Bourbaki

Appendix

Fixed points of groups of affine transformations

1. The case of solvable groups

Let E be a real vector space, and K a convex subset of E. A mapping u:K→Ku : K \to K such that for x,yx, y in K and for every real number tt in [0,1][0, 1], we have

u(tx+(1−t)y)=tu(x)+(1−t)u(y) u(t x + (1 - t) y) = t u(x) + (1 - t) u(y)

is said to be an affine transformation. From relation (1) we deduce that

u(∑i∈Itixi)=∑i∈Itiu(xi) u \left( \sum_{i \in I} t_i x_i \right) = \sum_{i \in I} t_i u(x_i)

for every finite set I, points xix_i in K and positive real numbers tit_i such that ∑i∈Iti=1\sum_{i \in I} t_i = 1.

Let uu and vv be two affine transformations on K, then the mapping u∘vu \circ v is an affine transformation on K. If v:E→Ev : E \to E is a linear mapping such that v(K)⊂Kv(K) \subset K, the mapping u:K→Ku : K \to K which coincides with vv on K is an affine transformation.

Theorem 1 (Markoff-Kakutani)

— *Let E be a Hausdorff locally convex vector space over the field R\mathbf{R}, and K a non-empty compact convex subset of E. Let Γ\Gamma be a set of continuous affine transformations on K, pairwise permutable. Then there exists a point a in K such that u(a)=au(a) = a for all u∈Γu \in \Gamma.

For every u∈Γu \in \Gamma, let KuK_u be the set of all x∈Kx \in K such that u(x)=xu(x) = x. We shall show that KuK_u is non-empty. Let xx be a point of K; for every integer n≥1n \geq 1, let xnx_n be the element 1n∑i=0n−1ui(x)\frac{1}{n} \sum_{i=0}^{n-1} u^i(x) of E. Since K is convex and stable under uu, the points xnx_n belong to K and since K is compact, there exists a limit point aa of the sequence (xn)n≥1(x_n)_{n \geq 1}. The mapping y↦u(y)−yy \mapsto u(y) - y from K into E is continuous, hence u(a)−au(a) - a is a limit point of the sequence (u(xn)−xn)n≥1(u(x_n) - x_n)_{n \geq 1}. But we have u(xn)−xn=1n(un(x)−x)u(x_n) - x_n = \frac{1}{n} (u^n(x) - x).

Since K is compact, hence also bounded (III, p. 3, prop. 2), the sequence (un(x)−x)n≥1(u^n(x) - x)_{n \geq 1} is bounded; consequently, the sequence (1n(un(x)−x))n≥1\left( \frac{1}{n}(u^n(x) - x) \right)_{n \geq 1} tends to 0 (III, p. 4, prop. 3), and since E is Hausdorff, we have u(a)−a=0u(a) - a = 0. Therefore a∈Kua \in K_u.

Each of the sets KuK_u is a closed and convex subset of the compact space K, and we shall prove that the intersection ⋂u∈ΓKu\bigcap_{u \in \Gamma} K_u is non-empty. For this it is enough to prove that, for n≥1n \geq 1, and u1,...,unu_1, ..., u_n in Γ\Gamma, the set Ku1∩...∩KunK_{u_1} \cap ... \cap K_{u_n} is non-empty. The case n=1n = 1 having been considered, we argue by induction on nn. Suppose n≥2n \geq 2 and put L=Ku1∩...∩Kun−1L = K_{u_1} \cap ... \cap K_{u_{n-1}}. By the hypothesis of induction, L is a non-empty compact convex subset of E. Since unu_n commutes with u1,...,un−1u_1, ..., u_{n-1}, we have un(L)⊂Lu_n(L) \subset L. Applying the first part of the proof to the affine transformation induced by unu_n on L, we conclude that there exists a point aa in L such that un(a)=au_n(a) = a; then aa belongs to Ku1∩...∩KunK_{u_1} \cap ... \cap K_{u_n}, which is then non-empty.

Corollary

— *Let G be a solvable group of continuous affine transformations on K. Then there exists a point in K which is invariant under G.

By the definition of a solvable group (A, I, § 6, No. 4) there exists a finite decreasing sequence (Gi)0≤i≤n(G_i)_{0 \leq i \leq n} of distinct subgroups of G, such that G0=GG_0 = G, Gn={e}G_n = \{e\} and such that the group Gi−1/GiG_{i-1}/G_i is commutative for 1≤i≤n1 \leq i \leq n. Let KiK_i denote the set of fixed points of GiG_i in K. Then Kn=KK_n = K. Moreover, for 1≤i≤n1 \leq i \leq n, every element of GiG_i induces the identity transformation on KiK_i; we thus deduce an action of the abelian group Gi−1/GiG_{i-1}/G_i on K if KiK_i is non-empty; it follows from th. 1 that the set Ki−1K_{i-1} of fixed points of Gi−1/GiG_{i-1}/G_i in KiK_i is non-empty. By descending induction on ii, we conclude that K0K_0 is non-empty, hence the corollary.

2. Invariant means

Let X be a topological space. Let B(X;R)\mathscr{B}(X; \mathbf{R}) denote the real vector space consisting of continuous bounded mappings from X into R\mathbf{R}. Endowed with the norm ∥f∥=sup⁡x∈X∣f(x)∣\|f\| = \sup_{x \in X} |f(x)|, this is a Banach space (GT, X, § 3, No. 1); it is also an ordered vector space, where the relation f≥gf \geq g means «f(x)≥g(x)f(x) \geq g(x) for all x∈Xx \in X».

Definition 1

A positive linear form μ\mu on the space B(X;R)\mathscr{B}(X; \mathbf{R}), where X is a topological space, for which ∥μ∥=1\|\mu\| = 1, is called a mean on X.

\* When X is compact, a mean on X is a positive measure on X such that μ(X)=1\mu(X) = 1. \*

Lemma 1

The set K of means on X is the subset of the unit ball of the dual of the Banach space E=B(X;R)E = \mathscr{B}(X; \mathbf{R}) whose elements are the linear forms μ\mu such that μ(1)=1\mu(1) = 1. It is a subset of E′E' which is convex and compact for σ(E′,E)\sigma(E', E).

Let μ\mu be a linear form on E, such that μ(1)=1\mu(1) = 1. For every function f∈Ef \in E, we define the function f′∈Ef' \in E by f′(x)=∥f∥−f(x)f'(x) = \|f\| - f(x) (x∈Xx \in X). First assume that μ\mu is a mean; for every f∈Ef \in E, we have f′≥0f' \geq 0, hence μ(f′)≥0\mu(f') \geq 0, i.e. μ(f)≤∥f∥\mu(f) \leq \|f\|; therefore ∥μ∥≤1\|\mu\| \leq 1.

Conversely, suppose μ\mu belongs to E′E', and that ∥μ∥≤1\| \mu \| \leq 1; for every positive function f∈Ef \in E, we have μ(f′)≤∥f′∥\mu(f') \leq \| f' \|, hence

∥f∥−μ(f)=μ(f′)≤∥f′∥≤∥f∥, \| f \| - \mu(f) = \mu(f') \leq \| f' \| \leq \| f \|,

and finally μ(f)≥0\mu(f) \geq 0; consequently, μ\mu is a mean.

It is clear that KK is convex; that it is compact for σ(E′,E)\sigma(E', E) follows from cor. 3 of III, p. 17.

Q.E.D.

Let Γ\Gamma be a set of continuous mappings from XX into XX which commute pairwise. Let γ∈Γ\gamma \in \Gamma. For every function f∈Ef \in E, we have f∘γ∈Ef \circ \gamma \in E; hence we can define an affine transformation uγu_{\gamma} on the set KK of means on XX, by

uγμ(f)=μ(f∘γ)(μ∈K,f∈E). u_{\gamma} \mu(f) = \mu(f \circ \gamma) \quad (\mu \in K, f \in E).

If KK is assigned the topology induced by σ(E′,E)\sigma(E', E), the mapping uγu_{\gamma} is continuous. If γ\gamma is a homeomorphism, uγμu_{\gamma} \mu can be deduced from μ\mu by transport of structure. Finally, we have uγuγ′=uγ′γuγu_{\gamma} u_{\gamma'} = u_{\gamma' \gamma} u_{\gamma} for all γ,γ′\gamma, \gamma' in Γ\Gamma. By the Markoff-Kakutani th. (IV, p. 39, th. 1), there exists a mean μ\mu on XX, such that uγμ=μu_{\gamma} \mu = \mu for all γ∈Γ\gamma \in \Gamma; in other words, μ\mu satisfies the relation μ(f)=μ(f∘γ)\mu(f) = \mu(f \circ \gamma) for f∈Ef \in E and γ∈Γ\gamma \in \Gamma.

In an analogous way, the corollary of th. 1 (IV, p. 40) implies the following result:

Proposition 1

Let XX be a topological space and GG a solvable group. We assume that GG operates on XX on the left, in such a way that for all g∈Gg \in G, the mapping x↦g.xx \mapsto g.x from XX into XX is continuous. Then there exists a mean on XX which is invariant under GG.

Corollary

Let GG be a solvable topological group. Then there exists a mean on GG which is invariant under the left and the right translations.

It is enough to apply prop. 1 to the solvable group G×GG \times G acting on GG by (g.g′).x=gxg′−1(g.g').x = {gxg'}^{-1}.

3. Ryll-Nardzewski theorem

In this section, EE denotes a normed space over the field R\mathbf{R} and T\mathscr{T} a Hausdorff locally convex topology on EE for which the norm of EE is lower semi-continuous. These hypotheses are in particular satisfied in the following cases :

a) T\mathscr{T} is the topology induced by the norm of the normed space EE. b) T\mathscr{T} is the weakened topology σ(E,E′)\sigma(E, E') of the normed space EE. c) EE is the dual of a normed space FF and T=σ(F′,F)\mathscr{T} = \sigma(F', F). d) There exist two normed spaces F1F_1 and F2F_2 such that E=L(F1;F2)E = \mathscr{L}(F_1 ; F_2) and T\mathscr{T} is the topology of simple convergence.

Unless otherwise expressly stated, the topological notions refer to the topology T\mathscr{T}.

Let KK be a convex subset of EE. Suppose that KK is compact (for the topology T\mathscr{T}), and that KK satisfies the first axiom of countability for the distance defined by the norm of EE.

Lemma 2

Suppose KK contains at least two points. For every ε>0\varepsilon > 0, there exists a partition of KK into two non-empty subsets K1K_1 and K2K_2, having the following properties :

a) K1K_1 is convex and compact ; b) we have ∥x1−x2∥<ε\| x_1 - x_2 \| < \varepsilon for every x1x_1 and x2x_2 in K2K_2.

Let LL be the closure of the set of all extremal points of KK. By the Krein-Milman th. (II, p. 55, th. 1), KK is the closed convex envelope of LL. Since KK contains at least two points, so does LL. For every x∈Lx \in L, let AxA_x be the set of all y∈Ly \in L such that ∥x−y∥≤ε/4\| x - y \| \leq \varepsilon / 4. By the hypothesis, on KK, there exists a countable subset DD of LL such that L=⋃x∈DAxL = \bigcup_{x \in D} A_x. Since the norm is lower semi-continuous, each of the sets AxA_x is closed. Applying Baire’s th. (GT, IX, § 5, No. 3, th. 1) to the compact space LL, we see that there exists a point aa in DD and an open subset UU in EE such that L∩UL \cap U is non-empty and is contained in AaA_a. Since LL contains at least two points, and since EE is Hausdorff, we can choose UU in such a way that L∉UL \notin U.

Let MM be the closed convex envelope of L∩∁UL \cap \complement U. For every real number tt such that 0<t<10 < t < 1, let MtM_t be the set of all vectors of the form tx1+(1−t)x2tx_1 + (1 - t) x_2 with x1∈Mx_1 \in M and x2∈Kx_2 \in K; this is a non-empty, compact convex subset of KK. We shall prove that Mt≠KM_t \neq K by reductio ad absurdum. Suppose that Mt=KM_t = K; then every extremal point xx in KK belongs to MtM_t, hence can be written in the form x=tx1+(1−t)x2x = tx_1 + (1 - t) x_2 with x1∈Mx_1 \in M and x2∈Kx_2 \in K. This implies that x=x1=x2x = x_1 = x_2, and so x∈Mx \in M. By Krein-Milman th. (II, p. 55, th. 1), we have K=MK = M, and KK is the closed convex envelope of L∩∁UL \cap \complement U. By II, p. 56, corollary, this implies that L⊂L∩∁UL \subset L \cap \complement U, which contradicts the relation L∩U≠∅L \cap U \neq \varnothing.

Put d=sup⁡x∈K,y∈K∥x−y∥d = \sup_{x \in K, y \in K} \| x - y \| and choose a real number tt such that 0<t<10 < t < 1 and t<ε/4dt < \varepsilon / 4d. Put K1=MtK_1 = M_t and K2=K−MtK_2 = K - M_t. By the preceding argument, the sets K1K_1 and K2K_2 are non-empty, and K1K_1 is convex and compact. Let M′M' be the closed convex envelope of L∩UL \cap U. Since KK is the closed convex envelope of the set L=(L∩∁U)∪(L∩U)L = (L \cap \complement U) \cup (L \cap U), it is also the closed convex envelope of M∪M′M \cup M'. Let x1x_1 and x2x_2 be two points in K2K_2; for i=1,2i = 1, 2, there exist yi∈My_i \in M, zi∈M′z_i \in M' and a real number αi\alpha_i such that 0≤αi≤10 \leq \alpha_i \leq 1 and xi=αiyi+(1−αi)zix_i = \alpha_i y_i + (1 - \alpha_i) z_i. If αi≥t\alpha_i \geq t, then

xi=tyi+(1−t){αi−t1−tyi+1−αi1−tzi} x_i = t y_i + (1 - t) \left\{ \frac{\alpha_i - t}{1 - t} y_i + \frac{1 - \alpha_i}{1 - t} z_i \right\}
; this contradicts the assumption that xi∉Mix_i \notin M_i. Hence αi<t\alpha_i < t, for i=1,2i = 1, 2, and so
∥xi−zi∥=∥αi(yi−zi)∥=αi∥yi−zi∥≤αid<dt<ε/4. \| x_i - z_i \| = \| \alpha_i (y_i - z_i) \| = \alpha_i \| y_i - z_i \| \leq \alpha_i d < dt < \varepsilon / 4 .
For every point zz in M′M', we have ∥z−a∥≤ε/4\| z - a \| \leq \varepsilon / 4, since L∩U⊂AaL \cap U \subset A_a, and so, in particular ∥zi−a∥≤ε/4\| z_i - a \| \leq \varepsilon / 4. Thus
∥x1−x2∥≤∑i=12(∥xi−zi∥+∥zi−a∥)<ε. \| x_1 - x_2 \| \leq \sum_{i=1}^2 (\| x_i - z_i \| + \| z_i - a \|) < \varepsilon .
This completes the proof.

Lemma 3

Let G be a group of continuous (for T\mathscr{T}) affine transformations on K. Suppose that K is non-empty and that ∥gx−gy∥=∥x−y∥\|gx - gy\| = \|x - y\| for all x, y in K and all g in G. Then there exists a point in K which is invariant under G.

Let J\mathfrak{J} be the family of non-empty subsets of K which are closed convex and stable for G. If (Lα)α∈I(L_\alpha)_{\alpha \in I} is a family of elements of J\mathfrak{J} which is totally ordered by inclusion, then the set L=⋂α∈ILαL = \bigcap_{\alpha \in I} L_\alpha belongs to J\mathfrak{J}. Consequently (S, III, § 3, No. 4, th. 2), there exists an element L in J\mathfrak{J} which is minimal for the relation of inclusion. We shall prove that L reduces to a point.

We argue by reductio ad absurdum, assuming that L contains at least two distinct points x1x_1 and x2x_2, put x=(x1+x2)/2x = (x_1 + x_2)/2 and ε=∥x1−x2∥/2\varepsilon = \|x_1 - x_2\|/2. The convex and compact set L satisfies the first axiom of countability for the distance defined by the norm (GT, IX, § 2, No. 8). Hence we can apply lemma 2 and find a compact and convex subset L1L_1 of L, distinct from ∅\varnothing and from L, having the following property :

(A) For every y1y_1 and y2y_2 in L−L1L - L_1, we have ∥y1−y2∥<ε\|y_1 - y_2\| < \varepsilon.

We shall prove by reductio ad absurdum that gx∈L1gx \in L_1 for all g∈Gg \in G. Let g∈Gg \in G be such that gx∈L−L1gx \in L - L_1 then we have

∥gxi−gx∥=∥xi−x∥=∥x1−x2∥/2=ε, \|gx_i - gx\| = \|x_i - x\| = \|x_1 - x_2\|/2 = \varepsilon ,

for i=1,2i = 1, 2. By property (A), we have gxi∈L1gx_i \in L_1. Since L1L_1 is convex, we conclude that gx=(gx1+gx2)/2gx = (gx_1 + gx_2)/2 belongs to L1L_1, which contradicts the assumption.

Let L′L' be the closed convex envelope of the orbit GxGx of x. The set L′L' belongs to J\mathfrak{J}. By the preceding argument, we have L′⊂L1L' \subset L_1, hence L′⊂L,L′≠LL' \subset L, L' \neq L. This contradicts the minimal character of L and the proof is complete.

Theorem 2 (Ryll-Nardzewski)

— *Let E be a normed space and K a non-empty convex subset of E, which is compact for the weakened topology σ(E,E′)\sigma(E, E'). Let G be a group of isometric affine transformations of K. Then there exists a point in K which is invariant under G.

For every g∈Gg \in G, let KgK_g denote the set of all points x in K such that gx=xgx = x; let K be assigned the weakened topology; each set KgK_g is convex and closed in the compact space K. We shall prove that the intersection ⋂g∈GKg\bigcap_{g \in G} K_g is non-empty; for this, it is enough to prove that the set Kg1∩...∩KgnK_{g_1} \cap ... \cap K_{g_n} is non-empty for every g1,...,gng_1, ..., g_n in G. Fix g1,...,gng_1, ..., g_n and let H be the subgroup of G generated by {g1,...,gn}\{ g_1, ..., g_n \}. Choose a point a in K and let L denote the closed convex envelope of the orbit Ha of a. Let D be the countable set of elements of the form λ1h1a+⋯+λmhma\lambda_1 h_1 a + \cdots + \lambda_m h_m a, where λ1,...,λm\lambda_1, ..., \lambda_m are positive rational numbers such that λ1+⋯+λm=1\lambda_1 + \cdots + \lambda_m = 1, and h1,...,hmh_1, ..., h_m are elements in H. The closure D‾\overline{D} of D for the strong topology, is convex, hence it is closed for σ(E,E′)\sigma(E, E') (IV, p. 4, prop. 2); therefore D‾=L\overline{D} = L and this proves that L is a metric space satisfying the first axiom of countability for the distance (x,y)↦∥x−y∥(x, y) \mapsto \|x - y\|. We can now apply lemma 2. There exists a point b in L which is invariant under H, hence b∈Kg1∩...∩Kgnb \in K_{g_1} \cap ... \cap K_{g_n}.

Corollary

— Let E be a reflexive Banach space, G a group of automorphisms of the normed space E, and K a subset of E. Suppose that K is non-empty, convex, closed, bounded and stable under G. Then there exists a point in K which is invariant under G.

Since E is reflexive, K is compact for σ(E,E′)\sigma(E, E') (IV, p. 15, th. 1). Moreover, every element of G belongs to L(E)\mathscr{L}(E).

4. Applications.

\* A) Unitary representations of groups : Let E be a complex hilbertian space, G a group and π\pi a unitary representation of G on E, i.e. a homomorphism from G into the group of automorphisms of E. Let EGE^G be the hilbertian subspace of E consisting of all vectors invariant under π(G)\pi(G). For every x∈Ex \in E, let KxK_x be the closed convex envelope of the orbit of x. Fix a point x in E.

We shall show that there exists a unique point in KxK_x which is invariant under π(G)\pi(G), namely the projection of x on EGE^G. By IV, p. 44, corollary (applied to the underlying real vector space to E), there exists a point in KxK_x which is invariant under π(G)\pi(G); let a be such a point; then a∈EGa \in E^G. Let P be the set of all y∈Ey \in E such that y−xy - x is orthogonal to EGE^G; we see immediately that P is closed, convex and invariant under π(G)\pi(G); therefore x∈Px \in P, hence Kx⊂PK_x \subset P and finally a∈Pa \in P. In other words, a−xa - x is orthogonal to EGE^G; consequently a is the projection of x onto EGE^G. \*

\* B) Trace of an operator in a hilbertian space : Suppose that the representation π\pi is irreducible, that is, that there exists no hilbertian subspace of E, distinct from {0}\{0\} and from E, which is invariant under π(G)\pi(G). Let F=L2(E)F = \mathscr{L}^2(E) be the hilbertian space of all Hilbert-Schmidt endomorphisms of E, with the scalar product ⟨u∣v⟩=Tr⁡(u∗v)\langle u|v \rangle = \operatorname{Tr}(u^*v). We define a unitary representation λ\lambda from G into F by the formula

λ(g).u=π(g) uπ(g)−1(u∈F, g∈G) \lambda(g).u = \pi(g)\ u\pi(g)^{-1} \quad (u \in F,\ g \in G)

The space FGF^G of all elements of E invariant under λ(G)\lambda(G) consists of the Hilbert-Schmidt endomorphisms uu of E which commute with π(g)\pi(g) for all g∈Gg \in G. By Schur’s lemma, such a uu is a homothety. Hence we must consider two cases : 1) if E is infinite dimensional, then FG={0}F^G = \{0\}; 2) if E is finite dimensional, then F=L(E)F = \mathscr{L}(E) and FG=C.1EF^G = \mathbf{C}.1_E.

Applying the result of A) to the unitary representation λ\lambda, we obtain the following theorem :

Let u∈L2(E)u \in \mathscr{L}^2(E), and let AuA_u be the closed convex envelope in L2(E)\mathscr{L}^2(E) of the set of endomorphisms π(g) uπ(g)−1\pi(g)\ u\pi(g)^{-1} of E, where g runs through G. If E is infinite dimensional, we have 0∈Au0 \in A_u. If E is finite dimensional with dimension d, there exists a unique homothety in AuA_u, namely the projection 1d Tr⁡(u). 1E\frac{1}{d}\ \operatorname{Tr}(u).\ 1_E of u onto the subspace C.1E\mathbf{C}.1_E of L2(E)\mathscr{L}^2(E). \*

C) Haar measure of a compact group:

Let G be a compact group and let E = C(G,R)\mathscr{C}(G, \mathbf{R}) be the Banach space of all real valued continuous functions on G, endowed with the norm

∥f∥=sup⁡x∈G∣f(x)∣. \| f \| = \sup_{x \in G} |f(x)| .

For all x∈Gx \in G, we define the automorphisms γx\gamma_x and δx\delta_x of E by the formulas

γxf(y)=f(x−1y),δxf(y)=f(yx) \gamma_x f(y) = f(x^{-1}y) , \quad \delta_x f(y) = f(yx)

(for y∈G, f∈Ey \in G,\ f \in E).

Let f∈Ef \in E, let Γf\Gamma_f (resp. Δf\Delta_f) denote the closed convex envelope i. E, of the set of all functions γxf\gamma_x f (resp. δxf\delta_x f) as x ranges over G. We shall prove that there exists a unique constant function μ(f)\mu(f) belonging to Γf\Gamma_f, a unique constant function μ′(f)\mu'(f) belonging to Δf\Delta_f, and that these constants are equal.

It is clear that a continuous function on G is invariant under the automorphisms γx\gamma_x (resp. δx\delta_x) of E if and only if it is constant. Now the set of all functions γxf\gamma_x f (resp. δxf\delta_x f) for x in G, is compact in E, since the mapping x↦γxfx \mapsto \gamma_x f (resp. x↦δxfx \mapsto \delta_x f) from G into E is continuous (GT, X, § 3, No. 4, th. 3). It follows (II, p. 25, prop. 3) that Γf\Gamma_f (resp. Δf\Delta_f) is a compact set in E for the topology defined by the norm, hence for σ(E,E′)\sigma(E, E'). By the Ryll-Nardzewski th. (IV, p. 43, th. 2), there exist constant functions in Γf\Gamma_f and Δf\Delta_f. It remains to prove that if c1∈Γfc_1 \in \Gamma_f and c2∈Δfc_2 \in \Delta_f are constants, then c1=c2c_1 = c_2.

Let ε>0\varepsilon > 0. By the hypothesis there exist points x1,...,xn, y1,...,ynx_1, ..., x_n,\ y_1, ..., y_n in G and positive real numbers λ1,...,λn, μ1,...,μm\lambda_1, ..., \lambda_n,\ \mu_1, ..., \mu_m such that

λ1+⋯+λn=μ1+⋯+μm=1. \lambda_1 + \cdots + \lambda_n = \mu_1 + \cdots + \mu_m = 1 .
sup⁡x∈G∣∑i=1nλif(xix)−c1∣≤ε, \sup_{x \in G} \left| \sum_{i=1}^n \lambda_i f(x_i x) - c_1 \right| \leq \varepsilon ,
sup⁡x∈G∣∑j=1mμjf(xyj)−c2∣≤ε. \sup_{x \in G} \left| \sum_{j=1}^m \mu_j f(xy_j) - c_2 \right| \leq \varepsilon .

Put r=∑i,jλiμjf(xiyj)r = \sum_{i,j} \lambda_i \mu_j f(x_i y_j). Then r−c1=∑j=1mμjajr - c_1 = \sum_{j=1}^m \mu_j a_j with aj=∑i=1nλif(xiyj)−c1a_j = \sum_{i=1}^n \lambda_i f(x_i y_j) - c_1; by (7), we have ∣aj∣≤ε|a_j| \leq \varepsilon for 1≤j≤m1 \leq j \leq m, hence ∣r−c1∣≤ε|r - c_1| \leq \varepsilon. Similarly, we prove the inequality ∣r−c2∣≤ε|r - c_2| \leq \varepsilon, hence ∣c1−c2∣≤2ε|c_1 - c_2| \leq 2\varepsilon. Since ε\varepsilon is arbitrary, we get c1=c2c_1 = c_2, as asserted.

By the definition of μ(f)\mu(f), for every ε>0\varepsilon > 0 we can find positive numbers λ1,...,λn\lambda_1, ..., \lambda_n with sum 1 and elements x1,...,xnx_1, ..., x_n in G such that ∣∑i=1nλif(xix)−μ(f)∣≤ε\left| \sum_{i=1}^n \lambda_i f(x_i x) - \mu(f) \right| \leq \varepsilon for all x∈Gx \in G.

It is immediate that for f,gf, g in E and for every scalar λ\lambda, we have Γf+g⊂Γf+Γg\Gamma_{f+g} \subset \Gamma_f + \Gamma_g and Γγf=λΓf\Gamma_{\gamma f} = \lambda \Gamma_f, hence we have the relations μ(f+g)=μ(f)+μ(g)\mu(f+g) = \mu(f) + \mu(g) and μ(λf)=λμ(f)\mu(\lambda f) = \lambda \mu(f). Therefore, the mapping μ:f↦μ(f)\mu : f \mapsto \mu(f) from E into R\mathbf{R} is a mean on the compact space G (IV, p. 40); in other words μ\mu is a positive measure on G such that μ(G)=1\mu(G) = 1.

It is immediate that μ\mu is invariant under the left translations of GG, and the equality μ(f)=μ′(f)\mu(f) = \mu'(f) implies that μ\mu is also invariant under right translations. \* In other words, μ\mu is a left and a right measure on GG (INT, VII, § 1, No. 2, def. 2). \*

*D) Existence of invariant measures :

Let XX be a Hausdorff topological space, μ\mu a positive bounded measure on XX, and GG a group of homeomorphisms of XX. Suppose that for all g∈Gg \in G, the measure g.μg.\mu, the image of μ\mu under the mapping g:X→Xg : X \to X is of base μ\mu. Let ugu_g be a positive μ\mu-integrable function on XX such that g.μ=ug.μg.\mu = u_g.\mu. Suppose also that there exist two positive μ\mu integrable functions ϕ\phi and ψ\psi on XX, which are not μ\mu-null and are such that ϕ≤ug≤ψ\phi \leq u_g \leq \psi μ\mu-almost everywhere for all g∈Gg \in G. We shall prove that there exists a positive bounded measure ν≠0\nu \neq 0 on XX, with base μ\mu, and invariant under GG.

Let PP be the subset of the Banach space E=L1(X,μ)E = L^1(X, \mu) consisting of classes of functions ff such that ϕ≤f≤ψ\phi \leq f \leq \psi μ\mu-almost everywhere. Then PP is compact for the weakened topology σ(E,E′)\sigma(E, E'). The mapping h↦h.μh \mapsto h.\mu from PP into the Banach space F=Mb(X)F = \mathscr{M}^b(X) of bounded real measures on XX, is a bijection from PP onto a subset P1P_1 of EE which is convex and compact for the topology σ(F,F′)\sigma(F, F'). By hypothesis, g.μ∈P1g.\mu \in P_1 for all g∈Gg \in G. Let KK be the closed convex envelope of the set of all measures g.μg.\mu. For all g∈Gg \in G, the mapping ν↦g.ν\nu \mapsto g.\nu is an isometric affine transformation of KK. By the Ryll-Nardzewski th. (IV, p. 43, th. 2), there exists a measure ν∈K\nu \in K which is invariant under GG. We have ϕ.μ≤ν\phi.\mu \leq \nu, hence ν≠0\nu \neq 0. \*

Exercises

Exercises

See the exercises for Appendix 0.

Transcribed from the printed volume. Not translated.