Kurdyka- ojasiewicz-Simon inequality for gradient flows in ...este.uca.ma/wedpta18/mazon.pdf · L.Ambrosio, N.Gigli, and G. Savar e.Gradient ows in metric spaces and in the space

Kurdyka- Lojasiewicz-Simon inequality for gradientflows in metric spaces

J.M. Mazonjoint works with Daniel Hauer (Sydney University)

EDP, Transport Optimal et Applications. Essaouira, 2018.

J.M. Mazon joint works with Daniel Hauer (Sydney University) K LS-inequality for gradient flows in metric spaces

Introduction

S. Lojasiewicz. Une propriete topologique des sous-ensembles analytiquesreels. In Les Equations aux Derivees Partielles (Paris, 1962), pages 87–89.Editions du Centre National de la Recherche Scientifique, Paris, 1963.

Every real-analytic function f : U → R defined on a neighborhoodU ⊆ RN of a, with f (a) = 0 satifies Lojasiewicz ( L-)inequality

‖∇f (x)‖ ≥ |f (x)|α for all x ∈ U (1)

for some exponent 0 < α ≤ 1.

If the function f : U ⊂ RN → R satisfies inequality (1), then thesolutions of the gradient system

v ′(t) +∇f (v(t)) = 0 for t > 0,

v(0) = v0,(2)

trend to equilibrium.


Introduction



‖∇f (x)‖ ≥ |f (x)|α for all x ∈ U (1)



v ′(t) +∇f (v(t)) = 0 for t > 0,

v(0) = v0,(2)



Introduction



‖∇f (x)‖ ≥ |f (x)|α for all x ∈ U (1)



v ′(t) +∇f (v(t)) = 0 for t > 0,

v(0) = v0,(2)



Introduction

K.Kurdyka. On gradients of functions definable in o-minimal structures.Ann. Inst. Fourier (Grenoble), 48(3):769–783, 1998.

Let f : U ⊂ RN → (0,∞) be a differentiable M-function where U is anopen and bounded subset of RN . Then there is c > 0, ρ > 0 a strictlyincreasing function θ ∈ C 1,1

loc (R) satisfying θ(0) = 0 such that

‖∇(θ f )(x)‖ ≥ c for all x ∈ U , f (x) ∈ (0, ρ) (3)

(3) is called Kurdyka- Lojasiewicz (K L-)inequality

Kurdyka- Lojasiewicz (K L-)inequality implies that the solutions of (2)have their length bounded by a constant independent of the trajectory.

L. Simon, Asymptotics for a class of nonlinear evolution equations, withapplications to geometric problems, Ann. of Math. (2) 118 (1983),525–571.


Introduction









Introduction









Introduction









Introduction









Introduction

Simon’s results imply that the solution of ut −∆u = f (u) in [0,∞× Ω

u = 0 on [0,∞× ∂Ω

converges of some equilibrium pint when f is analytic.

Lojasiewicz-Simon ( L-S-)inequality

A. Haraux, A. Jendoubi, R. Chill, S.-Z- Huang, Bolte, A. Danilidis, O.Lay, L. Mazet, A. Blanchet, Rybka, Fereisl etc

R.Jordan, D. Kinderlehrer, and F. Otto. The variational formulation ofthe Fokker-Planck equation. SIAM J. Math. Anal., 29(1):1–17, 1998.

“the classical linear heat flow in RN can be derived as the gradient flowof the Boltzmann entropy with respect to the 2-Wasserstein metric onthe space of probability measures on RN”.


Introduction


u = 0 on [0,∞× ∂Ω







Introduction


u = 0 on [0,∞× ∂Ω







Introduction


u = 0 on [0,∞× ∂Ω







Introduction


u = 0 on [0,∞× ∂Ω







Introduction

L.Ambrosio, N.Gigli, and G. Savare. Gradient flows in metric spaces andin the space of probability measures. Lectures in Mathematics ETHZurich. Birkhauser Verlag, Basel, second edition, 2008.

De Giorgi

curves of maximal slope

theory of minimizing moviments

(M, d) denotes a complete metric space. 1 ≤ p <∞ and p′ := pp−1 be

the Holder conjugate


Introduction


De Giorgi






Introduction


De Giorgi






Gradient flows in metric spaces

Definition (Strong upper gradient)

For a proper functional E : M→ (−∞,∞], a proper functionalg : M→ [0,+∞] is called a strong upper gradient of E if for every curvev ∈ AC(0,+∞;M), the composition function g v : (0,+∞)→ [0,∞] isBorel-measurable and

|E(v(t))− E(v(s))| ≤∫ t

s

g(v(r)) |v ′|(r) dr for all a < s ≤ t < b

where

|v ′|(t) := lims→t

d(v(s), v(t))

|s − t| the metric derivative

Definition

For a given functional E : M→ (−∞,∞], the descending slope|D−E| : M→ [0,+∞] of E is given by

|D−E|(u) :=

lim supv→u

[E(v)− E(u)]−

d(v , u)if u ∈ D(E),

+∞ if otherwise.



Definition (Strong upper gradient)

For a proper functional E : M→ (−∞,∞], a proper functionalg : M→ [0,+∞] is called a strong upper gradient of E if for every curvev ∈ AC(0,+∞;M), the composition function g v : (0,+∞)→ [0,∞] isBorel-measurable and

|E(v(t))− E(v(s))| ≤∫ t

s

g(v(r)) |v ′|(r) dr for all a < s ≤ t < b

where

|v ′|(t) := lims→t

d(v(s), v(t))

|s − t| the metric derivative

Definition

For a given functional E : M→ (−∞,∞], the descending slope|D−E| : M→ [0,+∞] of E is given by

|D−E|(u) :=

lim supv→u

[E(v)− E(u)]−

d(v , u)if u ∈ D(E),

+∞ if otherwise.



Definition

For a proper functional E : M→ (−∞,∞] with strong upper gradient g ,a curve v ∈ ACloc(0,+∞;M) is called a p-curve of maximal slope of E ifE v : (0,+∞)→ R is non-increasing and

d

dtE(v(t)) ≤ − 1

p′ gp′ (v(t))− 1

p |v′|p(t) for almost every t ∈ (0,+∞)

Definition (p-gradient flows in M)

For a proper functional E : M→ (−∞,∞] with strong upper gradient g ,a curve v ∈ ACloc(0,+∞;M) is called a p-gradient flow of E or also ap-gradient flow in M if v is a p-curve of maximal slope of E with respectto g .



Definition

For a proper functional E : M→ (−∞,∞] with strong upper gradient g ,a curve v ∈ ACloc(0,+∞;M) is called a p-curve of maximal slope of E ifE v : (0,+∞)→ R is non-increasing and

d

dtE(v(t)) ≤ − 1

p′ gp′ (v(t))− 1

p |v′|p(t) for almost every t ∈ (0,+∞)

Definition (p-gradient flows in M)

For a proper functional E : M→ (−∞,∞] with strong upper gradient g ,a curve v ∈ ACloc(0,+∞;M) is called a p-gradient flow of E or also ap-gradient flow in M if v is a p-curve of maximal slope of E with respectto g .



Proposition

Let E : M→ (−∞,∞] be proper functional with strong upper gradientg , and v ∈ ACloc(0,+∞;M). Then the following statements areequivalent.

(1) v is a p-gradient flow of E .

