* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download 8. Quantum field theory on the lattice
Noether's theorem wikipedia , lookup
Aharonov–Bohm effect wikipedia , lookup
Hidden variable theory wikipedia , lookup
Asymptotic safety in quantum gravity wikipedia , lookup
Quantum electrodynamics wikipedia , lookup
Quantum field theory wikipedia , lookup
Lattice Boltzmann methods wikipedia , lookup
Canonical quantization wikipedia , lookup
Feynman diagram wikipedia , lookup
Technicolor (physics) wikipedia , lookup
Ising model wikipedia , lookup
Scale invariance wikipedia , lookup
Higgs mechanism wikipedia , lookup
Path integral formulation wikipedia , lookup
Topological quantum field theory wikipedia , lookup
Gauge theory wikipedia , lookup
BRST quantization wikipedia , lookup
Gauge fixing wikipedia , lookup
History of quantum field theory wikipedia , lookup
Quantum chromodynamics wikipedia , lookup
Renormalization wikipedia , lookup
Renormalization group wikipedia , lookup
Introduction to gauge theory wikipedia , lookup
8. Quantum field theory on the lattice 8.1. Fundamentals • Quantum field theory (QFT) can be defined using Feynman’s path integral [R. Feynman, Rev. Mod. Phys. 20, 1948]: Z Y Z = Z S = x 4 dφ(x) exp[iS(φ)] (1) d xL(φ, ∂tφ) Here x = (x0, x1, x2, x3), and gµν = diag(+,-,-,-) • Physical observables can be evaluated if we add a source S → S + Jx φx . For example, 1- and connected 2-point functions are: ∂ 1Z hφx i = [dφ]φx exp[iS] log Z = i∂Jx J=0 Z ∂ ∂ 1Z hφx φy i = [dφ]φx φy exp[iS] − hφx i2 log Z = i∂Jx i∂Jy Z 1 • These can be computed using perturbation theory, if the coupling constants in S are small (see any of the oodles of QFT textbooks). • However: if – 4-volume is finite, and – 4-coordinate x is discrete (x ∈ aZ 4 , a lattice spacing), the integral in (1) has finite dimensionality and can, in principle, be evaluated by brute force ( = computers) 7→ gives fully non-perturbative results. • Need to extrapolate: V → ∞, a → 0 in order to recover continuum physics. 2 • But: - the dimensionality of the integral is (typically) huge - exp iS is complex+unimodular: every configuration {φ} contributes with equal magnitude 7→ extremely slow convergence in numerical computations (useless in practice). • This is (largely) solved by using imaginary time formalism (Euclidean spacetime). • Imaginary time also admits non-zero temperature T . 3 Units: Standard HEP units, where c = h̄ = kB = 1, and thus [length]−1 = [time]−1 = [mass] = [temperature] = [energy] = GeV. Mass m = rest energy mc2 = (Compton wavelength)−1 mc/h̄. (1 GeV)−1 ≈ 0.2 fm 4 8.2. Path integral in imaginary time Consider scalar field theory in Minkowski spacetime: S= Z 3 d x dt L(φ, ∂tφ) = Z 3 d x dt h µ 1 2 ∂µ φ ∂ φ − V (φ) i We obtain the (classical) Hamiltonian by Legendre transformation: H = = Z Z d3 x dt π φ̇ − L d3 x dt h h 1 2 2π i + 12 (∂iφ)2 + V (φ) i Here π is canonical momentum for φ: π = δL/δ φ̇. Quantum field theory: consider Hilbert space of states |φi, |πi, and field operators (φ → φ̂, π → π̂, H → Ĥ), with the usual properties: • φ̂(x̄)|φi = φ(x̄)|φi • Z Y • • [ x Z Y dφ(x̄)]|φihφ| = 1 [ x dπ(x̄)/(2π)]|πihπ| = 1 [φ̂(x̄), φ̂(x̄0)] = −iδ 3 (x̄ − x̄0 ) hφ|πi = exp [i Z d3 xπ(x̄)φ(x̄)] 5 Time evolution operator is U (t) = e−iĤt Feynman showed that the quantum theory defined by Ĥ and the Hilbert space is equivalent to the path integral (1). We shall now do this in imaginary time. Let us now consider the quantum system in imaginary time: t → τ = it , exp[−iĤt] → exp[−τ Ĥ] This makes the spacetime Euclidean: g = (+ − − −) → (+ + + +). We shall keep the Hamiltonian Ĥ as-is, and also the Hilbert space. Let us also discretize space (convenient for us but not necessary at this stage), and use finite volume: Z x̄ = an̄, n i = 0 . . . Ns 3X dx → a 3 x ∂i φ → 1 [φ(x̄ + ēi a) − φ(x̄)] ≡ ∆iφ(x̄) a 6 Thus, the Hamiltonian is: Ĥ = 2 3X 1 a 2 [π̂(x)] x + 2 1 2 [∆iφ̂(x̄)] + V [φ̂(x̄)] Let us now consider the amplitude hφB |e−(τB −τA )Ĥ |φA i: 1) divide time in small intervals: τB − τA = Nτ aτ . Obviously hφB |e−(τB −τA )Ĥ |φA i = hφB |(e−aτ Ĥ )Nτ |φA i. 2) insert 1-operators Z Y [ x dφi]|φi ihφi |, Z Y [ x dπi ]|πi ihπi | hφB |e−τ Ĥ |φA i = Z N Yτ Y i=2 x dφi(x) N Yτ Y i=1 x dπi (x) hφB |πNτ ihπNτ |e−aτ Ĥ |φNτ i ×hφNτ |πNτ −1ihπNτ −1 |e−aτ Ĥ |φNτ −1i . . . × hφ2 |π1 ihπ1 |e−aτ Ĥ |φA i 7 Thus, we need to calculate hπi |e −aτ Ĥ |φi i = × " 3 X 1 2 exp −a aτ 2 πi x hπi |φi i + O(a2τ ) + 2 1 2 (∆i φi ) + V [φi ] # and ia3 P πi (x̄)φi+1 (x̄) −ia3 P x πi (x̄)φi (x̄) e hφi+1|πi ihπi |φi i = e X = exp[ia3aτ πi (x̄)∆0φi (x̄)] x x where we have defined ∆0φi = We can now integrate over πi (x̄): Z Y [ = x " 1 (φi+1 − φi ). aτ πi (x̄)] exp[a3 aτ 2π a3 aτ #N 3 /2 S X x − 21 πi2 + (i∆0φi )πi] × exp[−a3 aτ 8 X1 2 2 (∆0 φi (x̄)) ] x (2) Repeating this for i = 1 . . . Nτ , and identifying the time coordinate x0 = τ = aτ i, we finally obtain the path integral hφB |e −τ Ĥ Z Y |φA i ∝ [ x dφ]e−SE where SE is the Euclidean (imaginary time) action X 1 ( 2 (∆µφ)2 x Z 4 d x( 21 (∂µφ)2 + SE = a 3 aτ → + V [φ]) V [φ]) and we have the boundary conditions φ(τA ) = φA , φ(τB ) = φB . The path integral is precisely in the form of a partition function for a 4-dimensional classical statistical system, with the identification SE ↔ H/(kB T ). For convenience, we make the system periodic in time by identifying φA = φB and integrating over φA . 9 We can now make a connection between the correlation functions of the “statistical” theory and the Green’s functions of the quantum field theory. First, note that we can interpret Tφi+1 ,φi = hφi+1 |e−aτ Ĥ |φi i as a transfer matrix. In terms of T the partition function is Z Z = [dφ]e−SE = Tr (T Nτ ) Let us label the eigenvalues of T by λ0 , λ1 . . ., so that λ0 > λ1 ≥ . . .. Note that λi = exp −Ei , where Ei are eigenvalues of Ĥ. Thus, λ0 corresponds to the state of lowest energy, vacuum |0i. If we now take Nτ to be very large (while keeping aτ constant; i.e. take ∆τ to be large), Z= X i Nτ Nτ τ λN i = λ0 [1 + O((λ1 /λ0 ) )] For example, a 2-point function can be written as (let i − j > 0) 1Z 1 [dφ]φiφj e−SE = Tr (T Nτ −i+j φ̂T i−j φ̂). hφi φj i = Z Z Taking now Nτ → ∞, and recalling aτ (i − j) = τi − τj , hφi φj i = h0|φ̂(T /λ0)i−j φ̂|0i = h0|φ̂(τi)φ̂(τj )|0i, 10 where we have introduced time-dependent operators φ̂(τ ) = eτ Ĥ φ̂e−τ Ĥ . Allowing for both positive and negative time separations of τi and τj , we can identify hφ(τi )φ(τj )i = h0|T [φ̂(τi )φ̂(τj )]|0i, where T is the time ordering operator. Mass spectrum: • Green’s functions in time τ : 1Z h0|Γ(τ )Γ (0)|0i = [dφ]Γ(τ )Γ†(0)e−S Z = h0|eĤτ Γ(0)e−Ĥτ Γ† (0)|0i † = h0|Γ(0) 11 X n |En ihEn |e−Ĥτ Γ† (0)|0i = X −E τ n e n −E0 τ → e |h0|Γ(0)|Eni|2 |h0|Γ(0)|E0i|2 as τ → ∞ where |E0 i is the lowest energy state with non-zero matrix element h0|Γ(0)|E0i. → measure masses (E0) from the exponential fall-off of correlation functions. 12 8.3. Finite temperature Connection Euclidean QFT ↔ classical statistical mechanics was derived for zero-temperature quantum system. However, this can be readily generalized to finite temperature: Quantum thermodynamics w. the Gibbs ensemble: Z Z = Tr e−Ĥ/T = [dφ] hφ|e−Ĥ/T |φi Expression is of the same form as the one which gave us the Euclidean path integral for T = 0 theory! The difference here is 1) Finite + fixed “imaginary time” interval 1/T 2) Periodic boundary condition: φ(1/T ) = φ(0). Repeating the previous derivation, the partition function becomes Z Z(T ) = [dφ] e −SE Z = [dφ] exp[− Thus, a connection between: – Quantum statistics in 3d: Z = Tr e−Ĥ/T 13 Z 1/T 0 dτ Z d3xLE ] – Classical statistics in 4d: Z = [dφ]e−S Euclidean P.I. is a very common tool for finite T field theory analysis [J. Kapusta, Finite Temperature Field Theory, Cambridge University Press] R 14 8.4. Some terminology: In numerical work, lattice is a finite box with finite lattice spacing a. In order to obtain continuum results, we should take 2 limits: • V →∞ • a→0 thermodynamic limit continuum limit Both have to be controlled – expensive! 1) Perform simulations with fixed a, various V . Extrapolate V → ∞. 2) Repeat 1) using different a’s. Extrapolate a → 0. 3) [ In QCD, one often has to extrapolate mq → mq,phys. . ] 15 T = 0: 1) V → ∞: Nτ , Ns → ∞, a constant. 2) continuum: a → 0. T > 0: 1) V → ∞: Ns → ∞, Nτ , a constant. 2) continuum: a → 0, 1 T = aNτ constant. 16 8.5. Scalar field Free scalar field on a finite d-dimensional lattice with periodic boundary conditions: xµ = anµ , nµ ∈ Z Action: S= X d 1X a 2 x µ h 1 2 2 1 2 d 1 (φ − φ ) + m φ φ = a x+µ x 2 x a2 2 x,y φy + 12 m2 φ2x (implicit sum over x, y), and the lattice d’Alembert operator is x,y φy = −∆2φ = X µ 2φx − φx+µ̂ − φx−µ̂ a2 The action is of form S = 12 φx Mx,y φy , and Z Z = [dφ]e−S = (Det M/2π)−1/2 17 i Fourier transforms: X f˜(k) = ad e−ikx f (x) x Since x = an, f˜(k + 2πn) = f˜(k), and we restrict k to Brillouin zone: −π/2 < kµ ≤ π/2 Inverse transform: Z π/a dd k f (x) = eikx f˜(k) −π/a (2π)d Note: often it is convenient to use dimensionless natural lattice units xµ ∈ Z, −π < kµ ≤ π. 18 Lattice propagator: The lattice propagator G(x, y) is defined to be the inverse of operator a−d M = ( + m2 ): X d a ( y x,y + m2 δx,y )G(y, z) = δx,z Take Fourier transform (G(x, y) = G(x − y)): X µ 2 (1 − cos kµ a) + m2 G̃(k) = 1 2 a which gives the lattice propagator G̃(k) = 1 , k̂ 2 + m2 where k̂µ = 4 2 kµ a . sin a2 2 Continuum limit: when a → 0, G̃(k) = 1/(k 2 + m2 ) + O(a2 ). 19 Generating function for Green’s functions: S → S(J) = Now X d h1 a 2 φx ( x Z + m2 )φx − Jx φx Z(J) = [dφ]e−S(J) = Z(0) exp[ X 2d 1 a 2 Jx G(x, y)Jy ] x,y N-point functions hφx . . . φy i = Z(0)−1 δ δ ... Z(J) δJx δJy J=0 Interactions: just modify (for example) L = Lfree + 20 i 1 4 λφ 4! 8.6. Gauge fields • Gauge field lagrangian in (Euclidean) continuum: 1 a F Fa 4 µν µν = 12 Tr (Fµν Fµν ) • Field tensor Fµν = [Dµ , Dν ] = ∂µ Aν − ∂ν Aµ + ig[Aµ, Aν ] where Aµ = Aaµ λa , and λa are the group generators. We shall consider only unitary groups U(1) [QED] and SU(N) [QCD, Electroweak]. For SU(N), the generators are orthonormalized Tr λa λb = 12 δ ab . • Put gauge potential Aµ directly on a lattice? Difficult to maintain gauge invariance. 21 • Good starting point for gauge field on a lattice is to consider consider gauge fields acting in a parallel transport: As field φ ∈ SU(N) is “parallel transported” along a path p, parametrized by xµ (s), s ∈ [0, 1], gauge fields rotate it: φ → U (s)φ Uφ • Define U (s) differentially via dU (s) dxµ = igAµ U (s), ds ds which can be formally solved: dxµ φ U (s) = P exp ig ds Aµ p ds • Here P = “path ordering”: in the power series expansion of the exponential one always takes A(x) in the order they are encountered along the path. • For U(1) (or any other Abelian group) path ordering is irrelevant. " 22 Z # • Alternatively, if we divide the path into N finite intervals of length ∆s, and xn = x(n∆s), n = 0 . . . N − 1: U (s) ≈ P exp ig X n Y dxnµ dxnµ n Aµ (x ) = exp ig∆s Aµ (xn) ∆s ds ds i Gauge transformations: • Gauge transformation Λ(x) is a (SU(N)) group element defined at every point. Gauge potential transforms as i Aµ → ΛAµ Λ−1 + Λ∂µ Λ−1 g • Field φ: φ(x) → Λ(x)φ(x) • Path p: U (p) → Λ(x1)U (p)Λ−1(x0), where x0 and x1 are the beginning and end of path. 23 • Closed loops: U (C) → Λ(x0)U (C)Λ−1(x0) • Trace of a closed loop: Tr U (C) is gauge invariant! 24 8.7. Lattice gauge fields • Variables: parallel transporters from one lattice site to a neighbouring one, Links: Uµ(x) = P exp ig Z x+µ x dxµAµ = exp igaAµ (x + 21 µ) + O(ga3 ) h i Uµ (x+ ν) Uν (x+ µ) Uν(x) • The lattice action has to be gauge invariant. Only traces of closed loops are gauge invariant; the simplest one is plaquette Uµ(x) Tr U = Tr Uµ (x)Uν (x + µ)Uµ† (x + ν)Uν†(x) • When a small (for SU(N)), Re Tr U = N − a4 g 2 a a Fµν Fµν + O(g 4 a6 ) 4 25 • → Wilson gauge action (K. Wilson, 1974): Sg Partition function 2N X 1 = 2 4−d 1 − Re Tr U g a N Z = dd x 12 Tr [Fµν Fµν ] + O(a2 g 2 ) " Z Y Z= [ x,µ # dUµ(x)]e−Sg Here the integral is over group elements Uµ (compact), not over algebra Aµ Common notation: β = βG ≡ 2N g2 for SU(N), 26 β= 1 g2 for U(1). 8.7.1. Continuum limit for U(1) (in 4d) • Uµ(x) = exp igaAµ (x) • The Wilson action for U(1) is 1 X † † 1 − Re U (x)U (x + µ)U (x + ν)U u(x) µ ν µ n g 2 x;µ<ν 1 X (1 − Re exp [iga(Aµ (x) + Aν (x + µ) − Aµ (x + ν) − Aν (x))]) = 2 g x;µ<ν S = Expanding Aν (x + µ) = Aν (x) + a∂µ Aν (x) + 21 a2 ∂µ2 Aν (x) + . . . h i 1 X 2 4 1 − Re exp iga (∂ A − ∂ A ) + O(a ) µ ν ν µ g 2 x;µ<ν " # 11 X 1 2 2 2 2 6 = 2 g a Fµν + O(g a ) g 2 x;µ6=ν 2 1Z 4 2 d xFµν = 4 S = 27 • Error in lagrangian O(a2 ) • Non-Abelian continuum limit comes in a similar way, but now one has to be careful with the commutators. Check it! (Campbell-BakerHausdorff formula eA eB = eA+B−[A,B]+... might help, but it also comes directly from the expansion of the exponents.) 28 8.8. Updating the gauge fields: Let us consider U(1) theory, for simplicity. The action is S = β X (1 − Re U (µ, ν; x)) x;µ<ν U (µ, ν; x) = Uµ (x)Uν (x + µ)Uµ† (x + ν)Uν†(x), where Uµ (x) = exp[iθµ(x)], 0 ≤ θµ(x) < 2π. To update Uµ(x) we need the part of the action which depends on it, i.e. the plaquettes which contain the link Uµ(x). This is Sµ (x) ≡ −βRe Uµ (x)Pµ(x) X Pµ (x) = Uν (x + µ)Uµ† (x + ν)Un†u(x) ν;|ν|6=µ where Pµ is called the sum of staples, for obvious reasons. 