1: \documentclass[11pt]{article}

2: 

3: \usepackage{epsfig}

4: 

5: \usepackage{latexsym}

6: \usepackage{amsfonts}

7: \usepackage{amsthm}

8: \usepackage{amsmath}

9: \usepackage{fullpage}

10: 

11: \newcommand{\vecc}[1]{{\overrightarrow{#1}}}

12: \newcommand{\reals}{\mathbb{R}}

13: \let\bd\partial

14: \def\A{\mathcal{A}}

15: \def\T{\mathcal{T}}

16: \def\L{\mathcal{L}}

17: \def\EE{\mathbf{E}}

18: \let\eps\varepsilon

19: \def\B{{\cal B}}

20: \def\C{{\cal C}}

21: \def\D{{\cal D}}

22: \def\R{{\cal R}}

23: 

24: 

25: \newcommand{\old}[1]{{}}

26: \newcommand{\floor}[1]{{\lfloor #1\rfloor}}

27: \newcommand{\ceil}[1]{{\lceil #1\rceil}}

28: 

29: \newcommand{\estie}[1]{\textbf{Estie:} #1}

30: \newcommand{\sandor}[1]{\textbf{Sandor:} #1}

31: \newcommand{\ferran}[1]{\textbf{Ferran:} #1}

32: \newcommand{\marc}[1]{\textbf{Marc:} #1}

33: \newcommand{\joe}[1]{\textbf{Joe:} #1}

34: \newcommand{\saurabh}[1]{\textbf{Saurabh:} #1}

35: \newcommand{\vera}[1]{\textbf{Vera:} #1}

36: 

37: \newtheorem{theorem}{Theorem}[section]

38: \newtheorem{lemma}[theorem]{Lemma}

39: \newtheorem{conjecture}[theorem]{Conjecture}

40: \newtheorem{corollary}[theorem]{Corollary}

41: \newtheorem{claim}[theorem]{Claim}

42: \newtheorem{observation}[theorem]{Observation}

43: \newtheorem{proposition}[theorem]{Proposition}

44: \newtheorem{fact}[theorem]{Fact}

45: \newtheorem*{claima}{Claim A}

46: \newtheorem*{claimb}{Claim B}

47: \newtheorem*{claimc}{Claim C}

48: \newtheorem*{propA}{Property A}

49: \newtheorem*{propB}{Property B}

50:  

51: \theoremstyle{definition}

52: \newtheorem{definition}[theorem]{Definition}

53: \newtheorem{remark}[theorem]{Remark}

54: 

55: \begin{document}

56: 

57: \title{On the Reflexivity of Point Sets}

58: \author{

59:   Esther M. Arkin%

60:     \thanks{Department of Applied Mathematics and Statistics,

61:             Stony Brook University, Stony Brook, NY 11794-3600,

62:             email: \{\texttt{estie}, \texttt{jsbm}\}\texttt{@ams.sunysb.edu}.}

63: \and

64:   S\'andor P. Fekete%

65:     \thanks{

66: Department of Mathematical Optimization, 

67: TU Braunschweig, Pockelsstr.~14,

68: D-38106 Braunschweig, Germany; \texttt{sandor.fekete@tu-bs.de}.}

69: \and

70:   Ferran Hurtado

71:     \thanks{Departament de Matem\`atica Aplicada II,

72:             Universitat Polit\`ecnica de Catalunya,

73:             Pau Gargallo, 5, E-08028, Barcelona, Spain, 

74: email: \{\texttt{hurtado}, \texttt{noy}, \texttt{vera}\}\texttt{@ma2.upc.es}.}

75: \and

76:   Joseph S. B. Mitchell\footnotemark[1]

77: \and

78:   Marc Noy\footnotemark[3]

79: \and

80:   Vera Sacrist\'an\footnotemark[3]

81: \and

82:   Saurabh Sethia

83:     \thanks{

84: Department of Computer Science,

85: Oregon State University, Corvallis, OR 97331, USA. \texttt{saurabh@cs.orst.edu}.

86: This work was conducted while S.~Sethia was at the Dept. of Computer Science, 

87: Stony Brook University.}

88: }

89: 

90: \date{}

91: 

92: \maketitle

93: 

94: \begin{abstract}

95:   We introduce a new measure for planar point sets $S$ that captures a

96:   combinatorial distance that $S$ is from being a convex set: The {\em

97:     reflexivity} $\rho(S)$ of $S$ is given by the smallest number of

98:   reflex vertices in a simple polygonalization of $S$.  We prove

99:   various combinatorial bounds and provide efficient algorithms to

100:   compute reflexivity, both exactly (in special cases) and

101:   approximately (in general).  Our study considers also some closely

102:   related quantities, such as the {\em convex cover} number

103:   $\kappa_c(S)$ of a planar point set, which is the smallest number of

104:   convex chains that cover $S$, and the {\em convex partition} number

105:   $\kappa_p(S)$, which is given by the smallest number of convex

106:   chains with pairwise-disjoint convex hulls that cover $S$.  We have

107:   proved that it is NP-complete to determine the convex cover or the

108:   convex partition number and have given logarithmic-approximation

109:   algorithms for determining each.

110: \end{abstract}

111: 

112: 

113: \section{Introduction}

114: 

115: In this paper, we study a fundamental combinatorial property of a

116: discrete set, $S$, of points in the plane: What is the minimum number,

117: $\rho(S)$, of {\em reflex vertices} among all of the {\em simple

118:   polygonalizations} of $S$?  A {\em polygonalization} of $S$ is a

119: closed tour on $S$ whose straight-line embedding in the plane defines

120: a connected cycle without crossings, i.e., a simple polygon.  A vertex

121: of a simple polygon is {\em reflex} if it has interior angle greater

122: than $\pi$.  We refer to $\rho(S)$ as the {\em reflexivity} of $S$.

123: We let $\rho(n)$ denote the maximum possible value of $\rho(S)$

124: for a set $S$ of $n$ points.

125: 

126: In general, there are many different polygonalizations of a point set

127: $S$.  There is always at least one: simply connect the points in

128: angular order about some point interior to the convex hull of $S$

129: (e.g., the center of mass suffices). A set $S$ has precisely one

130: polygonalization if and only if it is in convex position; in general,

131: though, a point set has numerous polygonalizations. Studying the set

132: of polygonalizations (e.g., counting them, enumerating them, or

133: generating a random element) is a challenging and active area of

134: investigation in computational geometry~\cite{aak-eotsp-01,ak-psotd-01,ah-hgrp-96,e-webpage,gnt-lbncf-95,zssm-grpgv-96}.

135: 

136: 

137: The reflexivity $\rho(S)$ quantifies, in a combinatorial sense, the

138: degree to which the set of points $S$ is in convex position.  See

139: Figure~\ref{fig:example} for an example.  We remark that there are

140: other notions of combinatorial ``distance'' from

141: convexity of a point set $S$, e.g., the minimum number of points to

142: delete from $S$ in order that the remaining point set is in convex

143: position, the number of convex layers, or the minimum number of

144: changes in the orientation of triples of points of $S$ in order to

145: transform $S$ into convex position.

146:   

147: \begin{figure}

148: \centerline{\psfig{file=example.eps,height=2.0in}}

149: \caption{Two polygonalizations of a point set, one (left) using 7 reflex vertices and one (right) using only 3 reflex vertices.}

150: \label{fig:example}

151: \end{figure}

152: 

153: We have conducted a formal study of reflexivity, both in terms of its

154: combinatorial properties and in terms of an algorithmic analysis of

155: the complexity of computing it, exactly or approximately.  Some of our

156: attention is focussed on the closely related {\em convex cover number}

157: of $S$, which gives the minimum number of convex chains (subsets of

158: $S$ in convex position) that are required to cover all points of~$S$.

159: For this question, we distinguish between two cases: The {\em convex cover

160: number}, $\kappa_c(S)$, is the smallest number of convex chains required 

161: to cover $S$; the {\em convex partition number}, 

162: $\kappa_p(S)$, is the smallest number of convex chains 

163: with pairwise-disjoint convex hulls required to cover $S$. 

164: Note that nested chains are feasible for a convex cover

165: but not for a convex partition.

166: 

167: \paragraph{Motivation.}

168: In addition to the fundamental nature of the questions and problems we

169: address, we are also motivated to study reflexivity for several

170: other reasons:

171: 

172: (1) An application motivating our original investigation is that of

173: meshes of low stabbing number and their use in performing ray shooting

174: efficiently.  If a point set $S$ has low reflexivity or a low convex

175: partition number, then it has a triangulation of low stabbing number,

176: which may be much lower than the general $O(\sqrt{n})$ upper bound

177: guaranteed to exist (\cite{a-rsoas-92,hs-parss-95,w-ggssn-93}).

178: For example, if the reflexivity is $O(1)$, then $S$ has a triangulation

179: with stabbing number $O(\log n)$.

180: 

181: (2) Classifying point sets by their reflexivity may give us some

182: structure for dealing with the famously difficult question of counting and

183: exploring the set of all polygonalizations of $S$. See

184: \cite{gnt-lbncf-95,zssm-grpgv-96} for some references to this problem.

185: 

186: (3) There are several applications in computational geometry in which

187: the number of reflex vertices of a polygon can play an important role

188: in the complexity of algorithms.

189: If one or more polygons are {\em given} to us, there are many problems

190: for which more efficient algorithms can be written with complexity in

191: terms of ``$r$'' (the number of reflex vertices), instead of ``$n$''

192: (the total number of vertices), taking advantage of the possibility

193: that we may have $r\ll n$ for some practical instances

194: (see, e.g., \cite{hm-ftprs-85,hn-tvgrv-96}).

195: The number of reflex vertices also plays an important role in convex

196: decomposition problems for polygons; see Keil~\cite{k-pd-00} for a recent

197: survey, and see Agarwal, Flato, and Halperin~\cite{afg-pdecm-02} for

198: applications of convex decompositions to computing Minkowski sums of

199: polygons.

200: 

201: (4) Reflexivity is intimately related to the issue

202: of convex cover numbers, which

203: has roots in the classical work of Erd\H{o}s and

204: Szekeres~\cite{es-cpg-35,es-sepeg-60}, and has been studied

205: more recently by Urabe~et~al.~\cite{HRU,hu-ndcqp-01,u-opcp-96,u-ppscp-97}.

206: 

207: (5) Our problems are related to some problems in curve (surface)

208: reconstruction, where the goal is to obtain a ``good''

209: polygonalization of a set of sample points 

210: (see, e.g.,~\cite{abe-cbscc-98,dk-spacr-99,dmr-crcdgr-99}).  

211:   

212: \paragraph{Related Work.}

213: The study of convex chains in finite planar point sets is the topic of

214: classical papers by Erd\H{o}s and

215: Szekeres~\cite{es-cpg-35,es-sepeg-60}, who showed that any point set

216: of size $n$ has a convex subset of size $t=\Omega(\log n)$.  This is

217: closely related to the convex cover number $\kappa_c$, since it

218: implies an asymptotically tight bound on $\kappa_c(n)$, the worst-case

219: value for sets of size $n$. There are still a number of open problems

220: related to the exact relationship between $t$ and $n$; see, for

221: example, \cite{dcg19} for recent developments.  

222: 

223: Other issues have been considered, such as the existence and

224: computation (\cite{deo-secp-90}) of large ``empty'' convex subsets

225: (i.e., with no points of $S$ interior to their hull); this is related

226: to the convex partition number, $\kappa_p(S)$. It was shown by

227: Horton~\cite{h-snec7g-83} that there are sets with no empty convex

228: chain larger than 6; this implies that $\kappa_p(n)\geq n/6$.  

229: 

230: Tighter worst-case bounds on $\kappa_p(n)$ were given by

231: Urabe~\cite{u-opcp-96,u-ppscp-97}, who shows that $\lceil (n-1)/4

232: \rceil \leq \kappa_p(n) \leq \lceil 2n/7\rceil$ and that

233: $\kappa_c(n)=\Theta(n/\log n)$ (with the upper and lower bounds having

234: a gap of roughly a factor of 2).  (Urabe~\cite{u-ppsdc-99} also studies the

235: convex partitioning problem in $\Re^3$, where, in particular, the

236: upper bound on $\kappa_p(n)$ is shown to be $\lceil 2n/9\rceil$.)

237: Most recently, Hosono and Urabe~\cite{hu-ndcqp-01} have obtained

238: improved bounds on the size of a partition of a set of points into

239: disjoint convex quadrilaterals, which has the consequence of improving

240: the upper bound on $\kappa_p(n)$: $\kappa_p(n)\leq \lceil 5n/18\rceil$

241: and $\kappa_p(n)\leq (3n+1)/11$ for $n=11\cdot 2^{k-1}-4$ ($k\geq 1$).

242: The remaining gaps in the constants between upper and lower bounds for

243: $\kappa_c(n)$ and $\kappa_p(n)$ (as well as the gap that our bounds

244: exhibit for reflexivity in terms of $n$) all point to the apparently

245: common difficulty of these combinatorial problems on convexity.

246: 

247: For a given set of points, we are interested in polygonalizations of the

248: points that are ``as convex as possible''. This has been studied

249: in the context of TSP (traveling salesperson problem) tours of a

250: point set $S$, where convexity of $S$ implies (trivially) the

251: optimality of a convex tour.  Convexity of a tour can be characterized

252: by two conditions.  If we drop the global condition (i.e., no crossing

253: edges), but keep the local condition (i.e., no reflex vertices), we

254: get ``pseudo-convex'' tours.  In \cite{fw-artp-97} it was shown that

255: any set with $|S|\geq 5$ has such a pseudo-convex tour. It is natural

256: to require the global condition of simplicity instead, and minimize

257: the number of local violations -- i.e., the number of reflex vertices.

258: This kind of problem is similar to that of minimizing the total amount

259: of turning in a tour, as studied by

260: Aggarwal~et~al.~\cite{ackms-amtsp-97}.

261: 

262: The number of polygonalizations on $n$ points is, in general,

263: exponential in~$n$; Garc{\'\i}a et al.~\cite{gnt-lbncf-95} prove a

264: lower bound of $\Omega(4.64^n)$.

265: 

266: Another related problem is studied by Hosono~et~al.~\cite{HRU}:

267: Compute a polygonalization $P$ of a point set $S$ such that the

268: interior of $P$ can be decomposed into a minimum number ($f(S)$) of

269: empty convex polygons. They prove that $\lfloor (n-1)/4\rfloor \leq

270: f(n) \leq \lfloor (3n-2)/5\rfloor,$ where $f(n)$ is the maximum

271: possible value of $f(S)$ for sets $S$ of $n$ points.  The authors

272: conjecture that $f(n)$ grows like $n/2$.  For reflexivity $\rho(n)$,

273: we show that $\floor{n/4}\leq \rho(n)\leq \ceil{n/2}$ and conjecture

274: that $\rho(n)$ grows like $n/4$, which, if true, would imply that

275: $f(n)$ grows like $n/2$.

276: 

277: We mention one final related problem.  A {\em convex decomposition} of

278: a point set $S$ is a convex planar polygonal subdivision of the convex

279: hull of $S$ whose vertices are $S$.  Let $g(S)$ denote the minimum

280: number of faces in a convex decomposition of $S$, and let $g(n)$

281: denote the maximum value of $g(S)$ over all $n$-point sets $S$. It has

282: been conjectured (\cite{RU}) that $g(n)=n+c$ for some constant $c$,

283: and it is known that $g(n)\le 3n/2$ (\cite{RU}) and that $n+2\le

284: g(n)$~(\cite{AK}).

285: 

286: \paragraph{Summary of Main Results.}

287: We have both combinatorial and algorithmic results on

288: reflexivity.  Our combinatorial results include

289: %

290: \begin{itemize}

291: \item Tight bounds on the worst-case value of $\rho(S)$ in terms of

292:   $n_I$, the number of points of $S$ interior to the convex hull of

293:   $S$; in particular, we show that $\rho(S)\leq \ceil{n_I/2}$ and that

294:   this upper bound can be achieved by a class of examples.

295: \item Upper and lower bounds on $\rho(S)$ in terms of $n=|S|$; in

296:   particular, we show that $\floor{n/4}\leq\rho(n)\leq \ceil{n/2}$.

297:   In the case in which $S$ has two layers, we show that $\rho(S)\leq

298:   \ceil{n/4}$, and this bound is tight.

299: \item Upper and lower bounds on ``Steiner reflexivity'', which is

300:   defined with respect to the class of polygonalizations that allow

301:   Steiner vertices (not from the input set $S$).

302: 

303: \end{itemize}

304: %

305: Our algorithmic results include

306: %

307: \begin{itemize}

308: \item We prove that it is NP-complete to compute the convex cover

309:   number ($\kappa_c(S)$) or the convex partition number

310:   ($\kappa_p(S)$), for a given point set~$S$.

311: \item

312:   We give polynomial-time approximation algorithms,

313: having approximation factor $O(\log n)$, for the problems

314: of computing convex cover number, convex partition number, or

315: Steiner reflexivity of $S$.

316: \item

317: We give efficient exact algorithms to test if $\rho(S)=1$ or $\rho(S)=2$.

318: \end{itemize}

319: %

320: 

321: In Section~\ref{sec:inflection} we study a closely related problem --

322: that of determining the ``inflectionality'' of $S$, defined to be the

323: minimum number of inflection edges (joining a convex to a reflex

324: vertex) in any polygonalization of $S$.  We give an $O(n\log n)$ time

325: algorithm to determine an inflectionality-minimizing polygonalization,

326: which we show will never need more than 2 inflection edges.

327: 

328: 

329: \section{Preliminaries}

330: 

331: Throughout this paper, $S$ will be a set of $n$ points in the

332: plane $\Re^2$.  A polygonalization, $P$, of $S$ is a simple polygon

333: whose vertex set is $S$.  

334: Let ${\cal P}$ be the set of all polygonalizations of $S$.

335: Note that ${\cal P}$ is not empty, since any point set $S$ having

336: $n\geq 3$ points has at least one polygonalization (e.g., the

337: star-shaped polygonalization obtained by sorting points of $S$

338: angularly about a point interior to the convex hull of~$S$).

339: 

340: Each vertex of a simple polygon $P$ is either {\em reflex} or {\em

341:   convex}, according to whether the interior angle at the vertex is

342: greater than $\pi$ or less than or equal to $\pi$, respectively.  We

343: let $r(P)$ (resp., $c(P)$) denote the number of reflex (resp., convex)

344: vertices of $P$.  We define the {\em reflexivity} of a planar point

345: set $S$ to be $\rho (S) = \min_{P \in {\cal P}} r(P)$.  Similarly, the

346: {\em convexivity} of a planar point set $S$ is defined to be

347: $\chi(S)=\max_{P\in{\cal P}} c(P)$.  Note that $\chi(S)=n-\rho(S)$.

348: We let $\rho(n)=\max_{|S|=n} \rho(S)$.

349: 

350: We let $\textrm{CH}(S)$ denote the {\em convex hull} of $S$.  The point set $S$

351: is partitioned into (convex) {\em layers}, $S_1,S_2,\ldots$, where the first

352: layer is given by the set $S_1$ of points of $S$ on the boundary of

353: $\textrm{CH}(S)$, and the $i$th layer, $S_i$ ($i\geq 2$) is given by the set of

354: points of $S$ on the boundary of $\textrm{CH}(S\setminus (S_1\cup\cdots\cup

355: S_{i-1}))$.  We say that $S$ has {\em $k$ layers} or {\em onion depth

356:   $k$} if $S_k\neq\emptyset$, while $S_{k+1}=\emptyset$.  We say that

357: $S$ is in {\em convex position} (or forms a {\em convex chain}) if it

358: has one layer (i.e., $S=S_1$).

359: 

360: A {\em Steiner point} is a point not in the set $S$ that may be added

361: to $S$ in order to improve some structure of $S$.  We define the {\em

362:   Steiner reflexivity} $\rho'(S)$ to be the minimum number of reflex

363: vertices of any simple polygon with vertex set $V\supset S$.  

364: We let

365: $\rho'(n)=\max_{|S|=n} \rho'(S)$.

366: %

367: The {\em

368:   Steiner convexivity}, $\chi'(S)$, is defined similarly.

369: %

370: A {\em convex cover} of $S$ is a set of subsets of $S$ whose union

371: covers $S$, such that each subset is a convex chain (a set in convex

372: position).  A {\em convex partition} of $S$ is a partition of $S$ into

373: subsets each of which is in convex position, such that the convex

374: hulls of the subsets are pairwise disjoint.  We define the {\em convex

375:   cover number}, $\kappa_c(S)$,

376: to be the minimum number of subsets in a convex cover of $S$.

377: We similarly define the

378: {\em convex partition number}, $\kappa_p(S)$.

379: We denote by $\kappa_c(n)$ and $\kappa_p(n)$ the

380: worst-case values for sets of size~$n$.

381: 

382: Finally, we state a basic property of polygonalizations of point sets.

383: 

384: \begin{lemma} \label{lem:CH-order}

385: In any polygonalization of $S$,

386: the points of $S$ that are vertices of the convex hull of $S$ 

387: are convex vertices of the polygonalization,

388: and they occur in the polygonalization in the same order

389: in which they occur along the convex hull.

390: \end{lemma}

391: 

392: \begin{proof}

393:   Any polygonalization $P$ of $S$ must lie within

394:   the convex hull of $S$, since edges of the polygonalization are

395:   convex combinations of points of $S$.  Thus, if $p\in S$ is a vertex

396:   of $\textrm{CH}(S)$, then the local neighborhood of $P$ at $p$ lies within a

397:   convex cone, so $p$ must be a convex vertex of~$P$.

398: 

399: Consider a clockwise traversal of $P$ and let $p$ and $q$ be

400: two vertices of $\textrm{CH}(S)$ occurring consecutively along $P$.  Then

401: $p$ and $q$ must also appear consecutively along a clockwise traversal of

402: the boundary of $\textrm{CH}(S)$, since the subchain of $P$ linking $p$ to $q$

403: partitions $\textrm{CH}(S)$ into a region to its left (which is outside the

404: polygon $P$) and a region to its right (which must contain all points of

405: $S$ not in the subchain).

406: \end{proof}

407: 

408: \section{Combinatorial Bounds}

409: 

410: In this section we establish several combinatorial results on

411: reflexivity and convex cover numbers.

412: 

413: \subsection{Reflexivity}

414: 

415: One of our main combinatorial results establishes an upper bound on

416: the reflexivity of $S$ that is worst-case tight in terms of the number

417: $n_I$ of points {\em interior} to the convex hull, $\textrm{CH}(S)$,

418: of $S$.  Since, by Lemma~\ref{lem:CH-order}, the points of $S$ that

419: are vertices of $\textrm{CH}(S)$ are required to be convex vertices in

420: any (non-Steiner) polygonalization of $S$, the bound in terms of $n_I$

421: seems to be quite natural.

422: 

423: \begin{theorem}

424: \label{thm:reflex-upper-bnd}

425: Let $S$ be a set of $n$ points in the plane, $n_I$ of which are

426: interior to the convex hull $\textrm{CH}(S)$.  Then $\rho(S)\leq \ceil{n_I/2}$.

427: % and this bound is tight. (next thm)

428: \end{theorem}

429: 

430: \begin{figure}

431: \centerline{\psfig{file=reflex-upper-bnd-new.eps,height=1.5in}}

432: \caption{Computing a polygonalization with at most $\ceil{n_I/2}$ reflex vertices.}

433: \label{fig:reflex-upper-bnd}

434: \end{figure}

435: 

436: \begin{proof} We describe a polygonalization in which at most half of

437:   the interior points are reflex. We begin with the polygonalization

438:   of the convex hull vertices that is given by the convex polygon

439:   bounding the hull.  We then iteratively incorporate interior

440:   points of $S$ into the polygonalization.  Fix a point $p_0$ that

441:   lies on the convex hull of $S$.  At a generic step of the algorithm,

442:   the following invariants hold: (1) our polygonalization consists of

443:   a simple polygon, $P$, whose vertices form a subset of $S$; and (2)

444:   all points $S'\subset S$ that are not vertices of $P$ lie

445:   interior to $P$; in fact, the points $S'$ all lie within the

446:   subpolygon, $Q$, to the left of the diagonal $p_0p_i$, where $p_i$

447:   is a vertex of $P$ such that the subchain of $\partial P$ from $p_i$

448:   to $p_0$ (counter-clockwise) together with the diagonal $p_0p_i$

449:   forms a convex polygon ($Q$).  If $S'$ is empty, then $P$ is a

450:   polygonalization of $S$ and we are done; thus, assume that $S'\neq

451:   \emptyset$.  Define $p_{i+1}$ to be the first point of $S'$ that is

452:   encountered when sweeping the ray $\vecc{p_0p_i}$ counter-clockwise

453:   about its endpoint $p_0$.  Then we sweep the subray with endpoint

454:   $p_{i+1}$ further counter-clockwise, about $p_{i+1}$, until we

455:   encounter another point, $q$, of $S'$.  (If $|S'|=1$, we can readily

456:   incorporate $p_{i+1}$ into the polygonalization, increasing the

457:   number of reflex vertices by one.)  Now the ray $\vecc{p_{i+1}q}$

458:   intersects the boundary of $P$ at some point $c\in ab$ on the

459:   boundary of~$Q$.

460:   

461:   As a next step, we modify $P$ to include interior points $p_{i+1}$ and $q$ (and

462:   possibly others as well) by replacing the edge $ab$ with the chain

463:   $(a,p_{i+1}$, $q$, $q_1,\ldots$, $q_k,b)$, where the points $q_i$ are interior

464:   points that occur along the chain we obtain by ``pulling taut'' the

465:   chain $(q,c,b)$.  In this ``gift wrapping''

466:   fashion, we continue to rotate rays counter-clockwise about each interior point

467:   $q_i$ that is hit until we encounter $b$.  This results in 

468:   incorporating at least two new interior points (of $S'$) into the

469:   polygonalization $P$, while creating only one new reflex vertex

470:   (at~$p_{i+1}$).  It is easy to check that the invariants (1) and (2)

471:   hold after this step.

472: \end{proof}

473: 

474: In fact, the upper bound of Theorem~\ref{thm:reflex-upper-bnd},

475: $\rho(S)\leq \ceil{n_I/2}$, is {\em tight} in the worst case, as we

476: now argue based on the special configuration of points, $S=S_0(n)$, in

477: Figure~\ref{fig:lower-bnd}.  The set $S_0(n)$ is defined for any

478: integer $n\geq 6$, as follows: $\ceil{n/2}$ points are placed in

479: convex position (e.g., forming a regular $\ceil{n/2}$-gon), forming

480: the convex hull $\textrm{CH}(S)$, and the remaining $n_I=\floor{n/2}$

481: interior points are also placed in convex position, each one placed

482: ``just inside'' $\textrm{CH}(S)$, near the midpoint of an edge of

483: $\textrm{CH}(S)$.  The resulting configuration $S_0(n)$ has two layers

484: in its convex hull.

485: 

486: 

487: \begin{figure}

488: \centerline{\psfig{file=lower-bnd.eps,width=2.0in}}

489: \caption{Left: The configuration of points, $S_0(n)$, which has reflexivity

490: $\rho(S_0(n))\geq \ceil{n_I/2}$. Right: A polygonalization having $\ceil{n_I/2}$ reflex vertices.}

491: \label{fig:lower-bnd}

492: \end{figure}

493: 

494: \begin{lemma}

495: \label{lem:lower-bnd}

496: For any $n\geq 6$, 

497: $\rho(S_0(n))\geq \ceil{n_I/2} \geq\floor{n/4}$.

498: \end{lemma}

499: 

500: \begin{proof}

501: Let $(x_1,x_2,\ldots,x_{\ceil{n/2}})$ denote the points 

502: of $S_0(n)$ on the convex hull,  in clockwise order, 

503: and let $(v_1,v_2,\ldots,v_{\floor{n/2}})$ denote the

504: remaining points of $S_0(n)$, with 

505: $v_i$ just inside the convex hull edge $(x_i,x_{i+1})$.

506: We define $x_{\ceil{n/2}+1}=x_1$.

507: 

508: Consider any polygonalization, $P$, of $S_0(n)$.  {From}

509: Lemma~\ref{lem:CH-order} we know that the points $x_i$ are convex

510: vertices of $P$, occurring in the order

511: $x_1$, $x_2,\ldots$, $x_{\ceil{n/2}}$ around the boundary of $P$.  Consider

512: the subchain, $\gamma_i$, of $\partial P$ that goes from $x_i$ to

513: $x_{i+1}$, clockwise around $\partial P$.  Let $m_i$ denote the number

514: of points $v_j$, interior to the convex hull of $S_0(n)$, that appear

515: along $\gamma_i$.  

516: 

517: If $m_i=0$, $\gamma_i=x_ix_{i+1}$.  If $m_i=1$, then

518: $\gamma_i=x_iv_ix_{i+1}$ and $v_i$ is a reflex vertex of $P$; to see

519: this, note that 

520: $v_i$ lies interior to the triangle determined by $x_i$, $x_{i+1}$,

521: and any $v_j$ with $j\neq i$.  If $m_i>1$, then we claim that (a)

522: $v_i$ must be a vertex of the chain $\gamma_i$, (b) $v_i$ is a convex

523: vertex of $P$, and (c) any other point $v_j$, $j\neq i$, that is a

524: vertex of $\gamma_i$ must be a reflex vertex of $P$.  This claim

525: follows from the fact that the points $x_i$, $x_{i+1}$, and any

526: nonempty subset of $\{v_j: j\neq i\}$ are in convex position, with the

527: point $v_i$ interior to the convex hull.  Refer to

528: Figure~\ref{fig:lower-bnd-proof}, where the subchain $\gamma_i$ is

529: shown dashed.

530: 

531: \begin{figure}

532: \centerline{\psfig{file=lower-bound-proof2.eps,width=1.5in}}

533: \caption{Proof of the lower bound:

534: $\rho(S_0(n))\geq \ceil{n_I/2} \geq\floor{n/4}$.}

535: \label{fig:lower-bnd-proof}

536: \end{figure}

537: 

538: Thus, the number of reflex vertices of $P$ occurring along $\gamma_i$

539: is in any case at least $\ceil{m_i/2}$, and we have

540: %

541: \begin{eqnarray*}

542: \rho(S_0(n)) &\geq& \sum\ceil{m_i/2}\\

543: &\geq& \left\lceil\sum (m_i/2)\right\rceil=\ceil{n_I/2}\geq\floor{n/4}.

544: \end{eqnarray*}

545: 

546: \end{proof}

547: 

548: Since $n_I\leq n$, the corollary below is immediate from

549: Theorem~\ref{thm:reflex-upper-bnd} and Lemma~\ref{lem:lower-bnd}.  The

550: gap in the bounds for $\rho(n)$, between $\floor{n/4}$ and

551: $\ceil{n/2}$, remains an intriguing open problem.  While our

552: combinatorial bounds are worst-case tight in terms of $n_I$ (the number of points

553: of $S$ whose convexity/reflexivity is not forced by the convex hull of

554: $S$), they are not worst-case tight in terms of~$n$.

555: 

556: \begin{corollary}

557: \label{cor:reflex-upper-bnd}

558: $\floor{n/4}\leq\rho(n)\leq \ceil{n/2}$.

559: \end{corollary}

560: 

561: Based on experience with a software tool developed by A.~Dumitrescu

562: that computes, in exponential time, the reflexivity of user-specified

563: or randomly generated point sets, as well as the proven behavior of

564: $\rho(n)$ for small values of $n$ (see Section~\ref{sec:small}),

565: we make the following conjecture:

566: 

567: \begin{conjecture}

568: \label{con:rho-upper}

569: $\rho(n)=\floor{n/4}$.

570: \end{conjecture}

571: 

572: \subsection{Steiner Points}

573: 

574: If we allow Steiner points in the polygonalizations of $S$, the

575: reflexivity of $S$ may go down substantially, as the example in

576: Figure~\ref{fig:Steiner-ex} shows.  In fact, the illustrated class of

577: examples shows that the use of Steiner points may allow the

578: reflexivity to go down by a factor of two. 

579: The Steiner reflexivity, $\rho'(S)$, of $S$ is the minimum number of reflex

580: vertices of any simple polygon with vertex set $V\supset S$. We conjecture that

581: $\rho'(S)\geq\rho(S)/2$ for any set $S$, which would imply that this

582: class of examples (essentially) maximizes the ratio

583: $\rho(S)/\rho'(S)$.

584: 

585: \begin{conjecture}

586: \label{con:rho.rho'}

587: For any set $S$ of points in the plane, $\rho'(S)\geq\rho(S)/2$. 

588: \end{conjecture}

589: 

590: \begin{figure}

591: \centerline{\psfig{file=Steiner-ex.eps,height=1.9in}}

592: \caption{Left: A point set $S$ having reflexivity $\rho(S)=r$.

593: Right: The reflexivity of $S$ when Steiner points {\em are} permitted

594: is substantially reduced from the no-Steiner case: $\rho'(S)=r/2$.}

595: \label{fig:Steiner-ex}

596: \end{figure}

597: 

598: We have seen (Corollary~\ref{cor:reflex-upper-bnd}) 

599: that $\floor{n/4}\leq\rho(n)\leq \ceil{n/2}$.  We now show that allowing

600: Steiner points in the polygonalization allows us to prove

601: a smaller upper bound, while still being able

602: to prove roughly the same lower bound:

603: 

604: \begin{theorem}

605: \label{thm:Steiner-upper}

606: $$\left\lceil\frac{n-1}{4}\right\rceil-1 \leq \rho'(n)\leq \left\lceil\frac{n}{3}\right\rceil.$$

607: \end{theorem}

608: 

609: \begin{proof}

610:   For the upper bound, we give a specific method of constructing a

611:   polygonalization (with Steiner points) of a set $S$ of $n$ points.

612:   Sort the points $S$ by their $x$-coordinates and group them into

613:   consecutive triples. Let $p_{n+1}$ denote a (Steiner) point with a

614:   very large positive $y$-coordinate and let $p_0$ denote a (Steiner)

615:   point with a very negative $y$-coordinate.  Each triple, together

616:   with either point $p_{n+1}$ or point $p_0$, forms a convex

617:   quadrilateral. Then, we can polygonalize $S$ using one reflex

618:   (Steiner) point per triple, as shown in

619:   Figure~\ref{fig:Steiner-upper}, placed very close to $p_{n+1}$ or

620:   $p_0$ accordingly.  This polygonalization has at most $\ceil{n/3}$

621:   reflex points.

622:   

623:   For the lower bound, we consider the configuration of $n$ points,

624:   $S$, used in Urabe~\cite{u-opcp-96} to prove that $\kappa_p(n)\geq

625:   \ceil{(n-1)/4}$.  For this set $S$ of $n$ points, let $P$ be a

626:   Steiner polygonalization having $r$ reflex vertices.  Then the

627:   simple polygon $P$ can be partitioned into $r+1$ (pairwise-disjoint)

628:   convex pieces; this is a simple observation of

629:   Chazelle~\cite{c-cgc-79} (see Theorem~2.5.1 of \cite{o-cgc-98}).

630:   The points $S$ occur as a subset of the vertices of these pieces;

631:   thus, the partitioning also decomposes $S$ into at most $r+1$

632:   subsets, each in convex position.  Since $\kappa_p(n)\geq \lceil

633:   (n-1)/4\rceil$, we get that $r\geq \lceil (n-1)/4\rceil -1$.

634: %

635: \end{proof}

636: 

637: \begin{figure}

638: \centerline {\input{Steiner-upper.pstex_t}}

639: \caption{Polygonalization of $n$ points using only $\ceil{n/3}$ 

640: reflex (Steiner) points. }

641: \label{fig:Steiner-upper}

642: \end{figure}

643: 

644: \subsection{Two-Layer Point Sets}

645: 

646: Let $S$ be a point set that has two (convex) layers.

647: It is clear from our repeated use of the example

648: in Figure~\ref{fig:lower-bnd} that this

649: is a natural case that is a likely candidate for worst-case behavior.

650: With a very careful analysis

651: of this case, we are able to obtain tight combinatorial bounds on

652: the worst-case reflexivity in terms of~$n$. 

653: 

654: \begin{theorem}

655: \label{thm:2-layer-bounds}

656: Let $S$ be a set of $n$ points having two layers.  Then $\rho(S)\leq

657: \ceil{n/4}$, and this bound is tight in the worst case.

658: \end{theorem}

659: 

660: \begin{proof}

661: Consider a set $S$ of $n$ points with onion depth two.  Let $h$ be the

662: number of points on the convex hull.  Thus there are $n - h$ points on

663: the interior onion layer.  Let the points on the convex hull be $a_0,

664: a_1,\dots, a_{h-1}$ in clockwise order.  Let the points on the

665: interior onion layer be $b_0, b_1,\dots, b_{n - h - 1}$ in clockwise

666: order.  All arithmetic involving the subscripts of the $a$'s and $b$'s

667: is done mod $h$ and mod $(n - h)$, respectively.

668: 

669: \begin{fact}

670: \label{fact:0}

671: In any polygonalization of $S$ each pocket has at least one

672:         reflex vertex.

673: \end{fact}

674: 

675: \begin{proof} 

676:   This follows from the fact that the polygon defined by the pocket

677:   and its convex hull edge must have at least three convex vertices

678:   (as does any simple polygon).

679: \end{proof}

680: 

681: Consider any point $b_i$ on the interior onion layer.  Let the

682: intersection of ray $\vecc{b_{i - 1} b_i}$ with the convex hull be $x_i\in

683: a_j a_{j+1}$.  We call the directed segment $b_i x_i$ the {\em spoke

684:   $s_i$ originating at $b_i$}, and we say that the spoke $s_i$ belongs

685: to the convex hull segment $a_j a_{j+1}$ and the segment $a_j a_{j+1}$

686: has the spoke~$s_i$.  Refer to Figure~\ref{fig:twolayer}.

687: 

688: The number of spokes a convex hull segment has is called its {\em

689:   spoke count}.  Let $c_j$ be the spoke count of convex hull segment

690: $a_j a_{j+1}$.  Clearly there are $n - h$ spokes $s_0, s_1,\dots, s_{n

691:   - h - 1}$ and each spoke belongs to exactly one convex hull segment

692: (assuming no degeneracy).  Also for all $j$, $0 \leq c_j \leq n-h$ and

693: $\sum_{j=0}^{h-1} c_j = n - h$.

694: 

695: \begin{figure}[htbp]

696: \centerline{\input{twolayer.pstex_t}}

697: \caption{The situation for a point set with two layers.}

698: \label{fig:twolayer}

699: \end{figure}

700: 

701: Consider any convex hull edge $a_j a_{j+1}$ with non-zero spoke count

702: $c_j$.  Assume that the originating points of its spokes are $b_i$,

703: $b_{i+1}$,$\dots$, $b_{i + c_j - 1}$ in clockwise order.  Then the pocket

704: $a_j$, $b_i$, $b_{i+1}$,$\dots$, $b_{i + c_j - 1}$, $a_{j+1}$ is called the {\em

705:   standard pocket} for the convex hull edge $a_j a_{j+1}$.  See

706: Figure~\ref{fig:standard-pockets}.

707: 

708: \begin{figure}[htbp]

709: \centerline{\input{standard_pockets.pstex_t}}

710: \caption{Standard pockets.}

711: \label{fig:standard-pockets}

712: \end{figure}

713: 

714: \begin{fact}

715: \label{fact:1} 

716: A standard pocket has exactly one reflex vertex.

717: \end{fact}

718: 

719: \begin{proof}

720:   Vertices $b_{i+1}, b_{i+2},\dots, b_{i + c_j - 2}$ are convex, since

721:   the angles at these vertices are the interior angles of the inner

722:   onion layer.  The vertex $b_{i + c_j - 1}$ is convex because point

723:   $a_{j+1}$ is on the right of directed line $b_{i + c_j - 2} b_{i +

724:     c_j - 1}$.  The vertex $b_i$ is reflex because there has to be at

725:   least one reflex vertex in any pocket.

726: \end{proof}

727: 

728: \begin{fact}

729: \label{fact:2}

730: No two standard pockets intersect each other, except possibly at their

731: endpoints.

732: \end{fact}

733: 

734: \begin{proof}

735:   The annulus between the two onion rings is divided into $n-h$

736:   disjoint regions by the spokes.  A standard pocket for segment $a_j

737:   a_{j+1}$ may share a region with each of the standard pockets (if

738:   any) for segments $a_{j-1} a_j$ and $a_j a_{j+1}$.  In such shared

739:   regions two pockets have a segment each.  The two segments do not

740:   intersect because the two segments are obtained by rotating two

741:   spokes about their points of origin until their intersection with

742:   the convex hull reaches a vertex.  The two segments can hence either

743:   remain disjoint or share an endpoint.

744: \end{proof}

745: 

746: We obtain the {\em standard polygonalization} by connecting all

747: standard pockets, in order, using convex hull segments with zero spoke

748: count.  See Figure~\ref{fig:standard-pockets}.

749: 

750: \begin{fact}

751: \label{lem1}

752: The number of reflex vertices in the standard polygonalization is at

753: most $\floor{n/2}$.

754: \end{fact}

755: 

756: \begin{proof} 

757:   The number of standard pockets can neither exceed the number of

758:   convex hull segments $h$ nor the number of internal points, $n-h$.

759:   Hence, this polygonalization has at most $\min\{h, n-h\}$ standard

760:   pockets and hence reflex vertices.  This can be at most

761:   $\floor{n/2}$.

762: \end{proof}

763: 

764: Consider a convex hull segment $a_j a_{j+1}$ that has a 

765: non-zero spoke count

766: $c_j$.  Assume that the origins of its spokes are $b_i, b_{i+1},\dots, b_{i +

767:   c_j - 1}$.  We call the pocket $a_j$, $b_{i-i}$, $b_i$, $b_{i+1},\dots$,

768: $b_{i + c_j - 1}$, $a_{j+1}$ the {\em premium pocket} for convex hull

769: segment $a_j a_{j+1}$.  Note that the premium pocket of a segment has

770: one more point than its standard pocket.  We obtain the {\em premium

771:   polygonalization} as follows: Start with the convex hull.

772: Process convex hull segments in clockwise order beginning anywhere.

773: If the convex hull segment $a_i a_{i+1}$ has spoke count greater than

774: or equal to two, replace the edge with its standard pocket.  If the

775: spoke count is zero, do nothing.  If the spoke count is one, move to

776: the next segment with non-zero spoke count and replace it with its

777: premium pocket provided that it is not already processed.  If the next

778: segment with non-zero spoke count was already processed, replace the

779: segment being processed with its standard pocket, and we are done.

780: One can verify that this process gives a valid polygonalization.

781: 

782: \begin{fact}

783: \label{lem2}

784: The number of reflex vertices in a premium polygonalization is at most

785: $\floor{n/3}$.

786: \end{fact}

787: 

788: \begin{proof}

789:   In the premium polygonalization, each pocket has at least one reflex

790:   and one convex vertex, with the exception of the last pocket

791:   created, which may have only a single reflex vertex.  Thus, the

792:   number of pockets created cannot exceed $\floor{(n-h+1)/2}$.  Also,

793:   the number of pockets cannot exceed the number of convex hull edges,

794:   $h$.  Thus, the number of pockets and, therefore, the number of reflex

795:   vertices cannot exceed $\min\{h, \floor{(n-h+1)/2}\}$, which can be

796:   at most $\floor{n/3}$.

797: \end{proof}

798: 

799: We now define the {\em intruding polygonalization}, as follows: Start

800: with the convex hull.  Process convex hull segments in clockwise order, 

801: beginning anywhere.  We consider cases, depending on the spoke count, $c_j$, of

802: the convex hull segment $a_j a_{j+1}$:

803: %

804: \begin{description}

805: \item[$c_j=0$] Do nothing.

806: 

807: \item[$c_j\geq 3$] Replace $a_j a_{j+1}$ with a standard pocket.

808:   

809: \item[$c_j=1$] Replace the next non-zero count unprocessed hull

810:   segment with its premium pocket.  (If no such non-zero count

811:   unprocessed segment exists, then replace $a_j a_{j+1}$ by its

812:   standard pocket and stop.)

813: 

814: \item[$c_j=2$]  We distinguish two subcases:

815: \begin{description}

816: \item[Case A] If this is the last segment to be processed or if the

817:   sequence of spoke counts following this segment begins with either

818:   0, or a count $\geq 4$, or an odd number of 1's, or an even number

819:   of 1's followed by a zero, do the following: Replace $a_j a_{j+1}$

820:   with its standard pocket.

821: \item[Case B] If there are an even number of 1's followed by a

822:   non-zero spoke count, or if the sequence begins with 2 or 3, do one

823:   of the following, depending on the current ``mode''; initially, the

824:   mode is ``normal.''

825:   

826:   {\bf Normal Mode.} Replace $a_{j+1} a_{j+2}$ with its premium

827:   pocket.  This leaves one of the originating points of $a_j a_{j+1}$;

828:   call a point $P$ of this type 

829:   the {\em pending point}.  Go into ``Point

830:   Pending Mode.''

831:   

832:   {\bf Point Pending Mode.} Let the origins of the spokes of $a_j

833:   a_{j+1}$ be $b_i, \dots, b_{i + c_j - 1}$.  Find the first point of

834:   intersection of the ray $\vecc{P b_i}$ with the current polygonalization.

835:   There are only two choices for where the point of intersection can

836:   lie.

837:   

838:   If the point of intersection lies on $a_j a_{j+1}$, treat ray $\vecc{P

839:   b_i}$ as a spoke of $a_j a_{j+1}$, thus increasing $c_j$ by 1.  Now

840:   replace $a_j a_{j+1}$ by its standard pocket, and return to the

841:   ``Normal Mode.''

842:   

843:   If the point of intersection lies on $a_j b_{i-1}$, replace $a_j

844:   b_{i-1}$ with $a_j b_i P b_{i-1}$.  This reduces $c_j$ from 2 to 1.

845:   Now handle as in Case~1.

846: \end{description}

847: \end{description}

848: 

849: 

850: \begin{fact}

851: \label{lem3}

852: The intruding polygonalization produces a valid (non-crossing) polygonalization of $S$, with at most $\floor{n/4}$ reflex vertices.

853: \end{fact}

854: 

855: \begin{proof}

856: It is straightforward to check each case to make certain that the

857: polygonalization is valid.

858: We prove the bound on the number of reflex vertices by charging each

859: reflex vertex created to a set of 4 input points.

860: 

861: In case $c_j=0$, we do not create any reflex points.

862: 

863: In case $c_j\geq 3$, we create a reflex vertex as part of the standard

864: pocket.  We charge this to the three internal vertices of the pocket

865: and the source vertex of the segment being processed.

866: 

867: In case $c_j=1$, we create a reflex vertex as part of the premium

868: pocket of the next segment having non-zero spoke count.  We charge

869: this to the sources of the segment being processed and the next

870: segment and to the (at least two) internal vertices of the pocket.

871: 

872: If $c_j=2$ and we are in Case A, we create a reflex vertex as part of

873: the standard pocket.  We charge this to the source of the segment

874: being processed and the two internal vertices.  We need to charge it

875: to one more point.

876: \begin{itemize}

877: \item If the next segment has zero spoke count, we charge it to its

878:   source.

879: \item If the next segment has spoke count greater than or equal to

880:   four, we charge it to one of its internal vertices.  (Recall that

881:   for a segment with spoke count greater than or equal to four, only

882:   three of its internal vertices will get charged by its own standard

883:   pocket.)

884: \item If there are some number of 1's followed by a 0, we charge the

885:   source of segment having spoke count 0.

886: \item If there are odd number of 1's followed by a non-zero count,

887:   then note that the segment following this sequence of 1's will be

888:   replaced by its premium pocket by our algorithm.  This segment will

889:   have a spoke count greater than or equal to 2.  We put the extra

890:   charge needed on one of the internal points of this premium pocket.

891:   (Recall that this premium pocket will have at least 3 internal

892:   points and the pocket itself will charge only two of them.)

893: \end{itemize}

894: 

895: If $c_j=2$ and we are in Case B, we consider each of the two modes separately:

896: %

897: \begin{description}

898: \item[Normal Mode] If the sequence begins with a 2 or 3, we charge the

899:   reflex vertex created to the internal points (at least 3) of the

900:   premium pocket and the source of segment being replaced.

901:   

902:   If the sequence has an even number of 1's followed by a non-zero

903:   spoke count, then we have a case similar to that above, but we have

904:   only two internal points to charge in the premium pocket.  However,

905:   since the first segment with spoke count 1 will be replaced by its

906:   premium pocket, there will be an odd number of 1's remaining.

907:   Hence, the segment following this sequence of one (which has spoke

908:   count $\geq 2$) will be replaced with a premium pocket, which has an

909:   extra internal point to which we can charge.

910:   

911:   Note that in Normal Mode, in both of the cases above, 

912:   the source of the

913:   segment being processed remains uncharged.  Also, the pending point

914:   remains to be incorporated into the polygonalization.

915: \item[Point Pending Mode] Depending on where ray $\vecc{P b_i}$ intersects in

916:   this case, there are two possibilities:

917:   

918:   If the point of intersection lies on $a_j a_{j+1}$, then we are

919:   creating a pocket with at least 2 internal points.  We charge the

920:   reflex vertex created to the two internal points and to the source

921:   of the segment being processed and to the uncharged point in the

922:   previous Normal Mode.  See Figure~\ref{fig:case2b1}.

923: 

924: 

925: \begin{figure}

926: \centerline{\input{case2b1.pstex_t}}

927: \caption{Subcase of Case B, $c_j=2$: Ray ${b_0 b_4}$ intersects segment $a_2 a_3$, but not segment $b_3 a_2$.}

928: \label{fig:case2b1}

929: \end{figure}

930: 

931: 

932: \begin{figure}

933: \centerline{\input{case2b2.pstex_t}}

934: \caption{Subcase of Case B, $c_j=2$: Ray ${b_0 b_4}$ intersects segment $b_3 a_2$, but not segment $a_2 a_3$.}

935: \label{fig:case2b2}

936: \end{figure}

937: 

938:   If the point of intersection lies on $a_j b_{i-1}$, then the

939:   subpocket has two internal points.  In this case we

940:   charge its reflex point

941:   (which is the pending point $P$) to these two internal points and to

942:   the uncharged point in the previous normal mode.  We need to make

943:   one more charge, which depends on the spoke count of next segment.

944:   If it is 2 or 3, we charge the extra internal point in its premium

945:   pocket.  If it is 1, note that since there were even number of 1's

946:   following, after we replace the next one with its premium pocket,

947:   there would be odd number of 1's remaining.  Hence, the segment

948:   following this sequence of 1's (which has non-zero spoke count),

949:   will be replaced with its premium pocket, by our algorithm.  Hence,

950:   we can charge its extra internal point.  See Figure~\ref{fig:case2b2}.

951: \end{description}

952: 

953: \end{proof}

954: 

955: This concludes the proof of Theorem~\ref{thm:2-layer-bounds}

956: \end{proof}

957: 

958: 

959: \paragraph{Remark.}

960: Using a variant of the polygonalization given in the proof of

961: Theorem~\ref{thm:2-layer-bounds}, it is possible to show that a

962: two-layer point set $S$ in fact has a polygonalization with at most

963: $\ceil{n/3}$ reflex vertices such that none of the edges in the

964: polygonalization pass through the interior of the convex hull of the

965: second layer.  (The polygonalization giving upper bound of

966: $\ceil{n/4}$ requires edges that pass through the interior of the

967: convex hull of the second layer.) This observation may be useful in

968: attempts to reduce the worst-case upper bound ($\rho(n)\leq

969: \ceil{n/2}$) for more general point sets~$S$.

970: 

971: \subsection{Convex Cover/Partition Numbers}

972: 

973: As a consequence of the Erd\H{o}s-Szekeres

974: theorem~\cite{es-cpg-35,es-sepeg-60}, Urabe has given bounds on the

975: convex cover number of a set of $n$ points:

976: Urabe~\cite{u-opcp-96} and Hosono and Urabe~\cite{hu-ndcqp-01}

977: have obtained bounds as well on the convex partition number

978: of an $n$-point set:

979: %

980: $$\left\lceil\frac{n-1}{4}\right\rceil \leq \kappa_p(n)\leq \left\lceil\frac{5n}{18}\right\rceil.$$

981: %

982: 

983: While it is trivially true that $\kappa_c(S)\leq \kappa_p(S)$

984: the ratio $\kappa_p(S)/\kappa_c(S)$ for a set $S$

985: may be as large as $\Theta(n)$; the set $S=S_0(n)$ (Figure~\ref{fig:lower-bnd})

986: has $\kappa_c(S)=2$, but $\kappa_p(S)\geq n/4$.

987: 

988: The fact that $\kappa_p(S)\leq \rho(S)+1$ follows easily by

989: iteratively adding $\rho(S)$ segments to an optimal polygonalization

990: $P$, bisecting each reflex angle. The result is a partitioning of $P$

991: into $\rho(S)+1$ convex pieces.

992: Thus, we can obtain a convex partitioning of $S$ by associating a

993: subset of $S$ with each convex piece of $P$, assigning each point of

994: $S$ to the subset associated with any one of the convex pieces that has

995: the point on its boundary.  

996: %

997: (This is the same observation of Chazelle~\cite{c-cgc-79} used in

998:   the proof of Theorem~\ref{thm:Steiner-upper}.)

999: 

1000: We believe that the relationship between reflexivity ($\rho(S)$)

1001: and convex partition number ($\kappa_p(S)$) goes the other way as well:

1002: A small convex partition number should imply a small reflexivity.

1003: In particular, we have invested considerable effort in trying to

1004: prove the following conjecture:

1005: %

1006: \begin{conjecture}

1007: \label{con:rho.kappa2}

1008: $\rho(S)=O(\kappa_p(S))$. 

1009: \end{conjecture}

1010: %

1011: The reflexivity can be as large as {\em twice} the

1012: convex cover number ($\rho(S)=2\kappa_p(S)$), as illustrated

1013: in the example of Figure~\ref{fig:ratio-two}; however,

1014: this is the worst class of examples we have found so far.

1015: 

1016: \begin{figure}[hbtp]

1017:    \begin{center}

1018:    \epsfxsize=.4\textwidth

1019:    \ \epsfbox{ratio.two.eps}

1020:    \caption{An example with $\rho(S)=2\kappa_p(S)$.

1021:    Each thick oval shape represents a numerous subset of points of $S$

1022: in convex position.}

1023:    \label{fig:ratio-two}

1024:    \end{center}

1025: \end{figure}

1026: 

1027: Turning briefly to Steiner reflexivity, it is not hard to see that

1028: $\rho'(S)=O(\kappa_p(S))$ (see the proof of

1029: Corollary~\ref{cor:reflexivity-approx}). 

1030: %

1031: Thus, a proof of

1032: Conjecture~\ref{con:rho.kappa2} would follow from the validity of

1033: Conjecture~\ref{con:rho.rho'}.

1034: 

1035: 

1036: 

1037: \subsection{Small Point Sets}

1038: \label{sec:small}

1039: 

1040: It is natural to consider the exact values of $\rho(n)$,

1041: $\kappa_c(n)$, and $\kappa_p(n)$ for small values of $n$.

1042: Table~\ref{tab:small} below shows some of these values, which we

1043: obtained through (sometimes tedious) case analysis.  Aichholzer and

1044: Krasser~\cite{ak-psotd-01} have recently applied their software that

1045: enumerates point sets of size $n$ of all distinct order types to

1046: verify our results computationally; in addition, they have obtained

1047: the result that $\rho(10)=3$.  (Experiments are currently

1048: under way for $n=11$; values of $n\geq 12$ seem to be

1049: intractable for enumeration.)

1050: 

1051: \begin{table}

1052: \small

1053: \centerline{%

1054: \begin{tabular}{|l||c|c|c|}

1055: \hline

1056: $n$ & $\rho(n)$ & $\kappa_c(n)$ & $\kappa_p(n)$ \\

1057: \hline \hline

1058: $\leq $ 3 &0 &1 & 1\\

1059:        4 &1 &2 & 2\\

1060:        5 &1 &2 & 2\\

1061:        6 &2 &2 & 2\\

1062:        7 &2 &2 & 2\\

1063:        8 &2 &2 & 3\\

1064:        9 &3 &3 & 3\\

1065:       10 &3 &- &- \\

1066: \hline

1067: \end{tabular}%

1068: }

1069: \vspace{1ex}

1070: \caption{\label{tab:small}

1071: Worst-case values of $\rho$, $\kappa_c$, $\kappa_p$ for small

1072: values of $n$.

1073: }

1074: \end{table}

1075: 

1076: \section{Complexity}

1077: \label{sec:npc}

1078: 

1079: We now prove lower bounds on the

1080: complexity of computing the convex cover number, $\kappa_c(S)$, and

1081: the convex partition number, $\kappa_p(S)$.

1082: %

1083: The proof for the convex cover number uses a reduction of the problem

1084: 1-in-3 SAT and is inspired by the hardness proof for the {\sc Angular

1085:   Metric TSP} given in~\cite{ackms-amtsp-97}.  The proof for the

1086: convex partition number uses a reduction from {\sc Planar 3\,Sat}.

1087: 

1088: 

1089: 

1090: \begin{theorem}

1091: \label{th:npc.disjoint}

1092: It is NP-complete to decide whether for a planar point set $S$

1093: the convex partition number $\kappa_p(S)$ is below some

1094: threshold $k$.

1095: \end{theorem}

1096: 

1097: 

1098: \begin{proof}

1099: We give a reduction from {\sc Planar 3\,Sat}, which was shown to be

1100: NP-complete by Lichtenstein (see~\cite{l-pftu-82}). A {\sc 3\,Sat} instance $I$

1101: is called a {\sc Planar 3\,Sat} instance, if the (bipartite) 

1102: ``occurrence graph''

1103: $G_I=(V_I,E_I)$ is planar, where each vertex of $V_I$ 

1104: corresponds to a variable or a clause, and two vertices

1105: are joined by an edge of $E_I$ if and only if the vertices

1106: correspond to a variable $x$ and a clause $c$ such that $x$ appears 

1107: in the clause $c$ in $I$.

1108: See Figure~\ref{fi:npc.disjoint}(a) for an example, 

1109: where a solid edge denotes

1110: an un-negated literal, while a dashed edge represents a negated literal

1111: in a clause.

1112: 

1113: \begin{figure}[h!btp]

1114:    \begin{center}

1115:    \epsfxsize=.20\textwidth

1116: \ \epsfbox{occur.eps}

1117: \hfill

1118:    \epsfxsize=.100\textwidth

1119: \ \epsfbox{polygon.eps}

1120: \hfill

1121:    \epsfxsize=.25\textwidth

1122: \ \epsfbox{replace.eps}

1123: \hfill

1124:    \epsfxsize=.25\textwidth

1125: \ \epsfbox{join.eps}

1126: \caption{(a) A straight-line embedding of the occurrence graph for the 

1127: {\sc 3\,Sat} instance $(x_1\vee\overline{x_2}\vee x_3) \wedge (x_2\vee{x_3}\vee\overline{x_4})\wedge (\overline{x_1}\vee x_2 \vee \overline{x_4})$;

1128: (b) a polygon for a variable vertex;

1129: (c) a point set $S_I$ representing the {\sc Planar 3\,Sat} instance $I$;

1130: (d) joining point sets along the odd or even polygon edges.}

1131: \label{fi:npc.disjoint}

1132:    \end{center}

1133: \end{figure}

1134: 

1135: The basic idea is the following: Each variable is represented

1136: by a set of points that can be partitioned into $s$ disjoint 

1137: convex chains in two different ways. One of these possibilities 

1138: will correspond to a setting of ``true'', the other to a setting of 

1139: ``false''.  Each clause is represented by a set of points, such

1140: that it can be covered by three convex chains disjoint from

1141: all other chains, if and only if at least one of the variables is set

1142: in a way that satisfies the clause.

1143: 

1144: So let $I$ be a {\sc Planar 3\,Sat} instance with $n$ variables 

1145: and $m=O(n)$ clauses. Consider a straight-line

1146: embedding of its occurrence graph $G_I$. In a first step, for 

1147: every vertex $v_x$ representing a variable $x$, draw 

1148: a small polygon $P_x$ with $2s\leq 2m$ edges 

1149: around $v_x$, where $s$ is bounded by the maximum degree of a variable

1150: vertex. (See Figure~\ref{fi:npc.disjoint}(b).) This is done

1151: such that no edge of $G_I$ passes through one of the corners of $P_x$,

1152: and only edges adjacent to $v_x$ intersect $P_x$; moreover, we

1153: choose edge orientations for all polygons that assure that 

1154: no two different polygons $P_x$ and $P_y$ have two collinear edges

1155: (See Figure~\ref{fi:npc.disjoint}(c), where

1156: we have $s=2$ and we have used rectangles as polygons to keep the figure clear.)

1157: Furthermore, we choose the polygons such that

1158: all of the line segments connecting $v_x$ to clauses where

1159: $v_x$ appears un-negated intersect ``even'' edges of the polygon, 

1160: and all the line segments connecting $v_x$ to clauses

1161: where $v_x$ appears negated intersect ``odd'' edges of the polygon.

1162: 

1163: In a second step,

1164: replace the polygons (and thus the variable vertices) by

1165: appropriate sets of $2s(kn^4+1)$ points

1166: -- see Figure~\ref{fi:npc.disjoint} for an example.

1167: Each of the corners $b_x^{(1)},\ldots,b_x^{(2s)}$ 

1168: of a polygon $P_x$

1169: is represented by a dense convex chain $B_x^{(i)}$

1170: of $kn^4$ points, forming

1171: a convex curve with an opening of roughly $\pi/s$, and an additional 

1172: ``pivot'' point $p_x^{(i)}$. 

1173: 

1174: Finally, we describe how to represent the clauses:

1175: Each vertex $v_c$ in $G_I$ representing a clause $c$ is replaced by

1176: three points forming a small triangle $T_c$; furthermore, for each

1177: of the three edges connecting some

1178: variable vertex $v_x$ to $v_c$, add a convex chain $C_{x,c}$

1179: of $k*n^2$ points within the polygon $P_x$. $C_{x,c}$

1180: is lined up with the triangle $T_c$

1181: such that $T_c\cup C_{x,c}$ is a convex chain.

1182: On the other hand, the opening of $C_{x,c}$ towards $T_c$

1183: is narrow enough to prevent any other point 

1184: from forming a convex chain with

1185: all points in $C_{x,c}$. We also avoid any other 

1186: collinearities in the overall arrangement. 

1187: 

1188: \smallskip

1189: Now we claim the following correspondence:

1190: 

1191: The resulting point set $S_I$ can be partitioned into at

1192: most $3m+sn$ disjoint convex chains, if and only if the {\sc Planar 3\,Sat}

1193: instance $I$ is satisfiable.

1194: 

1195: \smallskip

1196: To see that a satisfying truth assignment induces a feasible decomposition, 

1197: note that there are exactly two ways to cover the points for a

1198: polygon with $2s$ edges by at most $s$ disjoint convex chains.

1199: (One of these choices arises by joining the pairs of sets that belong

1200: to the ``even'' edges of a polygons box into one convex chain, 

1201: the other by joining the pairs of sets for the ``odd'' edges

1202: of a polygon. This is shown in Figure~\ref{fi:npc.disjoint}(d).)

1203: For a true variable, take the

1204: ``even'' choice, for a false variable, the ``odd'' choice.

1205: Now consider a clause $c$ and a variable $x$ that satisfies $c$.

1206: By the choice of chains covering the points for $x$,

1207: we join the triangle $T_c$ for $v_c$ with the chain

1208: $C_{x,c}$ into one convex chain, without intersecting any other chain. 

1209: The other two chains $C_{y,c}$ are each covered by separate chains.

1210: This yields a decomposition into $3m+sn$ disjoint convex chains, as claimed.

1211: 

1212: \smallskip

1213: To see the converse, assume we have a decomposition

1214: into at most $3m+sn$ disjoint convex chains.

1215: 

1216: We start by considering how the sets $B_x^{(i)}$ and the sets

1217: $C_{x,c}$

1218: (henceforth called ``gadget sets'' $G_j$, with 

1219: $j=1,\ldots,3m+2sn$) can be covered in such a solution.

1220: Associate each chain with all the $G_j$

1221: of which it covers at least $\Omega(n)$ points.

1222: It is straightforward to see that a convex chain

1223: that covers at least $\Omega(n)$ points from each of three

1224: different gadget sets must contain

1225: some other point of the set $S_I$ 

1226: in its interior, so this cannot occur in the given feasible

1227: decomposition. Therefore, no chain can be associated

1228: with more than two gadget sets. Moreover, a convex chain can

1229: contain at least $k*n^3-O(n)$ points from each of two different

1230: gadget sets without any pivot points $p_x^{(i)}$

1231: in its interior, only if these sets are some 

1232: $B_x^{(i)}$ and $B_x^{(i\pm 1)}$.

1233: (In particular, it is not hard to see that no chain that covers

1234: $\Omega(n)$ points of a set $C_{x,c}$ can cover points

1235: from any other gadget set.)

1236: Since there are $3m+sn=O(n^2)$ chains in total, there

1237: must be at least one chain associated with each

1238: gadget set, and a chain associated with a set $C_{x,c}$

1239: cannot be associated with any other set. This means that

1240: $3m$ of the chains are used to cover the $3m$ chains $C_{x,c}$.

1241: Therefore, the remaining

1242: $2sn$ gadget sets $B_x^{(i)}$ must be covered by the remaining

1243: $sn$ convex chains. None of these chains can cover

1244: more than two of these sets. It follows that the remaining chains form a 

1245: perfect matching on the set of $B_x^{(i)}$, implying that

1246: each chain covers a $B_x^{(i)}$ and a neighboring set $B_x^{(i\pm 1)}$.

1247: Therefore, the gadget sets for each variable are either covered

1248: by pairing along the ``odd'' edges, or by pairing along the ``even''

1249: edges of the associated polygon.

1250: 

1251: This describes the gadget sets associated with all the convex chains.

1252: Now it is straightforward to verify that a point from one of the triangles

1253: $T_c$ can only be part of one of the chains associated with a

1254: set $C_{x,c}$. This can only happen if this chain does not

1255: intersect a chain along an edge of the polygon for the variable $x$

1256: -- implying that this variable satisfies the clause $c$. 

1257: This completes the proof.

1258: \end{proof}

1259: 

1260: 

1261: \begin{theorem}

1262: \label{th:npc.nondisjoint}

1263: It is NP-complete to decide whether for a planar point set $S$

1264: the convex cover number $\kappa_c(S)$ is below some

1265: threshold $k$.

1266: \end{theorem}

1267: 

1268: \begin{proof}

1269: Our proof uses a reduction of the problem 1-in-3 SAT.

1270: It is inspired by the hardness proof for the {\sc Angular Metric

1271: TSP} given in \cite{ackms-amtsp-97}.

1272: 

1273: \begin{figure}[hbtp]

1274:    \begin{center}

1275:    \epsfxsize=.35\textwidth

1276: \ \epsfbox{1in3.eps}

1277: \caption{A point set $S_I$ for a 1-in-3 SAT instance $I$. Pivot points

1278: are shown for the clause $(x_1\vee \overline{x_i}\vee x_n)$.}

1279: \label{fi:npc.nondisjoint}

1280:    \end{center}

1281: \end{figure}

1282: 

1283: The construction is as follows:

1284: For a 1-in-3 SAT instance $I$ with $n$ variables

1285: and $m$ clauses, represent each clause

1286: $c_j$ by a triple of vertical columns, each one

1287: associated with a variable that occurs in $c_j$.

1288: Each variable $x_i$ is represented by a pair of 

1289: horizontal rows, the upper one corresponding to ``true'',

1290: the lower one corresponding to ``false''.

1291: This results in a grid pattern

1292: as shown in Figure~\ref{fi:npc.nondisjoint}.

1293: For each pair of a variable $x_i$ and a clause $c_j$, we get

1294: a 2x3 pattern of intersections points. If variable 

1295: $x_i$ appear in clause $c_j$,

1296: we add three ``pivot'' points to this pattern:

1297: If $x_i$ occurs un-negated in $c_j$, a pivot point is 

1298: added at the intersection of $x_i$'s ``true'' row with 

1299: the column of $c_j$ that corresponds

1300: to $x_i$; for both other two columns, a pivot point is added 

1301: at the intersection with the ``false'' row.

1302: If $x_i$ occurs negated in $c_j$, a pivot point is added at the 

1303: intersection of $x_i$'s ``false'' row with the column of $c_j$ 

1304: that corresponds to $x_i$; for both other two columns, a pivot point is 

1305: added at the intersection with the ``true'' row.

1306: 

1307: Finally, a horizontal ``staple'' gadget that consists

1308: of $\Omega(n^4)$ points is added to the rows for each 

1309: variable, and two nested vertical

1310: staple gadgets are added to the columns for each of the clauses.

1311: These are constructed in a way that a staple forms a

1312: convex chain that can cover all the pivot points in

1313: one row or one column, but not more than that,

1314: and no pivot points from any other clauses or

1315: variables. Thus, for each variable, we can 

1316: collect one of two rows, and for each clause, we can cover

1317: two of three rows.

1318: 

1319: Now it is possible to show the following:

1320: The 1-in-3 SAT instance $I$ has a satisfying truth assignment,

1321: if and only if the point set $S_I$ can be covered with not more than 

1322: $n+2m$ convex chains. 

1323: 

1324: It is easy to see that a satisfying truth assignment

1325: implies a small convex cover, by covering each staple gadget

1326: by one chain, and choosing the appropriate

1327: rows and columns of pivot points to be covered by the 

1328: staple gadgets: For each variable, choose

1329: the row corresponding to its truth assignment. For each clause,

1330: choose the two columns for the two variables that do not 

1331: satisfy it. Now it is straightforward to check that all pivot points

1332: are covered.

1333: 

1334: To see the converse, we can argue in a similar way as in the proof

1335: of Theorem~\ref{th:npc.disjoint} that each staple needs its

1336: own convex chain. Then we are left with a choice of one row

1337: for each variable, and two columns for each clause.

1338: This choice of rows induces a truth assignment to variables;

1339: it is not hard to check that the remaining uncovered pivot points

1340: lie in not more than two columns, if and only if this truth assignment

1341: is valid for the 1-in-3 SAT instance $I$.

1342: \end{proof}

1343: 

1344: 

1345: So far, the complexity status of determining the reflexivity of a

1346: point set remains open. However, the apparently close relationship

1347: between convex cover/partition numbers and reflexivity leads us to

1348: believe the following:

1349: 

1350: \begin{conjecture}

1351: \label{npc.reflex}

1352: It is NP-complete to determine the reflexivity $\rho(S)$

1353: of a point set.

1354: \end{conjecture}

1355: 

1356: \section{Algorithms}

1357: 

1358: We have obtained a number of algorithmic results on computing, exactly

1359: or approximately, reflexivity and convex cover/partition numbers.  We

1360: begin with the following theorem, which shows that one can compute

1361: efficiently a constant-factor approximation to the {\em convexivity}

1362: of~$S$:

1363: 

1364: \begin{theorem}

1365: \label{thm:approx-convexity}

1366: Given a set $S$ of $n$ points in the plane, in $O(n\log n)$ time one

1367: can compute a polygonalization of $S$ having at least 

1368: $\chi(S)/2$

1369: convex vertices, where $\chi(S)=n-\rho(S)$ is the convexivity of~$S$.

1370: \end{theorem}

1371: 

1372: \begin{proof} 

1373:   The proof of Theorem~\ref{thm:reflex-upper-bnd} is constructive,

1374:   producing a polygonalization of $S$ having at most $\ceil{n_I/2}\leq

1375:   \ceil{n/2}$ reflex vertices, and thus at least $\floor{n/2}$ convex

1376:   vertices (thereby giving a 2-approximation for convexivity).  In

1377:   order to obtain the stated time bound, we must implement the

1378:   algorithm efficiently.  This can be done using a data structure for

1379:   dynamic convex hulls, under pure deletion (see

1380:   Chazelle~\cite{c-clps-85} and Hershberger and

1381:   Suri~\cite{hs-asdch-92}).  At each main step of the algorithm, we

1382:   identify the vertices along the chain $(p_{i+1},q,q_1,\ldots,b)$ by

1383:   making repeated extreme-point queries in the convex hull data

1384:   structure, and then delete the points $p_{i+1},q,q_1,q_2,\ldots$

1385:   from the data structure, and repeat.  The dynamic data structure

1386:   supports deletions and queries in $O(\log n)$ time per operation,

1387:   for an overall time bound of $O(n\log n)$.

1388: %

1389: \end{proof}

1390: 

1391: 

1392: \begin{theorem}

1393: \label{thm:convex-cover-approx}

1394: Given a set $S$ of $n$ points in the plane, the convex cover number,

1395: $\kappa_c(S)$, can be computed approximately, within a factor $O(\log

1396: n)$, in polynomial time.

