Computation of determinants using contour integrals

Klaus Kirsten

Department of Mathematics, Baylor University, Waco, TX 76798

Paul Loya

Department of Mathematics, Binghamton University,

Vestal Parkway East, Binghamton, NY 13902
Abstract
It is shown how the pre-exponential factor of the Feynman propagator for a large class
of potentials can be computed using contour integrals. The prerequisites for this analysis
are accessible to advanced undergraduate students and involve only introductory courses in
ordinary dierential equations and complex variables.

Electronic address: klaus_kirsten@baylor.edu

Electronic address: paul@math.binghamton.edu

1
I. INTRODUCTION
The spectrum of certain dierential operators encodes fundamental properties of
dierent physical systems. Various functions of the spectrum, the so-called spectral
functions, are needed to decode these properties. One of the most prominent spec-
tral functions is the zeta function, which relates for example to partition sums, the
heat-kernel and the functional determinant; see, e.g., [? ]. Zeta functions are often
associated with suitable sequences of real numbers {
k
}
kIN
, which, for many applica-
tions, are eigenvalues of Laplace-type operators. In generalization to the zeta function
of Riemann,

R
(s) =

k=1
k
s
, (1.1)
one denes
(s) =

k=1

s
k
, (1.2)
where s is a complex parameter whose real part is assumed to be suciently large such
as to make the series convergent.
To indicate how zeta functions relate to other spectral functions, let us use the
functional determinant as an example because it is going to be the focus of the article.
For the purpose of relating zeta functions and determinants assume for the moment
that we talk about a sequence of nitely many numbers {
k
}
n
k=1
. Considering them as
eigenvalues of a matrix L, we have
det L =
n

k=1

k
,
which implies
ln det L =
n

k=1
ln
k
=
d
ds

s=0
n

k=1

s
k
.
In the notation of eq. (1.2) this shows
ln det L =

(0) or det L = e

(0)
. (1.3)
2
When the nite dimensional matrix is replaced by a dierential operator L, in general

k=1

k
will not be dened. However, as it turns out for many situations of rele-
vance, denition (1.3) makes perfect sense and has found important applications in
mathematics and physics; for the rst appearance of denition (1.3) see [? ? ? ].
In recent years a contour integral method has been developed for the analysis of
zeta functions [? ? ], which, although applicable in any dimension and to a variety of
spectral functions, shows its full elegance and simplicity when applied in one dimension,
and when applied to functional determinants. One of the main reasons for the relevance
of determinants is the fact that the evaluation of the Feynman propagator involves
this quantity. Because of its relevance a considerable number of articles in American
Journal of Physics has been devoted to this topic; see, e.g., [? ? ? ? ? ], and so we
decided to also concentrate on this topic. Our aim is to show how and why a contour
integral method is extremely well adapted for the evaluation of in particular functional
determinants. By demonstrating an additional way by which results may be obtained
we enlarge the arsenal of techniques by a component already proven useful in recent
research, see, e.g., [? ? ]. The probably most attractive feature of our approach is
that all prerequisites are known to advanced undergraduate students of physics and
mathematics. Namely, we only assume some working knowledge with Cauchys residue
theorem [? ] and some elementary facts of about ordinary dierential equations [? ].
The outline of this article is as follows. We explain the basic ideas of our approach
by looking at the zeta function of Riemann, and by evaluating

R
(0). This is identical
to the evaluation of the functional determinant of a free particle in an interval with
Dirichlet boundary conditions at the endpoints. We then will consider the case of
particles in a harmonic oscillator potential previously considered in [? ? ? ]. Results
will be trivially rederived. Finally, we show how particles moving in any potential
(satisfying reasonable conditions) and obeying quite general boundary conditions can
be analyzed. The Conclusions highlight the most important points of our contribution.
3
II. FUNCTIONAL DETERMINANT OF A FREE PARTICLE IN AN INTER-
VAL
A free particle in an interval is described by the operator O =
d
2
dt
2
together with some
boundary condition. In the context of the Feynman propagator Dirichlet boundary
conditions are quite common [? ] and this is what we rst concentrate on. It will
be convenient to make a rotation in the complex t-plane and to dene t = i. The
resulting operator P =
d
2
d
2
in terms of has positive eigenvalues. At the end of the
analysis we will have to undo this transformation of course.
So in order to evaluate the functional determinant associated with this situation we
consider the eigenvalue problem

