1: % Sept. 30, 2007: journal version (last edited by Adrian)

2: 

3: \documentclass[11pt]{article}

4: \usepackage{fullpage}

5: \usepackage{graphicx,amssymb,amsmath,theorem}

6: \usepackage{epsfig}

7: \usepackage{epsf}

8: \usepackage{times}

9: \usepackage{amsfonts}

10: \usepackage{latexsym}

11: \usepackage{bbm}     % blackboard characters (e.g. set of real numbers)

12: 

13: \newtheorem{theorem}{Theorem}

14: %[section]

15: \newtheorem{lemma}{Lemma}

16: \newtheorem{proposition}{Proposition}

17: \newtheorem{corollary}{Corollary}

18: \newtheorem{observation}{Observation}

19: \newtheorem{claim}{Claim}

20: \newtheorem{fact}{Fact}

21: \newtheorem{conjecture}{Conjecture}

22: \def \proof {\par \noindent {\bf Proof.}\hskip 5pt}

23: \def \endproof {\hfill $\Box$ \smallskip}%\linebreak \medskip}

24: 

25: \setlength{\oddsidemargin}{0in} \setlength{\evensidemargin}{0in}

26: \setlength{\textheight}{9in}

27: \setlength{\textwidth}{6.5in}

28: %\setlength{\topmargin}{-0.25in}

29: \setlength{\topmargin}{0in}

30: \setlength{\headheight}{0in}

31: \setlength{\headsep}{0in}

32: 

33: \long\def\ignore#1{}

34: \def\cut{\ignore}

35: \newcommand{\later}[1]{{}}

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

37: 

38: \newenvironment{proofof}[1]

39:       {\medskip\noindent{\bf Proof of #1:}\hspace{1mm}}

40:       {\hfill$\Box$\medskip}

41: 

42: 

43: \newcommand{\NN}{\mathbb{N}} %  set of natural numbers

44: \newcommand{\ZZ}{\mathbb{Z}} %  set of integer numbers

45: \newcommand{\RR}{\mathbb{R}} %  set of real numbers

46: \newcommand{\QQ}{\mathbb{Q}} %  set of rational numbers

47: \newcommand{\SH}{\mathbb{S}} %  set of unit vectors

48: 

49: \def\reals{{\mathbb R}}

50: 

51: \def\A{{\cal A}}

52: \def\B{{\cal B}}

53: \def\C{{\cal C}}

54: \def\L{{L}}

55: \def\N{{\cal N}}

56: \def\T{{T}}

57: \def\U{{\cal U}}

58: \def\bd{{\partial}}

59: \def\polylog{{\rm polylog}}

60: \def\eps{\varepsilon}

61: \def\etal{{\it et~al.}\,}

62: \def\uu{{\bf u}}

63: \def\vv{{\bf v}}

64: \def\xx{{\bf x}}

65: \def\kk{{\bf k}}

66: 

67: 

68: \def\marrow{{\marginpar[\hfill$\longrightarrow$]{$\longleftarrow$}}}

69: \def\adrian#1{{\sc Adrian says: }{\marrow\sf #1}}

70: \def\micha#1{{\sc Micha says: }{\marrow\sf #1}}

71: \def\csaba#1{{\sc Csaba says: }{\marrow\sf #1}}

72: \newcommand\comm[1]{\typeout{Used \string\comm...}\marrow

73: \quad{\bf [[#1]]}\quad}

74: 

75: 

76: \title{Extremal problems on triangle areas in two and three dimensions}

77: 

78: \author{Adrian Dumitrescu\thanks{%

79: Department of Computer Science,

80: University of Wisconsin-Milwaukee, USA,

81: email: {\tt ad@cs.uwm.edu.}

82: Research partially supported by NSF CAREER grant CCF-0444188.}

83: \and

84: Micha Sharir\thanks{%

85: School of Computer Science, Tel Aviv University,

86: Tel Aviv, Israel, and Courant Institute, New York University, New

87: York, NY, USA,

88: email: {\tt michas@tau.ac.il.}

89: Research partially supported by NSF Grant CCR-05-14079,

90: by a grant from the U.S.-Israeli Binational Science Foundation,

91: by grant 155/05 from the Israel Science Fund, Israeli Academy of Sciences,

92: and by the Hermann Minkowski--MINERVA Center for Geometry at Tel Aviv

93: University.}

94: \and

95: Csaba D. T\'oth\thanks{%

96: Department of Mathematics,

97: Massachusetts Institute of Technology, Cambridge, MA, USA,

98: email: {\tt toth@math.mit.edu.}}}

99: 

100: \begin{document}

101: 

102: \maketitle

103: \thispagestyle{empty}

104: 

105: 

106: \begin{abstract}

107: 

108: The study of extremal problems on triangle areas was initiated in

109: a series of papers by Erd\H{o}s and Purdy in the early 1970s.

110: % Erd\H{o}s-type extremal discrete geometry problems have attracted

111: % continued attention in the combinatorics and computational geometry

112: % communities, due to their relevance for computational problems (in

113: % range counting, pattern matching, motion planning, and others) and the

114: % rich techniques developed for improving extremal bounds (such as the

115: % $\eps$-net theory, the Crossing Lemma, quasi-planar graphs,

116: % etc.), which proved to be instrumental in many areas of computational

117: % geometry.

118: In this paper we present new results on such problems, concerning the

119: number of triangles of the same area that are spanned by finite point

120: sets in the plane and in 3-space, and the number of distinct areas 

121: determined by the triangles. 

122: 

123: In the plane, our main result is an $O(n^{44/19}) =O(n^{2.3158})$

124: upper bound on the number of unit-area triangles spanned by $n$

125: points, which is the first breakthrough improving the classical bound of

126: $O(n^{7/3})$ from 1992. We also make progress in a number of important

127: special cases: We show that (i) For points in convex position, there

128: exist $n$-element point sets that span $\Omega(n\log n)$ triangles of

129: unit area. (ii) The number of triangles of minimum (nonzero) area

130: determined by $n$ points is at most $\frac{2}{3}(n^2-n)$; there exist

131: $n$-element point sets (for arbitrarily large $n$)

132: that span $(6/\pi^2-o(1))n^2$ minimum-area

133: triangles.  (iii) The number of acute triangles of minimum area

134: determined by $n$ points is $O(n)$; this is asymptotically tight. (iv)

135: For $n$ points in convex position, the number of triangles of minimum

136: area is $O(n)$; this is asymptotically tight. (v) If no three points

137: are allowed to be collinear, there are $n$-element point sets that

138: span $\Omega(n\log n)$ minimum-area triangles

139: (in contrast to (ii), where collinearities are allowed and

140: a quadratic lower bound holds).

141: 

142: In 3-space we prove an $O(n^{17/7}\beta(n))= O(n^{2.4286})$ upper

143: bound on the number of unit-area triangles spanned by $n$ points,

144: where $\beta(n)$ is an extremely slowly growing function related to the

145: inverse Ackermann function. The best previous bound, $O(n^{8/3})$, is

146: an old result of Erd\H{o}s and Purdy from 1971. We further show,

147: for point sets in 3-space:

148: (i) The number of minimum nonzero area triangles is at most $n^2+O(n)$,

149: and this is worst-case optimal, up to a constant factor.

150: (ii) There are $n$-element point sets that span

151: $\Omega(n^{4/3})$ triangles of maximum area, all incident to a common

152: point. In any $n$-element point set, the maximum number of

153: maximum-area triangles incident to a common point is

154: $O(n^{4/3+\eps})$, for any $\eps>0$.

155: (iii) Every set of $n$ points, not all on a line, determines at least

156: $\Omega(n^{2/3}/\beta(n))$ triangles of distinct areas, which share a

157: common side.

158: \end{abstract}

159: 

160: 

161: \newpage

162: \setcounter{page}{1}

163: 

164: \section{Introduction}

165: \label{sec:intro}

166: 

167: Given $n$ points in the plane, consider the following equivalence

168: relation defined on the set of (nondegenerate) triangles  spanned by

169: the points: two triangles are {\em equivalent} if they have the

170: same area.

171: Extremal problems typically ask for the maximum cardinality of an

172: equivalence class, and for the minimum number of distinct equivalence

173: classes, in a variety of cases. A classical example is when we call

174: two segments spanned by the given points equivalent if they

175: have the same length. Bounding the maximum size of an equivalence class

176: is the famous {\em repeated distances} problem

177: \cite{bmp-05,e-46,sst-84,s-97}, and bounding the minimum number of distinct

178: classes is the equally famous

179: {\em distinct distances} problem \cite{bmp-05,e-46,kt-04,st-01,s-97,t-03}.

180: In this paper, we make progress on several old extremal problems on

181: triangle areas in two and in three dimensions. We also study some

182: new and interesting variants never considered before. Our proof

183: techniques draw from a broad range of combinatorial tools such as

184: the Szemer\'edi-Trotter theorem on point-line incidences~\cite{st-83},

185: the Crossing Lemma~\cite{acns-82,l-84}, incidences between curves and

186: points and tangencies between curves and lines, extremal graph

187: theory~\cite{kst-54}, quasi-planar graphs~\cite{aapps-97},

188: Minkowski-type constructions, repeated distances on the

189: sphere~\cite{pa-95}, the partition technique of Clarkson~\etal~\cite{ceg-90},

190: various charging schemes, etc.

191: 

192: In 1967, A.~Oppenheim (see \cite{ep-95}) asked the following question:

193: Given $n$ points in the plane and $A>0$, how many triangles spanned

194: by the points can have area $A$? By applying an affine transformation,

195: one may assume $A=1$ and count the triangles of {\em unit} area.

196: Erd\H{o}s and Purdy~\cite{ep-71} showed that a $\sqrt{\log n}\times

197: (n/\sqrt{\log n})$ section of the integer lattice determines

198: $\Omega(n^2 \log\log{n})$ triangles of the same area. They also showed

199: that the maximum number of such triangles is at most $O(n^{5/2})$. In

200: 1992, Pach and Sharir~\cite{ps-92} improved the exponent and obtained

201: an $O(n^{7/3})$ upper bound using the Szemer\'edi-Trotter

202: theorem~\cite{st-83} on the number of point-line incidences. We

203: further improve the upper bound by estimating the number of incidences

204: between the points and a 4-parameter family of quadratic curves. We

205: show that $n$ points in the plane determine at most

206: $O(n^{44/19})=O(n^{2.3158})$ unit-area triangles. We also consider the

207: case of points in convex position, for which we construct $n$-element

208: point sets that span $\Omega(n\log n)$ triangles of unit area.

209: 

210: Bra\ss, Rote, and Swanepoel~\cite{brs-01} showed that $n$ points in

211: the plane determine at most $O(n^2)$ minimum-area triangles, and they

212: pointed out that this bound is asymptotically tight. We introduce a

213: simple charging scheme to first bring the upper bound down to

214: $n^2-n$ and then further to $\frac{2}{3}(n^2-n)$. Our charging scheme

215: is also instrumental in showing that a $\sqrt{n} \times \sqrt{n}$

216: section of the integer lattice spans $(6/\pi^2-o(1))n^2$ triangles of

217: minimum area.  In the lower bound constructions, there are many collinear

218: triples and most of the minimum-area triangles are obtuse. We show

219: that there are at most $O(n)$ {\em acute} triangles of minimum

220: (nonzero) area, for any $n$-element point set.

221: Also, we show that $n$ points in (strictly) 

222: convex position determine at most

223: $O(n)$ minimum-area triangles---these bounds are best possible apart from

224: the constant factors. If no three points are allowed to be collinear,

225: we construct $n$-element point sets that span $\Omega(n\log{n})$

226: triangles of minimum area.

227: 

228: Next we address analogous questions for triangles in 3-space.

229: The number of triangles with some extremal property

230: might go up (significantly) when one moves up one dimension.

231: For instance, Bra\ss, Rote, and Swanepoel~\cite{brs-01} have shown that

232: the number of maximum area triangles in the plane is at most $n$

233: (which is tight). In 3-space we show that this number is at

234: least $\Omega(n^{4/3})$ in the worst case.

235: In contrast, for minimum-area triangles, we

236: prove that the quadratic upper bound from the planar case remains in

237: effect for 3-space, with a different constant of proportionality.

238: 

239: As mentioned earlier, Erd\H{o}s and Purdy~\cite{ep-71} showed that a

240: suitable $n$-element section of the integer lattice determines

241: $\Omega(n^2 \log\log{n})$ triangles of the same area. Clearly, this

242: bound is also valid in 3-space. In the same paper, via a forbidden

243: graph argument applied to the incidence graph between points and

244: cylinders whose axes pass through the origin, Erd\H{o}s and Purdy

245: deduced an $O(n^{5/3})$ upper bound on the number of unit-area triangles

246: incident to a common point, and thereby an $O(n^{8/3})$ upper bound on

247: the number of unit-area triangles determined by $n$ points in 3-space.

248: Here, applying a careful (and somewhat involved) analysis of the

249: structure of point-cylinder incidences in $\RR^3$, we

250: prove a new upper bound of $O(n^{17/7}\beta(n))=O(n^{2.4286})$,

251: for $\beta(n) = \exp(\alpha(n)^{O(1)})$,

252: where $\alpha(n)$ is the

253: extremely slowly growing inverse Ackermann function.

254: 

255: It is conjectured~\cite{bmp-05,bp-79,eps-82} that $n$ points in

256: $\RR^3$, not all on a line, determine at least $\lfloor(n-1)/2\rfloor$

257: distinct triangle areas. This bound has recently been established in

258: the plane \cite{p-07}, but the question is still wide open in $\RR^3$. 

259: It is attained by $n$ equally spaced points distributed evenly on two

260: parallel lines (which is in fact a planar construction).

261: We obtain a first result on this question and show that

262: $n$ points in $\RR^3$, not all on a line, determine at least

263: $n^{2/3}\exp(-\alpha(n)^{O(1)})=\Omega(n^{.666})$ triangles of

264: distinct areas.  Moreover, all these triangles share a common side.

265: 

266: 

267: \section{Unit-area triangles in the plane}

268: \label{sec:unit2}

269: 

270: \noindent {\bf The general case.} We establish a new upper bound on the

271: maximum number of unit-area triangles determined by $n$ points the

272: plane. 

273: 

274: \begin{theorem}\label{thm:unit2}

275: The number of unit-area triangles spanned by $n$ points in the plane

276: is $O(n^{2+6/19})= O(n^{2.3158})$.

277: \end{theorem}

278: %

279: \begin{proof}

280: Let $S$ be a set of $n$ points in the plane. Consider a triangle

281: $\Delta{abc}$ spanned by $S$. We call the three lines containing the

282: three sides of $\Delta{abc}$, {\em base lines} of $\Delta$, and the

283: three lines parallel to the base lines and incident to the third

284: vertex, {\em top lines} of $\Delta$.

285: 

286: \cut{For any integer $k\in \NN$, let $L_k$ denote the set of $k$-rich

287:   lines (that is, the lines containing at least $k$ points of $S$). By

288:   the Szemer\'edi-Trotter theorem~\cite{st-83}, we have

289:   $|L_k|=O(n^2/k^3+n/k)$. In particular, $|L_k|=O(n^2/k^3)$ for $k\leq

290:   \sqrt{n}$ and $|L_k| = O(n/k)$ for $k\geq \sqrt{n}$.}

291: %

292: For a parameter $k$, $1\leq k\leq \sqrt{n}$, to be optimized

293: later, we partition the set of unit-area triangles as follows.

294: %\begin{itemize}

295: 

296: \smallskip

297: \noindent $\bullet$ $U_1$ denotes the set of unit-area triangles where

298: one of the top lines is incident to fewer than $k$ points of $S$.

299: 

300: \smallskip

301: \noindent $\bullet$ $U_2$ denotes the set of unit-area triangles where

302: all three top lines are {\em $k$-rich} (i.e., each contains at

303: least $k$\\

304: \indent points of $S$).

305: 

306: \smallskip

307: %\end{itemize}

308: \noindent We derive different upper bounds for each of these

309: types of unit-area triangles.

310: 

311: \paragraph{Bound for  $|U_1|$.}

312: For any two distinct points, $a,b\in \RR^2$, let $\ell_{ab}$ denote

313: the line through $a$ and $b$. The points $c$ for which the triangle

314: $\Delta{abc}$ has unit area lie on two lines

315: $\ell^-_{ab},\ell^+_{ab}$ parallel to $\ell_{ab}$ at distances

316: $2/|ab|$ on either side of $\ell_{ab}$. The ${n\choose 2}$ segments

317: determined by $S$ generate at most $2{n\choose 2}$ such lines

318: (counted with multiplicity).

319: If $\Delta abc \in U_1$ and its top line incident to the fewest

320: points of $S$ is $\ell_{ab}'\in \{\ell^-_{ab},\ell^+_{ab}\}$, then

321: $\ell_{ab}'$ is incident to at most $k$ points, so the segment $ab$ is

322: the base of at most $k$ triangles $\Delta{abc}\in U_1$ (with

323: $c\in\ell_{ab}'$).  This gives the upper bound

324: %

325: $$|U_1|\leq  2{n\choose 2} \cdot k = O(n^2 k).$$

326: 

327: \paragraph{Bound for $|U_2|$.}

328: Let $L$ be the set of $k$-rich lines, and let $m=|L|$. By the

329: Szemer\'edi-Trotter theorem~\cite{st-83}, we have $m=O(n^2/k^3)$ for

330: any $k\leq \sqrt{n}$. Furthermore, the cardinality of the set

331: $I(S,L)$ of point-line incidences between $S$ and $L$ is

332: $|I(S,L)|=O(n^2/k^2)$.

333: 

334: \old{

335: If $\Delta{abc}\in U_2$, then its top lines

336: $\ell_{ab}',\ell_{bc}',\ell_{ac}'\in L_k$ form a triangle of area 4,

337: in which the points $a,b,c$ are the midpoints of the three sides. We

338: give an upper bound on the number of 4-tuples

339: $(\ell_1,\ell_2,\ell_3,a)\in L_k\times S$ where the lines

340: $\ell_1,\ell_2,\ell_3$ form a triangle of area 4 and $a$ is the

341: midpoint of a side. Since every triangle in $U_2$ corresponds to

342: three 4-tuples of this form, we obtain an upper bound on $|U_2$.

343: } % old

344: 

345: For any pair of nonparallel lines $\ell_1,\ell_2\in L$, let

346: $\gamma(\ell_1,\ell_2)$ denote the locus of points $p\in \RR^2$,

347: $p\not\in\ell_1\cup \ell_2$, such that the parallelogram that has a

348: vertex at $p$ and two sides along $\ell_1$ and $\ell_2$,

349: respectively, has area 2. The set $\gamma(\ell_1,\ell_2)$ consists

350: of two hyperbolas with $\ell_1$ and $\ell_2$ as asymptotes.

351: See Figure~\ref{fig:hyper}. For instance, if

352: $\ell_1: y=0$ and $\ell_2: y= ax$, then $\gamma(\ell_1,\ell_2) =

353: \{(x,y)\in \RR^2:xy=y^2/a+2\}\cup \{(x,y)\in \RR^2:xy=y^2/a-2\}$.

354: Any two nonparallel lines uniquely determine two such hyperbolas.

