Bourbaki

§ 2. Quotient of a space by a group; homogeneous spaces

1. General results

Let X be a locally compact space in which a locally compact group H operates on the right, continuously and properly, by (x,ξ)↦xξ(x, \xi) \mapsto x\xi (x∈X,ξ∈Hx \in X, \xi \in H). The equivalence relation in X defined by H is open (GT, III, §2, No. 4, Lemma 2) and X/H is Hausdorff (loc. cit., §4, No. 2, Prop. 3) hence locally compact (GT, I, §10, No. 4, Prop. 10). We denote by π\pi the canonical mapping of X onto X/H. The saturation of a subset Y of X is YH=π−1(π(Y))YH = \pi^{-1}(\pi(Y)). If K is a compact subset of X, then π(K)\pi(K) is compact and the saturation π−1(π(K))\pi^{-1}(\pi(K)) of K is closed in X. Every compact subset of X/H is the image under π\pi of a compact subset of X (GT, I, §10, No. 4, Prop. 10). We assume given once and for all a left Haar measure β\beta on H.

Let χ\chi be a continuous representation of H in R+∗\mathbf{R}_+^*. If a function g on X satisfies g(xξ)=χ(ξ)g(x)g(x\xi) = \chi(\xi)g(x) for all x∈Xx \in X and ξ∈H\xi \in H, its support S is invariant under H hence may be written π−1(π(S))\pi^{-1}(\pi(S)). We shall denote by Kχ(X)\mathscr{K}^\chi(X) the Riesz space formed by the continuous real-valued functions g on X that satisfy g(xξ)=χ(ξ)g(x)g(x\xi) = \chi(\xi)g(x) (x∈X,ξ∈Hx \in X, \xi \in H) and whose support is the saturation of a compact subset of X; we denote by K+χ(X)\mathscr{K}_+^\chi(X) the set of elements ⩾0\geqslant 0 of Kχ(X)\mathscr{K}^\chi(X). In particular, K1(X)\mathscr{K}^1(X) is none other than the set of continuous functions on X, constant on the orbits, whose support is the saturation of a compact subset.

Proposition 1

Let f be a continuous real-valued function on X whose support S has compact intersection with the saturation of every compact subset of X.

a) For every x∈Xx \in X, the function ξ↦f(xξ)\xi \mapsto f(x\xi) on H belongs to K(H)\mathscr{K}(H); one sets

fχ(x)=∫Hf(xξ)χ(ξ)−1 dβ(ξ). f^\chi(x) = \int_H f(x\xi)\chi(\xi)^{-1}\, d\beta(\xi).

b) The function fχf^\chi is continuous, is zero outside SH, and satisfies fχ(xξ)=χ(ξ)fχ(x)f^\chi(x\xi) = \chi(\xi)f^\chi(x).

c) If g is a continuous real-valued function on X and satisfies g(xξ)=χ(ξ)g(x)g(x\xi) = \chi(\xi)g(x), then (fg)χ=f1g(fg)^\chi = f^1g (f1f^1 being given by the formula (1) with χ\chi replaced by the representation ξ↦1\xi \mapsto 1 of H in R+∗\mathbf{R}_+^*).

d) If η∈H\eta \in \mathbf{H}, then (δ(η)f)χ=χ(η)ΔH(η)−1fχ(\delta(\eta)f)^{\chi} = \chi(\eta)\Delta_{\mathbf{H}}(\eta)^{-1}f^{\chi}.

Let x0∈Xx_0 \in X and let VV be a compact neighborhood of x0x_0 in XX. The set of ξ∈H\xi \in \mathbf{H} such that VξV\xi intersects SS is also the set of ξ∈H\xi \in \mathbf{H} such that VξV\xi intersects S∩VHS \cap VH, hence is compact in H\mathbf{H} since S∩VHS \cap VH is compact and H\mathbf{H} operates properly in XX (GT, III, §4, No. 5, Th. 1); then Lemma 1 of §1, No. 1 proves a) and the continuity of fχf^{\chi}. The rest of b) is obvious. Finally, c) and d) result from the following calculations:

(fg)χ(x)=∫Hf(xξ)g(xξ)χ(ξ)−1 dβ(ξ)=∫Hf(xξ)g(x)χ(ξ)χ(ξ)−1 dβ(ξ) (fg)^{\chi}(x) = \int_{\mathbf{H}} f(x\xi)g(x\xi)\chi(\xi)^{-1}\,d\beta(\xi) = \int_{\mathbf{H}} f(x\xi)g(x)\chi(\xi)\chi(\xi)^{-1}\,d\beta(\xi)
=g(x)∫Hf(xξ) dβ(ξ)=g(x)f1(x) = g(x) \int_{\mathbf{H}} f(x\xi)\,d\beta(\xi) = g(x)f^{1}(x)
(δ(η)f)χ(x)=∫Hf(xξη)χ(ξ)−1 dβ(ξ) (\delta(\eta)f)^{\chi}(x) = \int_{\mathbf{H}} f(x\xi\eta)\chi(\xi)^{-1}\,d\beta(\xi)
=ΔH(η)−1∫Hf(xξ)χ(ξη−1)−1 dβ(ξ) = \Delta_{\mathbf{H}}(\eta)^{-1} \int_{\mathbf{H}} f(x\xi)\chi(\xi\eta^{-1})^{-1}\,d\beta(\xi)
=χ(η)ΔH(η)−1∫Hf(xξ)χ(ξ)−1 dβ(ξ). = \chi(\eta)\Delta_{\mathbf{H}}(\eta)^{-1} \int_{\mathbf{H}} f(x\xi)\chi(\xi)^{-1}\,d\beta(\xi).

Proposition 2

The mapping* f↦fχf \mapsto f^{\chi} of K(X)\mathscr{K}(X) into Kχ(X)\mathscr{K}^{\chi}(X) is linear, and the image of K(X)\mathscr{K}(X) (resp. K+(X)\mathscr{K}_{+}(X)) is Kχ(X)\mathscr{K}^{\chi}(X) *(resp. K+χ(X)\mathscr{K}_{+}^{\chi}(X)).

Linearity is immediate. It is clear that fχ⩾0f^{\chi} \geqslant 0 for f⩾0f \geqslant 0. It then suffices to apply the following lemma:

Lemma 1

— Let KK be a compact subset of XX, u a function of K+(X)\mathscr{K}_{+}(X) with u(x)>0u(x) > 0 for x∈Kx \in K. Let g∈Kχ(X)g \in \mathscr{K}^{\chi}(X) be such that Supp⁡g⊂KH\operatorname{Supp} g \subset KH. a) One has inf⁡x∈KHu1(x)>0\inf_{x \in KH} u^{1}(x) > 0. b) The function hh equal to g/u1g/u^{1} on KHKH, and to 00 on X−KHX - KH, belongs to Kχ(X)\mathscr{K}^{\chi}(X). c) g=(uh)χg = (uh)^{\chi}. One has u1(x)>0u^{1}(x) > 0 for x∈Kx \in K, therefore inf⁡x∈KHu1(x)=inf⁡x∈Ku1(x)>0\inf_{x \in KH} u^{1}(x) = \inf_{x \in K} u^{1}(x) > 0. Assertion b) follows from this at once. Finally, (uh)χ=u1h(uh)^{\chi} = u^{1}h by Prop. 1 c), and it is clear that u1h=gu^{1}h = g.

Let I be a relatively bounded linear form (Ch. II, §2, No. 2) on Kχ(X)\mathscr{K}^{\chi}(X). Then f↦I(fχ)f \mapsto I(f^{\chi}) is a relatively bounded linear form on K(X)\mathscr{K}(X), that is, a measure μI\mu_{I} on XX. The mapping I↦μII \mapsto \mu_{I} is injective by Prop. 2. The measures μI\mu_{I} on XX so obtained may be characterized as follows:

Proposition 3

Let* μ\mu be a measure on XX. *The following conditions are equivalent:

a) There exists a relatively bounded linear form I on Kχ(X)\mathscr{K}^\chi(X) such that I(fχ)=μ(f)I(f^\chi) = \mu(f) for all f∈K(X)f \in \mathscr{K}(X). b) δ(ξ)μ=χ(ξ)−1ΔH(ξ)μ\delta(\xi)\mu = \chi(\xi)^{-1} \Delta_H(\xi)\mu for all ξ∈H\xi \in H. c) For all f,gf, g in K(X)\mathscr{K}(X),

(2)

μ(f⋅g1)=μ(fχ⋅g). \mu(f \cdot g^1) = \mu(f^\chi \cdot g) .

d) If f∈K(X)f \in \mathscr{K}(X) is such that fχ=0f^\chi = 0, then μ(f)=0\mu(f) = 0.

a) ⇒\Rightarrow b): If μ(f)=I(fχ)\mu(f) = I(f^\chi) then, taking into account Prop. 1 d),

⟨δ(ξ)μ,f⟩=⟨μ,δ(ξ−1)f⟩=I((δ(ξ−1)f)χ)=I(χ(ξ)−1ΔH(ξ)fχ)=χ(ξ)−1ΔH(ξ)⟨μ,f⟩, \langle \delta(\xi)\mu, f \rangle = \langle \mu, \delta(\xi^{-1})f \rangle = I\left( (\delta(\xi^{-1})f)^\chi \right) = I(\chi(\xi)^{-1} \Delta_H(\xi)f^\chi) = \chi(\xi)^{-1} \Delta_H(\xi) \langle \mu, f \rangle ,

whence δ(ξ)μ=χ(ξ)−1ΔH(ξ)μ\delta(\xi)\mu = \chi(\xi)^{-1} \Delta_H(\xi)\mu.

b) ⇒\Rightarrow c): Suppose hypothesis b) is satisfied. Note that the functions (x,ξ)↦f(x)g(xξ)(x, \xi) \mapsto f(x)g(x\xi) and (x,ξ)↦f(xξ)g(x)(x, \xi) \mapsto f(x\xi)g(x) on X×HX \times H are continuous with compact support (because HH operates properly in XX); this established, Th. 2 of Ch. III, §4, No. 1 permits us to write:

∫Xf(x)dμ(x)∫Hg(xξ)dβ(ξ)=∫Hdβ(ξ)∫Xf(x)g(xξ)dμ(x)=∫Hdβ(ξ)∫Xf(xξ−1)g(x)χ(ξ)ΔH(ξ)−1dμ(x)=∫Xg(x)dμ(x)∫Hf(xξ−1)χ(ξ)ΔH(ξ)−1dβ(ξ)=∫Xg(x)dμ(x)∫Hf(xξ)χ(ξ)−1dβ(ξ), \int_X f(x) d\mu(x) \int_H g(x\xi) d\beta(\xi) = \int_H d\beta(\xi) \int_X f(x)g(x\xi) d\mu(x) = \int_H d\beta(\xi) \int_X f(x\xi^{-1})g(x)\chi(\xi)\Delta_H(\xi)^{-1} d\mu(x) = \int_X g(x) d\mu(x) \int_H f(x\xi^{-1})\chi(\xi)\Delta_H(\xi)^{-1} d\beta(\xi) = \int_X g(x) d\mu(x) \int_H f(x\xi)\chi(\xi)^{-1} d\beta(\xi),

which proves c).

c) ⇒\Rightarrow d): If c) is verified and if fχ=0f^\chi = 0, then μ(f⋅g1)=0\mu(f \cdot g^1) = 0 for all g∈K(X)g \in \mathscr{K}(X), thus μ(f)=0\mu(f) = 0 on choosing g∈K(X)g \in \mathscr{K}(X) such that g1=1g^1 = 1 on Supp f\mathrm{Supp}\, f (which is possible by Prop. 2 applied with χ=1\chi = 1).

d) ⇒\Rightarrow a): If condition d) is satisfied, there exists a linear form I on Kχ(X)\mathscr{K}^\chi(X) such that μ(f)=I(fχ)\mu(f) = I(f^\chi) for f∈K(X)f \in \mathscr{K}(X), and this form is relatively bounded by virtue of Prop. 2.

2. The case χ=1\chi = 1

If ff is a function on X/HX/H, then f∘πf \circ \pi is a function on XX constant on the orbits, continuous if and only if ff is continuous. The mapping f↦f∘πf \mapsto f \circ \pi defines in particular a bijection of K(X/H)\mathscr{K}(X/H) onto K1(X)\mathscr{K}^1(X).

We can then, in the case that χ=1\chi = 1, reformulate certain results of No. 1 in the following way:

Let ff be a continuous numerical function on XX whose support has compact intersection with the saturation of every compact subset of XX. The formula

fb(π(x))=∫Hf(xξ) dβ(ξ) f^b(\pi(x)) = \int_H f(x\xi)\, d\beta(\xi)

defines a continuous function fbf^b on X/HX/H. If gg is a continuous function on X/HX/H, then

(f⋅g∘π)b=fb⋅g. (f \cdot g \circ \pi)^b = f^b \cdot g .

If η∈H\eta \in H, then

(δ(η)f)b=ΔH(η)−1fb. (\delta(\eta)f)^b = \Delta_H(\eta)^{-1} f^b .

One must not forget that the definition of fbf^b depends on the choice of β\beta. If HH is compact and β\beta is normalized, the function fbf^b is sometimes called the orbital mean of ff.

If f∈K(X)f \in \mathscr{K}(X), then fb∈K(X/H)f^b \in \mathscr{K}(X/H). The mapping f↦fbf \mapsto f^b of K(X)\mathscr{K}(X) into K(X/H)\mathscr{K}(X/H) is linear, and the image of K(X)\mathscr{K}(X) (resp. K+(X)\mathscr{K}_+(X)) is K(X/H)\mathscr{K}(X/H) (resp. K+(X/H)\mathscr{K}_+(X/H)).

Remark 1

— We are going to show that the mapping f↦fbf \mapsto f^b is a strict morphism (GT, III, §2, No. 8) of K(X)\mathscr{K}(X) onto K(X/H)\mathscr{K}(X/H).

a) The mapping is continuous: it suffices to prove that, for every compact subset KK of XX, the restriction of f↦fbf \mapsto f^b to K(X,K)\mathscr{K}(X, K) is a continuous mapping of K(X,K)\mathscr{K}(X, K) into K(X/H,π(K))\mathscr{K}(X/H, \pi(K)) (TVS, II, §4, No. 4, Prop. 5); since HH operates properly in XX, the set PP of ξ∈H\xi \in H such that KξK\xi intersects KK is compact; one concludes from (3) that sup⁡x∈K∣fb(π(x))∣≤β(P)sup⁡x∈K∣f(x)∣\sup_{x \in K} |f^b(\pi(x))| \leq \beta(P) \sup_{x \in K} |f(x)|, and this proves our assertion.

b) Let K′K' be a compact subset of X/HX/H. Let us choose a compact subset KK of XX such that π(K)=K′\pi(K) = K', and let us show that the restriction of f↦fbf \mapsto f^b to K(X,K)\mathscr{K}(X, K) is a strict morphism of K(X,K)\mathscr{K}(X, K) onto K(X/H,K′)\mathscr{K}(X/H, K'). It suffices to construct a right inverse for this restriction (GT, III, §6, No. 2, Prop. 3). Now, by Lemma 1 of No. 1 (whose notations we adopt), one obtains such an inverse by composing the following mappings: α)\alpha) the mapping f′↦f′∘πf' \mapsto f' \circ \pi of K(X/H,K′)\mathscr{K}(X/H, K') into the set EE of functions of K1(X)\mathscr{K}^1(X) whose support is contained in KHKH; β)\beta) the mapping of EE into EE that, to every g∈Eg \in E, makes correspond the function equal to g/u1g/u^1 on KHKH, and to 0 on X−KHX - KH; γ)\gamma) the mapping of EE into K(X)\mathscr{K}(X) that, to every function h∈Eh \in E, makes correspond uhuh.

c) This established, if VV is a convex neighborhood of 0 in K(X)\mathscr{K}(X), then V∩K(X,K)V \cap \mathscr{K}(X, K) is a convex neighborhood of 0 in K(X,K)\mathscr{K}(X, K), therefore Vb∩K(X/H,K′)V^b \cap \mathscr{K}(X/H, K') is a convex neighborhood of 0 in K(X/H,K′)\mathscr{K}(X/H, K') by b), therefore VbV^b is a neighborhood of 0 in K(X/H)\mathscr{K}(X/H) (TVS, II, §4, No. 4). This completes the proof.

Proposition 4

a) Let λ\lambda be a measure on X/HX/H. There exists one and only one measure λ#\lambda^\# on XX such that

∫X/Hfbdλ=∫Xfdλ# \int_{X/H} f^b d\lambda = \int_X f d\lambda^\#
*for all f∈K(X)f \in \mathscr{K}(X). One has δ(ξ)λ#=ΔH(ξ)λ#\delta(\xi)\lambda^\# = \Delta_H(\xi)\lambda^\# for all ξ∈H\xi \in H.

b) Conversely, let μ\mu be a measure on XX such that δ(ξ)μ=ΔH(ξ)μ\delta(\xi)\mu = \Delta_H(\xi)\mu for all ξ∈H\xi \in H. There exists one and only one measure λ\lambda on X/HX/H such that μ=λ#\mu = \lambda^\#.

This is a special case of No. 1, Prop. 3.

Definition 1

With hypotheses and notations as in Prop. 4, λ\lambda is called the quotient of μ\mu by β\beta and is denoted μβ\frac{\mu}{\beta} or μ/β\mu/\beta.

The mapping λ↦λ#\lambda \mapsto \lambda^\# of M(X/H)\mathscr{M}(X/H) into M(X)\mathscr{M}(X) is none other than the transpose of the mapping f↦fbf \mapsto f^b of K(X)\mathscr{K}(X) into K(X/H)\mathscr{K}(X/H). Let F\mathfrak{F} be a filter on M(X/H)\mathscr{M}(X/H); to say that lim⁡λ,Fλ#(f)=0\lim_{\lambda,\mathfrak{F}} \lambda^\#(f) = 0 for all f∈K(X)f \in \mathscr{K}(X) is equivalent to saying that lim⁡λ,Fλ(f′)=0\lim_{\lambda,\mathfrak{F}} \lambda(f') = 0 for all f′∈K(X/H)f' \in \mathscr{K}(X/H); therefore the mapping λ↦λ#\lambda \mapsto \lambda^\# is, for the vague topologies, an isomorphism of M(X/H)\mathscr{M}(X/H) onto a linear subspace of M(X)\mathscr{M}(X). This subspace is vaguely closed, since it is the set of μ∈M(X)\mu \in \mathscr{M}(X) such that δ(ξ)μ=ΔH(ξ)μ\delta(\xi)\mu = \Delta_H(\xi)\mu for all ξ∈H\xi \in H. It is clear that the conditions λ⩾0\lambda \geqslant 0 and λ#⩾0\lambda^\# \geqslant 0 are equivalent.

The formula (6) may, by analogy with the usual notation for double integrals, be written

∫Xf(x)dλ#(x)=∫X/Hdλ(x˙)∫Hf(xξ)dβ(ξ)(x˙=π(x)). \int_X f(x) d\lambda^\#(x) = \int_{X/H} d\lambda(\dot{x}) \int_H f(x\xi) d\beta(\xi) \quad (\dot{x} = \pi(x)).
This involves an abuse of notation, the integral ∫Hf(xξ)dβ(ξ)\int_H f(x\xi) d\beta(\xi) being regarded as a function of x˙\dot{x} and not of xx; this manner of writing will be used frequently in what follows provided no confusion can arise.

Remark 2

— Let E be a locally convex vector space and let m be a vectorial measure on X/HX/H with values in E. The mapping f↦m(fb)f \mapsto m(f^b) of K(X)\mathscr{K}(X) into E is then a vectorial measure on X, with values in E, which we shall again denote by m#m^\#. The mapping m↦m#m \mapsto m^\# is again an isomorphism of L(K(X/H);E)\mathscr{L}(\mathscr{K}(X/H); E) onto a linear subspace A of L(K(X);E)\mathscr{L}(\mathscr{K}(X); E) (when these spaces are equipped with the topology of pointwise convergence). Moreover, since the mapping f↦fbf \mapsto f^b is a surjective strict morphism, the subspace A consists precisely of the vectorial measures n on X that are zero on the kernel N of the mapping f↦fbf \mapsto f^b. In order that n∈An \in A, it is therefore necessary and sufficient that the scalar measures z′∘nz' \circ n be zero on N for every z′∈E′z' \in E'. One then deduces from Prop. 3 that n∈An \in A if and only if δ(ξ)n=ΔH(ξ)n\delta(\xi)n = \Delta_H(\xi)n for all ξ∈H\xi \in H.

3. Another interpretation of λ#\lambda^\#

For every x∈Xx \in X, the mapping ξ↦xξ\xi \mapsto x\xi of H into X is proper (GT, III, §4, No. 2, Prop. 4), therefore β\beta admits an image measure on X under this mapping, which image is concentrated on the orbit xHxH (Ch. V, §6, No. 2, Cor. 3 of Prop. 2); since β\beta is left-invariant, this image measure depends only on the class u=π(x)u = \pi(x) of xx in X/HX/H, and will be denoted βu\beta_u. By definition, for f∈K(X)f \in \mathscr{K}(X),

∫Xf(y) dβu(y)=∫Hf(xξ) dβ(ξ)=f♭(u). \int_X f(y)\, d\beta_u(y) = \int_H f(x\xi)\, d\beta(\xi) = f^\flat(u).

We thus see that

(εu)#=βu. (\varepsilon_u)^\# = \beta_u.

Lemma 2

Let f be a function on X, with values in a topological space.

a) If f is a numerical function ≥0\geq 0 then, for x∈Xx \in X,

∫X∗f(y) dβx˙(y)=∫H∗f(xξ) dβ(ξ)(x˙=π(x)). \int_X^* f(y)\, d\beta_{\dot{x}}(y) = \int_H^* f(x\xi)\, d\beta(\xi) \quad (\dot{x} = \pi(x)).

b) For f to be βx˙\beta_{\dot{x}}-measurable, it is necessary and sufficient that the function ξ↦f(xξ)\xi \mapsto f(x\xi) on H be β\beta-measurable.

c) Suppose that f is a function on X, with values in a Banach space or in R‾\overline{\mathbf{R}}; then, for f to be βx˙\beta_{\dot{x}}-integrable (resp. essentially βx˙\beta_{\dot{x}}-integrable), it is necessary and sufficient that the function ξ↦f(xξ)\xi \mapsto f(x\xi) on H be β\beta-integrable (resp. essentially β\beta-integrable), in which case ∫Xf(y) dβx˙(y)=∫Hf(xξ) dβ(ξ)\int_X f(y)\, d\beta_{\dot{x}}(y) = \int_H f(x\xi)\, d\beta(\xi).

This follows from Ch. V, §4, Prop. 2, Prop. 3 and Th. 2.

Since f♭∈K(X/H)f^\flat \in \mathscr{K}(X/H) for f∈K(X)f \in \mathscr{K}(X), formula (8) proves that the mapping u↦βuu \mapsto \beta_u of X/HX/H into M(X)\mathscr{M}(X) is vaguely continuous, that the family (βu)(\beta_u) is λ\lambda-adequate1^1 for any positive measure λ\lambda on X/HX/H, and that

λ#=∫X/Hβu dλ(u), \lambda^\# = \int_{X/H} \beta_u\, d\lambda(u),

which furnishes a new interpretation of λ#\lambda^\#.

1^1 In the sense of the first edition of Ch. V, hence a fortiori in the sense of the second (cf. the footnote to the Example of Ch. VI, §1, No. 1).

Proposition 5

— Let λ\lambda be a positive measure on X/HX/H.

a) Let ff be a λ#\lambda^\#-measurable function on XX, with values in a topological space, constant outside a countable union of λ#\lambda^\#-integrable sets. Then, the set of x˙∈X/H\dot{x} \in X/H such that the function ξ↦f(xξ)\xi \mapsto f(x\xi) is not β\beta-measurable is locally λ\lambda-negligible.

b) Let ff be a λ#\lambda^\#-measurable function ≥0\geq 0 on XX, zero outside a countable union of λ#\lambda^\#-integrable sets. Then, the function x˙↦∫∗f(xξ)dβ(ξ)\dot{x} \mapsto \int^* f(x\xi) d\beta(\xi) on X/HX/H is λ\lambda-measurable, and

∫X∗f(x)dλ#(x)=∫X/H∗dλ(x˙)∫H∗f(xξ)dβ(ξ)(x˙=π(x)). \int_X^* f(x) d\lambda^\#(x) = \int_{X/H}^* d\lambda(\dot{x}) \int_H^* f(x\xi) d\beta(\xi) \quad (\dot{x} = \pi(x)).

c) Let ff be a λ#\lambda^\#-integrable function on XX, with values in a Banach space or in R‾\overline{\mathbf{R}}. Then, the set of x˙∈X/H\dot{x} \in X/H such that ξ↦f(xξ)\xi \mapsto f(x\xi) is not β\beta-integrable is λ\lambda-negligible; the function fbf^b on X/HX/H defined almost everywhere by the formula

fb(x˙)=∫Hf(xξ)dβ(ξ)(x˙=π(x)) f^b(\dot{x}) = \int_H f(x\xi) d\beta(\xi) \quad (\dot{x} = \pi(x))

is λ\lambda-integrable, and

∫X/Hfbdλ=∫Xfdλ# \int_{X/H} f^b d\lambda = \int_X f d\lambda^\#

and

∫X/H∣fb∣dλ≤∫X∣f∣dλ#. \int_{X/H} |f^b| d\lambda \leq \int_X |f| d\lambda^\#.

d) Let ff be a λ#\lambda^\#-measurable function on XX, with values in a Banach space or in R‾\overline{\mathbf{R}}, and zero outside a countable union of λ#\lambda^\#-integrable sets. Then, for ff to be λ#\lambda^\#-integrable, it is necessary and sufficient that

∫X/H∗dλ(x˙)∫H∗∣f(xξ)∣dβ(ξ)<+∞(x˙=π(x)). \int_{X/H}^* d\lambda(\dot{x}) \int_H^* |f(x\xi)| d\beta(\xi) < +\infty \quad (\dot{x} = \pi(x)).

Taking into account Lemma 2, the assertions a), b) and c) follow from Ch. V, §3, Prop. 4, Prop. 5 and Th. 1 (with the exception of (13), which follows from (12) because it is clear that ∣fb∣≤∣f∣b|f^b| \leq |f|^b); d) follows from b).

Proposition 6

— Let λ\lambda be a positive measure on X/HX/H.

a) Let NN be a subset of X/HX/H. For NN to be locally λ\lambda-negligible, it is necessary and sufficient that π−1(N)\pi^{-1}(N) be locally λ#\lambda^\#-negligible.

b) Let g be a function on X/H, with values in a topological space. For g to be λ\lambda-measurable, it is necessary and sufficient that g∘πg \circ \pi be λ#\lambda^\#-measurable.

c) Let h be a function on X/H, with values in a Banach space or in R‾\overline{\mathbf{R}}. For h to be locally λ\lambda-integrable, it is necessary and sufficient that h∘πh \circ \pi be locally λ#\lambda^\#-integrable, in which case (h⋅λ)#=(h∘π)⋅λ#(h \cdot \lambda)^\# = (h \circ \pi) \cdot \lambda^\#.

Suppose h∘πh \circ \pi is locally λ#\lambda^\#-integrable. For every f∈K(X)f \in \mathscr{K}(X), f⋅(h∘π)f \cdot (h \circ \pi) is λ#\lambda^\#-integrable, therefore (Prop. 5) the function (f⋅(h∘π))b=fb⋅h(f \cdot (h \circ \pi))^b = f^b \cdot h is λ\lambda-integrable and

∫X/Hfb⋅h dλ=∫Xf⋅(h∘π) dλ#. \int_{X/H} f^b \cdot h \, d\lambda = \int_X f \cdot (h \circ \pi) \, d\lambda^\#.

Since f↦fbf \mapsto f^b is a surjective mapping of K(X)\mathscr{K}(X) onto K(X/H)\mathscr{K}(X/H), this shows that h is locally λ\lambda-integrable and that

(h⋅λ)#=(h∘π)⋅λ#. (h \cdot \lambda)^\# = (h \circ \pi) \cdot \lambda^\#.

In particular, if π−1(N)\pi^{-1}(N) is locally λ#\lambda^\#-negligible, then φN∘π\varphi_N \circ \pi is locally λ#\lambda^\#-negligible, therefore (φN⋅λ)#=(φN∘π)⋅λ#=0(\varphi_N \cdot \lambda)^\# = (\varphi_N \circ \pi) \cdot \lambda^\# = 0, consequently φN⋅λ=0\varphi_N \cdot \lambda = 0 and N is locally λ\lambda-negligible. Now suppose that g∘πg \circ \pi is λ#\lambda^\#-measurable. Let K' be a compact subset of X/H. Let f∈K+(X)f \in \mathscr{K}_+(X) be such that fb=1f^b = 1 on K' (No. 1, Prop. 2), and let K=Supp⁡fK = \operatorname{Supp} f; one has π(K)⊃K′\pi(K) \supset K'. There exists a partition of K formed of a λ#\lambda^\#-negligible set M and a sequence (Kn)(K_n) of compact sets such that (g∘π)∣Kn(g \circ \pi)|K_n is continuous for every n. Then g∣π(Kn)g|_{\pi(K_n)} is continuous. Let P be the set of points of K' not belonging to π(K1)∪π(K2)∪…\pi(K_1) \cup \pi(K_2) \cup \ldots; then π−1(P)∩K\pi^{-1}(P) \cap K is contained in M, hence is λ#\lambda^\#-negligible; therefore f⋅φπ−1(P)f \cdot \varphi_{\pi^{-1}(P)} is λ#\lambda^\#-negligible; it follows (Prop. 5) that

0=∫Xf⋅φπ−1(P) dλ#=∫X/Hfb⋅φP dλ≥∫X/H∗φP dλ, 0 = \int_X f \cdot \varphi_{\pi^{-1}(P)} \, d\lambda^\# = \int_{X/H} f^b \cdot \varphi_P \, d\lambda \geq \int_{X/H}^* \varphi_P \, d\lambda,

therefore P is λ\lambda-negligible, and g is λ\lambda-measurable.

If N is locally λ\lambda-negligible, then π−1(N)\pi^{-1}(N) is locally λ#\lambda^\#-negligible (Ch. V, §6, No. 6, Cor. 1 of Prop. 10). If g is λ\lambda-measurable, then g∘πg \circ \pi is λ#\lambda^\#-measurable (ibid.). Finally, suppose h is locally λ\lambda-integrable. Then we already know that h∘πh \circ \pi is λ#\lambda^\#-measurable. For every f∈K+(X)f \in \mathscr{K}_+(X) we have, by Prop. 5,

∫X∗f(x)∣h∣(π(x)) dλ#(x)=∫X/H∗∣h∣(u)fb(u) dλ(u)<+∞, \int_X^* f(x)|h|(\pi(x)) \, d\lambda^\#(x) = \int_{X/H}^* |h|(u)f^b(u) \, d\lambda(u) < +\infty,

therefore h∘πh \circ \pi is locally λ#\lambda^\#-integrable.

Corollary 1

— Let λ,λ′\lambda, \lambda' be two positive measures on X/HX/H. For λ′\lambda' to have base λ\lambda, it is necessary and sufficient that λ′#{\lambda'}^\# have base λ#\lambda^\#. For λ\lambda and λ′\lambda' to be equivalent, it is necessary and sufficient that λ#\lambda^\# and λ′#{\lambda'}^\# be equivalent.

The first assertion follows from Prop. 6, a) and c). The second follows from the first.

Corollary 2

— Let λ\lambda be a positive measure on X/HX/H, and ff a λ#\lambda^\#-measurable numerical function on XX. Suppose that, for every ξ∈H\xi \in H, δ(ξ)f=f\delta(\xi)f = f locally λ#\lambda^\#-almost everywhere. Then, there exists a λ\lambda-measurable function gg on X/HX/H such that f=g∘πf = g \circ \pi locally λ#\lambda^\#-almost everywhere.

Replacing ff by f/(1+∣f∣)f/(1+|f|), one reduces to the case that ff is bounded, hence locally λ#\lambda^\#-integrable. Let μ=f⋅λ#\mu = f \cdot \lambda^\#. The hypothesis on ff implies that δ(ξ)μ=f⋅δ(ξ)λ#=ΔH(ξ)μ\delta(\xi)\mu = f \cdot \delta(\xi)\lambda^\# = \Delta_H(\xi)\mu for all ξ∈H\xi \in H. There then exists (Prop. 4) a measure λ′\lambda' on X/HX/H such that μ=λ′#\mu = {\lambda'}^\#. By Cor. 1, there exists a locally λ\lambda-integrable function gg on X/HX/H such that λ′=g⋅λ\lambda' = g \cdot \lambda. By Prop. 6, f⋅λ#=λ′#=(g∘π)⋅λ#f \cdot \lambda^\# = {\lambda'}^\# = (g \circ \pi) \cdot \lambda^\#, whence f=g∘πf = g \circ \pi locally λ#\lambda^\#-almost everywhere.

Corollary 3

— a) Let (λι)ι∈I(\lambda_\iota)_{\iota \in I} be a family of real measures on X/HX/H. For the family (λι)(\lambda_\iota) to be bounded above in M(X/H)\mathscr{M}(X/H), it is necessary and sufficient that the family (λι#)(\lambda_\iota^\#) be bounded above in M(X)\mathscr{M}(X), in which case

sup⁡(λι#)=(sup⁡λι)#. \sup(\lambda_\iota^\#) = (\sup \lambda_\iota)^\#.

b) Let λ\lambda be a real measure on X/HX/H. Then (λ+)#=(λ#)+(\lambda^+)^\# = (\lambda^\#)^+ and (λ−)#=(λ#)−(\lambda^-)^\# = (\lambda^\#)^-.

c) Let λ\lambda be a complex measure on X/HX/H. Then ∣λ∣#=∣λ#∣|\lambda|^\# = |\lambda^\#|.

Assume the family (λι)(\lambda_\iota) to be bounded above and let μ=sup⁡λι\mu = \sup \lambda_\iota. Since λ≥0\lambda \geq 0 implies λ#≥0\lambda^\# \geq 0, we have μ#≥λι#\mu^\# \geq \lambda_\iota^\# for all ι\iota, which shows that the family (λι#)(\lambda_\iota^\#) is bounded above and that

(sup⁡λι)#≥sup⁡(λι#). (\sup \lambda_\iota)^\# \geq \sup(\lambda_\iota^\#).

Conversely, assume the family (λι#)(\lambda_\iota^\#) to be bounded above and let ν=sup⁡(λι#)\nu = \sup(\lambda_\iota^\#). Since δ(ξ)λι#=ΔH(ξ)λι#\delta(\xi)\lambda_\iota^\# = \Delta_H(\xi)\lambda_\iota^\# for all ξ∈H\xi \in H, obviously δ(ξ)ν=ΔH(ξ)ν\delta(\xi)\nu = \Delta_H(\xi)\nu, therefore there exists a measure μ′∈M(X/H)\mu' \in \mathscr{M}(X/H) such that ν=μ′#\nu = {\mu'}^\#. Since λ#≥0\lambda^\# \geq 0 implies λ≥0\lambda \geq 0, we have μ′≥λι\mu' \geq \lambda_\iota for all ι\iota, which shows that the family (λι)(\lambda_\iota) is bounded above and that ν=μ′#≥(sup⁡λι)#\nu = {\mu'}^\# \geq (\sup \lambda_\iota)^\#, whence

sup⁡(λι#)≥(sup⁡λι)#, \sup(\lambda_\iota^\#) \geq (\sup \lambda_\iota)^\#,

which completes the proof of a). The assertion b) then follows at once since, for example, λ+\lambda^+ is none other than sup⁡(λ,0)\sup(\lambda, 0). To prove c), it suffices to note that ∣λ∣=sup⁡R(αλ)|\lambda| = \sup \mathscr{R}(\alpha \lambda) over the complex numbers α\alpha of absolute value 1, and on the other hand that R(μ#)=(Rμ)#\mathscr{R}(\mu^\#) = (\mathscr{R}\mu)^\# for every μ∈M(X/H)\mu \in \mathscr{M}(X/H).

Remark 1

Prop. 6 a) may be expressed by saying that λ\lambda is a pseudo-image measure of λ#\lambda^\# under π\pi (Ch. VI, §3, No. 2, Def. 1).

Remark 2

Suppose HH is compact and β\beta is normalized. The saturation of every compact subset of XX is compact. Therefore, if f∈K(X/H)f \in \mathscr{K}(X/H) then f∘π∈K(X)f \circ \pi \in \mathscr{K}(X); and, for every positive measure λ\lambda on X/HX/H, Prop. 5 c) gives

∫X(f∘π)(x) dλ#(x)=∫X/Hf(u) dλ(u). \int_X (f \circ \pi)(x)\, d\lambda^\#(x) = \int_{X/H} f(u)\, d\lambda(u).

In other words, λ\lambda is the image of λ#\lambda^\# under π\pi.

Remark 3

Cor. 3 c) of Prop. 6 shows at once that the results of this subsection remain valid in the case of complex measures (except for those that involve the upper integral).

Remark 4

Let mm be a vectorial measure on X/HX/H with values in EE, and let qq be a lower semi-continuous semi-norm on EE. For mm to be qq-majorizable (Ch. VI, §2, No. 3, Def. 3), it is necessary and sufficient that m#m^\# be so, in which case q(m#)=q(m)#q(m^\#) = q(m)^\#. This follows at once from the definitions and Cor. 3 a).

On the other hand, let μ\mu be a positive measure on X/HX/H. For mm to be scalarly of base μ\mu, it is necessary and sufficient that m#m^\# be scalarly of base μ#\mu^\#: this follows from Cor. 1.

Finally, if mm has base μ\mu, with density ff with respect to μ\mu (Ch. VI, §2, No. 4, Def. 4), then m#m^\# has base μ#\mu^\#, with density f∘πf \circ \pi: this follows from Prop. 6 c).

4. The case that X/HX/H is paracompact

If X/HX/H is paracompact, we shall first see that the vector spaces Kχ(X)\mathscr{K}^\chi(X), for variable χ\chi, are all isomorphic to each other, and in particular isomorphic to K1(X)\mathscr{K}^1(X).

Proposition 7

Assume X/HX/H paracompact. Let χ\chi be a continuous representation of HH in R+∗\mathbf{R}_+^*.

a) There exists on XX a continuous function rr, with values >0> 0, such that r(xξ)=χ(ξ)r(x)r(x\xi) = \chi(\xi)r(x) for all x∈Xx \in X and ξ∈H\xi \in H.

b) The mapping g↦g/rg \mapsto g/r is an isomorphism of the vector space Kχ(X)\mathscr{K}^\chi(X) onto the vector space K1(X)\mathscr{K}^1(X).

Let us apply Prop. 1 of No. 1 on taking ff to be a function ⩾0\geqslant 0 that is not identically zero on any orbit (this is possible by Lemma 1 of Appendix 1); then r=fχr = f^\chi satisfies the properties of a). The assertion b) is obvious.

Proposition 8

Assume X/HX/H paracompact. There exists a continuous function h⩾0h \geqslant 0 on XX, whose support has compact intersection with* the saturation of every compact subset of X, and is such that hb=1h^b = 1. For such a function, one has g=(h⋅(g∘π))bg = (h \cdot (g \circ \pi))^b for every continuous function gg on X/HX/H.

Let us apply Prop. 1 of No. 1, with χ=1\chi = 1, on taking for ff a function ⩾0\geqslant 0 that is not identically zero on any orbit. We have f1(x)>0f^1(x) > 0 at every point xx of XX. Set h=f/f1h = f/f^1. Then h1=f1/f1=1h^1 = f^1/f^1 = 1, therefore hb=1h^b = 1. It follows that if gg is a continuous function on X/HX/H, then (h⋅(g∘π))b=hb⋅g=g(h \cdot (g \circ \pi))^b = h^b \cdot g = g.

Remark 1

In particular, let XX be a locally compact space on which a discrete group DD operates continuously and properly on the right; suppose X/DX/D is paracompact. Then, there exists a continuous function h⩾0h \geqslant 0 on XX whose support has compact intersection with the saturation of every compact subset of XX, and is such that ∑d∈Dh(xd)=1\sum_{d \in D} h(xd) = 1 for every x∈Xx \in X (all terms of the sum being zero except for a finite number of them).

Remark 2

Let us conserve the hypotheses and notations of Prop. 8. The mapping g↦h⋅(g∘π)g \mapsto h \cdot (g \circ \pi) is a continuous mapping of K(X/H)\mathscr{K}(X/H) into K(X)\mathscr{K}(X) that is a right inverse of the mapping f↦fbf \mapsto f^b. Consequently, every bounded (resp. compact) subset of K(X/H)\mathscr{K}(X/H) is the image of a bounded (resp. compact) subset of K(X)\mathscr{K}(X). From this, one deduces immediately that the mapping λ↦λ#\lambda \mapsto \lambda^\# is again an isomorphism of M(X/H)\mathscr{M}(X/H) onto a closed linear subspace of M(X)\mathscr{M}(X) when these spaces are equipped with the topology of bounded (resp. compact) convergence.

Proposition 9

— We conserve the hypotheses and notations of Prop. 8. Let λ\lambda be a positive measure on X/HX/H.

a) The pair (π,h)(\pi, h) is λ#\lambda^\#-adapted, and ∫Xh(x)επ(x)dλ#(x)=λ\int_X h(x) \varepsilon_{\pi(x)} d\lambda^\#(x) = \lambda.

b) The mapping π\pi is proper for the measure h⋅λ#h \cdot \lambda^\#, and π(h⋅λ#)=λ\pi(h \cdot \lambda^\#) = \lambda.

c) Let kk be a function on X/HX/H, with values in a Banach space or in R‾\overline{\mathbf{R}}. For kk to be measurable (resp. locally integrable, essentially integrable, integrable) for λ\lambda, it is necessary and sufficient that h⋅(k∘π)h \cdot (k \circ \pi) be so for λ#\lambda^\#; and, if kk is essentially integrable for λ\lambda, then

∫X/Hk dλ=∫Xh⋅(k∘π) dλ#. \int_{X/H} k \, d\lambda = \int_X h \cdot (k \circ \pi) \, d\lambda^\#.

Let f∈K(X/H)f \in \mathscr{K}(X/H). Then h⋅(f∘π)∈K(X)h \cdot (f \circ \pi) \in \mathscr{K}(X) and

∫Xh(x)f(π(x))dλ#(x)=∫X/Hf(x˙)dλ(x˙)∫Hh(xξ)dβ(ξ)=∫X/Hf(x˙)dλ(x˙), \int_X h(x) f(\pi(x)) d\lambda^\#(x) = \int_{X/H} f(\dot{x}) d\lambda(\dot{x}) \int_H h(x\xi) d\beta(\xi) = \int_{X/H} f(\dot{x}) d\lambda(\dot{x}),

whence a). The assertion b) is proved similarly. The assertions of c) concerning measurability, essential integrability and formula (14) may then be obtained by applying the results of Ch. V (§ 4, Prop. 3, § 5, Th. 1, § 4, Th. 2). If kk is λ\lambda-integrable, then h⋅(k∘π)h \cdot (k \circ \pi) is λ#\lambda^\#-integrable (Ch. V, § 3, No. 3, Th. 1). If h⋅(k∘π)h \cdot (k \circ \pi) is λ#\lambda^\#-integrable, Prop. 5 proves that (h⋅(k∘π))b=hb⋅k=k(h \cdot (k \circ \pi))^b = h^b \cdot k = k is λ\lambda-integrable. If kk is locally λ\lambda-integrable, then h⋅(k∘π)h \cdot (k \circ \pi) is locally λ#\lambda^\#-integrable (Prop. 6). Finally, suppose h⋅(k∘π)h \cdot (k \circ \pi) locally λ#\lambda^\#-integrable; for every f∈H(X/H)f \in \mathscr{H}(X/H), h⋅(k∘π)⋅(f∘π)h \cdot (k \circ \pi) \cdot (f \circ \pi) has compact support, and

∣h⋅(k∘π)⋅(f∘π)∣≤M∣h⋅(k∘π)∣, |h \cdot (k \circ \pi) \cdot (f \circ \pi)| \leq M |h \cdot (k \circ \pi)|,
where M=sup⁡∣f∣M = \sup |f|; therefore h⋅((kf)∘π)h \cdot ((kf) \circ \pi) is λ#\lambda^\#-integrable, consequently kfkf is λ\lambda-integrable, by what has already been proved; this proves that kk is indeed locally λ\lambda-integrable.

Corollary

— The continuous linear mapping f↦fbf \mapsto f^b of L1(X,λ#)L^1(X, \lambda^\#) into L1(X/H,λ)L^1(X/H, \lambda) defined by Prop. 5 is surjective.

Suppose first that X/HX/H is paracompact and let hh be a function on XX satisfying the conditions of Prop. 8. If kk is a λ\lambda-integrable numerical function on X/HX/H, then h⋅(k∘π)h \cdot (k \circ \pi) is λ#\lambda^\#-integrable and (h⋅(k∘π))b=k(h \cdot (k \circ \pi))^b = k (Prop. 9).

In the general case, let u∈L1(X/H,λ)u \in L^1(X/H, \lambda). There exists a function f∈L1(X/H,λ)f \in \mathscr{L}^1(X/H, \lambda) with class uu and zero outside a countable union of compact sets KnK_n. Let us define recursively a sequence of relatively compact open sets UnU_n of X/HX/H such that Un+1⊃Kn∪U‾nU_{n+1} \supset K_n \cup \overline{U}_n, and let VV be the union of the UnU_n. Then VV is an open subset of X/HX/H, a countable union of compact subsets U‾n\overline{U}_n, hence is paracompact (GT, I, § 9, No. 10, Th. 5). Set Y=π‾−1(V)Y = \overline{\pi}^{-1}(V) and let λV\lambda_V (resp. λY#\lambda_Y^\#) be the measure induced by λ\lambda (resp. λ#\lambda^\#) on VV (resp. YY). It is clear that Y/HY/H may be identified with VV (GT, I, § 3, Prop. 10) and that λY#\lambda_Y^\# may be identified with (λV)#(\lambda_V)^\#. Moreover, ff is zero outside VV and belongs to L1(V,λV)\mathscr{L}^1(V, \lambda_V). Therefore, there exists g∈L1(Y,λY#)g \in \mathscr{L}^1(Y, \lambda_Y^\#) such that gb=fg^b = f almost everywhere in VV. Extending gg by 0 on X−YX - Y, one obtains a function g1∈L1(X,λ#)g_1 \in \mathscr{L}^1(X, \lambda^\#), and it is clear that the class of g1bg_1^b in L1(X/H,λ)L^1(X/H, \lambda) is none other than uu.

Remark 3

— Suppose X/HX/H paracompact, and let us retain the notations of Proposition 9. The mapping k↦h⋅(k∘π)k \mapsto h \cdot (k \circ \pi) of L1(X/H,λ)L^1(X/H, \lambda) into L1(X,λ#)L^1(X, \lambda^\#) is then isometric by (14) and is a right inverse of the mapping f↦fbf \mapsto f^b of L1(X,λ#)L^1(X, \lambda^\#) onto L1(X/H,λ)L^1(X/H, \lambda).

5. Quasi-invariant measures on a homogeneous space

Lemma 3

— Let G be a locally compact group, μ\mu a left Haar measure on G, ν\nu and ν′\nu' two nonzero quasi-invariant measures on G. If, for every s∈Gs \in G, the densities of γ(s)ν\gamma(s)\nu with respect to ν\nu and of γ(s)ν′\gamma(s)\nu' with respect to ν′\nu' are equal locally μ\mu-almost everywhere, then ν\nu and ν′\nu' are proportional.

Write ν=ρ⋅μ\nu = \rho \cdot \mu, ν′=ρ′⋅μ\nu' = \rho' \cdot \mu, where ρ,ρ′\rho, \rho' are locally μ\mu-integrable functions on GG and are everywhere nonzero (§1§ 1, No. 9, Prop. 11). For every s∈Gs \in G,

γ(s)ν=(γ(s)ρ)⋅μ,γ(s)ν′=(γ(s)ρ′)⋅μ, \gamma(s)\nu = (\gamma(s)\rho) \cdot \mu, \qquad \gamma(s)\nu' = (\gamma(s)\rho') \cdot \mu,

and the hypothesis implies that ρ−1⋅γ(s)ρ=ρ′−1⋅γ(s)ρ′\rho^{-1} \cdot \gamma(s)\rho = {\rho'}^{-1} \cdot \gamma(s)\rho' locally μ\mu-almost everywhere. Set σ=ρ′/ρ\sigma = \rho'/\rho, which is a μ\mu-measurable function on GG. For every s∈Gs \in G, γ(s)σ=σ\gamma(s)\sigma = \sigma locally μ\mu-almost everywhere. Therefore σ\sigma is equal to a constant locally μ\mu-almost everywhere, by Cor. 2 of Prop. 6 applied with X=H=GX = H = G.

Let GG be a locally compact group, HH a closed subgroup of GG. Consider the homogeneous space G/HG/H of left cosets with respect to HH, on which GG operates continuously on the left. We are going to show that there exists one and only one class of nonzero quasi-invariant measures on G/HG/H.

Note that HH operates on GG continuously and properly by right translations; and the quotient space, which is none other than G/HG/H, is paracompact (GT, III, §4§ 4, No. 6, Prop. 13). We may therefore apply the results of Nos. 1 to 4, with X=GX = G. We thus have mappings f↦fbf \mapsto f^b of K(G)\mathscr{K}(G) onto K(G/H)\mathscr{K}(G/H), and λ↦λ#\lambda \mapsto \lambda^\# of M(G/H)\mathscr{M}(G/H) into M(G)\mathscr{M}(G) (once a left Haar measure β\beta on HH has been fixed). The fact that GG operates on the left in G/HG/H gives rise to a supplementary property:

(15)γG/H(s)⋅fb=(γG(s)⋅f)b(s∈G, f∈K(G))(16)(γG/H(s)⋅λ)#=γG(s)⋅λ#(s∈G, λ∈M(G/H)). \begin{align*} (15) \qquad & \gamma_{G/H}(s) \cdot f^b = (\gamma_G(s) \cdot f)^b \qquad (s \in G,\ f \in \mathscr{K}(G)) \\ (16) \qquad & (\gamma_{G/H}(s) \cdot \lambda)^\# = \gamma_G(s) \cdot \lambda^\# \qquad (s \in G,\ \lambda \in \mathscr{M}(G/H)). \end{align*}

Indeed, for every x∈Gx \in G,

(γG/H(s)⋅fb)(π(x))=fb(s−1π(x))=fb(π(s−1x))=∫Hf(s−1xξ) dβ(ξ)=∫H(γG(s)f)(xξ) dβ(ξ)=(γG(s)f)b(π(x)), (\gamma_{G/H}(s) \cdot f^b)(\pi(x)) = f^b(s^{-1}\pi(x)) = f^b(\pi(s^{-1}x)) = \int_H f(s^{-1}x\xi)\, d\beta(\xi) = \int_H (\gamma_G(s)f)(x\xi)\, d\beta(\xi) = (\gamma_G(s)f)^b(\pi(x)),

whence formula (15), which implies formula (16).

Lemma 4

Let λ\lambda be a measure ≠0\neq 0 on G/HG/H, and μ\mu a left Haar measure on GG. The following properties are equivalent: a) λ\lambda is quasi-invariant under GG; b) for a subset AA of G/HG/H to be locally λ\lambda-negligible, it is necessary and sufficient that π‾−1(A)\overline{\pi}^{-1}(A) be locally μ\mu-negligible; c) the measure λ#\lambda^\# is equivalent to μ\mu.

Suppose this to be the case and let λ#=ρ⋅μ\lambda^\# = \rho \cdot \mu, where ρ\rho is a locally μ\mu-integrable function that is everywhere nonzero. Then, for every s∈Gs \in G, the density θs\theta_s of γG/H(s)λ\gamma_{G/H}(s)\lambda with respect to λ\lambda is such that

θs(π(x))=ρ(s−1x)ρ(x) \theta_s(\pi(x)) = \frac{\rho(s^{-1}x)}{\rho(x)}

locally μ\mu-almost everywhere on GG.

c) ⇒\Rightarrow b): This follows at once from Prop. 6 a).

b) ⇒\Rightarrow a): If property b) holds, then the set of locally λ\lambda-negligible subsets of G/HG/H is invariant under GG, thus λ\lambda is quasi-invariant under GG.

a) ⇒\Rightarrow c): Assume λ\lambda is quasi-invariant under GG; for every s∈Gs \in G, λ\lambda and γG/H(s)λ\gamma_{G/H}(s)\lambda are equivalent, therefore λ#\lambda^\# and γG(s)⋅λ#=(γG/H(s)⋅λ)#\gamma_G(s) \cdot \lambda^\# = (\gamma_{G/H}(s) \cdot \lambda)^\# are equivalent (Cor. 1 of Prop. 6); since λ#≠0\lambda^\# \neq 0, λ#\lambda^\# is equivalent to μ\mu (§ 1, No. 9, Prop. 11).

Moreover, for every s∈Gs \in G,

(θs∘π)⋅λ#=(θs⋅λ)#=(γG/H(s)λ)#=γG(s)λ# (\theta_s \circ \pi) \cdot \lambda^\# = (\theta_s \cdot \lambda)^\# = (\gamma_{G/H}(s)\lambda)^\# = \gamma_G(s)\lambda^\#
=(γG(s)ρ)⋅μ=γG(s)ρρ⋅λ#, = (\gamma_G(s)\rho) \cdot \mu = \frac{\gamma_G(s)\rho}{\rho} \cdot \lambda^\#,

whence (17).

The equivalence of a) and b) implies first the uniqueness result already announced, and even a more precise result:

Theorem 1

Let GG be a locally compact group, HH a closed subgroup of GG.

a) Any two nonzero quasi-invariant measures on G/HG/H are equivalent; the subsets of G/HG/H locally negligible for these measures are those whose inverse image in GG is locally negligible for a Haar measure.

b) Let λ,λ′\lambda, \lambda' be two nonzero quasi-invariant measures on G/HG/H. If, for every s∈Gs \in G, the densities of γG/H(s)λ\gamma_{G/H}(s)\lambda with respect to λ\lambda and of γG/H(s)λ′\gamma_{G/H}(s)\lambda' with respect to λ′\lambda' are equal almost everywhere for λ\lambda (or λ′\lambda'), then λ\lambda and λ′\lambda' are proportional.

The assertion a) follows at once from Lemma 4. Let λ\lambda and λ′\lambda' be two nonzero quasi-invariant measures satisfying the condition of b). Then, for every s∈Gs \in G, the densities of γG(s)λ#\gamma_G(s)\lambda^\# with respect to λ#\lambda^\# and of γG(s)λ′#\gamma_G(s){\lambda'}^\# with respect to λ′#{\lambda'}^\# are equal locally μ\mu-almost everywhere, therefore (Lemma 3) λ#\lambda^\# and λ′#{\lambda'}^\# are proportional, hence λ\lambda and λ′\lambda' are proportional.

On the other hand, Lemma 4 reduces the search for nonzero quasi-invariant measures on G/HG/H to that for the measures on GG equivalent to

Haar measure and of the form λ#\lambda^\#. On this subject we have the following lemma:

Lemma 5

Let μ\mu be a left Haar measure on GG, and ρ\rho a locally μ\mu-integrable function. For ρ⋅μ\rho \cdot \mu to be of the form λ#\lambda^\#, it is necessary and sufficient that, for every ξ∈H\xi \in H,

ρ(xξ)=ΔH(ξ)ΔG(ξ)ρ(x) \rho(x\xi) = \frac{\Delta_H(\xi)}{\Delta_G(\xi)} \rho(x)

locally μ\mu-almost everywhere on GG.

To say that ρ⋅μ\rho \cdot \mu is of the form λ#\lambda^\# amounts to saying that, for every ξ∈H\xi \in H, δ(ξ)(ρ⋅μ)=ΔH(ξ)ρ⋅μ\delta(\xi)(\rho \cdot \mu) = \Delta_H(\xi)\rho \cdot \mu (Prop. 4). Now,

δ(ξ)(ρ⋅μ)=(δ(ξ)ρ)⋅(δ(ξ)μ)=ΔG(ξ)(δ(ξ)ρ)⋅μ, \delta(\xi)(\rho \cdot \mu) = (\delta(\xi)\rho) \cdot (\delta(\xi)\mu) = \Delta_G(\xi)(\delta(\xi)\rho) \cdot \mu,

whence the lemma.

We can now establish the announced existence result, and even a more precise result:

Theorem 2

Let GG be a locally compact group, HH a closed subgroup of GG, μ\mu a left Haar measure on GG, and β\beta a left Haar measure on HH.

a) *There exist functions ρ\rho continuous and >0> 0 on GG, such that

ρ(xξ)=ΔH(ξ)ΔG(ξ)ρ(x) \rho(x\xi) = \frac{\Delta_H(\xi)}{\Delta_G(\xi)} \rho(x)

for all x∈Gx \in G and ξ∈H\xi \in H.

b) Given such a function ρ\rho, one can form the measure λ=(ρ⋅μ)/β\lambda = (\rho \cdot \mu)/\beta on G/HG/H, and λ\lambda is a nonzero positive measure quasi-invariant under GG.

c) For s,xs, x in GG, ρ(sx)/ρ(x)\rho(sx)/\rho(x) depends only on ss and π(x)\pi(x), hence defines a function χ\chi continuous and >0> 0 on G×(G/H)G \times (G/H) such that

χ(s,π(x))=ρ(sx)ρ(x). \chi(s, \pi(x)) = \frac{\rho(sx)}{\rho(x)}.

Then

γG/H(s)λ=χ(s−1,.)⋅λfor all s∈G. \gamma_{G/H}(s)\lambda = \chi(s^{-1}, .) \cdot \lambda \quad \text{for all } s \in G.

a) follows from Prop. 7. b) follows from Lemmas 5 and 4. c) follows from (17).

Remark 1

One deduces from Remark 1 of No. 3 that the nonzero quasi-invariant measures on G/H are none other than the pseudo-images under π\pi of a Haar measure on G.

Remark 2

If G is a Lie group, we shall see later that the function ρ\rho of Th. 2 can be chosen to be infinitely differentiable.

Under the conditions of Th. 2, certain results of Nos. 3 and 4 may be specialized as follows (on taking into account Ch. V, §4, Th. 2 and Prop. 2 for passing from properties relative to μ\mu to properties relative to ρ⋅μ\rho \cdot \mu):

a) Let ff be a μ\mu-measurable function on G, with values in a topological space, constant outside a countable union of μ\mu-integrable sets; then, the set of x˙∈G/H\dot{x} \in G/H such that the function ξ↦f(xξ)\xi \mapsto f(x\xi) is not β\beta-measurable is locally λ\lambda-negligible.

b) Let ff be a μ\mu-measurable function ⩾0\geqslant 0 on G, zero outside a countable union of μ\mu-integrable sets. Then, the function

x˙↦∫H∗f(xξ)dβ(ξ) \dot{x} \mapsto \int_{H}^{*} f(x\xi) d\beta(\xi)
on G/HG/H is λ\lambda-measurable and
∫G∗f(x)ρ(x)dμ(x)=∫G/H∗dλ(x˙)∫H∗f(xξ)dβ(ξ)(x˙=π(x)). \int_{G}^{*} f(x)\rho(x) d\mu(x) = \int_{G/H}^{*} d\lambda(\dot{x}) \int_{H}^{*} f(x\xi) d\beta(\xi) \quad (\dot{x} = \pi(x)).

c) Let ff be a ρ⋅μ\rho \cdot \mu-integrable function on G, with values in a Banach space or in R‾\overline{\mathbf{R}}. Then, the set of x˙∈G/H\dot{x} \in G/H such that ξ↦f(xξ)\xi \mapsto f(x\xi) is not β\beta-integrable is λ\lambda-negligible; the function x˙↦∫Hf(xξ)dβ(ξ)\dot{x} \mapsto \int_{H} f(x\xi) d\beta(\xi) is λ\lambda-integrable, and

∫Gf(x)ρ(x)dμ(x)=∫G/Hdλ(x˙)∫Hf(xξ)dβ(ξ). \int_{G} f(x)\rho(x) d\mu(x) = \int_{G/H} d\lambda(\dot{x}) \int_{H} f(x\xi) d\beta(\xi).

d) There exists a continuous function h⩾0h \geqslant 0 on G, whose support has compact intersection with the saturation KH of every compact subset K of G, and such that ∫Hh(xξ)dβ(ξ)=1\int_{H} h(x\xi) d\beta(\xi) = 1 for every x∈Gx \in G. For a function g on G/H to be measurable (resp. locally integrable, essentially integrable, integrable) for λ\lambda, it is necessary and sufficient that h⋅(g∘π)h \cdot (g \circ \pi) be so for ρ⋅μ=λ♯\rho \cdot \mu = \lambda^{\sharp}; and, when g is essentially integrable for λ\lambda, one has

∫G/Hg(u)dλ(u)=∫Gh(x)g(π(x))ρ(x)dμ(x). \int_{G/H} g(u) d\lambda(u) = \int_{G} h(x)g(\pi(x)) \rho(x) d\mu(x).

6. Relatively invariant measures on a homogeneous space

Let G again be a locally compact group, H a closed subgroup, β\beta a left Haar measure on H.

Lemma 6

Let λ\lambda be a measure on G/HG/H, χ\chi a continuous representation of G in C∗\mathbf{C}^*. The following properties are equivalent: a) λ\lambda is relatively invariant on G/HG/H with multiplier χ\chi; b) λ#\lambda^\# is relatively invariant on G with left multiplier χ\chi; c) λ#\lambda^\# is of the form aχ⋅μa \chi \cdot \mu (a∈Ca \in \mathbf{C}).

The condition a) means that, for every s∈Gs \in G,

γG/H(s)λ=χ(s)−1λ; \gamma_{G/H}(s)\lambda = \chi(s)^{-1}\lambda;

this is equivalent to (γG/H(s)λ)#=χ(s)−1λ#(\gamma_{G/H}(s)\lambda)^\# = \chi(s)^{-1}\lambda^\#, that is, to

γG(s)λ#=χ(s)−1λ#. \gamma_G(s)\lambda^\# = \chi(s)^{-1}\lambda^\#.

Whence the equivalence of a) and b). The equivalence of b) and c) follows from §1, No. 8, Cor. 1 of Prop. 10.

Theorem 3

Let G be a locally compact group, H a closed subgroup of G, μ\mu (resp. β\beta) a left Haar measure on G (resp. H), χ\chi a continuous representation of G in C∗\mathbf{C}^*.

a) In order that there exist on G/HG/H a nonzero measure relatively invariant under G and with multiplier χ\chi, it is necessary and sufficient that χ(ξ)=ΔH(ξ)/ΔG(ξ)\chi(\xi) = \Delta_H(\xi)/\Delta_G(\xi) for all ξ∈H\xi \in H.

b) This measure is then unique up to a constant factor; more precisely, it is proportional to (χ⋅μ)/β(\chi \cdot \mu)/\beta.

For there to exist on G/HG/H a nonzero measure relatively invariant under G with multiplier χ\chi, it is necessary and sufficient (Lemma 6) that χ⋅μ\chi \cdot \mu be of the form λ#\lambda^\#, hence (No. 2, Prop. 4) that δ(ξ)(χ⋅μ)=ΔH(ξ)(χ⋅μ)\delta(\xi)(\chi \cdot \mu) = \Delta_H(\xi)(\chi \cdot \mu) for all ξ∈H\xi \in H. This condition may also be written χ(ξ)χ⋅ΔG(ξ)μ=ΔH(ξ)χ⋅μ\chi(\xi)\chi \cdot \Delta_G(\xi)\mu = \Delta_H(\xi)\chi \cdot \mu, that is,

χ(ξ)=ΔH(ξ)/ΔG(ξ) \chi(\xi) = \Delta_H(\xi)/\Delta_G(\xi)

for all ξ∈H\xi \in H. Whence a). The assertion b) follows at once from Lemma 6 and the fact that the mapping λ↦λ#\lambda \mapsto \lambda^\# is injective.

We shall see in §3 (No. 3, Example 4) some very simple examples where the representation ξ↦ΔH(ξ)/ΔG(ξ)\xi \mapsto \Delta_H(\xi)/\Delta_G(\xi) cannot be extended to a continuous representation of G in C∗\mathbf{C}^*. In this case, there therefore does not exist any nonzero complex measure on G/HG/H relatively invariant under G.

Corollary 1

For there to exist on G/HG/H a nonzero positive measure relatively invariant under G, it is necessary and sufficient that there exist a continuous representation of GG in R+∗\mathbf{R}_+^* extending the representation ξ↦ΔH(ξ)/ΔG(ξ)\xi \mapsto \Delta_H(\xi)/\Delta_G(\xi).

Note that this condition is fulfilled when HH is unimodular.

Corollary 2

— For there to exist on G/HG/H a nonzero positive measure invariant under GG, it is necessary and sufficient that ΔG\Delta_G coincide with ΔH\Delta_H on HH.

Corollary 3

— Suppose that HH is unimodular and that there exists on G/HG/H a nonzero bounded positive measure ν\nu relatively invariant under GG. Then ν\nu is invariant, and GG is unimodular.

Let χ\chi be the multiplier of ν\nu. For every s∈Gs \in G, ν\nu and γ(s)ν\gamma(s)\nu have the same finite total mass (§ 1, No. 1, formula (6)); since γ(s)ν=χ(s)−1ν\gamma(s)\nu = \chi(s)^{-1}\nu, we have χ(s)=1\chi(s) = 1. Thus ν\nu is invariant. By Cor. 2, ΔG(s)=1\Delta_G(s) = 1 for all s∈Hs \in H. Let G′G' be the set of t∈Gt \in G such that ΔG(t)=1\Delta_G(t) = 1. This is a closed normal subgroup of GG containing HH. Let π\pi be the canonical mapping of G/HG/H onto G/G′G/G'. Then π(ν)\pi(\nu) is a nonzero, bounded positive measure invariant under GG. Therefore the left Haar measure of the group G/G′G/G' is bounded, so that G/G′G/G' is compact (§ 1, No. 2, Prop. 2). Consequently the image of GG under ΔG\Delta_G is a compact subgroup of R+∗\mathbf{R}_+^*; this subgroup is reduced to {1}\{1\}, thus ΔG=1\Delta_G = 1 on all of GG.

7. Haar measure on a quotient group

Proposition 10

— Let GG be a locally compact group, G′G' a closed normal subgroup, G′′G'' the group G/G′G/G', π\pi the canonical mapping of GG onto G/G′G/G', and α,α′,α′′\alpha, \alpha', \alpha'' left Haar measures on G,G′,G′′G, G', G''.

a) Multiplying α\alpha by a constant factor if necessary, we can suppose that α′′=α/α′\alpha'' = \alpha/\alpha'. In particular, if f∈K(G)f \in \mathscr{K}(G) then

∫Gf(x) dα(x)=∫G′′dα′′(x˙)∫G′f(xξ) dα′(ξ)(x˙=π(x)). \int_G f(x)\,d\alpha(x) = \int_{G''} d\alpha''(\dot{x}) \int_{G'} f(x\xi)\,d\alpha'(\xi) \quad (\dot{x} = \pi(x)).

b) One has ΔG(ξ)=ΔG′(ξ)\Delta_G(\xi) = \Delta_{G'}(\xi) for all ξ∈G′\xi \in G'; in particular, if GG is unimodular then so is G′G'.

c) The kernel of the representation ΔG\Delta_G of GG in R+∗\mathbf{R}_+^* is the largest unimodular closed normal subgroup of GG.

Applying Th. 3 of No. 6 with χ=1\chi = 1 (and in the knowledge that here, there exists a measure on G/G′G/G' invariant under GG, namely α′′\alpha''), we obtain a) and b); c) follows at once from b).

Proposition 11

— We maintain the notations of Prop. 10. Let uu be an automorphism of GG such that u(G′)=G′u(G') = G'. Let u′u' be the restriction of uu to G′G', and u′′u'' the automorphism of G′′G'' deduced from uu by passage to the quotients. Then

mod⁡G(u)=mod⁡G′(u′)mod⁡G′′(u′′). \operatorname{mod}_G(u) = \operatorname{mod}_{G'}(u') \operatorname{mod}_{G''}(u'').

For, if α′′=α/α′\alpha'' = \alpha/\alpha' then u′′(α′′)=u(α)/u′(α′)u''(\alpha'') = u(\alpha)/u'(\alpha'), that is,

mod⁡G′′(u′′)−1α′′=mod⁡G(u)−1α/mod⁡G′(u′)−1α′ \operatorname{mod}_{G''}(u'')^{-1}\alpha'' = \operatorname{mod}_G(u)^{-1}\alpha/\operatorname{mod}_{G'}(u')^{-1}\alpha'
=mod⁡G′(u′)mod⁡G(u)(α/α′)=mod⁡G′(u′)mod⁡G(u)α′′, = \frac{\operatorname{mod}_{G'}(u')}{\operatorname{mod}_G(u)} (\alpha/\alpha') = \frac{\operatorname{mod}_{G'}(u')}{\operatorname{mod}_G(u)} \alpha'',
whence the proposition.

Corollary

— For every x∈Gx \in G,

ΔG(x)=ΔG/G′(x˙)mod⁡(ix), \Delta_G(x) = \Delta_{G/G'}(\dot{x}) \operatorname{mod}(i_x),
where x˙\dot{x} denotes the canonical image of xx in G/G′G/G', and ixi_x the automorphism s↦x−1sxs \mapsto x^{-1}sx of G′G'.

This follows from Prop. 11, and formula (33) of §1, No. 4.

8. A transitivity property

Let XX be a locally compact space in which a locally compact group HH acts on the right, continuously and properly, by (x,ξ)↦xξ(x, \xi) \mapsto x\xi (x∈X, ξ∈Hx \in X,\ \xi \in H). Let H′H' be a closed subgroup of HH; then H′H' operates on the right in XX, continuously and properly. We shall denote by π,π′,p\pi, \pi', p the canonical mappings of XX onto X/HX/H, of XX onto X/H′X/H', and of HH onto H/H′H/H'.

Let β,β′\beta, \beta' be left Haar measures on H,H′H, H'; suppose that ΔH\Delta_H and ΔH′\Delta_{H'} coincide on H′H'; one can then form the measure β/β′\beta/\beta' on H/H′H/H', left-invariant under HH (No. 6, Th. 3). On the other hand, let μ\mu be a positive measure on XX such that

δ(ξ)μ=ΔH(ξ)μ \delta(\xi)\mu = \Delta_H(\xi)\mu
for ξ∈H\xi \in H; one can then form the measures μ/β\mu/\beta on X/HX/H and μ/β′\mu/\beta' on X/H′X/H' (No. 2, Prop. 4). We are going to write μ/β′\mu/\beta' as the integral, with respect to μ/β\mu/\beta, of a family of measures on X/H′X/H' indexed by the points of X/HX/H. When H′={e}H' = \{e\}, we shall find ourselves again in the situation of No. 3.

The mapping (x,ξ)↦π′(xξ)(x, \xi) \mapsto \pi'(x\xi) of X×HX \times H into X/H′X/H' is continuous; since π′(xξ)=π′(xξξ′)\pi'(x\xi) = \pi'(x\xi\xi') for all ξ′∈H′\xi' \in H' this mapping defines, by passage to the quotient, a continuous mapping of X×(H/H′)X \times (H/H') into X/H′X/H'; whence, for each fixed xx in XX, a partial mapping ωx\omega_x of H/H′H/H' into X/H′X/H', deduced by passage to the quotient from the mapping ψx:ξ↦xξ\psi_x : \xi \mapsto x\xi of H into X. Note that ψxξ=ψx∘γH(ξ)\psi_{x\xi} = \psi_x \circ \gamma_H(\xi), therefore that ωxξ=ωx∘γH/H′(ξ)\omega_{x\xi} = \omega_x \circ \gamma_{H/H'}(\xi) for all ξ∈H\xi \in H.

Lemma 7

Let K be a compact subset of X/H', and L a compact subset of X. Then ⋃x∈Lωx−1(K)\bigcup_{x \in L} \omega_x^{-1}(K) is relatively compact in H/H'.

Let K1K_1 be a compact subset of X such that π′(K1)=K\pi'(K_1) = K. Let K2K_2 be the set of ξ∈H\xi \in H such that LξL\xi intersects K1K_1. Then K2K_2 is compact (GT, III, §4, No. 5, Th. 1). Let ξ∈H\xi \in H be such that p(ξ)∈⋃x∈Lωx−1(K)p(\xi) \in \bigcup_{x \in L} \omega_x^{-1}(K). Thus, there exists an x∈Lx \in L such that ωx(p(ξ))∈K\omega_x(p(\xi)) \in K, in other words such that π′(xξ)∈K\pi'(x\xi) \in K. Since π′(K1)=K\pi'(K_1) = K, there exists ξ′∈H′\xi' \in H' such that xξξ′∈K1x\xi\xi' \in K_1. Then ξξ′∈K2\xi\xi' \in K_2, therefore p(ξ)=p(ξξ′)∈p(K2)p(\xi) = p(\xi\xi') \in p(K_2). We have thus shown that ⋃x∈Lωx−1(K)⊂p(K2)\bigcup_{x \in L} \omega_x^{-1}(K) \subset p(K_2).

This lemma shows first of all that the mapping ωx\omega_x is proper. One can therefore form the measure ωx(β/β′)\omega_x(\beta/\beta') on X/H′X/H', which is concentrated on ωx(H/H′)=π′(ψx(H))=π′(xH)\omega_x(H/H') = \pi'(\psi_x(H)) = \pi'(xH). If f∈K(X/H′)f \in \mathscr{K}(X/H'), Lemma 7 and §1, No. 1, Lemma 1 show that the function x↦⟨f,ωx(β/β′)⟩x \mapsto \langle f, \omega_x(\beta/\beta') \rangle is continuous on X; moreover, ⟨f,ωx(β/β′)⟩\langle f, \omega_x(\beta/\beta') \rangle is zero when Supp ff does not intersect π′(xH)\pi'(xH), in other words when π(x)\pi(x) does not belong to the canonical image of Supp ff in X/HX/H.

Moreover, if x∈Hx \in H then

ωxξ(β/β′)=ωx(γH/H′(ξ)(β/β′))=ωx(β/β′). \omega_{x\xi}(\beta/\beta') = \omega_x(\gamma_{H/H'}(\xi)(\beta/\beta')) = \omega_x(\beta/\beta').

The mapping x↦ωx(β/β′)x \mapsto \omega_x(\beta/\beta') of X into M(X/H′)\mathscr{M}(X/H') therefore defines by passage to the quotient a mapping u↦(β/β′)uu \mapsto (\beta/\beta')_u of X/HX/H into M(X/H′)\mathscr{M}(X/H'). The foregoing shows that, for every f∈K(X/H′)f \in \mathscr{K}(X/H'), the mapping u↦⟨f,(β/β′)u⟩u \mapsto \langle f, (\beta/\beta')_u \rangle is continuous with compact support. Consequently the mapping u↦(β/β′)uu \mapsto (\beta/\beta')_u is a vaguely continuous and (μ/β)(\mu/\beta)*-adequate family of measures on* X/H′X/H', with X/HX/H as index set.

Let x∈Xx \in X, and u=π(x)∈X/Hu = \pi(x) \in X/H. Let ff be a function on X/H′X/H', with values in a Banach space or in R‾\overline{\mathbf{R}}. By Ch. V, §4, Th. 2, for ff to be (β/β′)u(\beta/\beta')_u-integrable, it is necessary and sufficient that the function p(ξ)↦f(ωx(p(ξ)))=f(π′(xξ))p(\xi) \mapsto f(\omega_x(p(\xi))) = f(\pi'(x\xi)) on H/H′H/H' be (β/β′)(\beta/\beta')-integrable, in which case

(21)∫X/H′f(u′)d(β/β′)u(u′)=∫H/H′f(π′(xξ))d(β/β′)(ξ˙)(ξ˙=p(ξ)). (21) \quad \int_{X/H'} f(u') d(\beta/\beta')_u(u') = \int_{H/H'} f(\pi'(x\xi)) d(\beta/\beta')(\dot{\xi}) \qquad (\dot{\xi} = p(\xi)).

One has analogous properties for measurability, the upper integral and the essential integral.

Proposition 12

— With the foregoing notations,

∫X/H(β/β′)ud(μ/β)(u)=μ/β′. \int_{X/H} (\beta/\beta')_u d(\mu/\beta)(u) = \mu/\beta'.

Let f∈K(X)f \in \mathscr{K}(X), and let fb∈K(X/H′)f^b \in \mathscr{K}(X/H'), defined by

fb(π′(x))=∫H′f(xξ′)dβ′(ξ′). f^b(\pi'(x)) = \int_{H'} f(x\xi') d\beta'(\xi').

It suffices (cf. No. 2) to prove that fbf^b has the same integral with respect to the two members of (22). Now, ⟨μ/β′,fb⟩=⟨μ,f⟩\langle \mu/\beta', f^b \rangle = \langle \mu, f \rangle. On the other hand,

⟨∫X/H(β/β′)ud(μ/β)(u),fb⟩=∫X/H⟨(β/β′)u,fb⟩d(μ/β)(u). \left\langle \int_{X/H} (\beta/\beta')_u d(\mu/\beta)(u), f^b \right\rangle = \int_{X/H} \langle (\beta/\beta')_u, f^b \rangle d(\mu/\beta)(u).

Now, let x∈Xx \in X and u=π(x)u = \pi(x). We have

⟨(β/β′)u,fb⟩=⟨ωx(β/β′),fb⟩=∫H/H′fb(ωx(ξ˙))d(β/β′)(ξ˙) \langle (\beta/\beta')_u, f^b \rangle = \langle \omega_x(\beta/\beta'), f^b \rangle = \int_{H/H'} f^b(\omega_x(\dot{\xi})) d(\beta/\beta')(\dot{\xi})
=∫H/H′fb(π′(xξ))d(β/β′)(ξ˙) = \int_{H/H'} f^b(\pi'(x\xi)) d(\beta/\beta')(\dot{\xi})
=∫H/H′d(β/β′)(ξ˙)∫H′f(xξξ′)dβ′(ξ′) = \int_{H/H'} d(\beta/\beta')(\dot{\xi}) \int_{H'} f(x\xi\xi') d\beta'(\xi')
=∫Hf(xξ)dβ(ξ). = \int_H f(x\xi) d\beta(\xi).

Therefore

⟨∫X/H(β/β′)ud(μ/β)(u),fb⟩=∫X/Hd(μ/β)(u)∫Hf(xξ)dβ(ξ)=⟨μ,f⟩, \left\langle \int_{X/H} (\beta/\beta')_u d(\mu/\beta)(u), f^b \right\rangle = \int_{X/H} d(\mu/\beta)(u) \int_H f(x\xi) d\beta(\xi) = \langle \mu, f \rangle,

which proves the proposition.

Corollary 1

— a) *Let f be a function on X/H', with values in a Banach space or in R‾\overline{\mathbf{R}}, integrable for μ/β′\mu/\beta'. There exists a (μ/β)(\mu/\beta)-negligible subset N of X/H having the following property: if x∈Xx \in X is such that π(x)∉N\pi(x) \notin N, then the function f∘ωxf \circ \omega_x on H/H', that is, the function ξ˙↦f(π′(xξ))\dot{\xi} \mapsto f(\pi'(x\xi)), is integrable for β/β′\beta/\beta'. The integral ∫H/H′f(π′(xξ))d(β/β′)(ξ˙)\int_{H/H'} f(\pi'(x\xi)) d(\beta/\beta')(\dot{\xi}) depends only on x˙=π(x)\dot{x} = \pi(x), and is a (μ/β)(\mu/\beta)-integrable function of x˙\dot{x}; and

∫X/H′fd(μ/β′)=∫X/Hd(μ/β)(x˙)∫H/H′f(π′(xξ))d(β/β′)(ξ˙). \int_{X/H'} f d(\mu/\beta') = \int_{X/H} d(\mu/\beta)(\dot{x}) \int_{H/H'} f(\pi'(x\xi)) d(\beta/\beta')(\dot{\xi}).

b) Let ff be a function ≥0\geq 0 on X/H′X/H', measurable for μ/β′\mu/\beta' and zero outside a countable union of (μ/β′)(\mu/\beta')-integrable sets. Then

π(x)↦∫H/H′∗f(π′(xξ))d(β/β′)(ξ˙) \pi(x) \mapsto \int_{H/H'}^* f(\pi'(x\xi)) d(\beta/\beta')(\dot{\xi})

is (μ/β)(\mu/\beta)-measurable, and

∫X/H′∗fd(μ/β′)=∫X/H∗d(μ/β)(x˙)∫H/H′∗f(π′(xξ))d(β/β′)(ξ˙). \int_{X/H'}^* f d(\mu/\beta') = \int_{X/H}^* d(\mu/\beta)(\dot{x}) \int_{H/H'}^* f(\pi'(x\xi)) d(\beta/\beta')(\dot{\xi}).

c) Let ff be a function on X/H′X/H' with values in a Banach space or in R‾\overline{\mathbf{R}}, measurable for μ/β′\mu/\beta' and zero outside a countable union of (μ/β′)(\mu/\beta')-integrable sets. Then, for ff to be (μ/β′)(\mu/\beta')-integrable, it is sufficient that

∫X/H∗d(μ/β)(x˙)∫H/H′∗∣f(π′(xξ))∣d(β/β′)(ξ˙)<+∞. \int_{X/H}^* d(\mu/\beta)(\dot{x}) \int_{H/H'}^* |f(\pi'(x\xi))| d(\beta/\beta')(\dot{\xi}) < +\infty.

Corollary 2

— Let GG be a locally compact group, AA and BB closed subgroups of GG such that A⊃BA \supset B. Suppose that there exists, on the homogeneous space G/BG/B of left cosets with respect to BB, a nonzero positive measure α\alpha that is invariant under GG and is bounded.

a) The canonical image of α\alpha in G/AG/A is a nonzero positive measure, invariant under GG, and bounded.

b) ΔG\Delta_G coincides with ΔA\Delta_A on AA and with ΔB\Delta_B on BB.

c) There exists, on the homogeneous space A/BA/B of left cosets of AA with respect to BB, a nonzero positive measure invariant under AA and bounded.

The assertion a) is immediate. The assertion b) follows from a) and No. 6, Cor. 2 of Th. 3. By b), ΔA\Delta_A coincides with ΔB\Delta_B on BB, and one can therefore apply the results of the present subsection, on taking X=GX = G, H=AH = A, H′=BH' = B. The function 1 on G/BG/B is α\alpha-integrable. By a) of Cor. 1, the function 1 on A/BA/B is integrable for β/β′\beta/\beta', where β\beta and β′\beta' denote left Haar measures on AA and BB; thus β/β′\beta/\beta' is bounded.

9. Construction of the Haar measure of a group from the Haar measures of certain subgroups

Let GG be a locally compact group, XX and YY two closed subgroups of GG such that Ω=XY\Omega = XY contains a neighborhood UU of ee. Then Ω\Omega is open in GG; for, for any x0∈Xx_0 \in X and y0∈Yy_0 \in Y, XY=(x0X)(Yy0)⊃x0Uy0XY = (x_0X)(Yy_0) \supset x_0Uy_0, and x0Uy0x_0 U y_0 is a neighborhood of x0y0x_0 y_0; thus Ω\Omega is a neighborhood of each of its points.

When G is a Lie group with Lie algebra gg, the condition imposed on X and Y is satisfied if the subalgebras corresponding to X and Y have sum gg.

The group X×YX \times Y operates continuously in G on the left, by the law (x,y)⋅s=xsy−1(x, y) \cdot s = x s y^{-1} (x∈X,y∈Y,s∈Gx \in X, y \in Y, s \in G). Let Z=X∩YZ = X \cap Y. The stabilizer of e in X×YX \times Y is the subgroup Z0Z_0 of X×YX \times Y formed by the pairs (z,z)(z, z), where z∈Zz \in Z, a subgroup canonically isomorphic to ZZ. Thus the set Ω\Omega may be identified with the homogeneous space (X×Y)/Z0(X \times Y)/Z_0 of left cosets; more precisely, the mapping (x,y)↦xy−1(x, y) \mapsto xy^{-1} of X×YX \times Y onto Ω\Omega defines, by passage to the quotient, a continuous bijection of (X×Y)/Z0(X \times Y)/Z_0 onto Ω\Omega. We shall assume that this mapping is a homeomorphism. (This is notably the case if G is countable at infinity: cf. App. I.)

Proposition 13

— *Suppose in addition that Z is compact. Let μG,μX,μY\mu_G, \mu_X, \mu_Y be left Haar measures on G, X, Y, and Λ\Lambda the restriction of ΔG\Delta_G to Y. Then the restriction μ\mu of μG\mu_G to Ω\Omega is, up to a constant factor, the image of μX⊗(Λ−1⋅μY)\mu_X \otimes (\Lambda^{-1} \cdot \mu_Y) under the mapping (x,y)↦xy−1(x, y) \mapsto xy^{-1} of X×YX \times Y onto Ω\Omega (a mapping that is proper).

For x∈X,y∈Yx \in X, y \in Y,

γ((x,y))μ=δ(y)γ(x)μ=ΔG(y)μ. \gamma((x, y)) \mu = \delta(y) \gamma(x) \mu = \Delta_G(y) \mu .

Identifying Ω\Omega with the homogeneous space (X×Y)/Z0(X \times Y)/Z_0 and choosing a suitable Haar measure on Z0Z_0, one sees that μ#\mu^\# is the product of the left Haar measure of X×YX \times Y, namely μX⊗μY\mu_X \otimes \mu_Y, by the function (x,y)↦ΔG(y)−1(x, y) \mapsto \Delta_G(y)^{-1} (No. 6, Lemma 6). On the other hand μ\mu is, up to a constant factor, the image of μ#\mu^\# under the canonical mapping of X×YX \times Y onto Ω\Omega (No. 3, Remark 2).

Corollary

— *Let f be a function defined on Ω\Omega, with values in a Banach space or in R‾\overline{\mathbf{R}}. For f to be μ\mu-integrable, it is necessary and sufficient that the function (x,y)↦f(xy)ΔG(y)ΔY(y)−1(x, y) \mapsto f(xy) \Delta_G(y) \Delta_Y(y)^{-1} be (μX⊗μY)(\mu_X \otimes \mu_Y)-integrable, in which case

∫Ωf(ω)dμ(ω)=a∬X×Yf(xy)ΔG(y)ΔY(y)−1dμX(x)dμY(y), \int_\Omega f(\omega) d\mu(\omega) = a \iint_{X \times Y} f(xy) \Delta_G(y) \Delta_Y(y)^{-1} d\mu_X(x) d\mu_Y(y),

where aa is a constant >0> 0 independent of f.

By Prop. 13, and Ch. V, §4, No. 4, Th. 2, for f to be μ\mu-integrable, it is necessary and sufficient that the function (x,y)↦f(xy−1)(x, y) \mapsto f(xy^{-1}) be integrable for μX⊗(Λ−1⋅μY)\mu_X \otimes (\Lambda^{-1} \cdot \mu_Y), or again that the function (x,y)↦f(xy−1)ΔG(y)−1(x, y) \mapsto f(xy^{-1}) \Delta_G(y)^{-1} be integrable for μX⊗μY\mu_X \otimes \mu_Y, or again that the function (x,y)↦f(xy)ΔG(y)ΔY(y)−1(x, y) \mapsto f(xy) \Delta_G(y) \Delta_Y(y)^{-1} be integrable for μX⊗μY\mu_X \otimes \mu_Y. Formula (23) results from an analogous argument.

Proposition 14

— Suppose that the conditions of Prop. 13 are fulfilled and that, in addition, Y is normal.

a) The restriction of μG\mu_G to Ω\Omega is, up to a constant factor, the image of μX⊗μY\mu_X \otimes \mu_Y under the mapping (x,y)↦xy(x, y) \mapsto xy of X×YX \times Y onto Ω\Omega.

b) For x∈Xx \in X and y∈Yy \in Y,

ΔG(xy)=ΔX(x)ΔY(y)mod⁡(ix), \Delta_G(xy) = \Delta_X(x)\Delta_Y(y) \operatorname{mod}(i_x),

where ixi_x denotes the automorphism v↦x−1vxv \mapsto x^{-1}vx of Y.

We have ΔG=ΔY\Delta_G = \Delta_Y on Y (Prop. 10 b)), therefore a) follows from (23). Let x0∈X, y0∈Yx_0 \in X,\ y_0 \in Y. Denote by pp the mapping (x,y)↦xy(x, y) \mapsto xy of X×YX \times Y onto Ω\Omega. Since

xy(x0y0)−1=xx0−1(x0yy0−1x0−1)=xx0−1ix0−1(yy0−1), xy(x_0y_0)^{-1} = xx_0^{-1}(x_0yy_0^{-1}x_0^{-1}) = xx_0^{-1}i_{x_0^{-1}}(yy_0^{-1}),

we have

ΔG(x0y0)p(μX⊗μY)=δ(x0y0)p(μX⊗μY)=p(δ(x0)μX⊗ix0−1δ(y0)μY)=p(ΔX(x0)μX⊗ΔY(y0)(mod⁡ ix0)μY)=ΔX(x0)ΔY(y0)(mod⁡ ix0)p(μX⊗μY), \begin{align*} \Delta_G(x_0y_0)p(\mu_X \otimes \mu_Y) &= \delta(x_0y_0)p(\mu_X \otimes \mu_Y) \\ &= p(\delta(x_0)\mu_X \otimes i_{x_0^{-1}}\delta(y_0)\mu_Y) \\ &= p(\Delta_X(x_0)\mu_X \otimes \Delta_Y(y_0)(\operatorname{mod}\ i_{x_0})\mu_Y) \\ &= \Delta_X(x_0)\Delta_Y(y_0)(\operatorname{mod}\ i_{x_0})p(\mu_X \otimes \mu_Y), \end{align*}

whence b).

Remark

Prop. 14 applies in particular when G is the topological semi-direct product of X by Y (GT, III, §2, No. 10). In this case, Z={e}Z = \{e\} and Ω=G\Omega = G. Since yx=xix(y)yx = xi_x(y) for x∈X, y∈Yx \in X,\ y \in Y, the measure μG\mu_G is also, up to a constant factor, the image of (mod⁡ ix)μX⊗μY(\operatorname{mod}\ i_x)\mu_X \otimes \mu_Y under the mapping (x,y)↦yx(x, y) \mapsto yx of X×YX \times Y into G.

10. Integration on a fundamental domain

Let X be a locally compact space, H a discrete group operating on the right continuously and properly in X. Let π\pi be the canonical mapping of X onto X/HX/H. For every x∈Xx \in X, we denote by HxH_x the stabilizer of xx in H; this is a finite subgroup of H (GT, III, §4, No. 2, Prop. 4); its order will be denoted n(x)n(x). For every s∈H, Hxs=s−1Hxss \in H,\ H_{xs} = s^{-1}H_x s, therefore n(xs)=n(x)n(xs) = n(x). There exists an open neighborhood U of x such that U∩Us=∅U \cap Us = \varnothing for s∉Hxs \notin H_x (loc. cit., No. 4, proof of Prop. 8); for y∈Uy \in U, one has Hy⊂HxH_y \subset H_x; thus the function nn on X is upper semi-continuous. When X is countable at infinity, H is countable; for, let (K1,K2,…)(K_1, K_2, \ldots) be a covering of X by a sequence of compact subsets, and let x0∈Xx_0 \in X; the set of s∈Hs \in H such that x0s∈Kix_0 s \in K_i is finite (loc. cit., No. 5, Th. 1), whence our assertion.

Definition 2

Let F⊂XF \subset X. One says that FF is a fundamental domain (for HH) if the restriction of π\pi to FF is a bijection of FF onto X/HX/H* (in other words, FF is a system of representatives for the equivalence relation defined by HH).

Lemma 8

— Let FF be a fundamental domain. For every x∈Xx \in X,

∑s∈HφFs(x)=n(x). \sum_{s \in H} \varphi_{Fs}(x) = n(x).
(24)

Since φFs(xt)=φFst−1(x)\varphi_{Fs}(xt) = \varphi_{Fst^{-1}}(x) for all ss and tt in HH, the two members of (24) remain invariant when xx is replaced by xtxt. We can therefore suppose that x∈Fx \in F. We then have the equivalences

φFs(x)=1  ⟺  x∈Fs  ⟺  xs−1∈F  ⟺  xs−1=x  ⟺  s∈Hx, \varphi_{Fs}(x) = 1 \iff x \in Fs \iff xs^{-1} \in F \iff xs^{-1} = x \iff s \in H_x,
whence (24).

Proposition 15

Assume that XX is countable at infinity. Let μ\mu be a measure ≥0\geq 0 on XX. Let FF be a fundamental domain such that FsFs is μ\mu-measurable for every s∈Hs \in H. Let ff be a μ\mu-integrable function on XX, with values in a Banach space or in R‾\overline{\mathbf{R}}. Then the family of the*

∫Fsn(x)−1f(x)dμ(x)(s∈H) \int_{Fs} n(x)^{-1} f(x) d\mu(x) \qquad (s \in H)
is summable, and
∫Xf(x)dμ(x)=∑s∈H∫Fsn(x)−1f(x)dμ(x). \int_X f(x) d\mu(x) = \sum_{s \in H} \int_{Fs} n(x)^{-1} f(x) d\mu(x).

If AA is a finite subset of HH, then

∣∑s∈An−1fφFs∣≤n−1∣f∣∑s∈AφFs≤∣f∣ \left| \sum_{s \in A} n^{-1} f \varphi_{Fs} \right| \leq n^{-1} |f| \sum_{s \in A} \varphi_{Fs} \leq |f|
by Lemma 8. Lemma 8 also proves that ∑s∈An−1fφFs\sum_{s \in A} n^{-1} f \varphi_{Fs} converges pointwise to ff with respect to the increasing directed set of finite subsets of HH. Prop. 15 then follows from Ch. IV, §4, No. 3, Th. 2.

Theorem 4

Let XX be a locally compact space countable at infinity, HH a discrete group operating continuously and properly on the right in XX, π the canonical mapping of X onto X/H, μ a positive measure on X invariant under H, β the normalized Haar measure of H, and λ = μ/β. Let F be a μ-measurable fundamental domain.

a) The pair (π,n−1φF)(\pi, n^{-1}\varphi_F) is μ-adapted, and

∫Xn(x)−1φF(x)επ(x) dμ(x)=λ. \int_X n(x)^{-1} \varphi_F(x) \varepsilon_{\pi(x)} \, d\mu(x) = \lambda.

b) The mapping π is proper for n−1φF⋅μn^{-1}\varphi_F \cdot \mu, and π(n−1φF⋅μ)=λ\pi(n^{-1}\varphi_F \cdot \mu) = \lambda.

c) Let k be a function on X/H. For k to be λ-measurable (resp. λ-integrable), it is necessary and sufficient that n−1φF(k∘π)n^{-1}\varphi_F(k \circ \pi) be μ-measurable (resp. μ-integrable); and, if k is λ-integrable then

∫X/Hk dλ=∫Fn−1(k∘π) dμ. \int_{X/H} k \, d\lambda = \int_F n^{-1}(k \circ \pi) \, d\mu.

We have μ=λ♯\mu = \lambda^\sharp. Let f∈H+(X/H)f \in \mathscr{H}_+(X/H). Then n−1φF(f∘π)n^{-1}\varphi_F(f \circ \pi) is μ-measurable and ≥0\geq 0, and by Prop. 5 b) of No. 3 we have

∫X∗n(x)−1φF(x)f(π(x))dμ(x)=∫X/H∗f(x˙)dλ(x˙)∫H∗n(xξ)−1φF(xξ)dβ(ξ) \int_X^* n(x)^{-1} \varphi_F(x) f(\pi(x)) d\mu(x) = \int_{X/H}^* f(\dot{x}) d\lambda(\dot{x}) \int_H^* n(x\xi)^{-1} \varphi_F(x\xi) d\beta(\xi)

and ∫H∗n(xξ)−1φF(xξ)dβ(ξ)=n(x)−1∑ξ∈HφF(xξ)=1\int_H^* n(x\xi)^{-1} \varphi_F(x\xi) d\beta(\xi) = n(x)^{-1} \sum_{\xi \in H} \varphi_F(x\xi) = 1 by Lemma 8. Thus n−1φF⋅(f∘π)n^{-1}\varphi_F \cdot (f \circ \pi) is μ-integrable and

∫Xn(x)−1φF(x)f(π(x))dμ(x)=∫X/Hf(x˙)dλ(x˙). \int_X n(x)^{-1} \varphi_F(x) f(\pi(x)) d\mu(x) = \int_{X/H} f(\dot{x}) d\lambda(\dot{x}).

This proves a). The assertion b) is proved similarly. The assertion c) may be deduced from a) and from Ch. V, §4, Prop. 3 and Th. 2.

Corollary

— We maintain the hypotheses and notations of Th. 4. Let F' be a second μ-measurable fundamental domain. Let u be a function on X, with values in a Banach space or in R‾\overline{\mathbf{R}}, invariant under H. Suppose that u is μ-integrable on F. Then u is μ-integrable on F' and

∫Fu(x)dμ(x)=∫F′u(x)dμ(x). \int_F u(x) d\mu(x) = \int_{F'} u(x) d\mu(x).

Since u and n are invariant under H, there exists a function v on X/H such that v o π coincides with nu on F and on F'. Then n−1φF(v∘π)=φF′u,n−1φF′(v∘π)=φF′u.n^{-1} \varphi_F(v \circ \pi) = \varphi_{F'} u, \quad n^{-1} \varphi_{F'}(v \circ \pi) = \varphi_{F'} u. By hypothesis, n−1φF(v∘π)n^{-1} \varphi_F(v \circ \pi) is μ\mu-integrable. By Th. 4, vv is λ\lambda-integrable, φF′μ\varphi_{F'} \mu is μ\mu-integrable, and

∫Fudμ=∫X/Hvdλ=∫F′udμ. \int_F u d\mu = \int_{X/H} v d\lambda = \int_{F'} u d\mu.

For the existence of μ\mu-measurable fundamental domains, see Exer. 12.

Exercises

See the exercises for § 2.

Transcribed from the printed volume. Not translated.