Geometrization of 3-Manifolds via the Ricci Flow Michael T. Anderson
Introduction results are already being used by others in research The classification of closed surfaces is a milestone on related topics. These circumstances serve to jus- in the development of topology, so much so that tify the writing of an article at this time, which oth- it is now taught to most mathematics undergrad- erwise might be considered premature. uates as an introduction to topology. Since the so- The work of Perelman builds on prior work of lution of the uniformization problem for surfaces Thurston and Hamilton. In the next two sections by Poincaré and Koebe, this topological classifica- we discuss the Thurston picture of 3-manifolds and tion is now best understood in terms of the the Ricci flow introduced and analyzed by Hamil- geometrization of 2-manifolds: every closed surface ton. For additional background, in particular on the Σ admits a metric of constant Gauss curvature +1, Poincaré Conjecture, see Milnor’s Notices survey [14] 0, or −1 and so is uniformized by one of the stan- and references therein. For much more detailed dard space-form geometries S2, R2, H2. Hence any commentary and discussion on Perelman’s work, surface Σ is a quotient of either the 2-sphere, the see [13]. Euclidean plane, or the hyperbolic disc by a discrete The Geometrization Conjecture group Γ acting freely and isometrically. The classification of higher-dimensional mani- While the Poincaré Conjecture has existed for about folds is of course much more difficult. In fact, due one hundred years, the remarkable insights of to the complexity of the fundamental group, a Thurston in the late 1970s led to the realistic pos- complete classification as in the case of surfaces sibility of understanding and classifying all closed is not possible in dimensions ≥ 4. In dimension 3 3-manifolds in a manner similar to the classifica- this argument does not apply, and the full classi- tion of surfaces via the uniformization theorem. fication of 3-manifolds has long been a dream of To explain this, we first need to consider what topologists. As a very special case, this problem in- are the corresponding geometries in 3-dimensions. cludes the Poincaré Conjecture. In terms of Riemannian geometry, a geometric In this article we report on remarkable recent structure on a manifold M is a complete, locally ho- work of Grisha Perelman [15]-[17], which may well mogeneous Riemannian metric g. Thus, M may be \ have solved the classification problem for 3- described as the quotient Γ G/H, where G is the manifolds (in a natural sense). Perelman’s work is isometry group of the universal cover (M,g) and currently under intense investigation and scrutiny by Γ, H are discrete and compact subgroups of the Lie many groups around the world. At this time, much group G respectively. Thurston showed that there of his work has been validated by experts in the area. are eight such simply connected geometries G/H 1 Although at the moment it is still too soon to declare in dimension 3 which admit compact quotients. a definitive solution to the problem, Perelman’s ideas As in two dimensions, the most important geome- are highly original and of deep insight. Morever, his tries are those of constant curvature: hyperbolic 1The Thurston classification is essentially a special case Michael T. Anderson is professor of mathematics at the of the much older Bianchi classification of homogeneous State University of New York, Stony Brook. His email space-time metrics arising in general relativity; cf. [3] address is [email protected]. for further remarks on the dictionary relating these Partially supported by NSF Grant DMS 0305865. classifications.
