BROWN-HET-849

August 1992

The Entropy of the Gravitational Field

R. Brandenberger, T. Prokopec
arXiv:gr-qc/9208009v1 20 Aug 1992

Department of Physics, Brown University, Providence, Rhode Island 02912, USA

and

V. Mukhanov∗

Institute of Theoretical Physics, ETH Hönggerberg, CH-8093 Zürich, Switzerland

ABSTRACT

We derive a formula for the nonequilibrium entropy of a classical stochastic field

in terms of correlation functions of this field. The formalism is then applied to define

the entropy of gravitational perturbations (both gravitational waves and density fluc-

tuations). We calculate this entropy in a specific cosmological model (the inflationary

Universe) and find that on scales of interest in cosmology the entropy in both density

perturbations and gravitational waves exceeds the entropy of statistical fluctuations

of the microwave background.

The nonequilibrium entropy discussed here is a measure of loss of information

about the system. We discuss the origin of the entropy in our cosmological models

and compare the definition of entropy in terms of correlation functions with the

microcanonical definition in quantum statistical mechanics.

1
 98.80.Cq, 04.20.Cv, 05.20.Gg

Typeset Using REVTEX

2
 I. INTRODUCTION

The concept of entropy plays an important role in all branches of physics. Hence, it

is no surprise that entropy can also be defined in general relativity and cosmology. The
most famous application of entropy in general relativity is to black holes. Bekenstein’s [1]
realization that the area of a black hole behaves like an entropy led to the discovery of
black hole radiation [2]. Other important applications of entropy in cosmology relate to the

characterization of the initial state of the universe [3] and to the study of the post-bounce
state in bouncing cosmologies.
Entropy expresses the loss of information about the system under consideration [4]. Quite
a long time ago, Jaynes [4] argued that entropy expresses the extent of human ignorance
about a system and is therefore an antropomorphic concept. One can uniquely define entropy

only after having specified the position of the observer with respect to the system.
Once we accept that entropy measures the loss of information about a system, it becomes
important to ask whether there is a natural “coarse graining”, i.e. can it be uniquely specified
which part of the information about a system is lost? The problem of defining the entropy

becomes the task of determining the correct coarse graining.

We wish to address this question in order to define the entropy of the gravitational
field. From a naive point of view, gravitational instability, which is responsible for the
formation of structure in the Universe, should lead to an increasing entropy - in agreement

with the second law of thermodynamics. In the initial state in which the gravitational
field is (almost) uniform, the gravitational entropy (almost) vanishes. In contrast, the later
state of the gravitational field which results from gravitational instability can be viewed
as a particular realization of some stochastic process producing density perturbations and
gravitational waves. Hence, there should be an associated entropy which characterizes the

naturalness of the occurrence of the given distribution or, more quantitatively, measures the
probability of the distribution.
To characterize the measure of a state of gravitational radiation or density perturbations

3
in a quantitative manner, we need a well defined and well justified notion of entropy of
gravitational perturbations. Since the states we are interested in are far from thermal
equilibrium, our task will be to define an entropy for nonequilibrium systems in cosmology.

There exists already a considerable body of work on the definition of nonequilibrium

entropy, in particular in the context of cosmology. For example, Smolin [5] derived a formula
for the entropy of a “quantum field” state of gravitational radiation and applied it to estimate
the entropy of some astrophysical sources of gravitational radiation.

In a series of paper, Hu and Paron [6], Kandrup [7,8] and Kandrup and Hu [9] have
discussed the entropy of particles produced in an expanding universe. They gave a definition
of entropy based on the single particle distribution function (density matrix) and showed
that this entropy increases in time (if the initial state is free of correlations) as a result of

particle production. This definition of entropy (and in particular the role of coarse graining
and loss of correlations) was further discussed by Habib and Kandrup [10].
Some interesting speculations about the entropy Sg of the gravitational field have been
made by Penrose [3] and Hu [11], who propose that Sg is proportional to the integral of the

Weyl tensor squared C 2 over space. This definition expresses the expectation that metric
fluctuations should give rise to entropy. With this formulation, Penrose’s initial condition
criterion [3] C = 0 for the universe is equivalent to the assumption that the universe starts
in a state of vanishing gravitational entropy.
In this paper we use two quite different approaches to define the nonequilibrium entropy

of cosmological perturbations. One of them is based on the microcanonical ensemble [12]

of quantized gravitational perturbations while the other is a formula for entropy which
can be associated with the stochastic distribution which describes the state of the classical
gravitational field. We show that these two definitions are in agreement and discuss the

physical meaning of the entropy of gravitational fluctuations. Then, we apply our definitions
to estimate the entropy of gravitational waves and linear density inhomogeneities produced
by inflation [13]. We also indicate how to apply our methods to other cosmological models,
e.g., those based on phase transitions [14].

4
 Our analysis is based on the fact that the theory of gravitational waves and of linearized
density perturbations in an expanding universe can be reduced [15] to the study of a real
scalar field in an external classical background. Hence, we will investigate the more general

question of how to define the entropy of a scalar field in a nonequilibrium state. We can
either study the quantum theory of this field and use the microcanonical ensemble, or we
can view the classical field as a stochastic process and determine its entropy.
The paper is organized in the following manner: after some general comments about

nonequilibrium entropy in section 2, we give in sections 3 and 4 the quantum and classical
definitions of entropy. In section 5 we show the equivalence of the two definitions when
applied to situations when both are applicable. We also demonstrate that the entropy
is a result of coarse graining. In section 6 we briefly review the gauge invariant theory of

cosmological perturbations upon which our definition of the entropy of the gravitational field
is based. For a pedagogical introduction, the reader is referred to Ref. 16, for an extensive
review to Ref. 15. Section 7 contains the main applications of our work: the evaluation of
the entropy of gravitational waves and density perturbations in inflationary universe models.

We use units in which h̄ = kB = c = 1. Greek indices run over space-time variables,

Latin indices only over spatial variables.

II. NONEQUILIBRIUM ENTROPY

We will first develop a general definition of entropy for a system far from thermal equi-

librium, based on the microcanonical ensemble [12]. Let us assume that the state of some
physical system can be completely described by a set of discrete variable J = {I, i, j, . . .}.
If we know that the system is in a certain state J, then the information about the state of
the system is complete and hence the entropy should be zero, as follows from the general
definition of entropy in information theory according to which entropy means the loss of

information. If, on the other hand, we only know the probability distribution PJ for the
system, PJ being the probability to find the system in state J, then the associated entropy

5
is [4,12]

X
S=− PJ ln PJ (1)
J

Now let us assume that we are not interested in or cannot measure the complete (fine
grained) state of the system, but only some coarse grained characteristics, e.g., the value
of the variable I. The coarse grained state can be described by a distribution function PI ,
and the associated entropy is given by the analog of (1) where we sum only over the index

I. This entropy characterizes the measure of solutions of the dynamical system which leads
to the particular coarse grained state.
If the variable J in (1) is continuous, some complications arise. If DJ is the measure on
the space of state, then a probability density p(J) can be defined by

DP (J) = p(J)DJ, (2)

where DP (J) is the probability to find the system in the volume DJ around the state J.
For example, in a system of n particles with Cartesian coordinates x1 , . . . xn and momenta
Qn i
p1 , . . . pn , we would have J ≡ (x1 . . . , xn , p1 , . . . , pn ) and DJ = i=1 (dxi dp ). In a system
with an infinite number of particles, DJ becomes a measure in the space of functionals.
To derive the formula for the entropy of a system with continuous J starting from
Eq. (1), we first divide phase space into sufficiently small cells J1 , J2 . . . with volume elements
∆J1 , ∆J2 , . . .. The probability to find the system in cell n is

PJn ≃ p(Jn )∆Jn , (3)

and hence from (1) the entropy is

X X X
S=− PJn ln PJn ≃ − p(Jn )(ln p(Jn )) ∆Jn − p(Jn ) ln(∆Jn )∆Jn . (4)
n n n

This expression has no limit for ∆Jn → 0 because of the diverging second term which in
general depends on {p(Jn )}. This terms represents the information about the process of
coarse graining.

6
 However, for a simple coarse graining with ∆J1 = ∆J2 = . . ., the second term in (4)
does not depend on the probability distribution P (J) and can hence be neglected as some
irrelevant additive constant − ln ∆J to the entropy. Note that in a quantum dynamical

system, there is a natural choice ∆J = (2πh̄)n due to the uncertainty principle. We conclude
that in the case of a continuous probability distribution, the entropy is defined by coarse
graining and depends on the measure in the phase space of the system. Dropping the second
term in (4) and taking the continuum limit ∆J → 0 gives
Z
S=− p(J) ln p(J)DJ, (5)

where DJ is the functional measure for the variable J.

III. MICROCANONICAL DEFINITION OF ENTROPY

FOR A QUANTIZED FIELD

Let us return to a system whose phase space is described by a set of discrete variables.
Furthermore, we consider the case when the entire system consists of N identical subsystems
(e.g., N photons), each characterized by a discrete set of variable {I, i, j, . . .}. We assume

that there is some principal quantum number I which is completely distinguishable, i.e., any
two states with different I can be experimentally distinguished, and that the other numbers
i, j, . . . correspond to different but experimentally indistinguishable states with the same
value of I. As an example, for a gas of photons in a box we can take I to be the energy of

a photon, and i, j, . . . to correspond to different directions of motion.

The source of entropy in the above setup is the loss of information coming from the
indistinguishability of states with identical I but different i, j, . . . States labelled by I can
be assigned a degeneracy gI which equals the number of microphysical states with identical
quantum number I.

Let us now assume that n̄I subsystems have the same principal quantum number I.
For the moment we assume that the spectrum of the system {n̄I }, i.e., the number n̄I of

7
subsystems with principal quantum number I (for all I), is fixed. Our goal is to calculate
the number of possible microphysical states with a given spectrum which are in principle
distinguishable. The calculation is done for systems with Bose statistics, e.g., photons,

gravitons or scalar type cosmological perturbations.

The problem reduces to calculating the number of possible and distinguishable ways in
which n̄I subsystems can be distributed among gI cells. This number is

(gI − 1 + n̄I )!
Wn̄I = . (6)
n̄I !(gI − 1)!

To obtain this expression, note that there are (n̄I + gI − 1)! ways of dividing n̄I objects by
gI − 1 cell divisions. However, both the particles and the cell divisions are indistinguishable
and hence we must divide by n̄I !(gI − 1)!.

For a system of N subsystems with spectrum {n̄I } (obeying n̄I = N ), the phase
P
I

volume (number of possible states) will be

Y
Γ{n̄I } = Wn̄I . (7)
I

The next step is to assume that all possible states are equally probable. In this case, the
probability for any state α with the given spectrum {n̄I } is

1
P{n̄I } (α) = . (8)
Γ{n̄I }

From (1) it follows that the corresponding entropy of the system with definite spectrum

{n̄I } is

X X
S=− P{n̄I } (α) ln P{n̄I } (α) = ln Γ{n̄I } = ln Wn̄I , (9)
α I

taking into account the normalization condition

X
P{n̄I } (α) = 1. (10)
α

If n̄I ≫ 1, then Stirling’s formula can be applied to approximate Wn̄I in (9). In this
case, the entropy becomes

8
 X
S= gI [(nI + 1) ln(nI + 1) − nI ln nI ] (11)
I

where nI = n̄I /gI are the occupation numbers.

All that was assumed in the above considerations is that the spectrum {n̄I } is well
defined. At no point was thermodynamic equilibrium invoked. Hence, (11) gives a formula
for the entropy of a statistical system with definite spectrum which is valid both in and far

out of thermodynamical equilibrium. It is based on the microcanonical ensemble.

The simplest application of this formula for the entropy is to a black body spectrum of
photons with

2
nk = (12)
eβk −1

with β = 1/T and k = |k|. In this case, each photon is a subsystem. The principal
quantum number I is the energy k, and the other quantum numbers i, j, . . . correspond to
the directions of photon propagation. The degeneracy gk of level I = k is

4π 2
gk = V k dk (13)
3

where V is the volume of space. Substituting (12) and (13) into (11) we obtain the entropy
density of the black body background

S 4π 3
s= ∼ T . (14)
V 3

In the above example, the origin of the entropy is the absence of information about the
direction of propagation of the photons.
If the spectrum {n̄I } is not well known, there is an additional source of entropy. Let us
assume a probability distribution P ({n̄I }) for different spectra. Note that even the total
number N of subsystems need not be fixed. In this case, the space of all possible states is

the direct sum of states for the different spectra {n̄I }. The probability to find the system
in a state α{n̄I } with spectrum {n̄I } is

1
P (α{n̄I } ) = P ({n̄I }) , (15)
Γ({n̄I })

9
and (from (1)) the entropy will be

X X
S=− P (α{n̄I } ) ln P (α{n̄I } )
{n̄I } α{n̄I }
X X
= P ({n̄I }) ln Γ({n̄i }) − P ({n̄I }) ln P ({n̄I }) (16)
{n̄I } {n̄I }

where stand for summation over all possible spectra {n̄I }. If the spectrum is com-
P
{n̄i }

pletely specified, then P ({n̄I }) = 1(0) for {n̄I } = {n̄0I }({n̄I } =

6 {n̄0I }) and (16) reduces to

(9). In the general case, there are two contributions to the entropy. The first term in (16) is
due to the absence of information about the non-principal quantum numbers of the system
for a fixed spectrum, the second comes from our ignorance of the precise spectrum.
In the case of cosmological perturbations, the distribution function P ({n̄I }) for the

spectrum is well localized at a particular spectrum {n̄0I } and hence

S ≃ ln Γ({n̄0I }) . (17)

However, there are examples where the second term in (16) gives the main contribution

to the entropy. For example, in the case of a black hole it is impossible to have information
about the spectrum of the configuration making up the hole, since this information is hidden
behind the horizon. If W is the number of possible different spectra for a black hole of fixed
mass, and if we assume that all of these spectra have equal probability, then, neglecting the
first term in (16), we get

S ≃ ln W . (18)

If the black hole is quantized, then W is finite. In fact, counting the number of spectra of
a black hole with fixed mass gives a formula for the entropy in agreement with the classical

result [17].
Let us return to the discussion of the formula (11) for the entropy of a quantum system
with fixed spectrum. In the classical limit nI ≫ 1, the equation simplifies to

X
S≃ gI ln nI . (19)
I

10
In order to apply this formula, the notion of particles (which is required to be able to
determine nI ) must be well defined. However, when considering quantum fields in some
external field (e.g., cosmological perturbations in an expanding background space-time),

the notion of particles is not always well defined. It is hence desirable to have a formalism
which generalizes the definition of entropy to situations where the number representation is
not well defined.
For large occupation numbers nI , the classical limit should give a good description of

the dynamics of the system. It is therefore convenient to derive a formula for the entropy
directly in terms of the classical field. In the cases when occupation numbers can be defined
for this field, the new definition should reduce to (19).
In the next section, we will give a definition of entropy of a classical field based on the

theory of stochastic processes and show that in the region where both definitions of entropy
are applicable they agree.

