Abstract
Recently, W. Lao and M. Mayer (2008) developed U-max-statistics, where instead of averaging the values of the kernel over various subsets, the maximum of the kernel is considered. Such statistics often appear in stochastic geometry. Their limit distributions are related to the distributions of extreme values. In this paper, we begin to consider the limit theorems for the generalized perimeter (the sum of side powers) of a random inscribed polygon and U-max-statistics related to it. We describe extreme values of the generalized perimeter and obtain limit theorems for the cases where the side powers involved in determining the generalized perimeter do not exceed 1.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 1. INTRODUCTION
In this paper, we consider the limiting behavior of U-max-statistics appearing in stochastic geometry. Suppose ξ1, ξ2, … is a sequence of independent equally distributed random variables with values in measurable space (\(\mathfrak{X}\), \(\mathfrak{A}\)), and the real-valued symmetrical Boreal function h(x1, …, xm) called an m-degree kernel is defined in space \({{\mathfrak{X}}^{m}}\).
U-max-statistics are determined as follows:
where n ≥ m and the set J = {(i1, …, im) : 1 ≤ i1 < … < im ≤ n} is the set of ordered m-elemental permutations with many indices from the set {1, …, n}. U-min-statistics are determined analogously.
U-max-statistics are introduced independently by Lao and Mayer in their theses [1] and [2]. They developed a method of studying the limiting behavior of U-max-statistics using the Poisson approximation from the monograph [3]. Their basic limit theorem from [4] is presented as follows.
Theorem 1. Consider the U-max-statistics Hn = maxJh(\({{\xi }_{{{{i}_{1}}}}}\), …, \({{\xi }_{{{{i}_{m}}}}}\)) introduced above and determine the following functions for each z ∈ \(\mathbb{R}\):
then for all n ≥ m and for all z ∈ \(\mathbb{R}\) the following inequality is true:
In [5], Silverman and Brown offer the conditions under which the general theorem proved in [4], leads in the limit to the nonsingular Weibull law.
Theorem 2. In terms of Theorem 1, if for a certain sequence of transformations zn : T → \(\mathbb{R}\), T ⊂ \(\mathbb{R}\) for each t ∈ T the conditions
are fulfilled, the equality limn → ∞\(\mathbb{P}\)(Hn ≤ zn(t)) = \({{e}^{{ - {{\lambda }_{t}}}}}\) for each t ∈ T holds.
Remark 1. If m ≥ 2, condition (2) can be replaced with
for each r ∈ {1, …, m – 1}.
The main application field of the concept of U-max-statistics is currently related to stochastic geometry: we talk about maximum areas, perimeters, volumes, and similar metric characteristics of figures built by a set of random points on a plane or in spaces of a larger dimension. The form of a geometric figure is determined by the kernel h. Lao and Mayer consider kernels only of small degrees; mostly, they study triangles. Koroleva and Nikitin in [6] pass on to U-max-statistics of a more complex nature. In particular, they consider the maximum perimeter among all perimeters of convex m-sided polygons whose vertices are chosen among n points that are independent and uniformly distributed on a circumference. This study summarizes and develops the results of [6].
2 GENERALIZED PERIMETER OF A POLYGON
Consider the generalization of the concept of a perimeter. Assume that di, i = 1, …, m are the sides of an inscribed m-sided polygon with vertices U1, …, Um and suppose h(U1, …, Um) = \(\sum\nolimits_{i = 1}^m {{{d}_{i}}} \) is its perimeter. We propose to consider the value hy(U1, …, Um) = \(\sum\nolimits_{i = 1}^m {d_{i}^{y}} \) for y ∈ \(\mathbb{R}\) and call it the generalized perimeter of a polygon.
It is well known that among all the convex inscribed polygons, the largest perimeter (and the square area) has a regular polygon (see, e.g., ([7], Problem 57a)). This was known at least to Legendre [8].
It turns out that this property is retained also for generalized perimeters when 0 < y < 1, while for y < 0, in contrast, on a regular polygon, the minimum of the generalized perimeter is achieved. We have not found evidence of this fact, which is of independent interest, in the literature. Therefore, we present the proof of this fact here.
Lemma 1. Assume that there exist m points V1, …, Vm lying on a unit circumference. Then for y < 0 the function hy(V1, …, Vm) attains a minimum only at the vertices of a regular m-sided polygon and its minimum value is equal to \({{2}^{y}}m\,{{\sin }^{y}}\left( {\frac{\pi }{m}} \right)\).
Analogously, when y ∈ (0, 1), the function hy attains its maximum only at the vertices of a regular m-sided polygon and its maximum value is again equal to \({{2}^{y}}m\,{{\sin }^{y}}\left( {\frac{\pi }{m}} \right)\).
Proof. First, consider the case of negative y. Without detracting from the generality, assume that the points V1, …, Vm are located on the circumference in the assigned order; then the function hy is presented as hy(V1, …, Vm) = \(\sum\nolimits_{j = 1}^m {{{{\left| {{{V}_{j}}{{V}_{{j + 1}}}} \right|}}^{y}}} \), where Vm + 1 = V1.
Introduce the notation γi = \(\angle {{V}_{i}}O{{V}_{{i + 1}}}\), where i ∈ {1, …, m}. We have \(\left| {{{V}_{i}}{{V}_{{i + 1}}}} \right|\) = 2 sin\(\left( {\frac{{{{\gamma }_{i}}}}{2}} \right)\), where i ∈ {1, …, m}. It follows that
We can assume that γi ∈ (0, 2π), otherwise hy = ∞. Consider the function f(x) = siny x for x ∈ (0, π). We find its first two derivatives
It is clear that for negative y this function is strictly convex on (0, π). Then by Jensen’s inequality for any y1, …, yn ∈ (0, π), we obtain
provided \(\sum\nolimits_{i = 1}^n {{{\alpha }_{i}}} \) = 1, αi > 0. Note that
At the same time, because of the strict convexity, the equality is achieved only in the case where all γi are identical. Hence, in the case where y < 0, the function hy(V1, …, Vm) attains the minimum equal to \({{2}^{y}}mf\left( {\frac{\pi }{m}} \right)\) = \({{2}^{y}}m\,{{\sin }^{y}}\left( {\frac{\pi }{m}} \right)\) only at the vertices of a regular m-sided polygon.
The case where 0 < y < 1 is considered analogously. If y ∈ (0, 1), then f ''(x) < 0; therefore, the function is strictly concave. By similar reasons, the function hy attains its maximum only at the vertices of a regular m-sided polygon.
Suppose the points U1, …, Un are independently and uniformly distributed on a unit circumference S1. Then the following two limit theorems hold; they are the main results of this study.
Theorem 3. Suppose \(H_{{n,m}}^{y}\) = \(\mathop {\min }\limits_{1 \leqslant {{i}_{1}} < \ldots < {{i}_{m}} \leqslant n} {{h}_{y}}({{U}_{{{{i}_{1}}}}}, \ldots ,{{U}_{{{{i}_{m}}}}})\). Then when y < 0, for any t > 0 the following equality holds:
where K1(y) = \({{m}^{{\frac{3}{2}}}}\Gamma \left( {\frac{{m + 1}}{2}} \right){{\left( { - {{2}^{{y - 1}}}y\pi {{{\left( {\sin \frac{\pi }{m}} \right)}}^{{y - 2}}}\left( {1 - y\,{{{\cos }}^{2}}\frac{\pi }{m}} \right)} \right)}^{{\frac{{m - 1}}{2}}}}\).
Theorem 4. Suppose \(G_{{n,m}}^{y}\) = \(\mathop {\max }\limits_{1 \leqslant {{i}_{1}} < \ldots < {{i}_{m}} \leqslant n} {{h}_{y}}({{U}_{{{{i}_{1}}}}}, \ldots ,{{U}_{{{{i}_{m}}}}})\). Then when y ∈ (0, 1), for any t > 0 the following equality holds:
where K2(y) = \({{m}^{{\frac{3}{2}}}}\Gamma \left( {\frac{{m + 1}}{2}} \right){{\left( {{{2}^{{y - 1}}}y\pi {{{\left( {\sin \frac{\pi }{m}} \right)}}^{{y - 2}}}\left( {1 - y\,{{{\cos }}^{2}}\frac{\pi }{m}} \right)} \right)}^{{\frac{{m - 1}}{2}}}}\).
