the geometry of the space of measures and its applications
TRANSCRIPT
![Page 1: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/1.jpg)
The geometry of the space of measures andits applications
John LottUC-Berkeley
http://math.berkeley.edu/˜lott
May 23, 2009
![Page 2: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/2.jpg)
Singer : 33 Children
![Page 3: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/3.jpg)
Singer : 94 Grandchildren
![Page 4: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/4.jpg)
Singer : 18 Great-grandchildren
![Page 5: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/5.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 6: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/6.jpg)
Infinite-dimensional spaces
Geometry and topology of infinite-dimensional spaces
Example : the space of connections modulo gaugetransformations (Atiyah-Singer, . . . )
Today : the space of probability measures.
![Page 7: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/7.jpg)
Infinite-dimensional spaces
Geometry and topology of infinite-dimensional spaces
Example : the space of connections modulo gaugetransformations (Atiyah-Singer, . . . )
Today : the space of probability measures.
![Page 8: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/8.jpg)
Questions
The motivation comes from questions about finite-dimensionalspaces.
How can we understand Ricci curvature?
Does it make sense to talk about Ricci curvature for nonsmoothspaces?
Perelman introduced an important monotonic quantity in Ricciflow, the reduced volume. Where does this come from?
Claim : These questions can be answered in terms of optimaltransport, or the geometry of the space of probability measures.
![Page 9: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/9.jpg)
Questions
The motivation comes from questions about finite-dimensionalspaces.
How can we understand Ricci curvature?
Does it make sense to talk about Ricci curvature for nonsmoothspaces?
Perelman introduced an important monotonic quantity in Ricciflow, the reduced volume. Where does this come from?
Claim : These questions can be answered in terms of optimaltransport, or the geometry of the space of probability measures.
![Page 10: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/10.jpg)
Questions
The motivation comes from questions about finite-dimensionalspaces.
How can we understand Ricci curvature?
Does it make sense to talk about Ricci curvature for nonsmoothspaces?
Perelman introduced an important monotonic quantity in Ricciflow, the reduced volume. Where does this come from?
Claim : These questions can be answered in terms of optimaltransport, or the geometry of the space of probability measures.
![Page 11: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/11.jpg)
Partly joint work with Cedric Villani (ENS-Lyon).
Related work was done by Karl-Theodor Sturm (University ofBonn).
LV = Lott-Villani, S = Sturm
![Page 12: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/12.jpg)
Partly joint work with Cedric Villani (ENS-Lyon).
Related work was done by Karl-Theodor Sturm (University ofBonn).
LV = Lott-Villani, S = Sturm
![Page 13: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/13.jpg)
Partly joint work with Cedric Villani (ENS-Lyon).
Related work was done by Karl-Theodor Sturm (University ofBonn).
LV = Lott-Villani, S = Sturm
![Page 14: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/14.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 15: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/15.jpg)
Dirtmoving
Given a before and an after dirtpile, what is the most efficientway to move the dirt from one place to the other?
Let’s say that the cost to move a gram of dirt from x to y isd(x , y)2.
![Page 16: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/16.jpg)
Gaspard Monge
Memoire sur la theorie des deblais et des remblais (1781)
Memoir on the theory of excavations and fillings (1781)
![Page 17: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/17.jpg)
Gaspard Monge
![Page 18: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/18.jpg)
Wasserstein space
Let (X ,d) be a compact metric space.
NotationP(X ) is the set of Borel probability measures on X.
That is, µ ∈ P(X ) iff µ is a nonnegative Borel measure on Xwith µ(X ) = 1.
DefinitionGiven µ0, µ1 ∈ P(X ), the Wasserstein distance W2(µ0, µ1) isthe square root of the minimal cost to transport µ0 to µ1.
![Page 19: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/19.jpg)
Wasserstein space
W2(µ0, µ1)2 = inf{∫
X×Xd(x , y)2 dπ(x , y)
},
where
π ∈ P(X × X ), (p0)∗π = µ0, (p1)∗π = µ1.
![Page 20: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/20.jpg)
Wasserstein space
Fact :(P(X ),W2) is a metric space, called the Wasserstein space.
The metric topology is the weak-∗ topology, i.e. limi→∞ µi = µ ifand only if for all f ∈ C(X ), limi→∞
∫X f dµi =
∫X f dµ.
So to one compact metric space (X ,d), we’ve assignedanother one (P(X ),W2).
Note : There is an isometric embedding X → P(X ) by x → δx .
![Page 21: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/21.jpg)
Wasserstein space
Fact :(P(X ),W2) is a metric space, called the Wasserstein space.
The metric topology is the weak-∗ topology, i.e. limi→∞ µi = µ ifand only if for all f ∈ C(X ), limi→∞
∫X f dµi =
∫X f dµ.
So to one compact metric space (X ,d), we’ve assignedanother one (P(X ),W2).
Note : There is an isometric embedding X → P(X ) by x → δx .
![Page 22: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/22.jpg)
Wasserstein space
Fact :(P(X ),W2) is a metric space, called the Wasserstein space.
The metric topology is the weak-∗ topology, i.e. limi→∞ µi = µ ifand only if for all f ∈ C(X ), limi→∞
∫X f dµi =
∫X f dµ.
So to one compact metric space (X ,d), we’ve assignedanother one (P(X ),W2).
Note : There is an isometric embedding X → P(X ) by x → δx .
![Page 23: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/23.jpg)
Displacement interpolations
To move dirt in the real world, we would transport it alongminimizing geodesics.
Take a snapshot at time t . We get a family of meaures{µt}t∈[0,1], called a displacement interpolation. We would like tosay that this is a “geodesic” in P(X ).
![Page 24: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/24.jpg)
Displacement interpolations
To move dirt in the real world, we would transport it alongminimizing geodesics.
Take a snapshot at time t . We get a family of meaures{µt}t∈[0,1], called a displacement interpolation. We would like tosay that this is a “geodesic” in P(X ).
![Page 25: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/25.jpg)
Displacement interpolations
To move dirt in the real world, we would transport it alongminimizing geodesics.
Take a snapshot at time t . We get a family of meaures{µt}t∈[0,1], called a displacement interpolation. We would like tosay that this is a “geodesic” in P(X ).
![Page 26: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/26.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 27: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/27.jpg)
Smooth measures
If (M,g) is a compact connected Riemannian manifold, letP∞(M) ⊂ P(M) be the smooth probability measures withpositive density.
P∞(M) = {ρ dvolM : ρ ∈ C∞(M), ρ > 0,∫
M ρ dvolM = 1}.
![Page 28: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/28.jpg)
Otto’s formal Riemannian metric on P∞(M)
Given µ = ρ dvolM ∈ P∞(M), consider an infinitesimally nearbymeasure µ+ δµ, i.e.
δµ = (δρ) dvolM ∈ TµP∞(M).
Solve δρ = −∑
i ∇i(ρ∇iφ) for φ ∈ C∞(M), unique up to anadditive constant.
Definition :〈δµ, δµ〉 =
∫M|∇φ|2 ρ dvolM .
This is the H−1 Sobolev metric, in terms of ρ.
![Page 29: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/29.jpg)
Otto’s formal Riemannian metric on P∞(M)
Given µ = ρ dvolM ∈ P∞(M), consider an infinitesimally nearbymeasure µ+ δµ, i.e.
δµ = (δρ) dvolM ∈ TµP∞(M).
Solve δρ = −∑
i ∇i(ρ∇iφ) for φ ∈ C∞(M), unique up to anadditive constant.
Definition :〈δµ, δµ〉 =
∫M|∇φ|2 ρ dvolM .
This is the H−1 Sobolev metric, in terms of ρ.
![Page 30: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/30.jpg)
Corresponding energy of a curve
Say c : [0,1]→ P∞(M) is a smooth curve.
Write c(t) = ρ(t) dvolM .
Fact : We can solve
∂ρ
∂t= −
∑i
∇i(ρ∇iφ)
for φ ≡ φ(t) ∈ C∞(M).
![Page 31: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/31.jpg)
Corresponding energy of a curve
Say c : [0,1]→ P∞(M) is a smooth curve.
Write c(t) = ρ(t) dvolM .
