Abstract
When classical systems fail to explore their entire configurational space, intriguing macroscopic phenomena like aging and glass formation may emerge. Also closed quantomechanical systems may stop wandering freely around the whole Hilbert space, even if they are initially prepared into a macroscopically large combination of eigenstates. Here, we report numerical evidences that the dynamics of strongly interacting lattice bosons driven sufficiently far from equilibrium can be trapped into extremely longlived inhomogeneous metastable states. The slowing down of incoherent density excitations above a threshold energy, much reminiscent of a dynamical arrest on the verge of a glass transition, is identified as the key feature of this phenomenon. We argue that the resulting longlived inhomogeneities are responsible for the lack of thermalization observed in large systems. Such a rich phenomenology could be experimentally uncovered upon probing the outofequilibrium dynamics of conveniently prepared quantum states of trapped cold atoms which we hereby suggest.
Introduction
The ergodicity axiom in classical statistical mechanics states that, during its time evolution, a closed macroscopic system uniformly explores the entire phase space compatible with conservation laws, so that the time average of any observable comes to coincide with the microcanonical ensemble average and, when the observable is local, also with the canonical Gibbs ensemble average. Nonetheless, ergodicity can be violated in classical systems, a noticeable example being glasses^{1}. Quantum effects might also spoil ergodicity by preventing the wave function from diffusing within all available configurations. This phenomenon is actually known to occur in the presence of disorder and manifests itself either by singleparticle^{2} or manyparticle^{3,4,5} wave function localization. However, alike classical models for glassy behavior, ergodicity breakdown in the quantum dynamics may not necessarily require disorder and it could instead be entirely due to frustrating dynamical constraints^{6,7}. This issue is currently attracting great interest^{8,9}, since well controlled realizations of closed quantum systems have become feasible upon trapping cold atomic species^{10}. Indeed, similarly to what can be done in numerical simulations, one can prepare atoms in a given initial state and probe their time evolution under a Hamiltonian whose parameters are fully under control, thus offering the unique opportunity to monitor the ergodicity principle at work in the quantum realm^{11}.
In this work, we report numerical evidences that an isolated system of strongly interacting bosons, modeling atoms in optical lattices, can be trapped during its evolution into longlived inhomogeneous metastable states, provided that its internal energy exceeds a certain threshold. We argue that the slowing down of highenergy incoherent excitations in the strongly correlated system is the key feature responsible for this dynamical arrest, much resembling a kind of glass transition. By formulating the problem in a different language, we explicitly show that a system initially prepared in a inhomogeneous state is unable to diffuse within the entire configurational space; such a dynamical localization in the manybody Hilbert space looks intriguing and may represent a kind of manybody Anderson localization^{3} that occurs without disorder. The above phenomenon is put in further relation with and deemed responsible for the lack of ergodicity observed in large finite size systems. This belief is confirmed by means of a novel timedependent variational Monte Carlo method that we introduce hereby and an experimental set up to uncover this very rich phenomenology is also suggested.
One of the simplest models that can be realized in experiments is the BoseHubbard Hamiltonizan^{10,12}:characterized by the amplitude J for a bosonic species of an atom to hop between nearestneighboring wells of an optical lattice and by a local repulsion U among atoms localized in the same potential well. The operators and b_{i} create and destroy, respectively, a boson on site i, and is the density operator^{36}. Experiments are often performed with anisotropic lattices that realize a collection of almost uncoupled chains, a fortunate case for numerical simulations that we shall mainly consider hereafter, apart from a brief excursion in two dimensions towards the end of the paper.
In one dimension, obstacles to ergodicity can arise in integrable models^{13}. However, the Hamiltonian (1) is not integrable and, indeed, there are experimental evidences that its dynamical evolution may succeed in fast relaxing to a thermal state. Specifically, in a recent experiment a system of ^{87}Rb atoms, well described by the Hamiltonian (1), has been prepared in a state in which the sites of the optical lattice were alternatively empty and singly occupied^{14}. This state was let evolve for different experimental conditions corresponding to different ratios U/J. Even at the largest value , the initial density profile (…1,0,1,0,…) was found to rapidly relax to the homogeneous thermal one , with half a boson per site, much faster than the integrable counterparts of noninteracting or infinitelyinteracting (i.e., hardcore) bosons and consistently with the increased number of relaxation channels that opens once integrability is lost.
Numerical simulations of the above experiment successfully reproduce the observed thermal behavior^{14}. However, there are also numerical evidences pointing out a breakdown of ergodicity in the same model (1) but within a different region of the parameter space, specifically when the number of bosons is one per site. This case, as well as any other one with integer density, is special because, at equilibrium and at zero temperature, the model (1) undergoes a quantum phase transition into a Mott insulator above a critical U/J. Even though the phase transition is washed out by thermal fluctuations, nonetheless its influence on the spectrum seems to prevent thermalization above a certain U/J in numerical simulations of finite size systems^{26,29}; a result that may^{27} or may not^{28} prelude to a true breakdown of ergodicity in the actual thermodynamic limit.
Here, we propose an experiment a lot alike the one previously described^{14}, which may distinguish very sharply a change from chaotic to nonchaotic dynamics in the BoseHubbard model (1). Our proposal starts by observing that, in the gapless phase next to the Mott insulator at one boson per site, lowenergy itinerant Bogoliubov quasiparticles should coexist with highenergy incoherent excitations, which, for , can be identified as sites occupied by more than a single boson. It is well possible that the overall relaxation depends critically upon the population of those highenergy excitations^{20}, which can be assessed by tailoring initial states with lots of doubly occupied sites rather than none as in the experiment of Ref.^{14}. This is indeed what we propose and simulate by means of different and complementary numerical tools, such as exact diagonalization, timeevolving block decimation^{15} and a novel timedependent variational Monte Carlo algorithm, as discussed in more detail in the Supplementary Material.
Results
Inhomogeneous initial states and dynamical localization
We start from analyzing the model at density n = 1, when at equilibrium a Mott transition occurs at a critical ^{16}. We imagine to prepare an initial state where all the sites are either empty or doubly occupied. In particular we consider the states depicted in Fig. 1, namely with clusters of doubly occupied sites of variable size S_{d}. These states are let evolve with the spatially homogeneous Hamiltonian (1) for different U/J, below and above the critical value. While for small U/J the density profile rapidly reaches the equilibrium configuration (…1,1,1,1,…), for large U/J, it stays close to its initial value for a remarkably long time. Eventually, since the system is finite, the density profile approaches the homogeneous plateau, with small residual oscillations that get damped as the system size increases.
We can define a relaxation time τ_{R} (whose inverse is shown in Fig. 2, for S_{d} = 1), as the first time for which the local density approaches its homogeneous value. We highlight that τ_{R}, at a specific , has a sudden step up, which becomes sharper and sharper as the system size increases. The above results show that above the system has the tendency to stay dynamically trapped into longlived inhomogeneous configurations.
We emphasize that such a behavior is all the more remarkable not because doubly occupied sites seem unable to decay, which is known to require for very long times and large system sizes^{18}, but rather because they are not capable to move, hence restore translational symmetry. We can better understand this surprising result by an effective Hamiltonian that can be derived following the same reasoning of Refs.^{19,20}. For a sufficiently large interaction, it is justified to project the evolution onto states with the same potential energy per site U/2, at least for time scales shorter than U^{2}/J^{3}. One realizes that states with the same potential energy but with triply and singly occupied sites start to contribute only at order J^{4}/U^{3}, so that, with accuracy J^{2}/U, the number of doubly occupied sites is conserved. If we associate a fictitious spin up or down to a doublon (doubly occupied site) or a holon (empty site), respectively, we find that the effective Hamiltonian that controls the evolution reads:which describes a hardaxis ferromagnetic Heisenberg model^{19}. We note that the evolution of an XXZ spin chain starting from an antiferromagnetic ordered initial state has been studied numerically in Ref.^{21} with DMRG. Here we consider the effective dynamics (2) in a regime when the anisotropy would favor a ferromagnetic ground state and when the initial state contains clusters of up and down spins of increasing lenght. The smaller cluster size of corresponds to a Neel initial state considered in Ref ^{21}. In the right panel of Fig. 2 we show the timedependence of the clusters density in the large U regime, evolved according to the effective Hamiltonian (2). Remarkably, we see that even for small clusters the system fails to restore the spatial homogeneity up to very long time scales, which turn to be far beyond those currently accessible in typical experimental setups.
The slowing down of the dynamics can be traced back to the effective attraction among doublons^{20}, which makes their aggregates hard to break up. In other words, what seems to matter more is the dissociation of clusters of doublons, rather than the decay of a single one. Indeed we have checked (not shown) that in the large U regime the system has the tendency to get dynamically stuck into clusters of doublons of finite size, whereas a much faster annihilation and recombination rate is observed in the small U limit.
To get further insights into the dynamical behavior of the system, we recast the problem in a different language. Starting from the initial state, denoted as 0〉, we can generate an orthogonal basis set i〉, i = 0, 1,…, by repeatedly applying the Hamiltonian (see Supplementary Material). In this Lanczos basis of manybody wave functions, the Hamiltonian has the form of a tightbinding model on a semiinfinite chain. Each site i = 0, 1,… corresponds to a manybody state, it has an onsite energy and is coupled only to its nearest neighbors by hopping elements t_{i}_{→i+1} and t_{i}_{→i−1}^{22}. It is easy to realize that the unitary evolution of the original many body problem is thus fully equivalent to the dynamics of a single particle, initially sitting at site 0, that is then let propagate along such a tightbinding chain of manybody states. We note that, both _{i} and t_{i}_{→i+1} largely fluctuate from site to site, therefore resembling an effective Anderson model, even though those parameters are in reality deterministic, see Fig. 3. In the same figure, we also show the mean distance traveled by the particle after time t starting from the first site of the chain, which corresponds to an initial state 0〉 ≡ (… 2, 0, 2, 0,…), and for different U/J. We observe that, for small values of U/J, the particle diffuses and its wavepacket finally spreads over the whole chain, in a rather uniform way. On the contrary, above a certain critical value of the interaction , the particle stays localized near the origin for arbitrarily long times. A closer look to the structure of the on site energies reveal the existence of a potential well at the edge of the chain. This is crucial in order to understand the observed localization transition. Since the potential well cannot induce a true bound state below the bottom of the spectrum, at most a resonance will form in the spectrum, from which the particle could in principle escape in a finite time. This indeed happens at small U but apparently not at large U, where the increased depth of the well and the effective randomness of the onsite energies conspire together to keep the particle localized close to the edge, preventing the the states in the well to hybridize with other states along the chain. We now see how this result connects with the previous analysis on the density relaxation times. In the small U/J regime the particle is able to escape from the well and to explore larger portions of the chain, thus resulting into a fast density relaxation. As opposite, for large U/J, the particle bounces back and forth inside the well, finding hard time to escape from it. This lack of diffusion results into a very longtime scale for the density to relax to its homogeneous value.
The above results show explicitly that some kind of localization in the manybody configurational space does occur, at least in the finite system^{3,23}. While such an intriguing behavior might well be a subtle effect due to the finite size spectrum, it could also signal the onset of a genuine localization that survives in the thermodynamic limit.
We conclude the discussion on the inhomogeneous initial states by studying the case at density n < 1, when at equilibrium there is no longer a Mott transition. For example, we consider n = 2/3 and an initial density profile (…2, 0, 0, 2, 0, 0,…). Interestingly, we find quite a different behavior for the density relaxation times τ_{R} (see Supplementary Material), with a much smoother crossover from small to large values of U/J and no evidence of any increase in the relaxation times with the system size. This fact suggests a deeper connection between the observed dynamical behavior and the zero temperature Mott transition that occurs at equilibrium and at integer filling, as suggested by calculations with infinitecoordination lattices^{24,25}.
Homogeneous initial states and Quantum Quenches
In light of the previous results, one may question that by choosing an inhomogeneous configuration of doublons we pick up a rather specific initial state in the Hilbert space. We are now going to show that the above findings strongly affect the dynamics starting from a perfectly homogeneous state. In this respect, a particularly interesting class of initial states are the ground states of for given values of the interaction U_{i}, which are let evolve under the Hamiltonian dynamics after a sudden increase of the interaction to a final value U_{f} > U_{i}, the socalled quantum quench. Kollath and coworkers^{26} reported evidence for the existence of two separated regimes in which either thermal or nonthermal behavior is observed for local observables. The origin of the nonthermal behavior in the large U_{f} region and the possibility of an ergodicity breaking in the thermodynamic limit is still highly debated^{29,30}.
In the following, we focus on an average density n = 1 and we show that signatures of long lived metastable states of doblons can be identified in the dynamics after a quantum quench. At variance with the previous numerical calculations, now both the initial state and the quantum Hamiltonian do preserve the spatial homogeneity and, therefore, the quest for possible signatures of ergodicity breaking requires a different approach. Since we have identified density relaxation as the slowest process in the problem, we monitor the dynamics of the system by measuring the autocorrelation of the density averaged over all sites, namely through
For any finite size system, since the densities decorrelate at very long times. Indeed, as shown in Fig. 4, for small U_{f} ≪ U_{i} this quantity has a very fast transient to zero. On the contrary, for U_{f} ≫ U_{i} the density autocorrelation gets stuck into a longlived finite value plateau , before approaching zero only on a much longer time scale. If we extract a relaxation time from , we find a similar behavior as in Fig. 2, i.e., a dramatic increase of the relaxation times above a threshold value of the final interaction strength.
In agreement with the previous analysis, the appearance of such a longlived metastable state characterized by a finite plateau of the density autocorrelation function might indicate an excess of double occupancies that have no channel to relax. In other words, the dynamical constraints brought by the interaction severely slow down density excitations, whose characteristic time scales increase abruptly after a critical threshold. The main phenomenological traits of this dynamical arrest characterized by longlived inhomogeneous states closely remind the physics of glassy materials.
Variational description, lack of thermalization and higher dimensions
The previous discussion revealed the existence of a threshold energy above which a steep increase of the relaxation time of density fluctuations takes place. In order to assess the relevance of this phenomenon for the dynamics of larger systems or even for higher spatial dimensions, it is desirable to devise a comprehensive alternative framework able to catch its very characteristics. Here, we introduce an approach based on realtime variational Monte Carlo (see Supplementary Material for details), which has two important advantages: it allows us to follow the evolution for times comparable to those accessible experimentally, which are much longer than tDMRG^{31,32}; it can be easily extended to higher dimensions. We mention that meanfieldlike variational approaches to the realtime dynamics of correlated systems have been developed in recent years^{24,25}.
However, although they seem to capture well the main features of the dynamical evolution, these methods are unable to describe important aspects such as damping and relaxation. On the contrary, our approach is sufficiently rich to account for damping and relaxation of local observables. It is based on a very simple and transparent outofequilibrium extension of the Jastrowlike variational wave function that was shown to describe quite accurately the equilibrium phase diagram of the BoseHubbard model^{33}:where Ψ_{0}〉 is the initial state and V_{ij}(n_{i}, n_{j}; t) is a Jastrow factor that depends on the occupancies n_{i} and n_{j} of two sites i and j and varies with time so to maintain the time evolution as close as possible to the true evolution via the Schroedinger equation (see Supplementary Material). The comparison between our approach and the tDMRG^{34} is reported in the Supplementary Materials, demonstrating the high accuracy of the timedependent variational Monte Carlo.
Results for the time evolution of a local observable, such as the potential energy, after a sudden quench from U_{i} = 2J to a final U_{f} are shown in Fig. 5. The values of the thermal averages have been computed in the grandcanonical ensemble by means of finitetemperature quantum Monte Carlo calculations^{35}, with the effective temperature fixed by the average energy of the initial state, which we take as the best variational approximation for the ground state at U_{i} = 2J.
As shown in Fig. 5, in the region of small U_{f} we observe a damping of the average potentialenergy, which approaches a quasisteady stationary value in contrast to the simple Gutzwiller wave function^{24,25}. In this regime, damping is mainly due to a densitydensity Jastrow factor of the form V_{ij}(n_{i}, n_{j};) = v_{ij}(t) n_{i} n_{j}, which already at equilibrium was shown to provide a satisfactory description of the physical behavior^{33}. This fact enlightens the relevance of the Bogoliubov modes whose dephasing during the time evolution allows to approach the stationary state. Remarkably, the steady state averages coincide with the thermal ones; a signal that the dynamics is ergodic.
In the region of large interactions U_{f}, a simple densitydensity Jastrow factor does not account for all relaxation pathways, which will now mainly result from specific correlations among doublons, holons and between holons and doublons. The effective Hamiltonian (2) indeed explicitly shows that doublons attract each other as well as holons do, while doublons repel holons. These correlations, as well as other among higher onsite densities, can be easily implemented via the the Jastrow factor in Eq. (4) and indeed substantially improve the dynamics. Interestingly, the effective interaction between doblons that results from the dynamical variational calculation turns to be attractive, therefore leading to a consistent determination of the anticipated dynamical effects that drive the dynamics in this regime. As we see from Fig. 5, in the region of very large U_{f} the potentialenergy expectation values do show a damping to a nonthermal quasisteady state on a time of the order of τ_{D} ∼ 1/U_{f}. This fast time scale must be put in comparison with the much longer one, τ_{R} of Fig. 4. It is natural to identify τ_{R} with the time scale that controls the eventual escape from the quasisteady state, hence the approach to thermal equilibrium. Whether this time scale does truly diverge in the thermodynamic limit, or rather saturates to a very large value which is still out of reach for state of the art numerics, it is certainly an important issue that cannot be definitively solved. However, we can safely state that a large finite system of actual experimental relevance will get stuck for a quite long time into highly inhomogeneous metastable states, which we revealed to be on the verge of a spatial symmetry breaking.
The above results have been obtained for outofequilibrium one dimensional systems and therefore leave the questions concerning the dependence on the dimensionality of the problem still open. However, the anomalously longtime relaxation of the density autocorrelation points towards a kind of glassy behavior that should be observable even in higher dimensions, provided that the interaction induces sufficiently strong dynamical constraints. In this regard, we have studied the twodimensional case by means of our timedependent variational scheme, verifying that a similar behavior occurs even in two dimensions. In Fig. 6, we show the results of the potentialenergy expectation values as a function of the time. As before, if the repulsion is weak, the time average coincides with the thermal ones, while for strong repulsion there is a clear difference between thermal and time averages, giving rise to a nonergodic dynamics. We therefore argue that the phenomenology we have hereby identified is almost independent on the dimensionality and it is rather due to the existence in strongly correlated systems of highenergy incoherent excitations that do not have channels to relax efficiently. This scenario is also consistent with the observed strong doping dependence of the relaxation time (see Supplementary Material).
Discussion
In conclusion, we have found that the dynamical constraints brought by a strong interaction can trap the evolution of repulsive bosons hopping on a lattice into metastable states that lack translational symmetry, provided that the energy stored into the initial state is above a threshold. We pointed out that a selfinduced effective attraction among doublons is one of the major processes that can effectively freeze the dynamics on long time scales. Such a mechanism is recognized to play a role in the density relaxation processes of purely homogeneous systems through a dynamical arrest visible in timedependent density correlations. The main features of this intriguing behavior, namely the slowing down of density excitations and the longlived inhomogeneous pattern, resembles closely a kind of glass transition.
Moreover, we have shown that the time evolution of the manybody problem can be mapped onto that of a particle moving from the edge of a semiinfinite tightbinding chain with nearestneighbor hopping, where each site represents a manybody wave function. This model looks like an Anderson model, since both the onsite energy and the hopping vary from site to site, with a potential well at one edge due to the highenergy content of the initial state. Interestingly, we find a delocalizationlocalization transition in this problem, with the particle being unable to diffuse on the whole chain above a certain value of the well depth. We consider this analogy quite suggestive and potentially constitute an even stronger indication of ergodicity breaking in the manybody space which is worth to be further investigated.
References
 1.
Parisi, G., Mezard, M. & Virasoro, M. A. Spin glass theory and beyond (World Scientific, Singapore, 1987).
 2.
Anderson, P. W. Absence of diffusion in certain random lattices. Phys. Rev. 109, 1492–1505 (1958).
 3.
Basko, D., Aleiner, I. & Altshuler, B. Metalinsulator transition in a weakly interacting manyelectron system with localized singleparticle states. Annals of Physics 321, 1126–1205 (2006).
 4.
Oganesyan, V. & Huse, D. A. Localization of interacting fermions at high temperature. Phys. Rev. B 75, 155111 (2007).
 5.
Pal, A. & Huse, D. A. Manybody localization phase transition. Phys. Rev. B 82, 174411 (2010).
 6.
Biroli, G. & Mézard, M. Lattice glass models. Phys. Rev. Lett. 88, 025501 (2001).
 7.
Foini, L., Semerjian, G. & Zamponi, F. Quantum BiroliMézard model: Glass transition and superfluidity in a quantum lattice glass model. Phys. Rev. B 83, 094513 (2011).
 8.
Rigol, M., Dunjko, V. & Olshanii, M. Thermalization and its mechanism for generic isolated quantum systems. Nature 452, 854–858 (2008).
 9.
Polkovnikov, A., Sengupta, K., Silva, A. & Vengalattore, M. Non equilibrium dynamics of closed interacting quantum systems. Rev. Mod. Phys. 83, 863–883 (2011).
 10.
Bloch, I., Dalibard, J. & Zwerger, W. Manybody physics with ultracold gases. Rev. Mod. Phys. 80, 885–964 (2008).
 11.
von Neumann, J. Beweis des ergodensatzes und deshtheorems in der neuen mechanik. Z. Physik 57 (1929).
 12.
Jaksch, D., Bruder, C., Cirac, J. I., Gardiner, C. W. & Zoller, P. Cold bosonic atoms in optical lattices. Phys. Rev. Lett. 81, 3108–3111 (1998).
 13.
Calabrese, P. & Cardy, J. Quantum quenches in extended systems. J. Stat. Mech. Theor. Exp. P06008 (2007).
 14.
Trotzky, S. et al. Probing the relaxation towards equilibrium in an isolated strongly correlated 1d bose gas. arXiv: 1101.2659 (2011).
 15.
Vidal, G. Efficient simulation of onedimensional quantum manybody systems. Phys. Rev. Lett. 93, 040502 (2004).
 16.
Kühner, T. D. & Monien, H. Phases of the onedimensional BoseHubbard model. Phys. Rev. B 58, R14741–R14744 (1998).
 17.
Timeevolving block decimation open source code. URL http://physics.mines.edu/downloads/software/tebd/.
 18.
Kollath, C. Roux, G. Biroli, G., Läuchli, A. M. Statistical properties of the spectrum of the extended BoseHubbard model . J. Stat. Mech. 2010, P08011 (2010).
 19.
Petrosyan, D., Schmidt, B., Anglin, J. R. & Fleischhauer, M. Quantum liquid of repulsively bound pairs of particles in a lattice. Phys. Rev. A 76, 033606 (2007).
 20.
Rosch, A., Rasch, D., Binz, B. & Vojta, M. Metastable superfluidity of repulsive fermionic atoms in optical lattices. Phys. Rev. Lett. 101, 265301 (2008).
 21.
Barmettler, P., Punk, M., Gritsev, V., Demler, E. & Altman, E. Quantum quenches in the anisotropic spin1/2 Heisenberg chain: different approaches to manybody dynamics far from equilibrium. New J. Phys. 12, 055017 (2010).
 22.
Lanczos, C. An iteration method for the solution of the eigenvalue problem of linear differential and integral operators. J. Res. Natl. Bur. Stand. 45, 225–280 (1950).
 23.
Canovi, E., Rossini, D., Fazio, R., Santoro, G. E. & Silva, A. Quantum quenches, thermalization, and manybody localization. Phys. Rev. B 83, 094431 (2011).
 24.
Schiró, M. & Fabrizio, M. Timedependent mean field theory for quench dynamics in correlated electron systems. Phys. Rev. Lett. 105, 076401 (2010).
 25.
Sciolla, B. & Biroli, G. Quantum quenches and offequilibrium dynamical transition in the infinitedimensional BoseHubbard model. Phys. Rev. Lett 105, 220401 (2010).
 26.
Kollath, C., Läuchli, A. M. & Altman, E. Quench dynamics and nonequilibrium phase diagram of the BoseHubbard model. Phys. Rev. Lett. 98, 180601 (2007).
 27.
Cassidy, A. C. Mason, D. Dunjko, V. Threshold for Chaos and Thermalization in the OneDimensional MeanField BoseHubbard Model . Phys. Rev. Lett. 102, 025302 (2009).
 28.
Biroli, G. Kollath, C. Läuchli, A. M. Effect of Rare Fluctuations on the Thermalization of Isolated Quantum Systems . Phys. Rev. Lett. 105, 250401 (2010).
 29.
Roux, G. Quenches in quantum manybody systems: Onedimensional BoseHubbard model reexamined. Phys. Rev. A 79, 021608(R) (2009).
 30.
Biroli, G., Kollath, C. & Läuchli, A. M. Effect of rare fluctuations on the thermalization of isolated quantum systems. Phys. Rev. Lett. 105, 250401 (2010).
 31.
White, S. R. & Feiguin, A. E. Realtime evolution using the density matrix renormalization group. Phys. Rev. Lett. 93, 076401 (2004).
 32.
Daley, A. J. et al. Timedependent densitymatrix renormalizationgroup using adaptive effective hilbert spaces. J. Stat. Mech. Theor. Exp. P04005 (2004).
 33.
Capello, M., Becca, F., Fabrizio, M. & Sorella, S. Superfluid to mottinsulator transition in BoseHubbard models. Phys. Rev. Lett. 99, 056402 (2007).
 34.
Kollath, C. Private Communication (2011).
 35.
Bauer, B. et al. The alps project release 2.0: Open source software for strongly correlated systems. J. Stat. Mech. P05001 (2011).
 36.
In the experimental setup, an additional confining potential is usually present. We do not expect the latter to play a major role in what we shall discuss.
Acknowledgements
We acknowledge discussions with G. Biroli, J.F. Carrasquilla, D. Huse, A. Parola, E. Tosatti, F. Zamponi, and C. Kollath also for providing us with tDMRG numerical data. Computational support from CASPUR through the standard HPC Grant 2011 is also acknowledged.
Author information
Affiliations
SISSA – Scuola Internazionale Superiore di Studi Avanzati and CNRIOM DEMOCRITOS Simulation Center, Via Bonomea 265 I34136, Trieste, Italy
 Giuseppe Carleo
 , Federico Becca
 & Michele Fabrizio
Princeton Center for Theoretical Science and Department of Physics, Princeton University, Princeton, New Jersey. Tel: 08544, USA
 Marco Schiró
ICTP – The Abdus Salam International Center for Theoretical Physics, P.O. Box 586, I34151, Trieste, Italy
 Michele Fabrizio
Authors
Search for Giuseppe Carleo in:
Search for Federico Becca in:
Search for Marco Schiró in:
Search for Michele Fabrizio in:
Contributions
All authors conceived and designed the research, equally contributing to the preparation of the manuscript. G.C. also designed and carried out the numerical calculations.
Competing interests
The authors declare no competing financial interests.
Corresponding author
Correspondence to Giuseppe Carleo.
Supplementary information
PDF files
 1.
Supplementary Information
Supplementary Infromation
Rights and permissions
This work is licensed under a Creative Commons AttributionNonCommercialShareALike 3.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/byncsa/3.0/
To obtain permission to reuse content from this article visit RightsLink.
About this article
Further reading

Restricted Boltzmann machines in quantum physics
Nature Physics (2019)

Weak ergodicity breaking from quantum manybody scars
Nature Physics (2018)

Test of quantum thermalization in the twodimensional transversefield Ising model
Scientific Reports (2016)

Quantum manybody systems out of equilibrium
Nature Physics (2015)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.