Let E1 be a closed vector subspace of E, stable under u∗u. If E1={0}, then there exists a vector of norm 1 in E1∩P.
Since B is weakly compact, so is the weakly closed subspace B∩E1 of B. Hence there exists (GT, IV, § 6, No. 1, th. 1) a point x0 in B∩E1 such that Q(x0)≥Q(x) for all x∈B∩E1. If Q(x0)=0, we have Q(x)=0 and so u(x)=0 for all x∈B∩E1. Thus E1⊂P and lemma 5 follows in this case.
Suppose now that Q(x0)>0, then x0=0. Since the vector ∥x0∥−1⋅x0 belongs to B∩E1, we have
Q(x0)≥Q(∥x0∥−1⋅x0)=Q(x0)/∥x0∥2 i.e. ∥x0∥=1. We shall prove that
x0 belongs to P; let
y∈E be orthogonal to
x0. It is enough to prove that
u(y) is orthogonal to
u(x0). But since
y is the sum of a vector of
E1 and a vector orthogonal to
E1, and both orthogonal to
x0 (since
x0∈E1), it is enough to consider the following two cases :
a) y is orthogonal to E1: since E1 is stable under u∗u, u∗u(x0)∈E1, hence 0=⟨y∣u∗u(x0)⟩=⟨u(y)∣u(x0)⟩.
b) y belongs to E1: for all t∈R, the vector x(t)=(x0+ty)/∥x0+ty∥ belongs to B∩E1. We have Q(xt)=f(t)/g(t) with
f(t)=∥u(x0)∥2+2tℜ⟨u(x0)∣u(y)⟩+t2∥u(y)∥2g(t)=1+t2∥y∥2. In view of the definition of
x0, we have
Q(x(0))≥Q(x(t)) for all real
t, hence
dtdQ(x(t)) is zero for
t=0. But
f(0)=∥u(x0)∥2,
g(0)=1,
f′(0)=2ℜ⟨u(x0)∣u(y)⟩,
g′(0)=0. Since
dtdQ(x(t))=g(t)2f′(t)g(t)−f(t)g′(t), we conclude that
f′(0)=0, that is,
ℜ⟨u(x0)∣u(y)⟩=0. When
K=R,
u(x0) is orthogonal to
u(y), when
K=C, the vector
iy belongs to
E1 and is orthogonal to
x0, hence
I⟨u(x0)∣u(y)⟩=−ℜ⟨u(x0)∣u(iy)⟩=0, and finally
u(x0) is orthogonal to
u(y). This proves lemma 5.
Now we prove th. 2. Applying th. 1 of S, III, § 4, No. 5 we see, as in V, p. 23, that there exists a set S which is maximal among the orthonormal subsets of E contained in P. Let E1 be the set of all vectors orthogonal to S. Let y∈E1; if x∈S, the vectors x and y are orthogonal, and since S⊂P, we conclude that u(x) and u(y) are orthogonal; then
⟨x∣u∗u(y)⟩=⟨u(x)∣u(y)⟩=0 and
u∗u(y) is orthogonal to
S. Hence
E1 is stable under
u∗u. If we had
E1={0}, there would exist a vector
x of norm 1 in
E1∩P (lemma 5) and
S∪{x} would be an orthonormal subset of
E contained in
P. This would contradict the maximal character of
S. Hence
E1={0} and
S is an orthonormal basis of
E. Q.E.D.
Bourbaki puts the proof directly after the statement, so what is above is the statement and its proof together.
Nothing.