Bootstrapping gravity: a consistent approach to energy-momentum self-coupling

Luke M. Butcher, Michael Hobson, and Anthony Lasenby

Astrophysics Group, Cavendish Laboratory, J J Thomson Avenue, Cambridge CB3 0HE, UK.
(Dated: October 14, 2009)
It is generally believed that coupling the graviton (a classical Fierz-Pauli massless spin-2 field) to its own
energy-momentum tensor successfully recreates the dynamics of the Einstein field equations order by order;
however the validity of this idea has recently been brought into doubt [1]. Motivated by this, we present a
graviton action for which energy-momentum self-coupling is indeed consistent with the Einstein field equations.
The Hilbert energy-momentum tensor for this graviton is calculated explicitly and shown to supply the correct
second-order term in the field equations; in contrast, the Fierz-Pauli action fails to supply the correct term.
A formalism for perturbative expansions of metric-based gravitational theories is then developed, and these
techniques employed to demonstrate that our graviton action is a starting point for a straightforward energy-
momentum self-coupling procedure that, order by order, generates the Einstein-Hilbert action (up to a classically
arXiv:0906.0926v2 [gr-qc] 14 Oct 2009

irrelevant surface term). The perturbative formalism is extended to include matter and a cosmological constant,
and interactions between perturbations of a free matter field and the gravitational field are studied in a vacuum
background. Finally, the effect of a non-vacuum background is examined, and the graviton is found to develop
a non-vanishing mass-term in the action.

PACS numbers: 04.20.Cv

I. INTRODUCTION where the numbers in parenthesis denote the powers of h

(1)
the term contains. Because G = 0 is the equation of mo-
It is a standard view in particle physics that the non-linearity tion for a massless spin-2 field h , the right-hand side of
of a field theory, such as those of Yang and Mills, can be (1) can be interpreted as this fields source. Thus a satisfying
equated with the notion that the field in question carries the physical picture suggests itself: the gravitational field h is
charge of the very interaction it mediates. This idea has been induced by the energy-momentum tensor of all fields T =
brought to bear on gravity many times, and various arguments matter + t , where t
T is gravitys own energy-momentum
[2, 3, 4, 5, 6, 7, 8] aim to derive general relativity from a lin- (2)
ear starting point by coupling gravity to the energy and mo- tensor, identified as G / . In actuality, however, this de-
mentum of all fields, including the gravitational field itself. scription cannot be formulated in a straightforward manner.
Despite the conventional wisdom that this self-coupling pro- Although the Fierz-Pauli action SFP is typically used to pre-
cess is already well understood, Padmanabhan has uncovered scribe the dynamics of a massless spin-2 field, its Hilbert
a number of serious problems with the standard arguments energy-momentum tensor2
[1]. Although we postpone an examination of Padmanabhans
1 SFP
analysis to appendix A, it suffices to express here what is, in t , (2)
our view, his most pertinent observation: one cannot start with
linear gravity, the Fierz-Pauli massless spin-2 action [1, 9],
(2)
and generate the higher-order corrections of general relativity is not proportional to G , and thus cannot be used as the
by coupling the gravitational field to its own Hilbert energy- source-term for the second-order field equations. As an alter-
momentum tensor. More succinctly: one cannot derive the native approach, one could introduce energy-momentum self-
Einstein equations by bootstrapping gravitons1 to their own coupling at the level of the action: because t is a function of
energy and momentum. h , adding the self-coupling term t h to the Lagrangian
To clarify the content of this observation, consider a per- yields a different result from adding t directly to the equa-
turbative expansion of the Einstein field equations G = tions of motion. Unfortunately, this procedure also fails to
T
matter about a Minkowski background: g
= + h .
Working to second-order in h , we obtain
2
Although other definitions of the energy-momentum tensor exist (see II C)
(1) (2)
G = G + T
matter
, (1) we must define t according to the Hilberts prescription (2) in order to
matter . This definition requires that S
maintain the analogy with T FP be co-
variantized (represented in arbitrary coordinates using a flat metric )
and a functional derivative taken with respect to the metric. It is important
to realise that even though is flat, the arbitrary variations required
[email protected]
to construct the functional derivative inevitably explore curved metrics in
1 In discussions of this nature, the word graviton is often used as a short- a neighbourhood of . Thus covariantization is not really sufficient:
hand for the classical massless spin-2 field. We follow this convention to the action must be generalised to a curved background spacetime. One
cohere with the literature, but stress that this graviton is in no way quantum of the key aims of this paper is to generalise SFP to curved spacetime in
mechanical. What is actually being referred to is a gravitational wave, a such a way that energy-momentum self-coupling is consistent with general
classical fluctuation in the geometry of spacetime. relativity.
 2

(2)
generate G / in the field equations. derived solely from the background geometry, and adopt the
Padmanabhan claims that these realizations bring to light a usual notational convenience of raising and lowering indices
previously neglected object S (see appendix A) which ap- with gab and gab .5
pears to codify the self-coupling of the gravitational field. Un- We posit that the dynamics, energy and momentum of the
fortunately, this object has many undesirable features: it is gravitational field hab , propagating in a background spacetime
not a tensor under general coordinate transformations, has no with metric gab , are all determined (to lowest-order) by the
clear physical interpretation, and fails to reveal any equiva- following action:
lence between the coupling of gravity to matter, and gravity to Z
itself. 1
S2 [gab , hab ] d4 x ghab (Gabcd + Habcd )hcd , (4)
We propose an alternative solution to this apparent incon- 2
sistency: the action for the graviton is not the Fierz-Pauli ac-
tion but is instead S2 given by (4), possessing a non-minimally where
coupled term that vanishes when the (vacuum) background
2 (ga(c gd)b gab gcd ) (c gd)(a b)
Gabcd 1 2
equations are enforced.3 We shall demonstrate that the
energy-momentum tensor of this action is the correct second- + 12 gab (c
d) + 1 gcd
2
(a
b) (5)
order contribution to the equation of motion, and furthermore,
that this action provides the starting point for a straightfor- is a differential operator representing the linearised Einstein
ward energy-momentum self-coupling procedure that gener- tensor (see appendix B) and
ates the Einstein-Hilbert action (modulo surface terms) to ar-
bitrary order. We conclude the discussion by extending our Habcd 1 1
formalism to non-vacuum spacetimes. 2 R(gac gdb + 2 gab gcd ) Rab gcd . (6)
Throughout the article we employ the abstract index nota- While Habcd has no obvious geometric interpretation, we in-
tion [10], with lower-case Roman indices indicating a tensors tend to show that its contribution to the action is necessary for
slots, and Greek indices serving to enumerate its compo- the consistency of energy-momentum self-coupling with gen-
nents in a particular coordinate system. The metric has sig- eral relativity. Further motivation for this ansatz is given in
nature (, +, +, +), 8 G/c4, and the Riemann and Ricci section III.
tensor are defined with the following conventions: Rabcd vb
Naturally, if we are to obtain general relativity without at
2[c d] va , Rab Rcacb .
first assuming it, we must begin by considering the graviton in
a flat background spacetime. Nevertheless, we will see from
the formalism of section III that (provided we use S2 to de-
II. THE GRAVITON ACTION scribe the graviton) energy-momentum self-coupling gener-
ates the Einstein-Hilbert action even when the background is
Contrary to the standard approach, we represent the gravita- not flat; gab need only satisfy the weaker condition
tional field as a perturbation hab of the inverse physical metric
gab from the background gab : Gab Rab 12 gab R = 0. (7)

gab = gab + hab. (3) While this equation expresses the generality of the analysis
that is to follow, it should be stressed that no knowledge of (7)
This expression is exact in that we have not neglected terms will be required to assemble the Einstein-Hilbert action order
O(h2 ); in contrast, the physical metric gab = gab hcd gca gdb + by order: a flat background will serve as a perfectly satis-
O(h2 ) . Following this convention, we use the contravariant
field hab , rather than hab , as the fundamental dynamical vari-
able of the action.4 In general we will write bars over tensors
order) infinitesimal gauge transformation hab = 2 (a b) . Fortunately, it
ab
is precisely for g {g ,gab } that the necessary energy-momentum self-
coupling is its most simple: hab tab (see III). These considerations provide
3 More precisely, S2 is the action for the graviton in a background spacetime no criteria for choosing the metric over its inverse as our expansion vari-
with metric in some small neighbourhood of the solutions of the vacuum able, and while this choice only trivially alters the perturbation theory at
field equations. We use the term vacuum to signify a region without matter; first-order (hab hab ) to second-order (the relevant order for S2 , tab , and
(2)
this does not necessarily imply the absence of spacetime curvature. Gab ) the two definitions of the h-field differ by a term of the form hac hbc .
4 Any metric theory Our choice of g = gab is preferable for this article because it simplifies
of gravity will have an ambiguity as to which variable
g {gab ,gab , ggab ,...} should be identified as the true gravitational the mathematics of the action and energy-momentum tensor. The reason
field. Such a distinction is of no physical consequence and is largely un- for this is explored in III E, and stems from the fact that any Lagrangian
necessary for a non-perturbative calculation; however for the present dis- for pure gravity must contain more factors of gab than gab in order that
cussion we are forced to single out a particular field variable for the ex- all the derivatives a be contracted; thus an expansion in g = gab will be
pansion g = g + h. Our aim is to connect gravity to the particle physics algebraically simpler. Indeed, this observation still holds when coupling
notion of a spin-2 field and elucidate a simple energy-momentum self- gravity to a scalar field or a 1-form Aa , and thus taking g = gab simplifies
coupling scheme that generates general relativity; to this end we are re- many of the calculations of the non-vacuum case also (see IV).
quired to pick g {gab ,gab } as it is only for these that h is a genuine 5 The only exception to this rule is the physical metric and its inverse, for
spin-2 field, i.e. a symmetric tensor (not a tensor density) with (lowest- which gab 6= gcd gac gdb , but rather gab gbc = ca .
 3

factory starting point.6 No matter which background we use, yields

however, it is absolutely crucial that we refrain from inserting
this particular metric (or even
R equation (7)) into the action,

thereby reducing S2 to 21 d4 x ghab Gabcd hcd . This is be- c hab
b
a hab
ah c c h c hcahb Rab
b c
cause we will need to be able to perform arbitrary variations habhcd Racdb . (10)
of gab , not just those consistent with Rabcd = 0 or Rab = 0, to
construct the energy-momentum tensor for hab . That said, it
will be instructive to temporarily ignore this advice so that we Thus we are forced to make a seemingly arbitrary choice: do
may relate S2 to the Fierz-Pauli action. we to covariantize (8) as written, or should we do so after per-
forming (9)? These two possibilities determine Lagrangians
which differ by hca hbc Rab + habhcd Racdb ; they lead to different
A. The Fierz-Pauli action (first-order) equations of motion if the background is curved,8
and determine different energy-momentum tensors even if the
background is flat.9 This last problem is discussed by Pad-
For a flat background, Habcd vanishes, and we can choose manabhan [1], and is one of his many non-trivial objections to
coordinates {x } such that g = and evaluate S2 as the conventional wisdom that general relativity is the unique
a functional of the components h . Integrating by parts energy-momentum self-coupled limit of the flat-space mass-
andR discarding surface terms, we find that S2 reduces to less spin-2 field.
1 4
2 d xLFP , where
A greater problem than this ambiguity, however, is that nei-
ther choice (nor an admixture) leads to general relativity after
2 h h 2 h h h h
1 1
LFP = coupling it to its own energy-momentum. As we shall see
+ h h (8) in section III, the contribution from hab Habcd hcd is necessary
to achieve this, and it is impossible to use the covariantizing
ambiguity to produce this tensor because it does not contain
is the Fierz-Pauli Lagrangian [1].7 Modulo surface terms and
hab hcd Racdb . Instead, the presence of Habcd represents a rather
an overall rescaling, LFP is the unique specially relativistic
different coupling ambiguity faced when moving from a flat
Lagrangian for a symmetric tensor field h that is invariant
background to a curved one. Typically we would invoke the
under the infinitesimal gauge transformation h = 2 ( )
Einstein equivalence principal to banish from the action terms
(see [1] for proof); hence it is the Lagrangian for the graviton
coupling matter fields and Ricci tensors; we would argue that,
(massless spin-2 field) in flat spacetime.
working in locally inertial coordinates about a point p, the La-
Starting from (8), we can covariantize LFP by making the
replacements g , and multiplying by g. grangian at p should have the same form as the Lagrangian in
flat spacetime. This amounts to a minimal coupling proce-
This process obviously generates a unique manifestly covari- dure: once we have covariantized a specially relativistic La-
ant Lagrangian density if gab is flat, as in this case the pro- grangian, the job of coupling the field to the gravity is com-
cedure is equivalent to representing the same Lagrangian in plete. However, while this rule may make sense to curve the
arbitrary coordinates. However, for the purposes of calculat- background spacetime of a spin-2 field that is just another
ing the energy-momentum tensor (via arbitrary variations of matter-field and has nothing to do with gravitation, it is far
gab ) it will be necessary to generalize LFP to arbitrary back- from clear that the principal should hold for the graviton, for
grounds, and for a curved metric the covariantization proce- which it was only ever a convenient fiction to think of as a
dure is ambiguous. To see this, observe that we can transmute tensor field propagating over a background geometry.
the third term of (8) by twice integrating by parts:
In summary, the Fierz-Pauli action is insufficient to deter-
mine S2 for an arbitrary background geometry; the principal of
h h h h . (9)
equivalence fails to give a unique solution, and cannot justify
all the contributions necessary for an energy-momentum self-
However this equivalence relies on the commutativity of par- coupling procedure consistent with general relativity. How-
tial derivatives, and does not occur for the covariant deriva- ever, it was never our aim to construct general relativity from
tives of a curved background; instead, integration by parts LFP , and we do not pretend to be able to derive a curved
spacetime theory of gravity from purely specially relativistic
concepts. S2 will serve as our starting point, and the only sig-
nificance we shall ascribe LFP is that of a special case.
6 Of course, once the self-coupling procedure is complete, and the Einstein-
Hilbert action has been assembled starting from the graviton on a flat back-
ground, we will be in a excellent position to justify (7), as this is precisely
the field equation (applied to the background) that we will have derived.
With hindsight, then, we can see there was nothing special about our flat- 8 The first-order field equation only describes the spacetime perturbations
space starting point: we may begin with any one solution to (7) and use of general relativity if the ambiguous term is covariantized to become
energy-momentum self-coupling to derive the action (and field equation) c hab
a h c ; see II B and Appendix B.
b
that defines all the others. 9 Note that all other terms of LFP are invariant under the operation that gen-
7 Here and elsewhere we use the customary shorthand h haa hab gab . erated (9) so do not introduce further ambiguity.
 4

B. Field equations define the energy-momentum tensor as a functional derivative

of the action with respect to the (background) metric:
Leaving the Fierz-Pauli action behind, we retrain our at- 1 S2
tention on S2 and begin the process of deriving its advertised tab , (14)
g gab
connection to general relativity. First, we shall calculate the
associated field equations. As usual, the equations of motion where hab (rather than hab or hab ) is to be held constant when
are derived from the condition that their solutions be station- taking this derivative, as this is the field we have taken to be
ary configurations of S2 with respect to variations in the dy- the fundamental dynamical variable.11
namical field hab . As we will have no cause to vary gab in As an aside, it is worth contrasting the variational definition
the derivation, we can enforce the background equations (7) (14) with Noethers (canonical) energy-momentum tensor:
immediately and discard Habcd . Next, observe that Gabcd is
self-conjugate: for any tensor fields Aab and Bab L
tcan h L , (15)
Z Z ( h )

d4 x gAab Gabcd Bcd = d4 x gBab Gabcd Acd , (11)
comprising the four conserved currents associated with the in-
variance of the Lagrangian L under rigid spacetime transla-
provided either Aab or Bab has compact support. Therefore, tions. The canonical tensor cannot be used in the present dis-
holding gab constant and performing a variation hab (a sym- cussion for a number of reasons. Firstly, it is not uniquely
metric tensor field with compact support) gives rise to a vari- determined by the action for hab : as it depends directly on

ation in the action the Lagrangian, we are free to alter tcan by adding a four-
Z
divergence to L , without changing either the dynamics of hab
1
S2 = d4 x g hab Gabcd hcd . (12) or S2 . Secondly, we require a symmetric tensor to act as the
source for the first-order field equation (13), but the canon-
ical tensor need not have this property.12 Lastly, Noethers
As Gabcd is already symmetric in its first two indices, we can
definition does not naturally generalize to curved spacetime
conclude that the equation of motion is
in such a way that tcan inherits a covariant conservation law
1 S2 [11]. None of these issues arise with tab , and in any case
= 1 Gabcd hcd = 0. (13) our aim has been to connect the coupling between matter and
g hab
gravity found in general relativity with a perturbative coupling
The centrally important feature of this equation is that of gravity to itself; it is the Hilbert energy-momentum tensor
(1) of matter, not the canonical tensor, that appears in the full
Gabcd hcd = Gab , the linear approximation to the Einstein ten- Einstein field equations as the gravitational source. For these
sor under the inverse metric expansion (3). This is particu- reasons we discard the canonical tensor and henceforth refer
larly easy to verify for the special case of a flat background to tab , following Hilberts prescription (14), as the energy-
in Lorentzian coordinates, but is shown to hold more gener- momentum tensor of hab .
ally for vacuum backgrounds in Appendix B. Thus S2 pre- To begin the calculation of tab , we divide the action into two
scribes the correct first-order equation of motion for the gravi- pieces S2 = S2G + S2H :
ton. In the next section we show that by adding the energy-
Z
momentum tensor tab of hab (determined by S2 ) to the right 1
S2G d4 x ghab Gabcd hcd , (16)
hand side of (13) we successfully generate the Einstein field 2
equations correct to second-order.10 Z
1
S2H d4 x ghab Habcd hcd . (17)
2
C. Energy-momentum tensor It will be convenient to perform the functional derivative (14)
on these two components separately. Focusing first on S2G ,
we integrate by parts13 so as to remove the second derivatives
We will now calculate the energy-momentum tensor of the
graviton and relate it to the second-order contribution to the
Einstein field equations. We follow Hilberts prescription and
11 2 to indicate
In later sections, the tensor written here as tab will be notated tab
that it is the energy-momentum contribution from the second-order action
S2 only. Here we need not make this distinction.
10 12 It is true that the canonical tensor can be made symmetric by adding to
Of course, the resulting field equation will no longer be a stationary config-
uration of the action S2 . In order that this self-coupled equation of motion it an identically conserved correction [ ] , a function of hab that

can be derived from the principle of stationary action it will be necessary cancels the antisymmetric part of tcan . However, if we allow this sort of
to introduce a third-order correction to the action S3 . Naturally, S3 will ad hoc adjustment of the energy-momentum tensor, we only exacerbate the
alter the energy-momentum tensor of hab by a term O(h3 ); however, seem- problem of non-uniqueness.
ingly by miracle, this will be precisely the third-order part of the Einstein 13 More precisely, one adds to the integrand a divergence of the form
a ) = g
field equations. This process continues indefinitely and is explained sys- a ( g[hh] a [hh]
a that alters S2 only by a function of the
tematically in III. For the moment we content ourselves with exploring fields on the boundary (or at infinity) and thus may be neglected for the
the theory to second-order only. purposes of functional variation.
 5

from the integrand: Meanwhile, S2H varies by

Z
1 ab e f
S2G = d4 x gc h d h Kab e f ,
c d
(18) 1
Z

2 S2H = d4 x g Rab
2    
for which we have introduced the abbreviation
12 gab 21 h2 + hcd hcd habh , (22)
1  cd
Kab ce f d g ga(e g f )b gcd gab ge f 2(ec g f )(a b)
d
2 
where we have used the background equation (7) (after the
+ (ec fd) gab + (a b) ge f
d c
(19)
variation) to remove the terms proportional to Rab ; these
= Kba ce f d = Kab c f e d = Ke f dab c . would only be significant if we intended to perform further
variations in the metric. Now, because
An infinitesimal variation in the inverse background metric
gab , vanishing on the boundary of the integral, induces a
Rab = 2 [cCc
b]a
variation in the action  
Z
" = 21 grs gap gqb + 12 (a
r s
b) g pq pr bs gaq s g pq ,
r
1 Kab ce f d
S2G = d x g g c h d h
4 pq ab ef when we (twice) integrate by parts to alleviate gab of its
2 g pq covariant derivatives, we generate a second-order differential
! # operator
1 ab (e
c d
g pq Kab e f + 4c h C sd h Kab e f ,
f )s c d
2
2 + 1 g pq
R pqab 21 ga(p gq)b 2
(a
b)
(a gb)(p
q) , (23)
where


Cabc b gcd +

1 ad c gbd d gbc with the property
2 g
 
= 21 2 pa (b
r
gc)q gar gbp gqc r g pq (20) Z Z

d4 x g Rab Aab = d4 x g g pq R pqabAab (24)
is the connection that arises from the variation of the covari-
ant derivative: g+ g = + C. We can move the covari-
ant derivatives off g pq in the connection term using integra- for all Aab . Therefore, we can conclude from (22) that
Rtion by parts, and arrive at an equation of the form S2G =
d4 x g pq [. . .] pq ; the tensor density in square brackets is then
S2H 1    
the functional derivative we seek:
= R pqab 12 gab 12 h2 + hcd hcd habh .
g g pq 2
S2G 1 ab e f
= c h d h (25)
g g pq 2
!
Kab ce f d 1
g pq Kab ce f d
g pq 2 Finally, we have only to combine equations (21) and (25),
  expand out all the products and derivatives, and assemble the
r
c hab K c
ab (p| f |q) h
rf outcome into a formula for tab as a function of
c hab . This is a
 straightforward but arduous calculation, and as such we chose
c r f f
+ Kab f (p
hq) Kab cr f (p hq) . (21) to complete it with a computer algebra package. The result is


t pq = 1
h
4 g pq
a b hab + 2hab a b h 2hab 2 hab h 2h 1
2 a h h 2 c hab h + c ha b h
a 5 c ab b ac

+ 2
a h
b hab + 1 h (p
q) h 1 h pq
2 h + 1 h 2 h a 1 hab
2 h pq + ha(p a b h pq + 1 h pq
a
b hab
4 2 4 q) 2 2
a
ha(p q) ha + 1 hab
b
b 2
q) hab 1 h
(p
2
a
(p h + 1
q) 4 a h h pq + 2 b hap h q 2 a h pq b h
a 1 b a 1 ab
b
+ 43
p hab b ha
q hab
(p q) ha 2 b h(p hq) + 2 b h p a h q .
b 1 1 a b
(26)
 6

It is possible to render this formula rather more manageable by working in a gauge with a hab = 0, h = 0:
 
t pq = g pq 41
ch b
b hac 5 c hab 1 hab
c hab 2 h a 1 hab
2 hab + ha(p b h pq hbc Rabc(p h a
a
a 8 2 q) 2 q)

+ 12 hab
(p
2 b hap h q + 4 p hab q h b h (p q) ha + 2 b h p a h q ,
q) hab + 1
b a 3 ab a b 1 a b
(27)

but we will not need this partially gauge-fixed result for this itself an exact solution of the vacuum field equations:
present article.14
(2) gab = gab + hab, (30)
Our task now is to compare tab with Gab and demonstrate
S[g]
that the energy-momentum self-coupling of hab (determined 0 = , (31)
by S2 ) is consistent with general relativity. Details of the cal- gab
(2)
culation of Gab can be found in Appendix B; the conclusion where , a dimensionless expansion parameter, is constant
is over spacetime.
Following (30), the action of the exact theory S[g] becomes
(2)
Gab = tab + O(h3 ), (28) a -dependent functional of gab and hab , which can be Taylor
expanded thusly:
and thus, to second-order, the vacuum Einstein field equations
are S[g] = S[g + h] = n Sn[g, h], (32)
n=0
Gabcd hcd = tab (29)
where Sn is the nth partial action given by
as advertised. 1 n

Sn [g, h] = S[g + h] =0 . (33)
As a corollary of (29), we can confirm Padmanabhans n!
observation that general relativity cannot be derived from
energy-momentum self-coupling the Fierz-Pauli Lagrangian. The derivative acts on each instance of hab in the inte-
Only once the contribution from Habcd is included will Ein- grand of S[g + h] by Leibnizs law, removing the factor of .
steins gravity result from an energy-momentum self-coupled The bare hab left behind may still be covered by spacetime
graviton. This realisation casts doubt on Mannheims recent derivatives a , but these can be moved onto the remainder of
treatment of gravitational energy-momentum [12], in which a the integrand by integration by parts. This operation generates
tensor is constructed by applying (14) to a covariantized Fierz- the usual functional derivative:
Z
Pauli Lagrangian, rather than S2 .
S[g + h] = d4 xhab (x) S[g + h]. (34)
gab (x)

III. PERTURBATIVE GRAVITY In truth, the left hand sideR of this equation differs from the
right by the surface term d4 xa J a created when integrating
by parts. As this is only a functional of the fields on the bound-
Here we develop the formalism to uncover the root cause of ary (or as x if the integral of S runs over the entire man-
the second-order energy-momentum self-coupling (29), and ifold) it will not contribute to equations of motion or energy-
reveal how the process continues to arbitrary order. The vast momentum tensors, the calculation of which are dependent
majority of this section applies to any metric theory of pure only on variations of the field that vanish on the boundary (or
gravity15 and can be generalized to include interactions with have compact support). Hence these surface terms may be
matter (see IV). Only in section III E will we commit to gen- neglected for our present purposes.
eral relativity, fix our action S = SEH , the Einstein Hilbert ac- It follows from the repeated application of (34) that
tion, and derive the formula (4) for S2 . Z n
We shall concern ourselves with an expansion of the inverse 4 ab
S[g + h] =
n
d xh S[g + h], (35)
metric gab about a non-dynamical background gab , which is gab
and thus the partial actions (33) are given by
Z n
1 4 ab
14 Gauge transformations are covered in III C; we note here only that be- Sn [g, h] = d xh S[g]. (36)
cause tab is not invariant under the infinitesimal gauge transformation
n! gab
hab = 2 (a b) , only the first formula (26) can be used in all gauges. Al-
An important consequence of this relation is that, using S2
though gauge invariance would be a highly desirable property if we in-
tended to argue that tab was a physically meaningful tensor in full general as our starting point, we can generate the entire set of partial
relativity, it is an impossible request to make of the tensor we seek, which actions {Sn : n 3} by calculating
(2)
should be proportional to the gauge dependent tensor Gab . Z n2
15 We require only that the dynamics are determined by an action that is a 2 4 ab
Sn [g, h] = d xh S2 [g, h], (37)
coordinate-independent integral of the metric and its derivatives. n! gab
 7

which is possible provided S2 is known in a neighbourhood where, because S2 is second-order in hab , Gabcd will be a linear
of whichever particular background (a solution of (31)) we differential operator dependent only on gab .17 The equation of
are interested in. Note that the first two partial actions do not motion (40) now takes the form
contribute to the dynamics of hab : S0 = S[g] is manifestly in-
dependent of hab , and S1 vanishes once the background equa- N+1 n
tion (31) has been enforced. We conclude, therefore, that S2
Gabcd hcd = Sn [g, h],
g hab n=3
(43)
contains all the information necessary to reconstruct the dy-
namical part of the action where it should be taken as given that terms O( N+1 ) have
been neglected. This is the N th -order approximation to the
Sdyn [g, h] n Sn[g, h], (38) equation of motion for hab that is consistent with the dynam-
n=2 ics of gab prescribed by the action S. The first-order contri-
which itself contains all the dynamical information of the full bution has been separated from the sum so as to evoke the
action S. This is absolutely key to the calculations of sec- picture of a wave equation Gabcd hcd = 0 with a source. In
tion II, in which we saw the first consequence of this recon- the next section we will see that the source term on the right
struction process, the recovery of the second-order equation of of (43) is indeed the energy-momentum tensor of the field hab ,
motion from an action that one would expect to encode only neglecting terms O( N+1 ).
first-order dynamics.
B. Energy-momentum tensor
A. Field Equations
First we shall demonstrate that the dynamical part of the
In general, we could let be a free parameter and, on de- action (38) can be generated from S2 by a simple energy-
manding S[g]/ gab = 0 for fixed gab , derive a -dependent momentum self-coupling procedure. Observe that, as a con-
equation of motion E [g, h] = 0 for our dynamical field hab . sequence of (36), we have
Any hab that solved this equation would correspond to a met- Z
ric gab = gab + hab that solved the field equations exactly.16 1 Sn1 [g, h]
Sn [g, h] = d4 xhab . (44)
However, if we are interested in approximating small varia- n gab
tions of the metric (i.e. the limit hab 0) we can choose
some order N to which we want the equation of motion to Defining the nth partial energy-momentum tensor tab n by ap-
th
plying Hilberts prescription to the n partial action,
hold:
S[g] 1 Sn [g, h]
= O( N+1 ). (39) n
tab , (45)
gab g gab
This is equivalent to
we conclude that
1 Sdyn [g, h]
N+1 Z
= O( N+1 ), (40) 1
hab Sn [g, h] = d4 x ghabtab
n1
. (46)
n
N+1
where Sdyn is defined by discarding from Sdyn those terms This makes manifest the energy-momentum self-coupling
that can be neglected in (39): procedure that allows us to generate the dynamical part of the
N+1 action (38) to arbitrary order, given only S2 . The nth partial
N+1
Sdyn [g, h] n Sn[g, h]. (41) action is nothing more than the integral of the contraction of
n=2 hab with the energy-momentum tensor of the previous partial
action (divided by n). The dynamical part of the action is
We shall adopt this N th -order approximation picture for the
therefore given by
development of our formalism, as we can always write N =
if we wish to discuss the exact theory. N+1
Sdyn [g, h] = 2 S2 [g, h]
For the sake of continuity with the previous section, we in-
Z N n+1t n
troduce the notation
 d4 x ghab ab
. (47)
S2 [g, h]  n=2 n + 1
 1 gGabcd hcd , (42)
h ab
S[g]/ gab =0 Note that, for the particular case of general relativity (S =
SEH ), the background equation (7) also sets S0 = 0, thus
Sdyn = SEH (modulo surface terms) and the energy-momentum
16
It is advisable to set = 1 before attempting to solve E [g,h] = 0, as this
constant can always be absorbed into the magnitude of hab . Although this
refinement was convenient for II, here we shall keep as it provides a
simple method for tracking the powers of hab in expressions and is useful 17 The operator Gabcd defined here coincides with the definition in (5) once
as a variable for differentiation. S = SEH has been fixed. This is shown in III E by deriving S2 .
 8

self-coupling procedure recovers the entire action of the full theory, and the right hand side is the energy-momentum tensor
theory, not just the dynamical part. prescribed by the action SdynN . This energy-momentum ten-
n in (47), it is
Because of factors of n + 1 dividing each tab sor is, to some extent, incomplete: it does not include the
not the case that in the action h couples directly to its (N th -
ab O( N+1 ) contribution from the highest-order partial action
order) total energy-momentum tensor, given by SN+1 . This contribution could be calculated, if so desired,
and added by hand to the field equations (53) so that the right
1 Sdyn hand side read Tab
N N N+1
, but this equation would no longer be
N
Tab = ntab
n
. (48)
g g ab
n=2
N+1
a stationary configuration of the action Sdyn . To remedy this,
we could introduce a correction to the action N+2 SN+2 that
Instead, the numerical denominators account for the n + 1 fac- would generate the extra term in the equation of motion; the
tors of hab in habtab
n , and ensure that the equations of motion
appropriate functional is given by (46) and couples hab to the
do indeed have Tab N as the source. To prove this, note that for N+1
highest-order partial energy-momentum tensor tab . But now
any symmetric field l ab (vanishing on the boundary, or with N+1
once again the energy-momentum tensor Tab is incomplete,
compact support) we have and we can apply this same line of reasoning anew. So long
N vanishes identically, this process
as there is no N for which tab
Z Z
Sn [g, h] l ab
can continue indefinitely, and as N the exact field equa-
d4 xl ab = d4 x n S[g + h] =0
hab n! hab tions are recovered, along with the action Sdyn = S S0 S1.



= 1
n! n S[g + (h + l)] =0 =0 All that remains is to connect our formalism to the specific

results of the previous section. For the sake of completeness,
= 1
n! n S[g + (h + l)] = =0 however, we shall first discuss the gauge symmetries of the

N+1
= 1
n! n ( S[g + h + l]) = =0 , theory, and deduce the conservation law for Tab .

where = . Thus,
C. Gauge transformations
Z 
Sn [g, h]
d4 xl ab = 1
n S[g + h + l]
hab n!

Because the action S[g] is a coordinate-system independent
integral, any diffeomorphism : M M gives rise to a
+ nn1 S[g + h + l]
= =0 gauge transformation of the theory through the action of ,


= 1
n1
S[g + h + l] the map comprising the pullback of on covector indices and
(n1)! = =0
the pushforward of 1 on vector indices:
= ( Sn1 [g + l, h]) =0
Z
Sn1 [g, h] S[ g] = S[g]. (54)
= d4 xl ab . (49)
gab
Taylor expanding both sides about gab and applying the back-
Hence we have the following important result: ground equation reveals the gauge invariance of the dynamical
part of the action:
Sn [g, h] Sn1 [g, h]
= . (50) N+1
[g, h ] = Sdyn
N+1
[g, h],
hab gab Sdyn (55)

Or, using definition (45), where

Sn [g, h] n1 hab gab gab. (56)

= gtab . (51)
h ab
In the context of an N th -order approximation, we must insist
Therefore the equation of motion (43) takes on the form that = 1 + O( ), otherwise these transformations will map
the small metric fluctuations hab onto fluctuations compa-
N+1
rable in magnitude to gab . We can write a general diffeo-
Gabcd hcd = 1 ntabn1, (52)
morphism of this form as = e L , where L is the Lie
n=3
derivative along a vector field a = O(1). The gauge transfor-
or, recalling (48), mations of the theory are hence given by
Gabcd hcd = Tab
N
. (53) hab hab = hab + hab ,
N ( L )n N1 ( L )n

We have derived the relation we sought, demonstrating that hab 1 gab + hab , (57)
any metric theory of pure gravity can be formulated as a first- n=1 n! n=1 n!
order wave equation with its own energy-momentum tensor
as a source. For every N 1, we can derive the equation of where we have discarded all terms O( N ), as these will only
motion (53) by applying the variational principle to the action contribute terms O( N+1 ) to the equation of motion, and
N+1
Sdyn ; the left hand side is the wave equation for the linearised terms O( N+2 ) to Sdyn
N+1
. If we wish we can let a = a , an
 9

infinitesimal vector field, and derive the infinitesimal gauge infinitesimal vector field a with compact support. Then (60)
transformation becomes
(
Z
" N+1 #
L gab + hab N 2, Sdyn Sdyn
N+1
h =ab
(58) 0= d x 4 ab
L g + ab
L h . (61)
2 (a b) N = 1. gab hab

Because these gauge transformations (infinitesimal or other- Clearly the second term vanishes (to O( N+1 )) if hab solves
N+1
wise) are symmetries of Sdyn , they map solutions of the equa- the equation of motion (53), and thus
tion of motion (53) to other solutions. We can therefore use
Z Sdyn
N+1
the equation of motion to deduce the transformation law for
N: 0 = d4 x a b + O( N+2 )

Tab gab
Z

= d4 x g b
a T N+1 + O( N+2 ). (62)
Tab
N N
Tab [g, h ] Tab
N
[g, h] = Gabcd hcd . (59) ab

As this equation holds for any a it follows that
This verifies the earlier remark that the energy-momentum
tensor is gauge dependent, except in the trivial case N = 1,
a T N+1 = 0 (63)
N = 0 by definition. It may come as a surprise ab
for which Tab
that the energy-momentum tensor does not inherit the gauge is valid up to and including O( N+1 ). Because this relation
invariance of the action from which it was derived. It should holds whenever hab solves its equation of motion, and because
N+1
be stressed, however, that Sdyn is not identically gauge in- gauge transformations map solutions to solutions, the conser-
variant: the relation (55) is only true when the background vation law is gauge invariant.
equation is obeyed. For general gab , the diffeomorphism in- It is important to recognize that (63) applies to the (N +
variance of S[g] only furnishes the gauge transformation law 1)th -order energy-momentum tensor: this is the highest-order
Sdyn
N+1
= S1 , the right-hand side of which has a non- approximation to the energy-momentum tensor that can be
N+1
vanishing energy-momentum tensor responsible for the varia- constructed from our truncated action Sdyn , and is a bet-
N . Equivalently, the gauge dependence of T N can be
tion in Tab N
ab ter approximation than the tensor Tab which features in the
seen to result from the non-commutativity of gauge transfor- equations of motion appropriate to this order. Of course, the
mations and the functional derivative / gab used to define conservation law for TabN follows from (63) by discarding the
N [13]; these operations would only commute if the gauge
Tab highest-order term, and ensures the consistency of the equa-
N+1
invariance of Sdyn extended to a neighbourhood of the solu- tion of motion (53) with the identity a Gabcd hcd = 0, which
tions of the background equation, rather than being confined ab
holds for all h once the background equation has been en-
to the solutions themselves. forced.

D. Conservation law E. Constructing the graviton action

N+1 It is now time to close the circle of our discussion and con-
It should be expected that Sdyn [g, h] inherits the diffeomor-
phism invariance of S[g], and that this symmetry endows the nect the abstract formalism to our earlier calculation. We shall
energy-momentum tensor with a covariant conservation law derive here the graviton action S2 , the ansatz of section II, by
with respect to the background metric. The derivation pro- applying the perturbative formalism to the particular case
ceeds in close analogy to the proof of a Tab matter = 0 from gen- Z

1
eral relativity. S[g] = d4 x gR SEH [g], (64)
We again appeal to the diffeomorphism invariance of the
action (54) but this time expand S[g] about gab (a solution of the Einstein-Hilbert action. To proceed, we will use equation
the background equation) and S[ g] about gab (which will (36) to derive S1 , and then S2 , by successive functional deriva-
also be a solution). The result, tives / gab acting on SEH [g]. The first derivative generates
Z
Sdyn [ g, h] = Sdyn
N+1 N+1
[g, h], (60) 1
S1 [g, h] = d4 x gGab hab , (65)

affirms that is diffeomorphism invariant.18 Now let be
N+1
Sdyn
an infinitesimal diffeomorphism: = 1 + L for an arbitrary which of course vanishes for all hab when gab solves the back-
ground equation Gab = 0. A second variation in gab gives rise
to
Z
1 

18 Note that diffeomorphism invariance is equivalent to being independent S1 = d4 x g Rab hab 12 hgab

of coordinate system, and is a distinct property from gauge invariance as

defined in III C. + gcd 21 hcd R hRcd gcd Gab hab .
 10

Replacing Rab gcd Rcdab in accordance with (24), we de- IV. MATTER
termine S1 / gab and assemble
Z To avoid over-complicating our discussion, we have so far
1 S1
S2 = d4 xhcd cd focused exclusively on pure gravity. Here we will go some
2 g way to remedy this simplification, and generalize the formal-
Z
1 

ism of the previous section to include the perturbations of mat-
= d4 x g hcd Rcdab hab 12 hgab
2 ter fields, and the effects of non-vacuum backgrounds.


+ 21 hcd hcd R hRcd gcd Gab hab In the most general case, let the action S be a functional
1
Z
of gab and a generic matter field A , where A will serve as a
= d4 x ghab (Gabcd + Habcd )hcd . (66) placeholder for any number of internal or spacetime indices.
2
We then expand S about a background (gab , A ) as follows:
In the last line we referred to the definitions (5) and (6), and
made use of the identity gab = gab + hab, (68)
A = A + A, (69)
Rabe f (ce d 21 ge f gcd ) Gabcd .
f
(67)
This completes the derivation of the graviton action (4) and S[g, ] = n Sn[g, h, , ], (70)
n=0
confirms that it can be used as the starting point of an energy-
momentum self-coupling procedure (46) that generates the where gab and A satisfy the background equations
Einstein field equations and the Einstein-Hilbert action (mod-
ulo surface terms) to arbitrary order. S[g, ] S[g, ]
The preceding calculation helps to reveal the advantage = 0, = 0. (71)
gab A
of using hab , a perturbation in the inverse metric, as our
fundamental degree of freedom. Had we instead taken the As before, each partial action can be calculated from the par-
usual approach, expanding gab = gab + hab and taking hab tial action at the previous order; with matter included, the ap-
as fundamental, the perturbative formalism would have un- propriate recurrence relation is
folded identically but for the placement of indices. However, Z
the calculation of S2 from SEH would have differed dramat- 1  
Sn = d4 x g habtab
n1
+ A jAn1 , (72)
ically. The Lagrangian of S1 would instead be proportional n
to Gab hab , and because the Ricci tensor is naturally covari-
ant, the variation of Gab = Rcd gca gdb 21 Rcd gcd gab under gab where
would have been complicated by the extra two factors of gab n 1 Sn 1 Sn
on the first term, compared to the relevant tensor in our ap- tab , jAn . (73)
g gab g A
proach: Gab = Rab 12 Rcd gcd gab . This trend continues at
every order; the hab convention leads to a greater prolifer- There are two aspects of this coupling scheme that differ from
ation of terms in each partial energy-momentum tensor be- pure gravity. The first is immediately apparent: the habtab
cause the Lagrangian of Sn has the form ( a )2 (hab )n so must term has been joined by an analogous coupling between mat-
be contracted with a further n + 1 factors of gab to render it a ter fluctuations A and its source current jA . The second
scalar.19 Each of these metric factors generates a term in the difference is hidden within the definitions of tab and jA ; be-
partial energy-momentum tensor, and thus act as compound cause the {Sn } now represent the partial actions for gravity
interest for the process of energy-momentum self-coupling. In and matter together, habtab and A jA are no longer just self-
comparison, our convention leads to Lagrangians of the form couplings, and will in general contain terms coupling hab to
( a )2 (hab )n , which only need only n 1 additional factors of A . In particular, tab
n should now be interpreted as the (nth -
gab .20 Clearly the inefficiency of the hab approach stems from order) energy-momentum tensor due to all the fields: hab , A ,
the natural covariance of derivative operators (a or a ) and
and the background matter A .
curvature tensors; the advantages of the contravariant expan- Proceeding as before, we can now demand that the dynam-
sion gab = gab + hab are therefore not peculiar to the Einstein ical fields hab and A solve the field equations of the action
n=2 Sn , and generate approximate solutions of the
= N+1
Hilbert action, and are expected to be even more distinguished N+1
Sdyn n
in higher derivative theories of gravity.
exact field equations (prescribed by S) accurate to O( N ). In-
stead of using the definition (42) for Gabcd , we write the gen-
eral form of S2 , modulo surface terms, as
Z
19 There are of course the instances of gab c gde in each a , but these occur 1 
equally in either convention. S2 = d4 x g hab Gabcd hcd /
20 2
This does not mean that all terms in such a Lagrangian will contain only 
n 1 additional factors of gab ; there will often be cases in which gab is 2habIabA A + AWAB B , (74)
contracted with ( a )2 and thus n + 1 factors of the metric (and its inverse)
will be present. These cases only represent a small proportion of all possi-
ble terms, particularly as n becomes large, and are no worse than the terms once the background equations (71) have been enforced. In
afforded by the hab convention. the above equation, Gabcd , IabA, and WAB are linear operators
 11

that depend only on background fields, Gabcd and WAB are self- This assumption will mean that the perturbative expansion of
conjugate, in the sense given by (11), and IabA is conjugate to S can be described by an energy-momentum coupling proce-

IAab : dure only. To see this explicitly, we expand the action about a
Z Z background (gab , 0):

d4 x gAab IabABA = d4 x gBA IAab Aab , (75)

S[g + h, ] = n Sgn [g, h] + Sn[g, h, ] , (82)
for all Aab or Bab ,
provided one has compact support. These n=0

definitions lead to equations of motion, accurate to O( N ), as

where each gravitational partial action
follows:
1 n

Gabcd hcd = Tab
N
+ IabA A , (76) Sgn [g, h] = Sg [g + h] =0
n! 
WAB B = JAN + I hab , (77) Z n
1 4 ab
Aab
= d xh Sg [g], (83)
where n! gab
N N much as before, and the matter partial actions
N
Tab ntabn , JAN n jAn . (78)
n=2 n=2

1 n
Sn [g, h, ] = S [g + h, ] =0
Although this formalism is quite general, it is probably too n!
general to be usefully employed. Indeed, the complications 1 n 2


= S [g + h, ] =0
involved in describing matter as a background field and a dy- n!
namical perturbation generally serve to obscure the physical 1

= n2 S [g + h, ] =0
interpretation of the mathematics. An interesting example of (n 2)!
this occurs when one tries to rederive
a T N+1 = 0 by apply- Z n2
4 ab
ab 1
ing the argument of section III D. The result that now follows = d xh S [g, ].
is (n 2)! gab
Z (84)
a T N+1 = 1
d4 x gJAN+1 L A , (79)
ab
2 g b Defining the partial energy momentum tensors for hab and A
as
the physical interpretation of which is far from clear. Rather
than continue with this formulation in its full generality, it will gn 1 Sgn n 1 Sn
therefore be more instructive to examine two special cases. tab , tab , (85)
g gab g gab
First, we set A = 0 and consider small matter fields A
interacting with hab . Second, by setting A = 0 we can study respectively, we see that the partial actions are coupled as
the effect of a background matter field A on the propagation
of the graviton. In principal, one could reach these special Z
!
t n1
gn1
4 ab tab
cases starting from the formalism we have just described, but Sn [g, h] = d x gh + ab . (86)
it will be simpler and more illuminating to build them up from n n2
scratch.
These partial actions lead to the N th -order equations of motion
A. Matter perturbations N  
Gabcd hcd = Tab n n
N gn
= tab + tab (87)
In a region where the matter fields are small enough that n=2

their effects on spacetime curvature can be described by small N
n
perturbations hab in the inverse metric, we can model the dy- WAB B
=
(n 1) g
n=2
namics by taking A = 0, and describe the matter field using Z 
A alone. As it is often the case for gravitational theories, n
d4 x ghabtab . (88)
let us suppose that the action S is the sum of a gravitational A
action Sg and a matter action S :
The first equation confirms that the energy-momentum tensors
S[g, ] = Sg [g] + S[g, ]. (80) of A and hab combine as the source for the graviton. The
second equation describes how the coupling between hab and

Moreover, for the sake of simplicity, we take A to be a free tab acts as a source for A . Note that even when the matter
field: field is not free, because S never contains terms linear in the
matter fields, IabA must be at least linear in A , so we will
S [g, ] = 2 S [g, ] gab , A . (81) always have IabA = 0 when A = 0.
 12

B. Non-vacuum background is the energy-momentum tensor of the background matter.

The second derivative yields the graviton action:
For a non-vacuum spacetime, we expect to be able to ap- Z
1 S1
proximate (at least to first-order) the behaviour of a gravita- S2 = d4 xhab ab
tional perturbation by ignoring the perturbations in the mat- 2 g
Z
ter field that it might induce. Alternatively, we may have in 1 h
= d4 x g hab Gabcd hcd
mind a particular non-vacuum solution of the field equations 2
(gab , A ) and wish to find nearby solutions (approximate or 

 2 L
exact) with precisely the same matter content. For these two Gab Tab hab h + 2 habhcd ab cd
g g
scenarios, we can set A = 0 and investigate the effect that the
1 i
background A has on the dynamics of hab . + R + 2 L 2 hab h 4 h
ab 1 2
. (93)
Considerations of this nature highlight an interesting fea-
ture of our prior discussion of the graviton action. In sec- This is the action we sought: the generalization of equation
tion II we saw the importance of a contribution to the action (4) to a non-vacuum background.
hab Habcd hab that vanished in the vacuum; the obvious ques- If we are only interested in the linear theory, and have no
tion to ask is whether a similar term exists in the non-vacuum wish to calculate the energy-momentum tensor, then we are
case, and whether or not it will vanish on the non-vacuum free to enforce the background equation
background equations. To answer these questions we will de-

rive the graviton action for a non-vacuum background, which Gab = Tab , (94)
will also include the cosmological constant as a special case.
Let us restrict our attention to general relativity in the pres- in the graviton action. In sharp contrast to the vacuum case,
ence of a matter field: however, the background equation does not reduce S2 to
1 R 4
2 d x ghab Gabcd hcd , or indeed any other covariantiza-
S[g, ] = SEH [g] + S[g, ], (89) tion of the massless spin-2 Fierz-Pauli action. Instead, it ap-
Z
pears as though the background matter has endowed the gravi-
S [g, ] 2 d4 x gL (gab , A , a A ). (90) ton with mass:
Z
The factor of two in the definition of the matter Lagrangian 1  
S2 = d4 x g hab Gabcd hcd + , (95)
L compensates for our slightly unusual normalization of 2
SEH .21 It should be noted that we have assumed that L does
where the mass-term is given by
not depend on derivatives of the metric. This is the case for the
Lagrangians of all the fields of the standard model except the  
spin- 21 fermion, which in any case should be coupled to grav- 21 M hab hab 21 h2 + Nabcd hab hcd , (96)
ity using the vierbein formalism, e.g. [14]; such an approach
is beyond the scope of this article. The results of this section with
can be generalized to allow Lm to depend on c gab without  
L 2 L
any great difficulty, but this is an added algebraic complica- M 2 L gab ab , Nabcd 2 ab cd . (97)
tion that seems to add little insight to our investigation. g g g
We proceed by expanding the action about a background
(gab , A ) just as in (68) and (69), but now, as A = 0, the cou- We refer to as a mass-term because it is quadratic in hab ,
pling scheme (72) reverts to the familiar energy-momentum free from derivatives, and has been added to the kinetic term
coupling of section III. Following precisely the same method hab Gabcd hcd in the Lagrangian. However, as we will see for
as section III E, we can compute S2 by two successive func- the specific case of the cosmological constant, does not by
tional derivatives (with respect to gab ) applied to S[g, ]. The itself determine whether the graviton is actually massive, i.e.
first derivative yields whether it propagates subluminally; the curvature of the back-
ground will play an equally important role in the field equa-
Z
1 
 tions. In particular, while it is tempting to identify a mass m
S1 = d4 x g Gab Tab hab , (91) for the graviton according to m2 = M (at least when Nabcd = 0)

we will soon see that the background matter often sets M < 0,
where so this idea is essentially untenable.
1 S [g, ] L To explore these issues, it will be instructive to calculate

Tab = = 2 ab + gabL (92) for a few simple examples. First, consider a scalar field
g g ab g background with Lagrangian

L = 12 gab a b V (); (98)

21 All our actions are twice as large as the usual definition. This normalization the mass-term is
has no effect on the classical equations of motion, but has allowed us to
define the energy-momentum tensor without a factor of two, simplifying
 
the algebra of II&III. = V () hab hab 21 h2 . (99)
 13

To ensure that the scalar field has positive energy density, we in a field equation that is identical in form to the first-order
must insist that V () 0; hence M 0 as previously warned. vacuum field equation (13) in this gauge. Of course, this does
Equation (99) can also be used to find the corresponding mass- not indicate that the cosmological constant has no effect on
term for a cosmological constant. In this case the Lagrangian the propagation of hab , only that these effects are limited to
is L = / , which we can reach from L by setting the constraints imposed on the background geometry by the
a = 0 and V = / . Clearly this gives background equation Rab = gab . For this reason, it does not
  seem particularly natural to interpret 2Rdabc hdc as endowing
= hab hab 12 h2 , (100) the graviton with a mass; equation (102) can instead be under-
stood as a (partially gauge-fixed) massless spin-2 field equa-
which similarly suffers from M < 0 if the cosmological con- tion that has been generalised to cosmological backgrounds.
stant is positive. Quite aside from this, there is also the technical issue of in-
At this point, the reader may be suspicious that the formu- terpreting the four-index tensor Rabcd as a mass: only if this
lae for and (with M < 0 and Nabcd = 0) signify that tensor can be defined in terms of a single scalar variable (and
hab is a tachyon in the presence of a scalar field background the background metric) could the argument be made that this
or a cosmological constant. Indeed, if the background were single variable described the gravitons mass. For a non-zero
flat and M constant over spacetime, we could derive the field cosmological constant, the only background with this prop-
equations from (95), observe that their divergence enforces erty is de Sitter space: Rdabc = 3 (gdb gac gdc gab ), thus the
the de Donder gauge condition gauge-fixed field equation (102) becomes
 
h 21 h = 0, 2 2 hab = 0.
(103)
3
and, substituting this back into the equations of motion, con-
If we were so inclined, we might interpret this as a field equa-
clude that the dynamics of the graviton were described by
tion for a graviton with m2 = 2/3, and note that this rela-

tion has the correct sign for positive , unlike the formula
2 M h = 0.
m2 = 2 suggested by our preliminary inspection of . In
truth, however, further investigation is needed before we can
This argument appears to justify the relation m2 = M for the
either adopt or discard this interpretation. This is not only
gravitons mass, and motivate the conclusion that M < 0 be-
because (102) (of which (103) is a special case) can be un-
trays tachyonic behaviour. It is important to realise, however,
derstood as a generalisation of a massless field equation to
that the field equation above is of little relevance to the ac-
cosmological backgrounds, but also because of the subtleties
tual physical system we were discussing. In reality, M will
involved in interpreting the wave operator 2 in curved space,
not be constant, and the presence of background matter will
and issues of whether or not to use a conformal coupling.
inevitably preclude background flatness. To understand how
Clearly, more work must be done to ascertain the physical
this last consideration alters the dynamics of the graviton, we
ramifications of , and the mass-term in general, be-
shall briefly examine the field equation for hab in the presence
fore we can understand the degree to which its effects can be
of a cosmological constant. First, we substitute (100) into (95)
thought of as giving mass to the graviton.
and derive the field equation
Although massive gravitons and the cosmological constant

were historically viewed as entirely separate concepts, re-
Gabcd hcd + hab 12 gab h = 0. (101)
cent work has brought to light a number of interesting con-
In contrast to the naive approach, the covariant divergence of nections between the two. Deser and Waldron [15] have
this equation vanishes identically, and so cannot be used to demonstrated that, in (anti-)de Sitter background spacetimes,
relate
b h and
a hab . In place of this, the gauge invariance a massive spin-2 field is stable if and only if m2 2/3, or
of the vacuum theory remains intact22 , and the field equation m = 0. While it is intriguing that our de Sitter background
may be simplified by setting h = 0, a hab = 0: field equation (103) suggests precisely the same special value
of m2 = 2/3, Deser and Waldrons analysis differs signifi-

2 hab 2Rdabchdc = 0. (102) cantly from our own, so this superficial observation may be
misleading. In particular, whereas our mass-term arises as
Surprisingly, the contribution from has been cancelled by a direct result of the perturbative expansion, Deser and Wal-
a term proportional to the background Ricci tensor, resulting dron add their mass-term to the action by hand. Thus it is
far from clear that the massive gravitons of their paper corre-
spond to the physical system considered above. In contrast,
Novello and Neves [16] claim to prove that m2 = 2/3,
22 If we wish to extend our discussion of gauge invariance (III C) to include with the implication that 0. This approach considers an
background matter in general, we would need to account for the gauge- unusual generalisation of the spin-2 field equation to curved
fixing implicit in our starting assumption A = 0, which is obviously not backgrounds, making a non-standard choice for the covari-
preserved by a (first-order) infinitesimal diffeomorphism A = L A .
However, because is constant over spacetime, no such difficulty arises antization ambiguous term discussed in section II A. Thus,
here, and the transformations hab = 2 (a b) remain a symmetry of the while their calculations arguably describe a spin-2 field, this
equations of motion. does not appear to be a natural way to describe the spin-2 field
 14

that results from perturbations of the metric (or its inverse) in The mass-terms induced by a scalar field (99), a cosmological
Einsteins theory. It is our intention to disentangle the connec- constant (100) and electromagnetism (105) have been calcu-
tions between these two approaches, and our own, in a later lated.
publication.
For the sake of completeness, we conclude this section with
an example of a mass-term that can have M > 0, and Nabcd 6= Acknowledgments
0. Unlike , however, we shall not attempt to derive any
of the implications for the equations of motion. Consider an L.M.B. is supported by STFC and St. Johns College, Cam-
electromagnetic 1-form background Aa , with Lagrangian bridge. The Mathematica package Ricci was used for the
bookkeeping part of the calculations that produced equations
LA = 41 F 2 = 41 gab gcd Fac Fbd , (104) (26) and (B15). We thank Stanley Deser and Tomas Ortn for
their helpful comments, and Thanu Padmanabhan for an en-
and note that Fab 2[a Ab] is independent of the metric. The lightening discussion.
calculation yields
 
A = 41 F 2 hab hab 12 h2 habhcd Fac Fbd , (105) APPENDIX A: PADMANABHANS ANALYSIS

which has the aforementioned properties. The recent article by Padmanabhan [1] unearths many sig-
nificant shortcomings of the well known arguments [2, 3, 4, 5]
that supposedly derive Einsteins equations by coupling the
V. CONCLUSION Fierz-Pauli graviton to its own energy-momentum tensor.
Here we attempt to summarize his observations, and explain
Contrary to the prevailing maxim, coupling the classical their relation to this present work.
Fierz-Pauli graviton to its own energy and momentum does In broad terms, Padmanabhans criticisms fall into three ar-
not recreate general relativity order by order. However, there eas:
is an alternative action for the graviton (4) for which energy-
momentum self-coupling is consistent with Einsteins theory. 1. The Einstein-Hilbert action consists of a bulk term (the
Using this action, the energy-momentum tensor of the gravi- 2 action) and a surface term. The latter includes a
ton (26), added as a source to the gravitons first-order equa- piece linear in h , so there can be no way to construct
tion of motion (13), builds a field equation consistent with the it from a self-coupling procedure that starts with an ac-
Einstein equations to second-order. Furthermore, the pertur- tion that is already quadratic in h .23
bative formalism developed in section III reveals that our ac-
tion provides sufficient information to reconstruct general rel- 2. The starting point, the Fierz-Pauli Lagrangian (8), de-
ativity to arbitrary accuracy: a simple recurrence relation (46) scribes a Lorentz invariant field theory, and yet the end
identifies the energy-momentum tensor at one order as the ap- result, general relativity, is generally covariant. It is
propriate contribution to the action at the next. To any order N, claimed that this metamorphosis only occurs because
this scheme assembles an action that dictates field equations general covariance has been assumed in the various
(53) in which the gravitons N th -order energy-momentum ten- derivations, in which case it is no big deal to obtain
sor is the source. Einsteins theory. More generally, the classic boot-
The formal machinery used to understand vacuum perturba- strapping arguments wield ideas developed in general
tions is easily extended to include matter, although the physi- relativity (such as Hilberts definition of the energy-
cal interpretation of the most general approach, in which mat- momentum tensor) or use knowledge of the end result
ter comprises both a background field and a small perturba- to achieve their goal. Hence they cannot be regarded as
tion, is less than transparent. Focusing on matter perturba- a derivation of general relativity from first principles.
tions separately from non-vacuum backgrounds serves to clar-
3. The first-order field equation can only take a symmetric
ify the formalism significantly. In a vacuum background, the
tensor as its source; the canonical energy-momentum
interactions between the graviton and perturbations of a free
tensor (15) is not necessarily symmetric, and although
matter field lead to a field equation (87) in which the source
it can be made to be so, this process is not unique.
for the graviton is the sum of gravitational and matter energy-
Therefore the energy-momentum self-coupling proce-
momentum. This interaction inevitably induces a source in
dure is ill-defined. The Hilbert definition is uniquely
the field equations for matter (88). Alternatively, one may
determined by the action, but to use it would violate
neglect matter perturbations and examine the consequences
of a non-vacuum background. In this case, the dynamics and
energy-momentum of the graviton are prescribed by the action
(93), generalizing our previous ansatz. Surprisingly, the back- 23 The argument given by Padmanabhan is phrased in terms of non-analyticity
ground matter appears to induce a mass-term in the graviton in a dimensionful coupling constant. This form of the argument depends
action, although it is currently unclear to what extent its inter- on his particular choice of normalization for h and SEH , but is essentially
pretation as a mass is valid at the level of the field equations. equivalent to the statement given here.
 15

criticism 2. Crucially, even if we allow ourselves to use inertial coordinate system; thus expressing physical equations
Hilberts definition, we still fail to recover the correct in coordinate invariant notation is an invaluable tool for de-
source term for the second-order field equation. scribing how their dynamics are modified by gravity. It is pos-
sible that when Padmanabhan refers to general covariance
It is to this very last crucial point that we have devoted the he is referring to the equivalence principle also. As the latter
bulk of this paper. We now wish to explain our position with is tantamount to identifying the gravitational field with a dy-
regards to the first two criticisms, and also Padmanabhans namical metric, he would certainly be correct to criticise any
proposed solution to the third. derivation that contained such a step; needless to say, we do
1. Our approach expressly avoids discussing surface terms. not appeal to the equivalence principle in our approach.
This has greatly streamlined our formalism, and because such General covariance aside, though, Padmanabhans objec-
terms are completely irrelevant for determining field equa- tion to the use of curved-space ideas is a valid one, indicating
tions or energy-momentum tensors, the only price to pay for that none of the classic arguments constitute a derivation from
this simplicity is that we can only claim to reconstruct the first principles. Our approach certainly makes use of curved-
Einstein-Hilbert action modulo surface terms.24 In this sense, space concepts; however our goals are perhaps not quite so
Padmanabhans first criticism still stands, although it is un- bold as the other derivations that Padmanabhan has scruti-
clear whether it has any great importance. If the action is an nized: we do not pretend to derive general relativity purely
integral over the whole manifold, and asymptotic conditions from the ideas of Lorentz-invariant field theory. It should be
apply to hab such that the surface term at infinity vanishes, stressed, however, that even if some of the kinematical content
then of course there is no distinction between the Einstein- of general relativity is in some way assumed (curved space-
Hilbert action and the action we have constructed. Even if the time, functional derivatives with respect to the metric, etc.)
action is an integral over a manifold with a boundary, so long it is still a big deal to derive the dynamical content of the
as we consider the action to be a functional over all fields with theory, Einsteins equations.
a particular boundary configuration (just as we might think of 3. We have already explained our position with regards
the action of a particle as a functional over all paths with par- to the definition of the energy-momentum tensor in section
ticular end-points) the two actions differ only by an irrelevant II C; the only reason that Hilberts definition is unpalatable to
constant. Besides, in situations where contributions from the Padmanabhan is that his aim is to start with as little curved-
boundary really are important, one does not typically use the space mathematics as he can. However, the failure of the
Einstein-Hilbert action anyway: the Gibbons-Hawking-York Hilbert energy-momentum tensor to give the correct second-
boundary term [17, 18] must be included to remove the de- order term for the Einstein field equations is a more signif-
pendence on second derivatives of the metric. This allows the icant stumbling-block. We have explained our remedy, the
field equations to be derived using a variational principle that use of a different starting action, in the body of this paper.
only demands that the variation in the fields (and not also their Padmanabhan, on the other hand, eschews energy-momentum
derivatives) vanish on the boundary. self-coupling and introduces a new object S that he defines
Padmanabhans major concern is that the surface term of with the following algorithm. Start with a Lorentz invariant
the Einstein-Hilbert action has some quantum mechanical sig- Lagrangian L ( , h , h ) expressed in Lorentzian co-
nificance. As the nature of quantum gravity has yet to be un- ordinates {x }. Replace every instance of with the metric
derstood, it remains unclear whether or not this is the case. We f(g , h , h ); note
g to produce a new Lagrangian L
stress once again that the analysis in this paper is purely clas-
that this is not the same as expressing L in an arbitrary coor-
sical, and that we make no claims as to a quantum mechani-
dinate system because the partial derivatives have not been
cal interpretation. Furthermore, it is not even known whether
upgraded to covariant derivatives . We can now define
the graviton is a useful theoretical object for describing quan-
tum gravity. We note again that the Gibbons-Hawking-York 
gL f
boundary term is usually included in quantum gravity investi-
S 2  . (A1)
gations for which the boundary is not negligible. g 
g=
2. It is our view that Padmanabhans concerns about gen-
eral covariance are unjustified: we take the position of Wein- The subscript reminds us that we must set g = after
berg [19], that general covariance by itself is empty of phys- taking the metric derivative, as we are supposedly working
ical content. Any theory (Lorentz invariant or not) can be in Lorentzian coordinates. Padmanabhan claims to be able to
expressed in arbitrary curvilinear coordinates, so the require- reconstruct the 2 action by coupling h to this new object
ment of general covariance cannot, in and of itself, constrain S . Unfortunately S has a number of highly undesirable
the sort of theory one might construct. Rather, the kinemati- properties, suggesting that it is a rather unnatural object, ill-
cal content of general relativity is encapsulated by the equiva- defined in its current form.25
lence principle, that the effect of gravity vanishes locally in an

25 In private communication, Padmanabhan has indicated that he shares our

24 Note that this does not nessesarily mean that we have constructed
the 2 concerns about S and does not believe it to be of any fundamental impor-
action, only that the integrand of the action differs from gR by some tance; hence we present the case against S for the sake of completeness
total divergence. rather than rebuttal.
 16

Firstly, as it has been constructed from a Lagrangian rather will reduce to S in Lorentzian coordinates. This expres-
than an action, S depends directly on surface terms. This sion gives us some insight into the geometrical meaning of
introduces a very large ambiguity, as S will depend on Padmanabhans half-covariantized algorithm; in particular it
whether we write the integrand of the action in the form ( h)2 , reveals that the derivative / g used to define S is in
as Padmanabhan does, in the form h 2 h, or as some arbitrary fact exploring geometries (infinitesimally close to Minkowski
combination of both. Each possibility defines a different S spacetime) with connections that are not metric compatible.27
and (presumably) leads to a different self-coupled limit for the It is perhaps unsurprising that this -constant derivative in-
graviton. It seems that the only remedy for this ambiguity is troduces a new layer of ambiguity to the procedure, as we
to artificially stipulate that L contain no second derivatives, can now alter Sab by adding terms proportional to 0 = c gab
although we note in passing that even this leaves us free to add to the Lagrangian. Although this might seem a rather con-
surface terms of the form ( A ) in theories for fields other trived objection, it is in fact a very common consideration.
than the graviton. For example, suppose the Lagrangian includes a term of the
The second troubling aspect to S is the half- form a ha ; should we calculate Sab by acting with / g|
b
covariantizing algorithm used to construct L f. It should be on a (gac hcb ), or should we first commute the metric past the
clear that this procedure has only been defined in Lorentzian covariant derivative, and act on gac a hcb instead? Note that
coordinates, thus the matrix S does not really constitute this issue would have been invisible in Lorentzian coordinates
the components of a tensor, as we have not explained how because
their values change when expressed in another coordinate sys-
tem.26 There are essentially two ways to extend the definition
(A1) to include curvilinear coordinates. The trivial solution c ge f 
 = 2(a b) ,
is to construct the tensor Sab S ( )a ( )b using the vec- gab  c(e f )
(A3)

tors {( )a }, partial derivatives with respect to the Lorentzian
coordinates used to calculate S in the first place. This ob-
viously defines a genuine tensor, so the components S of

which we would have automatically set to zero. It seems the
Sab in some curvilinear coordinate system {x } can be cal-

only way to avoid this uncertainty in Sab is to introduce an-

culated, and they will be related to S by the usual transfor- other artificial constraint on the Lagrangian: we insist that it
mation rules. It should be clear, however, that this solution be written in such a way that no derivatives act on the metric.
is rather unnatural: suppose we have a Lagrangian expressed This should be achieved by commuting covariant derivatives
in a curvilinear coordinate system, then the only way to calcu- through the metric, rather than integrating by parts, due to the
late the components S in that system is to first transform to

aforementioned issues with surface terms.
Lorentzian coordinates, calculate S according to (A1), and
We shall take our analysis of S no further at this time. It
then transform back to our original coordinate system. Also,
is still uncertain whether this object can be generalized, nat-
because this process picks out a special set of coordinates,
urally and uniquely, to form a genuine tensor; without such
there is also no reason to expect that Sab can be written as a
a generalization it is difficult to ascertain what sort of mathe-
tensorial function of hab , gab and
a . The natural way to pro-
matical object the matrix of functions S is supposed to rep-
ceed would be to generalize the definition (A1) in such a way
resent. Although we cannot claim to have exhausted all possi-
that we could calculate S working in any coordinate sys-

bilities, the evidence before us suggests, at the very least, that
tem. It might seem that a viable solution would be to define
this goal is not easily achieved.
the tensor
 Aside from these technical issues, we should also empha-
2 gL 
Sab , (A2) size that, unlike the energy-momentum tensor, S has no
g gab  apparent physical interpretation beyond its supposed role in
a graviton self-coupling scheme. Energy-momentum self-
where L = L (gab , hab , c hab ) is the fully covariant La-
coupling was justified by analogy with matter-gravity cou-
grangian, and the subscript indicates that the Christoffel sym-
pling, and advanced by the notion that the energy-momentum
bols abc are to be treated as independent of the metric and
of all fields should source gravitation. In contrast, the self-
held constant in the derivative. This expression generalizes
coupling scheme involving S only serves to set gravity apart
(A1) to define a tensor Sab in a coordinate invariant fashion;
from the other fields. Furthermore, our solution displays an
because the Christoffel symbols are held constant, no term
unusual symmetry between the coupling terms in the action
arises from a variation of the covariant derivatives, and Sab
and source terms generated in the field equations as a result
(see III B); this symmetry is broken by Padmanabhans self-
coupling procedure.
26 The insistence that we be able to calculate the components of this object in
arbitrary coordinates has nothing to do with curved spacetime or general
relativity. Rather, this reflects the perfectly reasonable expectation that we
should be able to express Padmanabhans self-coupling procedure in flat- 27 This is the same operation as the derivative used to acquire the Einstein
space spherical polar coordinates, for example, or any other coordinate equations from the Palatini action [20], although here we will have no cause
system we choose. to perform the complementary derivative / |g .
 17

APPENDIX B: EXPANSION OF Gab Thus,

(1) (1) (1)

Here we determine the first two terms of the expansion of Gab = Rab 21 gab Rcd gcd
the Einstein tensor (a h c + 1
c
= b) 2 2 a b h
2 hab + 1

(1) (2)
Gab = Gab + Gab + O(h3 ), (B1)  
1 gab
2
c
d hcd + 2h , (B12)
induced by a perturbation of the inverse metric about a vac-
uum background: which confirms that Gabcd , as defined in (5), represents the
ab ab ab linearised Einstein tensor:
g = g + h , (B2)
Gab = 0. (B3) (1)
Gabcd hcd = Gab . (B13)
The perturbation in the metric is of course fixed by the rela-
In particular, note that both sides of this equation agree on the
tionship gab gbc = ca , (a h c ; this is the descendant of
order of the derivatives in c
b)
gab = gab hab + hac hcb + O(h3 ). (B4) the covariantization ambiguous term discussed in section II A.
(2)
To find Gab , start with
To begin, introduce a connection E abc between the derivative
 
operators a and a: (2) (2) (2) (1)
Gab = Rab 21 gab Rcd gcd + Rcd hcd
E abc = 12 gab (
b gcd +
c gbd
d gbc ). (B5) (1)
+ 12 hab Rcd gcd , (B14)
This allow the Ricci tensor to be expressed as
and substitute equations (B10) and (B11), followed by (B8)
 
and (B9). The bookkeeping for this calculation is characteris-
Rab = 2 [c E c + E c E d
a]b d[c a]b . (B6)
tically laborious, but is easily accomplished using a computer
algebra package; the result is
From (B5) it is clear that
(2)
E
a(0)
= 0, (B7) Gab = tab + 12 hGabcd hcd , (B15)
bc
a(1)
E bc = 12 gad (2
(b hc)d
d hbc ), (B8) where tab is given by (26). As expounded in section II B, and
a(2) now confirmed by direct calculation (B13), the first-order ap-
E bc = 21 had (2
(b hc)d d hbc )
proximation to the Einstein field equation is Gabcd hcd = 0, so
+ 21 g (2
ad (b (hc)e h )
e
d
d (hbe he )).
c (B9) Gabcd hcd = O(h2 ) must hold true at second-order. Clearly it
follows from this that hGabcd hcd = O(h3 ), and hence (28) is
(1) (2) verified.
Hence the terms of the expansion Rab = Rab + Rab + O(h3 )
(2)
can be computed as follows: The third-order difference between Gab and tab exists
because the field
equationapproximated to second-order in
(1) c(1) (29) is actually gGab / g = 0; this is of course entirely
Rab = 2
[c E
a]b
(B10)
  equivalent to the usual form of the Einstein field equation
(2) c(2) c(1) d(1)
Rab = 2 [c E a]b + E d[c E a]b .
(B11) Gab = 0.

[1] T. Padmanabhan, Int.J.Mod.Phys. D 17, 367 (2008), URL ciety of London. Series A, Mathematical and Physi-
arXiv:gr-qc/0409089v1. cal Sciences 173, 211 (1939), ISSN 00804630, URL
[2] S. Deser, Gen. Rel. Grav. 1, 9 (1970). http://www.jstor.org/stable/97457.
[3] R. P. Feynman, F. B. Morinigo, and W. G. Wagner, Feynman [10] R. M. Wald, General Relativity (University of Chicago Press,
Lectures on Gravitation (Addison-Wesley, 1995), pp. 7488. 1984), p. 437.
[4] S. N. Gupta, Phys. Rev. 96, 1683 (1954). [11] K. Kuchar, Journal of Mathematical Physics 17, 801 (1976),
[5] R. H. Kraichnan, Phys. Rev. 98, 1118 (1955). URL http://link.aip.org/link/?JMP/17/801/1.
[6] S. Deser, Classical and Quantum Gravity 4, L99 (1987), URL [12] P. D. Mannheim, Physical Review D (Particles, Fields, Grav-
http://stacks.iop.org/0264-9381/4/L99. itation, and Cosmology) 74, 024019 (pages 6) (2006), URL
[7] D. G. Boulware and S. Deser, Annals of Physics 89, 193 http://link.aps.org/abstract/PRD/v74/e024019.
(1975), ISSN 0003-4916. [13] G. Magnano and L. M. Sokolowski, Classi-
[8] T. Ortin, Gravity and Strings (Cambridge University Press, cal and Quantum Gravity 19, 223 (2002), URL
2004), chap. 3.2. http://stacks.iop.org/0264-9381/19/223.
[9] M. Fierz and W. Pauli, Proceedings of the Royal So- [14] A. Lasenby, C. Doran, and S. Gull, Phil. Trans. R. Soc. Lond.
 18

A 356, 487 (1998), URL arXiv:gr-qc/0405033v1. (1977).

[15] S. Deser and A. Waldron, Physics Letters B 508, 347 (2001), [19] S. Weinberg, Gravitation and Cosmology (Wiley, 1972), pp.
ISSN 0370-2693. 9193.
[16] M. Novello and R. P. Neves, Classical and [20] M. P. Hobson, G. P. Efstathiou, and A. N. Lasenby, General
Quantum Gravity 20, L67 (2003), URL Relativity: An Introduction for Physicists (Cambridge Univer-
http://stacks.iop.org/0264-9381/20/L67. sity Press, 2006), chap. 19.10.
[17] J. W. York, Phys. Rev. Lett. 28, 1082 (1972).
[18] G. W. Gibbons and S. W. Hawking, Phys. Rev. D 15, 2752