6 views

Uploaded by Mauricio_Fried_8502

save

- Affine Object Tracking Using Kernel-Based Region Covariance Descriptors
- Post cat1.pdf
- Model- vs. design-based sampling and variance estimation
- Lars_Methods of Analysing Variability
- Small World
- 4703-07-Notes-0Ground Penetrating Radar Theory and Applications
- EfficientIndexes.pdf
- Proofs 2
- Autocorrelation_MR
- hw3sol
- Linear Least Squares Problems
- American Computer Science League (ACSL) Rules and Handbook
- Chap 5
- GraphEntropyEnsenadaneu
- Mincostflow Notes
- edgar-kalmanfilter
- 10.1.1.72
- Risk and Return for Portfolio
- Comsol Emt
- Topic03 Correlation Regression
- Kawai_A Flexible Shear Spinning of Axi-symmetrical Shells_JMPT_2007
- 173232298 a Guide to Modern Econometrics by Verbeek 151 160
- 6in Double Petal Shell
- Albert, R-Scale-Free Networks in Cell Biology-2008
- An All-Sky Analysis of Polarization in the Microwave Background
- 0580_w17_qp_43
- ansys_cmd
- Particle
- 10 Mathematics Impq Sa 2 7 Mensuration
- aPDS-II-UNIT -V-RN.pptx
- Bronkitis - Gejala, penyebab dan mengobati - Alodokter.pdf
- Group 9 Guardian Section C
- EGBE.docx
- Acil Tıp algoritma.pdf
- Um Breve Histrico Do Conceito de Funo
- Caso Anterior
- Kelsen
- Sop Penerimaan Pasien Baru
- 3 Tema. Managerii - definirea, activitățile și responsabilitățile.docx
- askep tbc
- 017.xlsx
- Cuenca 1 PDF
- PBL BLOK 28 LBP
- 4. Kommission Die Internationale Gemeinschaft Bekräftigt Ihre Unterstützung Für Die Marokkanische Sahara
- 4362-10903-1-PB.pdf
- El Llanto de Bebés Genera Respuesta Única en Las Personas Investigadores
- 360334450-Visualwize-4-0-PDF-Level-4.pdf
- González Casarrubias Arantza Yetzalen Examen de Word 3
- ACTIVIDAD 1
- Preguntas Septima Sesion 2018
- El Confort Térmico - Arquitectura y Energía
- Fifth Grade Music Curriculum
- Mat+II+Tema+09+Problemas+de+derivadas+(T08+PR)
- ch02 health diet.ppt
- Inequacoes1grau2013.doc
- Barthes_Roland_La_camara_lucida_Nota_sobre_la_fotografia.pdf
- Nueva Zelanda
- CANON DEL CUERO HUMANO
- PRATICA 1 - MEDICOES E ERROS.pdf
- 352171058-Harlem-Nocturne-pdf.pdf

You are on page 1of 17

CO] 9 Dec 2005

**RANDOM GEOMETRIC GRAPH DIAMETER IN THE UNIT BALL
**

ROBERT B. ELLIS, JEREMY L. MARTIN, AND CATHERINE YAN Abstract. The unit ball random geometric graph G = Gd (λ, n) has as its p vertices n points distributed independently and uniformly in the unit ball in d , with two vertices adjacent if and only if their ℓ -distance is at most λ. R p Like its cousin the Erd˝s-R´nyi random graph, G has a connectivity threshold: o e an asymptotic value for λ in terms of n, above which G is connected and below which G is disconnected. In the connected zone, we determine upper and lower bounds for the graph diameter of G. Speciﬁcally, almost always, diamp (B)(1− o(1))/λ ≤ diam(G) ≤ diamp (B)(1 + O((ln ln n/ ln n)1/d ))/λ, where diamp (B) is the ℓp -diameter of the unit ball B. We employ a combination of methods from probabilistic combinatorics and stochastic geometry.

1. Introduction A random geometric graph consists of a set of vertices distributed randomly over some metric space X, with two vertices joined by an edge if the distance between them is suﬃciently small. This construction presents a natural alternative to the classical Erd˝s-R´nyi random graph model, in which the presence of each edge is o e an independent event (see, e.g., [2]). The study of random geometric graphs is a relatively new area; the monograph [9] by M. Penrose is the current authority. In addition to their theoretical interest, random geometric graphs have many applications, including wireless communication networks; see, e.g., [4, 10, 12]. In this article, we study the unit ball random geometric graph G = Gd (λ, n), p deﬁned as follows. Let d and n be positive integers, B the Euclidean unit ball in Rd centered at the origin, λ a positive real number, and p ∈ [1, ∞] (that is, either p ∈ [1, ∞) or p = ∞). Let Vn be a set of n points in B, distributed independently and uniformly with respect to Lebesgue measure on Rd . Then G is the graph with vertex set Vn , where two vertices x = (x1 , . . . , xd ) and y = (y1 , . . . , yd ) are adjacent if and only if x − y p ≤ λ. (Thus the larger λ is, the more edges G has.) Here · p is the ℓp -metric deﬁned by x−y

p

=

|xi − yi |p max{|xi − yi | : 1 ≤ i ≤ d}

d i=1

1/p

for p ∈ [1, ∞), for p = ∞,

where the case p = 2 gives the standard Euclidean metric on Rd . When d = 1, G is known as a random interval graph. (Note that the value of p is immaterial when d = 1.) Random interval graphs have been studied extensively in the literature; the asymptotic distributions for the number of isolated vertices and the number of connected components were determined precisely by E. Godehardt and J. Jaworski [7]. The random Euclidean unit disk graph G2 (λ, n) was studied 2 by X. Jia and the ﬁrst and third authors [5].

1

d d d we deﬁne Bp (r) := Bp (O. y ∈ Vn is deﬁned to be the length of the shortest path between x and y in G. y) between two vertices x. the proof of which builds on the absolute upper bound (Theorem 10 of §6). 2. cf. d The origin of R is O := (0. . Thus B = B2 (1). JEREMY L. p p ∈ [1. We will generally avoid repeating the constraints on the parameters. We are interested in the asymptotic behavior of the connectivity and graph diameter of G as n → ∞ and λ → 0. but also comment along the way on the special case d = 1. diam(G). G has a connectivity threshold : roughly speaking. and its volume vol(X) is its Lebesgue measure. r) := ∪x∈X Bp (x. In fact. when the center of a ball is not explicitly given. We will make frequent use of the quantity (1) αd p d Γ vol(Bp (r)) := = d (r)) vol(B2 Γ p+1 p 3 d 2 d ·Γ 2+d 2 p+d p . where K is a constant depending only on d (Theorem 8 of §5). above the connectivity threshold. r) := {y ∈ Rd : x−y p ≤ r}. ∞] describes the metric. The graph distance dG (x. y ∈ Vn }. or ∞ if there is no such path. and n is the number of vertices. • an “absolute” upper bound diam(G) < K/λ.g. along p with the proofs of several other useful facts about ℓp -geometry. e. The graph diameter of G is deﬁned to be diam(G) := max{dG (x. Y ) between two sets X.2 ROBERT B. The calculation of αd . n). an expression for λ as a function of n.. d d while the ℓp -ball of radius r around a set X ⊆ Rd is Bp (X. . . above which G is connected and below which G is disconnected. Definitions and Notation As mentioned above. ∞]. deﬁned as diamp (X) := sup{ x − y p : x. This graph-theoretic quantity is not to be confused with the ℓp -diameter of a set X ⊆ Rd . AND CATHERINE YAN In the present article. when 2 ≤ p ≤ ∞. where d ≥ 1 is the dimension of the ambient unit ball B. The ℓp -ball of radius r centered at x ∈ Rd is deﬁned as d Bp (x. [11]). The ℓp -diameter of B is diamp (B) = 2d1/p−1/2 2 when 1 ≤ p ≤ 2. . (This behavior is ubiquitous in the theory of the Erd˝s-R´nyi random graph model.) We are interested primarily in the combinatorial graph diameter of G. Y ⊆ Rd is deﬁned as inf{ x− y p : x ∈ X. we focus on the case d ≥ 2 and p ∈ [1. y ∈ Y }. may be found in the Appendix at the end of the article. y) : x. 0). ·Γ where Γ is the usual gamma function (see. The boundary ∂X of X is its closure minus its interior (in the usual topology on Rd ). The distance dp (X. o e [2]. Our results include • a lower bound for diam(G) (Proposition 7 of §4). the main object of our study is the random geometric graph G = Gd (λ. y ∈ X}. MARTIN. r). r). ELLIS. . λ > 0 is the ℓp -distance determining adjacency. and • an asymptotically tight upper bound within a factor of the form (1 + o(1)) of the lower bound.

Let λ1 (n) = 2 n (log n + c + o(1)). λ) ⊆ B. where c is a constant. n) is conp p nected. 1 − d1/2 λ). Penrose. . Theorem 1 (Godehardt. Connectivity thresholds In order for G = to have ﬁnite diameter. λ)∩ B. p Suppose γ(n) is a nonnegative sequence such that limn→∞ γ(n) → ∞. by replacing c in Theorem 1 with a nonnegative sequence γ(n) → ∞. Let d ≥ 2. . for some absolute nonnegative constant c. ∞]. and let α = αd be the constant of (1). n) has no isolated vertices. Otherwise. G = Gd (λ. and that λ ≥ 1 αn 2(d − 1) 2 ln n + ln ln n + γ(n) d d 1/d . 3. λ) ∩ B is not less than 2 vol(Bp (vi . For this case we now quote parts of Theorems 10 and 12 of [7]. p Proof. then Bp (vi . we will use the fact that the connectivity threshold coincides with the threshold for the disappearance of isolated vertices. almost always G1 (λ1 (n). we mean. First we compute the threshold for isolated vertices. Proposition 2. Then n→∞ 1 n (log n Gd (λ. we seek a connectivity threshold – a lower bound on λ so that G is almost always connected. vi is isolated if and only if there are no other vertices in Bp (vi . vn } be the vertex set of G. Set X = X1 + X2 + · · · + Xn . For d ≥ 2. The case d = 1 is exceptional in that the thresholds for having isolated vertices and for connectivity are separated. n) p + c + o(1)) and λ2 (n) = −c lim Pr[G1 (λ1 (n). If vi − O 2 ∈ d [0. Xi = 1 if Ai occurs and 0 otherwise. By the notation f (n) = o(g(n)) and f (n) = O(g(n)).. lim Pr[G1 (λ2 (n). We will show that E[X] = o(1).a. Hence Pr[Ai ] ≤ (1 − d1/2 λ)d (1 − αλd )n−1 + (1 − (1 − d 1/2 d αλd (1 + O(λ)) λ) ) 1 − 2 n−1 . v2 . then 1 d d the volume of Bp (vi . n)has an isolated vertex ] = e−e p n→∞ . . and let Xi be the indicator of Ai . Therefore. if vi − O 2 ∈ (1 − d1/2 λ. to which we refer the reader for their precise determination of the asymptotic Poisson distributions of the number of isolated vertices and the number of connected components. which follows from two theorems of M.RANDOM GEOMETRIC GRAPH DIAMETER 3 We will say that the random graph G has a property P almost always. . let Ai be the event that vi is an isolated vertex. . all ℓp -metrics are identical. let p ∈ [1. or a. n)is connected ] = e−e p −c In particular. it must be connected. Then. respectively. which is easier to calculate. Let Vn = {v1 . . When d = 1. limn→∞ f (n)/g(n) = 0 and lim supn→∞ f (n)/g(n) ≤ c. n) has no isolated vertices and G1 (λ2 (n). Jaworski). that is. if n→∞ lim Pr [G has property P ] = 1. λ))(1 + O(λ)). almost always. We condition Pr[Ai ] on the ℓ2 -distance from vi to the origin O. For each vertex vi . 1]. d By deﬁnition.

and so Pr[X > 0] ≤ E[X] ≤ o(1). The number of isolated vertices below the threshold is easy to compute in certain special cases. For example. in particular. with the thresholds for G deﬁned as in (2a) and (2b). For d ≥ 2 and p ∈ (1. κ ≥ k + 1) and ρ(H. we have Pr[Ai ] ≤ (1 + o(1))e−αnλ + (dd1/2 λ + O(λ2 ))(1 + o(1))e−αnλ d d /2 . AND CATHERINE YAN Using 1 − x = e−x (1 + o(1)) as x → 0 and the binomial expansion. Sections 2 and 5 of [8]). JEREMY L. Let d ≥ 2 and p ∈ (1. n). it suﬃces to concentrate on the values of λ for which G has no isolated vertices. κ ≥ k + 1) = ρ(H. to determine the behavior of X more exactly would require d complicated integrals that describe the volume of Bp (vi . almost always. By linearity of expectation. Theorem 3 asserts that as λ increases (forcing more edges into the graph). and so E[X] ≤ o(1) + d3/2 nλ(1 + o(1))n−1+1/d (ln n)−1/d e−γ(n)/2 . ∞]. Theorem 1] yields X = (1 + o(1))n1−αc almost always. G is connected. ∞]. Let p ∈ (1. then a minor modiﬁcation of [5. so we have as an immediate corollary the following. We quote Penrose’s theorem [8. κ > k + 1) = min{λ | H has vertex connectivity κ ≥ k + 1}. However. Applying these results to G. MARTIN. The proof requires only a series of geometric and probabilistic arguments which hold in the unit ball as well as in the unit cube (see. ρ(H. λ) ∩ B near the boundary of B (cf. G has no isolated vertices. For our purposes. Corollary 4. Then. δ ≥ k + 1) are the same. δ > k + 1) = min{λ | H has minimum degree δ ≥ k + 1}. Chapter 8]). the connectivity threshold for the unit-cube random geometric graph coincides with the threshold for lacking isolated vertices. Theorem 7. 1. n) has no isolated vertices. Penrose shows that the limiting probability distributions for ρ(H. Here Theorem 3 does not apply. G = Gd (λ(n). 1]d rather than B. Then n→∞ When k = 0. [9. The second term is o(1). deﬁne (2a) (2b) ρ(H. ∞].1] after some supporting deﬁnitions. d = 2.7 provides a threshold for connectivity. lim Pr[ ρ(H. The ﬁrst term is o(n−1 ) for d ≥ 2. we can appeal to two general results about the behavior of a random geometric graph in an ℓp -metric space whose boundary is a compact (d−1)submanifold of Rd For such a graph. almost always. and 0 ≤ c < α−1 .2 of [9] provides a threshold for the disappearance of isolated vertices. except that its p vertices are points in [0. we obtain the following fact. Theorem 3 (Penrose). if p ∈ [1. Thm. and Theorem 13. that is. In the proof of Theorem 3 in [8]. E[X] = n · Pr[Ai ]. Deﬁne the unit d cube geometric graph H = Hp (λ. and let λ = λ(n) be suﬃciently large so that. We now consider the case that d ≥ 2 and p = 1. n) analogously to G = Gd (λ. δ ≥ k + 1) ] = 1. For any nonnegative integer k. H becomes connected simultaneously as the last isolated vertex disappears. In general. λ = c ln n/n. .4 ROBERT B. ∞] and let k ≥ 0 be an integer. almost always. almost p always. ELLIS.

p ∈ [1. κ ≥ k + 1)d log n = lim n→∞ nα ρ(G. ∞]. almost always.RANDOM GEOMETRIC GRAPH DIAMETER 5 Proposition 5. On the other hand. and assertion (ii) is implied by Proposition 5. almost always. and disconnected if c < (2(d − 1)/(dα)). Assertion (i) follows from combining Proposition 2 with Corollary 4. d We now collect the above results to present the connectivity thresholds that we will use in the rest of the paper. ∞]. chosen as in Figure 4. perpendicular to the line joining a and −a. Then. and G = Gd (λ(n). and λ = (c ln n/n)1/d for some constant c > 0. the lower bound in (ii) is implied by the stronger bound in (i). n→∞ lim nα ρ(G. and λ = 2(ln n + γ(n))/n. and let ±C be the spherical cap formed by slicing B with hyperplanes at distance h from ±a respectively. (When p > 1 and d ≥ 2. δ ≥ k + 1)d log n = 2(d − 1) . almost always. That is. Let G = Gd (λ. p ∈ [1. Proof. Proposition 7 (Diameter lower bound). γ(n) is a nonnegative sequence such that γ(n) → ∞ and λ ≥ 1 αn 2(d − 1) 2 ln n + ln ln n + γ(n) d d 1/d . p p (i) Suppose that d ≥ 2. (ii) Suppose that d ≥ 2. G = Gd (λ. p ∈ (1. (iii) Suppose that d = 1. Let d ≥ 2. p ∈ [1. Then. Let d ≥ 1 and p ∈ [1. If h = h(n) satisﬁes p (3) then. Then. ∞]. which together with the condition (3) on h implies that Pr[A] = o(1). n→∞ lim h(d+1)/2 n = ∞. and suppose that λ = λ(n) is suﬃciently large so that Theorem 6 guarantees that almost always. B will usually contain two vertices whose ℓp -distance is (asymptotically) diamp (B). ∞]. G almost always . almost always. almost always. G is connected if c > (2(d − 1)/(dα)). n) and let α = αd be the constant of (1).3. Then Pr[A] = 2 Pr[C ∩ Vn = ∅] = 2 1 − vol(C) vol(B) n ≤ 2 exp −n vol(C) vol(B) . Theorem 6. Let α = αd be the p p constant of (1). and let k ≥ 0 be an integer. Then. ∞]. n). n) is connected. vol(C)/vol(B) = O(h(d+1)/2 ) by (18) of §A. Let A be the event that at least one of the two caps ±C contains no vertex of Vn . the diameter of G will almost always be at least diamp (B)(1 − o(1))/λ. if p ≥ 2.) Assertion (iii) is implied by Theorem 1 and its accompanying remarks. The precise statement is as follows. G is connected. diam(G) ≥ 1−h diamp (B) = λ 2(1 − h)d1/p−1/2 /λ 2(1 − h)/λ if p ≤ 2. G is connected if γ(n) → ∞ and disconnected if γ(n) → −∞. Therefore. Let ±a be a pair of antipodes of the unit ball B. 4. A lower bound for diameter When G is connected.

λ/4) ∩ Vn = ∅.y (k) and connects x and y. but dependent on d. y) ≤ λ The proof is based on Proposition 9 below. almost always. Note that for all d ≥ 1. . Proposition 7 can be strengthened by identifying a collection of mutually disjoint antipodal pairs of caps of height h and showing that. Abbreviate Tx. Let d ≥ 2. diam(G) ≤ Kdiamp (B)/λ. y ∈ Vn is at most K x − y p /λ. since for any two points x. and suppose that λ = λ(n) is suﬃciently large so that Theorem 6 guarantees that almost always. dG (x. n→∞ d d Tx. n) is connected. MARTIN. y) between two vertices x. As a consequence. The proof of our next result uses ingredients from [9. First we construct a connected subset P ⊆ Tx. we cover the unit ball B with d-dimensional cubes. there exists a constant k0 > 0. y ∈ Rd . h can be chosen to satisfy both (3) and limn→∞ h/λ = 0. y ∈ Vn such that (i) at least one point is inside B2 (O. we have x − y 2 ≤ d1/2 x − y p .y (k) by Tx. y ∈ Vn there exists a constant K independent of n and p such that as n → ∞. This fact will be used repeatedly in the proof of the tight upper bound in Theorem 10 of §6.y such that d (i) Bp (P. For any two vertices x. lim Pr[An (k)] = 0. and for each z ∈ Ld . but does guarantee a short path between any pair of vertices. 285]. without loss of generality. For the case p = 2. y ∈ Vn . both caps in at least one pair contain a vertex. where K > 0 is a constant independent of n and p. Such a collection corresponds to an antipodally symmetric spherical code (see [3]). Also. 1 − (k + d)λ). where ǫ = 1/(4d). kλ) ∪ B2 (y. Theorem 8.y . First. Then for p any two points x. Proposition 9. denote the closed cube centered at z by Qz . supfor d pose y ∈ B2 (O. and d (iii) Bp (P. such that for all k > k0 . almost always. It is suﬃcient to prove the following Theorem 8.6 ROBERT B. if the limit in (3) is a nonnegative constant. adapted and extended for our present purposes. and (ii) there is no path of G that lies in Tx. Under the same assumptions as in Theorem 8. Suppose An (k) occurs√ a pair of vertices x. JEREMY L. 5. λ/4) ⊆ B. Proof. ELLIS. This will not be strong enough to meet (asymptotically) the lower bound in Proposition 7. G = Gd (λ. p. that is. let Thus Tx. Let Ld be the set of centers of these cubes. then limn→∞ Pr[A] > 0. K x−y 2 . kλ)) ∩ B. The absolute upper bound In this section we prove that when G is connected. AND CATHERINE YAN contains a vertex in each of C and −C.y is a “lozenge”-shaped region. vertices are not guaranteed in both caps. each of side length ǫλ. Let An (k) be the event that there exist √ d two vertices x. The result now follows from the deﬁnition of G and the lower bound (19) on the ℓp -distance between C and −C. the graph distance dG (x. √ (ii) diam2 (P ) ≥ (k − d)λ. Step 1. y. 1 − (k + d)λ).y (k) = convex closure of (B2 (x.

