Professional Documents
Culture Documents
Critical Behavior of The Three-Dimensional XY Universality Class
Critical Behavior of The Three-Dimensional XY Universality Class
⫽Ct ⫺ ␥ 共 1⫹a 0,1t⫹a 0,2t 2 ⫹•••⫹a 1,1t ⌬ ⫹a 1,2t 2⌬ ⫹••• the improved HT 共IHT兲 series allows us to substantially in-
crease the precision and the reliability of the results, espe-
⫹a 2,1t ⌬ 2 ⫹••• 兲 . 共4兲 cially of the critical exponents. As we shall see, in order to
determine the critical exponents, the extrapolation to  c of
The leading exponent ␥ and the correction-to-scaling expo- the IHT series, using biased integral approximants, is more
nents ⌬,⌬ 2 , . . . , are universal, while the amplitudes C and effective than the extrapolation L→⬁ of the FSS MC data.
a i, j are nonuniversal. For three-dimensional XY systems, the Moreover, we consider two improved Hamiltonians. The
value of the leading correction-to-scaling exponent is ⌬ comparison of the results from these two models provides a
⬇0.53,6,7 and the value of the subleading exponent is ⌬ 2 check of the errors we quote. The estimates obtained for the
⬇2⌬.9 two models are in very good agreement, providing support
The leading nonanalytic correction t ⌬ is the dominant for our error estimates and thus confirming our claim that the
source of systematic errors in MC and HT studies. Indeed, in systematic error due to confluent singularities is largely re-
MC simulations the presence of this slowly-decreasing term duced when analyzing IHT expansions.
requires careful extrapolations, increasing the errors in the We consider a simple cubic 共sc兲 lattice, two-component
final estimates. In HT studies, nonanalytic corrections intro- vector fields ជ x ⫽( (1)
x , x ), and two classes of models
(2)
duce large and dangerously undetectable systematic devia- depending on an irrelevant parameter: the 4 lattice model
tions in the results of the analyses. Integral approximants10 and the dynamically diluted XY 共dd-XY ) model.
共see, e.g., Ref. 11 for a review兲 can in principle cope with an The Hamiltonian of the 4 lattice model is given by
asymptotic behavior of the form 共4兲; however, in practice,
兺 ជ y⫹ 兺 关
they are not very effective when applied to the series of ជ x• ជ 2x ⫹ 共
ជ 2x ⫺1 兲 2 兴 .
moderate length available today. Analyses meant to effec- H 4 ⫽⫺  共5兲
具 xy 典 x
tively allow for the leading confluent corrections are based
on biased approximants, where the value of  c and the first The dd-XY model is defined by the Hamiltonian
nonanalytic exponent ⌬ are introduced as external inputs
共see, e.g., Refs. 12–17兲. Nonetheless, their precision is still
not comparable to that of the experimental result 共2兲, see,
Hdd⫽⫺  兺
具 xy 典
ជ y ⫺D 兺
ជ x•
x
ជ 2x , 共6兲
e.g., Ref. 14. The use of improved Hamiltonians, i.e., models
by the local measure
for which the leading correction to scaling vanishes 关 a 1,1
⫽0 in Eq. 共4兲兴,18 can lead to an additional improvement of
the precision, even without a substantial extension of the HT
series.
d 共 x 兲 ⫽d (1)
x dx
(2)
冋
␦ 共 (1)
x 兲 ␦共 x 兲⫹
(2)
1
2 册
␦ 共 1⫺ 兩 ជ x 兩 兲 ,
214503-2
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
TABLE I. Estimates of the critical exponents. See text for the explanation of the symbols in the second column. We indicate with an
asterisk ( * ) the estimates that have been obtained using the hyperscaling relation 2⫺ ␣ ⫽3 or the scaling relation ␥ ⫽(2⫺ ) .
Reference Method ␥ ␣
This work MC⫹IHT 1.3177共5兲 0.67155共27兲 0.0380共4兲 ⫺0.0146(8) *
This work MC 1.3177共10兲 * 0.6716共5兲 0.0380共5兲 ⫺0.0148(15) *
Reference 5 共2000兲 IHT 1.3179共11兲 0.67166共55兲 0.0381共3兲 ⫺0.0150(17) *
Reference 31 共1999兲 HT ⫺0.014(9), ⫺0.022(6)
Reference 14 共1997兲 HT, sc 1.325共3兲 0.675共2兲 0.037共7兲 * ⫺0.025(6) *
HT, bcc 1.322共3兲 0.674共2兲 0.039共7兲 * ⫺0.022(6) *
Reference 6 共1999兲 MC 1.3190共24兲 * 0.6723共11兲 0.0381共4兲 ⫺0.0169(33) *
Reference 32 共1999兲 MC 1.315共12兲 * 0.6693共58兲 0.035共5兲 ⫺0.008(17) *
Reference 33 共1996兲 MC 1.316共3兲 * 0.6721共13兲 0.0424共25兲 ⫺0.0163(39) *
Reference 34 共1995兲 MC 0.6724共17兲 ⫺0.017(5) *
Reference 35 共1993兲 MC 1.307共14兲 * 0.662共7兲 0.026共6兲 ⫺0.014(21) *
Reference 36 共1990兲 MC 1.316共5兲 0.670共2兲 0.036共14兲 * ⫺0.010(6) *
Reference 30 共2001兲 FT d⫽3 exp 1.3164共8兲 0.6704共7兲 0.0349共8兲 ⫺0.0112(21)
Reference 7 共1998兲 FT d⫽3 exp 1.3169共20兲 0.6703共15兲 0.0354共25兲 ⫺0.011(4)
Reference 7 共1998兲 FT ⑀ -exp 1.3110共70兲 0.6680共35兲 0.0380共50兲 ⫺0.004共11兲
Reference 1,3 共1996兲 4
He 0.67019共13兲 * ⫺0.01056(38)
Reference 37 共1993兲 4
He 0.6705共6兲 ⫺0.0115(18) *
Reference 38 共1992兲 4
He 0.6708共4兲 ⫺0.0124(12) *
Reference 39 共1984兲 4
He 0.6717共4兲 ⫺0.0151(12) *
Reference 40 共1983兲 4
He 0.6709共9兲 * ⫺0.0127(26)
In Table I we report our results for the critical exponents, analytic properties 共Griffiths’ analyticity兲, and take into ac-
i.e., our best estimates obtained by combining MC and IHT count the Goldstone singularities at the coexistence curve.
techniques—they are denoted by MC⫹IHT—together with From our approximate representations of the equation of
the results obtained from the analysis of the MC data alone. state we derive estimates of several universal amplitude ra-
There, we also compare them with the most precise experi- tios. The specific-heat amplitude ratio is particularly interest-
mental and theoretical estimates that have been obtained in ing since it can be compared with experimental results. We
the latest years. When only or ␣ is reported, we used the obtain A ⫹ /A ⫺ ⫽1.062(4), which is not in agreement with
hyperscaling relation 2⫺ ␣ ⫽3 to obtain the missing expo- the experimental result A ⫹ /A ⫺ ⫽1.0442 of Refs. 1 and 3. It
nent. Analogously, if only or ␥ is quoted, the second ex- is easy to trace the origin of the discrepancy. In our method
ponent was obtained using the scaling relation ␥ ⫽(2⫺ ) ; as well as in the analysis of the experimental data, the esti-
in this case the uncertainty was obtained using the mate of A ⫹ /A ⫺ is strongly correlated with the estimate of ␣ .
independent-error formula. The results we quote have been Therefore, the discrepancy we observe for this ratio is a di-
obtained from the analysis of the HT series of the XY model rect consequence of the difference in the estimates of ␣ .
共HT兲, by Monte Carlo simulations or by field-theory 共FT兲 Finally, we also discuss the two-point function of the or-
methods. The HT results of Ref. 14 have been obtained ana- der parameter, i.e., the structure factor, which is relevant in
lyzing the 21st-order HT expansions for the standard XY scattering experiments with magnetic materials.
model on the sc and the bcc lattice, using biased approxi- The paper is organized as follows. In Sec. II we present
mants and taking  c and ⌬ from other approaches, such as our Monte Carlo results. After reviewing the basic RG ideas
MC and FT. The FT results of Refs. 7,30 have been derived behind our methods, we present a determination of the im-
by resumming the known terms of the fixed-dimension g proved Hamiltonians and of the critical exponents. We dis-
expansion: the  function is known to six-loop order,41 while cuss the several possible sources of systematic errors, and
the critical-exponent series are known to seven loops.42 The show that the approximate improved models we use have
estimates from the ⑀ expansion have been obtained resum- significantly smaller corrections than the standard XY model.
ming the available O( ⑀ 5 ) series.43,44 A careful analysis shows that the leading scaling corrections
We also present a detailed study of the equation of state. are reduced at least by a factor of 20. We also compute  c to
We first consider its expansion in terms of the magnetization high precision for several values of and D; this is an im-
in the high-temperature phase. The coefficients of this expan- portant ingredient in our IHT analyses. Details on the algo-
sion are directly related to the zero-momentum n-point rithm appear in Appendix A.
renormalized couplings, which were determined by analyz- In Sec. III we present our results for the critical exponents
ing their IHT expansion. These results are used to construct obtained from the analysis of the IHT series. The equation of
parametric representations of the critical equation of state state is discussed in Sec. IV. After reviewing the basic defi-
that are valid in the whole critical region, satisfy the correct nitions and properties, we present the coefficients of the
214503-3
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
small-magnetization expansion, again computed from IHT present work we used a version of the algorithm that avoids
series. We discuss parametric representations that provide the flip to antiperiodic boundary conditions. For a detailed
approximations of the equation of state in the whole critical discussion see Appendix A 2.
region and compute several universal amplitude ratios. In
Sec. V we analyze the two-point function of the order pa-
rameter. Details of the IHT analyses are reported in Appen- B. Summary of finite-size methods
dix B. The definitions of the amplitude ratios we compute In this subsection we discuss the FSS methods we used to
can be found in Appendix C. compute the inverse critical temperature, the couplings *
and D * at which leading corrections to scaling vanish, and
II. MONTE CARLO SIMULATIONS the critical exponents and .
A. The lattice and the quantities that were measured
1. Summary of basic RG results
We simulated sc lattices of size V⫽L 3 , with periodic
boundary conditions in all three directions. In addition to The following discussion of FSS is based on the RG
elementary quantities such as the energy, the magnetization, theory of critical phenomena. We first summarize some basic
the specific heat or the magnetic susceptibility, we computed results. In the three-dimensional XY universality class there
so-called phenomenological couplings, i.e., quantities that, in exist two relevant scaling fields u t and u h , associated with
the critical limit, are invariant under RG transformations. the temperature and the applied field, respectively, with RG
They are well suited to locate the inverse critical temperature exponents y t and y h . Moreover, there are several irrelevant
 c . They also play a crucial role in the determination of the scaling fields that we denote by u i , i⭓3, with RG exponents
improved Hamiltonians. 0⬎y 3 ⬎y 4 ⬎y 5 ⬎•••.
In the present study we consider four phenomenological The RG exponent y 3 ⬅⫺ of the leading irrelevant scal-
couplings. We use the Binder cumulant U 4 and the similar ing field u 3 has been computed by various methods. The
quantity U 6 , defined by analysis of field-theoretical perturbative expansions7 gives
⫽0.802(18) ( ⑀ expansion兲 and ⫽0.789(11) (d⫽3 ex-
ជ 2 兲 j典
具共 m pansion兲. In the present work we find a result for that is
U2 j⬅ , 共9兲 consistent with, although less accurate than, the field-
ជ 2典 j
具m theoretical predictions. We also mention the estimate
where mជ ⫽1/V 兺 x
ជ x is the magnetization of the system. We ⫽0.85(7) that was obtained9 by the ‘‘scaling-field’’ method,
also consider the second-moment correlation length divided a particular implementation of Wilson’s ‘‘exact’’ renormal-
by the linear extension of the lattice 2nd /L. The second- ization group. Although it provides an estimate for that is
moment correlation length is defined by less precise than those obtained from perturbative field-
theoretic methods, it has the advantage of giving predictions
2nd⬅ 冑 /F⫺1
4 sin共 /L 兲 2
, 共10兲
for the irrelevant RG exponents beyond y 3 . Ref. 9 predicts
y 4 ⫽⫺1.77(7) and y 5 ⫽⫺1.79(7) (y 421 and y 422 in their no-
tation兲 for the XY universality class. Note that, at present,
where there is no independent check for these results. Certainly it
would be worthwhile to perform a Monte Carlo renormaliza-
⬅
1
V 冓冉兺 冊 冔
x
ជx
2
共11兲
tion group study. With the computational power available
today, it might be feasible to resolve subleading correction
exponents with a high-statistics simulation.
is the magnetic susceptibility and In the case of U 4 , U 6 , and 2nd /L we expect a correction
冓冏兺 冉 冊 冏 冔
caused by the analytic background of the magnetic suscepti-
1 2x1 bility. This should lead to corrections with y 6 ⫽⫺(2⫺ )⬇
F⬅ exp i ជx
2
共12兲
V x L ⫺1.962. We also expect corrections due to the violation of
rotational invariance by the lattice. For the XY universality
is the Fourier transform of the correlation function at the class, Ref. 47 predicts y 7 ⫽⫺2.02(1). Note that the numeri-
lowest nonzero momentum. cal values of y 6 and y 7 are virtually identical and should
The list is completed by the ratio Z a /Z p of the partition hence be indistinguishable in the analysis of our numerical
function Z a of a system with antiperiodic boundary condi- data.
tions in one of the three directions and the partition function We wish now to discuss the FSS behavior of a phenom-
Z p of a system with periodic boundary conditions in all di- enological coupling R; in the standard RG framework, we
rections. Antiperiodic boundary conditions in the first direc- can write it as a function of the thermal scaling field u t and
tion are obtained by changing sign to the term ជ x
ជ y of the of the irrelevant scaling fields u i . For L→⬁ and 
Hamiltonian for links 具 xy 典 that connect the boundaries, i.e., →  c (), we have
for x⫽(L,x 2 ,x 3 ) and y⫽(1,x 2 ,x 3 ). The ratio Z a /Z p can be
measured by using the boundary-flip algorithm, which was
applied to the three-dimensional Ising model in Ref. 45 and
generalized to the XY model in Ref. 46. As in Ref. 26, in the
R 共 L,  , 兲 ⫽r 0 共 u t L y t 兲 ⫹ 兺 r i共 u t L y 兲 u i L y ⫹•••,
i⭓3
t i 共13兲
214503-4
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
where we have neglected terms that are quadratic in the scal- R 共 L,  兲 ⫽R * , 共19兲
ing fields of the irrelevant operators, i.e., corrections of order
L 2y 3 ⬇L ⫺1.6. Note that we include here the corrections due to where the corrections vanish like L ⫺y t ⫹y 3 .
the analytic background 共with exponent L ⫺y 6 ⬇L ⫺2 ). In the
3. Locating * and D *
case of U 4 , U 6 , and 2nd /L 共but not Z a /Z p ), in Eq. 共13兲 we
have also discarded terms of order L y t ⫺2y h ⬇L ⫺3.5. In order to compute the value * for which the leading
The functions r 0 (z) and r i (z) are smooth and finite for corrections to scaling vanish, we use two phenomenological
z→0, while u t (  ,) and u i (  ,) are smooth functions of  couplings R 1 and R 2 . First, we define  f (L,) by
and . Note that, by definition, u t (  ,)⬃  ⫺  c (). In the
limit t→0 and u t L y t ⬃(  ⫺  c )L 1/ →0, we can further ex- R 1 共 L,  f , 兲 ⫽R 1,f , 共20兲
pand Eq. 共13兲, obtaining where R 1,f is a fixed value, which we can choose freely
within the appropriate range. It is easy to see that  f (L,)
R 共 L,  , 兲 ⫽R * ⫹c t 共  , 兲 L y t ⫹ 兺i c i共  , 兲 L y i →  c () as L→⬁. Indeed, using Eq. 共13兲, we have
r 1,3共 z f 兲
⫹O 关共  ⫺  c 兲 2 L 2y t ,L 2y 3 ,tL y t ⫹y 3 兴 , 共14兲  f 共 L, 兲 ⫺  c 共 兲 ⫽z f L ⫺y t ⫺ u 3 共  c 兲 L y 3 ⫺y t ⫹•••,
⬘ 共zf兲
r 1,0
where R * ⫽r 0 (0) is the value at the critical point of the 共21兲
phenomenological coupling.
where we have used y t ⬎ 兩 y 3 兩 and z f is defined as the solution
2. Locating  c of r 1,0(z f )⫽R 1,f . We have added a subscript 1 to make ex-
plicit that all scaling functions refer to R 1 . If R 1,f ⬇R * , we
We locate the inverse critical temperature  c by using
can expand the previous formula, obtaining
Binder’s cumulant crossing method. This method can be ap-
plied in conjunction with any of the four phenomenological R 1,f ⫺R * c 1,i
1
couplings that we computed.  f 共 L, 兲 ⬇  c 共 兲 ⫹ L ⫺y t ⫺ L ⫺y t ⫹y 3 . 共22兲
In its simplest version, one considers a phenomenological ⬘
c 1,t ⬘
c 1,t
coupling R(  ,L) for two lattice sizes L and L ⬘ ⫽bL. The Notice that for R 1,f ⫽R 1* the convergence is faster, and thus
intersection  cross of the two curves R(  ,L) and R(  ,L ⬘ ) we will always take R 1,f ⬇R 1* . Next we define
provides an estimate of  c . The convergence rate of this
estimate  cross toward the true value can be computed in the R̄ 共 L, 兲 ⬅R 2 共 L,  f , 兲 . 共23兲
RG framework.
By definition,  cross at fixed b, L, and is given by the For L→⬁ and R 1,f ⬇R 1* , we have
冉 冊
solution of the equation
⬘
c 2,t ⬘
c 2,t
R 共 L,  , 兲 ⫽R 共 bL,  , 兲 . 共15兲 R̄ 共 L, 兲 ⫽R *
2⫹
⬘
c 1,t
共 R 1,f ⫺R 1* 兲 ⫹ 兺i c 2,i ⫺
⬘
c 1,t
c 1,i L y i
Using Eq. 共13兲, one immediately verifies that  cross con-
verges to  c faster than L ⫺y t . Thus, for L→⬁, we can use
Eq. 共14兲 and rewrite Eq. 共15兲 as
⫽R̄ * ⫹ 兺i c̄ i共 兲 L y ,i 共24兲
214503-5
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
wish to see a good signal for the corrections. This means, in Let us consider a second phenomenological coupling R̄
⬘ c 2,3⫺c 2,t
particular, that in c 1,t ⬘ c 1,3 the two terms should add ⫽R(L,  f ,), which, for L→⬁, behaves as
up rather than cancel. Of course, also the corrections due to
the subleading scaling fields should be small. R̄⫽R̄ * ⫹c̄ 3 L ⫺ . 共31兲
4. The critical exponents The universality of the correction amplitudes implies that the
Typically, the thermal RG exponent y t ⫽1/ is computed ratio B/c̄ 3 is the same for the 4 model and the dd-XY
from the FSS of the derivative of a phenomenological cou- model and is independent of and D. Therefore, this ratio
pling R with respect to  at  c . Using Eq. 共13兲 one obtains can be computed in models that have large corrections to
scaling, e.g., in the standard XY model. Then, we can com-
R

冏 c
⫽r 0⬘ 共 0 兲 L y t ⫹ 兺 r i⬘共 0 兲 u i共  c 兲 L y ⫹y
i⫽3
i t
pute a bound on B for the 共approximately兲 improved model
from the known ratio B/c̄ 3 and a bound for c̄ 3 . This proce-
dure was proposed in Ref. 23.
⫹ 兺 r i共 0 兲 u ⬘i 共  c 兲 L y ⫹•••.
i⫽3
i 共28兲
C. The simulations
Hence, the leading corrections scale with L y 3 . However, in We simulated the 4 and the dd-XY model using the
improved models in which u 3 (  c )⫽0, the leading correction wall-cluster update algorithm of Ref. 23 combined with a
local update scheme. The update algorithm is discussed in
is of order L y 4 . Note that corrections proportional to L y 3 ⫺y t
Appendix A.
⬇L ⫺2.3 are still present even if the model is improved. In
Most of the analyses need the quantities as functions of  .
Ref. 23, for the spin-1 Ising model, an effort was made to
Given the large statistics, we could not store all individual
eliminate also this correction by taking the derivative with
measurements of the observables. Therefore, we did not use
respect to an optimal linear combination of  and D instead
the reweighting method. Instead, we determined the Taylor
of  . Here we make no attempt in this direction, since cor-
coefficients of all quantities of interest up to the third order
rections of order L ⫺2.3 are subleading with respect to those
in (  ⫺  s ), where  s is the value of  at which the simula-
of order L y 4 ⬇L ⫺1.8.
tion was performed. We checked carefully that this is suffi-
In practice it is difficult to compute the derivative at  c ,
cient for our purpose.
since  c is only known numerically, and therefore, it is more
Most of our simulations were performed at ⫽2.1 in the
convenient to evaluate R/  at  f 关see Eq. 共20兲兴. This pro-
case of the 4 model and at D⫽1.03 in the case of the dd-
cedure has been used before, e.g., in Ref. 33. In this case, Eq.
XY model. ⫽2.1 is the estimate of * of Ref. 6, and D
共28兲 still holds, although with different amplitudes that de-
⫽1.03 is the result for D * of a preliminary analysis of MC
pend on the particular choice of the value of R f .
data obtained on small lattices. In addition, we performed
The exponent is computed from the finite-size behavior
simulations at ⫽2.0 and 2.2 for the 4 model and D
of the magnetic susceptibility, i.e.,
⫽0.9 and 1.2 for the dd-XY model in order to obtain an
兩  f ⬀L 2⫺ . 共29兲 estimate of the derivative of the amplitude of the leading
corrections to scaling with respect to and D, respectively.
Also, here the corrections are of order L y 3 for generic mod- We also performed simulations of the standard XY model in
els, and of order L y 4 for improved ones. order to estimate the effect of the leading corrections to scal-
ing on the estimates of the critical exponents obtained from
5. Estimating errors caused by residual leading the FSS analysis.
scaling corrections
In Ref. 25, the authors pointed out that with the method D.  c and the critical value of phenomenological couplings
discussed in Sec. II B 3, the leading corrections are only ap- In a first step of the analysis we computed R * and the
proximately eliminated, so that there is still a small leading inverse critical temperature  c at ⫽2.1 and D⫽1.03, re-
scaling correction that causes a systematic error in the esti- spectively.
mates of, e.g., the critical exponents. The most naive solution For ⫽2.1 and D⫽1.03 we simulated sc lattices of linear
to this problem consists in adding a term L ⫺ to the fit size L from 4 to 16 and L⫽18, 20, 22, 24, 26, 28, 32, 36,
ansatz, i.e., in considering 40, 48, 56, 64, and 80. For all lattice sizes smaller than L
⫽24 we performed 108 measurements. For larger lattices, we
R

冏 f
⫽A L 1/ 共 1⫹B L ⫺ 兲 . 共30兲 collected approximately 107 measurements for L⬇40, and
approximately 106 for the largest lattices with L⬇80. A
measurement was performed after an update cycle as dis-
However, by adding such a correction term, the precision of cussed in Appendix A 1.
the result decreases, so that there is little advantage in using Instead of computing R * and  c from two lattice sizes as
共approximately兲 improved models. A more sophisticated ap- discussed in Sec. II B 2, we perform a fit with the ansatz
proach is based on the fact that ratios of leading correction
amplitudes are universal.48 R * ⫽R 共 L,  c 兲 , 共32兲
214503-6
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
TABLE II. Joint fits of the Z a /Z p data of the 4 model and the dd-XY model with the ansatz 共33兲. All
lattice sizes L min⭐L⭐L max are used in the fit. In column four we give the results of the fits for  c of the 4
model at ⫽2.1 and in column five the results for  c of the dd-XY model at D⫽1.03. Finally, in column six
we give the results for the fixed-point value (Z a /Z p ) * . The final results and an estimate of the systematic
errors are given in the text.
where R * and  c are free parameters. We compute R(L,  ) Our final result is obtained from the fit with L min⫽28 and
using its third-order Taylor expansion L max⫽80. The systematic error is estimated by comparing
this result with that obtained using L min⫽10 and L max⫽28.
R * ⫽R 共 L,  s 兲 ⫹d 1 共 L,  s 兲共  c ⫺  s 兲 ⫹ 21 d 2 共 L,  s 兲共  c ⫺  s 兲 2 The systematic error on  c is estimated by the difference of
the results from the two fits divided by 2.82.3⫺1, where 2.8
⫹ 16 d 3 共 L,  s 兲共  c ⫺  s 兲 3 , 共33兲 is the scale factor between the two intervals and y t ⫺y 3
⬇2.3. Estimating the systematic error by comparison with
where  s is the  at which the simulation was performed, the interval L min⫽14 and L max⫽40 leads to a similar result.
and R, d 1 , d 2 , and d 3 are determined in the MC simulation. We obtain  c ⫽0.509 150 7(6) 关 7 兴 for the 4 model at
First, we perform fits for the two models separately. We ⫽2.1 and  c ⫽0.562 797 5(7) 关 7 兴 for the dd-XY model at
obtain consistent results for R * for all four choices of phe- D⫽1.03. In parentheses we give the statistical error and in
nomenological couplings. In order to obtain more precise the brackets the systematic one. Our final result for the criti-
results for  c and R * , we perform joint fits of both models. cal ratio of partition functions is (Z a /Z p ) * ⫽0.3202(1) 关 5 兴 .
Here, we exploit universality by requiring that R * takes the Here the systematic error is computed by dividing the differ-
same value in both models. Hence, such fits have three free ence of the results of the two fits by 2.80.8⫺1.
parameters: R * and the two values of  c . In the following We repeat this analysis for 2nd , U 4 and U 6 . The final
we shall only report the results of such joint fits. results are summarized in Table III.
Let us discuss in some detail the results for R⫽Z a /Z p Next we compute  c at additional values of and D. For
that are summarized in Table II. In each fit, we take all data this purpose we simulated lattices of size L⫽96 and compute
with L min⭐L⭐L max into account. For L max⫽80, the  c using Eq. 共19兲. We use only R⫽Z a /Z p with the above-
2 /degree of freedom 共DOF兲 becomes approximately one reported estimate (Z a /Z p ) * ⫽0.3202(6). The results are
starting from L min⫽15. However, we should note that a summarized in Table IV.
2 /DOF close to one does not imply that the systematic er-
rors due to corrections that are not taken into account in the E. Eliminating leading corrections to scaling
ansatz are negligible. In this subsection we determine * and D * . For this pur-
pose, we compute the correction amplitude c̄ 3 for various
TABLE III. Summary of the final results for  c and R * . In
column one the choice of the phenomenological coupling R is TABLE IV. Estimates of  c from simulations of 963 lattices.  c
listed. In columns two and three we report our estimates of  c for is obtained from Eq. 共19兲 using Z a /Z p as phenomenological cou-
the 4 model at ⫽2.1 and the dd-XY model at D⫽1.03, and pling. In parentheses we give the statistical error and in brackets the
finally in column four the result for the fixed-point value of the error due to the error on (Z a /Z p ) * . ‘‘stat’’ gives 共the number of
phenomenological coupling. Note that the estimates of  c , based measurements兲/1000.
on the four different choices of R, are consistent within error bars.
Model ; D stat c
R  c , ⫽2.1  c , D⫽1.03 R*
4 2.07 545 0.5093853共16兲关8兴
Z a /Z p 0.5091507共6兲关7兴 0.5627975共7兲关7兴 0.3202共1兲关5兴 4 2.2 510 0.5083366共16兲关8兴
2nd /L 0.5091507共7兲关3兴 0.5627971共7兲关2兴 0.5925共1兲关2兴 dd-XY 0.9 720 0.5764582共15兲关9兴
U4 0.5091495共9兲关10兴 0.5627972共10兲关11兴 1.2430共1兲关5兴 dd-XY 1.02 1,215 0.5637972共12兲关9兴
U6 0.5091498共9兲关15兴 0.5627976共10兲关15兴 1.7505共3兲关25兴 dd-XY 1.2 665 0.5470377共17兲关9兴
214503-7
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
TABLE V. The quantity (R̄ 兩 D⫽0.9⫺R̄ 兩 D⫽1.2)L 0.8 for the dd-XY model. In the top row we give the choice
of R 1 and R 2 . For instance, U 4 at (Z a /Z p ) f means that R 1 ⫽Z a /Z p and R 2 ⫽U 4 .
choices of R 1 and R 2 for the 4 model at ⫽2.1 and the dd- In order to see whether we can predict the exponent , we
XY model at D⫽1.03. In order to convert these results into perform a fit with the ansatz
estimates of * and D * , we determine the derivative of the
correction amplitude c̄ 3 with respect to 共resp. D) at ⌬R̄⫽k L ⫺ , 共35兲
⫽2.1 共resp. D⫽1.03). We also simulated the XY model in
order to obtain estimates of the residual systematic error due
with k and as free parameters. From U 4 at Z a /Z p
to the leading corrections to scaling. Note that, in the follow-
⫽0.3202 we get ⫽0.795(9) with 2 /DOF⫽0.66, using all
ing, we always use as the value of R 1,f in Eq. 共20兲 the esti-
available data. This value is certainly consistent with field-
mates of R * given in Table III.
theoretical results. Note, however, that we would like to vary
the range of the fit in order to estimate systematic errors. For
1. Derivative of the correction amplitude with respect to or D
this purpose more data at larger values of L are needed.
For this purpose we simulated the dd-XY model at D In the following we need estimates of dc̄ 3 /dD 兩 D⫽1.03 and
⫽0.9 and D⫽1.2 on lattices of size L⫽5, 6, 7, 8, 9, 10, of the corresponding quantity for the 4 model, in order to
12, and 16. The 4 model was simulated at ⫽2.0 and determine D * and * . We approximated this derivative by a
⫽2.2 on lattices of size L⫽3, 4, 5, 6, 7, 8, and 9. In the finite difference between D⫽0.9 and D⫽1.2. The coefficient
case of the dd-XY model we performed 100⫻106 measure- c̄ 3 is determined by fixing ⫽0.8. Our final result is the
ments for each parameter set. In the case of the 4 model average of the estimates for L⫽10,12 and 16 in Table V. In
250⫻106 measurements were performed. a similar way we proceed in the case of the 4 model, aver-
In the following we discuss only the dd-XY model, since aging the L⫽8,9 results. The results are summarized in
the analysis of the 4 data is performed analogously. Table VI. We make no attempt to estimate error bars.
In Refs. 6,26 it was observed that subleading corrections Sources of error are the finite difference in D, subleading
to scaling cancel to a large extent when one considers the corrections, the error on , and the statistical errors. Note,
difference of R̄ at close-by values of . In order to get an however, that these errors are small enough to be neglected
idea of the size of the corrections, we report in Table V in the following.
TABLE VI. Estimates for dc̄ 3 /dD at D⫽1.03 共dd-XY ) and dc̄ 3 /d at ⫽2.1 ( 4 ). In the first row we
give the combination of R 1 and R 2 .
214503-8
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
TABLE VII. Results of fits with the ansatz 共36兲. The coefficients c̄ 3 are converted into D * and * using
Eq. 共37兲. All data with L min⭐L⭐L max are fitted.
The results of the fits for four different combinations of L min⫽20 and L max⫽80, including the statistical error, and
R 1 and R 2 are given in Table VII, where we have already therefore should take into proper account all systematic er-
translated the results for c̄ 3 into an estimate of * and D * , rors.
by using From these results, it is also possible to obtain a conser-
vative upper bound on the coefficient c̄ 3 for ⫽2.1 and D
* ⫽2.1⫺c̄ 3 冉 冊
dc̄ 3
d
⫺1
共37兲
⫽1.03. Indeed, using the estimates of * and D * and their
errors, we can obtain the upper bounds 兩 2.1⫺ * 兩 ⬍⌬
⫽0.08 and 兩 1.03⫺D * 兩 ⬍⌬D⫽0.04. Then, we can estimate
for the 4 model and the analogous formula for the dd-XY 兩 c̄ 3 (⫽2.1) 兩 ⬍⌬(dc̄ 3 /d), and analogously 兩 c̄ 3 (D
model, and the results of Table VI. ⫽1.03) 兩 ⬍⌬D(dc̄ 3 /dD). For U 4 at (Z a /Z p ) f ⫽0.3202, us-
A 2 /DOF close to 1 is reached for L min⫽10 and L max ing the results of Table VI, we have
⫽80 in the case of U 4 at (Z a /Z p ) f ⫽0.3202. This has to be
compared with L min⫽11, 11, and 14 in the case of U 6 at
(Z a /Z p ) f ⫽0.3202, U 4 at ( 2nd /L) f ⫽0.5925, and U 6 at 兩 c̄ 3 共 ⫽2.1兲 兩 ⬍0.004, 兩 c̄ 3 共 D⫽1.03兲 兩 ⬍0.005. 共38兲
( 2nd /L) f ⫽0.5925.
This indicates that U 4 at (Z a /Z p ) f ⫽0.3202 has the least
3. Corrections to scaling in the standard XY model
bias due to subleading corrections to scaling. Therefore we
take as our final result * ⫽2.07 and D * ⫽1.02 that is the We simulated the standard XY model on lattices with lin-
result of L min⫽12 and L max⫽80 in Table VII. Starting from ear sizes L⫽6, 8, 10, 12, 16, 18, 20, 22, 24, 28, 32, 48, 56,
L min⫽20 all results for * and D * are within 2 of our final and 64 at  s ⫽0.454 165, which is the estimate of  c of Ref.
result quoted above. 33. Here, we used only the wall-cluster algorithm for the
Our final results are * ⫽2.07(5) and D * ⫽1.02(3). The update. In one cycle we performed 12 wall-cluster updates.
error bars are such to include all results in Table VII with For L⭐16 we performed 108 cycles. For lattice sizes 16
214503-9
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
TABLE VIII. Corrections to scaling in the standard XY model TABLE IX. Fit results for the critical exponent obtained from
for R̄ with R 2 ⫽U 4 and R 1,f ⫽(Z a /Z p ) f ⫽0.3202. We use the ansatz the ansatz 共39兲. In all cases  f is fixed by (Z a /Z p ) f ⫽0.3202. We
共36兲 with ⫽0.8 fixed. analyze the 4 model at ⫽2.1, the dd-XY model at D⫽1.03, and
the standard XY model. We consider the slope of the Binder cumu-
L min L max 2 /DOF R̄ * c̄ 3 lant U 4 and of the ratio of partition functions Z a /Z p . We included
all data with L min⭐L⭐L max into the fit.
12 64 1.78 1.2432共1兲 ⫺0.1120(7)
16 64 0.73 1.2430共1兲 ⫺0.1087(13) L min L max 2 /DOF
20 64 0.38 1.2427共2兲 ⫺0.1048(22)
4 model: derivative of U 4
24 64 0.24 1.2429共2兲 ⫺0.1083(34)
7 80 1.17 0.67168共12兲
12 32 2.32 1.2433共1兲 ⫺0.1124(8)
9 80 0.79 0.67188共15兲
11 80 0.85 0.67181共19兲
16 80 0.98 0.67192共34兲
⭐L⭐64, we spent roughly the same amount of CPU time for 4 model: derivative of Z a /Z p
each lattice size. For L⫽64 the statistics is 2.73⫻106 mea- 12 80 3.01 0.67042共9兲
surements.
16 80 1.61 0.67104共15兲
We determine the amplitude of the corrections to scaling
20 80 1.04 0.67139共22兲
for R̄ with R 1,f ⫽(Z a /Z p ) f ⫽0.3202 and R 2 ⫽U 4 . Other 24 80 0.54 0.67194共32兲
choices lead to similar results. We fit our numerical results dd-XY model: derivative of U 4
with the ansatz 共36兲, where we fix ⫽0.8. The results are 7 80 2.06 0.67258共12兲
given in Table VIII.
9 80 1.13 0.67216共15兲
Note that the results for R̄ * are consistent with the result 11 80 1.19 0.67209共19兲
obtained from the joint fit of the two improved models. In 16 80 0.97 0.67154共31兲
Table VII we obtained, e.g., R̄ * ⫽1.243 01(8) with L min dd-XY model: derivative of Z a /Z p
⫽20 and L max⫽80. 12 80 1.89 0.67017共9兲
Corrections to scaling are clearly visible, see Fig. 1. From 16 80 1.60 0.67046共14兲
the fit with L min⫽20 and L max⫽64 we obtain c̄ 3 20 80 0.79 0.67099共21兲
⫽⫺0.1048(22). For the following discussion no estimate of 24 80 0.80 0.67113共30兲
the possible systematic errors of c̄ 3 is needed. Comparing XY model: derivative of U 4
with Eq. 共38兲, we see that in the 共approximately兲 improved 12 64 4.48 0.66450共28兲
models the amplitude of the leading correction to scaling is 16 64 1.30 0.66618共42兲
reduced by a factor of at least 20. Note, that even if this 20 64 0.54 0.66740共63兲
result was obtained by considering a specific observable, U 4 XY model: derivative of Z a /Z p
at fixed Z a /Z p , the universality of the ratios of the sublead- 12 64 1.33 0.67263共13兲
ing corrections implies the same reduction for any quantity. 16 64 0.69 0.67300共19兲
In the following section we will use this result to estimate the 20 64 0.30 0.67325共30兲
systematic error on our results for the critical exponents. 24 64 0.25 0.67327共41兲
R
兩 ⫽cL 1/ 共39兲
 f
for the 4 model at ⫽2.1, the dd-XY model at D⫽1.03,
FIG. 1. Corrections to scaling for the dd-XY model at D⫽0.9, and the standard XY model.
1.03, and 1.2, and for the standard XY model. We plot R̄ with We see that for the same L min and L max the statistical error
R 1,f ⫽(Z a /Z p ) f ⫽0.3202 and R 2 ⫽U 4 as a function of the lattice on the estimate of obtained from the derivative of Z a /Z p is
size. The dotted lines should only guide the eye. smaller than that obtained from the derivative of U 4 . On the
214503-10
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
other hand, for the two improved models, scaling corrections TABLE X. Results for the critical exponent from the FSS of
seem to be larger for Z a /Z p than for U 4 . the magnetic susceptibility. Fits with ansatz 共40兲. All data with
In the case of Z a /Z p , for both improved models, the re- L min⭐L⭐L max are taken into account.
sult of the fit for is increasing with increasing L min . In the
case of the Binder cumulant, it is increasing with L min for the L min L max 2 /DOF
4 model and decreasing for the dd-XY model. The fact that model
4
scaling corrections affect the two quantities and the two im- 20 80 2.44 0.0371共1兲
proved models in a quite different way suggests that system- 24 80 0.73 0.0375共1兲
atic errors in the estimate of can be estimated from the 28 80 0.94 0.0375共2兲
variation of the fits presented in Table IX.
32 80 0.41 0.0378共3兲
As our final result we quote ⫽0.6716(5) that is consis-
dd-XY model
tent with the two results from Z a /Z p at L min⫽24 and with
20 80 1.88 0.0371共1兲
the results from U 4 at L min⫽16.
24 80 1.19 0.0373共1兲
In the case of the standard XY model, the derivative of U 4
28 80 1.52 0.0374共2兲
requires a much larger L min to reach a small 2 /DOF than for
32 80 1.24 0.0376共2兲
the improved models. For the derivative of Z a /Z p instead a
XY model
2 /DOF⬇1 is obtained for an L min similar to that of the
improved models. Note that the result for from the deriva- 20 64 7.92 0.0325共2兲
tive of U 4 for L min⫽16 is by several standard deviations 24 64 1.81 0.0344共2兲
smaller than our final result from the improved models, 28 64 0.27 0.0340共3兲
while the result from the derivative of Z a /Z p is by several 32 64 0.06 0.0342共4兲
standard deviations larger. Again we have a nice example
that a 2 /DOF⬇1 does not imply that systematic errors due
to corrections that have not been taken into account in the fit by subleading corrections. Therefore, we consider 0.0375,
are small. which is the result of the fit with L min⫽24 in the 4 model,
Remember that in improved models the leading correc- as a lower bound of .
tions to scaling are suppressed at least by a factor of 20 with Finally, we perform a fit that takes into account the ana-
respect to the standard XY model. Since the range of lattice lytic background of the magnetic susceptibility. In Ref. 6, it
sizes is roughly the same for the XY model and for the im- was shown that the addition of a constant term to Eq. 共40兲
proved models, we can just divide the deviation of the XY leads to a small 2 /DOF already for small L min⬍10. Similar
results from ⫽0.6716(5) by 20 to obtain an estimate of the results have been found for the Ising universality class. This
possible systematic error due to the residual leading correc- ansatz is not completely correct, since it does not take into
tions to scaling. For the derivative of Z a /Z p we end up with account corrections proportional to L 2⫹y with y⬇⫺1.8,
0.0001 and for the derivative of U 4 with 0.0003. which formally are more important than the analytic back-
We think that these errors are already taken into account ground. However, the difference between these exponents is
by the spread of the results for from the derivatives of U 4 small, and a four-parameter fit is problematic. Therefore, we
and Z a /Z p and the two improved models. Therefore, we decided to fit our data with the ansatz
keep our estimate ⫽0.6716(5) with its previous error bar.
Next we compute the exponent . For this purpose we 兩  f ⫽c L 2⫺ ⫹b L , 共41兲
study the finite-size behavior of the magnetic susceptibility
at  f . In the following we fix  f by R 1,f ⫽(Z a /Z p ) f
⫽0.3202. Other choices for R 1,f give similar results. with fixed to 0.0 and to 0.2. The difference between the
In a first attempt we fit the data of the two improved results of the fits with the two values of will give an
models and the standard XY model to the simple ansatz estimate of the systematic error of the procedure. Results for
all three models are summarized in Table XI.
The value 2 /DOF is close to one for L min⫽8 for the 4
兩  f ⫽c L 2⫺ . 共40兲 model and L min⫽10 for the dd-XY model, and it does not
allow to discriminate between the two choices of . The
The results are summarized in Table X. values of are rather stable as L min is varied, although there
For all three models rather large values of L min are needed is a slight trend toward smaller results as L min increases; the
in order to reach a 2 /DOF close to one. In all cases the trend seems to be stronger for ⫽0.2. Moreover, the results
estimate of is increasing with increasing L min . For L min from the two models are in good agreement.
⫽24 the result for from the standard XY model is lower The fits for the XY model also give a good 2 /DOF for
than that of the improved models by an amount of approxi- L min⭓12; the value of is, however, much too small, and
mately 0.0030. Therefore, the systematic error due to leading shows an increasing trend. We can estimate from the differ-
corrections on the results obtained in the improved models ence between the XY model and the improved models at
should be smaller than 0.0030/20⫽0.000 15. Given this tiny L min⫽16 that the error on the value of obtained from im-
effect, it seems plausible that, for the improved models, the proved models, induced by residual leading scaling correc-
increase of the estimate of with increasing L min is caused tions, is smaller than 0.003/20⫽0.00 015.
214503-11
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
TABLE XI. Results for the critical exponent from the FSS of We analyze the HT series by means of integral approxi-
the magnetic susceptibility. Fits with ansatz 共41兲. All data with mants 共IA’s兲 of first, second, and third order. The most pre-
L min⭐L⭐L max are taken into account. cise results are obtain biasing the value of  c , using its MC
estimate. We consider several sets of biased IA’s, and for
Fit with ⫽0.0 Fit with ⫽0.2 each of them we obtain estimates of the critical exponents.
L min L max 2 /DOF 2 /DOF These results are reported in Appendix B 2. All sets of IA’s
4 model give substantially consistent results. Moreover, the results
8 80 0.72 0.0386共1兲 1.16 0.0391共1兲
are also stable with respect to the number of terms of the
series, so that there is no need to perform problematic ex-
10 80 0.68 0.0385共1兲 1.27 0.0388共1兲
trapolations in the number of terms in order to obtain the
12 80 0.75 0.0385共1兲 0.81 0.0388共1兲
final estimates. The error due to the uncertainty on * and
14 80 0.84 0.0386共2兲 0.92 0.0388共2兲
D * is estimated by considering the variation of the results
16 80 0.72 0.0384共2兲 0.73 0.0386共2兲
when changing the values of and D.
20 80 0.88 0.0384共3兲 0.88 0.0385共4兲
Using the intermediate results reported in Appendix B 2 ,
dd-XY model we obtain the estimates of ␥ and shown in Table XII. We
8 80 1.85 0.0387共1兲 3.06 0.0391共1兲 report on ␥ and three contributions to the error. The num-
10 80 0.95 0.0384共1兲 1.15 0.0388共1兲 ber within parentheses is basically the spread of the approxi-
12 80 0.99 0.0384共1兲 1.04 0.0386共1兲 mants at the central estimate of * (D * ) using the central
14 80 0.94 0.0384共2兲 1.03 0.0386共2兲 value of  c . The number within brackets is related to the
16 80 0.85 0.0383共2兲 1.14 0.0384共2兲 uncertainty on the value of  c and is estimated by varying  c
20 80 0.90 0.0381共4兲 0.90 0.0382共4兲 within one error bar at ⫽ * or D⫽D * fixed. The number
XY model within braces is related to the uncertainty on * or D * , and
12 64 0.64 0.0350共2兲 is obtained by computing the variation of the estimates when
16 64 0.48 0.0353共3兲 * or D * vary within one error bar, changing correspond-
20 64 0.37 0.0358共5兲 ingly the values of  c . The sum of these three numbers
should be a conservative estimate of the total error.
We determine our final estimates by combining the results
From the results for the improved models reported in for the two improved Hamiltonians: we take the weighted
Table XI, one would be tempted to take ⫽0.0384 as the average of the two results, with an uncertainty given by the
final result. However, as we can see from the results for the smallest of the two errors. We obtain for ␥ and
XY model, we should not trust blindly the good 2 /DOF of
these fits. Taking into account the decreasing trend of the ␥ ⫽1.3177共 5 兲 , 共42兲
values of for the improved models, we assign the conser-
vative upper bound ⬍0.0385. By combining it with the
lower bound obtained from ansatz 共40兲, we obtain our final ⫽0.671 55共 27兲 , 共43兲
result 0.0375⬍ ⬍0.0385, i.e., ⫽0.0380(5).
and by the hyperscaling relation ␣ ⫽2⫺3
III. HIGH-TEMPERATURE DETERMINATION
OF CRITICAL EXPONENTS
␣ ⫽⫺0.0146共 8 兲 . 共44兲
In this section we report the results of our analyses of the Consistent results, although significantly less precise 共ap-
HT series. The details are reported in Appendix B. proximately by a factor of 2兲, are obtained from the IHT
We compute ␥ and from the analysis of the HT expan- analysis without biasing  c 共see Appendix B 2兲.
sion to O(  20) of the magnetic susceptibility and of the From the results for ␥ and , we can obtain by the
second-moment correlation length. In Appendix B 2 we re- scaling relation ␥ ⫽(2⫺ ) . This gives ⫽0.0379(10),
port the details and many intermediate results so that the where the error is estimated by considering the errors on ␥
reader can judge the quality of our results without the need and as independent, which is of course not true. We can
of performing his own analysis. This should give an idea of obtain an estimate of with a smaller, yet reliable, error by
the reliability of our estimates and of the meaning of the applying the so-called critical-point renormalization method
errors we quote, which depend on many somewhat arbitrary 共CPRM兲 共see, e.g., Refs. 10 and references therein兲 to the
choices and are therefore partially subjective. series of and 2 . The results are reported in Table XII. We
TABLE XII. Estimates of the critical exponents obtained from the analysis of the HT expansion of the
improved 4 lattice Hamiltonian and dd-XY model.
␥ ␣
4 Hamiltonian 1.31780共10兲关27兴兵15其 0.67161共5兲关12兴兵10其 0.0380共3兲兵1其 ⫺0.0148(8)
dd-XY model 1.31748共20兲关22兴兵18其 0.67145共10兲关10兴兵15其 0.0380共6兲兵2其 ⫺0.0144(10)
214503-12
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
report two contributions to the error on , as discussed for ␥ with F(z)⬅ A/ z. Because of Griffiths’ analyticity, F(M )
and ; the uncertainty on  c does not contribute in this case. has also a regular expansion in powers of t for 兩 M 兩 fixed.
Our final estimate is Therefore, F(z) has the large-z expansion
⫽0.0380共 4 兲 . 共45兲
F 共 z 兲 ⫽z ␦ 兺
k⫽0
F ⬁k z ⫺k/  . 共54兲
Moreover, using the scaling relations we obtain
The function F(z) is defined only for t⬎0, and thus, in order
5⫺ to describe the low-temperature region t⬍0, one should per-
␦⫽ ⫽4.780共 2 兲 , 共46兲
1⫹ form an analytic continuation in the complex t plane.49,50 The
coexistence curve corresponds to a complex z 0 ⫽ 兩 z 0 兩 e ⫺i 
such that F(z 0 )⫽0. Therefore, the behavior near the coex-
 ⫽ 共 1⫹ 兲 ⫽0.3485共 2 兲 , 共47兲
2 istence curve is related to the behavior of F(z) in the neigh-
borhood of z 0 . The constants F ⬁0 and 兩 z 0 兩 can be expressed in
where the error on  has been estimated by considering the terms of universal amplitude ratios, by using the asymptotic
errors of and as independent. behavior of the magnetization along the critical isotherm and
at the coexistence curve:
IV. THE CRITICAL EQUATION OF STATE
共 C ⫹ 兲 (3 ␦ ⫺1)/2
A. General properties of the critical equation of state F ⬁0 ⫽ , 共55兲
of XY models 共 ␦ C c 兲 ␦ 共 ⫺C ⫹
4 兲
( ␦ ⫺1)/2
m3 f 共 x 兲 ⫽x ␥
兺
n⫽0
f ⬁n x ⫺2n  . 共59兲
⌬F⫽ A共 z 兲, 共51兲
g4
The coefficients f ⬁n can be expressed in terms of r 2n using
where m⫽1/ , is the second-moment correlation length, Eq. 共52兲. In particular, using Eqs. 共55兲 and 共56兲,
and g 4 is the zero-momentum four-point coupling. Note that
z⬀ 兩 M 兩 t ⫺  for t→0, thus the expansion of F̂( 兩 M 兩 t ⫺  ) for f ⬁0 ⫽R ⫺1 , 共60兲
兩 M 兩 t ⫺  →0 is equivalent to the small-z expansion of A(z): where R is defined in Appendix C. Griffiths’ analyticity
implies that f (x) is regular for x⬎⫺1.
1 1 1
A共 z 兲⫽ z 2⫹ z 4⫹
2 4! j⫽3 共 2兺 r z2 j.
j 兲! 2 j
共52兲 We want now to discuss the behavior of f (x) for x→
⫺1, i.e., at the coexistence curve. General arguments predict
that at the coexistence curve the transverse and longitudinal
Correspondingly, we obtain for the equation of state
magnetic susceptibilities behave respectively as
F共 M 兲 M ជ M M
ជ⫽
H ⬀ t ␦F共 z 兲, 共53兲 T⫽ , L⫽ ⬀H ⫺1/2. 共61兲
Mជ 兩M兩 H H
214503-13
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
TABLE XIII. Estimates of g 4 , r 6 , r 8 , and r 10 obtained from the analysis of the HT series for the two
improved Hamiltonians. Final results will be reported in Table XIV.
g4 r6 r8 r 10
In particular the singularity of L for t⬍0 and H→0 is 共51兲 and 共52兲. Indeed, these quantities can be expressed in
governed by the zero-temperature infrared-stable Gaussian terms of zero-momentum 2 j-correlation functions and of the
fixed point,51–53 leading to the prediction correlation length.
Details of the analysis of the HT series of g 4 , r 6 , r 8 , and
f 共 x 兲 ⬃c f 共 1⫹x 兲 2 for x→⫺1. 共62兲 r 10 are reported in Appendix B 3. We obtained the results
shown in Table XIII. In Table XIV we report our final esti-
The nature of the corrections to the behavior 共62兲 is less mates 共denoted by IHT兲, obtained by combining the results
clear. It has been conjectured,53–55 using essentially of the two models; we also compare them with the estimates
⑀ -expansion arguments, that, for y→0, i.e., near the coexist- obtained using other approaches. Note that our final estimate
ence curve, v ⬅1⫹x has a double expansion in powers of of g 4 is slightly larger than the result reported in Ref. 28 共see
y⬅HM ⫺ ␦ and y (d⫺2)/2. This implies that in three dimen- Table XIV兲. The difference is essentially due to the different
sions f (x) can be expanded in powers of v at the coexistence analysis employed here, which should be more reliable. This
curve. On the other hand, an explicit calculation56 to next-to- point is further discussed in Appendix B 3.
leading order in the 1/N expansion shows the presence of
logarithms in the asymptotic expansion of f (x) for x→⫺1. C. Parametric representations of the equation of state
However, they are suppressed by an additional factor of v 2 In order to obtain a representation of the equation of state
compared to the leading behavior 共62兲. that is valid in the whole critical region, we need to extend
It should be noted that for the transition in 4 He the analytically the expansion 共52兲 to the low-temperature region
order parameter is related to the complex quantum amplitude t⬍0. For this purpose, we use parametric representations
of helium atoms. Therefore, the ‘‘magnetic’’ field is not ex- that implement the expected scaling and analytic properties.
perimentally accessible, and the function appearing in Eq. They can be obtained by writing61–63
共57兲 cannot be measured directly in experiments. The physi-
cally interesting quantities are universal amplitude ratios of M ⫽m 0 R  m 共 兲 ,
quantities formally defined at zero external field, such as
U 0 ⬅A ⫹ /A ⫺ , for which accurate experimental estimates t⫽R 共 1⫺ 2 兲 ,
have been obtained. On the other hand, the scaling equation
of state 共57兲 is physically relevant for planar ferromagnetic H⫽h 0 R  ␦ h 共 兲 , 共63兲
systems.
where h 0 and m 0 are normalization constants. The variable R
is nonnegative and measures the distance from the critical
B. Small-M expansion of the equation of state in the high-
point in the (t,H) plane, while the variable parametrizes
temperature phase
the displacement along the lines of constant R. The functions
Using HT methods, it is possible to compute the first co- m( ) and h( ) are odd and regular at ⫽0 and at ⫽1. The
efficients g 2 j and r 2 j appearing in the expansion of the constants m 0 and h 0 can be chosen so that m( )⫽
Helmholtz free energy and of the equation of state, see Eqs. ⫹O( 3 ) and h( )⫽ ⫹O( 3 ). The smallest positive zero of
TABLE XIV. Estimates of g 4 , r 6 , r 8 , and r 10 obtained using the following methods: analyses of im-
proved HT expansions 共IHT兲, of HT expansions for the standard XY model 共HT兲, of fixed-dimension per-
turbative expansions (d⫽3 g-exp.兲, and of ⑀ expansions ( ⑀ -exp.兲. A more precise determination of r 10 will
be reported in Table XV.
214503-14
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
h( ), which should satisfy 0 ⬎1, corresponds to the coex- where is a free parameter.17,49 In the exact parametric
istence curve, i.e., to T⬍T c and H→0. The singular part of equation the value of may be chosen arbitrarily but, as we
the free energy is then given by shall see, when adopting an approximation procedure the de-
pendence on is not eliminated. In our approximation
⌬F⫽h 0 m 0 R 2⫺ ␣ g 共 兲 , 共64兲 scheme we will fix to ensure the presence of the Goldstone
where g( ) is the solution of the first-order differential equa- singularities at the coexistence curve, i.e., the asymptotic be-
tion havior 共68兲. Since z⫽ ⫹O( 3 ), expanding m( ) and
h( ) in 共odd兲 powers of ,
共 1⫺ 2 兲 g ⬘ 共 兲 ⫹2 共 2⫺ ␣ 兲 g 共 兲 ⫽ 关共 1⫺ 2 兲 m ⬘ 共 兲
⫹2  m 共 兲兴 h 共 兲 共65兲 m共 兲⫽⫹ 兺
n⫽1
m 2n⫹1 2n⫹1 ,
that is regular at ⫽1. In particular, the ratio A ⫹ /A ⫺ of the
specific-heat amplitudes in the two phases can be derived by
using the relation h共 兲⫽⫹ 兺
n⫽1
h 2n⫹1 2n⫹1 , 共71兲
g共 0 兲
A ⫹ /A ⫺ ⫽ 共 20 ⫺1 兲 2⫺ ␣ . 共66兲 and using Eqs. 共69兲 and 共70兲, one can find the relations
g共 0兲 among , m 2n⫹1 , h 2n⫹1 and the coefficients r 2n of the ex-
The parametric representation satisfies the requirements of pansion of A(z).
regularity of the equation of state. Singularities can appear Following Ref. 28, we construct approximate polynomial
only at the coexistence curve 共due, for example, to the loga- parametric representations that have the expected singular
rithms discussed in Ref. 56兲, i.e., for ⫽ 0 . Notice that the behavior at the coexistence curve51–53,56 共Goldstone singular-
mapping 共63兲 is not invertible when its Jacobian vanishes, ity兲 and match the known small-z expansion 共52兲. We will
which occurs when not repeat here in full the discussion of Ref. 28, which
should be consulted for more details. We consider two dis-
Y 共 兲 ⬅ 共 1⫺ 2 兲 m ⬘ 共 兲 ⫹2  m 共 兲 ⫽0. 共67兲 tinct schemes of approximation. In the first one, which we
denote by 共A兲, h( ) is a polynomial of fifth order with a
Thus, parametric representations based on the mapping 共63兲
double zero at 0 , and m( ) a polynomial of order (1
are acceptable only if 0 ⬍ l where l is the smallest posi-
⫹2n):
tive zero of the function Y ( ). One may easily verify that the
冉 冊
asymptotic behavior 共62兲 is reproduced simply by requiring n
that scheme 共 A兲 : m 共 兲 ⫽ 1⫹ 兺 c i 2i
i⫽1
,
h共 兲⬃共 0⫺ 兲 2
for → 0 . 共68兲
The functions m( ) and h( ) are related to F(z) by h 共 兲 ⫽ 共 1⫺ 2 / 20 兲 2 . 共72兲
214503-15
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
FIG. 2. The scaling function F(z). FIG. 3. The scaling function f (x). The MC curve is taken from
冉 冊
Ref. 64.
n
h 共 兲 ⫽ 共 1⫺ 2
/ 20 兲 2 1⫹ 兺 c i 2i
i⫽1
. 共73兲
quires an analytic continuation, which turns out to be effec-
Here h( ) is a polynomial of order 5⫹2n with a double tively performed by the approximate parametric representa-
zero at 0 . Note that for scheme 共B兲 tions we have considered.
Larger differences between the approximations given by
Y 共 兲 ⫽1⫺ 2 ⫹2  2 , 共74兲 schemes 共A兲 and 共B兲 for n⫽1 appear in the scaling function
⫺1 f (x), which is shown in Fig. 3, especially in the region x
independently of n, so that l ⫽(1⫺2  ) . Concerning
scheme 共A兲, we note that the analyticity of the thermody- ⬍0, which corresponds to t⬍0 and z imaginary. Note that
namic quantities for 兩 兩 ⬍ 0 requires the polynomial func- the sizeable differences for x⬎0 are essentially caused by
tion Y ( ) not to have complex zeroes closer to the origin the normalization of f (x), which is performed at the coexist-
than 0 . ence curve x⫽⫺1 and at the critical point x⫽0, by requir-
In both schemes the parameter is fixed by the require- ing f (⫺1)⫽0 and f (0)⫽1. Although the large-x region
ment 共68兲, while 0 and the n coefficients c i are determined corresponds to small values of z, the difference between the
by matching the small-z expansion of F(z). This means that, two approximate schemes does not decrease in the large-x
for both schemes, in order to fix the n coefficients c i we need limit due to their slightly different estimates of R 共see Table
to know n⫹1 values of r 2 j , i.e., r 6 , . . . ,r 6⫹2n . As input XV兲. Indeed, f (x)⬃R ⫺1 x ␥ for large values of x. In Fig. 3 we
parameters for our analysis we consider the estimates ob- also plot the curve obtained in Ref. 64 by fitting the MC
tained in this paper, i.e., ␣ ⫽⫺0.0146(8), ⫽0.0380(4), data.
r 6 ⫽1.950(15), r 8 ⫽1.44(10), r 10⫽⫺13(7). In Table XV we report the results for some universal ra-
Before presenting our results, we mention that the equa- tios of amplitudes. The notations are explained in Appendix
tion of state has been recently studied by MC simulations of C. The reported errors refer only to the uncertainty of the
the standard XY model, obtaining a fairly accurate determi- input parameters and do not include the systematic error of
nation of the scaling function f (x).64 In particular we men- the procedure, which may be determined by comparing the
tion the precise result obtained for the universal amplitude results of the various approximations. Comparing the results
ratio R 共see Appendix C for its definition兲, i.e., R for R and c f with the MC estimates of Ref. 64, we observe
⫽1.356(4), and for the constant c f , i.e., c f ⫽2.85(7), where that the parametrization 共A兲 is in better agreement with the
c f is defined in Eq. 共62兲. In the following we will take into numerical data. This is also evident from Fig. 3.
account these results to find the best parametrization within We also consider both schemes with n⫽2. If we use r 10
our schemes 共A兲 and 共B兲. to determine the next coefficient c 2 , scheme 共A兲 is not par-
By using the few known coefficients r 2 j —essentially r 6 ticularly useful because it is very sensitive to r 10 , whose
and r 8 because the estimate of r 10 is not very precise—one estimate has a relatively large error.28 This fact was already
obtains reasonably precise approximations of the scaling observed in Ref. 28, and explained by considerations on the
function F(z) for all positive values of z, i.e., for the whole more complicated analytic structure. One may instead deter-
HT phase up to t⫽0. In Fig. 2 we show the curves obtained mine c 2 by using the MC result R ⫽1.356(4). The estimates
in schemes 共A兲 and 共B兲 with n⫽1 that use the coefficients r 6 of the universal amplitude ratios obtained in this way are
and r 8 . The two approximations of F(z) are practically in- presented in Table XV. They are very close to the n⫽1 case,
distinguishable. This fact is not trivial since the small-z ex- providing additional support to our estimates and error bars.
pansion has a finite convergence radius that is expected to be Scheme 共B兲 is less sensitive to r 10 and provides reasonable
兩 z 0 兩 ⫽(R ⫹
4 ) ⬇2.7. Therefore, the determination of F(z) on
1/2
results if we use r 10 to fix the coefficient c 2 in h( ) and
the whole positive real axis from its small-z expansion re- impose the consistency condition 0 ⬍ l . The results are
214503-16
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
1.062共4兲 1.056共4兲 共Ref. 66兲 1.029共13兲 共Ref. 67兲 1.0442 共Refs. 1,3,4兲
1.055共3兲 共Ref. 28兲 1.067共3兲 共Ref. 39兲
1.058共4兲 共Ref. 40兲
1.088共7兲 共Ref. 68兲
shown in Table XV, where one observes that they get closer V. THE TWO-POINT FUNCTION OF THE ORDER
to the estimates obtained by using scheme 共A兲. PARAMETER IN THE HIGH-TEMPERATURE
As already mentioned, the most interesting quantity is the PHASE
specific-heat amplitude ratio A ⫹ /A ⫺ , because its estimate
The critical behavior of the two-point correlation function
can be compared with experimental results. Our results for
G(x) of the order parameter is relevant to the description of
A ⫹ /A ⫺ are quite stable and insensitive to the different ap-
scattering phenomena with light and neutron sources.
proximations of the equation of state we have considered,
In the HT critical region, the two-point function G(x)
essentially because they are obtained from the function g( ),
shows a universal scaling behavior. For k,m→0 (m⬅1/
which is not very sensitive to the local behavior of the equa-
and is the second-moment correlation length兲 with y
tion of state, see Eq. 共65兲. From Table XV we obtain the
⬅k 2 /m 2 fixed, we can write71
estimate
g 共 y 兲 ⫽ /G̃ 共 k 兲 . 共77兲
A ⫹ /A ⫺ ⫽1.062共 4 兲 . 共75兲
The function g(y) has a regular expansion in powers of y:
In Table XVI we compare our result 共denoted by IHT-PR兲 ⬁
with other available estimates. Note that there is a marginal g 共 y 兲 ⫽1⫹y⫹ 兺 c iy i. 共78兲
disagreement with the result of Ref. 28, i.e., A ⫹ /A ⫺ i⫽2
⫽1.055(3), which was obtained using the same method but Two other quantities characterize the low-momentum behav-
with different input parameters: ␣ ⫽⫺0.012 85(38) 共the ex- ior of g(y): they are given by the critical limit of the ratios
perimental estimate of Ref. 1兲, ⫽0.0381(3), r 6 ⫽1.96(2),
r 8 ⫽1.40(15) and r 10⫽⫺13(7). This discrepancy is mainly S M ⬅m gap
2
/m 2 , S Z ⬅ m 2 /Z gap , 共79兲
due to the different value of ␣ , since the ratio A ⫹ /A ⫺ is
particularly sensitive to it. This fact is also suggested by the where m gap 共the mass gap of the theory兲 and Z gap determine
phenomenological relation65 A ⫹ /A ⫺ ⬇1⫺4 ␣ . the long-distance behavior of the two-point function:
We observe a discrepancy with the experimental result Z gap ⫺m 兩 x 兩
reported in Refs. 1 and 3, A ⫹ /A ⫺ ⫽1.0442. However, we G共 x 兲⬇ e gap . 共80兲
note that in the analysis of the experimental data of Ref. 1 4兩x兩
the estimate of A ⫹ /A ⫺ was strongly correlated to that of ␣ ; If y 0 is the negative zero of g(y) that is closest to the origin,
indeed A ⫹ /A ⫺ was obtained by analyzing the high- and low- then, in the critical limit, S M ⫽⫺y 0 and S Z
temperature data with ␣ fixed to the value obtained from the ⫽ g(y)/ y 兩 y⫽y 0 .
low-temperature data alone. Therefore the discrepancy for The coefficients c i can be related to the critical limit of
A ⫹ /A ⫺ that we observe is again a direct consequence of the appropriate dimensionless ratios of spherical moments of
differences in the estimates of ␣ . G(x) and can be computed by analyzing the corresponding
As suggested in Ref. 69, one may consider the quantity HT series in the 4 and in the dd-XY models, which we have
calculated to 20th order.17,47 We report only our final esti-
1⫺A ⫹ /A ⫺ mates of c 2 and c 3 , i.e., c 2 ⫽⫺3.99(4)⫻10⫺4 , c 3
R ␣⫽ , 共76兲 ⫽0.09(1)⫻10⫺4 , and the bound 兩 c 4 兩 ⬍10⫺6 . As already ob-
␣
served in Ref. 47, the coefficients show the pattern
which is expected to be much less sensitive to the value of ␣ .
兩 c i 兩 Ⰶ 兩 c i⫺1 兩 Ⰶ•••Ⰶ 兩 c 2 兩 Ⰶ1 for i⭓3. 共81兲
Our analysis leads to the estimate R ␣ ⫽4.3(2), which com-
pares very well with the experimental estimate R ␣ ⬇4.19 that Therefore, a few terms of the expansion of g(y) in powers of
can be obtained from Refs. 1,3 and with the field-theoretical y provide a good approximation in a relatively large region
result reported in Ref. 70, i.e., R ␣ ⫽4.39(26). around y⫽0, larger than 兩 y 兩 ⱗ1. This is in agreement with
For the other universal amplitude ratios we quote as our the theoretical expectation that the singularity of g(y) closest
final estimates the results obtained by using scheme 共A兲 with to the origin is the three-particle cut 共see, e.g., Refs.
n⫽1: R ⫹ ⫽0.355(3), R c ⫽0.127(6), R ⫽1.35(7), R 4 47,72,73兲. If this is the case, the convergence radius r g of the
⫽7.5(2), F ⬁0 ⫽0.0302(3), c f ⫽4(2). These results are sub- Taylor expansion of g(y) is r g ⫽9S M . Since, as we shall
stantially equivalent to those reported in Ref. 28. see, S M ⬇1, at least asymptotically we should have c i⫹1 ⬇
214503-17
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
⫺ 19 ci . This behavior can be checked explicitly in the large-N small gain in performance was found with respect to the
limit of the N-vector model.47 single-cluster algorithm in simulations of the three-
Assuming the pattern 共81兲, we may estimate S M and S Z dimensional Ising model.
from c 2 , c 3 , and c 4 . We obtain Note that, since the cluster update does not change the
modulus of the field, identical routines can be used for the
S M ⫽1⫹c 2 ⫺c 3 ⫹c 4 ⫹2c 22 ⫹••• 共82兲 4 model and for the dd-XY model.
For the 4 model we sweep through the lattice with a
S Z ⫽1⫺2c 2 ⫹3c 3 ⫺4c 4 ⫺2c 22 ⫹•••, 共83兲 local updating scheme. At each site we perform a Metropolis
step, followed by an over-relaxation step and by a second
where the ellipses indicate contributions that are negligible Metropolis step. In particular, the over-relaxation step is
with respect to c 4 . In Ref. 47 the relation 共82兲 has been given by
confirmed by a direct analysis of the HT series of S M . From
Eqs. 共82兲 and 共83兲 we obtain S M ⫽0.999 592(6) and S Z ជ x•
共 ជ n兲
ជn
⫽1.000 825(15). These results improve those obtained in ជ x⬘ ⫽
ជ x ⫺2 , 共A1兲
Ref. 47 by using HT methods in the standard XY model and ជn
2
g 共 y 兲 ⫺1 ⫽
y
A1
1⫺ /2 冉 1⫹
y
A2
(1⫺ ␣ )/(2 )
⫹
y
A3
1/(2 ) 冊 . 共84兲
The local update of the dd-XY model is achieved by per-
forming at each site one Metropolis update followed by the
over-relaxation update 共A1兲. In the Metropolis update, the
proposal for the field ជ x at the site x is given by
The coefficients have been computed in the ⑀ expansion to
three loops,73 obtaining A 1 ⬇0.92, A 2 ⬇1.8, A 3 ⬇⫺2.7. In ជ ⬘x ⫽ 共 0,0兲 for 兩
ជ x 兩 ⫽1
order to obtain an interpolation that is valid for all values of
y, we will use a phenomenological function proposed by
Bray.73 This approximation has the correct large- and small- ជ x⬘ ⫽ 关 cos共 ␣ 兲 ,sin共 ␣ 兲兴 for 兩
ជ x 兩 ⫽0, 共A2兲
y behavior. It requires the values of the exponents , ␣ , and where ␣ is a random number with a uniform distribution in
, and the sum of the coefficients A 2 ⫹A 3 . For the exponents 关 0,2 ). One can prove that this Metropolis update leaves the
we use of course the estimates obtained in this paper, while Boltzmann distribution invariant.
the coefficient A 2 ⫹A 3 is fixed using the ⑀ -expansion predic- We summarize the complete update cycle: local update
tion A 2 ⫹A 3 ⫽⫺0.9. Bray’s phenomenological function pre- sweep; global field rotation, in which the angle is taken from
dicts then the constants A i and c i . We obtain: A 1 ⬇0.915, a uniform distribution in 关 0,2 ); 6 wall-cluster updates. The
A 2 ⬇⫺24.7, A 3 ⬇23.8, c 2 ⬇⫺4.4⫻10⫺4 , c 3 ⬇1.1⫻10⫺5 , sequence of the 6 wall-cluster updates is given by the wall in
c 4 ⬇⫺5⫻10⫺7 . The results for A 1 , c 2 , c 3 , and c 4 are in 1-2, 1-3 and 2-3 plane. In each of the three cases, we update
good agreement with the above-reported estimates, while A 2 separately each component of the field.
and A 3 differ significantly from the ⑀ -expansion results. No- We used ANSI C to implement our simulation programs.
tice, however, that, since 兩 ␣ 兩 is very small, the relevant quan- We used our own implementation of the G05CAF random-
tity in Eq. 共84兲 is the sum A 2 ⫹A 3 which is, by construction, number generator from the NAG-library. The G05CAF is a
equal in Bray’s approximation and in the ⑀ expansion. linear congruential random-number generator with modulus
m⫽2 59, multiplier a⫽1313 and increment c⫽0. Most of our
APPENDIX A: THE MONTE CARLO SIMULATION simulations were performed on 450-MHz Pentium III PC’s
running the Linux operating system.
1. The Monte Carlo algorithm
Concerning the efficiency of the Monte Carlo algorithm,
At present the best algorithm to simulate N-vector sys- we only mention that for the 4 model at ⫽2.1 and 
tems is the cluster algorithm proposed by Wolff75 共see Ref. ⫽0.509 15 and the dd-XY model at D⫽1.03 and 
76 for a general discussion兲. However, the cluster update ⫽0.5628, the integrated autocorrelation times 共in units of
changes only the angle of the field. Therefore, following update cycles兲 of the magnetic susceptibility slightly increase
Brower and Tamayo,77 we add a local update that changes with increasing L, and they are ⬇4 for the largest lattices.
also the modulus of the field.
We use the embedding algorithm proposed by Wolff75 2. Measuring Z a ÕZ p
with two major differences. First, we do not choose an arbi-
trary direction, but we consider changes of the signs of the One of the phenomenological couplings that we have
first and of the second component of the fields separately. studied is the ratio Z a /Z p of the partition function Z a of a
Second, we do not use the single-cluster algorithm to update system with antiperiodic boundary conditions 共a.b.c.兲 in one
the embedded model, but the wall-cluster variant proposed in direction and the partition Z p with periodic boundary condi-
Ref. 23. In the wall-cluster update, one flips at the same time tions 共p.b.c.兲 in all three directions. The a.b.c. are obtained
all clusters that intersect a plane of the lattice. In Ref. 23 a by multiplying the term ជ x
ជ y in the Hamiltonian by ⫺1 for
214503-18
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
all x⫽(L 1 ,x 2 ,x 3 ) and y⫽(1,x 2 ,x 3 ). This ratio can be ob- 0.510, 0.515, and 0.520. We computed R̄ with R 1,f
tained using the so-called boundary-flip algorithm, applied in ⫽(Z a /Z p ) f ⫽0.3202 and R 1,f ⫽( 2nd /L) f ⫽0.5925 and R 2
Ref. 45 to the Ising model and generalized in Ref. 46 to ⫽U 4 and R 2 ⫽U 6 for all these simulations, using a third-
general O(N)-invariant nonlinear models. order Taylor expansion. The results show that there is a large
In the boundary-flip algorithm, one considers fluctuating interval in which the method works: indeed, the results for R̄
boundary conditions, i.e., a model with partition function for  s ⫽0.505, 0.510 and 0.515 agree within two standard
Z fluct⫽Z p ⫹Z a ⫽ 兺
J b ⫽⫾1
冕 冋兺
D 关 兴 exp 
具 xy 典
ជ x
J 具 xy 典 ជy
deviations, although the variation of U 4 and U 6 at  s is
several hundred standard deviations. In addition, we have
gained information about the range of  where the extrapo-
⫹••• , 册 共A3兲
lation works with the desired accuracy:
Za
2 j⫽ 兺
x 2 , . . . ,x 2 j
具 ␣ 1 共 0 兲 ␣ 1 共 x 2 兲 ••• ␣ j 共 x 2 j⫺1 兲 ␣ j 共 x 2 j 兲 典 c
⫽具b典, 共A6兲 共B2兲
Zp
where the expectation value is taken with p.b.c. ( ⫽ 2 ). More precisely, we computed 4 to 18th order, 6
to 17th order, 8 to 16th order, and 10 to 15th order. The
series for the 4 Hamiltonian with ⫽2.07 and the dd-XY
3. Checks of the program
model with D⫽1.02 are reported in Tables XVII and XVIII.
The properties of the integration measure allow to derive The HT series of the zero-momentum four-point coupling g 4
an infinite set of nontrivial Schwinger-Dyson equations and of the coefficients r 2 j that parametrize the equation of
among observables of the model. We have used two such state can be computed using their definitions in terms of 2 j
equations to test the correctness of the programs and the and 2 , i.e.,
reliability of the random-number generator. For a more gen-
eral discussion of such tests, see Ref. 81. 3N 4
As a test of the MC program and of the analysis software, g 4 ⫽⫺ , 共B3兲
N⫹2 2 3
we simulated the 4 model for L⫽4 and ⫽2.1 at the fol-
lowing values of  :  ⫽0.485, 0.490, 0.495, 0.500, 0.505, and
214503-19
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
i 2 2 4
0 0 0.82195468340525626553069 ⫺0.18234682673209145113698
1 1.01341425235775258955047 2.02682850471550517910095 ⫺1.79856993883835218414950
2 4.99788354573054604609481 4.39836023278442865137831 ⫺10.1455394390643744635266
3 17.0521428795050699809477 9.45608303384639636168661 ⫺44.4379274874486579169902
4 50.2802232377961742954821 19.7509483421835061143003 ⫺166.911247402827967481247
5 136.081955345771888598704 41.0953224517675210993058 ⫺566.231817578861031037728
6 349.014986520228913452360 84.3784681474012731739965 ⫺1783.81074566131415934408
7 861.072204516234675501406 172.831419974338990927678 ⫺5314.73425458572561425498
8 2065.11115559310738635122 351.412467687273478895617 ⫺15153.8054671243945401658
9 4843.65801296852863958594 713.327141375620774069366 ⫺41706.6393990107185189143
10 11163.1410843891753269547 1441.29537543678062498269 ⫺111480.350225534613432415
11 25357.2828059634753398090 2908.62198285250042113961 ⫺290779.553330813045449802
12 56914.8160479537305800002 5851.22700855642135604880 ⫺742792.316757751368697227
13 126448.957588634753901037 11759.8313973078166535503 ⫺1863674.70918289087066236
14 278504.794338270260511244 23580.6969253332410696568 ⫺4603313.91229513488724197
15 608775.044038619834901124 47248.9211902315365277418 ⫺11214943.2818348822933596
16 1321948.27688188944339553 94508.3475046413360663308 ⫺26991439.1725436301683188
17 2853823.94933643220583008 188924.811397083165660961 ⫺64258568.1525362515312215
18 6128960.82003386821732524 377150.816225492759953104 ⫺151492730.104215500664052
19 13101467.5362982920867821 752534.725866450821199491
20 27889129.6761627014637264 1499898.13514730628043402
i 6 8 10
0 0.41197816889670188853628 ⫺2.04244608438484921752518 17.4966181373390399118217
1 8.09031485009238719234751 ⫺65.5324533503954629382215 827.907846284122870554505
2 80.5579033480687931078788 ⫺993.099544110386544944672 17583.1558423294442666385
3 569.040955700888943670354 ⫺10125.9430216092445934264 242162.737636025061958065
4 3235.75380473965447264355 ⫺79870.4410385967305759821 2508010.66864972545742050
5 15823.8724636340283846359 ⫺524859.679867495864300189 21148750.0733186453907577
6 69189.8243270623365490494 ⫺3005237.58408599956662949 152451568.855786252573448
7 277430.904691458150896928 ⫺15443601.8321215251366505 970471959.824139222337835
8 1038008.04844006612521747 ⫺72717890.9115560213984880 5582186363.05878798571648
9 3669720.02487068188513696 ⫺318510657.462768013289640 29507970388.4631661994252
10 12374182.7602063550019508 ⫺1312668767.17043587616989 145203489229.461390006077
11 40084338.4384102309965122 ⫺5135470803.97668705323439 671866031736.491009927361
12 125446404.887628912440242 ⫺19206423870.3058981804587 2946715041148.83369711661
13 381003987.146313729964559 ⫺69057827061.9724837730205 12330196249913.8306690405
14 1127156412.88952101681812 ⫺239824993302.080594778804 49488831955055.0830738902
15 3257906807.50792443987963 ⫺807540440278.527392223891 191378645936343.575972510
16 9223412391.97191941389912 ⫺2645013087720.90303565763
17 25631282620.7774958190658
5 共 N⫹2 兲 6 2 The formulas relevant for the XY universality class are ob-
r 6 ⫽10⫺ , 共B4兲 tained by setting N⫽2.
3 共 N⫹4 兲 24
2. Critical exponents
280共 N⫹2 兲 6 2 35共 N⫹2 兲 2 8 22
r 8 ⫽280⫺ ⫹ , In order to determine the critical exponents ␥ and from
3 共 N⫹4 兲 24 9 共 N⫹4 兲共 N⫹6 兲 34
the HT series of and 2 /  respectively, we used quasidi-
agonal first-, second-, and third-order integral approximants
7700共 N⫹2 兲 6 2 350共 N⫹2 兲 2 26 22
r 10⫽15 400⫺ ⫹ 共IA1’s, IA2’s, and IA3’s, respectively兲. Since the most pre-
共 N⫹4 兲 24 共 N⫹4 兲 2 44 cise results are obtained by using the MC estimates of  c to
1400共 N⫹2 兲 2 8 22 bias the approximants, we shall report only the results of the
⫹ biased analyses. We used the values of  c obtained in Sec.
3 共 N⫹4 兲共 N⫹6 兲 34 II D, i.e., those reported in Table IV,  c (⫽2.0)
35共 N⫹2 兲 3 10 32 ⫽0.509 904 9(15),  c (⫽2.1)⫽0.509 150 7(13) for the 4
⫺ . model, and  c (D⫽1.03)⫽0.562 797 5(14) for the dd-XY
3 共 N⫹4 兲共 N⫹6 兲共 N⫹8 兲 44 model 共see also Ref. 6兲.
214503-20
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
i 2 2 4
0 0 0.73497259946651881066155 ⫺0.12952381667498436104661
1 0.81027708294985783008938 1.62055416589971566017876 ⫺1.14235747481325709236407
2 3.57318872366283058263222 3.19240289872507821851086 ⫺5.86566452750871250322137
3 11.0006482367650529136739 6.24412164584056694746990 ⫺23.4964691362678107734860
4 29.3868020693268986850123 11.8078991080110710061038 ⫺80.6562687454557047131467
5 72.0537292098720051000559 22.2452167397051997713027 ⫺249.761461812242731292855
6 167.391323293772704647731 41.3083108337880277068933 ⫺717.220908333053642342640
7 373.956365385479749391430 76.5360051001560916440577 ⫺1945.68221342894368675090
8 811.915341705911027278243 140.679660885119763690958 ⫺5046.15791369304901147071
9 1723.52034635335232533650 258.178251744079668880881 ⫺12622.3575801936426660898
10 3594.34333565606104441219 471.478755943757026593209 ⫺30642.5601039304090869577
11 7386.64345245316356464020 860.010504587591344158158 ⫺72548.9796473949259402584
12 14997.4679478666032080397 1563.48753262138558216805 ⫺168135.618026803729643010
13 30136.9726456923668836816 2839.86987807993137143111 ⫺382565.790192788611922604
14 60029.1887398932204011553 5145.84503638874703629194 ⫺856629.943580595646246982
15 118656.316956262327168223 9317.67206591240832596126 ⫺1891351.24076972180490224
16 232979.699867454031093529 16841.0660076130187494565 ⫺4124166.11180668899412336
17 454746.664171304150538747 30421.5573167338465805825 ⫺8893532.37374656560556900
18 882960.924794534410812953 54875.4729390613106530869 ⫺18987953.9690154439591383
19 1706330.67007276458833100 98938.9870168970865371838
20 3283569.77023650242548276 178182.095750601905570976
i 6 8 10
where the functions P i (  ) and R(  ) are polynomials of or- where p k (  ) is a polynomial of order m k ⫺1.
der m i and l, respectively, which are determined by the 共ii兲 Since on bipartite lattices  ⫽⫺  c is also a singular
214503-21
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
P k 共  兲 ⫽ 共 1⫺  2 /  2c 兲 p k 共  2 兲 , 共B8兲 In the following we fix q⫽3 for the IA1’s and q⫽2 for
the IA2’s and IA3’s.
where p k (  ) is a polynomial in  2 of order m k ⫺1. We will For each set of IAk’s we calculate the average of the
denote them by bFCIAk’s. values corresponding to all nondefective IAk’s listed above.
In our analyses we consider diagonal or quasidiagonal Approximants are considered defective when they have sin-
approximants, since they are expected to give the most ac- gularities close to the real  axis near the critical point. More
curate results. Below, we give the rules we used to select the precisely, we consider those approximants defective that
quasidiagonal approximants. We introduce a parameter q have singularities in the rectangle
that determines the degree of off-diagonality allowed 共see
below兲. In order to check the stability of the results with x min⭐Re  /  c ⭐x max , 兩 Im  /  c 兩 ⭐y max . 共B16兲
respect to the order of the series, we also perform analyses in
which we average over the results obtained with series of The values of x min , x max , and y max are fixed essentially by
different length. For this purpose, we introduce a parameter p stability criteria, and may differ in the various analyses. One
and perform analyses in which we use all approximants ob- should always check that the results depend very little on the
tained from series of n̄ terms with n⭓n̄⭓n⫺p. chosen values of x min , x max , and y max , by varying them
within a reasonable and rather wide range of values. The
We consider the following sets of IAk’s:
domain 共B16兲 cannot be too large, otherwise only few ap-
共a兲 关 m 1 /m 0 /k 兴 bIA1’s with
proximants are left. In this case the analysis would be less
n⭓m 1 ⫹m 0 ⫹k⫹1⭓n⫺p, robust and therefore less reliable. We introduce a parameter s
such that
Max关 b 共 n⫺1 兲 /3c ⫺q,3 兴 ⭐m 1 ,m 0 ,k⭐ d 共 n⫺1 兲 /3e ⫹q.
共B9兲 x min⫽1⫺s, x max⫽1⫹s, y max⫽s. 共B17兲
共b兲 关 m 1 /m 0 /k 兴 b⫾ IA1’s with We obtain results for various values of s, checking their de-
pendence on s. We also discard some nondefective IA’s—we
n⭓m 1 ⫹m 0 ⫹k⭓n⫺ p, call them outliers—whose results are far from the average of
the other approximants. Such approximants are eliminated
Max关 b 共 n⫺1 兲 /3c ⫺q,3 兴 ⭐m 1 ,m 0 ,k⭐ d 共 n⫺1 兲 /3e ⫹q. algorithmically: first, we compute the average A and the
共B10兲 standard deviation of the results using all nondefective
共c兲 关 m 1 /m 0 /k 兴 bFCIA1’s with IA’s. Then, we discard those IA’s whose results differ by
more than n from A with n ⫽2. We repeat the procedure
n⭓m 1 ⫹m 0 ⫹k⫹1⭓n⫺p, on the remaining IA’s, by calculating the new A and , but
now eliminating the IA’s whose results differ by more than
Max关 b 共 n⫺1 兲 /3c ⫺q,3 兴 ⭐m 1 ,m 0 ,k⭐ d 共 n⫺1 兲 /3e ⫹q. n with n ⫽3. The procedure is again repeated, increasing
共B11兲 n by one at each step. This procedure converges rapidly
共d兲 关 m 2 /m 1 /m 0 /k 兴 bIA2’s with and, as we shall see, the outliers so determined are always a
very small part of the selected nondefective IA’s.
n⭓m 2 ⫹m 1 ⫹m 0 ⫹k⫹3⭓n⫺p, In the Tables XIX and XX, we present the results for the
critical exponents ␥ and respectively, obtained from the
Max关 b 共 n⫺3 兲 /4c ⫺q,2 兴 ⭐m 2 ⫺1,m 1 ,m 0 ,k⭐ d 共 n⫺3 兲 /4e ⫹q. HT analysis of the 4 and dd-XY models. There we also
共B12兲 quote the ‘‘approximant ratio’’ r a ⬅(g⫺ f )/t, where t is the
共e兲 关 m 2 /m 1 /m 0 /k 兴 b ⫾ IA2’s with total number of approximants in the given set, g is the num-
ber of nondefective approximants, and f is the number of
n⭓m 2 ⫹m 1 ⫹m 0 ⫹k⫹2⭓n⫺p, outliers that are discarded using the above-presented algo-
rithm; g⫺ f is the number of ‘‘good’’ approximants used in
Max关 b 共 n⫺3 兲 /4c ⫺q,2 兴 ⭐m 2 ⫺2,m 1 ,m 0 ,k⭐ d 共 n⫺3 兲 /4e ⫹q. the analysis; notice that gⰇ f , and g⫺ f is never too small.
共B13兲 For each analysis, beside the corresponding estimate, we re-
214503-22
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
TABLE XIX. Results for ␥ obtained from the analysis of the 20th-order HT series of for the 4 and dd-
XY models. The number n of terms used in the analysis is indicated explicitly when it is smaller than the
number of available terms (n⫽20). p⫽0 when its value is not explicitly given.
Approximants ra ␥
⫽2.00 bIA1s⫽1/2 (28⫺2)/48 1.31755共11兲关19兴
bIA2s⫽1/2 (82⫺5)/115 1.31749共9兲关17兴
⫽2.07 bIA1s⫽1/2 (28⫺2)/48 1.31785共10兲关29兴
bIA1p⫽3,s⫽1/2 (103⫺1)/172 1.31766共21兲关24兴
b⫾ IA1s⫽1/2 (21⫺1)/48 1.31789共22兲关28兴
bFCIA1s⫽1/2 (35⫺4)/48 1.31775共9兲关28兴
bIA2s⫽1/8 (99⫺7)/115 1.31780共9兲关27兴
bIA2s⫽1/4 (93⫺4)/115 1.31780共9兲关27兴
bIA2s⫽1/2 (87⫺4)/115 1.31780共8兲关28兴
bIA2s⫽1 (60⫺2)/115 1.31781共7兲关27兴
bIA2n⫽19,s⫽1/2 (48⫺6)/70 1.31777共10兲关28兴
bIA2n⫽18,s⫽1/2 (53⫺4)/62 1.31768共9兲关28兴
bIA2p⫽3,s⫽1/2 (277⫺18)/345 1.31772共14兲关25兴
b⫾ IA2s⫽1/2 (46⫺3)/100 1.31781共29兲关23兴
bFCIA2s⫽1/2 (91⫺2)/140 1.31780共11兲关29兴
bIA3s⫽1/2 (56⫺4)/61 1.31787共8兲关31兴
⫽2.10 bIA1s⫽1/2 (29⫺2)/48 1.31777共9兲关16兴
bIA2s⫽1/2 (92⫺2)/115 1.31773共6兲关15兴
bIA2p⫽3,s⫽1/2 (295⫺17)/345 1.31769共10兲关14兴
b⫾ IA2s⫽1/2 (49⫺5)/100 1.31774共20兲关15兴
bFCIA2s⫽1/2 (92⫺5)/140 1.31772共15兲关17兴
⫽2.20 bIA1s⫽1/2 (31⫺3)/48 1.31809共7兲关30兴
bIA2s⫽1/2 (94⫺6)/115 1.31807共3兲关27兴
port two numbers. The number in parentheses, e 1 , is basi- same type, obtained imposing different constraints. The
cally the spread of the approximants for  c fixed at the MC number in brackets, e 2 , is related to the uncertainty on the
estimate. It is the standard deviation of the results obtained value of  c and it is estimated by varying  c in the range
from all ‘‘good’’ IA’s divided by the square root of r a , i.e., 关  c ⫺⌬  c ,  c ⫹⌬  c 兴 .
e 1 ⫽ / 冑r a . Such a definition of e 1 is useful to compare re- The results of the analyses are quite stable: all sets of IA’s
sults obtained from different subsets of approximants of the give substantially consistent results. The comparison of the
214503-23
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
TABLE XX. Results for obtained from the analysis of the 19th-order HT series of 2 /  for the 4 and
the dd-XY models. The number n of terms used in the analysis is indicated explicitly when it is smaller than
the number of available terms (n⫽19). p⫽0 when its value is not explicitly given.
Approximants ra
results obtained using all available terms of the series with series. Therefore, we do not need to perform problematic
those using less terms 共in the Tables the number of terms is extrapolations in the number of terms, or rely on phenom-
indicated explicitly when it is smaller than the number of enological arguments, typically based on other models, sug-
available terms兲 and those obtained for p⫽3 共i.e., using gesting when the number of terms is sufficient to provide a
n,n⫺1,n⫺2, and n⫺3 terms in the series兲 shows that the reliable estimate.
results are also stable with respect to the order of the HT From the intermediate results reported in Tables XIX, and
214503-24
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
XX 共which, we stress, are determined algorithmically once TABLE XXI. Results for obtained using the CPRM: 共a兲 ap-
chosen the set of IAk’s兲, we obtain the estimates of ␥ and . plied to 2 /  and 共19 orders兲; 共b兲 applied to 2 and 共20 orders兲.
From the analyses for the 4 Hamiltonian at ⫽2.07, we
obtain Approximants ra
␥ ⫽1.317 80共 10兲关 28兴 ⫹0.003共 ⫺2.07兲 , 共B18兲 ⫽2.00 共a兲 bIA1s⫽1/2 33/37 0.02547共7兲
共a兲 bIA2s⫽1/2 (47⫺1)/70 0.0256共2兲
⫽0.671 61共 5 兲关 12兴 ⫹0.002共 ⫺2.07兲 . 共B19兲 共b兲 bIA2s⫽1/2 (99⫺9)/115 0.0251共3兲
⫽2.07 共a兲 bIA1s⫽1/2 37/37 0.02555共7兲
As before, the number between parentheses is basically the 共a兲 bIA2s⫽1/2 47/70 0.0257共2兲
spread of the approximants at ⫽2.07 using the central 共a兲 bIA3s⫽1/2 (20⫺1)/34 0.0255共2兲
value of  c , while the number between brackets gives the 共b兲 bIA2s⫽1/2 (96⫺8)/115 0.0252共3兲
systematic error due to the uncertainty on  c . Eqs. 共B18兲
共b兲 bIA3s⫽1/2 (51⫺2)/61 0.0253共5兲
and 共B19兲 show also the dependence of the results on the
⫽2.20 共a兲 bIA1s⫽1/2 33/37 0.02573共7兲
chosen value of . The coefficient is estimated from the re-
共a兲 bIA2s⫽1/2 (49⫺2)/70 0.0259共3兲
sults for ⫽2.2 and ⫽2.0, i.e., from the ratio 关 Q(⫽2.2)
共b兲 bIA2s⫽1/2 (95⫺11)/115 0.0252共3兲
⫺Q(⫽2.0) 兴 /0.2, where Q represents the quantity at hand.
Using * ⫽2.07(5), we obtain finally ␥ ⫽1.317 80(10) 关 28兴 D⫽0.90 共a兲 bIA2s⫽1/2 (45⫺1)/70 0.0252共3兲
兵 15其 and ⫽0.671 61(5) 关 12兴 兵 10其 , where the error due to the 共b兲 bIA2s⫽1/2 (84⫺3)/115 0.0248共9兲
uncertainty on * is reported between braces. D⫽1.02 共a兲 bIA1s⫽1/2 (22⫺1)/37 0.0256共9兲
Since for ⫽2.10 a more precise estimate of  c is avail- 共a兲 bIA2s⫽1/2 (37⫺1)/70 0.0257共3兲
able, it is interesting to perform the same analysis, using the 共a兲 bIA3s⫽1/2 (23⫺2)/34 0.0252共5兲
HT series of the 4 model at ⫽2.10. We obtain 共b兲 bIA2s⫽1/2 (93⫺3)/115 0.0252共8兲
共b兲 bIA3s⫽1/2 (59⫺3)/61 0.0253共4兲
␥ ⫽1.317 73共 10兲关 15兴 ⫹0.003共 ⫺2.10兲 , 共B20兲
D⫽1.20 共a兲 bIA2s⫽1/2 (33⫺2)/70 0.0263共3兲
⫽0.671 60共 5 兲关 8 兴 ⫹0.002共 ⫺2.10兲 , 共B21兲 共b兲 bIA2s⫽1/2 (96⫺4)/115 0.0259共8兲
214503-25
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
TABLE XXII. Results for g 4 , obtained from the analysis of the 17th-order series of  ⫺3/2g 4 (  ), for
⫽2.07 in the 4 model and D⫽1.02 in the dd-XY model.
Approximants ra g4
spread of the IA’s and the second one to the uncertainty on 共B10兲兴 and impose various constraints on the value of by
* , evaluated as before. Analogously, for the dd-XY model selecting bIA1’s with larger than a given non-negative
we find value.
In Table XXII we report the results obtained for g 4 using
⫽0.025 50共 40兲 ⫹0.004共 D⫺1.02兲 , 共B27兲 different sets of approximants. In this case the variation due
and therefore, using D * ⫽1.02(3), ⫽0.025 50(40) 兵 12其 , to the uncertainty of  c is negligible. Therefore, we report
where again the first error is related to the spread of the IA’s, only the average of the results of the ‘‘good’’ IA1’s and their
while the second one is related to the uncertainty on D * . standard deviation 共divided by 冑r a ) calculated at  c . In
Table XXII we also report the value of obtained from the
3. Amplitude ratios
selected IA1’s. The comparison of the results for different
values of and D shows that the errors due to uncertainty on
In the following we describe the analysis method we em- * and D * are small and negligible.
ployed to evaluate zero-momentum renormalized couplings, From the results of Table XXII we derive the estimates
such as g 4 and r 2 j . In the case of g 4 we analyzed the series g 4 ⫽21.15(6) and g 4 ⫽21.13(7), respectively for the 4
 3/2g 4 ⫽ 兺 i⫽0
17
a i i. Hamiltonian and the dd-XY model. We note that these re-
Consider an amplitude ratio A which, for t⬅  c /  ⫺1 sults are slightly larger than the estimates reported in Ref.
→0, behaves as 28. The difference is essentially due to the different analysis
employed. There, the analysis was based on Padé 共PA兲,
A 共 t 兲 ⫽A * ⫹c 1 t ⌬ ⫹c 2 t ⌬ 2 ⫹•••. 共B28兲 Dlog-Padé 共DPA兲 and IA1’s, selecting those without singu-
In order to determine A * from the HT series of A(t), we larities in a neighborhood of  c and evaluating them at  c .
consider biased IA1’s, whose behavior at  c is given by 共see, However, by analyzing the longer series that are now avail-
e.g., Ref. 10兲 able for the Ising model,83 we have realized that such proce-
dure is not very accurate and that the analyses using bIA1’s
IA1⬇ f 共  兲共 1⫺  /  c 兲 ⫹g 共  兲 , 共B29兲 are more reliable when a sufficiently large number of terms
is available. Moreover, when the series is sufficiently long,
where f (  ) and g(  ) are regular at  c , except when is a most 共and eventually all兲 PA’s, DPA’s and IA1’s become
non-negative integer. In particular, defective. Indeed, the functions we are considering do have
singularities at  c , although with a positive exponent.
P 0共  c 兲 R共 c兲 In the analysis of r 2 j , we also consider PA’s and DPA’s.
⫽ , g 共  c 兲 ⫽⫺ . 共B30兲
P 1⬘ 共  c 兲 P 0共  c 兲 We indeed expect that, when the series is not sufficiently
long to be asymptotic, the approximants obtained by biasing
In the case we are considering, is positive and therefore, the singularity at  c may not provide a robust analysis. For
g(  c ) provides an estimate of A * . Moreover, for improved comparison, we also use quasi-diagonal Padé approximants
Hamiltonians we expect ⫽⌬ 2 ⬇2⌬ and ⌬⬇0.5. In our 共PA’s兲 and Dlog-Padé approximants 共DPA’s兲, evaluating
analyses we consider bIA1’s and b⫾ IA1⬘s 关see Eqs. 共B9兲 and them at  c . For r 6 and r 8 the above PA’s and DPA’s give
214503-26
CRITICAL BEHAVIOR OF THE THREE-DIMENSIONAL . . . PHYSICAL REVIEW B 63 214503
214503-27
MASSIMO CAMPOSTRINI et al. PHYSICAL REVIEW B 63 214503
214503-28