THE RICCI FLOW

TALHA ALI

Abstract. The Ricci Flow is a means of taking a Riemannian manifold, and

improving its geometry so that it is more “even”. This is done by evolving
the metric by the heat-type equation ∂t gij = −2Rij , which has the effect of
evening out the manifold’s curvature distribution, with the hope of finding
an “optimal” or smoothest possible metric. We begin by defining tensors on
manifolds, connections, and various notions of curvature and distance. We
study heat equations, and relate them to the Ricci flow equation. Finally we
consider the Einstein-Hilbert functional and show how the Ricci flow can be
considered a modified gradient flow for this functional.

1. Introduction

The Ricci flow is a geometric flow, which evolves the metric on a Riemannian
manifold in a manner similar to heat diffusion; this process results in the smoothing
of the irregularities in the metric. Given a Riemannian manifold M , with metric
tensor gij , me may compute its Ricci tensor Rij , which is an average of sectional
curvatures. If we further let the metric tensor be dependent upon an external
variable t, usually called time, then we may define the Ricci flow with the equation,
∂gij
= −2Rij .
∂t

We shall illustrate a quick example. If M̄ is a Riemannian manifold with

Einstein metric gij , then its Ricci tensor is given by Rij = Cgij for some C ∈ R. If
we evolve this metric via the Ricci flow then,
∂gij
= −2Rij
∂t
= C̃gij
We may solve this differential equation, to obtain that gij = Aij eC̃t where the Aij ’s
are constants. A special case of this example is the round metric on Sn , where
sectional curvature is +1, and Ricci curvature Rij = (n − 1)gij . From the above
calculation, we can deduce that evolving the round metric on Sn via the Ricci flow,
in the limit will reduce Sn to a single point. In particular we have,
∂gij
= −2Rij
∂t
= −2(n − 1)gij

Date: 5/8/10.
1
2 TALHA ALI

Solving this we obtain that, gij = Aij e(−2n+2)t , and for n ≥ 2,

lim gij (t) = 0.
t→∞
For general Einstein manifolds, depending upon the sign of the constant C, evolving
the metric via the Ricci flow may either reduce the manifold to a single point or
make it infinitely large.

We will show in the paper, that the complicated formula for the Ricci tensor
actually reduces to an operator whose highest order component is a Laplacian. This
allows us to relate the Ricci flow to the heat flow. In essence, if given a general
manifold the Ricci flow distributes the curvature on the manifold, in a similar
manner to the heat flow. Areas of “concentrated” curvature are dispersed, and
areas of little or no curvature gain curvature. As the metric evolves the manifold
in turn becomes more “even” and smooth.

But the Ricci flow equation is not exactly a heat equation, among other tech-
nicalities it is not linear. Thus, there is no analytic theory which guarantees the
existence of solutions defined for all time. Unlike the heat flow, singularities can
form which cause the curvature to increase unboundedly. The Ricci flow equation
is in turn a reaction-diffusion equation for curvature on a manifold. As the metric
tensor evolves via the Ricci flow, it behaves like a heat equation plus some lower
order terms. In some cases, the lower order terms are negligible, however in other
cases the lower order are not negligible and the resulting metric has singularities.

We begin this paper by introducing some fundamental concepts of Riemannian

geometry. In section 2, we will cover some elementary tensor algebra, the Levi-
Civita connection, and the path length and energy functionals. These concepts are
necessary to understand the rest of the paper.

In section 3, we will introduce the geodesic equation which allows the reader
to grasp the concept of the shortest distance between two points on a manifold. For
Euclidean flat space, this a fairly straightforward idea because the shortest distance
between two paths is a straight line. However the notion of a straight line does not
make sense in curved space, and thus the geodesic equation helps to describe the
minimizing path between two points.

We will continue by introducing some customary coordinate systems. In sec-

tion 4 we discuss the exponential normal coordinates. These coordinates are very
convenient because as we shall see, under certain conditions, the metric can be
reduced to the Euclidean metric. Section 5 introduces the harmonic coordinates,
which we will use to show that the Ricci flow is similar to the heat flow.

We discuss the heat flow in section 6. The heat flow is a linear partial differ-
ential equation that describes the distribution of heat over time. The last section
of the paper will discuss the evolution of various geometric quantities such as the
Riemann tensor. The essential “game” in the Ricci flow is to control how these
quantities evolve. We will close the paper by discussing the Einstein-Hilbert func-
tional, which relates the Ricci flow to a gradient flow.

Special thanks to Brian Weber in the construction of this paper.

 THE RICCI FLOW 3

2. Riemannian Geometry

2.1. Tensor Algebra. The tensor product is the formal multiplication of vectors.
Let V and W be vector spaces over the field R, and let α ∈ R be a scalar and
v, v 0 ∈ V and w, w0 ∈ W be vectors. We can define the tensor product between v
and w as the formal symbol v ⊗ w, subject to the following linearity conditions:
v ⊗ (αw) = α(v ⊗ w) = (αv) ⊗ w
0
v ⊗ (w + w ) = v ⊗ w + v ⊗ w0
(v + v 0 ) ⊗ w = v ⊗ w + v 0 ⊗ w.
We define the tensor product between vector spaces V and W , denoted V ⊗ W , as
the linear span of tensors of the form v ⊗ w. Namely,

V ⊗ W = spanR {v ⊗ wv ∈ V, w ∈ W }.
We define the dual V ∗ of a vector space V to be the space of linear functionals
Ni
f : V → R. The definition of tensor product can be iterated. We can define V,
Ni,k
and V respectively by setting
Oi
V = V ⊗ V ⊗ ··· ⊗ V (i many copies of V )
Oi,k Oi Ok
V = V ⊗ V ∗ = V ⊗ V ⊗ · · · ⊗ V ⊗ V ∗ ⊗ V ∗ ⊗ · · · ⊗ V ∗.

