DESY 03-198 SFB/CPP-03-56 December 2003 4 0 Scaling test for Wilson twisted mass QCD 0 2 n a χ F Collaboration L J 6 Karl Jansena, Andrea Shindlera, 1 Carsten Urbacha,b and Ines Wetzorkea 2 v 3 1 a NIC/DESY Zeuthen 0 Platanenallee 6, D-15738 Zeuthen, Germany 2 1 3 0 b Institut fu¨r Theoretische Physik, Freie Universita¨t Berlin / t Arnimallee 14, D-14195 Berlin, Germany a l - p e h Abstract : v We presentafirstscaling test of twisted mass QCDwithpureWil- i X son quarks for a twisting angle of π/2. We have computed the vector r a meson mass mV and the pseudoscalar decay constant FPS for differ- ent values of β at fixed value of r m . The results obtained in the 0 PS quenched approximation are compared with data for pure Wilson and non-perturbatively O(a) improved Wilson computations. We show that our results from Wilson twisted mass QCD show clearly reduced lattice spacing errors, consistent with O(a) improvement and without the need of any improvement terms added. These results thus provide numerical evidence of the prediction in ref. [1]. 1 1 Introduction FormulatingQCDonaspace-timelatticeadmitsasubstantialamountoffree- dom in discretising the continuous derivative with restrictions coming from obeying principles such asgaugeinvariance, localityandunitarity. Naturally, as long as universality holds, all these formulations should provide consistent results in the limit when the discretisation is removed. The standard Wilson formulation of lattice QCD [2] is a simple realisation of such a discretised version of QCD and has been used for a long time in lattice simulations. However, it has been realised that this formulation has a number of severe problems: it shows large discretisation effects [3] that are linear in the lat- tice spacing a, it violates chiral symmetry strongly and develops unphysical small eigenvalues of the corresponding lattice Wilson-Dirac operator, even at rather large values of the quark mass. The problem of discretisation effects can be overcome following the Sy- manzik improvement program [4, 5], introducing the well known clover term [6]togetfull,on-shellO(a)improvedresults, iftheimprovement isperformed non-perturbatively [3]. In this approach then also the chiral properties are improved, although chiral symmetry breaking and O(a2) lattice spacing ef- fects are still left in the theory and have to be extrapolated away. Despite the fact that non-perturbatively improved Wilson fermions clearly diminish discretisation errors, they, unfortunately, show the same – if not worse – problem of the appearance of small eigenvalues of the lattice Dirac operator [7]. In order to solve the problem of small eigenvalues, it has been proposed in [8] to use the so-called twisted mass formulation of QCD. In this approach the mass term in the Dirac operator is chirally twisted [8], see also [9]. When using such a chirally twisted lattice action in combination with the clover term and a non-perturbatively tuned value of the improvement coefficient, the theory is O(a) improved and the corresponding lattice Dirac operator is safe against developing small eigenvalues. Recently, it has been realised [1] that all the above properties of non- perturbatively improved, twisted mass QCD can also be obtained when the clover term is completely omitted. If a special value of the twisting an- gle is chosen one theoretically obtains O(a) improved results without adding improvement terms. At the same time, the lattice Dirac operator is still protected against the appearance of small eigenvalues by construction. This curious observation receives a special importance for simulations with dy- namical quarks: In [10] it was found that dynamical fermion simulations with non-perturbatively improved Wilson fermions show signs of first order 2 phase transitions which render simulations very difficult and induce large cut-off effects. Although by varying the form of the gauge actions [11] the first order phase transition seems to vanish, it is unclear, whether eventually such phe- nomena will reappear. The problem of the presence of the first order phase transition may be related to the fact that the clover term in the fermion action generates an adjoint gauge action. Therefore, if O(a) improvement can be achieved without the clover term, as anticipated in [1], the problems connected to such phase transitions should be completely eliminated. As a consequence, the potential of twisted mass fermions in general may be very large. Since the small eigenvalues are regulated by the twisted mass pa- rameter, simulations at much lower quark masses than used today could be performed with promising, but hard to estimate, advantages to explore the chiral limit of lattice QCD. In this paper we provide a first test of the conjecture of O(a) improve- ment of Wilson twisted mass QCD in quenched numerical simulations. For this purpose we performed a scaling test of the vector meson mass m and V the pseudoscalar decay constant F at a fixed value of the physical pseu- PS doscalar mass m . We compare the results with those that have been ob- PS tained for standard Wilson fermions, see [12] and references therein, and non-perturbatively improved clover fermions [13]. 2 Twisted mass QCD with Wilson quarks In twisted mass QCD (tmQCD) as formulated in [8] the twisted mass action in the continuum reads as follows: S [ψ,ψ¯] = d4xψ¯(D γ +m +iµ γ τ3)ψ , (2-1) F µ µ 0 q 5 Z where D denotes the usual covariant derivative, m is the standard bare µ 0 quark mass, τ3 is the third Pauli matrix acting in flavour space and µ is the q twisted mass parameter, also referred to as the twisted mass. An axial transformation, ψ′ = exp(iωγ τ3/2)ψ, ψ¯′ = ψ¯exp(iωγ τ3/2) , (2-2) 5 5 witha realrotationangleω leaves theformoftheactioninvariant andmerely changes the mass parameters into m′ and µ′, 0 q m′ = m cos(ω)+µ sin(ω) 0 0 q µ′ = −m sin(ω)+µ cos(ω) . (2-3) q 0 q 3 The standard action (µ′ = 0) is obtained by setting tanω = µ /m . Note q q 0 that τ3 is traceless and therefore the transformation (2-2) does not couple to the fermion determinant anomaly. In order to have an O(a) improved twisted mass lattice action, it appears to be natural to discretise the Dirac operator adding appropriate improve- ment terms to the standard Wilson-Dirac operator. Indeed, it has been shown in [14, 15, 16, 17] that by adding the usual clover term, full on-shell O(a) improvement can be obtained. In[1]ithasbeenrealisedlaterthatusingsimplythestandardlatticeDirac operator D one can obtain O(a) improved physical observables without W adding improvement terms. More precisely, it is possible to obtain O(a) improved lattice results by employing only the standard massless Wilson- Dirac operator, 1 D = {γ (∇∗ +∇ )−ar∇∗∇ } (2-4) W 2 µ µ µ µ µ under the condition that one averages over physical observables that are obtained from simulations at positive and negative values of the Wilson pa- rameter r. A less general, but similar suggestion was made by the authors of refs. [18, 19]. Instead of using positive and negative values of r, quark masses with different signs may be used: the bare quark mass m can be 0 written as m = m (r)+m , with m (−r) = −m (r) , (2-5) 0 c q c c where m (r) is the critical quark mass, and m is the subtracted bare quark c q mass. Averaging physical observables obtained from simulations at positive and negative subtracted bare quark masses, again O(a)-improvement is ob- tained. Let us shortly sketch the arguments leading to this surprising result. One first has to observe that with ψ → ψ′ = γ ψ 5 R : (2-6) 5 (ψ¯ → ψ¯′ = −ψ¯γ5 the following combined transformation sp R ≡ R ×[r → −r]×[m → −m ] (2-7) 5 5 q q is a so called spurionic symmetry of the ordinary Wilson action. Another symmetry of the Wilson (and Wilson tmQCD) action is R × D . In the 5 d continuum, thetransformationD hastheeffectofchangingthesignofallthe d 4 space-time coordinates and multiplies each local term L in the Lagrangian i density by the factor (−1)d[Li], where d[L ] is the naive dimension of L . The i i lattice version of this transformation is more involved and we refer to ref. [1] for details. Taking now the parity properties of multiplicatively renormalisable op- sp erators under R and R × D into account, one can show – using the 5 5 d Symanzik expansion – that one gets O(a) improvement when averaging over two simulations with positive and negative Wilson coefficient r (Wilson av- erage (WA)). In addition, from the spurionic symmetry Rsp of the Wilson 5 (and Wilson tm) action one can obtain O(a) improved physical observables when averaging, at a fixed value of r, over two simulations with positive and negative m , asdefined ineq. (2-5)(massaverage (MA)),taking intoaccount q this time the R -parity of the operators. Studying the chiral properties of 5 the scalar condensate with Wilson fermions, a similar suggestion was made by the authors in ref. [20]. In the special case of choosing ω = ±π/2, such an averaging procedure is done automatically. A change of the sign of r is equivalent to ω → ω + π. Hence, for ω = ±π/2 all the quantities that are even under ω → −ω are automatically improved without any averaging procedure. It is the aim of this paper to check this conjecture in practical simulations by performing a scaling test for the vector meson mass m and the pseu- V doscalar decay constant F at ω = π/2. The main goal is to test whether PS the results for m and F are consistent with the anticipated leading O(a2) V PS behaviour and that the linear a dependence is indeed cancelled. In addition, it is an interesting and important question, what the size of the remaining lattice spacing effects arising in O(a2) will be. Let us list a few properties of the composite fields in the twisted mass formulation before going to the numerical results. Due to the transformation rule (2-2) one also has to transform the composite fields defined in the usual way, τα S0(x) = ψ¯(x)ψ(x), Pα(x) = ψ¯(x)γ ψ(x), 5 2 τα τα Aα(x) = ψ¯(x)γ γ ψ(x), Vα(x) = ψ¯(x)γ ψ(x) . (2-8) µ µ 5 2 µ µ 2 As an example we give here the relations for the axial and vector currents inthe“physicalbasis”(primedquantities) and“thetwisted basis”(unprimed quantities), cos(ω)Aα +ǫ3αβ sin(ω)Vβ (α = 1,2), A′α = µ µ (2-9) µ (A3µ (α = 3), 5 cos(ω)Vα +ǫ3αβ sin(ω)Aβ (α = 1,2), V′α = µ µ (2-10) µ (Vµ3 (α = 3). Note that in eq. (2-9) for α = 1,2 and ω = π/2 the role of the axial and vector currents are just interchanged. Of particular interest is the PCVC relation, which takes the following form in the twisted basis: ∂∗Vα = −2µ ǫ3αβPβ , (2-11) µ µ q where ∂∗ is the usual backward derivative. Through a vector variation of the µ action one obtains the point-split vector current as defined in [8, 16]. This current is protected against renormalisation and using the point-split vector current, the PCVC relation is an exact lattice identity. This implies that Z = Z−1, where Z is the renormalisation constant for the twisted mass P µ µ µ . This will become important in the extraction of the pseudoscalar decay q constant F as described below. PS 3 Numerical Tests In this section, we describe our numerical results for testing the scaling behaviour of Wilson tmQCD in the quenched approximation. We started our investigation by performing a comparative benchmark study of different solvers for obtaining the quark propagator. We found the CGS algorithm [21] to be superior to the BiCGstab and the CG algorithms. We therefore used the CGS algorithm throughout this work. Gauge field configurations were generated by standard heat-bath and over-relaxation techniques. 3.1 Mass average In order to test the predictions of ref. [1], we started with the mass average procedure. To this end, we selected a value of β = 5.85, set the Wilson parameter r = 1 and performed simulations on 123 ×24 lattices at positive and negative values of m = 1/2 (1/κ − 1/κ ). While for m = +0.02725, q c q the propagator computations went smoothly, for m = −0.02725 the com- q putation of the quark propagator was exceedingly expensive. The reason for this behaviour can be traced back to the spectrum of the Wilson tmQCD operator as can be seen in fig. 1. Comparing fig. 1(a) with fig. 1(b), in the case of negative m one has to q deal with extremely small eigenvalues of the operator (D +m )†(D +m ). W 0 W 0 6 10−04 10−05 x 10−06 a m λ / λj 10−07 10−08 10−09 0 2 4 6 8 10 12 14 16 t MC (a) m =+0.02725,κ=0.16025 q 10−04 10−05 x 10−06 a m λ / λj 10−07 10−08 10−09 0 2 4 6 8 10 12 14 16 t MC (b) m =−0.02725,κ=0.163099 q Figure 1: Monte Carlo time evolution of the eleven smallest eigenvalues λ of (D + W m0)†(DW +m0), normalised by the largest eigenvalue, at µq =0 and mq =±0.02725 on a 123×24 lattice (κ =0.161662(17),β=5.85). c 7 Clearly, these very small low-lying eigenvalues lead to a poor convergence of the solver employed and hence to very costly simulations. Projecting out these small modes does also not help in this respect since the computation of the eigenvalues is again costly. We therefore proceeded to the “self-averaging” case of choosing ω = π/2, for which it has been shown in [1] that one gets O(a) improvement even without the need of any averaging for all quantities that are even under ω → −ω. In our practical implementation we have used the twisted basis. Hence, the choice ω = π/2 corresponds to set m = 0 and µ 6= 0. In this q q situation, thecorresponding Wilson-Diracoperatorisprotectedagainstsmall eigenvalues and we do not expect difficulties with the simulations. 3.2 Scaling of the vector meson mass Inordertoverifythepredictionofref.[1]wecomputedthevectormesonmass m and the pseudoscalar decay constant F for the following values of β: V PS 5.85, 6.0, 6.1, 6.2. We used periodicboundaryconditions. The corresponding lattice volumes were 143 ×28, 163 ×32, 203×40 and 243 ×48, respectively. For our simulation in the twisted basis at m = m we had to determine 0 c the critical hopping parameter κ for each value of β. At all the β values of c our simulations we made our own determination of the value of κ from the c intercept in κ at zero pion mass. The values of κ are given in table 1. Note c that these critical values of κ have an intrinsic uncertainty of O(a). This is, however, sufficient for obtaining fully O(a) improved results in Wilson tmQCD [1]. β L T κ µ N c q meas 5.85 14 28 0.161662(17) 0.0376 400 6.0 16 32 0.156911(35) 0.03 388 6.1 20 40 0.154876(10) 0.025854 299 6.2 24 48 0.153199(16) 0.021649 215 Table 1: Parametersof the simulations. Note that the values for κc are obtained from a different set of measurements. In order to fix the physical situation in our scaling test, we kept r m 0 PS fixed for all values of β. For this purpose we determined the value of µ to q fix r m = 1.79. The corresponding values of µ for each value of β and all 0 PS q our simulation parameters are given in table 1. We computed the standard 2-point correlation functions at zero momen- tum for the pseudoscalar and axial operators (which in the twisted basis at 8 ω = π/2 gives the correct operator to extract the vector meson mass), fα(t) = hPα(x)Pα(0)i (3-1) P ~x X 3 1 fα(t) = hAα(x)Aα(0)i (3-2) A 3 i i i=1 ~x XX with Pα(x) and Aα(x) given in eqs. (2-8) and x = (~x,t). In order to obtain a non-vanishing result, theflavour index hastobethesame inthese correlation functions and we will choose α = 1 in the following. The pion mass could be extracted easily from the exponential decay of the correlation function fα(t). For the vector meson mass, we performed two mass fits for the ground P state mass and the first excited state. We checked the stability of the fit by changing the value of t where the fit started. As a cross check, we also min determined the effective ground state mass and found consistent results. All errors were computed by a jackknife analysis. The numerical results at our simulation points are collected in table 2. β r /a am am aF 0 PS V PS 5.85 4.067 0.4340(16) 0.656(11) 0.1147(11) 6.0 5.368 0.3329(21) 0.488(11) 0.0859(9) 6.1 6.324 0.2871(17) 0.427(9) 0.0717(8) 6.2 7.360 0.2438(16) 0.363(10) 0.0640(10) Table 2: Results for the vector meson mass and the pseudoscalar decay constant. In fig. 2 we show our results for the vector meson mass as a function of a2 represented by the open circles. In addition, we also show results from non-perturbatively O(a) improved Wilson fermions (filled circles) [13]. Finally, we added results for standard Wilson fermion simulations, see [12] and references therein, as they were available in the literature. We remark that the published data were not always at exactly the same value of r m 0 PS that we used for our Wilson tmQCD simulations. In such cases we performed an interpolation to the desired value of r m . The error from this (small) 0 PS interpolation is negligible for the results presented here. We performed a simple extrapolation of the Wilson tmQCD and the O(a) improved data of the form r m = r mcont+b·(a/r )2, with r mcont the con- 0 V 0 V 0 0 V tinuum value of the vector meson mass and r ≃ 0.5fm. For the pure Wilson 0 results we replaced the quadratic term with a term proportional to (a/r ) in 0 the extrapolation. Let us remark that our data for m for Wilson tmQCD V has about a factor of four less statistics than the data from O(a)-improved 9 3.0 2.8 2.6 V m 0 r 2.4 2.2 2.0 0.00 0.02 0.04 0.06 0.08 (a/r )2 0 Figure 2: Scalingbehaviourofthevectormesonmassasafunctionofthe latticespacing squared for fixed pion mass, r0mPS = 1.79. Open circles denote Wilson tmQCD, while filledcirclesarefromnon-perturbativelyimprovedWilsonfermions[13]. The Wilsondata without improvement (open squares) are collected from several sources in the literature. Note that the filled circles are slightly displaced for better visibility. Wilson fermions, which is reflected in the larger error bars. Nevertheless, it is evident that the Wilson tmQCD results show a very similar scaling be- haviour as the O(a)-improved Wilson fermions. This becomes even clearer when we compare with the unimproved pure Wilson data for m that we V show as open squares in fig. 2. Here large lattice artefacts are seen and the scaling behaviour is much worse than with Wilson tmQCD or O(a)-improved Wilson fermions. We also note that the data for Wilson tmQCD are rather flat as a function of a2 indicating that also higher order lattice spacing effects are suppressed. Clearly, it would be desirable to test these promising results in more precise simulations using a much higher statistics. 10