investigations in harmonic analysis h. j. reiter...
TRANSCRIPT
Investigations in Harmonic Analysis
H. J. Reiter
Transactions of the American Mathematical Society, Vol. 73, No. 3. (Nov., 1952), pp. 401-427.
Stable URL:
http://links.jstor.org/sici?sici=0002-9947%28195211%2973%3A3%3C401%3AIIHA%3E2.0.CO%3B2-X
Transactions of the American Mathematical Society is currently published by American Mathematical Society.
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available athttp://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtainedprior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content inthe JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained athttp://www.jstor.org/journals/ams.html.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printedpage of such transmission.
The JSTOR Archive is a trusted digital repository providing for long-term preservation and access to leading academicjournals and scholarly literature from around the world. The Archive is supported by libraries, scholarly societies, publishers,and foundations. It is an initiative of JSTOR, a not-for-profit organization with a mission to help the scholarly community takeadvantage of advances in technology. For more information regarding JSTOR, please contact [email protected].
http://www.jstor.orgMon Aug 6 17:27:23 2007
INVESTIGATIONS IN HARMONIC ANALYSIS(1) BY
H. J. REITER
This paper is concerned with the theory of ideals in the algebra L1 of integrable functions on a locally compact abelian group.
After some preliminaries an analytical proof is given of the known theo- rem that an analytic function of a Fourier transform represents again a Fourier transform (p. 406). Then, in part I , the continuous homomorphisms of closed ideals I of L' upon C, the field of complex numbers, are studied. Any such homomorphism is given by a Fourier transforin and, if lois its kernel, the quotient-algebra I/Io,normed in the usual way, is not only algebraically isomorphic, but also isometric with C (Theorem 1.2). Another result states that homomorphic groups have homomorphic L1-algebras and that a cor-responding property of isometry holds (Theorem 1.3).
In part 11, which may be read independently of part I , a theorem of S. Mandelbrojt and S. Agmon, which generalizes Wiener's theorem on the translates of a function in L1, is extended to groups (Theorem 2.2). Several generalizations of Wiener's classical theorem have been published in the past few years; references to the literature are given on p. 422. The rest of part I1 is devoted to some applications (pp. 422-425).
In conclusion it should be said that the work is carried out in abstract generality, with the methods, and in the spirit, of analysis, which is then ap- plied to algebra.
To Professors S. Mandelbrojt, B. L. van der Waerden, and A. Weil I owe my mathematical education. The inspiration which I have rcccivcd in their lectures, in letters, and above all in personal contact, is a t the base of this work; may I here express my gratitude.
Groups and Fourier transforms. The standard work of reference is A. Weil's book [14]; cf. also [4].
For the definition of a topological group see 114, p. 91. The additive group of real numbers (the "real axis") is denoted by R1, p-dii~lensional space by Rp.The group operation will always be denoted multiplicatively (it is hoped that this will not cause confusion in the case of R1). The identity (unit element) of the group is denoted by e.
We shall deal only with locally compact abelian groups G. By G we denote
Presented to the Society, December 28, 1951; received by the editors January 12, 1952. (1) This paper was submitted to the Faculty of the Rice Institute as a doctoral disserta-
tion. The author wishes to express his thanks for the generous award of a fellowship.
401
402 H. J. REITER [November
the dual group, by (x, x) the value for x E G of the character defined by xEG. For R l the symbol (x, x) may be regarded simply as an abbreviation of exp(2~ixx) and R1 is the same as R1.
The measure of a set S C G (SCG) is denoted by m(S) (m(S)). L1, L2, Lm denote, respectively, the space of complex-valued functions on G which are integrable, of integrable square, measurable, and essentially bounded. The last notation, a slight departure from Weil's, is used in [4]. Functions in L1, L2, Lm will be denoted, respectively, by f(x), . . . , F(x), . . . , +(x), . . . . We use a dot to indicate the conjugate of a complex number and we write f*(x), +*(x) for f'(x-l), +'(x-l).
The Faltung of two functions in L1 is
We shall also use the Faltung f*+(x), + E L m . Sometimes we shall write f(x) * + or even f(x) * +(x), according to convenience. If + l=f * +, then +T=f* * +*.
The Fourier transform of f EL1 is(2)
This is the definition adopted in [4, p. 871 rather than that used by Weil 114, p. 1121-the difference is irrelevant. For G =R1 it represents a slight, but convenient, deviation from the classical definition.
Banach spaces. For the definition and properties of Banach spaces B over the field of complex numbers, and the notation, we refer to Hille's book 181.
L1 and Lm are Banach spaces, the norms being, respectively,
The dual space of L1 is Lm; we shall write the general bounded linear func- tional on L1 in the form
The following "distance theorem" is a reformulation of a well known result 18, Theorem 2.9.41:
Let L be a closed linear subspace of B and dejine for x l E B
(2) Thus a function in Ll is indicated by a letter in boldface type and its Fourier transform by the same letter in italic type. This notation, considerably different from that used by French authors, was dictated by technical necessities of printing. Likewise for the use of the dot in- stead of the customary bar.
19.521 INVESTIGATIONS IN HARMONIC ANALYSIS 403
dist ( X I , L ] = inf 11x1 - xoli (XOE L).
If there is no bounded Linear functional x*(x) satisfying
for all xo E L,
then x l E L If there are bounded linear functionals satisfying condition (*), then
dist {xl , L ] = l/min IIx*II,
where x* ranges over all these functionals. This is an immediate consequence of Theorem 2.9.4 in [8] and the fact
that if x* satisfies (*) then x*(xl -xo) = 1, and hence IIx*II 2 l/IIxl -xol/. This theorem will be applied, in the space L1, to calculate the distance
in some cases (cf. 13, chap. 1111). The function +(x) is called orthogonal to f(x) : + l f , if Jf(x)+'(x)dx=O.
I t is called orthogonal to the linear subspace I C L 1 : +1I,if 4-l-f for all f E I . If 91L1 , then 9 vanishes almost everywhere.
The L1-algebra. L1 is not only a Banach space, but also a commutative Banach algebra [8], multiplication being defined by the Faltung.
A (closed) ideal in the algebra L1 is defined as a (closed) linear subspace of L1 which contains, with a function f , all products f JF g, for any g€L1. Closed ideals will be denoted by I.
A closed linear subspace is an ideal if and only if it is ((invariant under translation," i.e., contains, with a function f(x), all "translates" f(xy-I), for any (fixed) y E G . For the sufficiency see [6 , p. 1271. The necessity is proved here, for the sake of completeness.
Consider any f E I and any g E L 1 . Then, by hypothesis, f * g E I and hence if .ff(x)+'(x)dx = O for all f G I , then
Since g(y) is arbitrary, i t follows that
Since this holds for any +II,it follows that f(xy-I) €I,for any (fixed) y E G. The intersection of the zeros of the Fourier transforms of all functions in
404 H. J. REITER [November
I,which is a closed set in G, is called the co-spectrum of I by I,. Schn~artz [12]. We shall denote it by Zr.
By I(x, x)' we shall mean the "rotated" ideal obtained by replacing all f(x)€1 by f(x) (x, z) ' , x being any (fixed) element of G.
Let S be a compact, symmetrical neighborhood of the identity e E G and denote by S(x) the characteristic function of S :
This will be used as a standard notation throughout, and neighborhoods S,Sf,S,, etc., will always be supposed to have the above properties (without explicit mention).
Define now
Since S-l= S, S(x)is real and S(x-l) =S(x). If G =R1 and S= [-6/2, 6/21, S(x) = (sin mx) /TX; if G is compact and S= ( e 1 , then S(x) =1.
Consider now the Faltung
S * S(x) = m(xS A S).
By the theorem of Plancherel-Weil
In the applications we shall "normalize" the above Faltung by nieails of the factor l /m(S). The function thus obtained taltes values between 0 and 1 ; it is 1 for x = e and vanishes outside the neighborhood S2.
The function is non-negative and has the followi~~g p rope r t i~s (~ ) :
This follows immediately from the theorem of Plancherel-Weil, applied to S(x) and S(x).
Given E > O , for any (fixed) y E G
(ii) r n - l ( ~ ) / l S ( ~ - l x ) ~- s ( x ) ~ / ~< E ,
(3) We shall denotc the norm in L2 (and in the corresponding space L2) by l l ~ i l ~ = ( J I F(x)I 2 d ~ ] 1 / 2 .
19521 INVESTIGATIONS IN HARMONIC ANALYSIS 405
provided S is small ( S C U, a neighborhood of e depending on E and y ) . By Schwarz's inequality
m-l(~)lI 1~(y- 'x ) + S(x) ) { ~ ( y - l x )- S(x) ) I j 1 5 m-l(~)l j~(y- 'x)+ s(x)jj zjjS(y-lx) - s(x)lj2.
By the theorem of Plancherel-Weil this is equal to
m-112(~)lI~(x) l ) l j 21 (Y, x)' + 1 ) l12m-l12(s)lIs(x) 1(Y, x)* -
provided SCU, where the neighborhood U of e is such that I (y, x) -11 < ~ / 2for all xEU and the given y .
Take now two neighborhoods S and S' (cf. the convention concerning the notation on p. 404). Then the product SS' is a neighborhood of the same kind.
Consider the characteristic functions
S(x) and SS1(x)
(the use of two letters to denote a function is perhaps somewhat unusual, but convenient here). The normalized Faltung of these functions,
1%-I (S)S a SS1(x) = m-I (S)m(xS A SS'),
takes values between 0 and 1 ; it is 1 for xES1, vanishes outside S2S',and is the Fourier transform of
m-I (5') S(x) .SS'(x) . We shall need later the following estimate:
(iii) IIs(x)ssl(x) - B ( ~ ( s s ' ) -s ( x ) ~ ~ ~ I m(S) ]'12rn1/2(~).
Indeed,
I s (x ){s s l (x )- S(X)] [ I IB s (x)~~~I~ss ' (x)- s(x)lIz =
IlS(x)llzllSSf(x)- S(x)llz = m1I2(S) ( ~ ( s s ' ) - m(S) ) 112.
We shall also need a stronger form of property (ii): Given any (fixed) compact set C C G , then for any E > O
(ii') m - 1 ( ~ ) l l ~ ( y - 1 ~ ) 2S ( X ) ~ ~ I ~- < e
for all y € C, provided S is small ( S C U, a neighborhood of e which depends on E and C).
Referring to the proof of (ii), it is required to sholrr that there exists a neighborhood U of e such that 1 (y, x) -11 < ~ / 2 for all y E C and all xEU (this is obvious in the case G =R1; in the general case it is precisely the defini-
406 H. J. REITER [n'ovember
tion of the topology of the dual group given by Pontrjagin). The character (y, x) is a continuous function on G X G [14, p. 101]. Thus,
given E>O,, for each point y , E C there are neighborhoods V , of y , and U, of e such that I (y,x) - I I = I (y, x) - (y,, e ) I < ~ / 2 for all yE V,and all x E U,. Since C is compact, a finite number of the neighborhoods V, already cover C, and we may take for U any neighborhood of e contained in the intersection of the corresponding neighborhoods U,. Thus (ii') is proved(4).
Another estimate to be applied later is the following: Let C C G be a compact set and let E > O be given. Then
(iv) m-'(s)I I S(~-'X) SS ' (~-~X) ~ ( x ) < e(m(SS1)/m(S))'I2- SS'(X) 1 1 for all y E C , provided both S and S' are chosen small (namely, such that SS'C U, where U is some neighborhood of e depending on E and C).
The proof is analogous to that of properties (ii) and (ii'), U being such that I (y, % ) - I [ < ~ / 2 for all y E C , x E U .
We also have
( 4 m - l ( s ) l l ~ ( x ) ~ ~ ' ( x ) j j5 (m(SS')/m(S))1'2.
ANALYTIC TRANSFORMSFUNCTIONS OF FOURIER
THEOREM.Let f(x) be a funct ion in L1 and let CCG be a compact set. Denote by I' the set of points z , in the complex plane, of the form z =f(x), x E C . Let n o w A(z) be a (single-valued) analyt ic funct ion, regular at all points of r. T h e n there i s a funct ion in L1 whose Fourier transform i s equal to A(f(x)) for all x E C .
This theorem is proved in two steps: first it is proved locally, i.e., for a sufficiently small neighborhood of any (fixed) point xo€ C; then the exten- sion to the whole set C is carried out. The proof is modelled after that given by Carleman for the case G =R1 [3, chap. IV] (j).
Let xo be any (fixed) point of C and expand A(z) in a power series about zo=f(xO) .We may write, for x in some (small) neighborhood of xo.
m
A(f(x)) = A(f(x0)) + C cn(f(x) - f(x0))". n= 1
Consider now the Fourier transform (cf. p. 405)
S(X)= m-l(S)S(x) * SS1(x).
If both S and S' are small, the function
(4) The reader may now turn immediately to part I (p. 409) or part I1 (p. 417). (6) Professor A. IVeil has directed my attention to Segal's paper [13]. Theorem 3.8 in [13]
practically contains, as a special case, the theorem given here.
19521 INVESTIGATIOKS I N HARMOKIC ANALYSIS
exists for all x E G and coincides with A( j (x ) ) for x E x o S t The product
is the Fourier transform of
where
S ( X ) = m-l (S )S(x )SS1(x ) .
Write now
Hence
Given E >0 , let KC G be a compact set such that
By property (iv), pp. 406-408, the neighborhoods S , S' may also be so chosen that
for a l ly EK. Keeping S fixed, we may take S' so small that m ( S S 1 )<2 m ( S )(9). Under these conditions
5 ~{2l1~11fll l (cf. p. 406).+ 2(2)ll2j
Define now
q n ( ~ )= 4 * qn-~(x)r n > 1, q l ( x ) = q ( x ) .
Then
(6) This follows from the regularity of Haar measure (cf. [ 7 ] ) and the fact tha t for any open set 0 containing S there is an S' such tha t SS'CO.
H. J. REITER [November
llqnlll s llqll~l n 2 I.
By a suitable choice of the neighborhood., involved, me can thus obtain a function q(x) such that llqill <R, where R is the radius of convergence of the power series x,",,c,zn. Then the series
converges in L1-norm and its Fourier transform coincides with A(ffx)) for all xExoS'.
The theorem has now been proved locally; to extend it to the set C as a whole, we proceed as follows:
To every point x , E C there corresponds a (small) neighborhood x, TT,,V, being a neighborhood of e , such that there exists a function f , E L 1 whose Fourier transform f,(x) coincides with A Cf(x)) for all x ~ x , Vf. The season for considering first x,V, and then x,V: will appear shortly. Since C is com- pact, a finite number of those neighborhoods, say xl VI, xz Vz, . . . , x . ~VX, already cover C:
Take a neighborhood S so small that S C V, (1 5 r 5N) ; then
Define now a partition of the set CS as follows
n-1
cs1= cs nxlv:, CS, = cs n X,V; nc u cs, ( 2 5 n 5 x).
Thus 2
CS, c x , ~ , . , css c xrv,3 (1 5 r 5 N ) .
Consider the functions f r (x)EL1 (1 S r s N ) whose Fourier transforms f,(x) coincide "locally" with A(f(x)), i.e., for xEx,V:, respectively.
Let CS,(x) be the characteristic function of the set CS, (1 I r IN). Then the Faltung
is the Fourier transform of
(the notation being the usual), and vanishes outside CSrS. Moreover the sum
INVESTIGATIOKS IN HARMONIC AKALYSIS
N
m-l(S)S(x) * CSr(x) r=l
is just
m-l(S)S(x) * CS(x) = m-l(S)m(xS nCS)
and hence is 1 for x E C. Now the Fourier transform
coincides with ACf(x)) for all xEC. Indeed, if x E C , then in the product
either the Faltung in the bracket vanishes or else, if it does not, fr(x) coin- cides with ACf(x)), since CS,SCX,V~.
Thus, for all x E C , Z(x) =A(f(x)) . 1 and the theorem is proved. The following particular case should be especially noted: I f f (x )EL1 i s such that f(x) # O on the compact set CCG, then there i s a
.function g(x)EL1 such that
g(x) = l/f(x) for all x E C.
This theorem is a t the base of tile results that follow. For that reason it seems worthwhile to have a purely analytical proof; other proofs have been given by Godement [6, ThCor2me A] and Segal 113, Theorem 3.81.
COROLLARY.Let f(x), g(x), be in L1. I f f(x) vanishes outside some compact set and the zeros of g(x) are all interior to the set of zeros of f(x), then there i s a function h(x) EL1 such that
LEMMA1.1.1. I f f EL1 i s such that f ( e ) # 0 , then every solution +oELmof the integral equation
has the following property:
for any constant y, unless +O i s zero almost everywhere.
Consider the function ~ - ' ( S ) S ( X ) ~ . If S is small, we have, by the corol- lary above since f(x) # O in some neighborhood of e,
410 H. J. REITER [November
so that
rn - l (S )S (~ )~ = 0.$ +:(X)
Now
rn- l (S)S(~)~* 1 y + +o(x) I 2 S ess. sup I y + +o(x) 1 2 . The left-hand side may be written(')
rn- l (S)S(~)~* i: 1 y l 2 -/- 2~ [ys+o(x) 1 + / +o(x) 1'1 = 1 Y l 2 + ~ R [ ~ . ~ - ~ ( s ) s ( x ) ~* + o ] + rn - l (S )S (~ )~* I + o 12.
But the middle term vanishes and hence, unless + O is zero almost everywhere,
I y l 2 < ess. sup I y 4 +o(x) 1 2 , which was to be proved.
I a ~ l n ~ ~ ~ 0, then the integral equation 1.1.2. If f EL1 i s such that f ( e )=
f *+*(x) = 1
has no solution +ELw.
Suppose there is a + E L m such that
S ~ ( Y ) + . ( X - ~ Y ) ~ Y for all x E G.= 1
Multiply both sides by W - ~ ( S ) S ( X ) ~ and integrate with respect to x:
Since Jf(y)dy =0, the left-hand side is equal to
Given E > O , there is a compact set CCG such that
(7) R [ z ]=real part of z.
19521 INVESTIGATIONS IN HARRlONIC ANALYSIS
Furthermore, for any y € C,
provided S is small, by property (ii'), p. 405. Thus the last repeated ~211+11 ,+llfll 1 ~ 1 1 + 1 1integral is in absolute value smaller than ,. Since this
may be made arbitrarily small, no solution + E L * can exist. Given a function f E L 1 , denote by (f) the closed ideal generated by f , i.e.,
the smallest closed ideal in L1 containing f. I t may be defined explicitly as the closure of the set of all linear combinations of "translates" of f(x):
where al, az, . . . , a , ~are arbitrary complex numbers, yl, yz, . a . , YN are any elements of GI and N is any positive integer (cf. p. 403).
Define now the ideal (f), as the closure of the set of all linear combinations
where the coefficients b, satisfy the condition
The reason for the notation mill appear later. REMARK.I t is important to observe that (f), may also be defined as the
closure of the set of all linear combinations
with arbitrary coefficients a,.
THEOREM1.1. For any f (x)EL1
where N ranges over all positive integers, bl, bz, . . . , bH range over all complex numbers satisfying EL1bn =0, and yl, yz, . . . , y~ over the elements of G.
(If the equality sign is replaced by '(2 ", the statement becomes trivial.) Making use of the definition of the ideal (f),, the theorem may be stated
412 H. J. REITER [November
concisely thus:
dist I f , ( f ) , ] = ( f (c ) 1 . The proof is an application of the distance theorem (p. 402): me have
dist { f , ( f ) ,) = l/min / + I , , where + ranges over all functions in L* satisfying
(here the remark on p. 411 is used). If no such + exists, then fE(f ) , . Thus we have to investigate the solutions of the integral equation
Suppose first that y =lf(x)dx#0. Then every solution may be written in the form
where +0 satisfies f * +z(x) =O. By Lemma 1.1.1 the function +(x) = l /y ' is the solution with the smallest norm and hence
disi, { f , (f) ,) = j f (e) I If lf(x)dx = O , then, by Lemma 1.1.2, there is no + E L * satisfying
f * +*(x) =1 and hence fE(f) , . Thus the proof of the theorem is complete. Theorem 1.1 may be generalized. Let I be the closed ideal generited by
the functions f , g, . . . , i.e., the smallest closed ideal containing (f), (g), . . . ; we express this by writing
I = (f , g,.. . ) .
We define noxv the ideal I, as the smallest closed ideal containing (f),, (g),, . . . and write
I e = (f, g , . . . )e
Thus I, is the closure of the set of all finite sums of certain linear combina-tions (see p. 411) of the translates of f, g, . . . , respectively.
Then for any k € I
dist { k , I , ) = 1 k ( e ) 1 . Indeed, since k E I , there are, for any E > O , among the generating functions
f, g, . , a finite number, say f ~ ,£2, . . - , f~ such that the function
INVESTIGATIONS I N HARMONIC ANALYSIS
(where the choice of the coefficients a,, and the elements y,,E G depends on E )
satisfies the condition
llk(x) - kc(x) l l~< E. Consider now the function k, and the ideal (k,),: by Theorem 1.1
dist (k., (k.).] = lSk . (x)dxI . From the definition of I, it is clear that
(kt) 0 C I e
and since for all h E I0
it follows that
dist (k,, I.] = 1 Sk.(x)dx 1 This implies that
dist { k , I,) 5 Ilk - k.111 + / Sk.(x)dx 1
Now the inequality dist { k, I,] 2 I Jk(x)dxl is again trivial and hence, since E > O is arbitrary, the result is proved.
We proceed now to the final generalization of Theorem 1.1. If I= (f), we define the ideal I,= (f), as the closure of the set of all linear
combinations
where x is a fixed element of G, N is any positive integer, y,E G (1sn 5N), and the coefficients satisfy the condition
414 H. J. REITER [November
If I = ( f , 8, . . . ), we define I , = ( f , g , . . - ), as the smallest closed ideal containing ( f ) , , (g),, . . We have now (in the notation of p. 404):
dist { k ( x ) ,I , ] = dist { k ( x ) ( x , x ) ' , I , ( x , a ) ' ) .
But the ideal I , ( x , x)' bears to the ideal I ( x , x)' the relation
I,( . , x ) . = ( ~ ( x ) ( x ,x ) ' , g ( x ) ( x 7 x ) . , . . . ) e = I ( x , x)e,
and hence we may apply the first generalization of Theorem 1.1 to the func- tion k ( x ) ( x , x ) ' E I ( x , x)' and the ideal I(x,x ) ; , so that we finally obtain
dist { k , I , ) = I k ( x ) 1 . The results obtained may now be summarized as follows:
THEOREM1.2. Let I be a closed ideal in L1.Dejine, for a j i s e d s E G , the idea2 I , C I a s above. T h e n for a n y funct ion fE I
dist { f , I , ) = I f ( x ) I . Hence I , i s the kernel of the homomorphism
f -+ J(.) of I into C , the Jield of complex numbers. T h u s I , i s a nzasinzal ideal of I or coincides wi th I according a s x E e Z 1 or x E Z I . JIoreover, i f the guotient-algebra I / I , i s considered a s a Banach algebra(8) , then the i somorphism
preserves the norm. Conversely: ginen a cont inuous homomorphism
f -t P ( f )
of I u p o n C , then
where x , E G i s a uniquely determined element of CZr.
T o establish the converse, let s be any (fixed) element of G. I t is proved in [6, p. 1221 tha t
Pf(.-"x)) = (st ~ , ) . P ( f ) i
where x , is a uniquely determined element of G. Consider now the ideal I,,. I t is the kernel of the (continuous) homo-
morphism
(8) See, e.g. [8, Theorem 22.1 1.41.
INVESTIGATIONS IN HARMONIC ANALYSIS
f -+J f ( x ) ( x , %,)*ax.
Let I, be the kernel of the given homomorphism f -+p( f ) . If we can prove that IzpCI, ,it will follo~v that I zp=I ,and hence that p ( f ) =J f (x ) (x , x,)'dx.
Suppose there is a function g E l , , which is not in I,. Then p ( g ) # O and, for all s EG ,
while
Now p ( f ) , considered as a bounded linear functional on I, may be extended to the whole space L1,by the complex analogue of the Hahn-Banach theorem [a, Theorems 2.9.2 and 2.9.51. Hence
We have therefore for the function g ( x ) :
S s ( ~ - ~ x ) + : ( x ) ~ x (s , XJ .PCS) for all s E G ,=
S g ( x ) ( x , x , ) 'dx = 0.
If we write this in a slightly different form and apply Lemma 1 . 1 . 2 , we see that it is impossible. This completes the proof.
THEOREM the group G' i s homomorphic image of G , then the 1.3. I f a L1-algebra on G' i s a homomorphic image of the L1-algebra on G. The kernel I. of this homomorphism consists of all functions in L1(G) whose Fourier transforms vanish on the dual group G ' C G , and the isomorphism
preserves the norm (the measure on G' being properly normalized).
For any f ( x ) E L 1 ( G ) ,define
where the closed subgroup g C G is such that G 1 = G / g [14, p. 1 1 1 , and x' denotes the coset x g . Then fl(x')EL1(G1)and (with proper normalization of
416 H. J. REITER [November
the measure)
This result (immediate for G=R1) is proved in [I]. We have also
Furthermore, since GI= G/g, the dual group G' is contained in G and (y, x') =1 whenever yEg, x'EG' (cf. [14, pp. 108-1091). Thus for x =xlEG' ,
Since a function in L1(G') is entirely determined by its Fourier transform, it follows that T(f(x)) is determined by the values of f(x) for x =x'EG1. Hence we have
i.e., T is a homomorphism of L1(G) into L1(G'). The kernel of this homomorphism is the closed ideal I. consisting of all
functions in L1(G) whose Fourier transforms vanish on G'CG. We have
inf ((f(x) - fo(x) / (~I(\f '(xl)(\l (fo E 10).
To establish the opposite inequality, we use the distance theorem (cf. p. 402) :
dist ( f , 10)= l/min l l + l l m where
LEMMA.Any + l I o has the property that
almost everywhere (and conversely).
For any k(x) EL1(G), k(sx) -k(x) Elo,i.e.,
19.521 INVESTIGATIONS IN HARMONIC ANALYSIS
Since kEL1(G) and s E g are arbitrary, the assertion follows (the converse follows from the fact that f(x) $ +* =fl(x') * +'*-cf. below).
Thus the function $ 1 1 0 defines a function +'(xl) on G' and we have
Hence
ll+ ' l lml l f ' l l1 2 1
and thus
ll+(x)llm 2 l/llf'll 1.
This gives the opposite inequality. Thus the Banach quotient-algebra L1(G)/Io is not only algebraically iso-
morphic, but also isometric with the image of L1(G) in L1(G') under the homomorphism T. I t remains to show that this image is L1(G') itself.
Since L1(G)/Io is complete, it suffices to show that the functions T(f) are dense in L1(G'). But this follows from a proof given by A. Weil [14, pp. 42-43]. Thus the theorem is established.
11. A GENERALIZATION OF IVIENER'STHEOREM, WITH APPLICATIONS
By Wiener's theorem is meant Theorem IV in [IS]. I t was extended by Godement [6, p. 12.51 to locally compact abelian groups and reads:
Let I be a closed ideal in L1. In order that I= L1, it is necessary and suffi- cient that for every xoEG there exist an f(x) €I such that f(x0) t ' O .
In the course of the proof of Wiener's theorem a result is established which may be stated as follows:
An ideal I contains all functions fc(x) whose Fourier transforms vanish out- side some compact set C C 021.
(Actually this is stated here in a slightly more general form which is precisely what we shall need. The proof is that of Godement.)
For any x l E C there is a function £ , € I such that I fl(x) / 2 1 in some small neighborhood x,Sl of x,. Since C is compact, it may be covered by a finite number of those neighborhoods, x,S, (1 s r 5N). Then the Fourier transform of the function EL,f,(x) * f:(x) €1is strictly positive on C. If we now apply the corollary, p. 409, to fc(x) and this function, the result follows.
We shall suppose from now on that G has a denumerable fundamental system of neighborhoods of the identity (but cf. Appendix). We may then
418 H. J. REITER [November
choose in G a fundamental sequence (S,), n 2 1, of compact, symmetrical neighborhoods of e such tha t Sn+ICS,. Then every subsequence (S,,), r 2 1, will still be a fundamental system.
LEMMA2.1.1. For any function + E L m there exists (at least) one constant M such that, for a certain subsequence (n,),
~- '(S,,)S,,(X)~* + -t M (7 +w ) ,
for all x E G.
Since the numbers m-l(S,)JS,(y)2+(y)dy are all bounded, in absolute value, by l l + l l m , there exists a subsequence (n,) such that m-1(Sn,)J~,,(y)2 .+(y)dy tends to a limit, M, say, as r-+w.
We want to show that for any (fixed) x E G , ~- ' (S, , )S, , (X)~ * ++M (r+m). I t will be sufficient to prove that
for any (fixed) x E G . Now the integral may be written
and this, in absolute value, will be smaller than e/ j+/ lm,by property (ii), p. 404, if n, is sufficiently large. This completes the proof.
LEMMA2.1.2. Suppose that $11and that the identity e E G is not a limiting point of 21. Then, if the neighborhood S is small ( ~ ~ n z ~ n ~ ( e } =@),
m - l ( s ) S ( ~ ) ~* + = const.,
the constant being independent of S (subject to the above condition).
REMARK.This is of course trivial if G is compact, since then we may take S= ( e l . See also the Appendix a t the end of this paper.
Take a neighborhood S A T of the fundamental sequence of neighborhoods, so small that s ~ s H A z ~ ~ ~ ( ~ )Then the Fourier transform =@.
S * SSN(x)- S * SS,(x), n > N ,
vanishes outside the set S2S~nCS,(cf. p. 405; it may be helpful to plot a figure for the case G =R1). Applying the result on p. 417, we have
I t is our first aim to prove that
S(x) .SS,y(x) - S(x) E I.
19521 INVESTIGATIONS IN HARMONIC ANALYSIS
Indeed, we have
S(X).SSM(X)- S ( X ) ~= { S(X) .SSN(X) - S(X).ssn(x) ] + { ~ ( x ) ~ ( x ) ~ } ..ss*(~) -
The first part is in I (see above); the second is, in L1-norm, less than m(SS,) -m(S) } 1/2m1/2(S)(cf. property (iii), p. 405). But this can be made
arbitrarily small, by taking n sufficiently large(g), and thus the partial result is proved.
On the other hand, if we take S n C S , then
(cf. p. 405) the reader may again find it helpful to plot a graph of the Fourier transforms for the case G =R1).
But
(set S=S, in the partial result just proved) Hence finally
for all S,CS, i.e.,
+ being the function of the hypothesis. Hence we have
Let now (n,) be a subsequence such that
(Lemma 2.1.1). Since the left-hand side of the above equation is independent of n, i t follows that
rn - l (S )S (~ )~*+ = M = const.,
which proves the lemma. I t also entails that M is unique. In this case we write ~11+(x) ) instead of M and call i t the mean value of +(x).
This terminology is justified by the fact that , i f a ( x ) is not only in La,but also almost periodic in the sense of von Neumann, then
( 9 ) This follows from the regularity of the Haar measure (cf. [ 7 ] ) and the fact that, for any open set 0 containing S, the product SS, will be contained in 0 for large n.
420 H. J. REITEK [Kovember
exists and coincides with the usual mean value. Let A be tha t mean value. Then, given E > O , there exist positive numbers cl, 62, . . . , c , ~ ,whose sum is 1, and elements yl, yz, . . . , y x E G such that for all x E G
We may write the difference
in the form
But each term of the first sum tends to 0, as n+m, by property (ii), p. 404, and the second term is in absolute value less than E, for every n. Since E > 0 is arbitrary, the result is proved.
From Lemma 2.1.2 there follows immediately a general result:
THE ORE^^ 2.1. Let I be a closed ideal in L1 and suppose that x lEG is not a limiting point of 21.Suppose that + I I . Then the mean value of +(x)(x, xl)' exists and, if S is suficiently small ( x l S 2 n Z I n e { xl ] =@) ,
Indeed, we may reduce this case to Lemma 2.1.2 by writing +(x)(x, xl). I I ( x , xl)', the right-hand side denoting the "rotated" ideal (cf. p. 404).
Now we are ready to prove a generalization of Wiener's theorem (for groups G whose dual groups have a denumerable fundamental system of neighborhoods of the identity; cf. hon~ever the Appendix) :
THEOREM2.2. Let I be a closed ideal in L1, with co-spectrum Zr (p. 404). If the function k(x) EL1 is such that its Fourier transform k(x) satisjies the following conditions :
(i) Zk>Zr (21,denotes the set of zeros of k(x)) ; (ii) J(Zk) nJ(ZI) is denumerable (J(Z) denotes the frontier of Z) ,
then k(x) EI.
This theorem mas first proved, for the case G =R1, by S. n'fandelbrojt and S. Agmon 111, ThC.ori.rne 111, malting use of Carleman's analytic func- tions [3, p. 7.51. The reader will see for himself the connection between their proof and that given here.
19521 INVESTIGATIONS IN HARMONIC ANALYSIS 421
To prove the theorem, it suffices to show (cf, the distance theorem p. 40.2) that whenever +IIthen +Lk .
Consider a +IIand define the function +l(x) by +l(x) = k*(x) * +(x). If we can prove that +ILL1, then it will follow that k * +*(x) =0 , and + I k .
Let Il be the set of all f E L 1 such that f * + T = O ; this is a closed ideal. We shall prove:
(2) ZI, has no isolated points.
From (1) it follo\vs, under the hypothesis of the theorem, that 21,is de- numerable, and from (2) that it is perfect. This implies that ZI,=@, and hence (Wiener's theorem) Il=L1. Thus the theorem will be proved if we can establish (1) and (2).
(1) Since g * +*=O for any g € I , we have
Thus +lII,i.e., ICIl, and hence ZI,CZI. Now we shall show that no interior point of ZI, can belong to 21,.
Indeed, if xo is an interior point of Zk, let S be so small that xoS2CZk, and consider the function (x, ~ o ) S ( x ) ~ . This function is in Illsince [(x, x o ) ~ ( x ) ~ ] * +;= [(x, X O ) S ( X ) ~ ] (observe that * k * +*=O the Fourier transform of [(x, x ~ ) s ( x ) ~ ] * k vanishes identically). Thus xo is not in ZI, (cf. [ l l , Lemme I]).
Therefore 21,is in ZI and contains no point interior to Zk. Since Z I C Z ~ , the only points that can possibly belong to Zr, are those points of J(ZI) which are in J(Zk). Thus (1) is proved.
(2) Let us suppose that x lE r (Zr )n r (Zk) is not a limiting point of ZI,. Consider first the case xl=e. By Theorem 2.1
e.i., rn - l (S)S(~)~ * k * +* = MS{+l ) . Take now S=S,. By Lemma 2.1.1 there exists a subsequence (n,) such
that rn-l(S,,)S,,(~)~ * +* tends (boundedly) to a constant c ( r + w ) . But then the left-hand side of the last equation tends to cJk(x)dx=c.k(e) =O. Thus M'(+l)=0, i.e., S ( X ) ~ E I I , and hence e is not in 21,(see also Appendix).
The general case may be reduced to the preceding one by considering instead of +l(x) the function
The ideal corresponding to that function is just the "rotated" ideal (cf.
422 H. J. REITER [November
p. 404) Il(x, xl)', with co-spectrum x;'Zr,, and the co-spectrum of k(x)(x, xl)' is x;'Zk. Thus, by the special case just proved, e is not in x;'Zr,, i.e., xl is not in Zr,, and the proof of (2) is complete.
REMARKS.Condition (i) of the theorem is obviously necessary as well; moreover, for compact groups it is also sufficient, condition (ii) being then superfluous, since the dual group of a compact group is discrete. Thus for compact groups G there is a one-to-one correspondence between the closed ideals of L1 and the subsets of G which inverts the inclusion (cf. [lo, p. 3921). On the other hand L. Schwartz has proved [12] that condition (i) alone is not suffi- cient in general, if G is not compact. This shows the role of condition (ii).
The reader is referred to Mackey's survey [lo, pp. 392-3951 for a discus- sion of previous generalizations of Wiener's theorem. Besides the papers by Mandelbrojt-Agmon [ l l ] and Godement [ 6 ] ,we mention here in particular those of Ditkin [5], Segal [13], and Kaplansky [9](1°).
We shall now apply the preceding results to study the functions + E L m orthogonal to a given ideal I, when the co-spectrum ZI has a particularly simple structure.
If ZI is jinite, then any +IIis a linear combination, with constant coefi- cients, of the characters of G defined by the elements of Zr.
Let ZI consist of xl, xz, . . , XN. If k(x) is any function in L1, then (for S suitably small)
N
k(x) - k(x) * C (x, ~,)rn-l(S)S(x)~ n=l
will be in I (Theorem 2.2), i.e.,
for any +II.By Theorem 2.1 we may then write
k(x) * +*(x) -x ai(x, x,) = 0{ N
1n=l
where an =M(+(x) (x, x,)' ). Since k(x) is an arbitrary function in L1, i t follows tha t
(lo) It should be pointed out that some of these papers contain theorems proved for arbi- trary locally compact abelian groups. Since the present paper was submitted for publication, Dr. Henry Helson has kindly sent me a reprint of his paper Spectral synthesis of bounded func- tions, Arkiv for Matematik vol. 1 (1951) pp. 497-602. This paper contains also a generalization of Wiener's theorem (for arbitrary locally compact abelian groups). The theorem is stated there in a weaker form, but the proof would actually yield the result given here.
19521 INVESTIGATIONS IN HARMONIC ANALYSIS 423
almost everywhere. Cf. [6, Hypothkse A, p. 1361 and [9]. If 21is discrete, then any uniformly continuous +IIis almost periodic, witlz
spectrum in Zr (and conaersely). Let kc(x) be any function in L1 whose Fourier transform vanishes out-
side some compact set C. Then, in the same way as before,
kc(x) * +* = C kc(x,)a:(x, x,) (x, E ZI A c).
Now the functions of the type kc are dense in L1 [13, Theorem 2.61. Thus for any k E L 1 the function k * + is almost periodic, with spectrum in 21. In particular, set k(x) =m-l(V)V(x), where V(x) is the characteristic function of a neighborhood V of e. Given E > O , there is a V such thatI m-l(v) ~ ( x ) < E , for all x E G , by the assumed uniform con-*+-+(x)] tinuity of +(x). Hence +(x) itself is almost periodic, with spectrum in ZI (the converse follows from the approximation theorem for almost periodic functions).
For the case G =R1 this was proved by A. Beurling [2]. If now G =R*, let V, be a sequence of (#-dimensional) cubes, with center
a t the origin and sides of length l /n . Then m-l(V,) V,(x) *+ -+ +(x) almost everywhere (72-403). Since m-l(V,) V,(x) * + is almost periodic, there are trigonometric polynomials P,(x), with frequencies in 21,such that /m-l(V,)V,(x)*+-P,(x)/ < l / n , n 2 1 . Thus P,(x) -+ +(x) almost every- where (n-+co) and 1 P,(x)I <l/+i/m+l G=Rp, <f 21isfor all n. Hence, for discrete, then the set of all +IIconsists of all those fz~nctions in L* which are limits (almost everywhere) of sequences of uniformly bounded trigonometric polynomials with frequencies in Zr.
In the case G =R1 this was proved by Carleman [3, p. 1151, with an addi- tional restriction on 21.
If 21is discrete, then to any +IIthere corresponds (by Theorem 2.1) the "Fourier series"
where a, =M{+(x)(x, x,)' ) and the summation extends over all x ,EZr . The ('sum" is purely formal-no ordering of the terms, or convergence, is implied.
If all the Fourier coefficients are zero, then +(x) vanishes almost every- where ("uniqueness theorem"), since then + I L 1 .
LLBessel'sinequality" holds in the sense that
S , (X)~+(x)I a , 1 _I lim inf mn-l(S,) S 1 1 ~ d x , n-t rn
the set of all a,#O being denumerable. The proof is the same as in the clas- sical case.
The question now arises under what conditions it is true that
H. J. REITER
("Parseval's equation"). Consider the function
which is orthogonal to I. We have, for large n,
But, in view of Theorem 2.1, this is just
by an argument already familiar. The last sum (over all x,EZr) has actually only'a finite number of terms.
Set now S=Snand write correspondingly an(x) , A,(L, K) . We are going to indicate some restriction on 21which will guarantee the absolute convergence of
Suppose 21is such that, for some S,x S n 2 1 =x , for all X E Z I (21 inay then be called "uniformly discrete"). Then we have, for sufficiently large n ,
by the uniqueness theorem. Suppose now that 21is the union of a finite number N of sets, each uni-
formly discrete. Then for sufficiently large n , say n z n o , the following holds:
19521 INVESTIGATIONS IN HARMONIC ANALYSIS 425
if one of the indices L, K is kept fixed, then A,(L, K) > O for a t most N values of the other one.
Define now the function T(L, K) by
Then both x,T(L, K) and ELT(L, K) are a t most equal to N. Hence
Hence the series
xr,rAn(~i K)aLa,*
will converge absolutely, and (some order of the terms having been fixed) uniformly with respect to n. By the uniqueness theorem
Thus
M [ I + 1 2 } = lirn W-'(S,) S 1 +(XI 12dxS ~ ( X ) ~ n+ =
exists and
M { I + 121 = x I a1 1 2 .
Thus, i f Zr m a y be decomposed into a finite number of sets each of which i s uniforlnly discrete (p. 424), then "Parseval's equation" holds for all +II.I t follows that the Fourier series of + converges "inthe mean" to +.
For the case G= R1,this means that all + I I are almost periodic B2 (Besicovitch), since in this case M [ I + /2 ) , as defined here, coincides with the mean used by Besicovitch, in view of a theorem of Wiener [16, Theorem 211.
APPENDIX(")
I t is possible to modify the proof of Lemma 2.1.2, avoiding Lemma 2.1.1 entirely and dropping the assumption that the dual group possess a de-numerable fundamental system of neighborhoods of the identity (first axiom of countability). The modification is as follows:
In Lemma 2.1.2 it is proved that the function
(11) Added September 15, 1952.
426 H. J. REITER [November
for S sufficiently small, is independent of S, without any recourse to the first axiom of countability (it is convenient, of course, to change the notation, writing, say, Sf,S" instead of S,, SN,respectively, where SfCS").
To show that the function above is also independent of x, take x =e and set
Now the difference
is equal to
for any S'CS, and hence we have
5 m-'(~~)ljs '(x- 'Y)~- ~~(Y)~ljljj+lj ,
=< ~ l / + / l m ,
by property (ii), p. 404, if Sfis small enough. Thus Lemma 2.1.2 is established by a method which is more general as well as simpler.
The only other place where the first axiom of countability is used is on p. 421 where it is proved that in the equation
the constant M' (41) is zero. Keeping S fixed and setting
we have
k *$" = ~ ' { + l j
and also
Now in Lemma 1.1.2 (part I) i t is proved, without any countability restric- tion, that the last two equations imply
I t results that the proof of the extension of the theorem of Mandelbrojt- Agmon (Theorem 2.2) is valid for locally compact abelian groups in general.
19521 INVESTIGATIONS IN HARMONIC ANALYSIS 42 7
BIBLIOGRAPHY 1. W. Ambrose, Direct sunz theorenz for Haar measures, Trans. Amer. Math. Soc. vol. 61
(1947) pp. 122-127. 2. A. Beurling, S z ~ r zlne classe de fonctions presgue-pdriodiques, C.R. Acad. Sci. Paris vol.
225 (1947) pp. 326-328. 3. T. Carleman, L'intigrale de Fourier et qz~estions gz~ i s 'y rattachent, Uppsala, 1944. 4. H. Cartan and R. Godement, Thdorie de la dualitd et analyse harnzonigue dans les groupes
abdliens loca2enzent conzpacts, Ann. &ole Norm. vol. 64 (1947) pp. 79-99. 5. V. Ditkin, On the structure of ideals i n certain nornzed rings, cf. Mathematical Reviews
vol. 1 (1940) p. 336. 6. R. Godement, Thiordnzes taz~bk iens et thiorie spectrale, Ann. gcole Norm. vol. 64 (1947)
pp. 119-138. 7. P. R. Halmos, Measure theory, New York, 1950. 8. E. Hille, Functzonal analysis and semi-groups, Amer. Math. Soc. Colloquium Publica-
tions, vol. 31, 1948. 9. I. Kaplansky, Primary ideals i n group algebras, Proc. Nat. Acad. Sci. C.S.A. vol. 35
(1949) pp. 133-136. 10. G. W. Mackey, Functions on locally cotupact groups, Bull. Amer. Math. Soc. vnl. 56
(1950) pp. 385-412. 11. S. Mandelbrojt and S. Agmon, Une gindralisation d z ~ thiordtne tauberien de Wiener,
Acta Szeged vol. 12 B (1950) pp. 167-176. 12. L. Schwartz, Sur zlne propridtd de synthise spectrale dans les grozrpes non compacts, C.R.
Acad. Sci. Paris vol. 227 (1948) pp. 424-426. 13. I. E. Segal, The group algebra o f a locally conzpact group, Trans. Amer. Math. Soc. vol. 61
(1947) pp. 69-105. 14. A. Weil, L'intigration duns les groupes topologigues et ses applications, Paris, 1940. 15. N. Wiener, Taz~berian theorems, Ann. of Math. vol. 33 (1932) pp. 1-100. 16. ---, The Fourier integral and certain of its applications, Cambridge University
Press, 1933.
RICE INSTITUTE, HOUSTON,TEX.
UNIVERSITY OREGON,OF
EUGENE.OREGON