On possible deterioration of smoothness under the operation of convolution

On possible deterioration of smoothness under the operation of convolution

C. R. Acad. Sci. Paris, t. 333, Série I, p. 291–296, 2001 Analyse mathématique/Mathematical Analysis On possible deterioration of smoothness under th...

103KB Sizes 0 Downloads 18 Views

C. R. Acad. Sci. Paris, t. 333, Série I, p. 291–296, 2001 Analyse mathématique/Mathematical Analysis

On possible deterioration of smoothness under the operation of convolution Alexander IL’INSKII Department of Mathematics, Kharkov National University, 4, Svoboda Sq., 61077, Kharkov, Ukraine E-mail: [email protected] (Reçu le 18 juin 2001, accepté le 25 juin 2001)

Abstract.

Let µ be a finite non-negative Borel measure on the real line R. We give a condition which is necessary and sufficient for the existence of an entire function p satisfying the following conditions: (i) p(x)  0 for x ∈ R, (ii) p ∈ L1 (R), (iii) ess sup{(p ∗ µ)(x) : x ∈ I} = ∞ for every non-empty interval I ⊂ R. We give also a sufficient condition for the existence of an entire function p of finite order ρ > 1 and normal type satisfying conditions (i)–(iii).  2001 Académie des sciences/Éditions scientifiques et médicales Elsevier SAS

Sur la détérioration possible de régularité sous l’opération de convolution Résumé.

Soit µ une mesure borélienne, positive sur R et finie. Nous donnons une condition nécessaire et suffisante pour l’existence de fonctions entières p satisfaisant les conditions suivantes : (i) p(x)  0 pour x ∈ R, (ii) p ∈ L1 (R), (iii) ess sup{(p ∗ µ)(x) : x ∈ I} = ∞ pour chaque intervalle non vide I ⊂ R. Pour chaque ρ > 1, nous donnons aussi une condition suffisante pour l’existence de fonctions entières p d’ordre ρ > 1 et d’un type normal possédant les propriétés (i)–(iii).  2001 Académie des sciences/Éditions scientifiques et médicales Elsevier SAS

Version française abrégée On sait qu’en général, l’opération de la convolution améliore la régularité. Mais comme D.A. Raikov [2] l’a démontré, l’opération de convolution peut détériorer de la régularité. Il a construit des fonctions p1 , p2 , qui sont des restrictions à R de fonctions entières, dont la convolution p1 ∗ p2 n’est pas analytique partout sur R. M. Uluda˘g [3,4] a démontré que la détérioration peut être plus grande que celle de l’exemple de D.A. Raikov (voir aussi [1]). Le résultat suivant est démontré dans [1]. T HÉORÈME A. – Soit t ∈ C1 [0, +∞) une fonction telle que t (r) ↑ +∞ pour r ↑ ∞. Il existe une fonction entière p possédant les propriétés suivantes : (i) 0 < lim supr→∞ M (r, p) exp(−t(r)) < ∞, où M (r, p) = max{|p(z)| : |z| = r} ; (ii) p(x)  0 pour x ∈ R ; Note présentée par Jean-Pierre K AHANE. S0764-4442(01)02058-4/FLA  2001 Académie des sciences/Éditions scientifiques et médicales Elsevier SAS. Tous droits réservés

291

A. Il’inskii

(iii) p ∈ L1 (R) ; (iv) ess sup{(p ∗ p)(x) : x ∈ I} = ∞ pour chaque intervalle non vide I ⊂ R. En particulier, pour tout ρ > 1, il existe une fonction entière p d’ordre ρ et d’un type normal possédant les propriétés (ii)–(iv) du théorème A. Nous nous intéressons ici au problème suivant. Soit µ une mesure borélienne sur R. Quelles conditions garantissent l’existence de fonctions entières (ou entières d’ordre ρ et d’un type normal) p avec les propriétés suivantes : p(x)  0 pour x ∈ R, p ∈ L1 (R) et ess sup{(p ∗ µ)(x) : x ∈ I} = ∞ pour chaque intervalle non vide I ⊂ R ? Dans cette Note, nous indiquons de telles conditions. Introduisons quelques notations : M est l’ensemble de toutes les mesures de probabilités boréliennes sur R ; P est l’ensemble de toutes les fonctions non négatives sur R telles que p ∈ L1 (R) et p 1 = 1 ; Uµ est l’ensemble de toutes les fonctions p ∈ P possédant la propriété suivante : pour chaque intervalle I non vide, l’égalité ess sup{(p ∗ µ)(x) : x ∈ I} = ∞ est valable (µ ∈ M) ; E est l’ensemble de toutes les fonctions entières ; E ρ est l’ensemble de toutes les fonctions entières d’ordre fini ρ et d’un type normal ; D(µ) := {x ∈ R : µ({x}) > 0} ; Mµ (x) := sup{µ(Iε (x))/(2ε) : ε > 0} (où Iε (x) := (x − ε, x + ε), µ ∈ M et x ∈ R) est la fonction maximale de Hardy–Littlewood de la mesure µ. Le résultat essentiel de cette Note est le théorème suivant : T HÉORÈME 1. – Soit µ ∈ M. Alors les implications (2) et (3) sont vraies. Les théorèmes suivants sont des conséquences immédiates du théorème 1. T HÉORÈME 2. – Soit µ ∈ M, alors Uµ ∩ L∞ loc (R) = ∅ si et seulement si la condition (4) est vraie. T HÉORÈME 2 . – Soit µ ∈ M, alors Uµ ∩ E = ∅ si et seulement si la condition (4) est vraie. C OROLLAIRE 1. – Soit µ ∈ M et soit D(µ) non borné, alors Uµ ∩ E = ∅. Le résultat suivant donne une condition suffisante pour l’existence des fonctions entières p ∈ Uµ ∩ E ρ . T HÉORÈME 3. – Soient µ ∈ M, ρ > 1 et supposons qu’il existe une constante B > 1 et deux suites xn → ∞ et ∆n ↓ 0 telles que   lim ∆−1 n µ I∆n (xn ) = +∞ et

n→∞

  −ρ/(ρ−1) = +∞, lim x2n ∆−1 n exp −BCρ ∆n

n→∞

où Cρ := (ρ − 1)ρ−ρ/(ρ−1) . Alors Uµ ∩ E ρ = ∅. Remarque. – Nous rappelons que P ∩ E ρ = ∅ pour ρ < 1. Dans le cas ρ = 1, on peut indiquer une condition suffisante pour l’existence de fonctions p entières d’ordre 1 et d’un type maximal et telles que p ∈ Uµ . C OROLLAIRE 2. – Soit µ ∈ M, ρ > 1. Supposons D(µ) non borné et qu’il existe une suite {xn } telle que xn ∈ D(µ), xn → ∞ (n → ∞) et µ({xn })(log |xn |)(ρ−1)/ρ → ∞ (n → ∞). Alors Uµ ∩ E ρ = ∅.

1. Introduction and statement of results It is well known that for a given function f ∈ L1 (R) with a bounded support the convolution p ∗ f can only be smoother than p. This fact is widely used in analysis when one wants to improve the differential properties of p without substantial changes of its other properties. In 1939 D. Raikov [2] showed that the

292

On possible deterioration of smoothness

convolution p ∗ f may deteriorate the smoothness of p when f has an unbounded support. This effect was also studied later by M. Uluda˘g [3,4] and by the author [1]. The following fact was proved in [1]. T HEOREM A. – Let t ∈ C1 [0, +∞) be any real valued function such that t (r) ↑ +∞ for r ↑ ∞. There exists an entire function p satisfying the following conditions: (i) 0 < lim supr→∞ M (r, p) exp(−t(r)) < ∞, where M (r, p) = max{|p(z)| : |z| = r}; (ii) p(x)  0 for every x ∈ R; (iii) p ∈ L1 (R); (iv) ess sup{(p ∗ p)(x) : x ∈ I} = ∞ for every interval I ⊂ R. In particular, for all ρ > 1, there exists an entire fuction p of order ρ and normal type which satisfies conditions (ii)–(iv) of Theorem A. Let µ be a finite non-negative Borel measure on the real line R. It is reasonable to ask under which conditions we may guarantee the existence of an entire function p (or an entire function p with a prescribed growth on the complex plane C) such that p(x)  0 for all x ∈ R, p ∈ L1 (R), and   ess sup (p ∗ µ)(x) : x ∈ I = ∞

for all non-empty intervals I ⊂ R.