355: Let $\Gamma$ denote the set of these hyperbolas. Note that

356: $|\Gamma|=O(m^2)$. The family of such hyperbolas for all pairs of

357: nonparallel lines form a 4-parameter family of quadratic curves

358: (where the parameters are the coefficients of the defining lines).

359: 

360: For any triangle $\Delta{abc}\in U_2$, any pair of its top lines,

361: say, $\ell_{ab}'$ and $\ell_{ac}'$, determine a hyperbola passing

362: through $a$, which is incident to the third top line $\ell_{bc}'$;

363: furthermore $\ell_{bc}'$ is tangent\footnote{%

364:   For a quick proof, let $\uu$ (resp., $\vv$) be a unit vector

365:   along $\ell'_{ac}$ (resp., $\ell'_{ab}$). The point $a$ can be

366:   parametrized as $\xx= t\uu + \frac{\kappa}{t}\vv$, where

367:   $\kappa=2/\sin\theta$, and $\theta$ is the angle between

368:   $\ell'_{ac}$ and $\ell'_{ab}$. Hence the tangent to the

369:   hyperbola at $a$ is given by

370:   $\dot{\xx} = \uu-\frac{\kappa}{t^2}\vv \,\|\, t\uu-\frac{\kappa}{t}\vv =

371:   \vec{cb}$.}

372: to the hyperbola at $a$. See Figure~\ref{fig:hyper}. Any hyperbola in

373: this 4-parameter family is uniquely determined by two incident

374: points and the two respective tangent lines at those points.

375: %\micha{Has anyone ever checked this carefully?}

376: %\csaba{There are two trivial quadratic curve incident to points $p$ and $q$,

377: %and tangent to lines $\ell_p$ and $\ell_q$ at those points:

378: %The union of lines $\ell_p\cup\ell_q$ (a degenerate hyperbola),

379: %and the line segment $pq$ (a degenerate ellipse). Any other quadratic curve

380: %is the linear combination of these two (since the incident points and

381: %the tangent lines impose linear constraints on the coefficients of

382: %the general quadratic curve $ax^2+by^2+cxy+dx+ey=1$. In this family

383: %of linear combinations, just one quadratic curve corresponds to

384: %unit triangles.

385: 

386: \begin{figure}[htbp]

387:   \begin{center}

388: \epsfig{file=hyper.eps,width=3.9cm,clip=}

389:  \caption{\small One of the hyperbolas defined by the triangle

390:  $\Delta abc$.

391: \label{fig:hyper}}

392: \end{center}

393:  \end{figure}

394: 

395: 

396: \old{If three lines $\ell_1,\ell_2,\ell_3\in L_k$ form a triangle

397: $\Delta$ of area 4, then $\ell_3$ is tangent to

398: $\gamma(\ell_1,\ell_3)$, and the intersection point

399: $\gamma(\ell_1,\ell_2)\cap \ell_3$ is the midpoint of the side of

400: $\Delta$ along $\ell_3$. Any hyperbola in the 4-parameter family

401: $\Gamma$ is uniquely determined by two incident points and two

402: tangent lines at those points.

403: } % old

404: 

405: We define a topological graph $G$ as follows. For each point

406: $p\in S$, which is incident to $d_p$ lines of $L$, we create $2d_p$

407: vertices in $G$, as follows (refer to Figure~\ref{fig:graph}).

408: Draw a circle

409: $C_\eps(p)$ centered at $p$ with a sufficiently small radius

410: $\eps>0$, and place a vertex at every intersection point of the

411: circle $C_\eps(p)$ with the $d_p$ lines incident to $p$. The number

412: of vertices is $v_G=2|I(S,L)|=O(n^2/k^2)$. Next, we define the edges

413: of $G$. For each connected branch $\gamma$ of every hyperbola in

414: $\Gamma$, consider the set $S(\gamma)$ of points $p\in S$ that are

415: (i) incident to $\gamma$ and (ii) some line of $L$ is tangent to

416: $\gamma$ at $p$. For any two consecutive points $p,q\in S(\gamma)$,

417: draw an edge along $\gamma$ between the two vertices of $G$ that

418: (i) correspond to the incidences $(p,\ell_p)$ and $(q,\ell_q)$, where

419: $\ell_p$ and $\ell_q$ are the tangents of $\gamma$ at $p$ and $q$,

420: respectively, and (ii) are closest to each other along $\gamma$.

421: Specifically, the edge follows $\gamma$ between the

422: circles $C_{2\eps}(p)$ and $C_{2\eps}(q)$ and follows straight line

423: segments in the interiors of those circles.

424: %\micha{I dont see why you need this double circle rule. Isnt it enough

425: %to use just the inner circles $C_\eps(p)$? What are you afraid of?}

426: %\csaba{With the inner circle we can give exactly the location

427: %of the new vertices. If we place these vertices on the outer circle,

428: %then the segments in the figure would be too close to each other.}

429: Choose $\eps>0$ sufficiently small so that the circles $C_{2\eps}(p)$

430: have disjoint interiors and

431: the portions of the hyperbolas in the interiors of the circles

432: $C_{2\eps}(p)$, for every $p\in S$, meet at $p$ only. This

433: guarantees that the edges of $G$ cross only at intersection points

434: of the hyperbolas. The graph $G$ is {\em simple} because two points and

435: two tangent lines uniquely determine a hyperbola in $\Gamma$. The

436: number of edges is at least $3|U_2|-2m^2$, since every triangle in

437: $U_2$ corresponds to three point-hyperbola incidences in $I(S,\Gamma)$

438: (satisfying the additional condition of tangency with the respective

439: top lines);

440: and along each of the $2m^2$ hyperbola branches, each of its

441: incidences with the points of $S$ (of the special kind under

442: consideration), except for one, contributes one edge to $G$.

443: %

444: \begin{figure}[htbp]

445:   \begin{center}

446: \epsfig{file=graph.eps,width=14cm,clip=}

447:  \caption{\small On the left: a point $p\in S$ incident to three lines

448: of $L$ (dashed) and 8 hyperbolas, each tangent to one of those

449: lines. On the right: the 6 vertices of $G$ corresponding to the 3

450: point-line incidences at $p$, and the drawings of the edges along the

451: hyperbolas.\label{fig:graph}}

452: \end{center}

453:  \end{figure}

454: %

455: Thus $G$ is a simple topological graph with $v_G=2I(S,L)=O(n^2/k^2)$

456: vertices and $e_G \geq 3|U_2|-2m^2$ edges. Since in this drawing of

457: $G$, every crossing is an intersection of two hyperbolas, the

458: crossing number of $G$ is upper bounded by ${\rm cr}(G)= O(|\Gamma|^2) =

459: O(m^4)$. We can also bound the crossing number of $G$ from below via

460: the Crossing Lemma of Ajtai~\etal~\cite{acns-82} and

461: Leighton~\cite{l-84}. It follows that

462: %

463: $$\Omega\left( \frac{e_G^3}{v_G^2}\right) -4v_G \leq {\rm cr}(G) \leq

464: O(m^4).$$

465: %

466: Rearranging this chain of inequalities, we obtain $e_G^3 = O(m^4v_G^2

467: + v_G^3)$, or $e_G= O(m^{4/3}v_G^{2/3} + v_G)$. Comparing this bound

468: with our lower bound $e_G\ge 3|U_2|-2m^2$, we have $|U_2|=

469: O(m^{4/3}v_G^{2/3} + v_G+ m^2)$. Hence, for $k\leq \sqrt{n}$, we

470: have

471: %

472: $$|U_2|= O\left( \left(\frac{n^2}{k^3}\right)^{4/3}

473:   \left(\frac{n^2}{k^2}\right)^{2/3} +\frac{n^2}{k^2} +

474:   \left(\frac{n^2}{k^3} \right)^2\right) =

475: O\left(\frac{n^4}{k^{16/3}}+\frac{n^2}{k^2}\right)=

476: O\left(\frac{n^4}{k^{16/3}}\right).$$

477: %

478: The total number of unit-area triangles is $|U_1|+|U_2| = O(n^2k+

479: n^4/k^{16/3})$. This expression is minimized for

480: $k=n^{6/19}$, and we get $|U_1|+|U_2| = O(n^{44/19})$.

481: \end{proof}

482: 

483: 

484: \subsection{Convex position}

485: 

486: The construction of Erd\H{o}s and Purdy~\cite{ep-71} with many triangles

487: of the same area, the $\sqrt{\log n}\times (n/\sqrt{\log n})$ section of

488: the integer lattice, also contains many collinear triples. Here we consider

489: the unit-area triangle problem in the special case of point sets in

490: strictly convex position, so no three points are collinear.

491: We show that $n$ points in convex position in the plane can determine a

492: superlinear number of unit-area triangles. On the other hand, we do not know

493: of any subquadratic upper bound.

494: 

495: \begin{theorem} \label{thm:convex-unit}

496: For all $n \geq 3$, there exist $n$-element point sets in convex

497: position in the plane that span $\Omega(n\log n)$ unit-area

498: triangles.

499: \end{theorem}

500: %

501: \begin{proof}

502: We recursively construct a set $S_i$ of $n_i=3^i$ points on the

503: unit circle that determine $t_i=i3^{i-1}$ unit-area triangles, for

504: $i=1,2,\ldots$.  Take a circle $C$ of unit radius centered

505: at the origin $o$. We start with a set $S_1$ of 3 points along the

506: circle forming a unit-area triangle, so we have $n_1=3$ points and

507: $t_1=1$ unit-area triangles. In each step, we triple the number of

508: points, i.e., $n_{i+1}=3n_i$, and create new unit-area triangles, so that

509: $t_{i+1}=3t_i+n_i$. This implies $n_i=3^i$, and $t_i= i 3^{i-1}$,

510: yielding the desired lower bound.

511: The $i$-th step, $i \geq 2$, goes as follows. Choose a

512: generic angle value $\alpha_i$, close to $\pi/2$, say,

513: and let $\beta_i$ be the angle such that the

514: three unit vectors at direction $0$, $\alpha_i$, and $\beta_i$ from

515: the origin determine a unit-area triangle, which we denote by $D_i$

516: (note that $\beta_i$ lies in the third quadrant).

517: Rotate $D_i$ around the origin to each position where its $0$ vertex

518: coincides with one of the $n_i$ points of $S_i$, and add the other

519: two vertices of $D_i$ in these positions

520: to the point set. (With appropriate choices of $S_1$ and the angles

521: $\alpha_i$, $\beta_i$, one can guarantee that no two points of any

522: $S_i$ coincide.) For each point of $S_i$, we

523: added two new points, so $n_{i+1}=3n_i$. Also, we have $n_i$ new unit-area

524: triangles from rotated copies of $D_i$; and each of the $t_i$

525: previous triangles have now two new copies rotated by $\alpha_i$ and

526: $\beta_i$. This gives $t_{i+1}=3t_i+n_i$.

527: \end{proof}

528: 

529: 

530: \section{Minimum-area triangles in the plane}

531: \label{sec:minimum2}

532: 

533: \paragraph{The general case.} We first present a simple but

534: effective charging scheme that gives an upper bound of $n^2-n$ on the

535: number of minimum (nonzero) area triangles spanned by $n$ points in

536: the plane (Lemma \ref{lem:min-area}).

537: This technique yields a very short proof of the minimum area result

538: from \cite{brs-01}, with a much better constant of proportionality.

539: Moreover, its higher-dimensional variants lead to asymptotically tight

540: bounds on the maximum number of minimum-volume $k$-dimensional

541: simplices in $\RR^d$, for any $1 \leq k \leq d$

542: (see Section \ref{sec:minimum3} for the case $k=2,d=3$, and

543: \cite{dt-07b} for the case $k=3,d=3$; the generalization to arbitrary

544: $1\leq k \leq d$ will be presented in the journal version of \cite{dt-07b}).

545: 

546: \begin{lemma} \label{lem:min-area}

547:  The number of triangles of minimum (nonzero)

548:  area spanned by $n$ points in the plane is at most $n^2-n$.

549: \end{lemma}

550: \begin{proof}

551: Consider a set $S$ of $n$ points in the plane. Assign every triangle

552: of minimum area to one of its longest sides. For a segment $ab$,

553: with $a,b\in S$, let $R_{ab}^+$ and $R_{ab}^-$ denote the two

554: rectangles of extents $|ab|$ and $2/|ab|$ with $ab$ as a common side.

555: If a minimum-area triangle $\Delta{abc}$ is assigned to $ab$, then

556: $c$ must lie in the relative interior of the side parallel to $ab$ in either

557: $R_{ab}^+$ or $R_{ab}^-$. If there were two points, $c_1$ and $c_2$,

558: on one of these sides, then the area of $\Delta{ac_1c_2}$ would be smaller

559: than that of $\Delta{abc}$, a contradiction. Therefore, at most two

560: triangles are assigned to each of the ${n \choose 2}$ segments

561: (at most one on each side of the segments), and so there are at

562: most $n^2-n$ minimum-area triangles.

563: \end{proof}

564: 

565: \smallskip

566: We now refine our analysis and establish a $\frac{2}{3}(n^2-n)$ upper bound, which

567: leaves only a small gap from our lower bound

568: $(\frac{6}{\pi^2}-o(1))n^2$; both bounds are presented in

569: Theorem~\ref{thm:min-area} below. Let us point out again

570: that here we allow collinear triples of points. The maximum number of

571: collinear triples is clearly ${n\choose 3}=\Theta(n^3)$.

572: The bounds below, however, consider only nondegenerate triangles of

573: {\em positive} areas.

574: 

575: % A segment $ab$ spanned by $S$ is a {\em visibility segment} if it

576: % does not contain any point of $S$ in its interior (see~\cite{kpw-05}

577: % for results on the graphs formed by visibility segments). Observe

578: % that every edge of a minimum area triangle spanned by $S$ must be a

579: % visibility segment. We deduce a better upper bound on the number of

580: % minimum area triangles by analyzing its connection to the number of

581: % visibility segments.

582: 

583: \begin{theorem} \label{thm:min-area}

584:  The number of triangles of minimum (nonzero)

585:  area spanned by $n$ points in the plane is at most

586:  $\frac{2}{3}(n^2-n)$.

587: The points in the $\lfloor\sqrt{n}\rfloor\times

588: \lfloor\sqrt{n}\rfloor$ integer grid span

589: $(\frac{6}{\pi^2}-o(1))n^2 \gtrapprox .6079 n^2$ minimum-area

590: triangles.

591: \end{theorem}

592: %

593: \begin{proof}

594: We start with the upper bound.

595: Consider a set $S$ of $n$ points in the plane, and let $L$ be the

596: set of connecting lines determined by $S$. Assume, without loss of

597: generality, that none of the lines in $L$ is vertical.

598: Let $T$ be the set of minimum (nonzero) area triangles spanned by $S$,

599: and put $t=|T|$.

600: There are $3t$ pairs $(ab,c)$ where $\Delta{abc}\in T$, and we may

601: assume, without loss of generality, that for at least half of these

602: pairs (i.e., for at least $\frac{3}{2}t$ pairs) $\Delta{abc}$ lies above

603: the line spanned by $a$ and $b$.

604: 

605: For each line $\ell\in L$, let $\ell'$ denote the line parallel to

606: $\ell$, lying above $\ell$, passing through some point(s) of $S$, and

607: closest to $\ell$ among these lines. Clearly, if $c\in S$

608: generates with $a,b\in\ell$ a minimum-area triangle which lies above

609: $ab$ then (i) $a$ and $b$ are a closest pair among the pairs of points

610: in $\ell\cap S$, and (ii) $c\in\ell_{ab}'$

611: (the converse does not necessarily hold).

612: 

613: Now fix a line $\ell\in L$; set $k_1=|\ell\cap S|\ge 2$, and

614: $k_2=|\ell'\cap S|\ge 1$, where $\ell'$ is as defined above.

615: The number of minimum-area triangles determined by a pair of points in

616: $\ell$ and lying above $\ell$ is at most $(k_1-1)k_2$. We have

617: %

618: \begin{equation} \label{eq:k12}

619: {k_1\choose 2} + {k_2\choose 2} \ge (k_1-1)k_2 .

620: \end{equation}

621: %

622: Indeed, multiplying by $2$ and subtracting the right-hand side from

623: the left-hand side gives

624: $$

625: k_1^2 -k_1 + k_2^2 -k_2 - 2k_1k_2 + 2k_2 =

626: (k_1-k_2)^2 - (k_1-k_2) \ge 0,

627: $$

628: which holds for any $k_1,k_2 \in \ZZ$.

629: 

630: We now sum (\ref{eq:k12}) over all lines $\ell\in L$. The sum of the

631: terms ${k_1\choose 2}$ is ${n\choose 2}$, and the sum of the terms

632: ${k_2\choose 2}$ is at most ${n\choose 2}$, because a line $\lambda\in L$

633: spanned by at least two points of $S$ can arise as the line $\ell'$ for at most

634: one line $\ell\in L$. Hence we obtain

635: $$

636: \frac{3}{2}t \le \sum_{\ell\in L} (k_1-1)k_2 \le 2{n\choose 2} =

637: n(n-1) ,

638: $$

639: thus $t\le \frac23(n^2-n)$, as asserted.

640: 

641: \smallskip

642: We now prove the lower bound. Consider the set $S$ of points in the

643: $\lfloor \sqrt{n}\rfloor \times \lfloor\sqrt{n}\rfloor$ section of the

644: integer lattice. Clearly $|S| \leq n$. The minimum nonzero area of

645: triangles in $S$ is $1/2$ (by Pick's theorem). Recall that the

646: charging scheme used in the proof of Lemma~\ref{lem:min-area}

647: assigns each triangle of minimum area to one of its

648: longest sides, which is necessarily a {\em visibility segment}

649: (a segment not containing any point of $S$ in its relative interior).

650: We show that every visibility segment $ab$ which is not

651: axis-parallel is assigned to exactly two triangles of minimum area.

652: 

653: Draw parallel lines to $ab$ through all points of the integer

654: lattice. Every line parallel to $ab$ and incident to a point of $S$

655: contains equally spaced points of the (infinite) integer lattice.

656: The distance between consecutive points along each line is exactly

657: $|ab|$.

658: % The gap between consecutive lines is at most $\sqrt{2}/2<1$.

659: This implies that each of the two lines parallel to $ab$ and closest

660: to it contains a lattice point on the side of the respective rectangle

661: $R_{ab}^-$ or $R_{ab}^+$,

662: opposite to $ab$, and this lattice point is in $S$.

663: Finally, observe that there are no empty acute triangles in the integer

664: lattice. It follows that our charging scheme uniquely assigns empty

665: triangles to visibility segments. An illustration is provided in

666: Figure~\ref{fig:grid}.

667: 

668: \begin{figure}[htbp]

669:   \begin{center}

670: \epsfig{file=grid.eps,width=2.5in,clip=}

671:  \caption{\small In an integer lattice section, every visibility segment

672:  which is not axis-parallel is the longest side of two triangles of

673:  minimum area.}

674: \end{center}

675: \label{fig:grid}

676: \end{figure}

677: 

678: A non-axis-parallel segment $ab$ is a visibility segment if

679: and only if the coordinates of the vector $\overrightarrow{ab}$ are

680: relatively prime. It is well known that $6/\pi^2$ is the limit of the

681: probability that a pair of integers $(i,j)$ with $ 1 \leq i,j \leq m$

682: are relatively prime, as $m$ tends to infinity \cite{v-54}.

