1:
2: \section 2. Simulation parameters
3:
4: We consider the Wilson formulation of lattice QCD, optionally
5: O($a$)-improved, with a doublet of mass-degenerate sea quarks.
6: The notation and normalization conventions adopted in this paper
7: coincide with those already used in our previous paper [\ref{I}].
8: In particular, the parameters of the lattice theory are the inverse
9: bare coupling $\beta$, the sea-quark hopping parameter $\ksea$
10: and the coefficient $\csw$ of the Sheikoleslami-Wohlert improvement
11: term [\ref{SW},\ref{OaImp}].
12:
13: All simulations reported here
14: were performed using the DD-HMC
15: simulation algorithm [\ref{SchwarzIII}].
16: As suggested by the name,
17: the algorithm combines domain-decomposition ideas with
18: the HMC algorithm [\ref{HMC}].
19: More precisely, by dividing the lattice
20: into non-overlapping rectangular blocks,
21: a natural separation of the high-frequency
22: from the low-frequency modes of the fields is achieved.
23: Following
24: Sexton and Weingarten [\ref{SextonWeingarten}],
25: the different modes are then evolved
26: using different molecular-dynamics step sizes,
27: which results in a significant acceleration of the simulation.
28:
29: \input table1
30:
31: On a given lattice and at fixed coupling,
32: the simulations progressed
33: from the larger to the smaller quark masses,
34: normally skipping $1500$ molecular-dynamics
35: trajectories for thermalization.
36: The number $\Ntrj$ of trajectories generated after thermalization,
37: the separation $\Nsep$ (in numbers of trajectories) between
38: successive saved field configurations and the number $\Ncfg$ of
39: saved fields are given in table~1.
40: Different runs at the same lattice parameters
41: (such as $A_{3a}$ and $A_{3b}$) are distinguished by a lower-case
42: latin index.
43: In our previous paper [\ref{I}], only the runs
44: $A_{1a}$, $A_2$, $A_{3a}$, $A_{3b}$, $B_1$--$B_4$ and $D_1$--$D_5$
45: were included in the physics analysis.
46: The other runs listed in table~1
47: merely serve, in sections $3$ and $4$, to clarify
48: some technical issues.
49:
50: \input table2
51:
52: The DD-HMC simulation algorithm was implemented following the lines of
53: ref.~[\ref{SchwarzIII}]. In particular, for the solution of
54: the Dirac equation on the full lattice, the Schwarz-preconditioned
55: GCR solver described in ref.~[\ref{SchwarzII}] was used.
56: The so-called replay trick, however, was
57: switched off in the more recent simulations $A_{3b}$--$E_3$,
58: because trajectory replays would have been rare
59: and hardly worth the extra effort (see subsect.~3.3).
60:
61: No attempt was made to tune the DD-HMC algorithm
62: and most of its parameters were actually
63: set to some fixed values, the same as the ones already
64: chosen in ref.~[\ref{SchwarzIII}].
65: Among these were the trajectory
66: length $\tau=0.5$, the integration step numbers $N_0=4$ and $N_1=5$
67: associated to the gauge and block fermion forces as well as
68: the admitted tolerances
69: $(r_1,r_2,\tilde{r}_1,\tilde{r}_2)=(10^{-8},10^{-7},10^{-11},10^{-10})$
70: for the numerical solution of the Dirac equation on the
71: blocks and the full lattice%
72: \kern1.5pt\footnote{$\dagger$}{\footnotefont%
73: The trajectory length $\tau$ and thus the
74: integration step sizes $\tau/N_2$, etc., refer to
75: a particular normalization of the kinetic term
76: in the molecular-dynamics Hamiltonian.
77: Here the normalizations are the same as in
78: ref.~[\ref{SchwarzIII}], i.e.~the term is assumed to be equal to
79: $\frac{1}{2}(\Pi,\Pi)=\sum_{x,\mu}\tr\{\Pi(x,\mu)^{\dagger}\Pi(x,\mu)\}$,
80: where $\Pi(x,\mu)$ denotes the canonical momentum of the link variable
81: $U(x,\mu)$.
82: }.
83: The parameters of the Schwarz-preconditioned GCR solver
84: were fixed to the values quoted
85: in ref.~[\ref{SchwarzII}], except for
86: the number $n_{\rm kv}$ of Krylov vectors generated
87: before the GCR recursion is restarted, which was set to $32$
88: in run $D_5$ and to $24$ in all other runs.
89:
90: What remains to be specified are then
91: the size of the blocks on which the
92: algorithm operates and the integration step number $N_2$
93: associated to the block interaction term in the molecular-dynamics
94: Hamiltonian (see table~2). In practice the latter must be increased
95: as one moves to lighter quark masses in order to
96: preserve a high acceptance rate $P_{\rm acc}$.
97: The average number $\Ngcr$
98: of GCR solver iterations needed along the trajectories
99: also depends on $N_2$ (it decreases when $N_2$ goes up),
100: while the average number $\Ncg$ of conjugate-gradient iterations required
101: for the computation of the block terms in
102: the molecular-dynamics equations is largely
103: determined by the block size.
104:
105: With the chosen parameters, the reversibility of the molecular-dynamics
106: trajectories is guaranteed to high precision. In the tests
107: that we have performed, the average absolute deviation of the
108: components of the link variables after a return trajectory was
109: at most $3\times10^{-9}$, while in the case of the Hamiltonian
110: the observed differences were less than $4\times10^{-6}$. Deviations
111: larger than $10$ times the average occurred in less than $1\%$ of
112: the cases and never went beyond $100$ times the average.
113:
114: