Solving Classical Field Equations

Robert C. Helling
[email protected]
Ludwig-Maximilians-Universität München

Feynman graphs are often thought of as tools for computations in perturbative quan-
tum field theories. However, there is nothing particularly quantum about them and, in
fact, Feynman rules for tree diagrams also arise when one solves classical field equations of
interacting theories (i.e. non-linear PDEs) perturbatively. I explained this in my lectures
on “Introduction to Quantum Field Theory” and since I am not aware of a textbook treat-
ment of this material (although all this is pretty standard and known to all practitioners)
I decided to write up these lecture notes.

1. The Klein-Gordon equation for the free field

We start with the Klein-Gordon equation

( + m2 )φ = 0.

This arises from the action Z

S= dd x φ( + m2 )φ.

The plane wave Ansatz φ(t, ~x) = exp(i(ωt − ~k · ~x)) is a solution provided the dispersion
relation
ω 2 − ~k 2 = m2
holds. Upon the identification ω 7→ E and ~k 7→ p~ this is nothing but the relativistic
dispersion relation
E 2 = m2 + p~2
of a particle of rest mass m in our units where c = h̄ = 1. As the Klein-Gordon equation is
linear there is a superposition principle and any sum or multiple of solutions yields a new
solution. We use this when we write the general solution in terms of its Fourier modes
which are plane wave solutions:
Z Z
~
φ(t, ~x) = dω d~k φ̃(ω, ~k)δ(ω 2 − ~k 2 − m2 )ei(ωt−k·~x)

d~k  ~ i(ω(~k)t−~k·~x)
Z 
~ ~
= a(k)e + a∗ (~k)ei(−ω(k)t−k·~x) ,
2ω(~k)
p
where we defined ω(~k) = ~k 2 + m2 . The measure factor in the second line arises from the
δ-function transformation (δ(f (x)) = δ(x − x0 )/|f ′ (x0 )| for f (x0 ) = 0) which contributes
at ω = ±ω(~k).

1
2. The non-linear field equation
Ultimately, we would like to solve equations of interacting fields stemming from a general
potential V (φ):
( + m2 )φ = V ′ (φ)
For example, we are interested in monomials V (φ) = g n1 φn . For odd n, the potential is
unbounded from below and one would be able to gain in infinite amount of energy by
making φ very large (or very negative). Such a system would be unstable and thus we
restrict our attention to even n. For n = 0, the right hand side vanishes. n = 2 gives just
another mass term and we can deal with it by redefining m2 . n = 4 is the first interesting
case and we will start out studying this “φ4 -theory”.
The equation ( + m2 )φ = gφ3 is non-linear and there is no longer a superposition
principle. This makes it very hard (and in general impossible) to write down exact solutions
to this equation beyond the trivial φ = 0 (the appendix, however, discusses a special class
of exact solutions, the “kinks” or “domain-walls”).
Thus, in general, we will have to resort to an approximate procedure: For g ≪ 1, we
can treat the right hand side as a small perturbation of the free Klein-Gordon equation and
obtain approximations to the true solution by perturbing solutions to the Klein-Gordon
equation.

3. The inhomogeneous field equation and Green’s functions

To this end, let us step back a bit and first analyse the inhomogeneous Klein-Gordon
equation
( + m2 )φ(x) = u(x)
where in the right hand side we have an arbitrary function which we think of as given
rather than the RHS of φ4 -theory which is a function of the unknown φ. Already this
equation is not linear but the space of solutions is still affine: When we have two solutions
to two such equations ( + m2 )φ1 = u1 and ( + m2 )φ2 = u2 we obtain a solution for the
equation where the RHS’s add:
( + m2 )(φ1 + φ2 ) = (u1 + u2 ).
We can use this idea to solve this equation in the case where the RHS can be decomposed
into a sum of “elementary” functions:
X
u(x) = αi ui (x)
i

If we can solve these elementary equations ( + m2 )φi = ui , we obtain a solution as

X
φ= αi φi .
i

We can take this idea to the extreme by decomposing the function u(x) into a “sum” of
δ-functions: Z
u(x) = dy u(y)δ(x − y)

2
Here, u(y) plays the role of the coefficients αi and the only x dependent function on the
right hand side is the δ-function. So, once we have a solution to

( + m2 )φg (x) = δ(x)

we have a solution Z
φ(x) = dy u(y)φg (x − y)

as we can directly compute:

Z
2 2
( + m )φ = ( + m ) dy u(y)φg (x − y)
Z
= dy u(y)( + m2 )φg (x − y)
Z
= dy u(y)δ(x − y)

= u(x).

This trick to solve inhomogeneous equations is obviously not restricted to the Klein-Gordon
equation. Such a solution φg which solves an equation with a δ-inhomogeneity is called
“Green’s function” or “propagator” (in physics circles) and “fundamental solution” (by
mathematicians). Strictly speaking, φg is not really a function but in general a distribution
(like the δ distribution) and and φ is obtained as a convolution which then yields a proper
function if u is nice enough but we will not analyse this in more detail.
It remains to find a solution φg of ( + m2 )φg = δ. As the differential operator is
translation invariant (it does not contain x dependent coefficients), this can be done using
a Fourier decomposition as differentiation becomes a simple multiplication in momentum
space. We take the Fourier transform on both sides of the equation

dk dk
Z Z
2

ikx ikx
√ d ( + m )φg (x) e = √ d δ(x)e
2π 2π
dk 1
Z
2 2 ikx
√ d (−k + m )φg (x)e =√ d
2π 2π

This allows us to read off

1 1
φ̃g (k) = √ d
2π −k 2 + m2

for the Fourier transform φ̃g of φg :

dk dk e−ikx
Z Z
−ikx
φg (x) = √ d φ̃g (k)e = .
2π (2π)d −k 2 + m2

3
4. Solving the interacting theory perturbatively
Armed with this ability to solve arbitrary inhomogeneous equations we now come back to
the φ4 -equation
( + m2 )φ = gφ3 .
We want to view this as a family of equations parametrised by the coupling constant g.
Similarly, the solutions to all these equations will depend on g. Underlying the idea of
perturbation theory is the idea that these solutions can be written as a power-series in g,
i.e. that they are analytic in g around g = 0.
Unfortunately, this is not really the case as can be seen as follows: Power series (in
the complex plane) have a radius of convergence (which can be zero or infinite): Everywhere
inside a circle of this radius the power series converges and outside it diverges. Thus, if the
power-series would converge for any g > 0 it would as well have to converge for some g < 0.
But for g < 0, again, the potential is unbounded from below and the system is unstable:
Solutions will be radically different from solutions of the free equation and not be small
perturbations. In fact, as is shown in the appendix, the kink solutions have energy and
action scaling like 1/g which has a singularity at g = 0. In a path integral (which in a
stationary phase approximation reproduces the classical behavior), these solutions appear
as saddle points contributing e−S ≈ e−1/g . These contributions are exponentially small
for small g. In fact, this function has an essential singularity at g = 0 and is invisible in a
Taylor expansion around this point. Indeed, “solitoninc” solutions like the kink are believed
to be what is missed by the perturbative treatment. Their contributions are exponentially
small for small g and can thus be safely ignored if one is interested in solutions to a finite
precision.
Nevertheless, we will just proceed and pretend that solutions to the φ4 -equation can
be written as a power series
X∞
φ= φn g n
n=0

for some coefficient functions φn (x).

Now plug this Ansatz into the equation and collect powers of g:
∞
X ∞
X 3 ∞  X
X 
2 n n
( + m )φn g = g φn g = φk φl φm g n
n=0 n=0 n=0 k,l,m
k+l+m+1=n

Comparing coefficients we find

X
( + m2 )φn = φk φl φm .
k,l,m
k+l+m+1=n

This simple manipulation has helped us a lot: We can now work our ways up starting from
n = 0 to larger n. The important observation here is that this is a differential equation
for φn in terms of a right-hand side given in terms of φk , φl , and φm where all k, l, m < n.
That is, when computing φn we already know these φk , φl , and φm !

4
 Let’s see how this works out for the first couple of n:

( + m2 )φ0 = 0

Nothing to be done. We know the solution is given in terms of plane waves obeying the
dispersion relation. Next is
( + m2 )φ1 = φ30
That was simple. Using the Green’s function, we can write down the solution:
Z
φ1 (x) = dy φg (x − y)φ0 (y)3 .

Now comes
( + m2 )φ2 = 3φ20 φ1 .
The 3 arises as there are three possible assignments of two 0’s and one 1 to (k, l, m). The
solution is Z
φ2 (x) = 3 dy φg (x − y)φ0 (y)2 φ1 (y)
Z Z
= 3 dy dy ′ φg (x − y)φ0 (y)2 φg (y − y ′ )φ0 (y ′ )3

Now for n = 3:
( + m2 )φ3 = 3φ20 φ2 + 3φ0 φ21 .
The iterated solution gets longer and longer:
Z
dy φg (x − y) 3φ0 (y)2 φ2 (y) + 3φ0 (y)φ1 (y)


φ3 (x) =
Z Z Z
= 9 dy dy ′
dy ′′ φg (x − y)φ0 (y)2 φg (y − y ′ )φ0 (y ′ )2 φg (y ′ − y ′′ )φ0 (y ′′ )3
Z Z Z
+ 3 dy dy ′
dy ′′ φg (x − y)φ0 (y)φg (y − y ′ )φ0 (y ′ )3 φg (y − y ′′ )φ0 (y ′′ )3

5. Feynman graphs in position space

Obviously, continuing like this will be more and more cumbersome. However, we see a
simple pattern of these terms emerging: We can represent the solution for φ1 like this:

φ0

x y φ0
φg
φ0

5
We obtain the solution by bringing together three φ0 ’s at one point y and then transport
this to x using the Green’s function φg . At higher orders, this pattern is iterated. For
n = 2, we have

y’

3 y

where the factor 3 arises because the graph for φ1 can be substituted at any of the three
legs. At level n = 3, there are two different graphs

y’’

y’ y’

y
9 +3
y

y’’

again with “symmetry factors” indicating the number of possibilities of obtaining these
graphs.
In this graphical notation, it should be clear what we have to do to obtain the
expression for φn : We have to draw all possible graphs according to these rules:

• Draw n vertices for the expression for φn at order g n .

• Each vertex gets one in-going line at the left and three outgoing lines to the right.
• A line can either connect to the in-going port of another vertex or to the right-hand side
of the diagram.
• Write down an integral for the point of each vertex.
• For a line connecting two vertices at points y1 and y2 , write down a Green’s function
φg (y1 − y2 ).
• For a line ending on the right. write down a factor of φ0 evaluated at the point of the
vertex at the left of the line.
• Multiply by the number of permutations of outgoing lines at the vertices which yield
different diagrams (“symmetry factor”).

6
6. More general field equations

Looking back at how these rules came up, we can immediately guess the generalisation to
other field equations: The fourth order potential V (φ) = g4 φ4 resulted in a field equation
with a cubic right-hand side. The cube in the field equation became φk φl φm with the
constrained sum over k, l, and m in the equation for φn and eventually resulted in the
rule that each vertex has to have three outgoing lines. This suggests that for a potential
′
V (φ) = gp φp we would derive similar rules but now with (p − 1) outgoing lines at each
vertex.
If we have a potential which consists of a sum of more than one monomial there will
be a separate coupling constant for each monomial.

I
X gi
V (φ) = φpi
i=1
pi

As a result, we would express φ in a multi-dimensional power series over all coupling

constants:
X
φ= φn1 ···nI g1n1 · · · gInI .
n1 ···nI

Consequently, we now have I different types of vertices. To compute the solution for
φn1 ···nI we draw all diagrams with n1 vertices of type 1 (which have (p1 − 1) outgoing
legs), n2 vertices of type 2 (with (p2 − 1) outgoing legs) and so on to nI vertices of type I.
Of course, we can have also more than one type of field. In that case each field comes
with its own equation of motion which we solve perturbatively: Each field has its Green’s
function and thus, in the graphical notation, we have different types of lines denoting the
different fields and the vertices have “ports” connecting to the different types of lines. For
example in Quantum electrodynamics, there is an electron field denoted by a straight line
and a photon field denoted by a wavy line. There is a cubic term in the action which reads
eψ̄γ µ Aµ ψ. Here ψ is the electron field, Aµ is the photon field (which is a fancy name for the
vector potential of electromagnetism), e is the charge of the electron playing the role of the
coupling constant and γ µ is some matrix needed to write down the Dirac equation which
is the analogue of the Klein-Gordon equation for spin 1/2 fermions like the electron. The
bar indicates a conjugate (which is finally responsible for the difference between electrons
and positrons) which results in the electron lines having a direction which is indicated
by an arrow. This cubic term in the action ends up in quadratic right hand sides of the
field equations: The field equation for the photon has a term quadratic in the electron on
its right-hand side (namely the expression for the electromagnetic current) whereas the
electron has a right-hand side which is bilinear in the electron and the photon. A typical

7
diagram then looks like this:

7. Relation to the quantum theory

Looking back at the “Feynman rules” above for the φ4 -theory which describe the allowed
graphs we notice that these rule simply describe all graphs with 4-valent vertices (there
are in total four lines at each vertex) that do not contain any loops. In quantum field
theory, there is a similar perturbative expansion which similarly is most easily written
down in a graphical notation in which each Feynman diagram corresponds to an integral
expression. It turns out that the rules which translate graphs to integrals are identical to
our rules above and the only difference between our classical theory here and the quantum
theory is that we drop the distinction between in-going and outgoing lines at a vertex
and only require that there are four lines at each vertex (for φ4 -theory), no matter if in-
or out-going. Weakening this rule has the consequence that also graphs containing loops
are allowed. This fact that the perturbative solution in classical field theory and in the
quantum theory are so similar is then explained by the observation that each loop in a
Feynman graph contributes a factor h̄ and thus the contribution of the loop graphs vanish
in the classical limit of h̄ → 0.

8. Feynman rules in momentum space

For practical calculations (at least those in flat space), our Feynman rules which are based
in position space are not the simplest ones possible. It is easy to see that in momentum
space, we don’t have to do any integrals! (This will, however, change in the quantum
theory where one has to do one integral per loop). For simplicity, let us once more consider
φ4 -theory. Assume we have drawn all our diagrams and translated them to an integral
expression. Let us now zoom in to one particular vertex representing a space-time point
y. It has one in-going and three out-going lines either representing φg or φ0 . We can write
all these four functions in terms of their Fourier transform
dk e−ik(x−y)
Z Z
φg (x − y) = and φ0 (y) = dk φ̃0 (k)δ(k 2 − m2 )eiky .
(2π)d −k 2 + m2
Let us denote the momentum of the in-going line at the vertex by k1 and those of the
three outgoing lines by k2 , k3 , and k4 . Then the only occurrence of the point y is in the
combined exponential
e−iy(k1 −k2 −k3 −k4 ) .

8
The integral over the position y now yields a δ-function via
Z
dy e−iy(k1 −k2 −k3 −k4 ) = (2π)d δ(k1 − k2 − k3 − k4 ).

This now trivialises the k1 integration of the in-going line in terms of setting k1 7→ k2 +
k3 + k4 . This of course is nothing but momentum conservation. Once we have repeated
this procedure for all vertices, all the momenta on the “internal” lines representing Green’s
functions φg are specified in terms of the momenta of the external lines representing φ0 .
Thus we arrive at an algebraic expression without any integrals for the Fourier modes of
φn . In momentum space, we found the new Feynman rules:
• Draw the same diagrams as in the position space representation including symmetry
factors.
• Specify momenta satisfying the dispersion relation k 2 = m2 on the lines extending to
right end of the diagram.
• Use momentum conservation at the vertices to determine the momenta ki of the internal
lines connecting two vertices.
• write down a factor √ 1 d −k21+m2 for each internal line and a factor (2π)d for each
2π i
vertex.
These rules then determine the Fourier component φ̃n (p) where p is the sum of the
momenta of all external lines extending to the right end of the diagram.
The method of Feynman diagrams suggests a new perspective on particle-wave du-
ality: In the beginning, we set out to find solutions to classical field equations. But in that
investigation we were lead to a language of diagrams which is much more natural in a par-
ticle interpretation: The lines of the Feynman diagrams are just the world lines of particles
of mass m: The free, external solutions obey the dispersion relation for this rest mass and
at the vertices the particles collide and interact observing conservation of momentum. As
it turns out, in theories with additional symmetries also the Noether charges are conserved
at the vertices. Thus it is natural to ask if we are really dealing with a field/wave-theory
or rather with a particle theory. However, after a bit of contemplation one realises that
this dichotomy is not real: This theory describes both: When looking at it in terms of a
partial differential equation it looks more like a wave theory but the perturbative solutions
in terms of Feynman diagrams are best understood in terms of interacting particles. This
theory is really both, a wave- and a particle-theory even though it is classical and not
quantum!

9. Appendix: The solitonic kink

In this appendix, we will discuss properties of a special class of non-perturbative solutions
to φ4 theory. Specifically, we want to consider the potential
2
m4 m2 2 g 4 g m2

2
V (φ) = − φ + φ = φ − .
16g 2 4 4 4g
Here, we have written the quadratic (mass) term as part of the potential. Note that it has
an unusual sign but this is not a problem as the potential is unbounded from below and

9
in fact positive. The constant term does not influence the field equations but simplifies
expressions in the following. Thus, we are looking for solutions to

φ = −V ′ (φ) = m2 φ − gφ3 .

The right-hand side vanishes at the extrema of the potential: The minima are at φ1/2 =
± √mg and the unstable maximum is at φ3 = 0. The two minima are the two degenerate
ground-states of the system and φ sitting in one of these “vacua” is a trivial solution.
To have any solution of finite energy (and action) the field φ has to be in its ground-
state at large distances. However, as we have a degenerate ground-state it can be in a
different ground state in different directions. Specifically, we can consider the case d = 1
with only one spatial coordinate or, alternatively, solutions which only depend on one
coordinate x. A solution which is in the ground-state φ1 = − √mg for x → −∞, then passes
over the hill of the potential around the origin and goes to the other vacuum φ2 = + √mg
for x → +∞ is called a kink. Here we will derive such a solution explicitly.
We are assuming that φ only depends on one spatial coordinate x and thus the field
equation reads
d2 φ dV
= .
dx2 dφ
In analogy with conservation of energy in mechanics, we observe that if we solve the first
order equation
 2
dφ
= 2V
dx
we automatically solve the second order equation of motion as can be seen by taking the
x-derivative on both sides. Hence, we solve

m2
r  
dφ g 2
=± φ −
dx 2 g

by separation of variables r Z
dφ g
Z
2 2
=± dx.
φ − m /g 2
Thus √

m
φ(x) = ± √ tanh m(x − x0 )/ 2 .
g
The constant of integration x0 determines where φ passes over the hill in the poten-
tial.
It is instructive to compute the action for this solution. As φ depends only on x the
integrals over the other variables
√ yield irrelevant (infinite) factors and it remains (using
the first order equation φ′ = ± 2V several times)
Z " 2 #  2 Z φ2 Z φ2 p
dφ dφ dφ 8m3
Z
S = − dx + V (φ) = −3 dx = −3 dφ = −3 dφ 2V (φ) = .
dx dx φ1 dx φ1 g

10
 5

1.0

VHΦL
0.5
0 x

0.0

1
0
Φ
-1 -5
Fig. 1: The kink solution

We see that the action scales like 1/g. Thus, in a path integral (which gives the classical
theory in the h̄ → 0 saddle point approximation), these solutions contribute with a weight
e−1/g which is exponentially small for small g as we assume it in the perturbative analysis.
In fact, due to the essential singularity of this expression at g = 0, all g-derivatives vanish
at g = 0 and thus in a Taylor expansion (as we assume it for the perturbative Ansatz),
this function is indistinguishable from 0. Thus, these “solitonic” solutions are missed by a
perturbative expansion. It is believed that these missed solutions are the reason that the
perturbative expansion cannot be convergent to the exact solution.
The kink solution above has a parameter or “modulus”, the constant of integration
x0 . This could have been expected from the very beginning as the equation of motion is
translationally invariant but the solution is not. Thus there should be a whole “moduli-
space” parametrising the kink solution. In this simple case, the moduli space is just the
real line which we can think of as the configuration space of a “kink-particle”. If this
interpration of the kink as a particle holds it should also be possible for it to move. That
is we can consider solutions where x0 has a (slow) time dependence: x0 (t).
Although this is not a solution of the equation of motion anymore (one should not
confuse this with a kink √ to which a boost is applied, that is a solution in which x is
substituted by (x + vt)/ 1 − v 2 which is again a solution), we can work out the action for

m √

φ(x, t) = ± √ tanh m(x − x0 (t))/ 2 :
g

11
With the time dependent x0 , ∂φ/∂t does not vanish anymore, but we have

∂φ ∂φ dx0
=− .
∂t ∂x dt
The action for this Ansatz gets a new term from expanding ∂µ φ∂ µ φ
2
1 16m3
  Z   
∂φ
Z Z Z Z
2
S= dt dx = dt 2 dx V (φ) ẋ0 = dt ẋ20 .
∂t 2 3g
3
This is of course the action of a particle with mass M = 16m3g . This we should interpret
as the the mass of the kink. It diverges in the weak coupling g → 0 limit.

12