Strong-Disorder Renormalization Group Approach To The Anderson Model Using Raleigh-Schr Odinger Perturbation Theory
Strong-Disorder Renormalization Group Approach To The Anderson Model Using Raleigh-Schr Odinger Perturbation Theory
Strong-Disorder Renormalization Group Approach To The Anderson Model Using Raleigh-Schr Odinger Perturbation Theory
Abstract
Previous work proposed a strong-disorder renormalization approach for the Anderson model,
using it to calculate the density of states and the inverse participation ratio [Johri & Bhatt, Phys.
Rev. B 90 060205(R) (2014)]. This is interesting because of the potential for expansion to higher
dimensions and to interacting systems. The original proposal used a non-standard perturbation
theory approach which avoided degeneracies. We implemented the same structure but with stan-
dard Rayleigh-Schrödinger perturbation theory. Here degeneracies do arise, and we consider two
approaches, one in which renormalization is suppressed if degeneracy is present and a second in
which the most common form of degeneracy is handled using standard degenerate perturbation the-
ory. The version in which degeneracies are not handled performs similarly to the original proposal,
and the addition of degeneracies provides some improvement.
1
I. INTRODUCTION
2
Section II describes the method, including reviewing the original proposal and specifying
three versions we have implemented. Section III presents our results and discusses them
relative to several points of comparison.
II. METHOD
This section begins with a description of the model studied and the quantities calculated.
It then presents the renormalization procedure, reviewing the approach taken in JB and
highlighting where it diverges from standard perturbation theory. Two alternative solution
methods used as points of comparison are also described.
The Anderson model is a tight-binding model for non-interacting electrons with diagonal
disorder.
X X
H = Ei c†i ci + ti,i+1 (c†i ci+1 + c†i+1 ci ) (1)
i i
where c†i and ci are the creation and annihilation operators for site i, ti,i+1 is the hopping
amplitude between nearest neighbor sites (initially ti,i+1 = −1 ∀ i), and the site potentials Ei
are chosen from a flat distribution of width W . We focus here on one-dimensional systems
with periodic boundary conditions.
The main quantities calculated are the density of states (DOS) and the inverse participa-
tion ratio (IPR). The DOS is the probability distribution of the single-particle energies, and
the IPR is a measure of the localization of the single-particle states. Each single-particle
state may be expressed in the basis of Wannier states |ii ≡ c†i |vacuumi. For a particular
P
state |ψα i = i wα,i |ii, the IPR
|wα,i |4
P
Iα = P i . (2)
( i |wα,i |2 )2
The IPR as a function of energy for an ensemble of systems is generated by assigning Iα of
each state to a bin according to the energy of the state Eα , and then averaging the values
in each bin.
The renormalization procedure begins by evaluating, for each pair of adjacent sites, the
ratio of hopping amplitude to the difference in site potentials, mi,i+1 ≡ ti,i+1 /|Ei − Ei+1 |,
referred to as the bond strength. These bonds are then sorted into weak and strong, accord-
ing to whether they are above or below a chosen cutoff, m0 . The system is then divided into
3
clusters, where a cluster is defined as a collection of adjacent sites all connected by strong
bonds and linked to the rest of the system at each end by a weak bond.
Regarding the choice of the bond cutoff, m0 , a basic starting point is that it should be
less than one. Beyond this, smaller values generate larger clusters and correspondingly more
accurate results, while larger values generate smaller clusters and hence faster results.
As a point of comparison for the renormalization results, we also consider an approach
which stops here: The weak bonds are set to zero resulting in an ensemble of independent
clusters, and each cluster is then solved exactly.
In the renormalization, the next step is to identify the single-site cluster which is least
strongly coupled to its neighbors, specifically that for which the connectivity m2i,i+1 +m2i−1,i is
minimum. The state |Ci on the identified site and its associated energy EC are renormalized
by treating the couplings tL,C and tC,R to the left and right neighbor sites |Li and |Ri as a
perturbation using standard time-independent perturbation theory to first order in in t/∆E.
tL,C tC,R
|C 0 i = |Ci − |Li − |Ri (3)
EL − EC ER − EC
t2L,C t2C,R
EC0 = EC − − (4)
EL − EC ER − EC
The energy EC0 and the IPR I of |C 0 i are recorded, and the site is removed from the system.
The renormalized system includes the left and right neighbor sites similarly perturbed by
their coupling to the removed site and an effective coupling between them.
tL,C
|L0 i = |Li − |Ci (5)
EC − EL
t2L,C
EL0 = EL − (6)
EC − EL
tC,R
|R0 i = |Ri − |Ci (7)
EC − ER
t2C,R
ER0 = ER − (8)
EC − ER
tLR = hL |H|R0 i
0
1 1
= −tL,C tC,R + (9)
EC − EL EC − ER
This process continues until all one-site clusters have been removed. Up to this point, all
versions we have implemented are identical and the same as that in JB.
The next step is to proceed to clusters of two sites and then three sites, etc. Fig. (1) shows
a schematic diagram of the process for the case of a two-site cluster. For a cluster of n sites
4
FIG. 1: Schematic diagram of the procedure for the case of a two-site cluster. The cluster sites C1
and C2 are connected by a strong bond, with weak bonds to the left L and right R sites respectively.
The left and right states are renormalized, becoming L0 and R0 now connected by a new effective
coupling tLR .
Pn
(C1 to Cn ), the exact eigenstates |ψq i = j=1 cq,j |Cj i are determined and then, as for the
single-site case, these are renormalized by treating the hopping to the left and right neighbor
sites as a perturbation. It is at this point that the Rayleigh-Schrödinger approach diverges
from JB. Table I lists the equations for renormalization of the cluster states and energies as
well as the neighbor states and energies. The center column has the equations used by JB
and the right column shows the equations obtained from a standard Rayleigh-Schrödinger
calculation. We implement all renormalization approaches to linear order in t/∆E, and at
this order the basis remains orthogonal (see B). Weights on the original basis are used when
calculating IPR values.
Both sets of equations in Table I reduce to Eqns. (3)-(9) in the case n = 1. Nonetheless,
the two sets of equations differ significantly. Two differences in particular stand out. First,
the denominators: In both cases, the denominators contain the difference between the energy
of a neighboring site (EL or ER ) and an energy associated with the cluster. In the Rayleigh-
Schrödinger case, this second energy is a cluster-state energy Eq , whereas in the JB case this
second energy is the energy of the adjacent end-site in the cluster (E1 or En ). Second, the
renormalization of the neighbor states |Li and |Ri in the Rayleigh-Schrödinger case involves
the cluster state |ψq i (and hence all sites in the cluster), whereas in the JB case only the
sites |C1 i and |Cn i at the two ends of the cluster are involved.
It might seem surprising not to use standard Rayleigh-Schrödinger perturbation theory,
but there is a practical advantage to the JB approach. Namely, because the denominators
5
JB RG equations RS RG equations
t tn,R t tn,R
|ψq0 i = |ψq i + ... − ELL,1
−E1 cq,1 |Li − ER −En cq,n |Ri − ELL,1
−Eq cq,1 |Li − ER −Eq cq,n |Ri
t2 t2 t2 t2
Eq0 = Eq + ... − ELL,1 n,R
−E1 cq,1 − ER −En cq,n − ELL,1 2 n,R
−Eq cq,1 − ER −Eq cq,n
2
P
t n t
|L0 i = |Li + ... − E1L,1 − nq=1 Eq L,1
P
−EL q=1 c q,1 |C1 i −EL cq,1 |ψq i
t2 t 2
n Pn
EL0 = EL + ... − E1L,1 − q=1 Eq L,1 2
P
−EL q=1 cq,1 −EL cq,1
P
t n t
|R0 i = |Ri + ... − Enn,R − nq=1 Eqn,R
P
−ER q=1 cq,n |Cn i −ER cq,n |ψq i
t2 Pn t2
0 = E + ... − Enn,R − nq=1 Eqn,R 2
P
ER R −ER q=1 cq,n −ER cq,n
Pn c Pn
q=1 cq,n
− nq=1 tL,1 tn,R cq,1 cq,n Eq −E
q,1 1 1
−tL1 tnR Eq=1
P
tLR = 1 −EL
+ En −ER ‡ L
+ Eq −ER
TABLE I: Equations for the renormalized states and energies of the cluster and of its neighbors,
as well as the hopping amplitude between the left and right sites in the renormalized system.
Sites L and R are to the left and right, respectively, of cluster sites Cj , j = 1, n. tL,1 , and tn,R
are the hopping amplitudes between L and C1 and between Cn and R. EL , ER , E1 and En are
the (pre-renormalization) energies of the sites L, R, C1 , and Cn , respectively. Eq is the (pre-
renormalization) energy of the cluster eigenstate |ψq i = nj=1 cq,j |Cj i. ‡As discussed in the text,
P
JB does not provide an expression for tLR for n > 1. We use this expression in our implementation
of their work.
always involve precisely the bond strengths used to identify the cluster, there is never a case
of degeneracy. Meanwhile, in the Rayleigh-Schrödinger case, it is possible for a cluster state
energy Eq to be close to the energy of one of the neighboring sites, such that the terms
shown in Table I diverge and a degenerate perturbation calculation is required.
While the lack of degeneracy makes the JB approach simpler, it also limits its ability
to capture more extended states. With the intent to understand how much difference this
makes, we implement two versions of the Rayleigh-Schrödinger calculation. In one, SDRG-
RSnd, only clusters with no degeneracy are renormalized. Any cluster in which degeneracy
arises is not renormalized but instead handled as an independent cluster (as if the bonds at
the two ends of the cluster had zero strength). In the second, SDRG-RSdg, renormalization is
carried out, using standard degenerate perturbation theory, in clusters where a single cluster
state is degenerate with just one of the two neighbor sites (see A). In more complicated cases,
such as degeneracy of a cluster state with both neighbor sites or degeneracy of more than
6
one state of a given cluster, still no renormalization is done. The simple case accounts for
the bulk of the degeneracies, and the more complicated cases are very rare, as quantified
below in Section III.
The different approaches to the renormalization of states and energies also result in
distinct expressions for the hopping tLR between the left and right sights in the renormalized
system. JB defines
but provides no more explicit expression for it in either single-site or multi-site clusters.
In the case of a single-site cluster, both |L0 i and |R0 i have nonzero weight on the original
site |Ci, resulting in a nonzero value of tLR as shown in Eqn. (9). Moreover, in the case
of a multi-site cluster in the Rayleigh-Schrödinger version, both |L0 i and |R0 i have nonzero
weight on all sites in the cluster, again giving rise to a nonzero value of tLR . However, in
the JB approach, |L0 i has weight only on |C1 i, and |R0 i has weight only on |Cn i, such that
when n > 1, tLR = 0. Meanwhile, JB Fig. 3 shows the generation of nonzero bonds for all
cluster sizes, indicating that JB used something different from Eqn. (10) for tLR in the case
of multi-site clusters. The expression for tLR in the JB column of Table I is something we
have constructed following the form of the other JB equations which reduces to the one-
site cluster result in the limit n → 1 and produces results similar to those shown in JB as
discussed in Section III below.
In summary, we present the results of three renormalization approaches: SDRG-JB is
our implementation of the method described in JB to first order in t/∆E. SDRG-RSnd
uses standard Rayleigh-Schrodinger perturbation theory to first order in t/∆E, but for
clusters in which degeneracy arises no renormalization is performed. Finally, SDRG-RSdg
uses standard Rayleigh-Schrodinger perturbation theory to first order in t/∆E and handles
the most common form of degeneracy. We consider two choices of bond cutoff.
We compare the DOS and IPR results of these renormalization approaches to two alter-
natives: an exact solution and the ensemble of independent-clusters described above. While
the exact solution provides the ultimate benchmark, the independent-cluster solution gives
a sense for the impact of the renormalization.
As diagnostics of the renormalization procedures, and for comparison with JB, we also
calculate the fraction of the sites in the system which are removed at a given cluster size
7
Ncl , the distribution of bond strengths after all clusters of a given size have been removed,
and the number of degeneracies encountered.
0.5
0.04 0.09
Nss=10 0.4
0.02 0.08 Nss=100
0 2 4 6
Nss=1,000
Nss=10,000 0.3
0
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
ω ω
FIG. 2: Exact density of states and inverse participation ratio at disorder strength W = 10. Results
for a range of system sizes, with the product of the system size Nss and the number of disorder
configurations Ndc held constant at 107 .
This section begins with a brief examination of the exact results and then presents the
independent-cluster and renormalization results for two values of the bond cutoff.
Fig. 2 shows the exact DOS and IPR results for different system sizes, Nss . The number
of disorder configurations, Ndc , is chosen such that the total number of sites is the same for
all cases. The key message here is that for disorder strength W = 10 (the value used in JB),
the results are indistinguishable for Nss > 103 .
Fig. 3 shows the DOS and IPR results for disorder strength W = 10 from the exact,
independent-cluster, and renormalization approaches using bond cutoff m0 = 0.2, system
8
0.1 (a) (b) 0.8
0.5
0.04 0.09
exact
indep. clusters 0.4
0.08
0.02 0 2 4 6 SDRG-JB
SDRG-RSnd 0.3
SDRG-RSdg
0
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
ω ω
FIG. 3: Density of states and inverse participation ratio results at disorder strength W = 10 using
bond cutoff m0 = 0.2. Nss = 104 and Ndc = 103 .
size 104 , and 1000 disorder configurations. The independent-cluster results are qualitatively
similar to the exact results but clearly distinct quantitatively, both in magnitude and in the
location of the peaks.[16, 17] For the DOS, the SDRG-RSdg results are essentially the same
as those of SDRG-RSnd, and both are very close to the exact, while the SDRG-JB results
diverge, especially near the band center. For the IPR, all the renormalization methods give
similar results, following the exact results closely at the band edges but deviating in the
band center.
