Axiomatic quantum field theory in curved spacetime

Stefan Hollandsa∗, Robert M. Waldb†

a
arXiv:0803.2003v1 [gr-qc] 13 Mar 2008

School of Mathematics, Cardiff University, UK

b Enrico Fermi Institute and Department of Physics

University of Chicago, Chicago, IL, USA

March 13, 2008

Abstract
The usual formulations of quantum field theory in Minkowski spacetime make cru-
cial use of features—such as Poincare invariance and the existence of a preferred vacuum
state—that are very special to Minkowski spacetime. In order to generalize the formulation
of quantum field theory to arbitrary globally hyperbolic curved spacetimes, it is essential
that the theory be formulated in an entirely local and covariant manner, without assuming
the presence of a preferred state. We propose a new framework for quantum field theory, in
which the existence of an Operator Product Expansion (OPE) is elevated to a fundamental
status, and, in essence, all of the properties of the quantum field theory are determined by
its OPE. We provide general axioms for the OPE coefficients of a quantum field theory.
These include a local and covariance assumption (implying that the quantum field theory
is locally and covariantly constructed from the spacetime metric), a microlocal spectrum
condition, an "associativity" condition, and the requirement that the coefficient of the iden-
tity in the OPE of the product of a field with its adjoint have positive scaling degree. We
prove curved spacetime versions of the spin-statistics theorem and the PCT theorem. Some
potentially significant further implications of our new viewpoint on quantum field theory
are discussed.
∗ [email protected]
† [email protected]

1
1 Introduction
The Wightman axioms [27] of quantum field theory in Minkowski spacetime are generally
believed to express the fundamental properties that quantum fields possess. In essence,
these axioms require that the following key properties hold: (1) The states of the theory are
unit rays in a Hilbert space, H , that carries a unitary representation of the Poincare group.
(2) The 4-momentum (defined by the action of the Poincare group on the Hilbert space)
is positive, i.e., its spectrum is contained within the closed future light cone (“spectrum
condition”). (3) There exists a unique, Poincare invariant state (“the vacuum”). (4) The
quantum fields are operator-valued distributions defined on a dense domain D ⊂ H that is
both Poincare invariant and invariant under the action of the fields and their adjoints. (5)
The fields transform in a covariant manner under the action of Poincare transformations.
(6) At spacelike separations, quantum fields either commute or anticommute.
During the past 40 years, considerable progress has been made in understanding both
the physical and mathematical properties of quantum fields in curved spacetime. Although
gravity itself is treated classically, this theory incorporates some key aspects of general
relativity and thereby should provide a more fundamental base for quantum field theory.
Much of the progress has occurred in the analysis of free (i.e., non-self-interacting) fields,
but in the past decade, major progress also has been made in the perturbative analysis of
interacting quantum fields. Significant insights have thereby been obtained into the nature
of quantum field phenomena in strong gravitational fields. In addition, some important in-
sights have been obtained into the nature of quantum field theory itself. One of the key
insights is that—apart from stationary spacetimes or spacetimes with other very special
properties—there is no unique, natural notion of a “vacuum state” or of “particles”. In-
deed, unless the spacetime is asymptotically stationary at early or late times, there will
not, in general, even be an asymptotic notion of particle states. Consequently, it is essen-
tial that quantum field theory in curved spacetime be formulated in terms of the local field
observables as opposed, e.g., to S-matrices.
Since quantum field theory in curved spacetime should be much closer to a true theory
of nature than quantum field theory in Minkowski spacetime, it is of interest to attempt to
abstract the fundamental features of quantum field theory in curved spacetime in a man-
ner similar to the way the Wightman axioms abstract what are generally believed to be
the fundamental features of quantum field theory in Minkowski spacetime. The Wightman
axioms are entirely compatible with the focus on local field observables, as needed for
a formulation of quantum field theory in curved spacetime. However, most of the prop-
erties of quantum fields stated in the Wightman axioms are very special to Minkowski
spacetime and cannot be generalized straightforwardly to curved spacetime. Specifically,
a curved spacetime cannot possess Poincare symmetry—indeed a generic curved space-
time will not possess any symmetries at all—so one certainly cannot require “Poincare
invariance/covariance” or invariance under any other type of spacetime symmetry. Thus,
no direct analog of properties (3) and (5) can be imposed in curved spacetime, and the key
aspects of properties (1) and (2) (as well as an important aspect of (4)) also do not make
sense.

2
 In fact, the situation with regard to importing properties (1), (2), and (4) to curved space-
time is even worse than would be suggested by merely the absence of symmetries: There
exist unitarily inequivalent Hilbert space constructions of free quantum fields in spacetimes
with a noncompact Cauchy surface and (in the absence of symmetries of the spacetime)
none appears “preferred”. Thus, it is not appropriate even to assume, as in (1), that states
are unit rays in a single Hilbert space, nor is it appropriate to assume, as in (4), that the
(smeared) quantum fields are operators on this unique Hilbert space. With regard to (2),
although energy and momentum in curved spacetime cannot be defined via the action of a
symmetry group, the stress-energy tensor of a quantum field in curved spacetime should be
well defined as a distributional observable on spacetime, so one might hope that it might be
possible to, say, integrate the (smeared) energy density of a quantum field over a Cauchy
surface and replace the Minkowski spacetime spectrum condition by the condition that the
total energy of the quantum field in any state always is non-negative. However, this is not
a natural thing to do, since the “total energy” defined in this way is highly slice/smearing
dependent, and it is well known in classical general relativity that in asymptotically flat
spacetimes, the integrated energy density of matter may bear little relationship to the true
total mass-energy. Furthermore, it is well known that the energy density of a quantum field
(in flat or curved spacetime) can be negative, and, in some simple examples involving free
fields in curved spacetime, the integrated energy density is found to be negative. Conse-
quently, there is no analog of property (2) in curved spacetime that can be formulated in
terms of the “total energy-momentum” of the quantum field. Thus, of all of the properties
of quantum fields in Minkowski spacetime stated in the Wightman axioms, only property
(6) has a straightforward generalization to curved spacetimes!
Nevertheless, it has been understood for quite some time that the difficulties in the
formulation of quantum field theory in curved spactime that arise from the existence of
unitarily inequivalent Hilbert space constructions of the theory can be overcome by simply
formulating the theory via the algebraic framework [14]. Instead of starting from the pos-
tulate that the states of the theory comprise a Hilbert space and that the (smeared) quantum
fields are operators on this Hilbert space, one starts with the assumption that the (smeared)
quantum fields (together with the identity element 1) generate a *-algebra, A . States are
then simply expectation functionals h . iω : A → C on the algebra, i.e., linear functionals that
are positive in the sense that hA∗ Aiω ≥ 0 for all A ∈ A . The GNS construction then assures
us that given a state, ω, one can find a Hilbert space H that carries a representation, π,
of the *-algebra A , such that there exists a vector |Ψi ∈ H for which hAiω = hΨ|π(A)|Ψi
for all A ∈ A . All of the operators, π(A), are automatically defined on a common dense
invariant domain, D ⊂ H , and each vector |Ψi ∈ D defines a state via hAiΨ = hΨ|π(A)|Ψi.
Thus, by simply adopting the algebraic viewpoint, we effectively incorporate into quantum
field theory in curved spacetime the portions of the content of properties (1) and (4) above
that do not refer to Poincare symmetry.
It is often said that in special relativity one has invariance under “special coordinate
transformations” (i.e., Poincare transformations), whereas in general relativity, one has in-
variance under “general coordinate transformations” (i.e., all diffeomorphisms). Thus, one
might be tempted to think that the Minkowski spacetime requirements of invariance/covariance

3
under Poincare transformations could be generalized to curved spacetime by requiring a
similar “invariance/covariance under arbitrary diffeomorphisms”. However, such thoughts
are based upon a misunderstanding of the true meaning of “special covariance” and “general
covariance”. By explicitly incorporating the flat spacetime metric, ηab , into the formula-
tion of special relativity, it can easily be seen that special relativity can be formulated in
as “generally covariant” a manner as general relativity. However, the act of formulating
special relativity in a generally covariant manner does not provide one with any additional
symmetries or other useful conditions on physical theories in flat spacetime. The point is
that in special relativity, Poincare transformations are symmetries of the spacetime struc-
ture, and we impose a nontrivial requirement on a physical theory when we demand that
its formulation respect these symmetries. However, a generic curved spacetime will not
possess any symmetries at all, so no corresponding conditions on a physical theory can be
imposed. The demand that a theory be “generally covariant” (i.e., that its formulation is
invariant under arbitrary diffeomorphisms) can always be achieved by explicitly incorpo-
rating any “background structure” into the formulation of the theory. If one considers a
fixed, curved spacetime without symmetries, no useful conditions can be imposed upon a
quantum field theory by attempting to require some sort of “invariance” of the theory under
diffeomorphisms.
However, there is a meaningful notion of “general covariance” that can be very use-
fully and powerfully applied to quantum field theory in curved spacetime. The basic idea
behind this notion is that the only “background structure” that should occur in the theory
is the spacetime manifold and metric modulo diffeomorphisms, together with the time and
space orientations and (if spinors are present in the theory) spin structure. The quantum
fields should be “covariant” in that they should be constructed from only this background
structure. Indeed, since the smeared quantum fields are associated with local regions of
spacetime (namely, the support of the test function used for the smearing), it seems natural
to demand that the quantum fields be locally constructed from the background structure in
the sense that the quantum fields in any neighborhood O be covariantly constructed from
the background structure within O . This idea may be formulated in a precise manner as
follows [19, 20, 7].
First, in order to assure a well defined dynamics and in order to avoid causal pathologies,
we restrict consideration to globally hyperbolic spacetimes (M, gab ). (We consider theories
in arbitrary spacetime dimension D ≡ dim M ≥ 2.) If spinors are present in the theory, we
also demand that M admit a spin structure. It is essential that the quantum field theory
be defined on all D-dimensional globally hyperbolic spacetimes admitting a spin structure,
since in essence, we can only tell whether the quantum field is “locally and covariantly
constructed out of the metric” if we can see how the theory changes when we change the
metric in an arbitrary way. The “background structure”, M, of the theory is taken to consist
of the manifold M, the metric gab , the spacetime orientation—which may be represented
by a nowhere vanishing D-form, ea1 ...aD on M—and a time orientation—which may be
represented e.g. by the equivalence class of a time function T : M → R—i.e., we have

M = (M, g, T, e) . (1)

4
(If spinors are present in the theory, and M admits more than one spin structure, then the
choice of spin-structure over M also should be understood to be included M.) For each
choice of M, we assume that there is specified a *-algebra A (M) that is generated by
a countable list of quantum fields φ(i) and their “adjoints” φ(i)∗ . These fields may be of
arbitrary tensorial or spinorial type, and they are smeared with arbitrarily chosen smooth,
compact support fields of dual tensorial or spinorial type. In order to determine if the
quantum field theory and quantum fields φ(i) are “locally and covariantly constructed out of
the background structure M”, we consider the following situation: Let (M, g) and (M ′ , g′ )
be two globally hyberbolic spacetimes that have the property that there exists a one-to-one
(but not necessarily onto) map ρ : M → M ′ that preserves all of the background structure. In
other words, ρ is an isometric imbedding that is orientation and time orientation preserving
(and, if spinors are present, the choices of spin structure on M and M ′ correspond under ρ).
We further assume that ρ is causality preserving in the sense that if x1 , x2 ∈ M cannot be
connected by a causal curve in M, then ρ(x1 ) and ρ(x2 ) cannot be connected by a causal
curve in M ′ . We say that the quantum field theory is locally and covariantly constructed
from M (or, for short, that the theory is local and covariant) if (i) for every such M, M′ ,
and ρ we have a corresponding *-isomorphism χρ between A (M) and the subalgebra of
A (M′ ) generated by the quantum fields φ(i) and φ(i)∗ smeared with test fields with support
in ρ[M] and (ii) if ρ′ is a similar background structure and causality preserving map taking
M′ to M′′ , then χρ′ ◦ρ = χρ′ ◦ χρ . We further say that the quantum field φ(i) is locally and
covariantly constructed from M (or, for short, that φ(i) is local and covariant) if for every
such M, M′ , and ρ, we have


χρ φ(i) ( f ) = φ(i) (ρ∗ ( f )) , (2)

where ρ∗ ( f ) denotes the natural push-forward action of ρ on the tensor/spinor field f on

M.
Note that in contrast to the notion of Poincare invariance/covariance—which applies
to quantum field theory on a single spacetime (namely, Minkowski spacetime)—the no-
tion that a quantum field theory or quantum field is local and covariant is a condition that
applies to the formulation of quantum field theory on different spacetimes. Nevertheless,
the close relationship between these notions can be seen as follows: Suppose that we have
a local and covariant quantum field theory, with local and covariant quantum fields φ(i) .
Let M and M′ both be the background structure of Minkowski spacetime, and let ρ be a
proper Poincare transformation. Then ρ preserves all of the background structure, so for
each proper Poincare transformation, we obtain a *-isomorphism χρ : A → A , where A
here denotes the quantum field algebra for Minkowski spacetime. Furthermore, if ρ and
ρ′ are proper Poincare transformations, we have χρ◦ρ′ = χρ ◦ χρ′ . Thus, every local and
covariant quantum field theory in curved spacetime gives rise to a Poincare invariant theory
in this sense when restricted to Minkowski spacetime. Furthermore, if φ(i) is a local and
covariant quantum field, then in Minkowski spacetime it transforms covariantly via eq.(2)
under proper Poincare transformations.
Note also that, more generally, in any curved spacetime with symmetries, a local and
covariant quantum field theory will be similarly invariant under these symmetries, and a

5
local and covariant quantum field will transform covariantly under these symmetries. But
even for spacetimes without any symmetries, the requirement that the quantum field the-
ory and quantum fields be local and covariant imposes a very powerful restriction akin to
requiring Poincare invariance/covariance in Minkowski spacetime.
From these considerations, it can be seen that if we adopt the above algebraic frame-
work for quantum field theory in curved spacetime and if we additionally demand that the
quantum field theory and the quantum fields φ(i) be local and covariant, then we obtain sat-
isfactory generalizations to curved spacetime of properties (1), (4), and (5) of the Wightman
axioms in Minkowski spacetime. Since we already noted that (6) has a trivial generaliza-
tion to curved spacetime, only properties (2) and (3) remain to be generalized to curved
spacetime.
We have already noted above that there is no analog of property (2) in curved space-
time that can be formulated in terms of the “total energy-momentum” of the quantum field.
However, it is possible to reformulate the spectrum condition in Minkowski spacetime in
terms of purely local properties of the quantum fields. Specifically, the “positive frequency”
(and, thereby, positive energy) properties of states are characterized by the short-distance
singularity structure of the n-point functions of the quantum fields, as described by their
wavefront set. One thereby obtains a microlocal spectrum condition [24, 5, 6] that is for-
mulated purely in terms of the local in spacetime properties of the quantum fields. This
microlocal spectrum condition has a natural generalization to curved spacetime (see sec-
tion 2 below), thus providing the desired generalization of property (2) to curved spacetime.
Consequently, only property (3) remains to be generalized. In Minkowski spacetime,
the existence of a unique, Poincare invariant state has very powerful consequences, so it
is clear that a key portion of the content of quantum field theory in Minkowski spacetime
would be missing if we failed to impose an analogous condition in curved spacetime. How-
ever, as already mentioned above, one of the clear lessons of the study of free quantum fields
in curved spacetime is that, in a general curved spacetime, there does not exist a unique,
“preferred” vacuum or other state. Furthermore, even if a prescription for finding a unique
“preferred state” on each spacetime could be found, since generic curved spacetimes do not
have any symmetries and states on different spacetimes cannot be meaningfully compared,
there would appear to be no sensible “invariance” properties that such a preferred state
could have. We do not believe that property (3) can be generalized to curved spacetime by
a condition that postulates the existence of a preferred state with special properties.
In addition, we question the fundamental status of demanding the existence of a state
that is invariant under the symmetries of the spacetime. For example, it is well known
that the free massless Klein-Gordon field in two-dimensional Minkowski spacetime does
not admit a Poincare invariant state. However, there is absolutely nothing wrong with the
quantum field algebra of this field; the quantum field theory is “Poincare invariant” and the
quantum field is “Poincare covariant” in the sense described above. Furthermore, there is
no shortage of physically acceptable (“Hadamard”) states. Thus, the only thing unusual
about this quantum field theory is that it happens not to admit a Poincare invariant state.
We do not feel that this is an appropriate reason to exclude the free massless Klein-Gordon
field in two-dimensional Minkowski from being considered to be a legitimate quantum field

6
 theory. Similar remarks apply to the free Klein-Gordon field of negative m2 in Minkowski
spacetime of all dimensions. The classical and quantum dynamics of this field are entirely
well posed and causal, although they are unstable in the sense of admitting solutions/states
where the field grows exponentially with time. This instability provides legitimate grounds
for arguing that the free Klein-Gordon field of negative m2 does not occur in nature, but we
do not feel that the absence of a Poincare invariant state constitutes legitimate grounds for
rejecting this theory as a quantum field theory; see section 6 below for further discussion.
For the above reasons, we seek a replacement of property (3) that does not require the
existence of states of a special type. The main purpose of this paper is to propose that the
appropriate replacement of property (3) for quantum field theory in curved spacetime is to
postulate the existence of a suitable operator product expansion [29, 31, 26] of the quantum
fields. The type of operator product expansion that we shall postulate is known to hold in
free field theory and to hold order by order in peturbation theory on any Lorentzian curved
spacetime [15]. We thus propose to elevate this operator product expansion to the status of
a fundamental property of quantum fields1 . Although the assumption of the existence of
an operator product expansion in quantum field theory in curved spacetime is remarkably
different in nature from the assumption of the existence of a Poincare invariant state in
quantum field theory in Minkowski spacetime, we will show in sections 4 and 5 below that it
can do “much of the same work” as the latter assumption. It is shown in [21] how to exploit
consistency conditions on the OPE in a framework closely related to that presented here.
In particular, it is shown how perturbations of a quantum field theory can be characterized
and calculated via consistency conditions arising from the OPE.
We have an additional motivation for proposing to elevate the operator product expan-
sion to the status of a fundamental property of quantum fields. For free quantum fields
in curved spacetime, an entirely satisfactory *-algebra, A 0 , of observables has been con-
structed [5, 6, 19], which includes all Wick powers and time-ordered products. However,
the elements of A 0 correspond to unbounded operators, and there does not seem to be any
natural algebra of bounded elements (with, e.g., a C∗ -structure) corresponding to A 0 . Fur-
thermore, A 0 does not appear to have any natural topology (apart from a topology that can
be defined a postiori by using the allowed states as semi-norms). Fortunately, a topology
is not actually needed to define A 0 because the relations that hold in A 0 can be expressed
in terms of finite sums of finite products of fields. However, it is inconceivable that the
relations that define an interacting field algebra will all be expressible in terms of finite
sums of finite products of the fields. Thus, without a natural topology and without finitely
expressible relations, it is far from clear as to how an interacting field algebra might be
defined. We claim that an operator-product expansion effectively provides the needed “re-
lations” between the quantum fields, and we will propose in this paper that these relations
are sufficient to define a quantum field theory. In other words, we believe that an interact-
1 Various axiomatic approaches to conformal field theories based on the operator product expansion have been
proposed previously, see e.g. the one based upon the notion of "vertex operator algebras" given in [2, 23, 11].
However, in contrast to our approach, these approaches incorporate in an essential way the conformal symmetry of
the underlying space. In some approaches to quantum field theory on Minkowski space, the OPE is not postulated,
but instead derived [10, 3, 4].

7
 ing quantum field theory is, in essence, defined via its operator-product expansion. From
this perspective, it seems natural to view the operator product-expansion as a fundamental
aspect of the quantum field theory.
In the next section, we will describe our framework for quantum field theory in curved
spacetime. In particular, we will provide a precise statement of the what we mean by
an operator-product expansion and the properties that we will assume that it possesses. In
section 3, we will state our axioms for quantum field theory in curved spacetime and explain
how it is constructed from the operator-product expansion. Finally, in sections 4 and 5
we will show that our axioms have much of the same power as the Wightman axioms by
establishing “normal” (anti-)commutation relations and proving curved spacetime versions
of the spin-statistics theorem and PCT theorem. Some further implications of our new
perspective on quantum field theory are discussed in section 6.

2 General framework for the formulation of QFT

We will now explain in much more detail our proposed framework for defining quantum
field theory. We fix a dimension D ≥ 2 of the spacetime and consider all D-dimensional
globally hyperbolic spacetimes (M, gab ). As explained in the previous section, we will as-
sume that each spacetime is equipped with an orientation, specified by a nowhere vanishing
D-form ea1 ...aD on M, and a time orientation, specified by (the equivalence class of) a glob-
ally defined time function T : M → R. The set of background data specified this way will
be denoted
M = (M, g, T, e) . (3)
In certain cases, more background structure may be prescribed, such as a choice of bundles
in which the quantum fields live. For example, if spinors are present and if M admits more
than one spin structure, then a choice of spin-structure over M is assumed to be given as
part of the background structure2 , and is understood to also be part of M. It should be em-
phasized that two spacetimes with the same manifold and metric, but, e.g., with different
time-orientations define distinct background structures. Below, we will consider quan-
tum field theories associated with background structures, and we stress that, at this stage,
the quantum field theories associated with different background structures (e.g., ones that
merely differ in the choice of, say, time-orientation) need not have any relation whatsoever.
The quantum fields present in a given theory will be assumed to correspond to sections
of vector bundles over M. We will denote the various quantum fields by φ(i) , with i ∈ I
2 It is convenient to think of the background structure as a category [7], whose objects are the tuples M. Mor-

phisms in the category of tuples M are isometric, causality, orientation, and other background structure preserving
embeddings
ρ : M → M′ . (4)
Thus, ρ is a diffeomorphism M → M ′ such that g = ρ∗ g′ , such that ρ∗ T ′ represents the same time-orientation as
T , such that ρ∗ e′ represents the same orientation as e, and such that the causal relations in (M, g) inherited from
(M ′ , g′ ) coincide with the original ones. Furthermore, if M also includes the choice of a spin structure, then ρ must
also preserve the spin structures.

8
 where I is a suitable indexing set, and we will write V (i) for the vector bundle over M,
of which φ(i) corresponds to a section. It should be emphasized that i ∈ I labels all of
the quantum fields present in the theory, not just the “fundamental” ones. Thus, even if
we were considering the theory of a single scalar field ϕ, there will be infinitely many
composite fields of various tensorial types corresponding to all monomials in ϕ and its
derivatives, each of which would be labeled by a different index i. It will be convenient to
also include a field denoted φ(1) in the list of quantum fields, which will play the role of the
identity element, 1, in the quantum field algebra.
We assume that each field φ(i) has been assigned a Bose/Fermi parity F(i) = 0, 1 modulo
two.
We further assume that there is an operation
⋆:I→I i 7→ i⋆ , (5)
having the property that V (i⋆ ) = V (i), where for any vector space E, the vector space E
consists of all anti-linear maps E v → C, with E v denoting the dual space of E. In particular,
if i is associated with, say the vector bundle V (i) of spinors with P primed and U unprimed
indices, then V (i⋆ ) is the bundle of spinors with U primed and P unprimed spinor indices.
We demand that the star operation squares to the identity3 , i⋆⋆ = i. We also require that
∗
φ(1 ) = φ(1) , i.e. that 1∗ = 1.
As in many other approaches to quantum field theory, we will use the smeared fields
φ(i) ( f )—with i ∈ I , and f a compactly supported test section in the dual vector bundle
V (i)v to V (i)—to generate a *-algebra of observables, A (M). However, in most other
algebraic approaches to quantum field theory, A (M) is assumed, a priori, to possess a
particular topological and/or other structure (e.g., C*-algebra structure) and the algebraic
relations within A (M)—together, perhaps, with specified actions of symmetry groups on
A (M)—are assumed to encode all of the information about the quantum field theory under
consideration. In particular, since the state space, S (M), is normally taken to consist of all
positive linear maps on A (M), it is clear that S (M) cannot contain any information about
the quantum field theory that is not already contained in A (M). We shall not proceed in
this manner because it is far from clear to us what topological and other structure A (M)
should be assumed to possess a priori in order to describe the quantum field theory. Instead,
we shall view the theory as being specified by providing both an algebra of observables,
A (M), and a space of allowed states, S (M). Essentially the only information about the
theory contained in the algebra of obervables, A (M), will be the list of fields appearing
in the theory and the relations that can be written as polynomial expressions in the fields
and their derivatives. In our approach, the information normally encoded in the topology of
A (M) will now be encoded in S (M). Of course, the semi-norms provided by S (M) could
be used, a postiori, to define a topology on A (M), but it is not clear that this topology would
encode all of the information in S (M); in any case, we find it simpler and more natural to
consider the quantum field theory to be defined by the pair {A (M), S (M)}.
The key idea of this paper is that we will obtain the pair {A (M), S (M)} in a natural
(i.e., functorial) manner from the space of field labels I and another datum, namely, the
3 This operation gives I the structure of an involutive category.

9
collection of “operator product expansion (OPE) coefficients”. The OPE coefficients are a
family n o
C (M) ≡ C((ij)1 )···(in ) (x1 , . . . , xn ; y) : i1 , . . . , in , j ∈ I , n ∈ N , (6)
(i )···(in )
where each C( j)1 is a distribution on M n+1 , valued in the vector bundle
π
E = V (i1 ) × · · · ×V (in ) ×V ( j)v −→ M n+1 (7)
that is defined in some open neighborhood of the diagonal in M n+1 . Thus, given C (M), we
will construct both the algebra A (M) and the state space S (M)
S (M)
ր |
C (M) dual pairing
ց |
A (M)
Thus, in our framework, a quantum field theory is uniquely specified by providing a list of
quantum fields I and a corresponding list of OPE coefficients C (M). The OPE coefficients
will be required to satisfy certain general properties, which, in effect, become the “axioms”
of quantum field theory in curved spacetime. Most of the remainder of this section will be
devoted to formulating these axioms. However, before providing the axioms for C (M), we
briefly outline how the algebra A (M) and the state space S (M) are constructed from C (M)
for any background structure M = (M, g, T, e).
The algebra A (M) is constructed by starting with the free algebra Free(M) generated
by all expressions of the form φ(i) ( f ) with i ∈ I , and f a compactly supported test section in
the dual vector bundle to V (i) . We define an antilinear *-operation on Free(M) by requiring
that its action on the generators be given by
⋆
[φ(i) ( f )]∗ = φ(i ) ( f¯) , (8)

where f¯ ∈ Sect0 [V̄ (i)] is the conjugate test section to f ∈ Sect0 (V (i)). The *-algebra A (M)
is taken to be the resulting free *-algebra factored by a 2-sided ideal generated by a set of
polynomial relations in the fields and their derivatives. These relations consist of certain
“universal” relations that do not depend on the particular theory under consideration (such
as linearity of φ(i) ( f ) in f and (anti-)commutation relations) together with certain relations
that may arise from the OPE coefficients C (M). A precise enumeration of the relations that
define A (M) will be given at the beginning of section 3.
The state space S (M) is a subspace of the space of all linear, functionals ω : A (M) → C
that are positive in the sense that ω(A∗ A) ≡ hA∗ Aiω ≥ 0 for all A ∈ A (M). This subspace
is specified as follows: First, we require that for any state ω ∈ S (M), the OPE coefficients
in the collection C (M) in eq. (6) appear in the expansion of the expectation value of the
product of fields hφ(i1 ) (x1 ) · · · φ(in ) (xn )iω in terms of the fields hφ( j) (y)iω
D E D E
φ(i1 ) (x1 ) · · · φ(in ) (xn ) ≈ ∑ C( j)1 n (x1 , . . . , xn ; y) φ( j) (y) .
(i )...(i )
(9)
ω j ω

10
 Here “≈” means that this equation holds in a suitably strong sense as an asymptotic relation
in the limit that x1 , · · · , xn → y. A precise definition of what is meant by this asymptotic
relation will be given in eq. (39) below. Secondly, we require that ω satisfy a microlocal
spectrum condition that, in essence, states that the singularities of hφ(i1 ) (x1 ) . . . φ(in ) (xn )iω
are of “positive frequency type” in the cotangent space T(x∗ 1 ,...,xn ) M n . The precise form of
this condition will be formulated in terms of the wave front set [22] (see eqs. (23) and (24)
below).
We turn now to the formulation of the conditions that we shall impose on the OPE
coefficients C (M). As indicated above, in our framework, these conditions play the role
(i )···(i )
of axioms for quantum field theory. Each operator product coefficient C( j)1 n in C (M)
is a distribution4 on M n+1 , valued in the vector bundle V (i1 ) × · · · × V (in ) × V ( j)v that is
defined in some open neighborhood of the diagonal in M n+1 . We will impose the following
requirements on these coefficients:
C1) Locality and Covariance
C2) Identity element
C3) Compatibility with the ⋆-operation
C4) Commutativity/Anti-Commutativity
C5) Scaling Degree
C6) Asymptotic positivity
C7) Associativity
C8) Spectrum condition
C9) Analytic dependence upon the metric
Before formulating these conditions in detail, for each i ∈ I we define the dimension,
dim(i) ∈ R, of the field φ(i) by5

1 n o
(i)(i⋆ )
dim(i) := sup sd C(1) , (10)
2 backgrounds M

where "sd" denotes the scaling degree of a distribution (see appendix A) and it is understood
that the scaling degree is taken about a point on the diagonal. In other words, dim(i)
4 More precisely, each OPE coefficient is an equivalence class of distributions, where two distributions are
considered equivalent if their difference satisfies eq. (29) below for all δ > 0 and all T. Indeed, the OPE coefficients
are more properly thought of as a sequence of (equivalence classes of) distributions, such that the difference
between the nth and mth terms in the sequence satisfies eq. (29) for δ = min(m, n). However, to avoid such an
extremely cumbersome formulation of our axioms and results, we will treat each OPE coefficient as a distribution.
⋆
5 Note that, when V (i) is not equal to M × C, i.e., when φ(i) is not a scalar field, then quantities like C(i)(i )
(1)
are, by definition, distributions taking values in a vector bundle. What we mean by the scaling degree here and in
similar equations in the following such as eq. (20) is the maximum of the scaling degrees of all "components" of
such a bundle-valued distribution.

11
measures the rate at which the coefficient of the identity 1 in the operator product expansion
⋆
of φ(i) (x1 )φ(i ) (x2 ) blows up as x1 → x2 . It will follow immediately from condition (C3)
below that dim(i⋆ ) = dim(i). Note also that dim(1) = 0.
For distributions u1 and u2 on M n+1 , we introduce the equivalence relation
u1 ∼ u2 (11)
to mean that the scaling degree of the distribution u = u1 − u2 about any point on the total
diagonal is −∞. However, it should be noted that in the formulation of condition (C8)
below and in the precise definition of the operator product expansion, eq. (39), we will
need to consider limits where different points approach the total diagonal at different rates.
We will then introduce a stronger notion of equivalence that, in effect, requires a scaling
degree of −∞ under these possibly different rates of approach. It is this stronger notion that
was meant in eq. (9) above.

(C1) Covariance: Let ρ : M → M′ be a causality preserving isometric embedding pre-

serving the orientations, spin-structures, and all other background structure, i.e. ρ is a
morphism in the category of background structures. We postulate that for each member of
the above collection (6), we have
(i )...(in ) (i )...(in )
(ρ∗ × · · · ρ∗ × ρ∗−1 )C( j)1 [M′ ] ∼ C( j)1 [M] . (12)

(C2) Identity element: We require that

(i )...(1)...(in ) (i )...(ik )(ik+1 )...(in )
C( j)1 (x1 , . . . , xn ; y) = C( j)1 (x1 , . . . xk−1 , xk+1 , . . . xn ; y) , (13)
where the identity 1 is in the k-th place. We also impose the following additional conditions
(i)
on C( j) (x; y), since these coefficients should merely implement a Taylor expansion of the
operator with label i localized at x in terms of operators with labels j localized at point y.
As in a Taylor series, we demand that these coefficients depend only polynomially on the
Riemannian normal coordinates of x relative to y (and are thus in particular smooth), and
that
(i) (i)
C( j) (x; x) = δ( j) id(i) (x) , (14)
where id(i) (x) is the identity map in the fiber over x of the vector bundle V (i). Since a Taylor
expansion of an operator at x around another point y involves the derivatives of the operators
considered at y, and because derivatives tend to increase the dimension, we further demand
(i)
that C( j) = 0 if dim( j) < dim(i). Finally, if we Taylor expand a quantity at x1 successively
around a second point x2 , and then a third point x3 , this should be equivalent to expanding
it in one stroke around the third point. Thus, we require that we have
C( j) (x1 ; x3 ) = ∑ C(k) (x1 ; x2 )C( j) (x2 ; x3 ) ,
(i) (i) (k)
(15)
(k)

where we note that the sum is only over the (finitely many) field labels k such that dim(k) ≤
dim( j).

12
(C3) Compatibility with ⋆: This relation encodes the fact that the underlying theory
will have an operation analogous to the hermitian adjoint of a linear operator. The require-
ment is
(i )...(i ) (i⋆ )...(i⋆ )
C( j)1 n (x1 , . . . , xn , y) ∼ C( jn⋆ ) 1 ◦ π(x1 , . . . , xn , y) , (16)
where π is the permutation
 
1 2 ... n n+1
π= . (17)
n n− 1 ... 1 n+1

(C4) Commutativity/Anti-Commutativity: Let σ be the permutation

 
1 ... k k + 1 ... n+1
σ= , (18)
1 ... k+1 k ... n+1

and let F(i) be the Bose/Fermi parity of φ(i) . Then we have

(σi )...(σin ) (i )...(in )
C( j) 1 ◦ σ(x1 , . . . , xn , y) = −(−1)F(ik )F(ik+1 )C( j)1 (x1 , . . . , xn , y) (19)

whenever xk and xk+1 are spacelike separated (and in the neighborhood of the diagonal in
M n+1 where the OPE coefficients are actually defined).

(C5) Scaling Degree: We require that

n o
(i)( j)
sd C(k) ≤ dim(i) + dim( j) − dim(k) . (20)

(C6) Asymptotic Positivity: Let i ∈ I be any given index. Then, for D ≥ 3 we pos-
tulate that dim(i) ≥ 0, and that dim(i) = 0 if and only if i = 1. Note that, because we are
taking the supremum over all spacetimes in eq. (10), our requirement that dim(i) > 0 for
(i)(i⋆ )
i 6= 1 does not imply that the scaling degree of C(1) for i 6= 1 is positive for all space-
times, since the coefficient may e.g. “accidentally” happen to have a lower scaling degree
for certain spacetimes of high symmetry, as happens for certain supersymmetric theories
on Minkowski spacetime.
(i)(i⋆ )
On a spacetime M where the scaling degree of C(1) is dim(i), we know that if we scale
the arguments of this distribution together by a factor of λ, and multiply by a power of λ less
than 2dim(i), then the resulting family of distributions cannot be bounded as λ → 0. For
our applications below, it is convenient to have a slightly stronger property, which we now
explain. Let X a be a vector field on M locally defined near y such that ∇a X b = −δa b at y.
Let Φt be the flow of this field, which scales points by a factor of e−t relative to y along the
flow lines of X a . If f is a compactly supported test section in V (i), we set fλ = λ−D Φ∗log λ f .

13
This family of test sections becomes more and more sharply peaked at y as λ → 0. We
postulate that, for any δ > 0 and any X a as above, there exists an f such that
 
 2dim(i)−δ
Z
(i)(i⋆ ) 
lim λ C(1) (x1 , x2 , y) fλ (x1 ) fλ (x2 ) dµ1 dµ2  = ∞ ,
¯ (21)
λ→0 M×M

uniformly in y in some neighborhood. This statement is slightly stronger than the statement
that the scaling degree of our distribution is dim(i) on M, since the latter would only imply
that the rescaled distributions under the limit sign in eq. (21) contain a subsequence that is
unbounded in λ for some test section F(x1 , x2 ) in V (i) ×V (i⋆ ), not necessarily of the form
f (x1 ) f¯(x2 ).
The reason for the terminology “asymptotic positivity axiom” arises from lemma 2
below. An alternative essentially equivalent formulation of this condition, which is related
to “quantum inequalities”, is given in Appendix B.
In D = 2 spacetime dimensions, the above form of the asymptotic positivity condition
is in general too restrictive. The reason is that in D = 2, there are usually many fields
φ(i) of dimension dim(i) = 0 different from the identity operator. For example, for a free
Klein-Gordon field, the basic field ϕ and all its Wick-powers have vanishing dimension—
in fact, their OPE-coefficients have a logarithmic scaling behavior. A possible way to deal
with this example would be to consider only composite fields containing derivatives, as
this subspace of fields is closed under the OPE. For this subspace of fields, the asymptotic
positivity condition would then hold as stated. Another possibility is to introduce a suitably
refined measure of the degree of divergence of the OPE coefficients also taking into account
logarithms. Such a concept would clearly be sensible for free or conformal field theories in
D = 2, and it would also be adequate in perturbation theory (to arbitrary but finite orders).
A suitable refinement of the above asymptotic positivity condition could then be defined,
and all proofs given in the remainder of this paper would presumably still hold true, with
minor modifications. For simplicity, however, we will not discuss this issue further in this
paper, and we will stick with the asymptotic positivity condition in the above form.

(C7) Spectrum condition: The spectrum condition roughly says that the singularities
of a field product ought to be of “positive frequency type,” and is completely analogous the
condition imposed on states that we will impose below: We demand that, near the diagonal,
the wave front set (see Appendix A) of the OPE coefficient satisfies
(i )...(in )
WF(C( j)1 ) ⊂ Γn (M) × Z ∗ M , (22)

where the last factor Z ∗ M is the zero section of T ∗ M and corresponds to the reference
point y in the OPE, and where the set Γn (M) ⊂ T ∗ M n \ {0} is defined as follows. Consider
embedded graphs G(~x,~y,~p) ∈ G m,n in the spacetime manifold M which have the following
properties. Each graph G has n so-called “external vertices”, x1 , . . . , xn ∈ M, and m so-
called “internal” or “interaction vertices” y1 , . . . , ym ∈ M. These vertices are of arbitrary
valence, and are joined by edges, e, which are null-geodesic curves γe : (0, 1) → M. It
is assumed that an ordering of the vertices is defined, and that the ordering among the

14
external vertices is x1 < · · · < xn , while the ordering of the remaining interaction vertices
is unconstrained. If e is an edge joining two vertices, then s(e) (the source) and t(e) (the
target) are the two vertices γe (0) and γe (1), where the curve is oriented in such a way that it
starts at the smaller vertex relative to the fixed vertex ordering. Each edge carries a future
directed, tangent parallel covector field, pe , meaning that ∇γ̇e pe = 0, and pe ∈ ∂V + . With
this notation set up, we define

Γn,m (M, g) = (x1 , k1 ; . . . ; xn , kn ) ∈ T ∗ M n \ {0} | ∃ decorated graph G(~x,~y,~p) ∈ G m,n

such that yi ∈ J + ({x1 , . . . , xn }) ∩ J −({x1 , . . . , xn }) for all 1 ≤ i ≤ m,

such that ki = ∑ pe − ∑ pe for all xi and
e:s(e)=xi e:t(e)=xi

such that 0 = ∑ pe − ∑ pe for all yi , (23)
e:s(e)=yi e:t(e)=yi

where J ± (U ) is the causal future resp. past of a set U ⊂ M, defined as the set of points that
can be reached from U via a future resp. past directed causal curve. We set

Γn = Γm,n .
[
(24)
m≥0

Note in particular that the microlocal spectrum condition implies that the dependence of
our OPE coefficients (6) on the reference point y is smooth. When the spacetime is real
analytic, we require a similar condition for the "analytic" wave front set [22] WFA of the
OPE coefficient.
Our formulation of the microlocal spectrum condition is a weaker condition than that
previously proposed in [5], based on earlier work of [24]. The microlocal spectrum con-
dition of [5] is satisfied by the correlation functions of suitable Hadamard states in linear
field theory, but need not hold even perturbatively for interacting fields. In essence, our
formulation allows for the presence of interaction vertices, thus weakening the condition
relative to the free field case. Our condition can be shown to hold for perturbative interact-
ing fields [15].

(C8) Associativity: Following [15], a notion of associativity is formulated by consid-

ering configurations (x1 , . . . , xn ) of points in M n (where n > 2) approaching a point y ∈ M
at different rates. For example, if we have 3 points (x1 , x2 , x3 ), we may consider all points
coming close to each other at the same rate, i.e., assume that their mutual distances are of
order ε, where ε → 0. Alternatively, we may consider a situation in which, say, x1 , x2 ap-
proach each other very fast, say, at rate ε2 , while x3 approaches x1 , x2 at a slower rate, say at
rate ε. The first situation corresponds, intuitively, to the process of performing the OPE of
a triple product of operators ”at once”, while the second situation corresponds to first per-
forming an OPE in the factors corresponding to x1 , x2 , and then successively performing a
second OPE between the resulting fields and the third field situated at x3 . Obviously, for an

15
 arbitrary number n of points, there are many different possibilities in which configurations
can come close. We classify the different possibilities in terms of “merger trees,” T. Each
merger tree will give rise to a separate associativity condition.
For this, one constructs curves in M n parametrized by ε, which are in M0n (the space
M n minus all its diagonals) for ε > 0, and which tend to a point on the diagonal as ε → 0.
These curves are labeled by trees T that characterize the subsequent mergers of the points
in the configuration as ε → 0. A convenient way to formally describe a tree T (or more
generally, the disjoint union of trees, a “forest”) is by a nested set T = {S1 , . . . , Sk } of
subsets Si ⊂ {1, . . . , n}. “Nested” means that two sets are either disjoint, or one is a proper
subset of the other. We agree that the sets {1}, . . . , {n} are always contained in the tree
(or forest). Each set Si in T represents a node of a tree, i.e., the set of vertices Vert(T)
is given by the sets in T, and Si ⊂ S j means that the node corresponding to Si can be
reached by moving downward from the node represented by S j . The root(s) of the tree(s)
correspond to the maximal elements, i.e., the sets that are not subsets of any other set. If
the set {1, . . . , n} ∈ T, then there is in fact only one tree, while if there are several maximal
elements, then there are several trees in the forest, each maximal element corresponding
to the root of the respective tree. The leaves correspond to the sets {1}, . . . , {n}, i.e., the
minimal elements. For example for a configuration of n = 4 points, a tree might look like
in the following figure, and the corresponding nested set of subsets is also given.

S0

S1 S2

S3 S4 S6 S5

T = {S0 , S1 , . . . , S6 }
S0 = {1, 2, 3, 4}, S1 = {1, 2}, S2 = {3, 4}, S3 = {1}, S4 = {2}, S5 = {3}, S6 = {4}
Figure 1.

In the following, we will consider only T with a single root. The desired curves ~x(ε)
tending to the diagonal are associated with T and are constructed as follows. First, we
construct Riemannian normal coordinates around the reference point y, so that each point
in a convex normal neighborhood of y may be identified with a tangent vector v ∈ Ty M. We
then choose a tetrad and further identify Ty M ∼
= RD , so that v is in fact viewed as an element
in RD . With each set S ∈ T, we now associate a vector vS ∈ RD , which we collect in a tuple

~v = (vS1 , . . . , vSr ) ∈ (RD )|T| , T = {S1 , . . . , Sr } , (25)

and we agree that v{1,...,n} = 0, and where |T| is the number of nodes of the tree, i.e. the

16
 number of elements of the set T. For ε > 0, we define a mapping
ψT (ε) : (RD )|T| 7→ (RD )|T| , (vS1 , . . . , vSr ) 7→ (xS1 (ε), . . . , xSr (ε)) (26)

by the formula
∑ εdepth(S ) vS
′
xS (ε) = ′ (27)
S′ ⊂S
where depth(S′ ) is defined as the number of nodes that connect S′ with the root of the
tree T. For ε sufficiently small, and ~v in a ball B1 (0)|T| , the vectors xS (ε) ∈ RD may be
identified with points in M via the exponential map. If the vectors vS satisfy the condition
that, vS′ 6= vS′′ for any S′ , S′′ that are connected to a common S by an edge, then the vector
(x{1} (ε), . . . , x{n} (ε)) ∈ M n does not lie on any of the diagonals, i.e., any pair of entries are
distinct from each other. Its value as ε → 0 approaches the diagonal of M n . The i-th point
in the configuration x{i} (ε) is obtained starting from y by following the branches of the tree
towards the i-th leaf, moving along the first edge by an amount ε in the direction of the
corresponding vS then the second by an amount ε2 in the direction of the corresponding
vS , and so fourth, until the i-th leaf is reached. The curve (x{1} (ε), . . . , x{n} (ε)) ∈ M n thus
represents a configuration of points which merge hierarchically according to the structure
of the tree T, as ε → 0. That is, the outermost branches of the tree merge at the highest
order in ε, i.e., at rate εdepth of branch , then the next level at a lower order, and so fourth, while
the branches closest to the root merge at the slowest rate, ε. The following figure illustrates
our definition.

root = y S0

εv1 εv2

S1 S2

ε2 v3 ε2 v4 ε2 v5 ε2 v6

x1 (ε) x2 (ε) x4 (ε) x3 (ε)

