Applications of quantum Monte Carlo methods in condensed systems JindˇrichKolorenˇc Institute of Physics, Academy of Sciences of the Czech Republic, Na Slovance 2, 18221 Praha 8, Czech Republic I. Institut f¨ur Theoretische Physik, Universit¨atHamburg, Jungiusstraße 9, 20355 Hamburg, Germany [email protected] Lubos Mitas Department of Physics and Center for High Performance Simulation, North Carolina State University, Raleigh, North Carolina 27695, USA [email protected]
The quantum Monte Carlo methods represent a powerful and broadly applicable computa- tional tool for finding very accurate solutions of the stationary Schr¨odingerequation for atoms, molecules, solids and a variety of model systems. The algorithms are intrinsically parallel and are able to take full advantage of the present-day high-performance computing systems. This review article concentrates on the fixed-node/fixed-phase diffusion Monte Carlo method with emphasis on its applications to electronic structure of solids and other extended many-particle systems.
PACS numbers: 02.70.Ss, 71.15. m, 31.15. p − − To appear in Rep. Prog. Phys.
1 Introduction 1 4 Trial wave functions 15 1.1 Many-body stationary Schr¨odingerequation . . . 3 4.1 Elementary properties ...... 15 4.2 Jastrow factor ...... 16 2 Methods 4 4.3 Slater–Jastrow wave function ...... 16 2.1 Variational Monte Carlo ...... 4 4.4 Antisymmetric forms with pair correlations . . . 18 2.2 Diffusion Monte Carlo ...... 5 4.5 Backflow coordinates ...... 18 2.2.1 Fixed-node/fixed-phase approximation . .6 2.2.2 Sampling the probability distribution . . 6 5 Applications 19 2.2.3 General expectation values ...... 9 5.1 Properties of the homogeneous electron gas . . . 19 2.2.4 Spin degrees of freedom ...... 9 5.2 Cohesive energies of solids ...... 20 2.3 Pseudopotentials ...... 9 5.3 Equations of state ...... 21 3 From a finite supercell to the thermodynamic 5.4 Phase transitions ...... 22 limit 10 5.5 Lattice defects ...... 23 3.1 Twist-averaged boundary conditions ...... 11 5.6 Surface phenomena ...... 23 3.2 Ewald formula ...... 12 5.7 Excited states ...... 24 3.3 Extrapolation to the thermodynamic limit . . . . 13 5.8 BCS–BEC crossover ...... 25 3.4 An alternative model for Coulomb interaction energy ...... 14 6 Concluding remarks 26
1 Introduction The task of solving the Schr¨odingerequation for Many properties of condensed matter systems can be systems of electrons and ions, and predicting the quan- tities of interest such as cohesion and binding energies, arXiv:1010.4992v1 [physics.comp-ph] 24 Oct 2010 calculated from solutions of the stationary Schr¨odinger equation describing interacting ions and electrons. The electronic gaps, crystal structures, variety of magnetic grand challenge of solving the Schr¨odingerequation phases or formation of quantum condensates is noth- has been around from the dawn of quantum mechanics ing short of formidable. Paul Dirac recognized this and remains at the forefront of the condensed matter state of affairs already in 1929: “The general theory of physics today and, undoubtedly, for many decades to quantum mechanics is now almost complete . . . The come. underlying physical laws necessary for the mathemat-
1 ical theory of a large part of physics and the whole systems viable, allowing predictions that would be dif- chemistry are thus completely known, and the difficulty ficult or impossible to make otherwise. The quantum is only that the exact application of these laws leads Monte Carlo (QMC) methods described in this review to equations much too complicated to be soluble.”[1] provide an interesting illustration of what is currently Arguably, this is the most fundamental approach to possible and how much the computational methods the physics of condensed matter: Applications of the can enrich and make more precise our understanding rigorous quantum laws to models that are as close to of the quantum world. reality as currently feasible. Some of the ideas used in QMC methods go back The goal of finding accurate solutions for stationary to the times before the invention of electronic comput- quantum states is hampered by a number of difficulties ers. Already in 1930s Enrico Fermi noticed similarities inherent to many-body quantum systems: between the imaginary time Schr¨odingerequation and (i) Even moderately sized model systems contain the laws governing stochastic processes in statistical anywhere between tens to thousands of quantum mechanics. In addition, based on memories of his col- particles. Moreover, we are often interested in laborator Emilio Segr`e, Fermi also envisioned stochas- expectation values in the thermodynamic limit tic methodologies for solving the Schr¨odingerequa- that is usually reached by extrapolations from tion, which were very similar to concepts developed finite sizes. Such procedures typically require decades later. These Fermi’s ideas were acknowledged detailed information about the scaling of the by Metropolis and Ulam in a paper from 1949 [2], quantities of interest with the system size. where they outlined a stochastic approach to solv- ing various physical problems and discussed merits of (ii) Quantum particles interact and the interactions “modern” computers for its implementation. In fact, affect the nature of quantum states. In many this group of scientists at the Los Alamos National Lab- cases, the influence is profound. oratory attempted to calculate the hydrogen molecule by a simple version of QMC in the early 1950s, around (iii) The solutions have to conform to quantum sym- the same time when a pioneering work on the first metries such as the fermionic antisymmetry Monte Carlo study of classical systems was published linked to the Pauli exclusion principle. This is by Metropolis and coworkers [3]. In the late 1950s, a fundamental departure from classical systems Kalos initiated development of QMC simulations and and poses different challenges which call for new methodologies for few-particle systems and laid down analytical ideas and computational strategies. the statistical and mathematical foundations of the Green’s function Monte Carlo method [4]. Eventually, (iv) For meaningful comparisons with experiments, simulations of large many-particle systems became the required accuracy is exceedingly high, espe- practicable as well. First came studies of bosonic cially when comparing with precise data from fluids modelling 4He [5–7], and later followed investi- spectroscopic and low-temperature studies. gations of extended fermionic systems exemplified by 3 In the past, the most successful approaches to ad- liquid He [8, 9] and by the homogeneous electron gas dress these challenges were based mostly on reduction- [10, 11]. Besides these applications to condensed mat- ist ideas. The problem is divided into the dominant ter, essentially the same methods were in mid-seventies effects, which are treated explicitly, and the rest, which introduced in quantum chemistry to study small molec- is then dealt with by approximate methods based on ular systems [12, 13]. To date, various QMC methods variety of analytical tools: perturbation expansions, were developed and applied to the electronic structure mean-field methods, approximate transformations to of atoms, molecules and solids, to quantum lattice known solutions, and so on. The reductionist ap- models, as well as to nuclear and other systems with proaches have been gradually developed into a high contributions from many scientists. level of sophistication and despite their limitations, The term “quantum Monte Carlo” covers several they are still the most commonly used strategies in related stochastic methodologies adapted to determine many-body physics. ground-state, excited-state or finite-temperature equi- The progress in computer technology has opened a librium properties of a variety of quantum systems. new avenue for studies of quantum (and many other) The word “quantum” is important since QMC ap- problems and has enabled researchers to obtain re- proaches differ significantly from Monte Carlo methods sults beyond the scope of analytic many-body theories. for classical systems. For an overview of the latter see The performance of current large computers makes for instance [14]. QMC is not only a computational computational investigations of many-body quantum tool for large-scale problems, but it also encompasses a
2 substantial amount of analytical work needed to make limiting factor in further increase in accuracy. As we such calculations feasible. QMC simulations often uti- will see in section 5, the fixed-node error is typically lize results of the more traditional electronic structure rather small and does not hinder calculation of robust methods in order to increase efficiency of the calcu- quantities such as cohesion, electronic gaps, optical ex- lations. These ingredients are combined to optimally citations, defect energies or potential barriers between balance the computational cost with achieved accuracy. structural conformations. By robust we mean quanti- The key point for gaining new insights is an appro- ties which are of the order of tenths of an electronvolt priate analysis of the quantum states and associated to several electronvolts. Nevertheless, the fixed-node many-body effects. It is typically approached itera- errors can bias results for more subtle phenomena, tively: Simulations indicate the gaps in understanding such as magnetic ordering or effects related to super- of the physics, closing these gaps is subsequently at- conductivity. Development of strategies to alleviate tempted and the improvements are assessed in the next such biases is an active area of research. round. Such a process involves construction of zero- Fixed-node DMC simulations are computationally or first-order approximations for the desired quantum rather demanding when compared to the mainstream states, incorporation of new analytical insights, and electronic structure methods that rely on mean-field development of new numerical algorithms. treatment of electron-electron interactions. On the QMC methods inherently incorporate several types other hand, QMC calculations can provide unique in- of internal checks, and many of the algorithms used sights into the nature of quantum phenomena and possess various rigorous bounds, such as the varia- can verify many theoretical ideas. As such, they can tional property of the total energy. Nevertheless, the produce not only accurate numbers but also new un- coding and numerical aspects of the simulations are derstanding. Indeed, QMC methodology is very much not entirely error-proof and the obtained results should an example of “it from bit” paradigm, alongside, for be verified independently. Indeed, it is a part of the example, the substantial computational efforts in quan- modern computational-science practice that several tum chromodynamics, which not only predict hadron groups revisit the same problem with independent masses but also contribute to the validation of the software packages and confirm or challenge the results. fundamental theory [20, 21]. Just a few decades ago it “Biodiversity” of the available QMC codes on the scien- was difficult to imagine that one would be able to solve tific market (including QWalk [15], QMCPACK [16], the Schr¨odinger equation for hundreds of electrons by CHAMP [17], CASINO [18], QMcBeaver [19] and oth- means of an explicit construction of the many-body ers) provides the important alternatives to verify the wave function. Today, such calculations are feasible algorithms and their implementations. This is clearly a using available computational resources. At the same rather labourious, slow and tedious process, neverthe- time, there remains more to be done to make the less, experience shows that independently calculated methods more insightful, more efficient, and their ap- results and predictions eventually reach a consensus plication less labourious. We hope that this review and such verified data become widely used standards. will contribute to the growing interest in this rapidly developing field of research. In this overview we present QMC methods that The review is organized as follows: The remainder solve the stationary Schr¨odingerequation for con- of this section provides mostly definitions and nota- densed systems of interacting fermions in continuous tions. Section 2 follows with description of the VMC space. Conceptually very straightforward is the varia- and DMC methods. The strategies for calculation of tional Monte Carlo (VMC) method, which builds on quantities in the thermodynamic limit are presented explicit construction of trial (variational) wave func- in section 3. Section 4 introduces currently used forms tions using stochastic integration and parameter opti- of the trial wave functions and their recently devel- mization techniques. More advanced approaches rep- oped generalizations. The overview of applications resented by the diffusion Monte Carlo (DMC) method presented in section 5 is focused on QMC calculations are based on projection operators that find the ground of a variety of solids and related topics. state within a given symmetry class. Practical versions of the DMC method for a large number of particles 1.1 Many-body stationary Schr¨odinger require dealing with the well-known fermion sign prob- equation lem originating in the antisymmetry of the fermionic wave functions. The most commonly used approach to Let us consider a system of quantum particles such as overcome this fundamental obstacle is the fixed-node electrons and ions interacting via Coulomb potentials. approximation. This approximation introduces the Since the masses of nuclei and electrons differ by three so-called fixed-node error, which appears to be the key orders of magnitude or more, the problem can be sim-
3 plified with the aid of the Born–Oppenheimer approxi- are relevant for our discussion of quantum Monte Carlo mation, which separates the electronic degrees of free- methodology, since the latter uses the results of these dom from the slowly moving system of ions. The elec- approaches as a reference and also for construction tronic part of the non-relativistic Born–Oppenheimer of the many-body wave functions. Familiarity with hamiltonian is given by the basic concepts of the Hartree–Fock and density- functional theories is likely to make the subsequent ˆ 1 X 2 X ZI X 1 sections easier to follow, but we believe that it is not H = i + , (1) −2 ∇ − ri xI ri rj a necessary prerequisite for understanding our exposi- i i,I | − | jground state Ψ0 . and by formulating the Hartree–Fock (HF) theory, | ⟩ which correctly takes into account the Pauli exclu- Wave functions of interacting systems are non- separable, and the integration needed to evaluate EΨ2 sion principle [24, 25]. The HF theory replaces the T hard problem of many interacting electrons with a is therefore a difficult task. Although it is possible to system of independent particles in an effective, self- write these wave functions as linear combinations of consistent field. The theory was further developed by separable terms, this tactic is viable only for a limited Slater [26] and others, and it has become a starting number of particles, since the length of such expansions point of many sophisticated approaches to fermionic grows very quickly as the system size increases. The many-body problems. variational Monte Carlo method employs a stochastic For periodic systems, the effective free-electron the- integration that can treat the non-separable wave func- tions directly. The expectation value EΨ2 is written ory and the band theory of Bloch [27] were the first T crucial steps towards our present understanding of as the real crystals. In 1930s, Wigner and Seitz [28, 29] Z 2 ˆ ΨT( ) HΨT ( ) 3N performed the first quantitative calculations of the elec- EΨ2 = | R | R d T ΨT ΨT ΨT( ) R tronic states in sodium metal. Building upon the homo- ⟨ | ⟩ R N geneous electron gas model, the density-functional the- 1 X Hˆ ΨT ( i) EVMC = R , (3) ory (DFT) was invented by Hohenberg and Kohn [30] ≈ ΨT( i) and further developed by Kohn and Sham [31] who N i=1 R formulated the local density approximation (LDA) for where = (r1, r2,..., rN ) is a 3N-dimensional vec- the exchange-correlation functional. These ideas were tor encompassingR the coordinates of all N particles later elaborated by including spin polarization [32], by in the system and the sum runs over such vec- N constructing the generalized gradient approximation tors i sampled from the multivariate probability {R } 2 (GGA) [33, 34], and by designing a variety of orbital- density ρ( ) = ΨT( ) / ΨT ΨT . The summand R | R | ⟨ | ⟩ dependent exchange-correlation functionals [35–37]. EL( ) = Hˆ ΨT ( )/ΨT( ) is usually referred to as The DFT has proved to be very successful and has theR local energy. WeR assumeR spin-independent hamil- become the mainstream computational method for tonians, and therefore spin variables do not explic- many applications, which cover not only solids but itly enter the evaluation of the expectation value (3). also molecules and even nuclear and other systems This statement is further corroborated in section 4.1 [38, 39]. The density-functional theory together with where the elementary properties of the trial wave func- the Hartree–Fock and post-Hartree–Fock methods [40] tions ΨT are discussed. | ⟩
4 Equation (3) transforms the multidimensional in- estimation of the variance of EVMC is a non-trivial tegration into a problem of sampling a complicated affair since the random samples i generated by {R } probability distribution. The samples i can be means of the Markov chain are correlated. These cor- obtained such that they constitute a Markov{R } chain relations are not known a priori and depend on the with transitions i+1 i governed by a stochastic particular form of the transition matrix M that varies R ← R matrix M( i+1 i) whose stationary distribution from case to case. Nevertheless, it is possible to esti- coincides withR the← desired R probability density ρ( ), mate the variance without detailed knowledge of the R Z correlation properties of the chain with the aid of the ρ( ′) = M( ′ )ρ( ) d3N for all ′. (4) so-called blocking method [43]. R R ← R R R R The fluctuations of the local energy E are reduced After a period of equilibration, the members of the L as the trial wave function Ψ approaches an eigen- Markov sequence sample the stationary distribution T state of the hamiltonian, and| E⟩ becomes a constant regardless of the starting point of the chain, provided L when Ψ is an eigenstate. In particular, it is crucial the matrix M( ′ ) is ergodic. Inspired by the T to remove| ⟩ as many singularities from E as possible way the samplesR explore← R the configuration space, they L by a proper choice of the trial function. Section 4.1 il- are often referred to as walkers. lustrates how it is achieved in the case of the Coulomb The Markov chain can be conveniently constructed potential that is singular at particle coincidences. with the aid of the Metropolis method [3, 41]. The transition matrix is factorized into two parts, The total energy is not the only quantity of inter- ′ ′ ′ M( i) = T ( i)A( i), which cor- est and evaluation of other ground-state expectation respondR ← to R two consecutiveR ← R stochasticR ← R processes: A values is often desired. The formalism sketched so far ′ remains unchanged, only the local energy is replaced candidate for (i + 1)-th sample is proposed accord- R ′ by a local quantity AL( ) = AˆΨT ( )/ΨT( ) cor- ing to the probability T ( i) and this move is R R R R ← R ′ ˆ then either accepted with the probability A( i) responding to a general operator A. An important ′ R ← R or rejected with the probability 1 A( i). If difference between AL and the local energy is that − R ← R the move is accepted, the new member of the sequence fluctuations of AL do not vanish when ΨT is an ′ ˆ | ⟩ is i+1 = , otherwise it is i+1 = i. The length eigenstate of H. These fluctuations can severely im- ofR the chainR is thus incrementedR in eitherR case. The pact the efficiency of the Monte Carlo integration in ′ acceptance probability A( ), complementing ΨT Aˆ ΨT / ΨT ΨT , and the random error can decay i ⟨ | | ⟩ ⟨ | ⟩ some given T ( ′ ) andR ρ←( R) such that the sta- even slower than −1/2 [42]. The trial wave function i N tionarity conditionR ←(4) R is fulfilled,R is not unique. The cannot be altered to suppress the fluctuations in this choice corresponding to the Metropolis algorithm reads case, but a modified operator Aˆ′ can often be con- ˆ′ ˆ ′ ′ structed such that ΨT A ΨT = ΨT A ΨT while ′ T ( ) ρ( ) ⟨ | | ⟩ ⟨ | | ⟩ A( ) = min 1, R ← R R (5) the fluctuations of AL are substantially reduced [44– R ← R T ( ′ ) ρ( ) R ← R R 48]. and depends only on ratios of T and ρ. Consequently, normalization of the trial wave function ΨT is com- pletely irrelevant for the Monte Carlo evaluation| ⟩ of the quantum-mechanical expectation values. The freedom 2.2 Diffusion Monte Carlo ′ to choose the proposal probability T ( i) can be exploited to improve ergodicity of theR sampling,← R for The accuracy of the variational Monte Carlo method is instance, to make it easier to overcome a barrier of limited by the quality of the trial wave function ΨT . | ⟩ low probability density ρ separating two high-density This limitation can be overcome with the aid of the ′ regions. A generic choice for T ( i) is a Gaus- projector methods. In particular, the diffusion Monte R ← R sian distribution centered at i with its width tuned Carlo method [12, 49–51] employs an imaginary time to optimize the efficiency ofR the sampling. evolution The variational energy EVMC is a stochastic vari- able, and an appropriate characterization of the ran- ΨD(t) = exp Hˆ ET(t) t ΨT , (6) dom error EVMC E 2 is thus an integral part of the | ⟩ − − | ⟩ ΨT variational Monte− Carlo method. When the sampled local energies EL( i) are sufficiently well behaved [42], where the energy offset ET is introduced to main- R this error can be represented by the variance of EVMC. tain the wave-function norm at a fixed value. Formal In such cases, the error scales as −1/2 and is pro- properties of (6) can be elucidated by expanding the N portional to fluctuations of the local energy. Reliable trial function ΨT in terms of the hamiltonian eigen- | ⟩
5 states (2), which readily yields knowledge into constructive algorithms for the trial wave functions. The problem with the variable sign of ρ( , t) can ΨD(t) = exp E0 ET(t) t Ψ0 Ψ0 ΨT R | ⟩ − − | ⟩⟨ | ⟩ be circumvented by complementing the projection (6) ∞ with the so-called fixed-node constraint [13], X −(En−E0)t + e Ψn Ψn ΨT . (7) | ⟩⟨ | ⟩ n=1 ΨD( , t)ΨT( ) 0 for all and all t . (10) R R ≥ R The ground state Ψ0 is indeed reached in the limit of | ⟩ Doing so, limt→∞ ΨD(t) only approximates Ψ0 , large t as long as the trial function was not orthogonal since the projection| cannot⟩ entirely reach the ground| ⟩ to Ψ0 from the beginning. The requirement of a | ⟩ state if the initial wave function ΨT does not possess finite norm of ΨD(t) translates into a formula | ⟩ | ⟩ the exact nodes. The total energy calculated with this fixed-node method represents an upper-bound es- E0 = lim ET(t) (8) t→∞ timate of the true ground-state energy because the projection (6) is restricted to a subspace of the whole that can be used to obtain the ground-state energy. An Hilbert space when the constraint (10) is implemented alternative approach is to evaluate the matrix element [60–62]. The fixed-node approximation has proved E = Ψ (t) Hˆ Ψ / Ψ (t) Ψ that asymptot- ΨDΨT D T D T very fruitful in quantum chemistry [63, 64] as well as ically coincides⟨ with| | the⟩ ⟨ ground-state| ⟩ energy, since for investigation of the electronic structure of solids Ψ Hˆ Ψ / Ψ Ψ = Ψ Hˆ Ψ / Ψ Ψ . The in- 0 T 0 T 0 0 0 0 as testified by the applications reviewed in section 5. tegration⟨ | | in⟩ ⟨E | ⟩can⟨ be| performed| ⟩ ⟨ stochastically| ⟩ ΨDΨT In calculations of extended systems and espe- in analogy with the VMC method, cially metals, it is beneficial to allow for boundary conditions that break the time-reversal symmetry, Z ∗ Hˆ Ψ ( ) ΨD( , t)ΨT( ) T 3N since it facilitates reduction of finite-size effects (sec- EΨDΨT = R R R d ΨD(t) ΨT ΨT( ) R ⟨ | ⟩ R tion 3.1). The eigenfunctions are then complex-valued N 1 X and a generalization of the fixed-node approxima- EDMC = EL( i) , (9) ≈ R tion is required. The constraint (10) is replaced with i=1 iϕT N ΨD(t) = ΨD(t) e , where ϕT is the phase of the | | iϕT trial wave function ΨT = ΨT e [65]. The phase ϕT where the individual samples i are now drawn | | from a probability distribution definedR as ρ( , t) = is held constant during the DMC simulation to guar- ∗ R antee that ρ( , t) stays non-negative for all and t. ΨD( , t)ΨT( )/ ΨD(t) ΨT . R R R R ⟨ | ⟩ Additionally, a complex trial wave function ΨT causes | ⟩ 2.2.1 Fixed-node/fixed-phase approximation the local energy EL to be complex as well. The ap- propriate modification of the estimate for the total en- The Monte Carlo integration indicated in (9) is pos- ergy (9) coinciding with the asymptotic value of ET(t) sible only if ρ( , t) is real-valued and positive. Since then reads the hamiltoniansR we usually deal with are symmetric N with respect to time reversal, the eigenfunctions can 1 X EDMC = Re EL( i) . (11) be chosen real. Unfortunately, many-electron wave R i=1 functions must necessarily have alternating sign to N comply with the fermionic antisymmetry. In general, Analogous to the fixed-node approximation, the fixed- the initial guess ΨT will have different plus and mi- phase method provides a variational upper-bound es- | ⟩ nus sign domains (also referred to as nodal pockets timate of the true ground-state energy. Moreover, or nodal cells) than the sought for ground-state wave the fixed-phase approximation reduces to the fixed- function Ψ0 , which results in changing sign of ρ( , t). node approximation when applied to real-valued wave | ⟩ R In certain special cases, the correct sign structure of functions. the ground state can be deduced from symmetry con- siderations [52–54], but in a general interacting system 2.2.2 Sampling the probability distribution the exact position of the boundary between the pos- itive and negative domains (the so-called fermionic The unnormalized probability distribution that we node) is unknown and is determined by the quantum wish to sample in the fixed-phase DMC method, many-body physics [55]. A number of exact properties ∗ of the fermionic nodes have been discovered [56–59], f( , t) = ΨD( , t)ΨT( ) = ΨD( , t) ΨT( ) , but a lot remains to be done in order to transform this R R R | R || R |(12)
6 referred to as the mixed distribution, fulfills an equa- written as tion of motion ′ Gdrift( , τ) (16) R ← R ′ ′ ′ = 1 τ vD( ) δ vD( )τ ∂tf( , t) = − ∇ · R R − R − R − R + O(τ 2) , 1 2 f( , t) + vD( )f( , t) ′ − 2∇ R ∇ · R R Gdiff ( , τ) (17) h i R ← R + f( , t) Re EL( ) (1 + t ∂t)ET(t) (13) 1 ( ′)2 R R − = exp R − R , (2πτ)3N/2 − 2τ that is derived by differentiating (6) and (12) with ′ G ( , τ) (18) respect to time, combining the resulting expressions g/d R ← R h i ′ and rearranging the terms. The drift velocity vD in- = exp τ Re EL( ) ET(t) δ , − R − R − R troduced in (13) is defined as vD = ln ΨT and denotes the 3N-dimensional gradient∇ with| respect| and correspond to the three non-commuting oper- to∇ . The equation of motion is valid in this form ators from the right-hand side of (13) in the or- R 2 only as long as the kinetic energy is the sole non-local der: drift vD( ) , diffusion /2 and ∇ · R • −∇ • operator in the hamiltonian. Strategies for inclusion growth/decay Re[EL( )] (1 + t ∂t)ET(t) . The • R − of non-local pseudopotentials will be discussed later drift and diffusion Green’s functions preserve the nor- in section 2.3. The case of the fixed-node approxi- malization of f( , t) whereas the growth/decay pro- R mation is virtually identical to (13), except that the cess does not. local energy is real by itself. The following discussion The factorization of the exact Green’s function into therefore applies to both methods. the product of the short-time terms forms the basis of the stochastic process that represents the diffusion The time evolution of the mixed distribution Monte Carlo algorithm. First, samples i are f( , t) can be written in the form of a convolution M {R 2} R drawn from the distribution f( , 0) = ΨT( ) just like in the VMC method. Subsequently,R | thisR | set of Z walkers evolves such that it samples the mixed distri- f( , t) = G( ′, t)f( ′, 0) d3N ′ , (14) R R ← R R R bution f( , t) at any later time t. The probability distributionR is updated from time t to t + τ by multi- plication with the short-time Green’s function, 2 where f( , 0) = ΨT( ) and the Green’s function Z R ′ | Rˆ | ′ ′ ′ 3N ′ G( , t) = G(t) is a solution of (13) f( , t+τ) = Gst( , τ)f( , t) d , (19) withR the ← initial R condition⟨R| G|R( ⟩ ′, 0) = δ( ′). R R ← R R R R ← R R − R Making use of the Trotter–Suzuki formula [66, 67], which translates into the following procedure per- the Green’s function is approximated by a product of formed on each walker in the population: short-time expressions, ′ (i) a drift move ∆ drift = vD( )τ is proposed R R (ii) a diffusion move ∆ = χ is proposed, where ˆ ˆ ˆ ˆ M diff G(t) = Gg/d(τ) Gdiff (τ) Gdrift(τ) + O(τ) , (15) χ is a vector of GaussianR random numbers with | {z } variance τ and zero mean Gˆst(τ) (iii) the increment ∆ drift + ∆ diff is accepted if it complies withR the fixed-nodeR condition where τ denotes t/M and the exact solution of (13) Ψ ( ′)Ψ ( ′ + ∆ + ∆ ) > 0, other- is approached as this time step goes to zero. Con- T T drift diff wiseR the walkerR staysR at itsR original position; sequently, the DMC simulations should be repeated moves attempting to cross the node occur only for several sizes of the time step and an extrapola- due to inaccuracy of the approximate Green’s tion of the results to τ 0 should be performed in function (15), and they vanish in the limit τ 0; the end. For simplicity,→ we show in (15) only the the moves ∆ +∆ are accepted without→ simplest Trotter–Suzuki decomposition which has a drift diff any constraintR in the fixed-phaseR method time step error proportional to τ. Commonly used are 2 higher order approximations whose errors scale as τ (iv) the growth/decay Green’s function Gg/d is ap- or τ 3. The three new Green’s functions constituting plied; several algorithms devised for this purpose the short-time approximation Gˆst can be explicitly are outlined in the following paragraph
7 (v) at this moment, the time step is finished and the An alternative to the fluctuating population are simulation continues with another cycle starting • various flavours of the stochastic reconfigura- back at (i). tion [15, 70, 76–78]. These algorithms comple- ment each branched walker with high weighting After the projection period is completed, the algorithm factor W ( ) with one eliminated walker with samples the desired ground-state mixed distribution small W ( R), and therefore the total number of and the quantities needed for evaluation of various walkers remainsR constant. This pairing intro- expectation values can be collected in step (v). duces slight correlations into the walker popu- At this point we return to a more detailed lation that are comparable to those caused by discussion of several algorithmic representations of the population control feedback of the standard the growth/decay Green’s function Gg/d needed in branching/elimination algorithm [75]. Keeping step (iv). the population size fixed is advantageous for load balancing in parallel computational envi- The most straightforward way is to assign • ronments, since the number of walkers can be a weight w to each walker. These weights a multiple of the number of computer nodes are set to 1 during the VMC initialization (CPUs) at all times during the simulation. of the walker population and the applica- tion of Gg/d then amounts to a multiplication w(t + τ) = w(t)W ( ), where the weighting fac- R The branching/elimination process interacts in a subtle tor is way with the fixed-node constraint. Since the walk- h i ers are not allowed to cross the node, the branched W ( ) = exp τ Re EL( ) ET(t) . (20) R − R − and parent walkers always stay in the same nodal cell. If some of these cells are more favoured (that is, if Consequently, the formula for calculation of the they have a lower local energy on average), the walker total energy (11) is modified to population accumulates there and eventually vanishes from the less favoured cells. Such uneven distribution N −1 N X X of samples would introduce a bias to the simulation. EDMC = wi wi Re EL( i) (21) R i=1 i=1 Fortunately, it does not happen, since all nodal cells of the ground-state wave functions are connected by and the walkers remain distributed according particle permutations and are therefore equivalent, see 2 to ΨT( ) as in the VMC method. This algo- the tiling theorem in [56]. For general excited states rithm| isR referred| to as the pure diffusion Monte this theorem does not hold and the unwanted depopu- Carlo method [68, 69]. It is known to be intrinsi- lation of some nodal cells can indeed be observed. The cally unstable at large projection times where the problem is absent from the fixed-phase method, since signal-to-noise ratio deteriorates [70], but it is it contains no restriction on the walker propagation. still useful for small quantum-chemical systems The branching/elimination algorithm is just one [71–73]. of the options of dealing with the weights along the stochastic paths. Another possibility was introduced The standard DMC algorithm fixes the weights by Baroni and Moroni as the so-called reptation al- • to w = 1 and instead allows for stochastically gorithm [79], which recasts the sampling of both the fluctuating size of the walker population by path in the configuration space and the weight into branching walkers in regions with large weight- a straightforward Monte Carlo process, avoiding thus ing factor W ( ) and by removing them from some of the disadvantages of the DMC algorithm. The areas with smallR W ( ). The copies from high- reptation method has its own sources of inefficiencies probability regions areR treated as independent which can be, however, significantly alleviated [80]. samples in the subsequent time steps. The time dependence of the energy offset ET(t) provides a This concludes our presentation of the stochastic population control mechanism that prevents the techniques that are used to simulate the projection population from exploding or collapsing entirely operator introduced in (6). We would like to bring [50, 74]. The branching/elimination algorithm is to the reader’s attention that the algorithm outlined much more efficient in large many-body systems in this section is rather rudimentary and illustrates than the pure DMC method, although it also only the general ideas. A number of important perfor- eventually reaches the limits of its applicability mance improvements are usually employed in practical for a very large number of particles [75]. simulations, see for instance [74] for further details.
8 2.2.3 General expectation values spin-orbital interactions, since they lead to a non- trivial coupling of different spin configurations. In So far, only the total energy was discussed in con- fact, spin-dependent quantum Monte Carlo methods nection with the DMC method. An expression anal- were developed for studies of nuclear matter. A variant ˆ ogous to (9) can be written with any operator A of the Green’s function Monte Carlo method [87, 88] ˆ in place of the hamiltonian H. The acquired quan- treats the spin degrees of freedom directly in their full tity AΨ Ψ = ΨD Aˆ ΨT / ΨD ΨT , called the mixed D T ⟨ | | ⟩ ⟨ | ⟩ state space. Since the number of spin configurations estimate, differs from the pure expectation value grows exponentially with the number of particles, this ΨD Aˆ ΨD / ΨD ΨD unless Aˆ commutes with the ⟨ | | ⟩ ⟨ | ⟩ approach is limited to relatively small systems. More hamiltonian. In general, the error is proportional to favourable scaling with the systems size offers the aux- the difference between ΨD and ΨT . The bias can | ⟩ | ⟩ iliary field diffusion Monte Carlo method that samples be reduced to the next order using the following ex- the spin variables stochastically by means of auxiliary trapolation [7, 50] fields introduced via the Hubbard–Stratonovich trans- formation [89, 90]. Recently, a version of the auxiliary ΨD Aˆ ΨD ΨD Aˆ ΨT ΨT Aˆ ΨT field DMC method was used to investigate properties ⟨ | | ⟩ = 2 ⟨ | | ⟩ ⟨ | | ⟩ ΨD ΨD ΨD ΨT − ΨT ΨT of the two-dimensional electron gas in presence of the ⟨ | ⟩ ⟨ | ⟩ ⟨ | ⟩ 2 Rashba spin-orbital coupling [91]. ΨD ΨT + O p p . (22) ΨD ΨD − ΨT ΨT ⟨ | ⟩ ⟨ | ⟩ 2.3 Pseudopotentials Alternative methods that allow for a direct evaluation The computational cost of all-electron QMC calcula- of the pure expectation values have been developed, tions grows very rapidly with the atomic number Z such as the forward (or future) walking [50, 81, 82], of the elements constituting the simulated system. the reptation quantum Monte Carlo [79, 83, 84], or the Theoretical analysis [92, 93] as well as practical cal- Hellman–Feynman operator sampling [85, 86]. Due culations [94] indicate that the cost scales as Z5.5−6.5. to their certain limitations, these techniques do not Most of the computer time spent in these simulations fully replace the extrapolation (22)—they are usable is used for sampling of large energy fluctuations in the only for local operators and the former two become core region, which have very little effect on typical computationally inefficient in large systems. properties of interest, such as interatomic bonding The discussion of the random errors from the end and low-energy excitations. For investigations of these of section 2.1 applies also to the diffusion Monte Carlo quantities it is convenient to replace the core electrons method, except that the serial correlations among the with accurate pseudopotentials. A sizeable library of data produced in the successive steps of the DMC sim- norm-conserving pseudopotentials targeted specifically ulations are typically larger than the correlations in to applications of the QMC methods has been built the VMC data. Therefore, longer DMC runs are neces- over the years [95–100]. sary to achieve equivalent suppression of the stochastic Pseudopotentials substitute the ionic Coulomb po- uncertainties of the calculated expectation values. tential with a modified expression, Z V (r) + Wˆ (23) 2.2.4 Spin degrees of freedom − r → The DMC method as outlined above samples only the where V (r) is a local term behaving asymptotically as spatial part of the wave function, and the spin degrees (Z Zcore)/r with Zcore being the number of elim- − − ˆ of freedom remain fixed during the whole simulation. inated core electrons. The operator W is non-local This simplification follows from the assumption ofa in the sense that electrons with different angular mo- spin-independent hamiltonian that implies freezing of menta experience different radial potentials. Explicitly, ˆ spins during the DMC projection (6). This is indeed the matrix elements of the potential W associated with the current state of the DMC methodology as applied I-th atom in the system are to electronic-structure problems: In order to arrive at lmax l the correct spin state, a number of spin-restricted cal- ′ X X X Wˆ I = rˆiI lm ⟨R| |R ⟩ ⟨ | ⟩ culations are performed and the variational principle i l=0 m=−l is employed to select the best ground state candidate ′ ′ WI,l(riI )δ(riI r ) lm rˆ , (24) among them. × − iI ⟨ | iI ⟩ Fixing spin variables is not possible for spin- where lm are angular momentum eigenstates, riI is dependent hamiltonians, such as for those containing the distance| ⟩ of an electron from the I-th nucleus and
9 rˆiI is the associated direction riI /riI . Functions WI,l accurate enough for the localization error to be practi- vanish for distances riI larger than some cut-off ra- cally insignificant and nearly all applications listed in P dius rc, and the sum i therefore runs only over section 5 utilize this approximation. electrons that are sufficiently close to the particular A method that preserves the upper-bound property nucleus. of EDMC was proposed in the context of the DMC algo- Evaluation of pseudopotentials in the VMC method rithm developed for lattice models [104]. The non-local is straightforward, despite the fact that the local en- operator Wˆ is split into two parts, Wˆ = Wˆ + + Wˆ −, ergy EL itself involves integrals. As can be inferred such that Wˆ + contains those matrix elements, for ′ ′ − from the form of the matrix elements (24), these are which Wˆ ΨT( )ΨT( ) is positive, and Wˆ two-dimensional integrals over surfaces of spheres cen- contains⟨R| the elements,|R ⟩ R for whichR the expression is neg- tered at the nuclei. The integration can be imple- ative. Only the Wˆ + part is localized in order to obtain mented with the aid of the Gaussian quadrature rules the approximate hamiltonian, that employ favourably sparse meshes [101, 102]. The use of non-local pseudopotentials in the fixed- ˆ + − W ΨT( ) node DMC method is more involved, since the sam- Wˆ Wˆ + R . (27) → ΨT( ) pling algorithm outlined in section 2.2.2 explicitly as- R sumed that all potentials were local. Non-local hamil- One can explicitly show that the lowest eigenvalue of tonian terms can be formally incorporated by introduc- this partially localized hamiltonian is an upper bound ing an extra member into the Trotter break-up (15), to the lowest eigenvalue of the original fully non-local namely hamiltonian [104]. Recently, a stochastic represen- ′ Gnloc( , τ) tation of the non-local Green’s function (25) corre- R ← R sponding to Wˆ − was implemented also into the DMC ΨT( ) −τWˆ ′ = R′ e method for continuous space [105]. Apart from the ΨT( ) ⟨R| |R ⟩ R recovered upper-bound property, the new algorithm ′ ΨT( ) ˆ ′ 2 reduces fluctuations of the local energy for certain = δ( ) R′ τW + O(τ ) , (25) R − R − ΨT( ) ⟨R| |R ⟩ types of pseudopotentials. On the other hand, the R time step error is in general larger [105, 106], since the ˆ where W now combines the non-local contributions distinct treatment of the Wˆ + and Wˆ − parts of the from all atoms in the system. This alone is not the pseudopotential essentially corresponds to a Trotter desired solution, since the term involving the matrix splitting of the growth/decay Green’s function (18) ˆ element of W does not have a fixed sign and thus into two pieces. Very recently, a more accurate Trotter cannot be interpreted as a transition probability. break-up and other modifications improving efficiency To circumvent this difficulty the so-called localiza- of this method have been proposed for both continuous tion approximation has been proposed. It amounts to and lattice DMC formulations [107]. a replacement of the non-local operator in the hamil- The localization approximation is directly applica- tonian with a local expression [93, 102, 103] ble also to the fixed-phase DMC method. Adaptation of the non-local moves representing Wˆ − to cases in- Wˆ ΨT( ) Wˆ WL( ) = R . (26) volving complex wave functions has not been reported → R ΨT( ) R yet, nevertheless, the modifications required should be Consequently, the contributions from Wˆ are directly only minor. incorporated into the growth/decay Green’s func- tion (18) and no complications with alternating sign arise. Unfortunately, the DMC method with this ap- 3 From a finite supercell to the proximation does not necessarily provide an upper- thermodynamic limit bound estimate for the ground-state energy. The cal- culated total energy EDMC is above the lowest eigen- Quantum Monte Carlo methods introduced in the pre- value of the localized hamiltonian, which does not ceding chapter can be straightforwardly applied to guarantee that it is also higher than the ground-state physical systems of a finite size, such as atoms and energy of the original hamiltonian Hˆ . The errors in clusters of atoms. To allow investigation of bulk prop- the total energy incurred by the localization approx- erties of solids, the algorithms described so far have to imation are quadratic in the difference between the be complemented with additional techniques that re- trial function ΨT and the exact ground-state wave duce the essentially infinitely many degrees of freedom function [102].| The⟩ trial wave functions are usually into a problem of manageable proportions.
10 3.1 Twist-averaged boundary conditions and of an electron-electron contribution
In approximations that model electrons in solids as an X 1 Vˆee( ) = ensemble of independent (quasi-)particles, it is possible R ri rj j