MATHEMATICAL ISSUES IN THE MODELING OF

SUPERSOLIDS
A. AFTALION, X. BLANC, AND R. L. JERRARD
Abstract. Supersolids are quantum crystals having superuid properties.
This state of matter, predicted by Leggett, was achieved experimentally
only recently in solid
4
He. In the present paper, we study a model of
supersolids introduced in [14, 15]. This model is based on a Gross-Pitaevskii
energy with a nonlocal interaction term. We give a description of the ground
state in the limit of strong interaction, and relate it to a sphere packing
problem. We prove that, in dimension one, this ground state is periodic.
1. Introduction
Superuidity is a state of matter at the origin of dierent properties: the
disappearance of viscosity and hence ow without friction, or nonclassical be-
haviour when the sample is placed in a rotating container. When specic uids
such as Helium 3 or 4, are cooled below a critical temperature, they undergo
a superuid transition [16, 23]. Leggett predicted that a superuid behaviour
can also occur in a crystal, and this is called supersolidity. In particular, he
gave arguments indicating that nonclassical eects under rotation could oc-
cur in such crystals [21]. Several authors [2, 27, 6] pointed out microscopic
mechanisms that could explain supersolidity. This was only theoretical predic-
tions until 2004 when Kim and Chan [17, 18, 19] achieved a very controversial
experiment in which they produced a solid state with superuid properties.
This was the discovery of a new quantum phase of matter, then conrmed
by several groups (see [20, 25, 28] and also [29] for frictionless ow in solid
helium).
On the theoretical side, several approaches were proposed to account for
supersolidity. One of them relies on many-body quantum mechanics and the
design of an adapted ansatz for the many-body wave function. In these ap-
proaches, one usually allows for vacancies or defects in the crystal, and in-
terpret them as quantum particles which undergo condensation [3, 28]. We
refer for instance to [26] for a review on this subject. Many issues are still
debated in the physics community concerning the link between these theories
and experiments.
Another approach is based on the Gross-Pitaevskii (GP) equation, as pro-
posed by Josserand, Pomeau and Rica in [14, 15]. They use the GP energy with
1
2 A. AFTALION, X. BLANC, AND R. L. JERRARD
a nonlocal interaction potential, so that the (numerically computed) ground
state has a crystalline structure. Then they use it to derive properties under
rotation, giving numerical evidence of nonclassical eects. They check that
this model also provides mechanical behaviours typical of solids under small
stress.
The aim of the present article is to set the GP approach of [14, 15] on
a sound mathematical ground. In particular, the question of the crystalline
symmetry of the ground state will be addressed. Then considerations of the
eect of small rotation will be discussed. Some results of this paper have been
announced in a note for physicists [1].
1.1. The nonlocal Gross-Pitaevskii energy. We will call T R
d
the set
occupied by the supersolid, with d = 3 being the most physically relevant
choice. The wave function u is complex-valued, and [u[
2
provides the density
of atoms. In [14, 15], in order to model a supersolid, the following Gross-
Pitaevskii energy is introduced for u:
(1.1) E

(u) =
_
D
1
2
[u[
2
dx +
1
4
2
F([u[
2
),
where the function F is dened on the space of measures on

T, by
(1.2) F() =
_

D
(V )(x) (dx) =
_

D
_

D
V (x y)(dy) (dx).
The nondimensional parameter accounts for the strength of the interaction
and the interacting potential V is chosen to be
(1.3) V (x) =
_
1 if [x[ < 1
0 if not.
This specic choice of V corresponds to putting a hard core for each atom and
provides a dispersion relation with a roton minimum [14]. If instead, V is taken
to be the delta function, the energy E

is simply the usual Gross Pitaevskii

energy with an L
4
term describing interactions. This model bears an important
dierence with classical solids, in the sense that in classical solids, there is an
integer number of atoms per unit cell, while in this quantum solid model, the
average density is a free number, independent of the crystal parameters, which
provides an L
2
constraint on the wave function.
We assume that T is a bounded, open subset of R
d
with Lipschitz bound-
ary. The space H
1
0
(T; C) is, as usual, the space of H
1
functions satisfying
homogeneous Dirichlet boundary conditions. The functional is minimized in
the space
/
0
(T) :=
_
u H
1
0
(T; C) :
_
D
[u[
2
dx = 1
_
,
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 3
where
_
D
dx denotes the average
1
|D|
_
D
dx.
Physicists have identied two regimes for the ground state: for large , the
minimizer is constant, corresponding to a uniform density. On the other hand,
for small , it is localized on periodic sets which are almost on a triangular
lattice for large domains [14, 15], corresponding to a crystalline state. This is
what we want to characterize mathematically, and this is why we also introduce
periodic boundary conditions.
In the periodic setting, the domain T is a parallelogram, so that there
exist n linearly independent vectors v
1
, . . . , v
n
such that
T =
_

x
i
v
i
: 0 x
i
< 1 for all i
_
,
and V is understood to be the T-periodic extension of the potential V dened
in (1.3). Or equivalently, V is exactly as in (1.3), and [x y[ is taken to be
the distance in the periodic sense, so that
(1.4)
[xy[ = inf
_

x y +

n
i
v
i

: n
i
Z for all i
_
in the periodic setting.
The space H
1
per
(T; C) denotes T-periodic functions in H
1
loc
(R
d
; C) and the
energy is then minimized in
/
per
(T) :=
_
u H
1
per
(T; C) :
_
D
[u[
2
dx = 1
_
.
Under rotation along the vertical axis, the energy in the rotating frame
becomes
(1.5) E
,
(u) = E

(u)
_
D
A Im( uu) dx,
where is the rotational velocity, and A(x, y) = (y, x) when d = 2 and
A(x, y, z) = (y, x, 0) when d = 3.
When is small, inf E
,
can be expanded as inf E

(1/2)I
2
, where
I is the eective moment of inertia of the solid. Leggett [21] suggested as
a criterion for supersolidity the existence of a nonclassical rotational inertia
fraction dened as (I
0
I)/I
0
, where I
0
is the classical moment of inertia of
the crystal phase. Thus, we need to understand and estimate the ground state
for = 0 and for small .
Our aim is to characterize the minimizers of E
,
. When is small, mini-
mizers of E
,
are expected to be almost minimizers of
G

(u) =
_
D
1
2
[u[
2
dx
_
D
A Im( uu) dx
4 A. AFTALION, X. BLANC, AND R. L. JERRARD
in the set of minimizers of F since the energy (1.5) reads
E
,
(u) = G

(u) +
1
4
2
F
_
[u[
2
_
and thus the two terms are of dierent order.
1.2. Ground state of F and sphere packing problem. If S is any mea-
surable subset of any Euclidean space, we will use the notation
/(S) := nonnegative measures on S : (S) = [S[,
/
ac
(S) := /(S) : L
d
,
where L
d
is the Lebesgue measure of R
d
and [S[ := L
d
(S).
Our rst result is a complete description of the minimizers of F in /
ac
,
which is related to a sphere packing problem:
Theorem 1.1. Suppose that T is a bounded, open subset of R
d
with Lipschitz
boundary, and that diamT > 1. Dene
(1.6) n(T) := maxk : x
1
, . . . , x
k
T such that [x
i
x
j
[ > 1 i ,= j.
Then
min
Mac(D)
F = [T[
2
/n(T),
and a measure in /
ac
(T) minimizes F if and only if there exist n(T) pair-
wise disjoint closed sets A
1
, . . . , A
n(D)


T, such that
(1.7) dist(A
i
, A
j
) 1 if i ,= j,
and
(1.8)
_
A
i
dx =
[T[
n(T)
for all i.
An important point due to the choice of the interaction potential is that
the self interaction of an A
i
is a constant on each set, which eventually gets
added to the energy.
When the number n(T) is large, the optimal location of the points x
i
in
(1.6) is proved [8] to be close to a hexagonal lattice in 2D. In 3D, 2 congura-
tions are optimal: body centered cubic close packing and face centered cubic
close packing.
Remark 1. A description of minimizers of F in /, rather than /
ac
is some-
what easier and will be obtained as a byproduct of the proof.
Remark 2. In fact the boundary of T can be quite irregular. The proof we
give works as long as L
d
(T) = 0.
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 5
We will denote by /

ac
(T) the set of minimizers of F in /
ac
:
(1.9) /

ac
(T) := /
ac
(T) : A
1
, . . . , A
n
T satisfying (1.7), (1.8) .
It is clear that /

ac
(T) is nonempty.
1.3. Ground state with rotation. We will prove that the ground states
of E
,
are almost given by the minimizers of G

restricted to the space of

minimizers of F. For that purpose, we dene
(1.10) E
0,
(u) :=
_
G

(u) if [u[
2
/

ac
(T),
+ if not.
This functional arises as a sort of 0 limit of E
,
. The next result makes
this limit precise. We dene
/

0
(T) :=
_
u H
1
0
(T; C) : [u[
2
/

ac
_
,
/

per
(T) :=
_
u H
1
per
(T; C) : [u[
2
/

ac
_
.
Theorem 1.2. If u

minimizes E
,
() = G

() +
1
4
2
F([ [
2
) in /
0
(T),
then u

(0,1]
is strongly precompact in H
1
(T), and the limit of any con-
vergent subsequence minimizes G

in /

0
(T). Similarly, if u

minimizes E
,
in /
per
(T), then u

(0,1]
is strongly precompact in H
1
(T), and the limit of
any convergent subsequence minimizes G

in /

per
(T).
In particular Theorem 1.2 implies that the problem
(1.11) nd u
0
/

0
(T) such that G

(u
0
) = min
A

0
(D)
G

has a solution. This is actually the rst step of the proof of Theorem 1.2. The
rest of the proof is based on energy estimates, which imply that minimizers of
E
,
are bounded in H
1
, hence converge weakly in H
1
. This allows to pass to
the limit in the energy and the constraint. Since the convergence of the energy
implies that
_
[u

[
2
converges, we have strong convergence in H
1
.
1.4. The one-dimensional case. In the special case of dimension one, it
is possible to compute explicitly the minimizers of F and the limit points of
minimizers of E

. Note that in such a case no angular momentum terms are

present, so we will consider the functional
(1.12) E

(u) =
_
D
1
2
[u

[
2
+
1
4
2
(V [u[
2
)[u[
2
dx.
and its 0 limit
(1.13) E
0
(u) :=
_
_
D
1
2
[u

[
2
if [u[
2
/

ac
(T),
+ if not.
6 A. AFTALION, X. BLANC, AND R. L. JERRARD
Figure 1. The minimizer u
p
introduced in Proposition 1.1 The
bumps are of size h
p
and separated by a distance 1.
in the spaces /
per
(T) and /
0
(T), where T R is an open interval. We will
write G
0
(u) for the Dirichlet energy
_
1
2
[u

[
2
dx.
We assume that T = (0, L) for some L > 0, and consider the problem
(1.14) nd u
0
/

0
(T) such that E
0
(u
0
) = min
A

0
(D)
E
0
,
together with the corresponding periodic problem. We write L| to denote
the smallest integer that is greater than or equal to L.
Proposition 1.1. There is a unique nonnegative minimizer for problem (1.14)
with Dirichlet boundary conditions, and it is given explicitly by
u
0
(x) =
_

_
_
2L
h
0
n
0
sin(

h
0
(x x
i
)) if x (x
i
, x
i
+ h
0
)
for some i 0, . . . n
0
1
0 if not
where n
0
, h
0
and x
i
are dened by
n
0
= L|, h
0
= (L ( n
0
1))/ n
0
, x
i
= i(1 + h
0
).
Moreover, E
0
(u
0
) =
2
L/2h
2
0
.
Similarly, if we consider periodic boundary conditions, then there is a
positive minimizer that is unique modulo translations, and is given by
u
p
(x) =
_

_
_
2L
hp np
sin(

hp
(x x
i
)) if x (x
i
, x
i
+ h
p
)
for some i 0, . . . n
p
1
0 if not
where n
p
, h
p
and x
i
are dened by
n
p
= L| 1, h
p
= (L n
p
)/ n
p
, x
i
= i(1 + h
p
).
And E
0
(u
0
) =
2
L/2h
2
p
.
Note that for both the Dirichlet and the periodic problem, h
0
, h
p
are
bounded by C/L, so that the support of a minimizer is sharply concentrated
for L large. Note also that h
0
, h
p
tend to zero as L approaches an integer from
above.
Dimension one also allows for explicit computations of the next order in
the asymptotic expansion of E

. Indeed, we have the following

MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 7
Theorem 1.3. Assume that L > 2, and let T = [0, L] R with periodic
boundary conditions. Suppose that u

minimizes E

() in H
1
per
(T). Then there
exist C > c > 0 such that
E

(u
p
) C
2/5
E

(u

) E

(u
p
) c
2/5
.
The result in Theorem 1.3 is stated only in terms of energy, giving the
next order expansion in powers of . However, the proof indicates that the
support of the minimizer u

is concentrated at distance of order

2/5
from the
support of its limit u
p
, up to an error of size exp(C/). We consider only
the periodic case for simplicity, but similar results hold for the Dirichlet data.
Although we do not have results in dimension d 2, we believe that a
similar expansion for the energy and boundary layer should be true. However,
this would require a better understanding of the minimizers of F (that is
/

ac
(T)).
1.5. Large . Finally, we prove that in the case of weak interaction, that is,
when (and = 0), the minimizer is homogeneous (that is, [u[ = 1).
Note that this result holds only for the periodic boundary conditions, but is
valid in any dimension:
Proposition 1.2. Let T = [L/2, L/2]
d
with periodic boundary conditions.
If
(1.15) >

d
,
where
d
is the volume of the unit ball of R
d
, then
E
,0
(1) E
,0
(u)
for all u H
1
per
(T) such that
_
[u[
2
= [T[. (Here 1 denotes the constant
function.) Moreover, if equality holds then u = 1.
The proof of this result is based on the decomposition of u in Fourier
series, and on estimates of the energy in this basis.
1.6. Open problems. The main diculties, due to the nonlocal term in the
equation, are the uniqueness and periodicity of minimizers.
In the one-dimensional setting, we have an explicit expression of the limit
of u

as 0, and it is periodic, even with Dirichlet boundary conditions.

We also expect that, for any (or at least for small enough), the minimizer
of E

with periodic boundary conditions is periodic. The diculty one faces

in proving such a result is to have a uniqueness result for a nonlocal equation:
indeed, the Euler-Lagrange equation reads
u

+
1

2
_
V u
2
_
u =

u
8 A. AFTALION, X. BLANC, AND R. L. JERRARD
and we are not able to prove any kind of uniqueness or maximum principle.
This is also open in dimension 2 and 3. In the two-dimensional setting, it is
already a very dicult issue, related to the sphere packing problem [8, 11, 12]
and optimal partition and eigenvalue problems [5, 7, 13], to prove the unique-
ness (modulo symmetries) or periodicity of minimizers of the = 0 problem
(1.11), even when = 0. Apart from this problem, minimizing G

among
all minimizers of F should give information on the connected components A
i
.
For instance, their boundary is probably Lipschitz, and any boundary point
of A
i
should be at distance 1 from another A
j
. An interesting point of view in
proving this would be to consider the boundary of each A
j
as a free boundary,
and derive an Euler-Lagrange equation satised by A
j
.
Another question in the one-dimensional setting is to study the limit L
. This limit is justied by the fact that physically, the interaction length,
which is set to 1 here, is much smaller than the size of the sample. Thus, it
would be interesting to study the corresponding asymptotics. This is implicitly
what is done in [15]. As pointed out above, in this limit, the minimizers of
F become sums of Dirac masses, and the above results do not apply to this
case. The corresponding proofs would require a more detailed study of the
asymptotic expansion of the energy and of the boundary layer near each peak
of the density.
In the two-dimensional case, a dicult issue is the presence of vortices.
According to numerics [15], it seems that there are no visible vortices. How-
ever, the density is exponentially small between the A
j
. We give the proof
of this fact in dimension 1 (see Lemma 4.1 below), but it may be adapted
to dimension 2. Hence, investigating the presence of vortices in this region
will probably be very dicult, since one can add or remove a vortex with a
change in the energy of order e
C/
. However, proving that there is no vortex
in each A
j
is probably more tractable, and would use estimates on ground
states of Schrodinger operators with magnetic potential, in the spirit of [9, 10]
for instance.
1.7. Back to nonclassical rotational eects. We x > 0, d = 2, and
study the limit 0. The nonclassical rotational inertia fraction (NCRIF)
is dened by
NCRIF =
I I
0
I
0
,
where I is the moment of inertia of the system, that is the second derivative of
the energy with respect to , and I
0
is the classical moment of inertia, that is,
I
0
=
_
D
[x[
2
[u
0
[
2
, where u
0
is the minimizer of E
,0
. This minimizer satises an
elliptic equation, which allows to apply the Harnack inequality: u
0
is positive
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 9
in T. A perturbation argument then allows to prove that for small enough,
the same property holds for u
,
, the minimizer of E
,
, hence one can write
u
,
=
,
e
iS
,
.
If we assume uniqueness of the minimizer of E
,0
, which is likely to be a dicult
issue, as we have mentionned above, then we can derive an explicit expression
for NCRIF, since we can expand
,
and S
,
in powers of
2
, justify the
expansion and nd
NCRIF =
_
D
u
2
0
[S
0
x

[
2
_
D
[x[
2
u
2
0
=
inf
SH
1
(D)
_
D
u
2
0
[S x

[
2
_
D
[x[
2
u
2
0
.
In [1], we have derived the lower bound
NCRIF
1

2
16
e
2

2T/
.
This relies on the extra strong assumption that u
0
is periodic with a periodic
cell of diameter T, in order to apply the Harnack inequality and get
inf u
0

1

4
e

2T/
sup u
0
.
The issues of the uniqueness and periodicity are, to our opinion, a key point
in this analysis.
We give in Section 2 the proof of Theorem 1.1, together with a general-
ization to other interaction potentials. In Section 3, we prove the convergence
of minimizers in the limit 0 (Theorem 1.2). The one-dimensional case is
studied in details in Section 4, with the proof of Proposition 1.1 and Theo-
rem 1.3. Finally, Section 5 is devoted to the proof of Proposition 1.2.
2. Minimizers of the interaction
We give in this section the proof of Theorem 1.1. It covers both the
periodic case and the non-periodic case. The proofs are identical in the two
cases, the only dierence being that in the periodic case, [xy[ is understood
as in (1.4)
2.1. Preliminaries. In this subsection and part of the next, it is convenient
to work on a closed set; this is used in the proof of Lemma 2.2 below. We
will write such a set as

T, although in fact our arguments are valid for rather
general closed sets, not only for those that arise as the closure of open sets T
as in the statement of Theorem 1.1.
First we prove
10 A. AFTALION, X. BLANC, AND R. L. JERRARD
Lemma 2.1. The functional F as dened in (1.2), (1.3) is lower semicontin-
uous with respect to weak convergence of measures in /(

T), i.e.
if
k
/(T),
k
, then F() liminf
k
F(
k
)
Proof. The point is that
F() = ( )((x, y) : [x y[ < 1)
where denotes the product measure on

T

T. It is well-known (see [4]
Chapter 1, for example) that
k
if and only if
k

k
, and also
that if
k

k
then (O) liminf
k

k

k
(O) for every O that
is relatively open in

T

T. Since (x, y) : [x y[ < 1 is open, the lemma
follows.
Lemma 2.2. F has a minimizer in /(

T).
Proof. Let
k
/(

T) be a minimizing sequence such that F(
k
)
inf
M(

D)
F(). Standard facts about weak convergence of measures imply that
there exists a subsequence (still denoted by
k
) and a measure /(

T) such
that
k
. Then the previous lemma implies that F() inf
M
F().
Note that the above standard argument does not yield the existence of
minimizers in /
ac
(T), since this space is not closed with respect to weak
convergence of measures.
Next we nd optimality conditions:
Lemma 2.3. If minimizes F in /(

T), then
(2.1) [T[
1
F() = min
x

D
V (x)
and
(2.2) V = [T[
1
F() a. e.
(ie, the set S := x

T : F (x) > [T[
1
F() satises (S) = 0).
The same optimality conditions (2.1), (2.2) are satised by a measure
that minimizes F in /
ac
(T).
Proof. Let
0
minimize F in /(

T). Note that /(

T) is a convex set, so that
if
1
/(

T), then

=
1
+ (1 )
0
/(

T) for every [0, 1]. Thus
0 lim
0
1

(F(

) F(
0
)) =
_

D
V
0
(
1

0
),
In other words,
(2.3)
_

D
_
V
0

F(
0
)
[T[
_

1
0
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 11
for every
1
/(

T). In particular, taking
1
= [T[
x
for arbitrary x T, we
deduce that V
0
(x) [T[
1
F(
0
), and therefore that inf V
0
[T[
1
F(
0
).
Also, note that
_

D
_
V
0

F(
0
)
[T[
_

0
= F(
0
)
F(
0
)
[T[
_

0
= F(
0
) F(
0
) = 0.
However, we have shown above that V
0
[T[
1
F(
0
) 0, so we deduce
that V
0
(x) [T[
1
F(
0
) = 0 at
0
-a.e. x T.
Finally, if
0
minimizes F in /
ac
, then since /
ac
is convex, we nd as
before that
0
satises (2.3) for every
1
/
ac
. It is clear that V
0
is
continuous, so it follows that V
0

F(
0
)
|D|
everywhere. Then the remainder
of the proof is exactly as in the previous case.
2.2. Characterization of minimizers. In this section we present the proof
of Theorem 1.1. We rst provide a simpler proof with some (not so) restrictive
hypothesis in Section 2.2.1, then a full proof in Section 2.2.2
A trivial but useful fact is that
(2.4)
_

D
V (x) dx =
_

D
V (x y) dx (dy) =
__

D
V dx
__

D
(dy) = [T[
_

D
V dx.
Let us recall that n(T) is dened by (1.6). We also dene
(2.5) n(

T) = maxk : x
1
, . . . , x
k


T such that [x
i
x
j
[ 1 i ,= j.
From these denitions, it is clear that n(T) n(

T). In many cases, we have
in fact equality: for example, in dimension one, T = (0, L), and equality holds
i L is not an integer. In such a case, n(T) = n(

T) = [L] + 1, where [L] is
the integer part of L. On the contrary, if n is an integer, then n(T) = L and
n(

T) = L + 1.
2.2.1. The case n = n.
Proof of Theorem 1.1. This proof is valid only in the case n(T) = n(

T).
Throughout the proof we will write B(x) to denote the open ball B(x, 1)
of radius 1 centered at x.
Given a measure /
ac
(T), we will extend it to a measure on

T, still
denoted by , by specifying that (A) = (A T) for A

T. Thus we can
identify /
ac
(T) with a subset of /(

T).
Step 1: upper bound. Let /

ac
(T) be dened in (1.9), so that we can
associate to a family of sets A
1
, . . . , A
n
satisfying (1.7) and (1.8). Then (1.7)
immediately implies that
(2.6) x A
i
, B(x)
_

n
j=1
A
j
_
A
i
.
12 A. AFTALION, X. BLANC, AND R. L. JERRARD
Next, note that (1.7), (1.8) imply that (T A
j
) = 0. According to (2.6), we
infer that for any x A
i
,
V (x) = (B(x)) (A
i
) =
[T[
n
.
Computing the energy, we thus nd that F() [T[
2
/n, hence
(2.7) inf
Mac(D)
F sup
M

ac
(D)
F
[T[
2
n
.
We point out for future reference that the verication of (2.7) did not use the
assumption n = n.
Step 2: lower bound. Let be a minimizer of F in /(

T), which exists due to
Lemma 2.2. Then, it satises the Euler-Lagrange equations (2.1) and (2.2).
We claim that there exist n points x
1
, . . . , x
n


T such that
(2.8) [x
i
x
j
[ 1 whenever i ,= j, and V (x
i
) =
min
M
F(

T)
[T[
i.
In order to prove it, we dene
T := x

T : V (x) = inf
x
V = [T[
1
inf
M(

D)
F.
According to (2.1), T has a non-empty intersection with any set J such that
(J) > 0. In particular, T is not empty, and we can nd x
1
T. We then
argue by induction: suppose that we have found x
1
, . . . , x
j1
satisfying (2.8),
and assume that j n. Since V (x
i
) min
M
F(

T) inf
Mac(D)
F, we infer
from Step 1 that
(

T) = [T[ > [T[
j 1
n

j1

i=1
V (x
i
)
=
j1

i=1
(B(x
i
))
_

j1
i=1
B(x
i
)
_
.
Thus (

T (
j1
i=1
B(x
i
)) > 0, so we can nd x
j
T (
j1
i=1
B(x
i
)). Together
with the induction hypothesis, this implies that x
1
, . . . , x
j
satisfy (2.8). This
completes the induction proof and hence establishes the claim (2.8).
Next, the denition of n and the fact that n = n implies that if x
1
, . . . , x
n
are any points such that [x
i
x
j
[ 1 for all i ,= j, then B(x
j
)

T. So for
the points x
1
, . . . , x
n
satisfying (2.8), we nd that
[T[ = (

T) = (
n
i=1
B(x
i
))
n

i=1
(B(x
i
)) =
n

i=1
V (x
i
) =
n
[T[
inf
M(

D)
F.
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 13
Since /
ac
(T) /(

T), it follows from this and Step 1 that that
[T[
2
n
inf
M(

D)
F inf
Mac(D)
F
[T[
2
n
and thus that equality holds throughout when n = n. It also follows that F
attains its inmum in /
ac
, and that every measure in /

ac
is a minimizer.
Step 3: any absolutely continuous minimizer belongs to /

ac
(

T). We now as-
sume that /
ac
is a minimizer, and denote by x
1
, . . . , x
n
the points con-
structed in step 2. We then have (B(x
i
)) = [T[/n. We thus dene
A
j
= B(x
j
) supp , j = 1, . . . , n.
Then, according to (2.1) and (2.2), V = [T[/n in A
j
. Hence, (A
j
) = [T[/n:
(1.8) is satised. We next prove that (1.7) holds. In order to do so, rst note
that
(2.9) i ,= j, (B(x
i
) B(x
j
)) = 0.
Indeed, as remarked above, T B(x
i
), so (T)

(B(x
i
)) =

(A
i
) =
[T[ = (T), hence (B(x
i
)) =

(B(x
i
)), which implies (2.9). Since
B(x
i
)B(x
j
) is open and supp is closed, we deduce from (2.9) that supp
B(x
i
) B(x
j
) = whenever i ,= j, hence
(2.10) i ,= j, A
i
B(x
j
) = .
If y
i
A
i
, then V (y
i
) = (B(y
i
)) = [T[/n, according to the denition of
A
i
. Hence, by (2.10), the points (x
1
, . . . , x
n
y
i
) x
i
again satisfy (2.8),
where we know that inf F = [T[
2
/n. Thus, repeating the reasoning that led
to (2.10), we nd that A
j
B(y
i
) = for all j ,= i, ie [y
i
y
j
[ 1 for any
y
j
A
j
, j ,= i. Since y
i
was an arbitrary point in A
i
, this proves (1.7).
2.2.2. The case n ,= n. We write B(a, x) to denote the open ball of radius a
about the point x, with respect to the relevant distance (Euclidean or periodic)
so that B(a, x) := y T : [xy[ < a, with [xy[ understood in the suitable
way.
It is convenient to dene, for a > 0,
F
a
() :=
_

D
V
a
d where V
a
(x) :=
_
1 if [x[ < a
0 if not.
We rst note that the result of the last subsection (ie, Theorem 1.1 in the case
n = n) remains valid if F is replaced by F
a
.
Lemma 2.4. Dene
n(T, a) := maxk : x
1
, . . . , x
k
T such that [x
i
x
j
[ > a i ,= j
14 A. AFTALION, X. BLANC, AND R. L. JERRARD
and assume that a > 0 satises
(2.11) n(T, a) = maxk : x
1
, . . . , x
k


T such that [x
i
x
j
[ a i ,= j.
Then
min
Mac(D)
F
a
= [T[
2
/n(T; a)
and the minimum is attained. In fact, a measure in /
ac
(T) minimizes F
a
if
and only if there exist n(T; a) pairwise disjoint closed sets A
1
, . . . , A
n(D;a)
T,
such that
(2.12) dist(A
i
, A
j
) a if i ,= j,
and
(2.13) (A
i
) =
[T[
n(T; a)
for all i.
We will use the notation
(2.14)
/

ac
(T; a) := /
ac
(T) : A
1
, . . . , A
n
T satisfying (2.12), (2.13).
Note that if /

ac
(T; a) for some a > 0, the denitions easily yield
(2.15) (T A
j
) = 0.
When a = 1, this lemma is exactly the result proved in the last subsection,
and in particular assumption (2.11) is exactly the condition n = n. It is easy
to see that this assumption holds for all but countable many a.
Proof. This follows from the result of the previous subsection by a simple
rescaling.
We now use the above lemma to deduce the
proof of Theorem 1.1. Throughout the proof we will write n and n(a) instead
of n(T) (as dened in the statement of the theorem) and n(T; a) (as dened
in Lemma 2.4) respectively. Note that n = n(a) for a = 1. We will also write
n(a) to denote the right-hand side of (2.11). It is clear that n(a) n(a) for
all a, and also that n, n are nonincreasing functions of a.
We rst remark that there exists > 0 such that
(2.16) n = n(a) = n(a) for all a (1, 1 + ).
To see this, note that by the denition of n, there exist points x
1
, . . . , x
n
T
such that [x
i
x
j
[ > 1 whenever i ,= j, and so it is clear that n n(a)
whenever a < min[x
i
x
j
[ : i ,= j =: 1 + . Then properties of a noted
above imply that n(a) n(a) n(1) for a (1, 1 + ), proving (2.16).
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 15
Note that V
a
V almost everywhere as a 1. For any /
ac
, it
therefore follows from the dominated convergence theorem, Lemma 2.4, and
(2.16) that
F() = lim
a1
F
a
() liminf
a1
[T[
2
n(a)
=
[T[
2
n
.
On the other hand, when considering the case n = n, we proved that F()
|D|
2
n
for all /

ac
(T). (This proof did not use the assumption n = n; see
(2.7)). We deduce that min
Mac(D)
F =
|D|
2
n
, and this minimum is attained by
measures in /

ac
(T).
Finally, suppose that is a minimizer of F in /
ac
, so that F() =
|D|
2
n
.
Since the derivation of the optimality conditions (2.1), (2.2) in Lemma 2.3
holds for minimizers in /
ac
, we can repeat arguments from Step 2 of the
proof of Theorem 1.1 in the case n = n to nd points x
1
, . . . , x
n
T such
that V (x
i
) = F() and [x
i
x
j
[ 1 for i ,= j. We then dene A
j
=
B(x
j
) supp for j = 1, . . . , n, and by repeating the arguments from Step 3
of the n = n proof, we verify that these sets satisfy (1.7), (1.8), which shows
that /

ac
(T).
2.3. A generalization. The above results extend to the functional
(2.17) F
E
() =
_
D
V
E
d, where V
E
(x) =
_
1 if x E
0 if not
for suitable sets E.
Theorem 2.1. Suppose that E is a bounded, open subset of R
d
with 0 E,
and that E is even in the sense that x E x E.
Given a set S R
d
, dene n(T, S) by
n(T, S) := supk : x
1
, . . . , x
k


T such that x
i
x
j
, S i ,= j.
Then min
M
F
E
= [

T[
2
/n(T, E) and the minimum is attained. Moreover, a
measure in / minimizes F
E
if and only if there exist n(T, E) pairwise
disjoint closed sets A
1
, . . . , A
n(D,E)


T, such that if x
i
A
i
, x
j
A
j
, and
i ,= j, then x
i
x
j
, E; and in addition (A
i
) =
|

D|
n(D,E)
for all i.
If E also satises the condition L
d
(E) = 0, then min
Mac
F
E
= [

T[
2
/n(T,

E).
In this case a measure minimizes F
E
in /
ac
if and only if there exist n(T,

E)
pairwise disjoint closed sets A
1
, . . . , A
n


T, such that if x
i
A
i
, x
j
A
j
, and
i ,= j, then x
i
x
j
,

E; and in addition (A
i
) = [

T[/n(

E) for all i.
It is natural to assume that E is open, since then F
E
is lower semicon-
tinuous on the space of nonnegative measures, exactly as the proof of Lemma
2.1. (The same argument shows that if E is closed, then F
E
is upper semicon-
tinuous.)
16 A. AFTALION, X. BLANC, AND R. L. JERRARD
We omit the proof of Theorem 2.1, apart from the following remarks:
The part of the proof that deals with minimizers in / is a straightforward
adaptation of the proof of Lemma 2.4; it is just necessary to replace B(a, x)
by x + E throughout. The part of the proof concerning minimizers in /
ac
is
likewise an adaptation of the proof of Theorem 1.1; the point is that if is
absolutely continuous and L
d
(E) = 0, then F
E
() = F
E
() = lim
r0
F
Er
()
where E
r
:= x R
d
: dist(x, E) < r.
We do not know what are the most general hypotheses on a set E for a
result of the above sort to hold, and have only written the hypotheses that
allow the earlier proof to go through more or less without change.
3. Asymptotic 0: derivation of E
0,
In other sections of this paper, A denotes the innitesimal generator of
rotation about the z-axis. In this section we allow somewhat more general
vector elds A, since it makes absolutely no dierence to the proof. Thus,
throughout this section,
G

(u) :=
_
D
1
2
[u[
2
A Im( uu) dx
and A L

(T; R
d
) is a xed vector eld. In fact it will be apparent from our
proof that the result holds for a much larger class of functionals G.
Proof of Theorem 1.2. For concreteness we consider the Dirichlet problem; the
proofs for the periodic problem are almost exactly the same. Throughout the
proof we will write for example /
0
, /, and so on instead of /
0
(T), /(T)
when there is no possibility of confusion.
We will repeatedly use the fact that
(3.1) G

(u)
1
4
|u|
2
L
2
2
[T[|A|
2
L

(D)
, for u /
0
.
1. We rst show that G

attains its minimum in /

0
. Let u
k
/

0
be
a sequence such that G

(u
k
) inf
A

0
G

.
In view of (3.1), we can extract a subsequence (still labelled u
k
) and a
function u
0
H
1
0
such that such that u
k
u
0
strongly in L
2
and u
k
u
0
weakly in L
2
. In particular, it follows that [u
k
[
2
[u
0
[
2
in L
1
, and hence
[u
0
[
2
/
ac
. Also, standard lowersemicontinuity results (see also Step 4 below)
imply that G

(u
0
) inf
A

0
G

. So we only need to show that u

0
belongs to
/

0
. This however is easy, since the lowersemicontinuity of F implies that
F([u
0
[
2
) liminf
k
F([u
k
[
2
) = min
Mac
F,
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 17
since every [u
k
[
2
is a minimizer for F in /
ac
. Since [u
0
[
2
/
ac
, it follows that
it minimizes F in /
ac
. Hence we conclude from Theorem 1.1 that [u
0
[
2
/

ac
,
or equivalently that u
0
/

0
.
2. Now let u

minimize E
,
in /
0
, and let u
0
minimize G

in /

0
. Then
[u
0
[
2
minimizes F, by Theorem 1.1, so that
G

(u

) = E
,
(u

)
1
4
2
F([u

[
2
)
E
,
(u
0
)
1
4
2
F([u
0
[
2
)
= G

(u
0
). (3.2)
Thus we deduce from (3.1) that u

(0,1]
is uniformly bounded in H
1
. Simi-
larly
F([u

[
2
) = 4
2
[E
,
(u

) G

(u

)]
4
2
[E
,
(u
0
) G

(u

)]
= 4
2
_
G

(u
0
) +
1
4
2
F([u
0
[
2
) G

(u

)
_
min
Mac
F + C
2
. (3.3)
3. In view of (3.2), (3.1), and Rellichs compactness theorem, we may
extract a subsequence (still labelled u

) and a function u

H
1
0
such that
u

strongly in L
2
and u

weakly in L
2
. In particular, it follows
[u

[
2
[u

[
2
in L
1
, and also that [u

[
2
L
1
/
ac
. Since F is continuous
with respect to the L
1
norm, this convergence and (3.3) imply
F([u

[
2
) = limF([u

[
2
) = min
Mac
F.
It then follows from the characterization of minimizers of F in Theorem 1.1
that u

0
. Moreover, it follows from (3.2) and the lower semicontinuity of
G

that
min
A

0
G

(u

) liminf
0
G

(u

) limsup
0
G

(u

) G

(u
0
) = min
A

0
G

.
Therefore u

minimizes G

in /

0
, and
(3.4) lim
0
G

(u

) = G

(u

).
4. It only remains to prove that u

strongly in H
1
. This is easy, however,
because the weak H
1
convergence u

implies that
_
D
A Im( u

) dx
_
D
A Im( u

) dx,
18 A. AFTALION, X. BLANC, AND R. L. JERRARD
so we deduce from (3.4) that |u

|
L
2 |u

|
L
2, and this together with
weak convergence implies that u

strongly in L
2
.
4. One dimensional case
In this section we consider one-dimensional problems. It is easy to check
that E

([u[) E

(u), and similarly E

0
. Since we are interested in minimiz-
ers, we will therefore often implicitly restrict our attention to nonnegative
functions.
The analysis here is much simpler than in higher dimensions, but the
results we establish here presumably give at least some indication of what to
expect when n 2.
4.1. Minimizers of the = 0 problem. In this subsection we consider the
problem (1.14), which we recall here:
nd u
0
/

0
(T) such that E
0
(u
0
) = min
A

0
(D)
E
0,
.
Recall that L| to denote the smallest integer that is greater than or equal to
L. We now give the
Proof of Proposition 1.1. In the Dirichlet case, Theorem 1.1 implies that /
0
consists of functions supported in n(T) sets A
1
, . . . A
n(D)
, each one separated
from all others by distance at least 1. Hence there are numbers a
k
< b
k
,
k = 1, . . . , n(T) such that (after relabelling the A
i
if necessary) A
i
[a
i
, b
i
]
for all i, and b
i
+ 1 a
i+1
for i = 1, . . . , n(T) 1. Moreover, the choice of
n(T) makes it impossible for some A
j
to have nonempty intersection with the
convex hull of some other A
i
.
Then, since (1.8) implies that
_
b
i
a
i
[u[
2
=
L
n(D)
for every i,
_
D
[u

[
2
=
n(D)

i=1
_
b
i
a
i
[u

[
2

n(D)

i=1

2
(b
i
a
i
)
2
_
b
i
a
i
[u[
2
=
L
n(T)
n(D)

i=1

2
(b
i
a
i
)
2
,
with equality if and only if the restriction of u to each interval (a
i
, b
i
) is a
scaled and normalized sine function multiplied by a constant of modulus 1.
Moreover, Jensens inequality implies that

1
(b
i
a
i
)
2

n(T)
(n(T)
1

(b
i
a
i
))
2
n(T)/h
2
0
,
with equality if and only if b
i
a
i
= h
0
for every i. This implies that, if we
assume that u
0
0, it is the function dened in the statement of the theorem.
The proof of the periodic case is essentially the same.
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 19
4.2. Minimizers of E

for 0 < 1. This section is devoted to the proof

of Theorem 1.3. We rst prove three technical lemmas, then give the proof of
the theorem. We use the notation of Proposition 1.1, namely
(4.1) n
p
= L| 1, h
p
= (L n
p
))/ n
p
.
Recall also the notation
y
u(x) := u(x y).
The starting point of the proof of the theorem is the fact that a minimizer
u

of E

is close (after a suitable translation) to the minimizer u

p
of the = 0
problem. The rst lemma records a quantitative consequence of this fact. Thus
u

has n bumps separated by regions in which it is exponentially small. In

the proof of the theorem, we extract 3n parameters, related to a local mass,
kinetic energy, and a length-scale associated with each of these bumps. The
second lemma is used, in the proof of the theorem, to bound E

below by a
function depending only on these parameters, and the third lemma gives a
lower bound for this function of 3n variables.
Lemma 4.1. Given > 0, there exists
0
> 0 such that for 0 < <
0
, if u

is a nonnegative minimizer in H
1
per
(T) of E

and
(4.2) y |u
p

y
u

|
H
1 is minimized for y = 0
then there exists a constant C > 0 depending only on L such that
(4.3) u

(x) 2
_
1 +
L
h
p
_
e
C

5/2

for all x such that dist (x, supp u

p
) > .
Remark 3. Note that because T = (0, L] with periodic boundary conditions
is compact, the function y T |u
p

y
u

|
H
1 attains its minimum, and we
can arrange by a suitable translation that this minimum is attained at y = 0.
Remark 4. Tracking the dependence of the constants in the proof of Lemma 4.1,
it is possible to give an explicit expression for the constant C: C = L/(2

6h
p
n
p
).
Proof. Let u

be a sequence satisfying the assumptions of the theorem. We

rst point out that
1
2
_
L
0
(u

)
2
= E

(u

)
1
4
2
F(u

) E

(u
p
)
1
4
2
F(u
p
) =
1
2
_
L
0
_
u

p
_
2
=

2
L
2h
2
p
. (4.4)
Moreover, since
_
u
2

= 1, one can nd x T such that u

(x) 1. Hence,
(4.5) |u|
L

(D)
1 +

L
__
L
0
(u

)
2
_
1/2
1 +
L
h
p
.
20 A. AFTALION, X. BLANC, AND R. L. JERRARD
Next, we claim that u

u
p
in H
1
. Indeed, it follows from Theorem 1.2
(and also from inequality (4.4)) that u

is precompact in H
1
, and that every
limit of a convergent subsequence is a minimizer for the limiting problem.
However, every such minimizer is a translate of u
p
, and since |
y
u
p
u

|
H
1 =
|u
p

y
u

|
H
1 it follows from (4.2) that the only possible limit of a convergent
subsequence is u
p
.
Moreover, the convergence of u

in H
1
implies that
V u
2

V u
2
p
in H
1
.
It is easy to check that
x T,
_
V u
2
p
_
(x)
L
n
p
+ c [dist (x, supp(u
p
))]
3
,
where c is a constant independent of . Hence, if dist(x, supp u
p
) /2, we
can nd
0
> 0 such that, for any 0 < <
0
,
(4.6)
_
V u
2

_
(x)
L
n
p
+
c
2
[dist (x, supp(u
p
))]
3
Next we record the Euler-Lagrange equation for u

, which is
(4.7) u

=
1

2
(V u
2

)u

=
_
D
(u

)
2
+
1

2
V u
2

u
2

dx.
Note that
E

(u

) E

(u
p
) =
1
4
2
L
2
n
p
+

2
2h
2
p
L.
Clearly

<
4
L
E

(u

), so

<
L

2
np
+
2
2
h
2
p
. So it follows from (4.6), (4.7) that
(4.8)
2
u

+
_
c
2
dist(x, supp(u
p
))
3

2
2
h
2
p

2
_
u

0,
whenever dist (x, supp u
p
) /2. Taking
0
smaller if necessary, we may as-
sume that
2
< (h
2
p

3
)/(64
2
), so that (4.8) becomes
(4.9) u

+
c
4
2
dist(x, supp(u
p
))
3
u

0,
whenever dist (x, supp u
p
) /2. Without loss of generality, we may assume
that a connected component of x, dist(x, supp(u
p
)) /2 is an open interval
(a, b). On this connected component, we use the function
U(x) = |u
p
|
L

_
e

c
3/2
4

2
(xa)
+ e

c
3/2
4

2
(bx)
_
as a supersolution. Hence, on (a, b),
u

(x) U(x) 2|u

p
|
L
e

c
3/2
4

2
(dist(x,supp(up))

2
)
.
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 21
If dist(x, supp u
p
) > , we infer (4.3), with C =

c/8

2.
The next lemma quanties the fact that one cannot conne the entire mass
of a one-dimensional wave function in a small interval, if the kinetic energy is
bounded. As remarked earlier, it will be used to bound E

below by a function
of nitely many variables.
Lemma 4.2. Given an interval J = (a, b) R, let
p(m, k, J) := sup
__
J
u
2
dx :
_
R
u
2
dx = m,
_
R
u
2
dx = k
_
Then there exists a function f : [0, ) [0, ) such that
(4.10) p(m, k, J) = mf
_
[J[
_
k
m
_
,
and in addition f satises
f(s) 1 for all s, and f(s) = 1 if s , (4.11)
f(s) = 1
1
2
( s)
3
+ O
_
( s)
4
_
for s . (4.12)
As a result, there exists a constant c > 0 such that f(s) 1 c( s)
+ 3
for
every s 0.
Proof. 1. We will eventually prove the lemma when J = (
1
2
,
1
2
). We rst
show that the general case follows by a change of variables. Indeed, suppose
J = (a, b), and for m, k given, consider any u H
1
(R) such that |u|
2
L
2
=
m, |u

|
2
L
2
= k. Let x
0
=
1
2
(a + b) and L = b a, and dene a new variable
y =
xx
0
L
. Next dene U(y) = u(x), and let m

= |U|
2
L
2
= m/L, k

:=
|U

|
2
L
2
= Lk and J

= (
1
2
,
1
2
). Note that
_
J
u
2
dx = L
_
J

U
2
dy
Hence if the conclusion holds for J

= (
1
2
,
1
2
), then
p(m, k, J) = Lp(m

, k

, J

) = Lm

f([J

[
_
k

) = mf([J[
_
k
m
).
2. We henceforth x J = (
1
2
,
1
2
). By a scaling argument similar to that
given above, we can also assume that m = 1, so that in fact
(4.13) f(s) = sup
__
J
u
2
dx : |u|
L
2 = 1, |u

|
L
2 = s
_
.
22 A. AFTALION, X. BLANC, AND R. L. JERRARD
It is clear that f(s) 1 for all s. For s , dene
u(x) =
__
2s

cos(sx) if [x[ /2s

0 if not,
and note that |u|
L
2 = 1, |u

|
L
2 = s, and
_
J
u
2
= 1. Hence f(s) = 1 for s .
3. It remains to prove (4.12), which is the real content of the lemma. To
do this, x s < . We rst claim that the supremum in (4.13) is attained. To
see this, let u
k
be a sequence such that |u
k
|
L
2 = 1 and |u

k
|
L
2 = s, and such
that
_
J
u
2
k
f(s). Let u H
1
(R) be a function such that u
k
u locally in
L
2
, and u

k
u weakly in L
2
. Then
_
J
u
2
dx = f(s), |u|
L
2 1, |u

|
L
2 s.
Let us write a = |u|
L
2 and b =
1
s
|u

|
L
2, so that 0 < a, b 1. We consider
two cases:
case 1: suppose b a 1. Then dene U(x) = (ab)
1/2
u(ax/b). Then
one can check that |U|
L
2 = 1, |U

|
L
2 = s, and
_
J
U
2
dx =
1
a
_ a
2b

a
2b
u
2
dx
_
J
u
2
dx = f(s).
Then U satises the constraints and attains the supremum on the right-hand
side of (4.13). Note also that we get strict inequality, and hence a contradic-
tion, if a < 1.
case 2: suppose a < b 1. First, dene U
0
to be the symmetric re-
arrangement of u (see [22] for example), so that U
0
is even and nonincreasing
in (0, ). Then
_
J
U
2
0
dx
_
J
u
2
dx = f(s), |U
0
|
L
2 = a, |U

0
|
L
2 = b

s bs = |u

|
L
2.
If b

a then we are back to case 1, and this leads to a contradiction since

a < 1. If not, we dene U
1
(x) by
U
1
(x) :=
_
U
0
(0) if [x[
U
0
([x[ ) if [x[ ,
where is chosen so that |U
1
|
L
2 = b

. Note also that |U

1
|
L
2 = |U

0
|
L
2 = b

,
regardless of the choice of . Also, since U
0
attains its maximum at x = 0,
_
J
U
2
1

_
J
U
2
0
f(s). Finally, let U
2
=
1
b

U
1
. Then U
2
satises the constraints
|U
2
|
L
2 = 1, |U

2
|
L
2 = s, and
_
J
U
2
2
f(s), so the supremum on the right-hand
side of (4.13) is attained.
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 23
4. Now let u
s
denote a function such that
_
J
u
2
s
dx = f(s), |u
s
|
L
2 = 1, |u

s
|
L
2 = s.
Standard arguments show that u
s
> 0 everywhere. The Euler-Lagrange equa-
tion satised by u
s
is

J
u
s
= u
s
+ u

s
where , are Lagrange multipliers. Here
J
(x) = 1 if x J and 0 otherwise.
This equation cannot be satised if = 0, so it can be rewritten
(4.14) u

s
= (
J
+
J
c)u
s
for certain constants , , where J
c
denotes the complement of J. Note that
this is essentially the equation for a one-dimensional quantum particle in a
square well, so that if we like, we can nd a complete description of solutions
by consulting basic quantum mechanics textbooks. From this, or by solving
directly, we nd that (4.14) has square-integrable solutions only if > 0 and
< 0, so we can rewrite the equation as
(4.15) u

s
= (
2

J

2

J
c)u
s
.
We further see that L
2
solutions exist only if = tan(/2), and when this
holds, all square-integrable solutions are multiples of

(x) :=
_
cos x if [x[
1
2
cos(/2) exp
_

_
1
2
[x[
_
if [x[
1
2
.
It follows from this that for s < , u
s
=

/|

|
L
2 for some = (s) such
that

> 0 and |

|
L
2/|

|
L
2 = s. Note that

> 0 i < . We now

dene
g() := |

|
L
2/|

|
L
2, h() = |

|
2
L
2
_
J

dx.
It follows from what we have said that f(s) = h() for some (0, ) such
that g() = s. We will show that in fact g is one-to-one in (0, ], so that
f(s) = h(g
1
(s)), where g
1
denotes the inverse of the restriction of g to
(0, ].
5. We next explicitly compute g() and h(). We integrate to nd that
|

|
2
L
2 =
1
2
_
1 +
1

sin
_
+
1

cos
2

2
We substitute = tan

2
and write
1
2
sin = sin

2
cos

2
, then simplify to nd
|

|
2
L
2 =
1
2
+
1

cot

2
.
24 A. AFTALION, X. BLANC, AND R. L. JERRARD
Similarly, noting that
2

=
2

in J and
2

=
2

elsewhere, and rewriting

as above, we see that
|

|
2
L
2 =

2
2
(1 +
1

sin ) + cos
2

2
=

2
2
+ sin .
Thus
g() =
_
+ 2 sin
+ 2 cot

2
_
1/2
Similarly,
h() =
+ sin
+ 2 cot

2
.
It is easy to check that g
2
, and hence g, is an increasing function on (0, ), as
is h for that matter. Thus f(s) = h(g
1
(s)). Now (4.12) follows by calculus.
Indeed, we compute that g() = , g

() =
1
2
which implies that g
1
(s) =
+ 2(s ) + O((s )
2
) near s = . We also check that h() = 1, h

() =
h

() = 0 and h

() =
3
2
. Substituting the expansion for g
1
(s) into the
Taylor series for h at = yields (4.12).
The last technical lemma is
Lemma 4.3. Let L > 2 be as in Theorem 1.3, let c > 0 be given, and dene
n
p
, h
p
by (4.1). Then, for any positive numbers (m
i
)
1i np
, (
i
)
1i np
, (
i
)
1i np
such that
np

i=1
m
i
= L,
np

i=1

i
= n
p
h
p
,
we have
(4.16)
np

j=1
m
2
j
+ c(
j

j
)
+ 6
+ 2
2
m
j

2
j

L
2
n
p
+ 2
2

2
h
2
p
L C
2/5
.
Proof. Let us write q
0
((m
j
), (
j
), (
j
)) to denote the left-hand side of (4.16).
Fix (m
j
), (
j
), (
j
) that minimize q
0
(, , ) subject to the constraints on

m
i
and

i
. If we let m
j
= L/ n
p
,

j
= h
p
,
j
= /h
p
for all j, then
(4.17) q
0
((m
j
), (
j
), (
j
)) q
0
(( m
j
), (
j
), (

j
)) =
L
2
n
p
+ 2
2

2
h
2
p
L.
The constraint

m
i
= L implies that
np

i=1
m
2
i
=
L
2
n
p
+
np

i=1
(m
i

L
n
p
)
2
.
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 25
Thus
sup
j
(m
j

L
n
p
)
2

np

i=1
(m
j

L
n
p
)
2
2
2

2
h
2
p
L.
From (4.17) we also see that
sup
j
m
j

2
j

np

j=1
m
j

2
j


2
h
2
p
L.
Combining these, we deduce that m
j

2
j

L
np

2
j
O() for every j. Hence
q
0
((m
j
), (
j
), (
j
))
L
2
n
p
+
np

j=1
c(
j

j
)
+ 6
+ 2
2
L
n
p

2
j
C
3
.
It now suces to show
(4.18) q
1
((
j
), (
j
)) :=
np

j=1
c(
j

j
)
+ 6
+ 2
2
L
n
p

2
j
2
2

2
h
2
p
L C
2/5
.
To do this, let (

j
), (

j
) minimize q
1
subject to the constraint

j
= n
p
h
p
.
Clearly q
1
((

j
), (

j
)) 2
2
2
h
2
p
L, which implies that
j

j
C
1/3
. Since

j
= C, it follows that
j
C > 0 for all j.
The rst-order optimality conditions for (

j
), (

j
) are
6c
_

j
_
+5

j
= 4
2
L
n
p

j
6c
_

j
_
+5

j
=
for all j, where is a Lagrange multiplier. These imply that 4
2 L
np

2
j
=

j
,
and hence that
(4.19) 6c
_

4
2

L
n
p
_

j
_
3
_
+5

j
= .
We also deduce that
n
p
h
p
=

j
= 4
2
L
n
p

j
a
2
j
= O(
2
).
which implies that = O(
2
). For > 0, the equation 6c(
4
2

L
np
x)
+5
x =
has at most two roots, say x

< x
+
. Since = O(
2
), one of these roots must
x

= O(
2
). However, we have already shown that
j
C for all j, and so it
follows that

j
= x
+
for every j. In particular, all

j
are equal, which implies
that all

j
are equal. Thus

j
= h
p
for all j. It follows that
q
1
((
j
), (
j
)) n
p
inf

_
c( h
p
)
+6
+ 2
2
L
n
p

2
_
26 A. AFTALION, X. BLANC, AND R. L. JERRARD
Now it is easy to deduce (4.18).
Proof of Theorem 1.3. 1. upper bound: We dene a one-parameter family
of test functions v
h
, and we then optimize over h. Let
v
h
(x) =
_

_
_
2L
h np
sin(

h
(x x
i
)) if x (x
i
, x
i
+ h)
for some i 0, . . . n
p
1
0 if not
where n
p
and x
i
are dened by
n
p
= L| 1, x
i
= i(1 + h
p
), h
p
=
1
n
p
(L n
p
).
The assumption that L > 2 implies that
1
2
<
L
np
1. We have normalized v
h
so that
_
D
v
2
h
dx = 1. We easily see that
(4.20) |v

h
|
2
L
2 =
_

h
_
2
|v
h
|
2
L
2 =
_

h
_
2
L.
Let h
p
=
1
np
(L n
p
). If h h
p
then components of the support of v
h
are
separated by distance at least 1, and so
(4.21) F([v
h
[
2
) = min
Mac
F = F(u
p
) = L
2
/ n
p
On the other hand, if h > h
p
, then for x (0, h),
V v
2
h
(x) =
_
h
0
v
2
h
dx = L/ n
p
if h
p
h x < h
p
,
since then y supp u : x 1 < y < x + 1 = (0, h). If h
p
< x < h say, then
x + 1 1 + h, so
V v
2
h
(x)
_
h
0
v
2
h
dx +
_
1+h
1+hp
v
2
h
dx
L
n
p
+ C
(h h
p
)
3
h
3
.
The same bound holds for 0 < x < h h
p
. On these intervals, v
2
h
is bounded
by C(h h
p
)
2
/h
3
, so
_
h
0
V v
2
h
v
2
h
dx
L
n
p
_
h
0
v
2
h
+ C
(h h
p
)
3
h
3
_
(0,hhp)(h,hp)
v
2
h
dx

_
L
n
p
_
2
+ C
(h h
p
)
6
h
6
.
Summing the contributions from dierent components of the support of v
h
,
we nd that
(4.22) E

(v
h
)
1
4
2
_
L
2
n
p
+ Cn
p
[(h h
p
)
+
]
6
h
6
_
+

2
2h
2
L.
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 27
where (h h
p
)
+
= max0, h h
p
. Let h

minimize the right-hand side

of (4.22). One can then check by calculus that h

h
p

2/5
and that
E

(v
h
) E

(u
p
) c
2/5
, proving the upper bound.
2. lower bound. Let u

denote a minimizer of E

in H
1
per
(T) which
satises the hypotheses of Lemma 4.1. For i = 1, . . . , n
p
, let I
j
= (y
j
, y
j+1
),
where y
j
=
1
2
(x
j
+ x
j+1
) + h/2. Thus each y
i
is the midpoint between two
adjacent maxima of u
p
, and each I
i
is centered at a max of u
p
. We will use
the notation
m
j
=
_
I
j
u
2

dx, F
ij
:=
__
I
i
I
j
V (x y)u
2

(x)u
2

(y) dx dy
so that
_
D
V u
2
u
2
dx =

i,j
F
ij
. Note that for every j,
F
jj
= m
2
j

__
{x,yI
j
I
j
: |xy|1}
u
2

(x)u
2

(y) dx dy m
2
j
Ce
C/
by Lemma 4.1. Also, clearly F
ij
= 0 if [i j[ 2. For each j, we dene an
interval K
j
I
j
I
j+1
such that [K
j
[ = 1 and
_
K

j
u
2

dx =
_
K
+
j
u
2

dx =
1
2
_
K
j
u
2

dx =:
1
2

j
,
where K

j
:= K
j
I
j
and K
+
j
:= K
j
I
j+1
. Since V (xy) = 1 if x K

j
I
j
and y K
+
j
I
j+1
, we see that
F
j+1,j
= F
j,j+1

_
_
K

j
u
2

(x)dx
__
_
K
+
j
u
2

(y)dy
_
=
1
4

2
j
.
Thus
_
D
V u
2

u
2

dx

j
(m
2
j
+
1
2

2
j
) Ce
C/

j
[m
2
j
+
1
2
(

j1
+
j
2
)
2
] Ce
C/
. (4.23)
Note also that
(4.24)
1
2
(
j1
+
j
) =
_
K
+
j1
u
2

dx +
_
K

j
u
2

dx = m
j

_
J
j
u
2

dx.
for J
j
:= I
j
(K
+
j1
K

j
). We now change a little bit the function u

in
order to have u

H
1
0
(I
j
), so that, extending it by 0 outside I
j
, we may apply
Lemma 4.2. Indeed, recall that, according to Lemma 4.1, u

(y
j
) Ce
C/
.
Hence, one can nd an ane function v such that
|v|
W
1,
(I
j
)
Ce

, u

+ v H
1
0
(I
j
).
28 A. AFTALION, X. BLANC, AND R. L. JERRARD
Hence, denoting by w

the function equal to u

+ v in I
j
and 0 elsewhere, we
apply Lemma 4.2 to w

, nding that
_
J
j
w
2

m
_
_
1 c
_
[J
j
[
_
k
m
_
+ 3
_
_
,
where
k =
_
R
(w

)
2
, m =
_
R
w
2

.
Hence, using the fact that |w

|
W
1,
(I
j
)
Ce
C/
and that the function
s ( s)
+ 3
is uniformly continuous on R
+
, we have
(4.25)
_
J
j
u
2

dx m
j
_
1 c( [J
j
[
j
)
+3

+Ce

for
j
=
_
1
m
j
_
I
j
(u

)
2
dx
_
1/2
.
We combine (4.24) and (4.25) and substitute into (4.23) to obtain
_
D
V u
2

u
2

dx

j
_
m
2
j
+ c( [J
j
[
j
)
+6

Ce
C/
.
Since
_
D
(u

)
2
dx =

j
m
j

2
j
, we nally conclude that
(4.26) E

(u

j
_
1
4
2
_
m
2
j
+ c( [J
j
[
j
)
+6
_
+
1
2
m
j

2
j
_
Ce
C/
.
Note that

m
j
= [T[ = L, and

[J
j
[ = L n
p
= n
p
h
p
. This last fact follows
from the fact that T =
np
j=1
(J
j
K
j
), since [K
j
[ = 1 for every j and the union
is disjoint. Thus the lower bound follows from Lemma 4.3.
5. Minimizers of E

for large
We give in this section the
Proof of Proposition 1.2. Let us rst point out that

d
= |V |
L
1 = (V 1)(x), x T.
This will be useful in the sequel.
Fix u H
1
per
(T) such that
_
[u[
2
= [T[, and write u = 1 + h. We may
assume that u 0, since E

([u[) E

(u). Hence, h 1 and

_
D
(2h + h
2
) dx =
_
D
(1 + h)
2
dx
_
D
1 dx = 0.
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 29
Using this fact, one easily checks that
E

(u) E

(1) =
_
D
1
2
[h[
2
dx +
1
4
2
_
V (1 + h)
2
(1 + h)
2
V 1 1

dx
=
_
D
1
2
[h[
2
dx +
1
4
2
_
D
V (2h + h
2
) (2h + h
2
) dx.
Note also that
1
4
_
D
V (2h + h
2
)(2h + h
2
) dx
=
_
D
V h h + V h h
2
+
1
4
V h
2
h
2
dx
=
_
D
V h h + V (1 + h/2)
2
h
2
V 1 h
2
dx

_
D
V h hdx
d
|h|
2
L
2,
hence
(5.1) E

(u) E

(1)
_
D
1
2
[h[
2
dx +
1

2
_
D
V h hdx
1

d
|h|
2
L
2.
We are now going to prove that if (1.15) holds, then the right-hand side of
(5.1) is non-negative, and vanishes only when h 0. We will do this using
Fourier series, and we rst x notation: for f L
2
(T) we write

f(n) :=
1
L
d
_
D
f(x)e
inx/L
dx
so that
f(x) =

Z
d

f(n)e
inx/L
.
We have the usual identities, which scale in L as follows:
_
D
f(x)g(x) dx = L
d

Z
d

f(n) g(n),

(f g) (n) = L
d

f(n) g(n),
Using these we check that
(5.2)
1
2
_
D
[h[
2
dx+
1

2
_
D
V hhdx = L
d

Z
d
_

2
2
[n[
2
L
2
+
L
d

V (n)
_
[

h(n)[
2
.
30 A. AFTALION, X. BLANC, AND R. L. JERRARD
Next, we write TV () =
_
R
d
V (x)e
ix
dx for the Fourier transform on R
d
, so
that TV (n/L) = L
d

V (n). Since V is even and nonnegative,

TV () =
_
R
d
1
2
(V (x) + V (x))e
ix
dx =
_
R
d
V (x)
1
2
_
e
ix
+ e
ix

dx
=
_
R
d
V (x) cos(x ) dx =
_
{|x|<1}
cos(x ) dx.
In particular,
L
d

V (n)
d
inf
|s||n|/L
cos s
d
_
1

2
2
[n[
2
L
2
_
.
It follows from this, (5.2) and (5.1) that
E

(u) E

(1) L
d

Z
d
_

2
[n[
2
2L
2
+

d

2
_
1

2
[n[
2
2L
2
_

2
_
[

h(n)[
2
= L
d

Z
d
_

2
[n[
2
2L
2
_
1

d

2
_
_
[

h(n)[
2
0,
with equality i h 0. This completes the proof of the proposition.
Remark 5. In the preceding proof, we have used an exact decomposition of
the energy of 1 +h. However, the crucial point is that u 1 is a critical point
of E

under the constraint

_
u
2
= [T[, and that the second derivative of E

is
positive denite for suciently large. Under this condition, using the fact
that the minimizer of u

converges to 1 as tends to innity, it is possible to

argue by contradiction and prove that there exists
1
> 0 such that for any
>
1
, the unique minimizer is u 1. However, this kind of proof does not
give any information on
1
. In particular,
1
may depend on L.
Acknowledgement The third authors contributions to this paper were largely
carried out during a visit, funded by the CNRS, to the Laboratoire Jacques-
Louis Lions, and the support and hospitality of these institutions are gratefully
acknowledged. This work was supported by the French ministry grant ACI
Nouvelles interfaces des mathematiques.
References
[1] A. Aftalion, X. Blanc, R. L. Jerrard, Nonclassical rotational inertia of a supersolid,
Phys. Rev. Lett. 99, 135301, 2007.
[2] A. F. Andreev, I. M. Lifschitz, Sov. Phys. JETP., 29, 1107 (1969).
[3] S. Balibar, Supersolidity and Superuidity, Contemporary Physics 48, 31 (2007).
[4] P. Billingsley, Convergence of Probability Measures, Wiley, New York, 1968.
MATHEMATICAL ISSUES IN THE MODELING OF SUPERSOLIDS 31
[5] L. A. Caarelli, F. H. Lin, An optimal partition problem for eigenvalues, J. Sci. Comput.
31, 518, 2007.
[6] G. V. Chester, Speculations on Bose-Einstein condensation and quantum crystals, Phys.
Rev. A 2, 256258, 1970.
[7] M. Conti, S. Terracini, G. Verzini, An optimal partition problem related to nonlinear
eigenvalues, J. Funct. Anal. 198, 160196, 2003.
[8] J. H. Conway, N. J. Sloane, Sphere Packings, Lattices, and Groups, 2nd ed. New York:
Springer-Verlag, 1993.
[9] L. Erdos, Rayleigh-type isoperimetric inequality with a homogeneous magnetic eld,
Calc. Var. Partial Dier. Equ. 4, 283292, 1996.
[10] S. Fournais, B. Heler, Accurate eigenvalue asymptotics for the magnetic Neumann
Laplacian, Ann. Inst. Fourier, Genoble, 56, 167, 2006.
[11] T. C. Hales, The Sphere Packing Problem, J. Comput. Appl. Math 44, 41-76, 1992.
[12] T. C. Hales, Sphere packings, I, Discrete Comput. Geom., 17, 151, 1997.
[13] A. Henrot, Extremum problems for eigenvalues of elliptic operators, Frontiers in Math-
ematics, Birkhauser Verlog, Basel, 2006.
[14] C. Josserand, Y. Pomeau and S. Rica, Coexistence of ordinary elasticity and superu-
idity in a model of a defect-free supersolid, Phys. Rev. Lett. 98, 195301 (2007).
[15] C. Josserand, Y. Pomeau and S. Rica, Patterns ans supersolids, Euro. Phys. J. S. T.
146, pp 4761 (2007)
[16] P. Kapitza, Viscosity of liquid helium below the -point, Nature 141, 74, 1938.
[17] E. Kim and M.H.W. Chan, Probable observation of a supersolid helium phase, Nature
(London) 427, 225227 (2004);
[18] E. Kim, M.H.W. Chan, Observation of Superow in Solid Helium, Science 305, 1941
(2004);
[19] E. Kim, M.H.W. Chan, Supersolid Helium at high pressure, Phys. Rev. Lett. 97, 115302
(2006).
[20] M. Kondo, S. Takada, Y. Shibayama, K. Shirahama, Observation of Non-Classical
Rotational Inertia in Bulk Solid 4He, J. Low Temp. Phys. 148,, 695699, 2007.
[21] A. J. Leggett, Can a solid be superuid? Phys. Rev. Letters, 25, 15431546 (1970).
[22] E. H. Lieb, M. Loss, Analysis, Graduate Studies in Mathematics, 14. American Math-
ematical Society, Providence, RI, 1997.
[23] F. London The -phenomenon of liquid helium and the Bose-Einstein degeneracy, Na-
ture 141, 643, 1938.
[24] O. Penrose, L. Onsager, Bose-Einstein condensation and liquid helium, Phys. Rev. 104,
576584, 1956.
[25] A. Penzev, Y. Yasuta, M. Kubota, Annealing Eect for Supersolid Fraction in 4He,
ArXiv cond-mat 0702632.
[26] N. Prokofev, What makes a crystal supersolid? Advances in Physics 56, 381402, 2007.
[27] L. Reatto, Bose-Einstein Condensation for a Class of Wave Functions , Phys. Rev.,
183, 334-338, (1969).
[28] A. S. C. Rittner, J. D. Reppy, Observation of Classical Rotational Inertia and Non-
classical Supersolid Signals in Solid 4He below 250 mK, Phys. Rev. Lett. 97, 165301,
2006.
[29] S. Sasaki, R. Ishiguro, F. Caupin, H. J. Maris, S. Balibar, Superuidity of grain bound-
aries and supersolid behavior, Science 313, 10981100, 2006.
32 A. AFTALION, X. BLANC, AND R. L. JERRARD
Amandine Aftalion, CNRS, UMR 7598, Laboratoire J.-L. Lions, F-75005,
Paris, France
E-mail address: aftalion@ann.jussieu.fr
Xavier Blanc, UPMC Univ Paris 06, UMR 7598, Laboratoire J.-L. Lions,
F-75005, Paris, France
E-mail address: blanc@ann.jussieu.fr
Robert L. Jerrard, Department of mathematics, University of Toronto,
Toronto, Canada M5S 2E4
E-mail address: rjerrard@math.toronto.edu