PHYSICAL REVIEW D 72, 084013 (2005)

Information loss in black holes

S. W. Hawking
DAMTP, Center for Mathematical Sciences, University of Cambridge, Wilberforce Road, Cambridge CB3 0WA, United Kingdom
(Received 22 August 2005; published 18 October 2005)
The question of whether information is lost in black holes is investigated using Euclidean path integrals.
The formation and evaporation of black holes is regarded as a scattering problem with all measurements
being made at infinity. This seems to be well formulated only in asymptotically AdS spacetimes. The path
integral over metrics with trivial topology is unitary and information preserving. On the other hand, the
path integral over metrics with nontrivial topologies leads to correlation functions that decay to zero. Thus
at late times only the unitary information preserving path integrals over trivial topologies will contribute.
Elementary quantum gravity interactions do not lose information or quantum coherence.

DOI: 10.1103/PhysRevD.72.084013 PACS numbers: 04.70.Dy

I. INTRODUCTION mation could get out of a black hole. It is this question I

will address here.
The black hole information paradox started in 1967
when Werner Israel showed that the Schwarzschild metric II. EUCLIDEAN QUANTUM GRAVITY
was the only static vacuum black hole solution [1]. This
was then generalized to the no hair theorem; the only Black hole formation and evaporation can be thought of
stationary rotating black hole solutions of the Einstein- as a scattering process. One sends in particles and radiation
Maxwell equations are the Kerr-Newman metrics [2]. from infinity and measures what comes back out to infinity.
The no hair theorem implied that all information about All measurements are made at infinity, where fields are
the collapsing body was lost from the outside region apart weak and one never probes the strong field region in the
from three conserved quantities: the mass, the angular middle. So one cannot be sure a black hole forms, no
momentum, and the electric charge. matter how certain it might be in classical theory. I shall
This loss of information was not a problem in the show that this possibility allows information to be pre-
classical theory. A classical black hole would last forever served and to be returned to infinity.
and the information could be thought of as preserved inside I adopt the Euclidean approach [6]—the only sane way
it but just not very accessible. However, the situation to do quantum gravity nonperturbatively. One might think
changed when I discovered that quantum effects would one should calculate the time evolution of the initial state
cause a black hole to radiate at a steady rate [3]. At least by doing a path integral over all positive definite metrics
in the approximation I was using the radiation from the that go between two surfaces that are a distance T apart at
black hole would be completely thermal and would carry infinity. One would then Wick rotate the time interval T to
no information [4]. So what would happen to all that the Lorentzian.
information locked inside a black hole that evaporated The trouble with this is that the quantum state for the
away and disappeared completely? It seemed the only gravitational field on an initial or final spacelike surface is
way the information could come out would be if the described by a wave function which is a functional of the
radiation was not exactly thermal but had subtle correla- geometries of spacelike surfaces and the matter fields
tions. No one has found a mechanism to produce correla-

hij ;
; t; (1)
tions but most physicists believe one must exist. If
information were lost in black holes, pure quantum states where hij is the three metric of the surface,
stands for the
would decay into mixed states and quantum gravity would matter fields, and t is the time at infinity. However there is
not be unitary. no gauge invariant way in which one can specify the time
I first raised the question of information loss in 1975 and position of the surface in the interior. This means one
the argument continued for years without any resolution cannot give the initial wave function without already
either way. Finally, it was claimed that the issue was settled knowing the entire time evolution.
in favor of conservation of information by ADS-CFT. One can measure the weak gravitational fields on a
ADS-CFT is a conjectured duality between string theory timelike tube around the system but not on the caps at
in anti –de Sitter space and a conformal field theory on the top and bottom which go through the interior of the system
boundary of anti –de Sitter space at infinity [5]. Since the where the fields may be strong. One way of getting rid of
conformal field theory is manifestly unitary the argument the difficulties of caps would be to join the final surface
is that string theory must be information preserving. Any back to the initial surface and integrate over all spatial
information that falls in a black hole in anti –de Sitter space geometries of the join. If this was an identification under a
must come out again. But it still was not clear how infor- Lorentzian time interval T at infinity, it would introduce

1550-7998= 2005=72(8)=084013(4)$23.00 084013-1 © 2005 The American Physical Society

S. W. HAWKING PHYSICAL REVIEW D 72, 084013 (2005)
closed time like curves. But if the interval at infinity is the conserved global charges on such a three cycle. These
Euclidean distance  the path integral gives the partition would prevent correlation functions from decaying in to-
function for gravity at temperature   1 ; pologically trivial metrics. Indeed, one can regard the
unitary Hamiltonian evolution of a topologically trivial
Z metric as a global conservation of information flowing
Z  DgD
eI
g;
  TreH : (2)
through a three cycle under a global time translation. On
the other hand, nontrivial black hole topologies will not
There is an infrared problem with this idea for asymp- have any conserved quantity that will prevent correlation
totically flat space. The partition function is infinite be- functions from decaying. It is therefore very plausible that
cause the volume of space is infinite. This problem can be the path integral over a topologically nontrivial metric
solved by adding a small negative cosmological constant  gives correlation functions that decay to zero at late
which makes the effective volume of the space the order of Lorentzian times. This is born out by explicit calculations.
3=2 . It will not affect the evaporation of a small black The correlation functions decay as more and more of the
hole but it will change infinity to anti –de Sitter space and wave falls through the horizon into the black hole.
make the thermal partition function finite.
It seems that asymptotically anti –de Sitter space is the
IV. GIANT BLACK HOLES
only arena in which particle scattering in quantum gravity
is well formulated. Particle scattering in asymptotically flat In a thought provoking paper [7], Maldacena considered
space would involve null infinity and Lorentzian metrics, how the loss of information into black holes in AdS could
but there are problems with nonzero mass fields, horizons, be reconciled with the unitarity of the CFT on the boundary
and singularities. Because measurements can be made only of AdS. He studied the canonical ensemble for AdS at
at spatial infinity, one can never be sure if a black hole is temperature 1 . This is given by the path integral over
present or not. all metrics that fit inside the boundary S1  S2 where the
radius of the S1 is  times the radius of the S2 . For  
there are three classical solutions that fit inside the bound-
III. THE PATH INTEGRAL
ary: periodically identified AdS, a small black hole and a
The boundary at infinity has topology S1  S2 . The path giant black hole. If one normalizes AdS to have zero
integral that gives the partition function is taken over action, small black holes have positive action and giant
metrics of all topologies that fit inside this boundary. The black holes have very large negative action. They therefore
simplest topology is the trivial topology S1  D3 where D3 dominate the canonical ensemble but the other solutions
is the three disk. The next simplest topology and the first are important.
nontrivial topology is S2  D2 . This is the topology of the Maldacena considered two point correlation functions in
Schwarzschild anti –de Sitter metric. There are other pos- the CFT on the boundary of AdS. The vacuum expectation
sible topologies that fit inside the boundary but these two value hOxOyi can be thought of as the response at y to
are the important cases, topologically trivial metrics and disturbances at x corresponding to the insertion of the
the black hole. The black hole is eternal: it cannot become operator O. It would be difficult to compute in a strongly
topologically trivial at late times. coupled CFT but by AdS-CFT it is given by boundary to
The trivial topology can be foliated by a family of boundary Green functions on the AdS side which can be
surfaces of constant time. The path integral over all metrics computed easily.
with trivial topology can be treated canonically by time The Green functions in the dominant giant black hole
slicing. The argument is the same as for the path integral solution have the standard form for small separation be-
for ordinary quantum fields in flat space. One divides the tween x and y but decay exponentially as y goes to late
time interval T into time steps t. In each time step one times and most of the effect of the disturbance at x falls
makes a linear interpolation of the fields qi and their through the horizon of the black hole. This looks very like
conjugate momenta between their values on successive information loss into the black hole. On the CFT side it
time steps. This method applies equally well to topologi- corresponds to screening of the correlation function
cally trivial quantum gravity and shows that the time whereby the memory of the disturbance at x is washed
evolution (including gravity) will be generated by a out by repeated scattering.
Hamiltonian. This will give a unitary mapping between However the CFT is unitary, so theoretically it must be
quantum states on surfaces separated by a time interval T at possible to compute its evolution exactly and detect the
infinity. disturbance at late times from the many point correlation
This argument cannot be applied to the nontrivial black function. All Green functions in the black hole metrics will
hole topologies. They cannot be foliated by a family of decay exponentially to zero but Maldacena realized that
surfaces of constant time because they do not have any the Green functions in periodically identified AdS do not
spatial cross sections that are a three cycle, modulo the decay and have the right order of magnitude to be compat-
boundary at infinity. Any global symmetry would lead to ible with unitarity. In this paper I have gone further and

