Alexandrov meets Lott–Villani–Sturm

Anton Petrunin

arXiv: 1003.5948v3 · PDF · 12 September 2019
Categories: math.MG, math.DG, math.OC
Journal: Muenster Journal of Mathematics 4 (2011), 53--64
Licence: CC0 1.0 The site shows the text.

Abstract

Here I show compatibility of two definition of generalized curvature bounds — the lower bound for sectional curvature in the sense of Alexandrov and lower bound for Ricci curvature in the sense of Lott–Villani–Sturm.

Items 11

Main theorem

0SSK

\ifthenelse{\equal{0}{}}{\text{\rm Alex}^{m}}{\text{\rm Alex}^{m}[0]}\subset \ifthenelse{\equal{0}{}}{\text{\rm CD}}{\text{\rm CD}[m,0]}.

Remark

I1SW

Although \ifthenelse{\equal{\upkappa}{}}{\text{\rm CD}}{\text{\rm CD}[m,\upkappa]} is a very natural class of metric-measure spaces, some basic tools in Ricci comparison can not work there in principle. For instance, there are \ifthenelse{\equal{0}{}}{\text{\rm CD}}{\text{\rm CD}[m,0]}-spaces which do not satisfy the Abresch–Gromoll inequality, (see [AG]). Thus, one has to modify the definition of the class \ifthenelse{\equal{\upkappa}{}}{\text{\rm CD}}{\text{\rm CD}[m,\upkappa]} to make it suitable for substantial applications in Riemannian geometry.

Claim

Claim 1.1

5J9Q

Let A\in\ifthenelse{\equal{\upkappa}{}}{\text{\rm Alex}^{m}}{\text{\rm Alex}^{m}[\upkappa]} and f\:A\to \mathbb{R} be \uplambda-concave Lipschitz function. Then Laplacian \Delta f is defined and

\Delta f\leqslant m{\cdot}\uplambda{\cdot}\mathop{\rm vol}\nolimits.

In particular, \Delta^s f — the singular part \Delta f is nonpositive.

Moreover,

\Delta f =\text{\rm Trace}[\operatorname{Hess} f]{\cdot}\mathop{\rm vol}\nolimits+\Delta^s f.

Proof of Claim 1.1

Let us denote by F_t\:A\to A the f-gradient flow for time t.

Given a Lipschitz function u\:A\to\mathbb{R}, consider family u_t(x)=u\circ F_t(x). Clearly, u_0\equiv u and u_t is Lipschitz for any t\geqslant0. Further, for any x\in A we have \bigl|{\tfrac{d^+}{d t}}u_t(x)|_{t=0}\bigr|\leqslant\mathrm{Const}. Moreover

\tfrac{d^+}{d t}u_t(x)|_{t=0}\buildrel \hbox{\tiny\it a.e.} \over{=\joinrel=} d_x u(\nabla_x f)\buildrel \hbox{\tiny\it a.e.} \over{=\joinrel=}\<\nabla_x u,\nabla_x f\>.

Further,

\int\limits_A u_t{\cdot}\mathop{\rm vol}\nolimits=\int\limits_A u{\cdot}(F_t\#\mathop{\rm vol}\nolimits),

where \# stands for push-forward. Since |F_t(x)-F_t(y)|\leqslant e^{\uplambda t}{\cdot}|x-y| (see [petrunin:survey, 2.1.4(i)]), for any x,y\in A we have

F_t\#\mathop{\rm vol}\nolimits\geqslant \exp(-m{\cdot}\uplambda{\cdot} t){\cdot}\mathop{\rm vol}\nolimits.

Therefore, for any non-negative Lipschitz function u\:A\to\mathbb{R},

\int\limits_A u_t{\cdot}\mathop{\rm vol}\nolimits = \int\limits_A u{\cdot} (F_t\#\mathop{\rm vol}\nolimits) \geqslant \exp(-m{\cdot}\uplambda{\cdot} t)\cdot\int\limits_A u{\cdot}\mathop{\rm vol}\nolimits.

Therefore

\int\limits_A\<\nabla u,\nabla f\>{\cdot}\mathop{\rm vol}\nolimits = \!\left.\tfrac{d^+}{d t}\int\limits_A u_t{\cdot}\mathop{\rm vol}\nolimits\right|_{t=0} \geqslant -m{\cdot}\uplambda{\cdot}\int\limits_A u{\cdot}\mathop{\rm vol}\nolimits.

That is, there is a Radon measure \upchi on A, such that

\int\limits_A u{\cdot}\upchi = \int\limits_A\!\left[\<\nabla u,\nabla f\>+m{\cdot}\uplambda{\cdot} u\right]{\cdot}\mathop{\rm vol}\nolimits.

Set \Delta f=-\upchi+m{\cdot}\uplambda, it is a Radon sign-measure and \upchi=-\Delta f+m{\cdot}\uplambda\geqslant 0.