T = {S0 , S1 , . . . , S6 }
x1 (ε) = εv1 + ε2 v3 , x2 (ε) = εv1 + ε2 v4 , x3 (ε) = εv2 + ε2 v5 , x4 (ε) = εv2 + ε2 v6
Figure 2
Thus, the points are scaled towards the diagonal of M n , even though possibly at different
speeds, and the limiting element as ε → 0 is the element (y, . . . , y) on the diagonal6 .
Using the maps ψT (ε) we can define an asymptotic equivalence relation ∼T,δ for distri-
butions u defined on M |T| . For points within a convex normal neighborhood, and sufficiently
6 However, if the vector (x{1} (ε), . . . , x{n} (ε)) ∈ M n is alternatively viewed as an element of the “Fulton-
MacPherson compactification” Mcn of the configuration space M n , then its limiting value may be viewed alter-
natively as lying in the boundary ∂Mcn of the compactification, and the vectors ~v may be viewed as defining a

17
 small ε > 0, we can define the pull-back u ◦ ψT (ε). This may be viewed a distribution in
the variables vS ∈ RD , S ∈ T. We now define

u ∼T,δ 0 :⇐⇒ lim ε−δ u ◦ ψT (ε) = 0 for δ > 0 , (29)

ε→0+

in the sense of distributions defined on a neighborhood of the origin in (RD )|T| . We write
u ≈ 0 if u ∼T,δ 0 for all T and all δ. The condition that u ≈ 0 is stronger than the previously
defined condition u ∼ 0 [see eq. (11)], which corresponds to the requirement that u ∼T,δ 0
for all δ only for the trivial tree T = {{1}, . . . {n}, {1, . . . , n}}.
We can now state the requirement of associativity. Recall that if T is a tree with n leaves,
then ψT (ε) gives a curve in the configuration space of n points in M which representing the
process of a subsequent hierarchical merger of the points according to the structure of the
tree. If a subset of points in (x{1} (ε), . . . , x{n} (ε)) merges first, then one intuitively expects
that one should be able to perform the OPE in those points first, and then subsequently
perform OPE’s of the other points in the hierarchical order represented by the tree. We
will impose this as the associativity requirement. For example, if we have 4 points, and the
tree corresponds to the nested set of subsets T = {{1, 2}, {3, 4}, {1, 2, 3, 4}} as in the above
figure, the pairs of points x{1} (ε), x{2} (ε) respectively x{3} (ε), x{4} (ε) approach each other
at order ε2 , while the two groups then approach each other at a slower rate ε. We postulate
that7
(i )(i2 )(i3 )(i4 )
C(i51) (x1 , x2 , x3 , x4 ; y) ∼T

∑ C(i )
(i1 )(i2 ) (i )(i4 ) (i )(i7 )
6
(x1 , x2 ; x6 )C(i73) (x3 , x4 ; x7 )C(i56) (x6 , x7 ; y) . (30)
i6 ,i7

For the same product of operators, consider alternatively the tree T′ = {{1, 2, 3}, {1, 2, 3, 4}}.
The corresponding associativity relation for the OPE coefficient is now

∑ C(i )
(i )(i2 )(i3 )(i4 ) (i1 )(i2 )(i3 ) (i )(i4 )
C(i51) (x1 , x2 , x3 , x4 ; y) ∼T′ 6
(x1 , x2 , x3 ; x6 )C(i56) (x6 , x4 ; y) (31)
i6

It is important to note, however, that there is in general no simple relation between the right
hand sides of eqs. (30),(31) for different trees T and T′ .
The corresponding relation for arbitrary numbers of points, and arbitrary types of trees
is a straightforward generalization of this case, the only challenge being to introduce an
appropriate notation to express the subsequent OPE’s. For this, we consider maps~i : T → I
which associate with every node S ∈ T of the tree an element iS ∈ I , the index set labelling
coordinate system of that boundary, which thereby has the structure of a stratifold

∂Mcn ∼
[
= M[T] , (28)
T

with each face T corresponding to a lower dimensional subspace associated with a given merger tree [1].
7
Here, the distribution on the left side is viewed as a distribution in x1 , . . . , x7 , y with a trivial dependence on
x5 , x6 , x7 .

18
 the fields. If S ∈ T, we let S(1), S(2), . . . S(r) be the branches of this tree, i.e. the nodes
connected to S by a single upward edge. With these notations in place, the generalization
of eqs. (30) and (31) for an arbitrary number n of points, and an arbitrary tree T is as
follows. Let T be an arbitrary tree on n elements, and let δ > 0 be an arbitrary real number.
Then we have8
!


∑ ∏ C(iS )
(i1 )...(in ) (iS(1) )...(iS(r) )
C( j) (x1 , . . . , xn ; y) ∼T,δ xS(1) , . . . , xS(r) ; xS , (32)
~i∈Map(T,I) S∈T

where the sums are over~i with the properties that

i{1} = i1 , . . . , i{n} = in , i{1,...,n} = j . (33)
The sum over ~i is finite, with dim(iS ) ≤ ∆, where ∆ is a number depending on the tree T
and the real number δ. Furthermore, it is understood that x{1} = x1 , . . . , x{n} = xn , and that
x{1,...,n} = y.

(C9) Analytic and smooth dependence: Due to requirement (C1), the OPE coeffi-
cients may be regarded as functionals of the spacetime metric. We require that the distribu-
(i )...(i )
tions C( j)1 n have an analytic dependence upon the spacetime metric. For this, let g(s) be
a 1-parameter family of analytic metrics, depending analytically on s ∈ R. Then the corre-
(i )...(i )
sponding OPE-coefficients C( j)1 n are distributions in x1 , . . . , xn , y that also depend on the
parameter s. We demand that the dependence on s is "analytic". It is technically somewhat
(i )...(i )
involved to define what one precisely means by this, because C( j)1 n itself is not analytic,
but instead a distribution in the spacetime points. The appropriate way to make this def-
inition rigorous was provided in [20, 16]. Similarly, if the spacetime is only smooth, we
require a corresponding smooth variation of the OPE coefficients under smooth variations
of the metric.

3 Construction of the QFT from the OPE coefficients

Now that we have stated in detail all the desired properties of the OPE coefficients, we
are ready to give the precise definition of quantum field theory. A quantum field theory
in curved spacetime associated with a collection of OPE coefficients satisfying the above
properties is the pair {A (M), S (M)} consisting of a *-algebra A (M) and a space of states
S (M) on A (M) that is canonically defined by the operator product coefficients, C (M),
for any choice of the background spacetime structure M. The algebra A (M) is defined as
follows. To begin, let Free(M) be the free *-algebra generated by all expressions of the
form φ(i) ( f ) with i ∈ I , and f a compactly supported test section in the vector bundle V (i)
associated with the tensor or spinor character of φ(i) . The algebra A (M) is obtained by
factoring Free(M) by a set of relations, which are as follows.
8
Here, the distribution on the left is regarded as a distribution in xS ; S ∈ T, with a trivial dependence on the xS
with S not equal to {1}, . . ., {n}, {1, . . ., n}.

19
A1) Identity: We have φ(1) ( f ) =
R
f dµ · 1.

A2) Linearity: For any complex numbers a1 , a2 , any test sections f1 , f2 , and any field
φ(i) , we have φ(i) (a1 f1 + a2 f2 ) = a1 φ(i) ( f1 ) + a2 φ(i) ( f2 ). The linearity condition might be
viewed as saying that, informally,
Z
φ (f) =
(i)
φ(i) (x) f (x)dµ (34)
M

is a pointlike field that averaged against a smooth weighting function. We shall often use
the informal pointlike fields as a notational device, with the understanding that all identities
are supposed to be valid after formally smearing with a test function.

A3) Star operation: For any field φ(i) , and any test section f ∈ Sect0 (V (i)), let f¯ ∈
Sect0 (V̄ (i)), be the conjugate test section. Then we require that
⋆
[φ(i) ( f )]∗ = φ(i ) ( f¯) . (35)

A4) Relations arising from the OPE: Let K ⊂ I be a subset of the index set, and let
c(i) , i ∈ K be scalar valued differential operators [i.e., differential operators taking a section
V (i) to a scalar function on M], such that
n o
sd ∑ (c̄(i⋆ ) ⊗ c( j) )C(k) v(k) < 0 ,
(i⋆ )( j)
(36)
i, j∈K

for all v(k) ∈ V (k) and all k ∈ I . Then we impose the relation

0= ∑ c(i) φ(i) ( f ) (37)

i∈K

for all f ∈ C0∞ (M), where the differential operators act in the sense of distributions. This
relation can be intuitively understood as follows. Let O ( f ) be the smeared quantum field
defined by the right side of eq. (37). If we consider the OPE of the quantity hO (x1 )O (x2 )∗ iω
in some state, then the scaling degree requirement (36) implies that the limit of this quantity
as x1 , x2 → y is equal to 0. Heuristically, this implies that O (y) is a well-defined element
at a sharp point y, not just after smearing with a test function f as in eq. (37). Since
hO (y)O (y)∗ iω = 0 in all states, one would heuristically conclude that also O (y) should be
0 as an algebra element. This is what our requirement states.
The above requirement serves to eliminate any redundancies in the field content arising
e.g. from initially viewing, say, a field ϕ and ϕ as independent fields, or from initially
specifying a set of linearly dependent fields. More nontrivially, this requirement should
also serve to impose field equations in A (M). For example, in λϕ4 -theory, we expect that
a field equation of the form ϕ − m2 ϕ − λϕ3 = 0 should hold, where ϕ3 is a composite

20
field in the theory that should appear in the operator product expansion of three factors of
ϕ. If such a field equation holds, then clearly ϕ − m2 ϕ − λϕ3 should have a trivial OPE
with itself, i.e., all OPE coefficients of the product of this operator with itself should have
arbitrary low scaling degree. Thus, in this example, if we take c(1) =  − m2 , c(2) = −λ,
and φ(1) = ϕ, φ(2) = ϕ3 , then eq. (36) should hold. Our requirement effectively demands
that field equations hold if and only if they are implied by the OPE condition eq. (36).

A5) (Anti-)commutation relations: Let φ(i1 ) and φ(i2 ) be fields, and let f1 and f2 be
test sections corresponding to their respective spinor or tensor character, whose supports
are assumed to be spacelike separated. Then the relation

φ(i1 ) ( f1 )φ(i2 ) ( f2 ) + (−1)F(i1 )F(i2 ) φ(i2 ) ( f2 )φ(i1 ) ( f1 ) = 0 . (38)

holds in A (M).

Having defined the algebra A (M), we next define the state space S (M) to consist of all
those linear functionals h . iω : A (M) → C with the following properties:

S1) Positivity: The functional should be of positive type, meaning that hA∗ Aiω ≥ 0 for
each A ∈ A (M). Physically, hAiω is interpreted as the expectation value of the observable
A in ω.

S2) OPE: The operator product expansion holds as an asymptotic relation. By this we
mean more precisely the following. Let φ(i1 ) , . . . , φ(in ) be any collection of fields, let δ >
0 be arbitrary but fixed, and let T be any merger tree as described in the associativity
condition. Let ∼δ,T be the associated asymptotic equality relation between distributions of
n spacetime points that are defined in a neighborhood of the diagonal which was defined in
the associativity condition [see eq. (29]. Then we require that
D E D E
∼δ,T ∑ C( j)1 n (x1 , . . . , xn ; y) φ( j) (y)
(i )...(i )
φ(i1 ) (x1 ) · · · φ(in ) (xn ) , (39)
ω j ω

where the sum is carried out over all j such that d( j) ≤ ∆, where ∆ is a number depending
upon the tree T, and the specified accuracy, δ.

S3) Spectrum condition: We have

WF hφ(i1 ) (x1 ) . . . φ(in ) (xn )iω ⊂ Γn (M) , (40)

where the set Γn (M) ⊂ T ∗ M n \ {0} was defined above in eq. (23).

21
 As part of our definition of a quantum field theory, we make the final requirement that
there is at least on state, i.e.,
S (M) 6= 0/ for all M. (41)
If the state space were empty, then this is a sign that the OPE is inconsistent, and does not
define a physically acceptable quantum field theory.
Remarks: (1) The OPE coefficients enter the construction of the algebra A (M) only via
condition (A4). However, they provide a strong restriction on the state space S (M) via
condition (S2).
(2) If ω ∈ S (M) and A ∈ A (M), then ω(A∗ · A) is a positive linear functional on A (M) [i.e.,
satisfying (S1)] which can also be shown to satisfy (S2). This functional can be identified
with a vector state in the Hilbert space representation of A (M) obtained by applying the
GNS construction to ω, and is therefore in the domain of all smeared field operators. It
is natural to expect that, in a reasonable quantum field theory, the OPE [i.e. (S3)] should
hold in such states as well, but this does not appear to follow straightforwardly within our
axiomatic setting.
(3) There are some apparent redundancies in our assumptions in that commutativity/anti-
commutativity conditions have been imposed separately on the OPE coefficients and the
algebra (see conditions (C4) and (A5)), and microlocal spectrum conditions have been im-
posed separately on the OPE coefficients and the states (see conditions (C8) and (S3)). It
is possible that our assumptions could be reformulated in such a way as to eliminate these
redundancies, e.g., it is possible that condition (A5) might follow from condition (C4) with
perhaps somewhat stronger assumptions about states. However, we shall not pursue these
possibilities here.
The construction of the pair {A (M), S (M)} obviously depends only upon the data en-
tering that construction, namely the set of all operator product coefficients C (M), as well
as the assignments i 7→ V (i) and i 7→ F(i) of the index set enumerating the fields with
tensor/spinor character, and with Bose/Fermi character. Thus, any transformation on field
space preserving the OPE and the Bose/Fermi character will evidently give rise to a corre-
sponding isomorphism between the algebras, and a corresponding map between the state
spaces. We now give a more formal statement of this obvious fact, and then point out some
applications.
Let L : I → I , i 7→ i′ = Li be a map with the following properties. (a) L1 = 1 (b) L
preserves the ⋆-operation on I , (c) L preserves the assignment of Bose/Fermi character,
i.e., F(i′ ) = F(i). Furthermore, assume that there is an embedding map ψ : M → M ′ (not
necessarily an isometry at this stage), and for each index i, there is a bundle map ψ(i)
characterized by the following commutative diagram (where i′ = Li)
ψ(i)
V (i) −−−−→ V (i′ )
  (42)
πM y yπM′
ψ
M −−−−→ M ′
where πM respectively πM′ are the bundle projections associated with the vector bundles
V (i) and V (i′ ) over M respectively M ′ that characterize the spinor/tensor character of the

22
field labelled by (i). Recalling that V (i⋆ ) is required to be equal to the hermitian conjugate
bundle V (i), and denoting by con j the operation of conjugation mapping between these
bundles, we require as a consistency condition that
conjM′ ◦ ψ(i) = ψ(i)⋆ ◦ conjM , (43)
for all indices i ∈ I . We say that a collection of OPE coefficients C (M) on M = (M, g, T, e)
and a collection C ′ (M′ ) on M′ = (M ′ , g′ , T ′ , e′ ) are equivalent if
(Li )...(Lin ) (i )...(in )
(ψ∗(i1 ) × · · · × ψ∗(in ) × ψ−1 ′ 1
( j)∗ )C (L j) [M′ ] ∼ C( j)1 [M] , (44)
see eq. (11). By simply going through the definitions of the algebra A (M) and the state
space S (M) it is then clear that the following (almost trivial) lemma holds.
Lemma 1. Under the consistency conditions (43) and (44), and assuming that ψ preserves
all background structure (i.e., ψ is an isometric embedding preserving the causality rela-
tions, orientations, and spin structures) the map αL : A (M) 7→ A (M′ )
(i) (i′ )
αL : φM ( f ) 7→ φM′ (ψ(i)∗ f ), i′ = Li (45)
defines a linear *-homomorphism. The dual map αvL between the corresponding state spaces
defines a map S (M′ ) → S (M).
Another way of stating this result is to view the maps L as described above as morphisms
in the category whose objects are the OPE-coefficient systems C (M). The above lemma
then says that the constructions of S (M) and of A (M) from C (M) are functorial in nature.
We now discuss some applications of the lemma:

Application 1: Consider the case where L = id, ψ : M → M ′ is an isometric embedding

preserving orientations and any other background structure, and ψ(i) is the natural bundle
map associated with ψ. Then (43) obviously holds, while (44) holds because of the lo-
cality and covariance property of the OPE coefficients. The map αL , whose existence is
guaranteed by the lemma, then corresponds to the map χρ discussed in the introduction. In
particular, the assignment
Background Structures → Algebras , M → A (M) (46)
is functorial, in the sense that if ρ is an arrow in the category of background structures—
i.e., an orientation and causality preserving isometric embedding from one spacetime into
another—then χρ is the corresponding arrow in the category of *-algebras—i.e., an injective
*-homomorphism. Functoriality means that the assignment
Isometric Embeddings → Algebra Homomorphisms , ρ → χρ (47)
respects composition of arrows in the respective categories. Thus, in the terminology of
the introduction, the assumptions of our framework define a local, covariant quantum field
theory, and all fields φ(i) are local, covariant quantum fields. Furthermore, A : M → A (M)
is a functor in the sense of [7].

23
 Application 2: As the second application, consider a purely “internal” symmetry, L, i.e.,
consider the case that M = M′ , ψ = id. Then αL acts upon A (M) as a *-autmorphism, i.e.,
an internal symmetry. More generally, there could be an entire group G of such L with the
corresponding maps ψ(i) satisfying the composition law of the group, i.e., if we have L, L′ ,
then the map associated with L′′ = L ◦L′ and ψ′′ = ψ◦ψ′ is given by ψ′′(i) = ψ(i) ◦ψ′(i) . In this
case, we get an action of G on A (M) by ∗-automorphisms αL satisfying the composition
law αL ◦ αL′ = αL′′ .

Another simple consequence of our axioms is the following lemma. As above in (C6),
let X a be a vector field on M locally defined near y such that ∇a X b = −δa b at y. Let Φt be
the flow of this field9 , which scales points relative to y by a factor of e−t . If f is a compactly
supported test section in V (i), we set fλ = λ−D Φ∗log λ f for λ > 0.

Lemma 2. For i 6= 1, there exists an M, δ > 0, and test section f such that
Z
(i)(i⋆ )
lim λ2dim(i)−δ C(1) (x1 , x2 , y) fλ (x1 ) f¯λ (x2 ) dµ1 dµ2 = +∞ . (48)
λ→0 M×M
⋆
Proof: Let ω be an arbitrary state. Then we have hφ(i) ( fλ )φ(i ) ( f¯λ )iω ≥ 0, from the star
(i)(i⋆ )
axiom (C3), and the positivity of any state. Let M be such that the scaling degree of C(1)
equals 2dim(i). By the scaling degree and asymptotic positivity axioms (C5) and (C6), for
(i)(i⋆ )
sufficiently small δ > 0, the quantity 2dim(i) − δ is bigger than the scaling degree of C( j)
for any j 6= 1. Hence using eq. (39) and h1iω = 1, we have
  
(i⋆ ) ¯ (i)(i⋆ )
lim λ 2dim(i)−δ
φ ( fλ )φ ( fλ ) −C(1) ( fλ , fλ ) = 0 .
(i) ¯ (49)
λ→0 ω

By axiom (C6), we can choose f so that the second term tends to ∞ in absolute value as λ →
(i)(i⋆ )
0. However, the first term is always non-negative. Therefore, C(1) ( fλ , f¯λ ) → +∞.

4 Normal (anti-) commutation relations

In our axioms, we assumed that every field φ(i) was either a Bose or Fermi field, i.e., that
there was a consistent assignment i 7→ F(i) ∈ Z2 such that (19) holds. A priori, there
is no relation between F(i) and the Bose/Fermi character of the corresponding hermitian
conjugate field, i.e., F(i⋆ ). We will now prove that, in fact, F(i) = F(i⋆ ) as a consequence
of our axioms.

Theorem 1. We have “normal (anti-)commutation relations,” in the sense that

F(i) = F(i⋆ ) (50)

for all i ∈ I .
9 It follows that we can write Φlog λ (x) = y + λ−1(x − y) in a suitable coordinate system covering y.

24
 Proof: Let f and h be compactly supported test sections with support in a convex normal
neighborhood of a point y ∈ M. Let ω be a quantum state, i.e., a positive normalized linear
functional A (M) → C. Using (35), we see that positivity immediately implies that
D ⋆ ⋆
E
φ(i) ( f )φ(i) (h)φ(i ) (h̄)φ(i ) ( f¯) ≥ 0 . (51)
ω

Assume now that the supports of f , h are spacelike separated. Then, using the (anti- )com-
mutation relations eq. (38), it follows that
D ⋆ ⋆
E
p φ(i) ( f )φ(i ) ( f¯)φ(i ) (h̄)φ(i) (h) ≥ 0 , (52)
ω

⋆ ⋆ 2
where p = (−1)F(i)F(i )+F(i ) . Clearly, if we could show that the expectation value in this
expression were positive for some test sections, f , h, in some spacetime, then it would
follow that p = +1, i.e.
F(i)F(i⋆ ) + F(i⋆ )2 = 0 mod 2, (53)
and by reversing the roles of i and i⋆ , it would also follow that

F(i)F(i⋆ ) + F(i)2 = 0 mod 2, (54)

from which the statement F(i) = F(i⋆ ) modulo 2 would follow. Clearly, it suffices to show
that the expectation value is asymptotically positive for a 1-parameter family of test sections
fλ , hλ whose supports are scaled towards y ∈ M as λ → 0.
To show this, we consider the particular merger tree T of figure 1, and the corresponding
associativity condition. This tree corresponds to the scaling map ΨT (ε) :~x(1) →~x(ε), with


x1 (ε) = Expy εv1 + ε2 v3


x2 (ε) = Expy εv1 + ε2 v4


x3 (ε) = Expy εv2 + ε2 v5


x4 (ε) = Expy εv2 + ε2 v6


x5 (ε) = Expy εv1


x6 (ε) = Expy εv2
x7 (ε) = y . (55)

The corresponding associativity condition together with (S2) yields

 
(i⋆ ) (i⋆ )
lim ε −δ
φ (x1 (ε))φ (x2 (ε))φ (x3 (ε))φ (x4 (ε))
(i) (i)
ε→0 ω

∑
(i)(i⋆ ) (i⋆ )(i)
− C( j1 ) (x1 (ε), x2 (ε); x5 (ε))C( j2 ) (x3 (ε), x4 (ε); x6 (ε))
j1 , j2 , j3
  
( j )( j )
×C( j31) 2 (x5 (ε), x6 (ε); y) φ
( j3 )
(y) = 0 . (56)
ω

25
 This is to be understood in the sense of distributions in v1 , . . . , v6 . The sums go over all
indices with dim( jk ) ≤ ∆, where ∆ depends on δ. We now use axioms (C5) and (C6) to
analyze the scaling of the individual terms under the sum. It follows that
(i)(i⋆ ) (i⋆ )(i)
lim εα C( j1 ) (x1 (ε), x2 (ε); x5 (ε))C( j2 ) (x3 (ε), x4 (ε); x6 (ε))C( j31)
( j )( j2 )
(x5 (ε), x6 (ε); y) = 0
ε→0
(57)
if α > 8dim(i) − dim( j1 ) − dim( j2 ) − dim( j3 ) in the sense of distributions. Thus, the
term under the sum with the potentially most singular behavior as ε → 0 is the one where
dim( j1 ) = dim( j2 ) = dim( j3 ) is minimal, i.e. equal to 0, by axiom (C6). If these dimen-
sions vanish, then by axiom (C6), we have jk = 1.
Because h1iω = 1, and because the OPE coefficients involving only identity operators
are equal to 1 by the identity axiom, we have
 
(i⋆ ) (i⋆ )
lim ε 8dim(i)+δ
φ (x1 (ε))φ (x2 (ε))φ (x3 (ε))φ (x4 (ε))
(i) (i)
ε→0 ω

(i)(i⋆ ) (i⋆ )(i)
−C(1) (x1 (ε), x2 (ε), x5 (ε))C(1) (x3 (ε), x4 (ε), x6 (ε)) = 0 , (58)

for some δ > 0. We now integrate this expression against the test section f (v3 ) f¯(v4 )h̄(v5 )h(v6 ),
where f , h are of compact support, and change integration variables10 . Then we get for the
terms under the limit sign
 
(i⋆ ) ¯ (i⋆ ) (i)(i⋆ ) (i⋆ )(i)
φ ( fε )φ ( fε )φ (h̄ε )φ (hε ) −C(1) ( fε , f¯ε )C(1) (h̄ε , hε )
(i) (i)
(59)
ω

Here, we have defined fε (x) = ε−2D f ◦ α1 (ε, x) and hε (x) = ε−2D h ◦ α2 (ε, x) with αi (ε, . ) :
M → RD are the maps that are defined in a sufficiently small neighborhood of y by

α1 (ε, x) = ε−1 v1 + ε−2 Expy−1 (x) (60)

α2 (ε, x) = ε−1 v2 + ε−2 Expy−1 (x) . (61)

(i)(i⋆ )
Finally, we use the lemma 2 to conclude that there exist f , h such that both C(1) ( fε , f¯ε ) →
(i)(i⋆ )
+∞ and C(1) (hε , h̄ε ) → +∞ for some spacetime and some subsequence of ε → 0. In view
of eq. (59), it follows that
D ⋆ ⋆
E
lim φ(i) ( fε )φ(i ) ( f¯ε )φ(i ) (h̄ε )φ(i) (hε ) = +∞ , (62)
ε→0 ω

so the expectation value (52) is positive for the choice of test sections f , h given by fε , hε ,
see eq. (60), for sufficiently small ε. These test sections will have spacelike separated
support as long as Expy v1 and Expy v2 are spacelike, which we may assume to be the case.

10
We should also integrate against a test function in v1 , v2 . But the result (59) is already smooth in these
variables, so we can omit this smearing.

26
 5 The spin-statistics theorem and the PCT-theorem
In this section, we prove that appropriate versions of the spin-statistics theorem and the
PCT theorem hold in curved spacetime under our axiom scheme. We explicitly discuss the
case when the spacetime dimension is even, D = 2m and discuss the case of odd dimensions
briefly in remark (2) below the PCT theorem. The key ingredient in both proofs is a relation,
proven in in [16], between the OPE coefficients C (M) on the background structure M =
(M, g, T, e), and the OPE coefficients C (M) on the background structure

M = (M, g, −T, e) (63)

consisting of the same manifold M, the same metric g, the same orientation e, but the
opposite time orientation T . For even spacetime dimensions D = 2m, this relation is11 :
(
(i )...(i ) (i⋆ )...(i⋆ ) i−F( j) (−1)−U( j) ∏n iF(ik ) (−1)U(ik ) m even,
C( j)1 n [M] ∼ C( j1⋆ ) n [M] · −F( j)+U( j)−P( j) k=1 (64)
i ∏k=1 i k
n F(i )−U(ik )+P(ik ) m odd,
√
where i = −1. Here, we recall that with each quantum field φ(i) there is associated a
bundle V (i) over M corresponding to the tensor or spinor character of the quantum field. In
even spacetime dimensions D = 2m, such a bundle V (i) is a tensor product
⊗U(i) ⊗P(i)
V (i) = S− ⊗ S+ , (65)

where the first factor corresponds to the U (i) “unprimed-” and the second to the P(i)
“primed” spinor indices. More precisely, the bundles S± are defined as the ±1 eigenspaces
of the chirality operator12
1
Γ= (−1)(m−1)(2m−1)/2 ea1 a2 ...aD γa1 γa2 · · · γaD , Γ2 = idS , (66)
D!
acting on a 2m -dimensional complex vector bundle S over M of “Dirac spinors”. This bun-
dle S corresponds to a fundamental representation of the Clifford algebra (in the tangent
bundle) generated by the curved space gamma-matrices γa . There exists a linear isomor-
phism c : S → S, where S is the conjugate bundle of anti-linear maps Sv → C. Owing to
the relation Γ = (−1)m−1 c Γ c−1 , it follows that for even m, the bundles S± and S∓ are
isomorphic via c, while for odd m the bundles S± and S± are isomorphic, and we will
hence always identify these bundles. Thus, for even m the roles of primed and unprimed
spinor indices [i.e., respective tensor factors in eq. (65)] are exchanged when passing to the
⋆
hermitian adjoint φ(i ) of a quantum field φ(i) , while for odd m the roles are not exchanged.
We also note that the coefficients on the right side of eq. (64) are sections in the (tensor
product of the) spin bundles VM (i) referring to the time function T associated with M,
while the coefficients on the left side are sections in the spinor bundles VM (i⋆ ) defined via
11 Note that the “bar” symbol is referring to the PT -reversed background structure in the term on the left side,
while it means hermitian conjugation on the right side.
12 Here, the orientation D-form is normalized so that ga1 b1 . . . gaD bD e
a1 ...aD eb1 ...bD = −D!.

27
the opposite time orientation −T associated with M. As explained in [16], there is a natural
identification map between these bundles, and this identification map is understood in (64).
The proof of (64) makes use of the microlocal, analytical, and causal properties of the
OPE coefficients and proceeds via analytic continuation [16]. Since it is the main input
in the proofs of both the spin-statistics theorem and the PCT theorem, we now outline,
following [16], how (64) is proven within our axiomatic setting. We first consider the
case where g is analytic. Let y ∈ M and introduce Riemannian normal coordinates x =
(x0 , . . . , xD−1 ) ∈ RD about y. In this neighborhood of y, consider the 1-parameter family of
metrics g(s) for all |s| ≤ 1 defined by

g(s) = gµν (sx) dxµ dxν . (67)

Note that this family, in effect, interpolates between the given metric g = g(1) and the flat
Minkowski metric η = g(0) . We can expand g(s) in a power series in s about s = 0, which
takes the form ∞
gµν = ηµν + ∑ sn pµνβ1 ...βn−2 (y) xβ1 . . . xβn−2
(s)
(68)
n=2

where each p is a curvature polynomial p(y) = p[Rµνσρ (y), . . . , ∇(α1 · · · ∇α(n−2) ) Rµνσρ (y)]. It
can then be shown, using axiom (C1), that each OPE coefficient has an asymptotic expan-
sion of the form
∞
∑ qk (y) · (Wk )( j)
(i )...(in ) (i1 )...(in )
C( j)1 (x1 , . . . , xn , y) = (x1 , . . . , xn ) , (69)
k=0

where qk = (qk )µ1 ...µk is a curvature polynomial of the same general form as the p, and
where Wk = (Wk )µ1 ...µk are distributions defined on a neighborhood of 0 in (RD )n , valued
in the tensor product of (RD )⊗k with the spinor representation corresponding to the index
structure of the quantum fields in the operator product considered. They transform covari-
antly under the connected component SpinR (D − 1, 1)0 of the spin group of D-dimensional
Minkowski space.
Consider now the map ρ defined in a suitable convex normal neighborhood, O , of y
by (x0 , . . . , xD−1 ) 7→ (−x0 , . . . , −xD−1 ). In Minkowski spacetime, this map would define an
isometry which preserves spacetime orientation but reverses time orientation. In a general
curved spacetime, this map does not define an isometry. Nevertheless, we may view ρ as a
map
ρ : (O , g(s) , e, T ) → (O , g(−s) , e, −T ) . (70)
Viewed in this manner, it is easily seen that ρ preserves all background structure, i.e., it is
a causality preserving isometry that preserves orientations. Consequently, by the covari-
ance axiom (C1), the relation (64) is equivalent to a corresponding relation between the
OPE-coefficients on the spacetimes (M, g(s) , e, T ) and (M, g(−s) , e, T ), i.e., spacetimes with
different metrics but the same orientation and time orientation. If one now differentiates this
relation m-times with respect to s and puts s = 0 afterwards, then one can prove that (64) is

28
equivalent to the relation

Wk (x1 , . . . , xn ) = (71)
(
i−F( j) (−1)−U( j) ∏n iF(ik ) (−1)U(ik ) m even,
(−1)k π∗Wk (−xn , . . . , −x1 ) · −F( j)−U( j)+P( j) k=1
i ∏nk=1 iF(ik )−U(ik )+P(ik ) m odd,

for all k = 0, 1, 2, . . . . Here, π is the permutation (17), which acts by permuting the implicit
spinor/indices associated with the spacetime points xi .
We have thus reduced the proof of (64) to the proof of a statement about Minkowski
distributions Wk that transform covariantly under SpinR (D − 1, 1)0 . To prove it, one next
shows that Wk can be analytically continued, and that the analytic continuation transforms
covariantly under connected component of the identity in the complexified spin group
SpinC (D − 1, 1)0 . For this, one first proves, using the microlocal condition on the OPE-
coefficients, that, near 0, the analytic wave front set [22], WFA , of Wk satisfies

WFA (Wk ) ⊂ K. (72)

Here, K is a conic set defined in terms of the Minkowskian metric η and orientation e, T ,
by
 
K = (y1 , k1 ; . . . ; yn , kn ) ∈ T ∗ (×n Br ) \ {0}  ∃pi j ∈ V̄+ , n ≥ j > i ≥ 1:

ki = ∑ pi j − ∑ p ji for all i , (73)
j: j>i j: j<i

where V̄ + is the closure of the forward light cone V + in Minkowski space (defined with
respect to the time orientation T ),

V ± = {k ∈ RD | ηµν kµ kν > 0, kµ ∇µ T > 0}. (74)

The relation (72) is important because a theorem of [22] now guarantees that Wk is the
distributional boundary value

Wk (x1 , . . . , xn ) = B. V. Wk (x1 + iy1 , . . . , xn + iyn ) (75)

(y1 ,...,yn )∈K v →0

of a holomorphic function Wk (z1 , . . . , zn ) that is defined in the “half-space”

Wk : Br (0)n + iK v → C, some r > 0 (76)

where Br (0) is a ball of radius r in RD , where K v is the “dual cone” of all covectors
(y1 , . . . , yn ) ∈ (RD )n with the property that ∑ ki · yi > 0 for all (k1 , . . . , kn ) ∈ K. Using the
“edge of the wedge-theorem” [27], one proves that the holomorphic function Wk (z1 , . . . , zn )
transforms covariantly under the spin group SpinR (D − 1, 1)0 . As explained in more detail
in [16], one can use this in turn to prove the desired relation (71):
For D = 2m and m even, we consider the chirality element Γ in eq. (66) in flat space,
which is an element of the connected component of the identity of the complexified spin

29
group SpinC (D−1, 1)0 . It corresponds to the reflection element ρ : RD → RD , x 7→ −x of the
complexified Lorentz group SO(D − 1, 1; C) under the standard covering homomorphism
between these groups. This is an immediate consequence of the relation Γγa Γ−1 = −γa . Us-
ing the method of analytic continuation in overlapping patches, it can be shown that Wk may
be continued to a single valued analytic function on an extension of the domain in eq. (76),
and it can be shown that this continuation transforms covariantly under Γ. As explained
above, Γ acts as +idS+ on each tensor factor corresponding to a primed spinor index, and
as −idS− on each tensor factor associated with an unprimed spinor index. Therefore, if we
apply the transformation law of Wk under the element Γ, then we obtain relation (71) for
complex spacetime arguments, except that the order of the complex spacetime arguments
z1 , . . . , zn is reversed, and except for the factors relating to the Bose-Fermi character of the
fields involved. In order to be able to take the limit Im zi → 0 from within K v , we must pass
to so-called “Jost points” (z1 , . . . , zn ) in the extended domain of holomorphicity. For such
points, the (anti-)commutativity may be used, effectively allowing to permute the space-
time arguments in Wk in such a way that one can take the limit to real points from within K v
as required in eq. (75) afterwards. When permuting the arguments, we pick up the factors
related to the Bose/Fermi character of the fields.
For D = 2m and m odd, we consider instead the element iΓ of the connected component
of the identity of the complexified spin group SpinC (D − 1, 1)0 . This element again covers
the reflection ρ(x) = −x on D-dimensional Minkowski space. It acts as +i idS+ on each
tensor factor corresponding to a primed spinor index, and as −i idS− on each tensor factor
associated with an unprimed spinor index. Again it can be shown that Wk may be continued
analytically to a domain extending that in eq. (76), and that it transforms covariantly under
iΓ on the extended domain. The additional factor of i gives rise to the different factors in
eq. (71) compared to the case when m is even. The rest of the argument is identical to that
case.
This proves the PCT-relation (64) for analytic spacetimes, and even dimensions D =
2m. The validity of the corresponding relation for smooth spacetimes then follows from
the smoothness of the OPE-coefficients under smooth variations of the metric, since any
smooth metric can be viewed as the limiting member of a smooth 1-parameter family of
metrics g(λ) that are analytic for λ > 0 and smooth for λ = 0. The differences in the state-
ment and proof of the PCT-relation (64) for odd spacetime dimension D are described in
remark (2) below, following the proof of the PCT-Theorem.

We now are ready to state and prove the spin-statistics theorem within our framework.
The statement and proof of this theorem closely parallel the Minkowski spacetime version:
Theorem 2. (Spin-Statistics Theorem) If our axioms hold, then the spin statistics relation
F(i) = U (i) + P(i) mod 2 , (77)
also holds, i.e. fields with integer spin (= on half the number of primed + unprimed spinor
indices) have Bose statistics, while fields of half integer spin have Fermi statistics.
Proof: Let i ∈ I , and, as above, we restrict consideration to the even dimensional case
(i)(i⋆ )
D = 2m. Consider the PCT-relation (64) for the OPE-coefficient C(1) . This condition can

30
be written as
( ⋆ ⋆
(i)(i⋆ ) (i⋆ )(i) iF(i)+F(i ) (−1)U(i )+U(i) m even,
C(1) (x1 , x2 , y)M ∼ C(1) (x2 , x1 ; y)M · F(i)+U(i)−P(i)+F(i⋆ )+U(i⋆ )−P(i⋆ ) (78)
i m odd,

where we have used the hermitian conjutation axiom, and where we have used that F(1) = 0
since the identity is always a Bose field, by the identity axiom. When m is even, then
U (i⋆ ) = P(i) because conjugation of a spinor exchanges the number of primed and un-
primed indices. Furthermore, F(i) = F(i⋆ ) by Theorem 1, so we obtain
(i)(i⋆ ) (i⋆ )(i)
C(1) (x1 , x2 , y)M ∼ (−1)F(i)+U(i)+P(i)C(1) (x2 , x1 ; y)M . (79)

When m is odd, U (i⋆ ) = U (i) and P(i⋆ ) = P(i), because conjutation of a spinor does not
change the number of primed and unprimed spinor indices in that case. Using this, we
again obtain the expression eq. (79) when m is odd.
We now smear this expression with the test section fλ (x1 ) f¯λ (x2 ), where

fλ (x) = λ−D f [y + λ−1 (x − y)] . (80)

We are taking here an f of compact support in a sufficiently small neighborhood of y cov-

ered by some coordinate system, and y + λ−1 (x − y) is computed in this arbitrary coordinate
system. Denote by dim(i) the dimension of the field labelled by i. It follows from eq. (79)
that
Z
(i)(i⋆ )
λ2dim(i)−δ C(1) (x1 , x2 ; y)M fλ (x1 ) f¯λ (x2 ) dµ1 dµ2
Z
(i⋆ )(i)
− p λ2dim(i)−δ C(1) (x2 , x1 ; y)M fλ (x1 ) f¯λ (x2 ) dµ1 dµ2 → 0 (81)

as λ → 0, where p = (−1)F(i)+U(i)+P(i) . But the first term goes to +∞ by lemma 2, while

the second term goes to −p · ∞ for suitable f . Thus, these terms can only cancel for small λ
if we have p = +1, meaning that F(i) + P(i) +U (i) = 0 modulo 2. Hence the spin-statistics
relation must hold.

Next, we state and prove the PCT theorem, the formulation of which is quite differ-
ent from the Minkowski spacetime version (see the discussion below). Again, we restrict
consideration here to D = 2m, and describe the differences occurring in odd dimensions in
remark (2) below:

Theorem 3. (PCT-Theorem) Given a background structure M = (M, g, e, T ), we write M =

(M, g, e, −T ). In spacetime dimension D = 2m, define the anti-linear map θPCT
M : A (M) →
A (M) by (
(i) (i) iF(i) (−1)U(i) when m is even,
θM : φM ( f ) 7→ φM ( f ) · F(i)+U(i)−P(i)
PCT ∗
(82)
i when m is odd.

31
Then θPCT
M is an anti-linear *-isomorphism such that the diagram

M θPCT
A (M) −−−−→ A (M)
 
χρ y yχρ (83)
θPCT
′ ′
A (M′ ) −−−
M
−→ A (M )
as well as the diagram
′ θPCT
′
v
S (M ) −−M−−→ S (M′ )
  v
χvρ y yχρ (84)
θPCT
M
v
S (M) −−−−→ S (M)
commute for every isometric, causality and orientation preserving embedding ρ : M → M′ .
Here χvρ denotes the dual of the linear map χρ , and θPCT v denotes the dual of θPCT .

Proof: The proof of this theorem is, in essence, an application of lemma 1. In the notation
of lemma 1, we choose M = (M, g, e, T ), we choose M′ = M, we take ψ = id and we
choose L : i 7→ i⋆ . For any index i, we define ψ(i) to be the composition of the natural
anti-linear bundle map from VM (i) to VM (i⋆ ) that is implicit in the formula (82) with the
multiplication map by iF(i) (−1)U(i) when m is even and by iF(i)+U(i)−P(i) when m is odd.
From this definition we then have
( ⋆ ⋆
(−1)F(i)/2+F(i )/2+U(i)−U(i ) · ψ(i⋆ ) ◦ conjM m even,
conjM ◦ ψ(i) = F(i)+F(i⋆ )−U(i)−U(i⋆ )+P(i)+P(i⋆) (85)
i · ψ(i⋆ ) ◦ conjM m odd,

where con jM is the anti-linear map that sends a spinor to the hermitian conjugate spinor
on M. Now for even m the number of unprimed spinor indices associated with V (i) is
precisely equal to the number of primed indices P(i⋆ ) associated with V (i⋆ ), because V (i⋆ )
is assumed to be equal to V (i). Thus, P(i) = U (i⋆ ). Furthermore, we have F(i) = F(i⋆ ) by
Theorem 1, and F(i) = U (i) + P(i) mod 2 by the spin-statistics theorem. Consequently, the
compatibility condition for the *-operation holds when m is even, i.e.,

conjM ◦ ψ(i) = ψ(i⋆ ) ◦ conjM . (86)

When m is odd, then U (i) = U (i⋆ ) and P(i) = P(i⋆ ), and the compatibility condition again
holds because of F(i) = F(i⋆ ) and the spin-statistics theorem. Thus, we have shown that
the first input in lemma 1 holds. The second input is the compatibility of the OPE coeffi-
cients on M and M′ , eq. (44). That condition is essentially equivalent to the relation (64),
except that the latter relation also involves an additional complex conjugation of the OPE-
coefficient. However it is immediately seen that this will result only in the following dif-
ference in the conclusion of lemma 1: Instead of the linear *-homomorphism as provided
by this lemma, we now find that the PCT-map θPCT defined by eq. (82) yields an anti-linear
*-homomophism.

32
Remarks: (1) The above formulation of the PCT-theorem was suggested in [16]. As
noted in [9] the theorem can be stated in the language of functors by saying that the functors
M → A (M) and M → A (M) = A (M) are equivalent.
(2) In odd spacetime dimensions D = 2m + 1, the chirality operator Γ of eq. (66) is pro-
portional to the identity in S. Thus, in this case there is no decomposition S = S+ ⊕ S− as
in the even dimensional case, and there is consequently no difference between “primed”
and “unprimed” spinors. If we denote the number of spinor indices of a quantum field by
N(i) (i.e., the bundle associated with the label i the field is V (i) = S⊗N(i) ), then the factors
in formula (64) are now i−F( j)−N( j) ∏ iF(ik )+N(ik ) . In the proof of this formula, one must
now consider the map ρ : (x0 , x1 , . . . , xD−1 ) 7→ (−x0 , −x1 , . . . , +xD−1 ). This corresponds
again to a change of time orientation in Minkowski spacetime which leaves the spacetime
orientation invariant. The rest of the proof is similar.
We now explain the relation of the above formulation of the PCT theorem to to the usual
PCT theorem in Minkowski spacetime (see e.g. [27]). Changing T → −T while keeping e
unchanged is equivalent to changing parity (i.e., the spatial orientation s of a Cauchy surface
Σ induced by e and T via dT ∧ s = e) and time (i.e., the time function). Furthermore, the
field appearing on the right side of eq. (82) is usually referred to as the “charge conjugate
(i)
field” to φM ( f ). Thus our formulation of the PCT theorem asserts that the theory is indeed
invariant under simultaneous PCT-reversal in the sense that the theory on M is “the same”
as the theory on M with the fields replaced by their charge conjugates. However, note that
our PCT theorem relates theories on the two different background structures, M and M. By
contrast, the usual PCT theorem in Minkowski spacetime provides a symmetry of the theory
defined on a single background structure, namely Minkowski spacetime with a fixed choice
of orientation and time orientation. Indeed, the usual formulation of the PCT theorem in
Minkowski spacetime asserts the existence of an anti-unitary operator Θ : H → H on the
Hilbert space, H , of physical states such that, if ρ denotes the isometry on the, say, even-
dimensional Minkowski spacetime defined by

ρ : (x0 , x1 , . . . , xD−1 ) 7→ (−x0 , −x1 , . . . , −xD−1 ), (87)

and if φC is the charge conjugate field associated with φ defined by eq. (82), then AdΘ φ(ρ(x)) ≡
Θφ(ρ(x))Θ† = φC (x).
The relationship between these formulations can be seen as follows. Start with our for-
mulation of the PCT theorem. The isometry ρ maps Minkowski spacetime M = (RD , η, e, T )
with a given choice of orientation e and time orientation T , to M = (RD , η, e, −T ). Thus, by
“application 1” of lemma 1, we know that there is a *-homomorphism χρ : A (M) → A (M)
mapping the quantum fields φM ( f ) on M to the “same” quantum fields φM (ρ∗ ( f )) on M̄.
Thus, if we define
AdΘ := χ−1ρ ◦ θM .
PCT
(88)
we obtain a result that is essentially equivalent to the usual Minkowski version (as suitably
reformulated in an algebraic setting). Conversely, if we start with the usual formulation of
the PCT theorem and if we define quantum field theory on A (M) in terms of quantum field

33
 theory on M by means of the map χρ , then we obtain a version essentially equivalent to our
formulation by setting
θPCT
M = χρ ◦ AdΘ . (89)
Although the above formulations are essentially equivalent in Minkowski spacetime, in
a general spacetime, there does not exist any discrete isometry analogous to ρ. Thus, in
general we only have a PCT theorem describing the relation between the theory defined
on different backgrounds M and M. Of course in the case of a spacetime that admits
an isometry ρ mapping (e, T ) to (e, −T ) (as occurs, e.g. in Schwarzschild and deSitter
spacetimes), then a “same background structure” version of the PCT theorem can be given
via eq. (88).
The example of a Robertson-Walker spacetime

g = −dt 2 + a(t)2 dx · dx (90)

with a(t) a strictly increasing function of t, may be useful for clarifying the physical mean-
ing of our formulation of the PCT theorem in a general curved spacetime. If we choose
the time orientation T = t, then the above metric describes an expanding universe, while if
we take T = −t, it describes a corresponding contracting universe (with opposite choice of
spatial orientation since we keep e fixed). In essence, our formulation of the PCT theorem
relates phenomena/processes occuring in the expanding universe eq. (90) to corresponding
processes (involving the charge conjugate fields and also a reversal of parity) in the corre-
sponding contracting universe. Since the metric eq. (90) has no time reflection isometry ρ,
there are no relations implied by the PCT theorem between phenomena/processes occurring
in the expanding universe, eq. (90) with T = t. As a concrete illustration of this, suppose
that it were possible to give a definition of “particle masses” in curved spacetime—although
it is far from obvious that any such useful notion exists. The PCT theorem would then im-
ply that the mass of a particle in an expanding universe must be equal to the mass of the
corresponding antiparticle in a contracting universe. However, it would make no statement
about the masses of particles and antiparticles in the same universe13 .
Finally, it is worth emphasizing the nature of the action of our PCT map θPCT M′
v on

states. In Minkowski spacetime, the states of interest are normally assumed to lie in a sin-
gle Hilbert space H , and one often considers scattering states. In the usual formulation
of the PCT theorem in Minkowski spacetime, if ω describes an incoming scattering state
|p1 , . . . , pn ; ini ∈ H , then the corresponding state ω̄ under the PCT map may be identified
with an outgoing scattering state | − p1 , . . . , −pn ; outi ∈ H in the same Hilbert space. How-
ever, in our formulation of the PCT theorem, if ω is an in-state in M, then ω̄ is an in-state
in M. If the spacetime admits a time reflection isometry ρ, then ω̄ may be identified with
13 It is worth noting that the “third Sakharov necessary condition” for baryogenesis in the early universe (namely,

“interactions out of thermal equilibrium”) is based upon the (now seen to be unjustified) assumption that particle
and antiparticle masses are equal in an expanding universe. However, to the extent that particle and antiparticle
masses might differ in an expanding universe (even assuming that a useful notion of “particle mass” can be defined)
as a result of the lack of a time reflection symmetry, it would probably not even be possible to define a notion of
“thermal equilibrium” as a result of the lack of a time translation symmetry.

34
a corresponding out-state on M. However, for the Robertson-Walker spacetime eq. (90),
which does not admit a time reflection isometry, the PCT theorem relates in-states in an
expanding universe to in-states in the corresponding contracting universe.

6 Conclusions and outlook

In this paper we have proposed a new axiomatic framework for quantum field theory on
curved spacetime. We demonstrated that our new framework captures much of the same
content as the Wightman axioms by proving curved spacetime analogs of the spin-statistics
theorem and PCT theorem. In this section, we discuss some of the implications and poten-
tial ramifications of the viewpoint suggested by this new framework.
First, we address the issue of why we even seek an axiomatic framework for quantum
field theory in curved spacetime at all. Since gravity is being treated classically, quantum
field theory in curved spacetime cannot be a fundamental theory of nature, i.e., it must have
a limited domain of validity. In particular, we have focused considerable attention in this
paper on the OPE of quantum fields in curved spacetime, but the OPE is a statement about
the arbitrarily-short-distance singularity structure of products of fields. One would not
expect that quantum field theory in curved spacetime would give an accurate description
of nature at separations smaller than, say, the Planck scale. Consequently, why should
one seek a set of mathematically consistent rules governing quantum field theory that are
rigorously applicable only in a regime where the theory is not expected to be valid?
Our response to this question is that an exactly similar situation arises for classical field
theory. Classical field theory also is not a fundamental theory of nature, and its descrip-
tion of nature makes essential use of differentiability/smoothness properties of the classical
fields at short distance scales; one could not even write down the partial differential equa-
tions governing the evolution of classical fields without such short-distance-scale assump-
tions. However, if quantum field theory is any guide, the description of physical fields as
smooth tensor fields is drastically wrong at short distance scales. Nevertheless, classical
field theory has been found to give a very accurate description of nature within its domain
of validity, and we have obtained a great deal of insight into nature by obtaining a math-
ematically precise formulation of classical field theory. It is our belief that there exists a
mathematically consistent framework for quantum field theory in curved spacetime, and
that by obtaining and studying this framework, we will not only get an accurate description
of nature within the domain of validity of this theory, but we will also get important insights
and clues concerning the nature of quantum gravity.
We began our quest for the mathematical framework of quantum field theory in curved
spacetime by seeking to generalize the Wightman axioms to curved spacetime in as con-
servative a manner as possible. As described much more fully in the Introduction, there
are three key ingredients of the Wightman axioms that do not generalize straightforwardly
to curved spacetime: (1) Poincare invariance; (2) the spectrum condition; (3) existence of
a Poincare invariant state. We have seen in this paper that quantum field theory can be
generalized to curved spacetime by replacing these ingredients by the following: (1’) quan-
tum fields are locally and covariantly defined; (2’) the microlocal spectrum condition; (3’)

35
existence of an OPE. Although the formulation of conditions (1’) and (2’) differs signifi-
cantly from the formulation of conditions (1) and (2), the basic content of these conditions
is essentially the same. Indeed, there would be no essential difference in the formulation
of axiomatic quantum field theory in Minkowski spacetime if one replaced (1) and (2) with
(1’) and (2’). By contrast, as we shall elucidate further below, the replacement of (3) by
(3’) leads to a radically different viewpoint on quantum field theory.
The most important aspect of this difference is that the existence of a “preferred state”
no longer plays any role in the formulation of the theory. States are inherently non-local
in character, and the replacement of (3) by (3’)—along with the replacements of (1) with
(1’) and (2) with (2’)—yields a formulation of quantum field theory that is entirely local in
nature. In this way, the formulation of quantum field theory becomes much more analogous
to the formulation of classical field theory. Indeed, one can view a classical field theory as
being specified by providing the list of fields φ(i) occuring in the theory and the list of
local, partial differential relations satisfied by these fields. Solutions to the classical field
theory are then sections of the appropriate vector bundles that satisfy the partial differential
relations as well as regularity conditions (e.g., smoothness). Similarly, in our framework, a
quantum field theory is specified by providing the list of fields φ(i) occuring in the theory
and the list of local, OPE relations satisfied by these fields. Thus, the OPE relations play a
role completely analogous to the role of field equations in classical field theory. States—
which are the analogs of solutions in classical field theory—are positive linear maps on the
algebra A defined in section 3 that satisfy the OPE relations as well as regularity conditions
(in this case, the microlocal spectrum condition). We note that in classical field theory, the
field equations always manifest all of the symmetries of the theory, even in cases where
there are no solutions that manifest these symmetries. Similarly, in our formulation of
quantum field theory, the OPE relations that define the theory should always respect the
symmetries of the theory [30], even if no states happen to respect these symmetries.
Our viewpoint on quantum field theory is more restrictive than standard viewpoints in
that we require the existence of an OPE. On the other hand, it is less restrictive in that
we do not require the existence of a ground state. This latter point is best illustrated by
considering a free Klein-Gordon field ϕ in Minkowski spacetime

(2 − m2 )ϕ = 0 , (91)

where the mass term, m2 , is allowed to be positive, zero, or negative. In the standard
viewpoint, a quantum field theory of the free Klein-Gordon field does not exist in any
dimension when m2 < 0 and does not exist in D = 2 when m2 = 0 on account of the non-
existence of a Poincare invariant state. However, there is no difficulty is specifying OPE
relations that satisfy our axioms for all values of m2 and all D ≥ 2. In particular, for D = 4
we can choose the OPE-coefficient C of the identity in the OPE of ϕ(x1 )ϕ(x2 ) to be given
by

C(x1 , x2 ; y) = (92)
 
1 1
+ m j[m ∆x ] log[µ (∆x + i0t)] + m h[m ∆x ] ,
2 2 2 2 2 2 2 2
4π2 ∆x2 + i0t

36
 where ∆x2 = (x1 − x2 )2 and t = x01 − x02 . Here µ is an arbitrarily chosen mass scale and
√
j(z) ≡ 2i1√z J1 (i z) is an analytic function of z, where J1 denotes the Bessel function of
order 1. Furthermore, h(z) is the analytic function defined by
∞
(z/4)k
h(z) = −π ∑ [ψ(k + 1) + ψ(k + 2)] . (93)
k=0 k!(k + 1)!

with ψ the psi-function. This formula for the OPE coefficient—as well as the corresponding
formulas for all of the other OPE coefficients—is as well defined for negative m2 as for
positive m2 . Existence of states satisfying all of the OPE relations for negative m2 can be
proven by the deformation argument of [12], using the fact that such states exist for positive
m2 .
Although, in our framework, the Klein-Gordon field with negative m2 now joins the
ranks of legitimate quantum field theories, this theory is not physically viable because, in
all states, field quantities will grow exponentially in time14 . The potential importance of
the above example is that it explicitly demonstrates that the local OPE coefficients can have
a much more regular behavior under variations of the parameters of the theory as compared
with state-dependent quantities, such as vacuum expectation values. The OPE coefficients
in the above example are analytic in m2 . On the other hand, the 2-point function of the
global vacuum state is, of course, defined only for m2 ≥ 0 and is given by

h0|ϕ(x1 )ϕ(x2 )|0i = (94)

 
1 1
+ m j[m ∆x ] log[m (∆x + i0t)] + m h[m ∆x ] .
2 2 2 2 2 2 2 2
4π2 ∆x2 + i0t

This behaves non-analytically in m2 at m2 = 0 on account of the log m2 term. In other

words, in free Klein-Gordon theory, vacuum expectation values cannot be constructed per-
turbatively by expanding about m2 = 0—as should be expected, since no vacuum state
exists for m2 < 0—but there is no difficulty in perturbatively constructing the OPE coeffi-
cients by expanding about m2 = 0.
The above considerations raise the possibility that the well known failure of conver-
gence of perturbation series in interacting quantum field theory may be due to the non-
analytic dependence of states on the parameters of the theory, rather than any non-analytic
dependence of the fields themselves, i.e., that the OPE coefficients may vary analytically
with the parameters of the theory. In other words, we are suggesting the possibility that the
perturbation series for OPE coefficients may converge, and, thus, that, within our frame-
work, it may be possible to perturbatively construct15 interacting quantum field theories. In
14 In this respect, the quantum field of the Klein-Gordon field with negative m2 behaves very similarly to the
corresponding classical theory. The classical Klein-Gordon field with negative m2 has a well posed initial value
formulation and causal propagation (despite frequently expressed claims to the contrary). However, the classical
Klein-Gordon field with negative m2 is not physically viable since it is unstable, i.e., it admits solutions that grow
exponentially with time.
15 However, we are not suggesting that it should be possible to perturbatively construct states of the theory. Even

if one had the complete list of OPE coefficients, it would not be obvious how to construct states.

37
 order to do so, it will be necessary to define the basis fields φ(i) appropriately (see below
and [18]) and also to parametrize the theory appropriately (since a theory with an analytic
dependence on a parameter could always be made to appear non-analytic by a non-analytic
reparametrization). Aside from the free Klein-Gordon example above, the only evidence we
have in favor of convergence of perturbative expansions for OPE coefficients is the exam-
ple of super-renormalizable theories, such as λϕ4 -theory in two spacetime dimensions [17].
Here, only finitely many terms in a perturbative expansion can contribute to any OPE co-
efficient up to any given scaling degree, so convergence (up to any given scaling degree)
is trivial. By contrast, for λϕ4 -theory in two spacetime dimensions, the rigorously con-
structed, non-perturbative ground state n-point functions can be proven to be non-analytic
at λ = 0 [25].
In cases—such as free Klein-Gordon theory above—where the OPE coefficients can
be chosen to be analytic in the parameters of the theory, it seems natural to require that
the theory be defined so that this analytic dependence holds. This requirement has some
potentially major ramifications. Since vacuum expectation values of a products of fields
(i.e., a correlation function) would be expected to have a non-analytic dependence on the
parameters of the theory, it follows that if the OPE coefficients have an analytic dependence
on these parameters, then, even in Minkowski spacetime, some of the fields appearing
on the right side of the OPE of a product of fields must acquire a nonvanishing vacuum
expectation value, at least for some values of the parameters. This point is well illustrated
by the above Klein-Gordon example. It is natural to identify the next term in the OPE of
ϕ(x1 )ϕ(x2 ) [i.e., the term beyond the identity term, whose coefficient is given by eq. (92)]
as being ϕ2 (with unit coefficient), i.e.,
ϕ(x1 )ϕ(x2 ) ∼ C(x1 , x2 ; y)1 + ϕ2 (y) + ... . (95)
This corresponds to the usual “point-splitting” definition of ϕ2 , except that C(x1 , x2 ; y) now
replaces16 h0|ϕ(x1 )ϕ(x2 )|0i. If we take the vacuum expectation value of this formula (for
m2 ≥ 0, when a vacuum state exists) and compare it with eq. (94), we obtain
m2
h0|ϕ2 (y)|0i = − log(m2 /µ2 ) . (96)
16π2
Thus, we cannot set h0|ϕ2 |0i = 0 for all values of m2 . A similar calculation for the stress-
energy tensor of ϕ yields
m4
h0|Tab (y)|0i = log(m2 /µ2 )ηab . (97)
64π2
As in other approaches, the freedom to choose the arbitrary mass scale µ in eq. (92)
gives rise to a freedom to choose the value of the “cosmological constant term” in eq. (97).
16 The point-split expression using h0|ϕ(x1 )ϕ(x2 )|0i yields the “normal ordered” quantity : ϕ2 :. From the point
of view of quantum field theory in curved spacetime it is much more natural define ϕ2 via eq. (95) than by normal
ordering, since there is no generalization of normal ordering to curved spacetime that is compatible with a local and
covariant definition of ϕ2 [19] Indeed, it follows from the results of [19] that eq. (95) is the unique way to define ϕ2
compatible with desired properties, with the only ambiguities in the definition of ϕ2 arising from different allowed
choices of C(x1 , x2 ; y).

38
However, unlike other approaches, there is no freedom to adjust the value of the cosmo-
logical constant when m2 = 0 (i.e., we unambiguously obtain h0|Tab |0i = 0 in Minkowski
spacetime in that case), and the m2 -dependence of the cosmological constant is fixed (since
µ is not allowed to depend upon m2 ).
A much more interesting possibility arises for interacting field theories, such as non-
abelian gauge theories. In such theories, it is expected that there are “non-perturbative
effects” that vary with the coupling parameter g as exp(−1/g2 ). Such non-perturbative ef-
fects can potentially be very small compared with the natural scales appearing in the theory.
If such non-perturbative terms occur the vacuum expectation values of products of fields
and if—as we have speculated above—the OPE coefficients have an analytic dependence
on the coupling parameter, then composite fields—such as the stress-energy tensor—must
acquire nonvanishing vacuum expectation values that vary as exp(−1/g2 ). This possibility
appears worthy of further investigation.

Acknowledgments: This research was supported in part by NSF Grant PHY04-56619 to

the University of Chicago.

A Definition of the scaling degree and wave front set

of a distribution
In this appendix, we recall the notion of scaling degree and of the wave front set of a
distribution, which play an important role in the body of the paper. Quite generally, let u
be an distribution on Rn . We say that u has scaling degree d at x = 0, if d is the smallest
real number such that λδ u( fλ ) → 0 as λ → 0+, for all δ > d. Here, fλ (x) = λ−n f (λ−1 x)
denotes the function of compact support that is obtained by rescaling a smooth test function
f around x = 0, making it more and more sharply peaked at that point. The scaling degree
at an arbitrary point is obtained by simply translating the distribution u or the test function
f by the desired amount. We write sdx (u) = d for the scaling degree at a point x.
We next recall the definition of the wave front set of a distribution u on Rn . Let χ be
any smooth function of compact support. Then χu is evidently a distribution of compact
support, and its Fourier transform χ cu(k) defines an entire function of k ∈ Rn . For any
distribution v of compact support, we define its corresponding “singular set”, Σ(v) as the
collection of all k ∈ Rn such that

v(λk)| ≥ CλN ,
|b (98)

for some C > 0, and some N, and all λ > 0. We define the wave front set WFx (u) at a point
x ∈ Rn as the intersection
Σ(χu) ,
\
WFx (u) = (99)
χ:x∈supp χ

and we define WF(u) as the union

G
WF(u) = WFx (u) . (100)
x∈Rn

39
Each set WFx (u) is a conic set, in the sense that if k ∈ WFx (u), then so is tk for any t > 0,
and k = 0 is never in WFx (u). It immediately follows from the definition that WFx (u) = 0/
if and only if u can be represented by a smooth function in an open neighborhood of x.
In this sense, the wave-front set tells one at which points a distribution is singular. It also
contains information about the most singular directions in local momentum space, which
are represented by k ∈ WFx (u).
It turns out that both the scaling degree of a distribution at a point x, as well as the
wave front set at x are invariantly defined. By this one means the following. Let ρ : V →
U be a smooth diffeomorphism between open sets U,V ⊂ Rn . Let u be a distribution
supported in U , and let ρ∗ u be the pulled back distribution in V , where the pull-back is
defined by analogy with the pull back of a smooth function. Then it is easy to show that
sdx (ρ∗ u) = sdρ(x) (u). Furthermore, if x′ = ρ(x), and if we define ρ∗ (x′ , k′ ) = (x, k), where
k = [dρ(x)]v k′ , then one can show
WFx (ρ∗ u) = ρ∗ WFρ(x) (u) . (101)
These relations imply that the scaling degree and the wave front set can be invariantly
defined on an arbitrary manifold X , and that the wave-front set should be viewed as a
subset of T ∗ X .
In the body of the paper, we frequently consider the case X = M n+1 , and the scaling
degree at the point (y, y, . . . , y), i.e., points on the total diagonal. To save writing, this is
simply denoted sd u.

B Equivalent formulation of condition (C6)

In this appendix, we relate the scaling degree (C5) and asymptotic positive (C6) assump-
tions to other properties of the quantum field theory. Our first is just a repetition of a result
obtained in [8]:
Theorem 4. For any x ∈ M, and any (scalar) field T not equal to a multiple of the identity
operator, we define the convex set Sx ⊂ R by
Sx = {hT (x)iΦ | Φ a normalized state} (102)
Then, Sx = R for at least one spacetime M.

Remark: The statement means that pointlike hermitian fields T (x) are unbounded from
above and below, even though their classical counterpart (if the theory has a classical limit)
might be manifestly non-negative, such as the Wick square T = ϕ2 , or the energy density
T = Tab ua ub of a free Klein-Gordon field ϕ.
Proof: Choose any state Φ. We may assume that hT ( f )iΦ = 0, because if not, we just need
to consider instead T (x) by T (x) − hT (x)iΦ 1. Define
T(f)
A = cos α 1 + sin α 1/2
, (103)
hT ( f )T ( f )iΦ

40
and define a new normalized state by h .iα = hA . A∗ iΦ . Then

hT ( f )iα = a sin 2α + b(1 − cos 2α) (104)

where
1/2 1 hT ( f )T ( f )T ( f )iΦ
a = hT ( f )T ( f )iΦ , b= . (105)
2 hT ( f )T ( f )iΦ
Minimizing over α gives p
inf hT ( f )iΨ ≤ b − a2 + b2 . (106)
Ψ

Replacing f by − f also gives

p
sup hT ( f )iΨ ≥ −b + a2 + b2 . (107)
Ψ
R
Now, consider a test function f with f dµ = 1, f (x) 6= 0. By lemma 2, there exists a
spacetime and a state such that hT ( f¯λ )T ( fλ )iΦ → +∞ for some appropriately chosen f ,
where fλ (y) = λ−D f (x + λ−1 (y − x)). Hence a → ∞ as λ → 0. and therefore infΨ hT ( fλ )iΨ
becomes arbitrarily small as λ → 0. On the other hand fλ → δx in this limit, so the set Sx is
not bounded below. It similarly follows that it cannot be bounded above either.
Our next result is in some sense a converse to the above result:

Theorem 5. Let the set Sx be equal to R for a given spacetime M, and all hermitian op-
erators T not equal to a multiple of the identity. Then for any i, k ∈ I with i 6= 1, and any
sections v(i) of V (i) we have that

(k)(k⋆ ) (k)(k⋆ )
sd[C(i) (v(k) ⊗ v̄(k⋆ ) )] < sd[C(1) (v(k) ⊗ v̄(k⋆ ) )] . (108)

Remark: In generic spacetimes, we expect that the scaling degree of the right side is
equal to 2dim(k), where dim(k) is the dimension of the field φ(k) ; see the scaling degree
axiom (C5). We also expect the quantity on the left side to be equal to 2dim(k) − dim(i);
see again (C5). Thus, the result tells us that, in this situation, dim(i) > 0 unless φ(i) is the
identity field. Thus, in this sense, the assumption of the theorem implies the asymptotic
positivity axiom (C6), or—stated differently—the asymptotic positivity axiom is inconsis-
tent with not having Sx = R.
Proof: By assumption, we can find a state Φ such that hT (x)iΦ > A for each A ∈ R. Con-
sider an arbitrary, but fixed, finite collection φ(1) , . . . , φ(n) of fields. Each field is valued in
some vector bundle V (i). The set of expectation values of this collection of fields forms a
subset which we denote
n  o Mn

Kx = (hφ(1) (x)iΦ , . . . , hφ(n) (x)iΦ ) states Φ ⊂ V (i)x =: Vx . (109)
i=1

41
Because the set of states is convex (i.e., any convex linear combination of normalized states
is again a normalized state), the set Kx is a convex subset of Vx . We claim that, in fact,
Kx = Vx . Assume that this were not the case. Then, since any proper convex subset of
a finite dimensional vector space can be enclosed by a collection of planes, there exists a
collection of dual vectors c(i) ∈ V (i)v and an A ∈ R such that

Re ∑ c(i) v(i) < A , for all (v(1) , . . . , v(n) ) ∈ Kx . (110)

(i)

However, this would mean by definition that if T = Re ∑ c(i) φ(i) , then hT (x)iΦ < A for all
states Φ, a contradiction.
Assume that the statement of the theorem is not true. Let i = 1, . . . , n ∈ I be the field
labels, with i 6= 1 for which the inequality (108) does not hold. From what we have just
shown, if we define Kx as above, then Kx is equal to Vx . In particular, we may find states
Φ, Ψ, and nonzero v(i) with the property that

hφ(i) (x)iΨ = v(i) , hφ(i) (x)iΦ = −v(i) . (111)

Let fλ (x) = λ−D f (y + λ−1 (x − y)) be a test section in the dual of the space V (k), and let δ
be a real number which is bigger than 2dim(k), but smaller than the left side of eq. (108).
⋆
Using the fact that hφ(k) ( fλ )φ(k ) ( f¯λ )iΨ ≥ 0 for all λ, we find from the operator product
expansion that
λδ ∑ C(i) ( f¯λ , fλ ; x)v(i) → +∞ ,
(k⋆ )(k)
(112)
(i)

for at least one f and a subsequence of λ tending to 0. Applying a similar argument to the
state Φ gives that
− λδ ∑ C(i) ( f¯λ , fλ ; x)v(i) → +∞ ,
(k⋆ )(k)
(113)
(i)

for this subsequence of λ tending to 0. This is a contradiction, so the inequality (108) must
hold.

References
[1] S. Axelrod and I. M. Singer, “Chern-Simons perturbation theory. 2,” J. Diff. Geom.
39, 173 (1994)
[2] R. E. Borcherds, “Vertex Algebras, Kac-Moody Algebras, And The Monster,” Proc.
Nat. Acad. Sci. 83, 3068 (1986).
[3] H. Bostelmann, “Operator product expansions as a consequence of phase space prop-
erties,” J. Math. Phys. 46, 082304 (2005);
[4] H. Bostelmann: “Phase space properties and the short distance structure in quantum
field theory,” J. Math. Phys. 46, 052301 (2005)

42
 [5] R. Brunetti, K. Fredenhagen and M. Köhler: “The microlocal spectrum condition
and Wick polynomials on curved spacetimes,” Commun. Math. Phys. 180, 633-652
(1996)
[6] R. Brunetti and K. Fredenhagen: “Microlocal Analysis and Interacting Quantum Field
Theories: Renormalization on physical backgrounds,” Commun. Math. Phys. 208,
623-661 (2000)
[7] R. Brunetti, K. Fredenhagen and R. Verch, “The generally covariant locality prin-
ciple: A new paradigm for local quantum physics,” Commun. Math. Phys. 237, 31
(2003), [math-ph/0112041]; see also K. Fredenhagen, “Locally covariant quantum
field theory,” [arXiv:hep-th/0403007].
[8] C. J. Fewster: "Energy inequalities in quantum field theory," Proceedings of XIVth
International Congress on Mathematical Physics, ed. J.-C. Zambrini, 559 (2003)
[9] K. Fredenhagen: "Locally covariant quantum field theory," Proceedings of XIVth
International Congress on Mathematical Physics, ed. J.-C. Zambrini, 29 (2003)
[10] K. Fredenhagen and J. Hertel, “Local Algebras Of Observables And Point - Like Lo-
calized Fields,” Commun. Math. Phys. 80, 555 (1981); K. Fredenhagen and M. Jorss,
“Conformal Haag-Kastler nets, point - like localized fields and the existence of oper-
ator product expansions,” Commun. Math. Phys. 176, 541 (1996).
[11] I. Frenkel, J. Lepowsky and A. Meurman, “Vertex operator algebras and the Monster,”
Academic Press, Boston (1988)
[12] S.A. Fulling, F.J. Narcowich, and R.M. Wald: “Singularity Structure of the Two-
Point Function in Quantum Field Theory in Curved Spacetime, II,” Ann. Phys. 136
243 (1981).
[13] W. Fulton and R. MacPherson: “A compactification of configuration spaces,” Ann.
Math. 139 183 (1994)
[14] R. Haag and D. Kastler, “An Algebraic Approach To Quantum Field Theory,” J. Math.
Phys. 5, 848 (1964).
[15] S. Hollands, “The operator product expansion for perturbative quantum field theory
in curved spacetime,” Commun. Math. Phys. 273, 1 (2007) [arXiv:gr-qc/0605072].
[16] S. Hollands, “A general PCT theorem for the operator product expansion in curved
spacetime,” Commun. Math. Phys. 244, 209 (2004) [arXiv:gr-qc/0212028].
[17] S. Hollands and C. Kopper: in progress
[18] S. Hollands and R.M. Wald: in progress
[19] S. Hollands and R. M. Wald: “Local Wick Polynomials and Time Ordered Products
of Quantum Fields in Curved Space,” Commun. Math. Phys. 223, 289-326 (2001),
[gr-qc/0103074]
[20] S. Hollands and R. M. Wald: “Existence of local covariant time-ordered-products of
quantum fields in curved spacetime,” Commun. Math. Phys. 231, 309-345 (2002),
[gr-qc/0111108]

43
[21] S. Hollands, “Quantum field theory in terms of consistency conditions I: General
framework, and perturbation theory via Hochschild cohomology,” arXiv:0802.2198
[hep-th].
[22] L. Hörmander: The Analysis of Linear Partial Differential Operators I, Berlin,
Springer-Verlag 1983
[23] V. Kac, “Vertex algebras for beginners,” Providence, USA: AMS (1996) 141 p. (Uni-
versity lectures series. 10)
[24] M. J. Radzikowski, “Micro-local approach to the Hadamard condition in quantum
field theory on curved space-time,” Commun. Math. Phys. 179, 529 (1996).
[25] V. Rivasseau, “From perturbative to constructive renormalization,” Princeton, USA:
Univ. Pr. (1991) 336 p. (Princeton series in physics)
[26] B. Schroer, J. A. Swieca and A. H. Volkel, “Global Operator Expansions In Confor-
mally Invariant Relativistic Quantum Field Theory,” Phys. Rev. D 11, 1509 (1975).
[27] R. F. Streater and A. A. Wightman: PCT, Spin and Statistics and All That, New York,
Benjamin 1964
[28] R. M. Wald: Quantum Field Theory on Curved Spacetimes and Black Hole Thermo-
dynamics, The University of Chicago Press, Chicago (1990)
[29] K. G. Wilson, “Nonlagrangian Models Of Current Algebra,” Phys. Rev. 179, 1499
(1969).
[30] See appendix in: C. Bernard, A. Duncan, J. LoSecco, S. Weinberg: “Exact spectral-
function sum rules”, Phys. Rev. D 12, 792 - 804 (1975)
[31] W. Zimmermann, “Normal Products And The Short Distance Expansion In The Per-
turbation Theory Of Renormalizable Interactions,” Annals Phys. 77, 570 (1973) [Lect.
Notes Phys. 558, 278 (2000)].

44