1: \documentclass{article}

2: \usepackage{graphicx}

3: \usepackage{amssym}

4: 

5: \newcommand{\rbox}{\begin{flushright}

6:         \vspace{-8mm}

7:         \qed

8:         \vspace{-1mm}

9:         \end{flushright}

10: }

11: 

12: 

13: 

14: \oddsidemargin .25in

15: \setlength\topmargin{-.4in}

16: \setlength\textheight{8.7in}

17: \setlength\textwidth{6in}

18: 

19: 

20: 

21: 

22: \def\trp{{trapezoid}}

23: \def\Then{ ~\underline{Then}~}

24: \def\If{\underline{If}~}

25: \def\Begin{\underline{Begin}~}

26: \def\For{\underline{For}~}

27: \def\Repeat{\underline{Repeat}~}

28: \def\Do{\underline{Do}~}

29: \def\Return{\underline{Return}~}

30: \def\Else{\underline{Else}~}

31: \def\End{\underline{End}~}

32: \def\While{\underline{While}~}

33: \def\Loop{\underline{Loop}~}

34: \def\Until{\underline{Until}~}

35: 

36: 

37: 

38: \newcommand{\figlab}[1]{\label{fig:#1}}

39: \newcommand{\figref}[1]{Figure \ref{fig:#1}}

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

41:    \def\alon#1{{\sc Alon says: }{\marrow\sf #1}}

42: \def\eps{\varepsilon}

43: \def\B{{\cal B}}

44: \def\T{{\cal T}}

45: \def\D{{\cal D}}

46: \def\E{{\cal E}}

47: \def\Cov{\mbox{\sl Cov}}

48: \def\bt{\beta}

49: \def\al{\alpha}

50: \def\bd{{\partial}}

51: \def\R{{\cal R}}

52: \newcommand{\comm}[1]

53: {\marginpar{$\longleftarrow$}$\,\mbox{}^{\rm com}${\sf #1}$\mbox{}_{\rm ment}\,$}

54: \newcommand{\Fr}{Fr\'{e}chet\ }

55: \newcommand{\Frd}{Fr\'{e}chet distance\ }

56: \newcommand{\f}{\ensuremath{d_F}}

57: \newcommand{\wf}{\ensuremath{d_{\tilde{F}}}}

58: \newenvironment{proof}{\noindent{\bf Proof:}}{{\rbox}}

59: \newenvironment{sketch}{\noindent{\bf Proof Sketch:}}{{\rbox}}

60: \def\reals{{I\!\!R}}

61: \def\Reals{{\reals}}

62: 

63: \newtheorem{Def}[section]{Definition}

64: \newtheorem{Claim}{Claim}[section]

65: \newtheorem{Lemma}[Claim]{Lemma}

66: \newtheorem{theorem}{Theorem}\newtheorem{corollary}[Claim]{Corollary}

67: 

68: 

69: \newcommand{\euclidean}{{\Bbb E}}

70: \renewcommand{\reals}{{\Bbb R}}

71: \newcommand{\sphere}{{\Bbb S}}

72: \newcommand{\integers}{{\Bbb Z}}

73: \newcommand{\naturals}{{\Bbb N}}

74: \newcommand{\complex}{{\Bbb C}}

75: \newcommand{\rationals}{{\Bbb Q}}

76: \newcommand{\seg}[1]{\overline{#1}}

77: \def\TP{{\sl TP}\ }

78: \def\S{{\cal S}}

79: \def\maxs{{\sl max}(\sum}

80: \def\maxs{{\sl max\_}\sum}

81: \def\F{{\cal F}}

82: \def\epsilon{\varepsilon}

83: \def\Pos{{\sl Pos}}

84: \def\Neg{{\sl Neg}}

85: \long\gdef\boxit#1{\vspace{5mm}\begingroup\vbox{\hrule\hbox{\vrule\kern3pt

86: \vbox{\kern4pt#1\kern3pt}\kern3pt\vrule}\hrule}\endgroup}

87: \newcommand{\qed}{\vspace{.15cm}\rule{6pt}{6pt}}

88: 

89: 

90: \newcommand{\comment}[1]{}

91: 

92: \setcounter{section}{-1}

93: 

94: \title{Computing Homotopic Shortest Paths Efficiently

95: }

96: 

97: \author{

98: Alon Efrat\\

99: Department of Computer Science\\

100: University of Arizona\\

101: {\tt alon@cs.arizona.edu}

102: \and

103: Stephen G.~Kobourov\\

104: Department of Computer Science\\

105: University of Arizona\\

106: {\tt kobourov@cs.arizona.edu}

107: \and

108: Anna Lubiw\\

109: Department of Computer Science\\

110: University of Waterloo\\

111: {\tt alubiw@uwaterloo.ca}

112: \and

113: }

114: 

115: 

116: \begin{document}

117: \def\polylog{{{\sl polylog }}}

118: 

119: \maketitle

120: \begin{abstract} \small\baselineskip=9pt

121: 

122: This paper addresses the problem of finding shortest paths homotopic

123: to a given disjoint set of paths that wind amongst point obstacles in

124: the plane. We present a faster algorithm than previously known. 

125: 

126: \end{abstract}

127: 

128: 

129: 

130: 

131: 

132: 

133: 

134: 

135: 

136: 

137: 

138: 

139: 

140: 

141: 

142: 

143: 

144: 

145: \section{Introduction} \label{section:introduction}

146: 

147: 

148: Finding Euclidean shortest paths in simple polygons is a well-studied

149: problem. The funnel algorithm of Chazelle~\cite{FOCS82*339} and Lee

150: and Preparata~\cite{LP84} finds the shortest path between two points

151: in a simple polygon. Hershberger and Snoeyink~\cite{HerSno94} unify

152: earlier results for computing shortest paths in polygons. They

153: optimize a given path among obstacles in the plane under the Euclidean

154: and link metrics and under polygonal convex distance functions.

155: Related work has been done in addressing a classic VLSI problem, the

156: continuous homotopic routing problem~\cite{cs-rrew-84,lm-artr-85}. For

157: this problem it is required to route wires with fixed terminals among

158: fixed obstacles when a sketch of the wires is given, i.e., each wire

159: is given a specified homotopy class. If the wiring sketch is not given

160: or the terminals are not fixed, the problem is

161: NP-hard~\cite{Leiserson:1983:OPR,pinter83river-routing,Richards84b}.

162: 

163: Some of the early work on continuous homotopic routing was done by

164: Cole and Siegel~\cite{cs-rrew-84} and Leiserson and

165: Maley~\cite{lm-artr-85}. They show that in the $L_\infty$ norm a solution

166: can be found in $O(k^3 \log n)$ time and $O(k^3)$ space, where $n$ is

167: the number of wires and $k$ is the maximum of the input and output

168: complexities of the wiring. Maley~\cite{maley-thesis} shows how to

169: extend the distance metric to arbitrary polygonal distance functions

170: (including Euclidean distance) and presents a $O(k^4 \log n)$ time and

171: $O(k^4)$ space algorithm.  The best result so far is due to Gao {\em

172: et al.}~\cite{Storb88} who present a $O(kn^2 \log(kn))$ time and

173: $O(kn^2)$ space algorithm. Duncan {\em et al.}~\cite{dekw-dwfe01} and Efrat {\rm et al.}~\cite{eksw-gfg-02}

174: present an $O(kn+n^3)$ algorithm for the related fat edge drawing

175: problem: given a planar weighted graph $G$ with maximum degree 1 and

176: an embedding for $G$, find a planar drawing such that all the edges

177: are drawn as thick as possible and proportional to the corresponding

178: edge weights. 

179: 

180: 

181: The topological notion of homotopy formally captures the notion of

182: deforming paths. Let $\alpha,\beta:[0,1]\longrightarrow\Reals^2$ be two continuous

183: curves parameterized by arc-length.  Then $\alpha$ and $\beta$ are

184: {\it homotopic\/}

185: with respect to a set of obstacles $V\subseteq \reals^2$

186: if $\alpha$ can be continuously deformed into $\beta$ while avoiding the obstacles;

187: more formally,

188: if there exists a continuous

189: function $h:[0,1]\times [0,1]

190: \rightarrow \Reals^2$ with the following three properties:

191: \newpage

192: \begin{enumerate}

193: \item $h(0,t)=\alpha(t)$ and $h(1,t)=\beta(t)$, for $0\leq t \leq 1$

194: \item $h(\lambda,0)=\alpha(0)=\beta(0)$ and $h(\lambda,1)=\alpha(1)=\beta(1)$ for $0\leq

195: \lambda \leq 1$

196: \item $h(\lambda,t) \notin V$ for $0 \leq \lambda \leq 1, 0 < t<1$

197: \end{enumerate}

198: 

199: Let $\Pi=\{\pi_1, \pi_2, \dots, \pi_n\}$ be a set of disjoint, simple

200: polygonal paths and let the endpoints of the paths in $\Pi$ define the

201: set $T$ of at most $2n$ fixed points in the plane.

202: Note that we allow a path to degenerate to a single point.

203: We call the fixed points of

204: $T$ ``terminals,'' and call the interior vertices of the paths ``bends,''

205: and use ``points'' in a more generic sense, e.g.  ``a point in the

206: plane,'' or ``a point on a path.''

207: We assume that no two terminals/bends lie on the same vertical line.

208: 

209: Our goal is to replace each path $\pi_i\in \Pi$ by a shortest path $\sigma_i$

210: that

211: is homotopic to $\pi_i$ with respect to the set of obstacles $T$, see Fig.~\ref{fig-complexity} and Fig.~\ref{fig-monotone}.

212: Note that $\sigma_i$ is unique.  Let $\Sigma = \{\sigma_1, \ldots

213: \sigma_n\}$ be the set of resulting paths.

214: Observe

215: that these output paths may [self] intersect by way of segments

216: lying on top of each other, but will be

217: {\it non-crossing\/} in the sense that a slight perturbation of the bends

218: makes the paths simple and disjoint.

219: 

220: Let $k_{\it in}$ be the number of edges in all the paths of $\Pi$.

221: Let $k_{\it out}$ be the number of edges in all the paths of $\Sigma$.  We

222: will measure the complexity of our algorithms in terms of $n$ and $k =

223: \max\{k_{\it in}, k_{\it out}\}$.  In fact, the only relationship guaranteed

224: among the parameters $n, k_{\it in}, k_{\it out}$ is $k_{\it out} \le

225: nk_{\it in}$. Note that $k$ may be arbitrarily large compared to $n$.

226: Clearly this is the case since, for example, a path can wind around a

227: set of terminals arbitrarily many times. However, there exist

228: non-trivial cases in which $k$ can be much larger than $n$. Even after

229: shortest paths have been computed for each wire, $k$ can be as large

230: as $k=\Omega(2^n)$, see Fig.~\ref{fig-complexity}.

231: The algorithm of Hershberger and

232: Snoeyink~\cite{HerSno94} runs in time $O(nk)$. The algorithm presented in

233: this paper runs in time $O(n^{2+\eps} + k \log^2 n)$, which is an

234: improvement for $n<k$.

235: 

236: \begin{figure}[t]

237: \begin{center}

238: {\includegraphics[width=12cm]{complexity6.eps}}

239: \caption{\small\sf

240:     An example with exponential complexity: $k=\Omega(2^n)$: On the left

241:     is the initial wiring sketch, and on the right is the wiring after the shortest paths have been computed. The number of edge segments in the shortest paths $\sigma_1, \sigma_2, \sigma_3, \sigma_4$ is $1,2,4,8$, respectively. In general, wire $\sigma_i$ has $2^{i-1}$ edge segments. Note that on the right many edge segments are parallel. }

242: \label{fig-complexity}

243: \end{center}

244: \end{figure}

245: 

246: 

247: 

248: 

249: 

250: 

251: Although $k$ can be arbitrarily large compared to $n$, one easily

252: forms the intuition that, because the paths are simple and disjoint,

253: $k$ can be large in a non-trivial way only because path sections are

254: repeated over and over.  For example, a path may spiral arbitrarily

255: many times around a set of points, but each wrap around the set is the

256: same.

257: 

258: Our method makes essential use of this observation.  We do not begin

259: by explicitly searching for repeated path sections---this seems

260: difficult modulo homotopic equivalence.  Instead we begin in section 2

261: by applying vertical shortcuts to the paths (homotopically) so

262: that each left and each right local

263: extreme point occurs at a terminal.  These terminals must then be part of the

264: final shortest paths, and we have decomposed the paths into $x$-monotone

265: pieces with endpoints at terminals.

266: In section 3 we argue that the number of homotopically

267: distinct $x$-monotone

268: pieces is at most $O(n)$.

269: We will bundle together all the homotopically equivalent pieces.

270: Routing one representative from each such bundle using the

271: straightforward ``funnel'' technique takes $O(k + n^2)$ time total.  In

272: section 3 we reduce this using a ``shielding technique'' where we

273: again exploit the fact that the paths are disjoint and use the

274: knowledge gained in routing one shortest path to avoid repeating work

275: when we route subsequent paths.

276: The final step of the algorithm is to unbundle, and recover the final

277: paths by putting together the appropriate pieces.

278: We summarize the main steps of the algorithm in Fig.~\ref{fig-summary}.

279: 

280: \begin{figure}[t]

281: \medskip

282: \noindent

283: \begin{center}

284:  \fbox{

285:  \begin{minipage}{9.5in}

286:     \baselineskip 1mm

287:     \begin{tabbing}

288:        \hspace{6mm}\=\hspace{6mm}\=\hspace{6mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\= \kill

289: {\bf Main Algorithm}\\

290: \> 1. shortcut paths to divide into monotone pieces\\

291: \> 2. bundle homotopically equivalent pieces\\

292: \> 3. find the shortest path for each bundle\\

293: \> 4. unbundle to recover final paths

294:        \end{tabbing}

295:        \end{minipage}

296:        }

297: \end{center}

298: \smallskip

299: \caption{\small\sf Summary of the algorithm.}

300: \label{fig-summary}

301: \end{figure}

302: 

303: Before we turn to the remainder of the paper, which consists of one

304: section for each of steps 1, 2, and 3, it is worth noting that we rely on

305: some powerful techniques.  In step 1, we use simplex range search

306: queries to perform homotopic simplifications on the paths.  To do

307: this, we use Chazelle's cutting trees data

308: structure~\cite{Chazelle93b}.  In step 2 we need to identify

309: homotopically equivalent monotone path pieces.  We use the efficient

310: trapezoidization algorithm of Bar-Yehuda and

311: Chazelle~\cite{BC-tdjc-94} to perform this step.  Step 3 uses ideas

312: from ``funnel'' algorithms for shortest paths~\cite{LP84}.

313: Step 4 is straight-forward.

314: 

315: We will find it convenient to regard each terminal as a small diamond.

316: Otherwise when a vertical shortcut goes through a terminal we will

317: need to specify whether the path actually goes to the left or the right of

318: the terminal, which is awkward to say and to draw.

319: 

320: 

321: 

322: 

323: 

324: 

325: 

326: 

327: \section{Shortcutting to Divide Paths into Monotone Pieces}

328: 

329: 

330: We begin by applying vertical shortcuts to reduce each path to a

331: sequence of $x$-monotone path sections, see Fig.~\ref{fig-monotone}.

332: A {\it vertical

333: shortcut\/} is a vertical line segment $ab$ joining a point $a$ on

334: some path $\pi$ with a point $b$ also on $\pi$, and with the property that

335: the subpath of $\pi$ joining $a$ and $b$, $\pi_{ab}$, is homotopic to the

336: line segment $ab$.

337: 

338: \begin{figure}[h]

339: \begin{center}

340: {\includegraphics[width=12cm]{monotone3.eps}}

341: \caption{\small\sf

342: (a)

343: Two paths $\pi_1$ and $pi_2$ joining terminals $t_1$ to $t_2$ and $t_3$

344: to $t_4$, respectively. (b) The two paths after performing vertical

345: shortcuts. Note that path $\pi_1$ now consists of 3 monotone pieces: $\mu_1$ from $t_1$ to $t_3$, $\mu_2$ from $t_3$ to $t_4$, and $\mu_3$ from $t_4$ to $t_2$; path $\pi_2$ now

346: consists of one $x$-monotone piece, $\mu_4$, homotopically equivalent to $\mu_2$; (c) The final homotopically equivalent

347: shortest paths, $\sigma_1$ and $\sigma_2$.  }

348: \label{fig-monotone}

349: \end{center}

350: \end{figure}

351: 

352: 

353: 

354: We will only do {\it elementary vertical shortcuts\/} where the subpath

355: $\pi_{ab}$ consists of [portions of] 2 line segments or

356: 3 line segments with the middle one

357: vertical, and the other two non-vertical.

358: We distinguish {\it left shortcuts\/} which are elementary vertical

359: shortcuts where $\pi_{ab}$ contains a point to the left of the line through $ab$;

360: {\it right shortcuts\/} where $\pi_{ab}$ contains a point to the right of

361: the line through $ab$; and {\it collinear shortcuts\/} where $\pi_{ab}$

362: lies in the line through $ab$, see Fig.~\ref{fig-left} and

363: Fig.~\ref{fig-shortcuts}.

364: Collinear shortcuts are always applied after left/right shortcuts

365: (and only then), and prevent consecutive vertical segments.

366: 

367: \begin{figure}[t]

368: \begin{center}

369: {\includegraphics[width=14cm]{left2.eps}}

370:  \caption{\small\sf

371: Maximal left shortcuts (bends are represented by circles and

372: terminals by diamonds):

373: (a) a 2 segment shortcut, maximal because $a$ is a bend; (b) a 3 segment

374: shortcut, maximal for the same reason; (c) a 3 segment shortcut,

375: maximal because $ab$ hits a terminal; (d) the result of the shortcut in (c).

376: }

377: \label{fig-left}

378: \end{center}

379: \end{figure}

380: 

381: 

382: \begin{figure}[t]

383: \begin{center}

384: {\includegraphics[width=14cm]{collinear2.eps}}

385:  \caption{\small\sf

386: A sequence of maximal elementary vertical shortcuts

387: (bends are represented by circles and terminals by diamonds).

388: Shortcuts $a_1b_1$ and $a_2b_2$ are left shortcuts; $a_3b_3$ is a right

389: shortcut; and $a_4b_4$ and $a_5b_5$ are collinear shortcuts.

390: }

391: \label{fig-shortcuts}

392: \end{center}

393: \end{figure}

394: 

395: 

396: 

397: We will in fact only apply {\it maximal\/} elementary vertical shortcuts,

398: where the

399: subpath $\pi_{ab}$ cannot be increased.  In particular, this means

400: that for left and right shortcuts, either $a$ or $b$ is a bend,

401: or the line segment $ab$ hits a terminal.