IV. ENTROPY OF A CLASSICAL FIELD

Formulas (11) and (19) for the entropy of a field are only applicable if the spectrum of

occupation numbers is well defined. In order for this to be the case, there must be a well
defined notion of particles for the field under consideration.
As will be shown in Section 6, the description of cosmological perturbations and gravi-
tational waves in the Universe can be reduced to the study of a scalar field ϕ(x, t) with a

time dependent effective mass: [15,16]

z ′′
ϕ′′ − c2s ∆ϕ − ϕ = 0. (20)
z

Here, cs is the speed of propagation of perturbations. For gravitational waves cs = 1, whereas

for cosmological perturbations in a Universe with hydrodynamical matter cs is the speed of

sound. The time dependent function z depends on the system and on the background. For
gravitational waves z = a, whereas z is a complicated function of the background parameters

11
in the case of cosmological perturbations (for details see Ref. 15 and Section 6). In the above
a prime denotes differentiation time η.
The quantization of a scalar field obeying (20) is equivalent to the quantization of a

scalar field in some external classical field. If z ′′ /z = 0 and c2s = const 6= 0, then there is
no coupling and the quantum theory for this field (in particular the notion of particle) is
well defined. As will be shown in Section 6, this applies for both cosmological perturbations
and gravitational waves in a radiation dominated Universe. However, this is not the general

case. The ratio z ′′ /z can be nonvanishing and c2s may be zero. For example, for cosmological
perturbations in a matter dominated Universe cs = 0. In this case the solutions of (20) do
not have oscillatory character and it is not possible to define the notion of particles. Hence it
is not possible to define occupation numbers and Equations (11) and (19) are inapplicable.

However, if the perturbations are sufficiently large, the scalar field ϕ can be treated
classically (in the case when occupation numbers are defined, the condition for classicality
is nI ≫ 1). In order to define a notion of entropy valid in this case, we address the more
general question of defining the entropy of a classical scalar field with an action which is

quadratic in field variable and canonical momentum (see also Ref. 18).
The source of the entropy is in this case the ignorance about the exact field configuration.
A state of the system at some time t is specified by the values of the field ϕ(x, t) and its
canonical momentum π(x, t) at all points x in space. We assume that all we know is the
probability distribution P (ϕ(x), π(x)) of the field and its canonical momentum, i.e., we view

ϕ(x) as a stochastic classical field.

The above situation is realized in many situations of interest in cosmology. For example,
in inflationary Universe models, the amplification of scalar field fluctuations during the
period of exponential expansion of the Universe and the nontrivial transition of quantum

fluctuations to classical ones leads to a squeezed state [19,20] for the scalar field starting from
the vacuum state at the beginning of inflation. The Gaussian random state is characterized
D E D E D E
by definite correlation functions ϕ(x, t)ϕ(y, t) , π(x, t)π(y, t) and ϕ(x, t)π(y, t) , where
hqi stands for the ensemble average of the quantity q, (which coincides with the space average

12
of q for a spatially homogeneous stochastic process).
If the initial state of the system is Gaussian, and if the Hamiltonian is quadratic, then
time evolution will preserve the Gaussian character of the state. We will assume, as is the

case for linear cosmological perturbations and gravitational waves, that the state is Gaussian
at all times. Therefore, the probability distribution P (ϕ(x), π(x)) can be expressed in terms
of the above two-point correlation functions. Hence, also the entropy of the system must be
expressible in terms of two point correlation functions.

In the following, we will derive a general expression for the entropy of a stochastic
Gaussian field in terms of its two-point correlation functions.
Starting point of the analysis is formula (5) for the entropy in terms of the probability
distribution. In our case, the continuous variable J stands for a point in phase space. To

justify the choice of the measure in (5), we divide phase space (ϕ, π) at every point x in
space into units of volume 2πh̄ (the smallest volume the fields can be localized in by the
uncertainty principle) and calculate the probability ∆P that the fields lie in the bin ∆J :
Z
∆PJ = P (ϕ(x), π(x))Dϕ(x)Dπ(x) . (21)
∆J

Here, the integration ranges over fields ϕ(x) and π(x) which lie in bin ∆J and Dϕ(x) Dπ(x)
denotes the functional integral measure for a scalar field.
Thus, from the analysis of Section 2 we conclude that (apart from an irrelevant constant),
the entropy of the stochastic classical field is given by
Z
S=− P (ϕ(x), π(x)) ln P (ϕ(x), π(x))Dϕ(x)Dπ(x) . (22)

For a Gaussian state, the probability distribution is

1 1
 
P (ϕ, π) = exp − (ϕx Axy ϕy + πx B xy πy + 2ϕx C xy πy ) (23)
W 2

where the normalization factor W , determined by

Z
P (ϕ, π)DϕDπ = 1 , (24)

13
is given by

1
Z  
W = exp − (ϕx Axy ϕy + πx B xy πy + 2ϕx C xy πy ) DϕDπ . (25)
2

Here, we have used the short hand notation ϕ(x) = ϕx and Axy = A(x, y), and the Einstein
“summation” convention for repeated indices
Z
xy
ρx A = dxρ(x)A(x, y) (26)

is implied. For a homogeneous Gaussian state, the kernels Axy , B xy and C xy depend only

on the difference of the arguments

Axy = A(x, y) = A(x − y) = A(y − x) . (27)

The kernel of the operator A−1 inverse to A will be denoted by A−1

xy (defined as usual by

A−1 zy
d3 zA−1 (x, z)A(z, y) = δ 3 (x − y)). Note that the kernel C does not in general
R
xz A =

vanish since in general ϕ and π are not statistically independent. A nonvanishing C reflects
a correlation between ϕ and π.
Our goal now is to express the kernels Axy , B xy and C xy in terms of the two point corre-
D E D E D E
lation functions ϕ(x)ϕ(y) , π(x)π(y) and ϕ(x)π(y) . As a first step, it is convenient to

rewrite the distribution function P (ϕ, π) in terms of new variables which are independent

ζx = ϕx + C zy A−1
yx πz

πx = πx (28)

The Jacobean of this transformation is 1 and hence

DϕDπ = DζDπ . (29)

The probability distribution (23) in terms of the new variables ζ and π is

1 1
 
P (ζ, π) = exp − (ζx Axy ζy + πx Γxy πy ) , (30)
W 2

where

14
 Γxy = B xy − C xu A−1
uv C
vy
. (31)

At this point, the correlation functions can be immediately expressed in terms of the kernels

hζx ζy i = A−1
xy

hπx πy i = Γ−1
xy

hζx πy i = 0 . (32)

This can be seen either by direct functional integration, or by calculating the generating
functional (characteristic functional)
Z  
ζ π
Φ(J , J ) = P (ζ, π) exp −iJxζ ζ x − iJxπ π x DζDπ (33)

and taking second derivatives of it, e.g.,

δ2Φ
hζx ζy i >= − |J ζ =J π =0 . (34)
δJxζ δJyζ

Substituting ζ from (28) into (32) and solving the resulting set of equations we obtain

−1
A−1
xy = hϕx ϕy i − hϕx πu ihπu πv i hπv ϕy i

−1
Bxy = hπx πy i − hπx ϕu ihϕu ϕv i−1 hϕv πy i
−1
Cxy = hϕx πy i − hπx πu ihϕu πv i−1 hϕv ϕy i . (35)

Deriving the above formulas we took into account that for a spatially homogeneous Gaussian

process the correlation functions depend only on the difference of the arguments.
To calculate the entropy, we substitute (30) into the general formula (22) using the fact
that the Jacobean of the transformation (28) is unity and obtain
Z
S=− P (ζ, π) ln P (ζ, π)DζDπ

= V δ 3 (0) + ln W
1
Z  
3
= V δ (0) + ln exp − (ζx Axy ζy + πx Γxy πxy ) DζDπ
2
1
= V δ 3 (0) + ln det(A−1 Γ−1 ) (36)
2

15
where V is the volume of space. Dropping the irrelevant constant contributions to the
entropy and inserting (32) and (35) in the above, we get the following expression for the
entropy in terms of correlation functions

1 1
S= ln det (hϕx ϕz ihπz πy i − hϕx πz ihπz ϕy i) = ln det D xy (37)
2 2

Thus, the problem of calculating the entropy has been reduced to the evaluation of a
determinant of the operator
Z D E D E
xy
D = D(x − y) = (hϕ(x)ϕ(z)i π(z)π(y) − hϕ(x)π(z)i π(z)ϕ(y) )d3 z . (38)

This determinant can be calculated by ζ function regularization (see Appendix A). The
result is
 Z 
3
det D ∝ exp V d k ln Dk (39)

where

1
Z
Dk = d3 xe−ik·x D(x) (40)
(2π)3

is the spectral density of the operator D(x) = D(x − y). Substituting (39) in (37) we obtain
the following expression for the entropy per unit volume

S
Z
s= = d3 k ln Dk
V Z D ED E D E
= d3 k ln |ϕk |2 |πk |2 − |ϕk |2 |πk |2 . (41)

where we expressed the entropy in terms of the spectral density of correlation functions and
omitted irrelevant contributions which do not depend on the spectrum of the scalar field.

V. GENERAL COMMENTS

In this section we shall focus on two issues: the connection between the two definitions
of entropy given in Sections 3 and 4, and a further discussion of the origin of entropy.

16
 To concretize the analysis we consider a quantized free scalar field ϕ̂(x, η) with time
dependent effective mass. The field is assumed to start out in its initial vacuum state. As
is well known, this state evolves into a squeezed state [19,20]. Such a state is highly excited

in the sense that the expectation value of the number operator is large.
If the mass is constant at the beginning and at the end, then the notion of particles is
well defined for the in-state and for the out-state, and in the corresponding time intervals the
field operator ϕ̂ can be expanded either in terms of in-creation and annihilation operators

â+ and â− or in terms of the corresponding out-operators ĉ+ and ĉ− :

d3 k
Z  
ϕ̂(x, η) = eik·x uin ∗ −
k (η) âk + e
−ik·x in
uk (η)â+
k
(2π)3/2
Z
d3 k  
= eik·x uout ∗ −
k (η) ĉk + e
−ik·x out
uk (η)ĉ+
k (42)
(2π)3/2

where uin out

k (η) and uk (η) are the positive frequency mode functions in the in- and out-states

respectively. For the particular case under consideration, the operators ĉ+ −
k and ĉk are related

to â+ −
k and âk via a Bogoliubov transformation

c− − − 2iϕk
k = ak chrk − a−k e shrk

c+ +
k = −a−k e
−2iϕk
shrk + a+
k chrk . (43)

The real functions rk (η) and ϕk (η) are called squeeze parameter and squeeze angle respec-
tively [21,22]. The initial state is taken to be the vacuum |0 >in defined by â−
k |0 >in = 0 for

all k.
The expectation value of the number of particles at late times in the k mode is

< nk >=< 0in |c+ − 2

k ck |0in >= sh rk . (44)

Hence, for large values of the squeeze parameter rk , we should expect that the quantum field

ϕ̂ can with good accuracy be described as a classical field ϕcl , since if rk ≫ 1, the condition
< nk >≫ 1 to be in the region of applicability of the classical limit is satisfied. In this limit,
the correlation functions of the classical field ϕcl should coincide with the corresponding
expectation values of the operator ϕ̂, e.g.,

17
 < ϕcl (x, η)ϕcl (y, η) >≃< 0in |ϕ̂(x, η)ϕ̂(y, η)|0in >≡< ϕ̂(x, η)ϕ̂(y, η) > . (45)

Hence, in the classical limit both of our formulas for the entropy should be applicable and
give the same result. In the following we show that this is indeed true.
First of all, we need to calculate the correlation functions of ϕ̂ in terms of the squeeze

parameters. Taking into account that at late times (in the out-state)
q
uout
k (η) ∼ e
iωk η
, ωk = k 2 + m2out (46)

where mout is the mass in the out-phase, and using (43), we find the following expressions

for the late time correlation functions [23]

d3 k ik·(x−y) 1
Z
< ϕ(x, η)ϕ(y, η) > = e [2sh2 rk + 1 − sh2rk cos 2δk ]
(2π)3 2ωk
d3 k ik·(x−y) ωk
Z
< π(x, η)π(y, η) > = e [2sh2 rk + 1 + sh2rk cos 2δk ]
(2π)3 2
d3 k ik·(x−y) i
Z
< ϕ(x, η)π(y, η) > = e [1 − ish2rk sin 2δk ] , (47)
(2π)3 2

where the angle δk is

Z
δk (η) = dη(ωk (η) − ϕk ) . (48)

Hence, it follows immediately that

Z
d3 z < ϕ(x)ϕ(z) >< π(z)π(y) >
d3 k ik·(x−y) 1
Z
= e [(2sh2 rk + 1)2 − sh2 2rk cos2 2δk ] (49)
(2π)3 4

and
Z
d3 z < ϕ(x)π(z) >< π(z)ϕ(y) >
d3 k ik·(x−y) 1
Z
= e [(2sh2 rk + 1)2 − sh2 2rk cos2 2δk ] (50)
(2π)3 4

Taking into account (45) we can substitute (49) and (50) into (38) to determine the
entropy of the field in the classical limit. The two terms (49) and (50) evidently cancel

18
each other, giving vanishing entropy. But this is no surprise since we started in a pure state
whose entropy must vanish and since the evolution of the system is unitary, thus preserving
the entropy. The information about the final state is complete.

To associate entropy with the final state we must neglect some information. Typically,
this will be information which is very sensitive to any kind of perturbation, either of the
system or of the state. In our example, the phases δk will depend sensitively on a pertur-
bation, whereas the amplitudes will not. In the language of Section 3, the amplitude is the

principal quantum number and the phases are averaged over.

The nature of the perturbation mentioned above is a subject of independent analysis.
There are several mechanisms; which one is realized will depend on the particular system
under investigation. Interactions with other fields (which can be viewed as an environment)

will induce stochasticity of the phases. Weak self interactions of the scalar field may also
induce large changes in phases. Approximating the state of the system (e.g., neglecting
decaying modes in examples in which the mass is constant at early and late times but
changes in between) will have the same effect.

Returning to our example, let us define the coarse grained entropy by at first averaging
the two point correlation functions (47) over the phases δk and substituting the thus obtained
“reduced” correlation functions into (38) to find the reduced operator Dred which will be
d3 k
Z
D(x − y) = 3
sh2 rk (1 + sh2 rk )eik·(x−y) . (51)
(2π)
Then, according to (41), the classical field definition of entropy gives
Z
s∼
= d3 k ln sh2 rk (52)

in the classical limit sh2 rk ≫ 1. Taking into account (44) we obtain

Z
s≃ d3 k ln nk (53)

for nk ≫ 1, in agreement with the result for the entropy obtained in Section 3 (see (19)).
We conclude that in the cases in which both of our formulas for the entropy from Sections
3 and 4 are applicable, the results coincide as they should.

19
 VI. COSMOLOGICAL PERTURBATIONS

Our goal is to apply the definition of nonequilibrium entropy developed in the previous

sections to the gravitational field. Specifically, we will calculate the entropy of a stochastic
background of gravitational waves and of linearized density perturbations. To set the stage,
we briefly review the theory of linearized cosmological perturbations [24].
We consider linearized perturbations of metric and matter fields about a homogeneous

and isotropic background cosmological model. There are three types of fluctuations—scalar,
vector and tensor perturbations—which are distinguished by their transformation properties
under background space coordinate changes. Vector modes decay and are irrelevant for
cosmology. Hence, we shall focus on scalar modes (density perturbations) which couple to
density and pressure and tensor modes (gravitational waves).

Although it may at a first glance seem that both density perturbations and gravitational
waves are described by several independent fields, the analysis can in both cases be reduced
to the theory of a single scalar field [15,16]; and hence the formalism developed in previous
sections to determine the entropy of a nonequilibrium dynamical system becomes applicable.

A. Gravitational Waves

Gravitational waves are linearized purely gravitational fluctuations about a homogeneous

(0)
and isotropic background metric gµν . For simplicity we will consider a spatially flat Universe
given by the invariant line element

ds2 = a2 (η)(dη 2 − δij dxi dxj ) , (54)

η being conformal time and a(η) denoting the scale factor. The metric perturbation δgµν =
a2 hµν for gravitational waves is transverse and traceless and satisfies h00 = h0i = 0.

The action Sgr for gravitational waves can be obtained by expanding the Einstein action

1
Z
√ 8πG
S=− R −gd4 x , l2 = (55)
6l2 3

20
to second order in the perturbation variables, with the result

1
Z h i
′ ′
Sgr = d4 xa2 hik hki − hik,e hki,e , (56)
24l2

where the derivative with respect to η is denoted by a prime. Varying this action yields the
following equation of motion for hij :

a′
hij ′′ + 2 hij ′ − ∆hij = 0 . (57)
a

In order to reduce the study of gravitational waves to the analysis of a single scalar field,
we expand hij in a Fourier series [15]

d3 k ik·x i
Z
hij (x, η) = e Gj (k)a−1 uk (η) , (58)
(2π)3

where Gij (k) is the polarization tensor of a gravitational wave with wave vector k. The mode
functions uk (η) can be used to define a scalar field ϕ

1 1/2 d3 k ik·x  i 1/2

Z
j
ϕ(x, η) = ( ) e G j (k)G i (k) uk (η) (59)
12ℓ2 (2π)3

in terms of which the action (56) becomes

a′′
" #
1
Z
′2
Sgr = d4 x ϕ − ϕ,i ϕ,i + ϕ2 . (60)
2 a

This action coincides with the action of a free scalar field with time dependent mass m(m2 =

−a′′ /a) in flat space-time.

Based on the above equivalence between the gravitational radiation field and the scalar
field ϕ(x, η), we can use the general definitions of entropy developed in Sections 3 and 4 to
define the entropy in gravitational radiation. This will be done in Section 7.

B. Density Perturbations

The theory of linearized density perturbations can also be reduced to the analysis of a
single scalar field [25]. However, the reduction process is more complicated than in the case
of gravitational waves.

21
 Density perturbations are scalar type metric fluctuations which couple to energy density
and pressure. At first sight, the most general scalar type metric perturbation δgµν (x, η) can
be expressed in terms of four free functions. However, two of these functions describe pure

gauge modes [26], i.e., inhomogeneities which correspond to a change of the background
space-time coordinates. The easiest and most physical way to avoid gauge artifacts is to
adopt a manifestly gauge invariant formalism [26] (for a pedagogical introduction see Ref. 16,
for a comprehensive review see Ref. 15). In this approach, scalar type metric perturbations

are characterized by two functions which are, via the linearized Einstein equations, coupled
to matter inhomogeneities. The two gauge invariant functions Φ(x, η) and Ψ(x, η) can easily
be identified by transforming to longitudinal gauge (the system of coordinates in which δgµν
is diagonal):
 
2Φ 0
δgµν = a2   (61)
0 2Ψδij
For scalar matter and for an ideal gas, Φ = Ψ as a consequence of the i 6= j Einstein equation
(for these forms of matter δTij ∼ δij ).
At this point, the description of density perturbations has been reduced to prescribing
a gauge invariant combination v of matter perturbation and metric fluctuations in terms of

which the action for perturbations can be expressed in the form [27]:

z ′′
" #
1
Z
4 ′2
Sv = dx v − c2s v,i v,j δ ij + v2 (62)
2 z

where c2s = 1 for scalar field matter and c2s = p/ρ for ideal gas matter. The variable z(η) is

a combination of background dependent factors. For a perfect fluid and for a scalar field as
matter

aβ 1/2
z= , β = H2 − H′ , (63)
Hcs

where H = a′ /a.
All gauge invariant variables (like Φ) can be expressed in terms of the variable v via the
Einstein equations. For example, the metric potential Φ which characterizes the amplitude

22
of metric perturbations in longitudinal gauge (see (61)) is expressed in terms of v in the
following manner
s  ′
3 β v
∆Φ = − ℓ (64)
2 Hc2s z

(valid for both scalar field and hydrodynamical matter).

In conclusion, we have been able to reduce the action for scalar metric perturbations
to that of a free scalar field with mass m2 = −z ′′ /z. However, in contrast to the case of

gravitational perturbations, the spatial gradient term has a prefactor c2s . In the radiation
dominated period of the evolution of the Universe this term implies that density perturba-
√
tions propagate with the speed of sound cs = 1/ 3. In the matter dominated period, cs = 0.
This has important consequences for our ability to define the notion of “particle number”

of a given field configuration for different equations state.

VII. ENTROPY OF COSMOLOGICAL PERTURBATIONS

In the previous section we demonstrated that the action for both gravitational waves
and for density perturbations can be reduced to the action of a free scalar field. Hence

the formalism of sections 3 and 4 can be used to define the nonequilibrium entropy of the
gravitational field.
In order to use the quantum definition of entropy of Section 3 we need a well defined
notion of particles. For gravitational waves described by the action (60) and in the special

case when a′′ = 0 (which is realized in the radiation dominated period when a(η) ∼ η) the
mass term in (60) vanishes, and hence the notion of particles is well defined and the particle
number remains constant. In contrast, during an inflationary period a(η) = −1/(Hη), and
′′
this leads to a negative time dependent effective square mass m2 = − aa in (60) which is

2
m2 = − (65)
η2

Because of this time dependent mass, there will be in this case particle production and

23
corresponding increase in entropy. Focusing on modes with comoving wavenumber k, the
time dependent mass induces an effective potential
!
1 2 2
Vef f,k (ϕ) = k − 2 ϕ2 . (66)
2 η

Hence, the increase in particle number is significant on scales which satisfy

√
2 √
k<− = 2Ha, (67)
η
√
which (up to a factor of 2) is exactly the condition for the wavelength k −1 a to be larger
than the Hubble radius H −1 . Similarly, in the matter dominated period when a(η) ∼ η 2
there is a nonvanishing effective potential. In terms of η the potential takes precisely the
form (65). Hence, (66) and (67) also apply in this case.

In the following, we shall evaluate the entropy for the spectrum of gravitational waves pro-
duced in an inflationary Universe. From the previous discussion, we see that the magnitude
of the entropy is generated by “parametric amplification” [28] of the number of gravitons in
each mode which occurs during inflation on scales larger than the Hubble radius. On scales

which enter the Hubble radius after teq , the time of equal matter and radiation, an additional
amplification takes place during the matter dominated epoch. However, we emphasize that
although the magnitude of the “potential” entropy is set by the inflationary phase, in order
to get any nonvanishing entropy there needs to be a loss of correlations due to some coarse
graining (see Section 5).

The situation for density perturbations is similar. During the period of radiation
domination, it follows from (6.12) that z ′′ = 0. Hence, as in the case of gravitational
waves, the notion of particles is well defined, the particle number is time independent, and
hence the entropy per mode remains constant. During inflation, it follows from (63) that

z ′′ /z ≃ a′′ /a = 2/η 2 , and that therefore like for gravitational waves particle creation for
modes with wavelengths larger than the sound horizon H −1 occurs, which leads to an in-
crease in the occupation number of each mode. The analysis of density perturbations is
more complex in the matter dominated period since c2s = 0. Hence no occupation number

24
can be defined. In this case we must use the classical framework of Section 4 in order to
define the entropy.
In the following, we will first calculate the entropy per mode of gravitational waves and

density perturbations which were produced during inflation in a cosmological model in which
the Universe is radiation dominated at late times and argue that the result for the more
realistic case is the same.

A. Gravitational Radiation

To estimate the entropy of gravitational radiation produced during inflation, it is con-

venient to express the expectation value < nk > of the number operator in each mode k
(or, correspondingly, the squeeze parameter rk ) in terms of the spectral density of the two
point correlation function of the metric hij . Since this spectrum |δ h (k)| was calculated in

the particular models we are interested in [29], we can then use these results to estimate the
entropy via formulas (52) or (53) (the result will be the same since on the scales of interest
nk ≫ 1).
We can only justify the appearance of entropy for perturbations on scales smaller than

the Hubble radius, since the mode functions do not oscillate on larger scales. Hence, when
calculating the entropy in gravitational radiation at some late time, we will only consider
scales smaller than the Hubble radius at that time.
Combining (58) and (59) with the standard mode expansion of the quantum operator ϕ̂

associated with ϕ, we obtain the following result for the operator ĥij :
Z
d3 k −1/2 Gij (k) h i
ĥij (x, η) 2 1/2 −1
= (6l ) a k e−ik·x out
u k (η) ∗
c k + eik·x out
u k (η)c +
k
(2π)3/2 (Gm n
n (k)Gm (k))
1/2

(68)

where c+ −
k and ck are creation and annihilation operators respectively of particles with co-

moving wave vector k at some late time and uout

k (η) ≃ e
ikη
for perturbations on scales smaller
than the Hubble radius (kη ≫ 1).

25
 The spectrum δh (k) of gravitational radiation is defined in terms of the two point function
of ĥij by

D E Z ∞ dk sin(kr)
0in |ĥij (x, η)ĥji (x + r, η)|0in = |δh (k)|2 (69)
0 k kr

with r = (r) and k = |k| and where |0in > is the initial vacuum state defined by a−
k |0in >= 0.

On the other hand, this two point function can be evaluated in terms of occupation numbers
using (68)

6l2 −2 ∞ dk sin kr 2
D E Z
ψ|ĥij (x, η)ĥji (x + r, η)|0 = a k (2 < nk > +1) (70)
2π 0 k kr

where we took into account that < 0in |ĉ+ −

k ĉk |0in >=< nk >. Comparing (69) and (70) yields

2π |δh (k)|2
2 < nk > +1 = 2 . (71)
6l (k/a)2

Equation (71) can be applied to calculate the entropy of gravitational radiation in any

cosmological model. Given the spectrum δh (k), the occupation numbers < nk > are deter-
mined for each mode by (71), and the entropy per mode follows from (53).
To be specific, we evaluate the entropy in an inflationary Universe. The spectrum of
gravitational radiation originating in quantum fluctuations during the phase of exponential

expansion has been calculated many times [29] (for an explicit derivation using the notation
of this paper see Ref. 15). The result is

(t0 k)−2 t−1 −1 1/2
0 < k < t0 zeq

ℓH


δh (k) = √ (72)
2π 
 (t0 k)−1 zeq
−1/2 1/2 −1
zeq t0 < k < t−1 −1/4 t0 −1/2
0 zeq ( tR ) ,


where t0 is the present time and tR corresponds to the end of inflation. zeq is the redshift
at the time of equal matter and radiation. The top line corresponds to scales which enter
the Hubble radius after teq , the bottom to those which enter during the period of radiation
domination. Hence, from Equation (71)

0 k)
−4
t−1
 (t −1 1/2
0 < k < t0 zeq

H2

2 < nk > +1 = (73)
6πk 2 
 (t0 k)−2 zeq
−1 1/2 −1
zeq t0 < k < t−1 −1/4 t0 1/2
0 zeq ( tR ) ,


26
H being the Hubble expansion rate during the period of inflation.
Choosing H = 1013 GeV , a value for which fluctuations from inflation have the right
order of magnitude to seed galaxies [30], and comparing the entropy sk per mode on galactic

scales k −1 ∼ 10Mpc, we conclude from (73) that

sk ∼ 100 ln 10 . (74)

Formula (74) can also be interpreted as giving the entropy in a volume V = k −3 of

gravitational waves with wave number of the order k. This quantity can be compared with
the statistical fluctuations of the entropy of the cosmic microwave background (CMB). These
fluctuations ∆sCMB (k) scale as

∆sCMB (k) ∼ (kλγ )3/2 , (75)

where λγ is the characteristic wavelength of the CMB. Hence, for large k −1 (such as the
scales mentioned above), the entropy in gravitational waves exceeds that in the statistical
fluctuations of the CMB. This reflects the result that on cosmological distance scales the

fluctuations produced by inflation are much larger than the Poisson noise in the background
fields.
However, the total entropy density for the gravitational background is smaller than that
of the CMB for which

4π 3
SCMB ≃ T . (76)
3 0

From (53) it follows that the entropy density in gravitational radiation is

4π 3
Sgr ≃ k (77)
3 c

where kc is determined by

n(kc ) ≃ 1 . (78)

From (73) it follows that

27
 t0 1/2
kc ∼ t−1 −1/4
0 zeq ( ) (79)
tR

and hence
!3/2
H
Sgw ∼ T03 (80)
mpl

Comparing (77) and (80) we see that the total entropy density in gravitational waves is
suppressed compared to the entropy of the CMB by a factor (H/mpl )3/2 . Nevertheless, as

seen above, on large length scales, gravitational radiation dominates over the fluctuations
of the entropy of the CMB.

B. Density Perturbations

The approach for calculating the entropy of density perturbations follows what was done
above for gravitational waves. Starting point is the expansion of the operator ϕ̂ associated
with the scalar field v of (62) [which contains all the information about density perturbations]
in terms of late time creation and annihilation operators c+ −
k and ck respectively:

1 d3 k h ik·x out
Z i
− −ik·x out +
ϕ̂(x, η) = √ e vk (η)c k + e vk (η)c k (81)
2 (2π)3/2

To simplify the consideration and to justify the notion of entropy, we will estimate the
entropy on scales smaller than the Hubble radius at late times, assuming that in this late
time interval the Universe is radiation dominated. In this case, the notion of particles is
unambiguous since the mode functions vk (η) take the form

1
vk (η) = √ eiωη (82)
ω
√
with ω = cs k = k/ 3.
Using the relation (64) between the gauge invariant relativistic potential Φ and the scalar
field v valid when c2s 6= 0 we obtain [15]

3 1/2 β 1/2 d3 k h ik·x ∗

Z i
− −ik·x +
Φ̂(x, η) = ( ) l e u k (η)c k + e u k (η)c k (83)
4 a (2π)3/2

28
with

z vk ′
uk = ( ) . (84)
k2 c s z

As was done previously in the case of gravitational waves, the next step is to derive the
relation between the spectrum δΦ (k) of density perturbations

∞ dk sin kr
D E Z
0in |Φ̂(0, η)Φ̂(r, η)|0in = |δΦ (k)|2 (85)
0 k kr

and the occupation number < nk > defined with respect to the creation and annihilation
operators introduced in (81). Using (83) it follows that
D E 3 2 β Z d3 k
0in |Φ̂(0, η)Φ̂(r, η)|0in = l 2 (2 < nk > +1)u∗k uk eik·r . (86)
4 a (2π)3

Making use of (82) and (84), and evaluating z during the period of radiation domination,
we obtain
√
9 3 2 −4 −2 ∞ −1 sin kr
D E Z
0in |Φ̂(0, η)Φ̂(r, η)|0in = l η a dkk (87)
(2π)2 0 kr
(kη)2
!
−2
× (2 < nk > +1)k 1+ .
3

Comparing (85) and (87) yields

!−1
(2π)2 (kη)2
2 < nk > +1 = √ (kη)2 (aη)2 l−2 1 + |δΦ (k)|2 , (88)
9 3 3

which expresses the occupation numbers in terms of the spectrum of the relativistic potential
for density perturbations.

The entropy for density perturbations can now be determined by combining (88) and
(53). To be specific, we evaluate the entropy per mode in a model of chaotic inflation [31]
with potential

1
V (ϕ) = m2 ϕ2 . (89)
2

In this case, the spectrum of density perturbations immediately after inflation on scales
which are larger than the Hubble radius is given by [15]

29
 √
2 λ
|δkΦ | ≃ lm ln( ) (90)
3π λγ

where λγ is the characteristic wavelength of the CMB. To obtain the late time spectrum, the
decay of the amplitude of Φ on scales inside the Hubble radius must be taken into account.

Assuming that the Universe is radiation dominated after inflation we obtain [15]
√ ! !2
Φ 2 λ λ
|δk | ≃ ln ln (91)
3π λγ t0

for kη > 1 at time t0 .

Inserting (91) into (88) yields the following result for kη ≫ 1:

! !4
2 2 λ λ
2 < nk > +1 ∼ (mt0 ) ln . (92)
λγ t0

Up to the logarithmic factor, this result agrees with the corresponding result (73) for gravi-

tational waves.
For m ∼ 1013 GeV (the upper bound on m from constraints on the anisotropy of the
CMB) we hence obtain the same entropy density per mode

sk ∼ 100 ln 10 . (93)

We conclude that the entropy per mode of density perturbations on large scales (λ ≫ λγ ),
in particular on scales of galaxies, clusters and large-scale structure, exceeds the statistical
fluctuations of the entropy of the CMB. This supports the important role of the entropy of
the gravitational field in the galaxy formation process.

The quantum approach of calculating the entropy of density perturbations breaks down
during the matter dominated era. However, based on the considerations of Section 5, it is
easy to extend the analysis.
As discussed in Section 4, the classical definition (41) can easily be applied during the

epoch of matter domination. The results obtained above will not change significantly. How-
ever, there are some interesting questions concerning coarse graining and time dependence
of the entropy in a matter dominated Universe which we shall consider elsewhere.

30
 Formulas (71) and (88) can be used to calculate the entropy of gravitational waves and
density perturbations in any cosmological model in which the spectra δh (k) and deltaΦ (k)
are known. In particular, the entropy of the gravitational field in topological defect models

of structure formation [14] can easily be determined.

VIII. CONCLUSIONS

We have presented two general definitions of nonequilibrium entropy and applied them

to calculate the entropy in gravitational radiation and to give a measure of the entropy of
linear density fluctuations and gravitational waves in a Friedmann Universe.
Our first definition of entropy is based on the microcanonical ensemble and is applicable
to systems with well defined occupation numbers. The origin of the entropy is coarse grain-
ing: ignoring correlations in the form of information about quantum numbers other than

the principal quantum number which is usually taken to be the energy.

The second definition of entropy applies to any stochastic classical field and expresses
the entropy in terms of two point correlation functions. The physical origin of entropy in
this case is also due to coarse graining. We have shown that the entropy of scalar fields in

an expanding Universe satisfies the second law of thermodynamics. Any increase in entropy
during Hamiltonian evolution is a consequence of coarse graining.
We have used the gauge invariant theory of cosmological perturbations to give a consis-
tent and unified definition of entropy of cosmological perturbations. On scales of galaxies

and larger, this entropy is larger than the statistical fluctuations in the entropy of CMB
photons on these scales. Hence, this entropy is important for structure formation in the
Universe. However, the total entropy in density perturbations and in gravitational waves is
smaller than the total entropy of the CMB.
It is our hope that the methods presented here can be used in many different situations.

In particular, they might allow a definition of entropy of density perturbations beyond linear
theory.

31
 ACKNOWLEDGEMENTS
For stimulating discussions we thank Andy Albrecht, Leonid Grishchuk, Jim Hartle, Tony
Houghton, Bei-Lok Hu and Henry Kandrup. Two of us (R.B. and T.P.) thank the ITP of

the University of California in Santa Barbara for hospitality during the completion of this
work. V.M. thanks the Swiss National Science Foundation for financial support. At Brown,
this research was supported by DOE grant DE-AC02-76ER03130 Task A and by an Alfred
P. Sloan Foundation Fellowship to R.B.. Financial support from NSF grant PHY89-04035

at the ITP in Santa Barbara is acknowledged.

Calculation of the Determinant of D(x − y) ≈ D xy

The determinant of the operator D(x − y) ≈ D xy arising in (38) can be calculated using
the ζ-function method [32]. More generally, consider an operator D xy with a discrete set of
positive real eigenvalues ai and eigenfunctions

f (i) (x) ≈ fx(i)

Z
D xy fy(i) = d3 yD(x − y)f (i) (y) = ai δ xy fy(i) . (94)

This set of operators includes any operator with positive spectral density
Z
D(k) = d3 ze−ik·z D(z) (95)

which has support in a finite volume V , i.e., D(z) = 0 if z ∈

/ V.
The ζ-function associated with such an operator is

X 1
ζD (s) = , (96)
i asi

where the sum extends over all nonzero eigenvalues. It follows that

dζD (s)
ζD′ (0) = ln ai e−sai |s=0 = − ln (Πi ai )
X
|x=0 = − (97)
ds i

and hence

32
 ′
det D = Πi ai = e−ζD (0) . (98)

Thus, the calculation of the determinant of D has been reduced to the evaluation of ζD′ (0).
Let us now introduce the heat kernal G(x, y, τ ) associated with D,

∗
e−ai τ fx(i) fy(i) ,
X
G(x, y, τ ) = (99)
i

which in the case under consideration depends only on x − y and τ and satisfies the equation
Z
∂G(x − y, τ )
D(x − z)G(z − y)d3 z = − (100)
∂τ

with boundary condition

G(x − y, 0) = δ(x − y) , (101)

a consequence of the completeness of the set of eigenfunctions. It is easy to check by explicit

integration that

1 ∞
Z Z
s−1
ζD (s) = dτ τ d3 xG(x, x, τ ) . (102)
Γ(s) 0

Thus, in order to compute det D we need to solve Eq. (A7) for G(x, y, τ ) given the

boundary condition (A8). It is convenient to work in Fourier space where Eq. (A7) takes
the form

∂
D(k)G(k, τ ) = − (G(k, τ )) (103)
∂τ

with boundary condition

G(k, 0) = 1 , (104)

where
Z
G(k, τ ) = G(z, τ )eik·z d3 z . (105)

As a result we obtain the following solution of (A7):

33
 Z
G(x − y, τ ) = d3 keD(k)τ eik(x−y) . (106)

Substituting (A13) into (A9) on finds

Z
ζD (s) = V d3 ke−s ln D(k) (107)

and correspondingly
 Z 
3
det D = exp V ln D(k)d k (108)

where V is the volume of the support of D.

34
 REFERENCES

∗
On leave of absence from Institute for Nuclear Research, Academy of Sciences, Moscow,
Russia.

[1] J. Bekenstein, Phys. Rev. D7, 2333 (1973).

[2] S. Hawking, Comm. Math. Phys. 43, 199 (1975).

[3] R. Penrose, in ‘General Relativity: An Einstein Centenary Survey’ ed. by S. Hawking

and W. Israel (Cambridge Univ. Press, Cambridge, 1979).

[4] C. Shannon, Bell System Tech. J. 27, 379, 623 (1948);

E. Jaynes, Phys. Rev. 106, 620 (1957).

[5] L. Smolin, Gen. Rel. Grav. 17, 417 (1985).

[6] B. Hu and D. Pavon, Phys. Lett. 180B, 329 (1986).

[7] H. Kandrup, J. Math. Phys. 28, 1398 (1987).

[8] H. Kandrup, Phys. Rev. D37, 3505 (1988).

[9] B. Hu and H. Kandrup, Phys. Rev. D35, 1776 (1987).

[10] S. Habib and H. Kandrup, Ann. Phys. 191, 335 (1989).

[11] B. Hu, Phys. Lett. 97A, 368 (1983).

[12] K. Huang, ‘Statistical Mechanics’ (Wiley, New York, 1963);

R. Kubo, ‘Statistical Mechanics’ (Elsevier, Amsterdam, 1965);
L. Landau and E. Lifshitz, ‘Statistical Physics’ (Pergamon Press, London, 1959).

[13] A. Guth, Phys. Rev. D23, 347 (1981).

[14] T.W.B. Kibble, J. Phys. A9, 1387 (1976);

A. Vilenkin, Phys. Rep. 121, 263 (1985).

35
[15] V. Mukhanov, H. Feldman and R. Brandenberger, Phys. Rep. 215, 203 (1992).

[16] R. Brandenberger, H. Feldman and V. Mukhanov, ‘Gauge Invariant Cosmological Per-

turbations’, Brown preprint BROWN-HET-845 (1992), to be published in the proceed-
ings of ICGC-91 (Wiley Eastern Ltd. , New Delhi, 1992).

[17] V. Mukhanov, Pisma Zh. Eksp. Fiz. 44, 50 (1986).

[18] E. Calzetta and B. Hu, Phys. Rev. D37, 2878 (1988).

[19] L. Grishchuk and Yu. Sidorov, Class. Quant. Grav. 6, L161 (1989);
L. Grishchuk and Yu. Sidorov, Phys. Rev. D42, 3413 (1990).

[20] L. Grishchuk, ‘Quantum mechanics of the primordial cosmological perturbations’, to be

published in the proceedings of the 6th Marcel Grossmann meeting, Kyoto, 1991.

[21] C. Caves, et al. Rev. Mod. Phys. 52, 341 (1980).

[22] B. Schumaker, Phys. Rep. 135, 317 (1986).

[23] T. Prokopec, Brown Univ. preprint BROWN-HET-861 (1992).

[24] see e.g., S. Weinberg, ‘Gravitation and Cosmology’ (Wiley, New York, 1972);
P.J.E. Peebles, ‘The Large-Scale Structure of the Universe’ (Princeton Univ. Press,
Princeton, 1980);
Ya. B. Zel’dovich and I. Novikov, ‘The Structure and Evolution of the Universe’ (Chicago

Univ. Press, Chicago, 1983).

[25] G. Chibisov and V. Mukhanov, ‘Theory of Relativistic Potential: Cosmological Pertur-

bations’, preprint No. 154 of P.N. Lebedev Physical Institute (1983).

[26] U. Gerlach and U. Sengupta, Phys. Rev. D18, 1789 (1978);

J. Bardeen, Phys. Rev. D22, 1882 (1980).

[27] V. Mukhanov, Zh. Eksp. Teor. Fiz. 94, 1 (1988).

36
[28] L. Grishchuk, Zh. Eksp. Teor. Fiz. 67, 825 (1974).

[29] A. Starobinsky, Pisma Zh. Eksp. Teor. Fiz. 30, 719 (1979);
V. Rubakov, M. Sazhin and A. Veryaskin, Phys. Lett. 115B, 189 (1982);
V. Mukhanov, PhD thesis, Lebedev Physical Institute (1982). R. Fabbri and M. Pollock,
Phys. Lett. 125B, 445 (1983);

L. Abbott and M. Wise, Nucl. Phys. B244, 541 (1984);

L. Abbott and D. Harari, Nucl. Phys. B264, 487 (1986);
B. Allen, Phys. Rev. D37, 2078 (1988);
V. Sahni, Phys. Rev. D42, 453 (1990).

[30] J. Bardeen, P. Steinhardt and M. Turner, Phys. Rev. D28, 679 (1983);
R. Brandenberger and R. Kahn, Phys. Rev. D29, 2172 (1984);
V. Mukhanov, Pisma Zh. Teor. Fiz. 41, 402 (1985);

D. Lyth, Phys. Rev. D31, 1792 (1985).

[31] A. Linde, Phys. Lett. 129B, 177 (1983).

[32] see e.g. P. Ramond, ‘Field Theory: A Modern Primer’ (Benjamin, Reading, 1981).

37