On the dimensionality of spacetime

Max Tegmark
Institute for Advanced Study, Olden Lane, Princeton, NJ 08540; [email protected]

likely than faint ones to be sampled [3]. Below we will ar-

Some superstring theories have more than one effective
gue that if observers can only exist in a world exhibiting a
low-energy limit, corresponding to classical spacetimes with
certain minimum complexity, predictability and stability,
different dimensionalities. We argue that all but the 3+1-
dimensional one might correspond to “dead worlds”, devoid then all such ensemble theories may actually predict that
of observers, in which case all such ensemble theories would we should find ourselves inhabiting a 3+1-dimensional
actually predict that we should find ourselves inhabiting a spacetime with 100% certainty, as illustrated in Figure 1,
3+1-dimensional spacetime. With more or less than one and that the Bayesean prior probabilities of quantum-
time-dimension, the partial differential equations of nature mechanical origin are completely irrelevant. We will first
arXiv:gr-qc/9702052v2 5 Apr 1997

would lack the hyperbolicity property that enables observers review some old but poorly known results regarding the
to make predictions. In a space with more than three di- number of spatial dimensions (when m = 1), then present
mensions, there can be no traditional atoms and perhaps no some new arguments regarding the number of time di-
stable structures. A space with less than three dimensions al- mensions. In both cases, we are not attempting to rigor-
lows no gravitational force and may be too simple and barren ously show that merely (n, m) = (3, 1) permits observers.
to contain observers.
Rather, we are simply arguing that it is far from obvi-
ous that any other (n, m) permits observers, since radical
qualitative changes occur in all cases, so that the burden
of proof of the contrary falls on the person wishing to
criticize ensemble theories with fine-tuning arguments.
I. INTRODUCTION

Many superstring theories have several stable (or ex-

tremely long-lived) states that constitute different effec-
tive low-energy theories with different spacetime dimen-
sionalities, corresponding to different compactifications
of the many (e.g., 11 or 26) dimensions of the fundamen-
tal manifold. Since the tunneling probabilities between
these states are negligible, such a theory for all prac-
tical purposes predicts an ensemble of classical n + m-
dimensional spacetimes, and the prediction for the di-
mensionality takes the form of a probability distribution
over n and m [1]. There are also inflationary models
predicting a Universe consisting of parts of exponentially
large size having different dimensionality [2]. In this pa-
per, we argue that this failure to make the unique predic-
tion (n, m) = (3, 1) is not a weakness of such theories, but
a strength. To compute the theoretically predicted prob-
ability distribution for the dimensionality of our space-
time1 , we clearly need to take into account the selection
effect arising from the fact that some of these states are
more likely than others to contain self-aware observers
such as us. This is completely analogous to the famil- FIG. 1. When the partial differential equations of na-
iar selection effect in cosmological galaxy surveys, where ture are elliptic or ultrahyperbolic, physics has no predictive
we must take into account that bright galaxies are more power for an observer. In the remaining (hyperbolic) cases,
n > 3 may fail on the stability requirement (atoms are unsta-
ble) and n < 3 may fail on the complexity requirement (no
gravitational attraction, topological problems).
1
Here and thoughout, we let n and m refer to the number of
non-compactified space and time dimensions, or more gener-
ally to the effective spacetime dimensionality that is relevant
to the low-energy physics we will be discussing.

Published in Classical and Quantum Gravity, 14, L69-L75 (1997).

 ble over time, and thus probably cannot contain stable
observers.
What about n < 3? It has been argued [7] that organ-
isms would face insurmountable topological problems if
n = 2: for instance, two nerves cannot cross. Another
problem, emphasized by Wheeler [8], is the well-known
fact (see e.g. [9]) that there is no gravitational force in
General Relativity with n < 3. We will not spend more
time listing problems with n < 3, but simply conjecture
that since n = 2 (let alone n = 1 and n = 0) offers vastly
less complexity than n = 3, worlds with n < 3 are just
too simple and barren to contain observers.

III. WHY IS TIME ONE-DIMENSIONAL?

FIG. 2. The two body problem in four-dimensional space:
the light particles that approach the heavy one at the cen-
ter either escape to infinity or get sucked into a cataclysmic In this section, we will present an argument for why a
collision. There are no stable orbits. world with the same laws of physics as ours and with an
n + m-dimensional spacetime can only contain observers
if the number of time-dimensions m = 1, regardless of
II. WHY IS SPACE THREE-DIMENSIONAL? the number of space-dimensions n. Before describing
this argument, which involves hyperbolicity properties of
As was pointed out by Ehrenfest back in 1917 [4], nei- partial differential equations, let us make a few general
ther classical atoms nor planetary orbits can be stable comments about the dimensionality of time.
in a space with n > 3, and traditional quantum atoms What would reality appear like to an observer in a
cannot be stable either [5]. These properties are re- manifold with more than one time-like dimension? Even
lated to the fact that the fundamental Green function of when m > 1, there is no obvious reason for why an ob-
the Poisson equation ∇2 φ = ρ, which gives the electro- server could not nonetheless perceive time as being one-
static/gravitational potential of a point particle, is r2−n dimensional, thereby maintaining the pattern of having
for n > 2. Thus the inverse square law of electrostatics “thoughts” in a one-dimensional succession that char-
and gravity becomes an inverse cube law if n = 4, etc. acterizes our own reality perception. If the observer is
When n > 3, the two-body problem no longer has any a localized object, it will travel along an essentially 1-
stable orbits as solutions [6]. This is illustrated in Fig- dimensional (time-like) world line through the n + m-
ure 2, where a swarm of light test particles are incident dimensional spacetime manifold. The standard General
from the left on a massive point particle (the black dot), Relativity notion of its proper time is perfectly well-
all with the same momentum vector but with a range defined, and we would expect this to be the time that
of impact parameters. There are two cases: those that it would measure if it had a clock and that it would sub-
start outside the shaded region escape to infinity, whereas jectively experience.
those with smaller impact parameters spiral into a sin- Needless to say, many aspects of the world would
gular collision in a finite time. We can think of this as nonetheless appear quite different. For instance, a re-
there being a finite cross section for annihilation. This is derivation of relativistic mechanics for this more general
of course in stark contrast to the familiar case n = 3, case shows that energy now becomes an m-dimensional
which gives either stable elliptic orbits or non-bound vector rather than a constant, whose direction determines
parabolic and hyperbolic orbits, and has no “annihila- in which of the many time-directions the world-line will
tion solutions” except for the measure zero case where continue, and in the non-relativistic limit, this direction
the impact parameter is exactly zero. A similar disas- is a constant of motion. In other words, if two non-
ter occurs in quantum mechanics, where a study of the relativistic observers that are moving in different time di-
Schrödinger equation shows that the Hydrogen atom has rections happen to meet at a point in spacetime, they will
no bound states for n > 3 [5]. Again, there is a finite inevitably drift apart in separate time-directions again,
annihilation cross section, which is reflected by the fact unable to stay together.
that the Hydrogen atom has no ground state, but time- Another interesting difference, which can be shown by
dependent states of arbitrarily negative energy. The sit- an elegant geometrical argument [10], is that particles
uation in General relativity is analogous [5]. Modulo the become less stable when m > 1. For a particle to be
important caveats in the discussion section, this means able to decay when m = 1, it is not sufficient that there
that such a world cannot contain any objects that are sta- exists a set of particles with the same quantum numbers.
It is also necessary, as is well-known, that the sum of

2
their rest masses should be less than the rest mass of where the matrix A (which we without loss of generality
the original particle, regardless of how great its kinetic can take to be symmetric), the vector b and the scalar c
energy may be. When m > 1, this constraint vanishes are given differentiable functions of the d coordinates, it
[10]. For instance, is customary to classify it depending on the signs of the
eigenvalues of A. The PDE is said to be
• a proton can decay into a neutron, a positron and
a neutrino, • elliptic in some region of Rd if they are all positive
or all negative there,
• an electron can decay into a neutron, an antiproton
and a neutrino, and • hyperbolic if one is positive and the rest are negative
(or vice versa), and
• a photon of sufficiently high energy can decay into
any particle and its antiparticle. • ultrahyperbolic in the remaining case, i.e., where at
least two eigenvalues are positive and at least two
In addition to these two differences, one can concoct
are negative.
seemingly strange occurrences involving “backward cau-
sation” when m > 1. Nonetheless, although such un- What does this have to do with the dimensionality of
familiar behavior may appear disturbing, it would seem spacetime? For the various covariant field equations of
unwarranted to assume that it would prevent any form nature that describe our world (the wave equation u;µµ =
of observer from existing. After all, we must avoid the 0, the Klein-Gordon equation u;µµ + m2 u = 0, etc.3 ), the
fallacy of assuming that the design of our human bod- matrix A will clearly have the same eigenvalues as the
ies is the only one that allows self-awareness. Electrons, metric tensor. For instance, they will be hyperbolic in a
protons and photons would still be stable if their kinetic metric of signature (+ − −−), corresponding to (n, m) =
energies were low enough, so perhaps observers could still (3, 1), elliptic in a metric of signature (+ + + + +), and
exist in rather cold regions of a world with m > 12 . ultrahyperbolic in a metric of signature (+ + −−).
There is, however, an additional problem for observers One of the merits of this standard classification of
when m > 1, which has not been previously emphasized PDEs is that it determines their causal structure, i.e.,
even though the mathematical results on which it is based how the boundary conditions must be specified to make
are well-known. If an observer is to be able to make any the problem well-posed. Roughly speaking, the problem
use of its self-awareness and information-processing abil- is said to be well-posed if the boundary conditions de-
ities, the laws of physics must be such that it can make termine a unique solution u and if the dependence of
at least some predictions. Specifically, within the frame- this solution on the boundary data (which will always
work of a field theory, it should by measuring various be linear) is bounded. The last requirement means that
nearby field values be able to compute field values at the solution u at a given point will only change by a fi-
some more distant space-time points (ones lying along nite amount if the boundary data is changed by a finite
its future world-line being particularly useful) with non- amount. Therefore, even if an ill-posed problem can be
infinite error bars. If this type of well-posed causality formally solved, this solution would in practice be useless
were absent, then not only would there be no reason for to an observer, since it would need to measure the initial
observers to be self-aware, but it would appear highly un- data with infinite accuracy to be able to place finite error
likely that information processing systems (such as com- bars on the solution (any measurement error would cause
puters and brains) could exist at all. the error bars on the solution to be infinite).
Although this predictability requirement may sound Elliptic equations allow well-posed boundary value
modest, it is in fact only met by a small class of partial problems. On the other hand, giving “initial” data for an
differential equations (PDEs), essentially those which are elliptic PDE on a non-closed hypersurface, say a plane,
hyperbolic. We will now discuss the classification and is an ill-posed problem. This means that an observer in
causal structure of PDEs in some detail. This mathe- a world with no time dimensions (m=0) would not be
matical material is well-known, and can be found in more able do make any inferences at all about the situation in
detail in [12]. Given an arbitrary second order linear par- other parts of space based on what it observes locally.
tial differential equation in Rd ,
 
d X d d
X ∂ ∂ X ∂
 Aij + bi + c u = 0, 3
Our discussion will apply to matter fields with spin as well,
i=1 j=1
∂x i ∂x j i=1
∂x i
e.g. fermions and photons, since spin does not alter the causal
structure of the solutions. For instance, all four components
of an electron-positron field obeying the Dirac equation satisfy
the Klein-Gordon equation as well, and all four components of
2
It is, however, far from trivial to formulate a quantum field the electromagnetic vector potential in Lorentz gauge satisfy
theory with a stable vacuum state when m > 1 [11]. the wave equation.

3
 Asgeirsson’s theorem applies to the ultrahyperbolic
case as well, showing that initial data on a hypersurface
containing both spacelike and timelike directions leads
to an ill-posed problem. However, since a hypersurface
by definition has a dimensionality which is one less than
that of the spacetime manifold (data on a submanifold
of lower dimensionality can never give a well-posed prob-
lem), there are no spacelike or timelike hypersurfaces in
the ultrahyperbolic case, and hence no well-posed prob-
lems. 4
Since a mere minus sign distinguishes space from time,
the remaining case (n, m) = (1, 3) is mathematically
equivalent to the case where (n, m) = (3, 1) and all par-
ticles are tachyons [14] with imaginary rest mass. Also
in this case, an observer would be unable to make any
predictions, since as described in more detail in [15], well-
FIG. 3. The causality structure for hyperbolic and ul-
tra-hyperbolic equations. posed problems require data to be specified in the non-
local region outside the lightcones.
Hyperbolic equations, on the other hand, allow well- Above we discussed only linear PDEs, although the full
posed initial-value problems. For example, specifying ini- system of coupled PDEs of nature is of course non-linear.
tial data (u and u̇) for the Klein-Gordon equation on the This in no way weakens our conclusions about only m = 1
shaded disc in Figure 3 determines the solution in the vol- giving well-posed initial value problems. When PDEs
umes bounded by the two cones, including the (missing) give ill-posed problems even locally, in a small neighbor-
tips. A localized observer can therefore make predictions hood of a hypersurface (where we can generically approx-
about its future. If the matter under consideration is of imate the nonlinear PDEs with linear ones), it is obvious
such low temperature that it is nonrelativistic, then the that no nonlinear terms can make them well-posed in a
fields will essentially contain only Fourier modes with larger neighborhood.
wave numbers |k| ≪ m, which means that for all practi-
cal purposes, the solution at a point is determined by the
IV. DISCUSSION
initial data in a “causality cone” with an opening angle
much narrower than 45◦ .
In contrast, if the initial data for a hyperbolic PDE Our conclusions are graphically illustrated in Figure 1:
is specified on a hypersurface that is not spacelike, the given the other laws of physics, it is not implausible that
problem becomes ill-posed. Figure 3, which is based on only a 3+1-dimensional spacetime can contain observers
[12], provides an intuitive understanding of what goes that are complex and stable enough to be able to under-
wrong. A corollary of a remarkable theorem by Asgeirs- stand and predict their world to any extent at all, for the
son [13] is that if we specify u in the cylinder in Fig- following reasons.
ure 3, then this determines u throughout the region made • More or less than 1 time dimension: insufficient
up of the truncated double cones. Letting the radius predictability.
of this cylinder approach zero, we obtain the disturb-
ing conclusion that providing data in a for all practical • More than 3 space dimensions: insufficient stabil-
purposes one-dimensional region determines the solution ity.
in a three-dimensional region. Such an apparent “free
• Less than 3 space dimensions: insufficient complex-
lunch”, where the solution seems to contain more infor-
ity.
mation than the input data, is a classical symptom of
ill-posedness. The price that must be paid is specifying Thus although application of the so-called weak an-
the input data with infinite accuracy, which is of course thropic principle [16] does in general not appear to give
impossible given real-world measurement errors. Clearly,
generic boundary data allows no solution at all, since it is
not self-consistent. It is easy to see that the same applies
when specifying “initial” data on part of a non-spacelike 4
The only remaining possibility is the rather contrived case
hypersurface, e.g., that given by y = 0. These proper- where data is specified on a null hypersurface. To measure
ties are analogous in n+ 1-dimensions, and illustrate why such data, an observer would need to “live on the light cone”,
an observer in an n + 1-dimensional spacetime can only i.e., travel with the speed of light, which means that it would
make predictions in time-like directions. subjectively not perceive any time at all (its proper time
would stand still).

