QUANTUM FIELD THEORY IN CURVED SPACETIME

STEFAN HOLLANDS AND ROBERT M. WALD

A BSTRACT. We review the theory of quantum fields propagating in an arbitrary, classical, globally
hyperbolic spacetime. Our review emphasizes the conceptual issues arising in the formulation of the
theory and presents known results in a mathematically precise way. Particular attention is paid to the
distributional nature of quantum fields, to their local and covariant character, and to microlocal spectrum
conditions satisfied by physically reasonable states. We review the Unruh and Hawking effects for
free fields, as well as the behavior of free fields in deSitter spacetime and FLRW spacetimes with an
exponential phase of expansion. We review how nonlinear observables of a free field, such as the
stress-energy tensor, are defined, as well as time-ordered-products. The “renormalization ambiguities”
involved in the definition of time-ordered products are fully characterized. Interacting fields are then
perturbatively constructed. Our main focus is on the theory of a scalar field, but a brief discussion
of gauge fields is included. We conclude with a brief discussion of a possible approach towards a
arXiv:1401.2026v1 [gr-qc] 9 Jan 2014

nonperturbative formulation of quantum field theory in curved spacetime and some remarks on the
formulation of quantum gravity.

1. T HE NATURE OF Q UANTUM F IELD T HEORY IN C URVED S PACETIME

Quantum field theory in curved spacetime (QFTCS) is the theory of quantum fields propagating in a
background, classical, curved spacetime (M , g). On account of its classical treatment of the metric,
QFTCS cannot be a fundamental theory of nature. However, QFTCS is expected to provide an ac-
curate description of quantum phenomena in a regime where the effects of curved spacetime may be
significant, but effects of quantum gravity itself may be neglected. In particular, it is expected that
QFTCS should be applicable to the description of quantum phenomena occurring in the early universe
and near (and inside of) black holes—provided that one does not attempt to describe phenomena oc-
curring so near to singularities that curvatures reach Planckian scales and the quantum nature of the
spacetime metric would have to be taken into account.
It should be possible to derive QFTCS by taking a suitable limit of a more fundamental theory wherein
the spacetime metric is treated in accord with the principles of quantum theory. However, this has
not been done—except in formal and/or heuristic ways—simply because no present quantum theory
of gravity has been developed to the point where such a well defined limit can be taken. Rather, the
framework of QFTCS that we shall describe in this review has been obtained by suitably merging
basic principles of classical general relativity with the basic principles of quantum field theory in
Minkowski spacetime. As we shall explain further below, the basic principles of classical general re-
lativity are relatively easy to identify and adhere to, but it is far less clear what to identify as the “basic
principles” of quantum field theory in Minkowski spacetime. Indeed, many of the concepts normally
viewed as fundamental to quantum field theory in Minkowski spacetime, such as Poincare invariance,
do not even make sense in the context of curved spacetime, and therefore cannot be considered as
“fundamental” from the viewpoint of QFTCS. By forcing one to re-think basic concepts, such as the
notions of “vacuum state” and “particles,” QFTCS has led to deep insights into the nature of quantum
field theory—and one may hope that it will provide significant guidance towards the development of
quantum gravity itself.

Date: 10th January 2014.

1
2 HOLLANDS AND WALD

The fundamental ideas upon which classical general relativity is based are that (i) all aspects of space-
time structure are described by the topological and differential (i.e., manifold) properties of events to-
gether with a Lorentz signature metric g, and (ii) the metric and matter fields1 are dynamical; further-
more their evolution is locally determined. More precisely, the metric and the tensor (and/or spinor)
fields describing matter satisfy partial differential equations—namely, Einstein’s equation together
with the equations of motion for the matter fields—that have a well posed initial value formulation,
so that these fields are uniquely determined (up to “gauge”) from their initial data within a suitable
domain of dependence. In particular, in classical general relativity, there is no non-dynamical, “back-
ground structure” in the laws of physics apart from the manifold structure of events. This lack of
background structure in classical general relativity is usually referred to as the “covariance” or “co-
ordinate invariance” of the theory; a “preferred set of coordinates” defined independently of the metric
would provide non-dynamical, background structure.
It is much more difficult to identify the fundamental ideas upon which quantum field theory in
Minkowski spacetime is based. One can attempt to formulate the quantum theory of a field in
Minkowski spacetime by decomposing the field into modes and applying the rules of quantum mech-
anics to each mode. For a free field, each mode is an independent harmonic oscillator and one can
obtain a mathematically sensible quantum field theory in this manner, although even here one encoun-
ters infinite expressions for quantities that are nonlinear in the fields. A well known example of this
general phenomenon is that one obtains an infinite expression for the total energy (and energy density)
of the field, as can be seen by adding the zero-point energies and/or energy densities of the infinite
number of modes. The situation is considerably worse for interacting (i.e., nonlinear) fields, wherein
one immediately encounters ill defined and/or infinite expressions in the calculation of essentially all
physical quantities, arising from the fact that modes of arbitrarily high energies seemingly contrib-
ute to low energy processes. Historically, it it appears to have been generally assumed in its earliest
days that the quantum field theory description of nature would break down at, say, the energy scale
of elementary particles, and there was no reason to presume that it was a mathematically consistent
theory. However, starting from the early 1950s, it was gradually understood how to give mathem-
atically consistent rules to produce well defined expressions for physical quantities to all orders in
perturbation theory for renormalizable theories such as quantum electrodynamics. This process cul-
minated in the works of Bogliubov, Parasiuk and of Hepp and Zimmermann, with important practical
improvements (dimensional regularization) being given later by ‘t Hooft and Veltmann. It was also
seen that the predictions of quantum field theory give truly excellent agreement with experiment—as
they have continued to do through the present, LHC era. In the 1950s and 1960s, major progress was
made toward putting quantum field on a mathematically sound footing via the development of the
axiomatic [67], algebraic [35], and constructive [34] approaches. Nevertheless, the prevailing attitude
toward quantum field theory today is not very different from what it was in its earliest days, namely,
that it is not a fundamental theory but merely a valid description of quantum field modes up to some
cutoff in energy (now assumed to be at a much higher energy scale than would have been assumed
in its earliest days). At the present time, relatively little attention is generally paid to the issue of
whether quantum field theory can be given a mathematically precise and consistent formulation—as
compared with such issues as the “fine tuning” that would be necessary to give small values to the
cosmological constant and Higgs mass if one views quantum field theory as the quantum theory of
the modes of fields lying below some energy cutoff.
Our view is that it is very important to determine if quantum field theory can be given a mathematically
precise and consistent formulation as a theory in its own right—and to provide such a formulation if
1One of the truly remarkable aspects of general relativity is that no new “matter field” need be introduced to describe
gravitation, i.e., all physical phenomena normally attributed to “gravity” are, in fact, described by g.
 QUANTUM FIELDS IN CURVED SPACETIME 3

it can be given. This is not because we believe that quantum field theory should be a “final” theory
of nature; indeed, we do not believe that a quantum theory of the spacetime metric can be formulated
within the existing framework of quantum field theory. However, even if quantum field theory has
only a limited domain of validity, it is important to understand precisely what questions are well
posed within its framework and how the answers to these questions are to be obtained. In this way,
the predictions of quantum field theory can be made with clarity and precision, and hints may be
provided for some of the features that might be expected to survive in a more fundamental theory that
supersedes quantum field theory.
What are the “basic principles” of quantum field theory in Minkowski spacetime? The observables
of the theory are the tensor fields representing the fundamental constituents of matter, together with
“composite fields,” such as the stress-energy tensor, derived from these matter fields. A key basic
principle of quantum field theory is that each observable field, Φ(x), at each spacetime point x should
be represented as an operator. These operators will satisfy nontrivial algebraic relations, such as com-
mutation relations. However, there are two important caveats to this statement that Φ(x) is represented
as an operator.
The first caveat is that, even in the case of a free Klein-Gordon scalar field φ —where, as already men-
tioned above, a quantum field theory can be formulated by ordinary quantization of the independent
modes—it can be seen that one cannot make mathematical sense of φ (x) as an operator at a sharply
defined point x, since modes of arbitrarily high frequency and short wavelength contribute to φ (x).
However, φ (x) does make sense as a distribution, i.e., by “averaging” φ (x) with a smooth function of
compact support, f (x), one effectively eliminates the arbitrarily high frequency and short wavelength
oscillations and thereby obtains a well defined expression for the quantum field. Thus, quantum fields
are operator-valued distributions. The distributional nature of quantum fields is the source of most,
if not all, of the mathematical difficulties arising in quantum field theory. Nonlinear operations in-
volving distributions are intrinsically ill defined, and one will typically get infinite answers if one
attempts to evaluate nonlinear functions of a distribution via mode expansions or other procedures.
The second caveat is that the word “operator” presumes that there is some unique underlying Hilbert
space of states on which this operator will act. However, even for a free field, there are an infinite
number of unitarily inequivalent representations of the fundamental commutation relations. (This
contrasts sharply with the situation for a quantum mechanical system with a finite number of degrees
of freedom, where the Stone-von Neuman theorem asserts that, under mild additional assumptions,
all such representations are unitarily equivalent.) In Minkowski spacetime, a preferred representa-
tion normally can be chosen based upon the additional requirement that the representation contain a
Poincare invariant state (“the vacuum”). However, no criterion analogous to this can be applied in
a general curved spacetime. As discussed further at the end of this section, it would therefore seem
much more natural to view the algebraic relations satisfied by the field observables—rather than the
choice of representation—to be fundamental. Thus, in quantum field theory, we will take as a basic
principle that the quantum fields Φ(x) are distributions valued in an algebra.
Another basic principle of quantum field theory in Minkowski spacetime that the fields should “trans-
form covariantly” under Poincare transformations. The Poincare group is the isometry group of the
metric, η, of Minkowski spacetime, but a general, curved spacetime will not admit any isometries.
Nevertheless, “Poincare invariance” may be viewed as a special relativistic version of the above gen-
eral relativistic requirement of “covariance,” i.e., that quantum field theory in curved spacetime be
constructed out of the classical spacetime metric g and the fundamental quantum fields φ , without
any additional “background structure.” Furthermore, this construction should be local in nature. We
4 HOLLANDS AND WALD

will take as a basic principle that quantum field theory should be locally and covariantly constructed.
We will give precise meaning to this statement in section 3.
An additional basic principle of quantum field theory in Minkowski spacetime is the requirement of
positivity of energy. Since the notion an energy operator for a quantum field theory in Minkowski
spacetime is normally defined in terms of the transformation properties of the field under time trans-
lations, this requirement cannot be straightforwardly generalized to QFTCS. Nevertheless, one can
formulate local conditions on the quantum field theory—known as microlocal spectral conditions—
that correspond to positivity of energy in Minkowski spacetime and make sense in curved spacetime.
We will take as a basic principle that the quantum field theory should satisfy suitable microlocal
spectral conditions. We will give precise meaning to this statement in Appendix A.
Finally, an additional principle of quantum field theory in MInkowski spacetime that is usually taken
to be fundamental is the existence of a unique, Poincare invariant state. However, this condition has
no analog in a general curved spacetime as a condition on existence or uniqueness of states2, and we
will not attempt to impose any condition of this nature.
Thus, we seek a formulation of quantum field theory in curved spacetime that implements the three
basic principles written in italics above. In the remainder of this section, we briefly describe some
of the standard approaches that have been used to formulate quantum field theory in Minkowski
spacetime and explain why they do not appear suitable for the formulation of QFTCS. We will then
describe the approach that we shall adopt.
Many discussions of quantum field theory are based upon a notion of “particles,” and focus almost
entirely on the calculation of the S-matrix, describing the scattering of particles. For a free field in
Minkowski spacetime, a notion of “vacuum state” and “particles” can be defined in a natural and
precise manner. If an interacting field behaves like a free field in the asymptotic past and future,
one can define asymptotic particle states. The S-matrix provides the relationship between the “in”
and “out” particle descriptions of the states, and thereby directly yields the dynamical information
about the interacting field that is most relevant to laboratory experiments in high energy physics.
However, the use of an S-matrix description as a fundamental ingredient in the formulation of QFTCS
is unsuitable for the following reasons. Although natural notions of “vacuum state” and “particles”
can be defined for a free field in stationary spacetimes, no such natural notions exist in a general
curved spacetime. The difficulty is not that a notion of “particles” cannot be defined at all in a general
curved spacetime but rather that many notions exist and none appears preferred. Although it may
be possible (and useful) to define an S-matrix in spacetimes that become asymptotically stationary in
a suitable manner in the past and future, many of the spacetimes of greatest interest in QFTCS are
cosmological spacetimes or spacetimes describing gravitational collapse, where singularities occur in
the asymptotic past and/or future. If one wishes to apply QFTCS to such spacetimes, it clearly would
be preferable to formulate it in a manner that does not require one to define a notion of “particles” near
singularities before one can even pose a well defined question. Furthermore, even if the spacetime
of interest is suitably asymptotically stationary in the past and future, many of the most interesting
physical questions are concerned with the local dynamical behavior of the fields at finite times rather
than the particle-like description of states at asymptotically early and late times. For example, one
may wish to know the expected stress-energy tensor of a quantum field in order to estimate the “back
reaction” effects of the quantum field on the dynamics of the spacetime. An S-matrix would not be
useful for such a calculation.

2As argued in [50], the existence of an operator product expansion may be viewed as a generalization of this condition
to curved spacetime; see subsection 4.1 below.
 QUANTUM FIELDS IN CURVED SPACETIME 5

In many discussions of quantum field theory in Minkowski spacetime, Euclidean methods play an
important role in both the formulation of the theory and in calculational techniques. Minkowski
spacetime can be viewed as a real 4-dimensional section of a complex 4-dimensional manifold with
complex metric, which contains another 4-real-dimensional section (of “imaginary time”) on which
the metric is positive definite. If one can define a quantum field theory in a suitable manner on this
“Euclidean section,” a quantum field theory on Minkowski spacetime can then be obtained via analytic
continuation. Since it is much easier to make sense of formal expressions in the Euclidean setting
than in the Lorentzian setting, Euclidean methods have been employed in most of the attempts to
rigorously define interacting quantum field theories and in most of the methods employed to regularize
and renormalize quantities in perturbative quantum field theory [34]. Euclidean methods can be
generalized so as to apply to static, curved spacetimes, where the transformation “t → it” takes one
from a static Lorentzian spacetime to a Riemannian space. However, a general curved spacetime
will not be a real section of a complex manifold that also contains a real section on which the metric
is Riemannian. Thus, although it should be possible to define “Euclidean quantum field theory” on
curved Riemannian spaces [57], there is no obvious way to connect such a theory with quantum field
theory on Lorentzian spacetimes. Thus, if one’s goal is to define quantum field theory on general
Lorentzian spacetimes, it does not appear fruitful to attempt to formulate the theory via a Euclidean
approach.
Finally, by far the most prevalent approach taken towards the formulation of quantum field theory in
Minkowski spacetime is to write down a formal functional integral expression for an effective action.
Suitable functional derivatives of this expression are then interpreted as providing the correlation
functions of the quantum field in its vacuum state. Thus, one will have defined the quantum field
theory if one can make sense of this functional integral and its functional derivatives. The difficulty
with using a functional integral approach to formulate QFTCS is that, in effect, it requires one to
single out a preferred state in order to define the theory—namely, the state for which the correlation
functions are being given. This is not a difficulty in Minkowski spacetime, where Poincare invariance
naturally selects a preferred state and, furthermore, Euclidean methods are available to make sense of
the functional integral for this preferred state. However, as previously indicated above, no analogous
notion of a preferred state exists in a general curved spacetime without symmetries. As in the above
discussion of S-matrix approaches to the formulation of QFTCS, we do not believe that it will be
fruitful to formulate QFTCS via an approach that requires one to define a preferred state in order to
define the theory3.
For the above reasons, we shall adopt the “algebraic viewpoint” for the formulation of QFTCS. The
basic idea of this approach is to take the relations satisfied by the quantum fields—such as commut-
ation relations and field equations—as the fundamental starting point of the theory. To define the
theory, one must specify the complete set of algebraic relations satisfied by the fundamental field and
composite fields. As we shall see, this can be done for a free field. In addition, we can naturally
define time ordered products, enabling one to give a perturbative construction of interacting quantum
field theory. However, it is far less clear as to how to define appropriate algebraic relations so as to
give a non-perturbative definition of interacting quantum field theory; we will return to this issue in
subsection 4.1. Once the algebraic relations have been given, states are defined to be positive linear
functions on the algebra of quantum fields. The GNS construction then shows that every state in this

3Note that this objection does not apply to the formulation of quantum field theory in (complete) Riemannian spaces,
where one has a unique Green’s functions for Laplacian-like operators, and, thus, a “preferred state" for a free field.
However, as discussed above, there does not appear to be any way of relating quantum field theory in Riemannian spaces
to quantum field theory on Lorentzian spacetimes.
6 HOLLANDS AND WALD

sense arises as a vector in a Hilbert space that carries a representation of the field algebra, thus con-
necting the algebraic notion of states with usual notions of states in quantum theory. The key point
is that one can formulate QFTCS via the algebraic approach in a manner that does not require one to
single out a preferred state in order to define the theory.
We begin in the next section by formulating QFTCS for a free scalar field, taking into account only
the fundamental field observables. We also discuss some key applications, including the Hawking
and Unruh effects and quantum effects arising from inflation. In section 3, we show how a wide class
of nonlinear field observables can be defined for a free scalar field. We then describe the perturbative
construction of QFTCS for interacting fields. In section 3.3, we discuss the construction of QFTCS
for gauge fields. Finally, in section 4 we discuss ideas that may lead towards a nonperturbative
formulation of interacting QFTCS and we make some remarks on the formulation of quantum gravity
from the perspective of QFTCS.

2. F REE QUANTUM FIELDS

In this section, we provide a precise formulation of the theory of a free Klein-Gordon field in curved
spacetime, insofar as the fundamental field observable, φ , is concerned. (Observables that are nonlin-
ear in φ —i.e., “composite fields”—will be introduced in the next section.) We then discuss some key
applications of the theory, namely, the Unruh effect, quantum field theory in deSitter spacetime, the
Hawking effect, and cosmological perturbations.

2.1. Formulation of linear QFTCS via the algebraic approach (without nonlinear observables).
We now describe how to define the quantum field theory of a real, linear Klein-Gordon field φ on a
d-dimensional, curved, Lorentzian spacetime (M , g) along the lines sketched in the previous section.
We begin with the classical Klein-Gordon field, which satisfies
(g − m2 )φ = 0 (1)
where g = gµν ∇µ ∇ν is the D’Alambertian operator associated with g. In order that (1) have a
well-posed initial value formulation, we restrict consideration—here and throughout this article—
to globally hyperbolic spacetimes. By definition, a globally hyperbolic spacetime is a time oriented
spacetime that possesses a “Cauchy surface,” Σ, i.e., a smoothly embedded (d −1)-dimensional space-
like submanifold with the property that if γ : R → M is any inextendible causal curve, then γ intersects
Σ precisely once. 4 The classical Klein-Gordon equation with source j
(g − m2 )φ = j (2)
(where j is an arbitrary, fixed, smooth function on spacetime) has a well posed initial value formu-
lation on globally hyperbolic spacetimes in the following sense. Let n denote the unit normal to
Σ. Then, given any pair ( f0 , f1 ) of smooth functions on Σ, there exists a unique solution φ to the
Klein-Gordon equation (2) such that
φ |Σ = f0 , nµ ∇µ φ |Σ = f1 . (3)
Furthermore, solutions to the initial value problem have a causal dependence upon the initial data and
the source in the sense that if x ∈ J + (Σ), then the solution φ to the above initial value problem will not
change at x if we change the initial data ( f0 , f1 ) outside J − (x) ∩ Σ, or if we change the source outside
of J − (x) ∩ J + (Σ). Here
J ± (S) ≡ {x ∈ M | ∃ causal future/past directed curve from y ∈ S to x}, (4)
4It is a theorem that any globally hyperbolic spacetime is topologically a direct product, M ∼ Σ × R.
=
 QUANTUM FIELDS IN CURVED SPACETIME 7

denotes the “causal future/past” of a set S ⊂ M . Similarly, if x ∈ J − (Σ), then φ (x) will not change
if we change the initial data ( f0 , f1 ) outside J + (x) ∩ Σ, or if we change the source outside of J + (x) ∩
J − (Σ). Finally, the solution φ depends continuously on ( f0 , f1 ) and j in a suitable sense.
We can define the retarded and advanced propagators E ± of the Klein-Gordon equation as follows.
For j ∈ C0∞ (M ), the advanced solution E − j is the unique solution to the Klein-Gordon equation with
source j such that E − j(x) = 0 for all x in some Cauchy surface Σ lying towards the future of supp j,
with the opposite definition for the retarded solution E + j. The propagators may be viewed as maps
E ± : C0∞ (M ) → C∞ (M ) , (5)
or alternatively as distributional kernels on M × M . As distributions, E±
satisfy the differential
equation
(g − m2 )E ± (x, y) = δ (x, y) , (6)
where the Klein-Gordon operator acts on the first variable, x, in the sense of distributions. The support
properties are
supp E ± ⊂ {(x, y) ∈ M × M | y ∈ J ± (x)}. (7)
The advanced and retarded propagators are related by exchanging x with y. The anti-symmetric
combination
E = E+ − E− (8)
5
is called “causal propagator” .
As described in the Introduction, we will formulate QFTCS for a Klein-Gordon field by defining a
suitable algebra, A (M , g) of quantum observables. In this section, we will consider only the algebra
of observables generated by the fundamental field φ . An enlarged algebra that includes “composite
fields” will be defined in the next section. The construction of A (M , g) will take into account: 1) the
distributional nature of the field φ (x), 2) the field equation, 3) the real character of φ and and 4) the
symplectic structure of the classical phase space of this theory. We construct A (M , g) by starting
with the free *-algebra generated by a unit 1 and elements φ ( f ), with f ∈ C0∞ (M ), and factoring by
the following relations:
(1) Linearity: φ (c1 f1 + c2 f2 ) = c1 φ ( f1 ) + c2 φ ( f2 ) for all c1 , c2 ∈ C.
(2) Field equation: φ ((g − m2 ) f ) = 0,
(3) Hermitian field: φ ( f )∗ = φ ( f¯),
(4) Commutator: [φ ( f1 ), φ ( f2 )] = iE( f1 , f2 )1.
Item 1) incorporates the distributional character of the field; informally we write
Z
φ( f ) = φ (x) f (x) dvg (9)
M
and we think of φ (x) as an A -valued distribution. We refer to the algebra A as “abstract,” because no
reference has been made to any representation. In fact, we are not supposed to think, a priori, of the
elements of A as operators on a particular Hilbert space, in the same way as an abstract Lie-algebra is
defined irrespective of a particular representation. 2) incorporates the field equation in a distributional
sense. 3) expresses that the field is real, and 4) implements the usual quantum mechanical relationship
between classical Poisson-brackets and commutators. It also incorporates Einstein causality, because
E(x, y) = 0 if x, y are spacelike related.
Rather than working with fields φ ( f ) that smeared with test functions f on M , we can also equi-
valently view φ as being “symplectically smeared” with solutions, F, having initial data of compact
5Further discussion concerning the construction and local expansion of E ± can e.g. be found in [7].
8 HOLLANDS AND WALD