683: Hence, a fraction of about

684: $6/\pi^2$ of the ${|S| \choose 2} \leq {n \choose 2}$ segments

685: spanned by $S$ are visibility segments which are not axis-parallel.

686: Each of these $(\frac{6}{\pi^2}-o(1)){n\choose 2}$ segments corresponds

687: to two unique triangles of minimum area, so $S$ determines at least

688: $(\frac{6}{\pi^2}-o(1))n^2$ minimum-area triangles.

689: \end{proof}

690: 

691: 

692: \subsection{Special cases}

693: 

694: In this subsection we consider some new variants of the minimum-area

695: triangle problem for the two special cases (i) where no three points are

696: collinear, and (ii) where the points are in convex position.

697: We also show that the maximum number of {\em acute} triangles

698: of minimum area, for any point set, is only linear.

699: 

700: \paragraph{Acute triangles.}

701: We have seen that $n$ points in an integer grid may span

702: $\Omega(n^2)$ triangles of minimum area. However, in that

703: construction, all these triangles are obtuse (or right-angled).

704: Here we prove that for any $n$-element point set in the plane, the

705: number of {\em acute} triangles of minimum area is only linear.

706:     This bound is attained in the following simple example.

707:     Take two groups of about $n/2$ equally spaced points on two parallel

708:     lines: the first group consist of the points

709:     $(i,0)$, for $i=0,\ldots,\lceil n/2 \rceil -1$, and the second

710:     group of the

711:     points $(i+1/2,\sqrt{3}/2)$, for $i=0,\ldots,\lfloor n/2 \rfloor -1$.

712:     This point set determines $n-2$ acute triangles of minimum area.

713: 

714: \begin{theorem} \label{thm:acute}

715: The maximum number of acute triangles of minimum area determined by $n$

716: points in the plane is $O(n)$. This bound is asymptotically tight.

717: \end{theorem}

718: %

719: \begin{proof}

720: Let $S$ be a set of $n$ points in the plane, and let $\T$ denote the

721: set of acute minimum-area triangles determined by $S$. Define a

722: geometric graph $G=(V,E)$ on $V=S$, where $uv \in E$ if and only if

723: $uv$ is a shortest side of a triangle in $\T$. We first argue that

724: every segment $uv$ is a shortest edge of at most two triangles in

725: $\T$, and then we complete the proof by showing that $G$ is planar

726: and so it has only $O(n)$ edges.

727: 

728: Let $\Delta{a_1 b_1 c_1} \in \T$ and assume that $b_1 c_1$ is a

729: shortest side of $\Delta{a_1 b_1 c_1}$.

730: % We show that $b_1 c_1$ is a shortest side of at most two acute

731: % triangles of minimum area.

732: Let $\Delta{a_2 b_2 c_2}$ be the triangle such that the midpoints of its

733: sides are $a_1,b_1,c_1$; and let $\Delta{a_3 b_3 c_3}$ be the

734: triangle such that the midpoints of its sides are $a_2,b_2,c_2$.

735: Refer to Figure~\ref{acute-obtuse}(a). Since $\Delta{a_1 b_1 c_1}$

736: has minimum area, then, in the notation of the figure,

737: each point of $S \setminus \{a_1,b_1,c_1\}$ lies in one of the

738: (closed) regions $R_1$ through $R_6$ or on one of the lines

739: $\ell_2$, $\ell_4$ or $\ell_5$; also, no point of $S \setminus

740: \{a_1,b_1,c_1\}$  lies in the interior of $\Delta{a_3 b_3 c_3}$.

741: Similarly, any point $a\in S$ of

742: a triangle $\Delta{ab_1c_1} \in \T$ must lie on $\ell_1$ or

743: $\ell_3$. Thus $a=a_1$ and $a=a_2$ are the only possible positions

744: of $a$. This follows from the fact that the triangles of $T$ are

745: acute: any point on, say, $\ell_1\cap\bd R_2$ or $\ell_1\cap\bd R_6$

746: forms an {\em obtuse} triangle with $b_1c_1$.

747: 

748: Consider two acute triangles $\Delta{a_1b_1c_1},\Delta{xyz} \in \T$

749: of minimum area with shortest sides $b_1 c_1\in E$ and $xy\in E$,

750: respectively. Assume that edges $b_1c_1$ and $xy$ cross each other.

751: %

752: We have the following four possibilities: (i) $x$ and $y$ lie in two

753: opposite regions $R_iR_{i+3}$, for some $i \in \{1,2,3\}$; (ii)

754: $x=a_1$ and $y \in R_4$; (iii) $x \in \ell_4$ and $y \in R_4$; (iv)

755: $x \in \ell_5$ and $y \in R_4$. Since $xy$ is a shortest side of

756: $\Delta{xyz}$, the distance from $z$ to the line through $x$ and $y$ is at

757: least $\sqrt{3}/2|xy|$. But then, in all four cases $\Delta{xyz}$

758: cannot be an acute triangle of minimum area, since it contains one

759: of the vertices of $\Delta{a_1 b_1 c_1}$ in its interior, a

760: contradiction. (For instance if $x \in R_1$ and $y \in R_4$,

761: $\Delta{xyc_1}$ would be obtuse and $\Delta{xyz}$ contains $c_1$ in

762: its interior, or if $x=a_1$ and $y \in R_4$, $\Delta{xyz}$ contains

763: either $b_1$ or $c_1$ in its interior.)

764: \end{proof}

765: 

766: %

767: \begin{figure}[htbp]

768: \centerline{\epsfxsize=3in \epsffile{acute.eps}

769: \hskip 0.3in

770: \epsfxsize=3in \epsffile{obtuse.eps}}

771: \centerline{\hfill (a)\hskip 2.4in (b)\hfill}

772: \caption{\small (a) Acute triangles: the graph $G$ is planar.

773: (b) Convex position: the graph $G$ is quasi-planar.}

774: \label{acute-obtuse}

775: \end{figure}

776: %

777: 

778: 

779: \paragraph{Convex position.}

780: For points in strictly convex position we prove a tight $\Theta(n)$

781: bound on the maximum possible number of minimum-area triangles.

782: Note that a regular $n$-gon has $n$ such triangles,

783: so it remains to show an $O(n)$ upper bound. Also,

784: $n$ points equally distributed on two parallel

785: lines (at equal distances) give a well-known quadratic lower bound, 

786: so the requirement that the points be in strictly convex position

787: is essential for the bound to hold.

788: 

789: \begin{theorem} \label{thm:convex-min}

790: The maximum number of minimum-area triangles determined by $n$ points in

791: (strictly) convex position in the plane is $O(n)$. 

792: This bound is asymptotically tight.

793: \end{theorem}

794: %

795: \begin{proof}

796: The argument below is similar to that in the proof of

797: Theorem~\ref{thm:acute}. Since there can be only $O(n)$

798: acute triangles of minimum area, it is sufficient to consider

799: right-angled and obtuse triangles (for simplicity,

800: we refer to both types as obtuse), even though

801: the argument also works for acute triangles. We use a

802: similar notation: now $\T$ denotes the set of obtuse triangles of

803: minimum area. We define a geometric graph $G=(V,E)$ on $V=S$, where

804: $uv \in E$ if and only if $uv$ is a shortest side of a triangle in

805: $\T$. See Figure \ref{acute-obtuse}(b).

806: 

807: Let $\Delta{a_1 b_1 c_1} \in \T$ with $b_1c_1$ a shortest side. By

808: convexity, at most four triangles in $\T$ can have a common shortest

809: side $b_1c_1$: at most two such triangles have a third vertex on

810: $\ell_1$ and at most another two of them have a third vertex on $\ell_3$.

811: A graph drawn in the plane is said to be {\em quasi-planar}

812: if it has no three edges which are pairwise crossing; it is

813: known~\cite{aapps-97} (see also \cite{at-07})

814: that any quasi-planar graph with $n$ vertices

815: has at most $O(n)$ edges.

816: We now show that $G$ is quasi-planar, which will complete the proof

817: of the theorem.

818: 

819: Consider the triangles

820: $\Delta{a_2 b_2 c_2}$ and $\Delta{a_3 b_3 c_3}$, defined as in the

821: proof of Theorem~\ref{thm:acute}.

822: Each point of $S \setminus \{a_1,b_1,c_1\}$ lies in one of the

823: (closed) regions $R_1$ through $R_6$;

824: in particular no such point

825: lies in the interior of $\Delta{a_3 b_3 c_3}$. 

826: (Here, unlike the previous analysis, strict convexity rules out points

827: on any of the three middle lines, such as $\ell_2$.)

828: In addition, by

829: convexity, the regions $R_1$, $R_3$ and $R_5$ are empty of points.

830: Assume now that $b_1 c_1$, $xy$, $uv$ form a triplet of pairwise

831: crossing edges, where $xy$ and $uv$ are distinct shortest sides of

832: two triangles $\Delta{xyz} \in \T$ and $\Delta{uvw} \in \T$. 

833: It follows that each of the two edges $xy$ and $uv$ must have one

834: endpoint at $a_1$ and the other in $R_4$ (since each crosses $b_1 c_1$). 

835: Thus two edges in this triplet have  a common

836: endpoint, and so they do not cross, which is a contradiction.

837: \end{proof}

838: 

839: \paragraph{No three collinear points.}

840: We conjecture that if no three points are collinear, then the maximum

841: number of triangles of minimum area is close to linear. It is not

842: linear, though: It has been proved recently~\cite{dpt-05} that there

843: exist $n$-element point sets

844: in the plane that span $\Omega(n \log{n})$ empty congruent triangles.

845: Here, we show that one can repeat this construction such that there is

846: no collinear triples of points and that the $\Omega(n \log{n})$ empty

847: congruent triangles have minimum (nonzero) area.

848: However, we do not know of any sub-quadratic upper bound.

849: 

850: 

851: \begin{theorem} \label{thm:general}

852:   For all $n \geq 3$, there exist $n$-element point sets in the plane that

853:   have no three collinear points and span $\Omega(n\log n)$ triangles of

854:   minimum (nonzero) area.

855: \end{theorem}

856: %

857: \begin{proof}

858: The construction is essentially the one given in \cite{dpt-05}, and

859: we provide here only a brief description. We then specify the

860: additional modifications needed for our purposes. First, a point set

861: $S$ is constructed with many, i.e., $\Omega(n \log{n})$, pairwise

862: congruent triples of collinear points, which can be also viewed as

863: degenerate empty congruent triangles. Then 

864: this construction is slightly perturbed to obtain a set of points $S$

865: with no collinear triples, so that these degenerate triangles become

866: non-degenerate empty congruent triangles of minimum (nonzero) area. 

867: The details are as follows (see~\cite{dpt-05}).

868: 

869: Let $n=3^k$ for some $k\in \NN$.

870: Consider $k$ unit vectors $b_1,\ldots,b_k$, and for

871: $1\le i\le k$, let $\beta_i$ be the counterclockwise angle from the

872: $x$-axis to $b_i$.

873: % We choose the angles $\beta_i$ independently and

874: % uniformly at random from the interval $(0, \pi/{2})$.

875: Let $\lambda \in (0,1)$ be fixed and let $a_i=\lambda b_i$. Consider

876: now all $3^k$ possible sums of these $2k$ vectors, $a_i$ and $b_i$,

877: $1\le i\le k$,

878: with coefficients $0$ or $1$, satisfying the condition that for each

879: $i$, at least one of $a_i$ or $b_i$ has coefficient $0$.

880: Let $S$ be the set of $3^k$ points determined by these vectors.

881: Clearly, each triple of the form

882: ($v$, $v+a_i$, $v+b_i$), where $v$ is a subset sum that does not

883: involve $a_i$ or $b_i$, consists of collinear points.

884: For such a triple, denote by $s_i(v)$ the segment whose endpoints are

885: $v$ and $v+b_i$.  We say that the collinear triple $(v, v+a_i, v+b_i)$

886: is of type $i$, $i=1,\ldots,k$.

887: For each $i$ there are exactly $3^{k-1}$

888: triples of type $i$, therefore a total of $ k 3^{k-1} = (n

889: \log{n})/(3 \log{3}) = \Omega(n \log{n}) $ triples of collinear

890: points. Clearly, all these triples form degenerate congruent

891: triangles in $S$.

892: Denote by $\ell_i(v)$ the line supporting the segment $s_i(v)$,

893: and by $\L$ the set of lines corresponding to these triples.

894: 

895: We need the following slightly stronger version of Lemma 1

896: in~\cite{dpt-05}. The proof is very similar to the proof of

897: Proposition 1 in~\cite{dpt-05}, and we omit the details.

898: 

899: \begin{lemma} \label{lem:exist}

900: There exist angles $\beta_1,\ldots,\beta_k$,

901: and $\lambda \in (0,1)$, such that

902: {\rm (i)} $S$ consists of $n$ distinct points;

903: {\rm (ii)} if $u,v,w \in S$ are collinear (in this order), then

904: $v=u+a_i$ and $w=u+b_i$.

905: \end{lemma}

906: 

907: Let $\eps$ be the minimum distance between points

908: $p \in S \setminus \{v, v+a_i, v+b_i\}$

909: and lines $\ell_i(v) \in \L$, over all pairs $(v,i)$.

910: By Lemma~\ref{lem:exist}, we have $\eps>0$. Now instead of choosing

911: $a_i$ to be collinear with $b_i$, slightly rotate $\lambda b_i$

912: counterclockwise from $b_i$ through a sufficiently small

913: angle $\delta$

914: about their common origin, so the collinearity disappears.

915: This modification is carried out at the

916: same time for all vectors $a_i$, $i=1,\ldots,k$,

917: that participate in the construction.

918: By continuity, there exists a sufficiently small $\delta=\delta(\eps)>0$,

919: so that (i) each of the triangles $\Delta(v, v+a_i, v+b_i)$

920: remains empty throughout this small perturbation, (ii) the point set

921: $S$ is in general position after the perturbation, and (iii) the

922: congruent triangles $\Delta(v, v+a_i, v+b_i)$ have minimum area.

923: This completes the proof.

924: \end{proof}

925: 

926: \vspace{-\baselineskip}

927: 

928: \section{Unit-area triangles in 3-space}

929: \label{sec:unit3}

930: 

931: Erd\H{o}s and Purdy~\cite{ep-71} showed that a $\sqrt{\log n}\times

932: (n/\sqrt{\log n})$ section of the integer lattice determines

933: \linebreak $\Omega(n^2 \log\log{n})$ triangles of the same area.

934: Clearly, this bound is also valid in 3-space. They have also

935: derived an upper bound of $O(n^{8/3})$ on the number of unit-area

936: triangles in $\RR^3$. Here we improve this bound to 

937: $O(n^{17/7}\beta(n)) = O(n^{2.4286})$.

938: We use $\beta(n)$ to denote any function of the form

939: $\exp(\alpha(n)^{O(1)})$, where $\alpha(n)$ is the extremely slowly

940: growing inverse Ackermann function. Any such function $\beta(n)$ is also

941: extremely slowly growing.

942: %In this section we obtain the following upper bound:

943: 

944: \begin{theorem}\label{thm:unit3}

945:   The number of unit-area triangles spanned by $n$ points in $\RR^3$

946:   is $O(n^{17/7}\beta(n)) = O(n^{2.4286})$.

947: \end{theorem}

948: 

949: The proof of the theorem is quite long, and involves several technical steps.

950: Let $S$ be a set of $n$ points in $\RR^3$. For each pair $a,b$ of

951: distinct points in $S$, let $\ell_{ab}$ denote the line passing

952: through $a$ and $b$, and let $C_{ab}$ denote the cylinder whose axis

953: is $\ell_{ab}$ and whose radius is $2/|ab|$. Clearly, any point $c\in

954: S$ that forms with $ab$ a unit-area triangle, must lie on $C_{ab}$.

955: The problem is thus to bound the number of incidences between

956: ${n\choose 2}$ cylinders and $n$ points, but it is complicated for two

957: reasons: (i) The cylinders need not be distinct.  (ii) Many distinct

958: cylinders can share a common generator line, which may contain many

959: points of $S$.

960: 

961: \vspace{-1mm}

962: \paragraph{Cylinders with large multiplicity.}

963: Let $\C$ denote the multiset of the ${n\choose 2}$ cylinders

964: $C_{ab}$, for $a,b\in S$. Since the cylinders in $\C$ may appear

965: with multiplicity, we fix a parameter $\mu=2^j$, $j=0,1,\ldots$,

966: and consider separately incidences with each of the sets $\C_\mu$,

967: of all the cylinders whose multiplicity

968: is between $\mu$ and $2\mu-1$. Write $c_\mu=|\C_\mu|$.

969: We regard $\C_\mu$ as a set (of distinct cylinders), and will multiply

970: the bound that we get for the cylinders in $\C_\mu$ by $2\mu$,

971: to get an upper bound on the number of

972: incidences that we seek to estimate. We will then sum up the resulting

973: bounds over $\mu$ to get an overall bound.

974: 

975: Let $C$ be a cylinder in $\C_\mu$. Then its axis $\ell$ must contain

976: $\mu$ pairs of points of $P$ at a fixed distance apart (equal to $2/r$,

977: where $r$ is the radius of $C$). That is, $\ell$ contains

978: $t>\mu$ points of $S$. Let us now fix $t$ to be a power of $2$, and

979: consider the subset $\C_{\mu,t} \subset \C_\mu$ of those cylinders in $\C_\mu$

980: that have at least $t$ and at most $2t-1$ points on their axis.

981: By the Szemer\'edi-Trotter Theorem~\cite{st-83} (or, rather, its

982: obvious extension to 3-space), the number of lines containing at

983: least $t$ %(and at most $2t-1$)

984: points of $S$ is $O(n^2/t^3 + n/t)$.

985: Any such line $\ell$ can be the axis of many cylinders in $\C_\mu$

986: (of different radii).  Any such cylinder ``charges'' $\Theta(\mu)$

987: pairs of points out of the $O(t^2)$ pairs along $\ell$,

988: and no pair is charged more than once.

989: Hence, for a given line $\ell$ incident to at least $t>\mu$ and at

990: most $2t-1$ points of $S$, the number of distinct cylinders in

991: $\C_\mu$ that have $\ell$ as axis is $O(t^2/\mu)$. Summing over

992: all axes incident to at least $t$ and at most $2t-1$ points

993: yields that the number of distinct cylinders in $\C_{\mu,t}$ is

994: %

995: \begin{equation}\label{nut}

996: c_{\mu,t}=

997: O\left(\left(\frac{n^2}{t^3} + \frac{n}{t}\right)\frac{t^2}{\mu}\right)=

998: O\left(\frac{n^2}{t\mu} + \frac{nt}{\mu}\right).

999: \end{equation}

1000: %

1001: We next sum this over $t$, a power of 2 between $\mu$ and $\nu$,

1002: and conclude that the number of distinct cylinders

1003: in $\C_\mu$ having at most $\nu$ points on their axis is

1004: %

1005: \begin{equation} \label{cmu}

1006: c_{\mu,\le\nu} = O\left( \frac{n^2}{\mu^2} + \frac{n \nu}{\mu} \right).

