]>NUPHB13823S0550-3213(16)30215-210.1016/j.nuclphysb.2016.07.031The Author(s)High Energy Physics – TheoryFig. 1Fermion number n as a function of chemical potential μ at βU=30 with and without reweighting.Fig. 1Fig. 2Average phase function 〈Θ(zη,z‾η)〉η as a function of chemical potential μ at βU=30.Fig. 2Fig. 3Partially integrated distribution Py and its moment y5Py at α=1.4.Fig. 3Fig. 4Expectation value of iz as a function of α, with and without reweighting. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)Fig. 4Fig. 5Expectation value of (iz)2 as a function of α, with and without reweighting. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)Fig. 5Complex saddle points and the sign problem in complex Langevin simulationTomoyaHayataa⁎hayata@riken.jpYoshimasaHidakabYuyaTanizakicaRIKEN Center for Emergent Matter Science (CEMS), Wako, Saitama 351-0198, JapanRIKEN Center for Emergent Matter Science (CEMS)WakoSaitama351-0198JapanbTheoretical Research Division, Nishina Center, RIKEN, Wako, Saitama 351-0198, JapanTheoretical Research DivisionNishina CenterRIKENWakoSaitama351-0198JapancRIKEN BNL Research Center, Brookhaven National Laboratory, Upton, NY 11973, USARIKEN BNL Research CenterBrookhaven National LaboratoryUptonNY11973USA⁎Corresponding author.Editor: Stephan StiebergerAbstractWe show that complex Langevin simulation converges to a wrong result within the semiclassical analysis, by relating it to the Lefschetz-thimble path integral, when the path-integral weight has different phases among dominant complex saddle points. Equilibrium solution of the complex Langevin equation forms local distributions around complex saddle points. Its ensemble average approximately becomes a direct sum of the average in each local distribution, where relative phases among them are dropped. We propose that by taking these phases into account through reweighting, we can solve the wrong convergence problem. However, this prescription may lead to a recurrence of the sign problem in the complex Langevin method for quantum many-body systems.1IntroductionPrecise analysis of thermodynamic properties of a quantum many-body system, in particular, precise determination of its phase diagram is one of great challenges in theoretical physics. An ab initio simulation based on lattice field theory, in particular, so called Monte Carlo simulation is the most powerful tool for this. In many interesting cases, however, Monte Carlo simulation is hindered by the notorious sign problem. The importance sampling, using the Boltzmann weight e−S, breaks down when the action becomes complex. In hadron physics, lattice quantum chromodynamics (QCD) simulation suffers from the sign problem at finite quark densities [1,2], which is important to study quark matter inside neutron stars [3,4]. The sign problem occurs also in condensed matter systems [5–7]. Important examples are the fermionic Hubbard model away from half-filling, and geometric frustration in spin systems. A method to overcome the sign problem attracts a broad interest for application to the aforedescribed quantum many-body systems.There have been a lot of attempts to tackle the sign problem. Among them, idea of complexification of the integration variables is one promising way to solve the sign problem. Theoretical attempts along this line are classified into two approaches, that is, the Lefschetz-thimble and the complex Langevin methods. The Picard–Lefschetz theory gives a generalization of the steepest descent method, and Lefschetz thimbles are steepest descent paths in the extended complex plane [8–10]. This method is formulated on rigorous mathematics, but it needs some approximation when applied to quantum many-body systems [11–16]. On the other hand, the complex Langevin method is an extension of the Langevin equation to a complex Boltzmann weight [17–20]. The numerical implementation of this is possible based on lattice field theory. The complex Langevin method has been widely applied from condensed matter to hadron physics [21–26]. There is a formal proof [27,28] on the correctness of the complex Langevin method, where it has been shown that the complex Langevin method correctly gives physical observables if the distribution obtained from the Langevin equation damps exponentially fast around infinities and singular points. This method is, however, known to give wrong results for some cases, where distribution does not show the exponentially fast decay, and thus the formal proof cannot be applied (for recent discussions, see also [29–35]). Therefore, it is important to unveil what properties of the classical action cause the wrong convergence of the complex Langevin method.In this paper, we show within the semiclassical analysis that complex Langevin simulation converges to a wrong result, when path-integral weight at complex saddle points has different phases. This includes the case of the breakdown due to a singular drift term, e.g., the lattice QCD at finite density. We reveal that complex Langevin simulation breaks down more generic case where the Langevin drift term has no singular point. With the help of semiclassical analysis, we find that reweighting by the complex phase can partially solve the wrong convergence problem. However, the reweighting leads, in general, to a severe cancellation of the reweighting factor in many-body systems, which is nothing but a sign problem in terms of the complex Langevin method.2Complex Langevin method and its failureFor simplicity, we discuss an oscillatory integral of one variable x, which can be extended to multiple integrals in a straightforward way,(1)〈O(x)〉=1Z∫Rdxe−S(x)/ħO(x), where Z is the normalization factor. The action S(x) is complex valued in general, which makes the Monte Carlo simulation of Eq. (1) difficult because of the sign problem. One proposal to calculate Eq. (1) for a complex valued action is the so-called complex Langevin method [18–20]. In this method, we solve the Langevin equations for complex values z=x+iy along the fictitious time direction θ,(2)∂θzη(θ)=−S′(zη(θ))+ħη(θ), where η(θ) is real Gaussian noises satisfying 〈η(θ)〉η=0, and 〈η(θ)η(θ′)〉η=2δ(θ−θ′). Since the action S(x) is complex, the right-hand side of Eq. (2) is also complex. Thus, complexification of the variable z is unavoidable, which is the reason why this method is called the complex Langevin method. In the real Langevin method i.e., if S(x) is real, the ensemble average 〈O(xη(θ))〉η can be shown to converge to Eq. (1) as θ→∞. It has been shown that the complex Langevin method also converges to Eq. (1) for the complex action S(x) when the tail of distribution obtained from Eq. (2) damps exponentially fast [27,28]. However, it has not been understood yet that what behavior is required to actions for the success of the complex Langevin simulation. We first show based on the semiclassical analysis that the complex Langevin method gives wrong results if there are several dominant saddle points with different complex phases. After that, we propose a new prescription to evade this breakdown.Ito calculus shows the following: If the expectation value of a holomorphic operator O˜(zη(θ)) converges as θ→∞, the derivative of O˜(zη(θ)), O(zη(θ))=O˜′(zη(θ)), must satisfy the Dyson–Schwinger (DS) equation,(3)〈O(zη)S′(zη)〉η=ħ〈O′(zη)〉η. Here the argument θ=∞ of z is omitted. When ħ=0, the DS equation can be solved by complex saddle points S′(zσ)=0 (σ∈Σ). Even at finite ħ, the contour integral on the steepest descent path Jσ around each zσ solves the DS equation [36–38],(4)∫Jσdze−S(z)/ħO(z)S′(zη)=∫Jσdze−S(z)/ħħO′(z). In general, any solutions of the DS equation are represented by a linear combination of the contour integrals on the steepest descent paths [38]. Therefore, ensemble average of a holomorphic operator O(zη) at θ→∞ can be represented as(5)〈O(zη)〉η=1Z∑σ∈Σdσ∫Jσdze−S(z)/ħO(z). Here, dσ is a complex number. The Lefschetz-thimble method [9,10,39] is useful to connect those steepest descent integrals with the original one (1). If and only if dσ is an intersection number 〈Kσ,R〉 between a steepest ascent path Kσ and the original contour R, the original integral (1) is recovered.We analyze Eq. (5) in the semiclassical limit ħ→+0. For this, we expand S(z) around each complex saddle point zσ as(6)S(zσ+δz)=Sσ+ωσ2δz2+O(δz3). Let us first analyze the left hand side of Eq. (5). If Reωσ>0, the solution of the equation of motion (2) can converge into zσ as θ→∞ in ħ→0. On the other hand, it cannot converge for Reωσ≤0. Then in the semiclassical approximation, we have(7)〈O(zη)〉η≃∑σcσO(zσ), where cσ≥0, and cσ=0 if Reωσ≤0. Next let us analyze the right hand side of Eq. (5). In the semiclassical approximation, the integral along the thimble becomes(8)∫Jσdze−S(z)O(z)≃2πħωσe−Sσ/ħO(zσ). Also the denominator Z≃Zsemi can be evaluated using the semiclassical analysis by setting O(zσ)=1 in the above discussions. Now, from the comparison of the both sides of Eq. (5) for arbitrary operators, we reach(9)cσ=1Zsemi2πħωσe−Sσ/ħdσ for dominantly contributing saddle points. Note that cσ≥0. However, dσ needs to be an integer 〈Kσ,R〉 to recover (1). These two statements contradict with one another in general. As a result, we can conclude the following at least for semiclassical analysis: The complex Langevin method cannot reproduce the original integral (1) if there are several dominant saddle points with different complex phases.Equation (9) is completely sure for dominant saddle points. The above contradiction must be taken into account if the dominant saddle points have different complex phases. In other words, complex Langevin method may fail if there is some relative phase between the dominant saddle points. For subdominant saddle points, the ambiguity of Borel resummation of large order perturbations can give nontrivial cancellations [40–47]. Therefore, we cannot judge from our argument whether the complex Langevin method gives a correct result if there is only one dominant saddle point. This subtlety needs further studies. For a Gaussian action, it is easy to check that Eq. (9) is satisfied and the complex Langevin method works well. On the other hand, there exists a model with power-law tail, where the complex Langevin method does not work but there is only one dominant saddle point (see, e.g., [33]).Let us give a few comments on previous studies. There is a formal proof [27,28] on the correctness of the complex Langevin method, but it relies on several nontrivial assumptions.11One of the most nontrivial assumptions would be the C0 semigroup property generated by Fokker–Planck-type partial differential operators. Combined with a recent study [34], they have shown that the formal proof breaks down if the complex Langevin distribution does not decay exponentially fast around infinities and singular points. Our analysis suggests without accessing details of the complex Langevin distribution that the breakdown happens if the dominant complex saddle points have different phases. Even in many-body systems, we can obtain saddle points by numerically solving Eq. (2) without random noises. This situation would naturally bring us to the conjecture that the complex Langevin distribution has a polynomial tail around infinities or singular points if several dominant saddle points contribute with different phases. It would be an important future study to check this conjecture in order to achieve a deeper understanding of the complex Langevin method.3PrescriptionLet us propose a prescription to circumvent this inconsistency. We denote the equilibrium distribution of the complex Langevin method by P. The expectation value is given as(10)〈O(zη)〉η=∫dxdyP(x,y)O(x+iy). In the semiclassical limit, P will be represented, by using a sum of distributions Pσ localized at complex saddle points zσ, as P=∑σPσ, which gives the expectation value (7) i.e.,(11)∫dxdyPσ(x,y)O(x+iy)≃cσO(zσ). By defining (nonholomorphic) functions χσ satisfying χσPτ≃δστPτ, we define a phase function Θ by(12)Θ(z,z‾):=1Z∑σ1cσ〈Kσ,R〉2πħωσe−Sσ/ħχσ(z,z‾). If Pσ does not overlap with others, χσ can be chosen as the characteristic function of supp(Pσ). This is not true in general, and we must find χσ satisfying the condition with a good approximation. The expectation value of a holomorphic operator O(z) is given, by reweighting with Θ, as(13)〈Θ(zη,z‾η)O(zη)〉η/〈Θ(zη,z‾η)〉η. Even when the random noise or equivalently ħ correction is included, so long as P is well localized around each saddle point, this prescription seems to work nicely. Note that this replacement does not break DS equations (3) so far as the semiclassical analysis is valid.Now our question is “What dσ, or cσ, is adopted in the complex Langevin method?” If Reωσ≤0 means 〈Kσ,R〉=0, the following dσ is consistent with cσ≥0:(14)dσ=ωσ|ωσ|eiImSσ/ħ〈Kσ,R〉. If this is true, the complex Langevin method gives an extension of the so-called phase quenched approximation to include complex saddle points:(15)〈O(z)〉η≃1Z∑σ〈Kσ,R〉2πħ|ωσ|e−ReSσ/ħO(zσ). We adopt it as a working hypothesis in the following sections, although this is not the unique solution for consistency. Using this hypothesis, the phase function Θ is given by(16)Θ(z,z‾):=1Z∑σ|ωσ|ωσe−iImSσ/ħχσ(z,z‾), and the reweighting formula (13) is available for practical use.22A similar improvement of complex Langevin method by reweighting with saddle-point phases has been discussed in Ref. [48].4Numerical simulationWe test our proposal by applying it to two models with and without a singular drift term. We numerically solved (2) with the fictitious time step ε=5.0×10−6 and 5.0×10−7 for the models with and without the singular drift term, respectively. We adopted a higher order algorithm [49]. Errors were estimated by using the jackknife method, and each quantity is computed by using 5.0×105 configurations. Below we set ħ=1.4.1One-site fermion modelFirst, as a nontrivial example with a singular drift term, we analyze a one-site fermion model. This is the simplest model to suffer from the sign problem same as that in lattice QCD simulations [50–53]. To demonstrate that the modified complex Langevin method can simulate the Silver Blaze like feature [54] in the one-site model is a good landmark to show its applicability to the sign problem in many-body systems.After introducing a Hubbard–Stratonovich field φ, we explicitly integrate out the original fermionic fields. The partition function reads [50](17)Z=∫dφbge−S(φbg), with the action,(18)S(φbg)=β2Uφbg2−2ln(1+eβ(iφbg+μ+U/2)), where φbg=∫0βdτφ(τ)/β is the zero Matsubara mode of φ. U(>0), μ, and β=1/T are the on-site repulsive interaction, chemical potential, and inverse temperature, respectively. We dropped nonzero Matsubara modes of φ, since they do not couple to μ [50]. The auxiliary field φbg is related to the fermion number density n by(19)n=−〈∂S/∂(βμ)〉=Im[〈φbg〉/U], where we used the equation of motion to obtain the last expression. The integral (17) is analytically calculable, but instead, we shall apply the complex Langevin method. Due to the logarithmic term, the action has infinitely many saddle points, which appear in the period of 2πT. Since the Lefschetz-thimble method is valid even with these logarithmic singularities [55,56], all the discussions in previous sections are available in order to conclude the failure of the complex Langevin method.In the large βU limit, the saddle points φm are given as [50](20)φm=i(μ+U2)+T(2πm+iln32U−μ12U+μ)+O(T2) with m∈Z. The classical action at φm reads [50](21)S0≃−βU2(μU+12)2,(22)Re[Sm−S0]≃2π2βUm2,(23)ImSm≃2πm(μU+12). In Eqs. (22) and (23), we have calculated only the m-dependent leading terms in the large βU expansion. In this model, β−1 plays a role of ħ but the classical action (18) depends on it in a nontrivial way. Therefore, Re[Sm] becomes a function of β and m, which make difficult to judge the dominance of saddle points. According to Eq. (22), saddle points with m2≲βU would give dominant contributions in the large βU limit. Thus, the zero temperature limit, which corresponds to the classical limit in Sec. 2, is not described by the unique saddle point and the condition dσ=〈Kσ,R〉 is not trivially recovered. According to Eq. (23), these different saddle points have different complex phases, and thus the complex Langevin simulation may fail except for special cases.We show the fermion number n as a function of the chemical potential μ in Fig. 1. The standard complex Langevin method predicts the wrong linear μ-dependence. This wrong behavior is also obtained from the mean field or the one-thimble approximation [50]. To find a reweighting factor, we use approximate expressions on the saddle points in the leading order of the large βU expansion in Eqs. (21)–(23) [50]. The saddle points are in between singular points of the logarithm φms=i(μ+U2)+T(2πm+1/2), namely, Reφms<Reφm<Reφm+1s. The distribution generated by solving the complex Langevin equation is localized around φm and decays by a power law as it is getting close to φms along the real part direction. For the imaginary part direction the distribution exponentially decays. Then we put χm(φbg,φ‾bg)=θ(Reφms<Reφbg<Reφm+1s). The residual sign coming from ωσ turns out to be negligible for βU=30. Now Θ is given explicitly as(24)Θ=∑m∈Ze−2πi(μ/U+1/2)mθ(2m−1<βReφbg/π<2m+1), where θ(x) is the step function. We also show the fermion number after reweighting in Fig. 1. The result becomes much better, but we may still need an improvement of Θ for exact agreement. The number density seems to linearly decrease in each plateaux as chemical potential increases. This behavior is incorrect from the view point of the thermodynamics stability since the compressibility must be non-negative. There might exist the physics not included in our weighting prescription.We show the average phase factor 〈Θ〉η as a function of the chemical potential μ in Fig. 2. It becomes small near jumping points of n at μ/U=0 and 1, and is getting close to one near the half filling μ/U=1/2. If we apply the conventional reweighting by the Monte Carlo method to the original integral (17), however, the severe sign problem appears for every μ/U>−1/2 [50]. The cancellation of the phase function in the modified complex Langevin method is milder than that in the reweighting by the Monte Carlo method. The same cancellation may happen near phase transition points in many-body systems. If 〈Θ〉η becomes exponentially small as the system size increases, it is also true that the sign problem is still obstinate in the complex Langevin method.4.2Double-well potential modelNext, we consider a model without a singular drift term, whose action is given by(25)S(x)=x4/4−x2/2−iαx, with α>0. This action has three saddle points on the complex plane. Only two of them have positive Reωσ, and contribute to the semiclassical analysis. These two saddle points (z1 and z2) are, respectively, located on the first and second quadrant planes (Rez1>0 and Rez2<0). They have different complex phases, except when α=0. The complex Langevin simulation may fail at finite α from our semiclassical analysis given in Sec. 2.The distribution of this model seems to have the power law behavior. We show the partially integrated distribution(26)Py(y)=∫dxP(x,y)/∫dxdyP(x,y), and its fifth moment y5Py in Fig. 3. The distribution may behave as Py∼y−5 at y→∞. The power law implies that the expectation value of a higher power of z, e.g., zn (n≥4) diverges.33Recently, it is mathematically shown that the power law is always true for any polynomial model if we use the complex noise instead of the real one [57]. For the real noise, it seems to depend on a model whether the distribution shows the power law. Thus the complex Langevin simulation apparently breaks down, as expected. Remark here that since zn (n≥1) does not satisfy the DS equation (3) if the power law exponent is true, our argument based on the DS equation in Sec. 2 is no longer available. Nevertheless the complex Langevin method actually breaks down, and our prescription works well for lower dimensional operators as is seen in the following.We show the expectation values of iz and (iz)2 as a function of α in Figs. 4 and 5. The complex Langevin simulation converges to a wrong result (red squares). Based on our prescription, we put χ1(z,z‾)=θ(Rez) and χ2(z,z‾)=θ(−Rez). The result of the reweighting is also shown in Figs. 4 and 5 with blue circles. The reweighting works perfectly, and we resolve the wrong convergence problem. This is also true for (iz)3. For a diverging higher power of z, (iz)n (n≥4), our prescription does not work, and the expectation values suffer from the large fluctuations before and after reweighting.5Concluding remarksWe have analytically shown within the semiclassical approximation that complex Langevin method gives wrong results, when there are several dominant saddle points with different complex phases. Since the interference of these complex phases is an essential ingredient to understand the Silver Blaze phenomenon [50], the usual complex Langevin method might not be reliable in order to tackle the cold and dense nuclear matters. Moreover, this interference is also of great importance in order to study the dynamical phenomena, such as a particle production, using the real-time path integral [58–60]. For more general situation where the semiclassical analysis breaks down, we need further study to show the failure of the complex Langevin method.The next problem is to modify the distribution so as to reproduce the expectation values in the original theory. We proposed a reweighting prescription by introducing a working hypothesis, which is consistent with the semiclassical analysis. This must be justified or revised in future study. Also the correct treatment of subdominant saddle points must be clarified. Our prescription is numerically confirmed for two models with and without singular drift terms. In particular, we succeeded to simulate the nonanalytic behavior of the one-site fermion model at low temperatures.If our prescription were proven or revised, the modified complex Langevin method could provide a way to perform numerical simulations on multiple Lefschetz thimbles. However, it requires us to get complete knowledge on complex saddle points to assign correct phase function. Furthermore, our prescription causes a large cancellation of relative phases among saddle points, although it is somewhat milder than that of the conventional reweighting by the Monte Carlo method. This implies the sign problem possibly occurs in the modified complex Langevin method. To find more efficient prescription must be an important future study.AcknowledgementsT.H. thanks A. Yamamoto for stimulating discussions. Y.T. was supported by Grants-in-Aid for the fellowship of Japan Society for the Promotion of Science (JSPS) (No. 25-6615) when he belonged to the University of Tokyo, and is supported by Special Postdoctoral Researchers Program of RIKEN. Y.H. is partially supported by JSPS KAKENHI Grants Numbers 15H03652. This work was partially supported by the RIKEN interdisciplinary Theoretical Science (iTHES) project, and by the Program for Leading Graduate Schools of Ministry of Education, Culture, Sports, Science, and Technology (MEXT), Japan.References[1]S.MuroyaA.NakamuraC.NonakaT.TakaishiLattice QCD at finite density: an introductory reviewProg. Theor. Phys.1102003615668arXiv:hep-lat/0306031[2]G.AartsComplex Langevin dynamics and other approaches at finite chemical potentialPoSLATTICE20122012017arXiv:1302.3028 [hep-lat][3]M.G.AlfordA.SchmittK.RajagopalT.SchäferColor superconductivity in dense quark matterRev. Mod. Phys.80200814551515[4]K.FukushimaT.HatsudaThe phase diagram of dense QCDRep. Prog. Phys.742011014001arXiv:1005.4814 [hep-ph][5]E.DagottoCorrelated electrons in high-temperature superconductorsRev. Mod. Phys.661994763840[6]A.W.SandvikComputational studies of quantum spin systemsAIP Conf. Proc.129712010135338http://scitation.aip.org/content/aip/proceeding/aipcp/10.1063/1.3518900[7]L.PolletRecent developments in quantum Monte Carlo simulations with applications for cold gasesRep. Prog. Phys.752012094501arXiv:1206.0781 [cond-mat.quant-gas][8]F.PhamVanishing homologies and the n variable saddlepoint methodProc. Symp. Pure Math.vol. 40(2)1983AMS319333[9]E.WittenAnalytic continuation of Chern–Simons theoryChern–Simons Gauge Theory: 20 Years AfterAMS/IP Stud. Adv. Math.vol. 502010347446arXiv:1001.2933 [hep-th][10]E.WittenA new look at the path integral of quantum mechanicsarXiv:1009.6032 [hep-th][11]AuroraScience CollaborationM.CristoforettiF.Di RenzoL.ScorzatoNew approach to the sign problem in quantum field theories: high density QCD on a Lefschetz thimblePhys. Rev. D862012074506arXiv:1205.3996 [hep-lat][12]A.MukherjeeM.CristoforettiLefschetz thimble Monte Carlo for many body theories: application to the repulsive Hubbard model away from half fillingPhys. Rev. B902014035134arXiv:1403.5680 [cond-mat.str-el][13]G.AartsLefschetz thimbles and stochastic quantisation: complex actions in the complex planePhys. Rev. D882013094501arXiv:1308.4811 [hep-lat][14]H.FujiiD.HondaM.KatoY.KikukawaS.KomatsuT.SanoHybrid Monte Carlo on Lefschetz thimbles – a study of the residual sign problemJ. High Energy Phys.13102013147arXiv:1309.4371 [hep-lat][15]F.Di RenzoG.EruzziThimble regularization at work: from toy models to chiral random matrix theoriesPhys. Rev. D922015085030arXiv:1507.03858 [hep-lat][16]K.FukushimaY.TanizakiHamilton dynamics for the Lefschetz thimble integration akin to the complex Langevin methodarXiv:1507.07351 [hep-th][17]G.ParisiY.-s.WuPerturbation theory without gauge fixingSci. Sin.241981483[18]J.R.KlauderCoherent-state Langevin equations for canonical quantum systems with applications to the quantized hall effectPhys. Rev. A29198420362047[19]G.ParisiOn complex probabilitiesPhys. Lett. B1311983393395[20]P.H.DamgaardH.HuffelStochastic quantizationPhys. Rep.1521987227[21]F.KarschH.W.WyldComplex Langevin simulation of the SU(3) spin model with nonzero chemical potentialPhys. Rev. Lett.5519852242[22]J.AmbjornM.FlensburgC.PetersonThe complex Langevin equation and Monte Carlo simulations of actions with static chargesNucl. Phys. B2751986375[23]G.AartsI.-O.StamatescuStochastic quantization at finite chemical potentialJ. High Energy Phys.092008018arXiv:0807.1597 [hep-lat][24]G.AartsCan stochastic quantization evade the sign problem? The relativistic Bose gas at finite chemical potentialPhys. Rev. Lett.1022009131601arXiv:0810.2089 [hep-lat][25]D.SextySimulating full QCD at nonzero density using the complex Langevin equationPhys. Lett. B7292014108111arXiv:1307.7748 [hep-lat][26]T.HayataA.YamamotoComplex Langevin simulation of quantum vortex nucleation in the Bose–Einstein condensatearXiv:1411.5195 [cond-mat.quant-gas][27]G.AartsE.SeilerI.-O.StamatescuThe complex Langevin method: when can it be trusted?Phys. Rev. D812010054508arXiv:0912.3360 [hep-lat][28]G.AartsF.A.JamesE.SeilerI.-O.StamatescuComplex Langevin: etiology and diagnostics of its main problemEur. Phys. J. C7120111756arXiv:1101.3270 [hep-lat][29]G.AartsF.A.JamesOn the convergence of complex Langevin dynamics: the three-dimensional XY model at finite chemical potentialJ. High Energy Phys.082010020arXiv:1005.3468 [hep-lat][30]J.M.PawlowskiC.ZielinskiThirring model at finite density in 0+1 dimensions with stochastic quantization: crosscheck with an exact solutionPhys. Rev. D872013094503arXiv:1302.1622 [hep-lat][31]G.AartsP.GiudiceE.SeilerLocalised distributions and criteria for correctness in complex Langevin dynamicsAnn. Phys.3372013238260arXiv:1306.3075 [hep-lat][32]A.MollgaardK.SplittorffComplex Langevin dynamics for chiral random matrix theoryPhys. Rev. D882013116007arXiv:1309.4335 [hep-lat][33]G.AartsL.BongiovanniE.SeilerD.SextySome remarks on Lefschetz thimbles and complex Langevin dynamicsJ. High Energy Phys.14102014159arXiv:1407.2090 [hep-lat][34]J.NishimuraS.ShimasakiNew insights into the problem with a singular drift term in the complex Langevin methodPhys. Rev. D922015011501arXiv:1504.08359 [hep-lat][35]S.TsutsuiT.M.DoiAn improvement in complex Langevin dynamics from a view point of Lefschetz thimblesarXiv:1508.04231 [hep-lat][36]C.PehlevanG.GuralnikComplex Langevin equations and Schwinger–Dyson equationsNucl. Phys. B8112009519536arXiv:0710.3756 [hep-th][37]G.GuralnikC.PehlevanEffective potential for complex Langevin equationsNucl. Phys. B8222009349366arXiv:0902.1503 [hep-lat][38]G.GuralnikZ.GuralnikComplexified path integrals and the phases of quantum field theoryAnn. Phys.325201024862498arXiv:0710.1256 [hep-th][39]Y.TanizakiT.KoikeReal-time Feynman path integral with Picard–Lefschetz theory and its applications to quantum tunnelingAnn. Phys.3512014250arXiv:1406.2386 [math-ph][40]E.B.BogomolnyCalculation of instanton-anti-instanton contributions in quantum mechanicsPhys. Lett. B913–41980431435[41]J.Zinn-JustinMulti-instanton contributions in quantum mechanicsNucl. Phys. B19211981125140[42]G.BasarG.V.DunneM.UnsalResurgence theory, ghost-instantons, and analytic continuation of path integralsJ. High Energy Phys.13102013041arXiv:1308.1108 [hep-th][43]A.ChermanD.DorigoniM.UnsalDecoding perturbation theory using resurgence: Stokes phenomena, new saddle points and Lefschetz thimblesarXiv:1403.1277 [hep-th][44]D.DorigoniAn introduction to resurgence, trans-series and alien calculusarXiv:1411.3585 [hep-th][45]A.BehtashT.SulejmanpasicT.SchäferM.ÜnsalHidden topological angles in path integralsPhys. Rev. Lett.11542015041601arXiv:1502.06624 [hep-th][46]A.BehtashE.PoppitzT.SulejmanpasicM.ÜnsalThe curious incident of multi-instantons and the necessity of Lefschetz thimblesarXiv:1507.04063 [hep-th][47]A.BehtashG.V.DunneT.SchaeferT.SulejmanpasicM.UnsalToward Picard–Lefschetz theory of path integrals, complex saddles and resurgencearXiv:1510.03435 [hep-th][48]H. Fujii, Y. Kikukawa, T. Sano, unpublished.[49]G.AartsF.A.JamesComplex Langevin dynamics in the SU(3) spin model at nonzero chemical potential revisitedJ. High Energy Phys.012012118arXiv:1112.4655 [hep-lat][50]Y.TanizakiY.HidakaT.HayataLefschetz-thimble analysis of the sign problem in one-site fermion modelNew J. Phys.182016033002arXiv:1509.07146 [hep-th][51]H.FujiiS.KamataY.KikukawaLefschetz thimble structure in one-dimensional lattice Thirring model at finite densityarXiv:1509.08176 [hep-lat][52]H.FujiiS.KamataY.KikukawaMonte Carlo study of Lefschetz thimble structure in one-dimensional Thirring model at finite densityarXiv:1509.09141 [hep-lat][53]A.AlexandruG.BasarP.BedaqueA Monte Carlo algorithm for simulating fermions on Lefschetz thimblesarXiv:1510.03258 [hep-lat][54]T.D.CohenFunctional integrals for QCD at nonzero chemical potential and zero densityPhys. Rev. Lett.912003222001arXiv:hep-ph/0307089[55]Y.TanizakiLefschetz-thimble techniques for path integral of zero-dimensional O(n) sigma modelsPhys. Rev. D912015036002arXiv:1412.1891 [hep-th][56]T.KanazawaY.TanizakiStructure of Lefschetz thimbles in simple fermionic systemsJ. High Energy Phys.15032015044arXiv:1412.2802 [hep-th][57]D.P.HerzogJ.C.MattinglyNoise-induced stabilization of planar flows IarXiv:1404.0957 [math.PR][58]C.K.DumluG.V.DunneThe Stokes phenomenon and Schwinger vacuum pair production in time-dependent laser pulsesPhys. Rev. Lett.1042010250402arXiv:1004.2509 [hep-th][59]C.K.DumluG.V.DunneInterference effects in Schwinger vacuum pair production for time-dependent laser pulsesPhys. Rev. D832011065028arXiv:1102.2899 [hep-th][60]C.K.DumluG.V.DunneComplex worldline instantons and quantum interference in vacuum pair productionPhys. Rev. D842011125023arXiv:1110.1657 [hep-th]