You are on page 1of 16

G8099 – Selected Topics in Gravity 1

General Relativity from


Lorentz Invariance

1 The problem with masslessness


1.1 Massive particles
Roughly speaking, the state of a massive spin J particle is identified by its momentum p and by
the value of Jz , which is an integer between −J and J. A massive particle thus have 2J +1 distinct
spin states. In fact we can choose to parameterize the particle’s spin state in its rest-frame, where
rotations act like in non-relativistic QM. But to appreciate the peculiarities we will encounter
for massless particles, we have to be more precise than this. I will follow Weinberg’s treatment,
specifically sects. 2.5, 5.3, and 5.9 of ref. [1], with a somewhat lighter notation.

1.1.1 Single-particle states


First of all, for a Lorentz-invariant theory the space of single-particle states for a given particle
species must be the basis for an irreducible representation of the Lorentz group. Indeed, if the
representation were reducible, there would be different states that cannot be brought into one
another by any Lorentz transformation. In this case it is natural to talk of different particle
species rather than of different states of the same particle. Now, in the space of single-particle
states, consider an eigenstate of the four-momentum operator with eigenvalue pµ , and labeled by
some other internal degree of freedom (as we will see, the ‘spin’) σ,

|
p, σ . (1)

Notice that p is enough to determine the full pµ , for a given mass m. We want to see how a
Lorentz-transformation Λ acts on such a state. It is convenient to decompose Λ into a rotation in
the particle’s rest frame and everything else. Namely

Λ = L(Λp) · W (Λ, p) · L−1 (p) , (2)

where:

• L(p) is a standard Lorentz transformation that brings the particle from rest to momentum
p. That is, if k µ = (m, 0) is the particle’s four momentum in its rest frame, we need

Lµ ν (p)k ν = pµ . (3)

There are infinitely many such Lorentz transformations—but we have to choose a standard
one, once for all. For instance, we can pick a boost along p.
2 General Relativity from Lorentz Invariance

• W (Λ, p) is defined as
W (Λ, p) ≡ L−1 (Λp) · Λ · L(p) , (4)
and acts trivially on the rest-frame four momentum k µ = (m, 0):
 µ
W (Λ, p)µν k ν = L−1 (Λp) α (Λp)α = k µ , (5)

by definition of L.

Therefore, W (Λ, p) belongs to the little group of k µ , which is defined as the subgroup of Lorentz
transformations under which k µ is invariant. This is the group of 3D rotations, SO(3), because
k µ only has a non-vanishing time component: rotations act trivially on k µ , whereas a boost along
any direction would change the time component of k µ by a γ factor, and also create some spatial
component. W (Λ, p) is called the “Wigner rotation” associated with Λ and p.
Now, applying the generic Lorentz transformation (2) to our single particle state |p, σ we get
     
p, σ = U L(Λp) U W (Λ, p) U −1 L(p) |
U(Λ)| p, σ . (6)

We define the extra label σ as being unchanged by our standard Lorentz boosts L(p). More
precisely, we define the state |
p, σ as just the boosted version of the zero-momentum state |0, σ,
 
|
p, σ ≡ U L(p) |0, σ (7)

In a sense, this is saying that the internal degree of freedom σ is defined in the rest frame of the
particle, and its physical meaning for generic particle momenta may not be obvious. For instance,
in general it will not be the eigenvalue of Jz for p = 0. With this definition eq, (6) becomes
   
p, σ = U L(Λp) U W (Λ, p) |0, σ .
U(Λ)| (8)
 
W (Λ, p) is an element of the little group of k µ = (m, 0), so U W (Λ, p) will not change the
 
momentum of |0, σ—it will only act on the internal label σ. More precisely, U W (Λ, p) will be
a unitary operator in an irreducible representation of the little group of k µ , which is SO(3). Once
again, the representation has to be irreducible because otherwise there would be single-particle,
zero-momentum states that cannot be brought into each other by a rotation, in which case we
would talk about different particle species rather than different states. But from non-relativistic
QM, we know the irreducible representations of SO(3) (or of its universal covering SU(2)): they
are labeled by the total spin J = 0, 12 , 1, . . . , and each is 2J + 1 dimensional.
 