1007: \end{equation}

1008: %

1009: 

1010: \vspace{-1mm}

1011: \paragraph{Restricted incidences between points and cylinders.}

1012: We distinguish two {\em type}s of incidences, which we count

1013: separately. An incidence between a point $p$ and a cylinder $C$ is of

1014: {\em type 1} if the generator of $C$ passing through $p$ contains at

1015: least one additional point of $S$; otherwise it is of {\em type 2}. We

1016: begin with the following subproblem, in which we bound the number of

1017: incidences between the cylinders of $\C$, counted with multiplicity,

1018: and {\em multiple} points that lie on their generator lines, as well

1019: as incidences with cylinders with ``rich'' axes.

1020: Specifically, we have the following lemma.

1021: %

1022: \begin{lemma} \label{107}

1023:   Let $S$ be a set of $n$ points and $\C$ be the multiset of the

1024:   ${n\choose 2}$ cylinders $C_{ab}$, for $a,b\in S$ (counted with

1025:   multiplicity). The

1026:   total number of all incidences of type 1 and all incidences

1027:   involving cylinders having at least $n^{14/45}$ points on their axis

1028:   is bounded by $O(n^{107/45}\polylog(n))=O(n^{2.378})$.

1029: \end{lemma}

1030: %

1031: \begin{proof}

1032: Let $L$ denote the set of lines spanned by the points of $S$. Fix

1033: a parameter $k=2^i$, $i=1,\ldots$,

1034: and consider the set $L_k$ of all lines that contain at least $k$ and

1035: at most $2k-1$ points of $S$.

1036: We bound the number of incidences between

1037: cylinders in $\C$ that contain lines in $L_k$ as generators

1038: and points that lie on those lines. Formally, we bound the

1039: number of triples $(p,\ell,C)$, where $p\in S$, $\ell\in L_k$,

1040: and $C\in\C$, such that $p\in\ell$ and $\ell\subset C$.

1041: Summing these bounds over $k$ will give us a bound for the

1042: number of incidences of type 1. Along the way, we will also dispose of

1043: incidences with cylinders whose axes contain many points.

1044: 

1045: As already noted, the Szemer\'edi-Trotter Theorem~\cite{st-83} implies that

1046: ${\displaystyle

1047: \lambda_k:=|L_k|=O\left( \frac{n^2}{k^3} + \frac{n}{k} \right)}$.

1048: 

1049: \vspace{-1mm}

1050: \paragraph{Line-cylinder incidences.}

1051: Consider the subproblem of bounding the number of

1052: incidences between lines in $L_k$ and cylinders in $\C$, where

1053: a line $\ell$ is said to be incident to cylinder $C$ if $\ell$

1054: is a generator of $C$. We will then multiply the resulting

1055: bound by $2k$ to get an upper bound on the number of

1056: point-line-cylinder incidences involving $L_k$, and then

1057: sum the resulting bounds over $k$.

1058: 

1059: \vspace{-1mm}

1060: \paragraph{Generator lines with many points.}

1061: Let us first dispose of the case $k>n^{1/3}$. Any line

1062: $\ell\in L_k$ can be a generator of at most $n$ cylinders (counted

1063: with multiplicity), because, having fixed $a\in S$, the point

1064: $b\in S$ such that $C_{ab}$ contains $\ell$ is determined

1065: (up to multiplicity $2$). Hence the number of incidences between

1066: the points that lie on $\ell$ and the cylinders of $\C$ is

1067: $O(nk)$. Summing over $k=2^i>n^{1/3}$ yields the overall bound

1068: $$

1069: O\left( \sum_k nk\lambda_k \right) =

1070: O\left( \sum_k \left( \frac{n^3}{k^2} + n^2 \right) \right) = O(n^{7/3}) .

1071: $$

1072: Hence, in what follows, we may assume that $k\le n^{1/3}$.

1073: In this range of $k$ we have

1074: \begin{equation} \label{lambdak}

1075: \lambda_k=O\left( \frac{n^2}{k^3} \right) .

1076: \end{equation}

1077: 

1078: \vspace{-1mm}

1079: \paragraph{Axes with many points.}

1080: Let us also fix the multiplicity $\mu$ of the cylinders under

1081: consideration (up to a factor of $2$, as above).

1082: The number of distinct cylinders in $\C_\mu$

1083: having between $t>\mu$ and $2t-1$ points on their axes, is

1084: $O(n^2/(t\mu) + nt/\mu)$; see (\ref{nut}). While the first term

1085: is sufficiently small for our purpose, the second

1086: term may be too large when $t$ is large. To avoid this difficulty, we

1087: fix another threshold exponent $z < 1/2$ that we will optimize later,

1088: and handle separately the cases $t\ge n^{z}$ and $t< n^{z}$.

1089: That is, in the first case, for $t\ge n^{z}$ a power of 2, we seek an

1090: upper bound on the overall number of incidences between the points of

1091: $S$ and the cylinders in $\C$ whose axis contains between $t$ and $2t-1$

1092: points of $S$. (For this case, we combine all the multiplicities

1093: $\mu<t$ together.) By the Szemer\'edi-Trotter theorem~\cite{st-83},

1094: the number of such axes is $O(n^2/t^3+n/t)$.

1095: 

1096: Fix such an axis $\alpha$. It defines $\Theta(t^2)$ cylinders,

1097: and the multiplicity of any of these cylinders is at most $O(t)$. Since no

1098: two distinct cylinders in this collection can pass through the same

1099: point of $S$, it follows that the total number of incidences

1100: between the points of $S$ and these cylinders is $O(nt)$.

1101: Hence the overall number of incidences under consideration is

1102: $O(n^2/t^3+n/t) \cdot O(nt)= O(n^3/t^2+n^2)$.

1103: Summing over all $t\ge n^z$,

1104: a power of $2$, we get the overall bound $O(n^{3-2z})$.

1105: 

1106: Note that this bound takes care of {\em all} the incidences between

1107: the points of $S$ and the cylinders having at least $t\ge n^z$ points

1108: along their axes, not just those of type 1 (involving multiple points

1109: on generator lines).

1110: 

1111: %\paragraph{Incidences with low multiplicity.}

1112: \paragraph{Cylinders with low multiplicity.}

1113: We now confine the analysis to cylinders having fewer than $n^z$

1114: points on their axis, and go back to fixing the multiplicity $\mu$,

1115: which we may assume to be at most $n^z$.

1116: We thus want to bound the number of

1117: incidences between $\lambda_k$ distinct lines and $c_{\mu,\le n^z}$

1118: distinct cylinders in $\C_\mu$, for given $k\le n^{1/3}$, $\mu \leq n^z$.

1119: Note that a cylinder can contain a line if and only if it is parallel to the

1120: axis of the cylinder, so we can split the

1121: problem into subproblems, each associated with some direction

1122: $\theta$, so that in the $\theta$-subproblem we have a set of some

1123: $c_\mu^{(\theta)}$ cylinders and a set of some

1124: $\lambda_k^{(\theta)}$ lines, so that the lines and the cylinder axes

1125: are all parallel (and have direction $\theta$);

1126: we have $\sum_\theta c_\mu^{(\theta)} = c_{\mu,\le n^z}$,

1127: and $\sum_\theta \lambda_k^{(\theta)} = \lambda_k$.

1128: 

1129: For a fixed $\theta$, we project the cylinders and lines in the

1130: $\theta$-subproblem onto a plane with normal direction $\theta$,

1131: and obtain a set of $c_\mu^{(\theta)}$ circles and a set of

1132: $\lambda_k^{(\theta)}$ points, so that the number of line-cylinder

1133: incidences is equal to the number of point-circle incidences.

1134: By \cite{ANPPSS,ArS,MT},\footnote{%

1135:   The bound that we use, from \cite{MT}, is slightly better than

1136:   the previous ones.}

1137: the number of point-circle incidences between $N$ points and $M$ circles in the plane is

1138: $O(N^{2/3}M^{2/3} + N^{6/11}M^{9/11}\log^{2/11}(N^3/M)+N+M)$.

1139: It follows that the number of such line-cylinder incidences is

1140: \begin{equation} \label{eq:theta}

1141: O\left(

1142: (\lambda_k^{(\theta)})^{2/3}

1143: (c_\mu^{(\theta)})^{2/3} +

1144: (\lambda_k^{(\theta)})^{6/11}

1145: (c_\mu^{(\theta)})^{9/11}

1146:   \log^{2/11}((\lambda_k^{(\theta)})^3/c_\mu^{(\theta)})+

1147: \lambda_k^{(\theta)} +

1148: c_\mu^{(\theta)} \right) .

1149: \end{equation}

1150: Note that, for any fixed $\theta$, we have

1151: $\lambda_k^{(\theta)} \le n/k$ and

1152: $c_\mu^{(\theta)} \le n^{1+z}/\mu$.

1153: The former inequality is trivial.  To see the latter inequality,

1154: note that an axis with $t<n^z$ points

1155: defines ${t\choose 2}$ cylinders. Since we only consider cylinders

1156: with multiplicity $\Theta(\mu)$, the number of distinct such

1157: cylinders is $O(t^2/\mu)$, and the number of lines

1158: (of direction $\theta$) with about $t$

1159: points on them is at most $n/t$, for a total of at most $O(nt/\mu)$

1160: distinct cylinders. Partitioning the range $\mu < t\le n^{z}$ by

1161: powers of $2$, as above, and summing up the resulting bounds, the bound

1162: $c_\mu^{(\theta)} \le n^{1+z}/\mu$ follows.

1163: 

1164: Summing over $\theta$, and using H\"older's inequality, we have

1165: (here $x$ is a parameter between $2/11$ and $6/11$

1166: that we will fix shortly)

1167: $$

1168: \sum_\theta

1169: (\lambda_k^{(\theta)})^{6/11}

1170: (c_\mu^{(\theta)})^{9/11} \le

1171: \left(\frac{n}{k}\right)^{6/11-x}

1172: \left(\frac{n^{1+z}}{\mu}\right)^{x-2/11}

1173: \sum_\theta

1174: (\lambda_k^{(\theta)})^{x}

1175: (c_\mu^{(\theta)})^{1-x} \le

1176: $$

1177: $$

1178: \frac{n^{(4-2z)/11+xz}}{k^{6/11-x}\mu^{x-2/11}}

1179: \left( \sum_\theta

1180: \lambda_k^{(\theta)} \right)^{x}

1181: \left( \sum_\theta

1182: c_\mu^{(\theta)} \right)^{1-x} =

1183: \frac{n^{(4-2z)/11+xz}}{k^{6/11-x}\mu^{x-2/11}}

1184: \lambda_k^x c_{\mu,\le n^z}^{1-x} .

1185: $$

1186: We need to multiply this bound by $\Theta(k\mu)$.

1187: Substituting the bounds $\lambda_k=O(n^2/k^3)$ from (\ref{lambdak}),

1188: and $c_{\mu,\le n^z} = O(n^2/\mu^2 + n^{1+z}/\mu)$ from (\ref{cmu}),

1189: we get the bound

1190: \begin{eqnarray*}

1191: && O\left(

1192: n^{(4-2z)/11+xz} k^{5/11+x}\mu^{13/11-x}

1193: \left( \frac{n^2}{k^3} \right)^x

1194: \left( \frac{n^2}{\mu^2} + \frac{n^{1+z}}{\mu} \right)^{1-x}

1195: \log^{2/11} n\right)\\

1196: &=&O\left( k^{5/11-2x} \left(

1197: n^{2+(4-2z)/11+xz} \mu^{x-9/11} +

1198: n^{(15+9z)/11+x} \mu^{2/11} \right)

1199: \log^{2/11} n \right) .

1200: \end{eqnarray*}

1201: Choosing $x=5/22$ (the smallest value for which the exponent of $k$

1202: is non-positive), the first term becomes

1203: $O( n^{2+4/11+z/22} \log^{2/11} n)$, which we need to balance with

1204: $O(n^{3-2z})$; for this, we choose $z=14/45$ and obtain the bound

1205: $O(n^{107/45}\log^{2/11}n)=O(n^{2.378})$; for this choice of $z$,

1206: recalling that $\mu<n^z$, the second term is dominated by the first.

1207: Summing over $k,\mu$ only adds logarithmic factors, for a resulting

1208: overall bound $O(n^{2.378})$.

1209: 

1210: Similarly, we have (with a different choice of $x$, soon to be made)

1211: $$

1212: \sum_\theta

1213: (\lambda_k^{(\theta)})^{2/3}

1214: (c_\mu^{(\theta)})^{2/3} \le

1215: \left(\frac{n}{k}\right)^{2/3-x}

1216: \left(\frac{n^{1+z}}{\mu}\right)^{x-1/3}

1217: \sum_\theta

1218: (\lambda_k^{(\theta)})^{x}

1219: (c_\mu^{(\theta)})^{1-x} \le

1220: $$

1221: $$

1222: \frac{n^{(1-z)/3+xz}}{k^{2/3-x}\mu^{x-1/3}}

1223: \left( \sum_\theta

1224: \lambda_k^{(\theta)} \right)^{x}

1225: \left( \sum_\theta

1226: c_\mu^{(\theta)} \right)^{1-x} =

1227: \frac{n^{(1-z)/3+xz}}{k^{2/3-x}\mu^{x-1/3}}

1228: \lambda_k^x c_{\mu,\le n^z}^{1-x} .

1229: $$

1230: Multiplying by $k\mu$ and arguing as above, we get

1231: \begin{eqnarray*}

1232: &&O\left(

1233: n^{(1-z)/3+xz} k^{1/3+x}\mu^{4/3-x}

1234: \left( \frac{n^2}{k^3} \right)^x

1235: \left( \frac{n^2}{\mu^2} + \frac{n^{1+z}}{\mu} \right)^{1-x}

1236: \log^{2/11} n\right)\\

1237: &=&

1238: O\left( k^{1/3-2x} \left(

1239: n^{2+(1-z)/3+xz} \mu^{x-2/3} +

1240: n^{1+(1+2z)/3+x} \mu^{1/3} \right)

1241: \log^{2/11} n\right) .

1242: \end{eqnarray*}

1243: We choose here $x=1/6$ and note that, for $z=14/45$ and $\mu < n^z$,

1244: the bound is smaller than $O(n^{7/3})$, which is dominated by the

1245: preceding bound $O(n^{2.378})$.

1246: 

1247: Finally, the linear terms in (\ref{eq:theta}), multiplied by $k\mu$,

1248: add up to 

1249: $$

1250: k\mu \sum_{\theta} O\left(\lambda_k^{(\theta)}+c_\mu^{(\theta)}\right)=

1251: O\left( k\mu \left( \lambda_k + c_{\mu,\le n^z} \right)\right) =

1252: O\left( \frac{n^2\mu}{k^2} + \frac{n^2k}{\mu} + n^{1+z}k \right) ,

1253: $$

1254: which, by our assumptions on $k$, $\mu$, and $z$ is also dominated by

1255: $O(n^{2.378})$. Summing over $k,\mu$ only add logarithmic factors, for

1256: a resulting overall bound $O(n^{2.378})$.

1257: This completes the proof of Lemma~\ref{107}.

1258: \end{proof}

1259: 

1260: It therefore remains to count point-cylinder incidences of type 2,

1261: involving cylinders having at most $n^{14/45}$ points on their axes.

1262: 

1263: \vspace{-1mm}

1264: \paragraph{The intersection pattern of three cylinders.}

1265: We need the following technical lemma, whose proof is borrowed from a

1266: yet unpublished work \cite{EPS:cyl}, and is presented in the appendix.

1267: 

1268: \begin{lemma}\label{3int}

1269:   Let $C,C_1,C_2$ be three cylinders with no pair of parallel axes.

1270:   Then $C\cap C_1\cap C_2$ consists of at most 8 points.

1271: \end{lemma}

1272: 

1273: \paragraph{Point-cylinder incidences.}

1274: Using the partition technique~\cite{ch-05,ps-04}

1275: %\micha{Just making sure that this is the right REF?}

1276: %\csaba{This is a handbook summary of the partition technique.

1277: % For such a general method, it is better to cite a handbook, than research papers.}

1278: for disjoint cylinders in $\RR^3$, we show the following:

1279: 

1280: \begin{lemma}\label{lem:cut}

1281:   For any parameter $r$, $1\leq r\leq \min\{m,n^{1/3}\}$,

1282:   the maximum number of incidences

1283:   of type 2 between $n$ points and $m$ cylinders in 3-space satisfies

1284:   the following recurrence:

1285: \begin{equation}\label{eq:inci}

1286:   I(n,m)=O(n+mr^2\beta(r)) + O(r^3\beta(r))\cdot

1287:   I\left(\frac{n}{r^3},\frac{m}{r}\right),

1288: \end{equation}

1289: for some slowly growing function $\beta(n)$, as above.

1290: \end{lemma}

1291: %

1292: \begin{proof}

1293: Let $\C$ be a set of $m$ cylinders, and $S$ be a set of $n$ points.

1294: Construct a $(1/r)$-cutting of the arrangement $\A(\C)$.  The

1295: cutting has $O(r^3\beta(r))$ relatively open pairwise disjoint

1296: cells, each crossed by at most $m/r$ cylinders

1297: and containing at most $n/r^3$ points of $S$~\cite{ceg-89}

1298: (see also~\cite[p.~271]{as-95}); the first property is by definition

1299: of $(1/r)$-cuttings, and the second is enforced by subdividing cells

1300: with too many points.  The number of incidences between

1301: points and cylinders {\em crossing} their cells is thus

1302: $$

1303:   O(r^3\beta(r))\cdot I\left(\frac{n}{r^3},\frac{m}{r}\right) .

1304: $$

1305: (Note that any incidence of type 2 remains an incidence of type 2 in

1306: the subproblem it is passed to.)

1307: 

1308: It remains to bound the number of incidences between the points of

1309: $S$ and the cylinders that {\em contain} their cells. Let $\tau$ be

1310: a (relatively open) lower-dimensional cell of the cutting.

1311: If ${\rm dim}(\tau)=2$ then we can assign any point $p$ in $\tau$

1312: to one of the two neighboring full-dimensional cells, and count

1313: all but at most one of the incidences with $p$ within that cell.

1314: Hence, this increases the count by at most $n$.

1315: 

1316: If ${\rm dim}(\tau)=0$, i.e., $\tau$ is a vertex of the cutting,

1317: then any cylinder containing $\tau$ must cross or define one of the

1318: full-dimensional cells adjacent to $\tau$. Since each cell has at

1319: most $O(1)$ vertices, it follows that the total number of such

1320: incidences is $O(r^3\beta(r))\cdot (m/r) = O(mr^2\beta(r))$.

1321: 

1322: Suppose then that ${\rm dim}(\tau)=1$, i.e., $\tau$ is an edge of

1323: the cutting. An immediate implication of Lemma~\ref{3int} is that

1324: only $O(1)$ cylinders can contain $\tau$, unless $\tau$ is a line,

1325: which can then be a generator of arbitrarily many cylinders.

1326: 

1327: Since we are only counting incidences of type 2, this implies that

1328: any straight-edge 1-dimensional cell $\tau$ of the cutting generates

1329: at most one such incidence with any cylinder that fully contains

1330: $\tau$.  Non-straight edges of the cutting are contained in only

1331: $O(1)$ cylinders, as just argued, and thus the points on such edges

1332: generate a total of only $O(n)$ incidences with the cylinders. 

1333: Thus the overall number of incidences in

1334: this subcase is only $O(n+r^3\beta(r))$. Since $r\le m$,

1335: this completes the proof of the lemma.