1397: \end{theorem}

1398: 

1399: 

1400: 

1401: \begin{proof}

1402:   We use a greedy set cover heuristic.  At each stage, we need to

1403:   compute a largest convex subset among the remaining (uncovered)

1404:   points of $S$.  This can be done in polynomial time using the

1405:   dynamic programming methods of~\cite{mrsw-ccppp-95}.

1406: \end{proof}

1407: 

1408: 

1409: \begin{theorem}

1410: \label{thm:disjoint-convex-cover-approx}

1411: Given a set $S$ of $n$ points in the plane, the convex

1412: partition number, $\kappa_p(S)$, can be computed approximately, within

1413: a factor $O(\log n)$, in polynomial time.

1414: \end{theorem}

1415: 

1416: \begin{proof} 

1417:   Let $C^*=\{P_1,\ldots,P_{k^*}\}$ denote an optimal solution,

1418:   consisting of $k^*=\kappa_p(S)$ disjoint convex polygons whose

1419:   vertices are the set $S$.  

1420:   

1421:   By Theorem~2 of~\cite{ers-ccsno-90}, we know that there are $k^*$

1422:   pairwise-disjoint convex polygons $P_1',\ldots,P_{k^*}'$, having a

1423:   total complexity of $O(k^*)$, with $P_i\subseteq P_i'$, for each

1424:   $i$.  Furthermore, the sides of the polygons $P_i'$ can be assumed

1425:   to be segments lying on the $O(n^2)$ lines determined by pairs of

1426:   points of $S$.  Let ${\cal S}$ be the convex polygonal subdivision

1427:   of $C=CH(P_1'\cup\cdots P_{k^*}')$ obtained by decomposing the

1428:   region $C\setminus (P_1'\cup\cdots P_{k^*}')$ into convex polygons

1429:   (e.g., a triangulation suffices).  Then, ${\cal S}$ has total

1430:   complexity $O(k^*)$, its vertices are among the set $V$ of $O(n^4)$

1431:   vertices in the arrangement of $O(n^2)$ lines determined by $S$, and

1432:   its faces are convex polygons.  (Note that some faces of ${\cal S}$

1433:   may be empty of points of $S$.)

1434:   

1435:   We now decompose each face of ${\cal S}$ into a set of vertical

1436:   trapezoids by erecting vertical cuts through the vertices of each

1437:   face, within each face.  (Some of these trapezoids may be triangles,

1438:   which we can consider to be degenerate trapezoids.)  Finally, for

1439:   each such trapezoid $\tau$ we decompose it using vertical cuts into

1440:   $O(\log n^4)=O(\log n)$ {\em canonical trapezoids}, whose

1441:   $x$-projection is one of the $O(n^4)$ canonical $x$-intervals

1442:   determined by a segment tree on $V$.  The resulting canonical

1443:   trapezoidalization, ${\cal T}$, has $O(k^*\log n)$ faces, each of

1444:   which is a canonical trapezoid.  An important property of ${\cal T}$

1445:   is that it has the following binary space partition property: For

1446:   any canonical trapezoid, $\tau$, the subdivision of $\tau$ induced

1447:   by ${\cal T}$ is such that either (a) there exists an edge of ${\cal

1448:     T}$ that cuts $\tau$ in two, extending from its left side to its

1449:   right side, or (b) the vertical cut that splits $\tau$ into two

1450:   canonical subtrapezoids lies entirely on the edge set of ${\cal T}$.

1451:   This property allows us to optimize recursively using dynamic

1452:   programming, in much the same way as was done in

1453:   \cite{as-sagp-98,m-aagsp-93} for problems involving optimal

1454:   separation and surface approximation.

1455:   

1456:   In particular, for any canonical trapezoid, $\tau$, we desire to

1457:   compute the quantity $f(\tau)$, defined to be the minimum number of

1458:   faces in a partitioning of $\tau$ into canonical trapezoids, such

1459:   that within each face of the partitioning, the subset of points of

1460:   $S$ within the face is in convex position.  (The empty set is

1461:   considered trivially to be in convex position.)  Then $f(\tau)$

1462:   obeys a recursion,

1463: %

1464:   $$f(\tau)=\min\{\min_{\ell}\{f(\tau_\ell^+)+f(\tau_\ell^-)\}, 

1465:   f(\tau_{left})+f(\tau_{right})\},$$

1466: %

1467:   where the minimization over $\ell$ considers all choices of lines

1468:   $\ell$, intersecting $\tau$ on both of its vertical sides,

1469:   determined by two vertices of $V$; $\tau_\ell^+$ (resp.,

1470:   $\tau_\ell^-$) denotes the portion of $\tau$ lying above (resp.,

1471:   below) the line $\ell$.  We have used $\tau_{left}$ (resp.,

1472:   $\tau_{right}$) to indicate the canonical trapezoid obtained by

1473:   splitting $\tau$ by a vertical line at the $x$-median value (among

1474:   the $x$-coordinates of $V$ that lie in the vertical slab defined

1475:   by~$\tau$).  We leave it to the reader to write the boundary

1476:   conditions of the recursion, which is straightforward.

1477:   

1478:   Our algorithm gives us a minimum-cardinality partition of $S$ into a

1479:   disjoint set, $C'$, of (empty) convex subsets whose $x$-projections

1480:   are canonical intervals.  Since the optimal solution, $C^*$, can be

1481:   converted into at most $k^*\cdot O(\log n)$ such convex sets, we

1482:   know we have obtained an $O(\log n)$-approximate solution to the

1483:   disjoint convex partition problem.

1484: \end{proof}

1485: 

1486: 

1487: 

1488: \begin{corollary}

1489: \label{cor:reflexivity-approx}

1490: Given a set $S$ of $n$ points in the plane, its Steiner reflexivity,

1491: $\rho'(S)$, can be computed approximately, within a factor $O(\log

1492: n)$, in polynomial time.

1493: \end{corollary}

1494: 

1495: 

1496: 

1497: \begin{proof} 

1498:   Let $P^*$ denote an optimal solution, a simple polygon having

1499:   $\rho'(S)$ reflex vertices.  Then, we know that $P^*$ can be

1500:   decomposed into at most $\rho'(S)+1$ convex polygons, each of which

1501:   corresponds to a subset of $S$.  This gives us a partition of $S$

1502:   into at most $\rho'(S)+1$ disjoint convex sets; thus,

1503:   $\kappa_p(S)\leq \rho'(S)+1$.  By

1504:   Theorem~\ref{thm:disjoint-convex-cover-approx}, we can compute a

1505:   set, $C$, of $k\leq O(\log n)\cdot \kappa_p(S)$ disjoint convex

1506:   sets.  We can polygonalize $S$ by ``merging'' these $k$ polygons of

1507:   $C$, using a doubling of a spanning tree on $C$.  The important

1508:   property of the embedding of the spanning tree is that it consists

1509:   of ($k-1$) line segment bridges (with endpoints on the boundaries of

1510:   polygons $C$) that are pairwise non-crossing and do not cross any of

1511:   the polygons $C$. (One way to determine such a tree is to select one

1512:   point interior to each polygon of $C$, compute a minimum spanning

1513:   tree of these $k$ points, and then utilize the portions of the line

1514:   segments that constitute the tree that lie outside of the polygons

1515:   $C$ to be the set of $k-1$ bridging segments.)  A simple

1516:   polygonalization is obtained by traversing the boundary of the union

1517:   of the polygons $C$ and the $k-1$ bridging segments, while slightly

1518:   perturbing the doubled bridge segments.  Since each bridge segment

1519:   is responsible for creating at most 4 new (Steiner) points, this

1520:   results in a polygon, with at most $4(k-1)$ Steiner points, each of

1521:   which may be reflex.  (All other vertices in the polygonalization

1522:   are convex.)  Thus, we obtain a polygonalization with at most

1523:   $4(k-1)\leq O(\log n)\cdot \kappa_p(S)\leq O(\log n)\cdot \rho'(S)$

1524:   reflex vertices.

1525: \end{proof}

1526: 

1527: 

1528: For small values of $r$, we have devised particularly efficient

1529: algorithms that check if $\rho(S)\leq r$ and, if so, produce a witness

1530: polygonalization having at most $r$ vertices.  Of course, the case

1531: $r=0$ is trivial, since that is equivalent to testing if $S$ lies in

1532: convex position (which is readily done in $O(n\log n)$ time, which is

1533: worst-case optimal).  It is not surprising that for any

1534: fixed $r$ one can obtain an $n^{O(r)}$ algorithm: enumerate 

1535: over all combinatorially distinct (with respect to $S$) convex subdivisions

1536: of $\textrm{CH}(S)$ into $O(r)$ convex faces and test that the subsets of $S$

1537: within each face are in convex position, and then check all

1538: possible ways to order these $O(r)$ convex chains

1539: to form a circuit that may form a simple polygon.

1540: The factor in front of $r$ in the exponent, however, is not

1541: so trivial to reduce.  In particular, the straightforward method applied

1542: to the case $r=1$ gives $O(n^5)$ time.  With a more careful analysis

1543: of the cases $r=1,2$, we obtain the next two theorems.

1544: 

1545: \begin{theorem}

1546: \label{thm:small-reflexivity-1}

1547: Given a set $S$ of $n$ points in the plane, in $O(n\log n)$ time one

1548: can determine if $\rho(S)=1$, and, if so, produce a witness

1549: polygonalization.  Further, $\Omega(n\log n)$ is a lower bound on the

1550: time required to determine if $\rho(S)=1$.

1551: \end{theorem}

1552: 

1553: 

1554: \begin{proof}

1555:   First notice that if a point set $S$ is polygonalized with exactly 1 reflex

1556:   vertex, then (a) the two supporting lines from the endpoints of the lid of

1557:   the pocket to the second convex layer of $S$  are supported

1558:   by the same point $p$ of the layer, which is the reflex vertex of the

1559:   polygonalization, and (b) the vertices of the pocket appear in

1560:   angular order around $p$. The upper bound involves computing the convex

1561:   hull and the second layer of the onion (the entire onion can be computed

1562:   in time $O(n\log n)$), and then performing a careful case

1563:   analysis for how the single pocket must be.

1564: 

1565:   In fact, if the second convex layer is empty, or has 1 or 2

1566:   points, the solution is trivial. If the second convex layer has three

1567:   points or has all the internal points, then for each edge $e$ of the

1568:   convex hull, trace the supporting lines from its endpoints to the second

1569:   layer. In the cases where the two tangents are supported by the same

1570:   point $p$, check whether or not the angular order of the interior points

1571:   gives a pocket with lid $e$ and only one reflex vertex $p$. The cost of

1572:   this step is $O(n\log n)$: if the layer is a triangle, computing the

1573:   supporting lines can be done in constant time, and the complexity comes

1574:   from sorting the interior points around the vertex $p$ of the

1575:   triangle; if the layer has all the interior points, then the sorted

1576:   order is given, and the complexity comes from the computation of the

1577:   supporting lines (in fact, the supporting lines can be computed in

1578:   overall $O(n)$ time by a ``rotating calipers''-like technique). In all

1579:   the remaining cases, the point set cannot be polygonalized with one reflex

1580:   vertex.

1581:   

1582:   The lower bound follows from convexity testing: determining if a set

1583:   of $n$ points is in convex position.  Given a set $S$ of $n$ points,

1584:   we compute (in $O(n)$ time) one edge, $e=v_1v_2$, of the convex hull

1585:   of $S$.  We then determine (in $O(n)$ time) the point $v_3\in S$

1586:   furthest from the line through $v_1$ and $v_2$; thus, $v_3$ is also

1587:   a vertex of $\textrm{CH}(S)$.

1588:   Next we let $p_1\not\in S$ be a point within $\Delta v_1v_2v_3$ that

1589:   is closer to edge $v_1v_2$ than is any point of $S$.  We also select

1590:   points $p_2$ and $p_3$ within $\Delta p_1v_1v_2$ in such a way that

1591:   $v_3$ lies within the convex cone of apex $p_3$ defined by the rays

1592:   $\vecc{p_3p_1}$ and $\vecc{p_3p_2}$ and that $v_1$ lies within the

1593:   convex cone of apex $p_2$ defined by the rays $\vecc{p_2p_1}$ and

1594:   $\vecc{p_2p_3}$.  (Points $p_1,p_2,p_3$ can be determined in $O(n)$

1595:   time.)  Refer to Figure~\ref{fig:lower-bnd-rho1}.  Then,

1596:   $\vecc{p_3p_1}$ and $\vecc{p_3p_2}$ intersect distinct

1597:   edges of $\textrm{CH}(S)$, as do $\vecc{p_2p_1}$ and $\vecc{p_2p_3}$, while

1598:   $\vecc{p_1p_2}$ and $\vecc{p_1p_3}$ both intersect the edge $v_1v_2$

1599:   of $\textrm{CH}(S)$.  This implies that the only way that a polygonalization

1600:   of $S'=S\cup\{p_1,p_2,p_3\}$ can have only a single reflex vertex is

1601:   if that reflex vertex is $p_1$ and the corresponding pocket has lid

1602:   $v_1v_2$ and vertices $p_1,p_2,p_3$ (with $p_2$ and $p_3$ both being

1603:   convex in the polygonalization).  Thus, the only way to have

1604:   $\rho(S')=1$ is if the points $S$ are in convex position.  We

1605:   conclude that determining if $\rho(S')=1$ is equivalent to solving

1606:   the convex position problem on the input points~$S$.

1607: \end{proof}

1608: 

1609: 

1610: \begin{figure}

1611: \centerline{\psfig{file=lower-bnd-rho1.eps,height=1.5in}}

1612: \caption{Proof of the $\Omega(n\log n)$ lower bound

1613: for determining if $\rho(S)=1$.}

1614: \label{fig:lower-bnd-rho1}

1615: \end{figure}

1616: 

1617: 

1618: \begin{theorem}

1619: \label{thm:small-reflexivity-2}

1620: Given a set $S$ of $n$ points in the plane, in $O(n^3\log n)$ time one

1621: can determine if $\rho(S)=2$, and, if so, produce a witness

1622: polygonalization.

1623: \end{theorem}

1624: 

1625: 

1626: \begin{proof}

1627: 

1628:   We distinguish two cases according to whether or not the two 

1629: reflex vertices belong to the same pocket.

1630: 

1631:   {\bf Two pockets:}

1632:   Assume that it is possible to polygonalize $S$ with 2 reflex vertices,

1633:   $p$ and $q$, each one in a different pocket.

1634:   Since each pocket only has one reflex vertex, its convex hull

1635:   (including the lid) is a triangle. We distinguish three

1636:   subcases: (i) the two triangles lie entirely on the same side of the

1637:   line $pq$, (ii) the two triangles lie entirely on different sides of

1638:   the line $pq$, and (iii) at least one of the triangles intersects the

1639:   line~$pq$.

1640: 

1641:   In the first subcase, since all of the interior points lie on one side of

1642:   the line $pq$, the segment $pq$ is an edge of the second convex layer

1643:   of $S$. In addition, the vertices of the pocket containing $p$

1644:   (resp. $q$) appear in angular order about $p$ (resp. $q$). This gives

1645:   a possible algorithm to detect whether such a polygonalization is possible

1646:   for $S$: For each edge $pq$ of the second convex layer of $S$,

1647:   explore all of the points of $S$ in angular order around $p$, starting on

1648:   the side of $pq$ that does not contain any interior point. Once the first

1649:   interior point is found, we have entered the possible pocket of $p$.

1650:   Check the convexity of all (but one) of the interior points found

1651:   before the next external point, which will be the endpoint of the lid

1652:   of the pocket. Proceed symmetrically from $q$ (if the angular order

1653:   around $p$ was checked clockwise, the order around $q$ must be

1654:   checked counter-clockwise). End by making sure that no interior

1655:   points are left unexplored. Since the second convex layer of $S$ has

1656:   $O(n)$ edges, and the checking for each one of them takes linear time,

1657:   this case is checked in overall $O(n^2)$ time.

1658: 

1659:   In the second subcase, the segment $pq$ may not be an edge of the

1660:   second convex layer, making the previous algorithm impossible

1661:   to apply. On the other hand though, all of the vertices of the

1662:   polygonalization lying on one side of the line $pq$ are angularly sorted

1663:   about $p$, while all of those lying on the other side are angularly sorted

1664:   about $q$. This gives an algorithm to detect whether such a

1665:   polygonalization is possible for $S$. In a first stage we 

1666:   construct a data structure as follows. For each interior point $p$ and

1667:   for each oriented line $\ell$ through $p$, our structure will store

1668:   the following information: (a) the points lying to the left of

1669:   $\ell$, angularly sorted from $p$, (b) a label indicating whether the

1670:   angular order produces a correct polygonalization to the left

1671:   of $\ell$, and (c) in the affirmative, a pointer to the first

1672:   interior point $q$ that $\ell$ will hit when rotated clockwise

1673:   around $p$. This structure can be built in $O(n^2)$ time:

1674:   for each point $p$, it can be initialized at any arbitrary line

1675:   through $p$ in $O(n)$ time, and then all lines through $p$ can be

1676:   explored by rotation around $p$. Every time that a new point

1677:   is found, it is added/eliminated at one end of the ordered list of

1678:   points to the left of $\ell$, checking for convexity of at most one

1679:   vertex (notice that the actualization of the pointer $q$ can be done

1680:   in amortized $O(n)$ time per point $p$).

1681:   The second phase of the algorithm is straightforward. Explore

1682:   all of the pairs $(p,\ell)$ having a satisfactory left

1683:   polygonalization. For each one of them, a pointer indicates its possible

1684:   complementary pair $(q,-\ell)$. Checking whether or not the two

1685:   partial polygonalizations connect properly can be done in constant

1686:   time. Hence, this case can be checked in overall $O(n^2)$ time.

1687: 

1688:   Finally, in the third subcase, none of the previous good properties

1689:   apply ($pq$ is not necessarily an edge of the second convex

1690:   layer of $S$, and the polygonalization may not be in angular order

1691:   around $p$ or $q$ on one side of the line $pq$), but there is

1692:   at least one side of the line $pq$ where the polygonalization appears

1693:   in angular order both around $p$ and $q$. This gives a possible

1694:   algorithm to detect whether such a polygonalization is possible for

1695:   $S$. Consider all pairs of interior points, $(p,q)$. Angularly check

1696:   around $p$ or $q$ whether all of the points on one side of the line

1697:   $pq$ form a convex chain. If so, keep turning around $p$ and

1698:   around $q$ separately and in opposite senses on the other

1699:   halfplane, until the two possible pockets are found. The

1700:   remaining chain between $p$ and $q$ can be checked from any of the

1701:   two points. Since we perform linear time work for each pair of

1702:   interior points, this case is checked in overall $O(n^3)$ time.

1703: 

1704:   {\bf One pocket:}

1705:   Assume that it is possible to polygonalize $S$ with 2 reflex vertices,

1706:   $p$ and $q$, both belonging to the same pocket, with lid $ab$.

1707:   The pocket is formed by three convex chains: $ap$, $pq$ and

1708:   $qb$. We will distinguish two subcases, depending on whether or not the

1709:   chains $ap$ and $qb$ are separable by a line through

1710:   the point $p$. In the first subcase, the vertices of the chain $ap$

1711:   appear in angular order around $p$ before finding any point of the

1712:   remaining chains. In the second subcase, since the chains $ap$ and $qb$

1713:   are convex, they are linearly separable by a line defined by one point

1714:   $l\in ap$ and one point $r\in qb$. Such a line must intersect the lid

1715:   of the pocket.

1716:   In addition, we will distinguish the subcase in which the convex hull of

1717:   the pocket, including the lid endpoints $a$ and $b$ is a

1718:   quadrilateral from the case in which it is a triangle. In the fist subcase,

1719:   the segment $pq$ is an edge of the second convex layer of $S$, while

1720:   in the second subcase it is not. These observations give an

1721:   algorithm to detect whether such a polygonalization is possible for~$S$.

1722: 

1723:   The subcase {\em separable-quadrilateral} can be detected in $O(n^2)$ time.

1724:   For each edge $pq$ of the second convex layer of $S$, compute the

1725:   lid $ab$ of the possible pocket, if it exists, by intersecting the

1726:   convex hull of $S$ with the prolongations of the edges of the second

1727:   layer incident in $p$ and $q$. From $p$, explore in angular order the

1728:   interior points until the first left turn is reached. Then

1729:   check whether the remaining interior points behave properly when

1730:   explored in angular order from~$q$.

1731: 

1732:   The subcase {\em separable-triangle} can be detected in $O(n^2\log n)$ time.

1733:   For each edge $ab$ of the convex hull, find the reflex vertex $p$

1734:   of the possible pocket, if it exists, by computing the supporting

1735:   lines from $a$ and $b$ to the second convex layer of $S$.

1736:   From $p$, explore in angular order the interior points until the

1737:   first left turn is reached. Then, check whether the remaining

1738:   interior points, together with $p$ and $b$, form a set $S'$ such

1739:   that $\rho(S')=1$. In this case we do not have a candidate $q$ to

1740:   help us, and the complexity of this procedure comes from

1741:   Theorem~\ref{thm:small-reflexivity-1}.

1742:   

1743:   The subcase {\em not-separable-quadrilateral} can be detected in

1744:   $O(n^3)$ time.  For each pair $(l,r)$ of points of $S$, intersect

1745:   the line $lr$ with the convex hull of $S$ to compute the two

1746:   possible lids. Let $ab$ be a possible lid for $lr$; we will call $s$

1747:   the opposite intersection point of $lr$ and the convex hull of $S$.

1748:   Compute the supporting lines from $a$ and $b$ to the second convex

1749:   layer to obtain the candidate points $p$ and $q$ associated with the

1750:   lid. Explore all of the points to the left of $lr$, together with

1751:   $s$ and $l$ (resp. $r$), in angular order around $p$. Analogously,

1752:   do this for the points to the right, around $q$. If a suitable

1753:   polygonalization is possible in each halfplane, check the connection

1754:   between them. For each of the $O(n^2)$ pairs of points $lr$, we have

1755:   performed linear time work.

1756:   

1757:   The subcase {\em not-separable-triangle} can be detected in

1758:   $O(n^3\log n)$ time.  The only difference from the previous subcase

1759:   is that, to the right of the line $lr$ we do not have a point $q$ to

1760:   be used to perform the checking in angular order. But we can check

1761:   whether the interior points to the right of $lr$, together with $b$

1762:   and $s$, form a set $S'$ such that $\rho(S')=1$. The complexity of

1763:   this procedure ($O(n\log n)$) comes from

1764:   Theorem~\ref{thm:small-reflexivity-1}.

1765: \end{proof}

1766: 

1767: \section{Inflectionality of Point Sets}

1768: \label{sec:inflection}

1769: 

1770: Consider a clockwise traversal of a polygonalization, $P$, of $S$.

1771: Then, convex (resp., reflex) vertices of $P$ correspond to {\em right}

1772: (resp., {\em left}) turns.  In computing the reflexivity of $S$ we

1773: desire a polygonalization that minimizes the number of left turns.  In

1774: this section we consider the related problem in which we want to

1775: minimize the number of {\em changes} between left-turning and

1776: right-turning during a traversal that starts (and ends) at a point

1777: interior to an edge of $P$.  We define the minimum number of such

1778: transitions between left and right turns to be the {\em

1779:   inflectionality}, $\phi(S)$, of $S$, where the minimum is taken over

1780: all polygonalizations of $S$. 

1781: %

1782: (An alternative definition is based on

1783: defining an {\em inflection edge} of $P$ to be an edge connecting a

1784: reflex vertex and a convex vertex; the inflectionality is the minimum

1785: number of inflection edges in any polygonalization of $S$.)

1786: %

1787: Clearly, $\phi(S)$ must be an even integer; it is zero if and only if $S$

1788: is in convex position.  Somewhat surprisingly, it turns out that

1789: $\phi(S)$ can only take on the values 0 or~2:

1790: 

1791: \begin{theorem}

1792: \label{th:inflex}

1793: For any finite set $S$ of $n$ points in the plane, $\phi(S)\in

1794: \{0,2\}$, with $\phi(S)=0$ precisely when $S$ is in convex position.

1795: In $O(n\log n)$ time, one can determine $\phi(S)$ as well as a

1796: polygonalization that achieves inflectionality~$\phi(S)$.

1797: \end{theorem}

1798: 

1799: \begin{proof}

1800:   If $S$ is in convex position, then trivially $\phi(S)=0$.  Thus,

1801:   assume that $S$ is not in convex position.  Then $\phi(S)\neq 0$, so

1802:   $\phi(S)\geq 2$.  We claim that $\phi(S)=2$.  For simplicity, we

1803:   assume that $S$ is in general position.

1804:   

1805:   Consider the $\ell$ nested convex polygons, $C_1,C_2,\ldots,C_\ell$,

1806:   whose boundaries constitute the $\ell$ layers (the ``onion'') of the

1807:   set $S$; these can be computed in time $O(n\log

1808:   n)$~\cite{c-clps-85}.

1809:   

1810:   We construct a ``spiral'' polygonalization of $S$ based on taking

1811:   one edge, $ab$, of $C_1$, and replacing it with a pair of

1812:   right-turning chains from $a$ to $p\in S\cap C_\ell$ and from $b$ to

1813:   $p$.  The two chains exactly cover the points of $S$ on layers

1814:   $C_2,\ldots,C_\ell$.  A constructive proof that such a

1815:   polygonalization exists is based on the following claim:

1816: 

1817: \begin{claim}

1818:   For any $1\leq m<\ell$ and any pair, $a,b\in S$, of

1819:   vertices of $C_m$, there exist two purely right-turning chains,

1820:   $\gamma_a=(a,u_1,u_2,\ldots,u_i,p)$ and

1821:   $\gamma_b=(b,v_1,v_2,\ldots,v_j,p)$, such that the points of $S$

1822:   interior to $C_m$ are precisely the set $\{u_1,u_2,\ldots,u_i, p,

1823:   v_i,v_2,\ldots,v_j\}$.

1824: \end{claim}

1825: 

1826: \begin{proof}[Proof of Claim]

1827: We prove the claim by induction on $m$.  If $m=\ell-1$,

1828: the claim follows easily, by a case analysis

1829: as illustrated in Figure~\ref{fig:easy-case}.

1830: 

1831: \begin{figure}[hbtp]

1832:    \begin{center}

1833:    \epsfxsize=.85\textwidth

1834: \ \epsfbox{easy-case.eps}

1835: \caption{Simple case in the inductive proof: $m=\ell-1$.  There are

1836: four subcases, left to right: (i)

1837: $C_\ell$ is a single point; (ii) $C_\ell$ is a line segment

1838: determined by two points of $S$; (iii) $C_\ell$ is a triangle

1839: determined by three points of $S$; or (iv) $C_\ell$ is a convex polygon

1840: whose boundary contains four or more points of $S$.}

1841: \label{fig:easy-case}

1842:    \end{center}

1843: \end{figure}

1844: 

1845: 

1846: Assume that the claim holds for $m\geq k+1$ and consider the case

1847: $m=k$.  If $C_{k+1}$ is either a single vertex or a line segment

1848: (which can only happen if $k+1=\ell$), the claim trivially follows;

1849: thus, we assume that $C_{k+1}$ has at least three vertices.  We

1850: let $u_1$ be the vertex of $C_{k+1}$ that is a {\em left tangent}

1851: vertex with respect to $a$ (meaning that $C_{k+1}$ lies in the closed

1852: halfplane to the right of the oriented line $au_1$); we let $v$ be the

1853: left tangent vertex of $C_{k+1}$ with respect to $b$.  

1854: Refer to Figure~\ref{fig:inflectionality-proof2}.

1855: If $v=u_1$, we

1856: define $v_1$ to be the vertex of $C_{k+1}$ that is the

1857: counter-clockwise neighbor of $u_1$; otherwise, we let $v_1=v$.  Let

1858: $a'$ be the counter-clockwise neighbor of $v_1$.  Let $b'$ be the

1859: counter-clockwise neighbor of $u_1$.  (Thus, $b'$ may be the same

1860: point as $v_1$.)  By the induction hypothesis, we know that there

1861: exist right-turning chains, $\gamma_{a'}$ and $\gamma_{b'}$, starting

1862: from the points $a'$ and $b'$, spiraling inwards to a point $p$

1863: interior to $C_{k+1}$.  Then we construct $\gamma_a$ to be the chain

1864: from $a$ to $u_1$, around the boundary of $C_{k+1}$ clockwise to $a'$,

1865: and then along the chain $\gamma_{a'}$.  Similarly, we construct

1866: $\gamma_b$ to be the chain from $b$ to $v_1$, around the boundary of

1867: $C_{k+1}$ clockwise to $b'$, and then along the chain $\gamma_{b'}$.

1868: \end{proof}

1869: 

1870: 

1871: 

1872: \begin{figure}[hbtp]

1873: \centerline{\hfill\psfig{file=inflectionality-proof2.eps, width=0.25\textwidth}\hfill 

1874: \psfig{file=inflectionality.eps, width=0.25\textwidth}\hfill}

1875: %   \begin{center}

1876: %   \epsfxsize=.35\textwidth

1877: %\ \epsfbox{inflectionality-proof.eps}

1878: \caption{Left: Constructing the spiraling chains $\gamma_a$ and $\gamma_b$.