support on some Cauchy surface Σ. The correspondence is the following. If f ∈ C0∞ (M ), then F = E f
is a source free solution having initial data (3) of compact support. Conversely, given a solution F with
initial data of compact support, there exists a test function f —unique up to addition of (g − m2 )h
for h ∈ C0∞ —such that F = E f . Defining φ [F] = φ ( f ) under this correspondence between F and f ,
we can then informally write the field as
Z
φ [F] = (F∇µ φ − (∇µ F)φ )nµ dS . (10)
Σ
We can think of π = nµ ∇µ φ |Σ , ϕ = φ |Σ as canonically conjugate variables and the commutation
relation (4) then corresponds to [π(x), ϕ(y)] = iδΣ (x, y)1; for details, see, e.g. [22] or lemma 3.2.1
of [71].
A physical state, ω, is simply an “expectation value functional,” i.e., a linear map ω : A (M , g) → C
satisfying the normalization condition ω(1) = 1, and positivity, ω(a∗ a) ≥ 0 for all a ∈ A . Any state
is, by construction, specified by the collection (Wn )n≥1 of its “n-point functions”,
Wn ( f1 , . . . , fn ) := ω(φ ( f1 ) · · · φ ( fn )) . (11)
The KG-equation, condition 1), is translated into the fact that Wn is a distributional solution in each
entry. Condition 2) implies that Wn is an n-times multilinear functional on C0∞ (M ), which one nor-
mally requires to be distributions (i.e. continuous in the appropriate sense). The commutator condition
4) is translated into a linear condition which in the simplest case n = 2 is
W2 (x, y) −W2 (y, x) = iE(x, y) . (12)
Positivity is translated into a rather complicated hierarchy of conditions on the multi-linear functionals
Wn , the simplest of which is
W2 ( f¯, f ) ≥ 0 for all f ∈ C0∞ . (13)
Given two states ω, ω 0 , one can form a new state by forming any convex linear combination λ ω +(1−
λ )ω 0 , where 0 ≤ λ ≤ 1. A state which cannot be written as a non-trivial convex linear combination
of others is called pure.
The notion of algebraic state is in principle sufficient to answer all physical questions about the field
observables. In particular, the specification of a state ω directly yields the expected values of all
powers of φ ( f ) for all, say real, test functions f . If ω is such that ω(φ ( f )n ) ≤ K n n! for some constant
K, it follows from the classical “Hamburger moment problem” that there is a unique probability
measure dν(λ ) such that
Z
λ n dν(λ ) = ω(φ ( f )...φ ( f )) for all n ≥ 1. (14)
R | {z }
n
The probability that an observation of φ ( f ) will yield a value within [a, b] when the field is in state ω
is then given by Z
Pa,b = λ dν(λ ) . (15)
[a,b]
Similar results hold for the probability distribution for φ ( f1 ) · · · φ ( fn ).
The relationship between states is the algebraic sense as defined above and the usual notion of states
as vectors in a Hilbert space can be seen as follows: First, assume that we have a representation
π : A (M , g) → H on a Hilbert space with dense invariant domain D ⊂ H . If Ψ is a non-zero
vector in this domain, then
(Ψ, π(a)Ψ)
ωΨ (a) = (16)
(Ψ, Ψ)
 QUANTUM FIELDS IN CURVED SPACETIME 9

defines a state in the algebraic sense. More generally, any sufficiently regular (with respect to D)
density matrix on H —i.e., a non-negative, trace-class operator ρ ∈ I1 (H )—defines an algebraic
state ωρ . Conversely, given any algebraic state ω, there is a simple construction—known as the
GNS-construction, or Wightman reconstruction argument—that yields a Hilbert space H , a rep-
resentation π of A on H , and a vector Ω ∈ H such that the algebraic state corresponding to Ω is
ω. As we shall explain further below, for the case of a Gaussian state, the GNS construction yields a
Fock representation of A , with Ω being the vacuum vector of this Fock space.
The above correspondences show that the algebraic and Hilbert space formulations of QFTCS are
essentially equivalent. However, there is one important difference: There are many unitarily inequi-
valent6 representations of the field algebra A . The Hilbert space formulation requires one to chose a
“preferred representation” at the outset, while the algebraic formulation does not. For this reason, we
feel that the formulation of QFTCS via the algebraic approach is conceptually superior.
The algebra A admits states with rather pathological properties of their n-point functions Wn . How-
ever, there is a natural criterion to select physically reasonable states that is motivated from a variety
of closely related considerations, including that (i) the ultra-high-frequency modes of the field should
be essentially in their ground state, (ii) the short distance singular structure of the n-point functions
Wn should be similar to that of the n-point functions of the vacuum state in Minkowski spacetime,
and (iii) the singular structure of the Wn ’s should be of “positive frequency type”. A mathematically
precise implementation of these requirements in a general curved spacetime can be formulated in the
language of wave front sets discussed in Appendix A, as first proposed in [60, 61] and [14]. The
criterion is that the 2-point function have a wavefront set of the form
WF(W2 ) ⊂ {(x1 , k1 ; x2 , k2 ) ∈ T ∗ M 2 \ 0 | k1 ∈ V̇ + , k2 ∈ V̇ − , k1 ∼ −k2 } , (17)
and that each “connected n-point function” ωnc (φ (x1 ), ..., φ (xn )) for n 6= 2 (defined in eq. (19) below) is
smooth7. Here V + ⊂ T ∗ M is the collection of all non-zero, future directed time-like or null co-vectors
(and similarly V − ), and the relation ∼ holds between two covectors if they can be parallel transported
into each other along a null-geodesic in (M , g). Such states are called “Hadamard states”. As we
will see in the next section, it will be necessary to restrict to Hadamard states in order to extend the
action of the state to an enlarged algebra of field observables that includes all polynomials in the
field and its derivatives at the same spacetime point. In particular, it will be necessary to restrict to
Hadamard states when considering the perturbative expansion of interacting QFTCS.
One can prove [60] that the above definition of a Hadamard state is equivalent to the following local
condition [54] on the 2-point function together with the requirement that, globally, there be no sin-
gularities at spacelike separations: For every convex, normal neighborhood U ⊂ M , the two-point
function has the form (in d = 4 dimensions; similar expressions hold in arbitrary d):
" ! #
N
1 u
W2 = + ∑ vn σ n log(σ + i0+t)
4π 2 σ + i0+t n=0
+ some N-times continuously differentiable function RN,ω (18)
=: HN + RN,ω .
Here, the spacetime arguments are understood as (x, y) ∈ U ×U, and we mean that the above formula
should hold for every N, with (different) remainders RN,ω ∈ CN (U × U). σ is the signed squared
geodesic distance, and t = T (x) − T (y) for some (in fact, any) global time function T : M → R. The
6Two representations are said to be unitarily equivalent if there is an isometry U : H → H 0 such that Uπ(a)U ∗ = π 0 (a)
for all a ∈ A .
7 The last requirement may actually be shown to be a consequence of the first condition [63].
10 HOLLANDS AND WALD

functions u, vn are determined by certain local “transport equations” [18] and also appear in similar
local forms of the advanced and retarded Green’s function for the operator g − m2 . In particular,
HN is locally and covariantly defined in terms of the metric, and is hence the same for any Hadamard
state. By contrast, the remainder RN,ω depends on the state.
The Hadamard condition does not single out a particular state, but a class of states. Existence of
a large class of Hadamard state on any globally hyperbolic spacetime can be established by a de-
formation argument [32] combined with microlocal techniques, or by methods from the theory of
pseudo-differential operators [33, 53].
We now discuss two important classes of states: Gaussian states and thermal states.

Gaussian States: Gaussian states (also called “quasi-free states”) are defined by the condition that the
“connected n-point functions” ωnc (φ ( f1 ), ..., φ ( fn )) vanish for all n > 2, where ωnc : A × ... × A → C
is defined by
∂n
 
c
ωn (a1 , . . . , an ) = log ω exp(∑ ti ai ) , ai ∈ A . (19)
∂t1 . . . ∂tn i ti =0

Here the exponential is to be understood in the sense of a formal series. Thus, the n-point functions
of Gaussian states can be expressed entirely in terms of their 1- and 2-point functions. For a Gaussian
state, positivity will hold if and only if (13) is satisfied. Thus, there exist a wide class of Gaussian
states.
Any Gaussian state, ω, can be expressed as the “vacuum state” in a Fock representation of A . To see
this explicitly, let W2 be the 2-point function of ω. On the complex linear space C0∞ (M , C) of smooth
complex-valued functions on M of compact support, define the inner product h f |hi = W2 ( f , h). This
is hermitian and positive, h f | f i ≥ 0, but contains degenerate vectors, such as elements of the form
f = (g − m2 )h. Let h be the factor space of C0∞ (M , C), divided by the degenerate vectors. (It is
relatively easy to see that the elements in this space can be identified more concretely with complex
valued smooth solutions to the KG-equation, or alternatively, with (complex) initial data on Σ.) Its
completion, denoted by the same symbol, is a Hilbert space, usually referred to as the “1-particle
space”. Let H be the bosonic Fock space over h,

H = C⊕
M
(h ⊗S · · · ⊗S h) , (20)
| {z }
n≥1 n

where ⊗S is the symmetrized tensor product. A representation, π, of A on H can then be defined by

π[φ ( f )] = a([ f ])† + a([ f ]) (21)

where a([ f ]) is the annihilation operator associated with the equivalence class of f in h. The vacuum
vector Ω given by the element |1, 0, 0, . . . i in Fock-space then corresponds to ω, and, as already
mentioned above, (H , π, Ω) is precisely the GNS triple arising from the GNS construction.
A related construction often used in practice to construct Gaussian pure states is the following. Sup-
pose that we have a set of smooth, complex-valued “mode functions” uξ (x) that are solutions to
eq. (1), and which are labelled by ξ ∈ X in some measure space (X, dµ). We assume that for each
f ∈ C0∞ (M , R), the map
Z
X 3 ξ 7→ K f (ξ ) := uξ (x) f (x) dvg ∈ C (22)
M
 QUANTUM FIELDS IN CURVED SPACETIME 11

is in L2 (X, dµ), and in fact that the (real linear) span of such vectors is dense in L2 (X, dµ). We also
assume that the mode functions are such that
ImhK f1 |K f2 iL2 (X,dµ) = 21 E( f1 , f2 ) for all f1 , f2 ∈ C0∞ (M ). (23)
(These properties are equivalent to the statement that the collection of modes (uξ )ξ ∈X is “complete in
the KG-norm”.). Then, clearly
Z
W2 ( f1 , f2 ) = hK f1 |K f2 iL2 (X,dµ) ⇒ W2 (x, y) = dµ(ξ )uξ (x)uξ (y) (24)
X
defines the 2-point function of a Gaussian state, which can be shown to be pure. Its GNS-representation
is thus constructed as above. More concretely, since K clearly maps to zero any f of the form
(g − m2 )h, it follows that it is defined on the equivalence classes [ f ], and it can also be shown
that it provides a bounded isomorphism K : h → L2 (X, dµ). Hence, in this case, we may consider H
as the bosonic Fock-space over L2 (X, dµ), and we may informally write the representative of the field
on this Fock space as Z
π(φ (x)) = dµ(ξ )[uξ (x)aξ + uξ (x)a†ξ ] (25)
X
This is the usual form of the field in the vacuum representation on Minkowski space, where ξ = k ∈
R3 = X, dµ(k) = d 3 k/(2π)3 , and uk (x,t) = (2ωk )−1/2 exp(−iωkt + ikx).
Given two pure Gaussian states ω, ω̃ with 2-point functions W2 respectively W̃2 , one may ask when
their associated GNS-representations are unitarily equivalent. A necessary condition is that there be
a constant c such that
c−1W2 ( f , f ) ≤ W̃2 ( f , f ) ≤ cW2 ( f , f ) (26)
for all f ∈ C0 (M , R). Since it is easy to construct states violating this condition, one sees that there
∞

is in general a large class of inequivalent representations. If the condition is satisfied, it follows that
there is a bounded linear operator S on h such that h[ f ]|[ f ]ih̃ = h[ f ]|S[ f ]ih . A sufficient criterion which
ensures that the two states give rise to unitarily equivalent representations is that
1
tr[(1 − S)(1 − S† )] 2 < ∞ . (27)
A sufficient and also necessary criterion for Gaussian states8 can be found e.g. in [4]. The operator
1 − S in essence characterizes the difference between the 2-point functions. It is therefore plausible
that if their difference is smooth – as happens e.g. if both states are Hadamard – and the manifold
M has compact slices, then (27) should hold, and the representations should be unitarily equivalent.
This can indeed be shown9 [69]. On the other hand, for the case of a non-compact Cauchy surface,
the representations can be unitarily inequivalent if the 2-point functions have sufficiently different
long-range behavior.

Thermal States: If the spacetime (M , g) has a complete time-like Killing vector field ξ , one can
define the notion of a thermal state relative to the time evolution generated by this Killing vector field.
There is an elegant version of this notion referred to as the “KMS-condition,” which can be formulated
directly in terms of the expectation value functional ω, without making reference to any Hilbert-
space representation. The KMS condition is formulated as follows: Let ϕt be 1-parameter group of
isometries ϕt : M → M generated by a Killing vector field ξ . We define an action αt : A → A of our
1-parameter family of isometries on the algebra A of fields on (M , g) by setting αt (φ ( f1 ) · · · φ ( fn )) =
8The case of non-Gaussian states is open.
9 For Gaussian states that are not pure the generalization of unitary equivalence is “quasi-equivalence” [4]. Ref. [69]
actually deals with this more general situation.
12 HOLLANDS AND WALD

φ ( f1t ) · · · φ ( fnt ), where f t (x) = f (ϕ−t (x)). Since the retarded and advanced fundamental solutions, and
the field equation, are invariant under ϕt , it follows that αt respects the algebraic relations in A , i.e.,
it is an automorphism. In fact, from the composition law for the isometries ϕt it immediately follows
that αt ◦ αs = αt+s . In this situation, a state ω is called a KMS-state at inverse temperature β with
respect to αt if the following two conditions are satisfied:
(1) For any collection of ai ∈ A , the function t = (t1 , . . . ,tn ) 7→ Fa1 ,...,an (t) defined by
Fa1 ,...,an (t) = ω(αt1 (a1 ) · · · αtn (an )) (28)
has an analytic continuation to the strip
Tβn = {(z1 , . . . , zn ) ∈ Cn | 0 < Im(z j ) − Im(zi ) < β , 1 ≤ i < j ≤ n} , (29)
This function is required to be bounded and continuous at the boundary.
(2) On the boundary, we have
Fa1 ,...,an (t1 , . . . ,tk−1 ,tk + iβ , . . . ,tn + iβ )
(30)
=Fak ,...,an ,a1 ,...,ak−1 (tk , . . . ,tn ,t1 , . . . ,tk−1 ) .
Note that the definition of a KMS-state only assumes an algebra A and the existence of a 1-parameter
family of automorphisms. The notion of a KMS-state is therefore not tied to the particular example
A = A (M , g) and the particular 1-parameter group of automorphisms αt considered here. It is
thus a definition of a very general nature, applicable to many quantum systems, see e.g. [11] for
further discussion. In the case of C∗ -algebras (algebras of bounded operators), the condition for n = 2
implies the remaining ones, but this is not generally the case for the case of unbounded operator
algebras considered here. It is however the case for the concrete algebra A considered here if we
restrict attention to Gaussian states. In this case, the condition n = 2 also implies that the state is
Hadmard [62].
Let us now motivate the above technical definition by explaining its relation to the usual notion of
thermal equilibrium state in statistical mechanics. Consider a self-adjoint Hamiltonian H defined on
a Hilbert space with spectrum bounded from below, and suppose that Zβ = tr e−β H < ∞. The standard
definition of a Gibbs state is ω(a) = tr(ae−β H )/Zβ , where a is e.g. any (say) bounded operator on H .
Let αt (a) be defined in this example by αt (a) = eitH ae−itH , i.e., it describes the usual time evolution
of an observables a in ordinary quantum mechanics. Then, using that the spectrum of H is bounded
below, we easily see that,
Fa,b (z) = Zβ−1 tr(aeizH be−izH e−β H ) (31)
is holomorphic in the strip 0 < Im(z) < β , because in this range e−β H provides a sufficient “damping”
to make the trace finite. Furthermore, using the cyclicity of the trace, we have
Fa,b (t + is) = Zβ−1 tr(e−β H ae(it−s)H be(−it+s)Ĥ ) (32)
= Zβ−1 tr(e−β H e(it+β −s))H be(−it−β +s)H a) . (33)
From the first line we see that Fa,b (t + is) → ω(aαt (b)) for s → 0+ , while we see from the second
line that Fa,b (t + is) → ω(αt (b)a) for s → β − . Thus, (1) and (2) hold (for n = 2), and therefore a
Gibbs state in the usual sense is a KMS-state in the sense of the above definition. The idea behind
the definition of a KMS-state is to turn this statement around and define thermal equilibrium states by
conditions 1) and 2).
A key technical advantage of the definition of a KMS-state is that it still makes sense when a density
matrix no longer exists, as usually happens when the Cauchy surface Σ is non-compact, as occurs in
 QUANTUM FIELDS IN CURVED SPACETIME 13

the usual ‘thermodynamic limit’. The standard example of this is Minkowski space. In the standard
GNS-representation (H , π, Ω) corresponding to the vacuum state (described e.g. by the mode func-
tions given above), the Hamiltonian does not have the property that e−β H is a trace-class operator on
H , i.e. no density matrix exists. Nevertheless, a Gaussian KMS-state can easily be defined in terms
of its 2-point function, given by
1
Z
D+ (t1 − t2 − i0+ , p) eip(x1 −x2 ) d 3 p ,
β
W2 (x1 , x2 ) = ω(φ (x1 )φ (x2 )) = (34)
(2π)3
β
where Minkowski points are labelled by x = (t, x) ∈ R1,3 , and where D+ is given by

β 1 cosh( 12 β ωp − iωpt) p
D+ (t, p) = , ωp = p2 + m2 . (35)
2ωp sinh( 12 β ωp )
The 1-parameter family of isometries is simply given by ϕT (t, x) = (t + T, x), i.e. time translations.
The verification of the KMS-condition boils down to the condition on the 2-point function, which in
turn boils down to showing that the z = t + is-dependent distribution Fx,y (t + is) = W2 (x, ϕt+is (y)) in
x, y ∈ R1,3 has distributional boundary values W2 (x, ϕt (y)) resp. W2 (ϕt (y), x) for s → 0+ resp. s → β − .
This is in turn directly seen to be a consequence of the functional relation
β β
lim D+ (t − is, p) = lim D+ (−t + is, p) . (36)
s→β − s→0+

2.2. Applications: deSitter space, Unruh/Hawking effect, inflationary perturbations. We now

discuss some concrete examples in order to illustrate the abstract ideas just given and to present some
of the important applications of QFTCS.

a) Unruh effect: A relatively simple and yet very important application of QFTCS arises if we
consider a “wedge” W of Minkowski spacetime and view it as a spacetime in its own right. Namely,
let
W = {x ∈ R1,3 | x1 > |x0 |}, (37)
and let W be equipped with the Minkowski metric. Of course, this is not a curved spacetime, but it is
a globally hyperbolic spacetime that differs in essential ways from Minkowski spacetime, e.g., all of
its timelike and null geodesics are incomplete. The spacetime (W, g) is called “Rindler spacetime”.
Writing U = −x0 + x1 and V = x0 + x1 , the metric of Rindler spacetime is
g = dUdV + dx22 + dx32
(38)
= ea(u+v) dudv + dx22 + dx32
where u and v are defined by U = eau and V = eav . Further introducing (η, ξ ) by
u = ξ − η, v = η +ξ, (39)
the metric takes the form
g = e2aξ (−dη 2 + dξ 2 ) + dx22 + dx32 . (40)
The coordinates η and ξ are related to the original global inertial coordinates (x0 , x1 , x2 , x3 ) of Minkowski
spacetime by
x0 = a−1 eaξ sinh aη
(41)
x1 = a−1 eaξ cosh aη.
14 HOLLANDS AND WALD

It is not difficult to see that the hypersurfaces, Ση , of constant η are Cauchy surfaces for Rindler
spacetime for all η ∈ R. Furthermore, for any t ∈ R, the transformation ϕt : η → η + t is an isometry
of the Rindler spacetime, which corresponds to Lorentz boosts of Minkowski spacetime. Indeed, the
key fact about Rindler spacetime is that the orbits of the Lorentz boosts are everywhere timelike and
are complete in Rindler spacetime. Thus, Rindler spacetime is a static, globally hyperbolic space-
time, where the notion of “time translations” is defined by Lorentz boosts. Note that each Lorentz
boost orbit in Rindler spacetime corresponds to the worldline of a uniformly accelerating observer in
Minkowski spacetime. The Lorentz boost orbits become null on the boundary of Rindler spacetime
and, indeed, the hypersurfaces U = 0 and V = 0 of Minkowski spacetime comprise a bifurcate Killing
horizon of the Lorentz boost Killing field, with surface gravity κ = a.
As on every globally hyperbolic spacetime, we quantize the field φ by viewing it (after smearing with
a test function) as an element of the associated abstract algebra A (W, g) defined by the relations 1)
to 4) above. Actually, in the present context, these relations are identical with those of the whole
Minkowski spacetime, because the advanced and retarded propagators, and hence E, are locally the
same. However, there is a difference in that the algebra A (W, g) only contains smeared elements
of the form φ ( f ) for test functions f that are compactly supported in W . Thus, A (W, g) may be
viewed as a proper subalgebra of the algebra associated with the entire Minkowski spacetime. Thus,
we can obtain a state on A (W, g) by restricting the usual vacuum state on Minkowski spacetime, and
view it as a state on Rindler space. The 2-point function of this Gaussian state is (taking m2 = 0 for
simplicity)
1
W2 (x, y) = 2 , (42)
2π (x − y − i0+ e)2
where e is any future directed fixed vector, and the distributional boundary value prescription is un-
derstood.
Since W2 (x, y) 6= 0 for any x ∈ W and y ∈ W 0 (where W 0 denotes the “opposite wedge” x1 < −|x0 |),
it follows that there are correlations between field observables in W and W 0 and that restriction of
the Minkowski vacuum to Rindler spacetime cannot yield a pure state. A key result is the following
theorem, which is a special case of the Bisognano-Wichmann theorem [35] of axiomatic quantum
field theory in Minkowski spacetime:

