y, then (x(t),yˉ) is a c-segment.
D E F I N I T I O N (L–Todeschi–Vialard '24)
(X×Y,c) is an NNCC space if for each (x0,x1,yˉ)∈X×X×Y, there exists a path x(⋅) from x0 to x1 such that ∀y∈Y, c(x(t),yˉ)−c(x(t),y)≤(1−t)[c(x0,yˉ)−c(x0,y)]+t[c(x1,yˉ)−c(x1,y)].
(x(t),yˉ)=: generalized c-segment.
X,Y two arbitrary sets, c:X×Y→R∪{+∞}.
(L–Todeschi–Vialard '24)
Some properties
Products preserve NNCC
If (X1×Y1,c1) and (X2×Y2,c2) are NNCC then so is ((X1×X2)×(Y1×Y2),c) with c((x1,x2),(y1,y2))=c1(x1,y1)+c2(x2,y2).
Certain projections preserve NNCC
(X×Y,c) NNCC space, π1:X→X, π2:Y→Y.c(x,y)=inf{c(x,y):π1(x)=x,π2(y)=y} Additional assumptions. Then (X×Y,c) is NNCC
(P2(Rn)×P2(Rn),W22) is an NNCC space.
More generally X, Y Polish spaces, c∈C(X×Y).
If (X×Y,c) is an NNCC space then so is (P(X)×P(Y),Tc).
T H E O R E M (L–Todeschi–Vialard '24)
The Wasserstein space is NNCC
A generalized c-segment for Tc between (μ0,ν) and (μ1,ν) is given by (μ(t),ν)
⏵ (T0,S) optimal coupling of (μ0,ν)
⏵ (T1,S) optimal coupling of (μ1,ν)
⏵ for each ω∈Ω, t↦(Tt(ω),S(ω)) is a c-segment
⏵ μ(t)=(Tt)#P
(Ω,F,P)
T0,T1:Ω→X
S:Ω→Y
Tc(μ,ν)=π∈Π(μ,ν)inf∫c(x,y)dπ
More examples of NNCC spaces
Bures–Wasserstein
Gromov–Wasserstein X=[X,f,μ] and Y=[Y,g,ν]
GW2(X,Y)=π∈Π(μ,ν)inf∫∣f(x,x′)−g(y,y′)∣2dπ(x,y)dπ(x′,y′).
Unbalanced OT
Application: EVIs
EVIs
c:X×X→R+, E:X→R
Convergence rates:
EVI
c=W22,Tc,GW2,BW2…
EVIs
⏵ (X×X,c) NNCC space
⏵ E convex on generalized c-segments (x(t),xn−1)
⏵ c satisfies t→0liminftc(x(t),x(0))=0.
Then EVI.
T H E O R E M (L–Todeschi–Vialard '24)
EVI
Thank you!
(Moka-10 2024-06-04) Spaces of measures with nonnegative cross-curvature
By Flavien Léger
(Moka-10 2024-06-04) Spaces of measures with nonnegative cross-curvature
- 251