1336: \end{proof}

1337: 

1338: \begin{lemma}\label{lem:distinct-cylinders}

1339:   The number of incidences of type 2 between $n$ points and $m$

1340:   cylinders in $\RR^3$ is

1341: \begin{equation} \label{177}

1342: O\left(\left( m^{6/7}n^{5/7}+m+n \right) \beta(n)\right).

1343: %O\left( m^{6/7}n^{5/7} \beta(n)+m+n \right).

1344: \end{equation}

1345: \end{lemma}

1346: %

1347: \begin{proof}

1348: Let $\C$ be a set of $m$ cylinders, and $S$ be a set of $n$ points.

1349: We first derive an upper bound of $O(n^5+m)$

1350: on the number of incidences of type 2 between $\C$ and $S$. We represent

1351: the cylinders as points in a dual 5-space, so that each cylinder

1352: $C$ is mapped to a point $C^*$, whose coordinates are the five

1353: degrees of freedom of $C$ (four specifying its axis and the fifth

1354: specifying its radius). A point $q\in \reals^3$ is mapped to a

1355: surface $q^*$ in $\reals^5$, which is the locus of all points dual

1356: to cylinders that are incident to $q$. With an appropriate choice of

1357: parameters, each surface $q^*$ is semi-algebraic of constant description

1358: complexity. By definition, this duality preserves incidences.

1359: 

1360: After dualization, we have an incidence problem involving $m$ points

1361: and $n$ surfaces in $\reals^5$.  We construct the arrangement

1362: $\A$ of the $n$ dual surfaces, and bound the number of their

1363: incidences with the $m$ dual points as follows.  The arrangement

1364: $\A$ consists of $O(n^5)$ relatively open cells of dimensions

1365: $0,1,\ldots,5$. Let $\tau$ be a cell of $\A$. We may assume that

1366: ${\rm dim}(\tau)\le 4$, because no point in a full-dimensional cell

1367: can be incident to any surface.

1368: 

1369: If $\tau$ is a vertex, consider any surface $\varphi$ that passes

1370: through $\tau$. Then $\tau$ is a vertex of the arrangement

1371: restricted to $\varphi$, which is a $4$-dimensional arrangement with

1372: $O(n^4)$ vertices. This implies that the number of incidences at

1373: vertices of $\A$ is at most $n\cdot O(n^4) = O(n^5)$.

1374: 

1375: Let then $\tau$ be a cell of $\A$ of dimension $\ge 1$, and let $u$

1376: denote the number of surfaces that contain $\tau$. If $u\le 8$ then each

1377: point in $\tau$ (dual to a cylinder) has at most $O(1)$ incidences

1378: of this kind, for a total of $O(m)$.

1379: 

1380: Otherwise, $u\ge 9$. Since ${\rm dim}(\tau)\ge 1$, it contains

1381: infinitely many points dual to cylinders (not necessarily in $\C$).

1382: By Lemma~\ref{3int}, back in the primal 3-space, if three cylinders

1383: contain the same nine points, then the axes of at least two of

1384: them are parallel. Hence all $u$ points lie on one line or on two

1385: parallel lines, which are common generators of these pair of

1386: cylinders. In this case, all cylinders whose dual points lie in $\tau$

1387: contain these generator(s). But then, by definition, the incidences

1388: between these points and the cylinders of $\C$ whose dual points lie

1389: on $\tau$ are of type 1, and are therefore not counted at all by

1390: the current analysis. Since $\tau$ is a face of $\A$, no other point

1391: lies on any of these cylinders, so we may ignore them completely.

1392: 

1393: Hence, the overall number of incidences under consideration is $O(n^5+m)$.

1394: 

1395: If $m>n^5$, this bound is $O(m)$. If $m<n^{1/3}$,

1396: we apply Lemma \ref{lem:cut} with $r=m$, which then yields that

1397: each recursive subproblem has at most one cylinder, so each point in a

1398: subproblem generates at most one incidence, for a total of $O(n)$

1399: incidences. Hence, in this case (\ref{eq:inci}) implies that the

1400: number of incidences between $\C$ and $S$ is

1401: $O(n+m^3\beta(m)) = O(n\beta(n))$.

1402: 

1403: Otherwise we have $n^{1/3} \leq m \leq n^5$, so we can apply

1404: Lemma~\ref{lem:cut} with parameter $r=(n^5/m)^{1/14}$; observe that

1405: $1 \leq r \leq \min\{m, n^{1/3}\}$ in this case. Using the above bound

1406: for each of the subproblems in the recurrence, we obtain

1407: $I(n/r^3,m/r)=O((n/r^3)^5+m/r)$, and thus the total number of

1408: incidences of type 2 in this case is

1409: $$

1410:   O(n+mr^2\beta(r)) + O(r^3\beta(r))\cdot O\left(

1411:     \left(\frac{n}{r^3}\right)^5 + \frac{m}{r}\right) = O\left(

1412:     \frac{n^5}{r^{12}} + mr^2 \right)\beta(r).

1413: $$

1414: The choice $r=(n^5/m)^{1/14}$ yields the bound (\ref{177}).

1415: Combining this with the other cases, the bound in the lemma follows.

1416: \end{proof}

1417: 

1418: We are now in position to complete the proof of

1419: Theorem~\ref{thm:unit3}.

1420: 

1421: \begin{proofof}{Theorem~\ref{thm:unit3}}

1422: We now return to our original setup, where the cylinders in $\C$ may

1423: have multiplicities. We fix some parameter $\mu$ and consider,

1424: as above, all cylinders in $C_\mu$, and recall our choice of $z=14/45$.

1425: The case $\mu\ge n^z$ is taken care of by Lemma~\ref{107}, accounting

1426: for at most  $O(n^{107/45}\polylog(n))$ incidences.

1427: In fact, Lemma~\ref{107} takes

1428: care of all cylinders that contain at least $n^z$ points on

1429: their axes. Assume then that $\mu < n^z$, and consider only those cylinders

1430: in $\C_\mu$ containing fewer than $n^z$ points on their axes.

1431: By (\ref{cmu}), we have $c_{\mu,\le n^z} = O(n^2/\mu^2)$.

1432: Consequently, the number of incidences with the remaining cylinders

1433: in $C_\mu$, counted with multiplicity, but excluding multiple points

1434: on the same generator line, is

1435: %

1436: $$

1437:   O\left(\mu\beta(n)\cdot \left( \left(\frac{n^2}{\mu^2}\right)^{6/7}

1438:    \cdot   n^{5/7} + \frac{n^2}{\mu^2} + n \right) \right)

1439:   = O\left(\left(\frac{n^{17/7}}{\mu^{5/7}}

1440:     + \frac{n^2}{\mu} + n\mu \right) \beta(n)\right) .

1441: $$

1442: Summing over all $\mu\leq n^z$ (powers of 2), and adding the

1443: bound $O(n^{107/45}\polylog(n))=O(n^{2.378})$ from Lemma~\ref{107} on

1444: the other kinds of incidences, we get the desired overall bound of

1445: $O(n^{17/7}\beta(n))=O(n^{2.4286})$.

1446: \end{proofof}

1447: 

1448: \noindent

1449: {\bf Remark.}

1450: In a nutshell, the ``bottleneck'' in the analysis is the case where

1451: $\mu$ is small (say, a constant) and we count incidences of type 2.

1452: The rest of the analysis, involved as it is, just shows that all the

1453: other cases contribute fewer (in fact, much fewer) incidences. One

1454: could probably simplify some parts of the analysis, at the cost of

1455: weakening the other bounds, but we leave these parts as they are, in

1456: the hope that the bottleneck case could be improved, in which case

1457: these bounds might become the dominant ones.

1458: 

1459: 

1460: \section{Minimum-area triangles in 3-space}

1461: \label{sec:minimum3}

1462: 

1463: Place $n$ equally spaced

1464: points on the three parallel edges of a right prism whose base is an

1465: equilateral triangle, such that inter-point distances are small along

1466: each edge. This construction yields $\frac{2}{3}n^2-O(n)$ minimum-area triangles,

1467: a slight improvement over the lower bound construction in the plane.

1468: Here is yet another construction with the same constant $2/3$ in the leading

1469: term: Form a rhombus in the $xy$-plane from two equilateral triangles

1470: with a common side, extend it to a prism in 3-space, and place $n/3$

1471: equally spaced points on each of the lines passing through the vertices

1472: of the shorter diagonal of the rhombus, and $n/6$ equally spaced

1473: points on each of the two other lines, where again the

1474: inter-point distances along these lines are all equal and small.

1475: The number of minimum-area triangles is

1476: $$ 2\left(\frac{1}{3 \cdot 3} + \frac{4}{3 \cdot 6}\right)n^2 -O(n)=

1477: \frac{2}{3}n^2-O(n). $$

1478: The following theorem shows that this bound is optimal up to a

1479: constant factor. No quadratic upper bound has previously been known for

1480: minimum-area triangles in $\RR^3$.

1481: 

1482: \begin{theorem} \label{thm:min3}

1483:   The number of triangles of minimum (nonzero) area spanned by $n$ points in

1484:   $\RR^3$ is at most $n^2+O(n)$.

1485: %This estimate is optimal up to a constant factor.

1486: \end{theorem}

1487: %

1488: \begin{proof}

1489: Consider a set $S$ of $n$ points in $\RR^3$, and let $T$ be the set of

1490: triangles of minimum (nonzero) area spanned by $S$.

1491: Without loss of generality, assume the minimum area to be $1$.

1492: Similarly to the planar case, we assign each

1493: triangle in $T$ to one of its longest sides, and argue that at most a

1494: constant number of triangles are assigned to each segment spanned by

1495: $S$. This immediately implies an upper bound of $O(n^2)$ on the

1496: cardinality of $T$. To improve the main coefficient in this bound, we

1497: distinguish between {\em fat} and {\em thin} triangles.

1498: A triangle is called fat (resp., thin) if the length of the height

1499: corresponding to its longest side is at least (resp., less than)

1500: half of the length of the longest side.

1501: We show that the number $N_1$ of thin triangles of minimum area is

1502: at most $2{n \choose 2}=n^2-n$, and that

1503: the number $N_2$ of fat triangles of minimum area is only $O(n)$.

1504: 

1505: Consider a segment $ab$, with $a,b\in S$, and let $h=|ab|$.  Every

1506: point $c\in S\setminus \{a,b\}$ for which the triangle

1507: $\Delta{abc}$ has minimum (unit) area must lie on a bounded cylinder $C$

1508: with axis $ab$, radius $r=2/h$,

1509: and bases that lie on the planes $\pi_a$ and $\pi_b$, incident to

1510: $a$ and $b$, respectively, and orthogonal to $ab$.  In fact, if

1511: $\Delta abc$ is assigned to $ab$ (that is, $ab$ is the longest side), 

1512: then $c$ must lie on a smaller

1513: portion $C'$ of $C$, bounded by bases that intersect $ab$ at points at

1514: distance $h-\sqrt{h^2-r^2}$ from $a$ and $b$, respectively.

1515: Assume for

1516: convenience that $ab$ is vertical, $a$ is the origin and

1517: $b=(0,0,h)$. Since $ab$ is the longest side of $\Delta abc$, the

1518: side of the isosceles triangle with base $ab$ and height $r$ must

1519: be no larger than $h$, i.e., $\frac14 h^2+r^2 \le h^2$, or

1520: $r^2\le \frac34 h^2$.

1521: Notice that the triangle formed by any two points

1522: of $S$ lying on $C'$ with either $a$ or $b$ is non-degenerate.

1523: 

1524: We first derive a simple formula that relates the area of any

1525: (slanted) triangle to the area of its $xy$-projection. Consider a

1526: triangle $\Delta$ that is spanned by two vectors $u,v$, and let

1527: $\Delta_0$, $u_0$, and $v_0$ denote the $xy$-projections of $\Delta$,

1528: $u$, and $v$, respectively. Write (where $\kk$ denotes, as usual, the

1529: vector $(0,0,1)$)

1530: $$

1531: u=u_0 + x\kk \quad\text{and}\quad v=v_0 + y\kk ,

1532: $$

1533: and put $A= {\rm area}(\Delta)$, $A_0= {\rm area}(\Delta_0)$.

1534: Then

1535: $$

1536: A^2 = \frac14 \|u\times v\|^2 =

1537: \frac14\| (u_0 + x\kk)\times (v_0 + y\kk) \| =

1538: \frac14\left( \| u_0\times v_0 \|^2 + \|yu_0 - xv_0\|^2 \right) ,

1539: $$

1540: or

1541: \begin{equation} \label{eq:pa}

1542: A^2 = A_0^2 + \frac14 \|yu_0 - xv_0\|^2 .

1543: \end{equation}

1544: 

1545: \vspace{-1mm}

1546: \paragraph{An initial weaker bound.}

1547: We claim that at most 10 triangles are assigned to $ab$. Assume, to

1548: the contrary, that this number is at least 11.  Divide $C$ into two

1549: equal slices by a horizontal plane orthogonal to $ab$ through its

1550: midpoint. Since more than 10 points of $S$ lie on $C$, at least

1551: 6 of them must lie on the same slice $C_0$, say the bottom slice.

1552: It follows that two points, $c$ and $d$,

1553: lie in some sector $\Upsilon$ of $C_0$ making a dihedral angle

1554: $\alpha$ at $ab$ of at most $360^\circ/6=60^\circ$.

1555: An illustration is provided in Figure~\ref{fig:min3}.

1556: 

1557: \begin{figure}[htbp]

1558:   \begin{center}

1559: \epsfig{file=min3.eps,width=3in,clip=}

1560:  \caption{\small Charging scheme for minimum-area triangles in

1561:    3-space; (a) the cylinder $C$; (b) the projection on $\pi_a$;

1562:    $c'$ and $d'$ are the respective projections of $c$ and $d$.}

1563: \end{center}

1564: \label{fig:min3}

1565: \end{figure}

1566: 

1567: We may assume, without loss of generality, that

1568: $$

1569: c=(r,0,x)=c_0+x\kk \quad\text{and}\quad

1570: d=(r\cos\alpha,r\sin\alpha,y)=d_0+y\kk ,

1571: $$

1572: where $0\le\alpha\le 60^\circ$ and $0\le x,y\le h/2$.

1573: Write $A={\rm area}(\Delta{acd})$. Using (\ref{eq:pa}), we have

1574: %The squared area of $\Delta acd$ is

1575: %$$

1576: %\frac14 \left|\overrightarrow{ac}\times\overrightarrow{ad}\right|^2 =

1577: %\frac14 \left| c_0\times d_0 + (yc_0-xd_0)\times\kk \right|^2 =

1578: %$$

1579: $$

1580: A^2=

1581: \frac14 \left\| c_0\times d_0\right\|^2 + \frac14\left\|yc_0-xd_0\right\|^2 =

1582: \frac{r^4\sin^2\alpha}{4} + \frac{r^2}{4}\left( x^2+y^2-2xy\cos\alpha\right).

1583: $$

1584: The expression $x^2+y^2-2xy\cos\alpha$ is the squared length of the

1585: third side of the triangle with sides $x$, $y$, with the angle

1586: $\alpha \leq 60^\circ$ between them. Since $x,y \leq h/2$, we clearly

1587: have $x^2+y^2-2xy\cos\alpha \leq h^2/4$.

1588: Thus, recalling that $r^2\le \frac34h^2$ and that $h^2r^2 = 4$, we have

1589: $$

1590: A^2 \le \frac{r^4\sin^2\alpha}{4} + \frac{r^2}{4} \cdot \frac{h^2}{4}=

1591: \frac{r^2}{4} \left( r^2\sin^2\alpha+ \frac{h^2}{4} \right)

1592: \leq \frac{r^2 h^2}{4} \left( \frac{9}{16} + \frac{1}{4} \right)=

1593: \frac{13}{16} <1,

1594: $$

1595: which contradicts the minimality of the area of $\Delta abc$.

1596: Hence, at most 10 triangles are assigned to each segment

1597: spanned by $S$.

1598: This already implies that there are at most $5(n^2-n)$

1599: minimum-area triangles.

1600: 

1601: \vspace{-1mm}

1602: \paragraph{A better bound.}

1603: We now improve the constant of proportionality, using a more careful

1604: analysis, which distinguishes between the cases in which the

1605: minimum-area triangles charged to the segment $ab$ are thin or fat.

1606: 

1607: \smallskip

1608: \noindent{\bf (a)} $r< \frac12h$ (thin triangles).

1609: We claim that in this case at most two triangles can be assigned

1610: to $ab$. Indeed, suppose to the contrary that at least three triangles

1611: are assigned to $ab$, so their third vertices, $c,d,e \in S$ lie on

1612: $C'\subset C$. 

1613: Write the $z$-coordinates of $c,d,e$ as $z_1h$, $z_2h$, $z_3h$,

1614: respectively, and assume, without loss of generality, that

1615: $0< z_1\le z_2\le z_3< 1$, and $z_2\le 1/2$.  Consider the

1616: triangle $\Delta{acd}$, and let $A$ denote its area.

1617: As before, write, without loss of generality,

1618: $$

1619: c=(r,0,z_1h) \quad\text{and}\quad

1620: d=(r\cos\alpha,r\sin\alpha,z_2h) ,

1621: $$

1622: for some $0\le\alpha\le 180^\circ$.  Using (\ref{eq:pa}), we get

1623: $$

1624: A^2=\frac14 r^4\sin^2\alpha + \frac14 r^2h^2 (z_1^2+z_2^2-2z_1z_2\cos\alpha).

1625: $$

1626: Thus, recalling that $r< \frac12h$ and that $h^2r^2 = 4$, we get

1627: \begin{equation} \label{eq:A2}

1628: A^2 <

1629: \frac14 r^2h^2 \left( \frac14\sin^2\alpha +

1630: z_1^2+z_2^2-2z_1z_2\cos\alpha \right) =

1631: \frac14\sin^2\alpha + z_1^2+z_2^2-2z_1z_2\cos\alpha .

1632: \end{equation}

1633: Let us fix $z_1,z_2$ and vary only $\alpha$. Write

1634: $$

1635: f(\alpha) =

1636: \frac14\sin^2\alpha + z_1^2+z_2^2 - 2z_1z_2\cos\alpha ,

1637: \quad\text{and}\quad

1638: f'(\alpha) =

1639: \frac12\sin\alpha\cos\alpha + 2z_1z_2\sin\alpha .

1640: $$

1641: $f$ attains its maximum at the zero of its derivative, namely at

1642: $\alpha_0$ that satisfies

1643: $$

1644: \cos\alpha_0 = - 4z_1z_2 .

1645: $$

1646: (Note that since $z_1\le z_2\le\frac12$, we always have $4z_1z_2\le 1$.

1647: Also, at the other zero $\alpha=0$, $f$ attains its minimum

1648: $(z_1-z_2)^2$.)

1649: 

1650: Substituting $\alpha_0$ into (\ref{eq:A2}),

1651: and using $z_1\le z_2\le\frac12$, we get

1652: $$

1653: A^2 < \frac{1-16z_1^2z_2^2}{4} + z_1^2 + z_2^2 + 8z_1^2z_2^2 =

1654: \frac14 + z_1^2 + z_2^2 + 4z_1^2z_2^2 =

1655: \left(\frac12 + 2z_1^2\right) \left(\frac12 + 2z_2^2\right) \le 1 ,

1656: $$

1657: which contradicts the minimality of the area of $\Delta abc$

1658: (recall that $\Delta acd$ is non-degenerate).