Theorem 1. The restriction of the Minkowski vacuum state to the Rindler algebra A (W, g) is a KMS-
state with respect to the 1-parameter group of isometries given by η → η +t. The inverse temperature
of this KMS-state is given by
2π
β= . (43)
a
To prove this claim, one has to verify the KMS-condition. This can be done in a completely straight-
forward manner. For a Gaussian state such as as that given here, it suffices to verify the KMS-
condition for a1 = φ ( f1 ), a2 = φ ( f2 ) with f1 , f2 having their support in W , which in turn boils down
to showing that the z-dependent distribution Fx,y (t + is) = W2 (x, ϕt+is (y)) in x, y ∈ W has distributional
boundary values W2 (x, ϕt (y)) resp. W2 (ϕt (y), x) for s → 0+ resp. s → β − . This is an elementary com-
putation done by transforming the 2-point function into Rindler coordinates.
The Minkowski vacuum state is, of course, regular on the Rindler horizon and is invariant under
Lorentz boosts, and it is the only Hadamard state on all of Minkowski spacetime that is Lorentz boost
invariant. In fact, the uniqueness and KMS property (but not necessarily existence) of an isometry
invariant Hadamard state can be proven to hold on any globally hyperbolic spacetime with a bifurcate
Killing horizon [54]. Important further examples will be provided in the next two subsections.
 QUANTUM FIELDS IN CURVED SPACETIME 15

The above theorem has an important physical interpretation, known as the Unruh effect: If the field
is in the Minkowski vacuum state, a uniformly accelerating observer in Minkowski spacetime—who
may also be viewed as a static observer in Rindler spacetime—will “feel himself” immersed in a
“thermal bath of particles” at inverse temperature (43). This can be explicitly seen by introducing
a model “particle detector” and showing that it will be suitably excited as a result of its interaction
with the quantum field. This provides an excellent illustration of why notions of “vacuum state”
and “particles” cannot be considered to be fundamental in the formulation of QFTCS. If the field
is in the Minkowski vacuum state, an inertial observer will naturally declare that no “particles” are
present, whereas the accelerating observer will naturally declare that the Rindler wedge is filled with
a thermal bath of particles. However, there is no actual disagreement between these observers: They
both agree that the field is in a Gaussian state with two-point function (42), and they will be in
complete agreement on the probabilities for measuring any field obserables.

b) deSitter spacetime: Four dimensional (global) deSitter space dS4 is the 4-dimensional hyperbol-
oid defined by the equation Y ∈ R5 ,Y ·Y = H −2 where H > 0 is the Hubble constant and the dot “·”
denotes the 5-dimensional Minkowskian inner product with signature (− + + + +). The metric is that
induced from the ambient space. d-dimensional deSitter space is defined in the same way.
From its definition as a hyperboloid in 5-dimensional Minkowski space, it is clear that deSitter has the
10-dimensional group O(4, 1) as its isometry group. Let us now consider Klein-Gordon quantum field
on deSitter spacetime. Since dS4 is globally hyperbolic, we can define the algebra of field observables
A (dS4 , g) by the general procedure above. For m2 > 0, a globally O(4, 1)-invariant state exists called
the Bunch-Davies (aka Hartle-Hawking, aka Euclidean) vacuum. To describe this it is convenient to
introduce a function Z : dS4 × dS4 → R by
Z(x, y) = H 2 Y (x) ·Y (y) (44)
in terms of the embedding Y : dS4 → R5 of deSitter space into five-dimensional Minkowski space.
This function is symmetric, deSitter invariant, and is related to the signed geodesic distance σ by the
formula √
cos(H σ ) = Z , (45)
where the square root is taken to be imaginary for time-like separated points. The causal relationships
between points can be put in correspondence with values of Z; see the conformal diagram, fig. 1.
In terms of Z, the 2-point function of the Bunch-Davies state is [1, 12, 13] (in d dimensions)
H d−2 Γ(−c)Γ(c + d − 1) 1 + Z − it0+
 
W2 (x, y) = 1 F2 −c, d − 1 + c; d/2; , (46)
(4π)d/2 Γ(d/2) 2
where the dimensionless constant c is defined by
r
d −1 (d − 1)2 m2
c=− + − 2 (47)
2 4 H
and the usual boundary value prescription has to be applied, with t = Y 0 (x) − Y 0 (y). It is relatively
easy to check that this 2-point function satisfies the KG-equation in each argument, and it can be
verified that its antisymmetric part satisfies (12). It is far less obvious that the 2-point function is
positive, but this be shown using the following rather elegant and non-trivial representation due to [12,
2
13] (assuming for simplicity a “principal series scalar field” characterized by µ 2 := m2 − (d−1) 4 H ≥
2

0)
d d−1 k
Z
W2 (x, y) = const. ∑ (Y (x) · ξ (k, l))c (Y (y) · ξ (k, l))c̄ p . (48)
l=± R
d−1 k2 + µ 2
16 HOLLANDS AND WALD

J + (x) = future of x

I+
Z>1

1
Z

1
−

=
=
=

Z
1
Z
Z < −1 |Z| < 1 x |Z| < 1

Z
=

=
−

1
1
Z>1
I−

F IGURE 1. Conformal diagram and values of the point-pair invariant Z = Z(x, y)

as y is varied and x is kept fixed. For the sake of easier visualization, we are giv-
ing the diagram in the case of d = 2 dimensional deSitter spacetime, where the left
and right vertical boundary are to be identified. For d > 2 dimensions, the diagram
would basically consist of only the shaded “left half”, with the vertical boundary lines
corresponding to the north- and south pole of the Sd−1 Cauchy surface.

Here, ξ (k, l) ∈ Rd+1 are the (d + 1)-dimensional vectors defined by

q
ξ (k, ±) = ( k2 + µ 2 , ±k, ±µ) . (49)
More precisely, W2 is defined again as the boundary value of the analytic function obtained by adding
to the time coordinate of y a small positive imaginary part. To check that it is Hadamard, one may use
the relationship between the wave-front set and distributional boundary values; see appendix A.
The following statements hold true concerning deSitter invariant states [1]:
• When m2 > 0, then the Bunch-Davies state is the unique deSitter invariant, pure, Gaussian,
Hadamard state, although a 1-parameter family of states exist if the Hadamard condition is
dropped.
• When m2 ≤ 0, no deSitter invariant state exist, although, as emphasized, in those cases an
infinite set of non-deSitter-invariant Hadamard states still exists. In particular, the algebra A
may always be defined for any value of m2 , although when m2 < 0 the n-point functions of
physically reasonable states will grow exponentially with time.
In deSitter space, there is a phenomenon reminiscient of the Unruh effect which takes place in the
“static chart;” see fig. 2. That chart can be defined as the intersection of dS4 with a wedge {X0 ± X1 >
0} in the ambient R5 . The static chart is again a globally hyperbolic spacetime in its own right, and
can also be defined as the intersection J + (i− ) ∩ J − (i+ ) of two points i± ∈ I ± which are at the “same
angle”. It can be covered by the coordinate system (t, r, ϕ, θ ) defined for t ∈ R, 0 < r < H −1 , in which
the line element takes the form
g = −(1 − H 2 r2 )dt 2 + (1 − H 2 r2 )−1 dr2 + r2 (dθ 2 + sin2 θ dϕ 2 ) . (50)
It can be seen from this form of the line element that, within this chart—but of course not in the
∂
full deSitter space—the metric is static, with timelike Killing field ξ = ∂t . The corresponding flow
ϕs : t 7→ t + s defines a 1-parameter group of isometries in the static chart, which correspond to a boost
in the X0 -X1 plane in the ambient R5 . The boundary H = H+ ∪ H− is formed from two intersecting
cosmological horizons, and is another example of a bifurcate Killing horizon, with surface gravity
 QUANTUM FIELDS IN CURVED SPACETIME 17

I+ i+

+
H
static chart

on
riz
ho
bifurcation bifurcation

ho
riz
on
H
−
I− i−

F IGURE 2. Conformal diagram for deSitter spacetime, and the static chart. Again,
we are drawing the case d = 2. The case d > 2 would correspond to the shaded square
having the bifurcation surface ∼
= Sd−2 in the middle. The vertical boudaries of the
shaded square correspond to the north- and south pole of the Cauchy surface Sd−1 .

κ = H. The restriction of the Bunch-Davies state to the static chart is seen to be a KMS-state at inverse
temperature β = H/2π by the same argument as given for the Unruh effect in Rindler spacetime
(see [54]). Note that the static orbit corresponding to r = 0 is a geodesic, and, by deSitter invariance,
any timelike geodesic in deSitter spacetime is an orbit of the static Killing field of some static chart.
In this sense, one may say that in the Bunch-Davies state in deSitter spacetime, every freely falling
observer will “feel himself” immersed in a thermal bath of particles at inverse temperature β = H/2π.
Of particular interest in cosmology is the behavior of the 2-point function of the Bunch-Davies state
for large time-like separation. A representation10 which is particularly useful for that purpose is, for
a principal series scalar field
dz Γ(d − 1 + c + z)Γ(−c + z)Γ(−z)
Z
1 + z


W2 (x, y) = cst. 2 (1 − Z + it0 ) sin[π(z + (d − 2)/2)] Γ(d/2 + z)
. (51)
C 2πi

Here, the contour C is parallel to the imaginary axis near infinity, and it leaves the poles −(d − 2)/2 +
N0 and N0 to the right, whereas it leaves the poles c − N0 and −(d − 1 + c) − N0 to the left, see fig. 3.
√
The location of the contour C implies that for large proper time-like separation τ = −σ [i.e. Z →
−∞, compare eq. (45) and fig. 1], the 2-point function decays exponentially like e−(d−1)Hτ/2 . This
behavior corresponds to the exponential dispersive effects of fields on deSitter space. It implies that
the 2-point function of any Hadamard state of the form Ψ := π[φ ( f1 ) · · · φ ( fn )]Ω, fi ∈ C0∞ (M ) in the
GNS-representation (H , π, Ω) of the Bunch-Davies state ω approaches that of the Bunch-Davies
state (at fixed proper time separation) in the distant future. Such states are, by construction, dense in
H . That behavior corresponds, physically, to the “no-hair property” of deSitter spacetime. As one
can show with considerably more effort, that behavior persists for interacting quantum field theories,
see [37–39, 59].

c) Hawking effect: The algebraic formalism can be used to give a conceptually clear explanation
of the Hawking effect. In fact, there are two closely related, but distinct, results that are commonly
referred to as the “Hawking effect.”

10The formula may be demonstrated by deforming C → C0 (see fig. (3)), evaluating the integral by the residue theorem,
and comparing the result to (46) which can be transformed and expanded using well-known identities for the hypergeo-
metric function.
18 HOLLANDS AND WALD

C Im(z)
C0

Im(c)

− d−1
2 Re(z)
− d−2
2 − d−4
2 − d−6
2

−Im(c)

F IGURE 3. Location of the contour C and the poles in the representation (51) for
the 2-point function. We assume a principal series scalar field characterized by µ 2 =
m2 − ( d−1 2
2 H) > 0.

singularity (r = 0) i+
+
H

exterior
I+
on
riz
ho

bifurcation S2 spatial infinity i0

ho

I−
riz
on
H
−

singularity (r = 0) i−

F IGURE 4. Conformal diagram of extended Schwarzschild spacetime (‘eternal black hole’).

The first result concerns maximally extended Schwarzschild spacetime (i.e., an “eternal black hole”).
As is well known, the exterior region, r > 2M, of Schwarzschild spacetime
ds2 = −(1 − 2M/r)dt 2 + (1 − 2M/r)−1 dr2 + r2 (dθ 2 + sin2 θ dϕ 2 ), M > 0, (52)

may be extended by introducing the Kruskal coordinates

U = e−u/4M , V = ev/4M , (53)
where
u = t − r∗ , v = t + r∗ , (54)
with r∗ = r + 2M log(r/2M − 1). In Kruskal coordinates, the line element takes the form
2 32M 3 e−r/2M
ds = dUdV + r2 (dθ 2 + sin2 θ dϕ 2 ) , (55)
r
By considering arbitrary U,V compatible with r > 0, one obtains the maximally extended Schwarz-
schild spacetime shown in the conformal diagram fig. 4. The surfaces U = 0 and V = 0 (corresponding
to r = 2M) comprise a bifurcate Killing horizon, H± , of the Killing field ξ = ∂ /∂t, analogous to the
bifurcate Killing horizons of the boost Killing field of Minkowski spacetime and the static Killing field
 QUANTUM FIELDS IN CURVED SPACETIME 19

of deSitter spacetime. In close analogy with those cases, there exists [63] a unique [54] Hadamard
state, ω, on extended Schwarzschild spacetime that is stationary i.e., invariant under time-translation
automorphisms, ω = ω ◦ αt . This state is known as the “Hartle-Hawking vacuum” and is analog-
ous to the Minkowski vacuum in Minkowski spacetime and to the Bunch-Davies vacuum in deSitter
spacetime. By the same argument as in those cases [54], when restricted to the original Schwarzschild
wedge, r > 2M, the Hartle-Hawking vacuum is a KMS state at the Hawking temperature
κ 1
=TH = , (56)
2π 8πM
where κ = 1/4M is the surface gravity of the Killing horizon.
Two other states of interest on extended Schwarzschild spacetime are the “Boulware vacuum” [10],
and the “Unruh vacuum” [68]. The Boulware vacuum is defined in the right wedge of extended
Schwarzschild spacetime, where it is a ground state with respect to the timelike Killing vector field
∂t . It is Hadamard in the right wedge, but cannot be extended as a Hadamard state beyond the right
wedge, i.e. it would become singular on the past and future horizons. The Unruh vacuum is defined
on the union of the right wedge and the interior of the black hole. In the right wedge, it can be thought
of as a KMS state with respect a subalgebra of A corresponding to the modes that are outgoing from
the white hole, whereas it the ground state with respect to a subalgebra corresponding to the modes
that are incoming from past null infinity. The Unruh vacuum has been rigorously constructed in [20],
and has been shown to be Hadamard on the union of the right wedge and the black hole regions. It
cannot be extended as Hadamard state beyond the past horizon.
The thermal nature of the Hartle-Hawking state suggests, but does not imply, a second key result:
Black holes formed by gravitational collapse will emit thermal radiation. To analyze this issue, one
must consider the much more physically relevant case of a spacetime in which gravitational collapse
to a Schwarzschild black hole occurs, rather than the maximally extended Schwarzschild spacetime
considered above. In the case of a black hole formed by collapse, one can show that if the state of
the quantum field is Hadamard and if it approaches the ground state near spatial infinity, then at late
times it contains quanta of radiation appearing to emanate from the black hole, distributed according
to a Planck-distribution with temperature (56).
We now give some details, following the argument given by Fredenhagen and Haag [28]. Let Fνlm T be a

solution to the Klein-Gordon equation with smooth initial data of compact support on the gravitational
collapse spacetime which has Ylm angular dependence and frequency peaked sharply near ν > 0 (with
respect to the timelike Killing field), and which corresponds at late times to an outgoing wave reaching
null infinity at retarded time centered about T . We normalize Fνlm T so that it has unit Klein-Gordon
11
norm . Then, in any state ω, the quantity ω(φ [Fνlm T ]∗ φ [F T ]) has the interpretation of being the
νlm
T
“number of particles” in the mode Fνlm as seen by a distant oberver at late times, as can be seen from
Fock representation formulas (see section 2.2 above) or by considering the behavior of model particle
detectors [31], [71]. We shall show that if ω is Hadamard and approaches the ground state near spatial
infinity, then
T ∗ T |Dl (ν)|2
lim ω(φ [Fνlm ] φ [Fνlm ]) = 2πν/κ , (57)
T →∞ e −1
where Dl (ν) is the amplitude for the absorption by the black hole of a mode of angular dependence
Ylm and frequency ν. This is precisely the expected number of particles that one would have for
black-body radiation “emitted” by the black hole at the Hawking temperature.
11The Klein-Gordon inner product between two solutions F, G of compact support on a Cauchy-surface Σ is defined as
R
Σ (F̄∇µ G − G∇µ F̄)n
(F, G) = i µ dS; compare (10).
20 HOLLANDS AND WALD

support of initial data of F T

i+
in
fin
ity
I +

+
H
no
i0
riz
ho
Σ0
collapsing star

− support of f−T
I
ity
fin f T decaying → 0 as T → ∞
in
support of f+T

i−

F IGURE 5. Conformal diagram of collapsing star spacetime.

To show this, we choose a partial Cauchy surface Σ0 intersecting the future horizon H+ at a 2-sphere
S ∼ = S2 outside of the collapsing star (see fig. 5). In the future domain of dependence D+ (Σ0 ) of Σ0 ,
the spacetime metric is precisely equal to that of the Schwarzschild metric. The conditions on the
state ω imply that
1) The Hadamard condition in the form (18) holds for the 2-point function W2 (x1 , x2 ) for x1 , x2
in an open neighborhood of the horizon H+ . Together with the fact that the coefficient ‘u’ in
the Hadamard expansion satisfies u(x1 , x2 ) ≡ 1 for all x1 , x2 ∈ H+ , one can infer that
1
W2 (x1 , x2 ) = 2 + lower order terms in σ near H+ . (58)
2π (σ + it0+ )
(Here, lower order refers to the “scaling degree”; see appendix A.)
2) In an open neighborhood of Σ0 , and for r1 , r2 → ∞, the 2-point function W2 (x1 , x2 ) approaches
that of the ground state.
T by a test function f T which is supported
The key idea in the derivation is to replace the solution Fνlm νlm
close to Σ0 . To do so, we let ψ be a smooth function which is equal to 1 slightly in the future of Σ0 ,
and equal to 0 slightly in the past of Σ0 , and we set
T T
fνlm = g [ψFνlm ]. (59)
Then we have φ ( f T ) = φ [F T ] (see eq. (10)). The key property of f T is that it decays uniformly for
large T in any region r1 < r < r2 where r1 > 2M [19]. Thus, in this limit, f T splits approximately into
two parts f+T + f−T (see fig. 5), where the first part, f+T , is supported close to the horizon S , and the
second part, f−T , is supported close to spatial infinity i0 near Σ0 . For f+T , one finds to the appropriate
accuracy for T  1:
 
T iν κT
f+ (U,V ) ∼ Dl (ω) ∂V ψ(V ) ∂U exp log(Ue ) (60)
κ
 QUANTUM FIELDS IN CURVED SPACETIME 21

where for simplicity, we have taken ψ to be a function of V only near S . Here the logarithmic
dependence on U in the exponent can be traced back to the relationship U = exp κu between “affine
time” U and “Killing time” u on the past horizon H− of extended Schwarzschild spacetime (see (53)).
The contribution of f+T to (57) is given by
Z
T ∗
lim ω(φ [Fνlm T
] φ [Fνlm ]) = lim W2 (x1 , x2 ) f+T (U1 ,V1 , x1 )∗ f+T (U2 ,V2 , x2 ) dv1 dv2 (61)
T →∞ T →∞
On the other hand assumption (2) implies that f−T makes no contribution in the limit T → ∞, either in
its direct terms or its cross-terms with f+T . Since f+T is supported increasingly close to H+ , we may
use assumption (1) to approximate W2 with (58). Using furthermore (60), a short calculation then
gives the thermal distribution formula (57).
The above argument corresponds closely to Hawking’s original derivation [36], showing that the
expected number of particles seen by a distant observer at late times corresponds to thermal emission
by the black hole. In fact, all aspects of this radiation are thermal [70]. The precise result is that
any state that is Hadamard and behaves like a ground near spatial infinity will approach the Unruh
vacuum state, in the sense that the n-point functions of the state approach those of the Unruh vacuum
in the exterior at late times.
We have presented the argument in the above manner to emphasize the following points: (i) The
derivation of the Hawking effect does not depend on introducing a notion of “particles.” (ii) No
assumptions need to be made on the initial state other than it is regular (Hadamard) and approaches
the ground state near spatial infinity (i.e., there is no “incoming radiation” at late times). (iii) The
Hawking effect follows from causal propagation of the quantum field outside of the black hole; one
does not need to make any assumptions about what is happening inside of the black hole. In particular,
any breakdown of known laws of physics in the high curvature regime near the singularity deep inside
the black hole should not affect the validity of the derivation. (iv) The details of the collapse are not
important; all that matters is that the spacetime metric asymptotically approach the Schwarzschild
metric12 by some sufficiently late “time” Σ0 .
Nevertheless, there is one potentially disturbing aspect of this derivation. For large T , it can be seen
from (60) that f+T is peaked near U ∼ exp(−κT ), and that the locally measured frequency of f+T —say,
as seen by observers who freely fall into the black hole from rest at infinity—diverges as exp(κT )
as T → ∞. For any reasonable detector frequency ν, this vastly exceeds the Planck frequency for
T  1/κ. We cannot expect QFTCS to be a good approximation to nature on transplackian scales,
but the above derivation of the Hawking effect appears to depend upon the the validity of QFTCS
at transplanckian frequencies. However, we claim now that this is not actually the case: For a given
T (r,t) at large T , instead of doing the analysis on the partial Cauchy surface Σ , we can work on
Fνlm 0
a partial Cauchy surface Σ1 , which is sufficiently early in time that the approximation leading to (60)
still holds, but is sufficiently late in time that f+T on Σ1 is not transplanckian. By formulating the
detector response at time T an evolution problem starting from Σ1 rather than Σ0 , one avoids13 any
elements of the derivation that allude to phenomena at transplanckian scales.
The derivation of the emission of thermal radiation by a black hole at the Hawking temperature (56)
brought to a culmination a remarkable relationship between certain laws of black hole physics and the
12The results can be straightforwardly generalized to asymptotic approach to other stationary black hole geometries,
with the only significant difference being that, for a rotating black hole, the horizon Killing field will now be a linear
combination of a time translation and rotation at infinity; see, e.g. [36, 70] for a discussion of the Kerr case.
13Of course, the Hadamard condition itself concerns arbitrarily short distance singularity structure and thus, in effect,
involves transplanckian scales. One might therefore question the validity of QFTCS for arguing that the Hadamard
condition is preserved under evolution. However this question could equally well be raised in Minkowski spacetime.
22 HOLLANDS AND WALD

