NOTES IN GEOMETRIC ANALYSIS

PROFESSOR ZHIQIN LU

Introduction
Let’s state the most important theorem in geometric analysis:
Theorem 0.1. All theorems in geometric analysis are also theorems in Calculus.
In Calculus, we study complicated function theory on relatively simple space: the Euclidean spaces.
In topology, we study relative simple function theory on complicated spaces. In differentially, we
create a kind of Calculus that takes the underlying topological spaces into account. As a consequence,
we have the following
1. Inevitably, the study of geometric analysis will have implications in topology;
2. We only study those Calculus properties that are covariant with the choices of local coordinates.
In this lecture, we will fully conform the above philosophy.

1. The Comparison theorems

Let X be a Riemannian manifold with the Riemannian metric ds2 . Let ω1 , · · · , ωn be one forms
such that
ds2 = ω12 + · · · + ωn2
Then by the theorem of Cartan, there are one forms {ωij } with ωij = −ωji , called the connection
forms, such that
dωi = −ωij ∧ ωj
1
dωij = −ωik ∧ ωkj + Rijkl ωk ∧ ωl
2
The tensor Rijkl is called the curvature tensor of the Riemannian metric ds2 .
Let U ⊂ Rn−1 be an open set. Let η2 , · · · , ηn be smooth one forms on U such that
(ds0 )2 = η22 + · · · + ηn2
defines a Riemannian metric on U . Let θ2 , · · · , θn be local coordinates of U . Let f (r, θ) be a positive
function on R+ × U . Define a Riemannian metric
X n
ds2 = dr2 + f 2 (r, θ) ηi2
i=2

on R × U . We are going to compute the curvature of ds2 .

+

Example of the above setting appears in the Euclidean metric under polar coordinates. For example
in R2 , the Euclidean metric can be written as
ds2 = dr2 + r2 dθ2 .
Of course, the curvature of the above metric is zero. We shall see that, even in the most general
setting, certainly components of the curvature tensor are quite simple.
We let
ω1 = dr
ωk = f (r, θ)ηk , k>1

Date: Lecture at Fudan University.

1
2 PROFESSOR ZHIQIN LU

Then
ds2 = ω12 + · · · + ωn2
Let e1 , · · · , en be the dual basis of ω1 , · · · , ωn , and let ηkl be the connection form for (ds0 )2 . Define
ω1l = −e1 (lg f )ωl
ωkl = ηkl − ek (lg f )ωl + el (lg f )ωk
for k, l > 1. Then we have
dωi = −ωij ∧ ωj .
∂
Note that e1 = ∂r
. A straightforward computation gives
ek ( ∂f
∂r
)
dω1l + ω1k ∧ ωkl = − ωk ∧ ωl
f
Proof. Since
∂ lg f
ω1l = −
ωl .
∂r
We have  
∂ ∂
dω1l = −ek (lg f ) ωk ∧ ωl − (lg f )dωl .
∂r ∂r
On the other hand, we have
∂
ω1k ∧ ωkl = − (lg f )ωk ∧ (ηkl − ek (lg f )ωl + el (lg f )ωk )
∂r
∂ ∂
= dηl + (lg f )ek (lg f )ωk ∧ ωl .
∂r ∂r
Since
∂ ∂f ∂
em ( lg f ) = f −1 em − em (lg f ) (lg f ),
∂r ∂r ∂r
we have
em ( ∂f
∂r
)
dω1l + ω1k ∧ ωkl = − ωm ∧ ωl .
f
Thus we have
em ( ∂f
∂r
) en ( ∂f
∂r
)
R1lmn = −δln + δlm .
f f
∂
The Ricci curvature at ∂r
direction is
1 ∂ 2f
R11 = R1l1l = −(n − 1) .
f ∂r2
More generally, we have the following Uhlenbeck’s trick: Let
ds2 = (dr)2 + hij (r, θ)dθi θj .
Ket ω1 = dr and ω2 , · · · , ωn be defined such that
X
hij dθi dθj = ωj2
j>1

is one forms, we can write

∂ωi
= aij ωj
∂r
for some matrix valued function (aij ). The Uhlenbeck’s trick told us that by an orthogonal change
of the co-frame, we can assume that (aij ) is symmetric.
 NOTES IN GEOMETRIC ANALYSIS 3

To prove rresult, we write the above equations into matrix form. let
ω2
 

ω =  ...  , A = (aij )
ωn
Then
∂ω
= Aω
∂r
Let A = B + C be the decomposition of the matrix A into symmetric and skew-symmetric parts.
Let Q be an orthogonal matrix valued function. We consider
∂ ∂Q
(Qω) = ω + QAω
∂r ∂r
Let B̃ be a symmetric matrix-valued function such that
∂
(Qω) = B̃Qω
∂r
If we let B̃ = QBQ> . Then the equation becomes
∂Q
= −QC
∂r
Given the above equation, we can verify that any solution must be orthogonal matrices:
∂(QQ> )
= −QCQ> − QC > Q = 0.
∂r
Now back to the computation of the curvature. Let (ηij ), (i, j > 1) be the connection forms for
Riemannian metric
hij dθi dθj .
Define
∂ωi
ωi1 = . ωij = ηij
∂r
Then we have
dωi = −ωij ∧ ωj
for i > 1 and
∂ωi
ωi1 ∧ ωi = ∧ ωi .
∂r
Since ∂ω
∂r
i
= aij ωj for symmetric (aij ), we have
ωi1 ∧ ωi = aij ωj ∧ ωi = 0
Thus (ωij ) is connection forms of (ds)2 .
We compute
1
dωi1 + ωik ∧ ωk1 = Ri1kl ωk ∧ ωl .
2
For the left hand side of above, the terms with {dr ∧ dθj } should be
∂ωi1 ∂ 2 ωi
dr · = dr · .
∂r ∂r2
In other words, we have
∂ 2 aij
dr ∧ dθj = Ri11l dr ∧ ωl
∂r2
It follows that
∂ ∂ ∂ 2 aij
Ric( , ) = R1i1i = −aij
∂r ∂r ∂r2
4 PROFESSOR ZHIQIN LU
p
Let f = log det hij . Then f = log(det aij ). From the above equation, we get
 2
∂ 2f ∂f ∂ ∂
2
+ = −Ric( , ) 
∂r ∂r ∂r ∂r
Using the above computation, we get the following comparison theorems.
Theorem 1.1. (Laplacian comparison theorem) Let X be a complete Riemannian manifold of di-
mension n. Let
Ric(X) = −(n − 1)k 2
Let N be an n-dimensional simply connected space form of constant sectional curvature −k 2 . Let
ρM , ρN be the distance functions to fixed reference points, respectively. If x ∈ X and y ∈ N such that
ρM (x) = ρN (y)
Then in the sense of distribution, we have
∆ρM (x) ≤ ∆ρN (y).
Proof. Outside the cut-locus and the reference points, the function ρM is smooth. Since ρM is a
distance function, |∇ρM | = 1. The Laplacian can be written as
1 ∂ ij √ ∂
∆= √ (g g )
g ∂xi ∂xj
Under the assumption that ds2 = dρ2 + hij dθi dθj , we have
∂ √ ∂f
∆ρ = lg g =
∂r ∂r
By our computation, we have
∂ 2f ∂f ∂ ∂
2
+ ( )2 = −Ric( , )
∂r ∂r ∂r ∂r
The corresponding function on N satisfies
∂ 2 f0 ∂f0 2
(1.1) 2
+( ) = (n − 1)k 2
∂r ∂r
The initial conditions on f and f0 are
∂f n − 1 ∂f0 n−1
∼ , ∼
∂r r ∂r r
Since
∂ ∂
−Ric( , ) ≤ (n − 1)k 2 .
∂r ∂r
By the maximum principle,
∂f ∂f0
≤ .
∂r ∂r0
This proves the comparison theorem at the smooth points of f .
Solving equation (??), we get
1
f (r) = · sinh kr
sinh kρ
Thus we have
n−1 n−1
∆N ρ = kρ coth kρ ≤ (1 + kρ)
ρ ρ
Thus the Laplacian comparison theorem can be written as
n−1
∆M ρ ≤ (1 + kρ)
ρ
 NOTES IN GEOMETRIC ANALYSIS 5

We shall prove that the above inequality is true even at singular points of ρ, in the sense of distribu-
tion. To see this, we let Ω be the domain in X such that ρ is smooth on Ω. Ω is a star-like domain
in X. Apparently
X = Ω ∪ Cut(ρ),
where Cut(ρ) is the cut-locus of X. Since ρ is at least continuous, and since the measure of Cut(ρ)
at least continuous, and since the measure of Cut() is zero we must have
Z Z
ρ∆ψ = ρ∆ψ
X

for any smooth function ψ. Let Ω be an exhaustion of Ω in the sense that Ω ⊂ Ω, Ω ⊂ Ω0 if  > 0
and lim→0 Ω = Ω. Then we have
Z Z Z
ρ∆ψ = − ∇ρ∇ψ = lim(−1) ∇ρ∇ψ
X X →0 Ω

Using the Green’s formula, we have

Z Z Z
∂ρ
− ∇ρ∇ψ = ∆ρ · ψ − ψ
Ω Ω ∂Ω ∂r

On the boundary ∂Ω , we have ∂ρ

∂r
≥ 0. Thus we have
n−1
Z Z Z
− ∇ρ∇ψ ≤ ∆ρ · ψ ≤ (1 + kρ)ψ
Ω Ω Ω ρ
Finally, we get
n−1 n−1
Z Z Z
ρ∆ψ ≤ lim (1 + kρ)ψ = (1 + kρ)ψ
X →0 Ω ρ X ρ
which finishes the proof of the theorem. 

Using exactyly the same method, we have the following volume comparison theorem of Bishop
Theorem 1.2. Let X be an n-dimensional complete Riemannian manifold. Let
Ric(X) ≥ −(n − 1)k 2 .
Then
vol(∂B(R))
volk (∂B(R))
is a decreasing function. Here volk (∂B(R)) is the volume of sphere of radius R of space form of
constant sectional curvature −k 2 , up to a constant, it is well defined as
volk (∂B(R)) − k sinh R.
The proof is exactly the same as that of the Laplacian comparison theorem. So we omit it.
Corollary 1.3. (Bishop) Let X be an n-dimensional Riemannian manifold. If Ric(X) ≥ (n − 1)k.
Then for any R > 0,
vol(BX (R))
vol(k, R)
is a decreasing function. In particular
volBX (k) ≤ V (k, R)
where V (k, R) is the volume of ball of radius R in space form of curvature k.
6 PROFESSOR ZHIQIN LU

Yesterday we proved the following: Let M be a complete Riemannian Manifold such taht Ric(M ) ≥
−(n − 1)k 2 . Then
1
∆r ≤ (1 + kr).
r
Here at non-smooth point, the inequality is true in the sense of distribution.
The main computation is as follows: Let the metric of M
ds2 = dr2 + hij dθi dθj
Then we have
2
∂ 2f

1 ∂f ∂ ∂
(1.2) + ≤ −Ric( , )
∂r2 n−1 ∂r ∂r ∂r
p
for f = log det(hij ). We have
∂f
∆r =
∂r
Thus using the equality (??) and the maximal principle we can prove the comparison result.
One important part of the computation is called Uhlenbeck’s trick, which is particularly useful in
computing the curvature of the Riemannian metric
dr2 + gr
where gr is the Riemannian metric on U .
Using Uhlenbeck’s trick, we get
1 ∂
(∆r)2 + ∆r + Ric(∇r, ∇r) ≤ 0
n−1 ∂r
However, the above inequality also follows from the so-called Ricci identity.
Let f be a smooth function, then by Ricci identity, we have
1 ∂
∆|∇f | = |∇2 f |2 + ∆f + Ric(∇f, ∇f )
2 ∂r
If we specialize f to be r, then |∇r| = 1 and ∆|∇r|2 ≡ 0. On the other hand
2

X 1 X 2 1
|∇2 f |2 ≥ fii2 ≥ fii = (∆f )2
i
n − 1 n − 1
and the inequality follows.
We implicitly used the first variational formula, because the special form
dr2 + hij dθi dθj
is implied by the Gauss-lemma, and we use the first variational formula to prove the Gauss Lemma.

2. Gradient estimates and maximal priciple

If X is a compact manifold, then any smooth real function on X reaches it maximum point. At the
maximum point, the first derivatives are zero, and the Hessian matrix at that point is non-positive.
For non-compact manifold, the above statement is not true in general. In order to estimate
a function we some times need a differential inequality. Usually such a differential inequality is
obtained by doing the so-called gradient estimate.
We start with the following generalized maximum principle:
Theorem 2.1. Let f be a positive smooth function on a complex non-compact Riemannian manifold
X. We assume that Ric(x) ≥ −(n − 1)k 2 for some number k. Let ψ1 , ψ2 be two smooth functions on
X with ψ1 bounded. Assume that there are constants α > 0, C1 , C2 such that
(2.1) ∆f ≥ C1 f 1+α + ψ0 ∇ψ1 ∇f − C2
 NOTES IN GEOMETRIC ANALYSIS 7

Then f is bounded. Furthermore, there is a constant C = C(α, C1 , C2 , ||ψ0 ||C 0 , ∇ψ1 ) such that
f ≤C
Proof. We first assume that f is bounded. That is sup f < +∞. We claim that there is sequence
{xk } in X such taht for any  > 0, if k is large enough, we have
f (xk ) > sup f − 
|∇f |(xk ) < 
∆f (xk ) < 
To prove the claim, we first take a sequence {yk } such that
lim f (yk ) = sup f
k→∞
Define the cut-off function r as follows
ψ:R→R
smooth, 0 ≤ ρ ≤ 1, ρ(t) = 1 for 0 ≤ t ≤ 1,ρ(t) = 0 for t > 2 and ρ0 (t) ≤ 0 for all t ∈ R. Let R be a
large number to be determined later.Let d(x) = dist(x, yk ) be the distance function. In general, the
function d(x) is only a continuous function. But let
 2 
d (x)
g(x) = ρ f (x)
R2
If the maximum point of g(x) happens to be not smooth, we can always perburb the reference point
by a little. So without loss of generality, we assume that g(x) is smooth. Let xk be the maximum
point of g(x) . By the definition of xk , we have
g(xk ) ≥ g(yk ) = f (yk ) ≥ sup f − 
Using the definition, we also have
d2 (xk )
 

1−ρ <
R2 sup f
Since
2
0 = ∇g(xk ) = ρ0 d∇df (xk ) + ρ∇f (xk )
k2
we have
−ρ0 2
|∇f |(xk ) = d|∇d|f (xk )
ρ k2
Because |∇d| = 1, if R is big enough, we have
|∇f |(xk ) < 
Finally, we have
0 ≥ ∆g(xk ) = ρf + 2∇ρ∇f + f ∆ρ
In order to prove the claim, we just need to prove that
 2 
d (x, yk )
∆ρ <
R2
for k large enough. A straight forward computations gives that
2 0 4d2 00 2
∆ρ =2
ρ + 4 ρ + 2 ρ0 d∆d
R R k
Since the Ricci curvature of X is bounded from below, by the Laplacian comparison theorem, we
have
d∆d ≥ −C,
for some constant C depending only on the lower bound of the Ricci curvature. Since ρ0 ≤ 0, for R
large enough we have ∆ρ <  for any . The claim is proved.
8 PROFESSOR ZHIQIN LU

Using the differentiable inequality ??, we have

 ≥ C1 f (xk )1+α − ||ψ0 ||||∇ψ1 || − C2
Thus s
1+α C2 + ||ψ0 ||||∇ψ1 ||
f (xk ) ≤
C1
Taking k → ∞,we get the effective bound for f .
Now we assume that sup f = +∞. Let
1
v =1−
(1 + f )β
for β > 0 to be determined later. Then by the previous result, there is sequence {xk } such that for
any  > 0, if k is large enough, we have
v(xk ) > 1 − 
|∇v|(xk ) < 
∆v(xk ) < 
By the definition of v we have
∆f 1+β
∆v = β − |∇v|2 (1 + f )β
(1 + f )(1+β) β
Applying the above inequality to (??), we get
C1 f 1+α − ||ψ0 ||||∇ψ1 || ·  − C2 1 + β
>β − · (1 + f )β
(1 + f )1+β β
Letting k → ∞, we get a uniform bound of sup f from the above inequality. 
Definition 2.2. Let f be a smooth function on a complete non compact manifold. We say that f
is harmonic function, if ∆f = 0
Then we have the following:
Theorem 2.3 (Yau). Let p > 1, if f ∈ Lp , ∆f = 0. Then f is constant.
For the sake of simplicity, we only prove the theorem for p = 2.
Proof. Let f be an L2 harmonic function. Take a fixed point x0 . Let
d(x) = dist(x, x0 )
Let ρ be a cut-off function such that Suppρ is compact. We consider
 
d(x)
g(x) = ρ f (x)
R
We have Z Z
2 2 1
ρ |∇f (x)| = −2 ρf ∇ρ∇f
M R
using Cauchy inequality, we get
Z Z 1/2 Z
2 2 2 2 2
ρ |∇f | ≤ ρ |∇f | |∇ρ|2 f
M R M M

Since |∇ρ| ≤ C. We get Z Z

2 1 2
ρ |∇f | ≤ C f2
M R M
Letting R → ∞, we get ∇f ≡ 0, so f is constant. 
 NOTES IN GEOMETRIC ANALYSIS 9

By the above theorem, for a complete manifold, the only interesting harmonic functions may be
bounded harmonic functions.
In order to study the bounded harmonic functions, we first introduce the Ricci identity.
Let f be a smooth function of M . Define the derivative of f using the following formula fi ωi = df .
Using the same idea, we define
fij ωj = dfi − fs ωsj .
The matrix (fij ) is called the Hessian matrix. We have
fij ωj ∧ ωi = dfi ∧ ωi − fs ωsj ∧ ωi = 0.
Thus the Hessian matrix is always symmetric.
The 3rd order co-variant derivatives are defined as
fijk ωk = dfij − fis ∧ ωsj − fsj ∧ ωsi
A careful computation gives
1
fijk ωk ∧ ωj = − fs Rsikj ωk ∧ ωj
2
Thus we have
fijk − fikj = fs Rsijk
In particular, we have the following Ricci identity
flik − fkli = −fs Rilsk
Remark 2.4. We have ∆f = fii
With the preparation above, we prove the following.
Theorem 2.5. Let M be a complete Riemannian manifold, dim M = n ≥ 2. Ric(M ) ≥ −(n −
1)k, k ≥ 0. Let u be apositive harmonic function. Then on any geodesic ball Ba (x), we have
1
!
|∇u| 1 + ak 2
≤ Cn
u a
where Cn is a constant only depends on n.
Proof. We first prove that
1 X
(2.2) ∆|∇u|2 ≥ u2ij − (n − 1)k|∇u|2
2 i,j

To see this, we do the following

X 
u2j = 2uj uji
i
Thus
1
∆|∇u|2 = (uj uji )i = u2ji + uj ujii
2
Using the Ricci identity, we have
1
∆|∇u|2 = u2ji + uj (∆u)j + Ric(∇u, ∇u)
2
Thus (??)follows frome the assumption on the Ricci curvature and harmonicity of u.
We now consider the points such that ∇u 6= 0. By changing a frame, we may assume that u1 6= 0,
uj = 0 for j > 1. From (??), we conclude that
2
∆|∇u|2 1 |∇|∇u|2 |
∆|∇u| = −
2|∇u| 4 |∇u|3
10 PROFESSOR ZHIQIN LU

using the above information, we get

!
1 X X
∆|∇u| ≥ u21j + u2jj − (n − 1)k|u1 |
|∇u| j>1 j>1

Since !2
X 1 X 1
u2jj ≥ ujj = u2
j>1
n−1 j>1
n − 1 11
we have
1 1 X 2
∆|∇u| ≥ u − (n − 1)k|u1 |
n − 1 |∇u| j 1j
Now we assume that ψ = |∇u|/u. Then we have
 
∆|∇u| 1 1
∆ψ = + 2∇|∇u|∇ + |∇u|∆
u u u
Since
1 1 ∇ψ∇u
2∇|∇u|∇ = 2∇(ψu)∇( ) = −2 − 2ψ 3
u u u
We have
∆|∇u| ∇ψ∇u
∆ψ = −2
u u
We have
∇ψ∇u 1
= −ψ 3 − 2∇|∇u|∇
u u
Thus if  > 0 is small enough
∆|∇u| ∇ψ∇u
− ≥ ψ 3
u u
Thus we get
∇ψ∇u
∆ψ ≥ −(n − 1)kψ − (2 − ) + ψ 3
u
Using the maximum principle, we have ψ is bounded. 
Corollary 2.6. Let M be a complete Riemannian manifold if Ric(M ) ≥ 0. Then any positive
harmonic function is a constant.
Proof. If Ric ≥ 0, then k = 0. We have
∇ψ∇u
∆ψ ≥ −(2 − ) + ψ 3
u
2
For  = n−1 . Thus using the generalized maximal principle, ψ ≡ 0 
Corollary 2.7 (Harnack Inequality). Let M be n-dimensional Riemannian manifold, Ric(M ) ≥
−(n − 1)k. If u is a positive harmonic function in Ba . Then
sup u ≤ C(n, a, k) inf u
Ba/2 Ba/2

where C(n, a, k) are constants depending only on n, a, k.

Proof. By the above theorem, we have
|∇u|
sup ≤ C(n, a, k)
Ba u
Thus
|∇ log u| ≤ C(n, a, k)
The conclusion follows. 
Corollary 2.8. Suppose Ric(M ) ≥ 0. Then any positive harmonic function must be constant.
 NOTES IN GEOMETRIC ANALYSIS 11

Supplementary for yesterday. Let the Riemann metric be written as

(dr)2 + hij (r, θ)dθi dθj
What is the asymptotic behavior of hj (r, θ) when r → 0.
Note that the Riemannian metric at the reference point is regular. Thus we can define (θ2 , · · · , θn )
as
s
xj X
θj = , r= x2j
r

Suppose ds2 = gij dxi dxj . Let’s compare

gij dxi dxj ∼ (dr)2 + hij dθi dθj
We have
gij dxi dxj = gij (dr · θi + rdθi )(dr · θj + rdθj )
= gij θi θj (dr)2 + 2gij xi drdθj + gij r2 dθi dθj
By comparison, we have
gij θ1 θj = 1
X
gij xi = 0
i

and
 
2 θi θj θi θj
hij = r gij − g1j − gi1 + g11 2
θ1 θ1 θ1
p
Thus det(hij ) = rn−1 · µ, for µ being a regular function (at least for fixed θi ). Thus
∂f n−1
∆ρ = ∼ + small terms.
∂r r
If we choose (gij ) to be normal, we can actually compute
 
θi θj 1
det δij + 2 = 2
θi θ1
p
Thus det(hij ) can be extended as a regular function near 0.

3. Eigenvalue problems
We first make different notation of the Laplace operator.
We assume that M is a Riemannian manifold with the Riemannian metric
X
ds2 = gij dxi dxj .

Let (g ij ) be the inverse matrix of (gij ) and let g = det(gij ). Then under the local coordinates
(x1 , · · · , xn ) we define
 
1 ∂ ij √ ∂
∆= √ g g
g ∂xi ∂xj
Apparently, we can write
∂2 1 √
∂
∆ = g ij + √ g ij g
∂xi ∂xj g ∂xj
12 PROFESSOR ZHIQIN LU

As before, we have the following computation

1 ∂ √
1 ∂ ∂ ij
√ g ij g = g ij log g + g
g ∂xi 2 ∂xi ∂xi
1 ∂gkl ∂gmn
= g ij g kl − g in g jm
2 ∂xi ∂xi
kl j
= −g Γkl
Thus we also have the following formula
∂2 ∂
∆=g ij
− g kl Γjkl
∂xi ∂xj ∂xj
From the above representation, we have the following third formula for the Laplace operator.
n
Let ds2 = ωj2 . With respect to the frame, the connection (ωij ) is well-defined. We are above
P
j=1
the define the covariant derivatives of function like the following
df = fi ωi
fij ωj = dfi − fs ωsi
using the above notations, we can define
n
X X
∆f = fii = ∇i ∇i f
i=1 i

Finally, we can define the Laplace operator on p-forms as follows:

Let d : ∧p (M ) → ∧p+1 (M ) be the ordinary differential operator, where ∧p (M ) be the space of
smooth p-forms. With respect to the Riemannian metric, ∧p (M ) becomes an infinite dimensional
inner product space.
Let δ be the formal dual operator with respect to d. Then
δ : ∧p (M ) → ∧P −1 (M )
we can prove that δ is also a differential operator of fist order. The Laplace operator can be defines
as
∆ = dδ + δd
In particular, on the space of functions, or 0-forms,
∆ = δd
We have the following
Theorem 3.1. (Weitzenböck formula) For function f , we have
X
∆f = δdf = − ∇i ∇i f

In general, we have the following

X
∆=− ∇i ∇i + curvature terms

Which is also called Weizenböck formula.

∇i ∇i is well defined on C ∞ (M )
P
WE go back to the Laplacian on functions. We know that ∆ =
Unfortunately, with the following L2 -inner product
Z
hf, gi = f gdV
M
 NOTES IN GEOMETRIC ANALYSIS 13

C ∞ (M ) is not a complete metric space. The complete metric space is L2 (M ). However, there is no
way that we can extend ∆ on L2 (M ).
Proof. The key pint is that any differential operator is a closed graph operator. Thus if ∆ is extend-
able, then by close graph theorem, ∆ has to be a bounded operator. however ∆ is not a bounded
operator, as by the example of Heaviside function

 0 x≤0
f (x) = continuous cure connected (0,0) and (1,1)
 1 x≥1

Thus we can only extend the operator into a densely defined self-adjoint operator. 
Recall that an operator ∆ is self-adjoint, if
Dom(∆) = Dom(∆∗ )
and
h∆f, gi = hf, ∆gi
for any f, g ∈ Dom(∆)
In functional analysis, we have the following theorem: Let
Z
Q(φ, ψ) = ∇φ∇ψ.

Then Q is a non-negative quadratic form defined on H01 (M ). Then there is a unique densely defined
operator A such that
Q(φ, ψ) = −(Aφ, ψ)
such an operator A is in fact called the Dirichlet Laplacian Operator.
As an exercises, we prove that on only manifold, L2 harmonic function must be constant.
Theorem 3.2. Let A be the Dirichlet extension of Laplacian ∆. A function f is called A-harmonic,
if f ∈ Dom(A) and Af = 0. If f ∈ L2 (M ), then f is a constant.
Proof. The key point is that
Q(ρ2 f, f ) = ρ2 f, ∆f = 0
Thus using the same method as before, f is a constant. 
If M is a compact manifold with no boundary. we still use ∆ to denote the Dirichlet extension of
the Laplace operator. By the elliptic regularity, the spectrum of ∆ are discrete. That is, there is a
sequence
0 = λ0 < λ1 ≤ λ2 ≤ · · ·
2
such that for any λi , there ∃fi ∈ L such that,
∇fi = −λi fi
We have similar results similar results for manifolds with boundary conditions. To be more precise,
the Laplacians acting on functions with the following boundary conditions.
• Dirichlet boundary condition: ∆ acting on f vanishing on the boundary.
∂f
• Neumann boundary condition: ∆ acting on function such that ∂n = 0.
For eigenvalues, we have the following minimax principle. Assume that M is a closed manifold, then
R
|∇f |2
λ1 = Rinf R 2
f =0 f
R
To prove the above result, we let ψ be any smooth function such that M ψ = 0. Then by the
definition of λ1 , we have
|∇(f + )|2
Z
R ≥ λ1 .
(f + ψ)2
14 PROFESSOR ZHIQIN LU

However, if we take the first order term, we get

Z
(∆f + λ1 f )ψ = 0

Note that Z
(∆f + λ1 f ) = 0
Then
∆f + λ1 f ≡ 0
and λ1 is the first eigenvalue.
By elliptic regularity, λ1 > 0. Thus we have the following Poincare-inequality. There exists a
constant C, such that Z Z
|∇f | ≥ C f 2
2

R
for any function with f = 0. Before going further, let’s prove the following C0 -area formula.
Theorem 3.3. Let M be a compact Riemannian manifold with boundary. Let f ∈ H 1 (M ). Then
Z Z+∞Z
g
g= dσ.
M {f =σ} |∇f |
−∞

Proof. Without loss of generality, we assume that |∇f | 6= 0. Thus by the implicit function theorem
{f = σ}is a smooth manifold.
Using cut-off function, we may assume that Supp g is contained in a coordinate chart. Thus we
may assume that the Riemannian metric is given under the global coordinates (x1 , · · · , xn ) as follows
X
ds2 = gij dxi dxj
by definition Z Z q
g= g det(gij )dx1 · · · dxn
M M
Since ∇f 6= 0, we can solve the equation
f =σ
by
x1 = x1 (σ, x2 , · · · , xn )
or in order word, by the implicit function theory (σ, x2 , · · · , xn ) is a local coordinate system as well.
The Jacobian of the transformation is
∂x1
dx1 ∧ · · · ∧ dxn = dσ ∧ dx2 ∧ · · · ∧ dxn
∂σ
On the other side, restricting to f = σ, the Riemann metric can be written as
∂x1 ∂x1 ∂x1 ∂x1
(g11 · + g1l + gk1 + gkl )dxk dxl
∂xk ∂xl ∂xk ∂xl
If we choose local coordinates such that gkl = δkl . Then we have
 
∂x1 ∂x1
· + δkl dxk dxl
∂xk ∂xl
The volume form of the above is

∂x1 ∂x1
 X  ∂x1 2 1 X ∂f
2
|∇f |2
(3.1) det δkl + · =1+ = =
∂xk ∂xl ∂xk ∂f ∂xk ∂f
∂x1 ∂x1

Thus we have
Z Z Z
∂x1
q q
g= g det(gij )dx1 · · · dxn = g det(gij ) dσ ∧ dx2 ∧ · · · ∧ dxn
M M M ∂σ
 NOTES IN GEOMETRIC ANALYSIS 15

By (??) we must have

∂f
∂x1
dVds2 = dVf =σ ·
|∇f |
Thus
Z Z∞ Z 
g
g= dσ
M f =σ |∇f |
−∞
Of course, in general, there are points such that ∇f = 0. But by a theorem of Sand. The set
{f (y) : ∇f (y) = 0}
is of zero measure. Using the standard covering technique, we can prove the same result.

As an application of the above co-area formula, we prove the following result of Sobolev inequality.
Theorem 3.4 (Sobolev inequality). let M be a compact manifold with boundary. Then there is
constant C > 0 such that
Z  n−1
n
Z
n
C |f | n−1 ≤ |∇f |
M M
R
for any smooth function f |∂M = 0(D-Condition) or f = 0(Neumann condition).
Theorem 3.5. (Isopermetric inequality) Let Ω be a domain in M, Ω ⊂⊂ M . Then there is a constant
independent to Ω such that
n−1
C vol(Ω) n ≤ vol(∂Ω)
We want to prove that the Isopermetric inequality is equivalent to the Sobolev inequality.
At least one-side of the implication was clear: assuming the sobolev inequality, if we let

 1, x ∈ Ω, d(x, ∂Ω) ≥ 
d(x,∂Ω)
f = 
, x ∈ M, d(x, ∂Ω) ≤ 
0, otherwise.


Then using the Sobolev inequality, the isopermetric inequality follows by letting  → 0.
In order to prove that the isopermetric inequality implies the Sobolev inequality, we use the co-area
formula. We assume that f ≥ 0. Then
Z Z∞
|∇f | = Area(f = σ)dσ
M
0

We also have
Z Z∞ Z∞
n n n n
|f | n−1 = vol(f n−1 > λ)dλ = vol(f > σ)σ n−1 dσ
M n−1
0 0
Using the isopermetric inequality, we have
Z Z∞ Z∞
n−1
|∇f | = Area(f = σ)dσ ≥ C vol(f > σ) n dσ
M
0 0

Thus in order to prove the Sobolev inequality, we just need to prove that
Z∞ Z∞
n−1 1 n−1
vol(f > σ) n dσ ≥ C( vol(f > σ)σ n−1 dσ) n
0 0
16 PROFESSOR ZHIQIN LU

We let
F (σ) = vol(f > σ)
Zt
n−1
φ(t) = F (σ) n dσ
0
 t  n−1
n
Z
1
ψ(t) =  F (σ)σ n−1 dσ 
0

Then φ(0) = ψ(0). Using the monotonicity of F (σ) we can prove that
n
φ0 (t) ≥ ψ 0 (t)
n−1
Thus
n
φ(∞) ≥ ψ(∞)
n−1
Corollary 3.6.
Z  n−p
np
Z  p1
np
p
f n−p ≤C |∇|
for any p > 1.
Definition 3.7. Let M be a compact Riemannian manifold if ∂M 6= ∅,
 
vol(∂Ω)
hD (M ) = inf | Ω ⊂⊂ M
vol(Ω)
if ∂M = ∅  
vol(H)
hN = inf | H is a hyper surface.
min(vol(M1 ), vol(M2 ))
Theorem 3.8. (Cheeger) For Dirichlet condition, we have
1
λ1 ≥ h2D
4
For Neumann codition
1
λ1 ≥ h2N (M )
4
Proof. We only prove the case for Dirichlet condition. We first observe that, if there is constant µ
such that Z Z
|∇φ| ≥ µ |φ|
M M
1 2
for any φ with φ|∂M = 0. Then λ1 ≥ To see this, we consider φ = f 2
4
µ.
Z Z Z  21 Z  12
2 2 2
µ f ≤2 |f | · |∇f | ≤ 2 f |∇F |
M M M
Thus Z Z
1 2 2
µ f ≤ |∇f |2
4 M M
Since the above is true for any function f , we must have
1
λ1 ≥ µ2
4

Finally, we prove a result which is well known but can’t readily be found in the literature.
 NOTES IN GEOMETRIC ANALYSIS 17

Theorem 3.9. Let M be a compact manifold with smooth boundary. Let

R
|∇f |2
µ1 = Rinf R
f =0 f2
Then we have the following result: let f be a minimizer, and let f be smooth. Then
∂f
=0
∂n R
Proof. Let φ be a smooth function with compact support such that φ = 0. Then we have
Z Z
|∇(f + φ)| ≥ µ1 (f + φ)2
2

Thus we have Z Z
∇f ∇φ = −µ1 fφ
M M
by Green’s formula Z Z Z
∂f
∇f ∇φ = φ − ∆f φ
M ∂M ∂n M
Thus since ∆f = µ1 f , we have Z
∂f
=0
∂M ∂n
∂f
and we must have ∂n
= 0. 
4. Eigenvalue Problems (II)
By the variational characterizing of the eigenvalues, we know that it is usually more difficult to
get the lower bound estimate of eigenvalues. Among all the eigenvalues, the lower bound of the first
eigenvalue is particularly important.
The Cheeger’s result did give a lower bound estimate of the first eigenvalues. But the bounds
are not ”computable”. In geometry, ”computable” bounds provide effective versions of Poincare and
Sobolev inequalities.
The following Lichnerowicz therem gives a good lower bound of the first eigenvalue for closed
manifold.
Theorem 4.1. (Lichnerowicz) Let M be a closed n-dimensional Riemannian manifold. Assume that
Ric(M ) ≥ (n − 1)k > 0
Then λ1 ≥ nk.
Proof. One line proof, let µ be the first eigenfunction. Then use the Ricci identity we have
1 X
∆|∇u|2 ≥ u2ij + ∇u∇∆u + Ric(∇u, ∇u)
2
We have
!2
X X 1 X λ1 2
u2ij ≥ u2ii ≥ uii = u
n n
Ric(∇u, ∇u) ≥ (n − 1)k|∇u|2
Thus we have
1 λ2
∆|∇u|2 ≥ 1 u2 − λ1 |∇u|2 + (n − 1)k|∇u|2
2 n
Taking integration on both sides, we get
λ21
− λ21 + (n − 1)kλ1 ≤ 0
n
The theorem follows. 
18 PROFESSOR ZHIQIN LU

In 1962. Obata proved, if λ1 = nk, then M has to be the standard sphere. For the rest of this
section, we use the gradient estimates to find ”computable” lower bounds of the first eigenvalue. We
prove the following theorem
k2
Theorem 4.2 (Li-Yau). let M be a closed manifold and Ric(M ) ≥ 0. Then λ1 ≥ 2d2
, d is the
diameter.
Proof. Let u be the first eigenfunction. After normalization, we may assume that
1 = sup u > inf u = −k ≥ −1
for some 1 ≥ k > 0. Let
1−k
u− 2
ũ = 1+k
2
The after this linear change of u, we have

 ∆ũ = λ1 (ũ + a)
sup ũ = 1
inf ũ = −1


for a = 1−k
1+k
, 1 > a ≥ 0.
1
Let g = 2 (|∇ũ|2 + (λ1 + )ũ2 ) for some  > 0 to be determined later. Assume that at x0 .
g(x0 ) = max g.
Using the maximum principle, at x0 , we have
ũj ũji + (λ1 + )ũui = 0
and
0 ≥ ∆g
= ũ2ij + Ric(∇ũ, ∇ũ) + ∇ũ∇∆ũ + (λ1 + )|∇ũ|2 + (λ1 + )ũ∆ũ
≥ ũij − λ1 |∇ũ|2 + (λ1 + )|∇ũ|2 − λ1 (λ1 + )ũ(ũ + a)
If at x0 , ∇ũ = 0. Then we have
|∇ũ|2 + (λ1 + )ũ2 ≤ (λ1 + )
In particular, we have
|∇ũ|2 + λ1 (1 + a)ũ ≤ λ1 (1 + a)
If ∇ũ(x0 ) 6= 0. Then using the Cauchy inequality
(ũi ũj ũij )2
ũ2ij ≥ 4
= (λ1 + )2 ũ2
|∇ũ|
Thus we have
0 ≥ ∇g(x0 )
≥ (λ1 + )ũ + |∇ũ|2 − λ1 (λ1 + )ũ2 − λ1 (λ1 + )a
≥ 2g − λ1 (λ1 + )a
For any  > λ1 a, the above gives
|∇ũ|2 + λ1 (1 + a)ũ ≤ λ1 (1 + a)
Let
f (t) = arcsin ũ(σ(t)),
where σ(t) is the arc-length curve connecting the minimal point and the maximum point of ũ. Then
by the above argument, we have p
|f 0 (t)| ≤ λ1 (1 + a)
 NOTES IN GEOMETRIC ANALYSIS 19

Let d be the diameter of the manifold, then we have

p Zd
d λ1 (1 + a) ≥ |f 0 (t)|dt ≥ arcsin 1, and − arcsin(−1) = π
0
1 π2 π2
Thus λ1 ≥ 1+a d2
. Since a < 1, this gives λ1 ≥ 2d2

We let θ = arcsin ũ. Then Zhong-Yau proved the following surprising theorem.
Theorem 4.3 (Zhong Jiaqing, Yang Hongchang). Let
4
ψ(θ) = (θ + cos θ sin θ) − 2 sin θ
π
Then
|∇ũ|2
≤ λ1 (1 + aψ(θ))
1 − ũ2
The method is maximal principle, very surprising and mysterious.
Using the above sharpened inequality, observed that ψ(θ) is an odd function, we can prove that
π2
λ1 ≥
d2
Assume that Ric(M ) ≥ −(n − 1)k for √ k > 0. Then by estimating |∇u|2 + λ1 (1 − u)2 , Li-Yau was
above to prove that λ1 ≥ dC2 exp(−C1 kd2 )
Yang Hongchang was able to modified the above and proved that
π2 √
λ1 ≥ 2 exp(−C kd2 )
d
When Ric(M ) > 0, or Ric(M ) ≥ (n − 1)k > 0, the above inequality is not optimal. In fact, it is
far from being optional. Let M = S n . Then by Lichnerowicz theorem, λ1 ≥ n ( in fact , λ1 = n).
d(S n ) = π. Thus Yang-Zhong gives
λ1 ≥ 1.
In this direction, we have the following Peter Li Conjecture.
Conjecture 4.4. (P.Li) If Ric(M ) ≥ (n − 1)k. Then
π2
λ1 ≥ + (n − 1)k
d2
Such a conjecture, if true, will sharpen both the result of Zhong-Yang and Lichnerowicz because
by Myer’s theorem
π2
≥k
d2
Not much was known to the proof of the conjecture. D.Yang proved that
π2 1
λ1 ≥ + (n − 1)k
d2 4
LingJun proved a bigger number
π2 1
λ1 ≥ 2 + α(n − 1)k, α >
d 4
On the other end, if Ric(M ) ≥ −(n − 1)k, k > 0. Then
π2
λ1 ≥
− (n − 1)k
d2
Recently, Wang Xiaodong and Hang Fengpi was able to prove that
π2
λ1 >
d2
20 PROFESSOR ZHIQIN LU

It is a very interesting result. We end this section by citing a result of Li and Croke.
Theorem 4.5. Let M be a manifold with boundary. Then there is a constant C = C(n, d, V, k) > 0
such taht the Sobolev constant is > C > 0.

5. Harmonic functions revisited

We assume that u is a positive harmonic function defined on B(a) ⊂ M , where M is complete
Riemannian manifold with Ric(M ) ≥ −(n − 1)k.
In this section, we re-prove the Harnark inequality using the de Giorgi-Nash-Moser estimates. Note
that our result is weaker than the differential Li-Yau Harnark inequality. however, the methods we
use here is useful in non-linear case, even we only use a linear problem as the example. The reference
book of this section is [?]. We first prove the following estimate.
Theorem 5.1. Under the above assumptions, then for any p > 0, 0 < θ < 1, there is a constant
C = C(n, p, θ, k) > 0 such that
Z  p1
sup u ≤ C up
B(aθ) B(a)

Proof. We first assume that p ≥ 2. Let ψ ≥ 0, ψ ∈ C0∞ (B(a)). Then since ∆u = 0, we have
Z
∇u∇ψ = 0
B(a)

Let ρ be a smooth function with compact support in B(a). Then we have

Z
∇u∇(ρ2 up−1 ) = 0
B(a)

Here p is a real number to be specialized later.

Expanding the above equality, we get
Z Z
2 p−2 2
(p − 1) ρ u |∇u| = −2 up−1 ρ∇u∇ρ
B(a) B(a)
Z
p p−2
= −2 (∇ρ · u 2 )(∇u · u 2 ρ)
B(a)

Using the Cauchy inequality, we get

Z Z
2 2 p−2 2
(p − 1) ρu |∇u| ≤ C |∇ρ|2 up
B(a) B(a)

Note that
4 p
up−2 |∇u|2 = 2
|∇u 2 |2
p
We have Z Z
p
2 2
ρ |∇u | ≤ C
2 |∇ρ|2 up
B(a) B(a)
If we allow C to be a litter bigger, we shall get
Z Z
p
2
|∇(ρu 2 )| ≤ C |∇ρ|2 up
B(a) B(a)

We let 2∗ = 2n
n−2
> 2. Using the Sobolev-inequality we have
Z  p 2∗  21∗ Z
(5.1) ρu 2 ≤C |∇ρ|2 up
B(a) B(a)
 NOTES IN GEOMETRIC ANALYSIS 21

We let Rk = a(θ + 1−θ

2k
). Let ρk ∈ C0∞ (B(Rk )), 0 ≤ ρk ≤ 1, ρk ≡ 1 on B(Rk+1 ). We further assume
that
2 2k+1
|∇ρk | ≤ =
Rk − Rk−1 (1 − θ)a
From (??), we have
! n−2
n
C · 4k
Z Z
np
u n−p ≤ up
B(Rk+1 ) (1 − θ)2 a2 B(Rk )

n k
Specializing p = pk , where pk = p( n−2 ) , we have
 p1
C · 4k

k
||u||Lpk +1 (B(Rk+1 )) ≤ ||u||Lpk (B(Rk ))
(1 − θ)2 a2
Iterating, we have
 p1
Y C · 4k k
||u||Lpk +1 (B(Rk+1 )) ≤ ||u||Lp (B(a))
(1 − θ)2 a2
We need to prove that
 p1
Y C · 4k k C
≤ ( Exercises)
(1 − θ)2 a2 ((1 − θ)2 a2 )n/2p
Since the right-hand side is independent of R, we let k → ∞ and get
C
||u||L∞ (B(θa)) ≤ ||u||Lp (B(a))
((1 − θ)a)n/p
This proves the theorem for p ≥ 2.
Now we assume that 0 < p < 2. Using the result for p = 2, we have
Z  12
C 1− p2 p
||u||L∞ (B(θa)) ≤ ||u||L∞ (B(a)) u
((1 − θ)a)n/2 B(a)

Using the Yong inequality, we get

1 C
||u||L∞ (B(θa)) ≤ ||u||L∞ (B(a)) + ||u||Lp (B(a))
2 ((1 − θ)a)n/p
Let ψ(s) = ||u||L∞ (B(sa)) , we get
1 C
ψ(s) ≤ ψ(t) + ||u||Lp (B(a))
2 (1 − s)n/p
∀0 < s < t ≤ a. Iterating again, we get
C
ψ(s) ≤ ||u||Lp (B(a)) , ( exercise ) 
((1 − s)a)n/p
Theorem 5.2. (weak Harnark inequality) There is a constant C > 0, p0 > 0 such that
Z  p1
1 p0
0
inf ≥ u
B(aθ) C B(a)

Here p0 , C only depends on a, (1 − θ)−1 and Sobolev constants.

Proof. Without loss of generality, we assume that u ≥  > 0. Otherwise we can use u +  in-place of
u. We also assume that a = 1. By a straight forward computation we get that
2|∇u|2
∆u−1 = ≥0
u3
22 PROFESSOR ZHIQIN LU

Thus u−1 is a subsolution. Using the above lemma, for any p, we have
Z
−p
sup u ≤ C u−p
B(θ) B(1)

Thus we must have

Z − p1
− p1 −p
inf u ≥ C u dx
B(θ) B1
Z Z − p1 Z  p1
− p1 −p p p
=C u u u
B1 B(1) B(1)

In order to prove the theorem, we just need to prove that for p > 0 small enough
Z Z
−p
u up ≤ C
B(1) B(1)
H
We let ω = log u − β, where β = log u. We shall establish
I
(5.2) ep|ω| ≤ C.
B(1)

for p > 0 small enough. We first prove that

I
|∇ω|2 ≤ C
B(σ)

for some σ > 1. To see this, let ρ be the cut-off function whose support is within B(σ̄) for some
σ̄ > σ. Since u is harmonic, we have
Z
∇u∇(u−1 ρ2 ) = 0
B(σ̄)

It follows that
Z Z
2 2
ρ |∇ω| ≤ ∇ω∇ρ2
B(σ̄) B(σ̄)
sZ sZ
≤2 ρ2 |∇ω|2 |∇ρ|2
B(σ̄) B(σ̄)

Since ρ only depends on σ, σ̄, we get the desired inequality. By the Poincare inequality, we have
Z
|ω|2 ≤ C
B(σ̄)

To get the estimate (??), we still use the Moser iteration. First observe that
∆ω = −|∇ω|2
Let ρ be a cut-off function to be determined later. Then we have
−ρ2 |ω|2q ∆ω = ρ2 |ω|2q |∇ω|2
It follows that
Z Z
2 2q 2
ρ |ω| |∇ω| = ∇ω∇(ρ2 |ω|2q )
Z Z
= 2 ρ∇ρ∇ω|ω| + 2q ρ2 |ω|2q−1 |∇ω|2
2q

We use the Young inequality to get

2q − 1 2q
(5.3) 2q|ω|2q−1 ≤ |ω| + (2q)2q−1
2q
 NOTES IN GEOMETRIC ANALYSIS 23

Inserting the above inequality into the equation, we get

Z sZ sZ
ρ2 |ω|2q |∇ω|2 ≤ (2q)2q + 2 ρ2 |ω|2q |∇ω|2 |∇ρ|2 |ω|2q

Using the Cauchy inequality ab ≤ a2 + 1 b2 , we get

Z  Z 
2 2q 2 2q 2q 2
ρ |ω| |∇ω| ≤ C (2q) + |ω| |∇ρ|

With a slightly larger constant C, we have

Z  Z 
2 q+1 2 2 2q 2q 2
ρ |∇|ω| | ≤ Cq (2q) + |ω| |∇ρ|

Thus we have Z  Z 
2 q+1 2 2q 2q+2 2
|∇(ρ |ω| )| ≤ 2Cq (2q) + |ω| |∇ρ|

Thus replacing q + 1 by q and choosing the suitable cut-off function, we get

Z  k1  Z 
2qk 2q −2 2 2q
|ω| ≤ C (2q) + τ q |ω|
B(δ) B(δ+τ )
n
Let k = ,q
n−1 i
= k i−1 , δ0 = σ̄ > 1,
σ̄ − 1
δi = δi−1 −
2i
Then we have
Z  k1 Z
2ki 2(i−1)ki−1 i i−1
|ω| ≤ Ck + C(4k) |ω|2k
B(δi ) B(δi−1 )
R  1
j 2kj
We let Ij = B(δj )
|ω|2k . Then we get
1 1 i
Ii ≤ C ki−1 k i−1 + C ki−1 (4k) 2ki Ii−1
Using the standard iteration we get
Ii ≤ C + Ck i
For any q ≥ 2. Let j be such that 2k j−1 ≤ q ≤ 2k j . Using Holder and ||ω||L2 ≤ C we get
||ω||Lq ≤ CIj ≤ C̃q, q ≥ 2.
Since q q ≤ eq q!, we have Z  
|ω|q ≤ C̃ q q q ≤ C̃eq q!
B1
For  small enough
Z ∞
X
|ω|
e ≤ (Ce)q ≤ C
B1 2
This proved the weak Harnack inequality. 
Theorem 5.3 (Harnack inequality). Let u be a positive harmonic function on B(a). Then for any
0 < θ < 1 we have
sup ≤ inf u
B(aθ) B(aθ)

where C only depends on n, (1 − θ)−1 , and the Sobolev constants.

24 PROFESSOR ZHIQIN LU

6. Spectrum on coplete non-compact manifolds

Unlike in the case of compact manifold, in general, a complete manifold doesn’t admit any eigen-
values. For example, there no L2 -eigenvalue on Rn . That is ∀λ ∈ R, if
∆f + λf = 0
and f ∈ L2 (R), then f ≡ 0.
Escobar proved that if M has a relational symmetric metric, then there is no L2 -eigenvalue.
Let ∆ be the Laplace operator on a complete non-compact manifold M. By the argument before,
∆ naturally extends to a self-adjoint densely defined operator, which we still denote ∆ for the sake
of simplicity.
It is well-known that there is a spectrum measure such that
Z∞
∆ = λdE
0

(Here we assume ∆ is the geometric Laplacian, which is a positive operator).

Define e−∆t for any t > 0. Obviously, it is a bounded operator. Thus by the Hahn-Banach theorem
−∆t
e is a bounded operator. The heat kernel is defined as
Z
−∆t
e f (x) = H(x, y, t)f (y)dy

The Green’s function is defined as

Z∞
G(x, y) = H(x, y, t)dt
0

Of course, we need to prove the existence of these functions when M is complete non-compact.
The pure point spectrum of ∆ are those λ ∈ R such that
• There exist a L2 function f 6= 0 such that ∆f + λf = 0.
• The multiplicity of λ is finite.
• In a neighborhood of λ, it is the only spectrum point.
We define
ρ(∆) = {y ∈ R|(∆ − y)−1 is a bounded operator }
σ(∆) = R−ρ(∆) is the spectrum set of ∆. From the above discussion, σ(∆) decomposes as the union
of pure point spectrum, and the so-called essential spectrum, which is by definition, the complement
of pure point spectrum.
Using the above definition, λ ∈ σ(∆) belongs to the set σess (∆), if either
• λ is an eigenvalue of infinite multiplicity, or
• λ is the limiting point of σ(∆).
The following theorems in functional analysis characterizing the essential spectrum.
Theorem 6.1. A necessary and sufficient condition for the interval (−∞, λ) to intersect the essential
some of an self-adjoint densely defined operator A is that, for all  > 0, there exists an infinite
dimensional subspace G ⊂ Dom(A), for which (Af − λf − f, f ) < 0.
Theorem 6.2. A necessary and sufficient condition for the interval (λ − σ, λ + σ) to intersect the
essential spectrum of A is that there exists an infinite dimensional subspace G ⊂ Dom(A) for which
||(A − λI)f || < σ||f ||, f ∈ G.
For reference, see Dormelly [?], Topology 20, 1-14,1981.
Using the above result, we give the following variational characterization of the lower bound of
spectrum and the lower bound of essential spectrum.
 NOTES IN GEOMETRIC ANALYSIS 25

Theorem 6.3. Let R

M
|∇f |2
λ0 = inf R
∞
f ∈C0 (M )
M
f2
and R
|∇f |2 M
λess = sup inf R
∞
k f ∈C0 (M \K) M
f2
where K is a compact set running through an exhaustion of the manifold. Then λ0 and λess are the
lower bound of σ(∆) and σess (∆) respectively.
Proof. We first prove the formula for λess . Let
λ0ess = inf σess (∆)
if σess (∆) = ∅, we define ,λ0ess = +∞. By the definition, ∀ > 0,
(−∞, λ0ess + ) ∩ σess (∆) 6= ∅.
Then we can find infinite dimensional space V such that for any f ∈ V
(6.1) h∆f − (λ0ess + )f, f i < 0
R
Without loss of generality, we may assume that f 2 = 1. Also, without loss of generality, we may
M
assume that all elements in V are smooth.
We leave as an exercise to prove that (??) implies
Z
|∇f |2 ≤ (λ0ess + )
M

Now we assume that K is a compact set. Let K 0 a larger ball containing K. Let ρ be the cut off
function such thatR ρ ≡ 1 on K but ρ ≡ 0 outside K 0 . We claim that (key point) that ∀ > 0, there
is an f ∈ V with f 2 = 1 but
Z
ρ2 f 2 < 
M
If the above is not true, then for any f ∈ V
Z Z
2 2
ρ f ≥ 0 , if f2 = 1
M
Since the set f ∈ V is of infinite dimensional, the set ρf is of infinite dimensional. Thus we can find
an orthogonal basis Z
ρ2 fi fj = 0, i 6= j
R
while we can still keeping f 2 = 1. We consider
Z Z Z
|∇(ρfi )| ≤ 2 |∇ρ| fi + 2 ρ2 |∇fi |2
2 2 2

≤ 2C + 2(λ0ess + )
Thus R
∇(ρfi )2 2C + 2(λess + )
R ≤
(ρfi )2 0
for infinitely dimensional space. This is a contradiction because on the compact set K 0 , the eigen-
values go to infinity. R
With the above preparation, we can prove out theorem ∀ > 0, we find an f with f 2 = 1 but
Z
ρ2 f 2 < 
26 PROFESSOR ZHIQIN LU

Consider ρ1 = 1 − ρ,
Z Z Z Z
2
|∇(ρ1 f | = ρ21 |∇f |2 +2 ρ1 f ∇ρ1 ∇f + f 2 |∇ρ1 |2 .

Using the above inequality, we have

Z Z Z
1
f 2 |∇ρ1 |2 ≤ C, 2 ρ1 f ∇ρ1 ∇f = ∆ρ21 f 2 ≤ C
2
Note that the support of ∇ρ1 ,∆ρ1 are within K 0 . Thus
Z
|∇(ρ1 f )|2 ≤ (λ0ess + ) + C
Z
ρ21 f 2 ≥ 1 − 

Using the definition, we get

λess ≤ λ0ess + 
and thus λess ≤ λ0ess . The other direction is easier to prove. 
Using the same method, we can prove the case for λ0 , provided that we need a similar theorem
like ??.
It is an interesting question to compute the set of essential. It is particularly interesting to get the
lower bound estimate for the essential spectrum because of the following Theorem.
Theorem 6.4. Suppose λ0 < λess , then λ0 is an eigenvalue of M with finite dimensional eigen space.
That is ,there exists an L2 function f 6= 0, such that
∆f = −λ0 f
which is very strong result.
In what follows we use R to elaborate our theorem. First let’s compute λess for R. We claim that
λess = 0. By definition, this is equivalent to say that ∀ > 0, there is a function f ∈ C0∞ (Rn − B(R))
such that Z Z
|∇f |2 <  f2

This is obvious: let B(x0 , R0 ) be a ball radius R0 with center x0 such that |x0 | > 2R + 1 + R0 . Then
B(x0 , R0 ) ⊂ Rn \ B(R). Let f be the first Dirichlet eigenfunction of B(x0 , R0 ). Then if R0 → ∞
R
|∇f |2
R <
f2
Zero extending f to R \ B(R) we get the result.
Using ??, we can even prove that
σess (∆) = [0, +∞)
q the above, we make the following observation ∀λ > 0, ∀m ∈ Z, we can find a square of
To prove
size mπ nλ such that
r r
λ λ λ
f = sin x1 · sin x2 · · · sin xn
n n n
is an eigenfunction with Dirichlet condition: ∆f + λf = 0.
However, we can’t use Theorem ?? directly, the reason is that f is not second differentiable near
the boundary. Thus we need to use cut off function. Without loss of generality, we may assume that
the square is in the first quadratic, we denote such a square to be S. Let S1 be a square with the
same center as S1 but with smaller size. We assume that the distance of the boundary of S1 to S is
 NOTES IN GEOMETRIC ANALYSIS 27

d, which is to be determined later. We choose a cut off function ρ such that ρ ≡ 1 on S1 and ρ ≡ 0
outside S. By the same argument as before we can prove that
||∆(ρf ) + λρf ||L2 < ||ρf ||L2
Thus σess (∆) = [0, ∞).
Unfortunately, the above argument doesn’t apply to the general case. Thus following result of
Wang Jiaping is very surprising and interesting
Theorem 6.5 (J-P. Wang). Let M be a complete manifold with non-negative Ricci curvature. Then
σess (∆) = [0, ∞).
Note that λess was known before, e.g. P.Li-Wang, Books.

7. Heat kernel and green’s functions on complete manifold

Let M be a Riemannian manifold. The Laplace operator ∆ can be extended as a densely defined
self-adjoint operator. Thus by spectrum theorem, we can write
Z
∆ = λdE

where E is the corresponding spectrum measure. Using functional analysis, we define

Z∞
−∆t
e = e−λt dE
0

to be the heat operator, and

Z∞
e−∆t dt
0
to be the Green’s operator.
The heat semi-group, or the heat operator, has a kernel.
Theorem 7.1. Let M be a complete Riemannian, then there is a heat kernel H(x, y, t) ∈ C ∞ (M ×
M × R+ ) such that
Z
−∆t


e f (x) = H(x, y, t)f (y)
M
2
∀f ∈ L (M ) such that
• H(x, y, t) = H(y, x, t).
• limt→0+ H(x, y, t) = δx (y)
∂
• (∆ − ∂t )H =R 0
• H(x, y, t) = H(x, z, t − s)H(z, y, s)dz
Remark 7.2. Let M be a compact manifold and let {fi } be an orthonormal basis of eigenfunctions.
Let λi be the corresponding eigenvalues. Then
∞
X
H(x, y, t) = e−λk t fk (x)fk (y)
k=1

As in the case of harmonic functions, for positive solutions of the heat equations
∂
(∆ − )u = 0
∂t
we also have the differentiable Harnack inequality. The theorem is as follows:
28 PROFESSOR ZHIQIN LU

Theorem 7.3. Let M be a compact Riemannian manifold with boundary, Ric(M ) ≥ 0. If ∂M 6= ∅,

we assume that ∂M is convex. In this case, we assume that u(x, t) satisfies the Neumann boundary
∂u
=0
∂n
∂
on ∂M × (0, ∞), where ∂n
is the outer normal direction. Then on M × (0, ∞), we have
|∇u|2 ut n
2
− ≤
u u 2t
For manifold with Ric(M ) ≥ −k, the estimates are more complicated but the same principle
applies.
The Harnack inequalities follows from the gradient estimates. Using the gradient estimates and
the Harnack inequality we obtain
Theorem 7.4. Let M be a complete manifold without boundary, Ric(M ) ≥ −k, k ≥ 0. Let H(x, y, t)
be the fundamental solution of the heat equation
∂
(∆ − )u(x, t) = 0
∂r
Then for any δ ∈ (0, 1), we have
−1 √ −1 √
 2 
r (x, y)
H(x, y, t) ≤ C(δ, n)Vx 2 ( t)Vy 2 ( t)exp − + C1 δkt
(4 + δ)t
For the proof, see the book of Yau and Schoen. We give some application of the above theorem.
Theorem 7.5. (Gromov) Let M be a compact manifold without boundary, let Ric(M ) ≥ 0 and d be
diameter of M. Then
C(n) 2
λk ≥
(k + 1) n
d2
where λk is the kth eigenvalue of the manifold.
Proof. We set x = y and k = 0 in the above theorem. Then we have
√
H(x, x, t) ≤ C(n)Vx−1 ( t)
Since
∞
X
H(x, y, t) = e−λk t fk (x)fk (y)
k=1
we have
∞
X
H(x, x, t) = e−λk t
k=1
As a result, we have
∞
X
−λk t
Z √
e ≤ C(n) Vx−1 ( t)dx
k=1 M
√ √ √
if t ≥ d, then Vx−1 ( x) = vol(M ), if t ≤ d, then by the Bishop volume comparison theorem
√  √ n
Vx ( t) t
≥
Vx (d) d
Thus
∞
(  n
d
X
−λk t √ t ≤ d2
(7.1) e ≤ C(n) t

k=1 1 t > d2
 NOTES IN GEOMETRIC ANALYSIS 29

For fixed k, by the monotonicity, we have

(  n
d
√ t ≤ d2
ke−λk t ≤ C(n) t
1 t > d2
√ 1
We let t = k − n d and the result follows. 

We can compare the above eigenvalue estimate with the result of Cheng-Li. From (??), we can
get
∞
n
X
e−λk t ≤ Ct− 2 , t ≤ d2
k=1

where C is the absolute constant. On the other hand, if we consider

Z Z
∂ 2
H(x, y, t) dy = 2 H(x, y, t)∆H(x, y, t)dy
∂t M M
Z
= −2 |∇y H(x, y, t)|2 dy
Z  n−2
n
2n
≤ −2C |H(x, y, t)| n−2 dy
M
R
where C is the Sobolev constant. Since for any t > 0, H(x, z, t)dz ≤ 1. So we have
Z n−2 Z  2+n
2n n n
|H(x, y, t)| n−2 dy ≥ |H(x, y, t)|2 dy
M M

Proof. Let f = H(x, y, t). Then

Z Z Z  n+2
2n
Z  n+2
4
2n 4 2n 4
2 +
f = f n+2 n+2 ≤ f n−2 f n+2

Thus we have
Z Z  2+n
∂ 2 2
n
H(x, y, t) dy ≤ −C H(x, y, t)
∂t M M

Since limt→0 H(x, x, t) = ∞. Thus

 − n2
4
H(x, x, 2t) ≤ Ct
n
Using the same method as before we have
  n2
k
λk ≥ C
vol(M )
The second application of the heat kernel estimate is following resolvent estimate.
(n/2)
Theorem 7.6. For any β > 0 there is n ∈ N , α < 0, C < ∞, such that the integral kernel gα (x, y)
n
of (∆ − α)− 2 satisfies
gα(n/2) (x, y) ≤ Cψ 2 (x)eβd(x,y)


where ψ(x) = v(B1 (x))1/2
30 PROFESSOR ZHIQIN LU

Proof. First, using the heat kernel H(x, y, t) we can write

Z∞
gα(n/2) (x, y) = Cn H(x, y, t)tn/2−1 eαt dt
0

By the volume comparison theorem, we have

v −1/2 (B√t (x))v −1/2 (B√t (y)) ≤ C2 ψ(x)2 sup(1, t−N/2 )eβ2 d(x,y)
By an element calculation, for anyβ3 > 0, we have
 2 
d (x, y)
exp − ≤ exp(−β3 d(x, y))exp(C1 β32 t)
C1 4t
Thus for any β > 0, there exist α > 0, C3 < ∞ such that
H(x, y, t) ≤ C3 ψ 2 (x)e−βd(x,y) sup t−N/2 , 1 e−(α+1)t


which implies
gα(n/2) (x, y) ≤ C4 ψ 2 (x)e−βd(x,y) . 
Cheeger-Yau’s Heat kernel comparison theorem.
Theorem 7.7. Let M be a complete Riemannian manifold such that Ric(M ) ≥ 0. Fixing x ∈
M, r0 > 0. Then the heat kernel H(x, y, t) in B(x, r0 ) and the heat kernel (r(x, y), t) in V (k, r0 )
satisfies the following
(r(x, y), t) ≤ H(x, y, t)
For both Dirichlet and Neumann conditions.
Proof. Using the property of the heat kernel, we have
Zt Z
d
H(x, y, t) − (x, y, t) = ((x, z, t − s)H(z, y, s))dzds
B(x,r0 ) ds
0
Zt Z
d
=− ( (r(x, z), t − s))H(z, y, s)dzds
B(x,r0 ) ds
0
Zt Z
d
+ (r(x, z), t − s) H(z, y, s)dzds
B(x,r0 ds
0
Zt Z
=− ˜
∆(r(x, z), t − s)H(z, y, s)dzds
B(x,r0 )
0
Zt Z
+ (r(x, z), t − s)∆H(z, y, s)dzds
B(x,r0 )
0

Using the Green’s formula, under either the Dirichlet or Neumann boundary condition, we have
Z Z
(r(x, z), t − s)∆H(z, y, s)dz = ∆(r(x, z), t − s)H(z, y, s)dz.
B(x,r0 ) B(x,r0 )

Since H(x, y, s) > 0, we just need to prove that

˜
∆(r(x, z), t − s) ≤ ∆(r(x, z), t − s)
 NOTES IN GEOMETRIC ANALYSIS 31

This essentially follows from the Laplacian comparison theorem: Let x = (r, ξ), ξ ∈ S n−1 . Then
2
˜ = ∂ + m(r) ∂ , ∂ p
∆ m(r) = log det g̃
∂r2 ∂r ∂r
∂2 ∂ ∂ √
∆ = 2 + m(r, ξ) , m(r, ξ) = log det g
∂r ∂r ∂r
Since Ric(M ) ≥ (n − 1)k, using the volume comparison theorem, we have m(r, ξ) ≤ m(r). Since
∂
∂r
< 0, we have
˜
∆(r, t − s) ≤ ∆(r, t − s)

In the above proof, we didn’t take the cut-locus into a count, using some kind of limiting process,
we can overcome the difficulty.
Theorem 7.8. Let M be a complete Riemannian manifold such that Ric(M ) ≥ (n−1)k, n = dim M .
We use B(x0 , r) to denote the ball centered at x0 with radius r. Let V (k, r) be the ball of radius r in
a simply connected space form. Then with the Dirichlet boundary condition we have
λ1 (B(X0 , R)) ≤ λ1 (V (k, r))
Proof. Let H(x, y, t) and (x, y, t) be the corresponding heat kernel. Then we have
X
H(x, y, t) = e−λ1 t φ2i (x)

X
(x, y, t) = eλ̃1 t φ̃2i (x)

If we let t → ∞, we get λ1 ≤ λ̃1 . 

S.Y.Cheng concretely computed the upper bounds of the eigenvalue.
• If Ric(M ) ≥ 0, then λ1 ≤ 2n(n+4)
d2
2
• If Ric(M ) ≥ n − 1, then λ1 ≤ n πd2 (P. Li conjecture)
• If Ric(M ) ≥ (n − 1)(−k), then
1 Cn
λ1 ≤ k + 2 , Cn = 2n(n + 4)
4 d
Unfortunately, for Neumann boundary condition, we don’t have the comparison theorem for the
1st eigenvalues directly. Under the Neumann boundary condition, the first eigenvalues are always 0.
The comparison theorem for that is trivial.
It is thus interesting to have the following result.
Theorem 7.9. If M is a compact manifold without boundary, then
d
λ1 (M ) ≤ λ1 (V (k, ))
2
where d is the diameter of M .

8. Green’s function and parobolicity

Let M be a complete non-compact Riemannian manifold. The Green’s function is a smooth
function on M × M \ diag(M ) such taht
• G(x, y) = G(y, x) and fixing y, we have ∆x G(x, y) = 0, ∀x 6= y.
• G(x, y) ≥ 0.
• Fixing y, when x → y, we have the following

ρy (x)2n (1 + o(1)), n>2
G(x, y) =
− log ρy (x)(1 + o(1)), n=2
32 PROFESSOR ZHIQIN LU

This last asymptotical expansion of the Green’s function also implies that
∆x G(x, y) = −δx,y .
From the asymptotical behavior we can find that n = 2 and n > 2, the Green’s functions are very
different. We make the following
Definition 8.1. A complete manifold is said to be parabolic, if and only if it doesn’t admit a positive
Green’s function. Otherwise it is said to be non-parabolic.
Definition 8.2. An End, E, with respect to a compact subset Ω ⊂ M an unbounded connected
component of M \ Ω. The number of ends with respect to Ω, denoted by NΩ (M ), is the number of
unbounded connected component of M \ Ω.
Definition 8.3. An end E is said to be parabolic, if it doesn’t admit a positive harmonic function
f satisfying
f ≡ 1 on ∂E
and
lim f (y) < 1
n→E(∞)
where E(∞) denotes the infinity of E. Otherwise, E is said to be non-parabolic and the function f
is said to be a barrier function of E.
We prove the following result:
Theorem 8.4. Let E be an parabolic end. Let A(R) = E ∩ ∂B(r) where B(R) is the ball of radius
R with respect to some reference point. Let f be a harmonic function on E such that
f |∂E = 1, f |A(R) = 0
Then Z
lim |∇f |2 → 0
R→∞ E
Proof. Using the Green’s formula, we have
Z Z
2 ∂f
|∇f | = −
E E ∂r
where ∂
∂r
is th outer normal direction, we claim that ∂f∂r
→ 0. To see this, we take a sequence
R1 < R2< · · · < Rk → ∞. The corresponding harmonic function fi = fRi . By the maximal
principle, fi are increasing sequence on any compact set of E. Let
lim fi = f
i→∞
Then f must be a positive harmonic function. by the parabolicity, f ≡ 1. By the maximal principle
again
∂fi
→0
∂r
as i → ∞. 
Example 8.5. R2 and Rn . Let R > 0 be a big number. Let
F (R) = {f ∈ X0∞ (R) : f = 1 for |x| < R, f rotational symmetric }
If n > 2, then for any c > 0, there exists an R0 such that for any R > R0
Z
|∇f |2 > C
Rn
for any f ∈ F (R). If n = 2, then for any  > 0 there exists R0 > 0 such that for any R > R0 , we
can find an fk ∈ F (R) for which Z
|∇f |2 < 
R2
 NOTES IN GEOMETRIC ANALYSIS 33

Proof. If n > 2, then

Z∞
1 1 1
dr =
rn−1 n−2R n−2
R
Thus we have
Z Z∞  2 Z∞
∂f 1
2
|∇f | ≥ (n − 2)CR n−2
r n−1
dr dr ≥ (N − 1)CrN −2 → +∞
Rn ∂r rn−1
R 0

However, for n = 2, we define fR = σR (|x|) such that


 1,
 t≤R

−1  log R 
σR (t) = 1 − logRR log t
− log R
R
, R < t < eR

0, t ≥ eR


Then a straightforward computation gives

Z∞
4 1
t|σR0 (t)|2 dt ≤ , for R >> 0. 
3 log R
0

The main result of this section is the following characterization of parabolicity.

Theorem 8.6. Let M be a complete manifold. If M is non-parabolic, then for any point p ∈ M , we
have
Z∞
dt
< +∞
Ap (t)
1
where Ap (r) denotes the area of ∂Bp (r).
Proof. For p ∈ M , Let G(p, y) be the Green’s function ( assuming that it exists). Let Gi (p, y) be the
Green’s function on Bp (Ri ) with the Dirichlet boundary condition, where Ri → +∞.
For any 1 < R < Ri , let’s denote that
Si (1) = sup Gi (p, y)
y∈∂Bp (1)

ii (R) = inf Gi (p, y)

y∈∂Bp (R)

Let f be the harmonic function defined on Bp (R) \ Bp (1) satisfying the boundary conditions
f (y) = si (1) on ∂Bp (1)
f (y) = Gi (p, y) on ∂Bp (R)
The maximum principle implies that
f (y) ≥ Gi (p, y) on Bp (R) \ Bp (1)
In particular, we have
∂f ∂Gi
≤ on ∂Bp (R)
∂r ∂r
On the other hand, since f (y) is harmonic, Stokes theorem implies that
Z Z Z
∂f ∂f
0= ∇f = −
Bp (R)\Bp (1) ∂Bp (R) ∂r ∂Bp (1) ∂r

Also, we observe that Z Z

∂Gi
= ∆Gi = −1
∂Bp (R) δr Bp (R)
34 PROFESSOR ZHIQIN LU

Thus we get Z
∂f
≤ −1
∂Bp (1) ∂r
Let’s now consider h to be the harmonic function defined on Bp (R) \ Bp (1) satisfying the boundary
conditions
h(y) = Si (1) on ∂Bp (1)
h(y) = ii (R) on ∂Bp (R)
Again the maximum principle implies that h(y) ≤ f (y) on Bp (R) \ Bp (1) and
∂h ∂f
≤ on ∂Bp (1)
∂r ∂r
Thus we have Z Z
∂h ∂h
= ≤ −1
∂Bp (R) ∂r ∂Bp (1) ∂r
Define the function
 −1
ZR ZR
dt  dt
g(r) = (Si (1) − ii (R))  + ii (R)
Ap (t) Ap (t)
1 r

Then g(r(y)) will have the same boundary conditions as h(y). The Dirichlet integral minimizing
property for harmonic functions implies that
Z Z
2
|∇h| ≤ |∇g|2
Bp (R)\Bp (1) Bp (R)\Bp (1)
  R −1 2
ZR Z
dt  1 
= (Si (1) − ii (R))   Ap (r)dr

Ap (t) Ap (r)
1 1
 R −1
Z
dt 
= (Si (1) − ii (R))2 
Ap (t)
1

On the other hand, integration by parts yields

Z Z Z
2 ∂h ∂h
|∇h| = ii (R) − Si (1) ≥ Si (1) − ii (R).
Bp (R)\Bp (1) ∂Bp (R) ∂r ∂Bp (1) ∂r

By taking the limit

ZR
dt
≤ sup G(p, y) − inf G(p, y). 
AP (t) y∈∂Bp (1) y∈∂Bp (R)
1

Corollary 8.7. Let M be a complete manifold such that

vol(B(R)) ≤ CR2
Then M has to be parabolic.
Corollary 8.8. Let M be a Riemann surface such that
Z
|K| < +∞
M
Then M is parabolic.
Proof of a theorem of Brascamp and Lieb.
 NOTES IN GEOMETRIC ANALYSIS 35

Theorem 8.9. Let Ω be a bounded convex domain of Rn . Let u be the first Dirichlet eigenfunction.
Let
∆u = −λ1 u, u > 0.
Then log u must be concave.
Proof. We choose any function u0 > 0, u0 |∂Ω = 0 such that − log u0 is concave. Such a function
always exists. For example, we can take the convex hull of graph of − log u.
Consider the flow
∂u
= ∆u + λ1 u, u|∂Ω = 0
∂t
We assume that ut → u, the first eigenfunction. We are going to use the maximum principle. Let
T be the biggest number such that det(−∇2 log u) is degenerated. Thus there is an x0 ∈ M and a
direction i such that
−(log u)ii (x0 ) = 0
and for other j, (− log u)jj (x0 ) ≥ 0. Let ϕ = − log u. Then the evolution of ϕ is
∂ϕ
= ∆ϕ − |∇ϕ|2 − λ1
∂t
By the maximum principle, ϕiik = 0, ∂ϕ ii
∂t
≤ 0 and ∆ϕii ≥ 0. Thus
∂ϕii
0≥ = ∆ϕii − 2ϕk ϕkii − 2ϕ2kl ≥ −2ϕ2ki .
∂t
However by convexity, ϕ2ki ≤ ϕii · ϕkk = 0, ϕki ≡ 0. The theorem follows from strong maximum
principle. 

9. The lower bound of λ2 − λ1

Let Ω be a smooth bounded domain in Rn , λ1 , λ2 be the first two eigenvalues with respect to the
Dirichlet boundary condition. Then
π2
λ2 − λ1 ≥ 2
4d
where d is the diameter.
Proof. Let u2 , u1 be the second and the first eigenfunctions, respectively. By the variational principle,
we must have u1 (x) > 0. Furthermore, near the boundary, we must have ∇u1 (x) 6= 0. Thus the
function
u2
v=
u1
is smooth up to the boundary. By a simple computation, we obtained
∆v = λv − 2(∇v · ∇ log u1 )
where λ = λ2 − λ1 > 0. Let G be the function Ω̄ → R
G = |∇v| + λ(µ − v), µ > sup v
Then G is a smooth function. Let x0 ∈ Ω̄ be the maximum point of G. Then we claim that
G ≤ sup λ(µ − v)2
Ω

In fact, if x0 ∈ ∂Ω. By choosing an orthonormal frame {e1 , · · · , en } such that e1 be the out normal
direction. If we let e1 |∂Ω = ∂x∂ 1 . Then
n
∂G X
(x0 ) = 2 vi vi1 − 2λv1 (µ − v)
∂x1 i=1
We claim that
∂G
(x0 ) ≥ 0
∂x1
36 PROFESSOR ZHIQIN LU

∂v
To see this, we fist observed that ∂x1
= 0. This can be proved using the variational principle. Thus
n
∂G X
(x0 ) = 2 vi vi1
∂x1 i=2

Using the definition of the second fundamental form, we get

X
vi1 = − hij vj

Thus we have
∂G X
0≤ (x0 ) = −2 hij vi vj
∂x1
Since Ω is assumed to be convex, all vi = 0. Thus
G(x) ≤ sup λ(µ − v)2
x∈Ω

Next we assume that x0 ∈ Ω. Then at the maximum point, we have

∇G(x0 ) = 0, ∆G(x0 ) ≤ 0
We apply the standard gradient estimate: assume that at x0 , ∇v 6= 0. Then we have
0 = Gi (x0 ) = 2vj vji − 2λ(µ − v)vi
X X X
0≥∆=2 vij2 + 2vi vjii + 2λ vi2 − 2λ(µ − v) vii

Since ∇v 6= 0, we can choose local orthonormal frame such that v1 6= 0, vi = 0(i > 1). Thus at x0 ,
we have
v11 (x0 ) = λ(µ − v)
vi1 (x0 ) = 0, 2 ≤ i ≤ n
Thus
X
0 ≥ ∆G = 2 vij2 + 2λ(µ − v)v − 4v12 (log u1 )11
Using a result of Brascamp and Lieb, log u1 is an concave function. Thus (lg u)11 (x0 ) ≤ 0. Thus
X
vij2 + λ2 (µ − v)v ≤ 0

Or in other word
v11 + λ2 v(µ − v)|x0 ≤ 0
which is not possible. Thus we must have ∇v = 0 at x0 and thus
G(x) ≤ sup λ(µ − v)2
we have
√ |∇v|2
λ≥ p
(sup v − inf v)2 − (sup v − v)2
Using the same method as in the estimate of the first eigenvalue, we get
π2
λ1 ≥
4d2

 NOTES IN GEOMETRIC ANALYSIS 37

Remark 9.1. Using the method of Zhong-Yang, Zhong-Yu was able to modified the above estimate
to
π2
λ2 − λ1 ≥ 2
d
However, even the above estimate is not optimal, Van der Berg ( see also Yau, Problem session)
conjectured that
3π 2
λ2 − λ1 ≥ 2
d
The estimate is asymptotically accurate for a very thin rectangular.
Not much progress was made in the direction of this conjecture. In JFA, 176, 368-399(2000),
Banuelos and Mendez-Hernandez proved that if Ω is a convex domain in R2 which is symmetric with
respect to both x− and y−axises, then the Van der Berg conjecture is true.
For a triangle, Lu and Rowlett proved the following
Theorem 9.2. Let ∆ be a triangle and let d = d(∆) be the diameter of ∆. That is, d is the longest
side of the triangle. Then ∀c > 0
(λ2 (∆) − λ1 (∆))d2 (∆) ≤ C
is a compact set.
Proof. When a sequence of triangles doesn’t converge, then the smallest angle must go to zero. We
assume that the smallest angle is απ and we assume that the diameter d = 1. The key part of the
proof is the following cutting lemma.
Let P2 , P1 be two points on BC such that |P1 C| = α , |P2 C| = 2α , where  < 29 . For the sake

of simplicity, we assume that ∠ACB = π2 . We are going to prove that the eigenvalues of ABC and
AQ2 P2 C are almost the same, thus cutting an acute angle won’t make too much difference.
38 PROFESSOR ZHIQIN LU

We let U = AQ1 P1 C and U 0 = AQ2 P2 C. Let fi be the eigenfunctions for λi = λi (ABC), i = 1, 2.

The height of V = ABC \ U is at most
(1 − α ) tan απ ≈ (1 − a )a
Above we have dropped a constant factor of π on the right side for simplicity in the arguments to
follow. By the one dimensional Poincare Inequality
R R
VR
|∇fi |2 1 V0
|∇fi |2 1
2
≥  2 2
, R ≥
f
V i
(1 − α ) α V0
|fi |2 (1 − 2d )2 α2
where as before V = ABC \ U , V 0 = ABC \ V 0 .
On the other hand, we always have
R 2
R
UR
|∇f i | 1 UR0
|∇fi |2 1
2
≥ 2
, 2
≥ 2.
f
U i
α f
U0 i
α
We normalize fi to be Z
fi2 = 1
ABC
R
Let fi2 = β. Then by the variational principle, we have
V
1−β
Z Z
β 2
 2 2
+ 2
≤ |∇fi | + |∇fi |2 = λi .
(1 − a ) α α V U
By the asymptotic estimate of the eigenvalues, using the Bessel functions, we have
1 C2
λi (ABC) ∼ 2 + 4/3
α α
Therefore we have
α2/3 C2 (1 − α )2 C2 2/3−
β≤  
≤ α
α (2 − α ) 2
For simplicity in the arguments to follow, we will replace the constant factor C2 /2 by a constant
factor 1, since we are considering α → 0.
Let ρ be a smooth compactly supported function so that
ρ|V ≡ 1, ρ|V 0 ≡ 0
We may also assume that
1 1
|∇ρ| ≤
, |∆ρ|, |∆ρ2 | ≤ 2
d d
For the arguments to follow, we use the sign convention for Euclidean Laplacian so that −∆ has
positive spectrum. Not that
(9.1) −(ρfi )∆(ρfi ) = λi ρ2 f1 − f12 ρ∆ρ − 2f1 ρ∇ρ∇fi
In the estimates to follow, we will absorb all constants multiplying factors of aδ for δ > 0 into a
factor one, since it is clear that as α → 0, no generality is lost by this assumption.
Estimate for λ1 (U 0 ). We may use ρf1 as a test function for the Rayleigh quotient on U 0 to estimate
λ1 (U 0 ) for above. By (??), we have
R 2
0 0 −ρ∆ρf1 − 2ρ∇ρf1 ∇f1
λ1 (U ) ≤ λ1 (ABC) + U R
U0
ρ2 f12
Since Z Z Z
1 2 2 1
ρ∇ρf1 ∇f1 = ∇ρ ∇f1 = − ∆ρ2 f12
U0 2 U0 2 U0
and ∇ρ, ∆ρ = 0 on U , we must have
Z Z
1
−ρ∆ρf1 − 2ρ∇ρf1 ∇f1 ≤ 2 f12 ≤ α2/3−3
U0 α U 0 −U
 NOTES IN GEOMETRIC ANALYSIS 39

Noting that Z Z
2
ρ f1 ≥ ρ2 f12 ≥ 1 − β ≥ 1 − α2/3−r
U0 U
we then have
α2/3−3
λ1 (U 0 ) ≤ λ1 (ABC) + ≤ λ1 (ABC) + α2/3−3
1 − α2/3−
By modifying the above argument, we are able to prove that
λ2 (U 0 ) − λ2 (ABC) ≤ α2/3−3 + (λ2 − λ1 )α2/3−3
Solving the above inequality, we have
λ2 − λ1 ≥ λ2 (U 0 ) − λ1 (U 0 ) − O(α2/3−3 )
By the gap theorem, and using the fact that the diameter of U 0 is at most 4α , we get
π2
λ2 (U 0 ) − λ1 (U 0 ) ≥
16α2
The compactness theorem follows. 

10. Essential spectrum on complete noncompact manifold

In this lecture, we study the essential spectrum of a complete non-compact manifold with non-
negative Ricci curvature. We are going to prove that the set of essential spectrum is [0, ∞). Then
reference papers for this lecture are [?] and [?].
Through this lecture, M is a complete Riemannian manifold with non-negative Ricci curvature.
We say the volume (M, g) grows uniformly sub-exponentially, if for any  > 0, there is a constant
C < ∞ such that for all r > 0 and all x ∈ M
ν(Br (x)) ≤ Cer ν(B1 (x))
Theorem 10.1. If the volume of (M, g) grows uniformly sub-exponentially, then the spectrum σ(∆p)
of ∆p acting on Lp (M ) is independent of p ∈ [1, ∞). In particular, it is a subset of the real line.
Note that by the Bishop volume comparison theorem, Ricci nonnegative implies uniformly sub-
exponentially volume growth.
The theorem can be proved using the resolvent estimates, which are based on the previous heat
kernel estimate.
We begin with the following
Lemma 10.2. If the volume of (M, g) grows uniformly sub-exponentially, then for any  > 0
Z
1 1
sup e−d(x,y) (ν(B1 (x)))− 2 (ν(B1 (y)))− 2 dν(y) < ∞
x∈M M

Proof. We take r = d(x, y). Then since B1 (y) ⊂ Br+1 (x), we must have
1 1 1 1
ν(B1 (y)))− 2 ≥ ν(Br+1 (x))− 2 ≥ Ce− 2 (r+1) ν(B1 (x))− 2
For any x. Thus the integration in the lemma if less than
Z
1
C e−r · e 2 (r+1) ν(B1 (x))−1 dy
M
Rr
We let f (r) = vol(∂Br (x)) and F (r) = f (t)dt. Then up to a constant, the above expression is less
0
than
Z∞
1
−1
(ν(B1 (x))) e− 2 r f (r)dr
0
n−1
By volume comparison again, f (r) ≤ Cr ν(B1 (x)). The lemma follows. 
40 PROFESSOR ZHIQIN LU

Lemma 10.3. For any β > 0, there is an n ∈ N, α < 0, C < ∞ such that the integral kernel
n/2
gα (x, y) of (∆ − α)−n/2 exists and satisfies
gαn/2 (x, y) ≤ Ce−βd(x,y) ϕ(x)2
1
where ϕ(x) = (ν(B1 (x)))− 2 .
The lemma was proved on page 40 theorem 9.6 , using the heat kernel estimates.
Before gonging further, let’s make some remarks on the kernel of an operator. Let A be an operator
on functions. If there is a function g(x, y) such that
Z
Af (x) = g(x, y)f (y)dy
M
then call g(x, y) the kernel of A. However, in general, the kernel doesn’t exist.
To see why the kernel in general not exist, we let ξ ∈ ρ(∆). To be more specific, ∆ is an operator
on L2 (M ), so we assume that ξ ∈ ρ(∆2 ). The operator (∆2 − ξ)−1 is called the resolvent. It is a
bounded operator form L2 (M ) → L2 (M ). (By Hahn-Banach theorem, it can be extended to whole
L2 (M )). However, in general, the kernel doesn’t exist, if not, we let fi → δ be a sequence converges
to the δ−function in L1 (M ). Then ∆2 fi could have been bounded. we need estimate to extend ∆2
form L2 (M ) to L1 (M ).
Lemma 2 told us that for α < 0, (∆ − α)−n/2 has a kernel. The operator can be extended to
L1 (M ). Further more, the kernel exponentially decays.
The Laplacian we used here is the geometric Laplacian. That is, it is a positive operator.
For our purpose, we just need to prove σ(∆1 ) ⊂ σ(∆2 ), which is also the major part of the paper
of Sturm.
Recall that x ∈ ρ(A), if (A−xI)−1 is a bounded operator. We define the spectrum σ(A) = R−ρ(A).
Note that σ(∆2 ) ⊂ [0, ∞). Then for α < 0, (∆2 − αI)−1 is bounded. Since
∆ − αI = (∆ − ξI)(1 − (α − ξ)(∆ − ξI)−1 )
we have
(∆ − ξI)−1 = (1 − (α − ξ)(∆ − ξ)−1 )(∆ − αI)−1
or for any n
(∆ − ξI)−n = (1 − (α − ξ)(∆ − ξ)−1 )n (∆ − αI)−n
From the above identity, the kernel for (∆ − ξI)−n exists. Let
A = (1 − (α − ξ)(∆ − ξ)−1 )n
Then
Ax (gα(n/2) (x, y))
is the kernel of (∆ − ξI)−n
Lemma 10.4. Let g(x, y) be the kernel of (∆ − ξI)−n . Then
|g(x, y)| ≤ Ce−d(x,y) ϕ(x)ϕ(y)
1
where ϕ(x) = ν(B1 (x))− 2 .
We omit the proof of the above lemma.
By lemma ?? Z
sup |g(x, y)|dy < +∞
x
Thus (∆2 − ξ)−n is an operator from L1 → L1 .
Since ∆2 and ∆1 are the same acting on C ∞ -functions, (∆1 − ξ)−n is a bounded operator from
L (M ) → L1 (M ).
1

Lemma 10.5. if (∆1 − ξ)−n is an bounded operator. Then ξ ∈ ρ(∆1 ).

 NOTES IN GEOMETRIC ANALYSIS 41

Proof. Since (∆1 − ξ)−n is bounded, there is a neighborhood of ξ such that for any ξ 0 in the neigh-
borhood, (∆1 − ξ 0 )−n is also bounded. By
(∆1 − ξ)−1 = (∆1 − ξ 0 )−1 (1 − (ξ − ξ 0 )(∆1 − ξ 0 )−1 )−1
and the fact that the latter can be expanded to a convergent series, (∆1 − ξ)−1 is bounded in L1 (M ).
Thus ξ ∈ ρ(∆1 ). 
From the above lemma, we have
σ(∆1 ) ⊂ σ(∆2 ) ⊂ [0, ∞)
The theorem thus follows from
Theorem 10.6. (Wang) Let M be a complete Riemannian manifold with non-negative Ricci curva-
ture. Then σ(∆1 ) = [0, ∞)
Before giving the proof, we use the following theorem.
Theorem 10.7. Let M be a complete non-compact Riemannian manifold. Ric(M ) ≥ 0. Then
vol Bp (R) ≥ C(n, volBp (1))R
Proof. Fixing x0 ∈ ∂Bp (R), using the comparison theorem we have ∆ρ2 ≤ 2n. For any ϕ ∈ C0∞ (M ),
ϕ ≥ 0, we have
Z Z
2
ϕ∆ρ ≤ 2n ϕ
M M
We choose a standard cut-off function ϕ = ψ(ρ(x)), where

 1, 0≤t≤R−1
1
ψ(t) = (R + 1 − t), R−1≤t≤R+1
 2
0, t≥R+1
By Stokes theorem
Z Z
2
ϕ∆ρ = − ∇ϕ∇ρ2
M
Z
= −2 ψ 0 ρ|∇ρ|2
Z
= ρ
Bx0 (R+1) \Bx0 (R−1)

≥ (R − 1)vol(Bx0 (R + 1) \ Bx0 (R − 1))

Thus Z
(R − 1)vol(Bx0 (R + 1) − Bx0 (R − 1)) ≤ 2n ϕ ≤ 2n vol Bx0 (R + 1)
Obviously
Bp (1) ⊂ bX0 (R + 1) \ bx0 (R − 1)
Thus
2n vol Bx0 (R + 1) ≥ (R − 1)vol Bp (1)
Since Bp (2(R + 1)) ⊃ Bx0 (R + 1), we have
2n vol Bp (2(R + 1)) ≥ (R − 1)vol Bp (1)
R−1
vol Bp (2(R + 1)) ≥ vol Bp (1).
2n

Wang modified the above argument and proved that
42 PROFESSOR ZHIQIN LU

Lemma 10.8. There is a constant C(n) such that for a ≤ r ≤ R

r
(10.1) Vq (r) ≤ Vq (R)
R
If we choose r = R such taht C < 12 , we have
1
Vq (R) < Vq (R)
2
Further more, we have
C
(10.2) Aa (r) ≤ (Vq (2r) − Vq (r))
r
This (??) is inverse Laplacian comparison theorem!
We pick a cut-off function ψ such that ψ(r) = 1 for 2 < r < 2, ψ(r) = 0 for r > 2, r < ,
1 ≤ ψ ≤ 1, |ψ 0 | + |ψ 00 | < C(δ). From now on, we fix  > 0. Let
√
 
r(x)
ϕk = ψ sin λr
k
Then {ϕk } forms an infinite dimensional vector space.
A straightforward computation gives
C
|∆ϕk + λϕk | ≤ + C|∆r|
k
n−1
We have known ∆r ≤ r
. Thus
n−1 n−1 2(n − 1)
|∆r| ≤ − ∆r + = − ∆r
r r r
Thus we have
C
|∆ϕk + λϕk | ≤ − C∆r
k Z
C
|∆ϕk + λϕk |L1 (M ) ≤ (V (2k) − V (k)) − C ∆r
k B(2k)−B(k)

On the other hand

|ϕk |L1 ≥ C (V (k) − V (2k))
By the volume comparison (??)
C C
(V (2k) − V (k))) ≤ |ϕk |L1
k k
On the other hand
Z Z Z
− ∆r = 1− 1
B(2k)−B(k) ∂B(k) ∂B(2k)
Z
≤ 1 = Aq (k)
∂B(K)

Using (??), we get the desired estimate.

R
Remark 10.9. The upper bound of |∆r| can be estimated, using the same method.
Further readings:Griffiths-Harris: principle of algebraic geometry.
 NOTES IN GEOMETRIC ANALYSIS 43

11. Final Exam

(1) Let φ(t) be a nonnegative continuous function on R. Assume that for any 0 < s < t < 1, we
have
1 C
φ(s) ≤ φ(t) +
2 (t − s)α
for some α > 0, C > 0. Then there is a constant C0 > 0 such that
C1
φ(s) ≤
(1 − s)α
for any 0 < s < 1.

This is a technical inequality in the Moser iteration.

(2) Let gt satisfy the Ricci flow equation:

∂gt
= −2Rict .
∂t
Let ωj (t) for j = 1, · · · , n be an orthonormal frame for gt . Prove that by an orthogonal
change of the frames, we may assume that
∂ωj
= −(Rict )ij ωi .
∂t

This is the Uhlenbeck’s trick in Ricci flow.

(3) Let M be a complete Riemannian manifold. Let u be a positive smooth function on M

satisfying the uniform Harnack inequality. That is, there is a constant C > 0 such that for
any R and any c < inf B(R) u,
sup u − c ≤ C( inf u − u).
B(R) B(R)

Prove that u is a constant.

In general, Harnack inequality implies Hölder continuity, which is weaker than the differ-
entiable Harnack inequality.

(4) Let f be a nonnegative smooth function on a complete noncompact Riemannian manifold.

We assume that f is subharmonic: ∆f ≥ 0 and f is in L2 (M ). Prove that f is a constant.

The above result implies Yau’s theorem on L2 harmonic functions. In the sense of distri-
bution, the positive part and the negative part of a harmonic function must be subharmonic.
44 PROFESSOR ZHIQIN LU

(5) Let M be a complete noncompact Riemann surface and let K be the Gauss curvature. We
assume that Z
|K|dVM < ∞.
M
Prove that M is a parabolic manifold.

By a theorem of Huber, topologically, M must be a compact Riemann surface with several

1
R
points removed. In general, 2π M KdVM is not an integer. If M is embedded into R3 , then
a surprising result of White stated that the second fundamental form is not integrable if
1
R
2π M
KdVM is not an integer.

(6) Let M be a complete noncomact Riemannian manifold. If

σ0 (∆) < σess (∆),
then by definition, the pure point spectrum exists. Prove directly that σ0 (∆) is an L2 eigen-
value. Here we use the variational definition of the quantities σ0 (∆), σess (∆). That is
R
M
|∇f |2
σ0 (∆) = inf R ,
f ∈C0∞ (M )
M
f2
and R
|∇f |2 M
σess (∆) = lim inf 2
,R
∞
K f ∈C0 (M \K)
M
f
where K is running over all exhaustions of the manifold.

In the above case, σ0 (∆) is called the ground state.

(7) Prove the Kato’s inequality: for any f ∈ H 1 (M ), we have

|∇|f || ≤ |∇f |
in the sense of distribution.

This is an inequality related to the Cauchy inequality.

References
1. S-T. Yau Lecture on differential geometry.
2. H. Wu Topics on Riemannian geometry.
3. P. Li Harmonic functions and applications to complete manifolds.
4. Chen Yazhe, Wu Lanchen, Elliptic partial differential equation and equation system of second order
5. Dormelly, Topology 20. 1-14,1981.
6. SturmJ. Funct. Anal. 118, 442-453, 1993.
7. Wang, Math. Research Letters, 4, 473-479, 1997.
8. Griffiths, HarrisPrinciple of algebraic geometry.
Zhiqin Lu
Department of Mathematics, UC. Irvine,
UC. Irvine, CA 92697
E-mail: zlu@uci.edu