402: 

403: A {\it local left [right] extreme\/} of a path is

404: a point or, more generally, a vertical segment,

405: where the $x$-coordinate of the path reaches a local min [max].

406: Observe that every left [right] shortcut ``cuts off'' a local left [right]

407: extreme.  Conversely, every left [right] local extreme provides a left

408: [right] shortcut {\it unless\/} the local extreme

409: {\it is locked at\/} a terminal,

410: meaning that the left [right] extreme contains the left [right] point of the

411: terminal's diamond.

412: Figure~\ref{fig-left} shows 4 cases of left extremes; in (a--c) they provide

413: left shortcuts, but in case (d) the left extreme is locked at a terminal.

414: 

415: We will use range searching to detect, for a local left extreme, $l$,

416: the maximal left shortcut that can be performed there.

417: In particular, given $l$, we first identify the maximal potential shortcut

418: in the absence of terminals.

419: To do this, take the segment preceding $l$ and the segment following $l$

420: along the path.  Take their right endpoints, and let $t$ be the leftmost

421: of these two points.  Then $t$ determines the maximal potential shortcut

422: that can be performed at $l$.

423: This potential shortcut forms either

424: a triangle (for example, triangle $abl$ in Figure~\ref{fig-left}(a))

425: or a trapezoid (for example, the trapezoid determined by $a$, $b$, and $l$,

426: in Figure~\ref{fig-left}(b)).

427: If a range query tells us that the triangle/trapezoid is free of terminals,

428: then it forms the true maximal shortcut from $l$.

429: Otherwise, the range query should tell us the leftmost terminal inside

430: the triangle/trapezoid, and that determines the maximal left shortcut from $l$.

431: Note that the ranges we need to

432: query are triangles with one vertical side, and trapezoids with two

433: vertical sides.

434: 

435: Our algorithm is to perform elementary vertical shortcuts until none

436: remain, at which time all local left and right extremes are locked at

437: terminals, so the path divides into monotone pieces.

438: See Section \ref{section:monotone-correctness} for justification.

439: Doing elementary vertical shortcuts in an arbitrary order may result in crossing

440: paths as shown in Fig.~\ref{fig-crossing}; to guarantee

441: non-crossing paths we will do the elementary vertical shortcuts

442: in alternating {\it left\/} and {\it right phases\/},

443: where a {\it left [right] phase\/} means we do

444: left [right] shortcuts (and the consequent collinear shortcuts)

445: until no more are possible.

446: See Section \ref{section:non-crossing} for justification.

447: We summarize these steps in Fig.~\ref{fig-pseudocode}.

448: 

449: 

450: 

451: 

452: \begin{figure}[thb]

453: \noindent

454:  \fbox{

455:  \begin{minipage}{9.5in}

456:     \baselineskip 1mm

457:     \begin{tabbing}

458:        \hspace{6mm}\=\hspace{6mm}\=\hspace{6mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\=\hspace{5mm}\= \kill

459: 

460: {\tt ShortCut}\\

461: \vspace{1cm}

462: 

463: traverse paths initializing\\

464: \> $L$ $\leftarrow$ $\{$local left extremes not locked at terminals$\}$\\

465: \> $R$ $\leftarrow$ $\{$local right extremes not locked at terminals$\}$\\

466: \Loop\\

467: \> {\tt LeftPhase}\\

468: \> {\tt RightPhase}\\

469: \Until $L$ and $R$ are empty\\ \\

470: \vspace{1cm}

471: 

472: {\tt LeftPhase}\\

473: \While $L$ is non-empty\\

474: \> \> $l \leftarrow$ remove an element of $L$\\

475: \> \> $\tau \leftarrow$ the triangle/trapezoid forming the maximal potential shortcut at $l$\\

476: \> \> RangeQuery($\tau$) \\

477: \> \>\If $\tau$ is empty of terminals\\

478: \> \>\>$s \leftarrow$ right side of $\tau$\\

479: \> \>\Else\\

480: \> \>\> $s \leftarrow$ leftmost terminal inside $\tau$\\

481: \> \>do a left shortcut of $\tau$ to $s$\\

482: \> \>perform consequent collinear shortcuts\\

483: \> \>\If any right local extremes disappear, remove them from $R$\\

484: \> \>\If the final vertical segment is a local left [right] extreme and not

485: locked at a terminal\\

486: \> \>\>add it to $L$ [$R$, resp.]\\ \\

487: \vspace{1cm}

488: 

489: {\tt RightPhase}\\

490: (defined analogous to {\tt LeftShortcut} but with ``left'' and ``right'' exchanged)

491: 

492: 

493:        \end{tabbing}

494:        \end{minipage}

495:        }

496: \medskip

497: \caption{\small\sf Pseudo-code for the shortcutting phase of the algorithm.}

498: \label{fig-pseudocode}

499: \end{figure}

500: 

501: \begin{figure}[t]

502: \begin{center}

503: {\includegraphics[width=10cm]{crossing.eps}}

504:  \caption{\small\sf

505: Performing shortcuts in arbitrary order may result in crossings. (a) Two disjoint paths $w_1$ and $w_2$; (b) After a left shortcut in $w_1$ and a right shortcut in $w_2$ there is a crossing. Note that no further shortcuts are possible for either path, and hence the crossing cannot be removed; (c) If both shortcuts are right [left] there are no crossings.

506: }

507: \label{fig-crossing}

508: \end{center}

509: \end{figure}

510: 

511: The remainder of this section consists of four subsections:

512: Subsection \ref{section:correctness} justifies organizing the above algorithm

513: around the sets of left and right extremes, rather than explicitly searching

514: each time for a shortcut to perform.

515: Subsection \ref{section:monotone-correctness} establishes the fact that

516: performing elementary vertical shortcuts enables us to divide the

517: paths into $x$-monotone pieces with endpoints at terminals.

518: Subsection \ref{section:non-crossing}

519: argues that doing elementary vertical shortcuts in left/right phases prevents

520: crossings.

521: Finally, Subsection \ref{section:implementation} deals with

522: implementation and run-time analysis.

523: 

524: 

525: 

526: \subsection{Correctness} \label{section:correctness}

527: 

528: The point of this section is to show that the above algorithm

529: correctly maintains the sets $L$ and $R$

530: of local left and right extremes not locked at terminals.

531: Consider a left phase.  The set $R$ is explicitly updated whenever it changes.

532: For the correctness of $L$ we use:

533: 

534: \begin{Claim} \label{claim:left-preservation}

535: Doing one left shortcut does not alter other local left extremes,

536: nor the shortcuts that will be performed there.

537: \end{Claim}

538: 

539: \begin{proof}

540: A left shortcut only removes left portions of non-vertical segments.

541: \end{proof}

542: 

543: We note that a similar claim fails for collinear shortcuts: doing one

544: may prevent others.

545: 

546: \subsection{Correctness of division into monotone pieces} \label{section:monotone-correctness}

547: 

548: \begin{Claim} \label{claim:shortcut-structure}

549: At the end of a left [right] phase of shortcuts, each local left

550: [right] extreme is locked at a terminal.

551: \end{Claim}

552: 

553: We note that more than one left and one right phase may be required, since

554: the right phase may add new members to $L$, see Fig.~\ref{fig-many-phases}.

555: 

556: 

557: \begin{figure}[t]

558: \begin{center}

559: {\includegraphics[width=12cm]{phases.eps}}

560: \caption{\small\sf

561: A path going through alternating left and right shortcutting phases: (a) the original path; (b) after a left phase; (c) after a right phase; (d) after a left phase; (e) after a right phase.

562:  }

563: \label{fig-many-phases}

564: \end{center}

565: \end{figure}

566: 

567: 

568: 

569: \begin{Claim} \label{claim:shortcut-correctness}

570: Let $\pi$ be a path, and let

571: $\mu$ be a result of performing left and right phases

572: of elementary vertical shortcuts on $\pi$ until no more are possible.

573: Suppose that the local left and right

574: extremes of $\mu$ are locked at the terminals $t_{i_1}, \ldots, t_{i_l}$

575: in that order.

576: Let $\sigma$ be a shortest path homotopic to $\pi$.

577: Then the local left [right] extremes of $\sigma$ are locked at exactly the same

578: ordered list of terminals, and furthermore, the portion of $\sigma$

579: between $t_{i_j}$ and $t_{i_{j+1}}$ is a shortest path homotopic to the portion

580: of $\mu$ between those same terminals.

581: 

582: \end{Claim}

583: 

584: \begin{sketch} Because $\pi$ and $\mu$ are homotopic, $\sigma$ is the

585: shortest path homotopic to $\mu$.  We can thus go from $\mu$ to $\sigma$

586: using ``rubber band'' deformations that only shorten the path, and such

587: deformations cannot loosen a left [right] extreme from the terminal

588: it is locked at.

589: \end{sketch}

590: 

591: \def\T{{\cal T}}

592: 

593: \subsection{Non-crossing paths}\ \label{section:non-crossing}

594: The purpose of this section is to prove the following:

595: 

596: \begin{Lemma} Each left [right] phase of shortcuts preserves the

597: property that paths are non-crossing.

598: \end{Lemma}

599: 

600: By symmetry, we can concentrate on a left phase.

601: Note that a left phase, as we have described it, is non-deterministic.

602: At each stage we choose one member $l$ from

603: the set $L$ of current local left extremes,

604: and use it to perform a left shortcut.

605: To prove the Lemma we will first show that for each left phase there is

606: {\it some\/} sequence of choices that preserves the property that

607: the paths are non-crossing.

608: We will then argue that the end result of a phase does not depend on

609: the choices made.

610: 

611: \begin{Claim} Suppose we have a set of non-crossing paths.

612: Let $L$ be the current set of local left extremes not locked at

613: terminals, and let $l \in L$ be a rightmost element of $L$.

614: Performing the shortcut for $l$ (together with any consequent collinear

615: shortcuts) leaves the paths non-crossing.

616: \end{Claim}

617: 

618: 

619: Observe that we can---at least in theory---complete a left phase of shortcuts

620: using this ``rightmost'' order.

621: In practice we choose not to

622: do this simply because of the extra time required to maintain a heap.

623: 

624: \begin{Lemma}

625: The end result of a left phase does not depend on the sequence of choices

626: made during the phase.

627: \end{Lemma}

628: 

629: \begin{proof}

630: We begin with the claim that the set $\cal L$ of all local left extremes

631: that appear in $L$ over the course of the phase is independent of the choices

632: made during the phase.

633: 

634: This implies that the set of left shortcuts performed during the

635: course of the phase is also independent of the choices made during the phase.

636: However, the set of collinear shortcuts is {\it not\/} independent.

637: In particular, the order in which we perform left shortcuts affects

638: the set of collinear shortcuts, see Fig.~\ref{fig-bigmess}.

639: 

640: \begin{figure}[t]

641: \begin{center}

642: {\includegraphics[width=9cm]{bigmess.eps}}

643: \caption{\small\sf

644: (a) A given path undergoes a right phase; (b)

645: Handling local left extreme $l_1$ first yields the shortcut $ab$. Handling the local left extreme $l_2$ first yields shortcut $cd$. The final result of handling $l_1$, $l_2$ and $l_3$ in any order is shortcut $ce$; (c) An ``unfolded'' version of part (b).

646: }

647: \label{fig-bigmess}

648: \end{center}

649: \end{figure}

650: 

651: 

652: Consider one left phase.  Let $L_0$ be the initial set of local left

653: extremes not locked at terminals.  Let $\cal L$ be the union of $L$

654: over the course of the phase.  Suppose two sequences of choices $C_1$

655: and $C_2$ during a left phase yield sets ${\cal L}_1$ and ${\cal

656: L}_2$.  We would like to show that ${\cal L}_1 = {\cal L}_2$.

657: Consider $l \in {\cal L}_1$.  We will prove $l \in {\cal L}_2$ by

658: induction on the number of left shortcuts performed in the phase

659: before $l$ enters ${\cal L}_1$.  If this number is 0 then $l \in L_0$

660: and we are done.  Otherwise, $l$ is added to $L$ as a result of some

661: left shortcut and consequent collinear shortcuts.  Any vertical

662: segment used in the collinear shortcuts may, in its turn, have arisen

663: as a result of some left shortcut and consequent collinear shortcuts.

664: Tracing this process, we find that $l$ is formed from a set of left

665: shortcuts linked by vertical segments, all in the same vertical line

666: as $l$, see Fig.~\ref{fig-bigmess}.

667: All these left shortcuts arose from

668: local left extremes that entered ${\cal L}_1$ before $l$ did, and

669: thus, by induction, are in ${\cal L}_2$. By

670: Claim~\ref{claim:left-preservation} left shortcuts are performed at

671: each of these during the choice sequence $C_2$---though possibly in a

672: different order than in $C_1$.  The consequent collinear shortcuts

673: will merge all the verticals forming $l$, and cannot merge more than

674: that because the segments attached before and after $l$ are not

675: vertical ($l$ is a local left extreme).  Thus $l$ is in ${\cal L}_2$.

676: 

677: 

678: 

679: 

680: 

681: 

682: This proves that the set of local left extremes, and thus the set

683: of left shortcuts is independent of choices made during the phase.

684: Any vertical segment that is in the final set of paths output

685: by the phase arises through left shortcuts plus consequent collinear

686: shortcuts.  Since any set of choices leads to the same set of left shortcuts,

687: though possibly in different order, the consequent collinear shortcuts

688: will arrive at the same final vertical shortcuts---i.e. the same

689: final paths.

690: 

691: \end{proof}

692: 

693: \subsection{Implementation and run-time analysis} \label{section:implementation}

694: 

695: \def\eps{\varepsilon}

696: 

697: In order to perform range queries we need the cutting trees

698: data structure of Chazelle~\cite{Chazelle93b}. The cutting

699: trees can be constructed in  $O(n^{2+\eps})$ time and space, where $\eps$ is an

700: arbitrarily small constant,

701: and they support simplex range queries in time

702: $O(\log n)$  ~\cite{Chazelle93b}.

703: (Note that a trapezoid is a union of two triangles.)

704: In case a triangle that we query is not empty,

705: we need to find the rightmost/leftmost terminal inside it.

706: To do this, create a segment tree $\T$ on the $x$-projections of the terminals,

707: and maintain a cutting tree on each node of $\T$.

708: This enables us to find the rightmost/leftmost

709: terminal inside a query triangle in time $O(\log^2n)$, without

710: increasing the asymptotic space and preprocessing time.

711: Details are standard and omitted from this abstract.

712: 

713: 

714: \begin{Claim} \label{claim:shortcut-bounds}

715: The number of elementary vertical shortcuts that

716: can be applied

717: to a set of paths with a total of $k$ segments is at most $2k$.

718: \end{Claim}

719: 

720: \begin{proof}

721: Assume that no two terminals and/or bends line up vertically.

722: Consider the set of vertical lines through bends and

723: through the left and right sides of each terminal's diamond.

724: An elementary shortcut operates between two of these vertical lines.  If

725: a left [right] shortcut has its leftmost [rightmost] vertical at a

726: bend, then after the shortcut, this vertical disappears forever, see Fig.~\ref{fig-shortcuts}.  There are thus at most $k$ such elementary

727: shortcuts.  Consider, on the other hand, a left shortcut that has its

728: leftmost vertical at a terminal.  This only occurs when two previous

729: left shortcuts are stopped at the terminal, and then combined in a collinear

730: shortcut.  See the right hand pictures of

731: Fig.~\ref{fig-shortcuts}.  Thus an original edge of the path has disappeared.

732: Note that elementary shortcuts never fragment an edge of the path into two

733: edges, but only shorten it from one end or the other.  Thus there are at most

734: $k$ elementary shortcuts of this type.  Altogether, we obtain a bound

735: of $2k$ elementary shortcuts.

736: \end{proof}

737: 

738: This implies the following simple corollary:

739: 

740: \begin{corollary}

741: The running time of the shortcutting phase, not counting the

742: preprocessing to construct cutting trees, is $O(k \log^2n )$.

743: \end{corollary}

744: 

745: Including preprocessing, step 1 takes time

746: $O(n^{2+\eps} + k \log^2 n)$.

747: 

748: \section{Bundling Homotopically Identical Paths}

749: 

750: 

751: Let $M=\{\mu_1\dots \mu_m \}$ be the set of $x$-monotone paths

752: obtained from step 1 of our algorithm. In the second step of

753: the algorithm we bundle homotopically equivalent paths in $M$.

754: More precisely, we take

755: one representative path for each equivalence class of homotopically

756: equivalent paths in $M$.

757: This is justified because the paths in each equivalence class have the

758: same homotopic shortest path.

759: Because the paths in $M$ are non-crossing and $x$-monotone, it is

760: easier to detect

761: homotopic equivalence: two paths are homotopically equivalent if they

762: have the same endpoints, and, between these endpoints

763: no terminal lies vertically above one path and vertically below the other.

764: 

765: In order to perform the bundling we use a

766: trapezoidization of $M$, see Fig.~\ref{fig-trapez}.  We apply the trapezoidization algorithm of

767: Bar-Yehuda and Chazelle~\cite{BC-tdjc-94} to the paths obtained after

768: the shortcuts of step 1, but before these paths are chopped

769: into monotone pieces.

770: 

771: \begin{figure}[t]

772: \begin{center}

773: {\includegraphics[width=4cm]{trapez.eps}}

774: \caption{\small\sf

775: A trapezoidization enables us to identify that $\mu_2$ and $\mu_4$ are homotopically equivalent.

776: }

777: \label{fig-trapez}

778: \end{center}

779: \end{figure}

780: 

781: \begin{Claim} We can perturb the paths output from step 1 so that

782: they become disjoint and the Bar-Yehuda and Chazelle

783: algorithm can be applied.

784: \end{Claim}

785: 

786: The trapezoidization algorithm takes time $O(k + n (\log n)^{1+\eps})$,

787: which is bounded by the time taken by step 1.

788: Once we have a trapezoidization of $M$, we bundle homotopically

789: equivalent paths as follows.  While scanning each $\mu\in M$, we

790: check if it is homotopically equivalent to the path ``below'' it,

791: by examining all the edges of the trapezoidization that are incident to

792: $\mu$ from below.  If all these trapezoidization edges reach the same path

793: $\mu_j$ and none pass through a terminal on the way to $\mu_j$,

794: and $\mu_i$ and $\mu_j$ have the same

795: terminals as endpoints, then we mark

796: $\mu_i$ as a duplicate. Let $R=\{\rho_1\dots \rho_r \}$

797: be the paths of $M$ that are not marked as duplicates.

