Preprint typeset in JHEP style - HYPER VERSION

Eternal inflation predicts that time will end

arXiv:1009.4698v1 [hep-th] 23 Sep 2010

Raphael Boussoa,b,c , Ben Freivogeld , Stefan Leichenauera,b and Vladimir

Rosenhausa,b
a Center for Theoretical Physics and Department of Physics
University of California, Berkeley, CA 94720-7300, U.S.A.
b Lawrence Berkeley National Laboratory, Berkeley, CA 94720-8162, U.S.A.
c Institute for the Physics and Mathematics of the Universe

University of Tokyo, 5-1-5 Kashiwa-no-Ha, Kashiwa City, Chiba 277-8568, Japan

d Center for Theoretical Physics and Laboratory for Nuclear Science

Massachusetts Institute of Technology, Cambridge, MA 02139, U.S.A.

Abstract: Present treatments of eternal inflation regulate infinities by imposing a

geometric cutoff. We point out that some matter systems reach the cutoff in finite
time. This implies a nonzero probability for a novel type of catastrophe. According
to the most successful measure proposals, our galaxy is likely to encounter the cutoff
within the next 5 billion years.
Contents
1. Time will end 1

2. The probability for time to end 6

3. Objections 9
3.1 Time cannot end in a late-time limit 9
3.2 The end of time is an artifact 10

4. The Guth-Vanchurin paradox 12

5. Discussion 13
5.1 Assumptions 14
5.2 Observation 15
5.3 Interpretation 17

1. Time will end

A sufficiently large region of space with positive vacuum energy will expand at an
exponential rate. If the vacuum is stable, this expansion will be eternal. If it is
metastable, then the vacuum can decay by the nonperturbative formation of bubbles
of lower vacuum energy. Vacuum decay is exponentially suppressed, so for a large range
of parameters the metastable vacuum gains volume due to expansion faster than it loses
volume to decays [1]. This is the simplest nontrivial example of eternal inflation.
If it does occur in Nature, eternal inflation has profound implications. Any type of
event that has nonzero probability will happen infinitely many times, usually in widely
separated regions that remain forever outside of causal contact. This undermines the
basis for probabilistic predictions of local experiments. If infinitely many observers
throughout the universe win the lottery, on what grounds can one still claim that
winning the lottery is unlikely? To be sure, there are also infinitely many observers
who do not win, but in what sense are there more of them? In local experiments such
as playing the lottery, we have clear rules for making predictions and testing theories.
But if the universe is eternally inflating, we no longer know why these rules work.

–1–
 To see that this is not merely a philosophical point, it helps to consider cosmological
experiments, where the rules are less clear. For example, one would like to predict or
explain features of the CMB; or, in a theory with more than one vacuum, one might
wish to predict the expected properties of the vacuum we find ourselves in, such as the
Higgs mass. This requires computing the relative number of observations of different
values for the Higgs mass, or of the CMB sky. There will be infinitely many instances
of every possible observation, so what are the probabilities? This is known as the
“measure problem” of eternal inflation.
In order to yield well-defined probabilities, eternal inflation requires some kind
of regulator. Here we shall focus on geometric cutoffs, which discard all but a finite
portion of the eternally inflating spacetime. The relative probability of two types of
events, 1 and 2, is then defined by
p1 hN1 i
= (1.1)
p2 hN2 i

where hN1 i is the expected number of occurences of the first type of event within
the surviving spacetime region. (We will drop the expectation value brackets below
for simplicity of notation.) Here, 1 and 2 might stand for winning or not winning
the lottery; or they might stand for a red or blue power spectrum in the CMB. The
generalization to larger or continuous sets of mutually exclusive outcomes is trivial.
There are different proposals for what spacetime region should be retained. Our
basic observation in this paper applies to all geometric cutoffs we are aware of, and
indeed seems to be an inevitable consequence of any simple geometric cutoff: Some
observers will have their lives interrupted by the cutoff (Fig. 1).
Let events 1 and 2 be the observation of 1 o’clock and 2 o’clock on an observer’s
watch. For simplicity, we will suppose that local physics can be approximated as
deterministic, and we neglect the probability that the observer may die between 1 and
2 o’clock, or that the clock will break, etc. Each observer is born just before his watch
shows 1 and dies just after it shows 2, so that no observer can see more than one event
of each type.
Conventionally, we would say that every observer sees both 1 o’clock and then
2 o’clock. But the figure shows that for some observers, 2 o’clock will not be under
the cutoff even if 1 o’clock is. A fraction 1 − N2 /N1 of observers are prevented from
surviving to 2 o’clock. The catastrophic event in question is evidently the cutoff itself:
the observer might run into the end of time.
One can imagine a situation where the relative number of observations of 1 o’clock
and 2 o’clock is relevant to predicting the results of an experiment. Suppose that the
observers fall asleep right after seeing 1 o’clock. They wake up just before 2 o’clock

–2–
 2 2

2 2
1
2 1
2

2 2 2 2 2 2
1 1
1 1
2 2 2 2
1 1 1 1 1 1
2 2 2 2 2 2

1 1
1 1

1 1 1 1 1 1

(a) Global Cutoff (b) Causal Patch Cutoff

Figure 1: A multiverse populated by infinitely many observers (vertical line segments) who
first see 1 o’clock (at events labeled “1”) and then 2 o’clock (“2”). A geometric cutoff selects a
finite set of events, whose relative frequency defines probabilities. Events that are not counted
are indicated by dashed lines. The left figure shows a global cutoff: all events prior to the
time t0 (curved line) are counted and all later events ignored. (The global time has nothing
to do with the observers’ clocks, which come into being at a rate dictated by the dynamics of
eternal inflation.) The right figure shows the causal patch cutoff, which restricts to the causal
past of a point on the future boundary. In both figures, the cutoff region contains observers
who see 1 o’clock but not 2 o’clock. Their number, as a fraction of all observers who see 1
o’clock before the cutoff, defines the probability of reaching the end of time between 1 and 2
o’clock.

with complete memory loss: they have no idea whether they have previously looked at
their watches before. In this case, they may wish to make a prediction for what time
they will see. Since N2 < N1 , by Eq. (1.1), p2 < p1 : the observation of 2 o’clock is less
probable than that of 1 o’clock. This is possible only if some observers do not survive
to 2 o’clock.
The conclusion that time can end obtains whether or not the observers have mem-
ory loss. Consider an observer who retains her memory. She is aware that she is about
to look at her watch for the first time or for the second time, so the outcome won’t be
a surprise on either occasion. But this does not contradict the possibility that some
catastrophic event may happen between 1 and 2 o’clock. The figure shows plainly that
this event does happen to a nonzero fraction of observers. The only thing that changes
when considering observers who remember is the type of question we are likely to ask.
Instead of asking about two alternative events (1 or 2), we may find it more natural to
ask about the relative probability of the two different possible histories that observers
can have. One history,“1−” , consists of seeing 1 o’clock and running into the cutoff.
The alternative, “12”, consists of seeing 1 o’clock and then seeing 2 o’clock. From
Fig. 1 we see that N12 = N2 and N1− = N1 − N2 . Since N1 > N2 , we have p1− > 0:
there is a nonzero probability for the history in which the observer is disrupted by the

–3–
end of time.

Outline and Frequently Asked Questions The probability for the end of time is
nonzero for all geometric cutoffs. Its value, however, depends on the cutoff. In Sec. 2
we compute the probability, per unit proper time, that we will encounter the end of
time.
A number of objections may be raised against our conclusion that time can end.

• Q: Can’t I condition on the end of time not happening?1

A: Certainly. This is like asking what the weather will be tomorrow, supposing
that it will not rain. It is a reasonable question with a well-defined answer: The
sun will shine with probability x, and it will snow with probability 1 − x. But
this does not mean that it cannot rain. If the end of time is a real possibility,
then it cannot be prevented just by refusing to ask about it.

• Q: In some measures, the cutoff is taken to later and later times. In this limit,
the probability to encounter the end of time surely approaches 0?
A: No. In all known measures of this type, an attractor regime is reached where
the number of all types of events grows at the same exponential rate, includ-
ing observers who see 1 o’clock. The fraction of these observers who also see 2
o’clock before the cutoff approaches a constant less than unity, as will be shown
in Sec. 3.1.

• Q: But as the cutoff is taken to later times, any given observer’s entire history is
eventually included. Isn’t this a contradiction?
A: No. We do not know which observer we are, so we cannot identify with any
particular observer. (If we could, there would be no need for a measure.) Rather,
we consider all observers in a specified class, and we define probabilities in terms
of the relative frequency of different observations made by these observers.
1
In the above example, this would force us to ask a trivial question (“What is the relative probability
of seeing 1 or 2, for an observer whose history includes both 1 and 2?”), which yields the desired answer
(p2 /p1 = 1). For a more interesting example, consider an experiment that terminates at different times
depending on the outcome, such as the Guth-Vanchurin paradox described in Sec. 4, or the decay of
a radioactive atom. In such experiments it is tempting to propose that the experiment should be
regarded to last for an amount of time corresponding to the latest possible (conventional) outcome,
regardless of the actual outcome; and that any outcome (other than the end of time) should be counted
only if the experiment has been entirely completed before the cutoff, in the above sense. This proposal
is not only futile (as described in the answer), but also ill-defined, since any event in the past light-cone
of the event P can be regarded as the beginning of an experiment that includes P .

–4–
 • Q: If I looked at what happened on Earth up to the present time (my “cutoff”), I
would find not only records of past clocks that struck both 1 and 2, but also some
recently manufactured clocks that have struck 1 but not yet 2. I could declare
that the latter represent a new class of clocks, namely clocks whose existence is
terminated by my “cutoff”. But I know that this class is fake: it wasn’t there
before I imposed the “cutoff”. Surely, the end of time in eternal inflation is also
an artifact that ought to be ignored?
A: Only a finite number of clocks will ever be manufactured on Earth. Proba-
bilities are given not in terms of the sample picked out by your “cutoff”, but by
relative frequencies in the entire ensemble. If every clock ever built (in the past or
future) strikes both 1 and 2, then the probability for a randomly chosen clock to
undergo a different history vanishes, so we may say confidently that the “cutoff”
has introduced an artifact. In eternal inflation, however, the cutoff cannot be
removed. Otherwise, we would revert to a divergent multiverse in which relative
frequencies are not well-defined. The cutoff defines not a sample of a pre-existing
ensemble; it defines the ensemble. This is further discussed in Sec. 3.2.

• Q: Why not modify the cutoff to include 2 o’clock?

A: This is a possibility. If we deform the cutoff hypersurface so that it passes
through no matter system, then nothing will run into the end of time. It is not
clear whether this type of cutoff can be obtained from any well-defined prescrip-
tion. At a minimum, such a prescription would have to reference the matter
content of the universe explicitly in order to avoid cutting through the world
volumes of matter systems. In this paper, we consider only cutoffs defined by a
purely geometric rule, which take no direct account of matter.

In Sec. 4, we discuss an apparent paradox that is resolved by the nonzero probability

for time to end.
Any conclusion is only as strong as the assumptions it rests on. The reader who feels
certain that time cannot end may infer that at least one of the following assumptions
are wrong: (1) the universe is eternally inflating; (2) we may restrict attention to a finite
subset of the eternally inflating spacetime, defined by a purely geometric prescription;
and (3) probabilities are computed as relative frequencies of outcomes in this subset,
Eq. (1.1). We discuss these assumptions in Sec. 5.1.
In Sec. 5.2, we discuss whether, and how, the nonzero probability for the end of time
may be observed. We point out that known predictions of various measures actually
arise from the possibility that time can end. On the problematic side, this includes

–5–
the famous youngness paradox of the proper time cutoff; on the successful side, the
prediction of the cosmological constant from the causal patch cutoff.
In Sec. 5.3, we discuss how the end of time fits in with the rest of physics. This
depends on the choice of cutoff. With the causal patch cutoff, there may be a relatively
palatable interpretation of the end of time which connects with the ideas of black hole
complementarity. The boundary of the causal patch is a kind of horizon, which can be
treated as an object with physical attributes, including temperature. Matter systems
that encounter the end of time are thermalized at this horizon. This is similar to an
outside observer’s description of a matter system falling into a black hole. What is
radically new, however, is the statement that we might experience thermalization upon
crossing the black hole horizon.
This work was inspired by discussions with Alan Guth, who first described to us
the paradox mentioned in section 4. We understand that Guth and Vanchurin will be
publishing their own conclusions [2]. In taking seriously the incompleteness of spacetime
implied by geometric cutoffs, our conclusion resembles a viewpoint suggested earlier by
Ken Olum [3].

2. The probability for time to end

The phenomenon that time can end is universal to all geometric cutoffs. But the rate
at which this is likely to happen, per unit proper time τ along the observer’s worldline,
is cutoff-specific. We will give results for five measures.

Causal patch The causal patch cutoff [4] restricts attention to the causal past of
the endpoint of a single worldline (see Fig. 1). Expectation values are computed by
averaging over initial conditions and decoherent histories in the causal patch. The end
of time, in this case, is encountered by systems that originate inside the causal patch
but eventually exit from it.
Our universe can be approximated as a flat FRW universe with metric

ds2 = −dτ 2 + a(τ )2 (dχ2 + χ2 dΩ2 ) . (2.1)

Observers are approximately comoving (dχ/dτ = 0). We assume that the decay rate
of our vacuum, per unit four-volume, is much less than t−4Λ . Then the decay can be
neglected entirely in computing where the boundary of the causal patch intersects the
equal time surfaces containing observers. The boundary is given by the de Sitter event
horizon: Z ∞
dτ 0
χE (τ ) = . (2.2)
τ a(τ 0 )

–6–
 We consider all observers currently within the horizon: χ < χE (τ0 ), with τ0 = 13.7
Gyr. This corresponds to a comoving volume Vcom = (4π/3)χE (τ0 )3 . Observers located
at χ leave the patch at a time τ 0 determined by inverting Eq. (2.2); in other words,
they reach the end of time at ∆τ ≡ τ 0 − τ0 from now. An (unnormalized) probability
distribution over ∆τ is obtained by computing the number of observers that leave the
causal patch at the time τ0 + ∆τ :

dp 4πχE (τ0 + ∆τ )2
∝ . (2.3)
d∆τ a(τ0 + ∆τ )

We compute a(τ ) numerically using the best-fit cosmological parameters from the
WMAP5 data combined with SN and BAO [5]. From the distribution (2.3), we may
obtain both the median and the expectation value for ∆τ . We find that the expected
amount of proper time left before time ends is

h∆τ i = 5.3 Gyr . (2.4)

Time is unlikely to end in our lifetime, but there is a 50% chance that time will end
within the next 3.7 billion years.

Light-cone time The light-cone time of an event is defined in terms of the volume
of its future light-cone on the future boundary of spacetime [6–8]. The light-cone time
cutoff requires that we only consider events prior to some light-cone time t0 ; then the
limit t0 → ∞ is taken. It can be shown that the light-cone time cutoff is equivalent to
the causal patch cutoff with particular initial conditions [7]. Thus, the probability for
an observer to encounter the end of time is the same as for the causal patch cutoff.

Fat geodesic The fat geodesic cutoff considers a fixed proper volume 4πd3 /3 near
a timelike geodesic [9]. To compute probabilities, one averages over an ensemble of
geodesics orthogonal to an initial hypersurface whose details will not matter. One
can show that the geodesics quickly become comoving after entering a bubble of new
vacuum. Since our vacuum is homogeneous, we may pick without loss of generality
a fat geodesic at χ = 0. We shall neglect the effects of local gravitational collapse
and approximate the universe as expanding homogeneously. Equivalently, we take the
proper distance d to be small compared to the present curvature scale of the universe
but large compared to the scale of clusters. These approximations are not essential,
but they will simplify our calculation and save us work when we later consider the scale
factor cutoff.
We should only consider observers who are currently (τ0 = 13.7 Gyr) within the
fat geodesic, with χ < d/a(τ0 ). An observer leaves the geodesic a time ∆τ later, with

–7–
χ = d/a(τ0 + ∆τ ). The unnormalized probability distribution over ∆τ is
dp d3 (da/dτ )τ0 +∆τ
∝ 4π . (2.5)
d∆τ a(τ0 + ∆τ )4
From this distribution, we find numerically that the expected amount of proper time
left before the end of time is 5 Gyr. There is a 50% chance that time will end within
the next 3.3 billion years.
While the result is similar, there is an important formal difference between the fat
geodesic and causal patch cutoffs. The boundary of the fat geodesic is a timelike hyper-
surface, from which signals can propagate into the cutoff region. Boundary conditions
must therefore be imposed. When a system leaves the fat geodesic, time ends from its
own point of view. But an observer who remains within the cutoff region continues to
see the system and to communicate with it. The image of the system and its response
to any communications are encoded in data specified on the timelike boundary. In
practice, the simplest way to determine these boundary conditions is to consider the
global spacetime and select a fat geodesic from it. This means that the fat geodesic
is not a self-contained description. The content of the causal patch, by contrast, can
be computed from its own initial conditions without reference to a larger spacetime
region.

Scale factor time Scale factor time is defined using a congruence of timelike geodesics
orthogonal to some initial hypersurface in the multiverse: dt ≡ Hdτ , where τ is the
proper time along each geodesic and 3H is the local expansion of the congruence. This
definition breaks down in nonexpanding regions such as dark matter halos; attempts
to overcome this limitation (e.g., Ref. [10]) remain somewhat ad-hoc. Here we use for
H the Hubble rate of a completely homogeneous universe whose density agrees with
the average density of our universe. This does not yield a precise and general cutoff
prescription, but it allows us to compute an approximate rate at which we are likely
to encounter the cutoff: in an everywhere-expanding timelike geodesic congruence, the
scale factor time cutoff is equivalent to the fat geodesic cutoff [9]. Hence it gives the
same rate for time to end as the fat geodesic cutoff.

Proper time In the proper time cutoff, the characteristic timescale is the shortest
Hubble time of all eternally inflating vacua. In a realistic landscape, this is microscopi-
cally short, perhaps of order the Planck time [11]. Thus, time would be overwhelmingly
likely to end in the next second:
dp
≈ t−1
Pl . (2.6)
d∆τ
This is the famous “youngness paradox” in a new guise. The cutoff predicts that our
observations have superexponentially small probability, and that most observers are

–8–
“Boltzmann babies” who arise from quantum fluctuations in the early universe. Thus,
this measure is already ruled out phenomenologically at a high level of confidence [11–
18].

3. Objections
Our intuition rebels against the conclusion that spacetime could simply cease to exist.
In the introduction, we answered several objections that could be raised against the
end of time. In this section, we will discuss two of these arguments in more detail.

3.1 Time cannot end in a late-time limit

In some measure proposals, such as the proper time cutoff [19,20], the scale factor time
cutoff [10], and the light-cone time cutoff [6], a limit is taken in which the cutoff region
is made larger and larger as a function of a time parameter t0 :

p1 /p2 = lim N1 (t0 )/N2 (t0 ) . (3.1)

t0 →∞

Naively one might expect the cutoff observers to be an arbitrarily small fraction of all
observers in the limit t0 → ∞. This turns out not to be the case.
One finds that the number of events of type I that have occurred prior to the time
t is of the form
NI (t) = ŇI exp(γt) + O(σt) , (3.2)
with σ < γ ≈ 3. Thus, the growth approaches a universal exponential behavior at
late times [7, 11, 21], independently of initial conditions. The ratio N1 /N2 appearing in
Eq. (1.1) remains well-defined in the limit t0 → ∞, and one obtains
p1 Ň1
= . (3.3)
p2 Ň2
The constants ŇI , and thus the relative probabilities, depend on how the time variable
is defined; we will discuss some specific choices below.
Suppose that observers live for a fixed global time interval ∆t. Then a person dies
before time t if and only if he was born before t − ∆t. Therefore the number of births
Nb is related to the number of deaths Nd by

Nd (t) = Nb (t − ∆t) (3.4)

Using the time dependence of Nb given in (3.2), this can be rewritten

Nd (t)
≈ exp(−γ∆t) , (3.5)
Nb (t)

–9–
up to a correction of order e(σ−γ)t which becomes negligible in the late time limit. Thus,
the fraction of deaths to births does not approach unity as the cutoff is taken to infinity.
The fraction of observers whose lives are interrupted by the cutoff is
Nc
= 1 − exp(−γ∆t) (3.6)
Nb
where Nc = Nb − Nd is the number of cutoff observers.
Since (3.6) is true for any time interval ∆t, it is equivalent to the following simple
statement: any system has a constant probability to encounter the end of time given
by
dp
=γ≈3 (3.7)
dt
This result can be interpreted as follows. Due to the steady state behavior of eternal
inflation at late times, there is no way to tell what time it is. The exponential growth
(3.2) determines a t0 -independent probability distribution for how far we are from the
cutoff, given by (3.7).

3.2 The end of time is an artifact

Could it be that observers who run into the cutoff are an artifact, not to be taken
seriously as a real possibility? Certainly they would not exist but for the cutoff. Yet,
we argue that cutoff observers are a real possibility, because there is no well-defined
probability distribution without the cutoff; in particular, only the cutoff defines the set
of allowed events. In order to convince ourselves of this, it is instructive to contrast
this situation with one where a cutoff may introduce artifacts. We will consider two
finite ensembles of observers, without reference to eternal inflation. We then restrict
attention to a portion of each ensemble, defined by a cutoff. We find that this sample
looks the same in both ensembles, and that it contains observers that run into the
cutoff. In the first ensemble, these observers are an artifact of sampling; in the second,
they are real. We will then explain why eternal inflation with a cutoff is different from
both examples.

A cutoff on a finite ensemble defines a sample Consider a civilization which

begins at the time t = 0 and grows exponentially for a finite amount of time. (We will
make no reference to a multiverse in this example.) Every person is born with a watch
showing 1 o’clock at the time of their birth, when they first look at it. One hour later,
when they look again, the watch shows 2 o’clock; immediately thereafter the person
dies. After the time t∗  1 hour, no more people are born, but whoever was born
before t∗ gets to live out their life and observe 2 o’clock on their watch before they
die. In this example, there is a well-defined, finite ensemble, consisting of all observers

– 10 –
throughout history and their observations of 1 and 2. The ensemble contains an equal
number of 1’s and 2’s. Every observer in the ensemble sees both a 1 and a 2, each with
100% probability. No observer meets a premature demise before seeing 2.
Now suppose that we do not know the full ensemble described above. Instead,
we are allowed access only to a finite sample drawn from it, namely everything that
happened by the time t, with 1 hour  t < t∗ . This sample contains many observers
who died before t; each of them will have seen both 1 and 2. We refer to these as
“histories” of type 12. It also contains observers (those who were born after t − 1 hour)
who are still alive. Each of them has seen 1 but not yet 2, by the time t, which we
refer to as a history of type 1−. What do we say about these observers? Should we
declare that there is a nonzero probability that an observer who sees 1 does not live
to see 2? In fact, a finite sample of a larger, unknown ensemble allows us to draw no
conclusion of this kind, because we have no guarantee that our sampling of the ensemble
is fair. The true set of outcomes, and their relative frequency, is determined only by
the full ensemble, not by our (possibly unfaithful) sample of it. Similarly, if we had
a considered a more complicated system, such as observers with watches of different
colors, etc., the relative frequency of outcomes in any subset need not be the same as
the relative frequency in the full ensemble, unless we make further assumptions.
If we examined the full ensemble, we could easily verify that every observer who
sees 1 also lives to see 2. Thus we would learn that 1− was an artifact of our sampling:
imposing a cutoff at fixed time produced a new class of events that does not exist (or
more precisely, whose relative frequency drops to zero) once we take the cutoff away.
Armed with this knowledge, we could then invent an improved sampling method, in
which the 1− cases are either excluded, or treated as 12 events.
As our second example, let us consider a civilization much like the previous one,
except that it perishes not by a sudden lack of births, but by a comet that kills everyone
at the time t∗ . This, too, gives rise to a finite, well-defined ensemble of observations.
But unlike in the previous example, there is a larger number of 1’s than 2’s: not every
observer who sees a 1 lives to see a 2. Thus, the probabilities for the histories 12 and
1− satisfy p1− > 0, p12 < 1. Indeed, if we choose parameters so the population grows
exponentially on a timescale much faster than 1 hour, most people in history who see
1 end up being killed by the comet rather than expiring naturally right after seeing 2;
that is, p12 = 1 − p1−  1 in this limit.
Again, we can contemplate sampling this ensemble, i.e., selecting a subset, by
considering everything that happened prior to the time t < t∗ . Note that this sample
will look identical to the finite-time sample we were given in the previous example.
Again, we find that there are apparently events of type 1−, corresponding to observers
who have seen 1 but not 2 by the time t. But in this example, it so happens that

– 11 –
(i) events of type 1− actually do exist in the full ensemble, i.e., have nonzero true
relative frequency; and (ii) assuming exact exponential growth, our sample is faithful:
the relative frequency of 1− vs. 12 in the sample (observers prior to t) is the same as
in the full ensemble (observers in all history, up to t∗ ).2
We learn from the above two examples that a subset of an ensemble need not yield
reliable quantitative information about the relative frequencies of different outcomes,
or even qualitative information about what the allowed outcomes are. All of this
information is controlled only by the full ensemble. In both examples, the set of events
that occured before the time t < t∗ contain events of type 1−. But in the first example,
these events are a sampling artifact and their true probability is actually 0. In the
second example, 1− corresponds to a real possibility with non-zero probability.

The cutoff in eternal inflation defines the ensemble Now let us return to eternal
inflation. In order to regulate its divergences, we define a cutoff that picks out a finite
spacetime region, for example the region prior to some constant light-cone time t.
Naively, this seems rather similar to the examples above, where we sampled a large
ensemble by considering only the events that happened prior to a time t < t∗ . But we
learned that such samples cannot answer the question of whether the histories of type
1− are real or artifacts. To answer this question, we had to look at the full ensemble.
We found in the first example that 1− was real, and in the second that 1− was an
artifact, even though the sample looked exactly the same in both cases. In eternal
inflation, therefore, we would apparently need to “go beyond the cutoff” and consider
the “entire ensemble” of outcomes, in order to decide whether 1− is something that
can really happen.
But this is impossible: the whole point of the cutoff was to define an ensemble.
An infinite set is not a well-defined ensemble, so the set we obtained by imposing a
cutoff is the most fundamental definition of an ensemble available to us. We can argue
about which cutoff is correct: light-cone time, scale factor time, the causal patch, etc.
But whatever the correct cutoff is, its role is to define the ensemble. It cannot be said
to select a sample from a larger ensemble, namely from the whole multiverse, because
this larger ensemble is infinite, so relative abundances of events are not well-defined. If
they were, we would have had no need for a cutoff in the first place.

4. The Guth-Vanchurin paradox

Another way to see that the end of time is a real possibility is by verifying that it
2
Actually it is faithful only in the limit as t is much greater than the characteristic growth timescale
of the civilization, because of the absence of any observers prior to t = 0.

– 12 –
resolves a paradox exhibited by Guth and Vanchurin [22]. Suppose that before you go
to sleep someone flips a fair coin and, depending on the result, sets an alarm clock to
awaken you after either a short time, ∆t  1, or a long time ∆t  1. Local physics
dictates that there is a 50% probability to sleep for a short time since the coin is fair.
Now suppose you have just woken up and have no information about how long you
slept. It is natural to consider yourself a typical person waking up. But if we look at
everyone who wakes up before the cutoff, we find that there are far more people who
wake up after a short nap than a long one. Therefore, upon waking, it seems that there
is no longer a 50% probability to have slept for a short time.
How can the probabilities have changed? If you accept that the end of time is a
real event that could happen to you, the change in odds is not surprising: although the
coin is fair, some people who are put to sleep for a long time never wake up because
they run into the end of time first. So upon waking up and discovering that the world
has not ended, it is more likely that you have slept for a short time. You have obtained
additional information upon waking—the information that time has not stopped—and
that changes the probabilities.
However, if you refuse to believe that time can end, there is a contradiction. The
odds cannot change unless you obtain additional information. But if all sleepers wake,
then the fact that you woke up does not supply you with new information.
Another way to say it is that there are two reference classes one could consider.
When going to sleep we could consider all people falling asleep; 50% of these people
have alarm clocks set to wake them up after a short time. Upon waking we could
consider the class of all people waking up; most of these people slept for a short time.
These reference classes can only be inequivalent if some members of one class are not
part of the other. This is the case if one admits that some people who fall asleep never
wake up, but not if one insists that time cannot end.

5. Discussion

Mathematically, the end of time is the statement that our spacetime manifold is ex-
tendible, i.e., that it is isometric to a proper subset of another spacetime. Usually, it
is assumed that spacetime is inextendable [23]. But the cutoffs we considered regulate
eternal inflation by restricting to a subset of the full spacetime. Probabilities are fun-
damentally defined in terms of the relative abundance of events and histories in the
subset. Then the fact that spacetime is extendible is itself a physical feature that can
become part of an observer’s history. Time can end.

– 13 –
5.1 Assumptions
We do not know whether our conclusion is empirically correct. What we have shown is
that it follows logically from a certain set of assumptions. If we reject the conclusion,
then we must reject at least one of the following propositions:

Probabilities in a finite universe are given by relative frequencies of events

or histories This proposition is sometimes called the assumption of typicality. It
forces us to assign a nonzero probability to encountering the end of time if a nonzero
fraction of observers encounter it.
Even in a finite universe one needs a rule for assigning relative probabilities to
observations. This need is obvious if we wish to make predictions for cosmological
observations. But a laboratory experiment is a kind of observation, too, albeit one
in which the observer controls the boundary conditions. A comprehensive rule for
assigning probabilities cannot help but make predictions for laboratory experiments in
particular. However, we already have a rule for assigning probabilities in this case,
namely quantum mechanics and the Born rule, applied to the local initial conditions
prepared in the laboratory. This must be reproduced as a special case by any rule
that assigns probabilities to all observations [11]. A simple way to achieve this is by
defining probabilities as ratios of the expected number of instances of each outcome in
the universe, as we have done in Eq. (1.1).

Probabilities in an infinite universe are defined by a geometric cutoff This

proposition states that the infinite spacetime of eternal inflation must be rendered
finite so that the above frequency prescription can be used to define probabilities.
Moreover, it states that a finite spacetime should be obtained by restricting attention
to a finite subset of the infinite multiverse.3 It is possible that the correct measure
cannot be expressed in a geometric form. Imagine, for instance, a measure that makes
“exceptions” for matter systems that come into existence before the cutoff, allowing
all events in their world volume to be counted. A purely geometric prescription would
have chopped part of the history off, but in this measure, the cutoff surface would
be deformed to contain the entire history of the system. Such a cutoff would depend
not only on the geometry, but also on the matter content of spacetime.4 A more
3
We have considered measures in which the cutoff is completely sharp, i.e., described by a hyper-
surface that divides the spacetime into a region we keep and a region we discard. In fact this is not
essential. One could smear out the cutoff by assigning to each spacetime event a weight that varies
smoothly from 1 to 0 over some region near the cutoff surface. There would still be a finite probability
for time to end.
4
We have not attempted to prove this statement, so it should be considered an additional as-
sumption. Because the metric has information about the matter content, we cannot rule out that a

– 14 –
radical possibility is that the measure may not involve any kind of restriction to a
finite portion of spacetime. For example, Noorbala and Vanchurin [24], who exhibit a
paradox similar to that described in Sec. 4, but do not allow for the possibility that
time can end, advocate a nongeometric type of measure. If such a prescription could be
made well-defined and consistent with observation (which seems unlikely to us), then
one might escape the conclusion that time can end. Similarly, Winitzki [25–27] defined
a measure where only finite spacetimes are considered, and in this measure there is no
novel catastrophe like the end of time.

The universe is eternally inflating To prove this proposition wrong would be a

dramatic result, since it would seem to require a kind of fundamental principle dictating
that Nature abhors eternal inflation. After all, eternal inflation is a straightforward
consequence of general relativity, assuming there exists at least one sufficiently long-
lived field theory vacuum with positive vacuum energy (a de Sitter vacuum). This
assumption, in turn, seems innocuous and is well-motivated by observation: (1) The
recent discovery of accelerated expansion [28, 29], combined with the conspicuous lack
of evidence that dark energy is not a cosmological constant [5], suggests that our own
vacuum is de Sitter. If this is the case, the universe must be eternally inflating. (2) Slow-
roll inflation solves the horizon and flatness problems. Its generic predictions agree well
with the observed CMB power spectrum. But slow-roll inflation requires a sufficiently
flat scalar field potential. Eternal inflation requires only a local minimum and so is less
fine-tuned. How could we consider slow-roll inflation, but exclude eternal inflation? —
There are also theoretical motivations for considering the existence of de Sitter vacua:
(3) In effective field theory, there is nothing special about potentials with a positive
local minimum, so it would be surprising if they could not occur in Nature. (4) String
theory predicts a very large number of long-lived de Sitter vacua [30–32], allowing for
a solution of the cosmological constant problem and other fine-tuning problems.

5.2 Observation
If we accept that time can end, what observable implications does this have? Should
we expect to see clocks or other objects suddenly disappear? In measures such as scale
factor time or light-cone time, the expected lifetime of stable systems is of order 5
billion years right now, so it would be very unlikely for the end of time to occur in, say,

geometric measure could be formulated whose cutoff surfaces never intersect with matter. It seems
unlikely to us that such a cutoff could select a finite subset of the multiverse. A related possibility
would be to define a global time cutoff such that typical observers live farther and farther from the
cutoff in the limit as t → ∞. This would invalidate our analysis in Sec. 3.1, which assumed exponential
growth in t.

– 15 –
the next thousand years. And even if it did occur, it would not be observable. Any
observer who would see another system running into the end of time is by definition
located to the causal future of that system. If the cutoff surface is everywhere spacelike
or null, as is the case for the light-cone time cutoff and the causal patch cutoff, then the
observer will necessarily run into the cutoff before observing the demise of any other
system.
Though the end of time would not be observable, the fact that time has not ended
certainly is observable. If a theory assigns extremely small probability to some event,
then the observation of this event rules out the theory at a corresponding level of
confidence. This applies, in particular, to the case where the event in question is time-
not-having-ended. For example, Eq. (2.6) shows that the proper time measure is thus
falsified.
An observation which indirectly probes the end of time is the value of the cosmo-
logical constant. For definiteness consider the causal patch measure, which predicts
a coincidence between the time when the observers live and the time when the cos-
mological constant beings to dominate the expansion of the universe, tΛ ∼ tobs . This
represents the most important phenomenological success of the measure, and we will
now argue that it is tied intimately to the end of time.
The most likely value of the cosmological constant is the one which leads to the most
observers inside the causal patch. We will assume that there are a constant number√of
observers per unit mass, and will imagine scanning the possible values of tΛ ∼ 1/ Λ
with tobs held fixed. It’s most useful to think of the distribution of values of log tΛ , where
the preferred value is largely determined by two competing pressures. First, since the
prior probability is flat in Λ, there is an exponential pressure in log tΛ toward lesser
values. Second, if tΛ < tobs there is an exponential pressure in tΛ (superexponential in
log tΛ ) toward greater values. This is a simple consequence of the fact that all matter is
expelled from the causal patch at an exponential rate after vacuum domination begins.
These two pressures lead to tobs ∼ tΛ .
The end of time is implicitly present in this argument. Suppose there are two
generations of observers, one living at tΛ and another at 10tΛ . Even if local physics
says that there are the same number of observers per unit mass in each generation, the
second generation must be atypical, and hence have fewer members, if the prediction
for the cosmological constant is to remain valid. Where are the missing members of the
second generation? The answer is that time has ended for them. They are not counted
for the purposes of any calculation, and so they do not exist. Clearly, the setup is
identical to the observers who see 1 o’clock and 2 o’clock discussed above.
In this paper, we have considered sharp geometric cutoffs. However, intuition from
AdS/CFT [6, 8, 33] suggests that the cutoff should not be a completely sharp surface,

– 16 –
but should be smeared out over a time of order tΛ . If the cutoff is smeared, there could
be observable consequences of approaching the end of time; the details would depend
on the precise prescription for smearing the cutoff.

5.3 Interpretation
The notion that time can come to an end is not completely new. Space and time break
down at at singularities, which are guaranteed to arise in gravitational collapse [34]. But
our conclusion is more radical: the world can come to an end in any spacetime region,
including regions with low density and curvature, because spacetime is incomplete.
One might speculate that semiclassical gravity breaks down on very large time
scales, say t3Λ , the evaporation time for a large black hole in de Sitter space, or exp(πt2Λ ),
the recurrence time in de Sitter space. But in the most popular measures, we are likely
to encounter the end of time on the much shorter timescale tΛ . Perhaps one could
invent a new cutoff that would push the end of time further into the future. But there
is no well-motivated candidate we are aware of, and, as we have discussed, one would
be likely to lose some of the phenomenological success of the measures in solving, e.g.,
the cosmological constant problem.
How can we make sense of our conclusion? Is there a way of thinking about it
that would make us feel more comfortable about the end of time? Does it fit in with
something we already know, or is this a completely new phenomenon? The answer to
this question turns out to depend somewhat on which cutoff is used.

All measures One way to interpret the end of time is to imagine a computer just
powerful enough to simulate the cutoff portion of the eternally inflating spacetime.
The simulation simply stops at the cutoff. If the measure involves taking a late time
limit, then one can imagine building larger and larger computers that can simulate the
spacetime until a later cutoff. These computers can be thought of as the definition of
the cutoff theory, much in the same way that lattice gauge theory is used. There is no
physical significance to any time after the cutoff.5 This is an interesting rephrasing of
the statement of the end of time, but it does not seem to mitigate its radicality.

Causal patch only Our result appears to admit an intriguing interpretation if the
causal patch measure is used. The original motivation for the causal patch came from
black hole complementarity [35]. Consider the formation and evaporation of a black
hole in asymptotically flat space. If this process is unitary, then the quantum state
5
Ken Olum has pointed out for some time that one way to interpret a geometric cutoff is that
“we are being simulated by an advanced civilization with a large but finite amount of resources, and
at some point the simulation will stop.” The above interpretation adopts this viewpoint (minus the
advanced civilization).

– 17 –
describing the collapsing star inside the black hole is identical to the state of the
Hawking radiation cloud. Since these two states are spacelike separated, two copies of
the quantum state exist at the same time. But before the star collapsed, there was
only one copy. This amounts to “quantum xeroxing”, which is easily seen to conflict
with quantum mechanics.
A way around this paradox is to note there is no spacetime point whose past light-
cone contains both copies. This means that no experiment consistent with causality
can actually verify that xeroxing has taken place. Thus, the paradox can be regarded
as an artifact of a global viewpoint that has no operational basis. A theory should
be capable of describing all observations, but it need not describe more than that.
Geometrically, this means that it need not describe any system that cannot fit within
a causal patch. What the xeroxing paradox teaches us is that we must not describe
anything larger than the causal patch if we wish to avoid inconsistencies in spacetimes
with black holes.
But once we reject the global description of spacetime, we must reject it whether or
not black holes are present. In many cosmological solutions, including eternal inflation,
the global spacetime is not accessible to any single experiment. This motivated the use
of the causal patch as a cutoff to regulate the infinities of eternal inflation [4, 36].
Let us return to the black hole spacetime and consider the causal patch of an
outside observer. This patch includes all of the spacetime except for the interior of the
black hole. As Susskind has emphasized, to an outside observer, the causal patch is a
consistent representation of the entire world. The patch has a boundary, the stretched
horizon of the black hole. This boundary behaves like a physical membrane, endowed
with properties such as temperature, tension, and conductivity. When another observer
falls into the black hole, the outside observer would say that he has been thermalized at
the horizon and absorbed into the membrane degrees of freedom. Later the membrane
evaporates and shrinks away, leaving behind a cloud of radiation.
It is very important to understand that this really is the unique and complete
description of the process from the outside point of view; the black hole interior does
not come into it. The process is no different, in principle, from throwing the second
observer into a fire and watching the smoke come out. Any object is destroyed upon
reaching the horizon. Yet, assuming that the black hole is large, the infalling observer
would not notice anything special when crossing the horizon. There is no contradiction
between these two descriptions, since they agree as long as the two observers remain
in causal contact. Once they differ, it is too late for either observer to send a signal to
the other and tell a conflicting story.
The end of time in the causal patch is an effect that fits well with the outside
observer’s description. When the infalling observer enters the black hole, he is leaving

– 18 –
the causal patch of the outside observer. In the language of the present paper, the
outside observer defines a particular causal patch, and the inside observer encounters
the end of time when he hits the boundary of this patch. We now see that there
is a different, more satisfying interpretation: the inside observer is thermalized at
the horizon. This interpretation invokes a relatively conventional physical process to
explain why the inside observer ceases to exist. Time does not stop, but rather, the
observer is thermalized. His degrees of freedom are merged with those already existing
at the boundary of the causal patch, the horizon.
If this interpretation is correct, it can be applied to black holes that form in the
eternally inflating universe, where it modifies the theory of the infalling observer. It is
no longer certain that an infalling observer will actually make it to the horizon, and
into the black hole, to perish only once he hits the future singularity. Instead, time
might end before he enters the black hole. How is this possible?
In the traditional discussion of black hole complementarity, one picks an observer
and constructs the associated causal patch. It is impossible, by construction, for an
observer to leave his own patch. In other words, time cannot end if we live in a causal
patch centered on our own worldline. In eternal inflation, however, one first picks a
causal patch; then one looks for observers in it. Some of these observers will be closer
to the boundary and leave the patch sooner than others, who happen to stay in the
patch longer. Equivalently, suppose we do want to begin by considering observers of a
given type, such as an observer falling towards a black hole. To compute probabilities,
we must average over all causal patches that contain such an observer. In some patches
the observer will be initially far from the boundary, in others he will hit the boundary
very soon. This yields a probability distribution for the rate at which time ends.
Suppose, for example, that we attempted to jump into a black hole of mass m in
our own galaxy (and neglect effects of gravitational tidal forces, matter near the black
hole, etc.). Using the ensemble of causal patches defined in Ref. [7], one finds that time
would probably end before we reach the horizon, with probability 1 − O(m/tΛ ). This
probability is overwhelming if the black hole is much smaller than the cosmological
horizon.

Acknowledgments
We would like to particularly thank Adam Brown and Alan Guth for very influential
discussions. We also thank David Berenstein, Steve Shenker, Lenny Susskind, and
Vitaly Vanchurin for helpful discussions. This work was supported by the Berkeley
Center for Theoretical Physics, by the National Science Foundation (award number
0855653), by the Institute for the Physics and Mathematics of the Universe, by fqxi

– 19 –
grant RFP2-08-06, and by the US Department of Energy under Contract DE-AC02-
05CH11231. VR is supported by an NSF graduate fellowship.

References
[1] A. H. Guth and E. J. Weinberg, “Could the universe have recovered from a slow
first-order phase transition?,” Nucl. Phys. B212 (1983) 321–364.

[2] A. Guth and V. Vanchurin. To appear.

[3] K. Olum. Private communication.

[4] R. Bousso, “Holographic probabilities in eternal inflation,” Phys. Rev. Lett. 97 (2006)
191302, hep-th/0605263.

[5] WMAP Collaboration, E. Komatsu et al., “Five-Year Wilkinson Microwave

Anisotropy Probe (WMAP) Observations: Cosmological Interpretation,”
arxiv:0803.0547 [astro-ph].

[6] R. Bousso, “Complementarity in the Multiverse,” Phys. Rev. D79 (2009) 123524,
arXiv:0901.4806 [hep-th].

[7] R. Bousso and I.-S. Yang, “Global-Local Duality in Eternal Inflation,” Phys. Rev. D80
(2009) 124024, arXiv:0904.2386 [hep-th].

[8] R. Bousso, B. Freivogel, S. Leichenauer, and V. Rosenhaus, “Boundary definition of a

multiverse measure,” arXiv:1005.2783 [hep-th].

[9] R. Bousso, B. Freivogel, and I.-S. Yang, “Properties of the scale factor measure,”
arXiv:0808.3770 [hep-th].

[10] A. De Simone, A. H. Guth, M. P. Salem, and A. Vilenkin, “Predicting the cosmological

constant with the scale-factor cutoff measure,” arXiv:0805.2173 [hep-th].

[11] R. Bousso, B. Freivogel, and I.-S. Yang, “Boltzmann babies in the proper time
measure,” Phys. Rev. D77 (2008) 103514, arXiv:0712.3324 [hep-th].

[12] A. Linde, D. Linde, and A. Mezhlumian, “Nonperturbative Amplifications of

Inhomogeneities in a Self-Reproducing Universe,” Phys. Rev. D 54 (1996) 2504–2518,
gr-qc/9601005.

[13] A. H. Guth, “Inflation and Eternal inflation,” Phys. Rep. 333 (1983) 555,
astro-ph/0002156.

[14] A. H. Guth, “Inflationary Models and Connections to Particle Physics,”

astro-ph/0002188.

– 20 –
[15] A. H. Guth, “Inflation,” astro-ph/0404546.

[16] M. Tegmark, “What does inflation really predict?,” JCAP 0504 (2005) 001,
astro-ph/0410281.

[17] A. Linde, “Towards a gauge invariant volume-weighted probability measure for eternal
inflation,” JCAP 0706 (2007) 017, arXiv:0705.1160 [hep-th].

[18] A. H. Guth, “Eternal inflation and its implications,” J. Phys. A40 (2007) 6811–6826,
hep-th/0702178.

[19] A. Linde, “Eternally Existing Selfreproducing Chaotic Inflationary Universe,” Phys.

Lett. B175 (1986) 395–400.

[20] A. Linde, “Sinks in the Landscape, Boltzmann Brains, and the Cosmological Constant
Problem,” JCAP 0701 (2007) 022, hep-th/0611043.

[21] J. Garriga, D. Schwartz-Perlov, A. Vilenkin, and S. Winitzki, “Probabilities in the

inflationary multiverse,” JCAP 0601 (2006) 017, hep-th/0509184.

[22] A. Guth and V. Vanchurin. Private communication.

[23] R. M. Wald, General Relativity. The University of Chicago Press, Chicago, 1984.

[24] M. Noorbala and V. Vanchurin, “Geocentric cosmology: a new look at the measure
problem,” arXiv:1006.4148 [hep-th].

[25] S. Winitzki, “A volume-weighted measure for eternal inflation,” Phys. Rev. D78 (2008)
043501, arXiv:0803.1300 [gr-qc].

[26] S. Winitzki, “Reheating-volume measure for random-walk inflation,” Phys. Rev. D78
(2008) 063517, arXiv:0805.3940 [gr-qc].

[27] S. Winitzki, “Reheating-volume measure in the landscape,” Phys. Rev. D78 (2008)
123518, arXiv:0810.1517 [gr-qc].

[28] Supernova Search Team Collaboration, A. G. Riess et al., “Observational Evidence

from Supernovae for an Accelerating Universe and a Cosmological Constant,” Astron.
J. 116 (1998) 1009–1038, astro-ph/9805201.

[29] Supernova Cosmology Project Collaboration, S. Perlmutter et al., “Measurements

of Omega and Lambda from 42 High-Redshift Supernovae,” Astrophys. J. 517 (1999)
565–586, astro-ph/9812133.

[30] R. Bousso and J. Polchinski, “Quantization of four-form fluxes and dynamical

neutralization of the cosmological constant,” JHEP 06 (2000) 006, hep-th/0004134.

– 21 –
[31] S. Kachru, R. Kallosh, A. Linde, and S. P. Trivedi, “De Sitter vacua in string theory,”
Phys. Rev. D 68 (2003) 046005, hep-th/0301240.

[32] F. Denef and M. R. Douglas, “Distributions of flux vacua,” JHEP 05 (2004) 072,
hep-th/0404116.

[33] J. Garriga and A. Vilenkin, “Holographic Multiverse,” JCAP 0901 (2009) 021,
arXiv:0809.4257 [hep-th].

[34] S. W. Hawking and G. F. R. Ellis, The large scale stucture of space-time. Cambridge
University Press, Cambridge, England, 1973.

[35] L. Susskind, L. Thorlacius, and J. Uglum, “The Stretched horizon and black hole
complementarity,” Phys. Rev. D 48 (1993) 3743, hep-th/9306069.

[36] R. Bousso, B. Freivogel, and I.-S. Yang, “Eternal inflation: The inside story,” Phys.
Rev. D 74 (2006) 103516, hep-th/0606114.

– 22 –