Zeta Function Regularization PDF

Zeta Function Regularization
Nicolas M Robles
Department of Theoretical Physics

Imperial College London
September 25th 2009
Supervior: Dr Arttu Rajantie
Submitted in partial fulfillment

of the requirements for the degree of Master of Science in
Quantum Fields and Fundamental Forces of Imperial College London
Abstract
The technique of zeta (ζ) regularization is reviewed in quantum mechanics and field theory.
After introducing the Riemann and Hurwitz ζ functions rigorously we compute the partition
function of bosonic and fermionic harmonic oscillators in quantum mechanics and study the
generating functional of a quantum field in the presence of a source J by considering the
determinant of its corresponding differential quadratic operator at 1-loop order.
The invariance of the potential in the bare Lagrangian of the φ4 theory at the solution of
the classical equations with source J is proved and explained and the partition function
of the harmonic oscillator in field theory is computed and explained in the limit of high
temperature.
The link between the ζ function, the heat kernel and the Mellin transform is explained and
the equivalence between ζ and dimensional regularizations is shown and explicitly derived
for the case φ4 theory.
Furthemore, the transformation on the 1-loop effective Lagrangian is also illustrated.
Finally, the Casimir effect is introduced.
1
The formula
1
1 + 1 + 1 + 1 + ··· = −
2
has got to mean something.
Anonymous
(...) the use of the procedure of analytic continuation through the zeta function requires a
good deal of mathematical work. It is no surprise that [it] has been often associated with
mistakes and errors.
E. Elizalde and A. Romeo

We may - paraphrasing the famous sentence of George Orwell - say that ’all mathematics
is beautiful, yet some is more beautiful than the other’. But the most beautiful in all math-
ematics is the zeta function. There is no doubt about it.
Krzysztof Maslanka
2
Acknowledgements
I should like to thank firstly my advisor Arttu Rajantie for accepting the supervision of this
thesis. His patience and comments have been most invaluable. In particular his suggestion
to go beyond the description of the zeta regularization technique and explore comparisons
with dimensional regularization proved to be the most challenging, yet rewarding, aspect
of this dissertation.
I was lucky enough to learn quantum field theory from Daniel Waldram and Arttu Ra-
jantie. Colin Atkinson, FRS, taught me complex variables. The theory of the Riemann
zeta function was wonderfully introduced to me by Marianna Csrnyei in a course on ana-
lytic number theory at UCL. To all of them I am grateful.
I am also extremely thankful to the MSc directors Jonathan Halliwell and Kellogg Stelle
for their advice and help throughout the programme.
Furthermore, my future PhD supervisor in the mathematics department, Yang Chen, has
always been a constant source of encouragement.
Finally, I would like to thank all other group members - students and staff alike - who
have contributed to the running of the MSc programme and who have created a stimulating
academic atmosphere in the theoretical physics group.
CONTENTS 3
Contents
1 Introduction 5
2 Introduction to the Riemann Zeta Function 11

2.1 The Gamma function Γ(s) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 The Hurwitz function ζ(s, a) . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Analytic continuation and the functional equation of ζ(s, a) . . . . . . . . . 22
2.4 Bernoulli numbers and the value of ζ(0) . . . . . . . . . . . . . . . . . . . . 26
2.5 The value of ζ 0 (0) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.6 The polygamma function ψ (m) (s) . . . . . . . . . . . . . . . . . . . . . . . . 30
2.7 Laurent series of the ζ(s) function . . . . . . . . . . . . . . . . . . . . . . . 31
2.8 Concluding remarks on the disbriution of prime numbers and the zeros of ζ(s) 32
3 Zeta Regularization in Quantum Mechanics 39

3.1 Path integral of the harmonic oscillator . . . . . . . . . . . . . . . . . . . . 39
3.2 Solution of the free particle . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 The bosonic partition function . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Zeta regularization solution of the bosonic partition function . . . . . . . . 45
3.5 Alternative solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.6 The fermionic partition function . . . . . . . . . . . . . . . . . . . . . . . . 47
3.7 Zeta regularization solution of the fermionic partition function . . . . . . . 50
4 Dimensional Regularization 53
4.1 Generating functional and probability amplitudes in the presence of a source J 53
4.2 Functional energy, action and potential and the classical field φcl (x) . . . . 56
4.3 Derivation of φ4 potential at φcl (x) . . . . . . . . . . . . . . . . . . . . . . . 61
4.4 Γ-evaluation of dimensional loop integrals . . . . . . . . . . . . . . . . . . . 67
4.5 Expansion of Z in Eucliean spacetime . . . . . . . . . . . . . . . . . . . . . 70
5 Zeta Regularization in Field Theory 73

5.1 Heat kernels and Mellin transforms . . . . . . . . . . . . . . . . . . . . . . . 73
5.2 Derivation of φ4 potential at φcl using ζ regularization . . . . . . . . . . . . 75
5.3 Coupling constants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.4 Partition functions in field theory . . . . . . . . . . . . . . . . . . . . . . . . 80
5.5 High temperature limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.6 Equivalence of ζ and dimensional regularization in φ4 theory . . . . . . . . 90
6 Casimir Effect 96
6.1 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.2 ζ regulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.3 Experimental evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7 Conclusion 99
A Appendix 103
A.1 Generalized Gaussian integrals . . . . . . . . . . . . . . . . . . . . . . . . . 103
A.2 Grassmann numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
A.3 The Mellin transform and the series expansion of log Γ(s + 1) . . . . . . . . 109
CONTENTS 4
B References 114
B.1 Books . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
B.2 Papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
B.3 Internet Articles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
1 INTRODUCTION 5
1 Introduction
Special functions have arisen constantly and systematically in mathematical and theoretical
physics throughout the XIX and XX centuries.
Indeed, the study of theoretical physics is plagued with special functions: Dirac, Legendre,
Bessel, Hankel, Hermite functions - to name a few - are abundant in the indices of most
modern treatises on physics. The most prominent ones have been the ever-present gamma
function and those which are solutions to differential equations that model physical systems.
The Riemann zeta function is defined for a complex variable s with Re(s) > 1 as [8]
∞
X 1 Y 1
ζ(s) := = .
n=1
n s
p
1 − p−s
The above definition can be extended for Re(s) < 1 as it can be analytically continued to
the whole complex plane except at s = 1 where it has a simple pole with residue 1. It will
be shown that the product runs over all primes p.
It is not a solution to any physically motivated differential equation [1] which sets it apart
from other special functions which have a more transparent physical meaning.
Traditionally, the Riemann ζ function has had its applications in analytic number theory
and especially in the distribution of prime numbers. As such it has been regarded mostly as
a function that fell completely within the realm of pure mathematics and it was temporarily
excluded.
We will begin by studying the Γ and ζ functions. The lack of a course on special functions
at Imperial College gives us a welcome opportunity to discuss these entities thoroughly.
The exposition will be rigorous definition-theorem-proof style and the only essential pre-
requisites are those of complex analysis: convergence, analytic continuation, residue calculus
and Fourier and Mellin transforms.
We have limited the discussion, however, to the essential aspects of the ζ functions that
we will need for the rest of the dissertation and thus commented the beautiful connection
between number theory and the ζ function only briefly.
The functional equation is of particular importance
sπ
ζ(s) = 2(2π)s−1 Γ(1 − s) sin ζ(1 − s),
2
as well as the formulas
1 1
ζ(0) = − , ζ 0 (0) = − log(2π),
2 2
which will be used time and again. Of course, they need a note of clarification. The ζ
function can be written as
∞
1 X (−1)n+1
ζ(s) = 1−s
.
1−2 n=1
ns
These two sums agree in the region where they converge. However, when Re(s) < 1 the RHS
is the step to analytic continuation and it goes to − 21 as s → 0 hence ζ(0) = − 21 in this sense.
During the last quarter of the XX century papers from S. Hawking [4], S. Elizalde, S.
1 INTRODUCTION 6
Odintsov and A. Romeo (EOR) [3] explained how the ζ regularization assigns finite values
to otherwise superficially divergent sums.
This is precisely what the two above formulas for ζ(0) and ζ 0 (0) accomplish.
One of the early uses of the ζ technique was made by Hendrik Casimir in 1948 to compute
the vacuum energy of two uncharged metallic plates a few micrometers apart [2].
Another of the first instances of the Riemann’s ζ function as a summation device comes
from Hawking’s paper [4]. Others before had used this device in connection with the renor-
malization of effective Lagrangians and vacuum energy-momentum tensors T µν on curved
spaces applied to a scalar field in a de Sitter space background. What Hawking accom-
plished was to show that the ζ function could be used as a technique for yielding finite
values to path integrals whose fields are curved. This, in turn, amounts to saying that the
ζ function can be used to compute determinants of quadratic differential operators.
It is interesting to note that at a more academic level however ζ regularization is hardly

ever mentioned in undergraduate quantum mechanics books, nor is it mentioned either in
Peskin and Schroeder [6] or in Weinberg [10] which are some of the standard books on
quantum field theory. It is precisely in QFT where the ζ function becomes apparent as a
serious competitor to dimensional renormalization.
Let us briefly explain how the technique works in broad strokes. The determinant of an
operator A can be written as the infinite product of its eigenvalues λn as [3], [4], [7]
Y X
log det A = log λn = Tr log A = log λn .
n n
The ζ function arises naturally by using

X 1
d X
ζA (s) := TrA−s =

, ζA (s) =− log λn ,
n
λsn ds s=0 n
and the functional determinant of the operator can then be written as

0
det A = exp(−ζA (0)).
When we use this definition we can find finite values for products which are otherwise di-
vergent because the spectrum of their eigenvalues is unbounded.
Now, we have mentioned operators must be differential and quadratic.

In the bosonic quantum mechanical case, we take this operator to be the harmonic oscillator
d2
AQM = − + ω2 .
dτ 2
Using ζ regularization we will show that the partition function of the oscillator defined as
Z(β) = Tr exp(−β Ĥ)
is given by [5] Z
1
Z(β) = dx hx| exp(−βH)|xi = ,
2 sinh(βω/2)
where β is a constant appearing in the partition function which accounts for the discretiza-
tion of time, i.e. β = N ε as N → ∞ and ε → 0 with N being number of time slices and ε
1 INTRODUCTION 7
the size of each time slice. Also, its inverse can be thought of as the temperature by setting
imaginary time β = iT .
A slightly more exotic ζ function is needed for fermions called the Hurwitz ζ function defined
as [8]
∞
X 1
ζ(s, a) = ,
n=0
(n + a)s
with a 6= 0, −1, −2, · · · . Since fermionic theories demand anti-commutation relations we
will need Grassmann numbers and therefore we will explain these in detail but only in so
much as is needed to compute partition functions for fermions.
The same technique applied with this ζ function to the partition function
Z
∗
−βH
= dθ∗ dθ −θ|e−βH | θ e−θ θ ,

Tre
where θ and θ∗ are conjugate Grassmann numbers yields [5]

∞
" 2 #
Y βω βω
Z(β) = 2 1+ = 2 cosh .
n=1
π(2n − 1) 2
In the quantum mechanical case the eigenvalues λn are explicitly known thus the compu-
tation of ζA is relatively easy compared to the field theoretical case. The key step in these
cases is to relate the ζA function of the operator to the standard ζ function, for instance in
the bosonic case [6]
∞ −2s 2s
X nπ β
ζboson (s) = = ζ(2s),
n=1
β π
whereas in the fermionic case
∞ −s s
X 2π(k − 1/2) β
ζfermion (s) = = ζ(s, 1/2).
β 2π
k=1
We will make use of formulas for ζ(0) and ζ 0 (0) showing the necessity of having discussed
the ζ function at length but more importantly also showing that in a certain sense regu-
larization in quantum theory can be thought of as a technique of complex analysis, namely
analytic continuation.
Because determinants of differential quadratic operators arise in field theory through path
integrals in the presence of a source J we have devoted a whole chapter to explain this
construction. This approach is complementary to the QFT/AQFT courses from the MSc.
The only field theory we shall consider is φ4 however it is important to note that ζ regu-
larization can be applied to more complex field theories and even to string theory [3]. We
shall keep this section brief and avoid topics such as Feynman diagrams whenever possible.
Equipped with all the tools we have developed in the preceding section we will see how
the ζ function can be used in field theory. This will be the culmination of the formulas we
have proved in Section 2, the techniques developed in Section 3 and the theory explained
in Section 4. Furthermore, this will encapsulate the spirit of the technique.
Quantum field theories of a scalar field in the presence of an external source J will be
studied. It will be shown how the generating functional or probability amplitude [6]
Z Z
ZE [J] = e−iE[J] = Ω|e−iHT |Ω J = Dφ exp i d4 x(L [φ] + Jφ)

1 INTRODUCTION 8
can be written in Euclidean spacetime as [3], [7]

NE0 exp {−SEuc [φ(x0 ), J]}
ZE [J] = q .
det (−∂¯µ ∂¯µ + m2 + V 00 [φ(x0 )])δ(x1 − x2 )

It is plainly visible that the determinant in the denominator already contains the seed of
the technique, namely a differential quadratic operator. By taking the potential V to be
λ 4
that of the φ4 theory ( 4! φ ), the operator becomes
λ 2
AFT = −∂ 2 + m2 + φ (x).
2 0
In contrast to the quantum mechanical case, the link to the ζA function is much more
complex and makes use of the heat kernel. With appropriate boundary conditions, the
solution to the heat equation
∂
Ax GA (x, y, t) = − GA (x, y, t)
∂t
is given by scalar product
X −tλn
GA (x, y, t) = x|e−tA |y = ψn (x)ψn∗ (y)

e
n
where λn are the eigenvalues of the operator and ψn the orthogonal eigenvectors. The trace
of the solution is Z X
d4 xGA (x, x, t) = e−tλn = Tr[GA (t)],
n
and the link with the ζA function comes from the Mellin transform [4]
Z∞
X 1
ζA (s) = λ−s
n = dtts−1 Tr[GA (t)].
n
Γ(s)
0
We will show that, by taking into account first order quantum corrections, the potential in
the bare Lagrangian of the field theory is renormalized as
λ2 φ4cl φ2cl

λ 25
V (φcl ) = φ4cl + log −
4! 256π 2 M2 6
where φcl is the classical field defined in terms of the source J and ground state Ω as
δ δ
φcl (x) = hΩ|φ(x)|ΩiJ = − E[J] = −i log Z.
δJ(x) δJ(x)
We will then discuss how coupling constants evolve in terms of scale dependence by exploring
the close analogy between field theory and statistical mechanics. Furthermore, we will be
able show that [7]
1 0 1
log Z = ζA (0) = − ωβ − log(1 − e−ωβ )
2 2
is the partition function of QFT harmonic oscillator. We can push the method further by
showing
3
π2 M2 βM 4

1 0 M µβ
log Z = ζB (0) = V + − + γ + log + · · ·
2 12π 90β 3 24β 32π 2 4π
1 INTRODUCTION 9
where
∂2

2 2 2 λ 2
B=− + β −∇ + m + φ̄
∂t2 2 0
in the limit of high temperatue β → 0. The ζ technique will also show how the effective
Lagrangian is affected.
As a means of a check to make sure the technique yields the same results as other reg-
ularization techniques we will derive the same results by using more standard methods such
as the one-loop expansion approach. However, we will do better than this by showing that
the two techniques are equivalent and we will show this explictily for the case of φ4 theory.
EOR devote a substantial amount of their book on the Casimir effect. Consequently there
is a brief introduction to the calculation done by Casimir in the late 1940s. This will con-
stitute the more ’experimental’ aspect of the thesis.
The conclusion contains a clarification on some of the prejudices regarding the use and
ill definiteness of the ζ function as well as a repertoire of analogous equations concerning
the distribution of the zeros of the ζ function, ζ regularization and dimensional regulariza-
tion.
The appendix contains some formulas that did not fit in the presentation and makes the
whole dissertation almost self-contained. References have been provided in each individual
chapter, including pages where the main ideas have been explored. Overall, the literature
on ζ regularization is sparse. The main sources for this dissertation have been Grosche and
Steiner, Nakahara, Kleinert, Ramond and the superb treatise by EOR.
In terms of the knowledge of physics, the only pre-requisite comes from field theory up
to the notes of QFT/AQFT from the MSc in QFFF.
It has been an objective to try to put results from different sources in a new light, by
clarifying proofs and creating a coherent set of examples and applications which are related
to each other and which are of increasing complexity.
Finally, a few remarks which I have not been able to find in the literature are now made
concerning a potential relationship between ζ functions and families of elementary parti-
cles. Different ζ functions come into play in quantum mechanics by computing partition
functions. As we have said above, the Riemann ζ(s) function is used in bosonic parti-
tion functions, whereas the Hurwitz ζ(s, a) function is needed for the fermionic partition
function, i.e.
Riemann ζ function ⇔ bosons Hurwitz ζ function ⇔ fermions.
It would be an interesting subject of research to investigate how the ζ function behaves with
respect to its corresponding particle, and what kind of knowledge we can extract about the
particle given its special ζ function. For instance, a photon has an associated ζ function, and
its partition function can be computed by using known facts of this ζ function, on the other
hand, known facts about the boson might bring clarifications to properties of this ζ function.
The celebrated Riemann hypothesis claims that all the nontrivial zeros of the Riemann
ζ function are of the form ζ( 12 + it) with t real. Another observation I have not been able
1 INTRODUCTION 10
to find would be to consider an operator which brings the ζ function to the form ζ( 12 + it)
and investigate the behaviour of ζ function with respect to this operator. This would mean
a ’translation’ of the Riemann hypothesis and a study of its physical interpretations.
REFENCES
- [1] J. Brian Conrey The Riemann Hypothesis

- [2] H. B. G. Casimir and D. Polder The Influence of Retardation on the London-van der
Waals Forces
- [3] Elizalde, Odintsov, Romeo, Bytsenk and Zerbini Zeta Regularization Techniques with
Applications
- [4] Stephen Hawking Zeta Function Regularization of Path Integrals in Curved Spacetime
- [5] Hagen Kleinert Path Integrals in Quantum Mechanics, Statistics, Polymer Physics and
Financial Markets
- [6] Michael E. Peskin and Daniel V. Schroeder An Introduction to Quantum Field Theory
- [7] Ramond, Pierre Field Theory: A Modern Premier
- [8] Edward Titschmarsh The Theory of the Riemann Zeta Function
- [9] Mikio Nakahara Geometry, Topology and Physics
- [10] Stephen Weinberg The Quantum Theory of Fields
2 INTRODUCTION TO THE RIEMANN ZETA FUNCTION 11
2 Introduction to the Riemann Zeta Function

2.1 The Gamma function Γ(s)
Often in mathematics, it is more natural to define a function in terms of an integral de-
pending on a parameter rather than through power series. The Γ function is one such case.
Traditionally, it can be approached as a Weierstrass product or as a parameter-dependent
integral. The approach chosen to introduce the Γ function follows from the courses in
complex analysis such delivered by Freitag and Busam [2], as well as Titchmarsch [6] and
Whitaker and Watson [7].
We adopt the notation s = σ + it which was introduced by Riemann in 1859 and which
has become the standard in the literature of the ζ function. Let us first define the Euler Γ
function.
Definition 1 The integral

Z ∞
Γ(s) := dtts−1 e−t (2.1)
0
is well defined and defines a holomorphic function in the right half complex plane, where
Re(s) = σ > 0.
The first lemma generalises the factorial function as follows
Lemma 1 For any n ∈ N we R ∞have Γ(n) = (n − 1)!.

Proof. Note that Γ(1) := 0 dte−t = 1 and by integration by parts we have
Z ∞ Z ∞
Γ(s + 1) = dtts e−t = −ts e−t |∞
0 + s dtts−1 e−t = sΓ(s)
0 0
for any s in the right half-plane. Now, for any positive integer n, we have
Γ(n) = (n − 1)Γ(n − 1) = (n − 1)!Γ(1) = (n − 1)! (2.2)
so the result is proved.
In order to have a complete view of the Γ function we need to extend it to a meromor-

phic function in the whole complex plane.
Lemma 2 Let cn where n ∈ Z+ be a sequence of complex numbers such that the sum
P ∞ +
n=0 |cn | converges. Furthermore, let S = {−n|n ∈ Z and cn 6= 0}. Then
∞
X cn
f (s) =
n=0
s +n
converges absolutely for s ∈ C−S and uniformly on bounded subsets of C−S. The function
f is a meromorphic function on C with simple poles at the points in S and the residues are
given by ress=−n f (s) = cn by for any −n ∈ S.
Proof. Let us start by finding upper bounds. If |s| < R, then |s + n| ≥ |n − R| for all
n ≥ R. Therefore, we have |s + n|−1 ≤ (n − R)−1 for |s| < R and n ≥ R. From this we can
deduce that for n0 > R we have

∞ ∞ ∞ ∞
X cn X |cn | X |cn | 1 X
≤ ≤ ≤ |cn |.

n=n s + n n=n |s + n| n=n n − R n − R n=n

0 0 0 0
P
As such the series n>R cn /(s + n) converges absolutely P and uniformly on the disk |s| < R
∞
and defines there a holomorphic function. It follows that n=0 cn /(s + n) is a meromorphic
function
P∞ on that disk with simple poles at the points of S on the disk |s| < R. Thus,
n=0 nc /(s + n) is a meromorphic function with simple poles at the points in S and for
any −n ∈ S we can write
cn X ck cn
f (s) = + = + g(s),
s+n s+k s+n
−k∈S−{n}
where g is holomorphic at −n. From this we see that residues are indeed ress=−n f (s) = cn .
This concludes the proof.
Equipped with this lemma we are in a position to extend the Γ function as we wanted.
Theorem 1 The Γ function extends to a meromorphic function on the complex plane.

It has simple poles at 0, −1, −2, −3, · · · . The residues of Γ at these points are given by
(−1)k
res Γ(s) = , (2.3)
s=−k k!
for any k ∈ Z+ .
Proof. Let us split the Γ function as
Z 1 Z ∞
Γ(s) = dtts−1 e−t + dtts−1 e−t ,
0 1
the second integral converges for any complex s and it is an entire function. Let us expand
the exponential function in the first integral
1 1 ∞ ∞ 1 ∞
(−1)k (−1)k (−1)k
Z Z Z
X X X 1
dtts−1 e−t = dtts−1 tk = dttk+s−1 = ,
0 0 k! k! 0 k! s+k
k=0 k=0 k=0
these operations are valid for s ∈ C as the exponential function is entire and converges
uniformly on compact sets of the complex plane. The Γ function can now be written in a
form where Lemma 2 can be used, i.e.
∞ ∞
(−1)k
Z X 1
Γ(s) = dtts−1 e−t + , (2.4)
1 k! s+k
k=0
for any s in the right half-plane. By Lemma 2, the RHS defines a meromorphic function
on the complex plane with simple poles at 0, −1, −2, −3, · · · . The residues are given as a
direct application of the lemma.
Theorem 2 For s ∈ C any we have
Γ(s + 1) = sΓ(s). (2.5)
Proof. This follows directly from Lemma 1 and Theorem 1.
The three-dimensional representation of the Γ function is shown in Figure 2.1.

Another important function related to the Γ function, and also discovered by Euler, is the
Figure 2.1: |Γ(x + iy)| for −5 ≤ x ≤ 3 and −1 ≤ y ≤ 1
Beta function which we proceed to develop as follows. Let Re(p), Re(q) > 0 and in the
integral that defines the Γ function, make the change of variable t = u2 to obtain
Z ∞ Z ∞
2
Γ(p) = dttp−1 e−t = 2 duu2p−1 e−u .
0 0
In an analogous form we have

Z ∞
2
Γ(q) = 2 dvv 2q−1 e−v .
0
Multiplying these two together we have

Z ∞Z ∞
2
+v 2 ) 2p−1 2q−1
Γ(p)Γ(q) = 4 dudve−(u u v ,
0 0
and switching to polar coordinates u = r cos θ, v = r sin θ, dudv = rdrdθ
Z∞ Zπ/2
2
Γ(p)Γ(q) = 4 drdθ e−r r2(p+q)−1 cos2p−1 θ sin2q−1 θ
0 0
  π/2 
Z∞

Z
2
= 2 dr e−r r2(p+q)−1  2 dθ cos2p−1 θ sin2q−1 θ
 
0 0
Zπ/2
= 2Γ(p + q) dθ cos2p−1 θ sin2q−1 θ.
0
The integral can be simplified by setting z = sin2 θ

Zπ/2 Z1
2p−1 2q−1
2 dθ cos θ sin θ = dz z q−1 (1 − z)p−1 .
0 0
Next we define
Z1
B(p, q) := dz z p−1 (1 − z)q−1
0
for Re(p), Re(q) > 0, and this gives the identity
Γ(p)Γ(q)
B(p, q) = B(q, p) = .
Γ(p + q)
We denote by B the Beta function. Moreover, if 0 < x < 1 we have
Z1
Γ(x)Γ(1 − x)
Γ(x)Γ(1 − x) = = B(x, 1 − x) = dz z x−1 (1 − z)−x .
Γ(1)
0
We will evaluate this integral with an appropriate contour, but first we need to make one
last change z = u/(u + 1) which yields
Z1 Z∞ −x Z∞
ux−1 ux−1

x−1 −x du u
dz z (1 − z) = 1− = du .
(u + 1)2 (u + 1)x−1 u+1 1+u
0 0 0
Lemma 3 For 0 < y < 1 we have

Z∞
u−y π
du = . (2.6)
1+u sin πy
0
Proof. Let us use a keyhole contour, which is accomplished by cutting the complex plane
along the positive real axis. On this region we define the function
s−y
f (s) =
1+s
with argument of s−y equal to 0 on the upper side of the cut. Furthermore, the function f
has a first order pole at s = −1 with residue e−iyπ . See Figure 2.2 below.
We are now to integrate this function along the path described in Figure 2.2: the path goes
along the upper side of the cut from > 0 to R, then along the circle CR of radius R centred
at the origin, then along the side of the cut from R to and at the end around the origin
via the circle C of radius also centred at the origin. An application of the Cauchy residue
theorem gives
ZR ZR
u−y z −y u−y z −y
I I
−πiy
2πie = du + dz − e−2πiy du − dz .
1+u 1+z 1+u 1+z
ε CR ε Cε
Figure 2.2: C = CR ∪ Cε ∪ [ε, R] ∪ [R, ε]
We can get rid of the integrals around the arcs by appropriate estimates. Note that for
z 6= 0 we have the following
|z −y | = |e−y log z | = e−yRe(log z) = e−y log |z| = |z|−y ,

−y −y
z
≤ |z|
|z|−y

1 + z |1 + z| ≤ ,
|1 − |z||
and the integrals can be estimated

−y 1−y −y 1−y
I I
z ≤ 2π R
z ≤ 2π ε
dz → 0, dz → 0;

1 + z R − 1 R→∞
1 + z 1 − ε ε→0
CR Cε
so that we are left with

Z∞
−2πiy u−y
(1 − e ) du = 2πie−πiy .
1+u
0
We may re-write this to obtain the final result

Z∞ Z∞
πiy −πiy u−y u−y π
(e −e ) du = 2πi ⇒ du = .
1+u 1+u sin πy
0 0
This proves the claim of the lemma.
Let us now make the concluding remarks.
Theorem 3 (Euler Reflection Formula) For all s ∈ C one has

π
Γ(s)Γ(1 − s) = . (2.7)
sin πs
Proof. The lemma we have just proved can be written as

π π
Γ(x)Γ(1 − x) = = ,
sin π(1 − x) sin πx
for 0 < x < 1. However, both sides of the equation above are meromorphic, hence we have
proved the theorem.
Corollary 1 One has

Z∞ Z∞
1 −1/2 −t 2 √
Γ = dt t e = dt e−t = π. (2.8)
2
0 −∞
Proof. This follows by substituting s = 1/2 in the Euler reflection formula (2.7).
Theorem 4 The Γ function has no zeroes.

