Mean-Field Dynamics of the Bose–Hubbard Model

in High Dimension
Shahnaz Farhat∗, Denis Périce†, Sören Petrat‡

School of Science, Constructor University Bremen,

Campus Ring 1, 28759 Bremen, Germany

January 10, 2025

arXiv:2501.05304v1 [math-ph] 9 Jan 2025

Abstract
The Bose–Hubbard model effectively describes bosons on a lattice with on-site interactions
and nearest-neighbour hopping, serving as a foundational framework for understanding strong
particle interactions and the superfluid to Mott insulator transition. This paper aims to rigorously
establish the validity of a mean-field approximation for the dynamics of quantum systems in high
dimension, using the Bose–Hubbard model on a square lattice as a case study. We prove a trace
norm estimate between the one-lattice-site reduced density of the Schrödinger dynamics and the
mean-field dynamics in the limit of large dimension. Here, the mean-field approximation is in the
hopping amplitude and not in the interaction, leading to a very rich and non-trivial mean-field
equation. This mean-field equation does not only describe the condensate, as is the case when the
mean-field description comes from a large particle number limit averaging out the interaction, but
it allows for a phase transition to a Mott insulator since it contains the full non-trivial interaction.
Our work is a rigorous justification of a simple case of the highly successful dynamical mean-field
theory (DMFT) for bosons, which somewhat surprisingly yields many qualitatively correct results
in three dimensions.

1 Introduction
One of the big aspirations of mathematical physics is to advance our rigorous understanding of phase
transitions. Within this research area lots of recent attention has been paid to the phenomenon of
Bose–Einstein–Condensation (BEC), a phase of matter of cold Bose gases that has been predicted in
1924 by Bose [6] and Einstein [14, 15]. Since then, BEC has been studied extensively by theoretical
physicists, and at least since the 1980’s also by mathematical physicists with more rigorous methods.
After the first experimental realizations in the labs of Cornell/Wieman [2] and Ketterle [12] in 1995
the study of BEC has received a new wave of attention throughout experimental, theoretical, and
mathematical physics. As a recent highlight in mathematical physics, let us mention the rigorous
derivation of the Lee–Huang–Yang formula by Fournais and Solovej [18, 19]. The motivation of our
work comes from different perspectives:

1. Lots of recent effort has been put into understanding BEC at zero temperature, e.g., in the Gross–
Pitaevskii limit [35] and thermodynamic limit [18, 19]. This yields insight into the behavior of the
condensate and its excitations, e.g., a rigorous proof of Bogoliubov theory in the Gross–Pitaevskii
regime [5], but is very far away from understanding the (thermodynamic) phase transition to
BEC.
∗
Email: sfarhat@constructor.university
†
Email: dperice@constructor.university, corresponding author.
‡
Email: spetrat@constructor.university

1
 2. A very useful and simple method for studying many-body systems is the mean-field approxima-
tion. For bosonic cold atoms, one scales down the interaction potential with the inverse of the
particle number N [16] (or density [13]), thus considering weak interaction. Then the inverse
particle number can be regarded as a small parameter, and the interaction can be effectively
replaced by its mean-field. In this mean-field limit, many physical effects can be rigorously
established regarding the dynamics and the low-energy properties. In particular, one can prove
the validity of Bogoliubov theory [28, 29], and perturbative expansions beyond Bogoliubov [7, 8].
However, these types of mean-field models do not describe phase transitions.

3. For Bose gases in the continuum, one would ultimately like to prove a thermodynamic phase
transition. However, Bose gases on the lattice offer a different possibility for a phase transition,
namely a quantum phase transition between a BEC and a localized state usually called a Mott
insulator; see, e.g., [37] for a review. There are only few mathematical rigorous works on this
topic, e.g., [1].

Our work addresses these points in the following way. 1: We study a limit that may describe a phase
transition. 2: Our limit is a mean-field limit, not for large particle number but for large dimension.
We hope and indeed show that some of the methods of large N mean-field limits are still relevant for
this case. Since in our model the averaging is done over the hopping terms, and the interaction is
treated non-perturbatively, our mean-field model is strongly interacting. 3: Our microscopic model
is the Bose–Hubbard model, which is a lattice model that has been successfully used to describe the
BEC-Mott transition.
Our main result is convergence of the reduced one-lattice-site density matrix of the many-body
Schrödinger dynamics to the mean-field dynamics, with an error bound that goes to zero as the dimen-
sion d goes to infinity. Other parameters such as the density and the coupling remain fixed. Thus, we
rigorously justify the validity of the mean-field approximation for a quantum system in large dimension.
We choose the Bose–Hubbard model to illustrate this statement both for its remarkable usefulness in
physics and for the technical simplifications it offers as a lattice model. The Bose–Hubbard model is a
popular model used to describe bosons on a lattice with on-site interactions, allowing hopping between
nearest-neighbor lattice sites. It is well-known for capturing strong interactions between particles [4]
and providing one of the simplest descriptions of the Mott transition to date, see [17] and later [26],
see also [22, 23].
A common technique to study models such as Bose–Hubbard is Dynamical Mean-field Theory
(DMFT). This theory is well-known for its description of the Mott insulator/superfluid phase transition
[25, 17]. It is usually formulated via a self-consistency condition for a Green’s function. DMFT is
typically justified in the physics literature by stating that mean-field theories become exact in the
limit of infinite dimensions [31]; see also [30] for fermions. A remarkable fact is that DMFT tends to
provide accurate results already in three dimensions [20]. In the literature, the effective equation we are
deriving here is often called the mean-field model of Fisher et al. [10], referring to [17]. Our equation
can be considered a simple case of DMFT. A more involved mean-field type equation is obtained in [10]
by scaling different parts of the hopping term in different ways. In the paper [17] the authors consider
the Bose–Hubbard model on a complete graph (the hopping term is of equal strength between all
vertices). In comparison, our model has only nearest-neighbor hopping. Rigorous justifications of the
effective thermodynamic behavior of the Bose–Hubbard model on a complete graph were obtained in
[9]. As mentioned above, in the mathematical literature, mean-field limits are typically considered as
many-particle limits for the Bose–Hubbard model [32] or, more generally, for continuous models where
the Hartree equation is obtained as effective dynamics (see, e.g., [3, 21] for reviews). This approach
requires dividing the interaction term by the number of particles to ensure that the kinetic energy and
the interaction energy of the ground state remain of the same order.
Our goal is to provide a rigorous justification, in the d Ñ 8 limit, that DMFT is a good approx-
imation of the Schrödinger equation in the context of the Bose–Hubbard model. It is interesting to
note that, in the large d limit, the roles of the kinetic energy and the interaction between particles are

2
inverted compared to the usual mean-field limit N Ñ 8. The terms we aim to average in our regime
are the hopping terms between nearest-neighbor sites which come from the kinetic energy. Since we
only consider on-site interactions, the interaction between particles acts as a one-site operator and
therefore does not contribute to correlations between two different lattice sites. For our setting, the
basic idea behind the mean-field approximation is that the coordination number of the lattice (the
number of nearest neighbors) increases with the dimension. This means we have a mean-field pic-
ture locally around every site, which allows us to control the correlations between sites. Note that
our main estimates are for the reduced one-lattice-site density matrix, and not for the one-particle
reduced density matrix that is usually used to describe convergence in the large N limit.

2 The Model and Main Results

2.1 Model
We consider the d-dimensional square lattice with periodic boundary conditions Λ – pZ{LZqd of
volume |Λ| – Ld , with L P N, L ě 2. We write x „ y if x, y P Λ are nearest neighbors. The
one-site Hilbert space is ℓ2 pCq and its canonical Hilbert basis is denoted by p|n⟩qnPN . We define the
standard creation and annihilation operators a˚ , a satisfying the CCR ra, a˚ s “ 1, ra, as “ 0 “ ra˚ , a˚ s;
explicitly,
?
a |0⟩ – 0, a |n⟩ – n |n ´ 1⟩ @n P N˚ ,
?
a˚ |n⟩ – n ` 1 |n ` 1⟩ @n P N.

The number operator is given as N – a˚ a. To simplify our notation in some later proofs, we introduce
an order on Λ such that @x P Λ

# ty P Λ|y ą x and x „ yu “ # ty P Λ|x ą y and x „ yu “ d.

For example, the lexicographic order does the job. The Fock space is
à 2
F – ℓ2 pCqb|Λ| – F` L2 pΛ, Cq – L pΛ, Cqb` k ,
` ˘
kPN

where b` denotes the symmetric tensor product. Given a one-site operator A and x P Λ, we define
˜ ¸ ˜ ¸
â â
Ax – 1 A 1 .
yăx yąx

In the following we define xx, yy to mean that x, y P Λ, x „ y and x ă y. The kinetic energy is given
by the negative second quantized discrete Laplacian
ÿ ÿ ` ˘ ÿ
´ dΓp∆d q – pa˚x ´ a˚y qpax ´ ay q “ ´ a˚x ay ` a˚y ax ` 2d Nx .
xx,yy xx,yy xPΛ

Furthermore, we denote by NF the number operator on Fock space, i.e.,

ÿ
NF – dΓp1q :“ Nx .
xPΛ

Given hopping amplitude J P R, chemical potential µ P R, and coupling constant U P R, we define

the Bose–Hubbard Hamiltonian
J U ÿ
Hd – ´ dΓp∆d q ´ µNF ` Nx pNx ´ 1q
2d 2 xPΛ
J ÿ ` ˚ ˘ ÿ U ÿ
“´ ax ay ` a˚y ax ` pJ ´ µq Nx ` Nx pNx ´ 1q. (1)
2d xPΛ
2 xPΛ
xx,yy

3
Here, we have scaled down the hopping term with the inverse of the dimension d. The time-dependent
Schrödinger equation for Ψd P F is
d
i Ψd ptq “ Hd Ψd ptq. (2)
dt
ś
The idea of the mean-field approximation is the lattice site product state ansatz Ψd « xPΛ φx where
φ P ℓ2 pCq is a one-lattice-site wave function. Such a state ś xPΛ φx is sometimes called Gutzwiller
ś
ś prod-
uct state [24, 34]. Our main results state that if Ψd p0q « xPΛ φx p0q, then also Ψd ptq « xPΛ φx ptq for
all times t ą 0, where « is meant in an appropriate reduced sense. We can guess the right mean-field
equation for φptq by writing ax “ xφptq, aφptqy ` r ax ptq, and neglect in the hopping term of (1) all
terms that are quadratic in r aptq. Then the corresponding mean-field equation is

iBt φptq “ hφptq φptq, (3)

where the nonlinear mean-field operator is

U
hφ – ´J αφ a˚ ` αφ a ´ |αφ |2 ` pJ ´ µqN ` N pN ´ 1q,
` ˘
(4)
2
with order parameter αφ :“ xφ, aφy. Roughly speaking, αφ “ 0 indicates a Mott insulator state,
whereas αφ ‰ 0 indicates a superfluid
ś state. The well-posedness of (3) is discussed in Section 3.
The approximation Ψd « xPΛ φx is not expected to hold in F, but in the sense of reduced lattice
site density matrices. Given Ψd , let us first define the corresponding positive trace one operator
γd P L1 pFq, which satisfies the von Neumann equation

iBt γd ptq “ rHd , γd ptqs . (5)

We define its first reduced one-lattice-site density matrix as

p1q 1 ÿ
γd – Tr pγd q .
|Λ| xPΛ Λztxu

p1q
The operator γd : ℓ2 pCq Ñ ℓ2 pCq is not to be confused with the reduced one-particle density matrix
p1q p1q
γparticle : L2 pΛq Ñ L2 pΛq defined via its integral kernel γparticle px, yq “ xΨd , a˚y ax Ψd y with x, y P Λ.
Given φ P ℓ2 pCq, let us also introduce the corresponding orthogonal projections

p “ pφ :“ |φyxφ|, and q “ qφ :“ 1 ´ pφ . (6)

p1q
In our main results we prove convergence of γd ptq to pφ ptq.

2.2 Main Results

p1q
Our main results are estimates for the trace-norm difference of γd ptq and pφ ptq, which we denote by
› p1q ›
›γ ptq ´ pφ ptq› 1 . We prove two similar estimates. Theorem 1 proves an error bound that holds for
d L
any value of the parameters J, µ, U of the Bose–Hubbard model (1). The convergence rate is slightly
worse than ?1d . However, we need to assume stronger conditions on the initial data, and the bound
contains a double exponential growth in time. On the other hand, Theorem 2 holds only for repulsive
interaction, i.e., U ą 0. However, it holds for a larger class of initial data, and the error bound only
grows exponentially in time. Our first main result is the following.

Theorem 1. Let γd be the solution to (5) with initial data γd p0q P L1 pFq, and let φ be the solution
to (3) with initial data φp0q P ℓ2 pCq such that ∥φ∥ℓ2 “ 1. Let pφ be defined as in (6). We assume that
there exist a, c ą 0 such that
n
´ ¯ n
p1q
@n P N, Tr ppφ p0q1N “n q ď ce´ a and Tr γd p0q1N “n ď ce´ a . (7)

4
Then for all t P R` we have
¨ ˛1
1 2
p1q ? p1q C2 eC1 t
` Tr ppφ p0qN q 2
γd ptq ´ pφ ptq ď 2 ˝ γd p0q ´ pφ p0q ` ´ ´a C1
¯a ¯‚
L1 L1
d C4 ` 2 t
2pa ` eqe 2 ` 1 lnpd ` 1q
ˆ ˆ? C1
˙ ? ˙
JC3 C4 `2 2pa`eqe 2 t `1 lnpd`1q t
e , (8)

with the following constants independent of d and t:

C1 – 2eJ max pTr ppφ p0qN q , 1q ,

` ˘
C2 – 4 cp1 ` aq ` e´1 p2 ` 4pa ` eqq ,
1
C3 – pTr ppφ p0qN q ` 1q 2 ,
1
C4 – 4Tr ppφ p0qN q 2 ` 2.

Note that the d-dependent terms on the right-hand side of (8) are small when d Ñ 8, since
1 ? ?
C lnpd`1qt lnpd`1qt´ 12 lnpdq´ 14 lnplnpd`1qq
¯1 e “ eC Ñ 0
dÑ8
´ a
2
d lnpd ` 1q

for any C, t ą 0. Our second main result is the following.

Theorem 2. Let γd be the solution to (5) with initial data γd p0q P L1 pFq, and let φ be the solution
to (3) with initial data φp0q P ℓ2 pCq such that ∥φ∥ℓ2 “ 1. Let pφ , qφ be defined as in (6). We assume
that there is C ą 0 such that
Trppφ p0qN 4 q ď C,
and that U ą 0. Then there exists CpJ, µ, U q ą 0 such that for all t P R` we have

˘¯ 1 1{2
ˆ ´ ˙
p1q 7q
p1q
ď CpJ, µ, U qeCpJ,µ,U qp1`t 2
`
γd ptq ´ pφ ptq Tr γ d p0q q φ p0qN q φ p0q ` q φ p0q ` . (9)
L1 d
ś
Note that for initial Gutzwiller product states Ψd p0q “ xPΛ φp0q, we have
´ ˘¯
p1q
Tr γd p0q qφ p0qN 2 qφ p0q ` qφ p0q “ 0.
`
(10)
´ ¯ ´ ¯
p1q p1q
More generally, assuming Tr γd p0qqφ p0qN 2 qφ p0q ď d´1 and Tr γd p0qqφ p0q ď d´1 , the estimate
(9) becomes
p1q 7 1
γd ptq ´ pφ ptq 1 ď CpJ, µ, U qeCpJ,µ,U qp1`t q ? .
L d
From the perspective of the law of large numbers, this is the expected optimal convergence rate.
(However, this convergence rate obviously does not explain why our approximation is so successful
even for d “ 3.) Note also that the bound (9) can be written in more detail as
p1q
γd ptq ´ pφ ptq 1
L
ˆ ˆ ˙ ˆ ´ ˙˙1{2
11 1 CpJ,µ,U q
ř7
tj p1q ` 2
˘¯ 1
` CpJ,
r µ, U q 1 ` e Tr γd p0q qφ p0qN qφ p0q ` qφ p0q ` ,
r
ď j“1
dU U2 d
(11)
where CpJ,
r µ, U q depends polynomially on the parameters J, µ, U of the Bose–Hubbard model. The
divergence for small U comes from our use of an energy estimate, as outlined below (around Equa-
tion (13)).

