Quantum Field Theory

Uwe-Jens Wiese Institute for Theoretical Physics University of Bern August 21, 2007

Contents
1 Introduction 2 From Mechanics to Quantum Field Theory 2.1 2.2 2.3 2.4 2.5 2.6 2.7 7 11

From Point Mechanics to Classical Field Theory . . . . . . . . . . 11 The Path Integral in Real Time . . . . . . . . . . . . . . . . . . . . 13 The Path Integral in Euclidean Time . . . . . . . . . . . . . . . . . 17 Spin Models in Classical Statistical Mechanics . . . . . . . . . . . . 18 Quantum Mechanics versus Statistical Mechanics . . . . . . . . . . 20 The Transfer Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Lattice Field Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 24 29

3 Classical Scalar Field Theory 3.1 3.2

Scalar Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 Noethers Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 35

4 Canonical Quantization of a Scalar Field 4.1 4.2 4.3

From the Lagrange to the Hamilton Density . . . . . . . . . . . . . 35 Commutation Relations for Scalar Field Operators . . . . . . . . . 36 Hamilton Operator in Scalar Quantum Field Theory . . . . . . . . 37 3

4 4.4 4.5

CONTENTS Vacuum and Particle States . . . . . . . . . . . . . . . . . . . . . . 38 The Momentum Operator . . . . . . . . . . . . . . . . . . . . . . . 40 41

5 Path Integral for Scalar Field Theory 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 5.9

From Minkowski to Euclidean Space-Time . . . . . . . . . . . . . . 41 Euclidean Propagator and Contraction Rule . . . . . . . . . . . . . 43 Perturbative Expansion of the Path Integral . . . . . . . . . . . . . 44 Dimensional Regularization . . . . . . . . . . . . . . . . . . . . . . 45 The 2-Point Function to Order . . . . . . . . . . . . . . . . . . . 46 Mass Renormalization . . . . . . . . . . . . . . . . . . . . . . . . . 49 Connected and Disconnected Diagrams . . . . . . . . . . . . . . . . 49 Feynman Rules for 4 Theory . . . . . . . . . . . . . . . . . . . . . 51 The 4-Point Function to Order 2 . . . . . . . . . . . . . . . . . . 53

5.10 Dimensional Regularization of J(p2 ) . . . . . . . . . . . . . . . . . 55 5.11 Renormalization of the Coupling Constant . . . . . . . . . . . . . . 56 5.12 Renormalizable Scalar Field Theories . . . . . . . . . . . . . . . . . 57 6 Canonical Quantization of Electrodynamics 6.1 6.2 6.3 6.4 61

From the Lagrange to the Hamilton Density . . . . . . . . . . . . . 61 The Hamilton Operator for the Photon Field . . . . . . . . . . . . 63 Vacuum and Photon States . . . . . . . . . . . . . . . . . . . . . . 64 Electromagnetic Momentum Operator . . . . . . . . . . . . . . . . 65 67

7 Path Integral for Scalar Electrodynamics 7.1 7.2

Gauge Fixing and Photon Propagator . . . . . . . . . . . . . . . . 67 Feynman Rules for Scalar QED . . . . . . . . . . . . . . . . . . . . 69

CONTENTS 8 Lattice Field Theory 8.1 8.2 8.3 8.4 8.5 8.6 8.7 8.8 8.9

5 73

Fermionic Path Integrals and Grassmann Algebras . . . . . . . . . 73 The Fermion Doubling Problem . . . . . . . . . . . . . . . . . . . . 74 The Nielsen-Ninomiya No-Go Theorem . . . . . . . . . . . . . . . . 76 Wilson Fermions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 Abelian Lattice Gauge Fields . . . . . . . . . . . . . . . . . . . . . 78 The Notion of Lattice Dierential Forms . . . . . . . . . . . . . . . 79 Wilson loops and the Lattice Coulomb Potential . . . . . . . . . . 81 Lattice QED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 Lattice QCD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

8.10 Connement in the Strong Coupling Limit . . . . . . . . . . . . . . 87 8.11 Connement in Compact Abelian Gauge Theory . . . . . . . . . . 88 8.12 The Monte Carlo Method . . . . . . . . . . . . . . . . . . . . . . . 94

CONTENTS

Chapter 1

Introduction
Quantum mechanics describes the behavior of particles at atomic and subatomic scales. It is the quantum version of Newtons classical mechanics and reduces to this theory when is sent to zero. Of course, Plancks quantum is a fundamental constant of Nature that is indeed non-zero. This has consequences not only for mechanics but for all of physics. In particular, Maxwells theory of classical electrodynamics also needs to be modied by incorporating the principles of quantum physics. It turned out that this is a rather nontrivial enterprise which kept people like Feynman, Schwinger, and Tomonaga busy for some time after world war II. Their success was awarded with the Nobel prize for quantum electrodynamics (QED) the quantum eld theory that describes the electromagnetic interactions of electrons and positrons resulting from the exchange of photons. The fundamental degrees of freedom in QED are elds not particles. In particular, photons the elementary particles of light emerge as quantum states of Maxwells electromagnetic eld. Similarly, electrons and positrons emerge as quanta of the 4-component spinor eld that Dirac rst introduced. In quantum eld theory, particles result as quantized eld uctuations. The most important feature of eld theory is locality. The eld degrees of freedom located at a given point of space are coupled only to the eld values at innitesimally neighboring points. The principle of locality is also at the heart of relativity theory which is inconsistent with instantaneous nonlocal interaction over a nite distance. Unifying relativity and quantum physics therefore naturally leads to quantum eld theory; it cannot be achieved within quantum mechanics alone. In relativity theory, particle interactions cannot be described by instantaneous potentials but must be mediated locally. This is possible if there are physical degrees of freedom everywhere is space. These degrees of freedom 7

CHAPTER 1. INTRODUCTION

are nothing but the eld values. Quantum eld theory has enormously advanced our understanding of physics. It was originally developed to describe elementary particles but it is equally useful in condensed matter physics. In fact, it is extremely useful for describing any local interaction of many degrees of freedom at microscopic scales. Our present understanding of fundamental physics is based on the eld concept. At each point of space there is a set of eld degrees of freedom whose uctuations lead to the propagation of a variety of elementary particles. Each particle has its own eld: the photon, the electron and positron, the gluons, the quarks and anti-quarks, the W- and Z-bosons, the neutrinos and anti-neutrinos, etc. These various elds are distinguished by their transformation properties under a variety of symmetries: space-time rotations, parity P , charge conjugation C, gauge transformations, etc. In fact, the various quantum eld theories are characterized by their symmetry properties. For example, QED is a gauge theory with the Abelian gauge group U (1)em which respects the symmetries C and P . Quantum chromodynamics (QCD) is the quantum eld theory that describes the strong interactions between quarks and anti-quarks mediated by the exchange of gluons. Its structure is similar to the one of QED. QCD is a gauge theory with non-Abelian gauge group SU (3)c which also respects the symmetries C and P . Both QED and QCD are incorporated in the standard model of particle physics which summarizes all we know about the fundamental forces of electromagnetism, as well as the weak and strong interactions (but not gravity). The standard model is a gauge theory with the non-Abelian gauge group SU (3)c SU (2)L U (1)Y which, however, strongly violates parity P and charge conjugation C. Even the combined symmetry CP is weakly broken. Unlike QED and QCD, the standard model is a so-called chiral gauge theory in which left- and right-handed particles carry dierent charges. The quantization of eld theories is a highly non-trivial issue. After all, quantum eld theories are quantum systems with innitely many degrees of freedom a given number per space point. This enormity of degrees of freedom gives rise to ultraviolet divergences which must be regularized and renormalized. Learning eld theory is a non-trivial and sometimes rather technical task, which will, however, throw you right in the middle of current research in particle and condensed matter physics. Quantum eld theory is such a powerful tool that it is indispensable in these elds of physics. You simply need to master quantum eld theory if you want to understand let alone contribute to these very exciting elds of current research. It will naturally take a while before you can call yourself a quantum eld theorist. It requires some time and a lot of hard work, but the awards are plenty. At the end of your studies you will have a basis for under-

9 standing all that is currently known about the microscopic quantum world. The regularization of chiral gauge theories like the standard model is a topic of current research. For example, only a few years ago, Lscher from CERN achieved u a breakthrough by quantizing a chiral gauge theory beyond perturbation theory. Naturally, when learning quantum eld theory, we will start with much simpler eld theories than the full standard model, or even QED. In fact, we will start with a simple relativistic theory of a scalar eld. Scalar elds play an important role in both particle and condensed matter physics. The yet to be discovered Higgs particle a corner stone of the standard model is described by a scalar eld. Similarly, scalar elds are used to describe Cooper pairs in superconductors or the order parameter in superuid helium. Later we will move on to fermionic elds as well as to gauge elds. Even at the end of the course, we will not have reached the quantization of non-Abelian gauge theories like QCD or the standard model. However, by then you will hopefully have a solid basis in quantum eld theory, which will allow you to proceed to these more advanced topics whenever a research project requires it.

10

CHAPTER 1. INTRODUCTION

Chapter 2

From Mechanics to Quantum Field Theory

This chapter provides a brief summary of the mathematical structure of quantum eld theory. Classical eld theories are discussed as a generalization of point mechanics to systems with innitely many degrees of freedom a given number per space point. Similarly, quantum eld theories are just quantum mechanical systems with innitely many degrees of freedom. In the same way as point mechanics systems, classical eld theories can be quantized with path integral methods. The quantization of eld theories at nite temperature leads to path integrals in Euclidean time. This provides us with an analogy between quantum eld theory and classical statistical mechanics. We also mention the lattice regularization which has recently provided a mathematically satisfactory formulation of the standard model beyond perturbation theory.

2.1

From Point Mechanics to Classical Field Theory

Point mechanics describes the dynamics of classical nonrelativistic point particles. The coordinates of the particles represent a nite number of degrees of freedom. In the simplest case a single particle moving in one spatial dimension we are dealing with a single degree of freedom: the x-coordinate of the particle. The dynamics of a particle of mass m moving in an external potential V (x) is described by Newtons equation
2 mt x = ma = F (x) =

dV (x) . dx

(2.1.1)

11

12

CHAPTER 2. FROM MECHANICS TO QUANTUM FIELD THEORY

Once the initial conditions are specied, this ordinary second order dierential equation determines the particles path x(t), i.e. its position as a function of time. Newtons equation results from the variational principle to minimize the action S[x] = dt L(x, t x), (2.1.2)

over the space of all paths x(t). The action is a functional (a function whose argument is itself a function) that results from the time integral of the Lagrange function m L(x, t x) = (t x)2 V (x). (2.1.3) 2 The Euler-Lagrange equation t L L = 0, (t x) x (2.1.4)

is nothing but Newtons equation. Classical eld theories are a generalization of point mechanics to systems with innitely many degrees of freedom a given number for each space point x. In this case, the degrees of freedom are the eld values (x), where is some generic eld. In case of a neutral scalar eld, is simply a real number representing one degree of freedom per space point. A charged scalar eld, on the other hand, is described by a complex number and hence represents two degrees of freedom per space point. The scalar Higgs eld a (x) (with a {1, 2}) in the standard model is a complex doublet, i.e. it has four real degrees of freedom per space point. An Abelian gauge eld Ai (x) (with a spatial direction index i {1, 2, 3}) for example, the photon eld in electrodynamics is a neutral vector eld with 3 real degrees of freedom per space point. One of these degrees of freedom is redundant due to the U (1)em gauge symmetry. Hence, an Abelian gauge eld has two physical degrees of freedom per space point which correspond to the two polarization states of the massless photon. Note that the time-component A0 (x) does not represent a physical degree of freedom. It is just a Lagrange multiplier eld that enforces the Gauss law. A non-Abelian gauge eld Aa (x) is charged and i has an additional index a. For example, the gluon eld in chromodynamics with a color index a {1, 2, ..., 8} represents 2 8 = 16 physical degrees of freedom per space point, again because of some redundancy due to the SU (3)c color gauge symmetry. The eld that represents the W - and Z-bosons in the standard model has an index a {1, 2, 3} and transforms under the gauge group SU (2)L . Thus, it represents 2 3 = 6 physical degrees of freedom. However, in contrast to the photon, the W - and Z-bosons are massive due to the Higgs mechanism and have three (not just two) polarization states. The extra degree of freedom is provided by the Higgs eld.

2.2. THE PATH INTEGRAL IN REAL TIME

13

The analog of Newtons equation in eld theory is the classical eld equation of motion. For example, for a neutral scalar eld this is the Klein-Gordon equation = dV () . d (2.1.5)

Again, after specifying appropriate initial conditions it determines the classical eld conguration (x), i.e. the values of the eld at all space-time points x = (t, x). Hence, the role of time in point mechanics is played by space-time in eld theory, and the role of the point particle coordinates is now played by the eld values. As before, the classical equation of motion results from minimizing the action S[] = d4 x L(, ). (2.1.6) The integral over time in eq.(2.1.2) is now replaced by an integral over spacetime and the Lagrange function of point mechanics gets replaced by the Lagrange density function (or Lagrangian) L(, ) = 1 V (). 2 (2.1.7)

A simple interacting eld theory is the 4 theory with the potential V () = m2 2 4 + . 2 4! (2.1.8)

Here m is the mass of the scalar eld and is the coupling strength of its selfinteraction. Note that the mass term corresponds to a harmonic oscillator potential in the point mechanics analog, while the interaction term corresponds to an anharmonic perturbation. As before, the Euler-Lagrange equation L L = 0, ( ) (2.1.9)

is the classical equation of motion, in this case the Klein-Gordon equation. The analogies between point mechanics and eld theory are summarized in table 2.1.

2.2

The Path Integral in Real Time

The quantization of eld theories is most conveniently performed using the path integral approach. Here we rst discuss the path integral in quantum mechanics

14

CHAPTER 2. FROM MECHANICS TO QUANTUM FIELD THEORY Point Mechanics time t particle coordinate x particle path x(t) action S[x] = dt L(x, t x) Lagrange function L(x, t x) = m (t x)2 V (x) 2 equation of motion L t (t x) L = 0 x Newtons equation 2 t x = dV (x) dx kinetic energy m (t x)2 2 harmonic oscillator potential m 2 x2 2 anharmonic perturbation 4! x4 Field Theory space-time x = (t, x) eld value eld conguration (x) action S[] = d4 x L(, ) Lagrangian 1 L(, ) = 2 V () eld equation L ( ) L = 0 Klein-Gordon equation = dV () d kinetic energy 1 2 2 mass term m 2 2 self-interaction term 4! 4

Table 2.1: The dictionary that translates point mechanics into the language of eld theory. quantized point mechanics using the real time formalism. A mathematically more satisfactory formulation uses an analytic continuation to so-called Euclidean time. This will be discussed in the next section. The real time evolution of a quantum system described by a Hamilton operator H is given by the time-dependent Schrdinger equation o i t |(t) = H|(t) . (2.2.1)

For a time-independent Hamilton operator the time evolution operator is given by i (2.2.2) U (t , t) = exp( H(t t)), such that Let us consider the transition amplitude x |U (t , t)|x of a nonrelativistic point particle that starts at position x at time t and arrives at position x at time t . Using x|(t) = (x, t) (2.2.4) we obtain (x , t ) = dx x |U (t , t)|x (x, t), (2.2.5) |(t ) = U (t , t)|(t) . (2.2.3)

2.2. THE PATH INTEGRAL IN REAL TIME

15

i.e. x |U (t , t)|x acts as a propagator for the wave function. The propagator is of physical interest because it contains information about the energy spectrum. When we consider propagation from an initial position x back to the same position we nd x|U (t , t)|x = =
n

i x| exp( H(t t))|x i | x|n |2 exp( En (t t)). (2.2.6)

We have inserted a complete set,

|n n| = , of energy eigenstates |n with (2.2.7)

H|n = En |n .

Hence, according to eq.(2.2.6), the Fourier transform of the propagator yields the energy spectrum as well as the energy eigenstates x|n . Inserting a complete set of position eigenstates we arrive at x |U (t , t)|x = = = i x | exp( H(t t1 + t1 t))|x i dx1 x | exp( H(t t1 ))|x1 i x1 | exp( H(t1 t))|x dx1 x |U (t , t1 )|x1 x1 |U (t1 , t)|x . (2.2.8)

It is obvious that we can repeat this process an arbitrary number of times. This is exactly what we do in the formulation of the path integral. Let us divide the time interval [t, t ] into N elementary time steps of size such that t t = N . (2.2.9)

Inserting a complete set of position eigenstates at the intermediate times ti , i {1, 2, ..., N 1} we obtain x |U (t , t)|x = dx1 dx2 ... dxN 1 x |U (t , tN 1 )|xN 1 ... (2.2.10)

x2 |U (t2 , t1 )|x1 x1 |U (t1 , t)|x .

In the next step we concentrate on one of the factors and we consider a single nonrelativistic point particle moving in an external potential V (x) such that H= p2 + V (x). 2m (2.2.11)

16

CHAPTER 2. FROM MECHANICS TO QUANTUM FIELD THEORY

Using the Baker-Campbell-Haussdor formula and neglecting terms of order 2 we nd xi+1 |U (ti+1 , ti )|xi ip2 i ) exp( V (x))|xi 2m 1 i ip2 = )|p p| exp( V (x))|xi dp xi+1 | exp( 2 2m i ip2 1 ) exp( p(xi+1 xi )) dp exp( = 2 2m i (2.2.12) exp( V (xi )). = xi+1 | exp(

The integral over p is ill-dened because the integrand is a very rapidly oscillating function. To make the expression well-dened we replace the time step by ia, i.e. we go out into the complex time plane. After doing the integral we take the limit a 0. Still one should keep in mind that the denition of the path integral required an analytic continuation in time. One nds xi+1 |U (ti+1 , ti )|xi = i m xi+1 xi 2 m exp( [ ( ) V (xi )]). 2i 2 (2.2.13)

Inserting this back into the expression for the propagator we obtain x |U (t , t)|x = i Dx exp( S[x]). (2.2.14)

The action has been identied in the time continuum limit as S[x] = dt [
0

= lim

m (t x)2 V (x)] 2 m xi+1 xi 2 [ ( ) V (xi )]. 2

(2.2.15)

The integration measure is dened as Dx = lim m 2i

N

dx1

dx2 ...

dxN 1 .

