1: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2: % lsst.tex

3: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

4: 

5: \documentclass[11pt]{article}

6: \usepackage{amsmath,amsfonts,amsthm,epsfig}

7: \usepackage{fancybox}

8: 

9: \textheight 8.5in

10: \topmargin -0.2in

11: \oddsidemargin 0.0in

12: \textwidth 6.5in

13: %\textwidth 6.8in

14: 

15: %\evensidemargin 0in

16: %\headheight 0pt

17: %\headsep 0pt

18: 

19: \input{prestuff}

20: 

21: \def\R{\mathbb{R}}

22: \def\Reals#1{\mathbb{R}^{#1}}

23: 

24: \def\defeq{\stackrel{\mathrm{def}}{=}}

25: 

26: \def\vs#1#2#3{#1_{#2},\ldots , #1_{#3}}

27: \def\orig#1{\hat{#1}}

28: \def\ind#1{\left[#1 \right]}

29: 

30: \def\form#1#2{\left<#1 | #2 \right>}

31: \def\form#1#2{#1^{T}#2}

32: 

33: \def\constBD{1}

34: 

35: \newcommand{\ceiling}[1]{\lceil#1\rceil}

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

37: 

38: \def\union{\cup}

39: \def\intersect{\cap}

40: \def\Union{\bigcup}

41: \def\Intersect{\bigcap}

42: 

43: \def\bdry#1{\partial \left(#1 \right)}

44: 

45: \def\edgbet#1{E \left(#1 \right)}

46: 

47: \def\vol#1{\mathrm{{vol}}\left(#1  \right)}

48: \def\area#1{\mbox{{\bf Area}}\left(#1  \right)}

49: \def\Gradius#1#2{\mathrm{rad}_{#1}\left(#2  \right)}

50: \def\cost#1{\mbox{cost}\left(#1  \right)}

51: 

52: 

53: \def\prob#1#2{\Pr_{#1}\left[ #2 \right]}

54: \def\expec#1#2{\mbox{\bf E}_{#1}\left[ #2 \right]}

55: \newcommand{\E}{\mbox{{\bf E}}}

56: 

57: \def\norm#1{\left\| #1 \right\|}

58: 

59: \def\gennorm#1#2{\left\| #2 \right\|_{#1}}

60: 

61: \def\setof#1{\left\{#1  \right\}}

62: \def\sizeof#1{\left|#1  \right|}

63: \def\abs#1{\left|#1  \right|}

64: 

65: \def\calS{\mathcal{S}}

66: 

67: \def\calL{\mathcal{L}}

68: \def\Gdist#1#2#3{\textrm{dist}_{#1} (#2, #3)}

69: \def\dist#1#2{\textrm{dist} (#1, #2)}

70: \def\tdist#1#2{\textrm{dist}_{#1} (#2)}

71: \def\stretch#1#2{\textrm{stretch}_{#1} (#2)}

72: \def\avestretch#1#2{\textrm{ave-stretch}_{#1} (#2)}

73: 

74: \newcommand{\BallShell}[2]{\ensuremath{{\mathop{\mathrm{BS}}\nolimits}({#1},

75: {#2})}}

76: \newcommand{\AS}[1]{\ensuremath{{\mathop{\mathrm{AS}}\nolimits}

77: ({#1})}}

78: \newcommand{\TS}[1]{\ensuremath{{\mathop{\mathrm{TS}}\nolimits}

79: ({#1})}}

80: \newcommand{\GraphSequence}[1]{\ensuremath{\sigma({#1})}}

81: \newcommand{\Contracted}[1]{\ensuremath{\widetilde{#1}}}

82: \newcommand{\OriginalVertices}[1]{\ensuremath{\chi ({#1})}}

83: \newcommand{\AvoidOverfull}[0]{\linebreak}

84: 

85: \newenvironment{fminipage}%

86:   {\begin{Sbox}\begin{minipage}}%

87:   {\end{minipage}\end{Sbox}\fbox{\TheSbox}}

88: 

89: \newenvironment{algbox}[0]{\vskip 0.2in

90: \noindent

91: \begin{fminipage}{6in}

92: }{

93: \end{fminipage}

94: \vskip 0.2in

95: }

96: \newenvironment{tengalgbox}[0]{\vskip 0.0in

97: \noindent

98: \begin{fminipage}{3.45in}

99: }{

100: \end{fminipage}

101: \vskip 0.1in

102: }

103: \newcounter{bean}

104: \newcounter{bean2}

105: 

106: \newenvironment{tightlist}%

107: {\begin{list}{}{\usecounter{bean}%

108: %  \setlength{\topsep}{0em}

109: %  \setlength{\parsep}{0em}

110:   \setlength{\partopsep}{0em}

111: %  \setlength{\itemsep}{0em}

112: %  \setlength{\leftmargin}{\labelwidth}

113:   \setlength{\leftmargin}{0em}

114: %  \setlength{\labelsep}{0em}

115:   \setlength{\labelwidth}{0em}

116: }}%

117: {\end{list}}

118: 

119: \newenvironment{tightlist2}%

120: {\begin{list}{}{\usecounter{bean}%

121:   \setlength{\topsep}{0em}

122:   \setlength{\parsep}{0em}

123:   \setlength{\partopsep}{0em}

124: %  \setlength{\itemsep}{0em}

125: %  \setlength{\leftmargin}{\labelwidth}

126:   \setlength{\leftmargin}{1.3em}

127: %  \setlength{\labelsep}{0em}

128: %  \setlength{\labelwidth}{1em}

129: }}%

130: {\end{list}}

131: 

132: \newenvironment{tightlistA}%

133: {\begin{list}{}{\usecounter{bean}%

134:   \setlength{\topsep}{0em}

135:   \setlength{\parsep}{0em}

136:   \setlength{\partopsep}{0em}

137: %  \setlength{\itemsep}{0em}

138: %  \setlength{\leftmargin}{\labelwidth}

139:   \setlength{\leftmargin}{1.7em}

140: %  \setlength{\labelsep}{0em}

141: %  \setlength{\labelwidth}{1em}

142: }}%

143: {\end{list}}

144: 

145: \newenvironment{tightlistB}%

146: {\begin{list}{}{\usecounter{bean}%

147:   \setlength{\topsep}{0em}

148:   \setlength{\parsep}{0em}

149:   \setlength{\partopsep}{0em}

150: %  \setlength{\itemsep}{0em}

151: %  \setlength{\leftmargin}{\labelwidth}

152: %  \setlength{\leftmargin}{0.5em}

153: %  \setlength{\labelsep}{0em}

154: %  \setlength{\labelwidth}{1em}

155: }}%

156: {\end{list}}

157: 

158: \newenvironment{tightlist3}%

159: {\begin{list}{}{\usecounter{bean}%

160: %  \setlength{\partopsep}{0em}

161:   \setlength{\leftmargin}{0.5em}

162:   \setlength{\labelwidth}{0.5em}

163: }}%

164: {\end{list}}

165: 

166: 

167:     \newcommand\xx{\boldsymbol{\mathit{x}}}

168: \newcommand\yy{\boldsymbol{\mathit{y}}}

169: 

170: \newcommand\xxc{\boldsymbol{\mathit{\widetilde{x}}}}

171: \newcommand\yyc{\boldsymbol{\mathit{\widetilde{y}}}}

172: 

173: \def\mbar{\widehat{m}}

174: 

175: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

176: \begin{document}

177: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

178: 

179: \title{Lower-Stretch Spanning Trees}

180: 

181: \author{

182: Michael Elkin\thanks{Part of this work was done in Yale University,

183: and was supported by the

184:  DoD University Research Initiative (URI)

185: administered by the Office of Naval Research under Grant

186: N00014-01-1-0795. The work was also partially supported by

187: the Lynn and William Frankel Center for Computer

188:  Sciences.}\\

189: Department of Computer Science\\

190: Ben-Gurion University of the Negev

191: \and

192: Yuval Emek\\

193: Department of Computer Science\\

194: and Applied Mathematics\\

195: Weizmann Institute of Science

196: \and

197: Daniel A. Spielman\thanks{

198: Partially supported by NSF grant CCR-0324914.

199: Part of this work was done at Yale University.}\\

200: Department of Mathematics \\

201: Massachusetts Institute of Technology\\

202: \and

203: Shang-Hua Teng\thanks{Partially supported by NSF grants CCR-0311430

204: and ITR CCR-0325630.}\\

205: Department of Computer Science\\

206: Boston University and\\

207: Akamai Technologies Inc.

208: }

209: 

210: \date{\ }

211: 

212: \maketitle

213: 

214: \begin{abstract}

215: {

216: We show that every

217:   weighted connected graph $G$

218:   contains as a subgraph a

219:   spanning tree into which the edges of $G$ can be embedded

220:   with average stretch

221:   $O (\log^{2} n \log \log n)$.

222: Moreover, we show that this tree can be constructed in time

223:   $O (m \log n + n \log^2 n)$ in general, and in time $O (m\log n)$ if

224:   the input graph is unweighted.

225: The main ingredient in our construction is a

226:    novel graph decomposition technique.

227: 

228: Our new algorithm can be immediately used to

229:   improve the running time of

230:   the recent solver for symmetric diagonally dominant linear systems

231:   of Spielman and Teng from

232: \[

233: m 2^{ \left(O (\sqrt{\log n\log\log n}) \right) }\ \mbox{to} \ m \log^{O (1)}n,

234: \]

235: %  to

236: %\[

237: %m \log^{O (1)}n,

238: %\]

239: and to

240:   $O (n \log^{2} n \log \log n)$ when the system

241:   is planar.

242: Our result can also be used to improve several earlier approximation

243:   algorithms that use low-stretch spanning trees.

244: }

245: \end{abstract}

246: 

247: % A category with the (minimum) three required fields

248: %A category including the fourth, optional field follows...

249: 

250: \vspace{1mm}

251: \noindent

252: {\bf Categories and Subject Descriptors:} F.2 {[Theory of Computation]}: {Analysis of Algorithms and

253: Problem Complexity}

254: 

255: \vspace{1mm}

256: \noindent

257: {\bf General Terms:} Algorithms, Theory.

258: 

259: \vspace{1mm}

260: \noindent

261: {\bf Keywords:} Low-distortion embeddings, probabilistic tree metrics,

262:   low-stretch spanning trees.

263: 

264: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

265: \section{Introduction}\label{sec:intro}

266: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

267: 

268: Let $G = (V,E,w)$ be a weighted connected graph, where $w$ is a function from $E$ into the positive reals. We define the length of each edge $e \in E$ to be the reciprocal of its weight:

269: \[

270: d (e) = 1/w (e).

271: \]

272: 

273: Given a spanning tree $T$ of $V$, we define the distance in $T$

274: between a pair of vertices $u,v \in V$, $\tdist{T}{u,v}$, to be the

275: sum of the lengths of the edges on the unique path in $T$ between $u$

276: and $v$. We can then define the stretch\footnote{Our definition

277:   of the stretch differs slightly from that used in \cite{AKPW}:

278:   ${\tdist{T}{u,v}}/{\tdist{G}{u,v}}$, where

279:   $\tdist{G}{y,v}$ is the length of the shortest-path between $u$ and $v$.

280:   See Subsection~\ref{subsec:applications} for a discussion of the difference.}

281:  of an edge $(u,v) \in E$ to be

282: \[

283: \stretch{T}{u,v} = \frac{\tdist{T}{u,v}}{d (u,v)},

284: \]

285: and the average stretch over all edges of $E$ to be

286: \[

287: \avestretch{T}{E} =

288:   \frac{1}{\sizeof{E}}

289:   \sum _{(u,v) \in E} \stretch{T}{u,v}.

290: \]

291: 

292: Alon, Karp, Peleg and West~\cite{AKPW}

293:   proved that every weighted

294:   connected graph $G = (V,E,w)$ of $n$ vertices and $m$ edges contains a

295: spanning tree $T$ such that

296: \[

297: \avestretch{T}{E} = \exp \left(O (\sqrt{\log n \log \log n}) \right),

298: \]

299: and that there exists a collection $\tau = \{T_1,\ldots,T_h\}$ of

300: spanning trees of $G$ and a probability distribution $\Pi$ over $\tau$

301: such that for every edge $e \in E$,

302: \[

303: \expec{T  \leftarrow \Pi}{\stretch{T}{e}} = \exp \left(O (\sqrt{\log n

304: \log \log n}) \right).

305: \]

306: 

307: 

308: The class of weighted graphs considered in this paper

309:   includes multi-graphs that may contain weighted self-loops

310:   and multiple weighted-edges between a pair of vertices.

311: The consideration of multi-graphs is essential

312:   for several results (including some in \cite{AKPW}).

313: 

314: The result of \cite{AKPW} triggered the study of low-distortion

315:   embeddings into {\em probabilistic tree metrics}.

316: Most notable in this context is the work of Bartal \cite{Bartal1, Bartal2}

317:   which shows that if the requirement that the trees $T$

318:   be {\em subgraphs} of $G$ is abandoned, then the upper bound of

319:   \cite{AKPW} can be improved by finding a tree whose distances

320:   approximate those in the original graph with average distortion

321:   $O(\log n \cdot \log\log n)$.

322: On the negative side, a lower bound of

323:   $\Omega(\log n)$ is known for both scenarios \cite{AKPW,Bartal1}. The

324:   gap left by Bartal was recently closed  by  Fakcharoenphol, Rao, and

325:   Talwar \cite{FRT}, who have shown a tight upper bound of $O(\log n)$.

326: 

327: However, some applications of graph-metric-approximation require trees

328:   that are subgraphs.

329: Until now, no progress had been made on reducing

330:   the gap between the upper and lower bounds proved in \cite{AKPW}

331:   on the average stretch of subgraph spanning trees.

332: The bound achieved in \cite{AKPW} for general weighted graphs

333:   had been the best bound known for {\em unweighted} graphs,

334:   even for {\em unweighted planar} graphs.

335: 

336: In this paper\footnote{

337: In the submitted version of this paper, we proved the weaker bound

338:   on average stretch of $O ((\log n \log \log n)^{2})$.

339: The improvement in this paper comes from re-arranging the arithmetic

340:   in our analysis.

341: Bartal~\cite{BartalComm} has obtained a similar improvement by

342:   other means.

343: }%

344: , we significantly narrow this gap by improving the upper

345:   bound of \cite{AKPW} from $\exp ( O(\sqrt{\log n\log\log n}))$ to

346:   $O(\log^{2} n \log\log n)$.

347: Specifically, we give an algorithm

348:   that for every weighted connected graph $G = (V,E, w)$, constructs a

349:   spanning tree $T \subseteq E$ that satisfies \AvoidOverfull{}

350:   \( \avestretch{T}{E} = O( \log^{2} n \log \log n) \).

351: The running time of our algorithm is

352:   \( O(m \log n + n\log^2 n) \) for weighted graphs,

353:   and \( O(m \log n) \) for unweighted.

354: Note that the input graph need not be simple and its number of edges

355:    $m$ can be much larger than $n \choose 2$.

356: However, as  proved in \cite{AKPW}, it is enough to

357:   consider graphs with at most $n (n+1)$ edges.

358: 

359: We begin by presenting a simpler algorithm that guarantees

360:   a weaker bound, $\avestretch{T}{E} = O(\log^3 n)$.

361: As a consequence of the result in \cite{AKPW} that the existence of a

362:   spanning tree with average stretch $f(n)$ for every weighted graph

363:   implies the existence of a distribution of spanning trees in which

364:   every edge has expected stretch $f (n)$,

365:   our result implies that for every weighted

366:   connected graph $G = (V,E,w)$ there

367:   exists a probability distribution $\Pi$ over a set  $\tau =

368:   \{T_1,\ldots,T_h\}$ of spanning trees ($T \subseteq E$ for every $T

369:   \in \tau$) such that for every $e \in E$, $\expec{T  \leftarrow \Pi

370:   }{\stretch{T}{e}} = O(\log^{2} n\log\log n).$

371: Furthermore, our

372:   algorithm itself can be adapted to produce a probability distribution

373:   $\Pi$ that guarantees a slightly weaker bound of $O(\log^3 n)$ in time

374:   $O(m \cdot \log^2 n)$.

375: So far, we have not yet been able to verify

376:   whether our algorithm can be adapted to produce the bound of \AvoidOverfull{}

377:   $O(\log^{2}  n \cdot \log\log n)$ within similar time limits.

378: 

379: \subsection{Applications} \label{subsec:applications}

380: 

381: For some of the applications listed below it is essential to define

382: the stretch of an edge \( (u,v) \in E \) as in \cite{AKPW}, namely,

383: \( \stretch{T}{u,v} = \tdist{T}{u,v} / \tdist{G}{u,v} \). The

384: algorithms presented in this paper can be adapted to handle this

385: alternative definition for stretch simply by assigning new weight \(

386: w'(u,v) = 1 / \tdist{G}{u,v} \) to every edge \( (u,v) \in E \) (the

387: lengths of the edges remain unchanged). Observe that the new weights

388: can be computed in a preprocessing stage independently of the

389: algorithms themselves, but the time required for this computation

390: may dominant the running time of the algorithms.

391: 

392: \subsubsection{Solving Linear Systems}

393: 

394: Boman and Hendrickson~\cite{BomanHendricksonAKPW}

395:   were the first to realize that low-stretch spanning trees

396:   could be used to solve symmetric diagonally dominant linear systems.

397: They  applied the spanning

398:   trees of~\cite{AKPW} to design solvers that run in time

399: 

400: \[

401: m^{3/2} 2^{O (\sqrt{\log n \log \log n})}

402:   \log (1/\epsilon ),

403: \]

404: where $\epsilon $ is the precision of the solution.

405: Spielman and Teng \cite{SpielmanTengPrecon} improved their results

406:   to

407: \[

408: m 2^{O (\sqrt{\log n\log\log n}) } \log (1/\epsilon).

409: \]

410: 

411: 

412: Unfortunately, the trees produced by the algorithms of Bartal \cite{Bartal1, Bartal2}

413:    and Fakcharoenphol, Rao, and Talwar \cite{FRT} cannot be used

414:    to improve these linear solvers, and it is currently not known

415:    whether it is possible

416:    to solve linear systems efficiently using trees that are not subgraphs.

417: 

418: By applying the low-stretch spanning trees developed in

419:   this paper, we can reduce the time for solving these linear systems

420:   to

421: \[

422: m \log^{O (1)}n \log (1/\epsilon),

423: \]

424:   and to $O (n \log^{2} n \log \log n \log (1/\epsilon))$ when the

425:   systems are planar.

426: Applying a recent reduction of Boman, Hendrickson and Vavasis \cite{BHV},

427:   one obtains a $O (n \log^{2} n \log \log n  \log (1/\epsilon))$ time algorithm

428:   for solving the linear systems that arise when

429:   applying the finite element method to solve two-dimensional elliptic

430:   partial differential equations.

431: 

432: 

433: 

434: \subsubsection{Alon-Karp-Peleg-West Game}

435: 

436: Alon, Karp, Peleg and West~\cite{AKPW} constructed low-stretch

437: spanning trees to upper-bound the value of a zero-sum two-player game

438: that arose in their analysis of an algorithm for the $k$-server

439: problem: at each turn, the {\em tree player} chooses a spanning tree

440: $T$ and the {\em edge player} chooses an edge $e\in E$,

441: simultaneously. The payoff to the edge player is $0$ if $e \in T$ and

442: $\stretch{T}{e}+1$ otherwise. They showed that if every

443: $n$-vertex weighted connected graph $G$  has a spanning tree $T$

444: of average stretch $f (n)$, then the value of this game is at most \( f

445: (n) + 1 \). Our new result lowers the bound on the value of this

446: graph-theoretical game from \( \exp \left(O (\sqrt{\log n \log \log

447: n}) \right) \) to \( O\left( \log^{2} n\log\log n \right) \).

448: 

449: 

450: \subsubsection{MCT Approximation}

451: 

452: Our result can be used to improve drastically the upper bound on the

453: approximability of the \emph{minimum communication cost spanning tree}

454: (henceforth, \emph{MCT}) problem. This problem was introduced in

455: \cite{Hu}, and is listed as [ND7] in \cite{GareyJohnson} and \cite{NPKann}.

456: 

457: The instance of this problem is a weighted graph $G = (V,E,w)$, and a

458: matrix $\{r(u,v) \mid u,v \in V\}$ of nonnegative requirements. The

459: goal is to construct a spanning tree $T$ that minimizes $c(T) =

460: \sum_{u,v \in V} r(u,v) \cdot \tdist{T}{u,v}$.

461: 

462: Peleg and Reshef \cite{PelReshef} developed a

463:   $2^{O(\sqrt{\log n\cdot \log\log n})}$ approximation

464:   algorithm for the MCT problem on metrics

465:    using the result of \cite{AKPW}.

466: A similar approximation ratio can be achieved for arbitrary graphs.

467: Therefore our result can be used to produce an

468:    efficient $O(\log^{2} n  \log\log n )$

469:   approximation algorithm for the MCT problem on arbitrary graphs.

470: 

471: \subsubsection{Message-Passing Model}

472: 

473: Embeddings into probabilistic tree metrics have been extremely useful

474: in the context of approximation algorithms

475: (to mention a few:

476: buy-at-bulk network design \cite{AwerbuchAzar}, graph Steiner problem

477: \cite{GargKonjevodRavi}, covering Steiner problem \cite{KonjRavi}).

478: However, it is not clear that these algorithms

479:   can be implemented in  the

480:   message-passing model of distributed computing (see \cite{PelegBook}).

481: In this model, every vertex of the input graph hosts a processor, and

482:   the processors communicate over the edges of the graph.

483: 

484: Consequently, in this model executing an algorithm that starts by

485:   constructing a non-subgraph spanning tree of the network, and then

486:   solves a problem whose instance is this tree is very problematic,

487:   since direct communication over the links of this ``virtual'' tree is

488:   impossible.

489: This difficulty disappears if the tree in this scheme is a

490:   {\em subgraph } of the graph.

491: We believe that our result will

492:   enable the adaptation of the these approximation algorithms to

493:   the message-passing model.

494: 

495: 

496: \subsection{Our Techniques}

497: 

498: We build our low-stretch spanning trees by recursively applying

499:   a new graph decomposition

500:   that we call a \textit{star-decomposition}.

501: A star-decomposition of a graph is a partition of the vertices

502:   into sets that are connected into a star: a central set is connected

503:   to each other set by a single edge (see Figure~\ref{fig:star}).

504: We show how to find star-decompositions that do not cut too many

505: short edges and such that the radius of the graph induced by the

506: star decomposition is not much larger than the radius of the

507: original graph.

508: 

509: Our algorithm for finding a low-cost star-decomposition applies a

510: generalization of the ball-growing technique of

511: Awerbuch~\cite{Awerbuch} to grow \textit{cones}, where the cone at a

512: vertex $x$ induced by a set of vertices $S$ is the set of vertices

513: whose shortest path to $S$ goes through $x$.

514: 

515: \subsection{The Structure of the Paper}

516: 

517: In Section \ref{sec:preliminaries}, we define our notation.

518: In Section \ref{sec:treelogcube}, we introduce

519:   the star decomposition of a weighted connected graph.

520:   We then  show how to use this decomposition to

521:     construct a subgraph spanning tree

522:     with average stretch $O (\log^{3} n)$.

523: In Section \ref{sub:cutting}, we present our

524:   star decomposition algorithm.

525: In Section \ref{sec:Improved}, we refine our construction and improve

526:   the average stretch to $O \left(\log^{2} n\log\log n \right)$.

527: Finally, we conclude the paper in Section

528: 6

529: %\ref{sec:conclusion}

530: and list

531:   some open questions.

532: %Due to space limitations, most proofs are deferred to the appendix.

533: 

534: 

535: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

536: \section{Preliminaries}\label{sec:preliminaries}

537: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

538: 

539: Throughout the paper, we assume that the input graph is a weighted

540: connected multi-graph \( G = (V, E, w) \), where \(w\) is a \emph{weight}

541: function from \(E\) to the positive reals. Unless stated otherwise, we let

542: \(n\) and \(m\) denote the number of vertices and the number of edges

543: in the graph, respectively. The \emph{length} of an edge \( e \in E \) is

544: defined as the reciprocal of its weight, denoted by \( d(e) = 1 / w(e)

545: \).

546: %Moreover, the graphs throughout the paper are not assumed

547: %to be simple, i.e., they may contain multiple edges and self

548: %loops. This may be crucial for some of our arguments.

549: 

550: 

551: \begin{tightlist}

552: \item For two vertices $u,v \in V$,

553:   we define $\dist{u}{v}$ to be the length of the shortest

554:   path between $u$ and $v$ in $E$.

555:   We write $\tdist{G}{u,v}$ to emphasize that the distance is in the

556:   graph $G$.

557: 

558: \item

559: For a set of vertices, $S \subseteq V$,

560:   $G (S)$ is the subgraph induced by vertices in $S$.

561:   We write $\tdist{S}{u,v}$ instead of $\tdist{G (S)}{u,v}$

562:   when $G$ is understood.

563:   $E (S)$ is the set of edges with both endpoints in $S$.

564: 

565: 

566: \item

567: For $S, T \subseteq V$,

568: $E (S,T)$ is the set of edges with one endpoint in $S$

569:   and the other in $T$.

570: 

571: \item The \emph{boundary} of a set $S$, denoted $\bdry{S}$, is the set of edges

572: with exactly one endpoint in $S$.

573: 

574: \item

575: For a vertex $x\in V$, $\tdist{G}{x,S}$ is the length of the shortest path

576:   between $x$ and a vertex in $S$.

577: 

578: \item A \emph{multiway partition} of $V$ is a collection

579:   of pairwise-disjoint sets

580:   $\setof{\vs{V}{1}{k}}$ such that

581:   $\bigcup_{i} V_{i}= V$.

582: The \emph{boundary} of a multiway partition, denoted

583:   $\bdry{\vs{V}{1}{k}}$, is the set of edges with endpoints

584:   in different sets in the partition.

585: 

586: \item

587: The \emph{volume} of a set \(F\) of edges, denoted $\vol{F}$, is the

588: size of the set $|F|$.

589: 

590: \item

591: The \emph{cost} of a set \(F\) of edges, denoted \( \cost{F} \), is

592: the sum of the weights of the edges in \(F\).

593: 

594: \item

595: The \emph{volume} of a set \(S\) of vertices, denoted \( \vol{S} \),

596: is the number of edges with at least one endpoint in \(S\).

597: 

598: \item

599: The \emph{ball} of radius $r$ around a vertex $v$, denoted $B

600: (r,v)$, is the set of vertices of distance at most $r$ from $v$.

601: 

602: \item

603: The \emph{ball shell} of radius \(r\) around a vertex \(v\), denoted

604: \( \BallShell{r}{v} \), is the set of vertices right outside \( B(r,v) \),

605: that is, \( \BallShell{r}{v} \) consists of every vertex \( u \in V - B(r,v)

606: \) with a neighbor \( w \in B(r,v) \) such that \( \dist{v}{u} =

607: \dist{v}{w} + d(u,w) \).

608: 

609: \item

610: For $v\in V$, $\Gradius{G}{v}$ is the smallest $r$ such that every

611: vertex of $G$ is within distance $r$ from $v$. For a set of vertices

612: \( S \subseteq V \), we write \( \Gradius{S}{v} \) instead of \(

613: \Gradius{G (S)}{v} \) when \(G\) is understood.

614: \end{tightlist}

615: 

616: 

617: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

618: \section{Spanning-Trees of \mbox{${\bf O ({\bf

619: \lowercase{log}}^{3}\lowercase{n})}$} Stretch}\label{sec:treelogcube}

620: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

621: 

622: We present our first algorithm that

623:   generates a spanning tree with average stretch $O

624:   (\log^{3}n)$.

625: %We will improve the stretch bound in Section \ref{sec:Improved}.

626: We first state the properties of the graph decomposition algorithm

627:   at the heart of our construction.

628: We then present the construction and analysis of the

629:   low-stretch spanning trees.

630: We defer the description of the graph decomposition algorithm

631:   and its analysis

632:   to Section~\ref{sub:cutting}.

633: 

634: \subsection{Low-Cost Star-Decomposition}\label{sub:structure}

635: 

636: %% DAS: we might want this back

637: %%

638: % Our graph decomposition algorithm will produce

639: %  star-decompositions of graphs, which we now define.

640: % We will state our definition and result for weighted connected graphs.

641: 

642: \begin{definition}[Star-Decomposition]

643: A multiway partition

644: $\{\vs{V}{0}{k}\}$ is a {\em star-decomposition} of a weighted

645:   connected graph $G$ with center

646:   $x_{0} \in V$

647:   (see Figure \ref{fig:star}) if $x_{0} \in V_{0}$ and

648: \begin{itemize}

649: \item [1.] for all $0\leq i\leq k$, the subgraph induced on $V_{i}$ is

650:    connected, and

651: \item [2.] for all $i\geq  1$, $V_{i}$ contains an \texttt{anchor}

652: vertex $x_{i}$ that is connected to a vertex $y_{i} \in V_{0}$ by an

653: edge

654:   $(x_{i},y_{i}) \in E$.

655:   We call the edge $(x_{i}, y_{i})$ the \texttt{bridge} between

656:   $V_{0}$ and $V_{i}$.

657: \end{itemize}

658: %Let $r = \Gradius{G}{x}$.

659: %For $\delta ,\epsilon \leq 1/2$, a star-decomposition

660: %  $\{\vs{V}{0}{k}\}$ is a $(\delta,\epsilon )$-star-decomposition with

661: %  center $x_{0}$ if

662: %  there exists an $r_{0}$ with $\delta r\leq r_{0}\leq (1-\delta ) r$

663: %      such that

664: %\begin{itemize}

665: %\item [a.] $\Gradius{V_{0}}{x}\leq r_{0}$,

666: %           that is,  every vertex in $V_{0}$ is at distance at

667: %           most $r_{0}$ from $x_{0}$, and

668: %\item [b.] for each $i \geq 1$,

669: %   there exists $r_{i}$ such that

670: %  $\Gradius{ V_{i}}{x_{i}}\leq r_{i}$ and

671: %\[

672: %r_{0}+ d (x_{i},y_{i}) +  r_{i} \leq (1+\epsilon ) r.

673: %\]

674: Let $r = \Gradius{G}{x_0}$, and $r_i = \Gradius{V_i}{x_i}$ for each

675:   $0 \le i \le k$.

676: For $\delta ,\epsilon \leq 1/2$, a star-decomposition

677: $\{\vs{V}{0}{k}\}$ is a $(\delta,\epsilon )$-star-decomposition if

678: \begin{itemize}

679: \item [a.]  $\delta r\leq r_{0}\leq (1-\delta ) r$, and

680: 

681: \item [b.]

682: $r_{0}+ d (x_{i},y_{i}) +  r_{i} \leq (1+\epsilon ) r$, for each $i \geq 1$.

683: \end{itemize}

684: The \texttt{cost} of the star-decomposition is

685:    $\cost{\bdry{\vs{V}{0}{k}}}$.

686: \end{definition}

687: 

688: \begin{figure}[htop]

689: \centering\epsfig{figure=star.eps,height=1.8in}

690: \caption{Star Decomposition. \label{fig:star}}

691: \end{figure}

692: 

693: 

694: Note that if

695:  $\setof{\vs{V}{0}{k}}$ is a $(\delta ,\epsilon )$ star-decomposition

696:   of $G$, then the graph consisting of the union of the induced subgraphs on

697:   $\vs{V}{0}{k}$ and the bridge edges $(y_{k}, x_{k})$ has radius at most

698:   $(1+\epsilon)$ times the radius of the original graph.

699: %See Figure \ref{fig:star} for an illustration.

700: 

701: 

702: In Section~\ref{sub:cutting}, we present an algorithm

703:   \texttt{StarDecomp} that satisfies the following

704:   cost guarantee.

705: Let $\xx= (\vs{x}{1}{k})$ and $\yy = (\vs{y}{1}{k})$.

706: 

707: \begin{lemma}[Low-Cost Star Decomposition]\label{lem:stardecomposition}

708: Let $G = (V,E,w)$ be a connected weighted graph and let \(x_0\) be a

709: vertex in \(V\). Then for every positive \( \epsilon \leq 1 / 2 \),

710: \[

711:   \left(\setof{\vs{V}{0}{k}},\xx,\yy\right)

712:   = \mathtt{starDecomp} (G, x_{0}, 1/3, \epsilon) ~ ,

713: \]

714: in time $O (m + n \log n)$, returns a  $(1/3, \epsilon)$-star-decomposition of

715:   $G$ with center $x_0$ of cost

716: \[

717: %\cost{\bdry{\setof{\vs{V}{0}{k}}}} \leq \frac{4 m \log_{2} m}{\alpha r}.

718: \cost{\bdry{V_0, \ldots, V_k}} ~ \leq ~ \frac{6 \, m \log_{2}

719: (m+1)}{\epsilon \cdot \Gradius{G}{x_{0}}}.

720: \]

721: On unweighted graphs, the running time is $O(m)$.

722: \end{lemma}

723: 

724: \subsection{A Divide-and-Conquer Algorithm}\label{sub:log3}

725: 

726: The basic idea of our algorithm is to use low-cost

727: star-decomposition in a divide-and-conquer (recursive) algorithm to

728: construct a spanning tree. We use \( \widehat{n} \) (respectively, \(

729: \widehat{m} \)) rather than \(n\) (resp., \(m\)) to distinguish the

730: number of vertices (resp., number of edges) in the original graph,

731: input to the first recursive invocation, from that of the graph

732: input to the current one.

733: 

734: Before invoking our algorithm we apply a linear-time transformation

735: from \cite{AKPW} that transforms the graph into one with at most \(

736: \widehat{n} (\widehat{n} + 1) \) edges (recall that a multi-graph of

737: \(\widehat{n}\) vertices may have an arbitrary number of edges), and

738: such that the average-stretch of the spanning tree on the original

739: graph will be at most twice the average-stretch on this graph.

740: 

741: We begin by showing how to construct low-stretch spanning trees

742:   in the case that all edges have length $1$.

743: In particular, we use the fact that in this case

744:   the cost of a set of edges equals the number of edges in the set.

745: 

746: \begin{algbox}

747: \noindent Fix $\alpha = (2 \log_{4/3} (\widehat{n}+6))^{-1}$.

748: \vskip 0.1in

749: \noindent \( T = \mathtt{UnweightedLowStretchTree} (G, x_{0}) \).

750: \vskip 0.1in

751: \begin{tightlistA}

752: \item [0.] If $\sizeof{V} \leq 2$, return $G$. (If \(G\) contains

753: multiple edges, return a single copy.)

754: \item [1.] Set $\rho = \Gradius{G}{x_{0}}$.

755: \item [2.]

756: $(\setof{\vs{V}{0}{k}},\xx,\yy)$~$=\mathtt{StarDecomp}(G, x_{0}, 1/3, \alpha)$

757: \item [3.]

758:   For $0 \leq i \leq k$, \\

759:   set  $T_{i} = \mathtt{UnweightedLowStretchTree} (G (V_{i}), x_{i})$

760: 

761: \item [4.] Set $T = \Union_{i} T_{i} \union \Union_{i} (y_{i},x_{i})$.

762: \end{tightlistA}

763: \end{algbox}

764: 

765: \begin{theorem}[Unweighted]\label{thm:Unweighted}

766: Let \( G = (V, E) \) be an unweighted connected graph and let \(x_0\) be a

767: vertex in \(V\). Then

768: \[

769: T = \mathtt{UnweightedLowStretchTree} (G, x_{0}) ~ ,

770: \]

771: in time $O (\widehat{m} \log \widehat{n})$, returns a spanning tree of

772: \(G\) satisfying

773: \begin{equation} \label{eq:RadiusUnweighted}

774: \Gradius{T}{x_{0}} \leq \sqrt{e} \cdot \Gradius{G}{x_{0}}

775: \end{equation}

776: and

777: \begin{equation} \label{eq:AverageStretchUnweighted}

778: \avestretch{T}{E} \leq O (\log^{3} \widehat{m}) ~ .

779: \end{equation}

780: \end{theorem}

781: \begin{proof}

782: For our analysis, we define a family of graphs that converges to $T$.

783: For a graph $G$, we let

784: \[

785:  (\setof{\vs{V}{0}{k}}, \xx, \yy  )

786:   = \mathtt{StarDecomp} (G, x_{0}, 1/3, \alpha)

787: \]

788: and recursively define

789: \[

790: R_{0} (G) ~=~ G \quad \mbox{and}\quad ~~

791: R_{t} (G) ~= ~  \Union_{i} (y_{i},x_{i}) \union \Union_{i} R_{t-1} (G

792: (V_{i})) ~ .

793: \]

794: The graph~$R_{t} (G)$ is what one would obtain if we forced

795:   \texttt{UnweightedLowStretchTree} to return its input graph after

796:   $t$ levels of recursion.

797: 

798: Because for all $\widehat{n}\geq 0$, $(2\log_{4/3}

799: (\widehat{n}+6))^{-1}\leq 1/12$, we have $(2/3+\alpha )\leq 3/4$.

800: Thus, the depth of the recursion in \texttt{UnweightedLowStretchTree}

801:   is at most $\log_{4/3} \widehat{n}$, and we have

802:   $R_{\log_{4/3} \widehat{n}} (G) = T$.

803: One can prove by induction that, for every \( t \geq 0 \),

804: \[

805:   \Gradius{  R_{t} (G)}{x_{0}}

806:   \leq (1+\alpha)^{t} \Gradius{G}{x_{0}}.

807: \]

808: %Of course, this is trivially true for $t = 0$.

809: %In our proof for general $t$, we will use the notation

810: %\[

811: %  \delta^{t}_{G} (u,v) \defeq  \tdist{R_{t} (G)}{u,v}.

812: %\]

813: %To prove the inequality for $t$, assuming it holds for

814: %  $t-1$,

815: %  we consider $\delta^{t}_{G} (x_{0},v)$ for any $v \in V$.

816: %If $v \in V_{0}$, then

817: %\begin{align*}

818: %  \delta^{t}_{G} (x_{0},v)

819: %  & = \delta^{t-1}_{G (V_{0})} (x_{0},v)\\

820: %  & \leq (1+\alpha)^{t-1} \delta^{0}_{G (V_{0})} (x_{0},v)

821: %  \text{ by induction }

822: %  \\

823: %  & \leq (1+\alpha)^{t-1} \Gradius{G}{x_{0}}.

824: %\end{align*}

825: %On the other hand, if $v \in V_{i}$, then

826: %\begin{align*}

827: %  \delta^{t}_{G} (x_{0},v)

828: %  & =   \delta^{t-1}_{G (V_{0})} (x_{0},y_{i})

829: %    + \dist{y_{i}}{x_{i}}

830: %    +   \delta^{t-1}_{G (V_{i})} (x_{i},v)

831: %  \\

832: %  & \leq (1+\alpha)^{t-1}  \delta^{0}_{G (V_{0})} (x_{0},y_{i})

833: %    + \dist{y_{i}}{x_{i}}

834: %    + (1+\alpha)^{t-1}  \delta^{0}_{G (V_{i})} (x_{i},v)

835: %  \text{ by induction }

836: %  \\

837: %  & \leq (1+\alpha)^{t-1} r_{0}

838: %    + \dist{y_{i}}{x_{i}}

839: %    + (1+\alpha)^{t-1} r_{i}

840: %  \\

841: %  & \leq

842: %   (1+ \alpha)^{t} r.

843: %\end{align*}

844: The claim in \eqref{eq:RadiusUnweighted} now follows from $(1+\alpha

845: )^{\log_{4/3} \widehat{n}} \leq  \sqrt{e}$.

846: To prove the claim in \eqref{eq:AverageStretchUnweighted}, we note that

847: \begin{eqnarray}

848: \label{eq:str1}

849: \sum_{(u,v) \in \bdry{\vs{V}{0}{k}}} \stretch{T}{u,v} & \leq &

850: \sum_{(u,v) \in \bdry{\vs{V}{0}{k}}} \left( \tdist{T}{x_{0},u} +

851: \tdist{T}{x_{0},v} \right) \\

852: \label{eq:str3}

853: & \leq  &

854:   \sum_{(u,v) \in \bdry{\vs{V}{0}{k}}}

855:    2 \sqrt{e} \cdot \Gradius{G}{x_{0}},

856:   \text{ by \eqref{eq:RadiusUnweighted}} \\

857: & \leq &

858:    2 \sqrt{e} \cdot  \Gradius{G}{x_0}

859: \left(  \frac{

860:       6 \, m \log_{2} (\widehat{m}+1)

861:        }{

862:         \alpha \cdot \Gradius{G}{x_0}

863:        }

864:  \right),

865:   \text{ by Lemma~\ref{lem:stardecomposition}. } \nonumber

866: %\\

867: % & \leq

868: %  16 \sqrt{e} m \log_{2} m \log_{4/3} (n+6).

869: \end{eqnarray}

870: Applying this inequality to all graphs at all $\log_{4/3} \widehat{n}$

871: levels of the recursion, we obtain

872: \[

873:   \sum_{(u,v) \in E}

874:   \stretch{T}{u,v}

875: ~ \leq~

876: 24 \, \sqrt{e} \, \widehat{m} \log_{2} \widehat{m} \log_{4/3} \widehat{n}

877: \log_{4/3} (\widehat{n}+6)

878: ~ =~

879:    O ( \widehat{m} \log^{3} \widehat{m}) ~ .

880: \]

881: \end{proof}

882: 

883: We now extend our algorithm and proof to general weighted connected

884:    graphs.

885: We begin by pointing out a subtle difference between general

886:    and unit-weight graphs.

887: In our analysis of

888:   \texttt{UnweightedLowStretchTree}, we used the facts that

889:   $\Gradius{G}{x_{0}} \leq n$ and that each edge length is

890:   1 to show that the depth of recursion is at most $\log_{4/3}n$.

891: In general weighted graphs, the ratio of $\Gradius{G}{x_{0}}$

892:   to the length of the shortest edge can be arbitrarily large.

893: Thus, the recursion can be very deep.

894: To compensate, we will contract all edges that are significantly

895:   shorter than the radius of their component.

896: In this way, we will guarantee that each edge is only active

897:   in a logarithmic number of iterations.

898: 

899: Let $e = (u,v)$ be an edge in $G= (V,E,w)$.

900: The contraction of $e$ results in a new graph

901:   by identifying $u$ and $v$ as a new vertex

902:   whose neighbors are the union of the neighbors of $u$ and $v$.

903: All self-loops created by the contraction are discarded.

904: We refer to  $u$ and $v$ as the preimage of

905:   the new vertex.

906: 

907: We now state and analyze our algorithm for

908:   general weighted graphs.

909: 

910: \begin{algbox}

911: \noindent Fix \( \beta = (2 \log_{4/3} (\widehat{n} + 32))^{-1} \).

912: \vskip 0.1in \noindent \( T = \mathtt{LowStretchTree} (G = (V, E,w),

913: x_{0}) \).

914: \vskip 0.1in

915: \begin{tightlistA}

916: \item [0.] If $\sizeof{V} \leq 2$, return $G$. (If \(G\) contains

917: multiple edges, return the shortest copy.)

918: \item [1.] Set $\rho = \Gradius{G}{x_{0}}$.

919: \item [2.] Let \( \Contracted{G} = (\Contracted{V}, \Contracted{E}) \)

920:   be the graph obtained by contracting

921:   all edges in $G$ of length less than \( \beta \rho / \widehat{n} \).

922: \item [3.]

923:   \( \left(\setof{\vs{\Contracted{V}}{0}{k}},

924:   \xxc  , \yyc \right) =

925:   \mathtt{StarDecomp} (\Contracted{G}, x_{0}, 1/3, \beta) \).

926: \item [4.] For each $i$, let $V_{i}$ be the preimage

927:   under the contraction of step 2

928:   of vertices in $\Contracted{V}_{i}$, and

929:   $(x_{i},y_{i}) \in V_{0} \times V_{i}$ be the edge of shortest

930:   length for which $x_{i}$ is a preimage of $\Contracted{x}_{i}$ and

931:   $y_{i}$ is a preimage of $\Contracted{y}_{i}$.

932: \item [5.]

933:   For $0 \leq i \leq k$,

934: set \( T_{i} = \mathtt{LowStretchTree} (G

935:   (V_{i}), x_{i}) \)

936: \item [6.] Set $T = \Union_{i} T_{i} \union \Union_{i} (y_{i},x_{i})$.

937: \end{tightlistA}

938: \end{algbox}

939: 

940: In what follows, we refer to the graph \( \Contracted{G} \),

941: obtained by contracting some of the edges of the graph \(G\), as the

942: \emph{edge-contracted graph}.

943: 

944: \begin{theorem}[Low-Stretch Spanning Tree]\label{thm:main}

945: Let  $G= (V,E,w)$ be a weighted connected graph and let \( x_0 \) be a

946: vertex in \(V\). Then

947: \[

948: T = \mathtt{LowStretchTree}(G,x_{0}) ~ ,

949: \]

950: in time $O (\widehat{m} \log \widehat{n} + \widehat{n} \log^2

951: \widehat{n})$, returns a spanning tree of $G$ satisfying

952: \begin{equation} \label{eq:RadiusWeighted}

953: \Gradius{T}{x_0} \leq 2 \sqrt{e} \cdot  \Gradius{G}{x_0}

954: \end{equation}

955: and

956: \begin{equation} \label{eq:AverageStretchWeighted}

957: \avestretch{T}{E} = O (\log^{3} \widehat{m})

958: \end{equation}

959: \end{theorem}

960: 

961: %and the main result of this section,

962: % Theorem \ref{thm:main}.

963: % % about its performance.

964: %We prove Theorem \ref{thm:main} by extending

965: %  the proof of Theorem \ref{thm:Unweighted}.

966: %The key observation in establishing Claim (2) is that each edge is

967: %  active (uncontracted and uncut) on at most \( O (\log n) \) levels of

968: %  the  recursion.

969: %The proof can be found in Appendix \ref{sec:proofMain}.

970: 

971: \begin{proof}

972: %We first apply a linear-time transformation from \cite{AKPW} that

973: %  transforms the graph into one with at most $n (n+1)$ edges,

974: %  and such that the average-stretch of the spanning tree on the

975: %  original graph will be at most twice the average-stretch on this graph.

976: 

977: 

978: We first establish notation similar to that used in the proof of

979: Theorem~\ref{thm:Unweighted}.

980: Our first step is to define a procedure, \texttt{SD}, that captures

981:   the action of the algorithm in steps 2 through 4.

982: We then define $R_{0} (G) = G$ and

983: \[

984:   R_{t} (G) =   \Union_{i} (y_{i},x_{i}) \union

985:                 \Union_{i} R_{t-1} (G (V_{i})),

986: \]

987: where

988: \[

989:  (\setof{\vs{V}{0}{k}}, \vs{x}{1}{k}, \vs{y}{1}{k})

990:   = \mathtt{SD} (G, x_{0}, 1/3, \beta).

991: \]

992: 

993: We now prove the claim in \eqref{eq:RadiusWeighted}.

994: Let \( \rho = \Gradius{G}{x_0} \). Let \( t = 2

995: \log_{4/3}{(\widehat{n} + 32)} \) and let \( \rho_{t} = \Gradius{R_{t}

996: (G)}{x_{0}} \). Each contracted edge is of length at most \( \beta

997: \rho / \widehat{n} \), and every path in the graph \( G(V_i) \)

998: contains at most \(n\) contracted edges, hence the insertion of the

999: contracted edges to \( G(V_i) \) increases its radius by an additive factor of at most \( \beta \rho \). Since \( (2 \log_{4/3} (\widehat{n} + 32))^{-1}

1000: \leq 1 / 24 \) for every \( n \geq 0 \), it follows that \( 2 / 3 + 2

1001: \, \beta \leq 3 / 4 \). Therefore, following the proof of

1002: Theorem~\ref{thm:Unweighted}, we can show that \( \rho_t \) is at most \(

1003: \sqrt{e} \cdot \Gradius{G}{x_{0}} \).

1004: 

1005: We know that each component of \(G\) that remains after \(t\) levels

1006: of the recursion has radius at most \( \rho (3 / 4)^t \leq \rho

1007: / \widehat{n}^{2} \). We may also assume by induction that for the

1008: graph induced on each of these components, \texttt{LowStretchTree}

1009: outputs a tree of radius at most \( 2 \sqrt{e} (\rho /

1010: \widehat{n}^{2}) \). As there are at most \(n\) of these components,

1011: we know that the tree returned by the algorithm has radius at most

1012: \[

1013: \sqrt{e} \, \rho + n \times 2 \sqrt{e} (\rho / \widehat{n}^{2}) ~ \leq ~ 2

1014: \, \sqrt{e} \, \rho ~ ,

1015: \]

1016: for $\widehat{n} \geq 2$.

1017: 

1018: We now turn to the claim in \eqref{eq:AverageStretchWeighted}, the

1019: bound on the stretch.

1020: In this part, we let \( E_{t} \subseteq E \) denote the set of edges

1021:   that are present at recursion depth \(t\).

1022: That is, their endpoints are

1023:   not identified by the contraction of short edges in step~2, and their

1024:   endpoints remain in the same component.

1025: We now observe that no edge can be present at more than \( \log_{4 /

1026: 3}{((2 \, \widehat{n} / \beta) + 1)} \) recursion depths. To see this,

1027: consider an edge \( (u, v) \) and let \(t\) be the first recursion

1028: level for which the edge is in \( E_{t} \). Let \( \rho_t \) be the

1029: radius of the component in which the edge lies at that time. As \(u\)

1030: and \(v\) are not identified under contraction, they are at distance

1031: at least \( \beta \rho_t / \widehat{n} \) from each other. (This

1032: argument can be easily verified, although the condition for edge

1033: contraction depends on the length of the edge rather than on the

1034: distance between its endpoints.) If \(u\) and \(v\) are still in the same graph on recursion level \( t + \log_{4 / 3}{((2 \, \widehat{n} / \beta) +

1035: 1)} \), then the radius of this graph is at most \( \rho_t / ((2 \,

1036: \widehat{n} / \beta) + 1) \), thus its diameter is strictly less than

1037: \( \beta \rho_t / \widehat{n} \), in contradiction to the distance

1038: between \(u\) and \(v\).

1039: 

1040: Similarly to the way that $ \sum_{(u,v) \in \bdry{\vs{V}{0}{k}}}

1041:   \stretch{T}{u,v}$ is upper-bounded in inequalities \eqref{eq:str1}-

1042: \eqref{eq:str3}

1043:  in the proof of Theorem~\ref{thm:Unweighted}, it follows  that

1044:   the total contribution to the stretch at depth $t$

1045:   is at most

1046: \[

1047: O (\vol{E_{t}} \log^{2}\widehat{m}).

1048: \]

1049: Thus, the sum of the stretches over all recursion depths

1050:   is

1051: \[

1052:   \sum _{t} O \left( \vol{E_{t}} \log^{2}\widehat{m} \right)

1053:   = O\left(\widehat{m} \log^{3}\widehat{m}\right).

1054: \]

1055: 

1056: We now analyze the running time of the algorithm. On each recursion

1057: level, the dominant cost is that of performing the \texttt{StarDecomp}

1058: operations on each edge-contracted graph. Let \(V_t\) denote the

1059: set of vertices in all edge-contracted graphs on recursion level

1060: \(t\). Then the total cost of the \texttt{StarDecomp} operations on

1061: recursion level $t$ is at most $O(\sizeof{E_{t}} + \sizeof{V_t} \log

1062: \sizeof{V_t})$. We will prove soon that \( \sum_{t}{\sizeof{V_t}} = O

1063: (\widehat{n} \log{\widehat{n}}) \), and as $\sum_{t} \sizeof{E_{t}} = O

1064: (\widehat{m} \log \widehat{m})$, it follows that the total running

1065: time is $O (\widehat{m} \log \widehat{n} + \widehat{n} \log^2 \widehat{n})$.

1066: Note that for unweighted graphs $G$, the running time is only

1067: $O(\widehat{m} \log \widehat{n})$.

1068: %factor $\log n$, that is, it is $O(m \log n)$.

1069: \end{proof}

1070: %{\bf

1071: %Depends on how detailed we wanted to be:

1072: %  To support fast decision for contraction and un-contraction of edges, we

1073: %  could presort the edges by their lengths and properly maintain

1074: %  links between new vertices and their preimages.

1075: 

1076: %Also, here I assume that $\mathtt{StarDecomp}$ takes $O (m\log n)$

1077: %  time, rather than the $O (m\log^{2}n)$ time as we stated currently.

1078: %}

1079: 

1080: The following lemma shows that even though the number of recursion

1081: levels can be very large, the overall number of vertices in

1082: edge-contracted graphs appearing on different recursion levels is at

1083: most $O(\widehat{n} \log \widehat{n})$. This lemma is used only for

1084: the analysis of the running time of our algorithm; a reader interested

1085: only in the existential bound can skip it.

1086: 

1087: \begin{lemma}

1088: \label{lm:num_vs}

1089: Let \( V_t \) be the set of vertices appearing in edge-contracted

1090: graphs on recursion level $t$. Then $\sum_t \sizeof{V_t} =

1091: O(\widehat{n} \log \widehat{n})$.

1092: \end{lemma}

1093: \begin{proof}

1094: Consider an edge-contracted graph \( \Contracted{G} = (\Contracted{V},

1095: \Contracted{E}) \) on recursion level \(t\) and let \(x\) be a vertex

1096: in \( \Contracted{V} \). The vertex \(x\) was formed as a result of a

1097: number of edge contractions (maybe \(0\)). Consequently, \(x\) can be

1098: viewed as a set of all original vertices that were identified together

1099: to form it, i.e., \(x\) can be viewed as a \emph{super-vertex}. Let \(

1100: \OriginalVertices{x} \) denote the set of original vertices that were

1101: identified together to form the super-vertex \( v \in \Contracted{V}

1102: \).

1103: 

1104: We claim that for every super-vertex \( x \in V_{t + 1} \), there

1105: exist a super-vertex \( y \in V_t \) such that \(

1106: \OriginalVertices{x} \subseteq \OriginalVertices{y} \). To prove it,

1107: note that every graph on recursion level \(t + 1\) corresponds to a

1108: single component of a star decomposition on recursion level \(t\).

1109: Moreover, an edge that was contracted on recursion level \(t + 1\)

1110: must have been contracted on recursion level \(t\) as well.

1111: Therefore we can consider a directed forest \( \mathcal{F} \) in

1112: which every node at depth \(t\), corresponds to some super-vertex in

1113: \( V_t \), and an edge leads from a node \(y\) at depth \(t\) to a

1114: node \(x\) at depth \(t + 1\), if \( \OriginalVertices{x} \subseteq

1115: \OriginalVertices{y} \). Note that the roots of \( \mathcal{F} \)

1116: correspond to super-vertices on recursion level \(0\) and the leaves

1117: of \( \mathcal{F} \) correspond to the original vertices of the

1118: graph \(G\).

1119: 

1120: In the proof of Theorem~\ref{thm:main}, we showed that every edge is

1121: present on \( O (\log{\widehat{n}}) \) recursion levels. Following a

1122: similar line of arguments, one can show that every super-vertex is

1123: present on \( O (\log{\widehat{n}}) \) (before it decomposes to

1124: smaller super-vertices, each contains a subset of its vertices). Since

1125: there are \( \widehat{n} \) vertices in the original graph \(G\),

1126: there are \( \widehat{n} \) leaves in \( \mathcal{F} \). Therefore the

1127: overall number of nodes in the directed forest \( \mathcal{F} \) is \(

1128: O (\widehat{n} \log{\widehat{n}}) \), and the bound on \(

1129: \sum_{t}{\sizeof{V_t}} \) holds.

1130: 

1131: %Consider a node $z$ on some level $t$ of the recursion.

1132: %The node $z$ was formed as a result of a number $k_z \ge 0$ of edge

1133: %contractions. Consequently, $z$ can be viewed as a set (or, {\em

1134: %  super-node}) of all original vertices that were glued together to

1135: %form $z$.

1136: %

1137: %Consider the graph $F$ whose vertices are all the

1138: %super-nodes. (Particularly, original nodes are singleton super-nodes.)

1139: %There is an edge between two super-nodes $z$ and $z'$ if and only if

1140: %$z$ was formed as a result of contracting an edge $e'$ with $z' \in

1141: %e'$, or vice versa.

1142: %

1143: %Observe that $F$ is a forest, and $z$ is the parent of $z'$ if they

1144: %are neighbors in $F$, and the set of nodes that form $z$ contains the

1145: %set of nodes that form $z'$. Observe also that every internal

1146: %super-node in $F$ has exactly two children, and that the number of

1147: %leaves in $F$ is $n$. Consequently, the overall number of super-nodes

1148: %is at most $2n-1$. Moreover, each super-node $z$ appears on at most

1149: %$O(\log n)$ different recursion levels. (This is because it is enough

1150: %that one of the edges incident to $v$ gets contracted for the

1151: %super-node $z$ to merge with the other endpoint of this edge, and to

1152: %disappear.)

1153: %Also, $z$ appears at most once on every level, and thus its overall

1154: %contribution to $\sum_t n_t$ is at most $O(\log n)$. Hence,

1155: %$\sum_t n_t = O(n\log n)$.

1156: \end{proof}

1157: 

1158: 

1159: 

1160: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1161: \section{Star decomposition}\label{sub:cutting}

1162: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1163: 

1164: Our star decomposition algorithm exploits two algorithms for growing

1165:   sets.

1166: The first, \texttt{BallCut}, is the standard ball growing technique

1167:   introduced by Awerbuch~\cite{Awerbuch}, and was the basis

1168:   of the algorithm of~\cite{AKPW}.

1169: The second, \texttt{ConeCut}, is a generalization of ball growing

1170:   to cones.

1171: So that we can analyze this second algorithm, we abstract the analysis

1172:   of ball growing from the works of~\cite{LeightonRao,AumannRabani,LLR}.

1173: Instead of nested balls, we consider \textit{concentric systems},

1174:   which we now define.

1175: 

1176: \begin{definition}[Concentric System]\label{def:concentric}

1177: A {\em concentric system } in a weighted graph $G= (V,E,w)$ is a

1178: family of vertex sets $\calL =\{L_{r} \subseteq V: r \in

1179: \Reals{+}\cup \{0 \} \}$ such that

1180: \begin{enumerate}

1181: \item \( L_0 \neq \emptyset \),

1182: \item $L_{r} \subseteq L_{r'}$ for all $r< r'$, and

1183: \item if a vertex $u \in L_{r}$ and $(u,v)$ is an edge in $E$ then $v

1184: \in L_{r+d(u,v)}$.

1185: \end{enumerate}

1186: \end{definition}

1187: 

1188: For example, for any vertex $x\in V$, the set of balls

1189:   $\setof{B (r,x)}$ is a concentric system.

1190: The {\em radius} of a concentric system $\calL $ is

1191:    $radius (\calL ) = \min \{r: L_{r} = V \}$.

1192: For each vertex $v\in V$, we define $\gennorm{\calL}{v}$

1193:   to  be the smallest $r$ such that

1194:   $v\in L_{r}$.

1195: 

1196: \begin{lemma}[Concentric System Cutting]\label{lem:cuttingconcentricsystem}

1197: Let $G = (V,E,w)$ be a connected weighted graph and let $\calL  =

1198: \{L_{r} \}$ be a concentric system. For every two reals \( 0 \leq

1199: \lambda < \lambda' \), there exists a real \( r \in [\lambda,

1200: \lambda') \) such that

1201: \[

1202: \cost{\bdry{L_{r}}}

1203: ~ \leq ~

1204: \frac{\vol{L_{r}} + \tau }{\lambda' - \lambda}

1205: \max \left[1,

1206:  \log_{2} \left(\frac{m+\tau }{\vol{E (L_{\lambda})} +\tau } \right) \right]

1207:  ,

1208: \]

1209: where $m = \sizeof{E}$ and

1210: \[

1211:   \tau  = \begin{cases}

1212: 1 & \text{if $\vol{E (L_{\lambda})} = 0$}\\

1213: 0 & \text{otherwise.}

1214: \end{cases}

1215: \]

1216: \end{lemma}

1217: 

1218: \begin{proof}

1219: Note that rescaling terms does not effect the statement of the

1220:   lemma.

1221: For example, if all the weights are doubled, then the costs are

1222:   doubled but the distances are halved.

1223: Moreover, we may assume that \( \lambda' \leq radius (\calL) \), since

1224: otherwise, choosing \( r = radius (\calL) \) implies that \(

1225: \bdry{L_r} = \emptyset \) and the claim holds trivially.

1226: 

1227: Let $r_{i} = \norm{v_{i}}$, and assume that the vertices are ordered

1228:   so that

1229:   $r_{1} \leq r_{2} \leq \dotsb \leq r_{n}$.

1230: We may now assume without loss of generality that each edge

1231:   in the graph has minimal length.

1232: That is, an edge from vertex $i$ to vertex $j$ has length

1233:   $\abs{r_{i} - r_{j}}$.

1234: The reason we may make this assumption is that it only increases

1235:   the weights of edges, making our lemma strictly more difficult to

1236:   prove.

1237: (Recall that the weight of an edge is the reciprocal of its length.)

1238: 

1239: Let $B_{i} = L_{r_{i}}$.

1240: Our proof will make critical use of a quantity $\mu_{i}$, which is defined

1241:   to be

1242: \[

1243: \mu_i ~=~

1244: %%DAS 10/25

1245: %1+

1246: %

1247: \tau  +

1248:  \vol{E(B_i)}

1249: %  \sum_{e \in B_{i}} 1

1250:  +

1251:   \sum_{(v_{j},v_{k}) \in E: j \leq  i < k}

1252:    \frac{r_{i} - r_{j}}{r_{k} - r_{j}}.

1253: \]

1254: That is, $\mu_{i}$ sums the edges inside $B_{i}$, proportionally counting

1255:   edges that are split by the boundary of the ball.

1256: The two properties of $\mu_{i}$ that we exploit are

1257: \begin{equation}\label{eqn:muiPlus}

1258:   \mu_{i+1} = \mu_{i} + \cost{\bdry{B_{i}}} (r_{i+1} - r_{i}),

1259: \end{equation}

1260: and

1261: \begin{equation}\label{eqn:muiVol}

1262: \tau +  \vol{E (B_{i})} \leq   \mu_{i} \leq \tau  + \vol{B_{i}}.

1263: \end{equation}

1264: The equality~\eqref{eqn:muiPlus}  follows  from the definition

1265: by a straight-forward calculation, as

1266: 

1267: \[

1268: \mu_i ~= ~ \tau

1269:   + \vol{ E (B_{i+1}) } - \vol{ \{ (v_j, v_{i+1}) \in E \mid j \leq i \} }

1270:    + \sum_{(v_j, v_k) \in E : j \leq i < k}{\frac{r_i - r_j}{r_k - r_j}}

1271: \]

1272: and

1273: \[

1274: \cost{\bdry{B_{i}}} (r_{i+1} - r_{i}) = \sum_{(v_j, v_k) \in E : j

1275: \leq i < k}{\frac{r_{i+1} - r_i}{r_k - r_j}} ~ .

1276: \]

1277: 

1278: % as $|E(B_{i+1})| - |E(B_i)| =

1279: %|\{(v_{i+1},v_l) : l \in \{1,2,\ldots,i\}|$,

1280: %and

1281: %\begin{eqnarray*}

1282: %&& \sum_{(v_j,v_k) \in E : j \le i < k} \frac{r_i - r_j}{r_k - r_j} -

1283: %\sum_{(v_j,v_k) \in E : j \le i+1 < k} \frac{r_{i+1} - r_j}{r_k - r_j}

1284: %~= \\

1285: %&&

1286: %(r_{i+1} - r_i) \cdot \sum_{e = (v_j,v_k) \in E: j \le i, i+1 < k} 1/d(e) -

1287: %\sum_{j \le i, e = (v_j,v_{i+1}) \in E} {{r_i - r_j} \over {d(e)}}~.

1288: %\end{eqnarray*}

1289: 

1290: %%DAS 10/25

1291: %The only trick to proving inequality~\eqref{eqn:muiVol} is to observe

1292: %  that the $+1$ in $\mu_{i}$ is offset by an edge from $v_{i}$ leaving

1293: %  $B_{i}$, which must exist because the graph is connected and $i < n$.

1294: %The point is that this edge does not contribute to the sums in the definition of

1295: %  $\mu_{i}$.

1296: 

1297: Choose $a$ and $b$ so that \( r_{a-1} \leq \lambda < r_{a} \) and \(

1298: r_{b} < \lambda' \leq r_{b+1} \). Let \( \nu = \lambda' - \lambda \).

1299: We first consider the trivial case in which $b < a$.

1300: In that case, there is no vertex whose distance from $v_{0}$

1301:   is between \(\lambda\) and \(\lambda'\).

1302: Thus every edge crossing \( L_{(\lambda + \lambda') / 2} \)

1303:   has length at least \(\nu\), and therefore cost at most \( 1 / \nu \).

1304: Therefore, by setting \( r = (\lambda + \lambda') / 2 \), we obtain

1305: \[

1306: \cost{\bdry{L_{r}}} \leq \vol{\bdry{L_{r}}} \frac{1}{\nu} \leq

1307: \vol{L_{r}} \frac{1}{\nu} ~ ,

1308: \]

1309: establishing the lemma in this case.

1310: 

1311: We now define

1312: \[

1313: \eta = \log _{2} \left(\frac{m+\tau }{\vol{E (B_{a-1})} + \tau } \right) ~ .

1314: \]

1315: Note that \( B_{a - 1} = L_{\lambda} \), by the choice of \(a\).

1316: A similarly trivial case is when $[a,b]$ is non-empty, and where

1317:   there exists an $i \in [a - 1, b]$ such that

1318: \[

1319: r_{i + 1} - r_{i} \geq \frac{\nu}{\eta} ~ .

1320: \]

1321: In this case, every edge in $\bdry{B_i}$ has cost at most \( \eta

1322: / \nu \), and by choosing $r$ to be \( \max\{r_i, \lambda\} \), we

1323: satisfy

1324: \[

1325:   \cost{\bdry{L_{r}}} \

1326: \leq

1327:   \sizeof{\bdry{L_{r}}} \frac{\eta}{\nu}

1328: \leq

1329:   \vol{L_{r}} \frac{\eta}{\nu} ~ ,

1330: \]

1331: hence, the lemma is established in this case.

1332: 

1333: In the remaining case that

1334:   the set $[a,b]$ is non-empty

1335:   and for all $i \in [a - 1, b]$,

1336: \begin{equation}\label{eqn:shortDist}

1337:   r_{i + 1} - r_{i} < \frac{\nu}{\eta} ~ ,

1338: \end{equation}

1339:   we will prove that there exists an $i \in [a - 1, b]$ such that

1340: \[

1341:   \cost{\bdry{B_{i}}} \leq \frac{\mu_{i} \eta}{\nu} ~ ,

1342: \]

1343: hence, by choosing \( r = \max\{r_i, \lambda\} \), the lemma is

1344: established due to \eqref{eqn:muiVol}.

1345: 

1346: Assume by way of contradiction

1347:   that

1348: \[

1349:   \cost{\bdry{B_{i}}} > \mu_{i} \eta / \nu

1350: \]

1351: for all $i \in [a - 1, b]$.

1352: It follows, by \eqref{eqn:muiPlus}, that

1353: \[

1354:   \mu_{i + 1} > \mu_{i} + \mu_{i} (r_{i + 1} - r_{i}) \eta  / \nu

1355: \]

1356: for all \( i \in [a - 1, b] \), which implies

1357: \begin{align*}

1358: \mu_{b+1} & > ~ \mu_{a-1} \prod_{i = a - 1}^{b} \left( 1 + (r_{i + 1} - r_{i})

1359: \eta / \nu \right) \\

1360: & \geq ~ \mu_{a-1} \prod_{i = a - 1}^{b} 2^{\left((r_{i + 1} - r_{i})

1361: \eta / \nu \right)} \text{ , by~\eqref{eqn:shortDist} and since \( 1 +

1362: x \geq 2^x \) for every \( 0 \leq x \leq 1 \)} \\

1363: & = ~ \mu_{a-1} \cdot 2^{(r_{b+1} - r_{a-1}) \eta / \nu} \\

1364: & \geq ~ \mu_{a-1} \cdot \left( (m+\tau ) / (\vol{E (B_{a-1})}+\tau )

1365: \right) \\

1366: & \geq ~ m +\tau  \text{ , by \eqref{eqn:muiVol},}

1367: \end{align*}

1368: which is a contradiction.

1369: \end{proof}

1370: 

1371: %We defer the proof of Lemma~\ref{lem:cuttingconcentricsystem}

1372: %   to Appendix \ref{sec:cuttingProof}.

1373: An analysis of the following standard ball growing algorithm

1374:   follows immediately by applying

1375:   Lemma~\ref{lem:cuttingconcentricsystem} to the concentric system

1376: $\setof{B (r,x)}$.

1377: 

1378: \begin{algbox}\label{alg:ballcut}

1379: \( r = \mathtt{BallCut} (G, x_{0}, \rho, \delta) \)

1380: \vskip 0.1in

1381: \begin{tightlist2}

1382: \item [1.] Set \( r = \delta \rho \).

1383: \item [2.] While $\cost{\bdry{B (r, x_{0})}} >

1384: \frac{\vol{B (r,x_{0})} +1}{(1 - 2 \, \delta) \rho }

1385:  \log_{2} (m+1)$,

1386: \begin{tightlistA}

1387: \item [a.] Find the vertex $v \not \in B (r, x_{0})$ that minimizes

1388: $\dist{x_{0}}{v}$ and set $r = \dist{x_{0}}{v}$.

1389: \end{tightlistA}

1390: \end{tightlist2}

1391: \end{algbox}

1392: 

1393: \begin{corollary}[Weighted Ball Cutting]\label{cor:weightedBallCutting}

1394: Let $G = (V,E,w)$ be a connected weighted graph, let $x\in V$, \(

1395: \rho = \Gradius{G}{x} \), \( r = \mathtt{BallCut} (G, x_{0},

1396: \rho, 1 / 3) \), and $V_{0} = B (r, x)$.

1397: Then \( \rho / 3 \leq r < 2 \, \rho / 3 \) and

1398: \[

1399: \cost{\bdry{V_{0}}} ~ \leq ~ \frac{3 \, (\vol{V_{0}}+1)

1400:   \log_{2} (\sizeof{E} + 1)} {\rho} ~ .

1401: \]

1402: \end{corollary}

1403: 

1404: 

1405: We now examine the concentric system that enables us to

1406:   construct $\vs{V}{1}{k}$ in Lemma \ref{lem:stardecomposition}.

1407: \begin{definition}[Ideals and Cones]\label{def:C}

1408: For any weighted graph $G = (V,E,w)$ and $S\subseteq V$,

1409:   the set of {\em forward edges induced by $S$} is

1410: \[

1411: F(S) = \{(u\rightarrow v): (u,v)\in E, \dist{u}{S} + d(u,v) = \dist{v}{S}\}.

1412: \]

1413: For a vertex \( v \in V \), the \emph{ideal of $v$ induced by $S$},

1414:   denoted \( I_{S} (v) \), is the set of vertices reachable from $v$ by

1415:   directed edges in $F(S)$, including $v$ itself.

1416: 

1417: For a vertex \( v \in V \), the {\em cone of width $l$ around $v$

1418:   induced by $S$}, denoted \( C_{S} (l,v) \),

1419:   is the set of vertices in $V$ that

1420:   can be reached from $v$ by a path, the sum of the lengths of whose

1421:   edges $e$ that do not belong to $F(S)$ is at most $l$.

1422: Clearly, $C_{S} (0,v) = I_{S} (v)$ for all $v \in V$.

1423: \end{definition}

1424: That is, $I_{S} (v)$ is the set of vertices that have shortest

1425:   paths to $S$ that intersect $v$.

1426: Also, $u \in C_{S} (l,v)$ if there exist

1427:   $a_{0}, \dots , a_{k-1}$ and $b_{1}, \dots , b_{k}$

1428:   such that $a_{0} = v$,

1429:   $b_{k} = u$,

1430:   $b_{i+1} \in I_{s} (a_{i})$, $(b_{i}, a_{i}) \in E$,

1431:   and

1432: \[

1433:   \sum_{i} d(b_{i}, a_{i}) \leq l.

1434: \]

1435: 

1436: We now establish that these cones form concentric systems.

1437: 

1438: \begin{proposition}[Cones are concentric]\label{pro:concentriccone}

1439: Let $G = (V,E,w)$ be a weighted graph and let $S \subseteq V$.

1440: Then for all $v \in V$, $\setof{C_{S} (l,v)}_{l}$ is a concentric

1441: system in $G$.

1442: \end{proposition}

1443: \begin{proof}

1444: Clearly, $C_{S} (l,v) \subseteq C_{S} (l',v)$ if $l<l'$.

1445: Moreover, suppose $u\in C_{S} (l,v)$ and $(u,w) \in E$.

1446: Then if $(u \rightarrow w) \in F$, then $w \in C_{S} (l,v)$ as well.

1447: Otherwise, the path witnessing that $u \in C_{S} (l,v)$ followed by

1448:  the edge $(u,w)$ to $w$ is a witness that $w\in C_{S} (l+d(u,w),v)$.

1449: \end{proof}

1450: \begin{algbox}\label{alg:conecut}

1451: \( r = \mathtt{ConeCut} (G, v, \lambda, \lambda', S) \)

1452: \vskip 0.1in

1453: \begin{tightlist2}

1454: \item [1.] Set \( r = \lambda \) and

1455: if $\vol{E (C_{S} (\lambda ,v))} = 0$,

1456: \begin{tightlistA}

1457: \item Set $\mu =

1458:  (\vol{C_{S} (r, v)}+1) \log_{2} (m+1)$

1459: \end{tightlistA}

1460: otherwise,

1461: \begin{tightlistA}

1462: \item Set $\mu =

1463: \vol{C_{S} (r, v)} \log_{2} (m/ \vol{E (C_{S} (\lambda ,v))}$

1464: \end{tightlistA}

1465: \item [2.] While \( \cost{\bdry{C_{S} (r, v)}} > \mu / (\lambda' - \lambda) \),

1466: \begin{tightlistA}

1467: \item [a. ] \hspace*{-0.25in} Find the vertex \( w \not \in C_{S} (r, v)

1468: \) minimizing \( \dist{w}{C_{S}(r, v)} \) and set \( r = r +

1469: \dist{w}{C_{S} (r, v)} \).

1470: \end{tightlistA}

1471: \end{tightlist2}

1472: \end{algbox}

1473: 

1474: \begin{corollary}[Cone Cutting]\label{cor:weightedConeCutting}

1475: Let $G = (V,E,w)$ be a connected weighted graph, let \(v\) be a vertex

1476: in \(V\) and let $S\subseteq V$. Then for any two reals \( 0 \leq

1477: \lambda < \lambda' \), \( \mathtt{ConeCut} (G, v,

1478: \lambda, \lambda', S) \) returns a real \( r \in [\lambda, \lambda')

1479: \) such that

1480: \[

1481: \cost{\bdry{C_S (r, v)}}  ~\leq ~

1482:  \frac{\vol{C_S (r, v)}+\tau }{\lambda' -

1483: \lambda}

1484: \max \left[1,

1485: \log_{2} \frac{m+\tau }{\vol{E (C_S (\lambda, v))}+\tau } \right] ~ ,

1486: \]

1487: where $m = \sizeof{E}$, and

1488: \[

1489: \tau ~ = ~

1490: \left\{

1491: \begin{array}{ll}

1492: 1, & \text{ if $\vol{E (C_S (\lambda, v))} = 0$} \\

1493: 0, & \text{ otherwise.}

1494: \end{array}

1495: \right.

1496: \]

1497: \end{corollary}

1498: 

1499: We will use two other properties of the cones $C_{S} (l,v)$:

1500:   that we can bound their radius (Proposition~\ref{pro:radiusconcentriccone}),

1501:   and that their removal does not increase the radius of the resulting graph

1502:   (Proposition~\ref{pro:deleteconcentriccone}).

1503: 

1504: \begin{proposition}[Radius of Cones]\label{pro:radiusconcentriccone}

1505: Consider a connected weighted graph \( G = (V, E, w) \), a vertex

1506: subset \( S \subseteq V \) and let \( \psi = \max_{v \in V} \dist{v}{S}

1507: \). Then, for every \( x \in S \),

1508: \[

1509: \Gradius{C_{S} (l,x)}{x}

1510:   \leq \psi + 2 l.

1511: \]

1512: \end{proposition}

1513: \begin{proof}

1514: Let $u$ be a vertex in $C_{S} (l,x)$, and let $a_{0}, \dots , a_{k-1}$

1515: and $b_{1}, \dots , b_{k}$ be vertices such that $a_{0} = x$,

1516: $b_{k} = u$, $b_{i+1} \in I_{s} (a_{i})$, $(b_{i}, a_{i}) \in E$, and

1517: \[

1518:   \sum_{i} d(b_{i}, a_{i}) \leq l.

1519: \]

1520: These vertices provide a path connecting $x$ to $u$ inside $C_{S}

1521: (l,x)$ of length at most

1522: \[

1523:   \sum_{i} d(b_{i}, a_{i})

1524: +

1525:   \sum_{i} \dist{a_{i}}{b_{i+1}}.

1526: \]

1527: As the first term is at most $l$, we just need to bound the second

1528:   term by $\psi + l$.

1529: To do this, consider the distance of each of these vertices from

1530:   $S$.

1531: We have the relations

1532: \begin{align*}

1533: \dist{b_{i+1}}{S} & = \dist{a_{i}}{S} + \dist{a_{i}}{b_{i+1}}\\

1534: \dist{a_{i}}{S} & \geq  \dist{b_{i}}{S} - d(b_{i}, a_{i}),

1535: \end{align*}

1536: which imply that

1537: \[

1538:   \psi ~ \geq ~  \dist{b_{k}}{S}

1539:   \geq

1540:   \sum_{i} \dist{a_{i}}{b_{i+1}} - d(b_{i}, a_{i})

1541:   ~ \geq ~

1542:   \left(  \sum_{i} \dist{a_{i}}{b_{i+1}} \right) - l ~ ,

1543: \]

1544: as desired.

1545: \end{proof}

1546: 

1547: In our proof, we actually use

1548: Proposition~\ref{pro:radiusconcentriccone2} which is a slight

1549: extension of Proposition~\ref{pro:radiusconcentriccone}.

1550: It's proof is similar.

1551: %The proof of Proposition~\ref{pro:radiusconcentriccone2}

1552: %  follows the proof of Proposition~\ref{pro:radiusconcentriccone},

1553: %  replacing distances from $S$ with distances from $x_{0}$.

1554: 

1555: 

1556: 

1557: %The cones $\{C_{S} (l,v) \}$

1558: %  form a concentric system

1559: %  (Proposition \ref{pro:concentriccone})

1560: %  with many useful properties.

1561: %The proof of Proposition~\ref{pro:radiusconcentriccone2}

1562: %  appears in Appendix~\ref{sec:coneProperties}.

1563: 

1564: 

1565: \begin{proposition}[Radius of Cones, II]\label{pro:radiusconcentriccone2}

1566: Consider a connected weighted graph \( G = (V, E, w) \), a vertex

1567: \( x_0 \in V \) and let \( \rho = \Gradius{G}{x_{0}} \). Consider a real

1568: \( r_0 < \rho \) and let \( V_0 = B(r_0, x_0) \), \( V' = V - V_0 \)

1569: and \( S = \BallShell{r_0}{x_0} \). Consider a vertex \( x_1 \in S \)

1570: and let \( \psi = \rho - \Gdist{G}{x_{0}}{x_{1}} \). Then the cones

1571: \( C_{S} (l,x_{1}) \) in the graph \( G(V') \) satisfy

1572: \[

1573: \Gradius{C_{S} (l,x_{1})}{x_{1}}

1574:   \leq \psi + 2 l.

1575: \]

1576: \end{proposition}

1577: 

1578: 

1579: \begin{proposition}[Deleting Cones]\label{pro:deleteconcentriccone}

1580: Consider a connected weighted graph \( G = (V, E, w) \), a vertex subset \( S

1581: \subseteq V \), a vertex \( x \in S \) and a real \( l \geq 0 \) and

1582: let \( V' = V - C_{S}(l, x) \), \( S' = S - C_{S}(l, x) \) and \( \psi =

1583: \max_{v \in V} \dist{v}{S} \). Then

1584: \[

1585: \max_{v \in V'}{\Gdist{V'}{v}{S'}} ~ \leq ~ \psi ~ .

1586: \]

1587: \end{proposition}

1588: \begin{proof}

1589: Consider some $v \in V'$.

1590: If the shortest path from $v$ to $S$

1591:   intersects $C_{S} (l,x)$, then

1592:   $v \in C_{S} (l,x)$.

1593: So, the shortest path from $v$ to $S$ in $V$

1594:   must lie entirely in $V'$.

1595: \end{proof}

1596: 

1597: %%Thus we can apply the following cone growing algorithm. Its analysis follows

1598: %%  immediately from Lemma \ref{lem:cuttingconcentricsystem} and

1599: %%  Proposition \ref{pro:concentriccone}.

1600: 

1601: The basic idea of $\mathtt{StarDecomp}$ is to first use

1602:   $\mathtt{BallCut}$ to construct $V_{0}$ and then repeatedly apply

1603:   $\mathtt{ConeCut}$ to construct $\vs{V}{1}{k}$.

1604: %We now  define algorithm $\mathtt{StarDecomp}$.

1605: %Lemma \ref{lem:stardecomposition} follows from

1606: %  Corollaries \ref{cor:weightedBallCutting} and

1607: %\ref{cor:weightedConeCutting}

1608: %  and two other basic properties of the cones $C_{S} (l,W)$:

1609: %  that we can bound their radius (Proposition~\ref{pro:radiusconcentriccone}),

1610: %  and that their removal does not increase the radius of the graph

1611: %  (Proposition~\ref{pro:deleteconcentriccone}).

1612: 

1613: %See Appendix \ref{sec:stardecompApp} for the proof of Lemma

1614: %\ref{lem:stardecomposition}.

1615: 

1616: \begin{algbox}\label{alg:StarDecomp}

1617: \( \left(\{\vs{V}{0}{k},\xx,\yy \} \right) =

1618: \mathtt{StarDecomp} (G = (V, E, w), x_{0}, \delta, \epsilon) \)

1619: \vskip 0.1in

1620: \begin{tightlist2}

1621: \item [1.] Set \( \rho = \Gradius{G}{x_{0}} \); Set \( r_{0} =

1622: \mathtt{BallCut} (G, x_{0}, \rho, \delta) \) and \( V_{0}= B

1623: (r_{0},x_{0}) \);

1624: \item [2.] Let \( S = \BallShell{r_0}{x_0} \);

1625: \item [3.] Set \( G'= (V',E',w') = G (V-V_{0}) \), the weighted graph

1626: induced by $V-V_{0}$;

1627: \item [4.] Set \( \left(\{\vs{V}{1}{k},\xx\} \right) =

1628: \mathtt{ConeDecomp} (G', S, \epsilon \rho / 2) \);

1629: \item [5.] For each $i\in [1:k]$, set $y_{k}$ to be

1630:     a vertex in $V_{0}$

1631:     such that $(x_{k}, y_{k}) \in E$ and $y_{k}$ is on a

1632:     shortest path from $x_{0}$ to $x_{k}$.

1633:     Set $\yy = (\vs{y}{1}{k})$.

1634: \end{tightlist2}

1635: \vskip 0.1in

1636: \( \left(\{\vs{V}{1}{k},\xx \} \right) = \mathtt{ConeDecomp}

1637: (G, S, \Delta) \)

1638: \vskip 0.1in

1639: \begin{tightlist2}

1640: \item [1.] Set $G_{0} =G $, $S_{0} = S$, and $k = 0$.

1641: 

1642: \item [2.] while $S_{k}$ is not empty

1643: \begin{tightlistA}

1644: 

1645: \item [a. ] Set $k = k+1$; Set $x_{k}$ to be a vertex of $S_{k-1}$;

1646:   Set \( r_{k} = \mathtt{ConeCut} (G_{k-1}, x_{k}, 0, \Delta, S_{k-1}) \)

1647: \item [b. ]

1648:   Set \( V_{k} = C_{S_{k-1}} (r_{k}, x_{k}) \); Set \( G_{k} = G(V -

1649: \union_{i=1}^{k}V_{k}) \) and \( S_{k} = S_{k-1}-V_{k} \).

1650: \end{tightlistA}

1651: \item [3.] Set $\xx = (\vs{x}{1}{k})$.

1652: \end{tightlist2}

1653: \end{algbox}

1654: 

1655: \begin{proof}[Proof of Lemma~\ref{lem:stardecomposition}]

1656: %The idea of $\mathtt{StarDecomp}$ is to first use

1657: %  $\mathtt{BallCut}$ to construct $V_{0}$ and then repeatedly apply

1658: %  $\mathtt{ConeCut}$ to construct $\vs{V}{1}{k}$.

1659: Let $\rho  = \Gradius{G}{x_{0}}$.

1660: By setting $\delta = 1/3$, Corollary~\ref{cor:weightedBallCutting}

1661:   guarantees \( \rho / 3 \leq

1662: r_{0} \leq (2 / 3) \rho \).

1663: Applying \( \Delta = \epsilon \rho / 2 \) and Propositions

1664:   \ref{pro:radiusconcentriccone} and \ref{pro:deleteconcentriccone},

1665:   we can bound for every $i$,

1666:   \( r_{0} + d (x_{i},y_{i}) + r_{i} \leq \rho + 2 \Delta =

1667:   \rho + \epsilon \rho \).

1668: Thus \( \mathtt{StarDecomp} (G, x_{0}, 1 / 3, \epsilon) \) returns a

1669:   \( (1 / 3, \epsilon) \)-star-decomposition with center $x_{0}$.

1670: 

1671: To bound the cost of the star-decomposition that the algorithm

1672: produces, we use Corollaries \ref{cor:weightedBallCutting} and

1673: \ref{cor:weightedConeCutting}.

1674: 

1675: \begin{align*}

1676: &\cost{\bdry{V_0}}  \leq  \frac{3 \,(1+ \vol{V_0}) \log_{2} (m+1)}{\rho},

1677: \quad \text{and}\\

1678: &\cost{\edgbet{V_j,V - \cup_{i=0}^{j}V_i}}

1679:     \leq  \frac{2 \, ( 1+\vol{V_j}) \log_{2} (m+1)}{\epsilon \rho}

1680: \end{align*}

1681: for every \( 1 \leq i \leq k \), thus

1682: \begin{align*}

1683: \cost{\bdry{\vs{V}{0}{k}}} & \leq ~ \sum_{j = 0}^{k}

1684: \cost{\edgbet{V_j,V - \cup_{i=0}^{j}V_i}} \\

1685: & \leq ~  \frac{2 \, \log_{2} (m+1)}{\epsilon \rho}

1686:           \sum_{j = 0}^{k} ( \vol{V_j} +1) \\

1687: & \leq ~ \frac{6 \, m \log_{2} (m+1)}{\epsilon \rho} ~ .

1688: \end{align*}

1689: 

1690: To implement $\mathtt{StarDecomp}$ in $O (m + n\log n)$ time,

1691:   we use a Fibonacci heap to implement steps (2) of

1692:   $\mathtt{BallCut}$ and  $\mathtt{ConeCut}$.

1693: If the graph is unweighted, this can be replaced by

1694:   a breadth-first search that requires $O(m)$ time.

1695: \end{proof}

1696: 

1697: %\begin{lemma}[Cone Decomposition]\label{lem:conecutting}

1698: %Let $G = (V,E,w)$ be a weighted graph, and let $S\subseteq V$.

1699: %If $\dist{v}{S}\leq r'$ for all $v \in V$,

1700: %  then for every $\Delta $, there exists

1701: %  a multiway partition of $V$ into $\vs{V}{1}{k}$ such that

1702: %\begin{tightlist2}

1703: %\item

1704: %$\cost{\bdry{\{\vs{V}{1}{k}\}}} \leq 2 m \log_{2} m / \Delta $, and

1705: 

1706: %\item  for each set $V_{j}$, there exists an $s_{j}\in V_{j} \cap S$

1707: %  such that $\dist{v_{j}}{s_{j}}

1708: %  \leq r' + 2\Delta $, for all $v_{j}\in V_{j}$.

1709: %\end{tightlist2}

1710: %\end{lemma}

1711: 

1712: 

1713: %\begin{proof}[Proof of Lemma~\ref{lem:conecutting}]

1714: 

1715: %We apply the following iterative construction.

1716: 

1717: %\begin{itemize}

1718: %\item [0.] Let $G_{0} =G $, and let $U_{0} = V$ and $E_0 = E$ be the

1719: %  vertices and edges of $G_{0}$, respectively.

1720: %  Let $S_{0} = S$. Let $k = 0$.

1721: 

1722: %\item [1.] while $S_{i-1}$ is not empty

1723: %\begin{itemize}

1724: 

1725: %\item [a. ] Let $k = k+1$

1726: %\item [b. ] Let $s_{k}$ be a vertex of $S_{k-1}$;

1727: %\item [c. ] Apply Corollary \ref{cor:weightedConeCutting} to construct

1728: %            $C (l_{k},s_{k})$ and let $V_{k} = C (l_{k},s_{k})$

1729: %($l_k$ is the number whose existence is guaranteed by Corollary

1730: %\ref{cor:weightedConeCutting}).

1731: %\item [d. ]  Let $G_{k} = G_{k-1} (U_{k-1}-V_{k})$ and let $U_{k}$ and

1732: %             $E_{k}$  be the vertices and edges, respectively, of $G_{k}$.

1733: %\end{itemize}

1734: %\item [2.] Return $\vs{V}{1}{k}$

1735: %\end{itemize}

1736: 

1737: %It follows from Corollary \ref{cor:weightedConeCutting},

1738: %\[

1739: %\cost{\bdry{\vs{V}{1}{k}}} \leq \sum_{i=1}^{k} \cost{\bdry{V_{i}}}

1740: %\leq \frac{2\sum_{i}^{k}\vol{V_{i}}\log_{2}m}{\delta }

1741: %\leq \frac{2\sizeof{E}\log_{2}m}{\delta }.

1742: %\]

1743: 

1744: %Let follows from Propositions \ref{pro:radiusconcentriccone} and \ref{pro:deleteconcentriccone}

1745: %  that $\Gradius{G (V_{i})}{s_{i}}\sum r+2\delta $.

1746: %\end{proof}

1747: 

1748: 

1749: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1750: \section{Improving the Stretch}\label{sec:Improved}

1751: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1752: 

1753: In this section, we improve the average stretch

1754:   of the spanning tree to

1755:   $O \left(\log^{2} n \log\log n \right)$ by

1756:   introducing a procedure

1757:   $\mathtt{ImpConeDecomp}$ which refines $\mathtt{ConeDecomp}$.

1758: This new cone decomposition trades off the volume of the cone

1759:   against the cost of edges on its boundary (similar to Seymour \cite{Seym}).

1760: Our refined star decomposition algorithm $\mathtt{ImpStarDecomp}$ is

1761:   identical to algorithm $\mathtt{StarDecomp}$, except that it

1762:   calls

1763: \[

1764: (\{\vs{V}{1}{k},\xx \}) = \mathtt{ImpConeDecomp} (G', S,

1765: \epsilon \rho / 2, t, \mbar)

1766: \]

1767:   at Step 4, where \(t\) is a positive integer that will be defined soon.

1768: 

1769: %Due to  space limitations, we will only state the algorithm and

1770: % a  lemma  about its performance.

1771: %The proof of the lemma can be found in Section \ref{sec:proofImpStar}.

1772: 

1773: \vskip 0.1in

1774: \begin{algbox}\label{alg:impConeDecomp}

1775: \( \left(\{\vs{V}{1}{k},\xx \} \right) =

1776: \mathtt{ImpConeDecomp} (G, S, \Delta, t, \mbar) \)

1777: \vskip 0.1in

1778: \begin{tightlist2}

1779: \item [1.] Set \( G_{0} = G \), \( S_{0} = S \) and \( j = 0 \).

1780: 

1781: \item [2.] while \( S_{j} \) is not empty

1782: \begin{tightlistA}

1783: 

1784: \item [a.] Set \( j = j + 1 \), set \( x_{j} \) to be a vertex of \(

1785: S_{j - 1} \), and set \( p = t - 1 \);

1786: \item [b.] while \( p > 0 \)

1787: \begin{tightlistB}

1788: \item [i.]

1789:     \( r_j = \mathtt{ConeCut} (G_{j - 1}, x_j,

1790:     \frac{(t - p - 1) \Delta}{t}, \frac{(t - p) \Delta}{t}, S_{j - 1}) \);

1791: \item [ii.]

1792:     {\bf if} \( \vol{E(C_{S_{j - 1}}(r_j, x_j))} \leq

1793:     \frac{m}{2^{\log^{p / t} \mbar}} \) {\bf then} {\bf exit} the loop;

1794:     {\bf else}  \( p = p - 1 \);

1795: \end{tightlistB}

1796: \item [c.] Set \( V_j = C_{S_{j - 1}}(r_j, x_j) \) and set \( G_{j} =

1797:            G(V - \union_{i=1}^{j}V_{i}) \) and \( S_{j} = S_{j - 1} - V_{j} \);

1798: \end{tightlistA}

1799: \item [3.] Set $\xx = (\vs{x}{1}{k})$.

1800: \end{tightlist2}

1801: \end{algbox}

1802: \vskip -0.1in

1803: 

1804: \begin{lemma}[Improved Low-Cost Star Decomp]\label{lem:ImpStarDecompn}

1805: Let $G$, $x_0$ and $\epsilon$ be as in Lemma

1806: \ref{lem:stardecomposition}, $t$ be a positive integer control

1807: parameter, and \( \rho = \Gradius{G}{x_{0}} \). Then

1808: \[

1809: (\{V_0,\ldots,V_k\},\xx,\yy) =

1810: \mathtt{ImpStarDecomp}(G, x_0, 1 / 3, \epsilon, t, \mbar) ~ ,

1811: \]

1812: in time $O (m + n \log n)$, returns a  $(1/3, \epsilon)$-star-decomposition of

1813:   $G$ with center $x_0$ that satisfies

1814: \[

1815: \cost{\bdry{V_{0}}} ~ \leq ~ \frac{6 \, \vol{V_{0}} \log_{2} (\mbar +1)}{\rho} ~ ,

1816: \]

1817: and for every index $j \in \{1,2,\ldots,k\}$

1818:    there exists $p = p(j) \in \{0,1,\ldots,t-1\}$ such that

1819: \begin{equation} \label{eq:cost}

1820: \cost{\edgbet{V_j,V - \cup_{i=0}^{j}V_i}}

1821: \leq  t \cdot \frac{4 \, \vol{V_j} \log^{(p + 1) / t} (\mbar+1)}

1822:                     {\epsilon \rho} ~ ,

1823: \end{equation}

1824: and unless \( p = 0 \),

1825: \begin{equation} \label{eq:volume}

1826: \vol{E(V_j)} ~ \leq ~ \frac{m}{2^{\log^{p/t} \mbar}} ~ .

1827: \end{equation}

1828: 

1829: %\begin{equation}

1830: %\label{eq:cost}

1831: %\vol{E(V_j)} ~\le~ {m \over {2^{\log^{p/t} \mbar}}} \quad \mbox{and}

1832: %\quad \cost{\bdry{V_j,V - \cup _{i=0}^{j}V_i}} ~\le~ t \cdot {{4 \vol{V_j}

1833: %  \log^{(p+1)/t} \mbar}\over {\alpha r}}.

1834: %\end{equation}

1835: \end{lemma}

1836: \begin{proof}

1837: In what follows, we call \( p(j) \) the \emph{index-mapping} of

1838: the vertex set \( V_j \).

1839: We begin our proof by observing that \( 0 \leq r_j < \epsilon \rho

1840: / 2 \) for every \( 1 \leq j \leq k \).

1841:  We can then show that \( \{\vs{V}{0}{k}\} \)

1842:   is a \( (1 / 3, \epsilon) \)-star decomposition as we did in the proof of

1843: Lemma \ref{lem:stardecomposition}.

1844: 

1845: We now bound the cost of the decomposition.

1846: Clearly, the bound on \( \cost{\bdry{V_{0}}} \) remains

1847:   unchanged from that proved in Lemma

1848:   \ref{lem:stardecomposition}, but here we bound

1849:   $\vol{V_{0}} + 1$ by $2 \vol{V_{0}}$.

1850: 

1851: Below we will use \( \Delta = \epsilon \rho / 2 \) as specified in the algorithm.

1852: 

1853: Fix an index \( j \in \{1,2,\ldots,k\} \), and let \( p = p(j) \) be

1854: the final value of variable \(p\) in the loop above (that is, the

1855: value of \(p\) when the execution left the loop while constructing \(

1856: V_j \)). Observe that \( p \in \{0,1,\ldots,t-1\} \), and that unless the

1857: loop is aborted due to \( p = 0 \), we have \( \vol{E(V_j)} \leq

1858: \frac{m}{2^{\log^{p/t} \mbar}} \) and inequality \eqref{eq:volume} holds.

1859: 

1860: For inequality \eqref{eq:cost}, we split the discussion to two

1861: cases. First, consider the case \( p = t - 1 \).

1862: Then the inequality

1863:   \( \cost{\edgbet{V_j, V - \cup_{i=0}^{j}V_i}} \leq

1864:   (\vol{V_j}+1)  \log (\mbar +1) (t/\Delta) \) follows directly from Corollary

1865:   \ref{cor:weightedConeCutting}, and inequality \eqref{eq:cost} holds.

1866: 

1867: Second, consider the case \( p < t - 1 \) and let \( r_{j}' \) be the

1868: value of the variable \( r_j \) at the beginning of the last

1869: iteration of the loop (before the last invocation of Algorithm

1870: \texttt{ConeCut}). In this case, observe that at the beginning of the

1871: last iteration, \( \vol{E(C_{S_{j - 1}}(r_{j}', x_j))} >

1872: \frac{m}{2^{\log^{(p + 1) / t} \mbar}} \) (as otherwise the loop

1873: would have been aborted in the previous iteration).

1874: By Corollary \ref{cor:weightedConeCutting},

1875: \[

1876:  \cost{\edgbet{V_j, V - \cup_{i=0}^{j}V_i}}

1877:  ~\leq ~

1878:   \frac{\vol{V_j}}{\Delta / t} ~ \times

1879: \max \left[1,\log_{2}{\left( \frac{m}{\vol{E \left(

1880: C_{S_{j - 1}} \left( \frac{(t - p - 1) \Delta}{t}, x_j \right) \right)}}

1881: \right)} \right],

1882: \]

1883: where \( V_j = C_{S_{j - 1}}(r_j, x_j) \).

1884: Since

1885: \[

1886: \frac{(t - p - 2)

1887: \Delta}{t} \leq r_{j}' < \frac{(t - p - 1) \Delta}{t},

1888: \]

1889: it follows that

1890: \[

1891: \vol{E \left( C_{S_{j - 1}} \left( \frac{(t - p - 1) \Delta}{t}, x_j

1892: \right) \right)}

1893: ~ \geq ~ \vol{E(C_{S_{j - 1}}(r_{j}', x_j))} ~ > ~

1894: \frac{m}{2^{\log^{(p + 1) / t} \mbar}} ~ .

1895: \]

1896: Therefore

1897: \[

1898: \log^{(p + 1) / t} \mbar ~ \geq ~

1899: \max \left[1,\log_{2}{\left( \frac{m}{\vol{E \left(

1900: C_{S_{j - 1}} \left( \frac{(t - p - 1) \Delta}{t}, x_j \right) \right)}}

1901: \right)} \right] ~ ,

1902: \]

1903: and

1904: \[

1905: \cost{\edgbet{V_j, V - \cup_{i=0}^{j}V_i}}  ~ \leq ~ \frac{\vol{V_j}

1906: \log^{(p + 1) / t} \mbar}{\Delta / t}

1907:  ~ = ~ t \cdot \frac{2 \, \vol{V_j} \log^{(p + 1) / t}

1908: \mbar}{\epsilon \rho} ~ .

1909: %\qed

1910: \]

1911: \end{proof}

1912: 

1913: Our improved algorithm \( \mathtt{ImpLowStretchTree} (G, x_{0}, t, \mbar) \),

1914: is identical to \texttt{LowStretchTree} except that in Step 3 it calls

1915: $\mathtt{ImpStarDecomp} (\Contracted{G}, x_{0}, 1/3, \beta, t,

1916: \mbar),$ and in Step 5 it calls \AvoidOverfull{}

1917: $\mathtt{ImpLowStretchTree} (G (V_{i}), x_{i}, t, \mbar)$. We set $t =

1918: \log\log n$ throughout the execution of the algorithm.

1919: %The proof of the following theorem is deferred to Section

1920: %\ref{sec:proofImproved}.

1921: 

1922: \begin{theorem}[Lower--Stretch Spanning Tree] \label{thm:main2}

1923: Let \( G = (V, E, w) \) be a connected weighted graph and let \(x_0\)

1924: be a vertex in \(V\). Then

1925: \[

1926: T = \mathtt{ImpLowStretchTree} (G, x_{0}, t, \mbar) ~ ,

1927: \]

1928: in time \( O (\widehat{m} \log \widehat{n} + \widehat{n} \log^2

1929: \widehat{n}) \), returns a spanning tree of \(G\) satisfying

1930: \[

1931: \Gradius{T}{x_0} \leq 2 \sqrt{e} \cdot  \Gradius{G}{x}

1932: \]

1933: and

1934: \[

1935: \avestretch{T}{E} = O \left(\log^{2} \widehat{n} \log\log \widehat{n}

1936: \right) ~ .

1937: \]

1938: %For every weighted connected graph $G = (V,E,w)$ on $n$ vertices and

1939: %  $m$ edges, \texttt{ImpLowStretchTree}, in $O (m\log^2 n)$ time,

1940: %  constructs a spanning tree $T$ satisfying

1941: %\( \Gradius{T}{x} \leq 2 \, \sqrt{e} \cdot \Gradius{G}{x} \), and \(

1942: %\stretch{T}{E} = O \left(\log^{2} n\log\log n \right) \).

1943: \end{theorem}

1944: \begin{proof}

1945: %We first apply a linear-time transformation from \cite{AKPW} that

1946: %  transforms the graph into one with at most $n (n + 1)$ edges,

1947: %  and such that the average-stretch of the spanning tree on the

1948: %  original graph will be at most twice the average-stretch on this graph.

1949: 

1950: The bound on the radius of the tree remains unchanged from that proved

1951: in Theorem~\ref{thm:main}.

1952: 

1953: We begin by defining a system of notations for the recursive process,

1954: assigning for every graph \( G = (V, E) \) input to some recursive

1955: invocation of Algorithm \texttt{ImpLowStretchTree}, a sequence \(

1956: \GraphSequence{G} \) of non-negative integers. This is done as

1957: follows. If \(G\) is the original graph input to the first invocation

1958: of the recursive algorithm, than \( \GraphSequence{G} \) is

1959: empty. Assuming that the halt condition of the recursion is not

1960: satisfied for \(G\), the algorithm continues and some of the edges in

1961: \(E\) are contracted. Let \( \Contracted{G} = (\Contracted{V},

1962: \Contracted{E}) \) be the resulting graph. (Recall that we refer to \(

1963: \Contracted{G} \) as the edge-contracted graph.) Let \( \{

1964: \Contracted{V}_0, \Contracted{V}_1, \dots , \Contracted{V}_k \} \) be

1965: the star decomposition of \( \Contracted{G} \). Let \( V_j \in V \) be

1966: the preimage under edge contraction of \( \Contracted{V}_j \in

1967: \Contracted{V} \) for every \( 0 \leq j \leq k \). The graph \( G(V_j)

1968: \) is assigned the sequence \( \GraphSequence{G({V}_{j})} =

1969: \GraphSequence{G} \cdot j \). Note that \( | \GraphSequence{G} | = h

1970: \) implies that the graph \(G\) is input to the recursive algorithm on

1971: recursion level \(h\). We warn the reader that the edge-contracted

1972: graph obtained from a graph assigned with the sequence \(\sigma\) may

1973: have fewer edges than the edge-contracted graph obtained from the

1974: graph assigned with the sequence \( \sigma \cdot j \), because the

1975: latter may contain edges that were contracted out in the former.

1976: 

1977: We say that the edge \(e\) is \emph{present} at recursion level \(h\)

1978: if \(e\) is an edge in \( \Contracted{G} \) which is the

1979: edge-contracted graph obtained from some graph \(G\) with \( |

1980: \GraphSequence{G} | = h \) (that is, it was not contracted out). An

1981: edge $e$ \emph{appears} at the first level \(h\) at which it is

1982: present, and it \emph{disappears} at the level at which it is present

1983: and its endpoints are separated by the star decomposition. If an edge

1984: appears at recursion level \(h\), then a path connecting its endpoints

1985: was contracted on every recursion level smaller than \(h\), and no

1986: such path will be contracted on any recursion level greater than

1987: \(h\). Moreover, an edge is never present at a level after it

1988: disappears. We define \( h(e) \) and \( h'(e) \) to be the recursion

1989: levels at which the edge \(e\) appears and disappears, respectively.

1990: 

1991: For every edge $e$ and every recursion level $i$ at which it is

1992: present, we let $U (e,i)$ denote the set of vertices $\Contracted{V}$

1993: of the edge-contracted graph containing its endpoints.

1994: If $h (e) \leq i < h' (e)$, then we let $W (e,i)$ denote the set of vertices

1995:   $\Contracted{V}_{j}$ output by \texttt{ImpStarDecomposition}

1996:   that contains the endpoints of $e$.

1997: 

1998: Recall that \( p(j) \) denote the index-mapping of the vertex set \(

1999: V_j \) in the star decomposition. For each index \( i \in

2000: \{0,1,\ldots,t-1\} \), let \( I_i = \{j \in \{1,2,\ldots,k\} \mid p(j)

2001: = i\} \). For a vertex subset \( U \subseteq V \), let \( \AS{U} \) denote the

2002: average stretch that the algorithm guarantees for the edges of

2003: $E(U)$. Let \( \TS{U} = \AS{U} \cdot | E(U) | \). Then by Lemma

2004: \ref{lem:ImpStarDecompn}, the following recursive formula applies.

2005: \begin{eqnarray}

2006: \TS{V} & \leq & \left( \sum_{j = 0}^k \TS{\Contracted{V}_j} \right)

2007: \nonumber \\

2008: & + & 4 \sqrt{e} \times \left( 6 \, \log (\widehat{m} + 1) \cdot

2009: \vol{\Contracted{V}_0} ~ + ~ 4 \, \frac{t}{\beta} \sum_{p = 0}^{t - 1}

2010: \log^{(p + 1) / t} ( \mbar+1) \sum_{j \in I_p} \vol{\Contracted{V}_j}

2011: \right) \nonumber \\

2012: & + & \left( \sum_{e \in E - \Contracted{E}}{\stretch{T}{e}} \right)

2013: \nonumber \\

2014: & = & \left( \sum_{j = 0}^k \TS{V_j} \right) \nonumber \\

2015: & + & 4 \sqrt{e} \times \left( 6 \, \log (\widehat{m} + 1) \cdot

2016: \vol{\Contracted{V}_0} ~ + ~ 4 \, \frac{t}{\beta} \sum_{p = 0}^{t - 1}

2017: \log^{(p + 1) / t} ( \mbar+1) \sum_{j \in I_p} \vol{\Contracted{V}_j}

2018: \right) ~ , \label{eq:TS}

2019: \end{eqnarray}

2020: where we recall $\beta = \epsilon = (2 \log_{4/3} (n + 32))^{-1}$.

2021: 

2022: For every edge \(e\) and for every \( h(e) \leq i < h'(e) \), let

2023: \( \pi_i(e) \) denote the index-mapping of the component \(

2024: W(e, i) \) in the invocation of Algorithm \texttt{ImpConeDecomp}

2025: on recursion level \(i\). For every index \( p \in \{ 0, \dots , t - 1

2026: \} \), define the variable \( l_p(e) \) as follows

2027: \[

2028: l_p(e) ~ = ~ \left| \left\{ h(e) \leq i < h'(e) \mid \pi_i(e) = p

2029: \right\} \right| ~ .

2030: \]

2031: 

2032: For a fixed edge \(e\) and an index

2033:   \( p \in \{ 0, \dots , t - 1 \} \),

2034:   every \( h(e) \leq i < h'(e) \) such that \( \pi_i(e) = p \)

2035:    reflects a contribution of

2036:   \( O (t / \beta) \cdot \log^{(p + 1) / t} (\mbar + 1) \)

2037:   to the right term in

2038:   \eqref{eq:TS}.

2039: Summing $p$ over $\setof{0, 1, \ldots , t-1}$, we obtain

2040: \[

2041: \sum _{p=0}^{t-1} O (t / \beta) l_{p} (e)\log^{(p + 1) / t} (\mbar + 1).

