Abstract
The goals of this chapter are to: • Introduce \(\mathfrak{N}\)-distances defined by zonoids, • Explain the connections between \(\mathfrak{N}\)-distances and zonoids.
Access provided by Autonomous University of Puebla. Download chapter PDF
Similar content being viewed by others
Keywords
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.
The goals of this chapter are to:
-
Introduce \(\mathfrak{N}\)-distances defined by zonoids,
-
Explain the connections between \(\mathfrak{N}\)-distances and zonoids.
Notation introduced in this chapter:
Notation | Description |
---|---|
h(K, u) | Support function of a convex body |
K 1 ⊕ K 2 | Minkowski sum of sets K 1 and K 2 |
S d − 1 | Unit sphere in ℝd |
1 Introduction
Suppose that \(\mathfrak{X}\) is a metric space with the distance ρ. It is well known (Schoenberg 1938) that \(\mathfrak{X}\) is isometric to a subspace of a Hilbert space if and only if ρ2 is a negative definite kernel. The so-called \(\mathfrak{N}\)-distance (Klebanov 2005) is a variant of a construction of a distance on a space of measures on \(\mathfrak{X}\) such that \({\mathfrak{N}}^{2}\) is a negative definite kernel. Such a construction is possible if and only if ρ2 is a strongly negative definite kernel on \(\mathfrak{X}\).
In this chapter, we show that the supporting function of any zonoid in ℝd is a negative definite first-degree homogeneous function. The inverse is also true. If the support of a generating measure of a zonoid coincides with the unit sphere, then the supporting function is strongly negative definite, and therefore it generates a distance on the space of Borel probability measures on ℝd.
2 Main Notions and Definitions
Here we review some known definitions and facts from stochastic geometry.Footnote 1
Let \(\mathfrak{C}\) (resp. \({\mathfrak{C}}^{{\prime}}\)) be the system of all compact convex sets (resp. nonempty compact convex sets) in ℝd. A set \(K \in {\mathfrak{C}}^{{\prime}}\) is called a convex body if \(K \in {\mathfrak{C}}^{{\prime}}\); then for each u ∈ S d − 1 there is exactly one number h(K, u) such that the hyperplane
intersects K, and ⟨x, u⟩ − h(K, u) ≤ 0 for each x ∈ K. This hyperplane is called the support hyperplane, and the function h(K, u), u ∈ S d − 1 (where S d − 1 is the unit sphere), is the support function (restricted to S d − 1) of K. Equivalently, one can define
Its geometrical meaning is the signed distance of the support hyperplane from the coordinate origin.
An important property of h(K, u) is its additivity:
where \(K_{1} \oplus K_{2} =\{ a + b :\; a \in K_{1},\,b \in K_{2}\}\) is the Minkowski sum of K 1 and K 2. For \(K \in {\mathfrak{C}}^{{\prime}}\) let Ǩ\(=\{ -k,k \in K\}\). We say that K is centrally symmetric if K ′ = Ǩ ′ for some translate K ′, i.e., if K has a center of symmetry.
The Minkowski sum of finitely many centered line segments is called a zonotope. Consider a zonotope
where a i > 0, \(v_{i} \in {\mathbb{S}}^{d-1}\). Its support function is given by
We use the notation \({\mathcal{K}}^{{\prime}}\) for the space of all compact subsets of ℝd with the Hausdorff metric
where \(\mathrm{dist}(x,K) =\mathop{\inf}_{z\in K}\|x - z\|\).
A set \(\mathcal{Z}\in {\mathfrak{C}}^{{\prime}}\) is called a zonoid if it is a limit in a d H distance of a sequence of zonotopes.
It is known that a convex body \(\mathcal{Z}\) is a zonoid if and only if its support function has a representation
for an even measure \(\mu _{\mathcal{Z}}\) on \({\mathbb{S}}^{d-1}\). The measure \(\mu _{\mathcal{Z}}\) is called the generating measure of \(\mathcal{Z}\). It is known that the generating measure is unique for each zonoid \(\mathcal{Z}\).