798: 

799: \begin{Lemma}

800: The number of paths in $R$ is bounded by $2n$.

801: \end{Lemma}

802: 

803: 

804: \begin{proof}

805: Note that every terminal $t$ is either a right endpoint of paths in $R$

806: or a left endpoint of paths in $R$, but not both.

807: This is because a terminal cannot have local left extremes and local 

808: right extremes locked at it without the paths crossing.

809: 

810: For each terminal $t \in T$, associate its bottommost incident path $\phi(t)$

811: (its ``floor''), and the first path hit by a vertical ray going up from $t$, 

812: $\gamma(t)$ (its ``roof'').  

813: We claim that every path in $R$ is $\phi(t)$ or $\gamma(t)$ for

814: some $t \in T$.  This proves that the number of paths in $R$ is at most 

815: $2n$.

816: 

817: Consider a path $\rho \in R$ with left and right terminals $s$ and $t$, 

818: respectively.   Suppose that $\rho$ is not a roof.  Then every vertical 

819: ray extending downward from a point of $\rho$ must hit the same path $\sigma$.

820: (If two rays hit different paths, then in the middle some ray must

821: hit a terminal.)

822: Furthermore, since $\rho$ is not the bottommost path incident to $s$,

823: $\sigma$ must hit $s$.  Similarly $\sigma$ must hit $t$.  But then 

824: $\sigma$ and $\rho$ are homotopically equivalent.

825: \end{proof}

826: 

827: 

828: \section{Shortest Paths}

829: 

830: In this section we find shortest paths homotopic to the $O(n)$

831: monotone paths $R = \{\rho_1, \ldots, \rho_r\}$ produced

832: in the previous section.

833: Let $\rho'_i$ denote the shortest path homotopic to $\rho_i$.

834: We route each path using a funnel technique.

835: The funnel algorithm of

836: \cite{FOCS82*339} and \cite{LP84}

837: operates on a triangulation of the $n$ points

838: (the terminals in our case), and follows the path through the triangulation

839: maintaining a current ``funnel'' containing all possible shortest paths

840: to this point.

841: The algorithm takes time proportional to

842: the number of edges in the path plus the

843: number of intersections between

844: the triangulation edges and the path.

845: 

846: Rather than a triangulation, we will use a trapezoidization formed by

847: passing a vertical line through each of the $n$ terminals, see Fig.~\ref{fig-shield}(a).

848: Then, since each path $\rho_i$ is $x$-monotone, it has $O(n)$ intersections with

849: trapezoid edges, and the funnel algorithm takes

850: time $O(n + k_i)$, where $k_i$ is the number of edges in $\rho_i$.

851: This gives a total over all paths of $O(n^2 + k)$.

852: Recall that $k$ is $\max\{k_{\it in}, k_{\it out}\}$, where

853: $k_{\it in}$ is the number of edges in all the input paths,

854: and $k_{\it out}$ is the number of edges in all the output paths.

855: Note that the number of edges in $R$ is bounded by $k_{\it  in\/}$.

856: 

857: 

858: \begin{figure}[t]

859: \begin{center}

860: {\includegraphics[width=12cm]{shield.eps}}

861: \caption{\small\sf

862: (a) A path $\rho_1$ (dashed), its funnel up to line $l$ (dotted) and the final homotopic shortest path $\rho'_1$; (b) After the shortest path $\rho'_1$ has been found, shielding allows us to route another path $\rho_2$ without examining vertical lines $l_1$ and $l_2$.

863: }

864: \label{fig-shield}

865: \end{center}

866: \end{figure}

867: 

868: With respect to our whole algorithm, this time of $O(n^2 +k)$ is dominated by

869: the $O(n^{2+\epsilon})$ time required to create the range query

870: data structure in step 1.

871: However, it is interesting to see how much improvement we can make

872: in this step alone.

873: In the remainder of this section we describe a randomized algorithm to

874: route the $O(n)$ monotone paths of $R$ in time $O(n \log n + k)$, and finally

875: mention a deterministic algorithm with $O(n \sqrt n + k)$ running time.

876: 

877: Both methods use a ``shielding'' technique.  We begin by describing this

878: idea intuitively.

879: First note that the $\rho_i$'s can be routed independently, since

880: none affects the others.

881: The initial paths $\rho_i$ are non-crossing, and so are the final

882: shortest paths, $\rho'_i$.  If $\rho_j$ is below $\rho_i$, and we have already

883: computed

884: $\rho'_j$, then $\rho'_j$ behaves as a barrier that ``shields''

885: $\rho_i$ from terminals that are vertically below $\rho'_j$, see Fig.~\ref{fig-shield}(b).

886: 

887: 

888: 

889: 

890: To utilize shielding we will modify the basic trapezoidization described above

891: as we discover shortest

892: paths $\rho'_i$.  In particular, the upward vertical ray

893: through terminal $t$, $u(t)$, will be truncated at the lowest shortest

894: path $\rho'_i$ that is strictly above $t$ and does not bend at $t$.

895: The downward vertical ray through terminal $t$, $d(t)$, will be truncated

896: in an analogous way.

897: 

898: The shortest paths found so far, together with the truncated vertical rays

899: $u(t)$ and $d(t)$ for each terminal $t$,

900: partition the plane into

901: trapezoids, each bounded from left and right by the vertical rays,

902: and from above and below by shortest paths.

903: To route a new path $\rho_i$ through this modified trapezoidization

904: we use the following algorithm.

905: 

906: \vspace{2mm}

907: \noindent{\bf Routing $\rho_i$ with shielding}

908: \begin{enumerate}

909: 

910: \item

911: Identify the first trapezoid that $\rho'_i$ will traverse.  This can

912: be done in $O(\log n)$ time because the shortest paths observe the

913: same vertical ordering as the original $\rho_j$'s.

914: 

915: \item

916: Traverse from left to right the sequence of trapezoids that $\rho'_i$

917: will pass through.  (Note that $\rho_i$ itself may pass through different

918: trapezoids, see Fig.~\ref{fig-shield}(b).)

919: We construct the funnel for $\rho'_i$ as we do this traversal.

