Nonrelativistic limit of the Dirac-Fock equations.
Maria J. Esteban and Eric Sere
CEREMADE (UMR C.N.R.S. 7534)
Universite Paris IX-Dauphine
Place du Marechal de Lattre de Tassigny
75775 Paris Cedex 16 - France
esteban or sere@ceremade.dauphine.fr
Abstract.
In this paper, the Hartree-Fock equations are proved to be the non relativistic limit of the Dirac-Fock equations as far as convergence of \stationary states" is concerned. This property is used to derive a meaningful denition of \ground state" energy and \ground state" solutions for the Dirac-Fock model.AMS Subject Classication:
49 S 05, 35 J 60, 35 P 30, 35 Q 75, 81 Q 05, 81 V 70, 81 V 45, 81 V 55.Key Words :
Relativistic quantum mechanics, nonrelativistic limit, quan-tum chemistry, ground state, nonlinear eigenvalue problems, Dirac-Fock equations, Hartree-Fock equations, variational methods, critical points, strongly indenite functionals, bifurcations from the essential spectrum.1. Introduction
In this paper we prove that solutions of Dirac-Fock equations converge, in a certain sense, towards solutions of the Hartree-Fock equations when the speed of light tends to innity.
This limiting process allows us to dene a notion of ground state for the Dirac-Fock equations, valid when the speed of light is large enough.
First of all, we choose units for which
m
=h
= 1, wherem
is the mass of the electron, andh
is Planck's constant. We also impose e24"0 = 1, with ?
e
the charge of an electron,"
0 the permittivity of the vacuum.The Dirac Hamiltonian can be written as
H
c = ?ic
r+c
2
;
(1)
where
c >
0 is the speed of light in the above units, = 1I 00?1I !
,
k =0k 0k
!
(
k
= 1;
2;
3) and the k are the well known Paulima-trices. The operator
H
cacts on 4-spinors, i.e. functions from IR3to CI4, andit is self-adjoint in
L
2(IR3;
CI4), with domainH
1(IR3;
CI4) and form-domainH
1=2(IR3;
CI4). Its spectrum is (?1
;
?c
2][[c
2;
+1).Let us consider a system of
N
electrons coupled to a xed nuclear charge densityeZ
, wheree
is the charge of the proton,Z >
0 the total number of protons and is a probability measure dened on IR3. Note that in theparticular case of
m
point-like nuclei, each one having atomic numberZ
i ata xed location
x
i;eZ
=Xmi=1
eZ
ixi andZ
=m
X
i=1
Z
i.In our system of units, the Dirac-Fock equations for such a molecule are given by 8 > > > > > > > < > > > > > > > :
H
c; k:=H
c k?Z
( 1 jx
j ) k+ ( 1 jx
j ) k ? Z IR 3R
(x;y
) k(y
) jx
?y
jdy
="
ck k (k
= 1;:::N
);
GramL 2 = 1IN (i:
e R IR 3 k l=kl;
1k;
lN):
(DFc)Here, = ( 1
;
;
N);
each k is a 4-spinor inH
1=2(IR3;
CI4) (bybootstrap, k is also in any
W
1;p(IR3) space, 1p <
3=
2), and (x
) := N X k=1 k(x
) k(x
); R
(x;y
) := N X k=1 k(x
) k(y
):
(2)We have denoted the complex line vector whose components are the
conjugates of those of a complex (column) vector , and
1 2 is the inner
product of two complex (column) vectors 1, 2. The
n
n
matrix Gram L2
is dened by the usual formulas (Gram ) :=Z
k(
x
) l(x
)dx :
Finally,
"
c1:::
"
cN are eigenvalues ofH
c;:
Each one represents theenergy of one of the electrons, in the mean eld created by the molecule. For physical reasons, we impose 0
< "
ck< c
2:
Note that the scalars"
ck canalso be seen as Lagrange multipliers. Indeed, the Dirac-Fock equations are the Euler-Lagrange equations of the Dirac-Fock energy functional
Ec( ) = N X k=1 Z IR 3 k
H
c k?Z
1 jx
j k k +12ZZ IR3 IR 3 (x
) (y
) ?trR
(x;y
)R
(y;x
) jx
?y
jdxdy
under the constraintsR IR
3
k l=
kl:
In [6] we proved that under some assumptions on
N
andZ
, there exists an innite sequence of solutions of (DFc). More precisely:Theorem 1 [6]
. LetN < Z
+ 1. For anyc >
=2+2=2 max(Z;
3N
?1),there exists a sequence of solutions of (DFc),
n c;jo j1
H
1=2(IR3)N , such that(i)
0<
Ec( c;j)< Nc
2 ,(ii)
jlim !+1 Ec( c;j) =Nc
2,(iii)
0< c
2?j< "
c;j 1:::
"
c;j N< c
2?m
j , withj> m
j>
0 independent ofc
.The constant =2+2=2 is related to a Hardy-type inequality obtained independently by Tix and Burenkov-Evans (see [15, 3, 16]), and which plays an important role in the proof of Theorem 1. With the physical value
c
= 137:
037:::
andZ
an integer (the total number of protons in the molecule), our conditions becomeN
Z; N
41; Z
124:
The constraintN
41is technical, and has no physical meaning.
Our result was recently improved by Paturel [13], who relaxed the condi-tion on
N
. Paturel obtains the same multiplicity result, assuming only thatN < Z
+ 1 and =2+2=2 max(Z;N
)< c
. Takingc
= 137:
037:::
, Paturel's conditions areN
Z
124 : they cover all existing neutral atoms. This isIn [6], the critical points c;j are obtained by a complicated min-max
argument (in [13], the argument is not simpler). Here, we do not give this min-max in its full detail. We just state the minimal information needed in the present paper. Let us denote
E
:=H
1=2(IR3;
CI4). Since (H
c) = (?1;
?c
2]
[[
c
2
;
+1)
;
the Hilbert space
E
can be split asE
=E
+cE
? c;
whereE
c := cE
, and c :=IR (H
c). The projectors c have a simpleexpression in the Fourier domain : d
c (
) = ^ c() ^(), with b c() := 12 1ICI 4 c+ c2 p c4+ c2jj2 !:
(4)Proposition 2 [6, 13]
. For everyj
1, letV
be any (N
+j
?1)dimen-sional complex subspace of
E
+c. Then, taking the notation of Theorem 1, we haveE
cj;DF :=Ec( c;j) sup 2(E ? c V ) N Gram L 2 1I N Ec( ):
(5)In the present paper, we prove three main theorems. We rst consider a sequence
c
n !+1 and a sequencef ngn of solutions of (DFcn). For all
n
, n = ( n1;:::;
nN), each nk is inH
1=2(IR3;
CI4), with ZIR3 k l
dx
= kl andH
c n ;n nk =
"
nk nk. Using the standard Hardy inequality, one can provethat the functions nk are in
H
1(IR3;
CI4) forc
n large enough. We assumethat ?1
<
lim n!+1 ("
n1?c
2n) lim n!+1 ("
nN ?c
2n)<
0:
(6)A (column) vector 2CI4can be written in block form = ?'
where
'
2CI2(respectively
2 CI2) consists of the two upper (resp. lower) components
of . This gives the splitting nk = ?'n k
n k
with
'
nk and nk inH
1(IR3;
CI2).Finally, n splits as? n
n
, where n := (
'
n1;:::;'
nN) andn:= (
n1
;:::;
nN).Our rst result is that n=? n
n
towards =? 0 , where = (
'
1;
;
'
N) 2H
1(IR3;
CI2)N is a solution of the Hartree-Fock equations:8 > > > > > > > > < > > > > > > > > : H
'
k = ?'
k 2 ?Z
1 jx
j'
k+ 1 jx
j'
k ? Z IR3R
(x;y
)'
k(y
) jx
?y
jdy
= k'
k; k
= 1;:::N;
Z IR3'
k'
ldx
=kl;
k= limn !+1 ("
nk?c
2n):
(HF)Here (as in the Dirac-Fock equations),
(x
) =XN l=1'
l(x
)'
l(x
); R
(x;y
) =XN l=1'
l(x
)'
l(y
):
Note that the Hartree-Fock equations are the Euler-Lagrange equations corresponding to critical points in
H
1(IR3;
CI2)Nof the Hartree-Fock en-ergy: EHF() := N X k=1 1 2jjr
'
kjj 2 L 2 ?Z
Z IR3 1 jx
j j'
kj 2dx
(7) +12ZZ IR3 IR 3 (x
)(y
)?tr(R
(x;y
)R
(y;x
)) jx
?y
jdxdy ;
under the constraintZ
IR3
'
k
'
l=kl ,i;j
= 1;:::N
.Theorem 3
. LetN < Z
+ 1. Consider a sequencec
n ! +1 and asequence f ngn of solutions of (DFc
n), i.e.
n = ( n1
;
;
nN), each nkbeing in
H
1=2(IR3;
CI4) , with ZIR3
k l
dx
= kl andH
cn;n nk =
"
nk nk .Assume that the multipliers
"
nk,k
= 1;:::;N;
satisfy (6). Then forn
large enough, nk is in
H
1(IR3;
CI4) , and there exists a solution of (HF),= (
'
1;
;
'
N), with negative multipliers, 1;:::;
N, such that, afterex-traction of a subsequence,
nk:="
nk?(c
n) 2 ?! n!+1 k; k
= 1;:::;N ;
(8)nk=
'
nknk ! ?! n!+1
'
k 0 ! inH
1(IR3;
CI2)H
1(IR3;
CI2);
(9) nk+i
2c
n(r)'
nk L 2 (IR 3 ;CI 2 ) =O
(1=
(c
n)3);
(10) and Ec n( n)?Nc
2n ?! n!+1 EHF():
(11)As a particular case, we have
Corollary 4
. Ifc
n! +1 andN; Z;
are xed, then for anyj
1 thesequencef c n;j
g
n of Theorem 1 satises the assumptions of Theorem 3 (see
(iii) in Theorem 1). So it is precompact in
H
1(IR3;
CI4)N . Up to extraction of subsequences, c n ;j k :="
c n ;j k ?c
2n?! j k<
0; k
= 1;:::;N
(12) cn;j ?! j 0 ! inH
1(IR3;
CI2)NH
1(IR3;
CI2)N (13) and j ='
j1;
;
'
jnis a solution of the Hartree-Fock equations with multipliers
j1;
;
j N. Moreover, Ec n( cn;j) ?Nc
2n ?! n!+1 EHF(j):
(14)Particular solutions of the Hartree-Fock equations are the minimizers of EHF() under the constraints Gram
L
2 = 1IN
:
They are called groundstates. Their existence was proved by Lieb and Simon [10] under the as-sumption
N < Z
+ 1 (see also [11] for the existence of excited states).It is dicult to dene the notion of ground state for the Dirac-Fock model, sinceEc has no minimum under the constraints
R
IR3
k l=
kl:
Oursecond main result asserts that the rst solution c;1 of (DFc) found in
[6], whose energy level will be denoted
E
c1;DF, can be considered, in some (weak) sense, as a ground state for (DFc). Indeed,E
c1;DF ?Nc
2 convergesto the minimum of EHF as
c
goes to innity. Moreover, forc
large themultipliers
"
c;1k associated to c;1 are theN
smallest positive eigenvalues ofTheorem 5
. LetN < Z
+ 1 andc
suciently large. With the above notations,E
c1;DF := Ec( c;1) = min Gram L 2=1IN EHF() +Nc
2+o
(1)c !+1 (15)Moreover, for any subsequence f c n;1 g n converging in
H
1(IR3;
CI4)N to some ? 1 0, 1 is a ground state of the Hartree-Fock model, i.e.
EHF( 1) = min GramL 2=1IN EHF()
:
(16)Furthermore, for
c
large, the eigenvalues corresponding to c;1 in (DFc),"
c;11;:::;"
c;1N are the smallest positive eigenvalues of the linear operatorH
c; c;1and the (
N
+1)-th positive eigenvalue of this operator is strictly larger than"
c;1N .Finally, we are able to show that, for
c
large enough, the function c;1can be viewed as an electronic ground state for the Dirac-Fock equations in the following sense: it minimizes the Dirac-Fock energy among all electronic congurations which are orthogonal to the \Dirac sea".
Theorem 6
. FixN;Z
withN < Z
+1 and takec
suciently large. Then c;1 is a solution of the following minimization problem:inffEc( ); Gram L 2 = 1IN
;
? = 0g (17) where ?=
(?1;0)(H
c; ) is the negative spectral projector of the operatorH
c; , and ? := (? 1;
;
? N):
The constraint ?= 0 has a physical meaning. Indeed, according
to Dirac's original ideas, the vacuum consists of innitely many electrons which completely ll up the negative space of
H
c; : these electrons formthe \Dirac sea". So, by the Pauli exclusion principle, additional electronic states should be in the positive space of the mean-eld Hamiltonian
H
c;:
The proof of Theorem 6 will be given in Section 4. This proof uses some other interesting min-max characterizations of c;1 (see Lemma 9).
2. The nonrelativistic limit
This section is devoted to the proof of Theorem 3. We rst notice that when
N < Z
+1,N;Z
xed, andc
is suciently large, any solution of (DFc)is actually in (
H
1(IR3))N. This follows from the fact that for
small, the operatorH
1?j
x
jis essentially self-adjoint with domain
H
1(IR3) (see [14]).We can also obtaina prioriestimates on
H
1 norms:Lemma 7
. FixN;Z
2ZZ
+, takec
large enough, and let c be a solutionof (DFc) . If the multipliers
"
ck associated to c satisfy 0"
ckc
2(
k
= 1;:::;N
), then c2(H
1(IR3
;
CI4))N , and the following estimate holds
jj c jj 2 2 + jjr c jj 2 2
K :
The constant
K
is independent ofc
(forc
large).Proof.
The normalization constraint GramL 2 c= 1I N implies jj c jj 2 2 =N :
(18)Using the (DFc) equation and the standard Hardy inequality
Z IR3
u
2 jx
j2 4 Z IR3 jru
j 2;
(19)one easily proves that cis in
H
1 , and satises:(
H
c c;H
c c) =c
4jj c jj 2 2+c
2 jjr c jj 2 2 (20)c
4 jj cjj 2 2 +`
(Z
2+N
2) jjr cjj 2 2+`c
2max(N;Z
) jjr cjj 2;
for some
` >
0 independent ofN;Z
andc
. The estimates (18) and (20)prove the lemma. tu
Proof of Theorem 3.
Let us split the spinors nk : IR3 !CI4 in blocks ofupper and lower components : nk=
'
nknk
!
denote
L
:=?i
(r):
Then we can rewrite (DFcn) in the following way: 8 > > > > > > > > > > > > > > > > > > > > > > > > > > > > < > > > > > > > > > > > > > > > > > > > > > > > > > > > > :
c
nL
nk?Z
1 jx
j'
nk+N X l=1( j'
nlj 2 +jnlj 2 ) 1 jx
j'
nk+ (c
n)2'
nk ? N X l=1'
nl(x
) Z IR3 ('
nl)(y
)'
nk(y
) + (nl)(y
)nk(y
) jx
?y
jdy
="
nk'
nkc
nL'
nk?Z
1 jx
j nk+N X l=1( j'
nlj 2 +jnlj 2 ) 1 jx
j nk?(c
n) 2nk ? N X l=1nl(
x
) Z IR3 ('
nl)(y
)'
ck(y
) + (cl)(y
)ck(y
) jx
?y
jdy
="
nknk Z IR3('
nk)'
nl+ (nl)nldx
= kl (21) Note thatkL
k L 2 = krk L 2 for all 2H
1(IR 3;
CI2)
:
So, dividing byc
n therst equation of (21), we get kr
nkk L 2 (IR 3 ;CI 2 ) =O
(1=c
n):
(22)Dividing by 2(
c
n)2 the second equation of (21), and using the fact that"
nk?(c
n)2 is a bounded sequence, we get nk ? 1 2c
nL'
nk L 2 (IR 3 ;CI 2 ) = 1(c
n)2O
XN l=1 knlk H 1 (IR 3 ;CI 2 ) !:
(23)The estimate (23) together with Lemma 7 implies
k
nkk L 2 (IR 3 ;CI 2 ) =O
(1=c
n):
(24)Combining this with (22), we obtain
k
nkk H 1 (IR 3 ;CI 2 ) =O
(1=c
n):
(25) So XN l=1 knlk H 1 (IR 3 ;CI 2) =
O
(1=c
n), and (23) gives the estimate nk ? 1 2c
nL'
nk L 2 (IR 3 ;CI 2 ) =O
(1=
(c
n)3):
(26)Now, the rst equation of (21), combined with (26), implies 8 > > > > > > > > > > > > > < > > > > > > > > > > > > > : ?
'
nk 2 ?Z
( 1 jx
j )'
nk+N X l=1 j'
nlj 2 1 jx
j nk ? N X l=1'
nl(x
) Z IR3 ('
nl)(y
)'
nk(y
) jx
?y
jdy
=nk'
nk+h
nk;
Z IR3('
nk)'
nl = kl+r
nkl;
withnk :="
nk?(c
n)2, and lim n!+1 jjh
nkjj H ?1 (IR 3 ) = 0;
nlim !+1 jr
nklj= 0 for allk;l
2f1;::: ;N
g:
Therefore n:= (
'
n1;:::;'
nN) is a Palais-Smale sequence for the Hartree-Fock problem, and the multipliers nk satisfy limn!+1nk<
0:
At thispoint, we just invoke an argument used in [11] to obtain the convergence in
H
1 norm of some subsequencef n0
g towards = (
'
1;
;
'
N), a solutionof the Hartree-Fock equations
8 > > < > > : H
'
k= k'
k; k
= 1;:::N
Z IR3'
k'
l = kl;
where k= limn0 !+1 n 0 k .Finally, let us prove thatEc n
0(
n0
)?
N
(c
n0)2 converges toEHF():
FromLemma 7 and the estimate (26), one easily gets
Ec n(
n)?
Nc
2n=EHF(n) +O
(1=
(c
n)2):
(27)
Since n0
converges in
H
1 norm to , the energy level EHF(n 0) con-verges to EHF()
:
So (27) implies the desired convergence. This ends the3. Ground state for Dirac-Fock equations in the nonrelativistic
limit.
The aim of this section is to prove Theorem 5. The estimate given in Proposition 2 on the energyEc( c;i) and the expression of ^
c given in (4),
will be crucial.
Proof of Theorem 5
. By Corollary 4, for any sequencec
ngoing to innity,cn;1is precompact in
H
1 norm. If it converges, its limit is of the form ( 1 0), and Ec n( cn;1) ?N
(c
n)2 converges toEHF(1):
As a consequence, lim c!+1 Ec( c;1)?Nc
2 inf GramL 2=1IN EHF():
(28)In order to prove (15) and (16) of Theorem 5, we just have to show that lim c!+1 Ec( c;1)?
Nc
2 inf Gram L 2=1I N EHF():
(29) Take = ('
1;
;'
N) 2H
1(IR3;
CI2)N , with GramL 2 = 1IN:
LetV
c be the complex subspace ofE
+c dened byV
c:= Span f+c( ' 1 0);:::;
+c( ' N 0 ) g:
(30)From formula (4) and Lebesgue's convergence theorem, one easily gets, for
k
= 1;:::;N ;
lim c!+1 k ? c(' k 0) k H 1 = 0:
(31)So, for
c
suciently large, we havedim
V
c =N :
(32) Hence, by (5),E
c1;DF :=Ec( c;1) sup 2(E ? V c) N Gram L 2 1I N Ec( ):
(33) Let +2(E
+c) N;
? 2(E
? c)Nsuch that GramL 2(
++ ?) 1I
N. By
the concavity property of Ec in the
E
?is large enough, we have Ec( ++ ?) Ec( +) + E 0 c( +) ? ? 1 4 N X k=1( ? k
;
p ?c
2 +c
4 ? k) Ec( +) +M
jj ? jj L 2 ?c
2 4jj ? jj2 L 2;
(34)for some constant
M >
0 independent ofc
. Hence, forc
large,E
c1;DF sup + 2D(V c) Ec( +) + (1)c !+1;
(35) whereD
(V
c) := n +2(V
c)N ; Gram L 2 +1I N o .If
c
is large enough, it follows from Hardy's inequality (19) that the map + !Ec( +) is strictly convex on the convex setA+:=f +2(
E
+c)N; Gram L 2 +1I N g:
Indeed, its second derivative at any point + of A+ is of the form E 00 c( +)[
d
+]2= 2 N X k=1(d
k;
pc
4?c
2d
k) L 2 +Q
(d
+)with
Q
a quadratic form on (H
1=2(IR3;
CI4))N bounded independently ofc
and +2A+
:
As a consequence, sup
+ 2D(V
c)
Ec( +) is achieved by an extremal point +max
of the convex set
D
(V
c) =A+\(V
c)N. Being extremal inD
(V
c);
the point+max satises
GramL
2 +max= 1I N
:
(36)
Since +k;max 2
V
c;
there is a matrixA
= (a
kl)1k ;lN such that, for all
l
,+l;max = X 1kN
a
kl +c(' k 0):
ThenA
Gram L 2 +c( 0)A
= GramL 2 +max = 1I N:
(37)Using the
U
(N
) invariance ofD
(V
c) and Ec;
and the polar decompositionof square matrices, one can assume, without restricting the generality, that
from (31), that GramL 2(+c ? 0 ) = 1IN+o(1)
:
So (37) impliesA
2= 1I N+o(1);
hence
A
= 1IN+ o(1):
Combining this with (31), we get k +k;max?( ' k 0) k H 1 =o
(1)c !+1:
Now, since +k;max 2
E
+c; H
c +k;max= pc
4?c
2 +k;max:
But pc
4?c
2c
2? 2:
This inequality is easily obtained in the Fourier domain: it follows from
p 1 +
x
1 + x2 (8x
0):
So we get N X k=1(H
c +k;max;
+k;max)L 2Nc
2+ 12 XN k=1 kr +k;maxk 2 L 2:
Combining this with (31), we nd
Ec( +max)
Nc
2+
EHF() +(1)c !+1
:
(38)
Finally, (35) and (38) imply
E
c1;DF Ec( +max) +(1)c !+1Nc
2+ EHF() +(1)c !+1:
(39)Since is arbitrary, (39) implies (29). The formulas (15), (16) of Theo-rem 5 are thus proved.
We now check the last assertion about the
"
c;1k;k
= 1;:::;N;
being the smallest eigenvalues of the operatorH
c; c;1 forc
large. By Corollary4, we can translate this statement in the language of sequences. We take a sequence
c
n ! +1 such that f cn;1 g n converges in
H
1(IR3;
CI4)N to some ? 1 0, for
n
large enough. LetH
n :=H
cn;cn;1 and H 1 := H 1
:
We haveH
n cn;1 k ="
nk cn;1 k and H 1'
1k= k'
1k;
with 0< "
n1:::
"
nN<
(c
n) 2;
1
:::
N<
0;
k= lim n!+1 ("
nk?(c
n) 2):
Let us denote
e
n1:::
e
ni the sequence of eigenvalues ofH
n;
in the interval (0
;c
2n);
counted with multiplicity. Similarly, we shall denote 1:::
i the sequence of eigenvalues of H(?1
;
0);
counted with multiplicity. Letz
2 CIn(H1)
:
Then forn
largeenough,
z
+ (c
n)2 2CIn(H
n);
and the resolventR
n(z
+ (c
n)2) :=
(
z
+ (c
n)2)I
?H
n?1
converges in norm towards the operator
L
(z
) : ?'! ?R(z)' 0
;
whereR
(z
) :=z I
?H 1 ?1 is the resolvent of H1
:
So, by the standard spectraltheory, limn
!+1
(
e
ni?(c
n)2) =
i for all
i
1
:
We know that 1 is a ground state of the Hartree-Fock model. So a
result proved in [1] tells us that
k= k for all 1k
N
, and N+1>
N:
But (
"
nN ?(c
n)2) converges to N;
and (e
nN+1?(c
n)2) converges to N+1;
as
n
goes to innity. So, forn
large enough,e
nN+1> "
nN;
hence"
nk=e
nk for all 1k
N :
This ends the proof of Theorem 5. tu4. Proof of Theorem 6
In this section, both and will denote
N
-uples of 4-spinors (functions from IR3 into CI4). As already noted above, in [6] the solution c;1 wasobtained by a complicated min-max argument. For
c
large, we will show that this argument can be simplied.First of all, we introduce the notion of projector \
"
-close to +c", where +c= 12jH
cj?1
(
H
c+jH
cjis the positive free-energy projector.
Denition 8
: LetP
+ be an orthogonal projector inL
2(IR3;
CI4), whoserestriction to
H
12(IR3
;
CI4) is a bounded operator onH
12(IR3
;
CI4) .Given
" >
0,P
+ is"
-close to +c if and only if, for all 2H
1 2(IR3;
CI4), ?c
2 +c
4 1 4P
+ ?+c L2(IR 3;CI4)"
?c
2 +c
4 1 4 L2(IR 3;CI4):
An obvious example of projector
"
-close to +c is +c itself. More inter-esting examples will be given below. Let us now give a min-max principle associated toP
+ :Lemma 9
. FixN;Z
withN < Z
+ 1. Takec >
0 large enough, andP
+a projector
"
-close to+c, for" >
0 small enough. LetP
?= 1I L 2 ?P+, and deneE
(P
+) := inf + 2(P +H 1 2) N Gram =1I sup 2(P ?H 1 2 Span( +))N Gram =1I Ec( ):
Then
E
(P
+) does not depend onP
+ and Ec( c;1)E
(P
+).Remark.
In the caseN
= 1;
Ecis the quadratic form (;H
) L2 associated
to the operator
H
=H
c?Z
1jxj
:
ThenE
(+c) coincides with the min-maxlevel
1(V
) dened in [4], forV
= ?Z
1
jxj
:
By Theorem 3.1 of [4], ifc > =
2 + 22=
, then1(V
) is the rst positive eigenvalue ofH :
Proof of Lemma 9
.The idea behind this lemma is inspired by [2]. Note that, under our assumptions,
E
(P
+)< Nc
2(1 +K"
) for someK >
0 independent ofc
and
"
. This follows from arguments similar to those used in the proof of Lemma 5.3 of [6]. In [6] the free energy projectorsc were used. With
these projectors, it was seen that
E
(+c)< Nc
2 (thanks to a careful choiceof +). When
P
+ is"
-close to +c, we then getE
(P
+)< Nc
2(1 +K"
).To continue the proof of the lemma we perform a change of physical units. In mathematical language, this change corresponds to a dilation in space by the factor
c
, and to dividing the energies byc
2. Let (d
c'
)(x
) =c
3=2'
(cx
)and e Ec() : = 1 c2 Ec
d
c = XN k=1 Z IR3'
k;
(?i
r+)'
k ?Z
c
e 1 jx
j j'
kj 2 + 12c
ZZ IR3 IR 3 (x
)(y
)?kR
(x;y
)k2 jx
?y
jd
3xd
3y
(40)where ~
(E
) =(c
?1E
) for any Borel subsetE
of IR3.The interest of this rescaled energy e
Ecis that for
c
large and Gram L 2 1I N, we have e Ec( ) = N X k=1 Z IR 3 k;
(?i
r+) k +O
1c
jj jj 2 (H 1=2 ) N:
(41) Let us denoteP
e :=d
c?1P
d
c, e :=d
c?1 cd
c= IR ?i:
r+:
Note that edoes not depend on
c
. Now,P
+ is"
-close to +c if and onlyif 8 > > < > > : ? + 1 1 4 e
P
+? e + L2(IR 3;CI4)"
? + 1 1 4 L2(IR3;CI4);
8 2H
1 2(IR3;
CI4):
(42)We denote
A
the right action of anN
N
matrixA
= (a
kl) 1k ;lNon an
N
-uple = ('
1;:::;'
N)2(L
2(IR3;
CI4))N:
More precisely,(
A
) := ( N X k=1a
k1'
k;:::;
N X k=1a
kN'
k):
(43) Given + = ('
+1;:::;'
+N) 2 eP
+H
1=2Nsuch that GramL 2 + = 1I N
;
and ? 2 eP
?H
1=2 N;
we deneg
+( ?) := (++ ?) h GramL 2 ( ++ ?) i ? 1 2 = (++ ?) h 1IN + Gram L 2 ? i ? 1 2:
(44)We obtain a smooth map
g
+
;
from eP
?H
1 2 N to + := n 2 eP
?H
1 2 Span('
+1;:::;'
+N) N=
GramL 2 = 1IN o:
In fact, the values of
g
+ lie in the following subset of + : 0 + := n 2 +
=
Gram L 2 eP
+>
0o:
Now, take an arbitrary 2 0
+
:
Then there is an invertibleN
N
matrix
B
such thatP
e+ = +B
. So we may writeB
?1= ++ eP
?B
?1:
As a consequence,g
+( eP
?B
?1) = (B
?1) h GramL 2 (B
?1) i ? 1 2:
One easily computes GramL 2 (
B
?1) = (B
)?1 GramL 2B
?1 = (B B
)?1:
Henceg
+( eP
?B
?1 ) = (B
?1 )(B B
)1=2 = (B
?1 (B B
)1=2);
and nally =g
(P
B
1)U ;
where
U
:= (B B
)?1=2B
2 U(
N
) is the unitary matrix appearing in thepolar decomposition of
B :
So we have proved that 0 + = [ ? 2( e P?H 1 2) N U2U(N)g
+( ?)U :
Now, Ec is invariant under the U(
N
) action \" , and 0+ is dense in
+ for the norm of (
H
1=2(IR3
;
CI4))N . Hence sup 2( e P?H 1 2 Span( +))N GramL 2 = 1IN e Ec( ) = sup ? 2( e P?H 1 2) N e Ecg
+( ? ):
(45)We now prove Lemma 9 in three steps.
Step 1
. Let + 2 ( eP
+H
1=2)N be such that GramL 2
+ = 1I
N and such
that e
Ec(+)
N
+, for some>
0 small. For"
small andc
large, thereis a unique ? 2 e
P
?H
1=2 N maximizing e Ecg
+ and lying in a small
neighborhood of 0. If we denote
k
(+) this maximizer, the mapk
is smoothfrom S+ = n + 2 e
P
+H
1=2N /GramL 2 + = 1I N;
e Ec( +) N + o to eP
?H
1=2 N, and equivariant for the U(
N
) action.Proof of Step 1
. Taker >
0. For";
small andc
large, if + 2 S+;
? 2( e
P
?H
1=2)N;
and k ? k H1=2 is not smaller than
r
, then e Ecg
+( ?)< N
?;
by (41). On the other hand, for
c
large enough, using (41) once again, one has e Ecg
+(0) = e Ec(+)N
? 2:
So, if we deneVr := n ? 2 eP
?H
1=2 N /k ? k H 1=2r
o , no maximizer of e Ecg
+ can be outsideVr. Moreover, choosing
r
small, and then takingc
large and"
small, the map ? 2Vr7?! e Ecg
+( ?)is strictly concave. Indeed, its second derivative at ?
2Vr is very close in
norm to the negative form ? 2( e
P
?H
1=2 )N 7?!?2 N X i=1 k ? i k2 H 1=2 ? 2 X 1i;jN ('
+j;'
+i)H 1=2( ? i;
? j )L 2:
Step 1 immediately follows from these facts. tu
Step 2
. The min-max levelE
(P
+) does not depend onP
+.Proof of Step 2
. Take two projectorsP
1+,P
2+, both"
-close to +c. Fori
= 1;
2, and +i2 eP
i+H
1=2N , with GramL 2+i= 1IN and e Ec(+i)N
+;
letJ
i(+i) := max ? 2( ~P ? i H 1=2)N Gram L 2 ?=1I N e Ecg
i + i (?) = e Ecg
i + ik
i(+i):
(46) Here,g
i + andk
i are the maps associated to
P
i+ in Step 1.By Ekeland's variational principle [5], there is a minimizingsequence
+1;n
n0
for
J
1, such that (J
1)0 +1;n ?! n?!+1 0 inH
?1=2 N . Let n:=g
1+ 1;nk
1(+1;n) . Then n is a Palais-Smale sequence for eEc in the manifold :=n 2
H
1=2N /GramL 2 = 1I N o;
with e Ec nN
?2
;
where>
0 is the constant of the rst step. So GramL 2 eP
2+ n>
0:
We denote 8 > < > : +2;n:=P
e+ 2 n h GramL 2 eP
2+ ni ? 1 2;
? 2;n :=P
e ? 2 n h GramL 2 eP
2+ ni ? 1 2:
(47)One easily checks that n =
g
2+ 2;n (? 2;n)
:
Since e Ec nN
? 2;
we have k ? 2;nk H 1=2r
, wherer >
0 is the same as in the proof of step 1. Sincen is a Palais-Smale sequence for e
Ec, the derivative of e
?
2;n converges to 0 as
n
goes to innity. So, by the concavity properties ofe Ec
g
2 + 2;n in the domain V2;r := n ? 2 eP
? 2H
1=2 N /k ? k H 1=2r
o(see the proof of step 1), we get
k ? 2;n?
k
2(+2;n) k H 1=2 ?! n!+1 0 and e Ec n ?J
2 +2;n ?! n!+1 0:
As a consequence,E
(P
1+) = inf + 1 2( ~P + 1 H 1=2)N Gram L 2 + 1=1INJ
1 +1 inf + 2 2( ~P + 2 H 1=2)N Gram L 2 + 2=1INJ
2(+2) =E
(P
2+):
Since 1
;
2 play symmetric roles in the above arguments, we conclude thatE
(P
+) does not depend onP
+, forc
large enough and"
small enough. tuStep 3
. Ec c;1E
+c, where c;1 is the rst solution of (D-F) found
in [E-S].
Proof of Step 3
.For
c
large enough, if ? 2 ? cH
1=2satises GramL 2 ? 1I N, it followsfrom Hardy's inequality that the map +!Ec( ++
?) is strictly convex on
W
( ?) := f + 2(+cH
1=2)N ; Gram L 2( ++ ?) 1I N g:
As a consequence, for an arbitrary
N
-dimensional subspaceV
of +cH
1=2,S
V( ?) := sup + 2W( ?) \V N Ec(+ + ?) is achieved by an extremal point
+max of the convex set
W
( ?)\
V
N. Being extremal, +max must satisfythe contraints GramL
2( +max+ ?) = 1I N
:
So we have sup 2( ? cH 1=2 V ) N Gram L 2 1I N Ec( ) = sup ? 2( ? cH 1=2)N Gram L 2 ? 1I NS
V( ?) = sup 2( ? cH 1=2 V ) N Gram L 2 =1IN Ec( ):
By proposition 2, Ec c;1 sup 2( ? cH 1=2 V ) N GramL 2 1I N Ec
:
Finally we get, for
c
large,Ec( c;1) inf + 2( + cH 1=2)N Gram L 2 +=1I N sup 2( ? cH 1=2 Span( +))N Gram L 2 =1IN Ec( ) =
E
(+c):
(The correspondence between + and
V
isV
= Span(+)):
This ends theproof of Step 3 and of Lemma 9. tu
Thanks to Lemma 9, we are able to write the following inequalities for
c
large, andP
+"
-close to +c,"
small :E
(P
+) =E
(+c) Ec( c;1) inf solution of (DFc) ? = 0 Ec( ) inf 8 > < > : 2(H 1=2)N Gram L 2 =1I N ? =0 Ec( ) (48)As announced before, we now give some important examples of projectors
"
-close to +c :Lemma 10
. FixN;Z
, and takec
large enough. Then, for any 2H
1=2N , with GramL 2 1I N, the projector + = (0;+1)H
c; is"
-close to +c.Proof of Lemma 10
. We adapt a method of Griesemer, Lewis, Siedentop [7] to the HamiltonianH
c;. Once again, it is more convenient to work in asystem of units such that
H
c; becomese
H
c;~ : 7!d
c?1H
c;d
c( ) = ?i
r+ ?Z
c
e 1 jx
j +1c
~x
1 1c
Z IRR
~(x;y
)x
(y
)y
dy
with
e(E
) =(c
?1E
), e (x
) =c
?3=2(c
?1x
). DenotingH
1 := ?i
r+, e +~ := (0;1) eH
c;~ , e+ := (0;1)(H
1),K
~:=c
eH
c;~?H
1, we nd, as in the proof of Lemma 1 of [7],
e +~? e + = 1
c
Z + 1 0dz
hH
21+z
2i ?1H
1K
~H
e c;~?z
2K
~ h eH
c;~2 +z
2i ?1;
and for any
2L
2(IR3;
CI4), following [7] (proof of Lemma 3), we get;
(? + 1)1=4( e +~? e +) L2M
c
kk L 2 k(? + 1)1=4 k L 2for
c
large enough (M
is a constant independent ofc
). As a consequence, ifc
is large enough and bigger thanM
"
, then + is"
-close to +c. This endsthe proof of Lemma 10. tu
Now, to end the proof of Theorem 6, we just need the following result :
Lemma 11
. FixN; Z
and takec >
0 large enough. If 2H
1=2N , GramL 2 = 1IN, ? = 0 and Ec()Nc
2, then Ec() = max n Ec( ); 2 h ?H
1=2Span() iN;
GramL 2 = 1I N o:
Proof of Lemma 11
. If ? = 0 and GramL 2 = 1IN, then 0 is a criticalpoint of the map ? 2 ?
H
1=2 N 7?!Ecg
( ?);
withg
( ?) = + ? h 1IN + Gram L 2 ? i ?1=2:
Take" >
0 small. By Lemma 10, + is"
-close to +cforc
large enough. From the proof of Lemma9 (Step 1), there is a unique critical point of Ec
g
in a small neighborhood Vrof 0 in?
(
H
1=2) and this critical point is the unique maximizer of Ecg
in ?
(
H
1=2). So, 0 is this maximizer. This proves Lemma 11. tuLet us explain why Theorem 6 is now proved. We know that, for
c
large enough,Nc
2> E
+c Ec( c;1) inf 8 > < > : 2(H 1=2)N Gram L 2 =1IN ? =0 Ec( );
hence inf 8 > < > : 2(H 1=2)N Gram L 2 =1I N ? =0 Ec( ) = inf 8 > > > < > > > : 2(H 1=2)N GramL 2 =1IN ? =0 Ec( )Nc 2 Ec( )
:
Take
" >
0. By Lemma 10, for any 2(H
1=2)N with Gram L2 = 1IN,
the projector + is
"
-close to +c, ifc
is large. HenceE
(+ ) =E
(+c) (by Lemma 9), if we have chosen"
small enough. But if also satises ?= 0
and Ec( )
Nc
2, then, from Lemma 11 and from the denition ofE
(+ ),we have
E
(+c) =E
(+ ) Ec( ). SoE
+c inf 8 > < > : 2(H 1=2)N Gram L 2 =1IN ? =0 Ec( );
and therefore,E
+c = Ec( c;1) = inf 8 > < > : 2(H 1=2)N GramL 2 =1IN ? =0 Ec( )and Theorem 6 is proved. tu
Aknowledgements
. The authors are grateful to Boris Buoni for explain-ing them the work [2], and suggestexplain-ing that it might be useful in the study of the Dirac-Fock functional. The proof of Lemma 9 is inspired by this paper.References
[1] V. Bach, E.H. Lieb, M. Loss, J.P. Solovej. There are no unlled shells in unrestricted Hartree-Fock theory. Phys. Rev. Lett.
72
(19) (1994), p. 2981-2983.[2] B. Buoni, L. Jeanjean. Minimax characterization of solutions for a semi-linear elliptic equation with lack of compactness. Ann. Inst. H. Poincare
10(4)
(1993), p. 377-404.[3] V.I. Burenkov, W.D. Evans. On the evaluation of the norm of an integral operator associated with the stability of one-electron atoms. Proc. Roy. Soc. Edinburgh, sect. A
128 (5)
(1998), p. 993-1005.[4] J. Dolbeault, M.J. Esteban, E. Sere. Variational characterization for eigenvalues of Dirac operators. Cal. Var. and PDE
10 (4)
(2000), p. 321-347.[5] I. Ekeland. On the variational principle. J. Math. Anal.
47
(1974), p. 324-353.[6] M.J. Esteban, E. Sere. Solutions for the Dirac-Fock equations for atoms and molecules. Comm. Math. Phys.
203
(1999), p. 499-530. [7] M. Griesemer, R.T. Lewis, H. Siedentop. A minimax principle for eigenvalues in spectral gaps : Dirac operators with Coulomb potentials. Doc. Math.4
(1999), p. 275-283.[8] I.W. Herbst. Spectral theory of the operator (
p
2+m
2)1=2?ze
2=r
.Comm. Math. Phys.
53
(1977), p. 285-294.[9] Y.-K. Kim. Relativistic self-consistent eld theory for closed-shell atoms. Phys. Rev.
154
(1967), p. 17-39.[10] E. H. Lieb, B. Simon. The Hartree-Fock theory for Coulomb sys-tems. Comm. Math. Phys.,
53
(1977), p. 185-194.[11] P.-L. Lions. Solutions of Hartree-Fock equations for Coulomb systems. Comm. Math. Phys.
109
(1987), p. 33-97.[12] A. Messiah. Mecanique quantique. Dunod, 1965.
[13] E. Paturel. Solutions of the Dirac equations without projector. To appear in A.H.P.
[14] B. Thaller. The Dirac equation. Springer-Verlag, 1992.
[15] C. Tix. Strict positivity of a relativistic Hamiltonian due to Brown and Ravenhall. Bull. London Math. Soc.