You are on page 1of 13




1. Introduction
arXiv:1705.03039v1 [math-ph] 8 May 2017

The single particle Anderson model is defined by

(1.1) H ` V
where is a kinetic term and V is an random potential. In the standard formulation, is the discrete
Laplacian and V is a diagonal matrix with i.i.d. random entries with bounded density . H acts on the
Hilbert space H 2 pq for Zd . For sufficiently small , H has only pure point spectrum with
exponentially localized eigenvectors. The eigenvalues are a random point process, so that eigenvalues
corresponding to eigenfunctions localized at sufficient distances are nearly independent.
Our interest is in dynamics of models in the Anderson localized phase possessing highly correlated
energies localized at large distances. Our objective here is to consider a simple example of the phe-
nomena. To this end, consider the graph 2 created from two copies of Zd and an edge connecting the
origins of both lattices. Sites on this lattice may be labeled px, iq for x P Zd and i 1, 2, a natural
lattice distance may be incorporated into 2 as
}x y}1 if i j
d2 ppx, iq, py, jqq .
}x}1 ` }y}1 if i g
A natural basis for 2 p2 q is |x, iy which takes a value of 1 at site px, iq and 0 at all other sites.
Let both copies of Zd be equipped with a copy of the Hamiltonian H with identical potentials V
and let g parametrize a hopping between the origins of the respective lattices. That is we define the
(1.2) hg HZd HZd ` gp|0, 1yx0, 2| ` |0, 2yx0, 1|q
acting on 2 p2 q.
Whereas the rescaled eigenvalues of (1.1) obey Poisson statistics for small [8],[5], the rescaled
eigenvalues of (2.1) attain a clustering eigenvalue process of eigenvalue pairs. The pairs become
degenerate at g 0 but eigenvalues are almost surely simple at g 0. The lattice distance between
sites px, 1q and px, 2q is 2}x}1 , however, the eigenvalues localized in the region of these sites are highly
At a first viewing, (1.2) may appear unrelated to the more prominent multiparticle Anderson model
[1]. However, we would like to draw attention to the similarities these models possess. Both models
should be expected to exhibit resonant tunneling behavior, the dynamics of which have not been
explored in the multiparticle Anderson model case.
In the multiparticle model N copies of (1.1) act on a Hilbert space bN i1 pq for N particles. The
term becomes a Laplacian on each copy of Zd and the operator V is the local potential acting
on each particle. An additional term U is added for interactions between particles. For 2 particles,
configurations of the positions of the model are given by px, yq for x, y P Zd . As discussed in [1],
the symmetrized metric is the proper for indistinguishable particles. However, if the particles are
distinguishable then it is meaningful to ask about the dynamical properties in the non-symmetrized
metric. If the particles are of different species, say having slightly different mass or charge then a
Date: November, 2016.

transition of px, yq py, xq will be strongly resonant, and it is interesting to ask how long it takes
for the particles to exchange positions.
In a many body setting, similar propagations of excitations may well be infinite dimensional. In a
forthcoming paper [7], we consider the localized phase of a model of a tracer particle interacting with
a field of oscillators. A localized particle will affect the excitations at arbitrary distances, and the
nature of the statistics of the motion of the excitations is a nontrivial question.

2. Model
The double lattice model above (1.2) may be naturally recast as (1.1) describing state of a spin 1{2
particle perturbed by a transverse magnetic potential near the origin. Thus, we equivalently define
the Hilbert space as, H2 2 pxPZd C2 q the Hamiltonian becomes,

1 0 0 |yx|
(2.1) hg p ` V q `g
0 1 |yx| 0
where P 2 pZd q, so that }}2 1 with bounded support, and g is coupling to the external magnetic
potential. We will alter the above notation of the basis to |x, iy for i 1 and x P Zd . We will slightly
abuse notation and write |, iy for a state with entirely spin i electron for an 2 function .
It is well known that there exists a 0 0 so that || 0 implies H as defined in (1.1) obtains
pure point spectrum with simple eigenvalues and exponentially decaying eigenvectors. Moreover, H
obeys the fractional moment bound, see for example [2]. That is, there exists a C 8 so that, for
any 0 s 1 and 0 so that 1s 1 and z P C`
s |xy|
` s C C
(2.2) 1
E xy|pH zq |xy
1s 1s
Fractional moment bounds for (2.1) may almost immediately be inferred from (2.2).
Theorem 2.1. For any 0 s 1{4, there exists a C1 8 so that, for any 0 so that 1s
and z P C`
s 21 d2 ppx,iq,py,jqq
` s C
E xy, j|phg zq |x, iy C1
One virtue of fractional moment bounds such as (2.2) is their use in obtaining statements of Anderson
Localization. In particular, the property of dynamical localization for (1.1) was first obtained by this
method. Dynamical localization for H states that, there exists C 8 and 0 so that

(2.3) E sup |xy|e |xy| Ce|xy| .

Dynamical localization is a strong form of Anderson Localization, indeed it implies spectral localization
and bounds on eigenfunction correlators [2]. No such strong form of Anderson localization follows for
(2.1) in the metric d2 . However, we can extract a bound on rate of flipping spins for bounded times.
Let us define the support of as supp tx P Zd : pxq 0u.
Corollary 2.2. There exists C 8, 0 and 0 0 so that for any 0 0
E xy, i|eithg |x, iy tCeL
for L distpx, supp q ` distpy, supp q
Bounds of this form may be practically more useful for studying split level resonances in Anderson
localized systems as it would be easier to obtain in many cases. Indeed, such a bound can be obtained
for the model introduced in [7]. On the other hand, for (2.1) we can study the behavior of states
transfering between split energy levels much more precisely.

Our main theorem is the the following. Let us define the localization center of a vector P H2 as a
point x P Zd so that |xex , 0|y| _ |x0, ex |y| is maximized. Let us define distance
d1 ppx, iq, py, jqq }x y}
Let Ppx project to states within distance |x|{2 of |x, `1y in the d1 distance, that is

(2.4) Ppx |u, jyxu, j|.
u,j:d1 ppu,jq,px,`1qq|x|{2

We will examine the dynamics of a nearly localized eigenvector with spin entirely `1.
Theorem 2.3. There is a sequence of vectors k P H with localization centers xk with the following
There is an 0 so that, for t e|xk | the packet is stable, almost entirely spin up
@ D
(2.5) k , `1|eithg |k , `1 1 e|xk |

There is a time t 8 so that the packet reflects almost perfectly to the spin down position
@ D
(2.6) k , 1|eithg |k , `1 1 e|xk | .

Finally, the support at all times is almost entirely contained in |xk |{2 pxk q

(2.7) inf }Ppxk eithg |k , `1y}2 1 e|xk |


REMARK 2.4. The purpose of (2.7) is to emphasize that the particle travels to the second layer
only by tunneling, as no significant portion of the wave packet is ever near the edge connecting the two
In fact, (2.7) can be improved to a dynamical localization statement. That is, a statement analogous
to (2.3) holds for hg , which can be seen from Lemma (3.2) which depends on SULE (3.1).
Let us describe the organization of the rest of the paper. In the following subsection 2.1 we demon-
strate that the model with spin hg may be reduced to rank one perturbations of the spinless model and
thus obtain the spectral decomposition of (2.1) through that of (1.1). In Section 3 we introduce SULE
localization and describe the effect of the SULE basis under a rank one perturbation. In Section 4.1
we recall the Minami estimate and in Section 4.2 we utilze the SULE basis and the Minami estimate
to match vectors after a rank one perturbation. Finally in Section 5 we study the dynamics of the
(2.1) through the inherited SULE basis and matched eigenvalues, this obtains the proof of Theorem
2.3. In Appendix A we conclude with the short proofs of Theorem 2.1 and Corollary 2.2.

2.1. Reduction to H. hg may be analyzed by considering the family of operators, acting on H,

defined by,
(2.8) Hg H ` gD
where g P R and D |yx|. We will relate the Anderson localized phase of hg to Hg .
Proposition 2.5. (1.) For any g 0 and 0 the phg q pHg q Y pHg q.
(2.) Given g 0 there is a 0 0 so that for 0 0 , hg has simple pure point spectrum with
exponentially localized eigenvectors in the d1 metric. Moreover, pp phg q pp pHg q Y pp pHg q, and the
eigenvectors of hg are of the form | , `1y | , 1y where is an eigenvector of Hg . Conversely,
if is an eigenvector of Hg then | , `1y | , 1y is an eigenvector of hg .
This result `essentially follows from the
fact that hg commutes with
` p`1q the rotation operator U on H2 ,
p`1q p1q p1q
defined by U | , `1y ` | , 1y | , 1y ` | , `1y .

Proof. Recall that generalized eigenvectors with associated generalized eigenvalues in this context
are formal solutions, h for h hg or h Hg , not necessarily in 2 , which are polynomially
bounded. The usual arguments imply that the spectrum is exactly the closure of the set of generalized
eigenvalues. In fact, the growth of the generalized eigenfunctions may be bounded by a power d{2.
For details, see [3].
Suppose | p`1q , `1y ` | p1q , 1y is a generalized eigenvector, with generalized eigenvalue ie
hg . Then we have H piq ` gD piq piq , for i 1, from which we have

(2.9) pH gDqp p`1q p1q q p p`1q p1q q.

At least one of p p`1q p1q q are nonzero, thus, any generalized eigenvalue of hg is a generalized eigen-
value of Hg or Hg . Conversely, if is a generalized eigenvector of Hg with associated generalized
eigenvalue , then | , `1y | , 1y is a generalized eigenvalue of hg with generalized eigenvalue .
Thus, the set of generalized eigenvalues of hg is exactly given by the union of generalized eigenvalues
of Hg and Hg .
We first prove part (1.). The spectrum of Hg (and hg ) is the closure of its generalized eigenvalues,
as the generalized eigenvalues of hg is the union of generalized eigenvalues of Hg and Hg , the claim
Let us now prove part (2.). For given g and sufficiently small 0, Hg has almost surely simple
point spectrum and all generalized eigenvectors are exponentially decaying eigenvectors. Moreover,
with probability 1 pp pHg q X pp pHg q H [6], thus, in (2.9) at least one of p0q p1q is identically
0. The symmetry of eigenvectors of hg follows. The converse statement is immediate. 

3. Rank One perturbations

Let us consider stability of eigen-systems for Hg over g P R. For small enough , the Hamiltonian
Hg ` gD ` V is almost surely Anderson localized. For any eigenvector define the center of
localization to be a site x P Zd such that |xd`1 pxq| }X d`1 }8 . Where this definition is ambiguous,
select x which is left-most in terms of a lexicographic ordering of elements px1 , .., xd q. We will always
pgq pgq
suppose R is finite and 0 P . For g P , let I pgq index the eigenpairs pi , i q with normalized
pgq pgq pgq pgq pgq
eigenvectors }i }2 1, so that pH ` gDqi i i . Finally, let xi P Zd denote the center of
localization of i selected above.

3.1. SULE Localization. We recall the statement of SULE localization [2](Theorem 7.4). Let R
be a fixed finite set.
Theorem 3.1. Given , there is 0 0, so that for any 0 0 there exists a set 1
so that Pp1 q 1 and for every P 1 the Hamiltonian Hg has pure point spectrum, dense in
r0, 4ds ` supppq, such that all eigenvectors are exponentially decaying.
SULE (Semi Uniformly Localized Eigenfunctions) There exists a constant A A p, q and 0
so that for every g P and i P I pgq
pgq pgq
(3.1) |i pxq| A p1 ` |xi |qd`1 e|xxi | .
Let us define the local index set, for g P and Zd , the set is defined as
! )
pgq pgq
(3.2) I : i : xi P

and the local spectrum,

! )
pgq pgq pgq
(3.3) : i : i P I .

For Zd let us write P xP |xyxx| and for g P we write the local eigenbasis projection
pgq pgq
P |i yxi |.

In the next section we compare projections P to P .

3.2. Concentration of indices of SULE localization. We first show a limit on the fluctuation of
concentration of localization centers. Let us define the neighborhood of a set ,

p pq :
B tu P Zd : |u x| u


the core of is the set

p pq Zd zB
p pqd
Z z
p may be empty.
Depending on the choice of and , the set
Let us define a parameter p 1 and a sufficiently large constant C depending on d, , p, and the
realization of the random field. For u P Zd let u logp pC ` |u|q, for Zd let supuP u , we
define the standard neighborhood and core as
(3.4) p p q ;
B : B p p q .

The following lemma relates the basis of I to the spatial basis p|xyqx .
Lemma 3.2. Let parameters p, be as above. Moreover, fix 1, then let C be sufficiently large
depending on p, , and . Let be any subset Zd and let .
We have,
(3.5) }p1 PBp pq qP }2 e .

which implies,
(3.6) trpP q trpPB q.
On the other hand,
(3.7) }p1 P qPp pq }2 e .

which implies,
(3.8) trpP q trpP q
Observe that (3.5) states P projects almost entirely to PB , similarly, (3.7) states P projects
almost entirely to P . This observation leads to the implied statements (3.6) and (3.8) whose proofs
are included. The necessity of the increasing widths of the buffers are due to the prefactors in (3.1).
As a first application of Lemma 3.2 we bound the number of localization centers contained in a box.
For any u P Zd any L 1 and the box L puq : tx P Zd ||x u| Lu, the nesting
pLlogp pC `|u|`Lqq puq B pL`logp pC `|u|`Lqq puq
implies a bound on the number of states in the box
(3.9) |pLlogp pC `|u|`Lqq puq| |IL puq | 1 |pL`logp pC `|u|`Lqq puq|.
Now we prove the lemma.

Proof. The proof of the concentration inequalities (3.8) and (3.6) will follow immediately from esti-
mates (3.5) and (3.7). Indeed, all that is required is a general calculation for a seperable Hilbert space
r Let Pra and Prb be geometric projectors associated to different orthonormal eigenbases. That is, let

tri u and tri u be o.n. bases of H.
r Let Pra N |rpaq yxrpaq | and Prb M |rpbq yxrpbq |
paq pbq
i1 i i j1 j j

trrPra s trrPrbPra s ` trrp1 Prb qPra s

We may extract a trrPra s from the second term,

trrp1 Prb qPra s trrp1 Prb qPra Pra s trrPra s}p1 Prb qPra }
Thus we have,

(3.10) trrPra s 1 }p1 Prb qPra } trrPrb Pra s trrPrb s

Now let us prove (3.5). Let i P I , a state with localization center x xi inside then, as
x ,
pgq pgq
pgq pgq
}p1 PB qP |i y}2 |i pyq|2 |i pyq|2
yRB y:|xy|

Let us use (3.1) to bound the tail of i ,


pgq pgq
}p1 PB qP |i y}2 Cd A2 p1 ` |x|q4 k d1 e2k e .

The final inequality follows if x logp pC ` |x|q is taken large enough. We obtain (3.6) via (3.10).
Now let us prove (3.7) for x P
}p1 P qP |xy}2 |i pxq|2 .

Let us first expand the sum on the right hand side at each y R , by the eigenfunctions with localization
centers at xi y,

pgq pgq pgq
(3.11) |i pxq|2 |i pxq|2 |i pxq|2 .
pgq y:|yx| iPI pgq y:|yx| pgq
iRI tyu

From (3.6) the number of states with localization center at site y is bounded by 1 |By | 1 Cd logpd pC `
|y|q. Moreover, we use the SULE localization bound for i P Ityu , thus we have,
(3.12) |i pxq|2 1 Cd logpd pC ` |y|qA2 p1 ` |y|q2d`2 e2|xy| .
pgq y:|yx|

Clearly we have 1 Cd logpd pC ` |y|qA2 p1 ` |y|q2d`2 A2 p1 ` |y|q2d`3 for a large constant A . For
large enough C , which implies x is large, we have,

(3.13) Ap2 p1 ` |y|q2d`3 e|xy| 1.


Which implies iRI
pgq |i pxq|2 e obtaining (3.7) using (3.10). 

4. Matching Eigenbases
We are able to match eigenbases under rank for indices in regions with SULE localization and well
separated eigenvalues. The first step is to probabilistically control the local separation of eigenvalues.
The second step is a relatively standard perturbation argument for a system assuming well separated
4.1. Minami Estimate. Let us recall the standard Minami estimate for the Anderson model H
` V on a box . For an operator A on H and a subset Zd let us define the restriction to
as A P AP . We are only interested in sets away from the support of so this is equivalent
to the local Minami estimates for H. That is, for L puq, so that X supp H, we have
pH ` gDq H . As we will consider only boxes with support away from the support of , we will
simply use H below without further comment. For P H let us define a restriction and normalization
to as,
(4.1) P .
}P }
Finally let us define the minimal separation for values in a finite set. For a finite set T tt1 , .., tN u,
we define,
min rT s : mint|ti tj | : i ju.
Now let us introduce Minamis estimate [2] for the local minimum separation of eigenvalues.
Theorem 4.1. For any interval J R and subset Zd ,
` 2
P tr PpH`gDq pJq 2 p}}8 |J|||q2 .
As an immediate corollary we find a probabilistic bound on min rpH qs.
Corollary 4.2. For with compact support, there is some finite C so that for any 0,
P pmin rpH qs q Cp}}8 ||q2 .
Let puk q8 d 8
k1 be a sequence of sites uk P Z such that, there is a corresponding sequence pLk qk1 with
elements Lk 1 so that Lk 8 and, for all k, k , |uk uk1 | Lk ` Lk1 . For k P N let k Lk puk q
and let k L2d1
k .
Corollary 4.3. There are infinitely many k P N so that min rpHk qs k .
Proof. The events tmin rpHk qs k u are independent, and obey
P pmin rpHk qs k q 1 CL1
k ,

