Commun. Math. Phys.

147, 485-526 (1992) Communications in

Mathematical
Physics
9 Springer-Verlag 1992

Determinants, Finite-Difference Operators

and Boundary Value Problems*
Robin Forman
Department of Mathematics, Rice University, Houston, TX 77251, USA
Received October 10, 1991

Abstract. We relate the determinants of differential and difference operators to the

boundary values of solutions of the operators. Previous proofs of related results
have involved considering one-parameter families of such operators, showing the
desired quantities are equal up to a constant, and then calculating the constant. We
take a more direct approach. For a fixed operator, we prove immediately that the
two sides of our formulas are equal by using the following simple observation
(Proposition 1.3): Let U ~SU(n, C). Write U in block form
U=( ull u12~,
\u21 u22/
where u~l and U22 aFe square matrices. Then

detull = detu22 .

O. Introduction

Motivated by questions in quantum field theory, there has been much recent
interest in the problem of calculating the determinant of differential operators (see,
for example, I-Ra] chapter III). Suppose L is a positive elliptic differential operator
acting on sections of a vector bundle over a compact manifold. Then L has
a discrete spectrum
)q < 2 z < . . . ~ o e .
Various methods have been used to make sense of
det L " = H2i".
Perhaps the most common method is the zeta-function regularization of Ray and
Singer JR-S], in which one defines log det L by analytically continuing the function
2/- s log 2,

* Partially supported by an NSF postdoctoral fellowship

486 R. Forman

from Re(s) large to s = 0. There are few general methods for evaluating such
determinants. In [ F o l ] we investigated one method of studying operators on
manifolds with boundary, based on a reduction of calculations to the boundary
of M.
A precursor to this general theory was the following example. Consider the
operators
d2 d2
L1 - dx 2 + Rl(x), L2 = - ~ + R2(x)

acting on functions f on the interval [0, t] which satisfy f ( 0 ) = f ( I ) = 0. The

operator L t L f i is of the form (Identity) + T where T is trace class, and hence
L ~ L f ~ has a well-defined determinant (that is, without any regularization, see
[G-K]). In fact ([D-D], [L-S])

det L I L I 1 [{flf(o)=f(1)=o} = y1(1)

y2(1) (0.1)

where, for i = 1, 2, y~(x) is the unique solution to

L~yl = 0

satisfying
y~(o) = 0, y~(o) -- 1 .

Note that ydl) "lives" on the boundary of [0, 1].

Once we know that, defining det L~ via zeta-function regularization,

det L_Ai= det(L1L~ l)

dct L2
( [ F o l ] Corollary 1.2) we can then reformulate (0.I) as

det - d-~ + R(x) = cy(1),

{flf(01 =f(1)= 0}
where y is the unique solution of

- ~ + R(x)
t
/
y(x) = O,

y'(0) -- 0, y'(0) = l ,

and c is a constant independent of R(x).

Another example of the general theory in [Fot], also considered in [B-F-K], is
d2
the example of the above operator L = - ~ x 2 + R(x) acting on functions on S 1.
This problem can be identified with L acting on functions f on [0, 1] satisfying
f(0) =f(1), f'(0) = f ' ( 1 ) 9
We then have

detL[{flf(o)=f(i),f,(o)=f,(i)}=cdet[(~ 01)_(z(1 ) y(1)~]

\z'0) y'(~)] '
Determinants of Differential Operators 487

where y(x) and z(x) satisfy

Ly=Lz=O,

y(O) = O y'(O) = 1,
z(O) : 1 z'(O) = o ,

and c is independent of R. (The constant c is studied in [B-F-K]).

What these formulae have in common is that the determinant of L, an operator
acting on an m dimensional space of functions is equated with the determinant of
a finite dimensional (1 x 1 or 2 x 2) matrix whose entries are constructed from the
boundary values of functions in the kernel of L, This may appear somewhat
mysterious. The existing proofs offer little insight.
One type of proof, seen in [D-D] and [B-F-K], involves varying the operator
analytically in a parameter z and observing that both sides of (0.1) are analytic in
z with the same zeros. Moreover, they have exponential growth order less than 1.
Thus, their ratio must be constant. This method exhibits a few disadvantages. First,
although it is clear that the two sides of (0.1) are zero at the same values of z, the
proof that the zeros have the same order is surprisingly difficult (although a some-
what simpler topological proof is offered in [Fo2] Lemma 1.2). Second, the method
gives no information about the constant of proportionality.
Another method, seen in [L-S] and [Fol], involves proving that the two sides
of (0.1) have the same logarithmic derivative. This involves using detailed informa-
tion about the resolvents of the differential operators. This method has the
advantage of being better suited to generalization ([Fol]), but again gives no
information about the constant of proportionality.
The goal of this paper is to "demystify" these results by presenting greatly
simplified proofs. The simplifications occur in two ways. First, we "generalize" the
above results to operators on finite dimensional spaces (formula (0.1) is then
derived by replacing the differential operators Li by suitable finite dimensional
approximations). Second, the proof uses no analysis, neither analytic functions nor
resolvents. Instead, we reduce the theorem to the following simple fact.
Proposition 1.3. Let U ~SU(n, C). Write U in block form

U~ ( ullUI2)
\U21 U22
where ua 1 and U22 are square matrices. Then

d e t u l l = detu22 1
Note that this proposition does indeed relate the determinant of an operator on
one space to the determinant of an operator on another space. Using Proposition
1.3 we prove directly that the two sides of (0.1) are equal. That is, unlike earlier
proofs, we do not vary the two sides at all. This new proof has the benefit of yielding
immediately, with no extra effort, the precise value of the constant of propor-
tionality.
Our general setting is the following. Suppose V is a finite dimensional vector
space equipped with an inner product, and
L: V--* V
488 R. Forman

is a linear map. For A c V satisfying

dim A = dim V - dim(kernel L)

we consider

LA = ~ALrcA: A ~ A ,

where ~zA denotes orthogonal projection onto A. Our main object of study is
det LA. In this vein, we factor LA into simpler operators. Suppose C is any space
transverse to kernel L. Then we can write

LA = A P~
... ~ C L ~ Image(L) ~ ~A,

where pcK is the projection onto C determined by the decomposition

V = kernel L + C . (0.2)

Thus
detLA = (det p~: A ~ C)(det L: C -~ Image(L))(det hA: Image(L) --+ A).
The right-hand side is expressed in terms of determinants of operators taking one
space to another. To make sense of these determinants we must choose volume
forms for these spaces. (This is explained more precisely in Sect. 1). The term
det ha: Image(L) ~ A (0.3)

is simple to compute. In addition, we can choose C so that

det L: C ~ Image(L) (0.4)

can be identified. We then apply Proposition 1.3 to prove

(detA p~ ; C) = (det C • p~ ~ kernel L ~ ~ A • , (0.5)

where pC denotes the projection onto kernel L induced by (0.2). Combining

(0.3), (0.4) and (0.5), evaluating d e t L a is reduced to calculating the determinant
of a map which factors through kernel L, a space which in our examples has
dimension I or 2.
Most of the results in the above mentioned papers follow immediately from this
formula. To indicate how this works, in Sect. 2 we present some examples. We
begin in part I with the simple application of the Laplacian L on a graph G. That is

V = {f: (Nodes of G) ~ C}

and
L: V ~ V.
Furthermore, L is hermitian and the kernel of L consists of the constant functions.
Since
dim kernel L = 1
Determinants of Differential Operators 489

the formula (0.5) is particularly easy to apply. A beautiful theorem of Kirchhoff

[Ki] (see also [B-D], [Fo2] and [Fr]) states

det L(kernelL)• [ = z,~pl~o(L)'~I = # ( N ~ ~ G) # (Spanning Trees in G) " (0'6)

If
M _ N = {Nodes of G}
is any subset, we define

(so that AN = (kernelL)• Then (0.5) easily generalizes (0.6) to yield

det LAM = # m" # (Spanning Trees in G).
In part II we consider general finite difference operators on an interval. As an
example, let
d2
E - dx z + ,R(x).

