This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
,r .
(1.5)
It turns out, as we explain in Sect. III that t(A) is, for A e W(.9), nothing other than the usual normalized trace on .!F applied to A. The basic ingredients of a non-commutative integration theory are an algebra of operators such as `B(2), and a linear functional on it such as T. Norms analogous
to the L° norms of commutative integration theory can now be introduced by Il A ll, = (r(A'A)p12)uo for
I
II A II , denotes the operator norm of A. Let 4S"(2) denote '6(i) equipped with the corresponding norm. Having expanded our horizons beyond Hilbert space, we can ask for hounds between II P,A Iig and II A II,, for different q, p and t. Our main result, Theorem 4, is
the optimal fermion hypercontractivity inequality; i.e. for all I < p 5 q < cz, and all A in 16P(9), IIP'A!Iq!5 11Ali
when
e-z'
(1.6)
and the t saturating the inequality on the right is the smallest for which the inequality on the left always holds. We prove this for it degrees of freedom with n an arbitrary finite integer. Since the estimate is independent of n. a theorem of Gross [Gr72] implies that it holds as well with infinitely many degrees of freedom. The result (1.6) was conjectured by Gross [Gr75] who proved it [Gr72] in the
special case p = 2, q = 4. The cases p = 2. q = 2m, where in is an integer were proved by Lindsay and Meyer [LiMe] following earlier work by Lindsay [Lin] on the case p = 2, q = 2". Since P, is self adjoint, duality yields a corresponding family of results in which q = 2. Until now all other cases (except when it = I or 2) had remained open. The optimal relation between t, p and q found here for fermion hypercontractivity is the same as that found by Nelson [Ne73] for boson hypercontractivity. By now there are many proofs of Nelson's inequality. Neveu's elegant proof [Nev], like Nelson's original proof, is based on probabilistic methods. Proofs based on geometric methods have been given by Carlen and Loss [CL90] and by Lieb [Li90] who considers generalizations in which the Mehler kernel (the kernel for the
153
With E. Carlen in Commun. Math. Phys. 155, 27-46 (1993) E.A. Carlcn and E.H. Lieb
30
boson oscillator semigroup) is replaced by an arbitrary Gaussian kernel. More references can be found in the bibliography to Gross's article [Gr89, Gr92]. However, none of the existing approaches to the boson problem has been found to solve the fermion problem. See also [Far]. As a corollary to the optimal fermion hypercontractivity inequality, we obtain the optimal fermion logarithmic Sobolev inequality:
t(IAI2InIAl2)-(IIAIIiInlIAll2 :52r(A*HOA),
(1.7)
where I AI = (A*A)1/2.
This inequality was conjectured by Gross, and proved by him in a weaker form with the constant on the right increased by a factor of In 3. In studying perturbations of Ho by multiplication operators V, this inequality plays the same role as does the usual Sobolev inequality in studying perturbations of - JA by multiplication operators V on L2(R", d"x) [Fe69]. (The multiplication operator associated to a self-adjoint element V of X8(2) is defined to be the average of left and right multiplication by V, and is denoted here by the same symbol V.) Again, in the standard Fock space setting, there is no natural way to formulate such an a-priori regularity inequality, or even to introduce the notion of a multiplication operator. This paper is organized as follows: In Sect. 11 we study the structure of 'e(2) for finite n. It is actually simpler, as well as technically advantageous, to consider it as a subalgebra of the algebra W(it) generated by the identity, the configuration observables Q1, ... , Q. and their conjugate momenta P1, ... , P. Since they both turn out to be certain Clifford algebras, their structure has been worked out long ago with the representation theory of the orthogonal groups. Thus, this section contains no new result but simply introduces notation and prepares the way for what follows. Of particular use are an explicit spin-chain representation of the and the Jordan-Wigner transform identifying it with the operator algebra algebra generated by n "hard core bosons." Section III concerns properties of the spaces and their norms. The main result here is an optimal uniform convexity inequality for'8'(. P ), I < p < 2, which is joint work with Keith Ball. We need only a special case of this inequality here, and a proof is provided in an appendix for the reader's convenience. Section IV introduces a convenient expression for P, in terms of the conditional expectation its of with respect to W(2). The main result in this section is an inequality for conditional expectations which enables us to prove that
sup{IIP,AII,:IIAII,,= 1) =sup{IIP,AII4:A _ 0 and IIA'Ii= I' Thus, to establish (1.6) in general, we need only consider positive A. In Sect. V we establish the optimal fermion hypercontractivity bounds and the corresponding optimal fermion logarithmic Sobolev inequality. This is done in several steps. First, using results collected in Sects. 11 and III we establish that (1.7) holds for I < p < 2 and q = 2. At t = 0 this is an equality; differentiating it there yields the logarithmic Sobolev inequality (1.7). Gross showed that (1.6) would follow from (1.7), if it were true (as we show here), for all self adjoins A. His result rests on a deep inequality he established for positive operators. Since it is not in general true that II P,AII9 5 IIP:IAIIIQ, as is trivially true in the commutative case, Gross's result does not allow us to conclude (1.6) for general (i.e. non-self adjoint)
A from (1.7). The results of the Sect. IV do, however, allow us to draw this conclusion.
154
Optimal Hypercontractivity for Fermi Fields and Related Non-Commutative Integration Inequalities Optimal Hypercontractivity for Fermi Fields
31
Finally, in Sect. VI we show that the same hypercontractivity relation (1.6) holds for a mixed system of bosons and fermions. Non-commutative probability theory has grown into a substantial branch of
analysis with a number of physical applications. The mathematical theory is reviewed and developed in [Me85] and [Me86], while other sorts of physical applications, besides those discussed here, are treated in [Da76] and [HuPa] for example. It
is a pleasure to thank Leonard Gross for discussing his results and
conjectures with us, and for encouraging us to take up the latter. We are indebted to Keith Ball for his collaboration on the subject of convexity inequalities that led to Theorem I [BCL], which is one of the key ingredients in the present work. Thanks are also due to G.-F. Dell'Antonio, A. Jaffe and A. Wightman for useful discussions. H. Fermions and the Clifford Algebra
We begin by recalling for later use some well known facts about fermions. The fundamental observables for a system of n fermion degrees of freedom are
configuration operators QA,, ... , Q. together with their conjugate moments operators P,, P2.... , P. all acting as operators on a complex Hilbert spaces" and satisfying the canonical anticommutation relations: PjPk + PkPj = 2bjk ,
Qj Qk + QkQj = 2bjk ,
PjQk + QkPj = 0 .
(2.1)
We denote the complex algebra generated by the identity and the configuration
observables by IC(J), and the complex algebra generated by the identity, the configuration observables and the momentum observables all together by 'G'(:?) is the object of primary interest; but many aspects of its structure are most readily seen within the larger algebra lifl_*'). This algebra can be concretely represented as the algebra of observables for a spins-chain as follows. We define the matrices
I=[0
1],
-1]' Q-[1
U=10,
fi
P=[-
Let .s denote the n-fold tensor product of CZ with itself-
- .C2®... ®Cz n
times
and on W' define the operators
Qj =
U®... U®Q®1®... ®1 , Pj =
U®...
U®P®I®... ®1 .
(2.2)
where the Q and the P occur in the j1° places.
The operators Q,, ... , P...... P. just defined are easily seen to satisfy the canonical anticommutation relations. Of great use in studying the algebra that they generate is the fact that it is also the algebra generated by n "hard core boson" degrees of freedom. More explicitly, put
Qj=I®...1pQ®1®...®J,
Pj=I®...1QPQ1Q...®1,
(2.3)
155
With E. Carlen in Commun. Math. Phys. 155, 27-46 (1993) E.A. Carlen and E.H. Lieb
32
and call the algebra generated by the operators Q,, ... ,
P,, ... , P. the hard
core boson algebra, and denote it by W(17Y). To see that the two algebras coincide, put
UJ=1®...1®U®1®...®1
Vk
and
k-1
= r[ U;.
(2.4)
1=1
Then since PJQ; = P,Q; = iU;, each Vk belongs to both the hard core boson algebra and
Moreover Yk = VkQk
and Pk = VkPk
(2.5)
with the inverse relation given as well by Qk = VkQk
and
Pk = VkPk .
(2.6)
Thus, Qk, Pk are in and Qk, Pk are in What we call the hard core boson algebra was initially introduced by Jordan
and Klein [JoKI] as a first attempt to implement the Pauli exclusion principle mathematically. The transformation of observables (2.6) was discovered by Jordan and Wigner [JoWi] and used by them to write the algebraic relations characteriz-
ing the algebra in the familiar covariant form (2.1); today it is known as the Jordan Wigner transform. It has been used many times since for example, in a solution of the two dimensional Ising model by Schultz, Mattis and Lieb, whose paper [SML] can be consulted for references to other applications. It is also the key to Brauer and Weyl's treatment [BrWe] of the Clifford algebra on which the exposition in the rest of this section is largely based. Following them, we now show is the full matrix algebra on e. that
First, we introduce a basis in .f. Let e,
=(0')
and
e- , = (0)
be the standard basis of V. For each j = 1, ... , n, let a; be either 1 or - I. Then the unit vectors
®e provide a natural orthonormal basis for .,(''. Next introduce the fermion creation
and
operators
annihilation
c* = 2(Q; - iP,) and
2(QJ + iPJ) and their hard core boson analogs c;* = }(Q; - iP,) and eJ = j(Q, + iP,). Now put L(a,, .... fi;-, B;. where B, - c if a = 1 and Bj = ej if a1 - - 1. Then L(a .....a) e,, .,, = S2 .
_
.
where Q
0
Y
`\
12.7)
n limt
is a distinguished unit vector in .,W called the ground state. Also, L(a,.... , annihilates all the other basis vectors. Moreover, L*(a,, ... e,, , .. ,
Thus.
L*(r,, .... 156
r
Optimal Hypercontractivity for Fermi Fields and Related Non-Commutative Integration Inequalities Optimal Hypercontractivity for Fermi Fields
33
and this operator annihilates all other basis vectors. Manifestly this operator belongs to
and hence to W(,*) as well, and the 2" operators of this kind form
a basis for the full matrix algebra on .0. This concrete description of 16(Jf') in terms of spin-chain observables is the most useful for many purposes. Still. it is also useful to have a characterization of IC(,*-) which is less dependent on coordinates; i.e. on the choice of fermi configura-
tion observables Q1, ... , Q. Toward this end, consider the standard n-dimensional Hilbert space V equipped with its standard inner product (., ) and complex conjugation. Let .)f' denote 44'" considered as a real 2n-dimensional Hilbert space equipped with the inner product <X.y).ar = R(x,y) .
Then complex conjugation on C induces an involutory orthogonal transformation Jon .t'. Let 2 and Y respectively denote the eigenspaces of J corresponding
to the eigenvalues + I and - 1. The bilinear form on *' given by ;(x,
is
symplectic so that A' is naturally endowed with the structure possessed by the classical phase space of a system of n linear degrees of freedom. 1 is called the configuration space, 9 the momentum space, and the complex conjugation J is usually identified with time reversal. ') is characterized up to automorphism as the algebra The Clifford algebra with unit I such that:
(i) There is a linear imbedding J:..*' -'6(A ), and /(.)Y) generates 'h (S'). (ii) For all x, y e Jf',
/(x)/(y)+.f(r),f(x)=<x.y>a
.
(2.8)
To make contact with our previous concrete description, let . q,.... , qn; be an orthonormal basis of II " consisting of purely real vectors. For each j, let p, = iq;. Evidently .1 is spanned by {q,, ... , q"} and is spanned by ; p,, ... , p,}. Any +1; p;. Using the notation x e .7Y can then be written as x , ;q; + introduced above, put .,f : J(' (('(J') by
;.
,
rt
Q 1= 1
1=1
Let ;x,, ... , x2,,} be any orthonormal basis of 'rS'(. ) Then the monomials Q
(x,d 'f (X,) ....f(.,)
together with I form a basis for the algebra. It is easy to see that the product of any
two such monomials is a third. Though the multiplication rule can be simply expressed in terms of certain contraction rules, its precise form is not useful to us here. What is useful to observe is that since the right side of (2.8) is invariant under orthogonal transformations of .*', the multiplication law of these monomials does not depend on the choice of the orthonormal basis. For this reason, any orthogonal transformation R of .*' induces an automorphism of K (.X') which we shall also denote by R. Indeed, with R: (1) defined by
R(/ (x, I ... ;Y(X, I) _ .f (R(x,.I) ... ,f(R(x,,)) R is evidently invertible and by the remarks made just above one easily sees that for all A. B c- '6 (. (, R(AB) = R(A) R(B). which is to say that R is an automorphism.
157
With E. Carlen in Commun. Math. Phys. 155, 27-46 (1993) E.A. Carlen and E.H. Lieb
34
Finally, we remark that W(-V') is a *-algebra; there is a unique conjugate linear involutory antiautomorphism A'-'A* which is the identity on /(.iV ). It is given by
R(/(xa,) . . d (xa,,))* = /(xaw) . . . f (xa,) . .
Of course, on regarded as a matrix algebra, this is just the usual adjoint. Evidently the automorphism R of ''(1) induced by an orthogonal transformation R of .7Y is a *-automorphism; i.e. R(A*) = (R(A))*.
The facts that orthogonal transformations of _V- induce automorphisms of that T(1) is a full matrix algebra, and that all automorphisms of full matrix algebras are inner, i.e. of the form A a-' SAS ` for some nonsingular matrix S, are the basis of Brauer and Weyl's treatment of the spin representations of the orthogonal groups [BrWe]. We will use the fact that all automorphisms of are inner several times in what follows. Ill. Analysis on the Clifford Algebra
Let .4 be a von Neumann algebra of operators on some finite dimensional Hilbert space. By a trace on .fat we shall mean a linear functional T which is positive in the sense that T(A*A) > 0 for all non zero A in fat, and cyclic in the sense that T(AB) = T(BA) for all A and B in .sp. Such a functional is evidently
unitarily invariant in the sense that whenever A and U belong to Q, and U is unitary, then T(U* AU) = T(A). Since W(-*') is a full matrix algebra, it contains
all unitaries. Hence any trace on ''(,t) must assign the same value to all rank one projections, and thus must be a scalar multiple of the standard trace Tr on the matrix algebra. Henceforth, T shall denote this trace normalized by the condition that T(1) = 1 and Tr shall denote the standard unnormalized trace.
In the non-commutative integration theories of Dixmicr [Di53] and Segal [Se53], the trace functional T is the non-commutative analog of the functional that
assigns to an integrable function its integral. When the Hilbert space is infinite dimensional, some further regularity properties are required of r in order to obtain
a useful analog. Since all of our estimations will be carried out in the finite dimensional setting, we shall not go into this here, but shall simply refer the reader to these original papers as well as the accounts in [Gr72] and [Ne74]. Norms on ''(it) which are the non-commutative analogs of the L° norms can
now be introduced; namely for I S p < oc we put II All p
(3.1)
= (T((A*A)pI2))lip
and denote the operator norm of A by II A II ')P(J() shall denote It'(,V') equipped with the norm II'llp; evidently W2 (,Y ) is the Hilbert space of 2" x 2" matrices equipped with the Hilbert-Schmidt norm. Consider the monomials E, a,.....a,:B...... a4) = Qa, .
where a,>
.
. Q., Pn, .
.
. PQ4 ,
>a;and/,>...>(Jkandj+k>0.Evidently Ela,...a,.Bi... dk)Elai.....a,.Pi.....flw1 =
158
Optimal Hypercontractivity for Fermi Fields and Related Non-Commutative Integration Inequalities
Optimal Hypercontractivity for Fermi Fields
35
and thus IIE1a,....,a,:p,.....Pk111p= I for all p. Moreover T(E121.....a,tpi...../3k) = 0 .
To see this, first consider the case in which j + k is odd. The inversion x -- - x on and hence there is Jr is orthogonal. Hence it induces an automorphism of an invertible S in W (X') so that .....pwl = SE(a,.....a,.p,.....pkls
Then, by using cyclicity of the trace we get the desired result. Next consider the case in which j + k is even and, say, j > 0. Then write E1a, .... p...... ph1 = Qa, X, and
note that by (2.1), Q., X = - XQa,. Again the desired conclusion follows from cyclicity of the trace. It is easy to see from this that
=0
T(Ei ...... a,.p...... pkl Eln.....
(3.4)
unless the two monomials coincide. Thus, together with the identity, the assemFinally observe blage of such monomials forms an orthonormal basis for 'Gr2(
that since Qe_ = e= 0
(3.5)
whenever j >_ 1, and, as indicated, k = 0. It now follows that, restricted to W(1), r(A) =
(3.6)
for all A in "if (.K). Formula (3.6) is very important for us. It permits us to calculate the "physically"
relevant quantityin terms of the apparently mathematically simpler quantity T(A). Many familiar inequalities for L° norms hold for the 'C" norms as well [Di53].
This is true in particular of the Holder inequality 1
Ii ABII.:! 11AIIp11BIq
l
1
r= n +
Certain optimal inequalities expressing the uniform convexity properties of the L° norms also hold for the WP norms, and this fact constitutes one cornerstone of our analysis. The modulus of convexity 6p of V(.)r) is defined by 1
6p(c) = inf
1 - 21A +
BIIp:
!IA)=IIBII,=1,
I A - Blip = t: }
(3.7)
}))norms
for 0 < e < 2. For I < p < oc, by is always positive which means these are uniformly convex. Useful geometric information is contained in the rate at which 6,(t) tends to zero with e. It is known [TJ74] that for 2 < p < x , bp(e) - rp, but that for I < p < 2, bp(E) r.2. An optimal expression of this fact is given by the following theorem which was proved jointly with Keith Ball [BCL]:
159
With E. Carlen in Commun. Math. Phys. 155, 27-46 (1993) E.A. Carlen and E.H. Lieb
.16
Theorem 1. (Optimal 2-uniform convexity for matrices). For all nix in matrices
A and B and all p for I Sp 52, TrIA + BI" 2+ TrIA - BI" 2"
> (TrIAI")2," + (p - I)(TrIBI")r" .
(3.8)
For I < p < 2, there is equality only when B = 0. This result, which we interpret here as a statement about W"(,X''), is proved in the appendix in the special case that both A + B and A - B are positive; this is the only case in which we shall use it here, and the proof is considerably simpler in this case. The full result is proved in [BCL], in which other geometric inequalities for trace norms are proved as well. The theorem implies that
1)(r)2
(P 2
2
for
I < p <2
as one sees by considering A = (C + D)/2, B = (C - D)/2, II C I1" = II D II" = I and 11 C - D 11" = E. It is easily seen that the constant (p - 1)/8 cannot be improved.
We make our main application of this result in Sect. V. There we will also need to know that the norms on (4 (.)f) are continuously differentiable away from the origin for I < p < x . This is known [Gr75], but a simple proof can be based on inequalities of the form S"(c) > Ke'P' such as we have found above for I < p < 2. This proof, moreover, gives the modulus of continuity of the derivative, and is sketched in the appendix as well. Again, these estimates are independent of the dimension and therefore apply to the case of infinitely many degrees of freedom.
IV. Conditional Expectations and the Fermion Oscillator Semigroup
We arc particularly concerned with the subalgebra 16(2) of '6'(X ), and the conditional expectation [Di53, Um54] with respect to it shall play a basic role in our investigation. For any A in 't (1 ). the conditional expectation ;r -,(A) of A with respect to le(2) is defined to be the unique element of (6'(..1) such that Otherwise said, it, is the orthor(B*n,(A)) = r(B*A) for all B in
gonal projection from 12(,f') onto 162(1). It is well known that the conditional expectation is positivity preserving; a familiar argument shows that ir,(A*A) ? lr,(A)*n,(A). We can use the conditional expectation to give a useful expression for the oscillator semigroup for fermion fields.
Let R. be the orthogonal transformation of 1 given by Re(gj) = (cosO)q; + (sinO)p,
(4.1)
for each j. Of course Ro gives the evolution at time () on phase space i generated (p; + q; ). Let RB denote by the classical oscillator Hamiltonian H(p,q) the automorphism of 16(1Y) generated by the orthogonal transformation R, as in the first section. For each t > 0. define 6(t) = arccos(e-`) and define the operator P, (.f) by on
P,A = it,
160
n.,.,A
.
(4.2)
Optimal Hypercontractivity for Fermi Fields and Related Non-Commutative Integration Inequalities
Optimal Hypercontractivity for Fermi Fields
37
Note that n, is the natural imbedding of '(2) into W(Y), and regarded as such, it is a s-automorphism. Formula (4.2) is the analog of the familiar expression for the boson oscillator semigroup on L2(2, (27E) -'/' a-q2(2 d"q), i.e. the Mehler semigroup
P,'°"°") A(q) = J A(e-'q + (I -
e-2')ijzp)(2rr)-`!2e-p'12d"p
Note that since all of the operators on the right in (4.2) are positivity preserving, so is P,. Also, since the first two operations on the right preserve the V-norms, and since the conditional expectation is readily seen to be a contraction from WP(,*^) to `'P(2) for each p, it is readily seen that P, possesses this property as well.
To obtain a more familiar expression for P note that Re(,)
I)= e
k' F_t,
.
, ,
akl + (terms annihilated by 1r,) .
Hence P, (E(.,..
.. Ql. I)
e - k,p. "E[,..... Tkl'
,k I) = kEl .. _ .. Qy 1. It Evidently { P,: t ? 0} is generated by Ho where is easy to see that under the unitary equivalence between 1if2(2) and fermion Fock space.F described by Segal [Se56], Ho is equivalent to the usual number operator, or in other words, the oscillator Hamiltonian on .5t. Our primary goal is to prove optimal hypercontractivity bounds for P,. That is,
given l < p < q < oo we want to show that for some finite t, P, is a contraction from Vp(2) to Wq(2), and to find the smallest such t. Let IIP,11p-.q=sup{IIP,AIiq:1iAllp= I} .
(4.4)
As a first reduction, we shall show that the supremum on the right in (4.4) can be restricted to the positive operators A with 11 A 11 p = 1. In the boson case this follows
immediately from the fact that, in ordinary probability theory, the absolute value of
a conditional expectation is no greater than the conditional expectation of the absolute value.
In general, matters concerning the absolute value in the non-commutative setting are more troublesome than in the commutative setting. An example is provided by the Araki-Yamagami inequality [ArYa] which, specialized to our with context, asserts that the map A i-+ I A I is Lipschitz continuous on constant instead of the constant 1 which we would have in the commutative setting.
Thus while the conditional expectation in an operator algebra has many properties analogous to those of the conditional expectation in ordinary probability theory [Um54], it is not in general true that I n,(A)l will be a smaller operator than rr,(IAI). The following theorem expresses a useful property in this direction
which does hold, and after proving it we shall show by example that stronger properties do not hold. The theorem and its proof are easily extended to a more general von Neumann algebra setting by the methods in [Ru72]. Theorem 2. (A Schwarz inequality for conditional expec(ations). For all A in
and all p with t < p < Xo, I':ne(A)Ilp IIn,(IAI)IiP`2IInj(IA'I)11P`2.
(4.5)
161
With E. Carlen in Commun. Math. Phys. 155, 27-46 (1993) E.A. Carlen and E.H. Lieb
38
Remark. If we let F(A) denote II n,A II p, then the same argument which we shall use
to prove Theorem 2 also establishes that F(A*B) 5 F(A*A)`I2 F(B*B)12 .
(4.6)
In this form, the term "Schwarz inequality," by which we referred to (4.5), is more evidently appropriate. Moreover, inequalities of the type (4.6) are well known in matrix analysis for many familiar functions; for example when F(A) is the determinant of A or the spectral radius of A. Further examples can be found in [MeDS]. In [Li76] it is shown that for a function F that satisfies (4.6), and which is monotone increasing; i.e. satisfies F(B) > F(A) for all B Z A z 0, the following inequalities hold:
F\; ( AjB!
n
1112
F
1ij2
> B'B'/
and
A) C
IAI IU2F(,>
F\,Ei
In particular, these inequalities hold for F(A) = II n,A 11 p. Specializing the last inequality to the case m = I then yields (4.5). In our present case however, the proof of the (4.6) is essentially the same as the direct proof of (4.5). Nonetheless, it should not be considered novel that by taking I A * 1 into consideration as well as I A I , we can obtain a suitable bound on 1 1 n, A 11 p .
Proof Let A = UTAI be the polar decomposition of A. Then I! nf(A)11 p = t(CUTAI) for some C in W(2) with II C U,,. = 1. Let C = VI CI be the polar decomposition of C.
Both V and ICI belong to W(2) as well. Thus Iln_,(A)II,, = r(CUTA1"2IA1112) = r(ICI''2UlAIu2IA11 .2 VIC11,2)
<= t(ICl"2 UTAIU*ICI"2)"2 r(ICIu2 V*IAI
VICI1;2)112
= r(ICI(UTAIU*))`/2t((VICI V*)IA1)''2 < Ilns(UTAIU*) 11 o;2I it
Finally, we note that UTAI U* = IA*l.
U
Example. Let A be the matrix A = [0
I
lAl=f[0 Note that I A
IA*1=42[1 1].
and
1]
is in W(.2), but A and A I are not. One easily finds
ne(A)=z[1 1], n3(IAI)= 11n:(A)11
162
]. Then
=1
,
Iln,(IAI)II
=
0
I]'
f
and
Iln.,(IA*I)II, _
(IAI)Ilpn.
Optimal Hypercontractivity for Fermi Fields and Related Non-Commutative Integration Inequalities
Optimal Hypercontractivity for Fermi Fields
39
Theorem 3. (P, has positive maximizers). The norm of P, from 16'1'(2) to 10(.2) is achieved on the positive operators; i.e. (4.7)
11 P,IID-q=sup{f1P,AlIq:A>Oand IIAIIP= 1}. Proof. Since R. and it* are both s-automorphisms, IR0 ° n;A I = Re(,) ' (R6r,>°n3A)*I=Ret,i°nrlA*I. Thus IIP,AII,SIIP,IAIIIa12IIP,IA*IIIq%2
AI and
and of
course IIIA*IIIP= IIIAIIIP= IIAIIP.
V. Hypercontractivity for Fermions
Our main result is the following theorem which is established in this section. Theorem 4. (Optimal fermion hypercontractivity). For 11 P, Ilp-q = I exactly when
all
I < p 5 q < or,,
e-2,
=q -I
The heart of the matter is the following lemma: Lemma. For all 1 < p 5 2,
11 P,11 ,
. 2 = 1 exactly when e- 22' < (p - 1).
Proof. Fix a positive element A of W(2). Pick a basis {q,.....
of 2, and let
W(.21 _ 1)) denote the Clifford algebra associated with the span of the first n - I of these basis elements. It is evident from the form of our standard basis of 'W(2) that
A can be uniquely decomposed as A = B + CQ,,, where B and C belong to Then
A=B+
Jordan-Wigner
using the Now write
transform
we
can
.*,= ,f°rn-1,©C2
write (5.1)
so that B, C and VC can be considered as operators on the first factor *''r° - ,. Let (e+ ± e_)/./2, so that Qu , = ± u , . Then if v is any vector in u
=,.0-,,.
<(v®ut),A(v(u*)>,r
(5.2)
We see from this that since A > 0, so are both B + CV 2t 0 and B - C V ? 0. Now let Tr, and Tr2 denote the partial traces over the first and second factors in (5.1), so that with Tr still denoting the full trace, we have Tr = Tr, Tr2. Now applying Theorem I in the special case which is proved in the appendix li A 112 p =
+ C V.&P (TrIB 2)2/p(Tr, IB +
I C 2r"
"
ZiP
+Tr,I B 2
2iP
\
1
/
((Tri IBI°)2ip + (p - I)(Tr, CIP)Zro)
since V. is unitary. Thus 1 1A1I 2 ? II B Il v + (p _I) II C II norms on
p,
where the norms on the right are all
163
With E. Carlen in Commun. Math. Phys. 155, 27-46 (1993) E.A. Carlen and E.H. Lieb
40
for
Now we make the inductive assumption that the lemma has been established This is clearly the case when t = 1. We then have IIAIIp
where e
2'
11P,BIIi +(p - 1)1!P,C112 = (p - 1). But clearly from (4.3)11 P, CQ Il i = (p
norms are once again norms on
- 1)11 P,C 11 22, where the
Moreover by (3.4) P1B and P,CQ are
U + (p - 1) II p, C 1I 22 = 11 P, (B + I12 2 = II P, A 112. 2 Theorem 5. (Optimal fermion logarithmic Sobolev inequality). For all A E W(1), orthogonal. Thus JI P, B II
1
z
(5.3) r(IAI2InFAI')-(IIA112In 11A11')<2. P r o o f . By the lemma, 11 P, A I I 2 < I I A 1 1 2 +..- i.) and there is equality at t = 0. Both sides arc continuously differentiable, and comparing derivatives at t = 0 we obtain the result. Indeed,
dpJAII,=vhIAI; °(r(IAI°1nIAl)-HAII;InIIAII°)
(5.4)
and of course drIIP'AEI2=- _
Gross refers to the quadratic form on the right side of (5.3) as the Clifford Dirichlet form since it shares many properties of Dirichlet forms in the ordinary commutative setting. An approach to the development of a theory of Dirichlet forms in the non-commutative setting can be found in [AIHK].
0 and I < p < .7 ,
Proof of Theorem 4. By a deep result of Gross. when A .5
(P%2)'
p-I
'
Replacing A in (5.3) by Ap/2 and using the inequality just quoted we obtain, following Gross's ideas [Gr75],
r(A°InA)-IIAIl°PInhIAIIP. By combining this with (5.4) a differential inequality is obtained which implies that is a decreasing function of t when q(t) = I + e2t(p - 1). This establishes II P,A the result for A z 0, and by Theorem 3 it is established in general. By (4.3), P, I = 1, and therefore II P, II°-.Q is always at least 1 for all p and q. That the inequality is best
possible follows from a direct computation with one degree of freedom. To be precise, II P,(I + Q))IIQ = 111 + e-'Q) IIq is easily computed and compared with 11 1 + Q 11 ° [Gr72]. The first quantity is greater than the second if e 2' > (p - 1)! (q - 1). VI. Hypercontractivity for Bosons and Fermions Together
As a result of the present work and of earlier work on bosons, we know that for
t given by a-2t = (p - l)/(q - 1), both the fermion and the boson oscillator
164
Optimal Hypercontractivity for Fermi Fields and Related Non-Commutative Integration Inequalities Optimal Hypercontractivity for Fermi Fields
41
semigroups are contractive from the appropriate p-spaces to the appropriate q-spaces, and that this value of t is optimal for each case separately. It is natural to expect that the same condition governs hypercontractivity in a situation in which we have bosons and fermions together. This is indeed the case, as we now show using Minkowski's inequality in an argument based on Segal's
method for showing that the optimal conditions for hypercontractivity with m boson degrees of freedom are the same as for one degree of freedom. Let p(dx) = (2n)-,"'2 a"12 dx be the unit Gauss measure on IRt. Then in our mixed setting, with m boson degrees of freedom and n fcrmion degrees of freedom, the relevant p-space is
which may be regarded as consisting of 16p(:21,) valued measurable functions x" A (x) such that IIIAIII;= I IIA(x)II°p(dx) R^'
is
finite. This equation defines the norm on M°. For p = 2, .V° is naturally
isomorphic to the tensor product of the symmetric tensor algebra over C' and the antisymmetric tensor algebra over C" as shown by Segal. On the latter space we
have the mixed oscillator semigroup generated by the sum of the boson and fermion number operators
l
;_
,
111
Considered as operators on M", the operators °, which constitute this semigroup are given by
Y,A(x)= J M,(x,x')P,[A(x')]p(dx'),
(6.1)
where M,(x,x') is the Mehler kernel; i.e., the positive integral kernel for the boson oscillator semigroup P(,b°'°") discussed in Sect. IV. Of course P, denotes the fermion oscillator semigroup studied throughout this paper. Now successively applying Minkowski's inequality, our theorem on optimal fermion hypercontractivity, and Nelson's theorem on optimal boson hypercontractivity, we have for e- 2' <_ (p - 1)/(q - 1): q
J M,(x,x')P,CA(x')]p(dx')
III>,A1114=
µ(dx)
R^`
M,(x,x')!IP,A(x')liµ(dx')
5 R^ f
R^'
5
IIA
Rm
g
µ(dx)
g M,(x,x')!IA(x')II,u(dx') Iµ(dx)
/
\ Il
p
µ(dx)
Jq!"
/
= 1) A )j ,
165
With E. Carlen in Commun. Math. Phys. 155, 27-46 (1993) E.A. Carlen and E.H. Lieb
42
Appendix
Proof of Theorem I when A ± B ? 0. Let Z and W be the 2m x 2m matrices given by
Z-[0 A]' W-[0B -B] Our goal is to establish that for all r with 0 < r < 1,
(Tr(A + rB)° +Tr(A -
(Tr(A)'+ r2(p - 1)(TrBI)21,
2
or what is the same, Tr(Z + rW)z;p z (Tr(Z)p)21p + r2(p - 1)(TrI W jp)z/p
(A. 1)
First, note that the null space of Z + r W is exactly the null space of Z for 0 < r < 1.
Thus by carrying out all of the following computations on the orthogonal complement of this fixed null space, we may freely assume that Z + rW > 0 for all 0 5 r < 1. Next, both sides of (A.1) agree at r = 0, and the first derivatives in r of both sides vanish there as well. We define fi(r) to be Tr(Z + rW )p. Then the second derivative in r of the left side of (A. 1) satisfies d2
dr2
(fi(r)) zip =>
2
P
fi(r)
d2
u
dr2
0(r) .
The second derivative on the right side is just 2(p - 1)(TrI WIP)z;p, and we are left with showing that z
p ,(r)(2
-p),pdr-0(r)?(p-
(A.2)
1)(TrlW1p)2rp
for all 0 < r < 1. By redefining Z to be Z + r W, it suffices to establish (A.2) at r = 0.
Now dr fi(r) = p(Tr(Z + r W) (p-' W), since A ± B ? 0, Z + r W >_ 0 for small d
r, and we can use the integral representation (Z +
t'p
t.
rW)U'
0
Fl]dt It t+(Z+rW) 1
to conclude that d2
dr2
1
(0) = pc4 J
0
It
+Z WI
1
I
W
]dt
.
(A.3)
Consider the right side as a function, f (Z), ofZ for fixed W. It is easy to sce that
f is convex in Z. (Simply replace Z by Z + tX, with X self-adjoint, and then differentiate twice with respect to t; the positivity follows from the Schwarz inequality for traces.) Also, f (UZU') = f (Z) provided U is unitary and U commutes with W. In a basis in which W is diagonal, we form the set 11 consisting of the
166
Optimal Hypercontractivity for Fermi Fields and Related Non-Commutative Integration Inequalities Optimal Hypercontractivity for Fermi Fields
43
22m distinct diagonal unitary matrices, each with + I or - I in each diagonal entry. Each of these clearly commutes with W. Then
f(Z) = 2-2m Y f (UZU*) z U. *
f(2-2m
\
Y UZU*)
f(Zd,.,) ,
(!elf
where Zdj.g is the matrix that is diagonal in the basis diagonalizing W, and whose diagonal entries are those of Z in this basis. Replacing Z by Zd;,, in (A.3), the integration can be carried out, and we obtain dz
drz
2m
Zip-2) wI
(0) ? P(P - 1) Y
l
i=t J where z; and w;, respectively, denote the j" diagonal entries of Z and Win a basis diagonalizing W. Now consider i(0) = Tr(ZP) as a function of Z. It is clearly convex, and thus by the averaging method just employed, we obtain 2m
0(0) 2:
ZP
ll
1=((
To establish (A.2), we are only left with showing that
(`')
Puv ( ' (z(,
z) x2 )
(2.'
Iw,IP1z,P,
(A.4)
C,
but this follows immediately from Holder's inequality.
/J
To complete the proof, observe that equality in (A.1) for r = I and 1
denotes the j" diagonal element of Z + rW; these are the numbers z; + rw;, where z; denotes the j" diagonal element of Z. Let us assume that w; * 0 for some j. Then equality in Holder's inequality (A.4) requires that the vector with positive components z + rw; be proportional to the vector with components Iw,I. Thus, for almost every r in [0, 1] we require
=i+n+',=c(r)Iw,I for some number c(r) that depends on r but not on j. The left side above is a linear
function, and thus c(r) = a + rb for some numbers a and b. But then clearly h = w; I w;1, and all non-zero eigenvalues of W would necessarily have the same sign. This is impossible since TrW = 0. We now give an application of the uniform convexity implied by this theorem to the differentiability of the WP(,*-) norms. First we recall that for 2 < p < x. , the
modulus of convexity is given by an analog of an inequality of Clarkson for integrals which Dixmier [Di53] established for traces. Specializing to 4k''(. f'), this inequality reads A + B
P P
A-B <2[IIA11P+IIBIIP] 2<_p< x 2 .
P
which implies that in this range (cf. (3.7))
p (OP
(A.5)
167
With E. Carlen in Commun. Math. Phys. 155, 27-46 (1993)
E.A. Carlen and E.H. Lieb
44
For any non-zero A in '8°(.f), define 9(A) by U*
9(A)=
where A = UTAI is the polar decomposition of A. Let p' be defined by 1/p + 1/p' = 1. Then for I < p < oo, II9(A)ll = I and r(9(A)A) = II A IIp. Moreover, 9(A) is the unique element of V (.fit') with this property. We call the map the gradient map on IW°(.7Y). The next theorem sharpens a result of Gross [Gr75]. Theorem 6. (Holder continuity of the gradient map on 4B°()r )). For all I < p < x , the gradient map is norm continuous. Moreover, p
I'IA -Blip (P-1) forI
(A.6)
and I19(A)-9(B)11p.<4(p-1)
IIA - Blk
fort
IIA+Blip) Proof. First observe that for all I < p < a, ,
.
(A.7)
119(A) + 9(B)Il p.II A + Blip _ 91r((9(A) + 9(B))(A + B)) = 2(11 A 11° + II B II p) - SRr ((9(A) - 9(B))(A - B))
2(11 A + B11p) - 119(A) - 9(B)Ilp.ilA - B11°.
Thus, .IA =BI1p
9(A)- 9(B) 2
IA+BIlp)
2
p
(A.8)
But, by (A.5) and (3.7), we have
9(A)-1(B)
9(A) + 9(B) 2
p
p
p
(A.9)
2
for I < p 5 2. By combining (A.9) and (A.8) we obtain (A.6). Similarly, by combining (3.9) with (A.8) we obtain (A.7).
The continuity of the gradient map for '6°(.7Y) has been established by Gross
[Gr72], but his proof is more involved and does not yield an estimate of the modulus of continuity. It is now easy to establish continuous differentiability of the (B°(,f) norms away from the origin since, with h(t) =IIA + rB 11°, with A different from 0 and with t and s sufficiently small, we have
92r(9(A + IB)B) <
h(r + s) - h(t)
< 9 (9(A + (t + s)B)B)
(A. 10)
N
when s is positive. To see this, observe that h(t) = Rr(9(A + tB)(A + 1B)), and that h(t + s) -> 8tr(9(A + tB)(A + (t + s)B)) by Holder's inequality. By subtracting the expression for h(r) from the estimate for h(t + s) and dividing by s, we obtain
the inequality on the left in (A.10). The inequality on the right is obtained in an
168
Optimal Hypercontractivity for Fermi Fields and Related Non-Commutative Integration Inequalities
Optimal Hypercontractivity for Fermi Fields
45
analogous manner. When s is negative, the inequalities are clearly reversed. Letting
s tend to zero, we obtain dt (JA +
tBp,lr_o = 93T(!Y(A)B) .
References
[AIHK] [ArYa]
[BCL] [BrWe] [CL91]
[Da76] [Di53]
[Far]
Albeverio, S., Hnegh-Krohn, R.: Dirichlet forms and Markov semigroups on C'187 (1977) algebras. Commun. Math. Phys. 56, Araki, H. Yamagami, S.: An inequality for the Hilbert -Schmidt norm, Commun. Math. Phys. 31, 89-96 (1981) Ball, K., Carlen, E.A., Lieb, E. EL preprint 1992 Brauer, R., Weyl, H.: Spinors in is dimensions. Am. J. Math. 57, 425 -449 (1935) Carlen, EA., Loss, M.: Extremals of functionals with competing symmetries. J. Func. Anal. > 437-456 (1991) Davies, E.B.: Quantum Theory of Open Systems. New York: Academic Press, 1976 Dixmier, J.: Formes lineaires sur un anneau d'operateurs, Bull. Soc. Math. France 81 222-245(1953)
Faris, W.: Product spaces and Nelson's inequality, Hely. Phys. Acta 48 721-730 (1975)
[Fe69]
Federbush, P.: A partially alternate derivation of a result of Nelson. J. Math. Phys. 10 5a-52 (1969)
[Gr72]
Gross, L.: Existence an uniqueness of physical ground states. J. Funct. Anal. 10.
[Gr75]
52.104 (1972) Gross, L.: Hypercontractivity and logarithmic Sobolev inequalities for Clifford-Dirichlet form. Duke Math. J. 43 383- 396 (1975)
[Gr89]
the
Gross, L.: Logarithmic Sobolev inequalities for the heat kernel on a Lie group and a bibliography on logarithmic Sobolev inequalities and hypercontractivity. In: White Noise Analysis, Mathematics and Applications. Hida et al. (eds.) Singapore: World Scientific, 1990, pp. 1.1fl -130
[Gr92]
[HuPa] [JoKl)
Gross, L.: Logarithmic Sobolev inequalities and contractivity properties of semigroups. 1992 Varenna summer school lecture notes (preprint) Hudson, R., Parthasarathy, K.R.: Quantum It6's formula and stochastic evolutions. Commun. Math. Phys. 93, 301-323 (1984) Jordan, P., Klein, 0.: Zum Mchrkorpcrproblem der Quantentheorie. Zeits. far Phys.
[JoWi]
45. 751 765 (1927) Jordan, P., Wigner, E.P.: Ober das Paulische Aquivalenzvcrbot. Zeits. fur Phys. 41, 631-651 (1928)
[Li76]
Lieb, EH_ Inequalities for some operator and matrix functions. Adv. Math. 20 174-178 (1976)
[Li90]
Lieb, ELLGaussian kernels have only Gaussian maximizers. Invent. Math. 1(12. 179-208 (1990)
[Lin] [LiMe] [MeDS]
Lindsay, M.: Gaussian hypercontractivity revisited. J. Funct. Anal. 21313 324 (1990) Lindsay, M.. Meyer, P.A.: preprint, 1991
[Me85]
442 448 (1975) Meyer, P.A.: Elements de probabilites quantiques, exposes I V. In: Sem. de Prob. XX, Lecture notes in Math. 1204, New York: Springer, 1985 pp. 186 312
[Me86] [Nc66]
Morris, R., Dias da Silva, J.A.: Generalized Schur functions. J. Lin. Algebra 3,
Meyer, P.A.: Elements de probabilites quantiques, exposes VI VIII. In: Sem. de Prob. XXI, Lecture notes in Math. 1247, New York: Springer, 1986 pp. 27-81 Nelson, E.: A quartic interaction in two dimensions. In: Mathematical Theory of Elementary Particles, R. Goodman, L Segal (eds.) Cambridge, MA MIT Press. 1966
[Ne73] [Ne74] [Nev]
Nelson, E.: The free Markov field. J. Funct. Anal. 12. 211 227 (1973) Nelson, E.: Notes on non-commutative integration. J. Funct. Anal. 15 103 -116 (1974) Neveu, J.: Sur 1'esperance conditionelle par rapport a un mouvement Brownien. Ann. Inst. I3. Poincare Sect. B. (N.S.) 12. 105 109 (1976)
169
With E. Carlen in Commun. Math. Phys. 155, 27-46 (1993)
E.A. Carlen and E.H. Lieh
46
[Ru72]
Ruskai, M.B.: Inequalities for traces on Von Neumann algebras. Commun. Math. Phys. 26, 280-289 (1972)
[SML] [Sc53]
Schultz, T.D., Mattis, D.C., Lieb, E. H.: Two dimensional Ising model as a soluble problem of many fermions. Rev. Mod. Phys. 36, 856- 871 (1964) Segal, I.E.: A non-commutative extension of abstract integration. Ann. Math. 57. 401-457 (1953)
[Se56] [Se70]
Segal, I.E.: Tensor algebras over Hilbert spaces It. Ann. Math. 63, 160 175 (1956) Segal, I.E.: Construction of non-linear local quantum processes: 1. Ann. Math. 92.
[TJ74]
462-481 (1970) Tomtzak-Jaegermann, N.: The moduli of smoothness and convexity and Rademacher
[Um54]
averages of trace classes S,(I 5 p < co ). Studia Mathematics 50, 163 182 (1974) Umegaki, H.: Conditional expectation in operator algebras I. Tohoku Math. J. 6, 177-181(1954)
Communicated by A. Jaffe
170
With K. Ball and E. Carlen in Invent. Math. 115, 463-482 (1994)
Invent math. 115, 463 482 (1994)
Inventions mathematicae
Sharp uniform convexity and smoothness inequalities for trace norms Keith
Eric A.
and Elliott H. Lieb3,***
Department of Mathematics, Texas A&M University, College Station, TX, 77843, USA : School of Mathematics, Georgia Institute of Technology. Atlanta, GA, 30332, USA Departments of Mathematics and Physics. Princeton University. P.O. Box 708, Princeton, NJ, 08544, USA Oblatum 7-VII-1993
Summary. We prove several sharp inequalities specifying the uniform convexity and uniform smoothness properties of the Schatten trace ideals C,, which are the analogs of the Lebesgue spaces L, in non-commutative integration. The inequalities are all precise analogs of results which had been known in L,, but were only known in C. for special values of p. In the course of our treatment of uniform convexity and smoothness inequalities for C, we obtain new and simple proofs of the known inequalities for L,. I Introduction
The concepts of uniform convexity and its dual property, uniform smoothness, play
an important role in analysis. After reviewing these concepts in the L,, function spaces, we shall consider their extension to the Schatten trace ideals, C,, i.e., the setting in which functions are replaced by operators, and integrals are replaced by traces. The emphasis throughout will be on the optimal constants appearing in the various inequalities. These optimal constants are "natural", as will be explained later in the introduction: they are the constants one would obtain from an informed guess using elementary calculus. However, as is often the case in such matters, no ready-made arguments suffice to validate the informed guesses. A normed space X is said to be uniformly convex if, for each e > 0, there is a d > 0 such that if x and y are unit vectors in X with I; x - y II z 2e, then the
average (x + y)/2 has norm at most I - d. A normed space X is said to be uniformly smooth if, for all e > 0, there is a r > 0 such that if x and y are unit vectors
in X with I{ x - y II < 2t, then the average (x + y)/2 has norm at least I - cr.
Work partially supported by US National Science Foundation grant DMS 88-07243 ** Work partially supported by US National Science Foundation grant DMS 92-07703 **' Work partially supported by US National Science Foundation grant PHY90.19433 A02 ,C) 1993 by the authors. Reproduction of this article, in its entirety, by any means is permitted for noncommercial purposes
171
With K. Ball and E. Carlen in Invent. Math. 115, 463-482 (1994) 464
K. Ball et at
Figuratively speaking, the unit ball of a uniformly convex space is uniformly free of "flat spots", and the unit ball of a uniformly smooth space is uniformly free of "corners". Since the unit ball of X *, the dual of X, is the polar conjugate of the unit ball of X, it is not difficult to show that X is uniformly convex (and hence reflexive) if and only if X * is uniformly smooth [D]. Many applications of uniform convexity and smoothness require quantitative versions of these notions. The function bx given by
bx(e):=inf{1-}Ilx+y11:IIx11=llyll= 1, 1 x-yl:>2e)
(1.1)
is called the modulus of convexity of X. (N.B. The function bx is frequently defined with c in place of 2x. The definition used here simplifies several of the formulae involving bx and fits more naturally with the definition of the modulus of smoothness given below.) Clearly, X is uniformly convex if and only if bx is strictly positive for every r > 0. It might seem natural to define the modulus of smoothness by setting it equal,
at r, to
sup 'I - IIx+yil:ilx11 = 11yll = 1,1Ix-yI 521}
(*)
Clearly, X is uniformly smooth if and only if this supremum is o(1) at r = 0. The definition (*), however, would not be well adapted to the duality between uniform convexity and uniform smoothness. Instead, the function px given by Px(r): = sup{
His + v11 + Ilu - rll
- 1:IIu11 = 1, ;lv;I = 11
(1.2)
is called the modulus of smoothness of X. This definition arises from (*) if we rewrite the quantity to be maximized there in terms of is = (x + y)/2 and v = (x - y)/2, and change the constraint from ll is + v II = 11 u - v 1; = 1 to simply II is I; = 1. For small 1,
there is no substantial difference, and it is easy to show (see [K6]) that X is uniformly smooth if and only if lim, .oPx(T)/T = 0.
Lindenstrauss [L] has shown that with these definitions, the modulus of convexity of a normed space X and the modulus of smoothness of its dual X * are related by p,.(-r) = sup {1e - bx(s):0 < c < 11. (1.3)
This is a quantitative versions of Day's duality theorem [D]. Uniform convexity was introduced by Clarkson [C] who proved that every Lo space with I < p < o c. both uniformly convex and uniformly smooth. Clarkson proved inequalities which give bounds of the form ?c.,(r.) >_ (e/Kr.,)'
(1.4)
where r=pfor 25p< oc,andr=p/(p-1)for I_ (e/C)' for some constant C. (After Eq. (2.6) below, we make an apparently more restrictive definition of r-uniform convexity. The two definitions will be shown to be consis-
tent in Proposition 7, and the present definition is the simplest to use in the introduction.) Clarkson's bounds (1.4) only show that L, is r-uniformly convex with
r > 2 for all p + 2 while, actually, L, is 2-uniformly convex for I < p < 2.
172
Sharp Uniform Convexity and Smoothness Inequalities for Trace Norms
465
Sharp uniform convexity
The 2-uniform convexity of L, for I < p < 2 follows from a result of Hanner [H], who proved an inequality from which 5,., can be easily computed. Hanner's result is recalled in part (a) of Theorem 2 below. The best constant K,,2 in (1.4) seems to have been first determined by Ball and Pisier [BP], who gave a simple direct proof, independent of Hanner's calculation, that L, is 2-uniformly convex for such p. Their optimal 2-uniform convexity inequality is:
6,,,(E) > 2 F.2 for
I < p < 2.
(1.5)
Because of the dual nature of the notions of uniform convexity and smoothness, the
modulus of smoothness of Lo for 2 < p < .x satisfies an inequality of the form pL,(r) (K, 2 t)2. Again, this is a better estimate than that which follows from Clarkson's inequalities. A more detailed history of these and related inequalities will be presented in Sect. 11 of our paper.
Less is known about the corresponding inequalities for the trace classes C. Clarkson's inequalities were extended to C, partly by Dixmier [Di], and fully by Klaus [Si], with precisely the same constants and exponents as in the L, case. Tomczak-Jaegermann later showed that, as with L, C, is actually 2-uniformly convex for I < p < 2: 6,,(e) ? (E/K,,2)2
for
1 < p < 2.
(1.6)
Her proof proceeds by establishing the C. analog of Hanner's inequality when p is an even integer, then deducing the 2-uniform smoothness of C, for all p >_ 2 from this by interpolation, and then using Lindenstrauss's duality result to obtain the 2-uniform convexity of C, for 1 < p <= 2. Implicit in her proof is the fact that when p = 2k/(2k - 1) for some positive integer k, the sharp constants K,,2 for C,, coincide with those of L,; i.e. K,,2 = (p - 1),12 for such values of p.
The principal results in our paper are the determination of the best possible constants for all p in Tomczak-Jaegermann's theorem, and the proof that the C,
analog of Hanner's inequality holds for 1 < p < 4/3. and in the dual range 4 <_ p < oc. Our two main theorems are the following (in which II - II, denotes the L, or the C. norm):
Theorem I (Optimal 2-uniform convexity) For I < p < 2, the inequality
(IX 2
+(p-1)1YI;.
(1.7)
holds in the following cases:
(a) X and Y are functions in L,. (b) X and Y are matrices in C,. if 2 < p <_ cc , the inequality is reversed. The title of this theorem will be explained more fully in Sect. II. the point, of course,
is that validity of the inequality (1.7) implies 2-uniform convexity. Part (a) is an unpublished result of Ball and Pisier, and the cases of part (b) for p = 2k/(2k - 1) are, as we have said before, implicit in the paper [TJ] of Tomczak-Jaegermann. The rest is new. The constant p - I in (1.7) is clearly seen to be optimal as well as natural from the point of view of elementary calculus: if X and Y are real numbers with I YI much smaller than X. then the two sides of (1.7) agree to second order
173
With K. Ball and E. Carlen in Invent. Math. 115, 463-482 (1994)
K. Ball et al.
466
in Y. The fact that (1.7) is true for all pairs of numbers is Gross's two-point
inequality [G]. Theorem 2 (Extension of Harmer's inequality to Cp) For 1 5 p 5 2, the inequality 11X + Yl1o+:1X - YllPp ? (IIXIIp+ I! Yllp)p+ IIIXIp - II Yilplp
(1.8)
holds in the following cases: (a) X and Y are functions in Lp. (b) p < 1 and X and Y are matrices in Cp.
(c) X and Y are matrices in Cp such that both X + Y and X - Y are positive semidefinite.
For 2 < p < oo , the inequality is reversed and the restriction in (b) becomes p ? 4,
and the restriction in (c) changes to the restriction that X and Y are positive semidefinite.
Part (a) is Hanner's inequality, and the cases of part (b) in which p = 2k are due to
Tomczak-Jaegermann [TJ]. The rest is new. As we explain in the first proof of Proposition 3 below, the inequality (1.8), whenever it holds, implies the inequality (1.7). Thus, if the conditions under which we establish (1.8) were not more restrictive
than those under which we establish (1.7), Theorem I would be a corollary of Theorem 2. The paper is organized as follows: In Sect. 11, we review the large number of inequalities bearing on the uniform convexity of Lp spaces. Thus, Sect. II consists largely of known results which are presented because of the light they shed on the problems solved in this paper regarding the uniform convexity of Cp, and those
that remain open. To our knowledge, such a systematic compendium of these inequalities has not appeared before, and we hope it will be found useful. There are however some new results and some new, simpler proofs. Finally in Sect. III we prove Theorem 1, and in Sect. IV we prove Theorem 2. Theorem 1 has been applied by Carlen and Lieb [CL] to prove a conjecture of Gross, which arose in his work on quantum field theory. Other applications of the kinds of the inequalities that we discuss here are given, for example, in Pisier's book
[P]. Although all of our theorems are stated and proved in the language of matrices,
the proofs go through without any change in the context of linear operators on a Hilbert space. By the results of Ruskai [Ru], they can even be extended to a natural Von Neumann algebra context.
II Uniform convexity and smoothness in L. While this section is largely focused on inequalities relating to LP spaces, we state certain definitions and prove certain results in the general normed space setting so that they are available to us in the next section. For the rest of the paper, q denotes the dual index of p, i.e., I /p + l 1q = 1. The notion of uniform convexity was introduced by Clarkson who proved four
inequalities. The two that imply the uniform convexity of Lp spaces are the
174
Sharp Uniform Convexity and Smoothness Inequalities for Trace Norms
467
Sharp uniform convexity
following, in which x and y are functions in L,:
x+y
it
2
pl"°
p
+
x2y /
p
(11x11',+ IlYllpl`rv 2 J
2 5 p 5 oo
(2.1)
)lip for I <_p<2.
(2.2)
for
and
x+y
2
x-y
q
+
2
p
(I1x'lo2 :IY'l
`
p
In the cases 1 5 p 5 2 and 2 5 p 5 oc, the inequalities in (2.1) and (2.2), respectively, hold in the reversed sense. These are the other two Clarkson inequalities - the ones which imply that the Lp spaces are uniformly smooth. They follow from (2.1) and (2.2) by an elementary duality argument. The inequality (2.1), which involves only p powers and not q powers as well, is simpler to prove, and is known as the "easy" Clarkson inequality. In fact, (2.1) is not only easier to prove, it is actually a consequence of (2.2). This is so because both
inequalities can be viewed as statements about the norms of certain linear operators. Viewed as such, (2.1) is weaker than the dual inequality to (2.2). More concretely, for 15 s, t S oc, equip L. x L. with the norm II' 11,., given by II(x,Y)11,.,=((Ilxlh+Ilyll`<)/2)'". Also, define the operator B:L,,,-'L,,, by B(x, y) = ((x + y)/2, (x - y)/2). Then (2.2) is equivalent to the statement that B is a bounded operator from L,,, to Lp., for 15 p < 2 with norm 2/0. But since B is self-adjoint, it has the same norm as an operator between the dual spaces; i.e. B is bounded from Lq.p to L9 with norm 211p for 1 <_ p 5 2. Finally, since p < q, 11(X,Y)1i°,,, 5 ;(x,y)11,.q, and hence B has norm 2" from L°,9 to L°,9 for 2 5 q 5 oo, which is clearly equivalent to (2.1). Since these bounds on the norm of B (which are equivalent to (2.1) and (2.2)) are log-linear in 11p, they can be proved by interpolation between the elementary cases
p = 1 (Minkowski's inequality for L,), p = 2 (the parallelogram law) and p = oo (Minkowski's inequality for L,,), as observed by Boas [Bo]. This same approach was later used by Klaus [Si] to establish the C, analogs of Clarkson's inequalities. It is convenient also to have the following inequality, obtained from (2.2) by rearranging some powers of 2. If x and y belong to L. where either s = p or s = q, then
l
Ilz+yll,+Ilx-yll:l'4=(11x11SP J
2
+;IYfl:)'rp for
1<=p52.
(2.3)
Replacing x and y respectively with x + y and x - y, and rearranging some powers of 2, we see that the inequality reverses when p and q are interchanged. That is,
IIx+YII°+I:x-ylls "p(Ilxll°+Ifyll;)'/° 2
for
1<-p<2.
(2.4)
It follows directly from (2.3) that if x and y are unit vectors in one of the spaces L, or L°, and Il x - y II = 2e then Ilx 2
y II < (1 -
E°)t/° 5 I
- E9
so that b,,,(e) >_ 0°/q. Similarly, from (2.4) one sees that p,_,(r)5 rp/p.
175
With K. Ball and E. Carlen in Invent. Math. 115, 463-482 (1994) K. Ball et at.
468
We next turn to the sort of inequalities that figure in our Theorem 1. We begin with some definitions and general considerations intended to clarify the relations among all the inequalities that we consider. Let X be a uniformly convex normed space, and suppose that bx(E) > (E/CY for some C and r > 1. Then with 11x 11 = II Y II = I and 11x - y II = 2E, we have that
x+yl 2
5 (1 - (E/CYY 5 1 -
)
r
(E/CY,
and thus, with K = C(1/t)-", 1/t + 1/r = 1,
x+y
x-y
2
2K
'511x11'+ I1y11' 2
for all x and y such that II x II = ll 1' II By replacing x with x + y and y with x - y, we
find that
IIx+YII'+ IIx-Y11' 2
11x11'+ IK`yll'
(2.6)
for allxandysuchthatllx+yll=IIx-YII. As promised in the introduction (after (1.4)), we now impose a definition of r-uniform convexity that may seem more restrictive than the one we gave before. Proposition 7 below shows that the two definitions are equivalent, up to constants. It is the constant in (2.6), figuring in the second definition, that is the main object of our attention. A normed space X is said to be r-uniformly convex for some r e [2, oo ) if there is
a constant K such that (2.6) holds for all x,y a X. The best constant K is called the r-uniform convexity constant of X. When X is r-uniformly convex, so that (2.6) and hence (2.5) hold, it is immediate from the latter that 5X(e) >_ (E/KY. Thus r-uniform convexity implies the validity of a lower bound of the form 6x(E) >_ (E/C)' for the modulus of convexity; i.e. the condition under which we called X r-uniformly convex in the introduction. Similarly, X is said to be t-uniformly smooth for some t E (1, 2] if
IIx+YII'+IIx-Y11'c11x11'+ 2
II KYII',
(2.7)
for some K and all x, y E X. The best constant K is called the t-uniform smoothness
constant of X. We shall show at the end of this section that the t-uniform smoothness constant of a normed space X equals the r-uniform convexity constant
of its dual X where, as usual, 1 /r + I /r = 1. When (2.7) holds, we have that for all x and y with II x I! = 1 and II YII = T
IIx+YI1 + ilx-YII
'Ix+y11'+ IIx-YII'
2
2
Hence, by (1.1), t-uniform smoothness implies an estimate of the form pX(T) 5 (CT)'.
Proposition 7 shows that the reverse implication holds as well. The parallelogram identity shows that Hilbert space is 2-uniformly convex and 2-uniformly smooth, and it is readily seen that the exponent 2 is the best that can occur for each property. Clarkson's inequality shows that when 1 S p 5 2 then
176
Sharp Uniform Convexity and Smoothness Inequalities for Trace Norms
469
Sharp uniform convexity
each of LP and LQ is q-uniformly convex and p-uniformly smooth. As we have remarked, these exponents are not in general the best possible, despite the fact that the constants in Clarkson's inequalities are always sharp. The actual situation is the following: For 1 < p 5 2, L. is 2-uniformly convex though no better than p-uniformly smooth while for 2 < q < oc , L, is 2-uniformly smooth as well as q-uniformly convex.
These facts follow from Hanner's inequality (Theorem 2(a) of the introduction) which determines exactly the moduli of convexity and smoothness of all L° spaces. The optimal 2-uniform convexity inequality is the following: Proposition 3 then
(Optimal 2-uniform convexity for Lp) If I S p < 2 and x and y e L,
.Ix+yiID+ Ilx - yll; z Ilx!II+(p_ I)Ilyll;.
(2.8)
2
For 2 <- p < oo , the inequality is reversed.
Remark. The inequality (2.8) holds for any normed space for which (1.8) holds, as we will soon show. Inequality (2.8) does not seem to appear in the literature in quite this form but it is probably folk-lore. Ball and Pisier noticed that it follows from
Gross's two-point inequality using arguments which (in the context of general Banach lattices) go back to Figiel [F]. The reader will note that (2.8) is not identical to (1.7), but we shall soon see that the validity of (2.8) for all L° spaces implies the validity of the apparently stronger (1.7).
First proof. To deduce (2.8) from Hanner's inequality recall that a special case of Gross's inequality [G] states that if 1 5 p 5 2 and a and h are real.
la+hl°+Ia-bI°"° 2
Now if x, Y e L.
rllx+yllo+Ilx-y'ilP)1J2 '>rllx+yllo+ Ilx - yllglIO 2
``
J
2
> E(xp + I'yll°)°+IIx1I°- Ily1l°I°1':° J
2
?(IIxi' +(p -
1)Ilyi:;)1;1
where we have used, in succession, Holder's inequality, Hanner's inequality (1.8). and Gross's inequality with a = II x iI ° and h = II y II P. O 177
With K Ball and E. Carlen in Invent. Math. 115, 463-482 (1994)
470
K. Ball et al.
Second proof. An alternative proof of Proposition 3 consists simply of showing that for II y II small,
IIx+YIIp+ Ilx - yllp, 2
Ilxllo+(p - 1)IlYllp+o(IIYII;),
(2.9)
and then observing that this infinitesimal form of (2.8) is equivalent to the full statement. That is, (2.9) is the same as
ds2x+sy11o2 11x-sy11P>tAp1)11x11p
11Y11;,
(2.10)
)ls=O
which is easy to establish for LP functions by elementary calculus. Proposition 3 follows from this by integration with respect to s E [0, 1]. Q As will be shown later, inequality (2.10) is also true for matrices in Cp, and this
will form the basis of the proof of Theorem 1 given in Sect. III. But there is an important difference between the commutative and non-commutative cases. For functions x and y in LP,
l
ds2(Ilx+syiIPP2 Ilx - syll°ll
= p(p - 1)f lxIP 21y12.
(2.11)
and the latter dominates p(p - 1) 11 x II v 211 Y 112 by Holder's inequality (since p < 2). For matrices, the analogue of (2.11) is false: one always has
d2IIX+sYIlo+11X-sYllpl dS2(
2
5p(p-1)TrIXIPlYl 2 )Id-o
and equality need not hold. The problem is to find a replacement for (2.11) in the non-commutative setting. Inequality (2.8), as we said, is apparently weaker than (1.7) since
IIx-y11°\ <(Ilx+y12+ I1 x-y11212 l 2 l (lix+y11P+2 for I < p < 2. However, a simple doubling argument shows that (1.7) is actually a consequence of the fact that (2.8) holds for all Lp spaces. This argument is also valid for CP, and we give it in detail in Sect. III, Eq. (3.5}{3.6). The reader can easily
translate the CP version into the L. version.
Note also that in the first proof of Proposition 3, which was based on Hanner's inequality, we actually arrived at (1.7) in an intermediate step. In the CP setting, we do not possess a full analog of Hanner's inequality, and so we shall prove Theorem I by adapting the second proof of Proposition 3 to the C. setting. The following diagram shows the relationships between the several
expressions mentioned above. Connecting lines indicate inequality between
the expressions. All of the indicated inequalities hold in both CP and Lp except for that indicated by the line labeled Hanner, which is only known to hold in C. in the special cases specified in Theorem 2. In each expression x and y are elements of L. or of Cp, and q is the index conjugate to p. For 1 5 p S 2, the quantities increase as one goes up the page; for 2 < p 5 oc , they decrease.
178
Sharp Uniform Convexity and Smoothness Inequalities for Trace Norms 471
Sharp uniform convexity
C
I1= + yllx,+Il=-YIl; 1
/a
J
2
i
1
Holder
i
(II=+yllp+Il=-yllo/r 2
BCL (Thm. 1)
Harmer (Thm. 2)
"strong" Clarkson
f (114, + HOOP + 111-11, - HOW '
l
Gross
X
/r
2
"strong" Clarkson
for numbers
(II=II; + (P-1)Ilyll;)'/2
(lI4p + I1Y111P
I/°
Fig. 1. Relationships among the inequalities
We turn now to a proof of Hanner's inequality for L, which yields Clarkson's inequalities along the way, and to a simple duality result which shows
that optimal constants obtained for q-uniform convexity of a normed space immediately yield optimal constants for p-uniform smoothness of its dual (and conversely).
Lemma 4 (Variational characterization of sums of p" powers) For I < p < x: define a = ap: [0, cc) [0, oo) by
a(r) = (1 + r)"_' + 11 - rl" sign(I - r). Then for all x, y E 68 Ix+yl°+Ix-yl°={()}{(r)lxVsup
+a(1/r)iyl°:0
the sup or inf being taken according as p < 2 or p > 2.
179
I.
With K. Ball and E. Carlen in Invent. Math. 115, 463-482 (1994)
472
K. Ball et al.
Proof Assume I < p 5 2; the proof for p ;>= 2 is similar. Plainly, it may be assumed
that 0 < y 5 1 = x. For r = y; one readily checks that a(r)+ a
yP
=
(1
+y)P+ (1 -y)P.
To see that (1 + y)P + (1 - yf a(r) + a(l/r) yP for all r, it suffices to check that the latter quantity attains its maximum when r = y. But dr(a(r) + a(l/r)y") = a'(r) - -Z a'(1/r)y'
=(P-1)L(1+r)P-z-I1-rIP-Z-Y1+llv-2
r z \\ r
)
- llP-Z/J
r 2
-I1-rlv
2 5 0 and I + r Z 11 - rl, the last factor is non-positive. Thus, the whole is non-negative for 0 < r < y and non-positive for r > y.
Proof of Theorem 2(a) (Hanner s inequali(y) Again assume 1 < p 5 2 and let x, y e L. Then
lix+y!IP+ Ilx - yFVP= f(Ix+ylP+Ix-ylP)= f sup {a(r)IxIP+a(l/r)Iy1P} ,>sup f (a(r)IxIP+a(1/r)IyIP)=sup{a(r)IlxllP+a(1/r)IlylIP} _ (Ilxll + 11YII)P+ 111 x!l - 11y111P
Proof of (2.2) Let us also show how Lemma 4 can be used to deduce the "hard" Clarkson inequality. This time we shall prove it in the uniform smoothness range, i.e. for Lq with 2 <- p < oe. Since I l x + yIIP+ Ilx - y l l P < a(r) I !
a(l/r) IIyliP
for all r, it is enough to find an r for which the right side equals 2(11 x 11 q + I y ii'f'q. Set u = II x II p, v = 11 y II;, r = v/u, and assume that v 5 u. Then
a(r) i x 11 P + a(1/r) 11 y 11 P = a(v/u) uP -' + a(u/v) vP ' = 2(u + v)P = 2(IIxilP+ 11y11P)p1q.
We now prove the duality results mentioned earlier. These results hold in general; no reference to LP or CP is made.
Lemma 5 (Duality for q-uniform convexity and p-uniform smoothness) Let X he a normed space with dual X *. The p-uniform smoothness constant of X (the constant K in (2.7)) is equal to the q-uniform convexity constant of X * (the constant K in (2.6).
Proof. Suppose that the q-uniform convexity of X * is K and let x, y e X. We denote norms in X and X * indiscriminately by 1I.11 and trust that the meaning will be clear. There are unit vectors i. and p in X * such that A(x + y) = 1; x + t 11
180
Sharp Uniform Convexity and Smoothness Inequalities for Trace Norms Sharp uniform convexity
473
and u(x - Y) =IIx -YII. Define 44e X * by O= Z -''9 II x + y II ° - A and with Z=(IIx + yII°+11x-yII°)/2. Then, II4I1°+ 4, 11 0 II V = 2 and we have IIx-YII°l'i°-4,(x+y)+2 r(x-y)=(+_I/.,)(x)+ 0
(iix±Yii"+2
(_:_)(y)
<(II0
1IKy11 P)
2 110114
IIJ11°1'1`(IIx11°+ 2
=(Ilxll°+
J
IIKYII°)'r°
11Ky11°)u°.
The first inequality is Holder's inequality for numbers, and the second is (2.5) with r = q. The other implication is similar. Lemma 6 (Duality for Hanner's inequality) Let X be a normed space with dual X
Let I < p S 2 and 1/p + 1/q = 1. Then the validity of
III+YII"+IIIO119:! (II0II+IIJ1II)9 +III0II-IoIII°
(2.12)
for all 0,41 a X* implies the validity of
IIy+zll°+ IIy-z11°?(Ily11 + 11z11)°+IIIy11 - IIz111°
(2.13)
for all y, z e X. Similarly, the validity of (2.13) in X implies the validity of (2.12) in X *.
Proof. Suppose first that (2.12) holds in X *. To establish (2.13), we first rewrite it in
terms of u = y + z and v = y - z, so that what we must show is: 2°(IIvII°+ IIv111)>=(IIu+v11 + IIu-v11)°+IIIu+v11 - IIu - vlll°. (2.14)
We may assume without loss of generality that
and that
11 u + v 11 = I,
r: = 11 u - v11 S 1. Then the right side of (2.14), which we call RP. can be rewritten (as
in Lemma 4) as
RP =allu+vII°+fIlu-v11°=(1 +r)°+(I -r)°, where 2=(I + r)°-' + (1 - r)°-' and # = r' -°[(1 + r)°-' - (I - r)° -']. As in the proof of Lemma 5, we choose unit vectors A and p in X* such that ).(u + v) =IIu + v 11 and u(u - v) = 11 u - vII . Then we define 4,=aR-°wllu+v11°-'A
and
4G=/3R-pjallu-vII°-'µ.
Thus,
R=4,(u+v)+i(u-v)=(0+4,)(u)+(¢-t,li)(v) 110+4111 (lull + 110+01111vII :5- (I10+0119+ 110-0Iu )''9(01ull°+
To complete the demonstration of T:=110+4119+110-4,119 24. By (2.12), TSR-°[allu+vll°-1
= R °[a +
ffr°-']9 + R °[a
(2.14),
we
have
to
show that
+flIlu-v11°-']q+R-°[allu+vll°-'
-(illu-v11°-1]q
- lr°-']4 = 29.
A similar proof works in the other direction to go from (2.13) to (2.12)
181
With K. Ball and E. Carlen in Invent. Math. 115, 463-482 (1994) K. Ball et at.
474
Remark. In our application of Lemma 6 to the proof of Theorem 2, we need the following refinement whose truth is evident from the proof given above. We take X = CD and X* = Cq. Then the validity of (2.13) with the extra constraint that y + z andy - z are positive semidefinite matrices implies the validity of (2.12) when 4) and 0 are positive semidefinite matrices. We close this section with a proposition showing the consistency of the two definitions that we have given for r-uniform convexity; i.e., the ones following (1.4) and (2.6).
Proposition 7 (Equivalence of definitions of r-uniform convexity) Let X he a normed space. Then (2.5) holds for some constant K and all x, y e X if and only if bx(E) z (E/C)' for some constant C. Similarly, (2.7) holds for some constant K and all
x, y e X if and only if px(r) 5 (Cr)' for some constant C. Proof. We have already seen that (2.5) and (2.7) imply the indicated bounds on bx and px respectively. Suppose first that px(r) 5 (Cr)' for some constant C. Of course, 1 < r 5 2. Then for all Il x 11 = 1 and Il y ll 5 1,
IIx + yll + lix - YII _ 1 2
(C11y11Y.
Define numbers h and ft by b:=IIx + Y11 + Ilx - YII
and
2
p:=
IIx + YII - IIx - YII
IIx+YII+IIx-Y1,
Then
Cllx+yll'+ Ilx-YII')"'_ IIx+yll+ IIx - YII J
2
b(I +/1)'+(I -p)' J R 2
2
1J.
(2.15)
LL
The function of fi on the right side in (2.15) vanishes quadratically at the origin. Thus, a simple estimation using Taylor's theorem shows that it is no greater than D,fl2 for some constant D, depending only on r. Then, since 1#1 < II y li /h
Ii y II and
1 5 r 5 2, we have from (2.15) and the assumption on px that
Ifx+yll'+IIx-YII'
-`
-
(2.16)
for all x and y with Ii y II < 11 x II = 1, where K, depends only on C and r. Therefore,
IIx+YII'+ I!x - yll'-<11x11'+K;iiy1'
(2.17)
2
for all x and y with II y Il 5 11x 11. Finally, since we may assume that K,
t
1, (2.17)
holds for all x and y. Next, suppose that bx(E) >_ (E/C)' for some constant C. Then by (1.3), px.(t) = Sup{tE - bx(E): 0 S E 5 1) 5 Sup{rF. - (F./CT: 0 5 E 5 00 } = (Cr)'
182
Sharp Uniform Convexity and Smoothness Inequalities for Trace Norms
475
Sharp uniform convexity
where I/r + 1/r' = 1.Then, by what we have shown above, there is a constant K so that (2.7) is valid in X *.But then by Lemma 5, (2.5) is valid in X with the same
constant K. 0 Ill Optimal 2-uniform convexity inequalities for trace norms Norms on the space of n x n matrices which are non commutative analogs of the L, norms can be defined in terms of the trace by IIXII,=(Tr((X*X)vi2)'w=(Tr((XX*)pr2)up
(3.1)
for 1 <_ p < oo. For p = oo, 11 X 11p denotes the operator norm of X, as usual. The
analogy can be made quite close, and it has been developed in a von Neumann
algebra context by Segal [Se] and Dixmier [Di] as part of their theories of non-commutative integration. Many familiar inequalities for L. norms also hold for the C, norms. This is true, in particular, of the Holder inequality IIXYU,.
11X11,11 YIIq. 1/r=
1/p+ l/q.
There are, however, other inequalities for L, norms which do not hold for the C. norms. Many examples are connected with the poor behavior of the map
X HIXI = (X*X)'n.
(3.2)
For example, if f and g are complex valued functions in some L, space, then II if I - Igl 11, :5 Ilf - g Il, This is not true for C,, and, when p = 2, the factor of f in the Araki-Yamagami inequality [ArY] 111 X I - I Y1112 < f 11 X - Y 112 is optimal whenever n z 2. As we have asserted in Theorems 1 and 2, however, almost all of the optimal inequalities expressing uniform convexity and smoothness properties of LP spaces have exact analogs which hold for the C, norms. Most of this section is devoted to the proof of Theorem 1. Before giving the proof we briefly discuss the history of uniform convexity inequalities for C, as we know it. The first such inequality was established by Dixmier [Di] who proved the C, analog of the "easy" Clarkson inequality (2.1) by means of interpolation. As with L,, this implies that for 2 5 p < oo, bc,(e) Z (1/p)ep.
Interpolation was later used to establish the analog of the "hard" Clarkson inequality (2.2) which implies the uniform convexity of Lp for 1 < p:5 2. Such a proof has been given by Martin Klaus, and is sketched in [Si]; to some extent it is modeled on Boas' proof [Bo] of (2.2) for L,. This result implies that for I < p < 2, C, is at least q-uniformly convex. Later, Tomczak-Jaegermann showed that for I < p 5 2, C, is actually 2-uniformly convex and Cq is 2-uniformly smooth. Moreover, she showed for q = 2k, that the sharp 2-uniform smoothness constants of Cq are the same as those for L. (so that the corresponding equalities of 2-uniform convexity constants hold by Lemma 5).
183
With K. Ball and E. Carlen in Invent. Math. 115, 463-482 (1994)
K. Ball et al.
476
Before we prove Theorem 1, note that t i-. tP12 is concave for 1 < p < 2 and therefore (1.7) immediately implies that
IIX+ YII;+IIX - YIIP> 2
IIXI12+(p- 1)I1Y11;,
(3.3)
which expresses the 2-uniform convexity of C, in the usual way. It follows that Sc,(E) >_ (P 2
for
1)s2
1 < p < 2,
(3.4)
and thus the analog of (1.5) holds for C,. We now observe that (1.7) is only formally stronger than (3.3). To see that (3.3) implies (1.7), consider the 2n xx 2n matrices given in block form by
Z = LO 01,
W= [ U
o r].
(3.5)
Then
TrIZ + WIP=TrIZ - WIP=(TrIX+ YIP+TrIX - YIP) and thus,
IIZ+ W(,-IIZ- W112=(IIX+YIIP+IIX- YIIP)2/P Since also 11ZII; = 221PIIXII2 and II W11P = 22/P11 YII2, (3.3) implies
IIXIIo+(P-1)11Y41;=2 2;°(11ZIIn+ (P-1)11W112
l
<
W112 + 11Z - W11PI 2
=(I;X+ YIIP+IIX - YI(P2`P 2
which is (1.7).
Proof of Theorem 1 First, we reduce to the case in which X and Y are self-adjoint. Consider the 2n x 2n matrices given in block form by
C - [X*
]' D=LY* ol.
Clearly, if (1.7) holds for the 2n x 2n0matrices C and D. it holds for X and Y. Since C and D are self-adjoint, it suffices to prove inequality (3.3) for such matrices. We therefore assume without loss of generality that X and Y are self-adjoint. Let Z and W be defined in terms of X and Y as in (3.5). Then we can rewrite (1.7) as
Tr(IZ + rWIP)2'P > (TrlZIP)2iP + r2(P -
(3.7)
1)(TrIWI°)2,P.
First, note that without loss of generality we may assume by continuity that the union of the ranges of Z and W span Ctn. Then det(Z + rW) is a polynomial of order exactly 2n in r, and it has at most 2n zeros for 0 <-- r <- 1. We will avoid these
values of r below in our computations. We define O(r) by
O(r) = TrIZ + rWIP = Tr((Z2 + r(ZW + WZ) +
184
r2W2)P'2).
Sharp Uniform Convexity and Smoothness Inequalities for Trace Norms 477
Sharp uniform convexity
Then i/i(r) is continuously differentiable and (p/2)Tr((Z2 + r(ZW + WZ) +
+ 2rW2).
W2)vl2-')((ZW + WZ)
r2
With the aid of the integral representation (Z2 + r(Z W + WZ) + r2 W2 )t"l2 - 1
= w f tw/2 -" U
1
t+(Zz +r(ZW+WZ)+rz W z )
dt
'
(3.8)
we see that di(r)/dr is again continuously differentiable. Now, both sides of (3.7) agree at r = 0, and the first derivatives in r of both sides vanish there as well. Moreover, the second derivative in r of the left hand side of (3.7) satisfies
ziv_2pe(r)
d2
,2-v)rod2
dr2O(r),
while the second derivative on the right side of (3.7) is just 2(p - 1) (Tr I WI')2ip. It therefore suffices to show that 10(r)(2 -v)rod 22O(r) >
(3.9)
(p - 1)(TrI WIO)2i0
for all 0 < r < 1. By redefining Z to be Z + rW, it suffices to establish (3.9) at r = 0. Since Z + rW is non-singular, after the redefinition, IZI will be strictly positive. We now claim that
d2TrIZ+rWI°
,-o> a2TrIIZI+rWI°I'
(3.10)
To see this, note that by the integral formula (3.8), dz
dr2TrIZ+rWIr ,
TrIZIV-2 W2_(p/2)/JvJta'2 'Tr
X ( Z2 + t (ZW + WZ) Z2 + (ZW + 1
WZ))dt.
3 . 11 )
1
The trace under the integral sign consists of four terms which, using the cyclicity of the trace, can be rewritten as
Tr WZ
I
Z2 + t
WZ Zz I +
I_
1
WZ2 t) + 3Trf1\ W_ Z2 + t Z2 + t).
Since only Z2 enters the second of these two terms, this term is unchanged when Z is replaced by IZI. Upon writing out the first term in a basis that diagonalizes Z, that term becomes 2I
+t
, ..1 + t
Iw;ll2z;z;.
Clearly this term, and hence the integral in (3.11), increases when Z is replaced by
IZI. The first term in (3.11), being a function of Z2, is invariant under the substitution, and the assertion (3.10) is established.
185
With K. Ball and E. Carlen in Invent. Math. 115, 463-482 (1994) K. Ball et al.
478
Therefore, without loss of generality we may assume that Z > 0. Then, of course, Z + rW > 0 for all r sufficiently small, and we no longer need to square Z + r W to obtain a positive operator whose powers can be expressed as an integral over its resolvent. Working directly with Z + rW, we can use the simpler integral representation (Z+rW)'P-n=Yrf t(P-1)
t+(Z+r
Ctt
o
W)dt
(3.12)
1
to conclude that
4,"(0)=PYvtly "TrI o
t+ZWt+ZWldt.
J
L
(3.13)
Consider the right side of (3.13) as a function of Z for fixed W. We claim that it is convex in Z. To prove this, it suffices to prove the following inequality for every self-adjoint matrix A: d(A). _
d2
1 ds2Tr
There
are
six
terms.
1
t+(Z+sA)Wt+(Z+sA)W If
we
define
C = (t + Z)- 12A(t + Z)"!2
and
D = (t + Z)- 1/2 W(t + Z)-''2, then the result of the computation is A(A) = 4Tr C2D2 + 2Tr CDCD.
But by the Schwarz inequality, ITr(CDCD)I < {Tr(CD2C)}112 {Tr(DC2D)}1'2 = TrC2D2.
Thus, A(A) > 0 and the integrand in (3.13) is a convex function of Z. Now fix W and t, and define
F(Z) = Tr
[t+ZWt+ZW]-
Clearly, when U is any unitary matrix that commutes with W, F(UZU*) = F(Z). be Let {e1 , ... , e2.1 be an orthonormal basis of eigenvectors of W. Let some enumeration of the 22n unitary matrices with the property that U;ek= ± ek for each k. Clearly each of these unitaries commutes with W. Thus, by the convexity
of F which we have established in the last paragraph, 22-
F(Z) = 2-2"
22
F(U;ZUj) >_ .1=1
UjZU.1= F(Zdhg)
F(2-2"
l
J=1
where Zdi,g is the matrix whose diagonal entries, in the basis specified above, are those of Z, and whose off-diagonal entries are all zero. Replacing Z by Zd;,g in (3.13), the integration can be carried out, and we obtain 1 V "(0)>=P(P-
1 1I
where zj and w;, respectively, denote the jth diagonal entries of Z and W in the W-basis specified above.
186
Sharp Uniform Convexity and Smoothness Inequalities for Trace Norms Sharp uniform convexity
479
Now consider i/,(0 = Tr(Zv) as a function of Z. It is clearly convex. Thus,by the averaging method just employed, we obtain z,
O(0) ?
Y. z;
I
To establish (3.9), it only remains to check that /Y' , z' \(z v)rv/, '
Zw-z)wj) >_
( ,Y` Iwirvl
l
inequality.
but this follows immediately from Holder's
z
0
IV Hanner's inequality for matrices This section is devoted to the proof of parts (b) and (c) of Theorem 2. We begin with the proof of Theorem 2(c), and then show how that implies Theorem 2(b).
Proof of Theorem 2(c) First, let Y be a fixed self-adjoint n x n matrix, and consider the set Mr of n x n self-adjoint matrices given by
Mr:= {X: X+ Y> O and X- Y>01. Clearly Mr is convex, and if X e Mr, then X > 0. We claim that G(X):= 11X + YIIp+ IIX - YIIp-211XIIp
(4.1)
is a convex function on Mr. By the averaging method employed in the proof of Theorem 1, this convexity would imply that
IIX+ Y{ip,+ IIX- YIIp,-aIIXllp,? IIXdi.,+ YIIp,+ IIXdi,1- YIIp-aliXdi,`IIp, (4.2)
for any 0 a 5 2, where Xdi,a denotes the diagonal part of X in a basis diagonalizing Y. (Note that if X e Mr, then Xdi,a a My.) By Lemma 4 and Hanner's inequality in !v.
IIXdi,1+ YIIp,+ IIXdi.Q- Yllp- a(r)lIXdi.QIIp,>= a(l/r)II YIIp,
for all r, where a(r) is the function defined in Lemma 4. (Here we are making use of the easily checked fact that for I < p < 2, a(r) and a(l/r) never exceed 2.) Combining this with (4.2), we would obtain IIX+ YIIPp + IIX - Ylip,za(r)11X11p+a(1/r)II YIIp.
Then, by another application of Lemma 4, the inequality (1.8) would be established
for 15 p 5 2 for all matrices X and Y such that X + Y and X - Y are positive semidefinite. By Lemma 6, and the remark that follows it, (1.8) would be established
for 2 <- p < oo and all positive semidefinite matrices X and Y. It remains to establish the convexity of G(X). We choose a self-adjoint matrix A and define
4(s)= II(X+sA)+ YIIp+11(X+sA)-
YIIp-21'(X+sA)11P.
(4.3)
187
With K. Ball and E. Carlen in Invent. Math. 115, 463-482 (1994) 480
K. Ball et al.
Then ds O
(s) = pT r [ ((X + sA) + Y)°- I + ((X + sA) - Y)°-
2(X + sA )o- t ] A.
Using the integral representation formula to compute the next derivative, we have
"(O)= py,o 0
Tr t+X+YAt+X+Y+t+X-YAt+X-Y -2I+XAt+X]A)dt.
This is positive by the convexity of I
t+XAt+X A, I
which we established in the last section.
El
Proof of Theorem 2(b) We use a power doubling argument inspired by the 2convexification method developed by Figiel and Johnson [FJ]. First, consider the case p > 4. As in the proof of Theorem 1, we can assume that X and Y are self-adjoint n x n matrices. The spectrum of the 2n x 2n matrix
XY Y
X)
consists of the union of the spectra of X + Y and of X - Y. Thus, the pth power of its CD norm equals the left side of (1.8). By the same spectral considerations, one sees
that the pth power of the C, norm of the 2 x 2 matrix II X II
Ii Y!i
II YII
II
X11l
(4.5)
equals the right side of (1.8). Thus, our problem is to show that the C, norm of the 2 x 2 matrix in (4.5) exceeds the C, norm of the 2n x 2n matrix in (4.4).
Now
X2+Y2 XY+YX
X Y)
( X Y + YX X2+ Y2) pre ( Y X 12p The second matrix is positive semidefinite, and it has the special block form (B 'A). Block matrices of this form are characterized by the fact that they commute with
(° o), where I is the n x n identity matrix. Evidently, all powers of a positive semidefinite block matrix of this special form have the same special form. Thus, if r is the index conjugate to p/2, there is a positive matrix (o °) whose C,-norm is I with the property that the norm in (4.6) is realized as Tr(X2 + YZ X Y + YX
XY+ YX X2 + YZ
)(C D)
DC
= 2Tr(X 2 + Y2)C + 2Tr(X Y + YX )D < 21.C11,(11X211#2 + 11 Y2,1a2) + 4I1DII.IIX!!,1! Y!,
by the Holder inequality for traces of matrices.
188
Sharp Uniform Convexity and Smoothness Inequalities for Trace Norms
Sharp uniform convexity
481
Let us define IICII:= IICII..11D11:= IIDIIr.IIX11:= I1XII,and II Y11:= 11
last expression is 211C1](IIX112 + II Y112)+411D11 IIXII II YII
IICII
= Tr(IIDII
IIDIII
11X112+IIY112
2 11XII II Y!1
21IXII II YII
11XII2 + II Y112
IICII)(
=11(IIDII IIDII)ILII(III
IIIXIII)I,
YII
The positivity of the matrix (p D) guarantees that both C + D and C - D are positive. Since 1 <_ r 5 2, Theorem 2(c) implies that IICII IIDII
=1.
(IIDII IICII) Consequently,
\Y X X)II,
IIXII II YII )'l! IIXI II
2
P
as required. Finally, by Lemma 6, we obtain the validity of (1.8) for I
p
Remark. For all p, (1.8) holds (with the appropriate direction of inequality) in C. when II Y11 = 11 X II since (1.8) is then a special case of the "easy" Clarkson inequality
(2.1) which was extended to C,, by Dixmier [Di]. It also holds to leading order for small Y, as one can verify using Theorem 1. We make the natural conjecture that (1.8) holds in C,, for I <- p < 2. without the restrictions imposed in part (c) of Theorem 2.
References
[ArYa]
[BP] [Bo] [C] [CL]
[D]
Araki, H., Yamagami, S.: An inequality for the Hilbert-Schmidt norm. Commun. Math. Phys. 81, 89-96 (1981) Ball, K., Pisicr, G.: Unpublished result; private communication. Boas, R.P.: Some uniformly convex spaces. Bull. Am. Math. Soc. 46, 304-311 (1940) Clarkson, J.A.: Uniformly convex spaces. Trans. Am. Math. Soc. 40, 396-414 (1936)
Carlen. E., Lieb, E.: Optimal hypercontractivity for fermi fields and related noncommutative integration inequalities. Commun. Math. Phys. "155, 27-46 (1993); for a slightly different presentation, see: Optimal two-uniform convexity and fermion hypercontractivity. In: Araki, H., Ito. K.R., Kishimoto, A., Ojima. I. (eds,) Quantum and non-commutative analysis. London New York. Kluwer (in press) Day, M.: Uniform convexity in factor and conjugate spaces. Ann. Math. 45, 375-385 (1944)
[Di]
[F]
Dixmier, J.: Formes lineaires sur un anneau d'operateurs. Bull. Soc. Math. Fr. 81, 222-245(1953) Figiel. T.: On the moduli of convexity and smoothness. Studia Math. 56. 121-155 (1976)
[FJ] [Gr] [H]
Figiel, T., Johnson, S.B.: A uniformly convex Banach space which contains no Cr. Compos. Math. 29. 179-190 (1974) Gross, L.: Logarithmic Sobolev inequalities. Am. J. Math. 97. 1061-1083 (1975) Hanner. 0.: On the uniform convexity of I P and )P. Ark. Math. 3. 239-244 (1956)
189
With K. Ball and E. Carlen in Invent. Math. 115,463-482 (1994) K. Ball et al.
482
[Ko]
Kothe, G.: Topologische lineare Raume, Die Grundlehren der mathematischen Wissen schaften in Einzeldarstellungen, Bd. 107, Springer Berlin Heidelberg New York: 1960
[L]
[P] [Ru] [Sc] [Si]
[TJ]
Lindenstrauss, J.: On the modulus of smoothness and divergent series in Banach spaces. Mich. Math. J. 10, 241-252 (1963) Pisier, G.: The volume of convex bodies and Banach space geometry. Cambridge: Cambridge University Press, 1989 Ruskai, M.B.: Inequalities for traces on Von Neumann algebras. Commun. Math. Phys. 26, 280-289 (1972) Segal, I.E.: A non-commutative extension of abstract integration. Ann. Math. 57, 401-457 (1953) Simon, B.: Trace ideals and their applications. (See p. 22) Cambridge: Cambridge University Press, 1979
Tomczak-Jacgermann, N.: The moduli of smoothness and convexity and Rademacher averages of trace classes Sp(l (1974)
190
p < co). Studia Math. 50. 163-182
With E. Carlen in in Advances in Math. Sciences, Amer. Math. Soc. Transl. (2), 189, 59-62 (1999)
Amer. Math. Soc. Tranel. (2) Vol. 189, 1999
A Minkowski Type Trace Inequality and Strong Subadditivity of Quantum Entropy Eric A. Carlen and Elliott H. Lieb Dedicated to M. Sh. Bir,nan on his seventieth birthday ABSTRACT. We consider the following trace function on n-tuples of positive operators:
lUPl
((n p(A1,A2, -, A,,)=Tr\\LA)/ 3=1
and prove that it is jointly concave for 0 < p < 1 and convex for p = 2. We then derive from this a Minkowski type inequality for operators on a tensor product of three Hilbert spaces, and show how this implies the strong subadditivity of
quantum mechanical entropy. For p > 2, p is neither convex nor concave. We conjecture that +p is convex for 1 < p < 2, but our methods do not show this.
I. Introduction Let Px denote the set of all positive semidefinite operators on a finite dimenThen, for any finite natural number sional Hilbert space 7{ with inner product n, any finite p > 0, and any finite n-tuple (A1, A2, ... , An) of elements of PR, define (1.1)
4)p(A1,A2,...,An)=Tr\EAP 1
I.
1=1
The main result of this paper is the following is a jointly concave function of its arguments. THEOREM 1. For 0 < p < 1, For p = 2, 'pp is jointly convex. For p > 2, 41p is neither convex nor concave.
We conjecture that 4p is jointly convex for 1 < p < 2. We state all of the theorems in a finite dimensional context, and some of our methods of proof explicitly 1991 Mathematics Subject Classification. Primary 47A63, 15A90. Copyright 1997 in image and content by the authors. Reproduction of this article in its entirety by any means is permitted. The work of the first-named author was supperted by U. S. National Science Foundation grant no. DMS 9500840. The work of the second-named author was supported by U. S. National Science Foundation grant no. PHY95-19433 A01. 59
191
With E. Carlen in in Advances in Math. Sciences, Amer. Math. Soc. Transl. (2), 189, 59-62 (1999) ERIC A. CARLF.N AND ELLIOTT H. LIEI3
60
involve this finite dimension. Nonetheless, the results themselves do not depend on the dimension, and therefore easily extend to the appropriate trace classes on an infinite dimensional Hilbert space. We note that the trace in Theorem 1 is essential; the asserted trace inequalities do not hold as operators inequalities. If they did, we would have, for example at
p = 2 that (A2 + B2)1/2 < A + B. This is of course not true in general positive operators, as is well known and easily checked.
We shall use Theorem 1 to derive a Minkowski type inequality for traces of operators on a product of three Hilbert spaces. To set this in perspective, recall that the Minkowski inequality says that for non-negative measurable functions f on the Cartesian product of two measure spaces (X, µ) and (Y, v), I/p
p (1.2)
(L(Jf(x,Y)dv)
fp(x,y)dµlI/pdv
dl/
/
L(L
for p > 1, and that the opposite inequality(holds for 0 < p < 1. A direct analog of (1.2) holds for positive operators A on the tensor product of two Hilbert spaces %1 ® 9't2. To state it, let Tt, A denote the positive operator on 9{2 that is given as a quadratic form by (v,Tr, Av) = 1: (u, (& v,A(u, ®v)),
where v E it2 and the ul constitute an orthonormal basis of it1. As is well known, the quadratic form on the left is independent of the choice of the orthonormal basis on the right. The operator Tr1 A so defined is called the partial trace of A over it,. It will be convenient, and generally clearer, in what follows to write Tr1 also to denote the usual trace on 11, for operators A on %I alone. The following is the tracial analog of (1.2). THEOREM 2. Let A be a positive operator on the tensor product of two Hilbert
spaces it, ® t2. Then for all p > 1, (Tr2(Tr, A)") 1/P < Tr1 ((Tr2 A") I/p)
(1.3)
and inequality (1.3) reverses for 0 < p < 1.
Returning to (1.2), note that it has a trivial extension to functions of three (or more) variables. Though trivial, it has an interesting consequence. If one considers a non-negative measurable function f (x, y, z) on the Cartesian product of three measure spaces (X, µ), (Y, v) and (Z, p), and simply holds z fixed as a parameter, one gets 1/p (1.4)
(1, U
f (x, y, z)dv pdµ )
5
J
(f
f p(x, y, z)dµ f
"Pd,
pointwise in z for p > 1. Integrating in z then yields
z)dv)pdµ J I /pdP 1z CJx \1Y f (x, y, for p > 1, and of course the inequality reverses for 0 < p < 1. (1.5)
fz I (I
fp(x, y,
z)dµl
I ipdvdp
J
Now, since (1.5) is an equality at p = 1, we get another inequality by differentiating (1.5) with respect to p at p = 1. This yields an entropy inequality. In
192
A Minkowski Type Trace Inequality and Strong Subadditivity of Quantum Entropy STRONG SUBADDITIVITY OF QUANTUM ENTROPY
61
fact, using the homogeneity of (1.5), we can normalize f so that it is a probability density. Recall that for any probability density p on any measure space (X, 14), the entropy S(p) is defined as
r
S(p)=-J pIn pdµ.
(1.6)
x We denote various marginal densities of f as follows:
f2.3(y,z) = f f(x,y,z)dti,
f3(z) =XfYf f
f1.3(x,z) = f f(x,y,z)dv, Y
(x, y,
z)dpdv.
Then the derivative of (1.5) at p = 1 is S(fl,3) + S(f2.3) > S(f1,2.3) + S(f3),
(1.7)
which is the strong subadditivity of the classical entropy; see (L75]. Now consider operators on the product of three Hilbert spaces, and a density matrix A; i.e., a positive operator on W, ®7-[2 ® H3 with Tr A = 1. The entropy S(A) of a density matrix A is defined by
S(A) = - T (A In A).
(1.8)
The operator analog of (1.7) is the Lieb-Ruskai LLRI strong subadditivity inequality for the quantum mechanical entropy: (1.9)
S(A1.3) + S(A2,3) ? S(A1,2,3) + S(A3),
where, in analogy with our notational conventions for marginal densities, we define A1.2,3 = A,
A2,3 = T 1 A,
A3 = T 1 Tr2 A
and so forth.
Thus, the differential form of Minkowski type inequality (1.7) is known to hold
at p = 1 for operators. It is therefore natural to enquire whether there exists an operator analog of the three-variable Minkowski inequality (1.7) for other values of p. Unfortunately, the methods at our disposal suffice to establish this only for
0
THEOREM 3. Let A be a positive operator on the tensor product of three Hilbert
spaces ?ll ®f2 ®113. Then (1.10)
T
3(` r2(T 1 A)P) 11" < Tr1,3((T
2 AP) "P)
for p = 2 and, trivially, p = 1, while the reverse inequality holds for 0 < p < 1. This is, nonetheless, enough to imply the strong subadditivity (1.9): one simply takes the left derivative at p = 1. It is readily seen by considering block-diagonal matrices that the inequality of Theorem 3 implies the convexity of 4P for p = 2, and the concavity of 4bp for 0 < p < 1. By the same token, (1.10) cannot hold in general for p > 2 since this would imply the convexity of for such p, and Theorem 1 precludes this. This is
193
With E. Carlen in in Advances in Math. Sciences, Amer. Math. Soc. Transl. (2), 189, 59-62 (1999)
ERIC A. CARLEN AND ELLIOTT H. LIEB
62
in contrast to Theorem 2, the Minkowski inequality for two spaces, which holds for
allp> 1. The fact that there is such an easy passage from the Minkowski inequality in two variables to that in three variables may leave one surprised that there should be any difficulty in making the same passage with operators. But difficulty there is. In fact, even the simple version in Theorem 2 seems to require a more intricate proof than does the corresponding statement for integrals-which after all is simply the statement that the unit ball in LP is convex for p > 1. In fact, we know of no previous proof of Theorem 2. We emphasize that there is no operator analog of the pointwise inequality (1.4). That is, if we omit Tr3 on both sides of (1.10), the result will be two operators on 713i and these two operators do not satisfy the corresponding operator inequality. We present a proof of Theorem 2 in Section II. Then in Section III we prove Theorem 1. In Section IV, we recast Theorem 1 into an equivalent form, from which Theorem 3 is readily derived in Section W. Section V contains a brief comment on a relation between the conjectured convexity for 1 < p < 2 and a very interesting trace inequality of Birman, Koplienko and Solomyak (BKS].
H. Proof of Theorem 2 The following proof of Theorem 2 is given for matrices, but is easily extended to operators as the statement is dimension independent.
Let A be a positive operator on PH,®x the tensor product of two finite dimensional Hilbert spaces. Suppose first that p > 1. We proceed by duality.
There is a positive operator B in PH, with (Tr2(Ba)I/9) = I with 1/q + I/p = 1 such that A) = Tt1,2((I ® B)A)
(Tr2(Tr1
F(u, ®v,, (10 B)A(u, (9 v, )) _ J>, ®Bvj, A(u, (& v,)) for any pair of orthonormal bases {u,} and {vj}. We now choose the {v,} to be a basis of eigenvectors of B, and let {A, } be the corresponding eigenvalues. Then the right hand side above becomes ®v.,A(u,®vj))1(u,8+vj,A(u,®v,))/p) I/p
(EA)
EA,(u,
11/p
Next, by the spectral theorem, for each i and j, (uj ®vj, A(u, (9 vj)) < ((u, ®vj, A'(u, ®v,)))
Using this, one arrives at (Trz(Tri A) P'
1/p <
(((uj ®v,, A'(u, ®vj)) ll ((u,,Tr,Apu,))1/p
194
I/p
A Minkowski Type Trace Inequality and Strong Subadditivity of Quantum Entropy STRONG SUBADDITIVITY OF QUANTUM ENTROPY
63
Now we choose the {u;} to be a basis of eigenvectors of Tr2 AP. Then
E((u:,`
r2APu1))1/P=E(u, (T2AP)1/Pui)=TrI(Tr2AP)1/P
and the desired inequality is proved for p > 1. Note that this part of the proof works for all p > 1, not only 1 < p < 2. Now suppose 0 < p < 1, and define r = 1/p and B = AP so that A = Br. Since
> (Tr2(Tr IB)r)I/r. r > 1, the inequality proved above says Tr Rewriting this in terms of A and p, and switching the roles of f1 and 1(2, one 1((Tr2Br)I/r)
O
obtains the desired result for 0 < p < 1.
III. Proof of Theorem 1 As before we give the proof for matrices. Consider first the case 0 < p < 1. The proof in this case proceeds by reduction to a theorem of Epstein [E) concerning the function
A , Tr((BAPB)'/P) on P,, where B is any given element of P1(. Epstein's theorem says that this function is concave for 0 < p < 1. To apply this, consider first the case n = 2 in (1.1), and define
A=[A10
0]
A2
and 1
o= [01 0] Then
AP+aAPa= I
Al
l
But
+A2 0 ] 0 AI + A2
\
\
/
AP+aAPa=2(12a)AP(I 2a I+2I Now define
1I±=
\ )AP(I2a). / 126
--I
and observe that these are complementary orthogonal projections. Thus, (3.1)
2Tr((A + A2 )'/P) = 21/PTr((n+APn+)'/P) + 2'/PTr((lI-APf-)1/P).
Epstein's theorem, with A = A and B = lI±, now implies that each term on the right hand side of (3.1) is a concave function of A, which means that the left hand side is a jointly concave function of AI and A2. This concludes the proof for n = 2. One now easily iterates this procedure to obtain the result for all dyadic powers n = 2'`, and hence for all n. To prove the convexity of 4>2 there are several ways to proceed, but the simplest was pointed out to us by S. Sahi. Namely, let n be given and consider the block
195
With E. Carlen in in Advances in Math. Sciences, Amer. Math. Soc. Transl. (2), 189, 59-62 (1999)
ERIC A. CARLEN AND ELLIOTT H. LIED
64
matrix A given by
Al
0
...
0
0
...
0
0
...
0
Then A2,... , An) = Tr JAI,
where IXJ is the usual operator absolute value; i.e.,
In other words
A2, ... , An) is simply the trace norm of A, is therefore clearly jointly convex
in A1,A2,...,An. Finally, we show that convexity fails to hold for p > 2. To see this, choose any pair A1,A2 E Ph, and any vector v such that (3.2)
(v, ((AP + AZ)/2)v) < (v, ((A1 + A2)/2)Pv)
Note the strict inequality here. It is always possible to find such A1,A2 and v for p > 2 since, for such p, X '-. XP is not operator convex. Now let H,, denote the orthogonal projection onto the span of v, and let [Il, _ I - H. denote its orthogonal complement. Then, for a large number A to be fixed below, put
B=IIv+AIll,. Then, if 4bP were convex, we would have
1/2 \
A2,B f - I VtA1, B) - IbP(tA2,B) I < 0 .
(3.3)
2
However, for small t \> 0,
tP
11(t
A1 + 2
A2 B) = TY(tP(A1+A2)"+BP)" /J
\
2/
`\
=TrB+_TYI
P
\
B-P(A1
\
2
A2) )
/ I +O(t2p)
and
24P(tAl, B) + Z4p(tA2, B) = Tr B + p G Tr
B'-PAP +
Tr B1-PAP + O(t2p).
/
2
Thus, limseuppi_'
(
pI tAl 2 A2, B I - 24'(tA1, B) - 2lb(tA2i B))
=Tr(B'-P(A12A2I (Ai +A2)PV )/2
/
i-I!TrBI-PA°+2TrB1-PAP)
((v'Aly) + (v'A2v) 2 2 )
+O(Al-P).
Now taking A sufficiently large, this last term on the right is strictly positive by (3.2). This contradicts (3.3), and thus convexity does not hold not even separately. El
196
A Minkowski Type Trace Inequality and Strong Subadditivity of Quantum Entropy STRONG SUBADDITIVITY OF QUANTUM ENTROPY
65
IV. Corollary of Theorem 1 and proof of Theorem 3 A corollary of Theorem I is obtained by writing the partial trace as an average, and exploiting the convexity and concavity established above. Let A be a positive operator on ?-(l 0 l2. Next, suppose the dimension of ?{2 is N, and fix some orthonormal basis { e 1 , r.2, ... , eN }. With respect to this basis, define the self-adjoint unitary operators U,., and V,,, on 7.12 by
U, j = I - E;., - Ej, + E;, + EJ
V;=I-2E,,,, where the i and j are a distinct pair of indices, and E,,, in this basis has the matrix with I in the i,jth place, and 0 elsewhere. Let G be the subgroup of the group of unitary operators on 7i2 that is generated by this family together with the identity. Each operator W in this group acts by Wee =
and some map s : 1, 2, ... , N .--+ 0, 1. Thus, the size of the group is 2^'N!, and the point about it is that any operator on 712 that for some permutation
commutes with every element of this group is necessarily a multiple of the identity on 7{2. Then 2NN!
F (10W')A(I So H7) = N Tr2(A) 0 1,,,. W'
This way of writing partial traces can be traced back to Uhlmann JUJ. From here one easily arrives at the following result:
THEOREM 4. For p > 0, let the map WJ,(A) from positive operators A on f1 x f2 to R+ be given by 'Y/,(A) = Trl (('112 A") I/p)
(4.2)
.
Then this map is concave for 0 < p < 1, convex for p = 2, and neither for p > 2. PROOF. We shall assume that the dimension of 7.12 is N so that we may apply the averaging formula introduced above. We then have
// = N'IP- 1
\ = N'/p-1 (
1
2'v N!
l l 21 1: (I
)
1
w'E9
W')Ap(I0W))l lip) )
TY1.21 1\ F ((10W')A(I0 97))")
\
The result now follows directly from Theorem 1.
/ p
197
With E. Carlen in in Advances in Math. Sciences, Amer. Math. Soc. Transl. (2), 189, 59-62 (1999) ERIC A. CARLEN AND ELLIOTT H. LIEB
66
Notice that the conclusion of Theorem 4 not only follows from Theorem 1, but also implies it. To see this, suppose that the A in Theorem 4 is block diagonal with Al 0
A2
... ...
0
0
... A
0
0 0
A
Then clearly T'p(A) =
1/p
p
1
1/p
= x'1,31 Tr2 I N Tr1 A ® IN,/ 1 )
= 4rp (N Try A ®IN,)
where the pair of spaces in the definition of Pp is taken to be 113 and HI ®9.13. Then by (4.1) and the convexity of lip established in Theorem 4,
*pI TrIA®1x,/ =41pl\2 N
>(1®W')A(I®W)l
N
IN N.
/
I
WEQ
q'p((10W')A(1®W) WEB
The last term above is 1/v I
N 2N
I
E T1,3 (Tr2(1 ®W)AP(1® W)))
we(
1/p
Y- Tr1,3 ((1® W') (Tr2 AP) (1(9 W) 1 2N N.I WEG Tr1.3((I®W')(Tr2A')I/p(1®W))
= 2NNf E
=T13T1I((T<'2A")1/p),
WEG
which is the desired result.
0
V. The BKS inequality and the 1 < p < 2 conjecture Birman, Koplienko and Solomyak JBKS] proved that for p > 1, and A and B positive semidefinite operators, (5.1)
Tr(BP - A")+/p > Tt(B - A)+,
where X, denotes the positive part of a selfadjoint operator X; i.e., X+ _ (X + JXJ)/2. In (5.1), neither B nor A needs to be bounded, but it is assumed that (BP - AP)1 p is trace class. Though the inequality in (5.1) is only one of
198
A Minkowski Type Trace Inequality and Strong Subadditivity of Quantum Entropy STRONG SUBADDITIVITY OF QUANTUM ENTROPY
67
several very interesting inqualities proved in [BKS], we refer to it here as the BKS inequality.
The proof is in two parts, the first of which is to reduce consideration to the case BP > AP in which case one has B = (AP + CP)'l" with C > 0. Then (5.1) becomes
Tr(C + A) > Tr(AP + CP)I/P
(5.2)
for all A > 0 and C > 0. It is (5.2) that interests us here. Clearly (5.2) can be rewritten as 4YA, C) < $p(A, 0) + 4$p(O, C),
(5.3)
which is a subadditivity property of $p for all p > 1. Since Op is homogeneous of degree 1, subadditivity and convexity are the same thing. Thus for p = 2, (5.3) is a special case of the convexity of proved in Theorem 1, and for 1 < p < 2, it would be a consequence of the conjectured convexity for these p. However, the BKS inequality holds for all p > 1, not only for 1 < p < 2. There is a simple proof of (5.2) for matrices. Let Al
-
AP/2
0 0
fCP/2
so that
Tr(MfM})Ilp = Tr(AP +C")I/v On the other hand, the spectrum of M; M± is the same as the spectrum of M± M;, so
Tr(AP+CP)I/P= It(llftllf})1/v One computes
MfM _
t
AP
tJl
tJ CPI
with J = AP12Cn12. Since X - Tr(XI/p) is concave for p > 1, one has that
Tr(A+C)=Tt
M+M++M_M_ Up 2
1
>Tr(AP+CP) 1/v
A recent application of the BKS inequality, and a different proof of (5.1) that holds in the case of unbounded operators, can be found in [LSS].
Acknowledgements. We thank T. Ando and F. Hiai for a careful reading of this paper, and for pointing out many misprints in an earlier draft.
References T. Ando and F. Hiai, Holder type inequalities for matrices, Preprint, 1997. (BKSJ M. S. Birman, L. S. Koplienko, and M. Z. Solomyak, Estimates for the spectrum of the difference between fractional powers of two selfadjoint operators, J. Soviet Math. 19 (1975), no. 3, 1-6. JEJ H. Epstein, On a concavity theorem of Lieb, Commun. Math. Phys. 31 (1973), 317-327. (L75( E. H. Lieb, Some convexity and subadditivtty properties of entropy, Bull. Amer. Math. Soc. 81 (1975), 1-13. JLRj E. H. Lieb, and M. B. Ruskai, Proof of the strong subadditwity of quantum-mechanical entropy, J. Math. Phys. 14 (1973), 1938-1941. (AHI
199
With E. Carlen in in Advances in Math. Sciences, Amer. Math. Soc. Transl. (2), 189, 59-62 (1999) ERIC A. CAREEN AND ELLIUTT H. LIES
68 [LSS)
E. H. Lieb, H. Siedentop, and J. P. Solovej, Stability and instability of relativistic electrons in magnetic fields, J. Stat. Phys. 89 (1997), 37-59.
IU)
A. Uhlmann, Satze nber Diehtematrizen, Wiss. Z. Karl-Marx Univ. Leipzig 20 (1971), 633-653. SCHOOL OF MATHEMATICS, GEORGIA INSTITUTE OF TECHNOLOGY, ATLANTA, GEORGIA 30332 DEPARTMENTS OF MATHEMATICS AND PHYSICS, PRINCETON UNIVERSITY, PRINCETON. NEW
JERSEY 08544-0708
200
Part III
Inequalities Related to the Stability of Matter
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
INEQUALITIES FOR THE MOMENTS OF THE EIGENVALUES OF THE SCHRODINGER HAMILTONIAN AND THEIR RELATION TO SOBOLEV INEQUALITIES
Elliott H. Lieb* Walter E. Thirring 1.
Introduction
Estimates for the number of bound states and their energies, ej < 0, are of obvious importance for the investigation of quantum mechanical Hamiltonians. If the latter are of the single particle form H = -A + V(x)
in R", we shall use available methods to derive the bounds
fdnxlvx+n/2.
lejly< Ly,n
y> max (0, 1-n/2) . (1.1)
Here, IV(x)l_=-V(x) if V(x) < 0 and is zero otherwise. Of course, in many-body theory, one is more interested in Hamiltonians
of the form - E A i + F v(xi - xj). It turns out, however, that the i
i>1
energy bounds for the single particle Hamiltonian yield a lower bound for
the kinetic energy, T, of N fermions in terms of integrals over the single particle density defined by p(x) = N
fI(x,x2,.,xN)I2dnx2 ... d"xN
,
(1.2)
where cli is an antisymmetric, normalized function of the N variables
xi ( R. Our main results, in addition to (1.1), will be of the form
Work supported by U. S. National Science Foundation Grant MPS 71-03375-A03. 269
203
With W. Thitring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
E. H. LIEB AND W. E. THIRRING
270
rN f y IVi0(x1 ... xN)I2 dnxl ... dnxN
T
i=1
2(p-1)/n
> Kn [fdnxPxP/P_1]
(1.3)
when max I n/2, 11 < p< 1+ n/2. For N = 1, p = n/2, (1.3) reduces to the well-known Sobolev inequalities. (1.3) is therefore a partial generalization of these inequalities, and we shall expand on this in Section 3. Our constants Kp n are not always the best possible ones, but nevertheless, they may be useful for many purposes. In particular, in ref. [1], a special case of (1.3) was used to give a simple proof of the stability of matter, with a constant of the right order of magnitude. The result for q species of fermions (2m = e = >S = 1) moving in the field of M nuclei with
positive charges Zl is M
2
H > - 1.31 q2/3 N 1 + M Z7 /3/N 1/2
(1.4)
j=1
In particular, if q = 2 (spin 1/2 electrons), we have a bound N, and if we set q = N, we get a bound ^- NS/3 if no symmetry requirement is imposed on the wave function; a fortiori this is a bound for bosons. Our bound implies stability of matter in its intuitive meaning such that the N (Bohr radius)3. To give a volume occupied by N particles will be formal demonstration of this fact, one might use a method which gives lower bounds for the radii of complex atoms (compare Equation (3.6, 38)
of ref. [201). As a first observation, one calculates the ground state energy of N electrons (with spin) in a harmonic potential. Filling the oscillator levels, one finds N 2
(
i=1
204
+
(02 z2) > w N4/3 34 3 (1 +
O(N-t/3))
.
(1.5)
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON
271
Next, take the expectation value of this operator inequality with the ground
state of H, set
344/3 <-n>
W
(1.6)
Na/3
and use the virial theorem
<-Ai>
1/2 2
M
-E0 < 2.08N 1+
L j=1
I
(1.7)
Altogether we find N
x>>
<
i=1
>
(3N)s/3
16<-A.>
16.2.08
<X? > 1 / 2 > C N1 /3
r
38/3 N5/3
(1.$)
Z?/3/Nl1/2
+
c
.75
(1.9)
(2z1'3/N) Cl +
J
Therefore, if the system is not compressed by other forces, so that the virial theorem is valid, it will not collapse, but will adjust its volume to a size proportional to the number of particles. Regarding the Z-dependence, we see that with Z = Zj = N/M we have (for large Z)
<x?>1/2 _
M1/32-1/3
That is, the mean atomic radius is predicted to be > Z-1 /3. A better result can hardly be expected since for M = 1, this is the correct
Z-dependence for large Z.
Although we have no results on the best possible constants, Kp n, except in a few special cases, experience drawn from computer calculations suggests that there is a critical value yc n above which the classical value gives a bound:
205
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
272
E. H. LIEB AND W. E. THIRRING
(I Jejly)ctassica, = (2.)-n
LY
fdnp dnx lp2 + V(x)IY
n f I V(x)Iy+n/2 dnx (1.10)
Y ? Yc,n
and where LY n, given by the above integral, is LY,n
2-n,,-n/2 r(y+ 1)/I'(y + I+ n/2) .
=
(1.11)
We conjecture yc,l = 3/2, Yc,3 a., .863 and yc n = 0, all n > 8. If this conjecture were to be true, the constants in (1.3, 1.4) could be further improved.
In the next section we shall deduce bounds for F, lejlY and use them i
in Section 3 to derive (1.3). In Section 4 we shall discuss our conjectures and support them for n = 1 with results from the Korteweg-de Vries equation. Section 5 contains new results added in proof. In Appendix A, generously contributed by J. F. Barnes, further evidence from computer studies is presented. We are extremely grateful to Dr. Barnes for taking an interest in this problem, for without his results we would have been hesitant to put forth our conjectures. 2.
Bounds for Moments of the Eigenvalues
In this section we shall deduce bounds of the form (1.1), and we shall compare our Ly n with the classical values which one gets by replacing IejlY
by
(2n)-n
f
dnx dnp IP2 + V(x)ly
For n - 3 and y - 1, the latter are smaller by about an order of magnitude.
206
Inequalities for the Moments of the Eigenvalues of the Schr&dinger Hamiltonian
273
THE EIGENVALUES OF THE SCHRODINGER HAMILTON
Our inequalities are based on the Birman-Schwinger (2, 3) method for
estimating NE, the number of bound states of H = -A + V(x) having an energy < E. Since NE = S(E - ej)
V we have 00
Yi JejlY = y J
daay-I N_a
(2.1)
o
Now, according to Birman-Schwinger [2, 3], for all a > 0, m > I and t ( [0, 1],
N_a < Tr(JV+(1-t)all/2(--L\+ta)-IIV+(1-t)aj//2)m
.
(2.2)
REMARKS ABOUT (2.2):
1. We are only interested in potentials such that V_ LY+n/2(Rn) for y > min (0, 1- n/2). For such potentials (2.2) is justified, and a complete discussion is given in Simon [4, 5]. Moreover, it is sufficient to consider V c Co(Rn) in (2.2), and in the rest of this paper, and then to use a limiting argument. Such potentials have the advantage that they have only a finite number of bound states [5]. 2. Since we are interested in maximizing E IejJ>/ f lVjY+n/2, we may as well assume that V(x) < 0, i.e. V = - V I _ . This follows from the max-min principle [4] which asserts that ej(V) > ej(-l V1 _), all j, including multiplicity.
To evaluate the trace in (2.2), we use the inequality Tr (BI /2 A B1 /2)m < Tr Bm/2 Am Bm/2
(2.3)
when A, B are positive operators and m > 1. When m is integral and A, B is of our special form, (2.3) is a consequence of Holder's inequality. For completeness, we shall give a more general derivation of (2.3) in Appendix B.
207
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
E. H. LIEB AND W. E. THIRRING
274 To calculate
TrIV+(1-t)alm(-A+ta)-m,
(2.4)
we shall use an x-representation where (-fl+ta)-m is the kernel G(ma)(x-y) = (2,,)-n
J
dnp (P2 +
ta)-m
eip(x-Y)
(2.5)
if m > n/2. Using 2fln/2
dnP =
2)
.0
dp pn-I ,
(2.6)
we easily compute
Gta)(0)=(2n)-n 2nn/2 (ta)-m+n/2
r m
(2.7)
> n/2. Thus, N-a < (4rr)-n/2 r (m-n/2 (ta)-m+n/2 r dnx IV(x)+(1-t)alm. (2.8) r(m) J
Next, we substitute (2.8) into (2.1). If we impose the condition that t < 1, it is easy to prove that one can interchange the a and the x integration. Changing variables a -. (1-t)-I I V(x)L _fl, leads to
je)(y < x
y(4n)-n12 t-m+n/2(1_t)m-y-n/2 1'(y-m+n/2)r(m-n/2) m
Ry+1+n/2)
V(x)jy+n/2
f dnx (2.9)
provided n/2 < m < n/2 + y, m > 1 and 0 < t < 1. The optimal t is
t = (m-n/2)/y. If we put our results together, we obtain the following (see note added in proof, Section 5).
208
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON
275
THEOREM 1. Let V_ c LY+n/2(Rn), y > max (0, 1- n/2). Let H
= -A + V(x), and let e) < 0 be the negative energy bound states of H. Then jejjY < Ly
n
J IV(x)ly+n/2
(2.10)
where
Ly n < Ly n = min (4rr)-n/2yy+l M
n I y+2+1
F (m - n) F(y + ` / \ 2
- ml /
,
(2.11)
and where F(x) = I'(x) x-x, max 11, n/21 < m < n/2 + y. REMARKS:
1. When y = 0, 11 ejj 0 means the number of bound states, including
zero energy states. For n > 2, our Lo,n = '. In Section 4, we shall discuss the y = 0 case further. See also Section 5. 2.
In (2.11), LY
is the bound we have obtained using the Birman-
n
Schwinger principle. We shall henceforth reserve the symbol LY the quantity Ly = sup t` le)jy/ J lVly+n/2 . n
Vr
n for
(2.12)
Optimization with respect to m in (2.11) can be done either numerically or analytically in the region where Stirling's formula F(x)
a-x jv7x
(2.13)
can be applied. In [11, for n = 3, y = 1, we used the value 2 for m. A marginal improvement can be obtained with m = 1.9. If (2.13) were exact, the best m would be m = n(y + n/2)/(n + y)
.
(2.14)
Note that as y - w, in- is bounded by n. Using IS, together with (2.13), which is valid when yn(y+n)-t is large,
209
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
276
E. H. LIEB AND W. E. THIRRING
L
y,n
-
(4r7)1-n/2
yye Y r n/2 1/2 I'(y+n/2) LY + n J
(2.15)
Finally, we want to compare our bounds with their classical values, Ly,n. From the results of Martin [6] and Tamura [7], one has the following THEOREM 2. if V(x) < 0 and V t Co (Rn), then lim
Iej(AV)ly /
f
lAVly+n/2 = LY,n
(2.16)
,J
COROLLARY.
Ly,n > LCC
(2.17)
Our Ly,n satisfies (2.17), in particular in the asymptotic region (2.15), we find y
os
[4nn(y+n/2)]1/2y-1/2
.
(2.18)
We conjecture in Section 4 that for y sufficiently large, the best possible Lynn should be Lynn, a result which does not follow from the Birman-Schwinger method employed here. For small y, we know that Lynn is not a bound. We conclude this section with a theorem about Ly,n which will be useful in the discussion of the one-dimensional case in Section 4. THEOREM 3. Let y > 1 + max (0, 1- n/2). Then Ly n < Ly_ 1 n [y/(y + n/2)]
(2.19)
.
PROOF. Choose e > 0. We can find a V c CD (Rn), with V < 0, such that
LY n(V) =
210
4 lej(V)IY/J IVIY+n/2 > Ly n -
E
.
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON
277
Let g e Co (Rn) be such that 0 < g(x) < 1, Vx, and V(x) a< 0 implies g(x) = 1. Let VA(x) = V(x) - Ag(x), A < 0. The functions Iej(VA)I are continuous and monotone increasing in A. Furthermore, there are a finite number of values -- < Al < A2 < < Ak < 0 with Aj being the value of A at which ej(VA) first appears. Al is finite because VA is nonnegative for A sufficiently negative. ej(VA) is continuously differentiable on A={AIO>A>A1,AAAi, and dej(VA)/dA = - J Ikbj(x: VA)I2 g(x)dnx by the Feynman-Hellman theorem. It is easy to prove that if f, ge Lp(Rn),
p > 1,
then
h(A) =
Jlf(x)_Ag(x)I!.dhix
is differentiable, VA and dh/dAIA=o = p
f If(x)I?-1 g(x)dnx
.
Thus LY,n(VA) is piecewise C' on A and its derivative, LY,n, is given by -1
LY,n=1
VAV+n/21 yy Iej(VA)IY-1fg(x) j(x:VA)I2dnx-(Y+n/2)LY n(VA) J
fvA(x)n/2_1 g(x)dnx By the stated properties of LY n, there exists a A e (A1, 0] such that (i) LY,n(VA) >_ 0; (ii) Ly n(VA) > LY
n-
2e.
Thus, using the properties of g,
0 < Y ej(V),)IY-1 - LY n(VA)(Y+n/2) J
IVAIY+n/2-1
(2.20)
Since a was arbitrary, (2.20) implies the theorem.
211
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
278
E. H. LIEB AND W. E. THIRRING
If we use (2.17) together with the fact that Ly,n = LY_l,n[y/(y+n/2)], we have
COROLLARY. If for some y > max (0, 1-n/2), Lynn = LC,n, then y
Ly+l,n = LC
i = 0, 1,2,3,-.-
Y+j,n,
REMARK.. By the same proof
Lynn < L,_l,n[y/(y+n/2)]
(2.21)
(see (3.1) for the definition of Lynn). 3.
Bounds for the Kinetic Energy In this section, we shall use Theorem 1 to derive inequalities of the
type (1.3). We recall the definition (2.14) and we further define Ly,n
sup Jelly/
I
IVI +n/2
(3.1)
Clearly, L1y,n
If
< Ly n
(3.2)
E HN,n q = the N-fold antisymmetric tensor product of L2(Rn;Cq),
with x] r Rn, we can write 0 pointwise as aj c 11, 2,.. , q4 and 0 -, -0/i if (xi, ai) is permuted with (xj, aj). q = 2 for spin 1/2 fermions. We can extend the definition (1.2) to
pa(x) = N
I ... ± I
a2=1
We also define
212
aN=1
10(x, x2,..., xN ; a. a2,..., aN)12 dnx2 ... dnxN . (3.3)
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON q
N
TAG =
279
I q
a1=1
(3.4)
f Iojo(X;o)12dnNX
j=1 a1=1
i
aN=1
f
Iq,(X,a)12dnNx
(3.5)
Our result is THEOREM 4. Let p satisfy max j n/2, 11 < p < 1 + n/2 and suppose
"
that Lp-n/2,n < If 110112 = 1, then, except for the case n = 2, p = 1, there exists a positive constant K p,n such that 2(p-1)/n
q
T
,
[fPx)P1P_1)dnx]
> Kp,n
(3.6)
a=1
K
p,n > - 21 np-2p/n(p_n/2)-1+2p/n(L1p-n/2,n /L p-n/2,n
)-1+2p/nL'
p-n/2,n
-2/n
(3.7)
Before giving the proof of Theorem 4, we discuss its relation to the well-known Sobolev inequalities [9, 101:
THEOREM 5 (Sobolev-Talenti-Aubin). Let V0 f Lt(Rn) with 1 < r < n. Let t = nr/(n-r). Then
f
INolr > Cr,n J 101t
r/t
(3.8)
for some Cr n > 0.
Talenti [11] and Aubin [21] have given the best possible Cr n (for n = 3, r = 2, t = 6, C2 ,n is also given in [8] and (121):
213
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
280
E. H. LIEB AND W. E. THIRRING
Cr,n
nrr =
r/2 (
n-r
r R1+ n-n/r)I'(n/r) r/n { I'(n)r(1+n/2) }
(3.9)
Our inequality (3.6) relates only to the r = 2 case in (3.8), in which case t = 2n/(n - 2). Consider (3.8) with r = 2 and 110 112 = 1. Using Holder's inequality on the right side of (3.8), one gets
fvI2 > C2,n[J
IliI2P/(P-i)12(p-i)/nIJ
I
iI21 -2(p-n/2)/n (3.10)
whenever n > 2 and p ? n/2. However, C2 ,n is not necessarily the best constant in (3.10) when p A n/2 (p = n/2 corresponds to r = 2 in (3.8)). Indeed, Theorem 4 says something about this question.
In the case that N = 1 and q = 1, Theorem 4 is of the same form as (3.10) (since p = I0I2 and 110 II2 = 1). We note two things: 1. For n > 2 and p = n/2, (3.6) agrees with (3.8) except, possibly, for a different constant. We have, therefore, an alternative proof of the usual Sobolev inequality (for the r = 2 case). As we shall also show Kn/2,n - C2 n, so we also have the best possible constant for this case. 2. If max In/2, 11 < p < 1 + n/2, Theorem 4 gives an improved version of (3.10), even it n = 1 or 2 (in which cases C2 ,n = 0, but Kp n > 0). For p > 1 + n/2, one can always use Holder's inequality on the p = I+ n/2 result to get a nontrivial bound of the form (3.10). However, in Theorem 4, the restriction p < 1 + n/2 is really necessary. This has to do with the dependence of T , on N rather than on n, as we shall explain shortly. Next we turn to the case N > 1. To illustrate the nature of (3.6), we may as well suppose q = 1. To fix ideas, we take a special, but important form for 0, namely G(xl,...,xN)
=
(N!)-t /2Dett(kr(x));N=1
(3.11)
and where the (k' are orthonormal functions in L2(Rn). Then, suppressing
the subscript o because q = 1,
214
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON
281
N P(x) _
Pi(x)
i=1
p'(x) = Id'(x)12
ti
=
J
IV.0ij2
Theorem 4 says that ti > Kp
,
(3.12)
.
p/(p-1)
f
n
C
Pi(x)1
2(p-1)/n (3.13)
dnx
If we did not use the orthogonality of the 95', all we would be able to conclude, using (3.6) with N = 1, N times, would be
t' > Kp
r n
()((3.14) 2(p-1)/n
fPixP/P_tdnx1
L
If p = n/2, then (3.14) is better than (3.13), by convexity. In the opposite case, p = 1 + n/2, (3.13) is superior. For in between cases, (3.13) is decidedly better if N is large and if the p' are close to each other (in the LP/(p-1)(Rn) sense). Suppose p'(x) = p(x)/N, i = Then the right side of (3.13) is proportional to N2p/n while the right side of (3.14) grows only as N. This difference is caused by the orthogonality of
the 0', or the Pauli principle. In fact, the last remark shows why p < I + n/2 is important in Theorem 4.
If p' = p/N, all i, then the best bound, insofar as the N dependence
is concerned, occurs when p is as large as possible. It is easy to see
215
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
E. H. LIEB AND W. E. THIRRING
282
by example, however, that the largest growth for TV due to the orthogonality condition can only be N(n+2)/n PROOF OF THEOREM 4. Let V(x) < 0 be a potential in Rn with at
least one bound state. If` e1 = min Ieil, then, for y c [0, 1],
,
JejIY > le,JY-1 I I IejI
Using the definition (2.14) and (3.2), we have that 1/y
IejI < Ay,n { J11 IVly+n/2 ((
(3.1S) )f
i
Ay,n = L y,n (L1y,n )-1+1/y
(3.16)
when 1 > y > max (0, 1 - n/2). (3.15) holds even if V has no bound state. be the projection onto the state a, i.e. for Let aa, a = (avgti) (x, a) = +/,(x, v) if a = v and zero otherwise. Choose 0 c L2(Rn; Cq), y= p - n/2. Let (polo= l be given by (3.3) and, for as > 0, or = 1, , q, define h
-A-
q
aapa(x)1/(y+n/2-1)aa
(3.17)
a=1
to be an operator on L2(Rn; Cq) in the usual way. Define N
HN =
hi
(3.18)
i=1
where hi means h acting on the i-th component of xN n q. Finally, let E = inf.spec HN. Now, by the Rayleigh-Ritz variational principle
216
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian
THE EIGENVALUES OF THE SCHRODINGER HAMILTON
283
q
E < (0, HN 0) = Tgb - 1 ao J
Po/(p-1)
(3.19)
a=1
On the other hand, E > the sum of all the negative eigenvalues of h
I 4
> - A y,n
1/y a aP/y
(3.20)
a
a=1
by (3.15). Combining (3.19) and (3.20) with
2(y-1)/n ao = S rppa/(P-1)
J
y
1
2y/n
{y+n 2AYnc
the theorem is proved.
Note that when p = 1 + n/2 (corresponding to y = 1 in the proof ), L1 In this case, the right side of (3.7) is the i,n does not appear in (3.7). best possible value of K1+n/2,nl as we now show. LEMMA 6. From (3.7), define
Li n - [n/(2K1+n/2,n)Jn/2 Then L1,n
=
(1+n/2)-1-n/2
L 1,n'
PROOF. By (3.7), we only have to prove that Lin > L1,n. Let V < 0, be the bound state V c C0 (Rn) and let H = -A + V. Let 10i, ei:N_ eigenfunctions and eigenvalues of H. Let 0 and p' be as defined in (3.11), (3.12). Then
Y
Ieil = - J VP - To <_ BVIIp11PNp/(p-1) - TO
with p = 1 + n/2. Using Theorem 4 for TIP, one has that
217
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon. A. Wightman eds.
E. H. LIEB AND W. E. THIRRING
284
11e;I < maxIIIVIIpY-K1+n/2,nY2p/°s = LI nIIVIP i
Y>O
We conclude with an evaluation of Kn/2
n
for n > 2 as promised. By
a simple limiting argument
(right side of (3.7)) .
lim
Kn/2,n >
pln/2
(3.21)
Our bound (2,11) on Lp-n/2,n shows that (Lp-n/2,n)-1+2p/n = 1
lim
.
(3.22)
pin/2
Hence
K
>
L1
-2/n
(3.23)
On the other hand, by the method of Lemma 6 applied to the N = 1 case, The value of is given in (4.24). To be Kn/2,n < honest, its evaluation requires the solution of the same variational problem as given in [8, 11, 12]. Substitution of (4.24) into (3.23) yields the required result Kn/2,n = C2 n = rrn(n-2)[I'(n/2)/I'(n)]2/n . (3.24) (L0',n)-2/n.
If we examine (3.23) when n = 2, one gets K1 2 > 0 since Lo 2 = W. This reflects the known fact [5] that an arbitrarily small V < 0 always has a bound state in two dimensions. This observation can be used to
show that
K1,2 = 0 .
(3.25)
When n = 1, the smallest allowed p is p = 1. In this case, (3.6) reads q
T
, ? K1.1 1 a=1
Using (3.7) and (4.20),
218
IIPoII;,
(3.26)
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON 1 K1.1 > [2L1/2.11-1 /2.1]
285
(3.27)
If one accepts the conjecture of Section 4 that L1/2,1
=
LI/2,1 ° 1/2,
then (3.28)
K1,1=1.
The reason for the equality in (3.28) is that K1.1 = 1 is well known to be the best possible constant in (3.26) when q = 1 and N = 1. 4.
Conjecture About Ly
n
We have shown that for the bound state energies (ej{ of a potential
V in n dimensions and with Ly,n(V) = t` Jejly/ f lV(y+n/2
(4.1)
then
Ly,n =
sup
Ly,n(V)
(4.2)
V t LY+n/2
is finite whenever y + n/2 > 1 and y > 0. The "boundary points" are
y1/2
n=1
y=0
n>2
(4.3)
We showed that for n = 1 ,
L1 /2
1 < -. For y < 1/2, n = 1, there
cannot be a bound of this kind, for consider VL(x) - -1/L for Ixl < L and zero otherwise. For L -. 0, this converges towards - 2S(x) and thus has a bound state of finite energy (which is -1 for -2S(x)). On the other hand, lim
fdx tVLI1/2+Y = 0
for
y < 1/2
L+ 0
For n = 2, y = 0 is a "double boundary point" and L0 2 = -, i.e. there is no upper bound on the number of bound states in two dimensions. (Cf. [5].)
219
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
E. H. LIEB AND W. E. THIRRING
286
For n > 3, LOrn is conjectured to be finite (see note added in proof, Section 5); for n = 3, this is the well-known f IVI3/2 conjecture on the number, NO(V), of bound states (cf. [51). The best that is known at present is that 4/3
NO(V) < c
If
IVI3/21
(4.4)
NO(V) < I (1 + 4 in Il I = 4(3rr2 31
,
/2)-1 fvI3I2
(4.5)
In (1.4) and (3.1), we introduced LC and L1 and showed that Ly,n > max(L4,n,L?,n)
(4.6)
A parallel result is Simon's [22] for n > 3: NO(V) < Dn e(IIV_IIE+n/2 + IIV-II_e+n/2)n/2
with Dn,r - oo as r - 0. In our previous paper [4], we conjectured that L1 3 = Li,3, and we also pointed out that Li, l > L1,1 A remark of Peter Lax (private communication), which will be explained presently, led us to the following:
CONJECTURE. For each n, there is a critical value of y,yc,n, such that
LYn= LCy,n 1 Lynn = Ly,n
220
Y?Yc
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian
THE EIGENVALUES OF THE SCHRODINGER HAMILTON
287
yc is defined to be that y for which LY,n = Ly n; the uniqueness of this yc is part of the conjecture. Furthermore, yc I = 3/2, Yc 2 ^ 1.2, Yc,3
.86 and the smallest n such that yc,n = 0 is n = 8.
(A) Remarks on Lynn We want to maximize
fu2
/f Y
V + IVsGI2]dnx)
IVly+n/2
(4.7)
with respect to V, and where f 1,p 12 = 1 and (-A+V),A = e1i . By the variational principle, we can first maximize (4.7) with respect to V, holding 0 fixed. Holder's inequality immediately yields V(x) =
-aIVi(x)I2/(y+n/2-1)
with a > 0. The kinetic energy, f IVI#12, is not increased if Vi(x) is replaced by IVi(x)I and, by the rearrangement inequality [13], this is not increased if 101 is replaced by its symmetric decreasing rearrangement. Thus, we may assume that IVI and Irlil are spherically symmetric, nonincreasing functions. By the methods of [8] or [11], (4.7) can be shown to have a maximum when y + n/2 > 1. The variational equation is
-A4i(x) -
agr(x)(Y+n/2+1)/(y+n/2-1)
with 1
a =
YIe1IY
= e1V,(x)
(4.8)
}1/(Y+n/2._1) (4.9)
(y + n/2) Ly n
Equation (4.8) determines 0 up to a constant and up to a change of scale in x. The former can be used to make f 02 = 1 and the latter leaves (4.7) invariant.
221
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds. E. H. LIEB AND W. E. THIRRING
288
Equation (4.8) can be solved analytically in two cases, to which we shall return later: (i) n = 1, all y>1/2
(ii) n > 3, y = 0. (B) The One-Dimensional Case Lax's remark was about a result of Gardner, Greene, Kruskal and
Miura (14] to the effect that C
1'3/2,1 =1'3/2,1 = 3/16.
(4.10)
To see this, we may assume V c C0 '(R), and use the theory of the Korteweg-de Vries (KdV) equation [14]: (4.11)
Wt = 6WWX - Wxxx
There are two remarkable properties of (4.11): (i) As W evolves in time, t, the eigenvalues of -d2/dx2 + W remain invariant.
(ii) j W2 dx is constant in time. Let W(x, t) be given by (4.11) with the initial data W(x, 0) = V(x)
.
Then L3/2 1(W( , t )) is independent of t, and may therefore be evaluated by studying its behavior as t w. There exist traveling wave solutions to (4.11), called solitons, of the form
W(x, t) = f(x - ct)
.
Equation (4.11) becomes
-cfx = -fxxx+6ffx
(4.12)
The solutions to (4.12) which vanish at m are fa(x) = - 2a2 cosh-2 (ax) c = 4a2
222
.
(4.13)
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian
289
THE EIGENVALUES OF THE SCHRODINGER HAMILTON
Any solution (4.13), regarded as a potential in the Schrodinger equation
has, as we shall see shortly, exactly one negative energy bound state with energy and wave function
e = -a2 0(x) = cosh-I (ax)
.
(4.14)
Now the theory of the KdV equation says that as t - -, W evolves into a sum of solitons (4.13) plus a part that goes to zero in L°°(R) norm (but not necessarily in L2(R) norm). The solitons are well separated since they have different velocities. Because the number of bound states is finite, the non-soliton part of W can be ignored as t -. Hence, for the initial V,
I
I jej13/2
a3
(4.15)
solitons while
f V(x)2 dx
I
ffa(thx
(4.16)
solitons
Since 4 JG cosh -4(x)dx = 16/3, we conclude that c 1'3/2,1 = 1'3/2,1 = 3/16
(4.17)
with equality if and only if W(x, t) is composed purely of solitons as w. For the same reason, t
L'3/2,1 = LC3/2,1
(4.18)
(cf. (4.21)).
Not only do we have an evaluation of L3/2,1, (4.17), but we learn something more. When y = 3/2, there is an infinite family of potentials for which L3/2,1 (V) = L3/2 1, and these may have any number of bound states = number of solitons. What we believe to be the case is that when y < 3/2, the optimizing potential for Ly n has only one bound state, and satisfies (4.8). When
223
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
E. H. LIEB AND W. E. THIRRING
290
y > 3/2, the optimizing potential is, loosely speaking, infinitely deep and has infinitely many bound states; thus Lynn = LY,n. An additional indication that the conjecture is correct is furnished by the solution to (4.8). When y = 3/2, this agrees with (4.14). In general, one finds that, apart from scaling, the nodeless solution to (4.8) is clr
(x) = r(Y)1 /2 n 1 /4 I'(y - 1/2)-1 /2 cosh-Y+1 /2(x)
Vy(x) = -(y2 - 1/4)cosh-2(x)
e1 = -(y-1/2)2
(4.19)
.
Thus,
LI
=
rr-1/2
1
Y.1
r(y+1)
/y-1/21Y+1/2
I'(y+1/2)\YY+
/
(4.20)
When Ly , is compared with LY.,, one finds that
Ly,1 > LY 1
y < 3/2
LyI
y>3/2
(4.21)
This confirms at least part of the conjecture. However, more is true. For y = 3/2, VY has a zero energy single node bound state
O(x) = tanh(x)
.
Since VY is monotone in y, it follows that VY has only one bound state for y < 3/2 and at least two bound states for y > 3/2. The (unnormalized) second bound state can be computed to be
O(x) = sinh(x) cosh-Y+1/2(x) e2 = - (y - 3/2)2
.
(4.22)
In like manner, one can find more bound states as y increases even further.
224
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON
291
Thus we see that the potential that optimizes the ratio telly/fIV1y+I/2 automatically has a second bound state when y > yc. Finally, we remark that Theorem 3, together with (4.10), shows that THEOREM 7. Ly 1 = LC
y = 3/2, 5/2, 7/2, etc.
for
1
An application of Theorem 7 to scattering theory will be made in Section 4(D). (C) Higher Dimensions
We have exhibited the solution to the variational equation (4.8) for Ly 1.
When n > 2 and y = 0, we clearly want to take e1 = 0 in order to maximize Lo,n(V). (4.8) has the zero energy solution
c(x) = V(x) =
ac(x)2/(y+n/2-1)
(1+lxl2)1-n/2
=
n(n-2)(1+IxI2)(2-n)/(n/2-1)
(4.23)
(note: (h r L2(Rn) if and only if n > 4, but V ( Ln/2(Rn) always). This
leads to
Lp'n =
[nn(n-2)]-n/2 r(n)/I'(n/2)
.
(4.24)
The smallest dimension for which Lo n < LO n is n = 8. If we suppose that the ratio Ly n/LY n is monotone decreasing in y (as it is when n = 1 and as it is when n = 3 on the basis of the numerical solution of (4.10) by J. F. Barnes, given in Appendix A), and if our conjecture is correct, then Ly n = LY n for n > 8. The value of yc obtained numerically is yc=1.165
n=2
yc = .863
n= 3
.
(4.26)
225
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
292
E. H. LIEB AND W. E. THIRRING
The other bit of evidence, apart from the monotonicity of Ly n/LY n, for the correctness of our conjecture is a numerical study of the energy levels of the potential 'k >0 ,
VX(x)_Ae-t'I,
in three dimensions. This is given in Appendix A. The energy levels of the square well potential are given in [15, 16]. In both cases, one finds that
lim Lt,3(Vx) = Li3
a»0
and the limit is approached from below. Unfortunately, it is not true, as one might have hoped, that Lt 3(Vk) is monotone increasing in \. (D) Bounds on One-Dimensional Scattering Cross-Sections In their study of the KdV equation, (4.11), Zakharov and Fadeev [ 17]
showed how to relate the solution W(x, t) to the scattering reflection
coefficient R(k) and the bound state eigenvalues lejl of the initial potential V(x). There are infinitely many invariants of (4.11) besides fW2 and these have simple expressions in terms of R(k), lejl. Thus, for any potential V,
J V3 + 2 Vx = -(32/5)
f V4+2VV+ 1 Vxx = (256/35)
Iej(512- 8 J
k4 T(k) dk
(4.28)
ejI7/2 -(64/5) f k6 T(k)dk (4.29)
where
T(k) = , r
226
In (1- I R(k)I2) < 0 .
(4.30)
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON
293
These are only the first three invariants; a recursion relation for the others can be found in [171.
Notice that 3/16, 5/32, 35/256 are, respectively L3,,2,1, Ls12,1,
Since f V' > 1IV(4.27) establishes that L3/2,1 = 1'3/2,1' 1 as mentioned earlier. For the higher invariants, the signs in (4.28) and (4.29) are not as fortunately disposed and we cannot use these equations to prove Theorem 7. But, given that Theorem 7 has already been proved, we can conclude that L712
THEOREM 8. For any nonpositive potential V(x),
fv
? -16 E k4T(k) dk .
(4.31)
For any potential V(x), 00
2
fvv2 + (1/5)
fv Xx < -(64/5) J
k6 T(k)dk
.
(4.32)
The first inequality, (4.31), is especially transparent: If V(x) is very smooth, it cannot scatter very much. 5.
Note Added in Proof
After this paper was written, M. Cwikel and Lieb, simultaneously and by completely different methods, showed that the number of bound states,
N0(V) for a potential, V, can be bounded (when n > 3) by
N0(V) < An
J
I V(x)i n/2 dnx
(5.1)
Cwikel exploits the weak trace ideal method of Simon [221; his method is more general than Lieb's, but for the particular problem at hand, (5.1), his
227
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
E. H. LIEB AND W. E. THIRRING
294
An does not seem to be as good. Lieb's method uses Wiener integrals and the general result is the following: 00
N_a(V) < Cdnx f
J
dt t-Ie-at(4nt)-n/2 f(tIV(x)I-)
(5.2)
0
[0, oo) satisfying
for any non-negative, convex function f : [0, oe)
1=
f
00t-1
f(t) a-t dt
(5.3)
.
0
For a = 0, one can choose f(t) = c(t-b), t > b, f(t) = 0, t < b. This leads to (5.1), and optimizing with respect to b, one finds that
A3 = 0.116,
(5.4)
A4 = 0.0191
and, as n . 00,
An/Loin = (nv)1/2 + O(n 1/2)
.
(5.5)
Note that A3/L0 ' 3 = 1.49, i.e. A3 exceeds Lo 3 by at most 49%. Since N_a(V) < No(- I V +al _), one can use (5.1) and (2.1) to deduce
that for r? 0 and n>3, Lynn < LY n(An/Lo n)
(5.6)
This is better than (2.11), (2.18). In particular, for n = 3, y = 1, the improvement of (5.6) over (2.11) with m = 2 is a factor of 1.83. The factor 1.31 in Equation (1.4) can therefore be replaced by 1.31 (1.83)-2/3 = 0.87.
228
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian
THE EIGENVALUES OF THE SCHRODINGER HAMILTON
295
APPENDIX A. NUMERICAL STUDIES
John F. Barnes Theoretical Division Los Alamos Scientific Laboratory Los Alamos, New Mexico 87545
Evaluation of LI n, n = 1, 2, 3 y, The figure shows the numerical evaluation of Lynn as well as Ly The latter is given in (1.11) 1.
n
2-n,7-n/2r(y+1)/C'(y+1+n/2)
LY,n =
The former is obtained by solving the differential equation (4.8) in polar coordinates and choosing a such that c(t(x) - 0 as Ixi - W. Note that
by scaling, one can take el = -1, whence (LI
y,n
)-1
=
,(y+n/2)
Iql(x)I(2y+n)/(y-l+n/2)dnx
In one dimension, Ly l is known analytically and is given in (4.20). Another exact result, (4.24), is Lo 3 =
4n-2 3-3/2
= 0.077997
The critical values of y, at which Ly n = Ly,n are: yc I = 3/2 yC 2 = 1.165 0.8627
229
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
E. H. LIEB AND W. E. THIRRING
296
J
0.0011
0
1
.2
1
.4
1
.6
I
1
1
.8
1.0
12
1
r
230
I
1
I
I
1
I
14 16 1.8 2.0 2.2 2.4 2.6 28
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON
11.
297
The Exponential Potential
To test the conjecture that L1,3 = L1,3, the eigenvalues of the potential VA _ -X exp(- Ixi) in three dimensions were evaluated for A = 5, 10, 20, 30, 40, 50, and 100. These are listed in the table according to angular momentum and radial nodes. These numbers have been corroborated by H. Grosse, and they can be used to calculate Ly 3(Vx) for any
y. The final column gives L1 3(Vx), since f1Vjj5/2 = kS12(64rr)/125. It is to be noted that the classical value L 3 = 0.006755,
is approached
from below, in agreement with the conjecture, but not monotonically.
VA= -Ae-r Q
lei
nodes
states
lei
x5/2 4 125
A= 5
4 = 10
A = 20
0
0.55032
0
1
0.55032 0.55032
2
2.2520
3
1.0022
5
3.2542
3
8.0584
8.6342
1
0.06963
1
2.18241
0
1
0.33405
0
0
0.00869
2
1.42562
1
6.62410
0
0.16327
1
2.71482
0
6
0 43136
0
1
0
1
2
.
5
2.1568
0 006120 .
0 006398 .
0 006551 .
18.8 94
231
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
E. if. LIEB AND W. E. THIRRING
298
VA = -Ae-t (continued) lei
A = 30
0.58894
2
3.83072
1
11.84999
0
1.39458
1
6.12302
0
0.00593
1
2.36912
0
3
0 07595
0
0
0.07676
3
1.86961
2
6.88198
1
17.53345
0
0.41991
2
3.35027
1
10.13596
0
0.93459
1
5.03378
54738
0
1
2
A = 40
nodes
1
2
3
.
1
.
states
lei
3
16.270
6
22.553
10
26
125
11.875
51.22 30
4
26.362
9
41.718
0
10
29.842
0
7
10.832
0 . 006461
0 006682 .
108.754
30
232
A5/4n
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian
299
THE EIGENVALUES OF THE SCHRODINGER HAMILTON
VA = -Ae-r (continued)
,lei
A = 50
f
lei
nodes
0
0.60190
3
3.66447
2
10.39110
1
23.53215
0
1.43321
2
5.81695
1
14.56904
0
0.07675
2
2.45887
1
8.19840
0
0.26483
1
3.61626
0
0.49009
0
1
2
3
4
states
4
38.190
9
65.458
15
53.670
14
27.168
ST
A = 100
0
0.39275
5
2.91408
4
8.29231
3
17.44909
2
32.07168
1
lei
A 5/2
4a
0.006643
233
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
300
E. H. LIEB AND N. E: THIRRING
VA = -Ae-t (continued) P
1
2
3
4
234
,
let
nodes
states
56.28824
0
6
117.41
1.10170
4
4.76748
3
11.62740
2
22.79910
1
40.45495
0
15
242.25
0.02748
4
2.04022
3
6.85633
2
15.22147
1
28.46495
0
25
263.05
0.22692
3
3.14743
2
9.13429
1
19.04073
0
28
220.85
0.52962
2
4.37856
1
11.56470
0
27
148.26
lei
A 512 64n
irs
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON
301
VA = - Ae-r (continued) iel jel
5
6
nodes
A5/2 4n
states
0.88997
1
5.69707
0
22
1.26789
0
f
let
72.46 6.48 108 0.76
0.006719
APPENDIX B: PROOF OF (2.3) THEOREM 9.
Let x be a separable Hilbert space and let A, B be
positive operators on J{. Then, for m > 1, Tr (B/2 A BI /2)m < Tr BM/2 Am Bm/2
.
(B.1)
REMARK. When .){ = L2(Rn) and A is a kernel a(x-y) and B is a multiplication operator b(x) (as in our usage (2.2)), Seiler and Simon [19] have given a proof of (B.1) using interpolation techniques. Simon (private communication) has extended this method to the general case. Our proof is different and shows a little more than just (B.1).
PROOF. For simplicity, we shall only give the proof when A and B are matrices; for the general case, one can appeal to a llimiting argument. For m = 1, the theorem is trivial, so assume m > 1. Let C = Am and f(C) g(C) - h(C), where g(C) = Tr (B1 /2 C1 /m B1 /2)m and h(C) = Tr Let M' be the positive matrices. Clearly M+ ) C -. h(C) is linear. Epstein [181 has shown that M+) C - g(C) is concave (actually, he showed this for m integral, but his proof is valid generally for m > 1). Write C = CD + CO where CD is the diagonal part of C in a basis in which B is diagonal. CA CD + A CO ='k C + (1-A) CD is in M+ for A E (0, 11,
235
Bm/2CBm/2.
With W. Thirring in Studies in Mathematical Physics, E. Lieb, B. Simon, A. Wightman eds.
302
E. H. LIEB AND W. E. THIRRING
because CD c M+. Then A f(CA) = R(A) is concave on [0, 11. Our goal is to show that R(1) < 0. Since [CD , B] = 0, R(O) = 0 and, by concavity, it is sufficient to show that R(A) < 0 for A > 0 and A small. h(CA) h(CD) for A c [0, 11. Since f(C) is continuous in C, we can assume that CD is nondegenerate and strictly positive, and that CA is positive when A _> - e for some E > 0. Then R(A) is defined and concave on 1/m is differentiable at A = 0 and its derivative at A = 0 11. A CA has zero diagonal elements. (To see this, use the representation Cl/m = K fo dx x-1+1/m . C(C+ xl)-I.) Likewise, the derivative of (BI /2(D + AO) Bt /2)m at A = 0 has zero diagonal elements when 0 has and when
D is diagonal. Thus
dR(A)/dAIx=0 = 0
.
Acknowledgment
One of the authors (Walter Thirring) would like to thank the Department
of Physics of the University of Princeton for its hospitality. ELLIOTT H. LIEB DEPARTMENTS OF MATHEMATICS AND PHYSICS PRINCETON UNIVERSITY PRINCETON, NEW JERSEY
WALTER E. THIRRING INSTITUT FUR THEORETISCHE PHYSIK DER UNIVERSITAT WIEN, AUSTRIA
REFERENCES [1]
E. H. Lieb and W. E. Thirring, Phys. Rev. Lett. 35, 687(1975). See Phys. Rev. Lett. 35, 1116 (1975) for errata.
[2]
M. S. Birman, Mat. Sb. 55(97), 125(1961); Amer. Math. Soc. Translations Ser. 2, 53, 23 (1966).
[3]
J. Schwinger, Proc. Nat. Acad. Sci. 47, 122 (1961).
[4]
B. Simon, "Quantum Mechanics for Hamiltonians Defined as Quadratic Forms," Princeton University Press. 1971.
236
Inequalities for the Moments of the Eigenvalues of the Schrodinger Hamiltonian THE EIGENVALUES OF THE SCHRODINGER HAMILTON
303
[51
B. Simon, "On the Number of Bound States of the Two Body Schrodinger Equation - A Review," in this volume.
[61
A. Martin, Hely. Phys. Acta 45, 140 (1972).
[7]
H. Tamura, Proc. Japan Acad. 50, 19 (1974).
[8]
V. Glaser, A. Martin, H. Grosse and W. Thirring, "A Family of Optimal Conditions for the Absence of Bound States in a Potential," in this, volume.
[9]
S. L. Sobolev, Mat. Sb. 46, 471(1938), in Russian.
[101
, Applications of Functional Analysis in Mathematical Physics, Leningrad (1950), Amer. Math. Soc. Transl. of Monographs, 7(1963).
[11] G. Talenti, Best Constant in Sobolev's Inequality, Istituto Matematico, University Degli Studi Di Firenze, preprint (1975). [12] G. Rosen, SIAM Jour. Appl. Math. 21, 30(1971).
[13] H. J. Brascamp, E. H. Lieb and J. M. Luttinger, Jour. Funct. Anal. 17, 227 (1974). [14] C. S. Gardner, J. M. Greene, M. D. Kruskal and R. M. Miura, Commun.
Pure and Appi. Math. 27, 97 (1974). [15] S. A. Moszkowski, Phys. Rev. 89, 474 (1953). [16] A. E. Green and K. Lee, Phys. Rev. 99, 772(1955).
[171 V. E. Zakharov and L. D. Fadeev, Funkts. Anal. i Ego Pril. 5, 18(1971). English translation: Funct. Anal. and its Appl. 5, 280 (1971).
[181 H. Epstein, Commun. Math. Phys. 31, 317 (1973).
[19] E. Seiler and B. Simon, "Bounds in the Yukawa Quantum Field Theory," Princeton preprint (1975). [201 W. Thirring, T7 Quantenmechanik, Lecture Notes, Institut filr Theoretische Physik, University of Vienna.
1211 T. Aubin, C. R. Acad. Sc. Paris 280, 279(1975). The results are stated here without proof; there appears to be a misprint in the expression for Cm. [221 B. Simon, "Weak Trace Ideals and the Bound States of Schrodinger Operators," Princeton preprint (1975).
237
With M. Aizenman in Phys. Lett. 66A, 427-429 (1978) Volume 66A, number 6
26 June 1978
PHYSICS LETTERS
ON SEMI-CLASSICAL BOUNDS FOR EIGENVALUES OF SCHRODINGER OPERATORS* Michael AIZENMAN Department of Physics. Princeton University. Princeton. NJ 08540. USA and
Elliott H. LIEB Departments of Mathematics and Physics. Princeton University Princeton, NJ 08540, USA Received 27 April 1978
Our principal result is that if the semiclassical estimate is a bound for some moment of the negative eigenvalues (as is known in some cases in one-dimension), then the serniclassical estimates are also bounds for all higher moments.
Bounds on the moments of energy levels of Schrtidinger operators have been the object of several studies 11, 2, 5-81. In [ I I such bounds were used to obtain a lower bound for the kinetic energy of fermions in terms of their one particle density and thereby prove the stability of matter.
In the notation of (21,
and d"z = d"x drip (21r)-n. R,,(-f. V) is the ratio of the moments of the binding energies of a quantum mechanical hamiltonian to the moments of its classical analog. The integral in (1) comes from doing the d"p integration in (2). In the notation of (2[ Rn(Y) = Ly, /L`y,n
(4)
-For V < 0, V E Cp (R") it is known 13,4,91 that
Z;lei(V)Ir
(I)
where ei(V) are the eigenvalues of -A + V(x) defined in L2 (R") and lyl_ = max(-y, 0). Lv n denotes the smallest number for which (1) holds independently of V. The case y = I is the one needed for the kinetic energy bound. It was shown in (21 that L y., <m for y > max(0, I - n12). Eq. (1) also holds for y = 0, n > 3 but the proof is quite different (see refs. (5, 6, 81). For ,y > 0 we use the notation
H(x,p)=p2+V(x) If d"z l H(x, p)l Rn(y, V) = E l ei(V)I;I
(2)
Rn(7') = sup (Rn(y. V)} ,
(3)
V
Rn(y,1t V) -v I
(5)
as X - se, which is the semiclassical limit. Thus Rn('7) > 1
.
(6)
In [2[ it was conjectured that Rn(y) = I for certain y and n, in particular for y = 1, n = 3 which is the case of primary physical interest. R3(I) - I would imply that the Thomas-Fermi theory of atoms and molecules (together with a modified treatment of the electronelectron repulsion) gives a lower bound to the true Schr6dinger ground state energy (see (I I). The only cases where the value of Rn(y) is known are n = 1, y = 3/2, 5/2, 7/2, ..., where R1(y) = I. Part (a) of the following theorem, together with (6), settles the question for n - 1, y > 3/2. Theorem:
Work partly supported by U.S. National Science Foundation grant MCS 75-21684 A02.
(a) For any n, R"(y) is a monotone nonincreasing function of y. 427
239
With M. Aizenman in Phys. Lett. 66A, 427-429 (1978) 26 June 1978
PHYSICS LETTERS
Volume 66A, number 6
(b) If, for some 'y > max(O, 1 - n/2), the supremum in (3) is attained, i.e. R"(y) = Rn(y, V) for some V is strictly decreasing with 1171 _ E L7+n/2, then from the left at y. In fact
R"(y, V) = R"(y)
liminf [R"(y - b) - R"(y)) /6 > 0 .
is strictly decreasing from the We shall prove that left at y by showing that
6-0+
(13)
In particular (13) implies that
0 > e0 = inf spec(-A + V) > ess inf {V(x)) .
(14)
lim inf [R"(y - 6) - R"(-I)1 /6
Proof
6-0'
(a) Fix V. For y > 0, 6 > 0, let
(15)
fd"zlH(x,P)I' rt,(Ieo/Hl) dXX- I+d I I-XI'r<«+.
1(y,6)=f
(7)
0
> Rn(7)
fd"zIH(x,p)I!
>0
where
By scaling, for any e E R
1
I,tl,(t)= f dX(1 -X)y/X>(1
lell'6 =1(y.8)-1 f dXX-I+a le+Xly .
-t)t+rl(l +y).
(16)
(8)
0
The key fact which will be used is that the integral in (9) can be cut off from above at leoI
Thus, for any SchrOdinger potential V,
For anyS>0 Z!I e1(V)l !+6
EIe/V)111 =/(y-6,6)
=I(y.6Y1 f d), ),-'+6 Ele/(V)+XI" . 0
(9)
J
However, a/(V) + X are the eigenvalues of the potential V(x) + ) . Therefore, by definition (3),
Filet(V)+Al
(10)
I
leol
Xf
dXX-1+6
0
(V)+X111 -6 J
Xfd"z f d),),-1+61H(x.p)+XI'r-6 0
and, by substitution in (9) and using (8),
=R"(y 6)fd"zIH(x,p)IV Eler(V)1'''`6
/
0
(11)
-Rn(7-6)I(y-6,6)-1
X f d"z I H(x, p) + X111 = Rn(y) f d"z IH(x, P)I11+6 . Hence
R"(y + 6, V)
(12) Rn(y +6)max(0, 1 - n/2) and assume that for some V with IVI_ EL7+"/2(R")
428
240
Xfd"z f
dXX`I+6IH(x,p)+XIy-6.
(17)
leol
Therefore, using (2), (13) and Rn(y - 6) > Rn(y)
,
On Semi-Classical Bounds for Eigenvalues of Schrodinger Operators
Volume 66A. number 6
PHYSICS LETTERS
(R.(7 - 6) - Rn(7)1 /6 > (R"(7)/61(7 - 6.6)1
X f d"z I H(x, p)l y H <eo
Xf
(18) dX 7t- '+' (1 - X)y-b'1 d"z I H(x. p)l r .
leo/HI
Using lim5_p. 61(7 - 6, 6) = 1 and Fatou's lemma, we obtain (IS). In view of (16) the theorem implies
Corollary 1: If for some y, R"(7) = I then Rn(y)
= I for ally > 7. Moreover, for j > 7 the supremum in (3) is not attained by any potential. This proves part of a conjecture made in (21 (another part of the conjecture was disproved, for n > 7, in (71). In one dimension we can say even more since it is
known, (21, that R t (312) = 1 (R i (7) > 1 for 7 < 3/2).
Corollary 2: For ?>312,Rl(7)= 1.
26 June 1978
stricted classes of potentials, V, as was done in 171
for the spherically symmetric ones (the constants thus obtained are no larger than R,(-y) but it is not known whether any of them are strictly smaller). The theorem and its proof extend to such bounds as long as the class of potentials is closed under the addition of constants. References
III E.H. Lieb and WE. Thvring, Phys. Rev. Lett. 35 (1975) 687. See Phys. Rev. Lett. 35 (1975) 1116 for errata. Also E.N. Lieb, Rev. Mod. Phys. 48 (1976) 553. (21 E.H. Lieb and WE. Thirring. in: Studies in mathematical physics, Essays in honor of V. Bargmann (Princeton Univ. Press, Princeton, N.J., 1976). 131 A. Martin, Helv. Phys. Acta 45 (1972) 140. 141 H. Tamura. Proc. Japan Acad. 50 (1974) 19. 151 M. Cwikel, Ann. of Math. 106 (1977) 93.
16) E.H. Lieb, Bull. Amer. Math. Soc. 82 (1976) 751. 171 V. Glaser, H. Grosse and A. Martin, Bounds on the Number of Eigenvalues of the Schrbdinger Operator, CERN preprint TH2432 (1977). 181 G.V. Rosenblum, The distribution of the discrete spectrum for singular differential operators, lsvestia Math.
164 No.I(1976)75. 191 M.S. Birman and V.Y. Borzov, On the asymptotics of the discrete spectrum of some singular differential operators, Topics in Math. Phys. 5 (1972) 19.
One may also study bounds like (3) for some re-
429
241
Proceedings of the Amer. Math. Soc. Symposia in Pure Math. 36, 241-252 (1980)
THE NUMBER OF BOUND STATES OF ONE-BODY SCHROEDINGER OPERATORS AND THE WEYL PROBLEM
Elliott H. Lieb1
ABSTRACT.
If R(0,A) is the number of eigenvalues of -A in a
domain 0 in a suitable Riemannian manifold of dimension n, we
derive bounds of the form R(tt,A)< DnAn/2IQ,
for aZZ S2,
A
,n
Likewise, if Na(V) is the number of nonpositive eigenvalues of -A + V(x) which are < a < 0, then
Na(V)< LnIM [v
af12
for all a and V and n > 3.
1.
INTRODUCTION AND BACKGROUND.
Two closely related problems will concern us here : One is to bound the nonpositive eigenvalues of the one-body Schroedinger operator H - -A + V(x).
(1.1)
The other is to find an upper bound for R(it,A) - number of eigenvalues of -A < A
(1.2)
in a domain t with Ditichlet boundary conditions.
In both cases the setting is a Riemannian manifold, M,and A is the Laplace-Beltrami operator.
The only way in which the properties of the mani-
fold will appear in our results will be through the fundamental solution of the heat equation
G(x, y; t) -
[
e x p(t A)] (x, y)
(1.3)
1980 Mathematics Subject Classification 35P15. I Work supported by U.S.National Foundation grants PHYS7825390 and TNT 78-01160.
241
243
Proceedings of the Amer. Math. Soc. Symposia in Pure Math. 36, 241-252 (1980)
ELLIOTT H. LIEB
242
evaluated on the diagonal x - y, or else through the Green function e)-1
(-A +
(x, y) - 10 dt a-et C(x, Y; t)
(1.4)
(Note: G is defined for the whole manifold, not the subdomain
for e > 0.
One could of course, use C defined for the domain
Q in the case of (1.2).
S2 in all the formulas, but then the dependence of the result on 0 will be complicated.
It is precisely to avoid this complication that we use the C. for
the whole of M).
Let us begin with the problem defined by (1.2), which we may term the Weyl problem.
Weyl [1] proved the asymptotic formula (for suitable domains)
MIX)
Cn 0/2IOI
(1.5)
where I0I is the Riemannian volume of Si, n is the dimension of M and C
=
(4n)-n/2
C( 1 + n/2 )-1
n
(2n) n Tn
(1.6)
,
is the volume of the unit ball in1Rn. The constant Cn is called T n the classical constant for reasons which will become clear later. This result
where
(1.5) is discussed in (21 and (3], §XIQ.15, vol.4.
The proof uses Dirichlet-
Neumann bracketing.
Polya's conjecture is that (1.5) holds in Rn for all A and Q, not just asymptotically. THEOREM 1.
Here we will prove For all
on the manifold M)
A and 0 there exist constants Dn and En (depending
such that
In/2In'
R(S2 A) < Dn if
C(x, x; t)
for some An <
n
< An t-n/2, Vx C M, Vt > 0
and
n
(1.8)
while
(n,a) < (Dn In/2 + En)Is1I -n/2 + Bn, G(x, x; t) S Ant
if D
(1.7)
are proportional to A
(1.9)
Vx C M, Vt > 0,
(1.10)
and Bn respectively. n
In particular (1.8) and (1.7) hold for IRn (with A
n
- (4n)-n/2) and for
many noncompact M, e.g. homogeneous spaces with curvature <- 0.
(1.10) and
(1.9) hold for compact M.
Next we turn to the Schroedinger problem (1.1). be the nonpositive eigenvalues of H on L2(M).
If we write
V - V+ - Vwith
V+ (x)
IV(x)I
(1.11) ,
when V(x) < 0
0, otherwise
244
Let EI(V) < E2(V) <_...5 0
(1.12)
The Number of Bound States of One-Body Schrodinger Operators and the Weyl Problem BOUND STATES OF SCHROEDINGER OPERATORS AND THE WEYL PROBLEM 243
then the negative spectrum of H is discrete if V _E: Ln/2, for example. Na(V) is the number of eigenvalues of H which are <- a <- 0.
DEFINITION.
Our main result is Suppose.(1.8) holds and suppose n > 3. Then
THEOREM 2.
N0(V) S Ln 41 V_(x)n/2 dx
(1.13)
for some constant Ln depending on M. There are many remarks to be made about Theorem 2 and its connection with Rosenbljum (4) first announced
First, the history of (1.13).
Theorem 1.
Unaware of this, Simon (5) proved an inequality of the form
Theorem 2 in 1972.
n/2+En/2
NO(V) < Sn,el II V-II n/2+c with Sn,c
c 4 0.
as
+
11 V+11
Also unaware of (4], Cwikel and myself (6,7) simulReed and Simon (3) call Theorem 2 the
taneously found a proof of Theorem 2.
Cwikel's method exploits some ideas in (5].
Cwikel - Lieb - Rosenbljum bound.
All three methods are different, Cwikel's and Rosenbljum's are applicable to a wider class of operators than the Schroedinger operator, but my method [7], based on Wiener integrals, which is the one presented here, gives the This result was announced in [7) and the proof was
best constant by far.
written up in [3], Theorem XIII.12 and in [8].
Because all the technical de-
tails can easily be found in [8], the presentation given here will ignore technicalities nical
.
Not only am I indebted to B.Simon for his help with the tech-
details, as just mentioned, but I also wish to acknowledge his role in
stimulating my interest in the problem, and for his constant encouragement while the ideas were taking shape.
The connection between the two theorems is Let
PROPOSITION 3.
a <- 0.
Then for all M
8(Q,X) 5 Na( (a-a) where
XO
)
(1.14)
)
is the characteristic function of
PROOF.
Let 1P j be the j th
on all of M, be ip j
Q.
eigenfunction of -A in I
in Q and zero outside.
and let j, defined
(1.14) is obtained by using the
j as variational functions for -A + (a - A) Xn
in the Rayleigh - Ritz varia-
QED.
tional principle.
Similarly, we have if 0 < B < -a , then for all M
PROPOSITION 4.
Na(V) < Na PROOF.
+
8(-[V + 8]-).
Same as for proposition 3.
(1.15)
Alternatively, one can remark that
V(x) >-(V + 61-(x) - B, and adding a positive operator can never decrease any eigenvalue.
QED.
245
Proceedings of the Amer. Math. Soc. Symposia in Pure Math. 36, 241-252 (1980)
ELLIOTT H. LIEB
244
In particular, Na(V) 5 N0(-[V - a]_) and
N0(V) 5 N0(-V_), whencz we
have If
COROLLARY 5.
and
3
n -.
Na(V) <
(1.13) holds, then
Ln IM [V - a]- (x)n/2 dx.
(1.16)
Moreover to prove (1.13) it is sufficient to consider V satisfying V(X) < 0, Yx.
Proposition 3 will be used to derive Theorem 1 from Theorem 2 (actually from a generalization of Theorem 2, namely Theorem 8 and (4.3), which holds for all n).
However, at this point we can, under the assumption that (1.8)
holds, deduce (Y.7) of Theorem 2 from proposition 3 and Theorem 2. Choosing
a - 0, we have
Ln
t7(Sl,A)
an/2
Xiln/2
jM
dx
(1.17)
- Ln n/2InI n ? 3.
for
It is to be emphasized that Theorem 2 is more delicate than Theorem 1.
For one thing Theorem 1 holds for all n, whereas Theorem 2 holds only for The analogue of Theorem 2 is definitely false for n - 1 and 2.
n 2 3.
Rn,
In
at least,an arbitrarily small, nonpositive potential always has a nega-
tive eigenvalue when n <3; cf. [3], Theorem XIII.11. best constant in (1.7) is
constant Cn (in Rn at least). in general.
Cn
However, the best constant
n
7;
for
is bigger than
Ln
It is easy to prove [9] (by considering a very "large" V and
using Dirichlet-Neumann bracketing) that Ln > Cn
For another thing, the
, according to the Polya conjecture, the classical
3 < n 5 7
Ln 2 Cn.
In [9] it was shown that
and recently [16] it was shown that
Ln > Cn
for
in both cases explicit examples were constructed.
It is somewhat ironic that although Theorem 2 starts to hold only for
n - 3, Theorem 1 is easiest to prove for Rn with
n - 1.
In that case the
only domain that need be considered is a finite interval, and there the eigenvalues of
-A
can be computed explicitly.
The Polya conjecture is easily
seen to be true.
The intuition behind Theorem 2, and thereby the reason for calling Cn classical constant is important.
the
In the semiclassical picture of quantum me-
chanics in Rn, which is similar to WKBJ theory, one has the mystical postulate that "each nice set in phase space (2
)n
r - {(p,x)l p C Rn, x E Rn}
can accomodate one eigenstate of H".
of volume
This postulate can be made more
precise by mean of the Dirichlet-Neumann bracketing method mentioned before. In any event, since the "eigenvalues" of of V are V(x), the postulate implies that
246
-A
are
p2
and the "eigenvalues"
The Number of Bound States of One-Body Schrodinger Operators and the Weyl Problem
BOUND STATES OF SCHROEDINGER OPERATORS AND THE WEYL PROBLEM 245
(27r) n Jr,.,Rn dpdx 0(a -(p2+V(x))
Nat (V) z
with 0(a) - 1, for a > 0
and - 0 fnr a < 0.
The
(1.18)
integration in (1.18) for
fixed x, is easy to do, namely 1
= Tn (V - a]_(x)n/2
dp
(1.19)
p2< a-V(x)
Thus, (1.18) yields Na(V)
= Cn f (V -
a)_(x)n/2
dx
(1.20)
.
While the chief purpose of this paper is to prove Theorems 1 and 2, quantities of no less interest are the moments of the nonpositive eigenvalues of H.
DEFINITION.
For y > 0 IY(V) - EIEi(V)IY
Y
r° IclY-1 Nada
(1.21)
10(V) is defined to he NO(V).
For n > 3 we can use
Corollary 5 and Fubini's theorem to obtain I0
Ialy-1(V
IY(V) < Y Ln fM dx
= Y Ln fM dx = L
y,n
_ a )_(x)n/2 da
da IaIY-1(V-(x) +
f°
)n/2
-V_(x)
JM V-(x)Y+n/2 dx
(1.22)
with
LY n = Ln r(y+1) r(l+n/2) r(l+y+n/2)-I There are several things to be said about (1.22).
(1.23)
Although it was derived
from Corollary S under the assumptions (1.13) and n > 3, it holds much more generally. For example it holds in ,n for n - 2, y > 0 and for n - 1, y > 1/2 provided a > 0.
This was first given in (9].
holds for n - 1 and Y - 1/2.
That was an error
In (9) it was stated that it ;
it is not known if (1.22)
holds for n - 1. y = 1/2 but it is known (9] that (1.22) does not hold for n - 1, Y < 1/2.
In section II
we shall briefly mention how to deduce (1.22).
The best constant LY n in (1.22) is not given by (1.23), as the foregoing remark already indicates.
If we use C
n
classical value of L
in place of LY n in (1.23) we have the
namely, Y ,n
LY n -
As in the case cof Ln,
(4v)-n/2
r(Y+1) r(l+y+n/2)-1
(1.24)
it is easy to prove that LY The classin cal constant LY n can also be "derived" from the semiclassical assumption as .
in (1.18), (1.19), namely
247
Proceedings of the Amer. Math. Soc. Symposia in Pure Math. 36, 241-252 (1980) ELLIOTT H. LIES
246
(2r)-n fRrRn
EIE1IY =
dpdx
2 I p
+ V(x)IY 0(-p2-V(x)).
(1.25)
If the p integration is done in (1.25), the result is (1.22) with 'Y
,n
An important question is when is
LY.n
C
-
L Y.n
for y large enough, depending on n.
It seems to be true that LY.n - LY.n
This is known to be true 19, 101 for n - 1 and Y ? 3/2. LY .n 0
In fact 1101, if
for some y0, then equality holds for all y > Y0. for 0 The case of primary physical interest is Y - 1, n - 3. where it is conLC
jectured (91 that equality holds.
If this were so, it would have important
consequences for physics and it is hoped that someone will be motivated to solve the problem.
We now turn to the proof of Theoremst and 2 in the next three sections.
U.
THE BIRMAN-SCHWINCER KERNEL As stated in Corollary 5 we can assume V(x)
<- 0.
Therefore write
V(x) - -U(x), U >- 0.
A useful device for studying the nonpositive eigenvalues of -A-U was discovered by Birman [111 and Schwinger (121.
If (-A-U)* - EV, E < 0, then
IEI)-1
4,
(-A +
(2.1)
UV+
Defining U1/2* - $ , and multiplying (2.1) by U1/2 , we have
(2.2)
0 - KIEI (U)O where Ke(U), for e > 0, is the positive
Birman-Schert.nger Kernel given explici-
tly by U(x)1/2(-A + e)-I(x. Y)U1/2(Y)
Ke(x. Y; U) -
(2.3)
.
What (2.2) says is that for every nonpositive eigenvalue, E, of -A-U, KIEI(U) has an eigenvalue 1. (3,8) for more details).
The converse is also easily seen to hold ( see
Ke(U) is to be thought of as an operator on L2
; we
will see that it is compact, when e > 0 at least, and U is in a suitable Lp space.
In addition to the advantage that the study of the E's reduces to the study of a compact operator, there is the following important fact: Since
(-A + 0-1 is operator monotone decreasing as a function of e, so is Ke Hence (with V - -U),
Na(V) - k101(U) a nwnber of eigenvalues of K1(U) 2 1.
(2.4)
(2.4) will be exploited in the following way. PROPOSITION 6.
forx21. 248
Let F : R+ + 1t+ be any function such that
F(x) ? 1
The Number of Bound States of One-Body Schrodinger Operators and the Weyl Problem
BOUND STATES OF SCHROEDINGER OPERATORS AND THE WEYL PROBLEM 247
Then
Tr F(Ke(U))
F(Z' (U)) i
(2.5)
> ke(U) - N -e(V) where Tr means trace and the le (U) are the eigenvalues of Ke(U). For example, consider lRn, n
<-
3, and F(x) = x2
.
Then
Tr Ke(U)2 = ff U(x) U(y) ((-A + e)-1(x - y)12 dxdy <
by Young's inequality.
IIUIIZ 11 (-A + e)-1(x)IIZ
(2.6)
The last factor is of the form (2.7)
II(-o + e)_1(x)112 = hne-2+n/2
This shows that Ke(U) is Hilbert-Schmidt when U C L2 and e > 0. h
n
- - but one can show that
higher power of K.M.
When n > 3,
K (U) is compact by considering the trace of a
e
cf. [9).
At this point we can derive the aforementioned bound, (1.22), on Iy(V).
If we use (2.5) and (2.6) and insert the latter in (1.21), the a integration will diverge.
The trick [9) is to use Proposition 4 with 28 = e = -a. Thus
N-e(V) < Tr Ke/2([V + e/2)-) <
II
(U - e/21+II2 hnle/2)
-2+n/2
(2.8)
Inserting (2.8) into (1.21), and doing the a integration first, we obtain ly(V) < 22-n/2 yhnfdx f0 da IajY-3+n/2[U(x) + a/2)2 2U(x)
Ly n fU(x)y+n/2dx
if 7
>
2 - n/2 and n
<-
(2.9)
3.
(2.9) can be extended to other values of n and Y (but with y > 1/2 for n = 1), cf.
There is no way, however, to make this method work with F(x) = xa when y = 0 and n 2 3.
Quiet reflection shows that if theorem 2 is to be provable
by this method then we need x-n/2 F(x) ed as x
m.
0 as x -+ 0 but x 1F(x) remains bound-
The tool we will use to bound TrF(K (U)) for such F's is the e
Wiener integral.
That is the subject of the next section, which is really the
main point of this paper.
M.
THE WIENER INTEGRAL. REPRESENTATION OF F(K
(II))
e
Let du x y.t be conditional Wiener measure on paths w(t) with w(')) = x and w(t) - y.
This measure gives a representation for C (cf. (3)) by fdux.y:t(w)
- C(x.y;t) = etA(x, Y).
(3.1)
C itself has the semi group property
249
Proceedings of the Amer. Math. Soc. Symposia in Pure Math. 36, 241-252 (1980)
ELLIOTT H. LIEB
248
fMG(x, y; t) G(y, z; a) dy - G(x, z; t+s).
(3.2)
The well known Feynman-Kac formula [131 is 1dux.Y;t(w)
e-t(-A + XU)(x,
exp [-aft U(w(s))ds1 -
(3.3)
U 2 0, multiply (3.3) by
e 2 0, A 2 0, and
Take
(Note the signs in (3.3).).
y).
U(x)1/2U(Y)1/2 exp (-et) and integrate with respect to A - U(x)1/2U(Y)1/2 N°dt e-etrdpx.Y:t(w) exp (-Aft
t.
The result is
U(w(s))ds)
. {Ul/2(-A + XU + e)-1 U1/2) (x, Y).
(3.4)
Now (-A +
e)-1
e)-l
- (-A + AU +
+ A(-A + XU + e)-1 U(-A + e)-
Multiplying (3.5) on both sides by
U1/2
(3.5)
we obtain
A - [K (U) [1 + AKe(U)}-1}(x, Y). (3.6) can be cast in a more general form.
If
(3.6)
g(x) - exp(-ax)
and
then
F(x) - x(1 + Ax)-I
F(x) - xf0 dy a-y g(xy)
(3.7)
and
U(x)1/2U(Y)1/2 (°Ddt a-et fdpx Y;t
(w) g(f
U(w(s))ds) (3.8)
F(Ke(U)) (x, Y).
Next we want to take the trace of both sides of (3.8). by setting
This is obtained
(To be precise, one must first take
and integrating.
y - x
U E C and then use a limiting argument.)
The point to notice is that the
x
dependence occurs through fMdx U(x) dux.x;t (w) - fM dx dux,x;t (w) U(w(0)). By the semigroup property of obtained (after the U(w(s))
for any
a-et
f'0 dt t-1
du
exp[-t(-A + AU)], however, the same result is
integration) if
0 5 s S t.
(3.9)
U(w(0))
in (3.9) is replaced by
Thus,
fMdx fdyx,x;t (w) f(ft U(w(s))ds) - Tr F(Ke(U)).
(3.10)
with f(x) - xg(x),
(3.11)
i.e.
F(x) - f0 dy
Now the relations between (3.8) and (3.10).
F,
f
and
y-1 a-Y f(xy) g
(3.12)
are linear, as are the relations
The latter therefore extend to a large class of f's.
Since
we are here not particularly interested in the operator version, (3.8), we will
250
The Number of Bound States of One-Body Schrodinger Operators and the Weyl Problem
BOUND STATES OF SCHROEDINGER OPERATORS AND THE WEYL PROBLEM 249
concentrate on the trace version (3.10). The g's of the form
exp(-Xx)
norm dense in the continuous functions which vanish at infinity. using the semigroup property of
are
Furthermore,
exp [-t(-A + U)], one can see explicitly that
(3.8) and (3.10) hold for f's of the form
xk exp(-Ax), with
k
a positive
By a monotone convergence arguement one arrives at
integer.
THEOREM 7.
Let
f
be a nonnegative lower semicontinuous function on
[0, -) satisfying:
(i)
f(O) = 0
(ii) xp.f(x) - 0
as
x + m for some
p<
U>-0 and UcLP+Lq with p=n/2 (n23), p>l (n=2), p-1
Let
(n = 1)
and p < q < m
Then (3.10) holds, with
sense that both sides may be
F
given by (3.12), in the
+
The reader is referred to (81 for details.
f's
Obviously the class of
in Theorem 7 is not the largest possible, but it is more than adequate for our intended application.
The remark that allows us to bound the left side of (3.10) is the following.
THEOREM 8. that
f
Suppose that
is also convex.
Tr F(Ke(U)) < Ip dt t-1 a PROOF.
satisfies the conditions of Theorem 7 and
f
Then et
fM dx G(x, x; t) f(tU(x)).
By Jensen's inequality, for any fixed path
t
(3.13)
w(t),
f(f0 U(w(s)) ds) = f(f0 (t-Ids) t U((s))) S fo (t-Ids) f(tU(w(s))). By the same remark as that preceding (3.10), fm dx fdux,x;t(w) h(w(s)) independent of (3.11).
s
for any fixed function
h.
is
Inserting this in (3.10) gives
QED.
APPLICATIONS OF THEOREM 8
IV.
PROOF OF THEOREM 1:
We use Proposition 3 and Proposition 6.
sen to be of the following form for some
0<x
- b(x - a),
a < x.
F
given by (3.12) is monotone, and the condition that
a
and
(1.7)
E1
(4.1)
F(x)
1
is that
are related by 1 = h(a, b) - bfa dy(1 -
where
is cho-
0 < a < b
f (x) - 0,
b
f
is the exponential integral.
a/y)e-'
= be -a - abEI(a)
(4.2)
Assuming (1.10), (1.9) will be proved;
follows from the special assumption
Bn - 0
in what follows.
251
Proceedings of the Amer. Math. Soc. Symposia in Pure Math. 36, 241-252 (1980) ELLIOTT H. LIEB
250
a - -e
In Proposition 3 write U - (e + ).)y
f(tU(x)) - 0
A) < (3.13)
so that [y(Sl,
with
The x integration can be done last in (3.13) in which case
.
For x c f2 we change variables to
x t 4.
if
t + t/(e + A).
Thus, A)n/2t-n/2
blOl fa dt(l - a/t)(An(e +
+ Bn)exp[-et/(e + A)]
(4.3)
> W(Sl, A)
when n 2 3
Bn - 0.
and
Proposition 6 is used again and the proof parallels
PROOF OF THEOREM 2:
to
t
-.
tU(x)-1
We change variables in (3.13)
(4.1) and (4.2) are assumed.
if
U(x) # 0.
e - 0
If
Ln - An JO dt
the result is (1.13) with
t-1-n/2f(t)
4Anbal-n/2[n(n-2)]-1
-
REMARKS:
(i)
If
e - ca, c > 0 and
But for all cases we can choose QED.
thereby prove Theorem 1.
that of Theorem 1.
This will work only
e = 0.
The simplest choice for e, which is arbitrary, is
QED.
.
(4.4)
then the only estimate we have for
a - -e # 0
is contained in Corollary 5, which is valid only for one could try to estimate (3.13) directly with
n
>-
3.
Na(V)
Alternatively,
e 0 0, but this is messy.
As
stated earlier, no inequality of the form (1.16) holds for all a, V when n - 1 or 2.
But recently Ito [14] has bounded (3.13) when
He uses the fact that bound for
Na(V)
(ii) If
This estimate for
f(x) < bx
in terms of
Dn
n > 3
for
and
x > a
and
n = 2.
and obtains a complicated upper
and
lIV-llnV-ll/2112.
B. - 0
we can choose
11vJ12
e # 0
is, of course, the same as
Ln
e = 0
in (4.3).
given by (4.4).
As an illustration of how good our bound is let us consider the case of
R3, where
A3 = (4a)-3/2.
E1(.25) - 1.0443
and
a - 0.25
We choose
b - 1.9315
in (4.1) and find that
according to (4.2).
D3 - L3 - 0.1156
This value of D3 can be used in (1.7).
Using (4.4), (4.5)
.
When compared with
is supposed to be the sharp constant, it is not very good. D3 can be improved by using (4.3) with
C3 = 0.0169, which The estimate for
e - cA, c > 0.
If, however, the same number, L3, is used in (1.13) the result is quite good.
As already stated, the best
L3 > C3.
In fact, by an explicit example,
L3 2 (3n)-3/2P(3)/P(3/2) - 0.0780,
252
(4.6)
The Number of Bound States of One-Body Schrodinger Operators and the Weyl Problem
BOUND STATES OF SCHROEDINGER OPERATORS AND THE WEYL PROBLEM 251 cf. (9), eqn. (4.24).
It is conjectured that the right side of (4.6) is, in
fact, the sharp constant in (1.13) for R3.
In any case our result, (4.5), is
off by at most 49%.
As stated in Section 1, a quantity of physical interest is I1, the sum of
the absolute values of the eigenvalues, in R3.
Using the bound (1.22),
(1.23), together with (4.5), we have L1, 3
(2/5)L3
.04624
(4.7)
This result was announced in [15].
BIBLIOGRAPHY
H. Weyl, "Das asymptotische Verteilungsgesetz der Eigenwerte Linearer 1. partieller Differentialgleichungen", Math. Ann. 71 (1911), 441-469. M. Kac, "Can one hear the shape of a drum?", Slaught Memorial Papers, 2. no. 11, Amer. Math. Monthly 73 (1966), no. 4, part 11, 1-23. M. Reed and B. Simon, Methods of Modern Mathematical Physics, Acad. 3. Press, N. Y., 1978. C. V. Rosenbljum, "Distribution of the discrete spectrum of singular 4. differential operators", Dokl. Aka. Nauk SSSR, 202 (1972), 1012-1015 (MR 45 The details are given in "Distribution of the discrete spectrum of 84216). singular differential operators", Izv. Vyss. Ucebn. Zaved. Matematika 164 [English trans. Sov. Math. (Iz. VUZ) 20 (1976), 63-71.) (1976), 75-86.
B. Simon, "Weak trace ideals and the number of bound states of 5. Schroedinger operators", Trans. Amer. Math. Soc. 224 (1976), 367-380. M. Cwikel, "Weak type estimates for singular values and the nunber of 6. bound states of Schroedinger operators", Ann. Math. 106 (1977), 93-100. 7. E. Lieb, "Bounds on the eigenvalues of the Laplace and Schroedinger operators", Bull. Amer. Math. Soc. 82 (1976), 751-753.
B. Simon, Functional Integration and Quantum Physics, Academic Press, 8. N. Y., to appear 1979. E. Lieb and W. Thirring, "Inequalities for the moments of the 9. eigenvalues of the Schroedinger equation and their relation to Sobolev inequalities", in Studies in Mathematical Physics: Essays in Honor of Valentine Bargmann (E. Lieb, B. Simon and A. Wightman eds.), Princeton Univ. Press, Princeton, N. J., 1976. These ideas were first announced in "Bound for the kinetic energy of fermions which proves the stability of matter", Phys. Rev. Lett. 35 (1975), 687-689, Errata 35 (1975), 1116. 10. M. Aizenman and E. Lieb, "On semi-classical bounds for eigenvalues of Schroedinger operators", Phys. Lett. 66A (1978), 427-429. 11. M. Birman, "The spectrum of singular boundary problems", Math. Sb. 55 (Amer. Math. Soc. Trans. 53 (1966), 23-80). (1961), 124-174. 12. J. Schwinger, "On the bound states of a given potential", Proc. Nat. Acad. Sci. U.S.A. 47 (1961), 122-129.
253
Proceedings of the Amer. Math. Soc. Symposia in Pure Math. 36, 241-252 (1980) ELLIOTT H. LIEB
252
13. M. Kac, "On some connections between probability theory and differential and integral equations". Proceedings of the Second Berkeley Symposium on Mathematical Statistics and Probability, Univ. of Calif. Press, Berkeley, 1951, 189-215.
14. K. R. Ito, "Estimation of the functional determinants in quantum field theories", Res. Inst. for Math. Sci., Kyoto Univ. (1979), preprint. 15. E. Lieb, "The stability of matter", Rev. Mod. Phys. 48 (1976), 553-569. 16. V. Glaser, H. Grosse and A. Martin, "Bounds on the number of eigenvalues of the Schroedinger operator", Comcun. Math. Phys. 59 (1978), 197-212.
DEPARTMENTS OF MATHEMATICS AND. PHYSICS PRINCETON UNIVERSITY JADWIN HALL P.O.BOX 708 PRINCETON, N. J. 08544
254
With S. Oxford in Int. J. Quant. Chem. 19, 427-439 (1981)
Improved Lower Bound on the Indirect Coulomb Energy* ELLIOTT H. LIEB AND STEPHEN OXFORD Departments of Mathematics and Physics, Princeton University, Princeton, New Jersey 08544, U.S.A.
Abstnet For a Coulomb system of particles of charge e, it has previously been shown that the indirect part of
the repulsive Coulomb energy (exchange plus correlation energy) has a lower bound of the form -Ce21s f p(x) Vs dx, where p is the single particle charge density. Here we lower the constant C from the 8.52 previously given to 1.68. We also show that the best possible C is greater than 1.23.
1. Introduction
In the study of quantum Coulomb systems of charged particles (atoms, molecules, and solids), it is frequently desirable to estimate various energies in terms of the (diagonal) single particle charge density p,,(x) belonging to a given state tli of N particles. We will be concerned with the repulsive Coulomb energy N
1=(GI L eie1Ix,-xillo),
(1)
,<4
where e, > 0 are the particle charges (by a trivial change we could assume all e; <0). 41 is any normalized state, not necessarily an eigenstate of any given Hamiltonian; it could also be a density matrix. Our goal is to find a lower bound to I4. [See note added in proof.] We define
D4=D(p*,p4)
(2)
to be the direct part of l4, where, in general, we have
D(! g) 21 J Jf(x)g(y)Ix-yl-' dxdy. The density p4 is the sum defined by
P4(x)=e;
(3)
of the N individual single particle charge densities
JI0(xt,...,xi-1, x, x;+1,...,xN,a1,...,aN)I2dx,.
(4)
Work partially supported by U.S. National Science Foundation Grant No. PHY-7825390 A01.
255
With S. Oxford in Int. J. Quant. Chem. 19, 427-439 (1981) LIEB AND OXFORD
428
Here the a, are any quantum numbers (e.g., spin) and dz, means integration over all variables except xi. The indirect part E4 of 14 is defined by (5)
E,, + D,,, = 4.
E4, includes all "exchange" and "correlation" effects. It was shown by one of us [1], that for all #P and C = 8.52, we have
E4,?-Cez/3 Jp,
(x)4/3dx.
(6)
A more complicated bound was given when the e; are not identical. The main result of this paper is given in Section 2. We show that N
dx]1/2r
eip'(x)1413
Ey?-C[J((L'
p4(X)413 dxl
1/2
If
(7)
even in the case that the e, may be distinct, with C = 1.68. With this improved value of C, the bound for E,1, may have computational value as well as theoretical value. Inequality (7) reduces to the form of Eq. (6) when all e; = e. Our proof differs from that in Ref. I chiefly in avoiding the use of the Hardy-Littlewood maximal function. We also show in Section 3 that the best (i.e., smallest) possible C, in Eq. (6), must be greater than 1.23 by explicitly calculating E for a two-particle wave function q/.
It should be noted that Eq. (6) is a special case of the lower bound for N-particle states I/f, EJ ?
p e (Sp .6)/(3p-3)
P1/(1p-3)N (3p 4)/(3p P)
(8P4
31,
1
which is a consequence of Thomas-Fermi theory for p>3 (see Refs. 1-3). Holder's inequality applied to Eq. (6) gives Eq. (8) for all p 3 with the same C as in Eq. (6). However, the best Cp in Eq. (8) will in general be smaller than this C and it would be interesting to have an estimate for it. [In Ref. 1 it was incorrectly stated that Thomas-Fermi theory gives Eq. (8) for p > j. This is wrong because Thomas-Fermi theory is bounded below only for p > z. Equation (8), for p = 3, was first given in Ref. 3, where it was used to prove the stability of matter.] The bound, equations (6) or (7), can be examined in a number of ways. One might try to find a constant C in Eq. (6) only for 0 symmetric or antisymmetric, or only for 0 with a particular spin value. However any such restriction on ./i cannot improve the constant in Eq. (6). Let all e, = e. Suppose we take an
arbitrary 0(x,, ... , xN; a(, ... , aN), where a,, ... , a,v are arbitrary quantum numbers. [.et f4(X1,
256
.. , XN) ' E 1'1'(x1..... XN, aI. ..
,
(X,N)I2,
Improved Lower Bound on the Indirect Coulomb Energy INDIRECT COULOMB ENERGY
429
and let the symmetrized version of fa be given as
F(x,,.
, XN) =
I
L fJ,(xP1,
N. PEsN
, xPN)
We can define a symmetric 4 such that E,,,s = E,,, Ir,s = Ir, and p,,s(x) = p, (x). We
merely take iIis(x1, ... , xN) = F(x1.... , This shows that C cannot be improved by excluding bosonic 41. In a similar way, we define an antisymXN)112.
metric '°, t° = F(x,, ... ,
xN)112B(x...... XN ),
where 0 is any antisymmetric function which takes on the values ± 1 except on a set
of measure zero. We see that we cannot improve the constant by excluding r' with Fermi statistics. A similar construction shows that restrictions on the spin quantum number cannot improve C. It will be noted that the right-hand side of Eq. (6) is of the form given by Dirac [4] to approximate the exchange energy. There is, however, a difference between the two if spin is taken into account. The Dirac approximation is -Cq-113
Ey =
f
p.(X)413 dx,
(9)
with C = 0.93 and q is the number of spin states (q = 2 for electrons). Dirac computed Eq. (9) using a plane wave determinant for 0. This determinant depends upon q. In view of Eq. (9) one might infer that a good lower bound to Ep
should have a q113 factor. This, as we have just noted above, is not the case. Another way to say this is the following: a "diagonal" operator, such as the Coulomb repulsion cannot distinguish the spin of a particle. To "see" the spin it is necessary to examine an off-diagonal operator such as the kinetic energy T,d. A lower bound for T,, does exist (3] and it does have a factor q-213. In order to have a lower bound for E,,, that measures q it would be necessary to use an expectation value such as T,,, that is "off -diagonal." A useful bound of this kind might exist, but
a bound such as ours that involves only the "diagonal" quantity p,y has no q dependence. It is true however, that the constant in Eq. (6) or (7) can be improved if we specify the number of particles. Consider the case of equal charges. Define CN to be the best constant for Eq. (6) when we consider only N-particle states 1(i, i.e., CN=
sup
N pan,d
,4
(- J p, (x)413 dx)
Ey.
In Appendix A we compute C, = 1.092, following a treatment by Gadre, Bartolotti, and Handy [5]. One notes that C, is less than the constant 1.23 computed for the variational two-particle wave function of Section 3, which is itself less than C2. In Appendix B, we show that this is a general phenomenon, i.e., CN <_ CN 1. Presumably, one might take advantage of this if one were only dealing with a fixed low number of particles. However, we have no method available for computing an upper bound for CN, N ? 2, except 1.68, which holds for all N.
257
With S. Oxford in Int. J. Quant. Chem. 19, 427-439 (1981)
LIEB AND OXFORD
430
Inequality (7) might be improved upon in another way in the case of unequal charges. One might try to prove that a universal constant C existed such that the following is true (for fixed 0
N
Er,
- C [ r ( Y_I
e;'
I -a
o
dx]
2 poi (x))
p(x)4/3 dx]
[J
(10)
It is an open question whether a bound exists when a = 1 in Eq. (10), called the symmetric form. Holder's inequality implies that for Eq. (10), (right-hand side, a = 1) a- (right-hand side, a s 1); thus the symmetric form gives the best lower bound. In Ref. I a bound with constant 8.52 was proved for a =', and in this paper we sacrifice even more "symmetry" in the bound to show that Eq. (10) holds for a = 2 with the improved constant 1.68. In Ref. 5, the comment was made that no one has yet produced any upper bound for E*. The following simple remark is relevant. There certainly cannot be any upper bound of form C f p(x)4/3 dx. The reason is that E,I,/f p(x)4/3 dx can be made arbitrarily positive simply by taking a G for two particles of the form 4, =ALk -(XI-x2)]exp(-1x112- 1x212-Aixl-x212,
where k is some fixed vector. As A tends to infinity, Ei,, will tend to +oo while f py4/3dx will remain bounded. This Vi is antisymmetric. 2. A Lower Bound for E,1,
We will use an argument similar to that given in Ref. I to derive Eq. (7) for C = 1.68.
We first fix charges e 1 > 0, ... , eN z 0. Let f (x 1, ... , xN) be the particle density associated with an N-particle wave function t/r, as given in Section 1. Let p,,
be the associated single-particle charge density, equation (4), and let x1,. .. , xN be distinct but otherwise arbitrary points in R3. We take µ(y) to be some function satisfying the following: (i)µ is non-negative; (ii) µ is spherically symmetric about the origin and µ (x) = 0 if Ix I > 1; and (iii) f µ (y) dy = 1. Let A be a positive constant to be determined later. We now define a function A.(Y)A3P.,(x)IL(Ap4(x)1/3(y
F<<(- ):
-x)),
(11)
if p,r(x) > 0, and µ, (y) w 0 if p,r(x) = 0. We see that µ, is a non-negative function which satisfies (i) its integral (with respect to y) is I if p,y(x) > 0; (ii) it is spherically
symmetric about x; (iii) µ,(y)=0 if Iy - xI>A
'p,(X)-113.
We observe that Lemma 1 of Ref. I may be applied to this choice of µ. Namely, we prove
Lemma 1: N
N
eejjx, -x1j ' -_ -D(P#, p,)+2 E D(p,, eiix,)- E D(eiµi,, e,µ=,). i-I i -I Isi<jsN E
(12)
258
Improved Lower Bound on the Indirect Coulomb Energy INDIRECT COULOMB ENERGY
431
Proof. It is a well-known fact [1] that the potential io generated by a non-negative spherically symmetric charge distribution of total charge 1 satisfies
ti(x) s 1 /jxj. In particular, taking x, to be the origin, f µ,,(y)jx - yj_' dy :s ix -x,j- '. Hence D(e,p.r,, erg,,) = iee! J A., MIX - yj-'µ,,(x) dx dy (13)
sJee; Jix -x;jie.,,(x)dxske,eaIx,-x;j'. We now observe that
D(P* -
N
N
e,Ez,,, po -
(14)
e,K=) L. 0
by the positive definiteness of the Coulomb kernel. Expanding the left-hand side of Eq. (14) and rearranging, one has that N
N
2 ,
,-I D(e,µ=ne,µx,)
(15)
The lemma follows by applying Eq. (13) to the left-hand side of Eq. (15). Now let 8,, be a point charge distribution of charge one centered at x;. Adding and subtracting terms on the right-hand side of inequality (12), we get F. e,e;jx,-x,j-'>--D(P#,p,r,)+2 E D(P,,.,e,8,,) 1-I
,<j
- (2
N
N
(16)
D(P*, e8,, - e,µ:) + E D (eµ,,, e,k.))
We now integrate Eq. (16) against j*(x1, ... , xN), whence N
I*
D(P,n P)-(E f 2D(Pa, 8,, -/a,,)P'r(x,) dx,
,-I
+ E f D(IL, µ=,)e,Py(x,) dx,)
(17)
We wish to find an upper bound of the appropriate kind to the expression in large parentheses in Eq. (17). Let us denote the first sum in large parentheses as (*) and the second as (**). We rewrite (*) as follows: (*) = J P*(x)2D(P*, S. -µ.x) dx = f f Pf(Y)FA(P#(x), Ix -Y1) dx dy,
(18)
where we define
FA(a, r)=[ar-'-Aa°".(Aa'/3r)],
(19)
and 0 is the potential generated by our fixed µ. Hereafter we require that k be bounded. In this case, F, (a, r), considered as a function of a on (0, oo), satisfies the
259
With S. Oxford in Int. J. Quant. Chem. 19, 427-439 (1981)
LIEB AND OXFORD
432
following for r > 0; (i) it is continuously differentiable; (ii) F, (0, r) = 0 and r) =0 if Aa'/3r> 1. These properties follow from the continuity and differentiability of 0 and the relation 46(x)= 1/Ixl for lxl> 1. For a >O, let
Xa(x)=9[p*(x)-a], where 9(t)=0 if ts0, and 0(t)=1 if t>0. X. is the characteristic function of the set p4(x)>a. By Fubini's theorem and the fundamental theorem of calculus, one has that
Jo da JXa(x)--F,,(a,r)dx= Jdx Jo
aaF,(a,r)da= JdxF.,(p(x),r), (20)
and thus M
.0
(*)=Jo Jo dadb JdxdYXa(x)Xb(Y)aaFA(a,Ix-y1) where we have used the representation p(y) fo dbXb(Y) We bound (*) as follows: Let (y). = y if y? 0, and (y), = 0 if y:5 0. Then
Jo dadb JdxdYXa(x)Xb(Y)\aaFA(a,Ix-Yl)
sJ
ayb
+J6
dadb J dxdyX.(x)(a F,(a,lx-yl))
(21)
as
dadb JdxdYxb(Y)(aFA(a,Ix-YI))+
By scaling properties of (a/aa)F,(a, Ix - yl), one has that J(aaF,(a,Ix-YI))+dx=J(aaF,,(a,Ix-YI))
dy=A-2Ka-2/3
(22)
where K = f [(a/aa)F1(1, lz l)]. dz and K only depends on the original choice of A. We, therefore, have that ao
co
a
(*):5A-2KJdx(J
0
b
Xb(x)dbj a-2/3da)
daXa(x)a-2/3J db+J 0
0
0
(23)
4A-2K Jdx JXa(x)a'/3da=3A-2K J py/3(x)dx, a
where we have used the representation p?3 (x) = (4/3)fo a113Xa(x) da. The second sum (**) in the large parentheses of (17) can be written
(**) = E J D(µ., µx)e,P*(x) dx =AD(k, A)
i-1 fr
N
260
(24)
erPm(x)P,(x)'/3 dx
J 4/3
l3/4r dxJ
1/4 py(x)4/3dx]
LJ
(25)
Improved Lower Bound on the Indirect Coulomb Energy INDIRECT COULOMB ENERGY
433
Equation (24) follows from simple scaling and Eq. (25) is the Holder inequality. Optimizing Eqs. (24) and (25) with respect to A yields 4/3
N
rfr
1/2
dx,
(E e;p'(x))
1/2
r(
p(x)4/3dx]
LJ
(26)
A variational argument shows that the optimum choice of µ would be the uniform ball if [(a/aa )F, (a, r)]+ were replaced by (a/aa)F,, (a, r) [in which case the constant in Eq. (26) would be 1.45]. However, trial and error indicates this choice is also
approximately best with the cutoff. We find that [aF,(1, r)/aa]+=aF,(1, r)/aa if and only if r :s R with R = (5"' -1)/2. Then K = 0.6489 and D(µ,µ) = 5. The constant in Eq. (26) is then 1.68. Thus we reach the conclusion that N
4/3
E, > -1.68[ J (E e;py(x))
1/2
1/2
{Jp4(x)4'3dx]
dx]
.
(27)
3. A Lower Bound for C2 We now exhibit a lower bound to C2 [and thus to the best possible C in Eq. (6)] which is greater than C1. We choose a singular O(x, y), and take e, = e2 = 1. Let t = Ix I and s =1yI for x,
y E R3, and let h and e he unit vectors e = x/IxI and h = y/I yl. We define 1l12[(t, e), (s, h)]s f =(15/4rr2)S(1- t -s)S(e
h + 1)9(1-t)9(1-s).
We check the following:
J f[(I,e),(s,h)]s2 dsdh=
15
'
4` J S(1
z t - s)sdx
J2r
J 8(cos '+ 1)sinlidiIidb9(1-t)
x
= 15 (1-()29(1-t)'(t). We have used spherical coordinates to evaluate the above integral with the north pole in the (fixed) "e" direction. Similarly, the s marginal is p2(s) = p'(s).
One checks that f p' =I and hence f is properly normalized. We have that p(t) = 2p1(t). Trivially, 1, = I since the particles are always one unit apart. We have that
(-(1 - t)8j3t2 I5)-"'f01 J p(x)413 dx = 4zr
dt = 2.084.
+r
By Newton's theorem D(p, p) _ (4 7r)2 J p(t)tJ p(s)s2 = 3.572. 0
0
261
With S. Oxford in Int. J. Quant. Chem. 19, 427-439 (1981) 434
LIES AND OXFORD
The result is that
C2_>-E*/ J p#,(x)4'3dx = 1.234.
Appendix A: Evaluation of C,
For one particle I,, =0. Therefore C1 is the maximum of H(p) given in Eq. (28). We prove the following theorem and then compute C1. This was done in Ref.
5. The purpose of this appendix is only to give a rigorous justification of the calculation done in Ref. 5. Theorem. There exists a symmetric decreasing function p which maximizes the functional
H(P) _ (J p(x)413 dx)
1
D(p, p)
(28)
over the set A ={p(x)p(x)O, p E L413(R3), f p(x) dx = 1). Proof. For am p at 0 and p E L`13(R3) (1 L'(R3), let (p),, be a scaled version of p, i.e.,(p) (x) = A p (Ax). It is simple to check that f (P )a (x) dx = f p (x) dx,
J (P)" (X)413 dx =A J P(x)4/3 dx, and
H[(P)a] = H(p).
(29)
Let p , E A, j = 1, 2, ... be such that lime H(pi) = H(p). This supremum may a priori be infinite, but we will see that it is finite. By scaling p, and using Eq. (29), we may assume henceforth that f p;(x)4'3 dx = 1. By the Riesz inequality for
symmetric decreasing rearrangements [6],t we have H(p*)>-H(p), where p* is the symmetric decreasing rearrangement of p. One also has that f p*(x) dx = JP(X) dx, f p*(x)4'3 dx =f p(x)'i3 dx; therefore, by replacing pi by p" if necessary, we may also assume that the pi are symmetrically decreasing. We now use an idea used in Ref. 7. By the symmetric decreasing property of pi, we have [writing pi(x)=pi(jx1)]
4a
p;(x)dxs Jpi(x)ds=1.
'3 R p1(R)s J -ISR
t The 3-dimensional proof can be found in Brascamp, Lieb, and Luttinger, J. Funct. Anal. 17, 227 (1974).
262
Improved Lower Bound on the Indirect Coulomb Energy INDIRECT COULOMB ENERGY
435
Hence
P;(R)skI/R3,all j.
(30)
Similarly,
43 R3p4i3(R)s j
Pi(x)4n
dz = 1,
Pi(R) s k2/R914, all j.
(31)
We define f(R) - min (k, /R3, k2/R9'4). Since the pi are symmetric decreasing and uniformly bounded by f (which is finite except at 0), by a variant of Helley's theorem [8], some subsequence of the pi (which we continue to denote by pi) converges pointwise almost everywhere to some symmetric decreasing p(x) and p(x) s f(x). We will see that p(x) # 0. We now show that the p we have found satisfies the conditions of the theorem.
By calculation D (f, f) < m. We therefore apply the dominated convergence theorem to conclude that
limD(pi,PI)=D(p,P)
(32)
0<sup H(p)=limD(pi,pi)
(33)
i
In particular, A
Furthermore, by Fatou's lemma we have that j p(x)4"' dx s lim f pi(x)4'3 dx = 1,
(34)
j p(x) dx s lim j pi(x) dx = 1.
(35)
Therefore by Eqs. (32)-(34), H(p)>_ 5UPPEA H(p). By Eq. (35) we can multiply p by a scalar A z I so that f Ap(x) dx = 1. By definition of A, Ap e A and
H(Ap)=A2'3H(P)>_H(P)zsupH(P) peA
(36)
It must be that the inequalities in Eq. (36) are actually equalities and thus that on that A = 1. We therefore have that p belongs to A and maximizes set.
We shall now show that the constant C, can be calculated. By usual variational arguments [7], one knows that the optimizing p satisfies the following variational equations: '0(x)_4CIP(x)I13+A
0
ifp(x)>0,
(37)
s0
if p (x) = 0,
(38)
263
With S. Oxford in Int. J. Quant. Chem. 19, 427-439 (1981)
436
LIEB AND OXFORD
where fi(x) is the potential generated by p,
d6(x)= J p(y)Ix-yj-' dy,
(39)
and where A is a real-valued Lagrange multiplier. We first note that Eqs. (37) and (38) imply that p is compactly supported. If
not, then one must have that p(x)>0 for all x by the symmetric decreasing property of p. By letting lxl i oo one has that A = 0, since p and-O tend to zero in
Eq. (37). We then would have by Eq. (37), p(x)=(constant)c6(x)3, where the constant is positive. For sufficiently large lxj, we see that p(x) ;2(constant)Ix This implies that p is not integrable, contradicting the fact that f p(x) dx = 1. Let ro be the distance at which p first vanishes. We now apply the Laplacian to Eq. (39) and use Eq. (37),
Ix1`ro, X ? r,.
ll0,
(40)
Let f(r) = (3n/C,)312[.0 (X) +A ]/4a. We rewrite Eq. (40) in spherical coordinates
d
_
r
dr2rf(r)=-f(r)3,
r
(41)
I
Equations (40) and (41) hold in the distributional sense. We now argue that f(r) is continuously differentiable and that f'(0) = 0. This will also imply that Eq. (41) is supplemented by f (ro) = 0. In spherical coordinates, one can write ,o
fi(r)=
4ar-' Ju p(s)s2ds+41r r
J,p(s)sds,
rsro,
(42)
r -- ro.
We apply Holder's inequality to the first integral and use the fact that p E L°'' 3/4
r
J p(s)s2 ds 0
s (J,p(s)4/3s2 ds) 0
,/4
(J s2 ds)
s (constant)r3/4.
n
This and a similar inequality satisfied by the second integral imply that fi(r) = O(r '/4) near the origin. By Eq. (37) one has that p(r) = 0(r-3/4) near r = 0. This in turn implies that 0 is bounded at the origin, and therefore by Eq. (37), p is also bounded at the origin. Since 0 is the potential of a bounded, compactly supported charge distribution, it is also continuous, hence p is continuous by Eqs. (37) and (38). One now can see that m (hence f) is C' by examining Eq. (42). Since p is continuous and bounded, the first term of Eq. (42) is of the form r-'g(r), where g(r) is continuously differentiable for r >0, g(r) = 0(r3) and g'(r) = 0(r) near the origin. Hence the first term is continuously differentiable for r - 0, and has vanishing derivative at r = 0. The preceding statement is true of the second term in
264
Improved Lower Bound on the Indirect Coulomb Energy
INDIRECT COULOMB ENERGY
437
Eq. (42) by inspection. Thus A(r) is continuously differentiable for r>0 and 0'(0) = 0. Equation (41) holds in the strong sense because its right-hand side is C'.
As first noted by Gadre, Bartolotti, and Handy [5], Eq. (41) is the Emden equation of order 3. One may rescale p(x)-. a3p(ax) to ensure that f(0) = 1. The two conditions f (0) = 1 and f'(0) = 0 uniquely determine the solution of the ordinary differential equation (41).
If ro is the first zero of the solution, we have that p(r) = 0 if r>_ ro and p(r)=(3a) 312C1/zf(r)3 if rsro. In Ref. 5 it was noted that this equation determines the constant CI. Namely, we have that 1=41r
Jrzf(r)3dr
J
p(r)rzdr=4,3-3/z17,-1;zC'/z
o
_ -4 3
0
3r2,7 uzC
1n
O°
r[rf(r)]° dr
0
3nrz
_ -4 3
I
f(ro)
(43)
Emden functions are tabulated [9]. We find that r o = 6.89684, f(r0) = -0.04243. Equation (43) then gives C, = 1.092. Appendix B: Monotonicity of CN
We show that CN s C,.,,. ,, where CN is defined in Section 1 as the best constant in Eq. (6) for an N-particle state. We consider the case e; = e. Let e > 0 be arbitrary but fixed. We let fN (x,, ... , .N) be an N-particle density
which vanishes for Jx,j> L for I <- i s N, where L is some finite number, and furthermore, let fN have the property that -1
-(e 213
JP/N(x)4/3dx)
E!N?CN-E.
(44)
A simple approximation argument using dominated convergence shows that L and fv can be found satisfying Eq. (44). Let x0 E R3 be chosen such that Ixol > L +2R, where R will be determined later. We define a one-particle density f,(x) _ (31rR3)-'6(R - Ix -xoi) and we also define the (N + 1)-particle density fv.I(xI, ... , XN.I) = fN(XI, ... , xN)f,(xN+,). One sees that pfN ,(x) =pr,(x)+ef,(x). Since PN and fI are never simultaneously nonzero, we have e
r 1 p/N(x)4/3 2/3 J P/N.,(x)4/1 dx = e2/31\J dX + e4/3J f1(x)4/3 dx) /!
= e2/3 r J
PfN(x)4/3
dx +
(45)
ez(3/4rr)'/3 R
265
With S. Oxford in Int. J. Quant. Chem. 19, 427-439 (1981)
438
LIES AND OXFORD
We also have that
IIN+e2 LfN(x1, ,xN)f1(xN+1)Ix(-XN+II 1dxl,.. ,dxN+l J CCN
i- I
(46)
s If, +e2N/R, by the definition of ft. The evident inequality DIN ?DIN together with Eq. (46) implies that EIN,,, <EIN +e2N/R. This and Eq. (45) imply that CN+
-
-(e 2/3 r
P/N.,(X)1/3
J
1
dx
1
EIN., e2N
e2(3/47r)1/31-1
(e2/3
J PIN
(X)413
dx +
)
R
(
(E,N + R) .
(47)
We now choose R so large that the right-most term in Eq. (47) is greater or equal to
-(e2/3
PIN(X)4/3 dx)
-I
J
EIN -e.
Recalling Eq. (44), we have the result that CN+1 z CN - 2e. Since e was arbitrary, CN+1 ? CN.
In the case of distinct ei's, one may define (for some fixed a, 0 < a s 1) 4/3
N
CN(el,
Sup
, eN) =
X
-LJ( e, /2ap/N(x)1
(J PIN (x)413 dx)
dx] 11
EIN.
A similar argument shows that these constants also increase,
i.e.,
CN(el, .. , eN) -- CN+t(e1, .. , eN, eN+1), where eN+1 ? 0 is arbitrary. Of course Section 2 shows that CN(el, ... , eN) < 1.68 for all N and e; when a = 2. Note added in proof:
In the text we proved the inequalities, Eqs. (6) and (7), when 0 is a wave function (pure state), and remarked that the inequalities also hold for a density
matrix. To prove this, note that any density matrix, µ, can be written as µ = Ea kYa > < 0a. In the definition, Eq. (4), simply regard f3 as just one more quantum number to sum over-on the same footing as the a's. The rest of the proof is then the same as in the pure state case. Bibliography [1] E. H. Lieb, Phys. Lett. 70A, 444 (1979). [2] E. H. Lieb, Rev. Mod. Phys. 48, 553 (1976).
[3] E. H. Lieb and W. E. Thirring, Phys. Rev. Lett. 35, 687 (1975); 35, 1116 (1975) (errata). [4] P. A. M. Dirac, Proc. Cambridge Philos. Soc. 26, 376 (1930).
266
Improved Lower Bound on the Indirect Coulomb Energy
INDIRECT COULOMB ENERGY
439
[5] S. R. Gadre. L. J. Bartolotti, and N. C. Handy. J. Chem. Phys. 72, 1034 (1980). (61 F. Riesz, J. London Math. Soc. S, 162 (1930). [71 E. H. Licb, Stud. Appl. Math. 57, 93 (1977).
[8] W. Feller, An Introduction to Probability Theory and its Applications, (Wiley, New York, 1966), Vol. 2, p. 261. [9] British Association for the Advancement of Science Mathematical Tables, (Office of the British Assoc.. Burlington House, London, 1932) Vol. 2.
Received June 10, 1980. Accepted for publication September 18, 1980
267
Int. J. Quant. Chem. 24, 243-277 (1983) INTERNATIONAL JOURNAL OF QUANTUM CHEMISTRY, VOL XXIV, 243-277 09H3)
Density Functionals for Coulomb Systems ELLIOTT H. LIEB Departments of Mathematics and Physics, Princeton University, P.O.B. 708, Princeton. New Jersey 08544, U.S.A.
Abstract
This paper has three aims: (i) To discuss some of the mathematical connections between N-particle wave functions 4, and their single-particle densities p(x). (ii) To establish some of the mathematical
underpinnings of "universal density functional" theory for the ground state energy as begun by Hohenberg and Kohn. We show that the HK functional is not defined for all p and we present several ways around this difficulty. Several less obvious problems remain, however. (iii) Since the functional mentioned above is not computable, we review examples of explicit functionals that have the virtue of yielding rigorous bounds to the energy.
Introduction
It is a pleasure to dedicate this article to Laszlo Tisza on the occasion of his seventy-fifth birthday. As a colleague at MIT he was a source of inspiration and encouragement, especially in drawing our attention to the importance of careful and precise thought in mathematical physics. The subject, if not the content, of this article may therefore not be inappropriate in a book dedicated to Professor Tisza (see the Acknowledgment).
The idea of trying to represent the ground state (and perhaps some of the excited states as well) of atomic, molecular, and solid state systems in terms of the diagonal part of the one-body reduced density matrix p(x) is an old one. It goes back at least to the work of Thomas [1] and Fermi [2] in 1927. In 1964 the idea was conceptually extended by Hohenberg and Kohn (HK) [3]. Since then many variations on the theme have been introduced. As the present article is not meant to be a review, I shall not attempt to list the papers in the field. Some recent examples of applications are Refs. 4 and 5. Some recent examples of theoretical papers which will play a role here are Refs. 6-12. A bibliography can be found in the recent review article of Bamzai and Deb [13]. This article has three aims:
(i) To discuss and prove some of the mathematical relations between Nparticle functions Us and their corresponding single-particle densities p. (ii) To discuss the mathematical underpinnings of general density functional theory along the lines initiated by HK. In that theory a universal energy functional F(p) is introduced. Despite the hopes of HK, F(p) is not defined for all p because
it is not true (see Theorem 3.4) that every p (even a "nice" p) comes from the ground state of some single-particle potential v (x). This problem can be remedied
by replacing the HK functional by the Legendre transform of the energy, as is done here. However, the new theory is also not free of difficulties, and these c 1983 John Wiley & Sons, Inc.
CCC 0020-7608/83/090243-35504.50
269
Int. J. Quant. Chem. 24, 243-277 (1983)
244
LIES
can be traced to the fact that the connection between v and p is extremely complicated and poorly understood. (iii) To present briefly another approach to the ground state energy problem by means of functionals that, while not exact, are explicitly computable and yield upper and lower bounds to the energy.
The analysis in this paper gives rise to many interesting open problems. It is my hope that the incompleteness of the results presented here will be partly compensated if others are encouraged to pursue some of the questions raised by them. It is not my intention to present a brief for HK theory. However, it deserves to be analyzed for at least two reasons: The HK theory is used by many workers and it gives rise to some deep problems in analysis. While it is my opinion that density functionals are a useful way to approach Coulomb systems, there are other approaches besides the HK approach [e.g., see (iii) above]. Apart from the difficulties mentioned above, the HK approach may be too general because all potentials have to be considered. Coulomb potentials are special and do lend themselves to a density functional approach; for example, Thomas-Fermi theory is asymptotically exact as Z -ao (see Sect. 5E and Ref. 14). In addition to this question of generality there is also the crucial point that the "universal functional" is very complicated and essentially uncomputable. If one is going to make uncontrolled approximations for this functional, then the general theory is not very helpful.
It is a pleasure to thank Barry Simon for some very helpful conversations and the proofs of Theorems 4.4 and 4.8. I also thank Haim Brezis for the proof of Theorem 1.3. 1. Single-Particle Densities
The first order of business is to describe the single-particle densities of interest. For simplicity we confine our attention to three dimensions whenever dimensionality is important. z = (x, will denote a space-spin variable, that is, x e R3 and o' a {1, . . . , q}.
q = 2 for electrons, of course, but one might wish to consider q = 1, which would mean that a ferromagnetic state is under consideration. We use the notation
J dz =oil Jdx. Let ii = 4r(z 1, ... , zN) be an N-particle function (which may be complex valued). To simplify notation we will not indicate N explicitly except where needed. However, the condition of fixed N is crucial and frequently glossed over. The density functionals that will be introduced later are explicitly N dependent in a highly nontrivial way (see Sect. 4A). rjr is assumed to be normalized: J I0IZ = 1,
270
(1.2)
Density Functionals for Coulomb Systems (a revised version of no. 144) COULOMB DENSrrY FUNCTIONALS
245
(with f = J dz,, ... , dzN) and to have finite kinetic energy, that is, N
(1.3)
T(ar)= E J IV,q, 12<00.
-1
Notes. f E L° means that f is a function satisfying hill, ={11 j°}"°
T(f) =
(27r)-3
(1.4)
J k2If(k)12 dk,
where f is the Fourier transform of f. Since f E L2. f exists and j E L2. H' is a Hilbert space with inner product (f, g) = f f *g + J V f * Vg. In most of the following it will be assumed that ' satisfies the Pauli principle,
that is, 0 is antisymmetric. However, some of the theorems are easier for symmetric (i.e., bosonic) +l' with q = 1, and occasionally this will be mentioned explicitly. In either case, the symmetry implies that
T(IG)=N J
(1.5)
IV,0 I2.
We define the single-particle density to be [see Eq. (A.1)]
p(x)=NE
JI'((x.oi).....(xN,o'N))12dx2...dxN.
(1.6)
Notice that f p (x) dx = N, not 1.
Determinants. If 46,W_ .. , ON(z) are orthonormal functions, we can form the determinantal (N!)-112 det (0,(z;)}, O(z,, ... , zO = (1.7) which is normalized. Then N
P(X)=
q
EE N
I'Yi(x,o)I2
(1.8)
J IVO+(x, r)12 dx.
(1.9)
q
T(J') = E E
Returning to the general case, the finiteness of T(ir) implies the following [15].
Theorem 1.1. p(x)''2EL2(R3)and Vp(x)'r2EL2(R3),that is,p(x)1/2EH'(R3). Moreover, J (Vp' i2)2 -- T(0).
Proof. pi/2EL2(R3) because Jp=N. Now Vp(x)=N f'(V,0)*oft+ *VIr(i, where J' means the integral in (1.6). By the Schwarz inequality,
Nf
IV11#12.
[Vp(x)]2-_ 4Np(x) J +
Thus
I (VP 1/1)2 dx = 11 (Vp)2p
,
dx <_ T(r(r).
271
Int. J. Quant. Chem. 24, 243-277 (1983)
LIEB
246
We know p' /2 a H' (R3) = { fjf e L2, V f E L2}. (Here we use the standard convention that {ABC} means the set of A such that condition C holds.) To discuss the converse of Theorem 1.1 some definitions are useful. Definition.IN = {p 1p (x) . 0, p' /2 a H' (R 3), f p (x) dx = N).
Definition. IN ={p Ip(x)30, jp(x)dx =N, p eL3(R3)}. 1N contains sN by the Sobolev inequality (see Ref. 16) because if f e H' (R3), then 1/3
J
lVf(X)J2
dx -- 3(r/2)4/3 hf If(X)E6 dX]
.
(1.10)
Equation (1.10) is true only in three dimensions, but analogous inequalities hold in other dimensions. By Theorem 1.1, T(I!/) z 3(zr/2)4131lp113
`1N is clearly a convex set ; that is, if p, and p2 a 9IN, then p 3 API + (1- A )p2 E
IN for all 0 --A -- 1. ON is also convex by the same proof as in Theorem 1.1; that is, by the Schwarz inequality (V,0)2
-- 4p[A(Vp1/2)2+(1-A)(VP21/2)2].
In particular, the functional f [Vp 112]2 is convex. The convexity of YN will be important in Sect. 3. Definition. A function (or functional) f is convex if
f(Ax +(1 -A)Y)
Theorem 1.2. Suppose p E'N. Then for either Bose or Fermi statistics there exists a 0 (which is a determinant in the fermion case) such that (1.6) holds and, moreover,
T(dr). J T(.i)
[VP112(X)]2
(bosons),
dx
(41r)2N2 J [Vp'/2(x )]2 dx
(1.11)
(fermions).
(1.12)
Proof. For bosons the proof is easy; simply take 41(x 1,...,XN)=
P(xi
n
))1/2
N
For fermions the construction is much more complicated. Some ideas from Ref. 17 will be used in the following. Write x = (x', x 2, x 3) and define
2a
f(x')=(N) J 272
ds
J.dt f 00
M
dup(s,t,u).
Density Functionals for Coulomb Systems (a revised version of no. 144) COULOMB DENSITY FUNCTIONALS
247
Then f is monotone increasing from 0 to 21r. For k = 0, ... , N -1 define .0 k (x) _ [p (x)IN ]'
12
exp [ikf (x' )]
It is easy to check that the 16k are orthonormal functions in L2(R3). (First do
the x2 and x3 integrations and then note that the overlap integral is of the form f m(df/dx')exp[i(Ak)f(x')]dx'={exp[i(Ak)f(oo)]-exp[f(Ak)f(-co)]}/ i (Ak) = 0. Furthermore,
NJ VI2= ((Optz)z+\r
)
Jg(s)6ds,
(1.13)
with
g(s)2=J Jdtdup(s,t,u). As in Theorem 1.1, we conclude that
gEH'(R')
and
s
z
J 1d g ) ds ` J
(Vpuz)2:A.
Since
K(s)2 = 2 JS g(y) (dd(y)) dy, we conclude by the Schwarz inequality that g(s)4--4[J g2][J (dg/dy)2]. Thus, the
last term in (1.13) is less than 4(2irk/N)2N2A. Finally, we take 41 to be a determinant as in (1.7) using the functions 46'k (x) x (spin up). Equation (1.12) follows by summing on k. Theorem 1.2 is closely related to the results of Gilbert [8] and Harriman [9]. For fermions, the extra factor N2 in (1.12) is noticeably different from the factor No in Theorem 1.1. Although (1.12) can be improved, it is not easy to do so. In any case, the conclusion is that the map from 41 to p'/2 given by (1.6) is a map from H'(R3N) onto H'(R3). But the map is clearly not 1 : 1; different Qr's can give the same p.
Question 1. Is this map continuous as a map from H'(R3N) to H'(R3)? That is, if 0 is fixed and Oj is a sequence (with corresponding p and p;) such that J10-41iI2i0 J1p'12-pi12I2 and does it follow that 0 and JIvpl/2-VP 1/212,0?
Question 2. Although the map is not invertible (since it is not 1: 1), we can ask the following: Given a sequence p, 12 that converges to p 11' in the above H'(R3) since, and given some I# satisfying (1.6) for p, does there exist a sequence up, [related to p, by (1.6)] that converges to u in the above H'(R) sense? [This
is equivalent to the statement that the map u yp 1/2 is "open," that is, the map takes open sets in H'(R3N) into open sets in H'(R3).]
273
Int. J. Quant. Chem. 24, 243-277 (1983)
LIER
248
Intuitively, the answer to both questions should be affirmative. The continuity
can indeed be proved, but the proof is not entirely elementary. A proof of Theorem 1.3, due to H. Brezis, is given in the appendix. Theorem 1.3. The map ill -p112 given by (1.6) is continuous as a map from HI(R3N) to H'(R3).
I cannot offer any proof of the openness of the map, however. The fact that these questions do not have simple answers should serve as a warning that the connection between y and p is not as obvious as one might intuitively think. 2. Single-Particle Density Matrices
If ' is given as before, we can define the single-particle density matrix
y(X,X')=NE 0((X,a1),...,(xN,QN)) X 0((x', o-1), ... , (XN, OW W dx2 ... dxN.
(2.1)
This definition is different from the usual one because we sum on a, in (2.1). Usually one defines the quantity j(x, a; x', o-'), so our y(x, x') = E. y(x, a; x', o). Clearly, p (x) = y (x, x). Theorem 2.1. y satisfies
(i) Try = f y(x, x) dx = N. (ii) As an operator, 0 _- y <_ qI, for fermions ; that is, 0 -_ (f, yf) < q (f, f ). For bosons,
0<_y-_ NI.
Proof. (i) is "obvious" but not trivial. The point is that if an operator K is given, then its kernel K(x, y) is defined only almost everywhere. In particular, K (x, x) can be anything. Thus, Tr K need not be f K (x, x) dz. However, (i) can be proved from (2.1). This is left as an exercise. To prove (ii) let M(x, x') = f (x) f (x')* be a one-particle operator with (f, f) = 1. Then A = E; _ i M (x,, x:) has as its largest eigenvalue on the antisymmetric space the value q. Moreover, A is clearly positive semidefinite. Thus, 0--(f, yf) = Tr yM = (IG, AO) -_q. Definition. Let -y(x, y) be any kernel. y is said to be admissible if Try = N and 0:5 y s ql (fermions) or 0:5 y s N (bosons). The set of admissible y is clearly convex; that is, if y and S are admissible, then so is ay + (1- a )S for O :s a s 1.
Now we come to a subtle point. If y is an admissible operator, we can ask two questions:
Question 3. Does an N-particle density matrix r always exist, where r = r(z,, ... , ZN; z'1, ... , z N), so that y is given by (2.1) with i4i1i* replaced by I'? (I' is a density matrix if o!5 r and Tr r = 1. r must also satisfy the appropriate symmetry.) Question 4. Does a 0 always exist so that (2.1) holds; that is, can r be chosen to be a pure state, namely, r= Ili)(ilr? 274
Density Functionals for Coulomb Systems (a revised version of no. 144)
COULOMB DENSITY Fl1NCTIONALS
249
The answer to question 4 is No! (for fermions). For bosons, the answer is Yes. The proof of question 3 (which we now call Theorem 2.2) has been known
for a long time. An explicit construction is given in Ref. 26. An example in which f fails to be of the form ili)(ili, for N = 2 and q = 1, is the case in which y has three nonzero eigenvalues 1, Z. To see this, let the normalized eigenvectors of y be f (x), g (x), and h (x), respectively; that is, Y(x, x') = f (x)f(x')* +ig(x)g(x')* + zh (x )h (x')*.
Let A = -y(x 1, x;) - y(x2, X2) be an operator on the antisymmetric states. Its lowest eigenvalue is -1 - 1/2 = -3/2, which is doubl7 degenerate. If I = 0)(4t,
then eG must be a ground state since Tr f'A = -Try = -1 -1 /4 -1/4 = -3/2. But every ground state is of the form ift = 2-1/2 det (f, p), where p = ag +bh, + lb 12 = 1. But then y = f)(f + p)(p, and this is never of the form f)(f + 1a 12
zg)(g + zh)(h.
The moral of all this is the following: On the one-particle level we can study density matrices y(x, x') or densities, p(x) = y(x, x). The former do not always come from pure states 41)(1. The latter do, as Theorem 1.2 shows. While y is more complicated than p (it has two variables), it has the distinct advantage that the map r-y is linear! The map 4i - p is nonlinear, and this, as will be seen, is the source of some difficulty.
The relation among ', r, y, and p can be summarized by the following diagram: lG
r
Y -'p,
(2.2)
by which we mean (i) the map rG H r =1b)(ty, (ii) r- y by (2.1) with lylf* replaced by i', (iii) y +y(x, x) = p(x ). (ii) and (iii) are linear while (i) is nonlinear.
Notation. We shall use the symbol 1/1yp (or any other combination such as y--'p) to indicate that 41 and p are related by the above maps. Technical remarks. Since y is self -adjoint and trace class, it can always be written in the form co
Y(x, X') = E A/,(x)fi(x')*, j-1
(2.3)
where the f; are orthonormal and 0 { Aj
N=Try= Z A;.
(2.4)
1=i
3. General Density Functional Theory The problem that will concern us in calculating the ground state energy for
N electrons interacting with each other via a repulsive Coulomb potential 275
Int. J. Quant. Chem. 24, 243-277 (1983) LIEB
250
Ixi - x; I-' and also interacting with a single-particle potential v (x ). If v = 0, the Hamiltonian is
Ho=K+
E Ixi-x/L-', 1si<j N
(3.1)
where K is the kinetic energy operator N
K
Ai
(3.2)
i=1
in units in which h2/2m = 1. Also of interest is the case where Ho = K alone (see Sec. 4C). Recall that N is fixed and will not be mentioned unless necessary.
Also, to simplify matters we shall confine our attention in the following to fermions. However, many of the following results have obvious analogs for bosons. The total Hamiltonian is
H,; =Ho+ V,
(3.3)
where CN
V = L v(Xi)
(3.4)
i-i
The ground state energy E(v) is defined to be E(v) = inf {(', HH4i)I4i E `W'N},
(3.5)
IVN ={4LIII,PII= 1, T(e(,)
(3.6)
where
Technical remark. Something should be said about the meaning of (Ilr, and about the class of v's under consideration. We shall always interpret (ii', H,O ) in the sense of a quadratic form; in particular, this means that (0, KO) -M T(O). It is not assumed that A* a L2. Since E H', it is easy to prove that (0, Ixi - x/I-'rfr) is finite for all i # j. The part containing v is J p (x )v (x) dx. Asp a L' and p e L 3
(since Op' /2 E L), p E L' for all 1 S p S 3. The integral is then well defined if V E L3J2 +L°°. This means that we consider v's that can be written as v = v3/2+ vw with V3/2 E
L3/2
and with Ivml a bounded function. This choice precludes v's that
go tom as Ix I m, such as the harmonic oscillator potential. Unbounded potentials can also be handled by the methods given here, but then we have to place additional restrictions on p so that J vp makes sense. We restrict ourselves here to L3/2+L°° for simplicity of exposition. The class includes Coulomb potentials because Ix I -' = 0(x )Ix I-' + [1- 9(x )]Ix I-' with O(x)= 1 if Jx I < 1, 6(x) _
0, IxI> 1. The two terms on the right are in L3/2 and in L°°, respectively. L 3/2 +L' is a Banach space with the norm NvII = inf {IIg1I312+IIhilclg +h = v}.
276
(3.7)
Density Functionals for Coulomb Systems (a revised version of no. 144) COULOMB DENSITY FUNCIIONALS
251
Technical remark. RN is a subset of the Banach space X = L3 fl L'. X*, the dual of X, is Y = L312+L°°. However, the dual of Y is not X because while L°°
is the dual of L', L' is not the dual of L'. However, X - Y*. The duality will be useful.
There may or may not be a minimizing I(i for (3.5), and if there is one it may
not be unique (for bosons it is unique because it is a positive function). Any minimizing rG (called a ground state) would satisfy
H.0 = E(v)i/i
(3.8)
in the distributional sense. The proof of this assertion is not difficult. For example,
a minimizing 0 will not exist if v is an attractive square well and if N is too large; the extra, unbound electrons will simply "leak away" to infinity. In such a case, E(v) would still have physical significance. It would be the ground state energy for fewer than N particles. There are three simple, but important, properties of E(v):
Theorem 3.1. (i) E(v) is concave in v: that is,
E(v)?aE(v,)+(1 -a)E(v2),
(3.9)
for all v,, v2, 0:5 a <_ I and v = av, +(1-a )v2.
(ii) E(v) is monotone decreasing: that is, if v1(x):5V2(x) for all x, then E(vl)<_E(v2) (iii) E(v) is continuous in the L312+L° norm and is, moreover, locally Lipschitz. In particular, E(v) is finite. Proof. (i) If 0 e 9WV, then
(+G, HH.) =a(.y, H°,iI0 +(1-a)(iI, H°,0)>aE(v,)+(1-a)E(v2). 00 (iii) Fix vo and let S = v - vo. We want to show that when 11811:5L/3, for some C, independent of v. [Here, L is the constant in (1.10).] Since E(v) is concave, it is sufficient to show that for some fixed D, E(v) - E(vo) y D whenever 11811=L13; because if 0 s y s 1,
y[E(vo+S)-E(vo)]<E(vo+yS)-E(vo)sy[E(vo)-E(vo-S)]. Let E(v,122) denote (3.5) with K replaced by K/2. Then
E(v)?E(vo, 2)+inf (r(i, [ZK+ES(x,)]II/).
The last term is bounded by -LN/2 because S = g + h with 119113/2 < L12 and llh lI= < L/2. Thus,
J So > - (L/2)[IIPII3+ N]. But (fir, Ki#)/2 ? (L/2)llp113 by (1.10) and Theorem 1.1. Finally, note that E(vo, z) -
E(vo) is a constant, D', independent of v.
277
Int. J. Quant. Chem. 24, 243-277 (1983) LIES
252
Now we begin the study of density functional theory in the manner of Hohenberg and Kohn. Their work is based on the following theorem [3]: Theorem 3.2. Suppose t4i (respectively, 02) is a ground state for v I (respectively
v2) and vI # v2+constant. Then pI 0P2 Proof. Suppose p, = P2 = P. 1#10 02 because they satisfy different Schrodinger equations, (3.8). [Note. To prove this we must know that V10 = 02* implies that
vt = v2. This, in turn, requires that 4s(x) does not vanish on a set of positive measure. This technical point is discussed in remark (ii) preceding Theorem 3.5.]
Moreover, 02 (respectively, ¢I) does not satisfy (3.8) for vt (respectively, v2). Therefore, E(vi)<(142,HH,*2)=E(v2)+ J (vI-V2)P
Likewise, E(v2)<E(vI)+J (v2-vI)p. This is a contradiction.
Hohenberg and Kohn assume that every p comes from some I that is a ground state for some v. For such p they define the functional
FHK(p)=E(v)-J
vP,
(3.10)
and we shall retain this definition for p E 9'N, where s?N = {pip comes from a ground state}.
(3.11)
siN 0 5N, as remarked earlier, and it is not convex (see Theorem 3.4)! The definition given by (3.10) requires Theorem 3.2, according to which there is a unique v (up to a constant) associated with p. We can also define `Y'N = {v IH has a ground state}.
(3.12)
It then follows easily that for v E `Y'N
E(v) = min I FHK(P) + J VPlpEdN}
(3.13)
This is the HK variational principle, but it is important to note that it holds only for v E `1N, which is unknown, and that the variation is restricted to the unknown set SG-
We also do not know what FHK is, and that is a very serious problem. But there are also conceptual problems, which will be addressed here. If F is to be used in a variational principle, it is clearly desirable that F be a convex functional. In particular, it should be defined everywhere on .1N, or at least on some known convex subset of .$ N. The domain of FHK (i.e., s1N) is not all of -ON and it is not convex. This last fact is closely connected with the following difficulty: One can define a functional
278
Density Functionals for Coulomb Systems (a revised version of no. 144)
COULOMB DENSITY FUNCTIONALS
253
for all p in 'N by* F(p) = inf {(0, Hotfi)I,, -P, 4, E WN}.
(3.14)
It then follows trivially that
E(v)=infIP(p)+J VPIPEJfN}.
(3.15)
pE.iN.
(3.16)
F(p)=FHK(p),
if
So far, so good. The difficulty is that P is not convex either. However, F has one important property that is proved in the appendix. Theorem 3.3. For each p in ON there is a ' E WN such that P(p) = (4', Ho4s). In other words, the infimum in (3.14) is a minimum.
The following functional F is one choice for "the density functional" that remedies the difficulties mentioned so far:
F(p)=sup IE(v)-J VP IV EL'n+L'°}.
(3.17)
We shall explore the properties of F, but it, too will be seen to have subtle difficulties of its own.
Remarks. (i) (3.17) defines F(p) for all p E RN, not just 3N, provided F is interpreted in the extended sense as a function that can have the value +oo. In fact, (2.17) defines F on the much larger set X = L3 fl L t, without the restrictions p (z) -- 0 and f p = N. As Theorem 3.5 shows, however, it is only necessary to consider F on the convex subset J r4 of X. (ii) Recall that F depends explicitly on N through E. (iii) Since F is the supremum of a family of linear functionals, it is convex. (iv) Theorem 3.8 shows that F(p) = +co if p Of JON. There is an alternative definition of F, namely, F, by which F is finite on the set
5 .s{pIp(x)2OandVpti2eL2}, without requiring f p = N. This is
F(p) _ (J p) F(pl J p)' F'(0)=0.
p4*O,
(3.18)
It is easy to check that the convexity and lower semicontinuity (a concept to be defined later) of F carry over to F. This definition has the virtue that F is finite * Levy [10] also defined F(p) which he called 0, and derived (3.15). He did not prove Theorem 3.3. but assumed the existence of a minimizing +'. Also, he did not establish the connection between
t and the Legendre transform, F (Theorem 3.7). In Ref. 11, Levy proved Theorem 3.4(11), independently and virtually at the same time as myself, using essentially the same construction. See Ref. 12 for additional remarks about Q.
279
Int. J. Quant. Chem. 24, 243-277 (1983) LIEB
254
on a dense subset of the set of nonnegative functions in X. However, this does not change the theory in any important way, so we shall continue to use the definition given by (3.17). (v) Other characterizations of F, directly in terms of F, are given in Theorem 3.7, and in Eqs. (4.5)-(4.7). There is an obvious relation between F and F, namely,
F(p)sF(p)
for all pe.ON,
(3.19)
since E(v) s F(p) +I vp for all p e JN. Furthermore, since F is convex and F is not convex (by Theorem 3.4), there are p's in ON for which F(p)
First we prove that not all p's come from ground states. The essential ingredient is the existence of v with a degenerate ground state. (Such v's, incidentally, preclude the existence of a map v Hp.) Theorem 3.4. Let N > q = number of spin states. Then (i) F(p)) is not convex (ii) There exists a p E.'N that does not come from a ground state .4'. Moreover this p is a convex combination of p's that do come from a ground state.
Proof. Let v be a spherically symmetric potential having a ground state and with the property that its ground state has orbital angular momentum L 1. We assume the degeneracy is no greater than necessary, namely M = 2L + 1. The orthonormal ground states are Mfrs, ... , e,&M and .; yp1. Under simultaneous rotation of all N coordinates, they transform as a basis for the M-dimensional irreducible representation of 0(3). The following fact is easy to prove: (a) If p = M-1gyp;, then #(x) is spherically symmetric: that is, p depends only on r = jx j. A second fact that will be needed is (b): if 44 is any ground state (and hence a linear combination of the 4r;) and 45'p, then p is not spherically symmetric.
This fact must follow from some group-theoretic agreement, but I have not found one. However, it is not hard to see that (b) is equivalent to (c): There exists a perturbation of v, v (x) - v (x) +A w (x), with w bounded and of compact support, so that to first order in A the M- fold degeneracy is broken. Such pairs
v and w certainly exist, so we can henceforth assume that (b) holds. [A proof that a v satisfying (b) exists is the following. First, take the case that Ho = K, that is, independent particles. The ground states are determinants. Choose v so that the ground state has L >_ 1, in which case (b) obviously holds. Next, consider
H = K +A Ejx; -x,L-' + V. Angular momentum is still conserved and for sufficiently small k the ground state will have the same L and, by continuity of the ground states, (b) will continue to hold for small A. We are interested in A = 1 but, under the scaling
x-.x/A,
v(x)
A-zv(x/A)=v'(x),
the v, A problem is converted into the v', A = 1 problem. Thus, v' has the desired properties. I thank B. Simon for this remark.]
280
Density Functionals for Coulomb Systems (a revised version of no. 144) COULOMB DENSITY FUNCTIONALS
255
Clearly F(p;) = constant = D = E(v) - f vp;. We claim F(p) > D, thereby proving lack of convexity. Obviously, F(p) z D, for otherwise we could use p instead
of p; in (3.15). (Note: f vp = f vp; = constant = C.) Suppose F(p) = D. Then p comes from some 41 that must be a ground state for v. But (a) and (b) show this
to be impossible. Thus F(p) > D. Moreover, p cannot come from any ground state ' for any other v'. If it did, then r
E(v')=F(p)+J v'p>M-'I(F(p;)+J vp) This implies that for some 1:5i:!:-:M, F(p4)+f v'p; <E(v'), which is a contradiction.
Remarks. (i) The foregoing proof holds just as well if (0, Hopi) is replaced by T(I') in the definition of F. [This functional will be denoted by l'(p) and the analog of (3.17) by T(p).] In other words, the interelectron Coulomb repulsion plays no role in Theorem 3.4 (see Sec. 4C). (ii) There are other p's that do not come from some v, namely, those p EIN that vanish on a nonempty open set. If V E L9J2+L° and aG is its ground state, then 0 cannot vanish in an open set by the unique continuation theorem [18]. (Strictly speaking, this theorem is only known to hold fore EL C,butitisbelievedto hold for L' 12 + L'.) Presumably, such p's can, in many cases, be obtained as limits in
which v - 30 on the open set. Therefore, if the set of allowed v's can be extended properly to include infinite v's, the existence of such p's may not have any particular importance. The question is very delicate, however, as Englisch and Englisch [7] showed recently. Even for one particle there are densities which never vanish but which do not come from any v, even if one allows density matrices (see Sect. 4B) instead of pure states. These densities have regions in which they are "small"so that the obvious v (defined by v =p "2Ap1/2) has the property that -A+v cannot be defined as a semibounded operator. Theorem 3.5.
E(v)=inf {F(p)+J pvlpEL3f1L'},
(3.20)
E(v)=inf {F(p)+J pv)pEJN}.
(3.21)
Remark. The right sides of (3.20) and (3.21) are automatically concave functionals, which is a property we already proved for E.
Proof. Let M-(v) [respectively, M'(v)] be the infimum in (3.20) [respectively, (3.21)]. Obviously, M-(v)!5M4(v). First, pick v,,. Clearly F(p) > E( t-()) - f vop = F, (p). Therefore, r
M (vlyinf{FI(p)+J pvJpEL3f1L'}.
281
Int. J. Quant. Chem. 24, 243-277 (1983) LIES
256
and hence M-(vo)aE(vo). Second, by (3.19), F(p)sP(p), so that M*(v)s inf {FL(p)+ f pv{p e.fN} = E(v).
Let us pause briefly to review the situation. Three functionals have been defined: FHK
F F
defined on siN C IN ;
defined on 5NCL3 flL'; defined on X=L3flL'.
Of these, only F is convex and only F and F satisfy the variational principle for all v.
The next step is to find out something of the nature of F. It is at this point that the analysis becomes complicated and where difficulties and incompleteness
arise. The basic reason is that the connection between v and p is anything but simple. We have X = L3 fl L' and its dual X * = L 3'2 +L-. Although X is not the dual of X *, it is a subset of X**, the dual of X *. Definitions. (i) A sequence p" e X is said to converge top e X (p" -' p) if and only if Ib -p113 i 0 and IIp" -phj, - 0. This is also called norm convergence. P.
converges weakly to p(p"--p) if and only if f v(p"-p)-'0 for all ve Y=X*. Clearly, strong convergence implies weak convergence.
(ii) A functional f on X is continuous (or norm continuous) if and only if p" -*p implies f (p.) - f (p ). Weak continuity requires the concept of nets to define but, if f is weakly continuous, then whenever p" -- p, f (p") -> f (p ). Weak continuity implies norm continuity.
(iii) A real functional f on X is lower semicontinuous (I.s.c.) if and only if p" -' p implies f (p) _< lim inf f (P" ). Weak lower semicontinuity requires nets to
define, but if f is weakly l.s.c. then p" -p implies f (p) s lim inf f (p" ). (Weak) lower semicontinuity is equivalent to the following: {p j f(p) s A } is (weakly) closed for all real A.
Remarks. (i) Weak lower semicontinuity always implies lower semicontinuity, but not conversely. It is a theorem of Mazur [19], however, that if f is convex and norm l.s.c., then it is automatically weakly l.s.c. (ii) The function p (x) m 0 is not in the L3 fl L' weak closure of 'N. The reader may be puzzled by all these definitions, especially lower semi-
continuity, because finite convex functions on R" are always continuous. Unfortunately, this is not true in infinite-dimensional spaces such as the space X we are considering. Even l.s.c. cannot be taken for granted. Theorem 3.6. F(p) is weakly (and hence also norm) lower semicontinuous. Proof. K.,
282
={pJF(p)sA}=IpIE(v)-J vpsAfor all vEY}.
Density Functionals for Coulomb Systems (a revised version of no. 144) ('OUI.OMB DENSITY FUNCTIONALS
257
Now if p -+ p in norm and P. E K then, for each v E Y.
E(v) -
J
vp=lim(E(v)-
Therefore KA is norm closed, so that F is norm l.s.c. Weak l.s.c. is a consequence of Mazur's theorem. Next we define the convex envelope (CE).
Definition. Let f be a real functional defined on a subset A of X. f(p) is allowed to be +oo, but not for all p E A. CE f is defined on all of X as follows: CE f(p)=SUP {g(p)lg is weakly l.s.c., g is convex on X, and g(p')s f(p') for all p'EA}. It is easy to check that CE f is convex and weakly l.s.c. and CE f(p)sf(p) for all p E A. However, CE f (p) may be +oo for some p. The function of interest is CE F with A =JN. Note that A is convex and that
F (and hence cE F) is finite on A by Theorem 1.2. Since CE FsF on A, it is obvious from (3.19) and Theorem 3.6 that F:5 CE F on X. On the other hand, suppose we use cF F instead of F in (3.15). This gives a new function, which we call F. Clearly E's E. Then, if E' is used in (3.17), we get a new function F, and F's F. However, an infinite-dimensional generalization of Fenchel's theorem [29] (which uses the Hahn-Banach theorem) states that if the original function (in our case, cF F) is convex and weakly l.s.c. on X, then its double Legendre transform (in our case F) is equal to the original function. Thus, F' = CE F and we have
Theorem 3.7. F(p)=cEF(p) for all pEL'flL'. The reader may wonder what Theorem 3.7 is good for; the following is an example of the usefulness of the foregoing functional analysis (see Theorem 4.3).
Theorem 3.8. For all p E L' fl L' let
G(p)=J (Vp(x)"2)2dx = +00
1fPEJN otherwise.
Then F(p) ?G(p), for all pEL'flL'. Proof. G is obviously convex on X [see the remark after (1.10)]. We claim that G is norm I.s.c. (Note: The norm in question is L3 fl L', not the H' norm on p 1/2.) If so, we are done because G is then weakly I.s.c. and, by Theorem
1.1, G<-F on IN: but then G CE F=F. To prove norm l.s.c., let p be any sequence in X with p - p ; that is, exists and is finite, and we have to show that G>G(p). We can also assume p(x) -e 0 a.e. because if p <0 on a set S of positive measure, then, for sufficiently large n, ILp -pII, i 0 and 11p,, -pf13 _ 0. We can assume that G = lim
283
Int. J. Quant. Chem. 24, 243-277 (1983)
L.IFB
258
p < 0 on some set of positive measure; hence p 0'N and G (p.) = oo. For a similar reason we can assume f p = N. Since oo, P. E 5N. Thus, if we define
g =p,'/2 and g =p'/2, we have: (a) g is bounded in H'; (b) g. -' g2 in L3 and L'. By the Banach-Alaoglu theorem there is an f E H' such that g -f and V f - Vf weakly in L2. Clearly f (x) ? 0. It is not hard to prove that if g -f in L and g - g2 in L', then g = f. Hence VF = Of, and thus Vg -Cg. But since f (Vg)2 is H'-norm continuous, it is H weakly l.s.c., so that lim G(p).
Theorem 3.8 is certainly not obvious. Among other things it says that if p0' N (and such p's can be quite smooth and innocent looking), then there exists such that The reader a sequence of potentials is asked to reflect on this fact. Another interesting fact is that F is convex and
finite on .ON, but infinite off ON. However, the complement of 'N (in X) is
dense (in the X norm) in 'N and JN is dense in the cone of nonnegative functions in X. The following upper bound complements Theorem 3.8. Theorem 3.9. If p E .ON, then F(p) ts- F(p):55 (4ir)2N2G(p)+z
JJ p(x)p(y)Ix -yl-' dxdy.
(3.22)
Proof. Use the definition (3.14). By Theorem 1.2 there is a determinantal 0, with O -p, such that (1.12) holds. With this 41 we can calculate the Coulomb
repulsion I = (IG, 11x; -x;I-'iIi). I has a direct term, given in (3.22), plus an exchange term. The latter is negative, as is well known, since Ix - y I-' is a positive
definite kernel. Thus F(p)s right side of (3.22). Then use (3.19). Remark. By one of Sobolev's inequalities,
D - Jf p(x)p(y)Ix - yI - ' dxdy s (const.)IIPIIFi5. By Holder's inequality IIPII6/5
D:5 (const.)N3/2G(p
)''2
<_ (const.)[N +N2G (p)].
To continue the study of F the following concept is needed.
Definition. Let f be a real functional on a subset A of a Banach space B, and let p E A. A linear functional I on B is said to be a tangent functional (TF) at p(, if and only if for all p E A
f(p)?f(po)-1(p -po).
(3.23)
1 may not be unique. If ! is continuous, then 1 is a continuous tangent functional at p,,.
284
Density Functionals for Coulomb Systems (a revised version of no. 144)
COULOMB DENSITY FUNCTIONAL S
259
I is a continuous linear functional on X if and only if it has the form I vp with v E X * = Y. If f is convex, then at every point po at which f is finite, f has at least one TF. This is guaranteed by the Hahn-Banach theorem. However, f may have no continuous TF at po. The functional of interest is obviously F. In general, Fs F but the following says something about those p for which F(p) = P(p). Theorem 3.10. Let Po E IN. The following are equivalent : (1) F(po) = F(po) and F has a continuous rF at po. (2) PO E SIN.
(3) F has a continuous TF at po; that is, F(p) a: P(po) - J v (p - po) for p E .4N. (4) (3) and (5) hold with the same v.
(5) E(v)=F(po)+J vpo for some v. (6) (5) holds and, in addition, V E ?N and (7) F has a continuous TF at po and v is unique up to a constant. Moreover, F has the same continuous TFS at po, and no others.
Proof. (1)x(3): ForpEJN,
F(p)?F(P)?F(po)- J v(p-po)=F(Po)- J v(p-pa). (3)'(4): Let F1(p)=F(po)-Jv(p-pa)sF(p).Then F(po)+ J vpo?E(v)?inf I
FI(P)+J PvIPEJSN} =F(Po)+ J vpo.
(4)'(5), (7)'(3), (6)x(5): All trivial. (5)x(1): F(P(,)+Ipnv=EMF(Pn)?F(Po)=>F(Po) F(po). Then, for all p E X, F(p) + f pv > E(v) = F(po) + J pov. (2)x(5): By (3.16).
(5)x(2), (6): By Theorem 3.3, F(po)=(II/,H(,&I!) for some Il/ with 0
Then E(v)=(+',HoJr)+f
ve'VN, and v-.po. Thus (1)-(6) are
equivalent and (7) x(3). Now we show that (1)-(6)x(7). If v is a continuous TF
for F, then v is a continuous TF for F [by the proof of (1)x(3)]. If v is a continuous TF for F, then F(p) - E(v) - f vp, so v is a continuous TF for F. Suppose F has two continuous TFS v and w with v - w 0 constant. Then E(v) _ F(po) + J vpo and E(w) = F(po) + f wpo. Since po E dN, this is impossible by Theorem 3.2.
It should be noted that the only place that the HK Theorem 3.2 entered in the analysis of F was in establishing the uniqueness (modulo constants) in (7). Now we turn to two important questions whose answers we cannot give but that are obviously important for the theory. We replaced FHK by F because FHK was not defined on all of 'N. Theorem 3.10 states that on .c4, where FHK is defined, F = F = FHK and F has an essentially unique continuous TF. Question 5. For which points of ON does F have a continuous TO Where there is one, is it unique (modulo adding a constant to v)?
285
Int. J. Quant. Chem. 24, 243-277 (1983)
LIEB
260
Question 6. If F has a continuous TF at poEIN given by some v E L312+LOO, is this v e `I1N?
Questions 5 and 6 have alternative formulations, given below. Theorem 3.11. Let po a .5N and v E for all p,
L3i2
+ L. v is not necessarily in `VVN. Then,
F(p) ? F(po) - J V (p - po)
(continuous TF)
(3.24)
[minimum in (3.21)].
(3.25)
if and only if
E(v) =F(po)+v J Proof. Assume (3.24) rand let
E(v) zinf For the converse,
Po
be its right side. Then
J Pv} =F(po)+ J vpoa, E(v). `
F(p) + J vp aE(v)=F(po)+ J
vpo. 0
Question 5 is equivalent to the following: For which po E IN is there a v such that (3.25) holds? Is this v unique (up to constants)? Question 6 is the following: If (3.25) holds, is v E VN? Some insight into the continuous TFS of F are provided by the Bishop-Phelps
theorem. We refer the reader to Ref. 20 for this as well as other interesting facts about convexity. A definition is needed.
Definition. Let F be a real functional on a real Banach space B with dual B* (the set of continuous linear functionals on B). b * E B * is said to be F-bounded
if there is a constant C (depending on b* but not on b) such that F(b)?b*(b)+C
for allbeB. In our case B = X and F is our density functional. Theorem 3.12. Every v e X * = L 312 + L°° is F bounded.
Proof. By Theorem 3.8, F(p) = oo if p L IN, so we only have to consider p E IN and prove that G(p)? I vp + C for some C. The proof of this is identical to the last part of the proof of Theorem 3.1. The Bishop-Phelps theorem is the following.
Theorem 3.13. Let F be a l.s.c. convex functional on a real Banach space B. (Note: Norm and weak l.s.c. are identical.) F can take the value +oo, but not everywhere. Then
(i) The continuous tangent functionals to F (over all of B) are B*-norm dense in the set of F-bounded functionals in B*
286
Density Functionals for Coulomb Systems (a revised version of no. 144) COULOMB DENSITY FUNCTIONALS
261
(ii) Suppose bo e B and b *o E B* with F(bo) < oo. For every e > 0 there exists b, E B and b,*. E B * such that IIb ,* - bo lle s e and e Jib. - bolls s F(bo) +b o* (bo) -inf {F(x) +bo* (x )l x e B}. Moreover, b *,
is tangent to F at b,, namely F(b) ? F(b1) - b? (b - b,) for all b.
The significance of Theorem 3.13(i) is the following. There are certainly many v's in Y that are not in `YN. (Example: Suppose v eL't2 and 11V 113/2
where L is the constant in (1.10). Then (4i, H,4')>0 for all 4i, but E(v)=O because we can always take a sequence 0 that "leaks away to infinity.") Let v9 `YN, whence P(p)+Jvp does not have a minimum. What Theorem 3.13 says is that there always exists a sequence v (not necessarily in YN) such that
(a) F(p)+ J pv has a minimum at some p E.ON and this minimum is (b) F(p) for all p. v,, (p (c) v - v in the L112 +L°° norm.
Point (c) means the following: v = v +g +h. with
and Ilh,,ll=- 0. In particular, if veL312 with IIvII312
One consequence of Theorem 3.13(ii) is the following. Theorem 3.14. Let POE 'N. Then there exists a sequence p E .S6N such that
(i) p -. po in L3 fl L' norm. (ii) F has a continuous TF at p,,.
Proof. Given n > 0, by (3.17) there exists v such that
J pov >
F(po) -1 In. Hence
F(P) a
J pv z F(po) - J v, (P - po) - 1 In.
Take e = I in Theorem 3.13. There exists W. E Y such that w is a continuous TF at some p and
lI' -PolIEF(Po)+ J with
Z=inf {F(p)+J By the above,
Z?F(po)+ 4. Additional Remarks about Density Functionals A. The N-Dependence of F
As was stressed earlier, any functional F that satisfies (3.20) or (3.21) must
depend explicitly on the particle number N. This fact is unavoidable and
287
Int. J. Quant. Chem. 24, 243-277 (1983)
LIFR
262
frequently overlooked. Let us denote the N dependence by F(.'V, p). It might be hoped that F is jointly convex in N and p in the sense that for N a- 2 F(N + 1, PO +F(N -1, p2) 2t 2F(N, ?p, + ?P2).
(4.1)
This convexity definitely does not hold as a general feature, as will be demonstrated. The importance of convexity is shown by the following. Theorem 4.1. Consider the following two statements about any two functionals, Fand E: (i) F(N, p) is jointly convex in Nand p in the sense of (4.1). (ii) E(N, v) is convex in N for all fixed v ; that is, for N ? 2
E(N+1, v)+E(N-1, v)?2E(N, v).
(4.2)
(a) If (3.17) holds, then (ii) implies (i). (b) If either (3.20) or (3.21) holds, then (i) implies (ii).
Proof. (a) For each v, E(N, v) -f vp is jointly (N, p) convex. By (3.17), F(N, p) is the supremum of such convex functions and hence is convex. (b) Pick e > 0. For N + 1 there is a p+ such that
A=F(N+1, p.)+I p.vsE(N+1, v)+e. Likewise,
B =F(N-1, p_)+J p_v sE(N- 1, v)+e. Then
For the N problem, define 2p = p. +p .
2{F(N,p)+J pv}.A+B. Since this holds for all e >0, (ii) is proved.
Equation (4.2) has a simple physical meaning. The ionization potential increases as the number of electrons is decreased. This is intuitively expected to be true, but if it is true, it must be because of some special property of the Coulomb repulsion. A non-Coulombic counterexample is given below. The kinetic energy functional T(N, p) is not even convex in p (Theorem 3.4), but the Legendre transform T(N, p) is jointly convex. This is so because E(N, v) is indeed convex in N for independent particles as the explicit expression for E, as the sum of the first N eigenvalues (counted with an extra multiplicity q) shows.
What about the convexity of F when the Coulomb repulsion is included? While it has been conjectured that E(N, v) is convex in N (for all v) in the case of Coulomb repulsion, this has never been proved. It has not even been proved
that E(3, v)+E(1, v) ?2E(2, v).
288
Density Functionals for Coulomb Systems (a revised version of no. 144)
COULOMB DENSITY FUNCTIONALS
263
Lest the reader think that convexity in N is a general feature, we present a counterexample. Replace Jx I-' by the hard-core repulsion O (x) = oo if Jx I < I and O (x) = O otherwise. Pick four distinct points xo, Y1, Y2, Y3 in
R3
such that Jy, - y; J > I
for all i 96j but Jx,, - y; J < I for all i. Let v (x) = -2A < 0 in small balls about the y,, v (x) = -3A in a small ball about xo, and v (x) = 0, otherwise. If the kinetic
energy be neglected, then E(1, v) _ -3A, E(2, v) _ -4A, and E(3, v) _ -6A. Convexity does not hold. This can be turned into a proper example by letting A be sufficiently large so that the kinetic energy can effectively, be neglected; it is also possible to replace the hard core by a soft core.
Remark. The foregoing example is not applicable if 0 is replaced by JxJ-', thereby keeping alive the hope that convexity holds in the Coulomb case. The reason is the following: Given any four points xo, y1, Y2, y3, let Ixo - y I I = maxi {Ixo - Y. I}.
Then
Ixo-Y1I
'+IYt-Y31
The proof of this is left as an exercise, as well as the implication that if the kinetic energy is neglected, then convexity holds in the Coulomb case.
Question 7. For the case of Coulomb repulsion, is F(N, p) jointly convex in N and p? B. Density Matrices
Another possible modification of the theory of Sect. 3 is to replace densities p(x) by single-particle admissible density matrices y(x, x'). (See Questions 3 and 4 in Sec. 2. We do not restrict ourselves to y's that come from pure states 0)(0.) This set of y's is convex, and F(y), defined analogously to (3.14), is convex [see the proof of Theorem 4.1(b)].
Despite the attractive feature just mentioned, there are three drawbacks to the approach: (i) The problems about continuous tangent functionals remain and may even be more complex than before. (ii) The original aim of the theory was to express the energy in terms of p(x) and not y(x, x'). (iii) While the set of admissible y's is well defined, it is not easy to identify. Given some y, it is easy to verify that Tr y = N, but it is difficult to verify that
0`y<_ql. Still another possible modification is to retain p but to consider all N-particle density matrices r instead of merely pure states 0) (0. In other words, consider
F-p instead of ,/iyp and define FDM(p) = inf {Tr Horlr-.p}
(4.3)
on .ON and FDM(p) = +oo otherwise. Because r --*p is linear, FOM is convex on 1N. (Note: The example in Theorem 3.4 does not yield nonconvexity of FI)M.)
289
Int. J. Quant. Chem. 24, 243-277 (1983)
LIEB
264
Obviously, the analog of (3.15) holds, namely,
E(v) = inf {FDM(P)+J
PV 1P E
IN}
(4.4)
Since FOM is convex, (4.4) can be used directly instead of (3.20) or (3.21). Both F and FOM are convex. The amusing fact is that
F(p)=FDM(P), (4.5) PEJON. Equation (4.5) is not at all obvious, but it does say that the modification does
not change the theory in any way. Equation (4.5) also yields another characterization of F. Equation (4.5) is proved in Theorem 4.3. First, I is admissible if and only if 00
.-I with 0 s A;, EA; = 1, and the ti, are orthonormal. If t; -+p1, then Tr Hor = EA. (+Gr, Ho+G. ).
Thus we conclude that for all p e -ON
FoM(P)=inf { E A;F(P,)IyA+P1=p,P.E'N,Ar?0,yAr=1}. r=1
(4.6)
A simpler expression (which has to be proved) is FDM(P)=inf {EAF(Pr)I EAP.=P,P.EJON,
0,EAr= 1},
(4.7)
where the sums in (4.7) are restricted to finite sums. In view of (4.5), (4.7) is an alternative characterization of F(p) for p E'N.
Theorem 4.2. Equation (4.7) is true. Proof. Pick e > 0. Using (4.6), let {A,, p, } be an infinite sequence satisfying MA,p; = p, p1 E ON, and FFM(p) ? EA;F(pr) - e. Since EA; = 1 and I A,F(p;) < oo, there
exists K such that A s EO-K Aj s e and B - E°-K A;F(p;) s e. Assume A > 0 for otherwise we are done. By Theorem 1.1 and the convexity of G(p)= I (Vp1"2)2 E :2t
K
E ArG(pi)?AG(PK) K
with p K = EK A,p,/A E .ON. By Theorem 3.9 and the remark following it;
F(PK)sC(N2G(pK)+N]. Therefore the finite sequence {A;, Pi }K t with (A K, PK ) = {A, P K } satisfies EA;F(pr) :5
FDM(p)+ECN(N+1)+e. Theorem 4.3. Equation (4.5) is true.
290
Density Functionals for Coulomb Systems (a revised version of no. 144) COULOMB DENSITY FUNCTIONALS
265
Proof. The easy part is that for P E IN, FDM(p) aF(p). By (4.4). E(v) s FDM(p) + j pt; for all v. Hence, by (3.17), FDM(p) ? F(p). The hard part is contained
in Corollary 4.5, which will be assumed for now. Then: (i) FDM(p)sP(p) by (4.6); (ii) FDM is convex and l.s.c. Hence FDM(p) s CE F(p) = F(p) by Theorem 3.7.
Theorem 4.4. Suppose {p } and p E 5N and p p weakly in L'. Then there exists a density matrix r, with r-p, such that Tr Hor s lim inf FDM(p ). The proof of Theorem 4.4, due to Barry Simon, is given in the appendix. Corollary 4.5. (i) FDM is (norm and weakly) l.s.c.
(ii) If P E 1N, there exists a density matrix r with ry p such that Tr Hor = FDM(p) (see Theorem 3.3).
Proof. (i) If p -+ p, FDM(p) s Tr Hors lim inf FDM(p ). Norm I.s.c. implies weak I.s.c.
(ii) Take p = p in Theorem 4.4. C. The Kinetic Energy Functional
Kohn and Sham (KS) [30] define a kinetic energy functional TKS(p). There are several other possible kinetic energy functionals and we shall explore their interrelations, as well as the fact that TKs does not have a property assumed by KS. Ks define the exchange and correlation functional E,c(p) by JJp(x)p(Y)Ix-y[_'dxdy+TKS(p)+Exc(p).
FHK(P)-21
(4.8)
FHK and TKS are defined on different subsets of .0N, so E,c is defined only on a third unknown subset of .ON. This difficulty can be remedied by using P and '
in (4.8), but there is another point that should be stressed: There is no reason to believe that E,, is convex on 'N. First, let us give some definitions. These use K instead of Ho but otherwise are self-explanatory (with the aid of the equation numbers on the left): (3.5):
E'(v)
on
L'"Z+L°°;
(3.10):
TKS(p)
on
sQN
(3.14):
t(p)
on
.IN;
(3.17):
T(p)
on
L3flL'.
(3.11); (4.9)
(T(ar) = (dr, KO) was defined in (1.3) but it is quite different from T(p) above. It is hoped that this notational lapse will not be confusing.) All the previous theorems [except for 3.9, wherein the last term in (3.22) should be omitted] carry over to these quantities. The primes on E'(v) and ON indicate that these are different from before. Since Theorem 3.4 still holds, sd'N is not AN. It is left as an exercise to show that ''N # sF .
Question 8. What is sL fl .sZQ'N?
291
Int. J. Quant. Chem. 24, 243-277 (1983)
LIED
266
There is one more kinetic energy functional that can be defined on.4N, namely, Tde,(p) = inf {(O, K41)1 4i -p, 0 e WN, 0 is a determinant).
(4.10)
Clearly, Tde,(p) a- t(p). The question to be addressed is whether TdE1= t. The answer is No!, not even on all of 4'N. Ks assumed implicitly that TKS(p) = Td,(p) for P E 4d ; any such p minimizes K + V, but it is not true that such a p (x) can always be written as N I Ioi(x)I2 with the t/i; being orthonormal functions on R3. (Spin is a complication that is ignored at this point for simplicity.) In other words, not every ground state of K + V is a determinant when degeneracy is present. I thank B. Simon for drawing my attention to this subtlety and for the construction in Theorem 4.8, which is reminiscent of the construction in Theorem 3.4. Of course, TKS = T on sd'N by definition. Also T = T on .9f by Theorem 3.10. The following shows that there are cases in which t = Tde,. Theorem 4.6. Suppose P E so that K + V has a ground state. If this ground state is nondegenerate, then TdG1(p) = T(p).
Proof. The 0 that minimizes (0, [K + V]O) is, of course, a determinant. The following analog of Theorem 3.3 will be needed for Theorem 4.8.
Theorem 4.7. Let p EON. Then there exists a determinant that minimizes (I(,, KI/i) under the condition that a(i H p, 0 E WN, and 41 is a determinant. Thus, (4.10) is actually a minimum.
Proof. Let D; be a sequence of determinants with D1 Hp
and
lim (D;, KD;) = Tde,(p ).
The proof of Theorem 3.3 shows that 0 exists such that (i) p; (ii) (IG, Ku/r) _ Tde,(p ); (iii) Di -. 41 strongly in L2. It suffices to show that #A is a determinant. i = 1, ... , N, be the orthonormal single-particle functions of Di. By the Let Banach-Alaoglu theorem, N functions f'. ... . fN exist so that (after passing to
a subsequence) f; -f' weakly. The f' are not necessarily orthonormal. The function
Pl(z1,...,2)V)=llfl(z,) then converges weakly to P = R f'. This so because any IG E L2(R3N) can be approximated in norm by sums of product functions. Therefore,
D; -(N!)"2 det [ f'(z1)]=D
weakly.
Theorem 4.8. Let N = 7 and q = 1. Then there is a p E .s1?N such that Tde,(p) > T(p).
Proof. Take v(x) _ IxI - ', the hydrogen potential. The eigenvalues of -:1+v
are -1/4 (onefold), -1/16 (fourfold), -1/36 (ninefold). All other eigenvalues are greater than -1/36. The ground state for N = 7 and q = I is (z) = 36-fold
292
Density Functionals for Coulomb Systems (a revised version of no. 144) COULOMB DENSITY FUNCTIONALS
267
degenerate, and a basis for this eigenspace consists of the determinants (7!)-I"2
det (IS, 2S, 2P,, 2P2, 2P3, f, g) where f and g are any orthonormal functions in the nine-dimensional space M spanned by S, PI, P2, P3, DI, ... , Ds (an orthonormal set for the 3S, 3P, and 3D waves). Let d (f, g) denote the above normalized determinant and let 31/21/i =d(S, DI)+d(D2, D3) +d(D4, Ds). Then 0 -+p with p = pa + Pb and 3
P0(x) = I1S(x)I2+I2S(x)I2+ E 12Pr(x)12, i-I
3pb(x) = IS(x)I2 + E IDi(x)I2.
I-I
Clearly P E .:&' since .' is a ground state.
If T ((p) = T (p ), then there exists a determinant 44 with 0 -p and such that must be a ground state. Therefore 0 = d (f, g) for some orthonormal f, g E M. Thus, (4.11)
If(x)I2+Ig(x)I2=Pb(z).
I claim that this is impossible. Write f = A + D and g = B +d, with A and B being linear combinations of S and the P; while D and d are linear combinations of the D;. Now the S, P, and D waves behave as Ix 1°, Ix I', 1x12, respectively, near the origin. By examining the behavior of (4.11) near the origin we conclude that 3
ID(x)I2 + Id (x )12 = E ID,(x )I2. I
Since all the D, waves have the same radial wave functions, this is really an equality about spherical harmonics The right side of the last equality is spherically symmetric, so the problem is to find two linear combinations F and G of the Y2, such that IF(fl)I2 + IG (fl) I2 = constant > 0.
This is impossible, and the proof is left as an exercise. (It is easily carried out if the following five basis functions are used: xyr-2, yzr 2, 3x2r-2 -1, 3y2r-2-1, with r2=x2+y2+z2.) xzr_ 2,
Remarks. (i) N = 7 is not special; it was chosen for convenience in the proof. (ii) An alternative way of viewing Theorem 4.8 is following. Suppose K + V
has a degenerate ground state, so that the ground eigenspace G is more than one-dimensional. Ib E G is a linear combination of determinants. Consider a perturbation w of v, namely, In first-order perturbation theory, V + A W picks out a subspace g of G as the new ground eigenspace. If g is one dimensional, then g consists of one determinant since the ground eigenspace of V + A W always contains determinants (see Theorem 4.6). Now we ask, 41o G
293
Int. J. Quant. Chem. 24, 243-277 (1983) LIEB
268
and Ikon *po, can w be chosen so that g is one dimensional and g = {41o}? Alternatively, can w be chosen so that min {f wp Illr * p and 0 a G} occurs uniquely for p = po? If so, +&o is a determinant. Theorem 4.8 says that there can be a po such that no w can pick it out uniquely.
Even though T6,(p) > T (p) for some p, Td., still satisfies the variational principle for E(v). Theorem 4.9. For all v e L3/2 +L°°
E'(v)=inf{TT.t(P)+ JPvlpeill.
(4.12)
Proof. Equation (4.12) is equivalent to the following:
E'(v) inf
[K + V]0)14i e `V'N}
= inf {(0, [K + V]i)I i/i E WN, 0 is a determinant)
-E(v). Clearly E'(v)sE(v). Consider the operator -A+v(x). We d e f i n e its "eigenvalues" e1 s e2, ... (here, spin degeneracy is included) by the min-max principle: where inf {(0, [-o+v]0)I4b E H', 11.0 11, = I
and 46 is orthogonal to 40,.... ,
From this definition, it follows by a standard argument that
EN(v)- E e; =inf { E (0i, [-,& +v)0;)I-0 ,, ... , 40N are orthonormal}. 1-,
i-1
(4.13)
But this least infimum equals
are orthonormal,
inf { 00
A; =N}.
0 s A; s 1 and r-,
This is easy to verify. Let oft E `V'N and let y = EA; f;)(f; be its one-particle density matrix (including spin and with the f'-orthonormal. 0 s A; s 1, EA; = N). Then (#, [K + V]I(r) = EAr(f,, [-A+v]f;).
Thus E'(v)_-EN(v). But EN(v)=E(v) by inspection. Remark. This proof gives a formula for E'(v), namely, EN(v).
294
Density Functionals for Coulomb Systems (a revised version of no. 144)
COULOMB DENSITY FUNCTIONALS
269
The situation is complicated, so let us summarize it. TKS is defined only on the set of p's that come from ground states for some v. .4;v has a smaller subset, .sdN, in which p comes from a determinantal ground state, s1., includes. but is larger than, >1N, the set of p's that come from nondegenerate ground states. (Note: By Theorem 3.2 any p comes from a unique v (up to constants). Thus, if p comes from a determinant in a degenerate ground eigenspace, then pit std,'.) On siN' we have sf1 ,
TKS(P) = Td<,(P) = t(P)-
Elsewhere on 4, Td.,(P) > TKS(p) = t(P)-
Thus, there are two choices for (4.8): either TKS or Td t. On 96,v, the complement of std'N, TKS is not defined (but t, Td.,, and T are defined). The preferred functional here is T(p) because it is convex and hence most manageable. On 91
probably strictly less than t(p)
Td.,(p)-at least this is so when T has a
continuous tangent functional (Theorem 3.10). Such points are dense (Theorem 3.14). In any case, since T, T, and Tdet can be interchangeably used in (4.12); it makes no difference which is used as far as E'(v) is concerned.
5. Some Density Functionals That Are Bounds In this section we forego the abstract functional theory of the previous sections
and instead expound a different philosophy. Rather than pursuing "the correct density functional," which seems to be uncomputable, we shall content ourselves here with finding upper and lower bounds to the various quantities of interest in terms of p(x). This latter program can provide rigorous bounds on ground state energies that, while they may not always be extremely accurate, do have a proper place in our conceptual scheme. Some of these bounds will be briefly displayed here; the interested reader is referred to the original papers for proofs. It should be remembered that if one has bounds on two quantities (e.g., T and I; see below) and even if these bounds are optimal, then, in general, the sum of the bounds is not optimal for the sum of the two quantities (e.g., T + 1). A. Kinetic Energy Lower Bound
Lieb and Thirring (LT) [211 (also see Ref. 16) proved (for fermions in three dimensions) that if 0 Hp, then (for all N) T(Ji)?K`(41r)-2isq-2/3
p(x)513dx,
(5.1)
where K` is the "classical" value (3/5)(6a2)213. LT conjectured that (5.1) holds in three dimensions with the (41r)-213 deleted. [Note: Although an analog of (5.1) holds in all dimensions, the corresponding constant is definitely less than
K` in one and two dimensions.] In Ref. 22 (also see Ref. 16) (41r)-2'3 was replaced by
1.496(41r)-2i3.
295
Int. J. Quant. Chem. 24, 243-277 (1983) LIEB
270
K Incidentally, the statement T(O) ? Kq -213 J p 313 for all op, all N, andLssome 2(R3) is equivalent to the following [21]. Let v be any nonpositive potential in and let e I s e 2 s - , be the negative eigenvalues (if any) of -A+ v (x) counting degeneracy, but not counting the q- fold degeneracy. Then -
Ee;? -L J
Iv(x)Isn dx
with K = (3/5)(2/5L)213. B. ,Kinetic Energy Upper Bound
There is, of course, no upper bound for T(O) in terms of p. March and Young (my) [17] proposed that for all p e 3N there is a determinantal 0, with 41 Hp, such that
T(#)sq-ziK J p(x) ("-z)r' dx+
[gyp
1'2 (X))z
dx,
(5.2)
where n is the dimension and K` = -rr2/3 for n = 1. (Compare (5.2) with Theorem 1.2.) They proved (5.2) for n = 1, but their proof for n > 1 has an error. Equation (5.2) for n > 1 is still an open problem. The my construction for n = 1 motivated the construction in the proof of Theorem 1.2. C. Lower Bound for the Indirect Part of the Coulomb Repulsion
Let I' be a density matrix (which may be a pure state, r = ilr)(o) with r --+ p. Let
I(r)=Tr{r
y-
Isi<jsN
1x;-xii-I}
(5.3)
be the Coulomb repulsive energy. The indirect part of this energy, E(r), is defined by
I(t)=D(p)+E(I ),
(5.4)
with
y'
(5.5)
E(r)>-C J P(X)413 dx
(5.6)
D(p)=z JJ p(x)p(y)(x - yl-' being the direct part. In Ref. 23 it was shown that
with C = 8.5. In Ref. 24 this was improved to C = 1.68. The sharp (i.e., best) C in (5.6) is not known, but it is larger than 1.23.
296
Density Functionals for Coulomb Systems (a revised version of no. 144)
COULOMB DENSITY FUNCTIONAI.S
271
It is well known that in any pure, determinantal state, E(F)<0. For other states, E(I) can be positive. Indeed, for any fixed p there is no upper bound for E(I) (see Ref. 24). There is no q-dependence in (5.6) and, indeed, (5.6) holds for all statistics (i.e., C does not depend on statistics). This is explained in Ref. 24. The Dirac
approximation has CDq -'" in (5.6) with C,=3(6/1r)"'/4=0.93, but this q dependence is an artifact of the particular q-dependent determinantal 0 used to evaluate E from (5.4). It should be noted that the bound
I(r)>_D(p)-CJ p`1'
(5.7)
is not convex in p. It is not even positive. These two faults lead to absurd conclusions when the right side of (5.7) is used in Thomas-Fermi-Dirac theory (see Ref. 25). Since F -p is linear,
1(p)=inf{I(r)Ir -p}
(5.8)
is convex in p. In other words, an optimal positive, convex lower bound must exist. Any reader who is devoted to abstract density functional theory, in the spirit of Sec. 3 or (5.8), should try to guess a plausible form for 1(p). (Proving it
is another matter.) It will quickly be seen that 1(p) must be extremely
complicated, and to say that it is "nonlocal" is an understatement. To see this, consider N = 2 and p consisting of two "bumps," p, and p2, very far apart. As long as J p I = f P2 = 1, 1(p) - D (p) -0, independently of p, and p2. But when f p l > 1, J p2 < 1, then t (p) - D(p) depends heavily on p, but not on p2. The reason is that in the former case the two electrons can be far apart in the two bumps; in the latter case the two electrons must partly be close together in the first bump.
A problem that is physically more relevant and that illustrates the hidden complexity of density functional theory is the following problem about induced
dipolar (or Van der Waals) forces raised in Ref. 25. When two atoms are a distance R apart, and R is large, there is an attraction -R -6 (neglecting retardation effects). This attraction comes from the Coulomb repulsion, but it is not a static effect. The atomic dipole moment is almost zero. (There are, in fact, tiny dipole moments, but these are opposite in sign by symmetry, and hence repulsive. They must exist by the Feynmann-Hellman theorem: dE/dR = electric potential at the nucleus. I thank C. Herring for this remark.) There is almost no static dipole moment because to create one would cost a polarization energy ad2. The _z attractive energy is -d2R -' and, if R < a, d = 0 for minimum energy. The cause of the -R-(' energy is more subtle, but it has a semiclassical basis: The electrons in each atom move in phase while maintaining the spherical symmetry about each atom. The energy cost is then ad 4 and the minimum energy occurs when 2ad2 = R Thus, the -R -6 attraction comes from the fact that the electron
297
Int. J. Quant. Chem. 24, 243-277 (1983) LIES
272
cloud cannot be thought of as a simple "fluid." This effect is somehow built into
1(p), but an explicit form of 1(p) that will produce this effect has yet to be displayed.
D. A Variational Principle
E(v), given by (3.5), satisfies (by definition) the well-known variational principle
E(v):s (0, H4)
(5.9)
Can an upper bound for E(v) be given in terms of p alone? If (5.2) were true, then, for any p E .ON,
E(v) s right side of (5.2)+D(p)+J VP.
(5.10)
[See the remark about E(I') for determinants in Sec. 5C.] An upper bound for E(v) can, indeed, be given in terms of the one-particle density matrices y(z, z') as follows [26]: Let y be any admissible one-particle density matrix (0 <_ y s 1, Tr y = N). (Note: y includes spin. It was called i in Sec. 2). Then
E(v)sTry(-A+v(x))+'2 K2(z, z')Ix -x'I-' dz dz',
(5.11)
where f dz = Eo f dx and K2(z, z') = y(z, z)y(z', z') - Iy(z, z')12.
(5.12)
The form (5.11) is well known if y came from a pure state ry)(I(i with 0 being
a determinant. The point about (5.11) is that it holds for all admissible y. Incidentally, the minimum of (5.11) over all admissible y occurs when y comes from a determinantal 0. In other words, the best Hartree-Fock function minimizes (5.11), but (5.11) is interesting precisely because this HF function is unknown.
E. Thomas-Fermi Theory This theory (see Ref. 25 for an exposition) does not yield bounds and therefore
does not properly belong here. However, it illustrates the usefulness of the bounds in Secs. 5A-C. The TF functional is it
F(P)=Kcq-2/3 r p5/3+D(P)+ j
vP,
(5.13)
while the TF Weizsaecker functional '(p) is the right-hand side of (5.10). If -C f p°" is added to the right-hand side of (5.13), the result is TF Dirac theory.
298
Density Functionals for Coulomb Systems (a revised version of no. 144) COULOMB DENSITY FUN('TIONAIS
273
The TF energy for N particles is defined by ETF = inf { 'TF (p )I j p =N1
(5.14)
and similarly for ETFw and ETI I).
Now suppose that v is an atomic or molecular potential, that is,
v(x)=- E ziIx-R;I-',
(5.15)
with the z; > 0. It is a fact [ 14] that under the scaling z; - Az1 and N -. AN, as A -0 CO
E'F/E(v) -.1,
(5.16)
where E(v) is the true ground state energy. Note that (5.16) also holds if E'F ETI,v or E'FD (see Ref. 25). is replaced by Thus we see that if the conjecture in Sec. 5A holds, then, combining (5.1) with (5.7), TFD theory is a lower bound that is asymptotically exact. Similarly, if (5.2) holds, then, as remarked in Sec. 5C TFW theory is an upper bound that is asymptotically exact. F. Two-Body Density Matrices
If one is willing to go beyond the one-body density p or one-body density matrix y and consider the two-body reduced density matrix y(2), then E(v) is directly and exactly expressible in terms of y121, since H. has only one- and two-body terms. The problem is that it is very difficult to decide when a given yi2) is. in fact, the reduction of an admissible N-body density matrix r. This is called the N-representability problem and it has not been solved. (This is to be compared with the fact that there is a simple necessary and sufficient condition for a one-body y to be N-representable; see Sec. 2). It is possible, however, to find some necessary conditions and some sufficient conditions for .12) to be N-representable. Using these, bounds on E(v) can be derived. Since this approach is outside the scope of this article, we refer the reader to the excellent review of Percus (27).
Appendix: Proofs of Theorems 1.3, 3.3, and 4.4
The following proof of Theorem 1.3 is due to H. Brezis (pivate communication.)
Proof. For simplicity of presentation we take N = 2 and q = 1 (no spin). Therefore we have ,/i(x, y) and lie
F(z) _ (J 14,(x, y)12 dy)
_
[p(x)/2]'/
that ii -0 in H' (R x R'); that is, 0 -. r, and V O. -of in L2. We want to show that F. -.F in L2(R3) and VF,, - VF in L2(R'). The former is trivial:
299
Int. J. Quant. Chem. 24, 243-277 (1983)
274
LD -H
By the Schwarz inequality, ),)12dy JI+G(x,y)I2dy)
\J
-i J
(x,Y)dy
+
dx J10.,(x,Y)-1(x,Y)12dy.
J
and the right-hand side converges to zero. The proof that VF,, -* VF is the difficult one. It is sufficient to prove conver-
gence for some subsequence n,, j = 1, 2..... (If
then there is some
subsequence and some e > 0 such that IIVF,,, - VFII > E. But then this subsequence
clearly does not have a subsequence which converges to VF.) Now since in H' there is some subsequence and some function G E H' such that I
(x, y )l ts G(x, y) and
(x, Y )I <- G (x, y),
a.e. in R6. (The proof of this fact is the same as the first half of the proof of the Riesz-Fischer lemma that L' is complete.) Henceforth, we shall replace n, by n. We shall also assume, for simplicity, that F. (x) and F(x) > 0 for all x (otherwise, an approximation argument can be used). Now (x)+c.c. with
and
As we saw, F. -'F in L2, so, by passing to a subsequence, we can assume F (x) -F(x) a.e. Furthermore, a.e. y); sG(x, y)2E L1(R6).
By passing to a subsequence we can assume Vift -* V O and c,, i,(< a.e. Thus, by dominated convergence, B - B in L' (R6). For this subsequence I B (x, y) B (x, y )l - 0, a.e. (R6). Then, for a.e. x, IB (x, y) - B (x, y )I --> 0 a.e. y. Thus, by dominated convergence y)-B(x, y)I-'O, a.e. x. In other words, for some subsequence, VF (x) - VF(x ), a.e. Finally, we note that, by the Schwarz inequality, J
Y )I2 dy 5 J G(x, Y)' dy =C(x )2.
Since C is a fixed L2 function, VF -. VF in L2 by dominated convergence.
300
Density Functionals for Coulomb Systems (a revised version of no. 144)
COULOMB DENSITY FUNCTIONAL S
275
Proof of Theorem 3.3. Let Oi (with be a minimizing sequence for F(p). The ,(r; are obviously bounded in H'(R3t'), so, by the Banach-Alaoglu theorem, there is a ,y a H'(R3N) such that Oi - 0 weakly in H' (R3N). Obviously,
,/i has the same symmetry as the it/. It is well known that under weak limits positive quadratic forms decrease. Thus
F(p) = lim (4,;, Ho,/ii) ? (0, Hod ). If we can show that '-+p, we are done. To do so it is sufficient to prove that ,y, -,6 strongly because if i H p we have, by the easy part of Theorem 1.3 that P
1/2
=p;112 ->p" f /2 InL2 ,sothat p=p.
Strong convergence will be proved by showin, that f =1. Let S be the characteristic function of some bounded set in R3 . By the Rellich-Kondrachov theorem [28] there is a subsequence (which can be chosen independent of S) of the ,k/ such that St/i1 converges strongly (in L2) to Sift. Pick e > 0 and let x be the characteristic function of a bounded set in R3 such that
e>.fp(1-x)° f Ii1121[1-x(x)]. But
E[1-x(xr)]a. 1-S, where S = nx (x; ). Thus, f Iii,12S 2t 1- e. Since l,/r; l2S - f It' I2S, we have that f
10I2
f 1012S?1-e foralle>0. Remark. The symmetry of ,/' was not needed in this proof provided one generalizes definition (1.6) to N
P(X)=EE f 10(Z11
(A.1)
,
The following proof of Theorem 4.4 is due to B. Simon (private communication). It is closely related to the proof of Theorem 3.3 just given.
Proof. Without loss, replace H o by h 2 = Ho+ 1 in the definitions. h -' is a g;alimg exists, and bounded operator. We can assume that g Tr Tr h 2 s g + 1 /n with r. -+p,,. Thus, y = h I' h is uniformly bounded in the trace norm. The dual of the compact operators, com, is the trace class operators r, and y e t takes A e com into Tr yA. A sequence y E t converges to y e t, in the weak* topology, if and only if Tr Tr -yA for all A E corn. The Banach-Alaoglu theorem states that a norm-closed ball of finite radius in t is compact in the weak* topology. For us this means that there exists y with
Try0 and therefore lim Tr y ? Tr y. Also, y obviously has the correct (Pauli) symmetry. If we can show that r - h -' yh -' (which is in trace class)
satisfies r -p, we are done. To do this we shall show that if r - p', then f (p, - p')f - 0 for any f e L. This would mean that p -p' weakly in L. But since p;, -p in L', p'= p.
301
Int. J. Quant. Chem. 24, 243-277 (1983) LIES
276
As in the proof of Theorem 3.3, for any e >0 there is a X (=characteristic function of a bounded set in R3) such that
J P(1 -X)<e Since p
P'(1 -X)<e.
and
p, j p (1- X) < e for n sufficiently large. If
0n(xI. .XN)=Te(x1. .xN;xl..
.xN)
(after summing on spins), and similarly for 46, we have (as in Theorem 3.3)
J
J*/)(1_S)<E
and
where s = nX (x; ). In view of this, it is sufficient to show that
f. P-
JOP
with
P=SY-f(x1).
Let P = P(x 1, . . . , xN) be any bounded functions of compact support and let Mp be the operator (in L2) of multiplication by P. It is a fact that A,, = It -'Mph -'
is compact. (This is essentially the same as the Rellich-Kondrachov theorem used in Theorem 3.3.) Therefore
Tr I'Mp = Tr yAp - Tr yAp = Tr I'Mp.
Acknowledgment This work was partially supported by U.S. National Science Foundation grant
No. PHY-7825390-A02. This paper is a revised version of a paper with the same title that appeared in Physics as Natural Philosophy : Essays in Honor of Laszlo Tisza on his 75th Birthday, H. Feshbach and A. Shimony, Eds. (M.I.T. Press, Cambridge, 1982), pp. 111-149.
Bibliography (1) L. H. Thomas, Proc. Camb. Phil. Soc. 23. 542 (1927). [2] E. Fermi, Rend. Accad. Naz. Lincei 6.602 (1927).
[3) P. Hohenberg and W. Kohn, Phys. Rev. B 136, 86411964). [4] M. M. Morel[, R. G. Parr and M. Levy, J. Chem. Phys. 62, 549 (1975). [5] R. G. Parr, S. Gadre and L. J. Bartolotti. Proc. Natl. Acad. Sci. USA 76, 2522 (19791. [6) R. A. Donnelly and R. G. Parr, J. Chem. Phys. 69,4431 (1978). (7] H. Englisch and R. Englisch, "Hohenbcrg-Kohn theorem and non-v- representable densities," Physica A, to he published. [8] T. L. Gilbert, Phys. Rev. B 6, 211 (1975). [91 J. E. Harriman, Phys. Rev. A 6, 680 (19811.
302
Density Functionals for Coulomb Systems (a revised version of no. 144)
COULOMB DENSITY FIJN(TIONALS
277
[10] M. Levy, Proc. Natl. Acad. Sci. USA 76, 6062 (1979). [11] M. Levy, Phys. Rev. A 26, 1200 (1982). [12] S. M. Valone, J. Chem. Phys. 73, 1344 (1980); ibid. 73, 4653 (1980).
[13] A. S. Bamzai and B. M. Deb, Rev. Mod. Phys. 33, 95 (1981). Erratum, 33, 593 (1981). [14] E. H. Lieb and B. Simon, Adv. Math. 23, 22 (1977). See also Thomas-Fermi theory revisited, Phys. Rev. Lett. 31, 681 (1973). See also Refs. 16 and 25. [15] M. Hoffmann-Ostenhof, and T. Hoffmann-Ostenhof, Phys. Rev. A 16, 1782 (1977). [16] E. H. Lieb, Rev. Mod. Phys. 48, 553 (1976). [17] N. H. March and W. H. Young, Proc. Phys. Soc. 72, 182 (1958).
[18] M. Reed and B. Simon, Methods of Modern Mathematical Physics (Academic, New York, 1978), Vol. 4. [19] S. Mazur, Studia Math. 4, 70 (1933). [20] R. B. Israel, Convexity in the Theory of Lattice Gases (Princeton U.P., Princeton NJ, 1979).
[21] E. H. Lieb and W. E. Thirring, "Inequalities for the moments of the eigenvalues of the Schrodinger hamiltonian and their relation to Sobolev inequalities," in Studies in Mathematical Physics, E. H. Lieb. B. Simon, and A. S. Wightman, Eds. (Princeton U.P., Princeton, NJ, 1976). See also Phys. Rev. Lett. 687 (1975); Errata, 35. 1116 (1975). [22] E. H. Lieb, Am. Math. Soc. Proc. Symp. Pure Math. 36, 241 (1980). [23] E. H. Lieb. Phys. Lett. A 70, 444 (1979). [24] E. H. Lieb and S. Oxford. Int. J. Quantum Chem. 19, 427 (1981). [25] E. H. Lieb, Rev. Mod. Phys. 53, 603 (1981); Errata, 54,311 (1982). [26] E. H. Lieb, Phys. Rev. Lett. 46, 457 (1981); Erratum. 47, 69 (1981). [27] J. K. Percus, Int. J. Quantum Chem. 13, 89 (1978). [28] R. A. Adams, Sobolev Spaces (Academic Press, New York, 1975). [29] W. Fenchel, Can. J. Math. 1, 23 (1949). [30] W. Kohn and L. J. Sham. Phys. Rev. A 140 1133 (1965).
Received October 19, 1982
Accepted for publication March 11, 1983
303
Commun. Math. Phys. 92, 473-480 (1984)
LOfi1munications in Commun. Math. Phys. 92, 473 -480 (1984)
Mattlerrtatk:el Pwsics © Springer-Verlag 1984
On Characteristic Exponents in Turbulence Elliott H. Lieb* Departments of Mathematics and Physics. Princeton University, P.O. Box 708, Princeton, NJ 08544, USA
Abstract. Ruelle has found upper bounds to the magnitude and to the number of non-negative characteristic exponents for the Navier-Stokes flow of an
incompressible fluid in a domain 0. The latter is particularly important because it yields an upper bound to the Hausdorff dimension of attracting sets. However, Ruelle's bound on the number has three deficiences: (i) it relies on
some unproved conjectures about certain constants; (ii) it is valid only in dimensions > 3 and not 2; (iii) it is valid only in the limit S2-ca. In this paper these deficiences are remedied and, in addition, the final constants in the inequality are improved.
Ruelle [1] has derived upper bounds on the magnitude and number of nonnegative characteristic exponents of the Navier-Stokes equation for the flow of an incompressible fluid in a domain Qe IR°. The bound on the number, N(u) [defined
in (42)], is particularly interesting because it leads to an upper bound on the Hausdorff dimension of a compact attracting set [I, Corollary 2.3]. Unfortunately, the bounds in [I] on N(p), unlike those on the magnitude, have certain deficiencies which are
(i) They rely for their validity on some conjectured, but as yet unproved, relations between the sharp constants in two known inequalities. (ii) They are valid only for d > 3. (iii) Because Weyl's asymptotic formula for the eigenvalues of the Laplacian in 9 is used, the inequalities are not valid for any fixed Q, but only in the limit 0-c-0. In this paper a different proof of Ruelle's inequality for the number will be given so that the above three deficiencies are remedied. The result is contained in Eqs. (40)-(43).
Let v : Q IR° denote a solution to the Navier-Stokes equation, and let i
? 1A2 > ...
be the characteristic exponents corresponding to a probability
measure g(dv) on the space of solutions that is ergodic with respect to the Navier*
Work partially supported by U.S. National Science Foundation grant No. PHY-8116101-A01
305
Commun. Math. Phys. 92, 473-480 (1984) E. H. Lieb
474
Stokes time evolution. Ruelle shows [1] that for all n - I nd
n
µ;<_<_-d
<e,>
-d<E.>
The brackets < - ) denote average with respect to p, E. _
(1)
e;, and the e. = e;(v) are
ordered such that e, 5 e2 5 ... and are the eigenvalues of the Schrodinger operator
H = - vd - w(x)
(2)
with Dirichlet boundary conditions on Q. Here, v is the kinematic viscosity and w(x)z0 with
w(x)2=[(d-1)/4d] Y(av;/ax;+avJ/ax;)2=[(d-1)/2vd]a(x).
(3)
The quantity t(x) is the rate of energy dissipation per unit mass in the flow v. In (2), (3) and henceforth, explicit dependence of the various quantities on v is understood but not explicitly indicated unless necessary.
One might try to take additional advantage of the fact that divv=0 but, as in [1], we shall merely assume that w is some given non-negative function. It will, however, be assumed, as in [1], that wE L'
. d, 2(12)
(4)
Remark. The definition (3) has a factor (d- 1)1d, which is an improvement over
that in [1]. The reason is the following : Ruelle starts with an operator on L2(Rd)®W given by Jr = - vdS;,+ Wa(x), where W1,(x) is the d x d symmetric matrix W,4x)=(at;;1ax;+av;/ax;)12. Ruelle notes that the eigenvalues of Jr will satisfy (1) if w(x) in (2) is the largest eigenvalue of the matrix Wj(x). This he estimates by (Tr W2)"2, and this leads to (3) without (d-1)/d. Since divv=O, however, Tr W = 0. If A, > A2 > are the eigenvalues of W, then Tr W 2 =Y A, and
TrW=YA,. But (d-1)i 2
/4
``2
=22, and hence (d-I)TrW2>_dA;. In
2
addition to the condition divv=0, Jr is supposed to be restricted to the space of divergenceless functions. This restriction might improve (1) but, as in [1], it will not be used here.
The domain QER" is assumed to be an open set of finite volume 101; boundedness is not .required. Condition (4) insures that the quadratic form on H,(Q), defined by
Q*)=V1ivol2- Jwo2,
(5)
is bounded below and thus defines H as a self-adjoint operator. (Integrals, here and
henceforth, are over Ii) For our purposes, self-adjointness is not important; the only important consideration is the max -min principle which can be used as a definition of the e; :
inf
Q(4;,4,,).
(6)
where is any L2 orthonormal set in H,(Q). It is, in fact, the right side of (6) that enters in the derivation of the bound (1).
306
On Characteristic Exponents in Turbulence 475
Characteristic Exponents in Turbulence
The goal is to find upper bounds on the following two quantities
y20.
E(y)= Y Ie;IY,
(a)
e,60
(7)
E. itself for fixed n.
(b)
(8)
It is a consequence of (1) that for y> I lu Sd
I<e;>I'
Y_
<e,)<0
Y,30
d<E(y)>
(9)
This is Karamata's theorem which, more generally, states that when f :IR-+IR is convex and non-decreasing then (1) implies that
df(-e;))
f(j;)_d
(10)
for all n. If, in addition, f(t)=0 for t<0, then Yof(!t,)Sd <e,Y
i!S
f(-e;)1.
(I1)
[Actually, Karamata's inequality gives the left-hand inequalities in (9}{11). The right-hand inequalities come from Jensen's inequalityf().] It is (9) that gives information about the magnitude of the p,. The bound used in [1] [except for the factor (d- 1)/d in (3)] was E(y)
d;2f
w(x)Y'd,zdx.
(12)
The present knowledge about (12) is the following: (1) LY, d < 00 for y > i (d =1), y > 0 (d = 2), y Z 0 (d 3). No such bound exists for y < 1(d =1) or y = 0 (d = 2). The case y = 2d =1 does not seem to have been settled.
(The claim in [2] that Lti2 t < x is not justified.) Bounds on L,, were first given by Lieb and Thirring in [3] and on L,., for y >0 (d=2.3) and y > 1(d (d= in [2]. Bounds on Lo,d, d>-3, were first given by Cwikel [4], Lieb [5, 6]. and Rosenbljum
[7]. The best upper bound for Lo,3 is in [6], namely 0.0780=4n-23-3r2SLo,3 <_0.1156. The lower bound is from [2, Eq. (4.24)]. Recently, by a simpler method. Li and Yau [8] derived upper bounds for Lo.d, d >- 3 which they claimed was better
than that in [6); unfortunately a numerical error was made in [8] and their bound for Lo, 3 is three times larger than that in [6]. (2) The sharp constant LY.d in (12) cannot depend on n, i.e. LY. d(Q)=L, d(IRd).
To see this, assume that OeQ and, given w on Rd, consider wa(x)=c2i0cx) on 0. Then let c-+ oo. This situation is in contrast with the Inl dependent bound for E, to be derived later. (3) There is a natural "guess" for LY d given by the semiclassical formula (2n)-dIf dpdxlvp2-w(Y)1L`rdt.-d;2 1 w(x)''d'2dx (13) E(y) with Ia1_. =max(0, -a). An easy integration gives
Lr=2
en a;zl(y+d
1)/!(y+I+d/2).
(14)
(4) It is a fact [2] that Li.d
(15)
307
Commun. Math. Phys. 92, 473-480 (1984)
E. H. l.ieb
476
In [2,3] it was conjectured that L1., = L , for d 2t 3. It is known [2] that for each d< 7 there is a y, > 0 such that LY , > VI., when y < y,. When d =1 or 2, )', > 1. It is also known [9] that L,. L`y for y 3/2. In fact [9] the ratio Ra(y) = L}. ,, L'7.d is monotone non-increasing in y; thus if R,(yo)=1 for some yo, then R,(y)=I for all y>yo. Glaser et al. [10] have shown that Lo,d>Lo,d for d? 7. They also evaluate Lo,4 exactly (it is a Sobolev constant) provided it, is restricted to be spherically symmetric. For related results see [ I I]. 1
(5) Inequality (12) for y=I is equivalent [2] to n
_ f IV4i(x)12 dx ? K, f pm(x)' - z;d dx
(16)
i= 1
where the (0,) is any L2 orthonormal set in H'(Rd) [or Ho'(Q)] and co"W
IOi(x)12.
(17)
The sharp constants in (12) and (16) are related by L1,,=[d/2K,]d!2 (1
+d/2)-, -ere
(18)
[Note: If it is specified then the sharp constant in (16) may depend on n, i.e. K,(n). K,, the sharp constant in (16), (18) is defined to be sup.K,(n).] Corresponding to L,., in (14) there is a classical value Kd given by (18):
Ka=4rrdf(1 +d/2)2'd/(2+d).
(19)
By (15), K,__<_Ka.
An inequality related to (16), and which will be used later in the event that K,Kan' + 2r4IQI - 2.d
(20)
(The strict inequality in (20) is, in fact, implied by the proof in [8].) Before turning to our estimate for E. let us make a few additional remarks about (12).
(a) Combining (3), (9), (12) we see that the right side of (12) is suitable for passing to the "infinite volume" limit, i.e. in some vague sense it is proportional to the volume. The upper bound we shall obtain later for the quantity introduced in
[I]'
N(w) = smallest it such that E.>0,
(21)
will also have this extensivity property. By (1), dN(w) is related to number of nonnegative characteristic exponents and an upper bound on N(w) will yield a bound on the number of non-negative characteristic exponents [see (43)]. ((3) The bound on N(w) in [I] relied on the fact that Lo d < :r (which is true if
and only if d>3) and on the conjecture that Ltd1, the best bound published so far [6] for L1 , is
L1 ,5(6.844)L; ,=0.04624,
(22)
and this exceeds L'0, =0.01689. However, the bound can be improved slightly to 0.04030 [sec (51) below].
308
On Characteristic Exponents in Turbulence
Characteristic Exponents in Turbulence
477
(y) In