BOSE-EINSTEIN CONDENSATION IN TWO-DIMENSIONAL TRAPS A Dissertation Presented by JUAN PABLO FERNÁNDEZ Submitted to the Graduate School of the University of Massachusetts Amherst in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY February 2004 Department of Physics c Copyright by Juan Pablo Fernández 2004 ° All Rights Reserved BOSE-EINSTEIN CONDENSATION IN TWO-DIMENSIONAL TRAPS A Dissertation Presented by JUAN PABLO FERNÁNDEZ Approved as to style and content by: William J. Mullin, Chair Suzan Edwards, Member Robert B. Hallock, Member Jonathan L. Machta, Member Jonathan L. Machta, Department Chair Department of Physics To the memory of Gonzalo Cortés González, César Pretol Castillo, and Jacob Ketchakeu ACKNOWLEDGMENTS On my very first semester at UMass I had the pleasure of attending Bill Mullin’s lectures on classical mechanics, and from the beginning I noticed his excellence as a teacher, his willingness to never stop learning, his kindness, and his modesty. That first impression was only strengthened during the five years or so that we worked together. I will forever be grateful to Bill for always putting aside whatever he was doing and listening to anything I had to say, for convincing me that my work was valuable, and for teaching me a great deal of physics. Having three outstanding scientists in one’s dissertation committee is an honor, but also, given their example and their standards, an enormous challenge. My fear of disappointing Suzan Edwards, Bob Hallock, and Jon Machta forced me to do my best, and for that alone they would deserve my thanks. My debt to each of them, however, goes far beyond: Suzan introduced me to Matlab, the programming language that I used to obtain the great majority of the results that I present here, and Jon lent me the computers in which I ran those programs and wrote this thesis; Bob lent me his support during difficult periods and gave me much valuable advice in his office across the hall. David Hall made it possible for me to at last observe a condensate and witness the phenomenon that I had been thinking about for so many years; besides, he generously provided me with the very first figure of the thesis. Eugene Zaremba and Markus Holzmann kindly let me reproduce figures from their published papers (Prof. Zaremba’s figure, regrettably, does not appear in this final version, but can still be admired in Ref. 91); Prof. Holzmann, moreover, lent me the program he wrote for Ref. 69, and by doing so helped me get seriously started on my research. I also benefited greatly from conversations and correspondence with Panayotis Kevrekidis and Brandon van Zyl. Werner Krauth wrote the Monte Carlo code that, with some adjustments, produced the results that I report in the last chapter. Stefan Heinrichs taught me how to run, and extract information from, Prof. Krauth’s program. Justin Herrmann and Tim Middelkoop allowed me to run the simulations in their computer cluster. The Department of Astronomy at UMass granted me access to its computational facilities during the initial stages of this work. Finally, I would like to thank my family, friends, and colleagues, in Amherst, in New Hampshire, in Vermont, and in Colombia, for their warmth, their companionship, and their sense of humor. v ABSTRACT BOSE-EINSTEIN CONDENSATION IN TWO-DIMENSIONAL TRAPS FEBRUARY 2004 JUAN PABLO FERNÁNDEZ Fı́sico, UNIVERSIDAD DE LOS ANDES Ph.D., UNIVERSITY OF MASSACHUSETTS AMHERST Directed by: Professor William J. Mullin The fact that two-dimensional interacting trapped systems do not undergo Bose-Einstein Condensation (BEC) in the thermodynamic limit, though rigorously proved, is somewhat mysterious because all relevant limiting cases (zero temperature, small atom numbers, noninteracting particles) suggest otherwise. We study the possibility of condensation in finite trapped two-dimensional systems. We first consider the ideal gas, which incorporates the inhomogeneity and finite size of experimental systems and can be solved exactly. A semiclassical self-consistent approximation gives us a feel for the temperature scales; diagonalization of the one-body density matrix confirms that the condensation is into a single state. We squeeze a three-dimensional system and see how it crosses over into two dimensions. Mean-field theory, our main tool for the study of interacting systems, prescribes coupled equations for the condensate and the thermal cloud: the condensate receives a full quantum-mechanical treatment, while the noncondensate is described by different schemes of varying sophistication. We analyze the T = 0 case and its approach to the thermodynamic limit, finding a criterion for the dimensionality crossover and obtaining the coupling constant of the two-dimensional system that results from squeezing a three-dimensional trap. We next apply a semiclassical Hartree-Fock approximation to purely two-dimensional finite gases and find that they can be described either with or without a condensate; this continues to be true in the thermodynamic limit. The condensed solutions have a lower free energy at all temperatures vi but neglect the presence of phonons within the system and cease to exist when we allow for this possibility. The uncondensed solutions, in turn, are valid under a more rigorous scheme but have consistency problems of their own. Path-integral Monte Carlo simulations provide an essentially exact description of finite interacting gases and enable us to study highly anisotropic systems at finite temperature. We find that our two-dimensional Hartree-Fock solutions accurately mimic the surface density profile and predict the condensate fraction of these systems; the equivalent interaction parameter is smaller than that dictated by the T = 0 analysis. We conclude that, in two-dimensional isotropic finite trapped systems and in highly compressed three-dimensional gases, there is a phenomenon resembling a condensation into a single state. vii TABLE OF CONTENTS Page ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . LIST OF FIGURES x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi CHAPTER 1. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 1.2 1.3 The creation myth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 BEC, two dimensions, and traps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 This thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2. THE IDEAL TRAPPED BOSE GAS . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.1 2.2 2.3 2.4 2.5 2.6 Introduction . . . . . . . . . . . . . . . . . . . . Exact results for the ideal gas . . . . . . . . . . The semiclassical approximation . . . . . . . . The off-diagonal elements of the density matrix Effects of anisotropy . . . . . . . . . . . . . . . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 15 21 26 29 35 3. MEAN-FIELD THEORY OF INTERACTING SYSTEMS . . . . . . . . . . . . . . 36 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . The effective interaction . . . . . . . . . . . . . . . . . . . . The Hartree-Fock-Bogoliubov equations . . . . . . . . . . . The Gross-Pitaevskiı̆ equation and the Thomas-Fermi limit Anisotropic systems at zero temperature . . . . . . . . . . . Finite temperatures: The semiclassical approximation . . . The interacting Bose gas in the Hartree-Fock approximation Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 37 38 42 46 49 51 61 4. THE TWO-DIMENSIONAL BOSE-EINSTEIN CONDENSATE . . . . . . . . . . 62 4.1 4.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 4.2.1 4.2.2 4.2.3 The Hartree-Fock-Bogoliubov equations . . . . . . . . . . . . . . . . . . . . . . 66 The thermodynamic limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 The free energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 viii 4.3 4.4 4.5 4.6 Numerical methods and results . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . Appendix: The Hartree-Fock excess energy Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 76 76 77 5. PATH-INTEGRAL MONTE CARLO AND THE SQUEEZED INTERACTING BOSE GAS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 5.1 5.2 5.3 5.4 5.5 5.6 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . Path integrals in statistical mechanics . . . . . . . . . . The Monte Carlo method and the Metropolis algorithm An algorithm for PIMC simulations of trapped bosons . The anisotropic interacting gas at finite temperature . . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 6. CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 80 82 89 97 106 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 APPENDICES A. MATHEMATICAL VADEMECUM . . . . . . . . . . . . . . . . . . . . . . . B. PERMUTATION CYCLES AND WICK’S THEOREM . . . . . . . . . . . C. SPECTRAL DIFFERENTIATION AND GAUSSIAN QUADRATURE D. MATHEMATICAL DETAILS OF PIMC SIMULATIONS . . . . . . . . . . . . . . . . . . . . . . . . . 115 118 123 128 BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 ix LIST OF TABLES Table Page 1.1 Summary of methods used and systems studied in this thesis . . . . . . . . . . . . . . 11 3.1 Typical temperatures and frequencies found in experiments . . . . . . . . . . . . . . . 49 x LIST OF FIGURES Figure Page 1.1 Experimental realization of Bose-Einstein condensation . . . . . . . . . . . . . . . . . .2 2.1 Density and number density of a condensed Bose gas . . . . . . . . . . . . . . . . . . . 19 2.2 The growing condensate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.3 Front view of Fig. 2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.4 Condensate fraction of a two-dimensional ideal trapped Bose gas . . . . . . . . . . . . 21 2.5 Chemical potential of a two-dimensional ideal trapped Bose gas . . . . . . . . . . . . . 22 2.6 Position of the noncondensate maximum . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.7 Condensate fraction of a two-dimensional ideal trapped Bose gas in the semiclassical approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.8 Density and number density of a condensed 2D gas in the semiclassical approximation . . 25 2.9 Eigenvalues of the two-dimensional isotropic trapped-ideal-gas density matrix . . . . . 28 2.10 The first seven radially symmetric eigenfunctions (including the ground state) of the two-dimensional isotropic harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . . 30 2.11 Condensate fraction of a three-dimensional ideal gas in a trap of increasing anisotropy . . 32 2.12 Surface density and surface number density of a three-dimensional ideal Bose gas in a trap of increasing anisotropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 3.1 Wavefunctions of two-dimensional isotropic trapped Bose gases at zero temperature . . 44 3.2 Typical forces exerted on an atom in a trap of increasing anisotropy . . . . . . . . . . 46 3.3 Condensate fraction of a three-dimensional Bose gas in the Hartree-Fock approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.4 Failure of the Hartree-Fock approximation at high temperatures 3.5 Chemical potential of a three-dimensional interacting Bose gas in the Hartree-Fock approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 3.6 Two views of the density of an interacting 3D trapped Bose gas at various temperatures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 xi . . . . . . . . . . . . 55 3.7 Number density of an interacting 3D trapped Bose gas at various temperatures . . . . 58 3.8 Densities and number densities of a two-dimensional isotropic ideal gas and the equivalent interacting system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 3.9 Eigenvalues of the one-body density matrix of ideal and interacting two-dimensional trapped gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 3.10 Ground state and azimuthally symmetric excited eigenfunctions of a two-dimensional isotropic interacting trapped gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 4.1 Condensate and noncondensate density profiles of a two-dimensional Bose-Einstein gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4.2 Total density profiles of a two-dimensional gas . . . . . . . . . . . . . . . . . . . . . . . 74 4.3 Free energy per particle for the condensed and uncondensed solutions . . . . . . . . . 75 5.1 The density matrix as a path integral . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 5.2 A simple game that illustrates the Metropolis algorithm . . . . . . . . . . . . . . . . . 86 5.3 A variation on the game of Fig. 5.2 that shows the effect of choosing better a priori probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 5.4 Boxes and interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 5.5 The two steps involved in moving part of a cycle . . . . . . . . . . . . . . . . . . . . . 94 5.6 Moving a complete cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 5.7 Interactions between particles of changing identity . . . . . . . . . . . . . . . . . . . . 94 5.8 Monte Carlo and exact number densities of an ideal two-dimensional trapped gas . . . 98 5.9 Extracting the condensate number from a PIMC simulation . . . . . . . . . . . . . . . 98 5.10 Surface number density of a condensed ideal Bose gas of varying anisotropy . . . . . 101 5.11 Front view of the rightmost plot of Fig. 5.10 . . . . . . . . . . . . . . . . . . . . . . . 101 5.12 Finding the condensate fraction of an ideal Bose gas of varying anisotropy . . . . . . 102 5.13 Aspect ratios, obtained by Monte Carlo simulation, of condensed ideal and interacting Bose gases of increasing anisotropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 5.14 Monte Carlo and mean-field number densities of a three-dimensional trapped gas . . 103 5.15 Monte Carlo number surface densities of a three-dimensional trapped gas of increasing anisotropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 xii 5.16 Monte Carlo number surface density and best-fit two-dimensional profile of an interacting three-dimensional Bose gas in a highly anisotropic trap . . . . . . . . . . . . . 105 5.17 Condensate fraction of a quasi-2D interacting Bose gas . . . . . . . . . . . . . . . . . 107 5.18 Occupation numbers and eigenfunctions of the one-body density matrix for a quasi2D Bose gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 D.1 The Lévy construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 xiii CHAPTER 1 INTRODUCTION We were experienced enough to know that when results in experimental physics seem too good to be true, they almost always are! —E. A. Cornell and C. E. Wieman [1] 1.1 The creation myth It happened, appropriately enough, in the early morning. At 6:04:29 a.m. on 29 September 1995, at W. Ketterle’s laboratory in Building 26 of the Massachusetts Institute of Technology (MIT), postdoc N. J. van Druten scrawled “BEC” at the very bottom of a page of a lab notebook and followed it with an exclamation point that was taller than the acronym, and almost as wide. “BEC” stood for Bose-Einstein condensation, the topic of this thesis, and Ketterle’s group had just created the first condensate of sodium atoms [2,3]. For the preceding five years, Ketterle and his collaborators had been repeating and perfecting a procedure that is now followed all over the world: they would i) produce a vapor of alkali atoms; ii) load it into a magneto-optical trap (MOT); iii) cool the vapor to microkelvin temperatures using lasers that, when tuned to an appropriate transition frequency, exerted a strong damping force on the atoms; iv) transfer the resulting “optical molasses” into a magnetic trap; v) cool the molasses down to nanokelvin temperatures by applying a radiofrequency field that, by inducing spin-flip transitions into untrapped states, encouraged the most energetic atoms to evaporate and let the remaining ones collide until they achieved equilibrium at a lower energy; and vi) turn off the trap and use a CCD camera to photograph the expansion of the liberated cloud and capture it on a screen. 1 1 This oversimplified account of the experiments does no justice to the ingenuity and effort required to bring atoms to the densities and temperatures necessary for BEC by harnessing their complexity. The interested reader may turn to Ref. 4, which includes a complete collection of references on the experimental aspects of BEC, or to the Nobel lectures just cited. Our wording above is similar to that used in a semipopular article [5], according to which “the possibility of seeing the long-sought Bose-Einstein condensation” had been brought “tantalizingly close” by recent developments. The article appeared three days after the first condensate was created. 1 Figure 1.1. Experimental realization of Bose-Einstein condensation. Condensates like the one shown here are routinely produced in D. S. Hall’s TOP trap at Amherst College. In Prof. Hall’s words, “This particular condensate was released from our trap and allowed to expand for about 16 ms before we took the image, and probably contains somewhere around 300,000–400,000 atoms. That image is 5 microns per pixel, and it came out of a cylindrically symmetric trap that was 90 Hz (axial) by 32 Hz (radial). The axial direction is up-down in the image—and is the direction of greatest expansion upon release.” Now Ketterle, van Druten, and the rest were seeing one of the pixellated, false-color images that since that year have graced all sorts of science-related posters, calendars, and textbook covers: out of a surface resembling a three-dimensional plot of a Gaussian there emerged a narrow, tall peak located right at the center. Similar images taken at higher temperatures had shown only the Gaussian, and lowering the temperature further yielded images where there was the peak and little else. Moreover, the peak did not appear gradually, but, as hoped, emerged suddenly and unambiguously at temperatures below a certain “critical” value Tc . The images obtained in the experiments rendered the momentum distribution of the expanding cloud of atoms, or, since the Fourier transform of a Gaussian is another Gaussian, they could also be interpreted as a depiction of the density of atoms in the cloud. The sharp, narrow peak showed that a high fraction of the sodium atoms had very low momentum, or, equivalently, that a macroscopic number of atoms had accumulated at the center of the trap; both of these interpretations were consistent with BEC having taken place. Van Druten’s exclamation point was followed, in turn, by a much smaller, timid question mark enclosed in parentheses: it was of course possible that a faulty apparatus, experimental noise, 2 stress, lack of sleep, and wishful thinking were conspiring to deceive them, but Ketterle and his collaborators had at least two reasons to be confident that what appeared on their screen was indeed a condensate. The first one was trivial: they had already been beaten to the finish line. On 5 June of that year a group led by E. A. Cornell and C. E. Wieman of the Joint Institute for Laboratory Astrophysics (JILA) in Colorado had created a similar, though much smaller, condensate of 2000 rubidium atoms [6] in a time-orbiting-potential (TOP) trap. (R. Hulet and his collaborators at Rice University in Texas had also created a condensate using lithium [7], but their evidence was inconclusive until months later.) The second reason was a real one, and was based on the fact that the trapping potentials were, in all cases, anisotropic: while a “classical” thermal cloud obeys the equipartition theorem and expands isotropically when freed from an anisotropic trap, the Bose-Einstein condensate, despite its macroscopic size, is a fully quantum-mechanical entity that obeys the uncertainty principle and, when released, should expand faster along the direction in which it had been confined more tightly (see Fig. 1.1). This is just what happened at MIT and had happened before at JILA. More than a year later, after Ketterle and his group had succeeded in imaging condensates nondestructively and observing them undergo the transition in situ, there came another breakthrough: the group used a laser to gently cut a condensate in two, separated the pieces, put them back together, and obtained a series of clean interference fringes [8] that showed that the atoms in the condensate were all in the same quantum state, exactly as Einstein had predicted some 70 years before. There remained no doubt that the Bose-Einstein condensate, until then a creature made out of ink, paper, and mathematics, had finally been given physical form. Some time before these events, but surely aware of their imminence, A. J. Leggett wrote a paragraph that, in our view, goes a long way toward assessing their importance (the emphasis is Leggett’s): [T]he temperature, roughly 3 K, of the cosmic black-body radiation [ . . . ] is believed to be the present temperature (in so far as one can be defined) of outer space, and according to almost all cosmologies in current fashion the universe as a whole was never cooler than this [ . . . ]. If this is correct, then when we cool bulk matter to temperatures below the point marked we are in effect creating new physics which Nature herself has never explored: indeed, if we exclude the possibility that on some planet of a distant star alien life-forms are doing their own cryogenics, then the new phases of matter which we create in the laboratory at low temperatures have never previously existed in the history of the cosmos. [9] What’s more, the JILA, MIT, and Rice experiments in a sense created a whole new field, removing Bose-Einstein condensation from its niche somewhere between atomic physics and 4 He studies and 3 putting it right at the forefront as a general-interest discipline that, in turn, has breathed life into many new subfields and opened the possibility of illuminating some of the dark corners still left in more-established ones. A macroscopic manifestation of quantum mechanics akin to laser action, superfluidity, and superconductivity, BEC had been known for decades to be at least partly responsible for these other phenomena; since 1995 it became possible to extricate pure BEC behavior from the other complications that characterize low-temperature systems and thus make further progress toward their understanding. Of great help in this endeavor is yet another astounding feature of the systems in which BEC was finally created, namely their (in many respects) textbook-problem simplicity—for the most part, they can be accurately described by ensembles of spinless point masses in perfect harmonic confinement and affected by very simple interaction potentials—which makes them an excellent laboratory in which to study the quantum-mechanical N -body problem using the methods that have been developed over time to treat other systems. Now, the particular characteristics of these systems—their finite size, the Bose statistics they obey, and their inhomogeneity, among others—bring with them a set of difficulties and subtleties that need to be dealt with. This dissertation will use two standard many-body methods, mean-field theory and path-integral Monte Carlo simulations, to approach and partially answer, in the context of these finite trapped systems, a question almost as old as Einstein’s original prediction: can Bose-Einstein condensation occur in two dimensions? 1.2 BEC, two dimensions, and traps Predicted by Einstein [10,11] based on S. N. Bose’s [12] reinterpretation of Planck’s law of blackbody radiation, Bose-Einstein condensation is a phase transition whereby, at the appropriate combination of temperature and density, a system composed of bosons2 will experience a sudden avalanche of particles into its ground state, resulting in a macroscopic accumulation. This phase transition is a consequence of the indistinguishability of the particles that compose the system, as can be seen from the following simple model [13]: Suppose we have two balls that we can place in any of two boxes, 2 Most textbooks and popular books and articles define bosons to be particles of integer spin. However, bosons were originally defined in terms of their statistical behavior, of which BEC is the fundamental characteristic at low temperatures. In fact, the spin-statistics theorem states that particles with half-integer spin cannot be bosons and particles with integer spin cannot be fermions. In any event, the integer-spin identification is adequate for our purposes, even though the “particles” that we will consider are far from elementary. For example, 87 Rb, where BEC was first produced, consists of 124 spin- 21 particles. 4 each of which can fit any number of balls (this last condition distinguishes Bose-Einstein statistics from Fermi-Dirac statistics). If we initially make the balls distinguishable by painting one of them white, we have four distinct ways |•◦k | |◦ k• | |• k◦ | | k•◦| (1.1) of placing them in the boxes, of which two (or 50% of the total) have two balls in the same box. On the other hand, if the balls are identical, we now have only three different ways |••k | |• k• | | k••| (1.2) of placing them in the boxes, of which two (or 67% of the total) have more than one ball in the same box. The indistinguishability of the balls has enhanced their chances of occupying the same box. If we identify the boxes with energy states and the balls with atoms, it follows that, as the temperature of the system (and hence its average energy) decreases, the atoms will tend to congregate in the state of minimum energy. As Bose and Einstein showed, this feature was implicit in the Bose-Einstein distribution function, a generalization of Planck’s blackbody distribution that prescribes the average number of atoms, hNk i, that occupy a given energy state Ek in a system at temperature T : hNk i = 1 ; eβ(Ek −µ) − 1 (1.3) here we have introduced the inverse temperature, β ≡ 1/kB T , where kB is Boltzmann’s constant, and the chemical potential µ, which ensures that the total number of particles N is, on average, conserved: hN i = X k hNk i. (1.4) Since the occupation numbers must be all positive, the chemical potential in (1.3) must be smaller than the ground-state energy: µ < E0 . For a three-dimensional ideal gas in a box, where E = p2 /2m, we can in principle turn the sum (1.4) into an integral [14,15] to obtain N= V (2π~)3 Z ∞ 0 4πp2 dp eβ(p2 /2m−µ) 5 −1 , (1.5) in which case the chemical potential must always be negative. The substitution x ≡ βp 2 /2m turns Eq. (1.5) into V 2π N= (2π~)3 µ 2m β ¶3/2 Z ∞ 0 x1/2 dx V ≡ 3 g3/2 (eβµ ), x−βµ e −1 λT (1.6) where we have introduced the de Broglie wavelength, λ2T ≡ 2π~2 β/m ≡ 1/ΛT , and the Bose-Einstein integral g3/2 (x) ≡ ∞ X xk , k 3/2 k=1 (1.7) whose definition and properties are reviewed in Appendix A. This implies the following general relation to be obeyed by the number density, n ≡ N/V , of the Bose gas: λ3T n = g3/2 (eβµ ). (1.8) As the temperature of the system decreases, β becomes larger. The chemical potential must correspondingly increase in order to keep the occupation numbers from becoming negative, and eventually, at some temperature Tc , becomes (infinitesimally smaller than) the ground-state energy of the system (zero in this case). The right-hand side of (1.8) becomes ζ(3/2) ≈ 2.612, where P∞ ζ(z) = gz (1) ≡ k=1 k −z is the Riemann zeta function, and we are left with Einstein’s criterion for condensation: nc = ζ(3/2) ≈ 2.612. (ΛTc )3/2 (1.9) An inconsistency then arises: the right-hand side of (1.9) is bounded, and yet there is nothing that prevents us in theory from fixing the temperature at a value lower than T c and increasing the density beyond nc by adding particles—especially since, with a vanishing chemical potential, it costs no energy to do so. To resolve this inconsistency, Einstein postulated that any extra particles would have to occupy the ground state, whose relative importance had been neglected when introducing the continuum approximation (1.5), and force the system to condense at those lower temperatures; BEC had been conceived. The exact same analysis in two dimensions yields nc = ζ(1), ΛTc (1.10) and here we meet our dissertation topic: ζ(1), also known as the harmonic series, diverges, and as a consequence this two-dimensional homogeneous ideal system can accommodate any number of particles at any temperature without the need to invoke a condensation: there is no Bose-Einstein condensation in two dimensions. (Or is there?) 6 Einstein accompanied his 1924 prediction of a condensation with a proposal of three systems where it could be detected [16]: hydrogen, helium, and the electron gas. The third was ruled out the next year with the introduction of Fermi-Dirac statistics.3 F. London’s observation [19] that the λ point of 4 He occurs at a temperature close to Einstein’s Tc alerted physicists to the importance of BEC in low-temperature systems and started a line of research that continues to this day; however, the particular characteristics of helium—the low mass of its atoms and their complicated interactions—make it a difficult system in which to study BEC directly. 4 Research into the first system received a boost when it was realized [21] that atomic hydrogen (kept that way by spin polarization) should remain in the gaseous phase at all temperatures. By the mid-1980s, two rival groups, one led by D. Kleppner, T. J. Greytak, and D. E. Pritchard at MIT and the other led by I. F. Silvera at Harvard University, had managed to stabilize and cool down spin-polarized hydrogen (in a container, using standard low-temperature methods) to a combination of density and temperature not far from the BEC régime [22]. These efforts literally ran into a wall, however, when it was found that the atoms stuck to the walls of the container and recombined into molecules; this was initially suppressed by heating the walls, but the improvement in density thus attained turned out to be short-lived. As a solution the experimenters decided to enclose the hydrogen atoms in magnetic traps. Now, when we put an ideal Bose-Einstein gas in a (harmonic isotropic) trap, a similar analysis to the one above leads to 1 N= (2π~)3 Z 3 dx Z ∞ 0 4πp2 dp eβ(p2 /2m+mω2 r2 /2−µ) − 1 = µ kB T ~ω ¶3 g3 (eβµ ). (1.11) The density now depends on the position: λ3T n(r) = g3/2 (eβ(µ−mω 2 2 r /2) ); (1.12) if we evaluate this density at x = 0 we recover Einstein’s criterion (1.9) (to quote from Ref. 23, “the chief effect of the trap is merely to concentrate the particles to the density at which BEC commences”), and below the (higher) Tc that results any additional atoms pile up in the oscillator 3 There has been, however, an active search for BEC in a similar system, the exciton gas in semiconductors [17,18]. 4 In early 2001, BEC was observed in a dilute gas of 4 He in the 23 S1 metastable state [20]. 7 ground state. More interesting for our purposes is the result we obtain when we repeat the analysis in two dimensions; we obtain N= µ kB T ~ω ¶2 g2 (eβµ ), (1.13) and g2 (x), like g3/2 (x), is bounded in the limit µ → 0, which leaves open the possibility of a p condensation in this system below a temperature kB Tc = ~ω 6N/π 2 [24]. (Note, however, that the system density once again yields g1 (x), and diverges in this limit.) This possibility became √ even more attractive when it was realized that the critical temperature was proportional to N in two dimensions, while in three it went as N 1/3 and was thus much lower; this fact will show up repeatedly in this thesis. The quest for Bose-Einstein condensation in hydrogen at MIT—which, at one point or another, involved Cornell, Hulet, Ketterle, and Wieman [3]—continued to be beset by difficulties, this time at the evaporative-cooling stage, but eventually came to a happy conclusion in 1998 [25]. We cannot help but close this section with Kleppner’s account [26] of the night when the first hydrogen condensate was created: Late one night last June, a phone call from the lab implored me to come quickly. I had a pretty good idea of what was up because BEC in hydrogen had seemed imminent for more than a week. As I drove in the deep night down Belmont Hill toward Cambridge, still dopey with sleep, the blackness of the sky suddenly gave way to a golden glow. I was not surprised because I had a premonition that the heavens would glow when BEC first occurred in hydrogen. Abruptly, streamers of Bose-Einstein condensates shot across the sky, shining with the deep red of rubidium and brilliant yellow of sodium. Small balls of lithium condensates flared and imploded with a brilliant red pop. Stripes of interference fringes criss-crossed the zenith; vortices grew and disappeared; blimp-shaped condensates drifted by, swaying in enormous arabesques. The spectacle was exhilarating but totally baffling until I realized what was happening: The first Bose-Einstein condensates were welcoming hydrogen into the family! 1.3 This thesis We have just seen that, while Bose-Einstein condensation does not occur in a two-dimensional ideal homogeneous gas, the introduction of a trap changes the character of the system and allows the transition to happen. However, when one restores the interactions between atoms their ability to condense is lost once again, at least in the thermodynamic limit; this result, though rigorously proved (by P. C. Hohenberg in 1967 [27], and adapted to the trapped case by W. J. Mullin in 1997 [28]), continues to be somewhat of a mystery, since every relevant limiting case (not just the ideal gas) seems to suggest otherwise: the equations of many-body theory consistently predict the existence of 8 BEC long-range order at zero temperature; for small numbers of particles, both experiments [29,30] and Monte Carlo simulations [31–33] show results consistent with the presence of BEC. On the other hand, these results may be due to other processes (or combinations thereof) that have been predicted to happen in other two-dimensional systems. The Kosterlitz-Thouless transition [34], for example has has been proved to take place [35] in two-dimensional systems that are free in one direction and harmonically confined in the other. There is the possibility of a transition into a “smeared” BEC [36–38], in which a band of states become populated in such a way that their total population is macroscopic. Finally, there is the possibility of a quasicondensation: the presence of coherence over a region that, though finite, might still be larger than that prescribed by the characteristic length of the system [39–41]. In this dissertation we will look into a few different questions regarding the existence of BoseEinstein condensation in two-dimensional traps. First of all, is it possible to find an approximation scheme that predicts the existence of a BEC in a finite system? If that is the case, what will this condensed system look like? Moreover, is this condensation into a single state? In other words, how large is the population of the condensate (that is, the ground state of the system) when compared to those of the low-lying states? Finally, the standard way of creating a two-dimensional condensate in the laboratory [30] consists of taking a three-dimensional system and squeezing it in one direction, creating in effect a pancake-shaped trap. In that case, how well can the resulting quasi-2D condensate be described by our two-dimensional scheme? Do the density profiles match? Are the condensate fractions consistent? How does the ground-state population compare, once again, to those of the excited states? Our scope is narrow: we will restrict ourselves to the depiction and analysis of density profiles, the calculation of condensate fractions, and the comparison of ground-state occupations to those of a few excited states. Except in the case of the ideal gas, we will avoid the transition itself, mainly because most of the methods we use are expected to break down—and indeed do—before the temperature reaches its critical value. We will not consider the dynamics of the system at all, though experiments are going in that direction. In particular, and being aware of the enormity of our omission, we will shy away from discussing superfluidity; the Kosterlitz-Thouless transition will receive little more than a passing mention. The gas will always be assumed to be at temperatures so low that it is at least partially condensed. It will also be invariably assumed to be in equilibrium; in particular, whenever we speak of raising or lowering the temperature of a system, we will assume that this is done slowly enough so that equilibrium is not disturbed. (And even then we leave aside the problem of fluctuations about equilibrium values.) 9 We have tried to be as thorough as possible within this restricted range, and to make our presentation self-contained. Though we have tried to relegate the goriest details to the appendices, the main text is still full of mathematics. We have striven to fill in the steps that, though taken for granted in research papers, are far from trivial. The reader will also note that our restricted subject in fact ramifies into a maze of subdivisions and special cases; moreover, these subtopics can be and have been addressed using one or more—and even combinations—of the methods we present. As we said above, the systems under study are finite, dilute, harmonically trapped Bose-Einstein gases below their transition temperatures. They will be either ideal or interacting, and either at absolute zero or at finite temperature. The traps that confine them will be of three types: threedimensional and isotropic, two-dimensional and isotropic, or three-dimensional and isotropic in the x- and y-directions, with a variable confinement along the z-direction always greater than those along the other two. (Table 1.1 on the following page contains a partial listing of the systems that we have studied, the methods we have used to study them, and the results we have obtained.) We begin by studying the ideal gas in Chapter 2. Though hardly a realistic system, the ideal gas has been an important reference and starting point from the very beginning. For our purposes, the ideal trapped gas is important because it can be treated exactly at every temperature, for any (finite) number of atoms, and in traps of any dimensionality or degree of anisotropy. This exact method will allow us to look at the transition in some detail, to define and exhibit the quantities that we will be calculating later, and to introduce a semiclassical approximation that will recur, with similar successes and limitations, in the rest of the thesis; we do note that this semiclassical approximation will be used to describe only the noncondensate; the condensate itself will receive the same fully quantum-mechanical treatment as in the exact theory. We will display various results that will enable us to get a feel for the temperature scales and occupation numbers involved and will acquaint us with the typical shapes of the density profiles and of the wavefunctions that characterize the various states in which the atoms can be. In Chapter 3 we begin our study of interacting gases using mean-field theory. We develop the standard Gross-Pitaevskiı̆ theory (also known as the Hartree-Fock-Bogoliubov equations) in as general and detailed a way as we can; immediately afterwards we go on to introduce diverse approximations of increasing crudity, settling eventually upon, and extracting results from, two of the simplest: While it is perhaps the least sophisticated approach to the study of BEC, being easier to solve than even the ideal gas, the zero-temperature Thomas-Fermi limit nonetheless yields predictions remarkably close to reality and is a standard tool for fitting experimental data. Though the assump- 10 Table 1.1. Summary of methods used and systems studied in this thesis 3D 2D Compressed 3D Works always Works always Works always Fails at high T Fails at high T Fails at high T Works always Works always Not tested enough Works always Works always Fails at high λ1 but can be adapted for that case ◦ HFB + GP2 Yields results similar to those of the HF approximation or else fails Always fails Not tested enough ◦ HF + GP3 Fails at high T Fails at high T Not tested enough ◦ HFB + TF4 Not tested enough Always fails Not tested enough Works where the HF approximation works Works where the HF approximation works Not tested enough Always fails Works always but has a high free energy Not expected to work Works always Works always Works always Works always Works but disagrees with the compressed-3D results Described well at high λ by the 2D HF approximation Ideal gas ◦ Exact solution ◦ Semiclassical approx. Mean-field theory ◦ T =0 ◦ Gross-Pitaevskiı̆ Eq. ◦ Thomas-Fermi limit ◦ T 6= 0 ◦ Thermodynamic limit ◦ HF + TF5 Uncondensed solution PIMC simulation ◦ Ideal gas ◦ Interacting gas 1 2 3 4 5 Compression ratio or anisotropy parameter, defined by λ ≡ ωz /ω, where ω ≡ ωx = ωy . Semiclassical Hartree-Fock-Bogoliubov approximation for the thermal cloud combined with the Gross-Pitaevskiı̆ equation for the condensate. Semiclassical Hartree-Fock approximation for the thermal cloud combined with the GrossPitaevskiı̆ equation for the condensate. Semiclassical Hartree-Fock-Bogoliubov approximation for the thermal cloud combined with the Thomas-Fermi limit for the condensate. Semiclassical Hartree-Fock approximation for the thermal cloud combined with the ThomasFermi limit for the condensate. 11 tions that lead to it lose validity when the system becomes very anisotropic, it is still possible to slightly modify the model in order to accommodate the necessary changes, and as a result we get an important connection between a squeezed 3D gas and a 2D gas. The Gross-Pitaevskiı̆ theory models three-dimensional interactions by considering two-body collisions, and the relevant parameter that describes them, the s-wave scattering length, is an experimentally measurable quantity. The two-dimensional version of the theory is quite different, and, even if we try to model 2D collisions using 3D methods, we still have to deal with the fact that the corresponding interaction parameter has different dimensions, and, consequently, cannot be readily expressed in terms of the standard three-dimensional coupling constant. Still, we do not expect the interactions in the two-dimensional gas to depend too strongly on parameters other than the scattering length, and, when the 2D system has been created by compressing a three-dimensional trap in one direction, some relation between the characteristic length in that direction to those in the others. The Thomas-Fermi limit provides that missing link. The other simplified scheme will be our main workhorse when we turn to the study of finitetemperature systems in the rest of the thesis: this method combines an exact description of the condensate by means of the Gross-Pitaevskiı̆ equation (exact in the sense that it is identical to that prescribed by more-rigorous theories) with a semiclassical, Hartree-Fock treatment of the noncondensate. After presenting all the theory, we proceed to discuss the numerical methods we used to solve the Hartree-Fock equations and show some of the results we obtained. We initially study one of the standard systems in the literature: a 104 -atom 3D isotropic gas with 87 Rb parameters. This will give us the chance to display the condensate fraction, the density, and the chemical potential of an interacting gas; in particular, we will verify that interactions affect the transition temperature, deplete the condensate, and cause a noticeable increase in the size of the system. This study will also allow us to see some limitations of the semiclassical approximation—notably its failure at high temperatures. After that, we go back to the 2D gas: repeating what we did in Chapter 2 for the ideal case, we will set up and diagonalize the off-diagonal one-body density matrix in the semiclassical Hartree-Fock approximation; the resulting eigenvalues and eigenvectors will show us the effects of interactions on the ground- and excited-state populations of system and on the shapes of the corresponding wavefunctions. The study of the strictly two-dimensional interacting trapped gas is taken up in more detail in Chapter 4. At that point we will be confronted with another interesting fact: the mean-field equations derived and used in Chapter 3 can also be solved for a 2D interacting finite trapped system 12 without having to invoke the presence of a condensate [42]. The approximation scheme that allows these “uncondensed” solutions is in principle more rigorous than the one we have been using, since, while still semiclassical, it does not make use of the Hartree-Fock approximation. When we try to solve this model in the presence of a condensate, we encounter phononlike quasiparticles at the low end of the energy spectrum; on the other hand, if we momentarily consider high temperatures and artificially remove the condensate, the phonons disappear from the spectrum—and uncondensed solutions can be found all the way down to T = 0. This would imply that phonons destroy the long-range order of the system and prevent it from condensing, in agreement with previous explanations [43]. However, these uncondensed solutions have consistency problems of their own (in particular, they do not reduce to the correct limit as T → 0) and are found to have a higher free energy than the Hartree-Fock condensed ones. This renews our faith in the validity of the HartreeFock solutions, along with the infrared energy cutoff they impose, and encourages us to put them through one last test. Chapter 5 looks at the shape and condensate fraction of a quasi-2D interacting system created by deforming a three-dimensional trap into the shape of a pancake. Our aim is to see how well the surface density profile and condensate fraction of that squeezed system can be described using the density profile and condensate fraction predicted for a pure 2D gas by the Hartree-Fock approximation. As a guide, and in the absence of experimental results, we will introduce and use another, different, method of analysis. Path-integral Monte Carlo simulations [44] are, for interacting gases, the closest we have to an exact method like the one we use in Chapter 2 to treat the ideal gas [45]; we will spend a good fraction of Chapter 5 justifying this last statement, explaining the method, and describing its implementation for the specific case of a trapped Bose gas. After checking the correctness of our simulations by reproducing results obtained in earlier chapters, we will proceed to display the results we obtained for squeezed interacting systems and compare them to our mean-field predictions. Chapter 6 will wrap up our considerations and suggest some directions for future work. 13 CHAPTER 2 THE IDEAL TRAPPED BOSE GAS The weakly interacting Bose gas can be treated using the mean-field approximation, though at the low densities likely to be of experimental interest, the corrections are not expected to be important. —V. Bagnato and D. Kleppner [24] 2.1 Introduction As it turns out, corrections due to interactions are much more important than Bagnato and Kleppner expected. When interpreting their data, experimentalists use the zero-temperature Thomas-Fermi limit (to be discussed later), not the ideal gas, as a first approximation. It has even been argued that, even though the phenomenon was discovered in the ideal gas, the very existence of a condensation is a consequence of interactions.1 There are, however, many reasons why the ideal gas is worth studying, and they are not restricted to its excellence as a pedagogical tool. The condensates created in the laboratory occur in systems that are both decidedly finite and very inhomogeneous; the ideal gas can be adapted to incorporate these nontrivial effects, many of which carry over to the interacting case without modification. Moreover, the noninteracting gas is becoming less of an idealization with every passing day: recent work in the study of Feshbach resonances has enabled experimentalists to tune the strength of interatomic interactions, and the possibility of creating an essentially ideal trapped gas is becoming progressively plausible. We begin the chapter by introducing the one-body density matrix as the quantity that includes all the information that can be gleaned about a finite system of trapped noninteracting bosons. After discussing some exact quantum results where the inhomogeneity and finite size are already present, we introduce a semiclassical approximation that enables us to set a temperature scale for the 1 Consider for example the following remarks by P. Nozières [38] (the emphasis is his): “[D]o the condensate particles accumulate in a single state? Why don’t they share between several states that are degenerate, or at least so close in energy that it makes no difference in the thermodynamic limit? The answer is non-trivial: it is the exchange interaction energy that makes condensate fragmentation costly. Genuine Bose-Einstein condensation is not an ideal gas effect: it implies interacting particles!” 14 system and that will reappear in a more sophisticated form when we treat the interacting gas in the following chapters. This semiclassical approximation does not include the condensate, which has to be added by hand; this limitation will provide us with a first opportunity to perform a self-consistent calculation. By explicitly diagonalizing the one-body density matrix of a gas in an isotropic 2D trap we can obtain the populations of the ground state of the system and of its first few excited states. We perform the diagonalization for distinguishable particles and for bosons and find that in the distinguishable case the ground state is not preferentially populated, while in the Bose case the occupation of the ground state is identical to that found by the more direct method. As we would expect, the eigenvectors for both cases are the same and correspond to the eigenfunctions of the harmonic oscillator. We also squeeze a three-dimensional system by increasing one of the trapping frequencies; this contributes to an enhancement of the density at the center of the trap and thus encourages the system to condense. Indeed, we will see that we can take a system at a temperature above the transition temperature and cause it to condense merely by confining it more tightly. 2.2 Exact results for the ideal gas We start by studying a system of N distinguishable noninteracting atoms, each of mass m, trapped by a σ-dimensional isotropic harmonic-oscillator potential characterized by an angular frequency ω (we will consider anisotropic traps in later sections). Since there are no interactions, each atom moves independently of the others, and we can consequently visualize the ensemble as being composed of N one-body systems. Let us restrict ourselves momentarily to one of these. In an isotropic trap, the motion of a single atom is described by the one-body Hamiltonian H1 = 1 p2 + mω 2 r2 , 2m 2 (2.1) where p and r are the momentum and position operators. The harmonic-oscillator potential sets the length and energy scales of the system, and it is in terms of these that we will henceforth express all physical quantities. As is well known [46], a single quantum-mechanical point mass confined by a one-dimensional harmonic trap has a minimum energy, given by ~ω/2, that corresponds to a 15 ground state of Gaussian shape characterized by a length x0 such that x20 = ~/mω. Introducing dimensionless coördinates x = r/x0 and momenta κ = x0 p/~, we can rewrite (2.1) as H1 ≡ 1 1 ~ω H̃1 = ~ω (κ2 + x2 ), 2 2 (2.2) whose eigenenergies in σ dimensions are given in our dimensionless units by ²n1 ···nσ ≡ 2 En ···n = 2n1 + · · · + 2nσ + σ, ~ω 1 σ (2.3) where n1 , . . . , nσ are integers. (This is for Cartesian coördinates; the eigenfunctions and eigenvalues in other coördinate systems are displayed in Appendix A). When the system is at a temperature T , quantum statistics postulates that all relevant information is contained in the canonical one-body density matrix, ρ1 = e−βH1 /Z1 , where β ≡ 1/kB T , kB is Boltzmann’s constant, and the one-body partition function Z1 ≡ Tr e−βH is a normalization factor that ensures Tr ρ1 = 1. When we evaluate the diagonal elements of the (unnormalized) one-body density matrix in real space we obtain hx|e−β̃ H̃1 |xi = 2 1 cschσ/2 2β̃ e−x tanh β̃ , (2π)σ/2 (2.4) where we have introduced the dimensionless inverse temperature β̃ ≡ t−1 ≡ 21 ~ωβ. This expression, first found by F. Bloch, can be derived by explicitly summing over Hermite polynomials [47], by using Feynman path integrals, by solving the Bloch equation H̃ρ1 = −∂ρ1 /∂ β̃ [15], or by purely algebraic methods [48]. Appendix A sketches a proof for σ = 2 using Laguerre polynomials. The one-body partition function, Z1 = (2 sinh β̃)−σ , can be found by integrating (2.4) with respect to x or by evaluating the trace in its energy eigenbasis. If we divide Eq. (2.4) by Z 1 we obtain the normalized diagonal density matrix, ñ(x) = 2 1 tanhσ/2 β̃ e−x tanh β̃ ; π σ/2 (2.5) we shall henceforth use the notation ñ ≡ xσ0 n to refer to densities in real space. PN When we have N atoms, the Hamiltonian becomes H̃ = j=1 H̃1,j , and the system will be described by a density matrix ρ= N Y e−β̃ H̃1,j 1 −β̃Σj H̃1,j = e , ZN Z1 j=1 16 (2.6) where we used the fact that, in the absence of interactions, ZN = (Z1 )N . The corresponding reduced one-body density matrix is defined by [49] ρ1 = N Tr ρ = N 2:N e−β̃ H̃1,1 Z1 (2.7) and is identical to (2.5) but normalized to N : ñ(x) = 2 N tanhσ/2 β̃ e−x tanh β̃ . σ/2 π (2.8) Up to now, the atoms, though identical, have been taken to be distinguishable: we have used labels to tag them and even selected the “first” one when we calculated the reduced one-body density matrix. When the system is composed of indistinguishable bosons, the situation is much more involved (see Appendix B and Ref. 50 for some details), but the result is transparent and intuitive. If we invoke the grand canonical ensemble, it is possible to show that the reduced onebody density matrix adopts a form that mimics the Bose-Einstein distribution: 2 ρ1 ≡ 1 eβ̃(H̃1 −µ̃) − 1 = ∞ X e−`β̃(H̃1 −µ̃) ; (2.9) `=1 the chemical potential µ̃ is determined by the subsidiary condition N = Tr ρ 1 . The fact that we can express Eq. (2.9) as a sum over ` will be of enormous convenience throughout this work, but its importance reaches far beyond; as Appendix B shows, it is a manifestation of Bose symmetry at its deepest level, the closest we will get to being able to define what we mean by bosons. The real-space diagonal density now becomes ∞ ∞ X 2 1 e`β̃(µ̃−σ) 1 X σ/2 `β̃ µ̃ −x2 tanh `β̃ hx|ρ1 |xi ≡ ñ(x) = e−x tanh `β̃ . e csch 2` β̃ e = (2π)σ/2 `=1 π σ/2 `=1 (1 − e−4`β̃ )σ/2 (2.10) It is worth emphasizing that it is the difference µ̃ − σ, and not µ̃ by itself, that tends to zero below the transition in the trapped case. This can be clarified further by noticing that, if we were to add another particle to the system, it would have to have an energy of at least ~ωσ/2, the ground-state 2 Note that we use exactly the same notation for distinguishable and Bose-Einstein density matrices. In the few instances where this could lead to confusion, in Chapter 5 especially, we will use the suffix B to denote quantities that obey Bose statistics. 17 energy of the oscillator, and the energy of the system would increase by that amount. We can readily integrate any of the expressions above with respect to x to obtain [51,52] ∞ ∞ X e`β̃(µ̃−σ) 1 X e`β̃ µ̃ = . N= σ σ 2 sinh `β̃ (1 − e−2`β̃ )σ `=1 `=1 (2.11) The first expression is especially interesting because it gives us the total particle number in terms of the canonical one-body partition function; proofs of this result appear in Appendix B and in Ref. 53. Equations (2.10) and (2.11) provide a complete description of the system, and, as Fig. 2.1 illustrates, they predict the occurrence of a condensation below a certain temperature. The Bose-Einstein condensate is a pure (albeit macroscopic) system that can be described by a single wavefunction. In the trapped ideal case, this wavefunction is the purely real ground state of the harmonic oscillator. The condensate density is the square of this wavefunction, and as a consequence the condensate density profile is the same at all temperatures: ñ0 = N0 −x2 ; e π σ/2 (2.12) only its normalization constant, which corresponds to the condensate number, varies. To calculate the condensate number we plug the ground-state energy ²0 = σ into the Bose-Einstein distribution function: N0 = 1 eβ̃(²0 −µ̃) − 1 = 1 eβ̃(σ−µ̃) − 1 = ∞ X e`β̃(µ̃−σ) . (2.13) `=1 We can then use (2.12) and (2.13) to resolve the total density (2.10) into its condensate and noncondensate components at any temperature and for any number of atoms. Figures 2.2 and 2.3 display the rapid growth of the condensate density with decreasing temperature and show that the thermal density attains a maximum at a point away from the origin; we will later see that this “delocalization” effect is enhanced by interactions. The Bose-Einstein transition is characterized by abrupt changes in various quantities. Figure 2.4 on page 21 shows the temperature dependence of the condensate fraction for different values of N . The onset of condensation, as one would expect, becomes more pronounced as N grows. It is also evident that the system is completely condensed at zero temperature; indeed, in the limit β̃ → ∞ we can neglect the exponential in the denominator of Eq. (2.11) and once again obtain (2.13), provided we assume limβ̃→∞ β̃(µ̃ − σ) to be constant; this confirms that µ̃ ≈ σ at low temperatures. The chemical potential stays essentially fixed at this value until the transition temperature, at which point it experiences a sharp decrease and quickly becomes negative (see Fig. 2.5). The location of 18 0.3 ñ(x)/N PSfrag replacements 0.2 0.1 5 0 0 2 4 x 6 25 8 50 75 T /Tc (%) 50 75 T /Tc (%) 100 PSfrag replacements 1 0.8 N (x)/N 0.5 0.6 0.4 5 0.2 0 0 2 4 x 6 25 8 100 Figure 2.1. Density and number density of a condensed Bose gas. Throughout this work we will depict the density of a Bose system below the transition temperature using one of these two modes of description. In this case, we have a two-dimensional ideal gas of N = 1000 atoms confined by an isotropic trap. The plot at top shows the actual density of the gas, ñ(x), that has appeared ourR discussion; the bottom panel shows the “number density” N (x) defined by N = Rthroughout ∞ dx N (x) = dσx ñ(x)—in two dimensions, N (x) = 2πx ñ(x); this is the quantity yielded directly 0 by the Monte Carlo simulations of Chapter 5. In each case, the total density (solid line) has been resolved into its condensed (dotted) and thermal (dashed) components and has been divided by N . 19 0.15 ñ(x)/N PSfrag replacements 0.1 0.05 5 0 0 2 95 4 x 90 6 85 8 T /Tc (%) 80 Figure 2.2. The growing condensate. This “closeup” view of Fig. 2.1 shows (when read from right to left) the fast growth of the condensate in the region just below the critical temperature. 0.12 ñ(x)/N ñ(x)/N 0.12 0.06 0 0 2 x 4 0 6 0.06 0 0 2 0 2 x 4 6 4 6 0.12 ñ(x)/N ñ(x)/N 0.12 PSfrag replacements 0.06 0 2 x 4 0.06 0 6 x Figure 2.3. Front view of Fig. 2.2. Note that the noncondensate (dash-dotted line) has a much larger spatial extent than the condensate (dashed line). The delocalization hump is clearly visible. 20 1 PSfrag replacements N0 /N 0.8 0.6 0.4 0.2 1.2 0 0 104 0.5 T /Tc 103 N 1 102 Figure 2.4. Condensate fraction of a two-dimensional ideal trapped Bose gas. The solid lines are the exact results obtained from solving Eq. (2.11) for µ̃ and then using (2.13) to extract N 0 ; the dashed lines show the infinite-N limit given by Eq. (2.17) below. We can see that already for N = 10 4 the infinite-N limit is quite accurate. the thermal maximum also changes abruptly at the transition: as Fig. 2.6 illustrates, the “hump” is suddenly shifted towards the origin, signalling that a sizable fraction of the particles have been displaced to the most localized state available. 2.3 The semiclassical approximation As written, the density matrix (2.10) gives an exact description of a finite trapped Bose gas of any dimensionality at any temperature. It is difficult, however, to extract relevant quantities from it in analytic form—the transition temperature is a case in point—and in order to do that we will introduce a semiclassical approximation that we will also be using for the interacting gas. The semiclassical approximation assumes that the temperature is high when compared to the spacing between energy levels (which is microscopically small for a standard trap); if k B T À ~ω, then β̃ ≡ ~ω/2kB T can be treated as a small parameter, and this allows us to expand the exponential in the second expression of Eq. (2.11). However, since ` is unbounded, by performing that expansion we are neglecting the sizable contribution to the sum of large-` values; we will soon see that these 21 PSfrag replacements 2.5 2 0.5 1.5 1 µ̃ 0.5 0 −0.5 −1 −1.5 −2 0 0.2 0.4 0.6 T /Tc 0.8 1 1.2 Figure 2.5. Chemical potential of a two-dimensional ideal trapped Bose gas. The values of µ̃ shown correspond to N = 100, 1000, and 10000 (solid, dashed, and dotted line respectively), just like in Fig. 2.4. 1.17 PSfrag replacements xmax 1.14 1.11 1.08 97 98 99 100 T /Tc (%) 101 102 103 Figure 2.6. Position of the noncondensate maximum, located at the hump shown in Fig. 2.3. In this case we have N = 5 × 104 , 105 , and 1.5 × 105 (dash-dotted, dashed, and solid line respectively). The inward shift caused by the temperature drop is more pronounced for higher N (compare to the behavior of the chemical potential in Fig. 2.5). 22 1 0.8 N0 /N 0.6 PSfrag replacements 0.4 0.2 0 0 0.2 0.4 T /Tc 0.6 0.8 1 Figure 2.7. Condensate fraction of a two-dimensional ideal trapped Bose gas in the semiclassical approximation. The figure shows the exact result for N = 1000 (solid line) along with the infinite-N limit (2.17) (dotted line), the first-order condensate fraction predicted by the self-consistent procedure sketched in (2.19) (dashed line). The last two are virtually indistinguishable for most temperatures and clearly overestimate the condensate. Also shown is the finite-N approximation (2.22) (dash-dotted line), which underestimates the condensate in 2D. correspond to N0 , and thus it is necessary to insert the condensate number by hand [28]. The expansion then yields N ≈ N0 + ≈ N0 + ∞ 1 X e`β̃(µ̃−σ) (2β̃)σ `=1 `σ (1 − σ`β̃) 1 (gσ (eβ̃(µ̃−σ) ) + β̃σgσ−1 (eβ̃(µ̃−σ) )) (2β̃)σ for the total number of atoms; gσ (x) = P∞ k=1 (2.14) xk /k σ is a Bose-Einstein integral, already introduced in Chapter 1, whose definition and properties are summarized in Appendix A. Equation (2.14) can also be obtained directly from the assumption that the harmonic-oscillator energy levels are so closely spaced that they may be taken to form a continuum; the subsequent replacement of sums by integrals helps elucidate the identification of the neglected terms with the condensate fraction [53,54], exactly as in the homogeneous gas [14,15]. As we remarked right below Eq. (2.10) on page 17, the transition can be characterized by the vanishing of the quantity µ̃ − σ; when that happens, we have 23 N ≈ N0 + ζ(σ) σ ζ(σ − 1) + , σ 2 (2β̃)(σ−1) (2β̃) (2.15) where ζ(x) is the Riemann zeta function, or, upon introducing the “critical” temperature defined (σ) by tc ≡ 2(N/ζ(σ))1/σ , N0 =1− N µ t (σ) tc ¶σ − σ −1/σ ζ(σ − 1) N 2 (ζ(σ))(σ−1)/σ µ t (σ) tc ¶σ−1 . (2.16) If we keep the expansion to first order, we obtain N0 =1− N µ t (σ) tc ¶σ =1− µ T (σ) Tc ¶σ , (2.17) (σ) which justifies in hindsight our definition of the critical temperature. For t > t c , the condensate number vanishes: the ground state has an occupation no different than that of any other state. Below the critical temperature, the condensate fraction obeys a simple power law in this limit, and increases with decreasing temperature until it includes every particle in the system. To find the corresponding approximation for the density we expand the exponential in the denominator of Eq. (2.10) and obtain ñ(x) ≈ ñ0 (x) + µ t 4π ¶σ/2 2 gσ/2 (eβ̃(µ̃−σ−x ) ), (2.18) where ñ0 is given once again by (2.12) but with N0 given by (2.17). By integrating both sides of (2.18) over x we obtain the first two terms of Eq. (2.14); to this order, then, the number of atoms is conserved and the approximation is consistent. As a result, Eqs. (2.18), (2.13), and (2.12) combined provide in principle a complete specification of the system like that discussed on page 18. The only unknown is the chemical potential, which we can calculate self-consistently: we initially take N0 = N , µ̃ = σ and loop N0 +σ N +1 µ 0 ¶σ t gσ (eβ̃(µ̃−σ) ) N0 = N − 2 µ̃ = t log (2.19) until µ̃ and N0 stop changing. Some results of this procedure are shown in Figs. 2.7 and 2.8. The second term in (2.16) is a correction [28,55,56] that takes into account the finite size of the system and vanishes as N → ∞, in what is called the thermodynamic limit (TDL). 3 The Riemann 3 It has been remarked [28,43,57] that, just as the true TDL of a homogeneous system is taken by simultaneously letting N → ∞ and V → 0 so that the density remains finite, when taking 24 0.8 N0 (x)/N , N 0 (x)/N PSfrag replacements ñ0 (x)/N , ñ0 (x)/N 0.3 0.2 0.1 0 0 1 2 x 3 0.2 2 4 6 4 6 x 0.4 N0 (x)/N , N 0 (x)/N ñ0 (x)/N , ñ0 (x)/N 0.4 0 0 4 0.15 0.1 0.05 0 0 0.6 1 2 x 3 0.3 0.2 0.1 0 0 4 2 x Figure 2.8. Density and number density of a condensed 2D gas in the semiclassical approximation. Here N = 1000; the panels at top correspond to T = 0.4 Tc and those at the bottom to T = 0.8 Tc . The exact results (solid lines) for the condensate and noncondensate densities are shown alongside the first-order (dashed) and second-order (dash-dotted) semiclassical approximations. As Fig. 2.7 showed, the former overestimates the condensate and the latter underestimates it. Both do a good job of describing the noncondensate density away from the origin but fail at small values of x. zeta function ζ(x) diverges at x = 1, so the finite-size correction does not make sense 4 as written when σ = 2; we can still evaluate it, however, because the Bose-Einstein integral g 1 (z) has a closed analytic expression: g1 (z) = − log(1 − z). The condensate fraction is of order N , so we can use Eq. (2.13) to obtain the true TDL of a harmonic system one should simultaneously increase the number of particles and decrease the spring constant of the trap in such a way that N ω σ remains finite. This is automatically taken into account by our system of units, given the factor 21 ~ω used to scale the temperature. 4 In one dimension, we cannot calculate a finite-size correction at all, but we can still evaluate the first term and find a “critical” temperature, tc = 2N/ log N , below which there is a substantial accumulation of atoms in the ground state [28,51,58]. This, by the way, is all we will say about the one-dimensional trapped gas, a very interesting system in its own right. We refer the interested reader to our own publication [59] on the subject. 25 eβ̃(µ̃−σ) = N0 1 1 1 = ≈1− ≈1− N0 + 1 1 + 1/N0 N0 N (2.20) as β̃(µ̃ − σ) → 0, and hence lim β̃(µ̃−σ)→0 g1 (eβ̃(µ̃−σ) ) = lim β̃(µ̃−σ)→0 − log(1 − eβ̃(µ̃−σ) ) = − log(1/N ) = log N, (2.21) giving us [28] N0 =1− N µ t (2) ¶2 log N −p N ζ(2) µ t (2) tc tc The corresponding second-order density in two dimensions is ñ(x) = ñ0 (x) + ¶ . (2.22) 2 1 1 t . g1 (eβ̃(µ̃−2−x ) ) + 4π 2π eβ̃(µ̃−2−x2 ) − 1 (2.23) Figures 2.7 and 2.8 also show the second-order results of the self-consistent procedure described above. The agreement is in general good, but not perfect. The delocalization effect, for one, does not appear, and the noncondensate density is in general misrepresented at the center of the trap. We can also see that, in two dimensions, the finite-size correction turns out to be too large (in three dimensions, on the other hand, its agreement with the exact results is remarkably good [28,54–56]; see Fig. 3.3). Another shortcoming of the method is that it breaks down at high temperatures; the semiclassical approximations to be discussed later also have this problem. 2.4 The off-diagonal elements of the density matrix Up to now we have been considering only the diagonal of (2.9); needless to say, the off-diagonal elements of the density matrix can give us more information about the system. In Appendix A we derive the two-dimensional version of the full one-body reduced density for the 2D isotropic harmonically trapped ideal gas [60,61], 0 0 ñ(x, x ) = hx|ρ1 |x i = 1 π σ/2 ∞ X `=1 e`β̃(µ̃−σ) (1 − e−4`β̃ )σ/2 1 e− 2 csch 2`β̃((x 2 +x02 ) cosh 2`β̃−2x·x0 ) , (2.24) calculated by using the definition [53] ñ(x, x0 ) = hx| 1 eβ̃(H̃1 −µ̃) − 1 |x0 i = X k 1 eβ̃(²k −µ̃) − 1 hx | kihk | x0 i ≡ X fk φ∗k (x)φk (x0 ), (2.25) k where k denotes a collective summation index that varies with the dimensionality of the system, f k is the kth Bose-Einstein factor, and φk (x) is the kth oscillator eigenfunction. The chemical potential is 26 once again found by using Tr ρ1 = N . If we multiply both sides of Eq. (2.25) by φj (x) and integrate with respect to x, we can immediately exploit the orthogonality of the eigenfunctions to obtain Z dσx0 ñ(x, x0 ) φj (x0 ) = fj φj (x), (2.26) where, taking into account the symmetry of the matrix, we have swapped x and x 0 . A glance at Eqs. (2.4) and (2.10) should convince us that the calculation for the distinguishable gas is analogous; we find that the eigenvectors of the density matrix are the usual oscillator wavefunctions, regardless of the statistics obeyed by the atoms, and the corresponding eigenvalues are the populations of these eigenstates. While the diagonal one-body density provided us with the population of the ground state on one hand and of the “rest of the world” on the other, by considering the off-diagonal elements we can actually compare these occupation numbers state by state. From our preceding considerations we expect the Bose-Einstein density matrix to have an uneven distribution of eigenvalues. The largest one corresponds to the condensate population and should be of order N , while the rest should be orders of magnitude smaller. This, in fact, is the standard definition of BEC, first postulated in a classic paper by O. Penrose and L. Onsager [49]. As preparation for the interacting problem that we will tackle in later chapters, we will sketch the numerical diagonalization procedure that we have used for the ideal case. First of all, we note that Eq. (2.24) depends almost exclusively on the radial coördinates; the angular dependence is contained only in the dot product x · x0 = x0 x cos(ϕ0 − ϕ). The matrix is in general a 2σ-dimensional array, and in order to turn it into a matrix that we can diagonalize we must average over or integrate out all but two of these dimensions; in particular, the process will be feasible only for isotropic traps. Finally, the integrals have to be approximated by discrete sums. We will restrict ourselves to a two-dimensional system. Since we are assuming an isotropic trap, we can take our x-axis in any direction, and we will choose it so that ϕ = 0; moreover, we will average out the angle ϕ0 by defining ρ̄(x, x0 ) ≡ 1 2π Z 2π dϕ0 ñ(x, x0 ), (2.27) 0 which, using the result (A.15), turns the `th term of (2.24) into 1 ρ̄` (x, x0 ) ∝ e− 2 (x 2 +x02 ) coth 2`β̃ 27 I0 (x0 x csch 2`β̃), (2.28) 1000 Population PSfrag replacements 100 10 7 1 0 1 2 3 4 Eigenstate 5 6 Figure 2.9. Eigenvalues of the two-dimensional isotropic trapped-ideal-gas density matrix. The white bars correspond to a 1000-atom Bose-Einstein gas at T = 0.8 Tc , while the black bars represent the populations of an equivalent distinguishable gas. Note that the scale on the y-axis is logarithmic. The open circles represent the populations predicted by Eqs. (2.25) and (A.10), f k = (exp(β̃[²k − µ̃]) − 1)−1 with ²k = 2(2k + 1); the straight line depicts the same function for continuous k. The corresponding eigenfunctions are shown in Fig. 2.10. where I0 (x) is a modified Bessel function. (An equivalent treatment in three dimensions yields a hyperbolic sine.) For large values of the argument (x ≥ 12, to be precise), the Bessel function has to be replaced with its asymptotic expansion [62] ex I0 (x) ∼ √ 2πx µ ¶ 1 9 1+ + + ··· . 8x 128x2 (2.29) We used two different methods to integrate the resulting density matrix. First we evaluated it on an equally spaced grid of 300 × 300 points and multiplied it by a replicated array of the coefficients of the alternative extended Simpson’s rule [63], Z xN x1 59 43 49 17 f1 + f2 + f3 + f4 + f 5 + · · · 48 48 48 48 43 59 17 49 · · · + fN −4 + fN −3 + fN −2 + fN −1 + fN ], 48 48 48 48 f (x) dx ≈ h [ (2.30) where fj = f (xj ) and h = x2 −x1 . We got identical results much more quickly (and more accurately, since the matrices involved are much smaller) by evaluating the matrix on a 30 × 30 grid of points 28 located at the zeros of the 30th Laguerre polynomial [64] and multiplying it by the Gauss-Laguerre weights [62,65] introduced in Appendix C. The condensate fraction that we obtain for the twodimensional gas is indistinguishable from that displayed in the previous section; one particular case is shown in Fig. 2.9. As we can see in that figure, this method also reproduces the Bose-Einstein distribution function to very high accuracy, at least for the lowest-lying energy eigenstates. The corresponding eigenvectors are indeed the same for the distinguishable and Bose-Einstein matrices, and, as we can see in Fig. 2.10, they are identical to the standard harmonic-oscillator eigenfunctions. It is also possible to evaluate the density matrix in the semiclassical approximation that we introduced in the last section. We start by noting that the argument of the exponential in Eq. (2.24) can be rewritten as 1 1 − csch 2`β̃((x2 + x02 ) cosh 2`β̃ − 2x · x0 ) = − sech `β̃(2(x2 + x02 ) sinh `β̃ + (x − x0 )2 csch `β̃) 2 4 1 = − ((x + x0 )2 tanh `β̃ + (x − x0 )2 coth `β̃). (2.31) 4 The second of these expressions is frequently encountered in the literature [60,61]; any of them can be expanded for small β̃, along with the exponential in the denominator of Eq. (2.24), to yield the semiclassical expression 0 0 ñ(x, x ) ≈ ñ0 (x)ñ0 (x ) + µ t 4π ¶σ/2 X ∞ `=1 e`β̃(µ̃−σ) −`β̃ 1 (x2 +x02 ) −(x−x0 )2 /4`β̃ 2 e e ; `σ/2 (2.32) once again it is necessary to insert the condensate density by hand. This result will be rederived in Section 3.6 by a different method. 2.5 Effects of anisotropy Our next task consists of allowing the confining potential to be anisotropic. Isotropic potentials are useful theoretical constructs, but they are far from realistic: as we saw in Chapter 1, the first traps to create condensates at JILA and MIT [2,6] were already anisotropic, and their anisotropy was in fact of fundamental importance in identifying the condensates as such; since then, experimental setups have become, if anything, even less isotropic. We are moreover interested in studying two-dimensional systems; though some successful 2D experiments have been carried out in atomic hydrogen adsorbed on liquid 4 He [29], the only way of creating pure low-dimensional condensates consists of taking a trapped gas and increasing or decreasing one or more of the confining frequencies, effectively creating 29 1 0 1 0 φ0 (x), . . . , φ6 (x) 1 0 1 0 1 0 1 PSfrag replacements 0 1 0 0 1 2 3 4 5 6 x Figure 2.10. The first seven radially symmetric eigenfunctions (including the ground state) of the two-dimensional isotropic harmonic oscillator. The solid lines depict the exact eigenfunctions √ 2 from Eq. (A.10) of Appendix A, φn (x) ≡ hx | ni = e−x /2 Ln (x2 )/ π, while the open circles are the result of diagonalizing the angle-averaged density matrix of an ideal Bose gas of 1000 atoms at T = 0.8 Tc . The matrix was calculated using a 30 × 30 Gauss-Laguerre grid and is known only at those points; its values everywhere else can be found by interpolation. All eigenfunctions were normalized to be unity at x = 0. Shown with crosses are the first three eigenfunctions of an ideal gas of 1000 distinguishable atoms at the same temperature. It is somewhat amusing that the boson matrix gives more accurate results than its simpler distinguishable counterpart. 30 “pancake” and “cigar” geometries. These have already been produced in the laboratory [30] and have been the subject of various studies [55,56,66,67]. The fact that the trap becomes anisotropic forces us to modify the expressions derived above. Throughout this work we will concentrate on one possible squeezed geometry, that of a threedimensional trap whose z-frequency can be increased to create quasi-2D systems. We will use the frequency ω ≡ ωx = ωy as a reference to scale all energies and (through an unchanged x0 ) all lengths: in particular, we will resolve the position vector into transverse and z-components: x = ξ + η. The trap anisotropy will appear through the parameter λ ≡ ωz /ω ≥ 1, to which we will occasionally refer as the “compression ratio.” With these identifications, the Schrödinger equation for the harmonic oscillator becomes ·µ 1 ∂ − ξ ∂ξ µ ¶ ¶ µ ¶¸ ∂ 1 ∂2 ∂2 2 2 2 ξ − 2 φnmp ≡ (H̃ξ + H̃η )φnmp = ²nmp φnmp ; +ξ + − 2 +λ η ∂ξ ξ ∂ϕ2 ∂η (2.33) its solutions, displayed in Eq. (A.9) on page 116, are normalized products of isotropic 2D oscillator √ eigenfunctions in ξ and 1D eigenfunctions with η → η λ; the corresponding energy eigenvalues are given by ²nmp = 2(2n + |m| + λp) + (2 + λ). (2.34) An important consequence of (2.34) is that the ground-state energy now depends on λ and grows as the system is squeezed; it is far from negligible when λ becomes large. The chemical potential also increases in value, and tends to 2 + λ at zero temperature; this makes sense if we once again identify the chemical potential as the energy gained by the system when one adds a particle. The one-body partition function of a distinguishable system can be calculated by splitting the 3D Hamiltonian in Cartesian coördinates and evaluating three separate one-dimensional sums, resulting in Z1 = 1 csch2 β̃ csch λβ̃; 23 (2.35) in our scaled units, then, the inverse temperature for the compressed third dimension acquires a factor of λ. Armed with these results, it is not difficult to find an expression for the one-body diagonal density of the Bose gas: ñ(x) = √ ∞ λ X π 3/2 δ` e−(ξ 2 tanh `β̃+λη 2 tanh λ`β̃) , (2.36) `=1 where we have defined δ` ≡ e`β̃(µ̃−(2+λ)) (1 − e−4`β̃ )(1 − e−4λ`β̃ )1/2 31 ; (2.37) 1 PSfrag replacements N0 /N 0.8 0.6 0.4 0.2 0 0 100 104 1 (3) T /Tc 10 1.79 λ 1 2.5 Figure 2.11. Condensate fraction of a three-dimensional ideal gas in a trap of increasing anisotropy. The system contains 100 bosons; as the anisotropy parameter λ increases, the initially isotropic trap becomes progressively flattened. The dashed lines show the thermodynamic-limit expressions (an inverted cubic at the left end and a parabola at the right) predicted by Eq. (2.17) for two and three (2) (3) dimensions respectively. For N = 100 the equivalent 2D temperature is Tc ≈ 1.786 Tc . the off-diagonal elements are found by a straightforward extension of this procedure. The chemical potential is determined by an expression analogous to (2.11): N= ∞ ∞ X e`β̃ µ̃ e`β̃(µ̃−(2+λ)) 1 X = . 2 3 2 sinh `β̃ sinh λ`β̃ (1 − e−2`β̃ )2 (1 − e−2λ`β̃ ) `=1 `=1 (2.38) Equations (2.36) and (2.38) are equivalent to (2.10) and (2.11) and specify the anisotropic system completely. Figure 2.11 shows how the condensate fraction evolves as the anisotropy increases. Beyond a certain value of λ, the condensate fraction acquires the parabolic shape, prescribed by Eq. (2.17), that we had already seen in Fig. 2.4 for the strictly two-dimensional gas: a crossover to lower dimensionality has taken place. It is important to note that the tightly squeezed 3D system behaves like a 2D gas but at a lower equivalent temperature; in fact, the transition temperatures are related through Tc(2) ≈ 0.829N 1/6 Tc(3) . (2.39) This result is not too surprising, and is a consequence of Einstein’s condition: the condensation, we recall, is prompted by a proper combination of temperature and density, and by compressing the 32 gas we effectively lower the temperature.5 In particular, it should be possible take a 3D gas at a (3) temperature above Tc and coax it into condensing by compressing it. The crossover in dimensionality should also be apparent in the density profiles. Now, our traps have ceased to be isotropic, and as a consequence the radial densities that we have been studying until now no longer make sense. A more meaningful quantity is the actual appearance of the gas when looked at from one direction; in other words, we will consider integrals of the density over one coordinate, which are equivalent to the “optical” or “column” densities measured by experimentalists. When dealing with a system of increasing anisotropy, one can either measure the optical density along the squeezed direction or along one of the directions that are left untouched. In the first case, the measurable quantity is the surface density, given by σ̃(ξ) ≡ Z ∞ 2 1 X e`β̃(µ̃−(2+λ)) e−ξ tanh `β̃ ; −4` β̃ 2π (1 − e ) sinh λ`β̃ `=1 ∞ ñ(x) dη = −∞ (2.40) as the anisotropy of the system grows, the surface density should be described progressively better by a two-dimensional isotropic density profile. In particular, the three-dimensional raw density √ N0 λ −(ξ2 +λη2 ) ñ0 (x) = 3/2 e π (2.41) is clearly anisotropic, but the corresponding surface density adopts an expression identical to the condensate density in two dimensions, regardless of the degree of anisotropy: σ̃0 (ξ) = N0 −ξ2 e ; π (2.42) its amplitude, given by the condensate fraction, will of course change. Figure 2.12 displays the behavior with increasing gas anisotropy of this surface density and of the corresponding surface number density Ñ (ξ) ≡ 2πξ σ̃(ξ) such that N= Z ∞ Ñ (ξ) dξ. (2.43) 0 Note that the isotropic system is uncondensed at this temperature and develops a condensate upon compression; the thermal density, in turn, starts as a Gaussian and eventually takes the shape seen before, including the delocalization hump. 5 Using the semiclassical approximation we find a density ñ(0) ≈ (t/4π) 3/2 g3/2 (eβ̃(µ̃−(2+λ) ) at the center of the trap. As λ grows, we see that it becomes easier to reach a point where Einstein’s criterion λ3T ñ(0) = ζ(3/2) is fulfilled. 33 0.15 σ̃(ξ)/N PSfrag replacements 0.1 0.05 0 0 2 100 4 ξ 10 λ 6 8 1 0.4 PSfrag replacements Ñ (ξ)/N 0.3 0.2 0.1 0 0 2 100 4 ξ 10 λ 6 8 1 Figure 2.12. Surface density and surface number density of a three-dimensional ideal Bose gas in a (3) trap of increasing anisotropy. In every case N = 100 and T = 1.4 Tc . As usual, the total densities are resolved into their condensed and uncondensed portions. As λ increases, the gas develops a condensate. After it appears, the condensate density keeps the same shape throughout the process; the thermal density develops a hump. 34 As we said above, we can also view the system from one of the uncompressed directions, and in this case we also have a meaningful measurable quantity in the aspect ratio of the system, which results from a straightforward calculation:6 q P p ∞ λ `=1 δ` coth2 `β̃ coth1/2 λ`β̃ hξ 2 i p≡ p = qP , ∞ 2hη 2 i δ coth `β̃ coth3/2 λ`β̃ `=1 (2.44) ` where δ` was defined in Eq. (2.37). In the high-temperature semiclassical approximation this becomes p = λ. At low temperatures, where the condensate dominates, the aspect ratio is smaller; in fact, it √ is evident from (2.41) that p → λ. This aspect ratio will reappear in the next chapter, and will be of help in Chapter 5 as a check on our Monte Carlo simulations. 2.6 Summary We begin this work by considering the ideal gas, a system worth studying for many reasons: it incorporates the inhomogeneity and the finite size of the systems studied in experiments, it is known to undergo a condensation in two dimensions, and it can be solved exactly for any finite number of atoms. This last feature lets us study the transition itself in detail, something that we cannot do for the interacting gas (except via Monte Carlo simulations), and also enables us to introduce and sharpen the tools that we will be using throughout the rest of the thesis. We introduce a semiclassical self-consistent approximation that, though not really necessary here, gives us a feel for the temperature scales involved. We study (and rule out) the possibility of a smeared condensation by examining the eigenvalues of the off-diagonal one-body density matrix. Finally, we squeeze a three-dimensional system in order to see its crossover into the two-dimensional régime. 6 The factor of 2 in the definition of p might look a bit confusing at first and deserves some justification. It has been introduced to counterbalance the fact that, while η can be negative, ξ is radial and therefore always positive: the 2 ensures that p = 1 in an isotropic trap. 35 CHAPTER 3 MEAN-FIELD THEORY OF INTERACTING SYSTEMS In his well-known papers, Einstein has already discussed a peculiar condensation phenomenon of the ‘Bose-Einstein’ gas; but in the course of time the degeneracy of the Bose-Einstein gas has rather got the reputation of having only a purely imaginary existence. —F. London [19] 3.1 Introduction In the preceding chapter we saw some of the effects that Bose-Einstein statistics and the presence of a trapping potential have on a collection of atoms. The third essential ingredient for a realistic description of a Bose-Einstein condensate is the introduction of interactions between the particles that compose it and between those and the rest of the system. Needless to say, an interacting system is a much more complicated entity than an ideal gas, and its description is consequently more involved. Various schemes of differing degrees of sophistication have been introduced to describe the behavior and characteristics of a Bose-condensed system, and among them mean-field theory has earned a preëminent place because of its relative simplicity, its predictive power, and the crystal-clear interpretability of its results. Even in its simplest versions, mean-field theory exhibits a remarkable resilience: not only has its agreement with experiment been excellent so far [68], it also routinely passes high-precision tests [69] and makes correct predictions even beyond its expected range of validity [70,71]. A majority of the theoretical inquiry into BEC [54] is based on mean-field theory, and this work will continue that tradition. Our treatment of the mean-field theory of interacting systems will roughly parallel that of the ideal gas in Chapter 2. After introducing the model interparticle interaction that we will use, we begin by reviewing an “exact” theory, the Hartree-Fock-Bogoliubov (HFB) equations, based on a model first considered by Bogoliubov [72] for the homogeneous imperfect Bose gas and later adapted to describe the dilute trapped atomic gases in which BEC was created [73,74]. We digress for a while to treat systems at zero temperature by means of the famous GrossPitaevskiı̆ (GP) equation [75,76] and study their approach to the thermodynamic limit in what 36 is known as the Thomas-Fermi limit [77]. We will introduce a variational Ansatz based on the Thomas-Fermi wavefunction to study the zero-temperature gas in a trap of increasing anisotropy; this will enable us to provide a simplified description of the compressed gas and will provide us with a criterion to find the conditions under which a crossover in dimensionality occurs in the presence of interactions. We then introduce a semiclassical model in which the HFB equations become easier to interpret and to solve. The Hartree-Fock-Bogoliubov theory prescribes separate (though coupled) equations for the condensate and for the surrounding thermal cloud, so we will have to resort to a self-consistent procedure. We will conclude by mentioning some of the methods that we have used to solve these equations and showing some results. 3.2 The effective interaction The success of mean-field theory in describing Bose-Einstein condensates is due in part to the characteristics of the system under study. The trapped gases in which Bose-Einstein condensation has been created remain in a gaseous state at such low temperatures because they are so dilute— and, paradoxically, they are at such low temperatures—that three-body collisions and other more complicated processes are extremely rare. Moreover, the low-energy collisions that dominate the interactions can be described quite well in terms of a single parameter, the s-wave scattering length a [78], also known as the “hard-core radius.” In Chapter 5 we will indeed assume that the boson gas is composed of hard spheres or rods of radius a; this approximation, however, is convenient for Monte Carlo simulations but is too complicated for our present purposes. Instead, we will use a popular model in which the interatomic potential is described by a contact interaction (also called, somewhat inaccurately, a pseudopotential) of the form [79] V (r1 − r2 ) = g δ (σ) (r1 − r2 ). (3.1) In three dimensions the constant g is directly proportional to a: g= 4π~2 a = (kB T )λ2T (2a); m (3.2) in the dimensionless, dimension-explicit units that we are using, the equivalent parameter is the “interaction strength” or “coupling constant” γ, related to g via g ≡ 12 ~ωxσ0 γ, and Eq. (3.2) adopts 37 the particularly transparent form γ = 8πa/x0 . The coupling constant, and the isotropy of the interaction, should be the same even when the atoms are confined by an anisotropic trap. In two dimensions, the interpretation of g is more subtle: the “constant” turns out to be (logarithmically) dependent on the relative momentum of the colliding atoms [35,80]. Mean-field expressions for the 2D interaction strength that take this difficulty into account [67,81,82] have been derived and used in the literature, and we will discuss them in Section 3.5. In general, though, we have found that these more elaborate expressions have a rather negligible effect on the density profile and condensate fraction of the system, and in what follows we will usually take γ to be constant. More importantly, since we want to know the extent to which a squeezed 3D trapped system can be described by a 2D gas as the confining potential becomes increasingly anisotropic, we need to be able to relate the two-dimensional constant to its three-dimensional, experimentally known counterpart. We will turn to this question in Section 3.5. 3.3 The Hartree-Fock-Bogoliubov equations The introduction of the potential (3.1) greatly simplifies the appearance of the many-body grandcanonical Hamiltonian, which in terms of second-quantized field operators becomes H̃ = Z dσx Ψ† (x)Λ̃Ψ(x) + γ 2 Z dσx Ψ† (x)Ψ† (x)Ψ(x)Ψ(x), (3.3) ˜ 2σ + x2 − µ̃ is the one-body Hamiltonian. where Λ̃ ≡ −∇ We will be working below the transition temperature, where the condensate is the most heavily populated state, and so we will start by defining the condensate wavefunction Φ̃ as the ensemble average of the field operator Ψ. The rest of the system will then be described by the fluctuation of the operator about this average: Ψ = hΨi + ψ̃ = Φ̃ + ψ̃. (3.4) It obviously follows that ψ̃, the fluctuation or noncondensate operator, obeys hψ̃i = 0. The condensate wavefunction Φ̃, like that of the ideal gas, is real and positive everywhere and prescribes a condensate density ñ0 (x) = Φ̃2 ; moreover, it is a function, not an operator, which reflects the imposition of long-range order in the system and restricts the grand canonical ensemble in order to prevent unphysically large fluctuations in the condensate number [59,83]. When we insert the decomposition (3.4) into Eq. (3.3) the integrands become Ψ† Λ̃Ψ = Φ̃∗ Λ̃Φ̃ + Φ̃∗ Λ̃ψ̃ + ψ̃ † Λ̃Φ̃ + ψ̃ † Λ̃ψ̃, 38 (3.5) Ψ† Ψ† ΨΨ = |Φ̃|4 + 2|Φ̃|2 (Φ̃∗ ψ̃ + Φ̃ψ̃ † ) + 4|Φ̃|2 ψ̃ † ψ̃ +(Φ̃∗2 ψ̃ ψ̃ + Φ̃2 ψ̃ † ψ̃ † ) + 2ψ̃ † (Φ̃ψ̃ † + Φ̃∗ ψ̃)ψ̃ + ψ̃ † ψ̃ † ψ̃ ψ̃. (3.6) At very low temperatures, when the condensate dominates the system, it is sufficient to neglect the terms in (3.6) that contain more than two noncondensate operators [73,78,81,84]; this fruitful approach allows one to study the time-dependent behavior of a condensed system [54], yielding, among other results, analytic solutions for its excitation spectrum [85,86] and a description of quantized vortices [73,78,87]. However, we are interested in studying trapped gases at finite temperatures, and for that reason we will use a scheme that includes these neglected terms in approximate form but produces equations that mimic the second-order ones and have a similar method of solution. This “self-consistent mean-field approximation” [74,88] takes the third- and fourth-order terms of (3.6) and expresses them as combinations of first- and second-order terms, thus effecting a “quadratization” of the Hamiltonian [89]; the coefficients are ensemble averages of binary contractions of the field operators weighted so that each approximate term will have the same mean value as the corresponding exact term. Moreover, we invoke the Popov approximation, which consists of neglecting the “anomalous” averages1 hψ̃ † ψ̃ † i and hψ̃ ψ̃i and leaving only hψ̃ † ψ̃i ≡ ñ0 , the noncondensate density. The high-order combinations2 then become ψ̃ † ψ̃ ψ̃ ≈ 2hψ̃ † ψ̃iψ̃ ≡ 2ñ0 ψ̃, ψ̃ † ψ̃ † ψ̃ ≈ 2hψ̃ † ψ̃iψ̃ † ≡ 2ñ0 ψ̃ † , ψ̃ † ψ̃ † ψ̃ ψ̃ ≈ 4hψ̃ † ψ̃iψ̃ † ψ̃ − 2hψ̃ † ψ̃ihψ̃ † ψ̃i ≡ 4ñ0 ψ̃ † ψ̃ − 2ñ02 , (3.7) and we obtain a Hamiltonian of the form K0 + K1 + K1† + K2 , where the subscripts stand for the number of field operators Ψ, Ψ† contained in each term. The first one, K0 = Z 1 d x Φ̃(Λ̃ + γ ñ0 )Φ̃ − γ 2 σ Z dσx ñ02 , (3.8) 1 These averages describe, respectively, a process through which two condensate atoms scatter each other and end up in the thermal cloud, and a process through which two thermal atoms collide and “fall” into the condensate [84]. These processes, while not unphysical, are expected to be quite uncommon [88]; mean-field calculations that include them [70] confirm their relative rarity even at near-transition temperatures. 2 The various factors of 2 that appear reflect the fact that “direct” and “exchange” terms are identical for the zero-range interaction that we are considering [74]. 39 gives the energy of the condensate; the term at the end ensures that the interaction energy is not overcounted (see Section 4.5 for an alternative explanation). The terms linear in ψ̃ † and ψ̃, K1 = Z dσx ψ̃ † (Λ̃ + γ(ñ0 + 2ñ0 ))Φ̃ (3.9) and its Hermitian conjugate, are required to vanish exactly if we are to have h ψ̃i = 0; thus Φ̃ must obey the generalized Gross-Pitaevskiı̆ equation [75,76] ˜ 2 Φ̃ + (µ̃ − x2 )Φ̃ − γ(ñ0 + 2ñ0 )Φ̃ = 0, ∇ σ (3.10) whose interpretation and solution we will discuss in Section 3.4. The quadratic term K2 = Z ¸ Z · ¸ · 1 1 dσx ψ̃ † (Λ̃ + 2γ ñ)ψ̃ + γ ñ0 (ψ̃ † ψ̃ † + ψ̃ ψ̃) ≡ dσx ψ̃ † Lψ̃ + γ ñ0 (ψ̃ † ψ̃ † + ψ̃ ψ̃) , (3.11) 2 2 as we will presently prove, can be diagonalized by expanding, à la Bogoliubov, in a set of Bose creation and annihilation operators ψ̃ = X j (uj αj − vj∗ αj† ) [αj , αj†0 ] = δj 0 j . with (3.12) We begin by inserting (3.12) into (3.11). The operator L introduced in (3.11) is clearly Hermitian (and, moreover, real), and consequently obeys [62] Z d σ x u∗j Lvk = Z dσx vk Lu∗j (3.13) for any pair of functions uj , vk . This enables us to find, after some rearrangement, that K2 = 1X 2 jk £ Z dσx (vj (Lvk∗ − γ ñ0 u∗k ) + vk∗ (Lvj − γ ñ0 uj )) αj αk† + (u∗j (Luk − γ ñ0 vk ) + uk (Lu∗j − γ ñ0 vj∗ )) αj† αk + ¤ (vj (Luk − γ ñ0 vk ) + uk (Lvj − γ ñ0 uj )) αj αk + (u∗j (Lvk∗ − γ ñ0 u∗k ) + vk∗ (Lu∗j − γ ñ0 vj∗ )) αj† αk† ; (3.14) if we now assume that the expansion coefficients uj , vj obey the coupled eigenvalue equations [73, 74,88] à Λ̃ + 2γ ñ −γ ñ0 −γ ñ0 Λ̃ + 2γ ñ !à 40 uj vj ! = ²j à uj −vj ! (3.15) we obtain K2 = 1X 2 jk £ Z dσx ¤ −(²j + ²∗k ) uj vk∗ αj αk† + (²∗j + ²k ) u∗j uk αj† αk + (²k − ²j ) uk vj αj αk + (²∗j − ²∗k ) u∗j vk∗ αj† αk† . (3.16) We next introduce [78] the matrices M = (Λ̃ + 2γ ñ) − γ ñ0 σ x and Uj = " uj vj # , (3.17) where σ x is a Pauli matrix, and the inner product hi | ji ≡ Z d x U†i σ z Uj σ = Z dσx (u∗i uj − vi∗ vj ), (3.18) which allow us to rewrite (3.15) as MUk = ²k σ z Uk ; (3.19) the matrix M is once again Hermitian, so we can multiply each side of (3.19) by U †j and prove that ²k U†j σ z Uk = U†j MUk = (MUk )† Uj = (MUj )† Uk = ²∗j U†j σ z Uk , (3.20) where we used the Hermiticity of M in the first step and switched the indices j and k in the second (this is valid only when integrated with respect to dσx and summed over j and k); this leads to (²∗j − ²k )hj | ki = 0, (3.21) which tells us that the eigenfunctions uj , vj are orthogonal under the inner product (3.18) (in fact, we can choose the normalization hj | ki = δjk ) and that the eigenvalues ²j are real. In a similar fashion we can prove that (²j + ²k ) Z dσx (uj vk − uk vj ) = 0. (3.22) This last result immediately shows us that the off-diagonal terms in Eq. (3.16) vanish: those with j = k vanish at once, while every term with j 6= k will combine with its “mirror image” to add up 41 to zero. Thus the Hamiltonian (3.11) is diagonal. We can now use the commutation relation (3.12) to show that K2 = 1X (²j + ²k ) 2 jk Z dσx (u∗j uk − vj∗ vk ) αj† αk − X ²j j Z dσx |vj |2 ≈ X ²j αj† αj . (3.23) j The neglected term describes the zero-temperature depletion of the condensate, which in trapped Bose gases accounts typically for less than 1% of the number of atoms [88]. We also neglect it when we calculate the noncondensate diagonal density, given by ñ0 (x) = hψ̃ † ψ̃i ≈ Xh jk i X¡ 2 2¢ u∗j uk hαj† αk i + vj vk∗ hαk† αj i = |uj | + |vj | f˜j (3.24) j with f˜j = hαj† αj i = 1 eβ̃²j −1 . (3.25) The form of (3.25) follows from Wick’s theorem [84,90] (see Appendix B) and reflects the fact that the “quasiparticles” created by the αj† are bosons; it could have also been obtained by minimizing the free energy H̃ − tS̃ (see Section 4.2.3). Equations (3.10), (3.15), and (3.24), along with the conservation of particles imposed by N = R dσx ñ, where ñ = ñ0 + ñ0 , are the Hartree-Fock-Bogoliubov equations in the Popov approximation; they form a complete set that can be solved self-consistently for the densities at any temperature and at every point in space. Although the full set of equations has been successfully solved in three [70,91,92] and (very recently) two dimensions [93], we will concentrate on a semiclassical approximation that turns a discrete set of variables into a continuum—a method not unlike the one discussed in Chapter 2. But first we digress to discuss the behavior of trapped Bose-Einstein systems at zero temperature. 3.4 The Gross-Pitaevskiı̆ equation and the Thomas-Fermi limit The Gross-Pitaevskiı̆ (GP) equation (3.10) has a transparent physical interpretation. To discuss it, let us write it again in a slightly different form: ˜ 2σ Φ̃ + x2 Φ̃ + γ(ñ0 + 2ñ0 )Φ̃ = µ̃Φ̃. −∇ (3.26) We have already remarked that the Bose-Einstein condensate can be described by a single wavefunction, which we have labelled Φ̃. The wavefunction should obey a Schrödinger equation with an 42 external potential, and that is precisely what the first two terms in (3.26) exhibit. In the GP approximation, the interparticle interactions become an additional local potential that has two components. The second one, with the factor of 2 that we already discussed in a footnote on page 39, describes the interaction of the condensate with the thermal cloud and will be studied in detail in subsequent sections. For now, however, we will concentrate on the zero-temperature case, where this term disappears and the equation becomes ˜ 2 Φ̃ + x2 Φ̃ + γ ñ0 Φ̃ = µ̃Φ̃. −∇ σ (3.27) The mean-field potential that remains describes the interaction of the condensate with itself and makes the equation nonlinear at every temperature; this nonlinearity was originally introduced by Gross [75] and Pitaevskiı̆ [76] to suppress an unphysical overdependence on boundary conditions exhibited by the ideal Bose gas [14]. The GP equation is the Euler-Lagrange equation that minimizes the functional J[Φ̃] = Z £ ¤ ˜ Φ̃)2 + x2 Φ̃2 + 1 γ Φ̃4 dσx (∇ 2 (3.28) subject to the constraint that the number of atoms is fixed: N= Z dσx Φ̃2 ; (3.29) in other words, it minimizes the thermodynamic potential H̃ − µ̃N , and the corresponding Lagrange multiplier, which appears in (3.27) as the eigenvalue, can be identified as the chemical potential of the system (and not the energy per atom, as we would have if the equation were linear). We initially attempted to solve the two-dimensional GP equation by treating it as an initial-value √ problem [94,95], a method that we soon abandoned. To that end, we introduced Φ ≡ Φ̃/ N , whose unit normalization made it easier to display and compare solutions, and rewrote the equation as 2 ˜ 2 Φ + (µ̃ − x2 )Φ − N γΦ3 = d Φ + 1 dΦ + (µ̃ − x2 )Φ − N γΦ3 = 0 ∇ 2 dx2 x dx (3.30) with the boundary conditions ¯ dΦ ¯¯ =0 dx ¯x=0 and Z d2x Φ2 (x) = 1. (3.31) To solve the equation for a given set of parameters, we would start by making initial guesses for A (the amplitude of the wavefunction at x = 0) and µ̃, which, depending on the parameters, were 43 0.6 PSfrag replacements Φ(x), ΦTF (x) 0.4 0.2 0 0 2 x 4 6 Figure 3.1. Wavefunctions of two-dimensional isotropic trapped Bose gases at zero temperature. The full lines show the solutions (normalized to unity) of the Gross-Pitaevskiı̆ equation for (from top to bottom at x = 0) N γ = 10, 100, and 1000, while the dashed lines display the corresponding Thomas-Fermi wavefunctions. The Thomas-Fermi limit provides an adequate description of a system over most of its extent for N γ as low as 100. The dash-dotted line displays the noninteracting ground state. given by either the noninteracting ground state or the Thomas-Fermi wavefunction discussed below. This initial condition was then evolved using a standard differential-equation solver—we obtained the same results using a fourth-order Runge-Kutta integrator and a predictor-corrector algorithm— until the wave function became negative (due to µ̃ being too large) or attained a minimum at a point where it was finite (when µ̃ was too small). We would then readjust µ̃ to counter this effect and iterate the procedure until we found a function with the right behavior (that is, vanishing with zero derivative) at “infinity” (which we usually took to be in the neighborhood of x = 6). The resulting function would have the right shape but not the right size, so we had to restart the process with a new initial amplitude until we reached self-consistency, at which point the chemical potential was “accurate” to 12 or more figures. Figure 3.1 shows the two-dimensional solutions that we obtained by this method. The noticeable increase in the condensate radius caused by the interactions, as well as the lowering of the central density, are present here, just like in three dimensions [54,87]. (Shortly after we found these solutions, a paper [95] appeared that had two-dimensional results consistent with ours.) This process of solution, however, is extremely slow and laborious, especially for the larger values of N γ that we 44 want to study (N γ ≈ 1080 for 10000 rubidium atoms): each of the curves displayed in Fig. 3.1 took more than a day to generate. For sizable values of N γ, though, we can use a much simpler model that nevertheless gives a good description of the condensed system at zero temperature (except for a narrow region around the condensate edge). Alongside the exact solutions just discussed we display the Thomas-Fermi wavefunctions [77,96], Φ̃2TF = ñTF = 1 (µ̃ − x2 ) Θ(µ̃ − x2 ), γ (3.32) which result from neglecting the kinetic energy of the condensate. (Expression (3.32) is valid for any number of dimensions.) The Heaviside step functions in (3.32) are there to ensure that the wavefunctions are everywhere real and positive, as required (see the discussion following Eq. (3.4)), 1/2 and force the densities to lie within the condensate radii RTF = µ̃TF . The chemical potential, in turn, is fixed by normalization: µ̃TF = · σ(σ + 2)Γ(σ/2) Nγ 4π σ/2 ¸2/(σ+2) , (3.33) where Γ(x) is the usual gamma function. For the particular cases of two and three dimensions, (2) µ̃TF = µ 2 Nγ π ¶1/2 (3) and µ̃TF = µ 15 Nγ 8π ¶2/5 . (3.34) The energy per atom can be found by inserting (3.32) into (3.28) and integrating. Since the system at zero temperature has no entropy, the energy found is also the free energy per atom and is related to the chemical potential through σ+2 ÃTF = µ̃TF ; N σ+4 (3.35) r (3.36) in two dimensions we obtain 2 ÃTF = N 3 2p N γ. π The zero-temperature GP equation and the Thomas-Fermi limit already agree very well with 3D experiments carried out at finite temperature. We have already mentioned in Chapter 2 that experimentalists use the Thomas-Fermi limit as a first approximation; indeed, they usually fit their observed density profiles with Thomas-Fermi functions to find how many atoms they have in a trap [30]. On the other hand, the first Monte Carlo simulations of trapped bosons were checked by finding the condensate fraction appropriate at a given temperature (using methods that we will 45 I y 6 F012 ¾ 61 2 F12 1 ¾ 6 ẑ 2 F0ext ωz ¿ ωz0 Fext Figure 3.2. Typical forces exerted on an atom in a trap of increasing anisotropy. As the trap is compressed more and more, the z-component of a typical interatomic force F 12 becomes negligible compared to the external force Fext even if the actual interaction is large. (To avoid clutter we have omitted the transversal components of the trapping force.) explain later), solving Eq. (3.27) with that number as the total N , and comparing the results obtained [44]. The Thomas-Fermi should then give reasonable results for systems in traps of varying anisotropy, and to that topic we now turn. 3.5 Anisotropic systems at zero temperature We once again employ the specialized notation of Section 2.5, in which we resolve the position vector into its transverse and z-components: x = ξ + η. A glance at Eq. (2.33) makes it clear that the Hamiltonian for the harmonic oscillator—and hence the Gross-Pitaevskiı̆ Hamiltonian—acquires a factor of λ2 in the z-direction. As we mentioned in Sec. 3.2, we expect that the interaction between atoms will still be isotropic and that the coupling constant γ = γ (3) will remain unaffected. Under these conditions, the Thomas-Fermi density profile becomes Φ̃2TF = 1 (µ̃ − ξ 2 − λ2 η 2 ) Θ(µ̃ − ξ 2 − λ2 η 2 ) γ (3.37) (3) and the chemical potential is simply µ̃ = λ2/5 µ̃TF . The step function in Eq. (3.37) again ensures that the density is positive everywhere, and this requirement prescribes values for the condensate radii in all directions and a corresponding aspect ratio p = λ, which, as would be intuitively expected, grows faster with increasing λ than it does for the ideal gas at zero temperature (see our discussion following Eq. (2.44)). Beyond a certain degree of anisotropy, however, the Thomas-Fermi wavefunction becomes inadequate as a description of the system. The simple classical force diagram in Fig. 3.2 suggests an explanation: The extreme anisotropy lines up the atoms in a narrow spatial band, and the component 46 in the compressed direction of the force exerted on a given atom by any other one becomes negligible, regardless of the actual magnitude of the interaction, when compared to the now-much-larger trapping force. But harmonic forces as they get larger cause systems to vibrate faster, thus increasing the kinetic energy and making it important again. In the squeezed direction, then, the atoms tend to behave like noninteracting trapped particles. Now, from Eq. (2.34) we can see that, as λ grows, so does the gap between the ground state and the first excited state; this gap eventually becomes insurmountable, and at that point the atoms will be effectively frozen in the harmonic-oscillator ground state. When that happens, the system has reached the quasi-2D régime [66,81]. We can now estimate a condition for this “régime change.” The crossover from three-dimensional to quasi-2D behavior occurs when the confinement energy in the z-direction, ~ω z , becomes larger than the chemical potential given right below Eq. (3.37): 1 1 ~ωz = ~λω ≥ µ = ~ω µ̃ = 2 2 µ a 15N λ x0 ¶2/5 (3.38) or [30] 152 λ ≥ 5 2 3 µ a x0 ¶2 225 N = 32 2 µ γ (3) 8π ¶2 N 2. (3.39) To describe a condensate in the quasi-2D régime we can introduce an Ansatz for the condensate wavefunction consisting of a Thomas-Fermi profile in the radial direction and an ideal- (though trapped-) gas profile along z: Φ = A (R2 − ξ 2 )1/2 Θ(R2 − ξ 2 ) e−λη 2 /2 , (3.40) where the constant A can be found by normalization: Φ= r √ 2 π 3/2 2 N (R2 − ξ 2 )1/2 Θ(R2 − ξ 2 ) e−λη /2 ; 2 R (3.41) similar proposals have been introduced previously [66,77]. The “Thomas-Fermi radius” R can be found by imposing that the wavefunction minimize the functional J[Φ] = Z ¤ £ ˜ 2 + Φ(ξ 2 + λ2 η 2 )Φ + γ Φ4 , d3x (∇Φ) 2 (3.42) with the understanding that the gradient is just the derivative with respect to η. Insertion of (3.41) into (3.42) and straightforward integration yield 47 1 1 1 2N J = λ + R2 + λ + N 2 3 2 3π µ λ 2π ¶1/2 γ , R2 (3.43) which can be minimized with respect to R to give R4 = µ 2N π λ 2π ¶1/2 γ; (3.44) note that, even in its reduced form, the gradient term in (3.42) did not contribute to this expression and could have been dropped altogether. The surface density defined in Eq. (2.40) is readily found to be 1 σ̃(ξ) = N γ (3) µ λ 2π ¶1/2 " 2N π µ λ 2π ¶1/2 γ (3) #1/2 − ξ2 , (3.45) with a step function implied, and coincides with the two-dimensional Thomas-Fermi density given by Eqs. (3.32) and (3.34) if we identify γ (2) ≡ µ λ 2π ¶1/2 γ (3) , (3.46) the key result of this section. This expression for the equivalent two-dimensional coupling constant is consistent with others found in the literature [97–99] and represents the first-order approximation to the more elaborate expression [39,67,81,82] γ (2) = µ λ 2π ¶1/2 γ (3) à 1+ µ λ 2π ¶1/2 γ (3) log 8π µ λ πq 2 x20 ¶!−1 , (3.47) where q is the relative momentum of the colliding atoms in the equivalent two-body problem. There is no consensus as to what should replace the argument of the logarithm in the N -body problem. The authors of Ref. 81 suggest to transform q 2 x20 → µ̃ and then identify µ̃ → γ (2) ñ0 ; others either interpret this as a self-consistent equation for γ (2) [82], replace µ̃ → µ̃ − x2 − γ (2) ñ0 [100], or offer an expression like that of Ref. 81 but with a different coefficient and absolute-value signs around the logarithm itself [67]. In general, however, we found that these coupling constants predict results that differ very slightly from those of Eq. (3.46). Finally, our variational wavefunction predicts an aspect ratio that does not increase as quickly with λ as the Thomas-Fermi expression but still does so faster than the ideal-gas result: p µ ¶3/8 λ5/8 2 (N γ (3) )1/4 √ . p≡ p = π 3 2hη 2 i hξ 2 i 48 (3.48) 3.6 Finite temperatures: The semiclassical approximation In the last section we saw that the condensate is well described by the Gross-Pitaevskiı̆ equation; moreover, Section 3.3 summarizes the prescription for the complete treatment of the system. We will soon put this prescription to work, but before we do so we will study the simplified scheme that we will use from now on to treat the noncondensed particles. As we saw in Section 2.3, it is possible to simplify the more nearly exact equations using an approximate semiclassical scheme that will have a more transparent physical interpretation than the exact formalism. Though there we derived our results in a different manner, it turns out that they could have been found by solving the harmonic-oscillator equation in the WKB approximation [46,88]. We carry out the approximation by assuming that kB T À ~ω (a condition that, as Table 3.1 shows, is reasonably fulfilled in experimental situations) and consequently that the eigenenergies ² j form a continuous spectrum; the functions uj and vj can then be taken to be of the form uj → ueiφ(x) , vj → veiφ(x) ; (3.49) the phase φ has to be the same to ensure that both equations of (3.15) are compatible, and u and v are assumed to be smooth functions. By constructing the appropriate probability flux [46] we are led ˜ to identify the gradient of this phase as an excitation quasimomentum [88]: κ ≡ ∇φ. Furthermore, we can use the vector identity ∇ · (ψa) = a · ∇ψ + ψ∇ · a and temporarily put back all the factors of ~ (though keeping the notation we have been using for scaled variables) to see that the Laplacian term in the first equation of (3.15) is − · ¸ ~2 2 iφ/~ ~2 2 i~ i~ 1 ∇ ue = − ∇ u − ∇u · ∇φ − u∇2 φ + (∇φ)2 eiφ/~ . 2m 2m m 2m 2m (3.50) In the semiclassical approximation we make ~ → 0 (or, equivalently, neglect all derivatives of u and v and second derivatives of φ, in what is also called the local-density approximation [93,101]), and the Table 3.1. Typical temperatures and frequencies found in experiments. The frequencies displayed here correspond in each case to the harmonic mean of the three trapping frequencies, ω ≡ (ωx ωy ωz )1/3 . The factors kB T /~ω oscillate between 13 and 100. Trap [Reference] JILA trap [6] MIT trap [2] 2D condensates [30] Temperature (nK) 170 2000 40 Mean frequency (Hz) 169.7 415.6 61.9 49 kB T (J) 2.35 × 10−30 2.76 × 10−29 5.52 × 10−31 ~ω (J) 1.12 × 10−31 2.76 × 10−31 4.10 × 10−32 coupled equations (3.15) become [54,88] (κ2 + x2 + 2γ ñ − µ̃) u(x, κ) − γ ñ0 v(x, κ) = ²(x, κ) u(x, κ) −γ ñ0 u(x, κ) + (κ2 + x2 + 2γ ñ − µ̃) v(x, κ) = −²(x, κ) v(x, κ), (3.51) with a Bogoliubov-like energy spectrum given by [102] ²2 = (κ2 + x2 + 2γ ñ − µ̃)2 − γ 2 ñ20 ≡ Λ2 − γ 2 ñ20 . (3.52) In order to maintain consistency with the normalization (3.18), we impose the condition u 2 − v 2 = 1 and obtain u2 = Λ+² 2² and v2 = Λ−² ; 2² (3.53) the density of excited particles takes the form ñ0 (x) = 1 (2π)σ Z dσκ (u2 + v 2 )f˜(κ, x) = 1 (2π)σ Z dσκ Λ˜ f (κ, x) ² (3.54) with f˜(κ, x) given by Eq. (3.25) after replacing ²j with ² given by (3.52). These equations, together with (3.10) and the imposition that N= Z dσx (ñ0 + ñ0 ), (3.55) form a self-consistent set of equations that can be solved to study the complete thermodynamics of the trapped system at any temperature, with no adjustable parameters, given N and a. As usual, we have neglected the zero-temperature depletion. The model can be further simplified by taking the Hartree-Fock approximation [83,96,103], in which we neglect the “quasihole” function v altogether. While Eq. (3.10) remains unchanged, equation (3.15) is now simply Λ̃ = ², the energy spectrum (3.52) turns into ² ≡ ²HF = κ2 + x2 + 2γ ñ(x), (3.56) and the thermal density becomes ñ0 (x) = 1 (2π)σ Z dσk f (x, κ) = 1 2π σ/2 (2π)σ Γ(σ/2) 50 Z ∞ 0 κσ−1 dκ e β̃(κ2 +x2 +2γ ñ−µ̃) −1 = µ t 4π ¶σ/2 gσ/2 (e−β̃(x 2 +2γ ñ−µ̃) ). (3.57) Equations (3.56) and (3.57) suggest that one can describe the thermal atoms as being acted upon by the effective Hamiltonian H̃eff = κ2 + x2 + 2γ ñ(x), (3.58) with a semiclassical kinetic energy that at this level of approximation commutes with the position operator and a potential energy that contains both the trapping potential and an effective local potential due to interactions. Indeed, we obtain Eq. (3.57) by invoking Eq. (2.9) with the new Hamiltonian: ñ0 (x) = hx| 1 eβ̃(H̃eff −µ̃) −1 |xi. (3.59) This identification also gives us an expression for the off-diagonal elements of the density matrix. From the definition ñ(x, x0 ) ≡ hΨ† (x)Ψ(x0 )i = Φ̃∗ (x)Φ̃(x0 ) + hψ̃ † (x)ψ̃(x0 )i, (3.60) where we once again neglected the anomalous averages, we obtain [73] ñ(x, x0 ) = Φ̃(x)Φ̃(x0 ) + ñ0 (x, x0 ) (3.61) with [61,89] ñ0 (x, x0 ) = µ t 4π ¶σ/2 X ∞ `=1 1 `σ/2 µ ¶ t 1 exp − (x − x0 )2 − `β̃( (x2 + x02 ) + γ(ñ(x) + ñ(x0 )) − µ̃) , 4` 2 (3.62) which results from neglecting the commutator of κ and x (which is valid for small β̃ [45]) and 2 0 2 using the free-particle off-diagonal density matrix [14,15]: hx|e−β̃κ |x0 i = e−(x−x ) /4β̃ /4π β̃. Equa- tion (3.62) will be used in the next section and in Chapter 5. 3.7 The interacting Bose gas in the Hartree-Fock approximation The solution of the semiclassical HFB equations is a self-consistent process like the one described earlier in Section 2.3: Initially, we solve the GP equation (3.10) with ñ0 = 0, just as we did in Section 3.4, and as a result we obtain initial values for the condensate density and the chemical 51 potential. These can then be fed into either Eq. (3.54) or Eq. (3.57) to obtain a thermal density, which we integrate (using, for example, the extended Simpson’s rule (2.30)) to yield the number of thermal atoms. The particle-conservation condition (3.55) then allows us to find a new value for the condensate number, to which the condensate density must now be normalized. At this stage we have values for ñ0 and ñ0 that we can insert into (3.10) to obtain a new self-consistent condensate density. All we need to do now is iterate the process until all quantities (in particular, the condensate fraction) stop changing. Most of the self-consistent procedure is straightforward enough to implement; the most difficult part by far is the solution of the Gross-Pitaevskiı̆ equation—a nonlinear eigenvalue problem with the mixture of Neumann and global boundary conditions exhibited in Eq. (3.31). In Section 3.4 we described the difficulties we encountered when we tried to solve the GP equation as an initial-value problem. For finite temperatures the situation was, predictably, even worse, since the density profiles had to be iterated to self-consistency—and this on top of the multiple iterations that we had to carry out to obtain a self-consistent condensate density at every step. At low temperatures, when the plain GP equation was already a good description of the condensate and was thus a reasonable starting point for the self-consistent calculation, some three iterations sufficed. At higher temperatures, the program needed about a week of full-time dedication in order to generate a single density profile. The agreement with other semiclassical results found in the literature [69,87,88], and even with the solutions of the full set of discrete equations in three dimensions [91], was encouraging, but the method was simply not viable, and we had to look for alternatives. The method we eventually used3 is based on the direct minimization of the functional [44,69] J[Φ̃] = subject to the constraint R Z £ ¤ ˜ 2 + Φ(x2 + 2γ ñ0 (x))Φ + γ Φ4 dσx (∇Φ) 2 (3.63) dσx Φ2 = 1. In an isotropic trap the problem is one-dimensional, and the gradient term becomes a plain derivative. The minimization is performed by setting a grid of n fixed abscissas x(n) j that represent the radial coördinate x; at the beginning we assign a set of ordinates f j ≡ Φ(x(n) j ), in terms of which (3.63) becomes a multivariate function, J = J[f 1 , . . . , fn ] ≡ J[f ], that can be minimized using standard optimization routines [63]. The resulting function is known at the x (n) j and can be evaluated everywhere else by interpolation. 3 We obtained equivalent results using the modified steepest-descents method of Ref. 87, which seems to be the most popular way of solving the problem [54,86,88]. In general, however, we found the minimization method to be superior in both efficiency and accuracy, and much easier to implement. 52 Now, a glance at Eq. (3.63) shows that interpolation is also essential during the minimization process, since for each evaluation of J we need to be able to calculate both derivatives and integrals of functions that we know only at a few points. Finite-difference schemes cannot be used, since any grid that would give reliable approximations makes the minimization impracticable. A much better approach, based on cubic-spline interpolation [44,69,104], yielded the results that we present (along with a more detailed exposition of the method) in Chapter 4: this lowered the time necessary to obtain a density profile from a week to a few hours, and, unlike the initial-value method described above, could run alone without having to be babysat. More recently, however, we developed a method based on spectral differentiation [64,65,105] and Gaussian quadrature [62,63] that we can consider, at least for our purposes, definitive. This scheme, like the one described in Section 2.4, uses as grid points the zeros of the nth Laguerre polynomial; the calculation of (3.63) is reduced (in 2D) to the matrix-vector product J[f ] = n ³ γfj4 ´ (n) x(n) 1 X (n) 2 0 (n) 2 2 (2πx(n) wj e j , j ) (Df )j + (xj fj ) + 2γ ñ (xj )fj + ∆ j=1 2∆ where ∆≡ n X (n) 2 (n) xj (2πx(n) j ) f j wj e (3.64) (3.65) j=1 (n) is the normalization factor; the wj(n) exj are the Gauss-Laguerre weights and inverse weighting function [62,63,65] at the grid points and D is a differentiation matrix [64,65,105]. (These quantities are derived in detail in Appendix C.) Not only was this method more robust than that based on spline interpolants, it was orders of magnitude faster: each of the density profiles that we show below took about 15 seconds to generate. The results that we present in this section and in Chapter 5 were all found using this procedure. (Other authors [70,106] have recently used similar grids to solve problems related to BEC, but to our knowledge nobody has used them to minimize (3.63).) These results also correspond, in their entirety, to the Hartree-Fock approximation characterized by the energy spectrum (3.56). In general we found that, in 3D, the purportedly more accurate HFB approximation based on (3.52) converged for a quite restricted range of parameters and, when it did converge, gave essentially the same results as the Hartree-Fock scheme; in 2D we were unable to find a single instance where the HFB scheme worked (see Chapter 4). Initially we consider a three-dimensional gas of 104 atoms in an isotropic trap. Figure 3.3 on the next page shows the behavior of the condensate fraction. The dotted line is the thermodynamiclimit prediction from Eq. (2.17), while the dashed line is the finite-size prediction, Eq. (2.16), whose 53 1 0.8 N0 /N 0.6 PSfrag replacements 0.5 0.4 0.2 0 0 0.2 0.4 T /Tc 0.6 0.8 1 Figure 3.3. Condensate fraction of a three-dimensional Bose gas in the Hartree-Fock approximation. In this case we have N = 104 atoms in an isotropic trap. Refer to the text for a detailed explanation. correction to the critical temperature is shown as a square on the x-axis. The diamond on the x-axis exhibits the total (finite-size plus interaction) correction4 to the critical temperature [102,107,108], (3) δTc (3) Tc ≈ −0.73 N −1/3 − 1.33 γ N 1/6 . 8π (3.66) The dots represent the result of running the self-consistent Hartree-Fock program with γ ≈ 0, while the solid line corresponds to an interacting gas of rubidium atoms with γ = 8π × 0.0043. Finally, the solitary point with an error bar at T = 0.7 Tc is the result of a Monte Carlo simulation (to be described in Chapter 5). The plot shows that interaction effects are greater than finite-size effects, and that the latter are already built into the Hartree-Fock approximation. On the other hand, it shows that, as expected, the method becomes inaccurate at high temperatures (Fig. 2.7 showed this for the ideal gas) and eventually breaks down. For this combination of parameters, we cannot go beyond T ≈ 0.88 Tc because the argument of the exponential in Eq. (3.57), as illustrated in Fig. 3.4, vanishes at a point close to the Thomas-Fermi radius. 4 Repulsive interactions lower the condensation temperature because they tend to decrease the density of the system, which then finds it more difficult to reach the Einstein condition. 54 8 ¯ ¯ ¯µ̃eff ¯ 6 4 PSfrag replacements 2 0 0 1 2 x 3 4 Figure 3.4. Failure of the Hartree-Fock approximation at high temperatures, due to the vanishing of the effective chemical potential (defined by the argument of the exponential in Eq. (3.57)). The figure shows the behavior, for the same 3D gas as above, of µ̃eff at T /Tc = 0.82, 0.84, 0.86, and 0.88 (from top to bottom at x = 0). The diamond marks the Thomas-Fermi radius. As we remarked previously, we know the function only at the Gauss-Laguerre points denoted by the open circles; the lines result from a spline interpolation. Figure 3.5 shows the behavior of the chemical potential of the interacting gas; it is significantly larger, and decreases less abruptly with increasing temperature, than its ideal-gas counterpart (which is stuck at around 3 for the whole range of temperatures seen in the figure). Figures 3.6 on page 57 and 3.7 on page 58 exhibit some representative densities and number densities (solid lines), once again resolved into condensate (dotted) and noncondensate (dash-dotted) fractions. The delocalization hump that already characterized the thermal component of the ideal trapped gas reappears here and is even more pronounced. This is to be expected, since once again the most localized state available dominates, and this time the repulsive interactions displace the thermal cloud even farther away from the center of the trap. We now go back to the two-dimensional gas. In Chapter 4 we will study its density and free energy in detail; in the following figures we will use it to illustrate the behavior of the off-diagonal interacting density matrix introduced in Section 3.6. Figure 3.8 on page 59 compares the densities and number densities (as usual, resolved into condensate and noncondensate) of an ideal 1000-atom Bose-Einstein gas (at left) at T = 0.7 T c with 55 14 12 µ̃ PSfrag replacements 10 8 0 0.2 0.4 T /Tc 0.6 0.8 1 Figure 3.5. Chemical potential of a three-dimensional interacting Bose gas in the Hartree-Fock approximation. The parameters are the same as in Fig. 3.3. the equivalent interacting system (at right) in the Hartree-Fock approximation. We have used the “rubidium” value γ = 8π × 0.0043 for the coupling constant. The expected effects are visible: the interactions spread out the system and make it less dense; they also cause a significant decrease of the condensate fraction. Figure 3.9 on page 59 shows the occupation numbers yielded by the diagonalization of the onebody density matrix. The white bars show the ideal-gas populations and the black bars display those of the interacting system. The depletion of the condensate due to interactions, and the complementary increase of the excited occupation numbers, can be seen here. On the other hand, the ground-state population is still orders of magnitude greater than that of any other state; in this case we do not seem have the “smeared” condensation predicted for other two-dimensional systems [37,80] (see also the footnote on page 14). The open circle at the left shows the condensate number yielded by the self-consistent procedure; the eigenvalue of the density matrix is about 4% greater, but the difference between the two is only around 1% of N . Figure 3.10 on page 60 exhibits the corresponding eigenfunctions. The solid lines represent the (spline-interpolated) eigenfunctions that result from diagonalizing the interacting off-diagonal density matrix; for comparison we show the ideal-gas eigenfunctions (dashed lines) previously exhibited in Fig. 2.10. Just as in that figure, we used a 30 × 30 Gauss-Laguerre grid to discretize the matrix. 56 PSfrag replacements ñ(x)/N 0.015 0.01 0.005 6 0 0 1 2 88 x 4 22 5 0.015 ñ(x)/N 0.015 ñ(x)/N 66 44 T /Tc (%) 3 0.01 0.005 0.01 0.005 PSfrag replacements 0 0 1 2 x 3 4 0 0 5 2 1 2 x 3 4 5 3 4 5 0.015 ñ(x)/N ñ(x)/N 0.015 1 0.01 0.005 0.01 0.005 6 0 0 1 2 x 3 4 0 0 5 x Figure 3.6. Two views of the density of an interacting 3D trapped gas at various temperatures. The parameters are the same as in Fig. 3.3. Though this is not apparent from the figure, in the last panel the condensate accounts only for some 10% of the atoms. 57 0.6 PSfrag replacements N (x)/N 0.4 0.2 0 0 2 4 88 66 6 x 44 8 10 0 N (x)/N 0.6 N (x)/N 0.4 0.2 0 0 2 4 x 6 8 0.4 2 4 2 4 x 6 8 10 6 8 10 0.4 0.2 0 0 0.2 0 0 10 N (x)/N N (x)/N 0.4 PSfrag replacements T /Tc (%) 22 2 4 x 6 8 0.2 0 0 10 x Figure 3.7. Number density of an interacting 3D trapped Bose gas at various temperatures. The parameters and temperatures are the same as in the preceding figure; in three dimensions, N (x) = 4πx2 ñ(x). The noncondensate is clearly seen to dominate the system at the highest temperature. 58 0.15 0.1 0.1 ñ(x)/N ñ(x)/N 0.15 0.05 PSfrag replacements 0 0 2 4 x 0.05 0 0 6 0.2 0 0 x 4 6 4 6 0.4 N (x)/N N (x)/N 0.4 2 2 4 x 0.2 0 0 6 2 x Figure 3.8. Densities (top) and number densities (bottom) of a two-dimensional isotropic ideal gas (at left) and the equivalent interacting system (at right). In both cases, N = 1000 and T = 0.7 T c ; the interacting gas has an interaction parameter γ = 8π × 0.0043. 1000 Population PSfrag replacements 100 10 7 1 0 1 2 3 4 Eigenstate 5 6 Figure 3.9. Eigenvalues of the one-body density matrix of ideal (white bars) and interacting (black bars) two-dimensional trapped gases. The parameters are the same as in Fig. 3.8. 59 1 0 1 0 φ0 (x), . . . , φ6 (x) 1 0 1 0 1 0 1 PSfrag replacements 0 1 0 6 0 1 2 3 4 5 x Figure 3.10. Ground state and azimuthally symmetric excited eigenfunctions (solid lines) of a two-dimensional isotropic interacting trapped gas. The parameters are the same as in Fig. 3.9. For comparison we show the corresponding ideal-gas eigenfunctions (dashed lines). 60 The open circles at the bottom show the Gauss-Laguerre points where the functions are known exactly; the function they depict, however, is not the ground state given by the density matrix but the condensate wavefunction that resulted from the self-consistent procedure and that we used to calculate the matrix using Eq. (3.62). Here we can see once more the competition between confinement and interaction effects: while the wavefunctions decrease in value and are flattened at the center of the trap, the radius of the system becomes larger; as the figure shows, both condensate and noncondensate are affected. 3.8 Summary Mean-field theory is the main tool we use in the study of interacting Bose-Einstein systems; it has been used with much success in three dimensions, and this lets us perform thorough checks on our methods of solution. Mean-field theory prescribes coupled equations for the condensate and the rest of the system: the condensate receives a full quantum-mechanical treatment from the beginning; the noncondensate, on the other hand, is described by various schemes of different degrees of sophistication, some of which we introduce. We undertake a fairly detailed analysis of the zero-temperature case, where the system is fully condensed; not only is this a good test system in which to study the approach to the thermodynamic limit, it is also amenable in that case to exact solution. We use it to find a criterion for the crossover in dimensionality of the interacting gas and to obtain an expression for the interatomic coupling constant for the two-dimensional system that results from squeezing a 3D trap. For systems at finite temperature, we settle upon a semiclassical Hartree-Fock treatment that, though it breaks down at high temperatures, predicts reasonable results in two dimensions. In particular, it shows that there is no smearing in 2D. 61 CHAPTER 4 THE TWO-DIMENSIONAL BOSE-EINSTEIN CONDENSATE1 One reason that impels physicists and applied mathematicians to look at systems in less than three dimensions is that very often they are simpler. There are two rather different reasons for this simplicity. The first is the rather obvious reason that space in lower dimensions is described by fewer coördinates, and so calculations are less laborious. The second is that the topology—the structure—of lower-dimensional spaces is simpler, and this enables various tricks to be used that cannot be used in higher-dimensional spaces. —D.J. Thouless [109] 4.1 Introduction In a recent paper [30], one of the groups that pioneered the formation and detection of Bose-Einstein condensation (BEC) in harmonically trapped atomic gases [2,6,7] reports the creation of (pseudo-) two-dimensional condensates. These have been produced by taking a three-dimensional condensate of 23 Na atoms and carrying out two independent processes on it: i) Initially, one of the confining frequencies (that in the z direction, ωz , in order to minimize the effects of gravity) is increased until the condensate radius in that dimension is smaller than the healing length associated with the interaction between atoms (taken here to be repulsive and parametrized by the two-body scattering length a). This is not sufficient to reduce the dimensionality, however, since the atoms, each of which has mass m, will literally squeeze into the third dimension if there are more than N= s 32~ 225ma2 1 s ωz3 ωx2 ωy2 (4.1) This chapter is a reprint of the paper “The two-dimensional Bose-Einstein condensate,” by J. P. Fernández and W. J. Mullin, Journal of Low Temperature Physics 128, 233 (2002), and as such it purports to be self-contained. Some repetition of previous material is therefore inevitable, and we have kept references to other parts of the thesis at a minimum. The figures have been left untouched. The original version of the paper contained an appendix that eventually got discarded; it reappears here as Section 4.5. 62 of them in the trap. ii) Consequently, the number of atoms in the condensate must be reduced; this is achieved by exposing the condensate to a thermal beam. The reduction in effective dimensionality becomes apparent when the aspect ratio of the expanding condensate, which is independent of N in 3D, starts to change as the number of atoms is gradually reduced. The condensates thus produced have a number of atoms that ranges between 104 and 105 . This constitutes an important experimental contribution to the long-standing debate about the existence of BEC in two dimensions. It is a well-known fact (and a standard textbook exercise) that BEC cannot happen in a 2D homogeneous ideal gas; a rigorous mathematical theorem [27] extends this result to the case where there are interactions between the bosons. When the system is in a harmonic trap, on the other hand, BEC can occur in two dimensions [24] below the p temperature kB Tc = ~ω 6N/π 2 introduced on page 24, but the theorem is once again valid when interactions between the bosons are considered [28]: while there is a BEC in the 2D system, it occurs at T = 0. The preceding discussion is valid only in the thermodynamic limit, which in the particular case of an isotropically trapped system consists of making N → ∞ and ω → 0 in such a way that N ω 2 remains finite [28,57]. The question remains whether a phenomenon resembling BEC—that is, the accumulation of a macroscopic number of particles in a single quantum state—occurs or not when the system consists of a finite number of particles confined by a trap of finite frequency, as is certainly the case in experimental situations. If there is such a phenomenon, one would like to know more about the process by which this “condensate” is destabilized at finite temperatures as N grows. Some authors [39,40] have considered, in the finite homogeneous 2D case, the possibility of a BEC. A similar analysis [81] was considered for a quasi-2D trapped gas; the latter reference finds that the phase fluctuations in the condensate vary with temperature and particle number as h(δφ)2 i ∝ T log N (4.2) which diverges for finite temperature as N → ∞, as one would expect. For finite N , on the other hand, the fluctuations are tempered at very low temperature; since the coherence length, though finite, is still larger than the characteristic length imposed on the system by the trap [41] or by walls, one can speak of a “quasicondensate.” It is this quasicondensation that we wish to study in this paper. Reference 29 reports the observation of a quasicondensate in a homogeneous system of atomic hydrogen adsorbed on 4 He; at this point we cannot tell if the condensates reported in Ref. 30 are actually quasicondensates. 63 One can approach the study of the 2D Bose gas by employing mean-field theory, which in this context refers to the Hartree-Fock-Bogoliubov (HFB) equations and various simplifications thereof that we will review in quantitative detail and classify in the present work. The HFB theory has been remarkably successful in treating the 3D case; though we give a few representative references below, we refer the reader to Ref. 54, which includes a comprehensive review and an exhaustive list of references, and will henceforth concentrate on the work that has been carried out in 2D. The HFB theory assumes from the outset that the system is partially condensed, and proposes separate equations to describe the condensate and the uncondensed (thermal) component. The condensate is described by a macroscopic wavefunction that obeys a generalized Gross-Pitaevskiı̆ (GP) equation [75,76], while the noncondensate consists of a superposition of Bogoliubov quasiparticle and “quasihole” excitations weighted by a Bose distribution. The expression for the noncondensate can be simplified by neglecting the quasihole excitations (the Hartree-Fock scheme) [23,103], by performing a semiclassical WKB approximation [110], or by combining both of these [69,88,102]. Each one of the above schemes can be further simplified, when the thermodynamic limit is approached, by neglecting the kinetic energy of the condensate: this is the Thomas-Fermi limit [77,96]. Finally one can neglect the interactions between thermal atoms, arriving at the semi-ideal model [111,112]. The semi-ideal model has been implemented in 2D [113,114], but has been found to yield unphysical results as the interaction strength becomes sizable. The full-blown HFB model does not appear to be much more successful, at least when considered in the semiclassical limit. In a previous paper [43] it was found that, below a certain temperature, the introduction of a condensate in the Thomas-Fermi limit (corresponding to the thermodynamic limit) renders the HFB equations incompatible, with the noncondensate density becoming infinite at every point in space. The singularity occurs at the low end of the energy spectrum, indicating that the condensate is being destabilized by long-wavelength phonons; this interpretation in terms of phase fluctuations had already been proposed for the homogeneous system [27,115]. In this work we also report our inability to find self-consistent semiclassical solutions to the HFB model when a finite trapped system is considered. Moreover, it was recently discovered [42,100] that it is possible, in the 2D case exclusively, to solve the HFB equations semiclassically at any temperature without even having to invoke the presence of a condensate (thus obtaining what we will call an “uncondensed” solution). In other words, it is possible to simply cross out the condensate component and solve for the system to a temperature close to T = 0. The solution thus obtained shows an accumulation of atoms at the center of the trap and yields a bulge in density similar to that caused by the presence of a condensate, even though no state is macroscopically occupied. 64 These results appear to reinforce the conclusion that BEC cannot happen in the two-dimensional trapped system. However, we are still confronted with the experimental results described above. Furthermore, two different Monte Carlo simulations [31,32] show significant concentrations of particles in the lowest energy state for finite N , though it must be said that this method provides little information about the types of excitations that contribute to the disappearance of the condensate, and that it is difficult to carry out such simulations on very large systems. The HFB model cannot be discounted at this stage either: when we restrict ourselves to the Hartree-Fock approximation, it is possible to find self-consistent solutions (henceforth referred to as “condensed”) involving a condensate, both for finite systems and in the thermodynamic limit, when using either the discrete set of equations [116]2 or the WKB approximation [117] to treat the noncondensate. The Hartree-Fock approximation neglects phononlike excitations, so it is not surprising that it yields solutions. However, one may be able to justify its usefulness. It is known that, in the infinite homogeneous system, infrared singularities occur but are renor√ malized by interactions, providing, in essence, a cutoff at a low wavenumber k 0 ≈ nmU [118–120], where n is the density, m the particle mass, and U the effective interaction strength. Indeed, it is possible to estimate the Berezinskii-Kosterlitz-Thouless (BKT) transition temperature by simply cutting off the ideal-gas density expression at this k0 [119]. Presumably a similar situation occurs in the trapped case. The Hartree-Fock approach provides a convergent theory by cutting off the singularities at a wavenumber similar to that of more rigorous theories. Whether such an approach gives a reasonable estimate of the BKT transition temperature, the superfluid density, or a quasicondensate density for the finite system will need to await a more rigorous theoretical approach to the interacting 2D trapped gas [121]. Given the above limitations, we analyze the character of the BEC in the 2D trapped system by solving the coupled equations of the theory. We find that, in the Hartree-Fock scheme, it is possible to find both condensed and uncondensed solutions for the two-dimensional equations. We also calculate the free energy corresponding to each one and find that the condensed solution has a lower free energy at all temperatures, which appears to imply that, at least at this level of approximation, the uncondensed solution is unphysical or metastable. The condensed solution will be “preferred” over the uncondensed one, and we suspect that the solution represents an approximation to a quasicondensate. 2 The discrete set of equations was very recently solved for two dimensions in the full HFB approximation [93]. 65 It is evident from our discussion that our approach to BEC in 2D trapped systems is a preliminary one and that a further analysis that takes fully into account the BKT transition is necessary. A start to answering this need has been presented in Ref. 121, and we intend to return to this problem ourselves in the future. 4.2 4.2.1 The model The Hartree-Fock-Bogoliubov equations Throughout this paper we use a dimensionless system of units in which all lengths are scaled by p the oscillator length x0 ≡ ~/mω and all energies are expressed in terms of the one-dimensional ground-state energy of the oscillator, ~ω/2. Dimensionless variables will in general carry a tilde: for example, the total density n becomes ñ ≡ x20 n and the chemical potential µ̃ ≡ µ/ 12 ~ω. The HFB equations [54,73,74,88] result from assuming that i) the (repulsive) interactions between atoms consist exclusively of two-body low-energy collisions that can be described by a delta-function pseudopotential [79] of strength g (related to its dimensionless counterpart γ through g ≡ 12 ~ωx20 γ), that ii) the many-body field operator Ψ can be decomposed via Ψ = hΨi + ψ̃ ≡ Φ̃ + ψ̃, (4.3) where the ensemble average Φ̃ is a real macroscopic wavefunction that describes the condensate (reflecting the imposition of macroscopic long-range order on the condensed system), and that iii) products of noncondensate operators can be simplified using a finite-temperature version of Wick’s theorem [84,90]. If we insert (4.3) into the many-body grand-canonical Hamiltonian H̃ = Z d2x Ψ† (Λ̃ + γ † † Ψ Ψ Ψ)Ψ, 2 (4.4) ˜ 2 + x2 − µ̃, and neglect anomalous averages via the Popov approximation [88], where Λ̃ = −∇ we obtain an expression that can be diagonalized and yields an infinite set of coupled differential equations. On one hand, the macroscopic wavefunction mentioned above is the square root of the dimensionless condensate density ñ0 and obeys the generalized Gross-Pitaevskiı̆ equation, Λ̃Φ̃ + γ(ñ0 + 2ñ0 )Φ̃ = 0, 66 (4.5) where ñ0 ≡ ñ−ñ0 is the noncondensate density. The factor of 2 in (4.5) and hereafter is a consequence of the direct (Hartree) and exchange (Fock) terms being identical, which follows from the fact that the delta-function interaction that we are considering has zero range [74]. (The term involving the condensate does not include that factor; this is a consequence of the restricted grand-canonical ensemble that we are using. See the end of the next section.) The noncondensate, in turn, is described by an infinite number of pairs of functions that obey [73] à Λ̃ + 2γ ñ −γ ñ0 −γ ñ0 Λ̃ + 2γ ñ !à uj vj ! = ²j à uj −vj ! (4.6) and which generate the noncondensate density via ñ0 (x) = X ¡¡¯ ¯2 ¯ ¯2 ¢ ¯ ¯ ¢ ¯ u j ¯ + ¯ vj ¯ f j + ¯ vj ¯ 2 , (4.7) j where we have introduced the Bose-Einstein distribution factor fj ≡ (e²j /t − 1)−1 , which appears when the free energy of the system is minimized, and the dimensionless temperature t ≡ k B T / 12 ~ω. The last term of (4.7) describes the zero-temperature depletion of the condensate, which accounts for less than 1% of the particles and is therefore negligible [88]. This self-consistent set of equations is closed, and the chemical potential found, by imposing that the total number of particles remain fixed: N = N0 + N 0 = Z d2x ñ(x) = Z d2x (ñ0 + ñ0 ) . (4.8) At temperatures such that kB T À ~ω, one can use the semiclassical (WKB) approximation [110] that results from taking uj ≈ u(x)eiφ and vj ≈ v(x)eiφ . The phase common to both defines a ˜ quasiparticle momentum through κ = ∇φ; u, v, and κ vary sufficiently slowly with x that their spatial derivatives can be neglected [88]. The distribution factor is now f j ≈ f (x, κ), infinite sums are transformed into momentum integrals—which in two dimensions can be solved in closed form [43]—and the equations become algebraic, yielding the Bogoliubov energy spectrum [88] ²HFB (κ, x) = q (κ2 + x2 + 2γ ñ − µ̃)2 − γ 2 ñ20 (4.9) and the following integral expression for the noncondensate density: ñ0 (x) = t 4π Z ∞ ((x2 +2γ ñ−µ̃)2 −γ 2 ñ20 )1/2 /t √ ¡ ¢ t dξ −( (x2 +2γ ñ−µ̃)2 −γ 2 ñ20 )/t = − log 1 − e . ξ e −1 4π 67 (4.10) At high enough temperatures there is no condensate, and the density on the left-hand side of (4.10) is just the total density. We thus have to solve a single self-consistent equation, ñ0 (x) → ñ(x) = − ¡ ¢ 2 t log 1 − e−(x +2γ ñ−µ̃)/t ; 4π (4.11) the chemical potential is once again calculated by requiring N to be fixed. Equation (4.11) has been found by the authors of Ref. 42 to be soluble at all temperatures, a result that we confirm in the present work. Reference 43 had mistakenly concluded that it was impossible to solve (4.11) below a certain temperature. The Hartree-Fock approximation [88,96,103] amounts to neglecting the v j in (4.6) and, when combined with the semiclassical treatment, results in the disappearance of the last term within the square root of Eq. (4.9). The energy spectrum now becomes ²HF ≡ ² = κ2 + x2 + 2γ ñ − µ̃ (4.12) and the noncondensate density turns into ñ0 (x) = − ¡ ¢ 2 t log 1 − e−(x +2γ ñ−µ̃)/t , 4π (4.13) an expression that differs from (4.11) in that the left-hand side corresponds only to the noncondensate density. The three-dimensional version of this equation, coupled with (4.5) and (4.8), has been frequently used in the literature to study the three-dimensional gas. Reference 69, for example, exhibits a detailed comparison of its predictions to those of Monte Carlo simulations and finds excellent agreement between the two. In two dimensions, the Hartree-Fock equations have been solved without resorting to the WKB approximation [116]. The authors of this reference succeeded in finding self-consistent density profiles and used them to study the temperature dependence of the condensate fraction. Our results [117], which do take advantage of the semiclassical approximation, agree quite well with theirs (see Fig. 4.1). 4.2.2 The thermodynamic limit When N is large we can neglect the kinetic energy of the system, which in 2D can be shown to be of order 1/N [43], and obtain the Thomas-Fermi approximation [77] γ ñ0 = (µ̃ − x2 − 2γ ñ0 )Θ(µ̃ − x2 − 2γ ñ0 ), 68 (4.14) 160 120 100 0 x2 n (x) , x2 n’ (x) 140 0 0 80 60 40 20 0 0 2 4 x=r/x 6 8 10 0 Figure 4.1. Condensate and noncondensate density profiles of a two-dimensional Bose-Einstein gas with N = 104 and γ = 0.1 at T = 0.7 Tc , where Tc is the ideal-gas transition temperature. The Goldman-Silvera-Leggett model (dashed lines) treats the condensate in the Thomas-Fermi limit (4.14), while the Hartree-Fock model (full lines) uses the full Gross-Pitaevskiı̆ equation (4.5). Both models use Eq. (4.13) to describe the noncondensate. where Θ(x) is the Heaviside step function, introduced to ensure that the density profile is everywhere real and positive. Also, since the thermodynamic limit requires that ω → 0, the WKB approximation becomes rigorous and can be used with confidence. Thus we can insert expression (4.14) in the Bogoliubov energy spectrum (4.9) and show that the latter reduces to ²HFB (κ, x) ≈ κ p κ2 + 2γ ñ0 ≈ κ p 2γ ñ0 (4.15) for small quasimomenta. The fact that it is linear clearly shows us that, in this approximation, the low end of the energy spectrum corresponds to phononlike quasiparticles. Now, if we introduce (4.14) into the noncondensate density (4.10), we can see that the argument of the logarithm vanishes at all temperatures, making the density diverge at every point in space. The first line of (4.10) shows that the divergence in the integral is caused at its lower limit; this restates the conclusion arrived at in Ref. 43: low-energy phonons destabilize the condensate in the two-dimensional thermodynamic limit when the noncondensate quasiparticles obey the Bogoliubov spectrum. In the Hartree-Fock approximation, on the other hand, the energy spectrum tends in this limit to ² ≈ κ2 + γ ñ0 and predicts single-particle excitations whose minimum energy is γ ñ 0 ; this can be 69 interpreted equivalently by assigning a minimum value κ2c = γ ñ0 for the excitation quasimomentum. This cutoff is consistent with those proposed in the past [39,40,118–120] and removes the infrared singularity in the HFB equations. This momentum cutoff is robust enough that it enables one to carry out Hartree-Fock calculations even in the thermodynamic limit: in fact, as was first found in Ref. 96, the introduction of this limit actually simplifies the calculations, and it is possible to find self-consistent solutions by simultaneously treating the noncondensate in the Hartree-Fock approximation and the condensate in the Thomas-Fermi limit [116,117]. This model cannot provide realistic density profiles at every point in space, since the condensate density (4.14) has a discontinuous derivative at its edge, but predicts quite reasonable results outside of this region, as can be seen in Fig. 4.1. 4.2.3 The free energy We have seen that in 2D the mean-field-theory BEC equations admit solutions both with and without a condensate. The unphysical solution should be that with the highest free energy, since equilibrium at finite temperatures occurs when the grand potential attains a minimum; in fact, the Bose-Einstein distribution factor in Eq. (4.7) comes from minimizing this quantity [103,122], which in our dimensionless units adopts the form 1 Ω = (U − µN − T S)/ ~ω ≡ 2 Z d2x (Υ − µ̃ñ − tΣ). (4.16) At this point we have to keep in mind that the grand-canonical free energy is a function of µ̃, not of N ; in order to make a meaningful comparison of these energies at the same N , then, we have to compare the Helmholtz free energies, given by à = Ω + µ̃N . The expressions given below have all been derived in the grand-canonical ensemble and thus contain the chemical potential; we will eliminate this dependence on µ̃ by adding µ̃N to the expressions we obtain. In the Hartree-Fock approximation, the grand-canonical energy density of the system is given by [88] Υ − µ̃ñ = Φ̃(Λ̃ + γ 1 ñ0 )Φ̃ + 2 (2π)2 Z d2κ ² − γ ñ02 ; e²/t − 1 (4.17) the first term corresponds to the condensate energy, while the second one is the sum, weighted by the Bose-Einstein distribution, of the energies of the excited states; this last expression includes an extra 70 term γ ñ02 that has to be subtracted explicitly, as is usually the case in Hartree-Fock calculations [84]. (See Section 4.5.)We can simplify (4.17) further by invoking Eq. (4.5): Υ − µ̃ñ = −γ ñ2 + 1 γ 2 ñ0 + 2 (2π)2 Z d2κ ² . −1 e²/t (4.18) The entropy of the system can be found from the combinatorial expression [48,84] S = −kB X i (fi log fi − (fi + 1) log(fi + 1)), (4.19) which in the WKB and Hartree-Fock approximations yields the entropy density [88] tΣ = 1 (2π)2 Z t d2κ ² − ²/t e − 1 (2π)2 Z d2κ log(1 − e²/t ). (4.20) The first term in (4.20) cancels with the last one in (4.18), and the other term can be integrated in closed form, yielding Ãc = µ̃N − Z ¶ µ t2 1 2 −(x2 +2γ ñ−µ̃)/t 2 g2 (e ) d x γ(ñ − ñ0 ) + 2 4π 2 (4.21) for the free energy of the condensed solution. As T → 0, this expression reduces to the correct value in the homogeneous case [23]. In the trapped case, it tends to the Thomas-Fermi value given in Eq. (3.36). The free energy of the uncondensed solution is found by simply crossing out the condensate density in (4.21): Ãu = µ̃N − Z 2 dx µ ¶ t2 −(x2 +2γ ñ−µ̃)/t γ ñ + g2 (e ) . 4π 2 (4.22) Equations (4.21) and (4.22), of which more-general versions are derived in Ref. 83 by a different method [103], can be easily seen to become identical at temperatures high enough that the condensate density can be neglected. On the other hand, their low-temperature limits differ, since Eq. (4.22) tends to a higher value than that attained by (4.21). This can be traced back to the fact that the grand-canonical ensemble has to be changed in order for it to correctly describe the particle-number fluctuations at low temperatures [83,123]. Now, a 2D Bose system has a condensate at least at T = 0, so the two free energies should coincide there; however, we must keep in mind that the semiclassical approximation used to derive Eq. (4.22) requires that kB T À ~ω; thus the method we are using cannot describe the appearance of the zero-temperature condensate in the uncondensed solution. At the end of the next section we will study further consequences of this distinction. 71 4.3 Numerical methods and results There is a time-honored prescription [74,88] for finding the self-consistent solution of the HartreeFock equations in the presence of a condensate: Initially, we assume that only the condensate is present and solve the Gross-Pitaevskiı̆ equation (4.5) for ñ0 = 0. The wavefunction and eigenvalue that result are fed into the Hartree-Fock expression for the density (4.13), which, when integrated over all space, yields also a value for the noncondensate fraction N 0 ; one can then readjust the condensate fraction and solve the Gross-Pitaevskiı̆ equation that results. The process is then iterated until the chemical potential and the particle fractions stop changing. By far the most difficult part of this process is the solution of the nonlinear eigenvalue problem (4.5). Different methods exist in the literature; we have obtained identical results by solving it as an initial-value problem [94] and, much more efficiently, by employing the method of spline minimization [44,69]. This method uses the fact that Eq. (4.5) is the Euler-Lagrange equation that minimizes the functional J[Φ̃] = Z ¤ £ ˜ Φ̃)2 + Φ̃(x2 + 2γ ñ0 (x))Φ̃ + γ Φ̃4 . d2x (∇ 2 (4.23) After setting a small, nonuniform grid of fixed abscissas that represent the coördinate x, we take the corresponding ordinates, which represent Φ̃, as the parameters to be varied until (4.23) attains its smallest possible value. Since we only have information about the value of the function at a discrete set of points, in order to calculate the necessary derivatives and to integrate we perform a cubic-spline interpolation; the integral is found using ten-point Gauss-Legendre quadrature. The minimization is carried out using the Nelder-Mead method. We checked our code by comparing its predictions in three dimensions to previously published results. For the case studied in Ref. 69, the condensate fractions we found differed from those in the paper by less than one part in 104 . The code also reproduced previously known results for the ideal gas, including finite-size effects [28]. Figure 4.1, already discussed above, shows one of the solutions that we have found using the Hartree-Fock approximation. The gas has N = 104 atoms; the coupling constant γ has been chosen so that the system has approximately the same radius as the three-dimensional gas studied in Ref. 69, where parameters resembling those of the original JILA trap [6] are used. The system is shown at T = 0.7 Tc , where Tc is the condensation temperature for the ideal gas. We have also found solutions for the Goldman-Silvera-Leggett model, which corresponds to the large-N limit of the GP equation. 72 This was done by treating the problem as a simultaneous system of nonlinear equations on a uniform grid and solving it with a least-squares method. When N = 104 , as in Fig. 4.1, it is not possible to find a self-consistent solution for the condensed equations beyond T ≈ 0.8 Tc , since ²HF becomes negative; this had already been noted in Ref. 69 for three dimensions, where it was interpreted as a finite-size effect, and occurs at even lower temperatures for 2D. The limitation becomes more severe as N increases: for N = 10 6 we cannot find solutions beyond T ≈ 0.5 Tc . The authors of Ref. 116 report predictions at temperatures very close to the transition by using a finite-size correction to the chemical potential [70]. Our inability to work above certain temperatures might be a consequence of using the semiclassical approximation, though we point out that the Popov approximation is expected to break down close to the transition temperature [54,88]. We also found self-consistent three-dimensional solutions using the more general Bogoliubov spectrum by applying the same method and using the Hartree-Fock solutions as a starting point for the iteration. We find that these solutions exhibit an enhanced depletion of the condensate, in agreement with those found by other authors [70]. It was impossible, however, to find this kind of solution in two dimensions, even for systems with N as low as 100: after a few iterations, the chemical potential became too large, the Bogoliubov energies became imaginary, and the noncondensate density diverged. When the possibility of phononlike excitations is allowed, then, the condensate is destabilized, just as had been found in the Thomas-Fermi limit [43]. The uncondensed case, on the other hand, gives us solutions in these conditions, and to it we now turn. Equation (4.11) along with the particle-conservation condition (4.8) for all temperatures is most easily solved by rewriting (4.11) as Ze−x 2 /t = 2 e−(π−γ)ν(x) sinh πν(x), (4.24) where have introduced the fugacity Z = eµ̃/t and ν(x) ≡ 2ñ(x)/t. Given Z, t, and γ it is possible to find ν at every point using a standard root-finding algorithm. One then wants to find the value of Z such that the total number of particles is N . It is better, however, to write Eq. (4.24) at the origin, Z = 2 e−(π−γ)ν0 sinh πν0 , (4.25) eliminate Z between (4.24) and (4.25), and solve for ν0 = ν(0), the density at the center of the trap, using the same root finder. 73 200 x0 n (x) 150 2 100 50 0 0 2 4 6 8 10 x = r / x0 Figure 4.2. Total density profiles of a two-dimensional gas with the same parameters as in Fig. 4.1. In this case we exhibit both the uncondensed (full line) and the condensed (dashed line) solutions; the latter we have broken once again into its condensate and noncondensate parts (dotted lines). In Fig. 4.2 we show the (Hartree-Fock) condensed and uncondensed solutions that we have obtained. They are similar in shape and exhibit identical behavior for large x. The uncondensed solution has a lower value at the origin and predicts a wider radial density profile. We also calculated the free energy corresponding to each solution; the results are shown in Fig. 4.3, which shows Ã/N as a function of temperature for both cases when N = 103 and N = 104 . The free energies appear to coincide at high temperatures; this was to be expected, since Eqs. (4.21) and (4.22) become identical at temperatures high enough for the condensate density ñ 0 to be neglected. As for the low-temperature limit, we have already noted that the free energy of the condensed solution tends to the value predicted by the zero-temperature Thomas-Fermi limit, while that of the uncondensed solution tends to a higher value. This value can actually be calculated: it is easy to show [42] that the low-T limit of Eq. (4.24) is 2γn(x) = µ̃ − x2 , (4.26) the Thomas-Fermi limit but with γ replaced by 2γ; a larger interaction strength, as we can see from (3.36), implies a higher free energy. Interestingly, when we compare a given Hartree-Fock solution to an uncondensed solution with half the interaction strength, we find that the density 74 40 20 A/N 0 −20 −40 −60 −80 0 0.1 0.2 0.3 0.4 0.5 T/T 0.6 0.7 0.8 0.9 1 c Figure 4.3. Free energy per particle, Ã/N , for the condensed (dotted line) and uncondensed (full line) solutions. The curves with higher value at T = 0 correspond to N = 10 4 atoms, while those with the lower value at T = 0 correspond to N = 103 . The coupling constant is γ = 0.1 in both cases. The open circles on the vertical axis are the Thomas-Fermi predictions given by Eq. (3.36) for T = 0. The free energies are seen to concur in the high-temperature limit; the low-temperature limit, on the other hand, shows the discrepancy that results from restricting condensate fluctuations (see discussion after Eq. (4.22)). Note that Tc , the transition temperature for the ideal 2D trapped gas, depends on N ; thus the x axis for N = 103 and that for N = 104 represent different actual temperatures. profiles become very similar (and are indistinguishable at T = 0), while the free energies coincide almost exactly for a wide range of temperatures; attractive as this possibility might be, however, it has a serious flaw: the free energies start to differ as the temperature increases, where they should coincide by definition. This tells us that the factor of 2, which turns out to be the same one discussed after Eqs. (4.5) and (4.22), has to be retained; the WKB approximate expression for the uncondensed state is valid only for kB T À ~ω, so dropping the factor of 2 in order to match the T = 0 condensate is invalid. (The authors of Ref. 42, in fact, omit this factor from their paper, although they address this question in a subsequent publication [100]). It is in fact this factor that guarantees that the uncondensed solution has a higher free energy than the condensed one at all temperatures; this, despite our inability to find condensed solutions using the Bogoliubov energy spectrum, leads us to conclude that, at least at this level of approximation, the uncondensed solution is unphysical and the two-dimensional finite trapped system will exhibit some sort of condensation at finite temperature. 75 4.4 Conclusion We have found solutions to the two-dimensional HFB equations in the Hartree-Fock approximation, both for finite numbers of atoms and in the thermodynamic limit; still, when we try to go beyond this scheme and consider the whole Bogoliubov excitation spectrum, the low end of which is described by phonons, we are unable to find self-consistent solutions for finite—even low—values of N . We have seen that it is possible to describe the system as having no condensate at all, but these solutions, at least for the parameter combinations that we have studied, have a higher free energy than their partly-condensed counterparts; this leads us to conclude that the 2D system will have some kind of condensate at low enough temperatures, insofar as the Hartree-Fock approach can successfully describe the quasicondensate that we expect to find in a finite system. One could argue that the whole mean-field approach we have adopted is wrong in two dimensions, since a condensation into a single state is being assumed from the start. However, an alternative treatment [124] that does not make this assumption ends up with equations identical to (4.6), so we are left none the wiser. We have also bypassed the fact that the interaction strength g is not really constant in 2D, but rather depends logarithmically on the relative momentum [35,81]; this, however, has been found to be of little consequence [100]. Another possibility is that condensation occurs into a band of states, forming a “smeared” or generalized condensate [36,38]; this alternative situation is not ruled out by the standard proof of Hohenberg’s theorem [80]. Monte Carlo simulations do seem to predict the presence of a condensate in two dimensions [32, 125]. Furthermore, two-dimensional condensates appear to have been produced in the laboratory [29, 30]. Presumably the MC simulations show both the effects of a quasicondensation of a finite system and the BKT transition to the superfluid state. Our attempt here has been to test the possibility of representing these Bose effects by a relatively simple set of HFB or Hartree-Fock equations. While no solutions can be found for the HFB set, even for a finite system, the Hartree-Fock equations do provide a description of a condensed state. We feel there is reason to believe that this description should be a fair representation of the actual situation. 4.5 Appendix: The Hartree-Fock excess energy The fact that the interaction energy is overcounted is solely a result of the Hartree-Fock approximation (i.e., it happens independently of the semiclassical approximation), and it occurs even in the absence of a condensate. To simplify matters, we will prove this assertion at a temperature high enough that the whole system can be described by the noncondensate field operator ψ̃. 76 The functions uj and vj are defined as the coefficients that result when the noncondensate field operator ψ̃ is expanded in creation and annihilation operators. In the Hartree-Fock approximaR 2 ∗ tion, the vj are neglected and the uj can be proved to be orthonormal [73]: d x uj uk = δjk . Equation (4.6) becomes Λ̃uj + 2γ X fk u∗k uk uj = ²j uj ; (4.27) k on multiplying both sides by fj u∗j , summing over j, and integrating over the whole volume we obtain X j fj ²j = X j fj Z dσx u∗j (Λ̃ + 2γ X fk u∗k uk )uj . (4.28) k On the other hand, we can expand ψ̃, insert it directly into the Hamiltonian (4.4), and use Wick’s theorem [84,90] in the form hαj† αk i = fj δjk , hαj† αk† αl αm i = fj fk (δjl δkm + δjm δkl ) (4.29) to obtain X 1 fj (U − µN )/ ~ω = 2 j Z dσx u∗j (Λ̃ + γ X fk u∗k uk )uj . (4.30) k The interaction term has clearly been overcounted in Eq. (4.28) and has to be subtracted explicitly. At low temperatures, only a term corresponding to the noncondensate density has to be subtracted. 4.6 Summary We apply Hartree-Fock-Bogoliubov mean-field theory to the study of a purely two-dimensional finite trapped Bose gas at low temperatures and find that, in the Hartree-Fock approximation, the system can be described either with or without the presence of a condensate; this continues to be true as the system grows in size enough to have reached in practice the thermodynamic limit. Of the two solutions, the one that includes a condensate has a lower free energy at all temperatures. However, the Hartree-Fock scheme neglects the presence of phonons within the system, and when we allow for the possibility of phonons we are unable to find condensed solutions; the uncondensed solutions, on the other hand, are valid also in the latter, more general scheme, but are found to have consistency problems of their own. 77 CHAPTER 5 PATH-INTEGRAL MONTE CARLO AND THE SQUEEZED INTERACTING BOSE GAS It is worth noticing that the possibility of making a close comparison between exact Monte Carlo simulations, experimental data, and meanfield calculations is a rather rare event in the context of interacting many-body systems and represents a further nice feature of BEC in traps. —F. Dalfovo, S. Giorgini, L. P. Pitaevskiı̆, and S. Stringari [54] 5.1 Introduction We have seen in Chapter 4 that, while the HFB equations are successful in providing a description of the finite two-dimensional interacting trapped Bose gas, their success is not beyond doubt or qualification. The semiclassical approximation prescribes an energy spectrum that is linear (i.e., phononlike) at low wavenumbers, and these long-wavelength phonons seem to destroy the longrange order of the system: We saw analytically that the low-wavenumber part of the spectrum contributes an infrared divergence when we attempt to calculate the thermal density of the gas in the thermodynamic limit [43], and we were unable to find a single instance where the full HFBWKB scheme worked for a finite system. Furthermore, we confirmed that these same equations are amenable to solution, quite robustly and at all temperatures, when the condensate density is crossed out and the whole system is treated as an uncondensed gas [42,93,100]. On the other hand, we found that the Hartree-Fock approximation, which results from neglecting the “quasihole” part of the Bogoliubov spectrum, provides a quasimomentum cutoff κ min = γ ñ0 that takes care of most of the problems just described and makes the condensate reappear, fortified: the density profiles start behaving as one would want, closely resembling those in 3D, tending gently to reasonable limits at low temperatures and high particle numbers, and having a lower free energy than their uncondensed counterparts. Moreover, the recently discovered [93] exact solutions that do not resort to the semiclassical approximation (and therefore reflect the true quantum-mechanical behavior of the system) have a discrete energy spectrum that admits no infrared divergences; this bolsters 78 our confidence in the HF solutions as a reasonable description of the two-dimensional condensed Bose gas. We have also calculated (at the end of Chapter 3) the off-diagonal density matrix corresponding to these Hartree-Fock equations: The largest eigenvalue of the matrix gives a condensate number consistent with that found directly in the course of the self-consistent procedure, and inspection of its other eigenvalues reveals that the other states individually have negligible populations. In this final chapter we want to see if a Bose gas confined in a trap of extreme anisotropy can be described consistently by these two-dimensional Hartree-Fock solutions. The ideal trapped Bose gas, as we had the chance to see in Chapter 2, obeys this very well: We performed the thought experiment of positioning a camera along the squeezed direction and we looked at the surface density profile that resulted from compressing the system. The ideal trapped 2D gas undergoes a condensation, and, upon compression, the 3D ideal Bose gas experiences a smooth crossover from three-dimensional to two-dimensional behavior: The surface density evolves smoothly into the expected 2D isotropic density profile, with the condensate density having exactly the same shape all through the process but becoming progressively more important, and the condensate fraction changes from a cubic to a parabola, exactly as one would expect. A close look at that graph suggested the predictable result that if we squeeze a system we can make it condense at higher temperatures. A satisfying feature of the theory of BEC in traps is that interactions can be well accounted for, in all their complexity, by the introduction of just one additional parameter, the s-wave scattering length a of the atoms under study, that then enters the dynamics of the system through the coupling constant g. In Section 3.2 we saw how the coupling constant could be expressed consistently in our system of units, in which it appears as γ. While in three dimensions γ is clearly interpretable in terms of a, in 2D the connection is less transparent; our results in Chapter 4, for example, were found using a feasible but arbitrary coupling constant. However, our study of the zero-temperature gas in Section 3.5 gave us an expression for the equivalent 2D coupling constant of a highly anisotropic 3D gas that provides the missing piece in our analysis. In this chapter, then, we will study the compressed 3D gas, and we will do so using path-integral Monte Carlo simulations [32,33,44,69], a method that has been very useful in the study of liquid helium [45,126,127] and that has also been used for trapped and homogeneous [128] dilute Bose systems. We will start by giving a review of the method and its implementation and show some of the results we obtained for isotropic traps. We then simulate a gas of increasing anisotropy and study the behavior of its surface density using the Hartree-Fock approximation. 79 5.2 Path integrals in statistical mechanics Perhaps the main difference between this method and the mean-field theory we developed and used in past chapters is the fact that in the Monte Carlo simulation we will not impose Bose symmetry by invoking the grand canonical ensemble but rather by employing a trick based on our considerations in Appendix B. In the canonical ensemble, the N -body density matrix of a system of distinguishable atoms is related to the Hamiltonian that governs the motion through ρ = e −β̃ H̃ . The inverse temperature β̃ is a scalar parameter, and if we break it up into summands, say β̃ = β̃1 + β̃2 , the relation ρ = e−β̃ H̃ = e−(β̃1 +β̃2 )H̃ = e−β̃1 H̃ e−β̃2 H̃ (5.1) holds exactly. We can repeat this process as many times as we want, dividing β̃ into M parts—which for convenience we will take to be equal, though they need not be—and if we interpose complete sets of N -body position eigenkets (|Rk i, where Rk ≡ [x1,k , . . . , xN,k ]) between each pair of operators we can write ρ(R, R0 ; β̃) = Z dNσR2 · · · Z dNσRM ρ(R, R2 ; τ ) · · · ρ(RM , R0 ; τ ). (5.2) We can gain considerable insight into Eq. (5.2) by interpreting β̃ as the total duration of a process.1 By dividing the time interval [0, β̃] into a series of discrete time steps, each of duration τ , we can picture each member of the sequence R ≡ R1 , R2 , . . . , RM , RM +1 ≡ R0 as being an intermediate configuration (or “time slice” [45]) adopted by the system as it evolves from R 1 to RM +1 . Indeed, if we make M → ∞ and τ → 0 while keeping β̃ finite, this sequence becomes a continuous path (hence the name). The density matrix can then be interpreted as a propagator that obeys the usual composition property [46], and Eq. (5.2) states that in its evolution the system adopts all possible intermediate configurations; the density matrix results from summing over all the time slices; see Fig. 5.1. Beside providing this physical picture, the path integral also constitutes a useful calculational tool: since β̃ ⇔ T and hence τ ≡ β̃/M ⇔ M T , each of the intermediate density matrices is to be evaluated at a high temperature, where it is amenable to accurate analytic approximation. The exact convolution property (5.2) conveniently lowers the temperature at each step and leaves us at the end with the low-temperature density matrix that we desire; at no point in the process 1 Rigorously speaking, this time is imaginary [46]. Caution must be had when interpreting this variable: For example, as the temperature of a system decreases, so do on average the kinetic energies of the particles that compose it; but this means that the uncertainty over their position increases, their wavefunctions are more spread out, and, as a consequence, they have larger displacements over τ —they move “faster.” [45] 80 x1,M +1 x2,M +1 R7 = R 0 β̃ ≡ 6τ 5τ R6 4τ R5 3τ R4 2τ R3 τ R2 R1 = R 0 x1,1 x2,1 Figure 5.1. The density matrix as a path integral. The one-dimensional two-particle system shown above is described by a density matrix that can be expanded into paths by means of Eq. (5.2); the figure shows two possible paths traversed by each particle. The inverse temperature β̃ of the system is split into M = 6 “time slices”; the density matrix at each of these corresponds to a system temperature ∝ M T and can therefore be calculated more easily. When this is done, one must “sum” over all possible paths. In the figure, all paths end at their starting points; these closed paths suffice to determine the thermodynamics of the system. do we have to make any essential approximations [45] other than those inherent in our choice of interatomic potential. The price one pays for this convenience comes in the form of an integral in N × M × σ dimensions. The thermodynamics of the distinguishable system depends only on the diagonal (R = R 0 ) elements of the matrix, or, in the language of path integrals, only on paths that close upon themselves (frequently called “closed polymers” [45]). This changes when we introduce Bose statistics into the system, which we do by combining Eqs. (5.2) and (B.2) and rewriting the density matrix as 1 X ρ(R, P R0 ; β̃) N! P Z Z 1 X = dNσR2 · · · dNσRM ρ(R, R2 ; τ ) . . . ρ(RM , P R0 ; τ ), N! ρB (R, R0 , β̃) = (5.3) P where we have summed over all possible permutations of the indices that label the distinguishable particles.2 The paths that yield the thermodynamic properties are now more complicated because they are no longer closed: the system does not have to end in its initial configuration anymore, but can do so in any other configuration where the atom labels have been permuted. 2 If N = 2, for example, we have to consider the two possibilities P R ≡ P [x1 , x2 ] = [x1 , x2 ] and P [x1 , x2 ] = [x2 , x1 ], add them together, and divide by 2! = 2; see Appendix B for a more detailed discussion. 81 This can be seen by considering a given permutation P , which, as we saw in Appendix B, can be decomposed into cycles. The closed paths from the distinguishable system will still be trod by the atoms whose labels are left untouched by the permutation and thus belong to 1-cycles. Any two particles belonging to an exchange (or 2-) cycle, say 17 and 134, will have their paths intertwined: x17 will be x134 at the end of the convolution, and x134 will become x17 , regardless of the shape of their paths in the intervening time slices. Atoms whose labels belong to longer cycles will lie on progressively “grand[er] ‘ring-around-the-rosy’ exchange loops” [53]. At high temperatures, the identity permutation dominates, and most atoms will be in 1-cycles; the behavior of the Bose system will then be well described by a Maxwell-Boltzmann distribution in which the indistinguishability of the atoms is accounted for by the Gibbs factor 1/N !. At lower temperatures, when the de Broglie wavelength and the σth root of the inverse density become comparable, the particles’ wavefunctions start to overlap and the longer exchange loops become progressively important. As we saw in Chapter 2, the appearance of long permutation cycles is a signature of BEC [44,53]. Thus the additional degree of freedom granted by the permutations gives rise to Bose-Einstein condensation [129], and Eq. (5.3) incorporates this feature exactly. The price paid to lower the temperature by means of the composition property is raised even further when we have to perform an additional sum over N ! possible label reshufflings, and the task of directly evaluating the density matrix (5.3) is not just daunting but outright impossible. However, every summand and integrand in (5.3) is positive, and thus it is possible to sample the density matrix using the Monte Carlo method and use it to calculate thermodynamic averages of other quantities. 5.3 The Monte Carlo method and the Metropolis algorithm Even if it were feasible to set up and monitor a large-enough uniform grid on which to carry out the function evaluations required for an integral like (5.3), such an effort would quickly lead nowhere. The following simple example [130] illustrates why this is so. In a 10-point grid like • ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ • (5.4) most of the evaluations will be performed in the interior of the grid—as it should be, since in order to be integrable over an unbounded domain the function must fall steeply to zero with increasing 82 values of the argument. As the dimensionality of the domain grows, the number of interior points plummets as a fraction of the total: in two dimensions, the grid • • • • • • • • • • • ◦ ◦ • ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ • ◦ ◦ ◦ • • ◦ ◦ • • ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ • ◦ ◦ ◦ ◦ ◦ • ◦ • • ◦ ◦ • ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ • ◦ ◦ ◦ • • ◦ ◦ • • • ◦ • ◦ • ◦ • ◦ ◦ ◦ • (5.5) • • • • has 64 interior points and 36 endpoints. The three-dimensional grid that results from stacking eight copies of (5.5) on top of each other and adding lids has 512 points in its interior and 488 at its edges—almost half the total. In four dimensions the fringes already dominate, and in ten dimensions the interior holds only 10% of the points. In general, a p-point grid has p − 2 interior points per dimension, and the fraction of points within a σ-dimensional hypercube is µ p−2 p ¶σ = µ 2 1− p ¶σ = eσ log(1−2/p) , (5.6) a number that decays to zero exponentially as σ grows. Hence any attempt to calculate (5.3) using a uniform grid is just a wasteful time of calculating the number zero. Gaussian quadrature methods like those reviewed in Appendix C and used extensively in this thesis improve this situation somewhat by using nonuniform grids and giving more weight to the interior points, but the sheer number of dimensions involved in (5.3) renders this method useless as well. The Monte Carlo method, in which the function is evaluated at random points, has been developed to deal with this problem. The classic example of a Monte Carlo method is also the simplest: to calculate π, we only have to enclose the unit circle in a square of side 2 and choose a million points at random inside the square; the ratio of the number of points that also fall within the circle to the total number of points will then be π/4. Alternatively, to calculate an integral of the form Rb f (x) dx we can choose M numbers x1 , . . . , xM at random in the interval [a, b]; the mean-value a theorem then tells us that Z b a M 1 X f (x) dx ≡ (b − a)hf i ≈ (b − a) f (xi ), M i=1 83 (5.7) and we say that the function f has been “sampled” over [a, b]. The error in the approximation falls √ as 1/ M [131], so for M large enough Eq. (5.7) becomes a good estimate of the integral. Trivial as it may seem from these two examples, the Monte Carlo method can be used for problems where the domain of integration is so complicated that other methods are considerably more difficult, or even impossible, to apply—and it is the only reliable method to calculate integrals in high-dimensional spaces. Note, however, that we managed to calculate these two integrals so easily because we knew the kind of random numbers that we had to generate: the points in the first example had to lie within the square, and every xi in the second obeyed a ≤ xi ≤ b. Let us concentrate on this second Rb example, and, moreover, let us view it not as the calculation of the integral a f (x) dx but as that of the average value hf i. We can rewrite Eq. (5.7) as 1 hf i = b−a Z b f (x) dx = a R∞ f (x) %(x) dx −∞ R∞ −∞ %(x) dx (5.8) where we have introduced the probability distribution %(x) = ( 1 0 if a ≤ x ≤ b, otherwise. (5.9) These equations can be interpreted in one of two ways. We could calculate the integral by generating random real numbers uniformly distributed along the complete real axis (however that may be done), sifting these through %(x), and only then calculating the average of f . It should be clear that this process would be difficult and wasteful. On the other hand, we can go through the process in the way we described above: We can select random numbers already distributed according to %(x) and use that new uniform distribution to sample the function. In general, the probability distribution % will not be uniform, and this process of importance sampling must reflect the shape of the distribution—in other words, we must generate more random numbers (or, in higher-dimensional spaces, configurations) lying within the intervals where %(x) is large than within the intervals where %(x) is small. At this point we can connect these considerations to those in the preceding section. Equation (5.8), rewritten as hf i = Z dx f (x) R %(x, x) Tr %f , = Tr % dx %(x, x) (5.10) shows that it is possible to identify the average of f with an ensemble average, provided that we take the probability distribution % to be the density matrix of the system. Thus given a large enough 84 number of xi we can find the ensemble average of any relevant function f by simply calculating an arithmetic mean. The problem is reduced to finding a sufficient number of configurations, distributed according to the density matrix, that i) will be in the region where the distribution—which, of course, now includes the interactions—is significant and ii) will still be diverse enough to constitute a representative sample. The high dimensionality makes it impossible to propose configurations “by eye,” as one can easily do in the examples above, and even if that were possible it would still be faced with the nontrivial task of normalizing the density matrix. There is, however, a method, originally introduced by Metropolis et al. [132] to study a classical system of interacting hard spheres, that takes care of these problems at a single stroke. Instead of setting up a different configuration every time, we start with a plausible configuration and make it walk through configuration space by having it change it at every step, in what is known in the literature as a Markov process.3 The Metropolis algorithm, illustrated in Fig. 5.2, consists of the following steps: 1. Start with a random configuration Cold and calculate its probability %old . 2. Create a new configuration Cnew by displacing one or more of the atoms at random and calculate its probability %new . 3. If %new > %old then assign Cold → Cnew . 4. If %new < %old generate a random number 0 ≤ ν < 1. If ν < %new /%old then once again assign Cold → Cnew ; otherwise keep Cold but count it as a new configuration. To justify the Metropolis algorithm further we introduce the transition probability π(R s → Rt ) that a system will move from the configuration Rs to Rt , and the probability distribution %(Rs ) obeyed by Rs . If the system is visiting Rs at a given instant, it obviously has to have attained its present configuration by having been in any other (not necessarily different) configuration in the immediate past and having undergone a transition to where it is now. This can be written as %(Rs ) = X t π(Rt → Rs ) %(Rt ). 3 (5.11) We assume that transitions happen ergodically; in other words, it is possible for the system, when it starts at a given configuration, to reach any other configuration in a finite (though not necessarily small) number of steps. 85 2 1 ♦ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Total Tests 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ 14 5 ? ? ♦ ? ♦ ♦ ♦ ? ♦ ? ? ♦ ♦ ♦ ? ? ♦ ♦ ? 0 1 ♦ ♦ ♦ ♦ 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 Total Tests ♦ ♦ ♦ ♦ ♦ ♦ ♦ ? 10 5 6 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ 18 4 ? ? ? ? ? ♦ ♦ ? ♦ ♦ ♦ ♦ 6 2 ♦ ♦ ♦ ♦ 2 2 Figure 5.2. A simple game that illustrates the Metropolis algorithm. A diamond, confined in the box shown at the top, is twice as likely to be in region 0 than in region 1 and cannot be anywhere else; it can move to the left or right with equal probability. The second column of the table displays a string of random bits: the first one is 0 and prescribes the initial condition; after that, a 0 will order a move to the left and a 1 will order a move to the right, as long as it is into a region of nonzero probability. All moves from region 1 to region 0 are accepted, while moves from region 0 to region 1 are subjected to a Metropolis test: the move is accepted if the next bit in the string is 1 and rejected if it is 0. The third column shows the first 48 configurations: of the 16 decisions (shown by question marks) that were taken, 9 ordered the diamond to stay in place; at the end, the diamond is at the left in 32 configurations—exactly 32 of the total. The fourth column shows the effect of not counting “failed” tests as new configurations: the diamond spends roughly half the time in each region, as if the distribution were uniform everywhere. Note that these results were obtained without ever needing to normalize the probability distribution. 86 In the next possible instant, the system will either stay in its present configuration or undergo a transition to any other possible configuration; hence X t π(Rs → Rt ) = 1. (5.12) Equations (5.11) and (5.12) imply that %(Rs ) X t π(Rs → Rt ) = X t π(Rt → Rs ) %(Rt ), (5.13) which restates the definition of % as an equilibrium distribution: the rates at which the system goes into and out of any state s should be equal [133]. A sufficient condition for (5.13) to be satisfied is the detailed-balance condition [130,133]: %(Rs ) π(Rs → Rt ) = %(Rt ) π(Rt → Rs ). (5.14) The detailed-balance condition also expresses the fact that the microscopic dynamics of the system obeys the time-reversal invariance demanded by quantum mechanics [133]. The Metropolis algorithm is equivalent to choosing the transition probability µ ¶ %(Rt ) π(Rs → Rt ) = min 1, . %(Rs ) (5.15) It is easy enough to verify that (5.15) obeys (5.14), as can be seen from the following table [134]: π(Rs → Rt ) %(Rs )π(Rs → Rt ) π(Rt → Rs ) %(Rt )π(Rt → Rs ) %(Rt ) > %(Rs ) %(Rt ) < %(Rs ) 1 %(Rs ) %(Rt )/%(Rs ) %(Rt ) %(Rs )/%(Rt ) %(Rs ) 1 %(Rt ) (5.16) Note that there is a nonzero probability that the system will make a transition from a more-probable system (i.e., a system with lower energy) to a less-probable one (with higher energy); this possibility is, of course, not unphysical, and during the program run it will prevent the system from getting stuck in a local energy minimum [63]. On the other hand, the inverse transitions occur with 100% likelihood, and this encourages the system to visit more configurations and hence sample % more efficiently. 87 Consider again the game we played in Fig. 5.2. The diamond can be either in region 0 or in region 1; we do not know beforehand the absolute probability with which the diamond will be in any of them, but we know that the configuration R0 (with the diamond at 0, that is) is twice as probable as the configuration R1 and thus %(R0 )/%(R1 ) = 2. At any point in the game, if the diamond is in region 1 and is told to move to the left, it will do so with probability min(1, 2) = 1; if it is in region 0 and is told to move to the right, the probability for it to do so is min(1, 21 ) = 1 2; hence the choice of a random bit to make the decision for us. Equivalently, we could have chosen a real number 0 ≤ ν < 1 at random and have the decision pend on whether the number was greater or smaller than 1 2. Had the diamond been three times as likely to be in region 1 than in region 2, we could have used a three-faced die instead of a coin to help us decide, or, equivalently, we could have generated a random number and compared it to 31 . Now, suppose we perform many measurements in parallel on a large number of systems that are in equilibrium and far away from each other. Consider any two states s and t, with %(R s ) < %(Rt ) for definiteness. In a given move, we will have Ns π(Rs → Rt ) = Ns systems going from s to t and Nt π(Rt → Rs ) = Nt %(Rs )/%(Rt ) going from t to s, so the net number of systems that undergo the transition Rt → Rs is Nt→s = Nt π(Rt → Rs ) − Ns π(Rs → Rt ) = Nt µ %(Rs ) Ns − %(Rt ) Nt ¶ . (5.17) It is clear that Nt→s will be positive until Ns /Nt exceeds %(Rs )/%(Rt ), at which point it will become negative; as more moves take place, the ratio Ns /Nt will oscillate about %(Rs )/%(Rt ) until equilibrium is reached, at which point the ratio Ns /Nt for every s and t will have precisely the value required by the equilibrium distribution %. We would have reached the same conclusion by looking at the net number of systems going in the opposite direction: Ns→t %(Rs ) = Ns %(Rt ) µ Nt %(Rt ) − %(Rs ) Ns ¶ . (5.18) At this point we should note that, since we are working in the canonical ensemble, and using at every step a system of distinguishable particles, at equilibrium we should reach the Gibbs distribution, %(Rs ) ∝ exp(−β̃E(Rs )), where E(Rs ) is the energy associated with the configuration Rs . An additional important feature of the Metropolis algorithm is that we do not need to calculate the total energy at every step, but rather the change in energy elicited by the transition: %(Rs ) = exp(−β̃(E(Rs ) − E(Rt ))) ≡ exp(−β̃∆E). %(Rt ) 88 (5.19) In our simulations we will not consider energy differences but ratios of density matrices; Eq. (5.19) tells us that whenever we make a Metropolis decision we need to take into account only the parts that changed when we generated the trial configuration by displacing the original one. This significantly reduces the number of computations that we must perform and makes the simulations run significantly faster. The reader might have noted that in our previous discussion of the Metropolis game we did not mention the moves (into “forbidden” regions) that always ended with the diamond staying in place. These can, of course, be interpreted as moves into regions with zero probability that invariably failed the Metropolis test. Though we did not include these moves in the calculation, we still obtained the expected results because the Metropolis algorithm depends only upon probability ratios. We cannot, however, dismiss these failed tests as a mere waste of bits because moves almost as improbable will always be present; but we can try to avoid them. To that end, we observe that the probability that a configuration Rs will undergo a transition into another configuration Rt is in fact a product of two factors: one one hand, there is the probability that a given transition will be attempted; on the other, there is the probability that the transition, once proposed, will actually take place. While the latter is still fixed by %, we can make use of everything we know about the system to try to tweak the former in order to enhance the occurrence of transitions and make the system walk faster through configuration space. Figure 5.3 on the next page shows the result of playing the simple game of Fig. 5.2 using better selection probabilities. Region 0 is still twice as likely to be visited by the diamond as region 1, but now we have an additional mechanism that pushes the diamond away from the fringes and into the region of interest, guaranteeing a richer sample of relevant configurations. (In fact, we could have played the game, and obtained the same results more cheaply, with an additional rule that would make it impossible for the diamond to even attempt moves outside of regions 0 and 1; this, however, is possible only if we know everything about the system.) The PIMC algorithm that we now describe makes extensive use of these ideas. 5.4 An algorithm for PIMC simulations of trapped bosons Research papers on computational methods usually devote most of their allotted space to the fundamental ideas behind the algorithms they discuss and very seldom give insight into the details of their calculations. It is virtually impossible for the nonexpert to reconstruct one of those programs just from studying the papers, and most textbooks on Monte Carlo methods deal almost exclusively 89 2 1 ♦ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Total Tests − − 1 0 − 0 0 0 0 0 0 0 0 1 0 2 1 + 1 2 + 1 + 2 + 2 2 + 1 + 1 + 2 ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ 15 2 0 1 ? ? ♦ 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 Total Tests ? ? ♦ ? ♦ ? ♦ ♦ ? ♦ ? ♦ ? ♦ ♦ 9 7 0 1 0 0 0 0 − − 0 0 − − − − − 0 1 2 + 2 + 1 1 1 + 2 2 1 2 2 2 1 2 − ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ 17 7 ? ♦ ? ♦ ? ? ? ♦ ♦ ♦ ? ? ? ? ? ? ♦ 6 4 Figure 5.3. A variation on the game of Fig. 5.2 that shows the effect of choosing better a priori probabilities. The rules are the same as before, but now there are two kinds of random numbers. The integers 0–2 order the diamond to move in such a way that it tends to be pushed out of the edges or into the regions with finite probability: when it is in region 0, the number 0 will order it to move to the left, while both 1 and 2 order a move to the right; when it is in region 1, both 0 and 1 will order a left, while 2 will make it go right. The random bits, now called + and −, prescribe the initial condition and decide the Metropolis tests. The diamond is at the left in 32 configurations once again, but now there are 25% more Metropolis decisions, and the diamond moves around much more often; had we had more than just two regions, we would have managed to sample more “admissible” configurations instead of staying at the edges. 90 with lattice models [130,133] or one-body systems [131]. This section hopes to fill that void in the literature by providing a “narrative” description of the PIMC simulations of trapped bosons pioneered by W. Krauth [32,33,44,69]; the more technical aspects of the calculation will be relegated to Appendix D. The fundamental parameters in the simulation are N , the number of atoms in the gas, a 0 , the hard-sphere radius, and σ, the dimensionality of the system. We have seen in Section 5.2 that the inverse temperature β̃ and the number of slices M are not independent, but are instead related through β̃ = M τ . The time step τ is of fundamental importance: it has to be small enough to guarantee that the high-temperature density matrix is accurate, and at the same time it has to be large enough to ensure manageable array sizes and reasonable run times. In his paper [44], Krauth recommends τ = 0.01 for N = 104 , a value arrived at after “extensive tests.” We shall use that value for that particle number, and we will increase it to τ = 0.02 for N = 103 , where the densities attained by the system are much lower (we will later see that this value is indeed acceptable). Since τ is fixed, the temperature enters the program in the number of slices. (One of the virtues of the program is precisely the large size of τ : the critical temperature for 104 atoms is reached with only five slices.) The anisotropy factor λ is the last fundamental parameter we need; it enters the simulation when we set up the initial configuration, when we generate new configurations, and during the calculation of interaction corrections. In all of these instances, λ will appear as part of an ideal-gas density matrix, whose general form in the presence of an anisotropic trap has been displayed in Section 2.5 and Appendix A; further details are presented in Appendix D. The high-temperature density matrices will be calculated by means of a pair-product approximation [135], ρ(R, R0 ) ≈ Y i ρ1 (xi , x0i ; τ ) Y i<j ρ2 (xi , xj , x0i , x0j ; τ ) , ρ1 (xi , x0i ; τ )ρ1 (xj , x0j ; τ ) (5.20) where ρ1 stands for a one-body density matrix and ρ2 is a two-body density matrix. This approximation treats two-body collisions exactly but neglects three- and higher-order processes (which should be rare in a dilute gas at the high temperature prescribed by a small-enough value of τ [45,136]). Even this simplified description, however, requires the laborious enumeration of interacting pairs and the detailed collision analysis that usually constitute the bulk of the work done in an N -body simulation. For a short-range potential like the one we consider, it is clear that most collisions will have negligible effects; in fact, it is possible to find a maximum interparticle separation—an “effective range”—beyond which the interaction becomes unimportant. This parameter must also be found 91 4057 274 275 6068 8214 254 255 2121 7382 Figure 5.4. Boxes and interactions. When given an atom number, the routine Interboxes finds the box (of side L) to which it belongs and scours the neighboring boxes searching for other atoms that might be close enough to it for the two to interact. As the figure shows, the search can involve between one (2121) and four (6068, 8214) boxes in two dimensions, even though each box has eight neighbors; the variable Range defines the circles’ radii. In general, we need to look at 3 σ boxes, of which 2σ will contain potential pairs. In the case depicted here there is only one interacting pair, with the atoms belonging to different boxes. The box labels are consistent with a 20 × 20 mesh in two dimensions, while the particle labels are typical of a 10000-atom gas; it should be obvious that the figure shows only a minute fraction of the atoms that would be present in a realistic simulation. by hand by studying the two-body density matrix quantitatively. In the end it turns out to have a value Range = 0.2. The enumeration of pairs can be streamlined significantly through the introduction of a σdimensional mesh that divides the system into boxes larger than the interaction range; more precisely, the cube side L must obey L > 2 × Range so that an atom located at the very center of one such box will have a sphere of radius Range around it that will be completely contained by the cube. (Krauth’s version has Nbox = 20 boxes per dimension for a total of 8000 boxes; every interior box has L = 4 9 ≈ 0.44, while those on the edges extend outward to infinity.) That way, instead of going through the complete list of atoms to see if they interact with a given particle, we only have to search within the box that contains the particle in question and within the neighboring boxes; in fact, we don’t even have to look at all the adjoining boxes, as Fig. 5.4 shows. A list of boxes is generated (by a subroutine called Interboxes) by literally drawing a sphere of radius Range around each atom at every slice and seeing what boxes are contained within each sphere; an important auxiliary array keeps tally of the boxes that have to be considered at every step. With these considerations in mind, we are ready to introduce the arrays whose evolution constitutes the simulation. The first one, Rinner, is an N × σ × (M + 1) array that contains the system configuration at every time slice at any given point in the run; Ibox, of dimension Nbox σ × Maxbox × 92 (M + 1), where Maxbox is a large number (or can be left to vary if the programming language allows for dynamic resizing of arrays), keeps track of the atoms contained in each box at each time slice. Finally, Iperm keeps at every step the permutation that characterizes the system (along with its inverse), and obviously has dimension N × 2. At the beginning of a typical run we assign the initial positions of the atoms by sampling the diagonal elements of the density matrix for distinguishable particles, ρ(x) ∝ e−(ξ 2 tanh β̃+λη 2 tanh λβ̃) . (5.21) With β̃ and λ as parameters, and given the Gaussian form of (5.21), it suffices to generate three normally distributed random numbers per atom, one for each direction, centered on the origin and with standard deviations (2 tanh β̃)−1/2 for x and y and (2λ tanh λβ̃)−1/2 for z. This array is replicated in the slice direction in order to initialize Rinner. A call to the boxing routine Interboxes assigns a box number to every atom, and a special-purpose routine creates the list of atoms contained in each box. The permutation of atom labels is initially taken to be the identity. The program then walks the system through configuration space by generating moves, subjecting the moved configurations to Metropolis tests, and taking data. There were usually between 105 and 106 such steps in a run; as in all other Monte Carlo simulations, it is necessary to “equilibrate” the system to make it lose memory of the initial, unrealistic configuration: we usually did some 105 of these “thermalizing” moves. Each move consists of a configuration shift followed by several permutation moves. We will start by discussing the former. To increase the probability of acceptance, we move only particles that belong to the same permutation cycle (they are already located close to each other and possibly quite entangled). The program chooses an atom at random and finds the cycle to which it belongs and the length of that cycle; our knowledge of the global permutation makes this straightforward. In order to make moves that will stand a chance of being accepted while walking the system through its enormous domain at a brisk enough pace, it is necessary to have a maximum number of atoms that can be moved in a given step. This maximum number can only be determined by trial and error, and a value Lmax = 5 has been found to be appropriate. Two things can happen: either the cycle length Len ≤ Lmax or Len > Lmax. When Len > Lmax we move only part of a cycle and leave its endpoints untouched, since they are connected to the atoms “right before” and “right after” in the closed cycle (this is illustrated in Fig. 5.5). When Len ≤ Lmax we move the whole cycle to a completely new random position: the endpoints are relocated using 93 ← ∨ ∨ ∨ ⇒ ⇒ ← ∧ ∧ ∧ Figure 5.5. The two steps involved in moving part of a cycle. Initially, the atoms in the cycle segment (in this case just one) are displaced to new random positions obeying the distribution (D.15) dictated by the endpoints (marked by wedges). In the second step, the intermediate time slices are displaced using the same distribution (dictated now by the positions just found) and threaded in. The endpoints are unaffected. β̃ 3β̃ 0 2β̃ ⇐⇒ ⇓ β̃ β̃ 0 0 Figure 5.6. Moving a complete cycle. To move the complete 3-cycle shown above we proceed as in Fig. 5.5, but first we have to move the endpoints to a completely new random position; when doing this we must take into account that a 3-cycle at inverse temperature β̃ (left) is equivalent to a 1-cycle at inverse temperature 3β̃ (right). 1 3 2 4 1 2 3 4 Figure 5.7. Interactions between particles of changing identity. The figure shows a four-particle gas in a configuration described by the permutation {1, 2, 3, 4} → {1, 3, 2, 4} or (1)(2, 3)(4); we have taken the temperature to be so high that only one time slice is necessary. At the beginning of the path, atom 2 interacts with atom 1 and atom 3 interacts with atom 4; at the end, by virtue of the exchange cycle, atom 2 interacts with atom 4 instead, and atom 3 with atom 1; this possibility has to be taken into account during every move. 94 the ideal-gas routine from the start of the run while keeping in mind that the particle belongs to an l-cycle; in other words, its probability distribution is now governed by the partition function Z 1 (lβ̃), as we saw in Appendix B, and hence we have to evaluate the standard deviations at an inverse temperature lβ̃. (Figure 5.6 illustrates the equivalence; see also Ref. 136.)4 In both instances we thread in the positions of the intermediate particles using the interpolation algorithm described in Appendix D and illustrated in Fig. D.1. Once we have made the move, we once again use the algorithm from Appendix D to thread in the atoms’ positions at the intermediate time slices. Since the particles have moved, it is quite likely that a few of them are now located in new boxes; Interboxes updates this information at the end of every move. When a move is proposed it is subjected to the Metropolis test. It is at this stage that the interactions enter the simulation. Recalling Eq. (5.20), we see that at a given time slice k the density matrix to be calculated is bracketed between the configuration Rk and Rk+1 , the configuration at the next time slice; the typical term will then look like this: ρ(Rk , Rk+1 ; τ ) = Y ρ1 (xi,k , xi,k+1 ; τ ) i × Y i<j Ξ(xi,k − xj,k ) ρHC,1/2 (xi,k − xj,k , xi,k+1 − xj,k+1 ; τ ) Ξ(xi,k+1 − xj,k+1 ), ρHO,1/2 (xi,k − xj,k , xi,k+1 − xj,k+1 ; τ ) (5.22) where we have used the notation of Appendix D. We need only to calculate the change in energy as prescribed by Eq. (5.19), so for the calculation of each energy it suffices to consider only the particles belonging to the cycle that was moved. The problem is then reduced to finding the pairs of particles that interact with those particles at each time slice, a task simplified and hastened by our knowledge of the boxes containing each atom, evaluating the pair-product density matrices in (5.22) using the expressions derived in Appendix D, and calculating the ratios. There are two subtleties that have to be taken care of during this procedure. The first one is rather obvious: if both members of the interacting pair are part of the cycle that was moved, their interaction will be counted twice (in other words, both (i, j) and (j, i) will appear as interacting pairs at the end of the search), and it is therefore necessary to either discard one of the pairs or divide both interactions by 2. The second one has to do with the fact that, in the last time slice, 4 Note that the higher inverse temperature implies a smaller standard deviation in the diagonal density matrix; this feature is responsible for the density enhancement at the center of the trap that characterizes our spatial BEC. 95 each particle has become the next particle in the cycle, and therefore we have to find the particles that interact with that one, as Fig. 5.7 illustrates. At the end of the Metropolis test, the new configuration will either be accepted or rejected; in the latter case, we have to restitute the earlier configuration and count it again. After the configuration move, the program then tries Nperm = 10 permutation moves, which consist simply of swapping the positions of two of the atoms (in the last slice, as Eq. (5.3) prescribes) and interpolating their positions in the intermediate time slices (this is necessary because otherwise the resulting configurations would be highly improbable and almost invariably rejected [33,45]). To increase the probability of acceptance, only permutations between atoms located in the same box are attempted. The Metropolis test happens in two stages: in the first, the program calculates the ideal-gas density matrices in the old and new configurations and compares them; if the configuration passes this “ideal” test, then the move is treated exactly as the configuration moves described above. In this thesis we have been interested in calculating two quantities: the density of the condensed gas and the condensate fraction. The one-particle density of the system is given by ñ(x) = * N X δ (3) i=1 (x − xi ) + = M N 1 X X (3) δ (x − xi,j ), M j=1 i=1 (5.23) where M refers, as in Eq. (5.7), to the number of configurations that we sampled, not the number of slices. To find the density we proceed as in Eq. (5.10): at every step we pick a particle at random and record its coördinates, and, at the end of the run, we are left with a representative sample of the particles’ positions, which we can use to calculate the density by generating histograms. Every certain number of steps (100 in this case), the program also looks at the permutation characterizing the system, finds its cycle structure, and records the length of the various cycles; the cycle structure can be used to determine the condensate fraction. In an isotropic trap, the most direct and accurate way of determining the density is through the radial number density N (x), defined by N= Z dσx ñ(x) = Z dx N (x), (5.24) that we introduced in Fig. 2.1. As we may recall, N (x) = 2πx ñ(x) in 2D, while in 3D N (x) = 4πx2 ñ(x). This quantity is extracted from the Monte Carlo simulation simply by counting how many atoms are at a given distance from the origin; the corresponding histogram yields N (x) directly. For 96 the anisotropic trap, we instead consider the surface number density, defined in Eq. (2.43), which is extracted by counting how many atoms are at a given transverse distance from the origin. Krauth’s original method of extracting the condensate fraction [44] uses the fact—already mentioned in Chapter 2—that, even though the condensate contains particles in cycles of all lengths, the longest cycles must contain condensed atoms exclusively. Recalling our identification on page 95 of an l-cycle at inverse temperature β̃ with a single particle at inverse temperature l β̃, we see that such a cycle will be distributed, in the ideal gas, according to ρ(x) ∝ e−x 2 tanh lβ̃ , (5.25) and for lβ̃ À 1 they will be in the oscillator ground state given by Eq. (2.12) [44]. Another method [33,137] uses the cumulative distribution of atoms as a function of the cycle lengths; we will explain it in more detail when we discuss particular cases in the next section. 5.5 The anisotropic interacting gas at finite temperature We begin, naturally, by studying the ideal gas, which as usual will provide a check for the programs and for our choice of parameters. Figure 5.8 on the next page shows the number densities for a two-dimensional ideal gas of N = 1000 atoms at T ≈ 0.4056 Tc . The solid lines depict the Monte Carlo results obtained with τ = 0.02 and five slices after 8 × 105 steps, of which 6.5 × 105 were used for equilibration. The dashed lines show the exact results obtained using the methods of Chapter 2. In Fig. 5.9 we see how we can extract the condensate fraction from this Monte Carlo simulation. The figure shows a sampling, averaged over many configurations, of the cumulative distribution of particles as a function of the lengths of the cycles to which they belong; this function, defined below and denoted by q(l), is, provided directly by the simulation, as we already saw. From Eq. (B.6) we know that every single configuration of the system may be characterized by a permutation that can in turn be decomposed into c1 1-cycles, c2 2-cycles, etc.; all cycle lengths and numbers are constrained by the general relation N 1 X lcl = 1 ≡ q(N ). N (5.26) l=1 Moreover, as we saw at the end of Section 5.2, the occurrence of BEC is signalled by the appearance of nonzero cl ’s for high values of l. In the plot we show, averaged over many configurations, the 97 0.8 N (x)/N 0.6 PSfrag replacements 0.4 0.2 0 0 1 2 x 3 4 5 Figure 5.8. Monte Carlo and exact number densities of an ideal two-dimensional trapped gas of N = 1000 atoms at T ≈ 0.41 Tc . The solid line shows the result of a Monte Carlo simulation, while the dashed lines show the exact number density and its condensate and noncondensate components. Number of particles/N PSfrag replacements 1 0.8 0.6 1.1 0.4 1 0.2 0.9 750 0 0 200 400 600 Cycle length 800 850 800 1000 Figure 5.9. Extracting the condensate number from a PIMC simulation. The gas shown above has the same parameters as that of Fig. 5.8. The solid line depicts a configuration average of the function q(l), which tells us how many atoms (divided by N ) are in permutation cycles of lengths 1, 2, . . . , l. We can either take N0 to be the highest cycle length with q(l) = 1 or we can perform a linear fit (dashed line) and extrapolate the line to 1; the inset zooms in on the intersection. 98 function q(l) = Pl l0 =1 l0 cl0 /N , which tells us how many atoms (divided by N ) are in permutation cycles of lengths 1, 2, . . . , l. This cumulative distribution smooths out statistical variations [137] and is normalized to 1, all of which make it better to work with than the plain c l ’s; moreover, it yields two different estimates for N0 : Krauth’s method, justified at the end of the last section, assumes that N 0 is the smallest value of l such that q(l) = 1—in other words, the longest cycle containing particles; this is easily found by inspection. The other estimate for N0 results from fitting the large linear section of the plot and extrapolating to the value where this line would be 1. The inset shows the intersection point and its immediate vicinity: Krauth’s method gives N0 = 800 for these parameters, while extrapolation predicts N0 = 822; the exact value is N0 = 812, and lies almost exactly in the middle between the two estimates. We next study the behavior of the condensed ideal gas with increasing trap anisotropy. Figure 5.10 on page 101 shows the Monte Carlo solutions we obtained and compares them with the exact results. The figure is similar to Fig. 2.12; as in that one, we have N = 100, though now the (3) range of compression factors is somewhat smaller; the temperature, T = 1.3 T c , is a bit lower as well, but the gas is still uncondensed in the isotropic trap and clearly shows the appearance and growth of the condensate as a result of the increase in trapping frequency. A total of 3 × 10 5 steps, of which 2 × 105 were required for thermalization, were needed to produce the jagged lines that represent the Monte Carlo results (the simulation used τ = 0.01762 and attained the given temperature using ten slices). Note that the equilibration is essentially complete for the smallest and largest anisotropy ratios, while a few tens of thousands of additional steps would have yielded more satisfactory profiles for the intermediate values. The smooth lines represent, as usual, the exact solutions for the surface number densities, resolved into condensate and noncondensate components. Furthermore, Fig. 5.11 compares the surface number density of the gas with the highest anisotropy to the exact prediction of Eq. (2.43) and to the number density profile of an isotropic two-dimensional (2) ideal gas at the temperature T = 0.7279 Tc predicted by Eq. (2.39); as expected, the two analytic expressions give indistinguishable results, and agree quite well with the Monte Carlo histogram. The corresponding condensate fractions can be extracted from the plots shown in Fig. 5.12. This figure, which corresponds to a progressively compressed gas with the same parameters as Fig. 5.10, should be interpreted in the same way as Fig. 5.9. The inset zooms in on the region of interest, and the four numbers on the x-axes represent the exact predictions for the condensate fractions corresponding to λ = 1, 5, 15, and 50. The number of atoms is much smaller than in that figure, and it does not make sense to make a linear fit; to find the condensate fraction in this case we 99 must then content ourselves with the upper bound given by Krauth’s method, though, as the figure shows, even a crude fit would provide a reasonable lower bound (except for the totally uncondensed isotropic gas). Figure 5.13 on page 102 displays, as a function of the increasing anisotropy parameter λ, the asp pect ratio (defined, as we recall from Eq. (2.44), by p ≡ hξ 2 i/2hη 2 i) of a gas of N = 1000 atoms at (3) a temperature T = 0.5316 Tc . The circles show the ideal-gas aspect ratios obtained by PIMC sim- ulation, while the solid line corresponds to the ideal-gas prediction (2.44). The dotted line displays √ the aspect ratio λ of a pure condensate, and describes the system well at high compression. The good agreement between the Monte Carlo data and the exact prediction show that the value τ = 0.02 employed in the former is appropriate for this set of parameters. The diamonds exhibit the aspect ratio of an interacting gas (discussed below) with the same parameters and the usual coupling constant corresponding to rubidium; as expected, the aspect ratio of the interacting gas grows much faster with λ. We continue our study of the interacting gas by looking at the density profile of the large threedimensional trapped system at finite temperature that we studied in Chapter 3. Figure 5.14 on page 103 is in fact two versions of the same plot: the top panel is Fig. 2 of Ref. 69, which we used as a guide and as a check on most of the methods that we developed during the research reported in this thesis; by reproducing it on the bottom panel we have obtained a final check that all of our algorithms are consistent with each other and work correctly.5 Both panels show the density profiles of an ideal and an interacting 3D gas of 104 rubidium atoms at a temperature T = 0.705 Tc calculated using Monte Carlo simulations (solid lines); the dashed lines display the exact solution for the ideal case and the Hartree-Fock approximation to the interacting system. To avoid clutter, we have not resolved any density into condensate and noncondensate. As we have seen in the previous chapters, and now confirm with this new method, the repulsive interactions lower the density at the center of the trap, deplete the condensate, and widen the particle distribution. Some authors of competing PIMC simulations of bosons [31,125] report having spent approximately 50 hours of CPU time to obtain a single condensate fraction for N = 1000. Krauth’s 5 This reassurance is reinforced by the title of that article: “Precision Monte Carlo test of the Hartree-Fock approximation for a trapped Bose gas.” Note that the authors of the reference do not include a Monte Carlo simulation of the ideal gas in their figure; this latter profile is not completely resolved after sampling 5 × 105 configurations, even though the first 3 × 105 steps were used exclusively to make the system lose memory of its initial, distinguishable distribution, and even though the interacting gas has converged to satisfaction with the same number of steps. 100 PSfrag replacements 0.5 Ñ (ξ)/N 0.4 0.3 0.2 0.1 0 0 50 4 ξ 15 8 5 12 λ 1 Figure 5.10. Surface number density of a condensed ideal Bose gas of varying anisotropy. The gas has N = 100 atoms and is at a temperature T = 1.3 Tc . The jagged lines show the Monte Carlo results, while the smooth lines represent the exact solutions, resolved into condensate and noncondensate components (dotted lines). 0.5 N (x)/N, Ñ (ξ)/N 0.4 PSfrag replacements 0.3 0.2 0.1 λ 0 0 2 4 x, ξ 6 8 Figure 5.11. Front view of the rightmost plot of Fig. 5.10. The figure displays both the number surface density of the squeezed gas (dash-dotted line) and the density profile of a 2D gas at T = (2) 0.7279 Tc (dashed line). 101 1 PSfrag replacements Number of particles/N 0.8 16.2 1.1 40 0.6 1 0.4 0.2 0.9 0.6 37.4 60 0 0 20 40 60 Cycle length 80 100 Figure 5.12. Finding the condensate fraction of an ideal Bose gas of varying anisotropy using the procedure first shown in Fig. 5.9. The parameters are the same as in Fig. 5.10; the lines correspond to (from left to right) λ = 1, 5, 15, and 50. The inset shows the region where the condensate numbers should lie and displays the corresponding exact predictions. 6 hξ 2 i/2hη 2 i 5 PSfrag replacements 4 p= p 0 3 2 1 1 3 5 7 9 11 λ Figure 5.13. Aspect ratios, obtained by Monte Carlo simulation, of condensed ideal (circles) and interacting (diamonds) Bose gases of increasing anisotropy. Each gas has N = 1000 atoms and is (3) at T = 0.5316 Tc . Also shown are the exact aspect ratio of the ideal gas (solid line) and that of a pure ideal condensate (dotted line). 102 N(r) 0.5 ideal gas 0.4 HF QMC 0.3 0.2 0.1 0.0 PSfrag replacements 0.0 2.0 4.0 0 2 4 6.0 8.0 10.0 6 8 10 r/a 0 0.5 N (x)/N 0.4 0.3 0.2 0.1 0 x Figure 5.14. Monte Carlo and mean-field number densities of a three-dimensional trapped gas. The top panel in the figure is Fig. 2 of Ref. 69, which we used as a check on our methods, and has its own legend. The bottom panel displays the total number density of a gas of 10 4 atoms at T = 0.705 Tc . The thin, tall curve corresponds to the ideal-gas profile obtained by both the exact method (dashed line) and a Monte Carlo simulation (solid line). The other curve, which displays the Monte Carlo (solid) and Hartree-Fock (dashed) density profiles for an identical gas with repulsive interactions, shows the expected effects: a smaller density at the center and a larger cloud size. The coupling constant is γ = 8π × 0.0043, corresponding to rubidium atoms. 103 algorithm is orders of magnitude faster,6 and can handle many more atoms, but suffers from the same limitation. It is designed to cope with up to 104 atoms [44] and cannot go beyond that, since any further increase in this parameter will make the computer run out of memory: the array Ibox grows very quickly with N , and the search for interacting pairs, which can involve some Maxbox × 2σ × Lmax × Nslice ≈ 105 pairs per step in the present conditions, becomes unmanageable. Moreover, the resulting enhanced density would require a smaller value of τ and (possibly many) more slices. This was a pressing concern in our study of the squeezed 3D gas, the final topic to which we turn. We first made sure that our choice of τ was appropriate, and, as Fig. 5.13 shows, τ = 0.02 turned out to be acceptable for N = 1000. Though we did a few simulations with N = 10 4 , the run times were so long (and the data files so large, though this was the case for every set of parameters) that we could not repeat an experiment enough times until we trusted the results. For that reason, we (3) will concentrate on N = 1000 and use the same temperature T = 0.5316 Tc from Fig. 5.13. In this system, the crossover to two dimensions, as predicted by Eq. (3.39) of Section 3.5, should occur with a compression factor λc ≈ 5.07. Figure 5.15 on the next page displays the surface number densities obtained for this system with λ ranging between the isotropic value λ = 1 and λ = 11 > 2λc . We can see that it loses its noncondensate tail as λ increases and then changes very little for λ ' λ c , showing, at least qualitatively, that a finite-temperature crossover has indeed occurred. We then proceed to inspect the number surface density profile at the highest anisotropy and try to fit it with a pure-2D isotropic Hartree-Fock number density at the same temperature (which, for the parameters we are using, (2) corresponds to T = 0.2028 Tc ). From the result (3.46) found in Section 3.5 we should expect a coupling constant γ (2) = µ λ 2π ¶1/2 √ γ (3) ≈ 0.0043 λ ≈ 0.143 (5.27) at this compression ratio. The fit was close but not perfect, and we found that we could improve it by using a smaller effective λ. Figure 5.16 on the following page shows the best fit we obtained using λeff = 4. The condensate fraction is extracted from the Monte Carlo simulation in Fig. 5.17; the two methods we have used in previous examples once again bracket the condensate fraction predicted 6 The interacting density profile in Fig. 5.14 was generated by a fast machine in 2 hours and 44 minutes; in ours it took 11 hours and 47 minutes to generate the same profile. The ideal-gas profile needed 7 minutes in the fast machine and 25 in ours. For completeness, we remark that our exact ideal-gas program took about a minute to run, while the Hartree-Fock calculation took less than 15 seconds. 104 Ñ (ξ)/N PSfrag replacements 0.6 0.4 0.2 0 0 2 ξ 4 6 3 1 5 λ 7 11 9 Figure 5.15. Monte Carlo number surface densities of a three-dimensional trapped gas of increasing anisotropy. This system, like the one in Fig. 5.13, comprises N = 1000 rubidium atoms at T = (3) 0.5316 Tc . Ñ (ξ)/N, N (x)/N 0.6 0.4 PSfrag replacements 0.2 5 0 0 2 ξx 4 6 Figure 5.16. Monte Carlo number surface density and best-fit two-dimensional profile of an interacting three-dimensional Bose gas in a highly anisotropic trap. The parameters are the same as in the figure above, and λ = 11 as in the rightmost curve there. The Monte Carlo result is superimposed on the mean-field profile obtained, as explained in the text, by taking an effective anisotropy parameter λeff = 4. The condensed and thermal components are shown explicitly. 105 by the two-dimensional isotropic Hartree-Fock gas, with the difference between the mean-field and the Monte Carlo predictions being at most 5%. The condensate fraction, incidentally, changed by less than one percent within the range of λ that we considered. We repeated the procedure for N = 104 and found that, for a compression ratio of λ = 30 > 23.5 ≈ λc , a fit with λeff = 15.5 gave a similar agreement in both density and condensate fraction. In neither case did we see any significant difference when we used any of the other expressions for the coupling constant discussed after Eq. (3.47). Finally, we calculated and diagonalized the one-body density matrix corresponding to the best-fit 2D system; the results are exhibited on Fig. 5.18. The eigenfunctions are not very different from those that appeared in Fig. 2.10 (note, however, that they are displayed at a different scale); on the other hand, the occupation numbers show, if anything, a greater difference between the population of the ground state and those of the excited states; this is not too surprising at this very low temperature, and, at least at this level of description, would seem to indicate that the condensation that we are witnessing is not smeared. 5.6 Summary Path-integral Monte Carlo simulations provide an essentially exact description of finite interacting condensed Bose-Einstein gases; in particular, the method enables us to study how the properties of a trapped system change as the anisotropy of its confinement increases beyond the crossover condition and into the two-dimensional régime. We first give a self-contained account of the general method and of its implementation in the trapped case. Then, after we reproduce the ideal-gas and isotropic interacting results from previous chapters, we go on to study the interacting anisotropic trapped gas at finite temperature. We find that the two-dimensional Hartree-Fock solutions derived in Chapter 3 and studied in detail in Chapter 4 mimic the surface density profile and predict the condensate fraction of highly anisotropic systems to very good accuracy; the equivalent interaction parameter is smaller than that dictated by the T = 0 analysis of Section 3.5. Once again, we find no evidence of smearing. 106 PSfrag replacements 1 Number of particles 0.8 0.6 932.4 1.05 0.4 1 0.2 0.95 0 0 200 884 400 600 Cycle length 934 800 1000 Figure 5.17. Condensate fraction of a quasi-2D interacting Bose gas with N = 1000 and T = (3) 0.5316 Tc , extracted using the same method as in Fig. 5.9. The inset zooms in on the region of interest and displays the two estimates yielded by the simulation (bottom axis) and the value by the best-fit 2D Hartree-Fock calculation (top axis). PSfragpredicted replacements Population 1000 7 100 10 1 0.1 0 1 2 3 4 Eigenstate 5 6 8 φ0 (x), . . . , φ3 (x) 0.4 0.2 0 −0.2 −0.4 0 1 2 3 x 4 5 6 Figure 5.18. Occupation numbers and eigenfunctions of the one-body density matrix for a quasi-2D Bose gas with the same parameters as in Fig. 5.17. The figure shows the seven highest occupation numbers and the first four of the corresponding eigenfunctions. The y-axis in the top panel is logarithmic; the eigenfunctions are spline-interpolated (on a mesh similar to that of Fig. 2.10) and have not been normalized. 107 CHAPTER 6 CONCLUSION From a certain temperature on, the molecules “condense” without attractive forces, that is they accumulate at zero velocity. The theory is pretty, but is there also some truth to it? —A. Einstein, in a letter to P. Ehrenfest, 29 November 1924 [16] At the beginning of this thesis we set out to find whether Bose-Einstein condensation can occur in two dimensions or not. The obvious answer should be “Yes,” given that two-dimensional condensates have been produced in the laboratory, though at this point we cannot clearly ascertain if these are, say, quasicondensates that will become unstable as the size of the system grows—in fact, they could even be the uncondensed solutions that we considered in Chapter 4, since those also go to a (though not the) Thomas-Fermi limit as T → 0. On the other hand, and though the experimenters [30] were explicitly attempting to compress their traps beyond the crossover condition (3.39), 1 one can always argue that those systems are not really two-dimensional, since the interactions between the atoms are still 3D, and the peculiarities of pure-2D scattering (which we barely touched upon in Section 3.5) are at least partly responsible for the absence of two-dimensional condensation in the thermodynamic limit. As should be clear by now, we do not have a definite answer to our question. The methods we have used to study the problem have had mixed success, and there is much still left to do if we are to gain a deeper understanding of the subject. In the following we will summarize our results and conclusions and mention some open questions that we have left unanswered and to which we might come back in the future. Perhaps a good way to begin our summary is to retrace in sequence the steps that we took. When we began to study the topic, two-dimensional BEC was in no good shape: the Hohenberg theorem had just been extended to include trapped systems [28]; the semi-ideal model (briefly discussed on page 64) worked well in three dimensions [112] but gave unphysical predictions for 1 As we saw at the beginning of Chapter 4, the crossover condition is reached in experiments by varying N , not λ. 108 two [113]. On the other hand, there were signs that the situation was more complicated: Monte Carlo simulations had predicted a substantial accumulation of particles in the ground state of a finite system [32]; moreover, it had been found that, while the HFB equations (in the WKB approximation) were inconsistent in 2D as the thermodynamic (Thomas-Fermi) limit was approached, the form they took for a finite system without a condensate also failed below a certain temperature, signalling the possible presence of a transition in the system [43], probably of the Kosterlitz-Thouless type [35] or perhaps into a fragmented condensate [37]. These latter findings, plus the fact that the ideal trapped 2D gas [24,28] was as well understood and as well-behaved as its 3D counterpart, were the initial motivation for our work. Since there was no doubt that the 2D Bose gas has a condensate at zero temperature, it seemed to us that a reasonable starting point for our research would be the simple Gross-Pitaevskiı̆ equation. Our progress in that subject was discussed in Section 3.4; the wavefunctions we display on Fig. 3.1 were the very first results we obtained (and have been plotted using the original data file); as we have already remarked, these same results appeared in print [95] shortly afterward. The cumbersome method of solution that we had at that point was particularly difficult to employ for high values of N (which make the GP equation highly nonlinear), and we believed that that could be a consequence of the Hohenberg theorem, since these last curves corresponded to the “standard” 3D parameters used in the literature [44,87]. As an interlude, we concentrated on the ideal gas. The slowly convergent series present in the exact theory of Section 2.2 led us to digress for a while in the study of methods to accelerate convergence, though we abandoned it when we realized a (to us) key point: unlike what happens in the homogeneous case, in the trapped ideal gas the chemical potential becomes positive at low temperatures, and is equal to the single-particle zero-point energy. This discovery (which is in fact well-known [88] but very seldom acknowledged) led directly to a good part of the results presented in Chapter 2. We then got back to the interacting problem and tried to find finite-temperature solutions. We started in 3D, where the results of various flavors of the HFB equations were already available [88]. With great difficulty (since we were still using the initial-value method of solution) we reproduced the semiclassical Hartree-Fock results of Ref. 69 and, still using the semiclassical Hartree-Fock method, obtained good agreement with the discrete HFB calculations of Ref. 91. The semiclassical HFB solutions were more elusive, but we eventually made them converge by using our Hartree-Fock solutions as starting guesses; the results were essentially the same. 109 At that point we realized that the method we had been using was unreasonably laborious and slow, and we then turned to the search for a reliable and accurate scheme for solving the GP equation. We studied the spectral methods [64,65,105] described in Appendix C, but still could not deal satisfactorily with the nonlinearity. Eventually we became familiar with the spline-minimization method [44] that finally made it feasible to calculate density profiles and, in particular, yielded all of the results of Chapter 4. We then started to find partial answers to some of our questions. The semiclassical HartreeFock equations were soluble in 2D for a fairly wide range of temperatures and atom numbers; when we tried to go one step up in sophistication by using the HFB equations (still in the semiclassical approximation), however, we were never able to find any solutions, even when we iterated the selfconsistent procedure with the Hartree-Fock profiles as starting guesses. That has been the situation to this day. The semiclassical HFB equations fail due to the same pathology, shown in Fig. 3.4, that ails the Hartree-Fock solutions (and which make their solution “quite difficult” [69]), exacerbated by the long-wavelength behavior of the spectrum; this seemed to confirm that phonons are destabilizing the system. At that point we made another interesting discovery: by finding solutions to the Goldman-SilveraLeggett model [96] we proved that the semiclassical Hartree-Fock equations do have a solution in precisely the same limit where the HFB scheme was known to explicitly fail. These results, along with their finite-number counterparts and a report on our inability to do the same with HFB, were summarized in Ref. 117. (Months later we found that many of these had already been discovered independently [116].) Shortly thereafter we learned that, in contradiction to the results of Ref. 43, it was indeed possible to solve the uncondensed equations at any (numerically available) temperature [42]. We reproduced those results using two different methods and became convinced of their validity; though that did invalidate one of the questions that we had considered at the beginning, it of course did not rule out the possibility of a transition occurring in the 2D trapped system. Moreover, we evaluated the free energy for both solutions and found that the condensed solutions would be preferred over the uncondensed ones, and that the latter also did not reduce to the correct Thomas-Fermi limit at zero temperature; this contradicted the fact that the system has a condensate at absolute zero [138] and made these solutions either unphysical or metastable. These considerations led to another question: why does the semiclassical Hartree-Fock approach succeed where other, more sophisticated schemes fail? It turns out that, by eliminating the quasihole excitations from the HFB description of the noncondensate, the Hartree-Fock approximation 110 effectively imposes an infrared cutoff in the Bogoliubov spectrum. This cutoff is consistent with others imposed by interaction renormalization in the infinite homogeneous 2D system [40,119,120] (a similar study for the trapped system [121] has also appeared); since these emerge from more elaborate theories that take into account the Kosterlitz-Thouless transition and that allow for the possibility of quasicondensates, we became confident that our Hartree-Fock equations can be a reasonable approximation to what is happening within the finite 2D trapped Bose gas. (The full-blown HFB equations, successfully implemented in 3D a few years back [91] and in 2D very recently, have by nature a discrete, quantum-mechanical energy spectrum and thus also feature a lower limit of quasiparticle energy; the paper where their solution is reported [93], however, does not give an idea of the magnitude of this “cutoff.”) By the time we obtained those results [104], the first realization of lower-dimensional condensates was reported [30]. Those were, naturally, created by squeezing a 3D condensate until it reached the crossover régime, and that, plus the fact that 2D condensates had already been predicted for finite systems by Monte Carlo simulations, led us to the study of Krauth’s PIMC algorithm, which works for traps of any symmetry. As part of that study, we produced a working version of the program in Matlab, the programming language that we have used almost exclusively since we started our research; unfortunately, that semester-long endeavor failed: though each of the subroutines we wrote could compete with (and sometimes beat) Krauth’s in efficiency, the main loop was that: a loop. Matlab is designed for vectorized computation, and loops in Matlab programs are to be avoided at any cost (Ref. 139 explains some of the reasons why this is so); and that, we learned the hard way, is absolutely impossible in a path-integral Monte Carlo simulation. 2 This, though we have not mentioned it before, is one of the important conclusions of this thesis. We eventually took up Krauth’s Fortran code and, after undoing some approximations that Krauth himself had made and that became inaccurate only at high anisotropies, used it to obtain the results exhibited in Chapter 5; as we have seen, our agreement with published results (both mean-field and Monte Carlo) for isotropic traps was quite good. This led in turn to two separate research tracks. On one hand, it would be of interest to extract the off-diagonal one-body density matrix from the simulation results [129,136] and, by diagonalizing it, obtain the populations of a few excited states; this would allow us to explore the possibility of a smeared condensation in the 2D gas. On the other hand, it would be desirable to have a semiclassical Hartree-Fock program that, unlike the spline-minimization routine, would work for an anisotropic trap. In the end, we did 2 Other simulation methods (diffusion Monte Carlo, for example) do allow vectorization [131,133]. 111 not find a completely satisfactory answer to either of these challenges, but the partial answers we obtained account for a good part of the results we have presented here. In particular, our attempts to calculate the off-diagonal matrix for the ideal gas while avoiding the enormous matrices that Simpson’s rule requires made us look into spectral methods again. When the ideal-gas results turned out to be satisfactory, we tried to apply the method—now combined with the minimization of Eq. (3.63)—to the Hartree-Fock procedure in the anisotropic case. We were partially successful: the programs gave reasonable results for moderate degrees of anisotropy but became very inaccurate for compression ratios still within the 3D régime. One day we decided to check if the inaccuracy was due to the minimization routine, and that is how we found a method that allowed us to redo two years of research in three hours (and also gave us the chance to explore the different quasi-2D coupling constants mentioned in Section 3.5 and to find that they had a negligible influence in the systems we studied). We then implemented the Hartree-Fock off-diagonal density matrix [73,89] using this method; the results appear in Chapters 3 and 5. We conclude in this thesis that, in two-dimensional isotropic finite trapped systems, and insofar as such (admittedly small) systems can be described to any degree of approximation by the semiclassical Hartree-Fock equations, there is a phenomenon resembling a condensation into a single state. These equations, moreover, provide in at least a few cases a reasonable description of a 3D system anisotropic enough as to have entered the quasi-2D régime. There are, of course, a few qualifications to this statement and more than a few questions left wide open. In the following we enumerate a few of both. Monte Carlo simulations are closer to being exact than mean-field solutions, and surely include more physics than we have explored at this point, as long as we can extract it. In particular, and as we mentioned above, we are aware that it is possible to obtain the off-diagonal density matrix directly from path-integral Monte Carlo [45,129], which can then be diagonalized to yield excitedstate occupations; this has been done in 3D [136], and the results were in complete agreement with the predictions of the Gross-Pitaevskiı̆ equation. Only very recently, after getting to know the Monte Carlo method well enough, have we begun to understand how this can be done. Now, it has to be kept in mind that Monte Carlo simulations are not a panacea. While they excel at high temperatures3 (precisely where mean-field theory breaks down) and would thus allow the study of the transition region, PIMC simulations work only for fairly moderate particle numbers, and the finite-size-scaling ideas that have proved successful in the study of the homogeneous Bose 3 This is not to say that they fail at low temperatures. 112 gas [128] cannot be readily applied in a trap. Hence we cannot really use this method to approach the thermodynamic limit, where the consequences of Hohenberg’s theorem would be apparent. We also ran a few 2D interacting PIMC simulations, based on the code discussed in detail on Ref. 33 and whose results were published in Ref. 32; this program uses Krauth’s algorithm but implements hard-disk interactions.4 This program does predict a condensation in two dimensions, and gives reasonable estimates for the superfluid fraction (see below), but predicts condensate fractions much smaller than those given by the Hartree-Fock method. (This is just in the interacting case; the ideal gas works exactly as expected, and in fact our figures 5.8 and 5.9 on page 98 were generated using this program.) We cannot really tell if there is something wrong with it, and we scarcely used it because we did not gain sufficient control over its inner workings, given that at that point we were more interested in compressed 3D systems. It might also be possible that there is indeed a fundamental difference between hard disks in two dimensions and the hard spheres in quasi-2D space that, as we saw, are reasonably mimicked by a pure two-dimensional Hartree-Fock approximation; that, if confirmed, would be an interesting result. Mean-field theory, with all its limitations, should still be considered the method of choice at high atom numbers and low temperatures. One logical next step would be to work harder on obtaining solutions to the exact, discrete HFB equations, though we are aware that, just like Monte Carlo simulations, these programs should become too cumbersome for large systems: the 3D results reported in [70] correspond to N = 2 × 104 , and we suspect that these are not likely to be superseded in the near future. (The recent exact solutions have N = 2000.) Even if we restrict ourselves to the Hartree-Fock approximation, we still need a consistent meanfield treatment of the anisotropic gas. Our current methods do not provide us with enough accuracy to determine, for example, aspect ratios. As we already said, the method we describe at the end of Chapter 3, which proved so successful for isotropic traps, was in fact an offshoot of a program we wrote to deal with the anisotropic case; we only need point out that the necessary matrices (even assuming radial symmetry) involve 900 × 900 terms as opposed to the 30 × 30 that suffice for the treatment of the isotropic gas, with consequent losses of both speed and accuracy. On the other hand, this problem is not insurmountable, and might soon be within our reach. 4 The two-body density matrix for that program results from a derivation almost identical to the one we present in Appendix D, and involves cylindrical Bessel functions instead of the spherical Bessel functions obtained there; however, the actual implementation of the interactions is more difficult, since the resulting two-body density matrix cannot be efficiently calculated using GaussHermite quadrature and it becomes necessary to calculate it to very high precision and store the results in a table. See Ref. 33 for details. 113 This desirable mean-field program should also be able to find higher-order eigenstates of the GP equation, since the best methods we developed gave only the ground state. We do have the higher eigenstates as given by the off-diagonal density matrix, but we still need to find them directly; and at this point, the only method that was giving us excited states was the very inefficient initial-value method from Chapter 3. Access to these excited states will constitute a royal road to comparison with experiment. Another open direction would be the study of vortices, superfluidity, and the Kosterlitz-Thouless transition that we have so far avoided. There exist ways of calculating the superfluid fraction, both in the Hartree-Fock approximation [88], where it is treated as a kinematic effect and is extracted by comparing the moment of inertia of the gas in a rotating trap to its classical counterpart [140], in path-integral Monte Carlo simulations, where it is related to the surface area swept by the paths depicted in Fig. 5.1 when looked at from above [45,141], and through variational calculations [99]. Now, explicit 2D Monte Carlo simulations for 4 He found no significant temperature dependence of the vorticity correlation function, and attributed this negative result to low data resolution [142]. For the two-dimensional trapped gas, simulations resembling ours [32,33] found no evidence of a jump in the superfluid density close to the point where the transition should occur. A more in-depth study of these matters would lead us to better understand the Hartree-Fock cutoff that we have mentioned so often in this chapter and that has been so crucial for our arguments. It is possible that what we have found in this research is evidence for a quasicondensate. More insight into this topic could be gained by calculating the coherence lengths of our systems, a quantity already (but not too readily) available from the off-diagonal density matrix [61], and studying the behavior of phase fluctuations, which has been (though not necessarily by that name) behind many of our considerations in the previous chapters. Finally, another area where analytic results could join forces with Monte Carlo simulation is the interpretation of experimental data. It should be within our reach someday to create a purely mathematical rendering, including the effects of finite size, finite temperature, interactions, and anisotropy, of Bose-Einstein condensates not unlike that first sodium one that exploded in a screen at MIT one early morning. 114 APPENDIX A MATHEMATICAL VADEMECUM The Bose-Einstein integrals The Bose-Einstein integrals are defined by gσ (e −x 1 )= Γ(σ) Z ∞ 0 ∞ X e−kx tσ−1 dt = , e(t+x) − 1 k=1 k σ (A.1) where the second expression results from expanding the denominator as a geometric series and integrating term by term. For σ = 1, either the integral or the series can be immediately evaluated to yield g1 (e−x ) = − log(1 − e−x ). For x = 0 and σ 6= 1 the integrals become ζ(σ), the Riemann zeta function; g1 (e−x ) diverges at that point. For small nonzero values of x, the series in (A.1) converges very slowly, making it a very inefficient procedure to calculate the integrals; the accumulation of roundoff errors, moreover, makes it inaccurate as well. Instead we turn to an expression due to J. E. Robinson [143]: gσ (e−x ) = Γ(1 − σ) xσ−1 + ∞ X (−1)k k=0 k! ζ(σ − k) xk . (A.2) When σ is not an integer, this expression can be used directly to calculate the integral, though the values of the zeta function for negative arguments have to be found by analytic continuation. Most of the time, however, we will be using integer values of σ, for which the first term and the term in the sum corresponding to k = σ − 1 diverge. But these divergences turn out to cancel each other, resulting in a finite expression, lim ((−1)σ−1 Γ(σ)Γ(1 − k)xk−σ + ζ(k − σ + 1)) = k→σ σ−1 X `=1 1 − log x, ` (A.3) which can be obtained using the properties of the polygamma function [62]. Finally, we can use the following results about the Riemann zeta function [144], 1 ζ(0) = − , 2 ζ(−2m) = 0, 115 ζ(1 − 2m) = − B2m , 2m (A.4) where Bp is the pth Bernoulli number [62,144], to find gσ (e −x σ−2 X (−1)σ−1 (−1)n ζ(σ − n)xn + )= n! (σ − 1)! n=0 Ãσ−1 X1 `=1 ` ! − log x xσ−1 − ∞ X B2k x2k+1 (−1)σ σ x + . 2σ! 2k (2k + 1)! k=σ−1 (A.5) The function g2 , for example, can be approximated by g2 (e−x ) = π2 x2 x3 x5 x7 x9 − (1 − log x)x − + + + + + ··· 6 4 72 14400 1270080 87091200 (A.6) The harmonic-oscillator eigenfunctions The eigenfunctions of the harmonic-oscillator Hamiltonian are well known and can be found in any textbook on quantum mechanics: hx | ni = p 1 x2 /2 Hn (x) √ e 2p p! π (A.7) In three dimensions, we can simply multiply three one-dimensional functions together or we can solve the Schrödinger equation in spherical coördinates, in which case the angular dependence is carried by spherical harmonics: hx | n, l, mi = s 2 2n! l+ 1 e−x /2 xl Ln 2 (x2 )Ylm (ϑ, ϕ) 1 (n + l + 2 )! (A.8) To study the effect of anisotropy, we can write them down in cylindrical coördinates [145,146] 1 hx | n, m, pi = π s n! (n + m)! µ ¶1/4 √ 2 2 1 λ 1 p e− 2 (ξ +λη ) ξ m Lm (ξ 2 )eimϕ Hp (η λ); n π 2p p! (A.9) where we have used the scaling introduced in Section 2.5. In all of the above equations, the L m n are associated Laguerre polynomials and the Hp are Hermite polynomials. In a strictly two-dimensional isotropic trap the eigenfunctions are particular cases of (A.9), 1 hx | n, mi = √ π s 2 n! 2 imϕ e−x /2 xm Lm , n (x )e (n + m)! (A.10) and their corresponding eigenvalues are ²nm = 2(2n + |m| + 1). The first few radially symmetric functions (that is, hx | n, mi with m = 0) are shown in Fig. 2.10, and their associated eigenvalues 116 appear in Fig. 2.9. We will presently use these to derive an expression for the one-body density matrix of a distinguishable system. From Eq. (A.10) we have ñ(x, x0 ) = hx|ρ|x0 i = = ∞ ∞ X X n=0 m=−∞ hx|e−β̃ H̃1 |n, m, pi hn, m, p | x0 i ∞ 1 −2β̃ −(x2 +x02 )/2 X −2β̃|m| 0 m im(ϕ−ϕ0 ) e e e (x x) e π m=−∞ × ∞ X n! 2 m 02 e−4β̃n Lm n (x )Ln (x ). (n + m)! n=0 (A.11) The last sum can be evaluated by means of the identity [144] à p ! µ ¶ xyz (xyz)−α/2 x+y n! z n α α L (x)Ln (y) = exp −z Iα 2 , Γ(n + α + 1) n 1−z 1−z 1−z n=0 ∞ X (A.12) where Iα (x) is a modified Bessel function. We can now use the generating function for the modified Bessel functions, ∞ X Im (x) tm = e(x/2)(t+1/t) , (A.13) m=−∞ and the fact that x0 x cos(ϕ − ϕ0 ) = x · x0 to obtain ρ(x, x0 ) = 2 02 0 1 1 csch 2β̃ e− 2 csch 2β̃((x +x ) cosh 2β̃−2x·x ) , 2π (A.14) where we have recovered Eq. (2.4). Equation (A.13) can also be used to find an alternative definition of the modified Bessel function of order zero: I0 (x) = 1 2π Z 2π dϕ ex cos ϕ , 0 which supports our identification of the the averaged density matrix in (2.28). 117 (A.15) APPENDIX B PERMUTATION CYCLES AND WICK’S THEOREM Permutation cycles in the ideal gas Quantum particles are indistinguishable, and the wavefunction of a system of particles should be independent of the bookkeeping scheme that we adopt to label them. In the particular case of bosons, the wavefunction must remain unchanged when the particles (or, rather, the indices employed to pinpoint them) are permuted. Hence the only wavefunction that we can use to describe the system is the totally symmetric combination of the wavefunctions that would describe a system of distinguishable particles. For example, if we have three bosons, labelled 1 through 3, we must use the normalized combination 1 S|123i = √ (|1i|2i|3i + |2i|1i|3i + |1i|3i|2i + |2i|3i|1i + |3i|1i|2i + |3i|2i|1i), 3! (B.1) where the symmetrization operator S is Hermitian and obeys S 2 = S. This result can be generalized to the case of N particles: where P 1 X |P 1i · · · |P N i, S|1 . . . N i = √ N! P P (B.2) is the sum over all possible permutations of the numbers {1, . . . , N } and P i is the effect of one such permutation on the index i. The fundamental quantity that describes an ensemble of particles is the N -body density matrix, PN defined by ρN = e−β̃ H̃ /ZN , where H̃ is the N -body Hamiltonian, H ≡ i=1 H1,i ; the partition function is the trace ZN = Tr e−β̃ H̃ that normalizes the density matrix and is given by [15,147] 1:N ZN = Tr ρN = 1:N Z dσx1 · · · dσxN hx1 · · · xN |Se−β̃ H̃ S|x1 · · · xN i, (B.3) where, as usual, we evaluate the trace in coördinate space. Since the bosons cannot turn into fermions, the symmetrization operator is a constant of the motion and commutes with the Hamiltonian; moreover, since S 2 = S, only the bra or the ket need be symmetrized. The consequences of 118 this symmetrization are most easily exhibited by considering systems with small particle numbers. For N = 1, the partition function is the well-known result [14] Z1 = (2 sinh β̃)−σ . For N = 2, we have Z2 = 1 2! Z ¡ ¢ dσx1 dσx2 hx1 |e−β̃ H̃0 (1) |x1 ihx2 |e−β̃ H̃0 (2) |x2 i + hx2 |e−β̃ H̃0 (1) |x1 ihx1 |e−β̃ H̃0 (2) |x2 i . (B.4) The first integral is just a product of two one-body partition functions and gives Z 12 (β̃). To calculate the second one, we can insert two complete sets of oscillator eigenfunctions |ni, |n 0 i: Z Z dσx1 hx2 |e−β̃ H̃0 (1) |x1 i dσx2 hx1 |e−β̃ H̃0 (2) |x2 i Z XXZ σ −β̃ H̃0 (1) = d x1 hx2 |e |nihn | x1 i dσx2 hx1 |e−β̃ H̃0 (2) |n0 ihn0 | x2 i n = X n0 e−β̃²n = e −β̃²n = e−β̃²n = X n0 n X X e n0 n X e−β̃²n0 n0 n X X e Z dσx1 hx2 | nihn | x1 i σ 0 Z d x1 hn | x1 ihx1 | n i e−β̃²n0 hn | n0 ihn0 | ni = −β̃²n −β̃²n e −β̃²n0 Z X n Z dσx2 hx1 | n0 ihn0 | x2 i dσx2 hn0 | x2 ihx2 | ni e−β̃²n X e−β̃²n0 δn0 n n0 = Z1 (2β̃), (B.5) n where we have used the orthonormality of the oscillator eigenfunctions and the fact that the position eigenkets also form a complete set. This result is the basis for everything that follows. The important point here is that the permutation {1, 2} → {2, 1} is a closed cycle, a snake biting its own tail; that fact allowed us to move the brackets around and exploit the second completeness relation. Had we had three particles, we could have calculated the term h231|ρ|123i by inserting three sets of eigenfunctions and using the fact that δnn0 δn0 n00 δn00 n = δn0 n δnn00 ; the result would be Z1 (3β̃). If we have k particles, any cyclic permutation of the indices will give us a term Z 1 (k β̃). Now, every permutation of the N particles can be decomposed into cycles. Going back to three particles, we can see that {1, 2, 3} → {1, 3, 2} can be decomposed into (1)(23): 1 has been left alone (in its own 1-cycle) while 2 and 3 are now in a 2-cycle (otherwise known as an “exchange”). Another permutation, {1, 2, 3} → {3, 2, 1}, for example, can be decomposed into (1, 3)(2). In general, any given permutation of N particles will be broken up into c1 1-cycles, c2 2-cycles, and so on, up to cN N -cycles. Any combination of these c` that obeys N X `c` = N `=1 119 (B.6) is an acceptable permutation. At one end of the spectrum we will have the permutation consisting of N 1-cycles—the identity permutation—and at the other end we will have N − 1 cyclic permutations, each of them consisting of one N -cycle. It is easy to see why we have N − 1 of these, since we can put any number other than 1 at the start of the cycle (1 would, of course, give us the identity). Going back to the permutation {1, 2, 3} → {3, 2, 1}, we see that it can be put either as (1, 3)(2), (3, 1)(2), (2)(1, 3), or (2)(3, 1). In general, there are N! c1 !c2 ! . . . cN ! (B.7) ways of placing N particles in c1 1-boxes, c2 2-boxes, etc. Moreover, within each `-box we can put any one of the ` particles in the first slot, which gives us an extra factor of Π ` `c` in the denominator. Knowing this, we can rewrite the partition function (B.3) as ZN = 1 N! X c1 ,...,cN Y ` N! Z1 (`β̃)c` , c ` ! ` c` (B.8) subject to the restriction (B.6); for example, Eq. (B.4) really is Z2 (β̃) = 1 2! µ 2! 2! Z 2 (β̃) + Z 1 (2β̃) 2 1 1 1 2! 1 1! 2 | | {z } {z } {1,2}→{1,2}=(1)(2) c1 =2, c2 =0 ¶ . (B.9) {1,2}→{2,1}=(1,2) c1 =0, c2 =1 Finally, the constraint (B.6) can also be lifted by summing over all possible values of N —in other words, by invoking the grand canonical ensemble. Using the grand canonical partition function Z= ∞ X eβ̃ µ̃N ZN (B.10) N =1 we obtain Z= Y ` exp à Z1 (`β̃)e`β̃ µ̃ ` ! . (B.11) We can immediately apply the definition ∞ X ∂ eβ̃ µ̃ Z1 (`β̃), t log Z = N= ∂ µ̃ (B.12) `=1 along with the expression quoted above for the one-particle partition function, to recover Eq. (2.11). 120 Wick’s theorem Our treatment of interacting systems on Chapter 3 was based on the formalism of second quantization, in which the field operators are expanded in creation and annihilation operators, and often we had to calculate thermal averages of combinations of these. What follows is a brief discussion of a simultaneous simplification and generalization of Wick’s theorem [84,148], which expresses these combinations in terms of pairs of operators (also called “contractions”). It is a simplification because, just like everywhere else in this work, we will assume that the operators are independent of time; one has only to look at Wick’s original proof [148], with its painstaking attention to the time ordering of the operators, to see how much of a simplification this is. On the other hand, it is a generalization because it describes systems at finite temperature, a case not treated by Wick (this version of the theorem was in fact first proved by Matsubara [149]). Finally, we will concentrate on Bose statistics, even though the theorem (in a slightly modified form) applies also to fermions. We will illustrate the theorem rederiving some results obtained above for the ideal gas. The very fact that we can use second quantization to treat a many-body problem is based on the assumption that the many-body state vectors |n1 · · · nj · · ·i, where nj is the number of atoms in the jth state, form a complete orthonormal set. When we impose Bose statistics on the system, we know that a temperature t = 1/β̃ the populations must obey the Bose-Einstein distribution, nj = 1 eβ̃²j −1 ≡ fj , (B.13) and for that reason we must have hb†k bk i = hn1 · · · nk · · · | b†k bk |n1 · · · nk · · ·i = hn1 · · · nk − 1 · · · | √ √ nk nk |n1 · · · nk − 1 · · ·i = nk hn1 · · · nk − 1 · · · | n1 · · · nk − 1 · · ·i = nk = fk . (B.14) where we used the fundamental definition of the creation and annihilation operators [14,46,54]. Similarly, hb†j bk i = h· · · nj · · · nk · · · | b†k bk | · · · nj · · · nk · · ·i p p = h· · · nj − 1 · · · nk · · · | nj nk | · · · nj · · · nk − 1 · · ·i 121 = p nj nk h· · · nj − 1 · · · nk · · · | · · · nj · · · nk − 1 · · ·i = p nj nk δjk = nj δjk . (B.15) This procedure already tells us some of the properties that must be had by the combinations we want to average. First of all, all averages involving a single creation or annihilation operator will vanish, and so will all combinations with an odd number of operators; more precisely, the operator must have equal numbers of creation and annihilation operators. With this result we can derive some properties of the ideal gas, for which the field operator simply becomes Ψ(x) = X Ψ† (x) = bj φj (x), j X b†j φ∗j (x), (B.16) j where the φj (x) are the harmonic-oscillator eigenfunctions from Appendix A. The total atom number, for example, is readily found to be N= Z σ † 0 d x Ψ (x)Ψ(x ) = X b†k bj jk Z dσx φ∗j (x)φk (x0 ) = X b†k bj δjk = X b†j bj (B.17) j jk and thus hN i = X j hb†j bj i = X fj , (B.18) j which is also consistent with defining the one-body density as ñ(x) = hΨ † (x)Ψ(x)i. The off-diagonal one-body density matrix can also be found this way: ñ(x, x0 ) = hΨ† (x)Ψ(x0 )i = X X † fj φ∗j (x)φj (x0 ), hbj bk i φ∗j (x)φk (x0 ) = (B.19) j jk as we saw in Eq. (2.25). For interacting systems we must consider the next possible combination, which involves four operators: hb†j b†k bl bm i = h· · · nj · · · nk · · · nl · · · nm · · · | b†j b†k bl bm | · · · nj · · · nk · · · nl · · · nm · · ·i (B.20) p = nj nm h· · · nj − 1 · · · nk · · · nl · · · nm · · · | b†k bl | · · · nj · · · nk · · · nl · · · nm − 1 · · ·i p nj nk nl nm h· · · nj − 1 · · · nk − 1 · · · nl · · · nm · · · | · · · nj · · · nk · · · nl − 1 · · · nm − 1 · · ·i; = the last bracket will vanish unless either j = l and k = m or j = m and k = l. This immediately leads to hb†j b†k bl bm i = fj fk (δjl δkm + δjm δkl ), a result used in Eq. (4.29). 122 (B.21) APPENDIX C SPECTRAL DIFFERENTIATION AND GAUSSIAN QUADRATURE Interpolating polynomials Consider a function f whose values we know only at a set of n points x (n) j , 1 ≤ j ≤ n, assumed to be distinct. If the function is smooth and well-behaved, it should be amenable to approximation by means of the interpolant f (x) ≈ n X (n) f (x(n) j ) `j (x), (C.1) j=1 where the dependence on x is carried by the “cardinal function” [105] `(n) j (x), a polynomial of degree (n) (n) (n − 1) that obeys `(n) j (xi ) = δij . By construction, the interpolant is exact at the xj . We can write the cardinal function in terms of the Lagrange interpolating polynomial: `(n) j (x) = (n) (n) (n) (x − x(n) 1 ) · · · (x − xj−1 )(x − xj+1 ) · · · (x − xn ) (n) (n) (n) (n) (n) (n) (n) (x(n) j − x1 ) · · · (xj − xj−1 )(xj − xj+1 ) · · · (xj − xn ) = n Y x − x(n) k , (n) (n) x − x k j=1 j (C.2) j6=k which is easily seen to vanish at every x(n) i , i 6= j, since for every i we can find a vanishing monomial (n) (n) (x(n) i − xi ), and which equals unity for x = xj . Now, given a family of orthogonal polynomials u0 (x), u1 (x), . . . , un (x) such that Z I uj (x)uk (x) w(x) dx = kuk k2w δjk , (C.3) where we have introduced an interval I, a weight function w(x), and the norm ku k kw , we can find (n) another expression for the cardinal function. If we let x(n) 1 , . . . , xn be the roots of un (x), we can write the polynomial as (n) (n) un (x) = (x − x1 ) · · · (x − xn ) = n Y (n) (n) (x − xj ) = (x − xk ) j=1 123 n Y j=1 j6=k (x − xj(n) ); (C.4) its derivative will then be (n) (n) (n) (n) u0n (x) = (x − x(n) 2 ) · · · (x − xn ) + (x − x1 )(x − x3 ) · · · (x − xn ) + · · · = and, in particular, u0n (x(n) k )= n Y j=1 j6=k n n Y X j=1 j=1 j6=k (x − x(n) j ) (C.5) (n) (x(n) k − xj ). (C.6) If we look at the last expression for un (x) in (C.4), we can see that n Y x − x(n) un (x) j (n) ) = (x − x (n) (n) k ) − x u0n (x(n) x j k j=1 k (C.7) j6=k and hence `(n) k (x) = n Y x − x(n) un (x) k ; = 0 (n) (n) (n) (n) x − x u (x n k )(x − xk ) k j=1 j (C.8) j6=k (n) by means of L’Hôpital’s rule we can verify that `(n) k (xk ) = 1. Calculating orthogonal polynomials The polynomials themselves and all of their derivatives can be calculated via Eq. (C.4) at any point if we know their roots. These, in turn, can be calculated by exploiting the fact that, if a family of orthogonal polynomials obeys a three-term recurrence relation of the form un+1 (x) = (x − αn )un (x) − βn2 un−1 (x), (C.9) with u−1 = 0 and u1 (x) = 1, then it can be proved by induction that the roots of un (x) are the eigenvalues of the Jacobi matrix [63,64] α0 β1 J= β1 α1 β2 .. . βn−1 βn−1 αn−1 . For the particular case of the Laguerre polynomials, we have αn = 2n + 1 and βn = n2 . 124 (C.10) Differentiation matrices To calculate the derivative of f we do f 0 (x) = n X f (x(n) j ) j=1 d (n) ` (x), dx j (C.11) which, if we introduce the derivative matrix D(1) , whose elements are defined by [64,65,105] (1) dij ≡ can be written as f 0 (x(n) i )= d (n) ¯¯ ` (n) , dx j xi n X (C.12) (1) dij f (x(n) j ). (C.13) j=1 Straightforward differentiation of `(n) j (x) yields u0n (x)(x − x(n) d (n) j ) − un (x) , `j (x) = (n) 0 2 dx un (xj )(x − x(n) j ) (C.14) and the off-diagonal elements of the matrix can be found immediately: (1) dij = 1 u0n (x(n) i ) (n) (n) 0 un (xj ) xi − xj(n) for i 6= j. (C.15) To calculate the diagonal elements, we invoke L’Hôpital’s rule once again in order to avoid the indeterminacy: (1) djj = = lim (n) x→xj u0n (x)(x − x(n) j ) − un (x) 2 u0n (xj(n) )(x − x(n) j ) 0 0 u00n (x(n) u00n (x)(x − x(n) 1 j ) j ) + un (x) − un (x) = lim . (n) (n) 0 0 (n) un (xj ) x→xj 2(x − xj ) 2un (xj(n) ) (C.16) (n) T Hence, if we express f as a column vector, f = [f (x(n) 1 ), . . . , f (xn )] , we obtain f 0 = D(1) f . (C.17) It must be noted that, unlike those encountered in the finite-difference method, these differentiation matrices are not sparse. However, it can be proved [105] that, for smooth functions, the 125 accuracy attained grows much faster with n than with finite differences. In some cases, the error falls as e−n , while the usual finite-difference algorithm converges at a rate n−2 . This treatment can be extended to include higher-order derivatives [65] and many different boundary conditions [64]. We will not delve into those here because the functions we interpolate only require first derivatives, but we note that in some cases, and in particular in the case of Laguerre polynomials, the differentiation matrices obey D(k) = Dk . (C.18) Gaussian quadrature Given the interval and weight function introduced in Eq. (C.3), we can integrate our function f by a similar process [62,65]: Z f (x) w(x) dx = I n X (n) f (xj ) j=1 Z (n) I `j (x) w(x) dx ≡ n X (n) f (x(n) j ) wj , (C.19) j=1 (n) where we have introduced the weights wj(n) . To find the weights we can use the fact that `(n) j (xi ) = δij and write n X (n) (n) 1 (n) ` (xi )un−1 (x(n) (n) i )wi un−1 (xj ) i=1 j Z 1 `(n) = j (x) un−1 (x) w(x) dx un−1 (x(n) j ) I Z 1 un (x) = un−1 (x) w(x) dx. ) (un−1 u0n )(x(n) x − x(n) I j j wj(n) = (C.20) Now, we can determine the leading coefficient of each polynomial in the family. In other words, if un (x) = an xn + an−1 xn−1 + · · · and un−1 (x) = bn−1 xn−1 + bn−2 xn−2 + · · · (C.21) we can find the numbers an and bn−1 ; furthermore, we find the first term of the quotient un (x)/(x − x(n) j ) by long division and obtain un (x) an (n) un−1 + qj (x) ≡ Aj un−1 (x) + qj (x) = an xn−1 + · · · ≡ b x − x(n) n−1 j 126 (C.22) and, since qj is a polynomial of degree (n − 2) at most, it can be expanded in terms of u 1 , . . . , un−2 . But all of the resulting terms are orthogonal to un−1 , and thus we obtain (n) wj(n) = Aj (un−1 u0n )(xj(n) ) Z (n) un−1 (x) un−1 (x) w(x) dx = I Aj (un−1 u0n )(xj(n) ) kun−1 k2w . (C.23) Most of the time we will be concerned with the Laguerre polynomials, for which I = [0, ∞) and w(x) = e−x . It can be proved that [62] Ln (x) = ¶ n µ X (−1)k k n x , n−k k! (C.24) k=0 whence Aj = −(n − 1)!/n! = −1/n, and Z ∞ (Ln (x))2 e−x dx = 1. (C.25) 0 Combining these results we get 1 1 (n) 0 −1 wj(n) = − (Ln−1 (x(n) = (n) (L0n (xj(n) ))−2 , j )Ln (xj )) n xj where we used the recurrence relation xL0n = n(Ln − Ln−1 ) to arrive at the last expression. 127 (C.26) APPENDIX D MATHEMATICAL DETAILS OF PIMC SIMULATIONS The two-body density matrix The ratios in Eq. (5.20) can be calculated quite approximately because in these two-body collisions the interaction potential depends only on the relative distance. Let us consider the ratio corresponding to any two particles, which we will call 1 and 2. If we introduce the center-of-mass and relative coördinates and momenta X≡ 1 (x1 + x2 ), 2 x ≡ x 1 − x2 , K ≡ κ 1 + κ2 , κ≡ 1 (κ1 − κ2 ), 2 (D.1) we can rewrite the two-body Hamiltonian as H̃ = κ21 + κ22 + x21 + x22 + v(|x1 − x2 |) ¡1 2 ¢ ¡ ¢ 1 1 = K + 2X 2 + 2κ2 + v(x) + x2 ≡ H + h + x2 . 2 2 2 (D.2) Since the inverse temperature τ is small, we can use the “Trotter breakup” [44,45,89] (that is, we can neglect the commutators between momentum and position operators, as we did when we derived Eq. (3.62)) to turn the numerator of Eq. (5.20) into ρ2 (x1 , x2 , x02 , x02 ; τ ) = ρ2 (X, x, X0 , x0 ; τ ) = hX, x|e−τ H̃ |X0 , x0 i 1 2 1 2 ≈ hX|e−τ H |X0 ihx|e− 4 τ x e−τ h e− 4 τ x |x0 i = Ξ(x) hX|e−τ H |X0 ihx|e−τ h |x0 i Ξ(x0 ), 1 2 1 where Ξ(x) ≡ e− 4 τ x = e− 4 τ (ξ 2 +λ2 η 2 ) 1 (D.3) 2 . (Note that we separated e− 2 τ x into two terms in order to include contributions from both x and x0 ; this we also did when deriving (3.62).) The denominator 128 of Eq. (5.20) can also be calculated by defining center-of-mass and reduced coördinates [33]. The center-of-mass terms cancel out, and we are left with ρHC,1/2 (x, x0 ; τ ) ρ2 (x1 , x2 , x01 , x02 ; τ ) = Ξ(x) Ξ(x0 ), ρ1 (x1 , x01 ; τ )ρ1 (x2 , x02 ; τ ) ρHO,1/2 (x, x0 ; τ ) (D.4) where we have used the subscripts HC and HO to refer to the hard-core and harmonic-oscillator density matrices respectively; the 1/2 emphasizes that these are the reduced-mass matrices; in particular, the ideal-gas element in the denominator will have the form (2.31) with the 1 4 replaced by 81 . The hard-core density matrix for the relative coördinate can be found by assuming that the Hamiltonian has eigenfunctions of the form ψkl (x)Ylm (ϑ, ϕ), where the angular dependence is carried by spherical harmonics. The principal quantum number k is continuous, since the hard-sphere potential has no bound states, and thus we can try a partial-wave expansion of the form [127,150] 0 ρHC (x, x ; τ ) = ∞ X 2l + 1 4π l=0 Pl (cos γ) Z ∞ 2 dk e−2τ k ψkl (x)ψkl (x0 ), (D.5) 0 where we used the addition theorem for spherical harmonics to express the angles in terms of the relative angle γ defined by cos γ = cos ϑ cos ϑ0 + sin ϑ sin ϑ0 cos(ϕ − ϕ0 ). The density matrix (D.5) has to obey the Bloch equation [15] (see Section 2.2) hρHC = − ∂ρHC , ∂τ (D.6) and this requirement forces the eigenfunctions to obey ˜ 2 ψkl + 1 v(x)ψkl = − −∇ 2 µ d2 2 d l(l + 1) + − dx2 x dx x2 ¶ 1 ψkl + v(x)ψkl = k 2 ψkl , 2 (D.7) where we used the eigenvalue equation for the spherical harmonics to divide out the angular dependence. The hard-sphere potential v(x) is infinite for x ≤ α, where α ≡ a0 /x0 , and zero otherwise, and this transforms (D.7) into a free-particle Schrödinger equation with the boundary condition ψkl (α) = 0 [151]. In three dimensions, the solution to this Schrödinger equation is a linear combination of spherical Bessel and Neumann functions [62], ψkl (x) = A(cos δl jl (kx) − sin δl nl (kx)), 129 (D.8) and the phase shifts δl can be found by incorporating the boundary condition on the wavefunction: tan δl = jl (kα) . nl (kα) (D.9) The density matrix also has to obey the “initial” condition [15] ∞ ρHC (x, x0 ; 0) = δ (3) (x − x0 ) = X 2l + 1 1 δ(x − x0 ) Pl (cos γ); 0 xx 4π (D.10) l=0 comparison with the closure equation for the spherical Bessel functions [62], Z yields A = k p ∞ dk k 0 2 r 2 jl (kx) π r 1 2 jl (kx0 ) = 0 δ(x − x0 ), π xx (D.11) 2/π for the normalization constant. The Gaussian factor in (D.5) suggests that the numerical integration be performed using Gauss-Hermite quadrature, a process similar to the Gauss-Laguerre quadrature we study in Appendix C. (In fact, in order to save time the program calculates the matrix in a uniform mesh and then interpolates this table of values in subsequent steps.) Generating new configurations: The Lévy construction In their original paper [132], Metropolis et al. generated new configurations by selecting a single sphere and giving it a random displacement within an optimal range. In simulations like ours, where we have many more atoms than there were in Ref. 132, and where moreover we have multiple time slices, the only way to execute a calculation in a reasonable period of time is by generating multiparticle, multislice configurations, in effect “sending a ripple through the system” instead of moving just one particle at a time [152,153]; Krauth’s algorithmm uses a threading procedure, introduced by Pollock and Ceperley [45,127] and based on previous work by P. Lévy [154], that in the absence of interactions generates independent configurations that always pass the Metropolis test. The integrand of (5.3), we recall, is I ≡ ρ(R, R2 ; τ )ρ(R2 , R3 ; τ ) · · · ρ(Rk , Rk+1 ; τ ) · · · ρ(R`−1 , R` ; τ )ρ(R` , R0 ; τ ), 130 (D.12) ¾ σ3 x - 0 .. . .. . .. . x3 x2 x hx3 i Figure D.1. The Lévy construction can be used to generate the new positions in a configuration move. Given a point x2 and the (fixed) end position x0 , the Lévy construction prescribes the midpoint hx3 i and the width σ3 of the Gaussian distribution obeyed by x3 in the case of an ideal trapped gas. Each of the intermediate points depicted is a permissible, but the one marked with the larger empty circle and the dashed line will be preferred. The point x2 has been previously constructed using x (also fixed) and x0 ; x4 will be found based on x3 and x0 , and so on until the chain is complete. and we can rewrite it as [127] " #" # 0 0 ρ(R, R ; τ )ρ(R , R ; β̃ − τ ) ρ(R , R ; τ )ρ(R , R ; β̃ − 2τ ) 2 2 2 3 3 I = ρ(R, R0 ; β̃) × ··· ρ(R, R0 ; β̃) ρ(R2 , R0 ; β̃ − τ ) # " # " ρ(R` , R0 ; τ )ρ(R0 , R0 ; β̃ − `τ ) ρ(Rk , Rk+1 ; τ )ρ(Rk+1 , R0 ; β̃ − kτ ) ··· × ··· ; (D.13) ρ(Rk , R0 ; β̃ − (k − 1)τ ) ρ(R` , R0 ; β̃ − (` − 1)τ ) The terms that we added cancel in pairs, and a glance at the composition property (5.1) in the form ρ(R, R0 ; β̃) = Z dNσR00 ρ(R, R00 ; β̃ − β̃ 0 )ρ(R00 , R0 ; β̃ 0 ) (D.14) shows that every bracketed term is normalized to 1. Let us look more closely at one of those terms, Γk+1 " # ρ(Rk , Rk+1 ; τ )ρ(Rk+1 , R0 ; β̃ − kτ ) ≡ , ρ(Rk , R0 ; β̃ − (k − 1)τ ) (D.15) say, and think of it as a prescription for locating Rk+1 (or, more precisely, as N σ prescriptions for locating each of the points in that time slice). Apart from the index k and the usual parameters β̃ and τ , the only dependence is on Rk (the time slice immediately before Rk+1 ) and the final time slice R0 ; in principle, then, we should be able to locate Rk+1 given those two time slices. In that sense, and given its unit normalization, we can interpret (D.15) as a conditional probability distribution. 131 For the case of the ideal trapped gas, it can be shown after a lot of algebra that Eq. (D.15), upon substitution of the density matrix (2.31), becomes Γk+1 à !2 0 R csch 2τ + R csch 2( β̃ − kτ ) 1 k (D.16) ∝ exp − (coth 2τ + coth 2(β̃ − kτ )) Rk+1 − 2 coth 2τ + coth 2(β̃ − kτ ) in the x- and y-directions; in the anisotropic case, for the z-direction we must substitute 2 → 2λ everywhere except in the exponent at the end. Thus, as would be expected, the distribution (D.15) for the ideal gas is Gaussian, and if we relocate the time slice to a random configuration obeying the normal distribution (D.16) we will have a new configuration that will always pass the Metropolis test. In other words, it is possible to generate a multiparticle, multislice move that will exactly obey the probability distribution of the ideal gas by making each xi in the kth slice be a Gaussian random number with standard deviation σk = q 1 (D.17) coth 2τ + coth 2(β̃ − (k − 1)τ ) centered around hxi,k i = xi,k−1 csch 2τ + x0i csch 2(β̃ − (k − 1)τ ) . coth 2τ + coth 2(β̃ − (k − 1)τ ) 132 (D.18) BIBLIOGRAPHY [1] E. A. Cornell and C. E. Wieman, Rev. Mod. Phys. 74, 875 (2002). [2] K. B. Davis et al., Phys. Rev. Lett. 75, 3969 (1995). [3] W. Ketterle, Rev. Mod. Phys. 74, 1131 (2002). [4] D. S. Hall, Am. J. Phys 71, 649 (2003). [5] C. Foot, Nature 375, 447 (1995). [6] M. H. Anderson et al., Science 269, 198 (1995). [7] C. C. Bradley et al., Phys. Rev. Lett. 75, 1687 (1995). [8] M. R. Andrews et al., Science 275, 637 (1997). [9] A. J. Leggett, in The New Physics, edited by P. Davies (Cambridge University Press, Cambridge, 1989). [10] A. Einstein, Sitzber. Kgl. Preuss. Akad. Wiss. 261 (1924). [11] A. Einstein, Sitzber. Kgl. Preuss. Akad. Wiss. 3 (1925). [12] S. N. Bose, Z. Phys. 26, 178 (1924). [13] A. J. Leggett, Rev. Mod. Phys. 73, 307 (2001). [14] R. K. Pathria, Statistical Mechanics (Pergamon, New York, 1972). [15] R. P. Feynman, Statistical Mechanics (W. A. Benjamin, Reading, MA, 1972). [16] A. Pais, “Subtle is the Lord. . . ”—The Science and the Life of Albert Einstein (Oxford University Press, Oxford, 1982). [17] J. P. Wolfe, J.-L. Lin, and D. W. Snoke, in Bose-Einstein Condensation, edited by A. Griffin, D. W. Snoke, and S. Stringari (Cambridge University Press, Cambridge, 1993). [18] J.-L. Lin and J. P. Wolfe, Phys. Rev. Lett. 71, 1222 (1993). [19] F. London, Nature 141, 643 (1938). [20] F. Pereira Dos Santos et al., Phys. Rev. Lett. 86, 3459 (2001). [21] C. E. Hecht, Physica 25, 1159 (1959). [22] H. F. Hess et al., Phys. Rev. Lett. 51, 483 (1983). [23] V. Bagnato, D. E. Pritchard, and D. Kleppner, Phys. Rev. A 35, 4354 (1987). [24] V. Bagnato and D. Kleppner, Phys. Rev. A 44, 7439 (1991). [25] D. G. Fried et al., Phys. Rev. Lett. 81, 3811 (1998). [26] D. Kleppner, Phys. Today 52, 11 (1999). 133 [27] P. C. Hohenberg, Phys. Rev. 158, 383 (1967). [28] W. J. Mullin, J. Low Temp. Phys. 106, 615 (1997). [29] A. I. Safonov, S. A. Vasilyev, I. S. Yasnikov, I. I. Lukashevich, and S. Jaakkola, Phys. Rev. Lett. 81, 4545 (1998). [30] A. Görlitz et al., Phys. Rev. Lett. 87, 130402 (2001). [31] S. Pearson, T. Pang, and C. Chen, Phys. Rev. A 58, 4811 (1998). See the comment following Ref. 125. [32] S. D. Heinrichs and W. J. Mullin, J. Low Temp. Phys. 113, 231 (1998); ibid. 114, 571 (1999). [33] S. D. Heinrichs, Two dimensional Bose systems in a harmonic potential, Master’s thesis, University of Massachusetts (1997). [34] J. M. Kosterlitz and D. J. Thouless, J. Phys. C 6, 1181 (1973). [35] S. I. Shevchenko, Sov. Phys. JETP 73, 1009 (1991). [36] M. Girardeau, Phys. of Fluids 5, 1468 (1962). [37] M. Van den Berg and J. T. Lewis, Commun. Math. Phys. 81, 475 (1981). [38] P. Nozières, in Bose-Einstein Condensation, edited by A. Griffin, D. W. Snoke, and S. Stringari (Cambridge University Press, Cambridge, 1993), p. 15. [39] Y. Kagan, B. V. Svistunov, and G. V. Shlyapnikov, Sov. Phys. JETP 66, 314 (1987). [40] Y. Kagan, V. A. Kashurnikov, A. V. Krasavin, N. V. Prokof’ev, and B. V. Svistunov, Phys. Rev. A 61, 043608 (2000). [41] D. V. Fil and S. I. Shevchenko, Phys. Rev. A 64, 013607 (2001). [42] R. K. Bhaduri, S. M. Reimann, S. Viefers, A. Ghose Choudhury, and M. K. Srivastava, J. Phys. B 33, 3895 (2000). [43] W. J. Mullin, J. Low Temp. Phys. 110, 167 (1998). [44] W. Krauth, Phys. Rev. Lett. 77, 3695 (1996). [45] D. M. Ceperley, Rev. Mod. Phys. 67, 279 (1995). [46] J. J. Sakurai, Modern Quantum Mechanics (Addison-Wesley, Reading, MA, 1994), revised ed. [47] R. Kubo, Statistical Mechanics (North-Holland, Amsterdam, 1971). [48] L. D. Landau and E. M. Lifshitz, Statistical Mechanics (Addison-Wesley, Reading, MA, 1969), 2nd ed. [49] O. Penrose and L. Onsager, Phys. Rev. 104, 576 (1956). [50] R. M. Ziff, G. E. Uhlenbeck, and M. Kac, Phys. Rep. 32C, 169 (1977). [51] W. Ketterle and N. J. van Druten, Phys. Rev. A 54, 656 (1996). [52] R. K. Pathria, Phys. Rev. A 58, 1490 (1998). [53] W. J. Mullin, Am. J. Phys 68, 120 (2000). [54] F. Dalfovo, S. Giorgini, L. P. Pitaevskiı̆, and S. Stringari, Rev. Mod. Phys. 71, 463 (1999). [55] S. Grossman and M. Holthaus, Phys. Lett. 208, 188 (1995). 134 [56] T. Haugset, H. Haugerud, and J. O. Andersen, Phys. Rev. A 55, 2922 (1997). [57] K. Damle, T. Senthil, S. N. Majumdar, and S. Sachdev, Europhys. Lett. 36, 7 (1996). [58] N. J. van Druten and W. Ketterle, Phys. Rev. Lett. 79, 549 (1997). [59] W. J. Mullin and J. P. Fernández, Am. J. Phys 71, 661 (2003). [60] M. Naraschewski and R. J. Glauber, Phys. Rev. A 59, 4595 (1999). [61] B. P. van Zyl, J. Low Temp. Phys. 132, 207 (2003). [62] G. Arfken, Mathematical Methods for Physicists (Academic Press, San Diego, 1985), 3rd ed. [63] W. H. Press, S. A. Teukolsky, W. A. Vetterling, and B. P. Flannery, Numerical Recipes in C (Cambridge University Press, Cambridge, 1992), 2nd ed. [64] J. A. C. Weideman and S. C. Reddy, ACM Trans. of Math. Software 26, 465 (2000). [65] D. Funaro, Polynomial Approximation of Differential Equations (Springer, Berlin, 1992). [66] K. K. Das, Phys. Rev. A 66, 053612 (2002). [67] B. Tanatar, A. Minguzzi, P. Vignolo, and M. P. Tosi, Phys. Lett. A 302, 131 (2002). [68] L. V. Hau et al., Phys. Rev. A 58, R54 (1998). [69] M. Holzmann, W. Krauth, and M. Naraschewski, Phys. Rev. A 59, 2956 (1999). [70] T. Bergeman, D. L. Feder, N. L. Balazs, and B. I. Schneider, Phys. Rev. A 61, 063605 (2000). [71] N. V. Prokof’ev and B. V. Svistunov, Phys. Rev. A 66, 043608 (2002). [72] N. Bogoliubov, J. Phys. USSR 11, 23 (1947). [73] A. L. Fetter, Ann. Phys. (N.Y.) 70, 67 (1972). [74] A. Griffin, Phys. Rev. B 53, 9341 (1996). [75] E. P. Gross, Nuovo Cim. 20, 454 (1961). [76] L. P. Pitaevskiı̆, Sov. Phys. JETP 13, 451 (1961). [77] G. Baym and C. Pethick, Phys. Rev. Lett. 76, 6 (1996). [78] A. L. Fetter, in Proceedings of the Enrico Fermi International Summer School on Bose-Einstein Condensation, edited by M. Inguscio, S. Stringari, and C. E. Wieman (IOS Press, Washington, 1999). [79] K. Huang and C. N. Yang, Phys. Rev. 105, 767 (1957). [80] W. J. Mullin, M. Holzmann, and F. Laloë, J. Low Temp. Phys. 121, 263 (2000). [81] D. S. Petrov, M. Holzmann, and G. V. Shlyapnikov, Phys. Rev. Lett. 84, 2551 (2000). [82] M. D. Lee and S. A. Morgan, J. Phys. B 35, 3009 (2002). [83] M. Houbiers and H. T. C. Stoof, Phys. Rev. A 54, 5055 (1996). [84] A. L. Fetter and J. D. Walecka, Quantum Theory of Many-Particle Systems (McGraw-Hill, New York, 1971). [85] P. Öhberg et al., Phys. Rev. A 56, R3346 (1997). 135 [86] F. Dalfovo, S. Giorgini, M. Guilleumas, L. P. Pitaevskiı̆, and S. Stringari, Phys. Rev. A 56, 3840 (1997). [87] F. Dalfovo and S. Stringari, Phys. Rev. A 53, 2477 (1996). [88] S. Giorgini, L. P. Pitaevskiı̆, and S. Stringari, J. Low Temp. Phys. 109, 309 (1997). [89] M. Holzmann and Y. Castin, Eur. Phys. J. D 7, 425 (1999). [90] J. des Cloiseaux, in Problème à N corps (Many-Body Physics), Les Houches 1967, edited by C. DeWitt and R. Balian (Gordon and Breach, New York, 1968). [91] D. A. W. Hutchinson, E. Zaremba, and A. Griffin, Phys. Rev. Lett. 78, 1842 (1997). [92] R. J. Dodd, M. Edwards, C. W. Clark, and K. Burnett, Phys. Rev. A 57, R32 (1998). [93] C. Gies, B. P. van Zyl, S. A. Morgan, and D. A. W. Hutchinson (2003), cond-mat/0308177. [94] M. Edwards and K. Burnett, Phys. Rev. A 51, 1382 (1995). [95] S. K. Adhikari, Phys. Rev. E 62, 2937 (2000). [96] V. V. Goldman, I. F. Silvera, and A. J. Leggett, Phys. Rev. B 24, 2870 (1981). [97] T. Haugset and H. Haugerud, Phys. Rev. A 57, 3809 (1998). [98] L. Salasnich, A. Parola, and L. Reatto, Phys. Rev. A 65, 043614 (2002). [99] Y. Castin and R. Dum, Eur. Phys. J. D 7, 399 (1999). [100] B. P. van Zyl, R. K. Bhaduri, and J. Sigetich, J. Phys. B 35, 1251 (2002). [101] C. E. Campbell, J. G. Dash, and M. Schick, Phys. Rev. Lett. 26, 966 (1971). [102] S. Giorgini, L. P. Pitaevskiı̆, and S. Stringari, Phys. Rev. A 54, R4633 (1996). [103] D. A. Huse and E. D. Siggia, J. Low Temp. Phys. 46, 137 (1982). [104] J. P. Fernández and W. J. Mullin, J. Low Temp. Phys. 128, 233 (2002). [105] L. N. Trefethen, Spectral Methods in Matlab (SIAM, Philadelphia, 2000). [106] P. Muruganandam and S. K. Adhikari, J. Phys. B 36, 2501 (2003). [107] H. Shi and W.-M. Zheng, Phys. Rev. A 56, 1046 (1997). [108] R. V. E. Lovelace and T. J. Tommila, Phys. Rev. A 35, 3597 (1987). [109] D. J. Thouless, in The New Physics, edited by P. Davies (Cambridge University Press, Cambridge, 1989). [110] M. Van den Berg, Phys. Lett. 78A, 88 (1980). [111] M. Naraschewski and D. M. Stamper-Kurn, Phys. Rev. A 58, 2423 (1998). [112] A. Minguzzi, S. Conti, and M. P. Tosi, J. Phys.: Condens. Matter 9, 33 (1997). [113] M. Bayindir and B. Tanatar, Phys. Rev. A 58, 3134 (1998). [114] J. G. Kim and E. K. Lee, J. Phys. B 32, 5575 (1999). [115] L. Reatto and G. V. Chester, Phys. Rev. 155, 88 (1967). [116] J. G. Kim, K. S. Kang, B. S. Kim, and E. K. Lee, J. Phys. B 33, 2559 (2000). 136 [117] J. P. Fernández and W. J. Mullin, The Bose-Einstein condensate in two dimensions (2001), poster presented at the European Research Conference on Bose-Einstein Condensation. [118] D. S. Fisher and P. C. Hohenberg, Phys. Rev. B 37, 4936 (1988). [119] N. V. Prokof’ev, O. A. Ruebenacker, and B. V. Svistunov, Phys. Rev. Lett. 87, 270402 (2001). [120] V. N. Popov, Functional Integrals in Quantum Field Theory and Statistical Physics (Reidel, Dordrecht, 1983). [121] U. A. Khawaja, J. O. Andersen, N. P. Proukakis, and H. T. C. Stoof, Phys. Rev. A 66, 013615 (2002). [122] P. G. de Gennes, Superconductivity of Metals and Alloys (W. A. Benjamin, New York, 1966). [123] T. Bergeman, Phys. Rev. A 55, 3658 (1997). [124] S. I. Shevchenko, Sov. J. Low Temp. Phys. 18, 223 (1992). [125] S. Pearson, T. Pang, and C. Chen, Phys. Rev. A 58, 4796 (1998). It was later found that this paper underestimates the effect of interactions; see W. Krauth, cond-mat/9812328. [126] D. M. Ceperley and E. L. Pollock, Phys. Rev. Lett. 56, 351 (1986). [127] E. L. Pollock and D. M. Ceperley, Phys. Rev. B 30, 2555 (1984). [128] M. Holzmann and W. Krauth, Phys. Rev. Lett. 83, 2687 (1999). [129] D. M. Ceperley and E. L. Pollock, Can. J. Phys. 65, 1416 (1987). [130] K. Binder and D. W. Heermann, Monte Carlo Simulation in Statistical Physics (Springer, Berlin, 1997), 3rd ed. [131] H. Gould and J. Tobochnik, An Introduction to Computer Simulation Methods (AddisonWesley, Reading, MA, 1996), 2nd ed. [132] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, J. Chem. Phys. 21, 1087 (1953). [133] M. E. J. Newman and G. T. Barkema, Monte Carlo Methods in Statistical Phyisics (Clarendon Press, Oxford, 1999). [134] W. Krauth, “Introduction to Monte Carlo algorithms”, unpublished (1996). [135] J. A. Barker, J. Chem. Phys. 70, 2914 (1979). [136] M. Holzmann, La transition de Bose-Einstein dans un gaz dilué, Ph.D. thesis, Université Paris VI (2000). [137] W. J. Mullin, S. D. Heinrichs, and J. P. Fernández, Physica B 284–288, 9 (2000). [138] E. Lieb and R. Seiringer, Phys. Rev. Lett. 88, 170409 (2002). [139] V. Menon and K. Pingali, in Proceedings of the 2nd USENIX Conference on Domain-Specific Languages, edited by T. Ball (1999). [140] S. Stringari, Phys. Rev. Lett. 76, 1405 (1996). [141] P. Sindzingre, M. L. Klein, and D. M. Ceperley, Phys. Rev. Lett. 63, 1601 (1989). [142] D. M. Ceperley and E. L. Pollock, Phys. Rev. B 39, 2084 (1989). [143] J. E. Robinson, Phys. Rev. 83, 678 (1951). 137 [144] W. Magnus, F. Oberhettinger, and R. P. Soni, Formulas and Theorems for the Special Functions of Mathematical Physics (Springer, Berlin, 1966), 3rd ed. [145] S. Flügge, Practical Quantum Mechanics, vol. 1 (Springer, Berlin, 1971). [146] M. Linn and A. L. Fetter, Phys. Rev. A 60, 4910 (1999). [147] T. Matsubara, Prog. Theor. Phys. 6, 714 (1951). [148] G. C. Wick, Phys. Rev. 80, 268 (1950). [149] T. Matsubara, Prog. Theor. Phys. 14, 351 (1955). [150] S. Y. Larsen, K. Witte, and J. E. Kilpatrick, J. Chem. Phys. 44, 213 (1966). [151] S. Y. Larsen, J. Chem. Phys. 48, 1701 (1968). [152] M. Takahashi and M. Imada, J. Phys. Soc. Jpn. 53, 963 (1984). [153] P. K. MacKeown, Am. J. Phys 53, 880 (1985). [154] P. Lévy, Compositio Math. 7, 283 (1939). 138