Hyperbolic Measures on Infinite Dimensional Spaces

Localization and dilation procedures are discussed for infinite dimensional $\alpha$-concave measures on abstract locally convex spaces (following Borell's hierarchy of hyperbolic measures).


Introduction
The purpose of this note is to review some results about the localization techniques and hyperbolic measures on R n and to discuss possible extensions to the setting of abstract (infinite dimensional) locally convex spaces. As a starting point, let us recall the so-called "Localization Lemma" which is due to Lovász and Simonovits. (1.1) Then, for some points a, b ∈ R n and a positive affine function l on (0, 1), 1 0 u((1 − t)a + tb) l(t) n−1 dt > 0, There are some other variants of this theorem, for example, when the first integrals involving the function u are vanishing, cf. [K-L-S]. The approach of Lovász and Simonovits was based on the concept of a needle coming as result of a localization (or bisection) procedure. Later Fradelizi and Guédon [F-G1] proposed an alternative geometric argument with involvement of Krein-Milman's theorem, cf. also [F-G2].
Theorem 1.1 is a powerful tool towards certain integral relations in R n ; it allows reduction to related inequalities in dimension n = 1. It is therefore not surprising that this theorem has found numerous applications in different problems of multidimensional Analysis and Geometry, such as isoperimetric problems over convex bodies, log-concave and more general hyperbolic measures, as well as Khinchine and dilation-type inequalities (cf. [K-L-S], [G], [B1,2,3,6], [N-S-V], [B-N], [F], [B-M]). In many such applications, one considers integrals with respect to measures that are different than the Lebesgue measure on R n , and therefore a more flexible version of Theorem 1.1 involving other measures would be desirable. In addition, having in mind dimension free phenomena and applications to random processes with hyperbolic distributions, it is useful to avoid reference to the dimension and to obtain similar statements about spaces and measures of an infinite dimension.
A positive Radon measure µ on a locally convex space E is called α-concave (−∞ ≤ α ≤ ∞) if, for all non-empty Borel sets A and B in E and for all 0 < t < 1, Here, (1 − t)A + tB = {(1 − t)x + ty : x ∈ A, y ∈ B} stands for the Minkowski weighted sum, and µ * is the inner measure (for a possible case when (1 − t)A + tB is not Borel measurable). By the Radon property, (1.1) may equivalently be stated for all non-empty compact subsets of E, and then the inner measure is not needed. Any measure supported on a one-point set is ∞-concave. In all other cases, necessarily α ≤ 1. For example, the Lebesgue measure on R is 1-concave. More generally, the Lebesgue measure on R n is 1 n -concave, which is the content of the Brunn-Minkowski theorem. Inequality (1.3) strengthens with growing α. In the limit case α = −∞, (1.3) becomes corresponding to µ * (1 − t)A + tB ≥ min{µ(A), µ(B)}, (1.4) which describes the largest class. Such measures µ are called convex or hyperbolic. One important case is also α = 0, for which (1.3) is understood as Then the measure µ is called logarithmically concave, or just log-concave. The class of log-concave measures on R n was first considered by Prékopa [Pr] and previously in dimension one by other authors (cf. [I], [D-K-H]). The more general classes of α-concave measures (in the setting of an abstract locally convex space) were introduced by Borell [Bor1]. He studied basic properties of α-concave measures, including 0 − 1 law, integrability of norms, convexity properties of measures under convolutions. Borell also gave a characterization of the α-concavity in terms of densities of finite dimensional projections, cf. , and also [B-L].
Hyperbolic measures are known to satisfy many other important properties that are usually expressed in terms of various relations such as, for example, Khinchin-type inequalities for polynomials of a bounded degree. What is remarkable, most of them involve only the convexity parameter α and do not depend on the dimension of the underlying space E. It is therefore natural to state these relations without any restrictions on E where possible and anyway wider than in the popular setting of the Euclidean space E = R n . For example, returning to Theorem 1.1, it may be complemented with the following.
Theorem 1.2. Let µ be a finite α-concave measure on a complete locally convex space E, and let u, v : E → R be lower semi-continuous µ-integrable functions such that (1.5) Then, for some points a, b ∈ E and some finite α-concave measure ν supported on the segment Note that lower semicontinuous functions are bounded below on any compact set. Hence, their integrals over compactly supported finite measures such as (1.2) and (1.6) always exist. As an example, the indicator functions of open subsets of E are all lower semicontinuous.
The completeness assumption (meaning that every Cauchy net in E is convergent) is quite natural. It ensures that the closed convex hull of any compact set in E is also compact. In that case any finite Radon measure µ on E has a stronger property ( 1.7) This property is crucial in some applications, but without completeness it is not true in general.
