Sur les m´ethodes de Monte-Carlo quantique

L’article pr´esent´e [47] est une illustration du type de calculs num´eriques exacts que l’on peut effectuer tr`es simplement avec la nouvelle m´ethode de Monte-Carlo quantique que nous avons d´evelopp´ee avec Jean Dalibard et Iacopo Carusotto [36]. Il s’agit du calcul de la distribution de probabilit´e du nombre de particules dans un condensat en interaction, pour des temp´eratures allant de presque z´ero `a des valeurs sup´erieures `a la temp´erature critique.

En principe, les m´ethodes Monte-Carlo existantes (par int´egrale de chemin, Path In-tegral Monte Carlo) doivent pouvoir ˆetre utilis´ees. Cependant, l’observable consid´er´ee est tr`es complexe `a calculer avec ces m´ethodes, qui privil´egient l’espace des positions. Aussi les experts du domaine, qui maˆıtrisent la technique depuis des d´ecennies, n’avaient-ils en fait jamais calcul´e cette observable, qui est pourtant l’une des plus importantes physique-ment !

Par ailleurs, une comparaison est effectu´ee avec la th´eorie de Bogoliubov, le calcul de Bogoliubov de l’observable en question ´etant lui-mˆeme non trivial. Il y a excellent accord `a basse temp´erature. C’est probablement le test le plus s´ev`ere jamais effectu´e de la th´eorie de Bogoliubov, car l’observable mise en jeu n’est pas seulement `a un corps ou `a deux corps, mais `a N corps. Ceci a mis fin `a une pol´emique sur la validit´e de la th´eorie de Bogoliubov dans un syst`eme unidimensionnel, qui ´etait n´ee du fait que la th´eorie de Bogoliubov ne pr´edit qu’une seule branche d’excitation, alors que le calcul exact de Lieb et Lieniger en trouve deux.

Enfin, comme le montrent les simulations, une mesure de la distribution de probabilit´e du nombre de particules condens´ees permettrait de r´ev´eler tr`es directement la nature de type “´etat comprim´e du champ” de l’´etat fondamental du gaz, ce qui n’a pas encore ´et´e fait.

Notons que nous avons continu´e `a d´evelopper ces m´ethodes Monte-Carlo, en particulier pour voir la transition BCS d’un gaz unidimensionnel de fermions en interaction forte [57].

CHAPITRE 5. REPRODUCTION D’ARTICLES REPR ´ESENTATIFS 89

Condensate Statistics in One-Dimensional Interacting Bose Gases: Exact Results

Iacopo Carusotto and Yvan Castin

Laboratoire Kastler Brossel, Ecole Normale Supe´rieure, 24 rue Lhomond, 75231 Paris Cedex 05, France (Received 12 June 2002; published 22 January 2003)

Recently, a quantum Monte Carlo method alternative to the path integral Monte Carlo method was developed for solving the N-boson problem; it is based on the stochastic evolution of classical fields. Here we apply it to obtain exact results for the occupation statistics of the condensate mode in a weakly interacting trapped one-dimensional Bose gas. The temperature is varied across the critical region down to temperatures lower than the trap level spacing. We also derive the condensate statistics in the Bogoliubov theory: this reproduces the exact results at low temperature and explains the suppression of odd numbers of noncondensed particles at T’ 0.

DOI: 10.1103/PhysRevLett.90.030401 PACS numbers: 05.30.Jp, 02.70.Ss, 03.75.Hh

The first achievement of Bose-Einstein condensation in weakly interacting atomic gases in 1995 has renewed the interest on basic aspects of the Bose-Einstein condensa-tion phase transicondensa-tion [1]. In particular, an intense theo-retical activity has been recently devoted to the study of the occupation statistics of the condensate mode: although no experimental result is available yet, we ex-pect that it would provide a crucial test for many-body theories since, contrarily to more common one-body ob-servables such as the density, it involves arbitrarily high order correlation function of the quantum Bose field. While there are now well-established results for the ideal Bose gas [2], calculations for the weakly interacting Bose gas have been performed within the framework of mean-field approximation only [3–5]. The intermediate regime around the critical temperature where the mean-field theory fails has therefore been left unexplored. Some controversy is still open about the validity of the Bogoliubov approach even at temperatures much lower than the critical temperature [6]. In the absence of ex-perimental results, it is then very interesting to have exact theoretical results on the statistics of condensate occupation.

There exists an exact analytical solution to the bosonic N body problem [7], but restricted to the spatially homogeneous one-dimensional case. From the side of numerics, the path integral quantum Monte Carlo tech-nique is able to give exact predictions for any observable of the gas [8] and was successfully used to calculate the mean condensate occupation [9] and the critical tempera-ture [10]. In the path integral formulation, however, the position representation is priviledged which makes the calculation of highly nonlocal observables such as the condensate occupation probabilities rather involved.

In this Letter, we use instead a recently developed quantum Monte Carlo method based on the stochastic evolution of classical fields [11,12]. This new method has a much broader range of applicability than the standard path integral Monte Carlo technique: it can be applied to bosonic systems with complex wave functions, and even

to interacting Fermi systems [13]. As compared to the positive-P method of quantum optics [14], it has the decisive advantage of having been proven to be con-vergent. In this Letter, we perform the first nontrivial application of this new method, calculating for the first time the exact distribution function of the number of condensate particles in the presence of interactions, for temperatures across the Bose-Einstein condensation tem-perature [15].

An ultracold trapped interacting Bose gas in D dimen-sions can be modeled by the Hamiltonian

H X k  h2 k² 2m ^{a^a} y ka^a_k^g₂^0X r dV ^yr ^yr ^r ^r ^X r dVU_extr ^yr ^r; (1)

the spatial coordinate r runs on a discrete orthogonal lattice of N points with periodic boundary conditions; V is the total volume of the quantization box and dV  V=N is the volume of the unit cell of the lattice. U_extr is the external trapping potential, m is the atomic mass, and interactions are modeled by a two-body discrete delta potential with a coupling constant g₀. The field operators

^ 

r satisfy the usual Bose commutation relations  ^r; ^yr0  r;r0=dV and can be expanded on plane waves according to ^r P

ka^a_keikr=  V p

with k re-stricted to the first Brillouin zone of the reciprocal lattice. In order for the discrete model to correctly reproduce the underlying continuous field theory, the grid spacing must be smaller than the macroscopic length scales of the system like the thermal wavelength and the healing length.

In this Letter, we assume that the gas is at thermal equilibrium at temperature T in the canonical ensemble so that the unnormalized density operator is eq  eH with  1=kBT. It is a well-known fact of quan-tum statistical mechanics that this density operator can be obtained as the result of an imaginary-time evolution: P H Y S I C A L R E V I E W L E T T E R S _{24 JANUARY 2003}^{week ending} VOLUME90, NUMBER3

CHAPITRE 5. REPRODUCTION D’ARTICLES REPR ´ESENTATIFS 90

d_eq

d ^

1

2H _eq  _eq H (2) during a ‘‘time’’ interval   0 !  starting from the infinite temperature value _eq  0  1N. We have re-cently shown [12] that the solution of the imaginary-time evolution (2) can be written exactly as a statistical average of Hartree operators of the form

  jN:₁ihN:₂j; (3)

where, in both the bra and the ket, all N atoms share the same (not necessarily normalized) wave functions _ (  1; 2). For the model Hamiltonian (1) here consid-ered, this holds if each _ evolves according to the Ito stochastic differential equations:

dr  ^d 2  p2 2m^{ Uext}r  g0N  1^jrj2 jj_jj2 ^{g0N  1} 2 P r0dVj_r0j4 jj_jj4  _r  dB_r; (4)

with a noise term given by dB_r  i  d g₀ 2V s Q_  _r^X k>0 eikr
k c:c:  ; (5) where the projector Q_ projects orthogonally to _, the index k is restricted to a half space and the k are independent random angles uniformly distributed in 0; 2 .

Starting from this very simple but exact stochastic formulation, a Monte Carlo code was written in order to numerically solve the stochastic differential equation (4). As a first step, a sampling of the infinite temperature density operator has to be performed in terms of a finite number of random wave functions i. Since the effective contributions of the different realizations to the final averages can be enormously different, the statistics of the Monte Carlo results can be improved by using an importance sampling technique [8] in order to avoid wasting computational time. This is done using the iden-tities 1N^Z 1 D jN:ihN:j ^Z 1 P₀ D^jN:ihN:j P₀ ^; (6) where the integration is performed over the unit sphere jj2jj  hji ^P_rdV jj2 1 and where the a priori distribution function P₀ can be freely chosen in order to maximize the efficiency of the calculation. For the numerical calculations here reported, the following P₀ has been used

P₀  jjeh_GP=2jijj2N; (7)

since it joined the possibility of a simple sampling with a fast convergence. In this expression, hGP is the Gross-Pitaevskii Hamiltonian h_GP_2m^p2 U_extr  N g₀j₀rj2 , ₀ is the wave function which mini-mizes the Gross-Pitaevskii energy functional, and  is the corresponding chemical potential. With respect to the ideal Bose gas distribution function previously used [12], the present form (7) for P₀ has the advantage of taking into account the fact that the condensate mode can be strongly modified by interactions. More details on the actual sampling of P₀ can be found in [12].

Each realization ^i_1;2is then let to evolve according to the stochastic evolution in imaginary time (4) from its   0 value ^i_1;2 i to the inverse temperature of interest   . The expectation values of any observable at temperature T can then be calculated as averages over the Monte Carlo realizations; e.g., the partition function Tr is given by Tr  h₂j₁iN. In particular, the condensate wave function _BECr is the eigenvector of the one-body density matrix

h ^yr0 ^ri  ¹

Tr ^1r

2r0h₂j₁iN1; (8) corresponding to the largest eigenvalue, that is to the largest mean occupation number [16]. The complete probability distribution PN₀ for the occupation statis-tics of the condensate mode is obtained via the expression PN₀  Tr ^PP_N

0 Tr 1, where ^PP_N

0projects onto the subspace in which the condensate mode contains exactly N₀atoms; in terms of the ansatz (3), the expectation value of the projector can be written as

Tr ^PP_N0  ^N! N₀! N  N0! ^c  2c₁N0 h? 2j? 1iNN0; (9) where we have split _r as c_BECr  ?

r with ?

r orthogonal to the condensate wave function. In Figs. 1–3 we have summarized the results of the Monte Carlo calculations for a one-dimensional, har-monically trapped gas with repulsive interactions. In Fig. 1, we have plotted the condensate wave function density profile j_BECxj2for different values of the tem-perature, see solid lines. As the temperature decreases, the enhanced atomic density at the center of the cloud gives a stronger repulsion among the condensate atoms and therefore a wider condensate wave function. In this low temperature regime, we have found a good agreement between the Monte Carlo results and the Bogoliubov prediction for the condensate wave function including the first correction to the Gross-Pitaevskii equation due to the presence of noncondensed atoms [17]; see dashed lines in Fig. 1. This was expected, since the numerical examples in this paper are in the weakly interacting regime n"  15  1, where n is the density at the trap center and "   h2=2mg₀n1=2is the corresponding heal-ing length. For high temperatures, the wave function _BECx tends to the harmonic oscillator ground state.

P H Y S I C A L R E V I E W L E T T E R S week ending

24 JANUARY 2003 VOLUME90, NUMBER3

CHAPITRE 5. REPRODUCTION D’ARTICLES REPR ´ESENTATIFS 91

The wave function _BECx can then be used to deter-mine the occupation statistics of the condensate mode via (9). The signature of Bose condensation is apparent in Fig. 2: the occupation statistics of the condensate mode at high temperatures has the same shape as a noninteracting, thermally occupied mode, its maximum value being at N₀ 0. For decreasing T, PN₀ radically changes its shape; at low T, its maximum is at a nonvanishing value of N₀and its shape is strongly asymmetric with a longer tail going towards the lower N₀values, as already noticed in the mean-field approach of [5]. This transition ressem-bles the one occurring in a laser cavity for a pumping rate which goes from below to above threshold [18].

For even lower values of the temperature, see Fig. 3, most of the atoms are in the condensate and the

proba-bility distribution PN₀ tends to concentrate around val-ues of N₀close to N. However, interactions prevent the atomic sample from being totally Bose condensed even at extremely low temperatures. Moreover, the probability of having an odd number of noncondensed atoms is strongly reduced with respect to the probability of having an even number (see the strong oscillations in the kBT  0:4 h! curve of Fig. 3). A quantitative interpretation of this effect in terms of the Bogoliubov approximation will be given in the following part of the paper.

We now compare the exact results with the prediction of the number-conserving Bogoliubov approximation [17,19]. The Bogoliubov Hamiltonian has the form

H_Bog¹ 2^{dV  ~} y; ~  &L _~ ~  y  ; (10) with L  ^hGPQ₀Ng₀j₀j2Q₀ Q₀Ng₀2 0Q 0 Q 0Ng₀2 0Q₀ h_GPQ 0Ng₀j₀j2Q 0  : (11) The matrix & is defined according to

&  1 0 0 1

; (12)

and the projector Q₀projects orthogonally to the Gross-Pitaevskii ground state ₀. The operators ~ are defined according to  ~_r ^AA^y₀^{^}?r, where ^?r is the pro-jection of the field operator orthogonally to ₀and where

^ A

A₀jN₀:₀i  jN₀ 1:₀i if N₀> 0 and 0 otherwise. Neglecting the possibility of completely emptying the condensate mode, ^AA₀and ^AA^y₀ can be shown to commute. Under this approximation, the ~ are bosonic operators. The probability N of having N noncondensed atoms can be calculated within the Bogoliubov approxi-mation by means of the characteristic function [4]

f ^X N N ei N ’ Tr  ei  ^NN1 Z^eHBog  ; (13) with  ^NN  dV ~y ~ and Z  TreH_Bog ; when the probability of emptying the condensate mode is not

0 5 0 100 N₀ 0 0.02 0.04 0.06 0.08 P(N 0 ) (a) (b) (c) (d) (e)

FIG. 2. Solid lines: exact results for the distribution function PN0 of the number of condensate atoms for decreasing temperatures from left to right: kBT= h!  50 (a), 33 (b), 20 (c), 10 (d), 5 (e). Same parameters as in Fig. 1. For curves (c) –(e), comparison with Bogoliubov predictions (dashed lines); for curve (a), comparison with the ideal gas (diamonds). The Monte Carlo calculations were performed using 3  104

realizations on a grid of N  128 (a), (b) or N  64 points (c) –(e). 0 2 4 6 8 10 δN 0 0.2 0.4 0.6 0.8 Π (δ N) (a) (b)

FIG. 3. Solid lines: exact results (105 realizations on a N  16 point grid) for the distribution function N  PN0 N  N of the number of noncondensed atoms at low temperatures kBT= h!  0:4 (a) and 1 (b). Same parameters as in Figs. 1 and 2. Dashed lines: Bogoliubov predictions.

−3 −2 −1 0 1 2 3 x/a_ho 0 0.2 0.4 0.6 |φ^BEC (x)| 2 0 0.2 0.4 0.6 |φ^BEC (x)| 2 −3 −2 −1 0 1 2 3 x/a_ho (a) (b) (c) (d)

FIG. 1. Solid lines: exact results for the condensate wave function density profile j_BECxj2 for N  125 atoms in a one-dimensional harmonic trap at different temperatures kBT= h!  1 (a), 10 (b), 20 (c), 50 (d). ! is the trap frequency, jBECj2is in units of the inverse of aho 

 h=m! p

. The cou-pling constant is g₀ 0:08 h! a_ho. Comparison with Bogoliubov predictions (dashed lines) in (a) –(c) and with harmonic oscillator ground state (diamonds) in (d). In (a) the two lines are indistinguishable on the scale of the figure. Monte Carlo calculations were performed using 3  104realizations on a N  16 (a), 64 (b), (c) or 128 point (d) grid.

P H Y S I C A L R E V I E W L E T T E R S week ending

24 JANUARY 2003 VOLUME90, NUMBER3

CHAPITRE 5. REPRODUCTION D’ARTICLES REPR ´ESENTATIFS 92

totally negligible, the Bogoliubov prediction for N has to be truncated to the physically relevant N  N values and then renormalized to 1. The details of the calculation will be given elsewhere, we go here to the result f ’2N 1Y j1  _2* j 1 *_j ei1  *_j ₁₌₂ ; (14)

in which the *_j are the eigenvalues of M & tanhL=2; from such an expression, the occupation probabilities N are immediately determined by means of an inverse Fourier transform. The results are shown as dashed lines in Figs. 2 and 3 where they are compared to the exact results from Monte Carlo simula-tions; the agreement is excellent provided the noncon-densed fraction is small, i.e., at sufficiently low temperatures.

In order to understand the oscillating behavior shown by the lowest temperature curve of Fig. 3, we calculate the ratio of the probability oddof having an odd value of N to the probability evenof having an even value:

_{odd even}^{f0  f} f0  f^’ 1 R 1 R^{; (15)} where R Q j*¹⁼²_j  detM 1=2Q mtanh,m=2 and m runs over the N  1 eigenenergies ,_m of the Bogoliubov spectrum. In order to have oscillations in N, the ratio (15) must be small as compared to 1 which requires a temperature smaller than the energy of the lowest Bogoliubov mode. These oscillations are there-fore a property of the ground state of the system. In the Bogoliubov approximation, the Hamiltonian is quadratic in the field operators so that its ground state is a squeezed vacuum, which indeed contains only even values of N, a well-known fact of quantum optics [20]. From a con-densed matter point of view, several existing variational ansatz for the ground state corresponding to a condensa-tion of pairs also exhibit this property [21,22]. Physically, this is ultimately related to the fact that the leading interaction process changing the condensate occupation at zero temperature, that is the scattering of two conden-sate particles into two noncondensed modes and vice versa, does not change the parity of N, as it is apparent in the Bogoliubov Hamiltonian (10).

In conclusion, we have determined with a newly devel-oped stochastic field quantum Monte Carlo method the exact distribution function of the number N0of conden-sate particles in a one-dimensional weakly interacting trapped gas. The signature of Bose condensation is the appearance of a finite value for the most probable value of N0. At temperatures below the trap oscillation fre-quency, configurations with an odd number of noncon-densed particles are strongly suppressed, which we interpret successfully within the Bogoliubov approxima-tion. Possible extensions of this work are (i) the

determi-nation of critical temperature shift for an interacting Bose gas in the limit of a vanishing scattering length, still subject of some controversy [23], (ii) exact calculations for finite temperature Bose gases with vortices in rotating traps, and (iii) the determination of the BCS critical temperature in two-component Fermi gases.

This work was stimulated by interactions with I. Cirac. We acknowledge useful discussions with A. Leggett and J. Dalibard. I. C. acknowledges a Marie Curie grant from the EU under Contract No. HPMF-CT-2000-00901. Laboratoire Kastler Brossel is a Unite´ de Recherche de l’ENS et de l’Universite´ Paris 6, associe´e au CNRS.

[1] For a review, see F. Dalfovo, S. Giorgini, L. P. Pitaevskii, and S. Stringari, Rev. Mod. Phys. 71, 463 (1999). [2] M. Holthaus and E. Kalinowski, Ann. Phys. (N.Y.) 276,

321 (1999), and references therein.

[3] S. Giorgini, L. P. Pitaevskii, and S. Stringari, Phys. Rev. Lett. 80, 5040 (1998).

[4] V. Kocharovsky, Vl. Kocharovsky, and M. O. Scully, Phys. Rev. Lett. 84, 2306 (2000).

[5] A. Sinatra, C. Lobo, and Y. Castin, J. Phys. B 35, 3599 (2002).

[6] Z. Idziaszek, M. Gajda, P. Navez, M. Wilkens, and K. Rza¸z˙ewski, Phys. Rev. Lett. 82, 4376 (1999). [7] E. H. Lieb and W. Liniger, Phys. Rev. 130, 1605 (1963). [8] D. M. Ceperley, Rev. Mod. Phys. 71, S438 (1999). [9] W. Krauth, Phys. Rev. Lett. 77, 3695 (1996).

[10] P. Gru¨ter, D. Ceperley, and F. Laloe¨, Phys. Rev. Lett. 79, 3549 (1997).

[11] I. Carusotto, Y. Castin, and J. Dalibard, Phys. Rev. A 63, 023606 (2001).

[12] I. Carusotto and Y. Castin, J. Phys. B 34, 4589 (2001). [13] O. Juillet, Ph. Chomaz, D. Lacroix, and F. Gulminelli,

Phys. Rev. Lett. 88, 142503 (2002).

[14] A. Gilchrist, C.W. Gardiner, and P. D. Drummond, Phys. Rev. A 55, 3014 (1997).

[15] In [12] the condensate statistics is calculated for the ideal Bose gas only: the absence of noise terms makes the Monte Carlo method converge trivially. This is no longer the case in the presence of interactions.

[16] O. Penrose and L. Onsager, Phys. Rev. 104, 576 (1956). [17] Y. Castin and R. Dum, Phys. Rev. A 57, 3008 (1998). [18] M. O. Scully, Phys. Rev. Lett. 82, 3927 (1999). [19] C.W. Gardiner, Phys. Rev. A 56, 1414 (1997).

[20] D. F. Walls and G. J. Milburn, Quantum Optics (Springer-Verlag, Berlin, 1994).

[21] M. Girardeau and R. Arnowitt, Phys. Rev. 113, 755 (1959).

[22] A. J. Leggett, Rev. Mod. Phys. 73, 307 (2001).

[23] M. Wilkens, F. Illuminati, and M. Kraemer, J. Phys. B 33, L779 (2000); M. Holzmann, G. Baym, J.-P. Blaizot, and F. Laloe¨, Phys. Rev. Lett. 87, 120403 (2001); P. Arnold and G. Moore, ibid. 87, 120401 (2001); V. A. Kashurnikov, N.V. Prokof ’ev, and B.V. Svistunov, Phys. Rev. Lett. 87, 120402 (2001).

P H Y S I C A L R E V I E W L E T T E R S week ending

24 JANUARY 2003 VOLUME90, NUMBER3

CHAPITRE 5. REPRODUCTION D’ARTICLES REPR ´ESENTATIFS 93