y is simply connected. so that T contains x. y ∈ S. T ⊆ Tx. Then y is disconnected from x in G. If dp (P1 . Bp (P1 . As constructed. we show that there is a point w on ∂Tx. the probability d Pr[Qz ∩ Vn = ∅ for every Qz ⊆ Bp (P.y from x to y must pass through P1 . C2 ) ≥ λ/2. w) ≤ λ/2.y .RANDOM GEOMETRIC GRAPH DIAMETER 7 ∂B w P1 x u y ∂Tx. Since P is connected. λ/2) is (topologically) disconnected with x and y lying in diﬀerent connected components. λ/2) ∩ Vn = ∅. √ and derive from this that diam2 (P1 ) ≥ (k − d/2)λ provided that the ℓ2 -distance between u and w is at least kλ (see Figure 1 for an illustration). λ/2) containing x.1]. assume S. w) ≤ λ/2. Let ω be the set of points z ∈ Ld such that Qz ∩ P = ∅. We now show that when k is large enough. since Tx. and diam2 (P ) ≥ diam2 (P1 ) − λ/2 ≥ √ (k − d)λ.y Figure 1. C2 ) < λ/2. by deﬁnition of unicoherence. d Let VT = Vn ∩ Tx. without loss of generality.y \ Dx containing y. which happens with probability tending to zero by Theorem 6. The lozenge Tx.y . d Let Dx be the connected component of Bp (VT . To this end. To overcome this. Step 2. Furthermore. then P1 := S ∩ T is connected. let C1 = ∂B ∩ Tx. It follows that there is a d point w ∈ C2 such that dp (P1 . that is. P1 intersects the line segment joining x and y. and must also satisfy dp (P1 . In particular. we must avoid the case that w lies in the boundary of B. so it is unicoherent [9. Then Bp (VT . we let P be obtained from P1 by moving every point toward O by λ/4 under the transformation x → x − (λ/4)(x/ x 2 ). and their union is Tx. then C2 must be a subset either of S or of T .y be the set of vertices of G lying in Tx. and P1 separates x and y.y . Next. P1 may be too close to the boundary of B so that some cube Qz intersecting P1 might not lie entirely inside B. and let C2 = ∂Tx. ω is a ∗-connected subset of Ld .y \ ∂B. by deﬁnition of P1 as the intersection of S and T . To achieve this. Then both S and T are connected.y is the union of closed connected sets S. almost always. any path in Tx. Let T be the closure of Tx.y . The frontier P1 must intersect the line segment between vertices x. dp (P1 .y such that dp (P1 . w) ≤ λ/2 for some point w on the boundary ∂Tx. Hence. that is P satisﬁes conditions (i–iii). the union of the corresponding . Then P is connected.y \ C1 . y ∈ Vn at some point u. Since x ∈ T . Let u be one of the intersection points. λ/4)] tends to zero. Let S be the closure of the connected component of Tx.y \ S. Lemma 9.

The order bound in (5) is immediate by geometric series.i denote the collection of ∗-connected sets of ω ⊆ Ld of cardinality i. (k + d/2)λ).y has length g.y . AND CATHERINE YAN P x y Figure 2. then.3]). λ/4). and the resulting quantity is o(1) √ √ provided k > d + dα vol(B)(4d)d /(4 d(d − 1)). for example.y (k). where c is a constant and α = αd is the constant of (1). [9. It is known that the number of ∗-connected subsets of Zd of cardinality i containing the d origin is at most 23 i (see. and each must lie inside √ √ d d the convex closure of B2 (x. d d we have Cd. Then the ℓp -balls of radius λ/2 around every other vertex in the path must be pairwise disjoint. Since |Ld | ≤ (2/(ǫλ))d . Therefore Pr[An (k)] ≤ ≤ (4) (5) ≤ √ √ i≥4 d(k− d)) ω∈Cd. To justify inequality (4). we have 3d < ǫd (d − 1) ln n/(dα vol(B)). 1 − (k + d)λ). hence Qz ⊆ Bp (P. y lies in B2 (O. For each z ∈ ω. y ∈ Vn which are not connected by any path in Tx. (k + d/2)λ) ∪ B2 (y. MARTIN. Lemma 9. Let x and y be √ vertices two d in Vn with x − y p > λ. We show that the probability of such an event is o(1). and the “frontier” P separating them.i ≤ (2/(ǫλ))d 23 i ≤ 2d (ǫλ)−d e3 i . Proof of Theorem 8. By considering the ℓ2 √ √ diameter of P . Suppose the shortest path between x and y in Tx. Hence we d) √ have a ∗-connected subset ω ⊆ Ld with cardinality at least 4 d(k − d) such that Qz ∩ Vn = ∅ for all z ∈ ω. Let Cd. p when n is suﬃciently large. when d = 2. If at least one of x.i Pr[Vn ∩ (∪z∈ω Qz ) = ∅] i (ǫλ)d vol(B) n √ √ i≥4 d(k− d)) 2d (ǫλ)−d exp(3d i) 1 − = O n √ √ i≥4 d(k− d)) √ √ 1−(4 d(k− d)(d−1))/(dα vol(B)(4d)d ) cn exp(−iǫd (d − 1) ln n/ (dα vol(B))) . Two vertices x. The gray squares are the ∗-connected subset ω intersecting P of the set of squares covering B. ELLIS. JEREMY L. Fix k > k0 as in Proposition 9.8 ROBERT B. we see that ω contains at least 4 d(k − √ points. which proves the existence of k0 in the proposition. almost always. there is a path of G connecting x and y in Tx. set of cubes is (topologically) connected (see Figure 2). By comparing the . we have d Qz ∩ P = ∅ and ǫ ≤ 1/(4d1/p ).

where K1 and K2 are constants independent of n and p. y. λ/2). Suppose that λ = λ(n) is suﬃciently large so that Theorem 6 guarantees that. let r = max{(αd )1/d d. 1 − (k + d)λ) ∩ B(z. ∞]. y)/λ. y) ≤ 2|x − y|/λ. λ/2) ≤ vol B2 (x. That is. and suppose that λ = λ(n) is suﬃciently large so that Theorem 6 guarantees that almost always. B(x4 . almost always. Theorem K(d 8 follows from the fact that d2 (x. 6. so that the pin provides a “highway” through G. x1 )/λ steps. must be pairwise disjoint (else some xi is redundant). A pin consists of a collection of evenly spaced. diam(G) ≤ diamp (B)(1 + O((ln ln n/ ln n)1/d ))/λ. 1 − (k + d)λ). . G = Gd (λ. Hence |x−y| ≥ ⌈dG (x.y) = y. Having done this. almost always. and then from x1 to y in Kd2 (x1 . y)/λ steps. Let d ≥ 2 and p ∈ [1. . G1 (λ. The following deﬁnitions are illustrated in Figure 3. almost always. (k + d + 2r)λ). 1 − (k + d)λ)√ √ path of bounded length. 1] ⊂ R. B(x2 . y) · vol B2 (x. (k + d/2)λ) . √ which implies that g ≤ K1 + K2 d2 (x. (k + d/2)λ) 2 √ d−1 +d2 (x. we construct a pincushion so that every point of B is reasonably close to an ℓp -ball in one of its constituent pins. and then on to y. we can ensure that each intersection of consecutive balls contains a vertex in Vn . √ then we can travel from x to via a an intermediate vertex x1 just inside B(O. we improve the upper bound in Theorem 8. n) is connected. we can ﬁnd a point x1 ∈ Vn inside B(O. we obtain √ g d d vol Bp (x. y)/λ ≤ (K1 d + K2 )d2 (x. Applying this observation with z = x. x1 ) ≤ (k + d + 2r)λ. reducing the constant K to diamp (B) (asymptotically). The total length of the path is no more than √ 2 (x. .√ d If both x and y lie outside B2 (O. Then √ √ Pr[En (k)] ≤ n 1 − (1 − (k + d)λ)d (1 − ( dλ)d )n = o(1). p Then as n → ∞. λ/2). The proof uses the geometric ingredients of pins and pincushions. To this end. By the preceding argument we can ﬁrst travel from x to x1 in Kd2 (x. the balls B(x0 . (k + d + 2r)λ)) = ∅. n) is connected.RANDOM GEOMETRIC GRAPH DIAMETER 9 volume of the ℓp -balls of radius λ/2 to the volume of Tx. 1 − (k + d)λ) and V √ Vn ∩ (B(O. when 2 ≤ p ≤ ∞. The asymptotically tight upper bound In this section. 1 − √ √ (k + d)λ) ∩ B(x. Our main result is as follows: Theorem 10. diam(G) ≤ 2d1/p−1/2 + O (ln ln n/ ln n)1/d 2 + O (ln ln n/ ln n)1/d /λ /λ when 1 ≤ p ≤ 2. y) + 2d2 (x. overlapping ℓp -balls whose centers lie on a diameter of the Euclidean unit d-ball B. . y)/2⌉λ. the shortest path p between them clearly consists of a strictly increasing set of vertices x = x0 < x1 < x2 < · · · < xdG (x. d}.y . We brieﬂy discuss the case that d = 1. λ/2). and from this it is not hard to deduce that dG (x. x1 ))/λ. and let En (k) p √ be the event that there is a vertex z ∈ √n such that z ∈ B(O. so that B is the interval [−1. For any two vertices x. Moreover. By making suitable choices for the geometry.

(a) A pin in the unit circle in R2 . we must choose r so that the petal has positive volume: for example. Certainly. one centered at each point um . . . λ/2) : x. Fix d ≥ 2. Fix d ≥ 2. Unfortunately. σ. that is. . θ ∈ (−π/2. λ/2) ∩ Bp (y. x−y 2 = r}. π/2]d.. Deﬁnition 11 (Pins). A petal is nonempty if it contains a vertex of Vn . xd ) ∈ Rd . Deﬁnition 13 (Petals). A petal is the region of intersection of two overlapping ℓp -balls on U (the shaded regions in Figure 3). r. 2σ 2σ 2σ . θ. . 1 ≤ p ≤ ∞. AND CATHERINE YAN 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 111 000 1111 0000 111 000 111 000 111 000 111 000 111 000 1111 0000 111 000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 111 000 111 000 111 000 111 000 111 000 111 000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 111 000 111 000 111 000 111 000 111 000 111 000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 111 000 111 000 111 000 111 000 111 000 1111 0000 111 000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 r u1 u0 O O B 111 000 111 000 111 000 1111 0000 1111 0000 111 000 1111 0000 1111 0000 111 000 1111 0000 1111 0000 111 000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 111 1111 111 000 0000 111 1111 111 0000 000 000 0000 000 1111 000 111 000 111 111 000 111 1111 000 000 0000 000 111 111 000 111 000 111 111 000 111 000 111 111 000 000 1111 0000 111 111 000 000 1111 0000 111 111 000 000 1111 0000 1111 0000 1111 0000 111 111 000 000 1111 0000 1111 0000 1111 0000 111 111 000 000 1111 0000 1111 0000 1111 0000 111 111 000 000 1111 0000 111 000 1111 0000 111 000 1111 0000 1111 0000 1111 0000 111 000 1111 0000 111 000 1111 0000 111 000 1111 0000 1111 0000 111 000 111 000 1111 0000 111 000 1111 0000 111 1111 000 0000 111 1111 000 0000 111 000 1111 0000 111 000 1111 0000 111 000 111 1111 000 0000 111 000 111 000 111 000 1111 111 1111 111 111 0000 000 0000 000 000 0000 111 000 1111 0000 111 000 1111 111 111 111 000 000 000 111 000 111 000 1111 0000 111 111 000 000 111 111 000 000 111 000 111 000 1111 0000 111 1111 111 111 000 0000 000 000 111 000 111 1111 111 111 111 000 0000 000 000 000 1111 0000 111 1111 000 0000 1111 111 111 0000 000 000 111 000 1111 111 111 000 111 111 000 000 111 1111 0000 1111 111 111 0000 111 111 111 0000 000 000 111 000 0000 000 000 000 111 1111111 000 0000000 1111 0000 1111 000 000 111 0000 000 000 000 1111 0000 1111 111 111 000 1111 0000 111 111 000 111 1111111 000 0000000 111 000 1111 0000 1111 1111 0000 1111 0000 111 000 1111 1111 0000 111 000 111 1111 0000 000 0000 1111 111 1111111 000 0000000 111 000 1111 0000 1111 0000 1111 0000 111 0000 000 0000 111 000 1111 0000 111 1111 0000 000 0000 0000 111 1111111 000 0000000 111 000 1111 0000 1111 0000 1111 0000 1111 0000 0000 1111 111 1111 000 1111 1111 0000 1111 0000 111 000 111 1111111 000 0000000 111 000 1111 1111 0000 1111 0000 111 0000 000 1111 1111 0000 111 000 111 000 1111 0000 111 111 111 000 000 000 111 111 0000 111111 000000 1111 111 1111 0000 000 0000 1111 111 111 000 000 111 111 000 000 0000 111 111 000 111 1111 000 000 000 0000 111 111 111 111 000 000 111 111 000 000 000 111 111 000 111 1111 000 000 000 0000 000 000 111 000 111 111 000 000 000 111 111 111 000 111 000 111 000 111 111 111 000 000 000 111 000 111 000 111 1111 111 111 1111 000 0000 000 000 0000 111 000 111 111 111 000 000 000 111 000 111111 000000 111 1111 111 111 1111 000 0000 000 000 0000 111 000 111 000 1111 0000 1111 0000 111 1111 000 0000 111 1111 000 0000 1111 0000 1111111 0000000 111 000 1111 0000 111 1111 000 0000 111 1111 000 0000 1111 0000 1111111 0000000 111 000 1111 0000 1111 0000 1111 0000 1111 0000 0000111 111 000 1111 0000 1111 0000 111 000 1111 1111 0000 1111 0000 1111000 0000111 111 000 1111 0000 1111 0000 111 000 1111 1111 0000 1111 0000 1111 1111 0000000 111111 0000 000000 0000 1111 0000 1111 0000 111 000 1111 0000 1111 0000 1111 0000 111111 1111 000000 0000 111 000 111 111 1111 000 000 0000 111 000 111 000111 111 000 1111 000 111 0000 111 000 111 000 000 000 111 000 111 000 111 111 000 111 111 000 000 111 000 111 111 000 000 111 000 1111 0000 1111 0000 111 000 111 111 000 000 111 000 1111 0000 1111 0000 1111 0000 1111 0000 111 000 1111 0000 1111 0000 1111 0000 1111 0000 1111 1111 0000 0000 1111 0000 1111 0000 1111 1111 0000 0000 1111 0000 1111 0000 1111 0000 Figure 3. The petals are the shaded regions. Here p = 1. (b) A pincushion consisting of several pins. θ. the volume is diﬃcult to calculate exactly. so that the ℓp -balls are diamonds. p ∈ [1. which depends in turn on the parameters of the corresponding pin. it suﬃces to take r ≤ (diam2 (B)/diamp (B))λ. MARTIN. and σ ∈ Z+ . ∞]. put r um = um (r. π/2]. . π 2π (2σ − 1)π . x2 = sin θ i=3 sin φi . . θ) = + rm · (x1 . ELLIS. λ/2) := inf{vol Bp (x. p. The probability that a petal is nonempty depends on its volume. λ) : θ. . λ) is the set of (2σ)d−1 pins U := U (d. xj = cos φj i=j+1 sin φi for 3 ≤ j ≤ d. JEREMY L.. r > 0. d d d ξ = ξp (r. 2 where d d d x1 = cos θ i=3 sin φi . λ) consists of the points {um : m ∈ Z} ∩ B. . For m ∈ Z. . φ = (φ3 . . Let U be a pin. Even ﬁnding the minimum volume over all angles. φi ∈ 0.. φ. Deﬁnition 12 (Pincushions). r. y ∈ B.10 ROBERT B. φ. with angle θ. r. φd ) ∈ [0. 1 Note that the total number of ℓp -balls is 1 + 2⌊ r ⌋. p. The corresponding pincushion (with parameters d. and r is the ℓ2 -distance between consecutive centers (such as u0 and u1 ) of ℓ1 -balls. and λ.. together with a collection of ℓp -balls of radius λ/2. p. The corresponding pin U (d.