(Its validity remains unclear e.g. for Radon Gaussian measures.) One can also give a geometric variant of Theorem 1.1 together with a finer formulation of Theorem 1.2 in terms of extreme points of the set P α (u) of all α-concave probability measures supported on a convex compact set K ⊂ E and such that u dµ ≥ 0 (for a continuous function u on K). As we already mentioned, this interesting approach to localization was developed by Fradelizi and Guédon [F-G1]. It was shown there that in case E = R n and α ≤ 1 2 , any extreme point in P α (u) is either a mass point or it is supported on an interval ∆ ⊂ K with density l (1−α)/α (where l is a non-negative affine function on ∆). As will be explaned in Section 3, this property extends to general locally convex spaces, and then it easily implies Theorem 1.2.
One interesting application of Theorem 1.2 may be stated in terms of the following operation proposed in [N-S-V]. Given a Borel subset A in a closed convex set F ⊂ E and a number δ ∈ [0, 1], define where m ∆ denotes the normalized one-dimensional Lebesgue measure on ∆.
For example, if F = E and A is the complement to a centrally symmetric, open, convex set B ⊂ E, then A δ = E \ ( 2 δ − 1)B represents the complement to the corresponding dilation of B.
Theorem 1.3. Let µ be an α-concave probability measure on a complete locally convex space E supported on a closed convex set F (−∞ < α ≤ 1). For any Borel set A in F and for all δ ∈ [0, 1] such that µ * (A δ ) > 0, Here µ * denotes the outer measure (which is not needed, when E is a Fréchet space). This relation resembles very much the definition (1.3).
In the important particular case α = 0 (i.e., for log-concave measures), (1.8) becomes It was discovered by Nazarov, Sodin and Vol'berg [N-S-V]. The extension of this result to the class of α-concave measures in the form (1.8) is settled in [B-N] and [F], still for finite dimensional spaces. All proofs are essentially based on Theorem 1.1 or its modifications to reduce (1.8) to dimension one (although the one dimensional case appears to be rather delicate). Here we make another step removing the dimensionality of the space assumption, cf. Section 6. The organization of this note is as follows. In Section 2 we recall basic general facts about α-concave measures, including Borell's characterization of the α-concavity in terms of densities, and describe several examples. Sections 3-4 are devoted to the extension of Fradelizi-Guédon's theorem and Lovász-Simonovits' bisection argument. In particular, the existence of needles which we understand in a somewhat weaker sense is proved for probability measures on Fréchet spaces that satisfy the zero-one law. This can be used as an approach towards Theorems 1.1-1.2, but potenitally may have a wider range of applications. Finally, Sections 5-6 are devoted to Theorem 1.3, which is then illustrated in the problem of large and small deviations (Section 7).
We do not try to describe in detail results and techniques in dimension one, but mainly focus on their extensions to the setting of infinite dimensional spaces.

Support, dimension and characterizations
The support H µ = supp(µ) of any Radon measure µ on E is defined as the smallest closed subset of E of full measure, so that µ(E \ H µ ) = 0. If µ is hyperbolic, then the set H µ is necessarily convex, as follows from (1.4). This set has some dimension k = dim(µ) = dim(H µ ), finite or not, which is called the dimension of the hyperbolic measure µ. If it is finite, absolute continuity of µ will always be understood with respect to the k-dimensional Lebesgue measure on H µ .
First, let us recall an important general property of hyperbolic measures proven by Borell.
In particular, any µ-measurable affine subspace of E has measure either zero or one. In , Borell also gave a full description of α-concave measures. Similarly to (1.3), a non-negative function f defined on a convex subset H of E is called β-concave, if it satisfies for all t ∈ (0, 1) and all points x, y ∈ H such that f (x) > 0 and f (y) > 0. The right-hand side is understood in the usual limit sense for the values β = −∞, β = 0 and β = ∞.
Theorem 2.2 ( [Bor1]). If µ is a finite α-concave measure on R n of dimension k = dim(µ), then α ≤ 1 k . Moreover, µ is absolutely continuous with respect to Lebesgue measure on H µ and has density f which is positive, finite, and β-concave on the relative interior of H µ , where Conversely, if a measure µ on R n is supported on a convex set H of dimension k and has there a positive, β-concave density f with β ≥ − 1 k , then µ is α-concave.
Note that β is continuously increasing in the range [− 1 k , ∞], when α is varying in [−∞, 1 k ]. In the extremal case α = 1 k , the density f (x) = dµ(x) dx is ∞-concave and is therefore constant: Up to a factor, µ must be the k-dimensional Lebesgue measure on H µ .
More generally, if α ≤ 1 k , α = 0, the density has the form for some function V : Ω → (0, ∞) on the relative interior Ω of H µ , which is concave in case α > 0, and is convex in case α < 0. In particular, the formula As for general locally convex spaces, another theorem due to Borell reduces the question to Theorem 2.2. For special spaces in this characterization one may consider linear continuous maps T from a sufficiently rich family. For example, when E = C[0, 1] is the Banach space of all continuous functions on [0, 1] with the maximum-norm, the measure µ is α-concave, if and only if the image of µ under any map of the form T x = (x(t 1 ), . . . , x(t n )), x ∈ C[0, 1], t 1 , . . . , t n ∈ [0, 1], is an α-concave measure on R n . Simialrly, when E = R ∞ is the space of all sequences of real numbers (with the product topology), it is sufficient to consider the standard projections The next general observation emphasizes that infinite dimensional α-concave measures may not have a positive parameter of convexity. Apparently, it was not stated explicitly in the literature, so we include the proof. As usual, E ′ denotes the dual spaces of all linear continuous functionals on E.