In particular, P pmin rpHk qsq 1 so that the sum of the probabilities over k is infinite. Thus, the
corollary follows from the second Borel-Cantelli theorem. 
4.2. Matching Eigenbases. Let us now combine the above Minami estimate with a SULE localized
model. We will establish that the labelings are stable under the rank one perturbation.
Let C be a sufficiently large constant, and for x P Zd and Zd define x and as discussed in
Section 3.2. We will consider , so that, from the definitions in Section 3.2,
(4.2) p pq B
B p pq .

We say an eigenpair of H ` gD (indexed by i P I pgq ) p, rq-corresponds to an eigenpair of H ` g 1D

1 pgq pg 1 q
(indexed by j P I pg q ) if |i j | ,
pgq pg 1 q
(4.3) |xi , j y| 1 ,

pgq pg 1 q
and |xi xj | r. For any p, rq-corresponding eigenvecors we will always choose the phase so that
pgq pg 1 q pg,g 1 q
|xi , j y 1| . Let I p, rq be the pairs of indices for p, rq-corresponding eigenpairs with
localization centers contained in .
Theorem 4.4. Suppose SULE holds for H ` gD for all g P . Given Zd , let and let us
p pq and
write B B p pq .

pgq pgq
1. For all g P and every i P I , there is an eigenvalue pi of HB so that |i pi | e{4 .
Moreover, assume that HB has simple spectrum satisfying min rpHB qs for some e{8 . Then
we have the following.
pg 1 q pg 2 q
2. For g 1 , g 2 P , suppose there are i P I and j P I associated to the same P pHB q, that
pg 1 q pg 2 q
is |i |, |j | e{4, then
pg 1 q pg 2 q
(4.4) |xi |j y| 1 e{16
3. For any g P , the associated eigenvalues of HB are distinct. That is, for i, j P I so that
i j, we have i j Moreover, there is a minimum separation for the local spectrum at :

min {2.

4. For any g 1 , g 2 P ,
pg 1 ,g 2 q
(4.5) |I pe{16 , q| 2|| |B|
that is there at least 2|| |B| pairs of pe{16 , q-corresponding eigenstates with centers in
It follows immediately from Corollary 4.3 and 4.4 that infinitely many eigenvectors will be matchable
between models. For u0 P Zd zt0u, let uk 4k`1 u0 for all k P N and Lk 4k1 |u0 |, finally let
k Lk puk q and k k .
Corollary 4.5. Suppose SULE holds for H ` gD for g g 1 , g 2 P . There is an 0 0 so that for
pg 1 q pg 2 q
any satisfying 0 0 , there are infinitely many k P N so that there exist ik P Ik and jk P Ik
pg 1 q
|xi |
which index pe k , k q-corresponding eigenpairs.

Proof. Let c 0 be sufficiently small and for all k let lk cLk let Bk B p plk q L `l puk q and
k k k
p plk q 2
k k Lk lk puk q. Let k Lk , then Corollary 4.3 states there are infinitely many k P N so
that min rpHBk qs k .
For each such k we apply Theorem 4.4 with k , B Bk and k . From Theorem 4.4 there
are 2|k | |Bk | 2 |k | many pelk {16 , k q corresponding pairs of eigenvectors. From each such k
pgq pg 1 q
let ik P Ik and jk P Ik index one pair of these vectors. Finally, for large enough k, |xik | lk {16
which obtains the result. 
We now prove Theorem 4.4.
Proof. We begin by truncating the eigensystem on Zd to the subset B and comparing these to the
eigensystem of HB .
pgq pgq
For g P and i P I from (3.5) in Lemma 3.2 the normalization factors in (4.1) obey }i }
2}PB i }. Moreover, let us write the remainder term
}i }
pgq pgq pgq pgq
pHqB i pi qB pgq
|xy i pyq : Ri .
}PB i } xPBB yx:yRB

Now we may use (3.5), to bound the norm of the remainder,

pgq 2 }i } pgq
}Ri }2 pgq
}p1 PB qi }2 4e
}PB i }
On the other hand, observe that
pgq pgq pgq pgq
(4.6) |xpi q |pH i q2 |pi q y| }Ri }22
Thus, by the min-max theorem, there is an eigenvalue P pHB q, with an associated eigenvector
P CB , so that | i | 2e{2 e{4, which proves part 1.
Let us compare the truncated and normalized eigenvectors to the eigenvector of the truncated
pgq pgq
system. Continue to assume g P and i P I and that is the HB associated to i . Now let
0 0 1 2 N be the eigenvalues of pHB q2 , and let P j project to the eigenspace of
HB associated to j . Now, from the calculation in part 1, we have,
pgq pgq pgq
(4.7) xpi qB |pH q2 |pi qB y }pi qpi qB ` Ri }22 4e{2.
On the other hand we have,

2 pgq pgq
(4.8) xpi qB |pH q |pi qB y j xpi qB |P j |pi qB y 1 xpi qB |p1 P0 q|pi qB y.
pgq pgq
But the i , and therefore the pi qB are normalized, thus we may combine (4.7) and (4.8) to obtain
4e{2 pgq
(4.9) x|pi qB y2 .
We will obtain both conclusions 2 and 3 of the lemma from equation (4.9). Let us now suppose that
pHB q is simple and min ppHB qq . Under this assumption 1 2 .
We now prove part 2. From part 1, we have that every eigenvalue of is within e{4 of pHB q.
pg 1 q pg 2 q
For an eigenpair p, q of H and g 1, g 2 P let us suppose there are two indices i P Iu,L and j P Iu,L
pgq pg 1 q
so that |i |, |j | e{4 . From, (4.9) we have for p, gq pi, g 1q, pj, g 2 q,
4 {2
1 2
e x|ppgq q y2 .

pgq pgq pgq
Therefore, as is normalized, we may write p q x|p q y|y ` | y where } }2 42 e{2
and x|y 0. It follows that

pg1 q pg 2 q pg 1 q pg 2 q 4 {2 4 {2
xpi q |pj q y x|pi q yx|pj q y |xi |j y| 1 2 e 2e .

The conclusion, (4.4) follows from the lower bound of .
Let us now show part 3, the separation in the local spectrum. From part 1, we have that every
pgq pgq
eigenvalue of is within e{4 of pHB q. Thus, using g g 1 g 2 suppose there is i, j P I , then,
pgq pgq pgq pgq pgq pgq pgq pgq
|xpi q |pj q y| 4|xP i |P j y| 4|xi ` pP 1qi |j ` p1 P qj y|
pgq pgq
if i j, then xi |j y 0 so that, using (3.7)
pgq pgq
|xpi q |pj q y| 16e .
pgq pgq pgq pgq
On the other hand, if i and j are both associated to eigenvalue P pHB q then i , j both
pgq pgq
obey (4.4) which is clearly a contradiction. Thus every eigenvalue of i P is associated to a
distinct eigenvalue pi of pHB q, so that |i pi | e {4 . The desired minimum separation follows
pgq pgq
from the minimum separation of pHB q, that is |i j | 2e{4.

Finally, let us prove part 4, again we assume a minimal separation of of the spectrum of HB . From
part 3 every index of I is associated to a distinct eigenvalue of HB . Using the pigeon hole principle
pg 1 q pg 2 q
we see that there are |I | ` |I | |pHB q| pairs of indices associated to the same eigenvalue of HB .
pg 1 q pg 2 q
Of course |pHB q| |B|, on the other hand, from (3.8) we have, |I |`|I | 2| |. We claim that
pg 1 q pg 2 q
indices of I and I corresponding to the same eigenvalue of HB are p2e{4, q corresponding so
that (4.5) follows from (4.2).
pg 1 q pg 2 q
To obtain the claim, consider i P I and j P I both associated to the same eigenvalue P pHq.
Closeness of eigenvalues and eigenvectors follow from part 2 of the theorem: indeed, we obtain closeness
pg 1 q pg 2 q
of eigenvalues we have |i j | 2e{4 and condition (4.3) implies the largeness of the scalar
product. Finally, we show the closeness of the centers of localization by supposing the converse.
pg 1 q pg 2 q pg 1 q pg 2 q
Suppose, |xi xj | logp pC ` |xi | _ |xj |q it would then follow from (3.5) that,
pg 1 q pg 2 q
pg 1 q pg 2 q
(4.10) |xi |j y| e|xi xj |{8
pg 1 q pg 2 q pg 1 q pg 2 q
Indeed, writing y1 xi , y2 xj and r |xi xj | note that
2 2
pg 1 q pg 2 q 2 pg1 q pg 2 q pg1 q pg 2 q
}Py1 Py2 } Py1 1 PByr{4 Py2 ` Py1 1 PByr{4 Py2 .
1 2

From which we have

1 2
}Pypg1 q Pypg2 q }2 2epr{4q
which obtains (4.10).

5. Tunneling between corresponding eigenvectors

In this section we utilize the description of p, q-corresponding pairs of Hg and Hg to analyze the
evolution operator eithg on selected states initialized in the spin up state. That is, let us assume Hg
and Hg both obtain SULE localization.
Let us write the indices of the eigenbases of Hg as i P I pgq . The corresponding eigenvectors
pgq pgq
and eigenvalues are then i and i . By Proposition 2.5, the (normalized) eigenvectors of hg are
exactly the vectors
1 pgq E E
(5.1) i ? i , `1 i , 1
for i P I pgq .
pgq pgq
Proposition 5.1. Let P I index a given pe|x | , q-corresponding eigenpair. We have the
following for all t 0,

ith pgq 1 it
pgq pgq
(5.2) e g
|i` , `1y ? Je
` ` Ie
` e|x |{4

1 0
for J and I J 2 .
0 1
Proof. First let us consider the expression of |i` , 0y in the eigenbasis of hg . For j P I pgq we may
pgq pgq
write aj j ` using expression (5.1) we have
1 E E
? ipgq
, `1 `

i` , 1 aj j .
2 j PI pgq

Thus, we may readily represent a spin up state as

pgq 1 1
|` , `1y ? ` ` ? aj j
2 j PI pgq

Let us apply the evolution operator to this expansion,

itphg q pgq 1 1 itppgq pgq q
(5.3) e ` |
i` , `1y ?
` ` ? e j
aj j
2 2 j PI pgq

Expanding p` we find
itphg q pgq 1 pgq

e ` |` , `1y ? J ` Ie ` aj j .
2 j PI pgq

the result follows as


1 pgq
q pgq
? J ` Ie ` a 4 |aj |2 8e|x |
2 j j
j PI pgq ztp u j PI pgq ztp u

and A E
pgq pgq pgq pgq
p` ` p p 4e{2
by the eigenfunction correspondence assumption. 
pgq pgq
|x |
Proposition 5.2. Let P I index a given pe , q-corresponding eigenpair. For all 0 t
|x |{2
e `
pgq ithg pgq |x |{4
(5.4) xi` , `1|e |` , `1y 1 e ` .
On the other hand, for xp` is sufficiently far from the origin,
p ithg pgq |x |{4
(5.5) sup pI Pxp` qe |` , `1y e `
pgq pgq
Finally, for t |pi pi |1

ith pgq it
pgq pgq
(5.6) e g | , `1y e p | , 1y e|x |{4
i ` i `
Theorem 2.3 now follows by combining Corollary 4.5 with Theorem 5.2. Indeed, Corollary 4.5
guarantees an infinite sequence of index pairs ik , jk which have the correspondence property fulfilling
the conditions of Theorem 5.2. Thus for each ik set ik then set k which obtains the desired
sequence of Theorem 2.3. The result follows as |u| |u| for u far enough from the origin.
pgq pgq pgq
|x |{2 it it q pgq
Proof. For t e ` , by the correspondence of eigenvalues we have |e e ` | 2t|
pgq |x |{2
` | 2e ` . Therefore,

2 0 pgq
it pgq
(5.7) Je ` ` Ie 4e|x` |{2
0 0
Combining this with, (5.2) we have

ithg pgq pgq pgq
|x |{4
e |i` , `1y |i` , `1y 2e ` .
pgq ithg pgq pgq
|x |{4
xi` , `1|e |` , `1y 1 e ` ,

which is the statement of (5.4).

The localization statement follows from a combination of (3.5) and (5.2). From (3.5) we have for
all t 0, and for large enough |x` |

p pgq
|x` |{2

` 2e

applying this bound to (5.2) obtains the result.

pgq pgq
Finally, set t |p p` |1 (5.6) 

Appendix A. Fractional moments and Dynamical localization

Here we will recall the results in fractional moment bounds for the standard Anderson model and
the related dynamical localization model. In [2](Theorem 6.3) fractional moments are bounded in
terms of self avoiding random walks. The following fractional moment bound holds.
Theorem A.1. There is a finite constant C C,d so that for any 0 s 1 and any z P CzR
s |xy|
1 s C C
Ep|xx|pH zq |yy| q
1s 1s
Transition probabilities may be bounded in terms of fractional moments. For I pHq Theorem
7.7 of [2] states,
Theorem A.2. There is C 8 depending on s so that,

Ep|xx|e |yy|q C lim inf Ep|xx|pH pE ` iqq1 |yy|s qdE
0 I

From these statements we can easily demonstrate Theorem 2.1 and Corollary 2.2.

0 |yx|
Proof. Let us write D then the resolvent equation states, for z P CzR,
|yx| 0
(A.1) phg zq1 ph zq1 gph zq1 Dphg zq1 .
Thus, if i j, two applications of of the above equation implies,
xx, i|phg zq1 |y, iy xx|pH zq1 |yy ` g 2 xx|pH zq1 |yx, i|phg zq1 |, iyx|pH zq1 |yy.
Now we take the sth moment for 0 s 1{4. For the first term we may apply Theorem A.1 directly.
For the second, we apply Holder inequality to find,
E|xx, i|phg zq1 |y, iy|s E|xx|pH zq1 |yy|s `
g 2s pE|xx|pH zq1 |y|4s q1{4 pE|x, i|phg zq1 |, iy|2s y1{2 pE|x|pH zq1 |yy|4s yq1{4
Assume |x| |y|, then we may bound the second and third factors of the second term, by a constant,
using Theorem 1.1 of [4]. Thus we have, using Theorem A.1,
s |xy| s |xu|
1 s C C 2s C C
E|xx, i|phg zq |y, iy| `g C
1s 1s uPsupp
1s 1s
s |xy|^p|x|L q
p C
C ,
where L maxt|u| : u P supp u. As |x y| 2|x|,
s L s |xy|{2
1 s p C
E|xx, i|phg zq |y, iy| C
1s 1s
concluding the result in this case.

On the other hand, for i j, (A.1) implies

xx, i|phg zq1 |y, iy gxx, i|ph zq1 |, iyx, i|phg zq1 |y, iy
Again take the s-moment for 0 s 1{4, and apply Holders theorem,
` 1{2 ` 1{2
E|xx, 0|phg zq1 |0, yy|s g s E|xex , 0|ph zq1 |, 0y|2s E|x0, |phg zq1 |0, ey y|2s
Again, assume |x| |y| and bound the second term by a constant to obtain,
s |xu| s |x|L
1 s s C C C
E|xx, i|phg zq |y, iy| g C C1 ,
1s 1s 1s
which concludes the proof. 
Proof. The perturbation theorem for semigroups obtains,
ithg ith
(A.2) e e ig eishg Deiptsqh ds.
Now let us take the expectation of the transition probability of A.2,
E|xy, 1|e |x, `1y| g E|xy, 1|eishg |, 1yx, `1|eiptsqh|x, `1y|ds
gt sup E|x|eisH |xy|ds.

The Corollary now follows from Theorem (A.1) and Theorem (A.2). 
[1] Michael Aizenman and Simone Warzel. Localization bounds for multiparticle systems. Communications in Mathe-
matical Physics, 290(3):903934, 2009.
[2] Michael Aizenman and Simone Warzel. Random Operators: Disorder Effects on Quantum Spectra and Dynamics,
volume 168 of Graduate Studies in Mathematics. American Mathematical Society, 2015.
[3] H.L. Cycon, R.G. Froese, W. Kirsch, and B. Simon. Schr odinger Operators: With Application to Quantum Mechanics
and Global Geometry. Theoretical and Mathematical Physics. Springer Berlin Heidelberg, 2009.
[4] Alexander Elgart, Mira Shamis, and Sasha Sodin. Localisation for non-monotone schrodinger operators. Journal of
the European Mathematical Society, 16(5):909924, 2014.
[5] Francois Germinet and Frederic Klopp. Spectral statistics for random Schrodinger odinger operators in the localized
regime. arXiv preprint arXiv:1011.1832, 2010.
[6] Vojkan Jaksic and Yoram Last. Spectral structure of anderson type hamiltonians. Inventiones mathematicae,
141(3):561577, 2000.
[7] Rajinder Mavi and Jeffrey Schenker. Anderson localization in the holstein model. preprint, 2016.
[8] Nariyuki Minami. Local fluctuation of the spectrum of a multidimensional anderson tight binding model. Commu-
nications in mathematical physics, 177(3):709725, 1996.