Spectral theory v.s.

metric
geometry on Riemannian
manifolds

Chemnitz - 2025
Marcel Schmidt
 Contents

Chapter 1. Abstract 4
Chapter 2. Setup and preliminaries 5
1. Weighted Riemannian manifolds 5
2. Sobolev spaces and a self-adjoint Laplacian 8
3. Local regularity results and a local Harnack inequality 11
Chapter 3. Liouville theorems and essential self-adjointness 13
Chapter 4. The bottom of the spectrum 18
1. Hardy inequalities and the Agmon-Allegretto-Piepenbrink
theorem 18
2. The theorem of Brooks 20
Chapter 5. Probabilistic properties 23
1. Recurrence 24
2. Stochastic completeness 26
Bibliography 33

3
 CHAPTER 1

Abstract

In this series of lectures we give an introduction on how metric ge-

ometry influences the spectral theory of the Laplace-Beltrami operator
on a Riemannian manifold. We discuss self-adjoint realizations, bounds
for the bottom of the spectrum and some spectral theory on L∞ , with
the latter being related to global properties of Brownian motion. The
involved methods are so robust, that they generalize to large classes of
less smooth spaces (e.g. metric measure spaces and discrete graphs).
The spectral theory of the Laplace-Beltrami-Operator (or rather a
specific self-adjoint realization) on a Riemannian manifold is in close
relation with the geometry of the manifold. In this course we discuss
which features of the manifold viewed as a metric measure space influ-
ence spectral geometry.
It is impossible to cover such a vast subject in three lectures, in
particular when it comes to technical details. We aim to explain proof
techniques but keep technical details at a minimum. We refer to the
excellent book [4] and to the survey [3] for more details.

4
 CHAPTER 2

Setup and preliminaries

In this chapter we briefly introduce the setup used throughout this

mini course. We mostly follow [4] in our presentation. The first two
sections are mostly contained in [4, Chapter 4].

1. Weighted Riemannian manifolds

Let (M, g) be a smooth Riemannian manifold which is connected
and without boundary. Charts are denoted by (U, ψ). Here, U ⊆ M
and ψ(U ) ⊆ Rd are open and ψ : U → ψ(U ) is a homeomorphism
belonging to the chosen smooth structure. For f : M → A, where A is
any set, we use the notation f¯ = f ◦ ψ −1 : ψ(U ) → A. In this sense, f¯
is a local representation of f in the coordinates of (U, ψ).
Remark 2.1. For all results discussed in this text less smoothness is
sufficient but we will not discuss this in detail.
The volume measure on the Borel-σ-algebra of M is denoted by vol
and dx denotes the Lebesgue measure on (subsets of) Rd . In a chart
(U, ψ) we have Z q
vol(A) = det(g ij )dx
ψ(A)
for all Borel sets A ⊆ U . Instead of the volume measure itself, we we
consider the weighted measure µ = e−Φ vol with Φ ∈ C ∞ (M ). The
triplet (M, g, µ) is called weighted Riemannian manifold.
For a piecewise smooth curve γ : I → M we define its length by
Z
L(γ) = |γ̇|g dt.
I
The geodesic distance between x, y ∈ M is defined by
ϱ(x, y) = inf L(γ),
γ

where the infimum is taken over all piecewise smooth curves containing
x and y. The function ϱ is a metric on M that induces the topology of
M . The manifold M is called complete if (M, ϱ) is complete.
We write C ∞ (M ) for the smooth function M → R and Cc∞ (M ) =
D(M ) for the smooth functions with compact support. Elements of
D(M ) are called test functions. We say that a sequence (φn ) in D(M )
D
converges to φ ∈ D(M ) and write φn → φ if the following holds:
5
6 2. SETUP AND PRELIMINARIES

(a) There exists K ⊆ M such that supp φn ⊆ K for all n ∈ N.

(b) ∂ α φn → ∂ α φ uniformly in each chart (∂ α is defined with respect to
the chart).
We write X(M ) for the smooth vector fields and denote by D(M ⃗ )
the smooth vector fields of compact support. We equip the latter space
with a similar notion of convergence as in D(M ) by requiring the con-
vergence in (b) for each component of the vector fields in local coordi-
nates.
A sequentially continuous linear functional T : D(M ) → R is called
distribution and a sequentially continuous linear functional U : D(M
⃗ )→
R is called distributional vector field. We identify any f ∈ Lloc (M ) with
1

the distribution
Z
Tf : D(M ) → R, Tf (φ) = f φdµ.
M

Similarly, we identify X ∈ L
⃗ 1 (M ) with the distributional vector field
loc
Z
UX : D(M
⃗ ) → R, UX (ω) = hX, ωig dµ.
M

The space of all distributions is denoted by D′ (M ) and the space of all

⃗ ′ (M ).
distributional vector fields is denoted by D
∞
For f ∈ C (M ) we denote by ∇f ∈ X(M ) its gradient. It is the
unique vector field such that
df (X) = X(f ) = h∇f, Xig for all X ∈ X(M ).
In local coordinates it is given by
∂f ∂
∇f = g ij .
∂xj ∂xi
Here, we use two conventions: The Einstein notation and that (g ij ) is
the inverse of the matrix (gij ).
For each X ∈ X(M ) there exists a unique function divµ X ∈ C ∞ (M )
such that Green’s formula
Z Z
h∇φ, Xig dµ = − φdivµ Xdµ
M M

holds for all φ ∈ Cc∞ (M ).

In local coordinates with X = X i ∂x∂ i we have
1 ∂ p 
divX = p det gij X k
det gij ∂xk
∂ ∂ p
= k
X k + X k k log det gij .
∂x ∂x
 1. WEIGHTED RIEMANNIAN MANIFOLDS 7


1
It is readily verified that divµ X = e−Φ div e−Φ X . Using the product
rule, we infer
1

divµ X = −Φ div e−Φ X = divX − h∇Φ, Xig .
e
The weighted Laplace-Beltrami operator ∆µ is defined by ∆µ =
divµ ◦ ∇. For f ∈ C ∞ (M ) it is given by
∆µ f = ∆vol f − h∇Φ, ∇f ig = ∆vol f − ∇Φ(f ).
In local coordinates it takes the form
 
1 ∂ p kl ∂ ∂Φ ∂f
∆µ f = p k
det gij g l
f − g ij j i .
det gij ∂x ∂x ∂x ∂x
For T ∈ D′ (M ) we let
∇T : D(M
⃗ ) → R, ∇T (ω) = −T (divµ ω)
and for U ∈ D(M
⃗ )′ we let
divµ U : D(M ) → R, divµ U (φ) = −U (∇φ).
By Green’s formula these extensions to distributions are consistent
with the definition on smooth functions. Moreover, as on functions
we let ∆µ = divµ ◦ ∇. Hence, we obtain the Laplacian as an operator
D′ (M ) → D′ (M ).
Below we ALWAYS work with the weighted Laplacian ∆µ but for
convenience we drop the subscript µ.
Example 2.2 (Standard example). Let M = Ω ⊆ Rd open and A : Ω →
Rd×d with the following properties:
(a) For all x ∈ Ω the matrix A(x) is symmetric.
(b) Aij ∈ C ∞ (Ω), i, j = 1, . . . , d.
(c) hA(x)ξ, ξi > 0 for all x ∈ Ω and ξ ∈ Rd .
Using the canonical chart (Ω, id) we identify Tx Ω = Rd . Then
gx (ξ, η) = hA(x)ξ, ηi, x ∈ Ω, ξ, η ∈ Rd
defines a smooth Riemannian metric on Ω. In this case, the volume is
given by Z √
vol(B) = det Adx
B
for all Borel B ⊆ Ω.
The weighted Laplacian takes the form
1 √ 
∆µ f = √ dive det AA−1 ∇e f − hA−1 ∇e Φ, ∇e f i
det A


= dive A−1 ∇e f − hX, ∇e f i.
Here, dive and ∇e denote the standard divergence and gradient on
Euclidean space and X ∈ C ∞ (Ω, Rn ) is a suitable vector space. This
8 2. SETUP AND PRELIMINARIES

shows that ∆µ is an ’elliptic operator in divergence form with drift

term’. √
If we choose Φ = log det A, then dvol = dx and


∆µ f = dive A−1 ∇e f .
If one wants to avoid inverting A, one can modify the examples and
start with A−1 instead of A.
Example 2.3 (Model manifolds). We call a d-dimensional Riemannian
manifold (M, g) (with d ≥ 2) a model manifold if the following holds:
(a) For some R ∈ (0, ∞] we have M = {x ∈ Rd | |x| < R}.
(b) There exists ψ : (0, R) → (0, ∞) such that the map
M \ {0} → (0, R) × Sd−1 , x 7→ (|x|, |x|−1 x)
is a Riemannian isometry, where on (0, R) × Sd−1 the tangent space
Tr,θ ((0, R) × Sd−1 ) = R ⊕ Tθ Sd−1
is equipped with the metric gR ⊕ ψ(r)2 gSd−1 ,θ .
The function ψ is called scaling function of the model.
In this case, we have
Z r
vol(Br (0)) = ωd−1 ψ(t)d−1 dt,
0
where ωd−1 is the (d − 1)-dimensional volume of Sd−1 .
If f ∈ C ∞ (M ) with f (x) = h(|x|) for x 6= 0 and some h ∈
C ∞ ((0, ∞)), then
ψ ′ (r) ′
∆vol f (x) = h′′ (r) + (d − 1) h (r),
ψ(r)
where r = |x| and x 6= 0.
Example 2.4. (a) Rd is a model manifold with R = ∞ and scaling
function ψ(r) = r.
(b) The sphere without one point Sd \ {p} is isometric to a model
manifold with R = π and scaling function ψ(r) = sin r.
(c) Hyperbolic space Hd is isometric to a model manifold with R = ∞
and scaling function ψ(r) = sinh r.
2. Sobolev spaces and a self-adjoint Laplacian
In this section we introduce several function spaces on the weighted
manifold. Some depend on the choice of the density function Φ in the
measure µ and some do not. This is reflected in our notation.
The first order Sobolev space on (M, g, µ) is defined by
W 1 (M, µ) = {f ∈ L2 (M, µ) | ∇f ∈ L ⃗ 2 (M, µ)}.
Equipped with the norm

1/2
k·kW 1 : W 1 (M, µ) → [0, ∞), kf kW 1 = kf k22 + k∇f k22 ,
 2. SOBOLEV SPACES AND A SELF-ADJOINT LAPLACIAN 9

which is clearly induced by an inner product h·, ·iW 1 , it is a Hilbert

space. By a celebrated result of Meyers and Serrin [6] C ∞ (M ) ∩
W 1 (M, µ) is dense in W 1 (M, µ). By W01 (M, µ) we denote the clo-
sure of Cc∞ (M ) in W 1 (M, µ). We write Wc1 (M ) for the functions in
W 1 (M, µ) with compact support. It can be inferred as a consequence
to the Meyers Serrin theorem that Wc1 (M ) ⊆ W01 (M, µ).
We let
W 1 (M ) = {f ∈ L2 (M ) | ∇f ∈ L2 ⃗(M )}.
loc loc loc
Then f ∈ 1
Wloc (M ) if and only if for all relatively compact open Ω ⊆ M
there exists g ∈ W 1 (M, µ) with f = g a.s. on Ω. It follows from
the discussed properties that Wc1 (M ) and Wloc 1
(M ) are independent of
1
the choice of µ and every function in Wloc (M ) with compact support
belongs to Wc1 (M ).
The following result is the key tool in our analysis.
Theorem 2.5 (Rademacher). A function f : M → R is Lipschitz (with
respect to ϱ) if and only if f ∈ L1loc (M ) and ∇f ∈ L∞ (M ). In this
case, f ∈ Wloc
1
(M ) and
|f (x) − f (y)|
k∇f k∞ = sup =: Lip(f ).
x̸=y ϱ(x, y)
In particular, Lipschitz functions with compact support belong to Wc1 (M ).
Proof. The if statement and the inequality k∇f k∞ ≤ Lip(f ) is
contained in [4, Theorem 11.3]. For the reverse inequality and the only
if statement see [2, Theorem 4.5] and its proof. □
Corollary 2.6. For all x, y ∈ M we have
ϱ(x, y) = sup{|f (x) − f (y)| | f ∈ L1loc (M ) with k∇f k∞ ≤ 1}.
Remark 2.7. Rademacher’s theorem and its corollary have two conse-
quences, which make extensions of the presented results to non-smooth
spaces possible.
(a) We define the local Lipschitz constant of f : M → R at x ∈ M by
|f (x) − f (y)|
Lipx (f ) := inf sup ,
U x̸=y,x,y∈U ϱ(x, y)
where the infimum is taken over all open neighborhoods of x. From
localizing the equality in Rademacher’s theorem, we infer
|∇f |(x) = Lipx (f )
for f ∈ C 1 (M )-functions. Hence, local Lipschitz constants deter-
mine |∇f |. Since local Lipschitz constants are available in metric
spaces (without differentiable structure and Riemannian metric),
this paves the way towards an analysis on metric measure spaces
by considering Cheeger energies and the induced Laplacians, see
e.g. [1].
10 2. SETUP AND PRELIMINARIES

(b) The previous corollary shows that ϱ can be recovered from the
knowledge of all functions with k∇f k∞ ≤ 1. Hence, if in some
space one can give a meaning to the inequality k∇f k∞ ≤ 1 (e.g.
this is possible in regular Dirichlet spaces), one can define a metric
d through the formula
d(x, y) = sup{|f (x) − f (y)| | f ∈ L1loc (M ) with k∇f k∞ ≤ 1}.
Under relatively mild assumptions this metric satisfies Rademacher’s
theorem (or at least part of it), namely k∇f k∞ ≤ Lip(f ). We refer
to [8, 9, 7], where this Ansatz is used in the analysis of strongly
local Dirichlet spaces, and to [5], which discusses discrete spaces.
Corollary 2.8 (Existence of cut-off functions). For 0 < r < R and
o ∈ M consider the function
φ = φr,R : M → R, φ(x) = (1 − ϱ(x, Br (o))/(R − r))+ .
Then φ = 1 on Br (o), φ = 0 on M \ BR (o) and Lip(φ) ≤ 1/(R − r).
In particular, f ∈ Wloc
1
(M ) and k∇f k∞ ≤ 1/(R − r). If, moreover,
(M, ϱ) is complete, then φ ∈ Wc1 (M ).
Proof. All assertions are more or less trivial or consequences of
Rademacher’s theorem except ’moreover’-statement. It follows from
the Hopf-Rinow theorem, which states that (M, ϱ) is complete if and
only if Br (x) is compact for all r > 0 and x ∈ M . □
Remark 2.9. Instead of requiring completeness of M , for many ap-
plications the existence of a sequence of cut-off functions (φn ) with
0 ≤ φn ≤ 1, φn → 1 pointwise and k∇φn k∞ → 0 is sufficient. How-
ever, it turns out that the existence of such a sequence is equivalent to
completeness.
1
The following proposition summarizes several operations on Wloc (M )
that will be used below.
1
Proposition 2.10 (Product- and chain rules on Wloc (M )). (a) (Chain
′
rule) Let (Cn ) in C (R) such that supn kCn k∞ < ∞ and let C, D : R →
1

R such that Cn → C and Cn′ → D pointwise. Then for any

f ∈ Wloc
1
(M ) we have C ◦ f ∈ Wloc 1
(M ) and
∇(C ◦ f ) = (D ◦ f )∇f.
(b) (Truncation property) For all f, h ∈ Wloc
1
(M ) we have f ∧ h ∈
1
Wloc (M ) and
∇(f ∧ h) = 1{f ≤h} ∇f + 1{f >h} ∇h.
(c) (Product rule) Let f, h ∈ Wloc
1
(M ) such that f h ∈ Wloc
1
(M ). Then
∇(f h) = f ∇h + h∇f.
(d) (Ideal property) For all φ ∈ Lipc (M ) and f ∈ Wloc
1
(M ) we have
φf ∈ Wc (M, µ).
1
 3. LOCAL REGULARITY RESULTS AND A LOCAL HARNACK INEQUALITY11

Proof. All of the results except (b) are clear for smooth functions,
other functions have to be approximated (e.g. by using the Meyers-
Serrin theorem). To prove (b) one has to use (a) , f ∧ g = (f + g − |f −
g|)/2 and an approximation of | · | by smooth functions. The details
can be found somewhat scattered in [4]. □
We finish this section by constructing a non-negative self-adjoint
Laplacian on L2 (M, µ).
Theorem 2.11 (The Friedrichs extension). The operator ∆F defined
by
D(∆F ) = {f ∈ W01 (M, µ) | ∆µ f ∈ L2 (M, µ)}
and
∆F = ∆µ f
is a non-positive self-adjoint operator on L2 (M, µ) (i.e. σ(∆F ) ⊆
(−∞, 0]). It is the Friedrichs extension of the restriction of ∆µ to
Cc∞ (M ). In particular,
inf(σ(−∆F )) = inf{h−∆F φ, φi | φ ∈ Cc∞ (M ), kφk2 = 1}
Z
= inf{ |∇f |2g dµ | f ∈ W01 (M, µ), kφk2 = 1}.
M
For all α > 0 its resolvent (α − ∆F )−1 is positivity improving (f ⪈
0 implies (α − ∆F )−1 f > 0 a.s.) and Markovian (f ≤ 1 implies
α(α − ∆F )−1 f ≤ 1).
Proof. It follows from Green’s formula that the restriction of ∆µ
to Cc∞ (M ) is non-positive and symmetric. Hence, it has a Friedrichs
extension. The formula for ∆F then follows directly from the abstract
construction of the Friedrichs extension and the obsercation that the
adjoint of the restriction of ∆µ to Cc∞ (M ) is the restriction of ∆µ to
{f ∈ L2 (M, µ) | ∆µ f ∈ L2 (M, µ).
For the other properties we refer to [4, Chapter 5]. □
3. Local regularity results and a local Harnack inequality
Theorem 2.12 (Hypoellipticity). (a) Let α ∈ R and 1 < p < ∞.
Every u ∈ Lploc (M ) with −∆u+αu ∈ C ∞ (M ) satisfies u ∈ C ∞ (M ).
(b) If u ∈ D′ ((0, ∞) × M ) satisfies ∂t u = ∆u, then u ∈ C ∞ (M ).
Proof. See [4, Chapter 7]. □
Theorem 2.13 (Local Harnack inequality). Let λ ∈ R and let K ⊆ M
compact. There exists a constant C = C(K, λ) ≥ 0 such that for all
nonnegative u ∈ C ∞ (M ) with −∆u − λu = 0 the following inequality
holds:
sup u ≤ C inf u.
K K
The constants can be chosen such that R → (0, ∞), λ 7→ C(K, λ) is
increasing.
12 2. SETUP AND PRELIMINARIES

Proof. This is contained in [4, Theorem 13.11] under the assump-

tion that there exists λ′ ≥ λ and h > 0, h ∈ C ∞ (M ), such that
−∆h − λ′ h = 0. However, using the following ingredients this assump-
tion can be completely removed:
(a) The Agmon-Allegretto-Piepenbrink theorem on relatively compact
open subsets of M , which does not rely on Harnack inequalities but
on Sobolev embedding theorems.
(b) A domain monotonicity argument.
(c) The implication of the Agmon-Allegretto-Piepenbrink theorem, which
does not use Harnack inequalities.
Below in our proof of the Agmon-Allegretto-Piepenbrink theorem
we cheat a bit, because we do not prove (a) and (b) separately, but use
the Harnack inequality directly. □
 CHAPTER 3

Liouville theorems and essential self-adjointness

Theorem 3.1 (Karp’s Liouville theorem). Assume that M is complete

and let 1 < p < ∞. Let f ∈ Wloc 1
(M ) ∩ L∞loc (M ) satisfy f ≥ 0 and
∆f ≥ 0 (in the sense of distributions). If for some (all) o ∈ M and
some (all) r0 > 0 Z ∞
r
p dr = ∞,
r0 kf 1Br (o) kp
then f is constant.
Remark 3.2. Using local regularity theory for subharmonic functions
it can be proven that any f ∈ Wloc 1
(M ) with f ≥ 0 and ∆f ≥ 0
∞
automatically belongs to Lloc (M ). Such regularity is complicated and
beyond the scope of this lecture.
Before establishing Karp’s theorem we need one further result, which
is interesting on its own right.
Theorem 3.3 (Caccioppoli inequalities). Let 1 < p < ∞. Let f ∈
1
Wloc (M ) ∩ L∞
loc (M ) with f ≥ 0 and ∆f ≥ 0. For any φ ∈ Lipc (M ) and
n ∈ N we have
Z Z
2
f φ |∇f |g dµ ≤ −
p−2 2 2
f p−1 φh∇f, ∇φig dµ
M p−1 M

and Z Z
4
f p−2 2
φ |∇f |2g dµ ≤ f p |∇φ|2g dµ.
M (p − 1)2 M

Proof. Let n ∈ N. Instead of f we consider fn = f ∨ n−1 . Then

fn ∈ Wloc
1
(M ) ∩ L∞loc (M ), fn ≥ n
−1
and C : (0, ∞) → R, x 7→ xp−1
being locally Lipschitz imply fnp−1 ∈ Wloc
1
(M ).
For φ ∈ Lipc (M ) the product rule yields φ2 fnp−1 ∈ Wc1 (M ) ∩
L∞ (M ). The inequality ∆f ≥ 0 in the sense of distributions yields
Z
0≥ h∇f, ∇(φ2 fnp−1 )ig dµ,
M

after we approximate ∈ Wc1 (M ) by nonnegative functions in

φ2 fnp−1
Cc∞ (M )1. The chain rule, the product rule and the truncation property
1This is not entirely trivial.
13
14 3. LIOUVILLE THEOREMS AND ESSENTIAL SELF-ADJOINTNESS

yield
Z
h∇f, ∇(φ2 fnp−1 )ig dµ
M
Z Z
= (p − 1) φ fn h∇f, ∇fn ig dµ + 2
2 p−2
fnp−1 φh∇f, ∇φig dµ
ZM ZM
= (p − 1) φ2 f p−2 |∇f |2g dµ + 2 fnp−1 φh∇f, ∇φig dµ
{f ≥1/n} M

Combining both inequalities and letting n → ∞ yields

Z Z
2
φ f |∇f |g dµ ≤ −
2 p−2 2
f p−1 φh∇f, ∇φig dµ.
M p − 1 M

For the existence of both of the integrals (which is important for the
convergence as n → ∞) we used f ∈ L∞ loc (M ).
Next we use the elementary inequality |ab| ≤ εa2 + 1/(4ε)b2 to
estimate
Z Z Z
1
|f φh∇f, ∇φig |dµ ≤ ε
p−1
φ f |∇f |g dµ +
2 p−2 2
f p |∇φ|2g dµ.
M M 4ε M

Combining the two previous inequalities yields

Z Z
2ε 2
(1 − ) φ f |∇f |g dµ ≤
2 p−2 2
f p |∇φ|2g dµ.
p−1 M 4ε(p − 1) M
We obtain the statement by letting ε = (p − 1)/4. □
Remark 3.4. (a) The Caccioppoli inequality does not use the com-
pleteness of the manifold.
(b) Instead of assuming f ∈ L∞ loc (M ), it would be sufficient to require
Z
f p−1 |φ||h∇f, ∇φig |dµ < ∞,
M

which is used in one step of our proof. By the inequality 2|ab| ≤

a2 + b2 , we have
Z
f p−1 |φ||h∇f, ∇φig |dµ
M
Z Z
≤ f 2p−2
|∇φ| dµ +
2
|φ|2 |∇f |2g dµ.
M M

Hence, our proof works under the assumption f ∈ Lloc 2p−2

(M ). It is
automatically satisfied for f ∈ Wloc (M ), as long as 1 ≤ p ≤ 2.
1

To simplify notation we let Br = Br (o) and denote by

Ar,R = BR \ Br = {x ∈ M | r < ϱ(x, o) ≤ R}
the annulus with the radii 0 < r < R around o.
 3. LIOUVILLE THEOREMS AND ESSENTIAL SELF-ADJOINTNESS 15

Corollary 3.5 (Key estimate). Assume M is complete, let 1 < p < ∞

and 0 < r < R.Then for any f ∈ Wloc 1
(M ) ∩ L∞
loc (M ) with f ≥ 0 and
∆f ≥ 0 the following inequalities hold:
Z 2 Z
4 1
f φ |∇f |g dµ ≤
p−2 2 2
kf 1Ar,R kp
p
f p−2 φ2 |∇f |2g dµ
M (p − 1) (R − r)
2 2
Ar,R

and Z
4 1
f p−2 |∇f |2g dµ ≤ kf 1Ar,R kpp .
Br (p − 1) (R − r)
2 2

Here, φ = φr,R is the cut-off function of Corollary 2.8.

Proof. Let ε > 0 such that r +ε < R and set r̃ = r +ε. Since M is
complete, the cut-off function φr,R discussed in Corollary 2.8 belongs to
Lipc (M ). It is constant on the open sets Ur̃ and M \ BR and therefore
satisfies ∇φr̃,R = 0 on Ur̃ ∪ (M \ BR ). Moreover, by Rademacher’s
theorem |∇φr̃,R | ≤ Lip(φr̃,R ) ≤ 1/(R − r̃). Using these estimates,
0 ≤ φr̃,R ≤ 1 and φr̃,R = 1 on Br̃ , the Caccioppoli inequalities imply
the desired result (after another application of Cauchy-Schwarz for the
first inequality) with Ar,R replaced by BR \ Ur̃ . With this at hand, the
statement follows after letting ε & 0. □
Karp’s theorem. We choose R > 0 such that f 1BR 6= 0 and let
Rn = 2n R. We consider the quantities
Z
Qn = Qn (R) = f p−2 |∇f |2g dµ
BRn

and
Vn = kf 1ARn−1 ,Rn kpp .
We show Q1 = Q1 (R) = 0 for all R > 0, which implies
Z
f p−2 |∇f |2g dµ = 0,
M

that is f p−2
|∇f |2g
= 0 a.s. Since f ≥ 0, we have f = f+ and the
truncation property yields
∇f = ∇f+ = 1{f >0} ∇f.
Hence, f p−2 |∇f |2g = 0 a.s. implies ∇f = 0 a.s. and we infer that f is
constant from Rademacher’s theorem.
The assumption
Z ∞
r
p dr = ∞
r0 kf 1Br (o) kp

implies
X
∞
R2 n
= ∞.
n=1
Vn
16 3. LIOUVILLE THEOREMS AND ESSENTIAL SELF-ADJOINTNESS

We let C = 4/(p − 1)2 and φn = φRn−1 ,Rn (the cut-off function from
Corollary 2.8). Our key estimate shows
Z !2
Qn−1 + φ2n f p−2 |∇f |2g dµ
ARn−1 ,Rn
Z 2
= φ2n f p−2 |∇f |2g dµ
M
Z
CVn
≤ φ2n f p−2 |∇f |2g dµ
(Rn − Rn−1 ) ARn−1 ,Rn
2
Z
4CVn
= 2
φ2n f p−2 |∇f |2g dµ.
Rn ARn−1 ,Rn
Now assumeR that Q1 > 0 and hence Qn > 0 for all n ∈ N. If we
write Kn = AR ,R φ2n f p−2 |∇f |2g dµ, the previous inequality and Kn ≤
n−1 n
Qn − Qn−1 (which follows from 0 ≤ φn ≤ 1) yield
4CVn (Qn−1 + Kn )2 (Qn−1 + Kn )
2
≥ ≥ Qn−1
Rn Kn Kn
 
Qn−1 Qn−1 Qn
≥ Qn−1 +1 = .
Qn − Qn−1 Qn − Qn−1
Rearranging this inequality leads to
 
Rn2 1 1
≤ 4C −
Vn Qn−1 Qn
and summing it up shows
X∞
R2 4C
n
≤ < ∞,
n=2
Vn Q1
a contradiction. □
Corollary 3.6 (Yau’s Liouville theorem). Assume M that is complete
and 1 < p < ∞. Any f ∈ Lp (M, µ) with −∆f + αf = 0 for some
α ≥ 0 is constant. If α > 0, then f = 0.
Proof. By local regularity f ∈ C ∞ (M ) ⊆ Wloc1
(M ) ∩ L∞
loc (M ). We
choose a sequence of increasing convex functions (Cn ) in C 2 (R) with
kCn′ k∞ ≤ 1, Cn = 0 on (−∞, 0] and Cn (x) → x1[0,∞) (x) for all x ∈ R.
For example, the functions Cn defined by


0 if x ≤ 0
Cn (x) = x if x ≥ 1/n

x3 (3n4 x2 − 8n3 x + 6n2 ) if 0 < x < 1/n

do the job. The Chain rule for the Laplacian implies

∆Cn (f ) = Cn′′ (f )|∇f |2g + Cn′ (f )∆f = Cn′′ (f )|∇f |2g + αCn′ (f )f.
 3. LIOUVILLE THEOREMS AND ESSENTIAL SELF-ADJOINTNESS 17

We infer (∆Cn (f ), φ) ≥ 0 for all nonnegative φ ∈ D(M ) and, taking the

limit n → ∞, arrive at (∆f+ , φ) ≥ 0 for all nonnegative φ ∈ D(M ).
With this at hand, f+ being constant follows from Karp’s theorem.
Applying the result also to −f instead of f gives that f is constant.
Clearly, if α > 0, the only constant function f satisfying ∆f = αf
is f = 0. □
Corollary 3.7 (Gaffney’s theorem). Assume that M is complete. Then
the restriction of ∆ to Cc∞ (M ) is essentially self-adjoint on L2 (M, µ).
Its unique self-adjoint extension is ∆F and
D(∆F ) = {f ∈ L2 (M, µ) | ∆f ∈ L2 (M, µ)}
Moreover, W01 (M, µ) = W 1 (M, µ), i.e. Cc∞ (M ) is dense in H 1 (M, µ).
Proof. Let ∆c denote the restriction of ∆ to Cc∞ (M ). By defini-
tion we have to show that (∆c )∗ is self-adjoint.
It follows directly from the distributional definition of ∆ that
D((∆c )∗ ) = {f ∈ L2 (M, µ) | ∆f ∈ L2 (M, µ)}
and that (∆c )∗ f = ∆f .
Clearly, ∆c ⊆ ∆F , and by closedness of ∆F , also ∆c ⊆ ∆F . We
show (∆c )∗ ⊆ ∆F , which then implies
∆F = (∆F )∗ ⊆ (∆c )∗∗ = ∆c ⊆ ∆F .
Let f ∈ D((∆c )∗ ) and for α > 0 let g = (α − ∆F )−1 (α − ∆)f . Then
g ∈ D(∆F ) ⊆ W01 (M, µ) and
(α − ∆)g = (α − ∆)(α − ∆F )−1 (α − ∆)f = (α − ∆)f.
Hence, (α − ∆)(g − f ) = 0 and Yau’s Liouville theorem implies f =
g ∈ D(∆F ).
Recall that we equipped W 1 (M, µ) with the complete inner product
Z Z
hf, hiW 1 = h∇f, ∇hig dµ + f hdµ.
M M
It suffices to show Cc∞ (M )⊥ = {0}. Assume that hh, φiW 1 = 0 for all
φ ∈ Cc∞ (M ). Then
(−∆h + h, φ) = hh, φiW 1 = 0
for all φ ∈Cc∞ (M ),
i.e., −∆h + h = 0 in the sense of distributions. We
infer h = 0 from Yau’s Liouville theorem. □
 CHAPTER 4

The bottom of the spectrum

We study the bottom of the spectrum of the self-adjoint operator

−∆F on L2 (M, µ). We let λ0 (M ) = inf σ(−∆F ).
1. Hardy inequalities and the
Agmon-Allegretto-Piepenbrink theorem
We start with a description in terms of positive solutions and ground
states.
Theorem 4.1 (Agmon-Allegretto-Piepenbrink). Let M be complete.
For λ ∈ R the following assertions are equivalent.
(i) λ ≤ λ0 (M ).
(ii) There exists f ∈ C ∞ (M ) such that f > 0 on M and −∆f ≥ λf .
(iii) There exists a nonnegative f ∈ Wloc
1
(M ) with 1/f ∈ L∞
loc (M ) and
−∆f ≥ λf .
If M is not compact, the function in (ii) can be chosen to satisfy
−∆f = λf
The proof will require a weighted Hardy inequality, which is inter-
esting on its own right.
Theorem 4.2 (A weighted Hardy inequality). Let h ∈ Wloc 1
(M ) non-
∞
negative with 1/h ∈ Lloc (M ) such that −∆h ≥ w for some w ∈ L1loc (M ).
Then Z Z
w
|∇φ|g dµ ≥
2
|φ|2 dµ
M M h
∞
for all φ ∈ Cc (M ). If M is complete, this inequality extends to all
φ ∈ H 1 (M ).
Proof. Let ψ ∈ Wloc
1
(M ) such that ψ 2 , hψ, hψ 2 ∈ Wc1 (M ). The
product rule yields
h∇(hψ), ∇(hψ)ig − h∇(hψ 2 ), ∇hig = h2 |∇ψ|2 .
Clearly, this implies
Z Z
|∇(hψ)|g dµ ≥
2
h∇(hψ 2 ), ∇hig dµ.
M M
Approximating hψ by nonnegative test functions and using −∆h ≥ w
2

leads to Z Z
h∇(hψ ), ∇hig dµ ≥
2
ψ 2 hwdµ.
M M
18
1. HARDY INEQUALITIES AND THE AGMON-ALLEGRETTO-PIEPENBRINK THEOREM
19

To finish the proof it suffices to show that for φ ∈ Cc∞ (M ) the function
ψ = φ/h satisfies the assumptions leading to the previous inequality.
Since 1/h ∈ L∞ loc (M ), the local Lipschitz continuity of (0, ∞) →
(0, ∞), x 7→ 1/x and (0, ∞) → (0, ∞), x 7→ 1/x2 yields 1/h, 1/h2 ∈
1
Wloc (M ). For φ ∈ Cc∞ (M ) we infer φ/h, φ2 /h2 , φ/h2 ∈ Wc1 (M ). □

Remark 4.3. (a) The proof yields the more precise estimate

Z Z Z
2w φ
|∇φ|2g dµ ≥ |φ| dµ + h2 |∇ |2g dµ, φ ∈ Cc∞ (M ).
M M h M h

If −∆h = w, it is an equality.
(b) Using fine properties of functions 0 6= h ∈ Wloc
1
(M ) with h ≥ 0 and
∞
−∆h ≥ 0, the assumption 1/h ∈ Lloc (M ) can be dropped if w ≥ 0.

Proof. (i) ⇒ (ii): If λ < λ0 (M ), we can choose f = (−∆F −λ)−1 φ

with φ ∈ C ∞ (M ) ∩ L2 (M ) and φ > 0. Then (−∆ − λ)f = φ ∈ C ∞ (M )
implies f ∈ C ∞ (M ). Since the resolvent is positivity improving, we
obtain f > 0.
Case 1: M is compact. In this case, 1 ∈ L2 (M ) and −∆1 =
0 ∈ L2 (M ). Since compact manifolds are complete, this implies 1 ∈
D(−∆F ) and hence 0 ∈ σ(−∆F ), showing λ0 (M ) = 0.
Case 2: M is not compact. We choose an increasing sequence
∅ 6= Ωn ⊆ M of relatively compact open connected sets such that
Ωn ⊆ Ωn+1 and an increasing sequence λn < λ with λn % λ. We
further choose nonnegative φn ∈ Cc∞ (M ) with supp φn ⊆ M \ Ωn and
let gn = (−∆F − λn )−1 φn . As a consequence to the hypoellipticity we
have gn ∈ C ∞ (M ).
The condition on the support of φn implies (−∆F − λn )gn = 0 on
Ωn and since the resolvent is positivity improving and gn is smooth, we
have gn (x) > 0 all x ∈ M .
By the local Harnack inequality applied to the compact set Ωn in
the manifold Ωn+1 , there exist constants Cn > 0 such that

sup gk ≤ Cn inf gk , k ≥ n + 1.
Ωn Ωn

We let Kn = infΩ1 gn and fn = (1/Kn )gn . From the local Harnack

inequality we infer 0 ≤ fk ≤ Cn on Ωn for all k ≥ n + 1. Hence, (fk )
is bounded in L2 (Ωn , µ). Using weak compactness of balls in L2 -spaces
and an exhaustion argument we infer that there exists a subsequence
(fnk ) and f ∈ L2loc (M ) such that fnk → f weakly in L2loc (M ). For
20 4. THE BOTTOM OF THE SPECTRUM

φ ∈ Cc∞ (M ) this implies

((−∆ − λ)f, φ) = (f, (−∆ − λ)φ)
= lim (fnk , (−∆ − λnk )φ)
k→∞
= lim h(−∆ − λnk )fnk , φi
k→∞
1
= lim hφnk , φi
k→∞ Knk

= 0,
where for the last equality we use supp φ ⊆ Ωn for some large n and
supp φk ⊆ M \ Ωn for all k ≥ n. Since infΩ1 fn = 1, we also infer f 6= 0.
The strict positivity of f follows from the local Harnack inequality.
(ii) ⇒ (iii): This is obvious.
(iii) ⇒ (i): Using −∆f ≥ λf , the weighted Hardy inequality yields
Z Z
|∇φ|g dµ ≥ λ
2
|φ|2 dµ
M M

for all φ ∈ Cc∞ (M ). This implies λ0 (M ) ≥ λ. □

2. The theorem of Brooks

In this section we seek for upper and lower bounds of the spectrum
that involve Lipschitz functions.
α2
Theorem 4.4. Let f be 1-Lipschitz with ∆f ≥ α. Then λ0 (M ) ≥ 4
.
Proof. For φ ∈ Cc∞ (M ) we obtain
Z
(∆f, φ ) ≥ α
2
φ2 dµ.
M

Using k∇f k∞ ≤ Lip(f ) ≤ 1, we infer f ∈ Wloc

1
(M ) and
Z Z
(∆f, φ2 ) = − h∇f, ∇φ2 ig dµ = −2 h∇f, ∇φig φdµ
M M
Z 1/2 Z 1/2
≤2 |∇φ|2g dµ φ dµ2
.
M M

This implies the claim. □

Theorem 4.5. Assume that M is complete. If there exists a 1-Lipschitz
function f : M → R and β > 0 such that e−βf ∈ L1 (M, µ), then
2
λ0 (M ) ≤ β4 .

Proof. We consider the function h = e− 2 βf . It belongs to L2 (M, µ)

1

and satisfies
1
∇h = βh∇f.
2
 2. THE THEOREM OF BROOKS 21

Using k∇f k∞ ≤ 1, we infer

Z Z Z
β2 β2
|∇h| dµ =
2
h |∇f |g dµ ≤
2 2
h2 dµ.
M 4 M 4 M
This shows h ∈ W 1 (M, µ). Using the completeness of M , we infer
h ∈ W01 (M, µ) and arrive at λ0 (M ) ≤ β 2 /4. □
Remark 4.6. Instead of using completeness to deduce h ∈ W01 (M, µ),
one can also require limx→∞ f (x) = ∞.
Corollary 4.7 (Brooks). Assume that M is complete and let
log µ(Br (o))
β = lim sup .
r→∞ r
β2
Then λ0 (M ) ≤ 4 . In particular, if M has polynomial volume growth,
then λ0 (M ) = 0.
Proof. The condition implies that for given ε > 0 and R = R(ε) >
0 large enough we have
µ(Br (o)) ≤ e(β+ε)r , r ≥ R.
For α > β + ε consider the function f = e−αϱ(o,·) . By completeness of
M balls are relatively compact and we obtain that f is integrable over
any ball. For N > R we infer
Z ∞ Z
X
−αϱ(o,x)
e dµ(x) = e−αϱ(o,x) dµ(x)
M \BN (o) k=N Bk+1 (o)\Bk (o)

X∞
≤ e−αk µ(Bk+1 (o))
k=N
X∞
≤ e−αk e(k+1)(β+ε)
k=N
X
∞
=e β+ε
e(β+ε−α)k < ∞.
k=N

With this at hand the statement follows from the previous theorem. □
Remark 4.8. If µ(M ) = ∞, the stronger inequality
β2
inf σess (−∆F ) ≤
4
holds.
Example 4.9 (Hyperbolic space). We consider the hyperbolic space
Hd , d ≥ 2, and µ = vol. We claim λ0 (Hd ) = (d − 1)2 /4.
22 4. THE BOTTOM OF THE SPECTRUM

As discussed above, Hd is isometrically isomorphic to a d-dimensional

model manifold with scaling ψ(r) = sinh r. The formula for the volume
of balls around 0 in model manifolds implies
Z r
vol(Br (o)) = ψ d−1 (t)dt ≤ Ce(d−1)r ,
0
for some constant C > 0 and all r large enough. This shows
log vol(Br (o))
lim sup ≤d−1
r→∞ r
and we infer λ0 (Hd ) ≤ (d − 1)2 /4 from Brook’s theorem.
We now identify Hd with the model manifold (Rd , gψ ) with scaling
ψ = sinh and the appropriate metric gψ .
We consider the function f : Rd → R, f (x) = |x|, i.e., f (x) =
h(|x|) with h(r) = r, r > 0. Our formula for ∆ on radially symmetric
functions implies
ψ ′ (r) ′
∆f (x) = h′′ (r) + (d − 1) h (r) = (d − 1) coth(r) ≥ (d − 1)
ψ(r)
for all x 6= 0 and r = |x|. Since in a model manifold |x| = ρ(x, 0),
the function f is also 1-Lipschitz. Our lower estimate on λ0 yields
λ0 (Hd \ {o}) = λ0 ((Rd \ {0}, gψ )) ≥ (d − 1)2 /4, where o is the point
that gets mapped to 0 under the isometry between Hd and (Rd , gψ ).
Now there are two possibilities to show that the same estimate also
holds for λ0 (Hd ).
1. In a Riemannian manifold (M, g) of dimension at least 2 we have
λ0 (M ) = λ0 (M \ {o}) for any o ∈ M . This follows from the fact that
{φ ∈ Cc∞ (M ) | supp φ ⊆ M \ {o}} is dense in W01 (M ) in dimension
d ≥ 2, which can relatively easily be verified using local charts.
2. For any o ∈ Hd one can use polar coordinates starting in o to
obtain an isometry Φo : Hd → (Rd , gψ ) with Φo (o) = 0. Hence, our
discussion shows λ0 (Hd \ {o}) ≥ (d − 1)2 /4 for all o ∈ Hd .
Let φ ∈ Cc∞ (M ) and choose o ∈ Hd \ supp φ. Then its restriction
satisfies
φ|Hd \{o} ∈ Cc∞ (Hd \ {o})
and we obtain
Z Z
(d − 1)2
|∇φ|g dvol =
2
|∇φ|2g dvol ≥ λ0 (Hd \{o})kφk22 ≥ kφk22 .
Hd Hd \{o} 4
This shows λ0 (Hd ) ≥ (d − 1)2 /4.
 CHAPTER 5

Probabilistic properties

The Laplace operator −∆F generates an operator semigroup (Pt )

through the formula
 −n
t∆F t
Pt = e = lim 1 + ∆F .
n→∞ n
The semigroup and the resolvent are Markovian (i.e. 0 ≤ f ≤ 1 implies
0 ≤ Pt f ≤ 1 and 0 ≤ (1 + α∆F )−1 f ≤ 1). For measurable f ≥ 0 we
choose fn ≥ 0 in L2 (M, µ) with fn % f a.s. Then
Pt f := lim Pt fn
n→∞
and
(1 + α∆F )−1 f := lim (1 + α∆F )−1 fn
n→∞
exist (as an a.e. defined [0, ∞]-valued function) and are independent
of the chosen sequence (fn ).
If f ∈ L∞ (M )+ , then Pt f ∈ L∞ (M )+ by the Markov property. We
extend Pt to a linear operator on L∞ (M ) by letting Pt f = Pt f+ −
Pt f− . It can be proven that for f ∈ L2 (M, µ) + L∞ (M, µ) the function
u : (0, ∞) × M → R, u(t, ·) = Pt f satisfies the heat equation
∂t u = ∆u
in the sense of distributions. By the hypoellipticity of the heat operator,
this implies u ∈ C ∞ ((0, ∞) × M ) and u solves the heat equation in the
classical sense. Moreover, it satisfies the initial condition u(0, ·) = f in
the sense of L2loc (M ), i.e.,
lim u(t, ·) = f
t→0+

in L2loc (M ).
There is a deep connection between (minimal) Brownian motion
and the operator −∆F . More precisely, they are related through the
Feynman-Kac formula
Pt f (x) = Ex [1{t<τ } f (Bt )],
which holds for all f ∈ L∞ (M, µ) and x ∈ M . Here, Ex denotes the
expectation with respect to P(· | B0 = x) and τ denotes the lifetime of
(Bt ).
There are two interesting fundamental properties of Brownian mo-
tion that depend on the global geometry of M :
23
24 5. PROBABILISTIC PROPERTIES

(1) How much time does (Bt ) spend in bounded regions of M

(more precisely open relatively compact subsets). We call the
weighted manifold recurrent if the expectation of this time is
infinite for all relatively compact open sets and transient if it
is finite for all relatively compact open sets.
(2) Does 1 = Pt 1 = P(Bt ∈ M | B0 = ·) hold for all t > 0? If
this is the case, Brownian motion does not leave M in finite
time and we call the weighted manifold stochastically complete.
Otherwise, it is called stochastically incomplete.

1. Recurrence
We start by investigating the average time spent by Brownian mo-
tion in a relatively compact open set Ω ⊆ M . To this end, we compute
Z ∞ Z ∞
Ex 1{s∈R|Bs ∈Ω} (t)dt = Ex 1Ω (Bt )dt
0
Z0 ∞
= Pt 1Ω (x)dt
0
The map G : L (M, µ)+ → L (M, µ) defined by
1 +
Z ∞
Gf = Pt f dt
0
is called Green operator. With the help of the spectral theorem, it can
be proven that
Gf = lim (α − ∆F )−1 f,
α→0+
and this limit is monotone.
The weighted manifold is called recurrent if µ({0 < Gf < ∞}) = 0
for all f ∈ L1 (M, µ)+ . In contrast, it is called transient if for all
f ∈ L1 (M, µ)+ we have Gf < ∞ a.s. (note that f > 0 a.s. implies
Gf > 0 a.s.). First we give an analytic criterion for recurrence in terms
of a Liouville type result.
Theorem 5.1 (A characterization of recurrence). The weighted man-
ifold (M, g, µ) is recurrent if and only if every f ∈ Wloc
1
(M ) ∩ L∞ (M )
with ∆f ≤ 0 is constant.
Proof. Here we only show the if part.
Idea: If Gf < ∞ a.s., we have ∆Gf = −f ≤ 0. Hence, if Gf ∈
∞
L (M ), our assumption implies that Gf is constant and we arrive at
0 = ∆Gf = −f ≤ 0,
showing f = 0. Since in general Gf will not be essentially bounded
and Gf 6= ∞ is weaker than Gf < ∞ a.s., several approximations are
needed to make this argument rigorous.
We use the notation Gα = (α − ∆F )−1 .
Let f ∈ L1 (X, µ)+ with Gf 6= ∞. We show f = 0.
 1. RECURRENCE 25

We first prove Gf < ∞ a.s. The resolvent identity implies that for
α, β > 0 we have
Gα f = Gβ f + (β − α)Gβ Gα f.
Letting α → 0+, we infer (using monotonicity in α)
Gf = Gβ f + βGβ Gf.
Assume that there exists A ⊆ M with µ(A) > 0 and Gf = ∞ a.s. on
A. Then f 6= 0. Since Gβ is positivity improving, we have Gβ f > 0
a.s. and Gβ Gf ≥ nGβ 1A for all n ≥ 1 with Gβ 1A > 0 a.s. Together
with the previous identity, this implies
Gf ≥ nGβ 1A → ∞ a.s., as n → ∞.
Hence, Gf = ∞ a.s., which contradicts our assumption.
Without loss of generality we can assume f ∈ L1 (M, µ) ∩ L2 (M, µ)
(else consider f ∧ 1 and use {f = 0} = {f ∧ 1 = 0}). Moreover, we can
assume f Gf ∈ L1 (M, µ)+ ∩ L2 (M, µ) (else consider g = f /(Gf ∨ 1)
and use
f
gGg ≤ Gf ≤ f ∈ L1 (M, µ) ∩ L2 (M, µ)
Gf ∨ 1
and {f = 0} = {g = 0}).
The L2 -lower semicontinuity of the energy (which follows from H01 (M, µ)
being a Hilbert space) yields
Z Z
|∇(Gf )|g dµ ≤ lim inf
2
|∇(Gα f )|2g dµ
M α→0+ M
= lim infh−∆Gα f, Gα f i
α→0+
= lim inf (hf, Gα f i − αhGα f, Gα f i)
α→0+
Z
≤ f Gf dµ < ∞.
M
For the second equality we used Gα f ∈ W01 (M, µ), which allows the
application of Green’s formula. This implies Gf ∈ Wloc 1
(M ) (actually
one also has to prove Gf ∈ Lloc (M ), but this is a consequence of local
2

Poincaré inequalities - we refrain from giving details).

From Gf ∈ L2loc (M ), we infer Gα f → Gf in L2loc (M ). Using the
continuity of ∆ with respect to the weak-∗-topology on distributions,
we infer
∆Gf = lim ∆Gα f = lim (−f + αGα f ) = −f,
α→0+ α→0+

where we used Gf < ∞ a.s. for the last identity.

Let α > 0. We choose a C 2 -function C : R → R with the following
properties:
(1) Cn (t) = t for t ≤ α − 2/n and Cn (t) = α − 1/n for t ≥ α − 1/n.
(2) Cn is 1-Lipschitz.
26 5. PROBABILISTIC PROPERTIES

(3) Cn is increasing and concave.

Then Cn (t) → t ∧ α, as n → ∞, and Cn′ (t) → D(t), as n → ∞, with
D(t) = 1 if t < α and D(t) = 0 if t ≥ α. Using our chain rules, we
obtain Cn (Gf ) ∈ Wloc
1
(M ) and
∆Cn (Gf ) = Cn′′ (Gf )|∇f |2 + Cn′ (Gf )∆Gf
= Cn′′ (Gf )|∇f |2 − Cn′ (Gf )f.
Letting n → ∞ and using Cn′′ ≤ 0, we infer
∆(Gf ∧ α) ≤ −D(Gf )f ≤ 0.
Hence, Gf ∧ α ∈ 1
Wloc (M ) ∩ L∞ (M )with ∆(Gf ∧ α) ≤ 0. Our assump-
tion implies that Gf ∧ α is constant, leading to
0 = ∆(Gf ∧ α) ≤ −D(Gf )f ≤ 0.
Since D(Gf ) = 1 on {Gf < α}, Gf < ∞ a.s. and α > 0 was arbitrary,
we infer f = 0 a.s. □
Theorem 5.2 (Karp’s volume growth test for recurrence). Assume
that M is complete and that for some o ∈ M and r0 > 0 we have
Z ∞
r
dr = ∞.
r0 µ(Br (o))

Then (M, g, µ) is recurrent.

Proof. Let f ∈ Wloc 1
(M ) ∩ L∞ (M ) with ∆f ≤ 0. Without loss of
generality we assume kf k∞ ≤ 1. Then h = 1 − f ∈ Wloc
1
(M ) ∩ L∞
loc (M ),
h ≥ 0 and ∆h ≥ 0. Moreover,
k1Br (o) f k22 ≤ kf k2∞ µ(Br (o)).
Hence, our volume growth assumption yields
Z ∞
r
dr = ∞.
r0 k1Br (o) f k2
2

We obtain that f is constant from Karp’s Liouville theorem. □

Remark 5.3. The volume growth test is satisfied if there are R, C > 0
such that µ(Br (o)) ≤ Cr2 log r for all r ≥ R.

2. Stochastic completeness
In this section we study stochastic completeness of (M, g, µ). Here,
(M, g, µ) is called stochastically complete if Pt 1 = 1 for all t > 0.
Remark 5.4. It is quite remarkable that there are complete but stochas-
tically incomplete manifolds (e.g. model manifolds of a certain volume
growth, see [3]). In contrast, in the incomplete case it is clear that sto-
chastic incompleteness can occur. On open relatively compact domains
Ω ⊆ Rd the minimal Brownian motion that we consider is Brownian
motion that is killed upon hitting the boundary ∂Ω. Since Brownian
 2. STOCHASTIC COMPLETENESS 27

motion will will hit ∂Ω with probability one, the probability that it
stays in Ω for all times is strictly less than one.
In order to obtain criteria for stochastic completeness we consider
w : (0, ∞) × M → R, w(t, ·) = 1 − Pt 1. By our earlier discussion
w ∈ C ∞ ((0, ∞) × M ) and
(
∂t w = ∆w
.
w(0, ·) = 0 in the sense of L2loc (M )
In order to show w = 0, it suffices to prove that the previous equation
has unique bounded solutions.
Theorem 5.5 (Grigor’yan’s uniqueness class). Let M be complete and
let T > 0. Assume that u ∈ C ∞ ((0, T ) × M ) solves
(
∂t u = ∆u
.
lim u(t, ·) = 0 in Lloc
2
(M )
t→0+

Further assume that there is a monotone increasing f : (0, ∞) →

(0, ∞) with Z ∞
r
dr = ∞,
f (r)
such that for some o ∈ M and all R > 0 we have
Z TZ
|u(t, x)|2 dµdt ≤ exp(f (R)).
0 BR (o)

Then u = 0 on (0, T ) × M .
Corollary 5.6 (Grigor’yan’s volume growth test). Assume that M is
complete and that for some o ∈ M and r0 > 0 we have
Z ∞
r
♯
dr = ∞.
r0 log µ(Br (o))

Then (M, g, µ) is stochastically complete. Here, log♯ = max{log, 1}.

Remark 5.7. Stochastic completeness holds if µ(Br (o)) ≤ CeDr log r
2

for some C, D > 0 and all r large enough. If the Ricci tensor of (M, g)
is bounded from below, then by standard volume comparison results
vol(Br (o)) ≤ eDr for some D > 0 and all r > 0. Hence, in this case
(M, g, vol) is stochastically complete. This observation is due to Yau.
Proof. According to our previous discussion, it suffices to show
that any u ∈ Cb∞ ((0, ∞) × M ) with
(
∂t u = ∆u
lim u(t, ·) = 0 in L2loc (M )
t→0+

satisfies u = 0. Let
28 5. PROBABILISTIC PROPERTIES

S := sup |u(t, x)|

t>0, x∈M

and
f (r) := log♯ (S 2 T µ(Br (o))).
Our assumption yields
Z ∞
r
dr = ∞,
f (r)
and the choice of S and f implies
Z TZ
|u(t, x)|2 dµdt ≤ S 2 T µ(Br (o)) ≤ exp(f (R)).
0 BR (o)

With this at hand, u = 0 follows directly from Grigor’yan’s uniqueness

class. □
Next we establish Grigor’yan’s uniqueness class. The proof is taken
from ...
Lemma 5.8 (A priori estimate). Assume that M is complete. Let
0 ≤ a < b and u ∈ C ∞ ((a, b) × M ) satisfy ∂t u = ∆u and assume that
the limits
u(a, ·) := lim u(t, ·) and u(b, ·) := lim u(t, ·)
t→a+ t→b−

exist inL2loc (µ). If f : (0, ∞) → (0, ∞) is increasing such that

Z bZ
|u(t, x)|2 dmu(x)dt ≤ exp(f (R)) for all R > 0,
a BR (o)

then
Z Z
4
|u(b, x)| dµ(x) ≤
2
|u(a, x)|2 dµ(x) + , (♥)
BR (o) B4R (o) R2
as long as
R2
b−a≤ . (♢)
8f (4R)
Remark 5.9. This lemma says the following: Up to a small error one
can compare the size of solutions to the heat equation on a ball at a
later time with its size on a larger ball at an earlier time. The allowed
time difference depends on the sizes of the balls and the function f .
Proof. In this proof we use ρ for a 1-Lipschitz function (to be
specified later) and, as above, ϱ for the path metric on M .
Using the continuity of u and compactness of balls, we can assume
that u is smooth on an open neighborhood of [a, b] × M . For a 1-
Lipschitz function ρ : M → R and s 6∈ [a, b] we define ξ : [a, b] × M → R
 2. STOCHASTIC COMPLETENESS 29

by
ρ(x)2
ξ(t, x) := .
4(t − s)
We will choose ρ and s later. Rademacher’s theorem and the chain rule
yield
|ρ|
|∇ξ(t, ·)| ≤ .
2|t − s|
Moreover,
ρ2 (x)
∂t ξ(t, x) = − ,
4(t − s)2
which yields
∂t ξ + |∇ξ|2 ≤ 0.
For given R > 0 we define φ : M → R by
φ(x) := (3 − ρ(x, o)/R)+ ∧ 1.
It has the following properties:
• 0 ≤ φ ≤ 1 on M ,
• φ = 1 on B2R (o),
• φ = 0 on M \ B3R (o),
• φ is 1/R-Lipschitz.
Using the completeness of M we obtain φ ∈ Lipc (M ) ⊆ Wc1 (M, µ).
For fixed t the function uφ2 eξ is locally Lipschitz. Since φ has compact
support, we obtain uφ2 eξ ∈ Lipc (M ).
Multiplying the heat equation
∂t u = ∆u
by uφ2 eξ and integrating the result over [a, b] × M yields
Z bZ Z bZ
2 ξ
(∂t u)uφ e dµdt = (∆u)uφ2 eξ dµdt.
a M a M

Since u and ξ are smooth in a neighborhood of [a, b], we obtain

Z b Z
2 ξ 1 b
(∂t u)uφ e dt = ∂t (u2 )φ2 eξ dt
a 2 a
Z
1 2 2 ξ b 1 b
= u φ e a− (∂t ξ)φ2 u2 eξ dt.
2 2 a
We evaluate the integral over M with the help of Green’s formula
(use u(t, ·) ∈ C ∞ (M ) and approximate u(t, ·)φ2 eξ(t,·) ∈ Wc1 (M, µ) by
smooth compactly supported functions)
Z Z
(∆u)uφ e dµ = −
2 ξ
h∇u, ∇(uφ2 eξ )idµ.
M M
30 5. PROBABILISTIC PROPERTIES

The chain- and product rule yield

−h∇u, ∇(uφ2 eξ )i = − |∇u|2 φ2 eξ − h∇u, ∇ξiuφ2 eξ − 2h∇u, ∇φiuφeξ
≤ − |∇u|2 φ2 eξ + |∇u||∇ξ||u|φ2 eξ
 
1
+ |∇u| φ + 2|∇φ| u eξ
2 2 2 2
2
 
1
= − |∇u| + |∇u||∇ξ||u| φ2 eξ + 2|∇φ|2 u2 eξ .
2
2
Combining these inequalities and ∂t ξ + |∇ξ|2 ≤ 0, we infer
Z b Z bZ Z bZ
2 2 ξ 2 2 ξ
u φ e dµ = (∂t ξ)φ u e dµdt + 2 (∆u)uφ2 eξ dµdt
M a a M a M
Z bZ


≤ −|∇ξ|2 u2 − |∇u|2 + 2|∇u||∇ξ||u| φ2 eξ dµdt
a M
Z bZ
+4 |∇φ|2 u2 eξ dµdt
a M
Z bZ Z bZ
=− (|∇u| − |∇ξ||u|) dµdt + 4
2
|∇φ|2 u2 eξ dµdt.
a M a M

Hence,
Z b Z bZ
2 2 ξ
u φ e dµ ≤4 |∇φ|2 u2 eξ dµdt.
M a a M

Rademacher’s theorem yields |∇φ| ≤ 1/R2 . Using also the other

2

properties of φ, we obtain
Z Z
2 ξ(x,b)
u(b, x) e dµ(x) ≤ u(a, x)2 eξ(x,a) dµ(x)
BR (o) B4R (o)
Z bZ
4
+ 2 u2 eξ dµdt.
R a B4R (o)\U2R (o)

Now we choose ρ and s. Let ρ : M → R, ρ(x) := (ϱ(x, o) − R)+ and

s := 2b − a. For all t ∈ [a, b] we obtain
b − a ≤ s − t ≤ 2(b − a)
and, hence,
ρ(x)2 ρ(x)2
ξ(t, x) = − ≤− ≤ 0.
4(s − t) 8(b − a)
Moreover, ξ = 0 on BR (o) and ρ ≥ R on B4R (o) \ U2R (o), which implies
R2
ξ≤− on [a, b] × B4R (o) \ U2R (o).
8(b − a)
 2. STOCHASTIC COMPLETENESS 31

Plugged into the last integral inequality

Z Z
u(b, x) dµ(x) ≤
2
u(a, x)2 dµ(x)
BR (o) B4R (o)
 Z bZ
4 R2
+ 2 exp − u2 dµdt.
R 8(b − a) a B4R (o)

According to the assumptions u satisfies the growth bound

Z bZ
u2 dµdt ≤ exp(f (4R))
a B4R (o)

and our inequality further simplifies to

Z Z
u(b, x) dµ(x) ≤
2
u(a, x)2 dµ(x)
BR (o) B4R (o)
 
4 R2
+ 2 exp − + f (4R) .
R 8(b − a)
If (♢) holds, then
R2
− + f (4R) ≤ 0.
8(b − a)
This shows the a priori estimate. □
Proof of Grigor’yan’s uniqueness class. Let u ∈ C ∞ ((0, T )×
M ) solve
(
∂t u = ∆u
.
lim u(t, ·) = 0 in L2loc (M )
t→0+

We define u(0, x) := 0 for all x ∈ M .

Let R > 0 and 0 < t < T . We set Rk := 4k R, τ0 = 0 and
1 Rk2
τk := , k ∈ N.
128 f (Rk )
We define the times
( Pk Pk
t − l=0 τl if t − l=1 τl > 0,
tk :=
0 else.
If tk−1 6= 0, then 0 ≤ tk < tk−1 and
2 2
1 Rk2 1 16Rk−1 Rk−1
tk−1 − tk ≤ τk = = = .
128 f (Rk ) 128 f (4Rk−1 ) 8f (4Rk−1 )
Hence, the difference satisfies the condition (♢) of the previous Lemma.
Inequality (♥) yields
Z Z
4
u(tk−1 , x) dµ(x) ≤
2
u(tk , x)2 dµ(x) + 2 .
BR (o) BR (o) Rk−1
k−1 k
32 5. PROBABILISTIC PROPERTIES

Iterating this inequality yields

Z Z Xk
4
u(t, x) dµ(x) ≤
2 2
u(tk , x) dµ(x) + 2
BR (o) BRk (o) Rl−1
Z l=1
C
≤ u(tk , x)2 dµ(x) + .
BRk (o) R2
Here, C > 0 is a constant that is independent of k. If we show tk = 0
for some k ∈ N, then the claim follows from this inequality using
u(tk , ·) = u(0, ·) = 0 and letting R → ∞.
Since f is monotone increasing, the assume growth condition yields
Z ∞ X∞ Z Rk+1 X∞ 2
r r Rk+1
∞= dr ≤ dr ≤ .
R f (r) k=0 R k
f (r) k=0
f (R k)

From this we obtain

X
∞
τk = ∞,
k=0
PN
i.e., t − k=0 τk ≤ 0 for some N big enough, which leads to tN = 0. □
 Bibliography

[1] Luigi Ambrosio, Nicola Gigli, and Giuseppe Savaré. Calculus and heat flow in
metric measure spaces and applications to spaces with Ricci bounds from below.
Invent. Math., 195(2):289–391, 2014.
[2] Lawrence C. Evans and Ronald F. Gariepy. Measure theory and fine properties
of functions. Textbooks in Mathematics. CRC Press, Boca Raton, FL, revised
edition, 2015.
[3] Alexander Grigor’ yan. Analytic and geometric background of recurrence and
non-explosion of the Brownian motion on Riemannian manifolds. Bull. Amer.
Math. Soc. (N.S.), 36(2):135–249, 1999.
[4] Alexander Grigor’yan. Heat kernel and analysis on manifolds, volume 47 of
AMS/IP Studies in Advanced Mathematics. American Mathematical Society,
Providence, RI; International Press, Boston, MA, 2009.
[5] Matthias Keller, Daniel Lenz, and Radosław K. Wojciechowski. Graphs and
discrete Dirichlet spaces, volume 358 of Grundlehren der mathematischen
Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer,
Cham, [2021] ©2021.
[6] Norman G. Meyers and James Serrin. H = W . Proc. Nat. Acad. Sci. U.S.A.,
51:1055–1056, 1964.
[7] K. T. Sturm. Analysis on local Dirichlet spaces. III. The parabolic Harnack
inequality. J. Math. Pures Appl. (9), 75(3):273–297, 1996.
[8] Karl-Theodor Sturm. Analysis on local Dirichlet spaces. I. Recurrence, conser-
vativeness and Lp -Liouville properties. J. Reine Angew. Math., 456:173–196,
1994.
[9] Karl-Theodor Sturm. Analysis on local Dirichlet spaces. II. Upper Gaussian
estimates for the fundamental solutions of parabolic equations. Osaka J. Math.,
32(2):275–312, 1995.

33