Fact : We can solve
∂ρ
∂t= −
∑i
∇i(ρ∇iφ)
for φ ≡ φ(t) ∈ C∞(M).
![Page 32: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/32.jpg)
Benamou-Brenier variational problem
From {ρ(t)}t∈[0,1], we got {φ(t)}t∈[0,1].
Definition
E(c) =12
∫ 1
0
∫M|∇φ|2 ρ dvolM dt .
This is the energy of the curve c.
![Page 33: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/33.jpg)
Benamou-Brenier variational problem
From {ρ(t)}t∈[0,1], we got {φ(t)}t∈[0,1].
Definition
E(c) =12
∫ 1
0
∫M|∇φ|2 ρ dvolM dt .
This is the energy of the curve c.
![Page 34: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/34.jpg)
Geodesic distance equals Wasserstein distance
Theorem : (Otto-Westdickenberg 2005)
12
W2(µ0, µ1)2 = inf{E(c) : c(0) = µ0, c(1) = µ1}.
That is, the geodesic distance coming from Otto’s metric is theWasserstein distance W2, at least on P∞(M).
Note : the infimum may not be achieved. A minimizing c is asmooth displacement interpolation.
![Page 35: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/35.jpg)
Geodesic distance equals Wasserstein distance
Theorem : (Otto-Westdickenberg 2005)
12
W2(µ0, µ1)2 = inf{E(c) : c(0) = µ0, c(1) = µ1}.
That is, the geodesic distance coming from Otto’s metric is theWasserstein distance W2, at least on P∞(M).
Note : the infimum may not be achieved. A minimizing c is asmooth displacement interpolation.
![Page 36: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/36.jpg)
Geodesic distance equals Wasserstein distance
Theorem : (Otto-Westdickenberg 2005)
12
W2(µ0, µ1)2 = inf{E(c) : c(0) = µ0, c(1) = µ1}.
That is, the geodesic distance coming from Otto’s metric is theWasserstein distance W2, at least on P∞(M).
Note : the infimum may not be achieved. A minimizing c is asmooth displacement interpolation.
![Page 37: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/37.jpg)
Euler-Lagrange equations
The Euler-Lagrange equation for the functional E is
Hamilton-Jacobi equation
∂φ
∂t= − 1
2|∇φ|2.
We also had
Conservation equation
∂ρ
∂t= −
∑i
∇i(ρ∇iφ).
These are the equations for optimal transport and can besolved explicitly. (First worked out for Riemannian manifolds byRobert McCann 2001.)
![Page 38: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/38.jpg)
Euler-Lagrange equations
The Euler-Lagrange equation for the functional E is
Hamilton-Jacobi equation
∂φ
∂t= − 1
2|∇φ|2.
We also had
Conservation equation
∂ρ
∂t= −
∑i
∇i(ρ∇iφ).
These are the equations for optimal transport and can besolved explicitly. (First worked out for Riemannian manifolds byRobert McCann 2001.)
![Page 39: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/39.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 40: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/40.jpg)
Passing to metric geometry
Can we extend these statements from formal results aboutP∞(M) to rigorous results about P(M)?
Or, more generally, about P(X ) for a nonsmooth space X?
Use ideas from metric geometry.
![Page 41: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/41.jpg)
Passing to metric geometry
Can we extend these statements from formal results aboutP∞(M) to rigorous results about P(M)?
Or, more generally, about P(X ) for a nonsmooth space X?
Use ideas from metric geometry.
![Page 42: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/42.jpg)
Review of length spaces
Say (X ,d) is a compact metric space and γ : [0,1]→ X is acontinuous map.
The length of γ is
L(γ) = supJ
sup0=t0≤t1≤...≤tJ =1
J∑j=1
d(γ(tj−1), γ(tj)).
![Page 43: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/43.jpg)
Review of length spaces
Say (X ,d) is a compact metric space and γ : [0,1]→ X is acontinuous map.
The length of γ is
L(γ) = supJ
sup0=t0≤t1≤...≤tJ =1
J∑j=1
d(γ(tj−1), γ(tj)).
![Page 44: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/44.jpg)
Length spaces
Definition(X ,d) is a length space if the distance between two pointsx0, x1 ∈ X equals the infimum of the lengths of curves joiningthem, i.e.
d(x0, x1) = inf{L(γ) : γ(0) = x0, γ(1) = x1}.
A length-minimizing curve is called a geodesic.
![Page 45: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/45.jpg)
Length spaces
Definition(X ,d) is a length space if the distance between two pointsx0, x1 ∈ X equals the infimum of the lengths of curves joiningthem, i.e.
d(x0, x1) = inf{L(γ) : γ(0) = x0, γ(1) = x1}.
A length-minimizing curve is called a geodesic.
![Page 46: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/46.jpg)
Length spaces
Examples of length spaces :1. The underlying metric space of any Riemannian manifold.2.
Nonexamples :1. A finite metric space with more than one point.2. A circle with the chordal metric.
![Page 47: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/47.jpg)
Length spaces
Examples of length spaces :1. The underlying metric space of any Riemannian manifold.2.
Nonexamples :1. A finite metric space with more than one point.2. A circle with the chordal metric.
![Page 48: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/48.jpg)
Wasserstein space as a length space
Proposition : (LV,S)
If X is a length space then so is the Wasserstein space P(X ).
Hence we can talk about its (minimizing) geodesics {µt}t∈[0,1],called Wasserstein geodesics.
Proposition : (LV)
The Wasserstein geodesics are exactly the displacementinterpolations {µt}t∈[0,1].
![Page 49: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/49.jpg)
Wasserstein space as a length space
Proposition : (LV,S)
If X is a length space then so is the Wasserstein space P(X ).
Hence we can talk about its (minimizing) geodesics {µt}t∈[0,1],called Wasserstein geodesics.
Proposition : (LV)
The Wasserstein geodesics are exactly the displacementinterpolations {µt}t∈[0,1].
![Page 50: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/50.jpg)
Curvature of Wasserstein space
Formal calculation : (Otto 2001)
P(Rn) has nonnegative sectional curvature.
Theorem : (LV,S)
If a Riemannian manifold M has nonnegative sectionalcurvature then the length space P(M) has nonnegativecurvature in the Alexandrov sense.
Open question :
To what extent is P(M) an infinite-dimensional Riemannianmanifold?
![Page 51: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/51.jpg)
Curvature of Wasserstein space
Formal calculation : (Otto 2001)
P(Rn) has nonnegative sectional curvature.
Theorem : (LV,S)
If a Riemannian manifold M has nonnegative sectionalcurvature then the length space P(M) has nonnegativecurvature in the Alexandrov sense.
Open question :
To what extent is P(M) an infinite-dimensional Riemannianmanifold?
![Page 52: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/52.jpg)
Curvature of Wasserstein space
Formal calculation : (Otto 2001)
P(Rn) has nonnegative sectional curvature.
Theorem : (LV,S)
If a Riemannian manifold M has nonnegative sectionalcurvature then the length space P(M) has nonnegativecurvature in the Alexandrov sense.
Open question :
To what extent is P(M) an infinite-dimensional Riemannianmanifold?
![Page 53: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/53.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 54: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/54.jpg)
Review of Ricci curvature
Back to smooth manifolds.
Ricci curvature is an averaging of sectional curvature.
Fix a unit-length vector v ∈ TmM.
Definition
RicM(v,v) = (n − 1) · (the average sectional curvature
of the 2-planes P containing v).
Example : S2 × S2 has nonnegative sectional curvatures buthas positive Ricci curvatures.
![Page 55: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/55.jpg)
Review of Ricci curvature
Back to smooth manifolds.
Ricci curvature is an averaging of sectional curvature.
Fix a unit-length vector v ∈ TmM.
Definition
RicM(v,v) = (n − 1) · (the average sectional curvature
of the 2-planes P containing v).
Example : S2 × S2 has nonnegative sectional curvatures buthas positive Ricci curvatures.
![Page 56: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/56.jpg)
Ricci curvature
Regularity Issue :
To define RicM , we need a Riemannian metric which isC2-regular.
Can we make sense of Ricci curvature for nonsmooth spaces?Can we make sense at least of “nonnegative Ricci curvature”?
The analogous question for sectional curvature was solved byAlexandrov in the 1950’s.
![Page 57: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/57.jpg)
Ricci curvature
Regularity Issue :
To define RicM , we need a Riemannian metric which isC2-regular.
Can we make sense of Ricci curvature for nonsmooth spaces?Can we make sense at least of “nonnegative Ricci curvature”?
The analogous question for sectional curvature was solved byAlexandrov in the 1950’s.
![Page 58: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/58.jpg)
The relation of Ricci curvature to optimal transport
Say M is a compact Riemannian manifold.
Definition: The (negative) entropy functionalE : P(M)→ R ∪∞ is given by
E(µ) =
{∫M ρ log ρ dvolM if µ = ρ dvolM ,∞ if µ is not a.c. w.r.t. dvolM .
![Page 59: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/59.jpg)
Otto-Villani calculation
How does the entropy function behave along geodesics inP(M)?
Suppose that c(t) = ρ(t) dvolM is a smooth Wassersteingeodesic.
We defined φ(t) ∈ C∞(M) by
∂ρ
∂t= −
∑i
∇i(ρ∇iφ).
A local calculation : (Otto-Villani 2000)Along the geodesic c,
d2
dt2E(c(t)) =
∫M
[|Hess(φ)|2 + RicM(∇φ,∇φ)
]ρ dvolM .
![Page 60: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/60.jpg)
Otto-Villani calculation
How does the entropy function behave along geodesics inP(M)?
Suppose that c(t) = ρ(t) dvolM is a smooth Wassersteingeodesic.
We defined φ(t) ∈ C∞(M) by
∂ρ
∂t= −
∑i
∇i(ρ∇iφ).
A local calculation : (Otto-Villani 2000)Along the geodesic c,
d2
dt2E(c(t)) =
∫M
[|Hess(φ)|2 + RicM(∇φ,∇φ)
]ρ dvolM .
![Page 61: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/61.jpg)
Otto-Villani calculation
How does the entropy function behave along geodesics inP(M)?
Suppose that c(t) = ρ(t) dvolM is a smooth Wassersteingeodesic.
We defined φ(t) ∈ C∞(M) by
∂ρ
∂t= −
∑i
∇i(ρ∇iφ).
A local calculation : (Otto-Villani 2000)Along the geodesic c,
d2
dt2E(c(t)) =
∫M
[|Hess(φ)|2 + RicM(∇φ,∇φ)
]ρ dvolM .
![Page 62: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/62.jpg)
Convexity of entropy along Wasserstein geodesics
Corollary :If RicM ≥ 0 then d2E
dt2 ≥ 0, i.e. E is convex along anyWasserstein geodesic c.
Rigorous proof on P(M) :McCann-Erausquin-Cordero-Schmuckenschlager (2001)
Converse : von Renesse-Sturm (2005)
A new way of thinking about Ricci curvature :
Nonnegative Ricci curvature is equivalent to convexity of E (onP(M)).
We will use this property to define the notion of “nonnegativeRicci curvature” for a nonsmooth space.
![Page 63: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/63.jpg)
Convexity of entropy along Wasserstein geodesics
Corollary :If RicM ≥ 0 then d2E
dt2 ≥ 0, i.e. E is convex along anyWasserstein geodesic c.
Rigorous proof on P(M) :McCann-Erausquin-Cordero-Schmuckenschlager (2001)
Converse : von Renesse-Sturm (2005)
A new way of thinking about Ricci curvature :
Nonnegative Ricci curvature is equivalent to convexity of E (onP(M)).
We will use this property to define the notion of “nonnegativeRicci curvature” for a nonsmooth space.
![Page 64: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/64.jpg)
Convexity of entropy along Wasserstein geodesics
Corollary :If RicM ≥ 0 then d2E
dt2 ≥ 0, i.e. E is convex along anyWasserstein geodesic c.
Rigorous proof on P(M) :McCann-Erausquin-Cordero-Schmuckenschlager (2001)
Converse : von Renesse-Sturm (2005)
A new way of thinking about Ricci curvature :
Nonnegative Ricci curvature is equivalent to convexity of E (onP(M)).
We will use this property to define the notion of “nonnegativeRicci curvature” for a nonsmooth space.
![Page 65: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/65.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 66: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/66.jpg)
Gromov-Hausdorff topology
A topology on the set of all compact metric spaces (moduloisometry).
(X1,d1) and (X2,d2) are close in the Gromov-Hausdorfftopology if somebody with bad vision has trouble telling themapart.
![Page 67: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/67.jpg)
Example : a cylinder with a small cross-section isGromov-Hausdorff close to a line segment.
![Page 68: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/68.jpg)
Gromov’s precompactness theorem
Theorem : (Gromov 1981)
Given N ∈ Z+ and D > 0,
{(M,g) : dim(M) = N,diam(M) ≤ D,RicM ≥ 0}
is precompact in the Gromov-Hausdorff topology on{compact metric spaces}/isometry.
![Page 69: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/69.jpg)
Artist’s rendition of Gromov-Hausdorff space
Each point represents a compact metric space.Each interior point is a Riemannian manifold (M,g) withdim(M) = N, diam(M) ≤ D and RicM ≥ 0.
The boundary points are compact metric spaces (X ,d) withdimH X ≤ N and diam(X ) ≤ D. They are generally notmanifolds.(Example : X = M/G.)
In some moral sense, the boundary points are metric spaceswith “nonnegative Ricci curvature”.
![Page 70: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/70.jpg)
Artist’s rendition of Gromov-Hausdorff space
Each point represents a compact metric space.Each interior point is a Riemannian manifold (M,g) withdim(M) = N, diam(M) ≤ D and RicM ≥ 0.
The boundary points are compact metric spaces (X ,d) withdimH X ≤ N and diam(X ) ≤ D. They are generally notmanifolds.(Example : X = M/G.)
In some moral sense, the boundary points are metric spaceswith “nonnegative Ricci curvature”.
![Page 71: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/71.jpg)
Question :
What can we say about the Gromov-Hausdorff limits ofRiemannian manifolds with nonnegative Ricci curvature?
To answer this, it turns out to be useful to consider insteadmetric-measure spaces.
![Page 72: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/72.jpg)
Question :
What can we say about the Gromov-Hausdorff limits ofRiemannian manifolds with nonnegative Ricci curvature?
To answer this, it turns out to be useful to consider insteadmetric-measure spaces.
![Page 73: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/73.jpg)
Metric-measure spaces
DefinitionA metric-measure space is a compact metric space (X ,d)equipped with a given probability measure ν ∈ P(X ).
Canonical Example
If M is a compact Riemannian manifold then(
M,dM ,dvolMvol(M)
)is
a metric-measure space.
More generally, a smooth measured length space is a compactRiemannian manifold (M,g) equipped with a smooth probabilitymeasure dν = e−Ψ dvolM .
![Page 74: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/74.jpg)
Metric-measure spaces
DefinitionA metric-measure space is a compact metric space (X ,d)equipped with a given probability measure ν ∈ P(X ).
Canonical Example
If M is a compact Riemannian manifold then(
M,dM ,dvolMvol(M)
)is
a metric-measure space.
More generally, a smooth measured length space is a compactRiemannian manifold (M,g) equipped with a smooth probabilitymeasure dν = e−Ψ dvolM .
![Page 75: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/75.jpg)
Metric-measure spaces
DefinitionA metric-measure space is a compact metric space (X ,d)equipped with a given probability measure ν ∈ P(X ).
Canonical Example
If M is a compact Riemannian manifold then(
M,dM ,dvolMvol(M)
)is
a metric-measure space.
More generally, a smooth measured length space is a compactRiemannian manifold (M,g) equipped with a smooth probabilitymeasure dν = e−Ψ dvolM .
![Page 76: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/76.jpg)
Measured Gromov-Hausdorff limits
An easy consequence of Gromov precompactness :{(M,g,
dvolMvol(M)
): dim(M) = N,diam(M) ≤ D,RicM ≥ 0
}is precompact in the measured Gromov-Hausdorff topology on{compact metric-measure spaces}/isometry.
What can we say about the limit points? (Work ofCheeger-Colding 1996-2000)
What are the smooth limit points?
![Page 77: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/77.jpg)
Measured Gromov-Hausdorff limits
An easy consequence of Gromov precompactness :{(M,g,
dvolMvol(M)
): dim(M) = N,diam(M) ≤ D,RicM ≥ 0
}is precompact in the measured Gromov-Hausdorff topology on{compact metric-measure spaces}/isometry.
What can we say about the limit points? (Work ofCheeger-Colding 1996-2000)
What are the smooth limit points?
![Page 78: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/78.jpg)
Measured Gromov-Hausdorff (MGH) topology
Definitionlimi→∞(Xi ,di , νi) = (X ,d , ν) if there are Borel mapsfi : Xi → X and a sequence εi → 0 such that1. (Almost isometry) For all xi , x ′i ∈ Xi ,
|dX (fi(xi), fi(x ′i ))− dXi (xi , x ′i )| ≤ εi .
2. (Almost surjective) For all x ∈ X and all i , there is somexi ∈ Xi such that
dX (fi(xi), x) ≤ εi .
3. limi→∞(fi)∗νi = ν in the weak-∗ topology.
![Page 79: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/79.jpg)
Passage to P(X )
To one compact metric space we have assigned another.
(X ,d) −→ (P(X ),W2)
Proposition : (LV)
If Xi → X in the Gromov-Hausdorff topology thenP(Xi)→ P(X ) in the Gromov-Hausdorff topology.
We will use the properties of the Wasserstein space (P(X ),W2)to say something about the geometry of (X ,d).
In particular, we will define what it means for (X , ν) to have“nonnegative Ricci curvature” in terms of P(X ).
![Page 80: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/80.jpg)
Passage to P(X )
To one compact metric space we have assigned another.
(X ,d) −→ (P(X ),W2)
Proposition : (LV)
If Xi → X in the Gromov-Hausdorff topology thenP(Xi)→ P(X ) in the Gromov-Hausdorff topology.
We will use the properties of the Wasserstein space (P(X ),W2)to say something about the geometry of (X ,d).
In particular, we will define what it means for (X , ν) to have“nonnegative Ricci curvature” in terms of P(X ).
![Page 81: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/81.jpg)
Passage to P(X )
To one compact metric space we have assigned another.
(X ,d) −→ (P(X ),W2)
Proposition : (LV)
If Xi → X in the Gromov-Hausdorff topology thenP(Xi)→ P(X ) in the Gromov-Hausdorff topology.
We will use the properties of the Wasserstein space (P(X ),W2)to say something about the geometry of (X ,d).
In particular, we will define what it means for (X , ν) to have“nonnegative Ricci curvature” in terms of P(X ).
![Page 82: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/82.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 83: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/83.jpg)
Notation
X a compact Hausdorff space.
P(X ) = Borel probability measures on X , with weak-∗ topology.
Fix a background measure ν ∈ P(X ).
![Page 84: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/84.jpg)
Effective dimension
N ∈ [1,∞] a new parameter (possibly infinite).
It turns out that there’s not a single notion of “nonnegative Riccicurvature”, but rather a 1-parameter family. That is, for each N,there’s a notion of a space having “nonnegative N-Riccicurvature”.
Here N is an effective dimension of the space, and must beinputted.
![Page 85: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/85.jpg)
Entropy
Definition of the “negative entropy” function
EN : P(X )→ R ∪∞
Letµ = ρ ν + µs
be the Lebesgue decomposition of µ with respect to ν.
For N ∈ [1,∞), the “negative entropy” of µ with respect to ν is
EN(µ) = N − N∫
Xρ1− 1
N dν.
For N =∞,
E∞(µ) =
{∫X ρ log ρ dν if µ is a.c. w.r.t. ν,∞ otherwise.
![Page 86: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/86.jpg)
Entropy
Definition of the “negative entropy” function
EN : P(X )→ R ∪∞
Letµ = ρ ν + µs
be the Lebesgue decomposition of µ with respect to ν.
For N ∈ [1,∞), the “negative entropy” of µ with respect to ν is
EN(µ) = N − N∫
Xρ1− 1
N dν.
For N =∞,
E∞(µ) =
{∫X ρ log ρ dν if µ is a.c. w.r.t. ν,∞ otherwise.
![Page 87: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/87.jpg)
Entropy
Definition of the “negative entropy” function
EN : P(X )→ R ∪∞
Letµ = ρ ν + µs
be the Lebesgue decomposition of µ with respect to ν.
For N ∈ [1,∞), the “negative entropy” of µ with respect to ν is
EN(µ) = N − N∫
Xρ1− 1
N dν.
For N =∞,
E∞(µ) =
{∫X ρ log ρ dν if µ is a.c. w.r.t. ν,∞ otherwise.
![Page 88: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/88.jpg)
Entropy
Definition of the “negative entropy” function
EN : P(X )→ R ∪∞
Letµ = ρ ν + µs
be the Lebesgue decomposition of µ with respect to ν.
For N ∈ [1,∞), the “negative entropy” of µ with respect to ν is
EN(µ) = N − N∫
Xρ1− 1
N dν.
For N =∞,
E∞(µ) =
{∫X ρ log ρ dν if µ is a.c. w.r.t. ν,∞ otherwise.
![Page 89: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/89.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 90: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/90.jpg)
Convexity on Wasserstein space
(X ,d) is a compact length space.
ν is a fixed probability measure on X .
We want to ask whether the negative entropy function EN is aconvex function on P(X ).
That is, given µ0, µ1 ∈ P(X ), whether EN restricts to a convexfunction along a Wasserstein geodesic {µt}t∈[0,1] from µ0 to µ1.
![Page 91: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/91.jpg)
Convexity on Wasserstein space
(X ,d) is a compact length space.
ν is a fixed probability measure on X .
We want to ask whether the negative entropy function EN is aconvex function on P(X ).
That is, given µ0, µ1 ∈ P(X ), whether EN restricts to a convexfunction along a Wasserstein geodesic {µt}t∈[0,1] from µ0 to µ1.
![Page 92: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/92.jpg)
Nonnegative N-Ricci curvature
DefinitionGiven N ∈ [1,∞], we say that a compact measured lengthspace (X ,d , ν) has nonnegative N-Ricci curvature if :
For all µ0, µ1 ∈ P(X ) with supp(µ0) ⊂ supp(ν) andsupp(µ1) ⊂ supp(ν), there is some Wasserstein geodesic{µt}t∈[0,1] from µ0 to µ1 so that for all t ∈ [0,1],
EN(µt ) ≤ t EN(µ1) + (1− t) EN(µ0).
![Page 93: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/93.jpg)
Nonnegative N-Ricci curvature
DefinitionGiven N ∈ [1,∞], we say that a compact measured lengthspace (X ,d , ν) has nonnegative N-Ricci curvature if :
For all µ0, µ1 ∈ P(X ) with supp(µ0) ⊂ supp(ν) andsupp(µ1) ⊂ supp(ν), there is some Wasserstein geodesic{µt}t∈[0,1] from µ0 to µ1 so that for all t ∈ [0,1],
EN(µt ) ≤ t EN(µ1) + (1− t) EN(µ0).
![Page 94: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/94.jpg)
Nonnegative N-Ricci curvature
Note : We only require convexity along some geodesic from µ0to µ1, not all geodesics.
There’s also a notion of “N-Ricci curvature bounded below byK”.
![Page 95: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/95.jpg)
Nonnegative N-Ricci curvature
Note : We only require convexity along some geodesic from µ0to µ1, not all geodesics.
There’s also a notion of “N-Ricci curvature bounded below byK”.
![Page 96: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/96.jpg)
Nonnegative N-Ricci curvature
Note : We only require convexity along some geodesic from µ0to µ1, not all geodesics.
There’s also a notion of “N-Ricci curvature bounded below byK”.
![Page 97: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/97.jpg)
Main result
Theorem : (LV,S)
Let {(Xi ,di , νi)}∞i=1 be a sequence of compact measured lengthspaces with
limi→∞
(Xi ,di , νi) = (X ,d , ν)
in the measured Gromov-Hausdorff topology.
For any N ∈ [1,∞], if each (Xi ,di , νi) has nonnegative N-Riccicurvature then (X ,d , ν) has nonnegative N-Ricci curvature.
![Page 98: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/98.jpg)
What does all this have to do with Ricci curvature?
