{"http:\/\/dx.doi.org\/10.14288\/1.0071817":{"http:\/\/vivoweb.org\/ontology\/core#departmentOrSchool":[{"value":"Science, Faculty of","type":"literal","lang":"en"},{"value":"Physics and Astronomy, Department of","type":"literal","lang":"en"}],"http:\/\/www.europeana.eu\/schemas\/edm\/dataProvider":[{"value":"DSpace","type":"literal","lang":"en"}],"https:\/\/open.library.ubc.ca\/terms#degreeCampus":[{"value":"UBCV","type":"literal","lang":"en"}],"http:\/\/purl.org\/dc\/terms\/creator":[{"value":"Wu, Jinshan","type":"literal","lang":"en"}],"http:\/\/purl.org\/dc\/terms\/issued":[{"value":"2011-04-26T14:54:16Z","type":"literal","lang":"en"},{"value":"2011","type":"literal","lang":"en"}],"http:\/\/vivoweb.org\/ontology\/core#relatedDegree":[{"value":"Doctor of Philosophy - PhD","type":"literal","lang":"en"}],"https:\/\/open.library.ubc.ca\/terms#degreeGrantor":[{"value":"University of British Columbia","type":"literal","lang":"en"}],"http:\/\/purl.org\/dc\/terms\/description":[{"value":"To study transport properties, one needs to investigate the system of interest when coupled to biased external baths. This requires solving a master equation for this open quantum system. Obtaining this solution is very challenging, especially for large systems. This limits applications of the theories of open quantum systems, especially insofar as studies of transport in large quantum systems, of interest in condensed matter, is concerned.\n\nIn this thesis, I propose three efficient methods to solve the Redfield equation --- an example of such a master equation. The first is an open-system Kubo formula, valid in the limit of weak bias. The second is a solution in terms of Green's functions, based on a BBGKY (Bogoliubov--Born--Green--Kirkwood--Yvon)-like hierarchy. In the third, the Redfield equation is mapped to a generalized Fokker-Planck equation using the coherent-state representation. All three methods, but especially the latter two, have much better efficiency than direct methods such as numerical integration of the Redfield equation via the Runge-Kutta method. For a central system with a d-dimensional Hilbert space, the direct methods have complexity of d\u00b3, while that of the latter two methods is on the order of order of polynomials of log d. The first method, besides converting the task of solving the Redfield equation to solving the much easier Schr\u00f6dinger's equation, also provides an even more important conceptual lesson: the standard Kubo formula is not applicable to open systems.\n\nBesides these general methodologies, I also investigate transport properties of spin systems using the framework of the Redfield equation and with direct methods. Normal energy and spin transport is found for integrable but interacting systems. This conflicts with the well-known conjecture linking anomalous conductivity to integrability, and it also contradicts the relationship, suggested by some, between gapped systems (Jz > Jxy) and normal spin conductivity. I propose a new conjecture, linking anomalous transport to the existence of a mapping of the problem to one for non-interacting particles.","type":"literal","lang":"en"}],"http:\/\/www.europeana.eu\/schemas\/edm\/aggregatedCHO":[{"value":"https:\/\/circle.library.ubc.ca\/rest\/handle\/2429\/33955?expand=metadata","type":"literal","lang":"en"}],"http:\/\/www.w3.org\/2009\/08\/skos-reference\/skos.html#note":[{"value":"Quantum Transport Through Open Systems by Jinshan Wu B.Sc., The Beijing Normal University, 1999 M.Sc., The Beijing Normal University, 2002 M.Sc., The University Of British Columbia, 2006 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in The Faculty of Graduate Studies (Physics) THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) April, 2011 c\r Jinshan Wu 2011 Abstract To study transport properties, one needs to investigate the system of inter- est when coupled to biased external baths. This requires solving a master equation for this open quantum system. Obtaining this solution is very chal- lenging, especially for large systems. This limits applications of the theories of open quantum systems, especially insofar as studies of transport in large quantum systems, of interest in condensed matter, is concerned. In this thesis, I propose three e\u000ecient methods to solve the Red\feld equation | an example of such a master equation. The \frst is an open- system Kubo formula, valid in the limit of weak bias. The second is a solu- tion in terms of Green's functions, based on a BBGKY (Bogoliubov{Born{ Green{Kirkwood{Yvon)-like hierarchy. In the third, the Red\feld equation is mapped to a generalized Fokker-Planck equation using the coherent-state representation. All three methods, but especially the latter two, have much better e\u000eciency than direct methods such as numerical integration of the Red\feld equation via the Runge-Kutta method. For a central system with a d-dimensional Hilbert space, the direct methods have complexity of d3, while that of the latter two methods is on the order of polynomials of log d. The \frst method, besides converting the task of solving the Red\feld equation to solving the much easier Schrodinger's equation, also provides an even more important conceptual lesson: the standard Kubo formula is not applicable to open systems. Besides these general methodologies, I also investigate transport proper- ties of spin systems using the framework of the Red\feld equation and with direct methods. Normal energy and spin transport is found for integrable but interacting systems. This con\ricts with the well-known conjecture linking anomalous conductivity to integrability, and it also contradicts the relation- ship, suggested by some, between gapped systems (Jz > Jxy) and normal spin conductivity. I propose a new conjecture, linking anomalous transport to the existence of a mapping of the problem to one for non-interacting particles. ii Preface The work presented in Chapters 2 and 6 are a continuation of the work done for my M.Sc. thesis. This work is reported in the manuscript: \\Heat transport in quantum spin chains: the relevance of integrability\", Jinshan Wu and Mona Berciu, submitted to Physical Review B and posted as arXiv:1003.1559. The work presented in Chapter 3 is reported in Jinshan Wu and Mona Berciu 2010 Europhysics Letters 92 30003. The \frst part of the work presented in Chapter 4 is reported in Jinshan Wu 2010 New J. Phys. 12 083042. The second part, which is on the second- order BBGKY-like method, is currently under preparation for publication. Original work reported in other chapters will be prepared for publication in the near future. All the original work in this thesis was carried out by JW who also developed the conception, scope and methods of this research, with various degrees of consultation and editing support from MB. iii Table of Contents Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . ix Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x Chapter 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Two major theoretical approaches . . . . . . . . . . . . . . . 2 1.2 A big challenge for the system-baths scenario . . . . . . . . . 5 1.3 Our methods and their applications . . . . . . . . . . . . . . 7 1.3.1 Linear response theory . . . . . . . . . . . . . . . . . 7 1.3.2 The BBGKY-like hierarchy . . . . . . . . . . . . . . . 8 1.3.3 The GFPE in the coherent-state representation . . . 9 1.3.4 Thermal conductance of spin chains . . . . . . . . . . 9 1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Chapter 2 The projector technique and the Red\feld equation 12 2.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.2 The e\u000bective equation of motion for open systems . . . . . . 12 2.2.1 The projector operator technique . . . . . . . . . . . 14 2.2.2 An alternative derivation . . . . . . . . . . . . . . . . 19 2.2.3 QME in terms of creation and annihilation operators 20 2.3 Relaxation towards thermal equilibrium . . . . . . . . . . . . 22 2.3.1 One-site open system . . . . . . . . . . . . . . . . . . 22 2.3.2 Two-site open system . . . . . . . . . . . . . . . . . . 24 2.4 Various other extensions . . . . . . . . . . . . . . . . . . . . 28 iv Table of Contents 2.5 Evolution towards a NESS . . . . . . . . . . . . . . . . . . . 30 2.5.1 NESSs from the RE . . . . . . . . . . . . . . . . . . . 31 2.5.2 NESSs from the LOLE and the MQME . . . . . . . . 34 2.5.3 The general Lindblad form . . . . . . . . . . . . . . . 37 2.6 Major challenge for large systems . . . . . . . . . . . . . . . 38 Chapter 3 Linear Response Theory: Kubo formula for open systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.2 General Linear Response Theory . . . . . . . . . . . . . . . . 41 3.3 Limitations of the standard Kubo formula . . . . . . . . . . 43 3.4 OKF: the Kubo formula for open \fnite-size systems . . . . . 47 3.5 Comparison of the two OKFs . . . . . . . . . . . . . . . . . . 50 3.6 Summary and discussion . . . . . . . . . . . . . . . . . . . . 53 Chapter 4 BBGKY-like hierarchy for the Red\feld equation 55 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4.2 Derivation of the BBGKY-like hierarchy . . . . . . . . . . . 56 4.3 Non-equilibrium Wick's theorem . . . . . . . . . . . . . . . . 59 4.4 Truncation of the hierarchy . . . . . . . . . . . . . . . . . . . 62 4.4.1 Method 1: equilibrium substitution . . . . . . . . . . 62 4.4.2 Method 2: \frst-level cluster expansion . . . . . . . . 67 4.4.3 Method 2: second-level cluster expansion . . . . . . . 72 4.5 Conclusion and discussion . . . . . . . . . . . . . . . . . . . . 78 Chapter 5 Solving the Red\feld equation using coherent-state representations . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 5.2 The model and its e\u000bective equation of motion . . . . . . . . 81 5.3 Exact solution for non-interacting systems . . . . . . . . . . 85 5.4 Interacting systems: BBGKY-like hierarchy derived from the GFPE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 5.5 Approximate numerical solution for interacting systems . . . 90 5.6 Solving the LOLE via the GFPE method . . . . . . . . . . . 94 5.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 Chapter 6 Application: heat transport in spin chains . . . . 97 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 6.2 The model and its Red\feld equation . . . . . . . . . . . . . . 99 6.3 Numerical methods . . . . . . . . . . . . . . . . . . . . . . . 100 v Table of Contents 6.4 De\fnitions of the thermal current and local temperatures . . 101 6.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Chapter 7 Conclusions and discussions . . . . . . . . . . . . . 112 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 Appendix A Perturbational decomposition of the operators m\u0302 128 A.1 Expanded in eigenbasis of HS . . . . . . . . . . . . . . . . . 128 A.2 Perturbational expansion starting from non-interacting sys- tems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 Appendix B Second-order cluster expansion: G1, G2 . . . . . 134 B.1 Second-order cluster expansion . . . . . . . . . . . . . . . . . 134 B.2 Second-order equations of G1 and G2 . . . . . . . . . . . . . 135 Appendix C Estimation of convergence of the BBGKY-like methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 C.1 \u0001 E;(1) 1 from method 1 . . . . . . . . . . . . . . . . . . . . . . 141 C.2 \u0001 C;(1) 1 from method 2 . . . . . . . . . . . . . . . . . . . . . . 145 Appendix D Coherent states and coherent-state representa- tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 D.1 Coherent states for systems with a single mode . . . . . . . . 148 D.2 Coherent states for multi-mode systems . . . . . . . . . . . . 151 Appendix E FPE and Langevin equation . . . . . . . . . . . . 153 E.1 Langevin equation for standard FPEs . . . . . . . . . . . . . 153 E.2 Analytical solution for the Ornstein-Uhlenbeck process . . . 155 E.3 Stochastic di\u000berence equation for truncated GFPEs . . . . . 155 Appendix F Solving the RE and LOLE of bosonic systems with BBGKY-like hierarchy . . . . . . . . . . . . . . . . . . . 157 F.1 Exact solution on non-interacting systems . . . . . . . . . . . 158 F.2 Approximate solution on interacting systems . . . . . . . . . 159 vi List of Figures 1.1 Sketch of a typical setup of heat conduction. . . . . . . . . . 10 2.1 Sketch of a one-site system coupled to a heat bath. . . . . . . 23 2.2 Sketch of a two-site system coupled to a heat bath. . . . . . . 24 2.3 Sketch of a two-site system coupled to two heat baths. . . . . 27 2.4 Di\u000berence between non-equilibrium and equilibrium states . . 33 2.5 Stationary states from the local-operator Lindblad equation and the multi-mode quantummaster equation compared against those from the Red\feld equation . . . . . . . . . . . . . . . . 36 3.1 Di\u000berence between \u001a(1) (1) and \u001a(2) (1) for di\u000berent values of \u000e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.2 Di\u000berence between \u001a(1) (1) and \u001a(2) (1) for di\u000berent values of \u0015 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3.3 Di\u000berence between \u001a(1) (1) and \u001a(2) (1) for di\u000berent values of V0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 4.1 g E;(1) 1 compared with g Ex 1 for interacting systems at non- equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 4.2 g C;(1) 1 compared with g Ex 1 for non-equilibrium interacting sys- tems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 4.3 g C;(2) 1 compared with g Ex 1 for non-equilibrium interacting sys- tems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 4.4 g C;(2) 2 is compared with g Ex 2 for non-equilibrium interacting systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 4.5 Particle currents calculated from the second-, \frst- and the zeroth-order forms are compared . . . . . . . . . . . . . . . . 76 4.6 Heat currents calculated from the second-, \frst- and the zeroth- order forms are compared . . . . . . . . . . . . . . . . . . . . 77 5.1 Analytical solution compared against the BBGKY numerical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 vii List of Figures 5.2 The numerical simulation of the generalized Fokker-Planck equation for interacting systems . . . . . . . . . . . . . . . . . 93 5.3 A similar application to the local-operator Lindblad equation 96 6.1 Typical temperature pro\fles for N = 10 spin chains . . . . . 104 6.2 Current v.s. system size . . . . . . . . . . . . . . . . . . . . . 105 6.3 Normalized pro\fles . . . . . . . . . . . . . . . . . . . . . . . . 106 6.4 Typical temperature pro\fle for the cases of small Jz and small Bz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 6.5 Temperature pro\fles on Ising spin chains and fermionic chains 109 7.1 A sketch of an exactly solvable decoherence model and its variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 viii Acknowledgements It is a pleasure to thank those who made this thesis possible. I owe my deepest gratitude to my supervisor, Dr. Mona Berciu. This thesis would not have been possible without her great e\u000borts to educate and to help me in so many ways. She taught me not only physics, but also more importantly how to appreciate physics, and how to write up and explain things clearly and simply. She also encouraged and protected my interest in fundamental physical questions. Throughout my years of graduate study at UBC, she has always been a good advisor and a good friend to me. It is an honor for me to thank the many people with whom I discussed the related work and have been bene\fted greatly from such discussion: Ian A\u000feck, Keiji Saito, Philip Stamp, Tomaz Prosen, Peter Drummond, Hong Guo and Shouyong Pei. I would like to thank the many great teachers I ever had especially: Yue'e Yi, Jianjiang Zhang, Anshen Qi, Jiaruan Wang, Yongcheng Wang, Zhanru Yang, Shouyong Pei, Canbing Liang, Jingdong Bao, Zengru Di, Qiang Yuan, Andrew DeBenedictis, Leslie Ballentine, Ariel Zhitnitsky, Mark Van Raamsdonk and Ian A\u000feck. I am indebted to my many colleagues for the stimulating discussion and sometimes even excited and heated argument: Yong Sun, Mandy Wang, Kelly Cheung, Bayo Lau, Arthur James Charbonneau, Jennell Van Dongen, Kory Stevens, Si Chen, Junliang Song, Menghui Li, Liang Gao, Ying Fan, Dahui Wang, Limei Xu, Yougui Wang, Zhangang Han, Honggang Li. Lastly but most importantly, I wish to thank my wife, Qian Feng, for her unconditional love and support; and my daughters, Lindsay Wu and Elysia Wu for the happiness that they have already brought to this family. To them I dedicate this thesis. ix To Qian Feng, Lindsay Wu and Elysia Wu x Chapter 1 Introduction Theories of open quantum systems [1{3] usually start from an e\u000bective equa- tion of motion of the central systems of interest, which is usually derived from the exact equation of motion for the composite system | the cen- tral system plus the bath(s). Although various approximations are involved in deriving such e\u000bective equations of motion, which are sometimes called quantum master equations, they are believed to be reasonably reliable since they are built on relatively \frm grounds, i.e. an equation of motion derived from \frst-principles. Theories of transport [4, 5] in solid state physics deal with a speci\fc type of open system, where the central system is coupled to multiple baths held at di\u000berent temperatures, chemical potentials, etc. As a result, study of transport should be a branch of the theories of open quantum systems. In practice, di\u000berent theories based on other approaches have been developed for transport studies, such as the Kubo formula [6, 7], the Landauer-Buttiker formula [8{11] and the non-equilibrium Green's function method [4, 12{16]. One important reason why theories of open quantum systems are not widely used when studying transport is the computational di\u000eculty in applying them to large systems. In studies of relaxation times in nuclear magnetic resonance [17], where the theories of open quantum systems provide the standard tools, a typical central system has usually only a few degrees of freedom. In contrast, in discussions of transport, one studies quite often mesoscopic systems whose Hilbert space's dimension easily exceeds 220. It is the primary goal of this thesis to \fnd new ways of making the theo- ries of open quantum systems amenable for the study of transport in large(r) systems, by \fnding e\u000ecient calculation schemes. For non-interacting sys- tems there are straightforward methods to calculate stationary states and therefore physical quantities, even if the system is large. Thus, the real challenge is for systems with interactions, just as is the case in the study of equilibrium properties and dynamical processes [18]. We will propose various e\u000ecient and accurate approximations, and test them. This chapter serves as an introduction to the subject and also an overview of this thesis. Sections x1.1 and x1.2 brie\ry review the theories of open sys- 1 1.1. Two major theoretical approaches tems while focusing on their applications to the study of transport. Section x1.3 lists our contributions, including the development of new methodolo- gies and discovery of new physics from application of these methodologies to certain speci\fc systems. The rest of this thesis follows the same line of logic as this overview, but in more details. Chapter 2 reviews the projector technique used in theo- ries of open quantum systems. Examples are given to illustrate the basic procedure. The resulting e\u000bective equations of motion are solved via direct methods, which have very poor e\u000eciency. Then we present an approximate solution based on linear response theory in Chapter 3. This approxima- tion has better e\u000eciency than the direct methods but not good enough to be comparable to the non-equilibrium Green's function method. Chapters 4 and 5 introduce two very e\u000ecient and accurate approximations. One is based on a Bogoliubov{Born{Green{Kirkwood{Yvon (BBGKY)-like equa- tion hierarchy and the other is based on the coherent-state representation of density matrices. In Chapter 6 we apply the direct methods to discuss thermal transport of spin systems and the validity of Fourier's Law. At last, in Chapter 7 we summarize these results and list some of the potentially valuable questions to be investigated in future, with our newly developed methods. We also comment on the value and implications of the work presented in this thesis. 1.1 Two major theoretical approaches A rough picture of a typical experimental setup for measuring transport properties is that of a quasi 1-d system coupled to multiple biased baths. We want to calculate currents \rowing through the system given the parameters of the baths such as their temperatures, chemical potentials, etc. Such calculations are relevant for comparison against experimental measurements of charge, energy and spin transport, for example through quasi 1-d spin systems [19], quantum wires including nanotubes [20], molecular devices [21{ 24] and quantum dots [25, 26]. In particular, in recent years there have been many experiments looking at heat transport of spin systems. The heat conductance \u0014 = Jh\u0001T , where Jh is the heat current and \u0001T the applied temperature bias, is measured for a crystal with spin-chain structures. Typical crystals sizes are of roughly a few mm. In order to measure the heat conductance contributed by the spin chains, one has to subtract the conductance due to phonons and charge carriers. Usually this is achieved by using anisotropic spin-chain compounds 2 1.1. Two major theoretical approaches such as (Ca,Sr)14Cu24O41 [27], which contain S = 1 2 two-leg ladders with an intra-chain coupling much stronger than the inter-chain coupling. By measuring the di\u000berence between the conductance parallel to the chain and that perpendicular to the chain, one \fnds the pure spin conductance. The idea is that the parallel conductance is due to both the spin-chain and other sources, while the perpendicular one only has the contributions from other sources. Experimental results on spin chains in strong magnetic \felds have also been reported [28, 29]. However, at present, for spin chains, there is no available data regarding temperature pro\fles and \fnite size-scaling for the heat currents. This makes comparison with our theoretical results, discussed in Chapter 6, impossible at this time. In a recent measurement of thermal conductance of nanotubes [30], Wang et al. reported the size dependence of heat currents and concluded that it violates Fourier's law. For small, and especially also quantum, systems, it is not reliable to use phenomenological laws such as Fourier's law and Ohm's law to characterize their transport. To understand and model such systems, a theoretical frame- work based on \frst principle is necessary. Developing such a framework is not just a matter of theoretical interest, but also has obvious practical value. Various theoretical models have been used so far. Some studies are based on the phenomenological balance equation [31] or rate equation [32, 33], but the most common and systematic approaches are the Kubo for- mula [6, 34, 35], the non-equilibrium Green's function method [13{15], and the open quantum system approach [1{3] with the system-bath scenario [36{ 38]. In the approach based on the standard Kubo formula, the system itself is assumed to be in\fnite and homogeneous. There is no explicit use of the baths but instead the system is treated as being isolated. The current is calculated as a response to a presumed applied electrical potential [6] or temperature gradient drop [39], or uneven mass gradient distribution [40]. In the approach of non-equilibrium Green's function and Landauer formula, the system is modeled by three pieces: the left lead, the central system and the right lead. The leads are taken as half-in\fnite non-interacting systems and the central system is usually \fnite. There are no explicit baths in- cluded in this picture. The parameters of the baths, such as temperatures and chemical potentials, are attached to the corresponding leads. Finally, the theories of open quantum systems deal with a \fnite-size central system coupled explicitly to baths, whose degrees of freedom are then integrated out. There are no in\fnite leads in this picture. These three di\u000berent physical pictures are supposed to describe the same problem. Fisher and Lee [41] showed that for non-interacting systems the 3 1.1. Two major theoretical approaches \frst two are somewhat equivalent if the same three-piece system is used when applying the Kubo formula. We should note, however, that the ma- jority of studies of conductance via the Kubo formula are using an in\fnite homogeneous system [35, 42], or extrapolate from results on \fnite-size sys- tems [43, 44]. With that being said, let us ignore the question of whether these methods are equivalent, and focus only on the latter two approaches. Comparisons between them, or even settling how to make them directly comparable, has not been extensively studied. Non-equilibrium Green's functions is a perturbational theory starting from a well-de\fned piece-wise equilibrium state of the disconnected three- piece system. The left\/right lead is assumed to be in its own equilibrium state correspondingly de\fned by the left\/right temperature and chemical potential, while the central system is in an arbitrary equilibrium state [14]. From this, the Green's function of the disconnected system is calculated and later used to construct the full Green's function of the connected sys- tem, with the connection between the central system and the leads treated as perturbation. The standard diagrammatic perturbation theory does not apply here since the initial and \fnal states are qualitatively di\u000berent. It is an essential assumption of the standard diagrammatic perturbation theory that the states at both \u00061 in time are the same equilibrium state (or the ground state if T = 0) of the unperturbed system. At this point a non- equilibrium Keldysh formalism [12, 18] is used instead. Conceptually, the condition that the two leads are half-in\fnite is essential. With any \fnite- size leads, an equilibrium state { instead of the desired non-equilibrium stationary state { will be reached ultimately since the unbalanced energies or chemical potentials will be eventually dissipated over the whole system. However, even for in\fnite leads, theoretically there is still no guarantee that the proper stationary state can be calculated perturbationally [13]. In prac- tice, for interacting systems, the non-equilibrium Green's function method is often used together with the density functional theory [13, 16]. Addi- tional approximations, thus possibly additional errors [45, 46], are involved in dealing with the interactions. Theories of open quantum systems, on the other hand, are based on relatively \frmer grounds. They start from the \frst-principle equation of motion of the composite system, and arrive at an e\u000bective equation of mo- tion of the central system. Various equations can be derived depending on the level of approximations. Using the in\ruence functional theory [47, 48], exact time-dependent master equations have been derived for simple cen- tral systems [49]. Examples include a two-level system, one harmonic os- cillator and two coupled harmonic oscillators, which is the current limit of 4 1.2. A big challenge for the system-baths scenario such exact approaches [50]. When the Markovian approximation is used, time-independent master equations have also been derived from the exact ones [51]. For more general central systems, some direct brute-force gener- alization of the master equation of a simple two-site system have also been used in the literature. We will call one kind of such equations the local- operator Lindblad equation for reasons that will be clear later. It is in fact very hard to derive exact master equations from in\ruence functional theory for general central systems. Instead, approximate equations can be derived from the projector technique [51] or other similar techniques. Such a master equation is sometimes called the Red\feld equation [52{54]. Applications of the Red\feld equation to transport studies are in an early but fast develop- ing stage [36{38, 55{57]. Others prefer to use the local-operator Lindblad equation [58, 59], which turns out to be relatively easier to solve numerically. It is usually easier to deal with large systems using non-equilibrium Green's functions, than it is with theories of open quantum systems. In practice, quite often a set of few-particle Green's functions are enough for calculating the physical quantities of interest. The number of single-particle Green's functions, for instance, is much smaller than the number of ma- trix elements of the density matrix, which appears in the Red\feld equation. Consider for example an N -site 1-d spinless fermionic system. The set of all single-particle Green's functions has N2 unknowns while a density matrix has 22N unknowns. Therefore, unless an e\u000ecient calculation scheme whose complexity scales as N2 can be found, numerically the two approaches are not comparable at all. This explains why the non-equilibrium Green's func- tion method is much more widely used than other methods to study trans- port in solid state systems. To many, it does not even seem necessary and urgent to compare results from the two methods and determine which is more accurate or even which one is correct. This situation may change if e\u000ecient methods to solve the Red\feld equa- tion are found, as is the goal of this thesis. 1.2 A big challenge for the system-baths scenario Both the general Red\feld equation and the general local-operator Lindblad equation can be cast into the form of a generalized Liouville equation, d dt \u001a (t) = L\u001a (t) ; (1.1) where \u001a (t) is the density matrix of the central system and L is a linear superoperator acting on the density matrix. We are only interested in the 5 1.2. A big challenge for the system-baths scenario long-term stationary state and the steady current associated with it. The stationary state is de\fned by L\u001a (1) = 0; (1.2) where \u001a (1) is the long-term non-equilibrium stationary state. If the Hilbert space of the central system is d-dimensional, then the density matrix has d2 elements. If we represent the density matrix as a vector, for example, P = [\u001a11; \u001a12; \u0001 \u0001 \u0001 ; \u001add]T ; (1.3) then the linear superoperator L is a d2\u0002d2 matrix. The task of \fnding \u001a (1) thus implies \fnding the zeroth mode of L | the eigenvector corresponding to the zero eigenvalue. One can imagine that this becomes nontrivial when d > 210, as in the example of a fermionic chain with N > 10 mentioned earlier. The central task of this thesis is to \fnd e\u000ecient methods to solve this equation. Before we present our work, we brie\ry review the currently avail- able methods. Instead of solving directly for the zero mode of L, we may instead let the system evolve from an arbitrary initial state long enough to arrive at \u001a (1). Thus, the \frst class of solutions are based on propagators { this includes the direct Runge-Kutta method [37, 60], the Newtonian polynomial propagator [61], the short-iterative-Arnoldi propagator [62] and the short- time Chebyshev polynomial propagator [63]. All these methods have been compared in Ref. [64]. As for their e\u000eciency, they are all roughly on the same scale { their complexities scale roughly as d3. A more e\u000ecient approach is the stochastic wave-function method [2, 65, 66]. The basic idea is to replace the e\u000bective equation of motion by a stochastic Schrodinger equation for wave functions. The method has a complexity scaling as d2. On the local-operator Lindblad equation it has been shown that the stochastic wave-function method leads to accurate re- sults [67]. Breuer et al. showed that it can also be applied to the Red\feld equation [65], but its e\u000eciency is still not comparable with that of the non- equilibrium Green's functions method. For the local-operator Lindblad equation, in fact, there is another quite e\u000ecient method developed recently by Prosen et al. [59]. It is based on the time-dependent density matrix renormalization technique. This method makes it possible to analyze systems roughly of the same size as those that can be studied with the non-equilibrium Green's functions method. It is still 6 1.3. Our methods and their applications an open problem to extend this approach to the Red\feld equation. How- ever, we should note that the same central system might behave di\u000berently under evolution described by the local-operator Lindblad equation versus the Red\feld equation. In fact, this issue has not been settled because of the lack of comparison between the results of the two approaches. The lack of e\u000ecient methods for the Red\feld equation has certainly contributed to this state of a\u000bairs. With our newly developed methods, such a comparison can \fnally be undertaken. 1.3 Our methods and their applications In this section, we roughly summarize the ideas, performances and applica- tions of our newly developed methods. Their details will be discussed in the rest of this thesis. We have developed three di\u000berent methods, with di\u000berent strenghts and weaknesses. 1.3.1 Linear response theory The \frst method is a linear response theory based on the Red\feld equation. The basic idea is very simple. The equilibrium state of the central system (when unbiased) is known to be the Boltzmann distribution. The stationary state of the Red\feld equation under appropriate conditions, i.e. all baths at the same temperature and chemical potential, is indeed the Boltzmann distribution. This makes it possible to separate the operator L into a large part L0 and a small part \u0001L, where the former part is the operator describ- ing the equilibrium conditions. Then, the non-equilibrium distribution can be treated as a perturbational response to the small part \u0001L. This method has e\u000eciency comparable to the stochastic wave-function method. Besides the gain in e\u000eciency, which is not that remarkable, an impor- tant conceptual conclusion is drawn from this work. The standard Kubo formula [6], as we already knew, does produce a proper \frst-order correc- tion to equilibrium states when an additional potential is included into the Hamiltonian. However, our work shows that it fails to describe the non- equilibrium stationary states. If in\fnite-size systems are used as the cen- tral system, there are ways to get around this conceptual problem. For non-equilibrium stationary states in \fnite-size systems, the standard Kubo formula must be replaced by a Kubo-like formula based on the Red\feld equation { the standard Kubo formula is no longer valid. The key di\u000berence is that the standard Kubo formula uses Schrodinger's equation and the sys- tem is treated as isolated while in the new Kubo-like formula, the Red\feld 7 1.3. Our methods and their applications equation of an open system is used. Given that there are still many studies of transport based on various forms of the standard Kubo formula, this is a very important conclusion. 1.3.2 The BBGKY-like hierarchy The second method mirrors the ideas of the well-known BBGKY hierarchy. The original BBGKY hierarchy [68] is a set of coupled equations for equilib- rium correlation functions (Green's functions). For non-interacting systems, the hierarchy is decoupled, meaning that the equation of a single-particle Green's function only depends on other single-particle Green's functions, leading to a closed system of equations. Using Wick's theorem, which holds for correlation functions of non-interacting systems, higher order n-particle Green's functions can then be constructed from the single-particle Green's functions. However, when there are interactions in the system, equations for all orders of Green's functions are coupled together. Fortunately, other ap- proximations can be used to truncate the hierarchy and then one solves the truncated hierarchy. These approximations can be performed in such a way that they are equivalent to the diagrammatic perturbation theory [18, 69]. Here, we extend the same approach to non-equilibrium stationary states. A BBGKY-like hierarchy is derived from the Red\feld equation. We \fnd again that for non-interacting systems, the equations decouple and a similar Wick's theorem can be proved. All equations are coupled together for inter- acting systems. Two further approximations are introduced to truncate the hierarchy. The \frst one is based on the known equilibrium distribution of the interacting system. The second one is based on the cluster expansion [69]: the observation that higher-order Green's functions can be constructed from the lower-order ones through Wick's Theorem and an additional correlation term. Ignoring these correlation terms at a certain order leads to a trunca- tion of the hierarchy at the corresponding order. Our tests of the two approximations for small systems, where exact solu- tions needed for comparisons are also possible, indicate that they have very good accuracy. Moreover, they can be systematically improved by going to higher orders in the truncation schemes. The complexity of this approach scales only like polynomials of N , not d, therefore this is a huge improvement of the e\u000eciency. 8 1.3. Our methods and their applications 1.3.3 The general Fokker-Planck equation in the coherent-state representation The third method we propose is to map and solve the Red\feld equation in the coherent-state representation of quantum states [70, 71]. Consider a simple harmonic oscillator for example. The set of all coherent states j\u0018i, which are eigenvectors of the annihilation operator a j\u0018i = \u0018 j\u0018i, provides an overcomplete basis for quantum states. Expanding a density matrix in such a basis results in a \\distribution\" function P (\u0018) over the complex plane C. Here we use quotation marks because of the possibility that P (\u0018) could be a negative or even complex number. In this way, creation and annihilation operators become di\u000berential operators, and the Red\feld equation becomes a stochastic di\u000berential equation of the distribution function over CN for a general coupled interacting system of N harmonic oscillators. Instead of dealing with the huge size of the linear system, like with the direct meth- ods, now we need to solve an N -dimension stochastic di\u000berential equation, which generally can be simulated classically. For non-interacting systems, in fact, the equation can be solved analytically. We have obtained analytical expressions of non-equilibrium stationary states and compared them to the exact solutions. We \fnd perfect agreement, as expected. For interacting systems, once the classical simulations are performed, the same comparisons can be done. The number of variables increases linearly with the system size so this method is capable of looking at very large systems. A similar procedure [72{74] has been proposed and implemented for the study of equilibrium and pure dynamical systems. There the method is applied successfully for bosonic systems with roughly N = 106 modes [75]. In a way, our method can be regarded as an extension of that method to the study of non-equilibrium stationary states. Although this extension involves additional technical di\u000eculties, its e\u000eciency remains comparable to that of its counterpart for the study of equilibrium states and pure dynamical processes. 1.3.4 Thermal conductance of spin chains The physical question we discuss using this open-system approach is the validity of Fourier's Law of heat conduction, jh = \u0000\u0014rT; (1.4) where jh is the heat current density, T is the local temperature and \u0014 is the heat conductivity. This phenomenological law is found to be obeyed by 9 1.3. Our methods and their applications jh TL TR Figure 1.1: Sketch of a typical setup of heat conduction: left and right baths have di\u000berent temperatures and the conductor is a quasi 1D system. jh is the current density on the conductor. It is usually assumed that there is a linearly decreasing temperature pro\fle on the conductor. macroscopic systems, when measured experimentally in set-ups such as the one sketched in Fig. 1.1. The derivation of the law from the microscopic equation has been an open question for many decades [34, 76]. The question has been investi- gated in both classical [76] and quantum systems [34]. For classical systems, a plausible relation between chaotic behavior and the validity of Fourier's Law (also referred to as normal transport), has been observed. But there is still no consensus on the criteria required for a system to exhibit normal transport. The situation is even less clear-cut for quantum systems. Most experimental [19] and theoretical work [34] have focused on spin and energy transport of spin chains or spin ladders. It has been conjectured, based on investigations using the standard Kubo formula [34, 35], that integrability is the criterion for anomalous conductivity. Other studies based on di\u000ber- ent approaches, however, reach di\u000berent conclusions. For example, it has been suggested that the existence of gap leads to anomalous transport [77] and also that spin transport could be qualitatively di\u000berent from energy transport [59]. Experimental results on integrable systems range too from anomalous conductance [78{80] to normal conductance [81]. Most of the previous studies use the standard Kubo formula [34, 44, 82]. Some are based on the local-operator Lindblad equation [56, 59] and some are based on the Red\feld equation [36, 37, 83]. However, the spin systems studied via the Red\feld equation are either non-interacting (and thus triv- ially integrable) or non-integrable. We denote, for example, the XY chains as non-interacting since via Jordan-Wigner transformation [84] they can be mapped to systems of free fermions. In the same sense we regard the integrable Heisenberg chains to be interacting since by the same transfor- mation the resulting fermionic systems are interacting. Anomalous trans- port was found for (trivially) integrable systems and normal transport for non-integrable systems. Disordered spin systems, which are non-integrable, 10 1.4. Summary have also been investigated and found to have normal transport [37]. Prior to our work, there were no integrable interacting systems investigated via the Red\feld equation. We apply the Red\feld equation to investigate the energy transport of spin 1=2 chains. We also \fnd anomalous transport for XY chains, but normal transport for XXZ chains, which are integrable but interacting. Therefore, we conclude that integrability does not necessarily imply anomalous transport. We then investigate a wider class of systems and conjecture another criterion: interacting systems have normal transport. The reliability of this conjecture needs to be checked further. 1.4 Summary In summary, this chapter has presented a brief synopsis of the challenges that we are trying to address in this thesis, and the directions we have used to achieve these goals. The rest of the thesis will now discuss each of these issues in detail. 11 Chapter 2 The projector technique and the Red\feld equation 2.1 Outline In this chapter we review the projector technique for central systems coupled to single or multiple baths, and illustrate its application to small systems with only a few degrees of freedom, where the equations can be solved di- rectly. The challenge it faces when applied to large systems, and our solution to this challenge, will be the topics of other chapters. This chapter is organized as follows. Section x2.2 contains a brief re- view of an open system's dynamics based on the system-bath scenario, as a preparation for our method. In Section x2.3, we discuss how systems cou- pled to a single bath, or to multiple baths held at the same temperature and the same chemical potential, arrive at thermal equilibrium. As examples, the Red\feld equations of both a single-mode central system and a two-site central system are solved to illustrate the general procedure. Besides the Red\feld equation, there are other various forms of open-system quantum master equations. Usually they can be regarded as generalizations of the Red\feld equation of a single-mode central system. Two such equations, which are widely used in literature, are presented in Section x2.4. Then, in Section x2.5 we generalize the Red\feld equation to systems coupled to two or more baths of di\u000berent temperatures but with the same chemical potential, and discuss their heat conductance. The method is also valid for electrical and spin transport, where the chemical potentials of baths could also be di\u000berent. 2.2 The e\u000bective equation of motion for open systems In this section we derive a general e\u000bective equation of motion for a central system coupled to a single bath, starting from the Schrodinger's equation for 12 2.2. The e\u000bective equation of motion for open systems the total composite system. A similar derivation can be found in Refs. [1, 2, 51, 85]. Unless otherwise stated, in this thesis we use ~ = 1; kB = 1; c = 1. Consider a system S interacting with a single bath\/reservoir B. In the Schrodinger picture, the Liouville equation for the evolution of the density matrix of the total closed quantum system S +B is: i @\u001a\u0302T (t) @t = h H\u0302; \u001a\u0302T (t) i \u0011 L\u0302\u001a\u0302T (t) : (2.1) Here we denote the density matrix of the total composite system as \u001a\u0302T and the reduced density matrix of the central system (the bath) as \u001a\u0302 (\u001a\u0302B), \u001a\u0302 (t) = trB (\u001a\u0302T (t)) ; (2.2) \u001a\u0302B (t) = tr S (\u001a\u0302T (t)) ; (2.3) where trB (trS) stands for the operation of partial trace over the bath (the central system) degrees of freedom. The total Hamiltonian has the generic form: H\u0302 = H\u03020 (S;B) + V\u0302 (S;B) ; (2.4) where H\u03020 (S;B) = H\u0302S (S) I\u0302 (B) + I\u0302 (S) H\u0302B (B) (2.5) describes the system, respectively the bath, and V\u0302 (S;B) = X i L\u0302i (S) F\u0302i (B) (2.6) describes their interactions, with L\u0302i (S) and F\u0302i (B) being system, respec- tively bath operators. I\u0302 (B) and I\u0302 (S) are the identity operators in the Hilbert space of the bath and the central system, respectively. If we are interested only in the properties of S, we call it an open quantum system. Typically, we need to calculate the ensemble average of physical quantities depending on operators acting on S only, hAi = tr \u0010 [A\u0302S I\u0302 (B)]\u001a\u0302T \u0011 : (2.7) This can be trivially rewritten as: hAi = trS \u0010 A\u0302S \u001a\u0302 \u0011 : (2.8) Therefore, it is convenient to try to \fnd \u001a\u0302 (t) directly, and avoid solving for the total density matrix \u001a\u0302T (t), which is a much more complicated problem 13 2.2. The e\u000bective equation of motion for open systems due to the signi\fcant increase in the number of degrees of freedom. Instead we are looking for the evolution equation of the reduced matrix density in a form similar to Eq. (2.1): @\u001a\u0302 (t) @t = L\u0302 \u001a\u0302 (t) ; (2.9) but now L\u0302 is obviously neither h H\u0302; \u0001 i of the whole system nor h H\u0302S ; \u0001 i of the central system only. The task of this section is to derive the proper Liouville operator starting from knowledge of the total Hamiltonian H\u0302. There are two main types of approaches to discuss the dynamics of an open system: based on dynamic variables or based on density matrices. The \frst approach uses equations of motion for dynamic variables such as X\u0302; P\u0302 etc. The dynamic variables of the bath are integrated out, resulting in noise-sources terms in the Langevin-type equations for the dynamic variables of the system. If one can solve these Langevin equations, the full time dependence of the central system is known. From a quantum mechanics point of view, this approach is similar to the Heisenberg picture. The second approach focuses on distribution functions in the phase space of the system, such as \u001a (x; p) for classical systems and the density matrix \u001a\u0302 for quantum systems. Equations of motion are generated for them, the bath is projected out, and then the equation for the reduced density matrix is solved. From a quantum mechanics point of view, this approach is more similar to the Schrodinger picture. We will use the latter approach. In this section we summarize the general formalism of this method. There are two well-known techniques to derive the equation of motion of the reduced density matrix, namely the projector operator technique and the non-correlation approximation. We discuss the \frst one in reasonable detail and then shortly review the second one. 2.2.1 The projector operator technique and the reduced Liouville equation One way to derive the unknown L\u0302 in Eq. (2.9) is through the technique of projection. We de\fne the projector operator on the Hilbert space of the total system S +B, P : H ! H, such that P (\u001a\u0302T ) , trB (\u001a\u0302T ) \u001a\u0302eqB = \u001a\u0302 \u001a\u0302eqB (2.10) where \u001a\u0302eqB is the density matrix of the isolated bath in equilibrium, although this requirement could be relaxed. It is easy to verify that P is a projector, 14 2.2. The e\u000bective equation of motion for open systems i.e. that PP = P. We denote Q = I\u0000P, and rewrite the equation of motion for the total system, Eq. (2.1), as:\u001a @ @tP \u001a\u0302T (t) = PL\u0302P \u001a\u0302T (t) + PL\u0302Q\u001a\u0302T (t) @ @tQ\u001a\u0302T (t) = QL\u0302P \u001a\u0302T (t) +QL\u0302Q\u001a\u0302T (t) (2.11) The second equation is integrated formally and its solution is substituted into the \frst equation, to obtain: @ @t P \u001a\u0302T (t) = PL\u0302P \u001a\u0302T (t) + PL\u0302 Z t 0 d\u001ce(t\u0000\u001c)QL\u0302QL\u0302P \u001a\u0302T (\u001c) + PL\u0302etQL\u0302Q\u001a\u0302T (0) : (2.12) If we start from the reasonable initial condition \u001a\u0302T (0) = \u001a\u0302 (0) \u001a\u0302eqB ; (2.13) then Q\u001a\u0302T (0) = 0 (2.14) and the last term vanishes identically. Let LS ;LB;LV be the Liouville op- erators corresponding respectively to H\u0302S (S) ; H\u0302B (B) and V\u0302 (S;B). Then, L\u0302 = L\u0302S + L\u0302B + L\u0302V , L\u03020 + L\u0302V : (2.15) Since L\u0302B \u001a\u0302eqB = 0, we can rewrite the dynamic equation as @ @t \u001a\u0302 (t) \u001a\u0302eqB = PL\u0302S \u0002 \u001a\u0302 (t) \u001a\u0302eqB \u0003 + PL\u03021 \u0002 \u001a\u0302 (t) \u001a\u0302eqB \u0003 (2.16) +PL\u0302 Z t 0 d\u001ce(t\u0000\u001c)QL\u0302QL\u0302 \u0002\u001a\u0302 (t) \u001a\u0302eqB \u0003 : (2.17) The second term is identically zero if we require that for any eigenstate in the bath's Hilbert space, the expectation value of the bath operators involved in the bath-system coupling [see Eq. (2.6)] vanishes: hbj F\u0302i (B) jbi = 0: (2.18) In this case, we have: i hbj L\u03021 \u0002 \u001a\u0302 (t) \u001a\u0302eqB \u0003 jbi =X i hbj h L\u0302i (S) F\u0302i (B) ; \u001a\u0302 (t) \u001a\u0302eqB i jbi = X i \u0010 L\u0302i (S) \u001a\u0302 (t) hbj F\u0302i (B) \u001a\u0302eqB jbi \u0000 \u001a\u0302 (t) L\u0302i (S) hbj \u001a\u0302eqB F\u0302i (B) jbi \u0011 = 0 15 2.2. The e\u000bective equation of motion for open systems showing that indeed, PL\u0302V P = 0. This condition, which is satis\fed for most types of couplings, is not however necessary. If it is not obeyed, there will be a term of the form P i;b e\u0000\fEb Z D b \f\f\fF\u0302i\f\f\f bE hL\u0302i; \u0001i, which can be added to L\u0302S . It follows that this assumption is not essential, although all the coupling Hamiltonians we study satisfy it. With these assumptions and after tracing out the bath degrees of freedom, we \fnd that: @ @t \u001a\u0302 (t) = L\u0302S \u001a\u0302 (t) + trB \u001a PL\u0302 Z t 0 d\u001ce(t\u0000\u001c)QL\u0302QL\u0302P \u001a\u0302T (t) \u001b : (2.19) Subject to the two assumptions mentioned above, this equation is exact. Clearly, all the e\u000bects of the bath on the system's dynamics are contained in the second term on the right-hand side. This term can be further simpli\fed. One can verify that QL\u0302P = QL\u03020P +QL\u0302V P = QL\u0302V P = \u0000PL\u0302V P + L\u0302V P = L\u0302V P (2.20) where the second assumption has been invoked again. Now we use the \frst essential assumption, replacing: e(t\u0000\u001c)QL\u0302 ! e(t\u0000\u001c)QL\u03020 : (2.21) This is valid when the system-bath interaction is weak, since, as shown in the following, it is equivalent to second order perturbation in the \\coupling strength\" between system and bath. This follows because we can then sim- plify: e(t\u0000\u001c)QL\u03020 = e(t\u0000\u001c)(I\u0000P)L\u03020 = e(t\u0000\u001c)(L\u03020\u0000PL\u03020) = e(t\u0000\u001c)L\u03020 ; (2.22) where we make use of PL\u03020 = PL\u03020P + PL\u03020Q = 0. Therefore we obtain: @ @t \u001a\u0302 (t) = L\u0302S \u001a\u0302 (t) + trB \u001a L\u0302V Z t 0 d\u001ce(t\u0000\u001c)L\u03020L\u0302V \u001a\u0302 (\u001c) \u001a\u0302eqB \u001b ; (2.23) where indeed the second term is of second order in the interaction V\u0302 . This can be further simpli\fed in the interaction picture , where we de\fne \u001a\u0302I (t) = eiH\u0302St\u001a\u0302 (t) e\u0000iH\u0302St; (2.24a) L\u0302Ij (t) = e iH\u0302StL\u0302je \u0000iH\u0302St; (2.24b) F\u0302 Ij (t) = e iH\u0302BtF\u0302je \u0000iH\u0302Bt: (2.24c) 16 2.2. The e\u000bective equation of motion for open systems In this case, the \frst term cancels out, and the dynamic equation becomes: @\u001a\u0302I (t) @t = \u0000 X jl Z t 0 d\u001c nh L\u0302Ij (t) ; L\u0302 I l (t\u0000 \u001c) \u001a\u0302I (t\u0000 \u001c) i Gjl (\u001c) \u0000 h L\u0302Ij (t) ; \u001a\u0302 I (t\u0000 \u001c) L\u0302Il (t\u0000 \u001c) i Glj (\u0000\u001c) o (2.25) where Gjl (\u001c) = tr B \u0010 F\u0302 Ij (\u001c) F\u0302 I l (0) \u001a\u0302 eq B \u0011 : (2.26) are Green's functions which depend only on the bath characteristics and can be calculated directly. Depending on the speci\fc form of the Hamiltonian, Eq. (2.25) may or may not have a unique long-term stationary solution. For example, it may oscillate between several metastable states. The exact criteria guaranteeing the existence of such a steady-state are not known. Finding such criteria would be an interesting topic for further research. In the following, we take the pragmatic approach and assume that if a stationary steady-state solution has been found, then it is unique. Now comes our second essential assumption, the Markov approximation: \u001a\u0302I (t\u0000 \u001c) \u0019 \u001a\u0302I (t) for t\u001d \u001cc; (2.27) where \u001cc is the characteristic relaxation time of Gjl (\u001c) (i.e., of the bath). Since in this thesis we investigate the steady-state solution when t!1, the Markov approximation is likely justi\fed. Using the in\ruence functional tech- nique, similar but more complicated equations can be derived if the Markov approximation is relaxed [47{50, 86]. These more complicated equations are necessary in the study of transient process. We believe that for the purpose of study of stationary states, the Markov approximation is accurate enough. With this \fnal approximation, we \fnd the desired equation of motion for the reduced density matrix: @\u001a\u0302I (t) @t = \u0000 X jl Z t 0 d\u001c nh L\u0302Ij (t) ; L\u0302 I l (t\u0000 \u001c) \u001a\u0302I (t) i Gjl (\u001c) \u0000 h L\u0302Ij (t) ; \u001a\u0302 I (t) L\u0302Il (t\u0000 \u001c) i Glj (\u0000\u001c) o : (2.28) It can be transformed back to the Schrodinger picture, @\u001a\u0302 (t) @t = \u0000i [HS ; \u001a\u0302 (t)]\u0000 X jl Z t 0 d\u001c nh L\u0302j ; L\u0302 I l (\u0000\u001c) \u001a\u0302 (t) i Gjl (\u001c) \u0000 h L\u0302j ; \u001a\u0302 (t) L\u0302 I l (\u0000\u001c) i Glj (\u0000\u001c) o : (2.29) 17 2.2. The e\u000bective equation of motion for open systems This equation can also be projected in the eigenbasis of HS , denoted as fjsig such that HS jsi = Es jsi. In this case, the e\u000bective equation of motion becomes d dt \u001aIs0s (t) = X m;n Rs0smn (t) \u001a I mn (t) ; (2.30) with Rs0smn = \u0000ei\u0001Es0smnt \u0001 \"X k \u000esn\u0000 + s0kkm \u0000 \u0000+nss0m \u0000 \u0000\u0000nss0m + X k \u000es0m\u0000 \u0000 nkks # ; (2.31) where \u0001Es0smn = E S s0 \u0000ESs + ESn \u0000 ESm (2.32) and \u0000+mkln (t) = X ij hmj L\u0302i jki hlj L\u0302j jni Z t 0 e\u0000i(E S l \u0000ESn)\u001cGij (\u001c) d\u001c (2.33) and \u0000\u0000mkln (t) = \u0000 \u0000+nlkm (t) \u0001\u0003 . The above e\u000bective equation of motion is sometimes called a quantum master equation. If we focus only on the diag- onal terms for \u001aIss (t), then we obtain the secular quantum master equation, which requires the so-called secular approximation [1] in order to be consis- tent with Fermi's Golden Rule. The secular approximation is to ask that ESs0 \u0000 ESs + ESn \u0000 ESm = 0; (2.34) so as to eliminate the explicit time dependence of the right hand side of Eq. (2.31). Then, \fnally, d dt \u001aIss (t) = X m Wsm\u001a I mm (t)\u0000 X m Wms\u001a I ss (t) ; (2.35) where Wsm = \u0000 + mssm + \u0000 \u0000 mssm. This is now consistent with Fermi's Golden Rule. While there is a considerable body of work based on such secular quantum master equations, we choose to work directly with the full e\u000bective equation of motion Eq. (2.28). 18 2.2. The e\u000bective equation of motion for open systems 2.2.2 An alternative derivation: the non-correlation approximation This is a di\u000berent way to derive the above equation of motion for the reduced density matrix, using the so-called non-correlation approximation. Suppose that both H\u0302S and H\u0302B are solvable, and their eigenstates are respectively fjsig for ESs and fjbig for EBb . Then, a complete basis for the total Hilbert space can be chosen as fjsi jbig. In the interaction picture \f\f I (t)\u000b = ei(H\u0302S+H\u0302B)t j (t)i, the time evolu- tion of the operators describing the coupling is L\u0302Ii (t) = e iH\u0302StL\u0302i (t) e \u0000iH\u0302St (2.36) and F\u0302 Ii (t) = e iH\u0302BtF\u0302i (t) e \u0000iH\u0302Bt: (2.37) The dynamical equation of the whole system is: i @\u001a\u0302IT (t) @t = h V\u0302 I (t) ; \u001a\u0302IT (t) i : (2.38) Integrating this formally and substituting the resulting expression for \u001a\u0302I in the right-hand side of Eq. (2.38), we obtain: @\u001a\u0302IT (t) @t = \u0000i h V\u0302 I (t) ; \u001a\u0302IT (0) i \u0000 Z t 0 d\u001c h V\u0302 I (t) ; h V\u0302 I (\u001c) ; \u001a\u0302IT (\u001c) ii : (2.39) Integrating out the bath, this becomes: @\u001a\u0302I (t) @t = \u0000itrB \u0010h V\u0302 I (t) ; \u001a\u0302IT (0) i\u0011 \u0000 Z t 0 d\u001ctrB \u0010h V\u0302 I (t) ; h V\u0302 I (\u001c) ; \u001a\u0302IT (\u001c) ii\u0011 ; (2.40) which is still exact. Now we begin to introduce several approximations. First, we again assume hbj F\u0302i (B) jbi = 0; (2.41) so that the \frst term in Eq. (2.40) vanishes. As before, this is not an essential assumption. Second , we also assume, \u001a\u0302IT (\u001c) = \u001a\u0302 I (\u001c) \u001a\u0302IB (\u001c) ; (2.42) 19 2.2. The e\u000bective equation of motion for open systems such that the second term in Eq. (2.40) can be separated and simpli\fed. This is the essential approximation. In fact this is stronger than the corre- sponding approximation, Eq. Eq. (2.13), used in the previous section, where this factorization is assumed to be true only for the initial state. Third, we take, \u001a\u0302IB (\u001c) = \u001a\u0302 eq B = 1 ZB e\u0000\fH\u0302B : (2.43) With these assumptions, we get a simpli\fed dynamic equation for the re- duced density matrix identical to Eq. (2.25). If we further apply the Markov condition, we arrive at the same e\u000bective equation of motion as in Eq. (2.28). 2.2.3 The quantum master equation in terms of creation and annihilation operators From now on, we will work in the Schrodinger picture and directly deal with the reduced density matrix of the central system. In this thesis, we will mostly use the second quantization language. In terms of creation and annihilation operators, Eq. (2.28) can be simpli\fed further. Let us now derive such a simpli\fed version for a general central system, whose \u0017th site is coupled to the \u0017th bath with bath parameters T\u0017 ; \u0016\u0017 and coupling constant Vk;\u0017 and \u0015, HS =H \u0010 aj ; a y j \u0011 ; (2.44a) HB\u0017 = X k !\u0017 (k) byk;\u0017bk;\u0017 ; (2.44b) HSB = \u0015 X k;\u0017 \u0010 Vk;\u0017a y \u0017bk;\u0017 + h:c: \u0011 : (2.44c) Here H \u0010 aj ; a y j \u0011 can be any function of aj and a y j . \u0015 refers to the strength of the coupling between the central system and the baths. It could be absorbed into the coupling coe\u000ecients Vk;\u0017 but we separate it and use it as a relatively small dimensionless number for bookkeeping purpose. Here we assume that every site connected to a bath, is connected to its own bath. If two sites happen to share a bath, then the formula we derive below needs to be adjusted. First we identify L (1) k;\u0017 = \u0015Vk;\u0017a\u0017 ; F (1) k;\u0017 = b y k;\u0017 ;L (2) k;\u0017 = \u0015V \u0003 k;\u0017a y \u0017 ; F (2) k;\u0017 = bk;\u0017 (2.45) 20 2.2. The e\u000bective equation of motion for open systems and note that the only non-vanishing bath Green's functions are: G1;2k;\u0017;k;\u00170 (\u001c) = trB \u0010 F (1) k;\u0017 (\u001c)F (2) k;\u00170 (0) \u001aB \u0011 = hnk;\u0017i ei!\u0017(k)\u001c\u000e\u0017;\u00170 ; (2.46a) G2;1k;\u0017;k;\u00170 (\u001c) = trB \u0010 F (2) k;\u0017 (\u001c)F (1) k;\u00170 (0) \u001aB \u0011 = h1\u0000 nk;\u0017i e\u0000i!\u0017(k)\u001c\u000e\u0017;\u00170 : (2.46b) The \u000e\u0017;\u00170 comes from the assumption of independent baths for each site, which means that Green's functions among operators belonging to baths coupled to di\u000berent sites are zero. By de\fnition, hnk;\u0017i \u0011 hn (!\u0017 (k) ; T\u0017 ; \u0016\u0017)i = \u0012 e !\u0017 (k)\u0000\u0016\u0017 T\u0017 \u0000 1 \u0013\u00001 (2.47) is the average particle number in state k in bath \u0017. For simplicity of notation, hnk;\u0017i will be denoted as nk;\u0017 whenever it is safe to do so. After plugging these Green's functions into Eq. (2.28), we arrive at a quantum master equation @\u001a(t) @t = \u0000i[HS ; \u001a(t)] \u0000\u00152 X \u0017 nh ay\u0017 ; m\u0302\u0017\u001a(t) i + \u0002 a\u0017 ; \u0016\u0302m\u0017\u001a(t) \u0003 + h:c: o ; (2.48) with operators m\u0302 and \u0016\u0302m de\fned as follows: m\u0302\u0017 = X k jVkj2 Z 1 0 d\u001ca\u0017 (\u0000\u001c) e\u0000i!\u0017(k)\u001c h1\u0000 nk;\u0017i ; (2.49a) \u0016\u0302m\u0017 = X k jVkj2 Z 1 0 d\u001cay\u0017 (\u0000\u001c) ei!\u0017(k)\u001c hnk;\u0017i : (2.49b) Here a\u0017 (\u001c) = e iHSta\u0017e \u0000iHSt is expressed in the Heisenberg picture. Note that we have also changed the integral limit from t to 1. Again this is reasonable only if the stationary solution itself is of interest, not the whole evolution process. This quantum master equation in creation and annihila- tion operators is the central equation of this chapter and will be used later throughout the whole thesis. Eq. (2.48) is sometimes called the Red\feld equation [52, 53, 87]. In some cases, it is convenient to use the eigenstates of HS as a basis, fjEnig. In that case, using element-wise products in this basis, the expres- sions of the operators m\u0302 and \u0016\u0302m can be rewritten as m\u0302\u0017 = a\u0017 \u0001 \u0006\u0017 ; \u0016\u0302m\u0017 = ay\u0017 \u0001 \u0016\u0006\u0017 ; (2.50) 21 2.3. Relaxation towards thermal equilibrium where \u0006\u0017 = \u0019 X m;n jmi hnj [1\u0000 n (En \u0000 Em; T\u0017 ; \u0016\u0017)]J\u0017 (En \u0000 Em) ; (2.51a) \u0016\u0006\u0017 = \u0019 X m;n jmi hnjn\u0017 (Em \u0000 En; T\u0017 ; \u0016\u0017) J\u0017 (Em \u0000 En) ; (2.51b) and J\u0017 (\u000f) = X k jVkj2 \u000e(\u000f\u0000 !\u0017(k)) (2.52) is usually called the spectrum density function [36, 48] of the bath \u0017. 2.3 Relaxation towards thermal equilibrium We \frst apply this general procedure to the simplest example | a single-site fermionic system coupled to a bath | and show that the system evolves to the correct thermal equilibrium as t ! 1. We then brie\ry investigate a corresponding two-site system. 2.3.1 One-site open system Consider a one-state fermionic system in a strong magnetic \feld, describing for instance a single quantum dot, de\fned by a Hamiltonian: HS = \u000fa ya (2.53) where a; ay are fermionic operators. The system is coupled to a single bath (\u0017 = 1 only), which is regarded as an ensemble of fermions with \fxed T and \u0016, so that both particles and energy are exchanged. Using the general quantum master equation Eq. (2.48), we have @ @t \u001a = \u0000i h \u000faya; \u001a i \u0000 \u00152 nh ay; m\u0302\u001a i + \u0002 a; \u0016\u0302m\u001a \u0003 + h:c: o ; (2.54) where m\u0302 = \u0019J (\u000f) [1\u0000 n(\u000f)] a; (2.55a) \u0016\u0302m = \u0019J (\u000f)n(\u000f)ay; (2.55b) and n(\u000f) is shorthand for n(\u000f; T; \u0016) = \u0010 e \u000f\u0000\u0016 T \u0000 1 \u0011\u00001 , the average number of particles with energy \u000f in the bath (since there is a single bath, there is no 22 2.3. Relaxation towards thermal equilibrium B S Figure 2.1: Sketch of a one-site system coupled to a heat bath. need to explicitly write down T\u0017 and \u0016\u0017 everywhere). J (\u000f) can be de\fned accordingly for this speci\fc system using Eq. (2.52). Now we can solve the dynamical equation and check the stationary solu- tion. The density matrix of the above system can be written in the general form: \u001a (t) = 1 + p (t) 2 j0i h0j+ 1\u0000 p (t) 2 j1i h1j+ q (t) j0i h1j+ q\u0003 (t) j1i h0j ; (2.56) so that the trace is preserved. Then we get evolution equations for these parameters from the Red\feld equation: d dt p (t) = 2 (1\u0000 2n(\u000f))\u0000 2p (t) ; (2.57a) d dt q (t) = \u0000q (t) : (2.57b) This has a stationary solution, independent of the initial state: p (1) = 1\u0000 2n(\u000f) = e \f(\u000f\u0000\u0016) \u0000 1 e\f(\u000f\u0000\u0016) + 1 ; (2.58a) q (1) = 0; (2.58b) which leads indeed to the expected equilibrium grand-canonical distribution \u001a(1) = 1 e\u0000\f(\u000f\u0000\u0016) + 1 j0i h0j+ e \u0000\f(\u000f\u0000\u0016) e\u0000\f(\u000f\u0000\u0016) + 1 j1i h1j (2.59) for the two-level system. Note that in this example, the value of \u0015 itself is irrelevant: this equilibrium state will be reached irrespective of the strength of the coupling to the bath. As we show later, this is not true for non- equilibrium stationary states, which depend on the details of the coupling between the central system and the baths. 23 2.3. Relaxation towards thermal equilibrium 1 2 Bath Figure 2.2: Sketch of a two-site system coupled to a heat bath. 2.3.2 Two-site open system Let us now consider a two-site fermionic system coupled to one heat bath, see Fig. 2.2. The system's Hamiltonian is taken to be a simple hopping Hamiltonian: HS = \u0000\u000f \u0010 ay1a2 + a y 2a1 \u0011 ; (2.60) and we assume that only site 1 is coupled to the bath, so that: V = X k \u0010 \u0015Vka1b y k + \u0015V \u0003 k a y 1bk \u0011 : (2.61) Using the general e\u000bective equation of motion Eq. (2.48), we arrive at d dt \u001a = \u0000i [HS ; \u001a]\u0000 \u00152 nh ay1; m\u03021\u001a i + \u0002 a1; \u0016\u0302m1\u001a \u0003 + h:c: o ; (2.62) where m\u03021 = \u0019J (\u0000\u000f) [1\u0000 n (\u0000\u000f)] 2 (a1 + a2) + \u0019J (\u000f) [1\u0000 n (\u000f)] 2 (a1 \u0000 a2) ; (2.63a) \u0016\u0302m1 = \u0019J (\u0000\u000f)n (\u0000\u000f) 2 \u0010 ay1 + a y 2 \u0011 + \u0019J (\u000f)n (\u000f) 2 \u0010 ay1 \u0000 ay2 \u0011 ; (2.63b) and J (\u0006\u000f) can be calculated similarly as in Eq. (2.52) but now for the eigenenergies \u0006\u000f. Again, for simplicity we assume J(\u0006\u000f) = 1. Note that when \u000f = 0, the operators m\u03021 and \u0016\u0302m1 become m\u03021 = J (0) [1 + n (0)] a1; (2.64a) \u0016\u0302m1 = J (0)n (0) a y 1; (2.64b) 24 2.3. Relaxation towards thermal equilibrium in agreement with results in the previous section. This shows that only if \u000f is zero, then the operators m\u03021 and \u0016\u0302m1 depend only on a1 and a y 1. Otherwise, they also involve the creation and annihilation operators for other sites in the system (here, site 2). In fact, unless ay\u0017 happens to be the creation operator of an eigenmode of HS , which is generally not the case, \u0016\u0302m\u0017 involves other operators than a\u0017 ; a y \u0017 . We will revisit this issue in Section x2.4. The eigenmodes of this Hamiltonian are, HS = \u0000\u000fay+a+ + \u000fay\u0000a\u0000; (2.65) where a\u0006 = a1 \u0006 a2p 2 : (2.66) In terms of occupation states of these eigenmodes jn+n\u0000i, eigenstates are j00i, j10i, j01i, j11i with eigenvalues respectively \u000f1 = 0, \u000f2 = \u0000\u000f, \u000f3 = \u000f, \u000f4 = 0. A density matrix in this basis \u001a = 2664 p11 p12 p13 p14 p21 p22 p23 p24 p31 p32 p33 p34 p41 p42 p43 p44 3775 (2.67) can be cast into a 16-dimensional vector, P = [p11; p12; \u0001 \u0001 \u0001 ; p44]T : (2.68) 25 2.3. Relaxation towards thermal equilibrium In this basis, operators are represented by: a1 = 1p 2 2664 0 1 1 0 0 0 0 1 0 0 0 \u00001 0 0 0 0 3775 ; (2.69) a2 = 1p 2 2664 0 1 \u00001 0 0 0 0 \u00001 0 0 0 \u00001 0 0 0 0 3775 ; (2.70) m\u03021 = \u0019p 2 2664 0 1\u0000 n(\u0000\u000f) 1\u0000 n(\u000f) 0 0 0 0 1\u0000 n(\u000f) 0 0 0 \u00001 + n(\u0000\u000f) 0 0 0 0 3775 ; (2.71) \u0016\u0302m1 = \u0019p 2 2664 0 0 0 0 n(\u0000\u000f) 0 0 0 n(\u000f) 0 0 0 0 n(\u000f) \u0000n(\u0000\u000f) 0 3775 : (2.72) Plugging the above operators and density matrix into Eq. (2.62), we can derive the explicit form of matrix \u0000 and rewrite Eq. (2.62) in the following form, d d\u001c P = \u0000P; (2.73) where the matrix \u0000 has dimension 16\u000216, or generally 4N\u00024N for an N -site spinless fermionic system. Fortunately, since we are only interested in equilibrium states in this section, we know that the o\u000b-diagonal elements of the equilibrium density matrices will vanish. Therefore, we work with P dia = [p11; p22; p44; p44] T in this section and derive and solve only its evolution equation. In this case, the matrix \u0000 can also be reduced to a matrix \u0000dia. Using the above operators a1, a y 1, m\u03021, and \u0016\u0302m1, we \fnd \u0000dia = \u00152\u0019 2664 \u0000n(\u0000\u000f)\u0000 n(\u000f) 1\u0000 n(\u0000\u000f) 1\u0000 n(\u000f) 0 n(\u0000\u000f) n(\u0000\u000f)\u0000 n(\u000f)\u0000 1 0 1\u0000 n(\u000f) n(\u000f) 0 n(\u000f)\u0000 n(\u0000\u000f)\u0000 1 1\u0000 n(\u0000\u000f) 0 n(\u000f) n(\u0000\u000f) n(\u000f) + n(\u0000\u000f)\u0000 2 3775 ; (2.74) 26 2.3. Relaxation towards thermal equilibrium 21 Left bath Right bath Figure 2.3: Sketch of a two-site system coupled to two heat baths. and the reduced equation becomes d d\u001c P dia = \u0000diaP dia: (2.75) A general time-dependent solution of Eq. (2.75) with initial condition P dia0 is P dia (\u001c) = e\u0000 dia\u001cP dia0 : (2.76) It converges to a long term stationary state only when all eigenvalues of \u0000dia are non-positive. Its stationary state is the eigenvector of \u0000dia with eigenvalue equal to zero, \u0000diaP dia (1) = 0: (2.77) We veri\fed that \u0000dia indeed only has one zero eigenvalue, and that its cor- responding eigenvector P dia (1) leads to the proper thermal equilibrium state, P dia (1) = 1 Z h 1; e\f(\u0016+\u000f); e\f(\u0016\u0000\u000f); e\f2\u0016 iT ; (2.78) where Z = 1 + e\f(\u0016+\u000f) + e\f(\u0016\u0000\u000f) + e\f2\u0016. Up to this point, we have illustrated the general procedure for a system coupled to one bath and con\frmed that its stationary state is the expected thermal equilibrium. A straightforward extension is to consider a system coupled to multiple baths, for example di\u000berent sites in the system are locally coupled to di\u000berent baths. In fact, our general formula Eq. (2.48) has already been derived to be applicable for such situations. The only change to be made is to allow every bath to have its own operators m\u0302\u0017 and \u0016\u0302m\u0017 . 27 2.4. Various other extensions Speci\fcally, if as sketched in Fig.2.3, we couple site 2 to a bath 2 with the same bath parameters, then our e\u000bective equation of motion reads: d dt \u001a = \u0000i [HS ; \u001a]\u0000 \u00152 nh ay1; m\u03021\u001a i + \u0002 a1; \u0016\u0302m1\u001a \u0003 + h:c: o \u0000\u00152 nh ay2; m\u03022\u001a i + \u0002 a2; \u0016\u0302m2\u001a \u0003 + h:c: o ; (2.79) where the additional operators are de\fned as m\u03022 = \u0019 [1\u0000 n2 (\u0000\u000f)] 2 (a1 + a2)\u0000 \u0019 [1\u0000 n2 (\u000f)] 2 (a1 \u0000 a2) ; (2.80a) \u0016\u0302m2 = \u0019n2 (\u0000\u000f) 2 \u0010 ay1 + a y 2 \u0011 \u0000 \u0019n2 (\u000f) 2 \u0010 ay1 \u0000 ay2 \u0011 : (2.80b) Repeating the above calculation, we \fnd the same stationary state. It is now natural to ask what kind of stationary state will emerge if the two baths have di\u000berent parameters, for example di\u000berent temperatures and\/or chemical potentials. Ideally this should be the non-equilibrium sta- tionary state which can be used to calculate transport properties.1 We will use examples of simple systems to investigate this question in x2.5. Before that, however, we have a subtle issue to settle, in the next section. 2.4 Various other extensions of the single-site quantum master equation There are several forms of the e\u000bective equation of motion used in the lit- erature. Besides the multi-bath Red\feld equation derived above, a local- operator Lindblad equation and a multi-mode quantum master equation are also used quite often to study transport or relaxation processes. In this section, we will introduce the latter two and discuss their limitations. Another way to write Eq. (2.54) for a single-site system is: @ @t \u001a = \u0000i [HS ; \u001a]\u0000 J (\u000f) n (1\u0000 n(\u000f)) \u0010 aya\u001a+ \u001aaya\u0000 2a\u001aay \u0011 +n(\u000f) \u0010 aay\u001a+ \u001aaay \u0000 2ay\u001aa \u0011o : (2.81) 1What we have argued above shows only that it is natural to link this state to the non-equilibrium stationary state, since the proper equilibrium states are reached when the system is coupled to unbiased baths. However, we note that this is not a real \frst- principle derivation: thermal equilibrium of the baths has been explicitly assumed. The question of establishing thermal equilibrium from only \frst principles is not the scope of this thesis although many mathematicians and physicists are very much interested in this issue. 28 2.4. Various other extensions It is very tempting to generalize the above form directly to multi-site systems coupled to multiple local baths as follow, @ @t \u001a = \u0000i [HS ; \u001a] \u0000 X \u0017 J\u0017 (\u000f\u0017) n (1\u0000 n\u0017(\u000f\u0017)) \u0010 ay\u0017a\u0017\u001a+ \u001aa y \u0017a\u0017 \u0000 2a\u0017\u001aay\u0017 \u0011 +n\u0017(\u000f\u0017) \u0010 a\u0017a y \u0017\u001a+ \u001aa\u0017a y \u0017 \u0000 2ay\u0017\u001aa\u0017 \u0011o ; (2.82) i.e. as if each new bath adds the contribution it would if the site it is coupled to was isolated from the rest of the central system. Here, \u000f\u0017 is the on-site energy of site \u0017 and n\u0017(\u000f\u0017) is the occupation number for a mode of energy \u000f\u0017 given the parameters of bath \u0017. We call this simpli\fed equation the local-operator Lindblad equation. Because of its simplicity, it has been used often for the study of transport properties [58, 59, 88, 89]. We have seen, however, that Eq. (2.48) or more speci\fcally Eq. (2.79) can not be cast into this form because the operator m\u0302\u0017 usually involves not only a\u0017 , but also the operators for all other sites a\u00170 through formulas that depend on the particularHs of the system. Therefore, the two equations are di\u000berent. The question, then, is whether the stationary states of the two equations are qualitatively di\u000berent, and if yes, whether the solution of the local-operator Lindblad equation still captures most features of the non-equilibrium stationary states properly? These questions have not been answered yet. In this section, we will provide some preliminary results. Before that, we also introduce the multi-mode quantum master equation that arises from another extension of the Red\feld equation for a single- site system [70]. For multi-site systems, let us assume we can rewrite the Hamiltonian as follows: HS = X q \u000f (q) ayqaq; (2.83) We further assume that each eigenmode of the system is coupled to its own individual bath with temperature Tq: V\u0302 = \u0015 X k;q h Vqkaq byq;k + h:c: i (2.84) where creation and annihilation operators of bath q satisfy the fermionic commutation relation, fbq;k; byq0;k0g = \u000ek;k0\u000eq;q0 . Note that this is di\u000berent from our setup of Red\feld equation, where some of the sites, not eigenmodes, are coupled each to its own individual bath. 29 2.5. Evolution towards a NESS For this new setup, every eigenmode is independent so that the operators m\u0302q and \u0016\u0302mq involve only aq and a y q. Therefore, from Eq. (2.48) the e\u000bective equation of motion reads d dt \u001a = \u0000i X q \u000f (q) h ayqaq; \u001a i \u0000 X q Jq (\u000f (q)) n (1\u0000 nq(\u000f (q))) \u0010 ayqaq\u001a+ \u001aa y qaq \u0000 2aq\u001aayq \u0011 +nq(\u000f (q)) \u0010 aqa y q\u001a+ \u001aaqa y q \u0000 2ayq\u001aaq \u0011o ; (2.85) i.e. it is simply a sum over all eigenmodes, consisting of contributions similar to those in Eq. (2.54). We call this equation the multi-mode quantummaster equation. It has been used in quantum optics to study relaxation processes. We should emphasize that the two equations Eq. (2.82) and Eq. (2.85) are di\u000berent from each other, except for a single-site system in which case all three equations are the same. For example, for our two-site system, a\u0017 in Eq. (2.82) refers to a1; a2 while aq in Eq. (2.85) refers to a\u0006 = a1\u0006a2p2 . In the local-operator Lindblad equation a1; a y 1 does not mix with a2; a y 2, while in the Red\feld equation there are terms mixing them. Next, we investigate the di\u000berence between these three equations: the Red\feld equation, the local-operator Lindblad equation and the multi-mode quantum master equation. We will argue \frstly that while the local-operator Lindblad equation gives both wrong equilibrium states and wrong non- equilibrium stationary states, it may lead to qualitatively correct results for some physical quantities such as certain currents; and secondly that while the proper equilibrium states can be found from the multi-mode quantum master equation, it does not describe any proper non-equilibrium stationary states. In order to do this, we begin by \fnding the \\proper\" non-equilibrium stationary state from the Red\feld equation of a two-site system coupled to two baths. This solution is discussed in the next section. 2.5 Evolution towards a non-equilibrium stationary state In this section, we \frst solve the Red\feld equation for a two-site system locally coupled to two baths, and then compare it with solutions from the local-operator Lindblad equation and the multi-mode quantum master equa- tion. The general procedure presented in this section is the general method 30 2.5. Evolution towards a NESS underlying all the later work in this thesis. In a sense, all the later work focuses either on the technical side of this procedure or on applications of this procedure. Of course, these are non-trivial because solving the Red\feld equation directly for large systems is a di\u000ecult task. 2.5.1 Non-equilibrium stationary states from the Red\feld equation We begin from Eq. (2.79). Note that when the bath parameters are di\u000berent: T1 6= T2 and\/or \u00161 6= \u00162, unlike in the case of thermal equilibrium states, the o\u000b-diagonal elements of the density matrix are no longer necessarily zero. As a result, we need to \fnd all the elements of the density matrix. Correspondingly the matrix \u0000 has d4 = 28 elements. For such a large linear system we have to solve the problem numerically. We are interested in its stationary solution, i.e. \u0000P (1) = 0: (2.86) This can be solved by \fnding the eigenvector of L corresponding to the zero eigenvalue. Or, it can also be solved as the solution of the following linear system of equations, \u0016\u0000P (1) = V ; (2.87) where V = [1; 0; \u0001 \u0001 \u0001 ]T and \u0016\u0000 is de\fned by replacing the \frst row of \u0000 by the normalization condition tr (\u001a) = 1, i.e. P j Pj\u0003d+j = 1 with our notation. The advantage of this is that \u0000 is a singular matrix but \u0016\u0000 is not. Therefore, while Eq. (2.86) has to be treated as a computationally costly eigenvalue problem, Eq. (2.87) is a non-singular system of linear equations that can be solved more e\u000eciently. For simplicity of notation in later chapters, the above two equations de\fning the non-equilibrium stationary states are also written as L\u001a (1) = 0: (2.88) and \u0016L\u001a (1) = \u0017: (2.89) Here instead of superoperator \u0000; \u0016\u0000 and supervector P;V , we use directly the Liouvillian L; \u0016L and matrix \u001a; v, which are de\fned from mapping the supervector P;V in the above supervector representation back into the usual 31 2.5. Evolution towards a NESS density matrix representation. These two sets of equations are equivalent, but the later has simpler notation since it is still formulated in terms of the Liouvillian and density matrices. Let us summarize the general algorithm for this direct method: 1. Find all eigenvalues and eigenstates of HS , and express all the original operators a\u0017 and a y \u0017 in this basis of eigenstates; 2. Evaluate the correlation functions for every bath to calculate the ma- trices \u0006\u0017 and \u0016\u0006\u0017 de\fned in Eq. (2.51); 3. From these, derive the expressions of the operators m\u0302\u0017 and \u0016\u0302m\u0017 as in Eq. (2.50); 4. Construct \u0000 or \u0016\u0000 using Eq. (2.48) and the previous results; 5. Solve for the zero eigenvectors of \u0000 or the linear system de\fned by \u0016\u0000 in Eq. (2.89). Using this procedure we solve Eq. (2.79). Typical results are shown in the Fig. 2.4 for the speci\fc parameters \u000f = 1; \u0015 = 0:1; \u00161 = 0 = \u00162; T1 = T \u0000 1 + \u000e2 \u0001 ; T2 = T \u0000 1\u0000 \u000e2 \u0001 , and T = 2:0. We have calculated the di\u000berence between the general non-equilibrium stationary state \u001a (1) and the thermal equilibrium \u001aEq (T ) at temperature T . The results are presented separately for o\u000b-diagonal elements and diagonal elements, ddiag = vuuutPj \u0010\u001ajj \u0000 \u001aEqjj \u00112P ij (\u001aij) 2 ; d off = vuuutPi6=j \u0010\u001aij \u0000 \u001aEqij \u00112P ij (\u001aij) 2 : (2.90) The charge current operator is de\fned as J = \u0000ie\u000f \u0010 ay1a2 \u0000 ay2a1 \u0011 ; (2.91) and we also separate its expectation value into contributions from the diag- onal and o\u000b-diagonal parts, Jdiag = X j Jjj\u001ajj ; J off = X i6=j Jji\u001aij : (2.92) From Fig. 2.4(a) we see that compared to the diagonal part, the o\u000b-diagonal elements of the non-equilibrium stationary states are relatively small but 32 2.5. Evolution towards a NESS 0 0.5 1 1.5 2 \u03b4 0 0.05 0.1 0.15 0.2 d ddiag doff 0 0.5 1 1.5 2 \u03b4 0 0.005 J Jdiag Joff Figure 2.4: Density matrices for non-equilibrium stationary states are com- pared to those of equilibrium states. In (a) ddiag and doff are plotted vs. the temperature bias \u000e. Diagonal elements of the density matrices for non- equilibrium stationary states are found to be quite di\u000berent from those of equilibrium states. The di\u000berence between o\u000b-diagonal elements of non- equilibrium stationary states and equilibrium density matrices are much smaller. Note that the o\u000b-diagonal elements of equilibrium density matrices are zero always. (b) Charge currents, Jdiag and Joff , for the non-equilibrium stationary states, are plotted vs. \u000e. We see that only the o\u000b-diagonal terms of the density matrix contribute to the current. 33 2.5. Evolution towards a NESS non-zero. However, Fig. 2.4(b) shows that the diagonal part does not con- tribute at all to the current, Jdiag = 0; the current comes entirely from the o\u000b-diagonal part. So the seemingly small change in the o\u000b-diagonal part makes a big qualitative di\u000berence: o\u000b-diagonal elements are essential to de- scribe transport properties. This is not surprising since we know that if the time-reversal symmetry is not broken, the eigenstates of HS do not support currents, Jjj = 0. Next, we will perform a comparison between stationary states from the local-operator Lindblad equation\/multi-mode quantum master equation and these non-equilibrium stationary states obtained from the Red\feld equation, to see whether they agree or not. 2.5.2 Non-equilibrium stationary states from the local-operator Lindblad equation and the multi-mode quantum master equation We have claimed before that the local-operator Lindblad equation leads to wrong equilibrium states but possibly qualitatively correct non-equilibrium stationary states, while the multi-mode quantum master equation gives proper equilibrium states but not non-equilibrium stationary states. In this section, we will con\frm these statements by presenting results for the simple example of a two-site fermionic system coupled to two baths hold at di\u000berent temperatures but the same chemical potential. The Red\feld equation for this system has been written down in Eq. (2.79). Following the general form in Eq. (2.85) and Eq. (2.82), the local- operator Lindblad equation and the multi-mode quantum master equation for this speci\fc system are, respectively, @\u001aL @t = \u0000i[HS ; \u001aL]\u0000 \u00152\u0019 n h1 + n (0; T1; \u00161)i h ay1; a1\u001a L i + hn (0; T1; \u00161)i h a1; a y 1\u001a L i + h1 + n (0; T2; \u00162)i h ay2; a2\u001a L i + hn (0; T2; \u00162)i h a2; a y 2\u001a L i + h:c: o : (2.93) 34 2.5. Evolution towards a NESS and @\u001aM @t = \u0000i[HS ; \u001aM ]\u0000 \u00152\u0019 n h1 + n (\u0000\u000f; T; \u0016)i h ay+; a+\u001a M i + hn (\u0000\u000f; T; \u0016)i h a+; a y +\u001a M i + h1 + n (\u000f; T; \u0016)i h ay\u0000; a\u0000\u001a M i + hn (\u000f; T; \u0016)i h a\u0000; a y \u0000\u001a M i + h:c: o ; (2.94) where n (\u000f; T; \u0016) is the average particle number in state with energy \u000f in a bath with temperature T and chemical potential \u0016. For the multi-mode quantum master equation, in which every eigenmode | not every site | is treated independently, it is impossible to incorporate the local coupling (the fact that only certain local sites couple to their own individual local baths), so we have just set all Tq = T . This makes the multi-mode quantum master equation not applicable to calculation of non-equilibrium stationary states. We will \fnd so in the following numerical study. Following a similar procedure like in the last section, we solve for the stationary states, \u001aL (1) and \u001aM (1). We compare them against the non- equilibrium stationary states of the Red\feld equation, \u001aR (1) using the following measure for distances: dAB = vuuuut P i;j \u0010 \u001aAij \u0000 \u001aBij \u00112 P ij \u0010 \u001aBjj \u00112 : (2.95) Typical results are shown in Fig. 2.5. We see that the di\u000berence between the multi-mode quantum master equation and the Red\feld equation is rela- tively small compared to the di\u000berence between the local-operator Lindblad equation and the Red\feld equation. At equilibrium (\u000e = 0), the multi-mode quantum master equation and the Red\feld equation lead to the same sta- tionary state, which is of course the thermal equilibrium. The local-operator Lindblad equation is quite di\u000berent from the Red\feld equation for all values of \u000e, and in particular it fails to predict the proper thermal equilibrium for unabiased baths. It can be shown that this is always the case, even for more general systems. However, we cannot discard the local-operator Lindblad equation on this basis, when analyzing transport, since we just argued that transport is re- lated only to o\u000b-diagonal matrix elements. Fig. 2.5(b) shows that while 35 2.5. Evolution towards a NESS 0 0.5 1 1.5 2 \u03b4 0 0.1 0.2 0.3 0.4 0.5 d dMR dLR 0 0.5 1 1.5 2 \u03b4 0 0.01 0.02 J JR JM JL Figure 2.5: \u001aL (1) and \u001aM (1) are compared with \u001aR (1). In (a) we plot dLR and dMR vs. the temperature bias \u000e. We see that at equilibrium (\u000e = 0), the di\u000berence between \u001aM (1) and \u001aR (1) is zero but the di\u000berence between \u001aL (1) and \u001aR (1) is not. For all other cases, the di\u000berences are \fnite. However, as shown in (b), the steady-state charge current JL and JR, re- spectively calculated from \u001aL (1) and \u001aR (1), show qualitatively similar behavior. On the other hand, JM = 0 shows that the multi-mode quantum master equation does not capture non-equilibrium stationary states at all. 36 2.5. Evolution towards a NESS the charge current from the multi-mode quantum master equation is al- ways zero, the currents from the local-operator Lindblad equation and the Red\feld equation are qualitatively similar. The \frst fact con\frms that the multi-mode quantum master equation captures properly only the equilib- rium distribution, but cannot describe transport through the system, as should be obvious from its structure. The second fact suggests that the local-operator Lindblad equation may be used to study, qualitatively, trans- port properties. However, since the ratio between the two currents changes for di\u000berent parameters, it is impossible to link JR quantitatively to JL. This qualitative similarity, if con\frmed for more complicated problems, could still be valuable for qualitative studies, for example in the question of normal v.s. anomalous conductance. More investigation are needed to check how universal this similarity is, but they are hampered by the lack of e\u000ecient solutions for the Red\feld equation. If con\frmed, this similarity would be useful since it is much easier to solve the local-operator Lindblad equation than the Red\feld equation [59]. While not a proof, the results shown here for a 2-site system coupled to 2 baths clearly illustrate why the local-operator Lindblad equation and the multi-mode quantum master equation approaches are questionable when studying transport through open systems. This is why we focus on the Red\feld equation from now on. To summarize, up to this point we have derived the general e\u000bective equation of motion and the more speci\fc Red\feld equation, and we have also illustrated the general procedure of solving the Red\feld equation by brute force direct methods. Before we develop more e\u000ecient approaches in the next chapters, we would also like to discuss its relation to the most general evolution equation, the semigroup Lindblad form. This is the topic of the next section. 2.5.3 The general Lindblad form The evolution operators of a closed quantum system H, U (t) = e\u0000iHt, form a group. In general, however, the evolution operators of an open system form a semigroup so as to keep the density matrix always positive and normalized, but not necessary reversible, \u001a (t) = U (t) \u001a (0) (2.96) such that U (t1 + t2) = U (t1)U (t2) : (2.97) For a closed system, we have the additional condition: U (\u0000t) = U\u00001 (t) (2.98) 37 2.6. Major challenge for large systems which insures reversibility. Lindblad discussed the generators of such trace-preserving positive semi- groups [90]. As shown in a theorem by Gorini, Kossakowski and Sudarshan [91], a completely positive semigroup evolution of a quantum system in a n-dimensional Hilbert space can be characterized by a Lindblad operator that governs its evolution d dt \u001a (t) = L\u001a (t) ; (2.99) where generally: L\u001a (t) = \u0000i h H\u0302; \u001a i + 1 2 n2\u00001X ij Aij \u0010h Li; \u001aL y j i + h Li\u001a; L y j i\u0011 ; (2.100) where the operators fLig form an orthonormal basis of the operators on the system's Hilbert space, and the constants coe\u000ecient matrix A must be positive. Note that the local-operator Lindblad equation is of the Lindblad form, where Li = a\u000b and Aij = Ai\u000eij , and \u000b = L;R stands for the left and the right baths respectively. This is why, especially in practical usage, some [59, 75] refer to the local-operator Lindblad equation as the Lindblad equation. More importantly, the Red\feld equation of Eq. (2.48) is not of the Lind- blad form. In principle, this means that it may lead to non-physical results such as non-positive density matrices, however we have never encountered such problems in any of our simulations. 2.6 Major challenge for large systems As it is clear from the examples presented above, the major challenge in solving the Red\feld equation has to do with the dimension of the vector P . This is 22N for a N -site system of spinless fermions, (2S + 1)2N for an N -site chain of spins-S, and in\fnity for bosonic systems. With such exponential growth, it is impossible to deal with large systems by direct methods, in- cluding the Runge-Kutta method, the eigenvector corresponding to the zero eigenvalue, the solution of the linear system and other propagator-based methods [64]. Using these methods, current computational power allows us to deal with systems with up to N = 10 [37, 38]. There are interesting physi- cal systems well within this size which already show rich behavior. However, such limitations on the system's size does not give us a lot of freedom to 38 2.6. Major challenge for large systems separate reliably bulk behavior from boundary e\u000bects, for larger systems. In contrast, quite often the non-equilibrium Green's function methods [4, 13] are capable of dealing with system sizes of up to N \u0018 200. The various methods mentioned in the introduction are either not e\u000e- cient enough, or involve additional approximations. As we mentioned earlier, e\u000ecient stochastic wave-function methods [65] have been developed for both the local-operator Lindblad equation and the Red\feld equation. This has better e\u000eciency then the direct methods and it is capable of dealing with roughly N \u0018 20. For the local-operator Lindblad equation only, an e\u000ecient method has been developed based on the time-dependent density matrix renormalization [59], capable of dealing with N \u0018 100. However, we have argued that the local-operator Lindblad equation's predictions are quantita- tively di\u000berent from those of the Red\feld equation. In the rest of this thesis, we will search for more e\u000ecient methods to calculate non-equilibrium sta- tionary states of large systems for the Red\feld equation, and discuss some preliminary applications. This chapter served as a foundation for our theory. None of it, except for the comparison between the Red\feld equation, the local-operator Lind- blad equation and the multi-mode quantum master equation, is our original contribution. These start in the next chapter. 39 Chapter 3 Linear Response Theory: Kubo formula for open systems 3.1 Introduction Consider a \fnite size one-dimensional system whose ends are coupled to two baths held at di\u000berent temperatures and\/or di\u000berent chemical potentials, leading to \row of energy or charge through the system. The question of interest to us is to \fnd the stationary state and thus physical quantities, in particular the steady-state charge and heat currents. Since it is hard to solve exactly for the non-equilibrium stationary states of large systems, in this chapter we introduce an approximate solution for the Red\feld equation valid for small biases. We call this approach the open Kubo formula as it is somewhat similar to the standard Kubo formula for isolated systems. While the size limitation of direct methods is around N = 10 (measured in qubit), in certain circumstances detailed below the open Kubo formula is capable of dealing with systems of up to N = 20 sites, the same size limitation as with exact diagonalization of Hamiltonians. The open Kubo formula gives the \frst-order corrections towards non- equilibrium stationary states for \fnite-size systems, starting from the cor- responding equilibrium states. Two forms of such Kubo-like formulae are proposed. The \frst, while computationally less e\u000ecient, gives proper general density matrices. The more e\u000ecient second form produces correct average values only for certain special physical quantities. Conditions for these spe- cial cases will also be discussed. Besides e\u000eciency, this work is also motived by considerations of the ap- plicability of the standard Kubo formula to \fnite-size open systems. The standard Kubo formula has been widely used for both in\fnite [42] and \fnite- size systems [82], usually with periodic boundary conditions. Open bound- ary conditions have also been considered, leading to qualitatively di\u000berent 40 3.2. General Linear Response Theory results [44]. For \fnite-size systems, the standard Kubo formula leads to a summation of many Dirac \u000e-function peaks at di\u000berent energies. Var- ious ways to get around or to smooth out such \u000e peaks have been pro- posed [92, 93]. However, there is no guarantee that correct physical quanti- ties can be extracted via such procedures. More importantly, we show that the standard Kubo formula gives the \frst-order corrections towards equilibrium states of the perturbed systems, instead of the desired non-equilibrium stationary states. We also prove that the standard formulas for Drude weight and dc conductance only work for in\fnite-size systems, not for \fnite-size ones. One way to solve these issues is to explicitly take into consideration the coupling to baths. This is why we are interested in linear response theory based on the Red\feld equation. This chapter is organized as follows. In Section x3.2 we brie\ry review the general linear response theory, and in Section x3.3 we discuss the ap- plicability of the standard Kubo formula. In Section x3.4 we present the open Kubo formula based on the Red\feld equation, and \fnally, in Section x3.5, for small systems, we compare the approximate solution from this open Kubo formula with the direct numerical solution of the Red\feld equation, in order to gauge its validity. 3.2 General Linear Response Theory Consider a general linear equation of motion, @\u001a (t) @t = (L0 +\u0001L) \u001a; (3.1) where L0 is the \\large\" term and \u0001L is the \\perturbation\". Explicit forms for L0 and \u0001L will be given later. We are particularly interested in station- ary states of the above equation, \u001a (1), given by: (L0 +\u0001L) \u001a (1) = 0: (3.2) Let us assume that \u001a0 is known and it satis\fes L0\u001a0 = 0: (3.3) Then, to \frst order, \u000e\u001a = \u001a\u0000 \u001a0 is the solution of @\u000e\u001a (t) @t = L0\u000e\u001a+\u0001L\u001a0; (3.4) 41 3.2. General Linear Response Theory where we have neglected the second order term \u0001L\u000e\u001a. A general solution of the above equation is \u000e\u001a (t) = Z t t0 d\u001ceL0(t\u0000\u001c)\u0001L\u001a0 + \u000e\u001a (t0) ; (3.5) Assuming that \u0001L was turned on at t0 = \u00001 when the initial state was \u001a0 so that \u000e\u001a (t0) = 0, we have the steady-state solution at t = 0 to be, \u000e\u001a (t = 0) = Z 0 \u00001 dte\u0000L0t\u0001L\u001a0: (3.6) Changing the variable t to \u0000t, we then arrive at \u000e\u001a (1) = Z 1 0 dteL0t\u0000\u0011t\u0001L\u001a0: (3.7) Here we have inserted an in\fnitesimal positive number \u0011 (\u0011 ! 0+) to make sure that the expression converges. This \u0011 is necessary if the real part of any eigenvalues of L0 are zero. The stationary state is \u001a (1) = \u000e\u001a (1)+ \u001a0. This is the basic formulation of linear response theory. It is required that every eigenvalue of L0 must have a negative or zero real part, so that the unperturbed system described by L0 eventually reaches a stationary state. If every eigenvalue of L0 happens to have a negative real part, as will be the case in one of the methods described below, then L0 is invertible and \u000e\u001a (1) = \u0000 (L0 \u0000 \u0011)\u00001\u0001L\u001a0: (3.8) This can also be derived from Eq. (3.2) since L0 is non-singular, 0 = (L0 +\u0001L) (\u001a0 + \u000e\u001a (1))) \u000e\u001a (1) = \u0000L\u000010 \u0001L\u001a0: (3.9) This is a straightforward formula if L0 is indeed invertible but it becomes tricky if L0 has zero eigenvalues. In this latter case, we can still use Eq. (3.7), but not Eq. (3.8). We may transform L0 to a non-singular \u0016L0 using the fact that \u001a is normalized, tr (\u001a) = 1 as we have done in rewriting Eq. (2.88) as Eq. (2.89). Following that we \fnd a corresponding expression similar to Eq. (3.8), more speci\fcally \u000e\u001a (1) = \u0000\u0016L\u000010 \u0001\u0016L\u001a0: (3.10) 42 3.3. Limitations of the standard Kubo formula 3.3 Limitations of the standard Kubo formula The Liouville-von Neumann equation of motion for the density matrix \u001a(t) of a closed system is: @\u001a (t) @t = LH\u001a (t) = \u0000i[H; \u001a]; (3.11) where H is its Hamiltonian. Long-term stationary solutions of this equation are not unique. A special class of such states are the Boltzmann equilibrium distributions: \u001aeq (H) = 1 Z e\u0000\fH ; (3.12) where Z = tr \u0000 e\u0000\fH \u0001 , \f = 1=T . If H = HS + Vext, where HS is the isolated system and Vext is a static weak coupling to an external \feld, one can formally use the above general linear response theory to \fnd a stationary solution \u001a (1) = \u001a0+\u000e\u001a (1) near a state \u001a0 of the unperturbed system, LHS\u001a0 = 0: \u000e\u001a (1) = Z 1 0 dteL0t\u0000\u0011tLVext\u001a0: (3.13) If \u001a0 = \u001aeq(HS) describes the unperturbed system in equilibrium, then this leads to the standard Kubo formula [6, 40]: \u000e\u001a (1) = \u0000i Z 1 0 dte\u0000\u0011t \u0014 Vext (\u0000t) ; e \u0000\fHS Z \u0015 ; (3.14) where Vext(t) = e iHStVexte \u0000iHSt. We can use the identity [Vext(\u0000t); e\u0000\fHS ] = \u0000ie\u0000\fHS R \f0 d\u001c _Vext (\u0000t\u0000 i\u001c), where _Vext (\u0000t\u0000 i\u001c) = ddtVext (t) jt=\u0000t\u0000i\u001c , to rewrite: \u000e\u001a (1) = \u0000 Z 1 0 dte\u0000\u0011t Z 0 d\u001c\u001a0 _Vext (\u0000t\u0000 i\u001c) : (3.15) In terms of the eigenvectors of HS , HS jni = \u000fnjni, hmj _Vext (t) jni = i (\u000fm \u0000 \u000fn) hm jVextjni ei(\u000fm\u0000\u000fn)t; (3.16) leading to: \u000e\u001a (1) = X m;n \u000fm 6=\u000fn e\u0000\f\u000fm \u0000 e\u0000\f\u000fn Z hm jVextjni \u000fm \u0000 \u000fn \u0000 i\u0011 jmihnj : (3.17) 43 3.3. Limitations of the standard Kubo formula Incidentally, note that there is no contribution from states with \u000fn = \u000fm, for which hmj _Vext (t) jni = 0. This also follows directly from Eq. (3.14); if we write Vext = V 0 ext+V ? ext, where V 0 ext = P \u000fm=\u000fn hm jVextjni jm ihnj commutes with HS , then [Vext(\u0000t); \u001a0] = [V ?ext(\u0000t); \u001a0]. The \\diagonal\" part V 0ext of Vext does not contribute to \u000e\u001a (1), and consequently has no in\ruence on the static response functions. The lack of diagonal contribution is, however, very puzzling. The well- known Drude weight, which is derived from the standard Kubo formula and is used quite often when discussing charge or thermal transport [42, 44, 82] reads D = \u0019 L X m;n \u000fm=\u000fn e\u0000\f\u000fm Z jhmjJ\u0302 jnij2: (3.18) It has contributions only from states with \u000fm = \u000fn. To understand the reason for this di\u000berence, consider the derivation of Eq. (3.18) from Eq. (3.15), e.g. for spinless fermions in a one-dimensional chain (lattice constant a = 1), described by HS = \u0000t X l \u0010 cyl cl+1 + h:c: \u0011 + V0 X l nlnl+1; (3.19) where nl = c y l cl, plus a static electric potential Vext = X l Vlnl (3.20) induced by a homogeneous electric \feld E = \u0000rV . From the continuity equation: _Vext(t) = X l Vl d dt nl(t) = \u0000 X l Vl[Jl+1(t)\u0000 Jl(t)]; (3.21) where Jl = it \u0000 c+l+1cl \u0000 c+l cl+1 \u0001 is the local current operator. The sum can be changed to \u0000 X l [Vl\u00001Jl(t)\u0000 VlJl(t)] = \u0000EJ(t); (3.22) where J(t) = P l Jl(t) is the total current operator. Using _Vext(t) = \u0000EJ(t) (or, on a lattice, E = Vl\u00001 \u0000 Vl) in Eq. (3.15) gives \u000e\u001a (1) = E Z 1 0 dte\u0000\u0011t Z 0 d\u001c\u001a0J (\u0000t\u0000 i\u001c) : (3.23) 44 3.3. Limitations of the standard Kubo formula The dc conductivity is then \u001b = Z 1 0 dte\u0000\u0011t Z 0 d\u001c hJ (\u0000t\u0000 i\u001c) Ji ; (3.24) where hOi = Tr[\u001aeq(HS)O]. This expression can be further simpli\fed to arrive at Eq. (3.18). The only questionable step in this derivation, and the one responsible for going from a result with no contributions from states with \u000fn = \u000fm, to one with contributions only from these states, is the changeX l VlJl+1(t)! X l Vl\u00001Jl(t): (3.25) This is only justi\fed for an in\fnite system (where boundary terms are pre- sumed to be negligible), or a system with periodic boundary conditions and an external \feld with the same periodicity. This latter condition can only be achieved for a charge current in a \fnite system with periodic bound- ary conditions driven by a varying magnetic \rux through the area enclosed by the system. For a \fnite-size system (even one with periodic boundary conditions) in a static applied electric \feld this approach is not valid. The same is true for thermal transport, which cannot experimentally be induced in a system with periodic boundary conditions. In both cases, the physical relevance of the results of Eq. (3.18) are hard to fathom. There is an even more serious conceptual problem with the standard Kubo formula: the resulting distribution ~\u001a = \u001aeq(HS) + \u000e\u001a (1) is the \frst order perturbational expansion of \u001aeq (H), not the expected non-equilibrium stationary state. This statement is proved below, where for convenience, we assume that the diagonal part V 0ext = 0. If it is not, we simply remove the \\diagonal\" part V 0ext from Vext and add it to HS . Consider then the eigenstates of the full Hamiltonian, Hj~ni = ~\u000fnj~ni, to \frst order perturbation in Vext. Since hm jVextjni = 0 for all \u000fm = \u000fn, we can apply the \frst order perturbation theory for non-degenerate states to all the states, whether degenerate or not, to \fnd ~\u000fn = \u000fn +O \u0000 V 2ext \u0001 (3.26) and j~ni = jni+ X m;\u000fm 6=\u000fn hmjVextjni \u000fn \u0000 \u000fm jmi+O \u0000 V 2ext \u0001 : (3.27) 45 3.3. Limitations of the standard Kubo formula This immediately leads to \u001aeq (H) = X n 1 ~Z e\u0000\f~\u000fn j~nih~nj = \u001aeq(HS) + \u000e\u001a (1) +O(V 2ext); (3.28) where \u000e\u001a (1) is indeed given by Eq. (3.17). Intuitively this is easy to un- derstand. In a sense, we found the original thermal equilibrium distribution \u001aeq(HS) from the Liouville equation with H = HS . Then, it is not so sur- prising that a perturbation based on the same Liouville equation for the full H = HS + Vext leads to \u001aeq (H). This veri\fes our previous statement that the standard Kubo formula implies \u001a (1)! \u001aeq(H). This is rather problematic because normally { for example if invariance to time reversal symmetry is not broken { no currents are generated in a thermal equilibrium state and therefore no steady-state transport through the closed \fnite system can be described by this approach. However, because we only keep the \frst-order perturbational correction, the situation is less clear-cut. In principle, it is not impossible for \u000e\u001a of Eq. (3.17) to also be a \frst order approximation to the true non-equilibrium stationary state, or at least to capture a sizable part from it. This needs to be investigated in more detail, and we do so below. Before that, let us also note that the fact that the use of Eq. (3.18) and its equivalents for \fnite size systems is problematic can also be seen from the following technical considerations. The spectrum of a \fnite-size system is always discrete. As a result, any pair of degenerate eigenstates \u000fm = \u000fn gives a \u000e-function contribution to the response functions. Such a singular response is unphysical for \fnite-size systems (for an in\fnite system, the integration over the continuous spectrum removes these singularities). Various techniques have been proposed to smooth out these singular con- tributions in order to extract some \fnite values, such as use of imaginary frequencies [93] or averaging \u001b (!) over a small range of frequencies \u000e! and then taking \u000e! ! 0 [92]. These di\u000berent approaches may lead to di\u000berent results. Moreover, the order in which the various limits are approached, e.g., taking \u0011 ! 0 before L!1 or vice versa, also make a di\u000berence [92]. All of these subtleties of the standard Kubo formula are related to the potential divergence whenever \u000fm = \u000fn in Eq. (3.18). To summarize, using the standard Kubo formula blindly for \fnite-size systems is fraught with both conceptual and technical problems. Its solution gives the \frst order correction of the equilibrium state of the full Hamilto- nian, instead of the expected non-equilibrium stationary state. The derived Drude weight formula is applicable only for in\fnite-size systems; applying 46 3.4. OKF: the Kubo formula for open \fnite-size systems it directly to \fnite-size systems involves dealing with unphysical \u000e-function peaks. In order to \fx these di\u000eculties, we now discuss how to calculate the non-equilibrium stationary states from the Red\feld equation, which takes the connection between system and baths explicitly into consideration. 3.4 Open Kubo formula: linear response theory for open \fnite-size systems For concreteness, let us assume that the central system is coupled to thermal baths kept at temperatures TL=R = T \u0006 \u0001T2 and investigate the thermal transport in the resulting steady state. If \u0001T \u001c T , this will lead to a Kubo- like formula which replaces Eq. (3.15). This approach can be generalized straightforwardly to derive a Kubo-like formula for charge transport. The general Red\feld equation can be written as: @\u001a(t) @t = [LH + LL(TL) + LR(TR)] \u001a(t); (3.29) where LH\u001a = \u0000i[H; \u001a], just like for an isolated system, while LL=R are additional terms that describe the e\u000bects of the left\/right thermal baths (assumed to be in equilibrium at their corresponding temperatures TL=R) on the evolution of the system. We have seen from Eq. (2.48) how the expressions for LL=R depend on the Hamiltonian H of the system and on its coupling to the baths. Generally speaking, H could also contain an external potential due to coupling to the external \feld, Vext, such thatH = HS+Vext. Such a term is expected to appear for charge transport, but not for thermal transport, therefore in the following we set Vext = 0. As we have seen in Eq. (2.48), the temperature and chemical potential of the left (right) baths enter the Red\feld equation only via average particle numbers n (\u000f; TL; \u0016L) (n (\u000f; TR; \u0016R)). Therefore, if \u0001T \u001c T , we can Taylor expand the particle number into a large part and a small part. Thus we also expand the LL=R and re-arrange the Red\feld equation to read: @\u001a(t) @t = [LHS + LB(T ) + LP (\u0001T )] \u001a(t) = L\u001a(t); (3.30) where LB(T ) = LR(T ) + LL(T ) is the contribution from the thermal baths if both are kept at the same temperature, while LP (\u0001T ) collects the terms proportional to \u0001T . 47 3.4. OKF: the Kubo formula for open \fnite-size systems We are interested in the t ! 1, stationary state solution \u001a (1) of the above equation, satisfying both Eq. (2.88) and Eq. (2.89), which are rewrit- ten in the following form for convenience, L\u001a (1) = 0; (3.31) and \u0016L\u001a (1) = \u0017: (3.32) As we have shown in examples in the previous chapter, L has one non- degenerate zero eigenvalue and all its other (transient) eigenvalues have a negative real part, so \u001a (1) is unique for any value of \u0001T . We have also shown that for small systems, Eq. (3.31) and Eq. (3.32) can be solved numerically. We call this solution \u001aex and we use it to validate the solutions of various approximation schemes. A Kubo-like formula, which is potentially more e\u000ecient, can be obtained using linear response theory. The \frst step is to separate the Liouvillian L of Eq. (3.30) into a \\large\" plus a \\small\" part. There are two possible choices: ( L (1) 0 = LHS + LB(T ) \u0001L(1) = LP (\u0001T ) (3.33) or ( L (2) 0 = LHS \u0001L(2) = LB(T ) + LP (\u0001T ) (3.34) We begin with the \frst choice. Assume that L (1) 0 has eigenvalues n L (1) 0;\u0016 o and left\/right eigenvectors fjL\u0016)g, fjR\u0016)g. As discussed, the unique (zero order in perturbation theory) steady-state solution of L (1) 0 \u001a0 = 0 is \u001a0 = \u001aeq(HS). The deviation \u000e\u001a (1) K due to the perturbation \u0001L (1) is obtained like in Eq. (3.13): \u000e\u001a(1) (1) = X \u0016 Z 1 0 dteL (1) 0;\u0016t\u0000\u0011t jR\u0016) (L\u0016j\u0001L(1)\u001a0 = \u0000 X \u0016 jR\u0016) (L\u0016j L (1) 0;\u0016 \u0000 \u0011 \u0001L(1)\u001a0 = \u0000 X \u0016>0 jR\u0016) (L\u0016j L (1) 0;\u0016 \u0001L(1)\u001a0: (3.35) Note that the only divergent term, due to L (1) 0;0 = 0, disappears because (L0j\u0001L(1)\u001a0 = (\u001a0j\u0001L(1)\u001a0 = 0. To see why, we start from Eq. (3.31), 48 3.4. OKF: the Kubo formula for open \fnite-size systems L(\u001a0 + \u000e\u001a) = 0, project it on (\u001a0j and keep terms only to the \frst order, to \fnd 0 = (\u001a0j \u0010 L (1) 0 +\u0001L (1) \u0011 (\u001a0 + \u000e\u001a) = (\u001a0j\u0001L(1)\u001a0 since L(1)0 \u001a0 = 0. As a result, Eq. (3.35) has only regular contributions. A similar approach has been suggested in Ref. [88], but for the local-operator Lindblad equa- tion [90] instead of the Red\feld equation. Eq. (3.35) is di\u000ecult to use in practice: \fnding all eigenstates of L (1) 0 is a hard task unless the system has an extremely small Hilbert space. A computationally simpler solution is obtained from Eq. (3.32), where, in matrix terms, \u0016L is de\fned by replacing the \frst row of the equation L\u001a (1) = 0 by Tr (\u001a (1)) = 1, so that \u0017 is a vector whose \frst element is 1, all remaining ones being 0. As a result det \u0000 \u0016L \u0001 6= 0 while det (L) = 0. We can also solve it to obtain a Kubo-like formula by dividing \u0016L = \u0016L (1) 0 +\u0001 \u0016L(1). Again, the overbar shows that in matrix terms, \u0016L (1) 0 is obtained from L (1) 0 by replacing its \frst row with Tr (\u001a (1)) = 1, while \u0001\u0016L(1) is obtained from \u0001L(1) by replacing its \frst row with zeros. We \fnd \u000e\u0016\u001a(1) (1) = \u0000[ \u0016L(1)0 ]\u00001\u0001\u0016L(1)\u001a0: (3.36) This is much more convenient because inverting the non-singular matrix \u0016L (1) 0 is a much simpler task than \fnding all the eigenvalues and eigen- vectors of L (1) 0 . We have veri\fed that both schemes produce identical re- sults on systems on which both of them can be performed. We denote \u001a0 + \u000e\u0016\u001a (1) (1) = \u001a(1) (1). The second option is to take L (2) 0 = LHS and \u0001L (2) = LB(T )+LP (\u0001T ). In this case, we can still choose the stationary solution associated with L (2) 0 to be the thermal equilibrium state at T , \u001a0 = \u001aeq(HS). However, this solution is no longer unique, since any matrix \u001a0 that is diagonal in the eigenbasis of HS satis\fes L (2) 0 \u001a0 = 0. Expanding the corresponding analogue of Eq. (3.14) in the eigenbasis of HS , we now \fnd: \u000e\u001a(2) (1) = \u0000i X n;m hmj\u0001L(2)\u001a0jni \u000fm \u0000 \u000fn \u0000 i\u0011 jmihnj: (3.37) We call \u001a0 + \u000e\u001a (2) (1) = \u001a(2) (1). Note that unlike \u001a(1) (1) of Eq. (3.36), this solution has divergent contributions from states with \u000fn = \u000fm. As such, it is analogous to the standard Kubo formula for in\fnite systems. This is not an accident. As discussed, the standard Kubo formula for in\fnite systems always ignores the coupling to the leads. It also takes L0 = LHS and assumes that \u001a0 = \u001aeq(HS). Moreover, the driving force for transport 49 3.5. Comparison of the two OKFs is only the potential Vext added to HS , so that \u0001L ! LVext . With these assumptions, Eq. (3.37) maps straightforwardly into the standard Kubo formula of Eq. (3.17). 3.5 Comparison of the two open Kubo formulae To see which of these two solutions { the regular solution \u001a(1) (1) or the singular solution \u001a(2) (1) { gives a proper density matrix, we compare them against the exact numerical solution \u001aex of Eq. (3.31) in the limit \u0001T \u001c T . We do this for an N -site chain of spinless fermions HS = \u0000t N\u00001X l=1 \u0010 cyl cl+1 + c y l+1cl \u0011 + V0 N\u00001X l=1 cyl+1cl+1c y l cl: (3.38) coupled to two heat baths, modeled as collections of fermions: HB = X k;\u000b=L;R !k;\u000bb y k;\u000bbk;\u000b; (3.39) where \u000b indexes the left and right-side baths and we set ~ = 1; kB = 1, the lattice constant a = 1, and hopping t = 1. The system-baths coupling is chosen as: VSB = \u0015 X k; V \u000bk \u0010 cy\u000bbk;\u000b + c\u000bb y k; \u0011 ; (3.40) where the left (right) bath is coupled to the \frst (last) site: cL = c1 and cR = cN . Bath parameters, i.e. the temperature and chemical potential, are chosen to be (TL; \u0016) and (TR; \u0016) with TL=R = T \u0006 \u0001T2 . The corresponding Red\feld equation reads from Eq. (2.48), @\u001a(t) @t = \u0000i[HS ; \u001a(t)] \u0000\u00152 X \u000b=L;R nh cy\u000b; m\u0302\u000b\u001a(t) i + \u0002 c\u000b; \u0016\u0302m\u000b\u001a(t) \u0003 + h:c: o ; (3.41) where operators m\u0302\u000b and \u0016\u0302m\u000b are de\fned by Eq. (2.49), which is rewritten here for convenience: m\u0302\u000b = \u0019 X m;n jmihnjhmjc\u000bjni (1\u0000 n\u000b ( nm)) ; (3.42a) \u0016\u0302m\u000b = \u0019 X m;n jmihnjhmjcy\u000bjnin\u000b ( mn) : (3.42b) 50 3.5. Comparison of the two OKFs 0 0.2 0.4 0 0.02 0.04 0.06 0.08 D \u03b4T (\u03bb=0.1, V0=0.2) 0 40 80 120 160 D2 D1 (a) 0 0.2 0.4 \u03b4T (\u03bb=0.1, V0=0.2) 0 0.002 0.004 0.006 0.008 J Jex J(1) J(2) 0 0.2 0.4 2.5 3 3.5 -900 -450 0 N(2) Nex N(2) (b) Figure 3.1: (a) D1 (squares) and D2 (circles) of Eq. (3.43) vs. \u000e = \u0001T=2T , for N = 8; t = 1:0; V0 = 0:2; T = 2:0; \u0016 = \u00001:0; \u0015 = 0:1; \u0011 = 0:00001. (b) the steady-state electric current calculated with \u001aex (triangles) and \u001a (1;2) (1) (squares, circles and stars). The inset shows the corresponding particle numbers. Eq. (3.41) is an example of the general Eq. (3.29). Since TL=R enters only in the Fermi-Dirac distributions, it is easy to expand when TL=R = T \u0006 \u0001T2 ;\u0001T \u001c T to identify LB(T ) and LP (\u0001T ). We characterize the distance between the exact numerical solution \u001aex and the two possible Kubo solutions \u001a(i) (1), i = 1; 2 using the norm: Di = sX n;m \f\fhnj\u001aex \u0000 \u001a(i) (1) jmi\f\f2: (3.43) For the proper solution, this di\u000berence should be small but still \fnite, be- cause of higher-order perturbation terms. Results typical of those found in all the cases we investigated are shown for N = 8; V0 = 0:2; \u0015 = 0:1; \u0011 = 10 \u00005 in Fig. 3.1(a), where we plot D1;2 vs. \u000e = \u0001T=2T . We see that D2 (circles, axis on the right) is very large. In fact, because of the singular contributions from \u000fn = \u000fm states, D2 is divergent, with a magnitude controlled by the cuto\u000b \u0011. In contrast, D1 (squares, left axis) is small and independent of \u0011. Fig. 3.1(b) shows the electric current (J) and the total number of particles (N) calculated with \u001aex, and \u001a (1)(1;2)(triangles, squares, respectively circles). Both N (1) and J (1) are very close to the exact values N ex, Jex. However, N (2) is very di\u000berent from N ex while J (2) is close to Jex. These results con\frm that \u001a(1) (1) of Eq. (3.36) is the proper Kubo solution. They also show that \u001a(2) (1) can also be used, but only for quantities A for which hm jAjni = 0 51 3.5. Comparison of the two OKFs 0 0.05 0.1 0.15 0.2 \u03bb(\u03b4T=0.1, V0=0.2) 0 1 2 3 4 5 6 D D1 D2 (a) 0 0.05 0.1 0.15 0.2 \u03bb(\u03b4T=0.1, V0=0.2) 0 0.0005 0.001 0.0015 0.002 0.0025 0.003 J Jex J(1) J(2) 0 0.05 0.1 0.15 0.2 2.5 3 3.5 -50 -40 -30 -20 -10 0 N(2) Nex N(1) (b) Figure 3.2: (a)D1 (squares) and D2 (circles) of Eq. (3.43) vs. \u0015, for N = 8; t = 1:0; V0 = 0:2; T = 2:0; \u0016 = \u00001:0; \u000e = 0:1; \u0011 = 0:00001. (b) the steady-state electric current calculated with \u001aex (triangles) and \u001a (1;2) (1) (squares and circles). The inset shows the corresponding particle numbers. whenever \u000fm = \u000fn, so that the divergences in Eq. (3.37) disappear. This explains the previous success of this formula to be somewhat of an accident. Fig. 3.2 shows the same quantities as Fig. 3.1, at a \fxed bias \u0001T as a function of the strength of the system-bath coupling \u0015. It con\frms again that \u001a(1) (1) is the proper approximation of \u001aex, but that for the charge current \u001a(2) (1) works well too. This also shows that the non-equilibrium stationary states depend on the coupling strength \u0015. This is not surprising for a \fnite-size system: the intrinsic conductance of the system is added to comparable \\contact\" contributions from the interfaces between the system and the baths, and experiments measure the total conductance. It follows that quantitative modeling of transport in \fnite systems will require a careful consideration of the entire experimental set-up. Fig. 3.3 shows again the same quantities as Fig. 3.1, at a \fxed bias \u0001T as a function of the strength of the interaction V0. It con\frms again that \u001a(1) (1) is the proper approximation of \u001aex. We also \fnd that values of currents Jex are quite di\u000berent between V0 = 0 and V0 = 1 for example. The charge currents calculated from both \u001a(1;2) (1) indeed capture the major part of such di\u000berence. One may note a discontinuity in the data around V0 = 0:6. We think this is due to numerical reasons. The calculation depends very strongly on the values of energy di\u000berences among all energy levels. For di\u000berent values of V0, such energy di\u000berences may either shift continuously or change qualitatively. 52 3.6. Summary and discussion 0 0.2 0.4 0.6 0.8 1 V0(\u03bb=0.1, \u03b4T=0.1) 0 0.5 1 1.5 2 2.5 3 D D1 D2 (a) 0 0.2 0.4 0.6 0.8 1 V0(\u03bb=0.1, \u03b4T=0.1) 0.00065 0.0007 0.00075 0.0008 J Jex J(1) J(2) 0 0.2 0.4 0.6 0.8 1 -8 -4 0 4 Nex N1 N2 (b) Figure 3.3: D1 (squares) and D2 (circles) of Eq. (3.43) vs. V0, for N = 8; t = 1:0; T = 2:0; \u0016 = \u00001:0; \u0015 = 0:1; \u000e = 0:1; \u0011 = 0:00001. (b) the steady- state electric current calculated with \u001aex (triangles) and \u001a (1;2) (1) (squares and circles). The inset shows the corresponding particle numbers. 3.6 Summary and discussion To conclude, we \frst showed that the standard Kubo formula fails to provide an approximation for non-equilibrium stationary states, instead approximat- ing the equilibrium state of the whole Hamiltonian. Secondly, applying the standard Kubo formula to \fnite systems generically leads to divergences, which are unphysical. We then showed that taking into consideration the coupling to baths explicitly solves both problems and leads to a well-behaved Kubo-like formula. Finally, we showed that the improper solution similar to those used in literature can give correct average values but only for certain physical quantities. Although we only considered small systems so as to be able to calculate \u001aex, Eqs. (3.36) and (3.37) can be used for larger systems, with the latter being more e\u000ecient but not always valid. In fact, the above results give us enough grounds to question the validity of the standard Kubo formula, where the e\u000bect of the baths are taken care of only via an additional potential. We believe that this is not enough. There are two di\u000berent types of \\driving forces\" responsible for a steady-state current \row. The \frst force is the applied electric \feld, which gives rise to an additional interaction term Vext to be included in the total Hamiltonian H = HS + Vext of the system. The second one is the imbalance in the chemical potentials and\/or temperatures of the two baths, which results in a non-equilibrium distribution function \u001a (1). As we have shown above, the standard Kubo formula calculates the current from \u001aeq(H), the thermal equilibrium state of the full Hamiltonian, so it takes into account only the 53 3.6. Summary and discussion e\u000bect of Vext, but not the full non-equilibrium distribution function. We believe that this approach is problematic when calculating charge currents. This approach is even less reliable for thermal currents, where there is only one driving force: the non-equilibrium distribution. There is no thermal equivalent for the electric potential. Some additional assumptions such as local equilibrium [39] or use of gravitational potentials [40] have been used to make the situation similar to that of charge transport. Either way, an arti\fcial thermal potential term Vext is constructed and added to HS and then one uses \u001aeq(H) instead of the proper non-equilibrium stationary state, in direct analogy with the usual approach for charge currents. The work presented in this chapter points to a solution for these concep- tual problems, if the baths are explicitly taken into consideration. However, the resulting open Kubo formula can still only be used for system of up to N \u0018 20, which may be too small for transport in solid state systems. In the next two chapters, we present two e\u000ecient ways to solve the Red\feld equation for larger systems. Furthermore, unlike the open Kubo formula based on linear response theory, the new methods are applicable even when the baths' bias is not small. 54 Chapter 4 BBGKY-like hierarchy for the Red\feld equation 4.1 Introduction In this chapter we introduce a much more e\u000ecient way to solve the Red\feld equation, based on a BBGKY-like hierarchy. Its e\u000eciency depends on the level at which the hierarchy is truncated. At \frst order, where only single- particle Green's functions are involved, it leads to a linear system with N2 unknowns for a N -site spinless fermionic system. This makes it possible to analyze systems larger than those that can be treated with the non- equilibrium Green's function method. Like the diagrammatic perturbation theory [18], the BBGKY equation hierarchy for systems in equilibrium also provides an equivalent systematic approach to calculate many-particle correlation functions [68, 69]. For non- interacting systems, the hierarchy is decoupled, meaning that the equation of a single-particle Green's function is only related to other single-particle Green's functions, forming a closed system of equations. The same holds for general n-particle Green's functions denoted as Gn. However, when there are interactions in the system, equations of all orders of Green's functions are coupled together. Generically, equations for n-particle Green's functions, Gn, also involve Gn+1, so that the system of BBGKY equations becomes in\fnite. In this case, approximations such as the cluster expansions can be used to truncate and then solve the truncated hierarchy [69, 94]. Here we use exactly the same idea but now for the Red\feld equation. Not surprisingly, for non-interacting central systems we \fnd that the hierarchy is decoupled, just like for systems in equilibrium. In fact, in Ref. [36], the Red\feld equation with N \u0018 100 has been solved in terms of single-particle Green's functions. However, they did not regard this as a special case of the more general BBGKY-like equation hierarchy. From this point of view, our idea is basically to extend this Green's function based solution of the Red\feld equation from Ref. [36] to interacting systems, where equations for 55 4.2. Derivation of the BBGKY-like hierarchy Green's functions of various orders become coupled. Two systematic approximations are proposed to truncate and solve the hierarchy. As a test of reliability of the proposed methods, we apply them to small systems which can also be solved with exact direct methods. Consis- tent results are found. Just as the usual Wick's theorem of non-interacting systems provides a basis for the usual diagrammatic perturbation theory of interacting systems, one of the approximations requires a non-equilibrium Wick's theorem. Such a theorem is proved starting from the Red\feld equa- tion of non-interacting systems. This chapter is organized as follow. In Section x4.2, using a speci\fc exam- ple, a BBGKY-like equation hierarchy is derived from the Red\feld equation. In Section x4.3, non-interacting systems are discussed and a non-equilibrium Wick's theorem is proved. In Section x4.4, we introduce the two methods to truncate and solve the coupled hierarchy for interacting systems, and com- pare them against exact results. We \fnd that both methods are signi\fcantly more e\u000ecient than the direct methods. The \frst is capable of dealing with relatively small systems but with strong interactions, while the second can deal with much larger systems but with relatively weaker interactions. For the latter case, we demonstrate that a systematic increase in the level at which the truncation is performed leads to a systematic improvement of the results, as well as an increase in the range of interactions for which good accuracy is obtained. 4.2 Derivation of the BBGKY-like hierarchy For concreteness in presenting our general formulation, let us start from the same Red\feld equation describing an N -site chain of spinless fermions coupled with two fermionic baths. Our system of interest is described by: HS = \u0000t N\u00001X l=1 \u0010 cyl cl+1 + c y l+1cl \u0011 + V0 N\u00001X l=1 cyl+1cl+1c y l cl = H0 + VS ; (4.1) and its Red\feld equation, given by Eq. (3.41), is rewritten here for conve- nience: @\u001a(t) @t = \u0000i[HS ; \u001a(t)]\u0000 \u00152 X \u000b=L;R nh cy\u000b; m\u0302\u000b\u001a(t) i + \u0002 c\u000b; \u0016\u0302m\u000b\u001a(t) \u0003 + h:c: o ; (4.2) 56 4.2. Derivation of the BBGKY-like hierarchy where the operators m\u0302 and \u0016\u0302m are de\fned as follows: m\u0302\u000b = X k jV \u000bk j2 Z 1 0 d\u001cc\u000b (\u0000\u001c) e\u0000i!k;\u000b\u001c h1\u0000 n (!k;\u000b)i ; (4.3a) \u0016\u0302m\u000b = X k jV \u000bk j2 Z 1 0 d\u001ccy\u000b (\u0000\u001c) ei!k;\u000b\u001c hn (!k;\u000b)i : (4.3b) Here, \u000b = L;R refer to the two baths, c\u000b correspond to the two sites at the boundaries coupled each to its corresponding bath, !k;\u000b are the energies of the modes in bath \u000b, and hn (!k;\u000b)i are their average occupation numbers. If U (t) = e\u0000iHSt is known, then so are c\u000b (t) = U y (t) c\u000bU (t) and there- fore the operators m\u0302. Finding explicit forms for m\u0302 and \u0016\u0302m thus requires a full diagonalization of HS . Using its eigenvectors, one can perform the above integrals to get these operators. More details are included in Appendix A. Again we are only interested in the long-time steady-state solution \u001a (1) and we want to calculate the values of various correlation functions. For a physical quantity of the central system, denoted by the operator A, from Eq. (4.2) we \fnd: 0 = i h[A;H0]i+ i h[A; VS ]i+ \u00152 X nDh A; cy i m\u0302 E + [A; c\u000b] \u0016\u0302m \u000b\u0000 Dm\u0302y\u000b [A; c\u000b]E\u0000 D \u0016\u0302my\u000b hA; cy\u000biEo ; (4.4) where hAi = tr (A\u001a (1)) and m\u0302y\u000b( \u0016\u0302my\u000b) is the hermitian conjugate of m\u0302\u000b( \u0016\u0302m\u000b). Here we have used the cyclic property of the trace operator. All equations of Green's functions in the rest of this paper will be derived from this equa- tion. For example, the \frst and the second equation of the hierarchy for respectively single-particle Greens functions: G1 \u0010 my; n \u0011 = D cymcn E (4.5) and two-particle Greens functions: G2 \u0010 my; ny;m 0 ; n 0\u0011 = D cymc y ncm0 cn0 E (4.6) can be derived by using A = cymcn, respectively A = c y mc y ncm0 cn0 in Eq(4.4): 57 4.2. Derivation of the BBGKY-like hierarchy 0 = it D cym\u00001cn E + it D cym+1cn E \u0000 it D cymcn+1 E \u0000 it D cymcn\u00001 E (4.7a) \u0000iV0 D cymc y n\u00001cncn\u00001 E + iV0 D cym+1c y mcm+1cn E \u0000iV0 D cyn+1c y mcn+1cn E + iV0 D cymc y m\u00001cncm\u00001 E (4.7b) \u0000\u00152 X D \u000em\u000bcn \u0016\u0302m\u000b + \u000en\u000b \u0016\u0302m y \u000bc y m \u0000 \u000en\u000bcymm\u0302\u000b \u0000 \u000em\u000bm\u0302y\u000bcn E ; (4.7c) and 0 = it D cym\u00001c y ncm0 cn0 E + it D cym+1c y ncm0 cn0 E + it D cymc y n\u00001cm0 cn0 E +it D cymc y n+1cm0 cn0 E \u0000 it D cymc y ncm0 cn0+1 E \u0000 it D cymc y ncm0 cn0\u00001 E \u0000it D cymc y ncm0+1cn0 E \u0000 it D cymc y ncm0\u00001cn0 E +iV0 D cymc y ncm0 cn0 E\u0010 \u000em0+1;n0 + \u000em0\u00001;n0 \u0000 \u000em+1;n \u0000 \u000em\u00001;n \u0011 (4.8a) \u0000iV0 X l=m\u00061;n\u00061 D cyl c y mc y nclcm0 cn0 E +iV0 X l=m0\u00061;n0\u00061 D cyl c y mc y nclcm0 cn0 E (4.8b) \u0000\u00152 X n \u000em0 D cymc y ncn0 m\u0302 E \u0000 \u000en0 D cymc y ncm0 m\u0302 E +\u000em D cyncm0 cn0 \u0016\u0302m E \u0000 \u000en D cymcm0 cn0 \u0016\u0302m E + \u000en0 D \u0016\u0302my\u000bc y mc y ncm0 E \u0000\u000em0 D \u0016\u0302my\u000bc y mc y ncn0 E + \u000en D m\u0302y\u000bc y mcm0 cn0 E \u0000 \u000em D m\u0302y\u000bc y ncm0 cn0 Eo : (4.8c) Note that since the set of all polynomials of n cl; c y l o form a complete basis of the operator space, operators m\u0302 are certainly functions of polynomials of n cl; c y l o . Therefore, we expected G1 to depend on G2 from Eq(4.7b), and possibly also G3 or higher Green's functions from Eq(4.7c); while G2 is coupled to G3 from Eq(4.8b), and possibly also G4 or higher Green's functions from Eq(4.8c). Solving this full hierarchy is no easier than directly solving the Red\feld equation, unless VS = 0 so that the set of equations for G1 is closed, i.e. not coupled to G2. We may, however, solve these equations by truncating the hierarchy at a certain level using further approximations, such as the molecular-chaos 58 4.3. Non-equilibrium Wick's theorem assumption in the classical Boltzmann equation [51], or replacing higher order Green's functions by a cluster expansion of lower order ones [69, 94, 95]. We use the following two approximate methods: (1) substitution of certain higher-order Green's functions by their values at equilibrium; (2) expressing higher-order Green's functions as combinations of lower-order ones plus a correlation part via cluster expansion, and then ignoring the correlation part at a certain order. A rough estimate of the accuracy of these substitutions at di\u000berent orders can be found in Appendix C. Here we focus on the potential of this BBGKY-like formulation and discuss brie\ry the performance of the two approximations. 4.3 Non-interacting systems: proof for a non-equilibrium Wick's theorem In this section, we will prove that when V0 = 0, G2 \u0010 ky1; k y 2; k3; k4 \u0011 = G1 \u0010 ky1; k4 \u0011 G1 \u0010 ky2; k3 \u0011 \u0000G1 \u0010 ky1; k3 \u0011 G1 \u0010 ky2; k4 \u0011 ; (4.9) where G1 and G2 are de\fned as G1 \u0010 ky1; k2 \u0011 = D cyk1ck2 E ; (4.10) respectively G2 \u0010 ky1; k y 2; k3; k4 \u0011 = D cyk1c y k2 ck3ck4 E : (4.11) In a general non-equilibrium state, one may have hck3ck4i 6= 0. In that case, the above Wick's theorem should have a more general form. In all of our examples, our choice of the speci\fc system-baths coupling makes such Green's functions vanish. If there are terms involving two creation or two annihilation operators in the coupling between the central system and the baths, the situation is di\u000berent. Here it is more convenient to work in the momentum representation than the position representation, therefore we de\fne: ck = 1p N NX l=1 sin kl\u0019 N + 1 cl: (4.12) 59 4.3. Non-equilibrium Wick's theorem For a tight-binding chain with N sites and free ends (c0 = 0 = cN+1), the usual plane-waves of in\fnite-size systems eikl are replaced by the eigenstates sin kl\u0019N+1 , hence this transformation. In this representation, H0 = NX k=1 \u000fkc y kck; (4.13) where \u000fk = \u00002t cos \u0019k N + 1 : (4.14) Starting from Eq(4.4) with the above H0 in momentum space and using A = cyk1ck2 and A = c y k1 cyk2ck3ck4 , we \fnd the equations for G1 \u0010 ky1; k2 \u0011 and respectively G2 \u0010 ky1; k y 2; k3; k4 \u0011 as follows: 0 = i (\u000fk2 \u0000 \u000fk1)G1 \u0010 ky1; k2 \u0011 \u0000\u00152 2\u0019 N + 1 X sin k1\u0019l N + 1 sin k2\u0019l N + 1 (n (k1) + n (k2)) +\u00152 2\u0019 N + 1 X \u000b;k \u0014 sin k2\u0019l N + 1 sin k\u0019l N + 1 G1 \u0010 ky1; k \u0011 +sin k1\u0019l N + 1 sin k\u0019l N + 1 G1 \u0010 ky; k2 \u0011\u0015 (4.15) 60 4.3. Non-equilibrium Wick's theorem and 0 = i (\u000fk4 + \u000fk3 \u0000 \u000fk2 \u0000 \u000fk1)G2 \u0010 ky1; k y 2; k3; k4 \u0011 +\u00152 2\u0019 N + 1 X \u000b;k sin k1\u0019l N + 1 sin k\u0019l N + 1 G2 \u0010 ky; ky2; k3; k4 \u0011 +\u00152 2\u0019 N + 1 X \u000b;k sin k2\u0019l N + 1 sin k\u0019l N + 1 G2 \u0010 ky1; k y; k3; k4 \u0011 +\u00152 2\u0019 N + 1 X \u000b;k sin k3\u0019l N + 1 sin k\u0019l N + 1 G2 \u0010 ky1; k y 2; k; k4 \u0011 +\u00152 2\u0019 N + 1 X \u000b;k sin k4\u0019l N + 1 sin k\u0019l N + 1 G2 \u0010 ky1; k y 2; k3; k \u0011 +\u00152 2\u0019 N + 1 X sin k2\u0019l N + 1 sin k4\u0019l N + 1 G1 \u0010 ky1; k3 \u0011 (n (k2) + n (k4)) \u0000\u00152 2\u0019 N + 1 X sin k2\u0019l N + 1 sin k3\u0019l N + 1 G1 \u0010 ky1; k4 \u0011 (n (k2) + n (k3)) \u0000\u00152 2\u0019 N + 1 X sin k1\u0019l N + 1 sin k4\u0019l N + 1 G1 \u0010 ky2; k3 \u0011 (n (k1) + n (k4)) +\u00152 2\u0019 N + 1 X sin k1\u0019l N + 1 sin k3\u0019l N + 1 G1 \u0010 ky2; k4 \u0011 (n (k1) + n (k3)) (4.16) Each of these forms a closed linear system of equations and has a unique solution. Therefore, we only need to \fnd one solution and that must be the unique solution. We \frst apply Eq(4.9) to Eq(4.16) to expand G2 into products of G1. It is then easy to prove that the resulting equation is equiv- alent with Eq(4.15), meaning that a solution of Eq(4.15) is also a solution of Eq(4.16). For example, if we collect terms with G1 \u0010 ky2; k4 \u0011 together, we will have G1 \u0010 ky2; k4 \u0011n i (\u000f (k3)\u0000 \u000f (k1))G1 \u0010 ky1; k3 \u0011 \u0000\u00152 2\u0019 N + 1 X sin k1\u0019l N + 1 sin k3\u0019l N + 1 (n (k1) + n (k3)) +\u00152 2\u0019 N + 1 X \u000b;k \u0014 sin k3\u0019l N + 1 sin k\u0019l N + 1 G1 \u0010 ky1; k \u0011 +sin k1\u0019l N + 1 sin k\u0019l N + 1 G1 \u0010 ky; k3 \u0011\u0015\u001b ; (4.17) 61 4.4. Truncation of the hierarchy where the term in curly brackets is zero according to Eq(4.15). Therefore, the solutions of Eq(4.15) satisfy Eq(4.16) as well, as long as Eq(4.9), the non-equilibrium Wick Theorem, holds. Since the combined equation has a unique solution, the solution with Eq(4.9) valid is the only possible solution. Therefore, Eq(4.9) is proved. From this, we also know that for non-interacting systems, it is enough to solve only for the single-particle Green's functions, which form a linear system of equations with N2 unknowns. All higher order Green's functions can be factorized in products of various G1. 4.4 Truncation of the hierarchy As already shown, for interacting systems the equations for all Green's func- tions are coupled to one another. In order to solve Eq(4.7) explicitly, we \frst have to \fnd explicit forms of the operators m\u0302 in terms of the operatorsn cl; c y l o . In Appendix A, we present an exact numerical calculation and a perturbational calculation for these operators. Correspondingly, based on these two methods of \fnding the operators m\u0302, we discuss in this section two ways of turning Eq(4.7) into a closed equation by approximating the G2 terms. This is the lowest meaningful level of truncation. We will \frst discuss a more accurate method, which works for larger interactions V0, but which is computationally costly: perturbation based on two-particle Green's functions at equilibrium. Then we present a relatively less accurate but computationally much more e\u000ecient method, the non- equilibrium cluster expansion. The latter works only for relatively smaller V0 but can be applied to much larger systems. The values G1 \u0000 my; n \u0001 obtained from both methods will be compared against GEx1 \u0000 my; n \u0001 , the exact solution of Eq(4.2) via the linear system Eq. (2.89). Finally, we generalize the latter method by truncating it at the next level. This improves its accuracy and range of applicability, without sacri\fcing its e\u000eciency. 4.4.1 Method 1: using equilibrium Green's functions As explicitly worked out in Eq(A.2) of Appendix A, the operators m\u0302 can be written in terms of eigenmodes of HS , which can be found from an exact diagonalization of HS . This is a 2 N -dimensional eigenvalue problem. Then, in the language of a super-operator space [88], where operators are treated like vectors { so called super-vectors { m\u0302 can be expanded under the basis 62 4.4. Truncation of the hierarchy of all polynomials of n cl; c y l o as: m\u0302\u000b = X l d\u000b;lcl + V0D\u000b (4.18a) \u0016\u0302m\u000b = X l \u0016d\u000b;lc y l + V0 \u0016D\u000b; (4.18b) where using the de\fnition of inner product between super-vectors hhAjBii = tr \u0000 AyB \u0001 , we have d\u000b;l = 1 2(N\u00001) tr \u0010 cyl m\u0302 \u0011 ; (4.19a) \u0016d\u000b;l = 1 2(N\u00001) tr \u0000 cl \u0016\u0302m \u0001 : (4.19b) Here 2N\u00001 is a normalization constant to make d\u000b;l = 1 when m\u0302\u000b = cl. Also, d\u000b;l and \u0016d\u000b;l are the expansion coe\u000ecients at the linear order in cl and cyl ; operators V0D and V0 \u0016D are the remaining terms in the expansion of the operators m\u0302 and \u0016\u0302m, respectively. V0 is explicitly factorized out because of the fact that when V0 = 0, this remaining part vanishes. With these expressions for m\u0302, Eq(4.7) becomes, 0 = it D cym\u00001cn E + it D cym+1cn E \u0000 it D cymcn+1 E \u0000 it D cymcn\u00001 E +\u00152 X l; D \u000en \u0000 d\u000b;l + \u0016d \u0003 \u000b;l \u0001 cymcl + \u000em \u0000 \u0016d\u000b;l + d \u0003 \u000b;l \u0001 cyl cn E (4.20a) \u0000\u00152 X \u0002 \u000em\u000b \u0016d\u000b;n + \u000en\u000b \u0016d \u0003 \u000b;m \u0003 (4.20b) \u0000iV0 D cymc y n\u00001cncn\u00001 E + iV0 D cym+1c y mcm+1cn E \u0000iV0 D cyn+1c y mcn+1cn E + iV0 D cymc y m\u00001cncm\u00001 E (4.20c) \u0000\u00152V0 X D \u000em\u000bcn \u0016D\u000b + \u000en\u000b \u0016D y \u000bc y m \u0000 \u000en\u000bcymD\u000b \u0000 \u000em\u000bDy\u000bcn E : (4.20d) Note that every c0; c y 0; cN+1 and c y N+1 that appears in the equation should be set to 0. First let us replace all G2s in Eq(4.20c) by their values at equi- librium, denoted here as G E;(0) 2 where the superscripts indicate the use of thermal equilibrium (denoted by superscript E) as the zeroth order approx- imation (denoted by (0)) of the non-equilibrium G2. Using the \frst term as an example, G E;(0) 2 \u0010 my; n \u0011 = tr \u0010 cymc y n\u00001cncn\u00001\u001aeq (HS) \u0011 ; (4.21) 63 4.4. Truncation of the hierarchy where \u001aeq (HS) = 1 Z e \u0000 HS kBT . This requires knowing the eigenstates of HS . Similarly one can de\fne G E;(0) D from Eq(4.20d), using the \frst term as an example, G E;(0) D \u0010 my; n \u0011 = \u000em\u000btr \u0000 cn \u0016D\u000b\u001aeq (HS) \u0001 : (4.22) Next let us calculate G E;(1) 1 from Eq(4.20), where the superscript (1) means that the approximate calculation takes care of the \frst equation of the hi- erarchy, Eq(4.20). We organize all G E;(1) 1 \u0000 my; n \u0001 as a vector, g E;(1) 1 = h G1 \u0010 1y; 1 \u0011 ; G1 \u0010 1y; 2 \u0011 ; \u0001 \u0001 \u0001 ; G1 \u0010 N y; N \u0011iT ; (4.23) and then Eq(4.20) for given values of m;n is the equation occupying the (mN + n)th row and in total there areN2 such equations. After substituting G E;(0) 2 and G E;(0) D for the exact but unknown G2 and GD, the whole set of Eq(4.20) for all m;n becomes a linear system for g E;(1) 1 with dimension N 2, \u0000(1)g E;(1) 1 = iV0g E;(0) 2 + \u0015 2\u0017 + \u00152V0g E;(0) D ; (4.24) where the vector \u0017 comes from ordering Eq(4.20b) in the same way as g E;(1) 1 . The same holds for g E;(0) 2 and g E;(0) D correspondingly from ordering Eq(4.20c) and Eq(4.20d). The matrix \u0000(1) is extracted from Eq(4.20a). For example, assuming m and n are not at the boundaries, one may read from Eq(4.20), \u0017mN+n = X \u0002 \u000em\u000b \u0016d\u000b;n + \u000en\u000b \u0016d \u0003 \u000b;m \u0003 ; (4.25a) \u0000 (1) mN+n;(m\u00001)N+n = it: (4.25b) We can calculate single-particle equilibrium Green's functions, G E;(0) 1 , and organize them in the same way into a vector denoted as g E;(0) 1 . We de\fne the distance between two vectors A and B as, dAB = qP i jAi \u0000Bij2qP i jBij2 : (4.26) In order to gauge the accuracy, we compare dE;(0), the di\u000berence between the zeroth order g E;(0) 1 and the exact solution g Ex 1 , and d E;(1), the di\u000berence between the \frst order solution above, g E;(1) 1 and the exact solution g Ex 1 . Here GEx \u0000 my; n \u0001 = tr \u0010 cymcn\u001aex \u0011 , where \u001aex is the exact solution from Eq(4.2). 64 4.4. Truncation of the hierarchy Results First, we keep V0 = 0:2 as a constant, and check the accuracy of g E;(1) 1 for di\u000berent values of \u0001T . From Fig.4.1(a) we can see that the worst is d(1) \u0019 1%. Secondly, we set \u0001T = 0:4T as a constant, and check the accuracy of g E;(1) 1 for di\u000berent values of V0. The worst case is d (1) \u0019 0:3% as shown in Fig.4.1(b). Overall, the distance dE;(1) is always much smaller than dE;(0). We also compared the particle current calculated from this Green's functions, de\fned as: J = ie N \u0000 1 N\u00001X l=1 (G (l + 1; l)\u0000G (l; l + 1)) : (4.27) JE;(0), JE;(1), JEx are calculated respectively from g E;(0) 1 , g E;(1) 1 and g Ex 1 . From Fig.4.1(c) and (d) we see that in both cases, JE;(1) is very close to the exact value, JEx, while JE;(0), the current in the equilibrium state, is always zero. Very high accuracy is found especially for small \u0001T . This indicates that the approximation captures the essential part of the non- equilibrium stationary states. It is also worth mentioning that this method generates reasonable results for very large V0. Furthermore, it is likely that the approximation could be further improved, by expansions to higher order polynomials of cl; c y l and substitution of their values at equilibrium for the higher order unknown Green's functions in the higher-order equations of the hierarchy. Stopping the expansion of operators m\u0302 at linear order of V0 is compatible with solving only the \frst equation of the hierarchy. If further equations of the hierarchy are used then one should also expand operators m\u0302 to higher orders of V0. In order to estimate the accuracy of this level of approximation and also get a rough estimate of the accuracy of higher levels of truncation, we study the leading order of residues in terms of \u00152 and \u0001TT , both of which are assumed to be small in the following. Thus \u00152V0 \u001c V0, therefore we know that gD is smaller than the g2 term so we drop it. Similarly, since \u00152\u0001T \u001c \u0001T , we drop the \u00152\u0001T term in \u00152\u0017 in Eq(4.24), \u00152\u0017 = \u00152\u00170 (T ) + \u0015 2\u0001T\u0017;T ; (4.28) and keep only the large term, \u00152\u00170 (T ), which is independent of \u0001T . Here \u0017;T denotes formally a derivative of T on \u0017 | ddT \u0017. The general idea is then to write down equations for gEx1 and g E;(1) 1 (as shown in Eq. (C.2) of Appendix C), and then compare them to get an estimate for \u0001 E;(1) 1 = g E;(1) 1 \u0000 gEx1 . In 65 4.4. Truncation of the hierarchy 0 0.2 0.4 0.6 0.8 1 \u2206T\/T(V0=0.2) 0 0.02 0.04 0.06 0.08 d (a) (b) 0 0.5 1 1.5 2 V0(\u2206T\/T=0.4) 0 0.005 0.01 0.015 0.02 dE,(0) dE,(1) 0 0.2 0.4 0.6 0.8 \u2206T\/T(V0=0.2) 0 0.001 0.002 0.003 0.004 J JE,(0) JEx JE,(1) 0 0.5 1 1.5 2 V0(\u2206T\/T=0.4) 0 0.0005 0.001 0.0015 0.002(c) (d) Figure 4.1: g E;(1) 1 is compared with g Ex 1 for interacting systems at non- equilibrium. dE;(1) (dE;(0)) is the di\u000berence between g E;(1) 1 (g E;(0) 1 ) and g Ex 1 de\fned using Eq. (4.26). Both dE;(1) and dE;(0) are plotted vs. respectively \u0001T in (a) and V0 in (b). In both cases, d E;(1) is much smaller than dE;(0). In (c) and (d) JE;(0) and JE;(1), plotted vs. respectively \u0001T and V0, are compared against JEx. We see that JE;(0) is zero while JE;(1) is close to JEx even for relatively large V0. In all these sample calculations, t = 1:0, \u0015 = 0:1, \u0016 = \u00001:0. 66 4.4. Truncation of the hierarchy order to understand how an approximation to the next order improves the accuracy, we also want to compare \u0001 E;(1) 1 to \u0001 E;(0) 1 = g E;(0) 1 \u0000 gEx1 , which is estimated in the same way from the di\u000berence between the equations respectively for gEx1 and g E;(0) 1 , see Appendix C for further details. Here we summarize the results, namely \u0001 E;(0) 1 = \u0001T \u0010 \u0000 (1) 0 \u0011\u00001 \u0000 (1) ;T g Ex 1 + iV0 \u0010 \u0000 (1) 0 \u0011\u00001 \u0001 E;(0) 2 (4.29) and \u0001 E;(1) 1 = \u0010 \u0000 (1) 0 \u0011\u00001 \u0014\u0000V 20 \u0010\u0000(2)0 \u0011\u00001\u0001E;(0)3 +iV0\u0001T \u0010 \u0000 (2) 0 \u0011\u00001 \u0000 (2) ;T g Ex 2 + iV0\u0015 2 \u0010 \u0000 (2) 0 \u0011\u00001 \u0001 E;(0) 1 \u0015 : (4.30) Here \u0001 E;(0) n = g E;(0) n \u0000 gExn and \u0001E;(1)n = gE;(1)n \u0000 gExn for general n-particle Green's functions gn. We refer readers to Appendix C for de\fnitions of all \u0000 matrices. Most importantly here we see that (from the last term in Eq. (4.30)) \u0001 E;(0) 1 is multiplied by a small number \u0015 2V0 and then is included into \u0001 E;(1) 1 . Furthermore, this relation holds generally for higher-order levels of this approximation. Judging from this it follows that, as long as \u00152V0 \u001c t the method is very reasonable. As for the other two terms, they can be regarded as \u0000 V 20 g Ex 3 + V0g Ex 2 \u0001 \u0001T . Therefore, the maximum value of V0 where this method is still accurate is determined by \f\fgEx2 \f\f\u00001 or \f\fgEx3 \f\f\u0000 12 . Such limit could be much larger than 1 since roughly \f\fgExn \f\f = \f\fgEx\f\fn | smaller for larger n. This explains why this method is applicable even for V0 larger than t, as we see from Fig.4.1. Here, we note that since this approximation requires a full diagonaliza- tion of HS , it is necessarily not e\u000ecient for large systems. As a result, we now discuss another approach, that can be used for much larger systems. 4.4.2 Method 2: the non-equilibrium cluster expansion | \frst level A di\u000berent way to turn Eq(4.7) into a closed equation is to use the cluster expansion. Taking G2 as an example, this leads to: G2 \u0010 my; ny;m 0 ; n 0\u0011 = \u0000G1 \u0010 my;m 0\u0011 G1 \u0010 ny; n 0\u0011 +G1 \u0010 my; n 0\u0011 G1 \u0010 ny;m 0\u0011 +G2 \u0010 my; ny;m 0 ; n 0\u0011 : (4.31) 67 4.4. Truncation of the hierarchy If we set the part that describes correlations, G2 ! 0 (4.32) we have an approximative expression of G2 in terms of G1, which allows us to obtain a close set of equations for G1 (for more details, see below). This approach can be applied to higher-order Green's functions, for example using a similar expansion for G3 in terms of G1 and G2 and setting G3 = 0. In fact the non-equilibrium Wick's Theorem proved in Section x4.3 shows that indeed G2 = 0 when V0 = 0. This makes such expansions plausible for non-equilibrium Green's functions. Setting G2 = 0 with V0 6= 0 is similar to using the Hartree-Fock approximation. Depending on the system and the physical problem under investigation, one may need to go to the next level of approximation, i.e. keeping G2 but ignoring G3 to truncate the equation hierarchy at the level instead of the \frst one. Here, we will study both levels, beginning \frst with the lowest level approximation, i.e. setting G2 ! 0. However, the cluster expansion cannot be applied to the operators D de\fned in the previous section. Instead, we have to expand the operators m\u0302 in higher-order polynomials of n cl; c y l o . This can be done as follows. In order to avoid the costly exact diagonalization, the operators m\u0302 can also be found analytically using perturbation theory (see Appendix A for full details). The basic idea is to start by assuming cl (t) = c (0) l (t) + V0c (1) l (t) +O \u0000 V 20 \u0001 ; (4.33) and then derive and solve the equations of motion of c (0) l ; c (1) l from Heisen- berg's equation. In this way, one avoids the direct diagonalization of HS . This simpli\fes the calculation but its accuracy depends on the order of V0 at which the expansion stops. Stopping at the linear order of V0 is compat- ible with the cluster expansion for G2 (\frst level of approximation). If the cluster expansion in a higher-order Green's functions is applied, then the operators m\u0302 should also be expanded to higher orders of V0. Keeping only the \frst order, the operators m\u0302 become m\u0302\u000b = X m D\u000b;mcm + V0 X m1m2m3 D\u000b;m1m2m3cm1c y m2cm3 +O \u0000 V 20 \u0001 (4.34a) \u0016\u0302m\u000b = X m \u0016D\u000b;mc y m \u0000 V0 X m1m2m3 D\u000b;m1m2m3c y m3cm2c y m1 +O \u0000 V 20 \u0001 ; (4.34b) where the de\fnitions of D\u000b;m and D\u000b;m1m2m3 are given in Appendix A. 68 4.4. Truncation of the hierarchy With the \frst-order cluster expansion and the above expansion of oper- ators m\u0302 plugged into Eq(4.7), we have 0 = itG1 \u0010 (m\u0000 1)y; n \u0011 + itG1 \u0010 (m+ 1)y; n \u0011 \u0000itG1 \u0010 my; n+ 1 \u0011 \u0000 itG1 \u0010 my; n\u0000 1 \u0011 +\u00152 X l; h \u000en \u0000 D\u000b;l + \u0016D \u0003 \u000b;l \u0001 G1 \u0010 my; l \u0011 +\u000em \u0000 \u0016D\u000b;l +D \u0003 \u000b;l \u0001 G1 \u0010 ly; n \u0011i (4.35a) +\u00152V0 X \u000b;m1;m2 (D\u000b;nm2m1 \u0000D\u000b;m1m2n)G1 \u0010 my1;m2 \u0011 \u000em +\u00152V0 X \u000b;m1;m2 (D\u000b;mm2m1 \u0000D\u000b;m1m2m)G1 \u0010 my2;m1 \u0011 \u000en\u000b (4.35b) \u0000\u00152 X \u0000 \u000em\u000b \u0016D\u000b;n + \u000en\u000b \u0016D \u0003 \u000b;m \u0001 (4.35c) +\u00152V0 X \u000b;m1 (D\u000b;m1m1n\u000em\u000b +D\u000b;m1m1m\u000en\u000b) (4.35d) \u0000iV0G \u0010 my; (n\u0000 1)y; n; n\u0000 1 \u0011 +iV0G \u0010 (m+ 1)y;my; (m+ 1); n \u0011 \u0000iV0G \u0010 (n+ 1)y;my; n+ 1; n \u0011 +iV0G \u0010 my; (m\u0000 1)y; n;m\u0000 1 \u0011 : (4.35e) Here G2 should be interpreted according to Eq. (4.31) with G2 ! 0. This is the closed system of equations for the unknowns G1. Next we de\fne a vector g C;(1) 1 , where as before superscript C means cluster expansion and (1) symbolizes keeping only the \frst equation in the hierarchy. For simplicity we order Eq(4.35e) in the same way and denote it as g C;(1) 2 = \u0005 \u0010 g C;(1) 1 \u0011 , where \u0005 refers to the nonlinear function | summation of products | of g C;(1) 1 in Eq(4.35e). Then the above equation can be written in matrix form as: \u0010 \u0000 (1) 0 + \u0015 2V0\u0000 (1) D \u0011 g C;(1) 1 = \u0015 2\u00170 + \u0015 2V0\u00171 + iV0g C;(1) 2 ; (4.36) where the \fve terms are respectively the \fve sub equations in Eq(4.35), for 69 4.4. Truncation of the hierarchy example, (\u00170)mN+n = X \u0000 \u000em\u000b \u0016D\u000b;n + \u000en\u000b \u0016D \u0003 \u000b;m \u0001 : (4.37) This equation can be solved iteratively g [n+1] 1 = \u0010 \u0000 (1) 0 + \u0015 2V0\u0000 (1) D \u0011\u00001 \u0010 \u00152\u00170 + \u0015 2V0\u00171 + iV0\u0005 \u0010 g [n] 1 \u0011\u0011 ; (4.38) starting from the initial value g [0] 1 = \u0010 \u0000 (1) 0 \u0011\u00001 \u00152\u00170: (4.39) Here we start from g [0] 1 = g C;(0) 1 , which is the exact solution of Eq(4.36) when V0 = 0. Through the iteration de\fned above we get the solution g C;(1) 1 = limn!1 g [n] 1 . In practice we stop at large enough n such that g [n] 1 \u0000 g[n\u00001]1 is small enough. Results Using again Eq. (4.26), we de\fne dC;(0) as the relative distance between g C;(0) 1 and g Ex 1 , and d C;(1) as the relative distance between g C;(1) 1 and g Ex 1 . First, we keep V0 = 0:2 as a constant, and check the accuracy of g C;(1) 1 for di\u000berent values of \u0001T . From Fig.4.2(a) we see that the worst case is d(1) \u0019 1%. Secondly, we set \u0001T = 0:4T as a constant, and check the accuracy of g C;(1) 1 for di\u000berent values of V0. The worst case is d (1) \u0019 2% as shown in Fig.4.2(b). Overall, the di\u000berence dC;(1) is always much smaller than dC;(0). From Fig.4.2(c) we see that for a small V0, J C;(0) already provides most of the total current. However, Fig.4.2(d) shows that as V0 increases, the di\u000berence between JC;(1) and JC;(0) grows. We should also note that for larger V0, JC;(1) starts to deviate from JEx. This indicates that the approximation captures the essential part of the interaction but it is qunatitatively accurate only for small V0. Of course, this can be improved by going to the next level in the trucation, as shown in section 4.4.3. In order to estimate the accuracy of this approximation, let us assume that \u00152 and V0 are small. We de\fne \u0001 C;(0) n = g C;(0) n \u0000 gExn and \u0001C;(1)n = g C;(1) n \u0000 gExn . Again we start from the equations of the three: gC;(0)1 , gC;(1)1 and gEx1 (as we have done in Eq. (C.14) of Appendix C), and then compare them while ignoring certain higher-order, smaller terms such as those which 70 4.4. Truncation of the hierarchy 0 0.2 0.4 0.6 0.8 \u2206T\/T(V0=0.2) 0 0.01 0.02 0.03 0.04 d (a) (b) 0 0.2 0.4 0.6 0.8 1 V0(\u2206T\/T=0.4) 0 0.03 0.06 0.09 0.12 0.15 dC,(0) dC,(1) 0 0.2 0.4 0.6 0.8 \u2206T\/T(V0=0.2) 0 0.001 0.002 0.003 0.004 J JC,(0) JEx JC,(1) 0 0.2 0.4 0.6 0.8 1 V0(\u2206T\/T=0.4) 0.001 0.0012 0.0014 0.0016 0.0018 (c) (d) Figure 4.2: g C;(1) 1 is compared with g Ex 1 for non-equilibrium interacting systems. Both dC;(1) and dC;(0) are plotted vs. respectively \u0001T in (a) and V0 in (b). In both cases, d C;(1) is always much smaller than dC;(0). In (c) and (d) JC;(0) and JC;(1), plotted vs. respectively \u0001T and V0, are compared against JEx. From (c), where V0 = 0:2 and it is relatively small, we see that for a given value of V0, both J C;(0) and JC;(1) are consistent with JEx. From (d) we \fnd that for relatively larger V0, J C;(1) agrees much better with JEx than JC;(0). 71 4.4. Truncation of the hierarchy are proportional to \u00152V0, see Appendix C for more details. We arrive at: \u0001 C;(0) 1 = \u0000iV0 \u0010 \u0000 (1) 0 \u0011\u00001 gEx2 \u0018 V0 \f\fgEx1 \f\f2 ; (4.40) and \u0001 C;(1) 1 = \u0010 \u0000 (1) 0 \u0011\u00001 \u0014 iV 20 \u0010 \u0000 (2) 0 \u0011\u00001 gEx3 +\u00152V0 \u0010 \u0000 (2) 0 \u0011\u00001 \u0001 C;(0) 1 \u0015 \u0018 V 20 \f\fgEx1 \f\f3 + \u00152V 20 \f\fgEx1 \f\f2 : (4.41) We refer the readers to Appendix C for de\fnitions of all \u0000 matrices. This estimate agrees with the numerical results that \u0001 C;(0) 1 is propor- tional to V0 while \u0001 C;(1) 1 is proportional to V 2 0 . Most importantly, we see again that \u0001 C;(0) 1 is multiplied by a small number \u0015 2V0 and then be- comes a part of \u0001 C;(1) 1 . Since roughly \f\fgExn \f\f = \f\fgEx\f\fn, the other term, V 20 g Ex 3 \u0018 V 20 \f\fgEx1 \f\f3, is also much smaller than \u0001C;(0)1 \u0018 V0 \f\fgEx1 \f\f2. However, for large enough V0 the other approximation used in this method, the per- turbational expansion of the operators m\u0302, becomes invalid. Therefore, as long as V0 \u001c t this level of approximation is expected to be very reasonable. It should be noted that this method is capable of dealing with large systems since it does not require a direct diagonalization of a 2N -dimension matrix, like the previous method. Instead, it deals with vectors of dimension N2. 4.4.3 Method 2: second-level cluster expansion In the previous section, we showed that even truncation at \frst level leads to quite accurate results, and a very e\u000ecient method. Here, we discuss the truncation at the second level of the BBGKY hierarchy. This requires us to keep G2 as unknowns in the equations which now include the equations for G2, while letting G3 = 0 in order to truncate the resulting larger sys- tem. While the \frst-level approximation is equivalent to the Hartree-Fock approximation, this second-level form goes beyond that. The cluster expan- sion of G3 [94] expresses it in terms of G1, G2 and G3 (see Appendix B for the explicit expression): G \u0010 my1;m y 2;m y 3;m4;m5;m6 \u0011 = X P (\u00001)PG1G1G1 + X P (\u00001)PG1G2 +G3 \u0010 my1;m y 2;m y 3;m4;m5;m6 \u0011 : (4.42) 72 4.4. Truncation of the hierarchy Here P P (\u00001)PG1G1G1 is a short-hand notation for the various ways of pairing fmy1;my2;my3;m4;m5;m6g into three groups using G1 and taking care of the anti-commutation relations, for example G1 \u0010 my1;m6 \u0011 G1 \u0010 my2;m5 \u0011 G1 \u0010 my3;m4 \u0011 : (4.43) Similarly P P (\u00001)PG1G2 denotes all di\u000berent ways of pairing these into two groups using G1 and G2, for example G1 \u0010 my1;m6 \u0011 G \u0010 my2;m y 3;m4;m5 \u0011 : (4.44) Setting G3 ! 0, we aim to derive a close system of equations for G1 and G2 from Eq. (4.7) and Eq. (4.8). In order to be consistent, i.e. so that all terms lower than G3 should be included in both equations, bath operators m\u0302\u000b should be truncated to second order, i.e. at terms proportional to V 20 . However, including the V 2 0 terms greatly complicates both equations (see Appendix B for details). Therefore, here we use only the expression of operators m\u0302\u000b truncated at the \frst order of V0. In this case, Eq. (4.7) becomes an equation very similar to Eq. (4.35) but with a di\u000berent meaning for G2: G2 should be interpreted according to Eq. (4.31) with explicit G2 included. We denote this, using compressed notations similar to the ones in Eq. (4.24) and Eq. (4.36), as \u0000(1)g1 = \u0015 2\u00170 + \u0015 2V0\u00171 + iV0\u0019 (g1g1) + iV0g2: (4.45) Note that now we have an additional term iV0g2, so this is not a closed equation by itself. Here \u0019 (g1g1) is the short-hand notation for the combi- nations of products between g1 and g1 the cluster expansions in Eq. (4.31). Meanwhile, Eq. (4.8) becomes, \u0000(2)g2 = \u0000\u0000(2)\u0019 (g1g1) + \u00152g1 + \u00152V0g1 +iV0\u0019 (g1g1g1) + iV0\u0019 (g1g2) ; (4.46) such that we get a closed system of equations when we combine the two sets. The explicit form of this latter equation can be found in Appendix B. Here \u0019 (g1g2) is the short-hand notation for the combinations of products between g1 and g2 from Eq. (4.42). Same holds for \u0019 (g1g1g1). The matrices \u0000(1) and \u0000(2) come from the following terms \u0000(1) = \u0000 (1) H0 + \u00152\u0000 (1) B0 + \u00152V0\u0000 (1) B1 ; (4.47a) \u0000(2) = \u0000 (2) H0 + iV0\u0000 (2) V + \u0015 2\u0000 (2) B0 + \u00152V0\u0000 (2) B1 : (4.47b) 73 4.4. Truncation of the hierarchy 0 0.2 0.4 0.6 0.8 1 \u2206T\/T(V0=0.8) 0 0.05 0.1 d (a) (b) 0 0.2 0.4 0.6 0.8 1 V0(\u2206T\/T=0.4) 0 0.05 0.1 0.15 0.2 d1 (0) d1 (1) d1 (2) Figure 4.3: G1 calculated with di\u000berent levels of the approximations are compared. d (2) 1 is plotted in addition to the previous shown d (1) 1 and d (0) 1 . We see that the distance d (2) 1 is smaller than, but rather close to d (1) 1 , while both are much smaller than d (0) 1 . The various \u0000s are obtained by considering only the term indicated by their subscripts. For example, \u0000 (1) H0 is the one derived from H0 only, without using the potential V and bath operators m\u0302\u000b, while \u0015 2\u0000 (1) B0 (\u00152\u0000 (1) B1 ) is related to the component of the bath operators that is of the zeroth (\frst) order in V0. The combination of Eq. (4.45) and Eq. (4.46) is a closed system for g1 and g2. Solving it gives us results of the second-level cluster expansion, and provides \fnite values for G2. In other words, explicit correlations are built in, and thus this method goes well beyond the Hartree-Fock approximation. In the following we are investigate how large are these G2 and also what is the gain in accuracy when compared against exact numerical solutions, and against the \frst-level truncation. First, we de\fne d (2) 1 to be the di\u000berence between the G1 calculated from this second-order cluster expansion and the exact G1, similarly as d C;(1), which is now denoted as d (1) 1 . In Fig.4.3 we \fnd that d (2) 1 is smaller than but very close to d (1) 1 . This means that the second-order cluster expan- sion has better accuracy, but quantitatively does not lead to a signi\fcant improvement of G1. 74 4.4. Truncation of the hierarchy 0 0.2 0.4 0.6 0.8 1 \u2206T\/T(V0=0.8) 0 0.05 0.1 0.15 0.2 0.25 0.3 d (a) (b) 0 0.2 0.4 0.6 0.8 1 V0(\u2206T\/T=0.4) 0 0.1 0.2 0.3 0.4 d2 (0) d2 (1) d2 (2) Figure 4.4: G2 calculated with di\u000berent levels of the approximations are compared. d (2) 2 is plotted v.s. \u0001T and V0. We see that d (2) 2 is signi\fcantly smaller than d (1) 2 . Second, in the same way, we de\fne d (2) 2 and compare the di\u000berences be- tween G2s calculated from the various truncations to the exact G2. From Fig.4.4 we \fnd that the second order truncation leads to a sizable improve- ment in G2. To \fnd how signi\fcant is this di\u000berence for physical quantities, we com- pare the currents calculated from the approximations and exact calculation. In order to enhance these di\u000berences, we increase the coupling constant to \u0015 = 0:5 in the following \fgures. This is because the absolute value of the particle current depends very strongly on the value of \u0015 (J \/ \u00152). For the same purpose we have also set V = 0:8 instead of V = 0:4 when checking the \u0001T dependence. Figure 4.5 shows that the agreement between J (2) and JEx is much better than the one between J (1) and JEx, especially for larger V0. This proves that the correlations kept at the second-level truncation are important for an accurate description of the transport properties. Since the second-level approximation allows us to calculate G2 | the correlated part of G2 { we also consider the heat current, a physical quantity 75 4.4. Truncation of the hierarchy 0 0.2 0.4 0.6 0.8 1 \u2206T\/T(V0=0.8) 0 0.02 0.04 0.06 0.08 J JEx J(0) J(1) J(2) 0 0.2 0.4 0.6 0.8 1 V0(\u2206T\/T=0.4) 0.02 0.025 0.03 (a) (b) Figure 4.5: Particle currents, J (2); J (1); J (0), calculated respectively from the second-, \frst- and the zeroth-order truncations are compared against JEx, which is found from the exact diagonalization method. For all the range of V0 shown in this \fgure, J (2) is very close to JEx, while J (1) is accurate only for small V0. 76 4.4. Truncation of the hierarchy 0 0.2 0.4 0.6 0.8 1 \u2206T\/T(V0=0.8) 0 0.005 0.01 0.015 0.02 0.025 J Jh Ex Jh (0) Jh (1) Jh (2) 0 0.2 0.4 0.6 0.8 1 V0(\u2206T\/T=0.4) 0.004 0.005 0.006 0.007 0.008 0.009 0.01 (a) (b) Figure 4.6: Heat currents, J (2) h ; J (1) h ; J (0) h are compared against J Ex h . For all range of V0 shown in this \fgure, J (2) h is close to J Ex h . Jh depends strongly on G2 while particle current J in the previous \fgure relies on G1. making explicit use of G2: hJhi = \u00002 N \u0000 2 N\u00002X l=2 Im \u0002 t2G1 (l \u0000 1; l + 1) +V tG2 (l \u0000 1; l + 1; l; l + 1) + V tG2 (l \u0000 1; l + 1; l; l \u0000 1)] ; (4.48) which is derived from Jh = N\u00002X l=2 i \u0002 hBl\u00001; h B l \u0003 (4.49) and hBl = \u0000tcyl cl+1 \u0000 tcyl+1cl + V cyl clcyl+1cl+1: (4.50) Again, we \fnd from Fig.4.6 that the agreement between J (2) h and J Ex h is much better than the one between J (1) h and J Ex h , especially for larger V0. This indicates that the \fnite G2 are essential for getting proper values of heat currents. 77 4.5. Conclusion and discussion Keeping orders of V 20 in the bath operators m\u0302\u000b is consistent with trun- cation at G3 = 0 and may improve accuracy for the second-level truncation even further. However, including these terms greatly complicates the linear equation for G2, so that it has about 200 di\u000berent terms (see Appendix A for more details). An investigation of their e\u000bect on the overall results is left as future work. 4.5 Conclusion and discussion To conclude, a BBGKY-like equation hierarchy is derived from the Red\feld equation and two systematic approximations are suggested to truncate and solve the hierarchy. Using the \frst-level of the \frst method and both the \frst- and the second-levels of the second method, non-equilibrium station- ary states of interacting systems are calculated. It is found these are in good agreement with exact results, for small systems where the latter are available. We also estimated the accuracy of the two approximations. We \fnd that the \frst method works for strong interaction V0. However, because it requires an explicit diagonalization of the Hamiltonian, it is less computationally e\u000ecient. We imagine this method being used to study, for example, systems of a few strongly coupled quantum dots which have large charging energies. The second method is much more e\u000ecient and therefore it can be ap- plied to much larger systems, but its accuracy decreases with increasing interaction strength V0. However, truncation at a higher level is shown to signi\fcantly improve both its accuracy, and its range of validity. Even at the second level, good accuracy for physical quantities is seen up to rather high values of V0 and \u0001T , and the results have the correct trends even when they become quantitatively less accurate. It is likely that going to even higher levels of truncation further improves accuracy, although, of course, this also leads to signi\fcantly more involved computations. In future work, one can investigate the generalization to higher level trun- cations. Even more interesting, however, is the application of this method to physical problems. 78 Chapter 5 Solving the Red\feld equation using coherent-state representations 5.1 Introduction In quantum optics, a common technique to solve quantum master equations for density matrices is to use the coherent-state representation [70]. Quite of- ten a quantum master equation for a single photon (boson) mode is solved in the coherent-state representation, where the operator-form quantum mas- ter equation becomes a di\u000berential equation of c-numbers and a density operator becomes a \\distribution\" function. The problem of the Hilbert space's in\fnite dimension is bypassed. For example, a distribution function over N complex numbers is su\u000ecient to describe a N -site Bose-Hubbard model [96, 97]. The equations for these distribution functions can be solved analytically or through numerical simulations of certain equivalent stochas- tic di\u000berential equations such as Langevin equations (see Refs. [70, 98] and also Appendix E for the relation between generalized Fokker-Planck equa- tions and Langevin equations). A common coherent-state representation is the P -representation (see Ref [70] and also the Appendix D for details on various forms of coherent- state representations). In the P -representation coherent eigenvalues \u0018 and \u0018\u0003 are treated as conjugate complex numbers, thus one solves only the equa- tions for \u0018. Sometimes it is also useful to work with generalizations of the P -representation [70]. In fact, the generalized P -representation, where co- herent eigenvalues \u0018 and \u0018\u0003 are treated independently, has been used in sim- ulations of both thermal equilibrium and dynamical evolution of quantum systems (see for example [72] and references therein). It has been shown that this approach is capable of dealing with the pure dynamical evolution of an interacting quantum system with 150000 atoms and 106 momentum modes [75]. The work presented in this chapter can be regarded as an ex- 79 5.1. Introduction tension of this method from simulations of equilibrium and pure dynamical states to simulations of non-equilibrium stationary states. Thus, in this chapter we introduce another e\u000ecient method to solve the Red\feld equation: using the coherent-state representation, the Red\feld equation is mapped onto a generalized Fokker-Planck equation. For non- interacting systems, analytical results are obtained for the non-equilibrium stationary states. For interacting systems, we \fnd that the resulting equa- tions can be solved e\u000eciently via classical simulations. The existence of a general distribution function for systems in thermal equilibrium, which depends only on the Hamiltonian and the temperature, greatly simpli\fes the calculation of equilibrium properties. However, there is no known general distribution for non-equilibrium stationary states. For each problem, therefore, one has to calculate the distribution from a quan- tum master equation, such as the Red\feld equation, which is capable of taking into account the role of the baths, besides the central system. However, we have seen in the previous chapters that it is very hard to solve the Red\feld equation e\u000eciently. The fact that there is no analytical expression for non-equilibrium stationary states may simply be due to the technical di\u000eculty in solving the quantum master equation. A more e\u000e- cient way of solving this equation might lead to di\u000berent results, or maybe to a proof that there is no general expression describing non-equilibrium stationary states. In this chapter, using the coherent-state representation, we derive an analytical expression for non-equilibrium stationary states of non-interacting systems. The question of \fnding such non-equilibrium sta- tionary states for non-interacting systems has been numerically investigated via the time-dependent density matrix renormalization [99, 100], however for the local-operator Lindblad equation instead of the Red\feld equation. The focus of this chapter, however, is to study the e\u000eciency of this approach when applied to interacting systems. The chapter is organized as follows. In section 5.2, we introduce the model system, de\fne its Red\feld equation and then convert the operator- form Red\feld equation into a di\u000berential equation in the coherent-state rep- resentation. We \fnd the mapped equation to be in the form of a generalized Fokker-Planck equation. We then proceed to \fnd analytical solutions of the generalized Fokker-Planck equation for non-interacting systems in Section 5.3. In Section 5.4, we show that the same BBGKY-like hierarchy can be derived from the generalized Fokker-Planck equation as from the original Red\feld equation. This provides further support for the work presented in the previous chapter. In section 5.5, we solve the generalized Fokker- Planck equation via numerical simulations for interacting bosonic systems. 80 5.2. The model and its e\u000bective equation of motion The method is also applicable to other quantum master equations such as the local-operator Lindblad equation, as we demonstrate in section 5.6. In section 5.7, we conclude that this new method based on the coherent-state representation leads to analytical solutions for non-interacting systems and comment on its possible further development and applications. 5.2 The model and its e\u000bective equation of motion For concreteness, the system of interest in this chapter is a 1-D Bose- Hubbard chain [96, 97] coupled to a bath at each of its ends. The left (right) bath is maintained at temperature TL = T + \u0001T 2 (TR = T \u0000 \u0001T2 ). We are interested in the stationary state of the central system. The Hamiltonian of the central system (HS), the baths (HB\u000b) and the couplings (HSB) are, respectively: HS = H0 + VS = N\u00001X l=1 \u0010 ayl al+1 + a y l+1al \u0011 + U 2 NX l=1 nl (nl \u0000 1) ; (5.1a) HB = X k; !k; \u0010 byk;\u000bbk;\u000b + 1 \u0011 ; (5.1b) HSB = \u0015 X k;\u000b=L;R \u0010 V \u000bk a y \u000bbk;\u000b + h:c: \u0011 : (5.1c) Note that only the ends are coupled to baths (\u000b = L;R): aL = a1; aR = aN etc. The system of interest, the Bose-Hubbard chain, has been realized experimentally [97]. Neglecting the shift of the spectrum of the central system due to the coupling to the baths, the Red\feld equation reads, from Eq. (2.48): @\u001a(t) @t = \u0000i[HS ; \u001a(t)] \u0000\u00152 X \u000b=L;R nh ay\u000b; m\u0302\u000b\u001a(t) i + \u0002 a\u000b; \u0016\u0302m\u000b\u001a(t) \u0003 + h:c: o ; (5.2) 81 5.2. The model and its e\u000bective equation of motion where m\u0302L(m\u0302R) is related to a1(aN ) while \u0016\u0302mL( \u0016\u0302mR) is related to a y 1(a y N ), m\u0302\u000b = X k jV \u000bk j2 Z 1 0 d\u001ca\u000b (\u0000\u001c) e\u0000i!k;\u000b\u001c h1 + n (!k;\u000b)i; (5.3a) \u0016\u0302m\u000b = X k jV \u000bk j2 Z 1 0 d\u001cay\u000b (\u0000\u001c) ei!k;\u000b\u001c hn (!k;\u000b)i: (5.3b) Here hn (!k;\u000b)i = \u0012 e !k;\u000b\u0000\u0016 T\u000b + 1 \u0013\u00001 is the Bose-Einstein distribution for the bath modes. Note that later we will switch the summation over momentum k into an integral over energy \u000f, which involves a density of states D (\u000f). Generally both D (\u000f) and jV \u000bk j2 may depend on \u000f. However, for simplicity we set the density of state and also jV \u000bk j2 to constants and absorb them into the coupling constant \u0015. Next we map the Red\feld equation into a di\u000berential equation in the coherent-state representation. A similar approach has been used in earlier works mentioned in Ref. [72]. However, they are based on di\u000berent equa- tions of motion, such as the pure dynamical equation or the local-operator Lindblad equation [90] with bath temperature set at zero, while we use the Red\feld equation for \fnite bath temperatures. From a technical point of view, in the coherent-state representation it is easier to deal with the local- operator Lindblad equation than the Red\feld equation. However, we have shown in section 2.5 of chapter 2 that in order to get proper non-equilibrium stationary states, it is necessary to use the more complicated Red\feld equa- tion. As we demonstrated in Appendix A, for fermions the operators m\u0302 and \u0016\u0302m can be derived perturbationally to avoid direct diagonalization of HS . Here we apply the same procedure for bosons. Assuming that the Hubbard U is small, we may solve the Heisenberg equation order by order to get al (t) = X n Una (n) l (t) ; a y l (t) = X n Una y;(n) l (t) ; (5.4) and then carry out the integrals to get the operators m\u0302 and \u0016\u0302m. An important observation is that when U = 0 the problem is exactly solvable and al (t) (a y l (t)) is a linear combination of all am (a y m), al (t) = a (0) l (t) = 2 N + 1 X km sin \u0019kl N + 1 sin \u0019km N + 1 e\u0000i\u000fktam; (5.5) 82 5.2. The model and its e\u000bective equation of motion where \u000fk = 2 cos k\u0019 N+1 . This is because for \fnite-size systems with free ends, the Fourier transform is discrete and eikl is replaced by the sin \u0019klN+1 wave- function, see for example Eq. (A.4). For interacting systems, in principle it is possible to perform this pertur- bational calculation up to a large n, however here we will stop at the \frst order, a (1) l (t), a (1) l (t) = i Z t 0 d\u001c \u0010 a y;(0) l (\u001c) a (0) l (\u001c) a (0) l (\u001c) \u0011 : (5.6) After some tedious but straightforward algebra( see Ref. [101] and also Ap- pendix A for details), we \fnd, to the linear order in U , that: m\u0302\u000b = X m D\u000b;mam + U X m1m2m3 D\u000b;m1m2m3a y m1am2am3 +O \u0000 U2 \u0001 (5.7a) \u0016\u0302m\u000b = X m \u0016D\u000b;ma y m + U X m1m2m3 D\u000b;m1m2m3a y m3a y m2am1 +O \u0000 U2 \u0001 ; (5.7b) where D\u000b;m = \u0019 2 N + 1 X k sin \u0019kl N + 1 sin \u0019km N + 1 [1 + n (\u000fk; T\u000b; \u0016\u000b)] ; (5.8a) \u0016D\u000b;m = \u0019 2 N + 1 X k sin \u0019kl N + 1 sin \u0019km N + 1 n (\u000fk; T\u000b; \u0016\u000b) ; (5.8b) D\u000b;m1m2m3 = \u0019 \u0012 2 N + 1 \u00134 \u0001 X k;m;k1;k2;k2 n (\u000f (k2) + \u000f (k3)\u0000 \u000f (k1) ; T\u000b; \u0016\u000b)\u0000 n (\u000f (k) ; T\u000b; \u0016\u000b) \u000f (k2) + \u000f (k3)\u0000 \u000f (k1)\u0000 \u000f (k) \u0001 sin \u0019kl N + 1 sin \u0019km N + 1 sin \u0019k1m1 N + 1 sin \u0019k2m2 N + 1 \u0001 sin \u0019k3m3 N + 1 sin \u0019k1m N + 1 sin \u0019k2m N + 1 sin \u0019k3m N + 1 : (5.8c) Here n (\u000f; T; \u0016) is the average Bose-Einstein occupation number of a state with energy \u000f, temperature T and chemical potential \u0016. Next, as summarized in Appendix D, the creation and annihilation oper- ators ayl and al become di\u000berential operators in the coherent representation, for example: ayl \u001a$ \u0012 \u0018\u0003l \u0000 @ @\u0018l \u0013 P \u0010 ~\u0018 \u0011 : (5.9) 83 5.2. The model and its e\u000bective equation of motion Using the above approximation for the operators m\u0302\u000b and the transformation from creation\/annihilation operators to di\u000berential operators, Eq.(5.2) be- comes, in the coherent P -representation, a generalized Fokker-Planck equa- tion, @P @t = \u0010 LH0 + ULV + \u0015 2L (0) B + \u0015 2UL (1) B \u0011 P: (5.10) where LH0 = N\u00001X l=1 \u0014 i @ @\u0018l+1 \u0018l + i @ @\u0018l \u0018l+1 + c:c: \u0015 ; (5.11a) LV = NX l=1 \u0014 \u0000i @ 2 @\u00182l \u00182l 2 + i @ @\u0018l \u0000 \u0018\u0003l \u0018 2 l \u0001 + c:c: \u0015 ; (5.11b) L (0) B = X \u000b;m \u0014\u0000 D\u000b;m \u0000 \u0016D\u000b;m \u0001 @ @\u0018 \u0018m + \u0016D\u000b;m @2 @\u0018\u000b@\u0018\u0003m + c:c: \u0015 ; (5.11c) L (1) B = X \u000b;mi D\u000b;m1m2m3 \u0014 @2 @\u0018\u000b@\u0018\u0003m2 \u0018\u0003m1\u0018m3 + @2 @\u0018\u000b@\u0018\u0003m3 \u0018\u0003m1\u0018m2 \u0000 @2 @\u0018\u000b@\u0018m1 \u0018m2\u0018m3 + c:c: (5.11d) \u0000 @ 3 @\u0018\u0003\u000b@\u0018m2@\u0018m3 \u0018m1 + c:c: \u0015 : (5.11e) Here c:c: refers to the complex conjugate and it is performed formally, e.g. \u0000i @@\u0018\u0003l+1 \u0018 \u0003 l is the c:c: of i @ @\u0018l+1 \u0018l. The last term, Eq. (5.11e), of L (1) B is separated from the other terms intentionally for reasons that will become apparent soon. If only equilibrium states are of interest, then the \frst and second term can be discarded since equilibrium states commute with HS . However, for non-equilibrium stationary states those terms are important. Setting @@tP = 0, we get the stationary state P \u0010 ~\u0018;1 \u0011 , which in the following is denoted simply as P \u0010 ~\u0018 \u0011 . Next, we will calculate the stationary of Eq. (5.10). We \frst discuss non-interacting systems and then turn to interacting ones. 84 5.3. Exact solution for non-interacting systems 5.3 Exact solution for non-interacting systems For U = 0 only the terms LH0 and L (0) B contribute, and the stochastic equation is a Fokker-Plank equation [98], in terms of complex variables: @ @t P = X i;j \" \u0000 @ @\u0018i \u0000ij\u0018j + @2 @\u0018i@\u0018\u0003j Dij\u0003 + c:c: # P: (5.12) For our speci\fc case of a non-interacting system we have: \u0000lm = \u0000i (\u000el;m\u00001 + \u000el;m+1)\u0000 X \u000b;m \u00152\u0019\u000el\u000b\u000e\u000bm; (5.13) and Dlm\u0003 = 1 2 \u00152 X \u0000 \u0016D\u000b;l\u000em\u000b + \u0016D\u000b;m\u000el \u0001 : (5.14) Note that we have veri\fed that Dy = D . This is in fact a standard Fokker-Planck equation (FPE) for the Ornstein- Uhlenbeck process. Such a FPE can be solved analytically [70, 98]. Its stationary solution is found to be a Gaussian function: P \u0010 ~\u0018 \u0011 = 1 Z e\u0000 1 2 \u0018y\u001b\u00001\u0018; (5.15) where Z is a normalization constant and \u0000\u001b + \u001b\u0000y +D = 0; (5.16) which has the solution, \u001b = \u0000 X l;m 1 \u0014l + \u0014\u0003m h\u0017l jD j \u0017mi j\u0016lih\u0016mj : (5.17) Here \u0014; j\u0016i ; h\u0017j are respectively the eigenvalues, right eigenvectors and left eigenvectors of \u0000 and h\u0016j ; j\u0017i are the complex conjugate of the two cor- responding vectors. This expression is correct for both equilibrium and non-equilibrium states. The only di\u000berence is in the matrix D, in whether TL is the same as, or is di\u000berent from TR. The only needed numerical task is to \fnd an eigenvector decomposition of \u0000, an N -dimensional matrix. Once \u001b is known, all correlation functions can be calculated, in particular the single-particle Green's function is G \u0000 ly;m \u0001 = h\u0018\u0003l \u0018mi = 2\u001blm. 85 5.3. Exact solution for non-interacting systems Next we work out explicitly the distribution function for a two-site (N = 2) sample system to \frstly con\frm that it leads to the proper ther- mal equilibrium, and secondly to show the general procedure to \fnd such a distribution function. It should be emphasized that the dimension of the space grows linearly with system size, so we chose to work with N = 2 just because it is easier to get some intuitive picture, not because of algo- rithm complexity issues. In principle, for two-site systems we need to work with 4-dimensional matrices, but for non-interacting systems they reduce to 2-dimensional matrices. We have for this speci\fc case, \u0000 = \u0014 \u0000\u00152\u0019 \u0000i \u0000i \u0000\u00152\u0019 \u0015 ;D = \u0014 D11 D12 D12 D22 \u0015 ; (5.18) where D11 = \u0015 2\u0019 n(1; TL; \u0016L) + n(\u00001; TL; \u0016L) 2 ; D12 = \u0015 2\u0019 n(1; TL; \u0016L)\u0000 n(\u00001; TL; \u0016L) + n(1; TR; \u0016R)\u0000 n(\u00001; TR; \u0016R) 4 ; D22 = \u0015 2\u0019 n(1; TR; \u0016R) + n(\u00001; TR; \u0016R) 2 : (5.19) >From this we get the stationary distribution, \u001b = 1 4\u00152\u0019 \u0014 D11 +D22 2D12 2D12 D11 +D22 \u0015 + D11 \u0000D22 4 (\u00154\u00192 + 1) \u0014 \u00152\u0019 \u0000i i \u0000\u00152\u0019 \u0015 : (5.20) For equilibrium states TL = TR = T , \u0016L = \u0016R = \u0016 and therefore D11 = D22, so that P \u0010 ~\u0018 \u0011 \/ e\u0000 1n(1;T;\u0016) \u0018\u00031+\u0018\u00032p 2 \u00181+\u00182p 2 \u0000 1 n(\u00001;T;\u0016) \u0018\u00031\u0000\u0018\u00032p 2 \u00181\u0000\u00182p 2 ; (5.21) which is exactly the P -representation form of the Boltzmann distribution in terms of the eigenmodes a1\u0006a2p 2 with eigenenergies \u00061. Note that this is independent of the coupling constant \u0015 since the second term in \u001b vanishes and the \u00152 in the \frst term cancel out. This con\frms that the equilibrium state of the central system does not depend on the details of the coupling between the system and the baths. However, this is unique for equilibrium states. A general NESS does depend on \u0015. In fact, the non-equilibrium part D1;1\u0000D2;2 \/ \u00152. In the following calculation, a relatively large \u0015 is used in order to enhance this non-equilibrium part. 86 5.4. Interacting systems: BBGKY-like hierarchy derived from the GFPE Next we compare correlations from the non-equilibrium distribution with correlations calculated with the BBGKY-like method discussed in the pre- vious chapter [101]. Numerical exact solution for non-interacting systems have been reported in Refs. for the Red\feld equation. For non-interacting systems, the BBGKY method can be solved exactly. Here we use it as a reference to check the analytical solution for non-interacting systems. We denote the single-particle correlation functions calculated from the above analytical solution of generalized Fokker-Planck equation, respectively from the numerical solution of the BBGKY-like method as Gre;An1 and G re;B 1 . Similar notations are used for particle currents Jre;An and Jre;B. We mea- sure the di\u000berence between the two correlations, for example the analytical solution and the BBGKY solution, through, dre;AnB = rP l;m \f\f\fGre;An1 (ly;m)\u0000Gre;B1 (ly;m)\f\f\f2rP l;m \f\f\fGre;B1 (ly;m)\f\f\f2 : (5.22) In Fig.5.1, we compare this analytical solution and the BBGKY-like nu- merical results for the Red\feld equation. Since dre \u0019 0, the two methods are indeed equivalent. The inset shows that currents calculated from both meth- ods also agree. This validates the fact that the generalized Fokker-Planck equation approach produces accurate results at least for non-interacting sys- tems. In the same \fgure, we also plot dre;nuB and J re;nu, which are obtained from a numerical solution that will be discussed in detail in section 5.4. There is a small but \fnite distance between the two correlation functions, however this has no contribution to the current, which is found to be in good agreement with the exact solution. 5.4 Interacting systems: BBGKY-like hierarchy derived from the generalized Fokker-Planck equation For central systems with interactions, Eq.(5.10) is a generalized Fokker- Planck equation which has additional triple derivative terms added to the standard Fokker-Planck equation. It is still possible to solve directly this generalized Fokker-Planck equation via a set of equivalent stochastic di\u000ber- ence equations, which are generalizations of the Langevin equation [98, 102]. We will discuss this technique in the next section x5.5. Before doing that, 87 5.4. Interacting systems: BBGKY-like hierarchy derived from the GFPE 0 0.2 0.4 0.6 0.8 1 \u2206T\/T 0 0.1 0.2 d dre,AnB dre,NuB 0 0.4 0.8 \u2206T\/T 0 0.3 0.6 J Jre,B Jre,An JRe,Nu Figure 5.1: Analytical solution of the generalized Fokker-Planck equation is compared against the BBGKY numerical solution for non-interacting sys- tems. The red (circles) curve in the main plot shows the comparison between the above analytical solution and the BBGKY-like numerical solution of the Red\feld equation. Both Green's functions (dre;AnB , main panel) and particle currents (Jre;An and Jre;B, inset) exactly agree with each other. The blue (square) curve is a comparison between a numerical simulation, to be dis- cussed in x5.4, and the BBGKY solution. Here we \fnd that the numerical simulation agrees with analytical solution quite well. Other parameters are \u0015 = 0:5; T = 2:0; \u0016 = \u00002:0. 88 5.4. Interacting systems: BBGKY-like hierarchy derived from the GFPE we verify here that the same equation hierarchy as the one presented in the previous chapter can also be derived from this generalized Fokker-Planck equation. The derivation of the BBGKY equations for bosonic systems mir- rors that of the previous chapter, and can be found in Appendix F. We focus again only on the single-particle stationary state Green's func- tions G \u0000 ly;n \u0001 , de\fned as h\u0018\u0003l \u0018ni = R d~\u0018\u0018\u0003l \u0018nP \u0010 ~\u0018 \u0011 . Let us then derive a set of equations for them from the generalized Fokker-Planck equation, Eq. (5.10). For example, consider a \frst-order derivative term in Eq. (5.11b). After performing a partial integration, we have,X m Z d~\u0018\u0018\u0003l \u0018ni @ @\u0018m \u0010 \u0018\u0003m\u0018 2 mP \u0010 ~\u0018;1 \u0011\u0011 = \u0000i X m Z d~\u0018P \u0010 ~\u0018;1 \u0011 \u0018\u0003m\u0018 2 m @ @\u0018m (\u0018\u0003l \u0018n) = \u0000i Z d~\u0018P \u0010 ~\u0018;1 \u0011 \u0018\u0003l \u0018 \u0003 n\u0018 2 n = \u0000i \u0018\u0003l \u0018\u0003n\u00182n\u000b : (5.23) Therefore, setting @@th\u0018\u0003l \u0018ni to zero, we have 0 = \u0000ih\u0018\u0003l \u0018n+1i \u0000 ih\u0018\u0003l \u0018n\u00001i+ ih\u0018\u0003l+1\u0018ni+ ih\u0018\u0003l\u00001\u0018ni (5.24a) +\u00152 X \u000b;m \u0000 \u0016D\u000b;m \u0000D\u000b;m \u0001 [\u000en\u000bh\u0018\u0003l \u0018mi+ \u000el\u000bh\u0018\u0003m\u0018ni] (5.24b) +\u00152 X \u0000 \u0016D\u000b;n\u000el\u000b + \u0016D\u000b;l\u000en \u0001 (5.24c) +iUh\u0018\u0003l \u0018\u0003l \u0018l\u0018ni \u0000 iUh\u0018\u0003l \u0018\u0003n\u0018n\u0018ni (5.24d) +\u00152U X \u000b;mi D\u000b;m1m2m3 \u0002 \u000en \u0000h\u0018\u0003m1\u0018m2i\u000elm3 + h\u0018\u0003m1\u0018m3i\u000elm2\u0001 +\u000el \u0000h\u0018\u0003m2\u0018m1i\u000enm3 + h\u0018\u0003m3\u0018m1i\u000enm2\u0001\u0003 : (5.24e) This is exactly the same with Eq. (4.7). This con\frms that the BBGKY- like hierarchy method and the generalized Fokker-Planck equation in the coherent-state representation are equivalent, as expected. Next, we discuss how to solve numerically the generalized Fokker-Planck equation. Results from the BBGKY-like method will be used as a reference. 89 5.5. Approximate numerical solution for interacting systems 5.5 Approximate numerical solution for interacting systems The generalized Fokker-Planck equation, Eq.(5.10), has triple derivative terms. The Pawula theorem [98] states that a distribution function with negative values will arise from a generalized Fokker-Planck equation with a \fnite series of derivatives beyond the second order. This does not, how- ever, rule out the possibility that a generalized Fokker-Planck equation with an in\fnite series of derivatives may lead to a positively valued distribution function. Our Red\feld equation is, in principle, such a generalized Fokker- Planck equation with in\fnite derivatives, if all terms in the expansion of the m\u0302 operators are kept. A truncation at a certain \fnite order might provide a more accurate result than a truncation at the second order, but the resulting distribution function will have negative values in some small regions. Ref. [98] presented such an example. Of course, new techniques are required to solve such generalized Fokker-Planck equation. Besides the example from Ref. [98], Ref. [102] proposed a set of equivalent stochastic di\u000berence equations for a generalized Fokker-Planck equation with triple derivative terms. The authors showed that a better distribution function than the one obtained after truncating at second derivatives, is given by their technique, which is based on a speci\fc kind of stochastic di\u000berence equation [102]. However, this technique is still under development and discussion. One may totally avoid such triple derivative terms by using generalized Gaussian phase-space method [72] or applying the phase-space method to the whole composite system [103]. Especially with the latter approach, one can simulate the dynamics of the whole system instead of that of the central system only. Of course, in this case, the baths will be treated as large but \fnite systems. Another bene\ft of this approach is that one could also go beyond Markov approximation. On the other hand, one should also note the possibility that drop- ping such triple derivative terms may be appropriate in certain circum- stances [104]. For example, in Eq. (5.24) | the \frst-order BBGKY equa- tion derived above for two-variable (corresponding to single-particle) corre- lations, we see that these third-order terms are not in the equation explicitly. They appear only in equations of three-or-more-variable correlations, for ex- ample in the equation of four-variable correlations, which itself appears in the equation of two-variable correlation. Therefore such third order terms a\u000bect two-variable correlations, but indirectly via four-variable correlations. Following this line of logic we have assumed that odd-number-variable cor- 90 5.5. Approximate numerical solution for interacting systems relations vanish. Since the hierarchical cluster expansion method [101] generally relates higher order correlations to lower ones, if one truncates it at a certain order as we do here, one captures certain aspects of the interaction although not its whole e\u000bects. This is better than not taking the interaction in consideration at all. In this sense, neglecting the third order terms is equivalent to the \frst order of the cluster expansion, where the four-variable correlations are approximated by products of two-variable ones such that e\u000bects of the third order derivatives are neglected. To summarize, in the rest of this chapter we will neglect all third order derivatives in the generalized Fokker-Planck equation, Eq.(5.10). The e\u000bects of their inclusion will be the topic of further investigations. For our speci\fc two-site interacting system, a classical simulation of the generalized Fokker-Planck equation without the third order derivatives can be casted into the following form, @ @t P = X i;j \u0014 \u0000 @ @\u0018i \u0000 (2) ij \u0018j \u0000 @ @\u0011i \u0000 (2) i\u0003j\u0003\u0011j + @2 @\u0018i@\u0018j D (2) ij + @2 @\u0011i@\u0011j D (2) i\u0003j\u0003 (5.25) + @2 @\u0018i@\u0011j D (2) ij\u0003 + @2 @\u0011i@\u0018j D (2) i\u0003j \u0015 P; (5.26) where we have replaced \u0018\u0003i with \u0011i and treat them as independent variables. This treatment in fact means that we are now working in the positive-P representation [70, 73]. For non-interacting systems, the di\u000busion matrix D is always positive de\fnite so that we can work with the P representation. For interacting systems, however, D is not positive de\fnite therefore we have to use the more general representation. Here we use the the positive- P representation. We denote the coe\u000ecients as \u0000 (2) ij , \u0000 (2) i\u0003j\u0003 and D (2) ij , D (2) i\u0003j , D (2) ij\u0003 , D (2) i\u0003j\u0003 . In principle, there could also be terms such as \u0000 @@\u0018i\u0000 (2) ij\u0003\u0011j , but for this speci\fc problem such terms do not appear. The explicit de\fnitions of those coe\u000ecients are, \u0000 (2) ij = \u0000ij + iU\u0018i\u0011i\u000eij ;\u0000 (2) i\u0003j\u0003 = (\u0000ij) \u0003 \u0000 iU\u0018i\u0011i\u000eij ; (5.27) and D (2) ij = \u0000i U 2 \u00182i \u000eij \u0000 \u00152U 2 X \u000b;m1;m2 (D\u000b;j;m1;m2\u000e\u000b;i +D\u000b;i;m1;m2\u000e\u000b;j) \u0018m1\u0018m2 ; (5.28a) 91 5.5. Approximate numerical solution for interacting systems D (2) i\u0003j\u0003 = i U 2 \u00112i \u000eij \u0000 \u00152U 2 X \u000b;m1;m2 (D\u000b;j;m1;m2\u000e\u000b;i +D\u000b;i;m1;m2\u000e\u000b;j) \u0011m1\u0011m2 ; (5.28b) D (2) ij\u0003 = D (2) j\u0003i = Dij\u0003 + \u00152U 2 X \u000b;m1;m2 [(D\u000b;m1;j;m2 +D\u000b;m1;m2;j) \u000e\u000b;i +(D\u000b;m2;i;m1 +D\u000b;m2;m1;i) \u000e\u000b;j ] \u0011m1\u0018m2 : (5.28c) Eq. (5.26) can be solved through the following Langevin equation [98] in positive-P representation [70, 72], d\u0018i = X j \u0000 (2) ij \u0018jdt+ p 2dwi (t) ; (5.29a) d\u0011i = X j \u0000 (2) i\u0003j\u0003\u0011jdt+ p 2d \u0016wi (t) ; (5.29b) where dw (t) and d \u0016w (t) are stochastic processes satisfying, dwi (t) dwj \u0000 t0 \u0001 = D (2) ij \u000ett0dt; (5.30a) dwi (t) d \u0016wj \u0000 t0 \u0001 = D (2) ij\u0003 \u000ett0dt; (5.30b) d \u0016wi (t) dwj \u0000 t0 \u0001 = D (2) i\u0003j \u000ett0dt; (5.30c) d \u0016wi (t) d \u0016wj \u0000 t0 \u0001 = D (2) i\u0003j\u0003\u000ett0dt: (5.30d) We can de\fne a 2N -dimensional symmetric matrix D, D = 24 \u0010 D (2) ij \u0011 N\u0002N \u0010 D (2) ij\u0003 \u0011 N\u0002N\u0010 D (2) i\u0003j \u0011 N\u0002N \u0010 D (2) i\u0003j\u0003 \u0011 N\u0002N 35 : (5.31) Then these random variables can be simpli\fed further and organized as a linear transformations of 2N independent Wiener noises d~! [70], d~w = Bd~!; (5.32) where D = BBT : (5.33) This requires a diagonalization of the 2N\u00022N matrix D at every time step, which is the most time consuming part. However, it is still much simpler 92 5.5. Approximate numerical solution for interacting systems 0 0.2 0.4 0.6 0.8 1 \u2206T\/T 0 0.1 0.2 0.3 0.4 d dre,NuB 0 0.4 0.8 \u2206T\/T 0 0.3 0.6 J JB Jre,Nu Figure 5.2: The numerical simulation of the generalized Fokker-Planck equa- tion approach is compared against the BBGKY-like numerical solution of the Red\feld equation, for interacting systems. The calculated current Jre;Nu is very close to Jre;B while dre;NuB in the main panel is relatively large. Other parameters are set to be \u0015 = 0:5; U = 0:1; T = 2:0; \u0016 = \u00002:0. than solving directly the d2-dimension linear system for the central system with a d-dimensional Hilbert space, since d increases exponentially with N . We have seen in Fig.5.1 that the numerical simulation leads to accu- rate results for non-interacting systems, where analytical solutions are pos- sible. Next, we compare this numerical simulation against the BBGKY-like method for interacting systems, in Fig.5.2. The currents Jre;Nu and J re;B are in agreement with each other up to high temperature biases. However, dre;NuB | the di\u000berence between the two Green's functions { is relatively large. This indicates that a reasonable result can be obtained using this numerical simulation, but there is space its improvement. In the numerical simulation of the generalized Fokker-Planck equations we used the Euler method for the It\u0016o form of the stochastic di\u000berential equation. The Euler method is not always stable, especially for large U . In fact, more stable and e\u000ecient methods such as the semi-implicit method for the Stratonovich form have been developed to solve Fokker-Planck equa- 93 5.6. Solving the LOLE via the GFPE method tions [105]. Using such methods might improve stability and accuracy of the results. However, currently there is an additional technical di\u000eculty in implementing such methods for the Red\feld equation: matrices B are found numerically so it is hard to derive the corresponding Stratonovich form. The relatively large dre;NuB could also be dure to the truncation of the true generalized Fokker-Planck equation to the standard Fokker-Planck equation. This calls for new methods to solve the generalized Fokker-Planck equation. 5.6 Solving the local-operator Lindblad equation via the generalized Fokker-Planck equation method Instead of the Red\feld equation, some authors prefer to use the local- operator Lindblad equation in the study of transport [58, 59, 99]. In this section, we demonstrate how to \fnd NESSs from the local-operator Lindblad equation using the coherent-state representation. The Red\feld equation \fnds that the operators m\u0302\u000b for the sites coupled to baths typically involve more than just the operator at the coupling sites. The local-operator Lind- blad equation, on the other hand, chooses m\u0302\u000b \/ h1 + n (!k;\u000b)ia\u000b, which results in the much simpler form of the coupling to baths: @\u001a(t) @t = \u0000i[HS ; \u001a(t)] \u0000\u00152\u0019 n h1 + n (\u000f1; TL; \u0016L)i h ay1; a1\u001a(t) i + hn (\u000f1; TL; \u0016L)i h a1; a y 1\u001a(t) i + h:c: +h1 + n (\u000fN ; TR; \u0016R)i h ayN ; aN\u001a(t) i + hn (\u000fN ; TR; \u0016R)i h aN ; a y N\u001a(t) i + h:c: o : (5.34) In the case of this speci\fc two-site systems, we have chosen that N = 2 and \u000f1;2 = 0. Note that we get Eq. (5.34) from the Red\feld equation Eq. (5.2) by letting D\u000b;m = \u0019\u000e\u000bm [1 + n (\u000f\u000b; T\u000b; \u0016\u000b)] ; \u0016D\u000b;m = \u0019\u000e\u000bmn (\u000f\u000b; T\u000b; \u0016\u000b) ; D\u000b;m1m2m3 = 0: (5.35) These relations greatly simplify the resulting generalized Fokker-Planck equation. Following the same procedure as for the Red\feld equation, we 94 5.6. Solving the LOLE via the GFPE method derive a similar Langevin equation for Eq. (5.34), d\u0018n = \" \u0000i (\u0018n\u00001 + \u0018n+1)\u0000 iU\u00182n\u0011n \u0000 \u00152\u0019 X (\u000e\u000bn\u0018\u000b) # dt+B (n) 1\u0017 dw (n) \u0017 ; (5.36a) d\u0011n = \" i (\u0011n\u00001 + \u0011n+1) + iU\u0018n\u00112n \u0000 \u00152\u0019 X (\u000e\u000bn\u0011\u000b) # dt+B (n) 2\u0017 dw (n) \u0017 ; (5.36b) where the matrix B(n) = \u0010 B (n) \u0016\u0017 \u0011 2\u00022 satis\fes, B(n) \u0010 B(n) \u0011T = \u0014 \u0000iU\u00182n 2\u00152\u0019P\u000b n (\u000f\u000b; T\u000b; \u0016\u000b) \u000en 2\u00152\u0019 P \u000b n (\u000f\u000b; T\u000b; \u0016\u000b) \u000en\u000b iU\u0011 2 n \u0015 ; (5.37) and n dw (n) \u0017 o is a set of 2N independent Wiener noises. Note that similar equations have been used to describe pure isolated dynamics [75] or T = 0 systems driven by the local-operator Lindblad equation [73], for interact- ing bosonic systems. The 2N -dimensional matrices B are decoupled into N 2-dimensional matrices B(n), which can be diagonalized analytically. There- fore, it is technically easier to solve the local-operator Lindblad equation than the Red\feld equation and the Stratonovich form is readily supplied into the xmds software [105]. Results from this local-operator Lindblad equation have been plotted in Fig.5.3. For non-interacting systems, ana- lytical expressions can be derived, and one \fnds that the NESSs from the local-operator Lindblad equation follows the general Gaussian distribution in Eq. (5.15) and that the correlation matrix \u001blole becomes, \u001blole = n (0; TL; \u0016L) + n (0; TR; \u0016R) 4 \u0014 1 0 0 1 \u0015 + \u00152\u0019 [n (0; TL; \u0016L)\u0000 n (0; TR; \u0016R)] 4 (\u00154\u00192 + 1) \u0014 \u00152\u0019 i \u0000i \u0000\u00152\u0019 \u0015 : (5.38) In Fig.5.3, we have compared these analytical and numerical results against the BBGKY results. Results for both currents and Green's func- tions are found to be consistent for the two approaches. This con\frms that our generalized Fokker-Planck equation method works for both the Red\feld equation and the local-operator Lindblad equation. 95 5.7. Conclusions 0 0.2 0.4 0.6 0.8 1 \u2206T\/T 0 0.1 0.2 d dlole,AnB dlole,NuB 0 0.4 0.8 \u2206T\/T 0 0.2 0.4 J Jlole,B Jlole,An Jlole,Nu 0 0.2 0.4 0.6 0.8 1 \u2206T\/T 0 0.1 0.2 0.3 0.4 d dlole,NuB 0 0.4 0.8 \u2206T\/T 0 0.2 0.4 J Jlole,B Jlole,Nu Figure 5.3: Results from the generalized Fokker-Planck equation approach are compared against the BBGKY solution of the local-operator Lindblad equation for non-interacting(a) and interacting systems(b). Results for both calculated currents and the Green's functions are consistent. 5.7 Conclusions To summarize, in this section we map the Red\feld equation into a gener- alized Fokker-Planck equation, using the coherent-state representation. For non-interacting systems, the resulting generalized Fokker-Planck equation can be solved analytically and an expression of non-equilibrium stationary states is obtained. This avoids working with the quantum master equation in a space with an exponentially growing (fermions and spins) or in\fnite (bosons) dimension. For interacting systems, \frstly this approach is found to be consistent with the BBGKY-like method. Secondly, a preliminary classical simulation via the Langevin equation with 2N continuous complex numbers is discussed. It is found to work well for weakly interacting sys- tems, although more work is needed. The advantage of this method, if it can be made to be accurate at stronger interactions, is that the total number of variables (2N) grows linearly with the system size. The same method is applied to solve the local-operator Lindblad equa- tion. Here the method is even more e\u000ecient, since at every time step one needs to solve N 2-dimensional eigenvalue problems instead of a 2N - dimensional one, like for the Red\feld equation. Results are found to be consistent with those from the BBGKY-like method. Further development of the techniques for solving the generalized Fokker- Planck equation is required to make full use of the potential of this approach based on the coherent-state representation. 96 Chapter 6 Application: heat transport in spin chains 6.1 Introduction The physical question addressed in this chapter is the validity of the phe- nomenological Fourier's law of thermal transport. A system which obeys (disobeys) Fourier's law is said to have normal (anomalous) transport, so this question is equivalent with identifying the conditions for normal transport. Even for classical systems this issue is not completely settled (see Ref. [76] for a review), although it was found that normal transport is closely related to the onset of chaos. For quantum systems a similar relation has also been proposed [106], where the quantum chaos is characterized by features of level spacing statistics [37, 107{113]. A widely believed conjecture is that anomalous heat transport occurs in systems described by integrable Hamiltonians (see Ref. [34] for a review). In this context, integrability is usually taken to mean the existence of a Bethe ansatz solution [114]. The validity of this conjecture is still not clear for heat transport in quantum spin chains, despite considerable e\u000bort [19, 35, 36, 42{ 44, 56, 59, 60, 82, 83, 106, 115]. Most of this work [35, 42{44, 60, 82, 115] studied in\fnite and\/or periodic chains, and used the standard Kubo for- mula [6, 40] where \fnite\/zero Drude weight signals anomalous\/normal trans- port. For integrable systems, the standard Kubo formula always predicts anomalous heat transport. In fact, full integrability is not even necessary, all that is needed is commutation of the heat current operator with the total Hamiltonian [115]. However, in Chapter 3, we have shown that the validity of the standard Kubo formula is questionable in the study of ther- mal transport, and that one needs to consider explicitly a system connected to baths. We note that the \\integrability\" of the chain connected to baths may be lost even if the isolated chain is integrable, since the terms describing the coupling to the baths lead to a non-vanishing commutator between the heat current operator and the total Hamiltonian [116{118]. Both these facts 97 6.1. Introduction might invalidate the main argument for anomalous transport of integrable systems, based on the standard Kubo formula. For simplicity of terminol- ogy, here we call a system integrable if it is so when isolated, since the term is used in this sense in the above conjecture. Studies which explicitly consider the e\u000bects of the baths, while fewer, also give contradictory results. Using the local-operator Lindblad equation, Prosen et al. showed that an integrable gaped (Jz > Jxy) XXZ chain has normal spin conductivity while its energy transport is anomalous [59, 119], while using the same approach, Michel et al. claimed that when Jz > 1:6Jxy, spin chains have normal energy transport [56]. Particularly, a recent work in Ref. [119] shows that results from the local-operator Lindblad equation and from the standard Kubo formula are consistent. Such con\ricting results exist not only in theoretical studies, but also in experiments. Anomalous heat transport observed experimentally in systems described by integrable models, such as (Sr,Ca)14Cu24O41, Sr2CuO3 and CuGeO3, [79, 120, 121] seems to validate this conjecture, but Ref. [81] \fnds normal transport in Sr2CuO3 at high temperatures. Even if there was no con\rict between results based on the local-operator Lindblad equation [56, 59], as discussed in Chapter 2 this approach is less reliable than the Red\feld equation. For example, only the latter leads to the proper Boltzmann distribution if the baths are not biased. This is why here we investigate heat transport in \fnite spin chains coupled to ther- mal baths, using the Red\feld equation. There are already several other such studies [36, 83]. However, these are for spin systems that are either non-interacting (thus trivially integrable), or are non-integrable. (We call a spin chain \\non-interacting\" if it can be mapped, for example through a Jordan-Wigner transformation, to a Hamiltonian for non-interacting spinless fermions). Anomalous transport is found for the former and normal trans- port is found for the latter. Disordered spin systems have also been found to have normal transport [37]. There are no examples of clean, interacting but integrable systems investigated via the Red\feld equation. We study such systems here. We \fnd no direct relation between inte- grability and anomalous transport. Instead, based on our systematic inves- tigations, we propose a new conjecture: anomalous conductivity is observed in systems that can be mapped to non-interacting fermions. Those that map to interacting fermion models, whether integrable or not, show normal transport. This is in qualitative agreement with the demonstration by Sirker et al. that di\u000busion is universally present in interacting 1D systems [42]. It is important to note that, chronologically, this was the \frst project undertaken once the formalism of the Red\feld equation was adopted. This 98 6.2. The model and its Red\feld equation explains why the results shown below are derived using (ine\u000ecient) direct methods to solve the Red\feld equation, and therefore are restricted to rather small systems. This is also why our conjecture needs to be further examined, for larger systems. It was this computational ine\u000eciency that drove the subsequent research on more e\u000ecient methods, presented in the previous chapters. In future, we plan to use these e\u000ecient methods to extend this study to larger systems. The chapter is organized as follows. Section 6.2 de\fnes the system and its Red\feld equation. In Sections 6.3 we discuss the numerical methods and in Section 6.4 we de\fne the thermal current and local temperatures. We then proceed to present numerical results in Section 6.5, and then conclude. 6.2 The model and its Red\feld equation We consider an N -site chain of spins-12 described by the Hamiltonian: HS = N\u00001X i=1 \u0002 Jxs x i s x i+1 + Jys y i s y i+1 + Jzs z i s z i+1 \u0003\u0000 ~B \u0001 NX i=1 ~si (6.1) while the heat baths are collections of bosonic modes: HB = X k; !k;\u000bb y k;\u000bbk;\u000b (6.2) where \u000b = R=L indexes the right\/left-side baths. The system-baths coupling is taken as: V = \u0015 X k; V (\u000b) k s y \u0010 byk;\u000b + bk; \u0011 (6.3) where syL = s y 1 and s y R = s y N , i.e. the left (right) thermal bath is only coupled to the \frst (last) spin and can induce its spin-\ripping. This is because we choose ~B \u0001 ~ey = 0 while j ~Bj is \fnite, meaning that spins primarily lie in the x0z plane so that sy acts as a spin-\rip operator. From Eq. (2.29), the resulting Red\feld equation for \u001a (t) is: @\u001a(t) @t = \u0000i[HS ; \u001a(t)]\u0000 \u00152 X \u000b=L;R ([sy\u000b; m\u0302\u000b\u001a(t)] + h:c:) (6.4) where m\u0302\u000b = s y \u000b \u0001 \u0006\u000b. Here, (\u0001) refers to the element-wise product of two matrices, hnja \u0001 bjmi = hnjajmihnjbjmi. The bath matrices \u0006L;R are de\fned 99 6.3. Numerical methods in terms of the eigenstates of the system's Hamiltonian HS jni = Enjni as: \u0006\u000b = \u0019 X m;n jmihnj h \u0002( mn)n\u000b ( mn)D\u000b ( mn) jV (\u000b)kmn j2 +\u0002( nm) (1 + n\u000b ( nm))D\u000b ( nm) jV (\u000b)knm j2 i where mn = Em\u0000En = \u0000 nm and kmn is de\fned by !kmn;\u000b = mn, i.e. a bath mode resonant with this transition. Furthermore, \u0002(x) is the Heaviside function, n\u000b( ) = \u0002 e \u0000 1\u0003\u00001 is the Bose-Einstein equilibrium distribu- tion for the bosonic modes of energy at the bath temperature T\u000b = 1=\f\u000b, and D\u000b( ) is the bath's density of states. The product D\u000b ( mn) jV (\u000b)kmn j2 is the bath's spectral density function. For simplicity, we take it to be a constant independent of \u000b, m and n. 6.3 Numerical methods It is straightforward to use the Runge-Kutta method to integrate the Red- \feld equation, Eq. (6.4), starting from some initial states. The memory cost is proportional to 22N for an N -spin chain, but it takes a very long time for the integration to converge to the stationary state. This is not surprising since, in principle, the stationary state is reached only as t!1. As already discussed, another approach [58] is to solve directly for the stationary state from: L\u001a(1) = 0; (6.5) i.e. to \fnd the eigenstate for the zero eigenvalue of the 2N -dimensional matrix L. However, in this case, the memory cost is proportional to 42N , which is much worse than for the Runge-Kutta method. The method which has better memory e\u000eciency than this eigenvalue problem and also better time e\u000eciency than the Runge-Kutta method, is to convert the equation into a linear system and solve it via matrix-free methods like the Krylov space methods, which requires only matrix-vector multiplication but not explicitly the matrix. The eigenvalue problem can be rewritten as a linear system of equations after explicitly using the normal- ization condition tr (\u001a) = 1 such that \u0016L\u001a(1) = \u0017; (6.6) where \u0017 = [1; 0; \u0001 \u0001 \u0001 ]T and \u0016L is found from L by replacing the \frst row by tr(\u001a). Then we use for example the generalized minimal residual method 100 6.4. De\fnitions of the thermal current and local temperatures (GMRES) [122], which requires only the matrix-vector multiplication rule. This new way of solving the Red\feld equation has memory cost of \u0018 22N and time e\u000eciency of a linear system with dimension 22N . It is still a di- rect method so its e\u000eciency is not comparable with methods such as the BBGKY-like approach and the coherent-state representation method that we discussed previously, however, unlike them it gives an exact result. This is important until one can understand whether the further approximations made in the more e\u000ecient methods can, for example, destroy the integra- bility of the system. Also, because the commutation relations between spin operators are quite di\u000berent from those of fermions and bosons, the more e\u000ecient methods are not readily generalizable. Developing and applying them to this problem will be the topic of future work. 6.4 De\fnitions of the thermal current and local temperatures We rewrite HS = PN\u00001 i=1 hi;i+1 + PN i=1 hi, where hi;i+1 is the exchange be- tween nearest-neighbor spins and hi is the on-site coupling to the magnetic \feld. We can then de\fne a local site Hamiltonian h (S) i = 1 2 hi\u00001;i + hi + 1 2 hi;i+1 (6.7) with h0;1 = hN;N+1 = 0, and a local bond Hamiltonian h (B) i = 1 2 hi + hi;i+1 + 1 2 hi+1 (6.8) such that HS = NX i=1 h (S) i = N\u00001X i=1 h (B) i : (6.9) The local bond Hamiltonians can be used to derive the heat current operator from the continuity equation j\u0302i!i+1 \u0000 j\u0302i\u00001!i = rj\u0302 = \u0000@h (B) i @t = \u0000i h HS ; h (B) i i : (6.10) This results in j\u0302i!i+1 = i h h (B) i ; h (B) i+1 i (6.11) 101 6.4. De\fnitions of the thermal current and local temperatures for i = 1; \u0001 \u0001 \u0001 ; N \u0000 2. As expected, in the steady state we \fnd hj\u0302i!i+1i = tr \u0010 j\u0302i!i+1\u001a (1) \u0011 = J to be independent of i. Similarly, we de\fne the spin polarization hszi i and the spin current for XXZ model, Js = Jxy syi s x i+1 \u0000 sxi syi+1 : (6.12) Knowledge of the steady state heat current J , as such, is not enough to decide whether the transport is normal or not. Consider an analogy with charge transport in a metal connected to two biased leads. What shows whether the transport is anomalous or not is the pro\fle of the electric potential, not the value of the electric current. In anomalous transport (for clean, non-interacting metals) all the voltage drop occurs at the ends of the sample, near the contacts. Away from these contact regions, electrons move ballistically and the electric potential is constant, implying zero intrinsic resistance. For a dirty metal, scattering takes place everywhere inside the sample and the electric potential decreases monotonically in between the contact regions, i.e. the sample has \fnite intrinsic resistivity. In principle, the scaling of the current with the sample size, for a \fxed e\u000bective bias, also reveals the type of transport: for anomalous transport, the current is independent of the sample size once its length exceeds the sum of the two contact regions, while for normal transport it decreases like inverse length. The problem with this approach is that one needs to keep constant the e\u000bective bias, i.e. the di\u000berence between the applied bias and that in the contact regions. Furthermore, since we can only study relatively short chains, the results of such scaling may be questionable. It is therefore desirable to use the equivalent of the electric potential for heat transport and to calculate its pro\fle along the chain, in order to determine the type of transport. This, of course, is the \\local temperature\", which is a di\u000ecult quantity to de\fne properly. One consistency condition for any de\fnition is that if TL = TR = T , i.e. the system is in thermal equilibrium at T , then all local temperatures should equal T . We de\fne local site temperatures Ti which ful\fll this condition in the following way. Since we know all eigenstates of HS , it is straightforward to calculate its equilibrium density matrix at a given T , \u001aeq (T ) = 1 Z P n e \u0000\fEn jnihnj, where Z = P n e \u0000\fEn . Let then hh(S)i ieq (T ) = tr[\u001aeq (T )h(S)i ]. We de\fne Ti to be the solution of the equation: hh(S)i ieq (Ti) = tr[\u001a (1)h(S)i ]: (6.13) In other words, the steady-state value of the energy at that site equals the energy the site would have if the whole system was in equilibrium at Ti. 102 6.5. Results Of course, we can also use other \\local\" operators such as h (B) i to calculate a local bond temperature Ti+ 1 2 . We \fnd that when these de\fnitions are meaningful, the results are in very good agreement no matter what \\local\" operator is used. This type of de\fnition of Ti is meaningful only if a large magnetic \feld B is applied. For small B, the expectation values hh(S)i ieq (T ) are very weakly T -dependent, so that tiny numerical errors in the steady-state value can lead to huge variations in Ti. Addition of a large B is needed to obtain hh(S)i ieq (T ) which varies fast enough with T for values of interest, so that a meaningful Ti can be extracted. Since we could not \fnd a meaningful de\fnition for Ti when \f\f\f ~B\f\f\f ! 0, we cannot investigate such cases. Note, however, that most integrable models remain integrable under addition of an external \feld ~B = Be\u0302z. 6.5 Results In all our calculations, we take Bz = 1 and the exchange J \u0018 0:1. Temper- atures TL=R = T (1 \u0006 \u000e=2) should not be so large that the steady state is insensitive to the model, or so small that only the ground-state is activated. Reasonable choices lie between min(Jx; Jy; Jz) and NB, which are roughly the smallest, respectively the largest energy scales for the N -site spin chain. In Fig. 6.1 we show typical results for (a) local temperature pro\fles Ti; Ti+ 1 2 and (b) local spin polarization pro\fles Szi . We apply a large bias \u000e = (TL \u0000 TR)=T = 0:4 for clarity, but we \fnd similar results for smaller \u000e. For these values, it seems that the \\contact regions\" include only the end spins. The pro\fle of the central part of the chain is consistent with anomalous transport (\rat pro\fle) for the XX chain (Jx = Jy, Jz = 0) and shows normal transport (roughly linear pro\fle) in all the other cases. XY chains with Jx 6= Jy behave similarly with the XX chain. We \fnd similar results for ferromagnetic couplings. We also \fnd that the ratio between the e\u000bective temperature di\u000berence, T2 \u0000 TN\u00001, and the applied temperature di\u000berence TL \u0000 TR, is not a constant for di\u000berent system sizes N . Therefore, in our numerical calculation, it is not possible to keep the e\u000bective temperature di\u000berence as a constant by applying the same temperature di\u000berence on the ends while changing the system size N . This implies that the dependence of J v.s. N can not be used as an indicator of normal or anomalous transport. At most, it can provide a very rough qualitative picture. We have plotted several of the J vs. N in Fig.6.2. We see that J is independent of N for XX 103 6.5. Results 1 2 3 4 5 6 7 8 9 10 i 1.8 2 2.2 Ti XX XXZ(0.05) XXX XXZ(0.2) XYZ 1 2 3 4 5 6 7 8 9 10 i 0.1 0.12 0.14 Szi Figure 6.1: Plot of (a) local temperature pro\fle and (b) local spin polariza- tion pro\fle for chains with N = 10. In all cases TL = 2:4; TR = 1:6; \u0015 = 0:1; Jx = 0:1; Bz = 1:0. Other parameters are Jy = 0:1; Jz = 0:0(XX); Jy = 0:1; Jz = 0:05 (XXZ0.05); Jy = Jz = 0:1 (XXX); Jy = 0:1; Jz = 0:2 (XXZ0.2) and Jy = 0:2; Jz = 0:3 (XYZ). Only the XX chain shows \rat Ti and Szi pro\fles. All other models have almost linearly decreasing pro\fles of both temperature and Sz. 104 6.5. Results 4 5 6 7 8 9 10 N 0 0.0005 0.001 0.0015 0.002 J XX XYZ J=0.001*(N-1.9)-0.89 XXX J=0.0017*(N-0.23)-0.37 Figure 6.2: Heat currents are plotted as a function of system size for three models: XX, XY Z and XXX. J is independent of N for XX chains, and it decreases with N for the other two cases. Due to the uncertainty about the e\u000bective applied temperature di\u000berence, the \ftting curves should only be regarded as a guide for eyes. The same parameters as those in Fig.6.1 are used. chains while it decreases with increasing N for XY Z and XXX chains. Another way to examine size-dependence is to compare the temperature and Sz pro\fles for all available sizes. We normalize all the pro\fles for various models with di\u000berent sizes by harvesting only the data in the central regions and converting them into dimensionless values in [0; 1], for example, from (i; Ti) to ( i\u00002 N\u00003 ; Ti\u0000TN\u00001 T2\u0000TN\u00001 ); i = 2; 3; : : : ; N \u0000 1. In Fig.6.3, we plot such normalized temperature and Sz pro\fles. We see that curves for all values of N collapse onto one single straight line and furthermore, there is no di\u000berence between XY Z and XXX chains. Currently we can only study small systems, but from these normalized pro\fles, no qualitative di\u000berence has been found for di\u000berent sizes. All these are integrable models. The XX model is special because it can be mapped to non-interacting spinless fermions with the Jordan-Wigner transformation [84]. A \fnite Jz leads to nearest-neighbor interactions be- tween fermions. Eigenmodes for models with Jz 6= 0 can be found using Bethe's ansatz, but they cannot be mapped to non-interacting fermions. In order to investigate where this transition between anomalous and 105 6.5. Results 0 0.2 0.4 0.6 0.8 1 (i-1)\/(N-2) 0 0.5 1 (T i- T N -1 )\/( T 2 - T N -1 ) N=10,XXX9 8 7 6 5 4 10, XYZ 0 0.2 0.4 0.6 0.8 1 (i-1)\/(N-2) 0 0.5 1 (S z i - Sz 2)\/ (S z N -1 - Sz 2) Figure 6.3: Normalized temperature and Sz pro\fles are plotted. Curves for di\u000berent values of N collapse onto a single straight line and we see no di\u000berence between XY Z and XXX chains. normal transport occurs, we plot temperature pro\fles for systems with small Jz in Fig.6.4(a). We see that even for very small Jz the above observation still holds. For Jz = 0:01 there is a slight qualitative di\u000berence, namely that the contact re- gions seem larger than just the end spins, as can be seen from the normalized pro\fles. We think this may be due merely to limitations of the numerical accuracy. In Fig.6.4(b), we investigate what is the minimum value of Bz that we can use with con\fdence. As we pointed out before, our de\fnition of local temperature and also the idea of local a spin polarization Szi are only meaningful for su\u000eciently large Bz. We \fnd that roughly this implies Bz > 0:3. This is reasonable, since for this limiting value the energy scale due to coupling to the local Bz \feld (0:3=2) is comparable with the energy related to exchange (3\u0002 0:1=4). We found this generic behavior for a wide range of parameters. When \u0015 2 [0:03; 0:2], T 2 [0:3; 30:0] and \u000e \u0015 0:01, the spin chain has normal conductivity when Jz 2 [0:02; 0:5] and anomalous conductivity when Jz = 0. These results lead us to conjecture that it is the presence or absence of interactions, rather than integrability, that determines whether or not the heat transport is normal. We \fnd no di\u000berence between thermal transport 106 6.5. Results 2 4 6 8 10 i 1.8 1.9 2 2.1 2.2 B z =0.1 B z =0.3 0 0.4 0.8 0 0.5 1 2 4 6 8 10 i 1.8 1.9 2 2.1 2.2 T i J z =0.00 J z =0.01 J z =0.02 J z =0.03 J z =0.04 (a) (b) Figure 6.4: (a) Temperature pro\fles for small Jz values are plotted. Here Jx = Jy = 0:1; Bz = 1:0. The e\u000bective temperature drop becomes smaller for smaller Jz, but the slope is still \fnite if Jz \u0015 0:02. The inset plots normalized pro\fles, and shows no obvious di\u000berences for models with Jz \u0015 0:02. The case of Jz = 0:01 shows a \fnite slope but the normalized pro\fle indicates a slight di\u000berence: the contact regions seem to be more extended than in the other cases. (b) Temperature pro\fles of XXX chains with J = 0:1 and small values of Bz. A linear temperature drop is observed for Bz = 0:3, but a \rat pro\fle for Bz = 0:1. Other parameters are the same as in Fig.6.1. 107 6.5. Results and spin transport, unlike the results of Prosen et al. [59], based on the local-operator Lindblad equation. The conjecture may be tested further on various systems. One candidate is the Ising model in a transverse \feldBx. It maps to non-interacting spinless fermions [123] and if we add a Bz \feld, it becomes interacting. Another closely related model is a system of spinless fermions on a tight-binding chain, with the nearest-neighbor interaction: HS = \u000f NX l=1 cyl cl \u0000 t N\u00001X l=1 \u0010 cyl cl+1 + c y l+1cl \u0011 + V0 N\u00001X l=1 cyl+1cl+1c y l cl; (6.14) HB = X k;\u000b=L;R !k;\u000bb y k;\u000bbk;\u000b; (6.15) VSB = \u0015 X k; \u0010 cy\u000bbk;\u000b + c\u000bb y k; \u0011 : (6.16) The XXZ chains map exactly into this HS model, after using the Jordan- Wigner transformation. However, the coupling to the baths in this fermionic system is di\u000berent from that in the spin systems. For a spin system, the Jordan-Wigner transformation maps the \u001by operator into an operator much more complicated than the cy or c used in this fermionic model. Therefore, although the two models are closely related, they are not identical. Results on this model can be interpreted as another test of our conjecture or at least a check of whether the observations reported above are in\ruenced by the speci\fc model for the system-bath coupling. In Fig.6.5, we plot local temperature pro\fles along the Ising spin chains in panel (a) and for fermionic chains in panel (b). The results support our conjecture: interactions lead to normal transport. Note that \u000f is set to be much larger than t, to mirror the condition Bz \u001d J . When \u000f is comparable to t, both non-interacting and interacting systems show almost \rat temperature pro\fles. In summary, the \frst conclusion we draw from these results is that in- tegrability is not su\u000ecient to guarantee anomalous transport: several inte- grable models show normal heat transport, in agreement with other studies [42, 58, 59]. The second conclusion is that only models that map onto Hamil- tonians of non-interacting fermions exhibit anomalous heat transport. This is a reasonable su\u000ecient condition, since once inside the sample (past the contact regions) such fermions propagate ballistically. However, we cannot, at this stage, demonstrate that this is a necessary condition as well. We therefore can only conjecture that this is the criterion determining whether 108 6.5. Results 1 2 3 4 5 6 7 8 9 10 i 1.8 2 2.2 V0=0.0 V0=0.2 1 2 3 4 5 6 7 8 9 10 i 1.6 1.8 2 2.2 T i Ising x Ising xz (a) (b) Figure 6.5: Temperature pro\fles of (a) Ising spin chains and (b) the V -t model of fermionic chains. The Ising spin chains have anomalous transport when Bz = 0 (Isingx, circles) and normal transport when Bz = 1:0 (Isingxz, squares). The fermionic chains show a \rat temperature pro\fle when V0 = 0 (circles) but a linear temperature drop when V0 = 0:2 (squares). Other parameters are Jz = 0:1; Bx = 1:0; \u000f = 1:0; t = 0:1; TL = 2:4; TR = 1:6; \u0016 = \u00001:0; \u0015 = 0:1; N = 10. 109 6.5. Results the heat transport is anomalous. If we believe that our results are not ar- tifacts due to mainly the boundary e\u000bects of the limited system sizes, then this conjecture is the only consistent qualitative conclusion that we may draw from all above results. In this context, it is important to emphasize again the essential role played by the connection to the baths. In its absence, an isolated integrable model is described by Bethe ansatz type wavefunctions. Di\u000busion is impos- sible since the conservation of momentum and energy guarantees that, upon scattering, pairs of fermions either keep or interchange their momenta. For a system connected to baths, however, fermions are continuously exchanged with the baths, and the survival of a Bethe ansatz type of wavefunction becomes impossible. In fact, even the total momentum is no longer a good quantum number. We believe that this explains why normal transport in systems mapping to interacting fermions is plausible. Normal transport is also possible for non-interacting fermions, if they are subject to elastic scattering on disorder. This can be realized, for example, by adding to the XY model a random \feld Bz at various sites. In that case, as expected, normal conductance has been found in disordered systems [37]. On the other hand, anomalous transport can also occur in models which map to homogeneous interacting fermions if the bath temperatures are very low. Speci\fcally, consider the XXZ models. Because of the large Bz we use, the ground-state of the isolated chain is ferromagnetic with all spins up. The \frst manifold of low-energy eigenstates have one spin \ripped (single magnon states), followed by states with two spins \ripped (two magnon states), etc. The separation between these manifolds is roughly Bz, although because of the exchange terms each manifold has a fairly considerable spread in energies and usually overlaps partially with other manifolds. If both TL; TR \u001c Bz, only single-magnon states participate in the trans- port. We can then study numerically very long chains by assuming that the steady-state matrix elements \u001anm vanish for all other eigenstates (Sz;tot is a good quantum number for these models). In this case we \fnd anomalous transport for all models, whether integrable or not. This is reasonable, since the lone magnon (fermion) injected on the chain has nothing else to interact with, so it must propagate ballistically. We can repeat this restricted calculation by including the two-magnon, three-magnon, etc. manifolds in the computation. As expected, the results agree at low TL; TR, but di\u000berences appear for higher TL; TR, when these higher-energy manifolds become thermally activated. In such cases, the transport becomes normal for the models mapping to interacting fermions as soon as the probability to be in the two (or more) magnon sector be- 110 6.6. Conclusions comes \fnite. In other words, as soon as multiple excitations (fermions) are simultaneously on the chain, and inelastic scattering between them becomes possible. These results may explain the heat transport observed experi- mentally in compounds such as Sr2CuO3 [81], where at low temperature anomalous transport was found while at high temperature normal transport was reported. 6.6 Conclusions Based on an extensive study of quantum spin chains using the Red\feld equa- tion, we propose a new conjecture for what determines the appearance of anomalous heat transport at all temperatures in spin chains. Unlike pre- vious suggestions linking it to the integrability of the Hamiltonian or the existence of energy gaps, we propose that, for clean systems, the criterion is the mapping of the Hamiltonian onto a model of non-interacting fermions. Our results showing normal energy and spin transport in systems with \f- nite Jz contradict the major conclusions of Ref. [56] and Ref. [59], which, however, are based on the less trustworthy local-operator Lindblad equation. This conjecture should be checked for larger systems, where more reliable information on the relation between J and N can be extracted. The more e\u000ecient methods presented in the previous two chapters might allow us to undertake such a study. For discussion of the thermal conductance, which is related to higher order Green's functions | G2 and G3, not only G1 | extension of the methods to higher orders may be needed. 111 Chapter 7 Conclusions and discussions We began this project with a study of the thermal transport of spin chains in the framework of theories of open quantum systems. At that time, simi- lar questions had been investigated using either the standard Kubo formula, the non-equilibrium Green's functions method or the local-operator Lind- blad equation. However, the idea of using the Red\feld equation for these questions had just been proposed and it had been applied only to non- interacting systems or very small interacting but non-integrable systems. We \frst developed and applied the direct methods to the Red\feld equation for spin systems. From the preliminary results for short chains we found a new conjecture, linking anomalous transport to existence of mapping onto non-interacting fermions. This is the main result of Chapter 6. It then became necessary to study larger interacting systems to either con\frm or reject this observation. However, at the time there were no avail- able methods capable of solving the Red\feld equation for large interacting systems. Thus, we turned to investigate such methods. We \frst tried to make direct use of the standard Kubo formula. However, we found that the standard Kubo formula is not applicable to open systems explicitly cou- pled to baths and instead we had to derive a similar Kubo formula from the Red\feld equation using linear response theory. We call the result the open Kubo formula. This is the content of Chapter 3. Several forms of such OKFs were proposed and tested. The proper approach was shown to not su\u000ber from singularities, like the the standard Kubo formula, however it is not very e\u000ecient. A more e\u000ecient OKF was also identi\fed, however only for certain types of quantities. Neither approach is e\u000ecient enough to allow the study of systems of the same size as permitted by the non-equilibrium Green's functions method. Next, inspired by the non-equilibrium Green's function method[13{15] and also by Saito's work on non-interacting systems[36], we focused on Green's functions instead of the reduced density matrices and developed the BBGKY-like hierarchical method presented in Chapter 4. Further ap- proximations are needed to truncate this hierarchy, if there are interactions. We identi\fed two possible approaches. The \frst replaces, at a certain level of 112 Chapter 7. Conclusions and discussions the hierarchy, all higher-order Green's functions by their equilibrium values. This method is computationally less e\u000ecient but accurate up to large values of the interaction. The second method uses a cluster expansion, basically factorizing the higher-order Green's functions in terms of lower-order ones. This approach is applicable to much larger systems. While the simplest version (truncation at \frst level of the hierarchy) is accurate only for rather small interactions, we showed that systematic increase of the truncation level leads to both a systematic improvement of its accuracy, and a larger range of interactions where this holds. Also, truncation at second level gen- erates correlated two-particle Green's functions, beyond the Hartree-Fock approximation. In Chapter 5 we proposed another method: solving the Red\feld equation in the coherent-state representation. The resulting di\u000berential equation has the form of the generalized Fokker-Planck equation, so we call this method the generalized Fokker-Planck equation method. Similar methods have been applied to pure dynamical evolution or equilibrium states, where studies of large interacting systems with roughly 105000000-dimensional Hilbert spaces have been performed[72{74]. We extend the method to apply it to the Red- \feld equation and study also non-equilibrium stationary states. We have con\frmed that for non-interacting systems exact results can be found ana- lytically with this new approach, even for non-equilibrium stationary states. Other advantages are that, from a numerical point of view, in this approach interacting systems are handled similarly like non-interacting systems. Also, Green's functions of all orders can be calculated all at once. The challenge, still not fully settled, is to \fnd e\u000ecient and stable ways to solve these dif- ferential equations. We showed some preliminary results, based on classical simulations, that are promising yet require more work. The next step will naturally be to apply these methods to the studies of transport properties of various systems, this time for large systems. The question of validity of Fourier's Law is not yet answered de\fnitely. The second-order form of the BBGKY-like method seems to have a good poten- tial there. However, the study of spin systems requires di\u000berent kinds of BBGKY-like equations, and also to understand whether the cluster approx- imation interferes with the integrability of a model. Therefore, although the methods we proposed are in principle applicable to this problem, spe- ci\fc forms of hierarchical equations need to be re-derived. Future possible projects, thus, are to develop the BBGKY-like method for spin systems, or to use the current BBGKY-like method to discuss thermal transport of bosonic or fermionic systems, instead of spin systems. Alternatively, we can also use the generalized Fokker-Planck equation method. Currently the simulation 113 Chapter 7. Conclusions and discussions (a) gi\u03c3 z i \u03c3 z c LBL LBR Lc \u03bbL \u03bbR (b) Figure 7.1: (a) A sketch of exactly solvable decoherence model: a single spin (~\u001bc) is coupled to a bath of N spins f~\u001big. (b) A variation of (a) in that the central system is coupled to the baths only via the ends. is not stable enough and it works only for bosonic systems. Improving this, and also developing similar methods for fermionic systems could be another future project. There are also many small systems showing rich physics, such as sys- tems of a few coupled quantum dots or molecular devices coupled to reser- voirs. Application of the proposed methods to such systems should be very straightforward. The methods can also be applied to more general physical systems, not only to study transport but also other physical questions, such as deco- herence control and classical simulations of quantum computations. In all models studied in this thesis, the baths are assumed to be collections of eigenmodes j!\u0017 (k)i, connected to the central systems via \u0015V \u0017k (see for ex- ample Eq. (2.44)). We then assume V \u0017k to have certain properties. A more realistic and speci\fc setup is to connect the baths locally to the central system, as sketched in Fig7.1(b). In that case, V \u0017k is implicitly de\fned. The system sketched in Fig.7.1(a) is in fact a famous exactly solvable model of quantum decoherence [85, 124]: a single spin-12 is connected to a 114 Chapter 7. Conclusions and discussions bath of N spins-12 via Ising coupling with strength gi to bath i [125]. The dynamics of the whole composite system is exactly solvable and Zurek et al. showed that the o\u000b-diagonal elements (in \u001bz basis), thus the coherent part, die out very quickly. A related but rather ambitious question is to study a counterpart with a thermal bath and to ask the following questions: whether or not the reduced density matrix of the central system converges to the expected thermal equilibrium state; whether or not the o\u000b-diagonal elements disappear and if so, which \\diagonal\" and \\o\u000b-diagonal\" parts are relevant, and how that depends on the speci\fc forms of interaction? In fact, such a natural selection of the basis is called, in theories of quantum measurement, eiselection of the pointer states [85, 124]. Furthermore, how do the answers to the above questions depend on various parameters of the composite systems, such as size of the bath, coupling strength between the central system and the bath, and temperature of the bath and so on. We can ask all the questions for a similar setup but now with two baths, biased or not, connected to the systems locally, as sketched in Fig.7.1(b). These questions can be investigated using di\u000berent approaches: starting from the pure dynamics of the whole composite system, which usually how- ever is only applicable to non-interacting systems (both the central system and the baths); through the Red\feld equation, which assumes the Markovian approximation, weak coupling and idealized in\fnite-size baths; and \fnally by the non-equilibrium Green's function method, which is typically applied for con\fgurations similar to the variation in Fig.7.1(b). In fact, there is another approach | the exact master equation or the in\ruence functional method, which relaxes the assumption of the Markovian approximation and the weak coupling limit, but still takes the baths to be in\fnitely large. How- ever, it currently works only for non-interacting systems since it is hard to solve such an exact master equation for interacting systems. So let us focus on the above three methods. The generalized Fokker-Planck equation method makes it possible to im- plement the \frst approach even for large interacting systems. The second approach can be realized through either the BBGKY-like method or the generalized Fokker-Planck equation method. Implementations of the third method are already well developed. Note, however, that if we set the tem- peratures of the baths to be equal then the non-equilibrium Green's function arrives at the corresponding equilibrium state of the whole composite sys- tem, which does not guarantee that the reduced density matrix of the central system follows the proper equilibrium distribution. On the other hand, the reduced density matrix generally converges to an equilibrium distribution in the theories of open quantum systems approach. Therefore, it will not 115 Chapter 7. Conclusions and discussions be much of a surprise if the non-equilibrium Green's function leads to qual- itatively di\u000berent results. If we want to compare them, a proper model, on which the non-equilibrium Green's function method and the Red\feld equa- tion ideally should produce the same result, is required. Comparison be- tween the \frst two might shed some light on di\u000berences between the Marko- vian and the non-Markovian approximation, the validity of the in\fnite-size bath assumption, etc. Comparison between the latter two might help us understand them better or even improve both. Lastly, our e\u000ecient methods for the Red\feld equation can be applied as well to solving the local-operator Lindblad equation. This should make the local-operator Lindblad equation a more powerful tool in those cases where it is applicable. As we mentioned earlier, there are still many groups that use the local-operator Lindblad equation instead of the Red\feld equation. It should be very straightforward to implement the methods for the local- operator Lindblad equation. 116 Bibliography [1] G. Mahler and V.A. Weberru\u0019. Quantum Networks: Dynamics of Open Nanastructures. Springer-Verlag, 1998. [2] H.-P. Breuer and F. Petruccione. The theory of open quantum systems. Oxford, 2002. [3] U. Weiss. Quantum Dissipative Systems. World Scienti\fc, 1999. [4] S. Datta. Quantum Transport: Atom to Transistor. Cambridge Uni- versity Press, 2005. [5] D.K. Ferry and S. M. Goodnick. Transport in Nanostructures. Cam- bridge Univ. Press, 1997. [6] R. Kubo. Statistical-mechanical theory of irreversible processes 1. general theory and simple applications to magnetic and conduction problems. Journal of the Physical Society of Japan, 12(6):570{586, 1957. [7] R. Kubo, M. Yokota, and S. Nakajima. Statistical-mechanical theory of irreversible processes 2. response to thermal disturbance. Journal of the Physical Society of Japan, 12(11):1203{1211, 1957. [8] R. Landauer. Electrical resistance of disordered one-dimensional lat- tices. Philosophical Magazine, 21(172):863{867, 1970. [9] R. Landauer. Spatial variation of currents and \felds due to local- ized scatterers in metallic conduction. IBM Journal of Research and Development, 1(3):223{231, 1957. [10] M. Buttiker. 4-terminal phase-coherent conductance. Physical Review Letters, 57(14):1761{1764, 1986. [11] M. Buttiker, Y. Imry, R. Landauer, and S. Pinhas. Generalized many- channel conductance formula with application to small rings. Physical Review B, 31(10):6207{6215, 1985. 117 Bibliography [12] H. Haug and A. P. Jauho. Quantum Kinetics in Transport and Optics of Semiconductors. Springer-Verlag, 1996. [13] M. Koentopp, C. Chang, K. Burke, and R. Car. Density functional calculations of nanoscale conductance. Journal of Physics-Condensed Matter, 20(8):083203, 2008. [14] C. Caroli, R. Combesco, P. Norzieres, and D. Saintjam. Direct calcu- lation of tunneling current. Journal of Physics C-Solid State Physics, 4(8):916{929, 1971. [15] R. Combesco. Direct calculation of tunneling current 3: E\u000bect of localized impurity states in barrier. Journal of Physics C-Solid State Physics, 4(16):2611{2622, 1971. [16] J. Taylor, H. Guo, and J. Wang. Ab initio modeling of quantum transport properties of molecular electronic devices. Physical Review B, 63(24):245407, 2001. [17] A. G. Red\feld. Relaxation theory: density matrix formulation. In D. M. Grant and R. K. Harris, editors, Encyclopedia Nuclear Magnetic Resonance, pages 4085{4092. Wiley, 1996. [18] G. D. Mahan. Many-Particle Physics. Plenum Press, 1990. [19] A. V. Sologubenko and H. R. Ott. Energy transport in one-dimensional spin systems. In D. Baeriswyl and L. Degiorgi, editors, Strong Interac- tions in Low Dimensions, chapter 12, pages 383{417. Springer, 2005. [20] A. Yacoby. Transport in quatnum wire. In D. Baeriswyl and L. De- giorgi, editors, Strong Interactions in Low Dimensions, chapter 10, pages 321{346. Springer, 2005. [21] A. Nitzan and M. A. Ratner. Electron transport in molecular wire junctions. SCIENCE, 300(5624):1384{1389, 2003. [22] A Salomon, D Cahen, S Lindsay, J Tomfohr, VB Engelkes, and CD Frisbie. Comparison of electronic transport measurements on or- ganic molecules. Advanced Materials, 15(22):1881{1890, 2003. [23] N. J. Tao. Electron transport in molecular junctions. Nature Nan- otechnology, 1(3):173{181, 2006. [24] W. Goepel and C. Ziegler. Nanostructures Based on Molecular Mate- rials. Wiley-VCH, 1992. 118 Bibliography [25] L. P. Kouwenhoven, C. M. Marcus, P. L. Mceuen, S. Tarucha, R. M. Westervelt, and N. S. Wingreen. Electron transport in quantum dots. In L. L. Sohn, L. P. Kouwenhoven, and G. Schon, editors, Mesoscopic Electron Transport, volume 345 of NATO ADVANCED SCIENCE IN- STITUTES SERIES, SERIES E, APPLIED SCIENCES, pages 105{ 214, 1997. [26] D. Natelson. Handbook of Organic Electronics and Photonics. Amer- ican Scienti\fc Publishers, 2006. [27] A. V. Sologubenko, T. Lorenz, H. R. Ott, and A. Freimuth. Thermal conductivity via magnetic excitations in spin-chain materials. Journal of Low Temperature Physics, 147(3-4):387{403, 2007. [28] A. V. Sologubenko, K. Berggold, T. Lorenz, A. Rosch, E. Shimshoni, M. D. Phillips, and M. M. Turnbull. Magnetothermal transport in the spin-12 chains of copper pyrazine dinitrate. Phys. Rev. Lett., 98(10):107201, Mar 2007. [29] A. V. Sologubenko, T. Lorenz, J. A. Mydosh, A. Rosch, K. C. Short- sleeves, and M. M. Turnbull. Field-dependent thermal transport in the haldane chain compound nenp. Phys. Rev. Lett., 100(13):137202, Apr 2008. [30] C. W. Chang, D. Okawa, H. Garcia, A. Majumdar, and A. Zettl. Breakdown of fourier's law in nanotube thermal conductors. Phys. Rev. Lett., 101(7):075903, Aug 2008. [31] B. Soree, W. Magnus, and W. Schoenmaker. Energy and momen- tum balance equations: An approach to quantum transport in closed circuits. Physical Review B, 66(3):035318, 2002. [32] T. H. Stoof and Y. V. Nazarov. Time-dependent resonant tunneling via two discrete states. Physical Review B, 53(3):1050{1053, 1996. [33] O. N. Jouravlev and Y. V. Nazarov. Electron transport in a double quantum dot governed by a nuclear magnetic \feld. Physical Review Letters, 96(17):176804, 2006. [34] X. Zotos and P. Prelovsek. Transport in one dimensional quanutm systems. In D. Baeriswyl and L. Degiorgi, editors, Strong Interactions in Low Dimensions, chapter 11, pages 347{382. Springer, 2005. 119 Bibliography [35] X. Zotos, F. Naef, and P. Prelovsek. Transport and conservation laws. Physical Review B, 55(17):11029{11032, 1997. [36] K. Saito, S. Takesue, and S. Miyashita. Energy transport in the in- tegrable system in contact with various types of phonon reservoirs. Physical Review E, 61(3):2397{2409, 2000. [37] Y. Yan, C.-Q. Wu, G. Casati, T. Prosen, and B. Li. Nonballistic heat conduction in an integrable random-exchange ising chain studied with quantum master equations. Physical Review B, 77(17):172411, 2008. [38] J. Wu and M. Berciu. Heat transport in quantum spin chains: the relevance of integrability. arXiv:1003.1559, submitted to PRB, 2010. [39] H. Mori. Statistical-mechanical theory of transport in \ruids. Physical Review, 112(6):1829{1842, 1958. [40] J. M. Luttinger. Theory of thermal transport coe\u000ecients. Physical Review, 135(6):A1505{A1514, 1964. [41] D. S. Fisher and P. A. Lee. Relation between conductivity and trans- mission matrix. Physical Review B, 23(12):6851{6854, 1981. [42] J. Sirker, R. G. Pereira, and I. A\u000feck. Di\u000busion and ballistic trans- port in one-dimensional quantum systems. Physical Review Letters, 103(21):216602, 2009. [43] F. Heidrich-Meisner, A. Honecker, and W. Brenig. Thermal transport of the xxz chain in a magnetic \feld. Physical Review B, 71(18):184415, 2005. [44] M. Rigol and B. S. Shastry. Drude weight in systems with open bound- ary conditions. Physical Review B, 77(16):161101, 2008. [45] C. Toher, A. Filippetti, S. Sanvito, and K. Burke. Self-interaction er- rors in density-functional calculations of electronic transport. Physical Review letters, 95(14), 2005. [46] F Evers, F Weigend, and M Koentopp. Conductance of molecular wires and transport calculations based on density-functional theory. Physical Review B, 69(23), 2004. [47] R. P. Feynman and F. L. Vernon. the theory of a general quantum system interacting with a linear dissipative system. Annals of Physics, 24(1):118{173, 1963. 120 Bibliography [48] A. J. Leggett, S. Chakravarty, A. T. Dorsey, M. P. A. Fisher, A. Garg, and W. Zwerger. Dynamics of the dissipative 2-state system. Reviews of Modern Physics, 59(1):1{85, 1987. [49] B. L. Hu, J. P. Paz, and Y. H. Zhang. Quantum brownian-motion in a general environment - exact master equation with nonlocal dissipation and colored noise. Physical Review D, 45(8):2843{2861, 1992. [50] C.-H. Chou, T. Yu, and B. L. Hu. Exact master equation and quantum decoherence of two coupled harmonic oscillators in a general environ- ment. Physical Review E, 77(1):011112, 2008. [51] R. Kubo, M. Toda, and N. Hashitsume. Statistical physics II: nonequi- librium statistical mechanics. Springer, 2nd edition, 1991. [52] A. G. Red\feld. On the theory of relaxation processes. IBM Journal of Research and Development, 1(1):19{31, 1957. [53] A. G. Red\feld. The theory of relaxation processes. Adv. Magn. Reson., 1:1{32, 1965. [54] W. T. Pollard, A. K. Felts, and R. A. Friesner. The red\feld equation in condensed-phase quantum dynamics. Advances in Chemical Physics, 93:77{134, 1996. [55] J. M. Jean, R. A. Friesner, and G. R. Fleming. Application of a multi- level red\feld theory to electron-transfer in condensed phases. Journal of Chemical Physics, 96(8):5827{5842, 1992. [56] M. Michel, O. Hess, H. Wichterich, and J. Gemmer. Transport in open spin chains: A monte carlo wave-function approach. Physical Review B, 77(10):104303, 2008. [57] S. Andergassen, V. Meden, J. Splettstoesser, H. Schoeller, and M. R. Wegewijs. Charge transport through single molecules, quantum dots and quantum wires. Nanotechnology, 21(27), 2010. [58] M. Michel, M. Hartmann, J. Gemmer, and G. Mahler. Fourier's law con\frmed for a class of small quantum systems. European Physical Journal B, 34(3):325{330, 2003. [59] T. Prosen and M. Znidaric. Matrix product simulations of non- equilibrium steady states of quantum spin chains. Journal of Statistical Mechanics-Theory and Experiment, page P02035, 2009. 121 Bibliography [60] K. Saito. Strong evidence of normal heat conduction in a one- dimensional quantum system. Europhysics Letters, 61(1):34{40, 2003. [61] W. Huisinga, L. Pesce, R. Koslo\u000b, and P. Saalfrank. Faber and newton polynomial integrators for open-system density matrix propagation. Journal of Chemical Physics, 110(12):5538{5547, 1999. [62] W. T. Pollard and R. A. Friesner. Solution of the red\feld equation for the dissipative quantum dynamics of multilevel systems. Journal of Chemical Physics, 100(7):5054{5065, 1994. [63] H. Guo and R. Q. Chen. Short-time chebyshev propagator for the Liouville-Von Neumann equation. Journal of Chemical Physics, 110(14):6626{6634, 1999. [64] I. Kondov, U. Kleinekathofer, and M. Schreiber. E\u000eciency of di\u000berent numerical methods for solving red\feld equations. Journal of Chemical Physics, 114(4):1497{1504, 2001. [65] H. Wichterich, M. J. Henrich, H.-P. Breuer, J. Gemmer, and M. Michel. Modeling heat transport through completely positive maps. Physical Review E, 76(3):031115, 2007. [66] H.-P. Breuer and F. Petruccione. Stochastic dynamics of quantum jumps. Physical Review E, 52(1):428{441, 1995. [67] P Saalfrank. Stochastic wave packet vs direct density matrix solu- tion of liouville-von neumann equations for photodesorption problems. Chemical Physics, 211(1-3):265{276, 1996. [68] D. Y. Petrina. Mathematical foundations of quantum statistical me- chanics. Kluwer Acad. Publ., 1995. [69] L. P. Kadano\u000b and P. C. Martin. Theory of many-particle systems II: superconductivity. Physical Review, 124(3):670{697, 1961. [70] C. W. Gardiner and P. Zoller. Quantum Noise. Springer, 2000. [71] K. E. Cahill and R. J. Glauber. Density operators for fermions. Phys- ical Review A, 59(2):1538{1555, 1999. [72] P. D. Drummond, P. Deuar, and J. F. Corney. Quantum many-body simulations using gaussian phase-space representations. Optics and Spectroscopy, 103(1):7{16, 2007. 122 Bibliography [73] P. Deuar and P. D. Drummond. First-principles quantum dynamics in interacting bose gases: I. the positive p representation. Journal of Physics A-Mathematical and General, 39(5):1163{1181, 2006. [74] P. Deuar and P. D. Drummond. First-principles quantum dynamics in interacting bose gases ii: stochastic gauges. Journal of Physics A-Mathematical and General, 39(11):2723{2755, 2006. [75] P. Deuar and P. D. Drummond. Correlations in a bec collision: First- principles quantum dynamics with 150 000 atoms. Physical Review Letters, 98(12):120402, 2007. [76] S. Lepri, R. Livi, and A. Politi. Thermal conduction in classical low- dimensional lattices. Physics Reports-Review Section of Physics Let- ters, 377(1):1{80, 2003. [77] K. Damle and S. Sachdev. Spin dynamics and transport in gapped one- dimensional heisenberg antiferromagnets at nonzero temperatures. Physical Review B, 57(14):8307{8339, 1998. [78] A. V. Sologubenko, S. M. Kazakov, H. R. Ott, T. Asano, and Y. Ajiro. Di\u000busive energy transport in the s=1 haldane-gap spin chain compound agvp2s6. Journal of Magnetism and Magnetic Materials, 272(Suppl. 1):E689{E690, 2004. [79] A. V. Sologubenko, K. Gianno, H. R. Ott, A. Vietkine, and A. Revcolevschi. Heat transport by lattice and spin excitations in the spin-chain compounds srcuo2 and sr2cuo3. Physical Review B, 64(5):054412, 2001. [80] N. Hlubek, P. Ribeiro, R. Saint-Martin, A. Revcolevschi, G. Roth, G. Behr, B. Buechner, and C. Hess. Ballistic heat transport of quantum spin excitations as seen in srcuo2. Physical Review B, 81(2):020405, 2010. [81] K. R. Thurber, A. W. Hunt, T. Imai, and F. C. Chou. O-17 nmr study of q=0 spin excitations in a nearly ideal s=1\/2 1d heisenberg antiferro- magnet, sr2cuo3, up to 800 k. Physical Review Letters, 87(24):247202, 2001. [82] F. Heidrich-Meisner, A. Honecker, D. C. Cabra, and W. Brenig. Zero- frequency transport properties of one-dimensional spin-1\/2 systems. Physical Review B, 68(13):134436, 2003. 123 Bibliography [83] K. Saito, S. Takesue, and S. Miyashita. Thermal conduction in a quantum system. Physical Review E, 54(3):2404{2408, 1996. [84] P. Jordan and E. Wigner. ber das paulische quivalenzverbot. Zeitschrift fr Physik A Hadrons and Nuclei, 47(9-10):631, 1928. [85] M. A. Schlosshauer. Decoherence and the Quantum-To-Classical Tran- sition. Springer, 2007. [86] J.-H. An and W.-M. Zhang. Non-markovian entanglement dynamics of noisy continuous-variable quantum channels. Physical Review A, 76(4):042127, 2007. [87] K. Blum. Density Matrix Theory and Applications: Physics of Atoms and Molecules. Springer, 2nd edition, 1996. [88] M. Michel, J. Gemmer, and G. Mahler. Heat conductivity in small quantum systems: Kubo formula in liouville space. European Physical Journal B, 42(4):555{559, 2004. [89] M. Michel, G. Mahler, and J. Gemmer. Fourier's law from schrodinger dynamics. Physical Review Letters, 95(18):180602, 2005. [90] G. Lindblad. Generators of quantum dynamical semigroups. commu- nications in mathematical physics, 48(2):119{130, 1976. [91] V. Gorini, A. Kossakowski, and E. C. G. Sudarshan. Completely posi- tive dynamical semigroups of N-level systems. Journal of Mathematical Physics, 17(5):821{825, 1976. [92] U. Brandt and M. Moraweck. Rigorous results on the dc kubo- greenwood conductivity of disordered-systems. Journal of Physics C- Solid State Physics, 15(25):5255{5268, 1982. [93] D. J. Thouless and S. Kirkpatrick. Conductivity of the disordered linear-chain. Journal of Physics C-Solid State Physics, 14(3):235{245, 1981. [94] M. Kira and S. W. Koch. Many-body correlations and excitonic ef- fects in semiconductor spectroscopy. Progress in quantum electronics, 30(5):155{296, 2006. [95] P. C. Martin and J. Schwinger. Theory of many-particle systems i. Physical Review, 115(6):1342{1373, 1959. 124 Bibliography [96] D. Jaksch, C. Bruder, J. I. Cirac, C. W. Gardiner, and P. Zoller. Cold bosonic atoms in optical lattices. Physical Review Letters, 81(15):3108{3111, 1998. [97] M. Greiner, O. Mandel, T. Esslinger, T. W. Hansch, and I. Bloch. Quantum phase transition from a super\ruid to a mott insulator in a gas of ultracold atoms. NATURE, 415(6867):39{44, 2002. [98] H. Risken. The Fokker-Planck Equation: Methods of Solutions and Applications. Springer, 1989. [99] T. Prosen and B. Zunkovic. Exact solution of markovian master equations for quadratic fermi systems: thermal baths, open xy spin chains and non-equilibrium phase transition. New Journal of Physics, 12:025016, 2010. [100] S. R. Clark, J. Prior, M. J. Hartmann, D. Jaksch, and M. B. Plenio. Exact matrix product solutions in the heisenberg picture of an open quantum spin chain. New Journal of Physics, 12:025005, 2010. [101] J. Wu. Non-equilibrium stationary states from the equation of motion of open systems. New Journal of Physics, 12:083042, 2010. [102] L. I. Plimak, M. K. Olsen, M. Fleischhauer, and M. J. Collett. Beyond the fokker-planck equation: Stochastic simulation of complete wigner representation for the optical parametric oscillator. Europhysics Let- ters, 56(3):372{378, 2001. [103] Benoit Palmieri, Yuki Nagata, and Shaul Mukamel. Phase-space algo- rithm for simulating quantum nonlinear response functions of bosons using stochastic classical trajectories. Physical Review E, 82:046706, 2010. [104] S.E. Ho\u000bmann, J.F. Corney, and P. D. Drummond. Hybrid phase- space simulation method for interacting bose \felds. Physical Review A, 78:013622, 2008. [105] J. et al. Hope. Xmds: extensible multi-dimensional simulator, http:\/\/xmds.sourceforge.net. [106] C. Mejia-Monasterio, T. Prosen, and G. Casati. Fourier's law in a quantum spin chain and the onset of quantum chaos. Europhysics Letters, 72(4):520{526, 2005. 125 Bibliography [107] M. Distasio and X. Zotos. Connection between low-energy e\u000bective- hamiltonians and energy-level statistics. Physical Review Letters, 74(11):2050{2053, 1995. [108] Y. Avishai, J. Richert, and R. Berkovits. Level statistics in a heisenberg chain with random magnetic \feld. Physical Review B, 66(5):052416, 2002. [109] T. A. Brody, J. Flores, J. B. French, P. A. Mello, A. Pandey, and S. S. M. Wong. Random-matrix physics - spectrum and strength \ruc- tuations. Reviews of Modern Physics, 53(3):385{479, 1981. [110] T. C. Hsu and J. C. A. Dauriac. Level repulsion in integrable and almost-integrable quantum spin models. Physical Review B, 47(21):14291{14296, 1993. [111] K. Kudo and T. Deguchi. Unexpected non-wigner behavior in level- spacing distributions of next-nearest-neighbor coupled xxz spin chains. Physical Review B, 68(5):052510, 2003. [112] K. Kudo and T. Deguchi. Level statistics of xxz spin chains with discrete symmetries: Analysis through \fnite-size e\u000bects. Journal of the Physical Society of Japan, 74(7):1992{2000, 2005. [113] K. Kudo and T. Deguchi. Level statistics of xxz spin chains with a random magnetic \feld. Physical Review B, 69(13):132404, 2004. [114] V. E. Korepin, N. M. Bogoliubov, and A. G. Izergin. Quantum Inverse Scattering Method and Correlation Functions. Cambridge University Press, 1993. [115] A. Klumper and K. Sakai. The thermal conductivity of the spin- 1\/2 xxz chain at arbitrary temperature. Journal of Physics A- Mathematical and General, 35(9):2173{2182, 2002. [116] V. E. Korepin, N. M. Bogoliubov, and A. G. Izergin. Quantum Inverse Scattering Method and Correlation Functions. Cambridge University Press, 1993. [117] M. Shiroishi and M. Wadati. Integrable boundary conditions for the one-dimensional hubbard model. Journal of the Physical Society of Japan, 66(8):2288{2301, 1997. 126 [118] E. K. Sklyanin. Boundary-conditions for integrable quantum-systems. Journal of Physics A-Mathematical and General, 21(10):2375{2389, 1988. [119] T. Prosen. Open xxz spin chain: Nonequilibrium steady state and strict bound on ballistic transport. arXiv:1103.1350v2, 2011. [120] A. V. Sologubenko, E. Felder, K. Gianno, H. R. Ott, A. Vietkine, and A. Revcolevschi. Thermal conductivity and speci\fc heat of the linear chain cuprate sr2cuo3: Evidence for thermal transport via spinons. Physical Review B, 62(10):R6108{R6111, 2000. [121] Y. Ando, J. Takeya, D. L. Sisson, S. G. Doettinger, I. Tanaka, R. S. Feigelson, and A. Kapitulnik. Thermal conductivity of the spin-peierls compound cugeo3. Physical Review B, 58(6):R2913{R2916, 1998. [122] Y. Saad and M.H. Schultz. Gmres - a generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM Journal on Scienti\fc and Statistical Computing, 7(3):856{869, 1986. [123] R. V. Jensen and R. Shankar. Statistical behavior in deterministic quantum-systems with few degrees of freedom. Physical Review Let- ters, 54(17):1879{1882, 1985. [124] M. Schlosshauer. Decoherence, the measurement problem, and in- terpretations of quantum mechanics. Review of Modern Physics, 76(4):1267{1305, 2004. [125] F. M. Cucchietti, J. P. Paz, and W. H. Zurek. Decoherence from spin environments. Physical Review A, 72(5):052113, 2005. [126] G. Hu. Non-equilibrium Statistical Physics. unpublished, 2002. 127 Appendix A Perturbational decomposition of the operators m\u0302 In this appendix we will derive expressions of operators m\u0302 and \u0016\u0302m in two ways: using the eigenvectors of HS = H0+ VS and expanding the operators perturbationally in orders of V0 based on expressions of the operators for the corresponding non-interaction H0. A.1 Expanded in eigenbasis of HS From its de\fnition in Eq(4.3), in the representation of eigenvalues fEmg and eigenstates fjmig of HS , the operator m\u0302 can be written as hm jm\u0302\u000bjni = (m\u0302\u000b)mn = (c\u000b)mn X k jV \u000bk j2 Z 1 0 d\u001cei(En\u0000Em\u0000!k;\u000b)\u001c h1\u0000 n\u000b (!k;\u000b)i = (c\u000b)mn \u0019 Z d!D\u000b (!) jV \u000b (!) j2h1\u0000 n\u000b (!)i\u000e ( mn \u0000 !) = (c\u000b)mn \u0019D \u000b (En \u0000 Em) jV \u000b (En \u0000Em) j2h1\u0000 n\u000b (En \u0000 Em)i; (A.1) where we have used R1 0 d\u001ce i!\u001c = \u0019\u000e (!)+iP \u0000 1 ! \u0001 and neglected the principal value part. We have also assumed that it is possible to perform a change of variable on V \u000bk such that it becomes V \u000b (knm), where kmn is de\fned by !kmn;\u000b = mn = Em \u0000 En = \u0000 nm, i.e. a bath mode resonant with this transition. This limits the possible forms of V \u000bk and !k;\u000b. For example, for each given energy mn, there should be a unique value of V knm . In all the work in this thesis, we take V \u000bk as a constant so this condition is satis\fed. 128 A.2. Perturbational expansion starting from non-interacting systems D\u000b(!) is the bath's density of states. We arrive at m\u0302\u000b = \u0019 X m;n jmihnjhmjc\u000bjni (1\u0000 n\u000b ( nm))D\u000b ( nm) jV \u000bknm j2; (A.2a) \u0016\u0302m\u000b = \u0019 X m;n jmihnjhmjcy\u000bjnin\u000b ( mn)D\u000b ( mn) jV \u000bkmn j2: (A.2b) We furthermore set V \u000bknmD\u000b(!) as a constant and absorb it into \u0015 2. Sim- ilar expressions can be derived for bosonic systems with the Fermi-Diract distribution replaced by the Bose-Einstein distribution and (1\u0000 n\u000b ( nm)) replaced by (1 + n\u000b ( nm)). This approach involves a direct diagonalization of the isolated system HS . One can avoid the diagonalization by \fnding such operators m\u0302 perturbationally. A.2 Perturbational expansion starting from non-interacting systems To present the general idea concretely, let us work on a speci\fc system, the Hamiltonian HS = H0 + VS de\fned in Eq. (4.1), which we rewrite here for convenience, HS = \u0000t N\u00001X l=1 \u0010 cyl cl+1 + c y l+1cl \u0011 + V0 N\u00001X l=1 cyl+1cl+1c y l cl = H0 + VS : (A.3) Next assuming V0 is small and treating it perturbationally, we want to ex- press the operator m\u0302\u000b in terms of polynomials of fcmg operators and in various orders of V0. When V0 = 0 the system is a tight-binding open chain. The following basis transformation | discrete Fourier transform for open chains ck = 1p N NX l=1 sin kl\u0019 N + 1 cl; (A.4) diagonalizes H0, H0 = NX k=1 \u000fkc y kck; (A.5) where \u000fk = \u00002t cos \u0019k N + 1 : (A.6) 129 A.2. Perturbational expansion starting from non-interacting systems Therefore, c\u000b (t) is a linear function of all cm, cl (t) \u0011 c(0)l (t) = 2 N + 1 X km sin \u0019kl N + 1 sin \u0019km N + 1 e\u0000i\u000fktcm: (A.7) Hence, m\u0302\u000b is also a linear combination of all cm operators. One can imagine that for small V0, m\u0302L should not be too far from such a linear combination. Let us call cl (t) when V0 = 0 as c (0) l (t). Treating this as the zeroth order form for the full dynamical cl (t), and expanding cl (t) = X n V n0 c (n) l (t) ; (A.8) we derive a perturbation equation of c (n) l (t): _c (n) l = it \u0010 c (n) l\u00001 + c (n) l+1 \u0011 \u0000 id(n\u00001)l ; (A.9) where we de\fne, for non-negative integers n; n1; n2; n3, d (n) l = X n1;n2;n3P i ni=n n c (n1) l c y;(n2) l\u00001 c (n3) l\u00001 + c (n1) l c y;(n2) l+1 c (n3) l+1 o : (A.10) The solution of the above equation can be written as c (n) l (t) = \u0000i Z t 0 d\u001c 2 N + 1 X km sin \u0019kl N + 1 sin \u0019km N + 1 e\u0000i\u000fk(t\u0000\u001c)d(n\u00001)m (\u001c) : (A.11) Here the initial condition c(n) (0) = 0(8n \u0015 1) is used. As the expression of c (0) l is given in Eq. (A.7), explicitly the expression of c (1) l is c (1) l (t) = \u0012 2 N + 1 \u00134 X km kimi ei(\u000f(k2)\u0000\u000f(k1)\u0000\u000f(k3))t \u0000 ei\u000f(k)t \u000f (k1) + \u000f (k3)\u0000 \u000f (k2)\u0000 \u000f (k) \u0001 sin \u0019kl N + 1 sin \u0019k1l N + 1 sin \u0019km N + 1 sin \u0019k1m1 N + 1 sin \u0019k2m2 N + 1 sin \u0019k3m3 N + 1 \u0001 \u0014 sin \u0019k2 (l \u0000 1) N + 1 sin \u0019k3 (l \u0000 1) N + 1 + sin \u0019k2 (l + 1) N + 1 sin \u0019k3 (l + 1) N + 1 \u0015 \u0001cm1cym2cm3 : (A.12) 130 A.2. Perturbational expansion starting from non-interacting systems Similarly we can derive c (2) l , which will be shown explicitly later when we discuss terms proportional to V 20 in bath operators m\u0302. Plugging this gen- eral solution into Eq(4.3), and after straightforward but tedious algebra, we arrive at the decomposition of m\u0302\u000b and \u0016\u0302m\u000b in Eq(4.34) with expansion coe\u000ecients de\fned as follows, D\u000b;m = \u0019 2 N + 1 X k sin \u0019kl N + 1 sin \u0019km N + 1 [1\u0000 n (\u000fk; T\u000b)] ; (A.13a) \u0016D\u000b;m = \u0019 2 N + 1 X k sin \u0019kl N + 1 sin \u0019km N + 1 n (\u000fk; T\u000b) ; (A.13b) for the zeroth order expansion, D\u000b;m1m2m3 = \u0019 \u0012 2 N + 1 \u00134 X k;m;k1;k2;k2 \u0001n (\u000f (k) ; T\u000b)\u0000 n (\u000f (k1) + \u000f (k3)\u0000 \u000f (k2) ; T\u000b) \u000f (k1) + \u000f (k3)\u0000 \u000f (k2)\u0000 \u000f (k) \u0001 sin k\u0019l N + 1 sin k1\u0019m1 N + 1 sin k2\u0019m2 N + 1 sin k3\u0019m3 N + 1 sin k\u0019m N + 1 sin k1\u0019m N + 1 \u0001 \u0012 sin k2\u0019 (m+ 1) N + 1 sin k3\u0019 (m+ 1) N + 1 + sin k2\u0019 (m\u0000 1) N + 1 sin k3\u0019 (m\u0000 1) N + 1 \u0013 ; (A.14) 131 A.2. Perturbational expansion starting from non-interacting systems for the \frst order expansion. For the second order expansion we make use of c (2) l and we have D\u000b;m1m2m3m4m5 = \u0019 \u0012 2 N + 1 \u00137 X k;m;k0;m0;k1;k2;k3;k4;k5 \u0001 sin \u0019km N + 1 \u0012 \u00055i=0 sin \u0019kimi N + 1 \u0013 sin \u0019kl N + 1 sin \u0019k1l N + 1\u001a 1 \u000f (k1) + \u000f (k3)\u0000 \u000f (k2)\u0000 \u000f (k0) \u0014 n (\u000f (k))\u0000 n (\u000f (k5)\u0000 \u000f (k0)\u0000 \u000f (k4)) \u000f (k0) + \u000f (k4)\u0000 \u000f (k5) + \u000f (k) \u0000n (\u000f (k))\u0000 n (\u000f (k1) + \u000f (k3) + \u000f (k5)\u0000 \u000f (k2)\u0000 \u000f (k4)) \u000f (k2) + \u000f (k4)\u0000 \u000f (k1)\u0000 \u000f (k3)\u0000 \u000f (k5) + \u000f (k) \u0015 \u0001 sin \u0019k0l N + 1 X \u000e=\u00061 \u0012 sin \u0019k2 (l\u000b + \u000e) N + 1 sin \u0019k3 (l\u000b + \u000e) N + 1 \u0013 \u0001 X \u000e=\u00061 \u0012 sin \u0019k4 (l\u000b + \u000e) N + 1 sin \u0019k5 (l\u000b + \u000e) N + 1 \u0013 + 1 \u000f (k2) + \u000f (k4)\u0000 \u000f (k3)\u0000 \u000f (k0) \u0014 n (\u000f (k))\u0000 n (\u000f (k0) + \u000f (k1) + \u000f (k5)) \u000f (k)\u0000 \u000f (k0)\u0000 \u000f (k1)\u0000 \u000f (k5) \u0000n (\u000f (k))\u0000 n (\u000f (k1) + \u000f (k3) + \u000f (k5)\u0000 \u000f (k2)\u0000 \u000f (k4)) \u000f (k2) + \u000f (k4)\u0000 \u000f (k1)\u0000 \u000f (k3)\u0000 \u000f (k5) + \u000f (k) \u0015 \u0001 X \u000e=\u00061 \u0012 sin \u0019k0 (l\u000b + \u000e) N + 1 sin \u0019k4 (l\u000b + \u000e) N + 1 \u0019k5 (l\u000b + \u000e) N + 1 \u0001 \u0014 sin \u0019k2l N + 1 sin \u0019k3l N + 1 + sin \u0019k2 (l\u000b + 2\u000e) N + 1 sin \u0019k3 (l\u000b + 2\u000e) N + 1 \u0015\u0013 + 1 \u000f (k3) + \u000f (k5)\u0000 \u000f (k4)\u0000 \u000f (k0) \u0014 n (\u000f (k))\u0000 n (\u000f (k1)\u0000 \u000f (k0)\u0000 \u000f (k2)) \u000f (k) + \u000f (k0)\u0000 \u000f (k1) + \u000f (k2) \u0000n (\u000f (k))\u0000 n (\u000f (k1) + \u000f (k3) + \u000f (k5)\u0000 \u000f (k2)\u0000 \u000f (k4)) \u000f (k2) + \u000f (k4)\u0000 \u000f (k1)\u0000 \u000f (k3)\u0000 \u000f (k5) + \u000f (k) \u0015 \u0001 X \u000e=\u00061 \u0012 sin \u0019k0 (l\u000b + \u000e) N + 1 sin \u0019k2 (l\u000b + \u000e) N + 1 sin \u0019k3 (l\u000b + \u000e) N + 1 \u0001 \u0014 sin \u0019k4l N + 1 sin \u0019k5l N + 1 + sin \u0019k4 (l\u000b + 2\u000e) N + 1 sin \u0019k5 (l\u000b + 2\u000e) N + 1 \u0015\u0013\u001b : (A.15) Here we introduce a short-hand notation \u000e = \u00061 in (l\u000b + \u000e) to stand for the neighbor sites of l\u000b. In the last expression we have denoted n (\u000f (k) ; T\u000b) as 132 A.2. Perturbational expansion starting from non-interacting systems n (\u000f (k)) to make it \ft into the page width. Therefore up the the order of V 20 , the bath operators can be written down as m\u0302\u000b = X m D\u000b;mcm + V0 X m1m2m3 D\u000b;m1m2m3cm1c y m2cm3 +V 20 X m1m2m3m4m5 D\u000b;m1m2m3m4m5cm1c y m2cm3c y m4cm5 +O \u0000 V 30 \u0001 (A.16a) \u0016\u0302m\u000b = X m \u0016D\u000b;mc y m \u0000 V0 X m1m2m3 D\u000b;m1m2m3c y m3cm2c y m1 \u0000V 20 X m1m2m3m4m5 D\u000b;m1m2m3m4m5c y m5cm4c y m3cm2c y m1 +O \u0000 V 30 \u0001 : (A.16b) The same procedure is applicable for bosonic systems and similar ex- pressions can be derived. This has been used in Chapter 5 to calculate the operators m\u0302\u000b and \u0016\u0302m\u000b, and \fnd non-equilibrium stationary states for bosonic systems using the BBGKY-like method. 133 Appendix B Second-order cluster expansion and second-order equations of G1, G2 In this chapter we present explicitly forms of the second-order cluster ex- pansion and the second-order equations for G1, G2. B.1 Second-order cluster expansion First-order cluster expansion as in Eq. (4.31) expresses G2 in terms of G1 and neglecting G2. For convenience, we rewrite it here, G2 \u0010 my; ny;m 0 ; n 0\u0011 = \u0000G1 \u0010 my;m 0\u0011 G1 \u0010 ny; n 0\u0011 +G1 \u0010 my; n 0\u0011 G1 \u0010 ny;m 0\u0011 +G2 \u0010 my; ny;m 0 ; n 0\u0011 : (B.1) 134 B.2. Second-order equations of G1 and G2 A second-order cluster expansion similarly writes G3 in terms of G1; G2 and neglecting G3[94], G \u0010 my1;m y 2;m y 3;m4;m5;m6 \u0011 = G \u0010 my1;m6 \u0011 G \u0010 my2;m5 \u0011 G \u0010 my3;m4 \u0011 +G \u0010 my1;m5 \u0011 G \u0010 my2;m4 \u0011 G \u0010 my3;m6 \u0011 +G \u0010 my1;m4 \u0011 G \u0010 my2;m6 \u0011 G \u0010 my3;m5 \u0011 \u0000G \u0010 my1;m6 \u0011 G \u0010 my2;m4 \u0011 G \u0010 my3;m5 \u0011 \u0000G \u0010 my1;m5 \u0011 G \u0010 my2;m6 \u0011 G \u0010 my3;m4 \u0011 \u0000G \u0010 my1;m4 \u0011 G \u0010 my2;m5 \u0011 G \u0010 my3;m6 \u0011 (B.2a) +G \u0010 my1;m6 \u0011 G \u0010 my2;m y 3;m4;m5 \u0011 \u0000G \u0010 my1;m5 \u0011 G \u0010 my2;m y 3;m4;m6 \u0011 +G \u0010 my1;m4 \u0011 G \u0010 my2;m y 3;m5;m6 \u0011 +G \u0010 my2;m5 \u0011 G \u0010 my1;m y 3;m4;m6 \u0011 \u0000G \u0010 my2;m4 \u0011 G \u0010 my1;m y 3;m5;m6 \u0011 \u0000G \u0010 my2;m6 \u0011 G \u0010 my1;m y 3;m4;m5 \u0011 +G \u0010 my3;m4 \u0011 G \u0010 my1;m y 2;m5;m6 \u0011 \u0000G \u0010 my3;m5 \u0011 G \u0010 my1;m y 2;m4;m6 \u0011 +G \u0010 my3;m6 \u0011 G \u0010 my1;m y 2;m4;m5 \u0011 (B.2b) +G \u0010 my1;m y 2;m y 3;m4;m5;m6 \u0011 : (B.2c) The same equation has been written down in short-hand notation in Eq. (4.42). B.2 Second-order equations of G1 and G2 Using the perturbational expansion of bath operators Eq. (A.13), Eq. (A.14) and Eq. (A.15) together with cluster expansions Eq. (B.1) and Eq. (B.2), 135 B.2. Second-order equations of G1 and G2 Eq. (4.7) of G1 | the \frst equation of the equation hierarchy becomes 0 = itG \u0010 (m\u0000 1)y; n \u0011 + itG \u0010 (m+ 1)y; n \u0011 \u0000itG \u0010 my; n+ 1 \u0011 \u0000 itG \u0010 my; n\u0000 1 \u0011 +\u00152 X l; h \u000en \u0000 D\u000b;l + \u0016D \u0003 \u000b;l \u0001 G \u0010 my; l \u0011 +\u000em \u0000 \u0016D\u000b;l +D \u0003 \u000b;l \u0001 G \u0010 ly; n \u0011i +\u00152V0 X \u000b;m1;m2 (D\u000b;nm2m1 \u0000D\u000b;m1m2n)G \u0010 my1;m2 \u0011 \u000em +\u00152V0 X \u000b;m1;m2 (D\u000b;mm1m2 \u0000D\u000b;m2m1m)G \u0010 my1;m2 \u0011 \u000en +\u00152V 20 X \u000b;m1;m2;m3 [D\u000b;m3m2m1m3n +D\u000b;m3m3nm2m1 +D\u000b;nm2m3m3m1 \u0000D\u000b;m1m2m3m3n \u0000D\u000b;m3m3m1m2n \u0000D\u000b;m3m2nm3m1 ]G \u0010 my1;m2 \u0011 \u000em +\u00152V 20 X \u000b;m1;m2;m3 [D\u000b;mm1m3m3m2 +D\u000b;m3m3mm1m2 +D\u000b;m3m1m2m3m \u0000D\u000b;m3m1mm3m2 \u0000D\u000b;m3m3m2m1m \u0000D\u000b;m2m1m3m3m]G \u0010 my1;m2 \u0011 \u000en\u000b (B.3a) \u0000\u00152 X \u0000 \u000em\u000b \u0016D\u000b;n + \u000en\u000b \u0016D \u0003 \u000b;m \u0001 +\u00152V0 X \u000b;m1 (D\u000b;m1m1n\u000em\u000b +D\u000b;m1m1m\u000en\u000b) +\u00152V 20 X \u000b;m1;m2 (D\u000b;m1m1m2m2n\u000em\u000b +D\u000b;m1m1m2m2m\u000en\u000b) (B.3b) \u0000iV0G \u0010 my; (n\u0000 1)y; n; n\u0000 1 \u0011 + iV0G \u0010 (m+ 1)y;my; (m+ 1); n \u0011 \u0000iV0G \u0010 (n+ 1)y;my; n+ 1; n \u0011 + iV0G \u0010 my; (m\u0000 1)y; n;m\u0000 1 \u0011 +\u00152V 20 X \u000b;m1;m2;m3;m4 n G \u0010 my1;m y 2;m3;m4 \u0011 \u0001 [\u000em\u000b (D\u000b;m1m3nm4m2 +D\u000b;nm3m2m4m1 \u0000D\u000b;m2m4m1m3n) + \u000en\u000b (D\u000b;m3m1mm2m4 \u0000D\u000b;mm1m3m2m4 \u0000D\u000b;m4m2m3m1m)]g (B.3c) 136 B.2. Second-order equations of G1 and G2 Using the condensed \u0000 matrices notation introduced in Chapter 4, we can denote this equation as \u0010 \u0000 (1) H0 + \u00152\u0000 (1) B0 + \u00152V0\u0000 (1) B1 + \u00152V0\u0000 (1) B2 \u0011 g1 = \u00152\u00170 + \u0015 2V0\u00171 + \u0015 2V 20 \u00172 + \u0000 iV0 + \u0015 2V 20 \u0001 \u0019 (g1g1) + \u0000 iV0 + \u0015 2V 20 \u0001 g2: (B.4) Similarly we can derive an equation for g2 truncated at the order of V 2 0 using these bath operators. It is a 10-pages long tedious equation so we here just write down its formal form in condensed matrices notation,\u0010 \u0000 (2) H0 + \u00152\u0000 (2) B0 + \u00152V0\u0000 (2) B1 + \u00152V0\u0000 (2) B2 \u0011 (g2 + \u0019 (g1g1)) = \u00152g1 + \u0015 2V0g1 + \u0015 2V 20 g1 + \u0000 iV0 + \u0015 2V 20 \u0001 \u0019 (g1g1g1) + \u0000 iV0 + \u0015 2V 20 \u0001 \u0019 (g1g2) : (B.5) In the calculation illustrated in the main text, we truncated the bath oper- ators at the order of V0 instead of V 2 0 . In this case, the second equation of 137 B.2. Second-order equations of G1 and G2 the hierarchy becomes, 0 = itG \u0010 my; (n\u0000 1)y ;m0 ; n0 \u0011 + itG \u0010 my; (n+ 1)y ;m 0 ; n 0\u0011 +itG \u0010 (m\u0000 1)y ; ny;m0 ; n0 \u0011 + itG \u0010 (m+ 1)y ; ny;m 0 ; n 0\u0011 \u0000itG \u0010 my; ny;m 0 ; \u0010 n 0 + 1 \u0011\u0011 \u0000 itG \u0010 my; ny;m 0 ; \u0010 n 0 \u0000 1 \u0011\u0011 \u0000itG \u0010 my; ny; \u0010 m 0 + 1 \u0011 ; n 0\u0011\u0000 itG\u0010my; ny;\u0010m0 \u0000 1\u0011 ; n0\u0011 +iV0G \u0010 my; ny;m 0 ; n 0\u0011\u0010 \u000em0+1;n0 + \u000em0\u00001;n0 \u0011 \u0000iV0G \u0010 my; ny;m 0 ; n 0\u0011 (\u000em+1;n + \u000em\u00001;n) (B.6a) \u0000iV0 X l=m\u00061;n\u00061 G \u0010 ly;my; ny; l;m 0 ; n 0\u0011 +iV0 X l=m0\u00061;n0\u00061 G \u0010 ly;my; ny; l;m 0 ; n 0\u0011 (B.6b) \u0000\u00152 X n \u000em h G \u0010 ny;m 0\u0011 \u0016D\u000b;n0 \u0000G\u0010ny; n0\u0011 \u0016D\u000b;m0i \u0000\u000en h G \u0010 my;m 0\u0011 \u0016D\u000b;n0 \u0000G\u0010my; n0\u0011 \u0016D\u000b;m0i +\u000em0 h G \u0010 my; n 0\u0011 \u0016D\u000b;n \u0000G\u0010ny; n0\u0011 \u0016D\u000b;mi \u0000\u000en0 h G \u0010 my;m 0\u0011 \u0016D\u000b;n \u0000G\u0010ny;m0\u0011 \u0016D\u000b;mio (B.6c) +\u00152 X \u000b;l \u0000 D\u000b;l + \u0016D\u000b;l \u0001 \u0001 h \u000em\u000bG \u0010 ly; ny;m 0 ; n 0\u0011 + \u000en\u000bG \u0010 my; ly;m 0 ; n 0\u0011 +\u000em0\u000bG \u0010 my; ny; l; n 0\u0011 + \u000en0\u000bG \u0010 my; ny;m 0 ; l \u0011i (B.6d) \u0001 \u0001 \u0001 to be continued 138 B.2. Second-order equations of G1 and G2 continued +V0\u0015 2 X \u000b;m1 n \u000em0 h G \u0010 my; n 0\u0011 D\u000b;m1m1n \u0000G \u0010 ny; n 0\u0011 D\u000b;m1m1m +G \u0010 my1; n 0\u0011 (D\u000b;nm1m \u0000D\u000b;mm1n) i \u0000\u000en0 h G \u0010 my;m 0\u0011 D\u000b;m1m1n \u0000G \u0010 ny;m 0\u0011 D\u000b;m1m1m +G \u0010 my1;m 0\u0011 (D\u000b;nm1m \u0000D\u000b;mm1n) i +\u000em h G \u0010 ny;m 0\u0011 D\u000b;m1m1n0 \u0000G \u0010 ny; n 0\u0011 D\u000b;m1m1m0 +G \u0010 ny;m1 \u0011\u0010 D\u000b;n0m1m0 \u0000D\u000b;m0m1n0 \u0011i \u0000\u000en h G \u0010 my;m 0\u0011 D\u000b;m1m1n0 \u0000G \u0010 my; n 0\u0011 D\u000b;m1m1m0 +G \u0010 my;m1 \u0011\u0010 D\u000b;n0m1m0 \u0000D\u000b;m0m1n0 \u0011io (B.6e) +V0\u0015 2 X \u000b;m1;m2 n \u000em0 h D\u000b;m1n0m2G \u0010 my; ny;m1;m2 \u0011 \u0000 (D\u000b;m1m2m \u0000D\u000b;mm2m1)G \u0010 my2; n y;m1; n 0\u0011 +(D\u000b;m1m2n \u0000D\u000b;nm2m1)G \u0010 my2;m y;m1; n 0\u0011i \u0000 \u000en0 h D\u000b;m1m0m2G \u0010 my; ny;m1;m2 \u0011 \u0000 (D\u000b;m1m2m \u0000D\u000b;mm2m1)G \u0010 my2; n y;m1;m 0\u0011 +(D\u000b;m1m2n \u0000D\u000b;nm2m1)G \u0010 my2;m y;m1;m 0\u0011i + \u000em h D\u000b;m1nm2G \u0010 my1;m y 2;m 0 ; n 0\u0011 + \u0010 D\u000b;m1m2n0 \u0000D\u000b;n0m2m1 \u0011 G \u0010 ny;my1;m 0 ;m2 \u0011 \u0000 \u0010 D\u000b;m1m2m0 \u0000D\u000b;m0m2m1 \u0011 G \u0010 ny;my1; n 0 ;m2 \u0011i \u0000 \u000en h D\u000b;m1mm2G \u0010 my1;m y 2;m 0 ; n 0\u0011 + \u0010 D\u000b;m1m2n0 \u0000D\u000b;n0m2m1 \u0011 G \u0010 my;my1;m 0 ;m2 \u0011 \u0000 \u0010 D\u000b;m1m2m0 \u0000D\u000b;m0m2m1 \u0011 G \u0010 my;my1; n 0 ;m2 \u0011io (B.6f) Here G3 should be interpreted according to Eq. (4.42) but neglecting G3. 139 B.2. Second-order equations of G1 and G2 However, for certain systems with relatively stronger interaction, it might be necessary to use these new equations at the order of V 20 . 140 Appendix C Estimation of convergence of the two methods of truncating the BBGKY-like hierarchy In this chaprter we present our estimation of the leading order of G1 such as \u0001 E;(1) 1 and \u0001 C;(1) 1 , de\fned in Chapter 4. We will see that \u0001 E;(1) 1 in fact involves \u0001 E;(0) 2 , which in turn needs Eq. (4.8) | the equation of G2. A similar equation is needed for estimation of \u0001 C;(1) 1 . C.1 \u0001 E;(1) 1 from method 1 In order to estimate the accuracy of the \frst-order form of this approxima- tion and also to get an estimate of accuracy for higher orders, we study the leading order of residues in terms of \u00152 and \u0001TT , both of which are assumed to be small in the following. Thus \u00152V0 \u001c V0, therefore we know that in Eq. (4.24) the gD term is smaller than the other g2 term so we drop it. Similarly since \u00152\u0001T \u001c \u0001T , we drop the \u00152\u0001T term in \u00152\u0017 in Eq(4.24), \u00152\u0017 = \u00152\u00170 (T ) + \u0015 2\u0001T\u0017;T ; (C.1) and keep only the large term, \u00152\u00170 (T ), which is independent of \u0001T . Here \u0017;T denotes formally a derivative of T on \u0017 | d dT \u0017. The general idea is then to write down respectively equations for gEx1 and g E;(1) 1 , and then compare the two to estimate \u0001 E;(1) 1 = g E;(1) 1 \u0000 gEx1 . In order to understand how an approximation to the next order improves the accuracy, we also want to compare \u0001 E;(1) 1 to \u0001 E;(0) 1 = g E;(0) 1 \u0000gEx1 , which is estimated in the same way from the di\u000berence between the equations respectively for gEx1 and g E;(0) 1 . More generally we de\fne \u0001 E;(0) n = g E;(0) n \u0000gExn and \u0001E;(1)n = gE;(1)n \u0000gExn for n-particle Green's functions gn. 141 C.1. \u0001 E;(1) 1 from method 1 After dropping the gD term and the term which is proportional to \u0015 2\u0001T , and keeping only up to the linear order of \u0001T , gEx1 ; g E;(0) 1 and g E;(1) 1 respec- tively satisfy \u0010 \u0000 (1) 0 + \u0000 (1) ;T \u0001T \u0011 gEx1 = iV0g Ex 2 + \u0015 2\u00170; (C.2a) \u0000 (1) 0 g E;(0) 1 = iV0g E;(0) 2 + \u0015 2\u00170; (C.2b)\u0010 \u0000 (1) 0 + \u0000 (1) ;T \u0001T \u0011 g E;(1) 1 = iV0g E;(0) 2 + \u0015 2\u00170; (C.2c) where \u0000 (1) 0 +\u0000 (1) ;T \u0001T is the zeroth and \frst order in \u0001T from \u0000 (1) of Eq(4.24). \u0000 (1) ;T denotes formally a derivative of T on \u0000 (1) | ddT \u0000 (1). To consider \u0001 E;(0) 1 , one may use Eq(C.2a) and Eq(C.2b), \u0001 E;(0) 1 = \u0001T \u0010 \u0000 (1) 0 \u0011\u00001 \u0000 (1) ;T g Ex 1 + iV0 \u0010 \u0000 (1) 0 \u0011\u00001 \u0001 E;(0) 2 : (C.3) \u0001 E;(1) 1 can be estimated from Eq(C.2a) and Eq(C.2c), \u0001 E;(1) 1 = iV0 \u0010 \u0000 (1) 0 + \u0000 (1) ;T \u0001T \u0011\u00001 \u0001 E;(0) 2 ; (C.4) where \u0001 E;(0) 2 is required. We \fnd that roughly speaking \u0001 E;(1) 1 takes the second term of \u0001 E;(0) 1 but drops the \frst term. Therefore, next we only need to show that the second term, iV0 \u0010 \u0000 (1) 0 \u0011\u00001 \u0001 E;(0) 2 , is much smaller than the \frst term, or equivalently, smaller than the whole \u0001 E;(0) 1 . Estimation of \u0001 E;(0) 2 involves the second equation of the hierarchy, i.e. equation ofG2, which can be derived from substituting Eq(4.18) into Eq(4.8): 142 C.1. \u0001 E;(1) 1 from method 1 0 = it D cymc y ncm0 cn0+1 E + it D cymc y ncm0 cn0\u00001 E +it D cymc y ncm0+1cn0 E + it D cymc y ncm0\u00001cn0 E \u0000it D cymc y n\u00001cm0 cn0 E \u0000 it D cymc y n+1cm0 cn0 E \u0000it D cym\u00001c y ncm0 cn0 E \u0000 it D cym+1c y ncm0 cn0 E +\u00152 X l; n \u000en0\u000bd\u000b;l D cymc y ncm0 cl E + \u000em0\u000bd\u000b;l D cymc y nclcn0 E +\u000en\u000b \u0016d\u000b;l D cymc y l cm0 cn0 E + \u000em\u000b \u0016d\u000b;l D cyl c y ncm0 cn0 E +\u000en\u000bd \u0003 \u000b;l D cymc y l cm0 cn0 E + \u000em\u000bd \u0003 \u000b;l D cyl c y ncm0 cn0 E +\u000en0 \u0016d\u0003\u000b;l D cymc y ncm0 cl E + \u000em0 \u0016d\u0003\u000b;l D cymc y nclcn0 Eo +iV0 D cymc y ncm0 cn0 E\u0010 \u000em0+1;n0 + \u000em0\u00001;n0 \u0000 \u000em+1;n \u0000 \u000em\u00001;n \u0011 (C.5a) \u0000iV0 X l=m\u00061;n\u00061 D cyl c y mc y nclcm0 cn0 E +iV0 X l=m 0\u00061;n0\u00061 D cyl c y mc y nclcm0 cn0 E (C.5b) \u0000\u00152 X h \u000en\u000b \u0016d\u000b;m0 D cymcn0 E + \u000em\u000b \u0016d\u000b;n0 D cyncm0 E +\u000em0 \u0016d\u0003\u000b;n D cymcn0 E + \u000en0 \u0016d\u0003\u000b;m D cyncm0 E \u0000\u000en\u000b \u0016d\u000b;n0 D cymcm0 E \u0000 \u000em\u000b \u0016d\u000b;m0 D cyncn0 E \u0000\u000en0\u000b \u0016d\u0003\u000b;n D cymcm0 E \u0000 \u000em0\u000b \u0016d\u0003\u000b;m D cyncn0 Ei (C.5c) \u0000\u00152V0 X n \u000em0 D cymc y ncn0D E \u0000 \u000en0 D cymc y ncm0D E +\u000em D cyncm0 cn0 \u0016D E \u0000 \u000en D cymcm0 cn0 \u0016D E + \u000en0 D \u0016Dy\u000bc y mc y ncm0 E \u0000\u000em0 D \u0016Dy\u000bc y mc y ncn0 E + \u000en D Dy\u000bc y mcm0 cn0 E \u0000 \u000em D Dy\u000bc y ncm0 cn0 Eo : (C.5d) This can be written in a compact matrix form as \u0000(2)gEx2 = iV0g Ex 3 + \u0015 2gEx1 + \u0015 2V0g Ex D3; (C.6) 143 C.1. \u0001 E;(1) 1 from method 1 where vector gEx2 is de\fned as follows, gEx2 = h G2 \u0010 1y; 1y; 1; 1 \u0011 ; G2 \u0010 1y; 1y; 1; 2 \u0011 ; \u0001 \u0001 \u0001 ; G2 \u0010 N y; N y; N;N \u0011iT : (C.7) Note that some of the elements of gEx2 are naturally zero but we still keep them in this vector. gEx3 , g Ex 1 and g Ex D3 come from ordering respectively Eq(C.5b), Eq(C.5c) and Eq(C.5d) in the same way as gEx2 . The matrix \u0000 (2) can be read o\u000b from Eq(C.5a), for example assuming m;n;m 0 ; n 0 are all di\u000berent and not equal to 1 or N , we have \u0000 (2) mN3+nN2+m 0 N+n 0 ;mN3+nN2+m 0 N+n 0 +1 = it: (C.8) Since \u00152V0 \u001c V0 we drop the gExD3 term in the following estimation of the accuracy. Therefore, the exact solution and the equilibrium solution satisfy respectively, \u0010 \u0000 (2) 0 + \u0000 (2) ;T \u0001T \u0011 gEx2 = iV0g Ex 3 + \u0015 2gEx1 ; (C.9a) \u0000 (2) 0 g E;(0) 2 = iV0g E;(0) 3 + \u0015 2g E;(0) 1 ; (C.9b) where \u0000 (2) 0 and \u0000 (2) ;T \u0001T stand for the zeroth and \frst order in \u0001T in \u0000 (2). Now \u0001 E;(0) 2 can be analyzed: \u0001 E;(0) 2 = iV0 \u0010 \u0000 (2) 0 \u0011\u00001 \u0001 E;(0) 3 + \u0010 \u0000 (2) 0 \u0011\u00001 \u0000 (2) ;T \u0001Tg Ex 2 + \u0010 \u0000 (2) 0 \u0011\u00001 \u00152\u0001 E;(0) 1 : (C.10) Here, in fact \u0000(1) and \u0000(2) have di\u000berent dimensions. However, in this esti- mation of order of magnitudes, we ignore this di\u000berence and furthermore the matrices are regarded as constants with order 1. For the moment, let us focus on the last term of \u0001 E;(0) 2 . Recall that we want to compare iV0 \u0010 \u0000 (1) 0 \u0011\u00001 \u0001 E;(0) 2 against \u0001 E;(0) 1 . Focusing only on the last term, we have iV0\u0001 E;(0) 2 \u0019 iV0\u00152\u0001E;(0)1 ; (C.11) which is much smaller than \u0001 E;(0) 1 as long as V0\u0015 2 \u001c 1. All together we have arrived at: \u0001 E;(0) 1 = \u0001T \u0010 \u0000 (1) 0 \u0011\u00001 \u0000 (1) ;T g Ex 1 + iV0 \u0010 \u0000 (1) 0 \u0011\u00001 \u0001 E;(0) 2 (C.12) 144 C.2. \u0001 C;(1) 1 from method 2 while \u0001 E;(1) 1 = \u0010 \u0000 (1) 0 \u0011\u00001 \u0014\u0000V 20 \u0010\u0000(2)0 \u0011\u00001\u0001E;(0)3 + iV0\u0001T \u0010\u0000(2)0 \u0011\u00001 \u0000(2);T gEx2 +iV0\u0015 2 \u0010 \u0000 (2) 0 \u0011\u00001 \u0001 E;(0) 1 \u0015 : (C.13) Most importantly here we see that \u0001 E;(0) 1 is multiplied by a small number \u00152V0 and then becomes a part of \u0001 E;(1) 1 . Judging from this it follows that, as long as \u00152V0 is a small number compared with t the method is very reasonable. As for the other two additional terms, they can be regarded as\u0000 V 20 g Ex 3 + V0g Ex 2 \u0001 \u0001T . Therefore, the limit of maximum value of V0, where this method is still applicable, is determined by \f\fgEx2 \f\f\u00001 or \f\fgEx3 \f\f\u0000 12 . Such limit could be much larger than 1 since roughly \f\fgExn \f\f = \f\fgEx\f\fn | smaller for larger n. This explains why this method is applicable even for V0 larger than t, as con\frmed from Fig.4.1. C.2 \u0001 C;(1) 1 from method 2 In order to estimate the accuracy of this approximation, let us assume that \u00152 and V0 are small. We de\fne \u0001 C;(0) n = g C;(0) n \u0000 gExn and \u0001C;(1)n = gC;(1)n \u0000 gExn . Again we start from the equations of the three: g C;(0) 1 , g C;(1) 1 and g Ex 1 , and then compare the three equations while ignoring certain higher-order terms such as terms which are proportional to \u00152V0. In this case, after dropping the \u0000 (1) D term and terms which are propor- tional to \u00152V0, we \fnd that g Ex 1 , g C;0 1 and respectively g C;(1) 1 satisfy the following equations: \u0000 (1) 0 g Ex 1 = iV0g Ex 2 + \u0015 2\u00170; (C.14a) \u0000 (1) 0 g C;(0) 1 = \u0015 2\u00170; (C.14b) \u0000 (1) 0 g C;(1) 1 = iV0g C;(1) 2 + \u0015 2\u00170: (C.14c) From Eq(C.14b) and Eq(C.14a) we \fnd \u0001 C;(0) 1 = \u0000iV0 \u0010 \u0000 (1) 0 \u0011\u00001 gEx2 : (C.15) 145 C.2. \u0001 C;(1) 1 from method 2 Comparing Eq(C.14c) and Eq(C.14a), we get \u0001 C;(1) 1 = iV0 \u0010 \u0000 (1) 0 \u0011\u00001 \u0010 g C;(1) 2 \u0000 gEx2 \u0011 : (C.16) Note that the magnitude of \u0001 C;(1) 2 = \u0010 g C;(1) 2 \u0000 gEx2 \u0011 is in fact smaller than the magnitude of \u0001 C;(0) 2 = \u0010 g C;(0) 2 \u0000 gEx2 \u0011 . Both involve the second equation of the hierarchy, i.e. the equation for G2. So we may analyze the latter to get an upper bound of the former. In this case, we substitute Eq(4.34) into Eq(4.8). The resulting equation will have the same structure as Eq(C.5) but every d\u000b;l and \u0016d\u000b;l are replaced respectively by D\u000b;m and D\u000b;m, and a similar substitution for D\u000b and \u0016D\u000b. Ignoring terms which are proportional to \u00152V0, it follows that g Ex 2 and g C;(0) 2 are respectively the solutions of \u0000 (2) 0 g Ex 2 = iV0g Ex 3 + \u0015 2gEx1 ; (C.17a) \u0000 (2) 0 g C;(0) 2 = \u0015 2g C;(0) 1 : (C.17b) Comparing these two equations, we \fnd that \u0001 C;(0) 2 = \u0000iV0 \u0010 \u0000 (2) 0 \u0011\u00001 gEx3 \u0000 \u00152 \u0010 \u0000 (2) 0 \u0011\u00001 \u0001 C;(0) 1 : (C.18) Focusing only on the last term, we have iV0\u0001 C;(0) 2 \u0019 \u0000iV0\u00152\u0001C;(0)1 ; (C.19) which is much smaller than \u0001 C;(0) 1 as long as V0\u0015 2 \u001c 1. We arrived at: \u0001 C;(0) 1 = \u0000iV0 \u0010 \u0000 (1) 0 \u0011\u00001 gEx2 \u0018 V0 \f\fgEx1 \f\f2 ; (C.20) and \u0001 C;(1) 1 = \u0010 \u0000 (1) 0 \u0011\u00001 \u0014 iV 20 \u0010 \u0000 (2) 0 \u0011\u00001 gEx3 + \u0015 2V0 \u0010 \u0000 (2) 0 \u0011\u00001 \u0001 C;(0) 1 \u0015 \u0018 V 20 \f\fgEx1 \f\f3 + \u00152V 20 \f\fgEx1 \f\f2 : (C.21) This agrees with the numerical results that \u0001 C;(0) 1 is proportional to V0 while \u0001 C;(1) 1 is proportional to V 2 0 . Most importantly, we see again that \u0001 C;(0) 1 is multiplied by a small number \u00152V0 and then becomes a part of \u0001 C;(1) 1 . Since 146 C.2. \u0001 C;(1) 1 from method 2 roughly \f\fgExn \f\f = \f\fgEx\f\fn, the other term, V 20 gEx3 \u0018 V 20 \f\fgEx1 \f\f3, is also much smaller than \u0001 C;(0) 1 \u0018 V0 \f\fgEx1 \f\f2. However, for large enough V0 the other approximation used in this method, the perturbational expansion of the operators m\u0302, becomes invalid. Therefore, as long as V0 is a small number compared to t the method is very reasonable. It should be noted that this method is capable of dealing with large systems since it does not require a direct diagonalization of a 2N -dimension matrix. 147 Appendix D Coherent states and coherent-state representation To provide a convenient reference, in this appendix, we brie\ry describe the general idea of coherent-state representation. More details can be found in for example Ref. [70, 71]. Here we will only discuss bosonic coherent states. Fermionic ones can be de\fned similarly with the usual complex numbers replaced by Grassman numbers[71]. In the xD.1, coherent states for single-mode bosonic systems is intro- duced and the simplest P -representation is discussed. In the xD.2, we discuss correspondingly the coherent-state representations for multi-mode bosonic systems. We also brie\ry review the more general positive P -representation. D.1 Coherent states for systems with a single mode A single-mode bosonic system is de\fned by a Hamiltonian H \u0000 a; ay \u0001 , where a; ay are respectively annihilation and creation operators of a particle with the particular mode. Its Hilbert space, expanded in occupation-number representation, is fjnig, where n = 0; 1; 2; \u0001 \u0001 \u0001 . A coherent state j\u0018i is de\fned to be an eigenstate of annihilation oper- ator, a j\u0018i = \u0018 j\u0018i (D.1) with eigenvalue \u0018. Operator a is not a hermitian operator so \u0018 is not a real but a complex number (c-number). Let us de\fne a displacement operator D (\u0018) = e\u0018a y\u0000\u0018\u0003a; (D.2) which equals to D (\u0018) = e\u0000 j\u0018j2 2 e\u0018a y e\u0000\u0018 \u0003a; (D.3) 148 D.1. Coherent states for systems with a single mode utilizing the Baker-Campbell-Hausdor\u000b formula[70]. From there one can verify that j\u0018i = D (\u0018) j0i ; (D.4) where j0i is the vacuum state. This can be seen in the following. Let us express D (\u0018) j0i in terms occupation-number states jni and note jni = (a y) n p n! j0i, D (\u0018) j0i = e\u0000 j\u0018j 2 2 e\u0018a y j0i = e\u0000 j\u0018j2 2 1X n=0 \u0018n n! \u0010 ay \u0011n j0i = e\u0000 j\u0018j2 2 1X n=0 \u0018np n! jni : (D.5) Therefore aD (\u0018) j0i = e\u0000 j\u0018j 2 2 1X n=0 \u0018np (n\u0000 1)! jn\u0000 1i = \u0018e\u0000 j\u0018j2 2 1X m=0 \u0018mp (m)! jmi = \u0018D (\u0018) j0i : (D.6) Inner product between two coherent states j\u0018i and j\u0011i is, h\u0018 j \u0011i = e\u0018\u0003\u0011\u0000 j\u0018j 2 2 \u0000 j\u0011j2 2 ; (D.7) so jh\u0018 j \u0011ij2 = e\u0000j\u0018\u0000\u0011j2 ; (D.8) which is close to zero when \u0018 is very di\u000berent from \u0011. This implies that although no pair of the coherent states are orthogonal to each other but the overlap tends to be small if their corresponding eigenvalues are very di\u000berent. This suggests that maybe the set of coherent states can be used as a basis to expand operators and vectors. Due to the non-orthogonal relation, such forms of basis are not unique. One particular convenient and important choice is I = 1 \u0019 Z d2\u0018 j\u0018i h\u0018j : (D.9) 149 D.1. Coherent states for systems with a single mode This can be veri\fed by considering arbitrary jni such that jni = 1 \u0019 Z d2\u0018 j\u0018i h\u0018 jni ; (D.10) which becomes jni = X m 1 \u0019 Z d2\u0018 jmi hm j\u0018i h\u0018 jni = X m jmi 1 \u0019 Z d2\u0018e\u0000j\u0018j 2 \u0018mp (m)! (\u0018\u0003)np (n)! = X m jmi \u000emn = jni : (D.11) Here we have used 1 \u0019 Z d2\u0018e\u0000j\u0018j 2 \u0018mp (m)! (\u0018\u0003)np (n)! = \u000emn; (D.12) which is 1 only when m = n and zero otherwise. With this basis, a density matrix \u001a (or a more general physical observable operator A) can always be expanded as[126] \u001a = 1 \u00192 ZZ d2\u0018d2\u0011e\u0000 j\u0018j2 2 \u0000 j\u0011j2 2 R (\u0018\u0003; \u0011) j\u0018i h\u0011j : (D.13) where R (\u0018\u0003; \u0011) = e j\u0018j2 2 + j\u0011j2 2 h\u0018j \u001a j\u0011i : (D.14) This distribution function R (\u0018\u0003; \u0011) holds generally for all \u001a and it is called the R representation. However, since the overcompleteness and the non- uniqueness of the coherent-state representation, there are other expansions, which hold not for all but a large class of density matrices[70]. For exam- ple, the P representation and the Positive P representation are respectively de\fned as \u001a = Z d2\u0018P (\u0018\u0003; \u0018) j\u0018i h\u0018j ; (D.15) and \u001a = ZZ d2\u0018d2\u0011P (\u0018; \u0011) e j\u0018j2 2 + j\u0011j2 2 \u0000\u0018\u0011 j\u0018i h\u0011\u0003j : (D.16) 150 D.2. Coherent states for multi-mode systems Next we convert operator a and ay to be operators acting on such distribution function P , in for example, the P representation. For a\u001a, a\u001a = Z d2\u0018P (\u0018\u0003; \u0018) a j\u0018i h\u0018j = Z d2\u0018\u0018P (\u0018\u0003; \u0018) j\u0018i h\u0018j ; (D.17) so a! \u0018. For ay\u001a, noticing ay j\u0018i = dd\u0018 j\u0018i ay\u001a = Z d2\u0018P (\u0018\u0003; \u0018) \u0012 d d\u0018 j\u0018i \u0013 h\u0018j = Z d2\u0018 \u0012 \u0018\u0003 \u0000 d d\u0018 \u0013 P (\u0018\u0003; \u0018) j\u0018i h\u0018j ; (D.18) so ay ! \u0018\u0003 \u0000 dd\u0018 . Similarly we need to consider also \u001aa and \u001aay. Summarize all those operators, we have a\u001a$ \u0018P (\u0018\u0003; \u0018) ; \u001aay $ \u0018\u0003P (\u0018\u0003; \u0018) ; ay\u001a$ \u0012 \u0018\u0003 \u0000 @ @\u0018 \u0013 P (\u0018\u0003; \u0018) ; \u001aa$ \u0012 \u0018 \u0000 @ @\u0018\u0003 \u0013 P (\u0018\u0003; \u0018) : (D.19) D.2 Coherent states for multi-mode systems For a multi-mode (countably many, denoted as N) system, a coherent basis can be de\fned similarly as\f\f\f~\u0018E = D \u0010~\u0018\u0011 j0i = e~\u0018\u0001~ay\u0000~\u0018\u0003\u0001~a j0i = e\u0000j~\u0018j22 e~\u0018\u0001~ay j0i ; (D.20) where ~\u0018 = (\u00181; \u00182; \u0001 \u0001 \u0001 ; \u0018N )T and ~ay = \u0010 ay1; a y 2; \u0001 \u0001 \u0001 ayN \u0011T . ayl is the creation operator at site l, \u0018l is a c number and j0i is the vacuum state. Coherent states are over complete. One convenient choice of representation is the P -representation[70], which decomposes the identity as I = 1 \u0019 Z d2~\u0018 \f\f\f~\u0018ED~\u0018\f\f\f ; (D.21) 151 D.2. Coherent states for multi-mode systems and it maps a density matrix \u001a (a physical quantityA) to a function P \u0010 ~\u0018\u0003; ~\u0018 \u0011 (A \u0010 ~\u0018\u0003; ~\u0018 \u0011 ) such that \u001a = Z d2~\u0018P \u0010 ~\u0018\u0003; ~\u0018 \u0011 \f\f\f~\u0018ED~\u0018\f\f\f ; (D.22) hAi = Z d2~\u0018P \u0010 ~\u0018\u0003; ~\u0018 \u0011 A \u0010 ~\u0018\u0003; ~\u0018 \u0011 : (D.23) If A is in normal order of al; a y l . In this representation, operators a; ay become di\u000berential operators on P \u0010 ~\u0018\u0003; ~\u0018 \u0011 , al\u001a$ \u0018lP \u0010 ~\u0018\u0003; ~\u0018 \u0011 ; \u001aayl $ \u0018\u0003l P \u0010 ~\u0018\u0003; ~\u0018 \u0011 ; ayl \u001a$ \u0012 \u0018\u0003l \u0000 @ @\u0018l \u0013 P \u0010 ~\u0018\u0003; ~\u0018 \u0011 ; \u001aal $ \u0012 \u0018l \u0000 @ @\u0018\u0003l \u0013 P \u0010 ~\u0018\u0003; ~\u0018 \u0011 : (D.24) In this way, the operator-form QME can be mapped to a di\u000berential equa- tion. 152 Appendix E Fokker-Planck equation and Langevin equation A standard Fokker-Planck equation with time-independent coe\u000ecients reads, @ @t P \u0010 ~\u0018; t \u0011 = 24\u0000X j @ @\u0018j Aj \u0010 ~\u0018 \u0011 + 1 2 X i;j @2 @\u0018i@\u0018j Dij \u0010 ~\u0018 \u001135P \u0010~\u0018; t\u0011 : (E.1) When there are higher-order derivative terms besides the above \frst two terms, it is called a generalized Fokker-Planck equation. The Pawula theorem[98] states that in order to keep the distribution function (P \u0010 ~\u0018; t \u0011 ) non-negative, there has to be either only 2 terms or in\fnity terms. In the later case, unless the equation is exactly solvable usually the generalized Fokker-Planck equa- tion is truncated \frst and then solved. However, while truncation at the second order keeps the positiveness, a further truncation beyond the second order which destroys such positiveness in small regions, could result a more accurate distribution function[98]. In this appendix we discuss solutions of the Fokker-Planck equation and the generalized Fokker-Planck equation. Note here ~\u0018 is a vector of real random variables. It is not directly applicable for the generalized Fokker- Planck equation in c-numbers we derived in the previous chapters. E.1 Langevin equation for standard Fokker-Planck equations There are several approaches to solve the standard Fokker-Planck equation besides brute-force numerical solution in the whole space of n ~\u0018 o . Convert- ing it to a set of equivalent Langevin equation[98] with white noise and then numerically simulate the Langevin equation is a common and general method. A standard Langevin equation states, _\u0018i = hi \u0010 ~\u0018 \u0011 + gij \u0010 ~\u0018 \u0011 wj (t) ; (E.2) 153 E.1. Langevin equation for standard FPEs where we only consider the time-independent coe\u000ecientsAj \u0010 ~\u0018 \u0011 andDij \u0010 ~\u0018 \u0011 , and ~w (t) is a white noise, hwi (t)i = 0; hwi (t)wj \u0000 t0 \u0001i = \u000eij\u000e \u0000t\u0000 t0\u0001 : (E.3) Using Ito\u0302 calculus[98], one can de\fne a correspondence between the standard Fokker-Planck equation and the Langevin equation. Assuming Dij \u0010 ~\u0018 \u0011 is positive semide\fnite such that there is a matrix G satisfying D \u0010 ~\u0018 \u0011 = G \u0010 ~\u0018 \u0011 Gy \u0010 ~\u0018 \u0011 ; (E.4) then gij \u0010 ~\u0018 \u0011 = Gij \u0010 ~\u0018 \u0011 ; (E.5) hi \u0010 ~\u0018 \u0011 = Ai \u0010 ~\u0018 \u0011 : (E.6) In numerical simulation, one works with directly d\u0018i = hi \u0010 ~\u0018 \u0011 dt+ gij \u0010 ~\u0018 \u0011 dWj (t) ; (E.7) where hdWi (t)i = 0; hdWi (t) dWj \u0000 t0 \u0001i = \u000eijdt\u000et;t0 : (E.8) Here \u000et;t0 is the Kronecker \u000e not the Dirac \u000e. Another way to solve the Fokker-Planck equation is to convert it into equations of correlations by considering for example, hxli and hxlxmi etc.. Let us work out the \frst derivative term for the above one-variable average, d dt h\u0018li = \u0000h\u0018l X j @ @\u0018j Aii = \u0000h X j @ @\u0018j \u0018lAii+ hAi X j @ @\u0018j \u0018li = hAli: (E.9) Here we have used the following property: for arbitrary F \u0010 ~\u0018 \u0011 , as long as lim\u0018j\u0000!1 F \u0010 ~\u0018 \u0011 does not diverge, we have h @ @\u0018j F \u0010 ~\u0018 \u0011 i = Z d2~\u0018 @ @\u0018j h F \u0010 ~\u0018 \u0011 P \u0010 ~\u0018 \u0011i = 0: (E.10) In this way, one can derive equations for correlations of any number of variables. However, then one has to solve such possibly coupled equations if many-variable correlations are of interest. 154 E.2. Analytical solution for the Ornstein-Uhlenbeck process E.2 Analytical solution for the Ornstein-Uhlenbeck process For a special case { the Ornstein-Uhlenbeck process, there is an analytical solution. An Ornstein-Uhlenbeck process has a constant di\u000busion matrix D and a linear drift term ~A = \u0000~\u0018. We further require \u0000 is positive de\fnite in or- der to have a non-divergent distribution function. Although time-dependent solution is also possible, here we focus on stationary solution, P = (2\u0019)\u0000 N 2 (Det\u001b)\u0000 1 2 e\u0000 1 2 ~\u0018T \u001b\u00001~\u0018; (E.11) where \u001b = X l;m 1 \u0014l + \u0014\u0003m D vl jD j vm E julihumj ; (E.12) and \u0000 = X l \u0014l \f\f\fulihvl\f\f\f : (E.13) The last expression de\fnes vl \f\f (juli) as the left (right) eigenvector of \u0000 with eigenvalue \u0014l. Out of all above approaches, only the simulation via Langevin equation is generally applicable, even for generalized Fokker-Planck equa- tions. E.3 Stochastic di\u000berence equation for truncated generalized Fokker-Planck equations A generalized Fokker-Planck equation with time-independent coe\u000ecients truncated at the third-order derivatives reads, @ @t P \u0010 ~\u0018 \u0011 = 24\u0000X j @ @\u0018j Aj \u0010 ~\u0018 \u0011 + 1 2 X i;j @2 @\u0018i@\u0018j Dij \u0010 ~\u0018 \u0011 + 1 6 X i;j;k @3 @\u0018i@\u0018j@\u0018k Dijk \u0010 ~\u0018 \u001135P \u0010~\u0018\u0011 : (E.14) Now let us solutions of this generalized Fokker-Planck equation. For such generalized Fokker-Planck equation an equivalent stochastic di\u000berence equation[102] for numerical molecular dynamical simulations can be found. In that case, usually the positive P -representation[72] or the gauge P -representation is 155 E.3. Stochastic di\u000berence equation for truncated GFPEs used. The basic idea is to use proper random variables to mimic the e\u000bect of the third order term just like what was done for the second derivative terms. For example, @ @\u0018i Ai ! h\u0001\u0018ii = Ai\u0001T; (E.15) @2 @\u0018i@\u0018j Dij ! h\u0001\u0018i\u0001\u0018ji = Dij\u0001T; (E.16) @3 @\u0018i@\u0018j@\u0018k Dijk ! h\u0001\u0018i\u0001\u0018j\u0001\u0018ki = Dijk\u0001T: (E.17) Therefore, there will be terms like \u0011ijk (\u0001T ) 1 3 appearing in the stochastic di\u000berence equation to make sure the above average holds, \u0001\u0018i = hi\u0001t+ gi (\u0001t) 1 2 + ki (\u0001T ) 1 3 : (E.18) just like gi happens to be gijwj such that Eq. (E.16) is satis\fed, ki has to be de\fned properly to obey Eq. (E.17). Generally relation between ki and Dijk is still missing, but speci\fc forms for example equations can be found in Ref.[102]. 156 Appendix F Solving the Red\feld equation and the local-operator Lindblad equation of bosonic systems with BBGKY-like hierarchy The BBGKY-like method is based on the equation of motion of the Green's functions. The method starts from turning the Red\feld equation, Eq. (4.2) into equation of motion of Green's functions, which is de\fned by single- particle Green's function G \u0000 ly;n \u0001 = D ayl an\u001a (1) E (G1 for short), two-article Green's function G \u0000 iy; jy; l; n \u0001 = D ayia y jalan\u001a (1) E (G2 for short) and so on. Here again the density matrix \u001a (1) refers to a stationary state and it is de\fned by L\u001a (1) = 0. Then generally for interacting systems the stationary equation becomes an equation hierarchy, in the \frst order of which are equations for G1. But such G1 is coupled to G2. The next order in the hierarchy are equations for G2 and such G2 is coupled to G1 and G3, and so on. The hierarchy can be truncated and solved if one takes certain further approximations such as cluster expansion[101]. Since the di\u000berence between the Red\feld equation and the local-operator Lindblad equation is only at the bath operators, the following calculation is in fact valid for both equations. Of course, all those coe\u000ecients D\r;m should then be substituted properly for the local-operator Lindblad equation. The operator-form Red\feld equation reads @\u001a @t = LH0\u001a+ ULV \u001a+ \u00152L0B\u001a+ \u00152UL1B\u001a; (F.1) 157 F.1. Exact solution on non-interacting systems where LH0\u001a = \u0000i N\u00001X l=1 h ayl al+1 + a y l+1al; \u001a i (F.2a) LV \u001a = \u0000i 2 NX l=1 h ayl al \u0010 ayl al \u0000 1 \u0011 ; \u001a i (F.2b) L(0)B \u001a = \u0000 X \r=L;R n D\r;m h ay\r ; am\u001a i + \u0016D\r;m h a\r ; a y m\u001a i + h:c: o (F.2c) L(1)B \u001a = \u0000 X \r;m1;m2;m3 D\r;m1m2m3 nh ay\r ; a y m1am2am3\u001a i + h a\r ; a y m3a y m2am1\u001a i + h:c: o : (F.2d) Let us now transform the above Red\feld equation in the operator form into equations of Green's functions by multiplying ayl an at the RHS of Eq.(3.30) and then performing a trace, 0 = \u0000ihayl an+1i \u0000 ihayl an\u00001i+ ihayl+1ani+ ihayl\u00001ani (F.3a) +\u00152 X \r;m \u0000 \u0016D\r;m \u0000D\r;m \u0001 h \u000en\rhayl ami+ \u000el\rhaymani i (F.3b) +\u00152 X \u0000 \u0016D\r;n\u000el\r + \u0016D\r;l\u000en \u0001 (F.3c) +iUhayl ayl alani \u0000 iUhayl aynanani (F.3d) +\u00152U X \r;mi D\r;m1m2m3 h \u000en \u0010 haym1am2i\u000elm3 + haym1am3i\u000elm2 \u0011 +\u000el \u0010 haym2am1i\u000enm3 + haym3am1i\u000enm2 \u0011i : (F.3e) F.1 Exact solution on non-interacting systems Note that when U = 0 only Eq. (F.3a), Eq. (F.3b) and Eq. (F.3c) are there so that this equation is closed. In this section we consider only the U = 0 case. This is an N2-dimension linear system for unknown G (0) 1 , where (0) refers to U = 0, \u0000H0g (0) 1 + \u0015 2\u0000B0g (0) 1 = \u0015 2\u0017(0): (F.4) Here we arrange all unknownG1 into a vector g1 = (G1 (1; 1) ; G1 (1; 2) ; \u0001 \u0001 \u0001 ; )T . Elements of the two matrices and the vector can be read o\u000b from Eq. (4.7). 158 F.2. Approximate solution on interacting systems For example, (\u0000H0)l\u0003N+n;l\u0003N+n+1 = \u0000i; (F.5) and when l = 1 (or l = N , or n = 1; N), \u0017 (0) 1\u0003N+n = \u0015 2 \u0016D1;n; (F.6) otherwise, it is zero. Solving this N2-dimension linear system we \fnd all G1. F.2 Approximate solution on interacting systems Note that when U 6= 0 this equation is not closed. This means that one need to consider the next order equation in terms of G2 together to solve the whole Eq.(3.30). One can derive such equation by multiplying ayia y jalan and then performing a trace. In fact, if one do so, one will \fnd that equations of G2 is coupled to G3 and so on[101]. Therefore, one has to solve the whole hierarchy to solve Eq.(3.30) exactly, which is as hard as dealing with directly Eq. (3.30). So we have to come up with some further approximation to truncate the hierarchy and then solve it. In the present work, we will truncate the hierarchy at the \frst order by using the cluster expansion[101], hayl ayl alani = 2hayl alihayl ani+ o (U) : (F.7) So we will do not need explicitly the second order equations here. After taking the approximation of the cluster expansion, Eq. (F.3d) becomes, iU2hayl alihayl ani \u0000 iU2hayl anihaynani: (F.8) Therefore, Eq. (F.3) becomes a closed equation, \u0000H0g (1) 1 + \u0015 2\u0000B0g (1) 1 + \u0015 2U\u0000B1g (1) 1 = \u0015 2\u0017(0) + iU\u0019 \u0010 g (1) 1 \u0011 ; (F.9) where \u0019 \u0010 g (1) 1 \u0011 l\u0003N+n = 2 \u0010 g (1) 1 \u0011 l\u0003N+n \u0010 g (1) 1 \u0011 n\u0003N+n \u00002 \u0010 g (1) 1 \u0011 l\u0003N+l \u0010 g (1) 1 \u0011 l\u0003N+n : (F.10) 159 F.2. Approximate solution on interacting systems This equation can be solved iteratively, g [n+1] 1 = \u0010 \u0000H0 + \u0015 2\u0000 (0) D + \u0015 2U\u0000 (1) D \u0011\u00001 h \u00152\u0017(0) + iU\u0005 \u0010 g [n] 1 \u0011i ; (F.11) with initial solution g[0] = g (0) 1 for the corresponding non-interacting system. Then the solution is g (1) 1 = limn!1 g [n] 1 ; (F.12) which in practice stops at large enough n such that g [n] 1 \u0000 g[n\u00001]1 is small enough. 160","type":"literal","lang":"en"}],"http:\/\/www.europeana.eu\/schemas\/edm\/hasType":[{"value":"Thesis\/Dissertation","type":"literal","lang":"en"}],"http:\/\/vivoweb.org\/ontology\/core#dateIssued":[{"value":"2011-05","type":"literal","lang":"en"}],"http:\/\/www.europeana.eu\/schemas\/edm\/isShownAt":[{"value":"10.14288\/1.0071817","type":"literal","lang":"en"}],"http:\/\/purl.org\/dc\/terms\/language":[{"value":"eng","type":"literal","lang":"en"}],"https:\/\/open.library.ubc.ca\/terms#degreeDiscipline":[{"value":"Physics","type":"literal","lang":"en"}],"http:\/\/www.europeana.eu\/schemas\/edm\/provider":[{"value":"Vancouver : University of British Columbia Library","type":"literal","lang":"en"}],"http:\/\/purl.org\/dc\/terms\/publisher":[{"value":"University of British Columbia","type":"literal","lang":"en"}],"http:\/\/purl.org\/dc\/terms\/rights":[{"value":"Attribution-NonCommercial-NoDerivs 3.0 Unported","type":"literal","lang":"en"}],"https:\/\/open.library.ubc.ca\/terms#rightsURI":[{"value":"http:\/\/creativecommons.org\/licenses\/by-nc-nd\/3.0\/","type":"literal","lang":"en"}],"https:\/\/open.library.ubc.ca\/terms#scholarLevel":[{"value":"Graduate","type":"literal","lang":"en"}],"http:\/\/purl.org\/dc\/terms\/title":[{"value":"Quantum transport through open systems","type":"literal","lang":"en"}],"http:\/\/purl.org\/dc\/terms\/type":[{"value":"Text","type":"literal","lang":"en"}],"https:\/\/open.library.ubc.ca\/terms#identifierURI":[{"value":"http:\/\/hdl.handle.net\/2429\/33955","type":"literal","lang":"en"}]}}