1659: 

1660: We have thus shown that at most two thin triangles of minimum area can be

1661: assigned to any segment $ab$, so $N_1 \leq 2{n \choose 2}=n^2-n$.

1662: 

1663: 

1664: \smallskip

1665: \noindent{\bf (b)} $r\ge \frac12h$ (fat triangles).

1666: Recall that we always have $r\le \frac{\sqrt{3}}{2}h$. Multiplying

1667: these two inequalities by $h/2$, we get

1668: $$

1669: \frac{h^2}{4} \le 1 \le \frac{h^2\sqrt{3}}{4} ,

1670: \quad\text{or}\quad

1671: \frac{2}{3^{1/4}} \le h \le 2 .

1672: $$

1673: % Note that $r > 3/4$.

1674: Let $E$ denote the set of all segments $ab$ such that the minimum-area

1675: triangles charged to $ab$ are fat. Note that the length of each edge

1676: in $E$ is in the interval $[2/3^{1/4}, 2]$.

1677: 

1678: We next claim that, for any pair of points $p,q\in S$ with

1679: $|pq| < 1$, neither $p$ nor $q$ can be an endpoint of an edge in $E$.

1680: Indeed, suppose to the contrary that $p,q$ is such a pair and that $pa$

1681: is an edge of $E$, for some $a\in S$; by construction, $a\ne q$.

1682: Let $\Delta pab$ be a fat minimum-area triangle charged to $pa$.

1683: If $q$ is collinear with $pa$, then $\Delta pqb$ is a nondegenerate

1684: triangle of area strictly smaller than that of $\Delta pab$ (recall that

1685: $|pq|< 1 < |pa|$), a contradiction.

1686: If $q$ is not collinear with $pa$, $\Delta paq$ is a nondegenerate

1687: triangle of area $ \leq \frac{|pa| \cdot |pq|}{2} < \frac{2 \cdot 1}{2} =1$,

1688: again a contradiction.

1689: 

1690: Let $S' \subseteq S$ be the set obtained by repeatedly removing

1691: the points of $S$ whose nearest neighbor in $S$ is at distance

1692: smaller than $1$.  Clearly, the minimum inter-point distance in $S'$ is at

1693: least $1$, and the endpoints of each edge in $E$ lie in $S'$. This implies,

1694: via an easy packing argument, that the number of edges of $E$ incident

1695: to any fixed point in $S'$ (all of length at most $2$) is only $O(1)$.

1696: Hence $|E|=O(n)$. Since each edge in $E$ determines at most $10$

1697: minimum-area triangles, as shown in the first part of our proof,

1698: we conclude that $N_2=O(n)$, as claimed.

1699: 

1700: Hence there are at most $2{n\choose 2} + O(n) = n^2+O(n)$

1701: minimum-area triangles in total.

1702: \end{proof}

1703: 

1704: 

1705: \section{Maximum-area triangles in 3-space}

1706: \label{sec:maximum3}

1707: 

1708: \'Abrego and Fern\'andez-Merchant~\cite{af-02} showed that one can

1709: place $n$ points on the unit sphere in $\RR^3$ so that they determine

1710: $\Omega(n^{4/3})$ pairwise distances of $\sqrt{2}$

1711: (see also~\cite[p.~191]{pa-95} and~\cite[p.~261]{bmp-05}).

1712: % They also showed that one can obtain $\Omega(n^{4/3})$ congruent

1713: % copies of any given triangle $T$. By selecting $T$ as a right

1714: % angled isosceles triangle, we can also ensure that $T$ has maximum

1715: % area among all triangles, and thus obtain

1716: This implies the following result:

1717: %

1718: \begin{theorem}\label{thm:maximum3}

1719: For any integer $n$, there exists an $n$-element point set in $\RR^3$

1720: that spans $\Omega(n^{4/3})$ triangles of maximum area, all incident to

1721: a common point.

1722: \end{theorem}

1723: %

1724: \begin{proof}

1725: Denote the origin by $o$, and consider a unit sphere centered at

1726: $o$. The construction in~\cite{af-02} consists of a set

1727: $S=\{o\}\cup S_1\cup S_2$ of $n$

1728: points, where $S_1\cup S_2$ lies on the unit sphere,

1729: $|S_1|=\lfloor (n-1)/2\rfloor$, $|S_2|=\lceil (n-1)/2 \rceil$,

1730: and there are $\Omega(n^{4/3})$ pairs of orthogonal segments of the form

1731: $(os_i,os_j)$ with  $s_i \in S_1$ and $s_j \in S_2$.

1732: 

1733: Moreover, this construction can be realized in such a way that

1734: $S_1$ lies in a small neighborhood of $(1,0,0)$, and $S_2$ lies in a

1735: small neighborhood of $(0,1,0)$, say. The area of every right-angled

1736: isosceles triangle $\Delta os_is_j$ with $s_i\in S_1$ and $s_j\in

1737: S_2$ is $1/2$. All other triangles have

1738: smaller area: this is clear if at least two vertices of a triangle

1739: are from $S_1$ or from $S_2$; otherwise the area is given by

1740: $\frac{1}{2}\sin \alpha$, where $\alpha$ is the angle of the two

1741: sides incident to the origin, so the area is less than $1/2$ if

1742: these sides are not orthogonal.

1743: \end{proof}

1744: 

1745: \smallskip

1746: We next show that the construction in Theorem~\ref{thm:maximum3} is

1747: almost tight, in the sense that at most $O(n^{4/3+\eps})$ maximum-area

1748: triangles can be incident to any point of an $n$-element point

1749: set in $\RR^3$, for any $\eps>0$.

1750: 

1751: \begin{theorem}  \label{thm:max-3d}

1752:   The number of triangles of maximum area spanned by a set $S$ of $n$

1753:   points in $\RR^3$ and incident to a fixed point $a\in S$ is

1754:   $O(n^{4/3+\eps})$, for any $\eps>0$.

1755: \end{theorem}

1756: 

1757: Assume, without loss of generality, that the maximum area is $1$.

1758: Similarly to the proof of Theorem~\ref{thm:unit3}, we map

1759: maximum-area triangles to point-cylinder incidences. Specifically,

1760: if $\Delta{abc}$ is a maximum-area triangle spanned by a point set $S$,

1761: then every point of $S$ lies on, or in the interior of, the cylinder

1762: with axis $ab$ and radius $2/|ab|$ ($c$ itself lies on the cylinder).

1763: The following two lemmas give upper bounds on the number of

1764: point-cylinder incidences in this setting. First we prove a

1765: weaker bound (Lemma~\ref{lem:kst-cylinder}) which, combined with the

1766: partition technique, gives an almost tight bound

1767: (Lemma~\ref{lem:st-cylinder}).

1768: %\micha{Why is the bound tight?}\csaba{construction is given after the lemma.}

1769: %

1770: Our proof is somewhat reminiscent of an argument of Edelsbrunner and

1771: Sharir~\cite{es-91}, where it is shown that the number of point-sphere

1772: incidences between $n$ points and $m$ spheres in $\RR^3$ is

1773: $O(n^{2/3}m^{2/3}+n+m)$, provided that no point lies in the exterior

1774: of any sphere.

1775: 

1776: \begin{lemma}\label{lem:kst-cylinder}

1777:   Let $S$ be a set of $n$ points, and $\C$ a set of $m$ cylinders in

1778:   $\RR^3$, such that the axis of each cylinder passes through the

1779:   origin, and no point lies in the exterior of any cylinder. Then

1780:   the number of point-cylinder incidences is

1781:   $O(nm^{\frac{1+\eps}{2}}+m)$, for any $\eps>0$.

1782: \end{lemma}

1783: %

1784: \begin{proof}

1785:   Assume, without loss of generality, that the horizontal plane $h$

1786:   incident to the origin does not contain any point of $S$, and that

1787:   the points above $h$ participate in at least half of the

1788:   point-cylinder incidences. For simplicity, continue to denote by

1789:   $S$ the subset of the at most $n$ points lying above $h$.

1790:   Consider the 3-dimensional dual arrangement $(S^*,\C^*)$, where

1791:   the dual of a point $p\in \RR^3\setminus \{o\}$ is the cylinder

1792:   $p^*$ with axis $op$ and radius $2/|op|$; and the dual of a cylinder

1793:   $\gamma$ whose axis passes through the origin is a point $\gamma^*$

1794:   above $h$ that lies on the axis of $\gamma$ at distance

1795:   $2/{\rm radius}(\gamma)$ from the origin. Note that incidences

1796:   between points and cylinders are preserved, and that no point of

1797:   $\C^*$ lies in the exterior of any cylinder of $S^*$. It therefore

1798:   suffices to prove that the number of incidences between $S^*$ and 

1799:   $\C^*$ is

1800:   $I(\C^*,S^*)=O(nm^{\frac{1+\eps}{2}}+m)$.

1801: 

1802:   Consider the intersection $B$ of the interiors of all cylinders in

1803:   $S^*$. Since the interior of each cylinder is convex, $B$ is a convex

1804:   body homeomorphic to a ball, whose boundary is composed of patches of

1805:   cylinders. Faces, edges, and vertices of $B$ can be defined as

1806:   connected components of the intersections of one, two, and three

1807:   cylinders, respectively.

1808: %

1809: Each of the points of $\C^*$ that lie on faces of $\bd B$ contributes

1810: one incidence. Since all the cylinder axes pass through the origin, no

1811: edge of $\bd B$ can be straight, so it cannot be contained in any

1812: cylinder of $S^*$ other than the two defining it (recall

1813: Lemma~\ref{3int}). Hence the points of $\C^*$ that lie on faces

1814: or edges of $\bd B$ contribute at most $2m$ incidences.

1815: 

1816: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%5

1817: We are left with the task of bounding the

1818: number of point-cylinder incidences involving points at vertices 

1819: of $B$. Note that there may exist cylinders incident to a vertex $p$ 

1820: of $B$ and not containing any other points of $\bd B$ in the 

1821: vicinity of $p$. To account for such cylinders too, perturb

1822: the radii of each cylinder in $S^*$, so that each radius $r$ is 

1823: decreased to the radius $(1-\delta r)r$,

1824: for a sufficiently small $\delta>0$ (that is, the radii of larger 

1825: cylinders decrease by a larger factor).

1826: As a result, every cylinder incident to a vertex $p\in \bd B$ is

1827: replaced by a cylinder that defines a face in a sufficiently

1828: small neighborhood of $p$ (even though it is not incident to

1829: $p$ after this perturbation). The number of point-cylinder incidences 

1830: between $\C^*$ and the vertices of $\bd B$ is proportional to

1831: the number of vertices of the resulting $\bd B'$ after the perturbation. 

1832: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%5

1833:   By a result of Halperin and Sharir~\cite{hs-95}, the

1834:   complexity of a single cell in the arrangement of $n$ constant

1835:   degree algebraic surfaces in $\RR^3$ is $O(n^{2+\eps})$,

1836:   for any $\eps>0$.

1837:   % (see~\cite{b-03} for extensions and generalizations

1838:   % to higher dimensions).

1839:   Hence, we obtain an upper bound of

1840:   $I(S,\C)=O(m+n^{2+\eps})$, for any $\eps>0$.

1841: 

1842:   Partition $S$ into $\lceil n/\sqrt{m}\rceil$ subsets, each

1843:   containing at most $\sqrt{m}$ points. The preceding argument

1844:   implies that each subset $S'\subset S$ has at most

1845:   $I(S',\C)=O(m+(\sqrt{m})^{2+\eps}) = O(m^{1+\eps/2})$ incidences with

1846:   the cylinders of $\C$. Therefore, altogether there are at most

1847:   $\lceil n/\sqrt{m}\rceil \cdot O(m^{1+\eps/2})

1848:   =O(nm^{\frac{1+\eps}{2}}+m)$ incidences.

1849: \end{proof}

1850: 

1851: \begin{lemma}\label{lem:st-cylinder}

1852:   Let $S$ and $\C$ be as in the preceding lemma.

1853:   Then the number of point-cylinder

1854:   incidences is $O((n^{2/3}m^{2/3}+n+m)^{1+\eps})$, for any $\eps>0$.

1855: \end{lemma}

1856: %

1857: \begin{proof}

1858: If $m>n^2$, then Lemma~\ref{lem:kst-cylinder} gives an upper bound

1859: of $O(nm^{\frac{1+\eps}{2}}+m) = O(m^{1+\eps})$.

1860: We may therefore assume henceforth that $m\leq n^2$.

1861: 

1862:   For an integer $r\in \NN$, to be specified later, choose a random

1863:   sample $R\subset \C$ of $r$ cylinders, and let $B$ denote the

1864:   intersection of the interiors of the cylinders in $R$.

1865:   By~\cite{hs-95}, the combinatorial complexity of $B$ is

1866:   $O(r^{2+\eps})$, for any $\eps>0$. Hence, the convex body $B$ can

1867:   be partitioned into $O(r^{2+\eps})$ cells, each bounded by a

1868:   constant number of constant-degree algebraic surfaces.

1869:   (This can be done, e.g., by first partitioning $\bd B$ into

1870:   pseudo-trapezoidal cells, and then by

1871:   taking the convex hull of each cell on $\bd B$ with the origin.)

1872:   By the $\eps$-net theory (see, e.g.,~\cite[Chap.~10.3]{m-02}), with

1873:   constant probability, the interior of each cell intersects at most

1874:   $O(m \frac{\log r}{r}) = O(m/r^{1-\eps})$ cylinders of $\C$.

1875:   We may assume then that our sample $R$ has this property.

1876:   Similarly to the proof of

1877:   Lemma~\ref{lem:cut}, assign each point to a unique cell. Assign

1878:   every point in the interior of a cell $\sigma_i$ to $\sigma_i$;

1879:   assign every point on the boundary of several cells to the cell with

1880:   minimum index. Let $n_i$ denote the number of points assigned to

1881:   cell $\sigma_i$.

1882: 

1883:   Applying Lemma~\ref{lem:kst-cylinder} in each cell $\sigma_i$, we

1884:   get the upper bound

1885:   ${\displaystyle O\left(n_i

1886:     \left(\frac{m}{r^{1-\eps}}\right)^{\frac{1+\eps}{2}} +

1887:     \left(\frac{m}{r^{1-\eps}}\right)\right) }$

1888: on the number of incidences between points assigned

1889:   to $\sigma_i$ and cylinders intersecting the interior of $\sigma_i$.

1890:   Summing over all $O(r^{2+\eps})$ cells, we have

1891: %

1892:   $$

1893:   \sum_{i} O\left(n_i

1894:     \left(\frac{m}{r^{1-\eps}}\right)^{\frac{1+\eps}{2}} +

1895:     \left(\frac{m}{r^{1-\eps}}\right)\right) = O\left(n\left(

1896:       \frac{m}{r^{1-\eps}}\right)^{\frac{1+\eps}{2}} +

1897:     mr^{1+2\eps}\right) =

1898:   O\left(\frac{nm^{\frac{1+\eps}{2}}}{r^{\frac{1-\eps}{2}}} +

1899:     mr^{1+2\eps}\right)

1900:   $$

1901:   incidences of this kind. By choosing

1902:   $r=\min\left\{\lfloor n^{2/3}/m^{1/3}\rfloor,\,m\right\}$, 

1903:   this is at most

1904:   $O(n^{2/3+\eps'}m^{2/3+\eps'}+n^{1+\eps'})$, for another, still

1905: arbitrarily small, $\eps'>0$.  Finally, the number of incidences

1906:   between points assigned to one cell and cylinders that do not

1907:   intersect the interior of that cell can be bounded similarly to the

1908:   proof of Lemma~\ref{lem:cut}: This number is proportional to the

1909:   number of cells plus the number of points, which is

1910:   $O(n+r^{2+\eps}) = O(n^{1+\eps})$, as is easily checked.

1911: (In this final argument, we use the fact all axes pass through the

1912: origin, so no 1-dimensional edge of $\bd B$ can be contained in more

1913: than two cylinders; see also the proof of Lemma~\ref{lem:kst-cylinder}.)

1914: \end{proof}

1915: 

1916: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%5

1917: The upper bound of Lemma~\ref{lem:st-cylinder} is almost tight: For any 

1918: $n$ and $m$, there are $n$ points and $m$ cylinders with axes through 

1919: the origin and containing no points in their exterior, which determine 

1920: $\Omega(n^{2/3}m^{2/3}+n+m)$ point-cylinder incidences. To construct

1921: such a configuration, take $n$ points and $m$ lines on the plane 

1922: $\pi:z=1$ in $\RR^3$ with $\Omega(n^{2/3}m^{2/3}+n+m)$ 

1923: point-line incidences~\cite{st-83}. Project these points and lines centrally

1924: from the origin onto the unit sphere, to obtain a system of $n$ points 

1925: and $m$ great circles with the same number of incidences.

1926: Each great circle of the unit sphere lies in a unique cylinder of unit 

1927: radius whose axis passes through the origin, and every such cylinder 

1928: contains all the other points of the unit sphere in its interior. 

1929: This gives $n$ points on the unit sphere and $m$ cylinders of 

1930: unit radius whose axes pass through the origin

1931: (so that no point lies in the exterior of any cylinder),

1932: with $\Omega(n^{2/3}m^{2/3}+n+m)$ point-cylinder incidences.

1933: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%5

1934: 

1935: \begin{proofof}{Theorem~\ref{thm:max-3d}}

1936: Let $A$ denote the maximum triangle area determined by a set $S$ of

1937: $n$ points in $\RR^3$. For every point $a\in S$, consider the system

1938: of $n-1$ points in $S\setminus \{a\}$ and $n-1$ cylinders, each defined

1939: by a point $b\in S\setminus \{a\}$, and has axis $ab$ and radius

1940: $2A/|ab|$.

1941: Every point-cylinder incidence corresponds to a triangle of area $A$

1942: spanned by $S$ and incident to $a$. Since $A$ is the maximum area, no

1943: point of $S$ may lie in the exterior of any cylinder.  By

1944: Lemma~\ref{lem:st-cylinder}, the number of such triangles is

1945: $O(n^{4/3+\eps})$, for any $\eps>0$.

1946: \end{proofof}

1947: 

1948: Theorems \ref{thm:maximum3} and \ref{thm:max-3d} imply the

1949: following bounds on the number of maximum-area triangles in $\RR^3$:

1950: 

1951: \begin{theorem}\label{thm:allmax-3d}

1952: The number of triangles of maximum area spanned by $n$ points in

1953: $\RR^3$ is $O(n^{7/3+\eps})$, for any $\eps>0$.

1954: For all $n \geq 3$, there exist $n$-element point sets in $\RR^3$

1955: that span $\Omega(n^{4/3})$ triangles of maximum area.

1956: \end{theorem}

1957: 

1958: 

1959: \section{Distinct triangle areas in 3-space}\label{sec:dist3d}

1960: 

1961: Following earlier work by Erd\H{o}s and Purdy~\cite{ep-76}, Burton and

1962: Purdy~\cite{bp-79}, and Dumitrescu and T\'oth~\cite{dt-07a},

1963: Pinchasi~\cite{p-07} has recently proved that $n$ noncollinear points in

1964: the plane always determine at least $\left\lfloor

1965:   \frac{n-1}{2}\right\rfloor$ distinct triangle areas, which is

1966: attained by $n$ equally spaced points distributed evenly on two

1967: parallel lines. No linear lower bound is known in 3-space, and

1968: the best we can show is the following:

1969: 

1970: \begin{theorem}\label{thm:dist3d}

1971:   Any set $S$ of $n$ points in $\RR^3$, not all on a line, determines at

1972:   least $\Omega(n^{2/3}/\beta(n))$ triangles of distinct areas, for

1973:   some extremely slowly growing function $\beta(n)$.

1974:   Moreover, all these triangles share a common side.

1975: \end{theorem}

1976: 

1977: For the proof, we first derive a new upper bound

1978: (Lemma~\ref{lem:clarkson}) on the number of point-cylinder

1979: incidences in $\RR^3$, for the special case where the axes of the

1980: cylinders pass through the origin (but without the additional requirement

1981: that no point lies outside any cylinder).

1982: %

1983: Consider a set $\C$ of $m$ such cylinders.

1984: These cylinders have only three degrees of

1985: freedom, and we can dualize them to points in 3-space. We use a

1986: duality similar to that used in the proof of Lemma~\ref{lem:kst-cylinder}.

1987: Specifically, we fix some generic halfspace $H$ whose bounding plane

1988: passes through the origin, say, the halfspace $z>0$. We then map

1989: each cylinder with axis $\ell$ and radius $\varrho$ to the point on

1990: $\ell\cap H$ at distance $1/\varrho$ from the origin;

1991: and we map each point $p\in H$ to the cylinder whose axis is the line

1992: spanned by $op$ and whose radius is $1/|op|$.

1993: As argued above, this duality preserves point-cylinder incidences.

1994: 

1995: By (a dual version of) Lemma~\ref{3int}, any three points

1996: can be mutually incident to at most eight cylinders whose axes

1997: pass through the origin. That is, the

1998: bipartite incidence graph (whose two classes of vertices correspond to

1999: the points of $S$ and the cylinders of $\C$, respectively, and an edge

2000: represents a point-cylinder incidence) is $K_{3,9}$-free. It follows

2001: from the theorem of K\H{o}v\'ari, S\'os and Tur\'an~\cite{kst-54} (see

2002: also~\cite[p.~121]{pa-95})

2003: that the number of point-cylinder incidences is $O(nm^{2/3}+m)$.

2004: We then combine this bound with the partition technique of

2005: Clarkson~\etal~\cite{ceg-90}, to prove a sharper upper bound on

2006: the number of point-cylinder incidences of this kind. Specifically, we

2007: have:

2008: 

2009: \begin{lemma}\label{lem:clarkson}

2010:   Given $n$ points and $m$ cylinders, whose axes pass through the

2011:   origin, in 3-space, the number of point-cylinder incidences is

2012:   $O(n^{3/4}m^{3/4}\beta(n)+n+m).$

2013: \end{lemma}

2014: %

2015: \begin{proof}

2016: Let $\C$ be the set of the $m$ given cylinders, and $S$ be the set of

2017: the $n$ given points.

2018: Let $h$ be a plane containing the origin, but no point of $S$, and

2019: assume, without loss of generality, that the subset $S'$ of points

2020: lying in the positive hafspace $h^+$ contributes at least half of

2021: the incidences with $\C$.  If $m>n^3$, then the

2022: K\H{o}v\'ari-S\'os-Tur\'an Theorem yields an upper bound of

2023: $I(S',\C)=O(nm^{2/3}+m)=O(m)$. Similarly, if $m<n^{1/3}$,

2024: the duality mentioned above leads to the bound

2025: $I(S',\C)=O(mn^{2/3}+n)=O(n)$. For these two cases we have

2026: then $I(S,\C) \leq 2I(S',\C)=O(m+n)$.

2027: Assume henceforth that $n^{1/3}\leq m\leq n^3$.

2028: 

2029: We apply Lemma~\ref{lem:cut} with parameter 

2030: $r= \lfloor n^{3/8}/m^{1/8} \rfloor$, and

2031: use the K\H{o}v\'ari-S\'os-Tur\'an Theorem to bound the number of

2032: incidences between the at most $n/r^3$ points and $m/r$ cylinders

2033: in each subproblem.

2034: Note that $1 \leq r \leq m$ in the above range of $m$.

2035: The total number of incidences is thus

2036: %

2037: \begin{eqnarray*}

2038: I(S,\C) &=& O(n+mr^2\beta(r)) + O(r^3\beta(r))\cdot

2039: O \left( \frac{n}{r^3} \cdot \left(\frac{m}{r}\right)^{2/3}

2040: +\frac{m}{r} \right)\\

2041: &=&

2042: O \left( n+ \frac{m^{2/3}n}{r^{2/3}}\beta(n) + m r^2\beta(r)\right) =

2043: O\left(n+ n^{3/4}m^{3/4}\beta(n)\right).

2044: \end{eqnarray*}

2045: Putting all three cases together gives the bound in the theorem.

2046: \end{proof}

2047: 

2048: 

2049: \begin{proofof}{Theorem~\ref{thm:dist3d}}

2050:   If there are $n/100$ points in a plane but not all on a line, then

2051:   the points in this plane already determine $\Omega(n)$ triangles of

2052:   distinct areas~\cite{bp-79}. We thus assume, in the remainder of the

2053:   proof, that there are at most $n/100$ points on any plane.

2054: 

2055:   According to a result of Beck~\cite{b-83}, there is an absolute

2056:   constant $k\in \NN$ such that if no line is incident to $n/100$ points

2057:   of $S$, then $S$ spans $\Theta(n^2)$ distinct lines, each of which

2058:   is incident to at most $k$ points of $S$. Since each point of $S$ is

2059:   incident to at most $n-1$ of these lines, there is a point $a\in S$

2060:   incident to $\Theta(n)$ such lines. Select a point of $S\setminus \{a\}$

2061:   on each of these lines, to obtain a set $P$ of $\Theta(n)$ points.

2062: 

2063:   Let $t$ denote the number of distinct triangle areas determined by

2064:   $S$, and let $\alpha_1,\alpha_2,\ldots ,\alpha_t$ denote these

2065:   areas.  For each point $b\in P$ and $i=1,2,\ldots ,t$, we define a

2066:   cylinder $C(ab,\alpha_i)$ with axis (the line spanned by) $ab$ and radius

2067:   $2\alpha_i/|ab|$.

2068:   Every point $c\in S$ for which the area of the

2069:   triangle $\Delta{abc}$ is $\alpha_i$ must lie on the cylinder

2070:   $C(ab,\alpha_i)$. Let $\C$ denote the set of the $O(nt)$ cylinders

2071:   $C(ab,\alpha_i)$, for $b\in P$ and $i=1,2,\ldots,t$.

2072:   For each point $b\in P$, there are

2073:   $n-k=\Theta(n)$ points off the line through $ab$, each of which must

2074:   lie on a cylinder $C(ab,\alpha_i)$ for some $i=1,2,\ldots , t$.

2075:   Therefore, the number $I(S,\C)$ of point-cylinder incidences between

2076:   $S$ and $\C$ is $\Omega(n^2)$. On the other hand, by

2077:   Lemma~\ref{lem:clarkson}, we have

2078: %

2079: $$\Omega(n^2)\leq I(S,\C)\leq O(n^{3/4}(nt)^{3/4}\beta(n)+n+nt)

2080: = O(n^{3/2}t^{3/4}\beta(n)),$$

2081: %

2082: which gives $t=\Omega (n^{2/3}/\beta^{4/3}(n))=\Omega

2083: (n^{2/3}/\beta'(n))$, for another function $\beta'(n)$ of the same

2084: slowly growing type, as required.

2085: \end{proofof}

2086: 

2087: \section{Conclusion} \label{sec:conclusion}

2088: 

2089: We have presented many results on the number of triangles of specific areas

2090: determined by $n$ points in the plane or in three dimensions. Our results improve

2091: upon the previous bounds, but, most likely, many of them are not

2092: asymptotically tight. This leaves many open problems of closing the

2093: respective gaps. 

2094: Even in cases where the bounds are asymptotically tight, such as

2095: those involving minimum-area triangles in two and three dimensions,

2096: determining the correct constants of proportionality still offers

2097: challenges.

2098: 

2099: Here is yet another problem on triangle areas, of a slightly

2100: different kind, with triangles determined by lines, not points

2101: (motivated in fact by the question of bounding $|U_2|$ in the

2102: proof of Theorem \ref{thm:unit2}). Any three nonconcurrent, and

2103: pairwise non-parallel lines in the plane determine a triangle of

2104: positive area. What is the maximum number of unit area

2105: triangles determined by $n$ lines in the plane?

2106: 

2107: 

2108: \begin{theorem}\label{thm:lines}

2109: The maximum number of unit-area triangles determined by $n$ lines in 

2110: the plane is $O(n^{7/3})$, and for any $n \geq 3$, there are $n$ lines that

2111: determine $\Omega(n^2)$ unit-area triangles. 

2112: \end{theorem}

2113: %

2114: \begin{proof}

2115: {\em Lower bound}: Place $n/3$ equidistant parallel lines at angles 

2116: $0$, $\pi/3$, and $2\pi/3$, through the points of an appropriate section of 

2117: the triangular lattice, and observe that

2118: there are $\Omega(n^2)$ equilateral triangles of unit side (i.e., of

2119: the same area) in this construction.

2120: 

2121: {\em Upper bound}: Let $L$ be a set of $n$ lines in the plane. 

2122: We define a variant

2123: of the hyperbolas used in the proof of Theorem \ref{thm:unit2}:

2124: For any pair of non-parallel lines $\ell_1,\ell_2\in L$, let

2125: $\gamma(\ell_1,\ell_2)$ denote the locus of points $p\in \RR^2$,

2126: $p\not\in\ell_1\cup \ell_2$, such that the parallelogram that has a

2127: vertex at $p$ and two sides along $\ell_1$ and $\ell_2$,

2128: respectively, has area $1/2$. The set $\gamma(\ell_1,\ell_2)$ is the

2129: union of two hyperbolas with $\ell_1$ and $\ell_2$ as asymptotes 

2130: (four connected branches in total).

2131: Any two non-parallel lines uniquely determine two such hyperbolas.

2132: Let $\Gamma$ denote the set of the branches of these hyperbolas, and

2133: note that $|\Gamma|=O(n^2)$.

2134: Observe now that, if $\ell_1$, $\ell_2$, and $\ell_3$ determine a

2135: unit area triangle, then $\ell_3$ is tangent to one of the two

2136: hyperbolas in $\gamma(\ell_1,\ell_2)$.

2137: 

2138: We first derive a weaker bound. Construct two bipartite graphs 

2139: $G_1,G_2\subseteq L\times\Gamma$. We put an edge $(\ell,\gamma)$ 

2140: in $G_1$ (resp., $G_2$) if $\ell$ is tangent to $\gamma$ and 

2141: $\ell$ lies below (resp., above) $\gamma$. The edges of $G_1$ 

2142: and $G_2$ account for all line-curve tangencies. 

2143: Observe that neither graph contains a $K_{5,2}$, 

2144: that is, there cannot be five distinct lines in $L$ tangent 

2145: to two branches of hyperbolas from above (or from below). 

2146: Indeed, this would force the two branches to intersect at five

2147: points, which is impossible for a pair of distinct quadrics.

2148: It thus follows from the theorem of 

2149: K\H{o}v\'ari, S\'os and Tur\'an~\cite{kst-54} (see 

2150: also~\cite[p.~121]{pa-95}) that the number of line-hyperbola tangencies 

2151: between any $n_0$ lines in $L$ and any $m_0$ hyperbolas in $\Gamma$ 

2152: is $O(n_0 m_0^{4/5}+m_0)$. With $n_0=n$ and $m_0=O(n^2)$, this already

2153: gives a bound of $O(n \cdot n^{8/5}+n^2)=O(n^{13/5})$ on the number of

2154: unit-area triangles determined by $n$ lines in the plane. We next

2155: derive an improved bound.  

2156: 

2157: Let $L$ be the given set of $n$ lines, and let $\Gamma$ be the

2158: corresponding set of $m=O(n^2)$ hyperbola branches. 

2159: We can assume that no line in $L$ is vertical, and apply a standard

2160: duality which maps each line $\ell\in L$ to a point $\ell^*$. A

2161: hyperbolic branch $\gamma$ is then mapped to a curve $\gamma^*$,

2162: which is the locus of all points dual to lines tangent to $\gamma$;

2163: it is easily checked that each $\gamma^*$ is a quadric.

2164: Let $L^*$ denote the set of the $n$ dual points, and let

2165: $\Gamma^*$ denote the set of $m=O(n^2)$ dual curves. 

2166: A line-hyperbola tangency in the primal plane is then

2167: mapped to a point-curve incidence in the dual plane. 

2168: 

2169: We next construct a $(1/r)$-cutting for $\Gamma^*$, partitioning

2170: the plane into $O(r^2)$ relatively open cells of bounded description

2171: complexity, each of which contains at most $n/r^2$ points and is

2172: crossed by at most $m/r$ curves. By using the previous bound for each

2173: cell, the total number of incidences involving points in the interior

2174: of these cells is

2175: % 

2176: $$ O\left( 

2177: r^2 \left(\frac{n}{r^2} \left(\frac{m}{r}\right)^{4/5} + \frac{m}{r}\right) 

2178: \right)= O\left(n \left(\frac{m}{r}\right)^{4/5} + mr\right).

2179: $$ 

2180: We balance the two terms by setting $r=n^{5/9}/m^{1/9}$, and observe

2181: that $1\le r\le m$ if $m\le n^5$ and $n\le m^2$; since

2182: $m=\Theta(n^2)$, both inequalities do hold in our case. Hence, the

2183: total number of incidences under consideration is

2184: $O(m^{8/9}n^{5/9})=O(n^{7/3})$.

2185: 

2186: It remains to bound the overall number of incidences involving points 

2187: lying on the boundaries of at least two cells. A standard

2188: argument, which we omit, shows that the number of these incidences

2189: is also $O(n^{7/3})$, and thereby completes the proof of the theorem.

2190: \end{proof}

2191: 

2192: Some remarks are in order: The line variant of unit-area triangle 

2193: problems is {\em not} equivalent to the point variant, under the 

2194: standard point-line duality.  Specifically: Let $S$ be a set of $n$ 

2195: points in the plane having distinct $x$-coordinates. Consider the 

2196: duality transform that maps a point

2197: $p=(a,b)$ to the line $p^*:\, y=ax-b$, and vice versa.

2198: %

2199: It is easy to see that

2200: there is no absolute constant $A>0$ such that, for $p,q,r \in S$,

2201: triangle $\Delta{pqr}$ has unit area if and only if the triangle

2202: $\Delta {p^*q^*r^*}$ formed by the three dual lines has area $A$.

2203: %

2204: 

2205: Yet, there is a connection between the point- and the line-variants of

2206: the unit-area problem in the plane. 

2207: Go back to the notation in the proof of Theorem \ref{thm:unit2},

2208: where, for a parameter $k \leq n^{1/3}$, we had $|U_1|= O(n^2 k)$.

2209: Recall that $U_2$ denotes the set of unit-area triangles where

2210: all three top lines are {\em $k$-rich}, and that there are 

2211: $|L_k|=O(n^2/k^3)$ such lines. Observe that the three top lines 

2212: of each triangle in $U_2$ determine a triangle of area $4$. We thus 

2213: face the question of bounding the number of triangles of area 4 

2214: determined by the $k$-rich lines in $L_k$. 

2215: By Theorem~\ref{thm:lines}, there are most $O((n^{2}/k^{3})^{7/3})$

2216: such triangles. Balancing $|U_1|$ with $|U_2|$ yields

2217: $k=n^{1/3}$, thereby implying that $|U_1|+|U_2| = O(n^{7/3})$.

2218: 

2219: We note that the bound $O(n^{44/19})$ of Theorem \ref{thm:unit2}

2220: could be re-derived with this new approach, if the bound of

2221: Theorem~\ref{thm:lines} could be improved to $O(n^{11/5})$.

2222: Moreover, an $o(n^{11/5})$ bound for the line-variant would in turn

2223: lead to an improvement in our current bound for the classical

2224: point-variant of the unit area problem in the plane.

2225: 

2226: 

2227: \begin{thebibliography}{10}\topsep=0pt\itemsep=-1pt\parsep=1pt

2228: 

2229: \bibitem{af-02}

2230: B.~M.~\'Abrego and S.~Fern\'andez-Merchant, Convex polyhedra in

2231: $\RR^3$ spanning $\Omega(n^{4/3})$ congruent triangles, {\em J.

2232: Combinat. Theory, Ser.~A} {\bf 98} (2002), 406--409.

2233: 

2234: \bibitem{at-07}

2235: E. Ackerman and G. Tardos,

2236: On the maximum number of edges in quasi-planar graphs,

2237: {\em J. Combinat. Theory, Ser. A} {\bf 114} (2007), 563--571.

2238: 

2239: \bibitem{aapps-97}

2240: P.~K. Agarwal, B. Aronov, J. Pach, R. Pollack, and M. Sharir,

2241: Quasi-planar graphs have a linear number of edges,

2242: {\em Combinatorica} {\bf 17} (1997), 1--9.

2243: 

2244: \bibitem{ANPPSS}

2245: P.~K.~Agarwal, E.~Nevo, J.~Pach, R.~Pinchasi, M.~Sharir, and S.~Smorodinsky,

2246: Lenses in arrangements of pseudocircles and their applications,

2247: {\it J. ACM} {\bf 51} (2004), 139--186.

2248: 

2249: \bibitem {acns-82}

2250: M.~Ajtai, V.~Chv\'atal, M.~Newborn, and E.~Szemer\'edi,

2251: Crossing-free subgraphs,

2252: {\em Annals Discrete Math.} {\bf 12} (1982), 9--12.

2253: 

2254: %\bibitem{as-05}

2255: %R.~Apfelbaum and M.~Sharir, Repeated angles in three and four

2256: %dimensions, {\em SIAM J. Discrete Math.} {\bf 19} (2005), 294--300.

2257: 

2258: %\bibitem{aks-05}

2259: %B.~Aronov, V.~Koltun, and M.~Sharir, Incidences between points and

2260: %circles in three and higher dimensions, {\em Discrete Comput. Geom.}

2261: %{\bf 33} (2005), 185--206.

2262: 

2263: \bibitem{ArS}

2264: B.~Aronov and M.~Sharir,

2265: Cutting circles into pseudo-segments and improved bounds for incidences,

2266: \textit{Discrete Comput. Geom.} {\bf 28} (2002), 475--490.

2267: 

2268: \bibitem{b-03}

2269: S.~Basu,

2270: The combinatorial and topological complexity of a single cell,

2271: {\em Discrete Comput. Geom.} {\bf 29} (2003), 41--59.

2272: 

2273: \bibitem{Baumann}

2274: E.~Baumann,

2275: Intersection of cylinders,

2276: {\tt http://private.mcnet.ch/baumann/all\_e.htm}

2277: 

2278: \bibitem{b-83}

2279: J.~Beck,

2280: On the lattice property of the plane and some problems of

2281: Dirac, Motzkin and Erd\H{o}s in combinatorial geometry,

2282: {\em Combinatorica} {\bf 3} (1983), 281--297.

2283: 

2284: %\bibitem{bk-03}

2285: %P.~Bra\ss\ and C.~Knauer, On counting point-hyperplane incidences,

2286: %{\em Comput. Geom.} {\bf 25} (2003), 13--20.

2287: 

2288: \bibitem{bmp-05}

2289: P.~Bra\ss , W.~Moser, and J.~Pach,

2290: {\em Research Problems in Discrete Geometry},

2291: Springer, New York, 2005.

2292: 

2293: \bibitem{brs-01}

2294: P.~Bra\ss, G.~Rote, and K.~J.~Swanepoel,

2295: Triangles of extremal area or perimeter in a finite planar point set,

2296: {\em Discrete Comput.  Geom.} ~{\bf 26} (2001), 51--58.

2297: 

2298: \bibitem{bp-79}

2299: G.~R.~Burton and G.~Purdy,

2300: The directions determined by $n$ points in the plane,

2301: {\em J.~London Math. Soc.} {\bf 20} (1979), 109--114.

2302: 

2303: \bibitem{ch-05}

2304: B.~Chazelle,

2305: Cuttings,

2306: In {\em Handbook of Data Structures and Applications}

2307: (D. Mehta and S. Sahni, editors), chap.~25, Chapman and Hall/CRC Press, 2005.

2308: 

2309: \bibitem{ceg-89}

2310: B.~Chazelle, H.~Edelsbrunner, L.~J.~Guibas, and M.~Sharir,

2311: A singly-expenential stratification scheme for real semi-algebraic

2312: varieties and its applications,

2313: in {\em Proc. 16th ICALP}, vol.~372 of

2314: LNCS, Springer, Berlin, 1989, pp.~179--193.

2315: 

2316: \bibitem{ceg-90}

2317: K.~L.~Clarkson, H.~Edelsbrunner, L.~G.~Guibas, M.~Sharir, and

2318: E.~Welzl, Combinatorial complexity bounds for arrangements of curves

2319: and spheres, {\em Discrete Comput. Geom.} {\bf 5} (1990), 99--160.

2320: 

2321: \bibitem{dpt-05}

2322: A.~Dumitrescu, J.~Pach, and G.~T\'oth,

2323: The maximum number of empty congruent triangles

2324: determined by a point set,

2325: {\em Revue Roumaine de Math. Pures et

2326: %{\em Revue Roumaine de Math\'ematiques Pures et

2327: Appliqu\'ees} {\bf 50} (2005), 613--618.

2328: 

2329: \bibitem {dt-07a}

2330: A. Dumitrescu and Cs.~D.~T\'oth,

2331: Distinct triangle areas in a planar point set,

2332: in {\em Proc. 12th Conf. on Integer Programming and

2333: Combinat. Opt.}, vol. 4513 of LNCS, Springer, 2007,

2334: pp.~119--129.

2335: 

2336: \bibitem{dt-07b}

2337: A. Dumitrescu and Cs. D. T\'oth,

2338: On the number of tetrahedra with minimum, uniform, and distinct volumes

2339: in 3-space,

2340: in {\em Proc. 18th ACM-SIAM Symp. Discrete Alg.}, 2007,

2341: ACM Press, pp.~1114--1123.

2342: 

2343: \bibitem{es-91}

2344: H.~Edelsbrunner ans M.~Sharir,

2345: A hyperplane incidence problem with applications to counting distances,

2346: in {\em Applied Geometry and Discrete Mathematics},

2347: vol.~4 of DIMACS Ser. Discrete Math. \& Theoret. Comput. Sci., AMS,

2348: Providence, 1991, pp.~253--263.

2349: 

2350: \bibitem {e-46}

2351: P. Erd\H{o}s,

2352: On sets of distances of $n$ points,

2353: {\em American Mathematical Monthly} {\bf 53} (1946), 248--250.

2354: 

2355: \bibitem{ep-71}

2356: P.~Erd\H{o}s and G.~Purdy,

2357: Some extremal problems in geometry,

2358: {\em J. Combinat. Theory} {\bf 10} (1971), 246--252.

2359: 

2360: %\bibitem{ep-75}

2361: %P.~Erd\H os and G.~Purdy, Some extremal problems in geometry III,

2362: %{\em Proc. 6th South-Eastern Conf. on Combinatorics, Graph Theory, and

2363: %Computing}, 1975, 291--308.

2364: 

2365: \bibitem{ep-76}

2366: P.~Erd\H os and G.~Purdy,

2367: Some extremal problems in geometry IV,

2368: {\em Congressus Numerantium} {\bf 17} (Proc. 7th South-Eastern Conf.

2369: on Combinatorics, Graph Theory, and Computing), 1976, 307--322.

2370: 

2371: %\bibitem{ep-77}

2372: %P.~Erd\H os and G.~Purdy, Some extremal problems in geometry V,

2373: %{\em Proc. 8th South-Eastern Conf. on Combinatorics, Graph Theory, and

2374: %Computing}, 1977, 569--578.

2375: 

2376: \bibitem{ep-95}

2377: P.~Erd\H os and G.~Purdy,

2378: Extremal problems in combinatorial geometry.

2379: in {\em Handbook of Combinatorics}

2380: (R. Graham, M. Gr\"otschel and L. Lov\'asz, editors),

2381: Vol. 1, 809--874, Elsevier, Amsterdam, 1995.

2382: 

2383: \bibitem{eps-82}

2384: P.~Erd\H os, G.~Purdy, and E.~G.~Straus, On a problem in

2385: combinatorial geometry, {\em Discrete Mathematics} {\bf 40} (1982),

2386: 45--52.

2387: 

2388: \bibitem{EPS:cyl}

2389: E.~Ezra, J.~Pach, and M.~Sharir,

2390: Regular vertices on the union of cylinders,

2391: manuscript, 2006.

2392: 

2393: \bibitem{hs-95}

2394: D.~Halperin and M.~Sharir,

2395: Almost tight upper bounds for the single cell and zone problems

2396: in three dimensions,

2397: {\em Discrete Comput. Geom.} {\bf 14} (1995), 385--410.

2398: 

2399: \bibitem{kpw-05}

2400: J.~K\'ara, A.~P\'or, and D.~R.~Wood,

2401: On the chromatic number of the visibility graph of a set of

2402: points in the plane,

2403: {\em Discrete Comput. Geom.} {\bf 34} (2005), 497--506.

2404: 

2405: \bibitem{kt-04}

2406: N. H. Katz and G. Tardos,

2407: A new entropy inequality for the Erd\H os distance problem,

2408: in {\em Towards a Theory of Geometric Graphs}

2409: (J. Pach, ed.), vol. 342 of Contemporary Mathematics, AMS, Providence,

2410: RI, 2004, 119--126.

2411: 

2412: %\bibitem{Ko}

2413: %V.~Koltun,

2414: %Almost tight upper bounds for vertical decomposition in four dimensions,

2415: %{\it J.~ACM} {\bf 51} (2004), 699--730.

2416: 

2417: \bibitem{kst-54}

2418: T.~K\H{o}vari, V.~T.~S\'os, and P.~Tur\'an,

2419: On a problem of K.~Zarankiewicz,

2420: {\em Colloquium Math.} {\bf 3} (1954), 50--57.

2421: 

2422: \bibitem{l-84}

2423: T. Leighton,

2424: New lower bound techniques for VLSI,

2425: {\em Math. Systems Theory} {\bf 17} (1984), 47--70.

2426: 

2427: \bibitem{MT}

2428: A.~Marcus and G.~Tardos,

2429: Intersection reverse sequences and geometric applications,

2430: {\em J. Combinat. Theory Ser. A} {\bf  113} (2006), 675--691.

2431: 

2432: \bibitem{m-02}

2433: J.~Matou\v{s}ek,

2434: {\em Lectures on Discrete Geometry},

2435: Springer-Verlag, Berlin, 2002.

2436: 

2437: \bibitem{pa-95}

2438: J.~Pach and P.~K.~Agarwal,

2439: {\em Combinatorial Geometry},

2440: Wiley-Interscience, New York, 1995.

2441: 

2442: \bibitem{ps-92}

2443: J.~Pach and M.~Sharir,

2444: Repeated angles in the plane and related problems,

2445: {\em J. Combinat. Theory Ser. A} {\bf 59} (1992), 12--22.

2446: 

2447: %\bibitem{ps-98} J.~Pach and M.~Sharir, On the number of

2448: %  incidences between points and curves, {\em Combin.  Probab.

2449: %    Comput.\/} {\bf 7} (1998), 121--127.

2450: 

2451: \bibitem{ps-04} J.~Pach and M.~Sharir, Geometric incidences, in

2452:   {\em Towards a Theory of Geometric Graphs\/}, AMS, vol. 342 of {\em

2453:     Contemp. Math.\/}, 2004, pp. 185--223.

2454: 

2455: \bibitem{p-07}

2456: R.~Pinchasi,

2457: The minimum number of distinct areas of

2458: triangles determined by a set of $n$ points in the plane,

2459: manuscript, 2007.

2460: 

2461: \bibitem{as-95}

2462: M.~Sharir and P.~K.~Agarwal,

2463: {\em Davenport-Schinzel sequences and their geometric applications},

2464: Cambridge University Press, New York, 1995.

2465: 

2466: \bibitem {st-01}

2467: J. Solymosi and Cs. D. T\'{o}th,

2468: Distinct distances in the plane,

2469: {\em Discrete Comput. Geom.} {\bf 25} (2001), 629--634.

2470: 

2471: \bibitem{sst-84}

2472: J. Spencer, E.~Szemer\'edi and W.~T.~Trotter,

2473: Unit distances in the Euclidean plane,

2474: in {\em Graph Theory and Combinatorics}, (B. Bollob\'as, editor),

2475: Academic Press, London, 1984, pp.~293--303.

2476: 

2477: \bibitem{s-97}

2478: L.~Sz\'ekely,

2479: Crossing numbers and hard {Erd\H os} problems in discrete geometry,

2480: {\em Combinat. Probab. Comput.} {\bf 6} (1997), 353--358.

2481: 

2482: \bibitem{st-83}

2483: E.~Szemer\'edi and W.~T.~Trotter,

2484: Extremal problems in discrete geometry,

2485: {\em Combinatorica} {\bf 3} (1983), 381--392.

2486: 

2487: \bibitem {t-03}

2488: G. Tardos,

2489: On distinct sums and distinct distances,

2490: {\it Advances in Math.} {\bf 180} (2003), 275--289.

2491: 

2492: \bibitem {v-54} I. M. Vinogradov,

2493: {\it Elements of Number Theory}, Dover Publications, New York, 1954.

2494: 

2495: 

2496: \end{thebibliography}

2497: 

2498: 

2499: \newpage

2500: \section*{Appendix}

2501: 

2502: \begin{proofof}{Lemma~\ref{3int}}

2503: Let us recall from~\cite{EPS:cyl} the structure of the

2504: intersection curve between two cylinders.  Let $C$ and $C'$ be two

2505: cylinders with nonparallel axes, so each pair of axes are either skew

2506: to each other or concurrent.  Let $\gamma$ denote the curve of their

2507: intersection.

2508: 

2509: To simplify the analysis, we assume, without loss of generality, that

2510: the axis $\alpha$ of $C$ is the $z$-axis and that its radius is $1$.

2511: Let $\alpha'$ and $\rho'$ denote respectively the axis and radius of

2512: $C'$. Let $\pi$ be the plane passing through $\alpha'$ and through the

2513: shortest segment $e$ connecting the axes $\alpha,\alpha'$. If

2514: $\alpha,\alpha'$ are skew lines, $e$ and $\pi$ are well defined.  If

2515: $\alpha$ and $\alpha'$ are concurrent, we take $\pi$ to be the plane

2516: passing through $\alpha'$ and orthogonal to the plane spanned by

2517: $\alpha$ and $\alpha'$.

2518: 

2519: Let $\sigma$ denote the ellipse $C\cap\pi$. We use a cylindrical

2520: coordinate system $\theta,z$ on $C$, and write the equation of

2521: $\sigma$ as $z=a\cos\theta+b\sin\theta+c$, where $z=ax+by+c$ is

2522: the quation of $\pi$.

2523: 

2524: As shown in \cite{EPS:cyl}, the equation of $\gamma$ is

2525: $$

2526: z=\sigma(\theta)\pm

2527:   \frac{1}{\sin\beta}\sqrt{(\rho')^2-d^2(\sigma(\theta),\alpha')} ,

2528: $$

2529: where $\beta$ is the angle between the axes. Moreover,

2530: $d(\sigma(\theta),\alpha')$, being the distance, within $\pi$,

2531: of a point on the ellipse $\sigma$ from the line $\alpha'$,

2532: can also be expressed as $|p\cos\theta+q\sin\theta+r|$, for

2533: appropriate parameters $p,q,r$.

2534: 

2535: Let now $C,C_1,C_2$ be three cylinders with no pair of parallel axes.

2536: Suppose to the contrary that $|C\cap C_1\cap C_2|\ge 9$.

2537: Let $\gamma_i$ denote the intersection curve $C\cap C_i$, for

2538: $i=1,2$. Write the equations of $\gamma_1,\gamma_2$ as

2539: $$

2540: z=a_i\cos\theta+b_i\sin\theta+c_i \pm

2541:   \frac{1}{\sin\beta_i}\sqrt{(\rho_i)^2-(p_i\cos\theta+q_i\sin\theta+r_i)^2},

2542: $$

2543: for $i=1,2$, with the appropriate parameters as above.

2544: We can re-parameterize these curves by putting $t=\tan(\theta/2)$

2545: and $w=z(1+t^2)$, to obtain two equations of the form

2546: \begin{eqnarray*}

2547: w & = & Q_1(t)\pm\sqrt{K_1(t)} \\

2548: w & = & Q_2(t)\pm\sqrt{K_2(t)} ,

2549: \end{eqnarray*}

2550: where $Q_1,Q_2$ are quadratic polynomials and $K_1,K_2$ are

2551: quartic polynomials. We are given that these two equations have

2552: at least $9$ common roots (it is easy to check that distinct roots of

2553: the original system are mapped to distinct roots of the new system).

2554: 

2555: If $Q_1(t)\equiv Q_2(t)$ then the common roots must satisfy

2556: $K_1(t)=K_2(t)$. Since there are at least $9$ such roots and this

2557: is a quartic equation, we must also have $K_1(t)\equiv K_2(t)$.

2558: 

2559: We will get to this case soon, but let us first consider the case

2560: $Q_1(t)\not\equiv Q_2(t)$. After squaring, the

2561: equations become

2562: \begin{eqnarray*}

2563: (w-Q_1(t))^2 & = & K_1(t) \\

2564: (w-Q_2(t))^2 & = & K_2(t) .

2565: \end{eqnarray*}

2566: Hence

2567: $$

2568: w = -\frac{K_2(t)-K_1(t)}{2(Q_2(t)-Q_1(t))} + \frac{Q_1(t)+Q_2(t)}{2} ,

2569: $$

2570: so $t$ must satisfy the equation

2571: \begin{equation} \label{deg8eq}

2572: \left(

2573:  -\frac{K_2(t)-K_1(t)}{2(Q_2(t)-Q_1(t))} + \frac{Q_2(t)-Q_1(t)}{2}

2574: \right)^2 = K_1(t) ,

2575: \end{equation}

2576: which is a polynomial equation of degree at most $8$. Since it has $9$

2577: roots, it must vanish identically.

2578: 

2579: Since the left-hand side of (\ref{deg8eq}) is a square, $K_1$ must

2580: also be a square.  However, $K_1(t)$ is proportional to

2581: $$

2582: \biggl(\rho_1(1+t^2)\biggr)^2 -

2583: \biggl(p_1(1-t^2)+2q_1t+r_1(1+t^2)\biggr)^2 =

2584: $$

2585: $$

2586: \biggl(\rho_1(1+t^2)-(p_1(1-t^2)+2q_1t+r_1(1+t^2))\biggr) \cdot

2587: \biggl(\rho_1(1+t^2)+(p_1(1-t^2)+2q_1t+r_1(1+t^2))\biggr) .

2588: $$

2589: It follows that either each of these factors is a square, or they

2590: are multiples of each other. In the former case, we must have

2591: \begin{eqnarray*}

2592: q_1^2 & = & (\rho_1+p_1-r_1)(\rho_1-p_1-r_1) = (\rho_1-r_1)^2-p_1^2 \\

2593: q_1^2 & = & (\rho_1-p_1+r_1)(\rho_1+p_1+r_1) = (\rho_1+r_1)^2-p_1^2 ,

2594: \end{eqnarray*}

2595: implying that $\rho_1-r_1 = \pm (\rho_1+r_1)$, so either $\rho_1=0$

2596: or $r_1=0$. The first equality is impossible---our cylinders have

2597: positive radii. The second equality implies that

2598: $\rho_1^2=p_1^2+q_1^2$. However, as argued in \cite{EPS:cyl}, by

2599: shifting $\theta$, we may assume that $q_1=0$ and $p_1$ is half the

2600: major axis of $\sigma_1$. This implies that $\sigma_1$ is a circle

2601: (since its minor axis is always equal to $2\rho_1$),

2602: which can happen only when $\alpha_1$ is orthogonal to $\alpha$.

2603: Moreover, $r_1=0$ implies that $\alpha$ and $\alpha'$ are concurrent.

2604: 

2605: In the latter case, since $\rho_1\ne 0$, the two factors are

2606: proportional to each other only when $p_1(1-t^2)+2q_1t$ is a

2607: multiple of $1+t^2$, which can only happen when $p_1=q_1=0$,

2608: which again is impossible.

2609: 

2610: Since the only remaining case is that of orthogonal concurrent axes,

2611: it follows, using a

2612: symmetric argument, that in the only remaining case, the three axes

2613: $\alpha,\alpha_1,\alpha_2$ are concurrent, at a common point, and

2614: mutually orthogonal. It is easily checked that in this case the cylinders

2615: can intersect in at most $8$ points, contrary to assumption.

2616: (This special case of three intersecting cylinders has been studied

2617: a lot; see, e.g.,~\cite{Baumann}.)

2618: 

2619: Hence, $Q_1(t)\equiv Q_2(t)$ and $K_1(t)\equiv K_2(t)$. However, the

2620: first identity implies that $\sigma_1=\sigma_2$,

2621: so the plane containing the axis of $C_1$

2622: also contains the axis of $C_2$. Since these axes are nonparallel,

2623: they must be concurrent. Since the analysis is fully symmetric with

2624: respect to the three cylinders, it follows that all three axes are

2625: either coplanar or concurrent. If they are coplanar but not concurrent,

2626: then it is easy to check that the planes $\pi_1$ and $\pi_2$ (with

2627: respect to $C$ as the ``base'' cylinder) cannot be equal. If the

2628: three axes are concurrent then again the identity of the planes

2629: $\pi_1,\pi_2$ implies that both $\alpha_1$ and $\alpha_2$ must be

2630: orthogonal to $\alpha$, and the fact that the argument is fully symmetric

2631: implies that all three axes must be concurrent and mutually orthogonal,

2632: a case that we have already ruled out.

2633: %

2634: This completes the proof of Lemma~\ref{3int}.

2635: \end{proofof}

2636: 

2637: 

2638: \end{document}

2639: 

2640: