In this se tion, we briey outlinethe idea behind aLatti e QCD simulation
(with any kindof fermions) and des ribe the most widely used algorithmof
generating gauge eld ongurations the Hybrid MonteCarlo algorithm.
3.1.1 General idea
AsalreadystatedinSe tion1.1, omputinganyobservableinaLatti eQCD
simulation onsistsinapproximatelyevaluatinganintegraloftheform(1.16)
by aMonteCarlomethod. Thisisa high-dimensionalintegraloverall
possi-blegaugeeldandfermioneld ongurations. Fortunately,the dependen e
ontheGrassmann-valuedfermionelds an alwaysbeeliminated,leavingan
integral overonlythe gaugeelds, weighted by the Boltzmannfa tor
e −S ef f
,where
S ef f
is some ee tive a tiondependent on the algorithmused.More-over, foramajority ofgauge eld ongurations the a tion isverylarge and
hen e theirweight isnegligiblysmall. Therefore, one shouldperform
impor-tan e sampling, i.e. use analgorithm that ee tively hooses ongurations
that have ahigh Boltzmannfa tor. Thus, having alarge number(of the
or-der of several thousand) of su h ongurations, one an ompute the Monte
Carlo average of anobservable
O
, whi h wewill denotebyO ¯
:O = ¯ 1
N
N
X
i=1
O[U i ],
(3.1)where
O[U i ]
denotes theobservableO
omputedinaba kgroundgaugeeldU i
belongingtothe Markov hain of generated ongurations. Ifthe simula-tion is performed orre tly, inthe limitN → ∞
the Monte Carlo averageO ¯
will orrespondtothea tualensembleaverage
hOi
. The onditionsthathavetobesatisedinordertoobtainthe orre taveragearemeasurepreservation
and detailed balan e. The latter reads:
e −S[U ] P (U → U ′ ) = e −S[U ′ ] P (U ′ → U),
(3.2)where
P (U → U ′ )
denotesthe probabilityof transitionfrom ongurationU
to
U ′
.Let us now onsider the partition fun tion (1.17). After integrating out
the fermion elds, one obtains:
Z =
where
det( ˆ D i [U])
isthedeterminantoftheDira operatormatrixforfermionavour
i
. Su hformoftheintegrandimpliesthattheprobabilitydistribution that hastobesimulateddepends onahighlynon-lo alfermiondeterminant.The ost of al ulating this determinant(s) is by far the highest ost in a
Monte Carlo simulation. However, the rst approximation to the partition
fun tion ouldbetonegle tthefermiondeterminant,i.e. setittoa onstant.
Su h approximation is alled the quen hed approximation and it physi ally
onsists in negle ting the fermion loops. As su h, it is very rude. Still,
for many years it was very mu h used in simulations, sin e the
omputa-tional ost relatedtothe determinantwasjusttoohigh forthe generationof
omputers then available.
However, the omputational power has been in reasing for many years
and presently itis possibletoperformfully dynami alsimulations(i.e. with
the determinant in luded 1
),whi h isalsodue tomany algorithmi
improve-ments. In the next subse tion we des ribe the algorithmof hoi e for most
simulations with dynami alfermions the HybridMonte Carlo algorithm.
3.1.2 Hybrid Monte Carlo
The HybridMonteCarlo(HMC)algorithmwasoriginallyintrodu edby
Du-ane,Kennedy,PendletonandRoweth[70℄. It ombinesamole ulardynami s
update of gaugeelds witha Metropolisa ept/reje t step. Here we outline
the basi steps that need tobe performed inan HMCsimulation[71℄.
Given the a tiontosimulate
S(U)
, rst one onstru ts the Hamiltonian:H(π, U) = 1
Thedeterminantis not omputedexpli itlyoneusuallyrepresentsitinanindire t
way,e.g. byasetofpseudofermionelds, tobedis ussedlater.
where
π x,µ a
is a omponent of a momentum eld:π x,µ = π x,µ a t a
(3.5)onjugate to ea h latti e link
U x,µ
. In this way, the integral one wants toevaluate:
R DU O[U] exp(−S(U))
an be written in the equivalent form:R DUDπ O[U] exp(−H(π, U))
,sin e the additionalintegrationover momen-tum eldsπ
yields just a Gaussian integral and hen e produ es a onstantfa tor.
Thus, one obtains a lassi al Hamiltoniansystem. The evolution of this
system in a titiousMonteCarlo time
τ
an be al ulated fromtheHamil-ton's equations:
˙π x,µ = −F x,µ ,
(3.6)U ˙ x,µ = π x,µ U x,µ ,
(3.7)wherethedotdenotesdierentiationwithrespe ttothe titioustime
τ
andthe for e
F x,µ
is given by2:F x,µ = ∂S(U)
∂U x,µ
.
(3.8)Solving the above system of dierential equations, one obtains a traje tory
in phase spa e,i.e. the values of
U x,µ (τ )
andπ x,µ (τ )
for every value ofτ
.The steps inthe HMCalgorithmare the following:
1. Randomly generate the initial(
τ = 0
) momentum eldπ x,µ (0)
a ord-ing tothe distribution
exp(− 1 2 P
x,µ π x,µ a π x,µ a )
.2. Numeri allyintegrate(e.g. bytheso- alledleap-frogalgorithm)
Hamil-ton's dierential equations (3.6)-(3.7) to obtain
U x,µ (τ )
andπ x,µ (τ )
fromtheir initialvalues
U x,µ (0)
andπ x,µ (0)
, respe tively. By onstru -tion, su h evolution preserves the value of the Hamiltonian up to anumeri al integration error.
3. A ept the new ongurationwith probability:
P =
min1, e −∆H(τ ) ,
(3.9)where
∆H(τ) = H(π(τ), U(τ)) − H(π(0), U(0))
is ingeneral non-zero,whi hisdueonlytothenumeri alintegrationerror. Ifthe onguration
is reje ted, then
U(τ ) = U(0)
, i.e. the initial onguration does not hange.2
In(3.8)weusesymboli notationforaderivativeofthea tionwithrespe ttoalink
variable. The derivative with respe t to an SU(3) element an be formally dened as
∂S(exp(ω a (x,µ)t a )U(x,µ))
∂ω a (x,µ)
ω a (x,µ)=0
.
tions)as isdesired. The initialvaluesof the variables
U
andπ
for step2oftraje tory
N
arethevaluesofthesevariablesattheendofstep3ofthe pre edingtraje tory
N − 1
, i.e.:U
traje toryN (0) = U
traje toryN −1 (τ )
,π
traje toryN (0) = π
traje toryN −1 (τ )
.The above algorithm ould in prin iple be used to simulate QCD with
dynami al quarks. However, to make su h simulations pra ti al, one has to
over ome the omputational problem of ee tively al ulating the fermion
determinant. This is usually done with the pseudo-fermion method, whi h
onsists in repla ing the fermion elds by auxiliary bosoni elds