3 PROOF OF THE MAIN RESULTS
Proof of Theorem 3 is similar to the corresponding proof in [6] but has several distinctions.
Consider the points U1, …, Um ∈ S1. From Lemma 1 it follows that the minimum of the function hy is \({{2}^{y}}m\,{{\sin }^{y}}\left( {\frac{\pi }{m}} \right)\) and is achieved only at the vertices of a regular m-sided polygon. Introduce the notation βi = \(\angle {{U}_{1}}O{{U}_{{i + 1}}}\) (in the counterclockwise direction). Analogously to the work [6], when z > \({{2}^{y}}m\,{{\sin }^{y}}\left( {\frac{\pi }{m}} \right)\), we obtain
In what follows, we assume that U1, …, Um lie on the circumference in this order in the counterclockwise direction.
We call the angles of the form \(\angle {{U}_{i}}O{{U}_{{i + 1}}}\) central angles. Consider the largest and smallest central angles. Assume that these are the angles \(\frac{{2\pi }}{m}\) + 2φ1 and \(\frac{{2\pi }}{m}\) – 2φ2 where φ1, φ2 > 0.
Lemma 2. Suppose the inequality hy(U1, …, Um) < \({{2}^{y}}m\,{{\sin }^{y}}\frac{\pi }{m}\) + s holds. Then there are constants C, D > 0 depending only on y such that for 0 < s < D we have φ1, φ2 < \(C\sqrt s \). In other words, with small s, the central angles differ from \(\frac{{2\pi }}{m}\) by O(\(\sqrt s \)).
Proof. Consider another polygon, which has the same central angles but these two specified angles are adjacent to each other. Then the lengths of the sides of the polygon and values of the function hy do not vary, and the following inequality holds:
Replace the central angles \(\frac{{2\pi }}{m}\) + 2φ1 and \(\frac{{2\pi }}{m}\) – 2φ2 by two identical angles \(\frac{{2\pi }}{m}\) + φ1 – φ2. With this replacement, the function value could be reduced no more than by s; therefore,
Consider one of the terms. The Taylor expansion with the remainder in the Lagrange form gives
where i ∈ {1, 2} and θi ∈ \(\left[ {\frac{\pi }{m} - {{\varphi }_{2}},\frac{\pi }{m} + {{\varphi }_{1}}} \right]\). Consequently, due to (6) the following inequality holds:
Note that when y < 1 and x ∈ (0, π) we have (sin x)y – 2 ≥ 1 and y cos2 x – 1 < 0, along with the estimate |y cos2 x – 1| > min(|y – 1|, 1). Hence, for y < 0 the constant in (φ1 + φ2)2 can be estimated from below as C1(y) > 0. Consequently, C1(y) (φ1 + φ2)2 < s and then φ1 + φ2 < \(\frac{1}{{\sqrt {{{C}_{1}}(y)} }}\sqrt s \), and this proves Lemma 2.
In what follows, we need the following proposition.
Lemma 3. In terms of Theorem 1, it is true that
Proof. Introduce the notation αk = βk – \(\frac{{2\pi k}}{m}\), α0 = 0, and αm = 0. Recall that βi is uniformly and independently distributed on the interval [0, 2π]. Hence, αk are independent and uniformly distributed on the intervals \(\left[ { - \frac{{2\pi k}}{m},2\pi - \frac{{2\pi k}}{m}} \right]\), respectively. In the new designations, the function hy is presented as follows:
If for small s the condition hy(U1, …, Um) < \({{2}^{y}}m\,{{\sin }^{y}}\frac{\pi }{m}\) + s holds, then by Lemma 2 we have αk = \(O(\sqrt s )\). This allows us to expand each term by the Taylor formula with the remainder in the form of Lagrange at the point \(\frac{\pi }{m}\). We use (4), together with the third-derivative formula for siny(x). To shorten the writing, we put Δk := \(\frac{{{{\alpha }_{k}} - {{\alpha }_{{k - 1}}}}}{2}\). We obtain
where ηk ∈ \(\left[ {\frac{\pi }{m} - 2C\sqrt s ,\frac{\pi }{m} + 2C\sqrt s } \right]\) by Lemma 2.
Since the linear term disappears, the condition h(U1, …, Um) < \({{2}^{y}}m\,{{\sin }^{y}}\frac{\pi }{m}\) + s is equivalent to the relation
In turn, this is equivalent to the inequality
where
Consider the coefficient in the kth summand of the third order
Because \(\left| {{{\eta }_{k}} - \frac{\pi }{m}} \right|\) < \(2C\sqrt s \), for small s this coefficient in modulus is smaller than
therefore, the left-hand side in (7) can be estimated as follows:
Thus, we have
Analogously, we can obtain the inequality
Consider the quadratic form Q(α) = \(\frac{1}{2}\sum\nolimits_{k = 1}^m {({{\alpha }_{k}}} \) – \({{\alpha }_{{k - 1}}}{{)}^{2}}\) = \(\sum\nolimits_{k = 1}^{m - 1} {\alpha _{k}^{2}} \) – \(\sum\nolimits_{k = 2}^{m - 1} {{{\alpha }_{k}}{{\alpha }_{{k - 1}}}} \). Applying the reasoning from [6] and with the help of orthogonal transformation replace this quadratic form by the quadratic form W(Y) = \(\sum\nolimits_{k = 1}^{m - 1} {{{\lambda }_{k}}Y_{k}^{2}} \), where λk = 1 – cos\(\frac{{\pi k}}{m}\), k = 1, …, m – 1. It is clear that all λk > 0; hence, the condition W(Y) < s implies the condition Yk = \(O(\sqrt s )\). Therefore, αk = \(O(\sqrt s )\). It follows that there exists a constant L depending only on the quadratic form Q such that αk < \(L\sqrt s \). We can consider that L ≤ 2C; hence, the condition of the smallness of αk can be removed from under the probability sign. Transforming the previous relations, we obtain the inequality
We estimate the left-hand and right-hand sides of the inequality using the following lemma. The proof of this lemma is analogous to the reasoning from ([6], p. 104).
Lemma 4. For any positive F > δ > 0 and 0 < s < Z(δ) the following equality holds:
where Z(δ) is a certain positive constant depending only on δ.
We put in formula (9) the result of Lemma 4 and obtain the two-sided inequality
Remove the conditions of ordering angles using condition (5) and also divide by \({{s}^{{\frac{{m - 1}}{2}}}}\). We obtain the new two-sided inequality
Let s tend to 0. It is clear that the upper and lower estimates join each other; hence,
By substituting the value of D from (8), we obtain the result of Lemma 3.
Consider for t > 0 the function zn = \({{2}^{y}}m\,{{\sin }^{y}}\frac{\pi }{m}\) + \(t{{n}^{{ - \frac{{2m}}{{m - 1}}}}}\). Then
Suppose s = \(t{{n}^{{ - \frac{{2m}}{{m - 1}}}}}\); then \({{n}^{m}}{{s}^{{\frac{{m - 1}}{2}}}}\) = \({{t}^{{\frac{{m - 1}}{2}}}}\). Check condition (1) from Theorem 2:
The last limiting transition was performed due to Lemma 3. We prove that condition (3) also holds.
Lemma 5. For any r ∈ {1, …, m – 1}, the following relation holds:
Proof. Introduce the notation: βi = \(\angle {{U}_{1}}O{{U}_{{i + 1}}}\) when i ∈ {1, …, 2m – r – 1} and γi = \(\angle {{U}_{{m - r + 1}}}O{{U}_{{i + 1}}}\) when i ∈ {m – r + 1, …, 2m – r – 1}. It is clear that when i ≥ m, we have γi = (βi – βm – r) mod 2π. Suppose I = (i1, …, \({{i}_{{m - 1}}}\)) and J = ( j1, …, \({{j}_{{m - 1}}}\)) are two permutations of m – 1 elements. Introduce the event Q = {\({{\beta }_{{{{i}_{1}}}}}\) < … < \({{\beta }_{{{{i}_{{m - 1}}}}}}\), \({{\gamma }_{{m - r + {{j}_{1}}}}}\) < … < \({{\gamma }_{{m - r + {{j}_{{m - 1}}}}}}\)} and estimate the probability
By Lemma 2 we have \(\left| {{{\beta }_{{{{i}_{k}}}}} - \frac{{2\pi k}}{m}} \right|\) < \(2C\sqrt s \) for i < m. Analogously we obtain \(\left| {{{\gamma }_{{m - r + {{j}_{k}}}}} - \frac{{2\pi k}}{m}} \right|\) < \(2C\sqrt s \). Denote by l the number of the position of the element m – r in the permutation I and by ti, the number of the position of γi in the permutation J. Then for i ≥ m we obtain the inequality
Therefore, probability (10), due to the distribution of βi, can be estimated as \({{\left( {\frac{{4C\sqrt s }}{{2\pi }}} \right)}^{{m - 1}}}{{\left( {\frac{{8C\sqrt s }}{{2\pi }}} \right)}^{{m - r}}}\). Summing this estimate by all permutations and substituting s = \(t{{n}^{{ - \frac{{2m}}{{m - 1}}}}}\), we obtain the inequality
Return to the proof of Theorem 3. We can use Theorem 2, because we proved that its terms are met. Then limn→∞\(\mathbb{P}\)(\(H_{{n,m}}^{y}\) ≥ zn(t)) = \({{e}^{{ - {{\lambda }_{t}}}}}\) for any t ∈ T. Consequently,
Hence, for any t > 0 it is true that
Theorem 3 is completely proved.
Proof of Theorem 4 is similar to the proof of Theorem 3 and therefore is omitted. The condition y ∈ (0, 1) guarantees that the function (sin x)y has the strictly negative second derivative on [0, π] (it can be considered that at the ends of the segment the first two derivatives are determined by continuity); therefore, the strict convexity is replaced by the strict concavity. Then keep the reasoning (here, we change the signs in the probabilities, because now we consider the maximum, not the minimum).
Remark 2. These arguments are not applicable for y = 1, because the second derivative of the function sinyx is not separated from 0 on [0, π]. However, [6] presents a method for the proof of Lemmas 1 and 2 without this fact. Therefore, the obtained constant is applicable also for y = 1.
REFERENCES
W. Lao, Some Weak Limit Laws for the Diameter of Random Point Sets in Bounded Regions, PhD Thesis (Institut für Stochastik, Karlsruhe, 2010).
M. Mayer, Random Diameters and Other U-max-Statistics, PhD Thesis (Univ. of Bern, Bern, 2008).
A. D. Barbour, L. Holst, and S. Janson, Poisson Approximation (Oxford Univ. Press, London, 1992).
W. Lao and M. Mayer, “U-max-statistics,” J. Multivar. Anal. 99, 2039–2052 (2008).
F. B. Silverman and T. Brown, “Short distances, flat triangles, and Poisson limits,” J. Appl. Probab. 15, 815–825 (1978).
E. V. Koroleva and Ya. Yu. Nikitin, “U-max-statistics and limit theorems for perimeters and areas of random polygons,” J. Multivar. Anal. 127, 99–111 (2014).
I. M. Yaglom and V. G. Boltyanskii, Convex Figures (Holt, Rinehart and Winston, New York, 1961).
A. M. Legendre, Elements of Geometry and Trigonometry: With Notes (Oliver & Boyd, Edinburgh, 1822).
Funding
The work was supported by the Ministry of Science and Higher Education of the Russian Federation (agreement no. 075-15-2019-1619).
Author information
Authors and Affiliations
Corresponding author
Additional information
Translated by L. Kartvelishvili
About this article
Cite this article
Simarova, E.N. Limit Theorems for Generalized Perimeters of Random Inscribed Polygons. I. Vestnik St.Petersb. Univ.Math. 53, 434–442 (2020). https://doi.org/10.1134/S1063454120040093
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1134/S1063454120040093