184 NOTICES OF THE AMS VOLUME 51, NUMBER 2 geometry H3 of curvature −1, Euclidean geometry M R3 of curvature 0, and spherical geometry S3 of cur- S2 vature +1. The remaining five geometries are prod- ucts or twisted products with the 2-dimensional 2 geometries. Trivial S1 bundles over a surface of S S2 genus g>1 have H2 × R geometry, while nontriv- S2 ial bundles have SL (2, R) geometry; nontrivial S1 bundles over T 2 have Nil geometry, while nontrivial T 2 bundles over S1 have Sol geometry (or Nil or R3 geometry); finally, S1 bundles over S2 have S2 × R (or S3) geometry. For example, any Seifert L fibered 3-manifold, a 3-manifold admitting a locally K2 1 1 free S action, has such a geometric structure. K1 Geometric 3-manifolds, that is 3-manifolds ad- mitting a geometric structure, are the building S2×S1 blocks of more complicated 3-manifolds. For sim- plicity, we assume throughout the article that all manifolds M are orientable. The building blocks are Figure 1. Sphere decomposition. then assembled along 2-spheres S2, via connected sum, and along tori T 2. As a simple example of such an assembly, let {Mi } be a finite collection of Seifert fibered 3-manifolds over surfaces Σi with non- K empty boundary, so that ∂Mi consists of tori. These H2 tori may then be glued together pairwise by dif- S feomorphisms to obtain a closed 3-manifold or a 2 T S1 3 3-manifold with toral boundary. A 3-manifold as- T 1 T2 sembled in this way is called a graph manifold. (One H1 assigns a vertex to each Seifert fibered space and T S an edge to each torus connecting two such Seifert 4 3 spaces). A torus bundle over S1 is a graph mani- fold, since it is the union of two Seifert fibered spaces over S1 × I. Graph manifolds were intro- duced, and their structure completely analyzed, by Figure 2. Torus decomposition, (Si Siefert fibered, Hj torus- Waldhausen. irreducible). Conversely, let M be an arbitrary closed (replacing regions S2 × I by two copies of B3); see 3-manifold, as above always orientable. One then Figure 1 for a schematic representation. decomposes or splits it into pieces according to The K factors in (1) may also contain topologi- the structure of the simplest surfaces embedded cally essential tori. A torus T 2 embedded in M is in M, namely spheres and tori. Topologically, this called incompressible if the inclusion map induces is accomplished by the following classical results an injection on π1. A 3-manifold N is called torus- in 3-manifold topology. irreducible if every embedded incompressible torus Sphere (or Prime) Decomposition (Kneser, may be deformed to a torus in ∂N. Hence, if Milnor) ∂N =∅, then N has no incompressible tori. Let M be a closed 3-manifold. Then M admits a Torus Decomposition (Jaco-Shalen, Johannsen) finite connected sum decomposition Let M be a closed, irreducible 3-manifold. Then there is a finite collection, possibly empty, of dis- (1) M = (K #...#K )#(L #...#L )#(#r S2 × S1). 1 p 1 q 1 joint incompressible tori in M that separate M into The K and L factors here are closed irreducible a finite collection of compact 3-manifolds (with 3-manifolds; i.e. every embedded 2-sphere S2 toral boundary), each of which is torus-irreducible bounds a 3-ball. The K factors have infinite fun- or Seifert fibered. damental group and are aspherical 3-manifolds A coarser, but essentially equivalent, decompo- (K(π,1)’s), while the L factors have finite funda- sition is given by tori separating M into torus- mental group and have universal cover a homotopy irreducible and graph manifold components; see 3-sphere. Since M#S3 = M, we assume no L factor Figure 2. is S3 unless M = L = S3. The factors in (1) are then With the simple exceptions of S2 × S1 and its ori- Z 2 × 1 RP3 RP3 unique up to permutation and are obtained from ented 2 quotient S Z2 S # , essential M by performing surgery on a collection of essen- 2-spheres are obstructions to the existence of a geo- tial, i.e. topologically nontrivial, 2-spheres in M metric structure on a 3-manifold. The same is true
FEBRUARY 2004 NOTICES OF THE AMS 185 for essential tori, unless M happens to be a Seifert Independently and around the same time as fibered or a Sol 3-manifold. Thus, the sphere and Thurston, Gromov [6], [7] also studied the defor- torus decompositions divide M topologically into mation and degeneration of more general Rie- pieces where these known obstructions are re- mannian metrics with merely bounded curvature moved. in place of constant curvature. The idea is that Geometrization Conjecture (Thurston). Let M be one can control the behavior of a metric, or of a a closed, oriented 3-manifold. Then each component family of metrics, given a uniform bound on the 2 of the sphere and torus decomposition admits a Riemann curvature tensor Riem of the metric. geometric structure. This leads to the important Gromov compactness theorem, the structure theory of almost flat man- The geometrization conjecture gives a complete ifolds, and the theory of collapsing Riemannian and effective classification of all closed 3-manifolds, manifolds, worked out in detail with Cheeger and closely resembling in many respects the classifi- Fukaya. cation of surfaces. More precisely, it reduces the One version of these results is especially rele- classification to that of geometric 3-manifolds. vant for our purposes. Let (M,g) be a closed The classification of geometric 3-manifolds is rather Riemannian manifold, normalized to unit volume, simple and completely understood, except for the and suppose case of hyperbolic 3-manifolds, which remains an active area of research. (2) |Riem|≤Λ, As an illustration of the power of the Thurston for some arbitrary constant Λ < ∞. The metric g Conjecture, let us see how it implies the Poincaré Conjecture. If M is a simply connected 3-manifold, provides a natural decomposition of M into thick = ν ∪ then the sphere decomposition (1) implies that M and thin parts, M M Mν , where must be an L factor. The geometrization conjecture ν (3) M ={x ∈ M : volBx(1) ≥ ν}, implies that L is geometric, and so L = S3/Γ. Hence, ={ ∈ } M = L = S3. Mν x M : volBx(1) <ν ; Thurston’s formulation and work on the here B (1) is the geodesic ball about x of radius 1 and geometrization conjecture revolutionized the field x ν>0 is an arbitrary but fixed small number. Now of 3-manifold topology; see [18], [19] and further consider the class of all Riemannian n-manifolds of references therein. He recognized that in the class unit volume satisfying (2), and consider the corre- of all (irreducible) 3-manifolds, hyperbolic 3- sponding decompositions (3). Then the geometry manifolds are overwhelmingly the most prevalent, and topology of Mν is a priori controlled. For any given as is the case with surfaces, and developed a vast array of new ideas and methods to understand the ν>0, there are only a finite number, (depending on ν structure of 3-manifolds. Thurston and a number Λ and ν), of possible topological types for M . More- ν of other researchers proved the geometrization over, the space of metrics on M is compact in a nat- conjecture in several important cases, the most ural sense; any sequence has a subsequence con- 1,α celebrated being the Haken manifold theorem: if M verging in the C topology, α<1 (modulo is an irreducible Haken 3-manifold, i.e. M contains diffeomorphisms). For ν sufficiently small, the com- an incompressible surface of genus ≥ 1, then the plementary thin part Mν admits an F-structure in the geometrization conjecture is true for M. sense of Cheeger-Gromov; in dimension 3 this just An important ingredient in the Thurston ap- means that Mν is a graph manifold with toral (or empty) proach is the deformation and degeneration of boundary. In particular, the topology of Mν is strongly hyperbolic structures on noncompact manifolds (or restricted. A metric on Mν is highly collapsed in the the deformation of singular hyperbolic structures sense that the circles in the Seifert fibered pieces of on compact manifolds). The eight geometric struc- Mν and the tori glueing these pieces together have tures are rigid in that there are no geometries very small diameter, depending on ν; see Figure 3 for which interpolate continuously between them. a schematic picture. Moreover, for any fixed ν>0, Hence, on a composite 3-manifold M, the geomet- the distance between Mν and the arbitrarily thin part ric structure on each piece must degenerate in Mν becomes arbitrarily large as ν /ν → 0. passing from one piece to the next; there is no sin- We point out that similar results hold locally and gle structure or metric giving the geometrization for complete noncompact manifolds; thus the unit of all of M. For example, in Figure 2 the H pieces volume normalization above is not essential. may be hyperbolic 3-manifolds separated by tori from Seifert fibered pieces S. Although this split- 2The curvature tensor is a complicated (3,1) tensor ting is topologically well defined, the geometries expressed in terms of the second derivatives of the do not match in the glueing region, and metrically metric; in a local geodesic normal coordinate system there is no natural region in which to perform the at a given point, the components of Riem are given by l =−1 + − − glueing. Rijk 2 (∂i ∂kgjl ∂j ∂l gik ∂j ∂kgil ∂i ∂l gjk).
186 NOTICES OF THE AMS VOLUME 51, NUMBER 2 The Thurston approach to geometrization has made a great deal of progress on the “hyperbolic M ν M ν part” of the conjecture. In comparison with this, relatively little progress has been made on the M ν “positive curvature part” of the conjecture, for example the Poincaré Conjecture. It is worth Mν pointing out that among the eight geometries, the constant curvature geometries H3 and S3 are by far the most important to understand (in terms of Figure 3. Thick-thin decomposition. characterizing which manifolds are geometric); the other (mixed) geometries are much simpler in example, rescaling the Ricci flow (5) so that the comparison. volume of (M,g(t)) is preserved leads to the flow From the point of view of Riemannian geome- equation (4) with λ = 2 R, twice the mean value try, the Thurston conjecture essentially asserts the of the scalar curvature R. existence of a “best possible” metric on an arbitrary In a suitable local coordinate system, equation closed 3-manifold. In the case that M is not itself (5) has a very natural form. Thus, at time t, choose geometric, one must allow the optimal metric to local harmonic coordinates so that the coordinate have degenerate regions. The discussion and fig- functions are locally defined harmonic functions ures above suggest that the degeneration should in the metric g(t). Then (5) takes the form be via the pinching off of 2-spheres (sphere de- d composition) and collapse of graph manifolds along (6) gij = ∆gij + Qij(g,∂g), circles and tori (torus decomposition). dt where ∆ is the Laplace-Beltrami operator on func- The Ricci Flow tions with respect to the metric g = g(t) and Q is One method to find a best metric on a manifold is a lower-order term quadratic in g and its first- to find a natural evolution equation, described by order partial derivatives. This is a nonlinear a vector field on the space of metrics, and try to heat-type equation for gij. From the analysis of prove that the flow lines exist for all time and con- such PDE, one obtains existence and uniqueness of verge to a geometric limit. In case a flow line does solutions to the Ricci flow on some time interval, not converge, the corresponding metrics degener- starting at any smooth initial metric. This is the ate, and one then needs to relate the degeneration reason for the minus sign in (5); a plus sign leads with the topology of M. to a backwards heat-type equation, which has no There is essentially only one simple and natural solutions in general. vector field (or more precisely family of vector Here are a few simple examples of explicit so- fields) on the space of metrics. It is given by lutions to the Ricci flow. If the initial metric g(0) is of constant Ricci curvature, Ric = a · g, then the d (4) g(t) =−2Ricg(t) + λ(t) · g(t). evolution g(t) is just a rescaling of g(0): dt g(t) = (1 − 2at)g(0). Note that if a>0, then the flow contracts the metric, while if a<0, the flow Here Ric is the Ricci curvature, given in local co- = = k expands the metric, uniformly in all directions. ordinates by Rij (Ric)ij k Rikj , so that Ric is a trace of the Riemann curvature. The constant 2 Hence, if one rescales g(t) to have constant volume, is just for convenience and could be changed by the resulting curve is constant. The stationary rescaling the time parameter; λ(t) is a constant de- points of the volume-normalized Ricci flow are ex- pending on time t. The Ricci flow, introduced by actly the class of Einstein metrics, i.e. metrics of Hamilton [11], is obtained by setting λ = 0, i.e. constant Ricci curvature. In dimension 3, Einstein metrics are of constant curvature and so give the d H3, R3 and S3 geometries. (5) g(t) =−2Ric . dt g(t) More generally, if Ric(x, t) > 0, then the flow contracts the metric g(t) near x, to the future, The reason (4) is the only natural flow equation is while if Ric(x, t) < 0, then the flow expands g(t) essentially the same as that leading to the Einstein near x. At a general point, there will be directions field equations in general relativity. The Ricci of positive and negative Ricci curvature along which curvature is a symmetric bilinear form, as is the the metric will locally contract or expand. metric. Besides multiples of the metric itself, it is Suppose g(0) is a product metric on S1 × Σ, the only such form depending on at most the where Σ is a surface with constant curvature met- second derivatives of the metric, and invariant ric. Then g(t) remains a product metric, where the under coordinate changes, i.e. a (2, 0) tensor length of the S1 factor stays constant while the sur- formed from the metric. By rescaling the metric and face factor expands or contracts according to the time variable t, one may transform (5) into (4). For sign of its curvature.
FEBRUARY 2004 NOTICES OF THE AMS 187 Finally, the Ricci flow commutes with the action a general lower bound Ric ≥−λ, for λ>0. For the of the diffeomorphism group and so preserves all remainder of the paper, we assume then that isometries of an initial metric. Thus, geometric 3- dimM = 3. manifolds remain geometric. For the “nonpositive” In the Gromov compactness result and thick/thin mixed geometries H2 × R, SL (2,R), Nil, and Sol, decomposition (3), the hypothesis of a bound on the volume-normalized Ricci flow contracts the S1 |Riem| can now also be replaced by a bound on or T 2 fibers and expands the base surface factor, |Ric| (since we are in dimension 3). Further, on time while for the positive mixed geometry S2 × R, the intervals [0,t] where |Ric| is bounded, the metrics volume-normalized flow contracts the S2 and g(t) are all quasi-isometric to each other: expands in the R-factor. cg(0) ≤ g(t) ≤ Cg(0) as bilinear forms, where c,C Now consider the Ricci flow equation (5) in gen- depend on t. Hence, the arbitrarily thin region Mν, eral. From its form it is clear that the flow g(t) will ν<<1, can only arise, under bounds on |Ric|, in continue to exist if and only if the Ricci curvature arbitrarily large times. remains bounded. This suggests one should con- The discussion above shows that the Ricci flow sider evolution equations for the curvature, in- is very natural and has many interesting proper- duced by the flow for the metric. The simplest of ties. One can see some relations emerging with the these is the evolution equation for the scalar Thurston picture for 3-manifolds. However, the ij curvature R = trgRic = g Rij: first real indication that the flow is an important d new tool in attacking geometric problems is the (7) R = ∆R + 2|Ric|2. following result of Hamilton: dt • Space-form Theorem [8]. If g(0) is a metric of
Evaluating (7) at a point realizing the minimum Rmin positive Ricci curvature on a 3-manifold M, then of R on M gives the important fact that Rmin is the volume-normalized Ricci flow exists for all monotone nondecreasing along the flow. In par- time and converges to the round metric on S3/Γ, ticular, the Ricci flow preserves positive scalar cur- where Γ is a finite subgroup of SO(4) acting freely 3 vature (in all dimensions). Moreover, if Rmin(0) > 0, on S . d ≥ 2 2 Thus the Ricci flow “geometrizes” 3-manifolds then the same argument gives dt Rmin n Rmin , n = dimM, by the Cauchy-Schwarz inequality of positive Ricci curvature. Since this ground- | |2 ≥ 1 2 breaking result, it has been an open question Ric n R . A simple integration then implies whether it can be generalized to initial metrics n (8) t ≤ . with positive scalar curvature. 2R (0) min Although the evolution of the curvature along Thus, the Ricci flow exists only up to a maximal time the Ricci flow is very complicated for general ini- tial metrics, a detailed analysis of (9) leads to the T ≤ n/2Rmin(0) when Rmin(0) > 0. In contrast, in regions where the Ricci curvature stays negative following important results: definite, the flow exists for infinite time. • Curvature pinching estimate [10], [12]. Let g(t) The evolution of the Ricci curvature has the be a solution to the Ricci flow on a closed same general form as (7): 3-manifold M. Then there is a nonincreasing func- tion φ :(−∞, ∞) → R, tending to 0 at ∞, and a con- d (9) R = ∆R + Q . stant C, depending only on g(0), such that dt ij ij ij (10) Riem(x, t) ≥−C − φ(R(x, t)) ·|R(x, t)|. The expression for Q is much more complicated than the Ricci curvature term in (7) but involves only This statement means that all the sectional curva- quadratic expressions in the curvature. However, tures Rijji of g(t), where ei is any orthonormal Q involves the full Riemann curvature tensor Riem basis at (x, t), are bounded below by the right side of g and not just the Ricci curvature (as (7) in- of (10). volves Ricci curvature and not just scalar curvature). This estimate does not imply a lower bound on An elementary but important feature of dimen- Riem(x, t) uniform in time. However, when com- sion 3 is that the full Riemann curvature Riem is bined with the fact that the scalar curvature R(x, t) determined algebraically by the Ricci curvature. is uniformly bounded below, it implies that This implies that, in general, Ricci flow has a much |Riem|(x, t) >> 1 only where R(x, t) >> 1. Hence, better chance of “working” in dimension 3. For ex- to control the size |Riem| of the full curvature, it ample, an analysis of Q shows that the Ricci flow suffices to obtain just an upper bound on the scalar preserves positive Ricci curvature in dimension 3: curvature R. This is remarkable, since the scalar if Ricg(0) > 0, then Ricg(t) > 0, for t>0. This is not curvature is a much weaker invariant of the the case in higher dimensions. On the other hand, metric than the full curvature. Moreover, at points in any dimension > 2, the Ricci flow does not pre- where the curvature is sufficiently large, (10) shows serve negative Ricci curvature, nor does it preserve that Riem(x, t)/R(x, t) ≥−δ, for δ small. Thus,
188 NOTICES OF THE AMS VOLUME 51, NUMBER 2 if one scales the metric to make R(x, t) = 1, then A singularity can form for the Ricci flow Riem(x, t) ≥−δ. In such a scale, the metric then only where the curvature becomes unbounded. 2 has almost nonnegative curvature near (x, t). Suppose then that one has λi =|Riem|(xi ,ti ) →∞, • Harnack estimate [9]. Let (N,g(t)) be a solution on a sequence of points xi ∈ M , and times to the Ricci flow with bounded and nonnegative cur- ti → T<∞. It is then natural to consider the vature Riem ≥ 0, and suppose g(t) is a complete rescaled metrics and times ≤ Riemannian metric on N. Then for 0