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 smoot