Let (M,g) be a compact connected n-dimensional Riemannianmanifold.
We could take the Riemannian measure, but let’s be moregeneral and consider any smooth measured length space.
Say Ψ ∈ C∞(M) has ∫M
e−Ψ dvolM = 1.
Put ν = e−Ψ dvolM .
![Page 99: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/99.jpg)
What does all this have to do with Ricci curvature?
Let (M,g) be a compact connected n-dimensional Riemannianmanifold.
We could take the Riemannian measure, but let’s be moregeneral and consider any smooth measured length space.
Say Ψ ∈ C∞(M) has ∫M
e−Ψ dvolM = 1.
Put ν = e−Ψ dvolM .
![Page 100: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/100.jpg)
The N-Ricci tensor
For N ∈ [1,∞], define the N-Ricci tensor RicN of (Mn,g, ν) byRic + Hess(Ψ) if N =∞,Ric + Hess(Ψ) − 1
N−n dΨ⊗ dΨ if n < N < ∞,Ric + Hess(Ψ) − ∞ (dΨ⊗ dΨ) if N = n,−∞ if N < n,
where by convention∞ · 0 = 0.
RicN is a symmetric covariant 2-tensor field on M that dependson g and Ψ.
(If N = n then RicN is −∞ except where dΨ = 0. There,RicN = Ric.)
Ric∞ = Bakry-Emery tensor = right-hand side of Perelman’smodified Ricci flow equation.
![Page 101: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/101.jpg)
The N-Ricci tensor
For N ∈ [1,∞], define the N-Ricci tensor RicN of (Mn,g, ν) byRic + Hess(Ψ) if N =∞,Ric + Hess(Ψ) − 1
N−n dΨ⊗ dΨ if n < N < ∞,Ric + Hess(Ψ) − ∞ (dΨ⊗ dΨ) if N = n,−∞ if N < n,
where by convention∞ · 0 = 0.
RicN is a symmetric covariant 2-tensor field on M that dependson g and Ψ.
(If N = n then RicN is −∞ except where dΨ = 0. There,RicN = Ric.)
Ric∞ = Bakry-Emery tensor = right-hand side of Perelman’smodified Ricci flow equation.
![Page 102: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/102.jpg)
Abstract Ricci recovers classical Ricci
Recall that ν = e−Ψ dvolM .
Theorem : (LV, S)For N ∈ [1,∞], the measured length space (M,g, ν) hasnonnegative N-Ricci curvature if and only if RicN ≥ 0.
Classical case : Ψ constant, so ν = dvolvol(M) .
Then (Mn,g, ν) has abstract nonnegative N-Ricci curvature ifand only if it has classical nonnegative Ricci curvature, as soonas N ≥ n.
![Page 103: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/103.jpg)
Abstract Ricci recovers classical Ricci
Recall that ν = e−Ψ dvolM .
Theorem : (LV, S)For N ∈ [1,∞], the measured length space (M,g, ν) hasnonnegative N-Ricci curvature if and only if RicN ≥ 0.
Classical case : Ψ constant, so ν = dvolvol(M) .
Then (Mn,g, ν) has abstract nonnegative N-Ricci curvature ifand only if it has classical nonnegative Ricci curvature, as soonas N ≥ n.
![Page 104: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/104.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 105: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/105.jpg)
Smooth limit spaces
Had Gromov precompactness theorem. What are the limitspaces (X ,d , ν)? Suppose that the limit space is a smoothmeasured length space, i.e.
(X ,d , ν) = (B,gB,e−Ψ dvolB)
for some n-dimensional smooth Riemannian manifold (B,gB)and some Ψ ∈ C∞(B).
Theorem : (LV)
If (B,gB,e−Ψ dvolB) is a measured Gromov-Hausdorff limit ofRiemannian manifolds with nonnegative Ricci curvature anddimension at most N then RicN(B) ≥ 0.
Note : the dimension can drop on taking limits.
![Page 106: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/106.jpg)
Smooth limit spaces
Had Gromov precompactness theorem. What are the limitspaces (X ,d , ν)? Suppose that the limit space is a smoothmeasured length space, i.e.
(X ,d , ν) = (B,gB,e−Ψ dvolB)
for some n-dimensional smooth Riemannian manifold (B,gB)and some Ψ ∈ C∞(B).
Theorem : (LV)
If (B,gB,e−Ψ dvolB) is a measured Gromov-Hausdorff limit ofRiemannian manifolds with nonnegative Ricci curvature anddimension at most N then RicN(B) ≥ 0.
Note : the dimension can drop on taking limits.
![Page 107: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/107.jpg)
Smooth limit spaces
Had Gromov precompactness theorem. What are the limitspaces (X ,d , ν)? Suppose that the limit space is a smoothmeasured length space, i.e.
(X ,d , ν) = (B,gB,e−Ψ dvolB)
for some n-dimensional smooth Riemannian manifold (B,gB)and some Ψ ∈ C∞(B).
Theorem : (LV)
If (B,gB,e−Ψ dvolB) is a measured Gromov-Hausdorff limit ofRiemannian manifolds with nonnegative Ricci curvature anddimension at most N then RicN(B) ≥ 0.
Note : the dimension can drop on taking limits.
![Page 108: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/108.jpg)
Bishop-Gromov-type inequality
Theorem : (LV, S)
If (X ,d , ν) has nonnegative N-Ricci curvature and x ∈ supp(ν)then r−N ν(Br (x)) is nonincreasing in r .
![Page 109: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/109.jpg)
Lichnerowicz inequality
If (M,g) is a compact Riemannian manifold, let λ1 be thesmallest positive eigenvalue of the Laplacian −∇2.
Theorem : (Lichnerowicz 1964)
If dim(M) = n and M has Ricci curvatures bounded below byK > 0 then
λ1 ≥n
n − 1K .
![Page 110: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/110.jpg)
Sharp global Poincare inequality
Theorem : (LV)
If (X ,d , ν) has N-Ricci curvature bounded below by K > 0 andf is a Lipschitz function on X with
∫X f dν = 0 then∫
Xf 2 dν ≤ N − 1
N1K
∫X|∇f |2 dν.
Here|∇f |(x) = lim sup
y→x
|f (y)− f (x)|d(y , x)
.
(There are also log Sobolev inequalities and Sobolevinequalities.)
![Page 111: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/111.jpg)
Sharp global Poincare inequality
Theorem : (LV)
If (X ,d , ν) has N-Ricci curvature bounded below by K > 0 andf is a Lipschitz function on X with
∫X f dν = 0 then∫
Xf 2 dν ≤ N − 1
N1K
∫X|∇f |2 dν.
Here|∇f |(x) = lim sup
y→x
|f (y)− f (x)|d(y , x)
.
(There are also log Sobolev inequalities and Sobolevinequalities.)
![Page 112: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/112.jpg)
Sharp global Poincare inequality
Theorem : (LV)
If (X ,d , ν) has N-Ricci curvature bounded below by K > 0 andf is a Lipschitz function on X with
∫X f dν = 0 then∫
Xf 2 dν ≤ N − 1
N1K
∫X|∇f |2 dν.
Here|∇f |(x) = lim sup
y→x
|f (y)− f (x)|d(y , x)
.
(There are also log Sobolev inequalities and Sobolevinequalities.)
![Page 113: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/113.jpg)
Ricci O’Neill Theorem
Theorem (O’Neill 1966) Sectional curvature is nondecreasingunder a Riemannian submersion.
Theorem (JL 2003) RicN is nondecreasing under aRiemannian submersion.
LV : A synthetic proof of the Ricci O’Neill theorem usingdisplacement convexity.
![Page 114: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/114.jpg)
Ricci O’Neill Theorem
Theorem (O’Neill 1966) Sectional curvature is nondecreasingunder a Riemannian submersion.
Theorem (JL 2003) RicN is nondecreasing under aRiemannian submersion.
LV : A synthetic proof of the Ricci O’Neill theorem usingdisplacement convexity.
![Page 115: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/115.jpg)
Ricci O’Neill Theorem
Theorem (O’Neill 1966) Sectional curvature is nondecreasingunder a Riemannian submersion.
Theorem (JL 2003) RicN is nondecreasing under aRiemannian submersion.