4
very strong predictions for physical constants [17], its di- [13] L. Asgeirsson, Math. Ann. 113, 321 (1936).
mensionality predictions may indeed turn out to give the [14] G. Feinberg, Phys. Rev. 159, 1089 (1967).
narrowest probability distribution possible. Viewed in [15] M. Tegmark, preprint (1996).
this light, the multiple dimensionality prediction of some [16] B. Carter, in IAU Symposium 63, Ed. S. Longair (Reidel,
superstring theories is a strength rather than a weak- Dordrecht, 1974).
J. D. Barrow & F. J. Tipler, The Anthropic Cosmological
ness, since it eliminates the otherwise embarrassing dis-
Principle (Clarendon, Oxford, 1986).
crete fine-tuning problem of having to explain the “lucky Y. V. Balashov, Am. J. Phys. 59, 1069 (1991).
coincidence” that the compactification mechanism itself [17] G. Greenstein & A. Kropf, Am. J. Phys 57, 746 (1989).
happened to single out only a 3+1-dimensional space-
time.
Needless to say, we have not attempted to rigorously
demonstrate that observers are impossible for other di-
mensionalities. For instance, within the context of spe-
cific models, one might consider exploring the possibility
of stable structures in the case (n, m) = (4, 1) based on
short distance quantum corrections to the 1/r2 poten-
tial or on string-like (rather than point-like) particles.
We have simply argued that it is far from obvious that
any other combination than (n, m) = (3, 1) permits ob-
servers, since radical qualitative changes occur when n
or m are altered. For this reason, a theory cannot be
criticized for failing to predict a definitive spacetime di-
mensionality until the stability and predictability issues
raised here have been carefully analyzed.

ACKNOWLEDGMENTS

The author wishes to thank Andreas Albrecht, Di-

eter Maison, Harold Shapiro, John A. Wheeler, Frank
Wilczek and Edward Witten for stimulating discussions
on some of the above-mentioned topics.

[1] A. Albrecht, in The Birth of the Universe and Funda-

mental Forces, ed. F. Occionero (Springer, Berlin, 1994).
[2] A. D. Linde & M. I. Zelnikov, Phys. Lett. B, 215, 59
(1988).
[3] A. Vilenkin, Phys. Rev. Lett., 74, 846 (1995)
[4] P. Ehrenfest, Proc. Amsterdam Acad. 20, 200 (1917).
P. Ehrenfest, Ann. Physik 61, 440 (1920).
[5] F. R. Tangherlini, Nuovo Cim. 27, 636 (1963).
[6] W. Büchel, Physikalische Blätter 19, 547 (1963).
I. M. Freeman, Am. J. Phys. 37, 1222 (1969).
[7] G. J. Whitrow, Br. J. Phil. 6, 13 (1955).
[8] C. W. Misner, K. S. Thorne, & J. A. Wheeler, Gravita-
tion (Freeman, San Francisco, 1973), page 1205.
[9] S. Deser, R. Jackiw & G. T’Hooft, Ann. Physics 152, 220
(1984).
[10] J. Dorling, Am. J. Phys. 38, 539 (1969).
[11] A. D. Linde, Particle Physics and Inflationary Cosmology
(Harwood, Chur, Switzerland, 1990) p. 318.
[12] R. Courant & D. Hilbert, Methods of Mathematical
Physics (Interscience, New York, 1962).