cond-mat Entanglement Entropy of Random Fractional Quantum Hall Systems B. A. Friedman, G. C. Levine* and D. Luna Department of Physics, Sam Houston State University, Huntsville TX 77341 and 1 *Department of Physics and Astronomy, 1 0 Hofstra University, Hempstead, NY 11549 2 n (Dated: February 1, 2011) a J Abstract 8 2 The entanglement entropy of the ν = 1/3 and ν = 5/2 quantum Hall states in the presence of ] l e short range random disorder has been calculated by direct diagonalization. A microscopic model - r t of electron-electron interaction is used, spin polarized electrons are confined to a single Landau s . t level and interact with long range Coulomb interaction. For very weak disorder, the values of a m the topological entanglement entropy are roughly consistent with expected theoretical results. By - d n considering a broader range of disorder strengths, the entanglement entropy was studied in an o c effort to detect quantum phase transitions. In particular, there is a signature of a transition as [ a function of the disorder strength for the ν = 5/2 state. Prospects for using the density matrix 4 v 2 renormalization group to compute the entanglement entropy for larger system sizes are discussed. 0 2 4 PACS numbers: 03.67.Mn,73.43.Cd,71.10.Pm . 7 0 0 1 : v i X r a 1 I. INTRODUCTION This paper isa numerical study, using direct diagonalization, of theentanglement entropy of fractional quantum Hall systems in the presence of a delta correlated random potential. The entanglement entropy, quite distinct from the thermodynamic entropy, is the Von Neu- mann entropy of the reduced density matrix of a subsystem and is a quantitative measure of the entanglement of the subsystem with the system. Our interest in this subject is two-fold; firstly, it has been proposed that entanglement entropy can be used as a tool to characterize fractional quantum Hall states. More precisely, Kitaev and Preskill1 and Levin and Wen2 have shown, for a topologically ordered state, that the entanglement entropy of a subsystem obeys an asymptotic relation 1 S αL γ +O( )+... (1) ≃ − L where L is the linear size of the subsystem (the area law ) and γ is a universal quantity, the topological entanglement entropy, the natural logarithm of the quantum dimension. For this scaling law to apply, the system must be very large and the subsystem must be large (compared to a cutoff, but the subsystem must be small compared to the system). This is a rather formidable numerical requirement, however, there has been some success numerically3–8 using (1) to extract the topological entanglement entropy of quantum Hall states. One may hope, that by adding weak randomness, there may be less system size dependence and hence it will be easier to obtain the topological entanglement entropy. Of course, by adding randomness, momentum conservation is destroyed and one cannot treat as large systems by direct diagonalization. In any case, it is of interest to see if the topological entanglement entropy can be calculated in the presence of weak disorder and to see if the values obtained are consistent with previous numerical estimates. Thesecondmotivationtoundertakethisstudy, istoseewhethertheentanglemententropy can be used to detect transitions between phases of quantum Hall systems. For example, experimentally, it iswell known thatfractional quantum Hall statesareparticularly sensitive todisorder. Canthissensitivity bedetectedintheentanglement entropy? Thetwoquestions discussed above will be studied for 2 filling factors ν = 1/3 in the lowest Landau level, representative ofLaughlinstates, andthe5/2thstateinthesecond Landaulevel. Currently, there is goodevidence bothexperimentally and numerically9 that the essential physics of the 2 5/2 state is given by the Moore-Read wave function and thus the 5/2 state is representative of the more exotic states with non abelian statistics. Thepaperisthenorganizedasfollows: inthesecondsection, themodelandthenumerical method are briefly described and the results for the topological entanglement entropy for weak disorder are discussed. In the third section, the entanglement entropy is calculated as a function of disorder strength for a wider range of disorder to determine whether transitions between phases of Hall systems can be detected. In the fourth section, some preliminary resultsusing thedensity matrixrenormalizationgrouptocalculatetheentanglement entropy are described. The fifth section is a summary and gives conclusions. In the final section, a recent alternative method23 to obtain the topological entanglement entropy on the torus is discussed. II. EXTRACTING THE TOPOLOGICAL ENTANGLEMENT ENTROPY FOR WEAK DISORDER The numerical method we have used is direct diagonalization applied to square (aspect ratio1)clusterswithperiodicboundaryconditions(thesquaretorusgeometry). TheLandau gauge is used for the vector potential. Spin polarized electrons are confined to a single Landau level and interact with a pure Coulomb potential. One can approach the limit of very large system sizes through clusters of any fixed aspect ratio and since we are concerned with quantum liquid states, aspect ratio one has been chosen. This numerical approach has previously been used to study the entanglement entropy without a disorder potential5,7. The random potential10 U(r) is taken to be delta correlated i.e. < U(r)U(r ) >= U δ(r r ) and ′ 0 ′ − the disorder strength will be given in terms of a parameter U = 3U /2. Since momentum R q 0 is not conserved, one is limited to smaller system sizes then for a disorder free system. In particular, the largest system size treated for ν = 1/3 is 10 electrons in 30 orbitals with a state space of approximately 30X106 and 14 electrons in 28 orbitals for ν = 5/2 with a state space of approximately 40X106. (This is in contrast to the disorder free case, ν = 1/3 13 electronsin39orbitals, andν = 5/218electronsin36orbitals, arerelativelystraightforward to treat). To calculate the entanglement entropy, we take a subsystem consisting of l adjacent orbitals (recall in the Landau gauge, these orbitals consist of strips oriented along, say the y- 3 axis, ofwidthoforderthemagneticlength). Thereduceddensitymatrixisstraightforwardto compute from the ground state wave function. It is then diagonalized giving the eigenvalues λ from which the l-orbital entanglement entropy S(l) , S(l) = λ lnλ is obtained. j j j j −P This procedure is done for every realization of the random potential, the results are then averaged to give < S(l) > where <> denotes average over the random potential. The position of the subsystem has been fixed, that is, for say S(l = 3) the subsystem always consists ofthe 1st , 2nd and3rdorbitals. Forthesmallest systems (6 electrons in 18orbitals) we have averaged over 1000 realizations of the random potential, for the largest systems we have averaged over asfew as10 realizations. This choice was dictated by thetime consuming nature of the larger calculations. In figure 1, we have plotted the entanglement entropy vs. square root of l for 10 electrons in 30 orbitals. The green circles are for no disorder ( an average is taken over the 3 ground states with k = 5,15,25), while the blue and red circles are for disorder strength U = 0.05 y R averaged over 10 and 100 samples respectively. The error bars are given by the root mean √<S(l)2> <S(l)>2 square values of S(l) i.e. σ = − with N the number of samples. From the √Ns 1 s − relation (1) we expect linear behavior vs. √l for subsystems small compared to the system size; this behavior is seen in figure 1. In particular, the linear regime is larger for the disordered case, indicating a smaller finite size effect for a given system size. This suggests a linear fit to the initial part of the S(l) vs. √l curve to obtain the topological entanglement entropy as the negative of the y-intercept. The results of the fit are plotted in figure 2 for U = 0.05. (the number of fitted values of l was chosen to give R a local maximum in the value of R2). The topological entanglement entropy γ , is found to be 1.10 .070 while a similar fit for U = 0.01 gives γ = 1.13 .078 both values in excellent R ± ± agreement with the value for the Laughlin 1/3 state of 2(ln√3) 1.10,26 the factor of 2 ≈ coming from the 2 boundaries of the subsystem. However, as will be discussed below, the excellent agreement may be fortuitous in that for the small system sizes considered γ tends to be overestimated at this filling. The dependence of the topological entanglement entropy on system size for filling 1/3 is shown in figure 3 for U = 0.01. In this figure γ is plotted vs 1/N (N=number of orbitals). R Clearly, it would be desirable even with disorder, to be able to treat larger system sizes. Another approach to obtain the topological entanglement entropy is, for a given < S(l) >, to do a linear extrapolation in 1/N yielding < S (l) >. < S (l) > is then plotted vs √l , a ∗ ∗ 4 linear least squaresfit isperformedandtheyintercept gives -γ. Aplotof< S (l) >vs. √l is ∗ shown in figure 4, for ν = 1/3 using systems with 21 to 30 orbitals to get the extrapolations. The negative of the y intercept is given by 1.30 with an error of 0.24 (the 0.24 due to the deviation of the fit from a line. The error in the extrapolations to get < S (l) > was not ∗ taken into account.) This method also gives a topological entanglement entropy consistent with the Laughlin 1/3 state. Turning now to the ν = 5/2 filling at U = 0.01, figure 5 shows γ calculated from the R initial linear part of < S(l) > (i.e. figure 2) vs. 1/N. For the largest system size, 14 electrons in 28 orbitals γ 1.5 considerably less then the γ expected from the Moore-Read state ( ≈ γ 2.08 for 2 boundaries26). If the 1/N dependence is fit by a line, one finds at N = M.R. ≈ ∞ a γ of 2.34 with an error of 0.08. However, without knowing the answer, one does not know to extrapolate in figure 5 but not to extrapolate in figure 3, where the largest system sizes give acceptable answers without extrapolation. In figure 6 < S (l) > vs √l is plotted for ∗ ν = 5/2 where < S (l) > was obtained through extrapolation of system sizes 24,26,28. The ∗ y-intercept of the linear fit gives a γ of 2.58. Of course, this is no great success, however one obtains better agreement with the expected value if the ratio γ5/2 is considered. Using γ1/3 the < S (l) > method (< S (l) > obtained from the 3 largest system sizes) γ5/2 1.97 ∗ ∗ γ1/3 ≈ compared to γM.R. 1.89. γLaughlin ≈ In any case, it appears the expected more generic behavior with weak randomness is unable to overcome the advantage of additional system sizes available to disorder free cal- culations. That is, by using the S*(l) method, and 2 more system sizes (without disorder), reference5 was able to get agreement, with the expected theoretical results, within the error bars, for both ν = 1/3 and ν = 5/2. Without disorder, fitting a line to the initial part of the curve S(l) vs. √l is problematical for ν = 1/3, since, in addition to a linear increase, there is superimposed oscillation (see the FSS (finite size scaling result) of fig. 1 (a) of ref.5). We suggest that a similar oscillation, though less pronounced when disorder is present, causes difficulty in extrapolation of the data presented in figure 3. To get the points in figure 3, for the smallest system size, 6 electrons in 18 states, the first 4 l-values were used to get the best linear fit (i.e. minimize R2) while for 10 electrons in 30 states, the 6 initial l values were used. From the extrapolated values in figure 4 the oscillation causes an overestimate of γ in both cases. On the other hand, for ν = 5/2 there is less oscillation in S (l), see figure ∗ 3a of reference5 and figure 6 of the present paper. To obtain the points in figure 5, for 10 5 electrons, 7 l-values were used, while for 14 electrons, 8-l values were used in the fit. Due to less oscillation and a greater number of l-values used it is perhaps not surprising that extrapolation of γ, obtained from the initial part of S(l), is more successful for ν = 5/2 then an extrapolation at ν = 1/3. III. ENTANGLEMENT ENTROPY AS A FUNCTION OF DISORDER STRENGTH Inthissection, theentanglemententropyisstudiedforawiderrangeofdisorderstrengths. Entanglement entropy hasbeenusedpreviously toinvestigate thephasediagramofquantum Hall systems as a function of interaction potential8, 1 dimensional quantum spin systems with disorder11, one particle entanglement entropy for Anderson transitions12 and to study the phase diagram of the 1 dimensional extended Hubbard model13 . In figure 7 a,b l- entanglement entropy < S(l) > vs U is plotted for filling 1/3. In figures 7a < S(2) > is R shown, this figure being representative of small subsystems l. In figures 7 b < S(12) > is graphed, this figure characteristic of larger subsystems. Both graphs show a strong decrease in the entropy with disorder, with S(12) exhibiting a slightly sharper decrease. Especially in the S(12) graph, the entropy appears to decrease and then level out at a disorder strength of approximately U = 0.25. Although it is hard to make a definite conclusion, this is at R least consistent with Wan et al.22 that sees a vanishing of the mobility gap for U > 0.25. R Let us now turn to filling 5/2. The same sequence of graphs is presented in figure 8 a,b. Here there appears to be, especially for the large l graph, fig 8b, a transition at a disorder strength U 0.04. A natural interpretation of these graphs is a quantum phase R ≈ transition from the Moore-Read state for disorder strength U 0.04. Previous numerical R ≈ studies5,14,27 indicate that the ground state for pure Coulomb potential (no disorder) is topologically equivalent to the Moore-Read state. We therefore suggest that the sharp drop off in figure 8a and particularly 8b as contrasted to the smoother curves in 7a and 7b is a transition due to the destruction of the Moore-Read state by disorder. A possible picture of this transition is the destruction of p-wave superconductivity of composite fermions29 by disorder. That such a transition should happen at rather weak disorder is physically appealing30. In an effort to characterize possible phase transitions with disorder, we have calculated 6 the variance < S(l)2 > < S(l) >2. In figures 9 and 10, the variance for l = 12 is plotted − for ν = 5/2 and ν = 1/3, respectively. For ν = 5/2, figure 9, the variance is nominal through the transition region (other then for the anomalous behavior of 10 electrons in 20 orbitals). In contrast, for ν = 1/3, figure 10, there is a general increase of the variance starting at U = 0.05 and reaching a plateau at U 0.2 0.25 which may indicate a transition in this R R ≈ − range, consistent with figure 7, and consistent with reference22. IV. PRELIMINARY DMRG STUDIES OF THE ENTANGLEMENT ENTROPY A common theme of the previous sections is the benefit of finding a method to access larger system sizes. A possible method to do this, for quantum Hall systems, is to use the density matrix renormalization group (dmrg)15. One expects, the number of states kept in the dmrg blocks, needs to scale as the exponential of the entanglement entropy of the block, for an accurate calculation. Since by the area law entropy scales as the √s where s is the number of sites in the block, the number of states kept needs to scale as ec√s. The bad news is that this depends on the exponential of the √s , however,the good news is that it does not depend on the exponential of s as in direct diagonalization. Hence, at least in principle, one should (if one can avoid being stuck in local minimum) be able to treat larger system sizes for quantum Hall systems by dmrg16–19. In particular, reference18 was able to accurately calculate ground state energies for ν = 1/3 for up to 20 electrons and up to 26 electrons for ν = 5/2 in the spherical geometry. In the spherical geometry 14 electrons at ν = 1/3 and 20 electrons at ν = 5/2 are accessible to direct diagonalization. However, the excitation gap, a more difficult numerical quantity at ν = 5/2 was only accurately calculable by dmrg for up to 22 electrons, 1 ”non-aliased” system size larger then that accessible to direct diagonalization. In this section, dmrg will be used to calculate the entanglement entropy for quantum Hall systems without disorder. We will be content, in this preliminary study, to use dmrg to study a large system size still accessible to direct diagonalization, that is, 12 electrons in 36 orbitals in the n=0 and n=1 Landau levels. In table I we display, the ground state energy vs. m, the number of states kept in the block; the first column is for the lowest Landau level, the second for the second Landau level. (the Madelung energy, which can be calculated exactly, is not included). One sees for the lowest Landau level a fairly accurate result can be obtained even without 7 TABLE I: Comparison of Dmrg and Direct Diagonalization Energies m N=0 N=1 200 -3.3675 -2.4109 300 -3.3691 400 -3.3699 -2.4178 600 -3.3717 -2.4203 700 -3.3723 800 -2.4219 Extrapolation -3.3739 .005 -2.4252 .003 ± ± Exact -3.3734 -2.4254 extrapolation, for the n=1 Landau level, extrapolation is more important (we extrapolate in 1/m). Let us now consider the calculation of S(l) (recall l is the number sites in the subsystem used when calculating the entanglement entropy) . All S(l)s are computed at the end of the calculation when the left and right blocks have equal number of sites (in addition, there are two sites in the middle17). This makes the calculations more complicated (i.e. clearly it is easier to get S(l) when there are l sites in the block) but it is necessary to get reliable results. Figure 11 is a plot of S(l) vs. √l up to l=8 for differing number of states in the blocks for 1/3 filling. One notices that even for the smallest block sizes, dmrg does a good job in computing S(l). This is consistent with the dmrg calculations of the entanglement entropy done by Shibata20. Turning now to 1/3 filling in the second Landau level (ν = 7/3), figure 12 is a plot of S(l) vs. √l for this filling. One againsees, that differing from the first Landau level, extrapolation is very important to obtain an accurate result. The larger l values are underestimated (i.e. entanglement is underestimated) particularly for calculations with smaller number of states in the blocks. Of course, the energy is also less accurately calculated in the second Landau level by dmrg. This is not the whole story, since the 800 state calculation in the second Landau level does better for the energy (on a relative basis) then the 200 state calculation in the first Landau level. However, the 200 state calculation still does better in calculating S(l). Even though it seems possible to use more states in the blocks (reference18 uses up to 5000) it appears to be difficult to go much beyond direct diagonalization in calculating the 8 entanglement entropy in the second Landau level. A simple estimate shows, based on the above calculations, why this is the case. The computation for ν = 7/3 indicates at least 1000 block states (and this may be an under estimate) are necessary to get a fairly accurate result. In going from 36 to 48 sites (12 to 16 electrons) the ”worst” block goes from 18 to 24 sites (1/2 the system size, since the entanglement entropy of the system and environment are equal). Assuming that the number of states kept needs to scale as ec√s, the number of states needed for 48 sites is at least 1000√24/18 3000 states. ≈ V. CONCLUSION The entanglement entropy of the ν = 1/3 and ν = 5/2 quantum Hall states in the presence of short range disorder has been calculated by direct diagonalization. For very weak disorder, the value of the topological entanglement entropy ( a universal quantity) is roughly consistent with the expected theoretical results and disorder free calculations. However, ( in particular for ν = 5/2) the advantages of having less system size dependence with weak disorder are outweighed by the disadvantage of the inaccessibility of larger system sizes. To investigate the possibility of using the entanglement entropy to detect quantum phase transitions, the entanglement entropy has been calculated for a broader range of disorder strength. For ν = 1/3 , the l-orbital entanglement entropy (figures 7a,b) shows a strong decrease, and the variance (figure 10) shows a strong increase through the range U 0.1 0.25. For the range of disorder considered and the amount of averaging done, R ≈ − we suggest that this is a possible signature of a phase transition similar to that observed for the mobility gap in reference22 at U 0.25. For ν = 5/2 we see a much sharper transition R ≈ feature in the l-orbital entanglement entropy (figures 8a,b) and at a much smaller value of the disorder strength, U 0.04. Despite the sharper transition, there is no corresponding R ≈ feature in the variance (figure 9), as there is in the ν = 1/3 case. The sensitivity of the 5/2 state to disorder is well known from experimental studies where samples must have a high (zero field) mobility to see an incompressible state. Thus there is qualitative agreement with experiment, taken with due caution in that a quantitative comparison likely requires considering longer range disorder. In our study, one number, the entanglement entropy has been used to characterize the reduced density matrix. There is possibly additional information in the full spectrum of the reduced density matrix14, which has been shown to 9 be related to the conformal field theory describing the one dimensional edge state of the quantum Hall state8,14,24. It would definitely be of interest11 to study the entanglement spectrum in the present system. Even if the topological entanglement entropy (derived from the entanglement entropy) is a complete invariant25, numerically it may well be easier to see transitions using theentire spectrum8,14. Finally, we have displayed some preliminary results using dmrg to compute the entanglement entropy. These results indicate dmrg holds some promise in calculating the entanglement entropy in the lowest Landau level; it appears more difficult to do calculations in the second Landau level and to go much beyond systems that one can treat by direct diagonalization. This may indicate that potentially more powerful numerical methods, for example, tensor network states21 or the methods of reference28, will prove useful. VI. FINAL REMARKS After this manuscript was posted at arXiv.org, we became aware of an interesting paper that calculates the topological entanglement entropy using a different method in the flat torus geometry. (We thank Dr. Haque for bringing this reference to our attention.) In essence, ref.23 , calculates the entanglement entropy S(N/2) taking the subsystem to be half the system size. The scaling law S(N/2) c N 2γ is then used where α is the ∼ 1qα − aspect ratio and N is the number of orbitals in the system; this approach was also used by Shibata20. InthemethoddescribedinsectionII(seealso5,7 )thescalinglawS(l) c √l 2γ 2 ∼ − is used where l the number of orbitals in the subsystem is much smaller then N. In this method, the subsystem for fixed l becomes increasingly thin since the number of states per unit length (the magnetic length) scales as √N. Let us examine this point23 in greater detail. Imagine there is a subsystem consisting of a fixed number of orbitals l and N becomes very large. Consider the square torus geometry, a ”box” of dimensions aXa; here a = √2πN. The width of a box of l orbitals is l √2πN N i.e. l 2π so the width goes to zero as 1. However, at the same time the width goes to qN qN zero, the length goes as √2πN. Although the width and length are both ”singular” as N goes to infinity, the area is perfectly well defined, 2πl (again in units of the magnetic length squared). Since the area law relates the entanglement entropy to a linear dimension of the subsystem, it is reasonable that S(l) scales as the square root of the area, S(l) c√l, and ∼ 10