Discrete tori and trigonometric sums

Alexander Grigor’yan Yong Lin

Department of Mathematics YMSC
University of Bielefeld Tsinghua University
33501 Bielefeld, Germany Beijing, 100084, China
[email protected] [email protected]
Shing-Tung Yau
YMSC
Tsinghua University
Beijing, 100084, China
and
Department of Mathematics
Harvard University
Cambridge, Massachusetts, USA
[email protected]
March 2022

Dedicated to our dear friend Peter Li on the occasion of his birthday

Contents
1 Introduction 2

2 Discrete time heat kernels 5

2.1 Weighted graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Product of regular graphs . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Quotient of graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 The heat kernel on Zn . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.5 Heat kernels on discrete tori . . . . . . . . . . . . . . . . . . . . . . . . 9
2010 Mathematics Subject Classification. Primary: 60J35; Secondary: 11L03, 05C81, 60J10.
Key words and phrases. Trace formula, heat kernel, Markov chain, discrete torus, trigonometric
sum
AG is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - SFB
1283/2 2021 - 317210226. YL is supported by the National Science Foundation of China (Grant No.
12071245 and 11761131002)

1
3 Trigonometric sums 10
3.1 Eigenfunctions on discrete tori . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Main result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3 An example of computation . . . . . . . . . . . . . . . . . . . . . . . . 13

References 15

1 Introduction
The well know Poisson summation formula says that, for any positive real t,
r  2 2
X πX π n
−k2 t
e = exp − . (1.1)
k∈Z
t n∈Z
t

It can be proved by using the heat kernel pSt (x, y) on the unit circle S as follows. For
the trace of the heat operator
Z
Pt f (x) = pSt (x, y) f (y) dy,
S

acting in L2 (S), there are two expressions as follows:

∞
X
trace Pt = e−λj t , (1.2)
j=0

d2
where {λj } is the sequence of all the eigenvalues of the Laplace operator Δ = dx2
on S
counted with multiplicity, and
Z
trace Pt = pSt (x, x) dx. (1.3)
S

Comparing (1.2) and (1.3), using that that the sequence {λj } consists of the numbers
k 2 , k ∈ Z, and that X
pSt (x, y) = pRt (x + 2πn, y)
n∈Z

where !
1 |x − y|2
pRt (x, y) = exp −
(4πt)n/2 4t
is the Gauss-Weierstrass function, one obtains (1.1) (see, for example, [5, Exercise
10.18]).
Similar ideas have been widely used in the literature for obtaining various trace
formulas and estimates of eigenvalues of Riemannian manifolds, for example, in [ 1], [3],
[4], etc. In the framework of graphs we mention [2] where the above idea was applied
to the heat kernels pTt (x, y) on discrete tori T in Zn and, hence, a certain analogue of
the Poisson summation formula was obtained.

2
 In this paper we also work with discrete tori but use a discrete time heat kernel
qs (x, y), s ∈ Z+ , instead of the one with a continuous time t ∈ R+ . In fact, qs (x, y) is
the transition density of a simple random walk on the graph in question. As a result,
we obtain explicit formulas for some trigonometric sums that seems to be new.
Our results are stated in Theorem 3.2 and Corollary 3.4. To illustrate them, let us
present them for 2-dimensional discrete tori. For any 2 × 2 integer matrix M and for
any non-negative integer s, set
X s!
Cs (M ) = . (1.4)
z1 !z2 ! (|v1 | + z1 )! (|v2 | + z2 )!
v∈M Z , z∈Z+
2 2

|v1 |+|v2 |+2z1 +2z2 =s

Consider a 2 × 2 integer matrix A with m := det A>1. Then the lattice AZ2 contains
mZ2 so that the quotient
TA = AZ2 /mZ2
is well defined (in the sense of groups) and can be regarded as a discrete torus. Note
that TA contains m vertices.
Theorem 1.1. For any non-negative integer s, we have the identity
X  2πk 2πl
s
m
cos + cos = s Cs (M ) , (1.5)
m m 2
(k,l)∈TA

where M = m (A∗ )−1 and A∗ denotes the transpose of A.

Example 1.2. Consider the matrix
 
3 1
A=
−1 2
with m = det A = 7. The torus TA is shown on Fig. 1, and it contains the following 7
vertices: (0, 0), (1, 2), (2, 4), (4, 1), (3, 6), (6, 5), (5, 3) .

Figure 1: The lattice 7Z2 (double lines), the lattice AZ2 (single lines) and the torus
TA (shaded).

3
 Hence, the sum in the left hand side of (1.5) is equal to
 s  s  s
s 2π 4π 4π 8π 8π 2π
σ s := 2 + cos + cos + cos + cos + cos + cos
7 7 7 7 7 7
 s  s  s
6π 12π 12π 10π 10π 6π
+ cos + cos + cos + cos + cos + cos .
7 7 7 7 7 7

In this case the matrix M is equal to

 
2 1
M= ,
−1 3

and the lattice M Z2 is shown on Fig. 2.

Figure 2: The lattice M Z2

Computation by means of (1.4) is performed in Section 3.3 and results in

C1 (M ) = 0, C2 (M ) = 4, C3 (M ) = 6, C4 (M ) = 44, C5 (M ) = 130
7
By (1.5) we have σ s = C
2s s
(M ) which yields

21 77 455
σ 1 = 2, σ 2 = 7, σ 3 = , σ4 = , σ5 = .
4 4 16
The structure of this paper is as follows. In Section 2 we have collected all necessary
information about the Markov operators on weighted graph and their heat kernels,
including the heat kernels on Cartesian products and quotients of graphs. These facts
are rather elementary but they are hardly available in the literature in this concise form.
Section 3 contains the main results mentioned above, their proofs, and examples.

4
2 Discrete time heat kernels
2.1 Weighted graphs
We briefly outline some fact from [6] about heat kernels on weighted graphs. Let Γ
be a locally finite graph where we denote by Γ also the set of vertices of this graph.
We write x ∼ y if the vertices x, y of Γ are connected by an edge in Γ. Let μxy be a
symmetric non-negative function on pairs xy of vertices such that μxy > 0 ⇔ x ∼ y.
Define the weight on the vertices of Γ by
X X
μ (x) = μxy = μxy
{y∈Γ:y∼x} y∈Γ

and assume in what follows that μ (x) > 0 for all x ∈ Γ (that is, each vertex has at
least 1 edge).
Consider a Markov operator P = PΓ acting on functions f : Γ → R as follows:
1 X
P f (x) = f (y) μxy .
μ (x) y∈Γ

It is clear that P f ≥ 0 if f ≥ 0 and P 1 = 1. It follows that P acts in any space lr (Γ, μ)

with r ∈ [1, ∞] and satisfies the norm-bound kP k ≤ 1. Besides, P is self-adjoint in
l2 (Γ, μ).
The weight μxy is called simple if μxy = 1 for all x ∼ y. In this case, μ (x) = deg (x)
and X
1
P f (x) = f (y) .
deg (x) y∼x

For any s ∈ Z+ the power P s is well defined, and the sequence {P s }s≥0 is a reversible
Markov chain on Γ. It is easy to see that
X
P s f (x) = qs (x, y) f (y) μ (y) ,
y∈Γ

1
where the function qs (x, y) = qsΓ (x, y) is defined inductively by q0 (x, y) = δ ,
μ(y) x,y

μxy X
q1 (x, y) = and qs+1 (x, y) = qs (x, z) q1 (z, y) μ (z) .
μ (x) μ (y) z∈Γ

The function qs (x, y) is called the discrete time heat kernel or the transition density
of the Markov chain {P s } .

2.2 Product of regular graphs

A graph Γ is called d-regular if any vertex has exactly d neighbors, that is, deg (x) = d
for all x ∈ Γ.
Let {Γj }nj=1 be a finite sequence of graphs. Consider their Cartesian product

Γ = Γ1 Γ2 ...Γn .

5
The vertices of Γ are n-tuples x = (x1 , ...xn ) where xj ∈ Γj . We write for some
j = 1, ..., n
Γj
x∼y
if
xj ∼ yj and xk = yk for all k 6= j.
The edges x ∼ y in Γ are defined by the following rule:
Γj
x ∼ y ⇔ x ∼ y for some j = 1, ..., n. (2.1)

Assume further that each Γ j is dj -regular. Then Γ is d-regular with

d = d1 + ... + dn .

Let us endow all the graphs Γ j and Γ with a simple weight. We have then for the
Markov operator PΓ on Γ
n
1X 1 XX
PΓ f (x) = f (y) = f (y) .
d y∼x d j=1 Γ
j
y ∼x

Let us consider the Markov operator PΓj on Γj as acting also on functions f (x) on Γ
along the component xj , so that

1 X
PΓj f (x) = f (y) .
dj Γ
j
y ∼x

It follows that n X n
X X
d PΓ f (x) = f (y) = dj PΓj f (x) .
j=1 Γj j=1
y ∼x

Since all the operators PΓj commute on Γ, we obtain that, for any s ∈ Z+ ,
n
!s   n
X X s Q sj s j
s s
d PΓ = d j P Γj = d P , (2.2)
j=1 n
s1 , ..., sn j=1 j Γj
(s1 ,...,sn )∈Z+
s1 +...+sn =s

s

s!
where s1 ,...,sn
= s1 !...sn !
is a multinomial coefficient. Since
X
PΓs f (x) = dqsΓ (x, y) f (y) ,
y∈Γ

it follows that
X  
s Q
n
s +1 Γj
ds+1 qsΓ (x, y) = dj j q sj (xj , yj ) . (2.3)
s1 , ..., sn j=1
(s1 ,...,sn )∈Zn
+
s1 +...+sn =s

6
2.3 Quotient of graphs
Let (Γ, μ) be a weighted graph with μ (x) > 0 so that the Markov operator PΓ is well
defined. Let G be a group of weighted graph automorphisms of Γ, that is,

μgx,gy = μx,y ∀g ∈ G, ∀x, y ∈ Γ.

Then the vertex weight μ (x) is also G-invariant. It follows that the operator PΓ
commutes with G, that is,
PΓ (f ◦ g) = (PΓ f ) ◦ g,
because
1 X
(PΓ f ) ◦ g (x) = f (y) μgx,y
μ (gx) y∈Γ
1 X
= f (y) μx,g−1 y
μ (x) y∈Γ
1 X
= f (gz) μx,z
μ (x) z∈Γ
= PΓ (f ◦ g) (x) .

Consequently, also qs (x, y) commutes with G, that is,

qs (x, y) = qs (gx, gy) ∀g ∈ G.

Consider the quotient Q = Γ/G that consists of the equivalence classes [x] of vertices
x ∈ Γ under the equivalent relation

x ≡ y mod G ⇔ x = gy for some g ∈ G.

The quotient Q has a natural weight:

X
μQ
[x],[y] := μx,gy , (2.4)
g∈G


so that Q, μQ is a weighted graph. For example, if the weight μxy on Γ is simple then

μQ
[x],[y] = card {g ∈ G : x ∼ gy} .

However, the weight μQ may be not simple because the G-orbit of y may have more
than 1 vertex adjacent to x.
Observe that always
μ ([x]) = μ (x) (2.5)
because X XX X
μ ([x]) = μ[x],[y] = μx,gy = μx,z = μ (x) .
[y]∈Q [y]∈Q g∈G z∈Γ

Any G-periodic function f on Γ can be regarded as a function on Q by

f ([x]) = f (x) .

7
Clearly, PΓ f is also G-periodic. Let us verify that
PQ f ([x]) = PΓ f (x) . (2.6)
Indeed, we have
1 X
PQ f ([x]) = f ([y]) μ[x],[y]
μ ([x])
[y]∈Q
1 X X
= f (y) μx,gy
μ (x) g∈G
[y]∈Q
1 X
= f (z) μxz
μ (x) z∈Γ
= PΓ f (x) .
Lemma 2.1. We have for all x, y ∈ Γ and s ∈ Z+
X
qsQ ([x] , [y]) = qsΓ (x, gy) . (2.7)
g∈G

Proof. Clearly, the right hand side of (2.7) is G-periodic in x and y and, hence, can
be regarded as a function on Q × Q. For any G-periodic function f on Γ, we have by
(2.6)
X
PQs f ([x]) = PΓs f (x) = qsΓ (x, z) f (z) μ (z)
z∈Γ
XX
= qsΓ (x, gy) f (gy) μ (gy)
g∈G [y]∈Q
XX
= qsΓ (x, gy) f (y) μ (y)
[y]∈Q g∈G
!
X X
= qsΓ (x, gy) f ([y]) μ ([y])
[y]∈Q g∈G

whence (2.7) follows.

In what follows we simplify notation by writing x instead of [x] when this does not
cause confusion.

2.4 The heat kernel on Zn

It is known that the transition density qsZ (x, y) of a simple random walk on Z is given
by (

1 s
s−k , s ≥ k and s ≡ k mod 2,
qsZ (x, y) = 2 s+1
2 (2.8)
0, otherwise,
where k = |x − y| (see [6, Eq. (5.6)]). Let us determine qsZ (x, y) . By (2.3), we have
n

X s!
(2n)s+1 qsZ (x, y) = 2s1 +1 qsZ1 (x1 , y1 ) . . . 2sn +1 qsZn (xn , yn ) .
n

s !...sn !
s +...+s =s 1
1 n

8
Setting
ki = |xi − yi | ,
we obtain
X s! Qn s!
(2n)s+1 qsZ (x, y) =
i si +ki

n

s !...sn ! i=1 si −ki

s1 +...+sn =s 1 2
! 2 !
X Q
n 1
= s! si −ki

si +ki

,
s1 +...+sn =s i=1 2
! 2
!

where the summation indices s1 , ..., sn satisfy in addition

si ≥ ki and si ≡ ki mod 2.
si −ki
Changing ji = 2
, setting j = (j1 , ..., jn ), k = (k1 , ..., kn ), and using the multiindex
notation
P
n Q
n
|j| = ji and j! = ji ,
i=1 i=1

we obtain
1 X s!
qsZ (x, y) =
n
. (2.9)
(2n)s+1 j! (k + j)!
{j∈Zn+ : 2|j|+|k|=s}

2.5 Heat kernels on discrete tori

Let us fix some integer valued n × n matrix M with

m := det M > 1.

We regard M Zn as an additive group that acts on Zn by shifts. Consider a discrete

torus
T = Zn /M Zn (2.10)
that is a finite graph with m vertices.
Let μ be the weight on T that comes from the simple weight of Zn by (2.4). By
(2.5), we have
μ (x) = 2n for any x ∈ T.
By (2.7), the heat kernel on (T, μ) is given by
X
qsZ (x + v, y) .
n
qsT (x, y) =
v∈M Zn

Using (2.9) and setting x = y, we obtain

X 1 X X s!
qsZ (x + v, x) =
n
qsT (x, x) = . (2.11)
v∈M Zn
(2n)s+1 v∈M Zn j∈Zn
j! (v + j)!
+
2|j|+|v|=s

9
3 Trigonometric sums
3.1 Eigenfunctions on discrete tori
The following function is an eigenfunction of PZn for any w ∈ Rn :

fw (x) = e2πihw,xi .

Indeed, we have
n
1 X 1 X
PZn fw (x) = fw (y) = (fw (x + ek ) + fw (x − ek ))
2n y∼x 2n k=1
n n
!
1 2πihw,xi X 2πihw,ek i
1X
= e e + e−2πihw,ek i = cos 2πwk fw (x) ,
2n k=1
n k=1

where {ek } is a canonical basis in Rn . Hence, we obtain

P Zn fw = α w f w

with n
1X
αw = cos 2πwk . (3.1)
n k=1
Note that the functions fw0 and fw0 are equal if and only of w0 = w00 mod Zn so that
we can assume that w ∈ Rn /Zn . Consider a lattice

W := (M ∗ )−1 Zn /Zn

and a torus (2.10).

Lemma 3.1. The function fw is M Zn -periodic if and only if w ∈ W. Consequently,
for any w ∈ W , the function fw is an eigenfunction of PT with the eigenvalue (3.1).
Moreover, the family {fw }w∈W forms an orthogonal basis in l2 (T, μ) .
Proof. To prove the first claim, it suffices to verify that

fw (x) ≡ 1 for all x ∈ M Zn (3.2)

if and only if w ∈ W . If x = M y and w = (M ∗ )−1 z where y, z ∈ Zn then

fw (x) = e2πihw,xi = exp 2πih(M ∗ )−1 z, M yi


= exp 2πihz, M −1 M yi = exp (2πihz, yi) = 1.

If (3.2) is true, then, for all x = M y with y ∈ Zn , we have

hw, M yi ∈ Z.

Let the columns of M be u1 , ..., un . Then for y = ek we obtain M y = uk so that

hw, uk i = zk

10
for some zk ∈ Z. The matrix of this linear system is M ∗ , whence

w = (M ∗ )−1 z,

which finishes the proof of the first claim.

The fact that fw is an eigenfunction of PT follows from (2.6) and the fact that fw
is an eigenfunction of PZn as was verified above.
Let us verify that the family {fw }w∈W is orthogonal For all w0 6= w00 , we have
X X
hfw0 , fw00 i = fw0 (x) fw00 (x)μ (x) = exp (2πihw0 − w00 , xi) μ (x) = consthfw , 1i,
x∈T x∈T

where w = w0 − w00 . Since w is non-zero as an element of the torus W , the eigenfunc-

tion fw is orthogonal to the eigenfunction f0 = 1 because 0 is known to be a simple
eigenvalue of PT . Hence, fw0 ⊥fw00 as claimed.
Since the family {fw }w∈W is linearly independent and the number of elements in
this family is equal to det M ∗ = m, it follows that this family forms an orthogonal
basis in l2 (T, μ).

3.2 Main result

For any multiindex v = (v1 , ..., vn ) ∈ Zn set

|v| = |v1 | + ... + |vn | , v = (|v1 | , ..., |vn |) ,

and for v ∈ Zn+ set

v! = v1 !...vn !.
As above, let us fix an integer valued n × n matrix M with

m := det M > 1.

For any non-negative integer s, set

X s!
Cs (M ) = . (3.3)
v∈M Zn , z∈Zn
z! (v + z)!
+
|v|+2|z|=s

Now we can state and prove our main result.

Theorem 3.2. For the torus

W = (M ∗ )−1 Zn /Zn (3.4)

and for any non-negative integer s we have

n
!s
X X m
cos 2πwk = Cs (M ) . (3.5)
w∈W k=1
2s

11
Proof. Since αw with w ∈ W are the eigenvalues of PT , we obtain using (2.11)
X X
αsw = trace PTs = qsT (x, x) μ (x)
w∈W x∈T
2nm X X s!
=
(2n)s+1 v∈M Zn j∈Zn
j! (v + j)!
+
2|j|+|v|=s
m X X s!
=
(2n)s v∈M Zn j∈Zn
j! (v + j)!
+
2|j|+|v|=s
m
= Cs (M ) . (3.6)
(2n)s

Substituting the value of αw from (3.1), we obtain (3.5).

It is convenient to rewrite (3.3) in the form
X
Cs (M ) = Cs (v) , (3.7)
v∈M Zn , |v|≤s

where, for any v ∈ Zn and s ∈ Z+ ,

X s!
Cs (v) = . (3.8)
z∈Zn
z! (v + z)!
+
|z|= 12 (s−|v|)

Observe that the numbers Cs (v) do not depend on M . By (3.7), the number Cs (M )
is determined by the vertices v of the lattice M Zn lying in the l1 -ball in Zn of radius
s (see Fig. 3).

Figure 3: The nodes of a lattice M Zn lying in the l1 -ball of radius s (shaded).

It is clear from (3.8) that

if |v| 6≡ s mod 2 then Cs (v) = 0.

12
Consequently, the summation in (3.7) can be restricted to those v with |v| = s mod 2.
In the case n = 2 Theorem 3.2 can be reformulated as follows. By (3.4) we have

mW = m (M ∗ )−1 Zn /mZn .

The nodes of the torus mW have integer components. Indeed, the entries of the matrix
(M ∗ )−1 are obtained by dividing the minors of M ∗ by m = det M ∗ , which implies that
the matrix
A := m (M ∗ )−1 (3.9)
has integer entries. Clearly, we have det A = mn−1 . In particular, if n = 2 then

det A = m.

In this case, also the converse is true.

Lemma 3.3. For any 2 × 2 integer matrix A with m = det A > 1, there exists an
integer matrix M such that (3.9) is true.

Proof. Indeed, set

M = m (A∗ )−1 (3.10)
so that (3.9) is satisfied. Since m = det A∗ , it follows that M has integer entries, which
finishes the proof.
Now we reformulate Theorem 3.2 in the case n = 2.

Corollary 3.4. For any 2 × 2 integer matrix A with m = det A > 1 and for any
non-negative integer s, we have the identity
X  2πa1 2πa2
s
m
cos + cos = s Cs (M ) , (3.11)
2 2
m m 2
a∈AZ /mZ

where M = m (A∗ )−1 and Cs (M ) is defined by (3.3).

Proof. Indeed, defining W by (3.4), we see that

a
w∈W ⇔w=
m
where
a ∈ m (M ∗ )−1 Zn /mZn = AZ2 /mZ2 .
Hence, (3.11) follows from (3.5).

3.3 An example of computation

Example 3.5. Consider the matrix
 
3 1
A=
−1 2

13
with m = det A = 7. The torus TA = AZ2 /mZ2 is shown on Fig. 1. It contains the
following 7 different points

(0, 0) , (1, 2) , (2, 4) , (4, 1) , (3, 6) , (6, 5), (5, 3) .

Hence, the sum in the left hand side of (3.11) becomes

 s  s  s
s 2π 4π 4π 8π 8π 2π
σ s := 2 + cos + cos + cos + cos + cos + cos
7 7 7 7 7 7
 s  s  s
6π 12π 12π 10π 10π 6π
+ cos + cos + cos + cos + cos + cos
7 7 7 7 7 7
Let us compute the right hand side of (3.11). By (3.10) we have
 
∗ −1 2 1
M = 7 (A ) = .
−1 3

The lattice M Z2 is shown on Fig. 2. Let us compute the coefficients Cs (M ) for

s = 1, ..., 5. One can see from Fig. 2 that

v ∈ M Z2 : |v| = 1 or 2 = ∅,

v ∈ M Z2 : |v| = 3 = {± (2, −1)}

v ∈ M Z2 : |v| = 4 = {± (1, 3)}

v ∈ M Z2 : |v| = 5 = {± (3, 2) , ± (1, −4)} .

In all the sums below we have v ∈ M Z2 and z ∈ Z2+ . Using (3.7) and (3.8), we obtain
the following: X
C1 (M ) = Cs (v) = 0,
|v|=1

X X
C2 (M ) = C2 (v) + C2 (v)
|v|=0 |v|=2
XX 2! XX 2! X 2
= + = = 4,
z! (v + z)! z! (v + z)! z +z =1 z!z!
|v|=0 |z|=1 |v|=2 |z|=0 1 2

X X
C3 (M ) = C3 (v) + C3 (v)
|v|=1 |v|=3
XX 3! X 6 6
= = =2 = 6,
z! (v + z)! v! 2!1!
|v|=3 |z|=0 |v|=3

X X X
C4 (M ) = C4 (v) + C4 (v) + C4 (v)
|v|=0 |v|=2 |v|=4
X 4! XX 4!
= +
z!z! z! (v + z)!
|z|=2 |v|=4 |z|=0

14
 X 24 X 24 24 24 24
= 2 + = 2 +2 2 +2 = 44,
z1 +z2 =2
(z1 !z2 !) v! (1!1!) (2!0!) 1!3!
|v|=4

X X X
C5 (M ) = C4 (v) + C4 (v) + C4 (v)
|v|=1 |v|=3 |v|=5
XX 120 XX 5!
= +
z! (v + z)! z! (v + z)!
|v|=3 |z|=1 |v|=5 |z|=0
X  
120 120 120
=2 +2 +
z !z ! (z1 + 2)! (z2 + 1)!
z1 +z2 =1 1 2
3!2! 1!4!
   
1 1 120 120
= 240 + +2 + = 130.
3!1! 2!2! 3!2! 1!4!

By (3.11) we have
7
σs =
Cs (M ) .
2s
Substituting the above values of Cs (M ), we obtain
21 77 455
σ 1 = 2, σ 2 = 7, σ 3 = , σ4 = , σ5 = .
4 4 16

References
[1] Cheng S.Y, Li P., Heat kernel estimates and lower bounds of eigenvalues, Comment. Mat.
Helv., 56 (1981) 327-338.
[2] Chung F.R.K., Yau S.-T., A combinatorial trace formula, in: “Tsinghua Lectures on Geometry
and Analysis”, International Press, 1997. 107–116.
[3] Donnelly H., Li P., Lower bounds for the eigenvalues of negatively curved manifolds, Math.Z.,
172 (1980) 29-40.
[4] Grigor’yan A., Heat kernel upper bounds on a complete non-compact manifold, Revista
Matemática Iberoamericana, 10 (1994) no.2, 395-452.
[5] Grigor’yan A., “Heat kernel and Analysis on manifolds”, AMS-IP Studies in Advanced Math-
ematics 47, AMS - IP, 2009.
[6] Grigor’yan A., “Introduction to analysis on graphs”, AMS University Lecture Series 71, AMS,
2018.

15