Therefore, for a spin J particle U W (Λ, p) when applied to |0, σ, acts as a (2J + 1) × (2J + 1)
unitary matrix Dσσ (Λ, p) on the spin label:
  
U W (Λ, p) |0, σ = Dσσ (Λ, p) |0, σ   . (9)
σ
G8099 – Selected Topics in Gravity 3

For instance we can define σ as the eigenvalue of Jz in the particle’s rest frame, and then Dσσ (Λ, p)
tells us how the Wigner rotation W (Λ, p) mixes different eigenstates. Then eq. (8) reduces to

U(Λ)| p, σ =  σ
Dσσ (Λ, p) |Λp, (10)
σ

The transformation law for states uniquely determines the transformation properties of the
creation and annihilation operators under Lorentz transformations. Indeed, consider the creation
operator a†σ (
p) for a state of momentum p and spin-state σ,

p, σ = a†σ (
| p)|0 . (11)

Applying a Lorentz transformation we have


   
U(Λ)|p, σ = U(Λ) a†σ (
p) U −1 (Λ) U(Λ)|0 = U(Λ) a†σ (
p) U −1 (Λ) |0 , (12)

where in the last step we used that the vacuum |0 is Lorentz invariant. To satisfy eq. (10) we
must have 
U(Λ) a†σ (
p) U −1 (Λ) = Dσσ (Λ, p) a†σ (Λp)
 (13)
σ

Taking the Hermitian conjugate of this, and using the unitarity of U(Λ), i.e. U † = U −1 , we get
the transformation law for the annihilation operator aσ (
p):

p) U −1 (Λ) =
U(Λ) aσ ( ∗
Dσσ 
 (Λ, p) aσ  (Λp) (14)
σ

1.1.2 The field operator


Now, to describe our particles with a local covariant (scalar, spinor, vector, etc.) field operator, we
need a finite-dimensional representation of the Lorentz group to use as polarization spinor, vector,
etc. For instance, we know that to describe a spin-1 particle we need polarization four-vectors eµ .
Indeed, the spin-1 representation of SO(3) is that of three-dimensional vectors, with the three Jz
eigenstates being

e−1 = √1 (1, −i, 0) (15)


2
e0 = (0, 0, 1) (16)
e+1 = − √12 (1, +i, 0) (17)

So, in the rest-frame we define the polarization four-vector for a particle of polarization σ simply
as
eµσ ≡ (0; eσ ) , (18)
because a non-vanishing time-component would not transform under spatial rotations, thus making
the representation of SO(3) reducible. Then, for generic particle momenta p, the polarization four
4 General Relativity from Lorentz Invariance

vector is defined by applying the standard Lorentz transformation L(p) to the zero-momentum
polarization four-vector:
eµσ (p) ≡ Lµ ν (p) eνσ . (19)
For all momenta p we have
eµσ (
p) pµ = 0 , σ = −1, 0, +1 , (20)
since in the rest frame k µ = (m; 0) and eµσ = (0; eσ ), and we can reach any other momentum via
a suitable L(p). Now, as we will see, it is crucial that the rest-frame polarization four-vectors eµσ
belong to the same representation of the little group of k µ as the single particle states they are
supposed to describe. That is, we want that under a Wigner rotation W (Λ, p)

W (Λ, p)µ ν eνσ = Dσσ (Λ, p) eµσ (21)
σ

with exactly the same matrix Dσσ appearing in (9). Then, under a generic Lorentz transformation
Λ and for generic momentum p,
 µ 
Λµ ν eνσ (
p) = L(Λp) · W (Λ, p) · L(p) ν eνσ (
p) = Dσσ (Λ, p) eµσ (Λp)
 (22)
σ

For what follows, it is convenient to rewrite this by moving the Λ matrix to the r.h.s., and the D
matrix to the l.h.s.:    µ
D −1 σσ (Λ, p) eµσ (p) = Λ−1 ν eνσ (Λp)
 . (23)
σ

Recall that D(Λ, p) by definition is a unitary matrix,


   
D −1 σσ
= D † σσ = Dσ∗  σ . (24)

Therefore   µ
Dσ∗  σ (Λ, p) eµσ (
p) = Λ−1 ν eνσ (Λp)
 . (25)
σ

Out of the polarization four-vectors and of the creation and annihilation operators a†σ (
p), aσ (
p)
for particles of momentum p and polarization σ, we can construct the local field operator associated
with our particles:

µ 1  d3 p  µ ip·x µ∗ † −ip·x

V (x) ≡  eσ (
p)aσ (p) e + es (p)as (
p) e (26)
(2π)3/2 σ 2p0

(we are assuming for simplicity that our spin-1 particle is its own antiparticle). Given the trans-
formation properties of the polarization four-vectors, the field operator V µ (x) transform as a
G8099 – Selected Topics in Gravity 5

four-vector under a generic Lorentz transformation Λ:



µ −1 1  d3 p  µ  −1
 ip·x 
U(Λ)V (x)U (Λ) =  eσ (
p ) U(Λ)a σ (
p)U (Λ) e + h.c. (27)
(2π)3/2 σ 2p0

1  d3 p  µ ∗ ip·x

=  eσ (
p ) Dσσ  (Λ, p) a σ  ( 
Λp) e + h.c. (28)
(2π)3/2  2p0
σ,σ

1  −1 µ  d3 p  µ   eip·x + h.c.

= Λ ν  eσ  ( Λp)a σ  (Λp) (29)
(2π)3/2 σ
2p0
 µ
= Λ−1 ν V ν (Λx) , (30)

where in the first step we used that everything in V µ (x) is a c-number but a and a† , which are
thus the only factors that do not commute with U and U −1 ; in the second step we used the
transformation law for aσ (p), eq. (14); in the third step we reabsorbed the matrix D ∗ into the
polarization four-vector, via eq. (25); finally, to get the last line we changed integration variable
to p = Λp: the measure is Lorentz invariant, and as to the exponential we have eip·x = ei(Λp)·(Λx) .
To summarize: when we apply a Lorentz transformation U(Λ) to the field-operator, the only
objects that transform non-trivially are the creation and annihilation operators, which inherit the
transformation properties of the single-particle states they create or destroy. Since the polarization
four-vectors, when transformed covariantly according to the Lorentz index they carry, behave
exactly in the same way as single particle states when acted upon by U(Λ) (that is: p changes to
Λp, and the spin label transforms with the D matrix), then we can reabsorb the transformation of
the creation and annihilation operators into the polarization four-vectors via a Lorentz-covariant
transformation of the latter. We are then left with a Lorentz matrix acting on the field-operator,
with the usual Lorentz contraction of indices.
The fact that the field operator transforms covariantly under Lorentz transformations, ensures
that in order to construct a Lorentz invariant theory where our spin-1 particles interact with
others and among themselves, it is enough to write down an interaction Lagrangian where V µ
always appears in Lorentz invariant contractions, like (Vµ V µ )2 etc. That is, it is enough that
the interaction Lagrangian looks Lorentz-invariant. This sounds trivial, but as we will see in the
following, for massless particles this is not the case.

1.2 Massless particles


1.2.1 Single particle states
Massless particles have no rest-frame—thus there is no reference frame where we can apply our
non-relativistic QM knowledge about rotations and spin. To see what the ‘spin’ degrees of freedom
are in this case, it is convenient to perform a Lorentz transformation to align the particle’s spatial
momentum with the z-axis, and bring it to some reference energy k:

k µ = (k, 0, 0, k) . (31)
6 General Relativity from Lorentz Invariance

This will be our reference four-momentum—the analogue of the rest-frame four-momentum in the
massive case.
We are interested in the irreducible representations of the little group of k µ , because any
particle species must fall into one of them. By definition the little group of k µ is the subgroup
of Lorentz transformations that leave k µ invariant. Clearly the little group contains at least the
group of rotations around the z-axis, which is SO(2) = U(1). This is an abelian group, and as
such it only has one-dimensional irreducible representations. These are labeled by the helicity
h ≡ Jz = J · k̂ , (32)
which can take the values 0, ± 12 , ±1, etc. 1 Being the irreducible representations of SO(2) one-
dimensional, a massless particle just has one possible value of the helicity h. However, from its
definition it’s clear that the helicity is a pseudo-scalar—it is odd under space-inversions. Since both
the electromagnetic and the gravitational interactions conserve parity, it is customary to assign
the photon and the graviton two different helicity states each, h = ±1 and h = ±2 respectively,
although technically speaking every value of h can be thought of as labeling a different particle
species.
Now, we said ‘the little group contains . . . ’, because there is something more, which is going
to be the culprit for us. Besides SO(2), there are other Lorentz transformations that leave k µ
invariant. These are boosts in the xy plane, composed with a rotation that brings the particle’s
momentum back along the z axis, composed with a boost along z that brings the particle’s energy
back to k. For instance under a boost along x we have
k µ → Λ(x) µ ν k ν = (γk, γβ k, 0, k) ≡ k µ . (33)
We can bring back the momentum to point along the z-axis through a rotation around the y-axis
with cos θ = 1/γ, sin θ = β. We get
k µ → R(y) µ ν k ν = (γk, 0, 0, γk) ≡ k µ . (34)
γ 2 +1
Then, with a boost along −z with γ̃ = 2γ
we bring the particle’s energy back to k,
k µ → Λ(−z) µ ν k ν = (k, 0, 0, k) = k µ . (35)
The full Lorentz transformation we just performed, from k µ back to k µ is
Λ = Λ(−z) · R(y) · Λ(x) = (36)
 1  
2
(3 − 1/γ 2 ) 1 − 1/γ 2 0 − 12 (1 2
− 1/γ )
 
   
 1 − 1/γ 2 1 0 − 1 − 1/γ 2 
 
= 


 (37)
 0 0 1 0 
 
 
1 2
 1
2
(1 − 1/γ ) 1 − 1/γ 2 0 2
(1 + 1/γ 2 )
1
The helicity is quantized in units of 1/2, but not quite for the same reason as the the angular momentum is
quantized. See ref. [1], sect. 2.7 for details.
G8099 – Selected Topics in Gravity 7

In general, if we start with a boost along a generic direction in the xy plane, the complete Lorentz
transformation that leaves k µ invariant is conveniently parameterized as (see ref. [1], sect. 2.5)
 
1 + ζ α β −ζ
 α 1 0 −α 
Λ(α, β)µν = 
 β
 (38)
0 1 −β 
ζ α β 1−ζ

with ζ ≡ (α2 + β 2 )/2, and α and β generic real numbers. Our example above corresponds to
α = 1 − 1/γ 2 , β = 0.
In summary, a generic element of the little group is the composition of a Lorentz transformation
(38) and of an SO(2) rotation. At the infinitesimal level:

Λ = 1 + iαA + iβB + iθJ3 , (39)

where A and B are the generators of a transformation like eq. (38),


   
0 −i 0 0 0 0 −i 0
 −i 0 0 i   0 0 0 0 
Aµ ν = 
 0 0 Bµν =   (40)
0 0   −i 0 0 i 
0 −i 0 0 0 0 −i 0

and J3 is the generator of SO(2) rotations,


 
0 0 0 0
 0 0 −i 0 
(J3 )µ ν =
 0
 (41)
i 0 0 
0 0 0 0

To find the representations of the little group, it is instructive to look at the algebra:
     
J3 , A = +iB , J3 , B = −iA , A, B = 0 . (42)

A and B commute with each other, and are rotated into each other by SO(2) rotations. This is the
algebra of two-dimensional rotations and translations. The corresponding group is called ISO(2).
Now, the problem with this group is that if the massless particle’s state transform non-trivially
under A or B, say it is an eigenvector of both with eigevalues a and b respectively 2 , then by
applying rotations we discover a continuum of eigenvectors with eigenvalues (a cos θ − b sin θ) and
(a sin θ + b cos θ). To see this, just make use of the analogy with 2D rotations and translations.
A and B generate 2D translations, that is they are the 2D momentum operators. If we have
an eigenvector of the 2D momentum, then by rotational invariance all rotated versions of it are
also eigenvectors of the 2D momentum, with different eigenvalues—that is, different x and y
2
Since A and B commute, we can diagonalize them simultaneously.
8 General Relativity from Lorentz Invariance

components. Now, this is exactly what happens with ordinary momentum, and there it is no
problem at all. Here however we have already completely fixed the four-momentum, and particles
are not observed to have any other continuous degree of freedom. We therefore postulate that
massless particles’ states transform trivially under the non-SO(2) part of the little group, i.e. they
are annihilated by A and B:
A|ψ = 0 , B|ψ = 0 . (43)
We are thus left with the SO(2) part of the little group, whose representations are labeled by
the helicity. Since we are back where we started from, the detour we just took may sound like
pointless technical complications, but as we will see in a moment precisely these complications are
the cause of gauge- and diffeomorphism-invariance.
In summary, for a massless particle with given helicity h, there is just one single-particle state
with momentum k µ = (k, 0, 0, k), which we call |k. Given a standard Lorentz transformation
L(p) that brings the particle’s four-momentum from k µ to pµ ,

L(p)µ ν k ν = pµ , (44)

we define the state with momentum pµ as


 
p ≡ U L(p) |k .
| (45)

We decompose a generic Lorentz transformation Λ as we did in the massive case,

Λ = L(Λp) · W (Λ, p) · L−1 (p) , (46)

where W = L−1 (Λp) · Λ · L(p) is an element of the little-group of k µ . When acting on |p we get
   
U(Λ)|p = U L(Λp) U W (Λ, p) |k . (47)

W (Λ, p) will be the combination of an SO(2) rotation and of a Λ(α, β) transformation. The latter
acts trivially on |k, whereas under the former we just get a phase

W (Λ, p)|k = eihθ |k , (48)

by definition of helicity. Here θ is the SO(2) rotation angle, which depends on Λ and on p, because
W does. In fact we can apply exactly the same formulae that we had in the massive case, by
suppressing the spin variable and replacing everywhere the Wigner rotation matrix Dσσ with the
phase eihθ . We have
U(Λ)|p = eihθ |Λp
 (49)
and

U(Λ) a† (
p) U −1 (Λ) = eihθ a† (Λp)
 (50)
U(Λ) a( p) U −1 (Λ) = e−ihθ a(Λp)
 (51)
G8099 – Selected Topics in Gravity 9

1.2.2 The photon field operator


Now we try to construct a covariant field operator, like we did for the massive case. In order
to do so, we need covariant objects that describe the particle’s polarization state, like the eµ ’s
for spin-1 massive particles. For the two photon helicities h = ±1, we need two polarization
four-vectors. We stress again that the two photon helicities do not transform into each other
under proper Lorentz transformations. Therefore, it is just for matter of convenience that we
treat them simultaneously. In the standard reference frame where the photon’s four momentum
is k µ = (k, 0, 0, k), the polarization four-vectors must be eigenvectors of Jz , with eigenvalues ±1.
They are
eµ±1 = √12 (0, 1, ±i, 0) . (52)
As before, for generic momentum p we define the two polarization four-vectors as

eµ±1 (p) = Lµ ν (
p) eν±1 , (53)

where L(p) is our standard Lorentz trasformation that brings the particle’s momentum from
k = (0, 0, k) to p. In particular, we can take this transformation to be a boost along z that
makes the energy of the particle equal to |p|, composed with a rotation that makes the particle’s
momentum point along p̂. Since the original eµ ’s are purely spatial four-vectors with vanishing
z-component, they are unaffected by the boost. Then the rotation shuffles the spatial components
among themselves, but does not generate a time-component. Therefore

e0±1 (
p) = 0 . (54)

Also, in the ‘standard’ frame we have kµ eµ±1 = 0, which is a Lorentz invariant relation. Since e0s
vanishes in all frames, we have
p · e±1 (p) = 0 . (55)
Let’s then define the photon field operator as

1   d3 p  µ µ∗ †

µ ip·x −ip·x
A (x) ≡  eh (
p )ah (
p ) e + eh (
p )ah (
p ) e (56)
(2π)3/2 h=±1 2p0

Given eqs. (54, 55), in all reference frames we have

A0 (x) = 0 , ∇
 · A(x)
 =0. (57)

The vanishing of A0 in all reference frames shows that, despite the appearance of a Lorentz index,
Aµ cannot transform as a four-vector under Lorentz boosts. What went wrong?
The problem resides in the infamous part of the little group we discarded, eq. (38). Under
such transformations, our polarization four-vectors (52) in the standard reference frame transform
non-trivially:
Λ(α, β)µν eν±1 = eµ±1 + (α ± iβ) √12k k µ , (58)
10 General Relativity from Lorentz Invariance

whereas we postulated that our particle’s states are invariant. Likewise, consider the polarization
four-vectors eµ±1 (
p) in a generic frame, eq. (53). Under a generic Lorentz transformation Λ we
have  µ
Λµ ν eν±1 (
p) = L(Λp) · W (Λ, p) · L−1 (p) ν eν±1 (p) (59)
W (Λ, p) is a little group element—the combination of an SO(2) rotation, under which eµh trans-
forms according to its helicity h, and of a Λ(α, β) transformation, under which eµh shifts like in
eq. (58). We thus get
 
p) = e±iθ eµ±1 (Λp)
Λµ ν eν±1 (  + (α ± iβ) √1 (Λp)µ ,
2k
(60)
where θ, α, and β depend on Λ as well as on p. Rewriting this like for the massive case (see eq. (25)),
we see that a rotation phase e−ihθ can be reabsorbed into a covariant Lorentz-tranformation of
eν±1 (
p) plus a shift proportional to pµ ,
 µ
e∓iθ eµ±1 (
p) = Λ−1 ν eν±1 (Λp)
 + (α ± iβ) √1 pµ
2k
(61)
On the other hand, the creation and annihilation operators we used to define the field operator
above, inherit the transformation properties of the states they create or destroy, that is, they are
invariant under the special little-group transformations, and they just get a phase under the SO(2)
rotation—see eqs. (50, 51). Now, when we apply a generic Lorentz transformation Λ to the field
operator,
Aµ (x) → U(Λ)Aµ (x)U −1 (Λ) , (62)
everything commutes with U and U −1 but the operators ah (p) and a†h ( p). They transform ac-
cording to eqs. (50, 51): their momentum gets changed from p to Λp and they pick up a phase.
Like in the massive case, the phase can be absorbed into a Lorentz-covariant transformation of
the polarization vector, but now we have to pay an additional price: a shift proportional to pµ , as
in eq. (61). As a consequence, the field operator transforms as
 µ
U(Λ)Aµ (x)U −1 (Λ) = Λ−1 ν Aν (Λx) + ∂ µ λΛ (x) , (63)
where λΛ (x) is some Λ-dependent linear combination (i.e., integral in d3 p with p-dependent coef-
ficients) of creation and annihilation operators whose explicit form we are not interested in. In
conclusion: the field operator that we have defined does not transform covariantly under Lorentz
tranformations: Aµ is not a four-vector. Of course, the problem is that the polarization vectors
do not belong to the same representation of the little group of k µ as the single-particle states they
are supposed to describe.