d
2
d
2

n
() =
n

n
(),
n
(0) =
n
(L) = 0.
The eigenfunctions have the form

n
() = a sin(
_

n
) + b cos(
_

n
).
The appearance of the cosine is excluded by the boundary value
n
(0) = 0. The
eigenvalues are then found from the equation
sin(
_

n
L) = 0. (2.1)
This condition is simple enough to be solved for analytically and one determines

n
() = a sin(
_

n
),
n
=
_
n
L
_
2
, n IN,
with some normalization constant a.
Although in this particular case it is of course convenient to have an explicit expres-
sion for the eigenvalues, let us pretend the best we are able to obtain is an equation
like (2.1), namely eigenvalues are determined as the zeroes of some function F(). As
we will see, actually this is as convenient as having explicit eigenvalues, but of much
larger applicability.
For the given setting the natural choice F() = sin(

L) has to be modied as
= 0 satises F(0) = 0. In order to avoid F() having more zeroes than there are
4
actual eigenvalues we therefore dene
F() =
sin(

L)

. (2.2)
The next step in the contour integral formalism is to rewrite the zeta function using
Cauchys integral formula. Given that F() = 0 denes the eigenvalues
n
, then
d
d
ln F() =
F

()
F()
has poles exactly at those eigenvalues. Furthermore, expanding about =
n
we see
for F

(
n
) = 0 that
F

()
F()
=
F

(
n
+
n
)
F(
n
+
n
)
=
F

(
n
) + (
n
)F

(
n
) + ...
(
n
)F

() + (
n
)
2
F

(
n
) + ...
=
1

n
+ ...
and the residue at all eigenvalues is 1. (A variation of this argument shows that if m
n
is the multiplicity of
n
, the residue of F

()/F() at
n
is m
n
.) This shows, noticing
the appropriate behavior of F() at innity, that for s >
1
2
,

P
(s) =
1
2i
_

d
s
d
d
ln F(), (2.3)
where the contour is shown in Figure 1.
-
6
r r r r r r r r r r
-plane

FIG. 1: Contour
As is typical for complex analysis, the next step in the evaluation of a line integral is
a suitable deformation of the contour. Roughly speaking, deformations are allowed as
5
long as one does not cross over poles or branch cuts of the integrand. For the integrand
in (2.3), the poles are on the positive real axis, and there is a branch cut of
s
which
we dene to be on the negative real axis, as is customary. So as long as the bahavior at
innity is appropriate, we are allowed to deform the contour to the one given in Figure
2.
-
6
r r r r r r r r r r
-plane
-
FIG. 2: Contour after deformation
In order to better see the || behavior of F(), let us rewrite the sine in terms
of exponentials. We then have
F() =
1
2i

_
e
i

L
e
i

L
_
,
and for s >
1
2
all deformations are indeed allowed. We next want to shrink the contour
to the negative real axis. As approaches the negative real axis from above,
s
picks
up a phase (e
i
)
s
= e
is
, whereas as the limit from below produces (e
i
)
s
=
e
is
. Given the opposite direction of the contour above and below the negative real
axis, contributions add up to produce a sin(s). Taking care of the same kind of
argumentation in F() one obtains

P
(s) =
sin s

_
0
dx x
s
d
dx
ln
_
e

xL
2

x
_
1 e
2

xL
_
_
. (2.4)
6
Notice, that by shrinking the contour to the negative real axis a new condition for the
integral to be well dened, namely s < 1, has become necessary due to the behavior
about x = 0.
Let us rest for a moment to stress the nice features of equation (2.4) for the eval-
uation of determinants. If the integral were nite at s = 0, an evaluation of the
determinant would be trivial. It would amount to nding ln(...) at the limits of in-
tegration; even no integration needed to be done explicitly. Whereas this is exactly
what occurs when considering ratios of determinants, see Section 4, for absolute de-
terminants the situation is slightly more complicated. The reason is that (2.4) is only
well dened for
1
2
< s < 1 and a little more eort is needed. Notice, the problem is
caused by the x behavior which enforces the condition
1
2
< s. To analyze this
further we therefore split the integral accoding to
_
1
0
dx +
_

1
dx. Whereas from the
above remarks it follows that
_
1
0
dx can be considered to be in nal form, the
_

1
dx
needs further manipulation. The pieces needing extra attention are

_
1
dx x
s
d
dx
ln e

xL
=
L
2

_
1
dx x
s
1
2
=
L
2s 1
,

_
1
dx x
s
d
dx
ln
_
1
2

x
_
=
1
2

_
1
dx x
s1
=
1
2s
.
This shows

P
(s) =
Lsin s
2s 1

sin s
2s
+
sin s

_
1
dx x
s
d
dx
ln
_
1 e
2

xL
_
+
sin s

1
_
0
dx x
s
d
dx
ln
_
e

xL
2

x
_
1 e
2

xL
_
_
,
a form perfectly suited for the evaluation of

P
(0). We nd

P
(0) = L 0 ln
_
1 e
2L
_
+ ln
_
e
L
2
_
1 e
2L

_
ln L = ln(2L).
This, of course, agrees with the answer found from the well known values
R
(0) =
1
2
,

R
(0) =
1
2
ln(2):

P
(s) =

n=1
_
n
L
_
2s
=
_
L

_
2s

R
(2s)
7
=

P
(0) = 2 ln
_
L

R
(0) + 2

R
(0) = ln
_
L

_
ln(2) = ln(2L). (2.5)
III. FUNCTIONAL DETERMINANT FOR PARTICLES IN A HARMONIC
OSCILLATOR POTENTIAL
Let be the frequency of the harmonic oscillator, then the relevant operator to be
considered is
P
ho
=
d
2
d
2
+
2
,
with Dirichlet boundary conditions imposed at the endpoints = 0 and = L. Eigen-
values are then determined by the implicit equation
sin(
_

2
L) = 0. (3.1)
Instead of looking at the determinant of P
ho
itself, let us now consider the ratio
det(P
ho
)/ det(P), that is we consider the dierence of the associated zeta function.
Using the same strategy as before in Section 2, we have

P
ho
(s)
P
(s) =
1
2i
_

d
s
d
d
ln
_
sin(

2
L)
sin(

L)

2
_
,
the contour still given by Figure 1. Deforming as before, we obtain

P
ho
(s)
P
(s) =
sin s

_
0
dx x
s
d
dx
ln
_
sinh(

x +
2
L)
sinh(

xL)

x +
2
_
, (3.2)
where sin(iy) = i sinh y has been used. The technically simplifying consequence of con-
sidering ratios gets now apparent: as x tends to innity, the behavior of the integrand
has improved. In detail we have as x
sinh(

x +
2
L)
sinh

xL

x +
2
= e
L(

x+
2

x)

x +
2
1 e
2L

x+
2
1 e
2L

x
= 1 +
1
2

2
L

x
+ ...
and the integrand behaves like x
s3/2
. Noting that the x 0 behavior up to a
proportionality constant is as before, we see that (3.2) is well dened for
1
2
< s < 1,
in particular, it is well dened at s = 0. Thus, trivially,

P
ho
(0)

P
(0) = ln
_
sinh L
L
_
,
8
or, switching back to real time, replacing L = i(t
f
t
i
),
ln
det P
ho
det P
= ln
sinh(i(t
f
t
i
))
i(t
f
t
i
)
= ln
sin((t
f
t
i
))
(t
f
t
i
)
, (3.3)
the well known answer; see e.g. [? ? ].
Other boundary conditions can be dealt with basically with no extra eort. For
example let us consider quasi-periodic boundary conditions as they have been analyzed
for anyon-like oscillators [? ? ]. In this case the boundary condition reads

n
(L) = e
i

n
(0),

n
(L) = e
i

n
(0),
with some real parameter; = 0 corresponds to periodic boundary conditions,
whereas = gives antiperiodic boundary conditions typical for fermions. The general
form of eigenfunctions is

n
() = a sin
_
_

_
+ b cos
_
_

_
.
The boundary condition produces the equations, use
n
=

2
,
a sin
n
L + b cos
n
L = e
i
b,
n
b sin
n
L +
n
a cos L = e
i
a.
Under the assumptions that there are no zero eigenvalues,
n
= 0, and that sin
n
L = 0,
which excludes periodic boundary conditions for the following, we rst nd from the
rst equation
a =
b
sin
n
L
_
e
i
cos
n
L
_
,
which, when substituted into the second equation, after some simple manipulations
gives the condition for eigenvalues as
cos
n
L cos = 0.
Following the steps of the previous calculation, denoting the operator with quasi-
periodic boundary conditions as P
qp
ho
and P
qp
, the answer can essentially be simply
read o,

qp
P
ho

(0)
qp
P

(0) = ln
_
cosh L cos
1 cos
_
,
9
and agrees with [? ]. So in real time,
ln
det P
qp
ho
det P
qp
= ln
cos((t
f
t
i
)) cos
1 cos
.
For periodic boundary conditions an eigenfunction with zero eigenvalue occurs, namely
the constant, and we comment on this situation in the conclusions.
IV. FUNCTIONAL DETERMINANTS OF PARTICLES IN GENERAL PO-
TENTIALS
As the previous section made clear, the answer was obtained without ever worrying
what the actual eigenvalues of the operator in question might be. The only information
that entered was the implicit eigenvalue equation (3.1). Is there any way an equation
like (3.1) can be obtained for general potentials, such that the evaluation of determi-
nants is similarly trivial than the previous one? The answer is yes and elementary
knowledge of ordinary dierential equations is all that is needed [? ].
So let us say we were interested in the ratio of determinants of operators of the type
P
j
=
d
2
d
2
+ R
j
(), j = 1, 2,
where for convenience again Dirichlet conditions are considered. In the previous sec-
tions R
2
() = 0 was chosen, but no additional complication arises for this more general
case. As suggested by the previous examples the contour integral to be written down
should involve solutions to the equation
P
j

j,
() =
j,
(),
where , for now, is an arbitrary complex parameter. As is well known, for continuous
potentials R
j
() there will be two linearly independent solutions and every initial value
problem
j,
(0) = a,

j,
(0) = b, will have a unique solution. A contact with the original
boundary value problem is established by imposing
j,
(0) = 0; the condition on the
derivative is merely a normalization and for convenience we choose

j,
(0) = 1. The
eigenvalues for the boundary value problem are then discovered by imposing

j,
(L) = 0, (4.1)
10
considered as a function of . To see a little better how this works, consider the case
R
2
= 0. The unique solution of the initial value problem described is

2,
() =
sin

.
Eigenvalues follow precisely from the condition

2,
(L) = 0.
But having the implicit eigenvalue equation (4.1) at our disposal, the calculation is
basically done! Arguing as below (2.2) we write

P
1
(s)
P
2
(s) =
1
2i
_

d
s
d
d
ln

1,
(L)

2,
(L)
=
sin s

_
0
dx x
s
d
dx
ln

1,x
(L)

2,x
(L)
,
valid about s = 0 because the leading behavior as x of
j,x
(L) does not depend
on the potential R
j
(); as evidence see the analysis in Section 3. But so

P
1
(0)

P
2
(0) = ln

1,0
(L)

2,0
(L)
and
det P
1
det P
2
=

1,0
(L)

2,0
(L)
.
The ratio of determinants is determined by the boundary value of the solutions to the
homogeneous initial value problem
_

d
2
d
2
+ R
j
()
_

j,0
() = 0,
j,0
(0) = 0,

j,0
(0) = 1.
Even is no analytical knowledge about the boundary value might be available, they
can easily be determined numerically.
V. CONCLUSIONS
The main aim of this contribution was to show that the analysis of functional deter-
minants for a large class of operators is accessible to advanced undergraduate students.
11
The only prerequisites are elementary ordinary dierential equation theory and a basic
course in complex variables. The beauty of the approach is that it is easily adapted to
dierent situations. We have indicated how other boundary conditions than Dirichlet
ones can be dealt with. Indeed, general boundary conditions can be considered along
the same lines [? ].
We have mentioned that the presence of zero eigenvalues adds some extra com-
plication. The reason is that when deforming the contour to the negative real axis,
some contribution from the origin may result. But again, a minor modication of the
procedure allows for a complete analysis [? ].
Even systems of dierential equations can be considered with about the same eort
[? ].
The many advantages of this approach described show that it is optimally adapted
to the evaluation of determinants. Instead of struggling with the needed mathematical
manipulations the students should be able to easily get a grasp of this technique and
to concentrate on the underlying physics.
12