p ∈ [1. r. Let d ≥ 1. when 2 ≤ p ≤ ∞. d The same argument implies that z ∈ Bp (y.RANDOM GEOMETRIC GRAPH DIAMETER 11 requires integrals that are not easily evaluated (although for ﬁxed d and r. Then. That is. The bound on ξ is then a simple application of (14). together with (17). Lemma 14. ∞]. z ∈ Bp (x. implies that z − x p ≤ λ/2. when 2 ≤ p ≤ ∞. vol(B) Proof. (7) Pr[τ ≥ T ] ≤ E[|{U : τU ≥ T }|] ≤ (2σ)d−1 · Pr[τU ∗ ≥ T ]. λ/2) ∩ Bp (y. we work with the pincushion U deﬁned by (6a) and (6b) where (6c) ρ = ρ(n) = 2d1/p−1/2 (ln ln n/ ln n)1/d 2(ln ln n/ ln n)1/d when 1 ≤ p ≤ 2. τ ≤ σ (d−1)/2 2 exp r −nξ . . it is certainly true that ξ = Θ(λd )). ρ as just deﬁned. λ/2). By the triangle inequality. y ∈ B with x − y 2 = r. Lemma 15. Denote the right-hand side of the desired inequality by T . almost always. r′ ). and deﬁne τ = τ (n) = max{τU (n) : U ∈ U}. Let z ∈ B2 ( x+y . − r d 2 λ 1/2−1/p 2d r d λ 2 − 2 when 2 ≤ p ≤ ∞. We ﬁrst calculate an upper bound on τ . r = r(n) = λd1/2−1/p (1 − ρ(n)) λ(1 − ρ(n)) when 1 ≤ p ≤ 2. and the parameters σ. d Proof. By linearity of expectation. The easiest way to ﬁnd a lower bound for ξ is to inscribe another ℓp -ball in the petal. where r′ = In particular. With the assumptions of Theorem 10. for all x. d B2 x+y ′ 2 . 2 z−x 2 ≤ z− x+y 2 2 + −x 2 ≤ r′ + r . when 2 ≤ p ≤ ∞. ξ ≥ vol(B) λ 1/2−1/p 2d λ r 2 − 2 − r 2 when 1 ≤ p ≤ 2. 2 d which. x+y 2 when 1 ≤ p ≤ 2. λ/2). σ = σ(n) = ⌊(ln n)1/d ⌋ Let τU = τU (n) be the number of empty petals along the pin U ∈ U. λ > 0 and 0 ≤ r ≤ (diamp (B)/diam2 (B))λ. λ/2). For the remainder of this section.r d d ⊆ Bp (x.

Now U ∗ contains at most 2/r petals. (9) and (10) gives (11) var[X] ≤ + By Chebyshev’s inequality (cf. JEREMY L. Proof of Theorem 10. ELLIS.12 ROBERT B. we walk along the appropriate pin to points x2 . By deﬁnition of τ . When n is suﬃciently large. which implies the desired result. Xj ) for i = j is cov(Xi . Let Xi be the indicator random variable of the event that the ith petal of U ∗ is empty. Xj ). Let Rx be the (possibly empty) petal nearest to x. MARTIN. respectively. y − y1 2 ≤ dπ + r(τ + 1). there is another vertex x1 ∈ Vn . From each of x1 . y 2 ≥ (τ + 3/2)r. Now Pr[τU ≥ T ] = Pr[X ≥ T ]. Repeat these constructions for y to obtain an analogous vertex y1 . where the covariance cov(Xi . y ∈ Vn . 2σ . substituting this last bound into (7) gives Pr[τ ≥ T ] = o(1). Let x. vol(B) 4 o(1) · exp r2 −2nξ . (9) var[X] ≤ E[X] + i=j E[X] ≤ 2 r 1− ξ vol(B) n ≤ 2 exp r −nξ . the distance from x to Rx is at most dπ/2σ. but is closer to the origin. and let X = i Xi . vol(B) Combining (8). as well as any “detours” needed in case there are missing edges in the paths along the pins. Without loss of generality. We will then use Theorem 8 to construct a path from x2 to y2 . so by linearity of expectation. vol(B) cov(Xi . y2 near the origin and belonging to petals on the same pin as x1 and y1 . belonging to petals of the pincushion U. Therefore. We can now prove the main result of this section. [1]) and the bounds for E[X] and var[X] in (8) and (11). which also lies in a petal on Ux . and let Ux be the pin containing Rx . We will ﬁnd vertices x1 and y1 near x and y respectively. so that x − x1 2 ≤ dπ/2σ + r(τ + 1). Xj ) 1− (10) ξ vol(B) 2n = Pr[Xi Xj = 1] − Pr[Xi = 1] · Pr[Xj = 1] ≤ o(1) · exp 2 exp r −nξ vol(B) −2nξ . we may assume that (12) x 2. The justiﬁcation for this is deferred until the end of the proof. Pr X ≥ T ≤ ≤ Pr |X − E[X]| ≥ T − E[X] var[X] = o (T − E[X])2 1 σ d−1 . AND CATHERINE YAN where U ∗ is chosen so as to maximize Pr[τU ≥ T ]. (8) Also. Then x − x1 2 . y1 .

there is a vertex x2 that belongs to a petal Rx2 on Ux (indeed. and the bounds on ξ and τ (Lemmas 14 and 15). by Theorem 8. y) ≤ o(ρ/λ) + dG (x2 . almost always. then by Theorem 8. y) ≤ ≤ (13) ≤ (2τ + 3)rK . d lying on the same side of O along Ux ) with Rx2 ⊆ B2 (O. y2 ) + dG (y2 . λ λ λ diamp (B) x diam2 (B) + y 2 + O(ρ) . . x1 ). vol(B) ln n n where the second inequality follows from the assumption that λ is above the threshold for connectivity in Theorem 6. y) ≤ 2 and the theorem follows. Moreover. v) ≤ (2τ + 3)rK/λ = o(ρ)/λ. We now explain the assumption (12). If x 2 . In the worst case. we have dG (x1 . λ Plugging this bound for τ into (13) gives dG (x. y2 ) ≤ dG (x. r(τ + 3/2)). λ By deﬁnition of τ . it follows that ln ln n ln ln n ξ ≥ λd ≥ . x2 − y2 2 ≤ (2τ + 3)r. all empty petals in Ux occur non-consecutively between x1 and x2 . y) ≤ (diamp (B)/diam2 (B))( y 2 + O(ρ))/λ. dG (x. x2 ) + dG (x2 . x2 ) ≤ ≤ x1 2 2Krτ + r λ dπ 2Krτ x 2 + + . if x 2 ≤ d (τ + 3/2)r ≤ y 2 . y) x 2+ y 2 dπ dπ K + + (8τ + 5)r + r σ λ σr diamp (B) x diam2 (B) 2 + y 2 + O(ρ) + O(τ ) + O λ (ln n)−1/d λ . Therefore τ ≤ ≤ 2σ (d−1)/2 exp r −nξ vol(B) O (ln n)−(d+1)/(2d) 4 o(ρ) (ln n)(d−1)/(2d) (ln n)−1 = = . r 2σr λ The same construction goes through if we replace the x’s with y’s. dG (x. By the deﬁnitions of r and ρ given in (6b) and (6c).RANDOM GEOMETRIC GRAPH DIAMETER 13 and. dG (y. (τ + 3/2)r) belonging to some petal. so for n large enough. so the shortest path in G between x2 and y2 satisﬁes dG (x2 . x1 ) + dG (x1 . y1 ) ≤ dπ + r(τ + 1) 2σ K . there is a vertex x2 ∈ B2 (O. dG (u. λ Concatenating all the above paths. By the preceding argument. y 2 ≤ (τ + 3/2)r. On the other hand. then. y1 ) + dG (y1 . we ﬁnd that dG (x.

Volume of the ℓp -unit ball. We refer the reader to [6] for details. By induction on d. AND CATHERINE YAN We conclude with two remarks. Evaluating this integral as in [11. y) ≤ x − y p + O (ln ln n/ ln n)1/d /λ. Fix p ∈ [1. For d = 1 this is trivial. so that the graph distance approximates the ℓp -metric. ELLIS. Appendix A.and spherical geometry In the body of the article. None of these facts are diﬃcult. Joel Spencer and Dennis Stanton for various helpful discussions and comments.4] yields the desired formula (14). Acknowledgements The authors thank Thomas Schlumprecht. the technique of Theorem 10 can be extended to obtain the stronger result dG (x. First. the proof of Theorem 10 can easily be adapted to the case d = 1 to obtain the upper bound diam(G1 (λ. x = ru1/p . §12. we obtain d vol(Bp (r)) = 2d r d Γ pΓ p+1 p p+d−1 p d−1 1 0 (1 − u) d−1 p u 1−p p du. We can still bound τ by o(ρ)/λ. however.14 ROBERT B. n)) ≤ p (2 + O(ln ln n/ ln n)) /λ. Facts about ℓp . but now any two vertices x and y are close to the same pin. ∞] and an integer d ≥ 1. d A. For d > 1 we have d vol(Bp (r)) = 2 r 0 d−1 vol Bp ((rp − xp )1/p ) dx. we used various facts about the Euclidean and ℓp geometry of balls and spherical caps. dx = (r/p)u(1−p)/p du. (Note that when d = 1 a pincushion consists of just one pin. Make the substitution u = xp /rp . d We will show that the (d-dimensional) volume of Bp (r) is (14) d vol(Bp (r)) = (2r)d Γ Γ p+1 p d p+d p . where Γ is the usual gamma function [11]. MARTIN. for convenience we present them together here along with brief proofs. ·Γ . on which a short path from x to y is found. Each pin is replaced by σ d−1 evenly spaced parallel pins. JEREMY L. We also thank the anonymous referees for several helpful suggestions. Let Bp (r) d be the ℓp -ball centered at the origin in R with radius r: d d Bp (r) = x ∈ Rd : i=1 |xi |p ≤ rp .1. It follows that (15) αd p d Γ vol(Bp (r)) := = d vol(B2 (r)) Γ p+1 p 3 d 2 d ·Γ 2+d 2 p+d p .) Second.

On the other hand. if p = 2.RANDOM GEOMETRIC GRAPH DIAMETER 15 d A.) Substituting for yi in (16) gives νxp−1 = xi . Spherical caps. but the precise formula is awkward for large d (one has to evaluate the integral sind θ dθ). In particular yi = (−λ/µ)xi for every i. Using the method d p of Lagrange multipliers (with objective function ( x − y p ) = i=1 (xi − yi )p ). the ℓp -diameter of B is (17) diamp (B) = max 2. centered at the origin in R . (That is. In the ﬁrst case (−x) − x p = 2. for instance. If d = 1. Every pair of antipodes x. . Of course. y must satisfy x 2 = y 2 = 1. A pair of points of B at distance diamp (B) are called ℓp -antipodes. . so either x is a coordinate unit vector or else xi = d−1/2 for every i. then B is a line segment and the only antipodes are its endpoints. we can easily obtain a lower bound for vol(C) by inscribing in it a “hypercone” H of √ height h whose base is a (d − 1)-sphere of radius s = r2 − (r − h)2 = 2rh − h2 . ℓp -antipodes on the Euclidean sphere. diamp (B) = max { x − y p : x. Let d ≥ 2. Then the pairs of ℓp -antipodes on B are precisely the pairs {±x} satisfying the following additional conditions: |x1 | = · · · = |xd | = d−1/2 when 1 ≤ p < 2. xd ) ∈ B : r − h ≤ x1 ≤ r}. In summary: d Proposition 16. Let B = B2 (1) be the Euclidean d unit ball. . the set of coordinates {x1 . C = {(x1 . . d−1 For r − h ≤ x ≤ r. . and let B = B2 (1) be the Euclidean unit ball. 2d1/p−1/2 = 2d1/p−1/2 2 for 1 ≤ p ≤ 2. We wish to calculate the ℓp -diameter of B. centered d at the origin in R .3. we may assume xi ≥ 0 ≥ yi for every i. In particular. . The volume of C can be determined exactly. Let C be a spherical cap of B of height h. In particular. and let p = 2. x p=1 when p = 2. d A. i where ν is some nonzero constant. while in the second case (−x) − x p = 2d−1/2+1/p . centered d at the origin in R . we ﬁnd that for every i. so y = −x. with 0 ≤ h ≤ r. xd } can contain at most one nonzero value. where λ and µ are nonzero constants. so applying (14) gives vol(H) = 2d−1 sd−1 Γ hd−1 Γ 3 d−1 2 d+1 2 r r−h (r − x)d−1 dx . then the antipodes are the pairs ±x with x 2 = 1. . the cross-section of H at x = x1 is B2 (s(r − x)/h). (16) p(xi − yi )p−1 = 2λxi = −2µyi .2. that is. for 2 ≤ p ≤ ∞. x is a coordinate unit vector when 2 < p ≤ ∞. . Suppose that d > 1. Without loss of generality. Let B = B2 (r) be the Euclidean ball of radius r. y ∈ B} . every pair of ℓp -antipodes on B is a pair of ℓ2 -antipodes.

Random Structures Algorithms (to appear) [6] Ellis. and let ±C be the cap of height h centered at ±a. Yan. z at minimum distance has displacement parallel to a (see Figure 4). 2nd edn.D. = π(2r − h)(d−1)/2 h(d+1)/2 . Z) := inf{ y − z p : y ∈ Y. dp (Y. M. N.B. X. Lattices and Groups. Random Structures Algorithms 9 (1996) 137–161 [8] Penrose. for p ≥ 2.H.H.. z ∈ Z}. J. Oxford University Press (2003) . Z ⊆ Rd is Let B = let ±a be a pair of ℓp -antipodes on B. By deﬁnition. References [1] Alon. Jia.B. Note that every pair y. MARTIN. Jia.: Package routing algorithms in mobile ad-hoc wireless networks. the ℓp -distance between two sets Y. In: 2001 International Conference on Parallel Processing Workshops.: Random Geometric Graphs. −C) = 2(r − h) a r p d B2 (r)..4.J. 3rd edn. d · Γ d+1 2 Since 2r − h ≥ r. C. R.: Approximating ℓp -metrics with path distance in a random geometric graph.: Random Graphs. ELLIS. AND CATHERINE YAN y B 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 a 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 z 0000000000 1111111111 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 r 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 −a 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 1111111111 0000000000 C z 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 −a0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 1111 0000 h B r C −C y 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 a 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 −C h p 2 p 2 Figure 4. J. 2nd edn. (2001) 485–490 [5] Ellis.D. Springer-Verlag (1998) [4] Chen. John Wiley (2000) [2] Bollob´s. Antipodally aligned caps ±C centered at antipodes ±a of B ⊆ R2 for diﬀerent values of p below and above 2. X. N. = 2(r − h)d1/p−1/2 2(r − h) for 1 ≤ p ≤ 2.: On k-connectivity for a geometric random graph. Sloane..H. d · Γ d+1 2 πr(d−1)/2 · h(d+1)/2 . (preprint) [7] Godehardt.16 ROBERT B. M.: Sphere Packing. (This can be veriﬁed by another easy Lagrange-multiplier calculation. Jaworski.. we have the bound (18) vol(C) ≥ A. E.: On random points in the unit disk. J: On the connectivity of a random interval graph. JEREMY L. R. X. The ℓp -distance between opposite spherical caps.) Therefore (19) dp (C. Spencer. Cambridge University Press (2001) a [3] Conway.: The probabilistic method. B..A. Random Structures Algorithms 15 (1999) 145–164 [9] Penrose..

KS 66045-7523 E-mail address: jmartin@math. College Station.iit. Lawrence.tamu. Zunic. USA E-mail address: cyan@math. J. Syst. IEEE Trans. I.: A course of modern analysis. IL 60616. 13 (2002) 14–25 [11] Whittaker. Texas A&M University..N. H. Li.: Dominating sets and neighbor elimination-based broadcasting algorithms in wireless networks. G. E.T..ku. Parallel Distrib. University of Kansas.edu Department of Mathematics. Watson. Chicago. Seddigh. Cambridge University Press (1996) [12] Wu.. TX 778433368. J. USA E-mail address: rellis@math.edu ..edu Department of Mathematics. M.: A dominating-set-based routing scheme in ad hoc wireless networks. Illinois Institute of Technology.RANDOM GEOMETRIC GRAPH DIAMETER 17 [10] Stojmenovic. Telecommunication Systems 18 (2001) 13–36 Department of Applied Mathematics.

- Affine Object Tracking Using Kernel-Based Region Covariance DescriptorsUploaded byAlberto Kobon Ground
- Post cat1.pdfUploaded byShrey Grover
- Model- vs. design-based sampling and variance estimationUploaded byFanny Sylvia C.
- Lars_Methods of Analysing VariabilityUploaded bymarco.vegaromero
- Small WorldUploaded byjasonbrogan
- 4703-07-Notes-0Ground Penetrating Radar Theory and ApplicationsUploaded byFuad M. Ashour
- EfficientIndexes.pdfUploaded byG
- Proofs 2Uploaded byJeoff Libo-on
- Autocorrelation_MRUploaded bysandrageorgiaion
- hw3solUploaded byMohamad Abdallah
- Linear Least Squares ProblemsUploaded byAshoka Vanjare
- American Computer Science League (ACSL) Rules and HandbookUploaded bybellataco
- Chap 5Uploaded byAmsalu Walelign
- GraphEntropyEnsenadaneuUploaded byaaguileraslp
- Mincostflow NotesUploaded byttungl
- edgar-kalmanfilterUploaded bygatzke
- 10.1.1.72Uploaded bylanhcht01
- Risk and Return for PortfolioUploaded byferuzbek
- Comsol EmtUploaded byjcow
- Topic03 Correlation RegressionUploaded byVijay Phani Kumar
- Kawai_A Flexible Shear Spinning of Axi-symmetrical Shells_JMPT_2007Uploaded byiulika28
- 173232298 a Guide to Modern Econometrics by Verbeek 151 160Uploaded byAnonymous T2LhplU
- 6in Double Petal ShellUploaded byOscar Pamos
- Albert, R-Scale-Free Networks in Cell Biology-2008Uploaded bySantiago Sierra
- An All-Sky Analysis of Polarization in the Microwave BackgroundUploaded byraveneyesdead
- 0580_w17_qp_43Uploaded byabbasmohammed60
- ansys_cmdUploaded byAaisha Nazreen
- ParticleUploaded byAdeliza Mortalla
- 10 Mathematics Impq Sa 2 7 MensurationUploaded byVinoth Rocker
- aPDS-II-UNIT -V-RN.pptxUploaded byRanganathan Nagendran