(2.2.16)

This means that we integrate over the possible particle positions for each intermediate time ti . In this way we integrate over all possible paths of the particle starting at x and ending at x . Each path is weighted with an oscillating phase factor exp( i S[x]) determined by the action. The classical path of minimum action has the smallest oscillations, and hence the largest contribution to the path integral. In the classical limit 0 only that contribution survives.

2.3. THE PATH INTEGRAL IN EUCLIDEAN TIME

17

2.3

The Path Integral in Euclidean Time

As we have seen, it requires a small excursion into the complex time plane to make the path integral mathematically well-dened. Now we will make a big step into that plane and actually consider purely imaginary so-called Euclidean time. The physical motivation for this, however, comes from quantum statistical mechanics. Let us consider the quantum statistical partition function Z = Tr exp(H), (2.3.1)

where = 1/T is the inverse temperature. It is mathematically equivalent to the time interval we discussed in the real time path integral. In particular, the operator exp(H) turns into the time evolution operator U (t , t) if we identify = i (t t). (2.3.2)

In this sense the system at nite temperature corresponds to a system propagating in purely imaginary (Euclidean) time. By dividing the Euclidean time interval into N time steps, i.e. by writing = N a/ , and again by inserting complete sets of position eigenstates we now arrive at the Euclidean time path integral Z= 1 Dx exp( SE [x]). (2.3.3)

The action now takes the Euclidean form SE [x] = = dt [

a0

lim

m (t x)2 + V (x)] 2 m xi+1 xi 2 a[ ( ) + V (xi )]. 2 a

(2.3.4)

In contrast to the real time case the measure now involves N integrations Dx = lim m 2 a
N

a0

dx1

dx2 ...

dxN .

(2.3.5)

The extra integration over xN = x is due to the trace in eq.(2.3.1). Note that there is no extra integration over x0 = x because the trace implies periodic boundary conditions in the Euclidean time direction, i.e. x0 = xN . The Euclidean path integral allows us to evaluate thermal expectation values. For example, let us consider an operator O(x) that is diagonal in the position

18

CHAPTER 2. FROM MECHANICS TO QUANTUM FIELD THEORY

state basis. We can insert this operator in the path integral and thus compute its expectation value O(x) = 1 1 Tr[O(x) exp(H)] = Z Z 1 Dx O(x(0)) exp( SE [x]). (2.3.6)

Since the theory is translation invariant in Euclidean time one can place the operator anywhere in time, e.g. at t = 0 as done here. When we perform the low temperature limit, , the thermal uctuations are switched o and only the quantum ground state |0 (the vacuum) contributes to the partition function, i.e. Z exp(E0 ). In this limit the path integral is formulated in an innite Euclidean time interval, and describes the vacuum expectation value O(x) = 0|O(x)|0 = lim 1 Z 1 Dx O(x(0)) exp( SE [x]). (2.3.7)

It is also interesting to consider 2-point functions of operators at dierent instances in Euclidean time O(x(0))O(x(t)) = = 1 Tr[O(x) exp(Ht)O(x) exp(Ht) exp(H)] Z 1 1 (2.3.8) Dx O(x(0))O(x(t)) exp( SE [x]). Z

Again, we consider the limit , but we also separate the operators in time, i.e. we also let t . Then the leading contribution is | 0|O(x)|0 |2 . Subtracting this, and thus forming the connected 2-point function, one obtains
,t

lim O(x(0))O(x(t)) | O(x) |2 = | 0|O(x)|1 |2 exp((E1 E0 )t). (2.3.9)

Here |1 is the rst excited state of the quantum system with an energy E1 . The connected 2-point function decays exponentially at large Euclidean time separations. The decay is governed by the energy gap E1 E0 . In a quantum eld theory E1 corresponds to the energy of the lightest particle. Its mass is determined by the energy gap E1 E0 above the vacuum. Hence, in Euclidean eld theory particle masses are determined from the exponential decay of connected 2-point correlation functions.

2.4

Spin Models in Classical Statistical Mechanics

So far we have considered quantum systems both at zero and at nite temperature. We have represented their partition functions as Euclidean path integrals

2.4. SPIN MODELS IN CLASSICAL STATISTICAL MECHANICS

19

over congurations on a time lattice of length . We will now make a completely new start and study classical discrete systems at nite temperature. We will see that their mathematical description is very similar to the path integral formulation of quantum systems. Still, the physical interpretation of the formalism is drastically dierent in the two cases. In the next section we will set up a dictionary that allows us to translate quantum physics language into the language of classical statistical mechanics. For simplicity, let us concentrate on simple classical spin models. Here the word spin does not mean that we deal with quantized angular momenta. All we do is work with classical variables that can point in specic directions. The simplest spin model is the Ising model with classical spin variables sx = 1. (Again, these do not represent the quantum states up and down of a quantum mechanical angular momentum 1/2.) More complicated spin models with an O(N ) spin rotational symmetry are the XY model (N = 2) and the Heisenberg model (N = 3). The spins in the XY model are 2-component unit-vectors, while the spins in the Heisenberg model have three components. In all these models the spins live on the sites of a d-dimensional spatial lattice. The lattice is meant to be a crystal lattice (so typically d = 3) and the lattice spacing has a physical meaning. This is in contrast to the Euclidean time lattice that we have introduced to make the path integral mathematically well-dened, and that we nally send to zero in order to reach the Euclidean time continuum limit. The Ising model is characterized by its classical Hamilton function (not a quantum Hamilton operator) which simply species the energy of any conguration of spins. The Ising Hamilton function is a sum of nearest neighbor contributions H[s] = J sx sy B sx ,
x

(2.4.1)

xy

with a ferromagnetic coupling constant J < 0 that favors parallel spins, plus a coupling to an external magnetic eld B. The classical partition function of this system is given by Z= Ds exp(H[s]/T ) = exp(H[s]/T ).
x sx =1

(2.4.2)

The sum over all spin congurations corresponds to an independent summation over all possible orientations of individual spins. Thermal averages are computed by inserting appropriate operators. For example, the magnetization is given by sx = 1 Z sx exp(H[s]/T ).
x sx =1

(2.4.3)

20

CHAPTER 2. FROM MECHANICS TO QUANTUM FIELD THEORY

Similarly, the spin correlation function is dened by sx sy = 1 Z sx sy exp(H[s]/T ).

x sx =1

(2.4.4)

At large distances the connected spin correlation function typically decays exponentially sx sy s 2 exp(|x y|/), (2.4.5) where is the so-called correlation length. At general temperatures the correlation length is typically just a few lattice spacings. When one models real materials, the Ising model would generally be a great oversimplication, because real magnets, for example, not only have nearest neighbor couplings. Still, the details of the Hamilton function at the scale of the lattice spacing are not always important. There is a critical temperature Tc at which diverges and universal behavior arises. At this temperature a second order phase transition occurs. Then the details of the model at the scale of the lattice spacing are irrelevant for the long range physics that takes place at the scale of . In fact, at their critical temperatures some real materials behave just like the simple Ising model. This is why the Ising model is so interesting. It is just a very simple member of a large universality class of dierent models, which all share the same critical behavior. This does not mean that they have the same values of their critical temperatures. However, their magnetization goes to zero at the critical temperature with the same power of Tc T , i.e. their critical exponents are identical.

2.5

Quantum Mechanics versus Statistical Mechanics

We notice a close analogy between the Euclidean path integral for a quantum mechanical system and a classical statistical mechanics system like the Ising model. The path integral for the quantum system is dened on a 1-dimensional Euclidean time lattice, just like an Ising model can be dened on a d-dimensional spatial lattice. In the path integral we integrate over all paths, i.e. over all congurations x(t), while in the Ising model we sum over all spin congurations sx . Paths are weighted by their Euclidean action SE [x] while spin congurations are weighted with their Boltzmann factors depending on the classical Hamilton function H[s]. The prefactor of the action is 1/ , and the prefactor of the Hamilton function is 1/T . Indeed determines the strength of quantum uctuations, while the temperature T determines the strength of thermal uctuations. The kinetic energy 1 2 2 ((xi+1 xi )/a) in the path integral is analogous to the nearest neighbor spin coupling sx sx+1 , and the potential term V (xi ) is analogous to the coupling Bsx

2.6. THE TRANSFER MATRIX Quantum mechanics Euclidean time lattice elementary time step a particle position x particle path x(t) path integral Dx Euclidean action SE [x] Plancks constant quantum uctuations kinetic energy 1 ( xi+1 xi )2 2 a potential energy V (xi ) weight of a path exp( 1 SE [x]) vacuum expectation value O(x) 2-point function O(x(0))O(x(t)) energy gap E1 E0 continuum limit a 0 Classical statistical mechanics d-dimensional spatial lattice crystal lattice spacing classical spin variable s spin conguration sx sum over congurations x sx classical Hamilton function H[s] temperature T thermal uctuations neighbor coupling sx sx+1 external eld energy Bsx Boltzmann factor exp(H[s]/T ) magnetization sx correlation function sx sy inverse correlation length 1/ critical behavior

21

Table 2.2: The dictionary that translates quantum mechanics into the language of classical statistical mechanics. to an external magnetic eld. The magnetization sx corresponds to the vacuum expectation value of an operator O(x) and the spin-spin correlation function sx sy corresponds to the 2-point correlation function O(x(0))O(x(t)) . The inverse correlation length 1/ is analogous to the energy gap E1 E0 (and hence to a particle mass in a Euclidean quantum eld theory). Finally, the Euclidean time continuum limit a 0 corresponds to a second order phase transition where . The lattice spacing in the path integral is an artifact of our mathematical description which we send to zero while the physics remains constant. In classical statistical mechanics, on the other hand, the lattice spacing is physical and hence xed, while the correlation length goes to innity at a second order phase transition. All this is summarized in the dictionary of table 2.2.

2.6

The Transfer Matrix

The analogy between quantum mechanics and classical statistical mechanics suggests that there is an analog of the quantum Hamilton operator in the context of classical statistical mechanics. This operator is the so-called transfer matrix.

22

CHAPTER 2. FROM MECHANICS TO QUANTUM FIELD THEORY

The Hamilton operator induces innitesimal translations in time. In classical statistical mechanics, on the other hand, the analog of continuous time is a 1dimensional spatial lattice. Hence, the transfer matrix cannot induce innitesimal space translations. Instead it induces translations by the smallest possible distance namely by one lattice spacing. For a quantum mechanical system the transfer matrix transports us by one lattice spacing in Euclidean time, and it is given by a (2.6.1) T = exp( H). Now we want to construct the transfer matrix for the 1-dimensional Ising model without an external magnetic eld. The corresponding partition function is given by Z=
x sx =1

exp(J
x

sx sx+1 ).

(2.6.2)

The transfer matrix obeys Z = TrT N , (2.6.3)

where N is the number of lattice points, and its matrix elements are given by the Boltzmann factor corresponding to a nearest neighbor pair by sx+1 |T |sx = exp(Jsx sx+1 ). (2.6.4)

This is a 2 2 matrix. The eigenvalues of the transfer matrix can be written as exp(E0 ) and exp(E1 ). The energy gap then determines the inverse correlation length as 1/ = E1 E0 . (2.6.5) It is instructive to compute as a function of to locate the critical point of the 1-d Ising model. Here we will do the corresponding calculation for the 1-d xy-model. In the xymodel the spins are unit vectors (cos x , sin x ) in the xy-plane that are attached to the points x of a d-dimensional lattice. Here we consider d = 1, i.e. we study a chain of xy-spins. The standard Hamilton function of the xy-model is given by H[] = J (1 cos(x+1 x )). (2.6.6)

xy

In complete analogy to the Ising model the transfer matrix is now given by x+1 |T |x = exp(J(1 cos(x+1 x )), (2.6.7)

2.6. THE TRANSFER MATRIX

23

which is a matrix with an uncountable number of rows and columns, because there is a continuum of values for x and x+1 . Still, we can ask about the eigenvalues of this matrix. For this purpose we consider the Fourier representation x+1 |T |x = where x |m = exp(imx ), (2.6.9) are the eigenvectors of the transfer matrix. The eigenvalues are given in terms of modied Bessel functions exp(Em ) = exp(J)Im (J). (2.6.10) x+1 |m exp(J)Im (J) m|x , (2.6.8)

mZ

The energy gap between the ground state and an excited state is given by Em E0 = log I0 (J) , Im (J) (2.6.11)

which is nonzero for nite . In the zero temperature limit we have I0 (J) m2 1+ , Im (J) 2J such that = 1/(E1 E0 ) 2J . (2.6.13) Hence, there is a critical point at zero temperature. In the language of quantum mechanics this implies the continuum limit of a Euclidean lattice theory corresponding to a quantum mechanical problem. In the continuum limit the energies corresponding to the eigenvalues of the transfer matrix take the form Em E0 m2 . 2J (2.6.14) (2.6.12)

These energies are in lattice units (the lattice spacing was put to 1). Hence, to extract physics we need to consider energy ratios and we nd Em E0 m2 . E1 E0 (2.6.15)

These are the appropriate energy ratios of a quantum rotor a particle that moves on a circle. Indeed the xy-spins describe an angle, which can be interpreted as the position of the quantum particle. Also the eigenvectors of the transfer matrix are just the energy eigenfunctions of a quantum rotor. Hence, we just

24

CHAPTER 2. FROM MECHANICS TO QUANTUM FIELD THEORY

solved the Schrdinger equation with a discrete Euclidean time step using the o transfer matrix instead of the Hamilton operator. The fact that energy ratios approach physically meaningful constants in the continuum limit is known as scaling. Of course, the discretization introduces an error as long as we are not in the continuum limit. For example, at nite the energy ratio is Em log(I0 (J)/Im (J) = , E1 log(I0 (J)/I1 (J) (2.6.16)

which is dierent from the continuum answer m2 . This cut-o eect due to a nite lattice spacing is known as a scaling violation.

2.7

Lattice Field Theory

So far we have restricted ourselves to quantum mechanical problems and to classical statistical mechanics. The former were dened by a path integral on a 1-d Euclidean time lattice, while the latter involved spin models on a d-dimensional spatial lattice. When we quantize eld theories on the lattice, we formulate the theory on a d-dimensional space-time lattice, i.e. usually the lattice is 4dimensional. Just as we integrate over all congurations (all paths) x(t) of a quantum particle, we now integrate over all congurations (x) of a quantum eld dened at any Euclidean space-time point x = (x, x4 ). Again the weight factor in the path integral is given by the action. Let us illustrate this for a free neutral scalar eld (x) R. Its Euclidean action is given by SE [] = m2 2 1 ]. d4 x [ + 2 2 (2.7.1)

Interactions can be included, for example, by adding a 4! 4 term to the action. The Feynman path integral for this system is formally written as

Z=

D exp(SE []).

(2.7.2)

(Note that we have put = c = 1.) The integral is over all eld congurations, which is a divergent expression if no regularization is imposed. One can make the expression mathematically well-dened by using dimensional regularization of Feynman diagrams. This approach is, however, limited to perturbation theory. The lattice allows us to formulate eld theory beyond perturbation theory, which is very essential for strongly interacting theories like QCD, but also for the standard model in general. For example, due to the heavy mass of the top quark,

2.7. LATTICE FIELD THEORY

25

the Yukawa coupling between the Higgs and top quark eld is rather strong. The above free scalar eld theory, of course, does not really require a nonperturbative treatment. We use it only to illustrate the lattice quantization method in a simple setting. On the lattice the continuum eld (x) is replaced by a lattice eld x , which is restricted to the points x of a d-dimensional space-time lattice. From now on we will work in lattice units, i.e. we put a = 1. The above continuum action can be approximated by discretizing the continuum derivatives such that SE [] =
x,

1 (x+ x )2 + 2

m2 2 . 2 x

(2.7.3)

Here is the unit vector in the -direction. The integral over all eld congu rations now becomes a multiple integral over all values of the eld at all lattice points

Z=
x

dx exp(SE []).

(2.7.4)

For a free eld theory the partition function is just a Gaussian integral. In fact, one can write the lattice action as SE [] = 1 2 x Mxy y , (2.7.5)

x,y

where the matrix M describes the couplings between lattice points. Diagonalizing this matrix by a unitary transformation U one has M = U DU . Introducing = Uxy y x one obtains Z=
x

(2.7.6)

(2.7.7)

d exp( x

1 2

Dxx ) = (2)N/2 detD 1/2 , x x

(2.7.8)

where N is the number of lattice points. To extract the energy values of the corresponding quantum Hamilton operator we need to study the 2-point function of the lattice eld x y = 1 Z D x y exp(SE []). (2.7.9)

26

CHAPTER 2. FROM MECHANICS TO QUANTUM FIELD THEORY

This is most conveniently done by introducing a source eld in the partition function, such that Z[J] = D exp(SE [] + Jx x ).
x

(2.7.10)

Then the connected 2-point function is given by x y

2

2 log Z[J] |J=0 . Jx Jy

(2.7.11)

The Boltzmann factor characterizing the problem with the external sources is given by the exponent 1 1 1 M J = M JM1 J. 2 2 2 Here we have introduced Integrating over = M1 J. in the path integral we obtain 1 Z[J] = (2)N/2 detD 1/2 exp( JM1 J), 2 and hence (2.7.14) (2.7.13) (2.7.12)

1 x y = M1 . (2.7.15) 2 xy It is instructive to invert the matrix M by going to Fourier space, i.e. by writing x = 1 (2)d dd p (p) exp(ipx).
B

(2.7.16)

The momentum space of the lattice is given by the Brillouin zone B =] , ]d . For the 2-point function in momentum space one then nds (p)(p) = [

(2 sin(p /2))2 + m2 ]1 .

(2.7.17)

This is the lattice version of the continuum propagator (p)((p) = (p2 + m2 )1 . (2.7.18)

From the lattice propagator we can deduce the energy spectrum of the lattice theory. For this purpose we construct a lattice eld with denite spatial momentum p located in a specic time slice (p)t =
x

x,t exp(ip x),

(2.7.19)

2.7. LATTICE FIELD THEORY and we consider its 2-point function (p)0 (p)t = 1 2

27

dpd (p)(p) exp(ipd t).

(2.7.20)

Inserting the lattice propagator of eq.(2.7.17) one can perform the integral. One encounters a pole in the propagator when pd = iE with (2 sinh(E/2))2 =
i

(2 sin(pi /2))2 + m2 .

(2.7.21)

The 2-point function then takes the form (p)0 (p)t = C exp(Et), (2.7.22)

i.e. it decays exponentially with slope E. This allows us to identify E as the energy of the lattice scalar particle with spatial momentum p. In general, E diers from the correct continuum dispersion relation E 2 = p 2 + m2 . (2.7.23)

Only in the continuum limit, i.e. when E, p and m are small in lattice units, the lattice dispersion relation agrees with the one of the continuum theory.

28

CHAPTER 2. FROM MECHANICS TO QUANTUM FIELD THEORY

Chapter 3

Classical Scalar Field Theory

Scalar elds play an important role in various areas of physics. For example, the Higgs eld of the standard model of particle physics is a scalar eld that gives rise to the spontaneous breakdown of the SU (2)L U (1)Y gauge group to the U (1)em gauge group of electromagnetism. This is how particles obtain their masses in the standard model. The lightest strongly interacting particle is the pion that arises in QCD. The pion is a so-called pseudo-Goldstone boson associated with the spontaneous breakdown of the approximate global chiral symmetry of QCD. In chiral perturbation theory, a technique developed by Gasser and Leutwyler, the pion is described by a scalar eld. Furthermore, scalar elds are used to describe Cooper pairs of electrons in the condensed matter physics of superconductors. In that case, the scalar eld dynamics leads to the spontaneous breaking of the U (1)em gauge symmetry itself. Besides being physically relevant, scalar elds are simpler to handle theoretically than fermion elds or gauge elds. This is the main reason why we begin our investigation of quantum eld theory using scalar elds. It should be mentioned that there are reasons to believe that truly elementary scalar elds may not even exist. For example, the scalar eld describing Cooper pairs is composed of electron elds. Similarly, at a more fundamental level the pion eld of chiral perturbation theory is composed of quark-, antiquark-, and gluon elds. It is likely that the Higgs eld of the standard model is also not truly fundamental. Since the Higgs particle has not yet been observed, this is remains an open question.

29

30

CHAPTER 3. CLASSICAL SCALAR FIELD THEORY

3.1

Scalar Fields

Scalar elds transform trivially under space-time transformations. In particular, they are invariant under Lorentz transformations. As a consequence, unlike vector elds, scalar elds do not carry any Lorentz indices. Still, scalar elds may transform non-trivially under certain internal symmetries. Such symmetries give rise to conserved charges which could be coupled to gauge elds. The simplest scalar eld is neutral and has no additional indices. It is simply described by a single real number per space-time point x. A charged scalar eld, such as the one representing a Cooper pair of electrons, is described by a complex number per space-time point. The scalar Higgs eld of the standard model is a complex doublet. It is described by two complex (or alternatively four real) numbers per space-time point. Finally, the pion eld of chiral perturbation theory has three internal degrees of freedom. It is described by a special unitary 2 2 matrix U (x) SU (2) (i.e. with determinant 1) per space-time point x. Let us consider an N -component scalar eld i (x) with i {1, 2, ..., N }. A neutral scalar eld corresponds to N = 1, while a charged scalar eld corresponds to N = 2. In case of the standard model Higgs eld one has N = 4. The Lagrangian of the corresponding scalar eld theory is given by 1 L(, ) = 2 Here V () = m2 2 4 + , (x)2 = 2 4!
N i=1

i i V ().
N

(3.1.1)

i (x)2 ,
i=1

(3.1.2)

is a scalar potential that contains the so-called bare mass m of the eld as well as the bare coupling constant of its self-interaction. Both m and get renormalized in the quantized theory. At this point we consider the theory at the classical level. The classical equation of motion for scalar eld theory is given by L dV () L = i + = i + m2 i + 2 i = 0. i i di 3! (3.1.3)

The classical vacuum conguration, i.e. the conguration of lowest energy, is simply given by i (x) = 0. Due to the uncertainty principle, the vacuum of scalar quantum eld theory (i.e. its ground state) cannot just be given by i (x) = 0. This is in complete analogy to an anharmonic oscillator with potential V (x) = 1 4 2 2 2 m x + 4! x . While the energy of the classical oscillator is minimized for x = 0, the quantum ground state contains quantum uctuations around the

3.2. NOETHERS THEOREM

31

classical vacuum. For example, for the harmonic oscillator (with = 0) these uctuations are described by a Gaussian wave function. Let us also look for non-zero solutions of the equation of motion from above. Due to the nonlinearity of the equations this is a non-trivial task. The situation simplies signicantly when we ignore the self-interaction of the scalar eld and put = 0. The classical equation of motion of the resulting free eld theory
2 i + m2 i = t i i + m2 i = 0,

(3.1.4)

is known as the Klein-Gordon equation. It admits simple plane-wave solutions i (x) = i exp(i(k x t)), 0 which obey the relativistic dispersion relation 2 = k 2 + m2 . (3.1.6) (3.1.5)

Here E = and p = k can be identied as the energy and the momentum of a relativistic free particle with energy-momentum relation E 2 = p2 c2 + (mc2 )2 . (3.1.7)

Since we have put = c = 1 this is completely consistent with the previous equation. In particular, the parameter m in the Lagrangian can be identied as the mass of the resulting free particle. It should be pointed out that the KleinGordon equation admits solutions for both positive and negative energies. Upon quantization, the latter give rise to anti-particles.

3.2

Noethers Theorem

The Lagrangian of scalar eld theory has an O(N ) symmetry, i.e. it is invariant against rotations (x) = O(x). (3.2.1) Here O is an N N orthogonal rotation matrix, i.e. OT O = OOT = 1. In components the previous equations take the form
N N N T Oij Ojk j=1

(x) =
j=1

Oij (x),

=
j=1

Oji Ojk = ik .

(3.2.2)

Symmetries are always important because they give rise to conserved quantum numbers. It should be noted that the O(N ) symmetry of our scalar eld theory

32

CHAPTER 3. CLASSICAL SCALAR FIELD THEORY

is global, i.e. the symmetry transformations O do not depend on space or time. This is in contrast to gauge theories in which symmetries are realized locally. Gauge theories have conserved charges. For example, electric charge is conserved both in classical and in quantum electrodynamics. In classical electrodynamics charge conservation is encoded in the continuity equation j = 0 for the electromagnetic current j (x) = ((x), j(x)). Here (x) and j(x) are the charge and current densities, respectively. The O(N ) symmetry of scalar eld theory also gives rise to a conserved current. In order to derive this current we now consider local O(N ) transformations (x) = O(x)(x). (3.2.3)

Interestingly, the potential contribution to the Lagrangian V () is invariant even against these local O(N ) transformations. This follows simply from (x)2 = (x)O(x)T O(x)(x) = (x)2 . (3.2.4)

Next we will compute the variation of the kinetic contribution to the Lagrangian with respect to innitesimal local O(N ) transformations O(x) = 1 + (x). The orthogonality of O(x) implies (x)T = (x). O(x)T O(x) = [1 + (x)T ][1 + (x)] 1 + (x) + (x)T = 1 (3.2.6) (3.2.5)

The innitesimally transformed scalar eld takes the form (x) = [1 + (x)](x), and hence (x) = (x) + (x)(x) + (x) (x). Consequently, one now obtains = + + + T + T = + , (3.2.9) and the variation of the action under local innitesimal O(N ) transformations takes the form S[ ] S[] =
N

(3.2.7) (3.2.8)

d4 x [L( , ) L(, )] =
N

d4 x
i,j=1

ij j ij =

d4 x
i,j=1

ij j ij ,

(3.2.10)

3.2. NOETHERS THEOREM with the O(N ) Noether current given by j ij = i j j i .

33

(3.2.11)

Indeed, the current is conserved, i.e. j ij = 0, as a consequence of the classical equations of motion. This follows directly from j ij = [i j j i ] = i j j i dV () dV () = i + j j d di = i [m2 j + 2 j ] j [m2 i + 2 i ] = 0. 3! 3!

(3.2.12)

Note that the derivation of current conservation is similar to the one for the probability current in ordinary quantum mechanics.

34

CHAPTER 3. CLASSICAL SCALAR FIELD THEORY

Chapter 4

Canonical Quantization of a Scalar Field

Canonical quantization of eld theory is a rather tedious approach. Therefore we will concentrate on the path integral for the rest of this course. However, canonical quantization has the advantage that it is rather similar to Schrdingers o approach to quantum mechanics which we are very familiar with. Here we consider the canonical quantization of a free 1-component scalar eld, which is fairly easy to carry out. The complications of canonical quantization versus the path integral show up only when interactions are included.

4.1

From the Lagrange to the Hamilton Density

Let us consider a 1-component real scalar eld (x) with the Lagrange density m2 2 1 . L(, ) = 2 2 The canonically conjugate momentum to the eld (x) is given by (x) = L = 0 (x), 0 (x) (4.1.2) (4.1.1)

which is just the time-derivative of (x). The classical Hamilton density is given by 1 m2 2 1 . (4.1.3) H(, ) = 0 L = 2 + i i + 2 2 2 35

36

CHAPTER 4. CANONICAL QUANTIZATION OF A SCALAR FIELD

Here the index i runs over the spatial directions only. The classical Hamilton function is the spatial integral of the Hamilton density H[, ] = d3 x H(, ) = 1 1 m2 2 d3 x ( 2 + i i + ). 2 2 2 (4.1.4)

The Hamilton function is a functional of the classical eld (x) and its canonically conjugate momentum eld (x). Upon quantization the Hamilton function will turn into the Hamilton operator of the corresponding quantum eld theory.

4.2

Commutation Relations for Scalar Field Operators

In the canonical quantization of eld theory the eld values and their conjugate momenta become operators acting in a Hilbert space. As we discussed before, the eld value (x) is analogous to the particle coordinate x in quantum mechanics. Similarly, (x) is analogous to the momentum p of the particle. In quantum mechanics position and momentum do not commute [xi , pj ] = i ij , [xi , xj ] = [pi , pj ] = 0. (4.2.1)

Similarly, (now putting = 1) one postulates the following commutation relations for the eld operators (x) and (y) [(x), (y)] = i(x y), [(x), (y)] = [(x), (y)] = 0. (4.2.2)

It is important to note that these commutation relations are completely local. In particular, elds at dierent points in space commute with each other. In quantum mechanics the momentum operator is represented as the derivative with respect to the position . (4.2.3) pi = i xi Similarly, the eld operator (x) can be written as (x) = i , (x) (4.2.4)

i.e. as a derivative with respect to the eld value.

4.3. HAMILTON OPERATOR IN SCALAR QUANTUM FIELD THEORY 37

4.3

Hamilton Operator in Scalar Quantum Field Theory

After turning the classical elds into operators it is straightforward to turn the classical Hamilton function H[, ] into the quantum Hamilton operator H= d3 x 1 2 ( + i i + m2 2 ). 2 (4.3.1)

At this level it should be obvious that quantum eld theory is really just quantum mechanics with innitely many degrees of freedom (in this case one for each point x in space). As usual, solving this quantum theory amounts to diagonalizing the Hamiltonian. For this purpose it is convenient to go to momentum space. Hence, we introduce Fourier transformed elds (p) = d3 x (x) exp(ip x), (p) = d3 x (x) exp(ip x). (4.3.2)

Note that, unlike (x) and (x), (p) and (p) are not Hermitean but obey (p) = (p), (p) = (p). (4.3.3)

Using the commutations relations for (x) and (y) one derives the commutation relations between (p) and (q) as [(p), (q)] = i(2)3 (p + q), [(p), (q)] = [(p), (q)] = 0. Similarly, one can now write the Hamilton operator as H= 1 (2)3 d3 p 1 ( + (p2 + m2 ) ). 2 (4.3.5) (4.3.4)

This Hamiltonian is reminiscent of the one for the harmonic oscillator with = p2 + m2 playing the role of the frequency. This suggests to introduce creation and annihilation operators a(p) and a(p) as i i 1 1 a(p) = [ (p) + (p)], a(p) = [ (p) (p) ], 2 2 which obey the commutation relations i i [a(p), a(q) ] = [(p), (q)] [(p), (q)] = (2)3 (p q), 2 2 [a(p), a(q)] = [a(p) , a(q) ] = 0. (4.3.7) (4.3.6)

38

CHAPTER 4. CANONICAL QUANTIZATION OF A SCALAR FIELD

In terms of these operators, the Hamiltonian takes the form H= 1 (2)3 d3 p 1 p2 + m2 (a(p) a(p) + V ). 2 (4.3.8)

The volume factor arises from (p) = 1 (2)3 d3 x exp(ip x) (2)3 (0) = d3 x 1 = V. (4.3.9)

4.4

Vacuum and Particle States

In analogy to a single harmonic oscillator, the vacuum state |0 of the scalar eld theory is determined by a(p)|0 = 0, (4.4.1) for all p. The vacuum is indeed an eigenstate of the Hamiltonian from above with the energy 1 1 V d3 p p2 + m2 . E= (4.4.2) (2)3 2 The volume factor represents a harmless infrared divergence. It is natural in a eld theory that the energy of the vacuum is proportional to the spatial volume. However, even the energy density 1 1 E = V (2)3 2 d3 p p 2 + m2 = 1 4 2
0

dp p2

p 2 + m2 ,

(4.4.3)

is still divergent in the ultraviolet. This is a typical short-distance (i.e. highmomentum) divergence of eld theory. The theory must be regularized in order to make the vacuum energy density nite. This can be achieved, for example, by introducing a momentum cut-o . The regularized vacuum energy density = E 1 = 2 V 4
0

dp p2

p2 + m2 4 ,

(4.4.4)

of course, again diverges in the limit . The vacuum energy of eld theory gives rise to one of the greatest mysteries in physics the cosmological constant problem. When one couples classical gravity, i.e. Einsteins general relativity, to quantum eld theory, the vacuum energy manifests itself as a cosmological constant. Recent observations have shown that the cosmological constant in Nature is extremely small but still positive. This leads to an accelerated expansion of the Universe, instead of the deceleration expected for a matter-dominated Universe.

4.4. VACUUM AND PARTICLE STATES

39

Understanding the origin of the vacuum energy density is one of the most challenging questions in physics today. A naive consideration from eld theory would perhaps identify the cut-o with the Planck scale MP = 1/ G 1018 GeV, at which quantum eects of gravity become important. Here G is Newtons constant. Hence, a naive eld theoretical estimate of the cosmological constant is about 4 theory MP . (4.4.5) The observed cosmological constant, on the other hand, is observation (103 eV)4 , such that (4.4.6)

theory 10120 . (4.4.7) observation This is presumably the greatest discrepancy between theory and observation ever encountered in all of physics. To explain this discrepancy is the essence of the cosmological constant problem. From a pure particle physics point of view, i.e. ignoring the gravitational eect of the vacuum energy, the divergence of is rather harmless. In particular, the energies of particles (which are excitations above the vacuum) are dierences between the energy of an excited state and the ground state (the vacuum). In these energy dierences the divergent factor drops out. In particular, the single particle states of the theory are given by |p = a(p) |0 , with an energy E(p) = = p 2 + m2 . (4.4.9) This is indeed the energy of a free particle with rest mass m and momentum p. More precisely, E(p) is the nite energy dierence between the divergent energy of the excited state and the ground state. Multi-particle states can be obtained by acting with more than one particle creation operator on the vacuum state. For example, the 2-particle states are obtained as |p1 , p2 = a(p1 ) a(p2 ) |0 , Since [a(p1 ) , a(p2 ) ] = 0 one nds |p2 , p1 = |p1 , p2 , (4.4.11) (4.4.10) (4.4.8)

i.e. the 2-particle state is symmetric under particle permutation. The same is true for multi-particle states. This shows that the scalar particles of our eld theory are indeed indistinguishable identical bosons.

40

CHAPTER 4. CANONICAL QUANTIZATION OF A SCALAR FIELD

4.5

The Momentum Operator

Let us also consider the momentum operator of our theory. Energy and momentum are components of the so-called energy-momentum tensor = g L. The Hamilton density from before is given by H = 00 = 0 0 g00 L = 2 L. Similarly, the momentum density is given by P = 0i = 0 i g0i L = i . (4.5.3) (4.5.2) (4.5.1)

Accordingly, in quantum eld theory, a Hermitean momentum operator is constructed as Pi = = = d3 x 1 (2)3 1 (2)3 1 (i + i ) 2 1 d3 p ipi ((p)(p) + (p)(p)) 2 d3 p pi a(p) a(p). (4.5.4)

The vacuum |0 is an eigenstate of the momentum operator with eigenvalue 0. Hence, as one would expect, the vacuum has zero momentum and is, consequently, translation invariant. The single particle states are again eigenstates, P |p = p|p , which shows that p is indeed the momentum of the particle. (4.5.5)

Chapter 5

Path Integral for Scalar Field Theory

As we have seen, the quantum mechanical path integral is particularly well dened in Euclidean time. Even the path integral in Minkowski real-time requires an innitesimal excursion into the complex time plane. In addition, the Euclidean path integral is intimately related to quantum statistical mechanics. Hence, we now consider quantum eld theory in the framework of the Euclidean path integral. This has the additional advantage that this formulation also works beyond perturbation theory. In particular, lattice eld theory is also formulated in Euclidean time. Just like the paths of a particle contributing to the quantum mechanical path integral in Euclidean time, Euclidean elds themselves are not directly physical objects. Instead they just serve as integration variables which allow us to derive physical quantities such as particle masses and coupling constants. There is a rigorous connection between the Euclidean time path integral and the real world. In particular, it yields the same physical results as the path integral formulated in Minkowski space-time. However, one should not confuse the Euclidean eld congurations with time-evolutions of physical elds in real time.

5.1

From Minkowski to Euclidean Space-Time

Let us start from the Lagrangian of a scalar eld in Minkowski space-time L(, ) = 1 1 V () = (t t i i ) V (), 2 2 41 (5.1.1)

42

CHAPTER 5. PATH INTEGRAL FOR SCALAR FIELD THEORY

which gives rise to the action S[] = dtd3 x L(, ). (5.1.2)

The path integral in Minkowski space-time is given by the formal expression Z= D exp(iS[]). (5.1.3)

We now analytically continue the time coordinate to purely imaginary values x4 = it. The Lagrangian then takes the form 1 1 L(, ) = [ (4 4 + i i ) + V ()] = [ + V ()]. 2 2 (5.1.5) (5.1.4)

In Euclidean space-time the distinction between co- and contra-variant indices is no longer necessary because the metric is simply given by g = . Due to the time-integration the action picks up an additional factor i and now takes the form 1 (5.1.6) S[] = i d4 x [ + V ()], 2 where d4 x = dx4 d3 x. The path integral in Euclidean time now takes the form Z= with the Euclidean action given by

D exp(SE []),

(5.1.7)

SE [] =
0

dx4

1 d3 x [ + V ()]. 2

(5.1.8)

We have introduced a nite extent = 1/T of the periodic Euclidean time dimension, which puts the eld theory at the nite temperature T . The path integral Z is nothing but the corresponding partition function of quantum statistical mechanics. It should be noted that, as it stands, the path integral Z is a highly divergent formal expression which needs to be regularized and properly renormalized. We have already seen how this can be done nonperturbatively by using the lattice regularization. While it is unavoidable in nonperturbative calculations, for perturbative calculations the lattice regularization is not the most convenient choice. It is much simpler to use dimensional regularization, i.e. an analytic continuation in the dimension of space-time. This is what we will concentrate on in

5.2. EUCLIDEAN PROPAGATOR AND CONTRACTION RULE

43

what follows. From now on we will keep the space-time dimension d as a continuous complex parameter. Then the formal expression for the Euclidean action reads 1 SE [] = dd x [ + V ()], (5.1.9) 2 and the corresponding partition function including an external eld j(x) takes the form Z[j] = D exp(SE [] + dd x j). (5.1.10)

Of course, one should not forget that this approach is limited to perturbation theory, and does not dene quantum eld theory nonperturbatively.

5.2

Euclidean Propagator and Contraction Rule

In Euclidean eld theory, physical information is extracted from n-point correlation functions, which correspond to vacuum expectation values of time-ordered products of eld operators 0|T (x1 )(x2 )...(xn )|0 = 1 Z D (x1 )(x2 )...(xn ) exp(SE []). (5.2.1)

For example, the 2-point function can be obtained as 0|T (x1 )(x2 )|0 = = 1 D (x1 )(x2 ) exp(SE []) Z 1 d2 Z[j] |j=0 . Z dj(x1 )dj(x2 )

(5.2.2)

In complete analogy to the lattice calculation discussed before, for a free scalar eld one obtains Z[j] = Z exp[ 1 2 dd xdd y j(x)G(x y)j(y)], (5.2.3)

where G(x) is the Euclidean propagator. In momentum space it takes the form G(p) = dd x G(x) exp(ipx) = p2 1 . + m2 (5.2.4)

For a free eld the n-point functions are simply related to the 2-point function.

44

CHAPTER 5. PATH INTEGRAL FOR SCALAR FIELD THEORY

For example, the 4-point function takes the form 0|T (x1 )(x2 )(x3 )(x4 )|0 = 1 D (x1 )(x2 )(x3 )(x4 ) exp(SE []) = Z d4 Z[j] 1 |j=0 . Z dj(x1 )dj(x2 )dj(x3 )dj(x4 ) Using eq.(5.2.3) one nds 0|T (x1 )(x2 )(x3 )(x4 )|0 = G(x1 x2 )G(x3 x4 ) +

(5.2.5)

G(x1 x3 )G(x2 x4 ) + G(x1 x4 )G(x2 x3 ).

(5.2.6)

In a Feynman diagram the propagators are represented as lines connecting the external points xi . This is an example of a general contraction rule for n-point functions with even n (the n-point functions for odd n simply vanish) 0|T (x1 )(x2 )...(xn )|0 = G(xi1 xi2 )G(xi3 xi4 )...G(xin1 xin ).

contractions

(5.2.7) The sum extends over all partitions of the indices 1, 2, ..., n into pairs (i1 , i2 ), (i3 , i4 ), ..., (in1 , in ). There are (n 1)!! = 1 3 5 ... (n 1) such pairings. For n = 4 there are indeed (4 1)!! = 1 3 = 3 contractions.

5.3

Perturbative Expansion of the Path Integral

Let us divide the Euclidean action into a free and an interacting part SE [] = Sf [] + Si [] with Sf [] = and dd x 1 [ + m2 2 ], 2 dd x (5.3.1)

(5.3.2)

1 4 . (5.3.3) 4! The interaction term makes it impossible to compute the full path integral analytically because it is no longer Gaussian. In perturbation theory one assumes that the coupling constant is small and one expands Si [] = 1 exp(SE []) = exp(Sf [])[1 + Si [] + Si []2 + ...]. 2 (5.3.4)

5.4. DIMENSIONAL REGULARIZATION

45

Inserting this expansion into the path integral, the expression for the n-point function takes the form 1 Z

0|T (x1 )(x2 )...(xn )|0 =

k=0

D (x1 )(x2 )...(xn )

k (4!)k k!

dd v1 dd v2 ...dd vk (v1 )4 (v2 )4 ...(vk )4 exp(Sf []). (5.3.5)

This expression is similar to the one for the n-point function in the free theory. However, in addition to the external points x1 , x2 , ..., xn , we now also have k internal points v1 , v2 , ..., vk whose positions are independently integrated over all of space-time and which are known as interaction vertices. These are the points at which the eld experiences its self-interaction. As before, we can apply the contraction rule. However, we now have four elds (vi ) at each interaction vertex. Correspondingly, there are now contractions that connect a vertex back to itself via a propagator G(vi vi ) = G(0). In a Feynman diagram such propagators appear as internal lines, while the propagators connected to external points xi are denoted as external lines. In four dimensions the propagator at zero distance G(0) is an ultraviolet divergent quantity. This divergence is regularized by analytically continuing the dimension of space-time.

5.4

Dimensional Regularization

Let us consider the propagator in d-dimensional space-time 1 (2)d exp(ipx) . p 2 + m2

G(x) =

dd p

(5.4.1)

Using 1 = p 2 + m2
0

exp(t(p2 + m2 )),

(5.4.2)

46

CHAPTER 5. PATH INTEGRAL FOR SCALAR FIELD THEORY

we can write G(x) = = = =

1 exp(tm2 ) dd p exp(ipx tp2 ) (2)d 0 1 x2 exp(tm2 ) dd q exp(tq 2 ) (2)d 0 4t x2 1 td/2 exp(tm2 ) 4t (4)d/2 0 1 md2 (m|x|)1d/2 K1d/2 (m|x|). (2)d/2

(5.4.3)

Here K (z) is a Bessel function. For |x| = 0 the propagator is nite for any value of d. However, for |x| = 0 we have G(0) = 1 (4)d/2

td/2 exp(tm2 ) =

1 d md2 (1 ), d/2 2 (4)

(5.4.4)

which diverges for d = 2, 4, 6, ... but is regular for other values of the space-time dimension. It should be noted that the above integral over t converges only for d < 2. However, the result of the integral can still be analytically continued to general d, except to even integer dimensions. Near d = 4 the -function takes the form 2 d (1) 1 + O(d 4), (5.4.5) (1 ) = 2 d4 which reveals the ultraviolet singularity as a pole in the space-time dimension at d = 4. Of course, the analytic continuation in the space-time dimension is just a mathematical trick that makes the propagator well-dened in the ultraviolet limit. Unlike the lattice regularization, dimensional regularization is dicult to interpret physically. Still, it is a rather elegant way to make the formal expressions of continuum eld theory mathematically well-dened at least in perturbation theory. Furthermore, it yields the same results in the perturbative continuum limit as e.g. the lattice regularization or other regularization schemes, but is much easier to handle. It is remarkable and reassuring that the physics is ultimately regularization-independent.

5.5

The 2-Point Function to Order

We will now use perturbation theory to evaluate the 2-point function to order . One obtains 1 0|T (x1 )(x2 )|0 = D (x1 )(x2 )[1 dd v1 (v)4 ] exp(Sf []). Z 4! (5.5.1)

5.5. THE 2-POINT FUNCTION TO ORDER Also the partition function itself must be expanded accordingly, i.e. Z= which implies 0|T (x1 )(x2 )|0 = + Here Zf = D exp(Sf []), 1 Zf 1 Zf 1 Zf 1 Zf D (x1 )(x2 ) exp(Sf []) D (x1 )(x2 ) D 4! 4! dd v (v)4 exp(Sf []) D [1 4! dd v (v)4 ] exp(Sf []),

47

(5.5.2)

dd v (v)4 exp(Sf []) (5.5.3)

D (x1 )(x2 ) exp(Sf []).

(5.5.4)

is the partition function of the free theory. The rst term in eq.(5.5.3) is the free propagator, and the other terms are corrections of order . Let us use the contraction rule to evaluate these terms. In the second term (x1 ) can be contracted either with (x2 ) or with one of the factors in (v)4 , for which there are four possibilities. Hence, we obtain 1 Zf D (x1 )(x2 ) 1 Zf D 1 Zf dd v (v)4 exp(Sf []) = dd v (v)4 exp(Sf []) + D (x2 )(v)3 exp(Sf []). (5.5.5)

G(x1 x2 )

dd v G(x1 v)

In the last term there are three ways to contract (x2 ) with one of the factors in (v)3 such that 1 Zf D (x2 )(v)3 exp(Sf []) = 3G(v x2 )G(0). (5.5.6)

Similarly, by putting x2 = v in this expression, we obtain 1 Zf D (v)4 exp(Sf []) = 3G(0)2 . (5.5.7)

48

CHAPTER 5. PATH INTEGRAL FOR SCALAR FIELD THEORY

Inserting these relations into eq.(5.5.5) one nds 1 Zf D (x1 )(x2 ) dd v (v)4 exp(Sf []) = dd v G(x1 v)G(v x2 ). (5.5.8)

3G(x1 x2 )G(0)2

dd v 1 + 3G(0)

Similarly, the last term in eq.(5.5.3) gives rise to 1 Zf 1 Zf D dd v (v)4 exp(Sf []) dd v 1. (5.5.9)

D (x1 )(x2 ) exp(Sf []) = 3G(x1 x2 )G(0)2

It should be noted that, like another term obtained before, this term is infrared divergent, i.e. it goes to innity with the space-time volume dd v 1. However, the two infrared divergent terms cancel. Both terms arise from so-called vacuum bubble diagrams. In those diagrams there are pieces completely disconnected from any external point xi . In that case, the integral over the corresponding vertex vk is unsuppressed and diverges with the space-time volume. One can prove that the contributions of vacuum bubbles to general n-point functions always cancel. We will not present the formal proof for this, but we will use this result and always drop vacuum bubble diagrams. Putting everything together, one thus obtains 0|T (x1 )(x2 )|0 = G(x1 x2 ) G(0) 2 dd v G(x1 v)G(v x2 ). (5.5.10)

The integral over v corresponds to a convolution in coordinate space, which after a Fourier transform turns into a simple product in momentum space. Hence, in momentum space the previous equation takes the form G (p) = 1 1 1 G(0) 2 + O(2 ) = 2 + O(2 ). 2 2 )2 +m 2 (p + m p + m2 (5.5.11)

p2

Here G (p) is the full propagator to order and m is the mass also corrected to that order. One reads o m2 = m2 + d 1 md2 (1 ). G(0) = m2 + d/2 2 2 (4) 2 (5.5.12)

5.6. MASS RENORMALIZATION

49

5.6

Mass Renormalization

We have seen that the mass m of the free scalar particle changes to m when the 4 interaction is taken into account at leading order in . Although eq.(5.5.12) yields a nite value for m for almost all values of d, m still diverges when one approaches the physical space-time dimension d = 4. This is a manifestation of the ultraviolet divergences of eld theory which, unlike infrared divergences, do not simply cancel. In particular, as it stands, the expression for m still diverges in the physical limit. To cure this problem, in addition to regularizing the ultraviolet divergence (in this case, by an analytic continuation in the dimension of spacetime), we must now renormalize the bare mass parameter m. As we have seen, in the presence of the 4 interaction, to order , the physical mass is now m and no longer m. Hence, m is physical and should thus be nite, while m is a so-called bare (i.e. unrenormalized) parameter that appears in the Lagrangian, but has no direct physical meaning. Renormalization of the mass means that we let the unphysical bare parameter m depend on the cut-o parameter (i.e. in this case on d 4) such that the physical renormalized mass m remains xed. In particular, we now put m2 = m2 1 d G(0) = m2 md2 (1 ) d/2 2 2 (4) 2 2 1 = m2 md2 [ (1) 1 + O(d 4)], 2 (4)d/2 d4

(5.6.1)

which means that m depends on d. In particular, now m2 diverges (it has a pole at d = 4) while the physical mass m remains nite. At the end we have traded an unphysical bare parameter m for a physical mass m . However, the theory itself does not predict the physical value of the mass. Just as we could choose any value of m in the free theory, we can now choose any value of m for the interacting theory.

5.7

Connected and Disconnected Diagrams

It is useful to classify Feynman diagrams according to their topology. An npoint function 0|T (x1 )(x2 )...(xn )|0 is represented by diagrams with n external points xi , a number k of internal vertices at points vi , and a number of propagators represented by lines connecting the points. There is one line emanating from each external point and there are four lines running into each

50

CHAPTER 5. PATH INTEGRAL FOR SCALAR FIELD THEORY

vertex. Vertices that are not connected directly or indirectly to any external points belong to vacuum bubbles. Feynman diagrams that contain vacuum bubbles cancel and thus need not be considered. In the following we limit ourselves to diagrams without vacuum bubbles. Such diagrams decompose into various connected pieces. A connected piece is characterized by the set of external points that it contains. Those are connected to each other by propagators either directly or indirectly via some vertices. All vertices belong to some connected piece, otherwise there would be vacuum bubbles. A Feynman diagram that contains more than one connected piece is called disconnected. The contribution of such a diagram to the n-point function factorizes into contributions from each individual connected piece. Hence, the problem of computing a general n-point function reduces to the evaluation of connected diagrams. For example, the 2point function 0|T (x1 )(x2 )|0 automatically receives contributions from connected diagrams only. The 4-point function 0|T (x1 )(x2 )(x3 )(x4 )|0 , on the other hand, gets contributions from various connected pieces corresponding to the partitions [(x1 , x2 ), (x3 , x4 )], [(x1 , x3 ), (x2 , x4 )], [(x1 , x4 ), (x2 , x3 )], as well as from contributions in which all four external points are connected with each other. A useful category of connected diagrams are the 1-particle irreducible ones. Those remain connected when any single propagator line in the diagram is cut. Diagrams that fall apart into two disconnected pieces under the cutting operation are 1-particle reducible. The full 2-point function gets contributions from chains of 1-particle irreducible subgraphs connected by single propagators. These terms build a geometric series. The rst term of the series is the free propagator (p2 + m2 )1 . The second term contains the remaining 1-particle irreducible diagrams and takes the form (p2 + m2 )1 (p2 )(p2 + m2 )1 , where (p2 ) is the so-called self-energy. The third term is given by (p2 + m2 )1 (p2 )(p2 + m2 )1 (p2 )(p2 + m2 )1 , etc., such that the full propagator is given by G (p) = + = 1 1 1 2 (p2 ) 2 2 2 +m p +m p + m2 1 1 1 (p2 ) 2 (p2 ) 2 ... 2 + m2 2 p p +m p + m2 1 1 1 = 2 . 2 + m2 1 + (p2 )/(p2 + m2 ) 2 + (p2 ) p p +m p2

(5.7.1)

One can now read o the physical mass m of the particle from the value p2 = m2 for which the full propagator has a pole, i.e. m2 = m2 + (m2 ). (5.7.2)

5.8. FEYNMAN RULES FOR 4 THEORY The perturbative calculation from before implies (p2 ) = d 1 md2 (1 ) + O(2 ). d/2 2 (4) 2

51

(5.7.3)

The full propagator can be expanded around the pole at p2 = m2 and one obtains Z , (5.7.4) G (p) = 2 p + m2 where Z is the residue of the pole. Expanding (p2 ) = (m2 ) + (p2 + m2 ) (m2 ) + ... one identies Z= The factor 1 1 (m2 ) (5.7.5)

(5.7.6)

Z is known as the wave function renormalization constant.

5.8

Feynman Rules for 4 Theory

It is convenient to consider the Fourier transform of the n-point function dd x1 dd x2 ...dd xn 0|T (x1 )(x2 )...(xn )|0 (2)d (p1 + p2 + ... + pn )(p1 , p2 , ..., pn ). exp[i(p1 x1 + p2 x2 + ... + pn xn )] = (5.8.1)

The -function results from translation invariance. Dropping Feynman diagrams containing vacuum bubbles or disconnected pieces, in perturbation theory one then obtains (2)d (p1 + p2 + ... + pn )(p1 , p2 , ..., pn ) = 1 D (p1 )(p2 )...(pn ) exp(Sf []) Zf
k=0

()k (4!)k k!

dd v1 dd v2 ...dd vk (v1 )4 (v2 )4 ...(vk )4 .

(5.8.2)

Here (p) is the Fourier transform of the eld (x). Since we are interested in connected diagrams only, we must contract a eld (p) carrying an external

52

CHAPTER 5. PATH INTEGRAL FOR SCALAR FIELD THEORY

momentum p with a eld (v) at an internal vertex v. The resulting contraction takes the form 1 Zf exp(ipv) . D (p)(v) exp(Sf []) = 2 p + m2 (5.8.3)

In addition, the remaining elds (vi ) must be contracted among themselves, which implies 1 Zf D (vi )(vj ) exp(Sf []) = 1 (2)d dd q exp(iq(vi vj )) . p 2 + m2 (5.8.4)

The momentum q associated with the internal line connecting the vertices vi and vj must be integrated over. After performing the contractions, the positions of the vertices vi appear only in exponential factors. Hence, when one integrates over vi one generates a -function for the four momenta owing into that vertex. These observations lead to the following Feynman rules for the evaluation of connected n-point functions in a 4 theory: Consider all pairwise contractions in the product of internal and external elds (p1 )(p2 )...(pn )(v1 )4 (v2 )4 ...(vk )4 that lead to a connected Feynman diagram. Each diagram has a multiplicity factor that counts the number of pairings leading to the same topology of the diagram. Also one must take into account the factor 1/k!. Write down a propagator (p2 + m2 )1 for each oriented external line. i Associate a momentum qi with each oriented internal line and write down 2 a factor (qi + m2 )1 . Write down a momentum conserving -function (2)d (q1 + q2 + q3 + q4 ) for each vertex vi taking into account the orientation of the internal or external lines. To each vertex is also associated a factor /4!. Finally, integrate the resulting expression over all internal momenta, i.e. write dd q1 ...dd qI /(2)Id . The number I of internal lines can be determined as follows. There are four lines emanating from each vertex, n of which are external. Each internal line connects two vertices and thus 1 I = (4k n). (5.8.5) 2 There are I integrations and k -functions. However, one of the -functions reects translation invariance and will remain in the nal result. Hence, only k1

5.9. THE 4-POINT FUNCTION TO ORDER 2

53

-functions can be used to perform some integrations trivially. The remaining l = I (k 1) = k + 1 n 2 (5.8.6)

integrations determine the number of loops l in the diagram. The number of loops increases with the order k of the expansion and also depends on the number of external momenta n of the n-point function. For example, at order (i.e. for k = 1) a 2-point function (with n = 2) receives a 1-loop contribution, while at the same order a 4-point function has no contributions from loop diagrams. Diagrams without loops (i.e. with l = 0) are called tree diagrams. The main diculty in evaluating Feynman diagrams is to perform the loop-integrations over internal momenta. While tree diagrams are easy to evaluate, multi-loop diagrams are much harder to deal with.

5.9

The 4-Point Function to Order 2

Let us begin with something simple, the tree diagram for the 4-point function. There is just one vertex and all four external momenta ow into that vertex. There are four possible contractions for the external line with momentum p1 , three remaining possible contractions for the external line with momentum p2 , two remaining possible contractions for the external line with momentum p3 , and nally only one possible contraction for the external line with momentum p4 . Hence, the multiplicity factor is 4 3 2 = 4!. Following the Feynman rules we obtain (2)d (p1 + p2 + p3 + p4 )(p1 , p2 , p3 , p4 ) = 1 4! 2 2 )(p2 + m2 )(p2 + m2 )(p2 + m2 ) (p1 + m 2 3 4 () (2)d (p1 + p2 + p3 + p4 ), 4! such that (p1 , p2 , p3 , p4 ) = (p2 1 + m2 )(p2 2 + m2 )(p2 3 + m2 )(p2 + m2 ) 4 . (5.9.2)

(5.9.1)

In the next order k = 2 we need to study the 1-loop diagrams. Now there are two vertices. If all external lines would ow into the same vertex, the diagram would be disconnected. Hence, either three or two external lines may run into the same vertex. First, we consider the case where the external line with the

54

CHAPTER 5. PATH INTEGRAL FOR SCALAR FIELD THEORY

momentum p1 runs into one vertex and the other external lines (with momenta p2 , p3 , and p4 ) run into the other vertex. There are 8 possible contractions for the rst external line with momentum p1 . Then there are 4 3 2 possible contractions of the three other external lines at the other vertex. Finally, there are 3 contractions for an internal line with momentum q1 connecting the two vertices and one remaining internal line with momentum q2 leading from the rst vertex back to itself. Hence, the total multiplicity factor is 84323 = (4!)2 . Following the Feynman rules one obtains the contribution 1 (2)2d 1 1 2 + m2 )(p2 + m2 )(p2 + m2 )(p2 + m2 ) 2! (p1 2 3 4 1 ()2 (2)2d (p1 q2 )(q2 + p2 + p3 + p4 ) = 2 2 (q1 + m2 )(q2 + m2 ) (4!)2 dd q1 dd q2 (4!)2 2 G(0)/2 (2)d (p1 + p2 + p3 + p4 ). (5.9.3) (p2 + m2 )2 (p2 + m2 )(p2 + m2 )(p2 + m2 ) 1 2 3 4 The integration over q2 was performed trivially using a -function while the integration over the loop-momentum q1 gives the divergent (but dimensionally regularized) factor G(0). Similar expressions exist for diagrams where the loop is attached to the external lines with momenta p2 , p3 , and p4 . These four 1-loop diagrams just give rise to the renormalization of the mass of the incoming particles. Together with the tree diagram they lead to (p1 , p2 , p3 , p4 ) = (p2 1 + m2 )(p2 2 + m2 )(p2 3 + m2 )(p2 + m2 ) 4 . (5.9.4)

However, this is not the full answer to order 2 . We also need to take into account the 1-loop diagrams where two external lines run into the same vertex. Then there are again 8 possible contractions for the external line with momentum p1 . If the external line with momentum p2 runs into the same vertex there are 3 remaining possible contractions. The external line with momentum p3 has 4 possible contractions at the other vertex, and the external line with momentum p4 has 3 remaining contractions also at that other vertex. Finally there are 2 possible contractions for internal lines with momenta q1 and q2 connecting the two vertices. Hence, the total multiplicity factor for this diagram is again 8 3 4 3 2 = (4!)2 . The Feynman rules lead to the expression 1 1 2 + m2 )(p2 + m2 )(p2 + m2 )(p2 + m2 ) 2! (p1 2 3 4 2 1 () (2)2d (p1 + p2 q1 q2 )(q1 + q2 + p3 + p4 ) = 2 2 (q1 + m2 )(q2 + m2 ) (4!)2 1 (2)2d dd q1 dd q2 (4!)2 (p2 1 + m2 )2 (p2 2 2 J(s)/2 (2)d (p1 + p2 + p3 + p4 ), + m2 )(p2 + m2 )(p2 + m2 ) 3 4

(5.9.5)

5.10. DIMENSIONAL REGULARIZATION OF J(P 2 ) where J(p2 ) =

55

1 (2)d

dd q

1 (q 2 + m2 )((p q)2 + m2 )

(5.9.6)

We have introduced the so-called Mandelstam variable s = (p1 + p2 )2 . There are two other diagrams with similar topology which depend on the Mandelstam variables t = (p1 + p3 )2 and u = (p1 + p4 )2 . The total contribution to the 4-point function at the order 2 takes the form 2 [J(s) + J(t) + J(u)]/2 (p1 , p2 , p3 , p4 ) = 2 . (5.9.7) (p1 + m2 )(p2 + m2 )(p2 + m2 )(p2 + m2 ) 2 3 4

5.10

Dimensional Regularization of J(p2)

Just like G(0), J(p2 ) is divergent in four space-time dimensions and must be regularized. As before we choose dimensional regularization. In order to evaluate the corresponding integral we use a trick due to Feynman and Schwinger 1 = AB and we write J(p2 ) = = =
0 1

d
0

1 , [(1 )A + B]2 1 q)2 + m2 ) + m2 )

(5.10.1)

1 (2)d 1 (2)d
1

dd q
1

(q 2

m2 )((p [(1 (2 q

d
0

dd q dd q

)(q 2

1 (2)d

1 . + m2 )2

1 , + ((p q)2 + m2 )]2 (5.10.2)

In the last step we have shifted the integration variable from q to q = q p and we have introduced m2 = m2 (1 )p2 . Also using 1 = (2 + m2 )2 q

dt t exp(t(q 2 + m2 )),

(5.10.3)

and performing the d-dimensional Gaussian integral 1 dd q exp(t2 ) = (4t)d/2 , q (2)d one obtains J(p2 ) =
0 1

(5.10.4)

d
0

dt (4t)d/2 t exp(tm2 )
1 0

(2 d ) 2 (4)d/2

d md4 =

1 + ... 8 2 (d 4)

(5.10.5)

56

CHAPTER 5. PATH INTEGRAL FOR SCALAR FIELD THEORY

The expression has a pole at d = 4. The ellipses indicate the remaining nite piece. It should be noted that the divergent piece is independent of p2 .

5.11

Renormalization of the Coupling Constant

Since the divergence of the 4-point function is momentum-independent, it can be absorbed into a redenition (a renormalization) of the coupling constant . The coupling depends on the momenta of the interacting particles. Hence, in order to dene the physical coupling in a unique way one must specify a so-called renormalization condition. The specic form of the renormalization condition is entirely a matter of convention. It is conventional to dene the physical coupling at the center of the so-called Mandelstam triangle 4 s = t = u = m2 , 3 (5.11.1)

where m is the physical renormalized mass. We dene the nite part of the function J(p2 ) as 4 J(p2 ) = J(p2 ) J( m2 ), (5.11.2) 3 and we obtain (p1 , p2 , p3 , p4 ) = 2 [J (s) + J(t) + J(u)]/2 2 + m2 )(p2 + m2 )(p2 + m2 )(p2 + m2 ) , (p1 2 3 4 (5.11.3)

where the physical renormalized coupling constant is given by 3 4 = J( m2 )2 + O(3 ). 2 3 (5.11.4)

The cut-o is now removed by demanding that the physical coupling constant remains xed at the value that describes experiments, while the bare coupling diverges. It is remarkable that the renormalizations of m and are sucient to make all higher n-point functions nite without further adjustments. In other words, after xing the renormalized mass and coupling at their physical values m and , all other results of the theory are completely xed. This implies a large predictive power of the theory. Just like the classical 4 theory, the corresponding quantum eld theory has only two free parameters. Quantum eld theories with a nite number of such parameters are called renormalizable.

5.12. RENORMALIZABLE SCALAR FIELD THEORIES

57

5.12

Renormalizable Scalar Field Theories

Let us consider a scalar eld theory in d space-time dimensions with a general polynomial self-interaction potential V () =
p

gp p .

(5.12.1)

The corresponding Euclidean action takes the form S[] = 1 dd x [ + V ()]. 2 (5.12.2)

Since the action enters the Boltzmann factor exp(S[]) in the path integral as an exponent, it must be dimensionless. Consequently, the eld has the dimension d = d2 . 2 (5.12.3)

For example, in d = 2 a scalar eld is dimensionless (d = 0), while in d = 4 it has the dimension of a mass (d = 1). Similarly, the coupling constant gp has the dimension p (5.12.4) dgp = d pd = d(1 ) + p. 2 For example, the 2 term has dimension dg2 = 2 irrespective of the dimension 1 d. Of course, the prefactor g2 = 2 m2 of this term is indeed given by the mass squared of the scalar particle. It will turn out that theories containing couplings with negative mass dimension (i.e. dgp < 0) are not renormalizable. Hence, renormalizability requires p 2d d(1 ) + p 0 p . 2 d2 (5.12.5)

As a result, in d = 2 space-time dimensions any polynomial potential (p ) leads to a renormalizable scalar eld theory. In d = 3, on the other hand, renormalizability requires p 6, and in d = 4 dimensions p 4. In d = 6 we have p 3, i.e. a 3 interaction is formally still renormalizable. However, in the absense of a stabalizing 4 term, such a theory has an unstable vacuum and is thus inconsistent. Consequently, in more than d = 4 dimensions, renormalizable scalar eld theories are necessarily trivial, i.e. they are free eld theories because the potential is limited to at most quadratic terms, which lead to simple Gaussian integrals. Let us now derive eq.(5.12.5) by analyzing the degree of divergence of Feynman diagrams. For this purpose we choose a more physical regularization than

58

CHAPTER 5. PATH INTEGRAL FOR SCALAR FIELD THEORY

dimensional regularization, which is nothing more (and nothing less) than a very neat mathematical device. The most natural physical regularization is the one using a space-time lattice. Here we choose something that is more easy to handle in the calculation of Feynman diagrams, namely a regularization using a momentum cut-o . In this regularization the integrals over internal momenta qi are limited to |q| . Let us consider a connected Feynman diagram for an n-point function in a d-dimensional scalar eld theory with general polynomial self-interactions. There are n external lines, kp vertices of type p > 2 (with p lines emanating from them), and I internal lines. The total number of vertices is k = p kp and the total number of lines emanating from these vertices is p kp p. Hence, the number of internal lines is given by 1 kp p n). (5.12.6) I= ( 2 p As before, each vertex is associated with a momentum-conserving -function, and, as a manifestation of translation invariance, there is one remaining overall -function. Hence, again there are l = I k+1 (5.12.7)

non-trivial loop integrations. Each integration over an internal line momentum qi d1 contains a measure factor qi , and each internal line propagator takes the form 2 (qi + m2 )1 . We are interested in an upper limit on the degree of ultraviolet divergence of the Feynman diagram. Hence, we neglect m2 in the propagator 2 (whoose ultraviolet contributions are dominated by qi 2 ) and we replace the 2 propagator by qi . The so-called supercial degree of divergence is then given by d2 = dl 2I = ( kp p n) d kp + d. (5.12.8) 2 p p The integral in the corresponding Feynman diagram is convergent if < 0, i.e. if [kp (
p

d2 d2 p d)] < n d. 2 2

(5.12.9)

In order to prevent the supercial degree of divergence to become arbitrarily large for a large number kp of vertices, i.e. in order for a theory to be renormalizable, one needs 2d d2 pd0 p . (5.12.10) 2 d2 This is just the condition of eq.(5.12.5). Theories for which p < 2d/(d 2) for all vertices are called super-renormalizable. For example, 4 theory (with p = 4) is

5.12. RENORMALIZABLE SCALAR FIELD THEORIES

59

super-renormalizable in d = 3 space-time dimensions. In four dimensions 4 theory is still renormalizable because then p = 2d/(d 2). In a renormalizable (but not super-renormalizable) theory Feynman diagrams are supercially divergent if d2 n d. (5.12.11) 2 For example, in d = 4, 2- and 4-point functions are divergent, but all higher n-point functions are convergent, while in d = 3 dimensions 6-point functions are still divergent.

60

CHAPTER 5. PATH INTEGRAL FOR SCALAR FIELD THEORY

Chapter 6

Canonical Quantization of Electrodynamics

In order to get a good intuitive understanding of photons, we rst consider the canonical quantization of the free electromagnetic eld. Interactions with scalar elds will be discussed in the next chapter using the Euclidean path integral.

6.1

From the Lagrange to the Hamilton Density

Let us consider the Lagrange density of the free electromagnetic eld 1 L(A ) = F F , 4 where the eld strength tensor is given by F = A A . (6.1.2) (6.1.1)

The eld strength (and thus the action) is invariant against Abelian gauge transformation A = A + . (6.1.3) In order to isolate the physical gauge invariant degrees of freedom of the gauge eld we x the gauge by choosing A0 (x) = 0. This leaves a remnant invariance against time-independent (but still space-dependent) gauge transformations. In particular, after a time-independent gauge transformation we still have A = A0 + 0 = A0 = 0. 0 61 (6.1.4)

62CHAPTER 6. CANONICAL QUANTIZATION OF ELECTRODYNAMICS It is convenient to x the remnant gauge freedom by choosing the Coulomb gauge i Ai = 0, (6.1.5)

at an initial instance of time. Then the electric and magnetic eld strengths are given by 1 Ei = Fi0 = i A0 0 Ai , Bi = ijk Fjk = ijk j Ak , (6.1.6) 2 and the Lagrange density takes the form 1 L(Ai ) = (Ei Ei Bi Bi ). 2 (6.1.7)

The canonically conjugate momentum to the vector potential Ai (x) is given by i (x) = L = 0 Ai (x) = Ei (x), 0 Ai (x) (6.1.8)

which is just the electric eld. The classical Hamilton density is given by H(Ai , i ) = i 0 Ai L = 1 1 (i i + Bi Bi ) = (Ei Ei + Bi Bi ). 2 2 (6.1.9)

The classical Hamilton function is the spatial integral of the Hamilton density H[Ai , Ei ] = d3 x H(Ai , Ei ) = d3 x 1 2 [E + (ijk j Ak )2 ]. 2 i (6.1.10)

The Hamilton function is a functional of the classical eld Ai (x) and its canonically conjugate momentum eld i (x) = Ei (x). Upon quantization the Hamilton function will turn into the Hamilton operator of the corresponding quantum eld theory. As in the canonical quantization of scalar elds we now postulate commutation relations between the gauge eld variables Ai (x) and their canonically conjugate momenta i (x) = Ei (x) [Ai (x), Ej (y)] = iij (x y), [Ai (x), Aj (y)] = [Ei (x), Ej (y)] = 0. (6.1.11)

Again, these commutation relations are completely local, i.e. elds at dierent points in space commute with each other. The eld operator Ei (x) takes the form . (6.1.12) Ei (x) = i Ai (x)

6.2. THE HAMILTON OPERATOR FOR THE PHOTON FIELD

63

6.2

The Hamilton Operator for the Photon Field

We now turn the classical Hamilton function H[Ai , Ei ] into the Hamilton operator H= d3 x 1 2 [E + (ijk j Ak )2 ]. 2 i (6.2.1)

In order to diagonalize the Hamiltonian it is convenient to go to momentum space. We introduce Fourier transformed elds Ai (p) = which obey d3 x Ai (x) exp(ip x), Ei (p) = d3 x Ei (x) exp(ip x), (6.2.2)

Ai (p) = Ai (p), Ei (p) = Ei (p).

(6.2.3)

The corresponding commutations relations take the form [Ai (p), Ei (q)] = i(2)3 ij (p + q), [Ai (p), Aj (q)] = [Ei (p), Ei (q)] = 0, and the Hamilton operator can be written as H= 1 (2)3 d3 p 1 [E Ei + (ijk pj A )(ilm pl Am )]. k 2 i (6.2.5) (6.2.4)

It is important to note that one must also respect the Coulomb gauge constraint i Ai = 0 which in momentum space takes the form pi Ai = 0. (6.2.6)

Hence, the direction of Ai must be perpendicular to the direction of the momentum pi . For example, when the momentum points in the z-direction, i.e. when p = (0, 0, p), there are two linearly independent modes Ax and Ay , corresponding to the two polarization states of an electromagnetic wave or a photon. Using the gauge constraint we can write (ijk pj A )(ilm pl Am ) = (jl km jm kl )pj pl A Am ) = p2 A Ak . k k k (6.2.7)

Hence, keeping in mind that their are only two physical transverse degrees of freedom for the electromagnetic eld, the Hamiltonian takes the form H= 1 (2)3 d3 p 1 (E Ei + p2 A Ai ). i 2 i (6.2.8)

It should be noted that, for a given momentum p, the sum over i is restricted to the two polarization directions perpendicular to p.

64CHAPTER 6. CANONICAL QUANTIZATION OF ELECTRODYNAMICS In complete analogy to the canonical quantization of a scalar eld, using = |p|, we now introduce creation and annihilation operators ai (p) and ai (p) as 1 1 i i ai (p) = [ Ai (p) + Ei (p)], ai (p) = [ Ai (p) Ei (p) ], (6.2.9) 2 2 which obey the commutation relations i i [ai (p), aj (q) ] = [Ei (p), Aj (q)] [Ai (p), Ej (q)] = (2)3 ij (p q), 2 2 [ai (p), aj (q)] = [ai (p) , aj (q) ] = 0. (6.2.10) In terms of these operators, the Hamiltonian takes the form H= 1 (2)3 d3 p |p|(ai (p) ai (p) + V ). (6.2.11)

Here V is the spatial volume. Again, the sum over i is restricted to the two polarization directions perpendicular to the momentum vector p.

6.3

Vacuum and Photon States

In complete analogy to the case of a scalar eld, the vacuum state of the electromagnetic eld is determined by ai (p)|0 = 0, (6.3.1)

for all p and all polarizations i. The vacuum is an eigenstate of the Hamiltonian from above with the energy E= 1 V (2)3 d3 p |p|. (6.3.2)

As usual, the volume factor represents a harmless infrared divergence which disappears when one considers the vacuum energy density = 1 E = V (2)3 d3 p |p| = 1 2 2
0

dp p3 .

(6.3.3)

As for the scalar eld, the vacuum energy is still divergent in the ultraviolet. When we use a momentum cut-o, the energy density diverges as 4 .

6.4. ELECTROMAGNETIC MOMENTUM OPERATOR

65

The energies of photons are dierences between the energy of an excited state and the vacuum. In these energy dierences the divergent factor drops out. The single photon states are given by |p, i = ai (p) |0 , with an energy E(p) = = |p|. (6.3.5) This is the energy of a free particle with vanishing rest mass and momentum p. Multi-photon states are obtained by acting with more than one creation operator on the vacuum state. The 2-photon states are obtained as |p1 , i, p2 , j = ai (p1 ) aj (p2 ) |0 , Since [ai (p1 ) , aj (p2 ) ] = 0 one has |p2 , j, p1 , i = |p1 , i, p2 , j , which shows that photons are bosons. (6.3.7) (6.3.6) (6.3.4)

6.4

Electromagnetic Momentum Operator

Let us also consider the momentum operator of our theory. The energy-momentum tensor of the electromagnetic eld takes the form
= F F g L.

(6.4.1)

The Hamilton density from before is given by 1 2 H = 00 = F0i F0i g00 L = Ei L = (Ei Ei + Bi Bi ). 2 The momentum density is given by P = 0i = F0j Fij g0i L = 1 ijk Ej Bk . 2 (6.4.3) (6.4.2)

Accordingly, the Hermitean momentum operator of the photon eld theory is given by 1 P = d3 x E B. (6.4.4) 2 One can show that 1 d3 p p ai (p) ai (p). (6.4.5) P = (2)3

66CHAPTER 6. CANONICAL QUANTIZATION OF ELECTRODYNAMICS As expected, the vacuum |0 is an eigenstate of the momentum operator with eigenvalue 0. The single photon states are again eigenstates, P |p, i = p|p, i , which shows that p is indeed the momentum of the photon. (6.4.6)

Chapter 7

Path Integral for Scalar Electrodynamics

In this chapter we consider the perturbative quantization of scalar electrodynamics using the path integral in Euclidean space-time. This theory describes the interactions of a complex (i.e. 2-component) scalar eld with itself and with the electromagnetic eld. Scalar QED can be used to describe the physics of superconductors with the scalar eld representing the Cooper pairs consisting of two electrons. In particle physics the more relevant theory is spinor QED the theory describing the interactions between electrons, positrons, and photons. At rst, in order to avoid extra complications due to the fermionic nature of electrons and positrons (which are described by the Dirac equation), we concentrate on scalar quantum electrodynamics. As in the case of 4 theory we consider the phase in which the symmetry is unbroken.

7.1

Gauge Fixing and Photon Propagator

The Euclidean action of the free electromagnetic eld is given by SE [A] = where the eld strength is given by F = A A . 67 (7.1.2) dd x 1 F F , 4 (7.1.1)

68 CHAPTER 7. PATH INTEGRAL FOR SCALAR ELECTRODYNAMICS Both the eld strength and the action are invariant under gauge transformations A = A + . The path integral in Euclidean time now takes the form Z= DA exp(SE [A]). (7.1.4) (7.1.3)

Again, this formal expression must be properly dened by imposing an appropriate regularization. In addition, in a perturbatively quantized gauge theory one must also x the gauge. Otherwise there would be a divergence resulting from path integration over innitely many gauge copies with the same Euclidean action. Without gauge xing the photon propagator would take the form (p2 p p )1 . This is expression is a singular matrix (with an innite eigenvalue) in the space of Lorentz indices because (p2 p p )p = 0. (7.1.5)

In the perturbative path integral quantization of non-Abelian gauge theories gauge xing is a non-trivial issue. Interestingly, when non-Abelian gauge theories are quantized nonperturbatively, using a space-time lattice, gauge xing is not even necessary. Still, even on the lattice, the quantization of (non-compact) Abelian gauge theories does require gauge xing. Fortunately, in Abelian gauge theories gauge xing is much easier to implement than in non-Abelian gauge theories. In contrast to canonical quantization were we chose the Coulomb gauge, in the path integral quantization it is easier to work with a Lorentz-covariant gauge such as the Landau gauge A = 0. (7.1.6) The gauge xing is incorporated as an additional term in the Euclidean action which then becomes SE [A] = 1 1 dd x [ F F + ( A )2 ]. 4 2 (7.1.7)

In the limit 0 only those gauge elds that satisfy the Landau gauge xing condition have a nite action and thus contribute to the path integral. However, well keep the gauge xing parameter as an arbitrary free constant. As well see, physical results are independent of . In momentum space the Euclidean action takes the form SE [A] = 1 (2)d dd p 1 1 A [p2 p p (1 )]A , 2 (7.1.8)

7.2. FEYNMAN RULES FOR SCALAR QED

69

from which we read o the photon propagator [p2 p p (1 1/)]1 . This expression is no longer singular unless . A particularly simple situation emerges in the so-called Feynman gauge, = 1, for which the photon propagator reduces to p2 .

7.2

Feynman Rules for Scalar QED

Let us now couple the electromagnetic eld to charged matter described by a complex scalar eld = 1 + i2 . The total Euclidean action of scalar QED takes the form SE [, A] = 1 1 1 ( A )2 + D D + V (), dd x [ F F + 4 2 2 (7.2.1)

where the covaraint derivative is given by D (x) = ( ieA (x))(x), and the potential for the scalar self-interaction takes the familiar form 1 V () = m2 ||2 + ||4 . 2 4! (7.2.3) (7.2.2)

Since we now have both scalar elds and gauge elds we can consider mixed n-point functions consisting of n scalar elds and n photon elds. The corresponding n-point function is given by 0|T (x1 )...(xn )A1 (x )...An (x )|0 = 1 n 1 DDA (x1 )...(xn )A1 (x )...An (x ) exp(S[, A]). 1 n Z (7.2.4) Again, it is natural to consider the Fourier transform of the n-point function dd x1 ...dd xn dd x ...dd x 0|T (x1 )...(xn )A1 (x )...An (x )|0 1 n 1 n exp[i(p1 x1 ... + pn xn )] exp[i(p x ... + p x )] = 1 1 n n (2)d (p1 + ... + pn + p + ... + p )1 ...n (p1 , ..., pn , p , ..., p ). 1 n 1 n (7.2.5) Again, the -function results from translation invariance.

70 CHAPTER 7. PATH INTEGRAL FOR SCALAR ELECTRODYNAMICS From the terms in the action which are quadratic in the elds one reads o the propagators and from the non-quadratic terms one reads o the vertices. The covaraint derivative term 1 1 D D = ( + ieA ) ( ieA ), 2 2 (7.2.6)

gives rise to a 2-scalar-1-photon and a 2-scalar-2-photon vertex. One arrives at the following Feynman rules for scalar QED. Consider all pairwise contractions in the product of internal and external elds that lead to a connected Feynman diagram. Each diagram has a multiplicity factor that counts the number of pairings leading to the same topology of the diagram. Write down a propagator (p2 + m2 )1 for each oriented external scalar eld i line, and a propagator p 2 (in Feynman gauge) for each external photon line. i
Associate a momentum qi or qi with each oriented internal line and write 2 + m2 )1 for each scalar line and q 2 for each photon line. down a factor (qi i

Write down a momentum conserving -function (2)d (q1 + q2 + q3 + q4 ) for each 4-scalar vertex taking into account the orientation of the internal or external lines. To each vertex of this kind is also associated a factor /4!.
Write down a momentum conserving -function (2)d (q1 + q2 + q1 ) for each 2-scalar-1-photon vertex. To each vertex of this kind is associated a factor e(q1 q2 ). Write down a momentum conserving -function (2)d (q1 + q2 + q1 + q2 ) for each 2-scalar-2-photon vertex. To each vertex of this kind is associated a factor 2e2 .

Finally, integrate the resulting expression over all internal momenta, i.e. write dd q1 ...dd qI /(2)I d dd q1 ...dd qI /(2)I d . The numbers I and I of internal scalara and photon lines can be determined as follows. There are four scalar lines emanating from each of the k4 scalar selfinteraction vertices. There are two scalar and one photon line emerging from a 2-scalar-1-photon vertex of which there are k2 A , and nally there two scalar and two photon lines emanating from a 2-scalar-2-photon vertex of which there are k2 A2 . Hence, one obtains 1 1 I = (4k4 + 2k2 A + 2k2 A2 n ), I = (k2 A + 2k2 A2 n ). 2 2 (7.2.7)

7.2. FEYNMAN RULES FOR SCALAR QED

71

There are I = I + I integrations and k = k4 + k2 A + k2 A2 -functions. Again, one of the -functions reects translation invariance and remains in the nal result. Hence, only k 1 -functions can be used to perform some integrations trivially. The remaining l = I (k 1) = 1 (4k4 + 2k2 A + 2k2 A2 n ) 2 1 + (k 2 + 2k2 A2 n ) k4 k2 A k2 A2 + 1 2 A 1 = k4 + k2 A + k2 A2 n /2 n /2 + 1 (7.2.8) 2

integrations determine the number of loops l in the diagram.

72 CHAPTER 7. PATH INTEGRAL FOR SCALAR ELECTRODYNAMICS

Chapter 8

Lattice Field Theory

When we regularize individual Feynman diagrams, e.g. using dimensional regularization, we quantize a eld theory in the framework of perturbation theory. Even if we could calculate to all orders of perturbation theory (which is, of course, practically impossible), the sum of all perturbative diagrams would not dene the theory nonperturbatively. In order to quantize a eld theory beyond perturbation theory, we must regularize the entire theory at once. A natural nonperturbative regularization arises when we replace the Euclidean space-time continuum by a 4-dimensional hypercubic lattice. The lattice spacing then serves as an ultraviolet cut-o.

8.1

Fermionic Path Integrals and Grassmann Algebras

We have dened the path integral by using the classical action. Theories with fermions have no immediate classical limit, and the denition of the path integral needs special care. The rst step is to dene a so-called Grassmann algebra, which works with anticommuting classical variables i with i 1, 2, ..., N . A Grassmann algebra is characterized by the anticommutation relations {i , j } = i j + j i = 0. An element of the Grassmann algebra is a polynomial in the generators f () = f +
i

(8.1.1)

fi i +
ij

fij i j +
ijk

fijk i j k + ...

(8.1.2)

73

74

CHAPTER 8. LATTICE FIELD THEORY

The fij...l are ordinary complex (or sometimes real) numbers, which are antisymmetric in i, j, ..., l. One also denes formal dierentiation and integration procedures. The dierentiation rules are i = 1, i j = j , j i = j , i i i and integration is dened by di = 0, di i = 1, di dj i j = 1. (8.1.4) (8.1.3)

These integrals are formal expressions. One should not ask over which range of i we actually integrate. The Grassmann algebra we use to dene fermion elds is generated by Grass mann numbers x and x , which are completely independent. The index x runs over all space-time points as well as over all spin, avor or color indices. Let us consider the simplest (completely unrealistic) case of just two degrees of freedom and , and let us perform the Gaussian integral dd exp(m) = dd (1 m) = m. (8.1.5)

Note that the expansion of the exponential terminates because 2 = 2 = 0. When we enlarge the Grassmann algebra to an arbitrary number of elements the above formula generalizes to dx dx exp(x Mxy y ) = D D exp(M) = detM. (8.1.6)

In the two variable case we have dd exp(m) = 1, which generalizes to D D x y exp(My ) = M1 detM. ij (8.1.8) (8.1.7)

8.2

The Fermion Doubling Problem

In the continuum the Euclidean free fermion action is given by S[, ] = dd x (x)( + m)(x). (8.2.1)

8.2. THE FERMION DOUBLING PROBLEM

75

In Euclidean space the Dirac matrices are Hermitean. Again the partition function Z = D D exp(S[, ]) (8.2.2) is a formal expression that requires regularization. On the lattice the continuum fermion eld (x), (x) is replaced by variables x , x that live on the lattice points x. The continuum derivative is discretized by nite dierences, such that 1 S[x , x ] = (x x+ x+ x ) + m x x . (8.2.3) 2 x, x It is instructive to bring this expression to momentum space, because it allows us to read o the lattice fermion propagator (p)(p) = [i

sin p + m]1 .

(8.2.4)

In complete analogy to the scalar eld case we can analyze the fermionic 2-point function for its exponential decay. The energies E of lattice fermions with spatial momentum p show up as poles in the propagator. Here the lattice dispersion relation takes the form sinh2 E =
i

sin2 pi + m2 .

(8.2.5)

Like in the scalar eld case the continuum dispersion relation is recovered in the continuum limit for small E, p and m. However, there are other momenta where E becomes small for small m. These are at the corners of the Brillouin zone at which at least one component of the momentum vector takes the value pi = , because also then sin pi goes to zero. This is in contrast to the scalar eld case, and it is due to the fact that the fermionic action contains only a single derivative. As a consequence, the lattice dispersion relation leads to extra states in the spectrum that are absent in the continuum theory. The extra states do not disappear in the continuum limit, such that the naive lattice discretization of the fermionic action does not lead to the correct continuum theory. The extra states appearing in the lattice dispersion relation show up as extra physical particles. These so-called doubler fermions are a manifestation of a fundamental problem of lattice regularized fermionic theories. This so-called doubling problem leads to a multiplication of fermion species. The above lattice fermion propagator has 2d poles instead of just 1 as in the continuum. Hence, the naively discretized fermion theory contains 2d 1 extra fermion species. The origin of the doubling problem is deeply connected with chiral symmetry and can even be traced back to the Adler-Bell-Jackiw anomaly. The doubler fermions pose a severe problem in lattice eld theory. Without removing them we cannot describe QCD (or the Standard model) which has denitely less than 16 quarks.

76

CHAPTER 8. LATTICE FIELD THEORY

8.3

The Nielsen-Ninomiya No-Go Theorem

Before we try to eliminate the doubler fermions let us prove a general theorem due to Nielsen and Ninomiya: a chiral invariant free fermion lattice action, which is translation invariant, Hermitean, and local, necessarily has fermion doubling. The theorem is based on topology. In fact, it holds because the lattice momentum space (the Brillouin zone) is a torus. A general chiral and translation invariant lattice action leads to a propagator (p)(p) = [i

(p)]1 .

(8.3.1)

Of course, the mass term is now absent because it breaks chiral symmetry. The above action is Hermitean only if (p) is real. Locality of the action, i.e. exponential suppression of couplings at large distances, implies that (p) is regular in momentum space, i.e. there are no poles or discontinuities. Most important, translation invariance implies that (p) is periodic over the Brillouin zone. Poles of the propagator and hence physical or doubler fermions correspond to points p with (p) = 0 for all . Hence, the theorem states that with the above properties, (p) = 0 at more than just one point. Let us prove this rst in d = 1. Then there is a single regular periodic function 1 (p), which should at least have one zero to describe the physical fermion pole. Then the function is positive on one side of the zero and negative on the other side, such that it must go through zero again in order to satisfy periodicity. A double zero (with the function not changing sign) is not allowed, because this would lead to a wrong dispersion relation for the physical fermion. In higher dimensions the proof is analogous. In d = 2 there are two functions 1 (p) and 2 (p). The zeros of 1 (p) lie on a closed curve in the 2-dimensional Brillouin zone. This curve may be closed via the periodic boundary conditions. The zeros of 2 (p) lie on another closed curve that intersects the rst one in the pole position of the physical fermion. Due to the periodic boundary conditions of the Brillouin zone, the curves than necessarily must also intersect somewhere else. Again, the curves cannot just touch each other, because this would lead to a wrong dispersion relation. In d dimensions the zeros of (p) (with = 1, 2, ..., d) lie on d closed d 1-dimensional manifolds. Those cannot intersect in just one point. If they intersect once they necessarily intersect also somewhere else. This proves lattice fermion doubling for a chiral invariant lattice action. The theorem does not specify the number of doubler fermions. It is denitely possible to reduce the number of doublers from 2d 1 to 1, but it is impossible to eliminate doublers completely. Of course, one may try to evade the theorem by violating one of its basic assumptions. In fact, people have worked on random lattices because it has no translation invariance,

8.4. WILSON FERMIONS

77

and hence no periodic momentum space. However, there is not much one can do analytically on a random lattice, even for a free fermion, and it is unclear if the consequences of the no-go theorem are indeed evaded. When one violates Hermiticity one looses contact with Minkowski space and it is unclear what the Euclidean results mean. People have also worked with nonlocal actions. Then (p) is not a continuous function (it may have poles or discontinuities), and the theorem which relies on topology obviously does not apply. Still, it has turned out that most nonlocal actions have serious problems. For example, the resulting continuum theory may be nonlocal or not Lorentz invariant. In work on lattice gauge theory Wilson removed the fermion doublers in a very direct and radical way, simply by breaking chiral symmetry explicitly. Then the theorem is evaded because the propagator then contains terms without . Of course, this leads to all kinds of complications. In particular, it is nontrivial that chiral symmetry will be recovered in the continuum limit. For vector-like theories like QCD this is under control. For chiral theories like the Standard model, however, it is still unclear if we can dene them nonperturbatively due to the notorious fermion doubling problem.

8.4

Wilson Fermions

In his original work on lattice gauge theory Wilson has avoided the fermion doubling problem by breaking chiral symmetry explicitly. The so-called Wilson term gives the doubler fermions a mass at the order of the cut-o while the physical fermion remains massless. Hence, in the continuum limit chiral symmetry is restored in the physical sector. Wilsons modication of the naive fermion action takes the form of a discretized second derivative S[, ] =
x,

1 (x x+ x+ x ) + 2

x x
x

1 2

x,

(2x x x x+ x+ x ).

(8.4.1)

Then the lattice propagator takes the form (p)(p) = [i

sin p + m +

1 2

(2 sin(p /2))2 ]1 .

(8.4.2)

It is instructive to work out the corresponding dispersion relation, and in particular to determine the fermion mass as a function of the bare mass m. The Wilson term acts as a momentum dependent mass term. For small momenta it

78

CHAPTER 8. LATTICE FIELD THEORY

vanishes quadratically, and hence it does not eect the dispersion of the physical fermion at least in the continuum limit. For the doubler fermions, on the other hand, the Wilson term is nonzero, and eectively gives them a mass of the order of the cut-o. In the continuum limit the doubler fermions are hence eliminated from the spectrum of the theory.

8.5

Abelian Lattice Gauge Fields

Let us consider an Abelian gauge eld a in the continuum. A gauge transformation then amounts to a (y) = a (y) + (y). (8.5.1)

Since a gauge eld is described by a Lorentz vector the corresponding lattice eld is naturally associated with the links connecting neighboring lattice points. We denote the lattice gauge potential by A,x where x denotes the center of the corresponding lattice link. The gauge transformation equation for Abelian lattice gauge eld then takes the form A = A,x + x/2 x+/2 , ,x (8.5.2)

where x is the lattice gauge transformation associated with the lattice point x. The derivative of the continuum gauge transformation has been replaced by a nite dierence, and we have again put the lattice spacing to 1. The eld strength of the continuum gauge eld is f (y) = a (y) a (y), while the lattice eld strength is F,x = A,x+/2 A,x/2 A,x+/2 + A,x/2 . (8.5.4) (8.5.3)

Here x denotes the center of a lattice plaquette (an elementary lattice square) such that the lattice eld strength is the oriented sum of gauge links around the plaquette. In complete analogy with the continuum action the standard lattice gauge action takes the form S[A ] = 1 2 F . 4e2 ,x ,x (8.5.5)

Again one can construct a perfect gauge action by blocking from the continuum. The natural way to block a continuum gauge eld is A,x =
cx/2

dd y (1 + y x )a (y) +

cx+/2

dd y (1 y + x )a (y). (8.5.6)

8.6. THE NOTION OF LATTICE DIFFERENTIAL FORMS

79

It is instructive to prove that with this blocking a continuum gauge transformation (y) induces a lattice gauge transformation x =
cx

dd y (y).

(8.5.7)

The derivation of the perfect gauge action is in complete analogy to the scalar and fermionic cases, and will not be presented here. As in the scalar eld case, the perfect action turns out to be local.

8.6

The Notion of Lattice Dierential Forms

Lattice elds have a natural place to live on the lattice. For example, gauge potentials live on the links, while gauge transformations live on the lattice points and eld strength variables live on plaquettes. The notion of lattice dierential forms is based on this fact, and it allows to perform complicated manipulations of Abelian lattice elds in a simple and transparent manner. Let us consider a 4-dimensional hypercubic lattice. It decomposes naturally into k-dimensional oriented elementary cells ck . These are the lattice points c0 , the links c1 , the plaquettes c2 , the cubes c3 and the hypercubes c4 . The dual lattice is obtained by shifting the original lattice by half a lattice spacing in each direction. The dual lattice consists of (4 k)-dimensional dual cells ck . The points c4 of the dual lattice are dual to the hypercubes of the original lattice, the links c3 are dual to the original cubes, the plaquettes c2 are dual to the original plaquettes, etc. On the cells ck one denes elds the so-called k-forms (ck ). For example, a gauge transformation is a 0-form because it lives on points, while a gauge potential is a 1-form because it lives on links. To each k-form there is associated a dual (4 k)-form simply by

( ck ) = (ck ).

(8.6.1)

The exterior dierential d increases the rank of a form by one d(ck+1 ) =

ck ck+1

(ck ),

(8.6.2)

i.e. d is a (k + 1)-form dened on the cells ck+1 . It is a sum of the k-forms living on the oriented boundary of ck+1 . For example, the eld strength is a 2-form constructed as F = dA, (8.6.3)

80

CHAPTER 8. LATTICE FIELD THEORY

and a gauge transformation takes the form A = A + d. (8.6.4)

Applying the exterior dierential twice always gives d2 = 0. For example, the Abelian eld strength is gauge invariant because F = dA = dA + d2 = dA = F. (8.6.5)

The codierential = d lowers the rank of a form by one and is given by (ck1 ) =
c k ck1

(ck ).

(8.6.6)

For example, the continuum Landau gauge condition a (y) = 0 takes the form A = 0 on the lattice. The Laplacian = d + d leaves the rank of a form unchanged. Each form can be written as a Hodge decomposition = d1 + 1 d. One denes a scalar product of two k-forms and by (, ) =
ck

(8.6.7)

(ck )(ck ).

(8.6.8)

This induces a norm ||||2 = (, ). For example, the standard lattice gauge action takes the form S[A] = 1 1 ||F ||2 = 2 (dA, dA). 2e2 2e (8.6.9)

The rule for partial integration on the lattice simply reads (d, ) = (, ), (8.6.10)

where is a k-form, and is now a (k + 1)-form. For a gauge eld in the Landau gauge one can rewrite the gauge action as S[A] = 1 1 1 (A, dA) = 2 (A, dA + dA) = 2 (A, A). 2 2e 2e 2e (8.6.11)

The notion of dierential forms is very useful, in particular because it allows us to perform partial integrations on the lattice in a simple manner.

8.7. WILSON LOOPS AND THE LATTICE COULOMB POTENTIAL

81

8.7

Wilson loops and the Lattice Coulomb Potential

Let us consider an Abelian lattice gauge theory (a theory of free photons) coupled to external electric currents. The currents may represent innitely heavy charged particles. In the continuum the interaction of a gauge eld with an external current is described by dd y j (y)a (y). Expressed in form language the corresponding lattice expression reads (J, A). Let us consider the lattice current of a heavy charged particle moving along the links of the lattice. The current is then 1 on the links that form the world line of the particle, and zero on all other links. The Wilson loop observable is associated with the propagation of a charge-anticharge pair that is created and later annihilated. The corresponding world line is a closed loop C and the so-called Wilson loop is given by WC = exp(i(J, A)). (8.7.1)

The expectation value of a rectangular Wilson loop of spatial size R and temporal size T is related to the potential V (R) of a static charge-anticharge pair
T

lim WC exp(V (R)T ).

(8.7.2)

Let us calculate this quantity for an Abelian gauge theory. The lattice quantized path integral 1 (8.7.3) dA(c1 ) exp( 2 ||F ||2 ) Z= 2e c
1

is still undened because the implicit integration over all gauges results in an innite factor. The quantization of Abelian (noncompact) gauge elds on the lattice does indeed require gauge xing. Here we choose to work in the Landau gauge, i.e. we introduce a -function constraint in the functional integral

Z=
c1

dA(c1 )
c0

(A(c0 )) exp(

1 ||F ||2 ). 2e2

(8.7.4)

This expression is now well-dened. In the Landau gauge we can write the Wilson loop expectation value as WC = 1 Z

dA(c1 )
c1 c0

(A(c0 )) exp(

1 (A, A) + i(J, A)) 2e2 (8.7.5)

exp(

e2 2

(J, 1 J)).

Going to momentum space and performing the large T limit one can derive the lattice Coulomb potential from this expression. The lattice Coulomb potential is

82

CHAPTER 8. LATTICE FIELD THEORY

slightly distorted at short distances due to discretization artifacts, but it agrees with the ordinary Coulomb potential at large separations. The Wilson-loop expectation value is an order parameter of lattice gauge theory. In the case of an Abelian gauge theory it indicates that we are in an ordinary Coulomb phase with the typical interactions between static charged particles.

8.8

Lattice QED

To construct the action of lattice QED we have to add interaction terms to the free fermion (electron) and free gauge eld (photon) actions from before. In particular, the free Wilson fermion action S[, ] =
x,

1 (x x+ x+ x ) + 2

m x x
x

1 2

x,

(2x x x x+ x+ x )

(8.8.1)

is not gauge invariant. Under a U (1) gauge transformation the fermion eld transforms as = exp(ix )x , (8.8.2) x while the gauge eld transforms as A = A,x + x+/2 x/2 . ,x (8.8.3)

To make the fermion action gauge invariant we now introduce the parallel transporter eld U,x = exp(iA,x ) U (1). (8.8.4) Under a gauge transformation it transforms as
U,x = exp(ix/2 )U,x exp(ix+/2 ),

(8.8.5)

such that S[, , U ] =

x,

1 + (x Ux+/2, x+ x+ Ux+/2, x ) 2 m x x

+
x

1 2

x,

+ ) (2x x x Ux+/2, x+ x+ Ux+/2, x(8.8.6)

8.8. LATTICE QED

83

is indeed gauge invariant and has the correct naive continuum limit. Note that + here denotes complex conjugation. The concept of a parallel transporter connecting neighboring lattice sites in a gauge covariant way extends naturally to the continuum. In fact, one can dene a parallel transporter along any curve C connecting the points x0 and x1 as UC = exp(i
C

dy a (y)).

(8.8.7)

Under a continuum gauge transformation a (y) = a (y) + (y) the parallel transporter transforms as
UC = exp(i(x0 ))UC exp(i(x1 )).

(8.8.8)

(8.8.9)

The Wilson loop dened above is nothing but a parallel transport around a closed curve. When one adds the free photon action to the above expression for the electron-photon interaction one obtains the action of lattice QED. Lattice QED is of theoretical interest because it is unknown how QED behaves for large values of the electric charge. This can only be investigated in a nonperturbative formulation such as lattice gauge theory. Numerical simulations suggest that the renormalized electric charge never becomes large. In fact, some calculations indicate that QED is trivial, i.e. the renormalized electric charge vanishes in the continuum limit. To locate the continuum limit in such calculations it is essential to measure the electron mass, i.e. the inverse correlation length of the electron eld. The naive correlation function x y , however, is not gauge invariant, and therefore does not contain information about the electron mass. The reason for the gauge dependence is that the physical electron consists not just of the bare electric charge carried by the fermion eld, but also of the photon cloud surrounding the charge. The charged eld of the physical electron can be constructed as follows c,x = exp(i(Ex , A))x . (8.8.10)

Here Ex is a 1-form describing the electric eld of a static charge located at the point x, i.e. Ex = x , (8.8.11) where x is a 0-form which is 1 at x and zero otherwise. The eld Ex is localized in the specic time slice dened by xd , and it is zero in all other time slices. This is very essential, because otherwise the denition of the electron eld would

84

CHAPTER 8. LATTICE FIELD THEORY

require information about the Euclidean past or future. This would destroy the causal structure of the theory, and would prevent us from constructing a transfer matrix, and hence from rotating our physical results back into Minkowski spacetime. Under a gauge transformation one nds
c,x = exp(i(Ex , A ))x = exp(i(Ex , A + d)) exp(ix )x

= exp(i(Ex , )) exp(ix )c,x = c,x.

(8.8.12)

Here (Ex , ) = (x , ) = x has been used. Hence the physical charged eld is gauge invariant. Its correlation function c,x c,y does indeed decay exponentially with the electron mass.

8.9

Lattice QCD

The concept of the parallel transporter is even more important in non-Abelian gauge theories. Not even the pure gluon part of lattice QCD can be formulated in terms of vector potentials G,x . Instead, the whole action is formulated in terms of link parallel transporters Ux, = exp(iGa a ) = exp(iG,x ) SU (3). ,x (8.9.1)

Again, the parallel transporter has a meaning already in the continuum. In nonAbelian gauge theory it arises as a path ordered exponential UC = P exp(i dy G (y)) = lim
C

exp(iG (y)).
yC

(8.9.2)

It is instructive to prove that under a non-Abelian gauge transformation G (y) = g(y)(G (y) + )g(y)+ , the parallel transporter transforms as
UC = g(x0 )UC g(x1 )+ ,

(8.9.3)

(8.9.4)

in complete analogy to the Abelian case. Consequently, the quark-gluon interaction part of the lattice QCD action has exactly the same form as the electronphoton interaction part of the lattice QED action of eq.(8.8.6). Of course, now the parallel transporters are SU (3) matrices and the fermion elds have a color index (which is suppressed in our notation).

8.9. LATTICE QCD

85

The pure gluon part of the action, on the other hand, is quite dierent from the free photon action in QED. This is because the gauge transformation equation (8.9.3) cannot be discretized in a gauge covariant way, when one uses the vector potential G . Instead Wilson has constructed the pure gluon part of the lattice QCD action again using parallel transporters. First, he builds a parallel transporter around an elementary plaquette
+ + U,x = U,x/2 U,x+/2 U,x+/2 U,x/2 .

(8.9.5)

Then the action is constructed as a sum over all plaquettes S[U ] = 1 Re Tr(1 U,x ). 2 2gs ,x (8.9.6)

It is a good exercise to prove that this expression reduces to the correct continuum action 1 (8.9.7) S[G ] = dd y 2 TrG (y)G (y). 2gs . To fully dene the lattice QCD path integral we also must consider the measure. For the quarks we have the ordinary Grassmann variable integrations. For the gluons we must integrate over the congurations of parallel transporters U . To discuss this, let us concentrate on a pure gluon theory. The lattice partition function of that theory is given by Z=
x, SU (3)

dU,x exp(S[U ]).

(8.9.8)

We integrate independently over all link variables using the so-called Haar measure dU,x for each parallel transporter. The Haar measure is a left and right invariant measure, i.e. dU f (gL U ) =
SU (3) SU (3)

dU f (U gR ) =
SU (3)

dU f (U ),

(8.9.9)

for any function f and for any SU (3) matrices gL and gR . It is convenient to normalize the measure such that dU 1 = 1.
SU (3)

(8.9.10)

One can show that the Haar measure is unique and that it has the following additional properties dU Uij = 0,
SU (3)

1 dU Uij Ukl = jk il . 3 SU (3)

(8.9.11)

86

CHAPTER 8. LATTICE FIELD THEORY

Here ij determines a specic matrix element, i.e. i, j, k, l {1, 2, 3}. For compact groups like SU (3) the integration is over a nite domain. This is in contrast to QED in which we integrated the gauge potential A over all real numbers. This made it necessary to x the gauge in lattice QED. In lattice QCD, on the other hand, the path integral is nite from the beginning, and gauge xing is not necessary. This is a big advantage of the formulation using parallel transporters. The above pure gluon partition function contains a single parameter the bare gauge coupling gs . When we want to perform the continuum limit, we must search for values of gs for which the correlation length of the lattice theory diverges. As always, we are looking for a second order phase transition. To analyze the phase structure of a gauge theory, one needs to study order parameters like the magnetization sx in the Ising model. A simple local order parameter like U,x does not make sense in a gauge theory. This follows from Elitzurs theorem, which states that gauge non-invariant observables vanish in lattice gauge theory (formulated with parallel transporters). This is trivial to prove using eq.(8.9.11). A useful order parameter in a gauge theory must be gauge invariant and, in addition, nonlocal. A good order parameter in the pure gluon theory is the Wilson loop WC = Tr
,xC

U,x .

(8.9.12)

For a rectangular curve with side lengths R and T the Wilson loop describes the instantaneous creation and annihilation of a static quark-antiquark pair at distance R that lives for a time T . As in QED the Wilson loop is related to the static potential
T

lim WC exp(V (R)T ).

(8.9.13)

In QCD we expect quarks and antiquarks to be conned to one another by a potential rising linearly at large separations R, i.e.
R

lim V (R) R,

(8.9.14)

where is the string tension. In a connement phase the Wilson loop hence shows an area law
R,T

lim

WC exp(RT ).

(8.9.15)

If connement would be lost (by going through a phase transition) the Wilson loop would no longer show area law behavior. A Coulomb phase, for example, is characterized by a perimeter law.

8.10. CONFINEMENT IN THE STRONG COUPLING LIMIT

87

8.10

Connement in the Strong Coupling Limit

In lattice gauge theory it is straight forward to prove connement for large values of the bare gauge coupling gs . In the strong coupling region, however, the correlation length is small and we cannot take the continuum limit. In fact, due to asymptotic freedom we expect the continuum limit to be at gs 0. It is an open question if connement persists in the continuum limit. However, there is a lot of positive numerical evidence that this is indeed the case. Assuming that there is no phase transition between the strong and weak coupling regions, the derivation of connement in the strong coupling regime would carry over to the continuum limit. In the strong coupling expansion we expand in powers of 1/gs around gs = . To leading order the pure gluon action is then simply zero. The Wilson loop operator takes the form WC = U1ij U2jk U3kl ...UN mi , (8.10.1)

where N = 2(R + T ) is the number of links along the loop. Using the rst formula of eq.(8.9.11) then immediately implies WC = 0. The second formula of eq.(8.9.11) suggests that to get a nonzero result one needs to have the product of two matrix elements for each link variable on the Wilson loop. This can be achieved by expanding the Boltzmann factor of the action to higher orders in 1/gs . The lowest nonzero contribution comes from tiling the Wilson loop with plaquettes that result from expanding exp( 3 1 1 Re Tr(1 U,x )) = exp( 2 )(1 + 2 Re TrU,x ) 2 2gs 2gs 2gs (8.10.2)

for the plaquettes in the interior of the Wilson loop. Taking the U,x term for all these plaquettes and using the second formula of eq.(8.9.11) gives WC = 1
2 (gs )RT

(8.10.3)

Indeed we nd an area law and we read o the string tension as = log( 1 ). 2 gs (8.10.4)

Higher order corrections arise from deformations of the simple tiling from above. The leading correction comes from eliminating one of the RT plaquettes of the original tiling, and replacing it with ve plaquettes at the surface of an elementary lattice cube, such that the resulting surface bounded by the Wilson loop has no holes. The cube can be attached above and below the plane of the Wilson loop,

88

CHAPTER 8. LATTICE FIELD THEORY

and it can go out in the two possible orthogonal directions. This results in 4RT contributions, which all have four more plaquettes than the leading term, and 2 are hence suppressed by 1/(gs )4 . Up to that order one nds WC = 1
2 (gs )RT

(1 + 4RT

1 ). 2 (gs )4

(8.10.5)

The last two terms are the rst term in an exponential, and hence we can read o the corrected string tension as = log( 1 1 ) + 4 2 4. 2 gs (gs ) (8.10.6)

The string tension has been computed to higher orders. Still, one is unable to go to high enough orders to reach the physically interesting scaling region in which we can perform the continuum limit. At present, only numerical simulation techniques are powerful enough to enter that regime. Consequently, the above result for the string tension cannot yet be compared with experimental results. It is instructive to derive the glueball mass in a similar manner. For this purpose one considers the correlation function of two elementary plaquette operators separated in Euclidean time. Using the strong coupling expansion one can identify the exponential decay, and extract the correlation length whose inverse is the massgap or glueball mass.

8.11

Connement in Compact Abelian Gauge Theory

The above strong coupling calculations prove that lattice QCD connes at large values of the bare gauge coupling. However, the calculation does not really hint at the dynamical mechanism that connes color. It is a subject of ongoing research to understand the connement mechanism in non-Abelian gauge theories. In abelian gauge theories, on the other hand, the issue is completely understood, even analytically. Of course, as we saw earlier, the Wilson loop in an Abelian gauge theory yields the Coulomb potential, which does not conne in d = 4. It is, however, possible to modify the formulation of Abelian gauge theories, such that they also conne at strong coupling. This is achieved simply by working with parallel transporters in the pure gauge action, as we were forced to do in nonAbelian theories. This leads to the compact formulation of Abelian lattice gauge theory, which is very similar to Wilsons lattice formulation of non-Abelian gauge theory. The use of (compact) parallel transporters instead of (noncompact) vector potentials has drastic consequences in Abelian gauge theory. As we will see, the compact theory has magnetic monopoles as additional degrees of freedom, which

8.11. CONFINEMENT IN COMPACT ABELIAN GAUGE THEORY

89

turn out to be responsible for connement. In fact, the monopoles condense in the connement phase (like Cooper pairs condense in a superconductor), while they are ordinary charged particles in the Coulomb phase. As we will see, the connement mechanism in a compact Abelian gauge theory is the dual of the Meissner eect in a superconductor. An external magnetic eld is constricted to ux strings within the superconductor. The ux strings cost energy proportional to their length, i.e. they have a nonzero string tension. If there would be single magnetic monopoles, in a superconductor their magnetic ux would form a string ending in an antimonopole, and hence monopoles and antimonopoles would be conned to one another. In QCD color electric charges (quarks and gluons) are conned. This could be described by a dual Meissner eect. The Meissner eect in a superconductor arises because electric charges (Cooper pairs) are condensed. A dual superconductor would require the condensation of magnetic charges (monopoles). This is exactly what happens in compact QED. It is still unclear if the same mechanism is responsible for connement in QCD, although there is some positive numerical evidence. To understand the role of the monopoles let us now consider 4-d compact U (1) lattice gauge theory. This is lattice QED without electrons, but as we will see with magnetic charges. For simplicity we consider the theory in the so-called Villain formulation, for which the partition function takes the form

Z=
c1

d(c1 )
c2 n(c2 )Z

exp(

1 d + 2n 2 ), 2e2

(8.11.1)

where the gauge eld is a 1-form on the links c1 , n is a 2-form on the plaquettes c2 and e is the bare electric charge. First we replace the link integration in the path integral by an integration over the plaquette angles = dmod2. The Bianchi identity is implemented by a -function constraint

Z =
c2

d (c2 )
c3 k(c3 )Z

1 d k) 2

exp(
c2 n(c2 )Z

1 + 2n 2 ) 2e2 d (c2 )

=
c3 k(c3 )Z c2

exp(
c2 n(c2 )Z

1 + 2n 2 ) exp(i(d , k)). 2e2 (8.11.2)

90

CHAPTER 8. LATTICE FIELD THEORY

The constraint has been eliminated using the Poisson summation formula. Thereby we have introduced an integer valued 3-form k. Now we perform the integration over the variable F = + 2n, and we obtain

Z =
c3 k(c3 )Z c2

dF (c2 ) exp( dF (c2 ) exp(

c3 k(c3 )Z c2

1 F 2e2 1 F 2e2

) exp(i(dF, k)) ) exp(i(F, d k)) (8.11.3)

= =
c 3 k( c3 )Z

exp(

e2 2 d k ). 2

Eq.(8.11.3) describes a Z gauge theory with dual link variables k and with a quadratic action. The dual Z gauge theory is the limit of scalar QED with gauge group R. The dual noncompact gauge eld A R is a 1-form on the dual links c3 , and the complex scalar eld x = exp(i x ) is dened in terms of the dual 0-form ] , ]. To make the action periodic in we also introduce a dual integer valued 1-form k Z such that the partition function reads

Z =
c 3

d A( c3 )
c 4

d ( c4 )
c 3 k( c3 )Z

exp(

1 d A 2g2

d + 2 k A 2 ). 2

(8.11.4)

The scalar eld x carries the bare charge g. In the unitary gauge = 0 the partition function is

Z =
c 3

d A( c3 )
c 3 k( c3 )Z

exp(

1 d A 2g2

2 k A 2 ). 2

(8.11.5)

Now it is clear that for only congurations with A = 2 k contribute to the functional integral and that we thus recover the dual Z gauge theory with Z=
c 3 k( c3 )Z

exp(

e2 2 d k ) 2

(8.11.6)

for e = 2/g. This is exactly the Dirac quantization condition for the bare electric and magnetic charges. Since the scalar particle of the dual gauge theory carries

8.11. CONFINEMENT IN COMPACT ABELIAN GAUGE THEORY

91

the charge g, it is natural to identify it with the magnetic monopole of the original model. We will see later that such an identication is indeed justied. Lscher has derived an inequality between the bare and renormalized charge u in noncompact scalar QED in the Coulomb phase gr g. (8.11.7)

Here gr is the renormalized magnetic charge. The Dirac quantization condition relates the renormalized electric and magnetic charges by 2 er = . (8.11.8) gr Duality in the limit together with the Dirac quantization condition thus turns the inequality eq.(8.11.7) into an inequality between the bare and the renormalized electric charge in compact U (1) lattice QED er = 2 2 = e. gr g (8.11.9)

Consequently, magnetic monopoles lead to antiscreening of electric charge, as opposed to the screening of electric charges in ordinary QED. Now it becomes clear why the presence of monopoles can lead to connement of electric charges. Because monopoles enhance the renormalized electric charge, they can drive the theory to strong coupling, where Coulombs law breaks down and connement sets in. This is in fact what happens in compact lattice QED. At small bare electric charges the theory is in the Coulomb phase. Then the monopoles are stable magnetically charged particles, and external electric charges interact via Coulomb forces. For large bare electric charges the model is in the conned phase, in which the monopoles are condensed and magnetic charge is not a good quantum number. In the dual formulation the conned phase is equivalent to the Higgs phase of scalar QED, in which the scalar charged particle (the monopole) is condensed. The Coulomb phase of compact QED, on the other hand, corresponds to the Coulomb phase of noncompact scalar QED. In the dual formulation of the theory the monopole is an ordinary charged scalar eld. Therefore it is straightforward to construct a creation operator for it. As discussed before the bare charged eld x cannot create the physical charged state, because it is not gauge invariant. x only creates the bare monopole, but not the dual photon cloud surrounding it. The gauge dependence of x is xed by going to the Coulomb gauge. We perform a gauge transformation in the time slice containing the point x

A =

A + d , (8.11.10)

= ( + )mod2.

92

CHAPTER 8. LATTICE FIELD THEORY

Coulomb gauge xing generates the dual Coulomb eld surrounding the monopole. The appropriate gauge transformation at the point x is given by

x = ( Bx , A), Bx = x ,

(8.11.11) (8.11.12)

with where x is the Kronecker delta function on the dual lattice, i.e. x =
B

1 for y = x 0 otherwise.

(8.11.13)

x is the magnetic eld of a monopole located at the point x. It is nonzero only on the dual links located in the time slice containing the monopole creation point x. The explicit form of Bx is given by

Bx = d3 1 x , 3

(8.11.14)

where d3 is the 3-dimensional exterior dierential and 3 is the 3-dimensional Laplacian. In the Coulomb gauge the charged eld is c,x = exp(i ) = exp(i( Bx , A))x . x (8.11.15)

This has exactly the same form as the charge creation operator discussed earlier, except that we are now using a dual formulation. If we again go to the limit and x to unitary gauge x = 1, A = 2 k we nd c,x = exp(2i( Bx , k)). (8.11.16) This is the monopole eld from which one can construct all monopole Green functions. Now we will investigate the vacuum expectation value of the monopole eld 1 e2 2 c,x = c,x exp( d k ) (8.11.17) Z c 2
3

k( c3 )Z

after transforming back to the original compact QED. From there one arrives at a Coulomb gas representation of the correlation function, which makes the creation and annihilation of monopoles explicit. When one performs the duality transformation again one goes back to the original compact lattice QED. The expectation value of the monopole eld then takes the form 1 d(c1 ) c,x = Z c c
1 2

n(c2 )Z

1 exp( 2 d + 2n + 21 (B ) 2 ), 2e

(8.11.18)

8.11. CONFINEMENT IN COMPACT ABELIAN GAUGE THEORY

93

The integer valued 3-form Z is a remnant of the Dirac string which emanates from the monopole position x. It obeys

(d) = x

(8.11.19)

such that d(B ) = 0. B and give rise to a plaquette shift 21 (B ) which acts as an external background eld. In the present form it is unclear why eq.(8.11.18) describes the creation of monopoles. Therefore we now rewrite the theory as a Coulomb gas of monopole world lines. First we perform a Hodge decomposition of n n = d1 n + 1 dn = d1 n + 1 m, (8.11.20)

where the monopoles are described by the 3-form m = dn. The dual 1-form m describes monopole world lines on the dual links, which form closed loops because of the continuity equation for magnetic charge m =

dm =

d n = 0.

(8.11.21)

When n is shifted to n = n + dl, where l Z is a 1-form we still have m = dn = dn + d2 l = dn = m. One may thus eliminate n from the path integral in favor of m and l c,x = exp( 1 Z

(8.11.22)

d(c1 )
c3 m(c3 )Z,dm=0 c1 l(c1 )Z c1

1 d( + 2l + 21 n) + 21 (m + B ) 2 ). 2e2 (8.11.23)

Introducing a noncompact gauge eld A = + 2l + 21 n one now writes c,x = 1 Z

dA(c1 )
c3 m(c3 )Z,dm=0 c1

exp(

1 dA + 21 (m + B ) 2 ). 2e2

(8.11.24)

Using partial integration and d(m + B ) = 0 one nds c,x = 1 Z exp(

c3 m(c3 )Z,dm=dB

2 2 (m + B, 1 (m + B))) e2 (8.11.25)

94 with Z=

CHAPTER 8. LATTICE FIELD THEORY

exp(
c3 m(c3 )Z,dm=0

2 2 (m, 1 m)). e2

(8.11.26)

The expectation value is expressed as a ratio of two partition functions. Z in the denominator describes an ensemble of closed monopole world lines. The monopoles interact with each other via long-range Coulomb forces represented by the inverse Laplacian. In the numerator there is, in addition, one open monopole world line that starts at x. Now it is clear that a monopole has indeed been created at the point x. This shows explicitly that the expectation value, which was originally constructed in the dual scalar QED, does indeed describe the creation of the topological excitations of compact QED. Note that the invisibility of the Dirac string is now obvious, because has disappeared from the nal expression. Because of the Dirac quantization condition strong electric couplings e correspond to weak magnetic couplings g. Hence, the strong coupling limit of the original compact QED corresponds to the weak coupling limit of the dual noncompact QED. We can make use of this to show that monopoles are indeed condensed in the conned phase. For this purpose we investigate the monopole eld expectation value c,x in the strong coupling e = limit, which corresponds to the weak coupling g = 0 limit in the dual formulation. In this limit only the congurations with zero dual action contribute to the path integral. Then k = 0 and consequently c,x = 1, (8.11.27) such that monopoles indeed condense in the conned phase.

8.12

The Monte Carlo Method

The most interesting questions in lattice eld theory especially those that may eventually lead to a solution of QCD cannot be answered analytically. For example, the strong coupling expansion does not converge well in the weak coupling scaling region, in which we want to take the continuum limit of lattice QCD. Fortunately, the close analogy with classical statistical mechanics allows us to use other techniques developed there. A powerful numerical technique to solve problems in statistical mechanics is the so-called Monte Carlo method. The idea is to compute the partition function (do the path integral) by generating eld congurations numerically. Of course, the path integral is an extremely high dimensional integral, such that doing it with standard numerical integration techniques is completely hopeless. In the Monte Carlo method predominantly

8.12. THE MONTE CARLO METHOD

95

those eld congurations are generated that have the largest contribution to the path integral. In fact, the Boltzmann factor exp(S[]) is used as the probability to generate the eld conguration . In a Monte Carlo simulation one generates a sequence of eld congurations (1) (2) ... (N ) , (8.12.1)

which form a so-called Markow chain, by applying an algorithm that turns the conguration (i) into (i+1) . The initial conguration (1) is either picked at random or selected otherwise. Ultimately, nothing should depend on this choice. After a (possibly large) number M of Monte Carlo iterations (applications of the algorithm) an equilibrium is reached, and the system has forgotten about the initial congurations. Only the congurations generated after equilibration are used in the actual calculation. To estimate the expectation value of some observable one averages its values over all congurations of the Monte Carlo sample N 1 O[(i) ]. (8.12.2) O[] N M
i=M +1

In the limit N the approximation becomes exact. At nite N M one makes a calculable numerical error that decreases proportional to 1/ N M . Hence, to increase the numerical accuracy by a factor of two one must run the Monte Carlo algorithm four times as long. The Boltzmann factor exp(S[]) is not explicitly included in the above sum. It is implicitly included, because the congurations in the Markow chain occur with probability exp(S[]). To demonstrate that a particular Monte Carlo algorithm converges to the correct equilibrium distribution it is sucient to show that it is ergodic and obeys detailed balance. Ergodicity means that starting from an arbitrary initial conguration the algorithm can in principle reach any other eld conguration. This condition is obviously necessary, because the correct value for the path integral can be obtained only if all eld congurations (of nite action) are included. Detailed balance means that exp(S[])w[, ] = exp(S[ ])w[ , ]. (8.12.3)

Here w[, ] is the transition probability for the algorithm to turn the conguration into . A Monte Carlo algorithm is completely characterized by its w[, ]. Since the algorithm denitely generates a new conguration the proper normalization is D w[, ] = 1. (8.12.4)

96

CHAPTER 8. LATTICE FIELD THEORY

When the Monte Carlo algorithm converges to an equilibrium distribution p[] of eld congurations, this distribution is an eigenvector of w[, ] with eigenvalue 1 D p[]w[, ] = p[ ]. (8.12.5) Now we want to show that the canonical Boltzmann distribution p[] = exp(S[]) (8.12.6)

is indeed an eigenvector of w[, ] if the algorithm obeys detailed balance. We nd D exp(S[])w[, ] = D exp(S[ ])w[ , ] Dw[ , ] (8.12.7)

= exp(S[ ]) = exp(S[ ]).

Assuming ergodicity one can show that only one eigenvector with eigenvalue 1 exists, and that the equilibrium distribution is therefore unique. A simple example of an algorithm that is ergodic and obeys detailed balance is the so-called Metropolis algorithm. In this algorithm a new conguration is randomly chosen in the vicinity of the old conguration . If the action of the new conguration is smaller than the action of the old conguration, the new conguration is accepted, i.e. S[ ] < S[] w[, ] = 1. (8.12.8)

On the other hand, if the new action is larger, the new conguration is accepted only with a certain probability, i.e. S[ ] > S[] w[, ] = exp(S[ ] + S[]). (8.12.9)

Otherwise the old conguration is kept. This algorithm obeys detailed balance. Let us consider two congurations and . We can assume that S[ ] < S[] such that w[, ] = 1. Then of course, S[] > S[ ] such that w[ , ] = exp(S[] + S[ ]), and hence exp(S[])w[, ] = exp(S[]) = exp(S[ ]) exp(S[] + S[ ]) = exp(S[ ])w[ , ]. (8.12.10)

The Metropolis algorithm is particularly simple, but not very ecient. It turns out that subsequent congurations in the Markow chain are correlated with

8.12. THE MONTE CARLO METHOD

97

each other. Hence, to generate a new statistically independent eld conguration may require a large number of Monte Carlo iterations. The autocorrelation time of the Metropolis algorithm actually increases when one approaches a second order phase transition (or equivalently the continuum limit). At a second order phase transition the correlation length diverges (or equivalently the particle mass in lattice units vanishes). One nds so-called critical slowing down z , (8.12.11)

where z is a dynamical critical exponent characterizing the eciency of a Monte Carlo algorithm. For the Metropolis algorithm one nds z 2, which leads to a very bad critical slowing down behavior. In lattice gauge theory the best algorithm that is presently known (the so-called overrelaxation algorithm) has z 1. For simpler spin models so-called cluster algorithms exist which have z 0, and which hence eliminate critical slowing down. These algorithms are extremely ecient, and allow one to perform very accurate numerical simulations in these models. Unfortunately, for lattice gauge theory no algorithm with z 0 is known. The inclusion of quarks poses additional very hard problems in numerical simulations of QCD. As a matter of fact, computers cannot deal with Grassmann numbers directly. Therefore, one integrates out the fermions analytically and obtains a fermion determinant, that is a very complicated function of the gluon link variables. One includes the determinant in the eective gluon action, which then becomes highly nonlocal. To evaluate the change in action, which is necessary for the Metropolis step, then is a very time-consuming procedure. Simulations with dynamical quarks are therefore orders of magnitude more complicated than simulations of the pure gluon theory. It is a matter of ongoing research to nd better ways to handle quarks numerically. Due to the tremendous diculties to simulate dynamical quarks, people often use the so-called quenched approximation, in which the fermion determinant is simply put to 1. It is unclear how this very drastic approximation aects the physics. Especially for light quarks one expects that the quenched approximation should be bad. On the other hand, for somewhat heavier quarks the approximation seems to work reasonably well.