Annals of Mathematics

Two Characterizations of Real Hilbert Space

Author(s): Shizuo Kakutani and George W. Mackey
Source: Annals of Mathematics, Second Series, Vol. 45, No. 1 (Jan., 1944), pp. 50-58
Published by: Annals of Mathematics
Stable URL: http://www.jstor.org/stable/1969076 .
Accessed: 17/11/2014 20:50

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Annals of Mathematics is collaborating with JSTOR to digitize, preserve and extend access to Annals of
Mathematics.

http://www.jstor.org

This content downloaded from 192.231.202.205 on Mon, 17 Nov 2014 20:50:45 PM

All use subject to JSTOR Terms and Conditions
 ANNALs OF MATMEMATICS
Vol. 45, No. 1, January,1944

TWO CHARACTERIZATIONS OF REAL HILBERT SPACE

BY SHizuo KAKUTANII AND GEORGE W. MACKEY

(Received September30, 1942)

Introduction
Let X be a real Banach space, let 9Zbe the ringof all continuouslinear2trans-
formationsof X into itselfand let 2 be the latticeof all closed linearsubspaces
of X. Eidelheithas shown [1]3 that X is determinedup to an isomorphismby
the abstractringof which9Z is a realizationand one of us has proved [2] the
corresponding theoremfor 2. These theoremssuggestamong othersthe prob-
lem of findingwhat special propertiesserveto characterizethe ringsand lattices
of various particularkinds of Banach spaces among those of general Banach
spaces. It is the purpose of this paper to give a solution of the problemfor
spaces isomorphicto generalizedreal Hilbert spaces (i.e., real Hilbert spaces
which need not be separable on the one hand nor infinitedimensionalon the
other).
Our principalresultis to the effectthat in orderto be able to conclude that a
Banach space X is isomorphicto a generalizedHilbertspace it is sufficient to
knowthat eitherthe ring or lattice of X admits an involutaryanti-automor-
phism with a certainadditional property. We prove furthermore that if the
dimensionof X is at least three then the anti-automorphism in the lattice
(ring) correspondsto the operationof taking the orthogonalcomplement(ad-
joint operator)in the resultingHilbertspace. The exact statementsand proofs
of thesetheoremsare containedforthe case of the latticein SectionI and forthe
case of the ringin Section II.
Since in [2] both Eidelheit's theoremand the lattice isomorphismtheorem
have been shown to hold fornot necessarilycompletenormedlinear spaces it
would be desirableto extendour characterizationto one among spaces of this
largerclass. We have not been able to do this. We have, however,examples
showingthat the theoremsof SectionsI and II are not trueif the completeness
restrictionis removed. These are presentedin Section11?.

I. The Lattice Theorem

Let X be a generalizedreal Hilbertspace. For each closed linear subspace
M of X let M' be the orthogonalcomplementof M. As is well known the
1 It should be pointed out that ProfessorKakutani left the United States beforethe
finaldraftofthispaperhad beenpreparedand accordinglyis notresponsibleforany degree
of unsatisfactoriness which it may possess. The results in this paper were obtained
while the junior author was in residence at the Institute for Advanced Study as a
SheldonTravelingFellow fromHarvard University.
2 In this paper linear means additive and homogeneous.
3 The numbersin bracketsreferto the bibliography.
50

This content downloaded from 192.231.202.205 on Mon, 17 Nov 2014 20:50:45 PM

All use subject to JSTOR Terms and Conditions
 CHARACTERIZATIONS
OF REAL HILBERT SPACE 51

operationAl Al' has the followingproperties:

(1) If Al1 and 312 are closedsubspaceswithM1 C M1I2 thenMi1 Q M1.
(2) For each closedsubspaceM, Al" = Al.
(3) For each closed subspace M, Al' n Al = 0.
W0econcludeat once that if X is any-real Banach space whichis isomorphicto
a generalizedreal Hilbertspace then the latticeof closed linearsubspaces of X
admits an involutaryanti-automorphism which takes each element into a
complement. Our firstcharacterizationis a consequenceof the fact that the
converseof this last statementis true. In case the dimensionof X is at least
threewe can provea strongerresultand we beginby doingthis.
THEOREM 1. Let X be a real Banach space whosedimensionis at least three
and let2 be its latticeof closedsubspaces. If thereexistsan operationA1- 31'
from2 into2 whichsatisfies(1), (2), and (3) abovethenthereexistsan isomorphism
between X and a generalizedreal Hilbertspace of such a characterthatthecorre-
spondenceMI+- Al' is carriedoverintothecorrespondence between orthogonal com-
plements in theHilbertspace. In otherwordsit is possibleto introduce a positive
definite4 symmetric bilinearinnerproduct(x, y) into X such thatthenew norm
x IIj in X definedby theequation Hx H=
HII (xx) is equivalentto thegiven
normH I x II andfurthermoresuchthatforeachMlin 2, Al' is thesetofall y in X for
which(x, y) = Oforall x in Al.5
PROOF. Let Ml be a one dimensional linear subspace of X. It is clear that
AMis closedand it followsfromthe factthat M -+ M' is a latticeanti-automor-
phismthat Al' is a maximalclosedlinearsubspace of X. Hence (Cf. [2],p. 245)
thereexistsa (real) continuouslinearfunctionalf definedon X and not identi-
callyzerosuchthatM' is theset ofall x in X forwhichf(x) = 0. The functional
f is uniquely determined tip to a multiplicative real constant and the set of all
such f's together with the identically zero functional form a one dimensional
linear subspace of the conjugate X of X. We shall denote this one dimensional
subspace by Mo.
It is easy to see that the operationAl -1 Al' thus definedsets up a one-to-one
linear independencepreservingcorrespondencebetween the one dimensional
subspacesof X on theone hand and thoseof X on the other. Since X is at,least
threedimensionalwe may apply Lemma A of [2] and concludethe existenceof
a one-to-onelinearmappingT of X onto X whichis such that foreach one di-
mensionalsubspace Al of X, T(M) = Al'.
We now use T to definean innerproductin X. For each two elementsx and
y in X let (x, y) = f(x) wvheref = T(y). Clearly(x, y) is bilinear. To showthat

4To avoid circumlocutions we say that (x, y) is positive definite if (x, x) is positive
definitein the ordinary sense.
5 In [4] G. Birkhoffhas proved a theorem quite similar to ours. The chief differences
between the two lie firstin the fact that Birkhoff'stheorem characterizes generalized Hil-
bert spaces themselves rather than spaces isomorphic to them and second in the fact that
while he assumes that a specific notion of orthogonality which he defineshas certain proper-
ties we merely assume the existence of a notion of orthogonality with certain properties.

This content downloaded from 192.231.202.205 on Mon, 17 Nov 2014 20:50:45 PM

All use subject to JSTOR Terms and Conditions
52 SHIZUO KAKUTANI AND GEORGE W. MACKEY

it is symmetricwe observefirstthat (x, y) = 0 if and onlyif MZ _ M', where

by Mx and My we mean the one dimensionalsubspacesof X generatedby x and
y respectively. But by properties(1) and (2) of the operationM -> M cC
M',
My if and onlyif My 5 MX. Thus the statements(x, y) = 0 and (y, x) = 0
are equivalent. Now let x and y be arbitraryelementsof X with x $ 0. By
property(3) of the operationM -1 M', (x, x) 0. Hence thereexistsa real
numberXsuch that X(x,x) + (x, y) = 0. Thus (x, Xx + y) = 0 so that (Xx +
y,x) = 0; thatis, X(x,x) + (y,x) = 0. We concludeat once that (x, y) = (y, x).
We have just pointedout that (x, x) # 0 if x $ 0. It followsthat (x, x) is
one signed. In factif (x, x) < 0 and (y, y) > 0 thenforsome pair of real num-
bers X,j, with'XI 0 we have X2(x, X) + 2XIA(x,y) + ,A2(y, y) = 0 = (XX + Ay,
Xx + juy)so that x = (-A I X)y. Thus (x, x) = (A21XI) (y, y) and we conclude
that (A2 1 X2) < 0 whichis impossible. Hence, since we may replace T by - T
ifnecessary,we may supposethat (x, x) > 0 forx 0 0 so that (x, y) in addition
to being bilinearand symmetricis positivedefinite.
We show now that the normI x I is equivalentto the originalnormi x 11;
that is, that thereexist positivereal constantsCi and C2 such that C, IIx I <_
IIIX j? 0_2 Ix forall x in X. If it is impossibleto findCi thenforeach n =
1, 2, thereis an x, in X withi1 n1II = land 11 Xn 11> n. Letf be an arbi-
trarymemberof X. Let y be such thatf = T(y). Then foreach n = 1, 2, - *
If(x,) I = I (Xn, y) I -<iII Xn III IIIY III <_ Y 111.Hence {f(xn)} is bounded.
It followsfromThdoreme6 on page 80 of [3] that { Xn IIj is bounded and this
is a contradiction. In order to prove the existenceof C2 we prove firstthat
T is a bounded transformation from X into X. Let {X} be an arbitrary
bounded sequenceof elementsof X and for each n = 1, 2, * letfn= T(xn).
For each x in X, {I f.(x) II = II (XrXn)II = II (Xn, x) I} = { f(xn) } where
f = T(x). Since {JX} is bounded IIf(x,) l} is bounded. Hence { Ifn(x) Il is
bounded. Thus using Th6orbme5 on page 80 of [3] and the fact that X is
complete we conclude that IIn III = I IIT(xn) II is bounded and hence
that T is a bounded transformation.Now we may write x jfl = V (x,x) =
x/fx) ? x/Ilfl~l lx l ? '\/IIT(x)I tlxlI ? <-\1IlT IIIxI 1 /2II
= T 1lix fl
forall x in X wheref = T(x). Thus C2 exists and may be taken equal to
VII T 11.
In orderto completethe proofof the theoremwe have only to show that for
each M in 2, M' is theset of all x in X such that (y,x) 0 forall y in Mll. But
foreach y in M of
the set all x in X such that (y, x) 0 is the set of all x in
/uII;that is M' . Since M'-- M' a
is latticeanti-automorphism the intersection
of all of theseis simplyM'.
Since any two finitedimensionalnormedlinearspaces havingthe same dimen-
sion are isomorphicand since there is a generalizedHilbert space of every
finitedimensionthe firstpart of Theorem 1 is true withoutany dimensionre-
striction. Thus we have:
THEOREM 2. Let X be a real Banach space and let2 be its latticeofclosedsub-
spaces. X is isomorphicto a generalizedreal Hilbertspace if and onlyif there

This content downloaded from 192.231.202.205 on Mon, 17 Nov 2014 20:50:45 PM

All use subject to JSTOR Terms and Conditions
 CHARACTERIZATIONS OF REAL HILPERTr SPACE 53

existsan operationlMi- - 31' from2 into t whichsatisficstheconditions(1), (2),

and (3) of Theorem1.
II. The Ring Theorem
We make the resultsof this sectiondepend upon those of the precedingby
findinga certainnaturallatticein the ring9kwhichis isomorphicto 2.
Let 9i be any ring. For each subset ef of Rtwe definethe right(left) anni-
hilatorof (d as the set Cf ((3) of all elementsr of fRsuch that ar (ra) = 0 forall
- iI thenar .D@37,
a in d. It is clear that forany subsetsCfand gi of ?kif(d
af D iB, I' D (1, and (Ia' D (3. It followseasily that CrT"(Cflrl) = (G(l) and
hence that (i' (Mfrl) = (f if and only if (d = C(O) forsome subset C of S. Ac-
cordingly we call a subset ad of9? such that d1(0(i') = (d a right(left) annihilator.
It is easily verifiedthat the set of all right(left)annihilatorsformsa complete
latticewhenpartiallyorderedby inclusion.
LEMMA 1. Let X be a normedlinearspace and let 9k be theringof continuous
linear transformations of X intoitself. A subset(f of 9k is a rightannihilatorif
and onlyif thereexistsa closedsubspaceM ofX suchthat(f is thesetof all members
of UZwhoseranges6are containedin M.
PROOF. Suppose that d is a rightannihilator. Then A is a memberof (a
if and only if BA = 0 foreach B in ai; that is if and only if the rangeof A is
containedin the intersectionM of the null spaces of the membersof ai. M is
obviouslya closedlinearsubspace ofX. Conversely,let M be a closedsubspace
of X and let Cfbe the set of all membersof Rkwhoserangesare contaniedin M.
Suppose that A is in (ar but that forsome xo in X, A (xo) is not in M. By the
lemmaon page 57 of [3],sinceM is closed,thereexistsa continuouslinearfunc-
tionalfo definedon X such that fo(A(xo)) = 1 and fo(x) = 0 for all x in M.
ClearlythememberC ofR such that C(x) = fo(x)A(xo)forall x in X is a member
of (1. Thus CA = 0 and hencefo(A(xo))A(xo) = A(xo) = 0. Since thisis a
contradictionA mustbe in (a. Hence a =Cf1 and is consequentlya rightan-
nihilator.
COROLLARY. The latticeof closedlinear subspacesof X is isomorphicto the
latticeof rightannihilatorsof Ai.
The analogue of Theorem 2 followsalmost at once fromthiscorollary and
Theorem 2 itself.
THEOREM 3. Let X be a real Banach space and letfRbe theringof continuous
linear transformations of X into itself. Then X is isomorphicto a generalized
Hilbertspace if and onlyif thereis an operationT -* T' fromORintoaJwhichhas
thefollowingproperties:
(1) If T, and T2 are membersof gRthen(T T2)' = T' T' and (T1 + T2)' =
T1 + T2.
(2) If T is a memberofR thenT" = T.
(3) If T is a member ofgRsuchthatT'T = 0 thenT = 0.
T is theset ofall elementsofthe formT(x) where
6 The rangeofa lineartransformation
x is in X.

This content downloaded from 192.231.202.205 on Mon, 17 Nov 2014 20:50:45 PM

All use subject to JSTOR Terms and Conditions
54 SHIZUO KAKUTANI AND GEORGE W. MACKEY

PROOF. The necessityof the conditionfollowsat once fromthe factthat the

operationof formingthe adjoint of an operator,which may be definedin any
generalizedHilbert space, has the properties(1), (2), and (3). Suppose, con-
versely,that an operationhavingthese propertiesexists. For each subset aTof
9k let a' denote the set of all T' for which T is in ai. It followseasily from
properties(1) and (2) that (a" = (a,that CaT = al" and that art = (Vf1. Thus
(a'Tr)'r ar
= = Cairand in particular if (a is a right annihilator (CT,/)'r = af. In
otherwords the operationai > Car takes the lattice of rightannihilatorsof ZR
into itselfand has property(2) of the firstsection. Also it clearlyhas property
(1). Let afbe any subset of 9Rand let T be contained in both adand aT'. Then
T'T = 0. Thus by property(3) of T -> T', T = 0. In otherwords d__a'r
also has property(3) of the firstsection. Using the corollaryto Lemma 1 we
conclude at once that the lattice of closed subspaces ofX admits an operation
havingthe properties(1), (2), and (3) of sectionI and henceby Theorem2 that
X is isomorphicto a generalizedHilbertspace.
The analogue of Theorem1 can be proved withoutrestricting the dimension
of X.
THEOREM 4. Let X be a real Banach space and let'? be theringofcontinuous
linear transformations of X intoitself. If thereexistsan operationT -* T' from
9kintoR whichhas properties (1), (2) and (3) of Theorem3 so thatX is isomorphic
to a generalizedreal Hilbertspace thenthisisomorphism may be set up in such a
mannerthatthecorrespondence T <-+ T' goes overintothecorrespondence between
an operatorand its adjoint. In otherwordsX may be providedwitha positive
definitesymmetric bilinearinnerproduct(x, y) such thatthenew normI x in
X definedby the equation IIIx I = V/(x, x) is equivalent to the given norm x
and suchthatforall x and y in X (T(x), y) = (x, T'(y)).
PROOF. By Theorem3 an innerproduct (x, y)i may be introducedinto X
satisfyingall conditionsexcept perhapsthe last. For each memberT of ai let
T* be the adjoint of T computedwithrespectto -(x,y)1. Then T -? T'* is an
automorphismof 9R. By a theoremof Eidelheit,[1], thereexists an automor-
phism U of X such that T'* = UTU' forall T in ?R. For all x and y in X let
(X, y) = (U(x), yX .7 We shall show that (x, y) has the properties desired.
First of all if (x, x) 0 then x 0. In fact if thereexistsan x $ 0 with0 =
(x, x) = (U(x), x)1 = (x, U*(x))i = (U*(x), x)1 let E be the projectionwhose
range is the one dimensionalsubspace of multiplesof x. Then 0 = (U*(E(y)),
E(z))I = (E(U*(E(y))), z)1forall y and z in X so thatEU*E = 0 Thus E'E =
(U*-lEU*)E = U*-1 (EU*E) = 0 which contradictscondition (3). It now
follows,using the argumentof the correspondingpart of Theorem 1, that we
may suppose that (x, y) is positivedefinite. (x, y) is obviouslybilinear. The
fact that it is symmetricis an immediateconsequenceof the fact that U is self
adjoint and this is a resultof the relationshipT" = T. In fact, since T' =
U*-1T*U*, T = T" = U*-1(UTU-')U* = (U-lU*)-lT(U-IU*) for all T in R.
7 In connection with our introduction of a new inner product in this way and our later

introduction of new norms in a similar fashion attention might be called to work of Fried-
richs [5] and Eberlein [61where this sort of thing is discussed is some detail.

This content downloaded from 192.231.202.205 on Mon, 17 Nov 2014 20:50:45 PM

All use subject to JSTOR Terms and Conditions
 CHARACTERIZATIONS OF REAL I1LBERT SPACE 55

Thus U-'U* is a constant,X; that is U* = XU. Since U** = U, X2 = 1 and(

since (U(x), x) is positiveforall x $ 0, X cannot be -1. Hence X = 1 and
U* = U. Since (x, x) = (U(x), x)i _ UI U(x)liii1jjx '1 1
IxII
wherejjjx 1lii is the normassociated withthe (x, y)i, in orderto establishthe
equivalence of x and jj1x liiiand hence of x and Ixj1 we have only
to provethat whenever{x,} is a sequenceof elementsof X such that lIII xn jjj I
is bounded then IIX, lII i) is bounded. Suppose that {IIll x IlII is bounded.
Then for each y in X I111 y + x,,IIl and hence I(U(x") + U(y), x. + y)iI =
{(U(x.), x.) + (U(y), y) + 2(x", U(y))) is bounded. Hence, since {(U(x^),
an) = 111 U(y)), I is bounded. But the range of U is the whole
n 1112r}X{ (x X,
of X. Hence foreach linearfunctional f on X whichis continuouswithrespect
to III x 11i, IIf(x,) II is bounded. Hence, since X is completewith respect to
jjj x lII, ,it followsfromTheoreme5 on page 80 of [3] that I II x" Ilii
II is bounded.
Finally for all x and y in X and each T in Rk,(T(x), y) = (U(T(x)), y), j-
(x, T*(U(y)))i = (UL(U(x)), T*(U(y)))i = (U(x), U-'(T*(U(y))))i = (U(x),
'(y)), = (x, T'(y)) and the proofof the theoremis complete.
We closethissectionby remarking thatby meansof the corollaryto Lemma 1
one of the questionsraised at the end of [2] may be answered;that is, the ring
theoremof that paper followsfromthe latticetheoremwithoutthe intervention
of Lemma A or Lemma B.
III. Some CounterExamples
We nuwshowthat the completenessassumptionsof the firsttwo sectionscan-
not be dispensedwith by producingexamples of incompletenormed linear
spaces whose ringsand latticesadmit operationssatisfying(1), (2), and (3) of
theorems3 and 1 respectively. It mightbe expected that one could at least
concludethatthecompletionofsuch a space is a generalizedHilbertspace. We
shall showfurthermore, however,that whilethismay be the case it need not be.
To beginwithlet us returnto the proofof Theorem 1 and see what we can
conclude about X if we do not,know that it is complete. A scrutinyof this
proofshowsus that we did not use the completenessof X untilwe had proved
the following:There existsa bilinear,symmetric, positivedefiniteinnerproduct
V
(x, y) in X such that forall x in X we have I x 11 +/(x, x) = I x IIIand such
thatf is a continuouslinearfunctionalon X withrespectto II x II if and only if
thereexistsy in X such thatf(x) = (x, y) forall x in X. Now let us complete
X withrespectto thenormIII x I1 It is clearthatthecompletion JCisa general-
ized Hilbertspace. It is furtherclear that thereis a unique way of extending
IIx II to X so that Ix II < III x III forall x in X3Cand that when UCis normed
with II x II a linearfunctionalf on X is continuousif and only if thereexists y
in X such thatf(x) = (x, y) forall x in X. In otherwordswe see that X may
be obtainedfroma generalizedHilbertspace by the followingprocess:Choose
a normin X whichnowhereexceeds the originalnorm. Considerthe subspace
of X consistingof all elementswhichdefinelinearfunctionalscontinuouswith
respectto the new norm. Topologize thissubspace withthe new norm.
Suppose now that X is any normedlinearspace whichhas been constructed

This content downloaded from 192.231.202.205 on Mon, 17 Nov 2014 20:50:45 PM

All use subject to JSTOR Terms and Conditions
56 SHIZUO KAKUTANI AND GEORGE W. MACKEY

froma generalizedHilbertspace in the mannerjust indicated. Then thereis a

bilinear,symmetric, positivedefiniteinnerproduct(X,y) in X such that a linear
functionalf on X is continuousif and only if thereexists y in X withf(x) =
(x, y) forall x in X. This is all that we need in orderto prove that foreach T
in the ring9i of all continuouslineartransformations of X into itselfthereis a
unique T* such that (T(x), y) = (x, T*(y)) forall x and y in X. The onlypart
of the proofthat differsfromthe corresponding part of the standard Hilbert
space proofis the part dealingwvith the continuityof T* and thisis carriedout
as follows. If {x, is any bounded sequenceof elementsof X thenforall y in
X, {(xn, T(y))) is bounded. But {(xn, T(y))} = {(T*(xn), y)} so that this
sequence is also bounded. Thus {f(T*(xn)) is bounded forevery continuous
linearfunctionalf definedon X. Accordingly,by Th6orbme5 on page 80 of
[2], {T*(xn)} is boundedso that T* is continuous. ObviouslyT -+ T* has the
properties(1), (2), and (3) listedin Theorem3. We see fromthe argumentused
in Theorem3 that the existenceof the ringoperationimpliesthat of the lattice
operationwhetherX is completeor not. Hence since finitedimensionaland in
particularone and two dimensionalnormedlinear spaces are always complete
we have proved the followingtheorem.
THEOREM 5. If a normedlinear space X has any of thefollowingproperties
then.it has all of them:
(1) The lattice? of all closedlinearsubspacesof X admitsan operationof the
sort describedin Theorem1.
(2) The ring9i of all continuouslineartransformations of X intoitselfadmits
an operationofthesortdescribed in Theorem3.
(3) Thereexistsa generalizedreal Hilbertspace 7X and a normIII x III defined
throughout C suchthatIKx I? < V/(x, x) forall x in X and suchthatthenormed
linearspace X1 which:(a) has as elements thosemembers y of Xfor which(x, y)
as a functionof x is continuouswith1respectto x and (b) is normedwiththe
normIII x 111,is isomorphicto X.
It followsfroma theoremof Fichtenholz,[7],that X, is containedproperlyin
X if and onlyif the normIIIx is not equivalent to the norm /(x, x). Fur-
thermoreX1 is obviouslydense in X with respectto both norms. Hence in
orderto obtainboth of the counterexampleswe are afterwe have onlyto finda
generalizedHilbert space in which there is a properlysmallernorm. But a
properlysmaller norm exists in any infinitedimensionalgeneralizedHilbert
space. We have only to choose a bounded linear transformation T which has
an unboundedinverseand let III x III = (1/ II T I|) II T (x) I I .8 It is clear that
such transformations always exist. For examplewe may choosean orthonormal
sequence{pI in Sand let T be theuniqueboundedlineartransformation whichis
theidentityin theorthogonalcomplementof {an}and is such that T(Otn)= 0n
forn = 1, 2, . The examplesconstructedby means of transformations in
thisfashionare howeverin a sense only partial counterexamplesin as much as
theyobviouslyhave generalizedHilbertspaces fortheircompletions. An ex-
8 See footnote
7.

This content downloaded from 192.231.202.205 on Mon, 17 Nov 2014 20:50:45 PM

All use subject to JSTOR Terms and Conditions
 CHARACTERIZATIONS OF REAL HILBERT SPACE 57

ample whichdoes not even have this propertymay be obtained by taking the
Hilbert space 12 and definingthe norm j x jfl as follows: IIIX1, X2, * = .

l.u.b. l x
"_1,2,.*.*
We note in passingthat all of theseexamplesare pseudo reflexivein the sense
of [2].
ConcludingRemarks
Since complexHilbertspaces are discussedmuch more frequentlythan real
Hilbertspaces it is naturalto ask whetheror not the theoremsof thispaper hold
forcomplexBanach spaces. We do not knowthe answerto thisquestion. Our
methodbreaks down because of the fact that the complexfieldhas too many
automorphisms. Let X1 and X2 be two linearspaces over a field X. If X -A V
is any automorphism of fTthen it is clear thatif T is a one-to-onetransformation
of all of Xi into all of X2 whichis quasi linearin the sense that T(Xx + /uy)=
X'T(x) + u'T(y) forall X,j in Ifandall x and y in X1 thenT sets up a one-to-one
linear independencepreservingcorrespondencebetween the one dimensional
subspaces of XI and X2 respectively. ConverselyLemma A of [2] is true for
linearspacesoverany fieldiftheword"quasi" is insertedbeforethewords"linear
transformation".Thus if we tryto carrythroughthe proofof Theorem1 fora
complexBanach space we obtain an innerproduct (x, y) whichis linearin the
firstvariableand quasi linearin the second. It is not difficult to show that the
automorphisms involvedin thequasi linearityis not theidentitybut we have not
been able to provethatit is theoperationoftakingthe complexconjugaterather
than one of the morepathologicalautomorphisms of the complexnumberfield.
In fact it can be shown that if X is finitedimensional it need not be. We are
inclined to believe that in the case of infinite
dimensional X the continuityof
the automorphism can be established. In thisconnectionit mightbe mentioned
that in workof B. H. Arnold[81the principaltheoremof [1] is discussedin the
complexcase and thatalthoughwe have notgoneintothematterwe have reason
to believe that some of the methodsused therecan be used to establishat least
our ringtheoremin the complexcase.
In view of thesubjectmatterof [2]it is naturalto attemptto use our methods
to characterizethe group of automorphismsof a generalizedHilbert space.
This can be done but as the conditionsare a littleunnaturaland theproofseems
to requiretherepetitionofa good partof thelengthyproofof thegrouptheorem
[2] we shall not go into the matter.

INSTITUTE FOR ADVANCED STUDY.

BIBLIOGRAPHY

1. EIDELHEIT,M., On isoinorphismsof rings of linear operations, Studia Mathematica, vol. 9

(1940), pp. 97-105.
2. MIACKEY, G. W. Isomorphisms of normedlinear spaces, Annals of Mathematics, vol. 43
(1942), pp. 244-260.
3. BANACH, S., Theorie des Operations Lineaires, (1932).

This content downloaded from 192.231.202.205 on Mon, 17 Nov 2014 20:50:45 PM

All use subject to JSTOR Terms and Conditions
58 SHIZUTO KAKUTANI AND GEORGE W. MACKEY

4. BIIIKHOFF, G., Orthogonalit?in linear metricspaces, Duke Mathematical Journal, vol. 1

(1935), pp. 169-172.
5. FRIEDRICHS, K., Spektraltheorie halbbeschrdnkterOperatoren und Anwendung aulf die
Spektralzerlegurnqloon Differentialoperatoren,'Mathematische Annalen, vol. 109
(1934), pp. 465-487.
6. EBERLEIN, W. F., Harvard Thesis, (1942).
lin-aires continues au sens generalish,Matema-
7. FICHTENHOLZ, G., Sitr les fuwnctionnelles
tiche Sbornik, N. S., vol. 4 (1938), pp. 193-213.
8. ARNOLD, B. H., Rings of Operators on Vector Spaces, Annals of Mathematics vol. 45
(1944).

This content downloaded from 192.231.202.205 on Mon, 17 Nov 2014 20:50:45 PM

All use subject to JSTOR Terms and Conditions