(2) E v : (0,+∞)→ R is non-increasing and energy dissipationequality

E(v(s))− E(v(t)) = 1p

∫ t

s

|v ′|p(r) dr + 1p′

∫ t

s

gp′ (v(r)) dr

holds for all 0 < s < t < +∞.

Moreover, for every p-gradient flow v of E , one has that

d

dtE(v(t)) = −|v ′|p(t) = −gp′ (v(t)) = −(g v)(t) · |v ′|(t)

for almost every t ∈ (0,+∞),



Definition (constant speed geodesics and λ-geodesic convexity)

A curve γ : [0, 1]→M is said to be a constant speed geodesicconnecting two points v0, v1 ∈M if γ(0) = v0, γ(1) = v1 and

d(γ(s), γ(t)) = (t − s)d(v0, v1) for all s, t ∈ [0, 1] with s ≤ t.

A metric space (M, d) with the property that for every two elements v0,v1 ∈M, there is at least one constant speed geodesic γ ⊆M connectingv0 and v1 is called a length space. Given λ ∈ R, a functionalE : M→ (−∞,∞] is called λ-geodesically convex if for every v0,v1 ∈ D(E), there is a constant speed geodesic γ ⊆M connecting v0 andv1 such that E is λ-convex along γ, tha is,

E(γ(t)) ≤ (1− t)E(γ(0)) + tE(γ(1))− λ2 t(1− t)d2(γ(0), γ(1))

for all t ∈ [0, 1],



Proposition

For λ ∈ R, let E : M→ (−∞,∞] be a proper λ-geodesically convexfunctional on a length space M. Then, if E is lower semicontinuous, thenthe descending slope |D−E| of E is a strong upper gradient of E , and|D−E| is lower semicontinuous.

Definition

An element ϕ ∈M is called an equilibrium point (or also critical point)of a proper functional E : M→ (−∞,∞] with strong upper gradient g ifϕ ∈ D(g) and g(ϕ) = 0. We denote by Eg = g−1(0) the set of allequilibrium points of E with respect to strong upper gradient g .

We denote by argmin(E) the set of all global minimisers ϕ of E .

If for λ ≥ 0, E is λ-geodesically convex, then

argmin(E) =ϕ ∈ D(|D−E|)

∣∣∣ |D−E|(ϕ) = 0

= E|D−E|.



Proposition


Definition





∣∣∣ |D−E|(ϕ) = 0

= E|D−E|.



Proposition


Definition





∣∣∣ |D−E|(ϕ) = 0

= E|D−E|.



Proposition


Definition





∣∣∣ |D−E|(ϕ) = 0

= E|D−E|.



Definition

For a curve v ∈ C ((0,∞);M), the set

ω(v) :=ϕ ∈M

∣∣∣ there is tn ↑ +∞ s.t. limn→∞

v(tn) = ϕ in M

is called the ω-limit set of v .

We denote byIt0 (v) =

v(t)

∣∣ t ≥ t0

For a proper functional E : M→ (−∞,∞] and ϕ ∈ D(E), we call thefunctional E(·|ϕ) : M→ (−∞,∞] defined by

E(v |ϕ) = E(v)− E(ϕ) for every ϕ ∈M

the relative entropy or relative energy of E with respect to ϕ.



Definition


ω(v) :=ϕ ∈M


v(tn) = ϕ in M



v(t)

∣∣ t ≥ t0






Definition


ω(v) :=ϕ ∈M


v(tn) = ϕ in M



v(t)

∣∣ t ≥ t0





Trend to equilibrium in the entropy sense

Proposition

Let E : M→ (−∞,+∞] be a proper functional, g a strong uppergradient of E , and v a p-gradient flow of E . Then, the followingstatements hold.

(1) E is a strict Lyapunov function of v .

(2) (Trend to equilibrium in the entropy sense) If for t0 ≥ 0, Erestricted on the set It0 (v) is lower semicontinuous, then for everyϕ ∈ ω(v), one has ϕ ∈ D(E) and

limt→∞

E(v(t)) = E(ϕ) = infξ∈It0 (v)

E(ξ). (4)

(3) (ω-limit points are equilibrium points of E) Suppose for t0 ≥ 0, Erestricted on the set It0 (v) is bounded from below and g restrictedon the set It0 (v) is lower semicontinuous. Then the ω-limit set ω(v)of v is contained in the set Eg of equilibrium points of E .


Kurdyka- Lojasiewicz-Simon inequalities in metric spaces

Definition

A proper functional E : M→ (−∞,+∞] with strong upper gradient gand equilibrium point ϕ ∈ Eg is said to satisfy a Kurdyka- Lojasiewiczinequality on the set U ⊆ [g > 0] ∩ [θ′(E(·|ϕ)) > 0] if there is a strictlyincreasing function θ ∈W 1,1

loc (R) satisfying θ(0) = 0 and

θ′(E(v |ϕ)) g(v) ≥ 1 for all v ∈ U . (5)

In the case that there are α ∈ (0, 1] and c > 0 such that

θ(s) = cα |s|

α−1s for every s ∈ R,

Definition

A proper functional E : M→ (−∞,+∞] with strong upper gradient gand equilibrium point ϕ ∈ Eg is said to satisfy a Lojasiewicz-Simoninequality with exponent α ∈ (0, 1] near ϕ if there are c > 0 and a setU ⊆ D(E) with ϕ ∈ U such that

|E(v |ϕ)|1−α ≤ c g(v) for every v ∈ U . (6)



Definition



θ′(E(v |ϕ)) g(v) ≥ 1 for all v ∈ U . (5)


θ(s) = cα |s|


Definition





Definition



θ′(E(v |ϕ)) g(v) ≥ 1 for all v ∈ U . (5)


θ(s) = cα |s|


Definition




Trend to equilibrium in the metric sense

Theorem (Trend to equilibrium in the metric sense)

Let E : M→ (−∞,+∞] be a proper functional with strong uppergradient g , and v be a p-gradient flow of E with non-empty ω-limit setω(v). Suppose, E is lower semicontinuous on It(v) for some t ≥ 0 andfor ϕ ∈ ω(v) ∩ Eg , there is an ε > 0 such that

B(ϕ, ε) ∩ [E(·|ϕ) > 0] ⊆ [g > 0].

If there is a strictly increasing function θ ∈W 1,1loc (R) satisfying θ(0) = 0

and |[θ > 0, θ′ = 0]| = 0 such that E satisfies the Kurdyka- Lojasiewiczinequality (5) on

Uε = B(ϕ, ε) ∩ [E(·|ϕ) > 0] ∩ [θ′(E(·|ϕ)) > 0], (7)

then v has finite length and

limt→∞

v(t) = ϕ in M. (8)


Decay estimates and finite time of extinction

Theorem (Decay estimates and finite time of extinction)

Let E : M→ (−∞,+∞] be a proper functional with strong uppergradient g , and v be a p-gradient flow of E with non-empty ω-limit setω(v). Suppose, E is lower semicontinuous on It(v) for some t > 0, andfor ϕ ∈ ω(v) ∩ Eg there are ε > 0, c > 0, and α ∈ (0, 1] such that Esatisfies a Lojasiewicz-Simon inequality (6) with exponent α onB(ϕ, ε) ∩ D(E). Then

d(v(t), ϕ) ≤ cα (E(v(t)|ϕ))α = O

(t−α(p−1)

1−pα

)if 0 < α < 1

p

d(v(t), ϕ) ≤ c p (E(v(t)|ϕ))1p ≤ c p (E(v(t0)|ϕ))

1p e− t

pcp′ if α = 1p

d(v(t), ϕ) ≤

c (t − t)

α(p−1)pα−1 if t0 ≤ t ≤ t,

0 if t > t,if 1

p < α ≤ 1,



Theorem

where,

c :=

[[1

αα−1c

] p′−1α pα−1

α(p−1)

]α(p−1)pα−1

,

t := t0 + αα−1α(p−1) c

1α(p−1) α(p−1)

pα−1 (E(v(t0)|ϕ))pα−1α(p−1) ,

and t0 ≥ 0 can be chosen to be the “first entry time”, that is, t0 ≥ 0 isthe smallest time t0 ∈ [0,+∞) such that v([t0,+∞)) ⊆ B(ϕ, ε).

