§ 1. Convolution
1. Definition and examples
Recall (Ch. V, §6, Nos. 1 and 4; Ch. VI, §2, No. 10) that, if X and Y are locally compact spaces, μ a measure on X, and φ a mapping of X into Y, φ is said to be μ-proper if: a) φ is μ-measurable; b) for every compact subset K of Y, φ−1(K) is essentially μ-integrable. Then the image measure ν=φ(μ) on Y exists and has the following property: for a function f on Y, with values in a Banach space or in R, to be essentially integrable for ν, it is necessary and sufficient that f∘φ be so for μ, in which case,
∫Yf(y)dν(y)=∫Xf(φ(x))dμ(x).
Definition 1
Let X1,…,Xn be locally compact spaces, μi a measure on Xi (1≤i≤n); let X be the product of the Xi, μ that of the μi. Let φ be a mapping of X into a locally compact space Y. One says that the sequence (μi) is φ-convolvable, or that μ1,…,μn are φ-convolvable, if φ is μ-proper; in this case, the image ν=φ(μ) of μ under φ is called the convolution product of the μi for φ, and is denoted ∗φ(μi)1≤i≤n, or ∗i=1nμi, or μ1∗μ2∗⋯∗μn.
The last two notations are of course used only when there can be no doubt as to φ.
Let f be a function on Y, with values in a Banach space or in R. In order that f be essentially integrable for μ1∗⋯∗μn, it is necessary and sufficient that the function
(x1,…,xn)↦f(φ(x1,…,xn))
be essentially integrable for μ1⊗μ2⊗⋯⊗μn, in which case
(1)∫f d(μ1∗⋯∗μn)=∫f(φ(x1,…,xn))dμ1(x1)…dμn(xn),
a formula that may be regarded as defining μ1∗⋯∗μn when one takes f∈K(Y).
The definitions imply at once that the μi are convolvable if and only if the ∣μi∣ are. When this is the case,
∣φ(μ1⊗⋯⊗μn)∣≤φ(∣μ1⊗⋯⊗μn∣)=φ(∣μ1∣⊗⋯⊗∣μn∣) (Ch. VI, §2, No. 10), that is,
(2)∣∗iμi∣≤∗i∣μi∣.
If the μi are convolvable and positive, and if νi is a measure on Xi such that 0≤νi≤μi, then the νi are convolvable and
∗iνi≤∗iμi.
Suppose μ1,μ2,…,μn are convolvable, and that μ1′,μ2,…,μn are convolvable (μ1′ being a measure on X1). By Ch. V, §6, No. 3, Cor. 1 of Prop. 6, μ1+μ1′,μ2,…,μn are convolvable and
(μ1+μ1′)∗μ2∗⋯∗μn=μ1∗μ2∗⋯∗μn+μ1′∗μ2∗⋯∗μn.
Example
— 1) For any φ, the measures εxi, where xi∈Xi for 1≤i≤n, are always convolvable and have convolution product εy, with y=φ(x1,x2,…,xn). Consequently, if each of the μi has finite support, then the μi are convolvable and μ1∗⋯∗μn has finite support. In particular, let M be a monoid1 equipped with a locally compact topology; if one takes φ to be the law of composition in M then the measures on M with finite support form, for convolution, an algebra that is none other than the algebra of the monoid M (over R or over C, according as one considers real or complex measures) (A, III, §2, No. 6).
2) Let M be a monoid equipped with the discrete topology; assume that for each m∈M, there are only finitely many pairs (m′,m′′)∈M×M such that m′m′′=m; this amounts to saying that the law of composition in M is a proper mapping of M×M into M; the measures on M then form an algebra for convolution, an algebra that is none other than the *total
(1) Monoïde, in the sense of Exer. 17 of Ch. VII, §1.
algebra of the monoid M (A, III, §2, No. 10); we note the following two special cases:
a) M=N, the law of composition being addition. To every measure μ on N, let us associate the formal series
S(μ)=n=0∑∞μ({n})tn
in an indeterminate t. Then S(μ∗μ′)=S(μ)S(μ′). An analogous remark holds for formal series in any number of indeterminates.
∗b) M=N∗, the law of composition being multiplication. To every measure μ on N∗, let us associate the formal Dirichlet series
D(μ)=n=1∑∞μ({n})n−s.
Then D(μ∗μ′)=D(μ)D(μ′).
3) Let X,Y,Z be locally compact spaces, φ a continuous mapping of X×Y into Z. If x∈X and μ is a measure on Y, to say that εx and μ are φ-convolvable comes to saying that the mapping φ(x,⋅) of Y into Z is μ-proper. One then has εx∗μ=φ(x,⋅)(μ).
2. Associativity
The following lemma completes Prop. 11 of Ch. V, §8, No. 5:²
Lemma 1
For 1≤i≤n, let Xi,Yi be two locally compact spaces, μi a measure on Xi, and φi a continuous mapping of Xi into Yi. Let X=∏iXi, Y=∏iYi, μ=⨂iμi, and φ the mapping of X into Y that is the product of the φi. If φ is μ-proper and μi=0 for each i, then the φi are μi-proper and φ(μ)=⨂iφi(μi).
We can suppose that the μi are positive and n=2. Let f1∈K+(Y1). Since μ2=0, there exists an f2∈K+(Y2) such that f2∘φ2 is not μ2-negligible. The function (x1,x2)↦f1(φ1(x1))f2(φ2(x2)) is essentially μ-integrable and continuous, hence μ-integrable. Therefore there exists an x2∈X2 such that f2(φ2(x2))=0 and such that the function x1↦f1(φ1(x1))f2(φ2(x2)) is μ1-integrable. Therefore f1∘φ1 is μ1-integrable,
²The lemma follows by induction on part b) of the cited Prop. 11, which is the case n=2; the corresponding result in the first edition of Ch. V (§8, No. 3, Prop. 7) did not include the result of part b).
which proves that φ1 is μ1-proper. One argues similarly for φ2. Then φ(μ)=⨂iφi(μi) by Prop. 11 of Ch. V, §8, No. 5.
The following lemma completes Prop. 4 of Ch. V, §6, No. 3.\footnote{The assertion of this lemma is in fact part b) of the cited Prop. 4, a part that was not included in the first edition of Ch. V.}
Lemma 2
Let T, T', T'' be three locally compact spaces, μ a measure on T, π a μ-measurable mapping of T into T', π′ a continuous mapping of T' into T'', and π′′=π′∘π. If π′′ is μ-proper, then π is μ-proper, π′ is π(μ)-proper, and π′′(μ)=π′(π(μ)).
Let K' be a compact subset of T'. Then K′′=π′(K′) is compact, therefore π′′(K′′) is essentially μ-integrable, therefore π−1(K′)⊂π′′(K′′) is essentially μ-integrable, thus π is μ-proper. Then π′ is π(μ)-proper and π′′(μ)=π′(π(μ)) by Ch. V, §6, No. 3, Prop. 4.
Proposition 1
Let Xij (1≤i≤m,1≤j≤ni), Yi (1≤i≤m), and Z be locally compact spaces; for each i, let φi be a mapping of Xi=∏jXij into Yi; let φ be the product of the φi, mapping X=∏iXi into Y=∏iYi; let ψ be a mapping of Y into Z.
(i) *Let μij be measures given, respectively, on the Xij, such that, for each i, the μij (1≤j≤ni) are φi-convolvable, and such that the measures ∗∣μij∣ are ψ-convolvable; then the μij, for 1≤i≤m,1≤j≤ni, are (ψ∘φ)-convolvable and
∗i,jμij=∗i(∗jμij).
(ii) Assume ψ and the φi continuous, and let μij be measures =0 given, respectively, on the Xij and (ψ∘φ)-convolvable; then, for each i, the μij (1≤j≤ni) are φi-convolvable, the measures ∗j∣μij∣ are ψ-convolvable, and the formula (3) holds.
It suffices to consider the case that all of the measures in question are ≥0.
Let us place ourselves under the hypotheses of (i). The mapping φ is proper for ⨂i,jμij, and
φ(i,j⨂μij)=i⨂φi(j⨂μij)=i⨂(∗jμij)
(Ch. V, §8, No. 5, Prop. 11). The mapping ψ∘φ is proper for ⨂i,jμij, and
(ψ∘φ)(i,j⨂μij)=ψ(i⨂(∗jμij))=∗i(∗jμij) (Ch. V, §6, Prop. 4). Therefore the
μij (
1≤i≤m, 1≤j≤ni) are
(ψ∘φ)-convolvable and formula (3) holds.
Let us place ourselves under the hypotheses of (ii). First of all, Lemma 2 proves that φ is proper for ⨂i,jμij. Lemma 1 then proves that for every i, φi is proper for ⨂jμij, and that
φ(i,j⨂μij)=i⨂(∗jμij). By Lemma 2,
ψ is proper for
⨂i(∗jμij). Whence the proposition.
Corollary
— Let Xi,Xi′ (1≤i≤n), Y,Y′ be locally compact spaces; let φ,φ′ be continuous mappings of X=∏iXi into Y and of X′=∏iXi′ into Y′, respectively; let fi be continuous mappings of Xi into Xi′ (1≤i≤n) and g a continuous mapping of Y into Y′, such that φ′∘f=g∘φ, f being the mapping of X into X′ that is the product of the fi. Let μi be measures given respectively on the Xi, all =0. Then the following two assertions are equivalent: (i) fi is μi-proper for all i, and the measures fi(∣μi∣) are φ′-convolvable; (ii) the μi are φ-convolvable, and g is proper for ∗φ(∣μi∣).
Moreover, when these assertions are verified,
∗φ′(fi(μi))=g(∗φ(μi))=∗g∘φ(μi).
For, let h=φ′∘f=g∘φ. By Prop. 1, the conditions (i) and (ii) are each equivalent to the following condition: (iii) the μi are h-convolvable. When this is so,
∗φ′(fi(μi))=∗h(μi)=g(∗φ(μi)).
3. The case of bounded measures
Proposition 2
— Let X1,…,Xn,Y be locally compact spaces, μi a bounded measure on Xi (1≤i≤n), μ the product of the μi, φ a μ-measurable mapping of ∏i=1nXi into Y. Then the μi are φ-convolvable and
∥∗i=1nμi∥≤i=1∏n∥μi∥. If the
μi are moreover positive, then
∥∗i=1nμi∥=i=1∏n∥μi∥. For,
μi′=∣μi∣ is bounded and
∥μi′∥=∥μi∥ (Ch. III, §1, No. 8, Cor. 1 of Prop. 10). One has
∣μ1⊗⋯⊗μn∣=μ1′⊗⋯⊗μn′ (Ch. III, §4, Nos. 2, 4), therefore
μ1⊗⋯⊗μn is bounded and
∥μ1⊗⋯⊗μn∥=∥μ1∥⋯∥μn∥ (ibid., Prop. 4). Therefore
φ is
μ-proper (Ch. V, §6, No. 1, Remark 1), that is, the
μi are
φ-convolvable. One has
∥∗i=1nμi′∥=∥μ1′⊗⋯⊗μn′∥ (Ch. V, §6, No. 2, Th. 1), consequently
∥∗i=1nμi′∥=∥μ1′∥⋯∥μn′∥. Finally,
∣∗iμi∣≤∗iμi′ (No. 1, formula (2)), therefore
∥∗iμi∥≤∥∗iμi′∥=i=1∏n∥μi∥.
Proposition 3
— Let X1,…,Xn,Y be locally compact spaces, φ a continuous mapping of ∏i=1nXi into Y. Then the mapping
(μ,…,μn)↦∗φ(μi) of
∏i=1nM1(Xi) into
M1(Y) is a continuous multilinear mapping. This follows from Prop. 2 and what has been said in No. 1.
4. Properties concerning supports
Proposition 4
— Let X1,…,Xn,Y be locally compact spaces, μi a measure on Xi (1≤i≤n), Si its support, and φ a continuous mapping of ∏iXi into Y such that the restriction of φ to ∏iSi is proper. Then the μi are φ-convolvable.
For, let K be a compact subset of Y. The support of μ=μ1⊗⋯⊗μn is S=∏iSi (Ch. III, §4, No. 2, Prop. 2). Therefore φ(K)∩(∏iXi−S) is μ-negligible. On the other hand, φ(K)∩S is compact. Therefore φ(K) is μ-integrable.
Proposition 5
Let X1,…,Xn,Y be locally compact spaces, μi a measure on Xi (1≤i≤n), μ the product of the μi, φ a μ-proper mapping of ∏iXi into Y, and Si the support of μi.
a) *The support of ∗iμi is contained in the closure of φ(∏iSi).
b) If φ is continuous and the μi are positive, then the support of ∗iμi is the closure of φ(∏iSi).
Let S=∏iSi be the support of μ. The support of ∗iμi is contained in φ(S) by Ch. V, §6, No. 2, Cor. 3 of Prop. 2. If φ is continuous and the μi are positive, then the support of ∗iμi is φ(S) (loc. cit., Cor. 4 of Prop. 2).
Corollary
If φ is continuous and the μi have compact support, then the μi are convolvable and ∗iμi has compact support.
5. Vectorial expression of the convolution product
Proposition 6
Let X,Y,Z be locally compact spaces, φ a continuous mapping of X×Y into Z, and λ,μ measures on X,Y. For λ and μ to be φ-convolvable, it is necessary and sufficient that the mapping (x,y)↦εφ(x,y)=εx∗εy of X×Y into M(Z) be scalarly (λ⊗μ)-integrable for the topology σ(M(Z),H(Z)), in which case
λ∗μ=∫X×Y(εx∗εy)dλ(x)dμ(y).
To say that λ and μ are φ-convolvable signifies that, for every f∈H(Z), f∘φ is (λ⊗μ)-integrable, that is, for every f∈H(Z) the function (x,y)↦⟨f,εφ(x,y)⟩ is (λ⊗μ)-integrable, that is, again, that the mapping (x,y)↦εφ(x,y) of X×Y into M(Z) is scalarly (λ⊗μ)-integrable for σ(M(Z),K(Z)). If this is the case, then
⟨λ∗μ,f⟩=∫f(φ(x,y))dλ(x)dμ(y)=∫X×Y⟨εφ(x,y),f⟩dλ(x)dμ(y),
whence λ∗μ=∫X×Yεφ(x,y)dλ(x)dμ(y).
Proposition 7
Let X, Y, Z be locally compact spaces, φ a continuous mapping of X×Y into Z, and λ,μ measures on X, Y. Assume that for every x∈X, εx and μ are φ-convolvable. For λ and μ to be φ-convolvable, it is necessary and sufficient that the mapping x↦εx∗∣μ∣ of X into M(Z) be scalarly λ-integrable for the topology σ(M(Z),K(Z)), in which case λ∗μ=∫X(εx∗μ)dλ(x).
Suppose that λ and μ are φ-convolvable. For every f∈K(Z), f∘φ is (∣λ∣⊗∣μ∣)-integrable, therefore the function x↦∫Yf(φ(x,y))d∣μ∣(y)=⟨f,εx∗∣μ∣⟩ (which by hypothesis is defined for all x∈X) is λ-integrable; thus x↦εx∗∣μ∣ is scalarly λ-integrable for σ(M(Z),K(Z)), and
⟨f,λ∗μ⟩=∫Xdλ(x)∫Yf(φ(x,y))dμ(y)=∫X⟨f,εx∗μ⟩dλ(x),
whence λ∗μ=∫X(εx∗μ)dλ(x). Conversely, suppose that the mapping x↦εx∗∣μ∣ of X into M(Z) is scalarly λ-integrable for σ(M(Z),K(Z)). Let f∈K+(Z). Then the function (x,y)↦f(φ(x,y)) is continuous and (Ch. V, §8, No. 3, Prop. 5)
∬∗f(φ(x,y))d∣λ∣(x)d∣μ∣(y)=∫∗d∣λ∣(x)∫∗f(φ(x,y))d∣μ∣(y)=∫∗⟨f,εx∗∣μ∣⟩d∣λ∣(x)<+∞.
Therefore f∘φ is (λ⊗μ)-integrable, so that λ and μ are φ-convolvable.
Exercises
See the exercises for § 1.