ordinary laws of thermodynamics. It was already known prior to this derivation that classical black
holes satisfy mathematical analogs of the zeroth, first, and second laws of thermodynamics [8], with
mass, M, playing the role of energy, E; surface gravity, κ, playing the role of temperature T ; and
horizon area, A, playing the role of entropy, S. Even in classical physics, a hint that this relationship
might be more than a mathematical analogy is provided by the fact that M and E are, in fact, the same
physical quantity. However, the physical temperature of a classical black hole is absolute zero, thereby
spoiling this relationship in classical physics. The fact that, when analyzed from the perspective of
QFTCS, black holes have a finite temperature (56) proportional to surface gravity strongly suggests
that these laws of black hole physics must actually be the laws of thermodynamics applied to black
holes; see [71, 72] for further discussion. In particular, A/4 must represent the physical entropy of a
black hole. The ramifications of these ideas continue to be explored.
In the above derivation of thermal emission by a black hole, we considered a fixed, classical space-
time corresponding to the gravitational collapse of a body to a Schwarzschild black hole. However,
the quantum field has a stress-energy observable, Tµν (see the next section), and, in semiclassical
gravity, ω(Tµν ) should contribute to the right side of Einstein’s equation. It is easy to see that ω(Tµν )
contributes a positive energy flux to infinity. It follows from conservation of stress-energy together
with the approximate stationarity of ω at late times that ω(Tµν ) contributes a corresponding flux of
negative energy14 into the black hole. Consequently, if the black hole is isolated (so that there is no
other flux of stress-energy into the black hole), the black hole will slowly lose mass as a result of
the quantum field effects. An order of magnitude estimate of the mass loss of the black hole can be
obtained from the Stefan-Boltzmann law
dM 1 1
∼ AT 4 ∼ M 2 4 = 2 (62)
dt M M
leading to the prediction that a black hole should “evaporate” completely15 in a time of order M 3 (in
Planck units).
The prediction of black hole evaporation gives rise to an issue that has re-gained considerable attention
recently. In the analysis of quantum field theory on the gravitational collapse spacetime shown in fig.
5—where the black hole remains present forever—the field observables in an open neighborhood of
any Cauchy surface comprise the entire algebra16, A . Thus, a state that is pure at any initial “time”
(i.e., a neighborhood of an initial Cauchy surface, say, prior to the collapse) automatically will remain
pure at any final “time” (i.e., a neighborhood of a final Cauchy surface, say, after the black hole has
formed). However, the field observables in a small neighborhood of a partial Cauchy surface—i.e., a
hypersurface, such as Σ0 in the above figure, whose future domain of dependence includes the region
exterior to the black hole but not the interior of the black hole—do not comprise all observables
in A since there will be additional field observables inside the black hole. Furthermore, for any
Hadamard state, there always are strong correlations between the field observables at small spacelike
separations. In particular, in any Hadamard state, the observable φ ( f T ) on Σ0 in our above discussion
will be highly entangled with corresponding field observables inside the black hole. Since φ [F T ] =
φ ( f T ) (see (10)), this means that the Hawking radiation flux measured by a distant detector is highly
entangled with field observables inside the black hole. If the black hole subsequently completely
14Negative energy fluxes of stress-energy can occur in quantum field theory even for fields that classically satisfy the
dominant energy condition.
15Of course, the approximate description leading to this prediction should be valid only when M  M , where M
P P
denotes the Planck mass (∼ 10−5 gm), but modifications to the evaporation process at this stage (including the possibility
of Planck mass remnants) would not significantly alter the discussion below.
16This result, sometimes called “time-slice property”, continues to hold for the enlarged algebra W defined in the next
section and also for the algebra of interacting fields BI [23, 48].
 QUANTUM FIELDS IN CURVED SPACETIME 23

evaporates as discussed above, the field observables corresponding to the emitted radiation remain
entangled with observables inside the (now non-existent) black hole. In particular, the subalgebra of
observables that can be measured at late times—after the black hole has evaporated—do not comprise
a complete set of observables, and the restriction of the state ω to this subalgebra is a mixed state.
Thus, in the process of black hole formation and evaporation, an initial pure state will evolve to a final
mixed state. Such an evolution does not violate any principles of quantum theory or any known laws
of physics—indeed, it is derived by a straightforward application of QFTCS to a spacetime in which a
black hole forms and evaporates—but it is in apparent conflict with ideas suggested by the AdS/CFT
correspondence. In any case, there is a widespread belief that evolution from a pure state to a mixed
state should not happen, thereby requiring a drastic modification of QFTCS in a low curvature regime,
where, a priori, one otherwise would have very little reason to question its validity. In particular, in
order to avoid entanglement between observables outside the black hole and inside the black hole, the
state must fail to be Hadamard at the event horizon of the black hole, thereby converting the event
horizon to a singularity (a “firewall”). On the other hand, as we have seen above, the Hawking effect
itself is crucially dependent upon the state being Hadamard arbitrarily close to (but outside of) the
event horizon. Given that the event horizon is not locally determined—i.e., it requires knowledge of
the future evolution of the spacetime—it would seem a daunting task for a quantum field to know, with
the required precision, exactly when to stop obeying the laws of QFTCS, so that Hawking radiation is
maintained but its entanglement with field observables inside the black hole is broken. Nevertheless,
there presently is a widespread belief that a quantum field will somehow manage to do this—or that
the presently known local laws of physics will be violated near the horizon of a black hole in some
other way, so as to maintain the purity of the final state (see [3] and the subsequent literature citing
this paper for further discussion).

d) Cosmological perturbations: We would now like to investigate a massless (m2 = 0) Klein-Gordon

field propagating on an FLRW-spacetime with flat slices
g = −dt 2 + a(t)2 (dx12 + dx22 + dx32 ) . (63)
The isometry group of this spacetime is, for general a(t), the Euclidean group E(3) acting on the
spatial coordinates x ∈ R3 . We consider a scale factor of the form
( Ht
e 0  for t ≤ t0 ,
a(t) = p (64)
a0 tt0 for t > t0 ,
describing a universe which is expanding exponentially first (inflation), followed by an era with
power-law expansion factor, assumed for simplicity to last forever. To be precise, we should actu-
ally choose scale factor that interpolates smoothly, rather than just continuously, between the epochs,
but for the rough calculation this will not be needed. We could also add an earlier epoch of power law
expansion prior to the exponential expansion, but this will not affect our results provided that the era
of exponential expansion lasts sufficiently long.
We consider a massless quantum Klein-Gordon scalar field in the spacetime (64). This is a (slight)
simplification of the more physically relevant problem of starting with a classical solution of the
Einstein-scalar-field system with scalar factor of a form approximating (64)—as would occur if the
scalar field “slowly rolls” down an extremely flat potential—and then treating the linearized perturba-
tions of this system as quantum fields. In that case, the linearized perturbations decouple into “scalar
modes” and “tensor modes,” each of which behave similarly to a scalar field in the background space-
time (64). Thus, consideration of a scalar field in the background spacetime (64) suffices to derive the
general form of the power spectrum of perturbations resulting from inflation.
24 HOLLANDS AND WALD

Consider a Gaussian, pure, E(3)-invariant, Hadamard state of the scalar field on the spacetime (64).
The two-point function, W2 , of such a state may be described by a set of mode functions uk (t, x) =
χk (t) exp(ikx) as in eq. (24). Let us normalize the mode functions so that the Wronskian is i-times
unity, i.e.,  
3 d d
i = a(t) χk (t) χ k (t) − χ k (t) χk (t) . (65)
dt dt
This condition ensures that eq. (23) holds with X = R3 , dµ(k) = d 3 k/(2π)3 so that eq. (24) indeed
defines the 2-point function of a state,
1
Z
W2 (t1 , x1 ,t2 , x2 ) = χk (t1 )χk (t2 ) eik(x1 −x2 ) d 3 k . (66)
(2π)3
A quantity that (partly) characterizes W2 in a Robertson-Walker spacetime is its “power spectrum,”
P(t, k), which is defined in terms of the spatial Fourier-transform of the 2-point function at equal time
t,
Ŵ2 (t, k,t, p) = (2π|k|)3 δ 3 (k − p) P(t, k) . (67)
For an E(3)-invariant state, the power spectrum P(t, k) only depends upon the modulus k = |k| of the
wave number (and of course t).
We shall now show that for any t > t0 (i.e., after inflation has ended), the power spectrum P(t, k)
for modes for which17 k/a(t)  1/R(t), where R denotes the Hubble radius, R(t) ≡ a(t)/ȧ(t), is
approximately given by
P(t, k) ∝ H02 . (68)
Thus, the power spectrum is “scale free,” with amplitude set by the scale of inflation. To show this,
we observe that the Hadamard condition fixes the asymptotic behavior of χk (t) for large |k| → ∞ (to
all(!) asymptotic orders). In the inflationary epoch (t < t0 ), the general solution for χk giving rise to
an E(3)-invariant state of the massless field is
χk (η) = Ak fk (η) + Bk f k (η) , (69)
R
in conformal time η = dt/a(t), where
3 (2)
fk (η) = const. η 2 H (kη) , (70)
− 32
(2)
where Hα denotes a Hankel function, and the Wronskian condition imposes |Ak |2 − |Bk |2 = 1. The
Hadamard condition requires Ak → 1, Bk → 0 at large k. If we assume that Ak ≈ 1, Bk ≈ 0 for all “short
wavelength modes” (i.e., k/a(t1 ) > H0 ) at some time t1 < t0 during the inflationary era, then we have
Ak ≈ 1, Bk ≈ 0 for all modes18 whose physical wavelength is smaller than H0−1 exp[(t0 − t1 )H0 ] at the
end of inflation. For (t0 − t1 )H0 & 60, this encompasses all wavelengths relevant for cosmology.
The mode functions compatible with E(3)-invariance in the power law epoch with a(t) ∝ t p have the
form
χ̃k (η) = Ãk f˜k (η) + B̃k f˜k (η), (71)
where 1−3p
(2)
f˜k (η) = const. η 2(1−p) H 1−3p (kη) , (72)
− 2(1−p)

17Such modes having wavelength larger than the Hubble radius in the present universe are not of observational interest.
However, modes whose wavelength was larger than the Hubble radius at the end of inflation but is smaller than the Hubble
radius in the present universe are highly relevant to cosmology.
18 For m2 > 0, the analogous modes are obtained by setting the index of the Hankel function to α = ( 9 − m2 H −2 )1/2 .
4
The state with Ak = 1, Bk = 0 for all k is the deSitter invariant Bunch-Davies state [66]. However, the choice Ak = 1,
Bk = 0 for m = 0 would yield an infrared divergence in the two-point function (66)
 QUANTUM FIELDS IN CURVED SPACETIME 25

The coefficients Ãk , B̃k are subject to the Wronskian condition |Ãk |2 − |B̃k |2 = 1 and are determined by
matching the modes χk in eq. (69) to χ̃k at time η0 . To do this, we first note that, by assumption, we
are considering modes that satisfy R(t)k/a(t)  1. During the power law epoch, we have a(t) ∝ t p
with p < 1, and R(t) ∝ t, so the quantity R(t)k/a(t) becomes even smaller as we go back in time from
t to t0 . In conformal coordinates, 1  R(t)k/a(t) ∝ kη, so this means that the modes χ̃k (η) of interest
are essentially constant (“frozen”) during the power law epoch for all times before t. Thus, we may
assume that χ̃k (η) ∼ χ̃k (η0 ) = χk (η0 ) during that epoch, so the power spectrum at t is essentially the
same as that of the state in deSitter spacetime at time t0 for kη  1, i.e., that obtained using the modes
χk (η0 ). These modes may then be approximated by χk (η) ∝ k−3/2 H0 for kη  1, thus giving rise to
the desired power spectrum (68). At this level of approximation, the power spectrum is independent
of the power p and of the precise nature of the transition period, although the finer properties of the
power spectrum would depend on such details.
For H0 ∼ 1016 GeV, the amplitude of the power spectrum is macroscopically large, and it provides
an explanation of the observed temperature fluctuations in the cosmic microwave background as well
as of “structure formation” in the universe, i.e., it produces density perturbations appropriate to act
as “seeds” for the formation of clusters of galaxies and galaxies. The fact that, in the presence of
exponential expansion, the short distance quantum fluctuations of fields in the very early universe
can produce macroscopically observable effects in the present universe is one of the most remarkable
predictions of QFTCS.

3. P ERTURBATIVE INTERACTING QUANTUM FIELDS

In this section, we extend the algebra A (M , g) of the previous section to an an enlarged algebra
W (M , g). We then define the algebra of polynomial field observable B0 for the free field and we
define time ordered products. These results are then used in subsection 3.2 to give a perturbative con-
struction of the algebra of interacting field observables, BI . In section 3.3, we give a brief discussion
of the additional ideas that are needed to give similar constructions for gauge fields.

3.1. Construction of Nonlinear Observables for a Free Quantum Scalar Field. The construction
of the theory of a free quantum field in curved spacetime given in the previous section provides a
mathematically consistent and satisfactory formulation of QFTCS for fields obeying linear equations
of motion. However, even in this case, the theory is incomplete: The observables represented in the
algebra A (M , g) consist only of the smeared fields φ ( f ) and their correlation function observables
φ ( f1 ) . . . φ ( fn ). However, A does not include any observables corresponding to nonlinear functions
of the field φ , such as φ 2 or the stress-energy tensor, Tµν , of φ . For this reason alone, one would
like to enlarge the algebra of observables A (M , g) to an algebra W (M , g) that, at the very least,
includes smeared versions of all polynomial expressions in φ and its spacetime derivatives, as well as
correlation function observables of these expressions.
There is an additional reason of at least equal importance for wanting to significantly enlarge A : We
wish to formulate QFTCS for interacting quantum fields. We will discuss a possible approach toward
directly doing so in section 5, but, as an important first step, one would like to give a perturbative
construction of an interacting quantum field theory with Lagrangian of the form
L = L0 + L1 , (73)
where the Lagrangian,
L0 = 21 ((∇φ )2 + m2 φ 2 ) (74)
26 HOLLANDS AND WALD

corresponds to the free Klein-Gordon field, but the interaction Lagrangian, L1 , contributes nonlinear
polynomial terms to the classical equations of motion. A frequently considered example is
L1 = λ φ 4 . (75)
In order to define perturbative expressions for quantities of interest in this nonlinear theory, one must
define polynomial quantities like “φ 4 ” as (distributional) elements of the algebra of observables of the
free theory. One must also define notions of “time ordered products” of φ 4 and other polynomial ob-
servables as elements of the free field algebra, since such quantities appear in the formal expressions
arising in perturbative expansions. One therefore would like to to enlarge the algebra of observables
A (M , g) to an algebra W (M , g) that includes all time ordered products of all polynomial expres-
sions in φ and its spacetime derivatives.
The difficulties in defining nonlinear functions of φ as observables can be understood as arising dir-
ectly from the distributional nature of φ . One would like to define φ 2 (x) as the pointwise product of
φ (x) with itself, but pointwise products of distributions are, in general, intrinsically ill defined. This is
not merely some fussy mathematical point. It is clear from the following simple observation that some
sort of nontrivial “regularization” will be needed to define nonlinear observables, such as φ 2 . For a
Hadamard state ω, the two-point function W2 (x, y) = ω(φ (x)φ (y)) is a smooth function when x and
y are spacelike separated, so we might expect ω(φ 2 (x)) = limy→x W2 (x, y). However, it is easily seen
from (17) that this limit diverges. In essence, the fluctuations of quantum fields become arbitrarily
large at short distances, making a straightforward construction of nonlinear functions of a quantum
field impossible. Nevertheless, a free quantum field should admit nonlinear field observables such as
φ 2 , and we need to have such observables to define nonlinear QFTCS (perturbatively or otherwise).
How are such observables defined/constructed?
We “constructed” the observable φ ( f ) by writing down the free algebra of symbols of this form and
factoring by the relations (1)-(4) of section 2.1, thereby obtaining the algebra A . This worked because
(1)-(4) comprise the complete set of relations that characterize the linear field observables. We could
attempt to define the enlarged algebra W by adding symbols of the form φ 2 ( f ), etc., to the free algebra
and factoring by all of the relations that are satisfied by this enlarged set of observables. However, to
do this, we would need to know all of the relations satisfied by the enlarged set of observables, and it
is far from obvious, a priori, what this complete list of these relations should be. Remarkably, one can
bypass this difficulty and obtain the desired enlarged algebra, W , by proceeding as follows [23, 46].
The first step is to make a trivial enlargement of A to an algebra, A 0 , by allowing, in addition to
the original elements of the form φ ( f1 ) . . . φ ( fn ), elements corresponding to smearing φ (x1 ) . . . φ (xn )
with an arbitrary smooth, compact support test function Fn (x1 , . . . , xn ). In other words, A 0 is obtained
by starting the free algebra generated by the formal expressions
Z
F̂n = φ (x1 ) . . . φ (xn )Fn (x1 , . . . , xn ) (76)

and then factoring by the analogs of relations (1)-(4). The replacement of A by A 0 makes no essential
change to the free field theory.
The second step is to choose a Hadamard state, ω, on (M, g) and define the normal-ordered product
of fields relative to ω by
: φ (x1 ) . . . φ (xn ) :ω = ∑ ∏ φ (xk ) ∏ [−W2 (xi , x j )] (77)
P k∈P
/ (i, j)∈P

where P is a collection of disjoint ordered pairs (i, j) such that i < j, and where k ∈ / P denotes the
indices k not in that collection of pairs. W2 is as usual the two-point function of the state ω, and both
 QUANTUM FIELDS IN CURVED SPACETIME 27

sides are understood to be smeared with a test function Fn (x1 , . . . , xn ). We will denote the element of
A 0 defined by the smearing of (77) with Fn as : F̂n :ω . For example, we have
Z 
: F̂2 :ω = F̂2 − F2 ·W2 1 . (78)
M ×M
Using the analog of the commutation relation (condition (3)) on A 0 , it may then be seen that elements
of the form : F̂n :ω comprise a basis of A 0 , i.e., every element of A 0 of the form Ĝk can be expressed
as a linear combination of terms of the form : F̂n :ω with n ≤ k. Furthermore, the analog of the
commutation relation (condition (3)) on A 0 is effectively encoded by the following rule for calculating
products of normal ordered products
: Fbn :ω · : G
bm :ω = ∑ ⊗k Gm :ω
: Fn\ (79)
k≤min(n,m)

where the k times contracted tensor product is defined by

k n!m!
Z
Fn ⊗k Gm (x1 , ..., xn+m−2k ) = (−1) W2 (y1 , y2 ) · · · W2 (y2k−1 , y2k ) × (80)
k! ∑π M
2k

Fn (y1 , y3 , ..., xπ(1) , ..., xπ(n−k) )Gm (y2 , y4 , ..., xπ(n−k+1) , ..., xπ(n+m−2k) ).
The sum ∑π denotes the sum over all permutations π of {1, ..., n + m − 2k} with the property π(1) <
... < π(n − k) and π(n − k + 1) < ... < π(n + m − 2k). The product formula is a version of “Wick’s
theorem.”
The construction of the previous paragraph produces no change whatsoever to the algebra A 0 , i.e., it
merely rewrites A 0 in a new basis, with a product law that incorporates the relations already present
in A 0 . However, it puts us in position take the major step of enlarging A 0 to the desired algebra
W as follows: As noted above, the pointwise product of distributions is, in general, ill defined.
However, there is one significant exception to this statement: As discussed in the Appendix, if u, v are
distributions and if their wavefront sets, WF(u) and WF(v), are such that WF(u) + WF(v) does not
contain a zero cotangent vector, then the pointwise product uv is well defined in a natural manner. The
wavefront set properties of the two-point function W2 of a Hadamard state are such that the products
appearing on the right side of (79) make sense when the test function Fn is replaced by any distribution
whose wavefront set obeys the restriction
WF(Fn ) ∩ (V+n ×V−n ) = 0/ . (81)
Similarly Gm may be replaced by such a distribution. Thus, we may define a new algebra W generated
by expressions of the form : F̂n :ω with product law (80), where Fn is now an arbitrary distribution
of compact support whose wavefront set satisfies (81). Since, in particular, for any test function f ,
the distribution f (x1 )δ (x1 , . . . , xn ) is of the required form, it can be seen that the algebra W includes
elements that can be interpreted as representing : φ n ( f ) :ω . As we shall see below, the algebra W
is “large enough” to include all polynomial expressions in φ and its spacetime derivatives as well
as all time ordered products of such expressions. The product law (79) provides us with all of the
“relations” that hold between these new observables.
The above construction of W can be understood as taking a certain “closure” of A 0 , because distri-
butions Fn satisfying the wave front set condition (81) can be approximated, in a suitable topology,
by smooth Fn ’s, which in turn correspond to elements : F̂ :ω of A 0 . The topology on the Fn ’s thereby
naturally induces a topology on W , and different Hadamard states ω are easily shown to lead to the
same topology. It then follows that: (i) The definition of W is independent of the choice of ω. (ii)
A 0 is dense in W . (iii) Any continuous homomorphism on A 0 extends uniquely to a homomorphism
of W (since the *-operation and the product, eq. (79), can be seen to be continuous in the topology).
28 HOLLANDS AND WALD

(iv) Any Hadamard state on A 0 can be extended to a (continuous) state on W . Conversely, it can be
shown that the restriction to A 0 of any continuous state on W yields a Hadamard state on A 0 [45,64].
(v) For any Hadamard state, ω, any one-point distribution, ω(: φ n (·) :ω ), is smooth.
The above construction tells us, for any given globally hyperbolic spacetime (M , g), how to construct
the desired extended algebra of observables W = W (M , g). However, it does not tell us which ele-
ment of W (M , g) to associate with a given field observable, such as φ 2 . In particular, the observable
: φ 2 ( f ) :ω [i.e., in our above notation, the observable : F̂2 :ω with F2 (x1 , x2 ) = f (x1 )δ (x1 , x2 )] is not a
reasonable candidate to represent φ 2 since it depends upon an arbitrary choice of Hadamard state ω.
How does one determine which element of W (M , g) to associate with a given field observable, such
as φ 2 ( f )?
As we shall see, there will be some “local curvature ambiguities” in the definition of field observables.
However, the fixing of these ambiguities in the choice of prescription for, say, φ 2 will affect the
definition of, say, φ 4 . Therefore, in order to properly discuss the ambiguities in the definition of field
observables, it is important that we determine all field observables of interest “at once.” The task
at hand can then be formulated as follows. Consider the space of classical polynomial expressions,
Φ[(M , g); φ ], in the scalar field and its derivatives defined on arbitrary globally hyperbolic spacetimes
(M , g), i.e., for each globally hyperbolic spacetime (M , g), Φ(φ ) is a polynomial expression in φ and
its derivatives on (M , g) (with coefficients that may depend upon spacetime point). We are interested
in the Φ that are local and covariant in the sense that if ψ : M → M 0 is an isometric embedding (i.e.,
ψ ∗ g0 = g) that also preserves the causal structure—so that if x1 and x2 cannot be connected by a causal
curve in M , then ψ(x1 ) and ψ(x2 ) cannot be connected by a causal curve in M 0 —then Φ[(M , g), φ ]
must satisfy

ψ ∗ Φ[(M 0 , g0 ), φ ] = Φ[(M , g), ψ ∗ φ ] . (82)

Thus, for example, φ 2 and µν

R R ∇µµφ ∇ν φ (where R µν is the Ricci curvature) are local and covariant
expressions, whereas φ M R and v ∇µ φ (where vµ [M ] is a vector field independent of g introduced
on each manifold M ) are not local and covariant expressions. Note that one cannot tell if an expres-
sion Φ is local and covariant unless it is defined for all globally hyperbolic spacetimes. Let P(M )
denote the space of classical local and covariant polynomial expressions in φ . By the Thomas re-
placement theorem [52], any Φ ∈ P(M ) may only depend on derivatives of the metric in the form
of the Riemann tensor and its (symmetrised) covariant derivatives ∇(µ1 ...∇µk ) Rαβ γδ . Furthermore,
at any point x ∈ M , Φ may only depend on fields φ and their (symmetrised) covariant derivatives
∇(µ1 ...∇µk ) φ evaluated at the point x. This result allows us to assign a dimension to terms in P(M ),
by assigning ∇µ to have dimension 1, φ to have dimension 1, and Rµνσ ρ to have dimension 2.
The question of how to determine which element of W (M , g) to associate with a given field observ-
able, Φ, can now be reformulated as how to define a suitable map Q (“for quantization”) from P(M )
to distributions valued in W . Our strategy for obtaining Q is to write down the conditions that we
want Q to satisfy, and then determine existence and uniqueness. One obvious condition on Q is that
it map the classical field expression φ ∈ P to the W -valued distribution φ . A key condition that Q
should satisfy on general elements of P is that the definition of the quantum field observables should
be “local and covariant.” This implements the basic requirement, motivated by general relativity, that
there be no “background structure” appearing in the formulation of QFTCS other than the spacetime
manifold, M , and the metric, g, and that the laws of QFTCS be local in the metric and the quantum
fields.
 QUANTUM FIELDS IN CURVED SPACETIME 29

To define the local and covariant requirement on Q, we first note that our construction of W (M , g) is
local and covariant19 in the following sense: Let (M , g) and (M 0 , g0 ) be globally hyperbolic space-
times and suppose that there is a causality preserving isometric embedding ψ : M → M 0 . Then
there exists a corresponding canonical homomorphism α(ψ) : A 0 (M , g) → A 0 (M 0 , g0 ) defined by
mapping F̂n ∈ A 0 (M , g) as in eq. (76) to the element of A 0 (M 0 , g0 ) obtained by replacing the test
function Fn by ψ∗ Fn , (viewing φ now as the algebra-valued distribution on M 0 ). Since A 0 is dense in
W , α(ψ) uniquely extends to a homomorphism from W (M , g) to W (M 0 , g0 ), which we also denote
as α(ψ).
The desired local and covariant condition on the map Q is that for all test functions f on M and all
causality preserving isometric embeddings ψ : M → M 0 , we have
α(ψ)[Q(Φ)( f )] = Q(Φ)( f ◦ ψ −1 ) . (83)
Note that one cannot tell if the map Q satisfies this local and covariant condition without knowing
how Q is defined on all globally hyperbolic spacetimes.
If Q satisfies (83), then any W (M , g)-valued distribution (defined on all globally hyperbolic (M , g))
lying in the image of Q, will be called a local and covariant quantum field. Thus, Q(φ ) (which we
have been denoting—and will continue to denote—as φ ) is local and covariant. On the other hand,
it is not difficult to show that for any assignment of a Hadamard state ω to each globally hyperbolic
spacetime, the quantum field : φ 2 :ω is not local and covariant. In essence, “preferred states” cannot be
locally and covariantly constructed from the metric, so ω provides additional, unwanted background
structure. Thus, our condition (83) on Q precludes the definition Q(φ 2 ) =: φ 2 :ω for any assignment
of ω to (M , g).
In addition to (83), we can make a list of other properties that Q should satisfy, including (1) ap-
propriate commutation relations of Q(Φ) with φ , (2) appropriate continuous/analytic behavior under
continuous/analytic variation of the metric, and (3) scaling of Q(Φ) that, up to logarithmic terms,
agrees with classical scaling under scaling of the metric. We refer the reader to [46] for a complete
and precise statement of the required properties.
It was proven in [46] that the map Q defined by
Q(Φ) =: Φ :HN (84)
satisfies all of these properties where the “Hadamard normal ordering” operation : · :HN is defined
by the same formula as (77), with the two-point function W2 (x1 , x2 ) of the Hadamard state ω replaced
by a locally and covariantly constructed Hadamard distribution HN (x1 , x2 ) (see (18)). The expansion
order N must be chosen to be greater than the highest derivative in Φ, but is otherwise arbitrary.
Uniqueness up to addition of suitable “local curvature terms” can then be proven recursively, using
the commutation relations. Details of the existence and uniqueness proofs can be found in [46]. A
precise statement of the uniqueness result appears below as a special case (n = 1) of Theorem 2.
We define the extended algebra of field observables, B0 (M , g), for the free field to be the subalgebra
of W (M , g) generated by elements lying in the image of Q, i.e.,
B0 ≡ alg{Q(Φ)( f ) | Φ ∈ P(M , g), f ∈ C0∞ (M )} . (85)
In particular, B0 includes all field observables corresponding to local and covariant polynomial ex-
pressions in φ and its derivatives, such as the stress-energy tensor, Tµν . All Hadamard states, ω, yield
states on B0 and all one-point distributions in Hadamard states, such as the expected stress-energy
tensor ω(Q(Tµν )( · )), are smooth.
19An elegant category theoretic formulation of the notion of a local and covariant theory is given in [16].
30 HOLLANDS AND WALD

Example: Casimir effect in KG-theory on M = R × T3 : We now illustrate the definition of

composite operators by providing a calculation of the Casimir effect. We consider the spacetime
M = R × T3 equipped with the flat metric induced from Minkowski space via the identification
T3 = (R/2πL)3 , and we assume that m2 > 0. The ground state ω0 for the KG-field can e.g. be found
from the corresponding positive frequency modes uk (t, x) = [(2πL)3/2 2ωk/L ]−1 e−iωk/Lt+ikx/L , where
k ∈ Z3 . According to (24), the two-point function is found to be
1 e−iωk/L (t1 −t2 ) ik(x1 −x2 )/L
ω0 (φ (x1 )φ (x2 )) = ∑3 2ωk/L e
(2πL)3 k∈Z
1 e−iωp (t1 −t2 ) ip(x1 −x2 +2πLn) 3
Z
= e d p (86)
∑3
(2π)3 n∈Z 2ωp
= ∑3 D̂∞+(t1 − t2 − i0+, x1 − x2 + 2πLn)
n∈Z
where D∞ + was defined above in eq. (35) and a hat denotes a Fourier transform in the spatial variables;
explicitly  √ 
K m −t 2 + x2 + it0+
m 1
D̂∞
+ (t, x) = √ , (87)
4π 2 −t 2 + x2 + it0+
where K1 is a modified Bessel function. To go to the second line in eq. (86), we used the Poisson
summation formula, which is applicable here in the sense of distributions in t. The above 2-point
function defines a Gaussian state on B0 . We wish to define the composite field φ 2 in this theory and
compute its expectation value in the state ω0 . According to the prescription given above, we need to
know the Hadamard parametrix HN (see (18)) on the spacetime M . Since M is locally the same as
Minkowski space, it is actually locally identical to that in Minkowski space. In fact, we have, with
x2 = −t 2 + x2 , e = (1, 0, 0, 0), and N = 0:
 
1 1 1 2 + 2
H0 (x1 , x2 ) = 2 + m log(x1 − x2 − ie0 ) , (88)
2π (x1 − x2 − ie0+ )2 8
+
which is equal to the singular terms in D̂∞ + (t1 − t2 − it0 , x1 − x2 ). The “local and covariant Wick
power” is then φ 2 =: φ 2 :H0 . Its expectation value is found using the definition of the normal ordering
prescription:
ω0 (φ 2 (x)) = lim (ω0 (φ (x + tξ )φ (x − tξ )) − H0 (tξ , −tξ )) . (89)
t→0
Inserting the explicit formula for the 2-point function of ω0 yields
1
ω0 (φ 2 (x)) = − 2
m2 log m2 + ∑ D̂∞ + (0, 2πLn) (90)
16π n∈Z3 ,n6=0
1 2 2
where it should be noted that the term with n = 0 is cancelled, up to − 16π 2 m log m , by H0 , as

can be seen from the asymptotic expansion of D̂∞ + in eq. (87). As a consequence of the exponential
decay of K1 , the sum on the right side converges rapidly for m2 > 0 and can be efficiently evaluated
numerically. The sum is practically zero for 2πLm  1, i.e. for small Compton wavelength compared
to the identification scale. A precursor of this derivation of the Casimir effect in QFTCS was given
in [55].
The definition of φ 2 is not unique—in fact, according to the previous discussion, the alternative
definition φ 2 → φ 2 + (c1 R + c2 m2 )1 would be equally acceptable, where c1 , c2 are any real constants,
and where R is the scalar curvature. Since R = 0 in the present context, the ambiguity consists in
adding to φ 2 the operator c2 m2 1, which changes ω0 (φ 2 (x)) by c2 m2 . We can fix this ambiguity by
 QUANTUM FIELDS IN CURVED SPACETIME 31

fixing the expectation value for any given but fixed m = m0 , L = L0 . If one is considering a scalar field
of mass m0 , the usual choice of c2 would be c2 = log m20 /16π 2 , so that limL→∞ ω0 (φ 2 ) = 0. Once c2
has been chosen, the expectation value of φ 2 is uniquely determined for any other m, L, and in fact
for any other state, such as a finite temperature state. A similar calculation can be performed for the
energy density operator, and the corresponding expectation value is called the ‘Casimir energy’.

The situation with regard to defining time ordered products of polynomial expressions—as needed to
define perturbative interacting quantum field theory—is similar to the above problem of defining Q,
although it is considerably more complicated. The time ordered product of Φ1 (x1 ), . . . , Φn (xn ) should
be linear in each Φi , so it is natural to view time order products in n factors as a map, Tn , from the
n-fold tensor product P⊗n ≡ P ⊗ ... ⊗ P into distributions in n variables valued in W . We will denote
time ordered products by Tn (Φ1 (x1 ) ⊗ · · · ⊗ Φn (xn )). For one factor, we define

T1 (Φ) = Q(Φ) . (91)

In order to make the notation less cumbersome, we will omit writing Q (as we have already been
doing for the quantum field φ ) when it is clear that we are referring to the quantum field Q(Φ) rather
than the classical polynomial Φ, so, e.g., we will write T1 (Φ) = Φ.
We would like to define Tn by “time ordering” the product of fields, so, e.g., for n = 2 we would like
to set
(
/ J − (x2 )
Φ1 (x1 )Φ2 (x2 ) if x1 ∈
T2 (Φ1 (x1 ) ⊗ Φ2 (x2 )) = (92)
/ J − (x1 ) .
Φ2 (x2 )Φ1 (x1 ) if x2 ∈

(If x1 and x2 cannot be connected by a causal curve, then Φ1 (x1 ) and Φ2 (x2 ) commute, so either
formula may be used.) The problem is that Φi (xi ) are distributions, so (92) provides a definition
of T2 (Φ1 ( f1 ) ⊗ Φ2 ( f2 )) only when the supports of the test functions f1 and f2 satisfy the relations
supp f1 ∩ J − [supp f2 ] = 0/ or supp f2 ∩ J − [supp f1 ] = 0.
/ It is not difficult to see that this enables us to
straightforwardly define T2 (Φ1 ( f1 ) ⊗ Φ2 ( f2 )) whenever supp f1 ∩ supp f2 = 0. / However, (92) makes
no sense when supp f1 ∩ supp f2 6= 0. / Thus, we must extend the distribution (92) to the “diagonal” x1 =
x2 . This may seem like a relatively trivial problem, but the extension of the definition of general time
ordered products Tn (Φ1 (x1 ) ⊗ · · · ⊗ Φn (xn )) to the “total diagonal” x1 = · · · = xn is the main problem
of renormalization theory in flat and curved spacetime [9,24]. We refer to a definition/construction of
Tn as a “renormalization scheme.”
We proceed, as in the construction of Q above, by writing down a list of properties that Tn should
satisfy. We have already required T1 = Q. We require Tn to satisfy the appropriate generalization of
(92) (involving lower order time-ordered products) away from the total diagonal. As in the case of
Q, we require Tn to be locally and covariantly defined, to satisfy appropriate commutation relations
with φ , to have appropriate continuous/analytic dependence on the metric, and to have appropriate
scaling behavior (up to logarithmic terms) under scaling of the metric. We also require commutation
of Tn with derivatives. Finally, we impose a number of additional conditions on Tn , specifically, a
microlocal spectrum condition, a “unitarity” condition, and conditions that guarantee that the perturb-
atively defined interacting field (see subsection 3.2 below) (i) satisfies the interacting field equation
and (ii) has a conserved stress-energy tensor. The microlocal spectrum condition is given by eq. (149)
of appendix A. The unitarity condition is a version of the “optical theorem” and it guarantees that the
algebra of interacting fields has representations on a positive definite Hilbert space. It is formulated as
32 HOLLANDS AND WALD

follows: Let T̄n (⊗i Φi (xi )) = [Tn (⊗i Φi (xi )∗ )]† be the ‘anti-time-ordered’ product20. Then we require
T̄n ⊗ni=1 Φi (xi ) = ∑ (−1)n+ j T|I1| ⊗i∈I1 Φi(xi) . . . T|I j | ⊗ j∈I j Φ j (x j ) , (93)




I1 t ... t I j = n

where the sum runs over all partitions of the set n = {1, . . . , n} into pairwise disjoint subsets I1 , ..., I j .
The condition that guarantees satisfaction of the field equation for the interacting field is a version of
the “Schwinger Dyson equation”, and reads
 n  n  
2
O δ Φk (xk )
Tn+1 (g − m )φ (y) ⊗ Φi (xi ) = i ∑ Tn Φ1 (x1 ) ⊗ · · · ⊗ · · · Φn (xn ) . (94)
i=1 k=1 δ φ (y)
The condition guaranteeing conservation of the interacting field stress-energy is discussed in [49].
We refer the reader to [46], [47], and [49] for a more complete and extensive discussion of all of the
conditions imposed on Tn .
It was proven in [47]—key parts of which were based on [15]—that there exists a definition of Tn , that
satisfies all of the above conditions. Furthermore, Tn is unique up to “appropriate local and covariant
counterterms.” To explain this freedom in the choice of Tn , we must introduce a considerable amount
of additional notation: We denote by P(M n ) the space of all distributional local, covariant functionals
of φ (and its covariant derivatives), of g, and of the Riemann tensor (and its covariant
R derivatives),
which are supported on the total diagonal (i.e. of delta-function type). Let F = λ f Φ be an integrated
local functional Φ ∈ P(M ), and formally combine the time-ordered functionals into a generating
functional written
∞
1
T(exp⊗ (F)) = ∑ Tn (F ⊗n ) ∈ W [[λ ]] , (95)
n=0 n!
where exp⊗ is the standard map from the vector space of local actions to the tensor algebra over the
space of local action functionals, and W [[λ ]] denotes the algebra of formal power series expressions
in W . Let D denote a hierarchy Dn of linear functionals
Dn : P(M ) ⊗ · · · ⊗ P(M ) → P(M n ) . (96)
We similarly write D(exp⊗ (F)) for the corresponding generating functional obtained from Dn . We
then have the following theorem:
Theorem 2. [46] [47], [48] (Uniqueness) If Tn and T̂n are two different renormalization schemes,
both satisfying our conditions, then they are related by
 

T̂(exp⊗ (iF)) = T exp⊗ iF + iD(exp⊗ F) . (97)
R
for any F = λ f Φ, Φ ∈ P(M ) and f ∈ C0∞ (M ). The functionals Dn are maps as specified in (96)
and satisfy:
(i) D(eF⊗ ) = O(h̄) if we reintroduce h̄.
(ii) Each Dn is locally and covariantly constructed from g.
(iii) Each Dn is an analytic functional of g.
(iv) Each Dn (Φ1 (x1 ) ⊗ · · · ⊗ Φn (xn )) is a distribution that is supported on the total diagonal (=
‘contact term’ = ‘delta-function type’).
(v) The maps Dn are real.
(vi) Each Dn is symmetric.
(vii) Each Dn satisfies the natural dimension constraint.
(viii) Derivatives can be pulled into Dn .
20It can be shown that the anti-time-ordered product in satisfies the causal factorization property with the reversed
time-orientation.
 QUANTUM FIELDS IN CURVED SPACETIME 33

Conversely, if Dn has these properties, then any T̂ given by (97) defines a new renormalization scheme
satisfying our conditions.

The expressions Dn corresponds the “counterterms” that characterise the difference between the two
renormalization schemes. For the case Φ = φ 4 , relevant for the interaction Lagrangian (75), the
expression
λn
Z
δ L (y) = ∑ Dn (φ 4 (y) ⊗ φ 4 (x1 ) ⊗ · · · φ 4 (xn−1 ))dv1 ...dvn−1 (98)
n≥0 n! M n−1

corresponds to the finite counterterms in the Lagrangian that arise from a change of the renormaliza-
tion scheme, with the λ n -term corresponding to the contribution at n-th order in perturbation theory.
From the properties of the maps Dn , these will correspond to linear combinations of all possible local
covariant expressions of dimension 4, i.e. the monomials already present in the original Lagrangian
L , together with linear combinations of Rφ 2 and the ‘C-number’ terms m4 , m2 R, Rµν Rµν , .... For
other Φ’s, the Dn likewise characterize the ambiguities of composite interacting fields, which are
defined below. The following examples illustrate the allowed ambiguities in the choice of renormal-
ization scheme.

f φ 2 dV, the formula (97) gives

R
Example: For n = 1, and F =
T̂1 (φ 2 (x)) = T1 (φ 2 (x)) + D1 (φ 2 (x)) . (99)
and the ambiguity must have the form
D1 (φ 2 (x)) = b1 m2 + b2 R(x) , (100)
where b1 , b2 are real numerical constants. At the next order n = 2, the formula (97) gives
T̂2 (φ 2 (x) ⊗ φ 2 (y)) = T2 (φ 2 (x) ⊗ φ 2 (y)) + T1 (D2 (φ 2 (x) ⊗ φ 2 (y))) . (101)
Since the scaling degree (see appendix A) of the delta function is 4 and the dimension of φ is 1, the
conditions on D2 stated in Theorem 2 imply that it must take the form
D2 (φ 2 (x) ⊗ φ 2 (y)) = c0 δ (x, y) , (102)
for some real constant c0 . Similarly
D2 (φ 3 (x) ⊗ φ 3 (y)) = c1 δ (x, y)φ 2 (y) + (c2 R + c3 g )δ (x, y) , (103)
because the scaling degree of g δ (x, y) is 6, and the dimension of R is 2. Condition (viii) of the
theorem implies that c1 = 9c0 . An example with 3 factors is
D3 (φ 2 (x) ⊗ φ 3 (y) ⊗ φ 3 (z)) = c4 δ (x, y, z) , (104)
because the scaling degree of the delta function with three arguments is 8.

3.2. The algebra BI of interacting fields. We are now in a position to perturbatively define the
composite fields for the interacting field theory described by the Lagrangian density (73), and the
algebra BI of which they are elements. Recall that we have already defined B0 , which corresponds
to the case when interaction is turned off, λ = 0; see eq. (85). In the following, we denote the
composite fields of the free theory by Φ0 ( f ) ≡ Q(Φ)( f ) (i.e., we add the subscript “0"), in order to
distinguish them from the corresponding composite fields of the interacting theory, ΦI ( f ), which we
will define below.
The basic idea to construct the interacting fields, ΦI ( f ), is to initially “turn off” the interaction at
some finite time in the past, so that ΦI = Φ0 at sufficiently early times. We then evolve ΦI forward
in time into the region where the interaction is fully turned on. Finally, we take a limit where the
34 HOLLANDS AND WALD

“turn-on time” of the interaction is arbitrarily far in the past. However, this last step is problematical
because this limit need not exist. This difficulty will be overcome by modifying the limit so that rather
than fixing ΦI = Φ0 in the asymptotic past, we fix ΦI in regions of increasing size in the interior of
the spacetime.
To implement this strategy to define ΦI ( f ), we choose a cutoff function, θ , of compact support
on M which is equal to 1 on an open neighborhood of some globally hyperbolic open region V
R the property that Σ ∩ V is a Cauchy surface for V for some Cauchy surface Σ in M . For F =
with
λ M Φ(x) f (x) dvg with f ∈ C0∞ (M ), we define the local S-matrix to be
in
S(F) = T(exp⊗ (iF)) ≡ ∑ n! Tn(F ⊗ · · · ⊗ F) . (105)
n≥0

We define the relative S-matrix with respect to the cut-off interaction, L1 (θ ) = θ L1 , by

R
M

SL1 (θ ) (F) = S(L1 (θ ))−1 S(F + L1 (θ )) . (106)

Then the interacting field, for the interacting theory with cutoff interaction L1 (θ ) corresponding to
Φ is defined by [9]
1d
Φ( f )L1 (θ ) ≡ S (tΦ( f )) . (107)
i dt L1 (θ ) t=0
Equations (105)-(107) are to be understood as formal series expressions that define the interacting
field to any finite order in perturbation theory; no convergence properties are claimed. Note that the
definition of Φ(x)L1 (θ ) has been adjusted so that it coincides with the corresponding free field Φ(x)
before the interaction is “switched on,” can be seen explicitly by expressing it in terms of “totally
retarded products” [15, 48].
We now wish to remove the cutoff. Formula (107) will not, in general, make sense if we straightfor-
wardly attempt to take the limit θ → 1. Indeed if θ could be set equal to 1 throughout the spacetime
in eq. (107), then the resulting formula for Φ( f )L1 (1) would define an interacting field in the sense of
Bogoliubov [9], with the property that the interacting field approaches the free field in the asymptotic
past. However, even in Minkowski spacetime, it is far from clear that such an asymptotic limit of
the interacting field will exist (particularly for massless fields), and it is much less likely that any
such limit would exist in generic globally hyperbolic curved spacetimes that do not become flat in the
asymptotic past.
To remove the cutoff in such a way that the limit does exist, we need to know how the fields (107)
change under a change of the cutoff function θ . If θ1 and θ2 are two cutoff functions, each of which
are 1 in an open neighborhood of V as above, then there exists a smooth function h− of compact
support on M which is equal to θ1 − θ2 on the causal past of the region V , and whose support does
not intersect the causal future of V . The unitary U(θ , θ 0 ) defined by
U(θ1 , θ2 ) = SL1 (θ1 ) (L1 (h− )) (108)
is then independent of the particular choice for h− , and one has [14, thm. 8.6]
U(θ1 , θ2 ) Φ( f )L1 (θ ) U(θ1 , θ2 )−1 = Φ( f )L1 (θ2 ) , (109)
for all fields Φ and all smooth scalar densities f of compact support in V .
Following the ideas of [14], we now take a limit where the field remains fixed in regions of increasing
size in the interior of the spacetime. The construction makes use of the following geometric fact (see
lemma 3.1 of [48]): Let (M , g) be a globally hyperbolic spacetime. Then there exists a sequence of
compact sets, {Kn }, with the properties that (i) for each n, Kn ⊂ Vn+1 , where Vn+1 ≡ int(Kn+1 ) (ii)
 QUANTUM FIELDS IN CURVED SPACETIME 35

∪n Kn = M , and (iii) for each n, Vn is globally hyperbolic and Σ ∩Vn is a Cauchy surface for Vn , where
Σ is a Cauchy surface for M . For each n, let θn be a smooth function with support contained in Kn+1
such that θn = 1 on an open neighborhood of Kn . Let U1 = 1 and let Un = U(θn , θn−1 ) for all n > 1,
where U(θn , θn−1 ) was defined in eq. (108) above. Our definition of the interacting field is:
ΦI ( f ) ≡ lim U1U2 . . .Un Φ( f )L1 (θn ) Un −1Un−1
−1
. . .U1 −1 . (110)
n→∞
The existence of the limit follows from the properties of the time-ordered products, see [48].
The meaning of the sequence under the limit in eq. (110) for n = 1, 2, . . . , is easily understood as
follows. Since U1 = 1, the first element of this sequence is just the Bogoluibov formula for this
interacting field quantity with cutoff function θ1 . The second element of this sequence modifies
the Bogoliubov formula with cutoff function θ2 in such a way that, according to eq. (109) above,
the modified Bogoliubov formula with cutoff function θ2 agrees with the unmodified Bogoliubov
formula with cutoff function θ1 when the supports of all of the fi are contained within K1 . For
the third element of the sequence, the unitary map U3 first modifies the Bogoliubov formula with
cutoff function θ3 so that it agrees in region K2 with the Bogoliubov formula with cutoff function
θ2 . The action of the unitary U2 then further modifies this expression so that it agrees in region K2
with the modified Bogoliubov formula of the previous step. In this way, we have implemented the
idea of “keeping the interacting field fixed in the interior of the spacetime” as the cutoff is removed.
A particularly simple expression for the interacting field φI (x) is hence obtained if x ∈ K1 , because
then the unitaries Un are absent in the formula. Unraveling the definitions, we find that, for the case
L1 = φ 4 , the following formula holds for x ∈ K1
n n+m  n+m
(iλ )n+m
Z O  
4
φ 4 (yk ) ∏ θ1 (y j )dv j .
O
ΦI (x) = ∑ Tn φ (y j ) Tm+1 Φ(x) ⊗ (111)
m,n≥0 n!m! j=1 k=n+1 j=1

The interpretation of ΦI as interacting fields is reinforced by the fact that the interacting field equation
(g − m2 )φI (x) = 4λ φI3 (x) (112)
holds to all orders in λ , as a consequence of (94).
As before (see (95)), we denote by W (M , g)[[λ ]] the algebra of formal power series expressions in
W (M , g). We define the interacting field algebra BI (M, g) to be the subalgebra of W (M , g)[[λ ]]
generated by the interacting fields, namely
BI = alg{ΦI ( f ) | Φ ∈ P(M , g), f ∈ C0∞ (M )} . (113)
This definition of BI (M , g) as a subalgebra of W (M , g)[[λ ]] depends on the choice of a family of
compact sets, Kn , described above, as well as the choice of cutoff functions, θn . If we were to choose
a different family, K̃n , of compact sets and a corresponding different family, of cutoff functions, θ̃n ,
we will obtain a different subalgebra, B̃I (M , g). However, BI (M , g) and B̃I (M , g) can be shown
to be isomorphic, and hence, in so far as the abstract algebra of interacting fields is concerned, these
choices are irrelevant.
Example: Interacting field in KMS-state: As an illustration of these definitions we consider finite
temperature states (aka KMS-states) in Minkowski space for self-interacting fields. Such finite tem-
perature states in Minkowski spacetime should be good approximations to physically relevant states
of interacting quantum fields in the early universe in situations where the “equilibration time” is much
smaller than the Hubble time.
For finite temperature states in Minkowski spacetime, it has long been observed that naive perturbative
expansions (replacing e.g. the ordinary “vacuum” Feynman propagator by a “thermal” propagator)
lead to notorious infra-red divergences [2]. The root cause of these divergences is the fact that such
36 HOLLANDS AND WALD

naive treatments basically amount to defining a “free in-field” in a thermal bath, see e.g. [58] for a
discussion. We have already argued above that this is inappropriate in a general curved spacetime, but
here we see that it is also inappropriate even in flat space if we want to consider e.g. thermal states.
However, one can set up a perturbation theory for KMS-states of the interacting field which is free
of these infra-red divergences if we follow the ideas presented in the previous subsection together
with standard ideas from statistical mechanics concerning the construction of perturbed KMS-states.
We describe these first in a quantum mechanical setting; for details see [4, 11]. Consider a sys-
tem with Hamiltonian H0 , so that the Heisenberg time-evolution of an observable a is given by
α0,t (a) = eitH0 ae−itH0 . We assume that tr(e−β H0 ) < ∞, so that a KMS state can be defined by
ω0 (a) = tr(ae−β H0 )/ tr(e−β H0 ). Consider now a perturbed Hamiltonian H0 + λV , with sufficiently
well-behaved V such that the corresponding perturbed time evolution αI,t and KMS-state ωI (relative
to the perturbed time-evolution) can be defined as above. The original and perturbed time-evolution
are related by
αI,t (a) = U(t)α0,t (a)U(t)−1 , (114)
for a 1-parameter family of unitary operators U(t), characterized by the conditions
d
U(t + s) = U(t)α0,t (U(s)) , U(0) = 1 , U(t) = iλV . (115)
dt t=0

The first condition is called a “cocycle condition”21, and λV is the generator of this cocycle. The
original and perturbed KMS-state are then related through [5]
ω0 (a U(iβ ))
ωI (a) = (116)
ω0 (U(iβ ))
and U(iβ ) is given by
∞ Z
−β (H0 +λV ) β H0 n
U(iβ ) = e e = ∑ (−λ ) α0,iun (V ) · · · α0,iu1 (V ) d n u , (117)
n=0 0≤un ≤...≤u1 ≤β

as a consequence of (115).
We now wish to apply these basic formulas to the case where α0,t is the time evolution in the linear
KG quantum field theory on a static spacetime, ω0 a Gaussian KMS state in that theory, whereas αI,t
is the time evolution in the interacting KG quantum field theory [29]. The desired KMS state ωI in
the interacting quantum field theory is then described by the right side of eq. (116). For simplicity22,
we consider the flat spacetime M = R × T3 with periodically identified spatial directions, T3 =
(R/2πLZ)3 , L > 0, and we also let m > 0. As described above, the observables in the free theory are
the elements of B0 , whereas observables in the interacting theory are elements of BI . The free and
interacting time-evolutions are specified by
α0,t (Φ0 ( f )) = Φ0 ( f t ) , αI,t (ΦI ( f )) = ΦI ( f t ) , (118)
where f t (s, x) = f (s−t, x), i.e. they simply define the time-translate the free and interacting fields, re-
spectively. α0,t resp. αI,t are seen to be automorphism groups of B0 resp. BI . We wish to define first

21In this context, one sometimes writes U(t) = [Dω : Dω ] and refers to U(t) as the “Radon-Nikodym” derivative.
I 0 t
These quantities play a profound role in Tomita-Takesaki theory in the context of v. Neumann algebras.
22 The interacting thermal KMS state on Minkowski space M = R1,3 can be obtained e.g. in the end by taking L → ∞
in formula (122). The resulting limit is shown to exist for any order in perturbation theory, the essential point being that
the higher terms n 6= 0 in eq. (119) vanish in that limit when m > 0; details are given in [29].
 QUANTUM FIELDS IN CURVED SPACETIME 37

a KMS-state ω0 for the free theory. This state can be computed by simply going to the standard va-
cuum representation of B0 , and then by forming the density matrix as ω0 (a) = tr(ae−β H0 )/ tr(e−β H0 ),
where H0 is the free field Hamiltonian. The resulting 2-point function turns out to be given by:
1
∑3 D+(t1 − t2, k/(2πL)) eik(x1−x2)/L
β
ω0 (φ0 (x1 )φ0 (x2 )) =
(2πL)3 k∈Z
(119)
D̂+ (t1 − t2 − i0+ , x1 − x2 + 2πLn)
β
= ∑
n∈Z3

β
where D+ was defined above in eq. (35) and a hat denotes a Fourier transform in the spatial variables.
To go to the second line, we again used the Poisson summation formula, which is applicable here in
the sense of distributions in t. Comparing with eq. (119), it shows that the thermal 2-point function on
the torus is again obtained by an “image sum”23. The expectation value ω0 (a) for a general a ∈ B0 is
determined by the above 2-point function, we omit the details.
The desired KMS-state for the interacting theory, ωI is now given by equations (116), (117). Due to
the Gaussian nature of the state ω0 , one can show that eq. (116) can also be rewritten as in terms of
c [see eq. (19)] as
the connected parts ω0,n
∞ Z  
n c
ωI (a) = ∑ (−λ ) ω0,n+1 a,V (iun ), . . . ,V (iu1 ) d n u , (120)
n=0 0≤un ≤...≤u1 ≤β

where a is any observable in the interacting field algebra BI and V (iu) ≡ α0,iu (V ). To obtain the
n-point functions of the field in the interacting theory in our thermal state, we would take a =
φI (t1 , x1 ) · · · φI (tn , xn ), where the interacting field has been defined in the previous section. To make
concrete perturbative computations, one needs to unravel these definitions. In the present spacetime
with compact Cauchy surface Σ = T3 , we may take the sequence of cutoff functions to be simply
θ j (t, x) = θ (t/ j), where θ is e.g. a smooth function such that θ (t) = 1 if |t| ≤ T and = 0 if |t| > 2T ,
where T is taken sufficiently large to include the time interval of interest for calculations.
One must first identify the generator V appearing in eq. (117) using the definition (115). It is found
by comparing αI,t and α0,t in eq. (118), and may be expressed as
Z
V= φI4 (t, x) θ̇− (t) , (121)
R×T3

where θ̇− (t) = 1[0,−∞] (t) dtd θ (t) . Note that λV is similar to the naive expression for the interaction
Hamiltonian, but that it is averaged 24 also over time. This smearing effectively avoids the UV-
divergences in the naive Hamiltonian formulation of interacting quantum field theories. Also, note
that φI4 is the interacting field, and therefore itself an infinite formal power series in the coupling.
By further unraveling the definitions of the interacting fields in (121) and (120), one obtains a com-
pletely explicit (in principle) expression for the interacting KMS-state ωI (a) as given in eq. (120) in
terms of time-ordered products and the KMS-state in the free KG-theory. For a given set of times
t1 , ...,tn in a = φI (t1 , x1 ) · · · φI (tn , xn ), we obtain the simplest expression if we take T sufficiently large

23 This feature is characteristic for ground or KMS-states on Cauchy surfaces that are quotients such as R3 /Γ, H3 /Γ,
etc.
24Note that θ̇ ≥ 0 and R θ̇ dt = 1.
− −
38 HOLLANDS AND WALD

−T +T
Re(τ)
iu3 − i0+ iu3

iu2 − i0+ iu2

iu1 − i0+ iu1

iβ

Im(τ)

F IGURE 6. Shown is the ordering, indicated by arrows, of the complex ‘times’ τ =

sl + iu j ∈ C appearing in the last expression in (122). The number of zigzags is equal
to N, and for simplicity, we formally take a sharp cutoff function θ (t) = 1[−T,T ] (t) so
that the contours oscillate between ±T . The ordering is related to that in the “real-time
path” formalism, see e.g. [58]. The correspondence is formally seen by deforming the
above contour to the“Schwinger-Keldysh path” in the real time path formalism.

such that |ti | < T for all i. In that case, the interacting field is given by eq. (111). We find altogether
n N
(−λ )N
Z Z Z
n N 3N
ωI (a) = ∑ ∑ d u d s d y ∏ − ik ∏ θ (s j )
θ̇ (s )
n,N≥0 partitions of {1,...,N} ∏ j |I j |!|J j |! S T3N
n RN k=1 j=1
 n  (122)
4 + 4
O
c
× ω0,n+1 a, T̄|I j | (⊗l∈I j φ (sl + iu j − i0 , yl ))T|J j | (⊗k∈J j φ (sk + iu j , yk ))
j=1

On the right side, we mean partitions I1 ∪ J1 ∪ ...In ∪ Jn = {1, ..., N} which are order preserving and
such that J j 6= 0,
/ whereas
Sn = {0 ≤ un ≤ ... ≤ u1 ≤ β } (123)
is an n-dimensional simplex. The label ik is the first number in Jk , and T resp. T̄ are the time-
ordered resp. anti-time ordered products described above in eq. (93). The structure of (122) becomes
considerably more transparent if we observe that the sequence of complexified time-coordinates in
the last line are ordered in a characteristic fashion, which is due to the pattern of time- and anti-time
ordered products, and the simplex Sn , see fig. 6.
Formula (122) is our desired perturbation series for the KMS-state in the interacting theory: All
quantities on the right hand side, such as free KMS state ω0 [see (119)], and the time-ordered products
are quantities that are computable in the free theory. Since all ingredients in the final formula for
ωI (a) are well defined as distributions, and since all integrations are over a compact domain, (122) is
manifestly free of any UV and IR divergences! Furthermore, unlike many other approaches to thermal
QFT, our derivation was based entirely on the basic principles of statistical mechanics, combined with
the principles of perturbative QFTCS laid out in the previous section.

An important and nontrivial feature of the interacting field algebra BI (M , g) in any globally hyper-
bolic spacetime is its behavior under scaling of the spacetime metric. Classically, for any polynomial
Lagrangian, L , we can define a scaling of the field and the coupling constants so that the action is
 QUANTUM FIELDS IN CURVED SPACETIME 39

invariant under scaling of the spacetime metric, g → Λ2 g. For example, for the Lagrangian
L = 21 ((∇φ )2 + m2 φ 2 + ξ Rφ 2 ) + λ φ 4 (124)
the associated classical scaling is φ → Λ−1 φ , m → Λ−1 m, ξ → ξ , and λ → λ . Any monomials in
φ and its derivatives can similarly be assigned a scaling behavior. However, this classical scaling
behavior cannot be maintained in the interacting quantum field theory. Indeed, in our construction
of composite fields and time-ordered products for the free field theory, we already noted that the
classical scaling of these fields can be maintained only “up to logarithmic corrections,” thus implying
a nontrivial behavior under scaling for the interacting fields. Nevertheless, the following result can
be proven [48]: Let f be the collection of coupling parameters of the theory—such as, in our above
example, m2 , ξ , and λ . Then there exists an isomorphism ρΛ : BI (M , Λ2 g, f ) → BI (M , g, f (Λ)),
which maps local fields to local fields. Thus, a change of scale (distances) can be compensated by a
corresponding change of the coupling constants, thereby leaving the theory unchanged. This change
of the coupling constants, f 7→ f (Λ) is called the renormalization group flow. The flow can be
calculated order-by-order in perturbation theory. In our example theory, the renormalization group
flow takes place in the 3-dimensional space of coupling constants m2 , ξ , λ . Theories such as this
one where the flow takes place in a finite dimensional space of coupling constants are traditionally
called “renormalizable”. By contrast, theories where the flow necessarily takes place in an infinite
dimensional space of coupling constants are called “non-renormalizable”.
In Minkowski spacetime, the dilations (i.e., the diffeomorphisms associated with a rescaling of global
inertial coordinates) are conformal isometries with a constant conformal factor, so the renormalization
group flow can be reformulated in terms of the behavior of physical quantities, such as correlation
functions of the vacuum state, under rescaling of coordinates (or, equivalently, rescaling of momenta).
However, in a general curved spacetime, there will not be any conformal isometries of any kind, so
renormalization group flow must be defined in terms of behavior under scaling of the spacetime
metric, as described above. We emphasize that our definition of the renormalization group flow does
not involve any preferred states (which do not exist in a general, curved spacetime), nor the idea of
‘cutoffs’. Cutoffs play an important role in a somewhat different concept of renormalization group
flow in flat space often called “Wilsonian RG-flow”. We do not see any canonical way of imposing a
cutoff in general Lorentzian curved spacetime, and furthermore, any cutoff is in fundamental conflict
with the idea of defining the theory in a “local and covariant manner”.

3.3. Yang-Mills fields. For simplicity, up to this point we have restricted our discussion QFTCS
to scalar fields. However, there are other types of fields that occur in nature25, namely spinor and
gauge (Yang-Mills) fields. To treat spinor fields, one must replace the canonical commutation algebra
A of section 2.1 by a corresponding “canonical anti-commutation algebra”. However, aside from
this important change, spinor fields can be dealt with without introducing any major new conceptual
innovations. In particular, analogs of W , B0 , and BI can be defined for spinor fields in close parallel
with the scalar case; see e.g. [21] for details.
However, significant conceptual innovations are needed for the construction of the quantum field
theory of gauge fields. All of the difficulties are associated with the fact that the physical observ-
ables in a gauge theory are not the gauge fields themselves but equivalence classes of these fields
under gauge transformations. In particular, the classical equations of motion for gauge fields are not
deterministic—they do not determine the gauge—so one does not have unique advanced and retarded

25Indeed, the existence of a scalar field in nature has only recently been confirmed by the discovery of the Higgs
particle.
40 HOLLANDS AND WALD

propagators. One obvious idea would be to fix the gauge, thereby making the classical dynamics de-
terministic. However, the gauge cannot be (completely) fixed in a local and covariant manner, thereby
making it extremely difficult to impose the requirement that the theory be local and covariant26. An
alternative, relatively straightforward, strategy can be employed for the case of a free (i.e., Maxwell)
gauge field, Aµ , to construct a quantum field theory at the level of the algebra A : Instead of smearing
Aµ with arbitrary test vector fields f µ , one defines Aµ ( f µ ) only for f µ that satisfy
∇µ f µ = 0 . (125)
R R
Such observables are invariant under the gauge transformation δ A = dχ, since ∇µ χ f µ = − χ∇µ f µ =
0. Indeed, the restriction to smearing with such f µ is equivalent to working with the Maxwell field
tensor Fµν = 2∇[µ Aν] as the fundamental observable. This yields a satisfactory theory at the level of
the algebra A , see e.g. [25, 27, 65] for details. But it is far from clear how to construct an analog of
BI for models with interaction, such as Yang-Mills theory, by proceeding this manner.
Instead, we shall we shall proceed by an elegant trick for constructing gauge invariant observables
for Yang-Mills theory by introducing additional fields—the “Fadeev-Popov ghost fields,” c and c,
together with an “auxiliary field” b—and then eliminating the extra degrees of via the imposition
of a BRST-symmetry condition. The observables of the resulting theory are equivalent to the gauge
invariant observables of the original Yang-Mills theory. As has been shown in [40], one can implement
the procedure consistently obtaining in the end an algebra of gauge invariant quantum observables BI
in perturbative interacting Yang-Mills theory.
We now very briefly outline this construction. The dynamical field in Yang-Mills theory is a connec-
tion, Dµ , on a principal fiber bundle with Lie group G. For simplicity, we consider the trivial bundle
M × G, so that we may write
Dµ = ∇µ + iλ Aµ (126)
where Aµ is a one-form field on M valued in the Lie algebra, g, of G. We have introducted a coupling
constant λ in anticipation of the fact that we will later perform a perturbative expansion of the theory.
The field strength Fµν is the g-valued two form given by
i
F = dA + iλ [A, A] (127)
λ
The Yang-Mills Lagrangian is given by
1
L = hFµν , F µν i , (128)
4
where h . , . i denotes the Cartan-Killing form of g, and we assume that G is compact and semi-
simple, so that this Cartan-Killing form is positive-definite. This Lagrangian is invariant under the
infinitiesimal gauge transformations
δ A = dχ − iλ [χ, A] = Dχ , (129)
where χ is an arbitrary g-valued scalar field on M .
We now introduce additional g-valued scalar fields c, c, and b. The fields c, c̄ are independent and
are Grassmann-valued (as well as g-valued). We define an infinitesimal “BRST-transformation,” Q,
acting on the fields (A, c, c, b) by
QA = Dc , Qc = − 21 iλ [c, c] , Qc̄ = b , Qb = 0 . (130)
26The corresponding difficulty for gauge fields in Minkowski spacetime is that the gauge cannot be completely fixed
in a Poincare invariant manner.
 QUANTUM FIELDS IN CURVED SPACETIME 41

Note that the transformation law for c is exactly analogous to the gauge transformation (129), i.e.
c may be thought of as a Grassmann-valued counterpart of the gauge transformation function χ.
The Grassmann-property (and Jacobi-identity) are used to show that Q is nil-potent, Q2 = 0. Note
that Q changes the “Grassmann-grading”, i.e. maps Grassmann-even expressions to Grassmann-odd
expressions, and vice versa. We now modify the Yang-Mills Lagrangian (128) to
Lˆ = L + QΨ (131)
where the density Ψ is given by
Ψ = h∇µ Aµ , c̄i − 21 thb, c̄i (132)
with t ∈ (0, 1]. The choice t = 1 is called the “Feynman gauge,” while the limit t → 0 is called the
“Landau gauge.” Note that the new Lagrangian Lˆ is BRST-closed, i.e. QLˆ = 0: Indeed, QL = 0
because the action of Q on expressions involving only A corresponds exactly to a gauge transform-
ation (129), and Q(QΨ) = 0 because Q is nil-potent. A key difference between Lˆ and L is that
Euler-Lagrange equations arising from Lˆ are hyperbolic for A, c, and c̄ and are algebraic in b, so the
classical dynamics defined by Lˆ is deterministic. Of course, the theory defined by Lˆ is a different
theory, and we are interested in the theory defined by L . Nevertheless, there is a direct relationship
between gauge invariant observables in the original Yang-Mills theory defined by L and observables
in the new theory defined by Lˆ : The gauge invariant polynomials Φ = f (Dk1 F, . . . , Dkn F) (with f
an invariant multi-linear form on g) are BRST-invariant since on such observables, Q acts just like a
gauge-transformation. Futhermore, modulo BRST-exact observables, these are all observables with
vanishing “ghost number” in the new theory, see e.g. [6]. This suggests the following strategy to
define the quantum theory of a Yang-Mills field: First define the algebra of interacting fields B̂I of
the theory corresponding to Lˆ . This should be possible, since the Euler-Lagrange equations for Lˆ
do not suffer from the gauge indeterminism of the theory defined by L . Then define a quantum
version of Q on B̂I which acts as a nilpotent, graded, derivation. The cohomology (closed modulo
exact elements) at ghost number zero should naturally be viewed as the algebra of gauge invariant
polynomial observables of the original theory.
The difficult step is to define a quantum version of Q. The point is that such an object must be
defined so as to be compatible with renormalization. The outline of this construction is as follows;
further details and discussion can be found in [40]. First, one introduces, essentially for the sake of
a sophisticated bookkeeping, for each dynamical field in the gauge fixed theory Φ a corresponding
non-dynamical field called an “anti-field” denoted Φ̂. The anti-field has opposite form degree, Grass-
mann number etc. A bracket ( , ) is introduced between fields and anti-fields as if the anti-field was
“canonically conjugate” to the original field, and Q is extended to anti-fields in such a way that QΦ̂
is equal to theR equation
R
of motion for Φ. Finally, define Q̂ = Q − (QΨ, · ) and define a new action
Ŝ as Ŝ = S + QΨ − (QΦ, Φ̂), where S is the Yang-Mills action. Q̂ is still nil-potent. Then we
split the action Ŝ into a free and an interacting part, Ŝ = Ŝ0 + Ŝ1 , where Ŝ0 corresponds to λ = 0.
The BRST-transformation of the free theory is called Q̂0 . One can write Q̂ = Q̂0 + (Ŝ1 , · ), and the
BRST-invariance of the action Ŝ is expressed by the functional “zero-curvature relation”
Q̂0 Ŝ1 − 21 (Ŝ1 , Ŝ1 ) = 0 , (133)
Finally, we can also define an algebra W (M , g) for the dynamical fields Φ that is constructed by
analogy to the algebra for the scalar field, and the anti-fields by definition behave as “C-numbers”.
Q̂0 then turns out to be a graded derivation of that algebra. The relationship between this graded
derivation and any renormalization prescription taking values in W —satisfying straightforward ana-
logs of the conditions imposed we imposed for the case of a scalar field—is then expressed by the
42 HOLLANDS AND WALD

fundamental relationship, called “master Ward identity”27:

     
iG/h̄ iG/h̄ iG/h̄
Q̂0 T e⊗ = h̄i T (Q̂0 G − 21 (G, G)) ⊗ e⊗ + h̄i T A(eG
⊗ ) ⊗ e⊗ , (134)
R
where G = f Φ for any smeared local field with Φ ∈ P(M ), and any smooth f with compact support.
The anomaly is a series
1
A(eG ⊗) = ∑ An (G⊗n ) , (135)
n≥0 n!
where An obeys properties analogous to the functionals Dn [see thm. (2)].
It turns out that, in order to define a nil-potent quantum BRST-derivation on B̂I , one needs a prescrip-
tion having A = A(exp⊗ Ŝ1 ) = 0. To see this in an informal way, take G = Ŝ1 in the master Ward
identity, and assume that A = 0. Using also the zero-curvature equation, it then formally follows that
Q̂0 T(exp iŜ1 /h̄) = 0, i.e. BRST-invariance of the formal scattering matrix. As explained, we do not
expect the formal scattering matrix to make sense, but the condition that A = 0 still turns out to be
sufficient to define a nil-potent BRST-transformation on the algebras of interacting fields B̂I . The
key trick to show that A = 0 for a suitable renormalization prescription is to exploit a consistency
equation for the anomaly, which is:
     
Q̂0 A(eG
⊗ ) − G, A(eG
⊗ ) − A (Q̂ 0 G − 1
2 (G, G)) ⊗ eG
⊗ = A A(eG
⊗ ) ⊗ eG
⊗ , (136)

To see how this works, take G = Ŝ1 , and expand both sides in a power series in h̄. The right side of
the equation has the structure of an “anomaly of the anomaly”, and hence is always of higher order in
h̄ than the left side. Then using also the, the zero-curvature equation (133), the equation tells us that
leading order h̄ contribution of A has to be Q̂-closed. Solutions to this condition modulo Q̂-exact ones
can be classified by cohomlogical methods [6]. In the present case, it follows that A is necessarily
proportional to the “gauge anomaly”, which is a certain 4-form expression constructed out of c and
A that is integrated over M . It is furthermore shown that a Q̂-exact term can always be removed
by a change in the renormalization prescription. Thus, it remains to be seen that the proportionality
constant vanishes. For the present pure Yang-Mills theory, this is a simple consequence of consid-
erations about parity properties. If the Yang-Mills theory is coupled to spinor fields in a non-real
representation of g, the constant does not have to vanish. In such a case the quantum field theory
cannot be consistently defined. If the theory can be defined consistently, then states can be defined
perturbatively in basically the same manner as above. However, since observables are now defined as
cohomology classes, the states have to be BRST-invariant, too. A corresponding construction method
is also given in [40] which follows the idea of [30].

4. O PEN ISSUES

In this final section, we briefly discuss two fundamental outstanding issues related to QFTCS. In
subsection 4.1 we discuss a possible approach to the formulation of interacting QFTCS via operator
product expansions. In subsection 4.2, we consider the formulation of quantum gravity from the
viewpoint of QFTCS.

27In contrast to elsewhere in this review, we reintroduce h̄ here. In the algebraic approach, this is done by simply
replacing E → h̄E in the definition of A , and, similarly, in other places involving commutators).
 QUANTUM FIELDS IN CURVED SPACETIME 43

4.1. Nonperturbative formulation of interacting QFTCS. The perturbative construction of inter-

acting QFTCS given in the previous section attempts to represent the interacting quantum field as an
element of the algebra W of the free quantum field theory. While this can be done to any finite order,
the algebra of interacting fields cannot plausibly be isomorphic to a subalgebra of W , simply because
W by definition consists of elements that are, in a sense, of finite “polynomial order” in the free field,
whereas the order of the terms in the perturbation series clearly grows unboundedly. Thus, in order
to even ask the question about convergence of the series, one would have to go to a suitably enlarged
algebra and it is not clear to us what that object might be. The situation is even worse if one is calcu-
lating properties of special states, such as ground states. Even if the interacting fields were elements
of the same algebra W for all λ and had a common set of states, as Dyson noted more than 60 years
ago, for a theory such as the one with interaction Lagrangian λ φ 4 , a ground state ω0 (λ ) cannot be
expected to be analytic in λ at λ = 0, since no ground state can exist when λ < 0. Thus, perturbative
expressions for quantities such as the S-matrix should not converge—even if it were the case that
notions of “in” and “out” particle states could be defined. Thus, the perturbative approach described
in the previous section would appear to be a very poor way to formulate QFTCS for interacting fields.
What we would really like to do, of course, is directly write down the algebra, BI , of local field
observables for the interacting field, in analogy to our construction of the algebra B0 ⊂ W for the
free field. However, our ability to define the algebra B0 traces back to the explicit product formula
(79), which allowed us to express the product of any two elements of W as a finite sum of other
elements of W . This property, in turn, traces back to the fact that the extremely simple commutation
relations satisfied by φ0 ( f ), namely [φ0 ( f1 ), φ0 ( f2 )] ∝ 1. We cannot expect that any simple product
relations of this sort will hold for an interacting quantum field. It therefore would seem a hopeless
task to directly construct an interacting field algebra BI .
Nevertheless, we believe that a possible route towards a nonperturbative formulation of QFTCS for
interacting fields arises from the fact that it appears likely that interacting quantum fields satisfy
simple “product relations” in the limit of small separation of points. More precisely, it is known that
order-by-order in perturbation theory, interacting quantum fields satisfy relations known as operator
product expansions (OPEs). What is meant by this is the following. Let us first introduce a “basis”
of composite fields ΦA . The label incorporates both the kind of field (such as φ 2 , φ 6 , ... in the case of
a free KG-field), as well as any tensor/spinor indices (such as in φ ∇(µ ∇ν ∇σ ) φ ). The OPE states that
the product of any number of composite fields can be expanded in an asymptotic series of the form

ΦA1 (x1 ) · · · ΦAn (xn ) ∼ ∑ CCA1 ...An (x1 , ..., xn )ΦC (xn ) . (137)
C

The coefficients C appearing in the expansions are C-valued distributions that should be locally and
covariantly in terms of the spacetime geometry (M , g), and which also, of course, depend on the
particular theory under consideration. They are analogous in a rough sense to the structure constants
of a finite dimensional algebra. The asymptotically equal sign ∼ expresses that, if both sides of the
expansion are inserted into a suitably well-behaved state ω (a Hadamard state for the case of a free
KG-field), and if the sum over C is carried out to operators of a sufficiently high dimension, then the
right side should approximate the left side well for short distances, x1 , ..., xn−1 → xn . For the free
KG-field, the OPE coefficients can be found concretely in terms of the Hadamard parametrix, H. For
example, for the simple case ΦA = ΦB = φ , ΦC = 1, we have CCAB (x1 , x2 ) = H(x1 , x2 ). In the context
of a perturbative interacting theory in curved space, an algorithm for finding the coefficients was given
in [41]. In fact, for Euclidean perturbative φ 4 -theory, it has been shown that the series on the right
side actually converges in this sense, at least at arbitrary but fixed order in perturbation theory [44,56].
If, as we expect, this also were found to be true at spacelike related points for theories on Lorentzian
44 HOLLANDS AND WALD

curved spacetimes, then this would strongly reinforce the view that the OPE coefficients contain the
entire local information about the theory, even on a curved spacetime.
The OPE-coefficients are expected to satisfy a kind of “factorization rule” in a situation in which
a subset of points, say (x1 , ..., xm ) is much “closer to each other than the rest”. In that situation, a
relation of the type

CCA1 ...An (x1 , ..., xn ) ∼ ∑ CBA1 ...Am (x1 , ..., xm )CCBAm+1 ...An (xm , ..., xn ) . (138)
B

should hold. This factorization rule has been shown to hold in Euclidean φ 4 -theory in [42] to arbitrary
but finite order in perturbation theory; again the sum on the right side is shown to actually converge,
and “∼” can actually be replaced by an equality sign! Eq. (138) can be thought of as a version of the
associativity law for the quantum fields, holding at the level of the coefficients C.
A possible approach towards a formulation of QFTCS for interacting fields would be to define the
theory via its collection of OPE coefficients. The OPE coefficients would be required to satisfy the
factorization rule (138) together with a list of other natural properties; see [50] for further discussion.
The basic philosophy of this approach is that the OPEs would then fully determine the quantum
field theory in a local manner, even though one may not easily be able to construct the field algebra
and the state space from the OPEs. This would be closely analogous to the situation in classical
field theory, where the field equations uniquely determine the theory in a local manner, but one may
not be able to easily construct the space of all solutions from the field equations. The advantage
of formulating QFTCS in terms of OPEs is that to define OPEs, one must specify a collection of
“C-number” distributions. These can be specified without reference to a “background field algebra”
or “background states.” In particular, there any no obvious reasons why the OPE coefficients of an
interacting theory could not be constructed perturbatively, with a convergent perturbation series.
Unfortunately, it seems to be extremely difficult to find non-trivial solutions to the associativity law
from first principles rather than by indirect methods that rely on perturbation theory. However, there
is some hope that a self-consistent equation for the OPE-coefficients may be obtained. The idea is
that interacting quantum field theories are normally labeled by some coupling parameter, such as λ
in λ φ 4 -theory. The OPE-coefficients thus have to be functions of λ . In [43], it is shown that, again in
perturbation theory, there exists a relation of the type
∂ B
C (x1 , . . . , xN ) =
∂ λ A1 ...AN
Z  N
− CBΦA1 ...AN (y, x1 , . . . , xN ) − ∑ ∑ CCΦAi (y, xi ) CBA (x1 , . . . , xN ) (139)
1 ...Ai C...AN
b
y i=1 [C]≤[Ai ]

− ∑ CCA1 ...AN (x1 , . . . , xN ) CBΦC (y, xN ) ,
[C]<[B]

where Abi denotes omission of the corresponding index, where [A] indicates the dimension of the field
ΦA , and where the operator Φ (= φ 4 in this example) is the interaction. The spacetime integral
over y can be shown to converge absolutely. Although it is derived within perturbation theory (i.e.
in the sense of formal power series in λ ), the final form of (139) no longer makes any reference
to perturbation theory and thus may be viewed as a self-consistent relationship between the OPE-
coefficients. Eq. (139) is a first order differential equation in λ , and since λ = 0 corresponds to
the free field theory, the “initial values” of all OPE-coefficients are known at λ = 0. Thus, it is
conceivable that one could actually show that a unique solution to (139) must exist (beyond the level
 QUANTUM FIELDS IN CURVED SPACETIME 45

of formal power series). We view this as a promising approach to a non-perturbative definition of the
OPE-coefficients.
In summary, and as discussed further in [50], we believe that OPEs provide a promising approach
toward the formulation of QFTCS that is independent of, and different from, perturbative methods.

4.2. Quantum gravity. As we have seen in this review, the formulation of QFTCS assumes the
presence of a classical spacetime metric g. Consequently, QFTCS can, at best, be an approximation
to a more fundamental theory of “quantum gravity,” wherein the metric is no longer treated as a
classical field. We shall not attempt here to review—or even mention—the various approaches that
have been taken towards the formulation of a quantum theory of gravity, their successes, and their
difficulties. Suffice it to say that no approach has yet been formulated that has the conceptual clarity
and mathematical rigor of QFTCS. The question we do wish to briefly address here is whether QFTCS
provides any insights towards the formulation and properties of quantum gravity.
One can attempt to formulate a quantum theory of gravity by expanding the metric g about a back-
ground solution g0 , i.e., by writing
g = g0 + γ . (140)
The idea is now to treat γ as a quantum field propagating in the classical spacetime (M , g0 ), and to
formulate a quantum theory of γ by the procedures of QFTCS. However, in order to be in accord with
the fundamental principle of general relativity that no “background structure” appear in the theory
apart from the spacetime manifold M , one would like the final theory to be independent of the choice
of g0 appearing in the expansion (140). To impose this requirement, γ should satisfy a complicated
“gauge symmetry” that can be obtained by applying an arbitrary diffeomorphism ψ to g and then re-
expanding ψ ∗ g about g0 . Indeed, even if one does not care about “fundamental principles” one would
need to impose a gauge symmetry of this sort in order to have deterministic dynamics, since Einstein’s
equation does not determine g but only the diffeomorphism equivalence class of g. Thus, one cannot
get well defined “propagators” for γ without making a gauge choice, and it would be natural to insist
that the resulting theory does not depend upon the choice of gauge.
In this regard, the situation with regard to formulating QFTCS for γ may appear to be similar to the
situation for formulating QFTCS for a Yang-Mills field, A, which has a somewhat similar-looking
gauge symmetry. For the Yang-Mills field, an appropriate route toward formulating QFTCS was to
formulate the theory in terms of gauge invariant local observables, which can be done via the BRST
procedure discussed in subsection 3.3 above. However, it is clear that such a route cannot work in the
case of quantum gravity, simply because the nature of diffeomorphism gauge symmetry is such that
there do not exist any gauge invariant local observables. Diffeomorphisms do not merely “transform
fields at a point” but they “move the points themselves around,” so there cannot be any physical
meaning to the properties of the metric “at a given spacetime location.” Thus, one cannot hope to
formulate QFTCS for γ in terms of gauge invariant local observables. This difficulty is closely related
to the “problem of time” encountered in approaches to quantum gravity that attempt to provide the
quantum state of the metric at a given “time.” If the theory is to respect the diffeomorphism gauge
symmetry, the state cannot depend upon “time,” thereby making its interpretation extremely nebulous.
Although, as we have just argued, the above approach cannot yield a quantum theory of gravity,
we can successfully obtain a theory of linearized gravity off of an arbitrary background solution g0
by following the analog of the procedure for Maxwell fields discussed near eq. (125) above. For
linearized gravity, the gauge transformations of γ are simply γ → γ + Lξ g0 , where ξ is an arbitrary
vector field on M . One can then construct an analog of the algebra A of section 2.1 for the quantum
46 HOLLANDS AND WALD

field γµν by defining its “smearing” only with symmetric test tensor fields f µν that satisfy28
∇µ f µν = 0 . (141)
Note that the resulting smeared fields γ( f ) are gauge invariant since
Z Z Z
µν µν
(Lξ g0 )µν f =2 ∇(µ ξν) f = −2 ξν ∇µ f µν = 0 . (142)
M M M
The classical advanced and retarded are well defined when smeared with test fields satisfying (141),
and this is all that is needed to construct the analog of A for linearized gravity; for details see [26].
Consequently, all of the constructions and analyses given in section 2 for the free scalar field can
be repeated, in parallel, for the linearized gravitational field, including the Hawking effect and the
behavior of cosmological perturbations29. Thus, despite the lack of a quantum theory of gravity, one
can rather confidently make predictions about the “radiation of gravitons” by a black hole and the
generation of “tensor perturbations” by quantum field effects in cosmology.
Nevertheless, if is far from clear whether an analog of W can be defined for the linearized gravitational
field, and it is far from clear that perturbative QFTCS can be defined for γ—even at second order30.
Furthermore, even if a perturbative QFTCS could somehow be defined for γ, it is very difficult to see
how it could give rise to a theory that makes predictions about local observables in a manner that is
compatible with the basic principles of general relativity.
Thus, it seems clear that ideas that go beyond the principles of QFTCS as described in this review will
be needed to formulate a quantum theory of gravity. Nevertheless, as a consequence of its prediction
of such phenomena as the Hawking effect, QFTCS has already provided some remarkable insights
into the nature of quantum gravity, and we believe that it will continue to do so in the future.

Acknowledgements: The research of S.H. was supported in part by ERC grant QC & C 259562. The
research of R.M.W. was supported in part by NSF grant PHY 12-02718 to the University of Chicago.

A PPENDIX A. D ISTRIBUTIONS , SCALING DEGREE , AND WAVE FRONT SETS

The objects appearing in quantum field theory such as n-point functions, time-ordered products, etc.
are singular and therefore best viewed as distributions. A distribution u on a d-dimensional manifold

28In Minkowski spacetime, a symmetric tensor field f µν satisfying (141) can be written in the form f µν = ∂ ∂ U α µβ ν ,
α β
where U α µβ ν is antisymmetric in its first and last pairs of indices and also satisfies U α µβ ν = U β να µ . From this and the
linearized field equation, it can be seen that the smeared observables γ( f ) are equivalent to using the smeared linearized
Weyl tensor as the fundamental observables of the theory. However, if perturbing off of a nontrivial background, the
gauge invariant observables γ( f ) with f satisfying (141) are no longer equivalent to the linearized Weyl tensor (which,
itself, is no longer gauge invariant).
29To treat cosmological perturbations, one must linearize all fields off of a classical background Einstein-matter
solution.
30If one has appropriate “fixed asymptotic regions” of the spacetime, then it should be possible to define an S-matrix
to all orders in perturbation theory, since such a quantity would be “gauge invariant” in an appropriate sense. However,
as we have argued in the introduction, the quantities of main interest in gravitational physics are local observables, not
relations between “in” and “out” states (even when such notions can be defined). Furthermore, it is worth mentioning that,
as is well known, on account of the “non-renormalizability” of the gravitational action, the determination of S-matrix to
all orders will depend upon an infinite number of parameters of the renormalization scheme, and thus has relatively little
predictive power.
 QUANTUM FIELDS IN CURVED SPACETIME 47

X is a complex linear functional u : C0∞ (X) → C for which there is a constant cK and an N ∈ N0 for
each compact K ⊂ X such that
|u( f )| ≤ cK ∑ sup |Dk f (x)| (143)
k≤N x∈K
for any f ∈ C0∞ having support within K, where D is any derivative operator on X. For instance, the
delta “function” on R concentrated at 0, which is defined by δ ( f ) = f (0), evidently satisfies the above
estimate with N = 0. The n-th derivative δ (n)R satisfies the criterion with N = n. Also, any smooth
function u defines a distribution via u( f ) = u(x) f (x)dv for a given integration element dv on X.
Such a distribution is called smooth, and more generally, a distribution is called smooth at x0 ∈ X
if it can be represented in that way for f having support sufficiently close to x0 . The complement
of the set of all such x0 is called the “singular support” singsupp(u) ⊂ X. The notion of singular
support is not very informative since it gives no insight into the precise nature of the singularity at
a given x0 ∈ singsupp(u). This shortcoming can be dealt with by introducing more refined concepts
to characterize singularities. Two such concepts of particular relevance for QFTCS are that of the
scaling degree and that of the wave front set.
The scaling degree of a distribution u at a point x ∈ X basically describes the “degree of divergence”
at x, if any. It is defined more formally as follows. Choose an arbitrary chart (U, ψ) near x and
let uψ ( f ) = u( f ◦ ψ −1 ) the pull back of u to Rd , defined for f supported in ψ[U]. Without loss of
generality we may assume that ψ(x) = 0, and we define fε (y) = ε −d f (y/ε). The scaling degree is
given by
sdx (u) = inf{δ ∈ R | lim ε δ uψ ( fε ) = 0 for all f supported in ψ[U]} . (144)
ε→0+
It is easily checked that the definition is independent of the choice of chart (U, ψ). For example, the
scaling degree of the distributions (x + i0+ )−n on R at x = 0 is n, whereas it is 0 at any other point
x 6= 0. The scaling degree of the n-th derivative δ (n) of the delta distribution on R at x = 0 is likewise
n whereas it is −∞ for x 6= 0. The scaling degree “doesn’t see logarithms”: The scaling degree of
logn (x + i0+ ) is = 0 for any n at any point x ∈ R.
The concept of wave front set [51] does not characterize the strength of a singularity, but rather
its nature from the point of view of momentum space. To define the wave front set, assume first
a distribution u of compact support contained in some chart (U, ψ) of X. Then we may define the
Fourier transform in that chart by ûψ (k) = u[exp iψ( . ) · k]. If u is smooth within U, then it is easy to
see that there holds
|ûψ (k)| ≤ cN (1 + |k|)−N for all N ∈ N, (145)
for some constants cN . For a general distribution supported in U, we say that k0 6= 0 is a singular
direction if eq. (145) fails to hold for k in some open cone around k0 . If x0 ∈ X and if u is an arbitrary
distribution, we say that (ψ(x0 ), k0 ) is in the wave front set WF(uψ ) of uψ if k0 is a singular direction
for χu for any cutoff function χ supported in U such that χ(x0 ) 6= 0. The wave front set of uψ is a
subset of ψ[U] × (Rd \ {0}). The pull-back
(ψ −1 )∗ WF(uψ ) ⊂ T ∗ X \ 0 ,
[
WF(u) = (146)
charts (U,ψ)
can be shown to be invariantly defined (i.e. independent of the choice of atlas of X for the given
differentiable structure), and is simply called the “wave front set”.
The notion of wave front set is applied above in eq. (17) to characterize Hadamard 2-point functions
W2 (X = M × M in that example) of the free KG field. It can also be used to characterize the wave
front set of an n-fold time ordered product Tn (X = M × ... × M [14, 15] (n copies) in that case), or
of the n-point functions of n interacting fields or their OPE coefficients [42].
48 HOLLANDS AND WALD

One of the most important uses of wave-front sets in QFTCS is to characterize situations in which the
product of distributions is defined. In fact, the following theorem holds: Let u, v be distributions on
X. If WF(u) + WF(v) (element-wise addition) does not contain a zero cotangent vector in T ∗ X, then
the distributional product uv is defined. More generally, for a set of n distributions, if ∑ j WF(u j ) does
not contain a zero cotangent vector, then ∏ j u j is defined.
As an example, consider the distribution (x + i0+ )−1 , whose wave front set is found to be {(0, k) | k >
0}. The square – and in fact any power – is therefore well defined. Next, consider δ (x), which has
wave front set {(0, k) | k 6= 0}. Its square is therefore not well defined. One way to think about these
examples is that in the first case, (x + i0+ )−1 is, by definition, the boundary value of an analytic func-
tion. Whence its powers also are the boundary value of an analytic function, and hence automatically
defined. By contrast, the distribution δ (x) is not the boundary value of an analytic function, whence
its square is not automatically defined. More generally, the relationship between distributional bound-
ary values and the wave-front set is that if u(x + iy) is an analytic function in U × Γ, where U ⊂ Rd
and Γ is some open cone having finite scaling degree in y at y = 0 for any x ∈ U, then the wave front
set of the distributional boundary value u(x) = limy∈Γ,y→0 u(x + iy) is contained in
WF(u) ⊂ U × Γ∗ , (147)

where Γ∗ = {k ∈ (Rd )∗ \ 0 | hk, yi ≤ 0 ∀y ∈ Γ} is the dual cone. This criterion can also be applied in
any (analytic) manifold X by localizing u in a chart (ψ,U).
This relationship between wave front set and distributional boundary values is relevant in QFTCS,
because many distributions involve some “i0+ -prescription”. The wave front set of the two-point
function W2 of a Hadamard state for instance can be determined from (18) and (147), since the i0+ -
prescription effectively states that W2 is given in a sufficiently small open set U ⊂ M × M by a
distributional boundary value (with cone locally given by Γ = ∪(x1 ,x2 )∈U Vx+1 ×Vx−2 ). This can be used
to deduce the wave front condition (17).
Two important applications of the above product criterion for distributions are the following. Consider
first the k times contracted Wick-product eq. (80) entering in the definition of the algebra W . The
right side involves the product of the distributions Fn , Gm ,W2 . The first two distributions satisfy the
wave front condition (81), whereas W2 satisfies (17). It is easily seen that the product criterion is
satisfied, whence the k times contracted product is indeed well-defined.
As the second example, consider the time-ordered 2-point function W2T (x1 , x2 ) = ω(T2 (φ (x1 )⊗φ (x2 )))
associated with a Hadamard state ω, also called a “Feynman propagator”. The wave front set of the
time-ordered product is e.g. found to be

WF(W2T ) ={(x1 , k1 ; x2 , k2 ) ∈ T ∗ M 2 \ 0 | k1/2 ∈ V ∓ if x1/2 ∈ J + (x2/1 ), k1 ∼ −k2 }

(148)
∪ {(x1 , k1 ; x2 , k2 ) ∈ T ∗ M 2 \ 0 | k1 = −k2 , x1 = x2 } .

Because it contains e.g. the point (x, k; x, −k) for any k 6= 0, WF(W2T ) + WF(W2T ) contains the zero
co-vector. The product criterion is not fulfilled, and thus the square of W2T cannot straightforwardly
be defined as a distribution. This problem shows up precisely when one naively tries to define the
product via Fourier transform e.g. for the vacuum state in Minkowski space, and is directly related to
the logarithmic divergence of the “fish-graph” in Feynman diagram language. However, for x1 6= x2
the criterion is fulfilled and [W2T (x1 , x2 )]2 can be defined for such points. Thus, the “renormalization”
required to define the time ordered product ω(T2 (φ 2 (x1 ) ⊗ φ 2 (x2 ))) corresponds precisely to obtain-
ing an extension of this distribution to the “diagonal” of M × M , i.e. in some sense, no problems
arise other than for coincident points. This is a rather non-trivial point in curved spacetime, because
 QUANTUM FIELDS IN CURVED SPACETIME 49

x6
6
x5
4

time
5
embedding
3
1
x4
2 x3
abstract Feynman graph G x1
x2
spacetime M with embedded graph

F IGURE 7. Shown here is the wave-front set of the time-ordered products (149)
and its relationship with embedded Feynman graphs G in M . Through each line e
flows a ‘momentum’ pe indicated by →, which is a parallel transported, cotangent null
vector. At each vertex xi the corresponding vector ki ∈ Tx∗i M in the wave front set is
characterized by the ‘momentum conservation rule’ ki = ∑in pe − ∑out pe counting the
momenta associated with the incoming vs. outgoing edges e with opposite sign.

the behavior of null-geodesics (points where the “propagators” are singular) can be very different
from flat spacetime.
These considerations can be generalized to the construction of higher order time ordered products.
For instance, in order to define the expectation value ω(Tn (⊗ j φ k j (x j ))) in a Gaussian, Hadamard
state for mutually distinct points x j ∈ M , one may apply the Wick-product and “causal factorization”
formulas. This leads to an expression in terms of a product of Feynman propagators W2T (xi , x j ), where
e = (i j) run through the edges of an abstract Feynman graph G with incidence number of the j-th
vertex ≤ k j . To the product, we may again apply our criterion and conclude that it exists away from
all “diagonals”, i.e. for the open subset of M n of points such that xi 6= x j for all i 6= j. Again, the
important point is that the “extension” has to be performed only on the “small” subset of diagonals in
M n , and the potentially very complicated nature of the “null-related singularities” is taken care of by
the wave front set techniques.
An important aspect of the precise analysis [15, 47] is that the wave front set of the extension can
be controlled, including at the diagonals; it is characterized by the following “microlocal spectrum
property”: For Hadamard states ω, we have
  n
WF ω(Tn (⊗ j φ k j )) ⊂ (x1 , k1 ; ...; xn , kn ) ∈ T ∗ M n \ 0 ki = ∑ pe − ∑ pe
e∈G :s(e)=i e∈G :t(e)=i (149)
o
pe ∈ V ∓ if xs(e)/t(e) ∈ J + (xt(e)/s(e) ) .

Here, one is considering embeddings of the graph G into M such that its edges e = (i j) are associated
with null-geodesics. Their cotangent null vectors are called pe . These are future/past oriented depend-
ing on whether the edge e = (i j) (oriented so that s(e) := i < j =: t(e)) is future or past directed. An
illustration is given in fig. 7.

R EFERENCES
[1] B. Allen, “Vacuum States in de Sitter Space,” Phys. Rev. D 32, 3136 (1985).
50 HOLLANDS AND WALD

[2] T. Altherr, “Infrared Problem in gφ 4 Theory at Finite Temperature,” Phys. Lett. B 238, 360 (1990).
[3] A. Almheiri, D. Marolf, J. Polchinski, and J. Sully, “Black Holes: Complementarity or Firewalls?” JHEP 1302, 62
(2013) arXiv:1207.3123 [hep-th].
[4] H. Araki and S. Yamagami: “On the quasi-equivalence of quasifree states of the canonical commutation relations”,
Publ. RIMS, Kyoto U. 18 (1982)
[5] H. Araki, “Relative Hamiltonian for faithful normal states of a von Neumann algebra,” Publ. RIMS, Kyoto U. 9 (1),
165–209 (1973).
[6] G. Barnich, F. Brandt and M. Henneaux, “Local BRST cohomology in gauge theories,” Phys. Rept. 338, 439 (2000)
[hep-th/0002245].
[7] C. Bär, N. Ginoux and F. Pfäffle, “Wave equations on Lorenzian manifolds and quantization,” Zuerich, Switzerland:
Eur. Math. Soc. (2007) 194 p
[8] J. M. Bardeen, B. Carter and S. W. Hawking, “The Four laws of black hole mechanics,” Commun. Math. Phys. 31,
161 (1973).
[9] N. N. Bogoliubov, D. V. Shirkov: Introduction to the theory of quantized fields, Wiley (1980)
[10] D. G. Boulware, “Quantum Field Theory in Schwarzschild and Rindler Spaces,” Phys. Rev. D 11, 1404 (1975).
[11] O. Bratteli and D. W. Robinson, Operator Algebras and Quantum Sta- tistical Mechanics 2: Equilibrium States.
Models in Quantum Statistical Mechanics, Springer, 2002.
[12] J. Bros and U. Moschella, “Two point functions and quantum fields in de Sitter universe,” Rev. Math. Phys. 8, 327
(1996) [gr-qc/9511019].
[13] J. Bros, U. Moschella and J. P. Gazeau, “Quantum field theory in the de Sitter universe,” Phys. Rev. Lett. 73, 1746
(1994).
[14] R. Brunetti, K. Fredenhagen and M. Kohler, “The Microlocal spectrum condition and Wick polynomials of free
fields on curved space-times,” Commun. Math. Phys. 180, 633 (1996) [gr-qc/9510056].
[15] R. Brunetti and K. Fredenhagen, “Microlocal analysis and interacting quantum field theories: Renormalization on
physical backgrounds,” Commun. Math. Phys. 208, 623 (2000) [math-ph/9903028].
[16] R. Brunetti, K. Fredenhagen and R. Verch, “The Generally covariant locality principle: A New paradigm for local
quantum field theory,” Commun. Math. Phys. 237, 31 (2003) [math-ph/0112041].
[17] B. Chilian and K. Fredenhagen, “The Time slice axiom in perturbative quantum field theory on globally hyperbolic
spacetimes,” Commun. Math. Phys. 287, 513 (2009) [arXiv:0802.1642 [math-ph]].
[18] B. S. DeWitt and R. W. Brehme, “Radiation damping in a gravitational field,” Annals Phys. 9, 220 (1960).
[19] M. Dafermos and I. Rodnianski, “The Red-shift effect and radiation decay on black hole spacetimes,” Commun.
Pure Appl. Math. 62, 859 (2009) [gr-qc/0512119].
[20] C. Dappiaggi, V. Moretti and N. Pinamonti, “Rigorous construction and Hadamard property of the Unruh state in
Schwarzschild spacetime,” Adv. Theor. Math. Phys. 15, 355 (2011) [arXiv:0907.1034 [gr-qc]].
[21] C. Dappiaggi, T. -P. Hack and N. Pinamonti, “The Extended algebra of observables for Dirac fields and the trace
anomaly of their stress-energy tensor,” Rev. Math. Phys. 21, 1241 (2009) [arXiv:0904.0612 [math-ph]].
[22] J. Dimock, “Algebras of local observables on a manifold,” Commun. Math. Phys. 77, 219 (1980)
[23] M. Duetsch and K. Fredenhagen, “Algebraic quantum field theory, perturbation theory, and the loop expansion,”
Commun. Math. Phys. 219, 5 (2001) [hep-th/0001129].
[24] H. Epstein and V. Glaser, “The Role of locality in perturbation theory,” Annales Poincare Phys. Theor. A 19, 211
(1973).
[25] C. J. Fewster and M. J. Pfenning, “A Quantum weak energy inequality for spin one fields in curved space-time,” J.
Math. Phys. 44, 4480 (2003) [gr-qc/0303106].
[26] C. J. Fewster and D. S. Hunt, “Quantization of linearized gravity in cosmological vacuum spacetimes,” Rev. Math.
Phys. 25, 1330003 (2013) [arXiv:1203.0261 [math-ph]].
[27] F. Finster and A. Strohmaier, “Gupta-Bleuler Quantization of the Maxwell Field in Globally Hyperbolic Space-
Times,” arXiv:1307.1632 [math-ph].
[28] K. Fredenhagen and R. Haag, “On the Derivation of Hawking Radiation Associated With the Formation of a Black
Hole,” Commun. Math. Phys. 127, 273 (1990).
[29] K. Fredenhagen and F. Lindner, “Construction of KMS States in Perturbative QFT and Renormalized Hamiltonian
Dynamics,” arXiv:1306.6519 [math-ph].
[30] M. Duetsch and K. Fredenhagen, “A Local (perturbative) construction of observables in gauge theories: The Ex-
ample of QED,” Commun. Math. Phys. 203, 71 (1999) [hep-th/9807078].
[31] K. Fredenhagen and R. Haag, “Generally Covariant Quantum Field Theory and Scaling Limits,” Commun. Math.
Phys. 108, 91 (1987).
 QUANTUM FIELDS IN CURVED SPACETIME 51

[32] S. A. Fulling, F. J. Narcowich and R. M. Wald, “Singularity Structure of the Two Point Function in Quantum Field
Theory in Curved Space-time. II,” Annals Phys. 136, 243 (1981).
[33] C. Gerard and M. Wrochna, “Construction of Hadamard states by pseudo-differential calculus,” arXiv:1209.2604
[math-ph].
[34] J. Glimm and A. Jaffe: Quantum physics: A functional integral point of view, Springer, Berlin 1987
[35] R. Haag, Local quantum physics: Fields, particles, algebras, Berlin, Germany: Springer (1992) 356 p. (Texts and
monographs in physics)
[36] S. W. Hawking, “Particle Creation by Black Holes,” Commun. Math. Phys. 43, 199 (1975) [Erratum-ibid. 46, 206
(1976)].
[37] A. Higuchi, D. Marolf and I. A. Morrison, “On the Equivalence between Euclidean and In-In Formalisms in de Sitter
QFT,” Phys. Rev. D 83, 084029 (2011) [arXiv:1012.3415 [gr-qc]].
[38] S. Hollands, “Massless interacting quantum fields in deSitter spacetime,” Annales Henri Poincare 13, 1039 (2012)
[arXiv:1105.1996 [gr-qc]].
[39] S. Hollands, “Correlators, Feynman diagrams, and quantum no-hair in deSitter spacetime,” Commun. Math. Phys.
319, 1 (2013) [arXiv:1010.5367 [gr-qc]].
[40] S. Hollands, “Renormalized Quantum Yang-Mills Fields in Curved Spacetime,” Rev. Math. Phys. 20, 1033 (2008)
[arXiv:0705.3340 [gr-qc]].
[41] S. Hollands, “The Operator product expansion for perturbative quantum field theory in curved spacetime,” Commun.
Math. Phys. 273, 1 (2007) [gr-qc/0605072].
[42] J. Holland and S. Hollands, “Operator product expansion algebra,” J. Math. Phys. 54, 072302 (2013)
[arXiv:1205.4904 [math-ph]].
[43] J. Holland and S. Hollands, “A recursion relation for OPE coefficients in perturbation theory,” in preparation
[44] S. Hollands and C. Kopper, “The operator product expansion converges in perturbative field theory,” Commun. Math.
Phys. 313, 257 (2012) [arXiv:1105.3375 [hep-th]].
[45] S. Hollands and W. Ruan, “The State space of perturbative quantum field theory in curved space-times,” Annales
Henri Poincare 3, 635 (2002) [gr-qc/0108032].
[46] S. Hollands and R. M. Wald, “Local Wick polynomials and time ordered products of quantum fields in curved
space-time,” Commun. Math. Phys. 223, 289 (2001) [gr-qc/0103074].
[47] S. Hollands and R. M. Wald, “Existence of local covariant time ordered products of quantum fields in curved space-
time,” Commun. Math. Phys. 231, 309 (2002) [gr-qc/0111108].
[48] S. Hollands and R. M. Wald, “On the renormalization group in curved space-time,” Commun. Math. Phys. 237, 123
(2003) [gr-qc/0209029].
[49] S. Hollands and R. M. Wald, “Conservation of the stress tensor in interacting quantum field theory in curved space-
times,” Rev. Math. Phys. 17, 227 (2005) [gr-qc/0404074].
[50] S. Hollands and R. M. Wald, “Axiomatic quantum field theory in curved spacetime,” Commun. Math. Phys. 293, 85
(2010) [arXiv:0803.2003 [gr-qc]].
[51] L. Hormander, The Analysis of Linear Partial Differential Operators I: Distribution Theory and Fourier Analysis,
Springer-Verlag, 2nd edition 1990
[52] V. Iyer and R. M. Wald, “Some properties of Noether charge and a proposal for dynamical black hole entropy,” Phys.
Rev. D 50, 846 (1994) [gr-qc/9403028].
[53] W. Junker and E. Schrohe, “Adiabatic vacuum states on general space-time manifolds: Definition, construction, and
physical properties,” Annales Poincare Phys. Theor. 3, 1113 (2002)
[54] B. S. Kay and R. M. Wald, “Theorems on the Uniqueness and Thermal Properties of Stationary, Nonsingular,
Quasifree States on Space-Times with a Bifurcate Killing Horizon,” Phys. Rept. 207, 49 (1991).
[55] B. S. Kay, “The Casimir Effect Without MAGIC,” Phys. Rev. D 20, 3052 (1979).
[56] G. Keller and C. Kopper, “Perturbative renormalization of composite operators via flow equations. 2. Short distance
expansion,” Commun. Math. Phys. 153, 245 (1993).
[57] C. Kopper and V. F. Muller, “Renormalization proof for massive φ44 theory on Riemannian manifolds,” Commun.
Math. Phys. 275, 331 (2007) [math-ph/0609089].
[58] N. P. Landsman and C. G. van Weert, “Real and Imaginary Time Field Theory at Finite Temperature and Density,”
Phys. Rept. 145, 141 (1987).
[59] D. Marolf and I. A. Morrison, “The IR stability of de Sitter QFT: results at all orders,” Phys. Rev. D 84, 044040
(2011) [arXiv:1010.5327 [gr-qc]].
[60] M. J. Radzikowski, “Micro-local approach to the Hadamard condition in quantum field theory on curved space-time,”
Commun. Math. Phys. 179, 529 (1996).
52 HOLLANDS AND WALD

[61] M. J. Radzikowski, “A Local to global singularity theorem for quantum field theory on curved space-time,” Commun.
Math. Phys. 180, 1 (1996).
[62] H. Sahlmann and R. Verch, “Passivity and microlocal spectrum condition,” Commun. Math. Phys. 214, 705 (2000)
[63] K. Sanders, “Equivalence of the (generalised) Hadamard and microlocal spectrum condition for (generalised) free
fields in curved spacetime,” Commun. Math. Phys. 295, 485 (2010) [arXiv:0903.1021 [math-ph]].
[64] K. Sanders, “On the construction of Hartle-Hawking-Israel states across a static bifurcate Killing horizon,”
arXiv:1310.5537 [gr-qc].
[65] K. Sanders, C. Dappiaggi and T. -P. Hack, “Electromagnetism, local covariance, the Aharonov-Bohm effect and
Gauss’ law,” arXiv:1211.6420 [math-ph].
[66] C. Schomblond and P. Spindel, “Unicity Conditions of the Scalar Field Propagator Delta(1) (x,y) in de Sitter Uni-
verse,” Annales Poincare Phys. Theor. 25, 67 (1976).
[67] R. F. Streater and A. S. Wightman, PCT, spin and statistics, and all that,’ Princeton, USA: Princeton Univ. Pr. (2000)
207 p.
[68] W. G. Unruh, “Notes on black hole evaporation,” Phys. Rev. D 14, 870 (1976).
[69] R. Verch, “Local definiteness, primarity and quasiequivalence of quasifree Hadamard quantum states in curved
space-time,” Commun. Math. Phys. 160, 507 (1994).
[70] R. M. Wald, “On Particle Creation by Black Holes,” Commun. Math. Phys. 45, 9 (1975).
[71] R. M. Wald, Quantum field theory in curved spacetime and black hole thermodynamics, University of Chicago Press,
Chicago (1994).
[72] R. M. Wald “The Thermodynamics of Black Holes,” Liv. Rev. Rel. 4, 6 (2001) arXiv:9912119 [gr-qc].

U NIVERSITÄT L EIPZIG , I NSTITUT FÜR T HEORETISCHE P HYSIK , B RÜDERSTRASSE 16, D-04103 L EIPZIG , FRG
E-mail address: [email protected]

E NRICO F ERMI I NSTITUTE AND D EPARTMENT OF P HYSICS , U NIVERSITY OF C HICAGO , C HICAGO , IL 60637, USA
E-mail address: [email protected]