Bourbaki

Proposition 2

Tag 02TS, permanent. Cite it as [Bourbaki, Tag 02TS].

Théories spectrales, chapter III, § 5. Perturbation par une application linéaire compacte

Let u∈L(E)u\in \mathscr{L}(E) be such that ∥1E−π(u)∥<1\|1_E-\pi (u)\|<1 in the algebra C\mathscr{C}alk(E). Then uu is a Riesz endomorphism of E.

Let r⩾1r\geqslant 1 be an integer such that ∥1E−π(u)∥r<12\|1_E-\pi (u)\|^r<\frac{1}{2}. Let P∈C[X]P\in \mathbf{C}[X] be the polynomial 1−(1−X)rX\frac{1-(1-X)^r}{X}. Let v=1E−(1E−u)rv= 1_E-(1_E-u)^r. We thus have v=uP(u)v=uP(u) and ∥1E−π(v)∥<12\|1_E-\pi (v)\|<\frac{1}{2}. Since the endomorphisms uu and P(u)P(u) of E commute, it suffices to prove that vv is a Riesz endomorphism of E (III, p. 49, prop. 9).

Since ∥1E−π(v)∥<1/2\|1_E-\pi (v)\|<1/2, there exists, by definition of the quotient norm in the space C\mathscr{C}alk(E), a compact endomorphism hh of E and an endomorphism ww of E such that v=1E+h+wv= 1_E+h+w and ∥w∥<12\|w\|<\frac{1}{2}. According to Corollary 1 of I, p. 22, the element 1E+w1_E+w is an automorphism of E. Since hh is compact, v=(1E+w)+hv= (1_E+w) +h is a Fredholm endomorphism of E of index 0 (Cor. 2 of III, p. 74). For every integer n⩾0n\geqslant 0, denote by NnN_n the kernel of vnv^n. To prove that vv is a Riesz endomorphism of E, it suffices to prove that there exists an integer n⩾0n\geqslant 0 such that Nn=Nn+1(N_n= N_{n+1}(III, p. 46, def. 2 and III, p. 46, remark).

Reason by contradiction by supposing the sequence (Nn)(N_n) strictly increasing. For every n∈Nn\in \mathbf{N}, let pnp_n be the canonical mapping of E onto the normed space E/NnE/N_n. Let cc be a real number such that 2∥w∥<c<12\|w\|< c <1. Let n∈Nn\in \mathbf{N}. Since Nn+1N_{n+1} is different from NnN_n, there exists an element xn∈Nn+1x_n\in N_{n+1} such that ∥pn(xn)∥=c\|p_n(x_n)\|=c and such that ∥xn∥<1\|x_n\|<1 (indeed, there exists y∈Nn+1/Nny\in N_{n+1}/N_n of norm cc so, for every ε>0\varepsilon >0, there exists xn∈Nn+1x_n\in N_{n+1} such that pn(xn)=yp_n(x_n) =y and ∥xn∥⩽c+ε\|x_n\|\leqslant c+\varepsilon ).

Let m,nm,n be two integers such that m>n⩾0m > n\geqslant 0. We have

∥h(xm)−h(xn)∥=∥v(xm−xn)−(1E+w)(xm−xn)∥ \|h(x_m)-h(x_n)\|=\|v(x_m-x_n)-(1_E+w)(x_m-x_n)\|
⩾∥(v−1E)(xm−xn)∥−∥w(xm−xn)∥ \geqslant \|(v-1_E)(x_m-x_n)\| - \|w(x_m-x_n)\|
⩾∥(v−1E)(xm−xn)∥−2∥w∥(1) \geqslant \|(v-1_E)(x_m-x_n)\| -2\|w\| \tag{1}

since ∥xm∥⩽1\|x_m\|\leqslant 1 and ∥xn∥⩽1\|x_n\|\leqslant 1. Moreover, note that

(v−1E)(xm−xn)=v(xm−xn)+xn−xm (v-1_E)(x_m-x_n) =v(x_m-x_n) +x_n-x_m

But since n<mn < m and v(Nm+1)⊂Nmv(N_{m+1})\subset N_m, we have v(xm−xn)+xn∈Nmv(x_m-x_n) +x_n\in N_m, whence

∥v(xm−xn)+xn−xm∥⩾∥pm(xm)∥=c \|v(x_m-x_n) +x_n-x_m\|\geqslant \|p_m(x_m)\|=c

The inequality (1) therefore gives the lower bound

∥h(xm)−h(xn)∥⩾c−2∥w∥>0(2) \|h(x_m)-h(x_n)\|\geqslant c-2\|w\|>0 \tag{2}

The sequence (h(xm))m∈N(h(x_m))_{m\in\mathbf{N}} therefore has no value of adherence in E, which contradicts the fact that the image by hh of the unit ball of E is relatively compact, since hh is compact (remark 1 of III, p. 2).

Bourbaki sets the proof directly after the statement, and the corpus does not separate them, so what is above is the statement and its proof together.

Cited by

Nothing.

Cites

Nothing.

Machine translation of the English, by gpt-5-mini, gpt-5-6-mini, against glossary version 34. Not checked by a person. That model is a cut down version of the one the rest of this language was written by.