Embed Size (px)
Citation preview
A Practical Introduction to Quantum Field Theory
0.1 A Practical Introduction
The aim of these lectures is to provide a very brief introduction to the philosophy and some of the key tools ofquantum eld theory (QFT). Inevitably with such a short course a huge amount will be omitted. To minimisethe damage I will try to present a technical introduction to some of the methods used day-to-day in practical eldtheory, and will discuss the philosophy of the subject around the mathematics.
There are two reasons for this choice. First, QFT is fundamentally a practical, pragmatic subject. It hasunmatched predictive power, and typies the modern way we understand the universe - but at the same time itis riddled with philosophical and mathematical issues. In many cases we don't know why it works, it just does1.Second, there are many aspects of eld theory which are necessary for a full understanding but which aren't usedday-to-day in research. An illustrative example is the quantization of classical elds. When calculating scatteringamplitudes, for example, particles are described by quantum elds from the outset. Starting from the familiar ideaof classical elds and turning them into the more complicated quantum elds may help build intuition, and has itsplace in a longer course - but I have rarely seen it used outside the eld of quantum gravity, and it will not makeit into this `practical' guide.
One topic which is notably missing from this guide is second quantization. This is important to the philosophyof QFT and is also routinely used in research. I have found it possible to tell a self-contained story without it, andrefer the interested reader to the literature.
0.2 Note on Conventions
My convention is to use , to mean `equal to by denition'. The symbol ≡ is reserved for `is equivalent to', in thesense that cos2 (θ) + sin2 (θ) ≡ 1, whereas cos (θ) = 1 could be a specic solution to something (but cos (θ) and 1are not equivalent).
I will assume Einstein summation notation throughout. I will make some limited eort to ascribe covariantquantities lower indices and contravariant upper, but since they can simply be raised and lowered with a metric Imay well forget to do so; please interpret ∂µ as ηµν∂ν . Should it come up I'll use the metric (+−−−).
Lower case symbols denote 4-vectors, e.g. xµ = (t,x)µ. The inner (dot) product behaves as is appropriate for
the things it's dotting, so x · p ≡ xµpµ and x · p ≡ xipi. I will also dene p2 , p · p (etc.) and the d'Alembertian
, ∂2 = ∂2t −∇2 = ∂2
t −4.Fourier transforms and their inverses are dened as:
ϕ (p) =
d4x exp (ip · x)ϕ (x)
ϕ (x) =
d4p
(2π)4exp (−ip · x) ϕ (p)
where I will stick to the convention of writing twiddles over transformed quantities (in the literature both elds areusually given the same name).
0.3 Suggested Textbooks
There was until recently a famous absence of QFT textbooks aimed at people new to the subject. I believe thischanged last year with the release of Tom Lancaster and Steve Blundell's Quantum Field Theory for the Gifted
1See, for example, Haag's theorem, which proves you cannot even consistently dene the vacuum in an interacting eld theory! Formathematical issues see the many failed attempts to create an Axiomatic QFT.
1
Amateur [3]. It covers a wide range of material very clearly and can be considered the course textbook (althoughthese notes do not follow it directly).
The classic QFT text is Peskin and Schroeder [2], which is amusingly called an Introduction but which goesfurther than any course I've so far encountered. Although general, it leans slightly towards particle physics. Forthe more condensed matter-minded there's the equally brilliant Altland and Simons [4]. I found Zee [5] a niceread when I rst met QFT, but it's probably best approached as a technical popular science book rather than atextbook. Think of it as the QFT version of the Feynman lectures - if you read it all those technical details will tinto place, but you're probably not going to learn the subject from it rst time around.
The most relevant resource for this course, however, is provided by the notes I was given as an undergraduate.That course was prepared by John Chalker, a condensed matter theorist, and André Lukas, a high energy physicistand string theorist. The notes for their course are truly excellent, and I provide a link to them on my website. Theyare unmatched in their clarity and conciseness: almost all of the material in the notes you are reading is covered inSection 1.1 of Chalker and Lukas!
I will mention at the start of each section particularly relevant reading for the coming topics, taken from theReferences below.
References
[1] J. T. Chalker and A. Lukas, Theoretical Physics Lecture Notes (Oxford University), link available on my website(link)
[2] M. E. Peskin and D. V. Schroeder, An Introduction to Quantum Field Theory, Perseus Books, 1995
[3] T. Lancaster and S. J. Blundell, Quantum Field Theory for the Gifted Amateur, Oxford University Press, 2014
[4] A. Altland and B. Simons, Condensed Matter Field Theory, Cambridge University Press, 2006
[5] A. Zee, Quantum Field Theory in a Nutshell, Princeton University Press, 2010
[6] R. B. Dingle, Asymptotic Expansions: Their Derivation and Interpretation, available on Michael Berry's website(link)
Acknowledgments
First and foremost I would like to thank my entire cohort of friends from the Perimeter Institute's PSI programmefor bringing their expertise from all areas of physics to these notes. Particular thanks are owed to Bruno Amorim,Eduardo Tremea Casali, Lauren Greenspan, Sebastian Montes Valencia, James Reid, and Nick Jones for thoroughreadings of the rst draft with many helpful comments.
2
1 Actions and Lagrangians
The working of this section, and in fact Sections 2 and 3, follows very closely the working presented to me in my rstcourse on QFT in [1]. This rst chapter additionally takes a lot of working from Peskin and Schroeder, probablythe classic QFT textbook (mainly written from a particle physics point of view) [2].
1.1 Natural Units
You will rarely see ~ or c appear in a modern eld theory paper. We say both are `set to one'. In fact we just choosesensible units. In relativistic systems it becomes apparent that space and time are really the same thing, so shouldbe measured in the same units. This does away with c. In a similar manner all fundamental constants graduallydisappear as we realize energies are just inverse times (~), which are also temperatures (kB), etc. etc. - in the endall units are simply comparisons to scales humans nd accessible. They do help by providing added redundancy incalculations to check you've not made a mistake, but at some point more errors are introduced by trying to keeptrack of them.
It is convenient to leave one dimension in. A typical choice of dimension is Energy, E, and a typical choice ofunit is GeV. When asked for the dimension of some mathematical object we can state a number corresponding tothe power of Energy. For example, [Energy] = 1, because the dimension of energy is E1. Similarly, [Time] = E−1
so [Time] = −1, and [Length] = −1, [kB ] = [~] = [c] = 0, [m] = [p] = [∂µ] = 1 etc. These `natural units' are arewhat I will employ throughout this course. The argument goes that you can always `put in the units again later' -well, you could convert them, or you could not bother - but in the latter case you might need to state a length inGeV−1. In reality you probably won't be putting many numbers in anyway so the issue rarely comes up, and thealgebra is easier with no constants kicking about.
1.2 Examples of Actions
The key object in a QFT is the action. This completely denes the theory of interest. Some examples from realtheories are:
SKlein-Gordon [ϕ] =
d4x
(12∂µϕ∂
µϕ− 12m2ϕ2 + jϕ
)SMaxwell [Aµ] =
d4x
(−1
4FµνF
µν + jµAµ
), Fµν = ∂µAν − ∂νAµ (1)
SSchrödinger [ψ] =
d3xdt(ψ† (i∂t −H)ψ + jψ† + j†ψ
)SChern-Simons [Aµ] =
d3x (εµνρAµ∂νAρ + jµA
µ)
SKlein-Gordon [ϕ] =
d4p
(2π)4
(12ϕ(p2 −m2
)ϕ+ jϕ
).
Things to note:
• the Schrödinger action is not relativistically covariant. That's okay, we use it in condensed matter where thepresence of a lattice breaks Lorentz symmetry. I'll mainly consider general theories rather than specializingto condensed matter, particle physics, cosmology and so on - eld theory is ubiquitous.
• The actions are functionals, i.e. functions of functions. This is symbolised by the brackets in S [ϕ]. We'll seesome examples of functionals shortly.
• The elds themselves are functions. This could be a function of spacetime like ϕ (xµ), a function of spaceand time separately like ψ (x, t), a function of 4-momentum like ϕ (pµ), or whatever basis you choose to writethings in just like in normal QM. Strictly, the elds are operator-valued functions as they may not commute(for example it may be the case that [ϕ (xµ) , ϕ (yν)] 6= 0 ).
• All the examples given here feature only one kind of eld and are quadratic in that eld. This arguably makesthe corresponding theories classical as we'll see later in this section.
3
f(x)
x
cos(x)
0
Figure 1: Two possible paths a particle could take between two points, assigned functions. The phase accumulatedis a function of the function - i.e. a functional.
• Dierent types of eld appear. I will stick to this convention: ϕ is a real scalar (spin-0) eld, Aµ is a realvector (spin-1) eld, ψ is a complex spinor (spin-1/2) eld. There are many other types of eld, these are justsome common examples.
• The actions all include external currents j. The type of eld dictates the type of current; for example thevector eld Aµ requires a vector current with the same number of components, jµ. Complex elds needcomplex currents, and so on.
• The action in each case involves an integral over a `Lagrange density', which can be in real space or momentumspace. For example the Klein-Gordon action is an integral over the Klein-Gordon Lagrange density:
LKlein-Gordon = ∂µϕ∂µϕ−m2ϕ2 + jϕ.
• Symmetries more or less constrain the theory. We will not have time to ruminate too long on this point here,but the modern way to construct eld theories is to think of the symmetries a system obeys, then to write themost general Lagrange density compatible with those symmetries. In many cases, the resulting QFT makesvery accurate testable predictions.
1.3 Functionals and Derivatives
A functional is very much like a function except that its domain is the set of functions rather than the set of (e.g.)real numbers. It is probably already clear why they are useful in QFT by considering standard QM: the famousexplanation of the result of Young's two-slit experiment is that the photon takes every possible path between sourceand screen, and these paths interfere. A path through space is a function of space. For example, the straight-linepath could be the trivial function f (x) = 1. Another path might be f (x) = cos (x), and so on. This is shown inFigure 1. We want to `sum over histories' and weight each path by the quantum mechanical phase accumulatedalong it, multiplying with exp (iS [f ]).
In QFT we are considering elds ϕ (x, t) which are a function of both space and time, but otherwise everythingis the same. The action is a functional of the possible eld congurations, and we again weight each congurationby a phase exp (iS [ϕ]). In both cases the `sum' is actually a form of integral over possible elds, called a functionalintegral, but we'll come to that later on. We should learn to dierentiate before we integrate, especially since onlya very tiny number of functional integrals can even be carried out.
There is a simple generalization of normal derivatives to `functional derivatives', considered shortly, but I'veseen people come unstuck using it on more complicated functionals. For minimizations I nd the easiest methodis to explicitly vary the function. Say we consider paths with xed endpoints like those in Figure 1, and that we'dlike to know the shortest distance between the two points, with no additional weighting (no phases for example).The length L of a given path f (x) is the sum of the lengths of the innitesimal line segments:
L [f ] =x1∑x0
√dx2 + df2
dx→ 0 ↓
= x1
x0
dx√
1 + f ′2.
4
Now consider varying the path f (x) to f (x) + ε (x) where the function ε (x) is small and completely arbitrary,except that it vanishes at the endpoints (ε (x0) = ε (x1) = 0). In fact we'll consider a continuous parametrizationbetween the two `extremes': f (x) + λε (x), λ ∈ [0, 1]. The function f which minimizes L [f ] is the one where, forany ε, a tiny shift from λ = 0 leaves L unchanged. That is, if f minimizes L then
∂L [f + λε]∂λ
∣∣∣∣λ=0
= 0.
Note that the derivative is just a standard partial derivative, as λ is just a real number. If f were changed intosome more complicated object such as a vector eld Aµ, the variation would change to Aµ + λεµ - but λ is alwaysjust a real number. We nd in the present case that
∂L [f + λε]∂λ
∣∣∣∣λ=0
= 0 = x1
x0
dxε′1√
1 + f ′2
or, integrating ε′ by parts,
0 = x1
x0
dxεd
dx
(1√
1 + f ′2
).
This must be true for all ε, which means that
d
dx
(1√
1 + f ′2
)= 0.
The derivatives are total derivatives and this can be solved by simple integration to yield (with the boundaryconditions)
f (x) = (f (x1)− f (x0))x− x0
x1 − x0+ f (x0)
and we have shown that the shortest distance between two points, in Euclidean space, is a straight line.The method can be formalized somewhat by introducing the functional derivative, given the symbol:
δL [f ]δf
.
Aside from behaving basically like a normal derivative, but one acting on functions rather than variables, thereare three requirements of the functional derivative:
1.[δδf ,
dx]
= 0 (it commutes with integrals)
2.[δδf ,
ddx
]= 0 (it commutes with normal derivatives)
3. δf(x)δf(y) = δ (x− y).
Redoing the example from before (terms in parentheses indicate the rule used):
5
L [f ] = x1
x0
dx√
1 + f ′2
↓ (1)δL
δf (y)= −
x1
x0
dx1√
1 + f ′2δf ′ (x)δf (y)
↓ (2)δL
δf (y)= −
x1
x0
dx1√
1 + f ′2d
dx
δf (x)δf (y)
= x1
x0
dxd
dx
(1√
1 + f ′2
)δf (x)δf (y)
↓ (3)
=d
dy
1√1 +
(dfdy
)2
so as required we lose the integral. In many actual calculations, especially minimizations, it is often easier tocontinue to use the full form with λ.
1.4 Euler Lagrange Equations
The most important example of functional dierentiation for elds will be taking derivatives of the action withrespect to elds. Consider a general action
S [ϕ] =
d4xL (ϕ, ∂µϕ)
and minimize as before:
S [ϕ+ λε] =
d4xL (ϕ+ λε, ∂µϕ+ λ∂µε)
∂S [ϕ+ λε]∂λ
∣∣∣∣λ=0
= 0 =
d4x∂L
∂ (ϕ+ λε)∂ (ϕ+ λε)
∂λ+
∂L
∂ (∂µϕ+ λ∂µε)∂ (∂µϕ+ λ∂µε)
∂λ
∣∣∣∣λ=0
0 =
d4x
(∂L
∂ϕε+
∂L
∂ (∂µϕ)∂µε
)=
d4xε
(∂L
∂ϕ− ∂µ
∂L
∂ (∂µϕ)
)with an integration by parts in the nal step. As before the arbitrariness of the eld ε implies that for the integralto be zero the term in parentheses must be zero, and we have the Euler Lagrange equation for the eld minimizingthe action:
∂L
∂ϕ− ∂µ
∂L
∂ (∂µϕ)= 0. (2)
By varying the actions in Equation 1, or equivalently sticking their Lagrange densities into Equation 2, wearrive at the Euler Lagrange equations for the respective elds. We nd the equations which lent their names tothe actions in each case:
( +m2
)ϕ (xµ) = j (xµ) (Klein-Gordon)
∂νFµν (xν) = jµ (xν) (Maxwell)
(i∂t −H)ψ (x, t) = j (x, t) (Schrödinger).
6
The j acts as a source for the respective elds (most obvious perhaps in the Maxwell case, where the source is anelectromagnetic current). The standard equations are dened for j ≡ 0; the inhomogeneous Schrödinger equation,for instance, is a strange object, but as we will see later we could set j (x, t) = δ3 (x) δ (t) to nd the Green'sfunction, a very useful operator2. In the Schrödinger case the variation can be done with respect to either ψ or ψ†,but the resulting equations are merely hermitian conjugates of one another.
1.5 Propagators and Green's Functions
The method of Green's functions is a beautiful and powerful piece of mathematics. George Green lived in a windmilllike Jonathan Creek - but whereas Creek solved hammy and implausible mysteries in the late 1990s, Green solveddierential equations in the 1830s. Green realized that if you can nd the response of a system to a δ-functionimpulse you know the response to any driving function, as you can make any function from an integral over δs. Youcan therefore write down a general solution to the equation. Green's functions generally have a physical signicancein themselves, and we will see that in QFT they give the propagators for moving particles around.
Take as a simple example the Klein-Gordon equation. Rather than a general external current we would like tosolve the equation for a δ-function kick (assuming the eld lives in 3 + 1D):(
+m2)G (x) = δ4 (x) .
This is easily solved by Fourier transform, dening
G (x) ,
d4p
(2π)4exp (−ip · x) G (p)
where p · x , pµxµ, giving
( +m2
) d4p
(2π)4exp (−ip · x) G (p) = δ4 (x) .
We recall a particular representation of the Green's function, and pull the dierential operator through the pintegral which it doesn't act on:
d4p
(2π)4G (p)
( +m2
)exp (−ip · x) =
d4p
(2π)4exp (−ip · x)
so
d4p
(2π)4G (p)
(m2 − p2
)exp (−ip · x) =
d4p
(2π)4exp (−ip · x)
and nally we have the momentum-space Green's function
G (p) =−1
p2 −m2.
Fourier transforming back to real space gives the Green's function
G (x) = ∞
−∞
d4p
(2π)4−1
p2 −m2exp (−ip · x) .
The fact that the momentum-space function is easier to work with is a regular feature of QFTs, as we often careabout translationally invariant systems. It is also often the case that our boundary conditions constrain p ratherthan x making this basis particularly convenient. It's often easier to work in p and Fourier transform at the end ifnecessary. If we'd now like to solve the equation for some arbitrary but specied current j (x) we have the solutionby adding δ-functions as follows:
2A similar method applied instead to the time independent Schrödinger equation leads to the Lippmann-Schwinger equation inscattering theory.
7
( +m2
)G (x− y) = δ4 (x− y)
d4y j (y) → ↓
( +m2
) [d4yG (x− y) j (y)
]= j (x) .
That is, the Green's function `propagates' the disturbance in the eld caused by the input current j (y) to the eldsolving the Klein-Gordon equation:
ϕ (x) = ϕ0 (x) +
d4yG (x− y) j (y)
= ϕ0 (x) +
d4yG (x− y)(y +m2
) [d4zG (y − z) j (z)
]= ϕ0 (x) +
d4yG (x− y)
(y +m2
) [d4zG (y − z)
(z +m2
) [d4wG (z − w) j (w)
]]etc. etc.
where ϕ0 (x) was a solution before the current was introduced. The physical interpretation of the Green's function isthat it is the propagator for the ϕ eld3. Later on, when considering interacting theories, j may itself be inuencedby the eld conguration, i.e. j = j [ϕ]. In this case the propagator method outlined here becomes very useful indening a perturbative series for solving the problem.
Note that if we had the action written in momentum space we could just have `read o' the propagator as theinverse of the function appearing between the elds in the quadratic term:
SKleinGordon[ϕ, j ≡ 0
]=
d4p
(2π)4
(12ϕ(p2 −m2
)ϕ
)=
d4p
(2π)4
(−1
2ϕG−1ϕ
).
This is generically true: the inverse propagator is sandwiched inside the quadratic term in the action. The reasonbecomes apparent when we consider functional integrals later on.
1.6 Gauge Fixing*
This section will not be covered in lectures and is for the interest of people already familiar with gauge theories.Reading o the propagator for the Maxwell eld is a bit tricky as it requires re-arranging into the form
SMaxwell [Aµ] =
d4x12Aµ(G−1
)µνAν .
Starting from the form above
SMaxwell [Aµ, jµ] =
d4x
(−1
4(∂µAν − ∂νAµ) (∂µAν − ∂νAµ) + jµAµ
)integration by parts leads to the form
SMaxwell [Aµ, jµ] =
d4x
(12Aν (Aν − ∂µ∂νAµ) + jµAµ
)which can now be written in the desired form by the introduction of a metric:
SMaxwell [Aµ, jµ] =
d4x
(12Aν (ηµν− ∂µ∂ν)Aµ + jµAµ
)3Prof. Hannay points out that strictly the Green's function and propagator are temporal Fourier transforms of one another, with
the Green's function depending on energy and the propagator on time. In the literature the terms are often used interchangeably.
8
and in momentum space
SMaxwell
[Aµ, j
µ]
=
d4p
(2π)4
(12Aν(−ηµνp2 + pµpν
)Aµ + jµAµ
)where A (p) = A (−p) for simplicity. This suggests the propagator will come from inverting the Euler Lagrangeequation (
−ηµνp2 + pµpν)Gνρ = δρµ.
The problem with this is that the term in parentheses is non-invertible. It clearly has at least one zero eigenvaluein the form of pµ: (
−ηµνp2 + pµpν)pµ = 0
and since the determinant of an operator is the product of its eigenvalues the determinant must be zero, and theoperator non-invertible.
Our gauge eld Aµ has four degrees of freedom (µ ∈ [0, 3]), whereas real photons only have two degrees offreedom (the two possible polarizations). The additional redundancy is known as a gauge freedom, and needs to berestricted. This can be done by modifying the action, for example by adding
SGauge fix [Aµ] =
d4x12ξ
(∂µAµ)2.
When ξ → 0 the Lorenz gauge ∂µAµ = 0 is rigidly enforced. The eect of this additional term on the Euler Lagrange
equation is (−ηµνp2 +
(1− 1
ξ
)pµpν
)Gνρ = δρµ
which is now invertible for all ξ <∞. The solution is (it's a `stick in and check' job rather than a `calculate'):
Gµν =−1p2
(ηµν + (ξ − 1)
pµpν
p2
).
Picking a ξ now xes a gauge: ξ = 0 is the Lorenz gauge, ξ = 1 the Feynman gauge, ξ = 3 the Yennie gauge (usefulfor a specic problem). Physical results do not depend on the choice of gauge, which merely reects a mathematicalredundancy built into the equations. Gauge xing is a vital procedure in gauge theories, as this section has hopefullydemonstrated.
1.7 Interaction Terms
The eld theories considered so far have all been quadratic in the elds, with no theory featuring multiple eldtypes. We can easily (`easily' !) nd the propagators for such theories, and then we can nd the response to anyexternal current we choose to apply. There is no back-reaction on the current from the eld.
QFT comes into its own when we consider interactions. These could be between two elds, such as in quantumelectrodynamics (QED):
SQED [Aµ, ψ] =
d4x
[−1
4FµνF
µν + ψ (γµ (pµ − eAµ)−m)ψ]
or of a eld with itself, as in ϕ4 theory:
Sϕ4 [ϕ] =
d4x
[12∂µϕ∂
µϕ− 12m2ϕ2 +
λ
4!ϕ4
]or the whimsically-named sine-Gordon theory:
Ssine-Gordon [ϕ] =
d4x
[12∂µϕ∂
µϕ− 12m2ϕ2 + cos (ϕ)
].
There is a good reason to claim that all the non-interacting (i.e. Gaussian) theories, including Maxwell's theory oflight, are classical. This is heavily debated, and is a point we will return to in later chapters after we have moretools at our disposal.
9
2 Partition Functions
The working of this section again sticks very closely to the notes of the Oxford Theoretical Physics course by JohnChalker and André Lukas, which I have linked to on my website [1].
2.1 Gaussian Integrals
Consider the classic Gaussian integral
I = ∞
−∞dv exp
(−1
2av2
)which is solved by squaring to give a 2D Gaussian integral, changing variables to plane polar coördinates, andsquare rooting:
I2 = ∞
−∞dv1dv2 exp
(−1
2a(v21 + v2
2
))I2 =
2π
0
dθ
∞
0
drr exp(−1
2ar2)
I2 =2πa
I =
√2πa.
We could have written the rst line as
I2 = ∞
−∞d2|v〉 exp
(−1
2〈v|A|v〉
)for 2D vector |v〉 , (v1, v2)
Tand 2 × 2 matrix A2 ,
(a 00 a
). In this case the result is perhaps more naturally
written
I2 = (2π)1/2 (det (A2))−1/2
where the subscript indicates a 2D space. Generalizing, we see that
In = (2π)n/2 (det (An))−1/2
(the factor of n is taken care of naturally in the second term by the determinant). In eect we are solvingn-dimensional Gaussian integrals. We can also generalize to arbitrary non-diagonal matrices provided all theeigenvalues are positive denite, since a unitary transformation will diagonalize again:
J =
dn|v〉 exp(−1
2〈v|An|v〉
)↓ An = U†DnU
J =
dn|v〉 exp(−1
2(〈v|U†
)Dn (U |v〉)
)↓ |w〉 = U |v〉
J =
dn|w〉 |det (U)| exp(−1
2〈w|Dn|w〉
)= (2π)n/2 (det (Dn))
−1/2
≡ (2π)n/2 (det (An))−1/2
where the modulus of the Jacobian for the variable change is given by |det (U)| ≡ 1, and the determinant of thediagonal matrix Dn , diag (a1, a2, . . . , an) and the original matrix An are identical since U is unitary.
10
We can additionally introduce a complete set of orthonormal basis states via a resolution of the identity:
I ≡n∑i=1
|ei〉〈ei|
giving
(2π)n/2 (det (An))−1/2 =
dn|v〉 exp
n∑i,j
−12〈v|ei〉〈ei|An|ej〉〈ej |v〉
.
In fact there is nothing to stop us taking the dimension of the vectors n→∞; the matrix An→∞ then becomesan operator. The innite dimensional vectors become elds |v〉 → ϕ, the innite sums of basis vectors becomeprojections into a continuous basis (such as position space or momentum space; I'll pick position)
∑ni=1 |ei〉〈ei| →
dx|x〉〈x|, and the innite product of integrals is known as a functional integral. The combination is written
Z =
Dϕ exp(
dx
dy
(−1
2ϕ (x)A (x, y)ϕ (y)
)).
The symbol Z suggests a partition function from statistical mechanics, and indeed this is the name we give theobject4:
Z =
Dϕ exp (iS [ϕ]) partition function.
There is a deep link between statistical mechanics and quantum eld theory which we will not have time to gointo. In much the same manner that switching t → it in the Schrödinger equation produces the heat equation,switching t→ it in QFT produces statistical mechanics. The process is known as Wick rotation - a π/2 rotation inthe complex time plane invented by Gian-Carlo Wick.
2.2 The First Trick in the Book
Let's re-introduce the current into the partition function and work in momentum space:
Z[j]
=
Dϕ exp
(i
d4p
(2π)4
(−1
2ϕG−1ϕ+ jϕ
))and for deniteness we'll take the Green's function to be the Klein-Gordon propagator
G (p) =1
m2 − p2.
The current term is linear in the eld, so can be eliminated by completing the square (combining it intothe quadratic term). The neatest way to do this, and the one which generalizes most straightforwardly to morecomplicated elds, is to change eld variables:
ϕ (p) = Φ (p) + φ (p)
where Φ is some xed conguration, and φ is a variation about this conguration, so that
Dϕ = D φ.
The result is
Z[j]
=
D φ exp
(i
d4p
(2π)4
(−1
2ΦG−1Φ− 1
2φG−1φ− 1
2ΦG−1φ− 1
2φG−1Φ + j
(Φ + φ
))).
In the case of the Klein Gordon eld the inverse propagator G−1 = + m2 has two spacetime derivatives in therst term meaning that if we do two by-parts integrals we can show
4Eagle-eyed observers may note that an i has been introduced. The Gaussian integrals converge for complex arguments providedthe real part of the exponent is negative.
11
d4x
(−1
2φG−1Φ
)=
d4x
(−1
2ΦG−1φ
).
In fact this should be true in most cases: the quadratic term in the action almost always comes from the kineticenergy term in the Lagrangian, and this term is a function of the derivative of the eld rather than the eld itself,so is actually quadratic in the derivative of the eld. Continuing, then,
Z[j]
= exp
(i
d4p
(2π)4
(−1
2ΦG−1Φ + jΦ
))D φ exp
(i
d4p
(2π)4
(−1
2φG−1φ− ΦG−1φ+ jφ
))
and if we choose the xed, `mean' eld to be
Φ = jG
the nal two terms cancel, giving the result
Z[j]
= exp
(i
d4p
(2π)412jGj
)Z0 (3)
where Z0 , Z [j ≡ 0] is the standard Gaussian, which we can do. In fact, the innite number which comes fromZ0 is of little interest, and we treat Z0 as a normalization. This innite number derives from the innite `zeropoint energy' of the free eld. It's one of the many innities in QFT, but in this case a reasonable one: the totalenergy of an innitely large eld is going to be innity, but it's only energy dierences we're interested in.
Note that the rst trick in the book really just amounts to completing the square.
2.3 Functional Averages
We dene the functional average of a functional O [ϕ] to be
〈O〉 ,
DϕO [ϕ] exp (iS [ϕ])
Dϕ exp (iS [ϕ])= Z −1
0
DϕO [ϕ] exp (iS [ϕ]) . (4)
In the case of QFT, with S some action, such a functional average weighted by exp (iS) is known as the `vacuumexpectation value' (VEV). You will sometimes see it written
〈O〉 ≡ 〈0|O|0〉 ≡ 〈Ω|O|Ω〉 .
In the latter two cases the ket |0〉 or |Ω〉 corresponds to the vacuum state, suggesting the interpretation that theVEV is the amplitude for the operator O to take the vacuum state back to itself.
One functional average of particular interest is the `2-point correlator'
〈ϕ (x2)ϕ (x1)〉 = Z −10
Dϕϕ (x2)ϕ (x1) exp (iS [ϕ])
which represents the amplitude for a particle in the eld ϕ to propagate from spacetime point xµ1 to point xµ2 . Thereis a neat trick to evaluating such objects, making use of Equation 3. From the denition Z [j]
Z [j] =
Dϕ exp(−i1
2
d4x
d4yϕ (x)G−1 (x, y)ϕ (y) + i
d4xj (x)ϕ (x)
)where the form of the inverse Green's function has been generalized slightly to potentially be a function of twospacetime points (this is the most general form of propagator). We can carry out a functional dierentiation withrespect to j under the integral (a classic Feynman trick):
−i δZ [j]δj (x2)
=
Dϕϕ (x2) exp(−i1
2
d4x
d4yϕ (x)G−1 (x, y)ϕ (y) + i
d4xj (x)ϕ (x)
)and
12
x1 x2
-iG(x1,x2)
Figure 2: The 2-point correlator is equal to the propagator (up to a constant in this case through a choice ofnormalization). It gives the amplitude for a particle to propagate from spacetime point x1 to point x2.
(−i δ
δj (x1)
)(−i δ
δj (x2)
)Z [j] =
Dϕϕ (x2)ϕ (x1) exp
(−i1
2
d4x
d4yϕ (x)G−1 (x, y)ϕ (y) + i
d4xj (x)ϕ (x)
)meaning the 2-point correlator is given by(
−i δ
δj (x1)
)(−i δ
δj (x2)
)Z [j]Z0
∣∣∣∣j≡0
= 〈ϕ (x2)ϕ (x1)〉 .
From Equation 3 we have that
Z [j]Z0
= exp(i
d4x
d4y
12j (x)G (x, y) j (y)
)and the 2-point correlator evaluates to
〈ϕ (x2)ϕ (x1)〉 =(−i δ
δj (x1)
)(−i δ
δj (x2)
)exp
(i
d4x
d4y
12j (x)G (x, y) j (y)
)∣∣∣∣j≡0
=(−i δ
δj (x1)
) d4yG (x2, y) j (y) exp
(i
d4x
d4y
12j (x)G (x, y) j (y)
)∣∣∣∣j≡0
= −iG (x1, x2)
assuming that G (x, y) ≡ G (y, x) after the rst line5. Thus we see that the 2-point correlator is the propagatorfor the eld, which makes sense given that it propagates a particle excitation between 2 points. This is showndiagrammatically in Figure 2.
As an interesting aside, note the mathematical similarity of the j tricks to a Fourier transform:
Z −10
Dϕ
exp
(i
d4p
(2π)4
(−1
2
)ϕG−1ϕ
)exp
(i
d4p
(2π)4jϕ
)=
exp
(i
d4p
(2π)4
(+
12
)jGj
)
cf.
dx exp (is (x)) exp (ipx) = exp (is (p))
(curly brackets introduced to emphasise the quantity being transformed) and derivatives with respect to j bringdown a ϕ:
−i δ
δj (x)→ ϕ (x)
cf. − i∂
∂p→ x.
I've never seen much discussion of this, but I believe there is some use in loop quantum gravity.
5If we don't assume this we get − i2
(G (x1, x2) + G (x2, x1)) instead which is a bit clumsier to carry around.
13
2.4 Interaction Terms Again
Note that by the reasoning in the last section the functional average of an arbitrary functional O [ϕ] can be evaluatedby
〈O〉 ,
DϕO [ϕ] exp (iS [ϕ])
Dϕ exp (iS [ϕ])= O
[−i δδj
]Z [j]Z0
∣∣∣∣j≡0
since the functional derivative pulls down a ϕ each time it acts. The method is even more powerful than it mayappear, and it explains one of the mathematical paradoxes of QFT: given that pretty much the only functionalintegrals which are mathematically well-dened are Gaussians6, how can we construct interesting (i.e. interacting)QFTs? The answer is that we only ever need to do Gaussian integrals! We then employ the mean eld trick andfunctionally dierentiate under the functional integral to generate whatever functions we like.
To see how this works let's take another look at interaction terms. We'll take as an example ϕ4 theory:
W [j] ,
Dϕ exp (iS0 [ϕ] + iSint [ϕ] + iSj [ϕ])
Dϕ exp (iS0 [ϕ])= Z −1
0
Dϕ exp
(i
d4x
(−1
2∂µϕ∂
µϕ+m2
2ϕ2 +
λ
4!ϕ4 + jϕ
))where the normalization Z −1
0 is now included as standard. We could reinterpret this expression as being thefunctional average of the interaction term:
W0 =⟨
exp(i
d4x
λ
4!ϕ4
)⟩(where W0 , W [j ≡ 0]) in which case we can evaluate it as before using
W0 = exp
(i
d4x
λ
4!
(−i δδj
)4)
Z [j]Z0
∣∣∣∣∣j≡0
.
The answer hints at how we deal with interactions, since it suggests a series expansion in powers of λ involvingpropagators G. In general we have that
W0 = exp(iSint
[−i δδj
])Z [j]Z0
∣∣∣∣j≡0
.
The object W [j] is known as a `generating functional'; j (x) may have a physical interpretation, but its primaryuse is to allow us to generate quantities of interest by taking functional derivatives with respect to it. This is auseful trick throughout all of physics.
6Prof. Berry has worked out a set of other (generally quite strange) cases where the functional integral can be carried out exactlyusing the saddle-point approximation.
14
3 n-point Functions
The best reference for this section is again reference [1].
3.1 Wick's Theorem
To streamline the notation slightly let's dene
ϕ (x1) , ϕ1
j (x1) , j1δ
δj1, δj1
G (x1, x2) , G12
(the algebra gets a bit tedious otherwise). We already calculated the 2-point correlator and found that the resultwas the propagator for the eld. We found that
〈ϕ2ϕ1〉 = (−iδj1) (−iδj2) exp(i
d4x
d4y
12jxGxyjy
)∣∣∣∣j≡0
= −iG12.
What if we calculated the 3-point correlator? A quick check shows the answer to be zero, and the same for anyn-point function for odd n. The next non-zero value is the 4-point function:
〈ϕ1ϕ2ϕ3ϕ4〉 = (−iδj4) (−iδj3) (−iδj2) (−iδj1) exp(i
d4x
d4y
12jxGxyjy
)∣∣∣∣j≡0
=(−i δδj4
)(−i δδj3
)(−i δδj2
)(d4xGx1jx
)exp
(i
d4x
d4y
12jxGxyjy
)∣∣∣∣j≡0
=(−i δδj4
)(−i δδj3
)[−iG12 − i
d4xGx1jx
d4yGy2jy
]exp
(i
d4x
d4y
12jxGxyjy
)∣∣∣∣j≡0
...
= −G12G34 −G23G14 −G24G13
(if you follow the algebra you'll agree with the streamlined notation!). Up to a constant prefactor, then, the 4-pointfunction is the sum of possible contractions into 2-point correlators. In fact the result generalizes to all nonzerocorrelators. It is known as Wick's theorem:
〈ϕ1ϕ2 . . . ϕ2n〉 =i=2n, j=2n,...∑possiblepairwise
contractionsi = 1, j = 1, . . .
GijGkl . . . Gyz
3.2 The ϕ4 vacuum
We now have everything in place to introduce interactions formally. Consider again the object
W0 =⟨
exp(i
d4x
λ
4!ϕ4
)⟩where the exponential is dened through its Taylor series (and therefore the coupling λ is assumed small). Tosecond order in the coupling we have
15
x3 x4
x1 x2
x3 x4
x1 x2
x3 x4
x1 x2
x3 x4
x1 x2
= + +
Figure 3: The 4-point function decomposes into all possible 2-point functions. This is an instance of the moregeneral rule that n-point functions (i.e. any correlator one might care to imagine) decompose into products of2-point functions. This is known as Wick's theorem.
W0 =
⟨1 + i
λ
4!
d4xϕ4
x +(iλ
4!
)2 d4x
d4yϕ4
xϕ4y + . . .
⟩
= 1 + iλ
4!
d4x
⟨ϕ4x
⟩+(iλ
4!
)2 d4x
d4y
⟨ϕ4xϕ
4y
⟩+ . . .
which can be evaluated using Wick's theorem (and the results from before) to give
W0 = 1 + iλ
4!
d4x
(−3G2
xx
)+
12
(iλ
4!
)2 d4x
d4y
(32G2
xxG2yy + 4232GxxGyyG
2xy + 4!G4
xy
)= 1− i
λ
8
d4xG2
xx − λ2
d4x
d4y
(1
128G2xxG
2yy +
116GxxGyyG
2xy +
148G4xy
).
Let's look at the combinatorics regarding the coecients in the λ2 term, as this is the really ddly bit and probablywhence QFT gets its reputation for diculty. We want all possible contractions of
ϕxϕxϕxϕxϕyϕyϕyϕy
these split into three types:
ϕxϕx︸ ︷︷ ︸ ϕxϕx︸ ︷︷ ︸ ϕyϕy︸ ︷︷ ︸ ϕyϕy︸ ︷︷ ︸ϕxϕx︸ ︷︷ ︸ ϕyϕy︸ ︷︷ ︸ ϕxϕy︸ ︷︷ ︸ ϕxϕy︸ ︷︷ ︸ϕxϕy︸ ︷︷ ︸ ϕxϕy︸ ︷︷ ︸ ϕxϕy︸ ︷︷ ︸ ϕxϕy︸ ︷︷ ︸
in the rst case, the ϕx and ϕy combine separately; the rst ϕx has 3 others to choose from, then the other pair iscompletely constrained. The total number of ways of forming this contraction, including the ϕy terms, is therefore
32, and combined with the prefactor this gives 32
2×(4!)2= 1
128 . In the second case we want to form one ϕxϕx pair;
there are 6 ways to do this, and the same for the ϕyϕy. The remainder is ϕxϕxϕyϕy which is to be formed into
ϕxϕy pairs, and there are 2 ways to do this in total. So the total prefactor is 2×62
2×(4!)2= 1
16 . Finally we want to form
all possible ϕxϕy contractions. Let's allow the ϕx to choose partners from amongst the ϕy, just as the girls wereallowed to choose from the boys in a ceilidh I once attended. The rst ϕx has 4 ϕy to choose from, then the secondϕx has 3 ϕy, the third ϕx has 2 ϕy to choose from, and the last ϕx gets what she's given (in the ceilidh analogyshe would be partnered with myself). This gives 4! combinations, so the total prefactor this time is 4!
2×(4!)2= 1
48 .
The philosophy of what is happening here is very interesting. We are either nding the expectation value of theϕ4 interaction term with respect to the vacuum of the non-interacting theory, or equivalently we are nding theexpectation value of the vacuum with respect to the interacting ϕ4 theory. In the latter interpretation we see thatinteracting theories have innitely rich vacua: rather than being empty, the vacuum contains particles coming intoand out of existence in all possible ways which involve 4-vertices. Note that it's 4-vertices because we have a ϕ4
interaction. If we had instead the QED interaction term ∼ Aµψ†ψ the vertices would need one electron going in,
ψ†, one coming out, ψ, and one photon Aµ either coming in or going out. The interaction term species the onlytypes of vertex in the theory.
16
x
x
y
x y x yO(λ) O(λ2)
Gxx Gxx
Gxx
Gyy
Gxx
Gxx
GxxGyyGxy
Gxy
Gxy
Gxy
Gxy
Gxy
Figure 4: The ϕ4 vacuum to second order in the coupling. The vacuum consists of all possible ways to combinethe 4-vertex generated by the ϕ4 interaction.
x
O(λ)
Gxx Gxx
O(1)G12x1 x2
G12x1 x2
+
G1x
Gxx
Gx2
x1 x2
Figure 5: The 2-point correlator for ϕ4 theory, to rst order in the coupling. The rst of the O (λ) diagramscontains a bubble.
Another way to think of the result is that the functional integral
Dϕ varies the vacuum over all possible eldcongurations, with each conguration weighted by exp (iS [ϕ]). This is quite natural when thinking of QFT asan extension of the Feynman path integral method of QM: in that case we consider 1D paths f (x) between pointsin space and use the functional integral
Df to sum over all possible paths, with each path again weighted by
exp (iS [f ]). For this reason people sometimes refer to QM as a 0 + 1D QFT: it's a QFT with a `eld' x which is afunction only of time: x (t).
3.3 The ϕ4 n-point functions
In exactly the same way as we evaluated the vacuum in the last section we can evaluate n-point functions. The2-point function in ϕ4 theory, for example, is given by
〈ϕ (x2)ϕ (x1)〉λ =
Dϕϕ (x2)ϕ (x1) exp (iS0 + iSint)Dϕ exp (iS0)
where λ suggests the interacting theory rather than the theory with λ = 0, and to order λ2 this expands to
〈ϕ2ϕ1〉λ = 〈ϕ2ϕ1〉0 + iλ
4!
d4x
⟨ϕ2ϕ1ϕ
4x
⟩0
+(iλ
4!
)2 d4x
d4y
⟨ϕ2ϕ1ϕ
4xϕ
4y
⟩0
+O(λ3)
using the streamlined notation, and 〈. . .〉0 indicating the vacuum expectation value with respect to the non-interacting vacuum (the thing we calculated before) for clarity. Some work leads to the result (only rst orderin λ as second order already contains ten ϕ terms!)
〈ϕ2ϕ1〉λ = −iG12 + iλ
4!
d4x
(3iG12G
2xx + 12iG1xGxxGx2
)+O
(λ2).
The result is shown diagrammatically in Figure 5. The interaction term has added an innite number of correctionterms to the propagator, where terms with n vertices are suppressed by a factor
(λ4!
)n. These are known as `quantum
corrections'.So can non-interacting (i.e. Gaussian) theories be quantum? It depends on how you dene a quantum eld
theory as opposed to a classical one. One option might be to say that quantum theories are those where theprobability for an event is given by summing amplitudes for dierent processes before taking a modulus square.
17
a) b) c)
Figure 6: In non-interacting theories there are no vertices, so the only vacuum bubble possible is a closed loop(a). However, in the non-interacting Chern Simons theory the dierent embeddings of that loop count dierently,and the dierent possible knots sum as independent amplitudes: (b) shows a trefoil knot, and (c) a gure-of-eightknot, whereas (a) is known as the unknot.
In Gaussian theories there can only be one diagram per process: there can be no quantum corrections to anyn-point functions as there are no vertices7; similarly there is only one allowed vacuum bubble, a closed loop. So thisdenition doesn't distinguish quantum and classical for Gaussian models, although the lack of quantum correctionsdoes rather suggest Gaussian theories are classical. Another option is to consider second quantization which wewon't directly cover in this course. This is a rewriting of the elds as an innite set of harmonic oscillator creationand annihilation operators, each sat at a dierent point in space (or, more commonly, momentum space). It's arewriting of the eld as a collection of particles, where we interpret the mth excited state of the oscillator labelledby p as m particles of momentum p. Both interacting and Gaussian theories can be second quantized, but classicalelds cannot8. Probably the most agreed-upon denition of a quantum eld theory is that it obeys canonicalcommutation relations, [ϕ (x) , π (y)] = iδ (x− y), where π, the momentum conjugate to the eld ϕ, is dened to be
∂L
∂∂0ϕ (xµ)= π (xµ) .
This is ultimately the same as the second quantization argument. The question of whether Maxwell theory isquantum is still unresolved, though, because it is a very subtle matter to have it obey canonical quantization -involving gauge xing - and is beyond the scope of Peskin and Schroeder, even!
An advanced note for the interested reader: there is a class of Gaussian theories which contains inequivalentvacuum bubbles. How can this happen, given that the only allowed diagram is a closed loop? The answer is thatthis class of theories, called topological QFTs, cares how the loop is embedded in the space. That is, it can tell whatkind of knot the loop is tied in (see Figure 6). The dierent knots add as amplitudes, so this theory is undoubtedlyquantum. Chern Simons theory is the classic example (super-advanced note: to be Gaussian the theory must beeither Abelian, or non-Abelian in the Weyl gauge A0
a = 0).
3.4 Feynman Diagrams
Feynman diagrams have already been appearing throughout these notes as I assume most people have some famil-iarity with them already (and they're very intuitive).
Aside from giving a neat interpretation to the maths in terms of particle creation an annihilation, the realstrength of Feynman diagrams comes when you can use them to jump straight to the answer rather than having tocarry out the complicated combinatorics of the functional derivative method we've employed so far.
As noted in Section 3.1, Wick's theorem states that the vacuum expectation value of any combination of eldsis always a product of propagators (the VEV decomposes into the sum of the VEVs of possible pairs of elds). AnyVEV can therefore be expressed in terms of the propagator and vertices. First let's take the general case for a realscalar eld:
S [ϕ] =
d4x
d4y
12ϕ (x)G−1 (x, y)ϕ (y) +
d4x
∑n>2
αn1n!λnϕn
7Astute readers will have noticed that the 4-point function for a non-interacting theory has 3 diagrams; we will see shortly that noneof these contribute to physical processes.
8Reductio ad absurdum for the reader familiar with second quantization (say, pages 19-21 of Peskin and Schroeder): classical elds
commute with their conjugate momenta, [ϕ (x) , π (y)] = 0, implying their ladder operators commute,hak, a†
q
i= 0. But then
hH, a†
k
i= 0
and a†k is not a ladder operator. QED.
18
Figure 7: The gure-of-eight bubble diagram.
where the interaction term has been expressed as a Taylor series, which is always possible. The ϕ4 case hasα4 = 1, α6=4 = 0. Then:
• for each xed external eld at spacetime point xµi we assign a xed vertex
• for each appearance of an internal n-vertex we assign −iαnλn
• for each instance of an internal line between points xµ and yµ we assign Gxy (if there are multiple elds inthe theory each one will have a propagator; they may be trivial, i.e. just a number)
• we integrate over all internal spacetime variables
• for each diagram we multiply all of the above elements together and divide by the corresponding symmetryfactor.
You might wonder where (i) the factor from the Taylor expansion of the operator O and (ii) the 1n! from the
interaction have gone. For (i), at order m the Taylor series for O will have a prefactor 1/m!, but this will cancelwith the m! degeneracy due to interchanging the m vertices. For (ii), there will be n lines heading into each vertex,giving an n! degeneracy, cancelling the 1/n! included in the denition above (for this reason). Both prefactors aretherefore already accounted for.
For example, let's take the gure-of-eight bubble diagram appearing in Figures 4 and 5, reproduced in Figure7. This has no external elds, one 4-vertex (since we were considering ϕ4 theory this was the only allowed vertex),and two propagators going into (or out of) the same internal point (call it xµ, but it's a dummy variable as we'llintegrate over it since it's internal), so two factors of Gxx and an integral over d4x. The symmetry factor is thenumber of diagrams equivalent to the one drawn. In this case either of the two lines could be reversed, or the twolines could be exchanged, and we wouldn't know - so the factor is 2× 2× 2. Thus the diagram evaluates to
−λ8
d4xGxxGxx.
The method is generally a lot simpler in momentum space, where the propagator should be a simple function ofone variable (assuming the problem has translational symmetry, which it does unless lattice eects are important9).There's the added advantage that in experiments we'll generally x the ingoing and outgoing momenta rather thanpositions, so objects like
〈ϕ (p2) ϕ (p1)〉
are a lot more natural than
〈ϕ (x2)ϕ (x1)〉 .
There's a major health warning accompanying the use of Feynman diagrams - you have to be extremely procientto properly calculate the symmetry factors. It's far easier to make a mistake in the factors by this method thanit is in the lengthy combinatoric method, or the even lengthier full method of taking δj functional derivatives.The method I use is to draw all the diagrams I can think of at the given order in the coupling, then to applythe combinatoric method, then try to see the symmetries in the diagrams afterwards. I think this is probably astandard way to go about things.
9Lattice eects will always prove important at some point - in condensed matter the spacing is the atomic spacing, and in particlephysics it's around the Planck length. On these scales the eld picture breaks down, and we need to deal with the eects carefully.
19
a)
b)
c)
d) e)
Figure 8: Common conventions for drawing Feynman diagrams: (a) bosonic propagator (e.g. photon γ, phonon
ϕ, W or Z boson) , (b) fermionic propagator (e.g. electron ψ† or c†, proton), (c) gluon propagator, (d) QED-typevertex (say, an electron and positron combine to a photon), (e) ϕ4-type vertex (two ϕ particles scatter into twomore, or four annihilate at a point). Being strict we should strike through the propagators in (d) and (e) becausethey're not themselves included in the vertices. In practice we often don't bother. Quite disturbingly, cutting oexternal legs in this manner is called `amputation'.
The interpretation, due to Feynman and Stückelberg, of these diagrams as particles moving through spacetimeis powerful but easily abused. To tell the story properly we need to introduce the propagators for more-interestingelds, and some more vertices, since real scalar elds are their own antiparticles. Complex elds allow for an-tiparticles as the complex conjugate of the eld, i.e. if ϕ∗ (x) creates a particle at point x then ϕ (x) creates anantiparticle, or equivalently annihilates a particle. The propagator for fermion elds (spin-12 or potentially 3
2 ,52 etc.)
is drawn as a line with an arrow on it, indicating the ow of fermion number. Since fermion number is conservedthe divergence of arrows at a vertex is always zero (if one goes in, one must go out). Some common conventionsare shown in Figure 8.
Bosons need not be conserved, so there are no arrows on bosonic propagators10. The lack of an arrow means inpractice one has to add an arrow next to the line to indicate the direction of momentum transfer, otherwise signerrors may be introduced.
With all this in place we can interpret the diagrams as sequences of spacetime events. This is probably mosteasily done by considering some examples, which I do in Figure 9. An important point to note is that the additionto the diagrams of axes for space and time (in either orientation), which is commonly seen at undergraduate level,is totally misleading. There's no requirement that the spacetime events appearing in the diagrams be timelikeseparated, so the idea of a consistent time-ordering to the events is false. Internal spacetime points are integratedover anyway, so occur at all possible points in spacetime. The labels may be useful in certain limited cases, forexample if we x a number of external spacetime events in a given frame (probably the lab), but this is a rare case.In most practical cases the boundary conditions are 4-momenta anyway.
The interpretation of antiparticles as particles running backwards in time is acceptable but not necessary. Theargument derives from complex conjugating the time-dependent Schrödinger equation:
i∂tψ (x, t) = Hψ (x, t)−i∂tψ∗ (x, t) = Hψ∗ (x, t)
∴ i∂tψ∗ (x,−t) = Hψ∗ (x,−t)
i.e. if ψ solves the Schrödinger equation with variable t, ψ∗ is a solution with −t. Still, in relativistic theories theonly really sensible quantities are spacetime events, so the idea of time `owing' in either direction is an additional,arguably incorrect, assertion. The quantum eld might have a nonzero correlation 〈ϕ (xµ)ϕ (yν)〉 between twospacetime events xµ and yν , but the interpretation of this as a particle being created at one event and propagatingto the other, or an antiparticle moving between the points in the opposite sense, is additional. Note that the `timereversal ≡ complex conjugation' argument doesn't hold for the Klein-Gordon equation(
+m2)ϕ = 0.
That equation governs bosonic scalar elds - even if the elds are complex, time reversal has no eect.
10This includes gluons, which get a special propagator symbol. I'm not entirely certain why this is, but one important feature ofgluons is that their self-interactions (they have a 3-vertex and a 4-vertex) are extremely important because of quark connement and`asymptotic freedom'. It's possible the helical lines' vertices were considered easier to draw.
20
a)
b)
c)
xµyν
pµ
p'ν
pµ - p'ν
Figure 9: Three Feynman diagrams. In (a) a bosonic particle propagates from spacetime point xµ to point yν . Ifwe pick a frame of reference in which x0 < y0 the Feynman-Stückelberg interpretation would be to say the particletravels from x to y, or alternatively the antiparticle travels backwards in time from y to x. In (b) it is externalmomenta that are xed rather than spacetime locations. An electron (say) with 4-momentum pµ annihilates apositron with 4-momentum p′ν , creating a photon (say) with 4-momentum pµ − p′ν . Note that vertices alwaysconserve 4-momentum like this even when they're internal. An alternative interpretation could be that the electronemits the photon, and thus changes its 4-momentum. There is no need to interpret either fermion line as eitherparticle or antiparticle. In (c) we have a vacuum bubble for a ϕ4 interaction. We could say four particles arecreated at the vertex and annihilate in pairs; that two particle/antiparticle pairs are created then all annihilate;that a particle/antiparticle pair are created, separate then come together to annihilate, and as they do they hitinto another particle/antiparticle pair doing the same but coming from the future into the past, etc. etc. . In casethis isn't general enough to show that it's rather arbitrary how to interpret the diagrams, note that the vertex isinternal so gets integrated over - this could be done either in real space or momentum space with the same result(dummy variable).
a) b) c) d)
Figure 10: A connected diagram has all external legs connected. This does not rule out bubble diagrams: (a) aconnected diagram; (b) a disconnected diagram; (c) a connected diagram with a bubble; (d) a disconnected diagramwith a bubble. Unfortunately there's another usage of the word `disconnected' which does mean diagrams withbubbles.
3.5 Cancelling Disconnected Terms
When calculating physically measurable quantities such as scattering amplitudes, it turns out that only diagramswhere all external legs are connected together contribute. These diagrams are `connected'. We could simply throwaway disconnected terms, and often do in practice, but to be a little more mathematically rigorous we can insteadconsider a dierent generating functional which only generates connected diagrams. Note that the denition of`connected' does not rule out bubble diagrams. Some clarication is given in Figure 10.
The relevant generating functional turns out to be simply
V [j] , lnW [j]
where W is the full generating functional considered before. Note that this form causes:
δV [j]δj
=1
W [j]δW [j]δj
= W −1j
δW [j]δj
.
Now, recalling our streamlined notation:
δ1Zj = 〈ϕ1〉j
21
Figure 11: The connected 4-point function for ϕ4 theory consists of the full 4-point function minus the discon-nected diagrams. It is generated by the connected generating functional Vj .
and for the 4-point function
δ4δ3δ2δ1Zj = 〈ϕ1ϕ2ϕ3ϕ4〉jand so on for higher n-point functions. Well actually we're going to use even more streamlined notation where thesequantities now read
δ1Zj = 〈ϕ1〉j , 1
δ4δ3δ2δ1Zj = 〈ϕ1ϕ2ϕ3ϕ4〉j , 1234.
This notation is not totally standard, but you'll nd people including yourself will invent dierent ways to dealwith the more tedious calculations. Remember that 1, for example, is still a functional of j so has a nontrivialfunctional derivative: δ21 = 12 and so on.What happens when we consider instead our new generating functional Vj? Well, after some working:
δ4δ3δ2δ1 ln (Zj) = δ4δ3δ2Z−1j 1
= δ4δ3[−Z −2
j 2 · 1 + Z −1j 12
]... (skip to the end...)
= Z −1j
[1234−Z −1
j
(1 · 234 + 2 · 134 + 3 · 124 + 4 · 123
)−Z −1
j
(12 · 34 + 13 · 24 + 14 · 23
)+2Z −2
j
(1 · 2 · 34 + 1 · 3 · 24 + 1 · 4 · 23 + 2 · 3 · 14 + 2 · 4 · 13 + 3 · 4 · 12
)−2Z −3
j 1 · 2 · 3 · 4].
When we now send j to zero, and note that a choice of 〈ϕ〉 = 0 is sensible under normal circumstances (not inthe Higgs eld...), and return to the normally streamlined notation, we have
δ4δ3δ2δ1 ln (Zj) = 〈ϕ1ϕ2ϕ3ϕ4〉 − (〈ϕ1ϕ2〉 〈ϕ3ϕ4〉+ 〈ϕ1ϕ3〉 〈ϕ2ϕ4〉+ 〈ϕ1ϕ4〉 〈ϕ2ϕ3〉)
that is, using V instead of W gives exactly the same result but with the disconnected diagrams removed. The resultis given diagrammatically in Figure 11.
22
4 Renormalization
This topic is one of the cornerstones of QFT. In fact we've done a lot of the work already. The classic referencescover a lot of this as well [4, 2, 5], in particular Chapter 33 of [3]. The reader interested in the discussion ofasymptotic series in Section 4.1 could read the book by Dingle available for free on Michael Berry's website [6], orindeed any number of papers by Berry himself. Mode-Mode Coupling, in Section 4.3, is a specic example of amore general idea and is not of wide relevance in itself.
4.1 Motivation: Asymptotic Series
An innite number of mathematicians walk into a pub. The rst orders a pint. The second orders ahalf-pint, the third orders a quarter-pint, and so on. After a while the Landlord says I see where this isgoing. Here's two pints: sort it out amongst yourselves.
Say we want to nd the innite sum
S = 1 + r + r2 + r3 + . . .
the trick is to use the fact that subtracting one and dividing by r returns the original sum again:
S − 1r
= 1 + r + r2 + r3 + . . .
≡ S
therefore, rearranging,
S (1− r) = 1
S =1
1− r.
Hence for r = 12 , S = 2, and the Landlord was correct.
By all rights we can only expect the method to work up to the radius of convergence of the series, |r| = 1. To gobeyond the radius of convergence requires the techniques of analytic continuation. However, should you choose tosum the series with these techniques (e.g. Borel summation) for, say, r = 2, you in fact get the same result: in thatcase, S = −1. Summing divergent series is possible, and in fact it's often faster than summing convergent series.
An innite number of mathematicians walk into a pub. The rst orders a pint. The second orders twopints, the third four pints, and so on. After a while the Landlord says I see where this is going. Youowe me a pint.
The relevance to QFT is that the sum of the innite series of Feynman diagrams is not actually convergent. Infact it is not divergent either; it is known as an Asymptotic Series. It converges for some time, then begins todiverge. There is an excellent book available on asymptotic series on Michael Berry's website [6]. There's animportant question as to where we should truncate the series, since adding up as many terms as possible doesn'teven necessarily get a better answer. For practical QFT matters the options are generally these:
• truncate at zeroeth order (mean eld theory)
• truncate at rst order (quantum corrections)
• truncate at second order (more quantum corrections - often very tricky)
• sum a subset of the diagrams to innity
• truncate the series at a physically relevant point.
We have already dealt with the rst three points, at least in principle. In the following sections I will deal with thenal two.
23
a) b)
Figure 12: (a)The rst order quantum correction to the phonon propagator with an electron-phonon coupling
gϕψ†ψ. (b) Upon amputating the external legs (again, apologies for the macabre phraseology) the resulting diagramis known as the electronic susceptibility.
= +
+
+
+
Figure 13: The RPA renormalization of the phonon propagator with the electron-phonon coupling. The inniteseries of diagrams includes no correlations between internal electron-hole loops.
4.2 The Random Phase Approximation
The Random Phase Approximation, or RPA, is a method of summing a certain kind of Feynman diagram to innity.The name generally refers to problems with electron-phonon coupling, but the QED electron-photon vertex is almostidentical.
Say we want to change the phonon propagator by adding contributions from higher-order diagrams. Physicallywe would like to include some eect of the electrons. This is a form of `renormalization'. The rst-order correction,of order g2, is shown in Figure 12. If we amputate the external phonon lines the resulting electron-hole loop11 isknown as the electronic susceptibility χ.
If we dene the bare (unrenormalized) phonon propagator with `4-momentum' q = (Ω,q) to be D0 (q), and theelectron propagator G (k), the expression is
D (q) = D0 (q) + g2D0 (q)
[∑k
G (k)G (k + q)
]D0 (q)
= D0 (q) + g2D0 (q)χ (q)D0 (q) .
We can keep adding these loops, provided we ignore correlations between the internal electron lines which aredoing the renormalization. The electrons' phases are thus in some sense `randomized', although I think the nameRandom Phase Approximation is mostly an historical artefact. The diagrams are shown in Figure 13.The resulting series is then (dropping q dependence for convenience)
D = D0 + g2D0χD0 + g4D0χD0χD0 + g6D0χD0χD0χD0 + . . .
and we can do the sum to innity using the Landlord's technique:
g−2χ−1D−10 (D −D0) = D0 + g2D0χD0 + g4D0χD0χD0 + . . .
≡ D
where in general the propagators may be complicated matrices or tensors so cannot be assumed to commute.Re-arranging, the result is
11As before, we could equally interpret this an electron-electron loop and so on; the most common convention here is to interpret thediagram as a phonon breaking up into an electron and hole which then recombine.
24
Figure 14: The next-order (i.e. g4) diagrams above RPA, known as the mode mode coupling diagrams. A physicalreason for going beyond RPA might be that electron correlations (mediated by the phonon eld) are of particularimportance.
D =(1− g2D0χ
)−1D0
which is the RPA approximation for the innite renormalization of the phonon propagator. The phonon propagatorin momentum space takes the form (bare frequency Ω0 (q) and `Matsubara frequency' Ωn):
D0 (q) =−2Ω0
(iΩn)2 − Ω2
0
so in this case the RPA renormalized propagator reads
D =−2Ω0
(iΩn)2 − Ω2
0 + 2Ω0g2χ
=−2Ω0
(iΩn)2 − Ω2
RPA
where the RPA renormalized phonon frequency is given by
Ω2RPA (q) = Ω2
0 (q)− 2g2Ω0 (q)χ (q) .
The bare phonon frequency had some normalization, and the process of adding quantum corrections changed it- hence, renormalization. Renormalization goes a lot further than what we'll see in these notes, and in fact plays animportant part in allowing QFTs to generate accurate results even though they're mathematically poorly-dened.
4.3 Mode Mode: Beyond RPA
The RPA works very well in many cases, as might be expected of an innite sum. The nal option, on my list ofpractical ways to truncate Feynman diagram series, was to truncate at a physically relevant point. I will give as anexample a case where I've needed to go beyond RPA, still in a theory with electron-phonon coupling. In this casewe were arguing that the phonon eld was driving a phase transition, and the contributions of the phonons werevital. The lowest order in which phonons inuence the susceptibility itself is second order, and the correspondingdiagrams are known as the Mode-Mode Coupling approximation, MMA. They are shown in Figure 14.
Rather than go through the details of the calculation I will simply show the eect on the renormalization ofthe phonon frequency in Figure 15. It can be seen that the RPA renormalizes the frequency down; when therenormalized frequency hits zero the material becomes structurally unstable and a phase transition occurs (it costszero energy to excite a phonon of that frequency, so the corresponding lattice distortion is permanent). However,the MMA diagrams bring the frequency back up again. Physically the diagrams correspond to uctuations of thephonon eld, which make it harder for the phonons to introduce order.
As a nal, philosophical, note on renormalization, it is the renormalized quantities we actually deal with in thereal world. The bare propagators are mathematical abstractions, ideal particles which are completely undisturbedby the rest of the universe. Sometimes it is necessary to infer their properties from the particles we see and, asusual, we begin to encounter innities. To take a popular example which is part analogy, part reality (maybe `fable'is best): the electric eld of an electron diverges as you approach it even in the classical theory
E (r) =Q
r2
suggesting an innite self-energy at r = 0. The fable tells us that the charge is screened by vacuum polarization:the QED vertex allows renormalization of quantities such as the observed charge, and the measured result includesthe contributions of the renormalization. It's misleading to now start talking about Feynman diagrams since these
25
Phonon Dispersion
0
12
Energ
y (
meV
)
M
MMA
RPA
Figure 15: This plot, showing theoretical results on the material NbSe2, shows the physical eect of the dierentdiagrams to the phonon energy. The plot is against momentum-transfer by the phonons, q, along a particular high-symmetry direction in the Brillouin zone. The RPA diagrams lower the phonon energy towards a phase transition,but the MMA diagrams negate the eect to some extent by disordering the system through phonon uctuations.
are perturbative and you'd want to sum the full innite series to get a real result - it's even more misleading to startinterpreting the diagrams in terms of virtual particles, so we'll stay away from the discussion. What can be said isthat the innity you started from is quite unphysical; the only measurable quantity, the renormalized quantity, isnite.
4.4 Dyson Series: Mass Renormalization
The RPA result of Section 4.2 is actually a special case of a trick which works for arbitrary interactions and arbitrarypropagators. I will pick for deniteness the Klein-Gordon propagator
D0 (p) =−1
p2 −m2. (5)
Note that the propagator diverges whenever p2 = m2, i.e. when E2 = p2 +m2. Normally we would always expectthis to be true, but internal lines are not required to obey this condition. When they do they are said to be `on-shell'.It is a general property of propagators that they contain a divergence (actually a `pole' in the complex p plane)corresponding to the mass of the particle they represent.In general the corrections to the bare propagator are known as its self-energy, Σ (p). These can be more generalthan the RPA, although this is a particular form of self-energy correction. We can still dene a perturbation serieswith self-energy corrections just as in Section 4.2:
D = D0 +D0ΣD0 +D0ΣD0ΣD0 +D0ΣD0ΣD0ΣD0 + . . .
which in this general case is known as a Dyson series. Rearranging with the Landlord's technique (I can now revealthat the Landlord was in fact Freeman Dyson):
D = (1−D0Σ)−1D0
=1
D−10 − Σ
.
Substituting the specic bare propagator we nd in this case that
D (p) =−1
p2 −m2 + Σ. (6)
26
Comparing the bare and renormalized propagators (Equations 5 and 6) we see that the self-energy has shifted thelocation of the pole to m2 − Σ. Since the pole dictates the mass of the particle this means the self-energy hasrenormalized the mass. This is a general result: except in special cases, self-interactions of a particle shift its mass.
A brief note on QED: the same reasoning might suggest the mass of the photon may renormalize under interactionwith electrons, yet we know the renormalized photons we see are truly massless. In some cases symmetries protectthe propagators from renormalizing, or rather the renormalization causes no eect. This is the case with the photonpropagator, which is protected by a `Ward identity'. Ward identities are an important feature of QFTs, and amountto making the theory gauge invariant. We won't have time to go through them in this course.
27
5 Removing and Introducing Fields
In this last section I will introduce two important tricks. The rst is how to approximate a theory of multiple eldsby removing one of them, arriving at an `eective action'. This is a very common procedure mathematically, andgets to the heart of QFT philosophically. The second is a neat mathematical trick for introducing new elds whichproves of use with complicated interactions.
I'm not familiar with an especially good reference for Section 5.1 so the standard references will have to dotheir usual good job [4, 2, 5]. I have some excellent notes written by A. Melikyan and M. R. Norman on theHubbard-Stratanovitch transformation in Section 5.2. They are not available online so please email me if you wouldlike a copy.
5.1 Integrating Out Fields: the Eective Action
Say we have a QFT which contains two independent elds. I will use the QED-type coupling, but for two real scalarelds12. This is sometimes referred to as scalar-QED:
W = Z −1ϕ Z −1
η
DϕDη exp
(iS0ϕ + iS0
η + iSint [ϕ, η])
where the non-interacting partition functions for the two elds are Zϕ and Zη, and
S0ϕ [ϕ] ,
d4x
d4y
(−1
2ϕ (x)A (x, y)−1
ϕ (y))
S0η [η] ,
d4x
d4y
(−1
2η (x)B (x, y)−1
η (y))
Sint [ϕ, η] ,
d4x (gϕ (x) η (x) η (x))
so the propagators for the elds are A and B. I've used the symbol W for the interacting partition function asbefore.To integrate out one of the elds is to nd an eective action featuring only the other eld. This requires us toapproximate one of the theories to Gaussian form and perform the respective functional integral. First we willintegrate out the η eld.
Expanding the interaction term as a Taylor series, reverting to the shorter notation ϕ (x) = ϕx etc. the partitionfunction is
W = Z −1ϕ Z −1
η
DϕDη exp
(−i1
2
d4x
d4y
(ϕxA
−1xy ϕy + ηxB
−1xy ηy
)+ ig
d4xϕxηxηx
)= Z −1
ϕ Z −1η
DϕDη exp (iSϕ) exp (iSη)
(1 + ig
d4xϕxηxηx
+12!
(ig)2
d4x
d4yϕxηxηxϕyηyηy
+13!
(ig)3
d4x
d4y
d4zϕxηxηxϕyηyηyϕzηzηz + . . .
).
By the denition of the functional average we're hopefully getting quite used to by now this is equivalent to
W = Z −1ϕ
Dϕ exp (iSϕ)
(1 + ig
d4xϕx 〈ηxηx〉η
+12!
(ig)2
d4x
d4yϕxϕy 〈ηxηxηyηy〉η
+13!
(ig)3
d4x
d4y
d4zϕxϕyϕz 〈ηxηxηyηyηzηz〉η
)12Although I don't know of a physical system to which this theory applies it avoids the complications of gauge elds and Grassman
elds associated with photons and electrons respectively. This course is practical in terms of methods rather than applications!
28
with the η subscript indicating the functional average with respect to that eld. The results are
W = Z −1ϕ
Dϕ exp (iSϕ)
(1 + ig
d4xϕxBxx
+12!
(ig)2
d4x
d4yϕxϕy
(BxxByy + 2B2
xy
)+
13!
(ig)3
d4x
d4y
d4zϕxϕyϕz (BxxByyBzz
+2BxxB2yz + 2ByyB2
xz + 2BzzB2xy + 4BxyByzBzx
))or, rewriting slightly,
W = Z −1ϕ
Dϕ exp (iSϕ)
(1 + ig
d4xϕxBxx +
12!
(ig
d4xϕxBxx
)2
+13!
(ig
d4xϕxBxx
)3
+22!
(ig)2
d4xd4yϕxϕyB2xy
+ +43!
(ig)3
d4x
d4y
d4zϕxϕyϕzBxyByzBzx
+63!
(ig)3
d4x
d4y
d4zϕxϕyϕzBxxB
2yz + . . .
).
As the rst four terms suggest, the series is equivalent to the Taylor series expansion of an exponential to thirdorder; in fact this is true to all orders in the expansion, and allows us to `re-exponentiate':
W = Z −1ϕ
Dϕ exp (iSϕ) exp
(ig
d4xϕxBxx + i2g2
d4xd4yϕxϕyB
2xy
+(ig)3
d4x
d4y
d4zϕxϕyϕz
(BxxB
2yz +
23BxyByzBzx
)+O
(g4))
.
The linear and cubic terms disappear in this case if we assume the amplitude to propagate nowhere, Bxx, is zero13.
The linear term is in fact such a common occurrence it gets its own name: the Tadpole diagram (draw it to seewhy)14. You will encounter many pieces of otsam and jetsam like this when navigating the Dirac sea, and inpractical applications they are discarded immediately. Terms of this form cannot conserve momentum. In general,though, higher order terms in the series will lead to higher order interaction terms in the eective action, where
W ≈ Z −1ϕ
Dϕ exp (iSeff [ϕ])
that is, an approximation has been made in which the η eld has been eliminated. Examining the form for thisspecic interaction, to this order, we see
Seff =
d4xd4y
(−1
2ϕxA
−1xy ϕy + ig2ϕxϕyB
2xy
)+
23
(ig)3
d4x
d4y
d4zϕxϕyϕzBxyByzBzx +O
(g4)
= −12d4xd4yϕx
(A−1xy − 2ig2B2
xy
)ϕy +
23
(ig)3
d4x
d4y
d4zϕxϕyϕzBxyByzBzx +O
(g4)
so the lowest order eect here is a simple renormalization of the free ϕ propagator. The next order term introducesa cubic interaction between ϕ particles.
The eective action is ubiquitous in applications of QFT. Of particular interest is its application to many modernideas about the universe, including the standard model of particle physics. The general idea is that elds importantat high energy scales may lose relevance on everyday scales, and can be integrated out to give low-energy eective
13This isn't always the case, for example if the propagator is a constant.14Lancaster and Blundell [3] note that Sidney Coleman, who invented the name `tadpole', also suggested `spermion'.
29
theories. The particles we see may not be the `fundamental' particles, whatever they are, but the remains afterwe've approximated away interactions with other elds we don't deal with directly day-to-day. By way of anotherfable, think of the Brownian motion of a pollen particle in water. The particle drifts with time - we know thisis the result of a random walk caused by individual water molecules colliding with the particle. We can see thepollen move but not the water molecules responsible; they're too `high energy' in the sense of being too fast-movingand too small (recall [Length] = E−1 so small ∼ high energy, where `∼' means `you get the general idea'). Wecan construct an eective theory for the motion by integrating out the `water molecule eld' and just maintainingpollen, whose motion is renormalized by the process.
5.2 Inserting New Fields: the Hubbard Stratanovitch Transformation
Our nal useful trick allows us to simplify certain types of complicated interaction at the expense of introducing a neweld. The new eld can often then be ascribed a physical interpretation. As we're familiar with the manipulationsof QFT at this stage I'm going to break my one rule and move away from real scalar elds, introducing insteadfermionic (anticommuting) Grassman elds. The reason I think this is okay here is that the new elds will coupleinto a real bosonic eld, meaning their anticommuting property does not complicate things.
This time the action describes electrons on a lattice. The eld operator ψ†σi (t) creates an electron with spinσ ∈ ↑, ↓ on site i. As we're on a lattice there's no Lorentz invariance and the operators carry a time label t. Wewill introduce a repulsive interaction between electrons on the same site15:
S[ψ†, ψ
]=
dt∑i
ψ†iG−1i ψi + g
∑i
ψ†↑iψ↑iψ†↓iψ↓i.
Just like in the harmonic oscillator in ordinary QM the operator ni , ψ†i ψi is the number operator. Now we'reworking with elds there is a dierent one for each site i. The Pauli exclusion principle disallows double-occupationof the same site if the electrons have the same spin, i.e. nσinσi ≡ 0.
First we recall the denition of the free partition function, this time for a eld with a constant propagator g:
Z∆ =
D∆ exp
(− i
2g
dt∑i
∆2i
).
In an ordinary innite integral∞−∞ dx we are free to change variables to x′ = x + a and leave the limits
unchanged. The same applies to the functional integral, which is eectively always `innite', so we can shift theeld to
∆′i = ∆i +
∑σ
ψ†σiψσi
D∆′ = D∆
giving
Z∆ =
D∆′ exp
(− i
2g
dt∑i
(∆′i −∑σ
ψ†σiψσi
)(∆′i −∑σ
ψ†σiψσi
))and expanding (also relabel ∆′ back to ∆)
Z∆ =
D∆ exp
(− i
2g
dt∑i
∆i∆i
)exp
(ig
dt∑σi
∆iψ†σiψσi
)exp
(− i
2g
dt∑σσ′i
ψ†σiψσiψ†σ′iψσ′i
)
nally, noting that if σ = σ′ Pauli exclusion sends the quartic term to zero, then rearranging slightly we see
exp
(ig
dt∑i
ψ†↑iψ↑iψ†↓iψ↓i
)= Z −1
∆
D∆ exp
(− i
2g
dt∑i
∆i∆i
)exp
(ig
dt∑σi
∆iψ†σiψσi
).
15Unlike scalar elds the Grassman eld doesn't need a − 12normalization for the propagator.
30
a)
b)
c)
Figure 16: The Hubbard Stratanovitch transformation removes a 4-fermion vertex, (a), at the cost of introducinga real bosonic eld (b) and an interaction between the two, (c).
This is precisely what we need to eliminate the quartic term from the partition function of the original theory:
Z = Z −1ψ
Dψ†Dψ exp
(i
dt∑i
ψ†iG−1i ψi + ig
dt∑i
ψ†↑iψ↑iψ†↓iψ↓i
)
= Z −1∆ Z −1
ψ
Dψ†DψD∆ exp
(i
dt∑i
ψ†iG−1i ψi
)exp
(− i
2g
dt∑i
∆i∆i
)exp
(ig
dt∑σi
∆iψ†σiψσi
)
= Z −1∆ Z −1
ψ
Dψ†DψD∆ exp
(iSHS
[ψ†, ψ, ∆
])where the Hubbard-Stratanovitch action is given by
SHS
[ψ†, ψ, ∆
]=
dt∑i
ψ†iG−1i ψi −
12g
dt∑i
∆i∆i + g
dt∑σi
∆iψ†σiψσi.
Thus we see that we have swapped a quartic interaction between electrons for a real bosonic eld ∆ with atrivial propagator, and our old friend the QED-type interaction vertex. The situation is shown diagrammaticallyin Figure 16.
It is important to note that the equivalence between the theories is exact - no approximations have entered at all.
One use of the transformation is that the expectation value of the ∆ eld, ∆ =⟨∆⟩, has a natural interpretation
as a Cooper pair order parameter. This is easiest seen by noting that ∆ took the place of ψ†σψσ in a somewhatapproximate sense.
5.3 Epilogue
After discussion with various friends and colleagues I feel it necessary to mention some of the concerns they haveraised. First I should make clear that this introduction has assumed `perturbative' QFT. The most up-to-dateresearch goes beyond perturbative methods in various ways. For an important recent discovery in this directionlook up `amplitudihedrons'. Contrary to my opening claim that The key object in a QFT is the action, there existQFTs which have no Lagrangian description, particularly in M-theory. Numerical QFT, including lattice QCD, haslong been able to consider theories away from Gaussian xed points. Renormalization is the more modern way toview QFT, and provides an explanation and more rigorous justication for many features which appeared to workwithout our understanding, contradicting my claims in the introduction. In particular Haag's theorem, that onecannot consistently dene the vacuum in interacting theories, is now considered to have been the wrong way to lookat the problem, and largely disappears with considerations of RG ow.
31
