Let X be a Polish space and let R be an equivalence relation on X, such that the equivalence classes mod R are closed in X and the saturation (with respect to R) of each closed set in X is a Borel set. Then there is a Borel set in X which meets each equivalence class in exactly one point.
Consider a metric on X compatible with its topology, and with respect to which X is complete. By Lemma 3 of no. 5, there exists a sifting of X, defined by a sieve C=(Cn,pn) and a sequence of mappings (φn). For each c∈Cn, let gn(c) be the saturation of the closed set φn(c) with respect to R; by hypothesis, gn(c) is a Borel set in X.
Since each set Cn is countable, we can linearly order each Cn in such a way that the set of elements smaller than a given element is finite. For each c∈Cn we define a set hn(c) by induction on n, as follows. In the first place, for c∈C0, h0(c) is the intersection of φ0(c) and the sets X−g0(c′), where c′∈C0 and c′<c. For c∈Cn+1, hn+1(c) is the intersection of φn+1(c),hn(Pn(c)) and the sets X−gn+1(c′) for c′∈Cn+1, pn(c′)=pn(c) and c′<c. The hn(c) are clearly Borel sets.
We shall prove the following assertion: for each integer n≥0 and each equivalence class H mod R, there is a unique element c∈Cn such that hn(c) meets H, and we have
h(cn)∩H=φn(c)∩H, which is therefore a
closed set. For
n=0, consider the smallest element
c∈C0 such that
φ0(c) meets
H; then
φ0(c)∩H does not meet any set
g0(c′) for which
c′∈C0 and
c′<c; hence it is contained in
h0(c)∩H and consequently is equal to this set; moreover, we have
H⊂g0(c) and therefore
h0(c′)∩H is empty for
c′∈C0 and
c′>c; thus the assertion is proved for
n=0. We continue by induction on
n: if there exists
c∈Cn+1 such that
hn+1(c) meets
H, then it follows from the relation
hn+1(c)⊂hn(pn(c)) and the inductive hypothesis that
pn(c) is the unique element
d∈Cn such that
hn(d) meets
H. Observe that
hn(d), which is contained in
φn(d), is contained in the union of the sets
φn(c) for which
c∈pn−1(d), by the definition of a sifting; there is therefore a smallest element
c∈pn−1(d) such that
φn+1(c) meets
H. We have therefore
φn+1(c)∩H⊂φn(d)∩H=hn(d)∩H by the inductive hypothesis. Hence
φn+1(c)∩H⊂φn+1(c)∩hn(d), and since by definition
φn+1(c)∩H meets none of the sets
gn+1(c′) for which
c′∈pn−1(d) and
c′<c, it follows from the definition of
hn+1(c) that
φn+1(c)∩H=hn+1(c)∩H. Moreover, we have
H⊂gn+1(c) and therefore, if
c′∈pn−1(d) is such that
c′>c,
hn+1(c′)∩H is empty. Hence the assertion is proved for all
n.
For each integer n, let Sn be the union of the sets hn(c), where c runs through Cn. The set Sn is a Borel set, and we have Sn+1⊂Sn.
Let S be the intersection of the sets S_n, which is also a Borel set in X; we shall show that S meets each equivalence class H mod R in exactly one point. For each n, let c_n(H) be the unique element c \in C_n such that h_{c_n(H)} meets H; then S_n \cap H = \varphi_n(c_n(H)) \cap H, and S \cap H is the intersection of the sets \varphi_n(c_n(H)) \cap H. Since the sequence (c_n(H)) belongs to L.C., the decreasing sequence of closed sets \varphi_n(c_n(H)), whose diameter tends to 0, converges to a point x \in X, since X is complete. The intersection of the closed sets \varphi_n(c_n(H)) \cap H therefore consists of the point x alone, and the proof of Theorem 4 is complete.
Bourbaki puts the proof directly after the statement, so what is above is the statement and its proof together.