a

r
X
i
v
:
0
8
0
2
.
0
1
3
3
v
1


[
m
a
t
h
-
p
h
]


1

F
e
b

2
0
0
8
Essential selfadjointness of the graph-Laplacian
Palle E. T. Jorgensen
Department of Mathematics
University of Iowa
Iowa City, IA 52242-1419 USA
Abstract
We study the operator theory associated with such innite graphs G
as occur in electrical networks, in fractals, in statistical mechanics, and
even in internet search engines. Our emphasis is on the determination
of spectral data for a natural Laplace operator associated with the
graph in question. This operator will depend not only on G, but
also on a prescribed positive real valued function c dened on the edges
in G. In electrical network models, this function c will determine a
conductance number for each edge. We show that the corresponding
Laplace operator is automatically essential selfadjoint. By this we
mean that is dened on the dense subspace T (of all the real valued
functions on the set of vertices G
0
with nite support) in the Hilbert
space l
2
(G
0
). The conclusion is that the closure of the operator is
selfadjoint in l
2
(G
0
), and so in particular that it has a unique spectral
resolution, determined by a projection valued measure on the Borel
subsets of the innite half-line. We prove that generically our graph
Laplace operator =
c
will have continuous spectrum. For a given
innite graph G with conductance function c, we set up a system of
nite graphs with periodic boundary conditions such the nite spectra,
for an ascending family of nite graphs, will have the Laplace operator
for G as its limit.
The author was partially supported by a grant from the US National Science
Foundation.
Math Subject Classication (2000): 47C10, 47L60, 47S50, 60J20, 81Q15,
81T75, 82B44, 90B15
Keywords: Graphs, conductance, network of resistors, resistance metric, dis-
crete dynamical systems, Hermitian operator, unbounded operators, Hilbert space,
deciency indices, innite Heisenberg matrices.
1
Contents
1 Introduction 3
2 Assumptions 4
3 The Main Theorem 5
3.1 The Graph Laplacian . . . . . . . . . . . . . . . . . . . . . 5
3.2 Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4 Operator Theory 13
5 The Energy Form 15
5.1 Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
5.2 A matrix representation . . . . . . . . . . . . . . . . . . . 21
5.3 Example 5.2 revisited . . . . . . . . . . . . . . . . . . . . . 22
5.4 Banded Matrices (A Preview) . . . . . . . . . . . . . . . 23
5.5 Extended Hilbert Spaces . . . . . . . . . . . . . . . . . . 24
5.6 Lattice Models . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.7 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.8 The Resistance Metric . . . . . . . . . . . . . . . . . . . . 34
6 Finite Dimensional Approximation 36
6.1 Systems of Graphs . . . . . . . . . . . . . . . . . . . . . . 36
6.2 Periodic boundary conditions . . . . . . . . . . . . . . . . 40
7 Boundary Conditions 42
8 Appendix 45
A Heisenbergs Innite Banded Matrices 45
2
1 Introduction
The innite graphs we consider live on a xed countable innite set, say L.
Starting with such a set L (subject to certain axioms, listed below), we get
a notion of edges as follows: Select distinguished pairs of points in L, say x
and y, and connect them by a line, called edge. In physics, when a vertex
x is given, the set of vertices connected to x with one edge is called a set
of neighbors, or nearest neighbors. Initially we do not assign direction to
the edges. So, as it stands, an edge e is dened as a special subset x, y
for selected points x, y in L. Think nearest neighbors!
A direction is only assigned when we also introduce a function I on
edges e, and then this function I is assumed to satisfy I(x, y) = I(y, x).
In electrical networks, such a function I may represent a current induced
by a potential which is introduced on a graph with xed resistors. So only
if a current function I is introduced can we dene a direction to edges,
as follows: We specify source s(e) = x, and terminal vertex t(e) = y if
I(x, y) > 0. meaning that the current ows from x to y.
In this paper we study the operator theory of innite graphs G, with
special emphasis on a natural Laplace operator associated with the graph in
question. This operator will depend not only on G, but also on a positive real
valued function c dened on the edges in G. In electrical network models,
the function c will determine a conductance number for each edge e. If
e = (xy) connects vertices x and y in G, the number c(e) is the reciprocal of
the resistance between x and y. Hence prescribing a conductance leads to
classes of admissible ows in G. When they are determined from Ohms law,
and the Kirchho laws, it leads to a measure of energy, and to an associated
graph Laplacian. We identify the Hilbert space H(G) which oers a useful
spectral theory, and our main result is a theorem to the eect that the
graph Laplacian is essentially selfadjoint, i.e., that its operator closure is a
selfadjoint operator in H(G).
Let G = (G
0
, G
1
) be an innite graph, G
0
for vertices, and G
1
for edges.
Every x in G
0
is connected to a set nbh(x) of other vertices by a nite
number of edges, but points in nbh(x) are dierent from x; i.e., we assume
that x itself is excluded from nbh(x); i.e., no x in G
0
can be connected to
itself with a single edge. Let c be a conductance function dened on G
1
.
Initially, the graph G will not be directed, but when a conductance is
xed, and we study induced current ows, then these current ows will give
a direction to the edges in G. But the edges in G itself do not come with
an intrinsic direction.
We show that the Laplace operator =
c
is automatically essentially
3
selfadjoint. By this we mean that is dened on the dense subspace T
(of all the real valued functions on G
0
with nite support) in the Hilbert
space H = H(G):= l
2
(G
0
). The explicit formula for the graph Laplacian
=
(G,c)
is given in (3.6) in section 3 below which also discusses the
appropriate Hilbert spaces. The conclusion is that the closure of the operator
is selfadjoint in H, and so in particular that it has a unique spectral
resolution, determined by a projection valued measure on the Borel subsets
of the innite half-line R
+
; i.e., the spectral measure takes values in the
projections in the Hilbert space: = l
2
(G
0
). We work out the measure.
In contrast, we note that the corresponding Laplace operator in the con-
tinuous case is not essentially selfadjoint. This can be illustrated for example
with = (d/dx)
2
on the domain T of consisting of all C
2
-functions on the
innite half-line R
+
which vanish with their derivatives at the end points.
Then the Hilbert space is L
2
(R
+
).
So our graph theorem is an instance where the analogy between the
continuous case and the discrete case breaks down.
A second intrinsic issue for the operator theory of innite graphs G, is
that generically our graph Laplace operator =
c
will have continuous
spectrum. We prove this by identifying a covariance system which implies
that the spectrum of the corresponding Laplace operator will in fact be
absolute continuous with respect to Lebesgue measure on the half-line.
In a third theorem, for a given innite graph G with conductance func-
tion c, we set up a system of nite graphs with periodic boundary conditions
such the nite spectra, for an ascending family of nite graphs, will have
the Laplace operator for G as its limit.
2 Assumptions
In order to do computations and potential theory on innite graphs G, it
has been useful to generalize the continuous Laplacian from Riemannian
geometry [AC04] to a discrete setting [BHS05], [CS07], [Kig03], [HKK02].
However the innities for graphs suggest an analogy to non-compact Rie-
mannian manifolds, or manifolds with boundary.
Once the graph Laplacian is made precise as a selfadjoint operator it
makes sense to ask for exact formulas for the spectrum of . Our Laplace
operator =
c
is associated with a xed system (G, c) where vertices and
edges are specied as usual, G = (G
(0)
, G
(1)
); and with a xed conductance
function c : G
(1)
R
+
. See (3.6) below for a formula.
And as usual our Laplace operator, =
c
is densely dened in the
4
Hilbert space
2
(G
(0)
) of all square-summable sequences on the vertices of
G; and if G is innite,
c
is not dened everywhere in
2
, but rather it has
a dense domain T in
2
. We show in the next section that
c
is essentially
selfadjoint for all choices of conductance function c.
By a graph G we mean a set G
(0)
of vertices, and a set G
(1)
of edges.
Edges e consist of pairs s, y G
(0)
. We write e = (xy); and if (xy) G
(1)
we say that x y.
Assumptions
(i) x , x (i.e.; (xx) , G
(1)
).
(ii) For every x G
(0)
,
_
y G
(0)
[y x
_
is nite.
(iii) Points x, y G
(0)
for which there is a nite path x
0
, x
1
, x
2
, . . . , x
n
with x
0
= x, x
n
= y, and (x
i
x
i+1
) G
(1)
, i = 0, . . . n 1, are said to be
connected.
(iv) We will assume that all connected components in G
(0)
are innite;
or else that G
(0)
is already connected.
3 The Main Theorem
3.1 The Graph Laplacian
In this section we specify a xed graph G (innite in the non-trivial case)
and an associated conductance function c. The associated graph Laplacian

c
will typically be an unbounded Hermitian operator with dense domain.
Our assumptions will be as above, and when the Hilbert spaces have been
selected, our main theorem states that the graph Laplacian
c
is essentially
selfadjoint; i.e., the operator closure, also denoted
c
, is a selfadjoint oper-
ator. In sections 58 we obtain consequences and applications.
The interpretation of this results in terms of boundary conditions will be
given in section 7 below. It means that
c
has a well dened and unique (up
to equivalence) spectral resolution. Then the next objective is to nd the
spectrum of the operator
c
. And a method for nding spectrum is based on
covariance. Covariance is used on other spectral problems in mathematical
physics, and it oers useful ways of getting global formulas for spectrum. As
we will see, innite models typically have graph Laplacians with continuous
spectrum.
In the nite case, of course the spectrum is the set of roots in a charac-
teristic polynomial, but unless there is some group action, it is dicult to
solve for roots by bare hands; and even if we do, only the occurrence of
groups oers insight.
5
A second approach to the nding spectra of graph Laplacians is renor-
malization: Renormalization of hierarchical systems of electrical networks
comes into play each time one passes to a new scale (upwards or down-
wards). This requires additional structure, such as is found in iterated func-
tion systems (IFSs), (see [BHS05], [DJ07], [JP98], [Kig03]), i.e., specied
nite systems of ane transformations in Euclidean space that are then
iterated recursively.
When the mappings are so iterated on a given graph, the iterations
may then be interpreted as scales in an innite graph: (post-)composition
of similarity mappings takes us further down the branches of a tree like
structure in path space. We get martingale constructions as instances of
renormalization.
Theorem 3.1 The graph Laplacian =
(G,c)
is essentially selfadjoint.
Proof. To get started we recall the setting. Given:
G: a xed innite graph. (It may be nite, but in this case the
conclusion follows from nite-dimensional linear algebra.)
G =
_
G
(0)
, G
(1)
_
.
G
(0)
: the set of vertices in G.
G
(1)
: the set of edges in G.
If x, y G
(0)
is a given pair, we say that x y when e = (xy) G
(1)
.
For x G
(0)
, set
nbh(x) =
_
y G
(0)
[y x
_
. (3.1)
Our standing assumptions are as follows:
(a) nbh(x) is nite.
(b) x , nbh(x).
H =
2
_
G
(0)
_
= all functions v : G
(0)
C such that

xG
(0)
[v (x)[
2
< . (3.2)
6
Set
u, v): =

xG
(0)
u(x)v (x) , u, v
2
_
G
(0)
_
. (3.3)
By H we refer to the completed Hilbert space
2
_
G
(0)
_
.
T:= the set of all nitely supported v H; i.e., v is in T i F G
(0)
,
F = F
v
some nite subset such that
v (x) = 0, x G
(0)
F.
e
x
:=
x
= Dirac mass, dened by
e
x
(y) =
_
1 if y = x
0 if y ,= x.
(3.4)
c : G
(1)
R
+
is a xed function taking positive values. In network
models, the function c is conductance; i.e., the reciprocal of resistance.
Assumption (symmetry): c (xy) = c (yx) , (xy) G
(1)
.
=
(G,c)
(3.5)
is the Laplacian, and is dened on T as follows:
(v) (x) : =

yx
c (xy) (v (x) v (y)) , v T, x G
(0)
. (3.6)
3.2 Lemmas
We will need some lemmas:
Lemma 3.2 The operator is Hermitian symmetric on T, and it is posi-
tive semidenite. Specically, the following two properties hold:
u, v)

2 = u, v)

2 , u, v T; (3.7)
and
u, u)

2 0, u T. (3.8)
7
Proof. Both assertions are computations:
In (3.7),
u, v)

2 =

x,yG
(0)
xy
c (x, y)
_
u(x) u(y)
_
v (x)
=

xy
c (xy) u(x)v (x)

xy
u(y)c (xy) v (x)
=

xy
u(x)c (xy) v (x)

xy
u(x)c (xy) v (y)
=

xG
(0)
u(x)

yx
c (xy) (v (x) v (y))
= u, v)

2 .
Note that the summation may be exchanged since, for each x G
(0)
,
the set of neighbors nbh(x) is nite.
In (3.8),
u, u)

2
=

xy
u(x)c (xy) (u(x) u(y))
=

xG
(0)
B
c
(x) [u(x)[
2

xy
u(x)c (xy) u(y) ,
where
B
c
(x) =

yx
c (xy) , x G
(0)
. (3.9)
The second term in the computation may be estimated with the use of
Cauchy-Schwarz as follows: Setting
c
c
(u) : =

all x,y
s.t. xy
c (xy) [u(x) u(y)[
2
; (3.10)
we show that
2u, u) = c
c
(u) 0. (3.11)
Indeed using the conditions on c: G
(1)
R+
c (xy) = c (yx) , (xy) G
(1)
;
c (xx) = 0, x G
(0)
;
8
c (xy) > 0, (xy) G
(1)
,
we get
2u, u) = 2

xG
(0)
B
c
(x) [u(x)[
2
2

xy
u(x)c (xy) u(y)
= 2

xG
(0)
B
c
(x) [u(x)[
2
2 Re

xy
u(x)c (xy) u(y)
=

xy
c (x, y)
_
[u(x)[
2
u(x)u(y) u(y)u(x) +[u(y)[
2
_
=

xy
c (xy) [u(x) u(y)[
2
= c
c
(u) .
For the general theory of unbounded Hermitian operators and their ex-
tensions, we refer the reader to [Jr78], [Nel69], [Sto51].
Denition 3.3 If is an operator with dense domain T in a Hilbert space
H, we dene its adjoint operator

by:
A vector v is in the domain dom(

) i there is a constant K such that

[v, u)[ K|u| , u T. (3.12)
When (3.12) holds, then by Riesz, there is a unique w :=

v such that
w, u) = v, u), u T. (3.13)
Note that since T is dense in H, w(=:

v) is uniquely determined by
(3.12).
Lemma 3.4 In the case of =
(G,c)
and H =
2
_
G
(0)
_
, the vector

v
for v dom(

) is given by the expression

(

v) (x) =

yx
c (xy) (v (x) v (y)) . (3.14)
Proof. Since the sum in (3.13) is nite, the RHS is well dened if v
dom(

). Since

v H,

xG
(c)
[(

v) (x)[
2
< . (3.15)
9
Set w(x):=

yx
c (xy) (v (x) v (y)).
We claim that (3.12) then holds. Indeed
w, u)

2 =

xG
(0)
_

yx
c (xy)
_
v (x) v (y)
_
_
u(x)
=

x
v (x)

yx
c (xy) (u(x) u(y))
= v, u)

2
(by the exchange of summation and Lemma 3.2).
Lemma 3.5 Let (G, c) and =
(G,c)
here as in the previous lemma.
Then the equation
v = v (3.16)
does not have non-zero solutions v
2
_
G
(0)
_
.
Proof. It is immediate from (3.7) in Lemma 3.2 that eq. (3.16) does not
have non-zero solutions in T, but the assertion is that there are no non-zero
solutions in any bigger subspace.
Also note that every solution in
2
_
G
(0)
_
to eq. (3.16) must be in
dom(

), i.e., the domain of the adjoint of with T as domain.

If v : G
(0)
C is a solution to (3.16), then
v (x)v (x) = [v (x)[
2
, x G
(0)
(3.17)
which yields v (x)v (x) 0, x G
(0)
. Hence c
c
(v) 0; see (3.10)-(3.11).
But by (3.11), then c
c
(v) = 0.
It follows from (3.10) that v must be constant on every connected com-
ponent in G
(0)
. Since all the connected components are innite, v must be
zero.
Remark 3.6 We stress that (3.16) may have non-zero solutions not in
2
.
For these solutions v, the energy will be unbounded.
Example 3.7 Let a graph system (G, c) be determined as follows:
G
(0)
= N
0
= 0, 1, 2, . . .,
G
(1)
: nbh(0) = 1 ,
10
nbh(n) = n 1, n + 1 if n > 0, and
c (n, n + 1) = n + 1.
Then the Laplace operator
c
will be unbounded in
2
as follows from

c
=
_
_
_
_
_
_
_
_
_
_
_
_
1 1 0 0 0
1 3 2 0 0
0 2 5 3 0
0 0 3 7 4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
n 2n + 1 (n + 1)
.
.
.
_
_
_
_
_
_
_
_
_
_
_
_
Then
(u)
0
= u
0
u
1
, and
(u)
n
= (2n + 1) u
n
nu
n1
(n + 1) u
n+1
, n 1.
For solving (3.16), initialize v
0
= 1. Then
v
1
= 2v
0
= 2,
v
2
=
7
2
, and inductively
v
n+1
= 2v
n

_
n
n + 1
_
v
n1
.
We get v
1
< v
2
< < v
n1
< v
n
< and
v
n+1
>
_
2
n
n + 1
_
v
n
.
Hence for the truncated summations for
2
and c applied to this solution v;
we get
1
2
c
N
(v) =
N

k=0
v
2
k
< N
which tends to .
The following lemma is from the general theory of unbounded operators
in Hilbert space, [Nel69], [Sto51], [vN31].
11
Lemma 3.8 Let be a linear operator in a Hilbert space H and dened in
a dense domain T.
Then is essentially selfadjoint (i.e., has selfadjoint closure

) if the
following conditions hold:
(i ) u, u) 0, u T
(ii ) dimv dom(

) [

v = v = 0.
Proof. This is in the literature, e.g. [vN31]. The idea is the following, if (i)
is assumed, then there is a well dened bounded operator
T =
_
I +

_
1
precisely when (ii) is satised.
In our analysis of the graph Laplacian
c
in (3.6) we shall need one
more:
Lemma 3.9 Let
c
be as in (3.6). Then for all v T,

xG
(0)
(
c
v) (x) = 0. (3.18)
In fact, when v is xed, the number of non-zero terms in (3.18) is nite.
Proof. The niteness claim follows from the assumptions on (G, c) we listed
in section 2.
A direct computation yields the result:

xG
(0)
(
c
v) (x) =

yx
c (xy) (v (x) v (y))
=

x
v (x)

yx
c (xy)

y
v (y)

xy
c (xy)
=

x
v (x) B
c
(x)

y
v (y)

xy
c (yx)
=

x
v (x) B
c
(x)

y
v (y) B
c
(y)
= 0.
12
4 Operator Theory
Once the operator theoretic tools are introduced, we show in section 5 below
that class of innite graph systems (G, c) where G is a graph and c is a
conductance function (the pair (G, c) satisfying the usual axioms as before),
have the spectrum of the associated Laplace operator
c
continuous. This
refers to the
2
space of G
(0)
, i.e., the Hilbert space is
2
(G
(0)
) where as
usual G
(0)
denotes the set of vertices.
It is important that G
(0)
is innite. Otherwise of course the spectrum is
just the nite set of zeros of the characteristic polynomial. See Example 6.6
below.
We give an operator theory/spectral theory analysis, with applications,
of a class of graph Laplacians; and we have been motivated by a pioneering
paper [Pow76] which in an exciting way applies graphs and resistor networks
to a problem in quantum statistical mechanics. In one of our results we
establish the essential selfadjointness of a large class of graph Laplacians on
graphs of innite networks. (A Hermitian symmetric operator with dense
domain in Hilbert space is said to be essentially selfadjoint if its closure is
selfadjoint, i.e., if the deciency indices are (0, 0). See Denition 4.1 below!
There are many benets from having the graph Laplacian essentially
selfadjoint.
Here is a partial list:
(a) We get the benet of having the spectral resolution for the selfadjoint
closure, also denoted for notational simplicity.
(b) We get a spectral representation realization of the operator , i.e.,
a unitarily equivalent form of in which an equivalent operator

may
occur in applications. See e.g., [Arv02], [PS72].
(c) We get a scale of Hilbert spaces, H
s
for s in R, dened from the graph
of the operator
s
where the fractional power
s
is dened by functional
calculus applied to the selfadjoint realization of . See [Jor04].
(d) Gives us a way of computing scales of resistance metrics on electrical
networks realized on innite graphs, extending tools available previously
only for nite graphs; see [BD49].
(e) The case s = 1/2 yields an exact representation of the energy Hilbert
space associated with a particular system (G, c) and the corresponding graph
Laplacian =
(G,c)
.
(f) Gives us a way of computing fractional Brownian motion on graphs,
allowing an analytic continuation in the parameter s, and with s = 1/2
corresponding to the standard Brownian motion; see e.g., [DJ07], [Jor06].
In the course of the proofs of our main results, we are making use of
13
tools from the theory of unbounded operators in Hilbert space: von Neu-
manns deciency indices, operator closure, graphs of operators, operator
domains, operator adjoints; and extensions of Hermitian operators with a
dense domain in a xed complex Hilbert space. Our favorite references
for this material include: [AC04], [Jr77], [Jr78], [JP00], [Nel69], [vN31],
[Sto51]. For analysis on innite graphs and on fractals, see e.g., [BHS05],
[CS07], [DJ06], [HKK02], [Hut81], [JP98], [JKS07], [Kig03], [BD49].
Denition 4.1 Let be a Hermitian linear operator with dense domain T
in a complex Hilbert space H. Set
T

: = v

dom(

) [

= iv

,
where i =

1. Then the two numbers n

: = dimT

are called the de-

ciency indices.
Von-Neumanns theorem states that the initial operator is essentially
selfadjoint on T if and only if n
+
= n

= 0. It has selfadjoint extensions

dened on a larger domain in H if and only if n
+
= n

.
The following two conditions on a Hermitian operator, (A) and (B),
individually imply equal deciency indices, i.e., n
+
= n

:
(A) For all v T, we have the estimate
v, v) 0,
i.e., is semibounded.
(B) There is an operator J : H H satisfying the following four condi-
tions:
(i) J (u +v) = Ju + Jv, for u, v H, C
(ii) Ju, Jv) = v, u) u, v H; (J is called a conjugation!)
(iii) J maps the subspace T into itself, and
Jv = Jv, v T.
(iv) J
2
= id; J is of period 2.
Remark 4.2 There are many examples (see the Appendix) where either
(A) or (B) is satised but where the operator is not essentially selfadjoint.
14
Both conditions (A) and (B) hold for a graph Laplacians
c
, and The-
orem 3.1 states that
c
is essentially selfadjoint.
For Riemannian manifolds with boundary, there is a close analogue of
the graph Laplacian
c
above; but it is known (see section 7) that these
continuous variants are typically not essentially selfadjoint.
Indeed the obstruction to essential selfadjointness in these cases captures
a physical essence of the metric geometry behind the use of Laplace operators
in Riemannian geometry.
5 The Energy Form
5.1 Operators
In section 3 we proved essential selfadjointness of the graph Laplacians
c
.
This refers to sequence space
2
, the Hilbert space of all square-summable
sequences indexed by the points in G
(0)
, and equipped with the usual
2
-
inner product.
This means that the axioms for
c
are such that boundary conditions at
innity in G are determined by computations on nite subsets of the vertices
in G. (In the Appendix, we will contrast this state of aairs with related
but dierent boundary conditions from quantum mechanics.) Recall that

c
is generally a densely dened Hermitian and unbounded operator. So
in principle there might be non-trivial obstructions to selfadjointness (other
than simply taking operator closure.) Recall (Denition 4.1) that a given
Hermitian operator with dense domain is essentially selfadjoint if and only
if the dimension of the each of the two defect eigenspaces is zero.
So that is why we look at the minus 1 eigenspaces for the adjoint
operator,

u = u.
For potential theoretic computations we need an additional Hilbert space,
the Energy Hilbert space H
E
(details below.) For example the voltage po-
tentials associated with a xed graph Laplacian are typically not in
2
(G
(0)
)
but rather in an associated Energy Hilbert space. Our Laplace operator
is formally Hermitian in both the Hilbert spaces
2
and H
E
(the energy
Hilbert space). We show that the Laplace operator is essentially selfad-
joint both in
2
and in H
E
. In both cases, we take for dense domain T the
linear subspace of all nitely supported functions G
(0)
R.
Our setting and results in this section are motivated by [Pow76] and
[BD49].
There are several distinctions between the two Hilbert spaces: For ex-
ample, the Dirac functions
x
[x G
(0)
form an orthonormal basis (ONB)
15
in
2
, but not in H
E
. The implication of this is that our graph Laplacians
have dierent matrix representations in the two Hilbert spaces. In speaking
of matrix representation for an operator in a Hilbert space, we will always
be referring to a chosen ONB.
We shall need the operator in both guises. One reason for this is that
for innite graphs, typically the potential function v solving v =
x

y
,
for pairs of vertices will not be in
2
, but nonetheless v will have nite
energy, i.e., c(v) < , meaning that the energy form applied to v is nite.
Caution: The sequence v might not be in the
2
-space. Specics in Example
5.2 below!
When we study the Laplace operator , our questions concern its spec-
trum, and its spectral resolution. The spectrum will be contained in the
half-line [0, ), but (as we show in examples) it can be unbounded, and it
can have continuous parts mixed in with discrete parts. In case the conduc-
tance function is very unbounded, as an operator in
2
, it may be necessary
to pass to a proper operator extension in an enlarged Hilbert space to get a
dierent selfadjoint realization of .
These operator issues only enter in case is unbounded. The unbound-
edness of the operator is tied in closely with unboundedness of the con-
ductance function c which is used. Recall =
c
depends on the choice of
c. If c is unbounded at innity (on the set G
(1)
of edges), then the resis-
tors tend to zero at distant edges. Intuitively, this means that the current
escapes to innity, and we make this precise in the language of operator
theory. Our general setup here will be as in [Pow76].
We will need a second Hilbert space, the energy Hilbert space H
E
. Here
the inner product is the energy quadratic form. Since the energy form
evaluated on a function v is dened in terms of the square of dierences
v(x) v(y), it follows that the elements in H
E
are really sequences modulo
constants.
Let G =
_
G
(0)
, G
(1)
_
be a graph satisfying the axioms in section 3, with
vertices G
(0)
and edges G
(1)
. Let
c : G
(1)
R
+
(5.1)
be a xed function (called conductance.) If e = (xy) G
(1)
, we say that x
y, and the function c must satisfy c (xy) = c (yx), symmetry. In particular,
for a pair of vertices x, y, c (xy) is only dened if x y, i.e., if (xy) G
(1)
.
For every x G
(0)
, we assume that
nbh(x) =
_
y G
(0)
[y x
_
(5.2)
16
is nite, and that x , nbh(x).
Following eq. (3.9), we study functions v : G
(0)
C for which
c
c
(v) =

all x,y
s.t. xy
c (xy) [v (x) v (y)[
2
< . (5.3)
Clearly we must work with functions on G
(0)
modulo constants. Setting
c
c
(u, v) : =

all x,y
xy
c (xy)
_
u(x) u(y)
_
(v (x) v (y)) , (5.4)
we get
[c
c
(u, v)[
2
c
c
(u) c
c
(v) , u, v T, (5.5)
by Schwarz inequality.
Setting
u, v)
E
: = c
c
(u, v) (5.6)
we get an inner-product, and an associated Hilbert space H
E
of all functions
v for which (5.3) holds.
The triangle inequality
c
c
(u +v)
1
2
c
c
(u)
1
2
+c
c
(v)
1
2
(5.7)
holds; or equivalently
|u +v|
E
|u|
E
+|v|
E
, u, v H
E
.
In the next result we give a characterization of the Hilbert space H
E
directly
in terms of the selfadjoint operator
c
from section 3.
Recall
c
is the closure of the operator
c
with dense domain T in

2
_
G
(0)
_
. It will be convenient to write simply
c
for the closure. Since it is
selfadjoint, we have a Borel functional calculus; i.e., if f is a Borel function
on R, and if P () is a projection valued measure for
c
, then

c
=
_

0
P (d) , (5.8)
and
f (
c
) : =
_
f () P (d) . (5.9)
For the corresponding (dense) domains, we have
dom(
c
) =
_
v
2
_
G
(0)
_
[
_

0
[[
2
|P (d) v|
2
<
_
, (5.10)
17
and
dom(f (
c
)) =
_
v
2
_
G
(0)
_
[
_

0
[f ()[
2
|P (d) v|
2
<
_
. (5.11)
Theorem 5.1 Let G =
_
G
(0)
, G
(1)
_
and c : G
(1)
R
+
be as described, and
let H
E
be the energy Hilbert space. Let
c
be the selfadjoint graph Laplacian
in
2
_
G
(0)
_
from Section 3.
(a) Then
dom
_

1/2
c
_
= H
E

2
_
G
(0)
_
. (5.12)
(b) In general the right hand side is a proper subspace of
2
_
G
(0)
_
.
Proof. We proved in (3.11), Lemma 3.2, that
c
c
(u, v) = 2u,
c
v) (5.13)
for all functions u, v on G
(0)
for which the two sides in (5.13) are nite.
If u = v, then the expression on the RHS in (5.13) is 2
_
_
_
1/2
c
v
_
_
_
2
i v
dom
_

1/2
c
_
. This follows from (5.11) applied to
f () : =

, [0, ).
Since dom(
c
) dom
_

1/2
c
_

2
_
G
(0)
_
the desired conclusion (5.12)
holds.
To see that dom
_

1/2
c
_
may be a proper subspace of
H
E
= v[c
c
(v) < , (5.14)
consider the following example (G, c) built on the simplest innite graph
G
(0)
: = Z.
Example 5.2
G
(0)
= Z,
G
(1)
= (n, n 1) [n Z , and
c : G
(1)
R
+
, c 1 on G
(1)
.
The corresponding graph Laplacian is
(v) (n) = 2v (n) v (n 1) v (n + 1) , n Z. (5.15)
If k Z
+
is given, we claim that there is a unique function v H
E
solving
v =
0

k
. (5.16)
18
Existence: Set
v (n) : =
_
_
_
0 if n 0
n if 0 < n k
k if k n.
(5.17)
A substitution shows that the function v in (5.17) satises (5.16).
Uniqueness: Let w H
E
be a solution to (5.16). Then
c
c
(u, v w) = 0, u T.
Since G =
_
G
(0)
, G
(1)
_
is connected, T is dense in H
E
; and so the dierence
v w must be a constant function. But the Hilbert space H
E
is dened by
moding out with the constants. Hence, v = w in H
E
.
The following three facts follow directly from (5.17):
(i) v is non-constant;
(ii) v ,
2
(Z); and
(iii) c
c
(v) < .
In fact, an application of (5.13) yields
c
c
(v) = 2k. (5.18)
Proof of (5.18):
c
c
(v) = 2v, v) by (5.13)
= 2v,
0

k
)
= 2 (v (0) v (k))
= 2k by (5.17).
Theorem 5.3 Let G =
_
G
(0)
, G
(1)
_
and c : G
(1)
R
+
be a graph system
as in the previous theorem, and in section 3; i.e., we assume that the pair
(G, c) satises the axioms listed there. Let
c
be the corresponding graph
Laplacian with a choice c for conductance.
Let , G
(0)
be a xed pair of vertices. Then there is a unique function
v H
E
, i.e., c
c
(v) < satisfying

c
v =

. (5.19)
The solution to (5.19) is called a voltage potential. Moreover,
c
c
(v) = 2 (v () v ()) . (5.20)
19
Proof. The argument for uniqueness is the same as in the previous proof.
To prove existence, we will appeal to Riesz theorem for the energy
Hilbert space H
E
.
Hence, we must show that there is a nite constant K such that
[u() u()[ Kc
c
(u)
1
2
for all u T. (5.21)
Motivated by Ohms law, we set (e) : = c (e)
1
, e G
(1)
. By the
assumptions in section 3, we may pick a nite subset x
0
, x
1
, x
2
, . . . , x
n
in
G
(0)
such
_
x
0
= , x
n
= , and
e
i
= (x
i
x
i+1
) G
(1)
, i = 0, 1, . . . , n 1.
(5.22)
Then
[u() u()[
n1

i=0
[u(x
i
) u(x
i+1
)[

_
n1

i=0
(e
i
)
_
1
2
_
n1

i=0
c (e
i
) [u(x
i
) u(x
i+1
)[
2
_
1
2
(by Schwarz)

_
n1

i=1
(e
i
)
_
1
2
c
c
(u)
1
2
.
To get a nite constant K in (5.21), we may take the inmum over all
paths subject to conditions (5.22), connecting to .
An application of Riesz lemma to H
E
yields a unique v H
E
such that
for all u T, we have the following identity:
u() u() =
1
2
c
c
(v, u)
=
c
v, u) (by (5.13)).
Using again density of T in H
E
, we get the desired conclusion

c
v =

. (5.23)
Corollary 5.4 Let (G, c) satisfy the conditions in the theorem. Let ,
G
(0)
, and let v H
E
be the solution (potential ) to

c
v =

.
20
Then
c
c
(v) 2 inf
(e
i
)
n1

i=0
(e
i
) (5.24)
where e
0
, e
1
, . . . , e
n1
G
(1)
is a system of edges connecting to , i.e.,
satisfying the conditions listed in (5.22).
Proof. This follows from the previous proof combined with the fact that
sup
Ec(u)=1
[c
c
(u, v)[
2
= c
c
(v) . (5.25)
5.2 A matrix representation
While
c
may be understood as an operator, it is also an matrix.
Since the set nbh(x) G
(0)
is nite for all x G
(0)
,
c
is a banded matrix.
To see this, note that when x G
(0)
is xed, the summation
(
c
v) (x) =

yx
c (xy) (v (x) v (y)) (5.26)
is nite for all functions v : G
(0)
C.
Since G is assumed connected, the only bounded solutions v to the equa-
tion

c
v = 0 (5.27)
are the constants.
Solutions v to (5.27) are called harmonic, or c-harmonic.
There are examples of systems (G, c) which are connected and have un-
bounded non-constant harmonic functions, e.g., models with G
(0)
= Z
3
, or
tree-models.
In the general case, introducing
B
c
(x) : =

yx
c (xy) , x G
(0)
; (5.28)
we see that (5.26) takes the following form
(
c
v) (x) = B
c
(x) v (x)

yx
c (xy) v (y) . (5.29)
21
Hence eq. (5.27) may be rewritten as
v (x) =
1
B
c
(x)

yx
c (xy) v (y) . (5.30)
It follows that harmonic functions on G
(0)
satisfy a mean value property. At
every x G
(0)
formula (5.30) expresses v (x) as a convex combination of its
values on the set nbh(x).
In matrix language, x B
c
(x) represents the diagonal matrix-entries;
and c (xy) the o-diagonal entries. Since y G
(0)
[c (xy) ,= 0 is nite, we
say that the matrix for
c
is banded. It is clear that products of banded
matrices are again banded; and in particular that the summations involved
in matrix-products of banded matrices are all nite. Hence, each of the
operators
c
,
2
c
,
3
c
, . . ., is banded. Since by Theorem 3.1,
c
is selfadjoint
as an operator in
2
_
G
(0)
_
, the fractional power
1/2
c
is well dened by the
Spectral Theorem. The matrix-entries of
1/2
c
are the numbers

x
,
1/2
c

y
)

2 =
_

1/2
c

y
_
(x) , x, y G
(0)
. (5.31)
It can be checked that if G is innite, the matrix for
1/2
c
is typically not
banded. The same conclusion applies to
s
c
when s RN.
5.3 Example 5.2 revisited
The system (G, c) in Example 5.2 does not have non-constant harmonic
functions. This can be seen from the representation of (in Ex. 5.2) as a
Z Z double innite matrix, i.e.,
(v) (n) = 2v (n) v (n 1) v (n + 1)
= v (n) v (n 1) +v (n) v (n + 1)
=

mn
v (n) v (m) , n Z.
In matrix form, from Example 5.2 is as follows:
22
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
.
.
.
.
.
.
.
.
.
.
.
.
1 2 1 0 0 0 0 0 0
0 1 2 1 0 0 0 0 0
0 0 1 2 1 0 0 0 0
0 0 0 1 2 1 0 0 0
0 0 0 0 1 2 1 0 0
0 0 0 0 0 1 2 1 0
0 0 0 0 0 0 1 2 1
.
.
.
.
.
.
.
.
.
.
.
.
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
Using Fourier series
f (x) =

nZ
v (n) e
inx
L
2
(, ) ; (5.32)

nZ
[v (n)[
2
=
1
2
_

[f (x)[
2
dx; (5.33)
we arrive at the representation
_

f
_
(x) = 2 (1 cos x) f (x) (5.34)
= 4 sin
2
_
x
2
_
f (x) ,
proving that has Lebesgue spectrum, and
spec

2
() = spec
L
2
_

_
= [0, 4] . (5.35)
5.4 Banded Matrices (A Preview)
It is immediate from the matrix representation for
c
in Example 5.2 that
it has a banded form. We will take up banded innite matrices in detail in
section 8 below.
Since
c
is selfadjoint, its square-root
1/2
c
is a well dened operator.
However its matrix representation is typically not banded; see (5.31). For

1/2
c
in Ex. 5.2, one can check that the (m, n)-matrix entries are
_

1
2
c
_
m,n

1
4 (n m)
2
+ 1
.
23
5.5 Extended Hilbert Spaces
To understand solutions v to operator equations like

c
v =

as in (5.23), potential functions it is convenient to extend the Hilbert space

2
_
G
(0)
_
. Indeed we saw in Example 5.2 that the solutions v to equations
like (5.23) are typically not in
2
_
G
(0)
_
.
Denition 5.5 The space H
c
(s).
A function v : G
(0)
C is said to belong to the space H
c
(s) if there is
a nite constant K = K (s) such that the following estimate holds:

xG
(0)
v (x) (
s
c
u) (x)

2
K (s)

xG
(0)
[u(x)[
2
for all u T. (5.36)
If (5.36) holds, then by Riesz, there is a unique w
2
_
G
(0)
_
such that

xG
(0)
v (x) (
s
c
u) (x) = w, u)

2 for all u T (
2
); (5.37)
and we set
|v|
H
(s)
c
: = |w|

2
(G
0
)
. (5.38)
By abuse of notation, we will write
s
c
v = w when v H
c
(s).
If two functions v
i
for i = 1, 2 are in H
c
(s), and if
s
c
v
i
= w
i

2
_
G
(0)
_
,
we set
v
1
, v
2
)
Hc(s)
: = w
1
, w
2
)

2 (5.39)
=

xG
(0)
w
1
(x)w
2
(x) .
Remark 5.6 We proved in section 3 that

xG
(0)
(
c
u) (x) = 0 for u T.
Hence the constant function v
1
(x) 1 on G
(0)
is in H
c
(1), and |v
1
|
Hc(1)
=
0. Hence in considering the extension spaces, we shall work modulo the
constant functions on G
(0)
.
24
Theorem 5.7 For every s R, the space H
c
(s) is a Hilbert space when
equipped with the inner product (5.39), and the norm (5.38).
Proof. The idea in the proof follows closely the construction of Sobolev
spaces, by analogy to the continuous case. The key step in the verication
of completeness of H
c
(s) is the essential selfadjointness of
c
as an operator
in
2
_
G
(0)
_
. As before, we use the same notation
c
for the closure of
c
,
dened initially only on the subspace T in
2
_
G
(0)
_
. Formulas (5.9)(5.11)
above now allow us to dene the selfadjoint operator
s
c
; and this operator
is closed in the sense that its graph is closed in
2
_
G
(0)
_

2
_
G
(0)
_
. The
completeness of H
c
(s) now follows from this, and an application of Riesz;
see the estimate (5.37).
Corollary 5.8 Let (G, c) be an innite graph, and let c : G
(1)
R
+
be a
conductance function satisfying the axioms above. Let , G
(0)
, and let
v : G
(0)
C be a solution to
c
v =

; i.e., to (5.23).
Assume v H
E
. Then
v H
c
_
1
2
_
H
c
(1) ; (5.40)
and we have
|v|
2
1/2
=
1
2
c
c
(v) , (5.41)
and
|v|
2
1
= 2. (5.42)
Proof. To prove (5.40), we must check the a priori estimate (5.36) for
s = 1/2, and s = 1:
Verication of (5.36) for s = 1/2
Let v satisfy the stated conditions, and let u T. Then

xG
(0)
v (x)
_

1
2
c
u
_
(x)

xG
(0)
v (x)
c

1
2
c
u(x)

=
1
2

c
c
_
v,

1
2
c
u
_

1
2
c
c
(v)
1
2
c
c
_

1
2
c
u
_1
2
(Schwarz)
=
1

2
c
c
(v)
1
2
|u|

2
(G
(0)
)
,
25
where we used the identity
c
c
_

1
2
c
u
_
= 2 |u|
2

2
(G
(0)
)
= 2

xG
(0)
[u(x)[
2
valid for u T.
Verication of (5.36) for s = 1
With v and u as before, we must estimate the summation:

xG
(0)
v (x) (
c
u) (x)

xG
(0)
(
c
v) (x) u(x)

xG
(0)
(

(x)

(x)) u(x)

= [u() u()[
2 |u|

2
(G
(0)
)
, u T.
Once (5.40) has been checked, the exact formulas (5.41) and (5.42) follow:
Firstly,
|v|
2
1
2
=
_
_
_
_

1
2
c
v
_
_
_
_
2

2
=
1
2
c
v,
1
2
c
v)
=
1
2
c
c
(v) ;
and secondly
|v|
2
1
= |
c
v|
2

2
= |

|
2

2
= 2.
26
Remark 5.9 In conclusion (5.40) in Corollary 5.8 is not best possible. In
fact, the optimal range of the fraction s for which the potentials v are in
H
c
(s) may be computed explicitly in Example 5.2 and related examples. De-
tails in the next subsection.
In Example 5.2, G
(0)
= Z, G
(1)
= (n, n 1) [n Z, and c 1. Let
k N. The graph Laplacian is given in formula (5.15).
Let v be the unique solution to the potential equation
v =
0

k
. (5.43)
Then v H(s) if and only if s > 1/4.
Proof. Setting
v (z) =

nZ
v
n
z
n
, and z = e
ix
, (5.44)
we get
v (z) =
z
_
z
k
1
_
(z 1)
2
; (5.45)
and therefore
[v (x)[ =

sin
_
kx
2
_

sin
2
(x/2)
. (5.46)
Since, in the spectral representation, the graph Laplacian is multiplication
by 4 sin
2
(x/2), the question: For what exponents s is
v H(s) ? (5.47)
is decided by the asymptotics near x = 0 of the function (
s
v) (x). Using
(5.46), we see that
s
v is in L
2
(, ) if and only if x
2s1
L
2
near x = 0;
and this hold if and only if
s >
1
4
(5.48)
as claimed.
5.6 Lattice Models
Example 5.10 We proved that potential functions are often not in
2
_
G
(0)
_
,
but in general the problem is more subtle.
The setting is a follows:
G =
_
G
(0)
, G
(1)
_
a given graph;
27
c : G
(1)
R
+
a given conductance function;

c
= the corresponding graph Laplacian;
, G
(0)
a xed pair of vertices, ,= .
With this, we say that a function v : G
(0)
R is a potential if

c
v =

. (5.49)
In the next result we show that lattice models Z
D
with D > 2 have
2
potentials.
5.7 Preliminaries
By Z
D
we mean the rank D-lattice of vertex points n = (n
1
, n
2
, . . . , n
D
),
n
i
Z, i = 1, 2, . . . , D. Every point n Z
D
has 2D distinct nearest
neighbors
(n
1
, . . . , n
i
1, n
i+1
, . . . , n
D
) , (5.50)
so nbh(n) consists of these 2D points; and G
(1)
is the corresponding set of
edges. In the discussion below, we pick the constant conductance c 1, i.e.,
a system of unit-resistors arranged in nearest-neighbor congurations. See
Fig. 1 for an illustration of the simplest lattice conguration, D = 1, 2, and
3.
Fig. 1a: D = 1
28
Fig. 1b: D = 2
Fig. 1c: D = 3
Fig. 1. Lattice congurations in the rank-D lattices Z
D
with nearest-neighbor resistors.
Proposition 5.11 The potential functions v, i.e., solutions to (5.49) with
c 1 are in
2
(Z
D
) if D > 2.
Proof. Recall that the D-torus T
D
is the compact dual of the rank-D lattice.
Pick coordinates in T
D
s.t. x = (x
1
, . . . , x
D
), < x
i
, i = 1, 2, . . . , D.
29
Then, by Parseval,

2
_
Z
D
_
L
2
_
T
D
_
.
By the argument from Example 5.2, we see that has the following spectral
representation in L
2
((, ] (, ]
. .
D times
)
(v) (x) = 4
D

k=1
sin
2
_
x
k
2
_
v (x) . (5.51)
Introducing spherical coordinates in R
D
, we get the representation
dx =
D1
dS
1
(5.52)
where
: =
_
D

k=1
x
2
k
_
1
2
,
and whence dS
1
denotes the rotationally invariant measure on the sphere in
R
D
.
The question of deciding when the solution v to (5.49) is in
2
_
Z
D
_
can
be better understood in the spectral representation v (x) for x = (x
1
, . . . , x
D
)
close to 0, i.e., (x) 0.
Using (5.51)-(5.52), we see that the potential function v is in
2
if D > 2.
More generally, the argument from Example 5.2 proves that in Z
D
, the
potential function v is in H(s) if s >
2D
4
.
The next results illustrate new issues entering the analysis of Z
D
-graphs
when D > 1, compared to the D = 1 case.
Corollary 5.12 For the case D = 3 in the lattice model in Example 5.10,
consider k = (k
1
, k
2
, k
3
) Z
3
(0) xed, and let
v
k
: Z
3
R
the solution to the potential equation
v
k
=
0

k
. (5.53)
Then
lim
n
v
k
(n) = 0. (5.54)
30
Proof. Our notation is as follows: n = (n
1
, n
2
, n
3
) Z
3
and by n
we mean:
[n[ =
_
n
2
1
+n
2
2
+n
2
3
.
Moreover,

k
: Z
3
R
is the usual Dirac mass

k
(n) =
k
1
,n
1

k
i
,n
i

k
3
,n
3
. (5.55)
We proved in Proposition 5.11 (D = 3) that

nZ
3
[v
k
(n)[
2
_
= |v
k
|
2
_
< ; (5.56)
and so in particular, the conclusion (5.54) must hold.
Our next example illustrates that the potential equation (5.53) has un-
bounded solutions in case D 3. This will also provide concrete cases
of unbounded harmonic functions, i.e., functions w : Z
D
R for which
w = 0.
To aid the construction, we include the following lemma which is about
the general case of systems (G, c) as analyzed in sections 3-4 above.
Lemma 5.13 Let G = (G
(0)
, G
(1)
), and c : G
(1)
R
+
, be a graph system
as described in Theorem 5.1, and let
c
be the graph Laplacian.
Let , G
(0)
be given, ,= . Then there is a 1-1 correspondence
between two classes of functions v : G
(0)
R, and functions I : G
(1)
R,
where the two classes are given as follows:
Class 1.

c
v =

(5.57)
Class 2. I : G
(1)
R satisfying (Kirchos Laws):
(a)

yx
I (x, y) = (

) (x) , x G
(0)
; and
(b)

i
c (x
i
x
i+1
)
1
I (x
i
x
i+1
) = 0 for all x
0
, x
1
, x
2
, . . . , x
n
G
(0)
sub-
ject to x
0
= x
n
, and x
i
x
i+1
, i.e., all closed loops in G
(0)
.
The connection between the two classes is given by the following formula:
c (xy)
1
I (xy) = v (x) v (y) , (xy) G
(1)
. (5.58)
31
The function v is determined from I uniquely, up to a constant, when I
is known to satisfy (a)-(b). Moreover,

eG
(1)
(I (e))
2
c (e)
= c
c
(v) . (5.59)
Proof. Left to the reader. The arguments are included in the proof of
Theorem 5.1.
Example 5.14 The T = 2 lattice model; i.e., G
(0)
= Z
2
, edges given by
nearest neighbors as in Fig. 1b; and c 1.
We consider the equation (5.57) for = (0, 0) and = (1, 1). The two
dierent solutions v to (5.57) will be presented in the form of Class 2 in
Lemma 5.13, i.e., in terms of current functions dened on the edges in G.
First recall that the Laplace operator in the Z
2
-model is
(v) (m, n) = 4v (m, n) v (m1, n) v (m+ 1, n)
v (m, n 1) v (m, n + 1) , (m, n) Z
2
.
Eq. (5.56) then takes the form
v =
(0,0)

(1,1)
. (5.60)
We now describe the two current functions I which correspond to the
two solutions to (5.60).
32


> > > > < < <

> > > > < < <

> > > > < < <

< < < > < < <

< < < > (1,1) < < <

< < < (0,0) > > > >

< < < > > > >

< < < > > > >

< < < > > > >


Fig. 2. The function I for the rst solution v to (5.60).
Flow design for the current function I. The symbols > indicate
arrows in the direction of the current ow. An arrow points in the
direction of voltage drop.
33
And now the (dierent) function I for the second solution to (5.60):
.
.
.
.
.
.
0 0 0 0 0 0
0

1
2
0 0 0
1
2

0
0 0 0 0 <
1
2
<
1
2
0

1
2
0 0
1
2

1
2

0
0 0 0 <
1
2
<
1
2
0
0

1
2
0
1
2

1
2

0 0
0 <
1
4
<
1
2
<
1
2
0 0
0

1
4
1
4

1
2

0 0 0
0
1
4
> (1,1) <
1
4
0 0 0
0
1
2

1
4
1
4

0 0 0
0 (0,0) >
1
2
1
4
>
1
2
>
1
2
>
1
2
>
1
2
>
0 0 0 0 0 0 0
0 0 0 0 0 0
.
.
.
.
.
.
Fig. 3.
Flow design for the current function I. The symbols > indicate
arrows in the direction of the current ow. An arrow points in the
direction of voltage drop.
5.8 The Resistance Metric
Let G = (G
(0)
, G
(1)
) be a graph satisfying the axioms from section 2, and
let
c : G
(1)
R
+
be a conductance function. Let c
c
() be the corresponding energy form, and
let
c
be the graph Laplacian.
Pick a denite point 0 in the vertex set G
(0)
. Now for every x G
(0)
let
v
x
H
Ec
be the solution to

c
v
x
=
0

x
. (5.61)
34
Set
dist
c
(x, y) : = c
c
(v
x
v
y
)
1
2
(5.62)
= |v
x
v
y
|
Ec
for x, y G
(0)
. We say that x, y dist
c
(x, y) is the resistance metric on
G
(0)
. It is immediate from (5.61) that it satises the triangle inequality.
Proposition 5.15 The following formula holds for the resistance metric:
dist
c
(x, y) =

2 (v
x
(y) +v
y
(x) v
x
(x) v
y
(y))
1
2
.
Proof. In view of (5.62), it is enough to compute c
c
(v
x
v
y
) for pairs of
points x, y.
Let x, y G
(0)
be given, and let v
x
, v
y
be the potential functions from
(5.61). Then
c
c
(v
x
v
y
) = 2
c
(v
x
v
y
) , v
x
v
y
)

2
= 2
0

x
(
0

y
) , v
x
v
y
)

2
= 2 ((v
x
v
y
) (y) (v
x
v
y
) (x))
= 2 (v
x
(y) +v
y
(x) v
x
(x) v
y
(y)) .
Example 5.16 (See also Example 3.7) Let
c
be given by the following
matrix:
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
1 1 0 0 0
1 5 2
2
0
0 2
2
2
2
+ 3
2
3
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
n
2
n
2
+ (n + 1)
2
(n + 1)
2
.
.
.
.
.
.
0
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
.
So G
(0)
= N
0
, G
(1)
= (0, 1) , , (n 1, n) , (n, n + 1) , , and c (n, n + 1) =
(n + 1)
2
. The rst vertex has one neighbor, and the later two.
35
The potential equation ( 5.62) may be solved by inspection, and we get
the following formula for the resistance metre dist
c
in Proposition 5.15: If
m < n ( in N
0
) then
dist
c
(m, n)
_
1
(m+ 1)
2
+
1
(m+ 2)
2
+ +
1
n
2
_1
2
.
Since

k=1
1
k
2
=

2
6
, we conclude that
_
G
(0)
, dist
c
_
is a bounded metric
space.
Further, the resistance is bounded at innity; or equivalently the voltage
drop is very slow at innity for the current ow induced by the experiment
which inserts 1 amp at a particular place in G
(0)
= N
0
.
The reason is that the conductance is very unbounded, or equivalently
or more precisely, the resistance is O
_
n
2
_
for this particular (G, c) system.
Some conclusions: The nite-energy solution v to (5.60) is the function
v : Z
2
R, beginning with the values 0, 1/2, and 1 as follows: In
Figs. 23 we list the values of v on the points in the interior square in
G
(0)
_
= Z
2
_
. The three values are prescribed in the centered square; and
they then propagate into the quarter planes, with the value 1/2 in the
NW and the SE quarter planes.
6 Finite Dimensional Approximation
6.1 Systems of Graphs
Let G = (G
(0)
, G
(1)
) be an innite graph satisfying the axioms from section
2. In particular, we assume for every x in G
(0)
that x itself is excluded from
nbh(x); i.e., no x in G
(0)
can be connected to itself with a single edge. Let c
any conductance function dened on G
(1)
and satisfying our usual axioms.
In section 3 we showed that the corresponding Laplace operator =
c
is automatically essentially selfadjoint. By this we mean that when is
initially dened on the dense subspace T (of all the real valued functions
on G
(0)
with nite support) in the Hilbert space H :=
2
(G
(0)
), then the
closure of the operator is selfadjoint in H, and so in particular it has a
unique spectral resolution, determined by a projection valued measure on
the Borel subsets the innite half-line R
+
.
In contrast, we note (Example 7.1) that the corresponding Laplace op-
erator in the continuous case is not essential selfadjoint.
This can be illustrated with = (d/dx)
2
on the domain T of consist-
ing of all C
2
-functions on the innite half-line R
+
which vanish with their
36
derivatives at the end points. Then the Hilbert space is L
2
(R
+
).
So this is an instance where the analogy between the continuous case
and the discrete case breaks down.
In the study of innite graphs G =
_
G
(0)
, G
(1)
_
and the corresponding
Laplacians, it is useful to truncate and consider rst a nested system of nite
graphs G
N
; then compute in the nite case and, in the end, take the limit as
N . Our approximation results here continue work started in [Jr77],
[Jr78].
Denition 6.1 In this section we prove specic results showing that the
procedure works. While there are several candidates for designing the nite
approximating graphs G
N
= (G
(0)
N
, G
(1)
N
), we will concentrate here on the
simplest: Starting with an innite G = (G
(0)
, G
(1)
), pick nite subsets of
vertices as follows:
G
(0)
1
G
(0)
2
G
(0)
3
G
(0)
N
G
(0)
(6.1)
such that

_
N=1
G
(0)
N
= G
(0)
. (6.2)
Set H : =
2
(G
(0)
), and H
N
=
2
(G
(0)
). Then the projection P
N
of H
onto H onto H
N
is multiplication by the indicator function
G
(0)
N
; and the
projection onto the complement HH
N
is multiplication with
(G
(0)
N
)
c
where
(G
(0)
N
)
c
= G
(0)
G
(0)
N
is the complement of G
(0)
N
.
The edges G
(1)
N
in G
N
are simple the edges in G, for which the vertices
lie in G
(0)
N
; i.e., if x, y G
(0)
, then:
(xy) G
(1)
N
(xy) G
(1)
and x, y G
(0)
N
. (6.3)
If a system (G
N
)
NN
of graphs is given as in (6.1)-(6.3), and if c :
G
(1)
R
+
is a conductance function; we denote by c
N
the restriction of c
to G
(1)
N
.
Lemma 6.2 Let G = (G
(0)
, G
(1)
) and c : G
(1)
R
+
be given as above. Let
G
N
be a system of graphs determined subject to conditions (6.1)-(6.3).
Let
N
be the graph Laplacian associated to (G
N
, c
N
). Then
P
N
P
N
=
N
, for N N. (6.4)
37
Proof. For v T = nite linear combinations of
x
[x G
(0)
, we have
(P
N
P
N
v) (x) =
G
N
(x)

yx
c (xy) ((
G
N
v) (x) (
G
N
v) (y))
=

yx in G
N
c
N
(xy) (v (x) v (y))
= (
N
v) (x) ;
proving the formula (6.4).
Lemma 6.3 Let G = (G
(0)
, G
(1)
), and c : G
(1)
R
+
, be as in Lemma 6.2
and Denition 6.1. Then for all v T and x G
(0)
, we have the following
formula for the dierence operator
N
, N = 1, 2, . . . :
(v) (x) (
N
v) (x) =
G
c
N
(x)

yx
yG
N
c (xy) v (y) . (6.5)
In other words, the contribution to
N
comes from the boundary of
G
N
= the edges e G
(1)
s.t. one vertex in e is in G
(0)
N
and the other in the
complement.
Proof. Using the previous lemma, we get
(v) (x) (
N
v) (x) =

yx
(c (xy) c
N
(xy)) (v (x) v (y))
=
G
c
N
(x)

yx
yG
N
c (xy) v (y) .
Denition 6.4 Let G = (G
(0)
, G
(1)
), and c : G
(1)
R
+
be given as in
Theorem 5.1; and denote by =
c
the corresponding selfadjoint graph
Laplacian. Setting
S (t) : =
_

0
e
t
P (d) (see (5.9)-(5.10)) (6.6)
= e
t
, t R
+
;
we see that t S (t) is a contractive semigroup of selfadjoint operators in

2
(G
(0)
); in particular,
S (s +t) = S (s) S (t) , s, t R
+
and (6.7)
S (0) = I

2.
38
The semigroup consists of bounded operators while the innitesimal gen-
erator =
c
is typically unbounded, albeit with dense domain in
2
(G
(0)
).
Moreover, the semigroup helps us identify dynamics as innite graphs of
resistors.
Returning to approximations, as in Denition 6.1, we now get a sequence
of Laplacians
N
, N = 1, 2, . . ., and a corresponding sequence of dynamical
semigroups, S
N
(t) : = e
t
N
, N = 1, 2, . . ..
Let N be xed, and let G
N
be the boundary of G
N
(Denition 6.1).
Then the nite matrix
T
N
: = (c (xy))
x,yG
N
(6.8)
is positive, and has a Perron-Frobenius eigenvalue
N
=
N
(PF) = the
spectral radius of T
N
.
Theorem 6.5 Let (G, c) be a graph/conductance system, and let (G
N
)
NN
ascending system of graphs such that (6.2) is satised. Let S (t) , and S
N
(t),
N = 1, 2, . . ., be the corresponding semigroups of bounded operators.
Then for all v
2
(G
(0)
), we have the following estimate:
|S (t) v S
N
(t) v|

2
N
(PF) t |v|

2 , t R
+
, N = 1, 2, . . . . (6.9)
Proof. With the use of ( 5.8)-( 5.9), we get the integral formula:
e
t
N
e
t
=
_
t
0
e
(ts)
(
N
) e
s
N
ds. (6.10)
Since the operators on both sides in (6.10) are bounded, it is enough to
verify the estimate (6.9) for vectors v in the dense domain T.
Using new Lemma 6.3, we get the following estimates on the respective

2
-norms:
|S (t) v S
N
(t) v|

2
_
t
0
|(
N
) S
N
(s) v|

2 (by (6.10))

N
(PF)
_
t
0
|S
N
(s) v| ds (by Lemma 6.3 and (6.8))

N
(PF) |v|

2
_
t
0
ds
=
N
(PF) t |v|

2 ,
which is the desired conclusion.
39
6.2 Periodic boundary conditions
Example 6.6 We now compare Example 5.2 with an associated family of
nite graphs G
N
where N N. Let Z
N
= Z/NZ 0, 1, 2, . . . , N 1 be
the cyclic group of order N. Introduce nearest neighbors as in Example 5.2
(the Z-case) with the modication for G
N
given by 0 (N 1), in other
words that there is an edge connecting 0 to N 1.
It follows that the graph Laplacian
N
for G
N
is the given by the nite
matrix
_

_
2 1 0 0 0 0 1
1 2 1 0 0 0 0
0 1 2 1 0 0 0
.
.
.
.
.
.
0 0 2 1 0
0 0 1 2 1
1 0 0 1 2
_

_
.
The spectrum of
N
is as follows:
spec (
N
) =
_
2
_
1 cos
_
2k
N
__
[k = 0, 1, . . . , N 1
_
(6.11)
=
_
4 sin
2
_
k
N
_
[k = 0, 1, . . . , N 1
_
.
Comparing with (5.34)-(5.35), we see that the spectra converge in a natural
sense; with the innite model in Ex 5.2 being a limit of N-periodic boundary
condition as N .
40
Fig. 4. The vertices and edges in G
N
for N = 3, 4 and 5.
The spectrum of the cyclic graph Laplacian
N
of the graphs G
N
, N =
3, 4 and 6 (in Fig. 3) may have multiplicity; see (6.11). This holds in fact
for all values of N. Specically,
spec (
3
) = 0, 3 with = 3 having multiplicity 2.
spec (
4
) = 0, 2, 4 with = 2 having multiplicity 2.
spec (
6
) = 0, 1, 3, 4 now with = 1 and = 3 each having multi-
plicity 2.
Hence for D = 1, we get the following distinction between the spectral
theory of the cyclic graph Laplacians
N
for N < on the one hand and
in Example 5.2 on the other: The commutant of is an abelian algebra
of operators in
2
(Z), while the operators in
2
(Z
N
) which commute with

N
form a non-abelian algebra.
Proposition 6.7 (Cyclic graphs) Let N N, N 3; and let G
N
be the
corresponding cyclic graph with graph Laplacian
N
; i.e., with
G
(0)
N
= 0, 1, 2, . . . , N 1 .
41
Then the voltage potential v
2
(Z
N
) solving
N
v =
0

1
is
_

_
v
0
= 0
v
1
=
N1
N
v
2
=
N2
N
.
.
.
v
N2
=
2
N
v
N1
=
1
N
.
Proof. A direct computation; see also Fig. 4, and eq. (6.10).
7 Boundary Conditions
In the study of innite graphs G, boundary conditions play an important
role; for example if a current escapes to innity in nite time, conditions
must then be assigned at innity.
One way to do this is to rst do computations in a system of nite graphs
G
N
which exhausts the given graph G in a suitable way. Do computations on
each nite subgraph G
N
of the xed innite graph G, and then take the limit
as N tends to innity. There are several ways one may do the computations
on each individual G
N
, for example look for symmetry, or look for a suitable
periodicity, or similarity up to scale. In the simplest cases, this allows the
use of a nite Fourier transform, thus making G
N
periodic, or cyclic. The
case of G = Z (the rank-1 integer graph), and G
N
= the cyclic group of
order N is done in all detail in Example 6.6 above.
Some advantages of the cyclic approach: One, the spectrum comes out
given explicitly by a closed formula, thus making it clear how the limit
N works also for spectra, getting the continuous spectrum in the
innite limit.
Example 7.1 In this section we compare the two cases, continuous vs. dis-
crete. As noted, our graph Laplacians are second order (or more than second
order) dierence operators in a generalized sense.
They have spectrum contained in the half-line [0, ), so generalizing
(v) (x) : =
_
d
dx
_
2
v (x) (7.1)
with the Hilbert space H : = L
2
(0, ), and dense domain
T : =
_
v C
2
(0, ) [v, v

, v

L
2
(0, ) , and v (0) = v

(0) = 0
_
; (7.2)
42
i.e., with vanishing boundary conditions on v and v

(x) =
dv
dx
at x = 0.
We get the spectral estimate:
v, v)
L
2 0, v T. (7.3)
A simple verication shows that for the adjoint operator

we have:
dom(

) =
_
v L
2
(0, ) [v

, v

L
2
(0, )
_
. (7.4)
Comparing (7.2) and (7.4) we see that results from

by removing
the two boundary conditions which specify the domain T of .
Moreover, the defect space
T
+
: = v dom(

) [

v = v (7.5)
is one-dimensional; in fact,
T
+
= Ce
x
. (7.6)
The selfadjoint extensions of on T are parametrized by pairs of num-
bers A, B R, not both zero, such that
Av (0) +Bv

(0) = 0. (7.7)
Example 7.2 Let G = (G
(0)
, G
(1)
) be the following graph generalizing the
continuous example:
G
(0)
: = N
0
= 0, 1, 2, 3, . . . , (7.8)
G
(1)
: = (01) , (n, n 1) ; n N .
Pick > 1, and set
c (n, n + 1) =
n+1
. (7.9)
Then the corresponding graph Laplacian is unbounded; and
(v) (0) = v
0
v
1
; (7.10)
(v) (n) =
n
v
n1
+
n
(1 +) v
n

n+1
v
n+1
,n N. (7.11)
43
For domain T, we take all v
2
(N
0
) s.t. v
n
= 0 except for a nite set
of values of n. the matrix representation of is presented in Fig. 4:
_
_
_
_
_
_
_
_
_
_
_
_
0 0 0 0
(1 +)
2
0
0
2

2
(1 +)
3
0
.
.
.
n
0
n

n
(1 +)
n+1

0
n+1
.
.
.
_
_
_
_
_
_
_
_
_
_
_
_
Fig. 4.
By Parsevals formula, we have the isometric isomorphism
2
(N
0
) =
H
+
= the Hardy space of analytic functions on D = z C; [z[ < 1
v (z) : =

n=0
v
n
z
n
;
and
|v|
2
H
+
=

n=0
[v
n
[
2
. (7.12)
In the Hardy space representation we have
(v) (z) = (1 +) v (z) zv (z) z
1
v (z) (7.13)
on the dense space of functions v on C which extend analytically to D

: =
z C; [z[ < .
We now show that there are no non-zero solutions to

v = v, (7.14)
i.e., v dom(

); equivalently T
+
() = 0; the defect space for the
operator

is trivial. So this is a direct verication that

is essentially
selfadjoint; and contrasting with (7.6) above.
To see this, combine (7.13) and (7.14). It follows that every solution v
to (7.14) must have an innite-product representation given by
v (z) =
(z 1) (z 1)
z
v (z) ; (7.15)
44
and the limit of nite products as follows
(z 1)
n1

k=1
_

k
z 1
_
2
(
n
z 1)
z
n

n(n+1)
2
.
These products do not have a non-zero representation consistent with the
isomorphism (7.12), and with (7.12).
8 Appendix
A Heisenbergs Innite Banded Matrices
We proved in sections 3 through 5 that in general, graph Laplacians
c
are essentially selfadjoint operators in the
2
sequence-Hilbert space. Recall
that the axioms for our graph Laplacians include the following given data:
A graph G = (G
(0)
, G
(1)
) and a xed positive conductance function c dened
on the set of edges G
(1)
. Every vertex x of G is connected to a nite set
of neighbors in G
(0)
. For every xed x in G
(0)
, this implies niteness of the
set of y in G
(0)
for which c(xy) is nonzero. This means in turn that the
natural matrix representation of the operator
c
is banded; see section 5 for
the Denition. Note however that we place no boundedness restrictions on
the conductance function c.
Our proof of essentially selfadjoint for the operator
c
uses this band-
edness property in an essential way. In fact, starting with an innite by
innite matrix, it is generally dicult to turn it into a linear operator in a
Hilbert space unless it is assumed banded, see section 4, and the references
cited there.
The purpose of this section is three-fold.
First to make precise the operator theory of banded innite by innite
matrices; and second to show that the innite matrices used in represent-
ing the operator algebra generated by Heisenbergs quantum mechanical
momentum and position observables consists of (innite) banded matrices.
Thirdly, we use Heisenbergs (and Borns) computations to exhibit such
banded operators which are not essentially selfadjoint. The simplest such
matrix M is as follows: let P be Heisenbergs momentum operator and Q
the (dual) position operator. Then we show that the monomial M = QPQ
is banded, but not essentially selfadjoint. In fact, its deciency indices are
(1, 1).
45
Denition A.1 Let L be a countable (typically innite) set, and let m :
L L C be a function on L L. We say that m is banded i for every
x L, the set
y L[m(x, y) ,= 0 (A.1)
is nite.
Let
2
(L) be the sequence space with norm
|v|
2

2 : =

xL
[v (x)[
2
< . (A.2)
The sum on the right is the supremum of all the numbers

xF
[v (x)[
2
as F ranges over all nite subsets in L.
Let T be the dense subspace of all functions v : L C such that the
support set
x L[v (x) ,= 0 (A.3)
is nite. Equivalently, setting

x
(y) =
_
1 y = x
0 y ,= x;
(A.4)
the space T is then the linear span of the set of functions
x
[x L;
and these functions form an orthonormal basis for
2
(L). Moreover, every
Hilbert space H is isomorphic to
2
(L) for some set L. The set L is countable
if and only if H is separable.
Lemma A.2 Let m : LL C be a banded function. For v T
2
(L),
set
(Mv) (x) =

yL
m(x, y) v (y) . (A.5)
Then M denes a linear operator M : T T, with a well dened adjoint
operator M

. Moreover,
T dom(M

) (A.6)
where dom(M

) is the domain of M

.
Proof. When x L is xed, the sum in (A.5) is nite because the set (A.1)
is nite by assumption. Using niteness of both sets (A.1) and (A.3) we
conclude that Mv in (A.5) is in T if v is. And so, in particular, Mv
2
(L);
see (A.2) and (A.4).
46
To establish the inclusion in (A.6), we must show that for every
v T, there is a constant K = K (v) such that the following estimate
holds:
[Mu, v)

2 [ K|u|

2 , for u T. (A.7)
The expression on the left in (A.7) is

x,yL
m(x, y)u(y)v (x) . (A.8)
But the terms in this double-sum vanish outside a nite subset in LL an
account of assumptions (A.1) and (A.3).
The modulus-square of the sum in (A.8) is estimated by Schwarz by:

yL
[u(y)[
2

yL

x
m(x, y)v (x)

2
which yields the desired estimate (A.7).
Corollary A.3 Let M be a linear operator in a Hilbert space H. Then M
has a banded matrix representation if and only if there is an orthonormal
basis (ONB) in H, e
x
[x L such that the linear space T spanned by
(e
x
)
xL
is mapped into itself by M.
Corollary A.4 In that case the matrix entries of M are indexed by L L
as follows:
m(x, y) : = e
x
, Me
y
). (A.9)
Proof. Only the conclusion (A.9) is not contained in the lemma. Now
suppose some operator M in Hsatises the conditions, and let (e
x
)
xL
be the
associated ONB. Then Me
y
H
2
(L), so Me
y
=

xL
e
x
, Me
y
)
H
e
x
,
and
|Me
y
|
2
H
=

xL
[e
x
, Me
y
)[
2
(A.10)
holds by Parsevals formula. The conclusion (A.9) follows.
Corollary A.5 Let G = (G
(0)
, G
(1)
) and
c : G
(1)
R
+
47
be a graph system satisfying the axioms in section 2. Let
x
[x G
(0)
be the
canonical ONB in
2
(G
(0)
). Then the graph Laplacian has a corresponding
banded matrix representation as follows:

x
,
c

y
) =
_
_
_
c (xy) if y ,= x and y x
B
c
(x) if y = x
0 if y , x and y ,= x.
(A.11)
Proof. Recall the function
B
c
(x) : =

yx
c (xy) (A.12)
on the right-hand side in (A.11).
Since, for v T, we have
(
c
v) (x) : =

yx
c (xy) (v (x) v (y)) , (A.13)
setting v =
y
, we get
(
c

y
) (x) =
_
_
_
B
c
(x) if y = x
c (xy) if y x
0 if y , x
from which the desired formula (A.11) follows.
Heisenberg introduced matrix representations for the operators
of momentum P and position Q in quantum mechanics.
In the simplest case of one degree of freedom, they are as follows:
1
2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
0 1 0 0 0 0 0
1 0

2
0

2 0 . . .
0 0

3
0 0 0

n 2 0 0
.
.
.
.
.
.
.
.
. 0

n 1 0

n 1 0

n
0

n 0
.
.
. 0 0

n + 1
.
.
.
0
.
.
.
.
.
. 0 0
0 0
.
.
. 0 0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
48
and
1
2i
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
0 1 0
1 0

2
0

2 0

n 2 0 0
.
.
.
.
.
.
.
.
. 0

n 1 0

n 1 0

n
0

n 0
0 0

n + 1

.
.
.
.
.
.
.
.
.
.
.
.

_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
.
Set N
0
: = 0, 1, 2, . . . = Z
+
0, and H : =
2
(N
0
). Then the two
matrices P and Q are represented by the following second order dierence
operators, having the same form as our graph Laplacians (A.13).
(Pv) (n) =
1
2
_
n 1 v (n 1) +

n v (n + 1)
_
; (A.14)
and
(Qv) (n) =
1
2i
_
n 1 v (n 1)

n v (n + 1)
_
, (A.15)
for v T, n N
0
; where i =

1.
It is well known that both P and Q, as in (A.14) and (A.15), are essen-
tially selfadjoint.
It follows by the above lemma that
M : = QPQ (A.16)
is also a banded operator., referring to the canonical ONB e
n
[n N
0
in

2
(N
0
).
Caution: All the operators P, Q, and M are unbounded, but densely
dened; see [Jr77], [Jr78], [Sto51].
Proposition A.6 The operator M in (A.16) is Hermitian, and has de-
ciency indices (1, 1); in particular is not essentially selfadjoint. In fact, it
has many selfadjoint extensions; a one-parameter family indexed by T.
Proof. By the Stone-von Neumann uniqueness theorem, the two operators
P and Q in (A.14) and (A.15) are unitarily equivalent to the following pair
in the Hilbert space L
2
(R) of all square-integrable functions on the red line:
(Pf) (x) =
1
i
d
dx
f (x) , (A.17)
49
and
(Qf) (x) = xf (x) , for f L
2
(R) , x R. (A.18)
For domain T in (A.17) and (A.18), we may take T : = C

c
(R), or the span
of the Hermite functions.
From the representations (A.14)-(A.15), it follows that the operator M :
= QPQ in (A.16) commutes with a conjugation in the Hilbert space; and
so by von Neumanns theorem (see Remark 4.2), it has deciency indices
(n, n). We will show that n = 1. Hence we must show that each of the
equations M

= i v

has a one-dimensional solution space in H.

Taking advantage of Schr odingers representation (A.17)-(A.18), we ar-
rive at the corresponding pair of ODEs in L
2
(R):
x
d
dx
(xf) = f (x) . (A.19)
By symmetry, we need only to treat the rst one.
A direct integration shows that
f (x) =
_
exp(
1
x
)
x
if x > 0
0 if x 0
(A.20)
solves (A.19) in the case of + on the right hand side. Also note that
(A.20) is meaningful as all the derivatives of x
1
exp(
1
x
) for x R
+
tend
to 0 when x 0
+
. This means that the two separate expressions on the
right-hand side in (A.20) patch together dierently at x = 0.
By the reasoning alone, we conclude that M has indices (1, 1). As a result
of von Neumanns extension theory, the distinct selfadjoint extensions of M
are then indexed by T = z C[ [z[ = 1. If z T, and if f

are normalized
solutions to (A.19), then the extension M
z
is determined by
M
z
(f
+
+zf

) = i (f
+
zf

) .
Example A.7 Let P and Q be the canonical momentum and position op-
erators; see (A.14)-(A.15), and let
H : = P
2
Q
4
(A.21)
be the Hamiltonian of a particle-wave in one degree of freedom, corre-
sponding to a repulsive x
4
potential. Then the reasoning from above shows
that H is a banded matrix. As an operator in
2
(Z), H has deciency
indices (2, 2).
50
References
[AC04] Esteban Andruchow and Gustavo Corach. Dierential geometry of
partial isometries and partial unitaries. Illinois J. Math., 48(1):97
120, 2004.
[Arv02] William Arveson. A short course on spectral theory, volume 209
of Graduate Texts in Mathematics. Springer-Verlag, New York,
2002.
[BD49] R. Bott and R. J. Dun. Impedance synthesis without use of
transformers. J. Appl. Phys., 20:816, 1949.
[BHS05] Michael Barnsley, John Hutchinson, and

Orjan Steno. A fractal
valued random iteration algorithm and fractal hierarchy. Fractals,
13(2):111146, 2005.
[CS07] Mihai Cucuringu and Robert S. Strichartz. Self-similar energy
forms on the Sierpinski gasket with twists. Potential Anal.,
27(1):4560, 2007.
[DJ06] Dorin Ervin Dutkay and Palle E. T. Jorgensen. Methods from
multiscale theory and wavelets applied to nonlinear dynamics. In
Wavelets, multiscale systems and hypercomplex analysis, volume
167 of Oper. Theory Adv. Appl., pages 87126. Birkhauser, Basel,
2006.
[DJ07] Dorin Ervin Dutkay and Palle E. T. Jorgensen. Harmonic analysis
and dynamics for ane iterated function systems. Houston J.
Math., 33(3):877905, 2007.
[HKK02] B. M. Hambly, Jun Kigami, and Takashi Kumagai. Multifractal
formalisms for the local spectral and walk dimensions. Math. Proc.
Cambridge Philos. Soc., 132(3):555571, 2002.
[Hut81] John E. Hutchinson. Fractals and self-similarity. Indiana Univ.
Math. J., 30(5):713747, 1981.
[JKS07] Palle E. T. Jorgensen, Keri Kornelson, and Karen Shuman. Har-
monic analysis of iterated function systems with overlap. J. Math.
Phys., 48(8):083511, 35, 2007.
[Jr77] Palle E. T. Jrgensen. Approximately invariant subspaces for un-
bounded linear operators. II. Math. Ann., 227(2):177182, 1977.
51
[Jr78] Palle E. T. Jrgensen. Essential self-adjointness of semibounded
operators. Math. Ann., 237(2):187192, 1978.
[Jor04] Palle E. T. Jorgensen. Iterated function systems, representations,
and Hilbert space. Internat. J. Math., 15(8):813832, 2004.
[Jor06] Palle E. T. Jorgensen. Analysis and probability: wavelets, signals,
fractals, volume 234 of Graduate Texts in Mathematics. Springer,
New York, 2006.
[JP98] Palle E. T. Jorgensen and Steen Pedersen. Dense analytic sub-
spaces in fractal L
2
-spaces. J. Anal. Math., 75:185228, 1998.
[JP00] Palle E. T. Jorgensen and Steen Pedersen. Commuting self-adjoint
extensions of symmetric operators dened from the partial deriva-
tives. J. Math. Phys., 41(12):82638278, 2000.
[Kig03] Jun Kigami. Harmonic analysis for resistance forms. J. Funct.
Anal., 204(2):399444, 2003.
[Nel69] Edward Nelson. Topics in dynamics. I: Flows. Mathematical
Notes. Princeton University Press, Princeton, N.J., 1969.
[Pow76] Robert T. Powers. Resistance inequalities for the isotropic Heisen-
berg ferromagnet. J. Mathematical Phys., 17(10):19101918, 1976.
[PS72] K. R. Parthasarathy and K. Schmidt. Positive denite kernels,
continuous tensor products, and central limit theorems of prob-
ability theory. Springer-Verlag, Berlin, 1972. Lecture Notes in
Mathematics, Vol. 272.
[Sto51] M. H. Stone. On unbounded operators in Hilbert space. J. Indian
Math. Soc. (N.S.), 15:155192 (1952), 1951.
[vN31] J. von Neumann.

Uber Funktionen von Funktionaloperatoren.
Ann. of Math. (2), 32(2):191226, 1931.
52
This figure "Fig1a.jpg" is available in "jpg"format from:
http://arxiv.org/ps/0802.0133v1
This figure "Fig1b.jpg" is available in "jpg"format from:
http://arxiv.org/ps/0802.0133v1
This figure "Fig1c.jpg" is available in "jpg"format from:
http://arxiv.org/ps/0802.0133v1
This figure "Fig4.jpg" is available in "jpg"format from:
http://arxiv.org/ps/0802.0133v1