You are on page 1of 4

Monotonicity and its analytic and geometric implications

Tobias Holck Coldinga,1 and William P. Minicozzi IIb
a

Massachusetts Institute of Technology, Cambridge, MA 02139; and bJohns Hopkins University, Baltimore, MD 21218

Edited by Robion C. Kirby, University of California, Berkeley, Berkeley, CA, and approved July 6, 2012 (received for review June 1, 2012)

In this expository article, we discuss various monotonicity formulas for parabolic and elliptic operators and explain how the analysis of function spaces and the geometry of the underlining spaces are intertwined. After briefly discussing some of the well-known analytical applications of monotonicity for parabolic operators, we turn to their elliptic counterparts, their geometric meaning, and some geometric consequences.
elliptic equations ∣ parabolic equations ∣ geometric inequalities ∣ Ricci-flat ∣ Einstein manifolds

using the heat equation, and integrating by parts gives 1∕t times F , where F is Z n j∇uj 2 n F ðtÞ ¼ tF 0 ðtÞ − ¼ t − : [2] u 2 2 M (We multiplied F 0 by t so that S and F scale the same way). Following ref. 1, we set W ¼ S þ F and get that on a Ricci-flat manifold or, more generally, on a manifold with Ric ≥ 0, that W 0 ≤ 0*. In fact, a computation shows that W0¼ ðtF Þ 0 t

2  Z .

.

.

.

Hesslog u þ 1 g.

þRicð∇ log u. [3] ¼ −2t . ∇ log uÞ u.

2t .

then the Shannon entropy S0 and Fisher information F 0 are defined as follows: Z log f f . (2.1203856109  Z    Z n 1 j∇f j 2 n F0 − f log f ¼ log þ S0 ≥ 0. www. If f > 0 is a positive function on a Riemannian Z f 2 log f e − jxj 2 2 ≤ [4] Z F0 ¼ j∇f j 2 : f M (The point being that the gain in the exponent in the usual Sobolev inequality is dimension dependent. tÞ ¼ ð4πtÞ −2 exp − . The log-Sobolev inequality of ref. M Functional Inequalities and Monotonicity In this section. When Ric ≥ 0 this implies (as tF is obviously 0 for t ¼ 0) that F ≤ 0 and hence S ↓. Each has their own advantages and applications and we will describe some of those. manifold M with ∫ M f ¼ 1. S0 ¼ − M where g is the Riemannian metric and Hesslog u is the Hessian of log u.pnas.H. y. Many other interesting inequalities can be proven by monotonicity. performed research.M.org/cgi/doi/10. the log-Sobolev inequality is equivalent to that for any function f > 0 with ∫ f ¼ 1 n jxj 2 Similarly. we discuss some sharp monotonicity formulas and sharp gradient estimates for the heat and Laplace equations on manifolds and mention briefly some of the applications of the parabolic formulas. in particular. There is a second point and that is that as n → ∞ the measures dx1 ∧⋯∧dxn do not converge to a Radon measure whereas ð2πÞ −2 e − 2 dx1 ∧⋯∧dxn do. whereas the geometric ones imply compactness for the spaces in question and cone structure for those same spaces. We will next briefly see how one can use these quantities and formulas to prove some functional inequalities using monotonicity. 11]. For instance. Our parabolic Fisher information F will also differ from F 0 by a normalization that comes from the Euclidean heat kernel. refs. But our real focus is how these different monotonicity formulas are linked and all have both analytic and geometric implications.) By a change of variables [see. for instance. The next section will focus on the interplay between them and the local geometry and. 2. These formulas can roughly be divided into three groups: elliptic. log f 2nπe 2 2nπe 2 [5] Author contributions: T. and W. Another well-known instance is the log-Sobolev inequality. 2–9. and geometric. S is normalized so that S remains identically zero for all time when u is the heat kernel   j x − yj 2 n H ðx.edu. in the log-Sobolev inequality there is a gain of a log factor independent of the dimension. if u > 0 is a solution to the heat equation ð∂ t − ΔÞ u ¼ 0 with ∫ M u ¼ 1. we single out the Cauchy–Schwarz inequality and the logarithmic Sobolev inequality. This article is a PNAS Direct Submission. 4t on Euclidean space R n . we will touch upon how the parabolic monotonicity formulas imply functional inequalities and the elliptic ones imply uniqueness of blow-ups. PNAS Early Edition ∣ 1 of 4 MATHEMATICS T he aim of this survey is to explain some new and old monotonicity formulas and describe some of their applications. parabolic. *A space-time version of W plays an important role in Perelman’s work on the Ricci flow (25). SPECIAL FEATURE . for instance. see. 1 To whom correspondence should be addressed.1073/pnas. 10 is that for any function f > 0 on R n with ð2πÞ −2 ∫ f 2 e − n jxj 2 2 ¼ 1. Taking the derivative of the entropy S. E-mail: colding@math. As examples. The authors declare no conflict of interest. The Cauchy–Schwarz inequality is perhaps not only the simplest example of an inequality that can be proven from monotonicity using the heat equation.P.1) on page 693 of ref.mit. but the argument itself is also perhaps the easiest.C. Z j∇f j 2 e − 2 : jxj 2 Parabolic Operators. the question of uniqueness of blow-ups or blow-downs of the space. cf. also ref. then we define SðtÞ by n n SðtÞ ¼ S0 ðtÞ − logð4πtÞ − 2 2 Z n n log u u − logð4πtÞ − : ¼− 2 2 M [1] Here.

To show this inequality on R n . let more generally M be a manifold with Ric ≥ 0 and suppose that u solves the heat equation with initial condition uð·. 0Þ ¼ f . 2 2 2 2nπe [6] Δðj∇bj 2 uÞ ¼ . The above inequality will follow from showing that for all t   Fþn n n n 2 log þ S þ logð4πÞ þ ≥ 0.with equality for any Gaussian.

 .

2 .

2 1 .

.

Hessb 2 − Δb g.

∇b 2 Þ b −n : [10] n . þRicð∇b 2 .

2 .

We define a functional A by A ð r Þ ¼ r 1 −n Z b¼r To show this. þ 2tðF þ n n 2Þ [7] j∇bj 3 : [11] A computation gives that the derivative A 0 ðr Þ is equal to − r n−3 2 Z . take the derivative of the left hand side of [6] to get   n 2 2 0 F ð tF Þ ≤ 0. The right-hand side is always nonnegative on a smooth manifold M with Ric ≥ 0 and it vanishes if and only if M is Euclidean space and u is a multiple of the Green’s function.

.

 2 .

2 .

.

Hessb 2 − Δb g.

∇b 2 Þ b 2−2n : . þRicð∇b 2 .

n .

It is easy to see (more or less from the argument given above) that [5] does not hold on a general manifold M with Ric ≥ 0. If it is normalized in this way. A key point in both the Li–Yau gradient estimate and the Harnack inequality is that they are sharp on Euclidean space for the heat kernel. 1. M will be assumed to have Euclidean volume growth and hence will be nonparabolic because n ≥ 3. 2. In the main application later on. 12 or 13. one even has that  r →∞ M \B ðxÞ r lim sup j∇bj ¼ VM VolðB1 ð0ÞÞ  1 n−2 : [14] The function b was already considered in ref. This asserts that if u > 0 is a solution to the heat equation ð∂ t − ΔÞu ¼ 0 on a manifold with nonnegative Ricci curvature. the inequality follows from the formula for W 0 together with the Cauchy–Schwarz inequality. one can prove the sharp gradient estimate j∇bj ≤ 1 [12] Note. Elliptic Operators. This is automatically the case when M is nonparabolic. [8] as long as Ric þ Hessϕ ≥ cg and ∫ M e −ϕ ¼ 1. this quantity on a smooth manifold converges to the volume of the unit sphere in R n and. it converges to  Volð∂ B1 ð0ÞÞ VM VolðB1 ð0ÞÞ  2 n−2 : [13] Here. VM is a geometric quantity that measures the Euclidean volume growth at infinity and will be defined in the next section and B1 ð0Þ ⊂ R n is the unit ball in Euclidean space. Rather. then using [10] and the maximum principle. which is roughly equivalent to that the volume growth is faster than quadratic. Another important and. but by now it is known how to deduce parabolic gradient estimates from monotonicity using the Shannon entropy and Fisher information and the monotonicity of W . closely related result is the Li–Yau gradient estimate. then A ↓. in particular. use the monotonicity together with the asymptotic formula for solutions of the heat equation (solutions on R n converge. 15). Often we will assume that u is normalized so that it has the same asymptotics near the pole p as the Green’s function jxj 2−n on Euclidean space near the origin. Suppose that u > 0 is harmonic in a pointed neighborhood of p ∈ M and define b by on a manifold with Ric ≥ 0. namely if equality holds at one point on the manifold. that if M n has Ric ≥ 0. we have implicitly assumed that b is proper so that the integration is over a compact set. As r tends to 0. Uniqueness is a key question for the regularity of geometric partial differential equations (PDEs). Briefly.15 of refs. we get that b ¼ jxj. This material is from ref. we will throughout the rest of the paper assume that n ≥ 3.org/cgi/doi/10. In the next section. that the Li–Yau gradient estimate also implies that F ≤ 0 when Ric ≥ 0 and hence S ↓. then it is flat Euclidean space and u is the Green’s function. Colding and Minicozzi . In fact. after rescaling. refs. 18 where it was shown that on a manifold with Ric ≥ 0 b ¼ r →∞ ∂ B ðxÞ r r lim sup  1 n−2 VM : VolðB1 ð0ÞÞ [15] b 2−n ¼ u: On Euclidean space.pnas. see. Similarly to the Li–Yau gradient estimate. see. Li–Yau got a sharp Harnack inequality on manifolds with nonnegative Ricci curvature. − t 2 2 u u2 [9] In particular. In fact. it is sharp in a very strong sense. for instance. then Z M f 2 log f e −ϕ ≤ 1 c Z M j∇f j 2 e −ϕ . corollary 5. and 17. as r tends to infinity. if equality holds at one point on the manifold. 12.1073/pnas. As an almost immediate consequence of their gradient estimate. where u ¼ jxj 2−n is harmonic on R n \ f0g. b ≥r Here. 14 (see also ref. in particular. to get the log-Sobolev inequality on R n .1203856109 Geometric Inequalities and Monotonicity We turn next to geometric applications of monotonicity and. We will next turn to some elliptic analogs of the functionals and formulas defined in the previous subsection. for instance. we will give some applications of these. From the formula for A 0 it follows. In the definition of A. then   j∇uj 2 ut n n − ¼ −tΔ log u − ≤ 0. Li–Yau originally proved their gradient estimate using the maximum principle. how one can use the functionals and formulas of the previous section to prove uniqueness of blow-ups and blow-downs of Einstein manifolds. at infinity to constant multiples of the usual Gaussian). in particular. by ref. then it is flat Euclidean space and u is the heat kernel. (16). A ↓ is monotone nonincreasing on a manifold with Ric ≥ 0. as it turned out. if f > 0 and ∫ M f 2 e −ϕ ¼ 1. For simplicity. The function b satisfies the following crucial formula: 2 of 4 ∣ www.

where tangents are shown to exist almost everywhere for any Lipschitz function. is by definition the infimum over all ϵ > 0 such that X and Y isometrically embed in a larger metric space ðZ. The Gromov–Hausdorff distance between them. We will also sometimes write ð0. ðr 2 . 19. ‡ r −n VolðBr ðxÞÞ [16] is monotone nonincreasing in the radius r of the ball Br ðxÞ for any fixed x ∈ M . see also ref. we mean the distance between the annuli after the metrics are rescaled so that the annuli have unit size. In perhaps its simplest form. by ref. Our next goal is to explain see also section 1 of ref. There is a similar inequality involving W 0 and Θr . 19. one would most of all like to work directly with and show some kind of decay for (in the scale). In a sense. and complex and algebraic geometry. 21 to show that. r i−2 gÞ. On its own it implies volume doubling and hence metric doubling and this gives directly compactness for the space of manifolds of a given dimension and a given lower bound of Ricci curvature. On the other hand. dZ Þ and X lies within an ϵ-tubular neighborhood of Y and vice versa. Y Þ. this is not at all the case in the current setting where the Bishop–Gromov is of very low regularity and ill suited: The distance function is Lipschitz but is not even C 1 . dY Þ is the metric completion of the set ð0. A metric cone CðY Þ over a metric space ðY . A similar result holds for local tangent cones of noncollapsed limits of Einstein manifolds. when we discuss applications to uniqueness of blow-downs or blow-ups. this is that dGH ðB4r ðxÞ \ Br ðxÞ. has a subsequence that converges in the Gromov–Hausdorff topology to a length space. The actual inequality is slightly more complicated as in reality the right-hand side of this inequality is not to the power 2 rather to the slightly worse power 2 þ 2ϵ for any ϵ > 0. However. It is one of the earliest fractal curves to be described and. its derivative has the right properties. ∞Þ × Y with the metric 2 2 2 dC ðY Þ ððr 1 . see ref. Koch described what is now known as the Koch curve or Koch snowflake. when a set or a curve has a well-defined tangent or welldefined blow-up at every point.†. suppose that ðX. dX Þ is a metric space and Br ðxÞ is a ball in X . For instance. When Y itself is a complete metric space. In particular. (By scale invariant distance. by ref. the issue of uniqueness comes up already in a 1904 paper entitled “On a continuous curve without tangents constructible from elementary geometry” by the Swedish mathematician Helge von Koch. denoted by dGH ðX. There is a rich history of uniqueness results for geometric problems and equations. This is a major point. ∞Þ × Y adds only one point to the space. y2 Þ. ∞Þ × X . if M has Ric ≥ 0. However. This is a prime example of how the monotone quantities given purely analytically give information about the geometry of the spaces. In the next subsection. then any sequence of rescalings ðM. dY Þ if dGH ðX i . and the constant C also depends on ϵ. amongst others. where r i → ∞. Tangents at every point. we need to recall a natural metric on the space of metric spaces. dY Þ are two compact metric spaces. not only is it not clear that Θr is monotone. This is the Gromov–Hausdorff distance that is a generalization of the classical Hausdorff distance between two subsets of the same Euclidean space. or uniqueness of blow-ups. ∞Þ ×r Y for the metric cone. arise in a number of different fields. In contrast. However. the classical monotone quantities are highly regular and are well-suited to this type of argument. furthermore. As mentioned compactness follows from doubling that is implied by monotonicity of the scale invariant volume. For tangent cones at infinity of manifolds with Ric ≥ 0 and VM > 0. The quantity A described above was used in ref. it will be key that A ¼ AðrÞ is monotone A ↓ on a manifold with Ric ≥ 0 and for some positive constant C −A 0 ðr Þ ≥ C Θr2 : r [17] The constant C depends only on a lower bound for the dimension and the scale invariant volume ‡. 19. Θr and Geometric Meaning of A 0 and W 0 . taking the completion of ð0. as suggested by the title. estimating the rate of convergence seems to require either integrability and/or a great deal of regularity (such as analyticity).Scale Invariant Volume Monotonicity and Consequences. see the next subsection for the precise definition of a metric cone over a general metric space. The most basic monotonicity for Ricci curvature is that of the scale invariant volume. Einstein manifolds and Ricci-flat manifolds. Y Þ → 0.) To make this precise. for minimal surfaces or harmonic maps. for any Ricci-flat manifold with Euclidean volume growth. Colding and Minicozzi PNAS Early Edition ∣ 3 of 4 MATHEMATICS SPECIAL FEATURE . it converges to a nonnegative number VM . 20 for more material related to this subsection. the geometric meaning of A 0 and similarly of W 0 and for that we will need to recall what a metric cone is. If the limit is noncompact. Suppose that ðX. This one point is usually referred to as the vertex of the cone. let alone analytic. This is usually called the Bishop–Gromov volume comparison theorem. 12. y2 ÞÞ ¼ r 1 þ r 2 − 2r 1 r 2 cos dY ðy1 . If VM > 0. We will next define the scale invariant distance between an annulus and the annulus in a cone centered at the vertex that approximates the annulus best. Gromov’s compactness theorem is the result that any sequence of manifolds of a given dimension and a given lower Ricci curvature bound has a subsequence that converges in the Gromov– Hausdorff topology to a length space. the functional A is defined on the level sets of an analytic function (the Green’s function) and does depend analytically and. where Br ðvÞ ⊂ CðY Þ and v is the vertex of the cone. To explain this. Uniqueness of Tangent Cones. Let Θr ðxÞ > 0 be the infimum of all Θ > 0 such that † 2 A metric cone CðX Þ with cross-section X is a warped product metric dr 2 þ r 2 dX on the space ð0. tangent cones at infinity are unique as long as one tangent cone has a smooth cross-section (22). B4r ðvÞ \ Br ðvÞÞ < Θ r. see ref. including string theory. this quantity on a smooth manifold converges to the volume of the unit ball in R n and. but as a purely metric quantity it is even less regular than the scale invariant volume. and there is an extensive literature of examples. any tangent cone at infinity is a metric cone. while the approaches to uniqueness rely on showing that the monotone quantity approaches its limit at a definite rate. Suppose now again that u > 0 is harmonic in a pointed neighborhood of p ∈ M and define b by b 2−n ¼ u and A by Aðr Þ ¼ r 1−n ∫ b¼r j∇bj 3 . then much regularity is known to follow. Any such limit is said to be a tangent cone at infinity of M . as opposed to say Rademacher’s theorem. existence of tangent cones comes from monotonicity. shows that there are continuous curves that do not have a tangent in any point. A sequence of compact metric spaces ðX i . then the convergence is on compact subsets. is a “hard” analytical fact that most often is connected with a PDE. In many geometric problems. As r tends to 0. 12. Uniqueness of tangent cones is a key question for the regularity of geometric PDEs. if one could. y1 Þ. as r tends to infinity. general relativity. the scale invariant volume is already a regularization of Θr that. dX Þ and ðY . any cross-section is a length space with diameter ≤π. in particular. then we say that M has Euclidean volume growth and. For a manifold with Ric ≥ 0. In that paper. dX i Þ is said to converge in the Gromov–Hausdorff topology to a compact metric space ðY .

(Springer. Vol 30. Colding TH (1997) On the structure of spaces with Ricci curvature bounded below. Cambridge. arXiv:1206. Rev Mat Iberoam 22:683–702. Int Math Res Not 2011:2177–2216. Tao T (2008) The Brascamp–Lieb inequalities: Finiteness. Cheeger J. 1. 12. J Geom Anal 14:87–100. Maurey B (2011) Correlation and Brascamp– Lieb inequalities for Markov semigroups. Berlin). 2. (Cambridge Univ Press. Berlin). pp 177–206.. and NSF Focused Research Group grants DMS 0854774 and DMS 0853501. Proc Am Math Soc 138:2755–2769. UK). Ni L (2004) The entropy formula for linear heat equation.DG/0211159/. Q is monotone nonincreasing and k −1 −ϵ < ∞ : ∑ k [21] Eq. 11.org/abs/1111. Colding TH (1996) Lower bounds on Ricci curvature and the almost rigidity of warped products. for some α < 1. The first key point is that if A ¼ Aðr Þ is defined as above. Qð2r Þ 2−α ≤ CðQðr∕2Þ − Qð2r ÞÞ: [24] ½Qðr ∕2Þ − Qð8r ފ ≥ C Θr2 : [18] We claim that uniqueness of tangent cones is implied by showing that A converges to its limit at infinity at a sufficiently fast rate or.org/cgi/doi/10. J Differential Geom 46:406–480. that is. J Funct Anal 101:194–211. Old and New Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. we shall use that. ed J Rakosnk Vol 8 (Czech Academy of Sciences. 1992). and we have [17]. Bez N (2009) Closure properties of solutions to heat inequalities. The proof of this comes from an infinite dimensional Lojasiewicz–Simon inequality. A ↓. Lectures on probability theory (Saint-Flour. then A is monotone. for a slightly larger ϵ. Colding TH New monotonicity formulas for Ricci curvature and applications. Am J Mat 119:1355–1398. 20 follows.. Cordero-Erausquin D. Colding TH. Invent Math 118:493–571. let again Θr be the scale invariant Gromov–Hausdorff distance between the annulus B4r ðpÞ \ Br ðpÞ and the corresponding annulus centered at the vertex of the cone that best approximates the annulus. uniqueness is implied by proving that Θ k < ∞: ∑ 2 k [19] This. was in residence at the Mathematical Science Research Institute (MSRI) in Berkeley. Cheeger J. arxiv. for Q roughly equal to −rA 0 ðr Þ. Perelman G The entropy formula for the Ricci flow and its geometric applications. Cordero-Erausquin D.1203856109 Colding and Minicozzi . Lieb EH. pp 165–166. Finally. Vol 1581. Acta Mathematica. pp 1–114. Naber A (2012) Characterization of tangent cones of noncollapsed limits with lower Ricci bounds and applications. in press. 14. We would like to show that the tangent cone at infinity is unique. The authors were partially supported by National Science Foundation (NSF) Grants DMS 11040934 and DMS 0906233. LNM. II pp 299–308. II (1997) Large scale behavior of kernels of Schrödinger operators. Ledoux M (2010) The geometry of Euclidean convolution inequalities and entropy. Cheeger J. Christ M. Mathematical Institute. To show this. I.) In fact. in turn. 21. This decay follows with rather elementary arguments from showing that.pnas. 24. 16.4929v1. 22. CA. Minicozzi WP. while T. Acta Math 156:153–201. see ref. [20] 4 of 4 ∣ www. equivalently. 13. Bennett JM. that Q decays sufficiently fast to zero. Colding TH. 3. Proceedings of the Spring School. Am J Mat 97:1061–1083. Emery P (1985) Diffusions hypercontractives. (Springer. Ann Math 144:189–237. Namely. (Springer. Function Spaces and Applications.org/abs/0708. Ledoux M. by the triangle inequality. Vol 1123. 17. ACKNOWLEDGMENTS. arxiv. Comparison Geometry. To explain some of the key points of how one shows uniqueness of tangent cones. we let p ∈ M be a fixed point in a Ricci-flat manifold with Euclidean volume growth. 7. II On uniqueness of tangent cones of Einstein manifolds. 5. Li P.Proving Uniqueness. Bakry D. 24. Colding TH. 4.H. Ye R The log entropy functional along the Ricci flow.. 8. does not depend on the sequence of blow-downs. there exist a large number of examples of manifolds with nonnegative Ricci curvature and Euclidean volume growth and nonunique tangent cones at infinity. Carbery A. Carlen EA. Ledoux M (2006) A logarithmic Sobolev form of the Li–Yau parabolic inequality. structure and extremals. Bennett JM. Gross L (1975) Logarithmic Sobolev inequalities.4715. 21 for details.C. Carlen EA (1991) Superadditivity of Fisher’s information and logarithmic Sobolev inequalities. Namely. (Perelman’s monotone W functional is also potentially a candidate. 25.1073/pnas. Nonlinear Analysis. Bakry D (1994) L’hypercontractivité et son utilisation en théorie des semigroupes. but it comes from integrating over the entire space which introduces so many other serious difficulties that it cannot be used. Q ðr Þ ≤ C : ðlog r Þ 1þϵ [23] [22] All the work is then to establish this crucial decay for Q. Vol 338. arxiv.org/abs/math. Geom Funct Anal 17:1343–1415. J Geom Anal 19:584–600. GAFA. 23. < ∞. Bakry D. Vol. by [18]. 18. Yau ST (1986) On the parabolic kernel of the Schrödinger operator. Villani C (2009) Optimal transport. Colding TH (1998) Spaces with Ricci curvature bounds. Minicozzi WP. Barthe F. Carbery A (2007) The Brascamp-Lieb inequalities: Recent developments. 19. Praha). 20. during the fall of 2011. 10. 6. pp 9–34. 9. note that it is well known that uniqueness may fail without the two-sided bound on the Ricci curvature. Perelman G (1997) A complete Riemannian manifold of positive Ricci curvature with Euclidean volume growth and nonunique asymptotic cone. 23. Tian G (1994) On the cone structure at infinity of Ricci flat manifolds with Euclidean volume growth and quadratic curvature decay. Loss M (2004) A sharp analog of Young’s inequality on S N and related entropy inequalities.2008. Proceedings of the International Congress of Mathematicians. This material is based upon work supported by the NSF under Grant 0932078. Berlin). J Geom Anal 14:487–520. I. from showing that ½Qð2 k−1 Þ − Qð2 kþ3 ފk 1þϵ < ∞: ∑ This is implied by proving that. is implied by the Cauchy—Schwarz inequality by showing that for some ϵ > 0 Θ 2 k 1þϵ ∑ 2k k as 1. 15. see refs. Lecture Notes in Math.