Elucidating the sign problem through noise distributions
Abstract
Due to the presence of light pions in the theory, lattice QCD at finite densities suffers from issues with noise in both grand canonical and canonical formulations. We study two different formulations of the NambuJonaLasinio model reduced to 2+1 dimensions at large , where is the number of flavors. At finite chemical potential one formulation has a severe sign problem and a fermion correlator which displays a broad probability distribution with small mean. In the other we find no sign problem and a distribution amenable to the cumulant expansion techniques developed in Ref. [1, 2, 3].

Maryland Center for Fundamental Physics, Department of Physics, University of Maryland, College Park MD 207424111, USA

Institute for Nuclear Theory, University of Washington, Seattle, WA 981951550, USA

Email: amynn@umd.edu, grabow@uw.edu, dbkaplan@uw.edu
1 Introduction
While lattice QCD is currently the only tool for studying QCD from first principles in the nonperturbative regime, the study of QCD at finite densities on the lattice has been greatly hindered by two related problems. In the first, lattice QCD at fixed baryon chemical potential suffers from what is known as a sign problem, which arises due to the complexity of the fermion determinant one wishes to use as a probability measure for Monte Carlo calculations. A possible solution is to employ phase reweighting, in which the phase of the determinant is absorbed into the observable and the magnitude of the determinant is used as a probability measure; however, the expectation value of the phase using the same ensemble must also be calculated in order to reconstruct the original observable. One finds that for above some critical value, this expectation value vanishes exponentially and becomes swamped by statistical noise. The second problem is encountered in a canonical formulation, in which a fixed number of quark sources and sinks are separated by a long Euclidean time to project out the ground state. Again, the signaltonoise ratio is found to vanish exponentially with time.
A notable property of these two problems is that they are both related to the same dynamical property of QCD; namely, they both arise due to the presence of light pions and thus are intimately related to the phenomenon of chiral symmetry breaking. In the case of finite chemical potential, this can be seen by noting that for two degenerate flavors of quark the magnitude of the fermion determinant corresponds to chemical potential for isospin. Thus, at low temperatures for pion condensation would tend to occur, and the role of the phase must be to cancel this pion condensation [4, 5].
For the other case where finite baryon density is achieved using correlation functions of quark propagators, the role of the pion in the signaltonoise problem may be noted by considering the variance of an operator used to create nucleons. While at large Euclidean times the correlator for three appropriately contracted quark propagators projects out the ground state of the nucleon, , the variance will be composed of three quark and three antiquark propagators, projecting out the pion mass at large times [6, 7]. The signaltonoise ratio will thus be exponentially vanishing with the factor , where is the Euclidean time, for late Euclidean times.
Both of the noise problems presented above can be shown to arise from probability distributions which are broad, symmetric, and centered about zero. For the nucleon correlator, one may see this by looking at higher moments of the distribution function. Even moments consist of equal numbers of quark and antiquark propagators which may pair up to form pions, while odd moments have three unpaired quark propagators which will form a baryon [8]. Thus, because odd moments will always contain a baryon mass in the exponential, they will be suppressed relative to the even moments, giving a symmetric distribution with exponentially small mean. A similar argument may be made for the expectation value of the phase at finite chemical potential, where all odd moments are equivalent to the mean, and therefore exponentially small for , while all even moments give exactly (for a discussion of the distribution of the phase, see [9]).
In [1, 2, 3] it was shown that another type of noise problem exists in which the distribution displays a long tail, resulting in a sample mean which may be far from the true mean with deceptively small error bars. This problem is often referred to as an overlap problem, and has been known to occur for observables when reweighting is performed. This was discussed in a canonical formulation using lattice calculations of unitary fermions, a stronglyinteracting nonrelativistic system in which the fermion determinant is real and positive for even numbers of fermion flavors. The probability distributions for fermion correlators were shown to be nearly lognormal, and a possible solution to the overlap problem was postulated in the form of a truncated cumulant expansion for the correlator,
(1) 
where is some operator of interest on a background field , is the th cumulant of the logarithm of the operator, and is some suitable truncation order, determined by examining the convergence of the series. For a distribution that is sufficiently close to lognormal, this series may converge at a low enough order to be useful in practical applications. Recently it has also been reported that nearly lognormal distributions are also seen in lattice QCD calculations of correlation functions for intermediate times [10].
2 Two formulations for a large NJL model in dimensions
To understand the origin of these two types of noise problem in QCD and to further explore the link between signaltonoise and the presence of light pions, we turn now to a QCDlike model which displays chiral symmetry breaking without the added complication of confinement. For more details behind these calculations, see [11]. The model we will consider is the NJL model in dimensions for large number of fermion flavors, ,
(2) 
where are a flavor index summed over , indices, are summed over 3D coordinates , while Greek indices are summed over 4D coordinates . The gamma matrices are the usual matrices used in 4D, and so the Lagrangian represents flavors of 3D Dirac fermions. In the limit this theory has a chiral symmetry in 4D, which becomes a flavor symmetry in 3D; this “chiral” symmetry is spontaneously broken as in dimensions, giving rise to a Goldstone boson.
To study this model numerically one may introduce auxiliary fields to generate a fourfermion interaction, which can then be integrated over using Monte Carlo methods. We find two physically equivalent methods for doing so. The first, which we will call the formulation, is the conventional method which introduces scalar fields that are singlet under the flavor symmetry,
(3) 
The second formulation, which we will call the formulation, follows if one performs a Fierz rearrangement of the fourfermion interaction in eq. (2) before introducing auxiliary fields. This results in the introduction of matrix valued vector and axial vector auxiliary fields and , and the equivalent theory
(4) 
To determine whether these two formulations will cause a Monte Carlo sign problem at nonzero chemical potential, we should investigate the positivity of the fermion determinants. For the formulation, it is possible to define a real symmetric charge conjugation matrix satisfying , for and . Then the fermion operator for a single flavor in the grand canonical formulation satisfies , where , and complex eigenvalues of must come in conjugate pairs. Thus is real, and positive for even . This implies that there is no sign problem at finite density [12].
In the formulation, the fermion matrix at finite chemical potential is given by , which is similar in structure to the QCD Dirac matrix with nonzero , and its determinant is similarly complex. In fact, as in QCD, the magnitude of the fermion determinant for two degenerate families, , corresponds to isospin chemical potential, so that for , the phase is responsible for eliminating pion condensation in the ground state.
Chiral symmetry breaking is simple to see in the formulation, where the large expansion is equivalent to the semiclassical expansion. Upon integrating out the fermion fields we have the following effective action:
(5) 
Performing a mean field calculation leads to (using dimensional regularization and the MS subtraction scheme),
(6) 
where we have defined to be the chiral symmetry breaking minimum when , and is the constituent fermion mass. We may also find the and dispersion relations by expanding the effective action to second order about our mean field solution.
In the formulation we cannot use mean field theory; instead, to leading order in we may find the fermion propagator by solving the SchwingerDyson equation and finding the nonzero fermion mass, Eq. 2. Furthermore, one may derive the and meson propagators by solving the matrix equation shown in Fig. LABEL:fig:fourpoint. One finds . Thus we see that an interaction via channel exchange of and mesons is exactly equivalent to a single meson in the or channel in the formulation, corresponding to a valence fermion/antifermion pair or two valence fermions or antifermions, respectively.