Bourbaki

§ 4. Convolution of measures and functions

1. Convolution of a measure and a function

Let X be a locally compact space on which a locally compact group G operates on the left continuously. Let β\beta be a positive measure on X, quasi-invariant under G. Let χ\chi be a function >0> 0 on G×XG \times X, measurable for every measure on G×XG \times X, and such that, for every s∈Gs \in G, χ(s−1,⋅)\chi(s^{-1}, \cdot) is a density of γ(s)β\gamma(s)\beta with respect to β\beta:

(1)γ(s)β=χ(s−1,⋅)⋅β, (1) \qquad \gamma(s)\beta = \chi(s^{-1}, \cdot) \cdot \beta,

which, with the conventions of Ch. VII, §1, No. 1, may be written:

(1′)dβ(sx)=χ(s,x) dβ(x). (1') \qquad d\beta(sx) = \chi(s, x) \, d\beta(x).

These data will remain fixed in Nos. 1, 2, 3 (an exception being made in Remark 2 of No. 2).

Recall (§2, No. 5) that if χ\chi is continuous and β\beta has support XX, then χ\chi is a multiplier.

Let ff be a locally β\beta-integrable complex function on XX, and let μ\mu be a measure on GG. For every s∈Gs \in G, the measure γ(s)(f⋅β)\gamma(s)(f \cdot \beta) has base β\beta since β\beta is quasi-invariant. Therefore, if μ\mu and f⋅βf \cdot \beta are convolvable, then μ∗(f⋅β)\mu * (f \cdot \beta) has base β\beta (§3, No. 2, Prop. 10).

Definition 1

If μ\mu and f⋅βf \cdot \beta are convolvable, μ\mu and ff are said to be convolvable relative to β\beta. Every density of μ∗(f⋅β)\mu * (f \cdot \beta) with respect to β\beta is called a convolution product of μ\mu and ff relative to β\beta and is denoted μ∗βf\mu *_{\beta} f.

One omits β\beta when no confusion is possible. Convolution for several measures on GG and a function on XX is defined in an analogous manner.

The various convolution products of μ\mu and ff are equal locally β\beta-almost everywhere. If β\beta has support XX and if there exists a convolution product of μ\mu and ff that is continuous, then the latter is uniquely determined; it is then called the convolution product of μ\mu and ff relative to β\beta.

Let s∈Gs \in G and let ff be a locally β\beta-integrable complex function on XX. Then εs\varepsilon_s and ff are convolvable, and

εs∗(f⋅β)=γ(s)(f⋅β)=(γ(s)f)⋅(γ(s)β)=(γ(s)f)⋅χ(s−1,⋅)⋅β, \varepsilon_s * (f \cdot \beta) = \gamma(s)(f \cdot \beta) = (\gamma(s)f) \cdot (\gamma(s)\beta) = (\gamma(s)f) \cdot \chi(s^{-1}, \cdot) \cdot \beta,

therefore

(2)(εs∗f)(x)=χ(s−1,x)f(s−1x)=(γχ(s)f)(x) (2) \qquad (\varepsilon_s * f)(x) = \chi(s^{-1}, x) f(s^{-1}x) = (\gamma_{\chi}(s)f)(x)

locally β\beta-almost everywhere.

Lemma 1

Let μ\mu be a measure on GG. Then χ\chi is locally (μ⊗β)(\mu \otimes \beta)-integrable, and the image of μ⊗β\mu \otimes \beta under the homeomorphism (s,x)↦(s,s−1x)(s, x) \mapsto (s, s^{-1}x) of G×XG \times X onto G×XG \times X is χ⋅(μ⊗β)\chi \cdot (\mu \otimes \beta).

We may suppose that μ⩾0\mu \geqslant 0. Let F∈K+(G×X)F \in \mathscr{K}_+(\mathbf{G} \times \mathbf{X}). Then

∬F(s,s−1x)dμ(s)dβ(x)=∫dμ(s)∫F(s,s−1x)dβ(x) \iint F(s, s^{-1}x) d\mu(s) d\beta(x) = \int d\mu(s) \int F(s, s^{-1}x) d\beta(x)
=∫dμ(s)∫F(s,x)d(γ(s−1)β)(x)=∫dμ(s)∫F(s,x)χ(s,x)dβ(x). = \int d\mu(s) \int F(s, x) d(\gamma(s^{-1})\beta)(x) = \int d\mu(s) \int F(s, x)\chi(s, x) d\beta(x).

Now, the function (s,x)↦F(s,x)χ(s,x)(s, x) \mapsto F(s, x)\chi(s, x) has compact support and is (μ⊗β)(\mu \otimes \beta)-measurable. By Ch. V, §8, No. 3, Prop. 7, the preceding equality proves that this function is (μ⊗β)(\mu \otimes \beta)-integrable and that

∬F(s,s−1x)dμ(s)dβ(x)=∬F(s,x)χ(s,x)dμ(s)dβ(x). \iint F(s, s^{-1}x) d\mu(s) d\beta(x) = \iint F(s, x)\chi(s, x) d\mu(s) d\beta(x).
This proves at the same time both assertions of Lemma 1.

Proposition 1

Let μ\mu be a measure on G\mathbf{G}, ff a locally β\beta-integrable complex function on X\mathbf{X}. Suppose that the function s↦f(s−1x)χ(s−1,x)s \mapsto f(s^{-1}x)\chi(s^{-1}, x) is essentially μ\mu-integrable except for a locally β\beta-negligible set of values of xx, and that the function x↦∫∣f(s−1x)∣χ(s−1,x)d∣μ∣(s)x \mapsto \int |f(s^{-1}x)|\chi(s^{-1}, x) d|\mu|(s), defined locally almost everywhere for β\beta, is locally β\beta-integrable. Then μ\mu and ff are convolvable.

We may assume that f⩾0f \geqslant 0 and μ⩾0\mu \geqslant 0. Let h∈K+(X)h \in \mathscr{K}_+(\mathbf{X}). We are to prove that the function (s,x)↦h(sx)(s, x) \mapsto h(sx) is essentially integrable for μ⊗(f⋅β)=(1⊗f)⋅(μ⊗β)\mu \otimes (f \cdot \beta) = (1 \otimes f) \cdot (\mu \otimes \beta) (Ch. V, §8, No. 5, Prop. 10), that is, that ∬∙h(sx)f(x)φK(s)dμ(s)dβ(x)<+∞\iint^\bullet h(sx)f(x)\varphi_K(s) d\mu(s) d\beta(x) < +\infty (Ch. V, §5, No. 3, Prop. 3); it will clearly suffice to prove that there exists an a>0a > 0 such that for every compact subset KK of G\mathbf{G},

∬∙h(sx)f(x)φK(s)dμ(s)dβ(x)⩽a. \iint^\bullet h(sx)f(x)\varphi_K(s) d\mu(s) d\beta(x) \leqslant a.
By Lemma 1,
∬∙h(sx)f(x)φK(s)dμ(s)dβ(x) \iint^\bullet h(sx)f(x)\varphi_K(s) d\mu(s) d\beta(x)
=∬∗h(x)f(s−1x)φK(s)χ(s−1,x)dμ(s)dβ(x). = \iint^* h(x)f(s^{-1}x)\varphi_K(s)\chi(s^{-1}, x) d\mu(s) d\beta(x).
Now, the function (s,x)↦h(x)f(s−1x)φK(s)χ(s−1,x)(s, x) \mapsto h(x)f(s^{-1}x)\varphi_K(s)\chi(s^{-1}, x) is (μ⊗β)(\mu \otimes \beta)-measurable (Lemma 1) and has compact support. The preceding expression is therefore equal (Ch. V, §8, No. 3, Prop. 7) to
∫∗h(x)dβ(x)∫∗f(s−1x)φK(s)χ(s−1,x)dμ(s) \int^* h(x) d\beta(x) \int^* f(s^{-1}x)\varphi_K(s)\chi(s^{-1}, x) d\mu(s)
⩽(sup⁡h)∫∗dβ(x)∫∗f(s−1x)χ(s−1,x)dμ(s), \leqslant (\sup h) \int^* d\beta(x) \int^* f(s^{-1}x)\chi(s^{-1}, x) d\mu(s),

where S denotes the support of h. Whence the proposition.

Proposition 2

— Let μ\mu be a measure on G, f a locally β\beta-integrable complex function on X. Assume that one of the following conditions is satisfied: (i) f and χ\chi are continuous; (ii) G operates properly in X and f is zero on the complement of a countable union of compact sets; (iii) μ\mu is carried by a countable union of compact sets. If μ\mu and f are convolvable, then the function s↦f(s−1x)χ(s−1,x)s \mapsto f(s^{-1}x)\chi(s^{-1}, x) is essentially μ\mu-integrable except for a locally β\beta-negligible set of values of x, and one has, locally almost β\beta-everywhere,

(μ∗βf)(x)=∫Gf(s−1x)χ(s−1,x) dμ(s)=∫G(γχ(s)f)(x) dμ(s). (\mu *_{\beta} f)(x) = \int_G f(s^{-1}x)\chi(s^{-1}, x)\, d\mu(s) = \int_G (\gamma_{\chi}(s)f)(x)\, d\mu(s).

Let h∈H(X)h \in \mathscr{H}(X). Since μ\mu and f are convolvable, the function (s,x)↦h(sx)f(x)(s, x) \mapsto h(sx)f(x) is essentially (μ⊗β)(\mu \otimes \beta)-integrable. By Lemma 1, the function (s,x)↦h(x)f(s−1x)χ(s−1,x)(s, x) \mapsto h(x)f(s^{-1}x)\chi(s^{-1}, x) is essentially (μ⊗β)(\mu \otimes \beta)-integrable. Under hypothesis (i) or (ii) of the statement, one then deduces that this function is (μ⊗β)(\mu \otimes \beta)-integrable; for, in the first case it is continuous and one applies Prop. 3 of Ch. V, §1, No. 1, and in the second case it is zero outside a countable union of compact sets, and one applies Prop. 7, 2) of No. 2, loc. cit. By the Lebesgue–Fubini theorem,

∬h(sx) dμ(s) d(f⋅β)(x)=∬h(x)f(s−1x)χ(s−1,x) dμ(s) dβ(x)=∫h(x) dβ(x)∫f(s−1x)χ(s−1,x) dμ(s), \iint h(sx)\, d\mu(s)\, d(f \cdot \beta)(x) = \iint h(x)f(s^{-1}x)\chi(s^{-1}, x)\, d\mu(s)\, d\beta(x) = \int h(x)\, d\beta(x) \int f(s^{-1}x)\chi(s^{-1}, x)\, d\mu(s),

the function x↦g(x)=∫f(s−1x)χ(s−1,x) dμ(s)x \mapsto g(x) = \int f(s^{-1}x)\chi(s^{-1}, x)\, d\mu(s) being moreover locally β\beta-integrable. One thus sees that

⟨h,μ∗(f⋅β)⟩=⟨h,g⋅β⟩, \langle h, \mu * (f \cdot \beta) \rangle = \langle h, g \cdot \beta \rangle,

whence g=μ∗βfg = \mu *_{\beta} f.

Suppose now that μ\mu is carried by the union S of a sequence of compact sets. The function

(s,x)↦h(x)f(s−1x)χ(s−1,x)φS(s) (s, x) \mapsto h(x)f(s^{-1}x)\chi(s^{-1}, x)\varphi_S(s)

is essentially (μ⊗β)(\mu \otimes \beta)-integrable, and zero outside a countable union of compact sets, hence (μ⊗β)(\mu \otimes \beta)-integrable. Since μ=φS⋅μ\mu = \varphi_S \cdot \mu, the argument concludes as before.

Remark

The hypothesis (iii) of Prop. 2 is satisfied notably when μ\mu is bounded. For, for every n>0n > 0, there then exists a compact subset KnK_n of GG such that

∣μ∣(G−Kn)≤1n |\mu|(G - K_n) \leq \frac{1}{n}
(Ch. IV, §4, No. 7), and μ\mu is carried by the union of the KnK_n. More generally, let ρ\rho be a lower semi-continuous finite function >0> 0 on GG such that ρ(st)≤ρ(s)ρ(t)\rho(st) \leq \rho(s)\rho(t); if μ∈Mρ\mu \in \mathscr{M}^\rho, the hypothesis (iii) is satisfied; for, ρ⋅μ\rho \cdot \mu is bounded, and μ\mu is carried by the same subsets as ρ⋅μ\rho \cdot \mu since, on every compact subset of GG, ρ\rho is bounded below by a constant >0> 0.

2. Examples of convolvable measures and functions

In Props. 3 and 4, C′(G)\mathscr{C}'(G) and M(G)\mathscr{M}(G) are equipped with the topology of compact convergence in C(G)\mathscr{C}(G) and K(G)\mathscr{K}(G), respectively.

Proposition 3

— Assume χ\chi continuous. Let μ∈C′(G)\mu \in \mathscr{C}'(G), f∈C(X)f \in \mathscr{C}(X). Then: (i) μ\mu and ff are convolvable relative to β\beta. (ii) Formula (3) of No. 1 defines for every x∈Xx \in X a convolution product μ∗βf\mu *_{\beta} f that is continuous and is none other than the element γχ(μ)f\gamma_\chi(\mu)f defined by the continuous representation γχ\gamma_\chi of GG in C(X)\mathscr{C}(X); moreover, the mapping (μ,f)↦μ∗βf(\mu, f) \mapsto \mu *_{\beta} f is hypocontinuous relative to the equicontinuous subsets of C′(G)\mathscr{C}'(G) and the compact subsets of C(X)\mathscr{C}(X). (iii) If in addition f∈K(X)f \in \mathscr{K}(X), then the product μ∗βf\mu *_{\beta} f of (ii) belongs to K(X)\mathscr{K}(X) and the mapping (μ,f)↦μ∗βf(\mu, f) \mapsto \mu *_{\beta} f is hypocontinuous relative to the equicontinuous subsets of C′(G)\mathscr{C}'(G) and the compact subsets of K(X)\mathscr{K}(X).

We know that μ\mu and ff are convolvable (§3, No. 2, Prop. 8 (i)). On the other hand, with the notations of §2, we have

γχ(μ)f=∫(γχ(s)f)dμ(s)∈C(X) \gamma_\chi(\mu)f = \int (\gamma_\chi(s)f)d\mu(s) \in \mathscr{C}(X)
since C(X)\mathscr{C}(X) is complete. In particular, for every x∈Xx \in X,
(γχ(μ)f)(x)=∫(γχ(s)f)(x) dμ(s). (\gamma_\chi(\mu)f)(x) = \int (\gamma_\chi(s)f)(x)\,d\mu(s).
This, combined with Prop. 2 (i), and §2, No. 6, proves (ii). Finally, if f∈K(X)f \in \mathscr{K}(X) then μ∗(f⋅β)\mu * (f \cdot \beta) has compact support (§3, No. 2, Prop. 9), therefore μ∗βf∈K(X)\mu *_{\beta} f \in \mathscr{K}(X). For, let us consider the continuous representation UU of GG in the completion K(X)∼\mathscr{K}(X)^{\sim} obtained by extending by continuity the continuous operators γχ(s)\gamma_\chi(s) in K(X)\mathscr{K}(X) (§ 2, No. 1, Remark 3). Let SS be the support of μ\mu. The functions γχ(s)f\gamma_\chi(s)f, for s∈Ss \in S, have their support contained in a fixed compact set KK. The set K(X,K)\mathscr{K}(X, K) is a complete linear subspace of K(X)\mathscr{K}(X). Therefore U(μ)f∈K(X)U(\mu)f \in \mathscr{K}(X). One sees as before that U(μ)f=μ∗βfU(\mu)f = \mu *^\beta f, and (iii) again follows from § 2, No. 6.

Proposition 4

Assume that G operates properly in X and that χ\chi is continuous. Let μ∈M(G)\mu \in \mathscr{M}(G) and f∈K(X)f \in \mathscr{K}(X).

(i) μ\mu and ff are convolvable relative to β\beta.

(ii) Formula (3) of No. 1 defines for every x∈Xx \in X a convolution product μ∗βf\mu *^\beta f that is continuous.

(iii) The mapping (μ,f)↦μ∗βf(\mu, f) \mapsto \mu *^\beta f of M(G)×K(X)\mathscr{M}(G) \times \mathscr{K}(X) into C(X)\mathscr{C}(X) is hypocontinuous relative to the bounded subsets of M(G)\mathscr{M}(G) and the compact subsets of K(X)\mathscr{K}(X) that are contained in some subspace K(X,L)\mathscr{K}(X, L) (where LL is a variable compact subset of XX).

We know that μ\mu and ff are convolvable (§ 3, No. 2, Prop. 8 (ii)), and it is clear that the integrals occurring in (3) exist for every x∈Xx \in X. Let KK and LL be two compact subsets of XX. There exists a compact subset HH of GG such that the relations x∈Kx \in K and s−1x∈Ls^{-1}x \in L imply s∈Hs \in H; let φ∈K+(G)\varphi \in \mathscr{K}_+(G) with φ(s)=1\varphi(s) = 1 for s∈Hs \in H. Then, for f∈K(X,L)f \in \mathscr{K}(X, L) and x∈Kx \in K,

∫f(s−1x)χ(s−1,x) dμ(s)=∫f(s−1x)χ(s−1,x)φ(s) dμ(s)=((φ⋅μ)∗βf)(x). \int f(s^{-1}x)\chi(s^{-1}, x)\, d\mu(s) = \int f(s^{-1}x)\chi(s^{-1}, x)\varphi(s)\, d\mu(s) = ((\varphi \cdot \mu) *^\beta f)(x).

Consequently ∫f(s−1x)χ(s−1,x) dμ(s)\int f(s^{-1}x)\chi(s^{-1}, x)\, d\mu(s) is a continuous function of xx and defines a convolution product μ∗βf∈C(X)\mu *^\beta f \in \mathscr{C}(X). Moreover, the mapping μ↦φ⋅μ\mu \mapsto \varphi \cdot \mu of M(G)\mathscr{M}(G) into C′(G)\mathscr{C}'(G) is continuous for the topologies of compact convergence. Prop. 3 (iii) therefore implies that the mapping (μ,f)↦μ∗βf(\mu, f) \mapsto \mu *^\beta f of M(G)×K(X,L)\mathscr{M}(G) \times \mathscr{K}(X, L) into C(X)\mathscr{C}(X) is, for every compact subset LL of XX, hypocontinuous relative to the compact subsets of K(X,L)\mathscr{K}(X, L). In particular, the mapping (μ,f)↦μ∗βf(\mu, f) \mapsto \mu *^\beta f of M(G)×K(X)\mathscr{M}(G) \times \mathscr{K}(X) into C(X)\mathscr{C}(X) is separately continuous. Since K(X)\mathscr{K}(X) is barreled, this mapping is hypocontinuous relative to the bounded subsets of M(G)\mathscr{M}(G) (TVS, III, § 5, No. 3, Prop. 6).

Remark 1

— Under the hypotheses of Prop. 4, the mapping μ↦μ∗βf\mu \mapsto \mu *^\beta f of M+(G)\mathscr{M}_+(G) into C(X)\mathscr{C}(X) is continuous when M+(G)\mathscr{M}_+(G) is equipped with the vague topology, for every f∈K(X)f \in \mathscr{K}(X). For, let KK be a compact subset of XX, SS the (compact) support of ff; since GG operates properly in XX, the set of s∈Gs \in G for which there exists an x∈Kx \in K with s−1x∈Ss^{-1}x \in S is a compact subset LL of GG (GT, III, § 4, No. 5, Th. 1). Let ε\varepsilon be a number >0> 0, φ\varphi a function in K+(G)\mathscr{K}_+(G) equal to 1 on the compact set LL, μ0\mu_0 an element of M+(G)\mathscr{M}_+(G); the set W0W_0 of measures μ∈M+(G)\mu \in \mathscr{M}_+(G) such that

∣∫φ(s) dμ(s)−∫φ(s) dμ0(s)∣≤ε \left| \int \varphi(s)\, d\mu(s) - \int \varphi(s)\, d\mu_0(s) \right| \leq \varepsilon
is a neighborhood of μ0\mu_0 in M+(G)\mathscr{M}_+(G). On the other hand, the function (s,x)↦f(s−1x)χ(s−1,x)(s, x) \mapsto f(s^{-1}x)\chi(s^{-1}, x) is uniformly continuous on L×KL \times K, therefore there exists a finite number of points xi∈Kx_i \in K (1≤i≤n1 \leq i \leq n) such that for every x∈Kx \in K, there is an ii for which
∣f(s−1x)χ(s−1,x)−f(s−1xi)χ(s−1,xi)∣≤ε |f(s^{-1}x)\chi(s^{-1}, x) - f(s^{-1}x_i)\chi(s^{-1}, x_i)| \leq \varepsilon
for all s∈Ls \in L. Since μ(L)≤∫φ(s) dμ0(s)+ε\mu(L) \leq \int \varphi(s)\, d\mu_0(s) + \varepsilon for all μ∈W0\mu \in W_0, also
∣∫f(s−1x)χ(s−1,x) dμ(s)−∫f(s−1xi)χ(s−1,xi) dμ(s)∣≤ε(∫φ(s) dμ0(s)+ε) \begin{align*} \left| \int f(s^{-1}x)\chi(s^{-1}, x)\, d\mu(s) - \int f(s^{-1}x_i)\chi(s^{-1}, x_i)\, d\mu(s) \right| \\ \leq \varepsilon \left( \int \varphi(s)\, d\mu_0(s) + \varepsilon \right) \end{align*}
for every xx satisfying the preceding inequality and every μ∈W0\mu \in W_0. Now let WW be the neighborhood of μ0\mu_0 in M+(G)\mathscr{M}_+(G) formed by the measures μ∈W0\mu \in W_0 such that
∣∫f(s−1xi)χ(s−1,xi) dμ(s)−∫f(s−1xi)χ(s−1,xi) dμ0(s)∣≤ε \left| \int f(s^{-1}x_i)\chi(s^{-1}, x_i)\, d\mu(s) - \int f(s^{-1}x_i)\chi(s^{-1}, x_i)\, d\mu_0(s) \right| \leq \varepsilon
for 1≤i≤n1 \leq i \leq n. It is clear that for every measure μ∈W\mu \in W and every x∈Kx \in K,
∣∫f(s−1x)χ(s−1,x) dμ(s)−∫f(s−1x)χ(s−1,x) dμ0(s)∣≤ε(2∫φ(s) dμ0(s)+2ε+1), \begin{align*} \left| \int f(s^{-1}x)\chi(s^{-1}, x)\, d\mu(s) - \int f(s^{-1}x)\chi(s^{-1}, x)\, d\mu_0(s) \right| \\ \leq \varepsilon \left( 2 \int \varphi(s)\, d\mu_0(s) + 2\varepsilon + 1 \right), \end{align*}
and since ε\varepsilon is arbitrary, this proves our assertion.

Proposition 5

Assume χ\chi a continuous multiplier and each function χ(s,⋅)\chi(s, \cdot) bounded.* (i) The function s↦ρ(s)=sup⁡x∈Xχ(s−1,x)s \mapsto \rho(s) = \sup_{x \in X} \chi(s^{-1}, x) on GG is lower semi-continuous >0> 0 and satisfies ρ(st)≤ρ(s)ρ(t)\rho(st) \leq \rho(s)\rho(t) for all s,ts, t in GG. (ii) Let μ∈Mρ(G)\mu \in \mathscr{M}^\rho(G) and f∈L∞(X,β)f \in L^\infty(X, \beta).¹ *Then μ\mu and ff are convolvable and μ∗βf\mu *^\beta f is given locally almost everywhere by the formula (3) of No. 1. One has μ∗βf∈L∞(X,β)\mu *^\beta f \in L^\infty(X, \beta), and ∥μ∗βf∥∞≤∥μ∥ρ∥f∥∞\| \mu *^\beta f \|_\infty \leq \| \mu \|_\rho \| f \|_\infty.

¹For a function ff, the expression "f∈L∞(X,β)f \in L^\infty(X, \beta)" is an abuse of notation signifying that, depending on the context, the symbol ff is to be interpreted either as a function or the equivalence class of a function. In particular, the symbol μ∗βf\mu *^\beta f can stand for either a function defined locally β\beta-almost everywhere, or the equivalence class of such a function for the relation of equality locally β\beta-almost everywhere.

(iii) *If, moreover, f∈C∞(X)f \in \mathscr{C}^\infty(X) (resp. K(X)‾\overline{\mathscr{K}(X)}), then formula (3) of No. 1 defines for every xx a convolution product μ∗βf\mu *^\beta f that belongs to C∞(X)\mathscr{C}^\infty(X) (resp. K(X)\mathscr{K}(X)).

(iv) *If f∈K(X)f \in \mathscr{K}(X), then the convolution product μ∗βf\mu *^\beta f defined by (3) is none other than the element γχ(μ)f\gamma_\chi(\mu)f defined by the continuous representation γχ\gamma_\chi of GG in K(X)‾\overline{\mathscr{K}(X)}.

The identity χ(st,x)=χ(s,tx)χ(t,x)\chi(st, x) = \chi(s, tx)\chi(t, x) implies at once that ρ(st)≤ρ(s)ρ(t)\rho(st) \leq \rho(s)\rho(t). On the other hand, ρ\rho is lower semi-continuous, being the upper envelope of continuous functions.

Let μ∈Mρ(G)\mu \in \mathscr{M}^\rho(G). By Prop. 1 of No. 1, μ\mu and 1 are convolvable; Prop. 2 (i) shows that (∣μ∣∗β1)(x)≤∫Gρ(s)d∣μ∣(s)(|\mu| *^\beta 1)(x) \leq \int_G \rho(s) d|\mu|(s) locally β\beta-almost everywhere. Therefore, if ff is β\beta-measurable and ∣f∣≤1|f| \leq 1, then μ\mu and ff are convolvable and N∞(μ∗βf)≤∫ρ(s)d∣μ∣(s)N_\infty(\mu *^\beta f) \leq \int \rho(s) d|\mu|(s). Moreover, μ∗βf\mu *^\beta f is given locally almost everywhere by formula (3) of No. 1, because condition (iii) of Prop. 2 of No. 1 is satisfied. This implies (ii).

Suppose ff continuous and bounded by 1 in absolute value. It is clear that the integrals occurring in (3) exist for all x∈Xx \in X. Let us show that they depend continuously on xx. We can suppose μ≥0\mu \geq 0. Let x0∈Xx_0 \in X and ε>0\varepsilon > 0. Let KK be a compact subset of GG such that ∫G−Kρ(s)dμ(s)≤ε\int_{G - K} \rho(s) d\mu(s) \leq \varepsilon. There exists a neighborhood VV of x0x_0 in XX such that x∈Vx \in V implies

∣f(s−1x)χ(s−1,x)−f(s−1x0)χ(s−1,x0)∣≤ε/μ(K) |f(s^{-1}x)\chi(s^{-1}, x) - f(s^{-1}x_0)\chi(s^{-1}, x_0)| \leq \varepsilon / \mu(K)

for s∈Ks \in K. Then, for x∈Vx \in V,

∣∫f(s−1x)χ(s−1,x)dμ(s)−∫f(s−1x0)χ(s−1,x0)dμ(s)∣ \left| \int f(s^{-1}x)\chi(s^{-1}, x) d\mu(s) - \int f(s^{-1}x_0)\chi(s^{-1}, x_0) d\mu(s) \right|
≤2∫G−Kρ(s)dμ(s)+∫Kεμ(K)dμ(s)≤3ε, \leq 2 \int_{G - K} \rho(s) d\mu(s) + \int_K \frac{\varepsilon}{\mu(K)} d\mu(s) \leq 3\varepsilon,

whence our assertion. Suppose that in addition f∈K(X)‾f \in \overline{\mathscr{K}(X)}. Let HH be a compact subset of XX such that ∣f(y)∣≤ε|f(y)| \leq \varepsilon for y∉Hy \notin H. Let x∉KHx \notin KH. Then s−1x∉Hs^{-1}x \notin H for s∈Ks \in K, therefore

∣∫Gf(s−1x)χ(s−1,x)dμ(s)∣≤∫G−Kρ(s)dμ(s)+∫Kερ(s)dμ(s) \left| \int_G f(s^{-1}x)\chi(s^{-1}, x) d\mu(s) \right| \leq \int_{G - K} \rho(s) d\mu(s) + \int_K \varepsilon \rho(s) d\mu(s)
≤ε(1+∫Gρ(s)dμ(s)), \leq \varepsilon \left( 1 + \int_G \rho(s) d\mu(s) \right),

which completes the proof of (iii).

Finally, if f∈K(X)f \in \mathscr{K}(X) then, since εx∈M1(X)\varepsilon_x \in \mathscr{M}^1(X) for all x∈Xx \in X, we have

(γχ(μ)f)(x)=∫(γχ(s)f)(x)dμ(s), (\gamma_\chi(\mu)f)(x) = \int (\gamma_\chi(s)f)(x) d\mu(s),

thus γχ(μ)f\gamma_\chi(\mu)f is the convolution product μ∗βf\mu *^\beta f defined by (3).

Proposition 6

Assume χ\chi a continuous multiplier and each function χ(s,⋅)\chi(s, \cdot) bounded. Let ρ(s)=sup⁡x∈Xχ(s−1,x)\rho(s) = \sup_{x \in X} \chi(s^{-1}, x). Let pp and qq be two conjugate exponents (1≤p<+∞1 \leq p < +\infty). Let μ∈Mρ1/q(G)\mu \in \mathscr{M}^{\rho^{1/q}}(G) and f∈Lp(X,β)2f \in L^p(X, \beta)^2 Then: (i) μ\mu and ff are convolvable; (ii) the convolution product μ∗βf\mu *^\beta f is given locally β\beta-almost everywhere by the formula (3), and is equal locally β\beta-almost everywhere to a function g∈Lp(X,β)g \in L^p(X, \beta) such that ∥g∥p≤∥μ∥ρ1/q∥f∥p\|g\|_p \leq \|\mu\|_{\rho^{1/q}} \|f\|_p; (iii) gg is equal to the element γχ(μ)f\gamma_\chi(\mu)f defined by the continuous representation γχ\gamma_\chi of GG in Lp(X,β)L^p(X, \beta).

We have

∫∗∥γχ(s)f∥pd∣μ∣(s)≤(∫∗ρ(s)1/qd∣μ∣(s))∥f∥p<+∞ \int^* \|\gamma_\chi(s)f\|_p d|\mu|(s) \leq \left( \int^* \rho(s)^{1/q} d|\mu|(s) \right) \|f\|_p < +\infty
by §2, No. 5, formula (5). On the other hand, the mapping s↦γχ(s)fs \mapsto \gamma_\chi(s)f of GG into Lp(X,β)L^p(X, \beta) is continuous (§2, No. 5, Prop. 9). Therefore this mapping is μ\mu-integrable. Let
g=∫G(γχ(s)f)dμ(s)∈Lp(X,β). g = \int_G (\gamma_\chi(s)f) d\mu(s) \in L^p(X, \beta).
We have ∥g∥p≤(∫ρ1/q(s)d∣μ∣(s))∥f∥p\|g\|_p \leq (\int \rho^{1/q}(s) d|\mu|(s)) \|f\|_p. Applying the preceding remarks to ∣f∣|f|, one sees that the mapping s↦εs∗∣f∣s \mapsto \varepsilon_s * |f| of GG into Lp(X,β)L^p(X, \beta) is μ\mu-integrable, therefore that, for every h∈H(X)h \in \mathscr{H}(X), the mapping s↦⟨h,εs∗(∣f∣⋅β)⟩s \mapsto \langle h, \varepsilon_s * (|f| \cdot \beta) \rangle is μ\mu-integrable. Prop. 7 of §1, No. 5 then proves that μ\mu and f⋅βf \cdot \beta are convolvable. Moreover,
∫Xg(x)h(x)dβ(x)=∫Gdμ(s)∫X(γχ(s)f)(x)h(x)dβ(x)=∫G⟨h,εs∗(f⋅β)⟩dμ(s), \int_X g(x)h(x) d\beta(x) = \int_G d\mu(s) \int_X (\gamma_\chi(s)f)(x)h(x) d\beta(x) = \int_G \langle h, \varepsilon_s * (f \cdot \beta) \rangle d\mu(s),
and this last integral is equal to ⟨h,μ∗(f⋅β)⟩\langle h, \mu * (f \cdot \beta) \rangle by Prop. 7 of §1, No. 5. One therefore sees that gg is a convolution product of μ\mu and ff. This convolution product is given locally β\beta-almost everywhere by (3), by Prop. 2 and the Remark that follows it.

2For a function ff, the expression " f∈Lp(X,β)f \in L^p(X, \beta) " is an abuse of notation signifying that, depending on the context, the symbol ff is to be interpreted either as a function defined β\beta-almost everywhere, or as the equivalence class of such a function for the relation of equality β\beta-almost everywhere. Thus f∈Lpf \in L^p can symbolize either f∈Lpf \in \mathscr{L}^p or f˙∈Lp\dot{f} \in L^p.

By an abuse of notation, it is often one of the functions gg of the statement that is denoted μ∗βf\mu *^\beta f, which permits writing

∥μ∗βf∥p≤∥μ∥ρ1/q∥f∥p. \| \mu *^\beta f \|_p \leq \| \mu \|_{\rho^{1/q}} \| f \|_p .

If XX is countable at infinity, this style of notation is, moreover, entirely justified.

Corollary

Under the hypotheses of Prop. 6, the mapping* (μ,f)↦μ∗βf(\mu, f) \mapsto \mu *^\beta f defines on Lp(X,β)L^p(X, \beta) the structure of a left module over Mρ1/q(G)\mathscr{M}^{\rho^{1/q}}(G) (1≤p≤+∞)(1 \leq p \leq +\infty).

This follows from Props. 5 and 6 and the associativity of the convolution product.

Remark 2

— Let XX be a locally compact space on which a locally compact group GG operates continuously on the right by (x,s)↦xs(x, s) \mapsto xs. Let β\beta be a positive measure on XX. Let χ\chi be a function >0> 0 on G×XG \times X, measurable for every measure on G×XG \times X, such that δ(s)β=χ(s,⋅)⋅β\delta(s)\beta = \chi(s, \cdot) \cdot \beta for every s∈Gs \in G. Let ff be a locally β\beta-integrable function on XX and let μ\mu be a measure on GG. If f⋅βf \cdot \beta and μ\mu are convolvable (for the mapping (x,s)↦xs(x, s) \mapsto xs of X×GX \times G into XX), then (f⋅β)∗μ(f \cdot \beta) * \mu has base β\beta. We then say that ff and μ\mu are convolvable relative to β\beta; every density of (f⋅β)∗μ(f \cdot \beta) * \mu with respect to β\beta is called a convolution product of ff and μ\mu relative to β\beta, and is denoted f∗βμf *^\beta \mu or simply f∗μf * \mu.

Let G0G^0 be the group opposite to GG. By (s,x)↦xs(s, x) \mapsto xs, G0G^0 operates continuously on the left in XX. To say that ff and μ\mu are convolvable in the foregoing sense is equivalent to saying that μ\mu and ff are convolvable for G0G^0 operating on the left in XX; and the convolution products f∗βμf *^\beta \mu are none other than the convolution products μ∗βf\mu *^\beta f for G0G^0 operating on the left in XX. On the other hand, for s∈G0s \in G^0 one has γ(s)β=χ(s−1,⋅)⋅β\gamma(s)\beta = \chi(s^{-1}, \cdot) \cdot \beta. The results of Nos. 1 and 2 may then be translated immediately into results concerning the products f∗βμf *^\beta \mu. In particular:

1) If s∈Gs \in G and ff is locally β\beta-integrable, then ff and εs\varepsilon_s are convolvable and

(f∗εs)(x)=χ(s−1,x)f(xs−1). (f * \varepsilon_s)(x) = \chi(s^{-1}, x)f(xs^{-1}) .

2) If ff and μ\mu are convolvable and if one of the conditions (i), (ii), (iii) of Prop. 2 is fulfilled, then f∗βμf *^\beta \mu is given locally β\beta-almost everywhere by

(f∗βμ)(x)=∫Gf(xs−1)χ(s−1,x) dμ(s). (f *^\beta \mu)(x) = \int_G f(xs^{-1})\chi(s^{-1}, x)\, d\mu(s) .

We leave to the reader the task of translating the other statements. Note that if χ\chi is continuous and β\beta has support XX, then

χ(ts,x)=χ(s,xt)χ(t,x)(x∈X; s,t in G). \chi(ts, x) = \chi(s, xt)\chi(t, x) \quad (x \in X; \ s, t \text{ in } G) .

3. Convolution and transposition

Let us return to the hypotheses and notations of the beginning of No. 1, but let us assume in addition that β\beta is relatively invariant with multiplier χ\chi; χ\chi is therefore a continuous function on GG.

Proposition 7

Let ff be a locally β\beta-integrable function on XX, ν\nu a measure on XX, and μ\mu a measure on GG. Assume that: (i) μ\mu and ff are convolvable and formula (3) of No. 1 defines locally β\beta-almost everywhere a convolution product μ∗βf\mu *_{\beta} f. (ii) χ⋅μˇ\chi \cdot \check{\mu} and ν\nu are convolvable. (iii) The function g(s,x)=f(s−1x)χ(s−1)g(s, x) = f(s^{-1}x)\chi(s^{-1}) is (μ⊗ν)(\mu \otimes \nu)*-integrable. Then ff is essentially integrable for (χ⋅μˇ)∗ν(\chi \cdot \check{\mu}) * \nu, the function μ∗βf\mu *_{\beta} f defined by (3) is ν\nu*-integrable, and

ν(μ∗βf)=((χ⋅μˇ)∗ν)(f). \nu(\mu *_{\beta} f) = ((\chi \cdot \check{\mu}) * \nu)(f).

Since g(s,x)g(s, x) is integrable for μ⊗ν\mu \otimes \nu, the function f(sx)f(sx) is essentially integrable for (χ⋅μˇ)⊗ν(\chi \cdot \check{\mu}) \otimes \nu and ff is essentially integrable for (χ⋅μˇ)∗ν(\chi \cdot \check{\mu}) * \nu. By the Lebesgue–Fubini theorem, μ∗βf=∫g(s,x)dμ(s)\mu *_{\beta} f = \int g(s, x) d\mu(s) is ν\nu-integrable and

ν(μ∗βf)=∬f(s−1x)χ(s−1)dμ(s)dν(x)=∬f(sx)χ(s)dμˇ(s)dν(x)=((χ⋅μˇ)∗ν)(f). \begin{align*} \nu(\mu *_{\beta} f) &= \iint f(s^{-1}x)\chi(s^{-1}) d\mu(s) d\nu(x) \\ &= \iint f(sx)\chi(s) d\check{\mu}(s) d\nu(x) = ((\chi \cdot \check{\mu}) * \nu)(f). \end{align*}

Example

— 1) One can take f∈C(X)f \in \mathscr{C}(X), ν∈C′(X)\nu \in \mathscr{C}'(X) and μ∈C′(G)\mu \in \mathscr{C}'(G) by Prop. 3, and the Cor. of Prop. 5 of §1, No. 4. The formula (7) then means that the endomorphism ν↦(χ⋅μˇ)∗ν\nu \mapsto (\chi \cdot \check{\mu}) * \nu of C′(X)\mathscr{C}'(X) is the transpose of the endomorphism f↦μ∗ff \mapsto \mu * f of C(X)\mathscr{C}(X).

2) One can take f∈K(X)f \in \mathscr{K}(X), ν∈M(X)\nu \in \mathscr{M}(X) and μ∈C′(G)\mu \in \mathscr{C}'(G) by Prop. 3, Prop. 8 of §3, No. 2, and the remark that the support of the continuous function g(s,x)g(s, x) intersects the support of μ⊗ν\mu \otimes \nu in a compact set. The formula (7) then means that the endomorphism ν↦(χ⋅μˇ)∗ν\nu \mapsto (\chi \cdot \check{\mu}) * \nu of M(X)\mathscr{M}(X) is the transpose of the endomorphism f↦μ∗ff \mapsto \mu * f of K(X)\mathscr{K}(X).

3) If GG operates properly on XX, one can take f∈K(X)f \in \mathscr{K}(X), ν∈C′(X)\nu \in \mathscr{C}'(X) and μ∈M(G)\mu \in \mathscr{M}(G) by Prop. 4, Prop. 8 of §3, No. 2, and the same remark as in Example 2.

Proposition 8

Let ff and gg be two locally β\beta-integrable functions on XX and let μ∈M(G)\mu \in \mathscr{M}(G). Assume that:

(i) μ\mu and ff are convolvable and the formula (3) of No. 1 defines locally β\beta-almost everywhere a convolution product μ∗βf\mu *_{\beta} f.

(ii) χ⋅μˇ\chi \cdot \check{\mu} and gg are convolvable and the formula (3) of No. 1 (with μ\mu replaced by χ⋅μˇ\chi \cdot \check{\mu} and ff by gg) defines locally β\beta-almost everywhere a convolution product (χ⋅μˇ)∗βg(\chi \cdot \check{\mu}) *_{\beta} g.

(iii) There exists a function ψ\psi on GG, equal locally μ\mu-almost everywhere to 1, such that the function

h(s,x)=g(x)f(s−1x)χ(s−1)ψ(s) h(s, x) = g(x) f(s^{-1} x) \chi(s^{-1}) \psi(s)

is (μ⊗β)(\mu \otimes \beta)-integrable.

Then the functions g(x)((μ∗βf)(x))g(x)((\mu *_{\beta} f)(x)) and f(x)(((χ⋅μˇ)∗βg)(x))f(x)(((\chi \cdot \check{\mu}) *_{\beta} g)(x)) are essentially β\beta-integrable, and

(8)∫f(x)(((χ⋅μˇ)∗βg)(x))dβ(x)=∫g(x)((μ∗βf)(x))dβ(x). (8) \quad \int f(x)(((\chi \cdot \check{\mu}) *_{\beta} g)(x)) d\beta(x) = \int g(x)((\mu *_{\beta} f)(x)) d\beta(x).

For, by (iii) and the Lebesgue–Fubini theorem, the function

x↦g(x)∫f(s−1x)χ(s−1)ψ(s)dμ(s) x \mapsto g(x) \int f(s^{-1} x) \chi(s^{-1}) \psi(s) d\mu(s)

is β\beta-integrable, and

I=∬f(s−1x)g(x)χ(s−1)ψ(s)dμ(s)dβ(x)=∫g(x)dβ(x)∫f(s−1x)χ(s−1)ψ(s)dμ(s). \begin{align*} \mathrm{I} &= \iint f(s^{-1} x) g(x) \chi(s^{-1}) \psi(s) d\mu(s) d\beta(x) \\ &= \int g(x) d\beta(x) \int f(s^{-1} x) \chi(s^{-1}) \psi(s) d\mu(s). \end{align*}

But ψ⋅μ=μ\psi \cdot \mu = \mu, consequently

∫f(s−1x)χ(s−1)ψ(s)dμ(s)=(μ∗βf)(x) \int f(s^{-1} x) \chi(s^{-1}) \psi(s) d\mu(s) = (\mu *_{\beta} f)(x)

locally β\beta-almost everywhere. This shows that the function

x↦g(x)((μ∗βf)(x)) x \mapsto g(x)((\mu *_{\beta} f)(x))

is essentially β\beta-integrable and that

I=∫g(x)((μ∗βf)(x))dβ(x). \mathrm{I} = \int g(x)((\mu *_{\beta} f)(x)) d\beta(x).

On the other hand, Lemma 1 shows that the function

(s,x)↦g(sx)f(x)χ(s−1)ψ(s)(s, x) \mapsto g(sx)f(x)\chi(s^{-1})\psi(s)

is integrable for (χ⋅μ)⊗β(\chi \cdot \mu) \otimes \beta. Therefore the function (s,x)↦g(s−1x)f(x)ψ(s−1)(s, x) \mapsto g(s^{-1}x)f(x)\psi(s^{-1}) is integrable for μˇ⊗βˇ\check{\mu} \otimes \check{\beta}, and

I=∬g(s−1x)f(x)ψ(s−1) dμˇ(s) dβ(x)=∫f(x) dβ(x)∫g(s−1x)ψ(s−1) dμˇ(s). \begin{align*} \mathrm{I} &= \iint g(s^{-1}x)f(x)\psi(s^{-1})\, d\check{\mu}(s)\, d\beta(x) \\ &= \int f(x)\, d\beta(x) \int g(s^{-1}x)\psi(s^{-1})\, d\check{\mu}(s). \end{align*}

But ψˇ⋅μˇ=μˇ\check{\psi} \cdot \check{\mu} = \check{\mu} and so ∫g(s−1x)ψ(s−1) dμˇ(s)=((χ⋅μˇ)∗βg)(x)\int g(s^{-1}x)\psi(s^{-1})\, d\check{\mu}(s) = ((\chi \cdot \check{\mu}) *^\beta g)(x) locally β\beta-almost everywhere. This shows that the function

x↦f(x)(((χ⋅μˇ)∗βg)(x))x \mapsto f(x)(((\chi \cdot \check{\mu}) *^\beta g)(x))

is essentially β\beta-integrable and that

I=∫f(x)(((χ⋅μˇ)∗βg)(x)) dβ(x).\mathrm{I} = \int f(x)(((\chi \cdot \check{\mu}) *^\beta g)(x))\, d\beta(x).

This proves the proposition.

Example

— 4) One can take f∈C(X)f \in \mathscr{C}(X), g∈K(X)g \in \mathscr{K}(X) and μ∈C′(G)\mu \in \mathscr{C}'(G) (with ψ=1\psi = 1).

5) If G operates properly on X, one can take f∈K(X)f \in \mathscr{K}(X), g∈K(X)g \in \mathscr{K}(X) and μ∈M(G)\mu \in \mathscr{M}(G) (with ψ=1\psi = 1).

6) One can take f∈Lp(X,β)f \in \mathbf{L}^p(X, \beta), g∈Lq(X,β)g \in \mathbf{L}^q(X, \beta) and μ∈Mρ(G)\mu \in \mathscr{M}^\rho(G), where 1≤p<+∞1 \leq p < +\infty, 1p+1q=1\frac{1}{p} + \frac{1}{q} = 1, ρ=χ−1/q\rho = \chi^{-1/q}. The conditions (i) and (ii) are satisfied by Props. 5 and 6. Let us prove (iii). We have seen that μ\mu is carried by a set S that is a countable union of compact sets. Let us take for ψ\psi the characteristic function of S. The function hh is (μ⊗β)(\mu \otimes \beta)-measurable: for, the function (s,x)↦g(x)χ(s−1)ψ(s)(s, x) \mapsto g(x)\chi(s^{-1})\psi(s) is so, as is the function (s,x)↦f(s−1x)(s, x) \mapsto f(s^{-1}x) by Lemma 1. Moreover, gg being zero outside a countable union of β\beta-integrable sets, hh is zero outside a countable union of (μ⊗β)(\mu \otimes \beta)-integrable sets. We then have (Ch. V, §8, No. 3, Prop. 7):

J=∬∗∣g(x)f(s−1x)∣χ(s−1)ψ(s) d∣μ∣(s) dβ(x)=∫∗∣g(x)∣ dβ(x)∫∗∣f(s−1x)∣χ(s−1)ψ(s) d∣μ∣(s). \begin{align*} \mathrm{J} &= \iint^* |g(x)f(s^{-1}x)|\chi(s^{-1})\psi(s)\, d|\mu|(s)\, d\beta(x) \\ &= \int^* |g(x)|\, d\beta(x) \int^* |f(s^{-1}x)|\chi(s^{-1})\psi(s)\, d|\mu|(s). \end{align*}

But since gg (resp. ψ\psi) is zero outside a countable union of integrable sets, the upper integrals of the second member of (9) are equal to the essential upper integrals (Ch. V, §1, No. 2, Prop. 7). Now (Ch. V, §5, No. 3, Prop. 3)

∫∙∣f(s−1x)∣χ(s−1)ψ(s)d∣μ∣(s)=∫∙∣f(s−1x)∣χ(s−1)d∣μ∣(s) \int^\bullet |f(s^{-1}x)| \chi(s^{-1}) \psi(s) d|\mu|(s) = \int^\bullet |f(s^{-1}x)| \chi(s^{-1}) d|\mu|(s)

since μ=ψ⋅μ\mu = \psi \cdot \mu. By Prop. 6, this last integral is finite and is equal to (∣μ∣∗β∣f∣)(x)(|\mu| *^\beta |f|)(x) locally β\beta-almost everywhere. Therefore

J=∫∙∣g(x)∣(∣μ∣∗β∣f∣)(x)dβ(x), J = \int^\bullet |g(x)|(|\mu| *^\beta |f|)(x) d\beta(x),

and J is finite since g∈Lqg \in L^q and ∣μ∣∗β∣f∣∈Lp|\mu| *^\beta |f| \in L^p (Prop. 6). Therefore h is (μ⊗β)(\mu \otimes \beta)-integrable.

The formula (8) then means that the endomorphism g↦(χ⋅μˇ)∗gg \mapsto (\chi \cdot \check{\mu}) * g of Lq(X,β)L^q(X, \beta) is, for μ∈Mρ(G)\mu \in \mathscr{M}^\rho(G), the transpose of the endomorphism f↦μ∗ff \mapsto \mu * f of Lp(X,β)L^p(X, \beta).

4. Convolution of a measure and a function on a group

Let G be a locally compact group. Throughout Nos. 4 and 5, we fix a relatively invariant positive measure β≠0\beta \neq 0 on G; let χ\chi and χ′\chi' be its left and right multipliers (recall that χ′=χΔG\chi' = \chi \Delta_G). If μ\mu is a measure on G and ff is a locally β\beta-integrable function on G, the convolvability of μ\mu and ff and the products μ∗f\mu * f (resp. the convolvability of ff and μ\mu and the products f∗μf * \mu) may be defined by regarding G as operating on itself on the left (resp. on the right) by translations. Let us make explicit in this situation some of the preceding results:

1) Let μ\mu be a measure on G, ff a locally β\beta-integrable complex function on G. Assume verified one of the following conditions: (i) ff is continuous; (ii) ff is zero on the complement of a countable union of compact sets; (iii) μ\mu is carried by a countable union of compact sets. If μ\mu and ff are convolvable then, locally β\beta-almost everywhere,

(μ∗f)(x)=∫Gf(s−1x)χ(s−1)dμ(s). (\mu * f)(x) = \int_G f(s^{-1}x) \chi(s^{-1}) d\mu(s).

If ff and μ\mu are convolvable then, locally β\beta-almost everywhere,

(f∗μ)(x)=∫Gf(xs−1)χ′(s−1)dμ(s). (f * \mu)(x) = \int_G f(xs^{-1}) \chi'(s^{-1}) d\mu(s).

2) Let pp and qq be two conjugate exponents (1≤p≤+∞1 \leq p \leq +\infty). If μ∈Mχ−1/q(G)\mu \in \mathscr{M}^{\chi^{-1/q}}(G) and f∈Lp(G,β)f \in L^p(G, \beta), then μ\mu and ff are convolvable, and μ∗f\mu * f is equal locally β\beta-almost everywhere to a function in Lp(G,β)L^p(G, \beta); one has (with an abuse of notations already noted)

∥μ∗f∥p≤∥μ∥χ−1/q∥f∥p. \| \mu * f \|_p \leq \| \mu \|_{\chi^{-1/q}} \| f \|_p .

If μ∈Mχ′−1/q(G)\mu \in \mathscr{M}^{{\chi'}^{-1/q}}(G) and f∈Lp(G,β)f \in L^p(G, \beta), then ff and μ\mu are convolvable, and f∗μf * \mu is equal locally β\beta-almost everywhere to a function in Lp(G,β)L^p(G, \beta); one has ∥f∗μ∥p≤∥μ∥χ′−1/q∥f∥p\| f * \mu \|_p \leq \| \mu \|_{{\chi'}^{-1/q}} \| f \|_p.

3) The mappings (μ,f)↦μ∗f,(f,μ)↦f∗μ(\mu, f) \mapsto \mu * f , (f, \mu) \mapsto f * \mu define on Lp(G,β)L^p(G, \beta) the structures of a left module over Mχ−1/q(G)\mathscr{M}^{\chi^{-1/q}}(G) and a right module over Mχ′−1/q(G)\mathscr{M}^{{\chi'}^{-1/q}}(G). The two external laws on Lp(G,β)L^p(G, \beta) are permutable by the associativity of convolution.

4) If μ∗f\mu * f is continuous and is given at every point by (10), then

(μ∗f)(e)=∫f(s−1)χ(s−1)dμ(s). (\mu * f)(e) = \int f(s^{-1}) \chi(s^{-1}) d\mu(s) .

If f∗μf * \mu is continuous and is given at every point by (11), then

(f∗μ)(e)=∫f(s−1)χ′(s−1)dμ(s). (f * \mu)(e) = \int f(s^{-1}) \chi'(s^{-1}) d\mu(s) .

5. Convolution of functions on a group

We conserve the notations G,β,χ,χ′G, \beta, \chi, \chi' of No. 4.

Recall that if ff is a complex function on GG, the property of being locally β\beta-integrable is independent of the choice of β\beta. Let L(G)\mathscr{L}(G) be the set of functions having this property. If f∈L(G),g∈L(G)f \in \mathscr{L}(G) , g \in \mathscr{L}(G), the relation

« f⋅βf \cdot \beta and g⋅βg \cdot \beta are convolvable »

is independent of the choice of β\beta (§ 3, No. 1, Prop. 6). We shall then say that ff and gg are convolvable. By No. 1, (f⋅β)∗(g⋅β)(f \cdot \beta) * (g \cdot \beta) is of the form h⋅βh \cdot \beta with h∈L(G)h \in \mathscr{L}(G), hh being determined up to locally β\beta-negligible sets. We shall write h=f∗βgh = f *^{\beta} g and we shall say that hh is a convolution product of ff and gg relative to β\beta. (One omits β\beta when no confusion is possible.) If β\beta is replaced by ψ⋅β\psi \cdot \beta, ψ\psi being a continuous representation of GG in R+∗\mathbf{R}_+^*, hh does not change (§ 3, No. 1, Prop. 6); if β\beta is replaced by aβa \beta (a∈R+∗a \in \mathbf{R}_+^*), then hh is replaced by ahah. The convolution product of several functions on GG is defined in an analogous manner.

If one of the convolutions of ff and gg is continuous, it is uniquely determined since the support of β\beta is GG. It is then called the convolution product of ff and gg relative to β\beta.

It is clear that

f∗ββg=(f⋅β)∗ββg=f∗ββ(g⋅β). f *_{\beta}^{\beta} g = (f \cdot \beta) *_{\beta}^{\beta} g = f *_{\beta}^{\beta} (g \cdot \beta).

Proposition 9

Let f,gf, g be in L(G)\mathscr{L}(G). Assume that the function s↦g(s−1x)f(s)χ(s−1)s \mapsto g(s^{-1}x)f(s)\chi(s^{-1}) is essentially β\beta-integrable except for a locally β\beta-negligible set of values of xx, and that the function

x↦∫∣g(s−1x)f(s)∣χ(s−1) dβ(s), x \mapsto \int |g(s^{-1}x)f(s)|\chi(s^{-1})\, d\beta(s),

defined locally β\beta-almost everywhere, is locally β\beta-integrable. Then ff and gg are convolvable.

This follows from Prop. 1 of No. 1.

Proposition 10

Let f,gf, g be in L(G)\mathscr{L}(G). Assume that one of these two functions is continuous or is zero on the complement of a countable union of compact sets. If ff and gg are convolvable, then the function f∗gf * g is given locally β\beta-almost everywhere by

(f∗g)(x)=∫Gg(s−1x)f(s)χ(s−1) dβ(s)=∫Gf(xs−1)g(s)χ′(s−1) dβ(s). (f * g)(x) = \int_G g(s^{-1}x)f(s)\chi(s^{-1})\, d\beta(s) = \int_G f(xs^{-1})g(s)\chi'(s^{-1})\, d\beta(s).

This follows from Prop. 2 of No. 1, and the remarks in No. 4.

In particular, if f∗gf * g is continuous and is given at every point by (15), then

(f∗g)(e)=∫g(s−1)f(s)χ(s−1) dβ(s)=∫f(s−1)g(s)χ′(s−1) dβ(s). (f * g)(e) = \int g(s^{-1})f(s)\chi(s^{-1})\, d\beta(s) = \int f(s^{-1})g(s)\chi'(s^{-1})\, d\beta(s).

Still more particularly, if β\beta is a left and right Haar measure, and if f∗gf * g and g∗fg * f are continuous and are given at every point by (15) and the analogous formula for g∗fg * f, then

(f∗g)(e)=(g∗f)(e)=∫f(s)g(s−1) dβ(s). (f * g)(e) = (g * f)(e) = \int f(s)g(s^{-1})\, d\beta(s).

Proposition 11

— Let f,gf, g be in L(G)\mathscr{L}(G). Assume that one of the functions f,gf, g is continuous, and that one of the functions f,gf, g has compact support. Then ff and gg are convolvable. The formula (15) defines for all x∈Gx \in G a product f∗gf * g that is continuous. If f∈K(G)f \in \mathscr{K}(G) and g∈K(G)g \in \mathscr{K}(G), then f∗g∈K(G)f * g \in \mathscr{K}(G).

This follows from Props. 3 and 4 of No. 2.

Proposition 12

— Let pp and qq be two conjugate exponents (1≤p≤+∞1 \leq p \leq +\infty). If fχ−1/q∈L1(G,β)f \chi^{-1/q} \in L^1(G, \beta) and g∈Lp(G,β)g \in L^p(G, \beta), then ff and gg are convolvable, f∗gf * g is equal locally β\beta-almost everywhere to a function in Lp(G,β)L^p(G, \beta), and

∥f∗g∥p≤∥fχ−1/q∥1∥g∥p. \|f * g\|_p \leq \|f \chi^{-1/q}\|_1 \|g\|_p.
If f∈Lp(G,β)f \in L^p(G, \beta) and gχ′−1/q∈L1(G,β)g {\chi'}^{-1/q} \in L^1(G, \beta), then ff and gg are convolvable, f∗gf * g is equal locally β\beta-almost everywhere to a function in Lp(G,β)L^p(G, \beta), and
∥f∗g∥p≤∥f∥p∥gχ′−1/q∥1. \|f * g\|_p \leq \|f\|_p \|g {\chi'}^{-1/q}\|_1.
This follows from Props. 5 and 6 of No. 2 and the remarks in No. 4.

Proposition 13

— If fχ−1∈L1(G,β)f \chi^{-1} \in L^1(G, \beta) and g∈K(G)‾g \in \overline{\mathscr{K}(G)}, or if f∈K(G)f \in \mathscr{K}(G) and gχ′−1∈L1(G,β)g {\chi'}^{-1} \in L^1(G, \beta), then ff and gg are convolvable, and (15) defines for every x∈Gx \in G a product f∗gf * g that belongs to K(G)\mathscr{K}(G).

This follows from Prop. 5 of No. 2, and the remarks in No. 4.

Proposition 14

— If fχ−1∈L1(G,β)f \chi^{-1} \in L^1(G, \beta) and g∈L∞(G,β)g \in L^\infty(G, \beta), then the formula (15) defines for every x∈Gx \in G a product f∗gf * g that is bounded and is uniformly continuous for the right uniform structure of GG.

We already know that f∗gf * g belongs to L∞(G,β)L^\infty(G, \beta) (No. 2, Prop. 5); moreover, (f∗g)(x)=∫f(xs−1)g(s) dν(s)(f * g)(x) = \int f(xs^{-1})g(s)\, d\nu(s), on setting ν=χ′−1⋅β\nu = {\chi'}^{-1} \cdot \beta; ν\nu is a right Haar measure. Therefore

∣(f∗g)(x)−(f∗g)(x′)∣≤∥g∥∞∫∣f(xs−1)−f(x′s−1)∣ dν(s) |(f * g)(x) - (f * g)(x')| \leq \|g\|_\infty \int |f(xs^{-1}) - f(x's^{-1})|\, d\nu(s)
=∥g∥∞∫∣(f(s−1)−f(x′x−1s−1)) dν(s) = \|g\|_\infty \int |(f(s^{-1}) - f(x'x^{-1}s^{-1}))\, d\nu(s)
and the latter integral is arbitrarily small provided x′x−1x'x^{-1} is in a suitable neighborhood of ee (§ 2, No. 5, Prop. 8).

Proposition 15

— Let pp and qq be two conjugate exponents (1<p<+∞1 < p < +\infty). Assume that β\beta is left-invariant. Let f∈Lp(G,β)f \in L^p(G, \beta), g∈Lq(G,βˇ)g \in L^q(G, \check{\beta}). Then ff and gg are convolvable. The formula (15) defines, for every x∈Gx \in G, a product f∗gf * g that belongs to K(G)\mathscr{K}(G) and is such that

∥f∗g∥∞≤∥f∥p∥g∥q. \|f * g\|_\infty \leq \|f\|_p \|g\|_q.

For, we have g˙∈Lq(G,β)\dot{g} \in L^q(G, \beta), therefore the function s↦g(s−1x)f(s)s \mapsto g(s^{-1}x)f(s) is β\beta-integrable for every x∈Gx \in G. Moreover,

∫∣g(s−1x)f(s)∣dβ(s)≤(∫∣f(s)∣pdβ(s))1/p(∫∣g(s−1x)∣qdβ(s))1/q=∥f∥p(∫∣g˙(x−1s)∣qdβ(s))1/q=∥f∥p∥g˙∥q, \int |g(s^{-1}x)f(s)| d\beta(s) \leq \left( \int |f(s)|^p d\beta(s) \right)^{1/p} \left( \int |g(s^{-1}x)|^q d\beta(s) \right)^{1/q} = \|f\|_p \left( \int |\dot{g}(x^{-1}s)|^q d\beta(s) \right)^{1/q} = \|f\|_p \|\dot{g}\|_q,

therefore ff and gg are convolvable (Prop. 9). One sees at the same time that (15) defines for every xx a product f∗gf * g such that

∣(f∗g)(x)∣≤∥f∥p∥g˙∥q. |(f * g)(x)| \leq \|f\|_p \|\dot{g}\|_q.

For f,gf, g in K(G)\mathscr{K}(G), we have f∗g∈K(G)f * g \in \mathscr{K}(G) (Prop. 11); therefore, for f∈Lp(G,β)f \in L^p(G, \beta) and g∈Lq(G,β˙)g \in L^q(G, \dot{\beta}), the product f∗gf * g furnished by (15) is the uniform limit of functions in K(G)\mathscr{K}(G), hence belongs to K(G)‾\overline{\mathscr{K}(G)}.

Corollary

— Let f∈L2(G,β)f \in L^2(G, \beta), g∈L2(G,β)g \in L^2(G, \beta). Then ff and g˙\dot{g} are convolvable. One of the convolutions f∗g˙f * \dot{g} belongs to K(G)‾\overline{\mathscr{K}(G)} and its value at ee is ∫Gf(s)g(s)dβ(s)\int_G f(s)g(s) d\beta(s).

It suffices to take p=q=2p = q = 2 in Prop. 15 and to apply (16).

We no longer assume β\beta to be left-invariant. Let ρ\rho be a lower semi-continuous finite function >0> 0 on GG, such that ρ(st)≤ρ(s)ρ(t)\rho(st) \leq \rho(s)\rho(t) for all s,ts, t in GG. We denote by Lρ(G,β)L^\rho(G, \beta) the set of equivalence classes of the complex functions on GG that are integrable for ρ⋅β\rho \cdot \beta. By the mapping f↦f⋅βf \mapsto f \cdot \beta, Lρ(G,β)L^\rho(G, \beta) may be identified with the set of elements of Mρ(G)\mathscr{M}^\rho(G) that have base β\beta (a set that is independent of the choice of β\beta). If one sets

∥f∥ρ=∫G∣f(s)∣ρ(s)dβ(s) \|f\|_\rho = \int_G |f(s)| \rho(s) d\beta(s)

for f∈Lρ(G,β)f \in L^\rho(G, \beta), this identification is compatible with the norms, thus Lρ(G,β)L^\rho(G, \beta) appears as a complete normed subalgebra of Mρ(G)\mathscr{M}^\rho(G). It is even a two-sided ideal of Mρ(G)\mathscr{M}^\rho(G) by Prop. 10 of §3, No. 2. (For ρ=1\rho = 1, one recovers one of the assertions of No. 4.) In particular, L1(G,β)L^1(G, \beta) may be identified with a closed two-sided ideal of M1(G)\mathscr{M}^1(G).

Proposition 16

— Let UU be a continuous representation of GG in a Banach space EE. Set ρ(s)=∥U(s)∥\rho(s) = \|U(s)\| for all s∈Gs \in G. For every f∈Lρ(G,β)f \in L^\rho(G, \beta), set U(f)=U(f⋅β)U(f) = U(f \cdot \beta). Then f↦U(f)f \mapsto U(f) is a linear representation of the algebra Lρ(G,β)L^\rho(G, \beta) in EE, such that ∥U(f)∥≤∥f∥ρ\|U(f)\| \leq \|f\|_\rho.

This follows from §2, No. 6 and §3, No. 3, Prop. 11.

6. Applications

Proposition 17

— Let G be a locally compact group, A a subset of G, measurable and not locally negligible for a Haar measure. Then A⋅A−1A \cdot A^{-1} is a neighborhood of e.

Let β\beta be a left Haar measure. There exists a compact subset K of G such that B=A∩KB = A \cap K is integrable with measure > 0 for β\beta. Let us apply the Cor. of Prop. 15 with f=g=φBf = g = \varphi_B. The function F=φB∗φˇBF = \varphi_B * \check{\varphi}_B is continuous and > 0 at e. Therefore there exists a neighborhood V of e such that F(x)>0F(x) > 0 for x∈Vx \in V. Now,

F(x)=∫φB(s)φB(x−1s)dβ(s)=β(B∩xB). F(x) = \int \varphi_B(s) \varphi_B(x^{-1}s) d\beta(s) = \beta(B \cap xB).

Therefore, for x∈Vx \in V, one has B∩xB≠∅B \cap xB \neq \emptyset, whence x∈B⋅B−1x \in B \cdot B^{-1}. Thus V⊂B⋅B−1⊂A⋅A−1V \subset B \cdot B^{-1} \subset A \cdot A^{-1}.

Corollary 1

— Let H be a subgroup of G measurable for a Haar measure β\beta. Then H is either open or locally β\beta-negligible.

For, H=H⋅H−1H = H \cdot H^{-1}, therefore if H is not locally β\beta-negligible, then H contains a neighborhood of e (Prop. 17) hence is open (GT, III, §2, No. 1, Cor. of Prop. 4).

Corollary 2

— Let L be a subset of G stable for multiplication and whose complement is locally negligible for a Haar measure β\beta. Then L=GL = G.

For, L−1L^{-1} and L∩L−1L \cap L^{-1} have locally β\beta-negligible complements. Now, L∩L−1L \cap L^{-1} is a subgroup, hence is open (Cor. 1) and therefore closed. Therefore G−(L∩L−1)G - (L \cap L^{-1}), which is open and locally β\beta-negligible, is empty. Thus G=L∩L−1G = L \cap L^{-1}.

Proposition 18

— Let G be a locally compact group, Γ\Gamma a set equipped with a multiplication (u,v)↦uv(u, v) \mapsto uv and a Hausdorff topology such that: 1) the topology of Γ\Gamma is invariant under the translations; 2) the restriction of the multiplication to every compact subset of Γ×Γ\Gamma \times \Gamma is continuous.

Let f:G→Γf : G \to \Gamma be a mapping of G into Γ\Gamma such that f(xy)=f(x)f(y)f(xy) = f(x)f(y) for x,yx, y in G, and measurable for a Haar measure β\beta on G. Then ff is continuous.

Set g(x)=f(x−1)g(x) = f(x^{-1}) for x∈Gx \in G. Since ff and gg are β\beta-measurable, there exists a non β\beta-negligible compact subset K of G such that the restrictions of ff and gg to K are continuous. The mapping (x,y)↦f(xy−1)=f(x)g(y)(x, y) \mapsto f(xy^{-1}) = f(x)g(y) of K×KK \times K into Γ\Gamma is continuous because the multiplication of Γ\Gamma is continuous on f(K)×g(K)f(K) \times g(K); now, this mapping may be written as φ∘ψ\varphi \circ \psi, where ψ\psi is the mapping (x,y)↦xy−1(x, y) \mapsto xy^{-1} of K×KK \times K onto K⋅K−1K \cdot K^{-1}, and φ\varphi is the restriction of ff to K⋅K−1K \cdot K^{-1}. Let RR be the equivalence relation defined on K×KK \times K by ψ\psi. The mapping ψ′\psi' of (K×K)/R(K \times K)/R onto K⋅K−1K \cdot K^{-1} deduced from ψ\psi by passage to the quotient is continuous, therefore (K×K)/R(K \times K)/R is Hausdorff and ψ′\psi' is a homeomorphism. Since φ∘ψ\varphi \circ \psi is continuous, one sees that the restriction of ff to K⋅K−1K \cdot K^{-1} is continuous. Now, K⋅K−1K \cdot K^{-1} is a neighborhood of ee (Prop. 17), therefore ff is continuous at ee. For every x0∈Gx_0 \in G, f(x0x)=f(x0)f(x)f(x_0x) = f(x_0)f(x), thus ff is continuous at x0x_0 because the topology of Γ\Gamma is invariant under translations.

Corollary 1

Let GG be a locally compact group, β\beta a Haar measure on GG, EE a Hausdorff barreled locally convex space, UU a linear representation of GG in EE, such that U(s)∈L(E;E)U(s) \in \mathscr{L}(E;E) for all s∈Gs \in G, β\beta-measurable when L(E;E)\mathscr{L}(E;E) is equipped with the topology of pointwise convergence. Then UU is a continuous linear representation.

Let Γ\Gamma be the group of automorphisms of EE, equipped with the topology of pointwise convergence. This topology is Hausdorff and is invariant under translations. Let KK be a compact subset of Γ\Gamma. Then KK is bounded in L(E;E)\mathscr{L}(E;E) equipped with the topology of pointwise convergence, hence is equicontinuous (TVS, III, §4, No. 2, Th. 1); therefore the mapping (u,v)↦v∘u(u,v) \mapsto v \circ u of K×KK \times K into L(E;E)\mathscr{L}(E;E) is continuous (loc. cit., §5, No. 5, Cor. 1 of Prop. 9). Therefore, for every x∈Ex \in E, the mapping s↦U(s)xs \mapsto U(s)x of GG into EE is continuous (Prop. 18). Since EE is barreled, UU is continuous (§2, No. 1, Prop. 1).

Corollary 2

Let GG be a locally compact group, β\beta a Haar measure on GG, EE a separable Banach space, and UU a linear representation of GG in EE such that U(s)∈L(E;E)U(s) \in \mathscr{L}(E;E) for all s∈Gs \in G. Let (am)(a_m) be a total sequence in EE, and let (an′)(a'_n) be a dense sequence in the unit ball B′B' of the dual E′E' of EE, equipped with the weak topology. Assume that the functions s↦⟨U(s)am,an′⟩s \mapsto \langle U(s)a_m, a'_n \rangle on GG are β\beta-measurable. Then UU is a continuous linear representation.

Let us first show that for every z′∈E′z' \in E', the scalar functions

s↦⟨U(s)am,z′⟩ s \mapsto \langle U(s)a_m, z' \rangle

are β\beta-measurable; we may restrict ourselves to the case that ∥z′∥≤1\|z'\| \leq 1, and, since B′B' is metrizable for the weak topology (TVS, III, §3, No. 4, Cor. 2 of Prop. 6), there exists a subsequence (ank′)(a'_{n_k}) of (an′)(a'_n) that converges weakly to z′z'; the function

s↦⟨U(s)am,z′⟩ s \mapsto \langle U(s)a_m, z' \rangle

is thus the limit of a sequence of β\beta-measurable functions, whence our assertion. It follows that the mapping s↦U(s)ams \mapsto U(s)a_m of GG into EE is β\beta-measurable for every mm (Ch. IV, §5, No. 5, Prop. 10). On the other hand, there exists a sequence (bm)(b_m) of elements of EE, linear combinations of the aia_i, that is dense in the unit ball of EE. For every s∈Gs \in G, ∥U(s)∥=sup⁡m∥U(s)bm∥\|U(s)\| = \sup_m \|U(s)b_m\|, therefore s↦∥U(s)∥s \mapsto \|U(s)\| is measurable. Let KK be a compact subset of GG and let ε>0\varepsilon > 0. There exists a compact subset K0K_0 of KK such that β(K−K0)≤ε\beta(K - K_0) \leq \varepsilon and such that the restrictions to K0K_0 of the functions s↦U(s)ams \mapsto U(s)a_m and s↦∥U(s)∥s \mapsto \|U(s)\| are continuous. Then the U(s)U(s) for s∈K0s \in K_0 are equicontinuous, and the topology of pointwise convergence induces on U(K0)U(K_0) the topology of pointwise convergence in the set of ama_m (TVS, III, §3, No. 4, Prop. 5). Consequently the mapping s↦U(s)s \mapsto U(s) of K0K_0 into Ls(E;E)\mathscr{L}_s(E; E) is continuous. It then suffices to apply Cor. 1.

7. Regularization

Proposition 19

— *Let GG be a locally compact group, β\beta a relatively invariant positive measure ≠0\neq 0 on GG, B\mathscr{B} a base for the filter of neighborhoods of ee in GG, consisting of compact neighborhoods. For every V∈BV \in \mathscr{B}, let fVf_V be a continuous function ≥0\geq 0 on GG, with support contained in VV, such that ∫fVdβ=1\int f_V d\beta = 1. If μ\mu is a measure on GG then, in M(G)\mathscr{M}(G) equipped with the topology of compact convergence in H(G)\mathscr{H}(G),

μ=lim⁡V(μ∗fV)⋅β=lim⁡V(fV∗μ)⋅β, \mu = \lim_V (\mu * f_V) \cdot \beta = \lim_V (f_V * \mu) \cdot \beta,

the limit being taken with respect to the section filter of B\mathscr{B}.

For the topology of compact convergence in C(G)\mathscr{C}(G), fV⋅βf_V \cdot \beta tends to εe\varepsilon_e with respect to the section filter of B\mathscr{B} (§ 2, No. 7, Cor. 1 of Lemma 4). Therefore μ=lim⁡Vμ∗(fV⋅β)=lim⁡V(fV⋅β)∗μ\mu = \lim_V \mu * (f_V \cdot \beta) = \lim_V (f_V \cdot \beta) * \mu in M(G)\mathscr{M}(G) equipped with the topology of compact convergence in H(G)\mathscr{H}(G) (§ 3, No. 3, Cor. of Prop. 12).

Remark 1

We thus see that every measure on GG is the limit of measures admitting a continuous density with respect to every Haar measure (for the topology indicated in Prop. 19 and a fortiori for the vague topology).

Remark 2

If GG is metrizable, B\mathscr{B} can be taken to be a sequence (Vn)(V_n) of neighborhoods. Then μ\mu is the limit of the sequence of measures (μ∗fVn)⋅β(\mu * f_{V_n}) \cdot \beta with continuous densities. *If GG is a real Lie group, the fVnf_{V_n} can be taken to be infinitely differentiable; we shall see later on that the densities μ∗fVn\mu * f_{V_n} are then infinitely differentiable.

Proposition 20

— We conserve the hypotheses and notations of Prop. 19. Let p∈[1,+∞[p \in [1, +\infty[ and g∈Lp(G,β)g \in L^p(G, \beta). Then

g=lim⁡Vg∗βfV=lim⁡VfV∗βg g = \lim_V g *^\beta f_V = \lim_V f_V *^\beta g

in the sense of the norm NpN_p, the limit being taken with respect to the section filter of B\mathscr{B}.

It suffices to apply Prop. 6 (iii), and §2, No. 7, Cor. 3 of Lemma 4.

Remark 3

By Prop. 15, the functions g∗fVg * f_V, fV∗gf_V * g belong to K(G)\mathscr{K}(G).

Corollary

— Let WW be a closed linear subspace of L1(G,β)L^1(G, \beta). For WW to be a left (resp. right) ideal of L1(G,β)L^1(G, \beta), it is necessary and sufficient that WW be invariant under the left (resp. right) translations of GG.

Suppose that WW is a left ideal. Let s∈Gs \in G and g∈Wg \in W. We have εs∗g=lim⁡VfV∗(εs∗g)=lim⁡V(fV∗εs)∗g\varepsilon_s * g = \lim_V f_V * (\varepsilon_s * g) = \lim_V (f_V * \varepsilon_s) * g, and (fV∗εs)∗g∈W(f_V * \varepsilon_s) * g \in W, therefore εs∗g∈W\varepsilon_s * g \in W, thus γ(s)g∈W\gamma(s)g \in W. Conversely, if WW is invariant under the left translations, then μ∗βg∈W\mu *^\beta g \in W for μ∈M1(G)\mu \in \mathscr{M}^1(G) and g∈Wg \in W, therefore WW is a fortiori a left ideal of L1(G,β)L^1(G, \beta). One argues similarly for right ideals.

Example

We take G=RG = \mathbf{R}. Let us define a function Fn∈K(R)F_n \in \mathscr{K}(\mathbf{R}) by

Fn(x)=(1−x2)nif x∈[−1,1]Fn(x)=0if x∉[−1,1]. F_n(x) = (1 - x^2)^n \quad \text{if } x \in [-1, 1] \\ F_n(x) = 0 \qquad \text{if } x \notin [-1, 1].
Let An=∫−1+1Fn(x)dxA_n = \int_{-1}^{+1} F_n(x) dx, and Gn=An−1FnG_n = A_n^{-1} F_n. It is immediate that the measures Gn(x)dxG_n(x) dx satisfy the conditions of §2, No. 7, Cor. 1 of Lemma 4. Let μ\mu be a measure on R\mathbf{R} whose support is contained in [−1/2,1/2][-1/2, 1/2]. Then
(μ∗Gn)(x)=∫RGn(x−y)dμ(y)=An−1∫−1/21/2Fn(x−y)dμ(y). (\mu * G_n)(x) = \int_{\mathbf{R}} G_n(x - y) d\mu(y) = A_n^{-1} \int_{-1/2}^{1/2} F_n(x - y) d\mu(y).
If −1/2≤x≤1/2-1/2 \leq x \leq 1/2, then
(μ∗Gn)(x)=An−1∫−1/21/2[1−(x−y)2]ndμ(y), (\mu * G_n)(x) = A_n^{-1} \int_{-1/2}^{1/2} [1 - (x - y)^2]^n d\mu(y),
therefore μ∗Gn\mu * G_n coincides on [−1/2,1/2][-1/2, 1/2] with a polynomial. In particular, if ff is a continuous function with support contained in [−1/2,1/2][-1/2, 1/2], then f∗Gnf * G_n coincides in [−1/2,1/2][-1/2, 1/2] with a polynomial; moreover, by Prop. 5 (iv), and §2, No. 7, Cor. 3 of Lemma 4, f∗Gnf * G_n converges uniformly to ff. *If ff is of class CrC^r, the derivatives Ds(f∗Gn)D^s(f * G_n) tend uniformly to DsfD^s f for 0≤s≤r0 \leq s \leq r.

Exercises

See the exercises for § 4.

Transcribed from the printed volume. Not translated.