1 and L:9¢-+F is defined by L({a,})=ay, find the vector fg in 3 such that L(h) = (h, ho) for every h in #. 3. Let 9=—(N u (O}). (a) Show that if {a,}€/°, then the power series E%_a, 2” has radius of convergence > 1. (b) If |AJ< 1 and L:#C is defined by L({a,}) = L%-oa,X", find the vector hg in ¥@ such that L(h) = (h,ho) for every h in 2. (c) What is the norm of the linear functional L defined in (b)? 4, With the notation as in Exercise 3, define L: > € by L({a,})=E2 na y"!, where |A|< 1. Find a vector ho in #% such that L(h) = (h,ho) for every h in 3. 5, Let J be the Hilbert space described in Example 1.8. If 0 <1 < 1, define L: 9+ F by Lth) = A(z), Show that L is a bounded linear functional, find ||LI}, and find the vector hg in 3% such that L(h) = (h, fo) for all A in 3, 6 Let #=L7(0, 1) and let C™ be the set of all continuous functions on [0, 1] that have a continuous derivative. Let t€ [0,1] and define L:C”—F by L(h) = A(t). Show that there is no bounded linear functional on 3% that agrees with L on C),§4. Orthonormal Sets of Vectors and Bases It will be shown in this section that, as in Euclidean space, each Hilbert space can be coordinatized. The vehicle for introducing the coordinates is an orthonormal basis. The corresponding vectors in F4 are the vectors {€1,€5,-.., a}, where e, is the d-tuple having a 1 in the & th place and zeros elsewhere. 4.1. Definition. An orthonormal subset of a Hilbert space # is a subset & having the properties: (a) for e in &,|lel| = 1; (b) if e;,e) € 6 and e, #e3, then e; Le. A basis for # is a maximal orthonormal set. Every vector space has a Hamel basis (a maximal linearly independent set). The term basis for a Hilbert space is defined as above and it relates to the inner product on #. For an infinite-dimensional Hilbert space, a basis is never a Hamel basis. This is not obvious, but the reader will be able to see this after understanding several facts about bases. 4.2. Proposition. If € is an orthonormal set in #, then there is a basis for # that contains &. The proof of this proposition is a straightforward application of Zorn $ Lemma and is left to the reader. 4.3. Example. Let #= [0,27] and for n in Z define e, in # by e(t) = 7) exp(int). Then {e,: 2 ©Z} is an orthonormal set in #. (Here L2[0,22] is the space of complex-valued square integrable functions.) It is also true that the set in (4.3) is a basis, but this is best proved after a bit of theory. 4.4, Example. If #?=F4 and for 1 | let &, = {e € I: |(h,e)|2 I/n}. By Bessel $ Inequality, &, is finite. But U%_,é, = {e © &: (h, e,) # O}. . 4.10. Corollary. Zf€ is an orthonormal set and h © #, then Likh,ey? s tal? eed This last corollary is just Bessel $ Inequality together with the fact (4.9) that at most a countable number of the terms in the sum differ from zero. Actually, the sum that appears in (4.10) can be given a better interpreta- tion- a mathematically precise one that will be useful later. The question is,what is meant by D{h;: i€ 1} if h, © # and J is an infinite, possibly uncountable, set? Let ¥ be the collection of all finite subsets of J and order F by inclusion, so ¥ becomes a directed set. For each F in ¥, define =D {aie F}. Since this is a finite sum, h, is a well-defined element of #. Now {hy: “FeEF¥)isanetin #. 4.11. Definition. With the notation above, the sum L{h,: i € J} converges if the aet {4,: F € F} converges; the value of the sum is the limit of the net. If #?= F, the definition above gives meaning to an uncountable sum of scalars. Now Corollary 4.10 can be given its precise meaning; namely, L{| 0, there is an N such that D> y\(h,e, yi < ee. Let R= {€,---,@y-1} and let #= all the finite subsets of &. For F in F define hp= Eh, eye: e€ F}. If F and G © and both contain Fy, then Wap = hel? = L {Ich e)P: e € (F\G) U(G\ F)} s r KCA,enyP? n=N (b): Suppose h L@ and h ¥ 0; then €U{A/|All} is an orthonormal set that properly contains @, contradicting maximality. (b) = (c): By Corollary 2.11, V8=# if and only if &+ = (0). (b) = @): If A EH, then f =h-LX{ (h, e)e:e SF} is a well-defined vector by Lemma 4.12. If e, ©@, then (f, e;) = (h, e))— L{(h,e){e, e,): e€&} = (h, e,)— (h, e)) = O. That is, f€&+. Hence f = 0. (Is every- thing legitimate in that string of equalities? We dont want any illegitimate equalities.) (d) = (e): This is left as an exercise for the reader. (e) =(f): Since Wall? = (h, h), this is immediate. (f)= (a): If & is not a basis, then there is a unit vector eo ({leol| = 1) in 9% such that eg 1 &. Hence, 0 = L{|(e,e)|:e€@}, contradicting (A. . Just as in finite-dimensional spaces, a basis in Hilbert space can be used to define a concept of dimension. For this purpose the next result is pivotal. 4.14. Proposition. Zf # is a Hilbert space, any two bases have the same cardinality. Proor. Let 6 and F¥ be two bases for # and put e = the cardinality of &, 74 = the cardinality of #. If e or q is finite, then e = q (Exercise 15). Suppose both e and q are infinite. For e in @, let A={ feF:(e p # O}; so , is countable. By (4.13b), each fin F belongs to at least one set F,,e in &. That is, F=ULF:e€}. Hence n yn =—(=+ -1)". Pax) aril ae) (>) The functions P,(x) are called Legendre polynomials. 7. Ifthe sequence exe", We" /?, is orthogonalized in L2(—00, 00), the sequence e,,(x) = (2"nlVa}/7H,(x)e-* ” is obtained, where ° nf 2)" A(x) = (-1)"e"(Z) The functions H, are Hermite polynomials and satisfy H;(x) =2 nH, ( x) « 8. If the sequence e~*/?, xe~*/?, x?e~*/?,.. is orthogonalized in L? (0,00), the sequence e,(x) = e~*/7L, (x)! is obtained, where L(x) = e(s) (x"e"*). ‘The functions L,, are called Laguerre polynomials. 9, Prove Corollary 4,10 using Definition 4.11. 10. If {h,} is a sequence in Hilbert space and L{h,:2€N)} converges to h (Definition 4.11), then lim,Dj_ ,h, = h. Show that the converse is false.LL. If {A,} is a sequence in a Hilbert space and £%.,||h,|]<00, show that (h,: n&N} converges in the sense of Definition 4.11. 12. Let {a,,} be a sequence in F and prove that the following statements are equivalent: (a) L{ a.: nN} converges in the sense of Definition 4.11. (b) If is any permutation of N, then L%_,a,,,) converges (unconditional convergence). (6) Dilan] < 00. 13, Let & be an orthonormal subset of # and let M = Vé.If P is the orthogonal projection of #% onto M, show that Ph = L{ (h, ee: e €&} for every h in #, 14, Let A = Area measure on D and show that 1, z, z?, . . . are orthogonal vectors in L*(A). Find |[z"|,n> 0. If e, = |Iz"I|-'z", n> 0, is {e., e,.} a basis for L?(A)? 15. In the proof of (4.14), show that if either e or 7 is finite, then e= 9 16. If # is an infinite-dimensional Hilbert space, show that no orthonormal basis for # is a Hamel basis. Show that a Hamel basis is uncountable. 17. Let d> | and let p be a regular Borel measure on R4. Show that L?(p) is separable. 18. Suppose L?(X,2,q) is separable and { EB: i € J} is a collection of pairwise disjoint subsets of X,£,€, and 0 X such that (Uh, Ug) = ¢h, 8) for all h, g in J@. In this case # and X are said to be isomorphic. It is easy to see that if U: 9? # is an isomorphism, then so is U~?: X— #. Similar such arguments show that the concept of isomorphic is an equivalence relation on Hilbert spaces. It is also certain that this is thecorrect equivalence relation since an inner product is the essential ingredient for a Hilbert space and isomorphic Hilbert spaces have the Same inner product. One might object that completeness is another essential ingredient in the definition of a Hilbert space. So it is! However, this too is preserved by an isomorphism. An isometry between metric spaces is a map that preserves distance. 5. Proposition. [f V: 3°— X is a linear map between Hilbert spaces, then V is an isometry if and only if (Vh, Vg) = (h, g) for all h, g in #. Proor. Assume (Vh, Vg) = (h, g) for all h, g in 4. Then Val)? = #¢ that is a surjective isometry. That is, a unitary operator is an isomorphism whose range coincides with its domain, This may seem to be a minor distinction, and in many ways it But experience has taught me that there is some benefit in making such a distinction, or at least in being aware of it. 5.4. Theorem. Two Hilbert spaces are isomorphic if and only if they have the same dimension. Proor. If U:#— & is an isomorphism and @ is a basis for #, then it is easy to see that UP ={Ue:e €@} is a basis for #. Hence, dim #= dim %.Let # be a Hilbert space and let & be a basis for #. Consider the Hilbert space 1°(&). If he #, define h:@>F by h(e) = (h,e). By Parseval’s Identity he/?(&) and |jAl| = |||]. Define U: > 12(&) by Uh =k. Thus U is linear and an isometry. It is easy to see that ranU contains all the functions f in /2(&) such that f(e) = 0 for all but a finite number of e; that is, ranU is dense. But U, being an isometry, must have closed range. Hence U: #17(&) is an isomorphism. If X is a Hilbert space with a basis #,X is isomorphic to /7(F). If dim #= dim #,& and ¥ have the same cardinality; it is easy to see that P(@)and 1?(.F) must be isomorphic. Therefore # and X are isomorphic. . 5.5. Corollary. All separable infinite dimensional Hilbert spaces are isomor- phic. This section concludes with a rather important example of an isomor- phism, the Fourier transform on the circle. The proof of the next result can be found as an Exercise on p. 263 of Conway [1978]. Another proof will be given later in this book after the Stone-Weierstrass Theorem is proved. So the reader can choose to assume this for the moment. Let D = {z €C:|z|<1}. 5.6. Theorem. Zf f:@D—C is a continuous function, then there is a sequence { p,( 2, Z)} of polynomials in z and Z such that p,(2,2)— flz) uniformly on dD. Note that if z €9D,2= 27". Thus a polynomial in z and Z on 0D becomes a function of the form YL a2*. k= —m If we put z = e”, this becomes a function of the form n Y ae, m k: Such functions are called trigonometric polynomials, We can now show that the orthonormal set in Example 4.3 is a basis for L@[0, 27]. This is a rather important result. 5.7. Theorem. Zf for each n in Z, e(t) =(27) \/exp(int), then fe,: n€Z} is a basis for L2[0,27]. Proor, Let 7={LZ__,,e,: & EC, n> O}. Then FJ is a subalgebra of C,[0, 27], the algebra of all continuous C-valued functions on [0,27]. Note that if (ET, fO) = f(27). We want to show that the uniform closure of 7is @={ f EC. [0, 27]: (0) = f(27)}. To do this, let fE@ and define F: 8D >C by Fle’) = ft). F is continuous. (Why?) By (5.6) there is a sequence of polynomials in z and 2,{ p,(z,Z)}, such that p,(z,Z)—> F(z) uniformly on D. Thus p,(e",e~") > f(t) uniformly on [0.2m]. But DAee "ETF. Now the closure of @ in L2[0,27] is all of L2[0,27] (Exercise 6). Hence V{ e: nEZ}=L2[0,27] and {e,} is thus a basis (4.13). . Actually, it is usually preferred to normalize the measure on [0,22]. That is, replace dt by (27)~' dt, so that the total measure of [0,27] is 1. Now define e,(t) = exp(int). Hence {e,: n€Z} is a basis for #= L&((0, 27], (27) 1 dt). If f EH, then /(n) = = p(y 58 Fn) = (Seen) = a5 [Pde ae is called the nth Fourier coefficient of f,n in Z. By (5.7) and (4.134), « 59 f= YL flaje a where this infinite series converges to f in the metric defined by the norm of 3. This is called the Fourier series of f. This terminology is classical and has been adopted for a general Hilbert space. If # is any Hilbert space and @ is a basis, the scalars {¢h, e); e © &} are called the Fourier coefficients of h (relative to &) and the series in (4.13d) is called the Fourier expansion of h (relative to 8). . Note that Parseval $ Identity applied to (5.9) gives that D%_ ,If(n)/?< infinity. This proves a classical result. 5.10. The Riemann-Lebesgue Lemma. If f€L7[0,27], then {oF (the ™ dt > 0 as n > 400. If f © L2[0, 27], then the Fourier series of f converges to f in L?-norm. It was conjectured by Lusin that the series converges to f almost every- where. This was proved in Carleson [1966]. Hunt [1967] showed that if f €L€[0, 27], 1

C is called the Fourier transform of f the map U:L2[0,22]—> 1°(Z) defined by Uf = f is the Fourier transform. The results obtained so far can be applied to this situation to yield the following. 5.11. Theorem. The Fourier transform is a linear isometry from L2(0,27] onto 1?(Z).Proor. Let U: L2[0,27]->/°(Z) be the Fourier transform. That U maps L? = 12{0,27] into /2(Z) and satisfies ||Uf|| =|[fl| is a consequence of Parseval § Identity. That U is linear is an exercise. If {a,}€/?(Z) and a, = 0 for all but a finite number of n, then f = D%_ ,a,e,€ L?. It is n= ~ ook easy to check that f(n) = a, for all n, so Uf = {a, }. Thus ranU is dense in P. But U is an isometry, so ranU is closed; hence U is surjective. = Note that functions in L@[0,27] can be defined on AD by letting f(ei®) = f(8). The ambiguity for 8 = 0 and 2a (or e” = 1) might cause us to pause, but remember that clements of L& [0,27] are equivalence classes of functions-not really functions. Since {0,27} has zero measure, there is really no ambiguity. In this way L2[0,27] can be identified with L&( dD), where the measure on @D is normalized arc-length measure (normalized so that the total measure of AD is 1). So L2[0,27] and L2 ( AD) are (naturally) isomorphic). Thus, Theorem 5.11 is a theorem about the Fourier transform of the circle. The importance of Theorem 5.11 is not the fact that L?[0,27] and /?(Z) are isomorphic, but that the Fourier transform is an isomorphism. The fact that these two spaces are isomorphic follows from the abstract result that all separable infinite dimensional Hilbert spaces are isomorphic (5.5). EXERCISES 1. Verify the statements in Example 5.3. 2. Define V: L?(0, 00) > L?(0, 00) by (Vf\(t) = f(r + 1). Show that V is an isometry that is not sujective. . Define V: L?(R) > L?(R) by (Vf)(t) = f(t + L) and show that V is an isomor- phism (a unitary operator). 4. Let # be the Hilbert space of Example 1.8 and define U: 9 L?(0, 1) by uf = f’. Show that U is an isomorphism and find a formula for U7}. 5. Let (X, @,p) be a w-finite measure space and let u:X > IF be an ~measurable function such that sup{ |u(x)|: « € X) < co, Show that U:L?(X,2,p)> L?(X,Q, w) defined by Uf = uf is an isometry if and only if |u(x)|= 1 ae. (#4) in which case U is sujective. 6, Let @= {f ©C[0,27}: f(O) = fQm)} and show that € is dense in L?(0,27]. 7. Show that ((1/ V2), (1/ Var)cos nt,(1/ Vm )sin nt: | L?(n) by Vf = Yof. Show that V is a well-defined linear isometry and V is an isomorphism if and only if p< v (that is, p and v are mutually absolutely continuous).24 §6. The Direct Sum of Hilbert Spaces Suppose 9 and ¥ are Hilbert spaces. We want to define #@ X s0 that it becomes a Hilbert space. This is not a difficult assignment. For any vector spaces # and Y,%@® Y is defined as the Cartesian product 2 x Y where the operations are defined on 2X Y coordinatewise. That is, if elements of 2 Y are defined as {x ®@y:x EX, ye Y}, then (x, @ y,) +(x, @ yy) = (x, + X,)@(¥, + 2), and so on. 6.1. Definition. If and & are Hilbert spaces, #® H={h@k: hE H#H, kX} and (hy ® ky, hy ® ky) = (hy, ha) + Chis ka). It must be shown that this defines an inner product on #@ % and that #8 HX is complete (Exercise). Now what happens if we want to define #,®#,@--+ for a sequence of Hilbert spaces #,, %,...? There is a problem about the completeness of this infinite direct sum, but this can be overcome as follows. 6.2. Proposition. If 3, %,... are Hilbert spaces, let H= {(h ,)% h, € #, for all n and L2.,\\h,\\? < 00}. For h = (h,) and g = (g,) in #, dejine 6.3 (h, 8) = ¥ Chas &)- n=l Then <-,+) is an inner product on 3 and the norm relative to this inner product is \Al| = (2% ,\t,t(7.With this inner product # is a Hilbert space. Proor. If h = (h,) and g = (g,.) © #, then the CBS inequality implies LAs Bu 1S LU Mall Mall $ LUA gll?)'/7 (Lig all?)'/? < 00. Hence the series in (6.3) converges absolutely. The remainder of the proof is left to the reader. . 64. Definition. If #,,#%4,... are Hilbert spaces, the space # of Proposi- tion 6.2 is called the direct sum of #,,96,,... and is denoted by #= #, OH,O---. This is part of a more general process. If {#:i€ J} is a collection of Hilbert spaces, #=@{ #f:i€/} is defined as the collection of function: h: 1>U{ #@:i€ I} such that A(i) © % for all i and L{A(i)I)? i€ I} <0. Ih, g €H#, (h, 9) =L{(ACi), Qi): i EI}; H is a Hilber space.The main reason for considering direct sums is that they provide a way of manufacturing operators on Hilbert space. In fact, Hilbert space is a rather dull subject, except for the fact that there are numerous interesting ques- tions about the linear operators on them that are as yet unresolved. This subject is introduced in the next chapter. EXERCISES _ 1, Let {(X,,@,,p,):i 1) be a collection of measure spaces and define X, 2, and p as follows. Let X = the disjoint union of {X;:i€ 1) and let Q= {A Cc X: An X,€Q, for all ij. For A in @ put p(A)=,,(A n X,). Show that (X, &, p) is a measure space and L?( X, @, 1) is isomorphic to { L?(X,, 2, u,): ie I). 2. Let (X, 2) be a measurable space, let #;, #. be measures defined on (X, 2), and put #=p,+p,. Show that the map V:L7(X,2,p)>L?( X, 2,4) ® L?(X,Q,p5) defined by Vf = f;®@f,, where f, is the equivalence class of L?(X, 2,,) corresponding to f, is well defined, linear, and injective. Show that U is an isomorphism iff p, and p, are mutually singular.CHAPTER II Operators on Hilbert Space A large area of current research interest is centered around the theory of operators on Hilbert space. Several other chapters in this book will be devoted to this topic. There is a marked contrast here between Hilbert spaces and the Banach spaces that are studied in the next chapter. Essentially all of the information about the geometry of Hilbert space is contained in the preceding chapter. The geometry of Banach space lies in darkness and has attracted the attention of many talented research mathematicians. However, the theory of linear operators (linear transformations) on a Banach space has very few general results, whereas Hilbert space operators have an elegant and well- developed general theory. Indeed, the reason for this dichotomy is related to the opposite status of the geometric considerations. Questions concerning operators on Hilbert space dont necessitate or imply any geometric difficul- ties. In addition to the fundamentals of operators, this chapter will also present an interesting application to differential equations in Section 6. §1. Elementary Properties and Examples The proof of the next proposition is similar to that of Proposition 1.3.1 and is left to the reader. Ll. Proposition. Let # and X be Hilbert spaces and A: #—> X a linear transformation. The following statements are equivalent. (a) A is continuous. (b) A is continuous at 0.(c) A is continuous at some point. (d) There is a constant c > 0 such that ||AA|| 0: ||AA|| L*(u) by Myf = of. Then My B(L"(u)) and |Mgll = ll9llo- Proor. Here ||$l|,,. is the ~-essential supremum norm. That is, lIéllae = inf{sup{lo(x)|: x € NJ: N © @, p(N) = 0} = inf{c > 0: w({x © X: |(x)|> c}) = O}. Thus Il9ll . is the infimum of all c > 0 such that |p(x)j 0, the u-finiteness of the measure space implies that there is a set A in 2,0 0, we get that |IM,|| > lisll..» a The operator M, is called a multiplication operator. The function ¢ is its symbol. If the measure space (X, 92,) is not a-finite, then the conclusion of Theorem 1.5 is not necessarily valid. Indeed, let @ = the Borel subsets of [0, 1] and define » on 2 by p(4) =the Lebesgue measure of A if 0 ¢ A and p(4) =00 if0 € A. This measure has an infinite atom at 0 and, therefore, is not u-finite. Let @=x,o). Then ¢€ L*(u) and |l9ll,, = 1. If FE L*(n), then o> /If]? du =|f()I?u({0}). Hence every function in L?(#) vanishes at 0. Therefore M, =0 and ||M,J|<|ldll,.- There are more general measure spaces for which (1.5) is valid-the decomposable measure spaces (see Kelley [1966]). 1.6. Theorem. Let ( X, 9,4) be a measure space and suppose k: X x X>F isan Q X Q-measurable function for which there are constants cand c, such that fies, vildu(y)se, ae.[n], {ks ri du(x) ser ae. [a].If K: L(w) > L*(u) is defined by (KF )(x) = fee »)F(0) del), then K is a bounded linear operator and ||K||<(¢\C2)'”?. Proor. Actually it must be shown that Kf €L?(), but this will follow from the argument that demonstrates the boundedness of K. If FE L7(p), IK s fle, YIVOV dey) = flex, A(x, 970) der) 1/2| 1/2 «| fies an6o9] [fees rar) <6? [ fies mvorran(oy] ”. Hence SKF u(x) < ef flex, VOIP du(y) d(x) = flO? fies vide) du(y) < eel’. Now this shows that the formula used to define Kf is finite ac. [p], Kf © L(y), and KAP seycalfl’. The operator described above is called an integral operator and the function & is called its kernel. There are conditions on the kernel other than the one in (1.6) that will imply that Kis bounded. A particular example of an integral operator is the Volterra operator defined below. 1.7. Example. Let &: f0, 1] x [0,1] >R be the characteristic function of {(x, y): y < x}. The corresponding operator V:L?(0, 1) > L?(0, 1) defined by Vf (@) = fdk(x, fl y) dy is called the Volterra operator. Note that 4x) = L104. Another example of an operator was defined in Example 1.5.3. The nonsurjective isometry defined there is called the unilateral shift. It will be studied in more detail later in this book. Note that any isometry is a bounded operator with norm 1.EXERCISES 1. Prove Proposition 1.1. 2. Prove Proposition 1.2. 3. Suppose {e, } is an orthonormal basis for #% and A: 3% 2X is a linear transformation such that E|[Ae,||<00. Show that A is bounded. 4, Proposition 1.2 says that d(A, B) = ||A— Bll is a metric on B(#, X). Show that @(3, X) is complete relative to this metric. 5. Show that a multiplication operator M, (1.5) satisfies M2 = M, if and only if g is a characteristic function. 6. Let (X, @,q) be a measure space and let ky,k> be two kernels satisfying the hypothesis of (1.6). Define ki XX XE byk(x.y) = fhi(x,2)ka(z,¥) du(z). (a) Show that k also satisfies the hypothesis of (1.6). (b) If K, K,,K are the integral operators with kernels k, k,, kz, show that K = K, K,. What does this remind you of? Is more going on than an analogy? 7. If (X, @,p) is a measure space and k © L7(u X w), show that k defines a bounded integral operator. 8. Let (e,} be the usual basis for /? and let {a, } be a sequence of scalars, Show that there is a bounded operator A on /? such that Ae, = a@,e, for all n if and only if {a, } is uniformly bounded, in which case ||A|] = sup{ |a,|:n> 1). This type of operator is called a diagonal operator or is said to be diagonalizahle. 9. (Schur test) Let (@,,}%j.1 be an infinite matrix such that a, 2 0 for all i, j and such that there are scalars p, > 0 and B,y> 0 with we LY a,,p, < Bp,» i=l LX ap, <1, jr for all i, j = 1. Show that there is an operator A on P(N) with (Ae,,e, )= a), and |\A||° < By. 10. (Hilbert matrix) Show that (Ae,, @) = (i + j +1) 7! for 0