084013-2
INFORMATION LOSS IN BLACK HOLES PHYSICAL REVIEW D 72, 084013 (2005)
E0
shown that the path integral over topologically trivial e . This projects out the states with energy E0 ;
metrics like periodically identified AdS is unitary. Z
i1
So in the end everyone was right in a way. Information is ZE0   dZeE0 : (3)
i1
lost in topologically nontrivial metrics like black holes.
This corresponds to dissipation in which one loses sight of For E0 1=2 most of these states will correspond to
the exact state. On the other hand, information about the thermal radiation in AdS which acts like a confining box of
exact state is preserved in topologically trivial metrics. The volume 3=2 . However, there will be thermal fluctua-
confusion and paradox arose because people thought clas- tions which occasionally will be large enough to cause
sically in terms of a single topology for spacetime. It was gravitational collapse to form a small black hole. This
either R4 or a black hole. But the Feynman sum over black hole will evaporate back to thermal AdS. If one
histories allows it to be both at once. One cannot tell which now considers correlation functions on the boundary of
topology contributed to the observation, any more than one AdS, one again finds that there is apparent information loss
can tell which slit the electron went through in the two slits in the small black hole solution but in fact information is
experiment. All that observation at infinity can determine preserved by topologically trivial geometries. Another way
is that there is a unitary mapping from initial states to final of seeing that information is preserved in the formation and
and that information is not lost. evaporation of small black holes is that the entropy in the
box does not increase steadily with time as it would if
information were lost each time a small black hole formed
V. SMALL BLACK HOLES and evaporated.
Giant black holes are stable and will not evaporate away.
However, small black holes are unstable and behave like VI. CONCLUSIONS
black holes in asymptotically flat space if M 1=2
[8]. However, in the approach I am using, one can not just In this paper, I have argued that quantum gravity is
set up a small black hole and watch it evaporate. All one unitary and information is preserved in black hole forma-
tion and evaporation. I assume the evolution is given by a
can do is to consider correlation functions of operators at
Euclidean path integral over metrics of all topologies. The
infinity. One can apply a large number of operators at
integral over topologically trivial metrics can be done by
infinity, weighted with time functions, that in the classical
dividing the time interval into thin slices and using a linear
limit would create a spherical ingoing wave from infinity,
interpolation to the metric in each slice. The integral over
that in the classical theory would form a small black hole.
each slice will be unitary and so the whole path integral
This would presumably then evaporate away.
will be unitary.
For years, I tried to think of a Euclidean geometry that
On the other hand, the path integral over topologically
could represent the formation and evaporation of a single nontrivial metrics will lose information and will be asymp-
black hole, but without success. I now realize there is no totically independent of its initial conditions. Thus the total
such geometry, only the eternal black hole, and pair crea- path integral will be unitary and quantum mechanics is
tion of black holes, followed by their annihilation. The pair safe.
creation case is instructive. The Euclidean geometry can be How does information get out of a black hole? My work
regarded as a black hole moving on a closed loop, as one with Hartle [9] showed the radiation could be thought of as
would expect. However, the corresponding Lorentzian ge- tunnelling out from inside the black hole. It was therefore
ometry, represents two black holes that come in from not unreasonable to suppose that it could carry information
infinity in the infinite past, and accelerate away from out of the black hole. This explains how a black hole can
each other for ever. The moral of this is that one should form and then give out the information about what is inside
not take the Lorentzian analytic continuation of a it while remaining topologically trivial. There is no baby
Euclidean geometry literally as a guide to what an observer universe branching off, as I once thought. The information
would see. Similarly, the formation and evaporation of a remains firmly in our universe. I am sorry to disappoint
small black hole, and the subsequent formation of small science fiction fans, but if information is preserved, there is
black holes from the thermal radiation, should be repre- no possibility of using black holes to travel to other uni-
sented by a superposition of trivial metrics and eternal verses. If you jump into a black hole, your mass energy will
black holes. The probability of observing a small black be returned to our universe but in a mangled form which
hole, at a given time, is given by the difference between the contains the information about what you were like but in a
actions. A similar discussion of correlation functions on state where it can not be easily recognized. It is like
the boundary shows that the topologically trivial metrics burning an encyclopedia. Information is not lost, if one
make black hole formation and evaporation unitary and keeps the smoke and the ashes. But it is difficult to read. In
information preserving. One can restrict to small black practice, it would be too difficult to rebuild a macroscopic
holes by integrating the path integral over  along a object like an encyclopedia that fell inside a black hole
contour parallel to the imaginary axis with the factor from information in the radiation, but the information

