Faddeev Popov Ghosts

Felix Flicker
These notes are to accompany my talk on gauge xing in quantum eld theories (QFTs), and the mathematical
tricks associated with this procedure. In Section 1 I will look at gauge xing for Abelian elds, considering the
specic case of the Maxwell theory describing photons. In Section 2 I consider a non-Abelian generalization of the
Maxwell theory called Yang Mills theory, which describes vector bosons or gluons. I demonstrate the necessity of
introducing `ghost' elds in order to x the gauge. In Section 3 I look at how ghosts arise as a natural consequence
of the geometry of QFTs.

1 Abelian elds
The Maxwell action is

S [A] =

1
d x F F
4
4

(1)

where the eld strength (`curvature' in Maths language) is dened to be

F = A A .

(2)

Physical observables in QFT are given by the vacuum expectation values of gauge invariant quantities O [A]:

h|T O [A] |i =

DAO [A] exp (iS [A])

DA exp (iS [A])

(3)

where vacuum states (containing no particles) are denoted |i, and T is the time-ordering operator (which is not
important for what follows).
Just as in the case of classical electromagnetism, the eld A is only dened up to a choice of gauge. That is,
physical results should remain unchanged under a transformation
1
A
(x) = A (x) + (x)
e

(4)

with (x) a Lorentz scalar eld. This is the Lorentz covariant version of the classical statement that the magnetic
eld B = A is unaected if we change the magnetic vector potential A to A = A + 1e (x). The eld
strength F in QED acts a little like the magnetic eld in classical electrodynamics, and we can see that it is
indeed invariant to the gauge choice.
The problem with having the gauge freedom of Equation 4 is that the functional integration DA has an innite
degeneracy built into it: for each physically unique eld conguration there are an innite number of equivalent
congurations given by the choice of gauge. This situation is shown schematically in Figure 1.
Physically equivalent elds are said to lie on the same `gauge orbit'. We need to pick one eld from each gauge
orbit in order to remove the degeneracy in our functional integral. If the space of all elds is denoted A , we'd like
to consider a subset S A which contains only physically inequivalent elds. This situation is given schematically
in Figure 2.
First we dene a functional G [A] such that enforcing G [A] = 0 xes the gauge. In this talk I will use the
Feynman 't Hooft Landau gauge, given by
G [A] = A (x) + w (x) .

(5)

g( )

Figure 1: The space of all eld congurations, A , includes an innite degeneracy of physically equivalent elds
related by a choice of gauge. If two elds are related by a gauge transformation g () we say they sit on the same
`gauge orbit'. In this picture A has been drawn as a set of gauge orbits.

Figure 2: Gauge choices which select out a subset of all elds S A . The choice S1 shows a complete gauge
choice; examples include the Feynman 't Hooft Landau gauge G = A (x) + w (x). The choice S2 shows an
incomplete gauge choice, in which there is a residual gauge freedom remaining. Examples of this type are the axial
gauge G = A0 (x) and the Coulomb gauge G = i Ai (x) (with i [1, 3]).

Now we employ what I'll call the rst Faddeev Popov trick (FP1):

D (G [A ]) det

1=

G [A ]

(6)

where the delta function is understood to be innite dimensional (a delta functional, perhaps). Although I can't
prove Equation 6 I can give a plausibility argument for it: FP1 is the continuum limit of the N -dimensional equation
1=

N
Y

di (N ) (Gi (j )) det

i=1

Gi
j


,

(7)

i, j [1, N ]

and, if we take the simplest case of N = 1, the equation takes the trivial form

dG
.
(8)
d
In order to evaluate Equation 3 it is only necessary to evaluate the numerator, then set O = 1 for the denominator. Applying FP1 we have
1=

DAO [A] exp (iS [A]) =

d (G ())

DA

DO [A] (G [A ]) det

G [A ]

The determinant can be found explicitly given our choice of gauge (Equation 5):
G [A ]

1
A + + w
e

1

e


exp (iS [A]) .

(9)

where  , is the d'Alembertian. The important point to note is that the determinant (whatever it is) is
independent of both and of A, so can be pulled out of the functional integrals to give a constant prefactor. Thus
we have that

DAO [A] exp (iS [A]) = det

G [A ]

DA

DO [A] (G [A ]) exp (iS [A]) .

(10)

Furthermore, noting that both S [A] and O [A] are gauge invariant, we can replace them with S [A ] and O [A ]
respectively. The integration measure is also unchanged by gauge transformations (as they're just linear shifts), so
we have


G [A ]
DA DO [A ] (G [A ]) exp (iS [A ])
DAO [A] exp (iS [A]) = det

S



G [A ]
D
= det
DAO [A] (G [A]) exp (iS [A])

S


(11)

where we have simply relabelled A to A in the second line (it's a dummy function!). The functional integral over
now gives an innite constant, which will cancel out later on.
In Equation 11 we have achieved the desired result, in that the delta function enforces our gauge choice and
restricts the functional integral from the space of all elds A to the space of all physically ineqivalent elds S .
Before nishing with the Abelian case, though, there's a further trick which makes the physics of the problem more
apparent.

If we functionally integrate both sides of Equation 11 with a Gaussian distribution over w, Dw exp i d4 xw2 /2 ,
the left hand side (being independent of w) is simply multiplied by a function of . Bringing this across to the right,
and grouping the various prefactors together, we have that

DAO [A] exp (iS [A])

= N ()
A

or alternatively

with



1
DAO [A] ( A + w) exp i d4 x w2 exp (iS [A])
2
S


1
2
DAO [A] exp i d4 x ( A ) exp (iS [A])
(12)
2

Dw

= N ()

DAO [A] exp (iS [A]) = N ()

1
4

DAO [A] exp (iSGF [A])

d4 x F F

SGF =

1
2
( A )
2

(13)

(14)

the gauge xed action. Finally we have the desired result:

h|T O [A] |i =
1.1

DAO [A] exp (iSGF [A])

.
DA exp (iSGF [A])
A

(15)

The Maxwell propagator

Using integration by parts we can rewrite Equation 14 as

SGF =



1
1

d x A  + A
2

(16)

with the Minkowski metric. In momentum space this is

SGF =

d4 k 1
1

k 2 + k k k k A .
4 A


(2) 2

(17)

The term sandwiched between the gauge elds is the inverse of the photon propagator, (D )1 . The propagator
itself is uniquely dened by this, and can easily be veried to be
3

a) <AA>~
b) <AA>~

+ g2

+ g2

+g

Figure 3: a) Photons have no self-interactions, so the tree-level propagator is exact. b) Gluons (and vector bosons)
have 3-pt and 4-pt vertices, so there are loop contributions to the propagator.

i
(k) = 2
k + i

k k
(1 ) 2
k

(18)

where the small imaginary term i has been added to make things better dened in calculations (again, it's not
necessary in what follows). Note that if we hadn't applied the gauge x, the last term in the inverse propagator in
Equation 17 would not be present, and the inverse propagator would not have been invertible (so the propagator
wouldn't exist).
The propagator features a term proportional
in Equation 12 by functionally inte

to , which
was introduced
grating the gauge xing delta function with Dw exp i d4 xw2 /2 . Taking the limit 0 therefore rigidly
enforces the gauge xing condition. However, there's no requirement to do this in the quantum theory - after all,
we introduced arbitrarily. From the form of Equation 18 we see that has something to do with regulating
the transverseness of the photons' polarizations. If we choose to take 0 all photons (even virtual photons1 )
must be transverse. If takes a nite value, virtual photons are allowed to have a longitudinal component to
their polarizations, with photon polarizations having a Gaussian distribution of longitudinal components centred
on zero2 .
In the case of quantum electrodynamics (QED), which governs the coupling of photons to electrons, the unphysical nature of the longitudinal virtual photons is not a problem. All Feynman diagrams with contributions from
longitudinal photons cancel one another, as a result of the Ward identity, which follows from applying Noether's
theorem to the continuous symmetry associated with the invariance of DA to gauge transformations.

2 Non-Abelian elds
Yang Mills theory provides an extension of Maxwell theory to the case of non-Abelian elds. The action takes the
following form:
S [A] = tr

1
4

a
d4 x Fa F

(19)

where the eld strength is

a
F
= Aa Aa + gf abc Ab Ac .

(20)

The non-Abelian elds are now matrix valued: A = Aa ta , with ta the basis elements of the relevant group. For
example, if our group is SU (2) we could choose ta to be the Pauli matrices.

The trace would then be the trace of
these 2 2 matrices. The structure constants are dened by if abc tc , ta , tb .
One major dierence between Yang Mills theory and Maxwell theory, and consequently between gluons and
photons, is that gluons have self-interactions, as can be seen from the terms A2 A and A4 in Equation 19. This
makes things much more dicult. The situation is depicted in Figure 3.
Gauge transformations of the eld A now take the form
a

(A )

=
=

1
Aa + a + f abc Ab c
g
1
Aa + Dab b
g

(21)

1 Virtual photons appear as internal lines in Feynman diagrams. This requires an interacting theory, such as QED, rather than the
free Maxwell theory. Still, the point remains.
2 For interest, = 0 is called the `Landau gauge' and = 1 the `Feynman gauge'. I don't know where these names came from, but
the choice of (a Lagrange multiplier) is not a gauge choice, so they're misleading.

with
(22)

Dab , ab + gf abc Ac .

Equation 21 can be compared with Equation 4 of the Abelian case.

We proceed as before to evaluate the vacuum expectation values of gauge invariant operators. We again choose
the Feynman 't Hooft Landau gauge:
(23)

G = Aa + wa

and we need to evaluate

tr

h|T O [A] |i =

DAO [A] exp (iS [A])

.
tr A DA exp (iS [A])

(24)

Focussing on the numerator as before, we apply FP1 (Equation 6):

tr

DAO [A] exp (iS [A]) = tr

DA

D (G [A ]) det

G [A ]


O [A] exp (iS [A]) .

(25)

In the Abelian case we now used the fact that the determinant is independent of A to pull it out of the functional
integral. Let's see if this holds for the non-Abelian case.
G [A ]

1
Aa + Dab b + wa
g

1 ab
1
D =  + f abc Ac .
g
g

(26)

So the determinant in the non-Abelian case is a function of A, and we can't use the same simplication as before.
It is however independent of , meaning we can at least use the trick of changing the gauge invariant quantities DA,
O [A], and S [A] to DA , O [A ], and S [A ], and then relabelling A A. Thus Equation 25 can be rewritten
tr

DAO [A] exp (iS [A]) =






1 ab
D tr
DA Aa + wa det
D O [A] exp (iS [A]) .
g
A

(27)

One more trick carrying

over from the Abelian case is to functionally integrate both sides of the equation with
d4 xw2 /2 . Again the left hand side is just multiplied by a function of , and the delta functional
picks out the gauge xing condition:

Dw exp i

tr

DAO [A] exp (iS [A]) = N () tr

DA det


1 ab
D O [A] exp (iSGF [A])
g

(28)

with
SGF = tr




1 a
1
a 2
d x Fa F
A
.
4
2
4

(29)

How do we deal with the determinant? The most elegant solution was again proposed by Faddeev and Popov,
and is what I'll call the second Faddeev Popov trick (FP2):

det

1 ab
D
g


=





D cDc exp i d4 x
ca (x) Dab cb (x)

(30)

(the coupling g has been absorbed into the denition of c). The elds ca (x) and ca (x) are complex anticommuting
Lorentz scalar elds. To see where FP2 comes from, let's again consider the nite dimensional case, in which we
have Grassman variables ci and ci . Grassman variables for the case i = 1 are just anticommuting complex numbers.
If i [1, N ] then ci is an N component complex vector satisfying these anticommutation relations:
5

(31)

{ci , cj } = {
ci , cj } = {
ci , cj } = 0

with ci being the equivalent of the complex conjugate to ci . Calculus is dened over this eld by

dci ci = 1,

dci = 0

(32)

with the equivalent for the ci s. We are interested in evaluating

N
Y

dci

i=1

N
Y

dci exp ci Aij cj .

(33)

i=1

Note that the exponential, dened by its Taylor series, terminates at nite order due to the anticommutation
relations. For i [1, N ] the series terminates at the N th order. For simplicity we'll consider N = 1. The second
term in the series is 12 cAccAc = 12 cAccAc = 12 c2 A2 c2 = 0, and similarly for higher terms (A is a commuting
complex number here, or a commuting complex matrix in general). Therefore we have

dcdc exp (cAc) =

dcdc (1 cAc)

= A dc dc
cc

= A dcc dcc

(34)

= A

which can be contrasted with the case for commuting complex numbers z = x + iy :

dzdz exp (z Az) =

dxdy exp x2 + y 2 A

(35)

from which we see the necessity of using anticommuting rather than commuting elds.
Applying FP2 to our numerator (Equation 28) we have the nal result
tr

DAO [A] exp (iS [A]) = N () tr

DA

D cDcO [A] exp (iSF P [A, c, c])

(36)

with
SF P = tr

1
4

d4 x Fa F



2


1
Aa + ca Dab cb .
2

(37)

The nal form of our action suggests that gauge xing in the non-Abelian case requires the introduction of
additional (anticommuting complex scalar) elds. These elds are called Faddeev Popov ghosts. Expanding the nal
term in Equation 37 we have both a ghost propagator hcci and an antighost-gluon-ghost vertex, with corresponding
Feynman rules
a

i ab
k 2 + i

(38)
c

,b

k
a

= gf abc k .

(39)

+g

<AA>~

+g

+O(g2)

Figure 4: The 2-pt propagator for the gluon eld, hAAi, to order g (one loop). Gluons are given by coiled lines, and
ghosts by dashed lines.

These rules add extra Feynman diagrams to those in Figure 3 b. The modied 2-pt function for gluons is given in
Figure 4.
Note that the ghost eld describes anticommuting spin-0 (scalar) particles. This is in contradiction to the spin
statistics theorem, which states that fermionic particles (odd-half-integer spin) are antisymmetric under exchange
(the elds anticommute) whereas bosonic particles (integer spin) are symmetric under exchange (the elds commute).
It is a general property of ghosts in any QFT that they violate spin statistics. For this reason, and because they
only appear as internal lines in Feynman diagrams, we say that ghosts are not `real'. They are a mathematical trick
required to patch up perturbation theory.

3 BRST symmetry
In Section 1 we saw that virtual photons can have unphysical longitudinal polarizations, but that the contributions
to scattering amplitdes from these photons exactly cancel out. This is a result of the Ward identity (or the Ward
Takahashi identity if the photons are also allowed to move o-shell). Do we have a similar scenario for the nonAbelian case? The action (Equation 37) appears to allow unphysical longitudinal gluons. It turns out that the
Ward identity no longer cancels these contributions, but a new symmetry does. We call this BRST (Becchi Rouet
Stora Tyutin) symmetry, and it arises from the geometry of gauge elds.
Rewriting the Lagrangian density slightly we have



2


1
1
a 2
F

Aa + ca Dab cb
4
2
which we can rewrite by introducing a scalar commuting auxiliary eld B a like so:
L =

L0 =

1
2
a 2
F
+ (B a ) + B a Aa + ca Dab cb
4
2

(40)

(41)

where the equality follows from completing the square. The remaining term disappears upon functional integration
over the B eld:


  2



4
4
0
DAD cDc exp i d xL (A, c, c) =
DAD cDcDB exp i d xL (A, c, c, B) .
(42)
2
BRST is an additional symmetry of L 0 , not present in the bare Yang Mills Lagrangian, which comes about
upon the introduction of the ghosts and the auxiliary eld. For an innitesimal perturbation , with  a Grassman
variable, we have that L 0 = 0 for
Aa
ca
a

B a

= Dac cc
1
= gf abc cb cc
2
= B a
=

(43)

0.

The rst term in Equation 43 is simply a local gauge transformation with parameter a (x) = gca (x), which
means that the rst and second terms in Equation 41 are trivially invariant. With
some rearrangement the remaining


variation is found to be L 0 = 12 g 2 f abc fcde Ab cd ce + Ad ce cb + Ae cb cd , which vanishes because of the Jacobi
identity.
Dening the BRST operator Q such that = Q (for any of the elds), the geometrical nature of the
symmetry begins to take shape. First we note that applying the operator twice (the variation of the variation) gives
zero, showing that Q is `nilpotent':
Q2 = 0.

(44)

This property holds for the boundary operator in algebraic geometry, which maps (for example) a manifold M
to its boundary M. Nilpotent operators have the property that they split up the space on which they act into
three distinct subspaces. To make this statement precise, let's consider the canonically quantized Yang Mills theory
(with fermions, ghosts, and auxiliary elds). The continuous BRST symmetry must have a conserved charge Q (the
eigenvalues of the Q operator) by Noether's theorem. Being conserved we know the operator Q commutes with the
Hamiltonian:
Q = 0,

[Q, H] = 0.

(45)

Our Hilbert space H is now divided into three subspaces by Q. We can write it as a direct sum of these subspaces
H = H0 H1 H2 , where
|1 i H1 ,
|2 i H2 ,
|0 i H0 ,

such that
such that
such that

Q|1 i =
6 0
|2 i = Q|1 i ( Q|2 i = 0)
Q|0 i = 0, but |0 i =
6 Q|1 i.

(46)

In Maths-speak, we say that the vector space annihilated by Q is its kernel : H0 H2 = ker (Q), whereas the vector
space comprising vectors of the form Q|i is the image of Q: H2 = im (Q).
To see the relevance of this, consider the limit of zero coupling g 0, in which we approach the Abelian case. In
this case, Q acts as follows: it annihilates single ghosts, it converts forward-polarized components of A to ghosts, and
it converts antighosts to backward-polarized components of A3 . In terms of our Hilbert space decompostion, then,
we see that forward-polarized components of A and antighosts both live in H1 , backward-polarized components of
A and ghosts live in H2 , and the transverse components of A (the only physical bits of the whole theory) live in
H0 . Thus BRST symmetry shows that the unphysical polarization states of non-Abelian gauge theories are exactly
cancelled by the Faddeev Popov ghosts.
Let's look at this statement a dierent way. Again in the g 0 limit, choosing the Feynman `gauge' = 1, and
working in d dimensions, our partition function is








1
d
a 2
ab
=
DAD cDc exp i d x F + ca D cb
4





1
d
d

ab
D cDc exp i d x
ca  cb
=
DA exp i d x A  A
2
=

d/2

(det ())

(det ())

+2/2

(47)

This tells us that the ghosts act as `negative degrees of freedom', which cancel out the unphysical degrees of
freedom (unphysical polarization states) aorded us by the ambiguity in gauge. By construction we see that the
ghosts cancel exactly two degrees of freedom, which is intuitively correct in d = 4: we should have 2 transverse
polarizations, but A is a 4-vector, with 4 degrees of freedom. The ghosts cancel the other two.
3.1

The S-matrix

As a nal remark on BRST symmetry, I will prove that the S-matrix projected onto the physical H0 subspace is
unitary, meaning that we can work entirely in that subspace rather than in the whole Hilbert space H = H0 H1 H2
(in which S is automatically unitary).
From the denitions of the states |i i in Equation 46, and the Hermiticity of Q, it follows that the inner product
of any two states in H2 is zero:


h2A |2B i = h2A |Q|1B i = Q|2A i |1B i = 0

(48)

and similarly the product of any state in H2 with a state in H0 is also zero:
h2A |0B i = 0.

(49)

We need to show that S projected onto H0 is unitary, i.e.

3 Directly

it converts c B , but varying the action with respect to B to get the equations of motion, we nd that B a = Aa ,
showing that Q in fact acts to convert antighosts to backward-polarized components of A.

(50)

h0A |S |0C ih0C |S|0B i h0A |I|0B i

with I the identity. We can choose our initial state to be |in i = |0A i H0 . Now, since Q commutes with
the Hamiltonian, we know that if the initial state is an eigenvector of Q with eigenvalue zero then the nal (time
evolved) state must also be an eigenvector of Q with eigenvalue zero (as Q and H are simultaneously diagonalizable,
and the total S matrix is unitary). Therefore we have that
(51)

|out i = S|0A i H0 H2 .

We already know that the inner product of any |2 i state with either a |2 i or a |0 i state is zero, so any
contribution to the inner product of two nal states can only come from the inner product of |0 i states. That is,
h0A |S

(52)

X




S|0B i =
h0A |S |0C ih0C | S|0B i
C

which is exactly Equation 50.

References
[1] M. E. Peskin and D. V. Schroeder, An Introduction to Quantum Field Theory,
(1995)
[2] J. Polchinski, String Theory (Volumes I and II ), Cambridge
[3] L. D. Faddeev, Faddeev-Popov Ghosts, Scholarpedia,

Westview Press

University Press

4(4):7389

(2009)

(2005)