PROCEEDINGS OF THE AMERICAN MATHEMATICAL SOCIETY Volume 127, Number 8, Pages 2431{2436 S 0002-9939(99)04788-7 Article electronically published on April 8, 1999 LARGE DEVIATIONS FOR SUMS OF I.I.D. RANDOM COMPACT SETS RAPHAEL¨ CERF (Communicated by Stanley Sawyer) Abstract. We prove a large deviation principle for Minkowski sums of i.i.d. random compact sets in a Banach space, that is, the analog of Cram´er theorem for random compact sets. Several works have been devoted to deriving limit theorems for random sets. For i.i.d. random compact sets in Rp, the law of large numbers was initially proved by Artstein and Vitale [1] and the central limit theorem by Cressie [3], Lyashenko [10] and Weil [16]. For generalizations to noncompact sets, see also Hess [8]. These limit theorems were generalized to the case of random compact sets in a Banach space by Gin´e, Hahn and Zinn [7] and Puri and Ralescu [11]. Our aim is to prove a large deviation principle for Minkowski sums of i.i.d. random compact sets in a Banach space, that is, to prove the analog of the Cram´er theorem. We consider a separable Banach space F with norm . Wedenoteby (F) the collection of all nonempty compact subsets of F . For|| an || element A of (FK), we denote by co A the closed convex hull of A. Mazur’s theorem [5, p. 416] impliesK that, for A in (F ), co A belongs to co (F ), the collection of the nonempty compact convex subsetsK of F .Thespace (FK) is equipped with the Minkowski addition and the scalar multiplication: for A K;A in (F )andλarealnumber, 1 2 K A +A = a +a :a A;a A ,λA=λa : a A : 1 2 { 1 2 1 ∈ 1 2∈ 2} 1 { 1 1 ∈ 1 } The Hausdorff distance d(A1;A2)=max sup inf a1 a2 ; sup inf a2 a1 a A a2 A2 || − || a A a1 A1 || − || 1∈ 1 ∈ 2∈ 2 ∈ makes ( (F );d) a complete separable metric space (i.e., a Polish space). We endow (F ) withK the Borel σ–field associated to the Hausdorff topology. K We denote by F ∗ the topological dual of F and by B∗ the unit ball of F ∗.The Banach–Alaoglu theorem asserts that B∗ endowed with the weak topology w is ∗ ∗ compact [13]. Moreover the space (B∗;w ) is separable and metrizable. We denote ∗ by M (B∗) the set of Borel signed measures on B∗ (the σ–field being the σ–field generated by the weak topology). Let (Ω; ;P) be a probability space. A random compact set of F is a measurable∗ functionF from Ω to (F ), i.e., a random variable with values in (F ). K K Received by the editors September 10, 1997 and, in revised form, October 27, 1997. 1991 Mathematics Subject Classification. Primary 60D05, 60F10. Key words and phrases. Cram´er theorem, random sets, large deviations. c 1999 American Mathematical Society 2431 License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 2432 RAPHAEL¨ CERF We suppose that F is of type p>1, i.e., there exists a constant c such that n n E f p c E f p || i|| ≤ || i|| i=1 i=1 X X for any independent random variables f1;:::;fn with values in F and mean zero. Every Hilbert space is of type 2; the spaces Lp with 1 <p< are of type min(p; 2). However, the space of continuous functions on [0; 1] equipped∞ with the supremum norm is of type 1 and not of type p for any p>1. We denote by N∗ the set of positive integers. Theorem 1. Let ( ) be a sequence of i.i.d. random compact sets of such An n N∗ F that ∈ + α R E(exp(α sup a )) < + : ∀ ∈ a A || || ∞ ∈ 1 For a measure λ of M(B∗) we set Λ(λ)=lnEexp sup x∗(a) dλ(x∗) B a A1 Z ∗ ∈ and for a set U belonging to co (F ), K Λ∗(U)= sup sup x∗(x) dλ(x∗) Λ(λ) : λ M(B ) B x U − ∈ ∗ Z ∗ ∈ For a nonconvex set U in (F ) we set Λ∗(U)=+ . Then the law of the random set S =(A + +A )=nK satisfies a large deviation∞ principle with rate function n 1 ··· n Λ∗, i.e., for any subset of (F ) U K 1 inf Λ∗(U) lim inf ln P (Sn ) −U interior( ) ≤ n n ∈U ∈ U →∞ 1 lim sup ln P (Sn ) inf Λ∗(U) ≤ n n ∈U ≤−U closure( ) →∞ ∈ U (where interior( ) and closure( ) are the interior and the closure of with respect to the HausdorffU topology). U U Remark. The rate function Λ∗ is a “good” rate function, i.e., it is lower semicon- + tinuous and its level sets U (F):Λ∗(U) λ ,λ R , are compact. { ∈K ≤ } ∈ We first collect several results which are the main ingredients to prove Theorem 1. An embedding theorem. To a compact convex subset A of F we associate its support function sA : B∗ R defined by → x∗ B∗ s (x∗)=sup x∗(x):x A : ∀ ∈ A { ∈ } We denote by (B∗;w ) the set of continuous functions from B∗ endowed with the C ∗ weak topology to R. With the uniform norm , (B∗;w ) is a separable ∗ || ||∞ C ∗ Banach space (for f in (B∗;w ), f =supx B f(x∗)). Whenever A is C ∗ || ||∞ ∗∈ ∗| | compact, its support function sA belongs to (B∗;w ). The map s :co (F) C ∗ K + → (B∗;w ) has the following properties. For any A1;A2 in co (F )andtin R , C ∗ K s = s A = A ;A A s s ; A1 A2 ⇐⇒ 1 2 1⊂2⇐⇒ A 1 ≤ A 2 s A 1 + A 2 = s A 1 + s A 2 ;stA1 = tsA1 License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use LARGE DEVIATIONS FOR SUMS OF I.I.D. RANDOM COMPACT SETS 2433 and finally d(A1;A2)= sA1 sA2 . Hence, co (F ) is algebraically and topo- logically isomorphic to its|| image− under||∞ s, s(co K(F )), which is a subset of the K separable Banach space (B∗;w ). This embedding theorem was used in [1] and [7] to prove limit theoremsC for random∗ sets. In the context of normed spaces, this theorem is due to R˚adstr¨om [12] and H¨ormander [9]. A general Cram´er theorem. We state here a slightly weakened version of the general Cram´er theorem (see either [4, theorem 3:1:6 and corollary 3:1:7] or [6, theorem 6:1:3]). Let E be a separable Banach space and E1 a closed convex subset of E.Let (Xn)n be a sequence of i.i.d. random variables defined on (Ω; ;P) with values ∈N∗ F in E and set S =(X + +X )=n. Suppose that for any λ in the dual E∗ 1 n 1 ··· n Λ (λ)=lnEexp(λ(X )) < + E 1 ∞ and that the laws of (Sn)n are exponentially tight, i.e., for any positive L there ∈N∗ exists a compact subset KL of E1 such that lim supn (1=n)lnP(Sn KL) L. →∞ 6∈ ≤− Then the law of Sn satisfies a large deviation principle with rate function ΛE∗ (x)= sup λ(x) ΛE(λ); λ E − ∈ ∗ i.e., for any subset U of E 1 inf ΛE∗ (x) lim inf ln P (Sn U) −x interior(U) ≤ n n ∈ ∈ →∞ 1 lim sup ln P (Sn U) inf ΛE∗ (x): ≤ n n ∈ ≤−x closure(U) →∞ ∈ Moreover, the condition + α R E(exp(α X1 )) < + ∀ ∈ || || ∞ automatically ensures that the laws of ( ) are exponentially tight and that Λ Sn n N∗ E∗ is a “good” rate function (it is lower semicontinuous∈ and its level sets are compact; see either [4, theorem 3:3:11] or [6, exercise 6:2:21]). Distance to the convex hull. Next we introduce a quantity which measures the non–convexity of a set. Let A belong to (F), its inner radius is (see [15]) K r(A)= sup inf R : a1;:::;as A; a co a1;:::;as ; a coA { ∃ ∈ ∈ { } ∈ a a R; 1 i s : || − i|| ≤ ≤ ≤ } Obviously, r(A) is zero if and only if A is convex. For any A, r(A) 2 A = ≤ || || 2supa A a . In [11],∈ || Puri|| and Ralescu extended a result of Cassels [2] and proved the following inequality: for any A ;:::;A in (F ) 1 n K 1=p d(A + +A ;co A + +coA ) c1=p r(A )p + +r(A )p : 1 ··· n 1 ··· n ≤ 1 ··· n Of course, the exponent p is related to the fact that F is a Banach space of type p and the constant c is the one appearing in the functional inequality (see the definition just before Theorem 1). License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 2434 RAPHAEL¨ CERF Proof of Theorem 1. We suppose first that the sets (An)n are convex. We ∈N∗ apply the general Cram´er theorem with E = (B∗;w ), E = s(co (F )) and the C ∗ 1 K sequence of random functions (sA )n . By the Riesz representation theorem [14], n ∈N∗ the topological dual of E is the set M(B∗) of the signed Borel measures on (B∗;w ). By the hypothesis of Theorem 1, ∗ + α R E(exp α sA1 )=E(exp(α sup a )) < + ∀ ∈ || ||∞ a A || || ∞ ∈ 1 so that the law of (s + +s )=n satisfies a large deviation principle with rate A1 ··· An function ΛE∗ (defined on E).