1879: Right: An example of the resulting spiral polygonalization.}

1880: \label{fig:inflectionality-proof2}

1881: \end{figure}

1882: 

1883: The proof of the above claim is constructive; the

1884: required chains are readily obtained in $O(n\log n)$ time,

1885: given the convex layers.  This concludes the proof of the theorem.

1886: \end{proof}

1887: 

1888: \section{Conclusion and Future Work}

1889: 

1890: 

1891: We have introduced a new class of combinatorial and algorithmic

1892: problems related to simple polygonalizations of a planar point set.

1893: We have given lower and upper combinatorial bounds, settled

1894: the complexity status of some problem variants, and given

1895: some efficient algorithms, both exact algorithms and

1896: approximation algorithms.

1897: 

1898: There are a number of interesting open problems that our work

1899: suggests.  First, there are the four specific conjectures mentioned

1900: throughout the paper; these represent to us the most outstanding open

1901: questions raised by our work.  In addition, we mention three other areas

1902: of future study:

1903: \begin{enumerate}

1904: \item Instead of minimizing the number of reflex vertices, can we

1905:   compute a polygonalization of $S$ that minimizes the sum of the {\em

1906:     turn angles} at reflex vertices?  (The turn angle at a reflex

1907:   vertex having interior angle $\theta>\pi$ is defined to be

1908:   $\theta-\pi$.)  This question was posed to us by Ulrik Brandes.  It

1909:   may capture a notion of goodness of a polygonalization that is

1910:   useful for curve reconstruction.  The problem differs from the

1911:   angular metric TSP (\cite{ackms-amtsp-97}) in that the only turn

1912:   angles contributing to the objective function are those of reflex

1913:   vertices.

1914: \item What can be said about the generalization of the reflexivity

1915:   problem to polyhedral surfaces in three dimensions?  This may be of

1916:   particular interest in the context of surface reconstruction.

1917: \item There are a number of natural measures of ``near convexity'' for

1918:   point sets.  It would be interesting to do a systematic study of how

1919:   the various measures compare.

1920: \end{enumerate}

1921: 

1922: \subsection*{Acknowledgments}

1923: 

1924: We thank Adrian Dumitrescu for valuable input on this work, including

1925: a software tool for calculating reflexivity of point sets.  We thank

1926: Oswin Aichholzer for applying his software to search all

1927: combinatorially distinct small point sets.  This collaborative

1928: research between the Universitat Polit\`ecnica de Catalunya and Stony

1929: Brook University was made possible by a grant from the Joint

1930: Commission USA-Spain for Scientific and Technological Cooperation

1931: Project 98191.  E. Arkin acknowledges additional support from the

1932: National Science Foundation (CCR-9732221, CCR-0098172).  S.\ Fekete

1933: acknowledges travel support by the Hermann-Minkowski-Minerva Center

1934: for Geometry at Tel Aviv University.  F.\ Hurtado, M.\ Noy, and V.\ 

1935: Sacrist\'an acknowledge support from CUR Gen. Cat. 1999SGR00356, and

1936: Proyecto DGES-MEC PB98-0933.  J.~Mitchell acknowledges support from

1937: NSF (CCR-9732221, CCR-0098172) and NASA Ames Research Center

1938: (NAG2-1325).

1939: 

1940: 

1941: \nocite{ackms-amtsp-97,u-ppsdc-99,u-opcp-96,u-ppscp-97}

1942: \nocite{es-sepeg-60,es-cpg-35,deo-secp-90,gnt-lbncf-95}

1943: 

1944: \bibliographystyle{abbrv}

1945: 

1946: \begin{thebibliography}{10}

1947: 

1948: \bibitem{a-rsoas-92}

1949: P.~K. Agarwal.

1950: \newblock Ray shooting and other applications of spanning trees with low

1951:   stabbing number.

1952: \newblock {\em SIAM J. Comput.}, 21:540--570, 1992.

1953: 

1954: \bibitem{afg-pdecm-02}

1955: P.~K. Agarwal, E. Flato, and D. Halperin.

1956: Polygon decomposition for efficient

1957: construction of Minkowski sums.

1958: \newblock {\em Comput. Geom. Theory Appl.}, 21:39--61, 2002.

1959: 

1960: \bibitem{as-sagp-98}

1961: P.~K. Agarwal and S.~Suri.

1962: \newblock Surface approximation and geometric partitions.

1963: \newblock {\em SIAM J. Comput.}, 27:1016--1035, 1998.

1964: 

1965: \bibitem{ackms-amtsp-97}

1966: A.~Aggarwal, D.~Coppersmith, S.~Khanna, R.~Motwani, and B.~Schieber.

1967: \newblock The angular-metric traveling salesman problem.

1968: \newblock In {\em Proceedings of the Eighth Annual ACM-SIAM Symposium on

1969:   Discrete Algorithms}, pages 221--229, Jan. 1997.

1970: 

1971: \bibitem{aak-eotsp-01}

1972: O.~Aichholzer, F.~Aurenhammer, and H. Krasser.

1973: \newblock Enumerating order types for small point sets with applications.

1974: \newblock In {\em Proc. 17th Annu. ACM Sympos. Comput. Geom.}, 2001, pp.~11--18.

1975: 

1976: \bibitem{ak-psotd-01}

1977: O. Aichholzer and H. Krasser.

1978: \newblock The point set order type data base: a collection of applications

1979: and results. 

1980: \newblock In {\em Proc. 13th Canad. Conf. Comput.

1981: Geom.}, Waterloo, Canada, 2001, pp.~17--20.

1982: 

1983: \bibitem{AK}

1984: O.~Aichholzer and H.~Krasser.

1985: \newblock Personal communication, 2001.

1986: 

1987: \bibitem{abe-cbscc-98}

1988: N.~Amenta, M.~Bern, and D.~Eppstein.

1989: \newblock The crust and the $\beta$-skeleton: Combinatorial curve

1990:   reconstruction.

1991: \newblock {\em Graphical Models and Image Processing}, 60:125--135, 1998.

1992: 

1993: \bibitem{ah-hgrp-96}

1994: T.~Auer and M.~Held.

1995: \newblock Heuristics for the generation of random polygons.

1996: \newblock In {\em Proc. 8th Canad. Conf. Comput. Geom.}, pages 38--43, 1996.

1997: 

1998: \bibitem{c-cgc-79}

1999: B.~Chazelle.

2000: \newblock {\em Computational geometry and convexity}.

2001: \newblock Ph.{D}. thesis, Dept. Comput. Sci., Yale Univ., New Haven, CT, 1979.

2002: \newblock Carnegie-Mellon Univ. Report CS-80-150.

2003: 

2004: \bibitem{c-clps-85}

2005: B.~Chazelle.

2006: \newblock On the convex layers of a planar set.

2007: \newblock {\em IEEE Trans. Inform. Theory}, IT-31(4):509--517, July 1985.

2008: 

2009: \bibitem{dk-spacr-99}

2010: T.~K. Dey and P.~Kumar.

2011: \newblock A simple provable algorithm for curve reconstruction.

2012: \newblock In {\em Proc. 10th ACM-SIAM Sympos. Discrete Algorithms}, pages

2013:   893--894, Jan. 1999.

2014: 

2015: \bibitem{dmr-crcdgr-99}

2016: T.~K. Dey, K.~Mehlhorn, and E.~A. Ramos.

2017: \newblock Curve reconstruction: {C}onnecting dots with good reason.

2018: \newblock In {\em Proc. 15th Annu. ACM Sympos. Comput. Geom.}, pages 197--206,

2019:   1999.

2020: 

2021: \bibitem{deo-secp-90}

2022: D.~P. Dobkin, H.~Edelsbrunner, and M.~H. Overmars.

2023: \newblock Searching for empty convex polygons.

2024: \newblock {\em Algorithmica}, 5:561--571, 1990.

2025: 

2026: \bibitem{ers-ccsno-90}

2027: H.~Edelsbrunner, A.~D. Robison, and X.~Shen.

2028: \newblock Covering convex sets with non-overlapping polygons.

2029: \newblock {\em Discrete Math.}, 81:153--164, 1990.

2030: 

2031: \bibitem{es-cpg-35}

2032: P.~Erd{\H o}s and G.~Szekeres.

2033: \newblock A combinatorial problem in geometry.

2034: \newblock {\em Compositio Math.}, 2:463--470, 1935.

2035: 

2036: \bibitem{es-sepeg-60}

2037: P.~Erd{\H o}s and G.~Szekeres.

2038: \newblock On some extremeum problem in geometry.

2039: \newblock {\em Ann. Univ. Sci. Budapest}, 3-4:53--62, 1960.

2040: 

2041: \bibitem{e-webpage}

2042: J. Erickson.

2043: Generating random simple polygons.\\

2044: {\tt http://compgeom.cs.uiuc.edu/\~{}jeffe/open/randompoly.html}

2045: 

2046: \bibitem{fw-artp-97}

2047: S.~P. Fekete and G.~J. Woeginger.

2048: \newblock Angle-restricted tours in the plane.

2049: \newblock {\em Comput. Geom. Theory Appl.}, 8(4):195--218, 1997.

2050: 

2051: \bibitem{gnt-lbncf-95}

2052: A.~Garc{\'\i}a, M.~Noy, and J.~Tejel.

2053: \newblock Lower bounds for the number of crossing-free subgraphs of {$K_n$}.

2054: \newblock In {\em Proc. 7th Canad. Conf. Comput. Geom.}, pages 97--102, 1995.

2055: 

2056: \bibitem{hs-asdch-92}

2057: J.~Hershberger and S.~Suri.

2058: \newblock Applications of a semi-dynamic convex hull algorithm.

2059: \newblock {\em BIT}, 32:249--267, 1992.

2060: 

2061: \bibitem{hs-parss-95}

2062: J.~Hershberger and S.~Suri.

2063: \newblock A pedestrian approach to ray shooting: {Shoot} a ray, take a walk.

2064: \newblock {\em J. Algorithms}, 18:403--431, 1995.

2065: 

2066: \bibitem{hm-ftprs-85}

2067: S.~Hertel and K.~Mehlhorn.

2068: \newblock Fast triangulation of the plane with respect to simple polygons.

2069: \newblock {\em Inform. Control}, 64:52--76, 1985.

2070: 

2071: \bibitem{h-snec7g-83}

2072: J.~Horton.

2073: \newblock Sets with no empty convex 7-gons.

2074: \newblock {\em Canad. Math. Bull.}, 26:482--484, 1983.

2075: 

2076: \bibitem{HRU}

2077: K.~Hosono, D.~Rappaport, and M.~Urabe.

2078: \newblock On convex decompositions of points.

2079: \newblock In {\em Proc. Japanese Conf. on Discr. Comp. Geom. (2000)}, volume

2080:   2098 of {\em Lecture Notes Comput. Sci.}, pages 149--155. Springer-Verlag,

2081:   2001.

2082: 

2083: \bibitem{hu-ndcqp-01}

2084: K.~Hosono and M.~Urabe.

2085: \newblock On the number of disjoint convex quadrilaterals for a plannar point

2086:   set.

2087: \newblock {\em Comp. Geom. Theory Appl.}, 20:97--104, 2001.

2088: 

2089: \bibitem{hn-tvgrv-96}

2090: F.~Hurtado and M.~Noy.

2091: \newblock Triangulations, visibility graph and reflex vertices of a simple

2092:   polygon.

2093: \newblock {\em Comput. Geom. Theory Appl.}, 6:355--369, 1996.

2094: 

2095: \bibitem{k-pd-00}

2096: J.~M. Keil.

2097: \newblock Polygon decomposition.

2098: \newblock In J.-R. Sack and J.~Urrutia, editors, {\em Handbook of Computational

2099:   Geometry}, pages 491--518. Elsevier Science Publishers B.V. North-Holland,

2100:   Amsterdam, 2000.

2101: 

2102: \bibitem{l-pftu-82}

2103: D.~Lichtenstein.

2104: \newblock Planar formulae and their uses.

2105: \newblock {\em SIAM J. Comput.}, 11(2):329--343, 1982.

2106: 

2107: \bibitem{m-aagsp-93}

2108: J.~S.~B. Mitchell.

2109: \newblock Approximation algorithms for geometric separation problems.

2110: \newblock Technical report, Department of Applied Mathematics, SUNY Stony

2111:   Brook, NY, July 1993.

2112: 

2113: \bibitem{mrsw-ccppp-95}

2114: J.~S.~B. Mitchell, G.~Rote, G.~Sundaram, and G.~Woeginger.

2115: \newblock Counting convex polygons in planar point sets.

2116: \newblock {\em Inform. Process. Lett.}, 56:191--194, 1995.

2117: 

2118: \bibitem{o-cgc-98}

2119: J.~O'Rourke.

2120: \newblock {\em Computational Geometry in {C}}.

2121: \newblock Cambridge University Press, 2nd edition, 1998.

2122: 

2123: \bibitem{dcg19}

2124: J.~Pach, editor.

2125: \newblock {\em Special Issue Dedicated to Paul {Erd\"{o}s}}, volume~19 of {\em

2126:   Discrete Comput. Geom.}

2127: \newblock 1998.

2128: 

2129: \bibitem{RU}

2130: E.~Rivera-Campo and J.~Urrutia.

2131: \newblock Personal communication, 2001.

2132: 

2133: \bibitem{u-opcp-96}

2134: M.~Urabe.

2135: \newblock On a partition into convex polygons.

2136: \newblock {\em Discrete Appl. Math.}, 64:179--191, 1996.

2137: 

2138: \bibitem{u-ppscp-97}

2139: M.~Urabe.

2140: \newblock On a partition of point sets into convex polygons.

2141: \newblock In {\em Proc. 9th Canad. Conf. Comput. Geom.}, pages 21--24, 1997.

2142: 

2143: \bibitem{u-ppsdc-99}

2144: M.~Urabe.

2145: \newblock Partitioning point sets into disjoint convex polytopes.

2146: \newblock {\em Comput. Geom. Theory Appl.}, 13:173--178, 1999.

2147: 

2148: \bibitem{w-ggssn-93}

2149: E.~Welzl.

2150: \newblock Geometric graphs with small stabbing numbers: {Combinatorics} and

2151:   applications.

2152: \newblock In {\em Proc. 9th Internat. Conf. Fund. Comput. Theory}, Lecture

2153:   Notes Comput. Sci., Springer-Verlag, 1993.

2154: 

2155: \bibitem{zssm-grpgv-96}

2156: C.~Zhu, G.~Sundaram, J.~Snoeyink, and J.~S.~B. Mitchell.

2157: \newblock Generating random polygons with given vertices.

2158: \newblock {\em Comput. Geom. Theory Appl.}, 6:277--290, 1996.

2159: 

2160: \end{thebibliography}

2161: 

2162: 

2163: 

2164: \end{document}

2165: 

2166: