Perelman's Proof of The Poincare Conjecture PDF

PERELMAN’S PROOF OF THE POINCARÉ CONJECTURE: A
NONLINEAR PDE PERSPECTIVE

arXiv:math/0610903v1 [math.DG] 29 Oct 2006
TERENCE TAO
Abstract. We discuss some of the key ideas of Perelman’s proof of Poincaré’s

conjecture via the Hamilton program of using the Ricci flow, from the perspec-
tive of the modern theory of nonlinear partial differential equations.
1. Introduction
1.1. Perelman’s theorem. In three remarkable papers [28], [29], [30] in 2003,
Grisha Perelman significantly advanced the theory of the Ricci flow, which was a
quasilinear PDE introduced by Richard Hamilton [16] and then extensively studied
by Hamilton and others in a program to understand the topology of manifolds. In
particular, in [29] Perelman showed that in three spatial dimensions there was a well-
defined Ricci flow with surgery procedure, which is the Ricci flow PDE punctuated
at a discrete set of times by carefully chosen surgery operations which may possibly
change the topology of the manifold, with the following remarkable property:
Theorem 1.2 (Perelman’s theorem, finite time extinction case). [28], [29], [30] Let
M = (M0 , g0 ) be a closed smooth three-dimensional compact Riemmanian mani-
fold whose fundamental group is a free product of finite groups and infinite simple
groups1. Then the Ricci flow with surgery t 7→ Mt = (Mt , gt ) is well-defined for all
t ≥ 0, and becomes extinct in finite time (thus there exists T∗ such that Mt = ∅ for
all t > T∗ ). Furthermore, at all times t ≥ 0 the manifold Mt is topologically equiv-
alent to a finite union of connected sums of quotients of S 3 (which in particular
includes the projective space RP 3 as a special case), and S 2 -bundles over S 1 .
Applying the conclusion of this theorem at t = 0, and using standard topological
methods to detect which of the possible connected sums are simply connected,
one obtains as a corollary the Poincaré conjecture: every closed smooth simply
connected three-dimensional manifold is topologically equivalent to a sphere2. A
more difficult version of this theorem proven in [28], [29], [30] for general (not
necessarily simply-connected) closed smooth manifolds M = M0 does not have the
finite extinction time result (which is false), but through an analysis of the final
1This hypothesis is in fact quite natural, because the arguments also show conversely that finite
time extinction for Ricci flow is only possible when the fundamental group is a free product of
finite groups and infinite simple groups. For the purposes of proving just the Poincaré conjecture,
though, it is possible to work entirely in the category of simply connected manifolds throughout,
although this only simplifies the argument at one small point (in the proof of finite time extinction).
We thank John Morgan for clarifying this point.
2In fact, an additional argument, noting that surgery cannot destroy the simple-connectedness
of the manifold components, demonstrates that the Ricci flow with surgery from a simply con-
nected manifold only involves finite unions of disconnected spheres. We thank John Morgan for
clarifying this point.
1
2 TERENCE TAO
state (or “scattering state”) as t → +∞, enough control on the topology of Mt

at each time is obtained to prove the geometrisation conjecture of Thurston (see
[5], [22] for full details, including several clarifications and alternate derivations on
certain difficult parts of Perelman’s papers; see also the paper [35] which fleshes
out a key step in the geometrisation argument).
Perelman’s result thus completes Hamilton’s program and settles two major con-
jectures in geometry and topology, and shows the way for further progress in these
fields. But, as one can already see from the statement of Theorem 1.2, a large part
of Perelman’s work is actually conducted3 in the arena of nonlinear PDE. Indeed,
even if one deleted the final claim in Theorem 1.2 (thus removing the applications
of this theorem to geometry and topology), the remaining arguments (and its gen-
eralisation to non-simply-connected manifolds), which are almost entirely nonlinear
PDE arguments, would already be (in my opinion) the most technically impressive
and significant result in the field of nonlinear PDE in recent years; the fact that
this PDE result also gives the Poincaré conjecture and the more general geometri-
sation conjecture makes it (again in my opinion) the best piece of mathematics we
have seen in the last ten years. It is truly a landmark achievement for the entire
discipline.
1.3. Purpose of these notes. Perelman’s own description [28], [29], [30] of the
arguments, which appeared in 2003, are highly succinct, with many steps which
would take over a page if written out in full often compressed to a few lines; for a
few years it was not entirely clear whether the argument was complete and correct,
though it was evident from the very beginning that major breakthroughs had been
made regardless4. However, in most cases these steps were of a standard type
known to the experts, and could be fleshed out with reasonable work; only in a
few cases were there significant enough issues with one or two steps which required
more non-trivial effort to resolve5, though, as noted in [22, §1], all such issues
could be resolved using the general methods of Hamilton and Perelman. Very
3It is a remarkable fact, appreciated in the last few decades (see e.g. [36]), that the flows of
nonlinear PDE are an extraordinarily effective tool for understanding geometry and topology, and
in particular in placing topological objects in a geometric “normal form”. One way to view this
is that the continuous flows of PDE, especially when augmented with a surgery procedure, are a
geometrically natural successor to the more classical discrete combinatorial algorithms that were
employed to understand topology (e.g. by taking a triangulation of a manifold and then simplifying
this triangulation by a sequence of combinatorial maneuvres). While a PDE flow is in many ways
“dumber” than a combinatorial algorithm (though this can be partially rectified by an intelligent
choice of surgery procedure), if the flow is sufficiently geometrical in nature then the flows acquire
a number of deep and delicate additional properties, most notably monotonicity formulae, which
sometimes have no obvious analogue for discrete combinatorial algorithms. Indeed, the non-
geometric nature of most combinatorial arguments introduces a certain amount of artificial “noise”,
thus drowning out some of the more subtle monotonicity properties and making it more difficult
to ensure that any given algorithm actually terminates in finite time.
4In particular, it was soon realised that Perelman had discovered a new monotone quantity for
Ricci flow which was both critical (scale-invariant) and coercive (geometry-controlling). While
this by itself does not automatically prove the Poincaré conjecture, it made several previously
impossible-looking steps in Hamilton’s program now become potentially tractable. In comparison
the global regularity problem for Navier-Stokes (another major unsolved problem in nonlinear
parabolic PDE) is considered extremely difficult precisely because no such critical coercive quantity
for that flow is known.
5For instance, in [29] some minor corrections to [28] were noted.
PERELMAN’S PROOF OF POINCARÉ 3
recently (in 2006), three essentially simultaneous expositions6 of Perelman’s work

have appeared, by Kleiner-Lott [22], Morgan-Tian [27], and Cao-Zhu [5]. Each of
these works, which emerged from large collaborative projects, some of which were
in communication with each other and with Perelman, have performed an immense
service to the field by giving an complete and detailed account of the entirety of
Perelman’s arguments7, and it is now certain that Perelman’s original argument was
indeed essentially complete and correct in every important detail. In particular the
notes here draw heavily on these three expositions as well as on the original papers
of Perelman.
In these notes, I will attempt to give a high-level overview of the proof of Theorem
1.2, from the perspective of nonlinear evolution equations, while also attempting a
sort of “concordance” between Perelman’s papers and the three expository mono-
graphs of Kleiner-Lott, Morgan-Tian, and Cao-Zhu. The main point of this is to
try to convey some aspects of Perelman’s work on Hamilton’s program to those
mathematicians versed in nonlinear PDE, but are not experts in either geometry
or topology. In particular, I want to convey that Perelman’s work is in many ways
the epitome of a nonlinear PDE argument, in which all the major milestones in the
argument are familiar PDE milestones8, but where the execution requires extremely
delicate analysis and several major technical breakthroughs, many of which arise
by carefully exploiting the rich geometric structure of Ricci flow.
In order to keep things as simple as possible I will only focus on Theorem 1.2,
which is sufficient for Poincare’s conjecture, and not discuss the more complicated
version of this theorem which also yields the more general geometrisation conjecture.
This simpler theorem already captures a large fraction of the breakthroughs that
Perelman made, and should already be sufficient for conveying the points discussed
in the previous paragraph.
Because the focus of these notes will be on the nonlinear PDE aspects of the
work, the other aspects from geometry and topology will be downplayed. In par-
ticular, various geometric structural equations arising from Ricci flow shall be pre-
sented simply as miraculous identities, with no derivation or intuitive geometric
explanation given; also, the use of standard topological tools to convert one type of
topological control on a manifold (such as control of homotopy groups) to another
will be sketched only briefly. Thus these notes should not be in any way construed
as a balanced presentation of Perelman’s work or more generally of Hamilton’s pro-
gram, as it shall be by conscious design heavily slanted towards the PDE component
of the work. This is in part for reasons of space and time, but also because I am
not sufficiently qualified in geometry or topology to expertly comment on the other
components of these papers.
We will not attempt here to give anything resembling a complete or detailed
presentation of Perelman’s work, or even of just the nonlinear PDE aspects of this
6Preliminary versions of parts of [22] were available as early as June 2003, and influenced all
three subsequent expositions.
7The book [27] deals only with the Poincaré conjecture aspects of Perelman’s work, and not
on the more lengthy version of the argument establishing the geometrisation conjecture.
8Indeed, the argument even goes beyond standard PDE milestones, such as classification of
blowup singularities, but also moves onward to new milestones, such as describing a surgery
procedure to remove singularities and continuing to control the flow beyond these singularities,
which are still considered too ambitious a project for most other PDE (unless one is content with
only possessing very weak solutions for which one has little control).
4 TERENCE TAO
work; for such presentations, we refer the reader to [5], [22], [27]. In particular we
will often choose to sacrifice rigour for sake of exposition, and in fact will commit
a number of small lies9 in order to suppress certain technical issues which might
otherwise obscure the underlying ideas. In a similar spirit, we will often not discuss
the strongest version of any given component of the proof, but instead focus on a
much simpler model proposition which is not strong enough for the full proof of
Poincaré, but does indicate the main ideas that are used in the proof of the true
proposition. We will also not attempt to systematically survey the literature and
other context which Perelman’s work and Hamilton’s program builds upon and is
placed in; for this, we refer the reader to [25], [26], [1], [37]. Finally, it should be
noted that the arguments here are arranged in an ahistorical order, in order to
expedite the natural flow of the proof. In particular, recent arguments (such as
those of Perelman) will sometimes be juxtaposed with earlier arguments (such as
those of Hamilton), reflecting the fact that different parts of Hamilton’s program
to prove the Poincaré conjecture via Ricci flow were understood at different times.
As has been made clear in several other places (see e.g. the introduction of [5]),
the primary contributions here are those of Hamilton and Perelman. Of course,
as with all other modern mathematical achievements, many other mathematicians
in the field have also made important contributions, either by laying foundational
work, establishing tools, developing insights or suggestions, or by clarifying, refin-
ing, generalising, or giving alternate proofs of key results. In particular, as has
been mentioned earlier, the efforts in [5], [22], [27] to fully explicate all the details
of Perelman’s arguments have been a significant, and extremely non-trivial, service
to the field. Assigning credit to each particular idea or insight is of course difficult
to do with absolute accuracy. While we will try our best to attribute things here,
we will refer the reader to the careful and detailed expositions in [5], [22], [27] for
a more thorough treatment of each of the steps and their provenance.
1.4. Overview of argument. In the jargon of nonlinear PDE, the Ricci flow is a
quasilinear parabolic system, and Theorem 1.2 is (in part) a finite time blowup re-
sult10 from essentially arbitrary data for this system. But it is also much more than
merely a finite time blowup result; it gives quite precise control of the behaviour
near each blowup point, and consequently also allows one to continue the evolution
beyond each blowup point (which is a level of understanding which is currently not
enjoyed by many commonly studied nonlinear PDE, even those that are substan-
tially simpler11). The need to continue the evolution even after the appearance of
singularities is essential for the topological applications of the theorem, because the
topological structure of the evolved manifold Mt is only obvious at late times (in
9We will attempt to apologise for these lies using footnotes whenever feasible, and cross-
reference to the more rigorous statements in other papers when appropriate.
10In particular this suggests that the nonlinear component of this system is “focusing” in
nature, acting in the direction of blowup rather than in the direction of dissipation.
11This is perhaps an unfair comparison. While the Ricci flow is much more nonlinear than some
other model nonlinear equations, such as the semilinear heat, wave, and Schrödinger equations, it
is parabolic rather than hyperbolic or dispersive, and perhaps more importantly is an extremely
geometric equation, thus enjoying significantly more structure than most other model nonlinear
PDE. The combination of the parabolic and the geometric natures of the flow in particular leads to
a very large number of monotonicity formulae, which are essential to the analysis and which would
not be expected to be present for many other model nonlinear PDE. We thank Igor Rodnianski
for emphasising this point.
particular, after the extinction time T∗ ), and then to reconstruct the topology of
the initial manifold M0 one needs to understand the nature of the blowup at all
blowup times between t = 0 and the extinction time t = T∗ , and not just the first
such blowup time12.
The proof of Theorem 1.2 follows standard PDE paradigms, even if the execution
of these paradigms is extraordinarily non-trivial. Firstly, a local-in-time existence
result for the Ricci flow is needed. This was first established by Hamilton [16], with
a simpler proof given by DeTurck [14], and is sketched out in Section 4. With the
local existence result in hand, Theorem 1.2 then reduces to a structural result on
the blowup times, and on the nature of the blowup singularities at each of these
times.
To achieve the latter task, one applies a standard rescaling limit argument
around each blowup singularity to understand the asymptotic profile of the blowup.
These asymptotic profiles were established in special cases by Hamilton [20] before
Perelman’s work, and conjectured in general. One major obstacle to establishing
this conjecture was the need for critical (scale-invariant) and coercive (geometry-
controlling) quantities as one approached a singularity of Ricci flow. For instance, a
paper of Hamilton [20] largely classifies the asymptotic profiles of singularities un-
der the scale-invariant assumption that the renormalised curvatures and injectivity
radius stayed bounded. However, it was not known at that time how to attain such
scale-invariant assumptions.
Previously to Perelman’s work, the type of controlled quantities which were
known were either supercritical in nature, which meant that they became dra-
matically weaker for the purposes of controlling the solution as one approaches
the singularity, or non-coercive, thus not controlling all aspects of the geometry
(except under special assumptions such as positive curvature). This raises the pos-
sibility that the rescaled solutions around that singularity might not be sufficiently
bounded to extract a limit. The first major breakthrough in Perelman’s work is to
introduce a number of new critical monotone quantities, which are coercive enough
to provide adequate boundedness on the rescaled solutions (in particular, uniform
bounds on the Riemann curvature) to provide a limit (by Hamilton’s compactness
theory [19] for Ricci flow, based in turn on the Cheeger-Gromov theory [7] of limits
of Riemannian manifolds). This very important no-collapsing theorem of Perel-
man13 controls the renormalised geometry as one approaches blowup, and allows
one to classify the asymptotic profiles of blowup singularities; a refinement of these
arguments then allows one to understand the geometry of high-curvature regions
quite accurately also. The latter is then crucial for defining a surgery procedure
properly.
12In the groundbreaking earlier work of Hamilton [16], it was shown that if the manifold M
0
initially has positive Ricci curvature everywhere, then there is extinction in finite time with no
intervening blowups, and that the metric becomes asymptotically round.
13
Actually, Perelman establishes a variety of no-collapsing theorems with variations in the hy-
potheses and conclusions, and for Ricci flow with and without surgery, as do the authors following
him: see [28, §4,7,8], [29, §7], [27, §8], [22, §12, 25, 27, 78], [5, §3.3, 3.4, 7.6].
6 TERENCE TAO
The next major task is to establish finite extinction time14 for simply connected
manifolds. This was achieved simultaneously by Perelman [30] and by Colding-
Minicozzi [13], and also explained in detail in [27, §18]. While non-trivial, these
results are somewhat simpler in nature than those in [28], [29], relying on a different
monotone quantity from that considered earlier15 which was known to be both
decreasing and positive for Ricci flow, which can then be used to deduce finite
extinction time (after carefully attending to the behaviour of these quantities during
surgery).
Finally, there is the endgame in which one returns back from the final extinction
time to the original state of the Riemannian manifold, and verifies that the topology
is always a finite union of spheres. This turns out to be relatively easy once the
other two major tasks are accomplished.
1.5. Acknowledgements and history. The author thanks Igor Rodnianski for
helpful discussions and John Lott, Gang Tian and especially John Morgan for many
clarifications and insights. The material here is of course heavily based on several of
the papers in the references, but the author was particularly reliant on all three of
the excellently detailed expositions of Perelman’s work in [22], [27], [5], in addition
to Perelman’s original papers [28], [29], [30]. Indeed, while these papers clearly have
strong overlap (as should be obvious from the discussion below), they each have
slightly different perspectives, strengths, and weaknesses, and by reading all these
papers in parallel (in particular, switching from one paper to another whenever
I was stuck on a particular point) I was able to obtain a far richer view of the
argument than I could have obtained just from reading one of them.
An initial version of this manuscript was posted on my web site on October 9,
2006. After receiving several comments and corrections, a revised version was then
placed on the ArXiV on October 29, 2006. At this present time, I am not planning
to submit this manuscript for publication.
2. Notation
I will be using notation from Riemannian geometry, which I will define very
rapidly here without providing any motivation as to where these geometric defini-
tions and identities arise from. For reasons of personal preference I tend to express
tensors using coordinates, for instance the metric g would be expressed as gαβ ,
with the usual raising and lowering and summation conventions. In many other
treatments it is more customary to use coordinate-free notation.
Throughout these notes, all Riemannian manifolds M = (M, g) will be smooth
compact and three-dimensional; while a substantial portion of Perelman’s work
does extend to general dimension, we will not discuss those generalisations in order
to slightly simplify the discussion.
Given a Riemannianpmetric g = gαβ , one has the inverse metric g −1 = g αβ , and
the volume form dg = det(gαβ )dx. The Levi-Civita connection ∇ is then defined
14
This is the “shortcut” approach to proving the Poincaré conjecture. There is a more difficult
approach that gives not only this conjecture, but also the geometrization conjecture, which allows
for the Ricci flow-with-surgery to evolve for infinite time, but we will not discuss this more complex
approach here.
15
In particular, this new monotone quantity is no longer critical, but criticality is not as relevant
for blowup time bounds as it is for blowup singularity analysis, as rescaling is not involved.
in coordinates via the Christoffel symbol

1 αδ
Γα
βγ :=g (∂β gγδ + ∂γ gβδ − ∂δ gβγ ).
2
This in turn leads to the Riemann curvature tensor
Riemδαβγ := ∂α Γδβγ − ∂β Γδαγ + Γσβγ Γδασ − Γσαγ Γδβγ
which then contracts to form the Ricci curvature tensor
Ricαγ := Riemβαβγ
which in turn contracts to form the scalar curvature
R := g αγ Ricαγ .
From the point of view of dimensional analysis16, it is helpful to keep in mind the
schematic form of these tensors (thus omitting indices, signs, and constants):
Γ = O(g −1 ∂g)
Riem, Ric = O(g −1 ∂ 2 g) + O(g −2 (∂g)(∂g))
R = O(g −2 ∂ 2 g) + O(g −3 (∂g)(∂g)).
In three spatial dimensions, one also has dg = O(g 3/2 )dx. Of course, these schematic
forms destroys all the geometric structure, which is absolutely essential for such
tasks as establishing monotonicity formulae, but for the purposes of analytical as-
pects of the Ricci flow, such as the local existence theory, these schematic forms
are often adequate.
Given two orthonormal tangent vectors X = X α , Y = Y β at a point x, the
sectional curvature K(X, Y ) of the tangent plane spanned by X and Y is given by
the formula
K(X, Y ) = Riemδαβγ X α Y β X γ Yδ .
One can show that this curvature depends only on the plane spanned by X and Y ,
indeed it can be interpreted as the Gauss curvature of this plane at x. Schematically,
K has the same units as the Ricci scalar: K = O(g −2 ∂ 2 g)+O(g −3 (∂g)(∂g)). Indeed
one can think of R (up to an absolute constant) as the expected sectional curvature
of a randomly selected 2-plane, whereas the Ricci bilinear form Ricαβ X α X β for a
unit tangent vector X is the expected sectional curvature of a randomly selected
2-plane passing through X.
The Laplace-Beltrami operator ∆g is defined on scalar functions f as
∆g f = g αβ ∂α ∂β f − g αβ Γγαβ ∂γ f
or schematically as
∆g f = O(g −1 ∂ 2 f ) + O(g −2 (∂g)(∂f )).
It is worth noting that ∆g has the same scaling as the Ricci scalar R. As such, the
Ricci scalar often naturally arises as a lower order correction term to the Laplacian.
This is for instance evident in the formula for the conformal Laplacian ∆g + 81 R
16Very informally, one can think of this dimensional analysis as simply counting the number of
times the symbols g and ∂ appear in every expression. Ricci flow will give ∂t the same dimension
as O(g −2 ∂ 2 ). See Table 1.
8 TERENCE TAO
from conformal geometry, although this Laplacian does not seem to directly play a
role in Perelman’s work17.
The Ricci flow, introduced for the study of topology of manifolds by Hamilton
[16], is a flow which (assuming a suitable local existence theory) produces a curve
t 7→ (Mt , gt ) of three-dimensional Riemannian manifolds for all time t in some time
interval. It is usually described by keeping the manifold Mt fixed (thus Mt = M ,
though perhaps Mt = {t} × M would be more appropriate) and varying the metric
g = gt via the partial differential equation
∂t gαβ = −2Ricαβ .
The factor 2 is a traditional and convenient normalisation which is not of essen-

tial importance, however the minus sign − is crucial in order for Ricci flow to be
well-posed in the forward time direction. Intuitively, Ricci flow expands (and flat-
tens) the negative curvature portion of the manifold while contracting the positive
curvature portion.
In order to continue Ricci flow past a singularity, it is often convenient to allow for
a more general coordinatisation of the flow, in which the three-dimensional “spatial”
manifolds Mt are viewed as level sets of a time function t on a four-dimensional
“spacetime” manifold (possibly containing some singularities or boundaries caused
by surgery), and the time derivative ∂t is replaced by a Lie derivative LT along some
time vector field T α (with the normalisation LT t = 1). See [29, §4], [27, §3.8, §14],
[22, §67], [5, §7.3] for a formalisation of these “generalised Ricci flows”, as well as the
related (but subtly different) “Ricci flows with surgery”. To simplify the exposition
we shall not discuss these flows (though they are of course essential in the full
proof of the Poincaré conjecture) and instead speak (somewhat imprecisely) solely
about ordinary Ricci flow (in which the manifold M , and in particular its topology,
remains invariant) together with isolated surgery operations. Nevertheless it is still
useful to keep in mind the idea that at two different times t1 , t2 , the manifolds Mt1
and Mt2 are really distinct manifolds, even if they can in principle be identified with
a single manifold M (at least if there is no surgery operation intervening between
t1 and t2 ). To emphasise this, we shall sometimes refer to a point x on Mt as (t, x)
rather than just x, thus we view points as elements of spacetime rather than merely
of space.
In order to avoid discussing a lot of technicalities regarding how various small
parameters depend on other small parameters (which are actually quite important
to the detailed analysis, but which are too distracting for a high-level overview) we
shall adopt the following rather imprecise notation. We use X = O(Y ), X . Y ,
or Y & X to denote the assertion that X is dominated by a bounded multiple of
Y ; this bound may in fact depend on a number of minor parameters but should
at least be independent of things like the natural spatial scale or the time until
blowup. Similarly we use X ∼ Y to denote X . Y . X.
17On the other hand, to verify various monotonicity formulae here one does need to apply the
Laplace-Beltrami operator to forms as well as to scalar fields, and the Ricci tensor again arises
as a lower order term in those cases. However as we are not pursing the structural geometric
identities in detail here we shall not go into these issues.
3. Dimensional analysis
In the large data global theory of nonlinear evolution equations, the role of scale
invariance, and the subsequent distinction of conserved or monotone quantities as
subcritical, critical, and supercritical, plays a fundamental role. Indeed, it is not
too much of an exaggeration to say that with current technology, a large data
global existence result is only possible either in the presence of a critical or sub-
critical controlled quantity (by which we mean a conserved or monotone quantity).
Furthermore, global asymptotics and blowup profiles are only possible in the pres-
ence18 of a critical controlled quantity, or the combination of a subcritical and a
supercritical controlled quantity.
In the context of Ricci flow, the scale invariance is given in coordinates by
t
(1) gαβ (t, x) 7→ λ2 gαβ ( 2 , x)
λ
where λ > 0 is arbitrary. Comparing this with the (local) scaling diffeomorphism
1 x
(2) gαβ (t, x) 7→ 2 gαβ (t, )
λ λ
(which is just a change of coordinates, and which therefore will preserve every
local 19 scalar geometric quantity), one can also write the scale invariance locally in
the form
t x
(3) gαβ (t, x) 7→ gαβ ( 2 , )
λ λ
which is the familiar parabolic scaling.
A critical geometric quantity is one which is invariant under the scale invariance
(being geometric, it does not matter which of the above two definitions of scaling one
chooses). Such quantities will be essential for understanding behaviour of the flow
near singularities; subcritical quantities would be preferable, but unfortunately turn
out to be insufficiently coercive to control the geometry by themselves, while super-
critical quantities (such as the total volume of the manifold) become increasingly
useless as one approaches the singularity. In parabolic theory, one can sometimes
compensate for the lack of critical controlled quantities by relying instead on the
maximum principle; this was for instance achieved by Hamilton [20] when estab-
lishing global existence (and absence of singularities) for two-dimensional Ricci flow
assuming that the Ricci curvature is initially negative definite. However it appears
that the maximum principle by itself is not strong enough to control the evolution
of Ricci flow in the case of indefinite curvature.
One quick informal way to determine the criticality of an expression is to perform
a dimensional analysis. It is convenient to do this by using two units, a unit L for
the spatial coordinate scale, and a unit G for the metric gαβ . This allows us to
18Furthermore, these quantities need to be coercive, in that bounds on such quantities severely
restrict the possible range of solution behaviours. For example, a quantity which could be bounded
due to a cancellation between a large positive and large negative term is unlikely to be coercive,
unless control from other sources can somehow eliminate this cancellation scenario.
19The key word here is “local”. A global quantity such as the total volume R dg is geometric
M
but not invariant under this scaling diffeomorphism unless one is willing to rescale the domain
of the coordinate chart(s) of M to be arbitrarily large or small. On the other hand, a localised
2
volume such as M e−distMt (x,y) /4(t0 −t) dg would be asymptotically invariant under (2) in the
R
limit where t0 − t is small, as this integral is effectively localised to a small, scale-invariant, region
of space.
10 TERENCE TAO
compute the dimension of every quantity in these notes in terms of G and L; see
Table 1.
L−2 L−1 L0 L1 L2 L3
∂t , ∆g , R, Rmin ,
G−1 γ̇ g −1
|Riem|g , |Ric|g , K
G−1/2 |γ̇|g
Ṽ , l,
G0 Riem, Ric, π ∂, Γ, X x, x0 , γ
κ, f, δ
dist, L, r,
G1/2
h, ρ
G1 ∂2g
R
∂g g t, t0 , τ, dτ
G3/2
R
Vol, dg
Table 1. The dimension of various quantities encountered in this
paper, as powers of G and L. Note that scalar geometric quantities
all lie on the diagonal where the power of L is twice that of G; this
is a manifestation of the dilation symmetry (2). Tensor quantities
typically do not lie on this diagonal, although their magnitudes do.
π here refers to the deformation tensor of a vector field X.
It is the power of G which determines the criticality, as can be seen from (1).
A net negative power of G implies that the quantity is sub-critical (control of this
quantity becomes increasingly powerful at fine scales and close to blowup times,
or equivalently when the metric becomes large), whereas a positive power of G
implies a super-critical quantity (and hence of little use near a blowup, unless one
has a favourable sign). A quantity with no powers of G is critical and thus a good
candidate for establishing uniform control on rescaled solutions near a singularity
(provided of course that it is sufficiently “coercive”, in that control on the quantity
implies significant control on the geometry).
The power of L is linked to that of G via the local scaling (2); as long as the
quantity is scalar and somehow “local” in nature, as well as being geometric, the
power of L must be twice the power of G. There is also a relationship for tensor
quantities depending on the number of subscripts and superscripts in the tensor
but it is slightly more tricky to state; it will be left to the reader (perhaps using
the above table as guidance).
The reader is also invited to verify the dimensional correctness of all the equations
in these notes.
4. Local existence
Given that we are trying to establish a large data final state result for a nonlinear
PDE, it is of course natural to begin with the local existence theory. Here we run
into an initial obstacle, which is that the Ricci flow
∂t gαβ = −2Ricαβ
does not obviously have the structure of a quasilinear parabolic equation. Indeed,
if we expand the Ricci tensor in coordinates we obtain
−2Ricαβ = g γδ (∂γδ gαβ + ∂αβ gγδ − ∂βγ gαδ − ∂αγ gβδ ) + O(g −2 (∂g)(∂g))
(4)
= O(g −1 ∂ 2 g) + O(g −2 (∂g)(∂g)).
Thus the Ricci flow equation takes the schematic form
∂t g = O(g −1 ∂ 2 g) + O(g −2 (∂g)(∂g))
where the lead term O(g −1 ∂ 2 g) is not elliptic. This is not a manifestly parabolic
equation.
Despite this, it is still possible to establish local existence for the Ricci flow for
suitably regular initial data. There are two approaches. The first approach, due to
Hamilton [16], uses the Nash-Moser iteration method to compensate for the lack of
smoothing present in the equation. The second approach, due to de Turck [14], is a
“gauge fixing” approach, taking advantage of the geometric nature of the Ricci flow,
which in practical terms creates the “gauge invariance” formed by diffeomorphisms
of the manifold (i.e. changes of coordinates). We shall briefly discuss the latter
method here.
Recall that every smooth vector field X α = X α (x) on a manifold (M, g) generates
an infinitesimal diffeomorphism xα 7→ xα + εX α + O(ε2 ). Pulling back the metric
gαβ by this diffeomorphism creates a infinitesimally deformed metric gαβ 7→ gαβ +
επαβ + O(ε2 ), where
παβ := LX gαβ = ∇α Xβ + ∇β Xα
is the deformation tensor of the vector field X. Thus, for instance, if the deforma-
tion tensor vanishes, then X generates an infinitesimal isometry, and is known as
a Killing field.
The Ricci flow commutes with diffeomorphisms. Because of this, the Ricci flow
equation is gauge equivalent to the equation
∂t gαβ = −2Ricαβ + παβ
where παβ is the deformation tensor of an arbitrary (time-dependent) vector field
X. Indeed, the solution to this gauge transformed Ricci flow is (formally) equal to
the solution to the original Ricci flow with the same initial data, composed with
the diffeomorphism obtained by integrating the vector field X; we omit the details.
In practical terms (and ignoring some technical issues in justifying uniqueness, pre-
serving regularity, etc.) this means that we have the freedom to add any expression
of the form παβ = ∇α Xβ + ∇β Xα to the right-hand side of the Ricci flow equation.
Note from the form of the Christoffel symbols Γ = O(g −1 ∂g) that
παβ = ∂α Xβ + ∂β Xα + O(g −1 (∂g)X).
In particular if one takes20
1
Xα := g γδ (∂γ gαδ − ∂α gγδ ) = O(g −1 ∂g)
2
then we see that
παβ = g γδ (−∂αβ gγδ + ∂βγ gαδ + ∂αγ gβδ ) + O(g −2 (∂g)(∂g)).
20This is not a completely arbitrary choice of infinitesimal diffeomorphism; it turns out that
the one-parameter family of diffeomorphisms generated by this recipe is also the harmonic map
heat flow deformation of the identity map from M to itself. See [24] for further discussion.
12 TERENCE TAO
Thus this gives a gauge-transformed Ricci equation of the form

∂t gαβ = g γδ ∂γδ gαβ + O(g −2 (∂g)(∂g))
or in other words
∂t g = ∆g g + O(g −2 (∂g)(∂g)).
This is now more obviously a parabolic equation, being a nonlinear perturbation of
the heat equation. One can now use standard parabolic methods, such as iteration
methods, to construct21 local solutions.
It is worth remarking that while de Turck’s gauge transformed Ricci flow equation
is very convenient for establishing local existence, it appears that for the global
theory it is more useful to retain the original Ricci flow equation. One reason for
this is that the metric g, while of course essential in the local theory, is not really
the main actor in the global theory; it is other quantities derived from the metric,
such as the Ricci curvature tensor and the reduced length and volume, which play
a much more dominant role. The Ricci flow equation in its original form seems well
suited for studying these latter quantities.
On the other hand, there is a specific gauge transformed version of the Ricci flow
which is worth noting. Suppose that the vector field X is the negative gradient of
a potential function f : X α = −∂ α f . Then the Ricci flow equation transforms to
∂t gαβ = −2Ricαβ − 2Hess(f )αβ
where Hess(f )αβ = ∇α ∇β f is the Hessian of f . Now suppose that at time t = −1,
the manifold obeys the gradient shrinking soliton equation
1
(5) Ricαβ + Hess(f )αβ − gαβ = 0
2
then the Ricci flow extends backwards and forwards in time for all time t ∈ (−∞, 0),
with the solution at time t being diffeomorphic to the solution at time −1, with
metric scaled by −t. This type of solution is known as a gradient shrinking soliton,
and plays a key role in the analysis of blowup singularities, similar to the role played
by solitons in other equations (particularly equations with a critical conserved or
monotone quantity). There is also the gradient steady solitons, which solve the
equation
Ricαβ + Hess(f )αβ = 0
and which exist for all time and are constant in time up to diffeomorphism.
More refined analysis can give some explicit blowup criteria necessary for singu-
larities to form. For us, the most important one is that if the Ricci flow cannot be
continued past a certain time, then the Riemann curvature must be blowing up to
infinity at that time; see [33] or [12, Theorem 6.3]. Thus singularities of the Ricci
flow will be preceded by the appearance of regions of arbitrarily large curvature.
The above local existence and uniqueness theory is not only valid for compact
manifolds, but extends to non-compact manifolds of bounded curvature; see [10].
21One cautionary note regarding the use of gauge transforms is that uniqueness of the gauge-
transformed equation does not always imply uniqueness of the original equation, because of the
possible presence of exotic solutions for which the gauge transform is unavailable due to lack of
regularity. Nevertheless there is a satisfactory uniqueness theory for Ricci flow, see e.g. [10].
However one can still make the proof of the Poincaré conjecture work even if uniqueness was not
known; it makes a few minor facts harder to prove (such as the fact that spherically symmetric
data leads to spherically symmetric solutions) but is otherwise not important to the argument.
This turns out to be important later on when we construct “standard solutions”

which are needed to replace singularities in the surgery procedure.
5. Monotonicity formulae I: Lower bounds on curvature

Perhaps the primary tools for obtaining long-time control of parabolic equations
are monotonicity formulae22.
These formulae arise from a number of sources. Prior to Perelman’s work, it
was already known that maximum principles (including the classical scalar maxi-
mum principle, but also the more advanced tensor maximum principle of Hamilton)
yielded some very useful monotonicity formulas which controlled the curvature from
below, but unfortunately not from above - so the monotone quantities were not fully
coercive.
For instance, the minimal scalar curvature Rmin (t) := inf x∈Mt R(t, x) was a non-
decreasing function of t; this follows easily from the observation that this scalar
curvature obeys a nonlinear heat equation23
2
(6) ∂t R = ∆g R + 2|Ric|2g ≥ ∆g R + R2
3
which when combined with the standard maximum principle for scalar parabolic
equations yields the useful monotonicity formula
2 2
(7) ∂t Rmin ≥ R
3 min
If Rmin was initially positive (i.e. the manifold had strictly positive scalar curvature
to begin with), this monotonicity formula implies finite time blowup of Ricci flow,
a fact first observed by Hamilton [16]. When Rmin is negative, the monotonicity
formula is not quite as powerful, but is still useful as one approaches a blowup
singularity, because the Rmin is subcritical with respect to scaling. Indeed it is
then easy to see that if one rescales a Ricci flow around its blowup time, then the
scalar curvature becomes asymptotically non-negative. This is a weak version of
the Hamilton-Ivey pinching phenomenon, of which more shall be said shortly.
A variant of Rmin which is also monotone non-decreasing is the least eigenvalue of
−4∆g + R. In fact Ricci flow can be viewed as a gradient flow for this quantity: see
[28, §1,2], [29, §8], [22, §5-9], or [5, §1.5]. However we do not need
R this monotonicity
formula here. We also mention Hamilton’s entropy functional M R log R dg, which
is monotone in the case of positive curvature [16], although again this will not be
needed here.
The Ricci tensor Ric and Riemann tensor Riem also obey equations similar to
(6), provided one works in a suitably chosen orthonormal frame. Indeed one has
equations which essentially take the form
(8) ∂t Riem = ∆g Riem + O(Riem2 )
22This is in contrast with Hamiltonian equations such as wave or Schrödinger equations, in

which the primary tool is conservation laws. Indeed the only important conserved quantity of Ricci
flow is a discrete quantity, namely the topology of the underlying manifold (though of course this
conservation can be broken during surgery). Of course, without this conserved quantity the whole
program of trying to extract topological information from Ricci flow would be doomed from the
start!
23For proofs of the identities in this section, see e.g. [12], [5].
14 TERENCE TAO
in the orthonormal frame coordinates24. These equations can be combined with a

tensor version of the maximum principle to obtain some uniform lower or upper
bounds on the Riemann tensor, similar to those obtained for the Ricci scalar. Some
simple examples of this are the following:
• If the scalar curvature of a Ricci flow is non-negative at the initial time,
then it remains non-negative at all later times. (This is a special case of
the monotonicity of Rmin .)
• If the Ricci tensor of a Ricci flow is non-negative definite at the initial time,
then it remains non-negative definite at all later times.
• If the sectional curvatures of a Ricci flow are all non-negative25 at the initial
time, then they remain non-negative at all later times.
See for instance [27, §4] or [5, §2] for a complete discussion of these results,
which essentially are due to Hamilton [21]. An important variant of the second
fact is Hamilton’s rounding theorem [16], which asserts that if the Ricci curvature
is initially strictly positive, then not only does the Ricci flow blow up in finite time,
the metric also becomes asymptotically spherical as one approached the blowup
time, thus establishing the Poincaré conjecture in the positive Ricci curvature case.
There are more quantitative versions of these results based on understanding
the dynamics of the lowest eigenvalue of the Ricci tensor or Riemann tensor. They
roughly have the following form: if these tensors are nearly non-negative (so that
the lowest eigenvalue is not too negative, relative to the size of the tensor) at the
initial time, then they remain nearly non-negative at later times. In particular, even
if the curvature is blowing up in magnitude, the negative portion of the curvature
will not blow up as fast; this phenomenon is known as Hamilton-Ivey pinching
towards positive curvature. One (informal) statement of this pinching is a lower
bound on sectional curvatures K = K(X, Y ) which roughly takes the following
form: if it is known that K ≥ −O(1) at time zero, then for later times t (which are
not too large) one has a bound of the form
(9) K ≥ −o(1 + |R|)
where o(1 + |R|) denotes some quantity which goes to infinity slower than 1 + |R| as
|R| → ∞ (note that monotonicity of Rmin ensures that R ≥ −O(1)). This pinching
also lets one deduce an upper bound
|Riem|g , |Ric|g = O(1 + |R|),
because pinching prevents the scenario where large positive and large negative cur-
vature cancel to create a small scalar curvature. Thus the scalar curvature R in
fact controls in magnitude all the other curvature components, and is particularly
effective at controlling the negative components of curvature.
This pinching statement was proven in [21]. Roughly speaking, it is proven by
showing that for a carefully selected (and mildly time-dependent) choice of o(1+|R|)
expression in (9), the set of points obeying (9) enjoys a certain convexity-in-time
property; the hypotheses on initial conditions combined with a continuity argument
24The selection of such a frame modifies the dimensional analysis, basically by setting G = 1,
thus collapsing each of the columns of Table 1 to a single entry.
25An extremely minor remark: it would be better to replace the non-negativity of sectional
curvatures by the non-negativity of the Riemann curvature tensor, as then this claim extends
to all dimensions. However for three dimensions the two statements are equivalent, due to the
identification between two-forms and (scalar multiples of) planes in this setting.
then give the result. As a very crude first approximation, the pinching phenomenon
means that in high curvature regions, almost all the negative curvature has been
eliminated, leaving only non-negatively curved geometries to consider.
One common application of maximum principles is to show that vanishing of
a non-negative elliptic or parabolic object at one point implies vanishing at other
points. For instance, the classical maximum principle for harmonic functions implies
that if a non-negative harmonic function on a domain vanishes at an interior point,
then it vanishes everywhere. Hamilton’s tensor maximum principle implies a result
of a similar flavour: if a three-dimensional Ricci flow with non-negative sectional
curvatures has a vanishing sectional curvature at one point, then (locally, at least)
the flow must split into the product of a two-dimensional Ricci flow, and a flat line
R. This type of splitting result is used heavily in the classification of various types
of limiting Ricci solution later on in the argument; see [27, §4.4], [22, §40] or [5,
§2.2].
Another naturalR quantity which is controlled (though not monotone) is the total
volume Vol(t) := Mt dg(t). Indeed, using the the formula for the rate of growth
of the volume form under Ricci flow, followed by monotonicity of Rmin ,
∂t dg = −Rdg ≤ −Rmin(t)dg ≤ −Rmin (0)dg,
we see from Gronwall’s inequality that the volume of the manifold can grow at most
exponentially in time, with the rate determined by initial minimal scalar curvature
Rmin (0). Unfortunately the volume is a supercritical quantity and is not directly
useful for blowup analysis, although we will need this volume control later in order
to show that the blowup times are discrete (the key point being that each surgery
removes a significant portion of volume from the manifold).
If the sectional curvatures are all non-negative, then much stronger monotonicity
formulae are possible. For instance, the above analysis shows that the volume is
now monotone decreasing; some further monotonicity results of this nature are dis-
cussed in the next section. Furthermore, there are stronger pointwise monotonicity
formulae for the Ricci scalar R, arising from Hamilton’s Harnack inequality [17]
(inspired by an earlier formula of Li and Yau [23] for a different equation). There
are several formulations of this inequality. One is the following: if the Ricci flow
has non-negative bounded sectional curvature on [0, t], then one has the pointwise
monotonicity-type formula for the scalar curvature
1
∂t R + R + 2∂ α RVα + 2Ricαβ V α V β ≥ 0
t
for all vector fields V ; this inequality can be obtained after some effort from many
geometric identities and the maximum principle. If 0 < t1 < t2 , one can then
quickly deduce the pointwise inequality
t1 2
R(x2 , t2 ) ≥ e−dMt1 (x1 ,x2 ) /2(t2 −t1 ) R(x1 , t1 )
t2
which may help explain the terminology “Harnack inequality”. In particular, for an-
cient solutions (extending to arbitrarily negative times) with non-negative bounded
sectional curvature, then R(x, t) is a non-decreasing function of t, thus we have
pointwise monotonicity of the scalar curvature in this case. See [28, §7,9], [27,
§4.5.2], [22, Appendix F], or [5, §2.5] for details.
The above consequences of the maximum principle are quite robust. For instance,
they all extend (with minor modifications) to general dimensions. Furthermore,
16 TERENCE TAO
while many of the stated results and arguments are initially only valid for compact
manifolds, there are a number of ways to extend them to non-compact manifolds
given a suitable control on the geometry at infinity. We shall gloss over the technical
details (usually involving suitable weight functions to localise arguments to near a
compact set) in performing such extensions, but see [34] or [27, §12.2] for examples.
6. Monotonicity formulae II: minimal surfaces and finite time

extinction
In the previous section we saw how pointwise quantities such as curvature enjoy
monotonicity properties, largely thanks to the maximum principle. In geometry,
another fertile source of monotone (or at least interesting) quantities arises not
from points in the manifold, but from embedded curves and surfaces, particularly
those which are chosen to minimise some geometric functional such as length, area,
or energy. It turns out that the Ricci flow tends to contract26 these sorts of quanti-
ties, which leads to a useful way to force finite time blowup (an important aspect of
Theorem 1.2). The intuition seems to be that the Ricci flow “wants” to compress
all of the non-trivial higher-dimensional topology of the manifold (e.g. the homo-
topy groups π2 (M ) and π3 (M )) into nothingness. This compression effect can be
temporarily counteracted by large negative curvature, which can act to expand the
metric, but the pinching phenomenon described in the previous section eventually
weakens the negative curvature components of the geometry to the point where the
compression becomes unstoppable27.
In this section we discuss some monotone (or at least controlled) quantities
associated to appropriately minimising one-dimensional manifolds (curves) and two-
dimensional manifolds (surfaces). The quantity Rmin in the preceding section can
be
R thought of as a zero-dimensional case of these scheme, while the total volume
28
M
dg is a three-dimensional case .
To give an example of this compression effect for one-dimensional minimal man-
ifolds (i.e. geodesics), let us observe the following simple but useful fact:
Proposition 6.1 (Non-negative curvature implies distance contraction). Suppose
that a Ricci flow t 7→ Mt has non-negative sectional curvature. Then for fixed
x, y ∈ M , the distance function dt (x, y) is a non-increasing function of t.
Proof. The first variation formula for metrics, combined with the Ricci flow equa-
tion, yields
Z 1
1 2
∂t dt (x, y) = − Ricαβ (γ(s))γ̇ α (s)γ̇ β (s)ds
2 0
for some energy-minimising curve γ : [0, 1] → M from x to y. The claim follows.
26Strictly speaking, these this contraction phenomenon only starts kicking in once the minimal
Ricci curvature Rmin stops being too negative, so this is only a conditional monotonicity formula.
However, by combining these formulae with the monotonicity formula (7) for Rmin we can still
extract finite time blowup results.
27Actually, this is only true for the higher-dimensional topology. For the first homotopy group
π1 (M ), it is possible that the negative curvature never weakens to the point where compression
occurs; this can be seen for instance by looking at a constant negative curvature manifold (which
can have non-trivial π1 (M ) but not π2 (M ) or π3 (M )), in which Ricci flow expands the manifold
indefinitely. We thank John Morgan for this example.
28Perelman’s monotone quantities, discussed in the next section, have an interpretation which
resembles an “infinite-dimensional” quantity, though it is unclear to me how they exactly fit into
the energy-minimising philosophy.
Clearly one can extend and refine this type of estimate; see for instance [28, §8],
[27, §3.6], [5, §3.4], or [22, §26] for versions of the above compression phenomenon
of relevance to Perelman’s argument.
Using Proposition 6.1 one could then create a monotonic decreasing functional
in the non-negatively curved case, namely the minimal length of a non-contractible
closed loop. Unfortunately, this quantity is only non-zero when the homotopy group
π1 (M ) is non-trivial - and for the purposes of proving the Poincaré conjecture, this
is certainly not the case! However, let us note that this result will yield a finite time
extinction result for non-simply connected manifolds with strictly positive curvature
(though this was already known to Hamilton by other means). Thus we must move
to surfaces (such as embedded two-spheres or two-tori) to get some interesting
geometric quantities. Experience has shown that surfaces which minimise area or
energy29 are likely be of great use in geometry; to pick just one example, witness
the use of minimal surfaces in Gromov’s ground-breaking paper [15] on symplectic
non-squeezing. Actually, it turns out that minimising over surfaces only is able to
exploit π2 (M ) information; it lets one reduce to the case where π2 (M ) is trivial,
but this turns out to not be enough. One may think then to move to embedded
three-spheres to exploit π3 (M ) information, but this turns out to be difficult to do
directly. Fortunately, we can view embedded three-spheres as a loop of two-spheres,
or as a two-sphere of loops, and by taking an appropriate min-max over the two
parameters one obtains a useful monotone quantity that allows one to exploit π3 (M )
information, which turns out to be sufficient to force finite time extinction. The
loop-of-two-spheres approach was pursued by Colding and Minicozzi [13], whereas
the two-sphere-of-loops approach was pursued by Perelman [30]. We shall discuss
the former which is slightly simpler conceptually, even though for technical reasons
the proof of Poincaré’s conjecture turns out to be more easily pursued using the
latter; see [27, §18] for details30.
We remark that the idea of using minimal surface areas as controlled quantities
under Ricci flow originates from Hamilton [21].
Let us first discuss the π2 (M )-based quantities. Suppose that π2 (Mt ) is non-
trivial. Let β : S 2 → Mt be any immersed R sphere not homotopic to a point. Each
such sphere has an energy E(β, t) := 12 S 2 |dβ|2gt using the metric gt at time t; this
is at least as large as the area of β. Define W2 (t) to be the infimum of E(β, t) over
all such β. It turns out (from standard Sacks-Uhlenbeck minimal surface theory
[32]) that this infimum is actually attained (at which point the energy is the same
as the area). A computation (using the local structure of minimal surfaces, and the
Gauss-Bonnet formula) lets one obtain the differential inequality
1
∂t W2 (t) ≤ −4π − Rmin (t)W2 (t)
2
(where the derivative is interpreted in a suitably weak sense) whenever the Ricci
flow remains smooth. Combining this with (7) and some elementary Gronwall-type
29Area and energy of surfaces are closely related, but not quite the same. A surface β : R2 → M
has area R2 |∂x β ∧ ∂y β|g dxdy and energy 21 R2 |∂x β|2g + |∂y β|2g dxdy. By Cauchy-Schwarz, the
R R
area is thus less than the energy, with equality when β is conformal. In practice, surfaces such as
spheres can usually be made conformal after reparameterisation, and so minimising the area and
minimising the energy usually end up amounting to the same thing.
30The papers [5], [22] do not address the finite time extinction issue as they are following
Perelman’s proof of the geometrisation conjecture, which does not require such extinction.
18 TERENCE TAO
analysis one can show that W2 (t) becomes negative in finite time, which is absurd.
This forces blowup in finite time whenever π2 (M ) is non-trivial. It turns out that
a similar argument can also be extended for Ricci flow with surgery. The final
conclusion is that after finite time, the Ricci flow with surgery is either completely
extinct, or at every time each component of M trivial π2 .
A variant of this argument works for π3 (M ). If π3 (M ) is non-trivial, then there
exists a loop of two-spheres γ : [0, 1] × S 2 → M with γ(0) = γ(1) equal to a
point, which cannot be contracted to a point. For any such γ, define E(γ, t) :=
sup0≤s≤1 E(γ(s), t) be the largest energy obtained by a two-sphere on the loop, and
then let W3 (t) be the infimum of E(γ, t) over all such γ. A variant of the above
argument then gives31
1
∂t W3 (t) ≤ −4π − Rmin (t)W3 (t)
2
so once again we can demonstrate finite time blowup as long as π3 (M ) is non-trivial.
Again, it is possible (given sufficient care) to extend these arguments to the case of
Ricci flow with surgery.
There is the issue of π3 (M ) being non-trivial. Fortunately some homology theory
settles this fairly readily, at least in the simply connected case. If M is simply
connected, then π1 (M ) is trivial, which implies (by the Hurewicz theorem) that
the first homology group H1 (M ) is also trivial. Poincaré duality then gives that
H2 (M ) is also trivial, whereas the three-dimensionality32 of M gives that H3 (M ) =
Z. In short, M has the same homology groups as the three-sphere, and thus
by Whitehead’s theorem it is therefore homotopy equivalent to a sphere33. In
particular34 it quickly follows that π3 (M ) is non-trivial (indeed, it is isomorphic to
Z). The more general case when π1 (M ) is a free product of finite groups and infinite
cyclic groups can be handled by more sophisticated topological and homological
arguments of the above type; see [27, §18] for details.
As mentioned earlier, the Colding-Minicozzi approach requires a technical an-
alytical property which is expected to be true but whose proof is not currently
in the literature. There is an alternate approach of Perelman [30] for which the
corresponding analytical property is easier to verify (because one works with area-
minimising surfaces instead of arbitrary surfaces). Instead of viewing embedded
three-spheres γ in M as loops of two-spheres, one instead views them as two-spheres
of loops (with some natural boundary conditions). Each loop spans a minimal disk
(because of simple connectedness), which of course has an area (or energy). Taking
the maximal area obtained in this way along a two-sphere of loops, one obtains
31This is assuming some analytical facts concerning near-extremisers of W (t) which are widely
3
expected to be true, but have not yet been explicitly proven. More precisely, one needs a bubbling-
type concentration compactness property for minimising sequences of loops of spheres, which is
likely to be true (based on analogy with similar results for simpler variational problems). Never-
theless it is believed that this property will be established soon. We thank Gang Tian and John
Morgan for clarifying these points.
32Note that simply connected manifolds are necessarily orientable.
33Note that this is a substantially weaker statement than being homeomorphic to a sphere;
indeed, since non-trivial homology spheres are known to exist, the Poincaré conjecture cannot be
proven by homology theory alone!
34Here we are using the fact that the surgery procedure cannot destroy the simply connected
nature of Mt (in fact the fundamental groups of the post-surgery components can easily be shown
to be subgroups of the pre-surgery fundamental groups, which are trivial in the simply connected
case).
an energy Ẽ(γ, t) analogous to the energy considered earlier; maximising over t,

and then minimising over all non-contractible γ one obtains an analogue W̃3 (t) to
W3 (t), which ends up obeying a very similar differential inequality35 to W3 (t). For
full details see [27, §18].
In summary, modulo some important technical details (such as how to deal with
surgery), we have used the compression of minimal-area surfaces phenomenon to ob-
tain the finite time extinction property for Ricci flow-with-surgery. In the bounded
amount of time remaining to the flow, we need to control the singularities (or near-
singularities) well enough that we can correctly define a surgery procedure which
(a) does not disrupt other aspects of the argument, such as the compression phe-
nomenon, and (b) modifies the topology of the manifold in a controllable manner.
These (highly non-trivial) tasks will be the focus of the remaining sections of this
note.
7. Monotonicity formulae III: Perelman’s reduced volume

One of Perelman’s major advances in the subject was to introduce for the first
time in [28, §3,7] several useful critical and coercive monotone quantities for Ricci
flow, most notably36 a reduced volume Ṽ which is a “renormalized” version of the
total volume. This quantity is not only scale-invariant, but is also monotone under
Ricci flow with no assumptions on the initial signature of the Ricci curvature.
This is an absolutely essential breakthrough if one is to analyse behaviour near a
singularity, or more generally near a high curvature region of the manifold.
Now we discuss the reduced volume in more depth37. This quantity is consistent
with the standard “heat ball” method for generating monotonicity formulae for
geometric parabolic equations. Note that for a solution to the free heat equation
ut = ∆u in a Euclidean space38 R3 , and any point (t0 , x0 ) in spacetime, the quantity
1
Z
2
(10) e−distR3 (x,x0 ) /4τ
u(t0 − τ, x) dx
(4πτ )3/2 R3
is constant in time for τ > 0; this is clear from the fundamental solution of the
heat equation. It turns out that when one replaces the free heat equations by
other parabolic equations, that similar quantities to (10) can often turn out to be
monotone in time rather than constant. As a gross caricature, one can think of (10)
as essentially measuring an average of u(t0 − τ ) on a ball centred at x0 of radius
35A significant technical issue arises, which is that to demonstrate the above inequality it
becomes necessary to evolve the loops involved by curve shortening flow. This flow can become
singular even when the Ricci flow is smooth, requiring a (standard) lift of the problem to higher
dimensions to resolve the singularities. See [30], [27, §18.4-18.6] for details.
36Perelman also introduces a number of other interesting critical monotone quantities, such
as a modified entropy functional W. While these quantities are also very useful, particularly for
the geometrisation conjecture, they are not strictly necessary for Theorem 1.2 and will not be
discussed here.
37Details on the reduced volume can be found in [28, §7], [27, §6], [22, §14-23], or [5, §3].
38Actually, the monotonicity formulae given here are valid (and scale-invariant) in all spatial
dimensions, thus raising the possibility that Perelman’s methods should extend to analyse Ricci
flow on manifolds in every dimension; see for instance [10] for some preliminary progress in this
direction. However, at present several other aspects of Perelman’s work is restricted to three
spatial dimensions.
20 TERENCE TAO
√
O( τ ). Observe also that if we normalise (t0 , x0 ) to the origin, then (10) is also
invariant under the scaling u(t, x) 7→ u( λt2 , λx ), which is the analogue to (2).
Let (t0 , x0 ) be a point in spacetime of a solution M to Ricci flow in d spatial
dimensions, thus x0 lies on the 3-dimensional Riemmanian manifold (Mt0 , gt0 ).
Based on the above discussion, one might naively expect the expression
1
Z
−distMt −τ (x,x0 )2 /4τ
(11) 3/2
e 0 dgt0 −τ (x)
(4πτ ) Mt0 −τ
to have some monotonicity

√ in time; this is basically measuring the volume of the
ball of radius O( τ ) centred at x0 at time t0 − τ , divided by the volume of the
corresponding Euclidean ball. Thus this quantity will be expected to equal one
when M is flat, be less than one when M has positive curvature, and greater than
one when M has negative curvature. Note also when we normalise (t0 , x0 ) to be the
origin, then the quantity distMt0 −τ (x, x0 )2 /4τ is dimensionless (scale-invariant), as
is the quantity (11).
Unfortunately the above quantity is not quite monotone. Geometrically, this
can be explained by the fact that x0 “lives” on the manifold Mt0 at time t0 , and
really shouldn’t be used to measure distances on the manifold Mt0 −τ at time t0 − τ .
Remarkably, however, it turns out that a slight modification of the above functional
is monotone non-increasing (as a function of τ ) for smooth Ricci flow as τ → 0+ ,
namely the reduced volume 39
1
Z
Ṽ (τ ) = Ṽt0 ,x0 (τ ) := e−lt0 ,x0 (t0 −τ,x) dgt0 −τ (x)
(4πτ )3/2 Mt0 −τ
where the reduced length lt0 ,x0 (t0 − τ, x) is a variant of the naive Gaussian distance
distMt0 −τ (x, x0 )2 /4τ , defined as follows. Observe from the definition of geodesic
distance and elementary calculus of variations that
distMt0 −τ (x, x0 )2 √
Z τ
1
= inf √ |γ̇(τ ′ )|2gt0 −τ τ ′ dτ ′
4τ γ 2 τ 0
where γ : [0, τ ] → Mt0 −τ runs over all smooth curves with γ(0) = x0 and γ(τ ) = x.
(Indeed, the minimiser is attained
q when γ traverses the geodesic from x0 to x,
which at time τ ′ is at distance ττ distMt0 −τ (x, x0 ) from x0 .) As mentioned before,
′
it is not geometrically natural for x0 to be placed in Mt0 −τ when it really should

lie in Mt0 ; similarly, it is not natural to γ to be completely contained in Mt0 −τ ,
and instead it should be a curve from (t0 , x0 ) to (t0 − τ, x). This (at least partially)
motivates Perelman’s definition of reduced length, namely
1
lt0 ,x0 (t0 − τ, x) := √ Lt0 ,x0 (t0 − τ, x)
2 τ
and L is the length
Z τ √
(12) Lt0 ,x0 (t0 − τ, x) = inf [|γ̇(τ ′ )|2gt ′
+ R(t0 − τ ′ , γ(τ ′ ))] τ ′ dτ ′
γ 0 0 −τ
where γ : [0, τ ] ∈ M ranges over all smooth curves with γ(τ ′ ) ∈ Mt0 −τ ′ for 0 ≤
τ ′ ≤ τ , with γ(0) = x0 = (t0 , x0 ) and γ(τ ) = x = (t0 − τ, x), and R is the Ricci
scalar at the point (t0 −τ ′ , γ(τ ′ ) ∈ Mt0 −τ ′ . This type of “parabolic length” had been
39This definition differs from the one by Perelman by a harmless factor of (4π)3/2 ; I have
chosen this normalisation to give the flat metric a reduced volume of 1.
considered earlier in a different context than Ricci flow by Li-Yau [23]. The presence
of the Ricci scalar appears to be some sort of correction term, similar in spirit to
the Ricci curvature terms arising in other geometric formulae such as the second
variation formula for geodesic distance, the Bochner-Wietzenböck formula for the
Laplacian of energy densities, or the presence of the Ricci scalar in the conformal
Laplacian. See [28, §7], [5, §2.5] for some motivation of this length functional arising
naturally from a Li-Yau type inequality for the backwards heat equation.
One can easily verify that lt0 ,x0 (t0 −τ, x) is dimensionless, and hence the reduced
volume Ṽ (τ ) is also scale-invariant. Thus the monotonicity of this quantity is thus
a critical control on the geometry as one approaches a blowup point (t0 , x0 ); very
informally speaking, it asserts that as one approaches such a singularity, the net
curvature tends to increase, thus there is a trend from negative curvature to positive.
In some ways this can be viewed as a local version of the monotonicity formula for
Rmin .
The proof of this monotonicity is remarkably delicate; see [28, §7], [22, §22], [27,
§6.7], or [5, §3.2] for the details. In principle, one simply needs to differentiate the
reduced volume in time, use variation formulae to compute various derivatives of
the reduced length, and integrate by parts, using many identities from Riemannian
geometry, until one visibly sees the sign of the derivative. In practice, of course,
such a computation is very lengthy and seems highly miraculous in the absence
of geometric intuition. Since I do not have this intuition, I cannot really shed any
light as to why this monotonicity formula works (other than by analogy with simpler
equations). It does however appear to be instructive to show that the monotone
quantity is in fact constant in the case of a gradient shrinking soliton. In [28, §6],
[5, §3.1] an interpretation of reduced volume in terms of infinite dimensional Rie-
mannian geometry was given, relating this quantity to a matrix Harnack expression
considered by Hamilton [18]. A related (but not identical) geometric interpretation
for this expression also appears in [11].
Comparing the reduced volume against the informal interpretation for (11) leads
to the heuristic
(13) Ṽt0 ,x0 (τ ) ≈ τ −3/2 VolMt0 −τ (B(x0 , τ 1/2 )).
This heuristic turns out to be essentially accurate in the case where the ball
B(x0 , τ 1/2 ) has bounded normalised curvature on the time interval [t0 −τ, t0 ], in the
sense that |Riem|g = O(τ −1 ) (this is the natural amount of curvature predicted by
dimensional analysis, and is also consistent with (8)). Verifying this is not trivial,
but roughly speaking proceeds by exploiting the curvature control to ensure the
reduced length lt0 ,x0 (t0 − τ, x) stays bounded on B(x0 , τ 1/2 ), which in turn leads
to an estimate similar to (13).
A more refined analysis of the proof of the monotonicity formula also reveals
the precise solutions for which equality occurs, i.e. the reduced volume is constant
rather than monotone. Indeed, this occurs precisely when the solution is a gradient
shrinking soliton, as in (5) (if we normalise t0 = 0 and t = −1). This strongly
suggests (but does not formally imply) that as one rescales around a singularity
(t0 , x0 ) of a Ricci flow, that the flow should increasingly resemble a gradient shrink-
ing soliton (possibly after passing to a subsequence of rescalings). A result of this
22 TERENCE TAO
form will be essential later in the argument. Details can be found40 in [28, §7], [27,
§9.2], [22, §36], [5, §3].
We have normalised the reduced volume to equal 1 when the manifold is flat. If
the Ricci flow has bounded curvature and is smooth at the final point (t0 , x0 ), then
its rescalings become flat, and so the reduced volume is asymptotically 1 as t → t0 .
By monotonicity this means that the reduced volume is less than or equal to 1 at
all times t < t0 . By analysing when equality occurs in the monotonicity formula as
in the previous paragraph, one can in fact conclude that the reduced volume will
be strictly less than 1 unless the entire flow was flat. See [27, §9.2] for details.
8. Blowup analysis I: a no-local-collapsing theorem

Armed with a critical monotone quantity centred at a spacetime point of our
choosing, we can now hope to obtain control on the manifolds Mt of a Ricci flow41
as one approaches a blowup time, t → t∗ .
This will be done in several stages. The first major stage, which we cover in
this section, involves the relationship between the “local” geometry of the manifold
Mt , which is given by pointwise quantities such as the Riemann curvature tensor
Riem (and its components such as the Ricci tensor Ric and Ricci scalar R), and
more “global” geometry of the manifold, such as the injectivity radius or volume
of balls B(x, r) at a macroscopic radius r. The local quantities will be controllable
directly from equations arising from the Ricci flow such as (8), whereas it is the
global quantities which will be needed to control the asymptotic profile of blowup
solutions and perform surgery properly. Given an arbitrary smooth Riemannian
manifold there are some general partial relationships between the local and global
quantities. Let us first work at the macroscopic radius scale r = 1. Just to give
the flavour of things we shall use informal O() notation (thus suppressing the exact
dependence of various implicit constants on other implicit constants, which is an
important but technical detail) and not give the sharpest results:
• (Comparison theorem) If one has bounded curvature on a unit ball B(x0 , 1),
thus |Riem|g = O(1) on this ball, then for 0 ≤ r ≤ 1, the volume of B(x0 , r),
divided by r3 , is essentially non-increasing in r. In particular this implies
that B(x0 , 1) has volume O(1).
• (Reverse comparison theorem) If one has bounded curvature on a unit ball
B(x0 , 1), then the volume of the ball is ∼ 1 if and only if the injectivity
radius42 is & 1.
These results are facts from Bishop-Cheeger-Gromov comparison theory and rely
primarily on an analysis of the exponential map from x0 . The curvature bounds
ensure that this map has controlled derivatives, leaving the injectivity of the map
as only remaining hypothesis required to estimate the volume accurately. For the
40The argument in [28, §7] is slightly incorrect, however can be fixed in a number of ways; see
the other three references cited here.
41In order to fully prove Theorem 1.2, one has to consider generalised Ricci flows, in which
various surgeries take place between t = 0 and t = t∗ . This makes the analysis significantly more
complicated and non-trivial, but for sake of this high-level overview we shall focus on the much
simpler case of smooth Ricci flow (without surgeries).
42Roughly speaking, the injectivity radius at a point x is the radius of the largest ball B(x, r)
for which length-minimising geodesics are unique.
more precise versions needed for Perelman’s argument, see [31, §6.2], [27, §1.6-7],
or [8, §4.7].
Note that on a generic smooth Riemannian manifold it is possible to have col-
lapsing, in which the curvature is bounded but the volume of the ball is small (and
the injectivity radius is small too, by the reverse comparison theorem). A simple
example is given by an extremely small flat torus or flat43 cylinder, which has zero
curvature but for which the unit ball has unexpectedly small volume. Many more
exotic examples can of course be constructed. In contrast, in the non-collapsed
case, a famous finiteness theorem of Cheeger [6] limits the number of topologically
distinct such examples available. Thus we see that the non-collapsed property is
very powerful in constraining the behaviour of manifolds. As the small torus and
cylinder examples show, collapsing is associated with the presence of “little loops”
(i.e. short closed geodesics).
Of course one can rescale the above results to work at any radius scale r, to
obtain:
• (Comparison theorem) If one has bounded normalised curvature on a ball
B(x0 , r), thus |Riem|g = O(r−2 ) on this ball, then for 0 ≤ r′ ≤ r, the
volume of B(x0 , r′ ), divided by (r′ )3 , is essentially non-increasing in r′ . In
particular B(x0 , r) has volume O(r3 ).
• (Reverse comparison theorem) If one has bounded normalised curvature
on a ball B(x0 , r), then the volume of the ball is ∼ r3 if and only if the
injectivity radius is & r.
One consequence of the monotonicity aspect of the comparison theorem is that
if there is no collapsing at a scale r, then there is also no collapsing at any smaller
scale44.
A remarkable phenomenon concerning Ricci flow is that as one approaches √ a
blowup time t → t∗ , collapsing cannot actually √ occur at scales r = O( t ∗ − t)
which are finer than the natural length scale t∗ − t associated to the near-blowup
time t (this is the scale given by dimensional analysis). This no-local-collapsing
theorem was first established by Hamilton [20] in the case of non-negative sectional
curvature, and then in general by Perelman [28]. We state here a fairly weak and
imprecise version of this theorem (stronger versions are certainly available, see [28,
§4,7,8], [29, §7], [27, §8], [22, §12, 25, 27, 78], [5, §3.3, 3.4, 7.6]):
Theorem 8.1 (No local collapsing). [28] Let t 7→ Mt be a smooth Ricci flow for
0 ≤ t ≤ t0 = O(1), let x0 ∈ M , and suppose we have noncollapsed geometry at
time zero in the sense that the curvature is bounded by O(1) and the volume of any
unit ball is comparable to 1 at time zero. Then√we have local noncollapsing at time
t0 in the following sense: given any 0 < r < t0 , if one has bounded normalised
curvature on the parabolic cylinder [t0 − r2 , t0 ] × B(x0 , r), then the volume of the
ball B(x0 , r) at time t0 is comparable to r3 (or equivalently, the injectivity radius
of this ball at time t0 is comparable or larger to r).
43This is of course flatness in the intrinsic sense of Riemann, as opposed to any sort of extrinsic
sense; indeed, we are not actually embedding our manifolds into an ambient Riemannian space
with which to measure extrinsic curvatures such as second fundamental forms.
44It may be more intuitive to view this assertion in the contrapositive; if there is some break-
down of injectivity at a small scale caused by little loops, then this will replicate itself at larger
scales also, since little loops can be concatenated to form big loops.
24 TERENCE TAO
Proof. (Sketch) A rescaling√argument using the comparison theorem lets us reduce

to the case t0 ≥ 1 and r < t0 /2. Because curvature is bounded at time zero, one
can use (8) to show that it will stay bounded for short times 0 ≤ t ≪ 1. This will
give us enough control on the geometry to obtain a lower bound Ṽt0 ,x0 (0) & 1 for the
reduced volume at time 0. By monotonicity this implies Ṽt0 ,x0 (t0 − r2 ) & 1. Using
(13) (pretending for sake of argument that this heuristic is accurate; the actual
argument is more technical) and the comparison theorem we obtain the claim.
There are many sharper versions of this no local collapsing theorem in Perelman’s
work, and also in subsequent papers. In particular, in order to fully complete the
proof of Poincaré’s conjecture one needs to obtain a no local collapsing theorem
which is true for Ricci flow with surgery, and not just smooth Ricci flow; this turns
out to be significantly more complicated and delicate. But the flavour of all these
results remains unchanged: if there is no collapsing at initial times, then at later
times one still has no collapsing as long as one is working at local scales (less than
the natural length scale dictated by the time left until blowup) and the curvature
is controlled in a scale-invariant fashion.
9. Blowup analysis II: asymptotic profiles

One major reason for wanting a no-collapsing theorem is that it enables one
to exploit Hamilton’s compactness theorem [19]: a sequence of Ricci flows45 which
are uniformly non-collapsed with uniformly bounded normalised curvature at a
certain scale, have a convergent subsequence46, though the limit might be a non-
compact manifold. This is basically because the non-collapsing yields injectivity
radius lower bounds which can be used to give a common coordinate chart to all
the manifolds (cf. [7]); also some parabolic regularity estimates of Shi (see [33],
[24, Appendix B], [12]) give control on higher derivatives of curvature (essentially
obtained by differentiating equations such as (8) and then using the energy estimate
and Sobolev embedding) which are sufficiently strong to give compactness (cf. the
Arzela-Ascoli theorem). Such compactness, when combined with standard rescaling
arguments, allows one (at least in principle) to analyse the asymptotic behaviour
of singularities, which is an important milestone in the analysis of any nonlinear
PDE and in particular is essential for defining the surgery procedure for Ricci flow.
Let us first see how this would work in a best-case scenario. Suppose we have
a Ricci flow t 7→ Mt on [0, T∗ ) which blows up at a finite time T∗ . Since the
scalar curvature controls all the other curvatures due to pinching, we thus see that
there is a sequence tn of times going to T∗ and xn ∈ M such that the scalar
curvatures Rn := R(tn , xn ) go to infinity. We make the optimistic assumption47
45Strictly speaking, each Ricci flow needs to also come with an “origin” or “base point” to
allow one to properly define convergence; it suffices to have uniform bounds on the geometry on
an expanding sequence of balls around this origin, with radii going to infinity.
46We will be vague about exactly what “convergence” of a sequence of Ricci flows means, but
(due to the parabolic smoothing effects of Ricci flow) one can largely rely on smooth convergence
on compacta in the geometric sense, after picking a base point as mentioned earlier.
47Of course, by choosing x to be near the point where the maximum is attained, one can
n
obtain this assumption automatically. But to understand surgery, we will need eventually to
place xn next to where a specific singularity occurs; the problem is that as the curvature near
that singularity is becoming large, it may be that somewhere else there is another singularity
where the curvature is becoming even larger still. Fortunately this can be dealt with because the
effect of each high curvature region ends up being localised, though to make the entire argument
that at each time tn , the curvature Rn is close to the largest curvature that has yet
been encountered, thus R(t, x) . Rn for all 0 ≤ t ≤ tn . One consequence of this
−1/2
is that we automatically enjoy bounded normalised curvature at scales Rn and
below and all times up to tn . By the no local collapsing theorem we can then also
conclude that there is no collapsing at these scales or below up to time tn .
(n)
We can then create rescaled flows t 7→ Mt on the interval [−tn Rn , (T∗ − tn )Rn )
by setting
(n) (n) (n)
(Mt , gt ) := (Mtn +t/Rn , Rn gtn +t/Rn ).
These rescaled flows are also Ricci flows, and their backwards time of existence
extends to −∞ as n → ∞. Also, because of our rescaling, we now have bounded
normalised curvature and no collapsing at unit scales. Thus we may apply Hamil-
(n)
ton’s compactness theorem (declaring xn as the origin of each Mt ) and extract a
convergent subsequence whose limit is an ancient Ricci solution, i.e. one which has
existed on the entire interval (−∞, 0]. Furthermore this ancient solution is also non-
collapsed at unit scales and below; and in fact if the xn are all piled on top of each
other, one can compare the rescaled solutions to each other to obtain non-collapsing
at all scales for which the normalised curvature is under control. We caution that
even if the original manifold M is compact, it is perfectly possible for the limiting
ancient solution to be non-compact. Furthermore, the Hamilton-Ivey pinching phe-
nomenon ensures that the limiting ancient solution will have non-negative sectional
curvature everywhere.
Roughly speaking, what this analysis suggests is that every high-curvature region
of a Ricci flow should resemble a rescaled ancient non-collapsing solution to Ricci
flow. If we knew this, and we also knew enough about the geometry of ancient
non-collapsing solutions, then we could hope to perform a controllable surgery
on high-curvature regions in order to reduce their curvature while controlling the
change in topology (and not disrupting the finite time extinction argument which
will arise later). However, the above argument required that at any given time, the
high-curvature region one is studying has higher curvature than any other region
previously encountered. This is unfortunately not always the case (consider for
instance two simultaneous singularities, one with a faster curvature blowup rate
than the other). Fortunately, it is possible to induct on the curvature, starting
with the highest curvature region and controlling that, then moving onwards to the
next highest curvature region, and so forth. In order to make this induction work,
we cannot quite decouple the two major steps in the classification of curvature
regions, namely the approximation by ancient non-collapsing solutions and the
understanding of the geometry of these ancient solutions; they have to be applied
alternately because we need enough control on the very high curvature regions to
guarantee that they will not disrupt the analysis of the less high curvature regions.
For this reason it makes sense to talk about the ancient solutions first, before
explaining why they model singularities.
Interestingly, we will eventually have to perform a secondary blowup analysis
inside the first, in that the ancient solutions themselves are not the irreducible
components of asymptotic behaviour, and contain inside them an even more special
rigorous requires a slightly tricky downward induction on curvature; in other words, one needs to
control the very high curvature regions first before attending to the moderately high curvature
reasons. See the next few sections for more discussion.
26 TERENCE TAO
asymptotic class of solution, namely the gradient-shrinking solitons. The latter were
completely classified by Perelman [28, §11] (assuming, as we may, non-negative
sectional curvature and non-collapsedness), which leads to a qualitative description
of the ancient solutions, which in turn will lead to a qualitative description of
the high curvature regions. It seems that this two-step procedure is necessary to
fully understand high curvature regions. Consider for instance a high-curvature
region which resembles a rescaled cigar. Then while this cigar itself contains an
asymptotically cylindrical piece at infinity (and the cylinder is a gradient-shrinking
soliton), the compact end of this cigar does not have any model as a gradient-
shrinking soliton.
10. Blowup analysis III: the geometry of ancient non-collapsed

solutions
As discussed above, we shall skip for now the analysis of why high-curvature re-
gions resemble ancient non-collapsed solutions, and instead analyse these solutions
directly. Following Perelman[28, §11], [29, §1] (and [27, §9], [5, §6], [22, §37]), let us
define a48 κ-solution to be a non-collapsed ancient Ricci flow of non-negative sec-
tional curvature49, which is not flat; these will eventually be the solutions used to
model the high-curvature regions. The non-flatness will guarantee (in conjunction
with some maximum principle arguments) that the scalar curvature R is strictly
positive. There are three model classes of examples of κ-solutions (and some non-
examples) to keep in mind:
• (The sphere) For all −∞ < t < 0, let Mt be the standard sphere S 3
with metric gt = (1 − t)g0 , where g0 is the standard metric (rescaled to
have constant sectional curvature 1). This is an ancient solution (in fact,
a gradient shrinking soliton) starting with an infinitely large round sphere
shrinking down to the unit sphere; if one continued the Ricci flow to positive
times, then the sphere would shrink to a point at time one. More generally
one can replace the sphere S 3 by any other manifold of positive constant
sectional curvature (e.g. S 3 /Γ, where Γ is a finite rotation group which
acts freely on S 3 ).
• (The cylinder) For all −∞ < t < 0, let Mt be the standard cylinder S 2 × R
with metric gt = (1−t)g0 ⊕h, where g0 is the standard metric on S 2 (rescaled
to have constant sectional curvature 1/2) and h is the standard metric on
R. This is an ancient solution (in fact, a gradient shrinking soliton) starting
with an infinitely wide round cylinder shrinking down to a cylinder of unit
length; if one continued the Ricci flow to positive times, then the cylinder
would shrink to a line at time one. Note that if we made the cylinder
compact by replacing R by S 1 , then the solution is still ancient but has
become collapsed at large scales, because at large scales r ≫ 1 and at early
times t ≪ −r the volume of balls of radius r only grows like r2 instead of the
expected r3 , despite the fact that the curvature never exceeds O(1/r2 ) on
such balls. Thus we see that the crucial non-collapsing property obtained
48The name derives from a parameter κ used to quantify the exact nature of the non-collapsing,
but we are using the imprecise & notation to suppress this parameter.
49Strictly speaking, one also has to hypothesise that the scalar curvature is pointwise non-
decreasing, and that the curvature is bounded. But the former follows from Hamilton’s Harnack
inequality of Li-Yau type discussed in earlier sections, while the latter will be established shortly.
by Perelman can be used to significantly reduce the possible geometries

available for ancient solutions. It is also worth remarking that in two spatial
dimensions, the cylinder S 1 × R is also a gradient shrinking soliton and an
ancient solution, but is now collapsed at large scales due to the absence of
curvature in this case. (Also, we require κ-solutions to be non-flat.)
• (Cigar-type solitons) Let us temporarily work in two spatial dimensions
instead of three. Then, as first observed by Hamilton [17], there is an
ancient t 7→ Mt which at time t = 0 is the plane R2 with the metric dg02 =
dx2 +dy 2
1+x2 +y 2 (causing the plane to curve somewhat like one end of a cigar), and
which at negative times is a isometric to the same plane with metric (1−t)g0 .
This solution, known as the cigar soliton, is a gradient steady soliton with
non-negative curvature, however it is collapsed (it asymptotically resembles
the cylinder S 1 × R at infinity, and S 1 has zero curvature). Analogues of
the cigar soliton - which resemble S 2 × R at infinity, and which thus (by
the curvature of S 2 ) have enough curvature at infinity to be non-collapsed
- are known to exist in three dimensions, see e.g. [2], [3]. These objects
are also similar to the standard solutions used for the surgery procedure in
Section 14.
• (A non-example) Consider the product of a two-dimensional cigar soliton
and a line R. This is an ancient solution but is collapsed at large scales,
because the cigar is collapsed. It is vital that this example is ruled out from
the analysis, because a high-curvature region of Ricci flow which resembles
the product of a cigar and a line cannot be removed by surgery.
One of the key results of Perelman is that in fact all κ-solutions essentially
resemble one of the above examples in a certain quantitative sense, after rescaling
to normalise the scalar curvature R(x) to equal 1. More on this later.
The analysis of κ-solutions is rather lengthy and proceeds in several stages; we
now consider each stage in turn.
10.1. Bounded curvature. We shall begin with an important result on κ-solutions

due to Perelman [28, §11] (and elaborated in [27, §9.3], [22, §45], [5, §6.3]), namely
that all κ-solutions have bounded curvature at time t = 0 (and hence at all ear-
lier times, by pointwise monotonicity of the scalar curvature for ancient solutions).
This argument is specific to three spatial dimensions (a simpler version of it also
works in two dimensions).
The argument runs by contradiction; if one can find a sequence of points where
the curvature goes to infinity, then by looking at rescaled versions of the geometry
around these points (or more precisely, slight perturbations of these points where
the curvature is essentially a local maximum), and inspecting the rays connecting
these points to a fixed origin, one can arrive (after passing to a subsequence) to a
limit where one of the scalar curvatures has vanished; this leads to a factorisation
of the limit as the Riemannian product of a two-dimensional ancient Ricci flow and
a line. The two-dimensional ancient Ricci flows can be completely classified by a
variety of means; one can repeat the argument just stated to show that curvature
must be bounded, and in fact furthermore that the manifold is compact. Then the
rounding arguments of Hamilton show that the two-dimensional manifold must in
fact have constant curvature, so is basically a sphere. In short, the limit is basically
28 TERENCE TAO
a cylinder solution50. Undoing the scaling, we thus see that the original κ-solution
contains arbitrarily small regions which are approximate rescaled cylinders (the
precise term for this is an ε-neck, which we shall return to later). However the soul
theorem (see e.g. [31]) and the hypothesis of non-negative curvature can be used
to eliminate this possibility (see also [10] for an alternate argument).
10.2. Curvature decays spatially slower than scaling. Having just obtained
a uniform (in x) bound on the scalar curvature R(t, x), we now turn to a lower
bound in the asymptotic regime x → ∞ (of course this only makes sense for non-
compact solutions). Naive dimensional analysis (see Table 1) suggests that, for a
fixed origin x0 ∈ M , the scalar curvature R(t, x) should decay like 1/d(x, x0 )2 as
x → ∞. However, the curvature actually decays slower than this.
Proposition 10.3 (Infinite asymptotic scalar curvature ratio). [28, §11] If t 7→ Mt
is a non-compact κ-solution, then for each time t we have
lim sup R(t, x)d(x, x0 )2 = +∞
x→∞
for any x0 ∈ M (the exact choice of x0 is not relevant).
We remark that a slightly weaker version of this proposition was established
earlier by Hamilton [20].
Proof. (Sketch; for details see [27, §9.6], [22, §40], [5, §6.4]) Let us first rule out the
case when the limit superior is finite but non-zero at some time t. Then we can
find a sequence of radii rk going to infinity in which supd(x,xk )∼rk R(t, x) ∼ rk−2 ; in
particular, as κ-solutions are non-collapsed, the volume of the annulus {d(x, xk ) ∼
rk } is ∼ rk3 . We rescale the annuli {d(x, xk ) ∼ rk } by the rk and take a limit (using
Cheeger-Gromov theory) to obtain a Tits cone (the cone over the sphere at infinity)
whose maximum scalar curvature is ∼ 1. But the Tits cone has some vanishing
sectional curvatures. It is then possible to apply tensorial maximum principles to
the equation (8) and conclude in fact that the limit is flat, but this contradicts the
non-zero scalar curvature.
Now let us rule out the case when the limit superior is zero. The above argument
now shows that the Tits cone is flat, which makes the sphere at infinity round. This
makes the manifold asymptotically flat in a volume sense; using Bishop-Gromov
volume comparison theory and the non-negativity of scalar curvature one sees that
this is only possible if the manifold was completely flat, a contradiction.
10.4. Volume grows spatially slower than scaling. Positive curvature tends
to contract volume. Since we have seen that the curvature is larger than what is
predicted by scaling, it is perhaps not then surprising that the volume is smaller
than is predicted by scaling51:
Proposition 10.5 (Vanishing asymptotic volume ratio). [28, §11] If t 7→ Mt is a
non-compact κ-solution, then for each time t and origin x0 ∈ M the asymptotic
volume ratio limr→∞ Volt (B(x0 , r))/r3 (which is independent of x0 ) vanishes.
50There is another possibility, that the 2D manifold looks like RP 2 instead of S 2 , but this case
can be eliminated by reducing to oriented solutions (this is the approach in [28], [22], [5]), or by
some additional case analysis involving the soul theorem (this is the approach in [27]).
51This is consistent with the non-collapsing nature of κ-solutions, because non-collapsing re-
quires bounded normalised curvature in order to lower bound the volume, and this hypothesis will
not be true for extremely large balls with fixed origin.
Proof. (Sketch; for details, see [27, §9.6], [22, §40]) This is an induction on dimen-
sion argument; to illustrate the key inductive step let us assume that the analogous
claim for two dimensions has already been shown and deduce the three-dimensional
case.
Fix t, x0 . Using the infinite asymptotic scalar curvature ratio and a “point-
picking” argument of Hamilton [20], one can find xk and rk → ∞ with rk =
o(d(x0 , xk )) such that limk→∞ R(t, xk )rk2 = ∞, and such that xk essentially max-
imises the scalar curvature on the ball B(xk , rk ). If Volt (B(xk , rk )) = o(rk3 ) then
we can use comparison theory to conclude the argument, so suppose that this is not
the case. Rescaling these balls so that R(t, xk ) = 1, we then see that the rescaled
balls have curvature at most 1 + o(1), radius r̃k → ∞ and volume ∼ r̃k3 , which by
comparison theory implies an injectivity radius of & 1. One can then use Hamil-
ton’s compactness theorem (declaring xk as the origin of the B(xk , rk ), of course)
to extract a subsequence of these balls which converge to a manifold of scalar curva-
ture at most 1 (with equality attained at some origin x0 ), and non-zero asymptotic
volume ratio. One then inspects the rescaled annuli of this manifold (as was done
in the proof of Proposition 10.3) to show that the Tits cone splits as the product of
a line with a non-flat two-dimensional manifold with non-zero asymptotic volume
ratio. But this can be used to contradict the induction hypothesis52.
We have now seen some “asymptotic” connections between high curvature and
small volume. It turns out that one can modify these arguments (again by ex-
ploiting Hamilton’s compactness theorem) to also obtain some “local” connections.
The precise statements are technical (see [5, §6.4] or [22, §41]), but roughly the
gist (allowing for some white lies) is that for any given ball B(x, r) the following
statements are morally equivalent:
• B(x, r) has volume & r3 .
• The curvature on any point in B(x, r) is O(r−2 ).
• The curvature of at least one point in B(x, r) is O(r−2 ).
Note that these equivalences already imply Proposition 10.3 and Proposition 10.5
(and indeed these propositions are used in the proof), since the scalar curvature
is strictly positive. Note that it (morally) doesn’t really matter which curvature
tensor we use here, since the scalar curvature (essentially) bounds the Riemann and
Ricci curvatures thanks to pinching.
There are also a number of useful parabolic estimates which allow one to control
derivatives of the Riemann curvature tensor in a pointwise sense by the scalar
curvature R: indeed we have ∂tk ∇lx Riem = O(Rk+l/2+1 ) for any given k, l ≥ 0.
In summary, the scalar curvature R(x) at a point will control the geometry
more or less completely out to distances O(R(x)−1/2 ) away from x, without any
significant collapse of volume or increase in curvature. Of course R(x)−1/2 is the
natural length scale associated to R(x) by dimensional analysis. After that distance,
volume collapse will occur, and the curvature will fluctuate, though it has to drop
off slower than what scaling predicts.
One consequence of these very strong local controls on curvature (as well as
standard analytical tools such as the Rellich compactness theorem) is the following
useful compactness theorem of Perelman:
52There is a mild technical issue in ensuring that this two-dimensional manifold is also a
κ-solution, which we will gloss over here. See [27, §9.6] or [22, §40] for details.
30 TERENCE TAO
Theorem 10.6 (Perelman compactness theorem). [28, §11] (see also [27, §9.7], [5,
§6.4], [22, §45]) Let t 7→ Mtk be a sequence of κ-solutions (with uniform control on
the non-collapsing), and let xk ∈ M0k be points. We normalise so that the scalar
curvature Rk (0, xk ) of (0, xk ) on M0k is equal to 1. Then the sequence Mtk (with
xk identified as an origin) contains a convergent subsequence whose limit is also a
κ-solution.
Very roughly speaking, this compactness theorem gives automatic uniformity
on any result involving κ-solutions, so that once one obtains some control on the
geometry of each individual κ-solution, one automatically gets uniform control on
the geometry of all κ-solutions at once. We shall therefore gloss over uniformity
issues (which we are doing anyway, thanks to the use of our imprecise O() notation).
10.7. Asymptotic gradient shrinking solitons. It turns out that every κ-solution
has a (possibly non-unique) gradient shrinking soliton as its “asymptotic profile”
as t → −∞. Roughly speaking, this profile is obtained as follows. Pick any point
x(0) ∈ M0 , and then for any t < 0 pick a point x(t) ∈ Mt which is close to q in
reduced length: l0,x(0) (t, x(t)) ≤ 3/2. (Such a point can be shown to exist after
non-trivial effort by using variational formulae and Hamilton’s Harnack inequality
to deriving some differential inequalities for parabolic length, and then use standard
barrier techniques; see [28, §7.1], [22, §23], [5, §6.2], or [27, §7.2].) In fact there is
enough compactness to select x(t) to minimise this reduced length (see [27, §7.2]).
The point is that as t → −∞, the geometry around (t, x(t)), shall begin resembling
a gradient shrinking soliton after rescaling.
More precisely, let tk be a sequence of times going to negative infinity. We rescale
the Ricci flow to send tk to −1 and declare x(tk ) to be origin O to obtain a sequence
of pointed Ricci flows. We would like to extract a convergent subsequence from
this sequence. To do this we need uniform control of the geometry of this sequence.
Because we have bounds on the reduced length at the origin O, it is possible (thanks
to certain differential inequalities for the reduced length, combined with Hamilton’s
Harnack inequality) to then obtain bounds on the curvature at O, and hence by
the preceding sections we obtain bounds on curvature at nearby points as well
(and also lower bounds on volume, and hence lower bounds on injectivity radius).
This is enough control to apply Hamilton’s compactness theorem (or Perelman’s
compactness theorem) and extract a limiting Ricci flow t 7→ Mt∞ . Because each
manifold was non-flat, it had reduced volume strictly less than one at time t = −1;
this reduced volume will be decreasing in k thanks to monotonicity and thus tends to
a limit Ṽ∞ < 1. The reduced lengths lk for each of the manifolds in the sequence can
be shown (with some non-trivial analytic effort) to converge in a moderately strong
topology to a limit l∞ (which morally is the reduced length function of M ∞ , though
due to various defects in the compactness and initial lack of control on the regularity
of the limit flow M ∞ , this is not immediately obvious). We use this pseudo-reduced
length l∞ to define a pseudo-reduced volume Ṽ∞ (t) on M ∞ . Because the reduced
volumes of the rescaled manifolds were converging to a constant Ṽ∞ , one can show
by taking limits53 that Ṽ∞ (t) = Ṽ∞ , thus the pseudo-reduced volume of M̃t is
constant in t. Morally speaking, since reduced volume is only constant for gradient
shrinking solitons, this should imply that M∞ is a gradient shrinking soliton. There
53To justify this we need some additional bounds on reduced length away from the origin O,
but this can be done by using some arguments of Perelman; see [38], [5, §6.4], or [27, §9.21].
are however some annoying analytic issues related to possible defects of compactness
(e.g. the inequality in Fatou’s lemma need not be an equality). Nevertheless it is
possible to extend the above classification of constant reduced volume solutions
to constant pseudo-reduced volume by carefully going through the proof of the
classification interpreting everything in a distributional sense; see [5, §6.2], [22,
§38], or [27, §9.2]. Thus the limit manifold is indeed a gradient shrinking soliton.
10.8. Classification of gradient shrinking solitons. Now that we know that
κ-solutions contain rescaled copies of a gradient shrinking soliton as t → −∞, it
is of interest to classify these solitons. We already know two such examples, the
sphere S 3 and the cylinder S 2 × R54. It turns out that up to some minor details
such as quotients by finite group actions, the these are the only κ-solutions which
are gradient shrinking solitons. This important result is due to Perelman [28, §9]
but is mainly proven using earlier technology of Hamilton. There are several cases
to consider in the proof. If the soliton has at least one zero sectional curvature, it is
possible to use Hamilton’s maximum principle for the Ricci curvature to conclude
that the soliton is a quotient of the cylinder. If the soliton has positive sectional
curvature and is compact, one uses a result of Hamilton [16], which says that all
such manifolds become asymptotically round under Ricci flow, to conclude that
the gradient-shrinking soliton is perfectly round, and is thus the sphere soliton
or a quotient thereof. The difficult case is when the soliton has strictly positive
curvature but is non-compact. By integrating the equation (5) along gradient curves
of the potential function f appearing in (5), one can show that the scalar curvature
does not go to zero at infinity; standard splitting arguments as in the preceding
paragraph then let one show that the manifold is asymptotically a cylinder, with
some upper bound on the scalar curvature. On the other hand, one can show a
monotonicity formula for the area of level sets of f . Comparing the lower bounds
on these areas against the upper bounds on curvature, one eventually obtains a
contradiction to the Gauss-Bonnet theorem, thus excluding this case. See [5, §6.2]
or [27, §9.4] for details.
10.9. Canonical neighbourhoods. To summarise so far, we have learnt that the
space of κ-solutions is compact (modulo scaling), and that asymptotically these
solutions resemble (after rescaling) either a constant-curvature compact surface
(such as a sphere) or a round cylinder. This strongly suggests that any given κ-
solution at any given time is formed by something resembling the connected sum
of rescaled round cylinders and rescaled constant-curvature compact surfaces. We
now (partially) formalise this intuition.
Let us say that a point x = (t, x) in a Ricci flow lives in a rescaled canonical
neighbourhood if, after rescaling so that R(x) = 1, one of the following (somewhat
informally phrased) assertions55 is true:
• (ε-neck) After rescaling, x lies in the center of a region which is within
ε (in a smooth topology such as C k for suitably large k) of the cylinder
[−1/ε, 1/ε] × S 2 with the standard metric, for some suitably small ε. In
some cases, we need to additionally assume that similar statements hold
(with slightly wider cylinders) if we evolve backwards in rescaled time for a
54The three-dimensional cigar-type solutions, in contrast, are a gradient steady solitons; their
asymptotic gradient shrinking soliton is the cylinder S 2 × R.
55For more precise definitions, see [29, §1], [27, §9.8], [22, §57], or [5, §7.1].
32 TERENCE TAO
short (but fixed) duration. This neck models the cylinder solution to Ricci
flow. We also define an ε-tube to be a finite or infinite number of ε-necks
chained56 together, thus for instance the standard cylinder R × S 2 is a
doubly infinite ε-tube, while cigar-type solitons are a half-infinite ε-tube
with a (C, ε)-component (see below) glued on the end.
• (C-component) After rescaling, x lies in a compact manifold diffeomorphic
to S 3 (or RP 3 ) with diameter ∼ 1, and all sectional curvatures positive and
∼ 1. This component models (rather crudely) the sphere solution to Ricci
flow.
• (ε-round) After rescaling, x lies in a open manifold which is within ε (in
a smooth topology) to a manifold of constant curvature +1, such as the
sphere S 3 or a quotient S 3 /Γ, for some small ε. This is similar to the
C-component with more precise control, but possibly with much smaller
diameter (because Γ could have arbitrarily large).
• ((C, ε)-cap) After rescaling, x lies in an open region C, diffeomorphic to
S 3 − {pt} or RP 3 − {pt} of diameter O(1), scalar curvature positive and
∼ 1 (with Lipschitz bounds in space and time). Furthermore, the region
C decomposes into an ε-neck N (with one end the boundary of C), and
a remainder Y , whose boundary is the other boundary of the ε-neck, and
whose interior Y (known as the core) contains x, and has a non-collapsed
geometry at scale 1. This cap models cigar-like solutions to Ricci flow.
It turns out that every point in a κ-solution lies in a rescaled canonical neigh-
bourhood. In fact a more global qualitative description of κ solutions is possible:
Theorem 10.10 (Description of κ-solutions). Let t 7→ Mt be a κ-solution. Then
one of the following occurs:
• (Constant-curvature compact) Mt is the round sphere example (or quotient
thereof ).
• (Approximately sphere or RP 3 ) Mt is a C-component for all t.
• (Cylinder) Mt is the cylinder example, or a quotient thereof by an involu-
tion. (In the former case, Mt is diffeomorphic to a punctured R3 .)
• (Approximate cigar) For each t, Mt is diffeomorphic to R3 , has strictly
positive sectional curvatures, and is a half-infinite ε-tube with a (C, ε)-cap
glued on at one end.
• (Two-ended cigar) For each t, Mt is diffeomorphic to S 3 or RP 3 , has
strictly positive sectional curvatures, and is a finite ε-tube with (C, ε)-caps
glued on at both ends.
For more precise formulations of this result, see [28, §11], [29, §1], [27, §9.8], [22,
§58], [5, §7.1].
Proof. (Sketch) There are several cases to consider. If Mt has vanishing sectional
curvature at even just one point and in one direction, then application of Hamilton’s
maximum principle for Ricci flow allows one to show that Mt (or a two-sheeted
covering thereof) factorises as the product of R with a two-dimensional κ-solution.
It turns out that there is only one such solution, namely the shrinking two-sphere
(this is a consequence of the classification of the sphere S 2 as the unique two-
dimensional gradient shrinking soliton, combined with Hamilton’s observation that
56One has to be a little careful what “chaining” means here; see [27] for full details.
any positively curved compact Ricci flow becomes round at the blowup time), and
we get the cylinder (or a quotient thereof).
Henceforth we may assume Mt has strictly positive sectional curvatures. If it is
non-compact, then the soul theorem (see e.g. [31]) tells us that Mt is diffeomorphic
to R3 and any soul is a point. Pick one such point, say x0 . Away from this point,
one can use the equivalences in Section 10.4 to conclude that the curvature decays
slower than scaling, thus limx→∞ R(t, x)d(x, x0 )2 = +∞ at any time t. Using the
classification of asymptotic gradient solitons, one can then conclude that Mt is
asymptotically a cylinder at infinity, which turns out to be enough to obtain the
approximate cigar case.
One consequence of the above analysis is that every point in a non-compact κ-
solution either lies in a rescaled ε-neck or is in the core of a rescaled (C, ε)-cap; we
shall need this result shortly.
The only remaining case is the compact case. We fix a time t and normalise the
maximum scalar curvature to be +1. There are two cases depending on whether
the normalised diameter of M is very large (≫ 1) for all time or bounded O(1) for
at least one time. In the former case, a rescaling argument exploiting Perelman’s
compactness theorem and observation in the preceding paragraph shows that every
point in M is contained either in the rescaled core of a (C, ε)-cap or a rescaled
ε-neck. Some elementary topological arguments then show that the compact man-
ifold M , which is glued together from finitely many of these components, must be
homeomorphic to S 3 , RP 3 , RP 3 #RP 3 , or an S 2 -fibration over S 1 . On the other
hand, compact manifolds of positive curvature are known to have finite fundamental
group, which rules out the latter two possibilities, and then leads to the two-ended
cigar case.
Finally, we have to consider the case when the manifold is compact with bounded
normalised diameter at some time t. If at some other time the manifold ceases to
have bounded normalised diameter, then by the preceding discussion M is homeo-
morphic to S 3 or RP 3 and will then be a rescaled C-component (using Perelman’s
compactness theorem to obtain uniformity in t). So we may assume that we have
bounded normalised diameter throughout. We look at the asymptotic gradient-
shrinking soliton. It must also have bounded normalised diameter, and thus cannot
be a cylinder or a quotient; it must be the sphere example. But then Hamilton’s
rounding theorem for positively curved compact manifolds implies that Mt is also
the round sphere example (or a quotient thereof).
11. Blowup analysis III: approximation by ancient solutions

Having concluded our (lengthy!) analysis of κ-solutions, we return now to the
blowup analysis of singularities or high-curvature regions of more general solutions
to Ricci flow. Very roughly speaking, Perelman [28, §12], [29, §6] establishes the
following three claims regarding a Ricci flow, as one gets sufficiently close to a
blow-up time:
(i) If one rescales a point (t, x) of sufficiently high scalar curvature R(t, x) ≥ K
so that the scalar curvature is 1, then a large57 rescaled neighbourhood of
that point is very close to the corresponding neighbourhood of a κ-solution.
57By “large” I mean of radius ≫ 1 in both space and time in rescaled coordinates. In unrescaled
coordinates, this would be a neighbourhood of radius ≫ R(t, x)−1/2 in space and ≫ R(t, x)−1 in
time.
34 TERENCE TAO
(ii) If one rescales a point (t, x) of sufficiently high scalar curvature R(t, x) ≥ K
so that the scalar curvature is 1, then a large rescaled neighbourhood of
that point is a canonical neighbourhood.
(iii) If one rescales a point (t, x) of sufficiently high scalar curvature R(t, x) ≥ K
so that the scalar curvature is 1, then a large rescaled neighbourhood of
that point has bounded curvature.
For precise statements of these claims see [28, §12], [29, §6], [27, §10,11], [5,
§7], [22, §51-52]. Note that the arguments and statements in [28, §12] had some
inaccuracies, but these were essentially all corrected in [29, §6].
The three claims are inter-related. The claim (i) implies (ii) thanks to Theo-
rem 10.10. Furthermore, (iii) can be used58 to imply (i) because of the no-local-
collapsing theorem, which parlays the bounded curvature property into an injectiv-
ity radius property, which in turn lets59 us use Hamilton compactness (and Perel-
man compactness) to extract the claim. To conclude the proof of all these claims,
Perelman introduces a downward induction on the curvature threshold K, essen-
tially reducing matters60 to showing that (ii) (for a slightly larger value of K, say
4K) implies (iii). Iteration in K (from incredibly huge down to some moderately
large constant) then gives all three claims (i)-(iii).
It remains to show that (ii) for 4K implies (iii) for K. Thus, we are assuming
that points of really high curvature lie in rescaled canonical neighbourhoods, and
want to conclude that points of slightly less high curvature lie in rescaled neigh-
bourhoods without curvature blowup. Oversimplifying enormously, the argument
is as follows. Suppose for contradiction that the claim failed. Then one could find a
sequence of large neighbourhoods Un in rescaled Ricci flows where the centre of the
neighbourhood had scalar curvature 1, but that the curvature at other points of the
neighbourhood were unbounded. Furthermore, every point of scalar curvature at
least 4 was contained in a canonical neighbourhood. This means that there exists
a curve γn in each Un which starts at a point of curvature exactly 4, and passes
entirely through canonical neighbourhoods until it reaches a point of arbitrarily
large curvature. It is not hard to show that this scenario cannot happen with the
ε-round or C-component type of canonical neighbourhoods, so essentially this curve
γ has to be contained in a long chain of rescaled ε-necks (possibly with one cap on
the end, and with different rescaling on each component of the chain, in contrast to
an ε-tube; this is a truncated version of an ε-horn, which we shall see in the next
section). But the curvature is unbounded, which basically means that the necks
are getting narrower and narrower at one end. It turns out that there is enough
control on the geometry inside this chain that one can extract a limit, and obtain an
incomplete Ricci flow which is topologically a cylinder S 2 × (0, 1) in space, but has
bounded curvature on one end and asymptotically infinite curvature at the other,
58Conversely, one can use the material in Section 10.4 to deduce (iii) from (i), though this
direction of implication will not be of use to us.
59There is a technical issue here in that one needs to extend the control on the geometry of the
rescaled high-curvature region backwards in time somewhat in order to ensure that the limiting
solution is indeed ancient. This subtlety is not fully addressed in Perelman’s work, but see [22,
§51], [27, §11.2], [5, §7.1] for resolutions.
60There is the issue of establishing the claims (i)-(iii) for the “base case” when K is incredibly
large, but this can be dealt with trivially by truncating the Ricci flow just before the first singu-
larity, so that it is still bounded curvature, and then the claims are vacuously true for K large
enough.
and consists entirely of ε-necks chained together. One then approaches the infi-
nite curvature end, rescaling as one goes, and extracts a further limit (using some
Aleksandrov-Toponogov theory) to obtain an open non-flat cone (cf. the arguments
in Sections 10.2, 10.4). At this point there is a technical step in which one uses
parabolic theory to upgrade the convergence of this limit from Gromov-Hausdorff
convergence to a smoother type of convergence, which we gloss over. Now we use an
argument of Hamilton. Namely, one observes that cones have vanishing sectional
curvatures along their rays. Using Hamilton’s maximum principle once again, one
can show that the manifold locally splits as a product of a two-dimensional Ricci
flow with a line; in particular the scalar curvature stays constant along each ray
of the cone. But this contradicts the fact that the curvature is blowing up as one
approaches the apex of the cone.
Details of this argument can be found in [28, §12], [29, §4], [27, §10-11],[22,
§51-52], or [5, §7.1].
12. Horns
We now briefly mention an important technical issue regarding constants. We
have seen that at times close to a singularity, every high-curvature point in a Ricci
flow is contained in a canonical neighbourhood, such as an ε-neck. As it turns
out, the ε parameter here (which is measuring how close the ε-neck is to a round
cylinder) is merely a small absolute constant, such as 10−6 . For technical reasons
later on (having to do with making sure that Ricci-flow-with-surgery has the same
type of estimates as smooth Ricci flow) one actually needs to find necks that are
much closer to being round, i.e. to find δ-necks for some δ ≪ ε; these very round
necks turn out to be excellent locations for performing surgery. This turns out to
be possible by a variant of the arguments used previously. Very roughly, things
proceed as follows.
Let us first pass to the very final time T∗ of a smooth Ricci flow, where singulari-
ties have already begun to occur. There will however still be regions of the manifold
where the scalar curvature did not pointwise go to infinity; it turns out that in these
regions one can still define a limiting manifold Ω at time T∗ , although this manifold
will now be incomplete. On the other hand, we know that the curvature will always
go to infinity near any incomplete end of the manifold (in other words, the curva-
ture function is proper ). We also know that high curvature points are contained in
canonical neighbourhoods. If these neighbourhoods are ε-round or C-components,
then they are disconnected from the low-curvature regions of Ω; if the neighbour-
hood is an ε-cap, and is connected to a low-curvature region then this portion of the
manifold is not really close to an incomplete end of the manifold. Putting all this
together, one sees that each incomplete end of Ω is the infinite end of a half-infinite
chain of ε-necks, whose curvature goes to infinity (thus the necks get increasingly
narrow at the infinite end). This shape is known as an ε-horn. Thus the picture
of the limiting incomplete manifold Ω is a large (possibly infinite) number of high
curvature connected components (which in addition to ε-round and C-component
pieces, can also be double ε-horns, a capped ε-horn, or a doubly capped ε-tube), to-
gether with the components that contain at least some low-curvature points, which
consist of compact sets with finitely many ε-horns attached to them. See [29, §3],
[22, §66], [5, §7.3], [27, §11] for details.
36 TERENCE TAO
Now the key point is that ε-horns increasingly resemble a round cylinder as one
goes deeper into the infinite end of the horn. More precisely, given any δ > 0, it
turns out that every point in the horn of sufficiently high curvature (depending on
δ) will be contained in a δ-neck. The proof of this statement (like many previous
arguments here) is by contradiction; if not, then we can find points xn of increasingly
high curvature which are unable to be placed in δ-necks. Rescaling and taking limits
we end up with a solution M∞ (not necessarily ancient) which has non-negative
sectional curvature, is non-collapsed, and is non-flat, and with every point contained
in an ε-neck. By considering what happens to geodesic rays from the xn to some
reference low-curvature point and to a point on the incomplete boundary (i.e. more
Aleksandrov-Toponogov theory) we see that M∞ has at least two ends (thus this
rules out toroidal examples such as S 2 × S 1 ). It turns out that this, combined
with negative curvature, already forces M∞ to be the product of R with a two-
dimensional manifold from Riemannian geometry considerations (see [27, §2.4]).
This two-dimensional manifold will of course also obey Ricci flow. Given that M∞
already has a local structure of an ε-neck, we see that the two-dimensional manifold
must be topologically S 2 and have positive curvature. It turns out that one can
use curvature bounds to extend this solution back in time to an ancient solution,
at which point the rounding theory of Hamilton shows that this solution is in fact
a round sphere, and so M∞ is a round cylinder. In particular the neighbourhoods
of xn must increasingly resemble this cylinder, a contradiction. Again, see [29, §3],
[22, §70], [5, §7.3], [27, §11] for details.
13. Standard solutions

Now that we have fully analysed the singularities of the Ricci flow, we are almost
ready to perform the surgery procedure that will remove these singularities and
replace them with a smooth manifold which can be used to continue the Ricci flow-
with-surgery procedure. There is however the question of what exactly the smooth
manifold will be that will do this. This will be the standard solutions of Perelman
[29, §2] (the material in which was then expanded in [5, §7.3], [27, §12], or [22, §59-
62]). The advantage of using these solutions to replace the high curvature regions
of a Ricci flow is that their blowup time, and their asymptotic behaviour near that
blowup time, is completely understood (to a much greater extent than for arbitrary
singularities).
Recall from the previous section that high curvature regions will contain a long
ε-tube, which is thus close to a cylinder. The idea of surgery is to replace this
cylinder-like region with two cigar-type solutions (with the infinite end of each cigar
being one of the two ends of the cylinder); this potentially changes the topology
but will also turn out to take a substantial bite out of the volume, so that this
procedure is only performed finitely often (thanks to the finite extinction result
and the bounded growth of volume).
Thus we want a half-infinite solution which resembles a cylinder at one end
and is rounded off at the other. In light of all the preceding analysis, it is not
unreasonable to also request as desirable properties that the sectional curvatures
are all non-negative and that the solution is non-collapsed at all scales O(1). (It
turns out to be neither necessary nor practical to try to make this solution an
ancient solution or to enforce non-collapsing at large scales.) This solution will
then be referred to as the standard solution. It is far from unique; its role in the
theory is vaguely analogous to the role of a smooth cutoff function in harmonic

analysis, and so it is the quantitative properties of this solution (such as blowup
asymptotics) which are of importance, rather than the explicit form of the solution.
To construct the standard solution t 7→ Mt (which will last for time t ∈ [0, 1)),
we first construct it at time t = 0. What we do here is that we design the initial
manifold M0 to be a manifold of revolution in R4 (with the Riemannian metric
inherited from the ambient Euclidean space R4 ), which is equal to the half-infinite
cylinder R+ × S 2 (where S 2 is embedded as the unit sphere in R3 in the usual
manner) with a cap smoothly glued on to the end. It is not difficult (see [27, §12] or
[5, §7.3]) to write down an explicit choice of such a manifold of revolution with non-
negative sectional curvature and positive scalar curvature everywhere; this manifold
is then isometric with R3 with some rotationally invariant metric. We then evolve
this manifold by Ricci flow. Using the local existence and uniqueness theory61 for
non-compact manifolds (see [10]), this solution will exist62 up to some maximal time
0 < T∗ ≤ ∞ of existence and remain rotationally symmetric throughout. Since the
initial manifold is asymptotically equivalent to the unit cylinder, which is known to
become extinct at time t = 1, a simple limiting argument shows that the standard
solution t 7→ Mt must asymptotically approach the shrinking cylinder at all times
0 < t < min(1, T∗ ) and in particular must develop a singularity by time t = 1
at the latest; in other words, T∗ ≤ 1. Also an application of Hamilton’s maximal
principle (modified suitably to deal with the non-compactness of the manifold,
see [34], [27, §12.2], [10]) shows that the solution will always have non-negative
sectional curvature and positive scalar curvature (and after t > 0 the sectional
curvature becomes strictly positive). It is not hard to demonstrate non-collapsing
at scales O(1) at time t = 0 due to the explicit form of the manifold, and then by
Perelman’s monotonicity formulae we have non-collapsing at all later times also at
the same scales.
Now it is possible to show that blowup cannot occur at any time strictly less
than 1 by the following argument. Suppose for contradiction that T is strictly less
than 1. Using the theory of Ricci flows on incomplete manifolds as in the previous
section, one can show that as t → T , outside of a compact set K, the Ricci flow
remains close to that of the shrinking cylinder solution. Also, from Proposition 6.1
we know that the metric is shrinking, and thus metric balls are increasing. One
consequence of this is that the volume of a large ball B(x, C) for some large but
fixed C cannot collapse to zero as one approaches the blowup time T , because this
large ball will contain a big component outside of K at time t = 0 and hence at
all later times by the increasing nature of the metric balls. Combining this with
Bishop-Gromov comparison theory (cf. Section 8) and the non-negative sectional
curvature one can conclude the same statement for small balls: the normalised
volume of any ball B(x, r)/r3 for r = O(1) cannot collapse to zero as t → T , even
61It is possible to avoid use of this theory by approximating the non-compact manifold as a
limit of compact manifolds, and exploiting the rotational symmetry assumption heavily to reduce
the PDE analysis to what is essentially a 1 + 1-dimensional dynamics which can then be analysed
by fairly elementary (though lengthy) means. See [29, §2], [27, §12], [22, §59-65] for this approach.
The approach in [5], which goes through the non-compact local theory of [10], does not actually
require rotational symmetry assumptions.
62
It will also be unique, although this fact is not strictly necessary for the argument, as observed
in [22, §59] and attributed to Bernhard Leeb.
38 TERENCE TAO
if we also let r → 0 as well. In other words, the volume of balls cannot grow slower
than scaling at scales O(1).
On the other hand, because we are assuming T is the maximal time of existence,
then there must be arbitrarily high curvature regions near the blowup time. By
Section 11 these high curvature regions resemble rescaled κ-solutions. By Section
10.4 this means that the volume of certain rescaled balls grows slower than scaling.
This rescaled balls get increasingly small as the curvature gets higher, and so we
contradict the preceding paragraph. Thus we must have T = 1. (For precise details
of this argument, see [29, §2], [27, §12.6], [22, §61], or [5, §7.3].)
Once T = 1, one can repeat the above types of arguments and eventually obtain
a bound
1
Rmin (t) & ;
1−t
the basic reason for this is that if Rmin was ever small at some point x compared
1
to 1−t , then the analysis in Section 11 would imply that the curvature remains
bounded near x even up to the blowup time T = 1. One can then show that the
same holds at all other points, because if there was a transition between bounded
limiting curvature and unbounded limiting curvature, the above type of analysis
shows that there is no volume collapsing near the first type of point and volume
collapsing near the second type of point, a contradiction. This leads to a limiting
metric at time t = 1, which must have unbounded curvature since T = 1 is the
maximal time of existence, which implies (by the blowup analysis) that the limiting
metric contains arbitrarily small necks. But this turns out to be incompatible with
the positive curvature. (See [29, §2], [22, §61-62], [27, §12.6], or [5, §7.3] for details.)
Because of this bound, we know that at times close to t = 1 the entire manifold
consists of high-curvature regions. One can then use the classification of high
curvature regions and basic topology (analogous to the arguments in Section 11)
to show that near the blowup time of a standard solution, every point either lives
in a ε-neck or (C, ε)-cap; again, see [29, §2], [22, §61-62], [27, §12.6], or [5, §7.3] for
details.
14. Surgery
We can now informally describe how Ricci flow with surgery is performed.
We first normalise the initial manifold (M0 , g0 ) so that the curvature is O(1)
and that there is no collapsing at scales O(1). This is enough to ensure that Ricci
flow will remain smooth for a fixed time interval, say [0, 1]. We then split the
remaining future time interval [1, ∞) into dyadic pieces [2n , 2n+1 ) and define Ricci
flow with surgery with slightly different surgery parameters on each such dyadic
time interval (for technical reasons one needs the parameters on each interval to
be much more extreme than those on the previous intervals). In the case of the
Poincaré conjecture, when M0 is simply connected, the finite time extinction theory
from Section 6 (extended to the case of Ricci flows with surgery) ensure that only
finitely many of these intervals need to be considered before the flow becomes
completely extinct; however for the purposes of proving the full geometrisation
conjecture, finite time extinction is not known in general, and so one must then
consider all of these dyadic time intervals.
For the rest of this section we work on a single dyadic time interval [2n , 2n+1 ),
assuming that the flow with surgery has somehow been constructed up to time
2n already. We shall need two surgery parameters, a small canonical neighbour-

hood parameter rn , which has the dimension of length, and a small surgery control
parameter δn , which is dimensionless and measures the accuracy of the canonical
neighbourhoods.
If the flow is non-singular on all of [2n , 2n+1 ) then we simply perform Ricci flow
on the entire interval. Now suppose instead that the flow first becomes singular at
some time T∗ in the interval [2n , 2n+1 ). From the theory in Section 12 we know
that the high-curvature regions - the regions where the curvature is ≫ 1/(δn rn2 ) -
are either disconnected from the low-curvature regions, or else lie in one of a finite
number of horns. Furthermore, there exists a surgery scale parameter hn , which is
a small parameter with the dimension of length (and much smaller than δn or rn )
with the property that every point in the Ricci flow63 with scalar curvature at least
h−2
n in a horn is contained in the centre of a δn -neck.
The high-curvature components which are disconnected from the low-curvature
regions are easy to classify: they are either a single compact canonical neighbour-
hood, or the finite union of non-compact canonical neighbourhoods glued together
to form a compact manifold. Some topology chasing then limits the possibilities
for the topology to be S 3 or RP 3 (or a quotient thereof), or a S 2 bundle over S 1 ,
or a connected sum of two RP 3 ’s.
At the blowup time T∗ we perform the following surgery procedure. All the
high-curvature regions which are disconnected from the low-curvature regions are
removed. For each horn, we locate a point x0 deep inside the horn where the scalar
curvature is equal to h−2 n , and hence lies inside a δn -neck. In particular, the point
x0 is contained inside a two-sphere, which separates the singular part of the horn
from the boundary connecting the horn to the low-curvature regions. We then
move from this two-sphere deeper into the horn by a rescaled distance A for some
large absolute constant A (and thus move by Ahn in the non-rescaled metric) -
this is still well within the δn -neck, and so still gives another two-sphere. We then
cut off and discard everything beyond this two-sphere and replace it by a standard
solution (rescaled by hn , and also smoothly interpolated to deal with the fact that
the δn -neck is not a perfect cylinder). Now the manifold has become smooth again,
and we restart the Ricci flow from this point. (For more precise descriptions of this
surgery procedure, see [28, §4], [5, §7.3], [22, §67, 71, 72], [27, §13,14,15.4].)
As we will eventually want to deduce the topology of the initial manifold from
the topology at a much later point in the Ricci flow with surgery, it is important to
know how to recover the topology of a pre-surgery manifold from the post-surgery
manifold. (The smooth portion of the flow will of course not affect the topology.)
So let us reverse the surgery parameter and see what happens. Firstly, reversing the
surgery adds a lot of new disconnected high-curvature components to the manifold.
Each point in these components is contained in a canonical neighbourhood. Some
topology-chasing then lets us describe the geometry (and hence topology) of these
components as one of the following:
• A C-component, diffeomorphic to S 3 or RP 3 .
63Strictly speaking, one needs to extend the analysis of Section 12 from smooth Ricci flows to
Ricci flows with surgery. This turns out to be a remarkably lengthy and non-trivial task, and it is
here that it becomes crucial that the δn , rn are chosen extremely small compared to all previous
values of δ and r, but we shall gloss over these very important technical issues in order to focus
on the broader picture of the argument.
40 TERENCE TAO
• An ε-round component, diffeomorphic to a finite quotient of S 3 .

• An S 2 fibration over S 1 .
• An ε-tube capped off by two caps (and thus diffeomorphic to S 3 ).
• An ε-horn capped off at one end.
• A double ε-horn.
The first four cases have no singularities and will just add disconnected compo-
nents with a well-controlled topology to the pre-surgery manifold. The other two
cases will have singularities which will link up with the ends of the ε-horns attached
to the low-curvature components of the manifold. It is also possible that two ε-
horns connected to the low-curvature components of the manifold are attached to
each other. Moving backwards a little bit in time to desingularise the pre-surgery
manifold, we see that the effect of the horns is either to connect two components
of the post-surgery manifold together, to connect sum a capped cylinder (i.e. a S 3 )
to a component (which does nothing to the topology), or to connect sum a torus
S 2 × S 1 to a component. So if there are only finitely many surgeries before total
extinction, each connected component of the manifold at any intermediate time is
a connected sum of finitely many manifolds which are either finite quotients of S 3 ,
RP 3 , or S 2 fibrations over S 1 . More generally, one can show that if a post-surgery
manifold obeys Thurston’s geometrisation conjecture, then so does the pre-surgery
manifold. (For details, see [29, §3], [22, §72], [5, §7.3], [27, §15]. There are minor
variations between these papers due to slightly different definitions of surgery.)
To complete the proof of Theorem 1.2 it then suffices (oversimplifying somewhat)
to establish the following three claims:
(i) The Ricci flow with surgery becomes extinct in finite time.
(ii) The surgery times are discrete (and hence finite, thanks to finite time ex-
tinction).
(iii) The characterisation of high-curvature regions, as used in the above analy-
sis, remains valid even in the presence of surgery. In particular, the manifold
remains non-collapsed at all sufficiently small scales.
(For the geometrisation conjecture, (i) can fail64, but instead one needs to show
that the geometrisation conjecture is asymptotically true in the limit t → ∞.)
The discreteness (ii) is the easiest to establish. Without loss of generality we
may work in a single dyadic time interval [2n , 2n+1 ). At each surgery time, either
a connected component of the manifold is removed, or an ε-horn is replaced with a
standard solution at the scale hn . In the latter case, the total volume of the manifold
is reduced by & h3n , and the number of connected components increased by at most
one, for each horn. In the former case, the number of connected components are
decreased by at least one. On the other hand, because the Ricci curvature is
bounded from below, we know that the total volume of the manifold during the
smooth part of the Ricci flow can only grow exponentially, and thus stays bounded
on a compact time interval such as [2n , 2n+1 ]. Putting all this together we see that
there can only be finitely many surgeries on any compact interval, i.e. that the
surgery times are discrete. See [29, §4.4], [27, §17.2], [22, §72], [5, §7.4].
64Indeed, if the fundamental group is not a free product of finite groups and infinite cyclic
groups - which is the case for instance if any prime component of the manifold has a contractible
universal covering, such as a hyperbolic 3-manifold - then the topological control of surgery dis-
cussed earlier shows that the Ricci flow cannot become fully extinct in finite time. We thank John
Morgan for clarifying this point.
To establish (iii), the key point is to ensure that some analogue of the mono-
tonicity of the Perelman reduced volume (see Section 7) for the Ricci flow, con-
tinues to hold for Ricci flow with surgery. This turns out to be rather technical
but straightforward to verify; the main point is to verify that the curves γ which
nearly attain the parabolic length (12) do not pass through the regions that are
cut away by surgery, and so behave (locally) as if one had a smooth Ricci flow.
In order to achieve this, it becomes essential that the surgery parameters on the
dyadic time interval [2n , 2n+1 ] are chosen to be sufficiently small compared to the
surgery parameters on previous intervals. See [29, §5], [22, §76-79], [5, §7.4], [27,
§16] for details. One technical point that arises for this analysis is that the quanti-
tative bounds for the non-collapsing on the time interval [2n , 2n+1 ] get substantially
worse as n increases, although this ultimately does not cause a major difficulty since
it simply means that the surgery parameters on later intervals have to be chosen
even smaller.
Claim (i) is proven in [27, §18]. The strategy is to adapt the finite extinction
arguments for Ricci flow of Perelman [30] as sketched in Section 6 to Ricci flow with
surgery. One can begin with some topological arguments which ensure that after
finitely many “exceptional” surgeries, the surgeries are all among homotopically
trivial 2-spheres (and so no longer affect π2 (Mt )). One then observes that the
geometric invariant of the minimal energy W2 (t) of a homotopically non-trivial 2-
sphere in Mt is not only decreasing under Ricci flow (as already noted in Section 6)
but is also non-increasing under surgery. As a consequence one readily deduces that
all homotopically non-trivial 2-spheres cease to exist in finite time, thus π2 (Mt ) is
trivial for large t.
A similar (but lengthier, and considerably more technical) argument also works
for the min-max energy W̃3 (t) in Section 6 to ensure π3 (Mt ) also becomes trivial
for large t. As mentioned in Section 6, there is a significant new technical issue
unrelated to surgery arising from the need to understand curve shortening flow in
order to demonstrate the monotonicity of W̃3 (t) even in the absence of surgery.
Once π3 (Mt ) is trivial one can use topological arguments (combined with the ex-
isting control on π1 (Mt ) and π2 (Mt ) to force that Mt is empty, and thus we have
finite time extinction as desired. This finally establishes the Poincaré conjecture!
References
[1] M. Anderson, Geometrization of 3-manifolds via the Ricci flow, Not. Amer. Math. Soc. 51
(2004), 184–193.
[2] R. Bryant, Local existence of Ricci solitons, preprint.
[3] H. Cao, Existence of Kähler-Ricci solitons, Elliptic and Parabolic Methods in Geometry
(Minneapolis, MN, 1994), 1–16. AK Peters, Wellesley MA, 1996.
[4] H. Cao, B.L. Chen, X.P. Zhu, Ricci flow on compact Kähler manifolds of positive bisectional
curvature, C.R. Acad. Sci. Paris, 337 (2003), 781–784.
[5] H. Cao, X. Zhu, A Complete Proof of the Poincar and Geometrization Conjectures - applica-
tion of the Hamilton-Perelman theory of the Ricci flow, Asian Journal of Mathematics, June
2006.
[6] J. Cheeger, Finiteness theorems for Riemannian manifolds, Amer. J. Math. 92 (1970), 61–75.
[7] J. Cheeger, M. Gromov, Collapsing Riemannian manifolds while keeping their curvature
bounded I., J. Diff. Geom. 23 (1986), 309–346.
[8] J. Cheeger, M. Gromov, M. Taylor, Finite Propagation Speed, Kernel Estimates for Functions
of the Laplace Operator, and the Geometry of Complete Riemannian Manifolds, J. Diff.
Geom. 17 (1982), 15–53.
42 TERENCE TAO
[9] B.L. Chen, X.P. Zhu, Uniqueness of the Ricci flow on complete noncompact manifolds,
preprint, math.DG/0505447.
[10] B.L. Chen, X.P. Zhu, Ricci flow with surgery on four-manifolds with positive isotropic cur-
vature, math.DG/0504478.
[11] B. Chow, S.C. Chu, A geometric interpretation of Hamilton’s Harnack inequality for Ricci
flow, Math. Res. Lett. 2 (1995), 701–718.
[12] B. Chow, P. Lu, L. Ni, Hamilton’s Ricci flow, to appear.
[13] T. Colding, W. Minicozzi, II, Estimates for the extinction time of Ricci flow on certain
3-manifolds and a question of Perelman, J. Amer. Math. Soc. 18 (2005), 561–569.
[14] D. DeTurck, Deforming metrics in the direction of their Ricci tensors, J. Diff. Geom. 18
(1983), 157–162.
[15] M. Gromov, Pseudoholomorphic curves in symplectic manifolds, Invent. Math. 82 (1985),
307–347.
[16] R. Hamilton, Three-manifolds with positive Ricci curvature, J. Diff. Geom 17 (1982), 255–
306.
[17] R. Hamilton, The Ricci flow on surfaces, Contemp. Math. 71 (1988), 237–261.
[18] R. Hamilton, The Harnack inequality for the Ricci flow, J. Diff. Geom. 37 (1993), 225–243.
[19] R. Hamilton, A compactness property for solutions of the Ricci flow, Amer. J. Math. 117
(1995), 545–572.
[20] R. Hamilton, The formation of singularities in Ricci flow, Surveys in Diff. Geom. 2 (1995),
7–136. International Press.
[21] R. Hamilton, Non-singular solutions of the Ricci flow on three-manifolds, Comm. Anal.
Geom. 7 (1999), 695–729.
[22] B. Kleiner, J. Lott, Notes on Perelman’s Papers, preprint, math.DG/0605667.
[23] P. Li, S.-T. Yau, On the parabolic kernel of the Schrödinger operator, Acta Math. 156 (1986),
153–201.
[24] P. Lu, G. Tian, Uniqueness of standard solutions in the work of Perelman, preprint.
[25] J. Milnor, Towards the Poincaré conjecture and the classification of 3-manifolds, Notices
Amer. Math. Soc., November 2003.
[26] J. Morgan, Recent progress on the Poincaré conjecture and the classification of 3-manifolds,
Bull. Amer. Math. Soc. 42 (2005), 57–78.
[27] J. Morgan, G. Tian, Ricci Flow and the Poincaré Conjecture, preprint, math.DG/0607607.
[28] G. Perelman, The entropy formula for the Ricci flow and its geometric applications, preprint,
math.DG/0211159.
[29] G. Perelman, Ricci flow with surgery on three-manifolds, preprint, math.DG/0303109.
[30] G. Perelman, Finite extinction time for the solutions to the Ricci flow on certain three-
manifolds, preprint, math.DG/0307245.
[31] P. Petersen, Riemannian geometry, Graduate Texts in Mathematics 171, Springer-Verlag,
New York, 1998.
[32] J. Sacks, K. Uhlenbeck, The existence of minimal immersions of 2-spheres, Annals of Math.
(2) 113 (1981), 1–24.
[33] W.X. Shi, Deforming the metric on complete Riemannian manifold, J. Diff. Geom. 30 (1989),
223–301.
[34] W.X. Shi, Ricci deformation of the metric on complete non-compact Riemannian manifolds,
J. Diff. Geom. 30 (1989), 303–394.
[35] T. Shioya, T. Yamaguchi, Volume collapsed three-manifolds with a lower curvature bound,
preprint, math.DG/0304472.
[36] S.-T. Yau, The role of partial differential equations in differential geometry, in ‘Proceedings
of the International Congress of Mathematicians’, Helsinki 1978, 237–250.
[37] S.-T. Yau, Structures of three-manifolds, preprint. math.DG/0607821
[38] R. Ye, Notes on the reduced volume and asymptotic Ricci solitons of κ-solutions, preprint.
Department of Mathematics, UCLA, Los Angeles, CA 90024

E-mail address: tao@@math.ucla.edu

Perelman's Proof of The Poincare Conjecture PDF

Uploaded by

Copyright:

Available Formats

Perelman's Proof of The Poincare Conjecture PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Perelman's Proof of The Poincare Conjecture PDF

Uploaded by

Copyright:

Available Formats

PERELMAN’S PROOF OF THE POINCARÉ CONJECTURE: A

NONLINEAR PDE PERSPECTIVE

Abstract. We discuss some of the key ideas of Perelman’s proof of Poincaré’s

state (or “scattering state”) as t → +∞, enough control on the topology of Mt

recently (in 2006), three essentially simultaneous expositions6 of Perelman’s work

in coordinates via the Christoffel symbol

The factor 2 is a traditional and convenient normalisation which is not of essen-

Thus this gives a gauge-transformed Ricci equation of the form

This turns out to be important later on when we construct “standard solutions”

5. Monotonicity formulae I: Lower bounds on curvature

22This is in contrast with Hamiltonian equations such as wave or Schrödinger equations, in

in the orthonormal frame coordinates24. These equations can be combined with a

6. Monotonicity formulae II: minimal surfaces and finite time

an energy Ẽ(γ, t) analogous to the energy considered earlier; maximising over t,

7. Monotonicity formulae III: Perelman’s reduced volume

to have some monotonicity

it is not geometrically natural for x0 to be placed in Mt0 −τ when it really should

(13) Ṽt0 ,x0 (τ ) ≈ τ −3/2 VolMt0 −τ (B(x0 , τ 1/2 )).

8. Blowup analysis I: a no-local-collapsing theorem

Proof. (Sketch) A rescaling√argument using the comparison theorem lets us reduce

9. Blowup analysis II: asymptotic profiles

10. Blowup analysis III: the geometry of ancient non-collapsed

by Perelman can be used to significantly reduce the possible geometries

10.1. Bounded curvature. We shall begin with an important result on κ-solutions

11. Blowup analysis III: approximation by ancient solutions

13. Standard solutions

theory is vaguely analogous to the role of a smooth cutoff function in harmonic

2n already. We shall need two surgery parameters, a small canonical neighbour-

• An ε-round component, diffeomorphic to a finite quotient of S 3 .

Department of Mathematics, UCLA, Los Angeles, CA 90024

You might also like