LV : A synthetic proof of the Ricci O’Neill theorem usingdisplacement convexity.
![Page 116: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/116.jpg)
Some open questions
Do measured length spaces with nonnegative N-Riccicurvature admit isoperimetric inequalities?
To what extent does the Cheeger-Gromoll splitting principlehold for measured length spaces with nonnegative N-Riccicurvature?
![Page 117: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/117.jpg)
Some open questions
Do measured length spaces with nonnegative N-Riccicurvature admit isoperimetric inequalities?
To what extent does the Cheeger-Gromoll splitting principlehold for measured length spaces with nonnegative N-Riccicurvature?
![Page 118: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/118.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 119: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/119.jpg)
Ricci flow
M a compact, connected n-dimensional manifold.
Say (M,g(t)) is a Ricci flow solution, i.e. dgdt = − 2 Ric.
Fix t0 and put τ = t0 − t . Then dgdτ = 2 Ric.
An important tool : monotonic quantities.
![Page 120: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/120.jpg)
Ricci flow
M a compact, connected n-dimensional manifold.
Say (M,g(t)) is a Ricci flow solution, i.e. dgdt = − 2 Ric.
Fix t0 and put τ = t0 − t . Then dgdτ = 2 Ric.
An important tool : monotonic quantities.
![Page 121: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/121.jpg)
Reduced volume
Fix p ∈ M. Say γ : [0, τ ]→ M is a smooth curve with γ(0) = p.(The graph of γ goes “backward in time”.)
DefinitionL-length L(γ) =
∫ τ0√τ(|γ|2g(τ) + R(γ(τ), τ)
)dτ .
Definitionreduced distance Given q ∈ M, putL(q, τ) = inf{L(γ) : γ(0) = p, γ(τ) = q}.
Put l(q, τ) = L(q,τ)
2√τ
.
Definitionreduced volume V (τ) = τ−
n2∫
M e− l(q,τ) dvol(q).
![Page 122: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/122.jpg)
Reduced volume
Fix p ∈ M. Say γ : [0, τ ]→ M is a smooth curve with γ(0) = p.(The graph of γ goes “backward in time”.)
DefinitionL-length L(γ) =
∫ τ0√τ(|γ|2g(τ) + R(γ(τ), τ)
)dτ .
Definitionreduced distance Given q ∈ M, putL(q, τ) = inf{L(γ) : γ(0) = p, γ(τ) = q}.
Put l(q, τ) = L(q,τ)
2√τ
.
Definitionreduced volume V (τ) = τ−
n2∫
M e− l(q,τ) dvol(q).
![Page 123: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/123.jpg)
Reduced volume
Fix p ∈ M. Say γ : [0, τ ]→ M is a smooth curve with γ(0) = p.(The graph of γ goes “backward in time”.)
DefinitionL-length L(γ) =
∫ τ0√τ(|γ|2g(τ) + R(γ(τ), τ)
)dτ .
Definitionreduced distance Given q ∈ M, putL(q, τ) = inf{L(γ) : γ(0) = p, γ(τ) = q}.
Put l(q, τ) = L(q,τ)
2√τ
.
Definitionreduced volume V (τ) = τ−
n2∫
M e− l(q,τ) dvol(q).
![Page 124: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/124.jpg)
Reduced volume
Fix p ∈ M. Say γ : [0, τ ]→ M is a smooth curve with γ(0) = p.(The graph of γ goes “backward in time”.)
DefinitionL-length L(γ) =
∫ τ0√τ(|γ|2g(τ) + R(γ(τ), τ)
)dτ .
Definitionreduced distance Given q ∈ M, putL(q, τ) = inf{L(γ) : γ(0) = p, γ(τ) = q}.
Put l(q, τ) = L(q,τ)
2√τ
.
Definitionreduced volume V (τ) = τ−
n2∫
M e− l(q,τ) dvol(q).
![Page 125: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/125.jpg)
Reduced volume
Fix p ∈ M. Say γ : [0, τ ]→ M is a smooth curve with γ(0) = p.(The graph of γ goes “backward in time”.)
DefinitionL-length L(γ) =
∫ τ0√τ(|γ|2g(τ) + R(γ(τ), τ)
)dτ .
Definitionreduced distance Given q ∈ M, putL(q, τ) = inf{L(γ) : γ(0) = p, γ(τ) = q}.
Put l(q, τ) = L(q,τ)
2√τ
.
Definitionreduced volume V (τ) = τ−
n2∫
M e− l(q,τ) dvol(q).
![Page 126: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/126.jpg)
Monotonicity of the reduced volume
Theorem : (Perelman 2002)
V is nonincreasing in τ , i.e. nondecreasing in t .
An “entropy” functional for Ricci flow.
The only assumption : g(t) satisfies the Ricci flow equation.
Main application : Perelman’s “no local collapsing” theorem.
![Page 127: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/127.jpg)
Monotonicity of the reduced volume
Theorem : (Perelman 2002)
V is nonincreasing in τ , i.e. nondecreasing in t .
An “entropy” functional for Ricci flow.
The only assumption : g(t) satisfies the Ricci flow equation.
Main application : Perelman’s “no local collapsing” theorem.
![Page 128: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/128.jpg)
Perelman’s heuristic derivation
Put M = M × SN × R+.
Here N is a free parameter and τ is the coordinate on R+.
Put
g = g(τ) + 2NτgSN +
(N2τ
+ R)
dτ2.
Fact : As N →∞, Ric(M) = O(N−1).
Bishop-Gromov : r− dim vol(Br (p)) is nonincreasing in r ifRic ≥ 0.
Apply formally to M and take N →∞. Get monotonicity of V .
![Page 129: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/129.jpg)
Perelman’s heuristic derivation
Put M = M × SN × R+.
Here N is a free parameter and τ is the coordinate on R+.
Put
g = g(τ) + 2NτgSN +
(N2τ
+ R)
dτ2.
Fact : As N →∞, Ric(M) = O(N−1).
Bishop-Gromov : r− dim vol(Br (p)) is nonincreasing in r ifRic ≥ 0.
Apply formally to M and take N →∞. Get monotonicity of V .
![Page 130: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/130.jpg)
Perelman’s heuristic derivation
Put M = M × SN × R+.
Here N is a free parameter and τ is the coordinate on R+.
Put
g = g(τ) + 2NτgSN +
(N2τ
+ R)
dτ2.
Fact : As N →∞, Ric(M) = O(N−1).
Bishop-Gromov : r− dim vol(Br (p)) is nonincreasing in r ifRic ≥ 0.
Apply formally to M and take N →∞. Get monotonicity of V .
![Page 131: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/131.jpg)
Heuristic relation to optimal transport
We know how to characterize nonnegative Ricci curvatureusing optimal transport. Apply to M and translate down to M.
This should give an optimal transport problem on M with whichwe can derive the monotonicity of V .
We’ll describe a (re)proof of the monotonicity of V , usingoptimal transport methods.
![Page 132: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/132.jpg)
Heuristic relation to optimal transport
We know how to characterize nonnegative Ricci curvatureusing optimal transport. Apply to M and translate down to M.
This should give an optimal transport problem on M with whichwe can derive the monotonicity of V .
We’ll describe a (re)proof of the monotonicity of V , usingoptimal transport methods.
![Page 133: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/133.jpg)
Heuristic relation to optimal transport
We know how to characterize nonnegative Ricci curvatureusing optimal transport. Apply to M and translate down to M.
This should give an optimal transport problem on M with whichwe can derive the monotonicity of V .
We’ll describe a (re)proof of the monotonicity of V , usingoptimal transport methods.
![Page 134: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/134.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 135: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/135.jpg)
Say c : [0,1]→ P∞(M) is a smooth curve.
Write c(t) = ρ(t) dvolM .
Solve∂ρ
∂t= −
∑i
∇i(ρ∇iφ)
for φ ≡ φ(t) ∈ C∞(M).
From {ρ(t)}t∈[0,1], we got {φ(t)}t∈[0,1]. Put
E(c) =12
∫ 1
0
∫M|∇φ|2 ρ dvolM dt .
![Page 136: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/136.jpg)
The Euler-Lagrange equation for the functional E is
∂φ
∂t= − 1
2|∇φ|2.
Then
d2
dt2
∫Mρ log ρdvolM =
∫M
[|Hessφ|2 + RicM(∇φ,∇φ)
]ρdvolM .
![Page 137: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/137.jpg)
The Euler-Lagrange equation for the functional E is
∂φ
∂t= − 1
2|∇φ|2.
Then
d2
dt2
∫Mρ log ρdvolM =
∫M
[|Hessφ|2 + RicM(∇φ,∇φ)
]ρdvolM .
![Page 138: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/138.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 139: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/139.jpg)
Question
Can we do something similar for the Ricci flow?
Principle : Satisfying Ric = 0 in the Riemannian case is likesatisfying the Ricci flow equation in the spacetime case.
Optimal transport in a Ricci flow background was firstconsidered by Peter Topping, with application to anothermonotonic quantity (W-functional).
Note : The Ricci flow equation
dgdt
= − 2 Ric
impliesdvolM
dt= − R dvolM .
![Page 140: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/140.jpg)
Question
Can we do something similar for the Ricci flow?
Principle : Satisfying Ric = 0 in the Riemannian case is likesatisfying the Ricci flow equation in the spacetime case.
Optimal transport in a Ricci flow background was firstconsidered by Peter Topping, with application to anothermonotonic quantity (W-functional).
Note : The Ricci flow equation
dgdt
= − 2 Ric
impliesdvolM
dt= − R dvolM .
![Page 141: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/141.jpg)
E0 functional
Assume hereafter that (M,g(t)) satisfies the Ricci flowequation.
Given c : [t0, t1]→ P∞(M), write c(t) = ρ(t) dvolM . Solve
∂ρ
∂t= −
∑i
∇i(ρ∇iφ) + R ρ
for φ ≡ φ(t) ∈ C∞(M).
DefinitionE0(c) = 1
2
∫ t1t0
∫M
(|∇φ|2 + R
)ρ dvolM dt
Euler-Lagrange equation for E0 :
∂φ
∂t= − 1
2|∇φ|2 +
12
R.
![Page 142: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/142.jpg)
E0 functional
Assume hereafter that (M,g(t)) satisfies the Ricci flowequation.
Given c : [t0, t1]→ P∞(M), write c(t) = ρ(t) dvolM . Solve
∂ρ
∂t= −
∑i
∇i(ρ∇iφ) + R ρ
for φ ≡ φ(t) ∈ C∞(M).
DefinitionE0(c) = 1
2
∫ t1t0
∫M
(|∇φ|2 + R
)ρ dvolM dt
Euler-Lagrange equation for E0 :
∂φ
∂t= − 1
2|∇φ|2 +
12
R.
![Page 143: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/143.jpg)
E0 functional
Assume hereafter that (M,g(t)) satisfies the Ricci flowequation.
Given c : [t0, t1]→ P∞(M), write c(t) = ρ(t) dvolM . Solve
∂ρ
∂t= −
∑i
∇i(ρ∇iφ) + R ρ
for φ ≡ φ(t) ∈ C∞(M).
DefinitionE0(c) = 1
2
∫ t1t0
∫M
(|∇φ|2 + R
)ρ dvolM dt
Euler-Lagrange equation for E0 :
∂φ
∂t= − 1
2|∇φ|2 +
12
R.
![Page 144: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/144.jpg)
Convexity statement
PropositionIf c satisfies the Euler-Lagrange equation then
d2
dt2
∫M
(ρ ln ρ − φ ρ) dvolM =
∫M|Ric−Hessφ|2 ρ dvolM .
CorollaryIf c satisfies the Euler-Lagrange equation then∫
M (ρ ln ρ − φ ρ) dvolM is convex in t.
![Page 145: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/145.jpg)
Convexity statement
PropositionIf c satisfies the Euler-Lagrange equation then
d2
dt2
∫M
(ρ ln ρ − φ ρ) dvolM =
∫M|Ric−Hessφ|2 ρ dvolM .
CorollaryIf c satisfies the Euler-Lagrange equation then∫
M (ρ ln ρ − φ ρ) dvolM is convex in t.
![Page 146: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/146.jpg)
Corresponding optimal transport problem
Say we want to transport a measure µ0 (at time t0) to ameasure µ1 (at time t1).
Take the cost to transport a unit of mass from p to q to be
min{L0(γ) : γ(t0) = p, γ(t1) = q},
where
L0(γ) =12
∫ t1
t0
(|γ|2g(t) + R(γ(t), t)
)dt .
There is a corresponding notion of optimal transport,displacement interpolation, etc.
![Page 147: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/147.jpg)
E− functional
Fix t0 and put τ = t0 − t . The Ricci flow equation is
dgdτ
= 2 Ric .
Given c : [τ0, τ1]→ P∞(M), write c(τ) = ρ(τ) dvolM . Solve
∂ρ
∂τ= −
∑i
∇i(ρ∇iφ) − R ρ
for φ = φ(τ) ∈ C∞(M).
DefinitionE−(c) =
∫ τ1τ0
∫M√τ(|∇φ|2 + R
)ρ dvolM dτ
Euler-Lagrange equation for E− :
∂φ
∂τ= − 1
2|∇φ|2 +
12
R − 12τ
φ.
![Page 148: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/148.jpg)
E− functional
Fix t0 and put τ = t0 − t . The Ricci flow equation is
dgdτ
= 2 Ric .
Given c : [τ0, τ1]→ P∞(M), write c(τ) = ρ(τ) dvolM . Solve
∂ρ
∂τ= −
∑i
∇i(ρ∇iφ) − R ρ
for φ = φ(τ) ∈ C∞(M).
DefinitionE−(c) =
∫ τ1τ0
∫M√τ(|∇φ|2 + R
)ρ dvolM dτ
Euler-Lagrange equation for E− :
∂φ
∂τ= − 1
2|∇φ|2 +
12
R − 12τ
φ.
![Page 149: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/149.jpg)
E− functional
Fix t0 and put τ = t0 − t . The Ricci flow equation is
dgdτ
= 2 Ric .
Given c : [τ0, τ1]→ P∞(M), write c(τ) = ρ(τ) dvolM . Solve
∂ρ
∂τ= −
∑i
∇i(ρ∇iφ) − R ρ
for φ = φ(τ) ∈ C∞(M).
DefinitionE−(c) =
∫ τ1τ0
∫M√τ(|∇φ|2 + R
)ρ dvolM dτ
Euler-Lagrange equation for E− :
∂φ
∂τ= − 1
2|∇φ|2 +
12
R − 12τ
φ.
![Page 150: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/150.jpg)
E− functional
Fix t0 and put τ = t0 − t . The Ricci flow equation is
dgdτ
= 2 Ric .
Given c : [τ0, τ1]→ P∞(M), write c(τ) = ρ(τ) dvolM . Solve
∂ρ
∂τ= −
∑i
∇i(ρ∇iφ) − R ρ
for φ = φ(τ) ∈ C∞(M).
DefinitionE−(c) =
∫ τ1τ0
∫M√τ(|∇φ|2 + R
)ρ dvolM dτ
Euler-Lagrange equation for E− :
∂φ
∂τ= − 1
2|∇φ|2 +
12
R − 12τ
φ.
![Page 151: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/151.jpg)
Convexity statement
PropositionIf c satisfies the Euler-Lagrange equation then(
τ32
ddτ
)2(∫M
(ρ ln ρ + φ ρ) dvolM +n2
ln τ)
=
τ3∫
M
∣∣∣Ric + Hessφ− g2τ
∣∣∣2 ρ dvolM .
CorollaryIf c satisfies the Euler-Lagrange equation then∫
M (ρ ln ρ + φ ρ) dvolM + n2 ln τ is convex in the variable
s = τ−12 .
![Page 152: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/152.jpg)
Convexity statement
PropositionIf c satisfies the Euler-Lagrange equation then(
τ32
ddτ
)2(∫M
(ρ ln ρ + φ ρ) dvolM +n2
ln τ)
=
τ3∫
M
∣∣∣Ric + Hessφ− g2τ
∣∣∣2 ρ dvolM .
CorollaryIf c satisfies the Euler-Lagrange equation then∫
M (ρ ln ρ + φ ρ) dvolM + n2 ln τ is convex in the variable
s = τ−12 .
![Page 153: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/153.jpg)
Geometry of the space of probability measures
Motivation
Optimal transport
Formal Geometry of Wasserstein Space
Metric geometry of Wasserstein space
Ricci meets Wasserstein
Some more metric geometry
Generalized entropy functionals
Abstract Ricci curvature
Applications
Perelman’s reduced volume
Formulas from Riemannian optimal transport
Optimal transport for Ricci flow
Monotonicity of the reduced volume
![Page 154: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/154.jpg)
The transport problem
Take τ0 → 0, µ0 = δp and µ1 an absolutely continuous measure.
The displacement interpolation is along L-geodesics emanatingfrom p.
In this case, φ = l .
![Page 155: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/155.jpg)
From convexity to monotonicity
PropositionIn this case,
∫M (ρ ln ρ + φ ρ) dvolM + n
2 ln τ is nondecreasingin τ .
Proof.We know that it is convex in s = τ−
12 . As s →∞, i.e. as τ → 0,
it approaches a constant. (Almost Euclidean situation.) So it isnonincreasing in s, i.e. nondecreasing in τ .
![Page 156: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/156.jpg)
From convexity to monotonicity
PropositionIn this case,
∫M (ρ ln ρ + φ ρ) dvolM + n
2 ln τ is nondecreasingin τ .
Proof.We know that it is convex in s = τ−
12 . As s →∞, i.e. as τ → 0,
it approaches a constant. (Almost Euclidean situation.) So it isnonincreasing in s, i.e. nondecreasing in τ .
![Page 157: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/157.jpg)
Free energy
Trivial fact : The minimizer of∫M
(ρ ln ρ + φ ρ) dvolM +n2
ln τ,
as ρdvolM ranges over absolutely continuous probabilitymeasures, is
− ln(τ−
n2
∫M
e− φ dvolM
).
The minimizing measure is given by
ρ =e−φ∫
M e− φ dvolM.
![Page 158: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/158.jpg)
Monotonicity of reduced volume
Proposition
V (τ) = τ−n2
∫M
e− l dvolM
is nonincreasing in τ .
Proof : Say τ ′ < τ ′′. Recall that φ = l . Takeµ(τ ′′) = ρ(τ ′′) dvolM with
ρ(τ ′′) =e−φ(τ ′′)∫
M e− φ(τ ′′) dvolM.
Transport it to δp (at time zero). At the intermediate time τ ′ wesee a measure µ(τ ′) = ρ(τ ′) dvolM .
![Page 159: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/159.jpg)
Monotonicity of reduced volume
Proposition
V (τ) = τ−n2
∫M
e− l dvolM
is nonincreasing in τ .Proof : Say τ ′ < τ ′′. Recall that φ = l . Takeµ(τ ′′) = ρ(τ ′′) dvolM with
ρ(τ ′′) =e−φ(τ ′′)∫
M e− φ(τ ′′) dvolM.
Transport it to δp (at time zero). At the intermediate time τ ′ wesee a measure µ(τ ′) = ρ(τ ′) dvolM .
![Page 160: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/160.jpg)
Monotonicity of reduced volume
Proposition
V (τ) = τ−n2
∫M
e− l dvolM
is nonincreasing in τ .Proof : Say τ ′ < τ ′′. Recall that φ = l . Takeµ(τ ′′) = ρ(τ ′′) dvolM with
ρ(τ ′′) =e−φ(τ ′′)∫
M e− φ(τ ′′) dvolM.
Transport it to δp (at time zero). At the intermediate time τ ′ wesee a measure µ(τ ′) = ρ(τ ′) dvolM .
![Page 161: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/161.jpg)
Proof
Then
− ln(
(τ ′)−n2
∫M
e− φ(τ ′) dvolM
)≤∫
M
[ρ(τ ′) ln ρ(τ ′) + φ(τ ′) ρ(τ ′)
]dvolM +
n2
ln τ ′
≤∫
M
[ρ(τ ′′) ln ρ(τ ′′) + φ(τ ′′) ρ(τ ′′)
]dvolM +
n2
ln τ ′′
= − ln(
(τ ′′)−n2
∫M
e− φ(τ ′′) dvolM
).
End of proof
![Page 162: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/162.jpg)
Proof
Then
− ln(
(τ ′)−n2
∫M
e− φ(τ ′) dvolM
)≤∫
M
[ρ(τ ′) ln ρ(τ ′) + φ(τ ′) ρ(τ ′)
]dvolM +
n2
ln τ ′
≤∫
M
[ρ(τ ′′) ln ρ(τ ′′) + φ(τ ′′) ρ(τ ′′)
]dvolM +
n2
ln τ ′′
= − ln(
(τ ′′)−n2
∫M
e− φ(τ ′′) dvolM
).
End of proof
![Page 163: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/163.jpg)
Proof
Then
− ln(
(τ ′)−n2
∫M
e− φ(τ ′) dvolM
)≤∫
M
[ρ(τ ′) ln ρ(τ ′) + φ(τ ′) ρ(τ ′)
]dvolM +
n2
ln τ ′
≤∫
M
[ρ(τ ′′) ln ρ(τ ′′) + φ(τ ′′) ρ(τ ′′)
]dvolM +
n2
ln τ ′′
= − ln(
(τ ′′)−n2
∫M
e− φ(τ ′′) dvolM
).
End of proof
![Page 164: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/164.jpg)
Proof
Then
− ln(
(τ ′)−n2
∫M
e− φ(τ ′) dvolM
)≤∫
M
[ρ(τ ′) ln ρ(τ ′) + φ(τ ′) ρ(τ ′)
]dvolM +
n2
ln τ ′
≤∫
M
[ρ(τ ′′) ln ρ(τ ′′) + φ(τ ′′) ρ(τ ′′)
]dvolM +
n2
ln τ ′′
= − ln(
(τ ′′)−n2
∫M
e− φ(τ ′′) dvolM
).
End of proof
![Page 165: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/165.jpg)
Optimal transport and heat flow
Otto, Otto-Westdickenberg Suppose that a compactRiemannian manifold has Ric ≥ 0. If µ0(t) and µ1(t) are twosolutions of the heat flow on measures then W2(µ0(t), µ1(t)) isnonincreasing in t .
J.L., McCann-Topping Suppose that (M,g(t)) is a Ricci flowsolution. Suppose that µ0(t) and µ1(t) are two solutions of thebackward heat flow on measures
dµdt
= −∇2g(t) µ.
Then W2(µ0(t), µ1(t)) is nondecreasing in t .
Topping Extension to a statement about the L-transportdistance between µ0 and µ1 at distinct but related times.
![Page 166: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/166.jpg)
Optimal transport and heat flow
Otto, Otto-Westdickenberg Suppose that a compactRiemannian manifold has Ric ≥ 0. If µ0(t) and µ1(t) are twosolutions of the heat flow on measures then W2(µ0(t), µ1(t)) isnonincreasing in t .
J.L., McCann-Topping Suppose that (M,g(t)) is a Ricci flowsolution. Suppose that µ0(t) and µ1(t) are two solutions of thebackward heat flow on measures
dµdt
= −∇2g(t) µ.
Then W2(µ0(t), µ1(t)) is nondecreasing in t .
Topping Extension to a statement about the L-transportdistance between µ0 and µ1 at distinct but related times.
![Page 167: The geometry of the space of measures and its applications](https://reader036.vdocuments.mx/reader036/viewer/2022071602/613d55e6736caf36b75c1ce4/html5/thumbnails/167.jpg)
Optimal transport and heat flow
Otto, Otto-Westdickenberg Suppose that a compactRiemannian manifold has Ric ≥ 0. If µ0(t) and µ1(t) are twosolutions of the heat flow on measures then W2(µ0(t), µ1(t)) isnonincreasing in t .
J.L., McCann-Topping Suppose that (M,g(t)) is a Ricci flowsolution. Suppose that µ0(t) and µ1(t) are two solutions of thebackward heat flow on measures
dµdt
= −∇2g(t) µ.
Then W2(µ0(t), µ1(t)) is nondecreasing in t .
Topping Extension to a statement about the L-transportdistance between µ0 and µ1 at distinct but related times.