If E : M→ (−∞,∞] be a proper, lower semicontinuous, λ-geodesicallyconvex functional on a length space (M, d), with λ > 0 and is boundefrom below, then there is a unique minimiser ϕ ∈ D(E) of E and

E(v |ϕ) ≤ 1

2λ|D−E|2(v) for all v ∈ D(E).

Then, every gradient flow v of E satisfies

d(v(t), ϕ) = O(e−λt

)as t →∞.



Theorem

where,

c :=

[[1

αα−1c

] p′−1α pα−1

α(p−1)

]α(p−1)pα−1

,

t := t0 + αα−1α(p−1) c

1α(p−1) α(p−1)

pα−1 (E(v(t0)|ϕ))pα−1α(p−1) ,



E(v |ϕ) ≤ 1




)as t →∞.



Theorem

where,

c :=

[[1

αα−1c

] p′−1α pα−1

α(p−1)

]α(p−1)pα−1

,

t := t0 + αα−1α(p−1) c

1α(p−1) α(p−1)

pα−1 (E(v(t0)|ϕ))pα−1α(p−1) ,



E(v |ϕ) ≤ 1




)as t →∞.


Entropy-transportation inequality and K L inequality

Definition

A proper functional E : M→ (−∞,+∞] with strong upper gradient g issaid to satisfy locally a generalised entropy-transportation (ET-)inequality at a point of equilibrium ϕ ∈ Eg if there are ε > 0 and astrictly increasing function Ψ ∈ C (R) satisfying Ψ(0) = 0 and

infϕ∈Eg∩B(ϕ,ε)

d(v , ϕ) ≤ Ψ(E(v |ϕ)) (9)

for every v ∈ B(ϕ, ε) ∩ D(E). Further, a functional E is said to satisfyglobally a generalised entropy-transportation inequality at ϕ ∈ Eg if Esatisfies

infϕ∈Eg

d(v , ϕ) ≤ Ψ(E(v |ϕ)) for every v ∈ D(E). (10)

Assumption (E) Suppose, for the proper energy functionalE : M→ (−∞,+∞] with strong upper gradient g holds:

for all v0 ∈ D(E), there is a p-gradient flow v of E with v(0+) = v0.



Theorem (Global K L- and ET-inequality)

For λ ≥ 0, let E : M→ (−∞,+∞] be a proper, lower semicontinuous,λ-geodesically convex functional on a length space (M, d). Suppose, Eand the descending slope |D−E| satisfying Assumption (E) and forϕ ∈ E|D−E|, the set [E(·|ϕ) 6= 0] ⊂ [|D−E| > 0]. Then, the followingstatements are equivalent.

(1) (K L-inequality) There is a strictly increasing function θ ∈W 1,1loc (R)

satisfying θ(0) = 0 and |[θ 6= 0, θ′ = 0]| = 0, and E satisfies aKurdyka- Lojasiewicz inequality onU := [E(·|ϕ) > 0] ∩ [θ′(E(·|ϕ)) > 0].

(2) (ET-inequality)There is a strictly increasing function Ψ ∈ C (R)satisfying Ψ(0) = 0 and s 7→ Ψ(s)/s belongs to L1

loc(R) such that Esatisfies the generalised entropy-transportation inequality

infϕ∈argmin(E)

d(v , ϕ) ≤ Ψ(E(v |ϕ)) for all v ∈ D(E).



Corollary (Global LS- and ET-inequality)

For λ ≥ 0, let E : M→ (−∞,+∞] be a proper, lower semicontinuous,λ-geodesically convex functional on a length space (M, d). Suppose, Eand the descending slope |D−E| satisfy Assumption (E) and forϕ ∈ E|D−E|, the set [E(·|ϕ) > 0] ⊂ [g > 0]. Then, for α ∈ (0, 1], thefollowing statements hold.

(1) ( LS-inequality implies ET-inequality) If there is a c > 0

(E(v |ϕ))1−α ≤ c |D−E|(v) for all v ∈ D(E) (11)

then E satisfies

infϕ∈argmin(E)

d(v , ϕ) ≤ c

α(E(v |ϕ))α for all v ∈ D(E). (12)

(2) (ET-inequality implies LS-inequality) If there is a c > 0 such thatE satisfies (12), then E satisfies

(E(v |ϕ))1−α ≤ c

α|D−E|(v) for all v ∈ D(E). (13)


Aplications. The classical Hilbert space case

In the case (H, (., .)H) is a real Hilbert space and E : H → (−∞,+∞] aproper, lower semicontinuous and semi-convex functional. Then, thefollowing well-known generation theorem holds

Theorem

For every v0 ∈ D(E), there is a unique strong solution v ofv ′(t) + ∂E(v(t)) 3 0, t ∈ (0,∞),

v(0) = v0.(14)

The sub-differential ∂E of E is given by

∂E =

(v , u) ∈ H × H∣∣∣ lim inf

t↓0E(v+tw)−E(v)

t ≥ (u,w) for all w ∈ H.

For v ∈ D(|D−E|), the descending slope

|D−E|(v) = min‖u‖H

∣∣∣ u ∈ ∂E(v)

and |D−E| is a strong upper gradient of E .




Theorem


v(0) = v0.(14)


∂E =

(v , u) ∈ H × H∣∣∣ lim inf

t↓0E(v+tw)−E(v)




∣∣∣ u ∈ ∂E(v)

and |D−E| is a strong upper gradient of E .




Theorem


v(0) = v0.(14)


∂E =

(v , u) ∈ H × H∣∣∣ lim inf

t↓0E(v+tw)−E(v)




∣∣∣ u ∈ ∂E(v)

and |D−E| is a strong upper gradient of E .J.M. Mazon joint works with Daniel Hauer (Sydney University) K LS-inequality for gradient flows in metric spaces

Applications. The classical Hilbert space case

Corollary

Let E : H → (−∞,+∞] be a proper, lower semicontinuous andsemi-convex functional on a Hilbert space H. Suppose v is solution of(14) and there are c , ε > 0 and an equilibrium point ϕ ∈ ω(v) such thatE satisfies a Lojasiewicz-Simon inequality

|E(v |ϕ)|1−α ≤ c ‖u‖ for every v ∈ B(ϕ, ε) and u ∈ ∂E(v).

Then,

‖v(t)− ϕ‖H ≤ cα (E(v(t)|ϕ))α = O

(t− α

1−2α)

if 0 < α < 12

‖v(t)− ϕ‖H ≤ c 2 (E(v(t)|ϕ))12 ≤ c 2 (E(v(t0)|ϕ))

12 e−

t2c2 if α = 1

2

‖v(t)− ϕ‖H ≤

c (t − t)

α2α−1 if t0 ≤ t ≤ t,

0 if t > t,if 1

2 < α ≤ 1,

where, t0 ≥ 0 can be chosen to be the “first entry time”, that is, t0 ≥ 0is the smallest time t0 ∈ [0,+∞) such that v([t0,+∞)) ⊆ B(ϕ, ε).


