Compact Operators. Spectral Decomposition of Self-Adjoint Compact Operators
6.2 The Riesz–Fredholm Theory
clearlyTεsatisfies (3).
This kind of approximation is very useful, for example, to deduce Schauder’s fixed-point theorem from Brouwer’s fixed-point theorem (see, e.g., K. Deimling [1], A. Granas–J. Dugundji [1], J. Franklin [1], and Exercise 6.26). A similar construction, combined with the Schauder fixed-point theorem, has also been used in a surprising way by Lomonosov to prove the existence of nontrivial invariant subspaces for a large class of linear operators (see, e.g., C. Pearcy [1], N. Akhiezer–I. Glazman [1], A. Granas–J. Dugundji [1], and Problem 42). Another linear result that has a simple proof based on the Schauder fixed-point theorem is the Krein–Rutman theorem (see Theorem 6.13 and Problem 41).
Proposition 6.3.LetE,F, andGbe three Banach spaces. LetT ∈ L(E, F )and S∈K(F, G)[resp. T ∈K(E, F )andS∈L(F, G)]. ThenS◦T ∈K(E,G).
The proof is obvious.
Theorem 6.4 (Schauder).IfT ∈K(E,F ), thenT∈K(F,E). And conversely.
Proof. We have to show thatT(BF)has compact closure inE. Let (vn)be a sequence inBF. We claim that(T(vn))has a convergent subsequence. SetK = T (BE); this is a compact metric space. Consider the setH⊂C(K)defined by
H= {ϕn:x∈K−→ vn, x;n=1,2, . . .}.
The assumptions of Ascoli–Arzelà’s theorem (Theorem 4.25) are satisfied. Thus, there is a subsequence, denoted byϕnk, that converges uniformly onK to some continuous functionϕ∈C(K). In particular, we have
sup
u∈BE
vnk, T u −ϕ(T u)−→0
k→∞ . Thus
sup
u∈BE
vnk, T u − vn, T u−→0
k,→∞, i.e.,Tvnk−TvnE −→0
k,→∞. ConsequentlyTvnk converges inE.
Conversely, assume T ∈ K(F, E). We already know, from the first part, thatT ∈ K(E,F). In particular,T(BE)has compact closure inF. But T (BE)=T(BE)andF is closed inF. ThereforeT (BE)has compact closure inF.
Remark2.LetEandFbe two Banach spaces and letT ∈K(E,F ). If(un)converges weaklytouinE, then(T un)convergesstronglytoT u. The converse is also true if Eis reflexive (see Exercise 6.7).
6.2 The Riesz–Fredholm Theory
We start with some useful preliminary results.
Lemma 6.1 (Riesz’s lemma).LetEbe ann.v.s.and letM⊂Ebe a closed linear space such thatM =E. Then
∀ε >0 ∃u∈Esuch thatu =1and dist(u, M)≥1−ε.
Proof. Letv∈Ewithv /∈M. SinceMis closed, then d =dist(v, M) >0.
Choose anym0∈Msuch that
d ≤ v−m0 ≤d/(1−ε).
Then
u= v−m0
v−m0
satisfies the required properties. Indeed, for everym∈M, we have u−m =
v−m0 v−m0−m
≥ d
v−m0 ≥1−ε, sincem0+ v−m0m∈M.
Remark3.IfMis finite-dimensional (or more generally ifMis reflexive) one can chooseε=0 in Lemma 6.1. But this is not true in general (see Exercise 1.17).
•Theorem 6.5 (Riesz). Let E be an n.v.s. with BE compact. Then E is finite- dimensional.
Proof. Assume, by contradiction, that E is infinite-dimensional. Then there is a sequence (En) of finite-dimensional subspaces of E such that En−1 ⊂ En and En−1 = En. By Lemma 6.1 there is a sequence (un) withun ∈ En such that un =1 and dist(un,En−1)≥ 1/2. In particular,un−um ≥ 1/2 form < n.
Thus(un)has no convergent subsequence, which contradicts the assumption thatBE is compact.
•Theorem 6.6 (Fredholm alternative).LetT ∈K(E). Then (a)N (I−T )is finite-dimensional,
(b)R(I−T )is closed, and more preciselyR(I−T )=N (I −T)⊥, (c)N (I−T )= {0} ⇔R(I−T )=E,
(d) dimN (I −T )=dimN (I −T).
Remark4.The Fredholm alternative deals with the solvability of the equation u−T u=f. It says that
• eitherfor everyf ∈Ethe equationu−T u=f has a unique solution,
• orthe homogeneous equationu−T u=0 admitsnlinearly independent solutions, and in this case, the inhomogeneous equationu−T u=f is solvable if and only iff satisfiesnorthogonality conditions, i.e.,
6.2 The Riesz–Fredholm Theory 161
f ∈N (I−T)⊥.
Remark5.Property (c) is familiar in finite-dimensional spaces. If dimE < ∞, a linear operator fromE into itself is injective(= one-to-one) if and only if it is surjective(=onto). However,in infinite-dimensional spaces a bounded operator may be injective without being surjective and conversely, for example the right shift (resp. the left shift) in2 (see Remark 6). Therefore, assertion (c) is a remarkable property of the operators of the formI−T withT ∈K(E).
Proof.
(a) LetE1=N (I−T ). ThenBE1 ⊂T (BE)and thusBE1is compact. By Theorem 6.5,E1must be finite-dimensional.
(b) Letfn =un−T un→f. We have to show thatf ∈R(I−T ). Setdn=dist(un, N (I −T )). SinceN (I−T )is finite-dimensional, there existsvn∈N (I −T ) such thatdn= un−vn. We have
(4) fn=(un−vn)−T (un−vn).
We claim thatun−vnremains bounded. Suppose not; then there is a subse- quence such thatunk−vnk → ∞. Setwn=(un−vn)/un−vn. From (4) we see thatwnk −T wnk → 0. Choosing a further subsequence (still denoted bywnk for simplicity), we may assume thatT wnk → z. Thus wnk → zand z∈N (I −T ), so that dist(wnk,N (I −T ))→0. On the other hand,
dist(wn, N (I−T ))= dist(un, N (I−T )) un−vn =1 (sincevn∈N (I −T )); a contradiction.
Thusun−vnremains bounded, and sinceT is a compact operator, we may extract a subsequence such thatT (unk−vnk)converges to some limit. From (4) it follows thatunk−vnk →f+. Lettingg=f+, we haveg−T g=f, i.e.,f ∈R(I−T ). This completes the proof of the fact that the operator(I−T ) has closed range. We may therefore apply Theorem 2.19 and deduce that
R(I−T )=N (I −T)⊥, R(I−T)=N (I−T )⊥. (c) We first prove the implication⇒. Assume, by contradiction, that
E1=R(I−T ) =E.
ThenE1is a Banach space andT (E1)⊂ E1. ThusT|E1 ∈ K(E1)andE2 = (I−T )(E1)is a closed subspace ofE1. Moreover,E2 =E1(since(I−T )is injective). LettingEn=(I−T )n(E), we obtain a (strictly) decreasing sequence of closed subspaces. Using Riesz’s lemma we may construct a sequence(un) such thatun∈En,un =1 and dist(un, En+1)≥1/2. We have
T un−T um= −(un−T un)+(um−T um)+(un−um).
Note that ifn > m, thenEn+1⊂En⊂Em+1⊂Emand therefore
−(un−T un)+(um−T um)+un∈Em+1.
It follows thatT un−T um ≥dist(um, Em+1)≥1/2. This is impossible, since T is a compact operator. Hence we have proved thatR(I−T )=E.
Conversely, assume thatR(I−T )=E. By Corollary 2.18 we know thatN (I− T)=R(I−T )⊥= {0}. SinceT∈K(E), we may apply the preceding step to infer thatR(I −T)= E. Using Corollary 2.18 once more, we conclude thatN (I −T )=R(I−T)⊥= {0}.
(d) Setd = dimN (I −T ) andd = dimN (I −T). We will first prove that d ≤ d. Suppose not, thatd < d. SinceN (I −T )is finite-dimensional, it admits a complement inE (see Section 2.4, Example 1). Thus there exists a continuous projectionP fromEontoN (I−T ). On the other hand,R(I−T )= N (I−T)⊥has finite codimensiond(see Section 2.4, Example 2) and thus it has a complement (inE), denoted byF, of dimensiond. Sinced < d, there is a linear map : N (I −T ) → F that isinjectiveandnot surjective. Set S=T +◦P. ThenS∈K(E), since◦P has finite rank.
We claim thatN (I−S)= {0}. Indeed, if
0=u−Su=(u−T u)−(◦P u), then
u−T u=0 and ◦P u=0, i.e.,u∈N (I−T )andu=0. Therefore,u=0.
Applying (c) to the operatorS, we obtain thatR(I−S)=E. This is absurd, since there exists somef ∈ F withf /∈R(), and so the equationu−Su=f has no solution.
Hence we have proved thatd≤d. Applying this fact toT, we obtain dimN (I−T)≤dimN (I−T)≤dimN (I−T ).
ButN (I −T)⊃N (I−T )and therefored=d.