Fig. 4 shows the fraction of sites removed at each cluster size for the independent-cluster
approach, SDRG-JB, and both the SDRG-RS approaches. In general, one might expect that
this fraction is small at small cluster size (because, although there are many such clusters,
each contains only a small number of sites) and small at large cluster size (because, although
each contains many sites, there are only a few such clusters) with a peak in between. The
structure is indeed seen Fig. 4(a) with the notable exception of an enhanced value at cluster
size one. This enhancement reflects the fact that, when the bond cutoff approaches 1/W ,
9
ln(fraction of sites removed) 0 0
(a) (b)
bond cutoff 0.2 bond cutoff 0.4
-2 -2
-4 -4
indep. clusters
SDRG-JB
-6 SDRG-RSnd -6
SDRG-RSdg
-8 -8
0 2 4 6 8 10 0 2 4 6 8 10 12
cluster size cluster size
FIG. 4: The fraction of sites removed at each cluster size. Disorder strength W = 10, Nss = 104 ,
and Ndc = 103 .
bonds adjacent to a weak bond have an increased probability of also being weak: The two
sites surrounding a weak bond must be near opposite edges of the band. If mi,i+1 < m0 ,
then Ei and Ei+1 are near (opposite) band edges, and therefore mi−1,i and mi+1,i+2 are more
likely to also be weak.
For a given value of the bond cutoff, the result for the independent-cluster approach
represents the maximum rate of removal at small cluster sizes. Any renormalization scheme
producing renormalized bonds which are strong combines two smaller clusters into one larger
one, removing weight from the small cluster region of this diagram and shifting it to higher
cluster sizes. This shift is indeed visible in all three renormalization schemes. The results for
the two SDRG-RS approaches are very similar, and the SDRG-JB approach shows slightly
greater shift to large clusters (reduction in small clusters), corresponding to a higher rate of
generation of strong bonds.
Fig. 5 shows the log of the distribution of bond strengths for each approach. In the
independent-cluster approach, no new bonds are generated, so the distribution shows just
a monotonic decline in the weight below the bond cutoff. All the SDRG approaches do
generate new bonds, most of which are weak, reflected in the evolution of their distributions
10
0
indep. clusters SDRG-JB
log of distribution
-2 (a) (b)
Ncl=0
-4 Ncl=1
-6 Ncl=2
-8 Ncl=3
Ncl=4
-10 Ncl=5
0
(c) (d)
log of distribution
-2 SDRG-RSnd SDRG-RSdg
-4
-6
-8
-10
-20 -15 -10 -5 0 5 10 -15 -10 -5 0 5 10
log of bond strength log of bond strength
FIG. 5: Logarithm of the distribution of the log of the bond strengths at disorder strength W = 10
using bond cutoff m0 = 0.2. Nss = 104 and Ndc = 103 .
toward greater weight at low values and hence a flow toward the strong-disorder limit. The
rate at which weak bonds are generated is slower in SDRG-JB than in the two SDRG-RS
approaches, consistent with the data on the fraction of sites removed (Fig. 4).
For all quantities calculated with this bond cutoff, the difference between SDRG-RSnd
and SDRG-RSdg is negligible: handling cases of degeneracy makes very little difference. For
this relatively high disorder strength (W = 10) and low bond cutoff (m0 = 0.2), degener-
acy arises in less than 0.6% of states, and multiple degeneracy arises in 0.002% of cases.
Nonetheless, in Fig. 3(b) the SDRG-RSdg curve is very slightly lower (and closer to the ex-
act) at the band center, reflecting the particular significance of degeneracy in the formation
of more extended states.
A motivation for studying this method is the potential to extend it to interacting systems.
11
0.1 (a) (b) 0.8
0.5
0.04 0.09
exact
indep. clusters 0.4
0.08
0.02 0 2 4 6 SDRG-JB
SDRG-RSnd 0.3
SDRG-RSdg
0
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
ω ω
FIG. 6: Density of states and inverse participation ratio results at disorder strength W = 10 using
bond cutoff m0 = 0.4. Nss = 104 and Ndc = 103 .
Given that computational time grows exponentially with size in interacting systems, the
performance of the method at larger bond cutoff with correspondingly smaller cluster sizes
is of interest.
Figures 6, 4(b), and 7 show the corresponding results for bond cutoff, m0 = 0.4. In
Fig. 6(a), the independent-cluster approach results are significantly farther from exact than
for m0 = 0.2. The SDRG-JB results are also worse while both SDRG-RS results remain
quite close to the exact. For the IPR, Fig. 6(b), the independent-cluster results are very far
from the exact, and, unlike with m0 = 0.2, all the SDRG results are higher than the exact
results not just at the band center but all the way out to the peaks. The SDRG-JB and
SDRG-RSnd results are again similar to each other, while the difference between these and
the SDRG-RSdg results is larger than for m0 = 0.2. The impact of handling degeneracies is
more significant because there are more cases of degeneracy. Nonetheless, degeneracy still
only arises in less than 2% of states, and multiple degeneracy arises in 0.02% of cases.
In 4(b), the fraction of sites removed shows a monotonic decrease. For this larger value
12
0
indep. clusters SDRG-JB
log of distribution
-2 (a) (b)
Ncl=0
-4 Ncl=1
-6 Ncl=2
-8 Ncl=3
Ncl=4
-10 Ncl=5
0
(c) (d)
log of distribution
-2 SDRG-RSnd SDRG-RSdg
-4
-6
-8
-10
-20 -15 -10 -5 0 5 10 -15 -10 -5 0 5 10
log of bond strength log of bond strength
FIG. 7: Logarithm of the distribution of the log of the bond strengths at disorder strength W = 10
using bond cutoff m0 = 0.4. Nss = 104 and Ndc = 103 .
of the bond cutoff, the fraction of bonds which are weak is much larger, resulting in a
correspondingly greater proportion of small clusters. The renormalization approaches still
generate strong bonds, resulting in a shift of weight to larger cluster sizes. In Fig. 7,
the distribution of bond strengths shows more rapid generation of weak bonds than for
m0 = 0.2. In both the fraction of sites removed and the distribution of bond strengths, the
results for the two SDRG-RS approaches are similar, and SDRG-JB produces more strong
bonds resulting in greater shift toward larger cluster sizes in Fig. 4(b) than (a) and less
rapid generation of weak bonds in Fig. 7 than 5.
We now compare the results of our SDRG-JB approach with the results shown in JB.
Comparing Fig. 4(a) here with JB Fig. 1, an inconsistency is immediately apparent: The
data labeled “w = 10, m0 = 0.2” in JB show a monotonic decline unlike our Fig. 4(a).
13
Particularly puzzling, the fraction of sites removed at small cluster sizes in JB is much
larger than even the independent-cluster approach results in our Fig. 4(a). Similarly, the
rate of generation of weak bonds in JB Fig. 3 is much more rapid than in our Fig. 5. It was
partly in response to this discrepancy that we considered the case of bond cutoff m0 = 0.4.
Indeed, the SDRG-JB results in our Fig. 4(b) bear a strong resemblance to the curve labeled
“w = 10, m0 = 0.2” in JB Fig. 1, and correspondingly the SDRG-JB results in our Fig.
7 show a strong resemblance to JB Fig. 3. This suggests there may be a factor of two
difference between their implementation of the bond cutoff and ours. Consistent with this,
our IPR results for bond cutoff m0 = 0.4 (Fig. 6(b)) are similar to those labeled m0 = 0.2
in JB Fig. 2 (see C). In particular the vertical offset, which is maximum at the band center,
extends all the way to the peaks near the band edge. This is in contrast to our IPR results
for bond cutoff m0 = 0.2 (Fig. 6(a)) in which the difference from the exact results is only
readily apparent near the band center.
However, a factor of two difference in the bond cutoff does not fully resolve the differences
between our results and those in JB. In particular, the DOS results we find with the SDRG-
JB approach for bond cutoff m0 = 0.4 deviate strongly from the exact results, unlike the
DOS results shown in JB Fig. 2 (see C). Further, the results in JB labeled m0 = 0.05 are
puzzling: All initial bonds for disorder strength W = 10 would be greater than this cutoff,
such that the whole system should be a single cluster, and the results should be identical to
the exact results. This is true even if the cutoff is doubled. It is unfortunate that the JB
paper does not specify (i) the form of the renormalized hopping amplitude tLR in the case
of multi-site clusters, (ii) the approach they took to the loss of orthogonality in the basis,
or (iii) why they chose to address this given that it only arises beyond first order in t/∆E.
Nonetheless, setting aside the relatively minor quantitative differences between our results
and those of JB as well as between the different approaches we’ve implemented, an important
question is whether any of these provides a practical advantage. For the one-dimensional
non-interacting systems studied here, if the desire is to generate the DOS and IPR across
the full spectrum maximizing accuracy and speed, these SDRG approaches appear to be
less effective than simply applying exact methods to an ensemble of smaller systems. In
particular, comparing Fig. 3 and Fig. 6 with Fig. 2, we note that none of the RG approaches
is as close to the converged exact result as the exact results for Nss = 100. These exact
results for 100-site systems take only a few minutes of computational time, roughly two
14
orders of magnitude less than the exact results for systems of 104 sites, and also significantly
less than the time to run the renormalization code on the larger systems.
IV. CONCLUSION
15
schemes. Extension to interacting systems, in which exact solution of even small clusters is
numerically demanding, may still hold potential.
Acknowledgments
Appendix A: Degeneracy
1. Defining degeneracy
Pn
In our calculation, a cluster state |ψm i = j=1 wm,j |Cji with energy Em is determined
to be degenerate with the right neighbor site when
wm,n tnR
Em − R > 0.5 (A1)
Likewise, the state |ψm i is determined to be degenerate with the left neighbor site when
wm,1 tL1
Em − L > 0.5 (A2)
Each cluster falls into one of the categories listed in Table II.
The left and right degenerate renormalization procedures are described in the sections
below.
16
number of degeneracies renormalization procedure
with L with R SDRG-RSnd SDRG-RSdg
TABLE II: Categories of degeneracies and how they are handled in the SDRG-RSnd and SDRG-
RSdg code versions.
The following equations describe the renormalization of |ψqa i, Eqa , |Li, and EL .
p
y + y 2 + h2
α = q p (A3)
h2 + (y + y 2 + h2 )2
h
β = q p (A4)
h2 + (y + y 2 + h2 )2
Eqa + L
where x = (A5)
2
Eqa − L
y = (A6)
2
h = wqa ,1 tL1 (A7)
17
Energies
s 2
Eqa + L Eqa − L
E+0 = + + |wqa ,1 tL1 |2
2 2
∗ 2
|αwqa ,n tnR |2 X βwq,1 tL1
− (0)
− (0)
(A8)
R − E+ q6=qa Eq − E+
s 2
Eqa + L Eqa − L
E−0 = − + |wqa ,1 tL1 |2
2 2
∗ 2
|βwqa ,n tnR |2 X αwq,1 tL1
− (0)
− (0)
(A9)
R − E− q6=qa Eq − E−
States
We now address the hopping amplitude tLR and the IPR value of the degenerate cluster
state.
0 0
Because the renormalized left state may be either the ψ+ or ψ− , both cases are
considered.
0 0
When Eqa < L , y < 0, α < β, ψ+ → |L0 i and ψ− → ψq0 a . In this case, the hopping
amplitude
0 0 0
t+
LR = ψ+ H |R i (A12)
!
Eqa R
= αwq∗a ,n tnR 1− −
Eqa − R R − E+(0)
n
!
∗ ∗
X wq,1 wq,n X wq,1 wq,n
−βtL1 tnR +
q=1
Eq − R q6=q Eq − E+(0)
a
+O(t/∆E)2 (A13)
18
0
At exact degeneracy the first term is zero. As for the IPR, ψq0 a = ψ− , and using (A11)
4 4
Pn P αwq,1 tL1 4 βwqa ,n tnR
i=1 βwqa ,i + q6=qa Eq −E−
(0) wq,i +α +
R −E−
(0)
IP Rqa ,− = 2 2 !2 (A14)
Pn P αwq,1 tL1 βwqa ,n tnR
i=1 βw qa ,i + q6=qa (0) wq,i + α2 + (0)
Eq −E− R −E−
0 0
When Eqa > L , y > 0, α > β, ψ+ → ψq0 a and ψ− → |L0 i. In this case, the hopping
amplitude
t−
0 0 0
LR = ψ− H |R i (A15)
!
Eqa R
= βwqa ,n tnR 1− −
Eqa − R R − E−(0)
n
!
∗ ∗
X wq,1 wq,n X wq,1 wq,n
+αtL1 tnR +
q=1
Eq − R q6=q Eq − E−(0)
a
+O(t/∆E)2 (A16)
0
Again, at exact degeneracy the first term is zero. As for the IPR, ψq0 a = ψ+ , and using
(A10)
4 4
Pn P βwq,1 tL1 4 αwqa ,n tnR
i=1 αwqa ,i − q6=qa Eq −E+
(0) wq,i +β +
R −E+
(0)
IP Rqa ,+ = 2 2 !2 (A17)
Pn P βwq,1 tL1 αwqa ,n tnR
i=1 αwqa ,i − q6=qa (0) wq,i + β2 + (0)
Eq −E+ R −E+
19
The following equations describe the renormalization of |ψqa i, Eqa , |Ri, and ER .
p
y + y 2 + h2
α = q p (A18)
2 2
h + (y + y + h ) 2 2
h
β = q p (A19)
h2 + (y + y 2 + h2 )2
Eqa + R
where x = (A20)
2
Eqa − R
y = (A21)
2
h = wqa ,n tnR (A22)
Energies
s 2
Eqa + R Eqa − R
E+0 = + + |wqa ,n tnR |2
2 2
∗ 2
|αwqa ,1 tL1 |2 X βwq,n tnR
− (0)
− (0)
(A23)
L − E+ q6=qa Eq − E+
s 2
Eqa + R Eqa − R
E−0 = − + |wqa ,n tnR |2
2 2
∗ 2
|βwqa ,1 tL1 |2 X αwq,n tnR
− (0)
− (0)
(A24)
L − E− q6=qa Eq − E−
States
20
We now address the hopping amplitude tLR and the IPR value of the degenerate cluster
state.
0 0
Because the renormalized right state may be either the ψ+ or ψ− , both cases are
considered.
0 0
When Eqa < R , y < 0, α < β, ψ+ → |R0 i and ψ− → ψq0 a . In this case, the hopping
amplitude
0 0 0
t+
LR = hL | H ψ+ (A27)
!
L Eqa
= αwqa ,1 tL1 1 − (0)
−
L − E+ Eqa − L
n
!
∗
X wq,1 wq,n ∗
X wq,1 wq,n
−βtL1 tnR (0)
+
q6=qa Eq − E+ q=1
Eq − L
2
+O(t/∆E) (A28)
0
At exact degeneracy the first term is zero. As for the IPR, ψq0 a = ψ− , and using (A26)
4 4
∗ t
αwq,n
Pn P nR 4 βwqa ,1 tL1
i=1 βwqa ,i + q6=qa Eq −E (0) wq,i +α +
L −E−
(0)
−
IP Rqa ,− = 2 2 !2 (A29)
∗ t
αwq,n
Pn P nR βw t
qa ,1 L1
i=1 βwqa ,i + q6=qa (0) wq,i + α2 + (0)
Eq −E− L −E−
0 0
When Eqa > R , y > 0, α > β, ψ− → |R0 i and ψ+ → ψq0 a . In this case, the hopping
amplitude
t− 0 0 0
LR = hL | H ψ− (A30)
!
L Eqa
= βwqa ,1 tL1 1 − (0)
−
L − E− Eqa − L
n
!
∗
X wq,1 wq,n ∗
X wq,1 wq,n
+αtL1 tnR (0)
+
q6=qa Eq − E− q=1
Eq − L
2
+O(t/∆E) (A31)
0
Again, at exact degeneracy the first term here is zero. As for the IPR, ψq0 a = ψ+ , and
using (A25)
4 4
∗ t
βwq,n
Pn P nR 4 αwqa ,1 tL1
i=1 αwqa ,i − q6=qa Eq −E+
(0) wq,i +β +
L −E+
(0)
IP Rqa ,+ = 2 2 !2 (A32)
∗ t
βwq,n
Pn P nR αwqa ,1 tL1
i=1 αwqa ,i − q6=qa (0) wq,i + β2 + (0)
Eq −E+ L −E+
21
Appendix B: Orthogonality of the basis
When we renormalize a cluster i = C1 to Cn, we record its energies in the DOS and its
states in the IPR, after which the states |C1i to |Cni no longer appear explicitly in our
calculations. This is referred to as “removing” the cluster. We are left with a modified basis
of states which includes
Is this basis orthonormal? Only the states |L0 i and |R0 i have changed in the renormalization
process, so for all the other states if they were orthonormal before then they remain so.
Consider therefore just the inner products involving |L0 i and |R0 i.
Single-site cluster
0 0 tL1 tL1
hL |L i = hL| − h1| |Li − |1i (B3)
1 − L 1 − L
= 1 + O(t/∆E)2 (B4)
0 0 tL1 t1R
hL |R i = hL| − h1| |Ri − |1i (B5)
1 − L 1 − R
= 0 + O(t/∆E)2 (B6)
Multi-site cluster
!
X wq,1 tL1
hL0 |L0 i = hL| − hψq |
q
Eq − L
!
∗
X wq,1 tL1
× |Li − |ψq i (B8)
q
Eq − L
= 1 + O(t/∆E)2 (B9)
!
X wq,1 tL1
hL0 |R0 i = hL| − hψq |
q
Eq − L
!
∗
X wq,n tnR
× |Ri − |ψq i (B10)
q
Eq − R
= 0 + O(t/∆E)2 (B11)
22
Similar results obtain for |R0 i. Therefore, to first order in t/∆E, there is no loss of orthonor-
mality in the basis.
0.06 0.6
0.1
0.5
0.04 0.09
exact 0.4
SDRG-JB m0=0.2
0.02 0.08 SDRG-JB m0=0.4
0 2 4 6
Johri2014 Fig.2 m0=0.2 0.3
0
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
ω ω
FIG. 8: Density of states and inverse participation ratio at disorder strength W = 10, comparing
the results of our implementation of the procedure proposed in JB and the results presented in
JB. Exact and SDRG-JB results are for Nss = 104 and Ndc = 103 . The results from JB have been
extracted from Fig. 2 of their paper using the digitizing software WebPlotDigitizer. Specifically,
these are the data labeled m0 = 0.2 and represented by stars.
Fig. 8 provides a direct comparison between the results of our implementation of the
method described in JB and the results presented in JB. One conclusion from this figure is
that our calculation and that of JB are not the same. Our implementation is described in
detail in the body of the paper with key equations summarized in Table I. Potential points
are difference are (i) we keep only terms first order in t/∆E, and (ii) the form of tLR in
the case of multi-site clusters. JB does not provide an expression for tLR in this case. We
23
first tried strictly following the equations in JB , using tLR = hL0 | H |R0 i. However, this
is zero for clusters of more than one site, to first order in t/∆E, and JB Fig. 3 shows the
generation of nonzero bonds for all cluster sizes. We therefore constructed an alternative
expression, shown in Table I, which (i) follows the form of other JB equations, (ii) reduces
to the one-site cluster expression in the limit n → 1, and (iii) produces similar distributions
of bond strengths to those shown in JB Fig. 3.
Also apparent in this comparison between the results of our implementation SDRG-JB
and the results presented in JB is that the JB results are closer to the exact results for
the DOS but the SDRG-JB results are closer to the exact results for the IPR. Meanwhile,
as shown in the body of the paper, the procedures based on standard Rayleigh-Schrödinger
perturbation theory (SDRG-RSnd and SDRG-RSdg) produce DOS results much closer to the
exact results than SDRG-JB, and the SDRG-RSdg inverse participation results are slightly
better than those from SDRG-JB.
[1] P. Anderson, Absence of diffusion in certain random lattices, Physical Review 109 (1958) 1492.
[2] E. Abrahams, P. Anderson, D. Licciardello, T. Ramakrishnan, Scaling theory of localization:
Absence of quantum diffusion in two dimensions, Physical Review Letters 42 (1979) 673.
[3] S. Gopalakrishnan, S. Parameswaran, Dynamics and transport at the threshold of many-body
localization, Physics Reports 862 (2020) 1.
[4] D. Abanin, E. Altman, I. Bloch, M. Serbyn, Colloquium: Many-body localization, thermal-
ization, and entanglement, Reviews of Modern Physics 91 (2019) 021001.
[5] R. Nandkishore, D. Huse, Many-body localization and thermalization in quantum statistical
mechanics, Annual Review of Condensed Matter Physics 6 (2015) 15.
[6] S. Ma, C. Dasgupta, C. Hu, Random antiferromagnetic chain, Physical Review Letters 43
(1979) 1434.
[7] C. Dasgupta, S. Ma, Low-temperature properties of the random heisenberg antiferromagnetic
chain, Physical Review B 22 (1980) 1305.
[8] D. Fisher, Random antiferromagnetic quantum spin chains, Physical Review B 50 (1994) 3799.
[9] F. Igloi, C. Monthus, Strong disorder rg approach of random systems, Physics Reports 412
(2005) 277.
24
[10] F. Igloi, C. Monthus, Strong disorder rg approach — a short review of recent developments,
European Physical Journal B 91 (2018) 290.
[11] K. Wilson, J. Kogut, The renormalization group and the epsilon expansion, Physics Reports
12 (1974) 75.
[12] C. Monthus, T. Garel, Statistics of renormalized on-site energies and renormalized hoppings
for anderson localization in two and three dimensions, Physical Review B 80 (024203) (2009).
[13] H. J. Mard, J. A. Hoyos, E. Miranda, V. Dobrosavljevic, Strong-disorder renormalization-
group study of the one-dimensional tight-binding model, Physical Review B 90 (125141)
(2014).
[14] H. J. Mard, J. A. Hoyos, E. Miranda, V. Dobrosavljevic, Strong-disorder approach for the
anderson localization transition, Physical Review B 96 (045143) (2017).
[15] S. Johri, R. Bhatt, Large-disorder renormalization group study of the anderson model of
localization, Physical Review B 90 (2014) 060205(R).
[16] S. Johri, R. Bhatt, Singular behavior of eigenstates in anderson’s model of localization, Phys-
ical Review Letters 109 (2012) 076402.
[17] S. Johri, R. Bhatt, Singular behavior of anderson-localized wave functions for a two-site model,
Physical Review B 86 (2012) 125140.
25