Probability Measure in Multiverse Cosmology

Sergei Winitzki
Habilitationsschrift
des Departments f ur Physik, Ludwig-Maximilians Universitat, Theresienstr.
37, 80333 Munich, f ur das Fach Physik, vorgelegt November 6, 2008
von Dr. Sergei Winitzki
Abstract
In generic models of cosmological ination, the geometry of spacetime is highly inhomogeneous
on scales of many Hubble sizes, consisting of innitely many causally disconnected pocket uni-
verses. The values of cosmological observables and even of the low-energy coupling constants
and particle masses may vary among the pocket universes. String-theoretic landscape models
present a similar structure of a multiverse where an innite number of de Sitter, asymptoti-
cally at (Minkowski), and anti-de Sitter pocket universes are nucleated via quantum tunneling.
Since observers on Earth have no information about their location within the eternally inating
multiverse, the main question in this context has been that of obtaining statistical predictions
for quantities observed at a random location.
I discuss the long-standing technical and conceptual problems arising within this statistical
framework, known collectively as the measure problem in multiverse cosmology. After review-
ing various existing approaches and mathematical techniques developed in the past two decades
for studying these issues, I describe a new proposal for a measure in the multiverse, called the
reheating-volume (RV) measure. The RV measure is based on approximating an innite multi-
verse by a family of progressively larger but nite multiverses. Such multiverses occur seldom
but are allowed by all cosmological multiverse models. I give a detailed description of the new
measure and its applications to generic models of eternal ination of random-walk type and to
landscape scenarios. The RV prescription is formulated dierently for scenarios with eternal
ination of the random walk type and for landscape scenarios. In each case I show in a math-
ematically rigorous manner that the RV measure yields well-dened results that are invariant
with respect to general coordinate transformations, independent of the initial conditions at the
beginning of ination, and free of the youngness paradox and the Boltzmann brain problems
aecting some of the previously proposed measures.
For models of random-walk ination, the RV cuto considers events where one has a nite (al-
though large) total reheating volume to the future of an initial Hubble patch. I derive a general
formula for RV-regulated probability distributions that is suitable for numerical computations.
Explicit analytic computations are presented in a toy model having an eective potential with
an exactly at top.
For landscape scenarios, I propose to calculate the distribution of observable quantities in a
landscape that is conditioned in probability to nucleate a nite total number of bubbles to the
future of an initial bubble. A general formula for the relative number of bubbles of dierent
types can be derived. I show that the RV measure yields results independent of the choice of the
initial bubble type, as long as that type supports further bubble nucleation. As an illustration,
I present explicit results for a toy landscape containing four vacuum states and for landscapes
with a single high-energy vacuum and a large number of low-energy vacua.
This dissertation is submitted in partial fulllment of requirements for the degree of Dr. habil. in
physics. I certify that I have prepared this text through my own eort and no other sources or
quotations except those listed and attributed.
Selbststandigkeitserklarung
Hiermit erklare ich, dass ich die vorliegende Arbeit in allen Teilen selbstandig verfasst und keine
anderen als die angegebenen Quellen und Hilfsmittel (einschlielich elektronischer Medien und
Online-Quellen) benutzt habe.
Sergei Winitzki
Contents
Preface iii
1 Introduction 1
2 Eternal ination 5
2.1 Cosmological ination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 The inationary paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.2 Inationary models: slow roll . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Eternal ination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 Predictions in eternal ination . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.2 Physical justications of the semiclassical picture . . . . . . . . . . . . . . 15
3 Stochastic approach to ination 17
3.1 Random walk-type eternal ination . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.1.1 Fokker-Planck equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.2 Methods of solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.1.3 Gauge dependence issues . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Self-reproduction of tunneling type . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4 Predictions and measure issues 29
4.1 Presence of eternal ination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Observer-based measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3 Regularization for a single reheating surface . . . . . . . . . . . . . . . . . . . . . 33
4.4 Regularization for multiple types of reheating surfaces . . . . . . . . . . . . . . . 35
4.5 The Youngness paradox and the Boltzmann brains . . . . . . . . . . . . . . . . 37
5 A new measure for multiverse cosmology 39
5.1 Reheating-volume cuto . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2 RV cuto in slow-roll ination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6 The RV measure for random-walk ination 47
6.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.2 Overview of the results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2.2 Probability of nite ination . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.2.3 Finitely produced volume . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2.4 Asymptotics of (1;
0
) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2.5 Distribution of a uctuating eld . . . . . . . . . . . . . . . . . . . . . . . 55
6.2.6 Toy model of ination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.3 Derivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
i
Contents
6.3.1 Positive solutions of nonlinear equations . . . . . . . . . . . . . . . . . . . 56
6.3.2 Nonlinear Fokker-Planck equations . . . . . . . . . . . . . . . . . . . . . . 58
6.3.3 Singularities of g(z) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.3.4 FPRV distribution of a eld Q . . . . . . . . . . . . . . . . . . . . . . . . 65
6.3.5 Calculations for an inationary model . . . . . . . . . . . . . . . . . . . . 68
7 The RV measure for the landscape 75
7.1 Regulating the number of terminal bubbles . . . . . . . . . . . . . . . . . . . . . 77
7.2 Regulating the total number of bubbles . . . . . . . . . . . . . . . . . . . . . . . 82
7.3 A toy landscape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
7.3.1 Bubble abundances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
7.3.2 Boltzmann brains. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.4 A general landscape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.4.1 Bubble abundances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.4.2 Example landscape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.4.3 Boltzmann brains. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
7.4.4 Derivation of Eq. (7.103) . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.4.5 Eigenvalues of

M(z) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.4.6 The root of
0
(z) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8 Conclusion 111
Bibliography 113
ii
Preface
This dissertation is based on the authors original work on the so-called measure problem in
multiverse cosmology. The measure problem has resisted solution for more than two decades
since its original formulation in the 1980s, which was in the context of cosmological ination
driven by a scalar eld. At present, the renewed interest in the measure problem is due to the
discovery of the string-theoretic landscape, which allows a very large number of metastable vacua
with dierent physical laws. The need to extract predictions from these models has spurred a
urry of activity resulting in the appearance of several competing measure proposals and a
deeper understanding of the issues involved.
This dissertation gives an overview of the presently known approaches to the measure problem
and proceeds to describe the authors own measure proposal as developed in a series of recent
publications. Currently the measure problem in multiverse cosmology is an active area of re-
search, and more progress may be achieved in the near future. Rather than trying to anticipate
the forthcoming results, the author wishes to concentrate on the exposition of concepts and
mathematical methods that will remain useful for future research in this area.
Acknowledgments
The author is grateful to the Department of Physics at the Ludwig-Maximilians University in
Munich where the author has been employed for several productive years of research and to Prof.
V. F. Mukhanov for guidance and ample help. Numerous conversations with Andrei Barvinsky,
Martin Bucher, Cedric Deayet, Gabriel Lopes Cardoso, Jaume Garriga, Josef Ganer, Matthew
Johnson, Andrei Linde, Slava Mukhanov, Matthew Parry, Misao Sasaki, Takahiro Tanaka, Vitaly
Vanchurin, and Alexander Vilenkin, as well as with other colleagues have been valuable and
stimulated the authors thinking as well as corrected misconceptions on the authors part. Part
of this work was completed on a visit to the Yukawa Institute of Theoretical Physics (University
of Kyoto) and was supported by the Yukawa International Program for Quark-Hadron Sciences.
The entire text was typeset with the excellent L
Y
X and T
E
X document preparation system ex-
clusively on computers running Debian GNU/Linux. The author expresses profound gratitude
to the creators and maintainers of this outstanding free software.
Sergei Winitzki, October 2008
iii
1 Introduction
Cosmology is the branch of physics concerned with description of the Universe at large as
manifested by large-scale astronomical observations. Since the discovery of the expansion of
the Universe by Hubble and the development of the general relativistic model of non-stationary
homogeneous spacetime by Friedmann, the accepted point of view has been that the Universe
is expanding and has been in an extremely dierent state in the distant past. How the Universe
evolved to its present condition and how it began (if it had a beginning) are among the main
questions modern cosmology hopes to answer.
Although the goal of describing the entire Universe might seem to require a theory of ev-
erything, which perhaps will never be constructed but in any case is presently not available, it
turns out that many important cosmological observations could be explained in the framework
of currently available physical theories, in particular the known high-energy physics and classi-
cal gravity. G. Gamow introduced a model of the Universe expanding from an extremely hot
and dense state (hot Big Bang). In what concerns the evolution of the Universe after the Big
Bang, this scenario is in such a good agreement with observations that it is now considered to be
the standard cosmology. However, the standard cosmology leaves several important questions
unanswered. For instance, the initial hot state of the Universe turned out to be rather special,
and the standard cosmology failed to adequately explain its origin.
This is why one of the central problems of cosmology today is to describe the era before
the expansion described by the standard hot Big Bang scenario. Because of the present lack of
detailed knowledge of high-energy physics and quantum gravity, as well as of insucient precision
of available astrophysical observations, numerous cosmological models of the very early Universe
compete on more or less equal footing.
The currently popular and observationally well-supported models of the very early Universe are
inationary models. The scenario of ination assumes an epoch of an extremely fast expansion
(ination) of the Universe, followed by reheating to a hot thermal state. The latter becomes
the starting point of the standard hot Big Bang model.
It was found early on that in most of these models ination does not stop everywhere at the
same time. As a result, the universe at extremely large distance scales is divided into domains
with dramatically dierent properties: Some regions have already thermalized and developed
matter structures such as stars and galaxies, while other regions still undergo the inationary
expansion and are cold and empty. Therefore, a radical departure from homogeneity can be
expected at extremely large distance scales. Generically, at arbitrarily late times there exist
large domains that are still inating. This phenomenon was called eternal ination (more
precisely one can refer to future-eternal ination).
In some models, the various thermalized regions may also dier from each other in observable
cosmological parameters or even in values of the coupling constants observed in low-energy
physics. There are two problems with such models: First, a model that allows a wide range of
parameters to be observed in various regions of the Universe has little predictive power, since
we cannot determine which region we happen to inhabit. Second, the existence of a (typically
1
1 Introduction
innite) multitude of regions of space that are too far from us to be ever observable and/or the
assumption of the a priori unobservable many-universe ensemble are not directly comparable
with the experiment.
A similar situation is found in the recently discoveredlandscape of string theory. It was found
that string theory admits an exceedingly large number (of order 10
1000
) of disjoint, metastable
vacuum states. Dierent vacua have dierent values of the eective cosmological constant, cou-
pling constants of low-energy physics, and particle masses. Transitions between these states are
possible through bubble nucleation; the interior of each bubble appears to the interior observers
as an innite homogeneous open universe (if one disregards rare bubble collisions). For this
reason, the interior of a bubble has been called a pocket universe. The presently observed
universe is situated within a bubble where the vacuum state is, in some sense, randomly chosen.
Eternal ination occurs generically in this setting and produces an innite number of bubbles,
each containing a pocket universe in a particular vacuum state. This bewildering array of uni-
verses is currently referred to as a multiverse in order to stress the fact that dierent pocket
universes (as well as dierent regions within a single pocket universe) are causally disconnected
and are seen by observers as separate universes. It appears to be impossible to predict with
certainty the values of the cosmological parameters that we will measure in our present position
in the multiverse.
To overcome these problems, one may change focus and concentrate on obtaining the proba-
bility distribution of values for the measured cosmological observables, such as the cosmological
constant, coupling constants, or particle masses. Heuristically, one would like to compute prob-
ability distributions of the cosmological parameters as measured by an observer randomly lo-
cated in the spacetime. This idea, sometimes called the principle of mediocrity, has been rst
clearly formulated in the mid-1990s [1, 2, 3]. One hopes that this procedure not only provides
results that are in principle testable by observation, but also indirectly conrms the existence
of the otherwise unobservable regions of space. This was the program outlined in those early
works on eternal ination.
However, one runs into an immediate problem when one tries to extract statistical predictions
for cosmological observables in this setting. The main diuculty is due to the innite volume
of regions where an observer may be located. Inded, an eternally inating universe contains
an innite, inhomogeneous, and topologically complicated spacelike hypersurface (the reheating
surface) where observers may be expected to appear with a constant density per unit 3-volume.
In the landscape scenarios, one encounters a kind of innity that is in some sense more ill-
behaved than in the random-walk inationary scenarios. Not only each pocket universe may
contain innitely many observers, but also the number of dierent pocket universes in the entire
spacetime is innite. Pocket universes of dierent types are not statistically equivalent to each
other because they have dierent rates of nucleation of other pocket universes. There seems to
be no natural ordering on the set of all pocket universes throughout the spacetime, since most
of the pocket universes are spacelike separated.
In both these contexts random-walk type models and the landscape models one can
view eternal ination as a stochastic process that generates a topologically complicated and
noncompact locus of points where observers may appear. A random location of an observer
within that locus is a mathematically undened concept, similarly to the concept of an integer
number uniformly chosen among all the integers, or a real number uniformly chosen among
all the reals. This is the root cause of the technical and conceptual diculties known collectively
as the measure problem in multiverse cosmology. Nevertheless, one may try to formulate a
2
prescription for calculating observer-weighted probabilities of events. Such a prescription, also
called a measure, should in some sense correspond to the intuitive notion of probability of
observation at a random location in the spacetime. These issues are discussed in Chapter 4.
Several measure prescriptions have been proposed in the literature. Below in Sec. 4.2 these
proposals will be reviewed and their contrasting features will be characterized. Almost all of the
existing prescriptions are based on cutting o the innite volume of space by a certain geometric
construction. Thus, one considers a nite subset of the total volume where observers may appear;
the subset is characterized by a regulating parameter, such as the largest scale factor attained by
the observers or another geometric parameter. Then one gathers statistics throughout the nite
part of the volume, and nally takes the limit as the regulating parameter tends to innity. The
limit usually exists and yields a certain probability distribution for cosmological observables.
Unfortunately, it turned out that the limiting distribution depends sensitively on the choice
of the regularizing procedure. Since a natural mathematically consistent denition of the
measure is absent, one judges a measure prescription viable if its predictions are not obviously
pathological. Possible pathologies include the dependence on choice of spacetime coordinates,
the youngness paradox, and the Boltzmann brain problem, to be discussed in more detail
below.
The main goal of this dissertation is to report on a novel class of measure prescriptions that
were introduced in the authors recent publications, which form the basis of Chapters 5-7. This
class of measures is not based on regulating the spacetime by any geometric construction, but
rather on manipulating the events in the probability space in order to obtain a well-dened
subensemble of nite multiverses. A nite multiverse can be generated by rare chance if ina-
tion ends everywhere; the probability of this event is small but nonzero. One can characterize
the size of a nite multiverse in some way, e.g. by specifying the total volume of the reheating
regions (which will be nite), the total number of nucleated bubbles, or another such number
considered as a regulating parameter. In this way, one obtains a sequence of nite multiverses
that in a well-dened sense approximate the actual, innite multiverse as the regulating param-
eter tends to innity. It is then expected that the probability distribution of any observable will
tend to a well-dened limit. That limit is the prediction of the new measure.
I work out in detail the mathematical formalism necessary for computations in the new mea-
sure, both in the case of random-walk ination (Chapter 6) and in the case of a landscape
(Chapter 7). The computations turn out to be cumbersome since the possibility of creating a
nite multiverse is dicult to describe explicitly, especially in the limit of a very large size of
the nite multiverse. Nevertheless, it is possible to obtain direct results and general proofs for
various properties of the new measure. This is achieved by dierent methods in random-walk
eternal ination and in landscape scenarios. In each case I derived explicit formulas for the
predictions of the RV measure so as to make the nal computations more tractable if a specic
model is given.
In the Conclusion, I summarize the results obtained in this study and discuss some problems
and possibilities for future research.
Throughout most of the exposition I use the Planck units, c = = G = 1, which corresponds
to measuring energy in units of the Planck mass M
P
1.2 10
19
GeV, time in the Planck times
10
43
sec, and so on.
3
2 Eternal ination
Cosmological ination is a currently a dominant framework in theoretical cosmology. I will briey
outline its origins and main postulates as far as is necessary to set the context for the main part
of this work. For early reviews of ination, see e.g. [4] and the book [5]. Eternal ination and
the accompanying measure problem are reviewed and discussed in Refs. [6, 7, 8, 9, 10, 11].
2.1 Cosmological ination
We begin with the story of ination. While the hot big bang cosmological scenario has been
widely accepted after the detection of the cosmic microwave background (CMB) radiation and
explanation of the nucleosynthesis, several poorly explained and contradictory facts remained.
The main problems of the standard hot cosmology were the horizon problem and the atness
problem.
The horizon problem stems from the fact that the observed CMB is highly isotropic, with
relative temperature variation 10
5
(see, for example, [12]). The CMB radiation is coming
from the last scattering surface, which at the time of decoupling consisted of a large number
of causally disconnected horizon-size regions, each region occupying about 2

of todays sky.
However, observations of the high degree of homogeneity of the CMB temperature suggest
that all these regions had had a nearly equal temperature at the time of last scattering. This
absence of large-scale uctuations is dicult to explain unless we assume rather unnatural-
looking, extremely uniform initial conditions.
The atness problem is also in a sense a problem of initial conditions: the Universe must
initially have been unnaturally close to at. The density parameter evolves in such a way that
any deviation from = 1 grows with time,
[ 1[ t
2
4
3(1+w)
a
1+3w
, (2.1)
which can also be expressed through the temperature T as
[ 1[ T
(1+3w)
. (2.2)
Since the value of at present is of order 1, it must have been extremely close to 1 at early
times. For instance, at Planck temperatures T
P
10
19
GeV with w = 1/3 and T
now
= 3 K one
obtains the estimate
[
P
1[ = [
now
1[
_
T
now
T
P
_
2
10
60
. (2.3)
If one assumes a more natural initial condition for , for example, that 1 at Planck time,
then one nds that the Universe would have either collapsed within a few Planck times if > 1,
or cooled down to the present temperature of 3 K within 10
11
sec if < 1 [13]. It is hard to
explain this exceptional ne-tuning of the initial matter density.
5
2 Eternal ination
These problems are not the only faults of the standard scenario. For instance, explanation
of the origin of structure in the framework of the standard model is also problematic. The
description of growth of density uctuations of matter due to gravitational instability [14, 15, 16]
would provide an explanation for the formation of stars and galaxies if the initial uctuations
had a scale-invariant power spectrum at horizon crossing [14, 17, 18]
P (k) k
n
, n 1. (2.4)
However, the wavelength of a galaxy-scale uctuation must have been at early times larger
than the horizon size. Such a perturbation is dicult to explain by a causal mechanism. The
standard model simply assumes a homogeneous initial state and neither explains how the initial
uctuations occurred nor predicts their spectrum.
These and other shortcomings of the standard hot scenario were suciently compelling so
that the concept of cosmological ination was relatively quickly accepted when its advantages
were rst clearly advocated by A. Guth [19].
2.1.1 The inationary paradigm
Cosmological ination is a regime of fast expansion of the Universe with the expansion rate
a/a = H (t) given by a slowly changing function of t (so that the change of H during one
Hubble time is negligible,

HH
1
H). Then the scale factor is approximately exponentially
growing with time,
a (t) exp
__
H (t) dt
_
. (2.5)
The spacetime with this scale factor is similar to the de Sitter space of constant curvature
R 12H
2
. More precisely, ination is dened as a period of accelerated expansion, a(t) > 0.
This condition allows the Hubble rate H(t) to decrease as long as it does not decrease too
quickly. The earliest working proposal of an inationary model can be found in the works of
A. Starobinsky [20, 21]. In that scenario, a period of ination was driven by a modication of
gravity. However, this pioneering work remained relatively unappreciated until A. Guth [19]
pointed out that several major problems of standard cosmology would disappear at once if an
epoch of accelerated expansion were to precede the hot initial state of the standard scenario,
provided that the duration of the inationary epoch is large enough compared with the Hubble
time scale, such that the total expansion factor during ination is a exp (60). Since such
a large expansion cools the Universe to extremely low temperatures, a reheating must occur
prior to the onset of the radiation-dominated epoch. We shall now briey explain how the
problems outlined in the previous section are solved in this modied scenario and then describe
how ination was implemented in particular models.
The evolution of during the inationary epoch is given by
(t) 1 =
_
a
0
a (t)
_
2
(
0
1)
H
2
0
H
2
(t)
(
0
1) exp
_
_
2
t
_
t
0
H (t) dt
_
_
. (2.6)
Unlike the power-law expansion a t

with < 1, ination draws the value of nearer to

1. If the total expansion factor a/a
0
(the amount of ination) is large enough, i.e. at least
exp (60), which is called 60 e-foldings, then for a generic initial condition
0
1 the value of
at the end of ination will be as close to 1 as the observational constraints (2.3) require. This
6
2.1 Cosmological ination
solves the atness problem. In fact, the amount of ination in generic models is much larger
than 60 e-foldings, and would be typically driven so close to 1 by the end of ination that
it would not signicantly deviate from = 1 afterwards, during the radiation-dominated and
matter-dominated expansion. Thus, generic models of ination predict that we should observe
1 nowadays.
The horizon problem manifested by the observed homogeneity of the CMB is absent because,
according to the inationary scenario, the whole surface of last scattering has been before in-
ation a small patch well under horizon size, and one would expect inhomogeneities in a small
region to be small. An alternate way to express this is to say that the initial inhomogeneities
have beeninated away. In this light, the homogeneous state at the beginning of the radiation
era does not appear mysterious.
It has also been shown [22, 23, 24, 25, 26, 27] that vacuum uctuations of matter elds during
ination give rise to an approximately scale-invariant (n 1) spectrum of perturbations (2.4),
as necessary to explain structure formation.
2.1.2 Inationary models: slow roll
A large number of specic inationary models have been proposed in the early 1980s. The model
originally proposed by Starobinsky remains viable even in view of todays experimental data. The
model of Guth [19], sometimes called the old inationaryscenario, did not provide an adequate
explanation of the exit from ination (the graceful exit problem). Several newscenarios were
subsequently introduced by A. Linde [28] and others. A detailed review of inationary models
is beyond the scope of this work; we instead concentrate on the most general common features
of inationary models.
The inationary expansion must be supported either by a modication of gravity or by exotic
matter with an equation of state p <
1
3
. Since neither a detectable modication of Einsteins
General Relativity nor any matter eld with such an equation of state has been found in ex-
periments, models of ination necessarily hypothesize either a new eld or a modied theory of
gravity. The rst major type of inationary models considered here is a class of models where
the Hubble rate H(t) is a smooth function of time. Initially H is large (although always well
below the Planck scale, H 1), and ination ends when H approaches zero.
The easiest way to model such evolution is to assume that a scalar-eld with an eective
potential V () drives ination. We consider a prototypical model with the action
S =
_
d
4
x

g
_
R
16
+
1
2
(

)
2
V ()
_
, (2.7)
wehre the scalar eld is minimally coupled to Einstein gravity. We will assume the Friedmann
metric ansatz,
ds
2
= dt
2
a
2
(t)
_
dr
2
1 kr
2
+r
2
d
2
+r
2
sin
2
d
2
_
. (2.8)
Einsteins equations of motion for the scale factor a (t) and the eld (x, t) are
a
2
a
2
+
k
a
2
=
8
3
_
V () +
1
2

2
_
, (2.9)

+ 3
a
a

1
a
2

2
=
dV ()
d
. (2.10)
7
2 Eternal ination
Because of large value of the scale factor a, we can disregard the curvature term
_
k/a
2
_
and
the spatial gradients of in Eqs. (2.9)(2.10). An exact treatment of the resulting equations,
including a recipe of how to construct a potential V () that would yield a given evolution a (t)
of the scale factor can be found in Ref. [29].
The slow roll approximation is based on the assumptions that the evolution of the eld is
such that the potential V () does not change appreciably on the Hubble time scale H
1
a/ a.
More precisely, one assumes that
1
2

2
V () ,

dV ()
d

(2.11)
and disregards the kinetic term

2
/2 in Eq. (2.9) and the

term in Eq. (2.10). The latter
assumption means that the friction term 3H

in Eq. (2.10) balances the force term V

(),
and the evolution of is overdamped. Then the eective equations of motion become
a
2
a
2
H
2
() =
8
3
V () , (2.12)

=
1
4
dH ()
d
. (2.13)
Here we denoted by H () the function
_
8V () /3, as is common in the literature.
1
Eqs. (2.12)(2.13) are the desired equations describing the slow roll regime of the evolution
of . Using Eqs. (2.12)(2.13), we can express the conditions (2.11) through H () and obtain
equivalently
_
H

16H
_
2
1,

12H

1. (2.14)
These are the requirements on the potential V () necessary for the slow roll approximation to
be valid.
2
The rst of these conditions also guarantees that the relative change of V () in one
Hubble time is negligible:
H
1
d
dt
V () V () . (2.15)
For a given potential there is usually a range of for which the slow roll conditions (2.14)
are satised. It is usually the case that reheating begins when reaches values for which the
slow roll condition is violated.
3
For simplicity we will assume that reheating begins at the value
=

such that
H

) 16H(

). (2.16)
If the slow roll ination starts at =
0
and ends at =

, the total expansion factor

a (t

) /a (t
0
) can be found from
ln
a (t

)
a (t
0
)
=
_
t
t
0
H (t) dt =
_

0
H ()
d

= 4
_

0
H ()
H

()
d. (2.17)
1
Although in some inationary models, notably in the open ination, the function H () does not approximate
the Hubble expansion rate a/a, we shall keep the notation throughout this text.
2
Another implied assumption is V () H ()
2
1, since any classical description is only valid far from the
Planck energy scales.
3
A detailed theory of reheating has been worked out in Refs. [30, 31].
8
2.2 Eternal ination
There are several typical shapes of the potential V () that allow for ination. In models of
new ination, the potential has a nearly at top, and ination ends when the eld rolls down
to the bottom (see Fig. 2.1). In chaotic models the potential is usually of the power-law form
V ()
n
or of exponential form, V () e

. We will not need to specify the potential in what

follows. The considerations of eternal ination apply generally to every such model, although
specic calculations of course require the knowledge of the inaton potential.
V

0
*
Figure 2.1: The inaton potential for the new inationary scenario at T = 0. The neighbor-
hood of the maximum of the potential at = 0 is very at, so that the eld
changes very slowly. Thermalization occurs after the eld nally rolls down to the
thermalization point =

.
2.2 Eternal ination
Eternal ination, or the fact that ination never ends in the whole Universe, is a generic property
of inationary models. The general idea of eternally inating spacetime was rst introduced
and developed in the 1980s [32, 33, 34, 35] in the context of slow-roll ination. Let us begin by
reviewing the main features of eternal ination, following Ref. [8].
A prototypical model contains a minimally coupled scalar eld (the inaton) with an
eective potential V () that is suciently at in some range of . When the eld has values
in this range, the spacetime is approximately de Sitter with the Hubble rate
a
a
=
_
8
3
V () H(). (2.18)
(We work in units where G = c = = 1.) The value of H remains approximately constant on
timescales of several Hubble times (t H
1
), while the eld follows the slow-roll trajectory

sr
(t). Quantum uctuations of the scalar eld in de Sitter background grow linearly with
9
2 Eternal ination
time [36, 37, 25],

2
(t + t))

2
(t)) =
H
3
4
2
t, (2.19)
at least for time intervals t of order several H
1
. Due to the quasi-exponential expansion of
spacetime during ination, Fourier modes of the eld are quickly stretched to super-Hubble
length scales. However, quantum uctuations with super-Hubble wavelengths cannot maintain
quantum coherence and become essentially classical [37, 25, 38, 39, 40]; this issue is discussed in
more detail in Sec. 2.2.2 below. The resulting eld evolution (t) can be visualized [27, 32, 38]
as a Brownian motion with a random jump of typical step size H/(2) during a time
interval t H
1
, superimposed onto the deterministic slow-roll trajectory
sr
(t). A statistical
description of this random walk-type evolution (t) is reviewed in Sec. 3.1.
One can distinguish two possible regimes of evolution depending on whether the deterministic
change of in one Hubble time is smaller or larger than a typical uctuation. If the deterministic
change

H
1
dominates the uctuations, the slow roll regime proceeds essentially unmodied.
In the opposite regime,

H
1
, the random walk dominates the evolution of , which means
that steps toward larger and smaller H () are almost equally probable. Once a horizon-sized
region where is dominated by uctuations is formed, it will expand to form several horizon-
sized regions, most of which would contain the eld still in the uctuation-dominated range.
One may say that a uctuation-dominated region reproduces itself, regardless of the evolution
of the neighbor regions outside its horizon. As a result, the volume of the inating domain
grows exponentially with time, even though any given co-moving point will eventually enter a
thermalized region.
In models of new ination, the uctuation-dominated range is near the at top of the
potential where the deterministic change of is small. In chaotic models it is usually a
range of bounded from below by some value
uct
, and from above by the Planck boundary

P
, which leads to the range
uct
< <
P
if
uct
<
P
and to no uctuation-dominated
range (nite ination) otherwise. For the power-law potential V () =
n
with 1, one
obtains
uct

1/(n+2)

P

1/n
. Generically, in both new and chaotic ination
the regions in the self-reproducing stage expand at the fastest possible rate. We see how the
feature of self-reproduction helps solve the problem of initial conditions for ination: Wherever
a self-reproducing region is formed, it dominates the physical volume of the Universe, and all
other regions (including those with initial conditions unsuitable for ination) will occupy an
exponentially small fraction of the total volume.
The jumps at points separated in space by many Hubble distances are essentially un-
correlated; this is another manifestation of the well-known no-hair property of de Sitter
space [41, 42, 43]. Thus the eld becomes extremely inhomogeneous on large (super-horizon)
scales after many Hubble times. Moreover, in the semi-classical picture it is assumed
4
that
the local expansion rate a/a H() tracks the local value of the eld (t, x) according to the
Einstein equation (2.18). Here a(t, x) is the scale factor function which varies with x only on
super-Hubble scales, a(t, x)x H
1
. Hence, the spacetime metric can be visualized as having
a slowly varying, locally de Sitter form (with spatially at coordinates x),
g

dx

dx

= dt
2
a
2
(t, x)dx
2
. (2.20)
4
This assumption was made in Ref. [33] and is still subject to active research. See Sec. 2.2.2 for more discussion
on this issue.
10
2.2 Eternal ination
t
x
y
Figure 2.2: A qualitative diagram of self-reproduction during ination. Shaded spacelike do-
mains represent Hubble-size regions with dierent values of the inaton eld . The
time step is of order H
1
. Dark-colored shades are regions undergoing reheating
( =

); lighter-colored shades are regions where ination continues. On average,

the number of inating regions grows with time.
The deterministic trajectory
sr
(t) eventually reaches a (model-dependent) value

signifying
the end of the slow-roll inationary regime and the beginning of the reheating epoch (thermaliza-
tion). Since the random walk process will lead the value of away from =

in some regions,
reheating will not begin everywhere at the same time. Moreover, regions where remains in
the inationary range will typically expand faster than regions near the end of ination where
V () becomes small. Therefore, a delay of the onset of reheating will be rewarded by additional
expansion of the proper 3-volume, thus generating more regions that are still inating. This fea-
ture is calledself-reproductionof the inationary spacetime [34]. Since each Hubble-size region
evolves independently of other such regions, one may visualize the spacetime as an ensemble of
inating Hubble-size domains (Fig. 2.2).
The process of self-reproduction will never result in a global reheating if the probability of
jumping away from =

and the corresponding additional volume expansion factors are su-

ciently large. The corresponding quantitative conditions and their realization in typical models
of ination are reviewed in Sec. 4.1. Under these conditions, the process of self-reproduction of
inating regions continues forever. At the same time, every given comoving worldline (except
for a set of measure zero; see Sec. 4.1) will sooner or later reach the value =

and enter
the reheating epoch. The resulting situation is known as eternal ination [34]. More precisely,
the term eternal ination means future-eternal self-reproduction of inating regions [44].
5
To
emphasize the fact that self-reproduction is due to random uctuations of a eld, one refers
to this scenario as eternal ination of random-walk type. Below we use the terms eternal
self-reproduction and eternal ination interchangeably.
Observers like us may appear only in regions where reheating already took place. Hence, it
is useful to consider the locus of all reheating events in the entire spacetime; in the presently
considered example, it is the set of spacetime points x there (x) =

. This locus is called the

reheating surface and is a noncompact, spacelike three-dimensional hypersurface [45, 3]. It is
5
It is worth emphasizing that the term eternal ination refers to future-eternity of ination in the sense
described above, but does not imply past-eternity. In fact, inationary spacetimes are generically not past-
eternal [45, 46].
11
2 Eternal ination
important to realize that a nite, initially inating 3-volume of space may give rise to a reheating
surface having an innite 3-volume, and even to innitely many causally disconnected pieces of
the reheating surface, each having an innite 3-volume. This feature of eternal ination is at
the root of several technical and conceptual diculties, as will be discussed below.
Everywhere along the reheating surface, the reheating process is expected to provide ap-
propriate initial conditions for the standard hot big bang cosmological evolution, including
nucleosynthesis and structure formation. In other words, the reheating surface may be visu-
alized as the locus of the hot big bang events in the spacetime. It is thus natural to view
the reheating surface as the initial equal-time surface for astrophysical observations in the post-
inationary epoch. Note that the observationally relevant range of the primordial spectrum
of density uctuations is generated only during the last 60 e-foldings of ination. Hence, the
duration of the inationary epoch that preceded reheating is not directly measurable beyond
the last 60 e-foldings; the total number of e-foldings can vary along the reheating surface and
can be in principle arbitrarily large.
6
The phenomenon of eternal ination is also found in multi-eld models of ination [49, 50],
as well as in scenarios based on Brans-Dicke theory [51, 52, 53], topological ination [54, 55],
braneworld ination [56], recycling universe[57], and the string theory landscape [58]. In some
of these models, quantum tunneling processes may generate bubblesof a dierent phase of the
vacuum (see Sec. 3.2 for more details). Bubbles will be created randomly at various places and
times, with a xed rate per unit 4-volume. In the interior of some bubbles, additional ination
may take place, followed by a new reheating surface. The interior structure of such bubbles is
sketched in Fig. 2.3. The nucleation event and the formation of bubble walls is followed by a
period of additional ination, which terminates by reheating. Standard cosmological evolution
and structure formation eventually give way to a -dominated universe. Innitely many galaxies
and possible civilizations may appear within a thin spacelike slab running along the interior
reheating surface. This reheating surface appears to interior observers as an innite, spacelike
hypersurface [59]. For this reason, such bubbles are calledpocket universes,while the spacetime
is called amultiverse. Generally, the termpocket universerefers to a noncompact, connected
component of the reheating surface [60].
In scenarios of this type, each bubble is causally disconnected from most other bubbles.
7
Hence, bubble nucleation events may generate innitely many statistically inequivalent, causally
disconnected patches of the reheating surface, every patch giving rise to a possibly innite
number of galaxies and observers. This feature signicantly complicates the task of extracting
physical predictions from these models. This class of models is referred to as eternal ination
of tunneling type.
The fact that eternal ination is generic to many scenarios signicantly changes the global
picture of the Universe. It is likely that while ination ended in our neighborhood of the Universe
approximately 10
10
years ago, there still are and will always be very large domains where ination
goes on. In a sense, eternal ination is a reversal of the cosmological principle, yielding a picture
of the Universe which is extremely inhomogeneous on the ultra-large scale.
In the following subsections, I discuss the motivation for studying eternal ination as well as
physical justications for adopting the eective stochastic picture. Dierent techniques devel-
6
For instance, it was shown that holographic considerations do not place any bounds on the total number of e-
foldings during ination [47]. For recent attempts to limit the number of e-foldings using a dierent approach,
see e.g. [48]. Note also that the eects of random jumpsare negligible during the last 60 e-foldings of ination,
since the produced perturbations must be of order 10
5
according to observations.
7
Collisions between bubbles are rare [61]; however, eects of bubble collisions are observable in principle [62].
12
2.2 Eternal ination
domination
reheating
wall
wall
nucleation
Figure 2.3: A spacetime diagram of a bubble interior. The innite, spacelike reheating surface
is shown in darker shade. Galaxy formation is possible within the spacetime region
indicated.
oped for describing eternal ination are reviewed in Sec. 3. Section 4 contains an overview of
methods for extracting predictions and a discussion of the accompanying measure problem.
2.2.1 Predictions in eternal ination
The hypothesis of cosmological ination was invoked to explain several outstanding puzzles in
observational data [19]. However, some observed quantities (such as the cosmological constant
or elementary particle masses) may be expectation values of slowly-varying eective elds
a
.
Within the phenomenological approach, we are compelled to consider also the uctuations of
the elds
a
during ination, on the same footing as the uctuations of the inaton . Hence,
in a generic scenario of eternal ination, all the elds
a
arrive at the reheating surface =

with values that can be determined only statistically. Observers appearing at dierent points
in space may thus measure dierent values of the cosmological constant, elementary particle
masses, spectra of primordial density uctuations, and other cosmological parameters.
It is important to note that inhomogeneities in observable quantities are created on scales
far exceeding the Hubble horizon scale. Such inhomogeneities are not directly accessible to
astrophysical experiments. Nevertheless, the study of the global structure of eternally inating
spacetime is not merely of academic interest. Fundamental questions regarding the cosmological
singularities, the beginning of the Universe and of its ultimate fate, as well as the issue of the
cosmological initial conditions all depend on knowledge of the global structure of the spacetime
as predicted by the theory, whether or not this global structure is directly observable (see
e.g. [63, 64]). In other words, the fact that some theories predict eternal ination inuences
our assessment of the viability of these theories. In particular, the problem of initial conditions
for ination [65] is signicantly alleviated when eternal ination is present. For instance, it
was noted early on that the presence of eternal self-reproduction in the chaotic inationary
13
2 Eternal ination
scenario [66] essentially removes the need for the ne-tuning of the initial conditions [67, 34].
More recently, constraints on initial conditions were studied in the context of self-reproduction
in models of quintessence [68] and k-ination [69].
Since the values of the observable parameters
a
are random, it is natural to ask for the
probability distribution of
a
that would be measured by a randomly chosen observer. Un-
derstandably, this question has been the main theme of much of the work on eternal ination.
Obtaining an answer to this question promises to establish a more direct contact between sce-
narios of eternal ination and experiment. For instance, if the probability distribution for the
cosmological constant were peaked near the experimentally observed, puzzlingly small value
(see e.g. [70] for a review of the cosmological constant problem), the smallness of would be
explained as due to observer selection eects rather than to fundamental physics. Considerations
of this sort necessarily involve some anthropic reasoning; however, the relevant assumptions are
minimal. The basic goal of theoretical cosmology is to select physical theories of the early uni-
verse that are most compatible with astrophysical observations, including the observation of our
existence. It appears reasonable to assume that the civilization of Planet Earth evolved near a
randomly chosen star compatible with the development of life, within a randomly chosen galaxy
where such stars exist. Many models of ination generically include eternal ination and hence
predict the formation of innitely many galaxies where civilizations like ours may develop. It
is then also reasonable to assume that our civilization is typical among all the civilizations that
evolved in galaxies formed at any time in the universe. This assumption is called the principle
of mediocrity [3].
To use the principle of mediocrity for extracting statistical predictions from a model of
eternal ination, one proceeds as follows [3, 71]. In the example with the elds
a
described
above, the question is to determine the probability distribution for the values of
a
that a
random observer will measure. Presumably, the values of the elds
a
do not directly inuence
the emergence of intelligent life on planets, although they may aect the eciency of structure
formation or nucleosynthesis. Therefore, we may assume a xed,
a
-dependent mean number
of civilizations
civ
(
a
) per galaxy and proceed to ask for the probability distribution P
G
(
a
) of

a
near a randomly chosen galaxy. The observed probability distribution of
a
will then be
P(
a
) = P
G
(
a
)
civ
(
a
). (2.21)
One may use the standard hot big bang cosmology to determine the average number
G
(
a
)
of suitable galaxies per unit volume in a region where reheating occurred with given values of

a
; in any case, this task does not appear to pose diculties of principle. Then the computation
of P
G
(
a
) is reduced to determining the volume-weighted probability distribution 1(
a
) for
the elds
a
within a randomly chosen 3-volume along the reheating surface. The probability
distribution of
a
will be expressed as
P(
a
) = 1(
a
)
G
(
a
)
civ
(
a
). (2.22)
However, dening 1(
a
) turns out to be far from straightforward since the reheating surface
in eternal ination is an innite 3-surface with a complicated geometry and topology. The
lack of a natural, unambiguous, unbiased measure on the innite reheating surface is known
as the measure problem in eternal ination. Existing approaches and measure prescriptions
are discussed in Sec. 4, where two main alternatives (the volume-based and worldline-based
measures) are presented. In Sections 4.2 and 4.4 I give arguments in favor of using the volume-
based measure for computing the probability distribution of values
a
measured by a random
14
2.2 Eternal ination
observer. The volume-based measure has been applied to obtain statistical predictions for the
gravitational constant in Brans-Dicke theories [51, 52], cosmological constant (dark energy) [72,
73, 74, 75, 76, 77], particle physics parameters [78, 79, 80], and the amplitude of primordial
density perturbations [81, 74, 82, 77].
The issue of statistical predictions has recently come to the fore in conjunction with the
discovery of the string theory landscape. According to various estimates, one expects to have
between 10
500
and 10
1500
possible vacuum states of string theory [83, 84, 58, 85, 86]. The
string vacua dier in the geometry of spacetime compactication and have dierent values of
the eective cosmological constant (or dark energy density). Transitions between vacua may
happen via the well-known Coleman-deLuccia tunneling mechanism [59]. Once the dark energy
dominates in a given region, the spacetime becomes locally de Sitter. Then the tunneling process
will create innitely many disconnected daughter bubbles of other vacua. Observers like us
may appear within any of the habitable bubbles. Since the fundamental theory does not specify
a single preferred vacuum, it remains to try determining the probability distribution of vacua
as found by a randomly chosen observer. The volume-based and worldline-based measures
can be extended to scenarios with multiple bubbles, as discussed in more detail in Sec. 4.4. Some
recent results obtained using these measures are reported in Refs. [87, 7, 88, 89].
2.2.2 Physical justications of the semiclassical picture
The standard framework of inationary cosmology asserts that vacuum quantum uctuations
with super-horizon wavelengths become classical inhomogeneities of the eld . The calculations
of cosmological density perturbations generated during ination [22, 23, 24, 25, 26, 27, 90, 91] also
assume that a classicalization of quantum uctuations takes place via the same mechanism. In
the calculations, the statistical average

2
_
of classical uctuations on super-Hubble scales is
simply set equal to the quantum expectation value 0[

2
[0) in a suitable vacuum state. While
this approach is widely accepted in the cosmology literature, a growing body of research is
devoted to the analysis of the quantum-to-classical transition during ination (see e.g. [92] for
an early review of that line of work). Since a detailed analysis would be beyond the scope of the
present text, I merely outline the main ideas and arguments relevant to this issue.
A standard phenomenological explanation of the classicalization of the perturbations is as
follows. For simplicity, let us restrict our attention to a slow-roll inationary scenario with one
scalar eld . In the slow roll regime, one can approximately regard as a massless scalar eld
in de Sitter background spacetime [35]. Due to the exponentially fast expansion of de Sitter
spacetime, super-horizon Fourier modes of the eld are in squeezed quantum states with
exponentially large ( e
Ht
) squeezing parameters [93, 94, 95, 96, 97, 98]. Such highly squeezed
states have a macroscopically large uncertainty in the eld value and thus quickly decohere
due to interactions with gravity and with other elds. The resulting mixed state is eectively
equivalent to a statistical ensemble with a Gaussian distributed value of . Therefore one may
compute the statistical average

2
_
as the quantum expectation value 0[

2
[0) and interpret
the uctuation as a classical noise. A heuristic description of the classicalization [35] is
that the quantum commutators of the creation and annihilation operators of the eld modes,
[ a, a

] = 1, are much smaller than the expectation values

a
_
1 and are thus negligible.
A related issue is the backreaction of uctuations of the scalar eld on the metric.
8
According
8
The backreaction eects of the long-wavelength uctuations of a scalar eld during ination have been investi-
gated extensively (see e.g. [99, 100, 101, 102, 103, 104, 105]).
15
2 Eternal ination
to the standard theory (see e.g. [90, 91] for reviews), the perturbations of the metric arising due
to uctuations of are described by an auxiliary scalar eld (sometimes called the Sasaki-
Mukhanov variable) in a xed de Sitter background. Thus, the classicalization eect should
apply equally to the uctuations of and to the induced metric perturbations. At the same time,
these metric perturbations can be viewed, in an appropriate coordinate system, as uctuations
of the local expansion rate H() due to local uctuations of [25, 35, 106]. Thus one arrives
at the picture of a locally de Sitter spacetime with the metric (2.20), where the Hubble rate
a/a = H() uctuates on super-horizon length scales and locally follows the value of via the
classical Einstein equation (2.18).
The picture as outlined is phenomenological and does not provide a description of the quantum-
to-classical transition in the metric perturbations at the level of eld theory. For instance, a
uctuation of leading to a local increase of H() necessarily violates the null energy con-
dition [107, 108, 109]. The cosmological implications of such semiclassical uctuations (see
e.g. the scenario of island cosmology [110, 111, 112]) cannot be understood in detail within the
framework of the phenomenological picture.
A more fundamental approach to describing the quantum-to-classical transition of perturba-
tions was developed using non-equilibrium quantum eld theory and the inuence functional
formalism [113, 114, 115, 116]. In this approach, decoherence of a pure quantum state of into
a mixed state is entirely due to the self-interaction of the eld . In particular, it is predicted
that no decoherence would occur for a free eld with V () =
1
2
m
2
. This result is at variance
with the accepted paradigm of classicalization as outlined above. If the source of the noise is
the coupling between dierent perturbation modes of , the typical amplitude of the noise will
be second-order in the perturbation. This is several orders of magnitude smaller than the am-
plitude of noise found in the standard approach. Accordingly, it is claimed [117, 118] that the
magnitude of cosmological perturbations generated by ination is several orders of magnitude
smaller than the results currently accepted as standard, and that the shape of the perturbation
spectrum depends on the details of the process of classicalization [119]. Thus, the results
obtained via the inuence functional techniques do not appear to reproduce the phenomeno-
logical picture of classicalization as outlined above. This mismatch emphasizes the need for
a deeper understanding of the nature of the quantum-to-classical transition for cosmological
perturbations.
Finally, let us mention a dierent line of work which supports the classicalization picture.
In Refs. [38, 120, 121, 122, 123, 124], calculations of (renormalized) expectation values such as

2
),

4
), etc., were performed for eld operators

in a xed de Sitter background. The results
were compared with the statistical averages
_
P(, t)
2
d,
_
P(, t)
4
d, etc., (2.23)
where the distribution P(, t) describes the random walk of the eld in the Fokker-Plank
approach (see Sec. 3.1). It was shown that the leading late-time asymptotics of the quantum
expectation values coincide with the corresponding statistical averages (2.23). These results
appear to validate the random walk approach, albeit in a limited context (in the absence of
backreaction).
16
3 Stochastic approach to ination
The stochastic approach to ination is a semiclassical, statistical description of the spacetime
resulting from quantum uctuations of the inaton eld(s) and their backreaction on the met-
ric [32, 33, 125, 126, 127, 128, 129, 130, 131, 132, 133]. In this description, the spacetime remains
everywhere classical but its geometry is determined by a stochastic process. In the next sub-
sections I review the main tools used in the stochastic approach for calculations in the context
of random-walk type, slow-roll ination. Models involving tunneling-type eternal ination are
considered in Sec. 3.2.
3.1 Random walk-type eternal ination
The important features of random walk-type eternal ination can be understood by considering
a simple slow-roll inationary model with a single scalar eld and a potential V (). The
slow-roll evolution equation is

=
1
3H
dV
d
=
1
4
dH
d
v(), (3.1)
where H() is dened by Eq. (2.18) and v() is a model-dependent function describing the
velocity

of the deterministic evolution of the eld . The slow-roll trajectory
sr
(t), which
is a solution of Eq. (3.1), is an attractor [134, 135] for trajectories starting with a wide range of
initial conditions.
1
As discussed in Sec. 2.2.2, the super-horizon modes of the eld are assumed to undergo a
rapid quantum-to-classical transition. Therefore one regards the spatial average of on scales
of several H
1
as a classical eld variable. The spatial averaging can be described with help of
a suitable window function,
(x))
_
W(x y)(y)d
3
y. (3.2)
It is implied that the window function W(x) decays quickly on physical distances a [x[ of order
several H
1
. From now on, let us denote the volume-averaged eld simply by (no other eld
will be used).
As discussed above, the inuence of quantum uctuations leads to random jumps superim-
posed on top of the deterministic evolution of the volume-averaged eld (t, x). This may be
described by a Langevin equation of the form [33]

(t, x) = v() +N(t, x), (3.3)

where N(t, x) stands for noise and is assumed to be a Gaussian random function with known
correlator [33, 136, 137, 138]

N(t, x)N(

t, x)
_
= C(t,

t, [x x[ ; ). (3.4)
1
See Ref. [69] for a precise denition of an attractor trajectory in the context of ination.
17
3 Stochastic approach to ination
An explicit form of the correlator C depends on the specic window function W used for av-
eraging the eld on Hubble scales [138]. However, the window function W is merely a phe-
nomenological device used in lieu of a complete ab initio derivation of the stochastic ination
picture. One expects, therefore, that results of calculations should be robust with respect to the
choice of W. In other words, any uncertainty due to the choice of the window function must be
regarded as an imprecision inherent in the method. For instance, a robust result in this sense is
an exponentially fast decay of correlations on time scales t H
1
,
C(t,

t, [x x[ ; ) exp
_
2H()

_
, (3.5)
which holds for a wide class of window functions [138].
For the purposes of the present consideration, we only need to track the evolution of (t, x)
along a single comoving worldline x = const. Thus, we will not need an explicit form of
C(t,

t, [x x[ ; ) but merely the value at coincident points t =

t, x = x, which is computed in
the slow-roll inationary scenario as [33]
C(t, t, 0; ) =
H
2
()
4
2
. (3.6)
(This represents the uctuation (2.19) accumulated during one Hubble time, t = H
1
.) Due
to the property (3.5), one may neglect correlations on time scales t H
1
in the noise eld.
2
Thus, the evolution of on time scales t H
1
can be described by a nite-dierence form
of the Langevin equation (3.3),
(t + t) (t) = v()t +
_
2D()t (t), (3.7)
where
D()
H
3
()
8
2
(3.8)
and is a normalized random variable representing white noise,
) = 0,

2
_
= 1, (3.9)
(t)(t + t)) = 0 for t H
1
. (3.10)
Equation (3.7) is interpreted as describing a Brownian motion (t) with the systematic drift
v() and the diusion coecient D(). In a typical slow-roll inationary scenario, there will
be a range of where the noise dominates over the deterministic drift,
v()t
_
2D()t, t H
1
. (3.11)
Such a range of is called the diusion-dominated regime. For near the end of ination,
the amplitude of the noise is very small, and so the opposite inequality holds. This is the
deterministic regime where the random jumps can be neglected and the eld follows the
slow-roll trajectory.
2
Taking these correlations into account leads to a picture of color noise [139, 140]. In what follows, we only
consider the simpler picture of white noise as an approximation adequate for the issues at hand.
18
3.1 Random walk-type eternal ination
3.1.1 Fokker-Planck equations
A useful description of the statistical properties of (t) is furnished by the probability density
P(, t)d of having a value at time t. As in the case of the Langevin equation, the values (t)
are measured along a single, randomly chosen comoving worldline x = const. The probability
distribution P(, t) satises the Fokker-Planck (FP) equation whose standard derivation we
omit [141, 142],

t
P =

[v()P +

(D()P)] . (3.12)
The coecients v() and D() are in general model-dependent and need to be calculated in each
particular scenario. These calculations require only the knowledge of the slow-roll trajectory
and the mode functions of the quantized scalar perturbations. For ordinary slow-roll ination
with an eective potential V (), the results are well-known expressions (6.11) and (6.12). The
corresponding expressions for models of k-ination were derived in Ref. [69] using the relevant
quantum theory of perturbations [143].
It is well known that there exists a factor ordering ambiguity in translating the Langevin
equation into the FP equation if the amplitude of the noise depends on the position. Specif-
ically, the factor D() in Eq. (3.7) may be replaced by D( + t), where 0 < < 1 is an
arbitrary constant. With ,= 0, the term

(DP) in Eq. (3.12) will be replaced by a dierent

ordering of the factors,

(DP)

_
D

_
D
1
P
__
. (3.13)
Popular choices = 0 and =
1
2
are called the Ito and the Stratonovich factor ordering
respectively. Motivated by the considerations of Ref. [144], we choose = 0 as shown in Eqs. (3.7)
and (3.12). Given the phenomenological nature of the Langevin equation (3.7), one expects
that any ambiguity due to the choice of represents an imprecision inherent in the stochastic
approach. This imprecision is typically of order H
2
1 [145].
The quantity P(, t) may be also interpreted as the fraction of the comoving volume (i.e. co-
ordinate volume d
3
x) occupied by the eld value at time t. Another important characteristic
is the volume-weighted distribution P
V
(, t)d, which is dened as the proper 3-volume (as op-
posed to the comoving volume) of regions having the value at time t. (To avoid considering
innite volumes, one may restrict ones attention to a nite comoving domain in the universe
and normalize P
V
(, t) to unit volume at some initial time t = t
0
.) The volume distribution
satises a modied FP equation [35, 128, 130],

t
P
V
=

[v()P
V
+

(D()P
V
)] + 3H()P
V
, (3.14)
which diers from Eq. (3.12) by the term 3HP
V
that describes the exponential growth of 3-
volume in inating regions.
3
Presently we consider scenarios with a single scalar eld; however, the formalism of FP equa-
tions can be straightforwardly extended to multi-eld models (see e.g. Ref. [144]). For instance,
the FP equation for a two-eld model is

t
P =

(DP) +

(DP)

(v

P)

(v

P) , (3.15)
where D(, ), v

(, ), and v

(, ) are appropriate coecients.

3
A more formal derivation of Eq. (3.14) as well as details of the interpretation of the distributions P and PV in
terms of ensembles of worldlines can be found in Ref. [146].
19
3 Stochastic approach to ination
3.1.2 Methods of solution
In principle, one can solve the FP equations forward in time by a numerical method, starting
from a given initial distribution at t = t
0
. To specify the solution uniquely, the FP equations
must be supplemented by boundary conditions at both ends of the inating range of [132, 133].
At the reheating boundary ( =

), one imposes the exit boundary conditions,

[D()P]
=
= 0,

[D()P
V
]
=
= 0. (3.16)
These boundary conditions express the fact that random jumps are very small at the end of
ination and cannot move the value of away from =

. If the potential V () reaches

Planck energy scales at some =
max
(this happens generally in chaotic type inationary
scenarios with unbounded potentials), the semiclassical picture of spacetime breaks down for
regions with
max
. Hence, a boundary condition must be imposed also at =
max
. For
instance, one can use the absorbing boundary condition,
P(
max
) = 0, (3.17)
which means that Planck-energy regions with =
max
disappear from consideration [132, 133].
Once the boundary conditions are specied, one may write the general solution of the FP
equation (3.12) as
P(, t) =

P
()
() e
t
, (3.18)
where the sum is performed over all the eigenvalues of the dierential operator

LP

[v()P +

(D()P)] , (3.19)
and the corresponding eigenfunctions P
()
are dened by

LP
()
() = P
()
(). (3.20)
The constants C

can be expressed through the initial distribution P(, t

0
).
By an appropriate change of variables z, P() F(z), the operator

L may be brought
into a manifestly self-adjoint form [126, 128, 147, 129, 130, 145],

L
d
2
dz
2
+U(z). (3.21)
Then one can show that all the eigenvalues of

L are nonpositive; in particular, the (alge-
braically) largest eigenvalue
max
< 0 is nondegenerate and the corresponding eigenfunc-
tion P
(max)
() is everywhere positive [145, 69]. Hence, this eigenfunction describes the late-time
asymptotic of the distribution P(, t),
P(, t) P
(max)
() e
t
. (3.22)
The distribution P
(max)
() is the stationary distribution of per comoving volume at late
times. The exponential decay of the distribution P(, t) means that at late times most of the
comoving volume (except for an exponentially small fraction) has nished ination and entered
reheating.
20
3.1 Random walk-type eternal ination
Similarly, one can represent the general solution of Eq. (3.14) by
P
V
(, t) =

P
(

)
V
()e

t
, (3.23)
where
[

L + 3H()]P
(

)
() =

P
(

)
(). (3.24)
By the same method as for the operator

L, it is possible to show that the spectrum of eigenvalues

of the operator

L + 3H() is bounded from above and that the largest eigenvalue

max

admits a nondegenerate, everywhere positive eigenfunction P
( )
(). However, the largest
eigenvalue may be either positive or negative. If > 0, the late-time behavior of P
V
(, t) is
P
V
(, t) P
( )
()e
t
, (3.25)
which means that the total proper volume of all the inating regions grows with time. This
is the behavior expected in eternal ination: the number of independently inating domains
increases without limit. Thus, the condition > 0 is the criterion for the presence of eternal
self-reproduction of inating domains. The corresponding distribution P
( )
() is called the
stationary distribution [132, 133, 1, 53].
If 0, eternal ination does not occur and the entire space almost surely (i.e. with proba-
bility 1) enters the reheating epoch at a nite time.
If the potential V () is of new inationary type [28, 148, 37, 149, 150] and has a global
maximum at say =
0
, the eigenvalues and can be estimated (under the usual slow-roll
assumptions on V ) as [145]

V

(
0
)
8V (
0
)
H(
0
) < 0, 3H(
0
) > 0. (3.26)
Therefore, eternal ination is generic in the new inationary scenario.
Let us comment on the possibility of obtaining solutions P(, t) in practice. With the po-
tential V () =
4
, the full time-dependent FP equation (3.12) can be solved analytically via
a nonlinear change of variable
2
[151, 147, 152]. This exact solution, as well as an ap-
proximate solution P(, t) for a general potential, can be also obtained using the saddle-point
evaluation of a path-integral expression for P(, t) [153]. In some cases the eigenvalue equation

LP
()
= P
()
may be reduced to an exactly solvable Schrodinger equation. These cases include
potentials of the form V () = e

, V () =
2
, V () = cosh
2
(); see e.g. Ref. [145] for
other examples of exactly solvable potentials.
A general approximate method for determining P(, t) for arbitrary potentials [154, 155, 156]
consists of a perturbative expansion,
(t) =
0
(t) +
1
(t) +
2
(t) +..., (3.27)
applied directly to the Langevin equation. The result is (at the lowest order) a Gaussian ap-
21
3 Stochastic approach to ination
proximation with a time-dependent mean and variance [154],
P(, t)
1
_
2
2
(t)
exp
_

(
0
(t))
2
2
2
(t)
_
, (3.28)

2
(t)
H
2
(
sr
)

_

in
sr
H
3
H
3
d, (3.29)

0
(t)
sr
(t) +
H

2H

2
(t) +
H

4
_
H
3
in
H
2
in

H
3
H
2
_
, (3.30)
where
sr
(t) is the slow-roll trajectory and
in
is the initial value of . While methods based
on the Langevin equation do not take into account boundary conditions or volume weighting
eects, the formula (3.28) provides an adequate approximation to the distribution P(, t) in a
useful range of and t [156].
3.1.3 Gauge dependence issues
An important feature of the FP equations is their dependence on the choice of the time variable.
One can consider a replacement of the form
t , d T()dt, (3.31)
understood in the sense of integrating along comoving worldlines x = const, where T() > 0 is
an arbitrary function of the eld. For instance, a possible choice is T() H(), which makes
the new time variable dimensionless,
=
_
Hdt = ln a. (3.32)
This time variable is called scale factor time or e-folding time since it measures the number
of e-foldings along a comoving worldline.
The distributions P(, ) and P
V
(, ) are dened as before, except for considering the 3-
volumes along hypersurfaces of equal . These distributions satisfy FP equations similar to
Eqs. (3.12)(3.14). With the replacement (3.31), the coecients of the new FP equations are
modied as follows [145],
D()
D()
T()
, v()
v()
T()
, (3.33)
while the growth term 3HP
V
in Eq. (3.14) is replaced by 3HT
1
P
V
. The change in the
coecients may signicantly alter the qualitative behavior of the solutions of the FP equations.
For instance, stationary distributions dened through the proper time t and the e-folding time
= ln a were found to have radically dierent behavior [133, 1, 3]. This sensitivity to the choice of
the time gauge is unavoidable since hypersurfaces of equal may preferentially select regions
with certain properties. For instance, most of the proper volume in equal-t hypersurfaces is lled
with regions that have gained expansion by remaining near the top of the potential V (), while
hypersurfaces of equal scale factor will under-represent those regions. Thus, a statement such
as most of the volume in the Universe has values of with high potential energy is largely
gauge-dependent.
22
3.1 Random walk-type eternal ination
In the early works on eternal ination [132, 133, 1, 52], the late-time asymptotic distribution
of volume P
( )
V
() along hypersurfaces of equal proper time [see Eq. (3.25)] was interpreted
as the stationary distribution of eld values in the universe. However, the high sensitivity of
this distribution to the choice of the time variable makes this interpretation unsatisfactory.
Also, it was noted [157] that equal-proper time volume distributions predict an unacceptably
small probability for the currently observed CMB temperature. The reason for this result is
the extreme bias of the proper-time gauge towards over-representing regions where reheating
occurred very recently [158, 3]. This is a manifestation of the so-calledyoungness paradox[157].
One might ask whether hypersurfaces of equal scale factor or some other choice of time gauge
would provide less biased answers. However, it turns out [146] that there exists no a priori choice
of the time gauge that provides unbiased equal- probability distributions for all potentials
V () in models of slow-roll ination (see Sec. 4.3 for details).
Although the FP equations necessarily involve a dependence on gauge, they do provide a
useful statistical picture of the distribution of elds in the universe. The FP techniques can also
be used for deriving several gauge-independent results. For instance, the presence of eternal
ination is a gauge-independent statement (see also Sec. 4.1): if the largest eigenvalue is
positive in one gauge of the form (3.31), then > 0 in every other gauge [159]. Using the FP
approach, one can also compute the fractal dimension of the inating domain [160, 159] and the
probability of exiting ination through a particular point

of the reheating boundary in the

conguration space (in case there exists more than one such point).
The exit probability can be determined as follows [144, 69]. Let us assume for simplicity that
there are two possible exit points

and
E
, and that the initial distribution is concentrated at
=
0
, i.e.
P(, t = 0) = (
0
), (3.34)
where
E
<
0
<

. The probability of exiting ination through =

E
during a time interval
[t, t +dt] is
dp
exit
(
E
) = v(
E
)P(
E
, t)dt (3.35)
(note that v(
E
) < 0). Hence, the total probability of exiting through =
E
at any time is
p
exit
(
E
) =
_

0
dp
exit
(
E
) = v(
E
)
_

0
P(
E
, t)dt. (3.36)
Introducing an auxiliary function F() as
F() v()
_

0
P(, t)dt, (3.37)
one can show that F() satises the gauge-invariant equation,

_
D
v
F
_
F
_
= (
0
). (3.38)
This is in accord with the fact that p
exit
(
E
) = F(
E
) is a gauge-invariant quantity. Equa-
tion (3.38) with the boundary conditions
F(

) = 0,

(
D
v
F)

=
E
= 0, (3.39)
can be straightforwardly integrated and yields explicit expressions for the exit probability p
exit
(
E
)
as a function of the initial value
0
[69]. The exit probability p
exit
(

) can be determined simi-

larly.
23
3 Stochastic approach to ination
3.2 Self-reproduction of tunneling type
Until now, we considered eternal self-reproduction due to random walk of a scalar eld. Another
important class of models includes self-reproduction due to bubble nucleation.
4
Such scenarios
of eternal ination were studied in Refs. [161, 61, 162, 163, 164, 165].
In a locally de Sitter universe dominated by dark energy, nucleation of bubbles of false vacuum
may occur due to tunneling [166, 59, 43, 167]. The bubble nucleation rate per unit 4-volume
is very small [59, 168],
= O(1)H
4
exp
_
S
I


H
2
_
, (3.40)
where S
I
is the instanton action and H is the Hubble constant of the parent de Sitter back-
ground. Hence, bubbles will generically not merge into a single false-vacuum domain [61], and
innitely many bubbles will be nucleated at dierent places and times. The resulting daugh-
ter bubbles may again contain an asymptotically de Sitter, innite universe, which again gives
rise to innitely many grand-daughter bubbles. This picture of eternal self-reproduction was
called the recycling universe [57]. Some (or all) of the created bubbles may support a period
of additional ination followed by reheating, as shown in Fig. 2.3.
In the model of Ref. [57], there were only two vacua which could tunnel into each other.
A more recently developed paradigm of string theory landscape [84, 58, 85, 169] involves a
very large number of metastable vacua, corresponding to local minima of an eective potential
in eld space. A similar picture is found in the recycling universe [57] scenario. In the
landscape models, transitions between vacuum states are possible through bubble nucleation;
the interior of a bubble appears as an innite homogeneous open universe [59], if one disregards
the small probability of bubble collisions (see Refs. [62, 170] for analyses of bubble collisions).
The presently observed universe is situated within a bubble (called a pocket universe) of some
type. The value of the potential at each minimum is the eective value of the cosmological
constant in the corresponding vacuum. Figure 3.1 shows a phenomenologists view of the
landscape. Vacua with 0 do not allow any further tunneling
5
and are called terminal
vacua [175], while vacua with > 0 are calledrecyclableor transientsince they can tunnel to
other vacua with > 0 or 0. Bubbles with 0 cannot support any further tunneling and
are thus called terminal bubbles. Recyclable bubbles will give rise to innitely many nested
daughter bubbles. Conformal diagrams of the resulting spacetime are shown in Figs. 3.2 and
3.3. Of course, only a nite number of bubbles can be drawn; the bubbles actually form a fractal
structure in a conformal diagram [176].
A statistical description of the recycling spacetime can be obtained [57, 175] by considering
a single comoving worldline x = const that passes through dierent bubbles at dierent times.
(It is implied that the worldline is randomly chosen from an ensemble of innitely many such
worldlines passing through dierent points x.) Let the index = 1, ..., N label all the available
types of bubbles. For calculations, it is convenient to use the e-folding time ln a. We
are interested in the probability f

() of passing through a bubble of type at time . This

probability distribution is normalized by

= 1; the quantity f

() can be also visualized

as the fraction of the comoving volume occupied by bubbles of type at time . Denoting by

the nucleation rate for bubbles of type within bubbles of type , computed according
4
Both processes may be combined in a single scenario [57], but we shall consider them separately for clarity.
5
Asymptotically at = 0 vacua cannot support tunneling [171, 172, 173]; vacua with < 0 will quickly collapse
to a big crunch singularity [59, 174].
24
3.2 Self-reproduction of tunneling type

X
0
Figure 3.1: A schematic representation of the landscape of string theory, consisting of a large
number of local minima of an eective potential. The variable X collectively denotes
various elds and is the eective cosmological constant. Arrows show possible
tunneling transitions between vacua.
1
1
2
3
3
4
4
4 5
5
Figure 3.2: A conformal diagram of the spacetime where self-reproduction occurs via bubble
nucleation. Regions labeled 5 are asymptotically at ( = 0).
Figure 3.3: A conformal diagram of the future part of a landscape-driven spacetime (gure from
Ref. [176]).
25
3 Stochastic approach to ination
to Eq. (3.40), and by

the corresponding rate per unit time,

H
3

, we write
the master equation describing the evolution of f

(),
df

d
=

, (3.41)
where we introduced the auxiliary matrix M

. Given a set of initial conditions f

(0), one can

evolve f

() according to Eq. (3.41).

To proceed further, one may now distinguish the following two cases: Either terminal vacua
exist (some such that

= 0 for all ), or all the vacua are recyclable. (Theory suggests

that the former case is more probable [85].) If terminal vacua exist, then the late-time asymptotic
solution can be written as [175]
f

() f
(0)

+s

e
q
, (3.42)
where f
(0)

is a constant vector that depends on the initial conditions and has nonzero components
only in terminal vacua, and s

does not depend on initial conditions and is an eigenvector of

M

such that

= qs

, q > 0. This solution shows that all comoving volume

reaches terminal vacua exponentially quickly. (As in the case of random-walk ination, there
are innitely many eternally recycling points x that never enter any terminal vacua, but these
points form a set of measure zero.)
If there are no terminal vacua, the solution f

() approaches a constant distribution [177],

lim

() f
(0)

f
(0)

= 0, (3.43)
f
(0)

= H
4

exp
_

H
2

_
. (3.44)
In this case, the quantities f
(0)

are independent of initial conditions and are interpreted as the

fractions of time spent by the comoving worldline in bubbles of type .
We note that the description of spacetime in terms of the distribution f

() satisfying Eq. (3.41)

depends on the choice of the time variable in an essential way. A dierent choice (e.g. the
proper time t) would lead to an inequivalent master equation and qualitatively dierent dis-
tributions f

(t) [146].
6
One may adopt another approach and ignore the duration of time spent by the worldline
within each bubble. Thus, one describes only the sequence of the bubbles encountered along a
randomly chosen worldline [177, 178, 7]. If the worldline is initially in a bubble of type , then
the probability

of entering the bubble of type as the next bubble in the sequence after
is

. (3.45)
(For terminal vacua , we have

= 0 and so we may dene

= 0 for convenience.) Once

again we consider landscapes without terminal vacua separately from terminal landscapes. If
there are no terminal vacua, then the matrix

is normalized,

= 1, and is thus a
stochastic matrix [179] describing a Markov process of choosing the next visited vacuum. The
sequence of visited vacua is innite, so one can dene the mean frequency f
(mean)

of visiting
6
In the case of a landscape without terminal vacua, dierent choices of time lead only to a trivial transformation
of the probabilities f because these are the fractions of time spent by the worldline in the bubbles of type .
26
3.2 Self-reproduction of tunneling type
bubbles of type . If the probability distribution for the rst element in the sequence is f
(0)
,
then the distribution of vacua after k steps is given (in the matrix notation) by the vector
f
(k)
=
k
f
(0)
, (3.46)
where
k
means the k-th power of the matrix

. Therefore, the mean frequency of

visiting a vacuum is computed as an average of f
(k)
over n consecutive steps in the limit of
large n:
f
(mean)
= lim
n
1
n
n

k=1
f
(k)
= lim
n
1
n
n

k=1

k
f
(0)
. (3.47)
(It is proved in the theory of Markov processes that the limit f
(mean)

given by Eq. (3.47) almost

surely coincides with the mean frequency of visiting the state ; see e.g. Ref. [180], chapter
5, Theorem 2.1, and Ref. [181], Theorem 3.5.9.) It turns out that the distribution f
(mean)
is independent of the initial state f
(0)
and coincides with the distribution (3.43) found in the
continuous-time description [177].
If there exist terminal vacua, then almost all sequences will have a nite length. The distri-
bution of vacua in a randomly chosen sequence is still well-dened and can be computed using
Eq. (3.47) without the normalizing factor
1
n
,
f
(mean)
= (1 )
1
f
(0)
, (3.48)
but now the resulting distribution depends on the initial state f
(0)
[178, 7].
27
4 Predictions and measure issues
As discussed in Sec. 2.2.1, a compelling question in the context of eternal ination is how to
make statistical predictions of observed parameters. We begin by determining whether eternal
ination is present in a given model.
4.1 Presence of eternal ination
Since the presence of eternal self-reproduction in models of tunneling type is generically certain
(unless the nucleation rate for terminal vacua is unusually high), in this section we restrict our
attention to eternal ination of the random-walk type.
The hallmark of eternal ination is the unbounded growth the total number of independent
inating regions. The total proper 3-volume of the inating domain also grows without bound
at late times, at least when computed along hypersurfaces of equal proper time or equal scale
factor. However, the proper 3-volume is a gauge-dependent quantity, and one may construct time
gauges where the 3-volume decreases with time even in an everywhere expanding universe [146].
The 3-volume of an arbitrary family of equal-time hypersurfaces cannot be used as a criterion
for the presence of eternal ination. However, a weaker criterion is sucient: Eternal ination
is present if (and only if) there exists a choice of time slicing with an unbounded growth of the
3-volume of inating domains [146]. Equivalently, eternal ination is present if a nite comoving
volume gives rise to innite physical volume [157]. Thus, the presence or absence of eternal
ination is a gauge-independent statement. One may, of course, use a particular gauge (such as
the proper time or e-folding time) for calculations, as long as the result is known to be gauge-
independent. It can be shown that eternal ination is present if and only if the 3-volume grows
in the e-folding time slicing [146].
The presence of eternal ination has been analyzed in many specic scenarios. For instance,
eternal ination is generic in chaotic [182, 67, 34] and new [38] inationary models. It is
normally sucient to establish the existence of a diusion-dominated regime, that is, a range
of where the typical amplitude H of jumps is larger than the typical change of the
eld,

t, during one Hubble time t = H
1
. For models of scalar-eld ination, the condition
is
H
2
H

. (4.1)
Such a range of is present in most slow-roll models of ination. (For an example of an
inationary scenario where eternal ination is generically not present, see Ref. [183].) A strict
formal criterion for the presence of eternal ination is the positivity of the largest eigenvalue
of the operator

L + 3H(), as dened in Sec. 3.1.2.
The causal structure of the eternally inating spacetime and the topology of the reheating
surface can be visualized using the construction of eternal comoving points [159]. These are
comoving worldlines x = const that forever remain within the inating domain and never enter
the reheating epoch. These worldlines correspond to places where the reheating surface reaches
t = in a spacetime diagram (see Fig. 5.1). It was shown in Ref. [159] using topological
29
4 Predictions and measure issues
arguments that the presence of inating domains at arbitrarily late times entails the existence
of innitely many such eternal points. The set of all eternal points within a given three-
dimensional spacelike slice is a measure zero fractal set. The fractal dimension of this set can
be understood as the fractal dimension of the inating domain [160, 159, 146] and is invariant
under any smooth coordinate transformations in the spacetime.
The existence of eternal points can be used as another invariant criterion for the presence
of eternal ination. The probability X() of having an eternal point in an initial Hubble-size
region with eld value can be found as the solution of a gauge-invariant, nonlinear diusion
equation [159]
D

X +v

X 3H (1 X) ln (1 X) = 0, (4.2)
with zero boundary conditions, X(

) = 0 at the reheating boundary and X(

Pl
= 0) at the
Planck boundary if any. Eternal ination is present if there exists a nontrivial solution X() , 0
of Eq. (4.2).
4.2 Observer-based measures
In theories where observable parameters
a
are distributed randomly, one would like to pre-
dict the values of
a
most likely to be observed by a random (or typical) observer. More
generally, one looks for the probability distribution P(
a
) of observing the values
a
. As dis-
cussed in Sec. 2.2.1, considerations of this type necessarily involve some form of the principle
of mediocrity [3]. On a more formal level, one needs to construct an ensemble of the possible
observers and to dene a probability measure on this ensemble. In inationary cosmology,
observers appear only along the reheating surface. If eternal ination is present, the reheating
surface contains innitely many causally disconnected and (possibly) statistically inequivalent
domains. The principal diculties in the probabilistic approach are due to a lack of a natural
denition of measure on such surfaces.
1
Existing proposals for an observer-based measure fall in two major classes, which may be
designated as volume-based vs. worldline-based. The dierence between these classes is in
the approach taken to construct the ensemble of observers. In the volume approach [133, 3,
71, 185, 175, 186], the ensemble contains every observer appearing in the universe, at any time
or place. In the worldline approach [187, 178, 188, 189], the ensemble consists of observers
appearing near a single, randomly selected comoving worldline x = const; more generally, an
arbitrary timelike geodesic could also be used. If the ensemble contains innitely many observers
(this is typically the case for volume-based ensembles), a regularization is needed to obtain
specic probability distributions. Finding and applying suitable regularization procedures is a
separate technical issue explored in Sec. 4.3 and 4.4. I begin with a general discussion of these
measure prescriptions.
A number of previously consideredvolume-basedmeasure proposals were found to be lacking
in one aspect or another [133, 3, 145, 190, 191, 187]. One of the problems was the dependence on
the choice of the time gauge [132, 133]. A prescription manifestly free from gauge dependence
is the spherical cuto measure [71]. This prescription provides unambiguous predictions for
models of random-walk type eternal ination if the reheating condition =

corresponds to a
topologically compact and connected locus in the eld space ,
a
(see Sec. 4.3). For models
1
To avoid confusion, let us note that the recent work [184] proposes a measure in the phase space of trajectories
rather than an observer-based measure in the sense discussed here.
30
4.2 Observer-based measures
where the reheating condition is met at several disconnected loci in eld space (tunneling-
type eternal ination belongs to this class), one can use the recently proposed prescription of
comoving cuto [175, 186]. Recently, the equal-time cuto with e-folding time (the scale
factor cuto) has received renewed attention [88, 192, 193]. We refer to these prescriptions
simply as the volume-based measures for the purposes of rough classication.
Similarly, existing measure proposals of the worldline type appear to converge essentially
to a single prescription [178, 7] (however, see [194, 195] for the most recent developments). We
refer to this prescription as the worldline-based measure.
The main dierence between the worldline-based and volume-based measures is in their de-
pendence on the initial state. When considering the volume-based measure, one starts from a
nite initial spacelike 3-volume. (Final results are insensitive to the choice of this 3-surface in
spacetime or to its geometry.) The initial state consists of the initial values of the elds
a
within the initial volume, and possibly a label corresponding to the type of the initial bubble.
When considering the worldline-based measure, one assumes knowledge of these data at the
initial point of the worldline.
2
It turns out that the volume-based measures always yield results
that are independent of the initial conditions. This agrees with the concept of the stationarity
of a self-reproducing universe [132, 133, 1]; the universe forgets the initial state in the course of
eternal self-reproduction. In contrast, probabilities obtained using the worldline-based measure
always depend on the initial state (except for the case of a nonterminal landscape, i.e. a land-
scape scenario without terminal vacua). A theory of initial conditions is necessary to obtain a
specic prediction from the worldline-based measure.
At this time, there is no consensus as to which of the two classes of measures is the physi-
cally relevant one. The present author regards the two measures as reasonable answers to two
dierently posed questions. The rst question is to determine the probability distribution for
observed values of
a
, given that the observer is randomly chosen from all the observers present
in the entire spacetime. Since we have no knowledge as to the total duration of ination in
our past or the total number of bubble nucleations preceding the most recent one, it appears
reasonable to include in the ensemble all the observers that will ever appear anywhere in the
spacetime. The answer to the rst question is thus provided by the volume-based measure.
The second question is posed in a rather dierent manner. In the context of tunneling-type
eternal ination, upon discovering the type of our bubble we may wish to leave a message
to a future civilization that may arise in our future after an unspecied number of nested
bubble nucleations. The analogous situation in the context of random-walk type ination is a
hypothetical observer located within an inating region of spacetime who wishes to communicate
with future civilizations that will eventually appear when ination is over. The only available
means of communication is leaving a message in a sealed box. The message might contain
the probability distribution P(
a
) for parameters
a
that we expect the future civilization to
observe. In this case, the initial state is known at the time of writing the message. It is clear
that the message can be discovered only by future observers near the worldline of the box. It
is then natural to choose the ensemble of observers appearing along that worldline. Starting
from the known initial state, one would then compute P(
a
) according to the worldline-based
measure.
Calculations using the worldline-based measure usually do not require regularization (except
2
Naturally, it is assumed that the initial state is in the self-reproduction regime: For random-walk type models,
the initial 3-volume is undergoing ination rather than reheating and, for tunneling models, the initial 3-volume
is not situated within a terminal bubble.
31
4 Predictions and measure issues

(1)

(2)

1

2
Figure 4.1: Illustrative inationary potential with a at self-reproduction regime
1
< <
2
and deterministic regimes
(1)

< <
1
and
2
< <
(2)

.
for the case of nonterminal landscapes) because the worldline-based ensemble of observers is
almost surely nite [178]. For instance, in random-walk type models the worldline-based measure
predicts the exit probability distribution p
exit
, which can be computed by solving a suitable
dierential equation [see Eq. (3.38)]. However, the ensemble used in the volume-based measure is
innite and requires a regularization. Known regularization methods are reviewed in Sections 4.3
and 4.4.
A simple toy model [3] where the predictions of the volume-based measures can be obtained
analytically is a slow-roll scenario with a potential shown in Fig. 4.1. The potential is at in the
range
1
< <
2
where the evolution is diusion-dominated, while the evolution of regions
with >
2
or <
1
is completely deterministic (uctuation-free). It is assumed that the
diusion-dominated range
1
< <
2
is suciently wide to cause eternal self-reproduction of
inating regions. There are two thermalization points, =
(1)

and =
(2)

, which may be
associated to dierent types of true vacuum and thus to dierent observed values of cosmological
parameters. The question is to compare the volumes 1
1
and 1
2
of regions thermalized into these
two vacua. Since there is an innite volume thermalized into either vacuum, one looks for the
volume ratio 1
1
/1
2
.
The potential is symmetric in the range
1
< <
2
, so it is natural to assume that Hubble-
size regions exiting the self-reproduction regime at =
1
and at =
2
are equally abundant.
Since the evolution within the ranges
(1)

< <
1
and
2
< <
(2)

is deterministic, the
regions exiting the self-reproduction regime at =
1
or =
2
will be expanded by xed
amounts of e-foldings, which we may denote N
1
and N
2
respectively,
N
j
= 4
_

(j)

j
H()
H

()
d. (4.3)
Therefore the volume of regions thermalized at =
(j)

, where j = 1, 2, will be increased by

the factors exp(3N
j
). Hence, the volume ratio is
1
1
1
2
= exp (3N
1
3N
2
) . (4.4)
This answer is found in several volume-based measures except the scale-factor cuto where one
nds V
1
/V
2
= 1.
32
4.3 Regularization for a single reheating surface
Figure 4.2: A 1+1-dimensional slice of spacetime in a two-eld inationary model (numerical
simulation in Ref. [185]). Shades denote dierent values of the eld , which takes
values in the periodically identied interval [0, 2/]. The white region represents
the thermalized domain. The boundary of the thermalized domain is the reheating
surface (cf. Fig. 5.1), which contains all the possible values of the eld .
4.3 Regularization for a single reheating surface
The task at hand is to dene a measure that ascribes equal weight to each observer ever appearing
anywhere in the universe. As discussed in Sec. 2.2.1, it is sucient to construct a measure 1(
a
)
of the 3-volume along the reheating surface. The volume-based measure P(
a
) will then be given
by Eq. (2.22). In the presence of eternal ination, the proper 3-volume of the reheating surface
diverges even when we limit the spacetime domain under consideration to the comoving future
of a nite initial spacelike 3-volume. Therefore, the reheating surface needs to be regularized.
In this section we consider the case when the reheating condition is met at a topologically
compact and connected locus in the conguration space ,
a
. In this case, every connected
component of the reheating surface in spacetime will contain all the possible values of the elds

a
, and all such connected pieces are statistically equivalent. Hence, it suces to consider a
single connected piece of the reheating surface. A situation of this type is illustrated in Fig. 4.2.
A sketch of the random walk in conguration space is shown in Fig. 4.3.
A simple regularization scheme is known as the equal-time cuto. One considers the part
of the reheating surface formed before a xed time t
max
; that part is nite as long as t
max
is
nite. Then one can compute the distribution of the quantities of interest within that part of
the reheating surface. Subsequently, one takes the limit t
max
. The resulting distribution
can be found from the solution P
( )
V
() of the stationary FP equation,
[

L + 3H()]P
( )
V
= P
( )
V
, (4.5)
with the largest eigenvalue (see Sec. 3.1.2). However, both the eigenvalue and the distri-
33
4 Predictions and measure issues

Figure 4.3: A random walk in conguration space for a two-eld inationary model considered
in Ref. [185]. The center of the eld space is a diusion-dominated regime. The
reheating condition, =

, selects a compact and connected region (a circle) in

conguration space. The problem is to determine the volume-weighted probability
distribution for the values of at reheating.
bution P
( )
V
() depend rather sensitively on the choice of the equal-time hypersurfaces. Since
there appears to be no preferred choice of the cuto hypersurfaces in spacetime, the equal-time
cuto cannot serve as an unbiased measure. Also, it was shown in Ref. [146] that the unbiased
result (4.4) cannot be obtained via an equal-time cuto with any choice of the time gauge.
Recently, the equal-time cuto with t chosen as the number of e-foldings (the scale factor
time, t = ln a) received new attention [88, 192]. Previously this cuto was considered problematic
due to the arbitrariness of choosing the e-folding time rather than any other coordinate for t. A
tentative fundamental motivation of the scale factor cuto was presented in Ref. [196].
The spherical cuto measure prescription [71] regularizes the reheating surface in a dierent
way. A nite region within the reheating surface is selected as a spherical region of radius R
around a randomly chosen center point. (Since the reheating 3-surface is spacelike, the distance
between points can be calculated as the length of the shortest path within the reheating 3-
surface.) Then the distribution of the quantities of interest is computed within the spherical
region. Subsequently, the limit R is evaluated. Since every portion of the reheating
surface is statistically the same, the results are independent of the choice of the center point.
The spherical cuto is gauge-invariant since it is formulated entirely in terms of the intrinsic
properties of the reheating surface.
While the spherical cuto prescription successfully solves the problem of regularization, there
is no universally applicable analytic formula for the resulting distribution. Application of the
spherical cuto to general models of random-walk ination requires a direct numerical simulation
of the stochastic eld dynamics in the inationary spacetime. Such simulations were reported in
Refs. [160, 133, 197, 185] and used the Langevin equation (3.3) with a specic stochastic ansatz
for the noise eld N(t, x).
Apart from numerical simulations, the results of the spherical cuto method may be obtained
analytically in a certain class of models [185]. One such case is a multi-eld model where the
34
4.4 Regularization for multiple types of reheating surfaces
potential V (,
a
) is independent of
a
within the range of where thediusionin dominates.
Then the distribution in
a
is at when eld exits the regime of self-reproduction and resumes
the deterministic slow-roll evolution. One can derive a gauge-invariant Fokker-Planck equation
for the volume distribution P
V
(,
a
), using as the time variable [185],

P
V
=

_
D
v

P
V
_

_
v

P
V
_
+
3H
v

P
V
, (4.6)
where D, v

, and v

are the coecients of the FP equation (3.15). This equation is valid for the
range of where the evolution of is free of uctuations. By solving Eq. (4.6), one can calculate
the volume-based distribution of
a
predicted by the spherical cuto method as P
V
(

, ). Note
that the mentioned restriction on the potential V (,
a
) is important. In general, the eld
cannot be used as the time variable since the surfaces of constant are not everywhere spacelike
due to large uctuations of in the diusion-dominated regime.
4.4 Regularization for multiple types of reheating surfaces
Let us begin by considering a simpler example: an inationary scenario with an asymmetric
slow-roll potential V () having two minima
(1)

,
(2)

. This scenario has two possibilities for

thermalization, possibly diering in the observable parameters
a
. More generally, one may
consider a scenario with n dierent minima of the potential, possibly representing n distinct
reheating scenarios. It is important that the minima
(j)

, j = 1, ..., n are topologically dis-

connected in the conguration space. This precludes the possibility that dierent minima are
reached within one connected component of the reheating hypersurface in spacetime. Addition-
ally, the elds
a
may uctuate across each connected component in a way that depends on the
minimum j. Thus, the distribution P(
a
; j) of the elds
a
at each connected component of the
reheating surface may depend on j. In other words, the dierent components of the reheating
surface may be statistically inequivalent with respect to the distribution of
a
on them. To use
the volume-based measure for making predictions in such models, one needs a regularization
method that is applicable to situations with a large number of disconnected and statistically
inequivalent components of the reheating hypersurface.
In such situations, the spherical cuto prescription (see Sec. 4.3) yields only the distribution
of
a
across one connected component, since the sphere of a nite radius R will never reach
any other components of the reheating surface. Therefore, the spherical cuto needs to be
supplemented by a weighting prescription, which would assign a weights p
j
to the minimum
labeled j. In scenarios of tunneling type, such as the string theory landscape, observers may
nd themselves in bubbles of types = 1, ..., N. Again, a weighting prescription is needed to
determine the probabilities p

of being in a bubble of type . The duration of time spent in a

bubble is presumably irrelevant for the probability of being in the bubble, since civilizations can
appear only for a (cosmologically) short time after reheating (see Fig. 2.3).
Two dierent weighting prescriptions have been formulated, using the volume-based [175,
186] and the worldline-based approach [178], respectively. The rst prescription is called the
comoving cuto while the second the worldline or the holographic cuto. For clarity, we
illustrate these weighting prescriptions on models of tunneling type where innitely many nested
bubbles of types = 1, ...N are created and where some bubbles are terminal, i.e. contain no
daughter bubbles. In the volume-based approach, each bubble receives equal weight in the
35
4 Predictions and measure issues
t
x
Figure 4.4: Weighting prescriptions for models of tunneling type. Shaded regions are bubbles of
dierent types. Dashed vertical lines represent randomly chosen comoving geodesics
used to dene a nite subset of bubbles.
ensemble; in the worldline-based approach, only bubbles intersected by a selected worldline are
counted and given equal weight. Let us now examine these two prescriptions in more detail.
Since the set of all bubbles is innite, one needs to perform a regularization, that is, one
needs to select a very large but nite subset of bubbles. The weight p

will be calculated as
the fraction of bubbles of type within the selected subset; then the number of bubbles in
the subset will be taken to innity. Technically, the two prescriptions dier in the details of
the regularization. The dierence between the two prescriptions can be understood pictorially
(Fig. 4.4). In a spacetime diagram, one draws a nite number of timelike comoving geodesic
worldlines emitted from an initial 3-surface towards the future. (It can be shown that the results
are independent of the choice of these lines, as long as that choice is uncorrelated with the bubble
nucleation process [186].) Each of these lines will intersect only a nite number of bubbles, since
the nal state of any worldline is (almost surely) a terminal bubble. The subset of bubbles
needed for the regularization procedure is dened as the set of all bubbles intersected by at
least one line. At this point, the volume-based approach assigns equal weight to each bubble in
the subset, while the worldline-based approach assigns equal weight to each bubble along each
worldline. As a result, the volume-based measure counts each bubble in the subset only once,
while the worldline-based measure counts each bubble as many times as it is intersected by some
worldlines. After determining the weights p

by counting the bubbles as described, one increases

the number of worldlines to innity and evaluates the limit values of p

.
It is clear that the volume-based measure represents the counting of bubbles in the entire
universe, and it is then appropriate that each bubble is being counted only once. On the other
hand, the worldline-based measure counts bubbles occurring along a single worldline, ignoring
the bubbles produced in other parts of the universe and introducing an unavoidable bias due to
the initial conditions at the starting point of the worldline.
Explicit formulas for p

were derived for a tunneling-type scenario (with terminal vacua) in

36
4.5 The Youngness paradox and the Boltzmann brains
the volume-based approach [175],
p

H
q

, (4.7)
where

is the matrix of nucleation rates, q and s

are the quantities dened by Eq. (3.42),

and H

is the Hubble parameter in bubbles of type . The expressions for p

obtained from the

worldline-based measure are given by Eq. (3.48). As we have noted before, the volume-based
measure assigns weights p

that are independent of initial conditions, while the weights obtained

from the worldline-based measure depend sensitively on the type of bubble where the counting
begins.
In the case of a non-terminal landscape, both the volume-based and the world-line based
measures give identical results for p

, which coincide with the mean frequency (3.43) of visiting

a bubble of type [177, 178].
With the weighting prescriptions just described, the volume-based and the worldline-based
measure proposals can be considered complete. In other words, we have two alternative prescrip-
tions that can be applied (in principle) to arbitrary models of random-walk or tunneling-type
eternal ination. Further research is needed to reach a denite conclusion concerning the via-
bility of these measure prescriptions.
4.5 The Youngnessparadox and the Boltzmann brains
One way to choose between the competing measure prescriptions is to examine their predictions
for various paradoxical answers. At least two sources of problems for the measure prescriptions
have been uncovered (for more discussion, see Ref. [7]).
One possible problem is the so-called youngnessparadox [158, 71, 157, 198]. It is manifested
in some volume-based measures, notably in the equal-time cuto by proper time t, by the pre-
dominance of universes where ination ended very recently [157]. Indeed, a tiny and improbable
uctuation that delays reheating by a time t is rewarded by a huge additional expansion fac-
tor, exp (3H
big
t), where H
big
is a ducial value of the highest Hubble rate available in the
model. While H
big
must be below the Planck scale M
Pl
, values such as H
big
10
6
M
Pl
are
already sucient to predict that the most likely temperature of the CMB is much higher than
the observed 2.7

K. For this reason, measure prescriptions exhibiting the youngness paradox

are considered to be unsuitable.
Another pathology that appears in the results of several measure prescription is the so-called
Boltzmann brain problem [199, 200, 201, 88, 202, 203, 198, 204]. The root of the problem is
the thermal property of de Sitter space that behaves as a thermal bath [41] with the Gibbons-
Hawking temperature T =
1
2
H. Thermal uctuations are able to produce any classical system,
for instance, an Earth-like planet with observers on it, with a small but nonzero probability.
Such rare events will nevertheless occur in eternally inating de Sitter spacetime, in the interior
of every recyclable bubble. Hence, one can expect that innitely many observers will be created
who do not see a universe full of galaxies around them, but instead nd themselves surrounded
by empty de Sitter space. An extreme case is an isolated brain nucleating out of thermal
uctuaitons in empty space (a Boltzmann brain). It appears reasonable to reject a multiverse
measure that predicts a large probability for us being Boltzmann brains rather than ordinary
observers.
Recent works investigated these issues for many of the currently considered measure propos-
als [7, 88, 11, 198, 192, 205, 193]. Below I will investigate whether the youngness paradox and
37
4 Predictions and measure issues
the Boltzmann brain problem are present in a new class of multiverse measures, to be presented.
38
5 A new measure for multiverse cosmology
In a series of publications [206, 207, 208] I proposed and studied a new class of volume-weighted
probability measures applicable to any cosmological multiverse scenarios. The new measure,
called the reheating-volume (RV) cuto, calculates the distribution of observable quantities
on an ensemble of multiverses that are conditioned to be nite (e.g. have a nite reheating
hypersurface), as wil be fully explained below. In this chapter I introduce the RV measure and
discuss its general properties. By considering slow-roll inationary models with a scalar inaton,
I show that the RV measure is gauge-invariant, does not suer from the youngness paradox,
and is independent of initial conditions at the beginning of ination. The application of the RV
cuto to random-walk ination will be treated more thoroughly in Chapter 6. Then in Chapter 7
I will apply the RV measure to landscape scenarios with bubble nucleation.
As I described in the previous chapters, eternal ination gives rise to innitely many causally
disconnected regions of the spacetime where the cosmological observables may have signicantly
dierent values. One hopes to obtain probability distribution of the cosmological parameters as
measured by an observer randomly located in the spacetime. The main diuculty in obtaining
such probability distributions is due to the innite volume of regions where an observer may be
located.
Since the universe is cold and empty during inationary evolution, observers may appear only
after reheating. The standard cosmology after reheating is tightly constrained by current exper-
imental knowledge. Hence, the average number of observers produced in any freshly-reheated
spatial domain is a function of cosmological parameters in that domain. Calculating that func-
tion is, in principle, a well-dened astrophysical problem that does not involve any innities.
Therefore we focus on the problem of obtaining the probability distribution of cosmological
observables at reheating.
The set of all spacetime points where reheating takes place is a spacelike three-dimensional
hypersurface [45, 3, 209] called the reheating surface. The hallmark feature of eternal in-
ation is that a nite, initially inating spatial 3-volume typically gives rise to a reheating
surface having an innite 3-volume (see Fig. 5.1). The geometry and topology of the reheat-
ing surface is quite complicated. For instance, the reheating surface contains innitely many
future-directed spikes around never-thermalizing comoving worldlines called eternally inating
geodesics[159, 176, 195]. It is known that the set of spikeshas a well-dened fractal dimension
that can be computed in the stochastic approach [159]. Since the reheating surface is a highly
inhomogeneous, noncompact 3-manifold without any symmetries, a random location on such
a surface is mathematically ill-dened. This feature of eternal ination is at the root of the
technical and conceptual diculties known collectively as the measure problem.
To visualize the measure problem, it is convenient to consider an initial inating spacelike
region S of horizon size (an H-region) and the portion R R(S) of the reheating surface that
corresponds to the comoving future of S. If the 3-volume of R were nite, the volume-weighted
average of any observable quantity Q at reheating would be dened simply by averaging Q over
39
5 A new measure for multiverse cosmology
t
x
Figure 5.1: A 1+1-dimensional slice of the spacetime in an eternally inating universe (numerical
simulation in Ref. [185]). Shades of dierent color represent dierent regions where
reheating took place. The reheating surface is the line separating the white (inating)
domain and the shaded domains.
R,
Q)
_
R
Q

d
3
x
_
R

d
3
x
, (5.1)
where is the induced metric on the 3-surface R. This would have been the natural prescription
for the observer-based average of Q; all higher moments of the distribution of Q, such as

Q
2
_
,

Q
3
_
, etc., would have been well-dened as well. However, in the presence of eternal ination
1
the 3-volume of R is innite with a nonzero probability X(
0
), where =
0
is the initial
value of the inaton eld at S. The function X(
0
) has been computed in slow-roll inationary
models [159] where typically X(
0
) 1 for
0
not too close to reheating. In other words, the
volume of R is innite with a probability close to 1. In that case, the straightforward average (5.1)
of a uctuating quantity Q(x) over R is mathematically undened since
_
R

d
3
x = and
_
R
Q

d
3
x = .
The average Q) can be computed only after imposing a volume cuto on the reheating surface,
making its volume nite in a controlled way. What has become known in cosmology as the
measure problem is the diculty of coming up with a physically motivated cuto prescription
(informally called a measure) that makes volume averages Q) well-dened.
Volume cutos are usually implemented by restricting the innite reheating domain R to a
large but nite subdomain having a volume 1. Then one denes the regularized distribution
p(Q[1) of an observable Q by gathering statistics about the values of Q over the nite volume
1. More precisely, p(Q[1)1dQ is the 3-volume of regions (within the nite domain 1) where
the observable Q has values in the interval [Q, Q +dQ]. The nal probability distribution p(Q)
1
Various equivalent conditions for the presence of eternal ination were examined in more detail in Refs. [159, 146]
and [209]. Here I adopt the condition that X() is nonzero for all in the inating range.
40
is then dened as
p(Q) lim
V
p(Q[1), (5.2)
provided that the limit exists.
A cuto prescription is a specic choice of the compact subset 1 and of the way 1 approaches
innity when the cuto is removed. It has been found early on (e.g. [133, 3]) that p(Q) depends
sensitively on the choice of the cuto. Without a natural mathematical denition of the measure,
one judges a cuto prescription viable if its predictions are not obviously pathological. Possible
pathologies include the dependence on choice of spacetime coordinates [145, 190], the youngness
paradox [158, 71, 157, 198], and the Boltzmann brain problem [201, 88, 202, 203, 198, 204].
Several cutos have been proposed in the literature, diering in the choice of the compact
subset 1 and in the way 1 approaches innity. It has been found early on (e.g. [133, 3])
that probability distributions, such as p(Q), depend sensitively on the choice of the cuto.
This is the root of the measure problem. Since a natural mathematically consistent de-
nition of the measure is absent, one judges a cuto prescription viable if its predictions are
not obviously pathological. Possible pathologies include the dependence on choice of spacetime
coordinates [145, 190], the youngness paradox [158, 71], and the Boltzmann brain prob-
lem [199, 200, 201, 88, 202, 203, 198, 204].
The presently viable cuto proposals fall into two rough classes that may be designated as
worldline-based and volume-based measures (a more ne-grained classication of measure
proposals can be found in Refs. [7, 195]). Theworldlineor the holographicmeasure [178, 188]
avoids considering the innite total 3-volume of the reheating surface in the entire spacetime.
Instead it focuses only on the reheated 3-volume of one H-region surrounding a single randomly
chosen comoving worldline. This measure, by construction, is sensitive to the initial conditions
at the location where the worldline starts and is essentially equivalent to performing calculations
with the comoving-volume probability distribution. Proponents of the holographic measure
have argued that the innite reheating surface cannot be considered because the spacetime be-
yond one H-region is not adequately described by semiclassical gravity [188]. However, the
semiclassical approximation was recently shown to be valid in a large class of inationary mod-
els [210]. In my view, an attempt to count the total volume of the reheating surface corresponds
more closely to the goal of obtaining the probability distribution of observables in the entire
universe, as measured by a typical observer (see Refs. [211, 212, 213, 214, 198] for recent
discussions of typicality and accompanying issues). The sensitive dependence of holographic
proposals on the conditions at the beginning of ination also appears to be undesirable. Volume-
based proposals are insensitive to the initial conditions because the 3-volume of the universe is,
in a certain well-dened sense, dominated by regions that spent a long time in the inationary
regime.
2
Existing volume-based proposals include the equal-time cuto [51, 1, 52], the spherical
cuto[71], the comoving horizon cuto[175, 186, 177], the stationary measure[11, 215], the
no-boundarymeasure with volume weighting [216, 217, 218, 219], the pseudo-comoving mea-
sure [88, 193]. Here I propose a new volume-based measure called the reheating-volume (RV)
cuto [206]. The RV cuto is based on the notion of nitely-produced volume distribution,
which is explained in the next section.
2
It has been noted that 3-volume is a coordinate-dependent quantity, and hence statements involving 3-volume
need to be formulated with care [155]. Indeed there exist time foliations where the 3-volume of inationary
space does not grow with time. The issues of coordinate dependence were analyzed in Ref. [146].
41
5 A new measure for multiverse cosmology
5.1 Reheating-volume cuto
In the RV cuto, the reheating surface is not being restricted to an articially chosen domain.
Instead, one simply selects only those initial regions S that, by rare chance, evolve into compact
reheating surfaces R having a nite, xed volume Vol(R) = 1. The ensemble c
V
of such initial
regions S is a nonempty subset of the ensemble c of all initial regions S. The volume-weighted
probability distribution p(Q[c
V
) of a cosmological observable Q in the ensemble c
V
can be
determined through ordinary sampling of the values of Q over the nite volume 1. The RV
cuto denes the probability distribution p(Q) as the limit of p(Q[c
V
) at 1 , provided that
the limit exists.
To develop an approach for practical computations in the RV cuto, let us rst consider the
probability density (1;
0
)d1 of having nite volume Vol(R) [1, 1 + d1] of the reheating
surface R that results from a single H-region with initial value =
0
. This distribution is
normalized to the probability of the event Vol(R) < , namely
_

0
(1;
0
)d1 = Prob (Vol(R) < ) = 1 X(
0
). (5.3)
The probability density (1;
0
) is nonzero since X(
0
) < 1. I call this (1;
0
) the nitely
produced reheated volume (FPRV) distribution. This and related distributions constitute the
mathematical basis of the RV cuto.
Below I will use the Fokker-Planck (or diusion) formalism to derive equations from which
the FPRV distributions can be in principle computed for models of slow-roll ination with a single
scalar eld. Generalizations of (1;
0
) to multiple-eld or non-slow-roll models are straightfor-
ward since the Fokker-Planck formalism is already developed in those contexts (e.g. [69, 210]).
Let us now dene the FPRV distribution for some cosmological observable Q at reheating.
Consider the probability density (1, 1
Q
R
;
0
, Q
0
), where
0
and Q
0
are values of and Q in
the initial H-region, 1 is the total reheating volume, and 1
Q
R
is the portion of the reheating
volume where the observable Q has a particular value Q
R
. The distribution (1, 1
Q
R
;
0
, Q
0
)
as a function of 1
Q
R
at xed and large 1 is sharply peaked around a mean value

1
Q
R
[
V
_
corresponding to the average volume of regions with Q = Q
R
within the total reheated volume
1. Hence, although the full distribution (1, 1
Q
R
;
0
, Q
0
) could be in principle determined, it
suces to compute the mean value

1
Q
R
[
V
_
. One can then expect that the limit
p(Q
R
) lim
V

1
Q
R
[
V
_
1
= lim
V
_

0
(1, 1
Q
;
0
, Q
0
)1
Q
d1
Q
1 Prob(Vol(R) = 1)
(5.4)
exists and is independent of
0
and Q
0
. (Below I will justify this statement more formally.) The
function p(Q
R
) is then interpreted as the mean fraction of the reheated volume where Q = Q
R
. In
this way, the RV cuto yields the volume-weighted distribution for any cosmological observable
Q at reheating.
To obtain a more visual picture of the RV cuto, consider a large number of initially identical
H-regions having dierent evolution histories to the future. A small subset of these initial H-
regions will generate nite reheating surfaces. An even smaller subset of H-regions will have
the total reheated volume equal to a given value 1. Conditioning on a nite value 1 of the
reheating volume, one obtains a well-dened statistical ensemble E
V
of initial H-regions. For
large 1, the ensemble E
V
can be pictured as a set of initial H-regions that happen to be located
very close to some eternally inating worldlines but do not actually contain any such worldlines
42
5.1 Reheating-volume cuto
x
t
Figure 5.2: A schematic representation of the ensemble E
V
in comoving coordinates (t, x).
Lightly shaded vertical strips represent the comoving future of various initial H-
regions from E
V
; dark shades represent reheated domains; the boundary of the dark-
shaded domains is the reheating surface. Vertical dashed lines are the eternally
inating comoving worldlines that never cross the reheating surface. The 3-volumes
of the reheating surfaces in the comoving future of the pictured H-regions are large
but nite because these H-regions are located near eternal worldlines but do not
contain any such worldlines.
(see Fig. 5.2). In this way, the ensemble E
V
samples the total reheating surface near the spikes
where an innite reheated 3-volume is generated from a nite initial 3-volume. It is precisely
near these spikes that one would like to sample the distribution of observable quantities along
the reheating surface. Therefore, one expects that the ensemble E
V
(in the limit of large 1)
provides a representative sample of the innite reheating surface, despite the small probability of
the event Vol(R) = 1. In this sense, the ensemble E
V
at large 1 is designed to yield a controlled
approximation to the innite reheating surfaces R.
The RV cuto proposed here has several attractive features. By construction, the RV cuto is
coordinate-invariant; indeed, only the intrinsically dened 3-volume within the reheating surface
R is used, rather than the 3-volume within a coordinate-dependent 3-surface. The results of the
RV cuto are also independet of initial conditions. This independence is demonstrated more
formally below and can be understood heuristically as follows. The evolution of regions S
conditioned on a large (but nite) value of Vol(R) is dominated by trajectories that spend a
long time in the high-H inationary regime and thereby gain a large volume. These trajectories
forget about the conditions at their initial points and establish a certain equilibrium distribution
of values of Q on the reheating surface. Hence, one can expect that the distribution of observables
within the reheating domain R will be independent of the initial conditions in S.
The youngness paradox arises in some volume-based prescriptions because H-regions with
delayed reheating are rewarded by an exponentially large additional volume expansion. However,
the RV measure groups together the H-regions that produce equal nal reheated volume V ;
a delay in reheating is not rewarded but suppressed by the small probability of a quantum
uctuation at the end of ination. Therefore, most of these H-regions have normal slow-roll
evolution before reheating. For this reason, the youngness paradox is absent in the RV measure.
A more explicit calculation conrming this conclusion will be given in Sec. 7.1.
43
5 A new measure for multiverse cosmology
5.2 RV cuto in slow-roll ination
As a rst specic application, I implement the RV measure in a slow-roll inationary model with
a scalar inaton and the action
S =
_ _

R
16G
+
1
2
(

)
2
V ()
_

gd
4
x. (5.5)
If the energy density is dominated by the potential energy V (), the Hubble expansion rate H
is approximately given by
H
_
8G
3
V (). (5.6)
In the stochastic approach to ination,
3
the semiclassical dynamics of the eld averaged over
an H-region is a superposition of a deterministic motion with velocity

= v()
1
4G
H
,
(5.7)
and a random walk with root-mean-squared step size
_
)
2
=
H()
2

2D()
H()
, D
H
3
8
2
, (5.8)
during time intervals t = H
1
. A convenient description of the evolution of the eld at time
scales t H
1
is
(t +t) = (t) +v()t +(t)
_
2D()t, (5.9)
where (t) is (approximately) a white noise variable,
) = 0,

(t)(t

)
_
= (t t

), (5.10)
which is statistically independent for dierent H-regions [138]. This stochastic process describes
the evolution (t) along a single comoving worldline. For convenience, we assume that ination
ends in a given H-region when reaches a xed value =

.
Consider an ensemble of initial H-regions S
1
, S
2
, ..., where the inaton eld is homogeneous
and has value =
0
within the inationary regime. Following the spacetime evolution of the
eld in each of the regions S
j
along comoving geodesics, we arrive at reheating surfaces R
j
where =

. Most of the surfaces R

j
will have innite 3-volume; however, some (perhaps small)
subset of S
j
will have nite R
j
. The
0
-dependent probability, denoted

X(
0
) 1 X(
0
), of
having a nite volume of R
j
is a solution of the gauge-invariant equation [159]
D()

X
,
+v()

X
,
+ 3H()

X ln

X = 0, (5.11)
with the boundary conditions

X(

) = 1 and

X(
Pl
) = 1 at reheating and at Planck boundaries.
While

X() 1 is always a solution of Eq. (5.11), the existence of a nontrivial solution with
0 <

X() < 1 indicates the possibility of eternal ination. The gauge invariance of Eq. (5.11)
is manifest since a change of the time variable, (t)
_
t
T()dt, results in dividing the three
coecients D, v, H by T() [145], which leaves Eq. (5.11) unchanged.
3
See Refs. [32, 33, 35] for early works and Refs. [133] and [8] for pedagogical reviews.
44
5.2 RV cuto in slow-roll ination
The probability distribution (1;
0
) can be found by considering a suitable generating func-
tion. Let us dene the generating function g(z;
0
) by
g(z;
0
)

e
zV
_
V<

_

0
e
zV
(1;
0
)d1. (5.12)
(Note that the formal parameter z has the dimension of inverse volume. The parameter z can
be made dimensionless by a trivial rescaling which we omit.) The function g(z;
0
) is analytic
in z and has no singularities for Re z 0. Moments of the distribution (1;
0
) are determined
as usual through derivatives of g(z;
0
) in z at z = 0, while (1,
0
) itself can be reconstructed
through the inverse Laplace transform of g(z;
0
) in z.
The generating function g(z; ) has the following multiplicative property: For two statistically
independent H-regions that have initial values =
1
and =
2
respectively, the sum of the
(nitely produced) reheating volumes 1
1
+1
2
is distributed with the generating function
_
e
z(V
1
+V
2
)
_
=

e
zV
1
_
e
zV
2
_
=g(z;
1
)g(z;
2
). (5.13)
This multiplicative property is the only assumption in the derivation of Eq. (5.11) in Ref. [159].
Hence, g(z; ) satises the same equation (we drop the subscript 0 in
0
),
Dg
,
+vg
,
+ 3Hg ln g = 0. (5.14)
The boundary condition at

is g(z;

) = e
zH
3
()
since an H-region with =

is already
reheating and has volume H
3
(

). The boundary condition at the Planck boundary

Pl
(or
other boundary where the eective eld theory breaks down) is absorbing, i.e. we assume that
regions with =
Pl
disappear and never generate any reheating volume: g(z;
Pl
) = e
z0
= 1.
Note that the variable z enters Eq. (6.49) as a parameter and only through the boundary
conditions. At z = 0 the solution of Eq. (6.49) is g(0; ) =

X(). Explicit approximate solutions
of Eq. (6.49) can be obtained using the methods developed in Ref. [159].
Let us now consider FPRV distributions of cosmological parameters Q. The generating func-
tion for the distribution (1, 1
Q
R
; , Q) discussed above is
g(z, q; , Q)
__
e
zVqV
Q
(1, 1
Q
; , Q)d1d1
Q
. (5.15)
The equation for g(z, q; , Q) is derived similarly to Eq. (6.49) and is of the form
D

g
,
+D
Q
g
,QQ
+v

g
,
+v
Q
g
,Q
+ 3H g ln g = 0, (5.16)
where D

, D
Q
, v

, v
Q
are the suitable kinetic coecients representing the diusion and the
mean drift velocity of and Q. The boundary condition at =

is
g(z, q;

, Q) = exp
_
(z +q
QQ
R
) H
3
(

, (5.17)
where we use the delta-symbol dened by
QQ
R
= 1 if Qbelongs to a narrowinterval [Q
R
, Q
R
+dQ]
and
QQ
R
= 0 otherwise.
To obtain the distribution (5.4), we need to compute the average

1
Q
R
[
V
_
at xed 1. We
dene the auxiliary generating function
h(z; , Q)

1
Q
R
e
zV
_
V<
= g
,q
(z, q = 0; , Q). (5.18)
45
5 A new measure for multiverse cosmology
Note that g(z, q = 0; , Q) = g(z; ). The equation for h(z; , Q) then follows from Eq. (5.16),
D

h
,
+D
Q
h
,QQ
+v

h
,
+v
Q
h
,Q
+ 3H (ln g + 1) h = 0. (5.19)
This linear equation contains as a coecient the function g(z; ), which is the solution of
Eq. (6.49). The boundary condition for Eq. (5.19) is
h(z;

, Q) = e
zH
3
()
H
3
(

)(QQ
R
). (5.20)
Here we can use the ordinary -function instead of the symbol
QQ
R
because the -function
enters linearly into the boundary condition. An appropriate rescaling of the distribution h by
the factor dQ is implied when we pass from
QQ
R
to (QQ
R
).
Finally, the expectation value

1
Q
R
[
V
_
at a xed 1 and the limit (5.4) can be found using the
inverse Laplace transform of h(z; , Q) in z.
The computation just outlined allows one, in principle, to obtain quantitative predictions
from the RV measure. Further details and a more direct computational procedure will be given
elsewhere [207]. Presently, let us analyze the limit 1 in qualitative terms. The function
p(Q
R
; 1) is expressed as [cf. Eq. (5.4)]
p(Q
R
; 1)

1
Q
R
[
V
_
1
=
_
i
i
e
zV
h(z; , Q)dz
1
_
i
i
e
zV
g(z; )dz
. (5.21)
The asymptotic behavior of the inverse Laplace transform of h(z; , Q) at large 1 is determined
by the locations of the singularities of h(z; , Q) in the complex z plane. The dominant asymp-
totics of the inverse Laplace transform are of the form exp(z

1), where z

is the singularity
with the smallest [Re z

[. It can be shown that solutions of Eqs. (6.49) and (5.19) cannot diverge
at nite values of or Q. Thus g(z; ) and h(z; , Q) cannot have - or Q-dependent singulari-
ties in z. Moreover, the function g(z; ) cannot have pole-like singularities in z; the only possible
singularities are branch points where the function g(z; ) is nite but a derivative with respect
to z diverges. Furthermore, derivatives
n
z
h satisfy linear equations with coecients depending
on the derivatives
n1
z
g(z; ), which diverge at the singularities of g. Hence the singularities
of
n
z
h in the z plane coincide with those of g(z; ). For these reasons the limit in Eq. (5.4)
exists and is independent of the initial values , Q. A more detailed analysis justifying these
statements will be given in the next chapter.
46
6 The RV measure for random-walk ination
The focus of this chapter is a more detailed study of the RV measure in the context of eternal
ination of random-walk type. As a typical generic model I choose a scenario where ination
is driven by the potential V () of a minimally coupled scalar eld . In this model, there
exists a range of where large quantum uctuations dominate over the deterministic slow-roll
evolution, which gives rise to eternal self-reproduction of inationary domains. I extensively use
the stochastic approach to ination, which is based on the Fokker-Planck or diusionequations
(see Ref. [8] for a pedagogical review). The results can be straightforwardly generalized to
multiple-eld or non-slow-roll models are straightforward since the Fokker-Planck formalism is
already developed in those contexts [69, 210]. Applications of the RV measure to landscape
scenarios will be considered in Chapter 7.
6.1 Motivation
An attractive feature of the RV measure is that its construction lacks extraneous geometric
elements that could introduce a bias. An example of a biased measure is the equal-time cuto
where one considers the subdomain of the reheating surface to the past of a hypersurface of xed
proper time t = t
c
, subsequently letting t
c
. It is well known that the volume-weighted
distribution of observables within a hypersurface of equal proper time is strongly dominated by
regions where ination ended very recently. A time delay t in the onset of reheating due to a
rare quantum uctuation is overwhelmingly rewarded by an additional volume expansion factor
exp[3H
max
t], where H
max
is roughly the highest Hubble rate accessible to the inaton. This
is the essence of the so-called youngness paradox that seems unavoidable in an equal-time cuto
(see Refs. [157] and [198] for recent discussions).
Moreover, the results of the equal-time cuto are sensitive to the choice of the time coordinate
(time gauge). For instance, the proper time can be replaced by the family of time gauges labeled
by a constant , [145]
t
()

_
t
H

dt, (6.1)
which interpolate between the proper time ( = 0, t
(0)
t) and the e-folding time ( = 1,
t
(1)
= ln a). It has been shown that the results of the equal-time cuto depend sensitively on
the value of , and that no correct value of could be specied so as to remove the bias [146].
Since the time coordinate is an arbitrary label in the spacetime, we may impose the requirement
that a viable measure prescription be invariant with respect to choosing even more general time
gauges, such as

_
t
T()dt, (6.2)
where T() > 0 is an arbitrary function of the inaton eld (and possibly of other elds), and
the integration is performed along comoving worldlines x
1,2,3
= const.
47
6 The RV measure for random-walk ination
The spherical cuto [71] and the stationary measure [11] prescriptions were motivated by
the need to remove the bias inherent in the equal-time cuto. In particular, the spherical cuto
selects as a compact subset 1 the interior of a large sphere drawn within the reheating surface
R around a randomly chosen center. The spherical cuto is manifestly gauge-invariant since its
construction uses only the intrinsically dened 3-volume of the reheating surface rather than the
spacetime coordinates (t, x). Some results were obtained in the spherical cuto using numerical
simulations [185]. A disadvantage of the spherical cuto is that its direct implementation requires
one to perform costly numerical simulations of random-walk ination on a spacetime grid, for
instance, using the techniques of Refs. [133, 185, 138]. Instead, one would prefer to obtain a
generally valid analytic formula for the probability distribution of cosmological observables. For
instance, one could ask whether the results of the spherical cuto depend in an essential way
on the spherical shape of the region, on the position of the center of the sphere, and on the
initial conditions. Satisfactory answers to these questions (in the negative) were obtained in
Refs. [71, 185] in some tractable cases where results could be obtained analytically. However, it
is dicult to analyze these questions in full generality since one lacks a general analytic formula
for the probability distribution in the spherical cuto.
The RV measure is similar in spirit to the spherical cuto because the RV cuto uses only the
intrinsic geometrical information dened by the reheating surface. It can be argued that the RV
cuto is more natural than other cutos in that it selects a nite portion 1 of the reheating
surface without using articial constant-time hypersurfaces, spheres, worldlines, or any other
extraneous geometrical data. Instead, the selection of 1 in the RV cuto is performed using a
certain well-dened selection of subensemble in the probability space, which is determined by
the stochastic evolution itself.
The central concept in the RV cuto is the nitely produced volume. The basic idea is
that there is always a nonzero probability that a given initial H-region S does not give rise
to an innite reheating surface in its comoving future. For instance, it is possible that by
a rare coincidence the inaton eld rolls towards reheating at approximately the same time
everywhere in S. Moreover, there is a nonzero (if small) probability (1)d1 that the total volume
Vol(R) of the reheating surface R to the future of S belongs to a given interval [1, 1 +d1],
(1) lim
dV0
Prob Vol(R) [1, 1 +d1]
d1
. (6.3)
I call (1) the nitely produced volume distribution. This distribution is nontrivial because
the probability of the event Vol(R) < is nonzero, if small, for any given (non-reheated) initial
region S. The distribution (1) is, by construction, normalized to that probability:
_

0
(1)d1 = ProbVol(R) < < 1. (6.4)
The RV cuto consists of a selection of a certain ensemble E
V
of the histories that produce
a total reheated volume equal to a given value 1 starting from an initial H-region. In the limit
of large 1, the ensemble E
V
consists of H-regions that evolve almost to the regime of eternal
ination. Thus, heuristically one can expect that the ensemble E
V
provides a representative
sample of the innite reheating surface.
1
Given the ensemble E
V
, one can determine the volume-
weighted probability distribution p(Q[c
V
) of a cosmological parameter Q by ordinary sampling
1
Of course, this heuristic statement cannot be made rigorous since there exists no natural measure on the innite
reheating surface. We use this statement merely as an additional motivation for considering the RV measure.
48
6.1 Motivation
of the values of Q throughout the nite volume 1. Finally, the probability distribution p(Q) is
dened as the limit of p(Q[c
V
) at 1 , provided that the limit exists.
To clarify the construction of the ensemble E
V
, it is helpful to begin by considering the dis-
tribution (1) in a model that does not permit eternal ination. In that case, the volume of the
reheating surface is nite with probability 1, so the distribution (1) is an ordinary probability
distribution normalized to unity. In that context, the distribution (1) was introduced in the
recent work [209] where the authors considered a family of inationary models parameterized by
a number , such that eternal ination is impossible in models where > 1. It was then found
by a direct calculation that all the moments of the distribution (1) diverge at the value = 1
where the possibility of eternal ination is rst switched on. One can show that the nitely pro-
duced distribution (1) for < 1 is again well-behaved and has nite moments (see Sec. 6.2.4).
This FPRV distribution (1) is the formal foundation of the RV cuto. It is worth emphasiz-
ing that the RV cuto does not regulate the volume of the reheating surface by modifying the
dynamics of a given inationary model and making eternal ination impossible. Rather, nite
volumes 1 are generated by rare chance (i.e. within the ensemble E
V
) through the unmodied
dynamics of the model, directly in the regime of eternal ination.
Below I compute the distribution (1) asymptotically for very large 1 in models of slow-roll
ination (Sec. 6.2.4). Specically, I will compute the distribution (1;
0
), where
0
is the
(homogeneous) value of the inaton eld in the initial region S. To implement the RV cuto
explicitly for predicting the distribution of a cosmological parameter Q, it is necessary to consider
the joint nitely produced distribution (1, 1
Q
R
;
0
, Q
0
) for the reheating volume 1(R) and the
portion 1
Q
R
of the reheating volume in which Q = Q
R
. (As before,
0
and Q
0
are the values in
the initial H-region.) If the distribution (1, 1
Q
R
;
0
, Q
0
) is found, one can determine the mean
volume

1
Q
R
[
V
_
while the total reheating volume 1 is held xed,

1
Q
R
[
V
_
=
_
(1, 1
Q
R
;
0
, Q
0
)1
Q
R
d1
Q
R
(1;
0
, Q
0
)
. (6.5)
Then one computes the probability of nding the value of Q within the interval [Q
R
, Q
R
+dQ]
at a random point in the volume 1,
p(Q = Q
R
; 1)

1
Q
R
[
V
_
1
. (6.6)
The RV cuto denes the probability distribution p(Q) for an observable Q as the limit of the
distribution p(Q; 1) at large 1,
p(Q) lim
V

1
Q
R
[
V
_
1
. (6.7)
One expects that this limit is independent of the initial values
0
, Q
0
because the large volume
1 is generated by regions that spent a very long time in the self-reproduction regime and forgot
the initial conditions.
In the previous chapter I derived equations from which the distributions (1, 1
Q
R
;
0
, Q
0
)
and (1;
0
, Q
0
) can be in principle determined. However, a direct computation of the limit
1 (for instance, by a numerical method) will be cumbersome since the relevant probabilities
are exponentially small in that limit. One of the main results of the present chapter is an
analytic evaluation of the limit 1 and a derivation of a more explicit formula, Eq. (6.35),
for the distribution p(Q). The formula shows that the distribution p(Q) can be computed
49
6 The RV measure for random-walk ination
as a ground-state eigenfunction of a certain modied Fokker-Planck equation. The explicit
representation also proves that the limit (6.7) exists, is gauge-invariant, and is independent of
the initial conditions
0
and Q
0
.
It was argued qualitatively in the previous chapter that the RV measure does not suer from
the youngness paradox. In this chapter I demonstrate the absence of the youngness paradox in
the RV measure by an explicit calculation. To this end, I will consider a toy model where every
H-region starts in the uctuation-dominated (or self-reproduction) regime with a constant
expansion rate H
0
and proceeds to reheating via two possible channels. The rst channel consists
of a short period t
1
of deterministic slow-roll ination, yielding N
1
e-folds until reheating; the
second channel has a dierent period t
2
,= t
1
of deterministic ination, yielding N
2
e-folds.
(For simplicity, in this model one neglects uctuations that may return the eld from the slow-
roll regime to the self-reproduction regime, and thus the time periods t
1
and t
2
are sharply
dened.) Thus there are two types of reheated regions corresponding to the two possible slow-
roll channels. The task is to compute the relative volume-weighted probability P(2)/P(1) of
regions of these types within the reheating surface. (Essentially the same model was considered,
e.g., in Refs. [3, 71, 146, 11]. See Fig. 4.1 for a sketch of the potential V ()in this model.)
This toy model serves as a litmus test of measure prescriptions. The holographic or world-
line prescription yields P(2)/P(1) equal to the probability ratio of exiting through the two
channels for a single comoving worldline. This probability ratio depends on the initial condi-
tions. Thus, the worldline measure is (by design) blind to the volume growth during the slow-roll
periods. On the other hand, the volume-weighted prescriptions of Refs. [71, 11] both yield
P(2)
P(1)
=
exp(3N
2
)
exp(3N
1
)
, (6.8)
rewarding the reheated H-regions that went through channel j by the additional volume factor
exp(3N
j
). This ratio is now independent of the initial conditions. For comparison, an equal-time
cuto gives
P(2)
P(1)
=
exp [3N
2
(3H
max

2
) t
2
]
exp [3N
1
(3H
max

2
) t
1
]
. (6.9)
The overwhelming exponential dependence on t
1
and t
2
manifests the youngness paradox:
Even a small dierence t
2
t
1
in the duration of the slow-roll inationary epoch leads to the
exponential bias towards theyoungeruniverses. The bias persists regardless of the choice of the
time gauge [146], essentially because the presence of t
1
and t
2
in the ratio P(2)/P(1) cannot
be eliminated by using a dierent time coordinate.
2
One expects that the RV measure will be
free from this bias because the RV prescription does not involve the time coordinate t at all.
Below (Sec. 6.2.6) I will show that the ratio P(2)/P(1) computed using the RV cuto is indeed
independent of the slow-roll durations t
1,2
. The RV-regulated result [shown in Eq. (6.36) below]
depends only on the gauge-invariant quantities such as N
1
and N
2
and is, in general, dierent
from Eq. (6.8).
2
It should be noted that the youngness bias becomes very small, possibly even negligible, if one uses the number
N of inationary e-foldings as the time variable rather than the proper time t. I am grateful to A. Linde and
A. Vilenkin for bringing this to my attention.
50
6.2 Overview of the results
6.2 Overview of the results
In this section I describe the central results of this chapter; in particular, I develop simplied
mathematical procedures for practical calculations in the RV measure. For convenience of the
reader, the results are stated here without proof, while the somewhat lengthy derivations are
given in Sec. 6.3.
6.2.1 Preliminaries
I consider a model of slow-roll ination driven by an inaton with the action
_ _
R
16G
+
1
2
(

)
2
V ()
_

gd
4
x. (6.10)
In the semiclassical stochastic approach to ination,
3
the semiclassical dynamics of the eld
averaged over an H-region is regarded as a superposition of a deterministic slow roll,

= v()
V
,
()
3H()
=
H
,
4G
, (6.11)
and a random walk with root-mean-squared step size
_
)
2
=
H()
2

2D()
H()
, D
H
3
8
2
, (6.12)
during time intervals t = H
1
, where H() is the function dened by
H()
_
8G
3
V (). (6.13)
A useful eective description of the evolution of the eld at time scales t H
1
can be given
as
(t +t) = (t) +v()t +(t)
_
2D()t, (6.14)
where (t) is a normalized white noise function,
) = 0,

(t)(t

)
_
= (t t

), (6.15)
which is approximately statistically independent between dierent H-regions. This stochastic
process describes the evolution (t) and the accompanying cosmological expansion of space along
a single comoving worldline. For simplicity, we assume that ination ends in a given horizon-size
region when =

, where

is a xed value such that the relative change of H during one

Hubble time t = H
1
becomes of order 1, i.e.

H
,
vH
1
H

=
=

H
2
,
4GH
2

=
1. (6.16)
From the point of view of the stochastic approach, an inationary model is fully specied by
the kinetic coecients D(), v(), H(). These coecients are found from Eqs. (6.11)(6.13)
in models of canonical slow-roll ination and by suitable analogues in other models.
3
See Refs. [32, 33, 35] for early works on the stochastic approach and Refs. [133, 8] for pedagogical reviews.
51
6 The RV measure for random-walk ination
Dynamics of any uctuating cosmological parameter Q is described in a similar way. One
assumes that the value of Q is homogeneous in H-regions. The evolution of Q is described by
an eective Langevin equation,
Q(t +t) = Q(t) +v
Q
(, Q)t +
Q
(t)
_
2D
Q
(, Q)t, (6.17)
where the kinetic coecients D
Q
and v
Q
can be computed, similarly to D and v, from rst prin-
ciples. For simplicity we assume that the noise variable
Q
is independent of the noise used
in Eq. (6.14). A correlated set of noise variables can be considered as well (see e.g. Ref. [210]).
6.2.2 Probability of nite ination
Let us consider an initial H-region S where the inaton eld as well as the parameter Q are
homogeneous and have values =
0
and Q = Q
0
. For convenience we assume that reheating
starts when =

and the Planck energy scales are reached at =

Pl
independently of the
value of Q. (If necessary, the eld variables , Q can be redened to achieve this.)
Although eternal ination to the future of S is almost always the case, it is possible that
reheating is reached at a nite time everywhere to the future of S, due to a rare uctuation. In
that event, the total reheating volume 1 to the future of S is nite. The (small) probability of
that event, denoted by
Prob (1 < [
0
, Q
0
)

X(
0
, Q
0
), (6.18)
can be found as the solution of the following nonlinear equation,
D
H

X
,
+
D
Q
H

X
,QQ
+
v
H

X
,
+
v
Q
H

X
,Q
+ 3

X ln

X = 0, (6.19)

X(
Pl
, Q) = 1,

X(

, Q) = 1, (6.20)
where for brevity we dropped the subscript 0 in
0
and Q
0
. This basic equation, rst derived
in Ref. [159], is of reaction-diusion type and can be viewed as a nonlinear modication of the
Fokker-Planck equations used previously in the literature on the stochastic approach to ination.
While

X(, Q) 1 is always a solution of Eq. (6.19), it is not the correct one for the case of
eternal ination. A nontrivial solution,

X(, Q) , 1, exists and has small values

X(, Q) 1
for , Q away from the thermalization boundary. If the coecients D/H and v/H happen to be
Q-independent, the solution of Eq. (6.19) will be also independent of Q, i.e.

X(, Q) =

X(),
and thus determined by a simpler equation obtained from Eq. (6.19) by omitting derivatives
with respect to Q,
D
H

X
,
+
v
H

X
,
+ 3

X ln

X = 0. (6.21)
It is easy to see that Eqs. (6.19) and (6.21) are manifestly gauge-invariant. Indeed, a change of
time variable according to Eq. (6.2) results in dividing the coecients D, D
Q
, v, v
Q
, H by the
function T() [145], which leaves Eqs. (6.19) and (6.21) unmodied.
Some approximate solutions of Eq. (6.21) were given in Ref. [159], where it was shown that

X() is typically exponentially small for in the inationary regime. While small,

X() is
never zero; hence, there is a well-dened statistical ensemble of initial H-regions that have a
nite total reheating volume in the future. The construction of the RV measure relies on this
fact.
52
6.2 Overview of the results
6.2.3 Finitely produced volume
In a scenario where eternal ination is possible, we now consider the probability density (1;
0
)
of having a nite total reheating volume 1 to the comoving future of an initial H-region with
homogeneous value =
0
(focusing attention at rst on the case of ination driven by a single
scalar eld). The distribution (1;
0
) is normalized to the overall probability

X(
0
) of having
a nite total reheating volume,
_

0
(1;
0
)d1 =

X(
0
). (6.22)
The distribution (1;
0
) can be calculated by rst determining the generating function g(z;
0
),
which is dened by
g(z;
0
)

e
zV
_
V<

_

0
e
zV
(1;
0
)d1. (6.23)
This generating function is a solution of the nonlinear Fokker-Planck equation,

Lg + 3g ln g = 0, (6.24)
where the dierential operator

L is dened by

L
D
H

+
v
H

. (6.25)
In the case of several elds, say and Q, one needs to use the corresponding Fokker-Planck
operator such as

L =
D

+
D
QQ
H

Q

Q
+
v

+
v
Q
H

Q
. (6.26)
The boundary conditions for Eq. (6.24) are
g(z; , Q) = 1 for , Q Planck boundary, (6.27)
g(z; , Q) = e
zH
3
(,Q)

{,Q}reheating boundary
. (6.28)
Note that the parameter z enters the boundary conditions but is not explicitly involved in
Eq. (6.24). Also, the operator

L and Eq. (6.24) are manifestly gauge-invariant with respect to
redenitions of the form (6.2).
The generating function g plays a central role in the calculations of the RV cuto. It will
be shown below that the solution g(z; , Q) of Eq. (6.24) needs to be obtained only at an
appropriately determined negative value of z. This solution can be obtained by a numerical
method or through an analytic approximation if available.
6.2.4 Asymptotics of (1;
0
)
The nitely produced distribution (1; ) can be found through the inverse Laplace transform
of the function g(z; ),
(1; ) =
1
2i
_
i
i
dz e
zV
g(z; ), (6.29)
where the integration contour in the complex z plane can be chosen along the imaginary axis.
The asymptotic behavior of (1; ) at large 1 is determined by the right-most singularity of
53
6 The RV measure for random-walk ination
g(z; ) in the complex z plane. It turns out that the function g(z; ) always has a singularity at
a real, nonpositive z = z

of the type
g(z; ) = g(z

; ) +()

z z

+O(z z

), (6.30)
where z

and () are determined as follows. One considers the (z-dependent) linear operator

L

L + 3(ln g(z; ) + 1), (6.31)
where

L is the Fokker-Planck operator described above. For z > 0 this operator is invertible in
the space of functions f() satisfying zero boundary conditions. The value of z

turns out to be
the algebraically largest real number (in any case, z

0) such that there exists an eigenfunction

() of

L with zero eigenvalue and zero boundary conditions,

L() = 0, (

) = (
Pl
) = 0. (6.32)
The specic normalization of the eigenfunction () can be derived analytically but is unimpor-
tant for the present calculations.
The singularity type shown in Eq. (6.30) determines the leading asymptotic of (1; ) at
1 :
(1; )
1
2

()1
3/2
e
zV
. (6.33)
The explicit form (6.33) allows one to investigate the moments of the distribution (1; ). It
is clear that all the moments are nite as long as z

< 0. However, if z

= 0 all the moments

diverge, namely for n 1 we have
1
n
) =
_

0
(1; )1
n
d1
_

0
1
n3/2
d1 = . (6.34)
The case z

= 0 corresponds to the transition point analyzed in Ref. [209], corresponding to

= 1 in their notation. This is the borderline case between the presence and the absence of
eternal ination. The fact that z

= 0 in the borderline case can be seen directly by noting

that the Fokker-Planck operator

L + 3 has in that case a zero eigenvalue, meaning that the 3-
volume of equal-time surfaces does not expand with time (reheating of some regions is perfectly
compensated by inationary expansion of other regions). In that case, the only solution g(z =
0; ) =

X() of Eq. (6.21) is

X 1 because there are no eternally inating comoving geodesics.
Hence ln g(z = 0, ) = 0, and so the operator

L is simply

L =

L + 3. It follows that the
operator

L also has a zero eigenvalue at z = 0, and thus z = z

= 0 is the dominant singularity

of g(z; ). This argument reproduces and generalizes the results obtained in Ref. [209] where
direct calculations of various moments of (1; ) were performed for the case of the absence of
eternal ination.
We note that the only necessary ingredients in the computation of () is the knowledge
of the singularity point z

and the corresponding function g(z

; ), which is a solution of the

nonlinear reaction-diusion equation (6.24). Determining z

and g(z

; ) in a given inationary
model does not require extensive numerical simulations.
54
6.2 Overview of the results
6.2.5 Distribution of a uctuating eld
Above we denoted by Q a cosmological parameter that uctuates during ination but is in
principle observable after reheating. One of the main questions to be answered using a multiverse
measure is to derive the probability distribution p(Q) for the values of Q observed in a typical
place in the multiverse. I will now present a formula for the distribution p(Q) in the RV cuto.
This formula is signicantly more explicit and lends itself more easily to practical calculations
than the expressions rst shown in Ref. [206].
As in the previous section, we assume that the dynamics of the inaton eld and the
parameter Q is described by a suitable Fokker-Planck operator

L, e.g. of the form (6.26), and
that reheating occurs at =

independently of the value of Q. We then consider Eq. (6.24) for

the function g(z; , Q) and the operator

L

L + 3(ln g + 1); we need to determine the value z

at which g(z; , Q) has a singularity. The operator

L has an eigenfunction with zero eigenvalue

for this value of z. This eigenfunction f
0
(z

; , Q) needs to be determined with zero boundary

conditions (at reheating and Planck boundaries). Then the RV-regulated distribution of Q at
reheating is
p(Q
R
) = const
_
f
0
(z

; , Q)

e
zH
3
H
4
_
=,Q=Q
R
, (6.35)
where the normalization constant needs to be chosen such that
_
p(Q
R
)dQ
R
= 1. The derivation
of this result occupies Sec. 6.3.4.
We note that f
0
is the eigenfunction f
0
of a gauge-invariant operator, and that the result in
Eq. (6.35) depends on the kinetic coecients only through the gauge-independent ratio D/H
times the volume factor H
3
. The distribution p(Q
R
) is independent of the initial conditions,
which is due to a specic asymptotic behavior of the nitely produced volume distributions, as
shown in Sec. 6.3.4.
6.2.6 Toy model of ination
We now apply the RV cuto to the toy model described at the end of Sec. 6.1. We consider a
model of ination driven by a scalar eld with a potential shown in Fig. 4.1. For the purposes of
the present argument, we may assume that there is exactly zero diusion in the deterministic
regimes
(1)

< <
1
and
2
< <
(2)

, while the range

1
< <
2
is suciently
wide to allow for eternal self-reproduction. Thus there are two slow-roll channels that produce
respectively N
1
and N
2
e-folds of slow-roll ination after exiting the self-reproduction regime.
Since the self-reproduction range generates arbitrarily large volumes of space that enter both
the slow-roll channels, the total reheating volume going through each channel is innite. We
apply the RV cuto to the problem of computing the regularized ratio of the reheating volumes
in regions of types 1 and 2.
In this toy model it is possible to obtain the results of the RV cuto using analytic approxi-
mations. The required calculations are somewhat lengthy and can be found in Sec. 6.3.5. The
result for a generic case where one of the slow-roll channels has many more e-folds than the
other (say, N
2
N
1
) can be written as
P(2)
P(1)
O(1)
H
3
(
(2)

)
H
3
(
(1)

)
exp [3N
2
]
exp [3N
1
]
exp [3N
12
] , (6.36)
55
6 The RV measure for random-walk ination

(1)

(2)

1

2
Figure 6.1: A model potential with a at self-reproduction regime
1
< <
2
and deterministic
slow-roll regimes
(1)

< <
1
and
2
< <
(2)

producing N
1
and N
2
inationary
e-folds respectively. In the interval
1
< <
2
the potential V () is assumed to be
constant, V () = V
0
.
where we have dened
N
12


2

2H
2
0
(
2

1
)
2
, H
2
0

8G
3
V
0
. (6.37)
The pre-exponential factor O(1) can be computed numerically, as outlined in Sec. 6.3.5.
We note that the ratio (6.36) is gauge-invariant and does not involve any spacetime coordi-
nates. This result can be interpreted as the ratio of volumes e
3N
1
and e
3N
2
gained during the
slow-roll regime in the two channels multiplied by a correction factor e
3N
12
. The dimensionless
number N
12
can be suggestively interpreted (up to the factor

2) as the mean number of steps

of size
1
2
H
0
required for a random walk to reach the boundary of the at region [
1
,
2
]
starting from the middle point
0

1
2
(
1
+
2
). Since each of the steps of the random walk
takes a Hubble time H
1
0
and corresponds to one e-folding of ination, the volume factor gained
during such a traversal will be e
3N
12
. Note that the correction factor increases the probability
of channel 2 that was already the dominant one due to the larger volume factor e
3N
2
e
3N
1
.
Depending on the model, this factor may be a signicant modication of the ratio (6.8) obtained
in previously used volume-based measures.
6.3 Derivations
6.3.1 Positive solutions of nonlinear equations
It is not easy to demonstrate directly the existence of nontrivial solutions of reaction-diusion
equations such as Eq. (6.19). However, there is a connection between solutions of such non-
linear equations and solutions of the linearized equations. Rigorous results are available in the
mathematical literature on nonlinear functional analysis and bifurcation theory.
Heuristically, consider a solution of Eq. (6.21) that is approximately

X() 1. The equation
can be linearized in the neighborhood of

X 1 as

X = 1 () and yields the Fokker-Planck
(FP) equation
_

L + 3
_
= 0,

L
D
H

+
v
H

. (6.38)
The FP operator

L + 3 is adjoint to the operator
_

+ 3
_
P

_
D
H
P
_

(
v
H
P) + 3P, (6.39)
56
6.3 Derivations
which enters the FP equation for the 3-volume distribution P(, t) in the e-folding time param-
eterization. If eternal ination is allowed in a given model, the operator

L

+ 3 has a positive
eigenvalue. The largest eigenvalue of that operator is zero in the borderline case when eternal
ination is just about to set in. The spectrum of the operator

L + 3 is the same as that of the
adjoint operator

L

+3. Hence, in the borderline case the largest eigenvalue of the operator

L+3
will be zero, and there will exist a nontrivial, everywhere nonnegative solution of Eq. (6.38).
Thus, heuristically one can expect that a nontrivial solution

X() , 1 will exist away from the
borderline case, i.e. when the operator

L + 3 has a positive eigenvalue.
Following the approach of Ref. [209], one can imagine a family of inationary models param-
eterized by a label , such that eternal ination is allowed when < 1. Then Eq. (6.21) will
have only the trivial solution,

X() 1, for 1. The case = 1 where eternal ination
is on the borderline of existence is the bifurcation point for the solutions of Eq. (6.21). At the
bifurcation point, a nontrivial solution

X() , 1 appears, branching o from the trivial solution.
A rigorous theory of bifurcation can be developed using methods of nonlinear functional analysis
(see e.g. chapter 9 of the book [220]). In particular, it can be shown that a nontrivial solution
of a nonlinear equation, such as Eq. (6.21), exists if and only if the dominant eigenvalue of the
linearized operator

L + 3 with zero boundary conditions is positive.
There remains a technical dierence between the eigenvalue problem for the operator

L + 3
with zero boundary conditions and with the no-diusion boundary conditions normally used
in the stochastic approach,

[D()P()] = 0. (6.40)
It was demonstrated in Ref. [159] that the eigenvalue of

L+3 with the boundary conditions (6.40)
is positive if a nontrivial solution of Eq. (6.21) exists. In principle, the eigenvalue of

L +3 with
zero boundary conditions is not the same as the eigenvalue of the same operator with the
boundary conditions (6.40). One can have a borderline case when one of these two eigenvalues
is positive while the other is negative. In this case, the two criteria for the presence of eternal
ination (based on the positivity of the two dierent eigenvalues) will disagree. However, the
alternative boundary conditions are imposed at reheating, i.e. in the regime of very small uctu-
ations where the value of the eigenfunction P() is exponentially small compared with its values
in the uctuation-dominated range of . Hence, the dierence between the two eigenvalues is
always exponentially small (it is suppressed at least by the factor e
3N
, where N is the number
of e-folds in the deterministic slow-roll regime before reheating). Therefore, we may interpret the
discrepancy as a limitation inherent in the stochastic approach to ination. In other words, one
cannot use the stochastic approach to establish the presence of eternal ination more precisely
than with the accuracy e
3N
. Barring an extremely ne-tuned borderline case, this accuracy is
perfectly adequate for establishing the presence or absence of eternal ination.
The main nonlinear equation in the calculations of the RV cuto is Eq. (6.24) for the generating
function g(z; ). That equation diers from Eq. (6.21) mainly by the presence of the parameter
z in the boundary conditions. Therefore, solutions of Eq. (6.24) may exist for some values of z
but not for other values. Note that g(z = 0; ) =

X(); hence, nontrivial solutions g(z; ) exist
for z = 0 under the same conditions as nontrivial solutions

X() , 1 of Eq. (6.21). While it is
certain that solutions g(z; ) exist for z 0, there may be values z < 0 for which no real-valued
solutions g(z; ) exist at all. However, the calculations in the RV cuto require only to compute
g(z

; ) for a certain value z

< 0, which is the algebraically largest value z where g(z; ) has a

singularity in the z plane. The structure of that singularity will be investigated in detail below,
57
6 The RV measure for random-walk ination
and it will be shown that g(z

; ) is nite while g/z (z z

)
1/2
diverges at z = z

. Hence,
the solution g(z; ) remains well-dened at least for all real z in the interval [z

, +]. It follows
that g(z; ) may be obtained e.g. by a numerical solution of a well-conditioned problem with
z = z

+, where > 0 is a small real constant.

6.3.2 Nonlinear Fokker-Planck equations
In this section I derive Eq. (6.24), closely following the derivation of Eq. (6.19) in Ref. [159].
We begin by considering the case when ination is driven by a single scalar eld , such that
reheating is reached at =

. Let (1;
0
) be the probability density of obtaining the nite
reheated volume 1. We will derive an equation for a generating function of the distribution
of volume, rather than an equation directly for (1;
0
). Since the volume 1 is by denition
nonnegative, it is convenient to dene a generating function g(z;
0
) through the expectation
value of the expression exp(z1), where z > 0 is the formal parameter of the generating function,
g(z;
0
)

e
zV
_
V<

_

0
e
zV
(1;
0
)d1. (6.41)
Note that for any z such that Re z 0 the integral in Eq. (6.41) converges, and the events with
1 = +are automatically excluded from consideration. However, we use the subscript 1 <
to indicate explicitly that the statistical average is performed over a subset of all events. The
distribution (1;
0
) is not normalized to unity; instead, the normalization is given by Eq. (6.4).
The parameter z has the dimension of inverse 3-volume. Physically, this is the 3-volume
measured along the reheating surface and hence is dened in a gauge-invariant manner. If
desired for technical reasons, the variable z can be made dimensionless by a constant rescaling.
The generating function g(z; ) has the following multiplicative property: For two statistically
independent regions that have initial values =
1
and =
2
respectively, the sum of the
(nitely produced) reheating volumes 1
1
+1
2
is distributed with the generating function
_
e
z(V
1
+V
2
)
_
=

e
zV
1
_
e
zV
2
_
= g(z;
1
)g(z;
2
). (6.42)
We now consider an H-region at some time t, having an arbitrary value (t) not yet in the
reheating regime. Suppose that the nitely produced volume distribution for this H-region
has the generating function g(z; ). After time t the initial H-region grows to N e
3Ht
statistically independent, daughter H-regions. The value of in the k-th daughter region
(k = 1, ..., N) is found from Eq. (6.14),

k
= +v()t +
k
_
2D()t, (6.43)
where the noise variables
k
(k = 1, ..., N) are statistically independent because they describe
the uctuations of in causally disconnected H-regions. The nitely produced volume distribu-
tion for the k-th daughter region has the generating function g(z;
k
). The combined reheating
volume of the N daughter regions must be distributed with the same generating function as
reheating volume of the original H-region. Hence, by the multiplicative property we obtain
g(z; ) =
N

k=1
g(z;
k
). (6.44)
58
6.3 Derivations
We can average both sides of this equation over the noise variables
k
to get
g(z; ) =
_
N

k=1
g(z;
k
)
_

1
,...,
N
. (6.45)
Since all the
k
are independent, the average splits into a product of N identical factors,
g(z; ) =
_
_
g(z; +v()t +
_
2D())
_

_
N
. (6.46)
The derivation now proceeds as in Ref. [159]. We rst compute, to rst order in t,
_
g(z; +v()t +
_
2D())
_

= g + (vg
,
+Dg
,
) t. (6.47)
Substituting N = e
3Ht
and taking the logarithmic derivative of both sides of Eq. (6.46) with
respect to t at t = 0, we then obtain
0 =

t
ln g(z; )
= 3H ln g +
vg
,
+Dg
,
g
. (6.48)
The equation for g(z; ) follows,
Dg
,
+vg
,
+ 3Hg ln g = 0. (6.49)
This is formally the same as Eq. (6.19). However, the boundary conditions for Eq. (6.49) are
dierent. The condition at the end-of-ination boundary =

is
g(z;

) = e
zH
3
()
(6.50)
because an H-region starting with =

immediately reheats and produces the reheating

volume H
3
(

). The condition at Planck boundary

Pl
(if present), or other boundary where
the eective eld theory breaks down, is absorbing, i.e. regions that reach =
Pl
do not
generate any reheating volume:
g(z;
Pl
) = 1. (6.51)
The variable z enters Eq. (6.49) as a parameter and only through the boundary conditions. At
z = 0 the solution is g(0; ) =

X().
A fully analogous derivation can be given for the generating function g(z;
0
, Q
0
) in the
case when additional uctuating elds, denoted by Q, are present. The generating function
g(z;
0
, Q
0
) is dened by
g(z;
0
, Q
0
) =
_

0
e
zV
(1;
0
, Q
0
)d1, (6.52)
where (1;
0
, Q
0
) is the probability density for achieving a total reheating volume 1 in the
future of an H-region with initial values
0
, Q
0
of the elds. In the general case, the uctuations
of the elds , Q can be described by the Langevin equations
(t +t) = (t) +v

t +

_
2D

t +
Q
_
2D
Q
t, (6.53)
Q(t +t) = Q(t) +v
Q
t +

_
2D
Q
t +
Q
_
2D
QQ
t, (6.54)
59
6 The RV measure for random-walk ination
where the diusion coecients D

, D
Q
, and D
QQ
have been introduced, as well as the
slow roll velocities v

and v
Q
and the noise variables

and
Q
. The resulting equation for
g(z;
0
, Q
0
) is (dropping the subscript 0)

Lg + 3g ln g = 0, (6.55)
where the dierential operator

L is dened by

L
D

+
2D
Q
H

Q
+
D
QQ
H

Q

Q
+
v

+
v
Q
H

Q
. (6.56)
The ratios D

/H, etc., are manifestly gauge-invariant with respect to time parameter changes
of the form (6.2).
Performing a redenition of the elds if needed, one may assume that reheating is reached
when =

independently of the value of Q. Then the boundary conditions for Eq. (6.55) at
the reheating boundary can be written as
g(z;

, Q) = e
zH
3
(,Q)
. (6.57)
The Planck boundary still has the boundary condition g(z;
Pl
) = 1.
6.3.3 Singularities of g(z)
For simplicity we now focus attention on the case of single-eld ination; the generating function
g(z; ) then depends on the initial value of the inaton eld . The corresponding analysis for
multiple elds is carried out as a straightforward generalization.
By denition, g(z; ) is an integral of a probability distribution (1; ) times e
zV
. It follows
that g(z; ) is analytic in z and has no singularities for Re z > 0. Then the probability distribu-
tion (1; ) can be recovered from the generating function g(z; ) through the inverse Laplace
transform,
(1; ) =
1
2i
_
i
i
dz e
zV
g(z; ), (6.58)
where the integration contour in the complex z plane can be chosen along the imaginary axis
because all the singularities of g(z; ) are to the left of that axis. The RV cuto procedure
depends on the limit of (1; ) and related distributions at 1 . The asymptotic behavior
at 1 is determined by the type and the location of the right-most singularity of g(z; )
in the half-plane Re z < 0. For instance, if z = z

is such a singularity, the asymptotic is

(1; ) exp[z

1]. The prefactor in this expression needs to be determined; for this, a

detailed analysis of the singularities of g(z; ) will be carried out.
It is important to verify that the singularities of g(z; ) are -independent. We rst show
that solutions of Eq. (6.55) cannot diverge at nite values of . If that were the case and say
g(z; ) as
1
, the function ln ln g as well as derivatives g
,
and g
,
would diverge as
well. Then
lim

[lnln g] = lim

g
g ln g
= . (6.59)
It follows that the term g ln g is negligible near =
1
in Eq. (6.55) compared with the term

g and hence also with the term

g. In a very small neighborhood of =

1
, the operator

L can be approximated by a linear operator

L
1
with constant coecients, such as

L

L
1
A
1

+B
1

. (6.60)
60
6.3 Derivations
Since at least one of the coecients A
1
, B
1
is nonzero at =
1
, it follows that g(z; ) is
approximately a solution of the linear equation

L
1
g = 0 near =
1
. However, solutions of
linear equations cannot diverge at nite values of the argument. Hence, the function g(z; )
cannot diverge at a nite value of .
The only remaining possibility is that the function g(z; ) has singular points z = z

such
that g(z

; ) remains nite while g/z, or a higher-order derivative, diverges at z = z

. We
will now investigate such divergences and show that g(z; ) has a leading singularity of the form
g(z; ) = g(z

; ) +()

z z

+O(z z

), (6.61)
where z

is a -independent location of the singularity such that z

0, while the function ()

is yet to be determined.
Denoting temporarily g
1
(z; ) g/z, we nd a linear equation for g
1
,

Lg
1
+ 3 (lng + 1) g
1
= 0, (6.62)
with inhomogeneous boundary conditions
g
1
(

) = H
3
(

)e
zH
3
()
, g
1
(
Pl
) = 0. (6.63)
The solution g
1
(z; ) of this linear problem can be found using a standard method involving the
Greens function. The problem with inhomogeneous boundary conditions is equivalent to the
problem with zero boundary conditions but with an inhomogeneous equation. To be denite,
let us consider the operator

L of the form used in Eq. (6.49),

L =
D()
H()

+
v()
H()

. (6.64)
Then Eqs. (6.62)(6.63) are equivalent to the inhomogeneous problem with zero boundary con-
ditions,

Lg
1
+ 3 (ln g + 1) g
1
= DH
4
e
zH
3

), (6.65)
g
1
(z;

) = g
1
(z;
Pl
) = 0. (6.66)
The solution of this inhomogeneous equation exists as long as the linear operator

L+3(ln g +1)
does not have a zero eigenfunction with zero boundary conditions.
Note that the operator

L+3(ln g +1) is explicitly z-dependent through the coecient g(z; ).
Note also that g(z; ) ,= 0 by denition (6.41) for values of z such that the integral in Eq. (6.41)
converges; hence ln g is nite for those z. Let us denote by G(z; ,

) the Greens function of

that operator with zero boundary conditions,

LG+ 3 (ln g(z, ) + 1) G = (

), (6.67)
G(z;

) = G(z;
Pl
,

) = 0. (6.68)
This Greens function is well-dened for values of z such that

L + 3(ln g(z; ) + 1) is invertible.
For these z we may express the solution g
1
(z; ) of Eqs. (6.62)(6.63) explicitly through the
Greens function as
g
1
(z; ) =
D
H
4
e
zH
3

G(z; ,

=
. (6.69)
61
6 The RV measure for random-walk ination
Hence, for these z the function g
1
(z; ) g/z remains nite at every value of . A similar
argument shows that all higher-order derivatives
n
g/z
n
remain nite at every for these z.
Therefore, the singularities of g(z; ) can occur only at certain -independent points z = z

,
z = z

, etc.
Since the generating function g(z; ) is nonsingular for all complex z with Re z > 0, it is
assured that g
1
(z; ) and G(z; ,

) exist for such z. However, there will be values of z for

which the operator

L + 3(ln g + 1) has a zero eigenfunction with zero boundary conditions, so
the Greens function G is undened. Denote by z

such a value with the algebraically largest real

part; we already know that Re z

0 in any case. Let us now show that the function g

1
(z; )
actually diverges when z z

. In other words, lim

zz
g
1
(z; ) = for every value of .
To show this, we need to use the decomposition of the Greens function in the eigenfunctions
of the operator

L + 3(ln g + 1),
G(z; ,

) =

n=0
1

n
(z)
f
n
()f

n
(

), (6.70)
where f
n
(z; ) are the (appropriately normalized) eigenfunctions with eigenvalues
n
(z) and zero
boundary conditions,
_

L + 3(ln g(z; ) + 1)
_
f
n
(z; ) =
n
(z)f
n
(z; ), (6.71)
f
n
(z; ) = 0 for =

, =
Pl
. (6.72)
The decomposition (6.70) is possible as long as the operator

L is self-adjoint with an appropriate
choice of the scalar product in the space of functions f(). The scalar product can be chosen in
the following way,
f
1
, f
2
) =
_
f
1
()f

2
()M()d, (6.73)
where M() is a weighting function. One can attempt to determine M() such that the operator

L is self-adjoint,
f
1
,

Lf
2
) =

Lf
1
, f
2
). (6.74)
In single-eld models of ination where the operator

L has the form (6.64), it is always possible to
choose M() appropriately [145]. However, in multi-eld models this is not necessarily possible.
4
One can show that in standard slow-roll models with K elds
1
, ...,
K
and kinetic coecients
D
ij
=
H
3
8
2

ij
, v
i
=
1
4G
H

i
, H = H(
1
, ...,
K
), (6.75)
there exists a suitable choice of M(), namely
M(
1
, ...,
K
) =
G
H
2
exp
_
G
H
2
_
, (6.76)
such that the operator

L = H
1

i,j
D
ij

j
+H
1

i
v
i

i
(6.77)
4
I am grateful to D. Podolsky for pointing this out to me. The hermiticity of operators of diusion type in the
context of eternal ination was briey discussed in Ref. [221].
62
6.3 Derivations
is self-adjoint in the space of functions f() with zero boundary conditions and the scalar prod-
uct (6.73). However, the operator

L may be non-self-adjoint in more general inationary models
where the kinetic coecients are given by dierent expressions. We omit the formulation of
precise conditions for self-adjointness of

L because this property is not central to the present
investigation. In non-self-adjoint cases a decomposition similar to Eq. (6.70) needs to be per-
formed using the left and the right eigenfunctions of the non-self-adjoint operator

L+3(ln g+1).
One expects that such a decomposition will still be possible because (heuristically) the nondiag-
onalizable operators are a set of measure zero among all operators. The requisite left and right
eigenfunctions can be obtained numerically. We leave the detailed investigation of those cases
for future work. Presently, let us focus on the case when the decomposition of the form (6.70)
holds, with appropriately chosen scalar product and the normalized eigenfunctions
f
m
, f
n
) =
mn
. (6.78)
The eigenfunctions f
m
(z; ) can be obtained e.g. numerically by solving the boundary value
problem (6.71)(6.72).
In the limit z z

, one of the eigenvalues

n
approaches zero. Since linear operators such
as

L always have a spectrum bounded from above [145], we may renumber the eigenvalues
n
such that
0
is the largest one. Then we dene z

as the value with the (algebraically) largest

real part, such that
0
(z

) = 0. By construction, for all z with Re z > Re z

all the eigenvalues

n
are negative. Note that the (algebraically) largest eigenvalue
0
(z) is always nondegenerate,
and the corresponding eigenfunction f
0
(z; ) can be chosen real and positive for all , except at
the boundaries =

and =
Pl
where f satises the zero boundary conditions.
For z near z

, only the nondegenerate eigenvalue

0
will be near zero, so the decomposi-
tion (6.70) of the Greens function will be dominated by the term 1/
0
. Hence, we can use
Eqs. (6.69) and (6.70) to determine the function g
1
(z; ) approximately as
g
1
(z; )
f
0
(z; )

0
(z)
f
0
(z;

_
D
H
4
e
zH
3
_

. (6.79)
It follows that indeed g
1
(z; ) as z z

because
0
(z) 0.
This detailed investigation allows us now to determine the behavior of g(z; ) at the leading
singularity z = z

. We will consider the function g(z; ) for z near z

and show that the

singularity indeed has the structure (6.61).
We have already shown that the function g itself does not diverge at z = z

but its derivative

g
1
g/z does. Hence, the function g(z

; ) is continuous, and the dierence g(z

; ) g(z; )
is small for z z

, so that we have the expansion

g(z; ) g(z; ) g(z

; ) (6.80)
g
1
(z; ) (z z

) +O
_
(z

z)
2

. (6.81)
(Note that we are using the nite value g
1
(z; ) rather than the divergent value g
1
(z

; ) in the
above equation.) On the other hand, we have the explicit representation (6.79). Let us examine
the values of
0
(z) for z z

. At z = z

we have
0
(z

) = 0, so the (small) value

0
(z) for
z z

can be found using standard perturbation theory for linear operators. If we denote the
change in the operator

L by

L 3(ln g(z; ) ln g(z

; ))
3g(z; )
g(z

; )
, (6.82)
63
6 The RV measure for random-walk ination
we can write, to rst order,

0
(z)
_
f
0
,

Lf
0
_
=
_
[f
0
()[
2
3g(z; )
g(z

; )
d. (6.83)
Now, Eqs. (6.79) and (6.81) yield
g(z; )
z z

f
0
(z; )

0
(z)
_
f
0

D
H
4
e
zH
3
_

. (6.84)
Integrating the above equation in with the prefactor
[f
0
()[
2
3
g(z

; )
d (6.85)
and using Eq. (6.83), we obtain a closed equation for
0
(z) in which terms of order (z z

)
2
have been omitted,

0
(z)
z z

0
(z)
_
f
0

D
H
4
e
zH
3
_

_
[f
0
[
2
3g(z; )
g(z

; )
f
0
()d. (6.86)
It follows that
0
(z)

z z

and g
1
(z; ) (z z

)
1/2
, conrming the leading asymptotic
of the form (6.61).
Let us also obtain a more explicit form of the singularity structure of g(z; ). We can rewrite
Eq. (6.86) as

0
(z)
0

z z

+O(z z

), (6.87)
where
0
is a constant that may be obtained explicitly. Then Eq. (6.79) yields
g
1
(z; )
f
0
(z; )

z z

1
, (6.88)
with a dierent constant
1
. Finally, we can integrate this in z and obtain
g(z; ) = g(z

; ) + 2
1
f
0
(z; )

z z

+O(z z

). (6.89)
We may rewrite this by substituting z = z

into f
0
(z; ),
g(z; ) = g(z

; ) +()

z z

+O(z z

), (6.90)
() 2
1
f
0
(z

; ). (6.91)
The result is now explicitly of the form (6.61). It will turn out that the normalization constant
2
1
cancels in the nal results. So in a practical calculation the eigenfunction f
0
(z

; ) may be
determined with an arbitrary normalization.
As a side note, let us remark that the argument given above will apply also to other singular
points z

,= z

as long as the eigenvalue

k
(z) of the operator

L + 3(ln g + 1) is nondegenerate
when it vanishes at z = z

. If the relevant eigenvalue becomes degenerate, the singularity

structure will not be of the form
_
z z

but rather (z z

)
s
with some other power 0 < s < 1.
64
6.3 Derivations
Now we are ready to obtain the asymptotic form of the distribution (1; ) for 1 . We
deform the integration contour in the inverse Laplace transform (6.58) such that it passes near
the real axis around z = z

. Then we use Eq. (6.90) for g(z; ) and obtain the leading asymptotic
(1; )
1
2i
()
__
z

_

z
_

z z

e
zV
dz
=
1
2

()1
3/2
e
zV
. (6.92)
The subdominant terms come from the higher-order terms in the expansion in Eq. (6.90) and
are of the order 1
1
times the leading term shown in Eq. (6.92).
Finally, we show that z

must be real-valued and that there are no other singularities z

with
Re z

= Re z

. This is so because the integral

g
1
(z

; ) =
_

0
(1; )e
zV
1d1 = (6.93)
will denitely diverge for purely real z

if it diverges for a nonreal value z

= z

+iA. If, on the

other hand, the integral (6.93) diverges for a real z

, it will converge for any nonreal z

= z

+iA
with real A ,= 0 because the function (1; ) has the large-1 asymptotic of the form (6.92) and
the oscillations of exp(iA1) will make the integral (6.93) convergent.
6.3.4 FPRV distribution of a eld Q
In this section we follow the notation of Ref. [206].
Consider a uctuating eld Q such that the Fokker-Planck operator

L is of the form (6.56). We
are interested in the portion 1
Q
R
of the total reheated volume 1 where the eld Q has a value
within a given interval [Q
R
, Q
R
+dQ]. We denote by (1, 1
Q
R
;
0
, Q
0
) the joint probability
distribution of the volumes 1 and 1
Q
R
for initial H-regions with initial values =
0
and
Q = Q
0
. The generating function g(z, q; , Q) corresponding to that distribution is dened by
g(z, q; , Q)
_
e
zVqV
Q
(1, 1
Q
; , Q)d1d1
Q
. (6.94)
Since this generating function satises the same multiplicative property (6.42) as the generating
function g(z; , Q), we may repeat the derivation of Eq. (6.55) without modications for the
function g(z, q; , Q). Hence, g(z, q; , Q) is the solution of the same equation as g(z; , Q). The
only dierence is the boundary conditions at reheating, which are given not by Eq. (6.57) but
by
g(z, q;

, Q) = exp
_
(z +q
QQ
R
) H
3
(

, Q)

, (6.95)
where (with a slight abuse of notation)
QQ
R
is the indicator function of the interval [Q
R
, Q
R
+dQ],
i.e.

QQ
R
(QQ
R
)(Q
R
+dQQ). (6.96)
We employ this nite version of the -function only because we cannot use a standard Dirac
-function under the exponential. This slight technical inconvenience will disappear shortly.
The solution for the function g(z, q; , Q) may be obtained in principle and will provide com-
plete information about the distribution of possible values of the volume 1
Q
together with the
total reheating volume 1 to the future of an initial H-region. In the context of the RV cuto,
65
6 The RV measure for random-walk ination
one is interested in the event when 1 is nite and very large. Then one expects that 1
Q
also
becomes typically very large while the ratio 1
Q
/1 remains roughly constant. In other words, one
expects that the distribution of 1
Q
is sharply peaked around a mean value 1
Q
), and that the
limit 1
Q
) /1 is well-dened at 1 . The value of that limit is the only information we need
for calculations in the RV cuto. Therefore, we do not need to compute the entire distribution
(1, 1
Q
; , Q) but only the mean value

1
Q
[
V
_
at xed 1.
Let us therefore dene the generating function of the mean value

1
Q
[
V
_
as follows,
h(z; , Q)

1
Q
R
e
zV
_
V<
=
g
q
(z, q = 0; , Q). (6.97)
(The dependence on the xed value of Q
R
is kept implicit in the function h(z; , Q) in order to
make the notation less cumbersome.) The dierential equation and the boundary conditions for
h(z; , Q) follow straightforwardly by taking the derivative
q
at q = 0 of Eqs. (6.55) and (6.95).
It is clear from the denition of g that g(z, q = 0; , Q) = g(z; , Q). Hence we obtain

Lh + 3 (ln g(z; , Q) + 1) h = 0, (6.98)

h(z;

, Q) =
e
zH
3
(,Q)
H
3
(

, Q)

QQ
R
, (6.99)
h(z;
Pl
, Q) = 0. (6.100)
Note that it is the generating function g, not g, that appears as a coecient in Eq. (6.98).
Since the nite -function
QQ
R
now enters only linearly rather than under an exponential,
we may replace
QQ
R
by the ordinary Dirac -function (QQ
R
). To maintain consistency, we
need to divide h by dQ, which corresponds to computing the probability density of the reheated
volume with Q = Q
R
. This probability density is precisely the goal of the present calculation.
The RV-regularized probability density for values of Q is dened as the limit
p(Q
R
) = lim
V

1
Q
R
[
V
_
V<
1 (1; , Q)
= lim
V
_
i
i
e
zV
h(z; , Q)dz
1
_
e
zV
g(z; , Q)dz
. (6.101)
To compute this limit, we need to consider the asymptotic behavior of

1
Q
R
[
V
_
V<
at 1 .
This behavior is determined by the leading singularity of the function h(z; , Q) in the complex
z plane. The arguments of Sec. 6.3.3 apply also to h(z; , Q) and show that h cannot have a -
or Q-dependent singularity in the z plane.
Moreover, h(z; , Q) has precisely the same singular points, in particular z = z

, as the basic
generating function g(z; , Q) of the reheating volume. Indeed, the function h(z; , Q) can be
expressed through the Greens function G(z; , Q,

, Q

) of the operator

L+3(ln g+1), similarly
to the function g
1
(z; ) considered in Sec. 6.3.3. For z ,= z

, this operator is invertible on the

space of functions f(, Q) satisfying zero boundary conditions. Hence, h(z; , Q) is nonsingular
at z ,= z

and becomes singular precisely at z = z

.
Let us now obtain an explicit form of h(z; , Q) near the singular point z = z

. We assume
again the eigenfunction decomposition of the Greens function (with the same caveats as in
Sec. 6.3.3),
G(z; , Q,

, Q

) =

n=0
1

n
(z)
f
n
(z; , Q)f

n
(z;

, Q

), (6.102)
66
6.3 Derivations
where f
n
are appropriately normalized eigenfunctions of the z-dependent operator

L+3(ln g+1)
with eigenvalues
n
(z). The eigenfunctions f
n
must satisfy zero boundary conditions at reheating
and Planck boundaries. Similarly to the way we derived Eq. (6.79), we obtain the explicit solution
h(z; , Q) =

n=0
f
n
(z; , Q)

n
(z)
_
f
n

e
zH
3
H
4
_
,Q
R
. (6.103)
The value of h(z; , Q) for z z

is dominated by the contribution of the large factor 1/

0
(z)
(z z

)
1/2
, so the leading term is
h(z; , Q)
f
0
(z

; , Q)

0
(z)
_
f
0

e
zH
3
H
4
_
,Q
R
. (6.104)
The asymptotic behavior of the mean value

1
Q
R
[
V
_
V<
is determined by the singularity of
h(z; , Q) at z = z

. As before, we may deform the integration contour to pass near the real
axis around z = z

. We can then express the large-1 asymptotic of

1
Q
R
[
V
_
V<
as follows,

1
Q
R
[
V
_
=
1
2i
_
i
i
e
zV
h(z; , Q)dz

f
0
(z

; , Q)

1
e
zV
_
f
0

e
zH
3
H
4
_
,Q
R
, (6.105)
where
0
is the constant dened by Eq. (6.87).
We now complete the analytic evaluation of the limit (6.101). Since the denominator of
Eq. (6.101) has the large-1 asymptotics of the form
1
_

0
g(z; , Q)e
zV
d1 f
0
(z

; , Q)1

1
2
e
zV
, (6.106)
where f
0
is the same eigenfunction, the dependence on and Q identically cancels in the
limit (6.101). Hence, that limit is independent of the initial values and Q but is a function only
of Q
R
, on which h(z; , Q) implicitly depends. Using this fact

we can signicantly simplify the

rest of the calculation. It is not necessary to compute the denominator of Eq. (6.101) explicitly.
The distribution of the values of Q at =

is simply proportional to the Q

R
-dependent part of
Eq. (6.105); the denominator of Eq. (6.101) serves merely to normalize that distribution. Hence,
the RV cuto yields
p(Q
R
) = const
_
f
0
(z

; , Q)

e
zH
3
H
4
_
=,Q=Q
R
, (6.107)
where the normalization constant needs to be chosen such that
_
p(Q
R
)dQ
R
= 1. This is the nal
analytic formula for the RV cuto applied to the distribution of Q at reheating. The the value
z

, and the corresponding solution g(z

; , Q) of Eq. (6.24), and the eigenfunction f

0
(z

; , Q)
need to be obtained numerically unless an analytic solution is possible.
Let us comment on the presence of the factor D

in the formula (6.107). The diusion

coecient D

is evaluated at the reheating boundary and is thus small since the uctuation
amplitude at reheating is (in slow-roll inationary models)

H
2

=
_
8
2
D

H
v

10
5
. (6.108)
67
6 The RV measure for random-walk ination
Nevertheless it is not possible to set D

= 0 directly in Eq. (6.107). This is so because

the existence of the Greens function of the Fokker-Planck operator such as

L depends on the
fact that

L is a second-order dierential operator of elliptic type. If one sets D

= 0 near
the reheating boundary, the operator

L becomes rst-order in at that boundary. Then one
needs to use a dierent formula than Eq. (6.65) for reducing an equation with inhomogeneous
boundary conditions to an inhomogeneous equation with zero boundary conditions. Accordingly,
one cannot use formulas such as Eq. (6.69) for the solutions. Alternative ways of solving the
relevant equations in that case will be used in Sec. 6.3.5.
6.3.5 Calculations for an inationary model
In this section we perform explicit calculations of RV cuto for a model of slow-roll ination
driven by a scalar eld with a potential shown in Fig. 4.1. The kinetic coecients D() and
v() are such that D() = D
0
, v() = 0, and H() = H
0
in the at region
1
< <
2
, where
the constants D
0
and H
0
are
H
0
=
_
8G
3
V
0
, D
0
=
H
3
0
8
2
. (6.109)
In the slow-roll regions
1
< <
(1)

and
2
< <
(2)

, the coecient D() is set equal to

zero, while v() ,= 0 and H() is not constant any more. The number of e-folds in the two
slow-roll shoulders can be computed by the standard formula,
N
j
=
_

(j)

j
H
v
d = 4G
_

(j)

j
H
H

d, j = 1, 2. (6.110)
The rst step of the calculation is to determine the singular point z = z

of solutions g(z; )
of Eq. (6.49). We expect z

to be real and negative. The boundary conditions for g(z; ) are

g(z;
(1,2)

) = exp
_
zH
3
(
(1,2)

)
_
. (6.111)
In each of the two deterministic regions,
(1)

< <
1
and
2
< <
(2)

, Eq. (6.49) becomes

v
H

g + 3g ln g = 0, (6.112)
with the general solution
g(z; ) = exp
_
C exp
_
3
_

H
v
d
__
, (6.113)
where C is an integration constant. Since the equation is rst-order within the deterministic
regions, the solutions are xed by the boundary condition (6.111) in the respective region,
g(z; ) = exp
_
zH
3
(
(1,2)

) exp
_
3
_

(1,2)

H
v
d
__
. (6.114)
We may therefore compute the values of g(z; ) at the boundaries
1,2
of the self-reproduction
region as
g(z;
1,2
) = exp
_
zH
3
(
(1,2)

) exp (3N
1,2
)
_
. (6.115)
68
6.3 Derivations
Now we need to solve Eq. (6.49) with these boundary conditions in the region
1
< <
2
. The
equation has then the form
D
0
H
0

g + 3g ln g = 0. (6.116)
Exact solutions of Eq. (6.116) were studied in Ref. [159], to which the reader is referred for more
details. It is easy to show that Eq. (6.116) is formally equivalent to a one-dimensional motion
of a particle with coordinate g() in a potential U(g),
U(g) =
6
2
H
2
0
g
2
(2 ln g 1) , (6.117)
while plays the role of time. A solution g(z; ) with boundary conditions (6.115) corresponds
to a trajectory that starts at the given value g(z;
1
) with the initial velocity chosen such that
the motion takes precisely the specied time interval
2

1
and reaches g(z;
2
). For z < 0 the
boundary conditions specify g(z;
1,2
) > 1, i.e. the trajectory begins and ends to the right of the
minimum of the potential (see Fig. 6.2). Since the system is conservative, there is a constant of
motion E (the energy) such that E = U(g
0
) at the highest point of the trajectory g
0
where
the kinetic energy vanishes. The solution g(z; ) can be written implicitly as one of the two
alternative formulas,

_
g(z;
1,2
)
g(z;)
dg
_
2E(z) 2U(g)
=
1,2
, (6.118)
valid in appropriate intervals
1
< <
0
and
0
< <
2
respectively, where
0
is the value
of corresponding to the turning point g
0
= g(z;
0
). The value E = E(z) in Eq. (6.118) must
be chosen such that the total time is
2

1
,
_
_
g(z;
1
)
g
0
+
_
g(z;
2
)
g
0
_
dg
_
2E 2U(g)
=
2

1
. (6.119)
This condition together with E = U(g
0
) implicitly determine the values E = E(z) and g
0
=
g
0
(z).
The singularity z = z

of the solution g(z; ) is found by using the condition g/z .

Dierentiating Eq. (6.118) with respect to z and substituting Eq. (6.115) for g(z;
1,2
), we obtain
the condition

g
z
1
_
2E(z) 2U(g)
+
e
3N
1,2
H
3
(
(1,2)

)
_
2E(z) 2U(g(z;
1,2
))
E

(z)
_
g(z;
1,2
)
g(z;)
dg
[2E(z) 2U(g)]
3/2
= 0. (6.120)
It follows that g/z when
E(z) = U(g(z;
1,2
)). (6.121)
This condition is interpreted in the language of Fig. 6.2 as follows. As the value of z becomes
more negative, the initial and the nal values of g given by Eq. (6.115) both grow. The last
available trajectory starts from rest at =
2
and at the value of g such that U(g) = E.
69
6 The RV measure for random-walk ination

6
2
H
2
0
U(g)
0 0.5 1 1.5
g
E
g
0
g(z;
1
) g(z;
2
)
Figure 6.2: Potential U(g) given by Eq. (6.117) can be used to interpret solutions g(z; ) as
mechanical motion in time at constant total energy E. The potential vanishes
at g = 0 and g = e
1/2
and has a minimum at g = 1. The value g
0
is the turning
point where U(g
0
) = E. A trajectory corresponding to z < 0 will begin and end
with g > 1, i.e. to the right of the minimum of the potential. Solutions cease to exist
when z < z

; the solution g(z

; ), shown by the thin line with arrows, corresponds

to a trajectory that starts at rest (as demonstrated in the text). The energy of this
trajectory is E 0, and so the value g(z

;
2
) is close to e
1/2
while g(z

;
1
) is close
to 1.
70
6.3 Derivations
To obtain a specic result, let us assume that
2

1
is suciently large to provide self-
reproduction (
2

1
H
0
) and that the number of e-folds in channel 1 is smaller than that
in channel 2,
H
3
(
(1)

) exp (3N
1
) H
3
(
(2)

) exp (3N
2
) . (6.122)
Then the value g(z;
2
) will grow faster than g(z;
1
) as z becomes more negative. It follows
that g(z;
2
) will reach the singular point rst. Since the time
2

1
is large, the constant
E will be close to 0 so that the trajectory spends a long time near g = 0. Then the value
g(z

;
2
) will be close to e
1/2
. Hence the value of z

is approximately
z

1
2
H
3
(
(2)

) exp (3N
2
) . (6.123)
For this value of z

, the starting point of the trajectory will be

g(z

;
1
) exp
_
1
2
exp (3N
1
3N
2
)
_
1. (6.124)
Hence, the solution g(z

; ) at the singular point z = z

can be visualized as the thin line in

Fig. 6.2, starting approximately at g(z

;
1
) = 1 and nishing at g(z

;
2
) e
1/2
.
An approximate expression for g(z

; ) can be obtained by setting E 0 in Eq. (6.118); then

the integral can be evaluated analytically. In the range
0
< <
2
we obtain

2

_
g(z;
2
)
g(z;)
dg
_
2U(g)

H
0

12
2
_
1 2 ln g

2
, (6.125)
so the solution is
g(z

; ) exp
_
1
2

6
2
H
2
0
(
2
)
2
_
,
0
< <
2
. (6.126)
In the range
1
< <
0
we obtain within the same approximation
g(z

; ) exp
_
1
2

6
2
H
2
0
_

1
+
H
0

12
2
_
2
_
. (6.127)
These approximations are valid for within the indicated ranges and away from the turning
point
0
. The value of
0
can be estimated by requiring that the value of g(z

;
0
) obtained
from Eq. (6.126) be equal to that obtained from Eq. (6.127). This yields

0

1
2
_

2
+
1

H
0

12
2
_

2
+
1
2
. (6.128)
We note that the value g(z

;
0
) can be obtained somewhat more precisely by approximating the
solution g(z

; ) in a narrow interval near =

0
by a function of the form exp
_
A +B(
0
)
2

and matching both the values and the derivatives of g(z

; ) to the approximations (6.126)

and (6.127) at some intermediate points straddling =
0
. In this way, a uniform analytic
approximation for g(z

; ) can be obtained. However, the accuracy of the approximations (6.126)

and (6.127) is sucient for the present purposes.
Having obtained adequate analytic approximations for z

and g(z

, ), we can now proceed

to the calculation of the mean volumes

1
1,2
[
V
_
of regions reheated through channels 1 and 2
respectively, conditioned on the event that the total volume of all reheated regions is 1. We use
71
6 The RV measure for random-walk ination
the formalism developed in Sec. 6.3.4, where the variable Q now takes only the discrete values
1 and 2, so instead let us denote that value by j. The relevant generating function h
j
(z; ) is
dened by
h
j
(z; )

1
j
e
zV
_
V<
, j = 1, 2, (6.129)
and is found as the solution of Eq. (6.98), which now takes the form
_
D()
H()

+
v()
H()

+ 3(ln g(z; ) + 1)
_
h
j
(z; ) = 0, (6.130)
with boundary conditions imposed at the reheating boundaries,
h
1
(z;
(1)

) = H
3
(
(1)

)e
zH
3
(
(1)

)
, h
1
(z;
(2)

) = 0; (6.131)
h
2
(z;
(1)

) = 0, h
2
(z;
(2)

) = H
3
(
(2)

)e
zH
3
(
(2)

)
. (6.132)
In the present toy model the diusion coecient is set to zero at reheating, so the formalism
developed in Sec. 6.3.4 needs to be modied. We will rst solve Eq. (6.130) analytically in
the no-diusion intervals of and obtain the boundary conditions for h at the boundaries of
the self-reproduction regime [
1
,
2
] where D() ,= 0. Then the methods of Sec. 6.3.4 will be
applicable to the boundary value problem for the interval [
1
,
2
].
Implementing this idea in the rst no-diusion region
(1)

< <
1
, we use the solu-
tion (6.114) for g(z; ) and reduce Eq. (6.130) to

h
j
+
3H
v
_
1 zH
3
(
(1)

) exp
_
3
_

(1)

H
v
d
__
h
j
= 0. (6.133)
This equation is easily integrated together with the boundary conditions (6.131)(6.132) and
yields the values of h
j
at
1
,
h
j
(z;
1
) =
j1
H
3
(
(1)

) exp
_
3N
1
zH
3
(
(1)

)e
3N
1
_
. (6.134)
Similarly we can determine the values h
j
(z;
2
). Since the value z = z

is important for the

present calculation, we now nd the values of h
j
at z = z

using the assumption N

2
N
1
and
the estimate (6.123),
h
j
(z

;
i
)
ij
exp
_
3N
i
+
1
2

i2

H
3
(
(i)

)
, i, j = 1, 2. (6.135)
We have thus reduced the problem of determining h
j
(z; ) to the boundary-value problem for
the interval [
1
,
2
] where the methods of Sec. 6.3.4 apply but the boundary conditions are given
by Eq. (6.135).
The next step, according to Sec. 6.3.4, is to compute the eigenfunction f
0
(z

; ) of the operator

L
D
0
H
0

+ 3(ln g(z

; ) + 1) (6.136)
such that

Lf
0
= 0; f
0
(z

;
1,2
) = 0. (6.137)
72
6.3 Derivations

V ()

1

0

2
f
0
(z

; )

Figure 6.3: Sketch of the eigenfunction f

0
(z

; ) (dashed line) interpreted as the wavefunction

of a stationary state with zero energy in the potential

V () (solid line). Due to
exponential suppression by the potential barrier, the amplitude of f
0
in the right
region is exponentially larger than that in the left region.
As we have shown, this eigenfunction with eigenvalue 0 exists precisely at z = z

. Once this
eigenfunction is computed, the ratio of the RV-regulated mean volumes in channels 1 and 2 will
be expressed through the derivatives of f
0
at the endpoints and through the modied boundary
conditions (6.135) as follows,
P(2)
P(1)
=
h
2
(z

;
2
)
h
1
(z

;
1
)
[

f
0
(z

;
2
)[

f
0
(z

;
1
)
. (6.138)
The absolute value is taken to compensate for the negative sign of the derivative

f
0
at the
right boundary point (assuming that f
0
0 everywhere). Since h
1,2
(z

;
1,2
) are already known,
it remains to derive an estimate for f
0
(z

; ).
The eigenvalue equation

Lf
0
= 0 formally resembles a one-dimensional Schrodinger equation
with the coordinate and the potential

V ()
12
2
H
2
0
(ln g(z

; ) + 1) . (6.139)
The eigenfunction f
0
(z

; ) is then interpreted as the wavefunction of a stationary state with

zero energy and zero boundary conditions at =
1,2
. According to Eqs. (6.126) and (6.127),
the function

V () has a maximum at
0
(see Fig. 6.3), while its values at the endpoints are

V (
1
)
12
2
H
2
0
,

V (
2
)
18
2
H
2
0
. (6.140)
Using the terminology of quantum mechanics, there is a potential barrier separating two classi-
cally allowed regions near =
1
and =
2
. Since the potential well at =
2
is deeper,
the ground state is approximately the ground state of that one well, with an exponentially small
amplitude of being near =
1
. The shape of the eigenfunction is sketched in Fig. 6.3. The ex-
ponential suppression of the amplitude near =
1
can be found using the WKB approximation,
which yields
[

f
0
(z

;
2
)[

f
0
(z

;
1
)
= A
21
exp
_
_

1
_

V ()d
_
, (6.141)
73
6 The RV measure for random-walk ination
where

1,2
are the turning points such that

V (

1,2
) = 0. The pre-exponential factor A
21
is
of order 1 and can, in principle, be obtained from a more detailed matching of the WKB-
approximated solution across the barrier to the solutions in the classically allowed regions, or
by determining the solution f
0
(z

; ) numerically. However, we will omit this calculation since

the main result will consist of an exponentially large factor. That factor can be estimated using
Eqs. (6.126), (6.127), and (6.139) as
_

1
_

V ()d 2
_

0

12
2
H
2
0
ln g d
3
2

2H
2
0
(
2

1
)
2
. (6.142)
Hence, the ratio (6.138) is simplied to
P(2)
P(1)
= A
21
H
3
(
(2)

)
H
3
(
(1)

)
e
3N
2
+
1
2
e
3N
1
exp
_
3
2

2H
2
0
(
2

1
)
2
_
. (6.143)
This is the main result quoted above in Eq. (6.36).
74
7 The RV measure for the landscape
The existing measure proposals that apply directly to landscape scenarios are the holographic
measure [178, 188] (see also the recent proposal [196]), the comoving horizon cuto [175, 186,
177], the stationary measure [11, 215], the measure on transitions [194], and the pseudo-
comoving measure [88, 193, 205, 192]. In the absence of a unique denition of the measure,
one judges a cuto prescription viable if its predictions are not obviously pathological. Possible
pathologies include the dependence on choice of spacetime coordinates [145, 190], the youngness
paradox [158, 71], and the Boltzmann brain problem [199, 200, 201, 88, 202, 203, 198, 204].
Various measures have been used for predicting cosmological parameters, most notably the
cosmological constant, in the landscape scenarios (see, e.g., Refs. [87, 215, 222, 89, 223, 224, 225]).
In this chapter I extend the RV measure, originally formulated in the context of random-walk
ination, to landscape scenarios. The basic idea of the RV proposal is to select multiverses
that are very large but (by rare chance) have a nite total number of observers. In the context
of a string landscape scenario (or a recycling universe), this can happen if suciently many
anti-de Sitter or Minkowski bubbles nucleate everywhere, collide, and merge. In such a case,
there will be a nite time after which no de Sitter regions remain and no further nucleations can
occur. Hence, there will be a nite time after which no more observers are created anywhere. By
this construction, one obtains a subensemble of multiverses having a xed, nite total number
N
obs
of observers. These nite multiverses with very large N
obs
are regarded as controlled
approximations to the actual innite multiverse. The observer-weighted statistical distribution
of any quantity within a nite multiverse can be obtained by ordinary counting, since the total
number of observers within any such multiverse is nite. The limit of that statistical distribution
as N
obs
is the nal result of the RV prescription.
It was shown in Chapters 5 and 6 that the RV measure is gauge-invariant, independent of the
initial conditions, and free of the youngness paradox in the context of random-walk ination.
Presently I investigate whether the same features persist in an application of the RV measure
to landscape scenarios. In particular, it is important to obtain RV predictions with respect to
the Boltzmann brain problem that has been widely discussed.
In principle, the RV prescription can be extended to landscape models in dierent ways,
depending on the precise choice of the ensemble of nite multiverses. The ensemble of multiverses
with a xed total number of observers N
obs
(where one counts both the ordinary observers and
the Boltzmann brains) appears to be the natural choice. However, it is dicult to compute
the number of observers directly and unambiguously. Instead of the total number of observers,
I propose to x the total number n
tot
of bubbles nucleated to the future of an initial bubble.
The total number of ordinary observers in bubbles of a given kind is proportional to the
volume of the reheating surface in those bubbles. It is known that the square bubble approx-
imation [198], which neglects the eects of bubble wall geometry, is adequate for the purposes
of volume counting. Then the evolution of the landscape is well described by the approximate
model called ination in a box [160, 146, 206]. In that approximation, one keeps track only
of the number of new bubbles nucleated in previously existing bubbles, and each new bubble is
75
7 The RV measure for the landscape
assumed to be instantaneously nucleated exactly of Hubble size in comoving coordinates. Moti-
vated by this approximation, I study a simplied denition of the RV measure for a landscape
scenario (see Sec. 7.2 for details): One requires the total number of bubbles of all types, n
tot
, to
be nite and evaluates the statistical distribution of bubble types (or other cosmological observ-
ables) in the limit n
tot
. In principle, this limit can be calculated if the bubble nucleation
rates are known. Since this is a rst attempt to perform this techically challenging calculation,
I concentrate only on bubble abundances and on the relative abundance of Boltzmann brains.
I neglect the increased number of observers due to additional slow-roll ination within bubbles;
this eect was considered in Ref. [206] and requires additional complications in the formalism.
A landscape scenario may be specied by enumerating the available N types of vacua by a
label j (j = 1, ..., N) and by giving the Hubble rates H
j
within bubbles of type j. One can,
in principle, compute the nucleation rate
jk
describing the probability (per unit four-volume
of spacetime) of creating a bubble of type k within bubbles of type j.
1
It is convenient to work
with the dimensionless rates,

jk

4
3

jk
H
4
j
. (7.1)
The rate
jk
equals the probability of having a bubble of type k within a 3-volume of one
horizon in a bubble of type j, during a single Hubble time.
2
Explicit expressions for
jk
are
available in some landscape scenarios. In what follows, I assume that
jk
are known.
In Sec. 7.3 I apply the RV measure proposal to a toy model of the landscape with four vacua
(the FABI model of Ref. [175]). In this model, the vacua labeled F and I are de Sitter (dS) and
the vacua labeled A and B are anti-de Sitter (AdS) states. One assumes that only the transitions
F I, I F, F A, and I B are allowed, with known nucleation rates
FI
,
IF
, etc., per
unit Hubble 4-volume. I show in Eq. (7.88) that the RV-regulated bubble abundances depend
on the value of the dimensionless number

_

IB

FA
_
+1

FI

IF
, e
3
. (7.2)
Here the constant , introduced for convenience, is simply the number of statistically independent
Hubble regions after one e-folding. Barring ne-tuned cases, one expects that the value of
is either much larger than 1 or much smaller than 1, since the nucleation rates may dier
by exponentially many orders of magnitude. By relabeling the vacua (F I and A B)
if necessary, we may assume that e
6
. Then the bubble abundances are approximately
described by the ratios
p(I) : p(F) : p(A) : p(B)
1

2
:
1

: 1 :
_

1
_ 1
+1
. (7.3)
This result can be interpreted as follows. Each of the I bubbles produces bubbles of type
F, and each of the F bubbles produces bubbles of type A. The abundance of B bubbles is
neligible compared with other bubbles. Heuristically, the chain of transitions I F A can
be interpreted as the dominant chain in the landscape. The ne-tuned case, e
6
< < e
6
, is
considered separately, and the result is given by Eq. (7.79).
1
For some recent work concerning the determination of the bubble nucleation rates, see Refs. [226, 221, 227, 228,
229].
2
The notation
jk
, chosen here and in Ref. [208] for its visual clarity, corresponds to
kj
of Ref. [175] and to

jk
of Ref. [206].
76
7.1 Regulating the number of terminal bubbles
I then consider the abundance of Boltzmann brains (BBs) in the FABI landscape (Sec. 7.3.2).
The total number of BBs is proportional to the total number of Hubble regions (H-regions,
or 4-volumes of order H
4
) in de Sitter bubbles. The coecient of proportionality is the tiny
nucleation rate
BB
of Boltzmann brains, which is of order exp(10
50
) or smaller. In comparison,
ordinary observers occur at a rate of at least 1 per horizon volume. It turns out that (after
applying the RV cuto) the total number of H-regions of dS types F or I is approximately
equal to the total number of nucleated bubbles of the same type. Hence, the BBs are extremely
rare compared with ordinary observers.
In Sec. 7.4 I extend the same calculations to a general landscape with an arbitrary number of
vacua. The RV prescription predicts a denite ratio p(j)/p(k) between the number of bubbles of
types j and k. I derive a formula for the ratio p(j)/p(k) that involves all the parameters of the
landscape. With the help of mathematical results derived in Sections 7.4.5 and 7.4.6, it is possible
to show in full generality that the RV measure gives well-dened results that are independent of
the initial conditions. Nevertheless, actually performing the required calculations for an arbitrary
landscape remains a daunting task. To obtain explicit expressions in a semi-realistic setup, I
calculate the bubble abundances for a landscape that contains a single high-energy vacuum and
a large number of low-energy vacua. The result is an approximate formula [Eq. (7.149)] for the
ratio p(j)/p(k) expressed directly through the nucleation rates of the landscape.
I also demonstrate in Sec. 7.4.3 that the abundance of Boltzmann brains is negligible compared
with the abundance of ordinary observers in the same bubble type.
7.1 Regulating the number of terminal bubbles
The landscape scenarios where transitions between metastable vacuum states occur via bubble
nucleation, promise to explain the values of presently observed cosmological parameters, such
as the eective cosmological constant . For this reason it is important to be able to apply the
RV measure to landscape-type scenarios and, in particular, to compute the relative abundances
of dierent bubble types.
In the terminology of Ref. [175], terminal bubbles are those with nonpositive value of . No
further transitions are possible from such bubbles because bubbles with < 0 rapidly collapse
while bubbles with = 0 do not support tunneling instantons. The RV measure, as presently
formulated, can be used directly for comparing the abundances of terminal bubbles. (Extending
the RV prescription to non-terminal bubble types is certainly possible but is delegated to a
future publication.)
It was shown in Ref. [198] that the bubble volume calculations may use the simplifyingsquare
bubble approximation, which neglects the eects of bubble wall geometry. In this approxima-
tion, the evolution of the landscape is well described by the ination in a box model [146],
dened as follows. All the vacuum states are labeled by j = 1, ..., N of which the terminal states
are j = 1, ..., N
T
. During a time step t, an initial H-region of type j expands into n
j
e
3H
j
t
independent daughter H-regions of type j. Each of the daughter H-regions then has probability

jk
of changing into an H-region of type k (for convenience we dene
jj
1

k=j

jk
). The
process is repeated ad innitum for each resulting H-region, except for H-regions of terminal
types. A newly created H-region of terminal type will admit no further transitions and will
not expand (or, perhaps, will expand only by a xed amount of slow-roll ination occurring
immediately after nucleation). This imitates the behavior of anti-de Sitter or Minkowski vacua
that do not admit further bubble nucleations.
77
7 The RV measure for the landscape
To implement the RV cuto in this model, let us consider the probability p
j
(n, n

; k) of pro-
ducing a nite total number n of terminal H-regions of which n

are of type j, starting from

one initial H-region of (nonterminal) type k. A generating function for this distribution can be
dened by
g
j
(z, q; k)

n,n

=0
z
n
q
n

p
j
(n, n

; k)
_
z
n
q
n

_
n<
. (7.4)
One can show that this generating function satises the following system of nonlinear algebraic
equations,
g
1/n
k
j
(z, q; k) =
N
T

i=1

ki
zq

ij
+
N

i=N
T
+1

ki
g
j
(z, q; i), (7.5)
I will merely sketch the derivation of Eq. (7.5) here.
3
The generating function g
j
satises a
multiplicative property analogous to Eq. (6.42). This property applies to the n
k
independent
daughter H-regions created by expansion from an H-region of type k. Therefore, g
j
(z, q; k),
which is the expectation value of z
n
q
n

in an initial H-region of type k, is equal to the product

of n
k
expectation values of z
n
q
n

in the n
k
daughter H-regions (which may be of dierent types).
The latter expectation value is given by the right-hand side of Eq. (7.5). This yields Eq. (7.5)
after raising both sides to the power 1/n
k
.
If the generating functions g
j
are known, the distribution p
j
(n, n

; k) can be recovered by
computing derivatives of g
j
(z, q; k) at z = 0 and q = 0. Further, the mean fraction of H-regions
of type j at xed total number n of terminal H-regions is found as
p(j[n)
n

[
n
)
n
=

n
z

q
g
j
(z = 0, q = 1; k)
n
n
z
g
j
(z = 0, q = 1; k)
. (7.6)
Then the RV cuto denes the probability of terminal type j, among all the possible terminal
types, through the limit
p(j) lim
n
p(j[n) = lim
n

n
z

q
g
j
(z = 0, q = 1; k)
n
n
z
g
j
(z = 0, q = 1; k)
, (7.7)
similarly to Eq. (5.4). Again one expects that the limit exists and is independent of the initial
bubble type, as long as the initial bubble is not of terminal type.
As a specic example requiring fewer calculations, let us consider a toy model with only three
bubble types. There is one de Sitter ( > 0) vacuum labeled j = 3 that can decay into two
possible anti-de Sitter terminal bubbles labeled j = 1 and j = 2. The growth rate n
3
and
the nucleation probabilities
31
and
32
are assumed known. To mimick interesting features of
the landscape, let us also assume that there is a period of slow-roll ination inside the bubbles
1 and 2, generating respectively N
1
and N
2
additional e-folds of inationary expansion after
nucleation. Hence, the model is determined by the parameters n
3
,
31
,
32
, N
1
, and N
2
. For
convenience we dene
33
1
31

32
.
We now perform the calculations for the RV cuto in this simple model. There are only two
generating functions, g
1
(z, q; k) and g
2
(z, q; k), that need to be considered. The only meaningful
initial value is k = 3 (i.e., the initial bubble is of de Sitter type) since the two other bubble
3
This equation is similar to the equations for generating functions used in the theory of branching processes (see
e.g. the book [230] for a mathematically rigorous presentation). A more detailed derivation will be given in
Sec. 7.4.4.
78
7.1 Regulating the number of terminal bubbles
types do not lead to eternal ination. Hence, we will suppress the argument k in g
j
(z, q; k). We
also need to modify Eq. (7.5) to take into account the additional expansion inside the terminal
bubbles. Let us denote the volume expansion factors by
Z
1
e
3N
1
, Z
2
e
3N
2
. (7.8)
Then the functions g
1
and g
2
are solutions of
g
1/n
3
1
=
31
z
Z
1
q
Z
1
+
32
z
Z
2
+
33
g
1
, (7.9)
g
1/n
3
2
=
31
z
Z
1
+
32
z
Z
2
q
Z
2
+
33
g
2
. (7.10)
An explicit solution of these equations is impossible for a general n
3
(barring the special cases
n
3
= 2, 3, 4). Nevertheless, sucient information about the limit (7.7) can be obtained by the
following method. Introduce the auxiliary function F(x) as the solution F > 0 of the algebraic
equation
F
1/n
3
= x +
33
F, (7.11)
choosing the branch connected to the value F(0) = 0. (It is straightforward to see that Eq. (7.11),
has at most two positive solutions, and that F(x) is always the smaller solution of the two.)
The generating functions g
1
and g
2
are then expressed through F(x) as
g
1
(z, q) = F(
31
z
Z
1
q
Z
1
+
32
z
Z
2
), (7.12)
g
2
(z, q) = F(
31
z
Z
1
+
32
z
Z
2
q
Z
2
). (7.13)
The limit (7.7) involves derivatives of these functions of very high order with respect to z. We
note that the function F(x) is analytic; thus the functions g
1
and g
2
are also analytic in z.
To evaluate the high-order derivatives, we need an elementary result from complex analysis.
The asymptotic growth of high-order derivatives of an analytic function f(z) is determined by
the location of the singularities of f(z) in the complex z plane. The required result can be
derived by the following elementary argument. Consider the derivative d
n
f/dz
n
at z = 0, and
assume that f(z) admits an expansion around the singularity z

nearest to z = 0, such as
f(z) = c
0
+c
1
(z z

)
s
+..., (7.14)
where s ,= 0, 1, 2, ... is the power of the leading-order singularity, and the omitted terms are
either higher powers of z z

or singularities at points z

located further away from z = 0. The

singularity structure (7.14) yields the large-n asymptotics
d
n
f
dz
n

z=0
c
1
(z

)
s
(n s)
(s)
z
n

+... (7.15)
This formula enables one to evaluate large-n limits such as Eq. (7.7). It can be seen from this
formula that any other singular point z

located further away from z = 0 gives a contribution

that is smaller by the factor [z

/z

[
sn
. The contribution of a subdominant singularity of the
form (z z

)
s

, i.e. at the same point z = z

but with a higher power s

> s, is suppressed, in
comparison with the term in Eq. (7.15), by the factor
(n s

)
(n s)
(s)
(s

)

1
(n 1)
s

s
(s)
(s

)
. (7.16)
79
7 The RV measure for the landscape
It is clear that the terms omitted from Eq. (7.15) indeed give subleading contributions at large
n, and so Eq. (7.15) is indeed the leading asymptotic term at n .
To proceed, we need to determine the location of the singularities of F(x). Since F(x) is
obtained as an intersection of a curve F
1/n
3
and a straight line x +
33
F, there will be a value
x = x

where the straight line is tangent to the curve. At this value of x the function F(x) has
a singularity of the type
F(x) = F(x

) +F
1

x x

+O(x x

), (7.17)
where F
1
is a constant that can be easily determined; we omit further details that will not be
required below. The value of x

is found from the condition that dF/dx diverge at x = x

. The
value of dF/dx at x ,= x

is found as the derivative of the inverse function, or by taking the

derivative of Eq. (7.11),
dF
dx
=
1
1
n
3
F
1
n
3
1

33
. (7.18)
This expression diverges at the values
F(x

) = (n
3

33
)

n
3
n
3
1
, (7.19)
x

= (n
3
1)
33
F(x

) =
33
n
3
1
n
n
3
/(n
3
1)
3
. (7.20)
Note that
33
1 and n
3
1, hence x

is a constant of order 1.
Rather than compute the limit (7.41) directly, we will perform an easier computation of the
ratio of the mean number of bubbles of types 1 and 2 at xed total number n of terminal bubbles,
_
n

(1)

n
_
_
n

(2)

n
_ =

n
z

q
g
1
(z = 0, q = 1)

n
z

q
g
2
(z = 0, q = 1)
. (7.21)
The derivatives
q
g
1
and
q
g
2
can be evaluated directly through Eqs. (7.12)(7.13). For instance,
we compute
q
g
1
as
g
1
(z, q)
q

q=1
= F

(
31
z
Z
1
+
32
z
Z
2
)
31
Z
1
z
Z
1
. (7.22)
It is clear that the functions
q
g
1
and
q
g
2
have a singularity at z = z

corresponding to the
singularity x = x

of the function F(x), where z

is found from the condition

31
z
Z
1

+
32
z
Z
2

= x

. (7.23)
Let us analyze this equation in order to esimate z

. If the nucleation rates

31
and
32
dier
by many orders of magnitude, we may expect that one of the terms in Eq. (7.23), say
31
z
Z
1

,
dominates. Then the value z

is well approximated by
z

_
x

31
_
1/Z
1
. (7.24)
This approximation is justied if the rst term in Eq. (7.23) indeed dominates, which is the case
if
_

32
x

_
1/Z
2

31
x

_
1/Z
1
. (7.25)
80
7.1 Regulating the number of terminal bubbles
If the reversed inequality holds, we can relabel the bubble types 1 and 2 and still use Eq. (7.24).
If neither Eq. (7.25) nor the reversed inequality hold, the approximation (7.24) for z

can be used
only as an order-of-magnitude estimate. To be specic, let us assume that the condition (7.25)
holds.
We can now compute the ratio (7.21) asymptotically for large n using Eqs. (7.17) and (7.15).
The singularities of the functions g
1
and g
2
are directly due to the singularity of the function F.
Then the dominant singularity structure of the function (7.22) is found as
g
1
(z, q)
q

q=1

31
Z
1
z
Z
1
2
_

z
(
31
z
Z
1
+
32
z
Z
2
)

z=z
F
1

z z

. (7.26)
This ts Eq. (7.14), where f(z)
q
g
1
(z, q = 1) and s = 1/2. After canceling the common
n-dependent factors, we obtain
p(1)
p(2)
= lim
n
_
n

(1)

n
_
_
n

(2)

n
_ =

31
Z
1
z
Z
1

32
Z
2
z
Z
2

z=z
. (7.27)
This is the nal probability ratio found by applying the RV cuto to a toy model landscape
containing two terminal vacua. Substituting the approximation (7.24), which assumes the con-
dition (7.25), we can simplify Eq. (7.27) to a more suggestive form
p(1)
p(2)


31
Z
1

32
Z
2
_

31
x

_
1+Z
2
/Z
1
. (7.28)
We can interpret this as the ratio of nucleation probabilities
31
/
32
times the ratio of volume
expansion factors, Z
1
/Z
2
, times a certain correction factor. As we have seen, the correction
factor is actually a complicated function of all the parameters of the landscape. The correction
factor takes the simple form
_

31
x

_
1+Z
2
/Z
1
(7.29)
only if the condition (7.25) holds.
We note that the result (7.28) is similar to but does not exactly coincide with the results
obtained in previously studied volume-based measures. For comparison, the volume-based mea-
sures proposed in Refs. [175] and [11] both yield
p(1)
p(2)
=

31
Z
1

32
Z
2
, (7.30)
which is readily interpreted as the ratio of nucleation probabilities enhanced by the ratio of
volume factors. The holographic measure [178], which is not a volume-based measure, gives
the ratio
p(1)
p(2)
=

31

32
(7.31)
that does not depend on the number of e-folds after nucleation. While the discrepancy between
volume-based and worldline-based measures is to be expected, the correction factor that distin-
guishes Eq. (7.28) from Eq. (7.30) is model-dependent and may be either negligible or signicant
depending on the particular model.
81
7 The RV measure for the landscape
As a specic example, consider bubbles that nucleate with equal probability,
31
=
32
1,
but have very dierent expansion factors, Z
1
Z
2
. Then the condition (7.25) holds and the
correction factor is given by Eq. (7.29), which is an exponentially large quantity of order
1
31
.
Qualitatively this means that the RV measure rewards bubbles with a larger slow-roll expansion
factor even more than previous volume-based measures.
On the other hand, if Z
1
= Z
2
but
31

32
, the correction factor disappears and we
recover the result found in the other measures.
To conclude, we note that the result (7.27) does not depend on the durations of time spent
during slow-roll ination inside the terminal bubbles, but only on the number of e-folds gained.
This conrms that the RV measure does not suer from the youngness paradox.
So far we were able to apply of the RV measure to the comparison of the abundances of
terminal vacua. In the next section, the RV measure can be extended to arbitrary observables
in landscape models.
7.2 Regulating the total number of bubbles
The RV measure prescription as formulated in the previous section applies only to the calculation
of abundances of terminal bubbles. We will now extend the RV measure to computing arbitrary
statistics on any landscape.
We rst note that RV measure prescription can be applied, strictly speaking, only to land-
scapes that contain some terminal bubble types (i.e. vacua from which no further tunneling is
possible). However, this limitation is quite benign, for two reasons. First, a landscape without
any Minkowski or AdS states is not expected to be realized in any realistic string theory scenario
without an exceptional amount of ne-tuning. Second, the previously proposed volume-based
and the world-line based measure prescriptions agree for a landscape without terminal bubble
types [177, 178]. One may therefore consider the measure problem as solved in such landscapes
and turn ones attention to more realistic landscapes where terminal bubble types are present.
Let us take an initial bubble of a nonterminal type j and consider the statistical ensemble
E
n
(j) of all possible evolutions of the initial bubble such that the total number of nucleated
bubbles of all types is nite and equals n (not counting the initial bubble). The total number of
nucleated bubbles in a multiverse can be nite only if terminal bubbles nucleate everywhere and
merge globally to the future of the initial bubble. This can happen by rare chance; however, it
is important the total probability of all events in the ensemble E
n
(j) is always nonzero for any
given n, so that the ensembles E
n
(j) are well-dened and nonempty.
The ensemble E
n
(j) may be described in the language of transition trees used in Ref. [178].
The ensemble consists of all trees that have a total number n+1 of bubbles, including the initial
bubble of type j. The trees in E
n
(j) are nite because all the outer leaves are bubbles of
terminal types. The motivation for considering the ensemble E
n
(j) is that a nite but very large
tree (with n large) is a controlled approximation to innite trees that typically occur. Hence,
we are motivated to consider E
n
(j) with n nite but very large.
Note that the ensemble E
n
(j) diers from the ensemble dened in Ref. [206]; in E
n
(j) the total
number of bubbles of all types is equal to n, rather than the total number of terminal bubbles
as in Ref. [206]. Thus, the current proposal, which appears more natural, is an extension of that
of Ref. [206]. Future work will show whether this technical dierence is signicant; presently I
will investigate the consequences of the current proposal.
Once the ensemble E
n
(j) is dened, one may consider the statistical distribution of some
82
7.3 A toy landscape
cosmological observable within the multiverses belonging to the set E
n
(j). For instance, one can
count the number of bubbles of some type k, or the number of observers within bubbles of type k,
or the number of observations of some physical process, etc. In a very large multiverse belonging
to E
n
(j) with n 1, one may expect that the statistics of observations will be independent of
the initial bubble type j. Indeed, this will be one of the results of this chapter. Hence, let us
suppress the argument j and write simply E
n
.
Each multiverse belonging to E
n
has a naturally dened probability weight, which is sim-
ply equal to the probability of realizing that multiverse. This probability weight needs to be
taken into account when computing the statistical distributions of observables. The sum of all
probability weights of multiverses within E
n
is equal to the total probability of E
n
, which is
exponentially small for large n but always nonzero. Since the multiverses from the ensemble E
n
are by construction nite, i.e. each multiverse supports only a nite total number of possible
observers, we are assured that any statistics we desire to compute on E
n
will be well-dened.
We can now consider the probability distribution p(Q[E
n
) of some interesting observable Q
within the ensemble E
n
and take the limit n . One expects that the probability distribution
p(Q[E
n
) will have a well-dened limit for large n,
p(Q) lim
n
p(Q[E
n
). (7.32)
It was shown in previous work on the RV prescription [206, 207] that the distribution p(Q) is
well-dened for a simplest toy landscape as well as in the case of random-walk eternal ination.
In this work I extend these results to a general landscape scenario. Below (Sec. 7.4.5) I will prove
rigorously that the limit (7.32) indeed exists and is independent of the chosen initial bubble type
j as long as the initial bubble is not of terminal type and as long as the landscape is irreducible
(every vacuum can be reached from every other non-terminal vacuum by a chain of nucleations).
Thus, the distribution p(Q) is unique and well-dened. This distribution is the nal result of
applying the RV prescription to the observable Q.
In practice, it is necessary to compute the distribution p(Q[E
n
) asymptotically in the limit of
large n. A direct numerical calculation of probabilities in that limit by enumerating all possible
evolution trees is extremely dicult because of the exponential growth of the number of possible
evolutions. Instead, I derive explicit formulas for the distribution p(Q) in a generic landscape
by evaluating the limit (7.32) analytically. These formulas are the main result of the present
chapter.
7.3 A toy landscape
I begin by applying the RV measure to a toy landscape with very few vacua. Using this simple
example, I develop the computational techniques needed for the practical evaluation of the limit
such as Eq. (7.32). In Sec. 7.4 the same techniques will be extended to a more general landscape
with an arbitrary number of vacua.
In Sec. 7.1 we considered the simplest possible nontrivial landscape: a single dS and two AdS
(terminal) vacua. The next least complicated example that can be treated analytically is a
toy landscape having two nonterminal and two terminal vacua. This toy landscape was called
the FABI model in Ref. [175] and consists of the vacua labeled F, I, A, B with the transition
diagram A F I B. In other words, one assumes that the vacuum F (false vacuum)
can nucleate only bubbles of types A and I, the vacuum I (intermediate vacuum) can nucleate
bubbles of types F and B, while A and B are terminal vacua that do not have further nucleations.
83
7 The RV measure for the landscape
To describe the nitely produced probability in this landscape, I use the discrete picture
called the eternal ination in a box [146, 206], which is closely related to the square bubble
approximation [198]. In this picture, one considers the evolution of discrete, causally disjoint
homogeneous H-regions in discrete time. All possible vacuum types are labeled by j = 1, ..., N.
During one time step of order t = H
1
j
, where H
j
is the local Hubble rate in a given H-region
of type j, the evolution consists of expanding the H-region into
e
3H
j
t
= e
3
(7.33)
daughter H-regions of type j. Each of the daughter H-regions has then the probability
jk
of
changing immediately into an H-region of type k ,= j; this imitates a nucleation of a horizon-size
bubble of type k. If no transition has taken place, the daughter H-region retains its type j. For
convenience, we denote by

jj
1

k=j

jk
(7.34)
the probability of no transitions during one Hubble time. The process of expansion and transition
is continued ad innitum, independently for each resulting H-region. The H-regions of terminal
types will admit no further transitions and will not expand further (except, perhaps, by a xed
amount due to slow-roll ination occuring immediately after nucleation). An example simulation
is shown in Fig. 7.1.
In the remainder of this section I perform explicit calculations of the RV cuto in the FABI
model.
7.3.1 Bubble abundances
The rst task is to compute the relative abundance of bubbles of dierent types. Consider the
probability p(n
tot
, n
F
, n
I
, n
A
, n
B
; k) of having a nite total number n
tot
of bubbles of which
n
j
are of type j (where j = F, I, A, B), if one starts from a single initial H-region of type k
(k = F, I). By construction, p ,= 0 only for n
tot
= n
F
+ n
I
+ n
A
+ n
B
. A generating function
for this probability distribution can be dened by
g(z, q
j
; k)

n0,n
j
0
z
n
p(n, n
j
; k)

j=F,I,A,B
q
n
j
j

z
n
q
n
F
F
q
n
I
I
q
n
A
A
q
n
B
B
_
n<;k
, (7.35)
where the notation ...)
n<;k
stands for a statistical average restricted to events with a nite
total number n of bubbles nucleated to the future of an initial bubble of type k.
The generating function g plays a crucial role in the entire calculation. Since we will be only
interested in the initial bubbles of types F and I, let us denote
F(z, q
j
) g(z, q
j
; F), I(z, q
j
) g(z, q
j
; I). (7.36)
The generating functions F and I satisfy the following system of nonlinear algebraic equa-
tions [206],
F
1

= zq
A

FA
+zq
I

FI
I +
FF
F, (7.37)
I
1

= zq
B

IB
+zq
F

IF
F +
II
I. (7.38)
84
7.3 A toy landscape
Figure 7.1: An example simulation of eternal ination in a boxin two spatial dimensions. Bub-
bles (H-regions) are represented in comoving coordinates by squares. Dark shades
indicate bubbles of terminal types. Other shades correspond to nested bubbles of
various nonterminal (recyclable) types. For the purposes of visual illustration,
nucleation rates were chosen of order one, and colored lines were drawn at bubble
boundaries.
85
7 The RV measure for the landscape
Here we denoted for brevity
FF
1
FA

FI
and
II
1
IB

IF
; within our assumptions,

FF
1 and
II
1. In Eqs. (7.37)(7.38) the generating variable z multiplies only the terms
that correspond to changing the type of the H-region (which imitates the nucleation of new
bubbles) but not the terms
FF
F and
II
I that correspond to the eventuality of not changing
the type of the H-region during one Hubble time.
The nonlinear equations (7.37)(7.38) may have several real-valued solutions, as well as
complex-valued solutions that are certainly not of physical interest. In particular, for z = 1 and
q
j
= 1 there exists the trivial solution F = I = 1 as well as a nontrivial solution with F 1
and I 1. Similarly, for z near 0 there exists the solution that approaches F(0) = I(0) = 0,
F = (zq
A

FA
)

+O(z
21
), I (zq
B

IB
)

+O(z
21
), (7.39)
as well as the solution
F

1
FF
1, I

1
II
1 (7.40)
and solutions where F 1 and I 0 and vice versa. It is important to determine the solution
branch F(z), I(z) that has the physical signicance as the actual generating function of the
nitely produced distribution of H-regions.
The functions F(z), I(z) are solutions of algebraic equations and thus are continuous functions
of z that are analytic everywhere in complex z plane except for branch cuts. It is easy to see
that the solution F = I = 1 at z = 1, q
j
= 1 is continuously connected with the solution (7.40)
at z 0, while the solution (7.39) is continued to a solution with F(z) 1 and I(z) 1 for
all 0 < z < 1. The values F(1) and I(1) are the probabilities of the events that the evolution
of an initial bubble of type F or I ends globally. These probabilities are extremely small and
of order

FA
and

IB
respectively. This is easy to interpret because, for instance,

FA
is the
probability of nucleating terminal regions at once after one Hubble time within an H-region
of type F. Hence, the generating functions for the nitely produced distribution of bubbles are
given by the solution branch having F 1 and I 1 rather than by the solution F = I = 1 at
z = 1. More precisely, the physically meaningful solution F(z), I(z) is selected by the asymptotic
behavior (7.39). This argument removes the ambiguity inherent in solving Eqs. (7.37)(7.38).
Below we refer to the branch of solutions F(z), I(z) connected to the nontrivial solution (7.39)
near z = 0 as the main branch.
Once the main branch of the generating functions g(z, q
j
; k) are known, one can express the
mean number of bubbles of type i (i = F, I, A, B) at a xed total number n
tot
as
p(i[n
tot
)
n
i
)
ntot
n
tot
=

ntot
z

q
i
g(z = 0, q
j
= 1 ; k)
n
tot

ntot
z
g(z = 0, q
j
= 1 ; k)
. (7.41)
One expects that the limit of this ratio at n
tot
will be independent of the initial bubble
type k since the ensemble E
ntot
will consist of H-regions having a very long evolution, so that
the initial conditions are forgotten. Below I will show explicitly that this is indeed the case.
It is more convenient to compute the ratios of the number of bubbles of types i and i

at xed
n
tot
,
p(i[n
tot
)
p(i

[n
tot
)
=

ntot
z

q
i
g

ntot
z

q
i

z=0,{q
j
=1}
. (7.42)
Then one only needs to compute derivatives
q
i
g g
,q
i
evaluated at q
j
= 1. These derivatives
satisfy a system of linear equations that can be easily derived from Eqs. (7.37)(7.38). For
86
7.3 A toy landscape
instance, the derivatives F
,q
A
and I
,q
A
satisfy
1

F
1

1
F
,q
A
= z
FA
+z
FI
I
,q
A
+
FF
F
,q
A
, (7.43)
1

I
1

1
I
,q
A
= z
IF
F
,q
A
+
II
I
,q
A
. (7.44)
Rewriting these equations in a matrix form, we obtain
_
1

F
1

FF
z
FI
z
IF
1

I
1

II
_
_
F
,q
A
I
,q
A
_
=
_
z
FA
0
_
. (7.45)
The coecients of the z-dependent matrix

M(z)
_
1

F
1

FF
z
FI
z
IF
1

I
1

II
_
(7.46)
are the main branch of solutions of the nonlinear equations (7.37)(7.38) at q
j
= 1 but at
arbitrary z.
One can verify using Eq. (7.39) that the matrix

M(z) is invertible near z = 0. Hence the
solution of Eq. (7.45) can be written, at least within some range of z where

M(z) remains
invertible, as
_
F
,q
A
I
,q
A
_
=

M
1
(z)
_
z
FA
0
_
. (7.47)
Similarly, the derivatives F
,q
F
and I
,q
F
satisfy the equations

M(z)
_
F
,q
F
I
,q
F
_
=
_
0
z
IF
F
_
, (7.48)
whose solution is
_
F
,q
F
I
,q
F
_
=

M
1
(z)
_
0
z
IF
F
_
. (7.49)
Other generating functions can be expressed in the same manner.
One could in principle obtain a numerical solution for F
,q
A
(z), I
,q
A
(z), and all the other
generating functions at any given value of z. However, the numerical solution is not particularly
useful at this point because the next step in the calculation is the evaluation of Eq. (7.42) in the
limit of very large n
tot
, for instance,
p(A[n
tot
)
p(F[n
tot
)
=

ntot
z
F
,q
A

ntot
z
F
,q
F

z=0
. (7.50)
It is generally not feasible to compute the n-th derivative of a numerically obtained function in
the limit n , because the unavoidable round-o errors are amplied by a xed factor with
each successive numerical dierentiation. Therefore, we need a way to evaluate the ratio (7.42)
in the limit n
tot
without using numerics. Indeed we will be able to compute the ratio
p(A)/p(F). We will also show that, for instance,
p(A)
p(F)
= lim
n

n
z
F
,q
A

n
z
F
,q
F

z=0
= lim
n

n
z
I
,q
A

n
z
I
,q
F

z=0
; (7.51)
87
7 The RV measure for the landscape
in other words, that the nal RV-regulated ratio p(A)/p(F) is independent of whether the initial
bubble is of type F or of type I.
To proceed, we use the fact that F
,q
A
(z), I
,q
A
(z), F
,q
F
(z), etc. are analytic functions of the
parameter z. The asymptotic growth of high-order derivatives of an analytic function f(z) is
determined by the location of the singularities of f(z) in the complex z plane, as shown in
Eq. (7.15).
We now need to determine the location of the singularities of F
,q
A
(z), I
,q
A
(z), etc., as functions
of z. This task is much simplied once we observe that all these quantities are expressed through
the inverse matrix

M
1
(z), and hence it remains to analyze the singularities of

M
1
(z). That
matrix can be singular at some value z = z

either because some of the coecients of

M(z) are
singular, or because the matrix

M(z) is degenerate (noninvertible) at z = z

. The coecients
of

M(z) depend on solutions F(z), I(z) of algebraic equations (7.37)(7.38) at q
j
= 1 and thus
cannot be divergent as functions of z. These coecients can be singular only in that some
derivative in z diverges. The derivatives F
,z
and I
,z
satisfy the equations

M(z)
_
F
,z
I
,z
_
=
_

FA
+
FI
I

IB
+
IF
F
_
. (7.52)
Therefore, F
,z
and I
,z
diverge only for those z for which the matrix

M(z) is degenerate. For
these z, all the derivatives F
,z
, I
,z
, F
,q
A
, etc. will be divergent at the same time since they are
all proportional to the inverse matrix

M
1
(z).
We note that the matrix

M(z) is the Jacobian of the nonlinear system (7.37)(7.38). As long
as

M(z) is nondegenerate, all the dierent branches of the solutions F(z), I(z) do not meet
and remain smooth functions of z. As we noted above, the main branch F(z), I(z) is the one
connected to the nontrivial solution (7.39) at z 0. The matrix

M(z) is nondegenerate near
z = 0; therefore, the solutions F(z), I(z) remain smooth functions of z for all z such that

M(z)
is nondenegerate. We conclude that the only possible singularities of F(z), I(z) are those values
z = z

where det

M(z) = 0.
Below it will be shown that the behavior of det

M(z) near z = z

is
det

M(z) c
1

z, (7.53)
i.e. of the form (7.14) with c
0
= 0 and s =
1
2
. It now follows from Eqs. (7.47) and (7.49) that the
derivatives such as F
,q
A
(z), I
,q
A
(z), F
,q
F
(z), etc., all diverge at z = z

with the same asymptotic

behavior, namely proportional to (z

z)
1/2
. We may express the inverse matrix as

M
1
(z) =
1
det

M(z)

M(z), (7.54)

M(z)
_
1

I
1

II
z
FI
z
IF
1

F
1

FF
_
, (7.55)
where we introduced the algebraic cofactor matrix

M

M
1
det

M, which remains well-dened
at z = z

. It then follows from Eq. (7.47) that the asymptotic behavior of the functions F
,q
A
(z)
88
7.3 A toy landscape
and I
,q
A
(z) near z = z

is given by
_
F
,q
A
I
,q
A
_

1
c
1

M(z

)
_
z

FA
0
_
=
1
c
1

z
_ _
1

I
1

II
_
z

FA
z
2

IF

FA
_
. (7.56)
Similarly, using Eq. (7.49) we nd near z = z

_
F
,q
F
I
,q
F
_

1
c
1

M(z

)
_
0
z

IF
F
_
=
1
c
1

z
_
z
2

FI

IF
F
_
1

F
1

FF
_
z

IF
F
_
. (7.57)
Using Eq. (7.15) and noticing that all n-dependent factors are the same in
n
z
F
,q
A
and other
such derivatives, we conclude that
lim
n

n
z
F
,q
A

n
z
F
,q
F

z=0
= F
1

_
1

I
1

II
_

FA
z

FI

IF
F

, (7.58)
lim
n

n
z
I
,q
A

n
z
I
,q
F

z=0
=
z

FA
_
1

F
1

FF
_
F

. (7.59)
It is important that the two limits above are equal; this is so because the condition det

M(z

) = 0
yields
det

M =
_
_
F
1

FF
_
_
_
_
I
1

II
_
_
z
2

FI

IF
= 0, (7.60)
where we denoted F

F(z

), I

I(z

) for brevity. It follows that the ratio of A-bubbles to

F-bubbles is independent of whether the initial bubble is of type F or of type I. In other words,
the RV-regulated ratio of A-bubbles to F-bubbles is
p(A)
p(F)
=
z

FA
_
1

F
1

FF
_
F

(7.61)
independently of the initial bubble type. Below (Sec. 7.4.5) the independence of initial conditions
will be rigorously proved for a general landscape using mathematical techniques of the theory
of nonnegative matrices. Presently we have shown this independence using explicit formulas
available for the toy landscape under consideration.
In a similar way, we nd the RV-regulated ratio of A-bubbles to B-bubbles,
p(A)
p(B)
=
z

FA

IF
_
1

F
1

FF
_

IB
, (7.62)
89
7 The RV measure for the landscape
and the ratio of F-bubbles to I-bubbles,
p(F)
p(I)
=
_
1

F
1

FF
_
F

FI
I

. (7.63)
It remains to compute z

, F

, I

and to justify Eq. (7.53). We will do this using the explicit

form of the matrix

M(z). To determine z

, F

, and I

, we need to solve Eq. (7.60) simultaneously

with the equations
F
1

= z

FA
+z

FI
I

+
FF
F

, (7.64)
I
1

= z

IB
+z

IF
F

+
II
I

, (7.65)
the latter two being Eqs. (7.37)(7.38) after setting q
j
= 1. We are interested in the solution
z

closest to z = 0. By denition (7.35), the generating functions are nonsingular for [z[ 1;
hence, the only possible values of z

are in the domain [z[ > 1 in the complex plane.

Moreover, we can show that the solutions F(z), I(z) are growing functions of z for real z < z

.
Initially at z 0 these functions have the form (7.39) and hence are positive. The determinant
det

M shown in Eq. (7.60) is also positive for z < z

. One can then nd from Eq. (7.55) that

the algebraic cofactor matrix

M and hence the inverse matrix

M
1
has all positive elements for
those z. It is then evident from Eq. (7.52) that
z
F > 0 and
z
I > 0 as long as the values of
F and I are themselves positive. Therefore, the solutions F(z), I(z) are positive and growing
functions of z as long as det

M(z) > 0.
We may thus visualize the behavior of det

M(z) as z grows. Both F(z) and I(z) will grow
with z, so that the terms in square brackets diminish while the second term, z
2

FI

IF
, grows.
Eventually the product of the square brackets in Eq. (7.60) will be balanced by the second term,
and the determinant will vanish. We need to determine the smallest value z = z

for which
det

M(z) = 0.
Since the physically signicant branch of the solution involves always very small values F(z)
and I(z) for [z[ 1, while other (unphysical) branches have either F or I approximately equal
to 1, it is reasonable to assume that F(z

) 1 and I(z

) 1 also at z = z

(the self-consistency
of this assumption will be conrmed by later calculations). Then the terms
FF
F

and

II
I

can be disregarded in comparison with F

1/

and I
1/

in Eqs. (7.64)(7.65). In this

approximation, we can simplify Eqs. (7.64)(7.65) to
F
1

= z

FA
+z

FI
I

, (7.66)
I
1

= z

IB
+z

IF
F

, (7.67)
while Eq. (7.60) becomes
(
FA
+
FI
I

) (
IB
+
IF
F

) =
2

FI

IF
F

. (7.68)
The ratios (7.61)(7.63) are also simplied and can be written more concisely as
p(A) : p(I) : p(F) : p(B) =
FA
: (
FI
I

)
:

FA
+
FI
I

:
IB

FA
+
FI
I

IF
F

. (7.69)
90
7.3 A toy landscape
To determine z

, we will obtain an explicit approximation for the main branch F(z), I(z) for all
z. For small enough z, the solutions of Eqs. (7.64)(7.65) are well approximated by Eq. (7.39),
F(z) (z
FA
)

, I(z) (z
IB
)

. (7.70)
These solutions are obtained under the assumption that the terms z
FA
and z
IB
are numer-
ically small (z
FA
1, z
IB
1) and yet dominant in Eqs. (7.64)(7.65). These terms only
remain dominant as long as

FA

FI
I(z),
IB

IF
F(z). (7.71)
Substituting Eq. (7.39) for F(z) and I(z) into Eq. (7.71), we obtain the conditions
z
1
IB
_

FA

FI
_1

, z
1
FA
_

IB

IF
_1

. (7.72)
We need to check whether det

M(z) could vanish already for some z

within the range (7.72).

Using Eq. (7.68) in the regime (7.71), we nd
z

_
(
FA

IB
)
1

FI

IF
_

1
2
. (7.73)
This value is within the range (7.72) only if the following simultaneous inequalities hold,
e
6

2

_

IB

FA
_
+1

FI

IF

2
e
6
. (7.74)
Let us denote by the quantity in Eq. (7.74),

_

IB

FA
_
+1

FI

IF
, (7.75)
then Eq. (7.74) becomes simply [ln [ < 6. One would expect that is generically either very large
or very small, so the inequalities (7.74) can hold only in a ne-tuned landscape. Additionally,
we need to require
z

min
_
1

FA
,
1

IB
_
, (7.76)
which entails

IB

IF

,

FA

FI

. (7.77)
However, this requirement is weaker than Eq. (7.74) since typically
FA

FI
and
IB

IF
,
so
z

1
IB
_

FA

FI
_1

1
IB
; z
1
FA
_

IB

IF
_1

1
FA
. (7.78)
Let us assume, for the moment, that Eqs. (7.74) and (7.77) hold. Then we use Eq. (7.69) to
obtain
p(A) : p(I) : p(F) : p(B) 1 :

:
1

. (7.79)
91
7 The RV measure for the landscape
Due to the ne-tuning assumption (7.74), these ratios are all within the interval
_

2
,
2

=
_
e
6
, e
6

.
Having considered the ne-tuned case, let us now turn to the generic case. Generically one
would expect that the quantity is either extremely large or extremely small, and in any case
outside the logarithmically narrow range [e
6
, e
6
]. In that case, one of the inequalities (7.74)
does not hold; generically, either
2
or
2
. It follows that det

M(z) ,= 0 for all z
within the range (7.72). To reach the value z

at which det

M(z

) = 0, we need to increase z
further, until one of the terms z
FI
I or z
IF
F in Eqs. (7.64)(7.65) becomes dominant and the
solution (7.70) becomes invalid.
It is impossible that both the terms z
FI
I and z
IF
F are dominant in Eqs. (7.64)(7.65) at
z = z

because then Eq. (7.68) would yield a contradiction,

FI
I

IF
F

=
2

FI

IF
F

. (7.80)
Hence, the determinant det

M(z) rst vanishes at the value z = z

such that only one of those

terms is dominant in its respective equation. Without loss of generality, we may relabel the
vacua (F I, A B) such that the term z
IF
F becomes dominant in Eq. (7.65) while the
term z
FA
is still dominant in Eq. (7.64). This is equivalent to assuming
2
. In the range
of z for which this is the case,
z
FI
I(z) z
FA
1; z
IB
z
IF
F(z) 1, (7.81)
the approximate solution of Eqs. (7.64)(7.65) can be written as
F(z) (z
FA
)

; I(z) (z
IF
F)

2
FA

IF
z
(+1)
. (7.82)
With these values of F(z) and I(z), the consistency requirement (7.81) yields the following range
of z,

IB

IF

FA
z

2
1
FA

FI

IF
_

1
+1
. (7.83)
This range is nonempty if

_

IB

FA
_
+1

FI

IF
1, (7.84)
which is indeed one of the two possible ways that the ne-tuning (7.74) can fail. Having assumed
that the above condition holds, we need to determine the value of z

and check that it belongs

to the range (7.81). Using Eq. (7.68) in the regime (7.81), we obtain
z

IF

FI

2
1
FA
_

1
(+1)
. (7.85)
Substituting this value into the inequalities (7.83), we nd the condition

1

2
1, (7.86)
which holds identically under the current assumption,
2
. (The relabeling F I, A B
is necessary if the opposite case,
2
, holds.) Therefore, z

is within the regime (7.81), and

our approximations are self-consistent, yielding
F

(z

FA
)

=

FA
_

IF

FI
1
+1
, I

FA

FI
. (7.87)
92
7.3 A toy landscape
The ratios (7.69) become
p(A) : p(I) : p(F) : p(B) 1 :
1

2
:
1

:
_

1
_ 1
+1
. (7.88)
This is the result of applying the RV prescription to a generic FABI toy landscape in the second
regime.
It remains to justify the statement of Eq. (7.53). Below in Sec. 7.4.5 I will demonstrate that
the property (7.53) holds for a general landscape. Here only a simple argument is presented to
illustrate this property for Eqs. (7.66)(7.67). Using those equations, we can express I

through
F

and derive a closed algebraic equation for F

,
F

= z

[
FA
+
FI
z

(
IB
+
IF
F

f(z; F

). (7.89)
The solution F(z) is given by the intersection of the line y = F and the curve y = f(z; F) in the
y F plane. The function f(z; F) is convex in F for F > 0, z > 0; hence, there will be a value
z = z

for which the curve f(z

; F) is tangent to the line y = F, i.e. f

,F
(z

; F

) = 1. This value
of z

will then implicitly determine F

. For values F F

, z z

the dependence of F(z) on z

will exhibit the singularity behavior of the type

z

z. To see this formally, we may expand

F = f(z; F) F

+f
,F
(F F

) +
1
2
f
,FF
(F F

)
2
+f
,z
(z z

) . (7.90)
Since f
,F
(z

; F

) = 1, we obtain
F F

+
2f
,z
f
,FF

F,z

z. (7.91)
This shows explicitly the singularity structure of the form (7.14). The determinant of the matrix

M is a smooth function of F(z) and I(z) near z = z

. Expressing det

M as a function only of
F, we obtain for z z

the required formula (7.53),

det

M(z) (F F

)
d
dF

F
det

M

z. (7.92)
7.3.2 Boltzmann brains
Let us now use the same techniques to compute the relative abundance of Boltzmann brain
observers to ordinary observers.
We need to introduce an appropriate set of generating functions. Boltzmann brains can be
created with a xed probability per unit 4-volume, unlike ordinary observers who can appear
only within a narrow interval of time after creation of a given bubble. Let us therefore compare
the total number of bubbles n
j
with the total number of 4-volumes H
4
j
in bubbles of type j.
To be specic, let us x j = F (no Boltzmann brains can be expected in a terminal vacuum).
Consider the probability p(n
tot
, n
F
, N
F
; k) of having a nite total number n
tot
of bubbles of
which n
F
are of type F and N
F
is the total number of H-regions of type F, if one starts from
a single initial H-region of type k (k = F, I). The number of Boltzmann brains in bubbles of
type F is proportional to N
F
with a small proportionality constant,
BB
F
, which we will include
93
7 The RV measure for the landscape
at the end of the calculation. A generating function g for the probability distribution p can be
dened by
g(z, q, r; k)

n,n
F
,N
F
0
z
n
q
n
F
r
N
F
p(n, n
F
, N
F
; k)

z
n
q
n
F
r
N
F
_
n<;k
. (7.93)
For the initial bubble types k = F and k = I, let us denote for brevity
F(z, q, r) g(z, q, r; F), I(z, q, r) g(z, q, r; I). (7.94)
The generating functions F and I satisfy the following system of equations,
F
1

= z
FA
+z
FI
I +r
FF
F, (7.95)
I
1

= z
IB
+zrq
IF
F +
II
I. (7.96)
These equations dier from the analogous Eqs. (7.37)(7.38) in that the generating parameter
q appears when a new bubble of type F is created, while the parameter r appears every time a
new H-region of type F is created, which can happen via Hubble expansion of old F-bubbles as
well as through nucleation of new F-bubbles.
We would like to compare the mean number of H-regions of type F with the mean number
of new bubbles of type F, so we compute (e.g. starting with I-bubbles)
N
F
)
n;I
n
F
)
n;I
=

n
z

r
I

n
z

q
I

z=0,r=1,q=1
(7.97)
and take the limit n . Taking the limit as n of the ratio
n
z
F
,r
/
n
z
F
,q
will yield the
same result, but the limit of Eq. (7.97) is more straightforwardly analyzed.
As before, we rst obtain the equations for the derivatives F
,q
, I
,q
, F
,r
, I
,r
at q = r = 1 from
Eqs. (7.95)(7.96). The derivatives F
,q
and I
,q
satisfy the same equations as before, namely
Eq. (7.48), while F
,r
and I
,r
satisfy

M(z)
_
F
,r
I
,r
_
=
_

FF
F
z
IF
F
_
, (7.98)
whose solution is
_
F
,r
I
,r
_
=

M
1
(z)
_

FF
F
z
IF
F
_
. (7.99)
Using the same arguments as in the previous section, we evaluate the limit of Eq. (7.97),
lim
n

n
z

r
I

n
z

q
I

z=0,r=1,q=1
= F
1
1

FF
+ 1. (7.100)
To analyze this simple result, we do not actually need to use the complicated decision procedure
of Sec. 7.3. Since
FF
1 and F

< 1 in any case,

4
the value (7.100) is bounded from above by
+ 1, which is not a large number. Hence, the mean total number of H-regions of type F is at
4
Typically F 1 but we can do with a weaker bound F < 1 here.
94
7.4 A general landscape
most + 1 times larger than the mean total number of bubbles of type F. (Both numbers are
nite in nite multiverses, and the relationship persists in the limit n
tot
.)
Analogous results are obtained for regions of type I. We simply need to replace F with I and

FF
by
II
in Eq. (7.100).
The Boltzmann brains are created at a very small rate
BB
j
per H-region of type j, while
ordinary observers are created at a much larger rate per reheated 3-volume H
3
j
in bubbles of
the same type. We conclude that the RV-regulated abundance of Boltzmann brains is always
negligible compared with the abundance of ordinary observers in bubbles of the same type.
7.4 A general landscape
In the previous section we performed computations in a simple toy model of the landscape. Let
us now consider a general landscape containing N vacua, labeled j = 1, ..., N. The dimensionless
transition rates
jk
between vacua j and k are considered known. The main task is to compute
the RV-regulated ratio of abundances of bubbles of kinds j and k. We will also compare the
abundances of ordinary observers with that of Boltzmann brains.
7.4.1 Bubble abundances
It is convenient to denote by T the set of terminal bubble types and to relabel the vacua such
that j = 1, ..., N
r
are the recyclable(nonterminal) vacua. Thus, T = N
r
+ 1, ..., N. We start
by considering the probability p(n, n
j
; k) of having n
j
bubbles of type j, with total n =

j
n
j
bubbles of all types, to the future of an initial bubble of type k , T. The generating function
g(z, q
j
; k) for this probability distribution can be dened by a straightforward generalization
of Eq. (7.35),
g(z, q
l
; k)

n,{n
l
}0
p(n, n
l
; k)z
n

j
q
n
j
j
. (7.101)
Here z is the generating parameter for n, and q
j
are the generating parameters for n
j
. The
generating function can be written symbolically as the average
g(z, q
l
; k)

z
n
q
n
1
1
...q
n
N
N
_
n<;k
, (7.102)
where the subscript (n < ) indicates that only the events with a nite total number of bubbles
contribute to the statistical average.
For terminal bubble types k T, the denition (7.101) yields g(z, q
j
; k) = 1 since there are
no further bubbles to the future of terminal bubbles, thus n = n
j
= 0 with probability 1.
The generating function g(z, q
j
; k), k = 1, ..., N
r
satises the system of N
r
nonlinear equa-
tions
g
1

(z, q
j
; k) =

iT,i=k
zq
i

ki
g(z, q
j
; i)
+

iT
zq
i

ki
+
kk
g(z, q
j
; k), (7.103)
where e
3
as before, and the quantities
kk
dened by Eq. (7.34). A derivation of
Eqs. (7.103) will be given below in Sec. 7.4.4.
95
7 The RV measure for the landscape
The solution of Eqs. (7.103) can be visualized as N
r
analytic functions g(z, q
j
; k), k =
1, ..., N
r
, of the free parameters z and q
j
(j = 1, ..., N). Arguments similar to those of Sec. 7.3
show that the physically signicant solution of Eqs. (7.103) is the main branch that has the
following asymptotic form at z 0,
g(z; k) = z

iT
q
i

ki
_

+O(z
21
) (7.104)
(see also the argument at the end of Sec. 7.4.4).
Once the generating function g(z, q
j
; k) is determined, the RV measure gives the ratio of
the mean number of bubbles of type j to that of type k as
p(j)
p(k)
= lim
n

n
z

q
j
g(z, q
i
; i

n
z

q
k
g(z, q
i
; i

z=0,{q
i
}=1
, (7.105)
where, for clarity, we wrote explicitly the type i

of the initial bubble. We will now use the

methods developed in Sec. 7.3.1 to reduce Eq. (7.105) to an expression that does not contain
limits and so can be analyzed more easily. It will then become evident that the limit (7.105) is
independent of i

.
For a xed bubble type j, we rst consider the derivatives
q
j
g(z, q
i
; k), k = 1, ..., N
r
,
evaluated at q
i
= 1. Let us denote these N
r
derivatives by h
j
(z; k),
h
j
(z; k)

q
j

q
i
=1
g(z, q
i
; k), k = 1, ..., N
r
. (7.106)
These quantities are conveniently represented by an N
r
-dimensional vector, which we will denote
by [h
j
(z)) using the Dirac notation (although no connection to quantum mechanics is present
here). This vector satises an inhomogeneous linear equation that follows straightforwardly by
taking the derivative
q
j
at q
i
= 1 of Eqs. (7.103). That equation can be written in the matrix
form as follows,
N

i=1
M
ki
(z)h
j
(z; i) = z
kj
g(z; j), k ,= j, (7.107)
N

i=1
M
ki
(z)g
,q
j
(z; i) = 0, k = j, (7.108)
where we denoted by g(z; j) g(z; q
i
= 1; j) the solution of Eqs. (7.103) at q
i
= 1, written
as
g
1

(z; k) = z

iT,i=k

ki
g(z; i)
+z

iT

ki
+
kk
g(z; k), (7.109)
while the matrix M
ki
(z) is dened by

M(z) M
ki
(z) =
_
1

g
1

1
(z; k)
kk
, k = i;
z
ki
, k ,= i.
(7.110)
96
7.4 A general landscape
For convenience we rewrite Eqs. (7.107)(7.108) in a more concise form,

M(z) [h
j
(z)) = [Q
j
(z)) , (7.111)
where [Q
j
) is the vector with the components [Q
j
)
i
, i = 1, ..., N
r
given by the right-hand sides
of Eqs. (7.107)(7.108),
[Q
j
(z))
i
z
ij
g(z; j) [1
ij
] . (7.112)
The solution of Eq. (7.111) is found symbolically as
[h
j
(z)) =

M
1
(z) [Q
j
(z)) , (7.113)
provided that the inverse matrix

M
1
(z) exists. Of course, it is impractical to obtain the inverse
matrix explicitly; but we will never need to do that.
The next task is to determine z for which the matrix

M(z) remains nondegenerate, det

M(z) ,=
0. It will be shown in Sec. 7.4.5 that there exists an eigenvalue
0
(z) of

M(z) such that
0
(z

) = 0
at some real z

> 0, while
0
(z) > 0 for z < z

. Moreover, the (left and right) eigenvectors

corresponding to the eigenvalue
0
(z) are always nondegenerate and can be chosen with all
positive components. At the same time, all the other eigenvalues of

M(z) remain nonzero for
all z z

. It will also be shown (see Sec. 7.4.6) that the following asymptotic expansion holds
near z = z

0
(z) = c
1

z +O(z

z), c
1
> 0. (7.114)
One can see from the small-z asymptotics (7.104) and from Eq. (7.110) that, for small enough
z,

M(z) contains very large positive numbers on the diagonal and very small negative numbers
o the diagonal. Hence det

M(z) > 0 for all suciently small z. Since the eigenvalues of

M(z) remain nonzero except for

0
(z) that rst vanishes at z = z

, we conclude that det

M(z)
remains positive for all 0 < z < z

, and that z = z

is the smallest positive value of z for which

det

M(z) = 0.
Now we restrict our attention to 0 < z < z

, for which det

M(z) > 0 and

M
1
(z) exists.
We need to analyze the behavior of

M
1
near z z

. We treat

M(z) as an operator in N
r
-
dimensional real space V . The matrix

M is not symmetric and may not be diagonalizable.
Instead of diagonalizing

M(z), we split the space V into the 1-dimensional subspace correspond-
ing to the smallest eigenvalue
0
(z), and into the (N
r
1)-dimensional complement subspace
V
1
. We know that the eigenvalue
0
(z) is nondegenerate. Thus, we may symbolically write

M(z) =
0
(z)

v
0
(z)
_
u
0
(z)

+

M
1
(z), (7.115)
where

v
0
_
and

u
0

are the right and the left eigenvectors corresponding to

0
(z), normalized
such that

u
0
[v
0
_
= 1, (7.116)
and it is implied that

M
1
(z) vanishes on

v
0
_
and

u
0

but is nonsingular on the complement

space V
1
. In other words, we have

v
0
_
=
0
[v
0
) ,

u
0

M =
0

u
0

, (7.117)

M
1

v
0
_
= 0,

u
0

M
1
= 0. (7.118)
97
7 The RV measure for the landscape
There exists a matrix

M
1
1(V
1
)
that acts as the inverse to

M
1
when restricted to the subspace V
1
and again vanishes on

v
0
_
and

u
0

. Using that matrix, we may write the inverse matrix

M
1
explicitly as

M
1
(z) =
1

0
(z)

v
0
(z)
_
u
0
(z)

+

M
1
1(V
1
)
(z). (7.119)
Hence, the solution (7.113) can be written as
[h
j
(z)) =

u
0
[Q
j
_

0
(z)

v
0
_
+

M
1
1(V
1
)
[Q
j
) . (7.120)
Now it is clear from Eq. (7.114) that [h
j
(z)) diverges at z = z

as (z

z)
1/2
. It also follows
that [h
j
(z)) does not diverge at any smaller real z.
We can also show that
0
(z) cannot vanish at some complex value of z that is closer to z = 0
than z = z

. If
0
(z

) = 0 with a complex-valued z

, then a derivative of the generating function,

such as
z
g(z; j), would diverge at z = z

. Using the denition (7.101) of g and substituting

q
i
= 1, we nd that the following sum diverges,
g(z; j)
z

z=z

n1
_
z

_
n1
np(n; j) = . (7.121)
The sum of absolute values is not smaller than the above, and hence also diverges:
g(z; j)
z

z=|z

|
=

n1

n1
np(n; j) = . (7.122)
So
z
g has also a singularity at a real value z = [z

[. As we have shown, z = z

is the smallest
such real-valued singularity point; hence [z

[ z

. It follows that z

is equal to the radius of

convergence of the series (7.121), which is a Taylor series for the function
z
g. There remains the
possibility that a singularity z

is located directly on the circle of convergence, so that [z

[ = z

and z

= z

e
i
with 0 < < 2. This possibility can be excluded using the following argument.
The function
z
g can have only nitely many singularities on the circle [z[ = z

; innitely many
singularities on the circle would indicate an accumulation point which would be an essential
singularity, i.e. not a branch point. However, by construction g(z; j) is an algebraic function of
z that cannot have singularities other than branch points. Since g(z; j) is regular at z = z

, all
branch points of g must be of the form (z z

)
s
with s > 0. Using the explicit formula (7.15) for
the n-th derivative of an analytic function with a branch cut singularity of the form (z z

)
s
,
we nd that the coecients np(n; j) of the Taylor series (7.121) decay at large n asymptotically
as
np(n; j) =
1
n!

n
z

z
g
(n + 1 s)
(n + 1)
_
z

_
n
_
1 +O(n
1
)
_
n
s
_
z

_
n
_
1 +O(n
1
)
_
. (7.123)
The function g has a nite number of singularities z

= z

e
i
on the circle [z[ = z

, and each
singularity gives a contribution of the form (7.123). Hence, we may estimate (for suciently
large n, say for n n
0
)
np(n; j) = c
0
n
s
z
n

_
1 +O(n
1
)
_
, c
0
> 0. (7.124)
98
7.4 A general landscape
Then the partial sum for n n
0
of the series (7.121) is estimated by

nn
0
_
z

_
n1
np(n; j)

nn
0
_
z

_
n1
c
0
n
s
z
n

=
c
0
z

nn
0
n
s
e
in
. (7.125)
The latter series converges for s > 0 and 0 < < 2. (The neglected terms of order n
1s
build
an absolutely convergent series and hence introduce an arbitrarily small error into the estimate.)
Therefore, the series (7.121) also converges at z

,= z

, contradicting the assumption that other

singularities exist on the circle [z[ = z

.
We conclude from these arguments that the singularity of g(z; j) nearest to z = 0 in the
complex z plane is indeed at a real value z = z

, and all other singular points z

satisfy the
strict inequality [z

[ > z

. The same statement about the locations of singularities holds for the
generating functions g(z, q
i
; j) and hence for their derivatives such as [h
j
(z)).
To compute the nal expression (7.105), we need to evaluate the n-th derivative
n
z
of the
functions h
j
(z; i) at z = 0 and for very large n. To this end, we use the formula (7.15), which
requires to know the location z = z

of the singularity of h
j
(z; i) nearest to z = 0. We have just
found that this singularity is at a real value z = z

and has the form

h
j
(z; i)
1
c
1

u
0
[Q
j
_
v
0
i
, z z

, (7.126)
where v
0
i
is the i-th component of the eigenvector [v
0
). The value of the proportionality constant
c
1
is not required for computing the ratios (7.105). Using Eqs. (7.15), (7.112), and (7.126), we
evaluate the limit (7.105) as
p(j)
p(k)
=

u
0
[Q
j
_
v
0
i

u
0
[Q
k
) v
0
i

z=z
=

i=j
u
0
i
(z

)
ij

i=k
u
0
i
(z

)
ik
g(z

; j)
g(z

; k)
. (7.127)
Since the component v
0
i

cancels, we nd that the limit (7.105) is indeed independent of the

initial bubble type i

. We also note that the normalization of the eigenvector

u
0

is irrelevant
for the ratio.
The formula (7.127) can be simplied further for nonterminal types j, k if we use the rela-
tionship

u
0
(z

M(z

) = 0 together with the explicit denition (7.110) of

M:
0 =

i=j
u
0
i
(z

)M
ij
(z

) +u
0
j
(z

)M
jj
(z

)
=

i=j
u
0
i
(z

)z

ij
+u
0
j
(z

)
_
1

g
1

1
(z

; j)
jj
_
. (7.128)
The last term in the square brackets in Eq. (7.128) can be neglected since
jj
1 while
g(z

; j) 1. Hence, Eq. (7.127) is simplied to

p(j)
p(k)

u
0
j
(z

)
u
0
k
(z

)
_
g(z

; j)
g(z

; k)
_
1/
. (7.129)
99
7 The RV measure for the landscape
This is the main formula for the RV-regulated relative abundances of bubbles of arbitrary (non-
terminal) types j and k. For a terminal type k T, one needs to use Eq. (7.127) together with
g(z; k) 1.
We note that the expression (7.129) depends on the components u
0
i
of the left eigenvector
of the matrix

M(z

) with eigenvalue
0
(z

) = 0. Although we have been able to evaluate

the limit (7.105) analytically and obtained Eq. (7.129), the task of computing the values of
z

, g(z

; j), and u
0
i
(z

) remains quite dicult. The outline of the required computations is

as follows: One rst needs to determine g(z; j) (j = 1, ..., N
r
) as the main branch of the
solution of Eqs. (7.109) with the small-z asymptotic given by Eq. (7.104). The functions g(z; j)
determine the matrix elements M
ij
(z) using Eq. (7.110). Then one needs to nd the smallest
value z = z

> 0 such that the determinant of the matrix

M(z) vanishes. Finally, one needs
to compute a left eigenvector

u
0
(z

that corresponds to the eigenvalue

0
(z

) = 0 of the
matrix

M(z

). The mathematical construction shown below guarantees that z

exists and that

u
0
(z

is nondegenerate and has all positive components; the results are independent of the
normalization of

u
0

. However, a brute-force numerical computation of these quantities appears

to be impossible due to the huge number N
r
of the simultaneous equations (7.109) and to the
wide range of numerical values of the coecients
ij
in a typical landscape. Even the numerical
value of z

is likely to be too large to be represented eciently in computers. In the next section

we will consider an example landscape where an analytic approximation can be found.
7.4.2 Example landscape
We begin with some qualitative considerations regarding the behavior of the functions g(z; k).
To determine g(z

; k), we need to follow the main branch g(z; k) as the value of z is increased
from z = 0 until det

M(z) vanishes, which will determine the value z = z

. We note that the

asymptotic form (7.104) is valid in some range near z = 0. Since g(z; k) 1 at those z, the
matrix

M(z) is dominated by large positive diagonal terms g
1

1
(z; k)
kj
. As z increases to z

, all
the functions g(z; j) also increase while det

M(z) decreases monotonically to zero (this statement
is proved rigorously in Sec. 7.4.5). One can visualize the changes in the matrix elements of

M(z)
if one notes that the positive diagonal terms decrease with z while the negative o-diagonal
terms (M
jk
= z
jk
, j ,= k) grow in magnitude. Eventually det

M(z) vanishes at z = z

such
that the o-diagonal terms become suciently large at least in some rows and columns of the
matrix

M(z).
Let us determine an upper bound on z such that Eq. (7.104) remains a good approximation for
the generating function g(z; j) g(z, q
i
= 1 ; j). This will be the case if the term z

iT

ki
in the right hand side of Eq. (7.109) dominates over all other terms, for every k. Denoting for
brevity

kT

iT

ki
(7.130)
the total transition probability from k to all terminal vacua, we have therefore the conditions
(for every k)
z
kT
g(z; k), z
kT
z

iT,i=k

ki
g(z; i). (7.131)
These conditions are satised, consistently with Eq. (7.104), if z is bounded (for every k) simul-
100
7.4 A general landscape
taneously by
z
1
kT
, z
_
_

iT,i=k

ki

kT

iT
_
_

. (7.132)
Since typically the rate of transitions to terminal vacua is much smaller than the rate of transi-
tions to dS vacua, one can expect that this regime will include z = 1. Nevertheless, one expects
that det

M(z) remains nonzero for these z. The detailed consideration of the FABI model in
Sec. 7.3.1 showed that the value of z

lies outside the regime (7.132) unless the landscape param-

eters are ne-tuned. In a general landscape, one expects the analogous ne-tuning to be much
stronger or even impossible to satisfy. In other words, one expects that the functions g(z

; k) at
least for some k will violate the conditions (7.131), although these conditions might be satised
for other k.
Now we would like to estimate the value of z for which the matrix

M(z) rst becomes degen-
erate. We can use a general theorem for estimating the eigenvalues of a matrix (Gershgorins
theorem, see [179], chapter 7). For each of the diagonal elements M
kk
, k = 1, ..., N
r
one needs
to draw a circle in the complex plane, centered at the diagonal element M
kk
with radius

i=k
[M
ki
[ . (7.133)
Thus one obtains N
r
circles
[ M
kk
[
k
, k = 1, ..., N
r
, (7.134)
called the Gershgorin circles. The Gershgorin theorem says that all the eigenvalues of a matrix
M
jk
are located within the union of these circles. An elementary proof is as follows. An
eigenvector v
i
with eigenvalue can be normalized such that its component of largest absolute
value is equal to 1. Let v
1
= 1 be this component (renumbering the components if necessary),
then it is easy to show that is within a circle with center M
11
and radius
1
. Namely, we start
from the eigenvalue equation,

i=1
M
1i
v
i
+M
11
v
1
= v
1
, (7.135)
and use the properties v
1
= 1 and [v
i
[ 1 for i ,= 1:
[ M
11
[ =

i=1
M
1i
v
i

i=1
[M
1i
[ =
1
. (7.136)
Applying the theorem to the matrix

M(z) at small z such that Eq. (7.104) is a good approx-
imation for the generating function g(z; j), we nd that the Gershgorin circles are centered at
large positive values
M
kk
=
1

g
1

1
(z; k)
kk

1

1
(z
kT
)
1
1, (7.137)
while the radii of the circles are small,

k
= z

i=k,iT

ki
1. (7.138)
101
7 The RV measure for the landscape
Hence, for small z none of the circles can contain = 0, so the matrix

M(z) is nondegenerate.
As z increases, the radii
k
increase while the centers M
kk
move monotonically towards zero.
The circle closest to = 0 is the one with the center closest to zero and the largest radius.
This circle is labeled by k with largest rates
kT
and
ki
. This value of k corresponds to the
high-energy vacua in the landscape, for which tunneling to any other vacua is much easier than
for low-energy vacua. Therefore the eigenvalue (z

) = 0 will be located inside the Gershgorin

circle(s) centered at M
kk
(z

) for k corresponding to high-energy vacua. The Gershgorin circle

to which the eigenvalue belongs indicates the largest component of the eigenvector. Hence, one
expects that the eigenvector

u
0
(z

has the largest values of its components u

0
k
corresponding
to the high-energy vacua k. The Gershgorin theorem requires that
0
(z

) = 0 be inside the
circle centered at M
kk
(z

) with radius
k
(z

), i.e.
M
kk
(z

) =
1

g
1

1
(z

; k)
kk
z

i=k,iT

ki
. (7.139)
Disregarding the terms
kk
, which are small in comparison with the remaining terms, and
using Eq. (7.109) for g(z

; k) we obtain the following condition,

z

kT
+

i=k,iT

ki
g(z

; i)
_
1

i=k,iT

ki
. (7.140)
Proceeding further requires at least an estimate of g(z

; i) for all i, which is so far not available.

While we are as yet unable to obtain an explicit estimate of z

and

u
0

for an arbitrary
landscape, let us consider a workable example that is more realistic than the FABI model. In
this example there is a single high-energy vacuum, labeled k = 1, and a large number of dS
low-energy vacua, k = 2, ..., N
r
, as well as a number of terminal vacua, k = N
r
+ 1, ..., N. We
assume that the downward tunneling rates
1i
for i ,= 1 are much larger than the rates
i1
or
ij
for i, j ,= 1. Then g(z; 1) g(z; i) for i ,= 1, and so the rst Gershgorin circle is the one
closest to = 0. Hence we may expect that the eigenvalue
0
(z) always belongs to that circle.
Moreover, it is likely that z

is within the regime (7.131) for the low-energy vacua (k ,= 1) but

not for the high-energy vacuum k = 1. We will now proceed with the calculation and later check
that these assumptions are self-consistent.
If the regime (7.131) holds for every low-energy vacuum k ,= 1, we have
g(z; k) z

kT
, k = 2, ..., N
r
. (7.141)
For the high-energy vacuum k = 1 we use Eq. (7.109) directly to nd
g(z; 1) =
_
_

1T
+z

iT,i=1

1i
g(z; i) +
11
g(z; 1)
_
_

2
+
_
Nr

i=2

1i

iT
_

, (7.142)
where we have neglected the terms with
1T
and
11
. The condition that the regime (7.131)
holds for k ,= 1 is

iT
z

j=i,jT

ij
g(z; j), i = 2, ..., N
r
. (7.143)
102
7.4 A general landscape
In the matrix

M(z) the o-diagonal elements M
ij
for i ,= 1, j ,= 1 are much smaller than all
other elements. Hence, we can approximate

M(z) by a matrix that has only its rst row, rst
column, and the diagonal elements,

M(z)
_
_
_
_
_
M
11
z
12
z
13
z
1Nr
z
21
M
22
0 0
.
.
. 0
.
.
.
.
.
.
z
Nr1
0 0 M
NrNr
_
_
_
_
_
,
M
kk

1

g
1

1
(z; k). (7.144)
This approximate matrix is much easier to analyze; in particular, its determinant and the eigen-
vectors can be computed in closed form. The determinant of this matrix is
det

M(z) M
11
...M
NrNr
_
1
Nr

i=2
z
2

1i

i1
M
11
(z)M
ii
(z)
_
. (7.145)
Therefore, the condition det

M(z

) = 0 can be written as
M
11
(z

) z
2

Nr

i=2

1i

i1
M
ii
(z

)
. (7.146)
Using Eqs. (7.141), (7.142), and (7.144), we transform this condition to
z

2
_
Nr

i=2

1i

iT
_
1
Nr

i=2

1i

i1

1
iT
. (7.147)
However, one does not need this value apart from checking explicitly that the assumptions (7.143)
hold.
The left eigenvector

u
0
(z

corresponding to the eigenvalue 0 of the matrix

M(z

) is found
approximately as
u
0
1
= 1, u
0
k

z

1k
M
kk
(z

)
, (7.148)
where the normalization u
0
1
= 1 was chosen arbitrarily for convenience. A perturbative improve-
ment of this approximation along the lines of Refs. [87, 223, 224, 225] may be possible, but the
precision obtained from the present approximation is sucient for our purposes.
The bubble abundance ratios (7.129) for nonterminal types are then expressed as follows,
p(j)
p(k)


1j

1k
_

jT

kT
_

, j, k = 2, ..., N
r
, (7.149)
p(1)
p(k)

1

Nr
i=2

1i

iT

1k

kT
, k = 2, ..., N
r
. (7.150)
These equations are the main result of this section, yielding RV-regulated bubble abundances
for a landscape with a large number of low-energy vacua. The formula (7.150) agrees with that
obtained in Sec. 7.3.1 for the ratio p(F)/p(I) 1/ in the FABI landscape, which may be
considered a special case of the present model with N
r
= 2.
103
7 The RV measure for the landscape
7.4.3 Boltzmann brains
We now investigate the abundance of Boltzmann brains in a general landscape, relative to the
abundance of ordinary observers.
We rst need to derive the equations for the suitable generating functions, analogously to
Eqs. (7.95)(7.96). Let us introduce the generating function
g(z, q
i
, r
i
; j)
_
z
ntot

i
q
n
i
i
r
N
i
i
_
ntot<;j
, (7.151)
where n
i
is the total number of bubbles of type i and N
i
is the total number of H-regions of type
i. The number of BBs in bubbles of type i is proportional to N
i
with a proportionality constant

BB
i
, which is the (extremely small) nucleation rate of a BB per Hubble 4-volume H
4
i
.
The generating function g(z, q
i
, r
i
; j), which we will denote for brevity by g(..., j), satises
a system of equations analogous to Eq. (7.103),
g
1

(..., j) =

k=j

jk
zq
k
r
k
g(..., k) +
jj
r
j
g(..., j). (7.152)
Let us compute the RV-regulated ratio of the number of H-regions of type j to the number of
bubbles of the same type j. This ratio is given by
lim
n
N
j
)
n
n
j
)
n
= lim
n

n
z

r
j
g(..., i

n
z

q
j
g(..., i

z=0,r
i
=1,q
i
=1
, (7.153)
where i

is the initial bubble type. To evaluate the limit, we use the methods developed in
Sec. 7.4.1. The derivatives
q
j
g [h
j
(z)), as dened in Eq. (7.106), satisfy Eq. (7.111). The
vector of derivatives

r
j

r
i
=1,q
i
=1
g(..., k) [
j
(z))
k
(7.154)
satises the linear equations that follow from Eq. (7.152),

M(z) [
j
(z)) = [
j
(z)) , (7.155)
where [
j
) is the vector with the components [
j
)
i
, i = 1, ..., N
r
dened by
[
j
(z))
i
z
ij
g(z; j) [1
ij
] +
jj
g(z; j)
ij
. (7.156)
By the same considerations that lead to Eq. (7.127), we now obtain
lim
n

n
z

r
j
g(..., i

n
z

q
j
g(..., i

z=0,r
i
=1,q
i
=1
=

u
0
[
j
_
u
0
[Q
j
)

z=z
, (7.157)
where, as before,

u
0
(z

is the unique eigenvector of the matrix

M(z

) with eigenvalue 0. (The

normalization of

u
0

is again irrelevant.) Using the denitions (7.112) and (7.156) and the
relationship (7.128), we compute

u
0
[
j
_
u
0
[Q
j
)

z=z
=
u
0
j
(z

)
jj
+

i=j
u
0
i
(z

)z

ij

i=j
u
0
i
(z

)z

ij
=
1

g
1

1
(z

; j)
1

g
1

1
(z

; j)
jj
. (7.158)
104
7.4 A general landscape
The last ratio is always very close to 1 since typically g(z

; j) 1 while
jj
1. In particular,
g(z

; j) 1 for vacua j with low-energy Hubble scale, since for those vacua we may approximate
g(z

; j) by Eq. (7.104),
g(z

; j)
_
z

iT

ji
_

1. (7.159)
Therefore, the RV-regulated ratio N
j
/n
j
of the total number of H-regions of type j to the total
number of bubbles of type j is never large.
Using this result, we can now estimate the RV-regulated ratio of BBs to ordinary observers.
The number of ordinary observers per one H-region of type j is not precisely known but is
presumably at least of order 1 or larger, as long as bubbles of type j are compatible with life.
On the other hand, the number of Boltzmann brains per H-region, i.e. within a four-volume
H
4
j
of spacetime, is negligibly small. It follows that the abundance of BBs in the RV measure
is always negligible relative to the abundance of ordinary observers in the same bubble type.
7.4.4 Derivation of Eq. (7.103)
To derive Eq. (7.103), we need to consider the expansion of a single initial H-region during one
Hubble time. Within theination in a boxmodel, an initial H-region of type j is split after one
Hubble time t = H
1
j
into e
3
statistically independent daughter H-regions.
5
Each of the
daughter H-regions may change its vacuum type from j to k ,= j (k = 1, ..., N) with probability

jk
. The quantity
jj
was dened for convenience by Eq. (7.34) to be the probability of not
changing the bubble type j during one Hubble time.
The generating function g is dened by Eq. (7.101),
g(z, q
i
; j)

z
ntot
q
n
1
1
...q
n
N
N
_
ntot<;j
(7.160)
where n
i
is the number of bubbles of type i, while the notation ...)
ntot<;j
stands for a prob-
abilistic average evaluated for the initial H-region of type j on the sub-ensemble of nite total
number of bubbles n
tot
. Note that the initial bubble of type j is not counted in n
i
or n
tot
. Our
goal is to obtain a relationship between g(z, q
i
; j) and the generating functions g(z, q
i
; k)
with k ,= j.
To this end, we equate two expressions for the average z
ntot

i
q
n
i
i
)
ntot<;j
. The rst ex-
pression is the left-hand side of Eq. (7.160). The second expression is found by considering the
daughter H-regions evolved out of the initial H-region and by using the fact that the same
average for a daughter region of type k is equal to g(z, q
i
; k). However, two details need to be
accounted for: First, the generating functions g(z, q
i
; k) evaluated for the daughter regions do
not count the daughter bubbles themselves. Second, the type k of each of the daughter regions
is a random quantity. Before deriving a general relationship, let us illustrate the procedure using
an example.
It is possible that, say, only two of the daughter regions change their type to k while all other
daughter regions retain the initial bubble type j. Denote temporarily by p
kkj...j
the probability
of this event; binomial combinatorics yields
p
kkj...j
=
!
2!( 2)!

2
jk

2
jj
. (7.161)
5
To avoid considering a non-integer number of daughter regions, we may temporarily assume that is an
integer parameter. At the end of the derivation, we will set e
3
20.1 in the nal equations.
105
7 The RV measure for the landscape
Then the average of z
ntot

i
q
n
i
i
receives a contribution
p
kkj...j
z
2
q
2
k
[g(z, q
i
; k)]
2
[g(z, q
i
; j)]
2
(7.162)
from this event. The factor z
2
describes two additional bubbles that contribute to n
tot
; the factor
q
2
k
accounts for two additional bubbles of type k; no factors of q
j
appear since no additional
bubbles of type j are generated. Finally, the powers of g account for all the bubbles generated
in the daughter H-regions, but these generating functions do not count the daughter H-regions
themselves. Those daughter H-regions are explicitly counted by the extra factors z
2
q
2
k
.
To compute the average, we need to add the contributions from all the possible events of
this kind. Since all of the daughter H-regions are independent and statistically equivalent, the
average z
ntot

i
q
n
i
i
)
ntot<;j
splits into the product of averages, each evaluated over a single
daughter region.
The average over a single daughter region of type j has contribution from transitions to
other types k ,= j and a contribution from the event of no transition. Let us rst consider a
terminal type k. With probability
jk
a given daughter region becomes a vacuum of type k.
Thereafter, no more bubbles will be nucleated inside it; the average of z
ntot

i
q
n
i
i
)
ntot<;j
over
that daughter region is simply zq
k
, meaning that there is a total of one bubbles and only one
bubble of type k. Hence, the contribution of that event to the statistical average is
jk
zq
k
.
Now let us consider a nonterminal type k ,= j, k , T. The corresponding contribution
to the average is
jk
zq
k
g(z, q
i
; k). Since we have dened g(z, q
i
; k) 1 when k is a
terminal bubble type, we may write the contribution as
jk
zq
k
g(z, q
i
; k) for both terminal
and nonterminal types k ,= j.
Finally, we consider the case of k = j (the daughter region retains the original bubble type).
Since no new bubbles were nucleated, the contribution to the average is simply
jj
g(z, q
i
; j)
without any factors of z or q
k
.
Putting these ingredients together, we obtain an equation for g(z, q
i
; j),
g(z, q
i
; j)
=
_
_

k=j

jk
zq
k
g(z, q
i
; k) +
jj
g(z, q
i
; j)
_
_

. (7.163)
This is equivalent to Eq. (7.103). One can also verify that the binomial expansion of Eq. (7.163)
indeed yields all the terms such as the one given in Eq. (7.162).
We can now analyze the behavior of g(z, q
i
; j) in the limit z 0. By denition, the
generating function g(z, q
i
; j) is a power series in z whose coecient at z
n
is equal to the
probability of the event that a multiverse has exactly n bubbles to the future of the initial bubble
j. It is clear that this power series starts with the term z

, corresponding to the probability

that the initial bubble j expands exactly into terminal bubbles, signalling the global end of
the multiverse. The probability of having fewer than bubbles in the entire multiverse is equal
to zero.
6
The next term of the binomial expansion is of order z
21
since it is the product of
6
The property that there are exactly daughter bubbles is, of course, an artifact of the ination in a box
approximation. In the actual multiverse, one has bubbles of spherical shape that can intersect in complicated
ways, so a given bubble may end in one, two, or any other number of terminal bubbles. However, we are using
the box approximation to obtain results in the limit of very large total number of bubbles, so we disregard the
imprecision in the description of multiverses with a very small total number of bubbles. Also, the value of
may be considered a variable parameter of the box model; the nal results will not be overly sensitive to the
value of .
106
7.4 A general landscape
z
1
and g itself. Therefore, the small-z behavior of the generating function g(z, q
i
; j) must be
given by Eq. (7.104). This condition, together with analyticity in z, selects the unique physically
relevant solution of Eqs. (7.103).
7.4.5 Eigenvalues of

M(z)
According to the denition (7.110), the matrix

M(z) has positive elements on the diagonal and
nonpositive elements o the diagonal. Such a matrix can be rewritten in the form

M(z) =

1

A(z), (7.164)
where a constant > 0 is introduced, the notation

1 stands for an identity matrix, and

A(z) is a
suitable nonnegative matrix, i.e. a matrix with all nonnegative elements. For instance, we may
choose as the largest of the diagonal elements of

M. The theory of nonnegative matrices gives
powerful results for the eigenvalues of matrices such as

M and

A (see e.g. the book [179]). For
the present case, the most important are the properties of the algebraically smallest eigenvalue
of the matrix

M.
It will be convenient to drop temporarily the argument z since all the results of matrix theory
will hold for every xed z. By the Perron-Frobenius theorem (see [179], chapter 9), under the
condition of irreducibility
7
a nonnegative matrix

A has a unique nondegenerate, real eigenvalue

0
> 0 such that all the eigenvalues of

A (which may be complex-valued) are located within
the circle [[
0
in the complex plane. This dominant eigenvalue
0
has a corresponding
(right) eigenvector

v
0
_
that can be chosen with all strictly positive components v
0
i
> 0. The
same property holds for the relevant left eigenvector

u
0

(the matrix need not be symmetric, so

the left and the right eigenvectors do not, in general, coincide). Therefore it is possible to choose
the eigenvectors

u
0

and

v
0
_
such that the normalization

u
0
[v
0
_
= 1 holds. This normalization
will be convenient for further calculations, and so we assume that such eigenvectors have been
chosen.
It follows that

v
0
_
and

u
0

are also the right and left eigenvectors of the matrix

M with the
eigenvalue

0

0
, (7.165)
while all the other eigenvalues of

M are located within the circle [ [
0
in the complex
plane. Since
0
> 0, all the other eigenvalues of

M are strictly to the right (in the complex
plane) of the real eigenvalue
0
. In other words,
0
is the eigenvalue of

M with the algebraically
smallest real part.
Restoring now the argument z of the matrix

M, we nd that

M always has a real, nondegen-
erate eigenvalue
0
(z), which is at the same time the eigenvalue with the algebraically smallest
real part among all the eigenvalues of

M(z). We know that det

M(z) > 0 for suciently small
z; hence
0
(z) > 0 for those z. Moreover, det

M(z) will remain positive as long as
0
(z) > 0,
since no other eigenvalue can become negative unless
0
(z) rst becomes negative. We will now
show that det

M(z) cannot remain positive for all real z > 0. It will then follow by continuity
of
0
(z) that there will be a value z

such that
0
(z) > 0 for all 0 < z < z

but
0
(z

) = 0.
7
The irreducibility condition means that any two recyclable vacua in the landscape can be connected by a chain
of transitions with nonzero nucleation rates. This condition has been discussed in Refs. [175, 196]. If some
subset of vacua form a disconnected island in the landscape, such that transitions to and from the island
are forbidden, one can regard the island as a separate irreducible landscape and apply the same technique
to it. Hence, we consider only irreducible landscapes in this work.
107
7 The RV measure for the landscape
We will use the property that the inverse matrix

M
1
(z) has all positive elements as long as

0
(z) > 0 (equivalently if
0
< ). The derivation of this property is simple:

M
1
= (

1

A)
1
=
1

1 +
2

A +
3

A
2
+..., (7.166)
which yields explicitly a matrix with all nonnegative elements. [The matrix-valued series in
Eq. (7.166) converges because all the eigenvalues of

A are strictly smaller than by absolute
value.] Moreover, the irreducibility condition means that some chain of transitions will connect
every pair of recyclable vacua; this is equivalent to saying that for any vacua i, j there exists
some integer s such that

A
s
has a nonzero matrix element (

A
s
)
ij
. Hence, every matrix element
of

M
1
is strictly positive as long as
0
(z) > 0.
Further, we can deduce that g(z; j) is a strictly increasing, real-valued function of z for those
z for which
0
(z) > 0. To show this, we consider the vector [
z
g) whose components are the N
r
derivatives
z
g(z; j), j = 1, ..., N
r
. It follows from Eq. (7.103) that the vector [
z
g) satises the
inhomogeneous equation

M(z) [
z
g) = [) , (7.167)
where we denoted by [) the vector with the components

k
(z)

i=k

ki
g(z; i). (7.168)
The solution of Eq. (7.167) is
[
z
g) =

M
1
(z) [(z)) . (7.169)
Since all the matrix elements of

M
1
(z) are positive and all the components of [) are non-
negative as long as g(z; i) > 0, it follows that all the components of [
z
g) are strictly positive.
Equation (7.104) shows that g(z; i) > 0 for suciently small z > 0, and it follows that g(z; i)
will remain positive for all z > 0 such that
0
(z) > 0. Therefore, g(z; i), i = 1, ..., N
r
are strictly
increasing functions of z for all those z.
Nevertheless, the functions g(z; i) are bounded from above. To see this, consider the relation-
ship

u
0

M =
0

u
0

, (7.170)
written in components as
u
0
j
(z)
_
1

g
1

1
(z; j)
jj
_

i=j
u
0
i
(z)z
ij
=
0
u
0
j
. (7.171)
Since all the components u
0
i
are strictly positive (as long as
0
(z) > 0), it follows that
1

g
1

1
(z; j)
jj
> 0 (7.172)
and hence
g(z; j) < (
jj
)

1
<
1

. (7.173)
Furthermore, we can show that
0
(z) monotonically decreases as z grows. This follows from
the perturbation theory formula for nondegenerate eigenvalues, which allows us to express d
0
/dz
as a matrix product with normalized left and right eigenvectors,
d
0
(z)
dz
=

u
0
(z)

d

M
dz

v
0
(z)
_
. (7.174)
108
7.4 A general landscape
As we have just shown, g(z; i) grows with growing z, so d

M/dz is a matrix with all nonpositive
elements. Since the vectors

u
0

and

v
0
_
have strictly positive components while at least some
matrix elements of the nonpositive matrix d

M/dz are strictly negative, we obtain the strict
inequality
d
0
(z)
dz
=

u
0
(z)

d

M
dz

v
0
(z)
_
< 0. (7.175)
Similarly, we can show that det

M(z) monotonically decreases with z:
d
dz
det

M(z) =
_
det

M(z)
_
Tr
_

M
1
d

M
dz
_
< 0 (7.176)
since it was already found that the matrix

M
1
(z) has all positive elements while d

M/dz has
all nonpositive elements. However, the monotonic decrease alone of
0
(z) and of det

M(z) is not
yet sucient to establish that the matrix

M(z) actually becomes singular at some nite z.
The results derived so far the monotonic behavior of g(z; j) and
0
(z), the positivity of the
matrix elements of

M
1
, the bounds on g hold for all z for which
0
(z) > 0. Now we will
show that
0
(z) cannot remain positive for all real z > 0. We can rewrite Eq. (7.109) as
z

i=k

ki
g(z; i) = g
1

(z; k)
kk
g(z; k). (7.177)
Using the property g(z; j) > 0 and the bound (7.173), we obtain (for every k) an upper bound
on z,
z =
g
1

(z; k)
kk
g(z; k)

iT,i=k

ki
g(z; i) +
kT
<
g
1

(z; k)

kT
<

1

kT
. (7.178)
In other words, no real-valued solutions of Eq. (7.109) exist for larger z. Let us then show
that an upper bound on z contradicts the assumption that
0
(z) > 0 for all z. We know that
there exists a real-valued solution branch g(z; j) near z = 0 such that 0 < g(z; j) < and
0 <
z
g(z; j) < for all those z > 0 for which this solution branch remains real-valued. Hence,
g(z; j) can be viewed as a solution of a dierential equation
z
g(z; j) = [
z
g) with continuous
coecients and everywhere positive right-hand side. The solution of such dierential equations,
if bounded, will exist for all z > 0. Indeed, if the solution g(z; j) existed only up to some z = z
1
,
we would have, by assumption,
0
(z
1
) > 0 and hence a nite value g(z
1
; j) > 0 and a nite
derivative
z
g(z
1
; j) > 0. So the solution g(z; j) could then be continued further to some z > z
1
.
Therefore, the real-valued solution branch g(z; j) must exist for all z > 0. This is incompatible
with the bound (7.178).
We conclude that there exists a value z

> 0 such that

0
(z

) = 0 but
0
(z) > 0 for all
0 < z < z

. Within the range 0 < z < z

the functions g(z; j) grow monotonically but remain

bounded by Eq. (7.173), while
0
(z) and det

M(z) both decrease monotonically to zero.
7.4.6 The root of
0
(z)
It remains to establish that
0
(z) indeed has the form (7.114) near z = z

. We again restrict
our attention to the interval 0 < z < z

where
0
(z) > 0. For these z we expand
0
(z) in Taylor
series and express the value
0
(z

) 0 as
0 =
0
(z

) =
0
(z) +
d
0
(z)
dz
(z

z) +O[(z

z)
2
], (7.179)
109
7 The RV measure for the landscape
hence
d
0
(z)
dz
=

0
(z)
z

z
+O(z

z). (7.180)
We then use Eq. (7.174) to express d
0
/dz in another way,
d
0
(z)
dz
=

u
0
(z)

d
dz

M(z)

v
0
(z)
_
=

u
0
(z)

M
g(z; i)

z
g(z; i) +
z

M(z)
_

v
0
(z)
_
, (7.181)
where in the second line we interpreted

M(z) as a function of N
r
variables g(z; i), i = 1, ..., N
r
,
and explicitly of z, in order to express d/dz through /g and /z. Using Eq. (7.169), we then
nd

0
(z)
z z

u
0
(z)

_
_

i,k

M
g(z; i)

M
1
ik

k
+


M(z)
z
_
_

v
0
(z)
_
. (7.182)
Since we are only interested in the qualitative behavior of
0
(z) at z = z

, we do not need to
keep track of the complicated coecients in Eq. (7.182). Near z = z

we have, by Eq. (7.119),

M
1
(z)
1

0
(z)

v
0
_
u
0

+O(1), (7.183)
so the dominant singular terms in Eq. (7.182) near z = z

are

0
(z)
z

z

C
1

0
(z)
+O(1). (7.184)
Therefore we obtain

0
(z) = c
1

z +O(z

z). (7.185)
The positivity of
0
(z) for z < z

entails c
1
> 0. This concludes the derivation of Eq. (7.114).
110
8 Conclusion
In this dissertation I have explored the problem of extracting observational consequences from
inationary models with spatial variation of the physical constants in the universe. The approach
was based on the version of the anthropic principle proposed in Ref. [3]. The main issue in this
line of work is to construct a well-motivated prescription for a measure on a multiverse.
Models of ination are generically future-eternal in that ination never ends globally (Chap-
ter 2). Much of the work is performed in the context of ination driven by a hypothetical scalar
eld (the inaton), as is currently observationally favored. A second much researched cosmo-
logical scenario is the so called string-theoretic landscape where the vacuum state of elds is
metastable. Transitions between vacua can proceed via bubble nucleation.
The stochastic description of ination based on the Fokker-Planck (FP) equation for the
spatial probability distribution of the inaton eld is reviewed in Chapter 3. I explain the main
properties of the stochastic approach and also give an introduction to the stochastic description
of the string-theoretic landscape using the master equation. These mathematical tools allow one
to obtain a concrete stochastic process representing eternal ination and to study its properties
quantitatively.
I give a broad overview of the currently present measure proposals in Chapter 4. In all the
measure proposals, the innite volume of the multiverse is cut o in a controlled manner. How-
ever, the measures dier in the specic implementation of the volume cuto and in the way the
volume approaches innity when the cuto is removed. The existing measure prescriptions can
be categorized into volume-based and worldline-based. I characterize these prescriptions from
a physical standpoint and explain their main features. The volume-based measures compute
the observation of a randomly situated observer in a global innite multiverse, and are suitable
for predicting cosmological conditions, i.e. external conditions found on rst observation. The
worldline-based measures are suitable for sending a message to the future containing infor-
mation about the expected cosmological conditions in the future given the knowledge of the
present conditions. Since the message can be opened only by observers near the message box,
the worldline selects the correct statistical ensemble of observers. In this way I show that the
two classes of measures look for answers to very dierent physical questions.
I proceed to present a new multiverse measure proposal the RV measure (Chapter 5). The
RV measure is based on the construction of an ensemble of nite multiverses, which can be visu-
alized as a cuto in probability space, rather than a geometric cuto in the real spacetime. This
is an attractive feature of the RV proposal since no arbitrarily chosen geometric constructions
need to be used.
In chapters 6 and 7 I develop a mathematical framework for the technically dicult analysis
of the probability distributions of observables in the RV cuto. There are two main contexts in
which the RV cuto is applied: the slow-roll ination driven by a scalar eld, and the landscape
of string theory. In both contexts I show by rigorous methods that the RV cuto is well-dened
and independent of the choice of initial conditions at the beginning of ination. For slow-roll
ination I derive formulas that are suitable for direct numerical calculations. For the landscape, I
111
8 Conclusion
derive an analytic expression for relative bubble abundances of low-energy vacua. I also perform
explicit analytic calculations for a toy model of slow-roll ination and for several toy models of
the landscape.
These calculations conrm that the RV measure has the desirable properties expected of
a volume-based measure: coordinate invariance, independence of initial conditions, and the
absence of the youngness paradox. A calculation for a general landscape also shows that the
RV measure does not suer from the Boltzmann brain problem. Thus the RV measure is a
promising solution to the long-standing problem of obtaining probabilities in models of eternal
ination. Ultimately, the viability of the RV measure proposal will depend on its performance in
various example cases. In the calculations performed so far, it is found that RV measure yields
results that do not always coincide with the results of any other measure proposal. Hence, the
RV measure is not equivalent to any of the earlier proposals.
An extension of the RV measure to landscape scenarios can be achieved in several ways. In
Sec. 7.1 I considered the set of all possible future evolutions of a single nonterminal bubble and
dened the ensemble E
N
of evolutions yielding a nite total number N of terminal bubbles. In
Sec. 7.2 I used the ensemble of evolutions with a nite total number N of daughter bubbles (of
all types). One could also consider the ensemble E

N
of evolutions yielding a nite total number
N of observers in bubbles of all types. After computing the distribution of some desired quantity
by counting the observations made within the nite set of N bubbles (or observers), the cuto
parameter N can be increased to innity. It remains to be seen whether the limit distributions
are dierent for dierently dened ensembles, such as E
N
and E

N
, and if so, which denition is
more suitable. Future work will show whether some extension of the RV measure can provide a
satisfactory answer to the problem of predictions in eternal ination.
To conclude, the present work demonstrates that the RV measure has attractive features and
may be considered a viable candidate for the solution of the measure problem in multiverse
cosmology. More work is needed to investigate the dependence of the predictions on the precise
details of the denition of the ensemble E
N
. I have developed an extensive mathematical frame-
work for the calculations in the RV prescription and obtained rst results for specic landscapes.
However, a more powerful approximation scheme is desirable so that the predictions of the RV
measure can be more easily obtained for landscapes of general type. Ultimately, the viability
of the RV measure is to be judged by its predictions for cosmological observables in realistic
landscapes. These issues will be considered in future publications.
112
Bibliography
[1] J. Garcia-Bellido and A. D. Linde, Stationarity of ination and predictions of quantum
cosmology, Phys. Rev. D51, 429443 (1995), hep-th/9408023.
[2] A. Vilenkin, Predictions from quantum cosmology, Phys. Rev. Lett. 74, 846849 (1995),
gr-qc/9406010.
[3] A. Vilenkin, Making predictions in eternally inating universe, Phys. Rev. D52, 33653374
(1995), gr-qc/9505031.
[4] K. A. Olive, Ination, Phys. Rept. 190, 307403 (1990).
[5] A. D. Linde, Particle physics and inationary cosmology, (1990), hep-th/0503203.
[6] A. H. Guth, Ination and eternal ination, Phys. Rept. 333, 555574 (2000), astro-
ph/0002156.
[7] A. Aguirre, S. Gratton, and M. C. Johnson, Hurdles for recent measures in eternal ination,
(2006), hep-th/0611221.
[8] S. Winitzki, Predictions in eternal ination, Lect. Notes Phys. 738, 157191 (2008),
gr-qc/0612164.
[9] A. Vilenkin, A measure of the multiverse, J. Phys. A40, 6777 (2007), hep-th/0609193.
[10] A. H. Guth, Eternal ination and its implications, J. Phys. A40, 68116826 (2007),
hep-th/0702178.
[11] A. Linde, Towards a gauge invariant volume-weighted probability measure for eternal
ination, JCAP 0706, 017 (2007), arXiv:0705.1160 [hep-th].
[12] G. F. Smoot et al., Structure in the COBE dierential microwave radiometer rst year
maps, Astrophys. J. 396, L1L5 (1992).
[13] R. H. Dicke and P. J. E. Peebles, The big band cosmology - enigmas and nostrums, (1979),
In *General relativity: an Einstein centenary survey*, ed. by S. W. Hawking and W. Israel.
[14] P. J. E. Peebles and J. T. Yu, Primeval adiabatic perturbation in an expanding universe,
Astrophys. J. 162, 815836 (1970).
[15] J. M. Bardeen, Gauge invariant cosmological perturbations, Phys. Rev. D22, 18821905
(1980).
[16] W. H. Press and E. T. Vishniac, Tenacious myths about cosmological perturbations larger
than the horizon size, Astrophys. J. 239, 111 (1980).
113
Bibliography
[17] E. R. Harrison, Fluctuations at the threshold of classical cosmology, Phys. Rev. D1,
27262730 (1970).
[18] Y. B. Zeldovich, Gravitational instability: An approximate theory for large density per-
turbations, Astron. Astrophys. 5, 8489 (1970).
[19] A. H. Guth, The inationary universe: a possible solution to the horizon and atness
problems, Phys. Rev. D23, 347356 (1981).
[20] A. A. Starobinsky, Spectrum of relict gravitational radiation and the early state of the
universe, JETP Lett. 30, 682685 (1979).
[21] A. A. Starobinsky, A new type of isotropic cosmological models without singularity, Phys.
Lett. B91, 99102 (1980).
[22] V. F. Mukhanov and G. V. Chibisov, Quantum uctuation and nonsingular universe. (in
Russian), JETP Lett. 33, 532535 (1981).
[23] V. F. Mukhanov and G. V. Chibisov, The vacuum energy and large scale structure of the
universe, Sov. Phys. JETP 56, 258265 (1982).
[24] S. W. Hawking, The development of irregularities in a single bubble inationary universe,
Phys. Lett. B115, 295 (1982).
[25] A. A. Starobinsky, Dynamics of phase transition in the new inationary universe scenario
and generation of perturbations, Phys. Lett. B117, 175178 (1982).
[26] A. H. Guth and S. Y. Pi, Fluctuations in the new inationary universe, Phys. Rev. Lett.
49, 11101113 (1982).
[27] J. M. Bardeen, P. J. Steinhardt, and M. S. Turner, Spontaneous creation of almost scale-
free density perturbations in an inationary universe, Phys. Rev. D28, 679 (1983).
[28] A. D. Linde, A new inationary universe scenario: a possible solution of the horizon,
atness, homogeneity, isotropy and primordial monopole problems, Phys. Lett. B108,
389393 (1982).
[29] G. F. R. Ellis, (1991), Proceedings of the Ban Summer Institute, Ban, Alberta, 1990,
ed. by R. Mann and P. Wesson.
[30] L. Kofman, A. D. Linde, and A. A. Starobinsky, Reheating after ination, Phys. Rev.
Lett. 73, 31953198 (1994), hep-th/9405187.
[31] L. Kofman, A. D. Linde, and A. A. Starobinsky, Towards the theory of reheating after
ination, Phys. Rev. D56, 32583295 (1997), hep-ph/9704452.
[32] A. Vilenkin, The birth of inationary universes, Phys. Rev. D27, 2848 (1983).
[33] A. A. Starobinsky, Stochastic de Sitter (inationary) stage in the early universe, (1986), in:
Current Topics in Field Theory, Quantum Gravity and Strings, Lecture Notes in Physics
206, eds. H.J. de Vega and N. Sanchez (Springer Verlag), p. 107.
114
Bibliography
[34] A. D. Linde, Eternally existing self-reproducing chaotic inationary universe, Phys. Lett.
B175, 395400 (1986).
[35] A. S. Goncharov, A. D. Linde, and V. F. Mukhanov, The global structure of the ination-
ary universe, Int. J. Mod. Phys. A2, 561591 (1987).
[36] A. Vilenkin and L. H. Ford, Gravitational eects upon cosmological phase transitions,
Phys. Rev. D26, 1231 (1982).
[37] A. D. Linde, Scalar eld uctuations in expanding universe and the new inationary
universe scenario, Phys. Lett. B116, 335 (1982).
[38] A. Vilenkin, Quantum uctuations in the new inationary universe, Nucl. Phys. B226,
527 (1983).
[39] R. H. Brandenberger, Quantum uctuations as the source of classical gravitational per-
turbations in inationary universe, Nucl. Phys. B245, 328 (1984).
[40] A. H. Guth and S.-Y. Pi, The quantum mechanics of the scalar eld in the new inationary
universe, Phys. Rev. D32, 18991920 (1985).
[41] G. W. Gibbons and S. W. Hawking, Cosmological event horizons, thermodynamics, and
particle creation, Phys. Rev. D15, 27382751 (1977).
[42] S. W. Hawking and I. G. Moss, Supercooled phase transitions in the very early universe,
Phys. Lett. B110, 35 (1982).
[43] S. W. Hawking and I. G. Moss, Fluctuations in the inationary universe, Nucl. Phys.
B224, 180 (1983).
[44] A. Vilenkin, Eternal ination and chaotic terminology, (2004), gr-qc/0409055.
[45] A. Borde and A. Vilenkin, Eternal ination and the initial singularity, Phys. Rev. Lett.
72, 33053309 (1994), gr-qc/9312022.
[46] A. Borde, A. H. Guth, and A. Vilenkin, Inationary space-times are incompletein past
directions, Phys. Rev. Lett. 90, 151301 (2003), gr-qc/0110012.
[47] D. A. Lowe and D. Marolf, Holography and eternal ination, Phys. Rev. D70, 026001
(2004), hep-th/0402162.
[48] C.-H. Wu, K.-W. Ng, and L. H. Ford, Constraints on the duration of inationary expansion
from quantum stress tensor uctuations, (2006), gr-qc/0608002.
[49] A. D. Linde, Axions in inationary cosmology, Phys. Lett. B259, 3847 (1991).
[50] A. D. Linde, Hybrid ination, Phys. Rev. D49, 748754 (1994), astro-ph/9307002.
[51] J. Garcia-Bellido, A. D. Linde, and D. A. Linde, Fluctuations of the gravitational constant
in the inationary Brans-Dicke cosmology, Phys. Rev. D50, 730750 (1994), astro-
ph/9312039.
115
Bibliography
[52] J. Garcia-Bellido, Jordan-Brans-Dicke stochastic ination, Nucl. Phys. B423, 221242
(1994), astro-ph/9401042.
[53] J. Garcia-Bellido and D. Wands, General relativity as an attractor in scalar - tensor
stochastic ination, Phys. Rev. D52, 56365642 (1995), gr-qc/9503049.
[54] A. D. Linde, Monopoles as big as a universe, Phys. Lett. B327, 208213 (1994), astro-
ph/9402031.
[55] A. Vilenkin, Topological ination, Phys. Rev. Lett. 72, 31373140 (1994), hep-th/9402085.
[56] K. E. Kunze, Stochastic ination on the brane, Phys. Lett. B587, 16 (2004), hep-
th/0310200.
[57] J. Garriga and A. Vilenkin, Recycling universe, Phys. Rev. D57, 22302244 (1998),
astro-ph/9707292.
[58] L. Susskind, The anthropic landscape of string theory, (2003), hep-th/0302219.
[59] S. R. Coleman and F. De Luccia, Gravitational eects on and of vacuum decay, phys. rev.
D21, 3305 (1980).
[60] A. H. Guth, Eternal Ination, Ann. N.Y. Acad. Sci. 950, 6682 (2001), astro-ph/0101507.
[61] A. H. Guth and E. J. Weinberg, Could the universe have recovered from a slow rst order
phase transition?, Nucl. Phys. B212, 321 (1983).
[62] J. Garriga, A. H. Guth, and A. Vilenkin, Eternal ination, bubble collisions, and the
persistence of memory, (2006), hep-th/0612242.
[63] J. Garriga, V. F. Mukhanov, K. D. Olum, and A. Vilenkin, Eternal ination, black
holes, and the future of civilizations, Int. J. Theor. Phys. 39, 18871900 (2000), astro-
ph/9909143.
[64] J. Garriga and A. Vilenkin, Many worlds in one, Phys. Rev. D64, 043511 (2001), gr-
qc/0102010.
[65] D. S. Goldwirth and T. Piran, Initial conditions for ination, Phys. Rept. 214, 223291
(1992).
[66] A. D. Linde, Chaotic ination, Phys. Lett. B129, 177181 (1983).
[67] A. D. Linde, Eternal chaotic ination, Mod. Phys. Lett. A1, 81 (1986).
[68] J. Martin and M. A. Musso, Stochastic quintessence, Phys. Rev. D71, 063514 (2005),
astro-ph/0410190.
[69] F. Helmer and S. Winitzki, Self-reproduction in k-ination, Phys. Rev. D74, 063528
(2006), gr-qc/0608019.
[70] S. M. Carroll, The cosmological constant, Living Rev. Rel. 4, 1 (2001), astro-ph/0004075.
116
Bibliography
[71] A. Vilenkin, Unambiguous probabilities in an eternally inating universe, Phys. Rev. Lett.
81, 55015504 (1998), hep-th/9806185.
[72] J. Garriga, T. Tanaka, and A. Vilenkin, The density parameter and the Anthropic Prin-
ciple, Phys. Rev. D60, 023501 (1999), astro-ph/9803268.
[73] J. Garriga and A. Vilenkin, On likely values of the cosmological constant, Phys. Rev.
D61, 083502 (2000), astro-ph/9908115.
[74] J. Garriga, M. Livio, and A. Vilenkin, The cosmological constant and the time of its
dominance, Phys. Rev. D61, 023503 (2000), astro-ph/9906210.
[75] J. Garriga and A. Vilenkin, Testable anthropic predictions for dark energy, Phys. Rev.
D67, 043503 (2003), astro-ph/0210358.
[76] J. Garriga, A. Linde, and A. Vilenkin, Dark energy equation of state and anthropic
selection, Phys. Rev. D69, 063521 (2004), hep-th/0310034.
[77] J. Garriga and A. Vilenkin, Anthropic prediction for Lambda and the Q catastrophe,
Prog. Theor. Phys. Suppl. 163, 245257 (2006), hep-th/0508005.
[78] M. Tegmark, A. Vilenkin, and L. Pogosian, Anthropic predictions for neutrino masses,
Phys. Rev. D71, 103523 (2005), astro-ph/0304536.
[79] M. Tegmark, A. Aguirre, M. Rees, and F. Wilczek, Dimensionless constants, cosmology
and other dark matters, Phys. Rev. D73, 023505 (2006), astro-ph/0511774.
[80] L. J. Hall, T. Watari, and T. T. Yanagida, Taming the runaway problem of inationary
landscapes, D73, 103502 (2006), hep-th/0601028.
[81] M. Susperregi, Spectrum of density uctuations in Brans-Dicke chaotic ination, Phys.
Rev. D55, 560572 (1997), astro-ph/9606018.
[82] B. Feldstein, L. J. Hall, and T. Watari, Density perturbations and the cosmological
constant from inationary landscapes, Phys. Rev. D72, 123506 (2005), hep-th/0506235.
[83] W. Lerche, D. Lust, and A. N. Schellekens, Chiral four-dimensional heterotic strings from
selfdual lattices, Nucl. Phys. B287, 477 (1987).
[84] R. Bousso and J. Polchinski, Quantization of four-form uxes and dynamical neutralization
of the cosmological constant, JHEP 06, 006 (2000), hep-th/0004134.
[85] S. Kachru, R. Kallosh, A. Linde, and S. P. Trivedi, De Sitter vacua in string theory, Phys.
Rev. D68, 046005 (2003), hep-th/0301240.
[86] F. Denef and M. R. Douglas, Distributions of ux vacua, JHEP 05, 072 (2004), hep-
th/0404116.
[87] D. Schwartz-Perlov and A. Vilenkin, Probabilities in the Bousso-Polchinski multiverse,
JCAP 0606, 010 (2006), hep-th/0601162.
[88] A. Linde, Sinks in the landscape, Boltzmann brains, and the cosmological constant orob-
lem, JCAP 0701, 022 (2007), hep-th/0611043.
117
Bibliography
[89] R. Bousso, R. Harnik, G. D. Kribs, and G. Perez, Predicting the Cosmological Constant
from the Causal Entropic Principle, Phys. Rev. D76, 043513 (2007), hep-th/0702115.
[90] H. Kodama and M. Sasaki, Cosmological perturbation theory, Prog. Theor. Phys. Suppl.
78, 1166 (1984).
[91] V. F. Mukhanov, H. A. Feldman, and R. H. Brandenberger, Theory of cosmological
perturbations. Part 1. Classical perturbations. Part 2. Quantum theory of perturbations.
Part 3. Extensions, Phys. Rept. 215, 203333 (1992).
[92] R. H. Brandenberger, Quantum eld theory methods and inationary universe models,
Rev. Mod. Phys. 57, 1 (1985).
[93] D. Polarski and A. A. Starobinsky, Semiclassicality and decoherence of cosmological per-
turbations, Class. Quant. Grav. 13, 377392 (1996), gr-qc/9504030.
[94] C. Kiefer and D. Polarski, Emergence of classicality for primordial uctuations: Concepts
and analogies, Annalen Phys. 7, 137158 (1998), gr-qc/9805014.
[95] C. Kiefer, J. Lesgourgues, D. Polarski, and A. A. Starobinsky, The coherence of primordial
uctuations produced during ination, Class. Quant. Grav. 15, L67L72 (1998), gr-
qc/9806066.
[96] C. Kiefer, D. Polarski, and A. A. Starobinsky, Quantum-to-classical transition for uctu-
ations in the early universe, Int. J. Mod. Phys. D7, 455462 (1998), gr-qc/9802003.
[97] C. Kiefer, D. Polarski, and A. A. Starobinsky, Entropy of gravitons produced in the early
universe, Phys. Rev. D62, 043518 (2000), gr-qc/9910065.
[98] C. Kiefer, I. Lohmar, D. Polarski, and A. A. Starobinsky, Pointer states for primordial
uctuations in inationary cosmology, (2006), astro-ph/0610700.
[99] L. R. W. Abramo, R. H. Brandenberger, and V. F. Mukhanov, The energy-momentum
tensor for cosmological perturbations, Phys. Rev. D56, 32483257 (1997), gr-qc/9704037.
[100] N. Afshordi and R. H. Brandenberger, Super-Hubble nonlinear perturbations during in-
ation, Phys. Rev. D63, 123505 (2001), gr-qc/0011075.
[101] L. R. Abramo and R. P. Woodard, A Scalar Measure Of The Local Expansion Rate, Phys.
Rev. D65, 043507 (2002), astro-ph/0109271.
[102] L. R. Abramo and R. P. Woodard, No one loop back-reaction in chaotic ination, Phys.
Rev. D65, 063515 (2002), astro-ph/0109272.
[103] L. R. Abramo and R. P. Woodard, Back-reaction is for real, Phys. Rev. D65, 063516
(2002), astro-ph/0109273.
[104] G. Geshnizjani and R. Brandenberger, Back reaction of perturbations in two scalar eld
inationary models, JCAP 0504, 006 (2005), hep-th/0310265.
[105] G. Geshnizjani and N. Afshordi, Coarse-grained back reaction in single scalar eld driven
ination, JCAP 0501, 011 (2005), gr-qc/0405117.
118
Bibliography
[106] A. V. Frolov and L. Kofman, Ination and de Sitter thermodynamics, JCAP 0305, 009
(2003), hep-th/0212327.
[107] A. Borde and A. Vilenkin, Violation of the weak energy condition in inating spacetimes,
Phys. Rev. D56, 717723 (1997), gr-qc/9702019.
[108] S. Winitzki, Null energy condition violations in eternal ination, (2001), gr-qc/0111109.
[109] T. Vachaspati, Eternal ination and energy conditions in de Sitter spacetime, (2003),
astro-ph/0305439.
[110] S. Dutta and T. Vachaspati, Islands in the Lambda-sea, Phys. Rev. D71, 083507 (2005),
astro-ph/0501396.
[111] Y.-S. Piao, Is the island universe model consistent with observations?, Phys. Rev. D72,
103513 (2005), astro-ph/0506072.
[112] S. Dutta, A classical treatment of island cosmology, Phys. Rev. D73, 063524 (2006),
astro-ph/0511120.
[113] M. Morikawa, The origin of the density uctuations in de Sitter space, Prog. Theor. Phys.
77, 11631177 (1987).
[114] M. Morikawa, Dissipation and uctuation of quantum elds in expanding universes, Phys.
Rev. D42, 10271034 (1990).
[115] B. L. Hu and A. Matacz, Back reaction in semiclassical cosmology: The Einstein- Langevin
equation, Phys. Rev. D51, 15771586 (1995), gr-qc/9403043.
[116] E. Calzetta and B. L. Hu, Quantum uctuations, decoherence of the mean eld, and struc-
ture formation in the early universe, Phys. Rev. D52, 67706788 (1995), gr-qc/9505046.
[117] A. Matacz, A New Theory of Stochastic Ination, Phys. Rev. D55, 18601874 (1997),
gr-qc/9604022.
[118] A. Matacz, Ination and the ne-tuning problem, Phys. Rev. D56, 18361840 (1997),
gr-qc/9611063.
[119] H. Kubotani, T. Uesugi, M. Morikawa, and A. Sugamoto, Classicalization of quantum uc-
tuation in inationary universe, Prog. Theor. Phys. 98, 10631080 (1997), gr-qc/9701043.
[120] R. P. Woodard, A leading logarithm approximation for inationary quantum eld theory,
Nucl. Phys. Proc. Suppl. 148, 108119 (2005), astro-ph/0502556.
[121] N. C. Tsamis and R. P. Woodard, Stochastic quantum gravitational ination, Nucl. Phys.
B724, 295328 (2005), gr-qc/0505115.
[122] R. P. Woodard, Generalizing Starobinskiis formalism to Yukawa theory and to scalar
QED, (2006), gr-qc/0608037.
[123] S.-P. Miao and R. P. Woodard, Leading log solution for inationary Yukawa, Phys. Rev.
D74, 044019 (2006), gr-qc/0602110.
119
Bibliography
[124] T. Prokopec, N. C. Tsamis, and R. P. Woodard, Two loop scalar bilinears for inationary
SQED, Class. Quant. Grav. 24, 201230 (2007), gr-qc/0607094.
[125] S.-J. Rey, Dynamics of inationary phase transition, Nucl. Phys. B284, 706 (1987).
[126] K.-i. Nakao, Y. Nambu, and M. Sasaki, STOCHASTIC DYNAMICS OF NEW INFLA-
TION, Prog. Theor. Phys. 80, 1041 (1988).
[127] H. E. Kandrup, Stochastic ination as a time dependent random walk, Phys. Rev. D39,
2245 (1989).
[128] Y. Nambu and M. Sasaki, Stochastic approach to chaotic ination and the distribution of
universes, Phys. Lett. B219, 240 (1989).
[129] Y. Nambu, Stochastic dynamics of an inationary model and initial distribution of uni-
verses, Prog. Theor. Phys. 81, 1037 (1989).
[130] M. Mijic, Random walk after the Big Bang, Phys. Rev. D42, 24692482 (1990).
[131] D. S. Salopek and J. R. Bond, Stochastic ination and nonlinear gravity, Phys. Rev. D43,
10051031 (1991).
[132] A. D. Linde and A. Mezhlumian, Stationary universe, Phys. Lett. B307, 2533 (1993),
gr-qc/9304015.
[133] A. D. Linde, D. A. Linde, and A. Mezhlumian, From the Big Bang theory to the theory
of a stationary universe, Phys. Rev. D49, 17831826 (1994), gr-qc/9306035.
[134] D. S. Salopek and J. R. Bond, Nonlinear evolution of long wavelength metric uctuations
in inationary models, Phys. Rev. D42, 39363962 (1990).
[135] A. R. Liddle, P. Parsons, and J. D. Barrow, Formalizing the slow roll approximation in
ination, Phys. Rev. D50, 72227232 (1994), astro-ph/9408015.
[136] M. Bellini, H. Casini, R. Montemayor, and P. Sisterna, Stochastic approach to ination:
Classicality conditions, Phys. Rev. D54, 71727180 (1996).
[137] H. Casini, R. Montemayor, and P. Sisterna, Stochastic approach to ination. II: Classi-
cality, coarse-graining and noises, Phys. Rev. D59, 063512 (1999), gr-qc/9811083.
[138] S. Winitzki and A. Vilenkin, Eective noise in stochastic description of ination, Phys.
Rev. D61, 084008 (2000), gr-qc/9911029.
[139] S. Matarrese, M. A. Musso, and A. Riotto, Inuence of super-horizon scales on cosmolog-
ical observables generated during ination, JCAP 0405, 008 (2004), hep-th/0311059.
[140] M. Liguori, S. Matarrese, M. Musso, and A. Riotto, Stochastic Ination and the Lower
Multipoles in the CMB Anisotropies, JCAP 0408, 011 (2004), astro-ph/0405544.
[141] N. G. van Kampen, Stochastic processes in physics and chemistry, North-Holland, Ams-
terdam, 1981.
[142] H. Risken, The Fokker-Planck equation, Springer-Verlag, 1989.
120
Bibliography
[143] J. Garriga and V. F. Mukhanov, Perturbations in k-ination, Phys. Lett. B458, 219225
(1999), hep-th/9904176.
[144] A. Vilenkin, On the factor ordering problem in stochastic ination, Phys. Rev. D59,
123506 (1999), gr-qc/9902007.
[145] S. Winitzki and A. Vilenkin, Uncertainties of predictions in models of eternal ination,
Phys. Rev. D53, 42984310 (1996), gr-qc/9510054.
[146] S. Winitzki, On time-reparametrization invariance in eternal ination, Phys. Rev. D71,
123507 (2005), gr-qc/0504084.
[147] S. Matarrese, A. Ortolan, and F. Lucchin, Ination in the scaling limit, Phys. Rev. D40,
290 (1989).
[148] A. D. Linde, Coleman-Weinberg theory and a new inationary universe scenario, Phys.
Lett. B114, 431 (1982).
[149] A. D. Linde, Temperature dependence of coupling constants and the phase transition in
the Coleman-Weinberg theory, Phys. Lett. B116, 340 (1982).
[150] A. Albrecht and P. J. Steinhardt, Cosmology for grand unied theories with radiatively
induced symmetry breaking, Phys. Rev. Lett. 48, 12201223 (1982).
[151] H. M. Hodges, Analytic solution of a chaotic inaton, Phys. Rev. D39, 35683570 (1989).
[152] I. Yi, E. T. Vishniac, and S. Mineshige, Generation of nonGaussian uctuations during
chaotic ination, Phys. Rev. D43, 362368 (1991).
[153] Y. V. Shtanov, Functional approach to stochastic ination, Phys. Rev. D52, 42874294
(1995), gr-qc/9407034.
[154] J. Martin and M. Musso, Solving stochastic ination for arbitrary potentials, Phys. Rev.
D73, 043516 (2006), hep-th/0511214.
[155] S. Gratton and N. Turok, Langevin analysis of eternal ination, Phys. Rev. D72, 043507
(2005), hep-th/0503063.
[156] J. Martin and M. Musso, On the reliability of the Langevin pertubative solution in stochas-
tic ination, Phys. Rev. D73, 043517 (2006), hep-th/0511292.
[157] M. Tegmark, What does ination really predict?, JCAP 0504, 001 (2005), astro-
ph/0410281.
[158] A. D. Linde, D. A. Linde, and A. Mezhlumian, Do we live in the center of the world?,
Phys. Lett. B345, 203210 (1995), hep-th/9411111.
[159] S. Winitzki, The eternal fractal in the universe, Phys. Rev. D65, 083506 (2002), gr-
qc/0111048.
[160] M. Aryal and A. Vilenkin, The fractal dimension of inationary universe, Phys. Lett.
B199, 351 (1987).
121
Bibliography
[161] J. R. Gott, Creation of open universes from de sitter space, Nature 295, 304307 (1982).
[162] J. R. Gott and T. S. Statler, Constraints on the formation of bubble universes, Phys.
Lett. B136, 157161 (1984).
[163] M. Bucher, A. S. Goldhaber, and N. Turok, An open universe from ination, Phys. Rev.
D52, 33143337 (1995), hep-ph/9411206.
[164] M. Bucher, A. S. Goldhaber, and N. Turok, Omega(0) < 1 from ination, Nucl. Phys.
Proc. Suppl. 43, 173176 (1995), hep-ph/9501396.
[165] K. Yamamoto, M. Sasaki, and T. Tanaka, Large angle CMB anisotropy in an open uni-
versity in the one bubble inationary scenario, Astrophys. J. 455, 412418 (1995), astro-
ph/9501109.
[166] S. R. Coleman, The fate of the false vacuum. 1. Semiclassical theory, Phys. Rev. D15,
29292936 (1977).
[167] K.-M. Lee and E. J. Weinberg, Decay of the true vacuum in curved space-time, Phys.
Rev. D36, 1088 (1987).
[168] J. Garriga, Nucleation rates in at and curved space, Phys. Rev. D49, 63276342 (1994),
hep-ph/9308280.
[169] M. R. Douglas, The statistics of string / M theory vacua, JHEP 05, 046 (2003), hep-
th/0303194.
[170] R. Bousso et al., Future Foam, (2008), 0807.1947.
[171] E. Farhi and A. H. Guth, An obstacle to creating a universe in the laboratory, Phys. Lett.
B183, 149 (1987).
[172] E. Farhi, A. H. Guth, and J. Guven, Is it possible to create a universe in the laboratory
by quantum tunneling?, Nucl. Phys. B339, 417490 (1990).
[173] R. Bousso, Cosmology and the S-matrix, Phys. Rev. D71, 064024 (2005), hep-th/0412197.
[174] L. F. Abbott and S. R. Coleman, The collapse of an anti-de Sitter bubble, Nucl. Phys.
B259, 170 (1985).
[175] J. Garriga, D. Schwartz-Perlov, A. Vilenkin, and S. Winitzki, Probabilities in the ina-
tionary multiverse, JCAP 0601, 017 (2006), hep-th/0509184.
[176] S. Winitzki, Drawing conformal diagrams for a fractal landscape, Phys. Rev. D71, 123523
(2005), gr-qc/0503061.
[177] V. Vanchurin and A. Vilenkin, Eternal observers and bubble abundances in the landscape,
Phys. Rev. D74, 043520 (2006), hep-th/0605015.
[178] R. Bousso, Holographic probabilities in eternal ination, Phys. Rev. Lett. 97, 191302
(2006), hep-th/0605263.
[179] P. Lancaster, Theory of matrices, Academic Press, New York, 1969.
122
Bibliography
[180] J. L. Doob, Stochastic processes, Wiley, New York, 1953.
[181] D. Kannan, Introduction to stochastic processes, North-Holland, New York, 1979.
[182] A. D. Linde, Chaotic inating universe, JETP Lett. 38, 176179 (1983).
[183] X. Chen, S. Sarangi, S. H. Henry Tye, and J. Xu, Is brane ination eternal?, (2006),
hep-th/0608082.
[184] G. W. Gibbons and N. Turok, The measure problem in cosmology, (2006), hep-th/0609095.
[185] V. Vanchurin, A. Vilenkin, and S. Winitzki, Predictability crisis in inationary cosmology
and its resolution, Phys. Rev. D61, 083507 (2000), gr-qc/9905097.
[186] R. Easther, E. A. Lim, and M. R. Martin, Counting pockets with world lines in eternal
ination, JCAP 0603, 016 (2006), astro-ph/0511233.
[187] J. Garriga and A. Vilenkin, A prescription for probabilities in eternal ination, Phys.
Rev. D64, 023507 (2001), gr-qc/0102090.
[188] R. Bousso, B. Freivogel, and I.-S. Yang, Eternal ination: The inside story, (2006),
hep-th/0606114.
[189] R. Bousso, Precision cosmology and the landscape, (2006), hep-th/0610211.
[190] A. D. Linde and A. Mezhlumian, On Regularization Scheme Dependence of Predictions
in Inationary Cosmology, Phys. Rev. D53, 42674274 (1996), gr-qc/9511058.
[191] A. Vilenkin and S. Winitzki, Probability distribution for Omega in open-universe ination,
Phys. Rev. D55, 548559 (1997), astro-ph/9605191.
[192] A. De Simone et al., Boltzmann brains and the scale-factor cuto measure of the multi-
verse, (2008), 0808.3778.
[193] A. De Simone, A. H. Guth, M. P. Salem, and A. Vilenkin, Predicting the cosmological
constant with the scale-factor cuto measure, (2008), 0805.2173.
[194] A. Aguirre, S. Gratton, and M. C. Johnson, Measures on transitions for cosmology from
eternal ination, Phys. Rev. Lett. 98, 131301 (2007), hep-th/0612195.
[195] V. Vanchurin, Geodesic measures of the landscape, Phys. Rev. D75, 023524 (2007),
hep-th/0612215.
[196] J. Garriga and A. Vilenkin, Holographic Multiverse, (2008), 0809.4257.
[197] A. D. Linde and D. A. Linde, Topological defects as seeds for eternal ination, Phys. Rev.
D50, 24562468 (1994), hep-th/9402115.
[198] R. Bousso, B. Freivogel, and I.-S. Yang, Boltzmann babies in the proper time measure,
(2007), arXiv:0712.3324 [hep-th].
[199] L. Dyson, M. Kleban, and L. Susskind, Disturbing implications of a cosmological constant,
JHEP 10, 011 (2002), hep-th/0208013.
123
Bibliography
[200] A. Albrecht and L. Sorbo, Can the universe aord ination?, Phys. Rev. D70, 063528
(2004), hep-th/0405270.
[201] D. N. Page, Is our universe likely to decay within 20 billion years?, (2006), hep-th/0610079.
[202] A. Vilenkin, Freak observers and the measure of the multiverse, JHEP 01, 092 (2007),
hep-th/0611271.
[203] D. N. Page, Return of the Boltzmann brains, (2006), hep-th/0611158.
[204] J. R. Gott III, Boltzmann brainsId rather see than be one, (2008), arXiv:0802.0233
[gr-qc].
[205] R. Bousso, B. Freivogel, and I.-S. Yang, Properties of the scale factor measure, (2008),
0808.3770.
[206] S. Winitzki, A volume-weighted measure for eternal ination, Phys. Rev. D78, 043501
(2008), 0803.1300.
[207] S. Winitzki, Reheating-volume measure for random-walk ination, Phys. Rev. D78,
063517 (2008), 0805.3940.
[208] S. Winitzki, Reheating-volume measure in the landscape, (2008), 0810.1517.
[209] P. Creminelli, S. Dubovsky, A. Nicolis, L. Senatore, and M. Zaldarriaga, The phase
transition to Slow-roll Eternal Ination, (2008), arXiv:0802.1067 [hep-th].
[210] A. J. Tolley and M. Wyman, Stochastic ination revisited: non-slow roll statistics and
DBI ination, JCAP 0804, 028 (2008), 0801.1854.
[211] L. Pogosian and A. Vilenkin, Anthropic predictions for vacuum energy and neutrino masses
in the light of WMAP-3, (2006), astro-ph/0611573.
[212] I. Maor, L. Krauss, and G. Starkman, Anthropics and Myopics: Conditional Probabilities
and the Cosmological Constant, Phys. Rev. Lett. 100, 041301 (2007), arXiv:0709.0502
[hep-th].
[213] J. Garriga and A. Vilenkin, Prediction and explanation in the multiverse, (2007),
arXiv:0711.2559 [hep-th].
[214] J. B. Hartle and M. Srednicki, Are We Typical?, Phys. Rev. D75, 123523 (2007),
arXiv:0704.2630 [hep-th].
[215] T. Clifton, S. Shenker, and N. Sivanandam, Volume-weighted measures of eternal ination
in the Bousso-Polchinski landscape, JHEP 09, 034 (2007), arXiv:0706.3201 [hep-th].
[216] S. W. Hawking, The measure of the universe, AIP Conf. Proc. 957, 7984 (2007).
[217] J. B. Hartle, S. W. Hawking, and T. Hertog, The no-boundary measure of the universe,
(2007), arXiv:0711.4630 [hep-th].
[218] S. W. Hawking, Volume weighting in the no boundary proposal, (2007), arXiv:0710.2029
[hep-th].
124
Bibliography
[219] J. B. Hartle, S. W. Hawking, and T. Hertog, The classical universes of the no-boundary
quantum state, (2008), arXiv:0803.1663 [hep-th].
[220] I. Stakgold, Greens functions and boundary value problems, Wiley, New York, 1979.
[221] D. I. Podolsky, J. Majumder, and N. Jokela, Disorder on the landscape, JCAP 0805, 024
(2008), 0804.2263.
[222] R. Bousso and I.-S. Yang, Landscape predictions from cosmological vacuum selection,
Phys. Rev. D75, 123520 (2007), hep-th/0703206.
[223] D. Schwartz-Perlov, Probabilities in the Arkani-Hamed-Dimopolous-Kachru landscape, J.
Phys. A40, 73637374 (2007), hep-th/0611237.
[224] K. D. Olum and D. Schwartz-Perlov, Anthropic prediction in a large toy landscape, JCAP
0710, 010 (2007), arXiv:0705.2562 [hep-th].
[225] D. Schwartz-Perlov, Anthropic prediction for a large multi-jump landscape, (2008),
0805.3549.
[226] A. Aguirre, T. Banks, and M. Johnson, Regulating eternal ination. II: The great divide,
JHEP 08, 065 (2006), hep-th/0603107.
[227] M. C. Johnson and M. Larfors, Field dynamics and tunneling in a ux landscape, (2008),
0805.3705.
[228] B. Freivogel and M. Lippert, Evidence for a bound on the lifetime of de Sitter space,
(2008), 0807.1104.
[229] M. C. Johnson and M. Larfors, An obstacle to populating the string theory landscape,
(2008), 0809.2604.
[230] K. Athreya and P. Ney, Branching Processes, Springer-Verlag, 1972.
125