920: Suppose that our path enters trapezoid $\tau$.  To leave $\tau$ on the right

921: we have two cases.

922: If the right side of $\tau$ is a point, then it is a

923: terminal $t$, and we are locked between

924: two paths that terminate or bend at $t$.  Then the funnel collapses

925: to this point, and we proceed to the next trapezoid if $\rho_i$ continues.

926: Otherwise the right side of $\tau$ is a vertical through some terminal

927: $t$, and (unless $\rho_i$ ends at $t$)

928: we have a choice of two trapezoids to enter, the upper one with

929: left side $u(t)$ or the lower one with left side $d(t)$.

930: We follow path $\rho_i$ until it crosses the infinite vertical

931: line through $t$.  If it passes above $t$ then we enter the upper

932: trapezoid, and otherwise we enter the lower trapezoid.  We update the

933: funnel to include $t$.

934: 

935: \item

936: When we reach the right endpoint of $\rho_i$, the funnel gives

937: us the shortest homotopic path $\rho'_i$.

938: 

939: 

940: \item We update the trapezoidization as follows.

941: For each vertical segment $u(t)$ or $d(t)$ that is

942: intersected by $\rho'_i$

943: we chop  the segment

944: at its intersection point with $\rho_i'$,

945: provided that the intersection point is not $t$

946: (i.e. that $\rho_i'$ does not bend at or terminate at $t$).

947: 

948: \end{enumerate}

949: 

950: Without yet discussing the order in which we route the paths,

951: we can say a bit about the timing of this shielding method.

952: As we traverse a path $\rho_i$ we spend time proportional to the

953: size of $\rho_i$ plus the number of trapezoids traversed by $\rho'_i$.

954: When $\rho'_i$ leaves a trapezoid, say at the line segment $u(t)$

955: above terminal $t$, it may happen that $\rho'_i$ will bend at

956: $t$ or terminate at $t$.  In this case $t$ is part of the output path

957: $\rho'_i$, and we can charge the work for this trapezoid to the output

958: size.  If, on the other hand, $\rho'_i$ does not bend or terminate

959: at $t$, then it crosses $u(t)$ and we chop $u(t)$ there.

960: In this case we charge the work for this trapezoid to the chop.

961: Thus the total time spent by the algorithm is $O(k + C)$

962: where $C$ is the total number of chops performed at the $n$ verticals.

963: 

964: For shielding to be effective

965: we need to route paths in an order

966: that makes $C$ grow more slowly than $n^2$.

967: We first analyze the randomized algorithm where we

968: choose the next path to route at random with uniform

969: probability from the remaining paths.

970: 

971: \begin{Claim}

972: If paths are routed in random order then

973: the expected number of times we chop a segment $u(t)$ or $d(t)$ is

974: $O(\log r)$, and thus $C$ is $O(n \log n)$, and the routing takes time

975: $O(k + n \log n)$.

976: \end{Claim}

977: 

978: \begin{proof}

979: This follows from a  standard backward  analysis.

980: See for example  \cite{cg-book-00} for other proofs along these lines.

981: Let $l$ be a vertical line through a point $t$,

982: and let $u$ denote a ray emerging

983: vertically from $t$.  Assume that $m$ paths

984: intersecting $u$ have been inserted up to now, and we

985: are about to insert a new one.  Then $u$ will be chopped if and only if

986: the new path creates an intersection point below all existing

987: intersection points on $u$, but above $t$ itself.

988: Since the order of the

989: insertions is random, the probability of the new intersection point being

990: below all other intersection points is $1/m$. Summing

991: over all paths yields the claimed bound.

992: \end{proof}

993: 

994: Finally, we mention that we can achieve a routing time of $O(k + n \sqrt n)$

995: deterministically.  Suppose that the paths $\{\rho_1, \ldots, \rho_r\}$

996: are in order from top to bottom.  (We get this for free from step 2 of our algorithm.)  Partition the paths into blocks $B_1, \ldots, B_{\sqrt r}$ each

997: of size $\sqrt r$, and route them one block at a time from $B_{\sqrt r}$ to $B_1$. Within each block we process the paths in order. 

998: Since the largest increasing [decreasing] sequence with this ordering

999: has size $\sqrt r$, therefore the number of chops at each vertical

1000: is $\sqrt r$.  Thus $C$ is $O(n \sqrt n)$, and this routing method takes

1001: time $O(k + n \sqrt n)$.

1002: 

1003: \section{Conclusion and Open Problems}

1004: 

1005: For any set of $n$ disjoint paths joining pairs of terminals in the plane

1006: we can find shortest paths homotopic 

1007: with respect to the set of terminals

1008: in time $O(n^{2+\eps} + k\log^2 n)$ where $k$ is the sum of input and output 

1009: sizes of the paths.  

1010: If $k$ is larger than $n$ this is better than the 

1011: Hershberger-Snoeyink algorithm which runs in time $O(nk)$.

1012: 

1013: More generally, we can use any other range search method, and 

1014: obtain a running time of $O(P + kQ + n (\log n)^{1+\eps} + T)$

1015: where $P$ is the preprocessing time for the range search data structure

1016: on $n$ points, $Q$ is the time for a simplex optimization query (find the 

1017: minimum/maximum $x$-coordinate point in a triangle), and $T$ is the 

1018: time for the shortest path method of step 3---$O(n \log n)$ randomized,

1019: or $O(n \sqrt n)$ deterministic.

1020: Whether $k$ is large or small compared to $n$ determines which 

1021: trade-off between $P$ and $Q$ is preferred.

1022: For example, the partition tree method of 

1023: Matou{\v s}ek \cite{Mat93} yields an emptiness 

1024: query time of $O(\sqrt n)$ with near-linear time preprocessing,

1025: and would be preferable for $k=O(n^{3/2})$.

1026: With any approach to range searching, it may be possible to improve

1027: the query time by taking into account the fact that one of the edges

1028: of each query triangle is vertical.

1029: 

1030: Another open problem is whether we can do without range queries, and

1031: somehow use a trapezoidization of the original paths, since this can

1032: be found so efficiently with the Bar-Yehuda and Chazelle algorithm.

1033: 

1034: 

1035: 

1036: {\small

1037: \bibliographystyle{abbrv}

1038: \bibliography{stephen}

1039: }

1040: \end{document}

1041: 

1042: 

1043: 

1044: