Anderson Localization

Daniel Bruns, Rafael Haenel, and Gary Tom

(Dated: November 25, 2017)
Localization phenomena in disordered systems have attracted great theoretical interest since the
second half of the 20th century. In this article we give an introduction to the theory of Anderson-
localization. We perform numerical calculations of eigenstates of disordered tight-binding Hamilto-
nians in 1D and 3D and review recent experimental evidence of localization of bosonic wavefunctions
in a 1D-magneto-optical trap. Additionally, we provide a brief overview of the effect of Hubbard
electron-electron interactions and their implications on the Eigenstate Thermalization Hypothesis.

I. INTRODUCTION
105 200
W=1
In 1958, Philip Anderson published his seminal paper
100

Localization length
on how certain materials can undergo a phase transition
from conductor to insulator when the amount of disorder
in the system passes a certain threshold [1]. This sudden 0
-2 0 2
reduction in conduction is associated with the localiza-
tion of the electronic wavefunction; when the state no E
longer is extended, and instead is spatially localized. In E=0
1977, Anderson was awarded the Nobel Prize, in part for E=1
E=1.5
his work on localization phenomena.
Anderson localization, as it is now known, can be un- 100
derstood as an interference phenomenon. In the original 10-1 100 101
tight-binding model formulated by Anderson, electrons Disorder strength W
are able to tunnel between neighbouring lattice sites.
However, at high enough disorder in the lattice, the quan- FIG. 1. Log-log scale plot of localization length ξ in 1D as a
tum amplitudes associated with tunnelling paths cancel function of disorder W and energy E (inset). The data was
numerically calculated by solving Eq. (2) with the transfer-
each other, resulting in a localized wavefunction. An
matrix method. ξ is well fitted by ξ = αW −2.0 for all E and
equivalent wave model has been proposed, in which the is smallest at the edge of the energy band.
incoming wave is scattered off of potentials generated by
the disorder. At a high enough disorder, the scattered
wavelets interfere destructively in the forward direction, Bloch-states with energy dispersion E(k) P = 2t cos(ka).
causing the wave to decay exponentially [2]. If we have finite W , the eigenstates |Ψi = n ψ(n)ĉ†n |0i
Since his original paper, advances in computing tech- are determined from the recursion relation
nologies have allowed for precise numerical studies of An-
derson localization. Experimental efforts have largely Eψ(n) = n ψ(n) + ψ(n + 1) + ψ(n − 1). (2)
confirmed Anderson localization in disordered materials,
but direct observation of localized wavefunctions has only This may be rewritten using transfer-matrices Ti :
been done in 1D [3].
  L    
ψ(n + 1) Y ψ(1) E − i −1
= T̂i , T̂i = .
II. LOCALIZATION IN ONE DIMENSION
ψ(n) ψ(0) 1 0
i=1

We consider the Anderson Hamiltonian [1], a tight- By proving Fuerstenberg’s theorem for products of ran-
binding model with nearest-neighbor hopping and ran- dom matrices, it can be rigorously shown that solu-
dom on-site energy: tions ψ(n) of Eq. (2) decay exponentially for n → ∞
[4, 5], so that the overall shape of the wavefunction is
ψ(n) ∼ exp(−n/ξ). Here, ξ is the localization length in
X X
n ĉ†n ĉn + t ĉ†n ĉm + h.c. .


Ĥ = W (1)
n <n,m>
units of lattice constant a. As a consequence, eigenstates
are not extended over real-space but are exponentially
The on-site energy n ∈ [−1/2, 1/2] is usually drawn from confined to certain areas of the lattice. We will refer
a uniform random distribution and W is the disorder to these states as localized. Remarkably, we note that
parameter. In the future we set t = 1, thus effectively any state in a one dimensional system will be localized.
measuring all energies in units of t. This implies that any 1D material will lose its diffusion
To draw some analytical conclusions, let us first con- or conduction capabilities if the system size is L > ξ.
sider a one dimensional chain of arbitrary length L with We can also solve Eq. (2) numerically. For this,
lattice spacing a. In the limit W = 0, the solutions are we apply 106 transfer matrices to an initial vector
 2

(ψ(1), ψ(0)) = (1, exp(ik)), where k is the momentum 1

of a tightly bound electron with a given energy E. The d=3
resulting coefficients ψ(n) are averaged from 100 calcula-
tions with different realizations of randomness. We then 0
d=2
extract the localization length ξ by exponentially fitting

(g)
the coefficients. The results are displayed in Fig. 1. The -1
disorder dependence of the localization is well described d=1
by the power law ξ ∝ W −2.0 . Moreover, states at the
-2
edges of the energy band localize fastest.