084013-3
S. W. HAWKING PHYSICAL REVIEW D 72, 084013 (2005)
preserving result is important for microscopic processes hair theorem. One cannot ask when the information
involving virtual black holes. If these had not been unitary, gets out of a black hole because that would require the
there would have been observable effects, like the decay of use of a semiclassical metric which has already lost the
baryons. information.
There is a problem describing what happens because In 1997, Kip Thorne and I bet John Preskill that infor-
strictly speaking, the only observables in quantum gravity mation was lost in black holes. The loser or losers of the bet
are the values of the field at infinity. One can not define the were to provide the winner or winners with an encyclope-
field at some point in the middle because there is quantum dia of their own choice, from which information can be
uncertainty in where the measurement is done. What is recovered with ease. I gave John an encyclopedia of base-
often done is to adopt the semiclassical approximation in ball, but maybe I should just have given him the ashes.
which one assumes that there are a large number N of light
matter fields coupled to gravity and that one can neglect the ACKNOWLEDGMENTS
gravitational fluctuations because they are only one among
I am very grateful to my student, Christophe Galfard
N quantum loops. However, in ignoring quantum loops,
[10], for help and discussions. He is working on a proof
one throws away unitarity. A semiclassical metric is in a
that correlation functions decay in topologically nontrivial
mixed state already. The information loss corresponds to
metrics.
the classical relaxation of black holes according to the no

[1] W. Israel, Phys. Rev. 164, 1776 (1967); W. Israel, (2000); and see E. Witten, Adv. Theor. Math. Phys. 2, 253
Commun. Math. Phys. 8, 245 (1968). (1998) for the holographic point of view.
[2] B. Carter, Phys. Rev. Lett. 26, 331 (1971); S. W. Hawking, [6] For an introduction, see Euclidean Quantum Gravity,
Commun. Math. Phys. 25, 152 (1972); D. C. Robinson, edited by G. W. Gibbons and S. W. Hawking, (World
Phys. Rev. D 10, 458 (1974); D. C. Robinson, Phys. Rev. Scientific, Singapore, 1993).
Lett. 34, 905 (1975). [7] J. Maldacena, J. High Energy Phys. 04 (2003) 21.
[3] S. W. Hawking, Commun. Math. Phys. 43, 199 (1975). [8] S. W. Hawking and D. N. Page, Commun. Math. Phys. 87,
[4] S. W. Hawking, Phys. Rev. D 14, 2460 (1976). 577 (1983); E. Witten, Adv. Theor. Math. Phys. 2, 505
[5] J. M. Maldacena, Adv. Theor. Math. Phys. 2, 231 (1998); (1998).
Int. J. Theor. Phys. 38, 1113 (1999); S. S. Gubser, I. R. [9] J. B. Hartle and S. W. Hawking, Phys. Rev. D 13, 2188
Klebanov, and A. M. Polyakov, Phys. Lett. B 428, 105 (1976).
(1998); and a classic review: O. Aharony, J. Maldacena, [10] Electronic address: [email protected]
S. Gubser, H. O. Oguri, and Y. Oz, Phys. Rep. 323, 183

084013-4