Bourbaki

Exercise 10, § 4

Topological Vector Spaces, V, p. 76

Bourbaki marks this one as harder than the rest.

Let EE be a Hilbert space, and let uu be a positive endomorphism of E. a) Now prove that for every x∈Ex\in E, one has

∥u(x)∥2≤∥u∥⟨u(x)∣x⟩ \|u(x)\|^2\leq\|u\|\langle u(x)|x\rangle
(observe that ⟨u(x)∣u(x)⟩2≤⟨u(x)∣x⟩⟨u2(x)∣u(x)⟩\langle u(x)|u(x)\rangle^2\leq\langle u(x)|x\rangle\langle u^2(x)|u(x)\rangle by virtue of V, p.3, prop. 2). b) Let M be a closed vector subspace of E, and let M∘M^\circ be its orthogonal supplement. Let x∈Mx\in M, and let f(x)f(x) be the greatest lower bound of ⟨u(x+y)∣x+y⟩\langle u(x+y)|x+y\rangle when yy runs through M∘M^\circ. For every ε>0\varepsilon>0, let E(x,ε)E(x,\varepsilon) be the set of the y∈M∘y\in M^\circ such that ⟨u(x+y)∣x+y⟩≤f(x)+ε\langle u(x+y)|x+y\rangle\leq f(x)+\varepsilon. Now prove that E(x,ε)E(x,\varepsilon) is convex and that for every z∈M∘z\in M^\circ, one has, for y∈E(x,ε)y\in E(x,\varepsilon), ()
⟨u(x+y)∣z⟩2≤ε⟨u(z)∣z⟩ \langle u(x+y)|z\rangle^2\leq\varepsilon\langle u(z)|z\rangle
(consider the function g:t↦⟨u(x+y+tz)∣x+y+tz⟩g:t\mapsto\langle u(x+y+tz)|x+y+tz\rangle of the real variable tt, which attains its minimum at a point t0t_0 and note that g(t0)≥f(x)g(t_0)\geq f(x) and g(0)≤f(x)+ε˙g(0)\leq f(x)+\dot{\varepsilon}. c) For every integer n≥1n\geq 1, let yn∈E(x,1/n)y_n\in E(x,1/n); now prove that the sequence (u(x+yn))(u(x+y_n)) tends to a limit x1x_1 which belongs to M, and that the sequence (⟨u(yn)∣yn⟩)(\langle u(y_n)|y_n\rangle) is bounded (majorize the numbers ⟨u(yn−ym)∣yn−ym⟩\langle u(y_n-y_m)|y_n-y_m\rangle for m≥nm\geq n and ∣⟨u(x+yn)∣z⟩∣|\langle u(x+y_n)|z\rangle| for z∈M∘z\in M^\circ by means of the inequality (
)). d) Let (yn′)(y'_n) be a sequence of points of M0M^0 such that ⟨u(yn′−ym′)∣yn′−ym′⟩\langle u(y'_n-y'_m)|y'_n-y'_m\rangle is arbitrarily small as soon as mm and nn are sufficiently large, and such that the sequence (u(x+yn′))(u(x+y'_n)) has a limit x1′∈Mx'_1\in M; now prove that one has x1′=x1x'_1=x_1. (Putting, for brevity, Q(z)=⟨u(z)∣z⟩Q(z)=\langle u(z)|z\rangle, prove first that the number Q((yp−yp′)−(yq−yq′))Q((y_p-y'_p)-(y_q-y'_q)) is arbitrarily small as soon as pp and qq are sufficiently large, and deduce that the sequence (Q(yn−yn′))(Q(y_n-y'_n)) is bounded; using the fact that ⟨x1′−x1∣yp−yp′⟩=0\langle x'_1-x_1|y_p-y'_p\rangle=0 for every pp, prove that the sequence (Q(yn−yn′))(Q(y_n-y'_n)) tends to 0 and use a.).)

e) Deduce from d) that the point x1x_1 does not depend on the choice of the yn∈E(x,1/n)y_n \in E(x, 1/n), and that if one sets u1(x)=x1u_1(x) = x_1, u1u_1 is a linear mapping of MM into itself. Now prove that one has 0≤⟨u1(x)∣x⟩≤⟨u(x)∣x⟩0 \leq \langle u_1(x)|x \rangle \leq \langle u(x)|x \rangle for all x∈Mx \in M, and consequently that u1u_1 is continuous and is an endomorphism ≥0\geq 0 of MM. (Observe that ⟨u(x+yn)∣yn⟩\langle u(x + y_n)|y_n \rangle tends to 0 and ⟨u(x+yn)∣x+yn⟩\langle u(x + y_n)|x + y_n \rangle to f(x)f(x).) f) Let pMp_M be the orthogonal projector with image MM, and let u0=u1∘pMu_0 = u_1 \circ p_M. One has 0≤u0≤u0 \leq u_0 \leq u, MM is stable under u0u_0 and the restriction of u0u_0 to M∘M^\circ is null. Now prove that u0u_0 is the greatest element of the set of endomorphisms v≥0v \geq 0 such that v≤uv \leq u, that MM is stable under vv and that the restriction of vv to M∘M^\circ is null.

Solution

No solution has been written yet.

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