-3
III. SCALING THEORY 10-1 100 101
Conductance g
To assess localization properties in higher dimensions,
we will examine the effects of scaling on the conductance FIG. 2. Scaling function β for 1-3 dimensions. For d < 3 the
of a system. These scaling arguments were first presented conductance always flows to the stable fixed point at g = 0,
by Abrahams, Anderson, Licciardello, and Ramakrish- so that the system will be an insulator in the thermodynamic
nan, after the concept of Renormalization Group was in- limit. Only for d = 3 is there is an unstable fixed point when
β(gc ) = 0. Here, gc marks a metal insulator phase transition.
troduced [6].
Let us define the dimensionless conductance for a d-
dimensional cube of size Ld
IV. NUMERICAL ANALYSIS
T (L) OF THE ANDERSON MODEL IN 3D
g(L) = , (3)
1 − T (L)
where T (L) is the transmission in one direction. Intu- In Anderson’s tight-binding Hamiltonian (1), the
itively, T scales as T (bL) = T (L)b . Considering that the strength of disorder is measured by the width W of
cross-section perpendicular to the direction of transmis- the on-site energy probability distribution. Following
sion increases with b(d−1) , we obtain an expression for Eq. (7), we expect that, once W surpasses a critical value
the new conductance Wc , electronic eigenfuntions localize in 3D which leads
1 1 to a MIT. Analytical estimation of the critical disorder,
g(bL) = −b
= . (4) however, is problematic. For a uniform box distribution
1 − T (L) 1 − (1 + g −1 )b
over [−W/2, W/2], numerical analysis of the 3D Ander-
In the last equality, we have inverted Eq. (3) to replace T . son Hamiltonian shows that Wc ≈ 16.5 [10]. Since Wc is
We can now calculate the β-function that characterizes not small compared to the relevant energy scale given by
the response of g to changes in system size L. the bandwidth of delocalized electrons in a perfectly or-
d ln g dered 3D crystal (−6 ≤ E ≤ 6), perturbative approaches
= (d − 1) − (1 + g) ln 1 + g −1 .


β= (5) to determine Wc fail. Nonetheless, it is straightforward
d ln L
to demonstrate the disorder-induced Anderson transition
For g  1, β is dominated by the logarithmic term and in 3D numerically.
simplifies to ln(g). For large g, it asymptotically ap-
proaches the value d − 2. A plot of β is shown in Fig. 2.
In one or two dimensions, the β-function is always neg- A. Eigenstates
ative and, under scaling of the system, the conductance
will always decrease, i.e. flow to the stable fixed point at In order to simulate the Anderson transition in
g = 0. Hence, the system will be a perfect insulator in three dimensions, we diagonalize the Anderson Hamil-
the thermodynamic limit. We conclude that for d = 1, 2 tonian (1) for a single particle whose eigenstates |ψi =
all states in disordered systems are localized. P †
r ψ(r)cr |0i are defined on a cubic lattice given by
In three dimensions, however, there is an unstable fixed the set of vectors r = nx ex + ny ey + nz ez , with in-
point at a critical conductance defined by β(gc ) = 0. tegers 1 ≤ nx , ny , nz ≤ N . In this case, the resulting
Consequently, gc marks a metal-insulator phase transi- Schroedinger equation to be solved reads
tion (MIT). From Renormalization Group theory [7], we
obtain the critical exponent of this transition with W r ψ(r) +
X
ψ(r + eα ) + ψ(r − eα ) = Eψ(r), (8)
−1
ν = (gc β 0 (gc )) = 1.68, (6) α=x,y,z

so that the localization length behaves as with disorder strength W and uniformly distributed ran-
dom on-site energies, r ∈ [−1/2, +1/2].
ξ ∝ (W − Wc )−ν (7)
Applying periodic boundary conditions along all spa-
in the vicinity of the critical disorder with W < Wc [6, tial directions, ψ(r + N eα ) = ψ(r), Fig. 3 shows elec-
8, 9]. tronic eigenstates near the band center E ≈ 0 for three
 3

FIG. 3. Anderson localization in 3D of a single-particle inside a 17×17×17 cube. Plotted are eigenfunctions of the Schroedinger
Eq. (8) as a function of disorder W ∈ {2, 10, 17}. The size of colored spheres indicates the magnitude of probability density
|ψ(r)|2 at each grid point. Blue spheres display points where |ψ(r)|2 ≤ 0.003, yellow spheres show 0.003 < |ψ(r)|2 ≤ 0.02,
and red spheres correspond to 0.02 < |ψ(r)|2 . Radii in the localized regime W = 17 are plotted six times smaller than in the
delocalized regime W = 2(10).

disorder realizations of different strength. For weak dis-

order W = 2, the solution resembles an evenly spread
probability distribution. This result is not surprising be-
cause in the limit W = 0 we expect the eigenstates to
be Bloch states, such that |ψ(r)|2 = 1/N 3 equally for all
lattice sites. As we go to stronger disorder W = 10, first
signatures of localization become evident. Yet, we ob-
serve that the eigenstate still spreads over the whole cube
and therefore we still ought to be in the metallic regime.
Once we reach sufficiently strong disorder W = 17 which
is above the critical value, Wc = 16.5, the system un-
dergoes a MIT. In this regime, we can clearly make out
distinct lattice sites where the probability density |ψ(r)|2
differs by at least one order of magnitude.

B. Phase diagram
FIG. 4. Metal-insulator transition of the 3D Anderson model.
To give a more general picture of the 3D Anderson The color map represents the localization strength of elec-
tronic eigenstates depending on energy E and disorder W as
transition shown in Fig. 3, one can characterize eigen-
given by the inverse participation ratio I, cf. Eq. (9). Red
states of Eq. (8) by use of the inverse participation ratio indicates stronly localized states, whereas light blue shows
(IPR), completely delocalized states. Energy regions containing no
X eigenstates are colored dark blue. Disorder varies from W = 0
I= |ψ(r)|4 . (9) to W = 20 with spacing ∆W = 0.1.
r

For delocalized states spread uniformly over all lattice

sites, the IPR is on the order of I ∼ 1/N 3 . Con- strength W . Eigenstate energies are grouped into inter-
versely, localized states are expected to occupy only a vals of width ∆E = 0.1.
few sites and to have an exponentially small probability Figure 4 confirms that, for weakly disordered systems,
amplitude elsewhere. The IPR projects out those lattice eigenstates across the whole band are delocalized with
sites with large probability amplitude and, consequently, I ∼ 1/8000 ∼ 10−4 . As we approach the critical disorder
yields I  1/N 3 for localized states. In Fig. (4), we parameter Wc = 16.5 [10], I increases by three orders of
choose a cubic lattice with N 3 = 203 = 8000 sites and magnitude which demonstrates the expected transition
compute I as a function of disorder W and band en- from the metallic to the insulating regime. Furthermore,
ergy E. Since the form of eigenstates fed into Eq. (9) the numerical data shows that localization emerges in
depends on the particular disorder realization, we av- the energy band edges already below the critical disorder
erage the IPR over 100 random realizations for a fixed strength, W < Wc . In the context of Anderson local-
 4

ization, the threshold |E| < Ec that denotes a MIT for B. 1D Localization Length in Speckle Potential
a fixed strength of disorder is commonly referred to as
mobility edge. A 3D system of noninteracting electrons The localization length of the 1D system in a speckle
whose Fermi energy is smaller than the mobility edge, potential can be theoretically predicted. First, consider
EF < Ec , behaves as an insulator. a trapped 1D Bose-Einstein condensate (BEC) in a har-
monic potential Vho = mω 2 z 2 /2. At some large time
after release t  1/ω, atom-atom interactions are negli-
V. EXPERIMENTAL EVIDENCE gible, and the wavefunction becomes [13]
!
Early efforts to confirm the existence of localized states ψ(z/b(t), 0) imz 2 ḃ(t)
ψ(z, t) = p exp , (10)
focused on conductivity measurements around the MIT b(t) 2~b(t)
of disordered materials. More recent work around 1990s
has experimentally demonstrated Anderson localization where √b(t) is a scaling parameter, with b(0) = 1, and
in a variety of classical and quantum materials, affecting b(t) ≈ 2ωt for large times.
light, acoustic, and matter waves [3]. Here, we will fo- Now let the speckle potential be turned on after re-
cus on experimental evidence of Anderson localization in lease. The wavefunction is a superposition of plane-waves
matter waves of 1D Bose-Einstein condensates (BEC). Z
dk
ψ(z, t) = √ ψ̂(k, t) exp(ikz). (11)
2π
A. Experimental Setup According to Anderson [1], the exponential components
of Eq. (11) will become localized functions φk (z) due to
The direct observation of Anderson localization is in the speckle potential. The atomic density of the system
matter waves of 1D rubidium-87 BEC [11]. The BEC would then be
is first created in a magneto-optical trap, which employs Z kmax
Doppler cooling on atoms inside a magnetic trap. Evapo- n0 (z) = h|ψ(z)|2 i = dkD(k)h|φk (z)|2 i, (12)
rative cooling, in which higher energy atoms are “kicked- −kmax
out” of the magnetic trap using a RF signal, is used to
lower the temperature to ∼ 1µK [12]. The BEC is then where D(k) = |ψ̂(k)|2 is the momentum distribution of
confined to the transverse z-direction the an optical trap. the system, which is constant for long times t  1/ω,
A weak magnetic trap is used to confine the BEC in the as seen in the Fourier transform of Eq. (10). These lo-
longitudinal direction (Fig. 5). calized functions can be obtained using diagrammatic
perturbation theory, assuming a weak speckle potential
2 √
(VR σR  ~mk kσR ), and distances far from the localized
state (|z|  ξ), we have [14]

h|φk (z)|2 i ∝ e−2|z|/ξ , (13)

4 2
2~ k
ξ≈ . (14)
πm2 VR2 σR (1 − kσR )Θ(1 − kσR )

where VR and σR are the potential amplitude and corre-

lation length of the speckle, respectively, Θ is the Heav-
iside function, and ξ(k) is a length scale defined as the
localization length. See that ξ ∝ VR−2 , as expected in
the transfer-matrix calculation (cf. Fig. 1). Since ξ is
an increasing function of k, the asymptotic behaviour of
FIG. 5. Experimental setup of [2]. (a) A transverse optical the localized function at large |z| is dominated by the
trap is used to confine the BEC to the x and y-direction. A largest k values. Thus we expect localized states to fall
shallow longitudinal magnetic trap is used to hold the BEC exponentially at the tails
before release. (b) After release, the BEC is allowed to evolve
in the speckle potential, and the atomic density is measured. n0 (z) ∝ exp(−|z|/ξ(kmax )). (15)

In order to observe Anderson localization, a disorder

potential is generated using 1D speckle created by a laser C. 1D Anderson Localization of BEC
beam sent through a diffusing plate. The longitudinal
magnetic trap is then released, and the BEC is allowed Prior attempts of this experiment have demonstrated
to evolve in the potential. The atomic density is observed reduced transport in the BEC, however this was at-
through absorption and fluorescence imaging [11]. tributed to scattering and inter-atomic interactions [15,
 5

tion, the localization lengths from the fits should agree

with the values of ξ(kmax ) of Eq. (14), where kmax is
obtained by allowing the BEC to expand in the wave
guide with VR = 0. Indeed the results are in agreement,
demonstrating that 1D Anderson localization is possible,
and measurable (Fig. 7).

D. Experimental Evidence in Higher Dimensions

Experiments looking at higher dimensions of Anderson

localization have been done, and compared with theoret-
ical results of scaling theory. For example, in 2D, Ander-
son localization of light waves has been observed, with
specially engineered optical fibres acting as “photonic lat-
tices”. Materials with adjustable indices of refraction are
used to vary the level of disorder in the system.
FIG. 6. The atomic density profile of the BEC at t = 1s [11].
The pink profile shows the condensate trapped at t = 0s. The For localization in 3D, experiments have mainly fo-
blue lines are the exponential fits to the tails. The inset shows cused on electronic transport, the original field from
the rms width of the profile as a function of time for VR = 0 which Anderson derived his theory. While MITs are ob-
and VR 6= 0, demonstrating the localization of the state. The served in conductive materials, the critical exponent of
red diamond corresponds to the point at which the profile is the phase transition was found to be close to unity, in-
taken. compatible with the predicted value of ν = 1.68 in Eq.
(6) [17]. Electron-electron interactions were hypothesized
to be the cause of the disparity.
There have additionally been experiments involving
other types of waves, i.e. electromagnetic or ultrasonic,
that demonstrated transmission levels consistent with
those predicted by scaling theory.

VI. CONCLUSION

In the present work, we have explored the effects of

disorder in a tight binding model. In one dimension,
an arbitrarily small random on-site energy localizes all
eigenstates, so that the wavefunctions are confined to a
specific region of the lattice and the system is charac-
terized by an absence of diffusion in the thermodynamic
limit. Using scaling arguments, we have extended this
FIG. 7. Localization length as a function of disorder potential result to two dimensions. Only for d ≥ 3, there is a
from the experimental fit compared with theoretical expres- second-order metal insulator phase transition. We have
sion [11]. The blue points are the fit results; the error bars numerically calculated a phase diagram for this transi-
represent 95% confidence interval for fit values. The dash-
tion by exact diagonalization of the 3D-Hamiltonian of a
dot line represents the theoretical result in Eq. (14), with the
shaded area being uncertainty evaluations of kmax .
cubic lattice with 203 sites.
Finally, we have reviewed recent experimental evidence
of localization of a one dimensional Bose-Einstein con-
densate in an optical trap. Here, direct imaging of the
16]. To avoid this, the speckle amplitude is small to en- wavefunction reveals localization in good agreement to
sure localization is not due to scattering off of the po- numerical and analytical predictions.
tential, and the initial atomic density was reduced to
minimize atom-atom interactions [11]. At these limits,
the above Eqs. (13) and (14) are valid. VII. ACKNOWLEDGEMENTS
The atomic densities are measured at different times
as the matter wave evolves through the potential, shown We would like to thank Mona Berciu for the assignment
in Fig. 6. The tails were found to be best modelled with of this project. Computations were facilitated by high-
an exponential fit of the form of Eq. (15). performance computing resources of the UBC Stewart
If the localized state is a result of Anderson localiza- Blusson Quantum Matter Institute.
 6

[1] P. W. Anderson, Physical Review 109, 1492 (1958). Appendix A: Electron-Electron Interactions
[2] A. Aspect and M. Inguscio, Phys. Today 62, 30 (2009).
[3] A. Lagendijk, B. van Tiggelen, and D. S. Wiersma,
In the previous part, we have neglected electron-
Physics Today 62, 24 (2009).
[4] H. Furstenberg and H. Kesten, The Annals of Mathemat- electron interactions. This approximation is only rea-
ical Statistics 31, 457 (1960). sonable for systems that behave ideally, such as BEC.
[5] K. Ishii, Progress of Theoretical Physics Supplement 53, For an improved theoretical description, we will add an
77 (1973). on-site repulsion term to our Anderson Hamiltonian (1)
[6] E. Abrahams, P. W. Anderson, D. C. Licciardello, and and obtain the Bose-Hubbard Hamiltonian with random
T. V. Ramakrishnan, Physical Review Letters 42, 673 on-site potential
(1979).
[7] M. Kardar, Statistical physics of fields, 6th ed. (Cam- X X
n ĉ†n ĉn + t ĉ†n ĉm + h.c.


bridge University Press, Cambridge [etc.], 2013). Ĥ = W
[8] R. Beig, G. Englert, U. Frisch, P. Hänggi, K. Hepp, n <n,m>
W. Hillebrandt, D. Imboden, R. L. Jaffe, R. Lipowsky, UX
H. v. Löhneysen, I. Ojima, D. Sornette, S. Theisen, + n̂n (n̂n − 1). (A1)
2 n
W. Weise, J. Wess, J. Zittartz, T. Brandes, and S. Ket-
temann, Anderson Localization and Its Ramifications,
Vol. 630 (Springer Berlin Heidelberg, Berlin, Heidelberg, The study of localization phenomena of interacting sys-
2003). tems defines the area of many-body localization (MBL).
[9] Cord A. Mueller, Dominique Delande, “Disorder and in- This is an area of active research with many questions
terference: localization phenomena,” (2016). to be answered [18, 19]. In the following, we will discuss
[10] T. Brandes and S. Kettemann, Anderson Localization the single most notable implication of MBL - the failure
and Its Ramifications: Disorder, Phase Coherence, and of localized systems to thermalize.
Electron Correlations, Lecture Notes in Physics (Springer
Berlin Heidelberg, 2003).
[11] J. Billy, V. Josse, Z. Zuo, A. Bernard, B. Hambrecht,
P. Lugan, D. Clément, L. Sanchez-Palencia, P. Bouyer, 1. Eigenstate Thermalization Hypothesis (ETH)
and A. Aspect, Nature 453, 891 (2008), arXiv:0804.1621.
[12] W. Guerin, J. F. Riou, J. P. Gaebler, V. Josse, P. Bouyer, We consider an isolated quantum system that we parti-
and A. Aspect, Physical Review Letters 97 (2006), tion in parts A and B. Originally, this system is prepared
10.1103/PhysRevLett.97.200402, arXiv:0607438 [cond- in state
mat].
[13] Y. Kagan, E. L. Surkov, and G. V. Shlyapnikov, Physical X
|ψi = cab |aiA ⊗ |biB .
Review A 54, R1753 (1996), arXiv:9606001 [atom-ph].
ab
[14] L. Sanchez-Palencia, D. Clément, P. Lugan, P. Bouyer,
G. V. Shlyapnikov, and A. Aspect, Physical Re-
view Letters 98 (2007), 10.1103/PhysRevLett.98.210401,
We can readily write down the density matrix for sub-
arXiv:0612670 [cond-mat]. system A:
[15] D. Clément, A. F. Varon, M. Hugbart, J. A. Retter,
c∗0b c1b . . .
 
P. Bouyer, L. Sanchez-Palencia, D. M. Gangardt, G. V. |c0b |
Shlyapnikov, and A. Aspect, Physical Review Letters ρA = trB (|Ψi hΨ|) =
X .. .. .

 . .
95, 170409 (2005). b ∗
[16] C. Fort, L. Fallani, V. Guarrera, J. Lye, M. Modugno, cN b c c0b |cN b |
D. Wiersma, and M. Inguscio, Physical review letters (A2)
95, 170410 (2005). The Eigenstate Thermalization Hypothesis now states
[17] S. Katsumoto, F. Komori, N. Sano, and S.-i. Kobayashi, [20–22], that if
Journal of the Physical Society of Japan 56, 2259 (1987).
[18] R. Nandkishore and D. A. Huse, Annual Review of Con- 1. the diagonal terms of ρA are a smooth function of
densed Matter Physics 6, 15 (2015). energy E of state |Ψi
[19] D. A. Abanin and Z. Papić, Annalen der Physik 529,
1700169 (2017). 2. the off-diagonal terms decohere (i.e. interfere de-
[20] J. M. Deutsch, Physical Review A 43, 2046 (1991). structively since they are the sum of many oscillat-
[21] M. Srednicki, Physical Review E 50, 888 (1994). ing terms of different frequencies)
[22] M. Srednicki, “Does quantum chaos explain quantum sta-
tistical mechanics?” . - then, the reduced density matrix will be equal to a
[23] R. Islam, R. Ma, P. M. Preiss, M. E. Tai, A. Lukin,
statistical density matrix (i.e. canonical) after some eu-
M. Rispoli, and M. Greiner, Nature 528, 77 (2015).
[24] A. M. Kaufman, M. E. Tai, A. Lukin, M. Rispoli,
qilibration time. We can write this as
R. Schittko, P. M. Preiss, and M. Greiner, Science (New  
York, N.Y.) 353, 794 (2016). ρA = ρth
A = exp −β ĤA . (A3)
 7

5
0.7 W=20
2
W=0

S
4 1 0.6
0
0 2 4 6 0.5
Renyi entropy S

System size

Trace difference
3
0.4

2 0.3

0.2
1
0.1

0 0
0 2 4 6 8 10 0 1 2 3 4
Time t Time t

FIG. 8. (left) Time evolution of entanglement entropy for subsystems of size 1-3. The initial state |Ψ0 i is not entangled. Renyi
entropy shows an initial increase and then fluctuates around a saturated value. The inset plots the average saturated value over
system size. For small system sizes, the entropy follows a volume law (red dashed line). (middle) On-site statistics for a single
site at t = 10. Blue bars display the probability P (N ) of occupation number N . Red marks are predictions of a canonical
ensemble. The two
values show reasonable agreement, indicating that the system has thermalized. (right) Trace difference
∆ = tr ρA − ρthA
 for disorder W = 0, 20. For W = 0, ∆ decreases and then fluctuates at low values around ∼ 0.1, while for
W = 20 the saturation value is ∼ 0.4. Thus, for high disorder, the quantum mechanical description differs substantially from
the canonical prediction, indicating that the system is many-body localized.

The effective temperature β in the canonical description of 1 (blue), 2 (red), and 3 (green) sites. After an ini-
is obtained by requiring that the expectation values of tial growth, the entanglement entropy fluctuates around
the energy are the same: a saturated value. Fluctuations are higher for larger sub-
systems.
hΨ(0)| Ĥ |Ψ(0)i = tr(ρth Ĥ). As a consequence of the ETH, the Renyi entropy should
be equal to a statistical entropy. Then, since entropy is
Simply put, this means: A small isolated quantum sys- an extensive quantity, S should obey a volume law. To
tem may act as its own bath and thermalize. prove this requirement, we have plotted the saturated
This result may seem astonishing to the reader, and we value of the entropy as a function of system size in the
will not give further justification of this hypothesis and inset of Fig. A3. As expected, the entropy S follows a
refer to references [18, 21] for detailed discussion. We volume law (red dashed line) for small system sizes and
note however, that this hypothesis - if true - is of great then bends back due to symmetry.
benefit. The numerical calculation of density matrices Another way to observe the effects of thermalization, is
for isolated quantum systems is only possible for fairly to compare the on-site statistics predicted by exact diag-
small system sizes while the statistical treatment does onalization and the canonical density matrix. Blue bars
not require much effort for arbitrary dimension of the in the middle panel of Fig. A3 show probabilities P (N ) of
Hilbert-space. occupation number N for a one-site subsystem at t = 10.
To test the ETH, we simulate a bosonic chain of six Clearly, these values closely resemble the canonical pre-
sites at half filling with the Hamiltonian (A1). For dictions marked by red points.
now, all n = 0. We time involve the initial state Last, we directly measure the differences of density ma-
|Ψ(t = 0)i = |111111i. This, of course, is not an eigen- trices ρA and ρth A
by computing the trace difference
state and has an energy far from the ground level. All ∆ = tr ρA − ρth A . The red curve in the right panel

results are obtained by exact diagonalization of the 462- of Fig. A3 shows ∆ as a function of time. The ini-
dimensional Hilbert space. tially high trace difference decreases to low values around
The ETH requires off-diagonal terms in Eq. (A2) to de- ∼ 0.1. This is clear evidence of the validity of the ETH
cohere. This process will likely cause entanglement be- in the Bose-Hubbard model.
tween the subsystems, since ρ = ρ2 only holds for a di- At this point, we would like to note, that recent advances
agonal matrix with ρn,n = δn,k . Thus, we expect time in experimental techniques allow for accurate experimen-
evolution of entanglement to give valuable insight into tal measurements of all quantities in Fig. A3. The reader
the thermalization

process. We define the Renyi entropy may be referred to references [23, 24].
S = −tr ρ2 as a measure of entanglement.
The left graph of Fig. A3 shows the Renyi entropy S as a
function of time for three different subsystems consisting
 8

2. Many-Body Localization (MBL) the right panelof Fig. A3. The high values of ∼ 0.4
compared to the case W = 0 suggest that disorder has in
fact prevented the system from thermalization. This is a
strong indicator of MBL. Note that this analysis is only
Let us now introduce some disorder. We diagonalize meant to be of qualitative nature. For a more rigorous
the Hubbard-Hamiltonian for a disorder parameter W = argument, the trace difference ∆ needs to be averaged
20 and plot the trace difference ∆ as the blue curve in over many different realizations of random on-site energy.