Stochastic method with low mode substitution for nucleon isovector matrix elements
Yi-Bo Yang, Andrei Alexandru, Terrence Draper, Ming Gong, Keh-Fei Liu
SStochastic method with low mode substitution for nucleon isovector matrix elements
Yi-Bo Yang , Andrei Alexandru , Terrence Draper , Ming Gong , and Keh-Fei Liu ( χ QCD Collaboration) Department of Physics and Astronomy,University of Kentucky, Lexington, KY 40506, USA Department of Physics,The George Washington University,Washington, DC 20052, USA Institute of High Energy Physics and Theoretical Physics Center for Science Facilities,Chinese Academy of Sciences, Beijing 100049, China
We introduce a stochastic method with low-mode substitution to evaluate the connected three-point functions. The isovector matrix elements of the nucleon for the axial-vector coupling g A ,scalar couplings g S and the quark momentum fraction (cid:104) x (cid:105) u − d are calculated with overlap fermionon 2+1 flavor domain-wall configurations on a 24 ×
64 lattice at m π = 330 MeV with lattice spacing a = 0 .
114 fm.
PACS numbers: 11.15.Ha, 12.38.Gc, 12.39.Mk
I. INTRODUCTION
The proton isovector-axial coupling g A and quark mo-mentum fraction (cid:104) x (cid:105) u − d are important benchmarks tocheck whether the systematic uncertainties of latticeQCD simulation, such as finite lattice spacing, finite vol-ume, and chiral extrapolation, are under control, by acorrect reproduction of the corresponding experimentalresults. Since the noisy disconnected insertion contribu-tion to the isovector part of the nuclear matrix element iscanceled between two degenerate flavors, the values areobtained solely from the connected insertion and thus arerelatively cheaper to compute with high precision to beconsidered as benchmarks.Most attempts have resulted in values ∼
10% below theexperimental number for the axial-vector coupling [1–8],while a few claim that their results could be consistentwith experiment [9–12]. For the quark momentum frac-tion (cid:104) x (cid:105) u − d , overestimation by ∼
20 – 30% is common inmost of the calculations [3, 7, 14–16] except [8].Recently, attention has been paid to lattice QCD cal-culation of the isovector scalar matrix element g S in theproton [2, 11, 17, 18] due to its role in constraining pos-sible scalar interactions at the TeV scale [19].In this work, we calculate the isovector matrix elementsof the nucleon for the axial-vector and scalar couplingsand the quark momentum fraction with the valence over-lap fermion on 2 + 1 flavor domain-wall fermion (DWF)configurations [20]. Compared to simulations with otheractions, the overlap fermion provides the best control of the systematic errors since it is free of explicit chiral sym-metry breaking and gives small O ( a ) errors, whereas thenumerical work is more costly.In order to improve SNR, the 8-grid smeared Z noisesource with low-mode substitution (LMS) [23–27] hasbeen applied to the hadron two point correlator on the24 ×
64 lattice [28] which improves the error of the nu-cleon mass of a point source by a factor of 7 and thatof the 8-grid source without smearing by a factor of 2.5.In this work, we use a stochastic sandwich contractionmethod to remove the need of multiple inversions in thesink-sequential approach and use the current-sequentialmethod for the low modes in the propagator between thecurrent and the sink. This is an extension of the noisegrid smeared source with LMS to the three point func-tion. Such a many-to-all correlator with LMS is usefulwhen the low-eigenmode contributions are important inthe relevant time windows where the physical quantitiesare extracted.The structure of the rest of the paper is organized asfollows. The LMS technique with noise grid source forthe non-zero momentum case of the two point correla-tion function is provided in Sec. II. Sec. III discusses thepossibility of applying LMS on all the four quark propa-gators in the proton three-point function. The numericaldetails are provided in Sec. IV. In Sec. V, the results ofisovector matrix elements of the nucleon for the axial-vector g A , the scalar coupling g S and the quark momen-tum fraction (cid:104) x (cid:105) u − d are provided. A short summary andoutlook are presented in Sec. VI. a r X i v : . [ h e p - l a t ] J a n II. LOW MODE SUBSTITUTION WITH MIXEDMOMENTUM GRID SOURCE
Let’s first consider the nucleon two-point function(2pt) with the interpolation field of the nucleon [29], χ α ( x ) = (cid:15) abc ψ ( u ) aα ( x ) ψ ( u ) bβ ( x )( ˜ C ) βγ ψ ( d ) cγ ( x ) χ α (cid:48) ( x ) = − (cid:15) a (cid:48) b (cid:48) c (cid:48) ψ ( d ) c (cid:48) γ (cid:48) ( ˜ C ) γ (cid:48) β (cid:48) ψ ( u ) b (cid:48) β (cid:48) ( x ) ψ ( u ) a (cid:48) α (cid:48) ( x ) , (1)where ˜ C ≡ Cγ = γ γ γ in the Pauli-Sakurai gamma-matrix convention, used throughout this work. Thereare two kinds of the Wick contractions so the 2pt of thenucleon can be constructed in terms of the point-to-pointquark propagator S as C ( y, x ; Γ; S ( u ) , S ( d ) , S ( u ) ) = (cid:104) (cid:15) abc (cid:15) a (cid:48) b (cid:48) c (cid:48) Tr (cid:16) Γ S ( u ) aa (cid:48) ( y, x ) (cid:17) Tr (cid:16) S ( d ) bb (cid:48) ( y, x ) S ( u ) cc (cid:48) ( y, x ) (cid:17) (cid:105)−(cid:104) (cid:15) abc (cid:15) a (cid:48) b (cid:48) c (cid:48) Tr (cid:16) Γ S ( u ) ab (cid:48) ( y, x ) S ( d ) ba (cid:48) ( y, x ) S ( u ) cc (cid:48) ( y, x ) (cid:17) (cid:105) = (cid:104) (cid:15) abc (cid:15) a (cid:48) b (cid:48) c (cid:48) Tr (cid:16) Γ S ( u ) aa (cid:48) ( y, x ) (cid:17) Tr (cid:16) S ( d ) bb (cid:48) ( y, x ) S ( u ) cc (cid:48) ( y, x ) (cid:17) +Tr (cid:16) Γ S ( u ) aa (cid:48) ( y, x ) S ( d ) bb (cid:48) ( y, x ) S ( u ) cc (cid:48) ( y, x ) (cid:17) (cid:105) (2)where S is defined as ( ˜ CS ˜ C − ) T and Γ is the projectionoperator for the nucleon polarization.The quark propagator S in the above equation is theinverse of the operator ( D c + m ) [30, 31], where D c isdefined in terms of the overlap operator and is chiral,i.e. { D c , γ } = 0 [32]. The details will be discussed inSec. IV. As in Ref. [21, 28], we use the low lying eigen-values and eigenvectors of the overlap fermion, λ i and | i (cid:105) , satisfying D c | i (cid:105) = λ i | i (cid:105) to speed up the inversion andseparate the propagator into its low-mode and high-modeparts, S L ( y, x ) = (cid:88) | λ i | <(cid:15) c λ i + m | i (cid:105) y (cid:104) i | x ,S H ( y, x ) = S ( y, x ) − (cid:88) | λ i | <(cid:15) c λ i + m | i (cid:105) y (cid:104) i | x , (3) with (cid:15) c as the upper bound of the modulus of the eigen-values.The idea of using the Z noise grid source is to tiethe sources of the three quark propagators stochasticallyto each point (or a smeared point) on the grid so thatone can have a multi-to-all correlator from one inversion.LMS for the quark propagator with Z noise grid source(PropNG), be it point-grid (PG) [21] or smeared grid(SG) [28], has been used to improve the SNR for the nu-cleon correlator with significant success. This techniqueremoves the gauge non-invariant contributions of the low-mode contributions (defined below) from the cases inwhich three propagators are from different source sites,and restores the benefit of using PropNG.To construct the nucleon correlation function withLMS, PropNG S NG ( y ) should be split into its high-modeand low-mode pieces S NG ( y ) = (cid:88) x ∈ G θ ( x ) S ( y, x )= S HNG ( y ) + (cid:88) x ∈ G θ ( x ) S L ( y, x ) , (4)with S HNG ( y ) = (cid:80) x ∈ G θ ( x ) S H ( y, x ) and random Z phases θ ( x ) ∈ { , e i π , e − i π } for each point on a grid G .As in Ref. [28], we can expand the nucleon correlationfunction C ( y, x ; Γ; S ( u ) NG , S ( d ) NG , S ( u ) NG ) with the decomposi-tion in Eq. (4) (ignoring the indices for the sink position y and the projection matrix Γ), C LMS (cid:0) S NG , S NG , S NG (cid:1) == C ( S HNG , S
HNG , S
HNG ) + (cid:88) x ∈ G C (cid:0) θ ( x ) S L ( x ) , θ ( x ) S L ( x ) , θ ( x ) S L ( x ) (cid:1) + C (cid:0) (cid:88) x ∈ G θ ( x ) S L ( x ) , S HNG , S
HNG (cid:1) + C (cid:0) S HNG , (cid:88) x ∈ G θ ( x ) S L ( x ) , S HNG (cid:1) + C (cid:0) S HNG , S
HNG , (cid:88) x ∈ G θ ( x ) S L ( x ) (cid:1) + (cid:88) x ∈ G C (cid:0) θ ( x ) S L ( x ) , θ ( x ) S L ( x ) , S HNG (cid:1) + (cid:88) x ∈ G C (cid:0) θ ( x ) S L ( x ) , S HNG , θ ( x ) S L ( x ) (cid:1) + (cid:88) x ∈ G C (cid:0) S HNG , θ ( x ) S L ( x ) , θ ( x ) S L ( x ) (cid:1) = C ker (cid:0) S HNG , (cid:88) x ∈ G θ ( x ) S L ( x ) (cid:1) + (cid:88) x ∈ G C ker (cid:0) θ ( x ) S L ( x ) , S HNG (cid:1) (5)where C ker ( S , S ) = C ( S , S , S ) + C ( S , S , S )+ C ( S , S , S ) + C ( S , S , S ) . (6)The nucleon correlator with LMS here can be obtainedfrom the one in Ref. [28] with just one more step. Thelow-mode propagator (cid:80) x ∈ G θ ( x ) S L ( y, x ) is decomposedinto several terms as in the very last term in the RHS ofEq. 5 to improve the SNR.After the noise averaging, the nucleon correlation func-tion with PropNG should be a stochastic estimate of thesum of nucleon correlators from each of the grid points,i.e. (cid:88) (cid:126)y C grid ( (cid:126)y ) = (cid:88) i (cid:88) (cid:126)y C ( (cid:126)y, (cid:126)w i ) , (7)where the grid points (cid:126)w i are (cid:126)w i ∈ ( x + m x ∆ x , y + m y ∆ y , z + m z ∆ z ) . (8)with m x,y,z = (0 , , · · · , L s / ∆ x,y,z ) modulo the periodicboundary condition in the spatial directions. In this gridpattern, in addition to the zero momentum mode (0,0,0),one can obtain non-zero momentum modes from the nu-cleon correlation function with PropNG. For example, forthe PropNG with a regular (∆ x = ∆ y = ∆ z = L s /m )grid, the momentum mode p = ( ± n m, ± n m, ± n m )( n , , are integers) can be obtained. In this case, thereis a phase factor which needs to be taken into accountwhen the origin w = ( x , y , z ) is changed from config-uration to configuration, (cid:88) (cid:126)y C grid ( (cid:126)y ) e − i πLs (cid:126)y · p = e − i nπLs w · p (cid:88) i (cid:88) (cid:126)y C ( (cid:126)y, (cid:126)w i ) e − i πLs ( (cid:126)y − (cid:126)w i ) · p − i mπLs ( (cid:126)w i − (cid:126)w ) · ( n ,n ,n ) = e − i nπLs w · p (cid:88) i (cid:88) (cid:126)y C ( (cid:126)y, (cid:126)w i ) e − i πLs ( (cid:126)y − (cid:126)w i ) · p (9)The exponential term in the second line with the expo-nent proportional to (cid:126)w i − (cid:126)w does not contribute, sinceall components of the latter are proportional to L s /m and, as a result, the exponent is a multiple of 2 π .In order to obtain the other momentum modes, propa-gators with noise grid non-zero momentum source (Prop-NGM) are required. To cover a range of p modes andminimize the effect of the rotation symmetry breakingdue to the finite lattice spacing and volume, three kindsof PropNGM S p ( y ) = (cid:88) i θ ( (cid:126)w i ) S ( (cid:126)y, (cid:126)w i ) e i πLs (cid:126)w i · (1 , , ,S p ( y ) = (cid:88) i θ ( (cid:126)w i ) S ( (cid:126)y, (cid:126)w i ) e i πLs (cid:126)w i · (0 , , ,S p ( y ) = (cid:88) i θ ( (cid:126)w i ) S ( (cid:126)y, (cid:126)w i ) e i πLs (cid:126)w i · (0 , , (10) and related inversions are required for the proton case.It is trivial to confirm that one can obtain a momentummode like (1,1,0) from the contraction C ( S p , S p , S NG ),and (1,1,1) from C ( S p , S p , S p ).To reduce the cost, we can combine these three kindsof PropNGM together as the mixed PropNGM, S p ≡ S p + S p + S p = (cid:88) i θ ( (cid:126)w i ) S ( (cid:126)y, (cid:126)w i )( e i πLs (cid:126)w i · (1 , , + e i πLs (cid:126)w i · (0 , , + e i πLs (cid:126)w i · (0 , , ) , (11)with the origin of the grid (cid:126)w = ( x , y , z ) to be selectedrandomly for each configuration. r e l a t i v e e rr o r p smeared pointsmeared grid, no lmssmeared grid, lms FIG. 1: The plot shows the relative error of 2pt as a functionof the momentum squared p at t=8 in lattice units. The datapoints of the smeared grid cases have been shifted a bit onthe abscissa to make it easier to distinguish them. The SNRof the case with the noise smeared grid source (red squares)and LMS applied is better than the one with smeared pointsource (blue dots), while the one with the noise smeared gridsource but no LMS (black triangles) is even worse than theone with smeared point source. Fig. 1 shows the SNR of the proton effective mass atthe unitary point where the pion mass due to the valencequark is the same as that from the sea, on the ensembleof which details will be addressed in Sec. IV. When LMSis applied, the SNR of the 2pt with the noise smearedgrid source propagators (PropNG and mixed PropNGM,∆ x = ∆ y = ∆ z = L s /
2) is 2.3 times smaller than thatof the of the smeared point source at p = 0. This is again of 5.3 in statistics which is very good consideringthat the maximum possible gain is 8 for the ideal casewhere the independent nucleon propagators emerge fromeach of the 8 smeared grid points. On the other hand,if we don’t use LMS, the SNR of 2pt with grid sourceis worse than the smeared point source, even though thelatter has only 1/8 of the statistics of the former. Thisis understood as due to the fact that the Parisi-Lepageestimate of the SNR for the nucleon is modified to C N ( t, (cid:126)p = 0) σ N ( t ) ≈ (cid:114) NV e − ( m N − / m π ) t , (12)where N is the product of the number of noise andthe number of gauge configurations and V is the three-volume of the noise with its support on a time slice. Inour case, V = 8. It is this extra factor of √ V whichmakes the SRN of the 2pt from the noise smeared gridsource without LMS worse than that of the smeared pointsource. When LMS is employed, the situation is reversedand one gains a statistical factor almost as large as the number of the grid points. Thus, it is essential to haveLMS when the noise grid source is used for the nucleon. III. LMS OF THE CONNECTED THREE-POINTCORRELATOR
Generally, a nucleon three point function (3pt), from x to y , with a current ¯ ψ ( x ) ( u ) O ( z ) ψ ( x ) ( u ) (with currentoperator O such as γ i , γ i D j , etc.) inserted at z , includesfour kinds of Wick contractions, C u ( y, x ; Γ; ˆ S ( u ) , S ( u ) , S ( d ) , S ( u ) ) = (cid:104) (cid:15) abc (cid:15) a (cid:48) b (cid:48) c (cid:48) Tr (cid:16) Γ S ( u ) ad ( y, z ) O ( z ) S ( u ) da (cid:48) ( z, x ) (cid:17) Tr (cid:16) S ( d ) bb (cid:48) ( y, x ) S ( u ) cc (cid:48) ( y, x ) (cid:17) (cid:105) + (cid:104) (cid:15) abc (cid:15) a (cid:48) b (cid:48) c (cid:48) Tr (cid:16) Γ S ( u ) ad ( y, z ) O ( z ) S ( u ) da (cid:48) ( z, x ) S ( d ) bb (cid:48) ( y, x ) S ( u ) cc (cid:48) ( y, x ) (cid:17) (cid:105) + (cid:104) (cid:15) abc (cid:15) a (cid:48) b (cid:48) c (cid:48) Tr (cid:16) Γ S ( u ) aa (cid:48) ( y, x ) (cid:17) Tr (cid:16) S ( d ) bb (cid:48) ( y, x ) S ( u ) cd ( y, z ) O ( z ) S ( u ) dc (cid:48) ( z, x ) (cid:17) (cid:105) + (cid:104) (cid:15) abc (cid:15) a (cid:48) b (cid:48) c (cid:48) Tr (cid:16) Γ S ( u ) aa (cid:48) ( y, x ) S ( d ) bb (cid:48) ( y, x ) S ( u ) cd ( y, z ) O ( z ) S ( u ) dc (cid:48) ( z, x ) (cid:17) (cid:105) (13)and can be expressed in terms of the 2pt correlation func-tion C ( y, x ; Γ; S ( u ) , S ( d ) , S ( u ) ) defined in Eq. (2), C u ( y, x ; Γ; ˆ S ( u ) , S ( u ) , S ( d ) , S ( u ) )= C ( y, x ; Γ; ˆ S ( u ) , S ( d ) , S ( u ) )+ C ( y, x ; Γ; S ( u ) , S ( d ) , ˆ S ( u ) ) , (14)where ˆ S ( O , z ; y, x ) ≡ (cid:80) (cid:126)z S ( y, z ) O ( z ) S ( z, x ) is the cur-rent inserted propagator (PropCI). Similarly, the 3ptwith a current of d quark can be expressed as C d ( y, x ; Γ; ˆ S ( d ) , S ( u ) , S ( d ) , S ( u ) )= C ( y, x ; Γ; S ( u ) , ˆ S ( d ) , S ( u ) ) . (15)Fig. 2 shows PropCI as the product of the propagatorsin the shadowed region.Supposing S ( u ) = S ( d ) = S , Eq. 14 can be rewritteninto the contraction of PropCI ˆ S and the remaining partsdenoted as X u,d (Γ , S , S ), C u (Γ; ˆ S , S, S, S ) = (cid:104) Tr (cid:0) ˆ S X u (Γ , S, S ) (cid:1) (cid:105) ,C d (Γ; ˆ S , S, S, S ) = (cid:104) Tr (cid:0) ˆ S X d (Γ , S, S ) (cid:1) (cid:105) , (16) FIG. 2: The quark diagram of the proton correlation functionwith the connected insertion, from x to y , with an insertionat z . The product of the propagators in the shadowed regionis the current inserted propagator, ˆ S . The propagator fromthe current z to the sink y is decomposed into its low- andhigh-mode contributions ( S L and S H respectively) for furtherSNR/cost improvement from the advanced technique in thelatter discussion. See Sec. III B for more details. with X aa (cid:48) u (Γ , S , S ) = (cid:15) abc (cid:15) a (cid:48) b (cid:48) c (cid:48) (cid:0) ΓTr[ S bb (cid:48) S cc (cid:48) ] + S bb (cid:48) S cc (cid:48) Γ+Tr[Γ S cc (cid:48) ] S bb (cid:48) + Γ S cc (cid:48) S bb (cid:48) (cid:1) ,X bb (cid:48) d (Γ , S , S ) = (cid:15) abc (cid:15) a (cid:48) b (cid:48) c (cid:48) (cid:0) Tr[Γ S aa (cid:48) ] ˜ C − ( S cc (cid:48) ) T ˜ C + ˜ C − ( S aa (cid:48) Γ S cc (cid:48) ) T ˜ C (cid:1) (17)Based on the above definition, a typical 3pt correlationfunction for a point source on the t = 0 time slice, whensummed over the spatial indices of y and z becomes C ( t , t ) = (cid:88) (cid:126)y (cid:10) Tr[ ˆ S ( O , t ; (cid:126)y, t ,(cid:126) , X u,d ( (cid:126)y, t ,(cid:126) ,
0; Γ , S, S )] (cid:11) . (18) A. Sink-sequential method and Stochasticsandwich method
The typical problem of the connected 3pt is calculatingthe propagator from the current to the sink S ( (cid:126)y, t , (cid:126)z, t ).On the surface, it is an all-to-all propagator which wouldbe beyond the ability of the standard lattice inversionoperation.However, when the sink time t is fixed, the se-quential source method [33, 34] could be used, with γ X † u,d ( (cid:126)y, t ,(cid:126) , γ as the source of the matrix inversion,to construct S seq ( X u,d ; (cid:126)z, t , t ,(cid:126) ,
0) = (cid:88) (cid:126)y S ( (cid:126)z, t , (cid:126)y, t ) γ X † u,d ( (cid:126)y, t ,(cid:126) , γ . (19)Then, one can contract S seq with the standard quarkpropagator from t = 0 to t to construct the 3pt correla-tor, C ( t , t , O ) = (cid:88) (cid:126)z,i Tr[ γ S † seq ( X u,d , (cid:126)z, t , t ,(cid:126) , γ O ( (cid:126)z, t ) S ( (cid:126)z, t ,(cid:126) , , (20)taking the advantage of the relation γ S ( z, y ) † γ = S ( y, z ).The disadvantage of the sequential method is that ithas to calculate the sink-sequential propagator repeat-edly when X is changed for any reason, such as for:different momentum, different quark flavor or mass, ordifferent polarization projection of the baryon. This isexpensive when many momenta are needed.The number of inversions required in the sink-sequential method is 2 × × N p where the 2 is for the u and d flavors in the nucleon, 4 is for the polarization,and N p is the number of momentum projections. When many N p are required for nucleon form factors with mo-mentum transfer (hundreds are needed for | (cid:126)p | ≤ δ ( (cid:126)y , (cid:126)y ) at t = t ,1 N noi N noi (cid:88) i =1 (cid:88) (cid:126)y ,(cid:126)y ,(cid:126)z Tr (cid:2) θ ( i ) (cid:126)y S ( (cid:126)y , t , (cid:126)z, t ) O ( (cid:126)z, t ) S ( (cid:126)z, t ,(cid:126) , X ( (cid:126) , , (cid:126)y , t ) θ ( i ) † (cid:126)y (cid:3) −−−−−−→ N noi →∞ C ( t , t , O ) , (21)where N noi is the number of the noises and the noise θ satisfies 1 N noi N noi (cid:88) i =1 θ ( i ) (cid:126)y θ ( i ) † (cid:126)y −−−−−−→ N noi →∞ δ (cid:126)y ,(cid:126)y . (22)In other words, it uses the noise estimate of the all-to-allpropagator, S ( (cid:126)y , t , (cid:126)z, t ) ∼ = (cid:88) i θ ( i ) (cid:126)y γ ( S ( i ) noi ( (cid:126)z, t , t )) † γ (23)with S ( i ) noi ( (cid:126)z, t , t ) = (cid:88) (cid:126)y S ( (cid:126)z, t , (cid:126)y , t ) θ ( i ) † , (24)instead of the original S ( (cid:126)y, t , (cid:126)z, t ), to avoid the expen-sive calculation to construct the sink-sequential propaga-tor with inversion of 2 × × N p sources. B. Stochastic sandwich method (SSM) with LMS
SSM avoids the cost of the repeated inversion for manydifferent sequential sources, but it still requires multipleinversions for several noises, before the SNR can reach itsupper limit – that of the sequential method. In this work,the basic idea is to improve the SNR of the 3pt correlatorof SSM using the low lying eigenvectors of D c to constructthe long distance part of the all-to-all S ( (cid:126)y, t , (cid:126)z, t ) ( S L in Fig. 2, the single line from the current to the sink),and using the noise many-to-all propagator to estimatethe remaining high frequency part of S ( (cid:126)y, t , (cid:126)z, t ) ( S H in Fig. 2, the double line from the current to the sink).Thus, the propagator with LMS is written as S LMS S ( (cid:126)y , t , (cid:126)z, t ) = (cid:88) i θ ( i ) (cid:126)y γ ( S ( i ) ,Hnoi ( (cid:126)z, t , t )) † γ + (cid:88) i λ i + m v i ( (cid:126)y, t ) v † i ( (cid:126)z, t ) . (25)where λ i and v i are the low-lying eigenvalues and thecorresponding eigenvectors of D c . In other words, it isa technique to apply LMS to the sequential propagator S seq ( X u,d ; (cid:126)z, t , t ,(cid:126) ,
0) (LMS S ). It is expected to reducethe number of the noise propagators needed to reach theupper limit of SNR.When LMS S in Eq. (25) is applied to the PropCI inEq. (14), ˆ S comes from t = 0 to t = t through t = t ˆ S LMS S ( O , t ; (cid:126)y, t , t ,(cid:126) ,
0) == (cid:88) (cid:126)z S LMS S ( (cid:126)y , t , (cid:126)z, t ) O ( (cid:126)z, t ) S ( (cid:126)z, t ,(cid:126) , (cid:88) (cid:126)z (cid:0) (cid:88) i λ i + m v i ( (cid:126)y, t ) v † i ( (cid:126)z, t )+ θ ( i ) ( (cid:126)y, t ) (cid:88) (cid:126)z,i γ ( S ( i ) ,Hnoi ( (cid:126)z, t , t )) † γ (cid:1) O ( (cid:126)z, t ) S ( (cid:126)z, t ,(cid:126) , , (26) as shown in the shadowed area in Fig. 2.Then one can construct 3pt with LMS by constructingthe standard 2pt repeatedly (the projection matrix Γ issuppressed for clarity), C LMS,u ( ˆ S , S ) = C ker ( ˆ S H , ˆ S L , S HNG , S
LNG , S
HNG , S
LNG ) + (cid:88) x ∈ G C ker ( ˆ S L ( x ) , ˆ S H , θ ( x ) S L , S HNG , θ ( x ) S L , S HNG ) + C ker ( S HNG , S
LNG , S
HNG , S
LNG , ˆ S H , ˆ S L ) + (cid:88) x ∈ G C ker ( θ ( x ) S L , S HNG , θ ( x ) S L , S HNG , ˆ S L ( x ) , ˆ S H ) C LMS,d ( ˆ S , S ) = C ker ( S HNG , S
LNG , ˆ S H , ˆ S L , S HNG , S
LNG ) + (cid:88) x ∈ G C ker ( θ ( x ) S L , S HNG , ˆ S L ( x ) , ˆ S H , θ ( x ) S L , S HNG ) (27)where S LNG = (cid:88) x ∈ G θ ( x ) S L ( x ) , and C ker ( X , X , Y , Y , Z , Z ) = C ( X , Y , Z )+ C ( X , Y , Z ) + C ( X , Y , Z ) + C ( X , Y , Z )(28)and ˆ S H and ˆ S L ( x ) are the high- and low-mode parts ofˆ S LMS S in Eq. (26).This is the stochastic sandwich method with LMSwhich uses the low eigenmodes for the propagator fromthe current to the sink in PropCI, ˆ S LMS S with current in-sertion and the high modes for the same which originatesfrom the sink time slice. The construction of the PropCIwith low modes needs to be done for each current andmomentum transfer and t (if desired). In contrast, thecurrent-sequential method will need to do an inversionfor each current, momentum transfer, and t separately.To account for the amount of numerical work for differ-ent approaches to the 3pt CI correlators, we note the thetraditional sink-sequential method entails 2 × × N p inver-sions at a fixed sink time slice t , where the 2 and 4 referto the separate sources X in Eq. (17) labeled with u and d flavors and polarization directions (unpolarized and po-larization in 3 spatial directions). N p is the number ofsink momenta for the nucleon. For SSM without LMS, there are N noi inversions of the N noi noise vectors at thesink time t . How many N noi is needed for acceptableSNR depends on the observable. For the SSM with LMS,besides the noise propagator S Hnoi with N Hnoi inversion,there is an overhead for the low-mode portion of PropCI( ˆ S LMS S in Eq. (26)). It includes N times the low-modecontributions from N smeared grid source plus one high-mode contribution for the propagator from the source tothe current ( S HNG ). Each needs to be folded with thecurrent for different momentum transfer (cid:126)q . Thereforethe overhead is (cid:15) × ( N + 1) × N cu × N q where N cu /N q is the number of currents/momentum transfer, and (cid:15) isthe fraction of inversion time for constructing the low-mode portion of ˆ S LMS S for each current and momentumtransfer. We list the cost for the sink and current partsof the 3pt function in units of quark inversion in Table Ifor future reference. To evaluate the efficacy among thethree methods, one needs to compare costs in the table toreach the same precision for a given observable. For thecase of SSM with LMS, there is an additional gain fromthe noise grid source with LMS as discussed in Sec. IIwhich needs to be taken into account. TABLE I: The cost for the sink and current parts of the3pt function in units of quark inversion is listed for the sink-sequential method (Sequential), stochastic sandwich method(SSM), and SSM with LMS. N p is the number of sink nu-cleon momenta, N noi is the number of noise in SSM. N Hnoi isthe number of noise in SSM with LMS, and N cu /N q is thenumber of currents/momentum transfer in the constructionof of the low-mode part of PropCI. (cid:15) is the fraction of inver-sion time for constructing the low-mode portion of PropCIfor each current and momentum transfer and N p momenta( ∼ .
02 on the ensemble used in this work).Sequential SSM SSM+LMS S N p N noi N Hnoi + (cid:15) ( N + 1) N cu N q IV. NUMERICAL DETAILS
In this work, we use the valence overlap fermion on2 + 1 flavor domain-wall fermion (DWF) configurations[20] to carry out the calculation [21].The lattice we use has a size 24 ×
64 with latticespacing a − = 1 . r at the chiraland continuum limits [42]. The light sea u/d quarkmass m l a = 0 .
005 corresponds to m π ∼
330 MeV. Wehave calculated the isovector matrix elements of the nu-cleon for the axial-vector and scalar couplings and thequark momentum fraction at 6 valence quark mass pa-rameters which correspond to the renormalized masses m Rq ≡ m MS q (2GeV) ranging from 13 to 32 MeV after thenon-perturbative renormalization procedure in Ref. [41].They correspond to the pion mass in the range of 250-400 MeV. In order to enhance the signal-to-noise ratioin the calculation of three-point functions, we use twosmeared noise 12-12-12 grid sources at t i = 0 and 32(one is PropNG and the one is PropNGM) [28] and twonoise 2-2-2 grid point sources at positions t f which are8, 10, and 12 time-slices away from the sources on 203configurations.The effective overlap operator D c is chiral, i.e. { D c , γ } = 0 [32], and is expressed in terms of the overlapoperator D ov as D c = ρD ov − D ov / D ov = 1 + γ (cid:15) ( γ D w ( ρ )) , (29)where (cid:15) is the matrix sign function and D w is the WilsonDirac operator with a negative mass characterized by theparameter ρ = 4 − / κ for κ c < κ < .
25. We set κ =0.2which corresponds to ρ = 1 . D c ’s low mode eigenvectors used forthe deflation of the overlap operator inversion and LMS,on this 24 ×
64 lattice, is 200 pairs plus the zero modes,and the upper bound of the absolute value of the eigen-values is 0.154 which is over two times larger than thedimensionless strange quark mass.We check the efficacy of the sequential low-mode sub-stitution (LMS S ) in the PropCI by examining the 3ptfunctions for the isovector axial and scalar currents. Weplot the ratio of 3pt-to-2pt correlators as a function ofthe current insertion time t in Fig. 3 where the sinktime t is 10. The blue dots and black triangles show thecontributions where the current-to-sink part of PropCI isfrom the low modes and the noise-estimated high modesrespectively. Notice that the contribution from the lowmodes is much larger than that of the high modes whenthe current time slice is farther away from the sink (i.e.closer to the source with small t ) for both the axial andscalar cases, which reflects the fact that the low modesdominate the long-distance behavior of the PropCI be-tween t to t . When the current is closer to the sink withlarger t , we see that the high modes dominate for the ax-ial case which shows that the high modes are importantand dominate the short distance behavior of the propa-gator. However, the high-mode contribution is still smallfor the scalar current case when t is close to the sinkwhich shows that the high-mode contribution is smallfor the 3pt function for the scalar current.The red squares are the sum of the low- and high-mode contributions from the present hybrid scheme. Wehave also calculated the 3pt function without LMS S forthe PropCI, but instead use only the noise propagatoras the full propagator from t to t . These are shown asthe green triangles in Fig. 3. These correspond to thestochastic method introduced by the QCDSF Collabora-tion [37, 38] and the Cyprus group [39]. Since our LMS S replaces the long distance part of the current-to-sink partof PropCI with an exact all-to-all one, the larger its con-tribution the larger the improvement. As in Fig. 3, theblue dotd contribute over 80% in the g uS case and so theimprovement of LMS S is larger than in the g uA case. Theerror bars of SSM at the time slices t = 2 − ∼ g uA ( ∼ g uS ) larger than that ofusing LMS S in the present approach.The fact that the error of g A /g S in our approach issmaller than that of SSM with 2 noises by a factor of ∼ / (cid:15) = 0 .
02. Therefore, the overhead (cid:15) ( N + 1) N cu N q = 0 . N = 8 (smeared grid source), N cu = 4 to account forthe scalar current and A i for 3 spatial directions and N q = 1. Together with N Hnoi = 2, the cost is 2.72 inver-sions. This means that, to reach the same error, it wouldtake SSM 2.9 and 11.8 times more inversions than SSMwith LMS for g A and g S respectively. Furthermore, the g u A t allLHall (no LMS S ) g u S t allLHall (no LMS S ) FIG. 3: The 3pt-to-2pt ratio with LMS S (red squares) vs. theone without it (green inverted triangles). The source/sink islocated at 0/10, and the current dependence for the matrixelement with the current-to-sink part of PropCV includingjust the low- or high-mode parts are plotted as the blue dotsor black triangles respectively. The upper panel is for theaxial-vector current case and the lower panel is for the scalarcase. Notice that the contribution of the low-mode part islarger when the current time slice is farther away from thesink. smeared grid source with LMS has improved the statis-tics by a factor of 5.3 for N = 8 for the 2pt function.This additional factor of improvement is also expectedfor the 3pt function.To compare with the sink-sequential method, we as-sume that our results have reached the SNR of that ofthe sink-sequential method. This is consistent with thefact that in the range t = 2 − g S . In this case, the cost ofsink-sequential takes 16 inversions. Here, we have taken N p = 2 to include the (cid:104) x (cid:105) u − d calculation in addition to g A and g S . For the overhead in SSM + LMS, the numberof currents needed is N cu = 6 for these three quantities Z V p smeared pointsmeared grid, no lmssmeared grid, lms FIG. 4: The vector renormalization constant in therest/moving frame at the unitary point, as a function of themomentum squared p in lattice unit. The p = 0 , and the overhead is (cid:15) ( N + 1) N cu N q = 1 .
08. Therefore,besides the improvement from use of the grid source, thepresent method would be 16 / .
08 = 5 . N mass for different masses,and also N when the necessary LMS is applied on thesource of the sink-sequential propagator (as in Eq. 17),so SSM is much cheaper than the sink-sequential method.When the physical volume is increased, while keepingthe lattice spacing unchanged, and with a noise vectorcovering the entire spatial volume of the sink time slice,we expect that the region essentially contributing to 3ptwill not change, while the remaining region contributesonly to the noise. Such a simple argument hints thatthe noise required to reach the same SNR is proportionalto volume and we have confirmed it explicitly on the48 ×
96 lattice with similar lattice spacing [40]. At thesame time, the number of low modes will be proportionalto volume if we want to reach the same upper bound ofthe eigenvalues, so the SSM with LMS will not lose itsefficiency as compared to SSM without LMS, when thevolume is larger. But, since the number of inversions isfixed in the standard sequential method, the SSM withand without LMS will lose their comparative efficiencieswhen the volume is very large.Another issue we need to check is the effect of LMS inthe 3pt case. For the 3pt function, we check, for exam-ple, the vector charge renormalization constant from theforward matrix element at the unitary point for severalnucleon momenta. For p = 0 and 4, only the propagatorPropNG is involved, while the other cases involve Prop-NGM also. In the former cases, we find that the smearedgrid source with LMS improves the SNR by a factor of2.0 compared to that with a smeared point source with-out LMS, slightly smaller than what we found with the2pt function as discussed in Sec. II; whereas, the gain isonly 1.4 for the other p where the PropNGM is involved.We shall look into the possibility of improving the SNRfurther when PropNGM is involved. V. RESULTS
A standard 3pt/2pt ratio in the forward matrix ele-ment case is R ( t , t ,
0) = C ( t , t , /C ( t , (cid:80) i,j Z ( i ) f Z ( j ) i e − E ( i ) ( t − t ) − E ( j ) t (cid:104) χ ( i ) f | J | χ ( j ) i (cid:105) (cid:80) k Z ( k ) f Z ( k ) i e − E ( k ) t −−−→ t (cid:29) (cid:104) χ (0) f | J | χ (0) i (cid:105) + Z (1) f Z (0) f (cid:104) χ (1) f | J | χ (0) i (cid:105) e − ∆ E ( t − t ) + Z (1) i Z (0) i (cid:104) χ (0) f | J | χ (1) i (cid:105) e − ∆ Et + Z (1) f Z (1) i Z (0) f Z (0) i ( (cid:104) χ (1) f | J | χ (1) i (cid:105) − (cid:104) χ (0) f | J | χ (0) i (cid:105) ) e − ∆ Et + ..., (30)where E ( i ) and Z ( i ) are the energy and the overlap of theinterpolation field of the i th state and ∆ E = E (1) − E (0) .For t (cid:29) t (cid:29)
0, the contributions from all the termsin the right hand of Eq. (30) except the first term van-ish, and then one can use Eq. (30) to obtain the matrixelement.When t is fixed, one may fit the first term and thecombined second and third terms around t = t / t dependent. But since the fourth term in the right handside of Eq. (30), which is the difference of the matrixelement in the ground state and the first excited state,is independent of t just like the first term, one will notbe able disentangle them and, as a result, a systematicerror may be induced by its contribution which is sup-pressed by e − ( E (1) − E (0) ) t . To get a feeling for the sizeof the correction, let us suppose that the first excitedstate matrix element (cid:104) χ (1) f | J | χ (1) i (cid:105) is 30% different fromthe ground state matrix element (cid:104) χ (0) f | J | χ (0) i (cid:105) , and themass difference of the first excited state and the groundstate is about 500 MeV. Then the correction from sucha effect with t =8, 10 and 12 (with the nucleon sourceset at t = 0) is about 3%, 2% and 1% respectively. Toassess this error, we shall calculate the 3pt function at three values of t so that we can fit all four terms inEq. (30).In order to check the t dependence of the plateau,three sets of propagators with two noise-grid pointsources each at positions t = 8 ,
10 and 12 time-slicesaway from the nucleon source are generated, and all the t dependence of these three cases are plotted togetherfor comparison in Fig. 5 for the vector current case. Thesink-source separation dependence seems to be mild here,but in general the minimum separation required by otherquantities can be different. g V u + d / t-t /2t max = 8t max =10t max =12 FIG. 5: The nucleon sink-source separation dependence ofthe matrix element with the vector charge for u + d in theconnected insertion. Obviously, the larger t , the worse thesignal. The data points marked with the black squares ( t =8),the blue dots ( t =10) and the red triangles ( t =12) are con-sistent. To check the separation effect quantitatively, we ap-plied three kinds of fits to deduce the results:The first method is to fit the ratio as a function of t and t , R fit ( t , t ) = C + C e − ∆ m ( t − t ) + C e − ∆ mt + C e − ∆ mt (31)with C , , , and ∆ m as free parameters. C is theground state matrix element we want. Since the t de-pendence of R ( t , t ) is mild in some of the quantities like g V and g A , we take ∆ m as a common parameter for allthe quantities. This is what we mark as “2-state” in thefollowing discussion.In this work, we use the smeared source and the pointsink, so the excited-state contaminations are different inthe smaller and larger t ends. If the smeared sourcemakes the contaminations in the smaller t end small, orhas a different sign compared to that in the larger t end,the position of the plateau will be harder to determine,as in the case of g u + dV (Fig. 5) and g A (Fig. 7). Ap-plying the “2-state” fit on such a quantity is not stableand provides large uncertainties (and/or large χ /d.o.f. )0on the results. In this work, we constrain the mass dif-ference ∆ m to be the same for the different matrix ele-ments with the same quark mass value, and apply a cor-related joint 2-state fit. To suppress the contaminationfrom the excited state, we excluded the data points with t = 0 , t − t . One more data point at the larger t end is excluded since the excited-state contamination islarger there. Despite this, the fit is still not very good.Taking the unitary point as an example, the χ /d.o.f. with ∼
70 degrees of freedom is 1.45, the correspondingp-value is just 0.008. In addition, this method requiresa joint fit with several quantities and is not suitable forthe analysis of a single quantity without the informationof the other quantities.The second method is the sum method [43, 44] which isused in the disconnected insertion case, wherein a sum istaken over all the 3pt/2pt ratios in Eq. (30) with different t , SR ( t , t ,
0) = (cid:88) 4) GeV gives the value of the vector renormaliza-tion factor as 1.096(6) which is just slightly smaller thanthe value 1.105(4) obtained from the axial Ward identity[41]. g A t-t /2t max = 8t max =10t max =12 FIG. 7: The sink-source separation dependence of the matrixelement of the isovector axial-vector current. The data of theisovector case with t = 8 (black squares) are slightly smallerthan that from the other two separations, while the resultwith t =10 (blue dots) is consistent with that with t =12(red triangles). Then the renormalization of the vector current can beused to renormalize the axial-vector matrix element withpolarized projection, g bA ≡ (cid:80) i =1 , , Tr[Γ mi (cid:104) P | (cid:82) d xψ ( x ) γ γ i ψ ( x ) | P (cid:105) ]3Tr[Γ e (cid:104) P | P (cid:105) ] g RA ≡ g bA Z V = (cid:80) i =1 , , Tr[Γ i (cid:104) P | (cid:82) d xψ ( x ) γ γ i ψ ( x ) | P (cid:105) ]3Tr[Γ e (cid:104) P | (cid:82) d xψ ( x ) γ ψ ( x ) | P (cid:105) ] (37)where the superscript b/R stands for thebare/renormalized value respectively and Γ mi =(1 + γ ) γ i γ / g bV (instead of that from the axial Ward iden-tity for pion) to renormalize g A as in Eq. (37) could im-prove the signal of the renormalized g A by ∼ 20% sincethese two matrix elements are correlated. As observed in Fig. 7, the sink-source separation dependence for theisovector case is mild, while a curve is observable at theright side of the plateau due to a larger excited state con-tribution from the point interpolation field at the sink.This is in contrast to the flatter behavior to the left ofthe plateau where the excited-state contribution is ame-liorated by the smeared source. In Fig. 8, we plot theresults of the isovector axial-vector coupling g A from thethree fitting methods we mentioned. We note that thosefrom the “mixed” method are always between those fromthe other two methods, for all the data points in therange of m π ∈ (0 . , . 4) GeV. The values from the threemethods at the unitary point are listed in Table II. Sim-ilar to other lattice calculations at this pion mass (i.e. ∼ 300 MeV), irrespective of which fit is used, the isovec-tor axial-vector matrix element, g u − dA is ∼ 10% smallerthan the experimental value 1.2723(23)[45]. g A m π (GeV)2-statesummixed FIG. 8: The isovector axial-vector matrix element vs the pionmass, from three kinds of fitting method: 2-state fit (redsquares), summed slope (black triangles), and the mixed fitwhich combines those two methods (the blue dots). The re-sults from these different methods are consistent while thatfrom the mixed method provides the best signal. B. Scalar case Similarly, the renormalized scalar matrix element withthe unpolarized projection of the nucleon can be calcu-lated by, g S ≡ Z S Tr[Γ e (cid:104) P | (cid:82) d xψ ( x ) ψ ( x ) | P (cid:105) ]Tr[Γ e (cid:104) P | P (cid:105) ] , (38)where the renormalization constant Z S is obtained fromthe RI/MOM scheme and its value on the ensemble weuse here is calculated to be 1.1397(54) [41]. On the otherhand, if one just focuses on the πN σ term, 2 Z m m q Z S g bS ,the renormalizations of the quark mass Z m and that of2 g S t-t /2t max = 8t max =10t max =12 g S ( C I ) t-t /2t max = 8t max =10t max =12 FIG. 9: The separation dependence of the matrix element ofscalar current, for both the isovector and the CI part of thesinglet case. The dependence is mild for the isovector case(the upper panel), while obvious for the CI part of the singletcase (the lower panel). the scalar matrix element Z S are canceled and so the πN σ term is free of the renormalization.It is interesting to point out that the CI part of thescalar singlet matrix element has a strong sink-sourceseparation dependence, as seen in the lower panel ofFig. 9. At the same time, such a separation dependenceseems to be canceled between the u and d quarks, so thatthe isovector case in the upper panel of Fig. 9 has only amild separation dependence. The results for the isovec-tor scalar matrix element from the three fitting methodsare plotted in Fig. 10 and those at the unitary point arelisted in Table II. This shows that, despite the fact thatthere are 2 u valence quarks and only one d quark in theproton, the d contribution to the scalar matrix elementper quark is more than that of the u , as g uS,CI g dS,CI = 0 . u and d quark increase as m q decreases, but theisovector scalar matrix element is not far from unity overthe entire quark mass region from light to heavy. Thishas been interpreted to be related to the Gottfried sumrule violation [46] where it is found experimentally thatthere are more d antipartons than u antipartons.. g S m π (GeV)2-statesummixed FIG. 10: Isovector scalar matrix element vs. pion massfrom three kinds of fitting method: 2-state fit (red squares),summed slope (black triangles), and the mixed fit which com-bines those two methods (blue dots). The results from thesedifferent methods are slightly different. C. Quark momentum fraction The quark momentum fraction in the nucleon can becalculated with the traceless part of the energy momen-tum tensor, and it should be consistent between calcula-tions with two different operators. The first one uses thecombination of the diagonal temporal and spatial com-ponents of the energy momentum tensor, (cid:104) x (cid:105) E ≡ Tr[Γ e (cid:104) P | (cid:82) d xO E ( x ) | P (cid:105) ] E Tr[Γ e (cid:104) P | P (cid:105) ] . (40)where O E ( x ) = ψ ( x ) ( γ ←→ D − (cid:80) i =1 , , γ i ←→ D i ) ψ ( x ) isthe traceless part of the energy momentum tensor T and is a measure of the quark fraction of the nucleonmass or energy. The related matrix element can be cal-culated in the rest frame and, as a result, it will havea good signal. On the other hand, the operator T it-self can have mixing with lower dimension operators likethe dimension-3 scalar operator ψ ( x ) ψ ( x ). Nevertheless,such a mixing will be canceled due the subtraction of thediagonal spatial components in O E .The other approach uses the forward off-diagonal ma-trix components of the energy momentum tensor ( T i ) in3 Eu-d FIG. 11: The plateau fit values ( t = 10 case) of the isovectormomentum fraction for (cid:104) x (cid:105) E in the rest frame (red square)and also (cid:104) x (cid:105) P in a moving frame with different momenta (bluedots). The results from both the diagonal and off-diagonalcomponents (and also that from different momenta based onthe off-diagonal matrix components) are consistent, but thefirst approach provides much better SNR. a moving frame, (cid:104) x (cid:105) P ≡ Tr[Γ e (cid:104) P | (cid:82) d xψ ( x ) ( γ i ←→ D + γ ←→ D i ) ψ ( x ) | P (cid:105) ] p i Tr[Γ e (cid:104) P | P (cid:105) ] (41)with p i being the i -th component of the nucleon momen-tum. Therefore, it is a measure of the quark momentumfraction in a moving nucleon. Such a scheme is free ofmixing of the lower dimension operators due to its tensorstructure, while the corresponding matrix element is pro-portional to the momentum and is thus more noisy thanthat from the first approach, because mixed momentumsources are involved for the matrix element of the nucleonat non-zero momentum.Fig. 11 shows the plateau fit values of the t =10case for the quark isovector momentum fraction. Theyare (cid:104) x (cid:105) E from the diagonal components of the energy-momentum tensor with the nucleon in the rest frame andalso (cid:104) x (cid:105) P from the off-diagonal components in a movingframe with different momenta. The results from boththe diagonal and off-diagonal components (and also thosefrom different momenta) are consistent, but (cid:104) x (cid:105) E pro-vides much better SNR. The sink-source separation de-pendence is shown in Fig. 12, for both results based onthe diagonal components and off-diagonal components.It is interesting to observe that the separation depen-dence of the isovector quark momentum fraction basedon the off-diagonal components seems to be milder thanthat based on the diagonal ones, for the cases with t =8and 10. The (cid:104) x (cid:105) P case with t =12 seems to have some t dependence at the smeared source end, but it could bedue to the statistical fluctuation due to relatively poor < x > E u - d t-t /2t max = 8t max =10t max =12 < x > P u - d t-t /2t max = 8t max =10t max =12 FIG. 12: The sink-source separation dependence of the isovec-tor quark momentum fraction for the case of the diago-nal components of the energy momentum tensor (the upperpanel) and that of the off-diagonal ones (the lower panel). signal.As in Ref. [51], the renormalization factor for the en-semble we used has been obtained with the one-loop lat-tice perturbative theory, as 1.049(3), in the M S schemeat 2 GeV. The error is from the uncertainty of the latticespacing. The renormalized values of the isovector quarkmomentum fraction of (cid:104) x (cid:105) E from the three fitting meth-ods are plotted in Fig. 13, and those at the unitary pointare listed in Table II. VI. SUMMARY We have introduced a new method to calculate the nu-cleon matrix elements in the connected insertion. Thestochastic sandwich method (SSM) with low-mode sub-stitution (LMS) is an approach which uses low modes forthe all-to-all quark propagator between the current andthe sink and the corresponding high-mode contributionis taken care of by the noise propagator from the sink4 < x > u - d m π (GeV)2-statesummixed FIG. 13: The isovector quark momentum fraction ( (cid:104) x (cid:105) E inthe rest frame) vs. the pion mass, from three fitting methods:2-state fit (red squares), summed slope (black triangles), andthe mixed fit which combines those two methods (blue dots).The results from the three methods are consistent. to the current. We have shown that it is more efficientthan the sink- and current- sequential methods. How-ever, it does not scale well with volume which requiresmore low eigenmodes. It will lose its advantage whenthe overhead from calculating the LMS for all the quarkpropagators involved is more than the amount it savescompared with the sink-sequential or current-sequentialmethod. But this will occur only at volumes much largerthan that used here.We have used three fitting methods. One is a two-statefitting including the contamination from the excited-statetransition and the second is the summed-slope method.The third is a mix of these two methods.The proton isovector axial-vector coupling g A we ob-tain with the overlap fermion at the unitary point with m π =330 MeV is g A = 1 . ∼ 8% smaller than the experimental value.The separation dependence of this quantity is mild.Since it is smaller than the experimental value on thislattice, it is essential to repeat the calculation of g A onlarger volumes and with lighter quark masses.For the isovector scalar matrix element in the pro-ton, the renormalized value at M S (2GeV) at the unitarypoint is g S = 0 . . (43) This shows that, despite the fact that there are 2 u va-lence quarks and only one d quark in the proton, the d contribution to the scalar matrix element per quark ismore than that of the u , as g uS,CI g dS,CI = 0 . d antipartons than u antipartons.In the isovector quark momentum fraction case, thebare value we obtained at the unitary point on the en-semble mentioned above is (cid:104) x (cid:105) u − d = 0 . , (45)with the renormalization factor 1.049(3) from one-looplattice perturbative theory [51]. This value is similar tothose from most lattice calculations [3, 7, 14–16] and islarger than the experimental value. However, the O ( a )error has not been considered. It can be assessed byimposing the momentum and angular momentum sumrules at finite lattice spacing as is demonstrated in aquenched calculation [52]. We will return to this issuewhen the complete lattice simulation of the momentumand angular-momentum decompositions is carried out.We will perform calculations with physical sea quarkmasses in the future. Acknowledgments We thank the RBC and UKQCD Collaborations forproviding us their DWF gauge configurations. This workis supported in part by the U.S. Department of En-ergy under Grant No. DE-FG05-84ER40154, and de-sc0013065. A.A. acknowledges the support of NSF CA-REER through grant PHY-1151648. M.G. is partiallysupported by the National Science Foundation of China(NSFC) under the project No. 11405178 and the YouthInnovation Promotion Association of CAS (2015013).This research used resources of the Oak Ridge Leader-ship Computing Facility at the Oak Ridge National Lab-oratory, which is supported by the Office of Science ofthe U.S. Department of Energy under Contract No. DE-AC05-00OR22725. [1] B. J. Owen, J. Dragos, W. Kamleh, D. B. Leinweber,M. S. Mahbub, B. J. Menadue and J. M. Zanotti, Phys. Lett. B , 217 (2013) [arXiv:1212.4668 [hep-lat]].[2] T. Bhattacharya, S. D. Cohen, R. Gupta, A. Joseph, H. W. Lin and B. Yoon, Phys. Rev. D , no. 9, 094502(2014) [arXiv:1306.5435 [hep-lat]].[3] C. Alexandrou, M. Constantinou, S. Dinter, V. Drach,K. Jansen, C. Kallidonis and G. Koutsou, Phys. Rev. D , 014509 (2013) [arXiv:1303.5979 [hep-lat]].[4] C. Alexandrou et al. [ETM Collaboration], Phys. Rev. D , 045010 (2011) [arXiv:1012.0857 [hep-lat]].[5] S. Ohta [RBC and UKQCD Collaborations], PoS LAT-TICE , 274 (2014) [arXiv:1309.7942 [hep-lat]].[6] J. D. Bratt et al. [LHPC Collaboration], Phys. Rev. D , 094502 (2010) [arXiv:1001.3620 [hep-lat]].[7] S. Syritsyn et al. , PoS LATTICE , 134 (2015)[arXiv:1412.3175 [hep-lat]].[8] J. R. Green, M. Engelhardt, S. Krieg, J. W. Negele,A. V. Pochinsky and S. N. Syritsyn, Phys. Lett. B ,290 (2014) [arXiv:1209.1687 [hep-lat]].[9] S. Capitani, M. Della Morte, G. von Hippel, B. Jager,A. Juttner, B. Knippschild, H. B. Meyer and H. Wittig,Phys. Rev. D , 074502 (2012) [arXiv:1205.0180 [hep-lat]].[10] R. Horsley, Y. Nakamura, A. Nobile, P. E. L. Rakow,G. Schierholz and J. M. Zanotti, Phys. Lett. B , 41(2014) [arXiv:1302.2233 [hep-lat]].[11] G. S. Bali et al. , Phys. Rev. D , no. 5, 054501 (2015)[arXiv:1412.7336 [hep-lat]].[12] A. Abdel-Rehim et al. , arXiv:1507.04936 [hep-lat].[13] Y. Aoki, T. Blum, H. W. Lin, S. Ohta, S. Sasaki,R. Tweedie, J. Zanotti and T. Yamazaki, Phys. Rev. D , 014501 (2010) [arXiv:1003.3387 [hep-lat]].[14] G. S. Bali et al. , Phys. Rev. D , no. 7, 074510 (2014)[arXiv:1408.6850 [hep-lat]].[15] D. Pleiter et al. [QCDSF/UKQCD Collaboration], PoSLATTICE , 153 (2010) [arXiv:1101.2326 [hep-lat]].[16] Y. Aoki, T. Blum, H. W. Lin, S. Ohta, S. Sasaki,R. Tweedie, J. Zanotti and T. Yamazaki, Phys. Rev. D , 014501 (2010) [arXiv:1003.3387 [hep-lat]].[17] J. R. Green, J. W. Negele, A. V. Pochinsky, S. N. Syrit-syn, M. Engelhardt and S. Krieg, Phys. Rev. D ,114509 (2012) [arXiv:1206.4527 [hep-lat]].[18] M. Gonzlez-Alonso and J. Martin Camalich, Phys. Rev.Lett. , no. 4, 042501 (2014) [arXiv:1309.4434 [hep-ph]].[19] T. Bhattacharya, V. Cirigliano, S. D. Cohen, A. Fil-ipuzzi, M. Gonzalez-Alonso, M. L. Graesser, R. Guptaand H. W. Lin, Phys. Rev. D , 054512 (2012)[arXiv:1110.6448 [hep-ph]].[20] Y. Aoki et al. [RBC and UKQCD Collaborations], Phys.Rev. D , 074508 (2011) [arXiv:1011.0892 [hep-lat]].[21] A. Li et al. [xQCD Collaboration], Phys. Rev. D ,114501 (2010) [arXiv:1005.5424 [hep-lat]].[22] A. Alexandru, M. Lujan, C. Pelissier, B. Gamari andF. X. Lee, arXiv:1106.4964 [hep-lat].[23] T. A. DeGrand and S. Schaefer, Comput. Phys. Com-mun. , 185 (2004) [hep-lat/0401011].[24] L. Giusti, P. Hernandez, M. Laine, P. Weisz and H. Wit-tig, JHEP , 013 (2004) [hep-lat/0402002].[25] L. Giusti, P. Hernandez, M. Laine, C. Pena, J. Wennekersand H. Wittig, Phys. Rev. Lett. , 082003 (2007) [hep-ph/0607220].[26] J. Foley, K. Jimmy Juge, A. O’Cais, M. Peardon, S. M. Ryan and J. I. Skullerud, Comput. Phys. Com-mun. , 145 (2005) [hep-lat/0505023].[27] T. Kaneko et al. [JLQCD Collaboration], PoS LAT ,148 (2007) [arXiv:0710.2390 [hep-lat]].[28] M. Gong [XQCD Collaboration], A. Alexandru, Y. Chen,T. Doi, S.J. Dong, T. Draper, W. Freeman, M. Glatz-maier, A. Li, K.F. Liu, and Z. Liu, Phys. Rev. D , no.1, 014503 (2013) [arXiv:1304.1194 [hep-ph]].[29] W. Wilcox, T. Draper and K. F. Liu, Phys. Rev. D ,1109 (1992) [hep-lat/9205015].[30] T.-W. Chiu, Phys. Rev. D , 034503 (1999) [hep-lat/9810052].[31] K.-F. Liu and S.J. Dong, Int. J. Mod. Phys. A , 7241(2005) [hep-lat/0206002].[32] T.-W. Chiu and S. V. Zenkin, Phys. Rev. D , 074501(1999) [hep-lat/9806019].[33] C.W. Bernard, Gauge Theory on a Lattice, 1984, editedby C. Zachos et al., Argonne National Laboratory, Ar-gonne, IL (1984) 85; T. Draper, Ph. D. thesis, UMI-84-28507 (1984); C. W. Bernard, T. Draper, G. Hockney,A. M. Rushton and A. Soni, Phys. Rev. Lett. , 2770(1985).[34] G. Martinelli and C. T. Sachrajda, Nucl. Phys. B ,355 (1989).[35] T. Draper, R. M. Woloshyn and K. F. Liu, Phys. Lett.B , 121 (1990).[36] T. Draper, R. M. Woloshyn, W. Wilcox and K. F. Liu,Nucl. Phys. B , 319 (1989).[37] R. Evans, G. Bali and S. Collins, Phys. Rev. D , 094501(2010) [arXiv:1008.3293 [hep-lat]].[38] G. S. Bali et al. , PoS LATTICE , 271 (2014)[arXiv:1311.1718 [hep-lat]].[39] C. Alexandrou et al. [ETM Collaboration], Eur. Phys. J.C , no. 1, 2692 (2014) [arXiv:1302.2608 [hep-lat]].[40] T. Blum et al. [RBC and UKQCD Collaborations],arXiv:1411.7017 [hep-lat].[41] Z. Liu et al. [chiQCD Collaboration], Phys. Rev. D ,no. 3, 034505 (2014) [arXiv:1312.7628 [hep-lat]].[42] Y. B. Yang et al. , arXiv:1410.3343 [hep-lat].[43] L. Maiani, G. Martinelli, M. L. Paciello and B. Taglienti,Nucl. Phys. B , 420 (1987).[44] M. Deka, T. Streuer, T. Doi, S. J. Dong, T. Draper,K. F. Liu, N. Mathur and A. W. Thomas, Phys. Rev. D , 094502 (2009) [arXiv:0811.1779 [hep-ph]].[45] K. A. Olive et al. [Particle Data Group Collaboration],Chin. Phys. C , 090001 (2014).[46] K. F. Liu and S. J. Dong, Phys. Rev. Lett. , 1790(1994) [hep-ph/9306299].[47] P. Amaudruz et al. [New Muon Collaboration], Phys.Rev. Lett. , 2712 (1991).[48] Y. B. Yang, M. Gong, K. F. Liu and M. Sun, PoS LAT-TICE , 138 (2014) [arXiv:1504.04052 [hep-ph]].[49] P. Hasenfratz, S. Hauswirth, T. Jorg, F. Niedermayerand K. Holland, Nucl. Phys. B , 280 (2002) [hep-lat/0205010].[50] P. A. Boyle, arXiv:1411.5728 [hep-lat].[51] M. Glatzmaier, in preparation.[52] M. Deka et al. , Phys. Rev. D91