Given a manifold M , we define the tangent space at a point p ∈ M , denoted

Tp M , to be the space of all vectors based at p. The tangent bundle, denoted T M ,
is the collection of all tangent spaces of M . Namely,
[
TM = Tp M.
p∈M

2.2. The Connection. We can define a functional, the Levi-Civita Connection

∇ : T M × C ∞ (T M ) → C ∞ (T M )
by requiring that it satisfy the following axioms. For vector fields X, Y, Z, and
functions f, g the following are required
• ∇f X+gY Z = f ∇X Z + g∇Y Z
• ∇ X Y + Z = ∇X Y + ∇X Z
• ∇X f Y = X(f )Y + f ∇X Y
• XhY, Zi = h∇X Y, Zi + hY, ∇X Zi
• [X, Y ] = ∇X Y − ∇Y X
From these five axioms, it can be shown that the Koszul formula
2 h∇X Y, Zi = X hY, Zi+Y hX, Zi−Z hX, Y i+h[X, Y ], Zi−h[Y, Z], Xi+h[Z, X], Y i
holds. This implicitly defines the Levi-Civita Connection.

Also, we note that ∇ acts on ⊗n,k T M ,

On,k On,k
∇ : TM ⊗ TM → TM
4 TALHA ALI

by requiring that it obey the Leibnitz rule. Namely if T ∈ ⊗n,k T M and X ∈ T M ,

then ∇X T ∈ ⊗n,k T M . In particular
∇X T (η1 , η2 , . . . , ηn , X1 , . . . , Xk ) = X (T (η1 , η2 , . . . , ηn , X1 , . . . , Xk ))
+ T (∇X η1 , η2 , . . . , ηn , X1 , . . . , Xk )
+ ··· +
+ T (η1 , η2 , . . . , ηn , X1 , . . . , ∇X Xk ).

Choosing a local coordinate system {xi }, we may express the metric g as

g = gij dxi ⊗ dxj
where the gij = gij (x1 , x2 , . . . , xn ) are functions of the local coordinates. We can
define the Christoffel symbol Γkij as
1 ∂gjl ∂gil ∂gij lk
Γkij , ( + − )g .
2 ∂xi ∂xj ∂xl
Also, it can be shown from the Koszul formula that
∂ ∂
∇ ∂ i j = Γkij k
∂x ∂x ∂x
∇ ∂ i dxj = −Γjik dxk .
∂x
Nn,k Nn,k+1
Note that, with T ∈ , we have shown that ∇T ∈ . In coordinates, ∇T
has an additional lower index, which we demarcate with a comma. That is, if
...in ∂ ∂
T = Tji11ji22...j ⊗ · · · ⊗ in ⊗ dxj1 ⊗ · · · ⊗ dxjk ,
k
∂xi1 ∂x
then
∂ ∂
∇T = Tji11ji22...j
...in
k ,jk+1
⊗ · · · ⊗ in ⊗ dxj1 ⊗ · · · ⊗ dxjk ⊗ dxjk+1 .
∂xi1 ∂x
As an example, let T ∈ T ∗ M be a covector; we will compute the expression ∇T ∈
N0,2
in coordinates. By definition,
Ti,j = (∇∂/∂xj T )(∂/∂xi )
∂
= T (∂/∂xj ) + T (∇∂/∂xj ∂/∂xi )
∂xj
∂Tj
= + T (Γkij ∂/∂xk )
∂xi
∂Tj
= + Γkij Tk .
∂xi

2.3. Path Length and Energy. For a path

γ : (a, b) → M
we define the length functional L and energy functional E to be
Z br
dγ dγ
L(γ) = h , idt
a dt dt
Z b
dγ dγ
E(γ) = h , idt
a dt dt
 THE RICCI FLOW 5

Lemma 1. The L functional is invariant under reperimatrization. Specifically, if

f : [a, b] → R is a differentiable function with differentiable inverse, then L(γ(t)) =
L(γ(f (t))).

Proof. We start we with L(γ(s)), and perform a change of variables t = f −1 (s)

s
Z f (b) 
dγ dγ
L(γ(s)) = , ds
f (a) ds ds
Z f (b) s 
dt dγ dγ
= , ds
f (a) ds dt dt
Z f −1 (f (b)) s 
dγ dγ
= , dt
f −1 (f (a)) dt dt
Z b s 
dγ dγ
= , dt
a dt dt
= L(γ(t))

Theorem 2.1. Let γs : [a, b] → M be a smooth family of paths parameterized

by s, where |γ̇s | = 1, for every s ∈ (−, ). Then L(γ0 ) ≤ L(γs ) if and only if
E(γ0 ) ≤ E(γs ).

Proof. We show that E minimizers are L minimizers. Assume that E(γ0 ) ≤

E(γs ), and assume further that L(γ̄) ≤ L(γ0 ). We want to show that L(γ̄) = L(γ0 ).
Thus,
Z b q
L(γ̄) = ˙ γ̄idt
hγ̄, ˙
a
Z b
1
= p ˙ γ̄i
hγ̄, ˙ dt
a ˙ γ̄i
hγ̄, ˙
Z b
= ˙ γ̄i
hγ̄, ˙ dt
a
= E(γ̄)
≥ E(γ0 )
= L(γ0 )

Hence we have, L(γ̄) = L(γ0 ).

By a similar argument we show that L minimizers are E minimizers. Assume

that L(γ0 ) ≤ L(γs ), and assume further that E(γ̄) ≤ E(γ0 ). We want to show that
6 TALHA ALI

E(γ̄) = E(γ0 ). Therefore,

Z b
E(γ̄) = ˙ γ̄i
hγ̄, ˙ dt
a
Z b
1
= p ˙ γ̄i
hγ̄, ˙ dt
a ˙ γ̄i
hγ̄, ˙
Z b q
= ˙ γ̄idt
hγ̄, ˙
a
= L(γ̄)
≥ L(γ0 )
= E(γ0 )

Hence we have, E(γ̄) = E(γ0 ).

3. The Geodesic Equation

We take it for granted that the shortest path between two points is a straight
line. However, this notion only makes sense in Euclidean space, which happens to
be flat and simply connected. If we consider other geometries perhaps with nonzero
sectional curvature, then the shortest path between two points must to be described
in a different way. The following equation describes how this may be done.
Definition 1 (Geodesic). Given a manifold M , a path γ is called a is geodesic if
E(γ) = inf E(η),
η

where the infimum is taken over all paths η : [a, b] → M with η(a) = γ(a) and
η(b) = γ(b).

As we have proved in Theorem 2.1, minimizers of E also are minimizers of L.

Thus our geodesics realize the minimum distance between two points.
Theorem 3.1. Let γ : (−, ) → M be a path. Then γ is a geodesic if and only if
∇γ̇ γ̇ = 0.

Proof. Let γs (t), t ∈ (a, b) be smooth family of paths parametrized by s, such

that γs (a) = 0 and γs (b) = p, ∀s. We wish to find the conditions under which γs0
is the shortest path between the endpoints γs0 (a) = 0 and γs0 (b) = p.

There will be two vector fields involved in the calculation, if we fix t and vary s
d
we obtain the variation field ds ; likewise if fix s, and vary t we obtain the directional
d
field dt . Note that since the end points are fixed, there is no variation at the end
points. Theorem 2.1 tells us that E-minimizers and L-minimizers are the same up
to reparametrization, so we can minimize the E functional instead.
 THE RICCI FLOW 7

d

If γs0 is a geodesic then we must have E(γs0 ) ≤ E(γs ), which implies ds E(γs ) s=s0
 =
0. Computing this derivative, we get
Z b
d  d d d
E(γs ) = h , idt

ds s=s0 ds a dt dt
Z b Z b
d d d d d
= h , idt = 2 h5 d , idt
a ds dt dt a
ds dt dt

The final equality is just the compatibility of ∇ with the metric.

d d d d
We note that from the torsion free axiom we have 5 d dt − 5 d ds = [ dt , ds ],
ds dt
d d d d
however [ dt , ds ] = 0 since partials commute. Therefore 5 d dt = 5 d ds . Continu-
ds dt
ing our computation we have
Z b Z b
d  d d d d
E(γs ) = 2 , idt = 2
h5 d h5 d , idt

ds s=s0 a
ds dt dt
a
dt ds dt

Z b Z b
d d d d d
= 2 h , idt − 2 h , 5 d idt
a dt ds dt a ds dt dt
  Z b
d d b d d
= 2 ,  −2 h , 5 d idt
ds dt a a ds dt dt

Z b
d d
= 0−2 h , 5 d idt
a ds
dt dt

The first term is zero since we have assumed there is no variation at the end points,
so we have
Z b
d d d
E(γs )|s=s0 = −2 h , 5 d idt
ds a ds dt dt

Since γs0 is the shortest path between the end points, we must have
Z b
d d d
0= E(γs )|s=s0 = −2 h , 5 d idt
ds a ds dt dt
d d
Since this must hold for any variational field ds , we must have 5 d dt ≡ 0. Hence
dt
d
we have, γs0 is a geodesic if and only if 5 d dt ≡ 0.
dt

Theorem 3.2. Let {xi } be a local coordinate chart on M and let γ(t) = (x1 (t), x2 (t), . . . , xn (t))
be a path on M . Then γ is a geodesic if and only if

d2 xk dxi dxj
+ Γkij =0
dt dt dt

Proof. We start by noting,

dγ dxi ∂
=
dt dt ∂xi
8 TALHA ALI

If γ is a geodesic then

dγ
0 = ∇ dγ
dt dt
dxj ∂
= ∇ dxi ∂
dt ∂xi dt ∂xj
i
dx dxj ∂
= ∇ ∂i
dt ∂x dt ∂xj
dxi ∂ dxj ∂ dxi dxj ∂
= i j
+ ∇ ∂
dt ∂x dt ∂x dt dt ∂xi ∂xj
d dxk ∂ dxi dxj k ∂
= + Γ
dt dt ∂xk dt dt ij ∂xk
d2 xk dxi dxj k ∂
= ( 2 + Γ )
dt dt dt ij ∂xk

Hence we have
d2 xk dxi dxj
+ Γkij =0
dt dt dt

Theorem 3.3 (Existence of solutions). Given an initial condition, γ(a) ∈ M , and

a terminal condition γ(b) ∈ M , there exists a solution γ : [a, b] → M of the geodesic
equation.

Proof. This follows from the existence theorems of ordinary differential equa-
tions.

Example 1. Reparametrizing a geodesic by a constant is also a geodesic, in partic-

ular, if γv : (−, ) → M is a geodesic parametrized by t, then for s > 0 the curve
c : ( − 
s , s ) → M defined by c(t) = γv (st) is also a geodesic. In particular we have,

5ċ ċ = s2 5γ̇v γ̇v = 0.

Example 2. Consider Euclidean space Rn with the ordinary rectangular coordinates

{xi } and metric g(., .) = dx1 ⊗ dx1 + dx2 ⊗ dx2 + · · · + dxn ⊗ dxn . Our goal is to
find the geodesics of Rn . We may do this by solving the geodesic equation, with
gij = δij we compute.

1 ∂gjl ∂gil ∂gij lk

Γkij = ( + − )g ≡ 0
2 ∂xi ∂xj ∂xl
Hence we have
d2 xk
= 0 ⇒ xk (t) = Ck1 t + Ck2 ,
dt2
 THE RICCI FLOW 9

which implies
 
C11 t + C12
 C21 t + C22 
x(t) = 
 
.. 
 . 
Cn1 t + Cn2
Given initial conditions we may solve explicitly for C11 , C12 , C21 , C22 , ..., Cn1 , Cn2 .
We note that the solutions do indeed make sense because we would expect the
geodesics of Rn to be straight lines.

Example 3. Consider S2 the 2-sphere defined by the equation x2 + y 2 + z 2 = 1, our

goal is to find the geodesics. We may do this by solving the geodesic equation, but
first we need to put coordinates on the sphere and find the corresponding metric.
We can put coordinates on S2 via the map
x y
φ : S2 → R2 such that φ(x, y, z) = ( , ) = (a, b).
1−z 1−z

These are known as stereographic coordinates. Note that (x, y, z) are coordi-
nates on R3 , and (a, b) are stereographic coordinates on the sphere on R2 . It is clear
that we may take any point on the sphere except for the north pole and locate a
corresponding point on R2 via φ, the north pole is the point at infinity. φ is clearly
a bijection if we restrict its domain to all points except the north pole. It’s inverse
is,
2a 2b a2 + b2 − 1
φ−1 : R2 → S2 such that φ−1 (a, b) = ( , , ) = (x, y, z)
a2 + b2 + 1 a2 + b2 + 1 a2 + b2 + 1
and
N = lim φ−1 (a, b) = lim φ−1 (a, b) = lim φ−1 (a, b)
a→∞ b→∞ (a,b)→(∞,∞)

The coordinate fields are:

∂ ∂x ∂y ∂z
= ∂x + ∂y + ∂z = (1 − z − x2 )∂x + (−xy)∂y + (x − xz)∂z
∂a ∂a ∂a ∂a
∂ ∂x ∂y ∂z
= ∂x + ∂y + ∂z = (−xy)∂x + (1 − z − y 2 )∂y + (y − yz)∂z
∂b ∂b ∂b ∂b

Thus we can compute the metric by plugging in:

∂ ∂ 4
g( , ) =
∂a ∂a (a2 + b2 + 1)2
∂ ∂
g( , ) = 0
∂a ∂b
∂ ∂ 4
g( , ) =
∂b ∂b (a2 + b2 + 1)2
So the metric on the sphere is
4
g(·, ·) = (da ⊗ da + db ⊗ db),
(a2 + b2 + 1)2
10 TALHA ALI

the Christoffel symbols are

 
1 −2a −2b
Γ1ij =
a2 + b2 + 1 −2b 2a
 
1 2b −2a
Γ2ij
= 2 ,
a + b2 + 1 −2a −2b
and the Geodesic equation is:
d2 γ 1 1 dγ 2
1 2
1 dγ dγ
1
1 dγ 2
2
+ Γ 11 ( ) + 2Γ 12 + Γ 22 ( ) =0
dt2 dt dt dt dt
d2 γ 2 dγ 1 2 dγ 2 dγ 1 dγ 2 2
2
+ Γ211 ( ) + 2Γ212 + Γ222 ( ) =0
dt dt dt dt dt
Clearly, this a second order nonlinear system of differential equations, the general
solution to which will not be easily obtained. It is important to remember that
we have projected S2 onto R2 , hence the solutions to the above equations will be
on R2 . Knowing that the geodesics of the 2-sphere are great circles, we expect
the solutions to this system to be circles which are not necessarily unit and not
necessarily centered at the origin. We also expect exactly one solution to be the
unit circle, and we also expect lines passing through the origin to be geodesics
although not with the standard parametrization.

4. Exponential Map

The theorems for ODE’s guarantee the existence and uniqueness of solutions
to the geodesic equation, however the solutions may only exist for a very short
period of time.
Theorem 4.1. Given a Riemannian Manifold M, point p ∈ M and vector v ∈
Tp M , there is a unique geodesic γv : [0, ∞] → M with γ(0) = p, and γ̇(0) = v.

Proof. This follows from the existence and uniqueness theorems for ODE’s.


Definition 2 (Exponential Map). Given a Riemannian manifold M and point
p ∈ M , we define the exponential map centered at point p to be
expp : Tp M → M
such that expp (v) = γv (1) where γv the geodesic with γ(0) = p, and γ̇(0) = v.

We should note that exp is smooth, and defined globally. One should note that for
c > 0, expp (cv) = γcv (1) = γv (c).
Definition 3 (Exponential Normal Coordinates). Given a Riemannian Manifold
M and point p ∈ M , if we choose an orthonormal basis vi for Tp M , we can assign
the point expp (q i vi ) the coordinates (q1 , q2 , ..., qn ).

This construction only makes sense for a region S ⊆ Tp M , around the origin
of Tp M for which expp : S → M is diffeomorphism onto its image.
 THE RICCI FLOW 11

Lemma 2 (Gauss Lemma). Given a Riemannian manifold M, and point p ∈ M , if

ρ(t) = tv is a ray through the origin of Tp M and w ∈ Tρ(t) Tp M is perpendicular to
ρ0 (t) then dexpp : Tρ(t) Tp M → Tp M , then dexp(w) is perpendicular to dexp(ρ0 (t)).

The Gauss Lemma says that given a radial vector v in the tangent space, and
w orthogonal to v also in the tangent space, after applying the exp mapping, they
must still be orthogonal. We should note that this only holds true if v is a radial
vector, in general the exp map does not preserve orthogonality.

Corollary 1. If we base the normal coordinates at a point p ∈ M , then

∂ ∂
g( i
, j )|p = δij
∂x ∂x
and
∂gij
|p = 0
∂xk

This is useful because we can always put exponential normal coordinates at a

certain point and simplify the calculations, we shall use this fact through out the
paper.

Example 4. We shall put exponential normal coordinates on S2 the 2-sphere. We

choose to base the coordinates at the north Pole N = (0, 0, 1) ∈ R3 , and choose
the ordinary orthonormal vectors (e1 , e2 ) as a basis for R¯2 . Note that R¯2 = TN M .
If we choose v ∈ TN M such that v = (a0 , b0 ) = a0 e1 + b0 e2 .pWe may put (θ, φ)
coordinates on the sphere, where θ = tan−1 ( ab00 ) and φ = (a20 + b20 ), where θ
represents the angle measured from the positive x-axis and φ represents the angle
measure from the positive z-axis. These are exponential normal coordinates on the
sphere because v ∈ TN M is clearly a geodesic and expN (v) is obtained by mapping
the vector v onto a great circle in the direction θ and of length φ . We may also note
that if v ∈ TN M and |v| = 2nπ then expN (v) = (0, 0) = N , and if |v| = (2n + 1) π2
then expN (v) = (θ, π2 ) = S.

5. Harmonic Coordinates

5.1. N
The Laplacian.N If X, and Y are vector fields, and f is a function, that is
0,0 0,1
f∈ , then ∇f ∈ , in particular

∇f (X) = ∇X f = X(f )
N0,2
Likewise, ∇(∇f ) ∈ , in particular we have

(∇∇f )(X, Y ) = (∇Y ∇f )(X)

= (∇Y (∇f (X))) − ∇f ∇Y X
12 TALHA ALI

In coordinates we have,
∂ ∂
f,ij = (∇∇f )( , )
∂xi ∂xj
∂ ∂
= (∇ ∂ j (∇f ( i )))) − ∇f ∇ ∂ j
∂x ∂x ∂x ∂xj

∂2f ∂
= i j
− Γkij k
∂x x ∂x
Definition 4 (The Laplacian). Given the metric g, we define the Laplacian ∆ of
function f to be
∂2f ∂
∆f , g ij f,ij = g ij i j − g ij Γkij k
∂x x ∂x

5.2. Harmonic Coordinates.

Definition 5 (Harmonic Coordinates). A coordinate chart {x1 , x2 , . . . , xn } on a
Riemannian Manifold M with metric g, is called harmonic if ∆xi = 0 for i =
1, 2, . . . , n.
Theorem 5.1. Given a Riemannian manifold M with metric g and a local co-
ordinate chart {x1 , x2 , . . . , xn } on M , we have ∆xk = −g ij Γkij . In particular a
coordinate function xk is harmonic if and only if g ij Γkij = 0.

Proof. Note that for any coordinate function xk , we have

∂ 2 xk
=0
∂xi ∂xj
Therefore,
∂ 2 xk ij p ∂x
k
∆xk = g ij − g Γ ij
∂xi ∂xj ∂xp
k
∂x
= −g ij Γpij p
∂x
= −g ij Γpij δpk
= −g ij Γkij
If xk is harmonic, it follows
0 = ∆xk
= −g ij Γkij
Therefore, we must have g ij Γkij = 0


Theorem 5.2 (The existence of Harmonic Coordinates [4]). Given a Riemannian
manifold M with metric g, and let g be of class C k,α (for k ≥ 1) or respectively C ω
in a local coordinate chart about a neighborhood of some point p. Then there is a
neighborhood of p in which the harmonic coordinates exist, these new coordinates
being C k+1,α or respectively C ω functions of the original coordinates. Moreover, all
harmonic coordinate charts defined near p have this regularity.
 THE RICCI FLOW 13

Proof. Existence and regularity theorems for second order partial differential
equations are beyond the scope of this paper. We refer the reader to Kazdan-Deturk
[4], lemma 1.2 for a formal proof.

The above two theorems give a sufficient condition for harmonic coordinates
and there existence.

Example 5. Consider Euclidean 2-space R2 with the ordinary orthonormal coordi-

nates (x, y). Then the Laplacian ∆ is given by
∂2 ∂2
∆= 2
+ 2.
∂x ∂y
This implies
∂2x ∂2x
∆x = + 2
∂x2 ∂y
= 0
and
∂2y ∂2y
∆y = +
∂x2 ∂y 2
= 0
Hence, the coordinate chart (x, y) is harmonic.

We can reconsider Euclidean 2-space R2 , with the ordinary polar

p coordinates
(r, θ). Then in relation to the rectangular coordinates, we have r = x2 + y 2 , and
θ = arctan( xy ). This implies
p
∆r = ∆ x2 + y 2
p p
∂ 2 x2 + y 2 ∂ 2 x2 + y 2
= +
∂x2 ∂y 2
y2 x2
= 3 + 3
2 2
(x + y ) 2 (x + y 2 ) 2
2

1
=
r
6= 0
and
y
∆θ = ∆ arctan
x
∂ 2 (arctan xy ) ∂ 2 (arctan xy )
= +
∂x2 ∂y 2
−2xy 2xy
= + 2
(x2 + y 2 )2 (x + y 2 )2
= 0

Hence, the coordinate chart (r, θ) is not harmonic since ∆r 6= 0.

14 TALHA ALI

Theorem 5.3. Given a manifold M with a metric g that is at least C 2 . Then in

harmonic coordinates, we have
1
Rij = − ∆gij + Q(gij , ∂k gij )
2
∂gij
Where the Q(gij , ∂xk
) indicates some polynomial expression in the functions gij
∂g
and ∂xijk .

Proof. By the formula for the ricci curvature tensor, we have

Rij = ∂k Γkij − ∂i Γkkj + Γkij Γssk + Γksj Γsik

However, the last two terms involve one derivative of the metric so we may
write
∂gij
Rij = ∂k Γkij − ∂i Γkkj + Q(gij , )
∂xk

Calculating further, and leaving out the first order terms we have
∂gij
Rij = ∂k Γkij − ∂i Γkkj + Q(gij , )
∂xk
1 1 ∂gij
= ∂k ( g kl (∂i gjl + ∂j gil − ∂l gij ) − ∂i ( g kl (∂j gkl + ∂k gjl − ∂l gkj ) + Q(gij , )
2 2 ∂xk
2 2 2 2 2 2
1 kl ∂ gjl ∂ gil ∂ gij ∂ gkl ∂ gjl ∂ gkj ∂gij
= g ( k i+ k j
− k l
− i j
− i k
+ i l
) + Q(gij , )
2 ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂xk
1 ∂ 2 gij 1 ∂ 2 gil ∂ 2 gkl ∂ 2 gkj ∂gij
= − g kl k l + g kl ( k j − i j
+ ) + Q(gij , )
2 ∂x ∂x 2 ∂x ∂x ∂x ∂x ∂xi ∂xl ∂xk

Also, if we let Γk = g ij Γkij , we can compute

∂ 2 gil ∂Γk 1 kl ∂ 2 gkl ∂gij
g kl = gki + g + Q(gij , )
∂xk ∂xj ∂xj 2 i
∂x ∂x j ∂xk
and
2
∂ gkj ∂Γk 1 ∂ 2 gkl ∂gij
g kl = gkj + g kl i j + Q(gij , )
∂xi ∂xl ∂xi 2 ∂x ∂x ∂xk

Substituting into (4) we have,

1 ∂ 2 gij 1 ∂Γk ∂Γk ∂gij
Rij = − g kl k l + (gki j + gkj i ) + Q(gij , )
2 ∂x ∂x 2 ∂x ∂x ∂xk

Since we are working in harmonic coordinates, we have Γk = g ij Γkij = 0, hence

we have
1 ∂ 2 gij 1 ∂Γk ∂Γk ∂gij
Rij = − g kl k l + (gki j + gkj i ) + Q(gij , )
2 ∂x ∂x 2 ∂x ∂x ∂xk
1 ∂ 2 gij ∂gij
= − g kl k l + 0 + Q(gij , )
2 ∂x ∂x ∂xk
1 ∂gij
= − ∆gij + Q(gij , )
2 ∂xk
 THE RICCI FLOW 15

6. The Heat equation

We study the heat flow briefly because it helps us to understand the Ricci
flow. We shall start by describing the heat flow, and then derive the fundamental
solution.

For t ∈ R, and x ∈ U ⊂ Rn , where U is an open subset, and variable u :

U × [0, ∞) → R is the temperature at a point x at time t and f : U × [0, ∞) → R
is a given function.
Definition 6. Homogeneous Heat Problem

∂u
= ∆u
∂t
Definition 7. Nonhomogenuous Heat Problem

∂u
= ∆u + f
∂t

The Laplacian is with respect to the spatial coordinates {x1 , x2 , ...., xn } ∈ Rn .

6.1. Physical interpretation. This equation maybe derived from Fourier’s law.
If W ⊂ U is any smooth subregion, the rate of change of the total quantity within
W equals the negative of the net flux through ∂W .

Z Z
d
u dx = − F · N dS
dt W ∂W
From Gauss’s divergence theorem
Z Z
− F · N dS = − div F dV.
∂W W
Where F is the heat current density. Therefore,
∂u
(1) = −div F.
∂t
Since heat flows from higher concentration to lower concentration, F is proportional
to the gradient of u, however it points in the opposite direction. Hence, for a > 0
we have
F = −a∇u.
Taking a divergence and using 1, we have
∂u
= adiv(∇u) = a∆u.
∂t
Setting a = 1, we obtain the heat equation
∂u
= ∆u.
∂t
16 TALHA ALI

6.2. Fundamental Solution. The heat equation involves one derivative with re-
spect to time t and two derivates with respect to the spatial variable x ∈ Rn .
If u solves the heat equation then so does u(x, t) = u(λx, λ2 t) for some constant
λ ∈ R. This indicates that we are looking for a solution which is invariant under
the dilation scaling
(2) u(x, t) = λα u(λβ x, λt), ∀λ > 0, x ∈ Rn , t > 0
We want to look for a solution with a certain form; since t > 0, we can let λ = t−1
and v(y) = u(y, 1). Plugging into equation (2),
(3) u(x, t)λα u(λβ x, λt)
=
1 x
(4) = α u( β , 1)
t t
1 x
(5) = α v( β ).
t t
x
Inserting (5) into the homogeneous heat equation and letting y = tβ
, we have
∂u
(6) 0 = − ∆u
∂t
∂v
(7) = − ∆v
∂t
α βy ∂v(y) 1
(8) = − α+1 v(y) − α+1 − α+2β ∆v(y).
t t ∂t t
If we let β = 21 , then we can further simplify (8) to
1 1 ∂v(y)
0=− (αv(y) + y + ∆v(y))
tα+1 2 ∂t
which implies that
1 ∂v(y)
0 = αv(y) + y + ∆v(y).
2 ∂t
We simplify further by guessing that v is radial, meaning v only depends on the
distance from the origin |y|. So we let v(y) = w(|y|) = w(r), for some function
w : R → R. Therefore we have
r ∂w ∂ 2 w n − 1 ∂w
0 = αw(r) + + +
2 ∂r ∂r2 r ∂r
Note that since the laplacian is with respect to the spatial variables (x1 , x2 , . . . , xn ) ∈
∂2 n−1 ∂ n
Rn , we have ∆ = ∂r 2 + r ∂r . We reduce further by letting α = 2 . We have

n r ∂w ∂ 2 w n − 1 ∂w
w(r) +
0= + + .
2 2 ∂r ∂r2 r ∂r
We then multiply through by rn−1 ,
n n−1 rn ∂w ∂2w ∂w
(9) 0 = r w(r) + + rn−1 2 + (n − 1)rn−2
2 2 ∂r ∂r ∂r
∂ 1 n n−1 ∂w
(10) = ( r w(r) + r ).
∂r 2 ∂r
After we integrate (10), for some constant a, we have
1 n ∂w
a= r w(r) + rn−1 .
2 ∂r
 THE RICCI FLOW 17

Assuming further that the limr→∞ w(r) = 0, and limr→∞ w0 (r) = 0, we can con-
clude that a = 0. Therefore we have
dw 1
= − rw,
dr 2
which is a linear ordinary differential equation. We may solve by separation of
variables, to get
r2
w(r) = Ce− 4

1 x
for some constant C. Applying our assumptions, u(x, t) = tα v( tβ ), β = 12 , α = n
2
and r = |y| = | txβ |, we can conclude that
|x|2
u(x, t) = Ce− 4t

solves the homogenous heat equation.

Definition 8 (Fundamental Solution). Let Φ(x, t) be the function

 2
|x|

 1
(4πt)n/2
e− 4t x ∈ Rn , t > 0
Φ(x, t) =

0 x ∈ Rn , t < 0


This satisfies the heat equation and is known as the fundamental solution.

The choice of the constant is reflected in the following lemma.

Lemma 3 (Integral of fundamental solution). For every t > 0, and x ∈ Rn

Z
Φ(x, t)dx = 1
Rn

Proof. Note that |x|2 = (x1 )2 + (x2 )2 + (x3 )2 + · · · + (xn )2 , and dx =

dx dx2 dx3 . . . dxn . We compute
1
Z Z
1 |x|2
− 4t
Φ(x, t)dx = n/2
e dx
Rn Rn (4πt)
Z ∞ Z ∞
1 −
(x1 )2 +(x2 )2 ···+(xn )2
= . . . e 4t dx1 . . . dxn
(4πt)n/2 −∞ −∞
n Z ∞
1 Y (xi )2
= n/2
e− 4t dxi
(4πt) i=1 −∞
n Z ∞
1 Y i 2 1
= e−(z ) 2t 2 dz i
(4πt)n/2 i=1 −∞
n Z
1 Y ∞ −(zi )2 i
= e dz = 1.
π n/2 i=1 −∞


18 TALHA ALI

Definition 9 (Dirac Delta Function). We define the Dirac delta function, denoted
δx , as 
 ∞ x=0
δx =
0 x 6= 0

and constrained to Z ∞
δx dx = 1.
−∞

According to lemma 3 since the integral over the spacial variable x of the
fundamental solution Φ is 1, and if we take a limit as t → 0, then we may view the
fundamental solution as

 2
|x|

 1
(4πt)n/2
e− 4t t>0
Φ(x, t) =

δx t=0


6.3. Initial Value Problem.

Definition 10 (Convolution). We define the convolution of functions f and g,
denoted f ∗ g as Z ∞
(f ∗ g)(x) = f (y)g(x − y)dy.
−∞

The convolution has some important properties,

f ∗g =g∗f
(f ∗ g) ∗ h = f ∗ (g ∗ h)
f ∗ (g + h) = f ∗ g + f ∗ h
a(f ∗ g) = af ∗ g = f ∗ ag
d df dg
(f ∗ g) = ∗g =f ∗
dx dx dx
f ∗ δx = f
Theorem 6.1 (Solution to initial value problem). Let g ∈ C(Rn ) ∩ L∞ (Rn ) and
t > 0, then Z
u(x, t) = (Φ ∗ g)(x) = Φ(x − y, t)g(y)dy
Rn
also solves the heat equation.

∂Φ
Proof. Note that since Φ solves the heat equation, we have ∂t = ∆Φ. There-
fore,
∂ ∂Φ
(Φ ∗ g) = ( ∗ g)
∂t ∂t
= (∆Φ ∗ g) = ∆(Φ ∗ g)


 THE RICCI FLOW 19

Theorem 6.1 provides a systematic way to obtain solutions to the heat problem,
if we are given the initial value problem
 ∂u
 ∂t = ∆u in Rn × (0, ∞)

u=g in Rn × {t = 0}


we may obtain a solution by taking the convolution Φ∗g. We verify this construction
is valid. At t = 0, Φ = δx , therefore we have u = δx ∗ g = g. For t > 0 we may
apply Theorem 6.1.

7. The Einstein-Hilbert Functional

The proof of the following theorem illustrates a nice trick in computing the
evolution of various geometric quantities.
Theorem 7.1 (Variation of Christoffel symbols). If g(t) is a one parameter family
d
of metrics with dt gij = hij , then
d k 1
(11) Γij = g kl (hjl,i + hil,j − hij,l )
dt 2

Proof. We compute in normal  coordinates

 centered at an arbitrarily chosen
point p ∈ M . Note that ∂x∂ k gij p = 0, and Γkij p = 0. Therefore at p we have,
 
d k d 1 ∂gjl ∂gil ∂gij lk
Γ = ( + − )g )
dt ij dt 2 ∂xi ∂xj ∂xl
   
1 ∂ d ∂ d ∂ d lk 1 ∂gjl ∂gil ∂gij d lk
= gjl + gil − gij g + + − g
2 ∂xi dt ∂xj dt ∂xk dt 2 ∂xi ∂xj ∂xl dt
 
1 kl ∂hjl ∂hil ∂hij
= g + − +0
2 ∂xi ∂xj ∂xl
Since we have chosen normal coordinates, at p the Christoffel symbols are zero,
∂
thus at p we have hij,k = ∂x j hij .

We note that since both sides of (11) are components of tensors, equation (11)
is in fact a tensor equation, and therefore it holds for all coordinate systems, not
just normal coordinates.


Theorem 7.2 (Variation of Riemann curvature tensor). If g(t) is a one parameter
d
family of metrics with dt gij = hij , then we have
d l
R = Γ̇ljk,i − Γ̇lik,j .
dt ijk

Proof. By the ordinary formula for the Riemann curvature tensor we have
∂ l ∂ l
l
Rijk = Γ − Γ + Γpjk Γlip − Γpik Γljp .
∂xi jk ∂xj ik
20 TALHA ALI

By a direct calculation we have,

d l d ∂ l d ∂ l
R = Γ − Γ +
dt ijk dt
 ∂x i jk
 dt ∂xj ik     
d p d l d p d l
Γjk Γlip + Γpjk Γip − Γik Γljp − Γpik Γjp
dt dt dt dt
In normal coordinates centered at an arbitrarily chosen point p ∈ M , we have
Γkij p = 0. Therefore at p we have,
d l d ∂ l d ∂ l
R = Γ − Γ
dt ijk dt ∂xi jk dt ∂xj ik
∂ d l ∂ d l
= Γ − Γ
∂xi dt jk ∂xj dt ik
d l d
= Γ − Γl
dt jk,i dt ik,j
Since this is a tensor equation, it holds for all coordinate charts.


Theorem 7.3 (Variation of Ricci curvature tensor). If g(t) is a one parameter
d
family of metrics with dt gij = hij , then
d 1 1
Rij = g lp (hjl,ip + hil,jp − hij,lp ) − H,ij ,
dt 2 2
where H = g ij hij .

Proof. By Theorem 7.2, and the ordinary definition of the Ricci tensor we
have,
d ∂ d p ∂ d p
(12) Rij = Γij − Γ .
dt p
∂x dt ∂xi dt pj
Substituting (11), we have
d ∂ d p ∂ d p
Rij = p
Γij − i
Γpj
dt ∂x dt  ∂x dt    
∂ 1 pl ∂hjl ∂hil ∂hij ∂ 1 pl ∂hpl ∂hjl ∂hpj
= g + − − g + −
∂xp 2 ∂xi ∂xj ∂xl ∂xi 2 ∂xj ∂xp ∂xl
 2
∂ 2 hil ∂ 2 hij ∂ 2 hpl
  
1 pl ∂ hpj 1
= g l i
+ j p
− l p
− g pl
2 ∂x ∂x ∂x ∂x ∂x ∂x 2 ∂xi ∂xj
Computing in normal coordinates centered at an arbitrary point p ∈ M , we have
g ij |p = δ ij , therefore at p
1 ∂2H
 
d 1 ∂ ∂hjl ∂hil ∂hij
Rij = + − −
dt 2 ∂xl ∂xi ∂xj ∂xl 2 ∂xi ∂xj
1 1
= (hjl,il + ∂hil,jl − hij,ll ) − H,ij
2 2
Since this a tensor equation, it holds for all coordinate systems.


 THE RICCI FLOW 21

Theorem 7.4 (Variation of Scalar curvature tensor). If g(t) is a one parameter

d
family of metrics with dt gij = hij , then
d
R = −∆H + g lp hil,pj − hij Rij
dt
where H = g ij hij .

Proof. By the ordinary formula for R and Theorem 7.3,

d d ij
R = g Rij
dt dt
   
d ij ij d
= g Rij + g Rij
dt dt
1 ∂2H
   
d ij 1 ∂ ∂hjl ∂hil ∂hij
= − gij Rij + g + − −
dt 2 ∂xl ∂xi ∂xj ∂xl 2 ∂xi ∂xj
2
∂ hil
= −hij Rij + − ∆H
∂xl ∂xj

Definition 11 (Adjugate). For a given matrix A, with entries Aij , the adjugate
matrix of A, denoted adj(A) = A,
e with entries A
eij is defined as

eij = (−1)i−j det(A without its j th row and ith column)

A

Lemma 4. From any invertible matrix A, such that the entries Aij (t) are functions
of t, we have
d dA −1
det(A) = det(A) tr( A )
dt dt

Proof. For the ordinary formula of the determinant we have,

det(A) = i1 i2 ...in A1i1 A2i2 . . . Anin ,

where the  is the anti symmetric symbol. From this formula we can compute,
d d i1 i2 ...in
det(A) =  A1i1 A2i2 . . . Anin
dt dt
dA1i1 dA2i2 dAnin
= i1 i2 ...in ( A2i2 . . . Anin + A1i1 . . . Anin + · · · + A1i1 A2i2 . . . )
dt dt dt
X X dAij dA e
= A
g jk = tr ( A)
j
dt dt
i=k

Where A
e is the adjugate matrix of A. From Cramer’s rule it can be shown,

1 e
A−1 = A.
det A
22 TALHA ALI

Thus we have,
d dA e
det(A) = tr ( A)
dt dt
dA
= tr ( det(A)A−1 )
dt
dA −1
= det(A) tr( A )
dt


Theorem 7.5 (Variation of the volume form). Let {xi } be a positively oriented
local coordinate chart on a Riemannian manifold M , and let gij (t), be the metric
on M at time t. Then
d 1
dV = HdV,
dt 2
ij d
where H = g hij , and hij = dt gij .

Proof. From the ordinary formula for dV , we have

q
dV = det(gij )dx1 ∧ dx2 ∧ · · · ∧ dxn ,

and from lemma 4 we have

d dgsr
det(gij ) = det(gij )g sr
dt dt
= det(gij )g sr hsr = det(gij )H
we therefore can compute,
d d
q
dV = det(gij )dx1 ∧ dx2 ∧ · · · ∧ dxn
dt dt
 q 
d d
q
= det(gij ) dx1 ∧ dx2 ∧ · · · ∧ dxn + det(gij ) dx1 ∧ dx2 ∧ · · · ∧ dxn
dt dt
 
1 d
= p det(gij ) dx1 ∧ dx2 ∧ · · · ∧ dxn + 0
2 det(gij ) dt
1
= p det(gij )Hdx1 ∧ dx2 ∧ · · · ∧ dxn
2 det(gij )
1
q
= H det(gij )dx1 ∧ dx2 ∧ · · · ∧ dxn
2
1
= HdV
2
Definition 12 (Einstein-Hilbert functional). Let M be an oriented Riemannian
manifold without boundary and with metric g. We define the Einstein-Hilbert (total
scalar curvature) functional,
Z
L(g) = − R dV,
M

where R is the scalar curvature.

 THE RICCI FLOW 23

Theorem 7.6 (Variation of Einstein-Hilbert functional). Let M be an oriented

Riemannian manifold without boundary and with time depended metric g. Then
d
for dt gij = hij and H = g ij hij ,
Z  
d 1
(13) L=− h, Rg − Ric dV.
dt M 2

Proof. By a direct calculation

Z
d d
L = − R dV
dt dt M
Z
d
= − R dV
dt
ZM
d d
= − ( R)dV + R( dV )
dt dt
ZM
1
= − (−∆H + ∇i ∇j hij − hh, Rici)dV + ( RH) dV
2
ZM
1
= − −∆H + ∇i ∇j hij − hh, Rici + RH dV
2
ZM
1
= − −∆H + ∇i ∇j hij − hh, Rici + R hh, gi dV
2
ZM  
1
= − −∆H + ∇i ∇j hij + h, Rg − Ric dV
M 2

From the divergence theorem, we have

Z Z Z
∆HdV = div(∇H)dV = h∇H, νi dσ,
M M ∂M
and Z Z Z
∇i ∇j hij dV = div(div(h))dV = hdiv(h), νi dσ,
M M ∂M
where ν is the unit outward normal and dσ is the volume form on ∂M . Since we
assumed that M was a Riemannian manifold without boundary, we have
Z Z
∆HdV = h∇H, νi dσ = 0
M
Z Z∂M
∇i ∇j hij dV = hdiv(h), νi dσ = 0
M ∂M
Therefore we can conclude,
Z  
d 1
L= h, − Rg + Ric dV.
dt M 2

It is possible to relate the geometry of Riemannian manifolds to the physics of

General Relativity. General Relativity, Einstein’s
R geometrical theory of gravitation,
postulates a total Lagrangian of L(gij ) = (L − R)dV ol, where L is the matter
Lagrangian. The physical postulate is that the universe’s metric minimizes this
24 TALHA ALI

R
Lagrangian. We have already computed
R the Euler-Lagrange equation for (−R);
the Euler-Lagrange equation for L is Tij =R 0, where Tij is the stress-energy tensor
of field theory. The statement that L(g) = L − R is extremized is precisely that
1
Tij − Rgij + Ricij = 0.
2
This system of n2 equations is collectively known as Einstein’s Field Equations.

In Riemannian geometry, we have no matter fields (so L ≡ 0). In analogy

with physics however, we would like our manifolds to arrange themselvesR in an
optimum configuration, which corresponds to a minimum of the functional (−R).
This is obtained when − 12 Rg + Ric = 0. Given a manifold M n , it would seem to be
too difficult a problem to determine such an optimum metric. However, given an
arbitrary metric g on M , we can change g in the direction of an optimum metric.
That is, if we want to evolve the metric in such a way that L(g) decreases as quickly
as possibly, then equation (13) indicates we should let dgdt be a negative multiple of
− 21 Rg + Ric. We should therefore try to evolve the metric via the PDE
 
dgij 1
= −C − Rgij + Ricij
dt 2
where C > 0. Unfortunately, this PDE is not parabolic! One possible modification
is to remove the Ricci curvature term, giving ġij = Rgij , the equation for the
Yamabe flow. Another possible modification is to remove the scalar curvature term,
giving us the Ricci flow:
dgij
= −2Ricij .
dt
The idea behind the Ricci flow, therefore, is to try to approximate the gradient flow
for the Einstein-Hilbert functional. An actual minimizer of the functional would be
considered (via our analogy with General Relativity) to be an optimal geometrical
configuration for the manifold under consideration. The Ricci flow is an attempt
to move from an initial, non-ideal metric to an ideal metric by approximating the
gradient flow of this functional.

References
[1] B. Chow, P. Lu, L. Ni Hamilton’s Ricci Flow. American Mathematical Society Science Press,
Providence, Rhode Island, 2006
[2] J. Cheeger, D. Ebin Comparison Theorems in Riemannian Geometry.American Mathematical
Society Science Press, Providence, Rhode Island, 2008
[3] M. doCarmo Riemannian GeometryBirkhäuser, Boston, Massachusetts, 1992
[4] D. DeTurck and J. Kazdan, Some regularity theorems in Riemannian geometry, Annales sci-
entifiques de l École Norm. Sup.14 no. 3 (1981) 249–260