Measurement of the mobility edge for 3D Anderson localization
Anderson localization is a universal phenomenon affecting noninteracting quantum particles in disorder. In three spatial dimensions it becomes particularly interesting to study because of the presence of a quantum phase transition from localized to extended states, predicted by P.W. Anderson in his seminal work Anderson (), taking place at a critical energy, the socalled mobility edge. The possible relation of the Anderson transition to the metalinsulator transitions observed in materials Lee () has originated a flurry of theoretical studies during the past 50 years Lee (); Thouless (); Abrahams (); Vollhardt (); Kramer (); Mirlin (); Review (), and it is now possible to predict very accurately the mobility edge starting from models of the microscopic disorder Slevin (). However, the experiments performed so far with photons, ultrasound and ultracold atoms, while giving evidence of the transition Hu (); Sperling (); Chabe (); Kondov (); Jendrzejewski (), could not provide a precise measurement of the mobility edge. In this work we are able to obtain such a measurement using an ultracold atomic system in a disordered speckle potential, thanks to a precise control of the system energy. We find that the mobility edge is close to the mean disorder energy at small disorder strengths, while a clear effect of the spatial correlation of the disorder appears at larger strengths. The precise knowledge of the disorder properties in our system offers now the opportunity for an unprecedented experimenttheory comparison for 3D Anderson localization, which is also a necessary step to start the exploration of novel regimes for manybody disordered systems.
The initial theoretical work by Anderson Anderson () was motivated by the observation of localization phenomena in solidstate systems Feher (). Studying a specific model of a disordered lattice, he realized that a critical amount of disorder could lead to a localization transition for electrons, due to quantum interference on the wavefunction. The possible connection between the Anderson transition and metalinsulator transitions Mott () started a huge theoretical investigation, which eventually led to a consensus between numerical Thouless (); Kramer (); Mirlin (); Slevin (); Review () and analytical results Abrahams (); Vollhardt (); Lee (); Review () about the mobility edge and the critical properties at the transition. However, these results cannot be tested on electronic systems, where interactions turn the singleparticle Anderson problem into a much more complex one Review (). The universal nature of Anderson localization allows to study some aspects of these phenomena with sound and light waves Hu (); Sperling (), which have provided a test of the IoffeRegel criterion Ioffe (), or with atomic kicked rotors Chabe (), which have been employed to measure the critical properties of the transition. Ultracold atoms in disordered optical potentials can realize a system with full control on the microscopic Hamiltonian as in the Anderson theory Lewenstein (); Billy (); Roati1D (). However, recent experiments Kondov (); Jendrzejewski (), while demonstrating the occurrence of an Anderson transition, could not precisely locate the mobility edge because of the difficulty in controlling the system energy. We now develop a novel method to control the energy of an ultracold atomic system in disorder, and we use it to measure precisely the mobility edge and its dependence on the disorder strength. To realize the disorder we employ an optical speckle potential, for which the intensity distribution and spatial correlations can be precisely measured. Our energycontrol strategy is based on three key parts. First, we achieve fullylocalized, lowenergy samples by loading a BoseEinstein condensate almost adiabatically into the disorder, thanks to a slow cancellation of the atomatom interactions. Then, we estimate the actual energy distribution of the samples by combining measurements of their kinetic energy with numerical simulations of the low energy eigenstates. Finally, we use a timedependent modulation of the disorder to produce controlled excitations and we deduce the mobility edge from the measurement of the energy needed to break localization.
The ultracold sample is composed by about 10 K atoms, for which the interaction can be controlled via a Feshbach resonance Roati (). The sample is initially cooled down to BoseEinstein condensation with repulsive interaction in a harmonic trap. The disorder is generated by two coherent speckle fields Jendrzejewski () that are bluedetuned from the atomic transitions, hence producing a repulsive potential. As sketched in Fig.1, the two speckles cross each other at 90 with parallel linear polarizations, creating a 3D intensity distribution with short correlation lengths along all directions, whose geometric average is =1.3m. The two relevant energy scales are the disorder strength , which represents both the mean value of the potential and its standard deviation, and the correlation energy 73nK. can be controlled via the total speckles power, and it is accurately calibrated. The two speckles envelopes are Gaussian with waists of about 1300m, hence much wider than the typical atomic distributions, ensuring the homogeneity of .
To prepare a lowenergy system in the disorder, we slowly increase from zero to a finite value while reducing to zero both the interactions and the harmonic confinement. In order to characterize the diffusive or localized nature of the system, we then let it evolve in the disordered potential for a variable time and we finally image the spatial distribution using absorption imaging. We study in particular the onedimensional density distribution , obtained from the integration along two spatial dimensions (an example of the corresponding evolution along and is reported in Extended Data Fig.3). Measuring the integrated second moment of , we observe a strong dependence of the expansion dynamics on (Fig.2a). For few smaller values of , the evolution is purely diffusive, according to . The linear increase of turns into anomalous diffusion for intermediate disorder and eventually, for larger , a small initial increase of is followed by a plateau, indicating that any further expansion of the cloud is inhibited and the system is fully localized. Fig.2b shows a typical time evolution of in the large regime; there is an initial expansion of the tails that explains the shorttime increase of , while for longer times the shape of apparently stops changing.
Studying the shorttime evolution of the cloud, we also find that the breakdown of the diffusive transport for 10nK is associated to an abrupt reduction of the shorttime diffusion coefficient, down to . This is indeed the predicted transport regime where quantum interference should suppress diffusion and lead to localization (see Extended Data Fig.5).
The observation of a rather sharp transition to localization with increasing suggests that the energy spread of the system is narrow. We can experimentally access only the distribution of kinetic energy, which is measured through a standard timeofflight technique. Just after loading the sample into the speckles, we release it abruptly from the disordered potential and we reconstruct the onedimensional momentum distribution along the direction from the time evolution of the density (see Extended Data Fig.6). We find that in the localized regime, i.e. for 18nK, the mean kinetic energy always lies well below . This motivated us to perform a numerical study of the lowenergy eigenstates by exact diagonalization of the disorder Hamiltonian, to reconstruct the total energy distribution from the measured . Numerical simulations are notoriously hard for energies close to the mobility edge, due to the finite spatial size in the simulations, but we found that a box with side length of about 10 could give reliable results in the lowenergy range occupied by the system. We calculate the momentumspace density of the energy eigenstates , which allows to relate to the energy distribution function through . In practice, we search for the that best reproduces the experimental , as sketched in Fig.3. We find that an exponential form, , provides a very good agreement with the experimental data. The energy distribution is then determined as usual as . The typical is narrow and peaked at an energy not far from the lowest energy (Fig.3b), in agreement with the observation of fully localized systems.
The final step to determine the mobility edge, from now on , consists in producing a controlled energy excitation in the system, so as to promote the initial towards diffusive high energy states. After loading the atoms into the disorder, we apply a weak sinusoidal modulation to the laser power for 0.5s. This corresponds to a timedependent perturbation of the disordered potential , with and variable . In the small limit this procedure allows to excite a fraction of the atoms by exactly . The final energy distribution can be written as , where is the initial energy distribution and is the probability to excite an atom at energy to . Despite is in principle depending both on and , in the Supplementary Information we estimate that this dependence is weak in the relevant range of energies, so that we can take as a constant. At the end of the modulation sequence we leave the disorder at fixed for another 0.5s, allowing the atoms transferred to diffusive states to expand enough to be effectively not visible to our imaging system. The transfer to diffusive states is hence detected as atom losses.
Fig.4 shows an example of the final atom number measured for different modulation frequencies. To determine we fit the data with , where is the only free parameter. For the specific dataset in Fig.4 we obtain nK. The agreement between the data and the model is in general very good until a large regime where an unexpected increase of occurs. This can be justified considering the poor overlap in momentum space between deeply localized states and the essentially free states of the continuum, which reduces the excitation probability . In absence of a precise modelling, we exclude the data at high frequency from the fit. A test of the validity of our model is provided by the evolution of both system size and kinetic energy with (Fig.4bc). They indeed reach a maximum value around the same frequency for which the model indicates an optimal transfer to localized states just below , which therefore have the largest localization lengths and the largest energies (see the computed in the inset of Fig.4b). For , the peak of the excited part of crosses and an increasingly larger fraction of atoms diffuses away and gets effectively lost, not contributing to the size nor to the kinetic energy, which therefore decrease again.
We repeated this procedure for several disorder strengths; a summary of the measured trajectory for in the disorderenergy plane is shown in Fig.5. The most interesting regime for Anderson localization is the one where , since trapped states in individual wells of the disorder are extremely rare, and localization is caused by destructive interference on the singleparticle wavefunction over many wells and barriers. In this regime we observe an almost linear scaling of with , which is justified by the fact that is the only relevant energy scale in the system. When is increased above , the trajectory bends down. A saturation of for large might actually be expected. When , so that the tunneling through the random barriers is suppressed, the particle mean free path reaches a minimum value which is set by the mean distance between the barriers, i.e. . According to the IoffeRegel criterion Ioffe (), this corresponds to a maximum kinetic energy at the transition.
Our results are rather different from a previous experimental report of , relative to a speckle potential with different spatial correlations Kondov (). That experiment was however performed on a much shorter time scale and the analysis neglected the potential energy of the disorder, both approaches which do not allow a direct comparison with the present results. We also find only a qualitative agreement with theoretical determinations of for isotropic speckles Yedjour (); Piraud2 (); Delande (). In particular, our results for are about 20% higher than those based on selfconsistent approaches Yedjour (); Piraud2 (), or about 40% higher than those obtained via numerical computation Delande (). These deviations might be at least partially attributed to the different anisotropy of the speckles in experiment and theory, and are definitely worth further studies. We note that a comparison to lattice models Anderson () could be carried out only close to the band edge, where the lattice contributes only with an effective mass, but obtaining indisputable results in that regime has proven to be challenging Bulka ().
In conclusion, we have experimentally determined the mobility edge trajectory in a system with controlled microscopic disorder and tunable energy. This will allow for the first time a direct comparison of an experiment to numerical calculations and may also provide a test of analytical theories. A further narrowing of the energy distribution could allow us to measure in the future also the critical properties at the transition. Our technique is general for atoms and can be applied to other types of disorder. A full assessment of the noninteracting problem is the prerequisite for exploring challenging problems for interacting disordered systems, such as manybody localization Basko (), anomalous diffusion Cherroret (), or BoseEinstein condensation Huang (); Natterman (); Pilati (), where already making theoretical predictions becomes very hard.
References
 (1) Anderson, P. W. Absence of diffusion in certain random lattices. Phys. Rev. 109, 14921505 (1958).
 (2) Lee, P. A. & Ramakrishnan, T. V. Disordered electronic systems. Rev. Mod. Phys. 57, 287 (1985).
 (3) 50 Years of Anderson Localization. E. Abrahams ed. (World Scientific, 2012).
 (4) Edwards, J. T. & Thouless, D. J. Numerical studies of localization in disordered systems. J. Phys. C: Solid State Phys. 5, 807 (1972).
 (5) Abrahams, E. et al. Scaling theory of localization: Absence of quantum diffusion in two dimensions. Phys. Rev. Lett. 42, 673676 (1979).
 (6) Vollhardt, D. & Wölfle, P. in Electronic Phase Transitions (eds. Hanke, W. & Kopaev, Yu. V.) 178 (Elsevier, Amsterdam, 1992)
 (7) Kramer, B. & MacKinnon, A. Localization: theory and experiment. Rep. Prog. Phys. 56, 14691564 (1993).
 (8) Evers, F. & Mirlin, A. D. Anderson transitions. Rev. Mod. Phys. 80, 13551417 (2008).
 (9) Slevin, K. & Ohtsuki, T. Critical exponent for the Anderson transition in the threedimensional orthogonal universality class. New J. Phys. 16, 015012 (2014).
 (10) Hu, H. et al. Localization of ultrasound in a threedimensional elastic network. Nat. Phys. 4, 845848 (2008).
 (11) Sperling, T. et al. Direct determination of the transition to localization of light in three dimensions. Nat. Photonics 7, 4852 (2013).
 (12) Chabé, J. et al. Experimental observation of the Anderson metalinsulator transition with atomic matter waves. Phys. Rev. Lett. 101, 255702 (2008).
 (13) Kondov, S. S. et al. Threedimensional Anderson localization of ultracold matter. Science, 334, 6668 (2011).
 (14) Jendrzejewski, F. et al. Threedimensional localization of ultracold atoms in an optical disordered potential. Nat. Phys. 8, 398403 (2012).
 (15) Feher, G. & Gere, E. A. Electron spin resonance experiments on donors in silicon. II. Electron spin relaxation effects. Phys. Rev. 114, 1245 (1959).
 (16) Mott, N. F. Metalinsulator transitions. Phys. Today 31, 42 (1978).
 (17) Ioffe, A. F. & Regel, A. R. Noncrystalline, amorphous and liquid electronic semiconductors. Prog. Semicond. 4, 237291 (1960).
 (18) Lewenstein, M. et al. Ultracold atomic gases in optical lattices: mimicking condensed matter physics and beyond. Adv. Phys. 56, 243379 (2007).
 (19) Billy, J. et al. Direct observation of Anderson localization of matterwaves in a controlled disorder. Nature 453, 891 (2008).
 (20) Roati, G. et al. Anderson localization of a noninteracting BoseEinstein condensate. Nature 453, 895 (2008).
 (21) Roati, G. et al. K BoseEinstein condensate with tunable interactions. Phys. Rev. Lett. 99, 010403 (2007).
 (22) Yedjour, A. & Van Tiggelen, B. A. Diffusion and localization of cold atoms in 3D optical speckle. Eur. Phys. J. D 59, 249255 (2010).
 (23) Piraud, M., Aspect, A. & SanchezPalencia, L. Anderson localization of matter waves in tailored disordered potentials. Phys. Rev. A 85, 063611 (2012).
 (24) Delande, D. & Orso, G. Mobility edge for cold atoms in laser speckle potential. ArXiv eprint: arXiv:1403.3821 (2014).
 (25) Bulka, B., Schreiber, M. & Kramer, B. Localization, quantum interference and the metalinsulator transition. Z. Phys. B  Condensed Matter 66, 2130 (1987).
 (26) Cherroret, N., Vermersch, B., Garreau, J. C. & Delande, D. How nonlinear interactions challenge the threedimensional Anderson transition. ArXiv eprint: arXiv:1401.1038 (2014).
 (27) Basko, D. M., Aleiner, I. L. & Altshuler, B. L. Metalinsulator transition in a weakly interacting manyelectron system with localized singleparticle states. Annals of Physics 321, 11261205 (2006).
 (28) Huang, K. & Meng, H.F. Hardsphere Bose gas in random external potentials. Phys. Rev. Lett. 69, 644 (1992).
 (29) Nattermann, T. & Pokrovsky, V. L. BoseEinstein condensates in strongly disordered traps. Phys. Rev. Lett. 100, 060402 (2008).
 (30) Pilati, S., Giorgini, S. & Prokofev, N. Superfluid transition in a Bose gas with correlated disorder. Phys. Rev. Lett. 102, 150402 (2009).

*These two authors contributed equally.

Acknowledgements We acknowledge discussions with V. Josse, L. Pezzé and D. S. Wiersma. This work was supported by ERC (grants 247371 and 258325).

Author Contributions G.S and M.L. designed the experiment; G.S, M.L. and G.M. analysed the data and performed the numerical simulations; all the other authors participated to the experiment, data analysis, discussion of the results and writing of the manuscript.

Correspondence Correspondence and requests for materials should be addressed to G.M. (email: modugno@lens.unifi.it).
EXTENDED DATA FIGURES
References
 (31) Goodman, J. W. Speckle phenomena in optics: theory and applications. (Roberts and Company Publishers, 2007).
 (32) Landini, M. et al. Direct evaporative cooling of K atoms to BoseEinstein condensation. Phys. Rev. A 86, 033421 (2012).
 (33) Shapiro, B. Cold atoms in the presence of disorder. J. Phys. A: Math. Theor. 45, 143001 (2012).
SUPPLEMENTARY INFORMATION
Appendix A Characterization of the 3D disordered potential
In Extended Data Fig.2, we have fully characterized the two speckles beams generating the disordered potential. We now use those measurements to reconstruct the corresponding properties of the final 3D potential. When the beams cross at 90 and the interference pattern adds to the speckles modulation, the main axes of the problem are the ones in Fig.1. In the direction, interference fringes split the speckles into smaller substructures and the first zero in the correlation function is given by half the distance between two interference maxima, so that =0.09m. Along we calculate the average between the transverse correlation lengths of the two speckles =0.76m, while is the projection at 45 of the same average: =1.08m. From the geometric average on the three directions we get =0.41m.
The other main feature of the disordered potential is the intensity distribution . The field generated by two interfering speckles is predicted to preserve the same exponential distribution of the single speckles with (see Ref.[31]). An angle between the polarizations of the two speckles would reduce the contrast in the interference and hence modify at low . Considering the geometry of our experimental setup, we estimate . For small , the position of the maximum in is expected to move to , which in our case corresponds to . Such a small modification is not expected to affect localization properties in the system, since the typical energies in the experiment range from to , hence far from this low region.
Appendix B Background potential
In addition to the disordered speckle potential, the atoms are subjected to a weak additional optical and magnetic potential. We indeed use a verticallyoriented homogeneous magnetic field to control the interaction, while a magnetic field gradient compensates gravity. The two sets of magnetic field coils generate weak curvatures, which we partially compensate with a weakly focused laser beam in the vertical direction. The first nonnegligible terms of the resulting potential around the initial position of the atoms are HzHzHzm. Here the axes are the same as in Fig.1. The antitrapping curvature along is caused by nonperfect Helmoltz configuration of the Feshbach coils. In the direction an offcenter dipole trap cancels a magnetic field gradient along the same direction. The resulting potential has a cubic spatial dependence, flat around the atoms position to allow for a free expansion in the disorder. In the direction the same optical trap contributes to a weak trapping potential. By noticing that the typical energy scale for the system is of several tens on nK, we could define a spatial region in which the spurious fields stay below 5nK, so that if the system remains within this region, we can consider negligible the effect of the residual curvatures. The size of the region amounts to 144m along , 42m along and 112m along . In the fast diffusive regime (very small ), we indeed observe some deviations from the expected expansion when the cloud size approaches the region’s boundaries. We expect that the effect of such a background potential is minimal at the mobility edge, where the typical energies we explored in this work range from 20nK to 100nK. Note also that the potential corresponds to antitrapping in two directions, and to trapping in the third direction, suggesting that the net effect on the 3D problem is less than that in the individual directions.
Appendix C Modulation spectroscopy: the coupling coefficient
Using the numerical simulations, we try to estimate the dependence of the excitation probability on the initial energy of the atom , at least for low modulation frequencies (so as to stay in the lowenergy regime where the simulations are reliable). We find a typical linear scaling with and nK. If we consider this dependence in the calculations for , we find only a small shift of with respect to the one obtained for const. Actually, even a 3 times larger than the one we get in the simulations would not change by more than 2nK. This very weak dependence on the actual form of is due to the fact that the features of the spectra that are relevant to determine simply depend on the behavior of in a small range of energies, between and the peak in . We can therefore conclude that the approximation const provides reliable results for the mobility edge.