29 The sum over ν above goes over both positive and negative directions; here U−ν (x) = Uν† (x − ν). Now the link variable Uµ(x) can be updated with a suitable algorithm (heat bath, Metropolis, overrelaxation . . . ). For U(1), one staple is exp[i(θν (x + µ) − θµ (x + ν) − θn u(x))]; thus, the 0 sum of staples is a complex number: Pµ (x) = Ae−iθ , and the ’local action’ becomes Sµ (x) = −βRe Uµ (x)Pµ(x) = −βA cos(θµ(x) − θ0 ) Thus, the local action is just like with the XY model discussed earlier! We can apply the same update algorithms as discussed there. 30 In the above discussion we used the standard compact formalism for U(1) gauge fields: Uµ = eigaAµ ; S as given above. For U(1) (but not for general non-abelian gauge field) it is also possible to use non-compact formulation: let the link variable be αµ (x) = gaAµ (x), − ∞ < αµ (x) < ∞, and the action is now S NC = 1 X 1 [αµ (x) + αν (x + µ) − αµ (x + ν) − αν (x)]2 2 g x;µ<ν 2 The continuum limit for this action is the same as for the compact action (check it!). The difference in the continuum limit is in the higher order O(a2 ) terms. At finite a the physics can be quite different! Note: switch 21 [·]2 7→ (1 − cos[·]), and you recover the compact action. Non-compact U(1) gauge field theory is actually completely solvable (it is free theory, as the continuum theory). However, if we would couple the theory to matter this would not be the case. 31 8.9. Gauge transformations: • Gauge transform Λ(x) ∈ SU(N) lives on lattice sites • Link variable: Uµ (x) → Uµ0 (x) = Λ(x)Uµ(x)Λ†(x + µ) 7→ Aµ(x) → Λ(x)Aµ(x)Λ†(x) + gi Λ(x)∂muΛ†(x) • Trace of a closed loop is gauge invariant • Fundamental matter: φ(x) → Λ(x)φ(x), where φ is a N-component complex vector: operator φ† (x)UP (x 7→ y)φ(y) = φ† (x)Uµ(x)Uν (x + µ) . . . Uρ (y − ρ)φ(y) is gauge invariant. • We want only gauge invariant animals on a lattice: lattice action + 32 observables must consist of closed loops (gauge only quantities) or φ† UP φ - “strings” (matter). • For example, the matter field kinetic term (Dµ φ)† (Dµ φ), where Dµ = ∂µ + igAµ , becomes X † 1 † 2dφ φ − 2 φ (x)Uµ(x)φ(x + µ) a2 µ • Elitzur’s theorem: expectation values of gauge non-invariant object = 0: hUµ (x)i = 0. 33 8.9.1. Gauge fixing on the lattice • Something you don’t want to do • Necessary for perturbative calculations • Most of the (Euclidean) continuum gauges go over to lattice • Special gauge: maximal tree 34 • A tree which connects every point of the lattice, but does not have closed loops. • Link matrices Uµ (x) are fixed to arbitrary values (for example all = 0) in the tree • Expectation values of gauge invariant quantities do not change (proof: easy) • (Almost) complete gauge fixing 35 8.10. Confinement and Wilson loop • Potential between static charge and anticharge, separated by R: V (R → ∞) → ∞: finite: confinement no confinement • Standard probe: Wilson loop WRT . Let W be a rectangular path of size R × T along x1 (say) and x0 = τ directions. WRT = Tr P exp ig I W dsµ Aµ = Tr Y Uxy . <xy>∈W Now − log WRT gives the “free energy” of a static charge-anticharge (“quark-antiquark”) system separated by R and which evolves for time T : − loghWRT i = V (R)T valid as T → ∞ 36 T R • Perimeter law: − loghWRT i ∼ m(2R + 2T ) Free charges, m: “mass” of the charge due to gauge field • Area law: − loghWRT i ∼ σRT ; V (R) = σR Charges confined with linear potential, σ: string tension • In general, 1 T loghW i ∼ σR + const. + c/R + . . . 37 Motivation: • In temporal gauge A0 = 0 → U0 = 1 we can define gauge field Hamiltonian Ĥ [Kogut-Susskind] • Ψ: (trial) wave function of q q̄ pair at x̄, ȳ; spatial gauge transformations Ψab → Λac (x̄)Λ†bd (ȳ)Ψcd • Ĥ gauge invariant: hΨ0 |e−T Ĥ |Ψi = 0 unless Ψ, Ψ0 have similar gauge transformation properties. hΨ|e−T Ĥ |Ψi = X n |hΨn |Ψi|2 e−En T → |hΨ0 |Ψi|2 e−E0 T when T → ∞ • E0 = V (R) ground state energy of static charges (R = |x̄ − ȳ|) • Trial wave function: Ψab = Uab(x 7→ y)Ψvac , where Ψvac is the (gauge 38 invariant) vacuum wave function hΨ|e −T Ĥ 1Z |Ψi = [dU ] Tr [U †(T ; x̄ 7→ ȳ) U (0; x̄ 7→ ȳ)]e−S Z = hWRT i • WRT gauge invariant: can be measured in any gauge: − 1 loghWRT i → V (R) T as T → ∞. • If we can guess/construct Ψ ≈ Ψ0, T need not be very large. 39 8.10 12 . Measuring the string tension from Monte Carlo The Wilson loops become exponentially small (∼ e−σRT ) as the size increases. However, the statistical noise is ∼ constant in magnitude independent of loop size! Thus, we need to increase the number of measurements exponentially as the loop size increases. R Smearing improves the situation: instead of using only straight const-T legs in the loop, take an averaged sum over paths around the straight one. The smearing is to mimic the wave function of the desired q q̄ ground state. The hope is that now a smaller T -extent would be sufficient to get the asymptotic behaviour, i.e. look like the limit T → ∞. 40 T Smearing can be done recursively: let i ⊥ t, and Ui(x) ← Ui(x) + c X j⊥i,t Uj (x)Ui(x + ĵ)Uj†(x + î) = + i.e. substitute link matrices on a plane ⊥ to t by the weighted average of the link and the staples (which are also ⊥ to t). This can be repeated a few times. 41 Wilson loop in pure gauge SU(3) [Bali,Schilling, Wachter 1995] 483 × 64 quenched lattice, β = 6.8 1 V0(r)/GeV 0.5 0 -0.5 -1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 r/fm 42 1 1.1 For large loops hW i ∼ e−V (r)T Phenomenological form e 1 V (r) = V0 + σr − + f GL (~r) − r r " 43 # 8.11. Strong coupling expansion and the Wilson loop • Perturbation theory (weak coupling, g 1) is toothless: no confinement • Strong coupling: g 1. For simplicity, consider U(1) (β = 1/g 2 1). The link variable is Uµ = exp iθµ. S = β Z = X Z 2π 0 (1 − Re eiθ ) dθµ x,µ 2π Y Y exp[β cos θ ] Here θ = θ1 + θ2 − θ3 − θ4 around the plaquette . • Expand in powers of β? OK, but more convenient is the following character expansion (here In are modified Bessel functions) eβ cos θ = ∞ X n=−∞ In (β)einθ = A 1 + 44 ∞ X n=1 fn cos nθ where 1 1 A = I0(β) = 1 + β 2 + β 4 + . . . 4 64 1 fn = 2In(β)/I0(β) = n−1 β n + O(β n+2) 2 n! 1 3 1 5 f1 = β − β + β + . . . 8 48 Compare to the “character expansion” in the Ising model: eβsi sj = a(1 + bsi sj ) with a = cosh β, b = tanh β. In Ising gauge theory, (i, j; x) = si (x)sj (x + î)si(x + ĵ)sj (x) = ±1, and the expansion of eβ (i,j;x) is exactly analogous. 45 The partition function becomes Z Z= [ dθ N ]A 2π Y (1 + f1 cos θ + f2 cos 2θ + . . .) here N = 6V is the number of plaquettes (in 4d). • Integration: dθa cos nθ = 0, if θa ∈ 2 adjacent plaquettes: R Z . dφ cos(θ + θa) cos(−θ + θb) = 21 cos(θa + θb) 2π θ (also for cos nθ) a θ θb • Non-zero contributions only from closed surfaces: lowest non-trivial comes from 3-d cube: Z = AN [1 + 4V ( 21 f1)6 + O(β 10 )] 46 • Each plaquette on the surface contributes 12 fi • Expansion of free energy F = − log Z contains only connected graphs – cluster expansion 47 Wilson loop: To leading order, the Wilson loop is hWRT i = AN Z Z [ dθ Y iθa Y ] e [1 + 12 f1(eiθ + e−iθ ) + O(f2)] 2π a∈W • First contribution comes when we “tile” the loop area with plaquettes: hWRT i ≈ ( 21 f1)RT = ( 12 β)RT + O(β RT +2 T ) • We see confinement: − loghW i/T = log( 12 f1)R = σR • Next order: “bump” hWRT i ≈ ( 21 f1 )RT [1 + 4RT ( 21 f1)4] 48 R which gives −a2 σ = − 1 loghWRT i = log 12 f1 + 4( 21 f1 )4 + O([ 21 f1 ]6) RT • Order [ 12 f1]6 gets contributions from the disconnected surfaces and Z. 10 936 10 1 532 044 12 3 596 102 14 8 • −a2 σ = log u + 4u4 + 176 3 u + 405 u + 1 215 u + 5 103 u , where u = ( 21 f1) (see Montvay+Münster, for example) • For U(1) this is strong coupling artefact! When g 1, there ∃ a phase transition. As g → 0 we regain free theory. 49 Non-Abelian gauge fields • Same, but more difficult: use character expansion e1/N β Re Tr U = Aβ [1 + X bR (β) χR (U )] R which gives “orthogonal” integration rules. • Graphs are again surfaces – with complications! • SU(2): −a2 σ = log u + 4u4 + 176 8 3 u + . . ., with u = I2 (β)/I1(β) • SU(3): −a2 σ = log u + 4u4 + 12u5 − 10u6 − . . . with u = 31 (x − 21 x2 + 23 x3 + . . .), x = β/6. 50 SU(2) strong coupling expansion 3 • Convergence is good in finite range β < βmax . • Roughening transition for Wilson loops as β increases? 2 σa • Mass gap: plaquetteplaquette correlation function decays exponentially. Can be calculated using strong coupling, not in perturbation theory. 2 1 log u 0 u u u 12 u 14 u 0 1 2 β 51 8 10 3 4 8.12. Realistic continuum limit • When we change lattice spacing a, we have to adjust the bare coupling g in order to keep physics (at “long” distances) invariant. • Measuring g(a): choose some physical/renormalized quantity M (a, g(a)), (mass, string tension, Wilson loop of fixed size, vertex function . . . ) which can be measured/calculated. This we require to be constant as a is changed (“renormalization prescription”) → RG equation: d ∂ ∂ M a M (a, g(a)) = 0 = a − βL da ∂a ∂g ! ∂g ∂a tells how coupling g and lattice spacing a evolve. • β-function: βL ≡ −a • Fixed point βL (gF ) = 0. 52 • If βL (g) < 0 (> 0), g decreases (increases) as a is decreased. • If ∂βL(gF ) < 0, this is UV attractive fixed point and lima→0 g(a) = gF : continuum limit • For example: β-function at strong coupling, using expansion for SU(3) string tension σ = −a−2 log(1/(3g 2)) + O(g −2 ) d 2 σ = −2σ − βL 2 + . . . da ag 2 → βL = −g log(3g ) + . . . 0 = a This is < 0, so g & as a &: no continuum limit at strong coupling. 53 Weak coupling: • Use perturbation theory. For QCD (for example, from renormalized 3-vertex): βL (g) = −b0 g 3 − b1g 5 + . . . Coefficients b0 and b1 are universal: independent of the regularization and prescription. b0 b1 11N 2NF 1 − = 16π 2 3 3 2 2 1 10N N N (N − 1) 34N F F = − − 2 2 (16π ) 3 3 N ! Z g dg 0 βL−1(g 0 )] → 1 2 b1 /2b20 ](b g ) [1 + O(g 2 )] aΛL = exp[ 0 2 2b0g • Integrate: a = exp[− • UV fixed point g → 0 as a → 0: asymptotic freedom. (If b0 > 0!) 