Theorem 2.4. For α > 0, any α-concave finite measure µ on a locally convex space E has finite dimension and is compactly supported.
Proof. First, suppose to the contrary that µ is infinite dimensional. We may assume that H µ = supp(µ) contains the origin. Since H µ is not contained in any finite dimensional subspace of E, for each n, one can find linearly independent vectors v 1 , . . . , v n ∈ H µ . Each point x ∈ E has a representation x = c 1 (x)v 1 +· · ·+c n (x)v n +y with some c i ∈ E ′ , where y = y(x) is linearly independent of all v i (cf. [R], Lemma 4.21). Consider the linear map T (x) = (c 1 (x), . . . , c n (x)), which is continuously acting from E to R n . Then the image ν = µT −1 of µ is a finite α-concave measure on R n .
Let us see that ν is full dimensional. Otherwise, ν is supported on some hyperplane in R n described by the equation a 1 y 1 + · · · + a n y n = a 0 , where the coefficients a i ∈ R are not all zero. Moreover, since 0 ∈ H µ , any neighborhood of 0 has a positive µ-measure, so for any ε > 0. Hence, necessarily a 0 = 0. This implies that µ is supported on the closed linear subspace H of E described by the equation a 1 c 1 (x) + · · · + a n c n (x) = 0. Here, at least one of the coefficient, say a i , is non-zero. Since c i (v i ) = 1 = 0, we obtain that v i / ∈ H. But this would mean that H µ ∩ H is a proper closed subset of the support of µ, while H µ has a full µ-measure, a contradiction.
Thus, µ must be supported on a finite dimensional affine subspace H ⊂ E. To prove compactness of the support, we may assume that H = E = R n and dim(µ) = n. Then, µ is supported on an open convex set Ω ⊂ R n , where it has density of the form for some concave function V : Ω → (0, ∞). The case γ = 0 is possible, but then f (x) = c for some constant c > 0, which implies µ(R n ) = c |Ω|. Since µ is finite, Ω has to be bounded, and so H µ = clos(Ω) is compact. Now, let γ > 0. Suppose that Ω is unbounded (to justify several notations below). It is known (cf. e.g. [B5]) that f (x) → 0 uniformly as |x| → ∞ (x ∈ Ω). In particular, f is bounded, that is, A = sup x∈Ω f (x) is finite. Here, we may assume that the sup is asymptotically attained at x 0 = 0 for some sequence x l → 0, x l ∈ Ω. Choose r > 0 so that has a limit λ(x) = sup{λ : λ l ∈ Ω} > 1. Consider the convex functions Hence, ψ l (λ) ≥ Cλ, for all admissible λ ≥ 1, and letting l → ∞, we obtain But V is non-negative, so necessarily λ(x) ≤ 1 1−2 −1/γ . This proves boundedness of Ω.
Examples 2.5. 1. The normalized Lebesgue measure on every convex body K ⊂ R n is 1 n -concave. 2. Any Gaussian measure on a locally convex space E is log-concave. In particular, the Wiener measure on C[0, 1] is such.
4. Although the above density f n essentially depends on the dimension, the measure µ n has a dimension-free essense. All marginals of µ n coincide with µ 1 and moreover, there is a unique Borel probability measure µ on R ∞ (an infinite dimensional Cauchy measure) which is pushed forward to µ n by the standard projection T n from (2.2). This measure can also be introduced as the distribution of the random sequence where the random variables ζ, Z 1 , Z 2 , . . . are independent and all have a standard normal distribution. Thus, µ is (−1)-concave on R ∞ . 5. This example is mentioned in [Bor1]. Given d > 0 (real), let χ d be a positive random variable such that χ 2 d has the χ 2 -distribution with d degrees of freedom, i.e., with density Let W be the standard Wiener process (independent of χ d ) viewed as a random function in C[0, 1]. Then the random function has the distribution µ which is α-concave on C[0, 1] with α = − 1 d . It is called the Student measure (and also Cauchy in case d = 1 similarly to the previous example).

Extreme α-concave measures
Given a convex compact set K in a locally convex space E, denote by M α (K) the collection of all α-concave probability measures with support contained in K. For a continuous function u on K, we consider the subcollection together with its closed convex hull P α (u) in the locally convex space M(K) of all signed Radon measures on K endowed with the topology of weak convergence. The latter space is dual to the space C(K) of all continuous functions on K, and P α (u) is a convex compact subset of M(K).
What are extreme points of P α (u)? Using a general theorem due to D. P. Milman, one can only say that all such points lie in P α (u) (cf. [B-S-S], p.124, or [Ph] for a detail discussion of Krein-Milman's theorem). A full answer to this question is given in Fradelizi-Guédon's theorem, which we formulate below in the setting of abstract locally convex spaces.
Theorem 3.1. Given a continuous function u on K and −∞ ≤ α ≤ 1, any extreme point In particular, any α-concave probability measure supported on K, belongs to the closed convex hull of the family of all one-dimensional α-concave probability measures supported on K having density of the form (3.1).
We only consider the first assertion of the theorem. The second part is a purely one dimensional statement, and we refer to [F-G1].
Proof. Suppose that a measure µ ∈ P α (u) has the dimension dim(µ) ≥ 2. For simplicity, let the origin belong to the relative interior G of the support H µ of µ. Then one may find linearly independent vectors x and y such that ±x and ±y are all in G. On the linear hull L(x, y) of x and y (which is a 2-dimensional linear subspace of E), define linear functionals λ x and λ y by putting λ x (x) = λ y (y) = 1, They are continuous, so by the Hahn-Banach theorem, these functionals may be extended from L(x, y) to the whole space E keeping linearity and continuity. With these extended functionals, we can associate Λ θ = θ 1 λ x +θ 2 λ y , where θ = (θ 1 , θ 2 ) ∈ S 1 (vectors on the unit sphere of R 2 ). Note that these functionals are uniformly bounded on K, i.e., sup Now, following in essense an argument of [F-G1], define the map Φ : By the construction, the set {Λ θ = 0} ∩ H µ represents a proper closed affine subspace of H µ . So, µ{Λ θ = 0} = 0 according to Theorem 3.1 (the zero-one law for hyperbolic measures). Hence, using (3.2), we may conclude that the map Φ is continuous.
One can now return to Theorem 1.2.
Proof of Theorem 1.2. Due to the property (1.7), and by the assumption (1.5), for some convex compact set K ⊂ E and a constant c > 0. Moreover, since the function min(u, c) is lower semicontinuous and bounded, while µ is Radon, where the sup is taken over all continuous functions on K such that g ≤ min(u, c) (cf. e.g. [M], Chapter 2, or [Bog], Chapter 7). A similar identity also holds for min (v, c). This allows us to reduce the statement of the theorem to the case where both u and v are continuous on K.
In the latter case, let u 0 = u − K u dµ. Consider the functional T (µ) = K v dµ. It is linear and continuous on M(K), and therefore being restricted to P α (u 0 ) it attains maximum at one of the extreme points ν. Since µ ∈ P α (u 0 ), we conclude that so, K u dν > 0 and K v dν > 0 which is (1.6). It remains to apply Theorem 3.1.
A similar argument, based also on the second part of Theorem 3.1, yields Theorem 1.1. Indeed, the n-dimensional integrals (1.1) can be restricted to a sufficently large closed ball K ⊂ R n . The normalized Lebesgue measure on K is α-concave with α = 1 n . Hence, the extreme points in P α (u) are at most one dimensional and have densities of the form l n−1 (if they are not Dirac measures).

Bisection and needles on Fréchet spaces
The notion of a needle was proposed by Lovász and Simonovits for the proof of Theorem 1.1 (Localization Lemma, cf. also [K-L-S]). Previously, it appeared implicitly in [P-W] and may be viewed as development of the Hadwiger-Ohmann bisection approach to the Brunn-Minkowski inequality ([H-O], [B-Z], cf. also [G-M] for closely related ideas).
As shown in [L-S], starting from (1.1), one can construct a decreasing sequence of compact convex bodies K l in R n that are shrinking to some segment ∆ = [a, b] and are such that, for each l, Moreover, choosing a further subsequence (if necessary) and applying the Brunn-Minkowski inequality in R n , one gets in the limit for some non-negative concave function ψ on ∆. Here |K l | denotes the n-dimensional volume, while the integration on the right-hand side is with respect to the linear Lebesgue measure on the segment. In this way, one may obtain a slightly weaker variant of (1.2) with ψ in place of l, and with non-strict inequalities. An additional argument of a similar flavour was then developed in [L-S] to make ψ affine (while the strict inequalities in (1.2) are easily achieved by applying the conclusion to functions u − εw and v − εw, where w > 0 is integrable, continuous, and ε > 0 is small enough). The last step shows that for K l one may take infinitesimal truncated cylinders with main axis ∆; it is in this sense the limit one dimensional measure l n−1 (x) dx on ∆ may be considered a needle. The aim of this section is to extend this construction to the setting of Fréchet, i.e., complete metrizable locally convex spaces. For example, E may be a Banach space, but there also other important spaces that are not Banach, such as the space E = R ∞ . Note that any finite Borel measure on a Fréchet space is Radon.
While one cannot speak about the Lebesgue measure when E is infinite dimensional, the main hypothesis (1.1) may readily be stated like (1.5) with integration with respect to a given (finite) Borel measure µ on E.
The space of all finite Borel measures on E is endowed with the topology of weak convergence. In particular, µ l → µ (weakly), if and only if for any bounded continuous functions u on E. As was noticed in [Bor1], the class of all α-concave probability measures on E is closed in the weak topology.
Definition 4.1. Let µ be a finite Borel measure on E. A Borel probability measure ν will be called a needle of µ, if it is supported on a segment [a, b] ⊂ E and can be obtained as the weak limit of probability measures where K l is some decreasing sequence of convex compact sets in E of positive µ-measure such that Here, all µ l represent normalized restrictions of µ to K l . In particular, all needles of a given α-concave measure are α-concave, as well. We do not require that K l be asymptotically close to infinitesimal truncated cylinders.
Definition 4.2. One says that a Borel probability measure µ on E satisfies the zero-one law, if any µ-measurable affine subspace of E has µ-measure either 0 or 1.
For example, this important property holds true for all (Radon) Gaussian measures. More generally, it is satisfied by any hyperbolic probability measure, as follows from Borell's Theorem 2.1.
With these definitions, Theorem 1.2 admits the following refinement.
Theorem 4.3. Suppose that a Borel probability measure µ on a Fréchet space E satisfies the zero-one law. Let u, v : E → R be lower semi-continuous µ-integrable functions such that Then, these inequalities also hold for some needle ν of µ. Moreover, if µ is supported on a closed convex set F , then ν may be chosen to be supported on F , as well.
First assume that E is a separable Banach space with norm · , and let E ′ denote the dual space (of all linear continous functionals on E) with norm · * . Suppose that any proper closed affine subspace of E has µ-measure zero. In this case, for the proof of Theorem 4.3 we use the construction similar to the one from the proof of Theorem 3.1.
Given 3 affinely independent points x, y, z in E, define linear functionals λ x and λ y on the linear hull L z (x, y) of x − z and y − z (which is a 2-dimensional linear subspace of E), by putting λ x (x − z) = λ y (y − z) = 1, (4.1) By the Hahn-Banach theorem, these functionals may be extended by linearity to the whole space E without increasing their norms. This will always be assumed below.
Lemma 4.4. Let {(x n , y n , z n )} n≥1 be affinely independent points in the Banach space E such that x n → x, y n → y, z n → z, where x, y, z are also affinely independent. Then the corresponding linear functionals λ xn and λ yn have uniformly bounded norms, i.e., sup n≥1 λ xn * < ∞, sup n≥1 λ yn * < ∞.
Proof. Define the lines Then, for w ∈ L z (x, y), w ≤ 1, where we use the notation dist(w, A) = inf{ w − a : a ∈ A} (the shortest distance from a point to the set). The extended linear functionals should thus satisfy the above inequalities on the whole space E for all w ≤ 1, i.e., Next, by shifting, one may assume that z = 0, in which case x and y are linearly independent and in particular x > 0 and y > 0. Using (4.3), it is enough to show that dist(x n , L zn (y n )) ≥ c, for all n ≥ n 0 , with some n 0 and c > 0. Indeed, take an arbitrary point w = z n + r(y n − z n ) in L zn (y n ), r ∈ R. By the triangle inequality, where the last inequality holds whenever |r| ≥ 3 xn−zn yn−zn . Hence, by the convergence assumption, for |r| ≥ r 0 = 4 x y , n ≥ n 0 .
In case |r| ≤ r 0 , again by the triangle inequality, Here, the right-hand side is also separated from zero for sufficiently large n. By a similar argument, dist(y n , L zn (x n )) ≥ c, for all n ≥ n 0 .
Proof of Theorem 4.3. We begin with a series of reductions assuming without loss of generality that µ(E) = 1.
Any Fréchet space with Radon probability measure µ has a subspace E 0 such that µ(E 0 ) = 1, and in addition there exists a norm · on E 0 with respect to which E 0 is a separable reflexive Banach space whose closed balls are compact in E (see [Bog], Theorem 7.12.4).
In particular, all Borel subsets of E 0 are Borel in E. By the zero-one law (turning to a smaller subspace if necessary), we may assume that any proper affine subspace of E 0 which is closed for the topology of E 0 has measure zero. That is, for all l ∈ E ′ 0 , µ{l = c} = 0, c ∈ R.
(4.4) Second, it suffices to assume that the support of µ is compact, metrizable and convex. Indeed, by Ulam's theorem, there is an increasing sequence of compact sets K n ⊂ E 0 such that µ(∪ n K n ) = 1. The closed convex hull of any compact set in E 0 is compact (which is true in any Banach and more generally complete locally convex spaces, cf. e.g. [K-A]). Therefore, all K n may additionally be assumed to be convex. These sets will also be compact in E, so that the associated weak topologies in the spaces of Borel probability measures on K n coincide, as well. By the dominated convergence theorem, lim n→∞ Kn so that Kn u dµ > 0 and Kn v dµ > 0 for large n. Hence, an application of the theorem to µ restricted and normalized to K n would provide the desired one dimensional measure ν, a needle of µ n and therefore of µ itself. Thus, from now on, we may assume that E is a separable Banach space, and µ is a Borel probability measure on E which is supported on a convex compact set K ⊂ E and is such that (4.4) holds true for all l ∈ E ′ .
We need only to prove the existence of ν such that u dν ≥ 0 and v dν ≥ 0. Since in this case we may apply the superficially weaker result to u − ε and v − ε for an ε > 0 chosen small enough to preserve the hypothesis.
In addition, it suffices to prove the result when u and v are both continuous. To see this, take u n and v n to be sequences continous functions increasing to lower semicontinuous u and v respectively. By the monotone convergence, lim n→∞ u n dµ = u dµ > 0 and lim n→∞ v n dµ = v dµ > 0, so we can take the approximating functions u n and v n to be such that u n dµ > 0 and v n dµ > 0. The theorem produces needles ν n of µ supported on F and such that u n dν n > 0, v n dν n > 0.
Since u ≥ u n and v ≥ v n , every such measure ν n will be the required needle. Let us now turn to the construction procedure. Given 3 affinely independent points x, y, z in E, consider the linear continuous functionals λ x and λ y on E introduced before Lemma 4.4 via the relations (4.1)-(4.2) and the Hahn-Banach theorem. To each point θ ∈ S 1 = {(t, s) : t 2 + s 2 = 1} we can associate a linear functional Λ θ = tλ x + sλ y and define the function Since µ{ξ : Λ θ (ξ − z) = 0} = 0 (cf. (4.4)), this function is continuous on S 1 . In addition, we have the identity Hence, by the intermediate value theorem, there exists θ ∈ S 1 such that Also, so that at least one the last two integrals is positive. Let H + denote one of the hyperspaces Necessarily, µ(H + ) > 0, and we may consider the normalized restriction µ + of µ to H + and will have the property that This procedure can be performed step by step along a sequence {(x n , y n , z n )} n≥1 of affinely independent points, chosen to be dense in K ×K ×K. Let ν 1 = µ + be constructed according to the above procedure for (x 1 , y 1 , z 1 ) and with an associated point θ 1 = (t 1 , s 1 ) ∈ S 1 . Similarly, on the n-th step, given ν n , let ν n+1 = ν + n be constructed for the triple (x n , y n , z n ) and with the associated linear functional Λ θn = Λ (tn,sn) = t n λ xn + s n λ yn .
Since the space of all Borel probability measures on K is compact and metrizable for the weak topology, the sequence ν n has a sub-sequential weak limit ν. In particular, from (4.5) we derive the desired property It remains to show that dim(H ν ) ≤ 1. Suppose not, in this case there exists affinely independent x, y, z in the relative interior of H ν that also contains the points 2z − x and 2z − y. Without loss of generality, let z = 0, so that ±x and ±y belong to the relative interior of H ν . By the density property, there exists a subsequence, say (x k , y k , z k ) such that (x k , y k , z k ) → (x, y, z).
By the construction, the measure ν + k is supported on the half-space H + k , which is either {ξ : Λ (t k ,s k ) (ξ − z k ) ≥ 0} or {ξ : Λ (t k ,s k ) (ξ − z k ) ≤ 0}. For definiteness, let it be the first half-space. Since all H + k contain x and −x, we then have (4.6) and similarly for the point y.
Recall that by Lemma 4.4, we can obtain a uniform bound M such that Hence, Λ (t k ,s k ) (z k ) → 0 and Λ (t k ,s k ) (x k − x) → 0 as k → ∞. But then by (4.6), necessarily Λ (t k ,s k ) (x k ) → 0, as well. By the same argument, Λ (t k ,s k ) (y k ) → 0. On the other hand, according to the definition of Λ (t k ,s k ) via (4.1)-(4.2), for each k, thus implying that lim k→∞ t k = lim k→∞ s k = 0. But this is impossible since t 2 k + s 2 k = 1. This proves that dim(H ν ) ≤ 1.

Dilation and its properties
Before turning to Theorem 1.3, we first comment on the basic properties of the operation A → A δ , where δ ∈ [0, 1] is viewed as parameter.
Let F be a closed convex subset of a locally convex space E with respect to which this operation is defined: As before, m ∆ denotes a uniform distribution on ∆ (understood as the Dirac measure, when the endpoints coincide). In this definiton, by the intervals ∆ we mean closed intervals [a, b] connecting arbitrary points a, b in F . Moreover, the requirement x ∈ ∆ may equivalently be replaced by the condition that x is one of the endpoints of ∆.
Note that A 1 = A. If 0 ≤ δ < 1, as an equivalent definition one could put Indeed, in this case, if x ∈ F \ A, then m [x,x] (A) = 0 < 1 − δ meaning that x / ∈ A δ according to the second definition. Thus, for δ ∈ [0, 1), both definitions lead to the same set and we have the property A δ ⊂ A.
Lemma 5.1. a) If A ⊂ F is closed, then every set A δ is closed as well. b) If E is a Fréchet space and A is Borel measurable in F , then every set A δ is universally measurable.
Let us recall that a set in a Hausdorff topological space E is called universally measurable, if it belongs to the Lebesgue completion of the Borel σ-algebra with respect to an arbitrary Borel probability measure on E. In that case one may freely speak about the measures of these sets.
Proof. For a Borel set A in F , consider the function First assume that A is closed. Then, given a net x i → x, y i → y in F indexed by a semi-ordered set I, we have lim sup After integration this implies lim sup i∈I ψ(x i , y i ) ≤ ψ(x, y).
Indeed, the space L 1 [0, 1] is separable, so the above relation is only to be checked for increasing sequences i = i n in I. But in that case one may apply the Lebesgue dominated convergence theorem. This means that ψ is upper semicontinuous on F × F , and thus A δ represents the intersection over all y ∈ F of the closed sets {x ∈ A : ψ(x, y) ≥ 1 − δ}.
In part b), assume that E is a Fréchet space. If A is Borel, then the function ψ is Borel measurable on F × F , so the complement of A δ in A, represents the x-projection of a Borel set in E × E. But every Borel set in a Polish space is Souslin, and therefore both A \ A δ and A δ are universally measurable (cf. [Bog], Corollary 6.6.7 and Theorem 7.4.1).
There is an opposite operation representing a certain dilation or enlargement of sets. Given a Borel measurable set B ⊂ F and δ ∈ [0, 1), define Here the union is running over all intervals ∆ ⊂ F such that m ∆ (B) > δ. Note that B δ contains B (since all singletons in B participate in the above union).
Lemma 5.2. For any δ ∈ [0, 1) and any Borel set B ⊂ F , the complement A = F \ B satisfies the dual relations Proof. Given x ∈ F , the property x / ∈ A δ means that, for some interval ∆ ⊂ F containing x, we have m ∆ (A) < 1 − δ, that is, m ∆ (B) > δ meaning that ∆ ⊂ B δ . Therefore, x / ∈ A δ ⇔ x ∈ B δ . For the last assertion, it remains to recall Lemma 5.1.
Lemma 5.3. Let F be a convex closed set in E, and let T be a linear continuous map from E to another locally convex space E 1 . For any Borel set C ⊂ T (F ), where the operation C → C δ is understood with respect to the image T (F ).
Proof. For all a, b ∈ F , the map T pushes forward the unform measure m [a,b] to m [T a,T b] . Therefore, the pre-image B = T −1 (C) has measure m [a,b] When E 1 = R n and C is a polytope, the dilated set C δ is a polytope, as well. Hence, by Lemma 5.3, (T −1 (C)) δ represents the intersection of finitely many half-spaces.

The dual form and proof of Theorem 1.3
Following [B-N], let us reformulate Theorem 1.3 in terms of dilated sets. Putting B = F \ A and using Lemma 5.2, the inequalty More precisely, in case α < 0, the above expression is well-defined and represents a strictly concave, increasing function in p ∈ [0, 1]. For α = 0, it is understood in the limit sense as which is also strictly concave and increasing. In case 0 < α ≤ 1, R (α) (p) is defined to be (6.2) on the interval 0 ≤ p ≤ 1 − (1 − δ) 1/α (when the expression makes sense) and we should put Theorem 6.1. Let µ be an α-concave probability measure on a complete locally convex space E supported on a convex closed set F (−∞ < α ≤ 1). For any Borel subset B of F and for all δ ∈ [0, 1), For example, on the real line E = R for the Lebesgue measure µ on the unit interval F = [0, 1], we have α = 1, and (6.3) becomes For the Cauchy measures µ on R n and R ∞ (cf. Examples 2.5), we have α = −1, and then (6.2)-(6.3) with F = E yield .
Note that when E is a Fréchet space and B is Borel, B δ is universally measurable, so there is no need to use the inner masure.
Thus, both Theorem 1.3 and Theorem 6.1 do not loose generality by assuming that 0 < δ < 1 (and we do this below in this section).
Proof of Theorem 6.1. Using Theorem 1.2, let us show how to reduce the desired statement (6.3) to dimension one. Since the sets B δ may only become larger, when F is getting larger, one may assume that F = H µ , i.e., the support of µ. Fix 0 < δ < 1.
Step 1: First suppose that B is an open set in F such that the boundary ∂B δ of B δ in F has µ-measure zero. Fix an arbitrary p ∈ (0, 1). Using the continuity of the functions R which is exactly the condition (1.5) for u = 1 B − p and v = R (α) δ (p) − 1 D (where 1 A denotes the indicator function of a set A). These functions are lower-semicontinuous, so we may apply Theorem 1.2: For some one dimensional α-concave probability measure ν supported on an interval ∆ ⊂ F , we have (1.6), i.e., where (B ∩ ∆) δ is the result of the one dimensional dilation operation applied to B ∩ ∆ with respect to ∆. Hence, we obtain ν((B ∩ ∆) δ ) < ν (B) which contradicts the relation (6.3) in dimension one.
Step 2: Here we describe one class of open sets to which the previous step may be applied. Let B be a set of the form T −1 (C) ∩ F , where T : E → R n is a continuous linear map and C ⊂ R n is an open polytope (n ≥ 1 is arbitrary). Then (T −1 (C)) δ represents an intersection of finitely many open half-spaces (Lemma 5.3). If µ(B) > 0, then, by the zero-one law, the boundaries of these half-spaces have µ-measure zero and hence µ(∂B δ ) = 0, as well.
More generally, let B = T −1 (C) ∩ F , where C is a finite union of open polytopes in R n . Then C δ is also a finite union of open polytopes. Using Lemma 5.3, we obtain that clos(B δ ) ⊂ T −1 (clos(C δ )), so ∂B δ ⊂ T −1 (∂C δ ). Again ∂C δ is contained in finitely many hyperplanes of R n and thus µ(∂B δ ) = 0.
Step 3: B is an arbitrary open set in F , assuming that F is a convex, compact set. Denote by G the collection of all cylindrical sets in F described on the last step. Such sets constitute a base in the original topology on F , since the two coincides by the compactness assumption. Hence B = ∪ G, where the union is over all G ∈ G such that G ⊂ B. Since G is closed under finite unions, one may apply the Radon property which gives For any G as above, we have µ(G δ ) ≥ R (α) δ (µ(G)), by the previous steps. Hence, we obtain (6.3) for B, as well.
Step 4: B is an arbitrary Borel set in F . By the strengthened Radon property (1.7), it is sufficient to consider the case of a non-empty compact set B, and we may additionally assume that F is compact.
Any open set in F containing x ∈ B contains this point together with B(x) ∩ F , where B(x) = T −1 x (C(x)). Here T x : E → R n is a continuous linear map and C(x) is a Euclidean ball in R n (with some n depending on x). Using compactness of B, one can compose its finite covering by the sets of the form with full intersection being B. Let {U i } i∈I be a decreasing net indexed by a semi-ordered directed set I such that each U i represents the intersection of finitely many sets G as above. The latter guarantees that µ(U i ) ↓ µ(B) along the net. Now, let δ < δ ′ < 1. Given x ∈ F , the property x / ∈ B δ means that m [x,y] (B) = inf i∈I m [x,y] (U i ) ≤ δ for any y ∈ F . In that case, there is i ∈ I such that m [x,y] (U i ) < δ ′ , and hence the increasing sets cover F . By the construction, for each i, the function ϕ(y) = m [x,y] (U i ) is of the type ϕ(y) = mes t ∈ (0, 1) : ∀ k ≤ l ∃j ≤ N k (1 − t)T x kj (x) + tT x kj (y) ∈ C(x kj ) for some continuous linear maps T x kj : E → R n(x kj ) and some Euclidean balls C(x kj ) in R n(x kj ) . As the boundaries of Euclidean balls do not contain non-degenerate intervals, any such function ϕ must be continuous on F . Therefore, all the sets V i (x) are open in F , so that by compactness, V i (x) = F for some i = i(x). Thus, given x ∈ F \ B δ , we have m [x,y] (U i(x) ) < δ ′ for any y ∈ F , and hence F \ B δ is contained in

It follows that B δ contains the intersection of the open sets
and thus, by the Radon property, On the other hand, by Step δ ′ (µ(U i )), and taking the limit along the net we get µ * (B δ (B)). It remains to let δ ′ ↓ δ and use the contunuity of R (α) δ with respect to δ.

Large and small deviations
As is known, the dilation-type inequality (1.8) of Theorem 1.3 may equivalently be stated on functions (which is often more convenient in applications). Namely, with every Borel measurable function u on E with values in the extended line [−∞, ∞], one associates its "modulus of regularity" δ u (ε) = sup mes t ∈ (0, 1) : where the supremum is running over all points x, y ∈ E such that u(x) is finite. The behavior of δ u near zero is used to control the probabilities of large and small deviations of u under hyperbolic measures by involving the parameter α, only (cf. [B4], [B-N], [F]). In particular, there is the following recursive functional inequality, which is stated below, in the setting of an abstract complete locally convex space E.
We assume that µ is an α-concave probability measure on E with −∞ < α ≤ 1 and that u is a Borel measurable, µ-a.e. finite function on E.
(7.2) Note that for α ≤ 0, the assumption λ < ess sup |u| may be removed. If µ is supported on a convex closed set F in E, the inequalities (7.1)-(7.2) continue to hold when u is defined on F (rather than on the whole space). In that case, in the definition of δ u the supremum should be taken over all points x, y ∈ F .
Proof of Theorem 7.1. Let us recall a simple argument based on Theorem 1.3. The latter is applied with F = E to the set A = {x ∈ E : λε < |u(x)| < ∞}.
Remark. An inequality of the form (7.5) can also be obtained by using a transportation argument, cf. [B5]. With this argument, a slightly weaker variant of (7.4) is derived in [B4]. Now, let us turn to the problem of small deviations.
Finally, let us illustrate Corollaries 7.2-7.3 on the example of the semi-norms.