When is multimetric gravity ghost-free?
Kouichi Nomura and Jiro Soda
Department of Physics, Kyoto University, Kyoto 606-8502, Japan(Received 21 July 2012; published 26 October 2012)
We study ghosts in multimetric gravity by combining the minisuperspace and the Hamiltonian
constraint analysis. We first revisit bimetric gravity and explain why it is ghost-free. Then, we apply
our method to trimetric gravity and clarify when the model contains a ghost. More precisely, we prove
trimetric gravity generically contains a ghost. However, if we cut the interaction of a pair of metrics,
trimetric gravity becomes ghost-free. We further extend the Hamiltonian analysis to general multimetric
gravity and calculate the number of ghosts in various models. Thus, we find multimetric gravity with loop
type interactions never becomes ghost-free.
DOI: 10.1103/PhysRevD.86.084052 PACS numbers: 04.50.�h
I. INTRODUCTION
It is interesting to explore the possibility that a gravitonis massive from both a theoretical and phenomenologicalpoint of view. Theoretically, it is challenging because ofvarious difficulties in constructing a consistent model for amassive graviton. At the linear level, Fiertz and Paulisucceeded in constructing a ghost-free model for a massivegraviton [1]. However, it was soon recognized that thereis a tension between the theory and experiments, the so-called van Dam—Veltman-Zaharov discontinuity [2,3]. Itwas suggested that the nonlinearity resolves the vanDam—Veltman-Zaharov discontinuity [4]. Unfortunately,it turned out that the nonlinearity gives rise to a ghost, theso-called Boulware-Deser (BD) ghost [5]. Recently, deRham, Gabadadze, and Tolley have succeeded in construct-ing ghost-free nonlinear massive gravity theory [6,7] (seethe review [8] and references therein). Still, there remainsvarious theoretically intriguing issues to be explored.Phenomenologically, there is a chance to explain the cur-rent accelerating universe based on massive gravity. In fact,there appears to be an effective cosmological constantproportional to the square of graviton mass [9–14]. It isworth studying this possibility in detail.
One peculiar feature of a massive theory of the graviton isthe necessity of a reference metric which breaks the diffeo-morphism invariance. It is natural to promote this referencemetric to a dynamical variable, which is nothing but bimet-ric gravity. Bimetric gravity contains two metrics, g and f,interacting with each other. The history of bimetric gravityis long [15–17]. Curiously, bimetric gravity also suffersfrom the ghost problem. Thanks to the recent developmentin massive gravity, however, Hassan and Rosen have pro-posed ghost-free bimetric gravity [18–20]. A natural ques-tion arises whether or not we can construct ghost-freemultimetric gravity. Actually, a naive extension of bimetricgravity to the trimetric case was proposed in Ref. [21].There, three metrics, g, f, and h, have a pair interactionbetween ðg; hÞ, ðh; fÞ, and ðf; gÞ, which forms a loop struc-ture. In contrast to bimetric gravity, however, the presence
or the absence of the BD ghost remains unknown. Recently,Hinterbichler and Rosen showed that a large class of multi-vielbein gravity is ghost-free [22]. The relation to metrictheory is also discussed [23]. However, the relation to themodels presented in Ref. [21] is not clear. The differencebetween Ref. [21] and [22] comes from the loop typeinteraction. In fact, the proof by the vielbein method isnot applicable to the loop type interaction. Hence, weneed to study multimetric gravity with a different approach.In this paper, we propose a simple method to study the
ghost problem and clarify when multimetric gravity isghost-free. A method often used for the ghost analysis isto examine models in the decoupling limit. However, a morehonest way for probing ghosts is to use the Hamiltonianconstraint analysis using the Arnowitt, Deser, and Misner(ADM) formalism [24]. The difficulty in studying multi-metric gravity with the constraint analysis comes from theexistence of a shift vector. To avoid the difficulty, weemploy the minisuperspace approximation. The minisuper-space reduction of phase space makes the analysis simple.Nevertheless, it is sufficient to identify ghosts because thisreduction process does not fail to capture ghosts.The organization of the paper is as follows. In Sec. II, we
revisit bimetric gravity and explain our strategy for theghost analysis. In Sec. III, we investigate trimetric gravityusing our method and find that the loop type interactionallows ghosts. In Sec. IV, we further extend the analysis togeneral N -metric gravity. We clarify when ghosts appearin the spectrum. The final section V, is devoted to theconclusion.
II. BIMETRIC GRAVITY REVISITED
In this section, we revisit bigravity and explain ourmethod to probe a ghost. It is already known that bimetricgravity is ghost-free [19]. Here, we show the same conclu-sion can be obtained using a simpleminisuperspace approxi-mation. In the context of massive gravity, the decouplinglimit analysis turns out to be a useful method for the ghostanalysis. However, the most complete one is to use the
PHYSICAL REVIEW D 86, 084052 (2012)
1550-7998=2012=86(8)=084052(10) 084052-1 � 2012 American Physical Society
Hamiltonian constraint analysis and count physical de-grees of freedom. Our strategy is to use the Hamiltonianconstraint analysis in the minisuperspace.
The action of ghost-free bimetric gravity [20] is given by
Sbi ¼ M2g
Zd4x
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detgp
R½g� þM2f
Zd4x
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detfp
R½f�
þ 2m2M2gf
Zd4x
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detgp X4
n¼0
�nenðffiffiffiffiffiffiffiffiffiffiffig�1f
qÞ; (1)
where the first and the second terms are the Einstein-Hilbertaction for each metric g and f from which we can calculatethe scalar curvatures R½g�, R½f�. Here, we have two Planckmasses Mg and Mf. The last term describes the interaction
between two metrics and �n are dimensionless couplingconstants. The other constants m andMgf are introduced to
adjust the mass dimension. The square root of the matrix is
defined such thatffiffiffiffiffiffiffiffiffiffiffig�1f
p ffiffiffiffiffiffiffiffiffiffiffig�1f
p ¼ g��f��. The interactionterms are constructed by enðXÞ which we define, formatrix X,
e0ðXÞ¼1; e1ðXÞ¼ trX;
e2ðXÞ¼1
2ðtr2X� trX2Þ;
e3ðXÞ¼1
6ðtr3X�3trXtrX2þ2trX3Þ;
e4ðXÞ¼ 1
24ðtr4X�6tr2XtrX2þ3tr2X2þ8trXtrX3�6trX4Þ
¼detX; (2)
where we used the notation trnX¼ðtrXÞn and trXn¼trðXnÞ.It is useful to represent the interaction by the diagram inFig. 1. Note that there is the order between g and f which isdenoted by the arrow. It is known that the interactionproduces massless and massive gravitons and the spectrumis free of the Boulware-Deser ghost. This feature comesfrom a specific interaction form found in massive gravitytheory. Remarkably, there exists the diagonal diffeomor-phism invariance which indicates the presence of the mass-less graviton.
Now, let us perform the Hamiltonian constraint analysisbased on the ADM formalism. In particular, to make theanalysis tractable, we employ the minisuperspace approach.Namely, we assume spatial homogeneity and express met-rics in terms of ADM variables as
g��dx�dx� ¼ �NðtÞ2dt2 þ �ijðtÞdxidxj; (3)
where N is a lapse function and �ij is a spatial metric.
Similarly, we can take the following ansatz,
f��dx�dx� ¼ �LðtÞ2dt2 þ!ijðtÞdxidxj; (4)
where L is a lapse function and !ij is a spatial metric. It is
convenient to write them in a matrix form,
g�� ¼ �N2 0
0 �ij
!; g�� ¼ �1=N2 0
0 �ij
!;
f�� ¼ �L2 0
0 !ij
!; f�� ¼ �1=L2 0
0 !ij
!;
(5)
where �ij and!ij are inverse matrices of spatial metrics �ij
and !ij. Then, a basic part of interaction terms can be
calculated to be
ðg�1fÞ�� ¼ L2=N2 0
0 �il!lj
!;
ffiffiffiffiffiffiffiffiffiffiffig�1f
q¼ L=N 0
0ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
p !
:
(6)
When we count physical degrees of freedom, the follow-ing must be taken into account. In the vacuum cases, wecan diagonalize one of two spatial metrics using diagonalspatial coordinate transformations. Performing a spatialcoordinate transformation xi ! �ðt0Þijxj, we can set one
spatial metric at the time t ¼ t0, �ijðt0Þ, a unit matrix �ij.
Moreover, since the orthogonal transformation does notchange �ijðt0Þ ¼ �ij, we can diagonalize _�ijðt0Þ simulta-
neously by using this freedom. At this stage, homogeneousspatial coordinates are completely fixed. Now, �ij and _�ij
are diagonal at the time t ¼ t0 as an initial condition. Thenwe assume the diagonal form of �ijðtÞ at all times, and
insert it into equations obtained from variations of action.Any contradiction never occurs in vacuum. Thus, we con-clude that one spatial metric �ijðtÞ can be diagonalized
because of the uniqueness of the solution. Hence, thenumber of components of one of two metrics reducesfrom 6 to 3. This fact will be used later.In this paper, for simplicity, we assume that interactions
are minimal [21,25], namely,
�0¼3; �1¼�1; �2¼0; �3¼0; �4¼1: (7)
Clearly, this simplification does not lose any generalityconcerning the ghost analysis. Then, the Lagrangian reads
L ¼ M2g�
ij _�ij þM2fp
ij _!ij � NCN � LCL; (8)
where �ij, pij are the canonical conjugate momentum of�ij, !ij. Here, we have defined
CN ¼ M2gffiffiffiffiffiffiffiffiffiffi
det�p
��ij�ij � 1
2�i
i�jj
��M2
g
ffiffiffiffiffiffiffiffiffiffiffiffiffiffidet�ð3Þ
qR½��
þ a1ffiffiffiffiffiffiffiffiffiffidet�
p ðtrffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q� 3Þ; (9)
FIG. 1 (color online). The interactionffiffiffiffiffiffiffiffiffiffiffig�1f
pcan be repre-
sented by a simple diagram. Each blob describes a spacetimewith a given metric. The arrow indicates the order of the product.
KOUICHI NOMURA AND JIRO SODA PHYSICAL REVIEW D 86, 084052 (2012)
084052-2
CL ¼ M2fffiffiffiffiffiffiffiffiffiffiffi
det!p
�pijpij � 1
2pi
ipjj
��M2
f
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffidet!ð3Þ
pR½!�
þ a1ðffiffiffiffiffiffiffiffiffiffidet�
p � ffiffiffiffiffiffiffiffiffiffiffidet!
p Þ; (10)
where the first two terms of constraints CN and CM come
from the Einstein-Hilbert term in the action, so ð3ÞR½�� andð3ÞR½!� are spatial scalar curvatures computed from � and!, respectively. The last term of each constraint comes fromthe interaction (see the derivation in Appendix A), and weuse a1 ¼ 2m2M2
gf. Now, the Hamiltonian is given by
H ¼ NCN þ LCL: (11)
Since there are two Lagrange multipliers, there are twoprimary constraints:
CN ¼ 0; CL ¼ 0: (12)
Moreover, we need to impose consistency conditions forthem,
_CN ¼ fCN;Hg ¼ LfCN; CLg � LCNL � 0;
_CL ¼ fCL;Hg ¼ NfCL; CNg � NCLN � 0; (13)
where the Poisson bracket fF;Gg is defined by
fF;Gg ¼�@F
@�mn
@G
@�mn �@F
@�mn
@G
@�mn
�
þ�
@F
@!mn
@G
@pmn �@F
@pmn
@G
@!mn
�: (14)
Here, ‘‘� 0’’ means ‘‘¼ 0’’ on the constraint surface.Notice that fF; Fg ¼ 0 because of spatial homogeneity.