Proof. Since s → sin(πs) is an entire function, the RHS of Theorem 3 has no zeroes,
therefore Γ(s) = 0 only happens where s → Γ(1 − s) has poles. However, as we have argued
before, the poles of Γ are at 0, −1, −2, −3, · · · so it follows that Γ(1 − s) must have poles at
1, 2, 3, · · · . By the factorial formula, Γ(n + 1) = n! 6= 0 and so Γ has no zeroes.
Intrinsically connected to the Γ function is the Euler γ constant. Let us first define it
and prove its existence.
Lemma 4 If sn := 1 + 12 + · · · + n1 − log n, then limn→∞ sn exists. This limit is called

the Euler γ constant.
1
Proof. Consider tn = 1 + 12 + · · · + n−1 − log n geometrically. It represents the area of the
Rn
n − 1 regions between the upper Riemann sum and the exact value of 1 dx x−1 . Therefore
tn increases with n. We can write
n−1
X ∞
1 k+1 lim
X 1 1
tn = − log , n→∞ tn = − log 1 + .
k k k k
k=1 k=1
The series on the right converges to a positive constant since

1 1 1 1 1 1
0 < − log 1 + = 2 − 3 + 4 − ··· ≤ 2.
k k 2k 3k 4k 2k
Next, the following holds

1 1 1 1
sn+1 − sn = − log 1 + , tn+1 − tn = − log 1 +
n+1 n n n
which means that

1 1 1
< log 1 + < ⇒ sn+1 − sn < 0 < tn+1 − tn .
n+1 n n
Convergence now follows because sn decreases monotonically whereas tn increase monoton-

ically and the differences are negative. Hence sn is monotonically decreasing and bounded
below thus convergent.
The value of γ was computed by Mascheroni in 1790 to be
γ = 0.57721566490153286060651209008240243104215933593992 · · ·
and of course subsequent improvements have been made.
Lemma 5 Weierstrass product of the Γ function is

n
1 Y s −s/k
= lim sesγ 1+ e . (2.9)
Γ(s) n→∞ k
k=1
Proof. Using the fact that (1 − t/n)n → e−t as n → ∞ it can be shown that
Zn n Zn
s−1 t 1
Γ(s) = lim dt t 1− = lim n dt ts−1 (n − t)n ,
n→∞ n n→∞ n
0 0
and integrating by parts yields

Zn Zn
1 n s n−1 1 n(n − 1) · · · 1
Γ(s) = lim n dt t (n − t) = lim n dt ts+n−1
n→∞ n s n→∞ n s(s + 1) · · · (s + n − 1)
0 0
ns

1 2 n
= lim ··· .
n→∞ s s+1 s+2 s+n
Inverting both sides
n
1 s s Y s
= lim sn−s (s + 1) 1 + ··· 1 + = lim sn−s 1+ .
Γ(s) n→∞ 2 n n→∞ k
k=1
In order to compute the limit we need to insert the convergence factor e−s/k to obtain
n
1 Y s −s/k
= lim sn−s es(1+1/2+···+1/n) 1+ e
Γ(s) n→∞ k
k=1
" n #
s(1+1/2+···+1/n−log n)
Y s −s/k
= lim e s 1+ e .
n→∞ k
k=1
However by the use of Lemma 4, we know that the sum converges to γ so that we have
shown the result (2.9).
The derivative of the Γ function at −1 is Γ0 (1) = −γ, as it can be seen by taking the
logarithmic derivative of Weierstrass product (2.9)
∞ h ∞
Γ0 (s)

X s si 1 X 1 1
− log Γ(s) = log s + γs + log 1 + − ⇒− = +γ+ − ,
k k Γ(s) s k+s k
k=1 k=1
and hence
∞
0
X 1 1
Γ (1) = −1 − γ − − = −γ. (2.10)
k+1 k
k=1
2.2 The Hurwitz function ζ(s, a)

The Hurwitz ζ(s, a) function is initially defined for σ > 1 by the series
∞
X 1
ζ(s, a) := , (2.11)
n=0
(n + a)s
provided that n + a 6= 0 or a 6= 0, −1, −2, · · · . The reason why we work with a generalized
ζ(s, a) function, rather than with the ζ(s) function itself, is because fermions require this
special kind of ζ function for their regularization. Note however that bosons require the
Riemann ζ(s, 1) = ζ(s) function. For the special case of the Hurwitz ζ( 21 + it, a) function
the 3-d plot is shown in Figure 2.3. The discussion presented here of the properties of
the Riemann ζ function has its foundations in Titchmarsh [5]. Although the roots of the
functional equation go back to Riemann, the development of the Hurwitz ζ can be traced
back to Apostol [1] which in turn is taken from Ingham [3].
Let us now examine the properties of the Hurwitz ζ function.
Figure 2.3: |ζ( 12 + it, a)| for 1 ≤ t ≤ 50 and 1

2 ≤a≤2
Proposition 1 The series ζ(s, a) for converges absolutely for σ > 1. The convergence
is uniform in every half-plane σ ≥ 1 + δ with δ > 0 so ζ(s, a) is analytic function of s in the
half-plane σ > 1.
Proof. From the inequalities

∞
X ∞ ∞
X X
(n + a)−s = (n + a)−σ ≤ (n + a)−(1+δ) .

n=1 n=1 n=1
all the statements follow and this proves the claim.
The analytic continuation of the ζ function to a meromorphic function in the complex

plane is more complicated than in the case of the Γ function.
Proposition 2 For σ > 1 we have the integral representation

Z∞
xs−1 e−ax
Γ(s)ζ(s, a) = dx . (2.12)
1 − e−x
0
In the case of the Riemann ζ function, that is when a = 1, we have

Z∞
xs−1 e−x
Γ(s)ζ(s) = dx . (2.13)
1 − e−x
0
Proof. First we consider the case when s is real and s > 1, then extend the result to
complex s by analytic continuation. In the integral for the Γ function we make the change
of variable x = (n + a)t where n ≥ 0 and this yields
Z∞
−s
(n + a) Γ(s) = dte−nt e−at ts−1 .
0
Next, we sum over all n ≥ 0 and this gives

∞ Z ∞
X
ζ(s, a)Γ(s) = dte−nt e−at ts−1 ,
n=0 0
where the series on the right is convergent if Re(s) > 1. To finish the proof we need to
interchange the sum and the integral signs. This interchange is valid by the theory of
Lebesgue integration; however we do not proceed to prove this more rigorously because it
would take us too far from the subject at matter. Therefore, we may write
∞ Z ∞ Z∞ ∞
X X
−nt −at s−1
ζ(s, a)Γ(s) = dte e t = dt e−nt e−at ts−1 .
n=0 0 0 n=0
However, if Im(s) = t > 0 we have 0 < e−t < 1 and hence we may sum
∞
X 1
e−nt = ,
n=0
1 − e−t
by geometric summation. Thus the integrand becomes

Z∞ ∞ Z∞
X
−nt −at s−1 e−at ts−1
ζ(s, a)Γ(s) = dt e e t = dt . (2.14)
n=0
1 − e−t
0 0
Now we have the first part of the argument and we need to extend this to all complex s
with Re(s) > 1. To this end, note that both members are analytic for Re(s) > 1. In order
to show that the right member is analytic we assume 1 + δ ≤ σ ≤ c where c > 1 and δ > 0.
We then have
Z∞ −at s−1 Z∞ Z∞
 1 
e t −at σ−1 Z −at σ−1

≤ dt e t  dt + dt e t
dt = .
1 − e−t 1 − e−t 1 − e−t
0 0 0 1
Note the analogy of splitting the integral as in the proof of Theorem 1.

If 0 ≤ t ≤ 1 we have tσ−1 ≤ tδ and if t ≥ 1 we have tσ−1 ≤ tc−1 . Also since et − 1 ≥ t for
t ≥ 0 we then have
Z1 Z1 Z1
e−at tσ−1 e(1−a)t tδ e1−a
dt ≤ dt t ≤ e(1−a) dt tδ−1 = ,
1 − e−t e −1 δ
0 0 0
and
Z∞ Z∞ Z∞
e−at tσ−1 e−at tc−1 e−at tc−1
dt ≤ dt ≤ dt = ζ(c, a)Γ(c).
1 − e−t 1 − e−t 1 − e−t
1 1 0
This proves that the integral in the statement of the theorem converges uniformly in every
strip 1 + δ ≤ σ ≤ c, where δ > 0, and therefore represents an analytic function in every
such strip, hence also in the half-plane σ = Re(s) > 1. Therefore, by analytic continuation,
(1.14) holds for all s with Re(s) > 1.
Consider the keyhole contour C : a loop around the negative real axis as shown in Fig-
ure 2.4. The loop is made of three parts C1 , C2 , and C3 . The C2 part is a positively
oriented circle of radius ε < 2π above the origin, and C1 and C3 are the lower and upper
edges of a cut in the z plane along the negative real axis.
Figure 2.4: C = C1 ∪ C2 ∪ C3
This can be translated into the following parametrizations: z = re−πi on C1 and z = reπi
on C3 where r varies from ε to ∞.
Proposition 3 If 0 < a ≤ 1 the function defined by the contour integral

z s−1 eaz
Z
1
Υ(s, a) = dz (2.15)
2πi 1 − ez
C
is an entire function of s. Moreover, we have
ζ(s, a) = Γ(1 − s)Υ(s, a) (2.16)

if Re(s) = σ > 1.
Proof. Write z s = rs e−πis on C1 and z s = rs eπis on C3 . Let us consider an arbitrary
compact disk |s| ≤ M and we proceed to prove that the integrals along C1 and C3 converge
uniformly on every such disk. Since the integrand is an entire function of s, this will prove
that the integral Υ(s, a) is entire. Along C1 we have for r ≥ 1,
|z s−1 | = rσ−1 |e−πi(σ−1+it) | = rσ−1 eπt ≤ rM −1 eπM
since |s| ≤ M . The same on C3 gives
|z s−1 | = rσ−1 |eπi(σ−1+it) | = rσ−1 e−πt ≤ rM −1 eπM
also for r ≥ 1. Therefore, independently on which side of the cut we place ourselves, we
have that for r ≥ 1,
z e rM −1 eπM e−ar rM −1 eπM e(1−a)r
s−1 az
1 − ez ≤ = .

1−e −r er − 1
However, er − 1 > er /2 when r > log 2 so the integrand is bounded
R ∞ by MΩ1 rM −1 e−ar where
Ω1 is a constant depending on M but not on r. The integral ε dr r −1 e−ar converges
if ε > 0 so this proves that the integrals along C1 and C3 converge uniformly on every
compact disk |s| ≤ M and hence Υ(s, a) is indeed an entire function of s.
To prove the equation of the theorem, we have to split up the integral as
 
Z Z Z
2πiΥ(s, a) =  dz + dz + dz  z s−1 g(z)
C1 C2 C3
where g(z) = eaz /(1 − ez ). According to the parametrizations we have on C1 and C3 that
g(z) = g(−r) but on the circle C2 we write z = εeiθ , where −π ≤ θ ≤ π. This gives us
Zε Zπ Z∞
s−1 −iπs s−1 (s−1)iθ iθ iθ
2πiΥ(s, a) = dr r e g(−r) + i dθ ε e εe g(εe ) + dr rs−1 eiπs g(−r).
∞ −π ε
Divide by 2i and name the integrals Υ1 and Υ2

πΥ(s, a) = sin(πs)Υ1 (s, ε) + Υ2 (s, ε).
If we let ε → 0 we see that
Z∞
rs−1 e−ar
lim Υ1 (s, ε) = dr = Γ(s)ζ(s, a),
ε→0 1 − e−r
0
as long as σ > 1. In |z| < 2π the function g is analytic except for a first order pole at
z = 0. Therefore zg(z) is analytic everywhere inside |z| < 2π and hence is bounded there,
say |g(z)| ≤ Ω2 /|z|, where |z| = ε < 2π and Ω2 is a constant. We can then write
Zπ
εσ

Ω2
|Υ2 (s, ε)| ≤ dθ e−tθ ≤ Ω2 eπ|t| εσ−1 .
2 ε
−π
When we let ε → 0 and provided that σ > 1 we find that Υ2 (s, ε) → 0 hence we have
πΥ(s, a) = sin(πs)Γ(s)ζ(s, a). Finally, by the use of the Euler relfection formula (2.7) we
have a proof of (2.16).
2.3 Analytic continuation and the functional equation of ζ(s, a)

Now we have to extend the previous result for complex numbers such that σ ≤ 1. In the
statement that we have just proved the functions Υ(s, a) and Γ(1 − s) make sense for every
complex s, and thus we can use this equation to define ζ(s, a) for σ ≤ 1.
Definition 2 If σ ≤ 1 we define ζ(s, a) by the equation
ζ(s, a) = Υ(s, a)Γ(1 − s). (2.17)
This provides the analytic continuation of ζ(s, a) in the entire s plane.
Theorem 5 The function ζ(s, a) defined above is analytic for all s except for a simple
pole at s = 1 with residue 1.
Proof. The function Υ(s, a) is entire so the only possible singularities of ζ(s, a) must be
the poles of Γ(1 − s), and we have shown those to be the points s = 1, 2, 3, · · · . However
Theorem 1 shows that ζ(s, a) is analytic at s = 2, 3, · · · so s = 1 is the only possible pole
of ζ(s, a).
If s in an integer s = n the integrand in the contour integral for Υ(s, a) takes the same
value on both C1 and C3 and hence the integrals along C1 and C3 cancel, yielding
z n−1 eaz z n−1 eaz

Z
1
Υ(n, a) = dz z
= res .
2πi 1−e z=0 1 − ez
C2
In this case we have s = 1 and so

eaz zeaz z −1
Υ(1, a) = res = lim = lim = lim z = −1.
z=0 1 − ez z→0 1 − ez z→0 1 − ez z→0 e
Finally, the residue of ζ(s, a) at s = 1 is computed as
lim (s − 1)ζ(s, a) = − lim (1 − s)Γ(1 − s)Υ(s, a) = −Υ(1, a) lim Γ(2 − s) = Γ(1) = 1,

s→1 s→1 s→1
now the claim is complete: ζ(s, a) has a simple pole at s = 1 with residue 1.
Let us remark that since ζ(s, a) is analytic at s = 2, 3, · · · and Γ(1 − s) has poles at
these points, then (2.17) implies that Υ(s, a) vanishes at these points. Also we have proved
that the Riemann ζ(s) function is analytic everywhere except for a simple pole at s = 1
with residue 1.
Lemma 6 Let S(r) designate the region that remains when we remove from the s plane
all open circular disks of radius r, 0 < r < π with centres at z = 2nπi, n = 0, ±1, ±2, · · · .
Then if 0 < a ≤ 1 the function
eas
g(s) :=
1 − es
is bounded in S(r).
Proof. With our usual notation: s = σ + it we consider the rectangle H(r) with the circle
at n = 0. This rectangle has an indentation as follows
H(r) = {s : |σ| ≤ 1, |t| ≤ π, |s| ≥ r} ,
as shown in Figure 2.5 below.

Figure 2.5: The region H(r)
The set H so defined is compact so g is bounded on H. Also, because of the periodicity

|g(s + 2πi)| = |g(s)|, g is bounded in the perforated infinite strip
{s : |σ| ≤ 1, |s − 2nπi| ≥ r, n = 0, ±1, ±2, · · ·} .
Let us suppose that |σ| ≥ 1 and consider

as aσ
e
= e
eaσ
|g(s)| = ≤ .
1 − es |1 − es | |1 − eσ |
We can examine the numerator and denominator for σ ≥ 1 giving |1 − eσ | = eσ − 1 and

eaσ ≤ eσ so
eσ 1 1 e
|g(s)| ≤ σ = ≤ = .
e −1 1 − e−σ 1 − e−1 e−1
A similar argument when σ ≤ −1 gives |1 − eσ | = 1 − eσ and so
eσ 1 1 e
|g(s)| ≤ ≤ ≤ = ,
1 − eσ 1 − eσ 1 − e−1 e−1
and therefore, as we claimed |g(s)| ≤ e/(e − 1) for |σ| ≤ 1.
Definition 3 The periodic ζ function is defined as

∞
X e2πinx
ζperiodic (x, s) := , (2.18)
n=1
ns
where x is real and σ > 1.
Let us remark the following properties of the periodic zeta function. It is indeed a pe-
riodic function of x with period 1 and ζperiodic (1, s) = ζ(s).
Theorem 6 The series converges absolutely if σ > 1. If x is not an integer the series
also converges (conditionally) for σ > 0.
Proof. This is because for each fixed non-integral x the coefficients have bounded partial
sums.
Theorem 7 (Hurwitz’s formula) If 0 < a ≤ 1 and σ > 1 we have

Γ(s) −πis/2
ζ(1 − s) = (e ζperiodic (a, s) + eπis/2 ζperiodic (−a, s)). (2.19)
(2π)s
If a 6= 1 this is also valid for σ > 0.
Proof. Consider the function defined by the contour integral
z s−1 eaz
I
1
ΥN (s, a) := dz , (2.20)
2πi 1 − ez
C(N )
where C(N ) is the contour shown in Figure 2.6 and N is an integer. It is the same keyhole
contour as that of the Γ function, only that it has been rotated for convenience.
Figure 2.6: The keyhole contour C(N )
The poles are located on the y-axis, symmetric to the origin, at multiplies of 2πi. Let us
first prove that if σ < 0 then limN →∞ ΥN (s, a) = Υ(s, a). The method to prove this is to
show that the integral along the outer circle tends to 0 as N → ∞. On the outer circle we
have z = R eiθ , −π ≤ θ ≤ π, hence
|z s−1 | = |Rs−1 eiθ(s−1) | = Rσ−1 e−tθ ≤ Rσ−1 eπ|t| .
The outter circle is inside the domain S(r) described in Lemma 6, the integrand is bounded
by Ω3 eπ|t| Rσ−1 where Ω3 is the bound for |g(s)| implied by Lemma 6, hence the whole
integral is bounded by 2πeπ|t| Rσ which tends to 0 as R tends to infinity as long as σ < 0.
Now when we replace s by 1 − s we obtain
z −s eaz
I
1
lim ΥN (1 − s, a) = lim dz = Υ(1 − s, a)
N →∞ N →∞ 2πi 1 − ez
C(N )
for σ > 1. We are left with the problem of computing ΥN (1 − s, a) which we proceed to do
by the use of the Cauchy residue theorem. Formally,
XN N
X
ΥN (1 − s, a) = − res f (z) = − res f (z) + res f (z)
z=n z=n z=−n
n=−N,n6=0 n=1
where f (z) is the integrand of ΥN (1 − s, a) and residues are calculated as follows

−s az
z e z −s eaz
res f (z) = res z
= lim (z − 2nπi)
z=n z=2nπi 1−e z→2nπi 1 − ez
e2nπia z − 2nπi e2nπia
= s
lim z
=− ,
(2nπi) z→2nπi 1 − e (2nπi)s
which in turn gives
N N
X e2nπia X e−2nπia
ΥN (1 − s, a) = + .
n=1
(2nπi)s n=1 (2nπi)s
Now we make the following replacements i−s = e−πis/2 and (−i)−s = eπis/2 which allow us
to write
N N
e−πis/2 X e2nπia eπis/2 X e−2nπia
ΥN (1 − s, a) = +
(2π)s n=1 ns (2π)s n=1 (2nπi)s
and we let N → ∞
e−πis/2 eπis/2
Υ(1 − s, a) = s
ζperiodic (a, s) + ζperiodic (−a, s).
(2π) (2π)s
We have thus arrived at the following result
Γ(s) n −πis/2 πis/2
o
ζ(1 − s, a) = Γ(s)Υ(1 − s, a) = e ζ periodic (a, s) + e ζ periodic (−a, s) .
(2π)s
This proves the claim.
The simplest particular case - and the most important one - is when we take a = 1 this
gives us the functional equation of the Riemann ζ function
Γ(s) n −πis/2 πis/2
o Γ(s) πs
ζ(1 − s) = s
e ζ(s) + e ζ(s) = s
2 cos ζ(s). (2.21)
(2π) (2π) 2
This is valid for σ > 1 but it also holds for all s by analytic continuation. Another useful
formulation can be obtained by switching s with 1 − s
πs
ζ(s) = 2(2π)s−1 Γ(1 − s) sin ζ(1 − s). (2.22)
2
Let us now see the consequences of this equation. Taking s = 2n + 1 in (2.21) when n is an
integer the cosine factor vanishes and we find the trivial zeroes of ζ(s)
ζ(−2n) = 0 n = 1, 2, 3, · · · . (2.23)
Later we will need to use certain other values of the Riemann ζ function which we can now
compute. In particular the value of ζ(−n, a) can be calculated if n is a non-negative integer.
Taking s = −n in the formula ζ(−s, a) = Γ(1 − s)Υ(s, a) we find that
ζ(−n, a) = Γ(1 + n)Υ(−n, a) = n!Υ(−n, a), (2.24)
where
z −n−1 eaz

Υ(−n, a) = res . (2.25)
z=0 1 − ez
The evaluation of this residue requires special functions of their own (special type of poly-
nomials, rather) which are known as the Bernoulli polynomials.
2.4 Bernoulli numbers and the value of ζ(0)

Definition 4 For any complex s we define the functions Bn (s) as
∞
zesz X Bn (s) n
= z (2.26)
ez − 1 n=0 n!
provided that |z| < 2π.
A particular case of the polynomials are the Bernoulli numbers Bn = Bn (0) i.e.
∞
z X Bn (0) n
= z . (2.27)
ez − 1 n=0 n!
Pn
Lemma 7 One has thePfollowing equations Bn (s) = k=0 (nk ) Bk sn−k . In particular when
n
s = 1 this yields Bn = k=0 (nk ) Bk .
Proof. Using a Taylor expansion and comparing coefficients on both sides we have
∞ ∞
! ∞ !
X Bn (s) n z X Bn n X sn
sz n
z = z e = z z ,
n=0
n! e −1 n=0
n! n=0
n!
n
Bn (s) X Bk sn−k
=
n! k! (n − k)!
k=0
and by passing n! to the RHS we obtain the lemma.
Now we can write the values of the ζ function in terms of Bernoulli numbers.
Lemma 8 For every integer n ≥ 0 we have

Bn+1 (a)
ζ(−n, a) = − . (2.28)
n+1
Proof. This follows from the previous observation (2.24) so we just have to evaluate the
integral Υ by the Cauchy residue theorem.
−n−1 az −n−2 az
z e z ze
Υ(−n, a) = res z
= − res z
z=0 1−e z=0 e −1
∞
!
X Bk (a) k Bn+1 (a)
= res z −n−2 z =− ,
z=0 k! (n + 1)!
k=0
and multiplying by n! we have the end of the proof.

Lemma 9 The recursion Bn (s + 1) − Bn (s) = nsn−1 is valid for Bernoulli polynomials

if n ≥ 1 and in particular for Bernoulli numbers when n > 1: Bn (0) = Bn (1).
Proof. From the identity
e(s+1)z esz
z z −z z = zesz
e −1 e −1
it follows that
∞ ∞
X Bn (s + 1) − Bn (s) n X sn n+1
z = z ,
n=0
n! n=0
n!
and as we did before, we equate coefficients of z n to obtain the first statement and then set
s = 0 to obtain the second statement.
Using the definition, the first Bernoulli number is B0 = 1 and the rest can be computed by
recursion. We obtain the values listed on the table below. Building on from the Bernoulli
numbers we can construct the polynomials by the use of the lemmas, the first ones as
n Bn (0) Bn (s)
0 1 1
1 − 21 s − 21
1
2 6 s2 − s + 61
3 0 s − 23 s2 + 1s s
3
1 1
4 − 30 s − 2s3 + s2 − 30
4
5 0 ···
Note that for n ≥ 0 we have by setting a = 1 in (2.28)
Bn+1
ζ(−n) = − (2.29)
n+1
and because of the trivial zeroes of the Riemann ζ function ζ(−2n) = 0 we confirm our
observation that the odd Bernoulli numbers are zero, i.e. B2n+1 = 0 . Also note
1
ζ(0) = − . (2.30)
2
Finally we can write a compact formula for the even values of the ζ function in terms of
Bernoulli numbers.
Theorem 8 Suppose n is a positive integer, then
(2π)2n B2n
ζ(2n) = (−1)n+1 . (2.31)
2(2n)!
Proof. This follows from the functional equation by setting s = 2n
ζ(1 − 2n) = 2(2π)−2n Γ(2n) cos(πn)ζ(2n)
re-arranging we obtain
B2n
− = 2(2π)−2n (2n − 1)!(−1)n ζ(2n),
2n
from which the result follows.
Using the tabulated values given above we have the well-known Euler formulas
π2 π4
ζ(2) = , ζ(4) = . (2.32)
6 90
For instance the value of ζ(2) was used in the Quantum Information course in connection
with the result that the probability of two random intergers being coprime is
1 6
p= = 2,
ζ(2) π
a very useful result in cryptography; as well as in the Cosmology course in connection

to the entropy density of a species of particles in the ultra-relativistic limit. The second
value occurs in the computation of the total energy u radiated by a blackbody in quantum
mechanics
Z∞
8πk 4 T 4 x3 8πk 4 T 4
u= dx = 3!ζ(4),
c3 h3 ex − 1 c3 h3
0
as well as in the neutrino density (Fermi distribution) in the early history of the universe
Z∞
4π x3 7π 5
ρν = 3 dx = (kT )4 .
h ex/(kT ) + 1 30h3
0
Let us remark that no such formula for odd values is known and in fact the value of ζ(3)
remains one of the most elusive mysteries of modern mathematics. A significant advance
was achieved by Roger Apery who was able to show in 1979 that it is an irrational number.
2.5 The value of ζ 0 (0)

Theorem 9 One has
1
ζ 0 (0) = − log(2π). (2.33)
2
Proof. We may write (2.15) and (2.16) as
(−z)s−1 dz (−z)s
Z Z
ζ(s) 1 1
= dz z =
Γ(1 − s) 2πi e −1 2πi z ez − 1
C C
where C is the same contour as that of Figure 2.4, except shifted to the positive infinity
instead of negative infinity and to account for this we have (−z)s = exp[s log(−z)]. Let us
differentiate with respect to s and then set s = 1
Z
1 dz (−z) log(−z)
.
2πi z ez − 1
C
The integral on the RHS can be split as

Zε Z∞
dz (−z)(log z − iπ) dz (−z)(log ε + iθ − iπ)
I
1 1 1 dz (−z)(log z + iπ)
z
+ z
+
2πi z e −1 2πi z e −1 2πi z ez − 1
+∞ |z|=ε ε
and writing z = εei(φ+π) in the middle integral we have

Z∞ I Zπ
dz log ε dz z 1 z
− − − dφ · φ .
ez − 1 2πi z ez − 1 2πi ez − 1
ε |z|=ε −π
At this point we need to evaluate all three integrals. The first one can be expanded as
Z∞ Z∞ X ∞ ∞
z=∞
∞
dz −nz
X e−nz X (e−ε )n
− z
= dz e = − = −
e −1 −n n

ε ε n=1 n=1 z=ε n=1
ε2 ε3
ε
= log(1 − e−ε ) = log ε − + − · · · = log ε + log 1 − + · · · .
2 6 2
The second integral is solved using Cauchy’s integral theorem by noting that at z = 0 we
have z(ez − 1)−1 → 1, and therefore we have
I
log ε dz z
− = − log ε.
2πi z ez − 1
|z|=ε
Finally, the third integral goes to zero as ε → 0. Putting all these facts together we have
shown that Z
1 dz (−z) log(−z)
= log ε − log ε + O(εα ) = 0.
2πi z ez − 1
C
Re-arranging the functional equation (2.22)
ζ(s) πs
= 2(2π)s−1 ζ(1 − s) sin
Γ(1 − s) 2
and because the derivative at s = 1 is zero, as we have just shown, its logarithmic derivative
ζ 0 (1 − s) π cos(πs/2)
log(2π) − +
ζ(1 − s) 2 sin(πs/2)
must also be 0 at s = 1 and consequently
ζ 0 (0)
log(2π) = ,
ζ(0)
finally yielding the result of Theorem 9 by use of (2.30).
Theorem 10 One has

1 0 1 1
ζ 0, =0 and ζ 0, = − log 2. (2.34)
2 2 2
Proof. We have the following identity
∞
X 1 1
ζ(s, 1/2) + ζ(s) = 2s + = 2s ζ(s),
n=1
(2n − 1)s (2n)s
from which it follows that

ζ(s, 1/2) = (2s − 1)ζ(s)
and hence the first formula is shown. By differentiating the above equation with respect to
s we have the second formula of the theorem.
2.6 The polygamma function ψ (m) (s)

A complete discussion of the digamma and polygamma functions would take us too far into
the presentation of special functions therefore we will only list the definitions and properties
of these functions.
The Weierstrass product can be re-written in such a form as to leave the Γ function as
n!
Γ(s) = lim ns
n→∞ s(s + 1)(s + 2) · · · (s + n)
for s 6= 0, −1, −2, · · · . Thus

n!
log Γ(s + 1) = log sΓ(s) = log lim ns
n→∞(s + 1)(s + 2) · · · (s + n)
= lim [log(n!) + s log n − log(s + 1) − log(s + 2) − · · · − log(s + n)] ,
n→∞
bearing in mind that the logarithm of the limit is the limit of the logarithm. Differentiating
with respect to s we can define the digamma function ψ as

d 1 1 1
ψ(s + 1) := log Γ(s + 1) = lim log n − − − ··· − . (2.35)
ds n→∞ s+1 s+2 s+n
We can bring in the definition of the Euler constant

∞ ∞
X 1 1 X s
ψ(s + 1) = −γ − − = −γ + (2.36)
n=1
s+n n n=1
n(s + n)
where of course
ψ(1) = Γ0 (1) = −γ. (2.37)
The polygamma function is a generalization of the ψ defined as

∞
dm+1 X 1
ψ (m) (s + 1) := log(s!) = (−1) m+1
m! . (2.38)
dsm+1 n=1
(s + n)m+1
We note that ψ (0) (s) = ψ(s). Some properties include
dm+1 dm+1
ψ (m) (s) = log Γ(s) = ψ(s), (2.39)
dsm+1 dsm+1
ψ (m) (1) = (−1)m+1 m!ζ(m + 1), (2.40)
and the very useful MacLaurin expansion

∞ ∞
X sn (n−1) X (−1)n n
log Γ(s + 1) = ψ (1) = −γs + s ζ(n) (2.41)
n=1
n! n=2
n
convergent for |s| < 1.

Figure 2.7: |ψ(x + iy)| for −5 ≤ x ≤ 2 and −1 ≤ y ≤ 1
2.7 Laurent series of the ζ(s) function

We will lastly need to write the Laurent series of the ζ function. In order to do so, we will
use the following theorem from Titchmarsh [5], which we assume.
Theorem 11 Let f (x) be any function with a continuous derivative in the interval [a, b].
Then if [x] denotes the greatest integer not exceeding x,
X Zb Zb
f (x) = dxf (x) + dx(x − [x] − 21 )f 0 (x) + (a − [a] − 21 )f (a) − (b − [b] − 21 )f (b).
a<n≤b a a
If we take the case f (n) = n−s where a, b and n are integers and s 6= 1 then
b Zb 1
X
−s b1−s − a1−s x − [x] − 1
n = −s dx 2
+ (b−s − a−s ).
n=a+1
1−s xs+1 2
a
By setting a = 1 and letting b → ∞ with σ > 1 and adding 1 to each side we have
Z∞ 1
x − [x] − 2 1 1
ζ(s) = s dx + + . (2.42)
xs+1 s−1 2
1
This equation contains a remarkable amount of information. First, there is clearly a simple
pole at s = 1 with residue 1 (the numerator of the fraction). The RHS provides analytic
continuation up to σ = 0. Following Ivic [4] we have
Theorem 12 (Stieltjes’ representation) The Laurent series of the ζ function is

∞
1 X
ζ(s) = +γ+ γn (s − 1)n , (2.43)
s−1 n=1
where γn are constants independent of s.

Proof. The first term is explained easily as the pole is at s = 1, it is simple and the residue
is 1 as we know from Theorem 5. Furthermore
Z∞
x − [x] − 12

1 1
lim ζ(s) − = dx +
s→1 s−1 x2 2
1
 n 
−
Z
x [x]
= lim  dx + 1
n→∞ x2
1
 k+1

n−1
X Z dx
= lim  k − log n + 1
n→∞ x2
k=1 k
n−1
!
X 1
= lim − log n = γ,
n→∞ k+1
k=0
which shows the second term of (2.43) and the remaining terms are regular.
We will use this formula frequently. The constants γn are known as the Stieltjes constants
and it can be shown that
"m #
(−1)n X logn k logn+1 m
γn = lim − , (2.44)
n! m→∞ k n+1
k=1
however, we will not be needing these. Finally note that the following equation
∞ ∞ ∞ ∞ ∞
X (−1)n X 1 X 1 1−s
X 1 1 X (−1)n
s
+ s
= 2 s
= 2 s
⇔ ζ(s) = 1−2
(2.45)
n=1
n n=1
n (2k) k 1−2 n=1
ns
k=1 k=1
also provides analytic continuation for σ > 0.
2.8 Concluding remarks on the disbriution of prime numbers and

the zeros of ζ(s)
It would not be fair to end this chapter without commenting, briefly, on the link between
the ζ function and prime numbers. The reason for doing this is twofold.
Firstly, it is precisely the connection between the continuous (ζ) and the discrete (primes)
that makes this theory so powerful and elegant.
Secondly, the ζ function is - in a certain way - deeply connected to quantum theory, and
hence this indicates that quantum theory is also connected with the primes numbers. This
connection is worth investigating on a research level and many papers have been published
[J. Brian Conrey The Riemann Hypothesis] on the link between primes, quantum theory
and random matrices.
We shall proceed informally.
Euler discovered that for |x| < 1 we have (1 − x)−1 = 1 + x + x2 + · · · and with x = p−s
for p 6= 1 we have
1
= 1 + p−s + p−2s + · · · .
1 − p−s
Now, the Fundamental Theorem of Arithmetic states that every integer n can be expressed
uniquely as a product of only prime factors, hence
Y 1 1
1 + p−s + p−2s + · · · = 1 + s + s + · · ·

p
2 3
or equivalently,
∞
Y 1 X 1
= (2.46)
p
1 − p−s n=1
n s
for s > 1. In fact, the above equation is usually called the Euler product or the analytic
version of the Fundamental Theorem of Arithmetic as it is its analytic equivalent. Once we
have established analytic continuation of the ζ function as we have done in this Chapter
and let s → 1 we have the divergence of the harmonic series on the RHS which implies that
the LHS must also diverge. This can only happen if there is an infinite number of primes.
Interesting as this may be, we can go further by manipulating the equation for s = 1 as
follows
∞
X 1 Y 1 X 1 X 1
log = log = log = log 1 +
n=1
n p
1 − p−1 p
1 − p−1 p
p−1
and ex > 1 + x ⇒ x > log(1 + x) so that

X 1 X 1
log 1 + < ,
p
p−1 p
p−1
effectively showing that p (p − 1)−1 diverges. Finally we note that (pk − 1)−1 < p−1
P
k−1
where pk denotes the k th prime, hence
X1
(2.47)
p
p
diverges. This result was the first attempt to quantify the distribution of prime numbers
as it suggests that
X1
∼ log log x, (2.48)
p<x
p
where ∼ indicates that the ratio of the two functions tends to 1 as x → ∞. This observation
of Euler’s was the point of no return in analytic number theory and the distribution of prime
numbers.
During the XIX century, several mathematicians, including Gauss and Legendre, formulated
the prime number theorem,
Zx
X dt x
π(x) := 1∼ ∼ , (2.49)
log t log x
p≤x 2
where π(x) denotes the number of primes less than or equal to x and the integral is called the
logarithmic integral Li. As we have seen, the zeros ζ(s) are located at s = −2, −4, −6, · · ·
but these are only the trivial zeros. Riemann defined the function ξ as
1 s
ξ(s) := s(s − 1)π −s/2 Γ ζ(s) (2.50)
2 2
Figure 2.8: |ξ(x + iy)| for −10 ≤ x ≤ 10 and −2 ≤ y ≤ 2
in which case the functional equations takes the elegant form
ξ(s) = ξ(1 − s).
The Euler product shows that ζ(s) has no zeros in the halfplane Re(s) > 1 because a
convergent infinite product can be zero only if one of its factors is zero. Let ρ1 , ρ2 , · · · be
the zeros of ξ(s). It follows from the functional equation (2.22) that ζ(s) has no zeros for
ρ < 0 except for the trivial ones. This is because in the functional equation ζ(1 − s) has
no zeros for ρ < 0, and sin(πs/2) has simple zeros at s = −2, −4, · · · and Γ(1 − s) has no
zeros.
Precisely, the trivial zeros of ζ(s) do not correspond to zeros of ξ(s) since they are cancelled
by the poles of Γ(s/2). Therefore, it follows that ξ(s) has no zeros for ρ < 0 and for ρ > 1.
The zeros ρ1 , ρ2 , · · · lie in the strip 0 ≤ σ ≤ 1. However, these are the zeros of ζ(s) also
since s(s − 1)Γ(s/2) has no zeros in the strip except the one at s = 1 which is cancelled by
the simple pole of ζ(s).
This proves that ζ(s) has an infinite number of zeros ρ1 , ρ2 , · · · in the strip 0 ≤ σ ≤ 1 and
since
∞
X (−1)n+1
(1 − 21−s )ζ(s) = >0
n=1
ns
for 0 < s < 1 and ζ(0) = −1/2 6= 0 then ζ(s) has no zeros on the real axis between 0 and
1, i.e. the zeros ρ1 , ρ2 , · · · are all complex.
Here comes a critical observation. The zeros come in conjugate pairs: since ζ(s) is real on
the real axis and if ρ is a zero so is 1 − ρ, by the functional equation (2.22), and then so is
1 − ρ̄. If ρ = β + iγ then 1 − ρ̄ = 1 − β + iγ. Consequently the zeros either lie on σ = 1/2
or occur in pairs symmetrical about this line.
Figure 2.9: |ζ(σ + it)|−1 for −2 ≤ σ ≤ 2 and 0 ≤ t ≤ 50: the zeros become poles
Riemann gave a sketch of a proof of the approximation of the number N (T ) of these zeroes
for 0 < t < T as
T T T
N (T ) = log − + O(log T ). (2.51)
2π 2π 2π
This was subsequently shown by von Mangoldt. This equation has the same form as that of
the behaviour of the effective Lagrangian at 1-loop order (see Section 5.6 Equation (5.189)).
Riemann went even futher and conjectured that all the zeros are on the line Re(s) = 12 .
In 1914 Hardy showed that there is an infinite number of zeros on the critical line Re(s) = 12
but this does not mean that all the zeros are located there.
Riemann’s main accomplishement of his 1859 paper was the analytic continuation of the ζ
function and his proofs (he gave two) of the functional equation. Furthemore, by the use of
an innovative Fourier transform he showed that
Z∞
π(u)
log ζ(s) = s du
u(us − 1)
2
which enabled him to prove the following remarkable closed analytic formula for π(x)
∞
1 1 1 1 X µ(n)
π(x) = R(x) − R(x1/2 ) − R(x1/3 ) − R(x1/5 ) + R(x1/6 ) + · · · = R(x1/n )
2 3 5 6 n=1
n
where µ(n) is the Moebius function defined as 0 if n is divisible by a prime square, 1 if n is

a product of an even number of distinct primes and −1 if n is a product of an odd number
of distinct primes and
Z∞
X
ρ dt
R(x) = Li(x) − Li(x ) − log 2 + ,
t(t2 − 1) log t
ρ:ζ(ρ)=0 x
where each term is paired with its ’twin’, i.e. ρ ↔ 1 − ρ so that

X X
Li(xρ ) = Li(xρ ) + Li(x1−ρ ).
ρ Imρ>0
In 1896 Hadamard and de la Vallee Poussin proved, independently and almost simultane-
ously, the prime number theorem by showing that it is equivalent to ζ(1 + it) 6= 0, i.e. no
zeros on Re(s) = 1.
Riemann’s Hypothesis is equivalent to
Zx
dt
π(x) − = O(x1/2 log x)
log t
2
as x → ∞.
Finally, Hadamard also proved the product representation stated (without a valid proof)
by Riemann
eH s Y
s s/ρ
ζ(s) = 1− e (2.52)
2(s − 1)Γ 2s + 1

ρ
ρ:ζ(ρ)=0
where H = log 2π − 1 − γ/2.

The ξ(s) function also admits the product representation

1Y s
ξ(s) = 1− .
2 ρ ρ
Let us re-write the definition of ξ(s) as

s
ξ(s) = Π (s − 1)π −s/2 ζ(s)
2
where Π is Riemann’s notation for the shifted Γ function
Π(s − 1) := Γ(s).
The logarithmic derivative of ξ(s) is on the one hand
X d X
s 1
log 1 − =
ρ
ds ρ ρ
s − ρ
and on the other hand

d s 1 1 ζ 0 (s)
log Π − log π + + .
ds 2 2 s−1 ζ(s)
Evaluating both expressions at s = 0 yields
X1 1 1
= γ + log π + 1 − log 2π
ρ
ρ 2 2
or
X 1 1

1
+ = [2 + γ − log 4π] (2.53)
ρ 1−ρ 2
Imρ>0
since Π0 (0) = Γ0 (1) = −γ. Note that γ − log 4π is a mathematical constant that will show
up frequently in renormalization, indicating yet another link between quantum theory and
the zeros of the ζ function. This formula can be used to compute the zeros ρ. Some of the
first zeros ρ = 21 + iti are
Figure 2.10: |ζ( 12 + it)| for 0 ≤ t ≤ 50
that is,
t1 = 14.134725
t2 = 21.022040
t3 = 25.010858
t4 = 30.424878
t5 = 32.935057
REFERENCES
Here we have followed the course from
- [1] Tom Apostol’s Introduction to Analytic Number Theory for the presentation of the ζ
function and its analytic continuation. Apostol in turn follows
- [3] Distribution of Prime Numbers by Ingham.
The section of the Γ function follows from the course (and the exercises) on
- [2] Complex Analysis by Freitag and Busam, Titchsmarch’s Theory of Functions [6] and
Whitaker and Watson’s Modern Analysis [7].
There exist many sources for the development of the Riemann ζ function. The most com-
prehensive ones are by
- [4] Ivic’s Riemann’s Zeta Function, and - [5] E.C. Titchmarsh (revised by D.R. Heath-
Brown) The Theory of the Riemann Zeta Function.
3 ZETA REGULARIZATION IN QUANTUM MECHANICS 39
3 Zeta Regularization in Quantum Mechanics

3.1 Path integral of the harmonic oscillator
One of the first instances where the Riemann ζ function occurs in quantum physics is in
the path integral development of the harmonic oscillator; specifically it takes place when
computing the partition function of the spectrum of the harmonic oscillator.
We will follow Exercise 9.2 p 312 from Peskin and Schroeder [2] and the path integral
development from Chapter 2 of Kleinert [1]. The action of the one-dimensional harmonic
oscillator is given by
Ztf
S= dtL, (3.1)
ti
where the Lagrangian is

1 1
L= mẋ2 − mω 2 x2 . (3.2)
2 2
As we know from the functional approach to quantum mechanics, the transition amplitude
is the functional integral
Z
hxf , tf |xi , ti i = DxeiS[x(t)] . (3.3)
The extremum of S, xc (t), satisfies

δS[x]
= 0. (3.4)
δx x=xc (t)
We now proceed to expand the action around xc (t). This indicates that xc (t) is the classical
trajectory connecting both space-time points of the amplitude and therefore it satisfies the
Euler-Lagrange equation
ẍc + ω 2 xc = 0. (3.5)
The solution to the equation above with conditions xc (ti ) = xi and xc (tf ) = xf is
xc (t) = (sin ωT )−1 [xf sin ω(t − ti ) + xi sin ω(tf − t)], (3.6)
where T = tf − ti . We next plug this solution into the action S

mω
Sc := S[xc ] = [(x2 + x2i ) cos ωT − 2xf xi ]. (3.7)
2 sin ωT f
As we intended originally, we now expand S[x] around x = xc to obtain
δ 2 S[x]
Z Z
δS[x] 1
S[xc + z] = S[xc ] + dtz(t) + dt dt
1 2 z(t 1 )z(t 2 ) (3.8)
δx(t) x=xc 2! δx(t1 )δx(t2 ) x=xc
where z(t) satisfies the boundary condition
z(ti ) = z(tf ) = 0. (3.9)

The expansion ends at second order because the action is in second order in x. Noting that
δS[x]/δx = 0 at x = xc we are left with the first and last terms only. Because the expansion
is finite
δ 2 S[x]
Z
1
S[xc + z] = S[xc ] + dt1 dt2 z(t1 )z(t2 ) (3.10)
2! δx(t1 )δx(t2 ) x=xc
the problem can be solved analytically.

Let us now compute the second order functional derivative in the integrand; this can be
accomplished as follows
Ztf
d2

δ 1 1
dt mẋ(t)2 − mω 2 x(t)2 = −m + ω 2
x(t1 ). (3.11)
δx(t1 ) 2 2 dt21
ti
Next, using the rule
δx(t1 )
= δ(t1 − t2 ) (3.12)
δx(t2 )
we obtain the following expression for the second order functional derivative
δ 2 S[x]
2
d 2
= −m + ω δ(t1 − t2 ). (3.13)
δx(t1 )δx(t2 ) dt21
Plugging this back into the equation for the expansion, we can use the delta function to get
rid of the t variables
Z 2
m d 2
S[xc + z] = S[xc ] − dt1 dt2 z(t1 )z(t2 )δ(t1 − t2 ) +ω
2! dt21
Z
m
= S[xc ] + dt(ż 2 − ω 2 z 2 ), (3.14)
2!
where we have simplified the expression by using (3.9).
A crucial point to be made is that because Dx is invariant, we may replace it by Dz and
this will give us
Ztf
 
Z
m
hxf , tf | xi , ti i = eiS[xc ] Dz exp i dt(ż 2 − ω 2 z 2 ). (3.15)
2
z(ti )=z(tf )=0 ti
The fluctuation part (integral at z(0) = z(T ) = 0)

 
Z ZT
m
If := Dz exp i dt(ż 2 − ω 2 z 2 ) (3.16)
2
z(0)=z(T )=0 0
is computed as follows. First, let us shift the variables so that time start at 0 and ends at
T. We Fourier expand z(t) as
∞
X nπt
z(t) = an sin . (3.17)
n=1
T
This choice satisfies (3.9). The integral in the exponential gives

ZT ∞
2 2 2 T X 2 nπ 2 2
dt(ż − ω z ) = a −ω . (3.18)
2 n=1 n T
0
In order to have a well-defined transformation we must check that the number of variables
is the same before and after the transformation. Indeed, the Fourier transformation from
y(t) to an may be thought of as a change of variables in the integration. To check this, take
the number of the time slice to be N + 1, including both t = 0 and t = T , for which there
exist N − 1 independent zk variables. Therefore, we must set an = 0 for all n > N − 1.
Next, we compute the corresponding Jacobian. Denote by tk the k th time slice when the
interval [0, T ] is split into N infinitesimal parts, then

∂zk nπtk
JN = det = det sin . (3.19)
∂an T
We evaluate the Jacobian for the easiest possible case, which is that of the free particle.
Therefore, let us make a digression to evaluate the above mentioned probability amplitude.
3.2 Solution of the free particle

For a free particle, the Lagrangian is L = 21 mẋ2 . Carefully derived solutions to the free
particle can be found in Chapter 2 of Kleinert [1] as well as in Chapter 3 of Grosche and
Steiner [4]. The amplitude is computed by first noting that the Hamiltonian is given
p2
H = pẋ − L = , (3.20)
2m
so that
D E Z D E
−iĤT
hxf , tf | xi , ti i = xf |e | xi = dp xf | exp(−iĤT )| p hp| xi i
im(xf − xi )2
Z r
dp ip(xf −xi ) −iT p2 /(2m) m
= e e = exp (3.21)
2π 2πiT 2T
where T = tf − ti as noted before and ε here denotes the discretization of time ε = T /N .
Similarly to the theory of functional integration we have amplitude
" n #
m n/2 Z X m xk − xk−1 2
hxf , tf | xi , ti i = lim dx1 · · · dxn−1 exp iε . (3.22)
n→∞ 2πiε 2 ε
k=1
We now change the coordinates to zk = (m/(2ε))1/2 xk so that the amplitude becomes

" n #
m n/2 2ε (n−1)/2 Z X
hxf , tf | xi , ti i = lim dz1 · · · dzn−1 exp i (zk − zk−1 )2 .
n→∞ 2πiε m
k=1
(3.23)
In the appendix (A.1), we prove by induction that

" n # 1/2
(iπ)n−1
Z X 2
dz1 · · · dzn−1 exp i 2
(zk − zk−1 ) = ei(zn −z0 ) /n . (3.24)
n
k=1
The expression reduces to

m n/2 2πiε (n−1)/2 1 2
hxf , tf | xi , ti i = lim √ eim(xf −x0 ) /(2nε)
n→∞ 2πiε m n
r
m im
= exp (xf − xi )2 . (3.25)
2πiT 2T
Taking (3.25) into account we arrive at
1/2 1/2
1 im 2 1
hxf , T | xi , 0i = exp (xf − xi ) = eiS[xc ] . (3.26)
2πiT 2T 2πiT
When we write this in terms of a path integral we obtain
Ztf
 
Z
m
eiS[xc ] Dz exp i dtż 2 . (3.27)
2
z(0)=z(T )=0 ti
Now, from (3.18) we have

ZT N
m X a2n n2 π 2
dt ż 2 → m (3.28)
2 n=1
4T
0
and when we compare both path integral expressions one has the equality
Ztf
 
1/2 Z
1 m
= Dz exp i dtż 2 
2πiT 2
z(0)=z(T )=0 ti
−1
1/2 Z N
!
a2n n2 π 2

1 X
= lim JN da1 . . . daN −1 exp im . (3.29)
N →∞ 2πiε n=1
4T
Now comes the process of evaluating the Gaussian integrals

1/2 N/2 NY−1 1/2
1 1 1 4πiT
= lim JN
2πiT N →∞ 2πiε n=1
n π2
N/2 (N −1)/2
1 1 4πiT
= lim JN (3.30)
N →∞ 2πiε (N − 1)! π2
and from this we obtain a formula for the Jacobian
JN = 2−(N −1)/2 N −N/2 π N/1 (N − 1)! → ∞. (3.31)
N →∞
The Jacobian is divergent, but this divergence is not relevant because JN is combined with
other divergent factors.
Let us now return to the original problem of the probability amplitude of a harmonic
oscillator. The amplitude was
N/2
1
hxf , T | xi , 0i = lim JN eiS[xc ]
N →∞ 2πiT
N −1
Z " #
mT X 2 nπ 2 2
× da1 . . . daN −1 exp i a −ω . (3.32)
4 n=1 n T
As we did with the free particle, we carry out the computation of the Gaussian integrals
using the formula
Z 1/2 " 2 #−1/2
imT 2 nπ 2 2 4iT ωT
dan exp a −ω = 1− . (3.33)
4 n T πn2 nπ
This breaks the amplitude (3.32) into smaller parts

N/2 " N −11/2 # NY−1
" 2 #−1/2
N 1 4iT
Y ωT
hxf , tf | xi , ti i = lim JN eiS[xc ] 1−
N →∞ 2πiT k π n=1
nπ
k=1
1/2 N −1
" 2 #−1/2
1 iS[xc ]
Y ωT
= lim e 1− . (3.34)
N →∞ 2πiT nπ
n=1
It can be shown that this product is equal to

N
" 2 #
Y ωT sin ωT
lim 1− = . (3.35)
N →∞
n=1
nπ ωT
The divergence of Jn cancels the divergence of the other terms and therefore we are left with
a finite value confirming what we stated above concerning the irrelevance of J∞ . When we
insert the value of the product we arrive at the final result
ω 1/2
hxf , tf | xi , ti i = eiS[xc ]
2πi sin ωT
ω 1/2 iω 2 2

= exp (xf + xi ) cos ωT − 2xi xf . (3.36)
2πi sin ωT 2 sin ωT
3.3 The bosonic partition function

If we have a Hamiltonian Ĥ whose spectrum is bounded from below then, by adding a
positive constant to the Hamiltonian, we can make Ĥ positive definite, i.e.
spec(Ĥ) = {0 < E0 ≤ E1 ≤ · · · ≤ En ≤ · · ·} . (3.37)
Also we assume that the ground state is not degenerate. The spectral decomposition of
e−iĤt is
X
e−iĤt = e−iEn t |ni hn| (3.38)
n
and this decomposition is analytic in the lower half-plane of t, where we have Ĥ |ni = En |ni.
As we have done when evaluating the Gaussian integrals we introduce the Wick rotation
t = −iτ where τ is real and positive, this gives us ẋ = idx/dτ and e−iĤt = e−Ĥτ so that
Ztf Zτf " 2 # Zτf " 2 #

1 2 1 dx 1 dx
i dt mẋ − V (x) = i(−i) dt − m − V (x) = − dt m + V (x) .
2 2 dτ 2 dτ
ti τi τi
(3.39)
Consequently, the path integral becomes

D E
hxf , tf | xi , ti i = xf , tf |e−Ĥ(τf −τi ) | xi , ti
 τf " #
Z Z 2
1 dx
= D̄x exp − dτ m + V (x)  , (3.40)
2 dτ
τi
where D̄x is the integration measure in the imaginary time τ . Equation (3.40) shows the
connection between the functional approach and statistical mechanics.
Let us now the define partition function [1], [2], [4], [5] and [6] of a Hamiltonian Ĥ as
Z(β) = Tre−β Ĥ , (3.41)
where β is a positive constant and the trace is over the Hilbert space associated with Ĥ.
This partition can be written in terms of eigenstates of energy {|En i} with
Ĥ |En i = En |En i , hEm | En i = δmn . (3.42)
In this case
XD E X
X −βEn
Z(β) = En |e−β Ĥ | En = En |e−βEn | En = e , (3.43)
n n n
or in terms of the eigenvector |xi of the position operator x̂,

Z D E
Z(β) = dx x|e−β Ĥ | x . (3.44)
Initially we had an arbitrary β but if we set it to be β = iT we find that

D E D E
xf |e−iĤT | xi = xf |e−β Ĥ | xi , (3.45)
and from this we have the path integral expression of the partition function
 β 
Z Z Z
1
Z(β) = dz D̄x exp − dτ mẋ2 + V (x) 
2
x(0)=x(β)=z 0
Zβ
 
Z
1
= D̄x exp − dτ mẋ2 + V (x) , (3.46)
2
periodic 0
where the periodic integral indicates that the integral is over all paths which are periodic
in the interval [0, β].
When we apply this to the harmonic oscillator, the partition function is simply
∞
X
−β Ĥ
Tre = e−β(n+1/2)ω . (3.47)
n=0
Although there are a number of ways of evaluating the partition function, here we choose
one where the use of the ζ regularization is illustrated. Proceed as follows. Set imaginary
time τ = iT to obtain the path integral
d2 d2
Z Z Z Z
i 1
Dz exp dt z − 2 − ω 2 z → D̄z exp − dτ z − 2 + ω 2 z ,
2 dt 2 dt
z(0)=z(T )=0 z(0)=z(β)=0
(3.48)
in this case D̄z indicates the path integration measure with imaginary time. Suppose we
have an n × n Hermitian matrix M with positive-definite eigenvalues λk where 1 ≤ k ≤ n
then we show in the Appendix (A.4) that (see Rajantie [3])
 ∞  " #
n n
π n/2
Z
Y 1 X Y 1
 dxk  exp − xp Mpq xq = π n/2 √ = . (3.49)
2 p,q λk det M
k=1 −∞ k=1
This is a matrix generalization of the scalar Gaussian integral

Z∞ r
1 2π
dx exp − λx2 = , λ > 0.
2 λ
−∞
The next task is to define the determinant of an operator

Q O by the infinite product of its
eigenvalues λk . This is accomplished by setting DetO = k λk . Note that Det with capital
d denotes the determinant of an operator, whereas with a small d it denotes the determinant
of a matrix, same applies for Tr and tr.
3.4 Zeta regularization solution of the bosonic partition function

Using this we can we can write the integral over imaginary time as
−1/2
d2 d2
Z Z
1
D̄z exp − dτ z − 2 + ω 2 z = DetD − 2 + ω 2 (3.50)
2 dt dτ
z(0)=z(β)=0
here the D denotes the Dirichlet boundary condition z(0) = z(β) = 0.

Similarly to what we did with (3.18) we see that the general solution is
∞
1 X nπτ
z(τ ) = √ zn sin . (3.51)
β n=1 β
We are restricted to having the coefficients zn real as z is a real function. We are now in
a position to write formal expressions. Knowing that the eigenvalues of the eigenfuction
sin(nπτ /β) are λn = (nπ/β)2 + ω 2 we may write the determinant of the operator as
∞ ∞ ∞ ∞
" # " 2 #
2 2 Y
d2
Y
2
Y nπ 2
Y nπ βω
DetD − 2 + ω = λn = +ω = 1+ .
dτ n=1 n=1
β n=1
β m=1
mπ
(3.52)
It is now time to identify the first infinite product with the functional determinant, i.e.
∞ 2
d2
Y nπ
DetD − 2 ↔ . (3.53)
dτ n=1
β
Here is where the ζ function comes into play. Suppose O is an operator with positive-define
eigenvalues λn . Following [1], [4], [5] in this case, we take the log
Y X
log DetO = log λn = Tr log O = log λn . (3.54)
n n
Now we define the spectral ζ function as

X
ζO (s) := λ−s
n . (3.55)
n
The sum converges for sufficiently large Re(s) and ζO (s) is analytic in s in this region.
Additionally, it can be analytically continued to the whole s plane except at a possible
finite number of points. The derivative of the spectral ζ function is linked to functional
determinant by

dζO (s) X
=− log λn . (3.56)
ds s=0
n
And therefore the expression for DetO is

dζO (s)
DetO = exp − . (3.57)
ds s=0
The operator we are interested in is O = −d2 /dτ 2 so this yields

∞ −2s 2s
X nπ β
ζ−d2 /dτ 2 (s) = = ζ(2s). (3.58)
n=1
β π
As we proved in Chapter 1, the ζ function is analytic over the whole complex s plane except
at the simple pole at s = 1. The values (2.30) and (2.33)
1 1
ζ(0) = − ζ 0 (0) = − log(2π)
2 2
were also calculated, and we can use them now to obtain

0 β
ζ−d2 /dτ 2 (0) = 2 log ζ(0) + 2ζ 0 (0) = − log(2β). (3.59)
π
Putting this into the expression for the determinant with Dirichlet conditions we have
d2

DetD − 2 = elog(2β) = 2β (3.60)
dτ
and finally
∞
" 2 #
d2

Y βω
DetD − 2 + ω2 = 2β 1+ . (3.61)
dτ p=1
pπ
Note how the infinite product now becomes finite due to ζ(0) and ζ 0 (0).
Let us go back to the partition function
∞
" ∞
( 2 )#−1/2 1/2
−β Ĥ
X
−β(n+1/2)ω
Y βω π
Tre = e = 2β 1+
n=0 p=1
pπ ω tanh(βω/2)
−1/2 1/2
π π 1
= 2β sinh βω = . (3.62)
βω ω tanh(βω/2) 2 sinh(βω/2)
3.5 Alternative solution

It is important to note that there is a more direct way of computing this partition function
and which is more satisfying for solvable cases such as the harmonic oscillator but which
fails under more complicated Lagrangians.
This can be done by computing
Z D E
Z(β) = Tre−β Ĥ = dx x| exp(−β Ĥ)|x .
Recall that
1/2
ω iω 2
(xf + x2i ) cos ωT − 2xi xf ,

hxf , tf | xi , ti i = exp
2πi sin ωT 2 sin ωT
so that
1/2 Z
ω ω
Z(β) = dx exp i (2x2 cosh βω − 2x2 )
2πi(−i sinh βω) −2i sinh βω
1/2 1/2
ω π 1
= = .
2π sinh βω ω tanh βω/2 2 sinh(βω/2)
3.6 The fermionic partition function

The quantisation of bosonic particles is done by using commutation relations, however, the
quantisation of fermionic particles require a more different approach, namely that of anti-
commutation relations. This in turn requires anti-commuting numbers which are called
Grassmann numbers. We continue the presentation from [1], [4], [5] and [6].
In analogy with the bosonic harmonic oscillator which was described by the Hamiltonian,
H = 21 (a† a + aa† ), (3.63)
where a and a† satisfy the commutation relations
[a, a† ] = 1, [a, a] = [a† , a† ] = 0. (3.64)
The Hamiltonian has eigenvalues (n + 21 )ω where n is an integer with eigenvector |ni
H |ni = (n + 21 )ω |ni . (3.65)
From now on we drop the hat notation in the operators whenever there is no risk for
confusion with the eigenvalue.
The prescription for the fermionic Hamiltonian is to set
H = 12 (c† c − cc† )ω. (3.66)
This may be thought of as a Fourier component of the Dirac Hamiltonian, which describes
relativistic fermions. However, it is evident that if c and c† were to satisfy commutation
relations then the Hamiltonian would be a constant, and therefore it is more appropriate
to consider anti-commutation relations
†
{c, c} = c† , c† = 0.

c, c = 1, (3.67)
In this case the Hamiltonian becomes
H = 12 c† c − (1 − c† c) ω = N − 21 ω,

(3.68)
where N = c† c. The eigenvalues of N are either 0 or 1 since
N 2 = c† cc† c = N ⇔ N (N − 1) = 0.
Next we need a description of the Hilbert space of the Hamiltonian. To this end, let |ni be
an eigenvector of H with eigenvalue n (necessarily n = 0, 1). Then the following equations
hold
ω ω
H |0i = − |0i , H |1i = |1i . (3.69)
2 2
For the sake of convenience we introduce the spin-notation

0 1
|0i = , |1i = , (3.70)
1 0
c† |0i = |1i , c† |1i = 0, c |0i = 0, c |1i = |0i .

When the basis vectors of the space have this form then the operators take the matrix
representations

0 0 † 0 1 1 0 ω 1 0
c= , c = , N= , H= . (3.71)
1 0 0 0 0 0 2 0 −1
Instead of having the bosonic commutation relation [x, p] = i we now have [x, p] = 0. The
anti-commutation relation {c, c† } = 1 is replaced by {θ, θ∗ } = 0 where θ and θ∗ are anti-
commuting numbers, i.e. Grassmann numbers which we proceed to develop in further detail
in the appendix.
The Hamiltonian of the fermionic harmonic oscillator is H = (c† c − 1/2)ω, with eigenvalues
±ω/2. From our previous discussion of the partition function and Grassmann numbers we
know that
1
X
Z(β) = Tre−βH = n|e−βH | n = eβω/2 + e−βω/2 = 2 cosh(βω/2).

(3.72)
n=0
As with the bosonic case, we can evaluate Z(β) in two different ways using the path integral
formalism. Let us start with some preliminary results. Let H be the Hamiltonian of a
fermionic harmonic oscillator, its partition function is written as
Z
∗
−βH
= dθ∗ dθ −θ|e−βH | θ e−θ θ .

Tre (3.73)
We can show this by inserting (A.59) into the partition function (3.72), i.e.
Z
X
X ∗
−βH
dθ∗ dθ |θi hθ| e−θ θ n|e−βH | n

Z(β) = n|e |n =
n=0,1 n
XZ
∗ −θ ∗ θ
hn | θi θ|e−βH | n

= dθ dθe
n
XZ
dθ∗ dθ(1 − θ∗ θ) (hn | 0i + hn | 1i θ) 0|e−βH | n + θ∗ 1|e−βH | n .

=
n
Consequently we have
XZ
Z(β) = dθ∗ dθ(1 − θ∗ θ)
n
0|e−βH | n hn | 0i − θ∗ θ 1|e−βH | n hn | 1i + θ 0|e−βH | n hn | 1i + θ∗ 1|e−βH | n hn | 0i .

×
(3.74)
Note now that the last term of the integrand does not contribute to the integral and therefore
we may substitute θ∗ to −θ∗ which implies that
XZ
Z(β) = dθ∗ dθ(1 − θ∗ θ)
n
0|e−βH | n hn | 0i − θ∗ θ 1|e−βH | n hn | 1i + θ 0|e−βH | n hn | 1i − θ∗ 1|e−βH | n hn | 0i

×
Z
∗
= dθ∗ dθe−θ θ −θ|e−βH | θ .

(3.75)
Unlike the bosonic case, we have to impose an anti-periodic boundary over [0, β] in the trace
since the Grassmann variable is θ when τ = 0 and it is −θ when τ = β.
By invoking the expression
e−βH = lim (1 − βH/N )N (3.76)
N →∞
and inserting the completeness relation (A.59) at each step one has the following expression
for the partition function
Z
∗
dθ∗ dθe−θ θ −θ|(1 − βH/N )N | θ

Z(β) = lim
N →∞
−1
Z N
" N −1 #
Y X
∗ ∗ ∗
= lim dθ dθ dθk dθk exp − θn θn
N →∞
k=1 n=1
× h−θ|(1 − εH)| θN −1 i hθN −1 | · · · | θ1 i hθ1 |(1 − εH)| θi
Z YN
" N #
X
∗ ∗
= lim dθk dθk exp − θn θn
N →∞
k=1 n=1
× hθN |(1 − εH)| θN −1 i hθN −1 | · · · | θ1 i hθ1 |(1 − εH)| − θN i , (3.77)
where we have the usual conventions and we have been using all along
∗ ∗
ε = β/N and θ = −θN = θ0 , θ∗ = −θN = −θ0 . (3.78)
Matrix elements are evaluated (up to first order) as

hθk |H| θk−1 i
hθk |(1 − εH)| θk−1 i = hθk | θk−1 i 1 − ε
hθk | θk−1 i
= hθk | θk−1 i exp(−ε hθk |H| θk−1 i / hθk |θk−1 i)
∗ ∗ ∗
= eθk θk−1 e−εω(θk θk−1 −1/2) = eεω/2 e(1−εω)θk θk−1 . (3.79)
In terms of the path integral the partition function becomes (2.113)
N Z
" N # " N
#
Y X X
βω/2 ∗ ∗ ∗
Z(β) = lim e dθk dθk exp − θn θn exp (1 − εω) θn θn
N →∞
k=1 n=1 n=1
N Z " N
#
Y X
=e βω/2
lim dθk∗ dθk exp − {θn∗ (θn − θn−1 ) + εωθn∗ θn−1 } ,
N →∞
k=1 n=1
and upon simplification of the exponential [1], [6]

N Z
Y †
Z(β) = eβω/2 lim dθk∗ dθk exp(−θ · B · θ), (3.80)
N →∞
k=1
where we have the following vector and matrix elements

 
  1 0 ··· 0 −y
θ1  y
 1 0 ··· 0 

θ =  ...  , θ = θ1∗ · · · θN ···
†
∗
, BN =  0 y 1 0

, (3.81)
   
 .. .. .. .. .. 
θN  . . . . . 
0 0 ··· y 1
with y = −1 + εω.
The computation is ended by recalling that from the definition of the Grassmann Gaussian
integral one had
βω
Z(β) = eβω/2 lim det BN = eβω/2 lim 1 + (1 − βω/N )N = eβω/2 (1 + eβω ) = 2 cosh

.
N →∞ N →∞ 2
(3.82)
As with the bosonic partition function, we can arrive to the same result using the ζ function
(a generalization of it), and this will prove useful later.
Recall that we showed that
N Z
Y †
Z(β) = e βω/2
lim dθk∗ dθk exp(−θ · B · θ)
N →∞
k=1
Zβ
 
Z
d
= eβω/2 Dθk∗ Dθk exp − dτ θ∗ (1 − εω) + ω θ
dτ
0

βω/2 d
=e Detθ(β)=−θ(0) (1 − εω) +ω . (3.83)
dτ
The subscript θ(β) = −θ(0) indicates that the eigenvalue should be evaluated for the
solutions of the anti-periodic boundary condition θ(β) = −θ(0).
First, we expand the orbit θ(τ ) in the Fourier modes. The eigenmodes and the corresponding
eigenvalues are

πi(2n + 1)τ πi(2n + 1)
exp , (1 − εω) + ω, (3.84)
β β
where n runs as n = 0, ±1, ±2, · · · . The number of degrees of freedom is N (= β/ε) so the
coherent states are (over)complete. The presence ε in operator will account for the fact
that the infinite contribution of the eigenvalues is finite.
3.7 Zeta regularization solution of the fermionic partition function

Since one complex variable has two real degrees of freedom, we need to truncate the product
at −N/4 ≤ k ≤ N/4. Following this prescription, one has
N/4
βω/2
Y π(2n − 1)
Z(β) = e lim i(1 − εω) +ω
N →∞ β
k=−N/4
∞
" 2 #
Y
βω/2 −βω/2 2π(n − 1/2) 2
=e e +ω
β
k=1
∞ 2 ∞ " 2 #
Y π(2k − 1) Y βω
= 1+ . (3.85)
β n=1
π(2n − 1)
k=1
The trouble comes from the first infinite product, Ξ, which is divergent and as such it is in
need of ζ values to become finite. This can be accomplished as follows
∞
X 2π(k − 1/2)
log Ξ = 2 log , (3.86)
β
k=1
and we define the corresponding ζ function by (which is the Hurwitz ζ function)

∞ −s s
X 2π(k − 1/2) β
ζfermion (s) = = ζ(s, 1/2), (3.87)
β 2π
k=1
where, see Chapter 2 Eq (2.11),

∞
X
ζ(s, a) = (k + a)−s ,
k=0
with 0 < a < 1. This gives

0
Ξ = exp(−2ζfermion (0)). (3.88)
So now we are left with the issue of differentiating the ζfermion function at s = 0 which is
done as follows [1], [6]

0 β 1
ζfermion (0) = log ζ(0, 1/2) + ζ 0 (0, 1/2) = − log 2, (3.89)
2π 2
since we showed in Chapter 2 (Theorem 10) that
ζ(0, 12 ) = 0, ζ 0 (0, 21 ) = − 12 log 2.
Putting all of this together, we obtain the surprising result

0
Ξ = exp(−2ζfermion (0)) = elog 2 = 2. (3.90)
This result indicates that Ξ is independent of β once the regularization is performed. Finally,
the partition function is evaluated using all these facts
∞
" 2 #
Y βω βω
Z(β) = 2 1+ = 2 cosh , (3.91)
n=1
π(2n − 1) 2
by the virtue of the formula

∞
" 2 #
Y x x
1+ = cosh . (3.92)
n=1
π(2n − 1) 2
Note the similarity between (3.35) and (3.92) as well as (3.62) and (3.91) for the bosonic
and fermionic cases respectively.
Richard Feynman was an advocate of using solutions of known problems in unknown prob-
lems, quoting him ’The same equations have the same solutions’. The rationale behind this
statement is once we solve a mathematical problem, we can re-use the solution in another
physical situation. Feynman was skilled in transforming a problem into one that he could
solve. This is precisely what we have done in this case. We can prove (3.92) using path
integrals.
REFERENCES
The discussion of the role of the ζ function in quantum mechanics and zeta function regu-
larization as explained in the above chapter can be traced back to
- [1] Hagen Kleinert’s Path Integrals in Quantum Mechanics [pages 81 to 83, 161 to 163
and 600 to 614]
- [2] Peskin and Schroeder [pages 299 to 301] and
- [3] Advanced Quantum Field Theory course by Rajantie for the development of Grassmann
numbers.
It is also complemented from
- [4] C. Grosche and F. Steiner Handbook of Feynman Path Integrals [pages 37 to 44, 55 to
59 for the fermionic case] and
- [5] EOR’s [p 65 to 67] for the fermionic case.
- [6] The Grassmann and fermionic cases follow from Nakahara’s [p 47 to 51] approach.
4 DIMENSIONAL REGULARIZATION 53
4 Dimensional Regularization
4.1 Generating functional and probability amplitudes in the pres-
ence of a source J
From the quantum mechanical case we can build a generalization with several degrees of
freedom: a field theory. We will exclusively deal with φ4 , where φ(x) is a real scalar field.
Let us summarize standard results from field theory from Peskin and Schroder [2] as well
as Rajantie [3]. The action is built from the Lagrangian
Z
S = dx L (φ(x), ∂µ φ(x)), (4.1)
where it is understood that L is the Lagrangian density. The equations of motion (EOM)
are given by the Euler-Lagrange equation
∂ ∂L ∂L
=− . (4.2)
∂xµ ∂(∂µ φ(x)) ∂φ(x)
From the free scalar field Lagrangian
1
L0 (φ(x), ∂µ φ(x)) = − (∂µ ∂ µ φ + m2 φ2 ) (4.3)
2
we can derive the Klein-Gordon equation
(∂µ ∂ µ − m2 )φ = 0. (4.4)
When there is a source J present the vacuum amplitude has functional representation
Z Z
i 2
h0, ∞|0, −∞iJ = Z[J] = N Dφ exp i dx L0 + Jφ + εφ , (4.5)
2
with the artificial iε is added to make sure the integral converges. We can think of J as
driving force, i.e. at any time we are allowed to drive the system in any arbitrary way and
measure the response. Integrating by parts we obtain
Z Z
i 2
Z[J] = Dφ exp i dx L0 + Jφ + εφ
2
Z Z
1
= Dφ exp i dx µ 2 2

φ(∂µ ∂ φ − m )φ + iεφ + Jφ . (4.6)
2
In this case, the Klein-Gordon equation becomes the slightly more generalized equation
(∂µ ∂ µ − m2 + iε)φc = −J. (4.7)
Working in d dimensions and defining the Feynman propagator as
−1 eik(x−y)
Z
d
∆(x − y) = d k (4.8)
(2π)d k 2 + m2 − iε
the solution to the generalized Klein-Gordon equation becomes
Z
φc (x) = − dy∆(x − y)J(y). (4.9)
The Feynman propagator obeys
(∂µ ∂ µ − m2 + iε)∆(x − y) = δ d (x − y). (4.10)
Hence the vacuum amplitude can be written in terms of the source J as

Z
i
h0, ∞|0, −∞iJ = N exp − dxdyJ(x)∆(x − y)J(y) , (4.11)
2
or
Z
i
Z0 [J] = Z0 [0] exp − dxdyJ(x)∆(x − y)J(y) (4.12)
2
by setting h0, ∞|0, −∞iJ := Z0 [J]. The Feynman propagator is also computed by the
functional derivative of Z0 [J]
δ 2 Z0 [J]

i
∆(x − y) = . (4.13)
Z0 [0] δJ(x)δJ(y) J=0
In order to evaluate Z0 [0] (which is the vacuum to vacuum amplitude when there is no
source) we need to introduce imaginary time x4 = t = ix0 and operator ∂¯µ ∂¯µ = ∂τ2 + ∇2 so
that
Z Z
1 1
Z0 [0] = D̄φ exp dxφ(∂¯µ ∂¯µ − m2 )φ = q , (4.14)
2 ¯
Det(∂µ ∂¯µ − m2 )
with capital d, the determinant is the product of eigenvalues with corresponding boundary
condition. With term sources, the Lagrangian of the free complex scalar field takes the form
2
L0 = −∂µ φ∗ ∂ µ φ − m2 |φ| + Jφ∗ + J ∗ φ, (4.15)
consequently the generating functional becomes

Z Z
2
Z0 [J, J ∗ ] = DφDφ∗ exp i dx(L0 − iε |φ| )
Z Z
= DφDφ∗ exp i dxφ∗ (∂µ ∂ µ − m2 − iε)φ + Jφ∗ + J ∗ φ , (4.16)
and upon differentiation we obtain the propagator
δ 2 Z0 [J, J ∗ ]

i
∆(x − y) = . (4.17)
Z0 [0, 0] δJ ∗ (x)δJ(y) J=J ∗ =0
We may split the function by virtue of the Klein-Gordon equations (∂µ ∂ µ − m2 )φ = −J

and (∂µ ∂ µ − m2 )φ∗ = −J ∗
Z
Z0 [J, J ∗ ] = Z0 [0, 0] exp −i dxdyJ ∗ (x)∆(x − y)J(y) (4.18)
and by using another Wick rotation we have

Z Z
1
Z0 [0, 0] = DφDφ exp −i dxφ (∂µ ∂ − m − iε)φ =
∗ ∗ µ 2
. (4.19)
Det(∂¯µ ∂¯µ − m2 )
The presence of a potential in the Lagrangian
L (φ, ∂µ φ) = L0 (φ, ∂µ φ) − V (φ) (4.20)
comes at a double price: the form of the potential is limited by symmetry and renormal-
ization of the theory and this theory needs to be handled perturbatively. The potential is
α n
usually of the form V (φ) = n! φ where α is a real number that sets the strength of the
interaction and n > 2 is an integer. As with the free theory, the generating functional is
[2], [4]
Z Z
1
Z[J] = Dφ exp i dx φ(∂µ ∂ µ − m2 )φ − V (φ) + Jφ
2
Z Z Z
= Dφ exp −i dxV (φ) exp i dx (L0 (φ, ∂µ φ) + Jφ)
Z Z Z
δ
= exp −i dxV −i Dφ exp i dx (L0 (φ, ∂µ φ) + Jφ)
δJ(x)
∞ Z
(−i)k
Z
X δ δ
= dx1 · · · dxk V −i · · · V −i Z0 [J]. (4.21)
k! δJ(x1 ) δJ(xk )
k=0
The Green function (which is the vacuum expectation of the order time product of field
operators)
(−i)n δ n

Gn (x1 , · · · , xn ) := h0|T [φ(x1 ) · · · φ(xn )]|0i = Z[J] (4.22)
δJ(x1 ) · · · δJ(xn ) J=0
is generated by the generating functional Z[J].

However, we can see that this is the nth functional derivative of Z[J] around J = 0 and
therefore we may plug it into the Taylor expansion of the exponential above and we obtain
∞
" n Z # Z
X 1 Y
Z[J] = dxi J(xi ) h0|T [φ(x1 ) · · · φ(xn )]|0i = 0 T exp dxJ(x)φ(x) 0 .

k! i=1
k=1
(4.23)
Connected n-point functions are generated by
Z[J] = exp(−W [J]), (4.24)
and the effective action is defined by a Legendre transformation as follows

Z
Γ[φcl ] := W [J] − dτ dxi Jφcl (4.25)
where
δW [J]
φcl := hφiJ = . (4.26)
δJ
We will also see that Γ[φcl ] generates 1-particle irreducible diagrams (see Bailin and Love
[1] and Ramond [4]).
It is convenient now to derive the above discussion in a formal manner and with a closer
analogy to statistical mechanics. The generating functional of correlation functions for a

field theory with Lagrangian L is given by (4.5)
Z Z
Z[J] = Dφ exp i d4 x(L + Jφ) , (4.27)
where the time variable is contained between −T and T , with T → ∞(1 − iε). Furthermore
we have the following
h R i
T
Dφφ(x1 )φ(x2 ) exp i −T d4 xL
R
h0|T φ(x1 )φ(x2 )|0i = lim h R
T
i
T →∞(1−iε) Dφ exp i −T d4 xL
R

δ δ
= Z[J]−1 −i

−i Z[J] . (4.28)
δJ(x1 ) δJ(x2 ) J=0
Let us do some manipulations on the time variable; when we derived the path integral
formulation of quantum mechanics (see AQFT [3]) it was shown that the time integration
was tilted into the complex plane in the direction that would allow the contour of integration
to be rotated clockwise onto the imaginary axis. We assumed that the original infinitesimal
rotation gives the correct imaginary infinitesimal to produce the Feynman propagator.
Now, the wick rotation of the time coordinate t → −ix0 yields a Euclidean 4-vector product
2 2 2
x2 = t2 − |x| → −(x0 )2 − |x| = − |xE | , (4.29)
and similarly we assume that the analytic continuation of the time variables in any Green’s
function of a quantum field theory produces a correlation function invariant under the
rotational symmetry of four-dimensional Euclidean space.
4.2 Functional energy, action and potential and the classical field
φcl (x)
Let us now apply this to the φ4 theory. As we know the action in this case is
Z Z
1 1 λ
S = d4 x(L + Jφ) = d4 x (∂µ φ)2 − m2 φ2 − φ4 + Jφ , (4.30)
2 2 4!
and performing the Wick rotation

Z Z
4 4 1 2 1 2 2 λ 4
i d xE (LE − Jφ) = i d xE (∂Eµ φ) + m φ − φ − Jφ , (4.31)
2 2 4!
gives the Wick-rotated generating functional

Z Z
Z[J] = Dφ exp − d4 xE (LE − Jφ) . (4.32)
The functional LE [φ] is bounded from below and when the field φ has large amplitude or
large gradient the functional becomes large. These two facts would imply that LE [φ] has
the form of an energy and consequently it is a possible candidate for a statistical weight for
the fluctuations of φ.
Within this light, the Wick rotated functional Z[J] is the partition function describing the
statistical mechanics of a macroscopic system when approximating the fluctuating variable
as a field.
Finally, let us push this analogy between field theory and statistical mechanics further by
presenting the Green’s function of φ(xE ) [2]
d kE eikE ·(xE1 −xE2 )

Z 4
hφ(xE1 )φ(xE2 )i = 2 + m2 , (4.33)
(2π)4 kE
which is in fact the Feynman propagator evaluated in the spacelike region and this falls off
as exp(−m|xE1 − xE2 |). This correspondence between quantum field theory and statistical
mechanics plays an important part in understanding ultraviolet divergences.
Recalling the generating functional of correlation functions, we define an energy functional
E[J] by
Z Z
Z[J] = exp(−iE[J]) = Dφ exp i d x(L + Jφ) = Ω|e−iHT |Ω ,
4

(4.34)
with the constraints on time explained above. Note the −i factor in the exponential in
contrast to (4.24). The functional E[J] is, as we have said before, the vacuum energy as
a function of the external source J. Let us perform the functional derivative of E[J] with
respect to J(x)
δ δ
E[J] = i log Z
δJ(x) δJ(x)
Z Z −1 Z Z
=− Dφ exp i d x(L + Jφ)
4
Dφφ(x) exp i d x(L + Jφ)
4
(4.35)
and set
δ
E[J] = − hΩ|φ(x)|ΩiJ (4.36)
δJ(x)
the vacuum expectation value in the presence of a source J.
Next, we define
- the classical field as
φcl (x) = hΩ|φ(x)|ΩiJ , (4.37)
a weighted average over all possible fluctuations, and dependent on the source J ;
- the effective action as the Legendre transformation of E[J] i.e. as in (4.25)
Z
Γ[φcl ] := −E[J] − d4 x0 J(x0 )φcl (x0 ). (4.38)
By virtue of (4.36) we have the following

δJ(x0 )
Z
δ δ
Γ[φcl ] = − E[J] − d4 x0 φcl (x0 ) − J(x)
δφcl (x) δφcl (x) δφcl (x)
δJ(x0 ) δE[J] δJ(x0 )
Z Z
= − d4 x0 0
− d4 x0 φcl (x0 ) − J(x) = −J(x) (4.39)
δφcl (x) δJ(x ) δφcl (x)
which means that when the source is set to zero, the equation
δ
Γ[φcl ] = 0 (4.40)
δφcl (x)
is satisfied by the effective action. This equation has solutions which are the values of
hφ(x)i in the stable states of the theory. It will be assumed that the possible vacuum states
are invariant under translation and Lorentz transformations. This implies a substantial
simplification of (4.40) as for each possible vacuum state the corresponding solution φcl will
be independent of x, and hence it is just solving an ODE of one variable.
Thermodynamically, Γ is proportional to the volume of the spacetime region over which
the functional integral is taken, and therefore, it can be a large quantity. Consequently, in
terms of volume V and associated T of the region we may write
Γ[φcl ] = −(V T )Veff (φcl ), (4.41)
where Veff is the effective potential. In order that Γ[φcl ] have an extremum we need the
following to hold
δ
Veff (φcl ) = 0. (4.42)
δφcl
Each solution of (4.42) is a translation invariant state without source, i.e J = 0. Therefore,
the effective action (4.38) is −E in this case (Γ = −E) and consequently Veff (φcl ) evaluated
at the solution of (4.42) is the energy density of the corresponding state.
The effective potential defined by (4.41) and (4.42) yields a function whose minimization
defines the exact vacuum state of the field theory including all effects of quantum corrections.
The evaluation of Veff (φcl ) will follow from the path integral formulation. In order to
accomplish this we will follow Peskin and Schroeder’s method which in turn follows from
R. Jackiw [5] and dates back to 1974. The idea is to compute the effective action Γ directly
from its path integral definition and then obtain Veff by focusing on constant values of φcl .
Because we are using renormalized perturbation theory, the Lagrangian
1 1 λ0
L = (∂µ φ)2 − m20 φ2 − φ4 (4.43)
2 2 4!
ought to be split as
L = L1 + δL , (4.44)
which is analogous to the split of the Lagrangian in renormalized φ4 theory done in AQFT,
see Rajantie [3] and Bailin and Love Chp 7.4 [1], i.e. rescaling the field φ = Z 1/2 φr where
Z is the residue in the LSZ reduction formula
Z
iZ
d4 x hΩ | T φ(x)φ(0)| Ωi eip·x = 2 + (terms regular at p2 = m2 ), (4.45)
p − m2
with m being the physical mass. This rescale changes the Lagrangian into
1 1 λ0 1 1 δλ
L= (∂µ φr )2 − m20 φ2r − φ4r + δZ (∂µ φr )2 − δm φ2r − φ4r , (4.46)
2 2 4! 2 2 4!
by the use of δZ = Z − 1, δm = m20 Z − m2 and δλ = λ2 Z − λ where m and λ are
physically measured. The last three terms are known as the counter terms and they take
into account the infinite and unobservable shifts between the bare parameters and the
physical parameters.
At the lowest order in perturbation theory the relationship between the source and the
classical field is

δL
+ J(x) = 0. (4.47)
δφ φ=φcl
Because the functional Z[J] depends on φcl through its dependence on J our goal is to
compute Γ as a function of φcl . This will be the starting point.
Next, we define J1 to be the function that exactly satisfies the classical field equation above
for higher orders, i.e. when L = L1

δL1
+ J1 (x) = 0, (4.48)
δφ φ=φcl
and the difference between both sources J and J1 will be written as (see Peskin and
Schroeder 11.4)
J(x) = J1 (x) + δJ(x), (4.49)
where δJ has to be determined, order by order in perturbation theory by use of (4.37), that
is by using the equation hφ(x)iJ = φcl (x). We may now write (4.34) as
Z Z Z
e−iE[J] = Dφ exp i d4 x(L1 [φ] + J1 φ) exp i d4 x(δL [φ] + δJφ) , (4.50)
where all the counter terms are in the second exponential. Let us concentrate on the first
exponential first. Expanding the exponential about φ(x) = φcl (x) + η(x) yields
Z Z Z
δL1
d4 x(L1 [φ] + J1 φ) = d4 x(L1 [φcl ] + J1 φcl ) + d4 xη(x) + J1
δφ
δ L1
2
Z
1
+ d4 xd4 yη(x)η(y)
2! δφ(x)δφ(y)
δ 3 L1
Z
1
+ d4 xd4 yd4 zη(x)η(y)η(z) + ··· (4.51)
3! δφ(x)δφ(y)δφ(z)
and it is understood that the functional derivatives of L1 are evaluated at φcl .

The second integral on the RHS vanishes by (4.48) and therefore the integral over η is a
Gaussian integral, where the perturbative corrections are given by the cubic and higher
terms.
Let us assume that the coefficients of (4.51) (i.e. the successive functional derivatives of
L1 ) give well-defined operators. If we keep only terms up to quadratic order in η and we
only focus of the first integral of (4.50) we find that there is a pure Gaussian integral which
can be evaluated in terms of a functional determinant as we have computed in the Appendix
δ 2 L1
Z Z
1
Dη exp i d x(L1 [φcl ] + J1 φcl +
4 4 4
d xd yη η
2 δφ(x)δφ(y)
−1/2
δ 2 L1
Z
4
= exp i d x(L1 [φcl ] + J1 φcl ) det − , (4.52)
δφ(x)δφ(y)
the lowest-order quantum correction to the effective action is given by the determinant. If
we now consider the second integral of (4.50) which consists of the counter terms of the
Lagrangian and expanding as we have done before we have
(δL [φcl ] + δJφcl ) + (δL [φcl + η] − δL [φcl ] + δJη). (4.53)

The cubic and higher order terms in η in (4.51) produce Feynman diagram expansion of
the functional integral in (4.50) in which the propagator is the inverted operator
−1
δ 2 L1

−i , (4.54)
δφ(x)δφ(y)
and hence when the second term in (4.53) is expanded as a Taylor series in η the successive
terms give counter term vertices which can be included in the above mentioned Feynman
diagrams. The first term is a constant with respect to the integral over η thus it gives
additional terms in the exponent of (4.52).
Taking (4.52) with the contributions from higher order vertices and counter terms together
we obtain an expression for the functional integral (4.50). Feynman diagrams representing
the higher order terms can be arranged in such a way that they yield the exponential of the
sum of the connected diagrams, obtaining the expression for E[J]
δ 2 L1
Z
1
−iE[J] = i d4 x(L1 [φcl ] + J1 φcl ) − log det − + (connected diagrams)
2 δφ(x)δφ(y)
Z
+ i d4 x(δL [φcl ] + δJφcl ), (4.55)
and finally by virtue of (4.49) and (4.38) we finally have [2]

δ 2 L1
Z Z
i
Γ[φcl ] = d4 xL1 [φcl ] + log det − − i(connected diagrams) + d4 x(δL [φcl ]),
2 δφ(x)δφ(y)
(4.56)
or by (4.44)
δ 2 L1
Z
i
Γ[φcl ] = d4 xL [φcl ] + log det − − i(connected diagrams).
2 δφ(x)δφ(y)
This is indeed the expression we were seeking since Γ is a function of φcl , taking away the
J dependence. The Feynman diagrams in the expression for Γ have no external lines and
they all contain at least two loops. The last term of (4.56) gives the counter terms that
are needed for the renormalization conditions on Γ and cancel the divergences that appear
in the evaluation of the determinant and the diagrams. We shall ignore any one-particle
irreducible one-point diagram (these diagrams are cancelled by the adjustment of δJ).
As a side point, note how a calculation of the sort
dd k
Z Z d Z d
2 2 d kEuc 2 2 ∂ d kEuc 2 2 −x

log(−k + m ) = i log(k Euc + m ) = −i (k Euc + m )
(2π)d (2π)d ∂x (2π)d x=0

∂ 1 Γ(x − d/2) 1 Γ(−d/2)
= −imd

= −i
∂x (4π)d/2 Γ(x) m2(x−d/2) x=0 (4π)d/2
(4.57)
yields
X
log det(∂ 2 + m2 ) = tr log(∂ 2 + m2 ) = log(−k 2 + m2 )
k
d4 k
Z
Γ(−d/2) d
= (V T ) 4
log(k 2 + m2 ) = −i(V T ) m
(2π) (4π)d/2
with V (volume) and T (time) as explained before. We could call this Γ renormalization.
More will be said about it later on.
4.3 Derivation of φ4 potential at φcl (x)

In the Unification course, it was shown that spontaneous symmetry breaking requires the
development of a vacuum expectation value (VEV) from the scalar field. Furthermore, the
VEV is determined by the minimisation of the effective potential,
dV
= 0. (4.58)
dφcl
The effective potential is given exclusively by the potential V of the Lagrangian if no
quantum effects are taken into account. Perturbation theory allows us to place the quantum
terms, however this would clash with the non-perturbative nature of spontaneous symmetry
breaking. The alternative parameter is the loop expansion which we now describe.
Let us re-write the theory around (4.24) as follows. The generating functional X of the
connected Green functions in a scalar field theory is
Z Z
−1
Z[J] = exp(i~ X[J]) = N 0
Dφ exp i~ −1 4
d x(L + Jφ) (4.59)
with normalization constant N 0 chosen so that
Z[0] = 1 X[0] = 0. (4.60)
One-particle-irreducible Green functions Γ(n) are generated by the effective action Γ[φcl ] by
the use of [1], [2], [3], [4]
∞ n Z Z
X i
Γ[φcl ] = d x1 · · · d4 xn Γ(n) (x1 , · · · , xn )φcl (x1 ) · · · φcl (xn ),
4
(4.61)
n=1
n!
and we note the equation

(N ) 1 δ δ δ
G (x1 , x2 , · · · , xN ) = N ··· Z[J] .
i δJ1 δJ2 δJN J=0
In (4.59) the factor ~−1 multiplies the whole Lagrangian (not just the interaction part) each
of the V vertices in any diagram will carry a ~−1 factor and each of the I internal lines will
carry a ~ factor. Each of the E external lines in Green functions G̃(E) has a propagator.
With this information, the overall factor is
~−V +I+E = ~L+1−E , (4.62)
by virtue of L = I − V + 1 and noting that E is the number of external lines. Furthermore,

there is a factor of ~L−E in any diagram in expansion of ~−1 X. The one-particle-irreducible
Green functions Γ̃(E) have no propagators, hence the multiplying factor is only ~L . This
means that the power of the ~ indicates the number of loops. When there are no loops
(L = 0) the only non-vanishing Γ̃(E) are
Γ̃(2) (p, −p) = p2 − µ2 (4.63)
and
Γ̃(4) (p1 , p2 , p3 , p4 ) = −λ, (4.64)

which gives the approximation to the potential (see Bailin and Love 4.4)
∞ n
X i (n) 1 1
V (φcl ) = − Γ̃ (0, · · · , 0)φncl ⇒ V0 (φcl ) = µ2 φ2cl + λφ4cl , (4.65)
n=1
n! 2 4!
as classically expected without quantum corrections. Note, however, that the role of ~ is
not central to the discussion as no assumptions about its size have been made, it is only an
expansion parameter. Bailin and Love [201] give same effective action Γ[φcl ] with
Z Z
1 1
Γ[φcl ] = −i log N − λ dx[i∆F (0)]2 − dxφcl (x)(∂ µ ∂µ + µ2 )φcl (x)
8 2
Z Z
1 1
− λi∆F (0) dxφ2cl (x) − λ dxφ4cl (x) + O(λ2 ). (4.66)
4 24
The terms that contain a Feynman propagator ∆F (0) come from divergent loop integrals
and hence they do not contribute in zeroth order, consequently we may write
Z Z
1 1
Γ0 [φcl ] = − dxφcl (x)(∂ µ ∂µ + µ2 )φcl (x) − λ dxφ4cl (x) (4.67)
2 4!
again taking into account that N = 1 implying Γ[0] = 0.
The first order accuracy of ~, the loop expansion that is, of the effective potential V and
effective action Γ can be computed by writing
φ(x) = φ0 (x) + φ̃(x), (4.68)
with φ0 being the zeroth order approximation to φcl . Hence this shift in the functional
integration variable must satisfy the following EOM
λ 3
(∂ µ ∂µ + µ2 )φ0 (x) + φ (x) = J(x). (4.69)
6 0
When we plug this change into the Lagrangian density
1 µ 1 1
L (φ) = (∂ φ)(∂µ φ) − µ2 φ2 − λφ4 (4.70)
2 2 4!
we have the following integral
Z Z
d x(L + Jφ) = d4 x(L (φ0 (x)) + Jφ0 )
4
Z
4 µ 2 1 3
+ d x (∂ φ̃)(∂µ φ0 ) − µ φ̃φ0 − λφ̃φ0 + J φ̃
6
Z
1 3 1 4
+ d x L2 (φ̃, φ0 ) − λφ̃ φ0 − λφ̃
4
(4.71)
6 4
with L2 accounting for all the quadratic leftovers in φ̃ when the shift is performed, i.e.
1 µ 1 1
L2 = (∂ φ̃)(∂µ φ̃) − µ2 φ̃2 − λφ20 φ̃2 . (4.72)
2 2 4
Now, since φ̃ minimises the classical action then the linear term in φ̃ of (4.71) disappears.
Our next step is to re-scale the integration variable as follows
φ̃ = ~1/2 φ, (4.73)
and plug this back into (4.59) yielding

Z
0 −1
Z[J] = N exp i~ d4 x[L (φ0 ) + Jφ0 ]
Z Z
λ 1/2 3 1
× Dφ exp i d x L2 (φ, φ0 ) − ~ φ φ0 − λ~φ .
4 4
(4.74)
6 24
We can recover a Gaussian integral out of this by noting that we are only interested in
the first-order corrections, hence terms proportional to ~1/2 and ~ may be discarded. This
procedure gives
Z Z
1
4
d xL2 (φ, φ0 ) = − d4 xd4 x0 φ(x0 )A(x0 , x, φ0 )φ(x) (4.75)
2
with

1 2
0
A(x , x, φ0 ) = −∂x µ ∂x + µ + λφ0 δ(x0 − x),
0
µ 2
(4.76)
2
finally obtaining for the value of Z

Z
0 −1 4 1 0
Z[J] ≈ N exp i~ d x[L (φ0 ) + Jφ0 ] exp − tr log A(x , x, φ0 ) . (4.77)
2
By our conditions (4.60) and φ0 [0] = 0 we then obtain

0 i 0
Z[0] = 1 ≈ N exp − tr log A(x , x, 0) . (4.78)
2
Here is where we can choose method. We will, instead of using ζ regularization, work out
the terms in the expansion.
Comparison with (4.59) shows that by keeping the same definition of A we then have
Z
X0 [J] = d4 x[L (φ0 ) + Jφ0 ] (4.79)
and
i A(x0 , x, φ0 )
X1 [J] = tr log . (4.80)
2 A(x0 , x, 0)
The effective action is computed by expanding [1], [4]
Γ[φcl ] = Γ0 [φcl ] + ~Γ1 [φcl ] + O(~2 ). (4.81)
Now, since φcl is a functional of J, then
δX[J]
φcl (x) = , (4.82)
δJ(x)
we have the following expansion

Z Z Z
1 µ 1 1
Γ0 [φcl ] = X0 [J] − d4 xJφ0 = d4 xL (φ0 ) = d4 x (∂ φcl )(∂µ φcl ) − µ2 φ2cl − λφ4cl ,
2 2 4!
(4.83)
which is the same as (4.67). The additional term in the expanded effective action is
Z
~Γ1 [φcl ] = X0 [J] − Γ0 [φcl ] − d4 xJφcl + ~X1 [J]
A(x0 , x, φ0 )
Z Z
i~
= d x[L (φ0 ) + Jφ0 ] − d4 x[L (φcl ) + Jφcl ] + tr log
4
. (4.84)
2 A(x0 , x, 0)
Because φ0 is a solution of (4.69) the difference of the two integrals in (4.84) is of order
(φ0 − φcl )2 = O(~2 ), and we can interchange φ0 and φcl at this level of accuracy. Therefore
i A(x0 , x, φcl )
Γ1 [φcl ] = tr log . (4.85)
2 A(x0 , x, 0)
Next, the effective potential V (φcl ) can be derived from Γ[φcl ] by setting φcl constant, in
which case
Z
Γ[φcl ] = − d4 xV (φcl ). (4.86)
Delta functions allow us to diagonalise A(x0 , x, φcl ) which is a prerequisite to properly define
the logarithmic part of (4.85). This is done as follows

1
A(x0 , x, φcl ) = −∂x0 µ ∂xµ + µ2 + λφ2cl δ(x0 − x)
2
4
Z
d k 1 2 0
= −∂ 0
xµ x ∂ µ
+ µ2
+ λφ eik(x −x)
(2π)4 2 cl
d4 k
Z
1 2 0
= −k 2
+ µ 2
+ λφ eik(x −x)
(2π)4 2 cl
d4 k d4 k 0 ix0 k0
Z
1 2
= e −k 2
+ µ2
+ λφ δ(k 0 − k)e−ikx . (4.87)
(2π)2 (2π)2 2 cl
Hence performing log and trace operations

0 0
eix k e−ixk
Z
0 4 4 0 2 2 1 2 0
log A(x , x, φcl ) = d kd k log −k + µ + λφ cl δ(k − k) (4.88)
(2π)2 2 (2π)2
so that
d4 k
Z Z Z
4 4 0 0 0 4 2 2 1 2
tr log A = d xd x δ(x − x) log A(x , x, φcl ) = d x log −k + µ + λφcl .
(2π)4 2
(4.89)
This the formula for the determinant of A since
Tr log A = log det A. (4.90)
Summarizing we can now state the following: the one-loop order contribution to the effective
potential is
−k + µ2 + 21 λφ2cl
2
d4 k

−i
Z
V1 (φcl ) = log (4.91)
2 (2π)4 −k 2 + µ2
and while at this order, the effective potential is approximated by

d4 k 1 λφ2cl
Z
1 2 2 1 4 i
V (φcl ) ≈ V0 (φcl ) + V1 (φcl ) = µ φcl + λφcl − ~ log 1 − , (4.92)
2 4! 2 (2π)4 2 k 2 − µ2
where the Greek parameters are bare parameters, i.e. the ones present in the Lagrangian.
The integral term in the potential is ultraviolet divergent and it requires dimensional regu-
larization to deal with i.e. (see for instance Bailin and Love, p.80)
d2ω k 2
Z 2ω
d 6k 2
Z
−1 1
I(ω, µB ) = 2ω
(k − µ2
B + iε) = (6 k + µ2B )−1
(2π) i (2π)2ω
iµ2B µ2B

2 ω−2 1 0
= (M ) + Γ (1) + 1 − log + O(ω − 2) . (4.93)
16π 2 2−ω 4πM 2
By the use of the following transformations for the bare parameters
φB (x) = Z 1/2 φ(x) Zµ2B = µ2 + δµ2 Z 2 λB = λ + δλ, (4.94)
we can transform the above potential (4.92) to
d2ω k 1 λφ2cl
Z
1 1 1 1 i
V (φcl ) = µ2 φ2cl + λφ4cl + δµ2 φ2cl + δλφ4cl − ~ log 1 − ,
2 4! 2 4! 2 (2π)2ω 2 k 2 − µ2
(4.95)
with the new Greek parameters now being the renormalized parameters rather than the
bare. It is important to note, however, that these parameters have constraints imposed by
the MS scheme. In terms of Green functions, these develop singularities in ω − 2 and these
are neutralized by the counter terms which are [1]
∞
!
4−2ω
X ak (λ̂, M/µ)
δλ = M a0 (λ̂, M/µ, ω) + , (4.96)
(2 − ω)k
k=1
∞
!
2 2
X bk (λ̂, M/µ)
δµ = µ b0 (λ̂, M/µ, ω) + , (4.97)
(2 − ω)k
k=1
∞
X ck (λ̂, M/µ)
δZ = c0 (λ̂, M/µ, ω) + (4.98)
(2 − ω)k
k=1
with a0 , b0 and c0 regular as ω → 2 and λ̂ = λM 2ω−4 . Because we have only simple poles,
the k -integrals only have simple poles at 2 − ω. This means that ∀k > 1 ⇒ ak = bk = 0.
The one-point-irreducible Green functions of the renormalized theory behave as
λ̂µ2 µ2

1
Γ̃2 (p, −p) = p2 (1 + δZ1 ) − µ2 − δµ21 + − γ + 1 − log + O(ω − 2) .
32π 2 2 − ω 4πM
(4.99)
This can be shown by using

d4 k 2
Z
I4 (µB ) = (k − µ2B + iε)−1 (4.100)
(2π)4
and (4.93) on the Feynman propagator
d4 k 2 2
µ2
Z
2 −1 2ω−4 iµ 1
∆F (0) = (k − µB + iε) = M − γ + 1 − log + O(ω − 2) .
(2π)4 16π 2 2 − ω 4πM
(4.101)
This propagator, in turn, is present in the Green function as

1
Γ̃2 (p, −p) = p2 (1 + δZ1 ) + µ2 + λi∆F (0) + δµ21 + O(λ2 ), (4.102)
2
as it can be seen by writing the Feynman diagrams to leading order.

The computation of Γ̃4 is lengthier but follows the same lines; eventually, since Γ̃4 is finite
in any renormalisation scheme, we have that as ω → 2 the following holds
3λλ̂ 1
− δλ2 → const, (4.103)
32π 2 2 − ω
with
∞
X
δλ = δλk . (4.104)
k=2
The finite part of δλ2 is arbitrary.

Going back to our discussion of (4.96) to (4.98) now we see that the poles in 2 − ω are
neutralized by setting
3 2 1
a1 = λ̂ , b1 = λ̂. (4.105)
32π 2 32π 2
Therefore putting these back together in the potential one has

1 2 2 λ 3
V (φcl ) = µ φcl 1 + b0 − − γ + log 4π
2 32π 2 2
3λ2 3

1 4
+ φcl λ + a0 − − γ + log 4π
4! 32π 2 2
" 2 #
2 1 2 2
1 1 µ + λφ cl µ
+ µ2 + λφcl log 2
− µ4 log 2 . (4.106)
64π 2 2 M2 M
It turns out that in the MS scheme we may choose (see Rajantie or Bailin and Love)

3λλ̂ 1
δλMS = − γ + log 4π + O(λ3 ) (4.107)
32π 2 2−ω
3λ̂2 3λ̂2
aMS
0 = [−γ + log 4π] + O(λ3 ), aMS
1 = + O(λ3 ), (4.108)
32π 2 32π 2
This will get rid of the −γ + log 4π factors.
However, and here is the main objective of this comparison with the ζ function regular-
ization, we could also renormalize V (φcl ) by writing it as a function of physical mass and
coupling constant. This can be accomplished by setting aphys

0 and bphys
0 such that the
following hold
d2 V d4 V

2
=µ , = λ. (4.109)
dφ2cl φcl =0 dφ4cl φcl =M
In the case that µ2 is small there could be radiative corrections that generate spontaneous
symmetry breaking. Performing the second and fourth derivatives (4.109) and using
λM 2 8
log 2
=−
2µ 3
leads to the potential [1], [4]
λ2 φ4cl φ2

1 λ 25
V (φcl ) = µ2 φ2cl + φ4cl + log cl2 − . (4.110)
2 4! 256π 2 M 6
4.4 Γ-evaluation of dimensional loop integrals

Let us make a brief digression to see how one would compute integrals of the form
Z∞
Υβ (k) = dβ `F (`, k) (4.111)
−∞
where F behaves as `−2 or `−4 for large `. The key component behind dimensional reg-
ularization is that by lowering the number of dimensions over which one integrates the
divergences trivially disappear [4]. For example as F → `β=−4 in 2-D the integral above
converges at the ultraviolet end.
The precise technique runs as follows, let
Z∞
Υ(ω, k) = d2ω `F (`, k), (4.112)
−∞
where ω is a complex variable. We have to choose a domain where Υ has no singularities

in the ω plane. Carefully choose a function Υ0 which has well-defined singularities outside
the domain of convergence. By analytic continuation (yet again) Υ and Υ0 are the same
function. In order to accomplish this, we first establish a domain of convergence for the
loop integral in the ω plane. Then we construct a function which overlaps with the loop
integral in its domain of convergence but is defined in a large domain which encloses the
point ω = 2 and then take limit ω → 2.
We shall follow the steps of ’t Hooft and Veltman [6] which are simplified in [4]. We start
by splitting up the domain of integration as
d2ω ` → d4 `d2ω−4 `. (4.113)
By introducting polar coordinates and letting L be the length of the 2ω − 4 dimensional `

vector the integral becomes
Z∞
L2ω−5
Z Z
4
Υ= d ` dΩ2ω−4 dL . (4.114)
L2 + `2 + m2
0
By using the familiar result from AQFT [3]
2π ω−2
Z
dΩ2ω−4 = (4.115)
Γ(ω − 2)
the integral now reads

Z∞
2π ω−2 L2ω−5
Z
4
Υ= d ` dL . (4.116)
Γ(ω − 2) L2 + `2 + m2
0
Note that our critieria is not met. This is not well-defined because it is divergent for ω ≥ 1
and integrating over L diverges at the lower, i.e. when ω ≤ 2. Therefore, we do not have
an overlapping region where Υ is well-defined. However,
1 d
L2ω−6 = (L2 )ω−2 . (4.117)
ω − 2 dL2
If we use this to integrate by parts Υ yields
Z∞
ω−2
Z
ω−2 4 1 22 ω−2 d 1
Υ=π d ` dL (L ) − 2 + surface term
Γ(ω − 1) ω−2 dL L2 + `2 + m2
0
(4.118)
ignoring the surface term and simplying

Z∞
π ω−2
Z
d 1
Υ= d4 ` dL2 (L2 )ω−2 − 2 . (4.119)
Γ(ω − 1) dL L2 + `2 + m2
0
This is still too weak as there is no overlap: we have a divergence for ω ≤ 1 and another
one for ω ≥ 1. It would have been sufficient if the divergence had been logarithmic. This
can be cured by performing the same procedure again, that is lowering the dimension
Z∞
π ω−2
Z
4 2 2 ω−1 d d 1
Υ= d ` dL (L ) − 2 − 2 (4.120)
Γ(ω) dL dL L + `2 + m2
2
0
and this equation is well defined for 0 < ω < 1. The process of analytic continuation to the
point ω = 2 requires using the trick

1 ∂L ∂`µ
1= + , (4.121)
5 ∂L ∂`µ
and it is no suprise that integration by parts is the next step

Z∞
2π ω−2 (L2 )ω−1
Z
4 ∂ ∂
Υ=− d ` dL2 `µ + 2L2 2 + 1
5Γ(ω) ∂`µ ∂L (L + `2 + m2 )3
2
0
2 ω−2 Z∞
(L2 )ω−1
Z
3m 2π 4
=− d ` dL2 .
ω − 1 Γ(ω) (L2 + `2 + m2 )4
0
Let us note the pole at ω = 1. In this case, Υ diverges when ω ≥ 2 therefore repeating this
again to increase the exponent of the denominator yields
Z∞
2 · 3 · 4 · m4 π ω−2 (L2 )ω−1
Z
4
Υ=− d ` dL2 . (4.122)
(ω − 1)(ω − 2) Γ(ω) (L2 + `2 + m2 )5
0
This is exactly what we aimed for. The (simple) pole is now at ω = 2 and Υ now converges.
By the means of another well known loop integral identity (see AQFT, Peskin and Schroder,
Bailin and Love, Ramond)
d2ω ` Γ(1 − ω)
Z
1
2 2
= πω (4.123)
` +m Γ(1) (m )1−ω
2
and expanding the Γ function around the poles −n where n is an integer
(−1)n 1

Γ(−n + ε) = + ψ(n + 1) + O(ε) , (4.124)
n! ε
so that

1 1
Γ(1 − ω) = + ψ(2) + O(2 − ω) , (4.125)
2 2−ω
and consequently,
d2ω `
Z
2 2 1
lim = −π m + ψ(2) + O(2 − ω) . (4.126)
ω→2 `2 + m2 2−ω
Finally, we note that

n
X 1
ψ(n) = − γ, (4.127)
k
k=1
hence
3
ψ(2) = − γ, (4.128)
2
and therefore
d2ω `
Z
2 2 1 3
lim = −π m + − γ + O(2 − ω) . (4.129)
ω→2 `2 + m2 2−ω 2
The general integral is
d2ω ` 2 2ω−2n
Γ(n − ω)
Z
2 −n nµ
2ω
(` − µ + iε) = i(−1) ω
. (4.130)
(2π) (4π) Γ(n)
The RHS is regular at ω = 2 and the LHS is convergent in four dimensions. Therefore
dimensional regularization has renormalised the divergent integral while leaving convergent
integrals unaffected when ω → 2. Naturally, this is due to analytic continuation.
4.5 Expansion of Z in Eucliean spacetime

We now go back to the Euclidean space definition of the generating functional
Z Z
1 ¯ ¯µ 1 2 2
ZE [J] = NE Dφ exp − d x̄ ∂µ φ∂ φ + m φ + V (φ) − Jφ .
4
(4.131)
2 2
This will be evaluated by expanding the action.

In order to do this, set φ0 to be a field configuration, then
Z
1 1
SEuc [φ, J] := d4 x̄ ∂¯µ φ∂¯µ φ + m2 φ2 + V (φ) − Jφ
2 2
Z
4 δSEuc
= SEuc [φ0 , J] + d x̄ (φ − φ0 )
δφ
δ 2 SEuc
Z
1
+ d4 x̄1 d4 x̄2 (φ(x̄1 ) − φ0 (x̄1 ))(φ(x̄2 ) − φ0 (x̄2 )) + · · · .
2 δφ(x̄1 )δφ(x̄2 )
(4.132)
It is understood that the functional derivatives are evaluated at φ0 . We know from AQFT
[3] that the classical limit can be recovered, by taking SEuc to be stationary at φ0 and this
implies that φ0 satisfies the classical EOM with the source term

δSEuc
= −∂¯µ ∂¯µ φ(x0 ) + m2 φ(x0 ) + V 0 [φ(x0 )] − J = 0. (4.133)
δφ x0
Integrating by parts gives

Z
1 d
SEuc [φ(x0 ), J] = 4
d x̄ 2 − φ(x0 ) [−Jφ(x0 ) + V 0 [φ(x0 )]], (4.134)
2 dφ(x0 )
whereas the second derivative is an operator
δ2
SEuc = δ(x̄1 − x̄2 ) −∂¯µ ∂¯µ + m2 + V 00 [φ(x1 )] .

(4.135)
δφ(x1 )δφ(x2 )
We need to make digression now on how to evaluate integrals of the sort

Z
I := dx exp(−α(x)). (4.136)
This can be evaluated by expanding the exponential around a point x0 where α is stationary
1
α(x) = α(x0 ) + (x − x0 )2 α00 (x0 ) + O(x3 ). (4.137)
2
The I integral is then approximated
Z
1
I = exp(−α(x0 )) dx exp − (x − x0 )2 α00 (x0 ) , (4.138)
2
and we can recognize this as a Gaussian integral (when the higher derivatives are ignored).
The degree of this approximation depends obviously on α. When α is smallest then the
integrand is largest and the points away from the minimum do not add a substantial con-
tribution.
Equipped with this method we can now apply it to the functional to obtain a crucially
important result [4]
δ 2 SEuc
Z Z
1
ZE [J] = NE exp {−SEuc [φ(x0 ), J]} Dφ exp − d4 x̄1 d4 x̄2 φ(x̄1 ) φ(x̄2 )
2 δφ(x̄1 )δφ(x̄2 )
exp {−S Euc [φ(x 0 ), J]}
= NE0 q , (4.139)
¯ ¯

µ 2 00
det (−∂µ ∂ + m + V [φ(x0 )])δ(x1 − x2 )
where we have ignored the higher order terms. The precise derivation is in the Appendix.
Let us do some re-writing to make this expression easier to handle, first the determinant,
which we will call M can be taken care of by using the identity
det M = exp(tr log M), (4.140)
so that

1 ¯ ¯
ZE [J] = NE0 µ 2 00
exp −SEuc [φ(x0 ), J] − tr log(−∂µ ∂ + m + V [φ(x0 )])δ(x1 − x2 ) ,
2
(4.141)
the delta term accounts for the quantum perturbations (or corrections) to Z[J], whereas
the first term accounts for the classical contribution. Also note that we set by convention
the determinant of an operator to be the product of its eigenvalues and that because φ0
satisfies δSEuc /δφ|x0 = 0 then it is a functional of J.
The concluding remark is to find the equivalence of these results in terms of the classical
field φcl and explore the corresponding effective action. When we work in Euclidean space,
the classical field was defined as
δZE δSE
φcl (x̄) = − ≈− + O(~), (4.142)
δJ(x̄) δJ(x̄)
and by using (3.57) and (3.58) we can have φcl as a function (functional, rather) of J,
however at the cost of doing it order by order in λ. The term O(~) stands for quantum
corrections. This relationship can be inverted and we can find J(x̄) as a function of φcl .
This inversion can be carried out to give (see Ramond)
λ
J(x̄) = (∂¯2 − m2 )φcl (x̄) − φ3cl (x̄). (4.143)
3!
An attractive (and indispensable) feature is that there are no higher terms in λ, comparison
with

δSEuc
0= = −∂¯µ ∂¯µ φ(x0 ) + m2 φ(x0 ) + V 0 [φ(x0 )] − J (4.144)
δφ x0
gives
φcl (x̄) = φ0 (x̄) + O(~). (4.145)
By integrating J(x) = −δΓ[φcl ]/δφcl we see that to this order that

Z
4 1 ¯2 2 λ 4
Γeff [φcl ] = − d x̄ φcl (∂ − m )φcl − φcl (x̄) , (4.146)
2 4!
which is indeed an effective action.
REFERENCES
The above summary on quantum field theory has its roots several sources. Overall it follows
the presentation given by
- [1] Bailin and Love Introduction to Gauge Field Theory section 4.4,
- [2] Peskin and Schroeder [275 to 294, 306 to 308 and 365 to 372].
At times it is complemented by the courses of QED and AQFT from the MSc QFFF
- [3] Rajantie’s AQFT
- [4] Ramond for the expansion of the generating functional integrals in Eucliean spacetime.
However, it is broad enough to be explained in most QFT books. The papers quoted are
- [5] R. Jackiw, Phys. Rev. D9 1686 (1974)
- [6] ’t Hooft and Veltman, Nucl. Phys. 44B 189
5 ZETA REGULARIZATION IN FIELD THEORY 73
5 Zeta Regularization in Field Theory

5.1 Heat kernels and Mellin transforms
At this point we can put together some of the concepts acquired in the second and third
chapters. Firstly we will further develop the theory of quantum corrections by evaluating
the determinant of the M matrix and then see how this is related to the use of the ζ function
in quantum theory as we explained in Chapter 3. In what follows, we will extrapolate the
results found in the quantum mechanical section to field theory; this section borrows some
its contents from Harfield [2], Hawking [3] and Ramond [5]. As we mentioned earlier, the
determinant in the expression (4.139)
NE0 exp {−SEuc [φ(x0 ), J]}

ZE [J] = q (5.1)
det (−∂¯µ ∂¯µ + m2 + V 00 [φ(x0 )])δ(x1 − x2 )

must be interpreted as the product of the eigenvalues of the operator. In order to discretize
these eigenvalues we truncate the space (by use of a box). We then multiply the resulting
eigenvalues and then let the size of the box increase to infinity.
First we state some preliminary results that will become useful later on. As it is shown in
a course on partial differential equations, the heat function
X
G(x̄, ȳ, t) := exp(−λn t)ψn (x̄)ψn∗ (ȳ) (5.2)
n
satisfies the heat equation

∂
Ax̄ G(x̄, ȳ, t) = − G(x̄, ȳ, t), (5.3)
∂t
where Ax̄ is taken to act on the first argument of G and initial condition
G(x̄, ȳ, t = 0) = δ(x̄ − ȳ). (5.4)
The key step, the relationship between the heat kernel G and the ζ function, is the following
result
Z∞ Z
ζA (s)Γ(s) = dt ts−1 d4 x̄G(x̄, x̄, t), (5.5)
0
which we proceed to explain. Note the similarity with (1.15), this should already give us a
hint on how to proceed. The eigenfunctions ψn (x) of A satisfy the following orthogonality
relations
Z X
hψn , ψm i = dxψn∗ (x)ψm (x) = δnm ,
∗
ψn∗ (x)ψn (x0 ) = δ(x − x0 ). (5.6)
n
We can explain (5.5) by taking the trace of G

Z X
Tr(G) = dxG(x, x, t) = e−λn t (5.7)
n
(remembering Tr(G) is a function of t) multiplying it by e−λt and integrating it with respect

to t and then with respect to λ to obtain
Z Z∞ XZ 1 X
dλ dtTr(G)e−λt = dλ = log(λn + λ). (5.8)
n
λn + λ n
0
The determinant of A is then written as

Z∞
 
Z
det A = exp  dλ dt Tr(G)e−λt λ=0 

(5.9)
0
ignoring the factor that would show from the λ integration. Swapping the integrals yields
X Z∞
−λn t
det A = e =− dtt−1 Tr(G). (5.10)
n 0
The relationship between the trace of the heat kernel Tr(G) and the ζ function is given by
a Mellin transformation, see Hawking [3]
Z∞
1
ζA (s) = dtts−1 Tr(G), (5.11)
Γ(s)
0
which justifies (5.5). A solution of the heat equation
∂
−∂¯x2 G0 (x̄, ȳ, t) = − G0 , (5.12)
∂t
with the boundary condition G0 (x̄, ȳ, t = 0) = δ(x̄ − ȳ) is

1 1 2
G0 (x̄, ȳ, t) = exp − (x̄ − ȳ) , (5.13)
16π 2 r2 4r
also a classic result from partial differential equations.

Let us generalize some of the techniques we used in Chapter 3. Consider an operator A
with positive real discrete eigenvalues λi where i runs from 1 to n and its eigenfunctions
are ψn (x) i.e. Aψi (x) = λi ψi (x). From here we set
X
ζA (s) = λ−s
n , (5.14)
n
which we call the zeta function associated to the operator A. The sum is over all the
eigenvalues and A is a real operator. If the operator A were the one-dimensional harmonic
oscillator Hamiltonian, then ζA would be the Riemann zeta function (excluding the singular
zero-point energy).
The first observation is that [1], [2], [3], [4] and [5]

d X Y
log λsn exp(−s log λsn )|s=0 = − log

ζA (s) =− λn , (5.15)
ds s=0 n n
which gives the determinant

Y
0
det A = λn = exp(−ζA (0)), (5.16)
n
as we know from our discussion of chapter 3. The zeta-function ζA is not always singular
at s = 0 for physically interesting operators and hence the convenience of writing this
representation for det A.
Algorithmically we have a procedure to compute the determinant of A and it can be done
as follows. First we need to find the solution to the heat equation subject to the delta-
function initial condition (5.4). Second, once this is done we can insert the solution into the
ζ representation above and we have ζA . Finally evaluate at s = 0 and compute exp(−ζA0 (0)).
5.2 Derivation of φ4 potential at φcl using ζ regularization

In our case, our operator is
λ
A = −∂¯2 + m2 + φ20 (x̄), (5.17)
2
where φ0 (x̄) is a solution of the classical equations with source J as we discussed in Chapter
4. Also note that in (5.1) we have a V 00 factor in the determinant which takes the form
V 00 = (λ/2)φ20 in the A operator.
As we have pointed out a solution of the heat equation with the boundary condition
G0 (x̄, ȳ, t = 0) = δ(x̄ − ȳ) is

1 1 2
G0 (x̄, ȳ, t) = exp − (x̄ − ȳ) . (5.18)
16π 2 r2 4r
However, this is only a part of the operator as we want to find G0 (x̄, ȳ, t) subject to the
initial condition (5.4) which obeys the whole A operator

¯2 2 λ 2 ∂
−∂ + m + φ0 (x̄) G(x̄, ȳ, t) = − G(x̄, ȳ, t). (5.19)
2 ∂t
For arbitrary fields φ0 we need to expand the effective action as (see (4.81) or Bailin and
Love, [5])
(0) (1)
ΓE [φcl ] = ΓE [φcl ] + ~ΓE [φcl ] + · · · , (5.20)
the ~ indicates quantum terms. Given the fact that the effective potential at 1 loop is of
the form [2]
Z
1 1
Veff [φcl ] d4 x = log det(−∂ 2 + m20 + λ20 φ2cl ) − log det(−∂ 2 + m20 ), (5.21)
2 2
as we also argued in general terms in Chapter 4 (4.56) and (4.85), allows us to write
(1) 1
ΓE [φcl ] = − ζ 0 ¯2 2 λ 2 (0), (5.22)
2 [−∂ +m + 2 φ0 (x̄)]
by use of (5.1) and (5.16) combined with (5.17). Note that replacing φ0 by φcl is not a
problem as there are no new quantum errors, that is errors up to O(~). By the expansion
of the effective potential we can swap m and m0 . From the fact that the effective action
can be written as
Z
ΓE [φcl ] = d4 x̄ V (φcl (x̄)) + F (φcl )∂¯µ φcl (x̄)∂¯µ φcl (x̄) + · · ·

(5.23)
we can compute the quantum O(~) contribution to V (φcl (x̄)) by considering a constant field
configuration. That is, suppose we take φcl (x̄) = v where v is a constant independent of x̄,
then
Z
ΓE [φcl ] = d4 x̄V (v), (5.24)
and ΓE is proportional to the infinite volume element d4 x̄, since the Euclidean space R4
R
is unbounded. This can be temporarily solved by taking the space to be the sphere S4 then
the volume is just that of the 5-dimensional sphere and hence finite. While the radius is
finite we need not worry about the infrared divergence. We then let the radius of the sphere
tend to infinity.
Taking V out of the integral we have [2], [5]
Z
1
V (v) d4 x̄ = − ζ 0 ¯2 2 λ 2 (0). (5.25)
2 [−∂ +m + 2 v ]
We can proceed to integrate (5.19) when v is constant this yields
µ4 µ2 (x̄ − ȳ)2

2 λ 2 t
G(x̄, ȳ, t) = exp − exp −m + v , (5.26)
16π 2 t2 4t 2 µ2
where the µ factor needs to be explained: it has dimensions of mass so that t is dimensionless.
Using the Mellin transform (5.5) we obtain
Z∞
µ4
Z
1 λ 2 t
ζA (s) = dt ts−1 d4 x̄ exp −m2 + v
Γ(s) 16π 2 t2 2 µ2
0
!2−s
µ4 m2 + λ2 v 2 Γ(s − 2)
Z
= d4 x̄, (5.27)
16π 2 t2 µ2 Γ(s)
the volume element d4 x̄ is present because it is in (5.25) and here t has been rescaled
R
since the integration over t is valid when s > 2 however the ζ function is defined everywhere
by analytic continuation as we know from Chapter 2. Differentiating and comparing these
two equations we have
 !2−s 
µ4 d  1 m2 + λ2 v 2 
V (v) = −
32π 2 ds  (s − 2)(s − 1) µ2 
s=0
2 !
2 λ 2
1 λ m + v 3
= m2 + v 2 log 2
− . (5.28)
64π 2 2 µ2 2
Note that we have used

Γ(s − 2) 1
= . (5.29)
Γ(s) (s − 2)(s − 1)
Also note that applying (5.16) and working with a non-constant field configuration φcl yields
2
λ0 φ2cl
Z
1 2 2 λ0 φcl 3
log det A = m0 + log m0 + − d4 x̄. (5.30)
32π 2 2 2 2
Equipped with this functional form of V the effective potential can be written as
2 !
m2 + λ2 v 2

1 2 2 λ 4 ~ 2 λ 2 3
V (φcl ) = m φcl (x̄) + φcl (x̄) + m + φcl (x̄) log − , (5.31)
2 4! 64π 2 2 µ2 2
ignoring terms of order ~2 . We make a pause now to examine this result. Superficially the
first striking observation is that there is a strong dependence on the unknown and arbitrary
scale µ2 . This would seem to imply that the potential is therefore arbitrary. However V
depends on the parameters m2 and λ which are undefined, except for the fact that they are
included in the classical Lagrangian.
Let us take the special massless case. This yields the following
d2 V

= 0. (5.32)
dφ2 φ=0
Now we define the mass squared as the coefficient of the terms φ2 in the Lagrangian eval-
uated at φ = 0. To first quantum corrections the coefficient is zero, if it is classically zero.
The λ term is defined to be the coefficient of the fourth derivative of V evaluated at some
constant point φ = M, i.e.
d4 V

λ := . (5.33)
dφ4 φ=M
We cannot take φ = 0 as with the mass squared factor because of the infrared divergence
coming from the logarithm. When we differentiate (5.31), set m2 = 0 and use (5.33) we see
that the above condition requires
λM 2 8
log =− . (5.34)
2µ2 3
In this case, we may use M 2 instead of 2µ2 /λ and write the result as (Bailin and Love and
[5])
λ2 φ4cl φ2cl

λ 4 25
V (φcl ) = φcl (x̄) + log 2 − , (5.35)
4! 256π 2 M 6
which is exactly the same result we found in Section (4.3). This was proved by Coleman
and Weinberg in 1973 [5]. The main result that can be extracted is that we need to be
careful with how we define the input parameters in the Lagrangian if we are to take into
account quantum corrections. Again, superficially it seems that (5.35) depends on another
arbitrary scale M 2 , but in fact it does not. Given the normalization condition, if we change
the scale from M 2 to M 02 we simultaneously have to change at the same time λ to λ0 by
use of (5.33)
3λ2 M0
λ0 = λ + log . (5.36)
16π 2 M
Therefore the potential (see Ramond)
λ0 4 λ02 φ4cl φ2cl

25
V (φcl ) = φcl (x̄) + log 02
− (5.37)
4! 256π 2 M 6
is indeed invariant under the representation V (λ0 , M 0 ) = V (λ, M ). This proves that the
physics behind remains unchanged but our way of interpreting the coefficients changes.
5.3 Coupling constants

Let us now look more closely to the scaling of determinants and the coupling constants.
The ζ function technique just used allows us derive scaling properties for determinants.
First, we will need the computation of ζ function ζ ¯2 λ 2 (0). This can be accomplished
[−∂ + 2 φcl ]
by taking the asymptotic expansion of G(x̄, ȳ, t) at µ2 = 1,
2 ∞
e−(x̄−ȳ) /(4t) X
G(x̄, ȳ, t) = e−εt an (x̄, ȳ)tn , (5.38)
16π 2 t2 n=0
with ε > 0 as a convergence factor. The boundary condition (5.4) sets the condition
a0 (x̄, x̄) = 1. (5.39)
Additionally, when we insert (5.38) into the PDE (5.19) we find recursion relations for the
an coefficients
∂
(x̄ − ȳ)µ a0 (x̄, ȳ) = 0 (5.40)
∂ x̄µ
and for n = 0, 1, 2, · · ·

∂ ¯2 λ 2
(n + 1) + (x̄ − ȳ)µ an+1 (x̄, ȳ) = ∂x − φcl + ε an (x̄, ȳ). (5.41)
∂ x̄µ 2
When we compute the first terms we have

λ
a1 (x̄, x̄) = − φ2cl + ε, (5.42)
2
λ2 4 λ ε ε2
a2 (x̄, x̄) = φcl (x̄) − ∂¯x2 φcl (x̄) − λφ2cl (x̄) + . (5.43)
8 4 2 2
Let us now use these results. We work under a scale change A → A0 = ead A, where d is
the natural dimension of A. By the definition of the ζ function we have
ζA0 (s) = e−ads ζA (s) (5.44)
which implies that

0
det(ead A) = eadζA (0) det A. (5.45)
Let us illustrate this with an example. Under the transformation
xµ → x0µ = ea xµ φcl → φ0cl = e−a φcl (5.46)

the massless classical action

Z
1 λ
SE [φcl ] = − d x̄ φcl ∂¯2 φcl − φ4cl
4
(5.47)
2 4!
is unchanged. However, the path integral corresponding to this action is not scale invariant
since in the steepest descent approximation the change in the effective action to quantum
order is as follows
0
eff eff eff
SE [φcl ] → SE [φcl ] = SE [φcl ] − ~aζ λ (0). (5.48)
[−∂¯2 + 2 φ2cl ]
Plugging (5.38) into (5.3) and with the assistance of the first two terms a1 and a2 and
integrating out the ∂¯2 with the divergence theorem yields
Z 2
1 4 λ
ζ(0) = d x̄ φ4 (x̄). (5.49)
16π 2 8 cl
A small digression is now required to further explain this. In 4 dimensions and in the
presence of mass the heat kernel is
2
e−(x̄−ȳ) /(4t) λ0 φ2cl

1 2 2
G(x̄, ȳ, t) = exp − |x̄ − ȳ| exp − m0 + t (5.50)
16π 2 t2 4t 2
and like in (5.27) and (5.28)

2−s
λ0 φ2cl

Γ(s − 2)
Z
1
ζ(s) = m20 + d4 x̄. (5.51)
16π 2 2 Γ(s)
Furthermore, note that in the more restricted case B = −∂ 2 + ω 2 we have kernel
(x − x0 )2

1
G(x, x0 , t) = √ exp − exp(−ω 2 t) (5.52)
4πt 4t
and the determinant of B is

Z∞ Z Z Z∞
1 −3/2 ω
log det B = − √ dtt exp(−ω t) d x̄ = − √ 2 4 4
d x̄ dte−t t−3/2
4π 4π
0 0
∞
Z
ω h i
= −√ d4 x̄ t−1/2 e−t − 2π 1/2 , (5.53)
4π 0
where integration by parts was used in the last line. Ignoring the divergent factor we have
the simplification
Z
4
det B = exp ω d x̄ . (5.54)
We can justify the disregard of the infinite in the integral for log det B by expressing it in
terms of Γ functions in this divergence zone by analytic continuation
Z
ω 1
log det B = − √ d4 x̄ Γ − (5.55)
4π 2
where the Γ function is defined in this region by analytical continuation.

Finally, the divergence of Γ(s − 2) at s = 0 cancels that the divergence of Γ(0), and this
process makes ζ(0) regular.
In terms of the effective action (5.48)
λ2
Z
eff 0 eff
SE [φcl ] = SE [φcl ] − ~a d4 x̄φ4cl (x̄). (5.56)
128π 2
Consequently the effect of the transformation to quantum order has been to change the
coupling constant λ by the following
λ λ0 λ λ2 0 3λ2
→ = − ~a ⇔ λ = λ − ~a. (5.57)
4! 4! 4! 128π 2 16π 2
What this is means is that the dimensionless coupling constant λ evolves as a result of
quantum effects. This evolution is in terms of scale dependence. At large scales the coupling
constant decreases indicating that the non-interaction theory is a good approximation for
the asymptotic states. On the other hand, if the scale decreases, the coupling increases.
Independently of how small λ was at the beginning this increment might throw away the
results obtained in the perturbation of λ. Moreover, this scaling law is like the one we found
earlier, and they are both correct to quantum order.
5.4 Partition functions in field theory

We recall that for any given time t a quantum mechanical system with one degree of
freedom, q and canonically conjugate momentum p, is described in terms of the spectrum
of its Hamiltonian H(p, q). From the path integral iformulation
we know that if the system
at an initial time ti is measured
to be in the state q , then the probability that the system
will be found in the state q f , at a final time tf is exactly
D E D E
qtff qtii = q f e−i(tf −ti )H q i , (5.58)

and it can be written in terms of path integrals as

 tf 
D E Z Z Z
q f e−i(tf −ti )H q i = Dq Dp exp i dt[pq̇ − H(p, q)], (5.59)

ti
the factor Dq denotes integration between the initial and final configurations q i and q f ; the
dot over q denotes the derivative of q with respect to time.
Quantum field theory and statistical mechanics share certain common elements and precisely
this analogy will allow us to apply path integrals to the description of dynamical systems
at finite temperature (see [1] and [4]). The first step, as we did in Chapter 3, is to compute
the partition function
Z = Tr[e−βH ], (5.60)
where the constant is
β = (kT )−1 , (5.61)

taking into account that the trace is taken to be the sum over all the possible configurations
the system is allowed to take. Note that time is singled out. Now, the probability for the
system to be in state of energy E is identified with
P = Z −1 e−βH . (5.62)
The value of any function for the dynamical variable f (q, p) is given by
hf i = Tr(f P ) = Z −1 Tr(f e−βH ). (5.63)
This does show a special similarity with zero temperature quantum mechanics and QFT
but the degree of this similarity is not fully illustrated. However, we can push the analogy
to calculate partition functions, specially this one. Let us start with a system which can be
regarded as a field theory in zero space dimensions.
We can compare this expression to the partition function for the same system at temperature
β −1
X
Z = Tr[e−βH ] = q|e−βH |q .

(5.64)
q
Let us draw comparisons between (5.59) and (5.64). If we were to set i(tf − ti ) = β, or
alternatively set ti = 0, then itf = β since the origin of time is arbitrary. Next, set q f = q i
which means that the initial and final configurations are the same, and since the difference
is a β factor, the only requirement is that the relevant configuration be periodic in the
functional integrals
q(β) = q(0). (5.65)
Thus the functional integration Dq is over the space of periodic functions as stressed in
Chapter 3. In this case, the sum over q in (5.59) is implicit. When we do the comparison
we can write
 β 
Z Z Z
dq
Z = Tr(e−βH ) = Dq Dp exp  dτ ip − H , (5.66)
dτ
0
bearing in mind again that Dq is over periodic functions.

If we take a well-behaved potential V (q) we could scale the temperature dependence purely
into the q integral. To do this, we make the following transformations
τ̄ = τ β −1 , p̄ = pβ 1/2 , q̄ = qβ −1/2 , (5.67)
then the exponent of the integrand becomes

Z1
p̄2

dq̄ 1/2
dτ̄ ip̄ − − βV (β q̄) . (5.68)
dτ̄ 2
0
Furthermore we can drop all the bars because the path integral measure is invariant under
the changes (5.67) thus we can write the partition function as
 1 
Z Z Z
1
Z = Dq Dp exp  dτ ipq̇ − p2 − βV (qβ 1/2 ) . (5.69)
2
0
Next, set p0 = p − iq̇, so that the measures are equal
Dp0 = Dp, (5.70)
which in turn, by completing the square in the exponent, allows us to write

   1 
Z Z1 Z Z
1 1 2
Z = Dp0 exp − dτ p02  Dq exp − dτ q̇ + βV (qβ 1/2 ) . (5.71)
2 2
0 0
As we have done repeatedly in AQFT we can ignore the p0 integral (even though it is
infinite) because it is independent of β. We call this integral N, and the reason why it can
be ignored is because N is usually always present in the numerator and denominators of
correlation functions.
The only example we could tackle is that of an integral that can be evaluated, i.e. a Gaussian
integral. The integral becomes Gaussian when we take the harmonic oscillator potential
V (q) = 21 ω 2 q 2 , the partition function is
 
Z Z1
1 2 1 2 2 2 
Z=N Dq exp − dτ q̇ + β ω q . (5.72)
2 2
0
It is one of the very few types that can actually be integrated. By virtue of (5.65) we have
Z1 2 Z1
d2

1 dq 1
dτ =− dτ q q, (5.73)
2 dτ 2 dτ 2
0 0
since the extra surface term is eliminated and therefore

 
Z Z1 2

1 d
Z = N Dq exp − dτ q − 2 + ω 2 β 2 q . (5.74)
2 dτ
0
If we proceed by analogy with the discrete case we have

 
Z1 2
N0
Z
1 d
Dq exp − 2 2
dτ q − 2 + ω β q  = √ , (5.75)
2 dτ DetA
0
where N 0 is a constant, and A is the operator [1], [4], [5]
d2
A=− + ω2 β 2 (5.76)
dτ 2
with positive definite eigenvalues (it must not contain zero eigenvalues as these would create
infinities which have to be removed). In order to prove (5.75) as we did in Chapter 3 we
need to express q(τ ) in terms of its Fourier components (QFT course) then transform into
the normal modes of A and integrate each one using
Z∞ r
1 2π
dqn exp − an qn2 = . (5.77)
2 an
0
The operator A operates on periodic functions with unit period, which can all be expanded
in terms of the complete Fourier set {e2πinτ }. The eigenvalues of A are
(4π 2 n2 + ω 2 β 2 ); n ∈ Z. (5.78)
Note the analogy with the eigenvalues of the quantum mechanical operator (3.84).
Hence multiplying we have the determinant of the operator A,
Y
det A = (4π 2 n2 + ω 2 β 2 ). (5.79)
n∈Z
Setting x2 = ω 2 β 2 yields the following

d X 1 1 X 1
log det A = = + 2 . (5.80)
dx2 4π 2 n2 + x2 x2 4π 2 n2 + x2
n∈Z n≥1
Substituting the formula that was shown in the Appendix (A.15b)

1 2x X 1
coth πx = + , (5.81)
πx π x2 + n2
n≥1
d 1 x
log det A = coth , (5.82)
dx2 2x 2
and from here we can integrate to find
Z
det A x x ωβ
log = dx coth = 2 log sinh = 2 log sinh . (5.83)
C 2 2 2
Removing the logs we have the formula for the determinant
ωβ
det A = C sinh2 . (5.84)
2
When we clean the expression we arrive at
1 D 1 1
F =− log Z = − + ω + log(1 − e−ωβ ), (5.85)
β β 2 β
where D is another constant, the zero-point energy is identified at 1 = e−ωβ . This formula
is called the thermodynamic potential.
Let us go back to our discussion of the ζ function. The heat equation associated with the
A operator (5.76) is (see [1], [2], [5])
X
G(t, t0 , σ) = exp 2πin(t − t0 ) − (ω 2 β 2 + 4π 2 n2 )σ ,

(5.86)
n∈Z
and recalling our Mellin transform (5.5) for the ζ function
Z∞ Z1
1 s−1
X
exp −(ω 2 β 2 + 4π 2 n2 )σ .

ζA (s) = dσσ dt (5.87)
Γ(s)
0 0 n∈Z
Scaling σ by ω 2 β 2 + 4π 2 n2 leads nowhere as we simply come back to the expression

X 1
ζA (s) = . (5.88)
(ω 2 β 2 + 4π 2 n2 )s
n∈Z
The technique we need to use is somewhat messier, it involves expanding in powers of ωβ

then integrating and re-arranging the sums. With this in mind we have
X 2 2 X 2 2
e−4π n σ = 1 + 2 e−4π n σ , (5.89)
n∈Z n≥1
and integrating we have

∞ ∞ Z ∞
2 X (ωβ)2k X 2 2
ζA (s) = (ωβ)−2s + (−1)−k dσσ s+k−1 e−4π n σ . (5.90)
Γ(s) k! n=1
k=0 0
σ by 4π 2 n2 and we can also identify the sum over

Now it is the time when we can rescale P
∞
n with the Riemann ζ function ζ(2s) = n=1 n−2s , which gives
∞
2 2 X (ωβ)2k (−1)k
ζA (s) = (ωβ)−2s + ζ(2s) + Γ(s + k)ζ(2s + 2k). (5.91)
(4π 2 )s Γ(s) k! (4π 2 )s+k
k=1
In order to differentiate ζA at s = 0 we note that the sum is well behaved at s = 0 and as

s → 0 a non zero term arises from the derivative of Γ−1 (s). Recalling our formulas from
Chapter 2 (2.30) and (2.33),
1 1
ζ(0) = − , ζ 0 (0) = − log 2π, (5.92)
2 2
and making use of the following expansions
2
(ωβ)−2s + ζ(2s) = 2(log 2 + log π − log 2π − log ωβ)s + O(s2 ) (5.93)
(4π 2 )s
as well as
2 (ωβ)2k (−1)k (−1)k 21−2k
Γ(s + k)ζ(2s + 2k) = (ωβ)2k Γ(k)ζ(2k)s + O(s2 ) (5.94)
Γ(s) k! (4π 2 )s+k k! π 2k
it follows (by differentiating with respect to s) that
∞
0
X (−1)k (ωβ)2k
ζA (s) = −2 log ωβ + 2 ζ(2k) + O(s). (5.95)
k (4π 2 )k
k=1
0
Consequently we arrive at a neat expression for ζA (0)
∞
0
X (ωβ)2k (−1)k ζ(2k)
ζA (0) = −2 log(ωβ) + 2 . (5.96)
k(4π 2 )k
k=1
Recalling the formula for the even values of the Riemann ζ function in terms of Bernoulli
numbers (2.31)
(−1)k+1 (2π)2k
ζ(2k) = B2k , (5.97)
2(2k)!
we can continue simplifying (5.96)

∞
0
X (ωβ)2k 1
ζA (0) = −2 log(ωβ) − B2k . (5.98)
(2k)! k
k=1
Using another formula from the Appendix (A.15a)

∞
X (2x)2k−1
1
coth x = +2 B2k , (5.99)
x (2k)!
k=1
and integrating yields

∞
1 X (2x)2k
Z
dx coth x = log x + B2k , (5.100)
2 k(2k)!
k=1
and finally by comparing with (5.99) and setting x = 12 ωβ, we obtain

0 ωβ ωβ
ζA (0) = −2 log(ωβ) + 2 log − 2 log sinh = −ωβ − 2 log(1 − e−ωβ ). (5.101)
2 2
Solving for Z yields
1 0 1
log Z = ζ (0) = − ωβ − log(1 − e−ωβ ), (5.102)
2 A 2
which is the same result we found earlier by evaluating the determinant with the eigenvalue
method. Although this technique is lengthy it will enable us to show the connection between
the ζ function and interacting quantum field theories in the high temperature limit as β → 0.
5.5 High temperature limit

We consider a scalar field φ(x) = φ(t, xi ), and canonical conjugate Π(t, xi ), interacting with
itself. The Hamiltonian in this case is
Z Z
3 1 2 1
H= d x Π + (∇φ) + V (φ) = d3 xH .
2
(5.103)
2 2
Generalizing from the quantum mechanical case of Chapter 3, we immediately have
 β 
Z Z Z Z 
∂φ
Z = Dπ Dφ exp dt d3 x iΠ −H , (5.104)
 ∂t 
0
again taking the φ integral over fields periodic in time φ(t, xi ) = φ(t + β, xi ) and keeping
the space variables unbounded. Rescaling the temperature φ → φβ 1/2 and introducing the
change
∂φ
Π0 = Π − i (5.105)
∂t
allows us to perform the Π integral
 1 
Z Z Z Z 
∂φ 1 1
Z = Dπ Dφ exp dt d3 x iΠ − Π2 − β 2 (∇φ)2 − βV (φβ 1/2 )
 ∂t 2 2 
0
 1 ! 
Z  Z Z 2
1 ∂φ 1 2 
= N Dφ exp − dt d x 3 2
+ β (∇φ) + βV (φβ )1/2
. (5.106)
 2 ∂t 2 
0
Note that we have a Euclidean integral. It is essential to note a feature of quantum theory:
the temperature dependence also appears in (∇φ)2 . We are, of course, working in φ4 theory:
1 2 2 λ 4
V (φ) = m φ + φ .
2 4!
The technique for approximating integrals of this sort at second order developped in Chapter
4 Section 4.5 readily gives
Z Z ! Z Z 2
!
δS 1 δ S
S[φ, J] = S[φ0 , J] + dt d3 x (φ − φ0 ) dt d3 x (φ − φ0 )2

+ + ··· ,
δφ φ0 2 δφ2 φ0
(5.107)
where φ0 satisfies the classical equation of motion
∂ 2 φ0

δS λ
= − 2 − β 2 ∇2 φ0 + m2 β 2 φ0 + β 3 φ30 + Jβ 3/2 = 0. (5.108)
δφ φ0
∂t 3!
By taking the new operator B to be
δ2 S ∂2 λ
B= = − − β 2 ∇2 + m2 β 2 + β 2 φ20 (5.109)
δφ20 ∂t2 2
and shifting the integration variable from φ to φ − φ0 we obtain

 
Z Z1 Z 2

1 ∂ λ
Z = N e−S[φ0 ,J] Dφ exp − dt d3 xφ − 2 − β 2 ∇2 + m2 β 2 + β 3 φ20 φ
2 ∂t 2
0
N 0 e−S[φ0 ,J]

1 0
= √ = N 0 e−S[φ0 ,J] (det B)−1/2 = N 0 e−S[φ0 ,J] exp ζ (0) , (5.110)
det B 2 B
with N 0 unknown. As before, with the non-interacting massless case, we will only consider
constant φ0 which will give information about the part of the one loop correction which does
not depend on derivatives of φ0 (Ramond [5]). Scaling backwards the classical equation of
motion gives φ0 (t, xi ) = β 1/2 φ̄0 (βt, xi ) where φ̄0 is indepedent of β. In order to do this, we
need to split up the operator B in two parts
∂2
B=− + β2C 2 (5.111)
∂t2
with C being β independent
λ 2
C = −∇2 + m2 + φ̄ . (5.112)
2 0
The first step is to go back to the C -operated heat equation
∂
Cx GC (xi , y i , σ) = − GC (xi , y i , σ) (5.113)
∂σ
where Cx indicates that we are operating on x. The boundary condition is as usual
GC (xi , y i , σ = 0) = δ(xi − y i ). (5.114)

The solution we had for four dimensions can easily be generalized to d dimensions yielding
µd n µ
o λ 2 σ
GC (xi , y i , σ) = exp − (x i
− y i 2
) exp − m 2
+ φ̄ , (5.115)
(4πσ)d/2 4σ 2 0 µ2
where µ has the purpose of making σ dimensionless, i.e. it has mass dimension. As we
∂2
argued for the quantum mechanical case, the eigenvalues for − ∂t 2 over periodic functions
2 2
are 4π n hence the heat kernel is essentially the same as (5.86)
µd µ2 i

2 2 σ
GB (t, xi , t0 , y i , σ) = exp − (x − y i 2
) − M β
(4πσ)d/2 4σ µ2
∞
4π 2 n2
X
× exp − 2 2 σ + 2πin(t − t0 ) (5.116)
n=−∞
µ β
where
λ 2
M 2 = m2 + φ̄ . (5.117)
2 0
The trace of the kernel then becomes
Z Z1 Z
µ 0
Tr[GB ] = d xGB (t, x, t , x, σ) = dt dd xGB (0, x, 0, x, σ)
0
Z1 ∞
µd
Z
2 2 2 2
n2 σ/(µ2 β 2 )
X
= d/2
e−M β σ/µ dt dd x e−4π . (5.118)
(4πσ) n=−∞
0
Finally, putting this in the Mellin transform

Z∞
1
ζB (s) = dσσ s−1 Tr[GB (σ)]
Γ(s)
0
Z∞ Z1 ∞
µd
Z
1 2 2 2 X 2
n2 /(µ2 β 2 ))σ
= d/2
e−M β σ/µ dσσ s−1−d/2
dt dd x e−(4π .
Γ(s) (4π) n=−∞
0 0
(5.119)
We note the Rfollowing two points: (1) d = 0 is the quantum mechanical result and (2) the
volume V = dd x can be regularized by constraining the system to be a finite box. Scaling
by µ2 β 2 σ → σ simplifies the exponential in the sum
Z∞ ∞
V (µβ)2s X 2
n2 +M 2 β 2 )σ
ζB (s) = dσσ s−1−d/2 e−(4π . (5.120)
Γ(s) (4πβ 2 )d/2 n=−∞
0
As we did in (5.89) we split up the infinite sum as

∞ ∞
2
n2 σ 2
n2 σ
X X
e−4π =1+2 e−4π , (5.121)
n=−∞ n=1
and then we have for ζB (s)

∞
Z∞

∞
V (µβ)2s 1 
Z
2 2 2 2 X 2 2
ζB (s) = dσσ s−1−d/2 e−M β σ + 2 dσσ s−1−d/2 e−M β σ e−4π n 
(4πβ 2 )d/2 Γ(s) n=1
0 0
∞
 
∞
V (µβ)2s 1 
Z
d X 2 2 2 2
= Γ s − + 2 dσσ s−1−d/2
e−(4π n +M β )σ  . (5.122)
(4πβ 2 )d/2 Γ(s) 2 n=1
0
Specialising to d = 3 we set the task of finding the limit of high temperature β → 0
V M 3 µ 2s Γ(s − 32 )
ζB (s) =
8π 3/2 M Γ(s)
3/2 ∞
π (µβ)2s X (βM )2k (−1)k 3
+ 2V Γ(s − 2 + k)ζ(2s + 2k − 3). (5.123)
β2 Γ(s) k! (2π)2k+2s
k=0
Ramond [5] leaves the evaluation of these terms as an exercise and I welcome the opportunity
to provide my solution. We shall make use of
Γ(s)−1 = s + O(s2 ), (5.124)
√

1
Γ = π, (5.125)
2
u u u u
Γ s− =Γ − +Γ − ψ (0) − s + O(s2 ), (5.126)
2 2 2 2
and
V M 3 µ 2s V M3
3
M V µ
= + log s + O(s2 ), (5.127)
8π 3/2 M 8π 3/2 4π 3/2 M
V M 3 µ 2s Γ s − 32

V M3
= s + O(s2 ), (5.128)
8π 3/2 M Γ(s) 6π
d V M 3 µ 2s Γ s − 32

V M3
3/2
= + O(s). (5.129)
ds 8π M Γ(s) 6π
The first term of the sum is very similar, except that it needs
ζ(2s − 3) = ζ(−3) + 2ζ 0 (−3)s + O(s2 ), (5.130)
(αβ)±s = 1 ± s log(αβ) + O(s2 ), (5.131)
1
ζ(−3) = , (5.132)
120
yielding
π2 V π2 V

2 (0) 3
s+ 2 log µβ + 240ζ(−3) + π ψ − − 2 log π − 2 log 2 + γ s2 + O(s3 ).
45β 3 45β 3 2
(5.133)
Differentiating with respect to s gives
π2 V
. (5.134)
45β 3
1
The third term (the second term in the sum, that is) requires ζ(−1) = − 12 and the same
technique gives
V M2
− . (5.135)
12β
The term at k = 2
3/2
V M 3 µ 2s Γ(s − 32 ) π (µβ)2s (βM )2·2 (−1)2
ζB (s) = + 2V 2
8π 3/2 M Γ(s) β Γ(s) k! (2π)2·2+2s
× Γ(s − 23 + 2)ζ(2s + 2 · 2 − 3) (5.136)
is more delicate as at s = 0 we have a singularity coming from ζ(1). This can be evaluated
as follows by expanding each factor separately
(µβ)2s (βM )4 1 M 4V β M 4V β
2V π 3/2 β −3 Γ(s + 1
2 ) = + s
Γ(s) 2 (2π)4+2s 16π 2 16π 2

(0) 1
× γ − 2 log 2 − 2 log π + log µβ + ψ 0, s2
2
+ O(s3 ). (5.137)
We make use of the Laurent expansion (2.43) of ζ(s) instead of the Taylor expansion to
account for the simple pole at s = 0
1
ζ(2s + 1) = + γ − 2γ1 s + O(s2 ), (5.138)
2s
where γ1 is a Stieltjes’ constant. Then multiplying these two expansions together
(µβ)2s (βM )4 M 4V β M 4V β

1 µβ
2V π 3/2 β −3 Γ(s + 1
2 )ζ(2s + 1) = + γ + log s
Γ(s) 2 (2π)4+2s 32π 2 32π 2 4π
+ O(s2 ). (5.139)
Differentiating with respect to s we have
2s
(βM )4 M 4V β

d 3/2 −3 (µβ) 1 1 µβ
2V π β Γ(s + 2 )ζ(2s + 1) = γ + log + O(s).
ds Γ(s) 2 (2π)4+2s 16π 2 4π
(5.140)
Therefore we have
3
π2 M2 βM 4

1 0 M µβ
log Z = ζB (0) = V + − + γ + log + ··· (5.141)
2 12π 90β 3 24β 32π 2 4π
which amounts to a free energy per unit volume [5]
M3 π3 M2 M4

F µβ
=− − + − γ + log + ··· (5.142)
V 12πβ 90β 4 24β 2 32π 2 4π
in the limit of high temperature β → 0.
5.6 Equivalence of ζ and dimensional regularization in φ4 theory

Let us now first summarize the features of the ζ regularization technique. We shall start
with situations where we do not know explicitly the eigenvalues of a non-negative self-adjoint
operator A
Aψn (x) = λn ψn (x). (5.143)
In these cases, we consider the heat equation

∂
GA (x, y, t) + AGA (x, y, t) = 0 (5.144)
∂t
where A is taken to act on the first argument of GA . The initial condition is
GA (x, y, 0) = δ(x − y). (5.145)
The expression GA is the heat kernel and it accounts for the diffusion over a region of
spacetime of a unit of heat placed at y at t = 0, i.e.
X −tλn
GA (x, y, t) = x|e−tA |y = ψn (x)ψn∗ (y).

e (5.146)
n
When we set x = y and we integrate over spacetime we obtain

Z X
d4 xGA (x, x, t) = e−λn t = Tr[GA (t)]. (5.147)
n
The ζ function of A is connected to TrGA (t) by the Mellin transform

Z∞
X 1
ζA (s) = λ−s
n = dtts−1 Tr[GA (t)], (5.148)
n
Γ(s)
0
with the understanding that zero modes are not taken in the sum.
If we make the transformation A → A0 = α−1 A. The eigenvalues of A become λn → α−1 λn
and also the scale µ becomes µ0 . The new ζA is
0 2s
2s µ
ζA0 /(µ0 )2 (s) = α ζA/µ2 (s). (5.149)
µ
Accordingly we obtain

0 0 2
d
log det A /(µ ) =− ζA /(µ ) (s)
0 0 2
ds s=0
2
µ0

= log det(A/µ2 ) − log α2 ζA/µ2 (0) − log ζA/µ2 (0),
µ
the presence of ζA/µ2 (0) indicates that det A is modified by the transformation.
Going back to the heat kernel [1], [2] and [5] Tr[GA (t)] = n e−tλn and multplying it by
P
2
e−tm and then integrating over t gives
Z∞
2 X
dte−tm Tr[GA (t)] = (λn + m2 )−1 . (5.150)
0 n
Performing another integration, this time with respect to m2 , and swapping the integrals
yields
Z∞ Z∞ Z∞
−tm2
X m=∞
dm 2
dte Tr[GA (t)] = dtt−1 Tr[GA (t)] = log(λn + m2 )m=0 . (5.151)
0 0 0 n
Ignoring the upper limit we have arrived at
X Z∞
log det A = log λn = − dtt−1 Tr[GA (t)]. (5.152)
n 0
We need aslo, a formality, the introduction of a cutoff ε for evaluating log det A, i.e.
Z∞
− lim dtt−1 Tr[GA (t)], (5.153)
ε→0+
ε
this is called the proper-time cutoff. This procedure necessitates the evaluation of the above
integral by asympotitic expansion of Tr[GA (t)].
However, if the determinant needs ζ regularization then
Z∞
 
0 d 1
log det A = −ζA (0) = −  dtts−1 Tr[GA (t)] . (5.154)
ds Γ(s)
0 s=0
An expansion around s = 0 of 1/Γ(s) using Weierstrass product

∞
1 Y s −s/n
= seγs 1+ e = s(1 + γs) + O(s2 ), (5.155)
Γ(s) n=1
n
and the polygamma function ψ (0) (s)
Γ0 (s) d
= log Γ(s) = ψ (0) (s), (5.156)
Γ(s) ds
ψ (0) (s)

d 1 1 1 (2) 2 3
= = s(1 + γs) − − γ + ζ(2)s + ψ (1)s + O(s )
ds Γ(s) Γ(s) s 2
= −1 − 2γs + O(s2 ), (5.157)
by the use of (2.41)

∞
(0)
X (−1)n n
ψ (s + 1) = log Γ(s + 1) = −γs + s ζ(n), (5.158)
n=2
n
gives
Z∞ Z∞
 
log det A = lim (1 + 2γs) dtts−1 Tr[GA (t)] + s(1 + γs) dtts−1 log tTr[GA (t)] .
s→0
0 0
(5.159)
Note how the above expression for log det A without ζ function would follow if we ignored
the divergent intrgral for s = 0.
The comparision begins by now trying to obtain the same answer for log det A from dimen-
sional regularization. To this end, we generalize the heat equation to 2ω + 1 dimensions
noting that pole is now shifted to 2ω = 4. One way to do this is to take the product of the
initial 4 dimensional spacetime with 2ω − 4 flat dimensions. In this case, the heat kernel
Tr[GA (t)] is changed by a factor of (4πt)2−ω hence
Z∞
1
log det A = − dtt1−ω Tr[GA (t)] = −(4π)2−ω Γ(2 − ω)ζA (ω − 2). (5.160)
(4π)ω−2
0
A final expansion around ω = 2 by the use of

1
−(4π)2−ω Γ(2 − ω) = + γ − log 4π + O(ω − 2), (5.161)
ω−2
and
0
ζA (ω − 2) = ζA (0) − ζA (0)(ω − 2) + O((ω − 2)2 ), (5.162)
yields
ζA (0) 0
log det A = + (γ − log 4π)ζA (0) − ζA (0) + O(ω − 2). (5.163)
ω−2
This encapsulates a fundamental result which has been at the core of the ζ and dimensional
regularizations [1], namely there is a pole at 2ω = 4 with residue ζA (0) and finite part
0
−ζA (0) + (γ − log 4π)ζA (0). Consequently, there is an equivalence (agreement is a better
word) between the values of log Z derived by ζ and dimensional regularization. This equiv-
alence is up to a multiple of ζA (0) which can be absorbed in the normalization constant.
For the sake of completeness we finish the summary of the ζ technique by following the tech-
nique described in [1] but omitting their use of the spacetime metric. When we evaluate
path integrals in curved spacetimes we compute exressions of the form
Z
Z[φ] = Dφ exp(iS[φ]), (5.164)
where Dφ is a measure on the space of matter field and S[φ] is the classical action. Certain
boundary (or periodicity) conditions are satisfied by φ. For example, for temperature
T = 1/β the boson fields is periodic in imaginary time on some boundary at large distance,
with period β. Then Z is the partition function from statistical mechanics. The leading
contriubtion to the path integral will come from field configurations near the background
φ0 which satisfies the classical equations in addition to the boundary conditions. Setting
φ = φ0 + φ̃ the action can be expanded about the background fields
S[φ] = S[φ0 ] + S22 [φ̃] + · · · (5.165)
where S22 [φ̃] is quadratic in the fluctuations of φ. Therefore we have

Z
log Z = iS[φ0 ] + log D φ̃ exp(iS22 [φ̃]) + · · · . (5.166)
The quadratic term is also of the form

Z
1
S22 [φ̃] = − d4 xφ̃A2 φ̃ (5.167)
2
with A2 a second order differential operator constructed from the background field. (Note
that in fermionic fields the operator would be of first order). The condition on the metric
for A2 to be real (elliptic) and self-adjoint is that the background metric be Eucliean.
These attributes of A2 will guarantee the existence of a complete set of eigenfuction ψn and
spectrum λn such that
A2 φ n = λ n φ n , (5.168)
with orthogonality
Z
d4 xφn φm = δnm . (5.169)
The field fluctuation φ̃ can be expressed as

X
φ̃ = θ n φn (5.170)
n
where the measure on the field can be written in terms of these θn coefficients
Y
Dφ = µdθn . (5.171)
n
Here µ is a normalization constant with mass dimension. Putting all of this together yields
Z YZ
1
Z[φ̃] := Dφ exp(iS22 [φ̃]) = µdθn exp − λn θn2
n
2
1/2 −1/2
Y 2π 1
= µ = det A 2 . (5.172)
n
λn 2πµ2
This means that the quadratic contribution in the field fluctuations is computed by eval-
uating a determinant. The issue is that the convergence of the product is not obvious,
let alone guaranteed, therefore making this expression a sensible one is a difficult problem.
Moreover, the free energy is proportional to the log of Z which by the use of
0
det A2 = exp(ζA 2
(0)) (5.173)
becomes

1 0 1 1
log Z[φ̃] = ζ (0) + log ζA2 (0). (5.174)
2 A2 2 2πµ2
Finally, let us summarize how the effective Lagrangian is affected by ζ regularization through
the effective action for scalar fields φ4 . Following [1], let us consider a 2-dimensional space-
time area S with a constant electromagnetic field with field strength

0 B
(Fµν ) = (5.175)
−B 0
and potential
(Ax , Ay ) = (0, Bx). (5.176)
The quadratic differential operator we are interested in is
−∆2 = −(∂µ − iAµ )(∂ µ − iAµ ) = −∂x2 − (∂y − iBx)2 , (5.177)
−∆2 commutes with the second component of the momentum p̂y = −i∂y , hence taking the
eigenvalues py of p̂y we have
py 2
−∆2 = −∂x2 + B 2 x − . (5.178)
B
We can already see the form of the Hamiltonian of the harmonic oscillator shapping up. In
fact, the change of variables x → x0 = x − py /B gives twice this Hamiltonian with frequency
|B|. Consequently, in this case, the eigenvalues are

1
λpy ,n = 2|B| n + . (5.179)
2
Note that the independence of py indicates that all the levels are degenerate. Let us now
produce the heat kernel
X 2 SB −tm2 X −tB(2n+1) SB −tm2
Tr [G−∆2 +m2 (t)] = e−t(λ+m )
= e e = e csch(tB)
4π 4π
λ∈(−∆2 ) n≥0
(5.180)
because the degeneracy is SB

2π and we take B to be positive. Note that when B → 0 we
have csch(tB) ∼ (tB)−1 so that the free heat kernel becomes
SB −tm2 1 −tm2
Tr [G−∆2 +m2 (t)] ∼ e (tB)−1 = e S. (5.181)
B→0 4π 4πt
The associated ζ−∆2 +m2 then becomes
∞ −s
1 m2

SB X 1 2 SB −s
ζ−∆2 +m2 (s) = 2B n + +m = (2B) ζ s, + , (5.182)
2π n=0 2 2π 2 2B
where we have a Hurwitz ζ function. The 1-loop effective Lagrangian is
(1) 1 (1) 1 1 0
Leff = Seff = log Det(−∆2 + m2 ) = − ζ−∆ 2 +m2 (0). (5.183)
S 2S 2S
A similar argument as that used in the proof of Theorem 9 of Chapter 2 shows that
1
ζ (0, v) = − v, (5.184)
2
and

d 1
= ζ 0 (0, v) = log Γ(v) − log 2π,

ζ(s, v) (5.185)
ds s=0 2
in fact, note how at v = 1 we have the formula of Theorem 9. Hence
m2 + B m2 + B

(1)
Leff = π −1 2−2 B log(2B)ζ 0, − ζ 0 0,
2B 2B
2
2
1 m +B m +B 1
= π −1 2−2 B log(2B) − − log Γ + log 2π
2 2B 2B 2
2
B m +B
= log 2π − 2 log Γ − m2 log 2B . (5.186)
8π 2B
The physically interesting limits occur as m → 0 and B → 0 for which we have
(1) B
lim Leff = log 2. (5.187)
m→0 8π
The limit as B → 0 necessitates the Stirling formula (A.87) for log Γ(s + 1) which is proved
in the Appendix

1 1 1 1 1
log Γ(s + 1) = log 2π + s + log s − s + − 3
+ − ··· (5.188)
2 2 12s 360s 1260ss
so that
(1) 1
lim Leff = − (m2 log m2 − m2 ). (5.189)
B→0 16π
This concludes our summary of the ζ technique.
REFERENCES
The discussion of the role of the ζ function in quantum field theory follows from
- [1] EOR’s Zeta Regularization Techniques with Applications (first three chapters)
- [2] Hartfield’s Quantum Field Theory of Point Particles and Strings [611 to 616]
- [3] Hawking’s paper was a quoted in the context of the determinant of a general operator
A.
- [4] Kleinert’s Path Integrals
- [5] Ramond’s Field Theory [81 to 93]
Additionally, the exposition is the progression from the ideas presented in Chapter 3 (ζ
functions in quantum mechanics) combined with Chapter 4 (path integrals in quantum
fields). Finally, the second quoted paper is
- [6] S. Coleman and E. Weinberg Phys. Rev. D7 1888 (1973).
6 CASIMIR EFFECT 96
6 Casimir Effect
6.1 Experimental setup
Hendrik Casimir and Dirk Polder discovered the existence of the Casimir effect in 1948.
The Casimir effect is a force arising from a quantized field. For instance two uncharged
metallic plates in a vacuum, placed a few micrometers apart, without any external elec-
tromagnetic field, affect the virtual photons which constitute the field, and generate a net
force [1]: either an attraction or a repulsion depending on the specific arrangement of the
two plates.
Figure 6.1: Two uncharged metallic plates in a vacuum, placed a few micrometers apart
The strength of the force falls off rapidly with distance thus it is only measurable when
the distance between the objects is extremely small. We shall describe and compute the
Casimir effect in terms of the zero-point energy of a quantized field in the intervening space
between the objects instead of expressing it in terms of virtual particles interacting with
the objects.
The Casimir effect can be understood by the idea that the presence of conducting metals
alters the vacuum expectation value of the energy of the second quantized electromagnetic
field [2].
6.2 ζ regulator
Using E = ω/2 we can determine the vacuum expectation value of the energy of the
electromagnetic field in the cavity to be
1X
hEi = En (6.1)
2 n
with the sum running over all possible values of n accounting for the standing waves. Note
that this sum is divergent. Each energy level En depends on the shape consequently En (s)
is the energy level, and hE(s)i is the vacuum expectation value.
6 CASIMIR EFFECT 97
Figure 6.2: Virtual particles interacting with the plates
The force at point P on the wall of the cavity is equal to the change in the vacuum energy
if the shape Ω of the wall is perturbed infinitesimally, say by δΩ, at point P, i.e.

δ hE(Ω)i
F (P) = − . (6.2)
δΩ P
Casimir considered the space between a pair of conducting metal plates at distance r apart.
In this case, the standing waves can be calculated, since the transverse component of the
electric field and the normal component of the magnetic field must vanish on the surface of
a conductor.
Ignoring the polarization and the magnetic components and assuming the parallel plates lie
in the x -y plane, the standing waves are
ψn (xi , t) = e−iωn t ei(kx x+ky y) sin(kn z) (6.3)
where ψ stands for the electric component of the electromagnetic field. Here kx and ky are
the wave vectors in directions parallel to the plates, and
nπ
kn = (6.4)
a
is the wave-vector perpendicular to the plates. Here n is an integer resulting from the
requirement that ψ vanish on the metal plates. The energy of this wave is
r
n2 π 2
ωn = kx2 + ky2 + 2 . (6.5)
r
The vacuum energy is then the sum over all possible excitation modes
Z ∞
1 dkx dky X
hEi = 2 Aωn (6.6)
2 (2π)2 n=1
where A is the area of the metal plates, and a factor of 2 is introduced for the two possible
polarizations of the wave. This expression is divergent, so we introduce a ζ regulator to
6 CASIMIR EFFECT 98
make the expression finite, and in the end we will remove this regulator. The ζ regulated
version of the energy per unit-area of the plate is
∞
hE(s)i
Z
dkx dky X −s
= ωn |ωn | . (6.7)
A (2π)2 n=1
This integral is finite for s > 3. The sum has a pole at s = 3, however it may be analytically
continued to s = 0, where the expression is finite. We have
∞ 1/2 −s/2
hE(s)i n2 π 2 n2 π 2
Z
dkx dky X 2 2
2 2
= k + k + k + k +
A (2π)2 n=1 x y
r2 x y
r2
∞ (1−s)/2
n2 π 2
Z
1 X 2 2
= dkx dky kx + ky + r2
. (6.8)
4π 2 n=1
We now introduce polar coordinates κ2 = kx2 + ky2 and dkx dky = κdκdθ
Z∞ Z2π Z∞
2 n2 π 2 (1−s)/2 2 n2 π 2 (1−s)/2

hE(s)i 1 X 1 X
= dκdθκ κ + 2
= 2πdκκ κ + 2

A 4π 2 n r 4π 2 n r
0 0 0
!(3−s)/2
2
1 X |n| π 3−s 1 π 2−s X 3−s
= = − 3−s |n| . (6.9)
2π n r2 s−3 2r 3−s n
1
At s = 0 we have the Riemann ζ function and ζ(−3) = 120 from Chapter 1
hEi hE(s)i 1 π2 π2 1 π2
= lim = − 3 ζ(−3) = − 3 =− . (6.10)
A s→0 A 2r 3 6r 120 720r3
Note that in terms of the Planck constant and the speed of light the above translates to [3]
hEi ~cπ 2
=− . (6.11)
A 720r3
The Casimir force per unit area FCas /A for idealized, perfectly conducting plates with
vacuum between them is
FCas d hEi ~cπ 2
=− =− . (6.12)
A dr A 240r4
The minus sign indicates that the force is attractive, also it decreases faster than gravity
due to the r4 factor in the denominator.
6.3 Experimental evidence

The original form of the experiment, described above, successfully demonstrated the force
to within 15% of the value predicted by the theory [4].
REFERENCES
.1. The Force of Empty Space http://focus.aps.org/story/v2/st28
.2. E.L. Losada ’Functional Approach to the Fermionic Casimir Effect’
.3. The Casimir effect http://socrates.berkeley.edu/b̃udker/Physics250 Spring0607/Corsini Casimir.ppt.
.4. A Force from Empty Space: The Casimir Effect http://apod.nasa.gov/apod/ap061217.html
7 CONCLUSION 99
7 Conclusion
We have seen that the task of computing the generating functional Z[J]
Z Z
Z[J] = exp(−iE[J]) = Dφ exp i d4 x(L + Jφ) = Ω|e−iHT |Ω ,

(7.1)
of a scalar field theory with source J is reduced to

δ 2 L1
Z
4 1
−iE[J] = i d x(L1 [φcl ] + J1 φcl ) − log det − + (connected diagrams)
2 δφ(x)δφ(y)
Z
+ i d4 x(δL [φcl ] + δJφcl ), (7.2)
where the classical field is taken to be

δ
φcl (x) = hΩ|φ(x)|ΩiJ = − E[J]. (7.3)
δJ(x)
The lowest order quantum corrections to to the effective potential Γ[φcl ] is given by the
functional determinant because the Feynman diagrams contributing to it have no external
lines and the simplest ones turn out to have two loops hence [3]
Z
(0) (1)
Γ[φcl ] := −E[J] − d4 x0 J(x0 )φcl (x0 ) = ΓE [φcl ] + ~ΓE [φcl ] + · · ·
δ 2 L1

i
= log det − . (7.4)
2 δφ(x)δφ(y)
For the φ4 scalar theory, the operator in the integral above is the differential quadratic
operator [1], [2] and [3]
δ 2 L1
A=− = −∂¯µ ∂¯µ + m2 + V 00 [φcl ] (7.5)
δφ(x)δφ(y)
where the bar indicates that we have passed to Eucliean spacetime coordinates. This is an
essential step since it guarantees that A will be a real and self-adjoint operator.
The determinant of this operator can be achieved through dimensional regularization by
evaluating integrals of the type [1]
d2ω k 2
Z
I(ω, µB ) = (k − µ2B + iε)−1
(2π)2ω
iµ2B µ2B

2 ω−2 1 0
= (M ) + Γ (1) + 1 − log + O(ω − 2) (7.6)
16π 2 2−ω 4πM 2
which are computed by extensive use of the Γ function and in paricular using clever tech-
niques of analytic continuation. This is a generalization of the integral
d2ω `
Z
2 2 1 3
lim = −π m + − γ + O(2 − ω) . (7.7)
ω→2 `2 + m2 2−ω 2
However, the underlying mathematical restructure of these computations required the use
of Γ0 (1) = −γ which is a signal that the Riemann ζ function is behind the scenes.
By the ζ technique we can show that

−S[φ0 ,J] 1 0
Z = Ne exp ζ (0) (7.8)
2 A
7 CONCLUSION 100
on in terms of the effective action

(1) 1 0
ΓE [φcl ] = − ζA (0). (7.9)
2
Furthermore the equivalence between dimensional and ζ regularization is manifest in [2]
ζA (0) 0
log det A = + (γ − log 4π)ζA (0) − ζA (0) + O(ω − 2). (7.10)
ω−2
Note that the last term in the last two equations is not important as it can be absorbed in
the normalization constant.
This should not be surprising as the analytic continuation of both the Γ and ζ function
are intrinsically linked through the functional equation (2.22). This has enabled us to show
in two very different techniques (yet, necessarily equivalent from a mathematical point of
view) for instance that
λ2 φ4cl φ2cl

λ 25
V (φcl ) = φ4cl (x̄) + log − . (7.11)
4! 256π 2 M2 6
The latter technique has certain - conceputal and computational - advantages over the for-
mer which we proceed to explain now. According to Elizalde, Odintsov, and Romeo there is
somewhat of a distaste in using ζ function regulazation in important scientific journals and
prefer to use dimensional regularization because the former (ζ, that is) procedure seems
ambiguous and ill defined.
Quoting these authors [2]:

The situation is such that, what is in fact a most elegant, well defined, and unique - in
many aspects - regularization method, may look now to the non-specialist as just one more
among many possible regularization procedures, plagued with difficulties and illdefiniteness.
The rest of this conclusion is based on their defense of the ζ function regularization proce-
dure.
Let us suppose we have a proper-time Hamiltonian H of a quantum system with boundary
conditions in a background field. This is equivalent to a differential operator A with corre-
sponding boundary conditions. Irrespective of whether the spectrum of A may be computed
explictily or not, to any such operator, we can define ζA rigorously as
ζA (s) = TrA−s . (7.12)
As we have seen several times, when the eigenvalues λn of A form a discrete set and can be
computed explicitly (i.e. the eigenvalues of H with boundary conditions and background
field) we obtain
X
ζA (s) = λ−s
n . (7.13)
n
Next, comes the classification of the eigenvalues. If the are of the form an then we consider
the Riemann ζ function and if there are of the form a(n + b) we consider the Hurwitz ζ
function.
Depending on our physical magnitude of interest, we have to compute the ζ function at a
particular value of s. In field theory and quantum mechanics we have used s = 0 but for
example in the vacuum energy of the Casimir effect, which is the sum over the spectrum
1X
ECas = λn (7.14)
2 n
7 CONCLUSION 101
we have used s = −3 yielding
1 π2
ECas = − A ζ(−3). (7.15)
2r3 3
In general, series of the form (7.14) are divergent and this will call for analytic continuation
through the ζ function. In this view, as we stressed in the Introduction, this regularization
is a special case of the mathematical concept of analytic continuation. Since this concept
is defined uniquely and rigorously then so is the regularization procedure.
Let us work out a final example taken from Lang’s Elliptic Functions [5]. When we compute
the Casimir effect of a piecewise uniform closed string, inevitably we will run into a clearly
infinite sum
∞
X
(n + β). (7.16)
n=0
The eigenvalues in the sum λn = n + β are the transverse oscillatons of the string. As we
have pointed out above, this will necessitate the Hurwitz zeta function,
∞
X
ζA (s) = (n + β)−s , (7.17)
n=0
which is valid for Re(s) > 1 but can be analytically continued as a meromorphic function
to the whole complex plane. This was the jewel result of Chapter 2. Having said this, the
ζ regularization procedure assigns unambiguously the value
∞
X
(n + β) = ζ(−1, β) (7.18)
n=0
to our sum (7.16). A mistake would be to write

∞
X
(n + β) = ζ(−1) + βζ(0) (7.19)
n=0
which yields a different result. X. Li, X. Shi and J. Zhang showed [6] in 1991 the necessity
of using the Hurwitz ζ function instead of the Riemann ζ function.
The ζ regularization method can be viewed as one of many possibilities of analytic con-
tinuation in order to make sense of infinite sums. When considered under this light, it
shares some similarities with dimensional regularization. It has been argued that the two
methods also share similar flaws. However, we introduce ζ regularization to solve the prob-
lem of the dependence of the regularized result on the kind of extra dimensions added in
dimensional regularization. It is a fact that a function may not have two different analytic
continuations but the number of ways of defining different analytic continuations in endless.
What remains to be studied is the use that one can make of them. This does not imply,
however, that ζ regularization suffers from the same problem as dimensional regularization.
There exist endless analytical regularization procedures and both ζ and dimensional meth-
ods are but two examples. In the latter one may change any exponent at any place with
the condition that he recovers the starting expression for a particular value of exponent.
7 CONCLUSION 102
A problem that arises in ζ regularization, however, is that the point at which the ζ function
must be evaluated is (precisely) a pole of the analytic continuation. According to Elizalde,
Odintsov, and Romeo [2] one eventually has to use renormalization group techniques to
solve this issue.
These authors also ask the rethoric question: which regularization does Nature use?
The elegance and uniqueness of the ζ technique makes it a plausible candidate.
Analogies
The numbers N (T ) of zeros in the critical strip 0 ≤ σ ≤ 1 is
T T T
N (T ) = log − + O(log T ), (7.20)
2π 2π 2π
and the behaviour of the 1-loop effective Lagrangian as B → 0 is
(1) 1
lim Leff = − (m2 log m2 − m2 ). (7.21)
B→0 16π
The Laurent expansion of log det A around ω = 2 is
ζA (0) 0
log det A = + (γ − log 4π)ζA (0) − ζA (0) + O(ω − 2), (7.22)
ω−2
where as the sum of the zeros ρ of the Riemann ζ function is
X 1 1

1
+ = [2 + γ − log 4π]. (7.23)
ρ 1−ρ 2
Imρ>0
The structure of these equations indicates that there might be a connection between the
distribution of the zeros of the ζ function and the behaviour of quantum field theories worth
exploring.
REFERENCES
- [1] Bailin and Love Introduction to Gauge Field Theory
- [2] EOR’s Zeta Regularization Techniques with Applications
- [3] Michael E. Peskin and Daniel V. Schroeder An Introduction to Quantum Field Theory
- [4] H. B. G. Casimir and D. Polder The Influence of Retardation on the London-van der
Waals Forces
- [5] Serge Lang Elliptic Functions
- [6] X. Li, X. Shi and J. Zhang Phys. Rev. D44 (1991) 560
A APPENDIX 103
A Appendix
A.1 Generalized Gaussian integrals
Claim 1: The following holds
Z Z
dx1 · · · dxn exp iλ (x1 − a)2 + (x2 − x1 )2 + · · · + (b − xn )2

s
in π n

iλ 2
= exp (b − a) (A.1)
(n + 1)λn n+1
Proof. (by induction)

Assume it is true for n and show it is true for n + 1
Z Z
dx1 · · · dxn+1 exp iλ (x1 − a)2 + (x2 − x1 )2 + · · · + (b − xn )2

s
in π n
Z
iλ
= dxn+1 exp (xn+1 − a) exp[iλ(b − xn+1 )2 ]
2
(n + 1)λn n+1
2 Z
in π n

1 2 2
= dxn+1 exp iλ (xn+1 − a) + (b − xn+1 ) , (A.2)
(n + 1)λn n+1
the exponential in the integrand can be worked out as followsb by setting xn+1 − a = y
1 n+2 2
(xn+1 − a)2 + (b − xn+1 )2 = y − 2y(b − a) + (b − a)2
n+1 n+1
2
n+2 n+1 1
= y− (b − a) + (b − a)2 .
n+1 n+2 n+2
Finally, let λ − ((n + 1)/(n + 2))(b − a) = z so that the integral becomes
s s
in π n in+1 π n+1
Z
n+2 2 iλ 2 iλ 2
dz exp iλ z + (b − a) = exp (b − a) ,
(n + 1)λn n+1 n+2 (n + 1 + 1)λn+1 n+2
(A.3)
and this concludes the proof by induction.
Claim 2: If M is a symmetric N × N matrix with real-valued elements Mij and q and J

are N component vectors with components qi and Ji respectively, then
(2π)N/2
Z
N 1 T T 1 T −1
Z(J) = d q exp − q Mq + J q = √ exp J M J (A.4)
2 det M 2
Proof. The process is to diagonalize the matrix M as M = ΛM̃ΛT where the following
relations hold
 
m̃1 · · · 0
ΛT Λ = 1, det Λ = 1, M̃ =  ... .. ..  , (A.5)

. . 
0 · · · m̃N
A APPENDIX 104
the integral then becomes

Z
N 1 T T T
Z(J) = d q exp − q ΛM̃Λ q + J q , (A.6)
2
next we define the following q̃ = ΛT q and J̃ = ΛT J from which it follows that dN q =

dN q̃ det Λ = dN q̃ and
Z Z " #
1 T X 1
N T
Z(J) = d q exp − q̃ M̃q̃ + J̃ q̃ = d q̃i exp N
− m̃i q̃i + J˜i q̃i
2
2 i
2
 ∞ " #  " #!
N Z X 1 N r
˜2
Y Y 2π J
=  dN q̃i exp − m̃i q̃i2 + J˜i q̃i  = exp i
i=1 i
2 i=1
m̃i 2m̃i
−∞
!−1/2 !
Y X J2
i
= (2π)N/2 m̃i exp . (A.7)
i i
2m̃ i
Finally we note that the inverse of the diagonal matrix is

 −1 
m̃1 ··· 0
M̃−1 =  ... .. ..  (A.8)

. . 
0 · · · m̃−1
N
and therefore the last product is the determinant of the matrix

Y
m̃i = det M̃ = det M, (A.9)
i
and consequently the result follows
(2π)N/2 (2π)N/2

1 T −1 1 T −1
Z(J) = √ exp J̃ M J̃ = √ exp J M J . (A.10)
det M 2 det M 2
Claim 3: One has the following
δ 2 SEuc
Z
Z
1 4 4
ZE [J] = NE exp {−SEuc [φ(x0 ), J]} Dφ exp − d x1 d x2 φ(x1 ) φ(x2 )
2 δφ(x1 )δφ(x2 )
exp {−SEuc [φ(x0 ), J]}
= NE0 q . (A.11)
det (−∂¯µ ∂¯µ + m2 + V 00 [φ(x0 )])δ(x1 − x2 )

Proof.
This follows from Claim 2 with the action
Z
1 1
SE [φ, J] = d4 x̄ ∂¯µ φ∂¯µ φ + m2 φ2 + V (φ) − Jφ (A.12)
2 2
as expanded on (4.136) and using
δ2
SEuc = δ(x̄1 − x̄2 ) −∂¯µ ∂¯µ + m2 + V 00 [φ(x1 )] ,

(A.13)
δφ(x1 )δφ(x2 )
A APPENDIX 105
δ 2 SEuc
Z Z
1
ZE [J] = NE exp {−SEuc [φ(x0 ), J]} Dφ exp − d4 x1 d4 x2 φ(x1 ) φ(x2 )
2 δφ(x1 )δφ(x2 )
(A.14)
and this proves the claim.
The formulas
∞
X (2x)2k
1 1 2x X 1
coth x = + 2 B2k , coth πx = + . (A.15)
x (2k)k πx π x2 + n2
k=1 n≥1
are almost always quoted as Gradshteyn and Ryzhik, p.35. They represent the Laurent
series. There is very little added value in reproducing the proofs here.
A.2 Grassmann numbers

Let us introduce some notation first. The following presentation about Grassmann numbers
follows the notes of A. Rajantie and Peskin and Schroeder. Ordinary commuting numbers
will be denoted c-numbers (these can be real or complex). Now let n generators {θ1 , · · · , θn }
satisfy the anti-commutation relations
{θi , θj } = 0 ∀i, j. (A.16)
Then the set of the linear combinations of {θi } with the c-number coefficient is called the
Grassmann number and the algebra generated by {θi } is called the Grassmann algebra,
denoted by Λn . Let us take an arbitrary element g of this algebra expand it as
n
X X X 1 X
g(θ) = g0 + gi θi + gij θi θj + · · · = gi1 ,··· ,ik θi1 · · · θik , (A.17)
i=1 i<j
k!
0≤k≤n {i}
where g0 , gi , gij , · · · and gi1 ,··· ,ik are c-numbers that are anti-symmetric under the exchange
of two indices. Additionally, we can write g as
X
g(θ) = g̃k1 ,··· ,kn θ1k1 · · · θnkn . (A.18)
ki =0,1
It is impossible for the set of Grassmann numbers to be an ordered set because the generator
θk does not have a magnitude. The only number that is both c-number and Grassmann
number is zero, moreover, a Grassmann number commutes with a c-number. From the
discussion above it follows that
θk2 = 0, (A.19)
θk1 θk2 · · · θkn = εk1 k2 ···kn θ1 θ2 · · · θn , (A.20)
θk1 θk2 · · · θkm = 0 (m > n). (A.21)
The tensor εk1 k2 ···kn is the Levi-Civita symbol, defined as


 +1 if{k1 · · · kn } is an even permutation of {1 · · · n}
εk1 k2 ···kn = −1 if{k1 · · · kn } is an odd permutation of {1 · · · n} (A.22)
0 otherwise.

A APPENDIX 106
Functions of Grassmann numbers are defined in terms of Taylor expansions of the function.
If n = 1 we have the simple expression
eθ = 1 + θ (A.23)
since terms O(θ2 ) are zero.

Our next step is to develop the theory of differentiation and integration of Grassmann
variables, this theory has a few surprising facts, for instance differentiation is the same
process as integration.
We assume that the differential operator acts on a function from the left, let θi and θj be
two Grassmann variables then
∂θj ∂
= θj = δij . (A.24)
∂θi ∂θi
Similarly, we assume that the differential operator anti-commutes with θk . The product
rule has the slightly different form
∂ ∂θj ∂θk
(θj θk ) = θk − θj = δij θk − δik θj . (A.25)
∂θi ∂θi ∂θi
Moreover, the following properties hold
∂ ∂ ∂ ∂ ∂2
+ = 0 ⇒ 2 = 0, (A.26)
∂θi ∂θj ∂θj ∂θi ∂θi
the last equation is termed nil-potency; and finally

∂ ∂
θj + θj = δij . (A.27)
∂θi ∂θi
Let us now move to integration. To this end, we adopt the notation D for differentiation
with respect to a Grassmann variable and ∫ for integration. Let us suppose that these
operations satisfy the relations
∫ D = D ∫ = 0, D(A) = 0 ⇒ ∫ (BA) = ∫ (B)A, (A.28)
where A and B are arbitrary functions of Grassmann variables. The first part of the first
equation implies that the integral of a derivative gives a surface term and it is set to zero,
whereas the second part implies that the derivative of an integral vanishes. The last equation
implies that if the derivative of the function is zero then it can be taken out the integral.
The relations are satisfied when D is proportional to ∫ and for normalization purposes we
set ∫ = D and write
Z
∂f (θ)
dθg(θ) = . (A.29)
∂θ
From the previous definition it follows that
Z Z
∂1 ∂θ
dθ = = 0, dθ θ = = 1. (A.30)
∂θ ∂θ
Z
∂ ∂ ∂
dθ1 dθ2 · · · dθn g(θ1 θ2 · · · θn ) = ··· g(θ1 θ2 · · · θn ). (A.31)
∂θ1 ∂θ2 ∂θn
A APPENDIX 107
In a theory where differentiation is equivalent to integration we would expect some strange

behaviour. In order to see this behaviour we consider the simpler case when we have only
one generator and we change variables θ0 = aθ where a is a complex number, then one has
∂g(θ0 /a)
Z Z
∂g(θ)
dθg(θ) = = = a dθ0 g(θ0 /a), (A.32)
∂θ ∂θ0 /a
which implies that dθ0 = (1/a)dθ. The extension to the general case with n generators
yields θi → θi0 = aij θj and hence
Z
∂ ∂ ∂
dθ1 dθ2 · · · dθn g(θ) = ··· g(θ)
∂θ1 ∂θ2 ∂θn
n
X ∂θk0 1 ∂θ0 ∂ ∂
= · · · kn 0 · · · 0 g(a−1 θ0 )
∂θ1 ∂θn ∂θk1 ∂θkn
ki =1
n
X ∂ ∂
= εk1 ···kn ak1 1 · · · akn n 0 · · · 0 g(a−1 θ0 )
∂θk1 ∂θkn
ki =1
Z
= det a dθ10 · · · dθn0 g(a−1 θ0 ). (A.33)
Consequently the measure has the Jacobian
dθ1 dθ2 · · · dθn = (det a)dθ10 · · · dθn0 . (A.34)
In the case of a single variable, the delta function of a Grassmann variable is defined in a
similar fashion as with c-numbers defined as
Z
dθ δ(θ − z)g(θ) = g(z). (A.35)
However, in the case of Grassmann variables we can obtain a closed expression for the delta
function. If we set g(z) = a + bz in the definition of the delta function we have
Z
dθ δ(θ − z)(a + bθ) = a + bz (A.36)
and this means that
δ(θ − z) = θ − z. (A.37)
Again, we can extend this to n generators if we are careful about the order of the variables
δ n (θ − z) = (θn − zn ) · · · (θ2 − z2 )(θ1 − z1 ). (A.38)
We can find the integral of the delta function by considering complex Grassmann variables
which we proceed to develop later. Consider
Z Z
iξθ
dξ e = dξ (1 + iξθ) = iθ (A.39)
so that we have
Z
δ(θ) = θ = −i dξ eiξθ . (A.40)
A APPENDIX 108
One of the most crucial developments of the Grassmann variables is the Grassmann Gaussian
integral which will be fundamental when developing the path integral formalism of fermions.
Let us evaluate the following integral
 
Z X
I = dθ1∗ dθ1 · · · dθn∗ dθn exp  θi∗ Mij θj , (A.41)
ij
where it is important to stress that {θi } and {θi∗ } are two independent sets of Grassmann
variables. Since Grassmann variables θi and θi∗ anti-commute we can take the n×n c-number
matrix M to be anti-symmetric. The formula for P the transformation of the measure solves
the problem of the computation. Setting θi0 = j Mij θj yields
Z ! Z n
X
I = det M dθ1∗ dθ10 · · · dθn∗ dθn0 exp − θi∗ θi0 = det M ∗ 0 ∗
dθ dθ(1 + θ θ ) = det M.
i
(A.42)
Complex conjugation is defined as
(θi )∗ = θi∗ , (θi∗ )∗ = θi . (A.43)
In the case of Grassmann variables we have
(θi θj )∗ = θj∗ θi∗ . (A.44)
The reasoning behind (2.94) is that the real c-number θi θi∗ does not satisfy (θi θi∗ )∗ = θi θi∗ .
Let us recall that the annihilation and creation operators c and c† satisfy the anti-commutation
relations {c, c† } = 1 and {c, c} = {c† , c† } = 0 and that the number operator N = c† c has
eigenvectors |0i and |1i. We are now in a position to study the Hilbert space Ω spanned by
these vectors, i.e.
Ω = span {|0i , |1i} . (A.45)
An arbitrary vector |ωi ∈ Ω can be written in the form
|ωi = |0i ω0 + |1i ω1 , (A.46)
with ωi ∈ C where i = 1, 2.
Next we define the coherent states
|θi = |0i + |1i θ, hθ| = h0| + θ∗ h1| , (A.47)
where θ and θ∗ are Grassmann numbers.

The coherent states are eigenstates of c and c† respectively, that is
c |θi = |0i θ = |θi θ, hθ| c† = θ∗ h0| = θ∗ hθ| . (A.48)
It can be shown fairly easily that the following identities hold

0∗
hθ0 | θi = 1 + θ0∗ θ = eθ θ
, (A.49)
hθ | gi = g0 + θ∗ g1 , (A.50)
A APPENDIX 109
θ |c† | g = hθ | 1i g0 = θ∗ g0 = θ∗ hθ | gi ,

(A.51)
∂
hθ |c| gi = hθ | 0i g1 = hθ | gi . (A.52)
∂θ∗
Finally, we show how matrix elements are represented and the completeness relation. Let
p(c, c† ) = p00 + p10 c† + p01 c + p11 c† c, pij ∈ C (A.53)
be an arbitrary function of c and c† .

The complex matrix elements of p can be written in terms of scalar products as
h0 |p| 0i = p00 , h0 |p| 1i = p01 , h1 |p| 0i = p10 , h1 |p| 1i = p00 + p11 . (A.54)
From these scalar products we can form the more general product
∗ 0
hθ |p| θ0 i = (p00 + θ∗ p10 + p01 θ0 + θ∗ θ0 p11 )eθ θ
. (A.55)
Moreover, one has

Z Z
∗
dθ∗ dθ |θi hθ| e−θ θ = dθ∗ dθ (|0i + |1i θ) (h0| + θ∗ h1|) (1 − θ∗ θ), (A.56)
θ |c† | g = hθ | 1i g0 = θ∗ g0 = θ∗ hθ | gi ,

(A.57)
∂
hθ |c| gi = hθ | 0i g1 = hθ | gi . (A.58)
∂θ∗
and therefore the completeness relation is
Z
∗
dθ∗ dθ |θi hθ| e−θ θ = I. (A.59)
A.3 The Mellin transform and the series expansion of log Γ(s + 1)
We shall follow Titchmarsh’s Theory of Functions and Withaker and Watson’s Modern
Analysis. The Mellin transform connects two functions f (x) and Ψ(s) in the following way
Z∞ σ+i∞
Z
1
Ψ(s) = dxf (x)x s−1
, f (x) = dsΨ(s)x−s . (A.60)
2πi
0 σ−i∞
For example if we take f to be f (x) = e−x then clearly Ψ(s) = Γ(s) for σ > 0. We can also
recover our formula relating the Γ and ζ functions by using f (x) = (ex − 1)−1 in which case
Ψ(s) = Γ(s)ζ(s) for σ > 1. From the Weierstrass product we can write
∞
z Y −z/n z Γ(a)
1+ e 1+ = e−γz (A.61)
a n=1 a+n Γ(z + a)
and take the principal values of the logarithms to be

∞ X ∞
z Y z z
log 1 + + log e−z/n 1 + = e−z/n 1 +
a n=1
a+n n=1
a+n
A APPENDIX 110
∞ ∞ ∞ ∞
( ) ( )
(−1)m−1 (−1)m−1 z m

X −az X z X −az X
= + + + ,
n=1
n(a + n) m=2
m (a + n)m n=1
n(a + n) m=1
m am
(A.62)
and the absolute convergence of the series is assured for |z| < a since
∞
X a|z| |z| |z|
− log 1 − + (A.63)
n=1
n(a + n) a+n a+n
converges. Now, taking logarithms we have

∞ ∞
e−γz Γ(a) z X az X (−1)m−1 m
log = − + z ζ(m, a). (A.64)
Γ(z + a) a n=1 n(a + n) m=2 m
Next, we need to consider

πz s
I
1
− ds ζ(s, a) (A.65)
2πi s sin πs
C
where the contour C is like Figure 2.4 except that it encloses the points s = 2, 3, 4, · · · but
not the points 1, 0, −1, −2, · · · . The residue of this integral at s = m ≥ 2 is given by
πz s πz s
res ζ(s, a) = lim (s − m) ζ(s, a)
s=m s sin πs s→m s sin πs
m
πz s−m (−1)m m
= ζ(m, a) lim = z ζ(m, a). (A.66)
m s→m sin πs m
Since ζ(s, a) = O(1) as σ → ∞ then the integral converges if |z| < 1. By Cauchy’s residue
theorem we can change the sum involving the ζ(m, a) term for the integral
∞
e−γz Γ(a) πz s
I
z X az 1
log = − − ds ζ(s, a). (A.67)
Γ(z + a) a n=1 n(a + n) 2πi s sin πs
C
Taking the exponential outside the logarithm and using the well known formula
∞
d X z
log Γ(z + 1) := ψ (0) (z + 1) = −γ (A.68)
dz n=1
n(n + x)
we see that
Γ0 (a) πz s
I
Γ(a) 1
log = −z − ds ζ(s, a). (A.69)
Γ(z + a) Γ(a) 2πi s sin πs
C
We now let D be a semicircle of large radius N with center at s = 32 , the semicircle lying
on the right of the line σ = 32 . On this semicirle, ζ(s, a) = O(1) as well as |z| < 1 and
−π + δ ≤ arg z ≤ π − δ where δ > 0 and
πz s
= O(|z|σ e−δ|t| ), (A.70)
s sin πs
A APPENDIX 111
therefore
πz s
Z
lim ds ζ(s, a) = 0. (A.71)
N →∞ s sin πs
D(N )
It follows immediately that if |arg z| ≤ π − δ, and |z| < 1 then

3/2+i∞
Γ0 (a) πz s
Z
Γ(a) 1
log = −z + ds ζ(s, a). (A.72)
3/2−i∞
However, the above integral defines an analytic function of z for all values of |z| if |arg z| ≤
π − δ then by analytic continuation it is valid for all values of |z| when |arg z| ≤ π − δ. Let
us consider a cutoff on this integral of the sort
3/2±iR
Γ0 (a) πz s
Z
Γ(a) 1
log = −z + ds ζ(s, a) (A.73)
−n−1/2±iR
where n is a fixed integer and R → ∞. As we have seen the integrand is O(|z|σ e−δR Rτ (σ) )
where −n − 12 ≤ σ ≤ 32 and thus irrespective of which sign we take in the limits of the
integral, the integral goes to zero as R → ∞.
Applying Cauchy’s residue theorem again yields
−n−1/2+i∞
n
Γ0 (a) πz s πz s
Z
Γ(a) 1 X
log = −z + ds ζ(s, a) + res ζ(s, a).
Γ(z + a) Γ(a) 2πi s sin πs m=−1
s=−m s sin πs
−n−1/2−i∞
(A.74)
On this new path of integration we have

πz s

−n−1/2 −δ|t|τ (−n−1/2)

s sin πs ζ(s, a) < Kz
e |t|, (A.75)
with K independent of both z and t and τ (σ) the function used above. Moreover, since
Z∞
dte−δ|t|τ (−n−1/2) t < ∞ (A.76)
−∞
we obtain
n
Γ(a) Γ0 (a) X πz s
log = −z + res ζ(s, a) + O(z −n−1/2 ) (A.77)
Γ(z + a) Γ(a) m=−1 s=−m s sin πs
provided that |z| is large. We next need to compute the residues at s = 0 and s = −1.
Both computations require the expansions around s = 0 and s = −1 as follows
1
ζ(s, a) = ζ(0, a) + ζ 0 (0, a)s + O(s2 ) = − a + ζ 0 (0, a)s + O(s2 ), (A.78)
2
z s = 1 + s log z + O(s2 ), (A.79)

A APPENDIX 112
π 1
= + ζ(2)s + O(s2 ), (A.80)
sin πs s
so that

1 1 0
(1 + ζ(2)2 + · · · )(1 + s log z + · · · ) − a + sζ (0, a) + · · · , (A.81)
s 2
and the residue at s = 0 is given by
πz s

1 1 1
res ζ(s, a) = − a log z + ζ 0 (0, a) = − a log z + log Γ(a) − log 2π.
s=0 s sin πs 2 2 2
(A.82)
For the residue at s = −1 we use the following expansions
1 Γ0 (a)
ζ(s, a) = − + O(s − 1), (A.83)
s−1 Γ(a)
z s = z + (s − 1)z log z + O((s − 1)2 ), (A.84)
π 1
=− − ζ(2)(s − 1) + O((s − 1)2 ), (A.85)
sin πs s−1
so that the same technique yields
πz s Γ0 (a)
res ζ(s, a) = −z log z + z + z . (A.86)
s=−1 s sin πs Γ(a)
Finally, if |z| is large and |arg z| ≤ π − δ we have

n 0
1 1 X Bk+2 (a)
log Γ(z + a) = z+a− log z − z + log 2π + (−1)k−1 z −k
2 2 k(k + 1)(k + 2)
k=1
+ O(z −n−1/2 ), (A.87)
where we have used

πz s (−1)m z −m (−1)m z −m Bm+1 (a)
res ζ(s, a) = ζ(−m, a) = , (A.88)
s=−m s sin πs −m m m+1
and assumed a result on Bernoulli numbers
0
Bm+2 (a)
Bm+1 (a) = . (A.89)
(m + 1)(m + 2)
As a side remark, note that specializing for the case a = 1 of the Riemann ζ function and
replacing s by 1 − s we have
σ+i∞
Γ0 (1 + x) ζ(1 − s) s
Z
1
− log x = − ds x . (A.90)
Γ(1 + x) 2i sin πs
σ−i∞
A APPENDIX 113
Using an inverse Mellin transform with
Γ0 (1 + x) πζ(1 − s)
f (x) = − log x, Ψ(s) = − , (A.91)
Γ(1 + x) sin πs
we obtain (again swapping 1 − s for s)

Z∞
Γ0 (1 + x)

sin πs
ζ(s) = − dxx−s − log x (A.92)
π Γ(1 + x)
0
valid for 0 < σ < 1, a formula known as Kloosterman’s equation. With the assistance of
Z∞
Γ0 (x) 1 t
= log x − −2 dt (A.93)
Γ(x) 2x (t2 + x2 )(e2πt − 1)
0
we can write
Γ0 (1 + x) Γ0 (x) 1
− log x = + − log x
Γ(1 + x) Γ(x) x
Z∞
1 t
= − 2 dt 2
2x (t + x2 )(e2πt − 1)
0
Z∞
t 1 1
= −2 dt − . (A.94)
t 2 + x2 e2πt − 1 2πt
0
Following Titchmarsh we have

Z∞ Z∞
2 sin πs −s t 1 1
ζ(s) = dxx dt 2 −
π t + x2 2πt
e − 1 2πt
0 0
Z∞ Z∞
x−s

2 sin πs 1 1
= dt − t dx 2
π e2πt − 1 2πt t + x2
0 0
Z∞
sin πs 1 1
= dt − t−s
cos πs
2 e2πt − 1 2πt
0
Z∞
πs 1 1
= 2(2π) s−1
sin dt − u−s
2 eu − 1 u
0
s−1 πs
= 2(2π) sin Γ(1 − s)ζ(1 − s), (A.95)
2
i.e. the functional equation of the Riemann ζ function (2.2).
B REFERENCES 114
B References
B.1 Books
Apostol, Tom ”Introduction to Analytic Number Theory”
Bailin and Love ”Introduction to Gauge Field Theory”
Elizalde, Odintsov, Romeo, Bytsenk, Zerbini ”Zeta Regularization Techniques with Appli-
cations”
Freitag, Eberhard and Busam, Rolf ”Complex Analysis”
Grosche C. and Steiner F, ”Handbook of Feynman Path Integrals”
Ingham, A.E. ”The Distribution of Prime Numbers”
Ivic Riemann’s Zeta Function
Kleinert, Hagen ”Path Integrals in Quantum Mechanics, Statistic, Polymer Physics and
Financial Markets”
Hatfield, Brian ”Quantum Field Theory of Point Particles and Strings”
Lang, Serge Elliptic Functions
Nakahara, Mikio Geometry, Topology and Physics
Peskin, Michael E. and Schroeder, Daniel V. ”An Introduction to Quantum Field Theory”
Rajantie, Arttu ”Advanced quantum field theory Course Notes”
Ramond, Pierre ”Field Theory: A Modern Primer”
Rivers, Ray J. ”Path Integral Methods in Quantum Field Theory”
Titchmarsh, E. C. and Heath-Brown, Roger ”The Theory of the Riemann Zeta-Function”
Titchmarsh, E. C. ”Theory of Functions”
Weinberg, Steven ”The Quantum Theory of Fields: Volume I Foundations”
Whitaker and Watson ”A Course on Modern Analysis”
B.2 Papers
Casimir and Polder The Influence of Retardation on the London-van der Waals Forces
Conrey, J. Brian The Riemann Hypothesis
Hawking, Stephen Zeta Function Regularization of Path Integrals in Curved Spacetime
R. Jackiw, Phys. Rev. D9 1686 (1974)
’t Hooft and Veltman, Nucl. Phys. 44B 189
S. Coleman and E. Weinberg Phys. Rev. D7 1888 (1973)
X. Li, X. Shi and J. Zhang Phys. Rev. D44 (1991) 560
B.3 Internet Articles

The Force of Empty Space http://focus.aps.org/story/v2/st28
E.L. Losada ’Functional Approach to the Fermionic Casimir Effect’
The Casimir effect http://socrates.berkeley.edu/b̃udker/Physics250 Spring0607/Corsini Casimir.ppt.
Wikipedia: Casimir Effect http://en.wikipedia.org/wiki/Casimir effect
A Force from Empty Space: The Casimir Effect http://apod.nasa.gov/apod/ap061217.html
Wikipedia on the Gamma and Zeta functions
Sondow, Jonathan and Weisstein, Eric W. Riemann Zeta Function.” From MathWorld–A
Wolfram Web Resource

Zeta Function Regularization PDF

Uploaded by

Copyright:

Available Formats

Zeta Function Regularization PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Zeta Function Regularization PDF

Uploaded by

Copyright:

Available Formats

Zeta Function Regularization

Department of Theoretical Physics

September 25th 2009

Supervior: Dr Arttu Rajantie

Submitted in partial fulfillment

E. Elizalde and A. Romeo

2 Introduction to the Riemann Zeta Function 11

3 Zeta Regularization in Quantum Mechanics 39

5 Zeta Regularization in Field Theory 73

It is interesting to note that at a more academic level however ζ regularization is hardly

The ζ function arises naturally by using

and the functional determinant of the operator can then be written as

Now, we have mentioned operators must be differential and quadratic.

Z(β) = Tr exp(−β Ĥ)

where θ and θ∗ are conjugate Grassmann numbers yields [5]

can be written in Euclidean spacetime as [3], [7]

Riemann ζ function ⇔ bosons Hurwitz ζ function ⇔ fermions.

- [1] J. Brian Conrey The Riemann Hypothesis

2 Introduction to the Riemann Zeta Function

Definition 1 The integral

The first lemma generalises the factorial function as follows

Lemma 1 For any n ∈ N we R ∞have Γ(n) = (n − 1)!.

In order to have a complete view of the Γ function we need to extend it to a meromor-

Theorem 1 The Γ function extends to a meromorphic function on the complex plane.

Theorem 2 For s ∈ C any we have

Γ(s + 1) = sΓ(s). (2.5)

Proof. This follows directly from Lemma 1 and Theorem 1.

The three-dimensional representation of the Γ function is shown in Figure 2.1.

Figure 2.1: |Γ(x + iy)| for −5 ≤ x ≤ 3 and −1 ≤ y ≤ 1

In an analogous form we have

Multiplying these two together we have

and switching to polar coordinates u = r cos θ, v = r sin θ, dudv = rdrdθ

The integral can be simplified by setting z = sin2 θ

Lemma 3 For 0 < y < 1 we have

Figure 2.2: C = CR ∪ Cε ∪ [ε, R] ∪ [R, ε]

|z −y | = |e−y log z | = e−yRe(log z) = e−y log |z| = |z|−y ,

so that we are left with

We may re-write this to obtain the final result

This proves the claim of the lemma.

Let us now make the concluding remarks.

Theorem 3 (Euler Reflection Formula) For all s ∈ C one has

Proof. The lemma we have just proved can be written as

Corollary 1 One has

Theorem 4 The Γ function has no zeroes.

Lemma 4 If sn := 1 + 12 + · · · + n1 − log n, then limn→∞ sn exists. This limit is called

The series on the right converges to a positive constant since

Next, the following holds

which means that

Convergence now follows because sn decreases monotonically whereas tn increase monoton-

The value of γ was computed by Mascheroni in 1790 to be

and of course subsequent improvements have been made.

Lemma 5 Weierstrass product of the Γ function is

and integrating by parts yields

2.2 The Hurwitz function ζ(s, a)

Let us now examine the properties of the Hurwitz ζ function.

Figure 2.3: |ζ( 12 + it, a)| for 1 ≤ t ≤ 50 and 1

Proof. From the inequalities

n=1 n=1 n=1

all the statements follow and this proves the claim.