Applications. Finite Extinction time of the Dirichlet-TotalVariational Flow

vt = div

(Dv|Dv |

)in Ω× (0,+∞),

v = 0 on ∂Ω× (0,+∞),

v(0) = v0 on Ω,

(15)

Problem (15) can be rewritten as an abstract initial value problem (14)in the Hilbert space H = L2(Ω) for the energy functionalE : L2(Ω)→ (−∞,+∞] given by

E(v) :=

∫

Ω

|Dv |+∫∂Ω

|v | if v ∈ BV (Ω) ∩ L2(Ω),

+∞ if otherwise.(16)

Extinction time

T ∗(v0) := inf

t > 0∣∣∣ v(s) = 0 for all s ≥ t

.



vt = div

(Dv|Dv |

)in Ω× (0,+∞),

v = 0 on ∂Ω× (0,+∞),

v(0) = v0 on Ω,

(15)


E(v) :=

∫

Ω

|Dv |+∫∂Ω

|v | if v ∈ BV (Ω) ∩ L2(Ω),


Extinction time

T ∗(v0) := inf

t > 0∣∣∣ v(s) = 0 for all s ≥ t

.



vt = div

(Dv|Dv |

)in Ω× (0,+∞),

v = 0 on ∂Ω× (0,+∞),

v(0) = v0 on Ω,

(15)


E(v) :=

∫

Ω

|Dv |+∫∂Ω

|v | if v ∈ BV (Ω) ∩ L2(Ω),


Extinction time

T ∗(v0) := inf

t > 0∣∣∣ v(s) = 0 for all s ≥ t

.



Theorem

Suppose N ≤ 2 and for v0 ∈ L2(Ω), let v be the unique strong solution ofproblem (15). Then,

T ∗(v0) ≤

s + S1 |Ω|1/2 E(v(s)) if N = 1,

s + S2 E(v(s)) if N = 2,

for arbitrarily small s > 0, and

‖v(t)‖L2(Ω) ≤

c (T ∗(v0)− t) if 0 ≤ t ≤ T ∗(v0),

0 if t > T ∗(v0),(17)

where SN is the best constant in Sobolev inequality and c > 0.



Skech of proof By Sobolev’s inequality in BV (Ω), we have

‖v‖L1∗ (Ω) ≤ SN

(∫Ω

|Dv |+∫∂Ω

|v |)

for all v ∈ BV (Ω) (18)

‖v‖L2(Ω) ≤ C E(v |ϕ) for all v ∈ D(E), (19)

where the constant C = S1 |Ω|1/2 if N = 1 and C = S2 if N = 1.

(19) is a entropy-transportation inequality for Ψ(s) = C s, (s ∈ R),which by Corollary 15, is equivalent to the Lojasiewicz-Simon inequality

1 ≤ C |D−E|(v), (v ∈ [E > 0]).

In dimension N = 2, the extinction time

T ∗(v0) ≤ 1√2π

∫Ω

|Dv0|.




‖v‖L1∗ (Ω) ≤ SN

(∫Ω

|Dv |+∫∂Ω

|v |)





1 ≤ C |D−E|(v), (v ∈ [E > 0]).


T ∗(v0) ≤ 1√2π

∫Ω

|Dv0|.




‖v‖L1∗ (Ω) ≤ SN

(∫Ω

|Dv |+∫∂Ω

|v |)





1 ≤ C |D−E|(v), (v ∈ [E > 0]).


T ∗(v0) ≤ 1√2π

∫Ω

|Dv0|.




‖v‖L1∗ (Ω) ≤ SN

(∫Ω

|Dv |+∫∂Ω

|v |)





1 ≤ C |D−E|(v), (v ∈ [E > 0]).


T ∗(v0) ≤ 1√2π

∫Ω

|Dv0|.


The total variation on graphs

Let G = (V (G ),E (G )) be a locally finite weighted graph. For each edge(x , y) ∈ E (G ) (we will write x ∼ y if (x , y) ∈ E (G )), it is assigned apositive weight wxy = wyx . We consider that wxy = 0 if (x , y) 6∈ E (G ).

For x ∈ V (G ) we define the weight at the vertex x as

dx :=∑y∼x

wxy =∑

y∈V (G)

wxy ,

For each x ∈ V (G ) we define the following probability measures

mGx =

1

dx

∑y∼x

wxy δy .

νG defined as

νG (A) :=∑x∈A

dx , A ⊂ V (G )

is a invariant and reversible measure for this random walk.





dx :=∑y∼x

wxy =∑

y∈V (G)

wxy ,


mGx =

1

dx

∑y∼x

wxy δy .

νG defined as

νG (A) :=∑x∈A

dx , A ⊂ V (G )






dx :=∑y∼x

wxy =∑

y∈V (G)

wxy ,


mGx =

1

dx

∑y∼x

wxy δy .

νG defined as

νG (A) :=∑x∈A

dx , A ⊂ V (G )






dx :=∑y∼x

wxy =∑

y∈V (G)

wxy ,


mGx =

1

dx

∑y∼x

wxy δy .

νG defined as

νG (A) :=∑x∈A

dx , A ⊂ V (G )




We define its mG -total variation as

TVmG (u) =1

2

∫V (G)

∫V (G)

|u(y)− u(x)|dmGx (y)dνG (x)

=1

2

∑x∈V (G)

∑y∈V (G)

|u(y)− u(x)|wxy .

The functional the functional FmG : L2(V , νG )→]−∞,+∞] defined by

FmG (u) :=

TVmG (u) if u ∈ L2(V , νG ),

+∞ if u ∈ L2(X ) \ BVmG (V ),

is convex and lower semi-continuous.

Definition

We define in L2(V , νG ) the multivalued operator ∆mG

1 by (u, v) ∈ ∆mG

1

if and only if −v ∈ ∂FmG (u).




TVmG (u) =1

2

∫V (G)

∫V (G)


=1

2

∑x∈V (G)

∑y∈V (G)

|u(y)− u(x)|wxy .


FmG (u) :=


+∞ if u ∈ L2(X ) \ BVmG (V ),


Definition


1 by (u, v) ∈ ∆mG

1





TVmG (u) =1

2

∫V (G)

∫V (G)


=1

2

∑x∈V (G)

∑y∈V (G)

|u(y)− u(x)|wxy .


FmG (u) :=


+∞ if u ∈ L2(X ) \ BVmG (V ),


Definition


1 by (u, v) ∈ ∆mG

1




We have existence and uniqueness for the Cauchy problemut −∆mG

1 u 3 0 in (0,T )× V

u(0, x) = u0(x) x ∈ V .(20)

(u, v) ∈ ∆mG

1 ⇐⇒ there exists g ∈ L∞(V (G )×V (G ), νG⊗mGx ) antisymmetric

‖g‖∞ ≤ 1 such that∑

y∈V (G)

g(x , y)1

dxwxy = v(x) ∀ x ∈ V (G ),

and

g(x , y) ∈ sign(u(y)− u(x)) (νG ⊗mGx )− a.e. (x , y) ∈ V × V .



We have existence and uniqueness for the Cauchy problemut −∆mG

1 u 3 0 in (0,T )× V

u(0, x) = u0(x) x ∈ V .(20)

(u, v) ∈ ∆mG

1 ⇐⇒ there exists g ∈ L∞(V (G )×V (G ), νG⊗mGx ) antisymmetric

‖g‖∞ ≤ 1 such that∑

y∈V (G)

g(x , y)1

dxwxy = v(x) ∀ x ∈ V (G ),

and

g(x , y) ∈ sign(u(y)− u(x)) (νG ⊗mGx )− a.e. (x , y) ∈ V × V .



Theorem

Let G = (V (G ),E (G )) be a finite weighted connected graphG = (V ,E ). Then (mG , νG ) satisfies a 2-Poincare inequality, that is

inf

TVmG (u)

‖u‖2: ‖u‖2 6= 0,

∫V

u(x)dνG (x) = 0

> 0.

Consequently,

‖et∆mG

1 u0 − ν(u0)‖L2(V ,νG ) ≤

λ2

(mG ,νG ) (t − t) if 0 ≤ t ≤ t,

0 if t > t,

where t := 1λ2

(mG ,νG )

Fm(u0).


Gradient flows in spaces of probability measures.

Let (X ,B, d) be a Polish space equipped with their Borel σ-algebra B.

For s ∈ 0, 1, let πs : X × X → X be defined byπs(x , y) := (1− s)x + sy . For given measures µ0, µ1 ∈ P(X ), the set oftransport plans with marginals µ0 and µ1 is denoted by

Π(µ0, µ1) :=γ ∈ P(X × X )

∣∣∣ π0#γ = µ0, π1#γ = µ1

.

where πi#γ is the push-forward of γ through πiFor 1 ≤ p < +∞, the p-Wasserstein distance Wp,d(µ1, µ2) between µ0

and µ1 ∈ P(X ) is defined by

Wp,d(µ1, µ2) =

(inf

γ∈Π(µ0,µ1)

∫X×X

d(x , y)p dγ(x , y)

) 1p

.

Fixed x0 ∈ X , the space of finite p-moment

Pp,d(X ) :=

µ ∈ P(X )

∣∣∣ ∫X

d(x0, x)p dµ(x) < +∞,

The pair (Pp,d(X ),Wp,d) is called the p-Wasserstein space.





Π(µ0, µ1) :=γ ∈ P(X × X )

∣∣∣ π0#γ = µ0, π1#γ = µ1

.

where πi#γ is the push-forward of γ through πi

For 1 ≤ p < +∞, the p-Wasserstein distance Wp,d(µ1, µ2) between µ0


Wp,d(µ1, µ2) =

(inf

γ∈Π(µ0,µ1)

∫X×X


) 1p

.


Pp,d(X ) :=

µ ∈ P(X )

∣∣∣ ∫X

d(x0, x)p dµ(x) < +∞,






Π(µ0, µ1) :=γ ∈ P(X × X )

∣∣∣ π0#γ = µ0, π1#γ = µ1

.



Wp,d(µ1, µ2) =

(inf

γ∈Π(µ0,µ1)

∫X×X


) 1p

.


Pp,d(X ) :=

µ ∈ P(X )

∣∣∣ ∫X

d(x0, x)p dµ(x) < +∞,






Π(µ0, µ1) :=γ ∈ P(X × X )

∣∣∣ π0#γ = µ0, π1#γ = µ1

.



Wp,d(µ1, µ2) =

(inf

γ∈Π(µ0,µ1)

∫X×X


) 1p

.


Pp,d(X ) :=

µ ∈ P(X )

∣∣∣ ∫X

d(x0, x)p dµ(x) < +∞,






Π(µ0, µ1) :=γ ∈ P(X × X )

∣∣∣ π0#γ = µ0, π1#γ = µ1

.



Wp,d(µ1, µ2) =

(inf

γ∈Π(µ0,µ1)

∫X×X


) 1p

.


Pp,d(X ) :=

µ ∈ P(X )

∣∣∣ ∫X

d(x0, x)p dµ(x) < +∞,

The pair (Pp,d(X ),Wp,d) is called the p-Wasserstein space.J.M. Mazon joint works with Daniel Hauer (Sydney University) K LS-inequality for gradient flows in metric spaces


Consider the free energy E : Pp(RN)→ (−∞,+∞] composed by

E = HF +HV +HW (21)

of the internal energy

HF (µ) :=

∫RN

F (ρ) dx if µ = ρLN ,

+∞ if µ ∈ Pp(RN) \ Pacp (RN),

the potential energy

HV (µ) :=

∫RN

V dµ if µ = ρLN ,


and the interaction energy

HW (µ) :=

12

∫RN×RN

W (x − y) d(µ⊗ µ)(x , y) if µ = ρLN ,




We assume that the function(F) : F : [0,+∞)→ R is a convex differential function satisfying

F (0) = 0, lim infs↓0

F (s)

sα> −∞ for some α > N/(N + p), (22)

the map s 7→ sNF (s−N) is convex and non increasing in (0,+∞),(23)

there is a CF > 0 such that

F (s + s) ≤ CF (1 + F (s) + F (s)) for all s, s ≥ 0, and (24)

lims→+∞

F (s)

s= +∞ (super-linear growth at infinity);

(25)

(V) : V : RN → (−∞,+∞] is a proper, lower semicontinuous, λ-convexfor some λ ∈ R, and the effective domain D(V ) has a convex,nonempty interior Ω := int D(V ) ⊂ RN .

(W) : W : RN → [0,+∞) is a convex, differentiable, and even functionand there is a CW > 0 such that

W (x + x) ≤ CW (1 + W (x) + W (x)) for all x , x ∈ RN . (26)J.M. Mazon joint works with Daniel Hauer (Sydney University) K LS-inequality for gradient flows in metric spaces


Proposition

Suppose, the functions F , V and W satisfy the hypotheses (F), (V)and (W), and E : Pp(RN)→ (−∞,+∞] is the functional given by (21).Then, for µ = ρLN ∈ D(E), one has µ ∈ D(|D−E|) if and only if

PF (ρ) ∈W 1,1loc (Ω), ρ ξρ = ∇PF (ρ) + ρ∇V + ρ(∇W ) ∗ ρ (27)

for some ξρ ∈ Lp′ (RN ,RN ; dµ), where PF (x) := xF ′(x)− F (x) is theassociated “pressure function” of F . Moreover, the vector field ξρsatisfies

|D−E|(µ) =

(∫RN

|ξρ(x)|p′ dµ

) 1p′

. (28)



For every µ0 ∈ D(E), there is a p-gradient flow µ : [0,+∞)→ Pp(RN)of E with initial value limt↓0 µ(t) = µ0. Moreover, for every t > 0,µ(t) = ρ(t)LN with supp(ρ(t)) ⊆ Ω, and ρ is a distributional solution ofthe following quasilinear parabolic-elliptic boundary-value problem

ρt + div(ρUρ) = 0 in (0,+∞)× Ω,

Uρ = −|ξρ|p′−2ξρ in (0,+∞)× Ω,

Uρ · n = 0 in (0,+∞)× ∂Ω,

(29)

with PF (ρ) ∈ L1loc((0,+∞); W 1,1

loc (Ω)) and

ξρ =∇PF (ρ)

ρ+∇V + (∇W ) ∗ ρ ∈ L∞loc((0,+∞); Lp′ (Ω,RN ; dµ(·))),

where, n in (29) denotes the outward unit normal to the boundary ∂Ωwhich in the case Ω = RN needs to be neglected.If the function F ∈ C 2(0,+∞), then one has that

−Uρ = |F ′′(ρ)∇ρ+∇V +(∇W )∗ρ|p′−2 (F ′′(ρ)∇ρ+∇V + (∇W ) ∗ ρ) .



Problem (29) includes the

doubly nonlinear diffusion equation

ρt − div(|∇ρm|p′−2∇ρm) = 0

(V = W = 0, F (s) = m sq

q(q−1) for q = m + 1− 1p′−1 ,

1p′−1 6= m ≥ N−(p′−1)

N(p′−1) )

Fokker-Planck equation with interaction term through porousmedium

ρt = ∆ρm + div (ρ(∇V + (∇W ) ∗ ρ))

(p = 2, F (s) = sm

(m−1) for 1 6= m ≥ 1− 1N ).



Every equilibrium point ν = ρ∞LN ∈ E|D−E| of E can be characterised byPF (ρ∞) ∈W 1,1

loc (Ω) with

ξρ∞ =∇PF (ρ∞)

ρ∞+∇V + (∇W ) ∗ ρ∞ = 0 a.e. on Ω.

Further, for every p-gradient flow µ of E and equilibrium pointν ∈ E|D−E|, we have

ddt E(µ(t)) = −|D−E|p′ (µ(t)) = −Ip′ (µ(t)|ν),

where the generalised relative Fischer information of µ with respect to νis given by

Ip′ (µ|ν) =

∫Ω

−Uρ · ξρ dµ.



Every equilibrium point ν = ρ∞LN ∈ E|D−E| of E can be characterised byPF (ρ∞) ∈W 1,1

loc (Ω) with

ξρ∞ =∇PF (ρ∞)

ρ∞+∇V + (∇W ) ∗ ρ∞ = 0 a.e. on Ω.

Further, for every p-gradient flow µ of E and equilibrium pointν ∈ E|D−E|, we have

ddt E(µ(t)) = −|D−E|p′ (µ(t)) = −Ip′ (µ(t)|ν),

where the generalised relative Fischer information of µ with respect to νis given by

Ip′ (µ|ν) =

∫Ω

−Uρ · ξρ dµ.



Definition

We call a functional f : RN → (−∞,+∞] uniformly λ-p-convex for someλ ∈ R if the interior Ω = int(D(f )) of f is nonempty, f is differentiableon Ω and for every x ∈ Ω,

f (x)− f (x) ≥ ∇f (x) · (y − x) + λ |y − x |p for all y ∈ RN .

(V∗) V : RN → (−∞,+∞] is proper, lower semicontinuous function, theeffective domain D(V ) of V has nonempty interior Ω := int D(V ) ⊆ RN ,and V is uniformly λV -p-convex for some λV ∈ R;



Definition

We call a functional f : RN → (−∞,+∞] uniformly λ-p-convex for someλ ∈ R if the interior Ω = int(D(f )) of f is nonempty, f is differentiableon Ω and for every x ∈ Ω,

f (x)− f (x) ≥ ∇f (x) · (y − x) + λ |y − x |p for all y ∈ RN .

(V∗) V : RN → (−∞,+∞] is proper, lower semicontinuous function, theeffective domain D(V ) of V has nonempty interior Ω := int D(V ) ⊆ RN ,and V is uniformly λV -p-convex for some λV ∈ R;



Theorem

Suppose that the functions F , V and W satisfy the hypotheses (F),(V∗) with λV ∈ R and (W). Further, suppose F ∈ C 2(0,∞) ∩ C [0,+∞)and let E : Pp(RN)→ (−∞,+∞] be the functional given by (21). Then,the following statements hold.

(ET-inequality) For an equilibrium point ν = ρ∞LN ∈ E|D−E| of Ewith ρ∞ ∈W 1,∞(Ω) and inf ρ∞ > 0, and every µ = ρLN ∈ D(E),

λV W pp (µ, ν) ≤ E(µ|ν). (30)

(p-Talagrand transportation inequality) If λV > 0, thenentropy-transportation inequality (30) is equivalent to thep-Talagrand inequality

Wp(µ, ν) ≤ 1

λ1/pV

p√E(µ|ν) (31)

holding for an equilibrium point ν = ρ∞LN ∈ E|D−E| of E with

ρ∞ ∈W 1,∞(Ω) and all µ = ρLN ∈ D(E).



Theorem

(generalised LS-inequality) If λ > 0 then for every probabilitymeasures µ1 = ρ1LN , µ2 = ρ2LN ∈ Pac

p (Ω) with ρ2 ∈W 1,∞(Ω) andinf ρ2 > 0, one has that

E(µ2|µ1) + (λV − λ) W pp (µ1, µ2) ≤ p − 1

pp′

1

λ1/(p−1)|D−E|p′ (µ2).

(generalised Log-Sobolev inequality) If λV > 0, then for everyprobability measures µ1 = ρ1LN , µ2 = ρ2LN ∈ Pac

p (Ω) with

ρ2 ∈W 1,∞(Ω) and inf ρ2 > 0 and ν ∈ E|D−E|, one has that

E(µ2|µ1) ≤ p − 1

pp′

1

λ1/(p−1)V

Ip′ (µ2|ν).

(p-HWI inequality) For every probability measures µ1 = ρ1LN ,µ2 = ρ2LN ∈ Pac

p (Ω) with ρ2 ∈W 1,∞(Ω) and inf ρ2 > 0, one has

E(µ2|µ1) + λV W pp (µ1, µ2) ≤ I1/p′

p′ (µ2|ν) Wp(µ1, µ2).



Corollary (Equivalence between global ET-, LS- and Log-Sobolevinequality)

Suppose that the functions F , V and W satisfy the hypotheses (F), (V)and (W). Further, suppose F ∈ C 2(0,∞) ∩ C [0,+∞) and letE : Pp(RN)→ (−∞,+∞] be the functional given by (21). Then, thefollowing statements hold.

(1) If for ν = ρ∞LN ∈ E|D−E|, there is some λ > 0 such that E satisfiesentropy transportation inequality

Wp(µ, ν) ≤ λ (E(µ|ν))1p for all µ ∈ D(E), (32)

then E satisfies the Lojasiewicz-Simon inequality

E(µ|µ∞)1− 1p ≤ λ |D−E|(µ) for all µ ∈ D(|D−E|), (33)

or equivalently, E satisfies the Log-Sobolev inequality

E(µ|µ∞)1− 1p ≤ λ

1

1− 1p Ip′ (µ|ν) for all µ ∈ D(|D−E|), (34)



(2) If for ν = ρ∞LN ∈ E|D−E|, there is some λ > 0 such that E satisfiesLog-Sobolev inequality (34), then E satisfies entropy transportationinequality

Wp(µ, ν) ≤ λp (E(µ|ν))1p for all µ ∈ D(E),

Corollary (Trend to equilibrium and exponential decay rates)

Suppose that the functions F , V and W satisfy the hypotheses (F),(V∗) with λV > 0 and (W). Further, suppose F ∈ C 2(0,∞) ∩ C [0,+∞)and let E : Pp(RN)→ (−∞,+∞] be the functional given byE = HF +HV +HW . Then, there is a unique minimiserν = ρ∞LN ∈ E|D−E| of E and for every initial value µ0 ∈ D(E), thep-gradient flow µ of E trends to ν in Pp(Ω) as t → +∞ and for all t ≥ 0,

Wp(µ(t), ν) ≤ (p−1)1/p′

λ1/pV

(E(µ(t)|ν))1p ≤ (p−1)1/p′

λ1/pV

(E(µ0|ν))1p e−

tp1p

p−1λ1

p−1V .



(2) If for ν = ρ∞LN ∈ E|D−E|, there is some λ > 0 such that E satisfiesLog-Sobolev inequality (34), then E satisfies entropy transportationinequality

Wp(µ, ν) ≤ λp (E(µ|ν))1p for all µ ∈ D(E),

Corollary (Trend to equilibrium and exponential decay rates)

Suppose that the functions F , V and W satisfy the hypotheses (F),(V∗) with λV > 0 and (W). Further, suppose F ∈ C 2(0,∞) ∩ C [0,+∞)and let E : Pp(RN)→ (−∞,+∞] be the functional given byE = HF +HV +HW . Then, there is a unique minimiserν = ρ∞LN ∈ E|D−E| of E and for every initial value µ0 ∈ D(E), thep-gradient flow µ of E trends to ν in Pp(Ω) as t → +∞ and for all t ≥ 0,

Wp(µ(t), ν) ≤ (p−1)1/p′

λ1/pV

(E(µ(t)|ν))1p ≤ (p−1)1/p′

λ1/pV

(E(µ0|ν))1p e−

tp1p

p−1λ1

p−1V .


REFERENCES

arXiv:1707.03129

Kurdyka-ojasiewicz-Simon inequality for gradient flows in metric spacesDaniel Hauer, Jose M. Mazon


REFERENCES

arXiv:1707.03129

Kurdyka-ojasiewicz-Simon inequality for gradient flows in metric spacesDaniel Hauer, Jose M. Mazon



Let (X , d , ν) be a compact measure length space.

Let DC∞ be the set of all continuous convex functions U : [0,+∞)→ Rsatisfying U(0) = 0 and such that the function

λ 7→ ψ(λ) = eλ U(eλ) is convex on (−∞,+∞), U ′(∞) := limr→+∞

U(r)

r

Important prototype functions U ∈ DC∞ are given by

U∞(r) :=

r log r if r ∈ (0,+∞),

0 if r = 0,(35)

Um(r) =rm

m(m − 1)for every m > 1. (36)

For U ∈ DC∞, let the corresponding functional U : P(X )→ (−∞,+∞]be given by

U(µ) =

∫X

U(ρ) dν + U ′(∞)µs(X ) (37)

for every µ ∈ P(X ) with Lebesgue decomposition µ = ρ ν + µs withrespect to ν.






U(r)

r


U∞(r) :=

r log r if r ∈ (0,+∞),

0 if r = 0,(35)

Um(r) =rm

m(m − 1)for every m > 1. (36)


U(µ) =

∫X

U(ρ) dν + U ′(∞)µs(X ) (37)

for every µ ∈ P(X ) with Lebesgue decomposition µ = ρ ν + µs withrespect to ν.






U(r)

r


U∞(r) :=

r log r if r ∈ (0,+∞),

0 if r = 0,(35)

Um(r) =rm

m(m − 1)for every m > 1. (36)


U(µ) =

∫X

U(ρ) dν + U ′(∞)µs(X ) (37)

for every µ ∈ P(X ) with Lebesgue decomposition µ = ρ ν + µs withrespect to ν.J.M. Mazon joint works with Daniel Hauer (Sydney University) K LS-inequality for gradient flows in metric spaces


Proposition (Existence of p-gradient flows in Pp,d(X ))

For every U ∈ DC∞, let U : Pp,d(X )→ (−∞,+∞] be the energyfunctional given by (37) with effective domainD(U) = µ ∈ Pp,d(X ) | U(µ) <∞.If U is λ-geodesically convex for some λ ∈ R, then for every µ0 ∈ D(U),there is a p-gradient flow µ : [0,+∞)→ Pp,d(X ) of U with initial valueµ(0+) = µ0.

The property that U is λ-geodesically convex on P2,d(X ) is provided bythe notion of Ricci-curvature of metric length spaces (X , d , ν) in thesense of Lott-Villani-Sturm

Let p : DC∞ → [0,+∞) be defined by

p(r) :=

rU ′+(r)− U(r) if r > 0,

0 if r = 0,

for every U ∈ DC∞, where U ′+(r) stand for the right derivative of U.







p(r) :=

rU ′+(r)− U(r) if r > 0,

0 if r = 0,








p(r) :=

rU ′+(r)− U(r) if r > 0,

0 if r = 0,




Further, for K ∈ R, let λK : DC∞ → [−∞,+∞) be given by

λK (U) := infr>0

K · p(r)

r∈ [−∞,+∞).

for every U ∈ DC∞.

Then for K ∈ R, a compact measure length space (X , d , ν) is said tohave (p,∞)-Ricci curvature bounded from below by K and we write(X , d , µ) is CDp(K ,∞) if for every two probability measures µ0,µ1 ∈ Pp,d(X ), there is a constant speed geodesic µ : [0, 1]→ Pp,d(X )with endpoints µ(0) = µ0 and µ(1) = µ1 such that for every U ∈ DC∞,the associated functional U given by (37) satisfies

U(µt) ≤ (1− t)U(µ0) + t U(µ1)− 1

2λK (U)t(1− t)Wp(µ0, µ1)2

for every t ∈ [0, 1].



Further, for K ∈ R, let λK : DC∞ → [−∞,+∞) be given by

λK (U) := infr>0

K · p(r)

r∈ [−∞,+∞).

for every U ∈ DC∞.

Then for K ∈ R, a compact measure length space (X , d , ν) is said tohave (p,∞)-Ricci curvature bounded from below by K and we write(X , d , µ) is CDp(K ,∞) if for every two probability measures µ0,µ1 ∈ Pp,d(X ), there is a constant speed geodesic µ : [0, 1]→ Pp,d(X )with endpoints µ(0) = µ0 and µ(1) = µ1 such that for every U ∈ DC∞,the associated functional U given by (37) satisfies

U(µt) ≤ (1− t)U(µ0) + t U(µ1)− 1

2λK (U)t(1− t)Wp(µ0, µ1)2

for every t ∈ [0, 1].



Theorem

If for K ∈ R, the compact measure length space (X , d , ν) is CDp(K ,∞)then for every U ∈ DC∞ with λ = λK (U) ∈ R, the energy functionalU : Pp,d(X )→ (−∞,+∞] given by (37) satisfies ET-inequality

λ2 W 2

p,d(µ, ν) ≤ U(µ|ν) for every µ ∈ Pp,d(X ), (38)

and p-HWI inequality

U(µ|ν) ≤Wp,d(µ, ν) p′√IUp′ (µ|ν)− λ

2 W 2p,d(µ, ν) (39)

for every µ = ρν ∈ D(U) ∩ Pacp,d(X ) with positive ρ ∈ Lip(X ). Moreover,

for U the Log-Sobolev inequality

U(µ|ν) ≤ 12λ

(IUp′ (µ|ν)

)2/p′(40)

holds for every µ = ρν ∈ D(U) ∩ Pacp,d(X ) with positive ρ ∈ Lip(X ), if

λ > 0.



For every µ = ρν ∈ D(U) ∩ Pacp,d(X ) with positive ρ ∈ Lip(X ),

IUp′ (µ|ν) :=

∫X

ρ (U ′′(ρ))p′ |D−ρ|p′ dν =

∫X

U ′′(ρ)p′ |D−ρ|p′ dµ (41)

denote the generalised relative Fischer information of µ = ρ ν ∈ Pacp,d(X )

with respect to ν and |D−ρ|(x) the descending slope of density functionρ with respect to the metric Wp,d .

Since for the integrand U∞ given by (35), one has that

λK (U∞) = K for every K ∈ R,

the Boltzmann H-functional (relative to ν)

H(·|ν) := U∞(·|ν) =

∫X

ρ log ρdν for every µ = ρν ∈ Pacp,d(X ),

(42)is K -geodesically convex on Pp,d(X ) provided the compact length space(X , d , ν) is CDp(K ,∞).



For every µ = ρν ∈ D(U) ∩ Pacp,d(X ) with positive ρ ∈ Lip(X ),

IUp′ (µ|ν) :=

∫X

ρ (U ′′(ρ))p′ |D−ρ|p′ dν =

∫X

U ′′(ρ)p′ |D−ρ|p′ dµ (41)

denote the generalised relative Fischer information of µ = ρ ν ∈ Pacp,d(X )

with respect to ν and |D−ρ|(x) the descending slope of density functionρ with respect to the metric Wp,d .

Since for the integrand U∞ given by (35), one has that

λK (U∞) = K for every K ∈ R,

the Boltzmann H-functional (relative to ν)

H(·|ν) := U∞(·|ν) =

∫X

ρ log ρdν for every µ = ρν ∈ Pacp,d(X ),

(42)is K -geodesically convex on Pp,d(X ) provided the compact length space(X , d , ν) is CDp(K ,∞).


Corollary

If for K ∈ R, the compact measure length space (X , d , ν) is CDp(K ,∞), thenthe following statements hold.

(1) The energy H satisfies the entropy transportation inequality

K2W 2

p,d(µ, ν) ≤ H(µ|ν) for every µ ∈ Pp,d(X ). (43)

(2) The p-HWI-inequality

H(µ|ν) ≤Wp,d(µ, ν) p′√IHp′ (µ|ν)− K

2W 2

p,d(µ, ν)

holds for every µ = ρν ∈ D(H)∩Pacp,d(X ) with positive ρ ∈ Lip(X ), where

the relative Fisher information associated with H is given by

IHp′ (µ|ν) =

∫ρ>0

ρ

ρp′|D−ρ|p′ dν µ = ρν ∈ Pac

p,d(X )

(3) If K > 0, the p-logarithmic Sobolev inequality

H(µ|ν) ≤ 12K

(IHp′ (µ|ν)

)2/p′

holds for every µ = ρν ∈ D(H) ∩ Pacp,d(X ) with positive ρ ∈ Lip(X ).



Corollary (Decay rates and finite time of extinction)

Suppose for K > 0, the compact measure length space (X , d , ν) isCDp(K ,∞) and U ∈ DC∞ with λK (U) > 0. Then, for every µ0 ∈ D(U),the p-gradient flow µ : [0,+∞)→ Pp,d(X ) of U with initial valueµ(0+) = µ0 trends to the unique equilibrium µ∞ = U(1) ν of U and forevery t ≥ 0,

Wp,d(µ(t), µ∞) ≤

1√K

√U(µ(t)|µ∞) = O(t−

p−12−p ) if 1 < p < 2,

2√K

√U(µ(t)|µ∞) ≤ 2√

K

√U(µ(0)|ν) e−t K if p = 2,

c (t − t)p−1p−2 for 0 ≤ t ≤ t if 2 < p,

where

c :=[K 2(p′−1) p−2

p−1

] p−1p−2

and t = 1

21−2p−1

(1

2K

) 1p−1 p−1

p−2 (H(µ0|ν))p−2p−1 .



corolario tapao



Definition

Let (X , d) be a metric space. Then, for U ∈ DC∞ and K > 0, aprobability measure ν on X is said to satisfy a (generalised) p-logarithmicSobolev inequality with constant K > 0 and write ν satisfies LSI pU(K ) iffor the relative entropy U(·|ν) the p-logarithmic Sobolev inequality

U(µ|ν) ≤ 12K

(IUp′ (µ|ν)

)2/p′

holds for every µ = ρν ∈ D(U) ∩ Pacp,d(X ) with positive ρ ∈ Lip(X ).

Further, a probability measure ν is said to satisfy a (generalised)Talagrand entropy transportation inequality in Pp,d(X ) with constant Kand write ν satisfies T p

U(K ) if the relative entropy U(·|ν) satisfiesentropy-transportation inequality

K2 W 2

p,d(µ, ν) ≤ U(µ|ν) for every µ ∈ Pp,d(X ),

If p = 2 and U = U∞, we write LSI (K ) = LSI 2U∞

(K ) and

T (K ) = T 2U∞

(K ).



Definition

Let (X , d) be a metric space. Then, for U ∈ DC∞ and K > 0, aprobability measure ν on X is said to satisfy a (generalised) p-logarithmicSobolev inequality with constant K > 0 and write ν satisfies LSI pU(K ) iffor the relative entropy U(·|ν) the p-logarithmic Sobolev inequality

U(µ|ν) ≤ 12K

(IUp′ (µ|ν)

)2/p′

holds for every µ = ρν ∈ D(U) ∩ Pacp,d(X ) with positive ρ ∈ Lip(X ).

Further, a probability measure ν is said to satisfy a (generalised)Talagrand entropy transportation inequality in Pp,d(X ) with constant Kand write ν satisfies T p

U(K ) if the relative entropy U(·|ν) satisfiesentropy-transportation inequality

K2 W 2

p,d(µ, ν) ≤ U(µ|ν) for every µ ∈ Pp,d(X ),

If p = 2 and U = U∞, we write LSI (K ) = LSI 2U∞

(K ) and

T (K ) = T 2U∞

(K ).



Theorem (Equivalence between LSI pU(K ) and T pU(K ))

Suppose, the measure length space (X , d , ν) is CDp(K ,∞) for K ≥ 0and for U ∈ DC∞, there is a set D ⊆ D(U) ∩ Pac

p,d(X ) with the propertyfor every µ0 = ρ0 ν ∈ D and every p-gradient flow µ

of U with initial value µ(0+) = µ0 satisfies

µ(t) ∈ D for every t ≥ 0.

(44)

such that the descending slope |D−U| and the generalised Fisherinformation IUp′ satisfy the “equality”

|D−U|p′ (µ) = IUp′(µ|ν) for every µ ∈ D. (45)

Then the following hold.

(1) If ν satisfies LSI pU(K ) for some K > 0, then ν satisfies TPU (K ).

(2) If ν satisfies T pU(K ) for some K > 0, then ν LSI pU(K/8).



Corollary

Suppose the measure length space (X , d , ν) is CD(K ,∞) for a K ≥ 0.Then the following hold.

(i) If ν satisfies LSI (K ) for some K > 0, then ν satisfies T (K ).

(ii) If ν satisfies T (K ) for some K > 0, then ν LSI (K/8).


Documents

Kurdyka- ojasiewicz-Simon inequality for gradient flows in ...este.uca.ma/wedpta18/mazon.pdf · L.Ambrosio, N.Gigli, and G. Savar e.Gradient ows in metric spaces and in the space