5
 The trace norm convergence of Theorems 1 and 2 in particular implies convergence of the order
parameter α, meaning
1 ÿ
αmicro ptq :“ xΨd ptq, ax Ψd ptqy Ñ αφptq :“ xφptq, aφptqy as d Ñ 8. (12)
|Λ| xPΛ

Note that for initial data Ψd p0q with a fixed particle number the left-hand side of (12) is zero (since
Hd is particle number conserving), but in general our initial data live on Fock space where the particle
number is not fixed.
p1q
Both theorems are proven using a Gronwall estimate for Trpγd qφ q. This quantity heuristically
counts the average number of lattice sites that do not follow the product state ansatz. It is inspired by
the corresponding quantity for the weak coupling limit introduced by Pickl [33]. The main technical
challenge is then caused by the unboundedness of the creation and annihilation operators in the
hopping term, a bit analogous to the technical problems that arise when considering the weak coupling
p1q
limit with singular interactions; see, e.g., [27]. More concretely, we need to bound Trpγd qφ pN `1qqφ q
p1q
in terms of Trpγd qφ q or terms that go to zero as d Ñ 8. This we do in two different ways, leading
to the two main theorems. For Theorem 1, we introduce a new moment method. For this, we first
separate
´ ¯ ´ ¯ ´ ¯
p1q p1q p1q
Tr γd qφ pN ` 1qqφ “ Tr γd qφ pN ` 1q1N ăM qφ ` Tr γd qφ pN ` 1q1N ěM qφ .

Then the first term can simply be bounded by M , whereas we use moment estimates to bound the
second term by eM Cptq d´1 . Then it turns out to be possible to optimize in M to close the Gronwall
argument. For Theorem 2, we proceed using an energy estimate inspired by [28] (which deals with
proving Bogoliubov theory for the dynamics of the weakly interacting Bose gas). The idea is to write
the Hamiltonian (1) as ÿ
Hd “ hxφptq ` Hptq.
r
xPΛ

Here, Hptq
r describes the excitations around our product state ansatz. A similar splitting was used in
[28], where, after a unitary transformation, Hptq
r converges to a Bogoliubov Hamiltonian. The energy
of the excitations should now be defined as
˜ ¸
A ÿ E
E exc ptq :“ Ψd , Hptq
r ` qx hxφptq qx Ψd , (13)
xPΛ

where the first term corresponds to the kinetic energy, and the second term to the mean-field energy of
the excitations (including the interaction energy). The energy E exc ptq is not conserved since excitations
from the product state can be created and annihilated. However, it can be bounded using a Gronwall
argument. The crucial point here is that the interaction term only as
ÿ
qx Nx pNx ´ 1qqx
xPΛ

and hence the Gronwall argument does not produce higher powers than qx Nx2 qx . This ultimately
p1q
allows us to control terms involving qx Nx2 qx or qx Nx qx , in particular Trpγd qφ pN ` 1qqφ q.
The rest of the paper is organized as follows. We first prove global well-posedness of the mean-field
equation (3) in Section 3. Then, we discuss some preliminary estimates in Section 4. In particular,
we prove properties of a two-lattice-site reduced density matrix, preliminary bounds for the mean-
field and Bose–Hubbard energies, and conservation laws and propagation estimates for the mean-field
p1q
equation and the Bose–Hubbard dynamics. Furthermore, we compute Bt Trpγd ptqqφ ptqq, and prove
p1q
bounds on all the terms in this time derivative except for Trpγd qφ pN ` 1qqφ q. Then, Theorem 1 is
proven in Section 5, and Theorem 2 is proven in Section 6, each using a different method to control
p1q
Trpγd qφ pN ` 1qqφ q.

6
3 Global Well-posedness of the Mean-Field Dynamics
The goal of this section is to prove well-posedness of the mean field dynamics. The mean-field dynamics
can be written as #
iBt φ “ hαφ φ “ Aφ ` F pφq,
(14)
φp0q “ φ0 ,
where the linear operator A and the nonlinear operator F are defined as
U
A :“ pJ ´ µqN ` N pN ´ 1q, (15)
2
F pφq :“ ´J αφ a˚ ` αφ a ´ |αφ |2 φ.
` ˘
(16)

When examining the semilinear equation (14) above, we cannot directly apply fixed-point arguments
to study global well-posedness because both A and F are unbounded operators, and the nonlinear
operator F is not Lipschitz continuous. Therefore, a different approach is required. Our strategy is to
approximate the nonlinear term so that it becomes Lipschitz continuous. This allows us to establish
the existence of a unique solution to the approximated problem by standard methods. Then, we show
that the obtained solution converges to the solution of the untruncated mean-field equation.

3.1 Approximating the Mean-field Dynamics

Let M ą 0 and consider the truncated creation and annihilation operators

aM :“ a1N ďM , a˚M :“ 1N ďM a˚ (17)

and

αM :“ xφM , aM φM y, (18)

where φM is the solution to the approximated problem

#
iBt φM “ hαMM φM “ AφM ` FM pφM q,
(19)
φM p0q “ φ0 P DpN 2 q,

where we have introduced the approximated nonlinear operator FM as

FM pφM q :“ ´J αM a˚M ` αM aM ´ |αM |2 φM .

` ˘
(20)

The solution to (19) solves the weak form of the preceding nonlinear equation (19) which is usually
known as Duhamel formula
# şt
φM ptq “ φ̃0 ptq ´ i 0 e´ipt´sqA FM pφM psqq ds,
(21)
φ̃0 ptq :“ e´itA φ0 , φ0 P DpN 2 q.

Remark 3. Note the following:

1. For the weak formulation (21) of the approximated nonlinear equation (19), the existence of
a unique local solution can be established using fixed-point arguments for a broader class of
initial data, specifically φ0 P ℓ2 pCq, which leads to the existence of a unique local solution φM P
Cpr0, T s, ℓ2 pCqq. This follows from the fact that FM is a nonlinear bounded operator satisfying
FM pφM q P Cpr0, T s, ℓ2 pCqq. However, to extend the solution to global times, we rely on the
conservation laws, which require the initial data φ0 P DpN 2 q, as A remains an unbounded
operator.

7
 2. Since FM pφM q P Cpr0, T s, ℓ2 pCqq, to ensure the equivalence between (19) and (21), it is enough
to restrict our analysis to initial data φ0 P DpN 2 q. For further details, see [11, Lemma 4.1.1,
Proposition 4.1.6 and Corollary 4.1.8].

3. One could in addition truncate the unbounded linear term A. On the one hand, this approach
ensures equivalence between (19) and (21) for all initial data φ0 P ℓ2 pCq. On the other hand,
it allows us to obtain a unique global strong solution φM P CpR, ℓ2 pCqq. However, ensuring
convergence to the solution of the mean-field dynamics becomes more complicated.

3.2 Properties of the Approximate Solution

In this subsection, we state some conservation laws for the approximated problem.
Lemma 4. Assume that φM is a solution to (19) with }φ0 }ℓ2 “ 1. Then, the following holds:
(i) }φM ptq}ℓ2 “ }φ0 }ℓ2 “ 1.

(ii) xφM ptq, N φM ptqy “ xφ0 , N φ0 y.

(iii) xφM ptq, hαMM φM ptqy “ xφ0 , hαMM φ0 y.

(iv) |αM | ď }N 1{2 φ0 }ℓ2 .

(v) Assume that φ0 P DpN k q. Then, there exists a constant C ą 0 such that
˘j A k´ 2j
E j
xφM ptq, N k φM ptqy ď 2k´2
ř ` 1{2 φ } t
j“0 CJk}N 0 ℓ2 φ0 , pN ` jq φ0 j! .

Proof. Statement (i) is true by definition of the truncation. For (ii), note that

rhαMM , N s “ ´Jp´αM a˚M ` αM aM q. (22)

This gives
d
xφM , N φM y “ ixφM , rhαMM , N sφM y “ ´iJxφM , p´αM a˚M ` αM aM qφM y “ 0.
dt
For (iii), we have
d
xφM , hαMM φM y “ ixφM , rhαMM , hαMM sφM y ` xφM , Bt hαMM φM y
dt
“ ´J xφM , p´Bt αM a˚M ` Bt αM aM ´ Bt αM αM ´ Bt αM αM q φM y “ 0.

Statement (iv) follows directly from Cauchy–Schwarz and (i)-(ii). For (v), note that
d
xφM , N k φM y “ ´iJαM xφM , ra˚M , N k sφM y ´ iJαM xφM , raM , N k sφM y
dt
“ ´iJαM xφM , 1N ďM ra˚ , N k sφM y ´ iJαM xφM , ra, N k s1N ďM φM y
´ ¯
“ 2JIm αM xφM , 1N ďM ra˚ , N k sφM y (23)

ď 2k|J||αM ||xφM , 1N ďM a˚ N k´1 φM y|

1
ď 2k|J|}N 1{2 φ0 }ℓ2 xφM , pN ` 1qk´ 2 φM y.

Iterating this p2k ´ 2q times leads to

2k´2
ÿ ´ ¯j A E tj
j
xφM , N k φM y ď C|J|k}N 1{2 φ0 }ℓ2 φ0 , pN ` jqk´ 2 φ0 . (24)
j“0
j!

8
3.3 Global Well-posedness of the Approximated Problem
For the approximated problem, proving global well-posedness is straightforward due to the use of
standard techniques such as fixed-point arguments, particularly because the nonlinearity in this case
is Lipschitz. We have the following results.

Lemma 5. For any fixed M ą 0, we have the following statements:

(i) There exists a unique global strong solution φM p¨q P CpR, DpN 2 qq of the Duhamel formula (21).

(ii) There exists a unique global strong solution φM p¨q P CpR, DpN 2 qq X C 1 pR, ℓ2 pCqq of the approxi-
mated problem (19).

Proof. Let X “ Cpr0, T s, DpN 2 qq denote the space of continuous functions from r0, T s to DpN 2 q,
equipped with the norm

|||φ||| :“ sup }φptq}DpN 2 q , }φptq}2DpN 2 q “ }φptq}2ℓ2 ` }N 2 φptq}2ℓ2 .

tPr0,T s
` ˘
Note that DpN 2 q, } ¨ }DpN 2 q is a Banach space. For a fixed M ą 0, we define the map ΓM : X Ñ X
by żt
ΓM pφqptq :“ φ̃0 ptq ´ i e´ipt´sqA FM pφpsqqds,
0

with φ̃0 ptq :“e´itA φ0

and where A and FM are defined in (15)and (20). We can check that, for any
T ą 0, the map ΓM is Lipschitz-continuous. More precisely, we claim that for all φ1 , φ2 P X,

|||ΓM pφ1 q ´ ΓM pφ2 q||| ď CpM, J, T q|||φ1 ´ φ2 |||,

where CpM, J, T q ą 0 is defined as

CpM, J, T q :“ M T |J|p6c2 ` 6c2 M 2 ` 10c4 q,

and c ą 0 as
c :“ max sup }φi ptq}DpN 2 q ă 8.
i“1,2 r0,T s

To prove the claim, we need first to establish some useful estimates. To this end, we denote αM,i psq :“
xφi psq, aM φi psqy. We have, for k ě 0 and for i “ 1, 2,
1 1
}N k a7M φi psq}ℓ2 ď M k` 2 }φi psq}ℓ2 ď M k` 2 c, 7 P t , ˚u, (25)
? ?
|αM,i psq| ď M }φi psq}2ℓ2 ď M c2 , (26)
?
|αM,1 psq ´ αM,2 psq| ď 2 M c}φ1 psq ´ φ2 psq}ℓ2 , (27)
ˇ ˇ
ˇ|αM,1 psq|2 ´ |αM,2 psq|2 ˇ ď 4M c3 }φ1 psq ´ φ2 psq}ℓ2 . (28)

We have then for all t P r0, T s,

›ż t ›
› ´ipt´sqA
›
}ΓM pφ1 qptq ´ ΓM pφ2 qptq}ℓ2 “› e
› pFM pφ1 psqq ´ FM pφ2 psqqq ds››
0 ℓ2
ż tˆ› ›
›αM,1 psqaM φ1 psq ´ αM,2 psqaM φ2 psq›
› ›
ď |J|
0 ℓ2

` }αM,1 psqa˚M φ1 psq ´ αM,2 psqa˚M φ2 psq}ℓ2

˙
› 2 2
›
` |αM,1 psq| φ1 psq ´ |αM,2 psq| φ2 psq ℓ2 ds.
› ›

9
We begin by considering the first term. Using (25) with k “ 0, along with (26) and (27), we get
› ›
›αM,1 psqaM φ1 psq ´ αM,2 psqaM φ2 psq› 2
› ›
ℓ
ď |αM,1 psq ´ αM,2 psq|}aM φ1 psq}ℓ2 ` |αM,2 psq|}aM pφ1 psq ´ φ2 psqq}ℓ2
ď 3M c2 }φ1 psq ´ φ2 psq}ℓ2 .

Similarly, for the second term, applying (25) with k “ 0, along with (26) and (27), we obtain

}αM,1 psqa˚M φ1 psq ´ αM,2 psqa˚M φ2 psq}ℓ2 ď 3M c2 }φ1 psq ´ φ2 psq}ℓ2 .

Finally, for the last term, by using (25) with k “ 0, along with (26) and (28), we obtain
› ›
›|αM,1 psq|2 φ1 psq ´ |αM,2 psq|2 φ2 psq› 2
ℓ
ˇ ˇ
ď ˇ|αM,1 psq|2 ´ |αM,2 psq|2 ˇ }φ1 psq}ℓ2 ` |αM,2 psq|2 }φ1 psq ´ φ2 psq}ℓ2
ď 5M c4 }φ1 psq ´ φ2 psq}ℓ2 .

To summarize, we obtain

}ΓM pφ1 qptq ´ ΓM pφ2 qptq}ℓ2 ď M T |J|p6c2 ` 5c4 q|||φ1 ´ φ2 |||. (29)

More generally, for any k ě 0, we have

}N k pΓM pφ1 qptq ´ ΓM pφ2 qptqq }ℓ2 ď M T |J|p6M k c2 ` 5c4 q sup }φ1 psq ´ φ2 psq}DpN k q . (30)
r0,T s

Specifically, for the other component of the norm, we have

}N 2 pΓM pφ1 qptq ´ ΓM pφ2 qptqq }ℓ2 ď M T |J|p6M 2 c2 ` 5c4 q|||φ1 ´ φ2 |||. (31)

By combining the two estimates (29) and (31) above, we obtain

|||ΓM pφ1 q ´ ΓM pφ2 q||| ď M T |J|p6c2 ` 6c2 M 2 ` 10c4 q|||φ1 ´ φ2 |||.

Considering the semilinear equation of the form (19), and noting that our nonlinearity is Lipschitz
continuous (or can be made a contraction by choosing T sufficiently small), we can approach the
problem in two ways. First, we can apply the local well-posedness results from [11], specifically [11,
Lemma 4.3.2 and Proposition 4.3.3], to obtain a unique local solution. Then, we can extend this
solution globally using [11, Theorem 4.3.4] by employing conservation laws, including the norm and
the moment bounds (i)-(v) in Lemma 4. On the other hand, we can establish global well-posedness
by directly applying the Banach fixed point arguments. To this end, we consider the closed ball on
the Banach space X defined by

BX pφ̃0 , Rs :“ tφ P X; |||φ ´ φ̃0 ||| ď Ru.

Then, we check that for T ą 0 small enough and for R ą 0 large enough, the map ΓM satisfies the
condition of the Banach fixed point theorem, namely

• ΓM maps BX pφ̃0 , Rs into itself,

• ΓM : pX, ||| ¨ |||q Ñ pX, ||| ¨ |||q is a contraction map,

guaranteeing the existence of a fixed point (φM “ ΓM pφM q P X). The solution can then be extended
globally using the same conservation laws employed for the first approach.

10
Remark 6. In the above theorem, we can also apply the fixed point theorem in the Banach space
X “ Cpr0, T s, ℓ2 pCqq, which guarantees the existence of a unique local solution φM P Cpr0, T s, ℓ2 pCqq.
Subsequently, we can globalize the solution (which is equivalent to obtaining an estimate of the norm
}φM ptq}ℓ2 on r0, T s) for the set of initial data φ0 P DpN 2 q, ensuring that

}φM ptq}ℓ2 “ }φ0 }ℓ2 .

The above estimate implies that

lim }φM ptq}ℓ2 “ }φ0 }ℓ2 ă 8,
tÒT

which guarantees that the solution does not blow up in finite time and thus T “ 8.

3.4 Convergence
By (i) and (v) from Lemma 4, we have that pφM qM PN and pN k φM qM PN are bounded sequences in the
Hilbert space ℓ2 pCq. Then there exist a convergent subsequence still denoted by pφM qM PN such that

• φM converges weakly to φ and the limit is unique,

• N k φM converges weakly to N k φ for all k P R` .

As a consequence of this convergence, we have

}φ}ℓ2 ď lim inf }φM }ℓ2 ,

M Ñ8
}N k φ}ℓ2 ď lim inf }N k φM }ℓ2 .
M Ñ8

This in fact implies strong convergence.

Lemma 7. Let pφM qM be a sequence of solutions to (19) with }φ0 }ℓ2 “ 1 and φ its associated weak
limit. Then we have for all k ě 0,

}N k pφM ptq ´ φptqq}ℓ2 ÝÑ 0. (32)

M Ñ8

Proof. This follows from the weak˚ convergence in the Banach space L1 pℓ2 pCqq “ pKpℓ2 pCqqq˚ , where
L1 pℓ2 pCqq and Kpℓ2 pCqq denote the space of trace-class and compact operators, respectively. Let
pφM “ |φM yxφM | be the projection onto the state φM , so taht in particular p2φM “ pφM . We also have

TrppφM q “ }φM }2ℓ2 “ 1, TrpN k pφM q “ xφM , N k φM y ă 8.

The second bound is a consequence of part (v) of Lemma 4. This ensures the existence of a subsequence,
still denoted by ppφM qM , such that
˚
pφM á ν as M Ñ 8 weakly * in L1 pℓ2 pCqq,
˚
N k pφM á N k ν as M Ñ 8 weakly * in L1 pℓ2 pCqq.
For any compact operator B P Kpℓ2 pCqq, this implies

TrppφM Bq Ñ TrpνBq and TrpN k pφM Bq Ñ TrpN k νBq as M Ñ 8.

For k ě 0, this leads to

TrpN k pφM q “ TrpN ´1 N k`1 pφM q Ñ TrpN ´1 N k`1 νq “ TrpN k νq. (33)

11
Specifically, for k “ 0 we have
TrppφM q “ }φM }2ℓ2 Ñ Trpνq,
which implies Trpνq “ 1. Now, using results from [36], we obtain strong convergence for all k ě 0,
´ˇ ˇ¯
k k ˇ k k ˇ
}N pφM ´ N ν}L1 “ Tr ˇN pφM ´ N ν ˇ Ñ 0 as M Ñ 8.

Since both pφM and ν are bounded in norm, we conclude that p2φM “ pφM converges strongly to
ν “ ν 2 . Therefore, ν is a projection, and ν “ Pχ “ |χyxχ|. Additionally, we have

TrppφM νq “ |xφM , χy|2 Ñ Trpν 2 q “ Trpνq “ 1.

On the other hand, we also know

xφM , χy Ñ xφ, χy,
which implies |xφ, χy|2 “ 1, leading to Pφ “ Pχ . Therefore, by exploiting (33), we have

N k φM á N k φ and }N k φM }ℓ2 Ñ }N k φ}ℓ2 .

Since ℓ2 is a Hilbert space, these results imply strong convergence

}N k pφM ´ φq}ℓ2 Ñ 0 as M Ñ 8.

Next, we show that the limit indeed satisfies the corresponding mean-field equation.
Lemma 8. Let pφM qM be a sequence of solutions to (19) and assume }φ0 }ℓ2 “ 1. Then the limit φ
satisfies the Duhamel version of the mean-field dynamics (14),
żt
φptq “ φ̃0 ptq ´ i e´ipt´sqA F pφpsqq ds, (34)
0

with φ̃0 ptq :“ e´itA φ0 and where F is defined in (16).

Proof. Let us start by establishing some useful estimates. Since both φM and N k φM converge weakly
to φ and N k φ, respectively, we have

}φ}ℓ2 ď lim inf }φM }ℓ2 “ }φ0 }ℓ2 “ 1, (35)

M Ñ8
}N 1{2 φ}ℓ2 ď lim inf }N 1{2 φM }ℓ2 “ }N 1{2 φ0 }ℓ2 , (36)
M Ñ8
|αφ | “ |xφ, aφy| ď }φ}ℓ2 }aφ}ℓ2 ď }N 1{2 φ0 }ℓ2 . (37)

Moreover, we also have

|αM ´ αφ | “ |xφM , aM φM y ´ xφ, aφy|

ď |xφM , paM ´ aqφM y| ` |xφM , apφM ´ φqy| ` |xpφM ´ φq, aφy|
ď }φM }ℓ2 }a1N ąM φM }ℓ2 ` }a˚ φM }ℓ2 }φM ´ φ}ℓ2 ` }aφ}ℓ2 }φM ´ φ}ℓ2
ď 2}pN ` 1q1{2 φ0 }ℓ2 }φM ´ φ}ℓ2 ` E1 pM q,

where we have introduced E1 pM q as

E1 pM q :“ }a1N ąM φM }ℓ2 . (38)

Then we estimate
› żt ›
› ›
›φptq ´ φ̃0 ptq ´ i e´ipt´sqA F pφpsqq ds› (39)
› ›
0 ℓ2

12
 ď }φM ptq ´ φptq}ℓ2 (40)
› żt ›
› ´ipt´sqA
›
` ›φM ptq ´ φ̃0 ptq ´ i e
› FM pφM psqq ds›› (41)
0 ℓ2
›ż t żt ›
› ´ipt´sqA ´ipt´sqA
›
`› e
› FM pφM psqq ds ´ e F pφpsqq ds›› (42)
0 0 ℓ2

The first term (40) converges to zero by Lemma 7. The second term (41) is zero because φM is a
solution to the approximated problem (21). It remains to estimate the difference between the nonlinear
parts,
ż t´ › ›
(42) ď |J| ›αM psqaM φM psq ´ αφ psqaφpsq› 2 (43a)
› ›
0 ℓ
` }αM psqa˚M φM psq ˚
´ αφ psqa φpsq}ℓ2 (43b)
› › ¯
` ›|αM psq|2 φM psq ´ |αφ psq|2 φpsq›ℓ2 ds. (43c)

For (43a), we find

(43a) ď |αM ´ α|}aM φM }ℓ2 ` |α|}paM ´ aqφM }ℓ2 ` |α|}apφM ´ φq}ℓ2

´ ¯
ď }N 1{2 φ0 }ℓ2 2}pN ` 1q1{2 φ0 }ℓ2 }φM ´ φ}ℓ2 ` }N 1{2 pφM ´ φq}ℓ2 ` 2E1 pM q ,

with E1 pM q from (38). The first and the second term go to zero as M Ñ 8. It remains to check
E1 pM q Ñ 0 as M Ñ 8. Indeed by (24), we have for some C ą 0 that

E1 pM q “ }a1N ąM φM }ℓ2
ď }1N `1ąM pN ` 1q´1{2 }L }pN ` 1q1{2 aφM }ℓ2
E tj 1{2
˜ ¸
2 ´
1 ÿ ¯j A j
ď? 2C|J|}N 1{2 φ0 }ℓ2 φ0 , pN ` jq2´ 2 φ0 .
M j“0 j!
loooooooooooooooooooooooooooooooooooomoooooooooooooooooooooooooooooooooooon
ă8 since φ0 PDpN 2 q

So, the term E1 pM q goes to zero as M Ñ 8. Similarly, for (43b) we find that

(43b) ď |αM ´ α|}a˚M φM }ℓ2 ` |α|}pa˚M ´ a˚ qφM }ℓ2 ` |α|}a˚ pφM ´ φq}ℓ2
ď 2}pN ` 1q1{2 φ0 }2ℓ2 }φM ´ φ}ℓ2 ` }N 1{2 φ0 }ℓ2 }pN ` 1q1{2 pφM ´ φq}ℓ2
` }pN ` 1q1{2 φ0 }ℓ2 E1 pM q ` }N 1{2 φ0 }ℓ2 E2 pM q,

where we have introduced

E2 pM q :“ }1N ąM a˚ φM }ℓ2 . (44)
By the same arguments as for (43a), the term (43b) goes to zero as M Ñ 8. It remains to estimate
the last term (43c). We have
› ›
(43c) “ ›|αM |2 φM ´ |αφ |2 φ›ℓ2
ˇ ˇ
ď ˇ|αM |2 ´ |αφ |2 ˇ }φM }ℓ2 ` |αφ |2 }φM ´ φ}ℓ2
ď |αM pαM ´ αφ q ` αφ pαM ´ αφ q| ` |αφ |2 }φM ´ φ}ℓ2
ď p|αM | ` |αφ |q |αM ´ αφ | ` |αφ |2 }φM ´ φ}ℓ2 ,

which also converges to zero by the same arguments as for the previous two terms.

13
4 Preliminaries
4.1 Reduced Densities Matrices
Given a density matrix γd P L1 pFq we define a two-lattice-site reduced density matrix

p2q 1 ÿ 1 ÿ
γd – TrΛztx,yu pγd q “ Tr pγd q .
d |Λ| 2d |Λ| x,yPΛ Λztx,yu
xx,yy
x„y

Note that this two-lattice-site reduced density matrix is symmetrized over all interacting pairs of sites
and not over all pairs of sites. The normalization factor 2d |Λ| is indeed the number of interacting
p2q
pairs of sites. Note that γd is symmetric, i.e.,
´ ¯ ´ ¯
p2q p2q
@A, B P Lpℓ2 pCqq, Tr γd A b B “ Tr γd B b A ,

p1q
and reduces to γd , i.e.,
´ ¯ ´ ¯
p2q p2q p1q
Tr1 γd “ Tr2 γd “ γd .

Moreover, if C P Lpℓ2 pCqq and D P Lpℓ2 pCqb2 q, then it follows directly from its definition that
1 ÿ ´
p1q
¯
Tr pγd Cx q “ Tr γd C ,
|Λ| xPΛ
1 ÿ ´
p2q
¯
Tr pγd Dx,y q “ Tr γd D .
2d |Λ| x,yPΛ
x„y

Furthermore, the following standard results hold. If A P Lpℓ2 pCqq is self adjoint such that A ě 0
p1q
or γd A P L1 pℓ2 pCqq, then
« ffp1q
ÿ ” ı
p1q
Ax , γd “ A, γd . (45)
xPΛ

p2q
If B P Lpℓ2 pCqb2 q is self adjoint such that B ě 0 or γd B P L1 pℓ2 pCqb2 q, then
» fip1q
1 — ÿ ´”
p2q
ı¯ ´”
p2q
ı¯
Bx,y , γd fl “ Tr1 B, γd ` Tr2 B, γd . (46)
ffi
2d x,yPΛ
–
x„y

4.2 Energy Bounds

With the definitions of one and two-lattice-site density matrices we can rewrite the energy per lattice
site as
ˆ ˆ ˙˙
Tr pγd Hd q p1q U ´
p2q
¯
“ Tr γd pJ ´ µq N ` N pN ´ 1q ´ JTr γd a˚ b a . (47)
|Λ| 2
Note that the mean-field energy can be written as
´ ¯ U
xφ, hφ φy “ J xφ, N φy ´ |αφ |2 ´ µ xφ, N φy ` xφ, N pN ´ 1qφy . (48)
2
The following bounds allow us to control the Bose–Hubbard energy and the mean-field energy in terms
of moments of the number operator.

14
Lemma 9. Let γd P L1 pFq and φ P ℓ2 pCq. Then there exists C ą 0 such that, for U “ 0,

|xφ, hφ φy| ď C xφ, N φy , (49)

|Tr pγd Hd q| ´ ´
p1q
¯¯
ď C 1 ` Tr γd N , (50)
|Λ|
and, for U ‰ 0,

|xφ, hφ φy| ď C 1 ` xφ, N 2 φy ,

` ˘
(51)
|Tr pγd Hd q| ´ ´
p1q
¯¯
ď C 1 ` Tr γd N 2 . (52)
|Λ|
Proof. Using Cauchy–Schwarz’s inequality we have

|αφ |2 “ |xφ, aφy|2 ď ∥φ∥2 ∥aφ∥2 “ xaφ, aφy “ xφ, a˚ aφy “ xφ, N φy . (53)

Recalling (48), this immediately yields (49) and (51). In order to obtain (50) and (52), we estimate
the two-site term in (47) with Cauchy–Schwarz to obtain
ˇ ´ ¯ˇ ´ ¯1 ´ ¯1 ´ ¯
p2q p1q 2 p1q 2 p1q
ˇTr γd a˚ b a ˇ ď Tr γd N Tr γd pN ` 1q ď Tr γd N ` 1.
ˇ ˇ

4.3 Conservation Laws

For both the Bose–Hubbard model (1) and the mean-field model (4) the total particle number and the
total energy are conserved. Furthermore, one can control higher powers of the total particle number.
Let us show this first for the mean-field equation. The total particle number is conserved since

iBt ⟨φ, N φ⟩ “ ⟨φ, rN , hφ s φ⟩ “ ´J pαφ ⟨φ, rN , a˚ s φ⟩ ` αφ ⟨φ, rN , as φ⟩q

“ ´J pαφ ⟨φ, a˚ φ⟩ ´ αφ ⟨φ, aφ⟩q
´ ¯
“ ´J |αφ |2 ´ |αφ |2
“ 0. (54)

The energy is conserved since

iBt ⟨φ, hφ φ⟩ “ ⟨φ, Bt hφ φ⟩ “ ´J ⟨φ, pBt αφ a˚ ` Bt αφ a ´ αφ Bt αφ ´ αφ Bt αφ q φ⟩

“ ´J pαφ Bt αφ ` αφ Bt αφ ´ αφ Bt αφ ´ αφ Bt αφ q
“ 0.

Moreover, we can prove two different bounds for controlling powers of the number operator, which we
will use for our two main theorems.

Proposition 10. Let φ solve (3) with φp0q P ℓ2 pCq. Let k P N{2, k ě 1 and t P R` . Then
´ ¯ ´ ´ ¯ ¯ 1
Tr pφ ptqN k ď Tr pφ p0qN k ` e´1 k k e2eJkTrppφ p0qN q 2 t , (55)
2pk´1q
ÿ ˆ ˙ ¯l ´
´ ¯ 2k ´ 1 l
¯
Tr pφ ptqN k
ď JTr ppφ p0qN q 2 t Tr pφ p0q pN ` lqk´ 2 . (56)
l
l“0

Proof. Let n P N. Recalling the mean-field dynamics (3), we find

D E D ” ı E
iBt φ, pN ` nqk φ “ φ, pN ` nqk , hφ φ

15
 D ” ı E
“ ´J φ, pN ` nqk , αφ a˚ ` αφ a φ
D ” ı E D ” ı E
“ Jαφ φ, a˚ , pN ` nqk φ ` Jαφ φ, a, pN ` nqk φ
” D ” ı Eı
“ 2iJ Im αφ φ, a, pN ` nqk φ . (57)
` ˘
Now, let A P L1 ℓ2 pCq be the positive operator defined as

Ak´1 – pN ` n ` 1qk ´ pN ` nqk ď k pN ` n ` 1qk´1 . (58)

Since a pN ` nqk “ pN ` n ` 1qk a we find

D ” ı E D E D k 1 k 3
E
φ, a, pN ` nqk φ “ φ, Ak´1 aφ “ A 2 ´ 4 φ, A 2 ´ 4 aφ

so with Cauchy–Schwarz’s inequality,

ˇD ” ı Eˇ D E1 D E1
k k´ 12 2 ˚ k´ 32 2
ˇ φ, a, pN ` nq φ ˇ ď φ, A φ φ, a A aφ
ˇ ˇ
D 1
E1 D 3
E1
ď k φ, pN ` n ` 1qk´ 2 φ φ, a˚ pN ` n ` 1qk´ 2 aφ
2 2

D 1
E1 D 3
E1
“ k φ, pN ` n ` 1qk´ 2 φ φ, pN ` nqk´ 2 N φ
2 2

D 1
E
ď k φ, pN ` n ` 1qk´ 2 φ . (59)

Combining (57) with (59) and also (53) we conclude

ˇ D Eˇ 1
D E
k k´ 1
ˇBt φ, pN ` nq φ ˇ ď 2Jk ⟨φ, N φ⟩ 2 φ, pN ` n ` 1q 2 φ . (60)
ˇ ˇ

Proof of (56). By induction on k, we prove that, φp0q, N k φp0q ă 8 implies that for all n P N,

D E 2pk´1q
ÿ ˆ2k ˙ ´ 1
¯l D l
E
k
φptq, pN ` nq φptq ď J ⟨φp0q, N φp0q⟩ 2 t φp0q, pN ` n ` lqk´ 2 φp0q . (61)
l
l“0

Then (56) follows for n “ 0. The inequality is indeed true for k “ 1 since ⟨φ, pN ` nq φ⟩ is conserved,
see (54). For the induction step, we assume (61) holds for some k and that
D 1
E
φp0q, N k` 2 φp0q ă 8,

and we now prove (61) for k ` 21 instead of k. Using (60) with k ` 21 instead of k, and using the
conservation of ⟨φ, N φ⟩ we find
ˇ D Eˇ D E
k` 1 1
k
ˇBt φ, pN ` nq 2 φ ˇ ď Jp2k ` 1q ⟨φp0q, N φp0q⟩ 2 φ, pN ` n ` 1q φ .
ˇ ˇ

Integrating over time and inserting (61) we conclude

D 1
E
φptq, pN ` nqk` 2 φptq

D E żt D E
k` 12 1
ď φp0q, pN ` nq φp0q ` Jp2k ` 1q ⟨φp0q, N φp0q⟩ 2 φpτ q, pN ` n ` 1qk φpτ q dτ
0
D E
k` 12 1
“ φp0q, pN ` nq φp0q ` Jp2k ` 1q ⟨φp0q, N φp0q⟩ 2

16
 2pk´1q
ÿ ˆ ˙
2k ´ 1 l
¯ D E żt
k´ 2l
J ⟨φp0q, N φp0q⟩ 2 φp0q, pN ` n ` l ` 1q φp0q τ l dτ
l
l“0 0
D E
k` 12
“ φp0q, pN ` nq φp0q
2pk´1q
ÿ ˆ ˙ ¯l`1 D
2k ` 1 ´ 1 l
E
` J ⟨φp0q, N φp0q⟩ t
2 φp0q, pN ` n ` l ` 1qk´ 2 φp0q
l`1
l“0
D 1
E
“ φp0q, pN ` nqk` 2 φp0q
2pk´1q`1
ÿ ˆ ˙ ¯l D
2k ` 1 ´ 1 1 l
E
` J ⟨φp0q, N φp0q⟩ 2 τ φp0q, pN ` n ` lqk` 2 ´ 2 φp0q
l
l“1
2p k` 12 ´1
ÿ q ˆ 2 `k ` 1 ˘˙ ´ 1
¯l D 1 l
E
ď 2 J ⟨φp0q, N φp0q⟩ 2 t φp0q, pN ` n ` lqk` 2 ´ 2 φp0q .
l
l“0

which concludes the induction.

Proof of (55). Since
1 k
N ě 1 ùñ pN ` 1qk “ N k ek lnp1` N q ď N k e N
we notice that
N ě k ùñ pN ` 1qk ď eN k .
Next, we continue from (60) for n “ 0, and introduce a cutoff, to obtain
ˇ ´ ¯ˇ 1
´ ¯
k
ˇBt Tr pφ ptqN k ˇ ď 2JkTr ppφ p0qN q 2 Tr pφ ptq pN ` 1q
ˇ ˇ
1
´ ¯
“ 2JkTr ppφ p0qN q 2 Tr pφ ptq pN ` 1qk p1N ăk ` 1N ěk q
1
´ ´ ¯ ´ ¯¯
ď 2JkTr ppφ p0qN q 2 Tr pφ ptqk k ` eTr pφ ptqN k
1
´ ´ ¯¯
“ 2JkTr ppφ p0qN q 2 k k ` eTr pφ ptqN k . (62)

With Gronwall’s lemma we conclude that

´ ¯ ´ ´ ¯ ¯ 1
Tr pφ ptqN k ď Tr pφ p0qN k ` e´1 k k e2eJkTrppφ p0qN q 2 t .

p1q
For the Bose–Hubbard model (1) the total energy Tr pγd Hd q and the total particle number Trpγd N q
are conserved as well. Moreover, we can prove bounds analogous to the mean-field dynamics for powers
of the total number of particles. Note first that we can rewrite the Hamiltonian Hd as
ÿ J ÿ
Hd “ hφx ´ pa˚x ´ αφ q pay ´ αφ q , (63)
xPΛ
2d x,yPΛ
x„y

with αφ “ xφ, aφy. Then, by using (45) and (46) we find that the one-lattice-site reduced density
matrix satisfies
» fip1q
« ffp1q
p1q
ÿ J — ÿ
iBt γd “ rH, γd sp1q “ hφ
x , γd pa˚x ´ αq pay ´ αq , γd fl
ffi
´
2d
–
xPΛ x,yPΛ
x„y
” ı ´” ı¯
p1q p2q
“ hφ , γd ´ JTr2 pa˚ ´ αq b pa ´ αq ` pa ´ αq b pa˚ ´ αq , γd . (64)

We have the following propagation bounds.

17
 p1q p1q
Proposition 11. Let γd solve (64), let k P N{2, k ě 1, t P R` , and γd p0qN k P L1 pℓ2 pCqq. Then
´ ¯ ´ ´ ¯ ¯
p1q p1q
Tr γd ptqN k ď Tr γd p0qN k ` e´1 k k e2eJkt . (65)

Proof. Similarly to (58), let us define

Ak´1 – pN ` 1qk ´ N k ď k pN ` 1qk´1 .

Then Cauchy–Schwarz yields

ˇ ´ ¯ˇ ˇ ´ ” ı ¯ˇ
p1q k ˇ p2q k
Tr γ N 2J Tr γ a , N 1 a2 ˇ
ˇ ˇ ˇ
B
ˇ t d ˇ ď ˇ d 1
ˇ ´ ¯ˇ
p2q
“ 2J ˇTr γd a2 Ak´1 a
ˇ ˇ
1 1 ˇ
´ ¯1 ´ ¯1
p2q 2 p2q 2
ď 2JTr γd a2 A1k´1 a˚2 Tr γd a˚1 Ak´1 1 a 1
´ ¯1 ´ ¯1
ď 2JkTr γd pN1 ` 1qk´1 pN2 ` 1q Tr γd N k
p2q 2 p1q 2

” ı
Since pN1 ` 1qk´1 , pN2 ` 1q “ 0, by Young’s inequality,
ˆ ˙
k´1 1 1
pN1 ` 1q pN2 ` 1q ď 1´ pN1 ` 1qk ` pN2 ` 1qk .
k k

Introducing a cutoff similarly to (62), we conclude that

ˇ ´ ¯ˇ ˆˆ
1
˙ ´ ¯ 1 ´ ¯˙ 12 ´ ¯1
p1q k ˇ p2q k p2q k p1q 2
ˇBt Tr γd N ˇ ď 2Jk 1´ Tr γd pN1 ` 1q ` Tr γd pN2 ` 1q Tr γd N k
ˇ
k k
´ ¯1 ´ ¯1
“ 2JkTr γd pN ` 1qk Tr γd N k
p1q 2 p1q 2

´ ¯1 ´ ´ ¯¯ 1
p1q 2 p1q 2
ď 2JkTr γd N k k k ` eTr γd N k
´ ´ ¯¯
p1q
ď 2Jk k k ` eTr γd N k .

With Gronwall’s lemma we conclude that

´ ¯ ´ ´ ¯ ¯
p1q p1q
Tr γd ptqN k ď Tr γd p0qN k ` e´1 k k e2eJkt .

4.4 Gronwall Estimate

´ ¯
p1q
Both Theorems 1 and 2 are proven via a Gronwall estimate for the quantity Tr γd q . This is
directly related to the trace norm difference of reduced density matrices, analogous to the case of the
weak coupling limit [27], as the following Lemma shows.

Lemma 12. Let p be a rank one projection and γ a positive trace 1 operator on ℓ2 pCq and q – 1 ´ p.
Then
? a
2Tr pγqq ď ∥γ ´ p∥L1 ď 2 2 Tr pγqq. (66)

Proof. In order to get the upper bound in (66), we first notice that since γ ď 1 and Tr pγq “ Tr ppq “ 1,

∥pγp ´ p∥L1 “ Tr pp1 ´ γq pq “ 1 ´ Tr pγpq “ Tr pγp1 ´ pqq “ Tr pγqq ,

18
so

∥γ ´ p∥L1 “ ∥pp ` qqγpp ` qq ´ p∥L1

ď 2Tr pγqq ` 2 ∥qγp∥L1
a a
ď 2Tr pγqq ` 2 Tr pγqq Tr pγpq
a ´a a ¯
“ 2 Tr pγqq Tr pγqq ` 1 ´ Tr pγqq
? a
ď 2 2 Tr pγqq,
? ? ?
where we used x` 1 ´ x ď 2 for 0 ď x ď 1. The lower bound follows directly from

Tr pγqq “ Tr ppp ´ γq pq ď ∥γ ´ p∥L1 .

´ ¯
p1q
Next, we compute the time derivative of Tr γd q and estimate some of the appearing terms.
This is analogous to the estimates in the weak´ coupling limit,
¯ see, e.g., [33, Lemma 3.2]. The only
p1q
term that causes technical difficulties is Tr γd q pN ` 1q q , and Sections 5 and 6 are devoted to
controlling this term in different ways, leading to our two main theorems.
Proposition 13. Let γd solve (5) with normalized initial data γd p0q P L1 pFq and φ solve (3) with
normalized initial data φp0q P ℓ2 pCq. We define p – |φ⟩ ⟨φ| and q – 1 ´ p. Then
ˇ ´ ¯ˇ
p1q
ˇBt Tr γd q ˇ
ˇ ˇ

¯ 1 Tr ppN q 12
˜ ¸
1 1
´ ¯ ´ ¯1 ´
p1q p1q 2 p1q 2
ď J pTr ppN q ` 1q 2 8Tr ppN q 2 Tr γd q ` 4Tr γd q Tr γd q pN ` 1q q ` .
d
(67)

Proof. Computation of the time derivative. We introduce the self-adjoint operator

A – pa˚ ´ αφ q b pa ´ αφ q ` pa ´ αφ q b pa˚ ´ αφ q .

With (64), we start by computing

´ ¯ ´” ı ¯ ´” ı ¯ ´ ¯
p1q p1q p2q p1q
iBt Tr γd q “ Tr hφ , γd q ´ JTr A, γd q1 ` Tr γd rhφ , qs
´ ¯
p2q
“ JTr γd rA, q1 s
” ´ ¯ı
p2q
“ 2iJIm Tr γd Aq1 . (68)

Inserting resolution of identities 1 “ p ` q, we get

´ ¯
p2q
Tr γd Aq1
´ ¯ ´ ¯ ´ ¯ ´ ¯
p2q p2q p2q p2q
“ Tr γd p1 p2 Aq1 p2 ` Tr γd p1 p2 Aq1 q2 ` Tr γd p1 q2 Aq1 p2 ` Tr γd p1 q2 Aq1 q2
´ ¯ ´ ¯ ´ ¯ ´ ¯
p2q p2q p2q p2q
` Tr γd q1 p2 Aq1 p2 ` Tr γd q1 p2 Aq1 q2 ` Tr γd q1 q2 Aq1 p2 ` Tr γd q1 q2 Aq1 q2 .

Note that q1 p2 Aq1 p2 and q1 q2 Aq1 q2 are self adjoint and hence do not contribute to 68. This is
also the case for q1 p2 Aq1 q2 and q1 q2 Aq1 p2 which are each others complex conjugate. Furthermore,
p1 p2 Aq1 p2 “ 0 by definition of A. Then, by symmetry, we see that p1 q2 Aq1 p2 is also not contributing,
since
´ ¯ ´ ¯ ´ ¯
p2q p2q p2q
Tr γd p1 q2 Aq1 p2 “ Tr γd q1 p2 Ap1 q2 “ Tr γd p1 q2 Aq1 p2 .

19
Thus, we are left with
´ ¯ ” ´ ¯ı ” ´ ¯ı
p1q p2q p2q
iBt Tr γd q “ 2iJIm Tr γd p1 p2 Aq1 q2 ` 2iJIm Tr γd p1 q2 Aq1 q2 . (69)

Estimation of the p1 p2 Aq1 q2 term. Since pq “ 0,

´ ¯ ´ ¯
p2q p2q
Tr γd p1 p2 Aq1 q2 “ Tr γd p1 p2 pa˚1 a2 ` a1 a˚2 qq1 q2 ,

p2q
and by symmetry of γd ,
´ ¯ ´ ¯
p2q p2q
Tr γd p1 p2 a˚1 a2 q1 q2 “ Tr γd p1 p2 a1 a˚2 q1 q2 .

Then, we use Cauchy–Schwarz to estimate, for any ϵ ą 0,

ˇ ˇ
ˇ ˇ
ˇ ´
p2q
¯ˇ 1 ˇˇ ÿ ˇ
ˇTr γd p1 p2 Aq1 q2 ˇ “ Tr pγd px py a˚x ay qx qy qˇ
ˇ ˇ ˇ
ˇ
d |Λ| ˇx,yPΛ ˇ
ˇ ˇ
x„y
ˇ ˜ ¸ˇ
1 ÿ ˇˇ 1 1 ÿ ˇ
ˇTr qx γd2 ¨ γd2 px py a˚x ay qy ˇ
ˇ
ď
d |Λ| xPΛ ˇ yPΛ,x„y
ˇ
¨ ˛
1 ÿ ϵ ÿ ˚ ÿ ‹
ď Tr pqx γd q ` Tr ˚γd px py a˚x ay qy qz a˚z ax pz px ‹
2dϵ |Λ| xPΛ 2d |Λ| xPΛ ˝
yPΛ,x„y
‚
zPΛ,x„z
1 ´
p1q
¯ Tr ppN q ÿ ÿ
“ Tr γd q ` ϵ Tr pγd px py ay qy qz a˚z pz q
2dϵ 2d |Λ| x,yPΛ zPΛ,x„z
x„y
1 ´
p1q
¯ Tr ppN q ÿ ` ˘
“ Tr γd q ` ϵ Tr γd px py ay qy a˚y py
2dϵ 2d |Λ| x,yPΛ
x„y
Tr ppN q ÿ ÿ
`ϵ Tr pqy γd qz px py ay a˚z pz q .
2d |Λ| x,yPΛ zPΛ,x„z
x„y z‰y

The last two summands can be estimated as

ÿ ` ˘ ÿ ` ˘ ÿ
Tr γd px py ay qy a˚y py ď Tr γd px py ay a˚y py “ Tr pppN ` 1qq Tr pγd px py q
x,yPΛ x,yPΛ x,yPΛ
x„y x„y x„y

ď 2d |Λ| pTr ppN q ` 1q ,

and
ÿ ÿ ÿ ÿ ` ˘1 1
Tr pqy γd qz px py ay a˚z pz q ď Tr γd qz px py ay a˚y py 2 Tr pγd qy pz az a˚z pz q 2
x,yPΛ zPΛ,x„z x,yPΛ zPΛ,x„z
x„y z‰y x„y z‰y
ÿ ÿ 1 1
“ Tr pppN ` 1qq Tr pγd qz px py q 2 Tr pγd qy pz q 2
x,yPΛ zPΛ,x„z
x„y z‰y
ÿ ÿ 1 1
ď pTr ppN q ` 1q Tr pγd qz q 2 Tr pγd qy q 2
x,yPΛ zPΛ,x„z
x„y
˜ ¸2
ÿ ÿ 1
“ pTr ppN q ` 1q Tr pγd qy q 2

xPΛ yPΛ,x„y

20
 ÿ
ď 2d pTr ppN q ` 1q Tr pγd qy q
x,yPΛ
x„y
´ ¯
p1q
“ 4d2 |Λ| pTr ppN q ` 1q Tr γd q .
1 1
Using these two estimates and then choosing ϵ´1 – 2dTr ppN q 2 pTr ppN q ` 1q 2 , we obtain
ˇ ´
p2q
¯ˇ 1 ´
p1q
¯
ˇTr γd p1 p2 Aq1 q2 ˇ ď Tr γd q ` ϵTr ppN q pTr ppN q ` 1q
ˇ ˇ
2dϵ ´ ¯
p1q
` 2dϵTr ppN q pTr ppN q ` 1q Tr γd q
ˆ ˙
1 1
´
p1q
¯ 1
“ Tr ppN q pTr ppN q ` 1q
2 2 2Tr γd q ` . (70)
2d
Estimation of the p1 q2 Aq1 q2 term. Since pq “ 0,
´ ¯ ´ ¯ ´ ¯ ´ ¯
p2q p2q p2q p2q
Tr γd p1 q2 Aq1 q2 “ Tr γd p1 q2 a˚1 a2 q1 q2 ` Tr γd p1 q2 a1 a˚2 q1 q2 ´ αφ Tr γd p1 q2 a˚1 q1 q2
´ ¯
p2q
´ αφ Tr γd p1 q2 a1 q1 q2 .

We estimate
ˇ ´ ¯ˇ ´ ¯1 ´ ¯1
p2q p2q 2 p2q 2
ˇTr γd p1 q2 a˚1 a2 q1 q2 ˇ ď Tr γd p1 q2 N1 p1 Tr γd q2 q1 a2 a˚2 q2
ˇ ˇ

1
´ ¯1 ´ ¯1
p2q 2 p2q 2
“ Tr ppN q 2 Tr γd p1 q2 Tr γd q1 q2 pN2 ` 1q q2
1
´ ¯1 ´ ¯1
p1q 2 p1q 2
ď Tr ppN q Tr γd q Tr γd q pN ` 1q q ,
2

ˇ ´ ¯ˇ ´ ¯1 ´ ¯1
p2q ˚ p2q 2 p2q 2
ˇTr γd p1 q2 a1 q1 q2 ˇ ď Tr γd p1 q2 N1 p1 Tr γd q1 q2
ˇ ˇ

1
´ ¯1 ´ ¯1 1
´ ¯
p2q 2 p2q 2 p1q
“ Tr ppN q 2 Tr γd p1 q2 Tr γd q1 q2 ď Tr ppN q 2 Tr γd q ,

and similarly
ˇ ´ ¯ˇ 1
´ ¯1 ´ ¯1
p2q p1q 2 p1q 2
ˇTr γd p1 q2 a1 a˚2 q1 q2 ˇ ď pTr ppN q ` 1q 2 Tr γd q Tr γd qN q ,
ˇ ˇ
ˇ ´ ¯ˇ 1
´ ¯
p2q p1q
ˇTr γd p1 q2 a1 q1 q2 ˇ ď pTr ppN q ` 1q 2 Tr γd q .
ˇ ˇ

Inserting these estimates yields

ˇ ´ ¯ˇ
p2q
ˇTr γd p1 q2 Aq1 q2 ˇ
ˇ ˇ

1
ˆ ´ ¯ ´ ¯1 ´ ¯1 ˙
p1q p1q 2 p1q 2
ď 2 pTr ppN q ` 1q 2 |αφ | Tr γd q ` Tr γd q Tr γd q pN ` 1q q

1
ˆ
1
´ ¯ ´ ¯1 ´ ¯1 ˙
p1q p1q 2 p1q 2
ď 2 pTr ppN q ` 1q 2 Tr ppN q Tr γd q ` Tr γd q Tr γd q pN ` 1q q
2 . (71)

Conclusion. Inserting (70) and (71) into (69) we obtain

ˇ ´ ¯ˇ
p1q
ˇBt Tr γd q ˇ
ˇ ˇ
ˆ ˙
1 1
´
p1q
¯ 1
ď 2JTr ppN q 2 pTr ppN q ` 1q 2 2Tr γd q `
2d
1
ˆ
1
´ ¯ ´ ¯1 ´ ¯1 ˙
p1q p1q 2 p1q 2
` 4J pTr ppN q ` 1q 2 Tr ppN q 2 Tr γd q ` Tr γd q Tr γd q pN ` 1q q

21
 ¯ 1 Tr ppN q 12
˜ ¸
1 1
´ ¯ ´ ¯1 ´
p1q p1q 2 p1q 2
“ J pTr ppN q ` 1q 2 8Tr ppN q Tr γd q ` 4Tr γd q Tr γd q pN ` 1q q `
2 .
d

5 Proof of Theorem 1
´ ¯
p1q
In this Section we prove Theorem 1 by estimating the term Tr γd q pN ` 1q q from the Gronwall
estimate in Proposition 13 using a moment method. The main idea is to use the following estimates
obtained by iterating the Cauchy–Schwarz inequality, along with the moment bounds we obtained in
Section 4.3.

Lemma 14. Let k P N and γ, p P L1 pℓ2 pCqq. We assume that 0 ď γ ď 1, that p is a rank one
projection and pN k , γN k P L1 pℓ2 pCqq. Then

´k
´ k
¯2´k
Tr pγq pN ` 1q qq ď Tr pγqq1´2 Tr γq pN ` 1q2 q , (72)
´ ¯ ´ ¯ ´ ¯
Tr γqN k q ď 2Tr γN k ` 2Tr pN k . (73)

Proof. Proof of (72). We proceed by induction on k. The inequality is trivial for k “ 0. With the
Cauchy–Schwarz inequality,
´ k
¯ 1
´ k`1
¯1
Tr γq pN ` 1q2 q ď Tr pγqq 2 Tr γq pN ` 1q2 q ,
2

so assuming the result holds for given k P N we can bound

´k
´ k
¯2´k
Tr pγq pN ` 1q qq ď Tr pγqq1´2 Tr γq pN ` 1q2 q
´k ´pk`1q
´ k`1
¯2´pk`1q
ď Tr pγqq1´2 `2 Tr γq pN ` 1q2 q
´pk`1q
´ k`1
¯2´pk`1q
“ Tr pγqq1´2 Tr γq pN ` 1q2 q ,

i.e., the result holds for k ` 1, closing the induction argument.

Proof of (73). With the Cauchy–Schwarz inequality,
´ ¯ ´ ¯ ´ ¯ ´ ¯ ´ ¯
Tr γqN k q “ Tr γN k ´ Tr γpN k p ´ Tr γpN k q ´ Tr γqN k p
´ ¯ ´ ¯ b b
“ Tr γN k ´ Tr γpN k p ` 2 Tr pγpN k pq Tr pγqN k qq
´ ¯ ´ ¯ 1 ´ ¯
ď Tr γN k ` Tr γpN k p ` Tr γqN k q ,
2
so
´ ¯ ´ ¯ ´ ¯ ´ ¯ ´ ¯
Tr γqN k q ď 2Tr γN k ` 2Tr γpN k p ď 2Tr γN k ` 2Tr pN k .

5.1 The Moment Method

We will prove Theorem 1 by showing that the probability of having a large lattice site occupation
outside the product state structure is small. We use the following basic Calculus estimates.

22
Lemma 15. Let pun qnPN Ă R` . Then
n ÿ
Da ą 0 s.t. @n P N, un ď e´ a ùñ @k P N, nk un ď p1 ` aqak k!, (74)
nPN

and conversely,
8
ÿ ÿ M
Db ą 0 s.t. @k P N, nk un ď bk k! ùñ @M P N, pn ` 1qun ď p2 ` 4bqe´ 2b . (75)
nPN n“M

Proof. Proof of (74). The function

x
fa : R` Ñ R` , x ÞÑ xk e´ a

is increasing up to ak and decreasing afterwards. Thus, by series-integral comparison,

ÿ ż
fa pxqdx ` fa ptakuq ` fa praksq “ ak ak! ` f1 a´1 taku ` f1 a´1 raks
` ` ˘ ` ˘˘
fa pnq ď
nPN R`
˜ ˆ ˙k ¸
k k k
ď a pak! ` 2f1 pkqq “ a ak! ` 2 .
e

If k ě 1, inserting the Stirling lower approximation,

ˆ ˙k
? k
2πk ď k! (76)
e
yields
˜ c ¸
ÿ
´n 2
nk e a ď ak k! a ` ď p1 ` aq ak k!.
nPN
πk

The statement also holds for k “ 0 since

ÿ n ż
x
´a
e ď 1 ` e´ a dx “ 1 ` a.
nPN R`

1
Proof of (75). If 0 ă a ă b and M P N we find
˜ ¸
8
ÿ
aM
ÿ
an
ÿ panqk ÿ ak ÿ
k`1
ÿ
k
pn ` 1qun e ď pn ` 1qun e “ pn ` 1q un “ n un ` n un
n“M nPN n,kPN
k! kPN
k! nPN nPN
ÿ ak ´ ¯ ÿ´ ¯
ď bk`1 pk ` 1q! ` bk k! “ bpk ` 1qpabqk ` pabqk
kPN
k! kPN
b 1
“ ` .
p1 ´ abq2 1 ´ ab
1
Choosing a “ 2b yields
8
ÿ 1 ´ ab ` b ´aM M
pn ` 1qun ď e “ p2 ` 4bqe´ 2b .
n“M
p1 ´ abq2

With this we can prove our first main theorem.

23
 ´ ¯
p1q
Proof of Theorem 1. Controlling Tr γd q pN ` 1q 1N ěM q with moments. Let k P N. Applying
´ ¯
p1q
(74) from Lemma 15 first to un – Tr ppp0q1N “n q and then to un – Tr γd p0q1N “n while using the
assumption (7) from Theorem 1, we obtain directly
´ ¯
Tr pp0qN k ď cp1 ` aqak k!, (77)
´ ¯
p1q
Tr γd p0qN k ď cp1 ` aqak k!. (78)

For k ě 1, we use first (73) from Lemma 14, then the moment bounds (55) from Proposition 10 and
(65) from Proposition 11, then (77) and (78), and Stirling’s approximation (76), and find
ÿ ´ ¯ ´ ¯
p1q p1q
nk Tr γd ptqq1N “n q “ Tr γd ptqqptqN k qptq
nPN
´ ¯ ´ ¯
p1q
ď 2Tr γd ptqN k ` 2Tr pptqN k
´ ´ ¯ ¯ 1
ď 2 Tr pp0qN k ` e´1 k k e2eJkTrppp0qN q 2 t
´ ´ ¯ ¯
p1q
` 2 Tr γd p0qN k ` e´1 k k e2eJkt
´ ´ ¯ ´ ¯ ¯
p1q
ď 2 Tr pp0qN k ` Tr γd p0qN k ` 2e´1 k k eC1 kt
´ ¯
ď 4 cp1 ` aqak k! ` e´1 k k eC1 kt
ek´1
ˆ ˙
k
ď 4 cp1 ` aqa ` ? k!eC1 kt
2πk
˘´ ¯
ď 4 cp1 ` aq ` e´1 ak ` ek k!eC1 kt
`

˘k
ď 4 cp1 ` aq ` e´1 pa ` eqeC1 t k!.
` ˘`

This is also valid for k “ 0, so (75) from Lemma 15 implies

´ ¯ 8
ÿ ´ ¯
p1q p1q
Tr γd ptqqptq pN ` 1q 1N ěM qptq “ pn ` 1qTr γd ptqqptq1N “n qptq
n“M
˘ ´ M e´C1 t
ď 4 cp1 ` aq ` e´1 2 ` 4pa ` eqeC1 t e 2pa`eq
` ˘`
M
C1 t´ 2pa`eq e´C1 t
ď C2 e . (79)

Conclusion of the proof. Let M P N˚ . We use the beginning of the Gronwall estimate from
Proposition 13 while introducing a cutoff on N , and then Proposition 10 to find, for any ϵ ą 0,
ˇ ´ ¯ˇ
p1q
ˇBt Tr γd q ˇ
ˇ ˇ
ˆ
1
´ ¯ ´ ¯1 ´ ¯1
p1q p1q 2 p1q 2
ď JC3 8Tr ppN q 2 Tr γd q ` 4Tr γd q Tr γd q pN ` 1q p1N ăM ` 1N ěM q q
1
¸
Tr ppN q 2
`
d
¯ 1 Tr ppN q 12
˜ ¸
´ 1 ? ¯ ´ p1q ¯ ´ ¯1 ´
p1q 2 p1q 2
ď JC3 8Tr ppN q 2 ` 4 M Tr γd q ` 4Tr γd q Tr γd q pN ` 1q 1N ěM q `
d
1
˜ ¸
´ 1 ? ¯ ´
p1q
¯ ´
p1q
¯ Tr ppN q 2
ď JC3 8Tr ppN q 2 ` 4 M ` 4ϵ´1 Tr γd q ` ϵTr γd q pN ` 1q 1N ěM q ` .
d

24
Next, we insert (79) and use the conservation of the mean-field number of particles (see (54)). Then
M
e´C1 t
the choice ϵ – d´1 e 2pa`eq yields
ˇ ´ ¯ˇ
p1q
ˇBt Tr γd ptqqptq ˇ
ˇ ˇ
1
˜ ¸
´ 1 ? ¯ ´
p1q
¯
C t´ M
e ´C1 t Tr ppp0qN q 2
ď JC3 8Tr ppp0qN q 2 ` 4 M ` 4ϵ´1 Tr γd ptqqptq ` ϵC2 e 1 2pa`eq `
d
¯ C eC1 t ` Tr ppp0qN q 21
˜ ¸
´ 1 ? ´ M
e´C1 t
¯ ´
p1q 2
ď JC3 8Tr ppp0qN q 2 ` 4 M ` 4de 2pa`eq Tr γd ptqqptq ` .
d

Then we choose1
S ˜ ¸W
C1 t d
M – 2pa ` eqe ln a .
ln pd ` 1q

Observing that for d ě 1 we have

˜ ¸
d
ln a ď lnpd ` 1q,
ln pd ` 1q

this choice implies

˜ ˜ ¸ ¸1
? 2
M
´ 2pa`eq e´C1 t C1 t d a
M ` de ď 2pa ` eqe ln a `1 ` ln pd ` 1q
ln pd ` 1q
´a C1
¯a
ď 2pa ` eqe 2 t ` 1 lnpd ` 1q ` 1.

Consequently,
˜
ˇ ´ ¯ˇ ´ ´a C1
¯a ¯ ´ ¯
p1q p1q
ˇBt Tr γd ptqqptq ˇ ď JC3 2C4 ` 4 2pa ` eqe 2 t ` 1 lnpd ` 1q Tr γd ptqqptq
ˇ ˇ

1
¸
C2 eC1 t ` Tr ppp0qN q 2
` .
d

Noticing that the time dependent coefficients in the above expression are non-decreasing in time, we
can use Gronwall’s lemma to obtain
¨ ˛
1
´ ¯ ´ ¯ C t
C2 e ` Tr ppp0qN q
1 2
p1q p1q
Tr γd ptqqptq ď ˝Tr γd p0qqp0q ` ´ ´a C1
¯a ¯‚
d 2C4 ` 4 2pa ` eqe 2 t ` 1 lnpd ` 1q
ˆ ˆ? C1
˙? ˙
JC3 2C4 `4 2pa`eqe 2 t `1 lnpd`1q t
e .

Finally, using (66) from Lemma 12 proves Theorem 1.

1
Let us comment on the choice of the cutoff parameter. Optimizing in M requires to solve, for x ě 0,
? ´ x e´C1 t x ´C1 t x ´C1 t x ´C1 t d2 ´C1 t
x “ de 2pa`eq ðñ e a`e e x “ d2 ðñ e a`e e e “ e
a`e a`e
ˆ 2 ˙ ˆ 2 ˙
x ´C1 t d d
ðñ e “ W0 e´C1 t ðñ x “ pa ` eqeC1 t W0 e´C1 t ,
a`e a`e a`e
where W0 is the principal branch of the Lambert W function. Our choice of M comes from the fact that
ˆ ˙
x
W0 pxq “ ln ` op1q.
xÑ8 lnpxq

25
6 Proof of Theorem 2
In this section we prove Theorem 2 using an energy estimate. Recall that the Bose–Hubbard Hamil-
tonian Hd can be written as a sum of two time-dependent quantities,
ÿ αφ
Hd “ hx ptq ` H̃ptq,
xPΛ

α
where αφ ptq :“ xφptq, aφptqy. Here, hx φ ptq is the mean-field operator from (4), i.e.,

α
” ı U
hx φ ptq :“ ´J αφ ptqa˚x ` αφ ptqax ´ |αφ ptq|2 ` pJ ´ µqNx ` Nx pNx ´ 1q, (80)
2
and H̃ptq can be computed as
J ÿ ´ p2q p2q
¯
H̃ptq : “ ´ px ptqpy ptqKx,y qx ptqqy ptq ` px ptqqy ptqKx,y qx ptqpy ptq ` h.c.
2d ăx,yą
J ÿ p3q
´ px ptqqy ptqKx,y ptqqx ptqqy ptq ` h.c. (81)
d ăx,yą
J ÿ p4q
´ qx ptqqy ptqKx,y ptqqx ptqqy ptq,
2d ăx,yą

where
p2q
Kx,y :“ a˚x ay ` a˚y ax , (82)
p3q p2q
Kx,y ptq :“ Kx,y ´ αφ ptqa˚x ´ αφ ptqax , (83)
p4q
Kx,y p3q
ptq :“ Kx,y ptq ´ αφ ptqa˚y ´ αφ ptqay ` 2|αφ ptq|2 . (84)

piq
Here, the superscript i in the expression Kx,y refers to the number of q’s that accompany it in the
p2q p3q p4q
expression of H̃ in (81). Note that Kx,y does not depend on t whereas the other terms Kx,y and Kx,y
do through the term αφ ptq.
For our proof we define the quantities
C ˜ ¸ G
1 ÿ` αφ αφ ˘
f ptq :“ Ψd ptq, Hd ` qx ptqhx ptqqx ptq ´ hx ptq ` cqx ptq Ψd ptq (85)
|Λ| xPΛ

with c ą 0, and
1 ÿ@
Ψd ptq, qx ptqNx2 qx ptq ` qx ptq Ψd ptq .
` ˘ D
gptq :“ (86)
|Λ| xPΛ
The idea of the proof is@ the following. In the Gronwall
D estimate from Proposition 13, the problem-
1 ř 2
atic term was |Λ| xPΛ Ψd ptq, qx ptqNx qx ptqΨd ptq . Hence, one might want to attempt to do a joint
1 ř
Gronwall argument for this and the original quantity |Λ| xPΛ xΨd ptq, qx ptqΨd ptqy that we want to
estimate, i.e., a Gronwall argument for g. However, if one computes the time derivative of g, one
finds higher and higher powers qN k q that need to be controlled, so the Gronwall argument cannot be
closed. The trick is to do instead a Gronwall argument for f . Except for the cq term, f represents
the energy of deviations from the lattice product state structure. The technical advantage for the
Gronwall argument is that xΨd ptq, Hd Ψd ptqy is conserved, and

qhαφ q ´ hαφ “ ´hαφ p ´ phαφ ` phαφ p,

so the N 2 term from the interaction appears always together with at least one p projection. And all
powers of N can be controlled when traced out against p due to Proposition 10. Hence, we can close

26
a Gronwall estimate for f . Finally, one can prove that Cg ´ d´1 ď f ď Cg ` d´1 . Hence, g can
be estimated in terms of its initial data and an error d´1 , which, together with Lemma 12, implies
Theorem 2.
In the following, we start by proving the equivalence of f and g up to an error d´1 in Section 6.1.
Then, in Section 6.2, we prove the Gronwall estimate for f . We conclude with the proof of Theorem 2
in Section 6.3.
Notation. In the following estimates, we use the quantities C ą 0, CpJ, µ, U q ą 0, and C̃ptq ą 0
with the following definitions:

• C is a positive constant that is independent of all parameters of the model.

• CpJ, µ, U q is a positive constant that depends on the parameters J, µ, U only polynomially, and
is independent of the initial conditions and time t.

• C̃ptq is a positive quantity that may depend on CpJ, µ, U q, the initial data xφp0q, N j φp0qy for
j ď 4, and polynomially on time t.

For convenience, these quantities may change from one line to the next in the subsequent estimates.

6.1 Equivalence of f and g

We start by presenting an estimate for a slightly modified f .

Proposition 16. There exist C ą 0 such that for all ϵ ą 0 we have

ˇC ˜ ˙ ¸ Gˇˇ
1 ˇˇ ÿ ˆ αφ U
2
ˇ Ψd , H̃ ` qx hx ´ Nx qx Ψd ˇ
ˇ
|Λ| ˇ xPΛ
2 ˇ
˜ ¸
U 2
ˆ ˙ ˆ ˙
2 1 2 1 ÿ (87)
ďC 1`J ` J ´µ´ 1 ` ` xφp0q, N φp0qy xΨd , qx Ψd y
2 ϵ |Λ| xPΛ
1 ÿ 1
`ϵ xΨd , qx Nx2 qx Ψd y ` ,
|Λ| xPΛ d

Proof. Recalling the definition of H̃ in (81), we find

1 J 1 ÿ p2q
xΨd , H̃Ψd y “ ´ xΨd , px py Kx,y qx qy Ψd y ` h.c. (88)
|Λ| 2d |Λ| ăx,yą
J 1 ÿ p2q
´ xΨd , px qy Kx,y qx py Ψd y ` h.c. (89)
2d |Λ| ăx,yą
J 1 ÿ p3q
´ xΨd , px qy Kx,y qx qy Ψd y ` h.c. (90)
d |Λ| ăx,yą
J 1 ÿ p4q
´ xΨd , qx qy Kx,y qx qy Ψd y, (91)
2d |Λ| ăx,yą

p2q p3q p4q

where the terms Kx,y , Kx,y and Kx,y are defined in (82), (83) and (84). Let us estimate the above
equation term by term. The pp-qq term of (88) has already been estimated in the proof of Proposi-
tion 13. Here, we find it slightly more convenient to choose ϵ´1 – 2dTr ppN q pTr ppN q ` 1q, so instead
of (70) we arrive at
˘ 1 ÿ 1
|(88)| ď 1 ` J 2 xφp0q, pN ` 1qφp0qy2
`
xΨd , qx Ψd y ` . (92)
|Λ| xPΛ d

27
For the pq-qp term of (89) we find, using Cauchy–Schwarz,
ˇ ˇ
ˇ ˇ
ˇ J 1 ÿ ÿ ˇ
|(89)| ď 2 ˇ´ xax px qy Ψd , ay py qx Ψd yˇ
ˇ ˇ
ˇ 2d |Λ| ˇ
ˇ xPΛ yPΛ ˇ
x„y

J 1 ÿ ÿ
ď xφp0q, N φp0qy}qx Ψd }}qy Ψd }
d |Λ| xPΛ yPΛ
x„y

1 ÿ
ď 2Jxφp0q, N φp0qy xΨd , qx Ψd y.
|Λ| xPΛ

The pq-qq terms of (90) have already been estimated in the proof of Proposition 13. We find it
convenient to introduce ϵ ą 0, so continuing from (71) and using Cauchy–Schwarz, we find

J2
ˆ ˙
1 ÿ 1 ÿ
|(90)| ď 2 3J ` ` 4Jxφp0q, N φp0qy xΨd , qx Ψd y ` 2ϵ xΨd , qx Nx2 qx Ψd y. (93)
ϵ |Λ| xPΛ |Λ| xPΛ

Finally, the terms involving a˚ a in (91) can be estimated directly with Cauchy–Schwarz. We find
ˇ ˇ
ˇ ˇ
ˇ J 1 ÿ ÿ ˇ
xΨd , qx qy a˚x ay qx qy Ψd yˇ
ˇ ˇ
ˇ´
ˇ 2d |Λ| ˇ
ˇ xPΛ yPΛ ˇ
x„y

J 1 ÿ ÿ
b b
ď xΨd , qy qx Nx qx Ψd y xΨd , qx qy Ny qy Ψd y
2d |Λ| xPΛ yPΛ
x„y

1 1 ÿ ÿa
b
ď J}qx Ψd }}Nx qx Ψd } J}qy Ψd }}Ny qy Ψd }
2d |Λ| xPΛ yPΛ
x„y
˙1{2 ˆ 2 ˙1{2
1 1 ÿ ÿ J2
ˆ
2 ϵ 2 J 2 ϵ 2
ď }qx Ψd } ` }Nx qx Ψd } }qy Ψd } ` }Ny qy Ψd }
2d |Λ| xPΛ yPΛ 2ϵ 2 2ϵ 2
x„y
˜ ¸1{2 ˜ ¸1{2
J2 1 ÿ ϵ 1 ÿ J2 1 ÿ ϵ 1 ÿ
ď }qx Ψd }2 ` }Nx qx Ψd }2 }qx Ψd }2 ` }Nx qx Ψd }2
2ϵ |Λ| xPΛ 2 |Λ| xPΛ 2ϵ |Λ| xPΛ 2 |Λ| xPΛ
J2 1 ÿ ϵ 1 ÿ
ď xΨd , qx Ψd y ` xΨd , qx Nx2 qx Ψd y.
2ϵ |Λ| xPΛ 2 |Λ| xPΛ

The terms involving αφ can be estimated in the same way, using additionally that
a
|αφ ptq| “ |xφptq, aφptqy| ď }aφptq} “ xφp0q, N φp0qy. (94)

Combining these bounds yields

1 J2
ˆ ˙
2 1 ÿ 1 ÿ
|(91)| ď C ` ` J xφp0q, N φp0qy xΨd , qx Ψd y ` Cϵ xΨd , qx Nx2 qx Ψd y. (95)
ϵ ϵ |Λ| xPΛ |Λ| xPΛ

Thus, altogether, we get for some C ą 0 that

ˇ ˇ
ˇ 1
ˇ xΨd , H̃Ψd yˇ ď Cϵ 1 1
ˇ ÿ
xΨd , qx Nx2 qx Ψd y `
ˇ |Λ| ˇ |Λ| xPΛ d
ˆ ˙ (96)
` 2
˘ 1 2 1 ÿ
`C 1`J 1 ` ` xφp0q, N φp0qy ` xφp0q, N φp0qy xΨd , qx Ψd y.
ϵ |Λ| xPΛ

28
Similarly, we can use Cauchy–Schwarz and again (94) to show that for some C ą 0,
ˇ ˆ ˙ ˇ
ˇ 1 ÿ U 2 ˇ
αφ
, q h N q Ψ
ˇ ˇ
ˇ xΨ d x x ´ x y
x d ˇ
ˇ |Λ|
xPΛ
2 ˇ
ˇ Fˇˇ
ˇ 1 ÿB ˆ ˆ
U
˙ ˙
˚ 2
Ψd , qx ´Jαφ ax ´ Jαφ ax ` J|αφ | ` J ´ µ ´ Nx qx Ψd ˇ
ˇ ˇ
“ˇ
ˇ |Λ| 2 ˇ
˜ xPΛ ¸ (97)
U 2
ˆ ˙ ˆ ˙
2 1 1 ÿ
ďC 1`J ` J ´µ´ ` xφp0q, N φp0qy xΨd , qx Ψd y
2 ϵ |Λ| xPΛ
1 ÿ
`ϵ xΨd , qx Nx2 qx Ψd y.
|Λ| xPΛ

Combining both bounds yields (87).

This proposition allows us to prove the equivalence of f and g up to an error d´1 .

Proposition 17. Let f and g be defined as in (85) and (86). Then, for U ą 0 and for some C ą 0,
we have the equivalence
˜ ˙ ¸ˆ
U 2
ˆ ˙
U 1 2 1 2 1
g´ ďf ďC 1`J `U ` J ´µ´ 1 ` ` xφp0q, N φp0qy g ` . (98)
4 d 2 U d

Proof. We start by proving the lower bound on f from (98). From Proposition 16 we know that
C ˜ ˙ ¸ G
1 ÿ ˆ αφ U
2
Ψd , H̃ ` qx hx ´ Nx qx Ψd
|Λ| xPΛ
2
˜ ˙ ¸ˆ
U 2
ˆ ˙
2 1 1 ÿ 1
ě ´C 1 ` J ` J ´ µ ´ 1 ` ` xφp0q, N φp0qy2 xΨd , qx Ψd y ´ (99)
2 ϵ |Λ| xPΛ d
1 ÿ
´ϵ xΨd , qx Nx2 qx Ψd y.
|Λ| xPΛ

Hence,
C ˜ ˆ ˙ ¸ G
1 ÿ α U
f“ Ψd , H̃ ` qx hx φ ´ Nx2 qx Ψd
|Λ| xPΛ
2
U 1 ÿ c ÿ
` xΨd , qx Nx2 qx Ψd y ` xΨd , qx Ψd y
2 |Λ| xPΛ |Λ| xPΛ
ˆ ˙
U 1 ÿ
ě ´ϵ xΨd , qx Nx2 qx Ψd y
2 |Λ| xPΛ
˜ ˜ ˆ ˙2 ¸ ˆ ˙¸
U 1 1 ÿ 1
` c ´ C 1 ` J2 ` J ´ µ ´ 1 ` ` xφp0q, N φp0qy2 xΨd , qx Ψd y ´ .
2 ϵ |Λ| xPΛ d

Then the lower bound on f from (98) follows by choosing

˜ ˆ ˙2 ¸ ˆ ˙
2 U 1 2 U U
c“C 1`J ` J ´µ´ 1 ` ` xφp0q, N φp0qy ` , ϵ“ . (100)
2 ϵ 4 4

29
For the upper bound on f from (98), note that
ˇ ˇ
ˇ 1 ÿ ˇ
αφ
xΨd , qx hx qx Ψd yˇ
ˇ ˇ
ˇ
ˇ |Λ| ˇ
xPΛ
ˇ B ˆ ˙ Fˇˇ
ˇ 1 ÿ U
Ψd , qx ´Jαφ a˚x ´ Jαφ ax ` J|αφ |2 ` pJ ´ µqNx ` Nx pNx ´ 1q qx Ψd ˇ (101)
ˇ ˇ
“ˇ
ˇ |Λ| 2 ˇ
ˆ xPΛ ˇ ˇ˙
ˇ Uˇ 1 ÿ
ď C 1 ` |J| ` U ` ˇˇJ ´ µ ´ ˇˇ p1 ` xφp0q, N φp0qyq xΨd , pqx Nx2 qx ` qx qΨd y.
2 |Λ| xPΛ

Using this and the bound (96) from the proof of Proposition 16 for ϵ “ 1, the choice (100) for the
constant c yields
ˇ ˇ
ˇ 1 1 ÿ c ÿ ˇ
αφ
|f | “ ˇ xΨd , H̃Ψd y ` xΨd , qx hx qx Ψd y ` xΨd , qx Ψd yˇ
ˇ ˇ
ˇ |Λ| |Λ| xPΛ |Λ| xPΛ ˇ
˜ ˆ ˙2 ¸ ˆ ˙
2 U 1 2 1 ÿ
ďC 1`J `U ` J ´µ´ 1 ` ` xφp0q, N φp0qy xΨd , pqx Nx2 qx ` qx qΨd y
2 U |Λ| xPΛ
1
` .
d

6.2 Proof of Gronwall Estimate for f

In the computation of the time derivative of f we need to control in particular H, 9̃ the time derivative
of H̃. Its computation in straightforward but a bit lengthy. The key point is to write this time
α α
derivative in such a way that it contains the commutator rH̃, qx hx φ qx ´ hx φ s, which we will later use
for cancellations.
Proposition 18. The expectation of H9̃ can be written as
1 9̃ y “ ´ i ÿ xΨ , rH̃, q hαφ q ´ hαφ sΨ y ` R
xΨd , HΨ d d x x x x d
|Λ| |Λ| xPΛ
(102)
J i ÿ α α α α
“ xΨd , rH̃x,y , qx hx φ qx ´ hx φ ` qy hy φ qy ´ hy φ sΨd y ` R,
2d |Λ| ăx,yą
ř
with H̃x,y refers to the terms in (81) such that H̃ “ ăx,yą H̃x,y and where the rest term R ” Rptq is
given by
J i ÿ A ´
α p2q α p2q
¯ E
R :“ ´ Ψd , px hx φ py Kx,y qx qy ` qx hx φ px qy Kx,y px py Ψd ` h.c. (103)
d |Λ| ăx,yą
J i ÿ A ´
α α
¯
p2q
E
´ Ψd , qy px hx φ ` hy φ py Kx,y qx py Ψd ` h.c. (104)
d |Λ| ăx,yą
J i ÿ A ´
α α α p3q
` Ψd , qy phx φ px ´ px hx φ ´ px hy φ py qKx,y
d |Λ| ăx,yą
¯ E (105)
p3q αφ αφ
` px Kx,y ppx hx ` py hy q qx qy Ψd ` h.c.
J i ÿ p4q α
` xΨd , qx qy Kx,y px hx φ qx qy Ψd y ` h.c. (106)
d |Λ| ăx,yą
B ˆ ˙ F
J 1 ÿ 9 p3q 1 9 p4q
´ Ψd , px qy Kx,y qx qy ` qx qy Kx,y qx qy Ψd ` h.c.. (107)
2d |Λ| ăx,yą 2

30
Proof. We start by gathering some useful computations,

α9 φ “ iµαφ ´ iU xφ, N aφy , (108)

αφ α9 φ ` αφ α9φ “ 2U Im pxφ, N aφyαφ q , (109)
α
h9 x φ “ ´J α9 φ a˚x ´ J α9φ ax ` 2JU Im pxφ, N aφyαφ q , (110)
K9 p2q “ 0,
x,y (111)
p3q
K9 x,y “ ´α9 φ a˚x ´ α9φ ax , (112)
p4q
K9 x,y “ ´α9 φ pa˚x ` a˚y q ´ α9φ pax ` ay q ` 4U Im pxφ, N aφyαφ q . (113)
α α
Starting from the definition (81) of H̃, using these relations and ip9x “ rhx φ , px s, iq9x “ rhx φ , qx s, we
arrive at
ˆ
9̃ iJ ÿ α p2q α p2q p2q αφ p2q α
H“ rhx φ , px spy Kx,y qx qy ` px rhy φ , py sKx,y qx qy ` px py Kx,y rhx , qx sqy ` px py Kx,y qx rhy φ , qy s
2d ăx,yą
α p2q α p2q p2q α p2q α
` rhx φ , px sqy Kx,y qx py ` px rhy φ , qy sKx,y qx py ` px qy Kx,y rhx φ , qx spy ` px qy Kx,y qx rhy φ , py s
α p2q α p2q p2q α p2q α
` rhx φ , qx sqy Kx,y px py ` qx rhy φ , qy sKx,y px py ` qx qy Kx,y rhx φ , px spy ` qx qy Kx,y px rhy φ , py s
˙
αφ p2q αφ p2q p2q αφ p2q αφ
` rhx , qx spy Kx,y px qy ` qx rhy , py sKx,y px qy ` qx py Kx,y rhx , px sqy ` qx py Kx,y px rhy , qy s
ˆ
iJ ÿ α p3q α p3q p3q αφ p3q α
` rhx φ , px sqy Kx,y qx qy ` px rhy φ , qy sKx,y qx qy ` px qy Kx,y rhx , qx sqy ` px qy Kx,y qx rhy φ , qy s
d ăx,yą
˙
αφ p3q αφ p3q p3q αφ p3q αφ
` rhx , qx sqy Kx,y px qy ` qx rhy , qy sKx,y px qy ` qx qy Kx,y rhx , px sqy ` qx qy Kx,y px rhy , qy s
ˆ ˙
iJ ÿ αφ p4q αφ p4q p4q αφ p4q αφ
` rhx , qx sqy Kx,y qx qy ` qx rhy , qy sKx,y qx qy ` qx qy Kx,y rhx , qx sqy ` qx qy Kx,y qx rhy , qy s
2d ăx,yą
„ ȷ
J ÿ ` ˚
˘ ` ˚
˘
` px qy α9 φ ax ´ α9φ ax qx qy ` qx qy α9 φ ax ´ α9φ ax px qy
d ăx,yą
ˆ ˙
J ÿ ˚ ˚ 9
` qx qy α9 φ pax ` ay q ` αφ pax ` ay q ´ 4U Impxφ, N aφyαφ q qx qy .
2d ăx,yą

To obtain (102), we isolate the first part on the right-hand side of (102) and define the rest as the
remainder term R.

Next, we estimate the rest term in Proposition 18.

Proposition 19. The rest term Rptq in Proposition 18 satisfies the bound
˜ ¸
1 ÿ 1
|Rptq| ď C̃ptq xΨd ptq, pqx ptqNx qx ptq ` qx ptqqΨd ptqy ` ,
|Λ| xPΛ d

where
C̃ptq “ CpJ, µ, U q 1 ` xφp0q, N φp0qy2
` ˘

6 ´ E tj ¯
´ ÿ ¯j A j (114)
1` 8Jxφp0q, N φp0qy1{2 φp0q, pN ` jq4´ 2 φp0q ,
j“0
j!

with CpJ, µ, U q ą 0 depending polynomially on the model parameters J, µ and U .

31
Proof. We need to estimate each term in R. We start by explaining in detail how to estimate one of
the terms in (103). By Cauchy–Schwarz and Hölder’s inequality we have
ˇ ˇ
ˇ J ÿ ÿ αφ ˚
ˇ
ˇ
ˇ d|Λ| xΨd , px hx py ax ay qx qy Ψd yˇˇ
xPΛ yPΛ
x„y
› ˜ ¸ ›
J ÿ ›› ˚ ››› αφ ÿ
˚
›
ax qx Ψd ››hx px qy ay py Ψd ›
›
ď
d|Λ| xPΛ ›
yPΛ
›
x„y
˜ ¸1
2
2J ÿ a a › ˚ › 1 1 ÿ ÿ
ď xφp0q, pN ` 1qφp0qy xφ, phαφ q2 φy›ax qx Ψd › ` }qz Ψd }}qy Ψd }
|Λ| xPΛ 2d p2dq2 yPΛ zPΛ
y­“z,x„y x„z
˜ ¸1
2
2J ÿ a a › ˚ › 1 1 ÿ 1 ÿ
ď α 2
xφp0q, pN ` 1qφp0qy xφ, ph φ q φy›ax qx Ψd › ` }qy Ψd }2 ` }qz Ψd }2
|Λ| xPΛ 2d 4d yPΛ
4d zPΛ
y­“z,x„y x„z
˜
a a 1 ÿ ˚ 1
ď CJ xφp0q, pN ` 1qφp0qy xφ, phαφ q2 φy }ax qx Ψd }2 `
|Λ| x d
¸
1 1 ÿ ÿ 1 1 ÿ ÿ
` }qy Ψd }2 ` }qz Ψd }2
d |Λ| xPΛ yPΛ d |Λ| xPΛ zPΛ
x„y x„z
loooooooomoooooooon loooooooomoooooooon
ř
“2d
ř
}qx Ψd }2 “2d xPΛ }qx Ψd }2
xPΛ
˜ ¸
a a 1 ÿ 1
ď CJ xφp0q, pN ` 1qφp0qy xφ, phαφ q2 φy xΨd , pqx Nx qx ` qx qΨd y ` .
|Λ| xPΛ d

The other terms of (103) can be estimated analogously, so we arrive at

˜ ¸
a 1 ÿ 1
|(103)| ď CJ p1 ` xφp0q, pN qφp0qyq xφ, phαφ q2 φy xΨd , pqx Nx qx ` qx qΨd y ` .
|Λ| xPΛ d

In order to bound (104), we use Cauchy–Schwarz to find

ˇ ˇ
ˇ ˇ
ˇJ 1 ÿ ÿ ˇ
αφ ˚
xΨd , qy px hx ax ay qx py Ψd yˇ
ˇ ˇ
ˇ
ˇ d |Λ| ˇ
ˇ xPΛ yPΛ ˇ
x„y
ˇ ˇ
ˇ ˇ
ˇJ 1 ÿ ÿ ˇ
αφ ˚ ˚
xhx px ay qy Ψd , ax qx py Ψd yˇ
ˇ ˇ
“ˇ ˇ
ˇ d |Λˇ ˇ
ˇ xPΛ yPΛ ˇ
x„y

J 1 ÿ ÿa ´
αφ q2 φy }q Ψ }2 ` }N 1{2 q Ψ }2
¯1{2 ´
2 1{2 2
¯1{2
ď ˇ xφ, ph x d x x d }q y Ψd } ` }N y q y Ψd }
d |Λˇ xPΛ yPΛ
x„y
a 1 ÿ
ď CJ xφ, phαφ q2 φy xΨd , pqx Nx qx ` qx qΨd y.
|Λ| xPΛ

Estimating the other terms of (104) in an analogous way, we get

´ a ¯a 1 ÿ
|(104)| ď CJ 1 ` xφp0q, pN qφp0qy xφ, phαφ q2 φy xΨd , pqx Nx qx ` qx qΨd y.
|Λ| xPΛ

32
To bound (105), we use Cauchy–Schwarz to estimate
ˇ ˇ
ˇ ˇ
ˇJ 1 ÿ ÿ ˇ
αφ ˚
xΨd , qy hx px ax ay qx qy Ψd yˇ
ˇ ˇ
ˇ
ˇ d |Λ| ˇ
ˇ xPΛ yPΛ ˇ
x„y
ˇ ˇ
ˇ ˇ
ˇJ 1 ÿ ÿ ˇ
˚ αφ ˚
xay qy Ψd , hx px ax qx qy Ψd yˇ
ˇ ˇ
“ˇ
ˇ d |Λ| ˇ
ˇ xPΛ yPΛ ˇ
x„y

J 1 ÿ ÿ ˚ ´ ¯1
αφ 2 ˚ 2
ď }ay qy Ψd } xΨd , qx ax p ph
x x q p x a q Ψ
x x d y
d |Λ| xPΛ yPΛ looooomooooon
x„y “xφ,phαφ q2 φypx

J 1 ÿ ÿa ´ ¯1 ´ ¯1
2 2
ď xφ, phαφ q2 φy xΨd , qy pNy ` 1qqy Ψd y xΨd , qx pNx ` 1qqx Ψd y
d |Λ| xPΛ yPΛ
x„y
a 1 ÿ
ď CJ xφ, phαφ q2 φy xΨd , pqx Nx qx ` qx qΨd y,
|Λ| xPΛ
and similarly
ˇ ˇ
ˇ ˇ
ˇJ 1 ÿ ÿ ˇ
αφ ˚
, q p h p a a q q Ψ
ˇ ˇ
ˇ xΨd y x y y
y x y x y d ˇ
ˇ d |Λ| ˇ
ˇ xPΛ yPΛ ˇ
x„y