54 • ΛL: dimensionful integration constant, Λ-parameter; sets the scale. It is regularization dependent. • Dimensional transmutation: no mass scale in the original (classical) Lagrangian! • Thus, a physical mass m should behave as m 1 2 b1 /2b20 am = exp[ ](b g ) 0 ΛL 2b0g 2 In practice, this not well satisfied in QCD; g ∼ 1: Asymptotic scaling does not hold. • Scaling holds if mass ratios are ∼ constant: m1 m1 (a) = (0) + O(a2 ) m2 m2 This has been observed to be much better satisfied than asymptotic scaling. 55 Numerically: • Principle: measure a quantity – am, a2 σ – at several g’s. Then (am)(g1) a(g1) = (am)(g2) a(g2) gives a discrete sample of a(g1) − a(g2 ) = exp[− Z g 1 g2 dg 0 1/βL(g)] • Matching more than one quantity simultaneously: need more than one lattice coupling. • Monte Carlo renormalization group (MCRG): direct implementation of Wilson’s RG in numerical simulations. - Write the action S = a κa Sa , where Sa are a set of (local) contributions to action: plaquettes, larger loops . . . P d - With couplings κA a simulate on L lattice, and measure a set of Wilson loops. 56 - block the configurations by factor of 2: obtain (L/2)d lattice. Measure the Wilson loops. - Repeat blocking + measuring, until lattice is too small. - simulate directly on (L/2)d lattice, with κB . As before, repeat measuring Wilson loops and blocking. - compare Wilson loops at blocking level b from simulation B to loops at level b + 1 from A. Tune κB a , until the measurements match (requiring a new simulation). - repeat for (L/4)d etc. For the plaquette term, this gives function ∆β = β(a) − β(2a) = 6/g 2 − 6/g 02 • Violation of asymptotic scalings large at practical values of β (fig.)! 57 • Perfect action: no scaling violations; evolves along RG trajectory. Requires infinitely many coupling constants. • Classically perfect action: perfect action for classical configurations. Much easier to derive, and these have been used for various models. It’s also 1-loop quantum perfect action (perfect in the limit g → 0). In 3 dimensions: • Lattice coupling β = 2N . Dimensions of [g 2] = GeV. 2 g a • Renormalization: gR2 = g 2 + O(g 4 a) dimensionally • Continuum limit β = 2N + O[(gR2 a)0] 2 gR a 58 • Theory superrenormalizable: only finite number of divergent diagrams 59 8.13. Continuum limit in λφ4 model • Lattice lagrangian L = −κ X µ ϕx ϕx+µ + ϕ2x + λ(ϕ2x − 1)2 • In 4 dimensions, the RG trajectories of the bare couplings κ, λ have been solved [Lüscher, Weisz, NPB 295 (1988)] using hopping parameter expansion and perturbative RG equations: 60 κ 4d Ising model κc 1/8 free theory 0 oo λ Blue curves: RG trajectories of constant λR ; λR decreases to -; along κc we have λR = 0. ∂λ 3 2 = λ >0 ∂a 16π 2 • No continuum limit! RG trajectories hit the Ising limit λ = ∞ before the critical line κc (λ) is reached; i.e. at finite a. • Perturbatively βL (λ) = −a 61 • Ising model: smallest possible a for any fixed λR . • This is a generic feature of λφ4 models in 4d → Triviality bound for the Standard Model Higgs mass! • In 3 dimensions: there ∃ continuum limit: coupling constant is dimensionful [λ] = GeV. • λR = λ + O(λ2 a) • m2R = m20 + C1 λa−1 + C2λ2 + O(λ3 a) m2R has linear (1/a) and (log a/Λ) divergence. Perturbatively calculable, and can be subtracted. 62