Divide the interval [0, 1] into d equal segments and approximate/7,by the standard
finite difference approximation
L = - V 2 -'~ R: V ~ V,
where

1} c}
and V denotes the finite difference approximation for d/dx. We then restrict L to
various subspaces of V (corresponding to boundary conditions). Three special
cases of the general boundary conditions considered in this section are:
Dirichlet Boundary conditions = A~o = {f~ V If(O) = f ( 1 ) = 0} ,

NeumannBoundaryConditions=Ao={feV[f(O)=f(~),f(d@t)=f(1)},

PeriodicBoundaryConditions=Bl={f6Vlf(O)=f(@),f(~)=f(1)}.

Formula (0.5) yields

Theorem 2.2. det LA~ = d 2d- 1y(1), where y is the unique element in kernel L satisfy-
in9
y(O) = O, gy(O) = l .

Corollary 2.5. det LA0 = - -4 Vz , where z e kernel L satisfies

z ( O ) = 1, Vz(O) : o .
490 R. Forman

Theorem 2.6.

detLm--d4d-2det
Ii~ O1)-lvz(d@l
(zt t) Vy(~-)J]"
/ ltl
Note that in each case the left-hand side is the determinant of an operator on
a d - 1 dimensional space, since

dim A ~ = dim Ao = dim B1 = dim(kernel L) • = d - 1 .

As our approximations become finer, that is as d ~ o% this dimension goes to or.
On the other hand, the right-hand side does not become more complicated as
d ~ ~ . This is due to the fact that
dim A~ = dim A~ = dim B~ = dim kernel L = 2

does not grow with d. Note also that there are no unknown constants.
We investigate the effect of adding a first order term to the operator L in Sect.
2(II), observation 4.
In Sect. 3 we provide the analysis which allows us to deduce formula (0.1) for
differential operators from the corresponding formula for finite difference oper-
ators. Unlike Sects. 1 and 2, this section is very technical. The reader willing to
believe some technical results may skip this section without detriment.
The results in this paper may have relevance for the attempts to study quantum
field theory via finite dimensional approximations.

1. The Main Theorem

Let V be a finite dimensional vector space defined over R or C, with a (Euclidean or

Hermitian, respectively) inner product, which we denote by ( , ) . Let

L: V--, V
be any hcrmitian m a p (this restriction will be removed later). Then for any
subspac e
AcV
we define
L A = 7rALTCA: A~ A ,
where ~A denotes the orthogonal projection onto A. Our goal is to study det LA.
As an example, let K denote the kernel of L, and K • the orthogonal com-
plement. Then
det LK~ = 1~ 2,
2~spec(L)
2=4=O

where spec(L) denotes the set of eigenvalues of L taken with their appropriate
multiplicity.
Determinants of Differential Operators 491

In fact, for any A, L A factors through LK~:

L~a 7~a

LA: A ~K~ ~ K - - ~ K• ~A .

Thus, it follows that if dim A = dim K •

det LA
- - = det ~A~K~: A --+ K-- --, A . (1.1)
det LKI
Note that the right-hand side depends only on the metric and the spaces A and K •
It is convenient at this point to be able to speak of the determinant of a map
between two different spaces. Suppose X and Y are two spaces of the same
dimension, each equipped with an inner product. Then, for any linear m a p
H:X~ Y
we cannot define det H, but we set
I d e t H I = (det H ' H : X --* X) ~ , (1.2)
where H* denotes the adjoint of H with respect to the inner products. The
following 2 properties are immediate:
(i) IdetH] = [ d e t H * [ , (1.3)
(ii) Suppose X, Y, Z are 3 spaces of the same dimension, each equipped with an
inner product. Then, for any maps
HI:X ~ Y

H2: Y~ Z
we have
ldet H2 o H11 = Idet H21 Idet H~ [ . (1.4)
We note that for any two subspaces X and Y of V, the adjoint of the map
~zr: X ~ Y
is
~x: Y-~ X . (1.5)
Thus, if
dim X = dim Y
we have
[detrcr: X + Y] = ( d e t ~ x x y : X ~ Y + X ) } . (1.6)
Together, (1.1) and (1.6) yield
Theorem 1.1. I f
dim A = dim K •
then

det LA
-- ]det rcK~: A ~ K • 2 = [det ~a: K • ~ AI 2 . (1.7)
det LK~
492 R. Forman

In examples, K = kernel(L) tends to be much smaller than K J-. Thus, our next
goal is to restate the right-hand side of (1.7) in terms of K, rather than K -L.
Before proceeding, for future reference it will be useful to broaden our perspec-
tive. Suppose
H:X~ Y
is linear, where X and Y are two spaces of the same dimension, each equipped with
an inner product. Rather than dealing only with
[detHt~R + ,
if we have volume forms for X and Y we can define a value for
detH~C.
That is, suppose
2~=X 1 A ''" A xn~AnX,
= Yl A " " ' A y. c A" Y (1.8)
(n = dim X = dim Y) are unit length volume forms (for example, if {x,} and {y,}
are orthonormal bases). Then we can define det H with respect to 5f and ~ by
Hxl /x . . . / x Hx, = (det H ) ~ . (1.9)
The forms Y" and Yr are determined up to multiplication by complex numbers of
norm 1. Varying our choice of ~r and ~r may vary the phase of det H (although (1.9)
gives a well-defined value for Idet HI which, as we will see below, agrees with our
previous definition). Defining all determinants with respect to fixed volume forms
2~ and ~r we have the following refinements of (1.3) and (1.4).

(i) d e t H = d e t H * . (1.10)
(Note that this implies that the definition of det H is compatible with our
previous definition (1.2) of Idet/41).
Proof. If 5f and Yr have unit length, then ~r and ~r are of the form (1.8) for some
orthonormal bases {x~} and {y~}. Define an n x n matrix ~ by
~ , j = (/4x,, yj>.
Then, it follows immediately from (1.9) that
det H = det Jt ~ .

Furthermore

(H*y~, x~) = ( y,, Hx~) = ~ = (~*)~

so that
d e t H * = detJt ~* = detJ{ ~ = d e t H . []
(ii) Suppose X, Y and Z are three spaces of the same dimension with volume
forms f , ~162and ~ , respectively. If
Hi:X--* Y, H2: Y--*Z
Determinants of Differential Operators 493

are linear, then, defining all determinants with respect to X, q; and H , we have

det H2 ~Ha = det H2 det H1 .

The main tool in all of our future work is the following lemma, which will
enable us to relate the determinants of operators between different pairs of spaces.

L e m m a 1.2. Let
V = X + X ' = Y + Y•
be two orthogonal decompositions satisfying
dim X = dim Y.

Suppose further that X, X • Y and Y• are equipped with unit length volume forms f ,
~• qY and oy• respectively, which satisfy
X / ~ X z = ~d/~ oyl . (1.11)

Then, defining all determinants with respect to f , 3s ~ and oy•

detrcr: X ~ Y = detrcx~: Y• --*X • .
(From (1.5) and (1.10), the right-hand side is equal to
det r@l: X • ~ y l ) .
We will reduce this lemma to a statement about special unitary matrices.
Let

{xl,...,xk}, {x~,...,x?}, {Yl,''.,Yk}, {Y~,...,#}

be orthonormal bases of X, X • Y and Y~, respectively. Then

{xl . . . . , Xk, X ~ , . . . , X{}, { y l , . . . , Yk, Y~ . . . . . Y~-} (1.12)

are orthonormal bases of V. Let U denote the matrix corresponding to the change
of basis
(:q,..., -. .....

That is

'(x~, y~> l<__i,j<k

<x,,y#k> l <_iNk, k + l <=j<=k +l
Uij =
(X•-k, Yj) k + l < = j < k + l , i<=j<=k
]_ •
.<x~-k, yj-k> k+l<=i, jGk+l.
Then, from (1.12), U is unitary. F r o m (1.11), det U -- 1, so that

UeSU(k + I,C)
Write U in the block form

U121
I ull
U21 U22
494 R. Forman

where
hi11. = ( U i j ) l ~ i,j ~ k U12 = ( U i j ) i <=k,j >=]r + 1

1/21 = ( U i j ) i ~ k + l,j <=k u22 ~- ( U i j ) k + 1 ~ i,j ~ k + l 9

Then
d e t e r : X--+ Y = d e t u , t
det ~v~ : X " -+ Ya = detu22 ,
Thus, the l e m m a follows from the following proposition
Proposition 1.3. Let U ~ SU(n, C) be expressed in block form

\U21 U22,/
so that u , , and uzz are square matrices. Then

det u, 1 = det u22 9

Proof Since U e S U ( n , C) we have

~),U,=( ull /g12~( u~l u * ~ = ( l 0 ?). (1.14)

\u2~ u;2J\ut~ uS)
Thus, using the second column of (1.14), we find
b/ll /,/12~(1 U~I~ = (Ull 0|)
u21 u22/kO U~'eJ \u_,,
T a k i n g determinants yields
det u*2 = det ul 1
as desired.
Substituting L e m m a 1.2 into T h e o r e m 1.1 yields
Corollary 1.4. (i) I f
dim A = dim K ~ .

Then
det LA
- - = IdetzcA• K--+ A l l 2
det LK ~
(ii) I f
dim A = dim B = dim K •

and
det LB =~ 0

then
det L a Idet •A• K -~ A I ] 2
]det rcBl: K ~ B~I z = Idet ~AB: B • --+ K ~ A l l z , (1.15)
det L8
Determinants of Differential Operators 495

where

(T~B• i ~A •
CbAB = B 1 )K )A • .

(This is a m a t r i x version o f Theorem 1 o f [Fol].)

The value of formula (1.15) largely rests on our ability to choose a space B for
which we can calculate det Ln. In what follows, it will help us greatly to be able to
consider slightly more general operators for our denominator.
From this point on, we will no longer assume that
L: V ~ V
is hermitian. Let C c V be any linear subspace such that
dim C = dim V - dim(K - kernel(L)) (1.16)
and
C c~ K = {0}. (1.17)
Then we can write

V= C + K, (1.18)
although this decomposition is not necessarily orthogonal. Let
pc~: V--, C
denote the projection onto C induced by (1.18). The symbol rc will be reserved for
orthogonal projections, so that, for example, for any space A
A•
~A = PA .

Now, for any v e V

v = p (v) + k

for some k ~ K, so that

Lv = Lpc (V).
This implies that for every A satisfying

dimA = dim V - d i m K (1.19)

LA factors through L p g . Namely
pK L ~A
LA = A ' C , Im(L) ,A

(where Im(L) = Image(L)). Thus

d e t L a = (det p~:: A ~ C)(detL: C ~ Im(L))(det~a: Im(L) ~ A ) . (1.20)

Our last simplification is to note that p~ factors through K • Namely
~'K • p cK
p~: A ~ K • )C
496 R. Forman

so that
(det p~" A ~ C) = (det rcK~: A ~ K l ) ( d e t pcK: K • ~ C ) . (1.21)
F r o m L e m m a 1.2, with appropriately chosen volume forms
detrcK~:A ~ K • = detTza~: K ~ A • . (1.22)
Moreover

Kl P~ , C = K • (~)-' , C ,

so that
det pcK: K • ---, C = (det nK~: C ~ K• -1 = (det rCc~: K --* C• -1 (1.23)
by Corollary 1.4. Now, using that

K zC~>c A = K (Pf'I-~C A

we find that (1.23) is equal to

det pC. C A ~ K . (1.24)
C o m b i n i n g (1.21), (1.22) and (1.24) we see that

detp~:A~C:detC z )K )A • .

Substituting into (1.20) we find

Theorem 1.5. Let C be any linear subspace of V satisfying (1.16) and (1.17) and A any
subspace satisfying (1.19). Then, with respect to appropriate volume forms on A, A l,
C, C • and Ira(L) we have

d e t L a = (det C• pf, > K ~A~) A)(det L: C ~ I m ( L ) ) ( d e t 7"~A: I r a ( L ) ~ ~A A ) .

The significance of T h e o r e m 1.5 is d e m o n s t r a t e d in Sect. 2(II), in which, for

a large class of finite difference operators L, we find a space C (in the examples
C = Cauchy data) such that
d e t L : C ~ Ira(L)
can be readily calculated. The space Ira(L) is easily identifiable and hence
det ~A: Ira(L) ~ A
can be calculated. Thus, T h e o r e m 1.5 relates det LA to det rcA~pc: C • ~ A-. This
has the p r i m a r y a d v a n t a g e that dim A • dim C • is m u c h smaller than dim A.
Moreover, we identify rCA~PK.c"C • ~ A a with a classical analytic operator.

2. Applications

I. Laplacians on Graphs. Let G be a connected graph consisting of a finite set of

nodes N and a finite set of edges E. We allow multiple edges between vertices. Let
V = N * be the space of m a p s
f:N~C.
Determinants of Differential Operators 497

Define a Laplacian
L: V--* V
by setting, for f c V, n 9 N
(Lf)(n) = d(n)f(n) - ~ /(~), (2.1)
eeE(n)

where E(n) is the set of edges attached to n,

d(n) = # E(n)
and, for e 9 E(n), ~ denotes the other endpoint of e.
There is a canonical basis of V given by {fn}n~N, where, for n, m e N

f,(m)={10 if m = n
if m ~ n .
We define an inner product ( , ) on Vby declaring {f, } to be an orthonormal basis.
The Laplacian L is hermitian with respect to ( , ) .
It can be seen directly from (2.1) that the kernel of L is 1-dimensional, consisting
of the constant functions.
The operator L plays a crucial role in electrical network theory. It is in this
context that Krichhoff proved his beautiful theorem [Ki] (see also [B-D, Fr, and
Fo3]) that

det LK•
[-- Xe2P~
l
[I%(L)v 2 = # N . # {maximal trees in G} , (2.2)

where a tree is a subgraph of G containing no cycles.

For any subset M __G_N we define a space A~t c V by

AM={ f 9 n~eMY'f ( m ) = 0 } .

Then, if M = N we have AM = K ' , where K = kernel(L). If M = {n} then

AM = { f 9 Vlf(n) = 0}.
If G represents an electrical circuit, then A M corresponds to the possible potentials
on the vertices if {n} is grounded.
Applying Corollary 1.4 we learn
Theorem 2.1. For any M ~ N

det LA~ = ~ M" # {maximal trees in G} .

Proof From Corollary 1.4 and (2.2)
det LA~, = # N . # {maximal trees in G} [det ~AM:K ~ A~[ 2 . (2.3)
An orthonormal basis for K is given by fK where
1
fK(n) ( # N ) 4 for every h e N .
498 R. Forman

An orthonormal basis for A~t is given by fM where

0 n6M
fM(n) = 1
n~M .
(#M) ~
Thus

so that

det(xA~: K ~ A~M) = < f r , f u > = ~ f r ( n ) f m ( n )

n~M

1 (#M) ~
(2.4)
= .--~M ( # N ) ~ - ( # / ) 89-- ( # N ) ~ '

Substituting (2.4) into (2.3) yields Theorem 2.1. []

II. B o u n d a r y Value P r o b l e m s on an Interval. We approximate the unit interval by

d + 1 nodes

0 = no nt na 9 9 9 rid-1 nd= 1 ,

where nl = i/d. Let

N={no,...,nd}.

We approximate C r valued functions on the interval by

V-- N* = {f:N~C"}.

Let L be a linear operator on V of the form

~ cj(i)f(ni+j) if a - < i - < d - b

(Lf)(n~) = j= -a , (2.5)
0 if i < a or i>d-b

where a and b are non-negative integers, not both zero, and each cj(i) is an r • r
matrix. L is a finite difference approximation of a differential operator of order
a + b. We require L to be elliptic. This is reflected in the condition

d e t c _ a ( i ) # O, detcb(i) 4= 0

for all a _< i _< d - b. It follows directly from the definition of L that

dim(K = kernel L) = r(a + b ) .

In particular, f E K is completely determined by

f(no),f(n~), . . . ,f(na+b-1).

For example,
(L f ) (n,) = 0
Determinants of Differential Operators 499

implies
b-1
f(n,+b) = -- [cb(a)]-i cj(a)f(n~+j).
j~ -a

The other values o f f are determined inductively.

We define a space C c V by
C={feVlf(ni)=O i<a+b}.
Then

d i m C = r(d + 1 - (a + b)) = d i m K l
and
c /c = { o } .

Thus
V=K+C
and we have a corresponding projection
p~ V-~C .
Let A be any linear subspace of V with
d i m A = r(d + 1 - (a + b)).
Then, applying T h e o r e m 1.5 we have, with respect to appropriately chosen volume
forms
det L a = (det 7~A~pC: C • --, A• L: C ~ Image(L))(det teA" I m a g e ( L ) -~ A ) .
O u r goal is to identify the determinants on the right-hand side. We first note that
I m a g e ( L ) = { f ~ V[f(ni) = 0 if i < a or i > d - b} .
Thus, once A is given
det rCa: I m a g e ( L ) ~ A
can be explicitly calculated.
Let {el . . . . . e,} d e n o t e the standard basis of C'. Define a basis
{f/,j}i=0 ..... d , j = 1. . . . . r of V b y setting for nkeN

{;jifi=k
fi, j(nk) = if i =[=k .

We define volume forms ~ , 5r and ~f, for V, I m a g e ( L ) and C, respectively, by

~U= fo,, A fz,i A "'" A fd, i

i i i

~- +b,i A +b+l,i A ' " " A ,i 9

i= i= i
500 R. Forman

Expressing
L: C ~ Image(L)
as a matrix with respect to the above bases, L is a (block) lower triangle matrix
(~ cb(a) 0 O. . ,

b_l(a + 1) cb(a + 1) O. . . 0
b-2(7+2) cb-l(a+2) cb(a + 2 ) . . . 0

0 cb(d'-- b) J
This follows by inverting (2.5) to find
rain{d - b, k + a}
L(A,,) = <ck-i(i)ez, ej>f,j.
•177 j=l
Thus, with respect to the volume forms ~ and (g
d-b
det(L: C -~ Image(L)) = 1--[ det cb(i).
i=a

Lastly, we consider the operator

Z~A~pC:C • ~ K ---,A a .
We note that
C • = {flf(nO = 0 for i = a + b}
and

is the solution of the initial value problem for the finite difference operator L.
Namely, for/e C •
pC(f) = f a K ,
where J~e K is the unique element in the kernel of L such that

f(ni)=f(n~) fori<a+b.

Thus
pC: C" ~ K

is the standard identification of an element in the kernel of L with its initial

conditions.

A Special Case: Hill's Operator, We now specialize to the case of finite difference
approximations to Hill's operator:
dz
s -- dx 2 +/~(x): C~176 1], C r) ~ C~176 1], Cr),

where /~ is a continuous one-parameter family of r x r matrices 9 Fixing d, the

Determinants of Differential Operators 501

number of nodes, we place the node n~ at the point i/d and approximate the
operators d/dx and dZ/dx 2 by
V: V--* V,
V2: V ----~V,
where
(Vf)(nO = d [f(ni+ 1) - f ( n i ) ] ,
(VZ f)(nl) = d[(Vf)(n~) - Vf(ni-1)]
= dZ[f(n~+l) - 2f(ni) + f ( n ~ - l ) ] 9
Set

Then our finite approximation t o / ~ is

L=- V2 + R : V ~ V .
That is, for f ~ V,

f O d2f(n~-l) + [2d2 + R(n~)]f(nO - d2f(n,+l) if0<i<d

(Lf)(nz)
if i = 0 or i = d
or, in the notation of (2.5), a = b = 1, and for 0 < i < d,
-d if j = _+ 1
cj(i)= 2d z + R(n~) j = O
0 j# - 1,0,1.
In this case
C = { f e VIf(no) = f ( n l ) = 0},
Image(L) = { f e V[f(no) = f(nd) = 0 } ,
and
d-1 d-1
detL:C~Image(L)= I-[ detcl(i) = ~ ( - d 2 ) ' = ( - d 2 ) r(~-l) . (2.6)
i=1 i=1

We now consider different "boundary conditions" A.

i) Dirichlet Boundary Conditions. Let
A~ = Image(L) = {flf(no) =f(nd) ----0}.
This is the so-called Dirichlet boundary condition. (The notation A~ is chosen to
be consistent with the notation of part (ii).) Then

Image(L) ~A~ A~

is the identity, so
(det ~ZA' Image(L) --+ A~) = 1
502 R. Forman

as long as we take as our volume form ~r on A~o the same form s that we defined
for Image(L). Thus, from T h e o r e m 1.5 and (2.6),
det LA~ = (-- dZ)'(a-1)(detTzA~pc: C a ~ A ~ ) . (2.7)
N o w , for simplicity, we set r = 1. Define a basis {fi}i = o..... a of V by setting

f(nj) = i #: j .

O u r volume forms W, 5 a, ~r and cd of V, Image(L), A~o and C, respectively, reduce

to
d
v= A/,,
i=0
d-1

,:=1
d

i=2

Let

~e" =fo ^A
be a volume form for C • and
d~ =(- 1)~-lfo Afa
a volume form for A ~.
The sign is chosen so that
call ~, (r = d ~ ,x doo = ~ .

The m a p
ztA~pC: C • ~ A ~

takes fo 9 C to
(ko(no))fo + (ko(na))fi = f o + (ko(na))fa,
where ko = pCfo e K is determined by
ko(no) = f o ( n o ) = l, ko(nl) = f o ( n O = 0
and f l to
(k 1 ( n o ) ) f 0 -]- (k I (F/d))j~ = (kl (nd))fi,
where kl = pC f l 9 K is determined by
kl(no) = A ( n o ) = O, kl(nl) = A ( n l ) = 1 .
Thus ~ " = fo /x f l goes to
[-fo + ko(na)fa] /x kl(na)fa = ( - 1 ) a - l k l ( n a ) [ ( - 1)a-lfo A f t ]
= (-- 1 ) a - l k a ( n a ) d ~ .
Determinants of Differential Operators 503

Thus
(det xA~pC: C ~ A~) = (-- t ) a - l k ~ ( n a ) . (2.8)
Recall that
kl(no)=0, kl(no)=l
so that
Vk~ (no) = d .
Thus

y = ~kl (2.9)

satisfies
y(no) = 0, Vy(no) = 1 .

Combining (2.7) [setting r = 1], (2.8) and (2.9),

Theorem 2.2.
det LA= = d 2a- 1y(na),
where y is the unique element in the kernel of L satisfying
Y(no) = O, Vy(no) = l .
For r > 1 the above analysis can be generalized to yield
Theorem 2.3. For general r
det LA. = d r(zd- 1) det Y,
where Y is the r • r matrix whose i th column is given by yi(na), where Yi is the unique
element in K such that

y i ( n o ) = 0, V y i ( n o ) = e~ .

Observations. 1) Note that as d ~ o% detLA~ ~ ~ . This is appropriate since, as

d ~ ~ , LA~ approaches
d2
s -- dx2 ~- /~]Diriehlet
which has ~ many eigenvalues which tend to ~ .
2) Let
d2 d2
T"I -- dx 2 +/~1, /~2 - d x 2 -I- R2
be Hill's Operators associated to two different potentials. For fixed d, let L~, L2 be
the corresponding finite difference approximations. The relative version of
Theorem 2.2 is

y~(nd)
detL1 A~(L2 A~) - 1 - (2.10)
' ' y2(nd) '
504 R. Forman

where Yi is the unique element in the kernel of L~ satisfying

yi(no) = O, Vyi(no) = 1 .
Taking limits as d ~ oo we find
Yl(X)
det - d~5 +/~1 (x) - d~5 +/~2(x) oir~chlo,= )~2(x) ' (2.11)

where y~ is the unique solution to

- ~ + g,(x) ~(x) = o ,

k(o) = o ;;(o) = I ,

and the left-hand side of (2.11) is a well defined Fredholm (= perturbation)

determinant. The technical details required to justify this limiting process are
provided in Sect. 3. There is a corresponding formula for general r. The formula
(2.11) was first proved in [L-S].
3) The operator LA~ is hermitian, and hence has real eigenvalues.
Taking the sign of both sides of Theorem 2.2, using the fact that y ( n l ) is
positive, we find
[ # of negative eigenvalues of LA~o] = [ # of sign changes of y(nl) as i goes from
1 to d] (mod 2).
This is a weak version of the Morse index theorem [Mi] which states that in the
smooth case:

# of negative eigenvalues of s -- dx 2 t- ifflD~richxet = [ # o f zeros of

~(x) = the unique solution of/2 such that ,~(0) = 0, ~'(0) = 1].
4) We can also vary the first order term i n / 2 That is, consider
d2
s dx 2 + O~(x) + g(x) .

A complication arises in that there are 3 reasonable ways to approximate d/dx by

a finite difference operator, namely by the left, right or symmetric difference. That is

can be approximated by

and we will label the resulting finite difference operators LL, LR and Ls, respec-
tively. In any case, the analysis goes through as before, except possibly for the
formula (2.6).
Determinants of Differential Operators 505

In the first case the formula is unchanged and we learn (in the case of Dirichlet
boundary conditions)
det LL = d zd- 1yL(nd) ,
where YL is the unique solution of
LLyL = O, yL(O)=O, VyL(O)= 1.
In the case of LR we have

so that (2.6) is replaced by

detLR:C--+Im(LR)=dfil(--d2+Q(~)

=(-d) 2(a- 1) 1 -~Q ,

i=l

and we learn

detLR = d 2(d-1)
r H,( (5))]
Li=I
1 - ~Q yR(nd) 9

In the case of Ls we have

cl(i)= -d 2+~ d

so that (2.6) is replaced by

1 ~/i\d\ )
detLs:c~Im(Ls)=afili=l - d 2 + 2Q~d)

= ( - d 2 ) d-~ I~ 1-~-~Q .
i=1

Proceeding as before, we learn

det Ls = d 2d-1 1 - ~-~Q ~ ys(na).

1_i=1

It is fascinating to note that as d ~ 0%

dili=(1-d
l \ d J J --'e

[I 1- --*e-i! 5(x)dx
i=i

In fact, we proved in [ F o l ] (Corollary 3.5, suitably interpreted) that there is

a c such that
1 I ~
det/] = ce-~!O-(x)dx~(1)
506 R. Forman

(where the left-hand side is defined via zeta function regularization). The constant
c is independent of (~ and/~. Thus, it appears that Ls is the appropriate approxima-
tion. However, it is not at all clear how to deduce the above formula for/7 from the
finite difference case. For example, if/21 and/~2 have different first order terms, then
/71/2~ 1 does not have a well-defined Fredholm determinant. (The spectrum of
I~1/~,~1 looks looselylike { 1 + ~ 0~en~Z} and thus, the product does not
converge absolutely). Hence, we cannot use det L 1 L f ~ as we did in Observation 2.
However, it follows from the above that, in fact
T(d) d~o~ det/~1
det LI,~
act L(2a,)s det/~2"
A direct proof of this fact would be very interesting, and may shed some light on the
mysterious process of zeta function regularization.
(ii) General Local Boundary Conditions. Fix 3o, 81eR. In this section we
consider the operator/~ restricted to functions
f : [0, 13 ~ C
which satisfy
if(O) + 3of(O) = i f ( l ) + 81f(1) = 0.
Note that 8o = 81 = 0 corresponds to the Neumann boundary condition
if(0) = / ' ( 1 ) = 0
and, although here we require 3o, 31 < oo the case 30 = 81 = oe corresponds to the
Dirichlet boundary condition considered in (i). As before, after fixing d, the number
of nodes, I2 is replaced by L, and the boundary condition is replaced by
Vf(no) +8of(no) = Vf(n~-l) + 81f(nd) = O,
i.e.

f(nl) + ( ~ - l)f(no) = f(ne-1) + ( - @ - l)f(n~) -- 0 " (2.12)

Thus, we define
Aao, al = { f e Vlfsatisfies (2.12)} .
Note that an orthonormal basis for Aao, al is provided by

where

= I --fo + 1)f l = + 1) 2'

.=
Determinants of Differential Operators 507

We define a volume from also, 5, for Aso,~ by setting

Afd-2
A (fa-t-(@nt- 1)fa-1) -
To find the determinant of the map
~a: Image(L) ~ A
we note that
=a(f)=f 2<i<d--2,
~a(fl)=lfl,l(-fo+(@--l)fl))l(--fo+(~--l)fl)
0~0\ 0~0
1 fro

~A(frt-1)=(fa-l,!(f,~-]-(~'~+
oq\ 1)f,t-1))l(fg+(~-+- 1)fa-1)
1 (51
= ~ \-d + 1) [ l ( f a + ( ~ + 1)fa- 1)1 '

Thus

det(rcA: Image(L) ~ A) = ~o~1 \d- - + " (2.13)

Combining Theorem 1.5, (2.6) and (2.13),

det LA~o~~- (-d2)a-1 ( ~ ~ - 1 ) ( ~ + 1 "(det~Aio.~,pC: C • --+ A~o,~,). (2.14)

' 0C0~ 1 t •
Now

A~ao,~l= s p a n { ( ( ~ - 1 ) f 0 +fl),((~ + lOfa-fa-ll}.

We define a volume form dOo,O
- , by
• ( - 1)a-1((~o ) )
dOo,~= 0~00~1 --1 f o + f l /~ ((~1 ~__1)fd __fd_1) "
Again, the sign is chosen so that
• cg•
d~o,~ ~ A ~4~o,~~ = A (g= ~ =fo A fl A " " A fd.
The map
~Aio,,,pC: C • A6o,
• 61
508 R. Forman

takes fo 9 C" to

where ko is the unique element in K satisfying

ko(no) = 1, ko(nl) = 0 , (2.t5)
and fl e C • to

where kl e K is the unique element satisfying

kl(no) = 0, kl(nz) = 1 . (2.16)
Thus)Co /x fi goes to

(~j);S' [ (ko,(~ff -1)fo +fl) (kl, (~ + 1)fa - fa-~)

Define z 9 K by

z=(ko,(~-l)fo+fi)ki-(kl,(~+l)fo+fi)ko=(~ 1)kl - ko.

Then z 9 K is the unique element in K satisfying

z(no)=-l, z(nl)=#- 1

or, equivalently
z(no) = - 1, Vz(no) = 6o .

Moreover, from (2.17),

det ~A~o.~,pC: C • ~ A~o.~1 = - O~oO~i

- z, +1 fa-fa-1
- ~(o- ~ 11) a - i ( ( ~ + l ) z ( n ~ ) -
Z(FId_1))

(_ 1)a- 1
- - -
d~o~ i
((61 + d)z(na)- dz(na-1))
1)a- 1
- (_
--dc~o~i (61z(na)+ Vz(na-1)). (2.18)

Combining (2.14) and (2.18) yields

Determinants of Differential Operators 509

Theorem 2.4.

det LA,o.~, - 22 -1 + 1)(Vz(nd-1) + glZ(na)) ,

~oO~1

where

~= -1 +1, c~= +1 +1

and z ~ K is the unique element in the kernel of L satisfying

z(no) = - 1, Vz(no) = go 9

Recall that Ao, o denotes the Neumann boundary conditions. Theorem 2.4
provides the formula
Corollary 2.5.
dZa-3
detLAo,o - 4 Vz(na_ l ) ,

where z ~ K satisfies
z(no) = 1, Vz(no) = O .

/f
d2 d2
s -- dx2 + RI(x), s - dx 2 +/{2(x)

are two operators, then the relative version of Theorem 2.4 yields, upon taking the
limit d ~ o%

det - ~ +/~(x) - ~ x 2 +/~2(x) = ~i(1) - g~2(1) '

where

-4ao,al = {f: [0, 1] ~ C If'(0) + 5of(0) = i f ( l ) + g l f ( 1 ) = 0}

and 2i, i = 1, 2, is the unique solution to Li such that

~i(0)=-l, ~(0)=go.
iii) Periodic Boundary Conditions. In this section we consider the operator
/7 acting on functions on the circle S ~. We associate S 1 with [0, 12 under the
identification 0 = 1. Then, C ~ functions on S 1 can be identified with functions on
[0, 12 satisfying

f(O) =f(1), f'(O) = f ' ( 1 ) .

We think of a finite approximation of S 1 as given by the interval with d + 1
evenly spaced nodes under the identification

no ~ nd-l~ n l ~ nd 9

Then, C a functions on S 1 are approximated by functions f ~ V satisfying

f ( n o ) =f(nd-1), Vf(no) = Vf(n~_z)
510 R. Forman

or, equivalently,
f(no)=f(ne-1), f(nl)=f(nd). (2.18)
Fixing 6 ~ C, we generalize (2.18) to the boundary condition
Ba = { f e VI f(no) = 6f(ne_ 1), f(nl) = 6f(nd)}.
An orthonormal basis for Ba is provided by
1 1
~(afo + f,-~l, ~ ( 6 f l + f,),f2,f3 . . . . . f,-~ .

Define a volume form Na by

1
d 0 - 1~ +( ~161f ~ + f " - ~ ) /' (6f~ +f~) ^ f2 /, . . . A f~_~.

To find the determinant of the map

rcB~:Image(L) ~ Ba
we note that
rcB,(f~) = f~ 2<_i<d-2,
1 1
~,(f) = (fl, ~ ( ~ f l +fd))~(6fl -t-fd)
6
+fd),
1 + 1612(65

{
~,,,(f~_,) = f , _ ~ , ~ ( a f o , },
+f~-~) ~ ( 6 f o +f~-~)

1
1 + 1612(~2~
+fd-1) 9

Thus
~e=A A f ~ ^ . . . A f,,-1
goes to
fi
((6k + f~) ^ f2 ^ ' " ^ fd-2 /, (6s + fd-1))
(1 + 161~) 2
6
= ( _ 1)d- 2 _ _ ~a (2.19)
1 + 16f2 "

Combining Theorem 1.5, (2.6) and (2.19),

__ d Z d - 2 t5
detLB~- i +[-~T (det~zs}P c: C--+ BaL). (2.20)

Now
B~ = span{(fl - 6fa), (fo - 6 f d - 1)} 9
Determinants of Differential Operators 511

We define a volume form ~ by

1
1 + lO[2(fo - ~ f d - D A ( A -- ~fd)

so that
~ A ~=~=fo A''' AA.
The m a p
nB~pC: C • --~ B~
takes fo to
1 1
+ la~ 2 (ko,fo - 6 f d - ~ ) ( f o - bfa- 1) + 1 + { O [ ~ ( k o , f - b f d ) ( f ~ - 6fd),
1
where ko e K satisfies (2.15), and fz to
1 1
+ lal 2 ( k , , f o - (~fa-1)(fo -- 6 f a - t ) + ~
1 (+k lal
x'f~ - 6 f d ) ( A -- 6fd) ,
1
where kl e K satisfies (2.16). Thus ~ z =Jo A f l goes to
1
1+ 1612[(ko,fo - ~f d - 1 ) ( k x , f t -- a f d ) -- (k~,fo -- ~f d - ~ ) ( k o , f l -- 6 f d ) ] N~-

SO that
det (nB~pC: C • ~ B~)
1
-- t + [6[ 2 [ ( k ~ 1 7 6 -- c ~ f a - ~ ) ( k l , f t - 6fa) - ( k l , f o - 6 f a - ~ ) ( k o , f ~ - a f t ) ]

1
- 1 + 1512 [(1 - 6ko(na-D)(1 - 6kl(na)) + (C]kl(na-i))(- Oko(nd))]

-i
+
1
I-~detI(o
1
?) -6{k~
\ ko(na)
k,(na-,)~]
k,(na) I ] "
(2.21)

Thus, from (2.20) and (2.21), we have

Theorem 2.6.

detLBa-(t _,~,~,2)~det[(
_ ~ ~)-~{\ kO(nd-1)
__ d 2 a - 2 6

ko(n~)
kl(nd-1)x]~
kl(na) ) ] "
(2.22)

Unfortunately, the formula (2.22) is not convenient for taking limits as d -* Go,
but (2.22) is easily seen to be equivalent to

aotL o=(, + i l ) aO kk ~
where z = ko + kl, y = ~ kl a K satisfy

z(no) = 1, Vz(no) -= O,
y(no) = 0, Vy(no) = 1
512 R. Forman

If
d2 d2
/~1 = -- ~X2 -~ e l ( X ) , /~2 -- d x 2 q- I~2(X)

are two operators, then the relative version of (2.23) leads, upon letting d --~ 0% to

det(-~xZ+/~l(x))(-d--~+/~2(x)) B =--;
det
det/i ~i \,~[(1)y[(1)J
6 ~ ~ '
\el(l) yi(1)/

where zi, Yi are the unique solutions to/~i on [0, 1] satisfying

~,(o) = 1, ~;(0) = 0 ,

y,(o) = 0, y;(o) = 1 .

3. The Convergence as d - + oo

In this section we prove that as d ~ oo the formula (2.10) converges to the formula
(2.11). More generally, we prove that for any boundary condition, the relevant data
from the finite difference operators converges, as the approximation becomes finer,
to the corresponding data for the limiting differential operator. Thus, each formula
in Sect. 2 for the determinant of a finite difference boundary value operator yields,
upon taking limits, a corresponding formula for the determinant of a differential
boundary value operator. Our analysis involves two steps. For the right-hand side
we must prove the convergence of solutions to the initial value problem. This type
of result is standard, but a proof is included for completeness. For the left-hand
side, we must prove the convergence of the ratio of determinants. The author has
been unable to locate the desired results in the literature, and thus we include
a complete proof. Although the results in this section hold for any ordinary
differential operator, for simplicity, we restrict attention to 2 "a order operators of
the type considered in Sect. 2.

Convergence of Solutions of the Initial Value Problem. The right-hand side of (2.10)
is expressed in terms of the boundary values at x = 1 of solutions with prescribed
boundary values at x ; 0. Fix a, b e C. Let L denote the differential operator
d2
L -- - a T + R ( x ) ,

where
R. [0, 13 --* C

is continuous, and let f denote the unique solution to

L f = O, f(O) = a, f'(O) = b . (3.1)
Let L (a) denote the finite difference approximation to L corresponding to
Determinants of Differential Operators 513

a division of [0, 1] into d equal subintervals, and letf (a) denote the unique solution
to
L(a)f (a) = O, f(a)(O) = a, gf(a)(O) = b. (3.2)
Theorem 3.1. A s d ~ oo

f(a)(1) + f ( 1 ) , (3.3)

Proof Let

Then (3.3) is equivalent to

lim e(a)(1) = 0. (3.5)
d-'* co

From (3.1) we learn thatfis twice continuously differentiable. Thus, in particular,

if(i)= dlf(1 )-f(~-)l+ O(~).

Since
(d) / d - 1\ q

(3.4) is equivalent to

lira die(a)(1) - e ( a ) ( d @ l ) l = 0. (3.6)

d --, oo

From Taylor's theorem, for any x 9 (0, 1),

f(x- h) =f(x) - f ' ( x ) h + 5] f tt (x)h 2 + o(hZ),

f ( x + h) =f(x) + f'(x)h + ~1 f . (x)h 2 + o(h2) .

Adding these two equations yields

h2f"(x)-- - 2f(x) + f ( x + h) + f ( x - h) + o(h2) .
Setting x = i/d and h = 1/d, we find that for 0 < i < d,

= 0. (3.7)
514 R. Forman

On the other hand

=0.

Therefore

( R(~-z)) -(~-~d~-)-"e)/i-l\ +o(~)0 (3.8)

N o w we examine ~'s initial conditions. F r o m (3.1) and (3.2),

~(a)(O) = 0 (3.9)

and

b= Vf(a)(0) = d(f(a)(~)-f(0))=df(a)(~)-da
da+ O(~)
so that

Choose M large enough so that

~,~,(~)1~~ora~ ~3~0,
[R(x)l ___<M for all x e [ O , 11 . (3.11)

Let c~(d)/d2be an upper bound for the n o r m of the o(1/d2)term in (3.8), so that
l i m e (d) -- 0 .
d--~ cx3

We will see that for all d, and all i such that 0 < i __<d,

e(a) a+d i 1 + - - , (3.12)

+ ~ - ) . (3.13)

Setting i = d, and using

lim 1 + - - =e M+l<m,
d---~ 09
Determinants of Differential Operators 515

we see that (3.12) implies (3.5), and (3.13) implies (3.6). We will prove (3.12) and
(3.13) by induction. From (3.9) and (3.!0) both bounds are true for i = 1. Using (3.8)
we can write
= Ri

Thus, by induction and (3.11),

] (.,fi+ 1~ M M Mi_ M(ex/~_eX/-~i@l)}

+ ii 1 + @ + --I+Ix.
Analysing the two terms separately,

I = M e'fMii ( l + ~ -Me -'/MI1) < d e . j ~ ( e'ffM~, - 1)

Using that / < 1 we have

d-

=<I((I+M)( 1 +~)M§247 + Md~l)= )

These inequalities combine to yield a proof of (3.13). The inequality (3.12) easily
follows from (3.8) and the inductive hypothesis. []

Convergence of Determinants. Let Ro and R1 be continuous functions on [0, 1],

and
d2
Lo - dx 2 '}- Ro(X) ,

d2
Lt - dx 2 + RI(x)

two differential operators acting on functions on [0, 1]. Let ,4 be a boundary

condition given by
c11f(O) + c12f'(0) -I- c13f(1) + c14f'(1) = O,
c21f(0) + c22f'(0) + c23f(1) + c24f'(1) = 0, (3.14)
516 R. Forman

where
Cll C12 C13 C14~
czl czz cz3 c24/
has rank 2.
Let L(oa) and L~a) denote the finite difference approximations of Lo and
LL corresponding to the partition of [0, 1] into d equal subintervals. Approximate
A by the finite difference boundary condition

+ (3.15)

We will also denote the approximation by A. We will prove that

det L~a,)A
lim ~ = det LI,A(Lo,A)- 1 . (3.16)
d--.r oo ~wL x-~0, A

To describe our method of proof, for # ~ C, let

d2
Lu = dx 5 + Ro(x) + # ( R l ( x ) - Ro(x)) ,

and define
D(#) = det Lu, A(Lo,A)- 1 ,
r(a)
det L,, a
D(a)(#) = ~ .

Note that for each d

D(O) = D(a)(O) = i . (3.17)
The function D(#) is holomorphic (see [G-K]), and non-trivial, by (3.17), and thus
has isolated zeros. On the other hand, for each d, D(a)(#) is a polynomial, and hence
is holomorphic. The analysis of Sect. 2 shows that D(a)(#) can be expressed in terms
of the boundary values of the solutions to the initial value problem for L(a)(#).
Theorem 3.1 implies that as d ~ 0% the boundary values, and hence D(a)(#),
converge uniformly on compact sets in the #-plane. Thus
lim D(a)(#) (3.18)
d'-* oo

is holomorphic, and, by (3.17), is non-trivial. Therefore, the zeros of (3.18) are

discrete. We prove (3.16) by proving that for each # such that
D(#) #: 0 =~ lim D{d)(#)
d--* co

we have the equality

~log D(#) = --dlog lira D { a ) ( # ) 9 (3.19)

d# a-~o
Determinants of Differential Operators 517

We first examine the left-hand side of (3.19),

logD(#) = tr ~ L , , A (L,,A) -~ = tr(Rl(x) -- Ro(x))(L~,,a) -1 (3.20)

(see [G-K]). The operator (Lu.A)-t has a continuous kernel Gu, a(x, y), such that
for any continuous function f ~ A,
1

(Lu, a)-* f = j G~,a(x, y)f(y)dy .

0
~

Equivalently, G~,,A(X, y) is uniquely determined by the following 4 properties:

(i) Fix y e (0, 1). Then, as a function of x,
Gu,A(X , y ) ~ A .
(ii) For x 4= y
Lu Gu, A(x, y) = 0
(where L u acts on G as a function of x).
(iii) The kernel Gu,a(x , y) is continuous on [0, 1] x [0, 1].
(iv) For all ye(0, I),
d d
lim ~xGu,a(X, y) = lim ~xG,,a(x, y) + 1 .
x-+y~ x~y§

(Note that the kernel G is Lipschitz, which implies L~-1 is a nuclear operator
([G-K]) and thus the discussion in [G-K], Chap. IV, applies.)
Therefore,
(RI(x) -- Ro(x))(Lu,A)- 1
has the continuous kernel
(Rl(x) - Ro(X))G~,Ax, y)
with trace
1

y (/l(X) -- Ro(x))Gu,A(x, y) dx
0

1 a-*((~) (~)) (i ~)
= a-~lim~ _ ~,=1"= R1 - Ro Gu, A -~, . (3.21)

Now we consider the right-hand side of (3.19). The uniform convergence of D(a)(/0
implies

# \ a-~oo a-,~ d-~#l~ D(e)(#)

,-1
= a-+~lim ~d# u,AJt~'u,a~ 9

The operator (L(ua)A)- 1 can be expressed as a matrix

(3.22)
518 R. Forman

This matrix is uniquely determined by the conditions:

(i) For every v ~ A,
MveA and L~ue,)AMv=v.
(ii) For every v~A l,
Mv=O.
Now

d_ L(e) d
.,A = ~ = A L ~ ~ = ~A(R1 -- R o ) ,
d•
where R~ -- Ro is the diagonal matrix with
i=0 ord
(3.23)
.,(:) .00)
Lastly, we consider the matrix representing rcA. Let
l<i<_d-l.

Vf(O) az bz
= span (3.24)
f(1) 123 ' b3
a4 b4

denote the solution space to (3.15). Then (3.24) is equivalent to

rio) \ al \ lb,\
4) = span
al + d
bl + - ~

f(1) 123 b3

a4 b4
a 3 -- -~ b3 - d-i

= span

f( (b)t
b13
b3
al

a3

a3
+o +

span{(i) (!)t +O
Determinants of Differential Operators 519

as long as

det(; I bb:)@O" (3.25)

We will now assume this. It is not true, the necessary modifications in the argument
are clear, and, in fact, lead to some simplification.
Thus, from (3.25),

1,~ O)
7r,4=
11+0
0 o0) t
0 l 0 (3.26)

o0) o l, 0) 22
l~+O

Combining (3.22), (3.23) and (3.26) we see

1 (d i

+ 20
(i,j)~s O) (' ')
M 2'? ' (3.27)

where
s = {(1, 0), (1, 1/, (1, d - 1), (1, d), (d - 1, 0), (d - 1, 1), (d - 1, d - 1), (d - 1, d ) } .
Comparing (3.20) to (3.27) we see that the desired equality (3.19) follows from the
two lemmas:
L e m m a 3.2. There is a constant c (independent of d) such that for i, j e
{o, ~, d - 1,d},
.e)/i i\l

L e m m a 3.3. a) For j = 1, and d - 1,

i~(~) ~o(~) =o0)

b) F o r 2 < j < d - 2
i i 1 i

where the above bounds are uniform in i and d.

We will see that Lemmas 3.2 and 3.3 reduce to
Lemma 3.4. I f La is invertible, then there is a constant c > 0 (independent old) such
that for all d large enough
IIL~d)vlloo ~ cllvlloo for all v ~ A
520 R, Forman

We assume Lemma 3.4 for the moment, and prove Lemmas 3.2 and 3.3.

P r o o f o f Lemma 3.2. From (3.25) there are two vectors in A of the form

/1\
1
0
U 1 ~- + (3.28)
0

~0/ \ /

and thus A l is spanned by two vectors of the form

/ / O\

0
1)3 ~--- + 1)4 + / (3.29)
0
1

Thus, from the definition of M (d),

M(a)vl ~ A, (d)
Lu,A(M (4)vl) = Vl ,

r(d) [~/t(d),,
M(a)v2 ~A, ~u,At 1.~ ~2) ~ I)2 ,

M (a)v3 = 0 ,

M(a)v4 = 0 . (3.30)

From (3.28) and Lemma 3.4, there is a c such that

IIMy1 I[| IIM Y 2 II co ~ c 9

Thus we have the following bound (uniformly in i and d)

Determinants of Differential Operators 521

From (3.28) and (3.29)

1
/~(~ o)\
I( 0
1
0
-1
0

0
1
0
1 -
+0

which shows that

/~(~, o)\
= o(1),
M -1)

as desired. []
Proof of Lemma 3.3. a) We consider the case j = 1. Let

i, = ~(Mvl - My3).

From (3.28) and (3.29),

so it is enough to prove

From (3.30),

(3.31)

We know G(x,~)~A (as a function of x). That is,

(O(0,~)~ ~oO(~,~)~0,~),~ ~10(~~))

522 R. Forman

satisfies (3.14). In addition

d

d G x, = V(le)G ~ +0
dx x=l
(where the subscript 1 denotes the difference is taken with respect to the first
variable). Thus

Moreover, as in (3.7), for 2 < i < d - 1,

,/i
#,A ~d'
From Taylor series

G(~,~)=G(d,~)+ ~ lim+dG(x,~) + 1 ,,I/

x --->~-

,or.omo.~(~~)
Thus

lira+ d G ( x , ~ ) = VtG(~,~)-~-~R(zl)G(zl,~).

Similarly

lira

~orsome..~(0,~)
Thus

_1-_ ,i~+_~
~)_~
o(~, l~m~~)~
o(~,

(The error is o(1/d) since both R and G are continuous.)

Determinants of Differential Operators

Therefore,

L~d~ G 3'3 =
(o)
1

o + o(~).
523

(3.32)

0
This implies, using (3.27),

:,(
",A3 3' =~AL. 3 ~ 3'3

(3.33)

Let

Then, from (3.31) and (3.33),

and

(~0,~,0,,~,~,~(~))+0(~) ,334,
satisfies (3.15).
Let g(i/d) denote the unique function which has boundary values equal to the
O(1/d 2) term in (3.34) and is linear in i for 1 < i _< d - 1. Then

~:~-o(:)
Thus,
e -- g 6 A
and
524 R, F o r m a n

F r o m L e m m a 3.4

so that

as desired,

b) The case 2 < j < d - 2 is easier than j = 1, d - 1. Following as above to

(3.32) we learn that

=M 2,? -2~ 2,~

satisfies

0
0

0
so that

(where the 1 is in the jt~ spot), Continuing as in (a) we learn

as desired. []
We now complete this section with a p r o o f of L e m m a 3.4.

Proof of Lemma 3.4. Suppose the l e m m a is false. That is, suppose there is a
sequence
di --~ oo

and
viEA (dd
Determinants of Differential Operators 525

such that

,Iv~j,+ = sup v~(J~l=l, (3.35)

O<=j<d~ \Uil l
[tL(Aa')v~ 11~ -~ 0. (3.36)

Then define a sequence of C 2 functions gi by

L g i = L(aOvi ,

5,(0) = vi(O),
~;(0) = vv,(o).
Then a very slight modification of the proof of Theorem 3.1 yields

~i - Vvl ~ 0 (3.37)

uniformly in i andj. (For this conclusion, we need to know that vi(0) and Vv~(O) are
bounded uniformly in i. The bound on vi(0), as well as a bound on vi(1), follows
from (3.35). The bound on Vvi(O) follows from the bounds on vi(O) and v~(1),the fact
that vi ~ A, and our assumption (3.25).)
We note that, by definition,
L(aOvi(O) = L(ai)vi(1) = O .

Thus, as can be seen from the form of UA in (3.26), (3.36) implies

IIL(d')v~[[~ -+ 0
so that, in fact,
[ILf, II0o ~ 0. (3.38)
Since 17;(0) and ~[(0) are bounded, we can find an a, b and a subsequence of the
~ such that
5~(0) --+ a, ~[(0) --+ b.
Then (3.38) implies that these 5i approach ~, uniformly in C 1 [0, 1], where ~ is the
unique solution to
L5=0, 5(0)=a, 5'(0)=b.
The uniformity of the convergence implies
~,(1) -~ ~(1), ~7~(1) ~ Y(1).
Comparing with (3.37) we learn
v,(o) + ~(o) w,(o) + ~'(o),

vi(1) ~ ~(1)
526 R. Forman

Since each v~eA,

we must have g e A .
Lastly, we note that the uniform b o u n d

]'v~l'~>supj v i ( ~ ) =supi v,(~)+o(1)[

= 1 + o(1)

implies
g#O.
Thus g ~ A is a non-trivial solution to Lg = O, which c o n t r a d i c t s the hypothesis of
the lemma. []

References

[B-F-K] Burghelea, D., Friedlander, L., Kappeler, T.: On the Determinant of Elliptic Differential
and Finite Difference Operators in Vector Bundles over S 1. Commun. Math. Phys.
(1991) (to appear)
[B-D] Bott, R., Duffin, R.: On the Algebra of Networks. Trans. Am. Math, Soc. 74, 99-103
(1953)
[D-D] Dreyfuss, T., Dym, H.: Product formulas for the eigenvalues of a class of boundary
value problems. Duke Math. J. 45, 15-37 (1978).
[Fol] Forman, R.: Functional Determinants and Geometry. Inv. Math. 88, 447-493 (1987)
[Fo2] Forman, R.: Difference Operators, Covering Spaces and Determinants. Topology 28,
413-438 (1989)
[Fo3] Forman, R.: Determinants of Laplacians on Graphs, preprint
[Fr] Franklin, P.: The Electric Currents in a Network. J. Math. Phys. 4, 9%102 (1925)
[G-K] Gohberg, I.C., Krein, M.G.: Introduction to the theory of linear nonselfadjoint oper-
ators; Vol. 18 (Translation of Mathematical Monographs). Providence, Rhode Island:
American Mathematical Society 1969
[Ki] Kirchhoff, G.: Uber die Aufl6sung der Gleichungen auf welche man beider Untersuchen
der Linearen Verteilung galvanischer Str6me geffihrt wird. Annal. Physik Chemic 72,
495-508 (1847)
[L-S] Levit, S., Smilansky, U.: A theorem on infinite products of eigenvalues of Sturm-
Liouville type operators. Proc. Am. Math. Soc. 65, 299-302 (1977)
[Mi] Milnor, J.: Morse Theory. Annal Math. Stud., Vol. 51. Princeton, NJ: Princeton
University Press 1963
[Ra] Ramond, P.: Field Theory, A Modern Primer, Frontiers in Physics, Vol. 51. Reading,
MA: Benjamin/Cummings 1981
[R-S] Ray, D.B., Singer, I.M.: R-Torsion and the Laplacian on Riemannian manifolds. Adv.
Math. 7, 145-210 (1971)

Communicated by A. Jaffe