G 1{2
¨ ˛
C
J 1 ÿ ÿa αφ 2
xφp0q, N φp0qy}qy Ψd } ˝ Ψd , qx qy a˚y p y phy q py ay qy Ψd ‚
˚ ‹
ď
d |Λ| xPΛ yPΛ looooomooooon
x„y “xφ,phαφ q2 φypy
J 1 ÿ ÿa a `@ D˘1{2
“ xφ, phαφ q2 φy xφp0q, N φp0qy}qy Ψd } Ψd , qx qy a˚y py ay qy Ψd
d |Λ| xPΛ yPΛ
x„y
a a 1 ÿ
ď CJ xφp0q, N φp0qy xφ, phαφ q2 φy xΨd , pqx Nx qx ` qx qΨd y.
|Λ| xPΛ
For (106), we directly find
ˇ ˇ ˇ ˇ
ˇJ 1 ÿ ˇ ˇJ 1 ÿ ˇ
˚ αφ αφ
xΨ , q q a a p hx qx qy Ψd yˇ “ ˇ xhx px ax qx qy Ψd , ay qx qy Ψd yˇ
ˇ ˇ ˇ ˇ
ˇ d |Λ| ăx,yą d x y x y x
ˇ
ˇ ˇ d |Λ| ăx,yą ˇ
a 1 ÿ
ď CJ xφ, phαφ q2 φy xΨd , qx Nx qx Ψd y.
|Λ| xPΛ
Estimating the other terms in an analogous way, we obtain
a 1 ÿ
|(104) ` (105) ` (106)| ď CJ p1 ` xφp0q, N φp0qyq xφ, phαφ q2 φy xΨd , pqx Nx qx ` qx qΨd y.
|Λ| xPΛ
Using ´ a ¯
9 ď C |µ| xφp0q, N φp0qy ` U xφ, pN ` 1q3{2 φy
|α|
a
|α9φ αφ ` αφ α9 φ | ď CU xφp0q, N φp0qyxφ, pN ` 1q3{2 φy,
we furthermore find
ˆ b ˙
a 3{2
a
3{2
|(107)| ďCJ |µ ` U | xφp0q, N φp0qy ` U xφ, N φy ` U xφp0q, N φp0qy xφ, pN ` 1q φy
1 ÿ ` ˘
xΨd , qx Nx qx ` qx Ψd y,
|Λ| xPΛ

33
Combining all estimates, we arrive at the bound
˜ ¸
a 1 ÿ 1
|R| ď CpJ, µ, U q p1 ` xφp0q, N φp0qyq xφ, phαφ q2 φy xΨd , pqx Nx qx ` qx qΨd y ` ,
|Λ| xPΛ d

where CpJ, µ, U q is a polynomial in J, µ and U . We also have by Cauchy–Schwarz

C ˆ ˙2 G
U
xφ, phαφ q2 φy “ φ, ´J αφ a˚ ` αφ a ´ |αφ |2 ` pJ ´ µqN ` N pN ´ 1q φ
` ˘
2 (115)
ď CpJ, µ, U q 1 ` xφp0q, N φp0qy2 1 ` xφ, N 4 φy .
` ˘ ` ˘

The proposition is proven by using the propagation bound (56) from Proposition (10) for k “ 4, since
then a ´ a ¯
xφ, phαφ q2 φy ď CpJ, µ, U q p1 ` xφp0q, N φp0qyq 1 ` xφ, N 4 φy
ď CpJ, µ, U q p1 ` xφp0q, N φp0qyq
´ ¯j (116)
8Jxφp0q, N φp0qy1{2 t A
˜ ¸
6 E
4´ 2j
ÿ
1` φp0q, pN ` jq φp0q .
j“0
j!

With this proposition we can now prove a Gronwall estimate for f .

Proposition 20. For f as defined in (85), we have for all t P R,

żt´
şt 1 ¯ şt
f ptq ď e 0 C̃psqds f p0q ` 1 ` C̃psq e s C̃prqdr ds, (117)
d 0

with ˆ ˙
CpJ, µ, U q 1 2
C̃ptq “ 1 ` ` xφp0q, N φp0qy
U U
˜
6 ´ ¯j A E tj
¸ (118)
ÿ j
1` 8Jxφp0q, N φp0qy1{2 φp0q, pN ` jq4´ 2 φp0q ,
j“0
j!

where CpJ, µ, U q ą 0 is a polynomial in J, µ and U .

α 9̃ the time derivative of f can be computed as
Proof. Using H9 d “ 0 “ xPΛ h9 x φ ` H,
ř

i ÿ α α 1 ÿ α α
f9 “ xΨd , rHd , Hd ` qx hx φ qx ´ hx φ ` cqx sΨd y ` xΨd , pqx h9 x φ qx ´ h9 x φ qΨd y
|Λ| xPΛ |Λ| xPΛ
i ÿ α α
´ xΨd , rhx φ , qx hx φ qx ` cqx sΨd y
|Λ| xPΛ
i ÿ α α 1 ÿ α 9̃
“ xΨd , rH̃, qx hx φ qx ´ hx φ ` cqx sΨd y ` xΨd , pqx h9 x φ qx ` HqΨ d y.
|Λ| xPΛ |Λ| xPΛ

Using Proposition 18 we get

i ÿ 1 ÿ α
f9 “ xΨd , rH̃, cqx sΨd y ` xΨd , qx h9 x φ qx Ψd y ` R.
|Λ| xPΛ |Λ| xPΛ

34
For the first two terms of this expression, we find
ˇ ˇ
ˇ 1 ÿ ˇ
α
xΨd , qx h9 x φ qx Ψd yˇ
ˇ ˇ
ˇ
ˇ |Λ| ˇ
xPΛ
ˇ ˇ
ˇ 1 @ ` ˚
˘ Dˇ
“ˇ ˇ 9
Ψd , qx ´J α9 φ ax ´ J αφ ax ` 2JU Impxφ, N aφyαφ q qx Ψd ˇˇ
|Λ|
´ a ¯ 1 ÿ@
Ψd , qx Nx2 qx ` qx Ψd ,
` ˘ D
ď CpJ, µ, U q p1 ` xφp0q, N φp0qyq 1 ` xφ, N 2 φy
|Λ| xPΛ

and ˇ ˇ
ˇ 1 ÿ ˇ
xΨd , rH̃, cqx sΨd yˇ
ˇ ˇ
ˇ
ˇ |Λ| ˇ
xPΛ
ˆ ˙
1 2 1 ÿ@ D 1
ď CpJ, µ, U q 1 ` ` xφp0q, N φp0qy Ψd , pqx Nx2 qx ` qx qΨd ` .
U |Λ| xPΛ d
These two estimates, together with the estimate on R from Proposition 19 and the equivalence of f
and g up to an error d´1 from Proposition 17 imply
ˇ ˇ ˆ ˙
ˇd
ˇ f ptqˇ ď C̃ptq f ptq ` 1 ` 1 ,
ˇ
ˇ dt ˇ d d

where C̃ptq depends on the initial data and on the other parameters of our model as defined in (118).
With Gronwall’s lemma we arrive at (117).

6.3 Conclusion of the Proof

We combine the above results to prove our second main result.

Proof of Theorem 2. We use the equivalence of f and g up to an error d´1 from Proposition 17, and
the Gronwall estimate for f from Proposition 20 to find
1 ÿ
xΨd , qx Ψd y
|Λ| xPΛ
1 ÿ
xΨd , qx Nx2 qx ` qx Ψd y
` ˘
ď
|Λ| xPΛ
ˆ ˙
4 1
ď f`
U d
4 t´
ˆ ż ˙
4 t C̃psqds
ş 1 4 ¯ şt
C̃prqdr
ď e0 f p0q ` ` 1 ` C̃psq e s ds
U d U U 0
˜ ˙ ¸ˆ
U 2
ˆ ˙ ş
C 2 1 t
ď 1`J `U ` J ´µ´ 1 ` ` xφp0q, N φp0qy e 0 C̃psqds
2
U 2 U
ˆ żt ˙
1 ÿ@ 2
D 14 şt
C̃psqds
şt
C̃prqdr
Ψd p0q, pqx p0qNx qx p0q ` qx p0qqΨd p0q ` 1`e 0 ` p1 ` C̃psqqe s ds ,
|Λ| xPΛ dU 0
(119)
where C̃ptq is defined in (118). Now note that since Trppp0qN 4 q ď C, we get that C̃ptq satisfies
˜ ¸
ÿ6
j
C̃ptq ď CpJ, µ, U q 1 ` t
j“1

35
where CpJ, µ, U q ą 0 depends polynomially on the parameters of our model J, µ and U . Thus, (119)
can be estimated as
1 ÿ
xΨd ptq, qx ptqΨd ptqy
|Λ| xPΛ
ˆ ˙˜ ¸
11 ř7 j 1 1 ÿ@ 1
` CpJ, µ, U qeCpJ,µ,U q 1 |t| 1 ` 2 Ψd p0q, qx p0qNx2 qx p0q ` qx p0q Ψd p0q `
` ˘ D
ď ,
dU U |Λ| xPΛ d
(120)
and the Theorem follows from using (66) from Lemma 12.

References
[1] M. Aizenman, E. Lieb, R. Seiringer, J. Solovej, and J. Yngvason. Bose–Einstein quantum phase
transition in an optical lattice model. Phys. Rev. A, 70:023612, 2004.

[2] M. H. Anderson, J. R. Ensher, M. R. Matthews, C. E. Wieman, and E. A. Cornell. Observation

of Bose–Einstein condensation in a dilute atomic vapor. Science, 269(5221):198–201, 1995.

[3] N. Benedikter, M. Porta, and B. Schlein. Effective Evolution Equations from Quantum Dynamics.
SpringerBriefs in Mathematical Physics. Springer, 2016.

[4] I. Bloch, J. Dalibard, and W. Zwerger. Many-body physics with ultracold gases. Rev. Mod. Phys.,
80:885–964, 2007.

[5] C. Boccato, C. Brennecke, S. Cenatiempo, and B. Schlein. Bogoliubov theory in the Gross–
Pitaevskii limit. Acta Math., 222(2):219–335, 2019.

[6] Bose. Plancks Gesetz und Lichtquantenhypothese. Z. Phys., 26:178–181, 1924.

[7] L. Boßmann, S. Petrat, P. Pickl, and A. Soffer. Beyond Bogoliubov dynamics. Pure Appl. Anal.,
3(4):677–726, 2021.

[8] L. Boßmann, S. Petrat, and R. Seiringer. Asymptotic expansion of low-energy excitations for
weakly interacting bosons. Forum Math. Sigma, 9:e28, 2021.

[9] J.-B. Bru and T. C. Dorlas. Exact solution of the infinite-range-hopping Bose–Hubbard model.
J. Stat. Phys., 113, 2003.

[10] K. Byczuk and D. Vollhardt. Correlated bosons on a lattice: Dynamical mean-field theory for
Bose–Einstein condensed and normal phases. Phys. Rev. B, 77(235106), 2008.

[11] T. Cazenave and A. Haraux. An Introduction to Semilinear Evolution Equations. Oxford lecture
series in mathematics and its applications, 1998.

[12] K. B. Davis, M.-O. Mewes, M. R. Andrews, N. J. van Druten, D. S. Durfee, D. M. Kurn, and
W. Ketterle. Bose–Einstein condensation in a gas of sodium atoms. Phys. Rev. Lett., 75(22):3969–
3973, 1995.

[13] J. Dereziński and M. Napiórkowski. Excitation spectrum of interacting bosons in the mean-field
infinite-volume limit. Ann. Henri Poincaré, 15(12):2409–2439, 2014.

[14] A. Einstein. Quantentheorie des einatomigen idealen Gases. Sitzber. Kgl. Preuss. Akad. Wiss.,
pages 261–267, 1924.

[15] A. Einstein. Quantentheorie des einatomigen idealen Gases. Zweite Abhandlung. Sitzber. Kgl.
Preuss. Akad. Wiss., pages 3–14, 1925.

36
[16] L. Erdős and H.-T. Yau. Derivation of the nonlinear Schrödinger equation from a many-body
Coulomb system. Adv. Theor. and Math. Phys., 5(6):1169–1205, 2001.

[17] M. P. A. Fisher, P. B. Weichman, G. Grinstein, and D. S. Fisher. Boson localization and the
superfluid-insulator transition. Phys. Rev. B, 40:546–570, Jul 1989.

[18] S. Fournais and J. P. Solovej. The energy of dilute Bose gases. Ann. Math., 192:893–976, 2020.

[19] S. Fournais and J. P. Solovej. The energy of dilute Bose gases II: the general case. Invent. Math.,
232:863–994, 2022.

[20] A. Georges, G. Kotliar, W. Krauth, and M. Rozenberg. Dynamical mean-field theory of strongly
correlated fermion systems and the limit of infinite dimensions. Rev. Mod. Phys., 68:13–125, 01
1996.

[21] F. Golse. On the Dynamics of Large Particle Systems in the Mean Field Limit, chapter I, page
1–144. Springer International Publishing, 2016.

[22] M. Greiner, O. Mandel, T. Esslinger, T. W. Hänsch, and I. Bloch. Quantum phase transition
from a superfluid to a Mott insulator in a gas of ultracold atoms. Nature, 415:39–44, 2002.

[23] M. Greiner, O. Mandel, T. Rom, A. Altmeyer, A. Widera, T. Hänsch, and I. Bloch. Quantum
phase transition from a superfluid to a Mott insulator in an ultracold gas of atoms. Physica B:
Condensed Matter, 329-333:11–12, 2003. Proceedings of the 23rd International Conference on
Low Temperature Physics.

[24] M. C. Gutzwiller. Effect of correlation on the ferromagnetism of transition metals. Phys. Rev.
Lett., 10, 1963.

[25] W.-J. Hu and N.-H. Tong. Dynamical mean-field theory for the Bose–Hubbard model. Phys.
Rev. B, 80:245110, Dec 2009.

[26] D. Jaksch, C. Bruder, J. I. Cirac, C. W. Gardiner, and P. Zoller. Cold bosonic atoms in optical
lattices. Phys. Rev. Lett., 81:3108–3111, Oct 1998.

[27] A. Knowles and P. Pickl. Mean-field dynamics: singular potentials and rate of convergence.
Commun. Math. Phys., 298(1):101–138, 2010.

[28] M. Lewin, P. T. Nam, and B. Schlein. Fluctuations around Hartree states in the mean field
regime. Amer. J. Math., 137(6):1613–1650, 2015.

[29] M. Lewin, P. T. Nam, S. Serfaty, and J. P. Solovej. Bogoliubov spectrum of interacting Bose
gases. Commun. Pure Appl. Math., 68(3):413–471, 2015.

[30] W. Metzner and D. Vollhardt. Correlated lattice fermions in d “ 8 dimensions. Phys. Rev. Lett.,
62:324–327, 1989.

[31] E. Pavarini, E. Koch, A. Lichtenstein, and D. Vollhardt. Dynamical Mean-Field Theory of Cor-
related Electrons. Institute for Advanced Simulation, 2022.

[32] E. Picari, A. Ponno, and L. Zanelli. Mean field derivation of DNLS from the Bose-–Hubbard
model. Ann. Henri Poincaré, 23, 10 2021.

[33] P. Pickl. A simple derivation of mean field limits for quantum systems. Lett. Math. Phys.,
97(2):151–164, 2011.

[34] D. S. Rokhsar and B. G. Kotliar. Gutzwiller projection for bosons. Phys. Rev. B, 44, 1991.

37
[35] B. Schlein. Bose gases in the Gross–Pitaevskii limit: A survey of some rigorous results. In
R. Frank, A. Laptev, M. Lewin, and R. Seiringer, editors, The Physics and Mathematics of
Elliott Lieb, volume II, pages 277–305. EMS Press, 2022.

[36] B. Simon. Trace Ideals and Their Applications. Mathematical surveys and monographs, 2005.

[37] W. Zwerger. Mott–Hubbard transition of cold atoms in optical lattices. J. Opt. B, 5:9–16, 2003.

38