To check if a secondary constraint arises or not, we haveto calculate the Poisson bracket CNL. From the calculationpresented in Appendix B, we obtain
CNL¼fCN;CLg¼a1
�1
2M2
g�ii�M2
f
ffiffiffiffiffiffiffiffiffiffiffidet�
det!
s �1
2pi
itrffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q
� trðffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qp!Þ
��: (15)
This leads to one secondary constraint CNL � 0. The con-sistency condition for the secondary constraint reads
_CNL ¼ NfCNL; CNg þ LfCNL; CLg � 0: (16)
This condition determines one of two Lagrange multipliersN and L. The remaining multiplier describes the diagonaltime reparametrization invariance in bimetric gravity.
The number of components of two metrics and theircanonical conjugates is 24. Since we can diagonalize oneof the two metrics, we should subtract 6 from this number.Recall that there are two primary constraints and onesecondary constraint. Furthermore, as we have one firstclass constraint, we have to put one gauge condition.Thus, the total number of degrees of freedom should beð24� 6� 2� 1� 1Þ=2 ¼ 7 in configuration space, whichmatches degrees of freedom of one massless graviton and
one massive graviton. This proves that the BD ghost isabsent in bimetric gravity described by the action (1).
III. TRIMETRIC GRAVITY
Now, we apply the method explained in the previoussection to trimetric gravity. In contrast to the bimetric gravity,there are two kinds of interactions, namely, the tree type andthe loop type interactions. We discuss both cases, separately.The action for trimetric gravity [21] can be written as
Stri ¼ M2g
Zd4x
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detgp
R½g� þM2f
Zd4x
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detfp
R½f�
þM2h
Zd4x
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� dethp
R½h�
þ 2m21M
2gf
Zd4x
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detgp X4
n¼0
�nenðffiffiffiffiffiffiffiffiffiffiffig�1f
qÞ
þ 2m22M
2fh
Zd4x
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detfp X4
n¼0
�0nenð
ffiffiffiffiffiffiffiffiffiffiffif�1h
qÞ
þ 2m23M
2hg
Zd4x
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� dethp X4
n¼0
�00nenð
ffiffiffiffiffiffiffiffiffiffiffih�1g
qÞ; (17)
where �n, �0n, and �00
n are free parameters and R½g�, R½f�,and R½h� are scalar curvatures constructed from metrics g,f, and h, respectively. We also introduced new mass pa-rameters m1, m2, m3, Mfh, Mhg and a Planck mass Mh. It
should be noted that there exists the diagonal diffeomor-phism invariance in this trimetric theory which makes oneof the gravitons massless. As discussed in Refs. [21,22],if this trimetric gravity contains no extra degrees of free-dom, the total number of degrees of freedom should be2þ 5þ 5 ¼ 12, which comes from one massless gravitonand two massive gravitons. From now on, we use
a1¼2m21M
2gf; a2¼2m2
2M2fh; a3¼2m2
3M2hg (18)
for notational simplicity. If we have a1 � 0, a2 � 0, anda3 � 0, all pairs ðg; fÞ, ðf; hÞ, and ðh; gÞ interact and wecall it the loop type interaction. When one of aiði ¼ 1; 2; 3Þis set to zero, two of three pairs of interactions remain,which we call the tree type interaction. The case where oneinteraction is cut is already proved to be ghost-free usingvielbein formalism [22]; however, for the loop type inter-action no one shows the presence or the absence of ghosts.In this paper, we settle this issue.Apparently, the full Hamiltonian constraint analysis is dif-
ficult. To circumvent this difficulty,we take themethodused inthe previous section. Namely, we assume spatial homogene-ity and express metrics in terms of ADM variables as
g��dx�dx� ¼ �NðtÞ2dt2 þ �ijðtÞdxidxj; (19)
where N is a lapse function and �ij is a spatial metric.
Similarly, we can take the following Ansatze,
f��dx�dx� ¼ �LðtÞ2dt2 þ!ijðtÞdxidxj; (20)
WHEN IS MULTIMETRIC GRAVITY GHOST-FREE? PHYSICAL REVIEW D 86, 084052 (2012)
084052-3
and
h��dx�dx� ¼ �QðtÞ2dt2 þ �ijðtÞdxidxj; (21)
where L andQ are lapse functions and!ij and �ij are spatial
metrics. To perform the Hamiltonian constraint analysis, weneed the Lagrangian in the ADM variables
L¼M2g�
ij _�ijþM2fp
ij _!ijþM2h
ij _�ij
�NCN�LCL�QCQ; (22)
where �ij, pij, and ij are the canonical conjugate momen-tum of �ij, !ij, and �ij. Here, three Hamiltonian constraints
CN ¼ M2gffiffiffiffiffiffiffiffiffiffi
det�p
��ij�ij � 1
2�i
i�jj
��M2
g
ffiffiffiffiffiffiffiffiffiffiffiffiffiffidet�ð3Þ
qR½��
þ a1ffiffiffiffiffiffiffiffiffiffidet�
p ðtrffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q� 3Þ þ a3ð
ffiffiffiffiffiffiffiffiffiffidet�
p � ffiffiffiffiffiffiffiffiffiffidet�
p Þ;(23)
CL ¼ M2fffiffiffiffiffiffiffiffiffiffi
det!p
�pijpij � 1
2pi
ipjj
��M2
f
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffidet!ð3Þ
pR½!�
þ a2ffiffiffiffiffiffiffiffiffiffidet!
p ðtrffiffiffiffiffiffiffiffiffiffiffiffiffi!�1�
q� 3Þ þ a1ð
ffiffiffiffiffiffiffiffiffiffidet�
p � ffiffiffiffiffiffiffiffiffiffidet!
p Þ;(24)
and
CQ ¼ M2hffiffiffiffiffiffiffiffiffiffi
det�p
�ijij � 1
2i
ijj
��M2
h
ffiffiffiffiffiffiffiffiffiffiffiffiffiffidet�ð3Þ
qR½��
þ a3ffiffiffiffiffiffiffiffiffiffidet�
p ðtrffiffiffiffiffiffiffiffiffiffiffiffi��1�
q� 3Þ þ a2ð
ffiffiffiffiffiffiffiffiffiffidet!
p � ffiffiffiffiffiffiffiffiffiffidet�
p Þ(25)
emerge. The first line of each Hamiltonian constraint comes
from the Einstein-Hilbert term in the action, so ð3ÞR½��,ð3ÞR½!�, and ð3ÞR½�� are spatial scalar curvatures calculatedfrom �, !, and �, respectively. The other terms can bederived as explained in Appendix A. Then, the Hamiltoniancan be read off as
H ¼ NCN þ LCL þQCQ: (26)
Since there are three Lagrange multipliers, there arise threeprimary constraints:
CN ¼ 0; CL ¼ 0; CQ ¼ 0: (27)
Moreover, we need consistency conditions for them:
_CN ¼ fCN;Hg ¼ LfCN; CLg þQfCN; CQg � 0;
_CL ¼ fCL;Hg ¼ NfCL; CNg þQfCL; CQg � 0;
_CQ ¼ fCQ;Hg ¼ NfCQ;CNg þ LfCQ;CLg � 0:
(28)
To check if secondary constraints arise, we must calculatePoisson brackets. From the calculation in Appendix B, weobtain
CNL�fCN;CLg¼a1
�1
2M2
g�ii�M2
f
ffiffiffiffiffiffiffiffiffiffidet�
det!
s �1
2pi
itrffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q
� trðffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qp!Þ
��: (29)
By performing permutations among g ¼ ðN;�Þ, f ¼ ðL;!Þ,and h ¼ ðQ;�Þ, we also get
CLQ�fCL;CQg¼a2
�1
2M2
fpii�M2
h
ffiffiffiffiffiffiffiffiffiffidet!
det�
s �1
2i
itrffiffiffiffiffiffiffiffiffiffiffiffiffi!�1�
q
� trðffiffiffiffiffiffiffiffiffiffiffiffiffi!�1�
q�Þ
��(30)
and
CQN �fCQ;CNg¼a3
�1
2M2
hii�M2
g
ffiffiffiffiffiffiffiffiffiffidet�
det�
s �1
2�i
itrffiffiffiffiffiffiffiffiffiffiffiffi��1�
q
� trðffiffiffiffiffiffiffiffiffiffiffiffi��1�
q��Þ
��: (31)
In general, quantities inside the bracket do not vanish. Hence,the coefficients a1, a2, and a3 determine the consistencyconditions.
A. Tree type interaction
In this subsection, we consider the tree type interaction
a1 � 0; a2 ¼ 0; a3 � 0; (32)
which cuts interaction between f and h as in Fig. 2. In anycase, there are primary constraints (27). Since CQL ¼CLQ ¼ 0 trivially holds, consistency conditions (28) lead
to equations
LCNLþQCNQ�0; NCLN �0; NCQN �0: (33)
Hence, we have two secondary constraints,
CNL � 0; CNQ � 0: (34)
Moreover, we must impose consistency conditions
FIG. 2 (color online). The diagram represents the tree typeinteraction.
KOUICHI NOMURA AND JIRO SODA PHYSICAL REVIEW D 86, 084052 (2012)
084052-4
_CNL ¼ fCNL;Hg¼ NfCNL; CNg þ LfCNL; CLg þQfCNL; CQg � 0;
(35)
_CQL ¼ fCQL;Hg¼ NfCQL; CNg þ LfCQL; CLg þQfCQL; CQg � 0;
(36)
which determine two of three Lagrange multipliers N, L,and Q. The remaining multiplier is related to the gaugetransformation.
Eventually, we have five constraints and one gauge free-dom. In trimetric gravity, propagating modes are spatialmetrics. Each of them has six components, but as is alreadyexplained we can diagonalize one of them. Hence, trimetricgravity has 3þ 6þ 6 ¼ 15 degrees of freedom in configu-ration space and 15� 2 ¼ 30 in phase apace. Thus, thetotal number of degrees of freedom is ð30� 5� 1Þ=2 ¼ 12which matches the physical degrees of one massless andtwo massive gravitons. Therefore, no BD ghost exists in thespectrum. This conclusion is consistent with the one ob-tained by the vielbein method [22].
B. Loop type interaction
Now, we consider the more general loop type interactionrepresented by Fig. 3:
a1 � 0; a2 � 0; a3 � 0: (37)
It is obvious that
fCN;CLg�0; fCL;CQg�0; fCQ;CNg�0 (38)
even on the constraint surface. Hence, consistency condi-tions (28) do not generate any secondary constraint. Instead,it determines Lagrange multipliers N, L, and Q. However,due to the antisymmetric property of Poisson brackets
CNL ¼ �fCN; CLg ¼ �fCL; CNg ¼ CLN; (39)
CLQ ¼ fCL; CQg ¼ �CQL; (40)
CQN ¼ fCQ;CNg ¼ �CNQ; (41)
only two of them are determined. For example, choosing
L ¼ �CNQ
CNL
Q; N ¼ �CLQ
CLN
Q; (42)
all of consistency conditions (28) are satisfied.To conclude, we have three primary constraints and we
need one gauge condition to fix one undetermined Lagrangemultiplier which is associated with the time reparametriza-tion invariance. In trimetric gravity, as is already counted,there are 3þ 6þ 6 ¼ 15 degrees of freedom in configura-tion space and 15� 2 ¼ 30 in phase apace. In phase space,we have three constraints and one gauge condition, so thetotal number of degrees of freedom is ð30� 3� 1Þ=2 ¼13. If no BD ghost is present, there must be 2þ 5þ 5 ¼ 12degrees of freedomwhich come from one massless gravitonand two massive gravitons. Therefore, one extra degree offreedom exists and it should be a BD ghost. Thus, we haveproved the existence of a ghost in generic trimetric gravity.
IV. GENERAL MULTIMETRIC MODELS
Now, we are in a position to discuss more general cases.We explicitly calculate the number of ghosts if they exist.In this section, we consider N dynamical metrics gk
(k ¼ 1; 2; . . . ;N ) and interaction terms such as
XNk¼1
akffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detgk
p X4n¼0
�k;nenðffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffig�1k gkþ1
qÞ; (43)
where we define gNþ1 ¼ g1 and for later purpose we alsoneed g0 ¼ gN . Let us describe the interaction between twometrics gk and gkþ1 in terms of the ADM form of metrics
ds2k ¼ �N2kðtÞdt2 þ �k;ijðtÞdxidxj: (44)
Schematically, the interaction can be written as
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detgkp X4
n¼0
�k;nenðffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffig�1k gkþ1
qÞ
¼ NkFkð�k: �kþ1Þ þ Nkþ1Gkð�k: �kþ1Þ; (45)
whereFk andGk are some functions determined by parame-ters �k;n. Thus, the total interaction terms are given by
XNk¼1
akffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�detgk
p X4n¼0
�k;nenðffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffig�1k gkþ1
qÞ
¼XNk¼1
NkfakFkð�k:�kþ1Þþak�1Gk�1ð�k�1:�kÞg: (46)
The Hamiltonian becomes
H ¼ XNk¼1
NkCk; (47)
Ck ¼ C0kð�k; �kÞ � akFkð�k: �kþ1Þ
� ak�1Gk�1ð�k�1: �kÞ; (48)FIG. 3 (color online). The diagram represents the loop typeinteraction.
WHEN IS MULTIMETRIC GRAVITY GHOST-FREE? PHYSICAL REVIEW D 86, 084052 (2012)
084052-5
where C0k comes from the Einstein-Hilbert term for gk, so it
contains �k and its canonical conjugate momentum �k.Corresponding toN Lagrange multipliers, we haveN
primary constraints
Ck ¼ 0; ðk ¼ 1; 2; . . . ;N Þ: (49)
Next, we have to examine N consistency conditions
_C k ¼ Ck;k�1Nk�1 þ Ck;kþ1Nkþ1 � 0; (50)
where Ck;l ¼ fCk; Clg and Ck;l ¼ 0 if jk� lj ^ 2. In this
formula, N0 ¼ NN and NNþ1 ¼ N1 should be under-stood. Note that the explicit calculation gives rise to theimportant information
Ck;kþ1 / ak: (51)
The structure of this matrix depends on an odd or evennumber. For example, in the case N ¼ 4, we have
Ck;l ¼
0 C1;2 0 C1;4
�C1;2 0 C2;3 0
0 �C2;3 0 C3;4
�C1;4 0 �C3;4 0
0BBBBB@
1CCCCCA; (52)
while, in the case of N ¼ 5, we get
Ck;l ¼
0 C1;2 0 0 C1;5
�C1;2 0 C2;3 0 0
0 �C2;3 0 C3;4 0
0 0 �C3;4 0 C4;5
�C1;5 0 0 �C4;5 0
0BBBBBBBB@
1CCCCCCCCA: (53)
In the case of an odd number of metrics, we cannot split theequations into two independent sets, while, in the case ofan even number of metrics, we can split a set of equationsinto two independent groups of equations. Hence, we haveto discuss the two cases, separately.
A. Tree type interaction
First, we consider the tree type interaction.If we cut one of ðgk; gkþ1Þk¼1;2;...;N interactions as in
Fig. 4, for example, setting a1 ¼ 0, Eq. (50) leads toN �1secondary constraints
Ck;kþ1 � 0; ðk ¼ 2; 3; . . . ;N Þ; (54)
and their consistency conditions
_C k;kþ1 ¼XNl¼1
fCk;kþ1; ClgNl � 0; ðk ¼ 2; 3; . . . ;N Þ
(55)
determine N � 1 of Nk (k ¼ 1; 2; . . . ;N ), only oneLagrange multiplier remains undetermined. Therefore, thetotal number of degrees of freedom can be deduced as
1
2ð2ð3þ6ðN �1ÞÞ�N �ðN �1Þ�1Þ¼5ðN �1Þþ2;
(56)
which corresponds to N � 1 massive gravitons and onemassless. Therefore, there exists no BD ghost. This con-clusion is also consistent with the one obtained by thevielbein method [22].
B. Loop type interaction
Now, we come to our main point.If all of ðgk; gkþ1Þk¼1;2;...;N interactions exist as in Fig. 5,
the analysis gets a little complicated. We have to discussodd and even numbers, separately.
1. Odd number of metrics
First, we consider the case where N ¼ 2mþ 1, wherem is a natural number. In this case, we can classify Eq. (50)into the following four parts:
C2k;2k�1N2k�1 þC2k;2kþ1N2kþ1 ¼ 0 ðk¼ 1;2;3; . . . ;mÞ;(57)
C2k�1;2k�2N2k�2 þ C2k�1;2kN2k ¼ 0 ðk ¼ 2; 3; . . . ; mÞ;(58)
FIG. 4 (color online). The diagram represents the tree typeinteraction.
FIG. 5 (color online). The diagram represents the loop typeinteraction.
KOUICHI NOMURA AND JIRO SODA PHYSICAL REVIEW D 86, 084052 (2012)
084052-6
C1;2mþ1N2mþ1 þ C1;2N2 ¼ 0; (59)
C2mþ1;2mN2m þ C2mþ1;1N1 ¼ 0: (60)
Solving Eq. (57), we see all of N2kþ1 (k ¼ 1; 2; . . . ; m) canbe expressed by N1. Similarly, Eq. (58) can be used toexpress N2k (k ¼ 2; 3; . . . ; m) in terms of N2. Substitutingthese results into Eqs. (59) and (60), we obtain a singleequation which determines N2 by N1. Thus, Eq. (50)determines N � 1 Lagrange multipliers, and one multi-plier is left undetermined, which reflects the existence ofgauge symmetry.
In the case of an odd number of metrics, there is nosecondary constraint, while we need one gauge conditionto fix the gauge degree of freedom. In conclusion, the totalnumber of degrees of freedom can be calculated as
1
2ð2ð3þ 6ðN � 1ÞÞ �N � 1Þ
¼ 5ðN � 1Þ þ 2þN � 1
2: (61)
Here, the first two terms correspond to massive and mass-less gravitons, respectively. The last one should be BDghosts and the number of ghosts is given by ðN � 1Þ=2.
2. Even number of metrics
Next, we consider the case N ¼ 2mþ 2, where m is anatural number. In this case, we can split Eq. (50) into twoindependent sets of equations,
Ck;k�1Nk�1 þCk;kþ1Nkþ1 ¼ 0 ðk¼ 1;3;5; . . . ;2mþ 1Þ;(62)
Ck;k�1Nk�1 þCk;kþ1Nkþ1 ¼ 0 ðk¼ 2;4;6; . . . ;2mþ 2Þ:(63)
The first set (62) contains only Nk ðk¼2;4;6;...;2mþ2Þ,and the second set (63) contains Nk ðk¼1;3;5;...;2mþ1Þ.Here, if the component Ck;k�1 is in Eq. (62), Ck�1;k ¼�Ck;k�1 must be in Eq. (63) and vice versa. Therefore, in
each set, every component Ck;k�1 appears only once. Now,
we define
Di;j ¼ C2i�1;2j; Mj ¼ N2j
ði; j ¼ 1; 2; 3; . . . ; mþ 1Þ: (64)
Note that Dij � 0 only for i� j ¼ 0, 1. Then, Eq. (62) can
be written asXj
Di;jMj ¼ 0; ði ¼ 1; 2; 3; . . . ; mþ 1Þ; (65)
which we can split into
D1;1M1 þD1;mþ1Mmþ1 ¼ 0; (66)
Di;i�1Mi�1þDi;iMi¼0; ði¼2;3; . .. ;mþ1Þ: (67)
Using Eq. (67), we can solve all ofMj ði ¼ 2; 3; . . . ; mþ 1Þin terms of M1. However, the relation between M1 andMmþ1 obtained from Eq. (67) is not the same as Eq. (66)because Eq. (67) contains no D1;1 and D1;mþ1. So, we have
to impose a constraint so that we get nontrivial Lagrangemultipliers. This is a secondary constraint expressed by
detDij ¼ 0: (68)
Under this condition, m of Mj ðj ¼ 1; 2; . . . ; mþ 1Þ are
determined, and one is left undetermined.Now, we take the latter set (63) and define
Ei;j ¼ C2i;2j�1; Wj ¼ N2j�1
ði; j ¼ 1; 2; 3; . . . ; mþ 1Þ: (69)
The same argument applies, soweget a secondary constraintdetEij ¼ 0, and one of Wj ðj ¼ 1; 2; . . . ; mþ 1Þ is left un-determined. However, matrix Eij satisfies Eij ¼ �Dji.
Hence, detEij ¼ 0 is not a new constraint. Therefore, from
Eq. (50), we get one secondary constraint detDij ¼ 0 and
two undetermined Lagrange multipliers. Then, we mustimpose a consistency condition for the secondary constraint
d
dtdetDij ¼
XNk¼1
fdetDij; CkgNk � 0; (70)
which reduces the number of undetermined Lagrangemulti-pliers from two to one.To summarize, there areN primary constraints and one
secondary constraint and we need one gauge condition.Thus, we come to the conclusion that the total number ofdegrees of freedom is
1
2ð2ð3þ 6ðN � 1ÞÞ �N � 1� 1Þ
¼ 5ðN � 1Þ þ 2þN � 2
2: (71)
Here, again, the first two terms correspond to massive andmassless gravitons, respectively. Hence, the number of BDghosts should be ðN � 2Þ=2.
C. More general diagrams
In the previous sections, we have considered tree and looptype interactions. In the case of bimetric gravity, the inter-action type is unique; namely, there is only the tree typeinteraction. In the case of trimetric gravity, there are twopossibilities, the tree and the loop type interaction. In thecase of tetrametric gravity, there are many loop type inter-actions represented by Fig. 6(a). If we cut some of theinteraction, we can make the tree type interaction and thebroom type interaction represented by Fig. 6(b). From ouranalysis, it is apparent that if the interaction contains at leasta loop, then there are ghosts. For example, the model in
WHEN IS MULTIMETRIC GRAVITY GHOST-FREE? PHYSICAL REVIEW D 86, 084052 (2012)
084052-7
Fig. 6(c) contains a ghost. Therefore, in generic cases, thereexist ghosts in multimetric gravity. The number of ghostsdepends on the interaction pattern. To construct a viablemodel, we have to eliminate all of the loop type interactions.
V. CONCLUSION
We studied multimetric gravity by combining the minis-uperspace and the Hamiltonian constraint analysis. We firstrevisited bimetric gravity and explained why it is ghost-free.This proved the validity of our method. Then, we appliedour method to trimetric gravity and clarified when the modelcontains a ghost. We proved trimetric gravity genericallycontains a ghost. However, if we cut the interaction of a pairof metrics, trimetric gravity turned out to be ghost-free. Wefurther extended the Hamiltonian analysis to general multi-metric gravity and calculated the number of ghosts in vari-ous models. Thus, we found multimetric gravity with looptype interactions never becomes ghost-free. The number of
BD ghosts in the N -metric case turned out to beðN � 1Þ=2 or ðN � 2Þ=2, depending on whether the num-ber of metrics N is odd or even. Hence, the number of BDghosts increases by one every time two more metrics areintroduced. There are other models which may contain ghostsor may not contain any ghosts. It depends on the interactiontype. The number of ghosts can be calculated once thediagram characterizing the interaction pattern is given.Admittedly, what we have investigated is BD ghosts.
There may be other ghosts depending on the solutions[26,27]. In other words, the absence of BD ghosts is anecessary condition as a healthy model. In this paper, wehave studied interaction terms consisting of only pairs ofmetrics. However, as in Ref. [22], interactions of triplets orquadruplets may be allowed. We hope to study this possi-bility in the future. It is also interesting to extend ouranalysis to various models such as higher curvature theo-ries [28] or theories containing other fields [29].
ACKNOWLEDGMENTS
This work was supported in part by the Japan Society forthe Promotion of Science (JSPS) Grant No. 24-1693, theGrant-in-Aid for Scientific Research Fund of the Ministry ofEducation, Science and Culture of Japan No. 22540274, theGrant-in-Aid for Scientific Research (A) (No. 21244033,No. 22244030), the Grant-in-Aid for Scientific Research onInnovative Area No. 21111006, JSPS under the Japan-Russia Research Cooperative Program, and the Grant-in-Aid for the Global COE Program ‘‘The Next Generation ofPhysics, Spun from Universality and Emergence.’’
APPENDIX A: INTERACTION TERMS
In this appendix, we calculate the interaction terms separately. We use the following representations:
ðg�1fÞ�� ¼ L2=N2 0
0 �il!lj
!;
ffiffiffiffiffiffiffiffiffiffiffig�1f
q¼ L=N 0
0ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
p !
: (A1)
The first one is given by
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detgp
e1ðffiffiffiffiffiffiffiffiffiffiffig�1f
qÞ ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detg
ptr
ffiffiffiffiffiffiffiffiffiffiffig�1f
q¼ ffiffiffiffiffiffiffiffiffiffi
det�p
NðL=N þ trffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qÞ ¼ ffiffiffiffiffiffiffiffiffiffi
det�p ðLþ Ntr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qÞ: (A2)
Due to the combination g�1f, we got the linear terms with respect to the lapse functions. The second one becomes
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detgp
e2ðffiffiffiffiffiffiffiffiffiffiffig�1f
qÞ ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detg
p 1
2ðtr2
ffiffiffiffiffiffiffiffiffiffiffig�1f
q� trðg�1fÞÞ ¼ ffiffiffiffiffiffiffiffiffiffi
det�p
N1
2fðL=N þ tr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qÞ2 � L2=N2 � trð��1!Þg
¼ ffiffiffiffiffiffiffiffiffiffidet�
p �Ltr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qþ 1
2Nðtr2
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q� trð��1!ÞÞ
�: (A3)
Again, we obtained desired linearity for the lapse functions. The third one can be calculated as
FIG. 6 (color online). The diagram (a) represents the mostgeneral type interaction. The diagram (b) is the broom typeinteraction. The diagram (c) includes the loop type interaction,hence there should be a BD ghost.
KOUICHI NOMURA AND JIRO SODA PHYSICAL REVIEW D 86, 084052 (2012)
084052-8
ffiffiffiffiffiffiffiffiffiffiffiffiffiffi�detgp
e3ðffiffiffiffiffiffiffiffiffiffiffig�1f
qÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffi�detg
p 1
6ðtr3
ffiffiffiffiffiffiffiffiffiffiffig�1f
q�3tr
ffiffiffiffiffiffiffiffiffiffiffig�1f
qtrðg�1fÞþ2trðg�1fÞ3=2Þ
¼ ffiffiffiffiffiffiffiffiffiffidet�
pN1
6fðL=Nþ tr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qÞ3�3ðL=Nþ tr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qÞðL2=N2þ trð��1!ÞÞþ2ðL3=N3þ trð��1!Þ3=2Þg
¼ ffiffiffiffiffiffiffiffiffiffidet�
p �1
2Lðtr2
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q� trð��1!ÞÞþ1
6Nðtr3
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q�3tr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qtrð��1!Þþ2trð��1!Þ3=2Þ
�: (A4)
This is also linear with respect to the lapse functions. The last one is
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detgp
e4ðffiffiffiffiffiffiffiffiffiffiffig�1f
qÞ ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detg
pdet
ffiffiffiffiffiffiffiffiffiffiffig�1f
q¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� detfp ¼ L
ffiffiffiffiffiffiffiffiffiffiffidet!
p: (A5)
To sum up, the interaction terms readX4n¼0
�nenðffiffiffiffiffiffiffiffiffiffiffig�1f
qÞ ¼ N
ffiffiffiffiffiffiffiffiffiffidet�
p ��0 þ �1tr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qþ 1
2�2ðtr2
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q� trð��1!ÞÞ þ 1
6�3ðtr3
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q
� 3trffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qtrð��1!Þ þ 2trð��1!Þ3=2Þ
�þ L
� ffiffiffiffiffiffiffiffiffiffidet�
p ��1 þ �2tr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q
þ 1
2�3ðtr2
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q� trð��1!ÞÞ
�þ �4
ffiffiffiffiffiffiffiffiffiffiffidet!
p �: (A6)
Notice that all interaction terms are linear in L andN. This is the advantage of the minisuperspace model, which makes theHamiltonian constraint analysis simple.
In the paper, for simplicity, we always assume that interactions are minimal [21,25], namely,
�0 ¼ 3; �1 ¼ �1; �2 ¼ 0; �3 ¼ 0; �4 ¼ 1: (A7)
Clearly, this simplification does not lose any generality concerning the ghost analysis.
APPENDIX B: CONSTRAINTALGEBRA
In this appendix, we calculate a Poisson bracket. It is sufficient to look at the following:
fCN; CLg ¼�
M2gffiffiffiffiffiffiffiffiffiffi
det�p
�1
2�i
i�jj � �ij�ij
�;�a1
ffiffiffiffiffiffiffiffiffiffidet�
p �þ��a1
ffiffiffiffiffiffiffiffiffiffidet�
ptr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q;
M2fffiffiffiffiffiffiffiffiffiffiffi
det!p
�1
2pi
ipjj � pijpij
��
¼ a1
�M2
gffiffiffiffiffiffiffiffiffiffidet�
p� ffiffiffiffiffiffiffiffiffiffi
det�p
;1
2�i
i�jj � �ij�ij
��M2
f
ffiffiffiffiffiffiffiffiffiffiffidet�
det!
s �tr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q;1
2pi
ipjj � pijpij
��: (B1)
The point is that the result is proportional to a1. Each term can be manipulated as� ffiffiffiffiffiffiffiffiffiffidet�
p;1
2�i
i�jj � �ij�ij
�¼ @
ffiffiffiffiffiffiffiffiffiffidet�
p@�mn
@
@�mn
�1
2�i
i�jj � �ij�ij
�¼ 1
2
ffiffiffiffiffiffiffiffiffiffidet�
p�mnð�mn�
ii � 2�mnÞ ¼ 1
2
ffiffiffiffiffiffiffiffiffiffidet�
p�i
i (B2)
and
�tr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q;1
2pi
ipjj � pijpij
�¼ @tr
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
p@!mn
@
@pmn
�1
2pi
ipjj � pijpij
�
¼ 1
2ðffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q�1��1Þmnð!mnp
ii � 2pmnÞ
¼ 1
2pi
itrffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
q� trð
ffiffiffiffiffiffiffiffiffiffiffiffiffi��1!
qp!Þ; (B3)
where p represents a matrix with components pmn.In this case, CNL � 0 because there is an interaction between g and f, namely, a1 � 0. Thus, whether the Poisson
bracket is nontrivial or not is determined by the interaction pattern.
WHEN IS MULTIMETRIC GRAVITY GHOST-FREE? PHYSICAL REVIEW D 86, 084052 (2012)
084052-9
[1] M. Fierz and W. Pauli, Proc. R. Soc. A 173, 211 (1939).[2] H. van Dam and M. Veltman, Nucl. Phys. B22, 397 (1970).[3] V. I. Zakharov, JETP Lett. 12, 312 (1970).[4] A. Vainshtein, Phys. Lett. 39B, 393 (1972).[5] D. Boulware and S. Deser, Phys. Rev. D 6, 3368 (1972).[6] C. de Rham and G. Gabadadze, Phys. Rev. D 82, 044020
(2010).[7] C. de Rham, G. Gabadadze, and A. J. Tolley, Phys. Rev.
Lett. 106, 231101 (2011).[8] K. Hinterbichler, Rev. Mod. Phys. 84, 671 (2012).[9] K. Koyama, G. Niz, and G. Tasinato, Phys. Rev. Lett. 107,
131101 (2011).[10] K. Koyama, G. Niz, and G. Tasinato, Phys. Rev. D 84,
064033 (2011).[11] Th. M. Nieuwenhuizen, Phys. Rev. D 84, 024038 (2011).[12] C. de Rham, G. Gabadadze, L. Heisenberg, and D.
Pirtskhalava, Phys. Rev. D 83, 103516 (2011).[13] G. D’Amico, C. de Rham, S. Dubovsky, G. Gabadadze, D.
Pirtskhalava, and A. Tolley, Phys. Rev. D 84, 124046(2011).
[14] A. E. Gumrukcuoglu, C. Lin, and S. Mukohyama, J.Cosmol. Astropart. Phys. 11 (2011) 030.
[15] C. Isham, A. Salam, and J. Strathdee, Phys. Rev. D 3, 867(1971).
[16] T. Damour, I. I. Kogan, and A. Papazoglou, Phys. Rev. D66, 104025 (2002).
[17] T. Damour and I. I. Kogan, Phys. Rev. D 66, 104024(2002).
[18] S. Hassan, R.A. Rosen, and A. Schmidt-May, J. HighEnergy Phys. 02 (2012) 026.
[19] S. Hassan and R.A. Rosen, J. High Energy Phys. 04(2012) 123.
[20] S. Hassan and R.A. Rosen, J. High Energy Phys. 02(2012) 126.
[21] N. Khosravi, N. Rahmanpour, H. R. Sepangi, and S.Shahidi, Phys. Rev. D 85, 024049 (2012).
[22] K. Hinterbichler and R.A. Rosen, J. High Energy Phys. 07(2012) 047.
[23] S. Hassan, A. Schmidt-May, and M. von Strauss,arXiv:1204.5202.
[24] R. L. Arnowitt, S. Deser, and C.W. Misner, arXiv:gr-qc/0405109.
[25] S. Hassan and R.A. Rosen, J. High Energy Phys. 07(2011) 009.
[26] A. De Felice, A. E. Gumrukcuoglu, and S. Mukohyama,arXiv:1206.2080.
[27] A. E. Gumrukcuoglu, C. Lin, and S. Mukohyama,arXiv:1206.2723.
[28] M. F. Paulos and A. J. Tolley, J. High Energy Phys. 09(2012) 002.
[29] Q.-G. Huang, Y.-S. Piao, and S.-Y. Zhou,arXiv:1206.5678.
KOUICHI NOMURA AND JIRO SODA PHYSICAL REVIEW D 86, 084052 (2012)
084052-10