(1)

The aim of this paper is to give such conditions. Without loss of generality we may assume that measures µ and functions p satisfy conditions µ(R) = 1 and p 1 = 1. We shall use the following notation: M is the set of all non-negative finite Borel measures on R; P is the set of all non-negative functions p ∈ L1 (R) with p 1 = 1; Uµ is the set of all functions p ∈ P possessing property (1) (µ ∈ M); E is the set of all entire functions on the complex plane C; E ρ is the set of all entire functions of finite order ρ and normal type; D(µ) := {x ∈ R : µ({x}) > 0}; Mµ (x) := sup{µ(Iε (x))/(2ε) : ε > 0} where Iε (x) := (x − ε, x + ε). Mµ (x) is the maximal Hardy–Littlewood function of a measure µ. Evidently, Mµ (x) = ∞ if µ({x}) > 0, Mµ (x)  m(x) almost everywhere if µ is an absolutely continuous measure with a density m(x). The last inequality is valid for all continuity points x of the density m. Our main result is the following: T HEOREM 1. – Let µ ∈ M. Then the two following implications are valid: U µ ∩ L∞ loc (R) = ∅ =⇒ lim sup Mµ (x) = ∞;

(2)

lim sup Mµ (x) = ∞ =⇒ Uµ ∩ E = ∅.

(3)

x→±∞

x→±∞

As immediate consequences of Theorem 1 we have T HEOREM 2. – If µ ∈ M, then Uµ ∩ L∞ loc (R) = ∅ if and only if lim sup Mµ (x) = ∞.

(4)

x→±∞

T HEOREM 2 . – Condition (4) is necessary and sufficient for Uµ ∩ E = ∅. We note also the following C OROLLARY 1. – Let µ ∈ M and the set D(µ) be unbounded. Then Uµ ∩ E = ∅. The following theorem gives a condition which is sufficient for the existence of a function p possessing the following properties: p is an entire function of finite order ρ > 1 and normal type, and p ∈ Uµ .

293

A. Il’inskii

T HEOREM 3. – Let µ ∈ M and ρ > 1. Assume that there exist a constant B > 1, a real sequence xn → ∞, and a positive sequence ∆n → 0 such that   lim ∆−1 n µ I∆n (xn ) = +∞ and

n→∞

  −ρ/(ρ−1) = +∞, lim x2n ∆−1 n exp −BCρ ∆n

n→∞

where Cρ := (ρ − 1)ρ−ρ/(ρ−1) . Then Uµ ∩ E ρ = ∅. Remark. – As is well known, P ∩ E ρ = ∅ if ρ < 1. On the other hand, it may be stated a condition which is sufficient for the existence of an entire function p of order 1 and maximal type such that p ∈ Uµ . The following corollary can be regarded as an analogue of Corollary 1. C OROLLARY 2. – Let µ ∈ M, ρ > 1, and the set D(µ) be unbounded. If there exists a sequence xn ∈ D(µ), xn → ∞, such that   (ρ−1)/ρ µ {xn } log |xn | → ∞ (n → ∞), then Uµ ∩ E ρ = ∅. 2. Proof of Theorem 1 Necessity. – Assume that condition (4) is not satisfied. Then there exist numbers A > 1 and B > 1 such that Mµ (x)  B for all |x|  A. By the definition of Mµ we have µ((x − ε, x + ε))  B · 2ε for |x|  A and any ε > 0. Therefore µ(E)  B mes(E) for every Borel set E ⊂ R \ (−A, A), where mes(·) denotes the Lebesgue measure. Let p ∈ P ∩ L∞ loc (R) and let I ⊂ R be any interval on the real line. If we denote B  := ess sup{p(x − y) : x ∈ I, |y| < A}, then B  < ∞ because p ∈ L∞ loc (R). It follows that    (p ∗ µ)(x) = + p(x − y)µ(dy) R\(−A,A)

(−A,A)

   B · µ (−A, A) + B · 



p(x − y) dy  B  + B R\(−A,A)

for every x ∈ I. Thus, p ∈ / Uµ . Sufficiency. – Without loss of generality we may assume that lim supx→−∞ Mµ (x) = ∞. It follows that there exists a sequence {xk }∞ k=1 such that 0 > xk ↓ −∞ (k → ∞), 1  hk := Mµ (xk ) → +∞ (k → ∞). To avoid the infinite values of hk , we introduce h∗k

 :=

hk , hk < +∞, k, hk = +∞.

Obviously, 1  h∗k < ∞ for any k, h∗k → +∞ for k → ∞, M (xk )  h∗k for any k. We fix a sequence µ∞ ∞ {an }n=1 satisfying two conditions: an > 0 (n = 1, 2, . . .) and n=1 an < ∞. It follows from the property h∗k → ∞ (k → ∞) that there exists a sequence of indexes k(n) ↑ ∞ (n → ∞) such that an h∗k(n)  1

for all n = 1, 2, . . . .

(5)

Set yn := xk(n) , gn := h∗k(n) . Let ∆n be any positive number such that   µ I∆n (yn )  gn ∆n .

294

(6)

On possible deterioration of smoothness

We note that ∆n  1 for all n because µ is a probability measure and gn  1. Let us denote η(z) := π −1/2 exp(−z 2 ) (z ∈ C). It is clear that η ∈ P ∩ E and there exists ε0 > 0 such that η(x)  ε0 for −1  x  1. Let {sj }∞ j=1 be a countable set dense in R. For every j = 1, 2, . . . , we construct a function pj ∈ P ∩ E such that (pj ∗ µ)(sj ) = ∞.

(7)

Let us pj (z) := cj



  −1 an(m) ∆−1 n(m) η (z + yn(m) − sj )∆n(m) ,

(8)

m=1

∞ where c−1 j = m=1 an(m) and n(m) ↑ ∞ is a sequence of indexes depending on j. It is obvious that pj ∈ P for any choice of the sequence n(m). We shall define n(m) more exactly later. We prove now that (7) is valid for any choice of {n(m)}. Indeed,  ∞ ∞

  −1 cj an(m) ∆−1 (pj ∗ µ)(sj ) = n(m) η (−t + yn(m) )∆n(m) µ(dt) −∞

 cj



m=1

an(m) ∆−1 n(m)

m=1 ∞

 cj ε 0



yn(m) +∆n(m)

yn(m) −∆n(m)

  η (yn(m) − t)∆−1 n(m) µ(dt)



  an(m) ∆−1 µ I (y )  c ε an(m) gn(m) = ∞ ∆ j 0 n(m) n(m) n(m)

m=1

m=1

according to (5). Let us prove that pj is an entire function if the sequence n(m) tends to infinity sufficiently rapidly. First we observe that there exists a constant v0 such that   2 −1   ∆ η (x + iy)∆−1  π −1/2 exp − 3x 8

(9)

for all |x|  v0 , |y|  |x|/2, and 0 < ∆  1. Indeed, |η(x + iy)|  π −1/2 exp(−3x2 /4) if x, y ∈ R, |y|  |x|/2. Therefore |∆−1 η((x + iy)∆−1 )|  π −1/2 ∆−1 exp(−3x2 /(4∆2 )) for |y|  |x|/2 and ∆ > 0. We see that (9) follows from the inequality ∆2 log ∆−1  (3/8)(2 − ∆2 )x2 . Since 0 < ∆  1, the last inequality is a consequence of the following inequality (8/3)∆2 log ∆−1  x2 . Thus the existence of v0 is proved. It follows from the condition yn ↓ −∞ (n → ∞) and from (9) that we can take the sequence n(m) −1 η((z + y − s )∆ )|  1 for |z| = x2 + y 2  m and for all m = 1, 2, . . . . We see so that |∆−1 j n(m) n(m) n(m) from this and condition an < ∞ that the series (8) converges uniformly on each disk of C. Therefore, pj ∈ E. We now give a construction of p. Let us fix a sequence {γj }∞ j=1 such that γj > 0 (j = 1, 2, . . .), ∞ −j γ = 1, γ max{|p (z)| : |z|  j}  2 for j  2 and define j j j j=1 p(x) :=



γj pj (x).

j=1

Evidently, p ∈ P ∩ E. Taking into account (7), we get p ∗ µ(sj ) = ∞ for all j. Let us prove that p ∈ Uµ . The

nfunction (p ∗ µ)(x) is a pointwise limit of a non-decreasing sequence of continuous functions fn (x) := −n p(x − y) µ(dy). Therefore the set IM := {x ∈ R : (p ∗ µ)(x) > M } ∩ I

295

A. Il’inskii

is open for any M > 0 and any non-empty interval I. Since all points sj from I belong to IM , we obtain IM = ∅. Therefore mes (IM ) > 0. ✷ Acknowledgement. The author is grateful to I.V. Ostrovskii for several helpful discussions.

References [1] Il’inskii A., On convolution of entire probability densities, Complex Variables 36 (1998) 165–181. [2] Raikov D.A., On composition of analytic distribution functions, Doklady Akad. Nauk SSSR 23 (1939) 511–514 (in Russian). [3] Uluda˘g M., On possible deterioration of smoothness under the operation of convolution, C. R. Acad. Sci. Paris, Série I 322 (1996) 173–178. [4] Uluda˘g M., On possible deterioration of smoothness under the operation of convolution, J. Math. Anal. Appl. 227 (1998) 335–358.

296