ANNALES SOCIETATIS MATHEMAT1CAE POLONAE Series I : COMMENTATIONES MATHEMATICAE X I V (1970) R O C ZN IK I POLSKIEGO TO W A R ZYSTW A MATEMATYCZNEGO
Séria I : PRACE MATEMATYCZNE X I V (1970)
A. P
e l c z y n s k i(Warszawa)
O n Sikorski’ s trace formula
The purpose of this note is to present a proof of the following result (the terminology and notation are explained below)
T
h e o r e m. Let A, B , G be 2-absolutely summing operators on a Banach space X . Then
( 1 ) Ьгво{А) = tvA B (C) = Ьхал(В ).
Formula (1) was conjectured by Sikorski [7] and was proved by Grothendieck [1]. Sikorski originally assumed that the operators A , B, G are integral. Since every integral operator is 2-absolutely summing (cf.
[в]? [3], [6], [2]), our assumption is less restrictive.
We recall that a linear operator N : X -> Y (X , Y — normed linear spaces) is nuclear if it has a nuclear representation
X = ^ xn ® yn, П
i.e. there are sequences {x*) of bounded linear functionals in X* — the dual of the space X , and (yn) of elements of У such that
1Й1 112/J < + °°,
( 2 )
N(x) = J ? Яп(х )Уп f° r x in X . П
We put
q(N) = g.l.b. £ \K\\ \\Vn\\i n
where the g.l.b. is taken over all nuclear representations of the operator N.
If N = x* ®уn: X -> T is a nuclear operator and A : T X П
is a bounded linear operator, then we write tr^v (A) = x l ( A { y n)).
П
It is well known [8] that the quantity tr ^ J .) depends only on N;
120 A. P o lczy n ski
it does not depend on a particular choice of a nuclear representation of N.
We have
(3) |trlV( i ) | < S(J)||i[|.
Next we recall that a linear operator A : X -» Y { X, Y — normed linear spaces) is said to be 2-absolutely summing if for some (equivalently, for all) isometrically isomorphic embedding JA from X into a space C(Q a ) there exist a probability measure mA on QA and a bounded linear operator SA : Li(mA, QA) -> Y such that
-d $
Л^
m A J Л7
where : C(QA) -> L 2(mA , QA) is the formal identity map which assignes to a continuous function / in C(QA) its mA-equivalence class in L 2(mA, Qa ) (cf. [5], [2], [3], [6]. In [2] and [3] instead of “ 2-absolutely summing” the term “ 2-integral” is used).
By G(Q) we denote the space of scalar-valued ( = real-valued or complex-valued) continuous functions on a compact Hausdorff space Q.
If m is a probability measure on Q ( = a regular non-negative Borel measure of total mass 1), then L 2(m, Q) denotes the Hilbert space of m-equivalence classes of square-integrable scalar-valued functions on Q).
Observe that the set of all 2-absolutely summing operators from X into Y forms a linear manifold in the space of bounded linear operators from X into Y (cf. [6], [5]). Furthermore note that the assumption of our theorem implies that the numbers Ь твс {А),Ь та в (С),Ь тса (В) are well defined, because the composition of 2-absolutely summing operators is a nuclear operator. More precisely we have:
L
e m m a1. Let JB = S в I тв Jв • X — > Y and G — So I me Ac : Z — > X be 2-absolutely summing operators, where X , Y, Z are normed linear spaces,.
J b : X - > G { Q
b) and J o: Z - + C { Q c ) are isometrically isomorphic embed
dings, I mB: G (Q
b) - > L 2(mB, QB) and I mc: C(QG) L J mc , Qc) are formal identity maps, SB- L 2(mB, Q
b) Y and Sc : L 2(mc, Qc) -> N are bounded linear operators. Then the operator
G(Qc) -> Lz(mB,QB) is nuclear and
(d) < 1 ( 1 твТ в 8 с 1 тс) < л а д .
This lemma is an immediate consequence of [4], p. 56, Satz 2 and the fact that the operator I mB(JBSc): X 2(me , Qc) L 2(mB, Qв) is a Hilbert-Schmidt operator with the Hilbert-Schmidt norm equals to the operator norm of Sc (cf. [4], p. 56, Satz 1).
Our next lemma reduces the proof of the theorem to the case where
X is a G(Q) space. For a given Banach space X, K* denotes the unit ball
SikorsM ’s trace formula 121
of the dual space X* equipped with the weak-star topology and J: X -» C(K*) denotes the eannonical isometrically isomorphic embedding defined by
J{x){x*) = x*(x) for x z X and for x*eK*.
L
e m m a2. Let A = SAI mAJ , В = SBI mB J ? C = ®cImcJ be 2-absolute- ly summing operators on a Banach space X . Then
A J 8 A ^-mA ) B = C = J8BI mc are 2-absolutely summing operators on C(K*) and
tr.BC (A) = tvè ô {Â).
P r o o f . Observe that A = I mj J j- , where J j is the identity operator on C(K*), mA = m j and S j — JSA- Hence the operator A is
2-absolutely summing. The same argument shows that the operators.
Ê and C are
2-absolutely summing.
To prove the second part of Lemma
2we observe first that Lemma
1implies the nuclearity of the operator
N = SBI mBJSc I mo: C(K*) - + X . Hence it has a nuclear representation
N = ^ m n®xn -, mne(C{K*))* , xnc X {n =
1,
2, . . . ) . П
Since JN — BC and NJ = BG, we get the nuclear representations.
BC = 2 mn®J{ xn)l BC = 2
П П
where J*: (C(IT*))* X* denotes the adjoint operator of J. Thus, using the obvious identity JA ~ ÂJ, we get
t r j B c ( ^ ) = 2 (J*(тп))(ЛЫ ) = 2 mn(JAiXn))
n n
= 2 mn(A J (Xn)) = tr^
0(
2).
n
This completes the proof of Lemma
2.
In the sequel we shall assume that X = C{Q). In this case if A : G(Q) -> C{Q) is a
2-absolutely summing operator, then there is a probability measure on Q, say mA, and a bounded linear operator SA: L 2(mA , Q) -> C(Q) such that A = 8AI mA (we take for JA the identity operator on G(Q)). In fact, we have the following representation lemma:
L
e m m a3 . I f A = SAI mA: C(Q) -> C(Q) ia s 2-absolutely summing
122 A. P e lc z y n sk i
operator, then there exists a set of indices stf, an orthonormal system {aa)ae^
in L 2(mA, Q) and a family {aa)ae^ in C(Q) such that
(5) ll^ll = SUP ( ^ \йаШ2)У2’
<6) A ( f ) = f fd a dmA • aa.
аел/Q
P r o o f . Let (aa)aej# be an arbitrary complete orthonormal system in L 2{mA l Q)-
Then
I mA( f ) = / / « « dmA-aa for feC(Q).
aejtfQ Thus
A { f ) = SAI mA (f) = f fd a dmA ’ 8A{aa).
W e put cia = SA (««) for azsé. Then clearly we have (
6). Finally we have
\\S a \\ = sup sup I V tadaiq)I = sup sup I У y<xa( d = sup ( V |ae(ff)|*)%.
2\tal* = l aeQ aejtf ' VQ 2\ta\'^ 1 aTS'
This completes the proof.
Our next lemma is a slight modification of [4], p. 56, Satz
1.
L
e m m a4 . Let m be a probability measure on a compact Hausdorff space Q. Let (bp)peS$ be an orthonormal system in the space L 2(m, Q) and let (Cy)ye<e be a family of functions in C(Q) such that
(7) sup QT
1M g ) l 2)*' = s < + ° ° -
q e Q y e <jf
Then
( 8 ) ( 2 2 \ f cy(q)b/3(q)dmj2jy2 < *•
P e @ y e # l Q
P r o o f . Let Z2(^) denote the Hilbert space of scalar-valued functions t = {ty)ye<# such that |[t|l2 = £ \ty\
2< -j— oo. It follows easily from (7) that
y e #
the formula
8(1) = £ t yCy for teZa(^).
y e #
defined a linear operator from Z2(^) into C(Q) with the norm \\S\\ = s.
How, if I m: G(Q) -> L 2{m, Q) denotes the formal identity map, then,
Sihorslci’s trace formula
123
by [4], p. 56, Satz
1, the composition I mS is a Hilbert-Schmidt operator and its Hilbert-Schmidt norm is < ||$||. Thus, if (ey)ye<Sf denotes the unit- vector basis in 1г(Щ, then using the Parseval identity we get
* = IISII» (y>„S<e,)II*)* = Р Л Г )* 4
У s ' # y e #
V e V P e Æ Q
This completes the proof of the Lemma.
Now we are ready for the main lemma.
L
e m m a5 . Let В = 8в 1тв and G — 8 c l mc be 2-absolutely summing operators on G(Q) and let (according to Lemma 3) for fcC(Q)
B (f) = ^ f fbp dmB -bp', 0 ( f ) = Jfcydmc-Cy,
P e @ Q y t V Q
where (bp)p€& and (cY)Ve<g are orthonormal systems in L 2(mB,Q) and in L 2(mc , Q ) respectively, bpeC(Q) for (SeB and cyeC(Q) for y**#, and
sup 2 Ib« («)l
2= ll-Sull2; sup 2 lcr
( « ) ! 2= ll*cll2.
P e Æ Q* ® y e #
Then for every bounded linear operator A : C(Q) -+C(Q) we have
(9) tr BO(A) = E E S iybpdmB J A(bp)cvdm0 .
y e # Pe& Q Q
P r o o f . It follows from Lemma 4 that the double series in the right- hand side of (9) is absolutely convergent because for every q in Q we have
у к^(ь„))(8)|г < m r y ; i m «) i s< над» и р .
ре@ РеЗЗ
Hence for e >
0one may choose finite subsets of J* and of so that
(
1 0) 2 E S * pdmB j A(bp)Cydmc—
y e # PtSSQ Q
- E E S cybfsdmBj A(bp)cydmc \< e,
Y*V0 fic# 0 Q Q
(
1 1) sup . 2 |Ь/?(з
,) 1<'- sup |Cy((Z)l
2<C £2.
Pe£ÿ\&0
y e # \ # о124 A. P e lcz yn sk i
Let us set for feC(Q), S b 0 (/)
Next put
f fb(}dmB-bp;
P e
&0
QВо Вв^1тВ1 It follows from (11) that
So0 (/) = f fCydmc-Cy.
y^o Q
Go = SoJ.. 0 mc ‘
(12) ||$ b — $j?0ll < £j \\Bo~ $ c0ll < £-
Next observe that for a fixed bounded linear operator A : G(Q) -> C(Q) the quantity trBa(A) is a bilinear form on the linear manifold of all 2-absolutely summing operators on G(Q). Indeed, if mn®fn and U Pn® 9n
П П
are nuclear representations for operators BC1 and BC2, respectively ( В , С Х, С 2 — 2-absolutely summing operators on C(Q)), then ^
71
+ 2 Pn®9n is a nuclear representation for the operator В(СХ+ С 2). Hence П
tr щс 1+ с 2){А) = ^ m n( A { f n)) + ] ? p n(A{gn)) = tTBOl{A) + Ьт Bc2{ A ) .
A similar argument shows that tr^ + b y c (A) = t r ^ (A) + t r ^ e (A).
Thus, observing that ||# b 0|| < \\S b \ \ and using inequahties (3), (4) and (
1 2), we get,
(13) |tr^o(A) — trj5oCfo(A) I < \^(B- bq ) o {A)\ + |trBo(C'_r
70)(A)|
< ( ^ ( ( Б - Б о)С) + д(Б
0( 0 - О о)))||А||
< ||A|| \\8o\\ I l ^ - ^
0II+||A|| \\SBo\\ \\S o - S c 0\\
^\\A\\(\\8c \\+\\SB\\)e.
(observe that the operators C0 and B0 are finite dimensional and therefore they are
2-absolutely summing and nuclear).
Since the sets and &0 are finite, the formula
B0G0(f) = I A J > ydmc f cyb[idmB-bp for feG{Q) У€<&0 Q Q
gives a nuclear representation of the operator B0C0. Hence (14) tr£ocf
0(A) = £ £ f A ( b fi)cydmc f tvb^dmB.
y*V QPeâfQ Q Q
Combining (
1 0), (13) and (14) we get
I tre e (A) — S S I A{bp)Cydm0 f CybpdmB\ < e(l+||A||(||tfB||+ ||£ c ll))-
ye''£
Q QSikorski’s trace formula
125
Thus, letting e tend to zero, we get (9). This completes the proof of the lemma.
P r o o f of t he t h e o r e m . Let A : C{Q)-^C{Q) be a 2-absolutely summing operator satisfying conditions (5) and (
6) of Lemma 3. Let В and G be as in Lemma 5. Then, combining (
6) and (9), we get
(15) Ьтв с (А) - I I S tvbpdmB J f bpaadmA -a^Cydinc
yetf РеШ Q Q « 6 ^ Q
= 'У , У ! У ! f CybpdmB fbpdadmA f aaCydmc .
PeâS Q Q Q