2042: \]

2043: In a few moments, we will prove that

2044: \begin{equation}\label{eqn:sumLp}

2045: \sum _{e} \sum _{p=0}^{t-1} l_{p} (e)\log^{p  / t} (\mbar + 1)

2046:   \leq

2047: O (\mbar \log_{2} \mbar),

2048: \end{equation}

2049: which implies that the sum of the contributions of all edges $e$ in levels

2050:   \( h(e) \leq i < h'(e) \) to the right term in

2051:   \eqref{eq:TS}

2052:  is

2053: \[

2054:  O \left(

2055: \frac{t}{\beta} \cdot \mbar \log^{1 + 1 / t} \mbar \right)

2056: =

2057:  O \left(

2058: t \cdot \mbar \log^{2 + 1 / t} \mbar \right)

2059: \]

2060: 

2061: As  \( \vol{V_j} \) counts the internal edges of \(V_j \) as well as

2062:   its boundary edges, we must also account for the contribution of

2063:   each edge $e$ at level $h' (e)$.

2064: At this level, it will be counted twice---once in each component containing

2065:   one of its endpoints.

2066: Thus, at this stage, it contributes a factor of at most

2067:   \( O ((t/\beta ) \cdot \log \mbar) \) to the sum \( \TS{V} \).

2068: Therefore all edges

2069:   \( e \in E \)

2070: contribute an additional factor of

2071:  \( O (t \cdot \mbar \log^2 \mbar) \).

2072: Summing over all the edges, we find that all the contributions to

2073:   the right term

2074:   in

2075:   \eqref{eq:TS} sum to at most

2076: \[

2077:  O \left(

2078: t \cdot \mbar \log^{2 + 1 / t} \mbar \right) .

2079: \]

2080: 

2081: Also, every \( h(e) \leq i < h'(e) \) such that the edge \(e\) belongs

2082: to the central component \( \Contracted{V}_0 \) of the star

2083: decomposition, reflects a contribution of \( O (\log \mbar) \) to the

2084: left term in \eqref{eq:TS}. Since there are at most \( O (\log \mbar) \) such

2085: \(i\)s, it follows that the contribution of the left term in

2086: \eqref{eq:TS} to \( \TS{V} \) sums up to an additive term of \( O

2087: (\log^2 \mbar) \) for every single edge, and \( (\mbar \log^2 \mbar) \)

2088: for all edges.

2089: 

2090: It follows that \( \TS{V} = O (t \cdot \mbar \log^{2 + 1 / t}

2091: \mbar) \). This is optimized by setting \( t = \log\log \mbar \),

2092: obtaining the desired upper bound of \( O (\log^{2} \widehat{n} \cdot \log\log

2093: \widehat{n}) \) on the average stretch \( \AS{V} \) guaranteed by Algorithm

2094: \texttt{ImpLowStretchTree}.

2095: 

2096: We now return to the proof of \eqref{eqn:sumLp}.

2097: We first note that  \( l_0(e) \) is at

2098:   most \( O (\log \mbar) \) for every edge \(e\).

2099: We then observe that for each index

2100:   \( p > 0 \) and each \( h(e) \leq i < h'(e) \) such that \( \pi_i(e) = p  \),

2101:   $\vol{E(U (e,i))}/\vol{E(W (e,i))}$

2102:   is at least

2103:   \( 2^{\log^{p / t} \mbar} \) (by Lemma

2104:   \ref{lem:ImpStarDecompn}, \eqref{eq:volume}).

2105: For $h (e) \leq  i < h' (e)$, let

2106:   $g_{i} (e) = \vol{E(U (e,i+1))} / \vol{E(W (e,i))}$.

2107: We then have

2108: \[

2109: \prod _{1 \leq p \leq t-1}

2110: \left( 2^{\log^{p / t} \mbar} \right)^{l_p(e)} ~

2111: \leq ~ \mbar \prod _{h (e) \leq  i < h' (e)} g_{i} (e) ~ ,

2112: \]

2113: hence \( \sum _{p=1}^{t-1} l_p(e) \log^{p / t} \mbar \leq \log

2114: \mbar + \sum _{h (e) \leq  i < h' (e)} \log g_{i} (e)

2115:  \).

2116: We will next prove that

2117: \begin{equation}\label{eqn:sumgis}

2118: \sum _{e} \sum _{h (e) \leq  i < h' (e)} \log g_{i} (e)

2119: \leq \mbar \log \mbar,

2120: \end{equation}

2121: which implies \eqref{eqn:sumLp}.

2122: 

2123: Let $E_{i}$ denote the set of edges present at recursion level $i$.

2124: For every edge $e \in E_{i}$ such that $i < h' (e)$, we have

2125: \[

2126:   \sum _{e' \in E(W (e,i))} g_{i} (e') = g_{i} (e) \vol{E(W (e,i))}

2127:  = \vol{E(U (e,i+1))},

2128: \]

2129: and so

2130: $  \sum _{e \in E_{i} : i < h' (e) } g_{i} (e) = \vol{E_{i+1}}.$

2131: As each edge is present in at most $O (\log \mbar)$ recursion depths,

2132: $\sum _{i} \vol{E_{i}} \leq \mbar \log \mbar$, which proves \eqref{eqn:sumgis}.

2133: \end{proof}

2134: 

2135: %\begin{proof}

2136: %The proof is analogous  to that of Lemma \ref{lem:stardecomposition},

2137: %but Lemma \ref{lem:imp_conecutting} is used

2138: %instead

2139: %Lemma \ref{lem:conecutting}.

2140: %Algorithm $\mathtt{ImpStarDecomp}$ is identical to algorithm

2141: %$\mathtt{StarDecomp}$, except that it invokes the refined star

2142: %decomposition procedure

2143: %(that uses $\mathtt{ImpConeDecomp}$)

2144: %instead of the original one.

2145: %Also, algorithm $\mathtt{ImpStarDecomp}$ delivers the new parameter

2146: %$\mbar$ unchanged to the recursive invocations. (In the very first

2147: %root invocation, $\mbar = m$, but in all the other recursive

2148: %invocations, $m$ is the size of the edge set on which the

2149: %procedure is invoked, while $\mbar$ is the size of the edge set

2150: %of the entire original graph.)

2151: %\end{proof}

2152: 

2153: 

2154: %Remark: all the logarithms in this section have base 2.

2155: 

2156: %\begin{lemma}

2157: %\label{lem:imp_conecutting}

2158: %For $G$, $S$, $r'$ and $\Delta$ as in Lemma \ref{lem:conecutting},

2159: %a positive integer parameter $t$, and $\mbar \ge m = |E|$,

2160: %there exists a multiway partition of $V$ into $V_1,\ldots,V_k$

2161: %such that

2162: %\begin{itemize}

2163: %\item

2164: %for every index $j \in \{1,2,\ldots,k\}$ there exists

2165: %$p = p(j) \in \{0,1,\ldots,t-1\}$ such that

2166: %$\vol{E(V_j)} \le {m \over {2^{\log^{p/t} \mbar}}}$, {\bf and}

2167: %\begin{equation}

2168: %\label{eq:cost}

2169: %\cost{\bdry{V_j,V - S - V_j}} ~\le~

2170: %t \cdot {{2 \vol{V_j}} \over \Delta} \cdot \log^{(p+1)/t} \mbar

2171: %\end{equation}

2172: %\item

2173: % for each set $V_{j}$, there exists an $s_{j}\in V_{j} \cap S$

2174: %  such that $\dist{v_{j}}{s_{j}}

2175: %  \leq r' + 2\Delta $, for all $v_{j}\in V_{j}$.

2176: %\end{itemize}

2177: %\end{lemma}

2178: %\begin{proof}

2179: %An iterative procedure  similar to the one described in the proof

2180: %of Lemma \ref{lem:conecutting} is applied.

2181: %The only difference is in step

2182: %1(c)

2183: %%

2184: %%\ref{step:cone},

2185: %%For some reason I could not make this ref' to work.

2186: %%

2187: %which now grows into the following loop.

2188: %(This variant of the cone decomposition algorithm will be referred to

2189: %as $\mathtt{ImpConeDecomp}$.)

2190: 

2191: %We next show that the cone decomposition that is constructed by this

2192: %modified iterative procedure satisfies the statement of the lemma.

2193: 

2194: %\end{proof}

2195: 

2196: 

2197: %{\bf 6   Open Questions}

2198: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2199: \section{Conclusion}\label{sec:conclusion}

2200: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2201: 

2202: At the beginning of the paper, we pointed out that the definition

2203:   of stretch used in this paper differs slightly from that used

2204:   by Alon, Karp, Peleg and West~\cite{AKPW}.

2205: If one is willing to accept a longer running time, then this problem

2206:   is easily remedied as shown in Subsection \ref{subsec:applications}.

2207: If one is willing to accept a bound of $O (\log ^{3} n)$ on the stretch,

2208:   then one can extend our analysis to show that the natural randomized

2209:   variant \texttt{LowStretchTree},  in which

2210:   one chooses the radii of the balls and cones at random, works.

2211: 

2212: A natural open question is whether one can improve the stretch bound

2213:    from $O\left(\log^{2} n\log\log n \right)$ to $O

2214:   (\log n)$.

2215: Algorithmically, it is also desirable to improve the running time of

2216:   the algorithm to $O (m\log n)$.

2217: If we can successfully achieve both improvements, then

2218:    we can use the Spielman-Teng solver to solve planar diagonally dominant

2219:    linear systems in $O (n\log n \log (1/\epsilon ))$ time.

2220: 

2221: As the average stretch\footnote{In the context of the optimization

2222: problem of finding a spanning tree with the lowest average stretch,

2223: the stretch is defined as in \cite{AKPW}.} of any spanning tree in a weighted connected graph is $\Omega (1)$, our low-stretch tree algorithm also provides

2224:   an $O\left(\log^{2} n\log\log n \right)$-approximation

2225:   to the optimization problem of finding

2226:   the spanning tree with the lowest average stretch.

2227: It remains open (a) whether our algorithm has a better approximation

2228:   ratio and (b) whether one can in polynomial time find a spanning tree

2229:   with better approximation ratio, e.g., $O (\log n)$ or even $O (1)$.

2230: 

2231: 

2232: 

2233: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2234: \section{Acknowledgments}\label{sec:ack}

2235: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2236: We are indebted to David Peleg, who was offered co-authorship on this paper.

2237: 

2238: \begin{thebibliography}{10}

2239: 

2240: \bibitem{AKPW}

2241: Noga Alon, Richard~M. Karp, David Peleg, and Douglas West.

2242: \newblock A graph-theoretic game and its application to the {$k$}-server

2243:   problem.

2244: \newblock {\em SIAM Journal on Computing}, 24(1):78--100, February 1995.

2245: 

2246: \bibitem{AumannRabani}

2247: Yonatan Aumann and Yuval Rabani.

2248: \newblock An o(log k) approximate min-cut max-flow theorem and approximation

2249:   algorithm.

2250: \newblock {\em SIAM J. Comput.}, 27(1):291--301, 1998.

2251: 

2252: \bibitem{AwerbuchAzar}

2253: B.~Awerbuch and Y.~Azar.

2254: \newblock Buy-at-bulk network design.

2255: \newblock In {\em Proceedings of the 38th IEEE FOCS}, pages 542--547, 1997.

2256: 

2257: \bibitem{Awerbuch}

2258: Baruch Awerbuch.

2259: \newblock Complexity of network synchronization.

2260: \newblock {\em J. ACM}, 32(4):804--823, 1985.

2261: 

2262: \bibitem{Bartal1}

2263: Yair Bartal.

2264: \newblock Probabilistic approximation of metric spaces and its algorithmic

2265:   applications.

2266: \newblock In {\em Proceedings of the 37th IEEE FOCS}, pages 184--193, 1996.

2267: 

2268: \bibitem{Bartal2}

2269: Yair Bartal.

2270: \newblock On approximating arbitrary metrices by tree metrics.

2271: \newblock In {\em Proceedings of the 30th ACM STOC}, pages 161--168, 1998.

2272: 

2273: 

2274: \bibitem{BartalComm}

2275: Yair Bartal.

2276: \newblock Personal Communication, 2005.

2277: 

2278: \bibitem{BomanHendricksonAKPW}

2279: Erik Boman and Bruce Hendrickson.

2280: \newblock On spanning tree preconditioners.

2281: \newblock Manuscript, Sandia National Lab., 2001.

2282: 

2283: \bibitem{BHV}

2284: Erik Boman, Bruce Hendrickson, and Stephen Vavasis.

2285: \newblock Solving elliptic finite element systems in near-linear time with

2286:   support preconditioners.

2287: \newblock Manuscript, Sandia National Lab. and Cornell,

2288:   http://arXiv.org/abs/cs/0407022.

2289: 

2290: \bibitem{NPKann}

2291: P.~Crescenzi and V.~Kann.

2292: \newblock A compendium of {NP-hard} problems.

2293: \newblock Available online at http://www.nada.kth.se/theory/compendium, 1998.

2294: 

2295: \bibitem{FRT}

2296: Jittat Fakcharoenphol, Satish Rao, and Kunal Talwar.

2297: \newblock A tight bound on approximating arbitrary metrics by tree metrics.

2298: \newblock In {\em Proceedings of the 35th  ACM STOC}, pages 448--455, 2003.

2299: 

2300: \bibitem{GareyJohnson}

2301: M.R. Garey and D.S. Johnson.

2302: \newblock {\em Computers and Intractability: a Guide to Theory of

2303:   NP-Completeness}.

2304: \newblock 1979.

2305: 

2306: \bibitem{GargKonjevodRavi}

2307: N.~Garg, G.~Konjevod, and R.~Ravi.

2308: \newblock A polylogarithmic approximation algorithm for the group steiner tree

2309:   problem.

2310: \newblock In {\em Proceedings of the 9th ACM-SIAM SODA}, pages 253--259, 1998.

2311: 

2312: \bibitem{Hu}

2313: T.C. Hu.

2314: \newblock Optimum communication spanning trees.

2315: \newblock {\em SIAM Journal on Computing}, pages 188--195, 1974.

2316: 

2317: \bibitem{KonjRavi}

2318: G.~Konjevod and R.~Ravi.

2319: \newblock An approximation algorithm for the covering Steiner problem.

2320: \newblock In {\em Proceedings of the 11th ACM-SIAM SODA}, pages 338--344, 2000.

2321: 

2322: \bibitem{LeightonRao}

2323: Tom Leighton and Satish Rao.

2324: \newblock Multicommodity max-flow min-cut theorems and their use in designing

2325:   approximation algorithms.

2326: \newblock {\em J. ACM}, 46(6):787--832, 1999.

2327: 

2328: \bibitem{LLR}

2329: Nathan Linial, Eran London, and Yuri Rabinovich.

2330: \newblock The geometry of graphs and some of its algorithmic applications.

2331: \newblock {\em Combinatorica}, 15:215--245, 1995.

2332: 

2333: \bibitem{PelegBook}

2334: D.~Peleg.

2335: \newblock {\em Distributed Computing: A Locality-Sensitive Approach}.

2336: \newblock 2000.

2337: \newblock SIAM Philadelphia PA.

2338: 

2339: \bibitem{PelReshef}

2340: D.~Peleg and E.~Reshef.

2341: \newblock Deterministic polylogarithmic approximation for minimum communication

2342:   spanning trees.

2343: \newblock In {\em Proc. 25th International Colloq. on Automata, Languages and

2344:   Programming}, pages 670--681, 1998.

2345: 

2346: \bibitem{Seym}

2347: P.~D.~Seymour.

2348: \newblock Packing directed circuits fractionally.

2349: \newblock {\em Combinatorica}, 15(2):281--288, 1995.

2350: 

2351: \bibitem{SpielmanTengPrecon}

2352: Daniel~A. Spielman and Shang-Hua Teng.

2353: \newblock Nearly-linear time algorithms for graph partitioning, graph

2354:   sparsification, and solving linear systems.

2355: \newblock In {\em Proceedings of the 36th ACM STOC}, pages 81--90, 2004.

2356: 

2357: \end{thebibliography}

2358: 

2359: 

2360: \end{document}

2361: