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