To prove the second part of theorem, assume u is a non-negative Lipschitz function with support in a DC_0-chart U\to A, where U\subset \mathbb{R}^m is an open subset. Then

\begin{aligned} \int\limits_A\<\nabla u,\nabla f\> &= \int\limits_U \det g\cdot g^{i j}\cdot\partial_i u\cdot\partial_j f \cdot d x^1{\cdot}d x^2\cdots d x^m = \\ &=-\int\limits_U u\cdot\partial_i(\det g\cdot g^{i j}\cdot\partial_j f) \cdot d x^1{\cdot}d x^2\cdots d x^m, \end{aligned}

Thus

\Delta f = \partial_i(\det g\cdot g^{i j}\cdot\partial_j f)\cdot d x^1{\cdot}d x^2\cdots d x^m \buildrel \hbox{\tiny\it a.e.} \over{=\joinrel=} \text{\rm Trace}[\operatorname{Hess} f].

-6mm

Proposition-definition

Proposition-definition 1.2

PF86

Let A\in\ifthenelse{\equal{\upkappa}{}}{\text{\rm Alex}^{m}}{\text{\rm Alex}^{m}[\upkappa]}, \mathbb{I} be an open real interval, \uplambda\:\mathbb{I}\to\mathbb{R} be a continuous function and f_t\:A\to \mathbb{R}, t\in \mathbb{I} be \uplambda(t)-concave family.

Then for any x\in A and t_0\in \mathbb{I} there exists an f_t-gradient curve \upalpha which is defined in a neighborhood of t_0 and such that \upalpha(t_0)=x.

Moreover, if \upalpha,\upbeta\:\mathbb{I}\to A are f_t-gradient then for any t_0,t_1\in\mathbb{I}, t_0\leqslant t_1,

|\upalpha(t_1)-\upbeta(t_1)|\leqslant L{\cdot}|\upalpha(t_0)-\upbeta(t_0)|,

where L=\exp\!\left(\int_{t_0}^{t_1}\uplambda(t){\cdot}\,d t\right).

Used by Proof

Claim

Claim 1.3

R9GG

Let f_t\:A\to\mathbb{R} be a \uplambda(t)-concave family and F_{{t_0},{t_1}} be f_t-gradient flow. Let E\subset A be a bounded Borel set. Fix t_1 and consider the function v(t)=\mathop{\rm vol}\nolimits F^{-1}_{{t},{t_1}}(E). Then

\bigl.v\bigr|_t^{t_1} = \int\limits_t^{t_1}\Delta f_\text{\it \c t}\!\left[F^{-1}_{\text{\it \c t},{t_1}}(E)\right]{\cdot}\,d\text{\it \c t}.

Proof of Claim 1.3

Let u\:A\to\mathbb{R} be a Lipschitz function with compact support. Set u_t=u\circ F_{{t},{t_1}}. Note that the function (x,t)\mapsto u_t(x) is locally Lipschitz. Thus, the function

w_u\:t\mapsto \int\limits_A u_t{\cdot}\mathop{\rm vol}\nolimits

is locally Lipschitz. Further

w_u'(t)\buildrel \hbox{\tiny\it a.e.} \over{=\joinrel=} -\int\limits_A \<\nabla u_t,\nabla f_t\>{\cdot}\mathop{\rm vol}\nolimits = \int\limits_A u_t{\cdot}\Delta f_t.

Therefore

\bigl.w_u\bigr|_t^{t_1} = \int\limits_t^{t_1}d\text{\it \c t}\cdot\int\limits_A u_{\text{\it \c t}}{\cdot}\Delta f_{\text{\it \c t}}.

The last formula extends to an arbitrary Borel function u\:A\to\mathbb{R} with bounded support. Applying it to the characteristic function of E we get the result.

Claim

Claim 2.1

57VH

Let A\in\ifthenelse{\equal{0}{}}{\text{\rm Alex}^{m}}{\text{\rm Alex}^{m}[0]} and f_0\:A\to\mathbb{R} be function; suppose f_t=\mathop{\mathcal H\lefteqn{\phantom{I}}}\nolimits_t f_0 is defined for t\in(0,1). Assume \upgamma\:[0,1]\to A is a geodesic path which is an f_t-gradient curve for t\in(0,1) and \upalpha\:(0,1)\to A is another f_t-gradient curve. If for some t_0\in(0,1), \upalpha(t_0)=\upgamma(t_0), then \upalpha(t)=\upgamma(t) for all t\in(0,1).

Proof of Claim 2.1

Set \ell=\ell(t)=|\upalpha(t)-\upgamma(t)|; it is locally Lipschitz function defined on (0,1). We have to show that if \ell(t_0)=0 for some t_0 then \ell(t)=0 for all t.

According to 1.2, \ell(t)=0 for all t\geqslant t_0. In order to prove that \ell(t)=0 for all t\leqslant t_0, it is sufficient to show that

\ell'\geqslant- [\tfrac 1 t+\tfrac{2}{1-t}]{\cdot}\ell

for almost all t.

Since \upalpha is locally Lipschitz, for almost all t, \upalpha^+(t) and \upalpha^-(t) are defined and opposite (That is, |\upalpha^+(t)|=|\upalpha^-(t)| and \measuredangle(\upalpha^+(t),\upalpha^-(t))=\uppi) to each other.

Fix such t and set x=\upgamma(0), z=\upgamma(t), y=\upgamma(1), p=\upalpha(t), so \ell(t)=|p-z|. Note that function

f_t+\tfrac1{2\cdot(1-t)}{\cdot}\mathop{\rm dist}\nolimits^2_y \refstepcounter{equation}\eqno\text{\ifcase \@nameuse{c@equation}\or \ding{202}\or \ding{203}\or \ding{204}\or \ding{205}\or \ding{206}\or \ding{207}\or \ding{208}\or \ding{209}\or \ding{210}\or \ding{211}\else\@ctrerr \fi}

has a minimum at z. Extend a geodesic [z p] by a both-sides infinite unit-speed quasigeodesic (A careful proof of existence of quasigeodesics can be found in [petrunin:survey].) \upsigma\:\mathbb{R}\to A, so \upsigma(0)=z and \upsigma^+(0)={\uparrow}_{\![{z}{p}]}. The function f_t\circ\upsigma\:\mathbb{R}\to\mathbb{R} is \tfrac1t-concave and from ?,

f_t\circ\upsigma(s)\geqslant f_t(z)+\<\upgamma^+(t),{\uparrow}_{\![{z}{p}]}\>{\cdot} s-\tfrac1{2\cdot(1-t)}{\cdot} s^2.

It follows that

\begin{aligned} \<\upalpha^+(t),\upsigma^+(\ell)\> &\geqslant d_pf_t(\upsigma^+(\ell))=\\ &= (f_t\circ\upsigma)^+(\ell)\geqslant\\ &\geqslant \<\upgamma^+(t),{\uparrow}_{\![{z}{p}]}\> - [\tfrac1t+\tfrac2{1-t}]{\cdot}\ell. \end{aligned}

Now,

Vectors \upsigma^\pm(\ell) are polar; therefore \<\upalpha^\pm(t),\upsigma^+(\ell)\>+\<\upalpha^\pm(t),\upsigma^-(\ell)\>\geqslant 0.
Vectors \upalpha^\pm(t) are opposite; therefore \<\upalpha^+(t),\upsigma^\pm(\ell)\>+\<\upalpha^-(t),\upsigma^\pm(\ell)\>= 0.

Since \upsigma^-(\ell)={\uparrow}_{\![{p}{z}]}, we get

\<\upalpha^+(t),\upsigma^+(\ell)\>+\<\upalpha^+(t),{\uparrow}_{\![{p}{z}]}\>= 0.

Therefore

\ell' = -\<\upalpha^+(t),{\uparrow}_{\![{p}{z}]}\>-\<\upgamma^+(t),{\uparrow}_{\![{z}{p}]}\> \geqslant -[\tfrac1t+\tfrac2{1-t}]{\cdot}\ell.

-6mm

Uses Proposition-definition 1.2

Proposition

Proposition 2.2

11ND

Let A\in\ifthenelse{\equal{0}{}}{\text{\rm Alex}^{m}}{\text{\rm Alex}^{m}[0]}, f\:A\to\mathbb{R} be a bounded continuous function and let f_t=\mathop{\mathcal H\lefteqn{\phantom{I}}}\nolimits_t f. Assume \upgamma\:(0,a)\to A is an f_t-gradient curve which is also a constant-speed geodesic. Assume that function

h(t) \buildrel \hbox{\tiny\it def} \over{=\joinrel=} \text{\rm Trace}[\operatorname{Hess}_{\upgamma(t)}f_t]

is defined for almost all t\in(0,a). Then

h'\leqslant -\tfrac1m{\cdot} h^2

in the sense of distributions; that is, for any non-negative Lipschitz function u\:(0,a)\to\nobreak\discretionary{} {\hbox{=0pt }}{}\mathbb{R} with compact support

\int\limits_0^a \!\left(\tfrac1m{\cdot} h^2{\cdot}u-h{\cdot} u'\right){\cdot}\,d t\geqslant0.

Proof of Proposition 2.2

Since h is defined a.e., all \text{\rm T}_{\upgamma(t)} for t\in(0,a) are isometric to Euclidean m-space. From ?,

f_{t_1}(x)=\inf_{y\in A}\!\left\{f_{t_0}(y)+\frac{|x-y|^2}{2{\cdot}(t_1-t_0)}\right\}.

Thus, for a parallel transportation \uptau\:\text{\rm T}_{\upgamma(t_0)}\to \text{\rm T}_{\upgamma(t_1)} along \upgamma, we have

[\operatorname{Hess}_{\upgamma(t_1)}f_{t_1}](y,y) \leqslant [\operatorname{Hess}_{\upgamma(t_0)}f_{t_0}](x,x) +\frac{|\uptau (x)-y|^2}{t_1-t_0}

for any x\in\text{\rm T}_{\upgamma(t_0)} and y\in\text{\rm T}_{\upgamma(t_1)}. It remains to take the trace of the last formula with y=c{\cdot}\uptau(x) for appropriate c>0.

Source and licence

The text on this page is from arXiv:1003.5948v3 by Anton Petrunin, under the licence CC0 1.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.

The record of each item is in analysis/objects/1003/1003.5948.jsonl.