1.2.3 The graviton field operator


With minor modifications, the same remarks apply to massless spin 2 particles, i.e. gravitons.
Here we need two polarization tensors to describe the two helicity states. In the standard frame
where the momentum is k µ = (k, 0, 0, k) we can take
eµν µ ν
±2 = e±1 e±1 , (64)
G8099 – Selected Topics in Gravity 11

where eµ±1 are the photon polarization four-vectors defined above. Since the latter are eigenstates
of rotations around the z-axis with helicity ±1, the eµν
±2 have helicity ±2,

Rz (θ)µ ρ Rz (θ)ν σ eρσ


±2 = e
±2iθ µν
e±2 , (65)

as desired. Explicitly they are


 
0 0 0 0
1  0 1 ±i 0 
eµν =   . (66)
±2
2  0 ±i −1 0 
0 0 0 0

Notice that in terms of the standard ‘plus’ and ‘cross’ polarizations for gravitational waves, they
read eµν 1 µν µν
±2 = 2 (e+ ± i e× ). They correspond to circularly polarized gravitational waves. For generic
momentum p, the polarization tensors are just the Lorentz-transformed versions of these,

eµν p) ≡ Lµ ρ (p) Lν σ (p) eρσ


±2 (
µ
p)eν±1 (p) ,
±2 = e±1 ( (67)

where L(p) is our standard Lorentz transformation that brings the graviton’s momentum from k
to p.
Next, we naively construct the field operator as we did before

1   d3 p  µν µν ∗ †

µν ip·x −ip·x
h (x) ≡  eh (
p )ah (
p ) e + eh (
p)ah (
p ) e , (68)
(2π)3/2 h=±2 2p0

and as before we realize that this is not a good Lorentz-tensor. Indeed from eq. (67) and from the
properties of the photon polarization vectors (54, 55) we see that

hµ0 = 0 , ∂i hij (x) = 0 (69)

in all reference frames—clearly a non Lorentz-invariant set of constraints if we assume that hµν
is a tensor. Once again, the culprit are the special little-group transformations under which the
particle’s state are supposed to be invariant, but under which the polarization tensors (67) in
fact transform. Since our polarization tensors are just products of the photon’s helicity polariza-
tion four-vectors, we can use the results of last section. For generic p, under a generic Lorentz
transformation we have (see eq. (60))
  
Λµ ρ Λν σ eρσ p) = e±2iθ eµ±1 (Λp)
±2 (
 + (α ± iβ) √1 (Λp)µ eν (Λp)
2k ±1
 + (α ± iβ) √1 (Λp)ν
2k

±2iθ µν µ ν ν µ

= e e±2 (Λp) + (Λp) v± + (Λp) v± , (70)
µ
where v± are four-vectors that depend on Λ and on p:
µ
v± ≡ (α ± iβ) √12k eµ±1 (Λp)
 + 1
4k 2
(α ± iβ)2 (Λp)µ . (71)
12 General Relativity from Lorentz Invariance

If we bring the Λ matrices to the r.h.s. and the phase to the l.h.s. we get

e∓2iθ eµν p) = (Λ−1 )µ ρ (Λ−1 )ν σ eρσ


±2 (
 µ ν ν µ
±2 (Λp) + p ṽ± + p ṽ± , (72)

where ṽ± = Λ−1 · v± . Then, like for the photon field, if we now apply a generic Lorentz trans-
formation Λ to the field operator hµν , the rotation phases e±ihθ that we get from transforming
the creation and annihilation operators can be absorbed into a covariant Lorentz transformation
µ
of the polarization tensors if we also shift the latter by pµ ṽ±
ν
+ pν ṽ± . As a consequence, the field
operator transforms as

U(Λ)hµν (x)U −1 (Λ) = (Λ−1 )µ ρ (Λ−1 )ν σ hρσ (Λx) + ∂ µ ξΛν (x) + ∂ ν ξΛµ (x) , (73)

where ξΛµ (x) is a Λ-dependent, four-vector linear combination of creation and annihilation opera-
tors.

2 Gauge Invariance from Lorentz Invariance


Since—as we just found out—the photon and graviton field operators do not transform covariantly
under Lorentz transformations, we cannot just stick them in a Lorentz invariant-looking Lagrangian
to get a truly Lorentz-invariant theory. For instance, a photon quartic self-interaction like (Aµ Aµ )2
looks Lorentz-invariant, but given the transformation properties of Aµ (63), it isn’t. However,
eq. (63) itself provides us with the solution to the problem. If, besides looking Lorentz-invariant,
the Lagrangian is invariant under the formal replacement

Aµ (x) → Aµ (x) + ∂µ λ(x) , (74)

for a generic function λ(x), then the extra piece in the Lorentz-transformation of Aµ is harmless,
and the Lagrangian is Lorentz invariant. For instance, out of the photon field we can construct
the ‘electromagnetic field-strength’ operator

Fµν ≡ ∂µ Aν − ∂ν Aµ . (75)

This is unaffected by the replacement (74), and, as a consequence, it is a true Lorentz-tensor

Fµν (x) → U(Λ)Fµν (x)U −1 (Λ) = Λµ ρ Λν σ Fρσ (Λ · x) . (76)

This means that now we can use Fµν in a Lagrangian in the usual way, just by contracting indices
in a Lorentz-invariant fashion. For instance, the quartic term (Fµν F µν )2 now describes a truly
Lorentz-invariant photon self-interaction. In addition to using Fµν , we can couple directly Aµ to a
conserved ‘current’ J µ (x)—that is, a (genuine) four-vector made up of other fields (electron, etc.)
that obeys a local conservation equation

∂µ J µ = 0 . (77)
G8099 – Selected Topics in Gravity 13

If the Lagrangian term that describes the interaction between Aµ and these other fields is

Lint = Aµ (x)J µ (x) , (78)

then upon the replacement (74) this Lagrangian term in unchanged:

Lint → Lint + ∂µ λ J µ = Lint − λ ∂µ J µ = Lint , (79)

where we integrated by parts, and we used the conservation of J µ . As a result, Lint describes a
Lorentz-invariant interaction.
Likewise, for the graviton field hµν we get a Lorentz-invariant theory if the Lagrangian has all
indices contracted in a Lorentz invariant fashion and is also invariant under the replacement

hµν (x) → hµν (x) + ∂µ ξν (x) + ∂ν ξµ (x) , (80)

for a generic four-vector field ξµ (x). This way, the non-covariant part in the transformation of
hµν (73) has no effect on the Lagrangian, which then is Lorentz-invariant. Like for the photon,
we can construct a true Lorentz-tensor by taking suitable derivatives of hµν . The simplest is the
(linearized) ‘Riemann tensor’ (see next section):

Rµνρσ ≡ 12 (∂µ ∂ρ hνσ − ∂ν ∂ρ hµσ − ∂µ ∂σ hνρ + ∂ν ∂σ hµρ ) . (81)

It is straightforward to check that this is unaffected by eq. (80). Then, it is a true tensor, for the
non-covariant pieces in the Lorentz transformation of hµν cancel among different terms. We can
use the Riemann tensor to build Lorentz invariant interactions for the graviton. Alternatively,
just like we did for the photon, we can couple hµν directly to a conserved symmetric tensor T µν :
a symmetric true Lorentz tensor made up of other fields—as well as of the graviton field itself, as
we will see—obeying the conservation law

∂µ T µν = 0 . (82)

The interaction Lagrangian


Lint = hµν (x)T µν (x) (83)
is invariant under the replacement (80),

Lint → Lint + 2 ∂µ ξν T µν = Lint − 2 ξν ∂µ T µν = Lint , (84)

and is therefore Lorentz-invariant.

3 Intermezzo: Remarks on gauge invariance


Of course invariance under the formal replacements (74, 80) is what usually deserves the name of
‘gauge invariance’. However from the constructive viewpoint we have taken here, it is clear that
14 General Relativity from Lorentz Invariance

gauge invariance is not a symmetry on which we have any choice: it is forced upon us by combining
quantum mechanics and Lorentz-invariance. In particular, if we want to describe massless particles
through local field operators like Aµ (x) and hµν (x), then the theory is truly Lorentz invariant
only if it is gauge-invariant—because these field operators do not transform covariantly under
Lorentz transformations. However gauge-invariance is not a symmetry in any physical sense:
recall that the non-covariant piece in the transformation of Aµ and hµν stems from the special
little-group transformations under which the particles’ states are invariant. This means that a
gauge transformation like (74) or (80) is not associated with any transformation in the particles’
Hilbert space. Therefore, unlike ordinary symmetries, gauge-invariance does not yield any non-
trivial relation between different observables like e.g. different scattering amplitudes—because it
acts trivially on the particles’ states!
Gauge invariance in a sense is a redundancy in our description of massless particles: we have
a class of transformations under which the field operators transform, but the states don’t. It had
better be that the different field configurations that we span by applying such transformations
are physically all equivalent—that they describe the same state of the system—which we achieve
by demanding that the Lagrangian be invariant under those transformations. This redundancy
can be seen already at the level of the photon polarization four-vectors e±1µ (
p). Under a special
little-group tranformation, these transform as in eq. (60). But the photon’s state is invariant
under any such transformation. Therefore e±1 µ (
p) and

e±1
µ (
p) + c pµ (85)

must describe the same photon state, for any (complex) c. When we compute a scattering am-
plitude involving an incoming (or outgoing) photon, Feyman’s rules instruct us to contract the
p) (or e∗µ (
photon’s polarization eµ ( p)) with the rest of the diagram, which is a four-vector built
out of the the other particles’ momenta and spin degrees of freedom:

M = M µ (rest) eµ (p) . (86)

Now, the fact that eµ (


p) and eµ (
p) + c pµ describe the same photon state, tells us that they must
yield the same scattering amplitude M when contracted with M µ . This is the case if and only if

M µ pµ = 0 . (87)

The argument straightforwardly generalizes to the graviton case. The graviton polarization
tensor (67) is the product of two photon polarization vectors. Under a special little-group trans-
formation the latter both shift by a term proportional to pµ . On the other hand the graviton state
is invariant. Hence, e±1 p)e±1
µ ( ν (p) and

(e±1 p) + c pµ )(e±1
µ ( ν (
p) + c pν ) (88)

must describe the same graviton state, for all c’s. To compute an amplitude in which a graviton
is absorbed (or emitted), we have to contract the graviton polarization tensor with the rest of the
G8099 – Selected Topics in Gravity 15

diagram, which now is a symmetric two-index tensor made up of the other particles’ momenta
and polarizations:
M = M µν (rest) eµν (
p) . (89)
Since eµν = eµ eν and (eµ + c pµ )(eν + c pν ) describe the same graviton state, they must produce
the same amplitude M when contracted with M µν . For infinitesimal c, this is only possible if

M µν pµ = 0 . (90)

Eqs. (87, 90) are the manifestation of the gauge-redundancy in scattering amplitudes. They
teach us that for any process in which a photon or a graviton is emitted or absorbed, the amplitude
must vanish if we formally replace the emitted or absorbed particle’s polarization with its four-
momentum. Notice however that—as we will see in the next sections—in general this involves a
cancellation between different Feynman diagrams contributing to the same amplitude. That is,
eqs. (87, 90) only hold for the full amplitude, not at the level of single Feynman diagrams.
We could dispose of gauge invariance altogether, if in our Lagrangian we never used Aµ and
hµν alone, but only used truly covariant field operators like Fµν and Rµνρσ . In fact these are true
Lorentz tensors precisely because they are gauge invariant—so that all non-covariant pieces in
the Lorentz-trasformation law of Aµ and hµν cancel. By using true tensor we would not need to
introduce gauge tranformations at all. Any Lorentz invariant-looking Lagrangian would be truly
Lorentz-invariant. Put another away: with gauge-invariant fields, and gauge-invariant states,
gauge transformations would act trivially on everthing, so why bother introducing them?
It is useful to go back to the field expansion of Aµ (56), and see what it implies for Fµν ,

Fµν = ∂µ Aν − ∂ν Aµ (91)
1   d3 p   
s s ip·x
=  p) − ipν eµ (
ipµ eν ( p) as (
p) e + h.c. . (92)
(2π)3/2 s=±1 2p0

That is, Fµν describes a photon of helicity ±1 and momentum p through the polarization tensor

ipµ e±1 p) − ipν e±1


ν ( µ (
p) . (93)

This is obviously an eigenstate of rotations around p with helicity ±1, because the e±1
µ (
p) is, and
pµ is invariant. However now this polarization tensor is invariant under the infamous special little
group transformation (60)—the unwanted terms in the transformation proportional to pµ cancel
between the two terms in (93). Hence, all ingredients that make up Fµν transform as they should,
and as a result Fµν is a true tensor. Likewise, the field operator Rµνρσ employs the polarization
tensors   
ipµ e±1 p) − ipν e±1
ν ( p) ipρ e±1
µ ( p) − ipσ e±1
σ ( ρ (
p) , (94)
which have helicity ±2, and are invariant under the transformation (60). Hence, Rµνρσ transforms
as a tensor under Lorentz transformations 3 .
3
Actually eq. (94) is a good mnemonic to reconstruct the tensor structure of the Riemann tensor in terms of
16 General Relativity from Lorentz Invariance

Why not use the true tensors Fµν and Rµνρσ as the field operators for photons and gravitons,
then? The problem is that the explicit powers of pµ appearing in eqs. (93, 94) would make any
amplitude for emitting (or absorbing) a photon or a graviton proportional to some powers of the
energy of the emitted (or absorbed) particle. This becomes small at low energies, or, equivalently,
at large distances. For instance, the Coulomb and Newton force would be proportional to 1/r 4 and
1/r 6 , respectively. There would be nothing wrong in theory, but apparently Nature has chosen a
different route.

References
[1] S. Weinberg, “The Quantum theory of fields. Vol. 1: Foundations,” Cambridge, UK: Univ. Pr.
(1995) 609 p.

[2] S. Weinberg, “Photons And Gravitons In S Matrix Theory: Derivation Of Charge Conservation
And Equality Of Gravitational And Inertial Mass,” Phys. Rev. 135, B1049 (1964).

[3] S. Weinberg, “Photons and gravitons in perturbation theory: Derivation of Maxwell’s and Einstein’s
equations,” Phys. Rev. 138, B988 (1965).

derivatives of hµν . We just write formally

Rµνρσ ∼ Fµν Fρσ = (∂µ Aν − ∂ν Aµ )(∂ρ Aσ − ∂σ Aρ ) , (95)

expand the product, and in each term replace the two A fields with an h, (∂µ Aν ) (∂ρ Aσ ) → ∂µ ∂ρ hνσ . This yields
precisely eq. (81), apart from the overall normalization which is of course a matter of convention.

You might also like