3 \(\mathfrak{N}\)-Distances
Suppose that \((\mathfrak{X},\mathfrak{A})\) is a measurable space and \(\mathcal{L}\) is a strongly negative definite kernel on \(\mathfrak{X}\). Denote by \(\mathcal{B}_{\mathcal{L}}\) the set of all probabilities μ on \((\mathfrak{X},\mathfrak{A})\) for which there exists the integral
For \(\mu, \nu \in \mathcal{B}_{\mathcal{L}}\) consider
It is known (Klebanov 2005) that
is a distance on \(\mathcal{B}_{\mathcal{L}}\).
Described below are some examples of negative definite kernels.
Example 25.3.1.
Let \(\mathfrak{X} = {\mathbb{R}}^{1}\). For r ∈ [0, 2] define
The function \(\mathcal{L}_{r}\) is a negative definite kernel. For r ∈ (0, 2), \(\mathcal{L}_{r}\) is a strongly negative definite kernel.
For the proof of the statement in this example and the statement in the next example (Example 25.3.2), See Klebanov [2005].
Example 25.3.2.
Let \(\mathcal{L}(x,y) = f(x - y)\), where f(t) is a continuous function on ℝd, f(0) = 0, \(f(-t) = f(t)\). \(\mathcal{L}\) is a negative definite kernel if and only of
where Θ is a finite measure on ℝd. Representation (25.3.3) is unique. Kernel \(\mathcal{L}\) is strongly negative definite if the support of the measure Θ coincides with the whole space ℝd.
We will give an alternative proof for the fact that | x − y | is a negative definite kernel. For the case \(\mathfrak{X} = {\mathbb{R}}^{1}\) define
Then \(\mathcal{L}\) is a negative definite kernel.
Proof.
It is sufficient to show that max(x, y) is a negative definite kernel. For arbitrary a ∈ ℝ1 consider
It is clear that
Let F(a) be a nondecreasing bounded function on ℝ1. Define
For any integer n > 1 and arbitrary c 1, …, c n under condition \(\sum_{j=1}^{n}c_{j} = 0\) we have
But
Let us fix arbitrary A > 0 and apply the previous equality to the function
In this case, \(\mathcal{K}(x,y) = A -\max (x,y)\) for x, y ∈ [ − A, A], and, as A → ∞, we obtain that max(x, y) is a negative definite kernel.
Directly from the definition of a negative definite kernel and Example 25.3.1 we obtain the next example.
Example 25.3.3.
Let x, y ∈ ℝd, and f : ℝd → ℝ1. Define
Then \(\mathcal{L}\) is a negative definite kernel.
Of course, the mixture of negative definite kernels is again a negative definite kernel.
Example 25.3.4.
Let us choose and fix a vector \(\theta \in {\mathbb{S}}^{d-1}\) and consider the kernel
From previous considerations it is clear that \(\mathcal{L}_{\theta }\) is a negative definite kernel on ℝd, and for the σ-finite measure Ξ
is, again, a negative definite kernel.
Consider expression (25.3.2) constructed on the basis of (25.3.7). Let us rewrite (25.3.2) in a different form. Suppose that X and Y are two random vectors in ℝd with distributions μ and ν, respectively. We write \(\mathcal{N}(X,Y )\) instead of \(\mathcal{N}(\mu, \nu )\), so that
where \({X}^{{\prime}}\stackrel{d}{=}X\) and \({Y }^{{\prime}}\stackrel{d}{=}Y\) are independent copies of X and Y, respectively. Note that we use the sign \(\stackrel{d}{=}\) for the equality in a distribution. We have
Denote X θ = ⟨X, θ⟩, Y θ = ⟨Y, θ⟩. Then
But Eu a (X θ) = Pr{X θ < a}, and therefore
where F θ(a) = Pr{X θ < a}, G θ(a) = Pr{Y θ < a}. So finally we have
If the support of Ξ coincides with \({\mathbb{S}}^{d-1}\), then \(\mathfrak{N}(X,Y ) ={\Bigl ( \mathcal{N}{(X,Y )\Bigr )}}^{1/2}\) is a distance between the distributions of X and Y.
Let us return to the kernel
Choose arbitrary \(\theta _{o} \in {\mathbb{S}}^{d-1}\), and consider the measure
where \(\delta _{\theta _{o}}\) is the measure concentrated at point θ o . Then
Now, if we have an arbitrary even measure Ξ s on sphere \({\mathbb{S}}^{d-1}\), then
is a negative definite kernel. Let us note that the function
is the support function of a zonoid with generating measure Ξ s .
Summarizing all the preceding relations we may formulate the following result.
Theorem 25.3.1.
Each zonoid \(\mathcal{Z}\) generates a negative definite kernel on ℝd
This kernel is strongly negative definite if the support of \(\mu _{\mathcal{Z}}\) coincides with the whole sphere \({\mathbb{S}}^{d-1}\), and
is the square of a distance between measures \(\mu, \nu \in \mathcal{B}_{\mathcal{L}}\). This distance has the following representation:
where
According to Example 25.3.2, the function \(h_{\mathcal{Z}}(u)\) from (25.3.10) may be represented in the form (25.3.3). Let us investigate the connection between \(\mu _{\mathcal{Z}}\) in (25.3.10) and Θ in (25.3.3). To do so, we will use the following identity:
We have
So
If \(h_{\mathcal{Z}}(u)\) is a support function of a zonoid \(\mathcal{Z}\), then clearly
for all τ > 0 and u ∈ ℝd, and, as was shown previously, \(h_{\mathcal{Z}}(x - y)\) is a negative definite kernel. The inverse is also true.
Theorem 25.3.2.
Suppose that f is a continuous function on ℝd such that f(0) = 0, \(f(-u) = f(u)\). Then the following facts are equivalent:
-
Fact 1. f(τ ⋅ u) = τf(u) and f(x − y) is a negative definite kernel.
-
Fact 2. f is a support function of a zonoid.
Proof.
Previously we saw that Fact 2 implies Fact 1, and we must prove only that Fact 1 implies Fact 2. According to Example 25.3.2,
where
and Θ is the measure from (25.3.3).
We have
for any τ > 0, u ∈ ℝd. Substituting (25.3.15) into (25.3.16) and using the uniqueness of the measure Θ in (25.3.3) we obtain
and
We write here v = r ⋅w for r > 0 and \(w \in {\mathbb{S}}^{d-1}\). We have
and, finally, for τ = r,
It is clear that representation (25.3.15) for Θ1 of the form (25.3.17) coincides with (25.3.14).Footnote 2
Note that the \(\mathfrak{N}\)-distance can be bounded by the Hausdorf distance. Let \(\mathcal{Z}_{\mu }\) and \(\mathcal{Z}_{\nu }\) be two zonoids with generating measures μ and ν, respectively. The following inequality holds for their supporting functions \(h(\mathcal{Z}_{\mu },u)\) and \(h(\mathcal{Z}_{\nu },u)\):
Obviously, from this inequality it follows that
and therefore
Note that each \(\mathfrak{N}\)-distance generated by a zonoid is an ideal distance of degree 1 ∕ 2.
References
Beneš V, Rataj J (2004) Stochastic geometry: selected topics. Kluwer, Boston
Burger M (2000) Zonoids and conditionally positive definite functions. Portugaliae Mathematica 57:443–458
Klebanov LB (2005) \(\mathfrak{N}\)-distances and their applications. Karolinum Press, Prague
Schoenberg IJ (1938) Metric spaces and positive definite functions. Trans Am Math Soc 44:552–563
Ziegler GM (1995) Lectures on polytopes. Springer, New York
Author information
Authors and Affiliations
Rights and permissions
Copyright information
© 2013 Springer Science+Business Media, LLC
About this chapter
Cite this chapter
Rachev, S.T., Klebanov, L.B., Stoyanov, S.V., Fabozzi, F.J. (2013). Distances Defined by Zonoids. In: The Methods of Distances in the Theory of Probability and Statistics. Springer, New York, NY. https://doi.org/10.1007/978-1-4614-4869-3_25
Download citation
DOI: https://doi.org/10.1007/978-1-4614-4869-3_25
Published:
Publisher Name: Springer, New York, NY
Print ISBN: 978-1-4614-4868-6
Online ISBN: 978-1-4614-4869-3
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)