1: \documentclass[12pt]{article}

2: \usepackage{epsfig,latexsym,graphicx}

3: 

4: \title{Coin-Moving Puzzles}

5: \author{%

6:   Erik D. Demaine%

7:     \thanks{Department of Computer Science, University of Waterloo, Waterloo,

8:             Ontario N2L 3G1, Canada, email: \{\texttt{eddemaine},

9:             \texttt{mldemaine}\}\texttt{@uwaterloo.ca}}

10: \and

11:   Martin L. Demaine%

12:     \footnotemark[1]

13: \and

14:   Helena A. Verrill%

15:     \thanks{Institut for Matematiske Fag, Universitetsparken 5,

16:             DK-2100 K\o benhavn, Denmark, email: \texttt{verrill@math.ku.dk}}}

17: \date{}

18: 

19: % 1-inch margins, from fullpage.sty by H.Partl, Version 2, Dec. 15, 1988.

20: \topmargin 0pt

21: \advance \topmargin by -\headheight

22: \advance \topmargin by -\headsep

23: \textheight 8.9in

24: \oddsidemargin 0pt

25: \evensidemargin \oddsidemargin

26: \marginparwidth 0.5in

27: \textwidth 6.5in

28: 

29: % Avoid line breaks before citations (\cite) and references (\ref)

30: \let\latexcite=\cite

31: \def\cite{\nolinebreak\latexcite}

32: \let\latexref=\ref

33: \def\ref{\nolinebreak\latexref}

34: 

35: % Set the default figure placement to [!htbp].  Relies on LaTeX internals.

36: {\makeatletter \gdef\fps@figure{!htbp}}

37: 

38: % Theorem environments

39: \newtheorem{theorem}{Theorem}

40: \newtheorem{lemma}{Lemma}

41: \newtheorem{corollary}{Corollary}

42: \newtheorem{conjecture}{Conjecture}

43: 

44: % Proof environment, with an optional argument to \begin{proof} for specifying

45: % what you are proving.  For example: \begin{proof}[Theorem \ref{fermat}].

46: \makeatletter

47: \def\GrabProofArgument[#1]{ (#1): \egroup\ignorespaces}

48: \def\proof{\noindent\textbf\bgroup Proof%

49:            \@ifnextchar[{\GrabProofArgument}{: \egroup\ignorespaces}}

50: \def\endproof{\hspace*{\fill}$\Box$\medskip}

51: \makeatother

52: 

53: % Complex \xxx for making notes of things to do.  Use \xxx{...} for general

54: % notes, and \xxx[who]{...} if you want to blame someone in particular.

55: {\makeatletter

56:  \gdef\xxx{\@ifnextchar[\xxx@lab\xxx@nolab}

57:  \long\gdef\xxx@lab[#1]#2{{\bf [\marginpar{xxx}{\sc #1}: #2]}}

58:  \long\gdef\xxx@nolab#1{{\bf [\marginpar{xxx}#1]}}

59:  % This turns them off:

60:  \long\gdef\xxx@lab[#1]#2{}\long\gdef\xxx@nolab#1{}%

61: }

62: 

63: % Put figures and text together

64: \def\textfraction{0.01}

65: \def\topfraction{0.99}

66: \def\floatpagefraction{0.99}

67: \def\dblfloatpagefraction{0.99}

68: 

69: % Customize the fonts in the captions.  \def\captionfont to whatever you want,

70: % for example \em, \small, \em\small, \sf, or \sf\small.  Ditto for

71: % \captionlabelfont (the Figure or Table part of the caption).

72: % Relies on LaTeX internals.

73: \def\captionfont{\sf\small}

74: \def\captionlabelfont{\bf\small}

75: {\makeatletter

76:  \global\let\old@makecaption\@makecaption

77:  \long\gdef\@makecaption#1#2{%

78:    \old@makecaption{\captionlabelfont #1}{\captionfont #2}}}

79: 

80: %--------

81: % Specific to this document

82: \def\span{\mathop{\rm span}}

83: \def\applymove{\,/\,}

84: 

85: \def\coinscale{0.75}

86: 

87: 

88: % Don't want to change all the minus signs to \setminus... xxx

89: \let\setminus=-

90: 

91: 

92: \begin{document}

93: \maketitle

94: 

95: \begin{abstract}

96: We introduce a new family of one-player games, involving the movement of coins

97: from one configuration to another.  Moves are restricted so that a coin can be

98: placed only in a position that is adjacent to at least two other coins.  The

99: goal of this paper is to specify exactly which of these games are solvable.

100: By introducing the notion of a constant number of extra coins, we give tight

101: theorems characterizing solvable puzzles on the square grid and

102: equilateral-triangle grid.  These existence results are supplemented by

103: polynomial-time algorithms for finding a solution.

104: %

105: %---Old:

106: % We analyze a family of one-player games, involving the movement of coins from

107: % one configuration to another.  Moves are restricted so that a coin can be

108: % placed only in a position that is adjacent to at least two other coins.  The

109: % goal of this paper is to specify exactly which of these games are solvable.  We

110: % solve this problem with a polynomial-time algorithm for two boards, the

111: % equilateral-triangle and square grids, and consider the problem on a general

112: % graph.

113: \end{abstract}

114: 

115: \section{Introduction}

116: 

117: Consider a configuration of coins such as the one on the left of Figure

118: \ref{rhomboid}.  The player is allowed to move any coin to a position that is

119: determined rigidly by incidences to other coins.  In other words, a coin can be

120: moved to any position adjacent to at least two other coins.  The puzzle or

121: 1-player game is to reach the configuration on the right of Figure

122: \ref{rhomboid} by a sequence of such moves.  This particular puzzle is most

123: interesting when each move is restricted to \emph{slide} a coin in the plane

124: without overlapping other coins.

125: 

126: \begin{figure}

127: \centerline{\includegraphics[scale=\coinscale]{rhomboid.eps}}

128: \caption{\label{rhomboid}

129:   Re-arrange the rhombus into the circle using three slides,

130:   such that each coin is slid to a position adjacent to two other coins.}

131: \end{figure}

132: 

133: This puzzle is described in Gardner's Mathematical Games article on Penny

134: Puzzles \cite{Gardner-1975-penny}, in \emph{Winning Ways}

135: \cite{Berlekamp-Conway-Guy-1982-coins}, in \emph{Tokyo Puzzles}

136: \cite{Fujimura-1978-coins}, in \emph{Moscow Puzzles}

137: \cite{Kordemsky-1972-coins}, and in \emph{The Penguin Book of Curious and

138: Interesting Puzzles} \cite{Wells-1992-coins}.  Langman \cite{Langman-1951}

139: shows all 24 ways to solve the puzzle in three moves.  Another classic puzzle

140: of this sort \cite{Bolt-1984-pennies, Fujimura-1978-coins, Gardner-1975-penny,

141: Wells-1992-coins} is shown in Figure \ref{upside-down}.  A final classic puzzle

142: that originally motivated our work is shown in Figure \ref{original helena};

143: its source is unknown.  Other related puzzles are presented by Dudeney

144: \cite{Dudeney-1967-pennies}, Fujimura \cite{Fujimura-1978-coins}, and Brooke

145: \cite{Brooke-1963}.

146: % Brooke covers Figure \ref{rhomboid} and some related puzzles,

147: % but not Figure \ref{upside-down}.

148: 

149: \begin{figure}

150: \centerline{\includegraphics[scale=\coinscale]{upside_down.eps}}

151: \caption{\label{upside-down}

152:   Turn the pyramid upside-down in three moves,

153:   such that each coin is moved to a position adjacent to two other coins.}

154: \end{figure}

155: 

156: \begin{figure}

157: \centerline{\includegraphics[scale=\coinscale]{original_helena.eps}}

158: \caption{\label{original helena}

159:   Re-arrange the pyramid into a line in seven moves,

160:   such that each coin is moved to a position adjacent to two other coins.}

161: \end{figure}

162: 

163: The puzzles above always move the centers of coins to vertices of the

164: equilateral-triangle grid.  Another type of puzzle is to move coins on the

165: square grid, which appears less often in the literature but has significantly

166: more structure and can be more difficult.  The only published example we are

167: aware of is given by Langman \cite{Langman-1953}, which is also described by

168: Brooke \cite{Brooke-1963}, Bolt \cite{Bolt-1991-coins}, and Wells

169: \cite{Wells-1992-coins}; see Figure \ref{HOH}.  The first puzzle (H $\to$ O) is

170: solvable on the square grid, and the second puzzle (O $\to$ H) can only be

171: solved by a combination of the two grids.

172: 

173: \begin{figure}

174: \centerline{\includegraphics[scale=\coinscale]{HOH.eps}}

175: \caption{\label{HOH}

176:   Re-arrange the H into the O in four moves while staying on the square grid

177:   (and always moving adjacent to two other coins),

178:   and return to the H in six moves using both the equilateral-triangle and

179:   square grids.}

180: \end{figure}

181: 

182: In this paper we study generalizations of these types of puzzles, in which

183: coins are moved on some grid to positions adjacent to at least two other coins.

184: Specifically, we address the basic algorithmic problem: is it possible to solve

185: a puzzle with given start and finish configurations, and if so, find a

186: sequence of moves.  Surprisingly, we show that this problem has a

187: polynomial-time solution in many cases.  Our goal in this pursuit is to gain a

188: better understanding of what is possible by these motions, and as a result to

189: design new and interesting puzzles.  For example, one puzzle we have designed

190: is shown in Figure \ref{diagonal}.  We recommend the reader try this difficult

191: puzzle before reading Section \ref{Re-orienting L's} which shows how to solve

192: it.  Figures \ref{spindle}--\ref{tri2oline} show a few of the other puzzles we

193: have designed.  The last two puzzles involve labeled coins.

194: 

195: \xxx{A variation on `diagonal' is where the extra coins are labeled differently

196: from the others.  How much harder does this make the puzzle?  Try it and

197: mention it.}

198: 

199: \begin{figure}

200: \centerline{\includegraphics[scale=\coinscale]{diagonal.eps}}

201: \caption{\label{diagonal}

202:   A difficult puzzle on the square grid.

203:   The optimal solution uses 18 moves,

204:   each of which places a coin adjacent to two others.}

205: \end{figure}

206: 

207: \begin{figure}

208: \centerline{\includegraphics[scale=\coinscale]{spindle.eps}}

209: \caption{\label{spindle}

210:   Another puzzle on the square grid.

211:   The optimal solution uses 24 moves,

212:   each of which places a coin adjacent to two others.}

213: \end{figure}

214: 

215: \begin{figure}

216: \centerline{\includegraphics[scale=\coinscale]{vee.eps}}

217: \caption{\label{vee}

218:   Another puzzle on the square grid with the same rules.}

219:   %Solvable in 21 moves, each of which places a coin adjacent to two others.}

220: \end{figure}

221: 

222: \begin{figure}

223: \centerline{\includegraphics[scale=\coinscale]{swap5.eps}}

224: \caption{\label{swap5}

225:   A puzzle on the square grid involving labeled coins.

226:   Solvable in eleven moves, each of which places a coin adjacent to two

227:   others; see Figure \protect\ref{swap5 solution}.}

228: \end{figure}

229: 

230: \begin{figure}

231: \centerline{\includegraphics[scale=0.8]{tri2oline.eps}}

232: \caption{\label{tri2oline}

233:   A puzzle on the equilateral-triangle grid involving labeled coins.

234:   Solvable in eight moves, each of which places a coin adjacent to two others.}

235: \end{figure}

236: 

237: This paper studies two grids in particular: the equilateral-triangle grid, and

238: the square grid.  It turns out that the triangular grid has a relatively simple

239: structure, and nearly all puzzles are solvable.  An exact, efficient

240: characterization of solvable puzzles is presented in Section \ref{Triangular

241: Grid}.  The square grid has a more complicated structure, requiring us to

242: introduce the notion of ``extra coins'' to give a partial characterization of

243: solvable puzzles.  This result is described in Section \ref{Square Grid} after

244: some general tools for analysis are developed in Section \ref{General Tools}.

245: 

246: Before we begin, the next section defines a general graph model of the puzzles

247: under consideration.

248: 

249: %We begin in Section \ref{General Theory} with a general graph model of

250: %token-moving puzzles.  Section \ref{Square Grid} presents our main result about

251: %the square grid.  In Section \ref{Triangular Grid} we describe our result about

252: %the equilateral-triangle grid.  We conclude in Section \ref{Conclusion} with

253: %several open problems.

254: 

255: %\pagebreak

256: \section{Model}

257: \label{Model}

258: 

259: %This section provides some general concepts that we will find useful for

260: %analyzing the games.  We begin in Section \ref{Problem Statement} with a formal

261: %problem statement and definitions of terminology.  Section \ref{Span} presents

262: %the notion of the span of a configuration, which essentially specifies the

263: %reachable positions given an unlimited supply of extra coins.

264: %

265: %\subsection{Problem Statement}

266: %\label{Problem Statement}

267: 

268: We begin by defining ``token-moving'' and ``coin-moving'' puzzles

269: and related concepts.

270: %We begin with some basic concepts generic to token-moving and coin-moving

271: %puzzles.

272: The \emph{tokens} form a finite multiset $T$.  We normally think of

273: tokens as unlabeled, modeled by all elements of $T$ being equal, but another

274: possibility is to color tokens into more than one equivalence class (as in

275: Figure \ref{tri2oline}).  A \emph{board} is any simple undirected graph

276: $G=(P,E)$, possibly infinite, whose vertices are called \emph{positions}.  A

277: \emph{configuration} is a placement of the tokens onto distinct positions on

278: the board, i.e., a one-to-one mapping $C : T \to P$.  We will often associate a

279: configuration $C$ with its image, that is, the set of positions \emph{occupied}

280: by tokens.

281: 

282: A \emph{move} from a configuration $C$ changes the position of a single token

283: $t$ to an unoccupied position $p$, resulting in a new configuration.  This move

284: is denoted $t \mapsto p$, and the resulting configuration is denoted $C

285: \applymove t \mapsto p$.  We stress that moves are not required to ``slide''

286: the token while avoiding other tokens (like the puzzle in Figure

287: \ref{rhomboid}); the token can be picked up and placed in any unoccupied

288: position.

289: 

290: The \emph{configuration space} (or \emph{game graph}) is the directed graph

291: whose vertices are configurations and whose edges correspond to feasible moves.

292: A typical \emph{token-moving puzzle} asks for a sequence of moves to reach one

293: configuration from another, i.e., for a path between two vertices in the

294: configuration space, subject to some constraints.  A \emph{coin-moving puzzle}

295: is a geometric instance of a token-moving puzzle, in which tokens are

296: represented by \emph{coins}---constant-radius disks in the plane, and

297: constant-radius hyperballs in general---and the board is some lattice in the

298: same dimension.  If a token-moving or coin-moving puzzle with source

299: configuration $A$ and destination configuration $B$ is solvable, we say that

300: $A$ can be \emph{re-arranged} into $B$, and that $B$ is \emph{reachable} from

301: $A$.  This is equivalent to the existence of a directed path from $A$ to $B$

302: in the configuration space.

303: 

304: This paper addresses the natural question of what puzzles are solvable, subject

305: to the following constraint on moves which makes the problem interesting.  A

306: move $t \mapsto p$ is \emph{$d$-adjacent} if the new position $p$ is adjacent

307: to at least $d$ tokens other than the moved token $t$.  (Throughout,

308: \emph{adjacency} refers to the board graph $G$.)  This constraint is

309: particularly meaningful for $d$-dimensional coin-moving puzzles, because then a

310: move is easy to ``perform exactly'' without any underlying lattice: the new

311: position $p$ is determined rigidly by the $d$ coin adjacencies

312: (sphere tangencies).

313: 

314: The \emph{$d$-adjacency configuration space} is the subgraph of the

315: configuration space in which moves are restricted to be $d$-adjacent.

316: Studying connectivity in this graph is equivalent to studying solvable puzzles;

317: for example, if the graph is strongly connected, then all puzzles are solvable.

318: 

319: Here we explore solvable puzzles on two boards, the equilateral-triangle grid

320: and the square grid.  Because these puzzles are two-dimensional, in the context

321: of this paper we call a move \emph{valid} if it is 2-adjacent, and a position a

322: \emph{valid destination} if it is unoccupied and adjacent to at least two

323: occupied positions.  Thus a valid move involves transferring a token from some

324: source position to a valid destination position.  When the context is clear, we

325: will refer to a valid move just by ``move.''  A move is \emph{reversible} if

326: the source position is also a valid destination.

327: 

328: 

329: \section{Triangular Grid}

330: \label{Triangular Grid}

331: 

332: This section studies the equilateral-triangle grid, where most puzzles are

333: solvable.  To state our result, we need a simple definition.  Associated with

334: any configuration is the subgraph of the board induced by the occupied

335: positions.  In particular, a \emph{connected component} of a configuration is a

336: connected component in this induced subgraph.

337: 

338: \begin{theorem} \label{triangular grid}

339: On the triangular grid with the 2-adjacency restriction and unlabeled coins,

340: configuration $A$ can be re-arranged into a different configuration $B$

341: precisely if $A$ has a valid move, the number of coins in $A$ and $B$ match,

342: and at least one of four conditions holds:

343: 

344:   \begin{enumerate}

345:   \item $B$ contains three coins that are mutually adjacent (a triangle).

346:   \item $B$ has a connected component with at least four coins.

347:   \item $B$ has a connected component with at least three coins

348:         and another connected component with at least two coins.

349:   \item There is a single move from $A$ to $B$.

350:   \end{enumerate}

351: 

352: \noindent

353: The same result holds for labeled coins, except when there are exactly three

354: coins in the puzzle, in which case the labelings and movements are controlled

355: by the vertex 3-coloring of the triangular grid.

356: 

357: Furthermore, there is a polynomial-time algorithm to find a re-arrangement from

358: $A$ to $B$ if one exists.  Specifically, let

359: $n$ denote the number of coins and $d$ denote the maximum

360: distance between two coins in $A$ or $B$.

361: Then a solution with $O(n d)$ moves can be found in $O(n d)$ time.

362: \end{theorem}

363: 

364: The rest of this section is devoted to the proof of this theorem.

365: We begin in the next subsection by proving necessity of the conditions: if a

366: puzzle is solvable, then one of the conditions holds.  Then in the following

367: subsection we prove sufficiency of the conditions.

368: 

369: \subsection{Necessity}

370: 

371: Of course, it is necessary for $A$ to have a valid move and for $A$

372: and $B$ to have the same number of coins.  Necessity of at least one of the

373: four conditions is also not difficult to show, because Conditions 1--3 are so

374: broad, encompassing most possibilities for configuration $B$.

375: 

376: Suppose that a solvable puzzle does not satisfy any of Conditions 1--3, as in

377: Figure \ref{condition4}.  We prove that it must satisfy Condition 4, by

378: considering play backwards from the goal configuration $B$.  Specifically, a

379: \emph{reverse move} takes a coin currently adjacent to at least two others, and

380: moves it to any other location.  Because the puzzle is solvable, some coin in

381: configuration $B$ must be reverse-movable, i.e., must have at least two coins

382: adjacent to it.  Thus, some connected component of $B$ has at least three

383: coins.  Because Condition 2 does not hold, this connected component has exactly

384: three coins.  Because Condition 1 does not hold, these three coins are not

385: connected in a triangle.  Because Condition 3 does not hold, every other

386: component has exactly one coin.

387: 

388: Hence, one component of $B$ is a path of exactly three coins, say $c_1, c_2,

389: c_3$, and every other component of $B$ has exactly one coin, as in the left of

390: Figure \ref{condition4}.  Certainly at this moment $c_2$ is the only

391: reverse-movable coin.  We claim that after a sequence of reverse moves, $c_2$

392: will continue to be the only reverse-movable coin.  If we removed $c_2$, then

393: every coin would be adjacent to no others.  Thus, if we reverse move $c_2$

394: somewhere, then every other coin would be adjacent to at most one other

395: ($c_2$).  Hence, it remains that only $c_2$ can be reverse moved.

396: 

397: \begin{figure}

398: \centerline{\input condition4.pstex_t}

399: \caption{\label{condition4}

400:   Reverse-moving a configuration $B$ that does not satisfy any of

401:   Conditions 1--3.}

402: \end{figure}

403: 

404: Therefore, if we can reach $A$ from $B$ via reverse moves, we can do so in a

405: single reverse move of $c_2$ directly to where it occurs in $A$.  Thus

406: Condition 4 holds, as desired.

407: 

408: \subsection{Sufficiency}

409: 

410: Next we prove the more difficult direction: provided one of Conditions 1--3

411: hold, there is a re-arrangement from $A$ to $B$.  (This fact is obvious when

412: Condition 4 holds.)  All three cases will follow a common outline: we first

413: form a triangle (Section \ref{Getting Started}), then maneuver this triangle

414: (Section \ref{Triangle Maneuvering}) to transport all other coins (Section

415: \ref{Transportation}), and finally we place the three triangle coins

416: appropriately depending on the case (Section \ref{Finale}).

417: 

418: \subsubsection{Getting Started}

419: \label{Getting Started}

420: 

421: It is quite simple to make some triangle of coins.  By assumption, there is a

422: valid move from configuration $A$.  The destination of this move can have two

423: basic forms, as shown in Figure \ref{move tri}.  Either the move forms a

424: triangle, as desired, or the move forms a path of three coins.  In the latter

425: case, if there is not a triangle already with a different triple of coins, a

426: triangle can be formed by one more move as shown in the right of the figure.

427: 

428: \begin{figure}

429: \centerline{\input move_tri.pstex_t}

430: \caption{\label{move tri}

431:   Two types of valid destinations for a coin $c$.

432:   In the latter case, we show a move to form a triangle.}

433: \end{figure}

434: 

435: This triangle $T_0$ suffices for unlabeled coin puzzles.  However, for labeled

436: coin puzzles, we cannot use just any three coins in the triangle; we need a

437: particular three, depending on $B$.  For example, if $B$ satisfies Condition 1,

438: then the coins forming the triangle in $B$ are the coins we would like in the

439: triangle for maneuvering.  To achieve this, we ``bootstrap'' the

440: triangle $T_0$ formed above, using this triangle with the incorrect coins to

441: form another triangle with the correct coins.  Specifically, if we desire a

442: triangle using coins $t_1$, $t_2$, and $t_2$, then we move each coin

443: in the difference $\{t_1, t_2, t_3\} \setminus T_0$ to be adjacent to

444: appropriate coins in $T_0$.  There are three cases, shown in Figure

445: \ref{bootstrap tri}, depending on how many coins are in the difference.  If

446: ever we attempt to move a coin to an already occupied destination, we first

447: move the coin located at that destination to any other valid destination.

448: 

449: \begin{figure}

450: \centerline{\input bootstrap_tri.pstex_t}

451: \caption{\label{bootstrap tri}

452:   The three cases of building a triangle $\{t_1, t_2, t_3\}$ out of an

453:   existing triangle, depending on how many coins the two triangles share.

454:   {From} left to right, zero, one, and two coins of overlap.}

455: \end{figure}

456: 

457: \subsubsection{Triangle Maneuvering}

458: \label{Triangle Maneuvering}

459: 

460: Consider a triangle of coins $t_1$, $t_2$, and $t_3$.  The possible positions

461: of this triangle on the triangular grid are in one-to-one correspondence with

462: their centers, which are vertices of the dual hexagonal grid.  Moving one coin

463: (say $t_1$) to be adjacent to and on the other side of the others ($t_2$ and

464: $t_3$) corresponds to moving the center of the triangle to one of the three

465: neighboring centers on the hexagonal grid.  Thus, without any other coins on

466: the board, the triangle can be moved to any position by following a path in the

467: hexagonal grid.

468: 

469: This approach can be modified to apply when there are additional obstacle

470: coins; see Figure \ref{maneuvering triangle} for an example.  Conceptually we

471: always move one of the triangle coins, say $t_i$, in order to move the center

472: of the triangle to an adjacent vertex of the hexagonal grid.  But if the move

473: of $t_i$ is impossible because the destination is already occupied by another

474: coin $g_i$, then in fact we do not make any move.  There will be a triangle in

475: the desired position now, but it will not consist of the usual three coins

476: ($t_1$, $t_2$, and $t_3$); instead, $t_i$ will be replaced by the ``ghost

477: coin'' $g_i$.  Such a triangle suffices for our purposes of transportation

478: described in Section \ref{Transportation}.  One final detail is how the ghost

479: coins behave: if we later need to move a ghost coin $g_i$, we instead move the

480: original (unmoved) coin $t_i$.  Thus ghost coins are never moved; only $t_1$,

481: $t_2$, and $t_3$ are moved during triangle maneuvering (even if coins are

482: labeled).

483: 

484: \begin{figure}

485: \centerline{\input maneuvering_triangle.pstex_t}

486: \caption{\label{maneuvering triangle}

487:   An example of triangle maneuvering.  Dotted arrows denote conceptual moves,

488:   and solid arrows denote actual moves.}

489: \end{figure}

490: 

491: \subsubsection{Transportation}

492: \label{Transportation}

493: 

494: Triangle maneuvering makes it easy to \emph{transport} any other coin to any

495: desired location.  Specifically, suppose we want to move coin $c \notin \{t_1,

496: t_2, t_3\}$ to destination position $d$.  If $d$ is already occupied by another

497: coin $c'$, we first move $c'$ to an arbitrary valid destination; there is at

498: least one because the triangle can be maneuvered.  Now we maneuver the triangle

499: so that the (potentially ghost) triangle has two coins adjacent to $d$, so that

500: the third coin is not on $d$, and so that the triangle does not overlap $c$.

501: This is easily arranged by examining the location of $c$ and setting the

502: destination of the triangle appropriately.  For example, if $c$ is within

503: distance two of $d$, then there are four positions for the triangle that are

504: adjacent to $d$ and do not overlap $c$; otherwise,

505: the triangle can be placed in any of the six positions adjacent to $d$.

506: Finally, because $d$ is now a valid destination---it is adjacent to two coins

507: in the triangle---we can move $c$ to $d$.

508: 

509: \begin{figure}

510: \centerline{\input transportation_tri.pstex_t}

511: \caption{\label{transportation tri}

512:   Transporting coin $c$ to destination $d$ using triangle $\{t_1, t_2, t_3\}$.

513:   In both cases, we choose the location of the triangle

514:   so that it does not overlap $c$.}

515: \end{figure}

516: 

517: By the properties of triangle maneuvering, this transportation process even

518: preserves coin labels: the only actual coins moved are $t_1$, $t_2$, $t_3$,

519: $c$, and possibly a coin at $d$.  But any coin at position $d$ must not have

520: already been in its desired position, because $d$ is $c$'s desired position.

521: Thus, applying the transportation process to every coin except $t_1$, $t_2$,

522: and $t_3$ places all coins except these three in their desired locations.

523: 

524: \subsubsection{Finale}

525: \label{Finale}

526: 

527: Once transportation is complete, it only remains to place the triangle coins

528: $t_1$, $t_2$, and $t_3$ in their desired locations.  By the bootstrapping in

529: Section \ref{Getting Started}, we are able to choose the unplaced coins $\{t_1,

530: t_2, t_3\}$ however we like.  This property will be exploited differently

531: in the three cases.

532: % For the moment we ignore the labeling of $t_1$, $t_2$, and $t_3$.

533: 

534: \paragraph{Property 1.}

535: If there is a triangle in $B$, then we choose these three coins as the unplaced

536: coins $t_1$, $t_2$, and $t_3$, and use them to transport all other coins.  Then

537: we maneuver the triangle $\{t_1, t_2, t_3\}$ exactly where it appears in $B$.

538: Because all other coins have been moved to their proper location, in this

539: position the triangle will not have any ghost coins.

540: 

541: However, it may be that the coins $\{t_1, t_2, t_3\}$ are labeled incorrectly

542: among themselves, compared to $B$.  Assuming there are more than three coins in

543: the puzzle, this problem can be repaired as follows.  We maneuver the triangle

544: so that it does not overlap any other coins but is adjacent to at least one

545: coin $c$; for example, there is such a position for the triangle just outside

546: the smallest enclosing hexagon of the other coins.  Refer to Figure

547: \ref{relabel tri}.  Now two coins of the triangle, say $t_1$ and $t_2$, are

548: adjacent to three other coins each: each other, $t_3$, and $c$.  Thus we can

549: move $t_1$ to any other valid destination, and then move $t_2$ or $t_3$ to

550: replace it.  Afterwards we can move $t_1$ to take the place of $t_2$ or $t_3$,

551: whichever moved.  This procedure swaps $t_1$ and either $t_2$ or $t_3$.  By

552: suitable application, we can achieve any permutation of $\{t_1, t_2, t_3\}$,

553: and thereby achieve the desired labeling of the triangle.

554: 

555: \begin{figure}

556: \centerline{\input relabel_tri.pstex_t}

557: \caption{\label{relabel tri}

558:   Swapping coins $t_1$ and $t_2$ in a triangle, using an adjacent coin $c$.}

559: \end{figure}

560: 

561: \paragraph{Property 2 but not Property 1.}

562: Refer to Figure \ref{finale tri 2}.  If there is not a triangle in $B$, but

563: there is a connected component of $B$ with at least four coins, then there is a

564: path in $B$ of length four, $(p_1, p_2, p_3, p_4)$.  From {B} we reverse move

565: $p_2$ so that it is adjacent to $p_3$ and $p_4$.  If this position is already

566: occupied by a coin $c$, we first reverse move $c$ to any other unoccupied

567: position.  Now $p_2$, $p_3$, and $p_4$ are mutually adjacent, so we have a

568: new destination configuration $B'$

569: with Property 1.  As described above, we can re-arrange $A$ into $B'$.

570: Then we undo our reverse moves: move $p_2$ back adjacent to $p_1$ and $p_3$,

571: and move $c$ back adjacent to $p_3$ and $p_4$.  This procedure re-arranges

572: $A$ into $B$.

573: 

574: \begin{figure}

575: \centerline{\input finale_tri_2.pstex_t}

576: \caption{\label{finale tri 2}

577:   Reverse moving a configuration $B$ with Property 2 into a

578:   configuration with Property 1.}

579: \end{figure}

580: 

581: \paragraph{Property 3 but not Property 1.}

582: This case is similar to the previous one; refer to Figure \ref{finale tri 3}.

583: There must be a path in $B$ of length three, $(p_1, p_2, p_3)$, as well as a

584: pair of adjacent coins, $(q_1, q_2)$, in different connected components of $B$.

585: If both positions adjacent to both $q_1$ and $q_2$ are already occupied, we

586: first reverse move one such coin (call it $c$) to an arbitrary unoccupied

587: position.  This frees up a position adjacent to $q_1$ and $q_2$, to which we

588: reverse move $p_2$.  Now $\{q_1, q_2, p_2\}$ form a triangle, so Property 1

589: holds, and we can reach this new configuration $B'$ from $A$.  Then we undo our

590: reverse moves: move $p_2$ back adjacent to $p_1$ and $p_3$, and move $c$ back

591: adjacent to $q_1$ and $q_2$.  This procedure re-arranges $A$ into $B$.

592: 

593: \begin{figure}

594: \centerline{\input finale_tri_3.pstex_t}

595: \caption{\label{finale tri 3}

596:   Reverse moving a configuration $B$ with Property 3 into a

597:   configuration with Property 1.}

598: \end{figure}

599: 

600: This concludes the proof of Theorem \ref{triangular grid}.

601: 

602: %--- This is covered in the conclusion instead.

603: % \subsection{Conjectures}

604: 

605: % An interesting question is 

606: 

607: % \begin{conjecture}

608: % On the triangular grid with the 2-adjacency restriction and unlabeled coins, if

609: % there is a re-arrangement from $A$ to $B$ when the coins are unlabeled, then

610: % there is also a re-arrangement consisting of a sequence of \emph{slides}.

611: % \end{conjecture}

612: 

613: 

614: %--- Old stuff about span:

615: % Coin-moving puzzles have an entirely different structure with the triangular

616: % grid.  In particular, we have the following characterization of span:

617: 

618: % \begin{lemma}

619: % For a configuration $C$ on the triangular grid, if $C$ has a valid destination

620: % position, then $\span C$ is the entire triangular grid; otherwise, $\span C =

621: % C$.

622: % \end{lemma}

623: 

624: 

625: \section{General Tools}

626: \label{General Tools}

627: 

628: In this section we develop some general lemmas about token-moving puzzles.

629: Although we only use these tools for the square grid, in Section \ref{Square

630: Grid}, they apply to arbitrary boards and may be of more general use.

631: 

632: \subsection{Picking Up and Dropping Tokens}

633: \label{Picking Up and Dropping Tokens}

634: 

635: First we observe that additional tokens cannot ``get in the way'':

636: %First we observe that, for the purpose of showing connectivity of the

637: %configuration space, tokens can be removed freely because this modification

638: %only makes the puzzle more difficult.

639: 

640: \begin{lemma} \label{emulation}

641: If a token-moving puzzle is solvable, then it remains solvable if we add an

642: additional token with an unspecified destination, provided tokens are

643: unlabeled.  This result also holds if all moves must be reversible.

644: \end{lemma}

645: 

646: \begin{proof}

647: A move can be blocked by an extra token $e$ at position $p$ because $p$ is

648: occupied and hence an invalid destination.  But if ever we encounter such a

649: move of a token $t$ to position $p$, we can just ignore the move, and swap the

650: roles of $e$ and $t$: treat $e$ as the moved version of $t$, and treat $t$

651: as an extra token replacing $e$.  Thus, any sequence of moves in the original

652: puzzle can be emulated by an equivalent sequence of moves in the augmented

653: puzzle.  We are not introducing any new moves, only removing existing moves,

654: so all moves remain reversible if they were originally.

655: \end{proof}

656: 

657: This proof leads to a technique for emulating a more powerful model for solving

658: puzzles.  In addition to moving coins as in the normal model, we can

659: conceptually \emph{pick up} (remove) a token, and later \emph{drop} (add) it

660: onto any valid destination.  At any moment we can have any number of tokens

661: picked up.  While a token $t$ is conceptually picked up, we emulate any moves

662: to its actual position $p$ as in the proof of Lemma \ref{emulation}: if we

663: attempt to move another token $t'$ onto position $p$, we instead reverse the

664: roles of $t$ and $t'$.  To drop a token onto a desired position $p$, we simply

665: move the actual token to position $p$ if it is not there already.

666: 

667: Of course, this process may permute the tokens.  Nonetheless we will find this

668: approach useful for puzzles with labeled tokens.

669: 

670: One might instead consider the emulation method used implicitly in

671: Section \ref{Triangle Maneuvering} for triangular maneuvering: move original

672: coins instead of ghost coins.  This approach has the advantage that it

673: preserves the labels of the coins.  Unfortunately, the approach makes it

674: difficult to preserve reversibility as in Lemma \ref{emulation}, and so is

675: insufficient for our purposes here.

676: 

677: \subsection{Span}

678: \label{Span}

679: 

680: The \emph{span} of a configuration $C$ is defined recursively as follows.  Let

681: $d_1, \dots, d_m$ be the set of valid destinations for moves in $C$.

682: %--- Valid destination is now defined, so we don't need this:

683: %For example, under the $k$-adjacency restriction, these

684: %are all unoccupied positions that have at least $k$ occupied neighbors.

685: %--- This is a ``tighter'' definition of span, but is somehow less intuitive:

686: %Let $t_1 \mapsto d_1, \dots, t_m \mapsto d_m$ be the valid moves from

687: %$C$, subject to any restrictions imposed on the puzzle.

688: If $m=0$, the span of $C$ is just $C$ itself.  Otherwise, it is the span of

689: another configuration $C'$, defined to be $C$ with additional tokens at

690: positions $d_1, \dots, d_m$.  If this process never terminates, the span is

691: defined to be the limit, which exists because it is a countable union of finite

692: sets.

693: 

694: \begin{figure}

695: \centerline{\includegraphics[scale=\coinscale]{span_3x3diag.eps}}

696: \caption{\label{3x3 diagonal span}

697:   In this example, the span is the smallest rectangle enclosing the

698:   configuration.}

699: \end{figure}

700: 

701: The span of a configuration lists all the positions we could hope to reach, or

702: more precisely, the positions we could reach if we had an unlimited number of

703: extra tokens that we could drop.  In particular, we have the following:

704: 

705: \begin{lemma} \label{span monotone}

706: If configuration $A$ can be re-arranged into configuration $B$, then $\span A

707: \supseteq B$ and thus $\span A \supseteq \span B$.

708: \end{lemma}

709: 

710: In other words, valid moves can never cause the span of the current

711: configuration to increase.  Thus the most connected we could hope the

712: configuration space to be is the converse of Lemma \ref{span monotone}: for

713: every pair of configurations with $\span A \supseteq \span B$, $A$ can be

714: re-arranged into $B$.  In words, we want that every configuration $A$ can be

715: re-arranged into any configuration $B$ with the \emph{same or smaller span}.

716: 

717: We call a configuration \emph{span-minimal} if the removal of any of its tokens

718: reduces the span.  Span-minimal configurations are essentially the ``skeleta''

719: that keep configurations with the same span reachable.

720: %

721: %--- The following is a funny way to put the result... this result is about

722: %    the number of coins past span minimality, which is different from ``extra

723: %    coins.''  For example, in triangular grid this is meaningless, though the

724: %    lemma still holds.

725: % 

726: % \subsection{Zero Extra Tokens Can Be Insufficient}

727: % 

728: % At this point it is easy to justify having at least one extra token

729: % in the worst case for any board:

730: %

731: One general property of span-minimal configurations is the following:

732: 

733: \begin{lemma} \label{span-minimal fragile}

734: If a configuration is span-minimal, any move will reduce the span.

735: \end{lemma}

736: 

737: \begin{proof}

738: Suppose to the contrary that there is a move $t \mapsto p$ that does not reduce

739: the span of a span-minimal configuration $C$.  In particular, $p$ must be a

740: valid destination position in the subconfiguration $C-t$, because $t$ does not

741: count in the $d$-adjacency restriction.  Hence, adding a new token at position

742: $p$ to the configuration $C-t$ has no effect on the span of $C$.  But this

743: two-step process of removing token $t$ and adding a token at position $p$ is

744: equivalent to moving $t$ to $p$, so $\span (C \applymove t \mapsto p) = \span

745: (C-t)$.  But we assumed that $\span (C \applymove t \mapsto p) = \span C$, and

746: hence $\span C = \span (C-t)$, contradicting that $C$ is span-minimal.

747: \end{proof}

748: 

749: Under the 2-adjacency restriction, a \emph{chain} is a sequence of tokens with

750: the property that the distance (in the board graph $G$) between two successive

751: tokens is at most $2$.  We will use chains as basic ``units'' for creating a

752: desired span.

753: 

754: %----- Probably not useful for general $k$:

755: % Under the $k$-adjacency restriction, a \emph{chain} is a sequence of occupied

756: % positions with the property that the distance between two consecutive positions

757: % (in the board graph $G$) is at most $k$.  We will use chains as basic ``units''

758: % for creating a desired span.

759: 

760: %--- Old, and pretty useless:

761: %For each type of board, the basic approach of our analysis is as follows.

762: %First, we give some basic properties of the span.  Then we determine the

763: %meaning of the span in terms of transferability.

764: 

765: Notice that the notion of span is useless for the already analyzed triangular

766: grid: provided there is a valid move, the span of any configuration is the

767: entire grid.  Thus, for the triangular grid, a configuration is span-minimal

768: precisely if it has no valid moves.  For the square grid, however, the notion

769: of span and span minimality is crucial.

770: 

771: \subsection{Extra Tokens}

772: \label{Extra Tokens}

773: 

774: As described in the previous section, we can only re-arrange configurations

775: into configurations with the same or smaller span.  Unfortunately, the converse

776: is not true.  Indeed, the key problem situations are span-minimal

777: configurations; by Lemma \ref{span-minimal fragile}, such configurations

778: immediately lose span when we try to move them.  Hence, any two distinct

779: span-minimal configurations with the same span cannot reach each other.  An

780: example on the square grid is that the two opposite diagonals of a square are

781: unreachable from each other, as shown in Figure \ref{square grid 0 extra}.

782: 

783: \begin{figure}

784: \centerline{\includegraphics[scale=0.55]{sqrgrid_0extra.eps}}

785: \caption{\label{square grid 0 extra}

786:   Subgraph of configuration space reachable from full-span configurations

787:   (outlined in bold) with no extra coins.}

788: \end{figure}

789: 

790: Thus we explore the notion of \emph{extra tokens}, a set of tokens whose

791: removal does not reduce the span of the configuration.  Lemma \ref{span

792: monotone} and Figure \ref{square grid 0 extra} shows that we need at least one

793: extra token.  In fact, the two opposite diagonals on the square grid shown in

794: Figure \ref{square grid 0 extra} are difficult to reach from each other; as

795: shown in Figure \ref{square grid 1 extra}, even one extra token is

796: insufficient.  What is surprising is that a small number of extra tokens

797: seem to be generally sufficient to make the configuration space strongly

798: connected.  We prove this for the square grid in the next section.

799: 

800: \begin{figure}

801: \centerline{\includegraphics[scale=0.55]{sqrgrid_1extra.eps}}

802: \caption{\label{square grid 1 extra}

803:   Subgraph of configuration space reachable from full-span configurations

804:   (outlined in bold) with one extra token.}

805: \end{figure}

806: 

807: %--- Old view of extra coins; bleh.

808: % This configuration space can easily be disconnected.  Thus for the purpose of

809: % analysis we introduce the notion of \emph{extra tokens} or \emph{don't-care

810: % tokens}.  A puzzle specifies the initial positions of such tokens, but not

811: % their desired final positions.  In other words, configurations are considered

812: % equivalent if they place all \emph{important} tokens (nonextra tokens) in the

813: % same way.  Certainly if all tokens are extra, then the configuration space

814: % consists of a single equivalence class.  The basic question is this: how few

815: % extra tokens can we get away with while keeping the configuration space

816: % connected?  Our motivation is puzzle design: a connected configuration space

817: % means that all puzzles are solvable, even if they do not ``appear'' so.

818: 

819: % If there are too few extra tokens, any move of an important token can

820: % immediately make previously reachable configurations unreachable.  See Figures

821: % \ref{square grid 0 extra} and \ref{square grid 1 extra} for examples.

822: % Surprisingly, we show in several cases that just a constant number of extra

823: % tokens suffice to make the configuration space (essentially) connected.  The

824: % basic idea is that a configuration can be split into two components: the basic

825: % skeleton that keeps other configurations reachable, and some number of extra

826: % tokens that allow motions to be made while maintaining the skeleton.

827: % %If there are too few of these extra tokens, any move can immediately make

828: % %previously reachable configurations unreachable.

829: 

830: % %--- This isn't true!

831: % %In fact, extra tokens are not as unrestricted as we have described.  We will be

832: % %able to control the final position of an extra token subject to a simple and

833: % %natural restriction.

834: 

835: % While we have modeled extra tokens as having unspecified destinations, this is

836: % not necessary, and indeed undesired for real puzzles.  A puzzle can specify the

837: % final position of an extra token to be any valid destination position, i.e.,

838: % any position that is adjacent to at least two important tokens or previously

839: % placed extra tokens.  This fact is used, for example, in our two square-grid

840: % puzzles in Figures \ref{diagonal}--\ref{vee}.

841: %--- Should have stated it this way upfront.

842: 

843: %--- Very old figure:

844: % \begin{figure}

845: % {\tt

846: % \centerline{%

847: % \vbox{\hsize=0.25in

848: % X.\\

849: % X.

850: % }

851: % ~~~

852: % \vbox{\hsize=0.25in

853: % XX\\

854: % ..

855: % }

856: % ~~~

857: % \vbox{\hsize=0.25in

858: % .X\\

859: % .X

860: % }

861: % ~~~

862: % \vbox{\hsize=0.25in

863: % ..\\

864: % XX

865: % }}

866: 

867: % \bigskip

868: % \bigskip

869: 

870: % \centerline{%

871: % \vbox{\hsize=0.25in

872: % X.\\

873: % XX

874: % }

875: % $\leftrightarrow$

876: % \vbox{\hsize=0.25in

877: % XX\\

878: % .X

879: % }

880: % ~~~

881: % \vbox{\hsize=0.25in

882: % .X\\

883: % XX

884: % }

885: % $\leftrightarrow$

886: % \vbox{\hsize=0.25in

887: % XX\\

888: % X.

889: % }}}

890: % \caption{\label{trivial disconnected}

891: %   The 2-adjacency configuration space is disconnected in these

892: %   square-grid examples with two and three unlabeled coins.}

893: % \end{figure}

894: 

895: 

896: \section{Square Grid}

897: \label{Square Grid}

898: 

899: % Our main result is about the 2-adjacency configuration space for the square

900: % grid.  First note that one extra coin does not suffice to maintain the span,

901: % as shown in Figure \ref{square grid 1 extra}.

902: 

903: This section analyzes coin-moving puzzles on the square grid, using the tools

904: from the previous section.  In particular, we show that with just two extra

905: coins, we can reach essentially every configuration on the square grid with the

906: same or smaller span.  The only restriction is that the extra coins can only

907: be destined for positions that are adjacent to at least two other coins.

908: 

909: \begin{theorem} \label{square grid}

910: On the square grid with the 2-adjacency restriction and unlabeled coins,

911: configuration $A$ can be re-arranged into configuration $B$ if there are coins

912: $e_1$ and $e_2$ such that $\span (A \setminus \{e_1, e_2\}) \supseteq \span (B

913: \setminus \{e_1, e_2\})$ and each $e_i$ is adjacent to two other coins in $B$

914: (excluding $e_1$ or $e_2$).

915: Furthermore, there is an algorithm to find such a re-arranging sequence using

916: $O(n^3)$ moves and $O(n^3)$ time, where $n$ is the number of coins.

917: \end{theorem}

918: 

919: %--- Old theorem statement:

920: % \begin{theorem} \label{square grid}

921: % For the square grid, any configuration can be re-arranged into any

922: % configuration with the same or smaller span using two extra coins.

923: % There is an algorithm to find such a re-arranging sequence using $O(n^4)$ moves

924: % and $O(n^4)$ time, where $n$ is the number of coins.

925: % %Let the board $G$ be the square grid, and fix the coin set $T$.  If two

926: % %configurations $C$ and $C'$ satisfy $\span C \supseteq \span C'$, then $C$ can

927: % %be re-arranged into $C'$ using two extra coins, using $O(|T|^3)$ moves.

928: % %----

929: % %On the square grid, if $A$ and $B$ are configurations with the same span, then

930: % %$A$ is reachable from $B$ using two extra coins.

931: % \end{theorem}

932: 

933: We prove this theorem by showing that every configuration (in particular, $A$

934: and $B$) can be brought to a canonical configuration with the same span via a

935: sequence of (mostly) reversible moves.  As a consequence, we can move from

936: any configuration $A$ to any other $B$ by routing through this canonical

937: configuration.

938: 

939: Our proof uses the model of picking up and dropping coins, which can be

940: emulated as described in Section \ref{Picking Up and Dropping Tokens}.

941: However, we must be careful how we pick up and drop coins, so that the

942: resulting moves are reversible.  For example, initially we pick up the extra

943: coins $e_1$ and $e_2$, and then drop them temporarily wherever needed.  For

944: re-arranging the source configuration $A$ into the canonical configuration,

945: this step may not result in reversible moves, but fortunately this is not

946: necessary in this case.  For re-arranging the destination configuration $B$

947: into the canonical configuration, however, reversibility is crucial, and is

948: guaranteed by the condition in the theorem of each $e_i$ being adjacent to at

949: least two other coins.

950: 

951: %--- This is covered below.

952: % For convenience, whenever we find a coin other than $e_1$ or $e_2$ whose

953: % removal does not reduce the span, we pick it up.  This rule simply removes

954: % clutter from the board, making it easier to argue about making moves; we can

955: % effectively assume that most destinations are unoccupied.  Of course, we must

956: % be careful when we later drop these picked-up coins, so that the resulting

957: % moves are reversible.  This problem will be dealt with in Section \ref{Final

958: % Sweep}, after all other coins are in their canonical configuration.

959: 

960: \subsection{Basics}

961: \label{Basics}

962: 

963: We begin with some preliminary lemmas.  A \emph{rectangle} is the full

964: collection of coins between two $x$ coordinates and two $y$ coordinates.  The

965: \emph{half-perimeter} of a rectangle is the number of distinct $x$ coordinates

966: plus the number of distinct $y$ coordinates over all coins in the rectangle.

967: The \emph{distance} between two sets of coins is the minimum distance between

968: two coins from different sets.

969: 

970: \begin{lemma}

971: For the square grid, the span of any configuration is a disjoint union of

972: (finite) rectangles with pairwise distances at least $3$.

973: \end{lemma}

974: 

975: % \begin{proof}

976: % Because a single coin is a rectangle, $\span A$ is certainly a disjoint

977: % union of rectangles.  We also want to show that each connected component

978: % is a rectangle, and different connected components are distance $3$ or more

979: % apart.

980: 

981: % Suppose $C$ is a connected component of $\span A$.  

982: % Define 

983: % $(a_0,b_0),(a_1,b_1)$ by

984: % $$a_0=min\left\{a_i|

985: % [a_i,b_i]\in \span A\right\},\>

986: % b_0=min\left\{b_i|

987: % [a_i,b_i]\in \span A\right\}$$

988: % $$a_1=max\left\{a_i|

989: % [a_i,b_i]\in \span A\right\},\>

990: % b_1=max\left\{b_i|

991: % [a_i,b_i]\in \span A\right\}$$

992: % We may assume that $a_0\not=a_1$ and $b_0\not=b_1$, since in these cases

993: % $\span A$ is a rectangle, and the result is easy.

994: % We claim that $C=R_{(a_0,b_0),(a_1,b_1)}$.  Suppose not, then there

995: % is a position $(a,b)\in R_{(a_0,b_0),(a_1,b_1)}\setminus C$ with 

996: % one of the positions adjacent to $(a,b)$ in $C$.  We can follow round

997: % the boundary of $C$ from this position until we reach a position

998: % in $R_{(a_0,b_0),(a_1,b_1)}\setminus C$  with two neighbors in $C$.

999: % But then this position would also be in $C$.  Hence we have a

1000: % contradiction, and thus the connected components are rectangles.

1001: % Clearly any two disconnected rectangles must be at least distance

1002: % three apart, or else a coin could be added, contradicting that we have

1003: % added all possible coins in constructing $\span A$.

1004: % \end{proof}

1005: 

1006: \begin{lemma} \label{min chain size}

1007: For each rectangle (connected component) of the span,

1008: say with half-perimeter $h$, there must be at least $\lceil h/2 \rceil$ coins

1009: within that rectangle in the configuration.

1010: \end{lemma}

1011: 

1012: The following beautiful proof of this lemma has been distributed among several

1013: people, but its precise origin is unknown.  We first heard it from Martin

1014: Farach-Colton, who heard it from Peter Winkler, who heard it from Pete Gabor

1015: Zoltan, who learned of it through the Russian magazine \emph{Kvant}

1016: (around 1985--1987).

1017: 

1018: \medskip

1019: 

1020: \begin{proof}

1021: Consider how the (full) perimeter changes as we compute the span of the coins

1022: within the rectangle.  Initially we have $n$ coins, say, so the perimeter is at

1023: most $4n$.  Each coin that we add while computing the span satisfies the

1024: 2-adjacency restriction, so the perimeter never increases.  In the end we must

1025: have a rectangle with perimeter $2h$.  Hence $4n \geq 2h$, i.e., $n \geq h/2$,

1026: and because $n$ is integral, $n \geq \lceil h/2 \rceil$.

1027: \end{proof}

1028: 

1029: \subsection{Canonical Configuration}

1030: 

1031: Observe that a chain has span equal to its smallest enclosing rectangle.  We

1032: define an \emph{L} to be a particular kind of chain, starting and ending at

1033: opposite corners of the rectangular span, and arranged along two edges of this

1034: rectangle, with the property that it has the minimum number of coins.  See

1035: Figure \ref{L} for examples.  More precisely, if the half-perimeter of the

1036: rectangle (along which the L is arranged) is $2k$, then there must be precisely

1037: $k$ coins, every consecutive pair at distance exactly two from each other.  And

1038: if the half-perimeter is $2k+1$, then there must be precisely $k+1$ coins,

1039: every consecutive pair at distance exactly two from each other, except the last

1040: pair which are distance one from each other.  In general, for half-perimeter

1041: $h$, an L has $\lceil h/2 \rceil$ coins.

1042: 

1043: \begin{figure}

1044: \centerline{\includegraphics[scale=\coinscale]{L.eps}}

1045: \caption{\label{L}

1046:   Examples of L's.}

1047: \end{figure}

1048: 

1049: While L's can have any orientation, the \emph{canonical L} is oriented

1050: like the letter L, starting at the top-left corner, continuing past the

1051: lower-left corner, and ending at the bottom-right corner.

1052: 

1053: Given a configuration, or more precisely, given its span and the number of

1054: coins in each connected component of the span, we define the \emph{canonical

1055: configuration} as follows.  Refer to Figure \ref{canonical} for examples.

1056: Within each connected component (rectangle) of the span, say with

1057: half-perimeter $h$, we arrange the first $\lceil h/2 \rceil$ coins into the

1058: canonical L.  (Lemma \ref{min chain size} implies that there are at least this

1059: many coins to place.)  Any additional coins are placed one at a time, in the

1060: leftmost bottommost unoccupied position.

1061: 

1062: \begin{figure}

1063: \centerline{\includegraphics[scale=\coinscale]{canonical.eps}}

1064: \caption{\label{canonical}

1065:   Examples of the canonical configuration of $k$ coins within a

1066:   rectangular span.  (Left) One coin in addition to the canonical L.

1067:   (Middle) Four coins in addition to the canonical L.

1068:   (Right) All 24 additional coins.}

1069: \end{figure}

1070: 

1071: This definition of the canonical configuration is fairly arbitrary, but it has

1072: the useful property that each successive position for an additional coin is a

1073: valid destination, given the previously placed additional coins.  This allows

1074: us to focus on forming the canonical L, and then picking up all additional

1075: coins and dropping them in the order shown on the right of Figure

1076: \ref{canonical}.

1077: 

1078: \subsection{Canonicalizing Algorithm}

1079: 

1080: The main part of proving Theorem \ref{square grid} is to show an algorithm for

1081: converting any configuration into the corresponding canonical configuration,

1082: using a sequence of (mostly) reversible moves.  We will apply induction (or,

1083: equivalently, recursion) on the number of coins.  That is, we assume that any

1084: configuration with fewer coins can be re-arranged into its canonical

1085: configuration.

1086: 

1087: For now, we assume that there are no extra coins in addition to $e_1$ and

1088: $e_2$.  For if there were such a coin, we could immediately pick it up.  Then

1089: we have a simpler configuration: it has one fewer coin.  Thus we can apply the

1090: induction hypothesis, and re-arrange the remaining coins into their canonical

1091: configuration.  Finally we must drop the previously picked-up coin in the

1092: appropriate location.  This aspect is somewhat trickier than it may seem: if we

1093: are not careful, we may make an irreversible move.  We delay this issue to

1094: Section \ref{Final Sweep}.

1095: 

1096: The overall outline of the algorithm is as follows:

1097: 

1098:   \begin{enumerate}

1099:   \item Initialize the set of L's to be one for each coin.

1100:   \item Until the configuration is canonical:

1101:     \begin{enumerate}

1102:     \item Pick two L's whose bounding rectangles are distance at most two

1103:           from each other.

1104:     \item Re-orient the L's so that the L's themselves are distance at most

1105:           three from each other.

1106:     \item Merge the two L's.

1107:     \end{enumerate}

1108:   \end{enumerate}

1109: 

1110: Normally, each iteration of Step~2 decreases the number of L's by one, so the

1111: algorithm would terminate in at most $n$ iterations.  However, at any time we

1112: may find an extra coin in addition to $e_1$ and $e_2$, and pick it up.

1113: Fortunately, this operation can only split one L into at most two L's.

1114: Thus we can charge the cost of creating an extra L to the event of picking up

1115: an extra coin, which can happen at most $n$ times.  Hence, the total number of

1116: iterations of Step~2 is $O(n)$.

1117: 

1118: In the following two sections, we describe how Steps~2(b) and~2(c) can be done

1119: in $O(n^2)$ moves each.  These bounds result in a total of $O(n^3)$ moves.  The

1120: running time of the algorithms will be proportional to the number of moves.

1121: 

1122: \subsubsection{Re-orienting L's}

1123: \label{Re-orienting L's}

1124: 

1125: There are eight possible orientations for an L, depending at which corner it

1126: starts, and whether it hugs the top edge or bottom edge of the rectangular

1127: span.  We will only be concerned with four different types of orientations,

1128: depending on whether it looks like the letter L rotated $0$, $90^\circ$,

1129: $180^\circ$, or $270^\circ$.  In other words, we are not concerned with the

1130: parity issue of which corner might have two adjacent coins.

1131: 

1132: It is relatively easy to \emph{flip} an L about a diagonal, using two extra

1133: coins.  Figure \ref{flip 3L} shows how to do this in a constant number of moves

1134: for an L consisting of three coins.  Figure \ref{flip L} shows how to use these

1135: subroutines to flip an L of arbitrary size.  Basically, we use the flips of

1136: three-coin L's to ``bubble'' the kink in the L up to the top, repeatedly until

1137: it is all the way right.  The total number of moves is $O(n^2)$, and they can

1138: easily be computed in $O(n^2)$ time.

1139: 

1140: \begin{figure}

1141: \centerline{\includegraphics[scale=\coinscale]{flip_3L.eps}}

1142: \caption{\label{flip 3L}

1143:   Flipping an L consisting of 3 coins.  Extra coins are shaded.}

1144: \end{figure}

1145: 

1146: \begin{figure}

1147: \centerline{\includegraphics[scale=0.6]{flip_L.eps}}

1148: \caption{\label{flip L}

1149:   Flipping a general L, using the subroutines in Figure \protect\ref{flip 3L}.}

1150: \end{figure}

1151: 

1152: The more difficult re-orientation to perform is a \emph{rotation} of an L by

1153: $\pm 90^\circ$.  Perhaps one of the most surprising results of this paper is

1154: that this operation is possible with two extra coins.  One way to do it for a

1155: square span, shown in Figure \ref{rotate L 4x4}, is to convert the L into a

1156: diagonal, and then convert more and more of the diagonal into a rotated L.

1157: This is the basis for our ``diagonal-flipping'' puzzle in Figure

1158: \ref{diagonal}.

1159: 

1160: \begin{figure}

1161: \centerline{\includegraphics[scale=\coinscale]{rotate_L_4x4.eps}}

1162: \caption{\label{rotate L 4x4}

1163:   One method for rotating an L with a square span.  Although this example

1164:   places the extra coins in the final configuration, this is not necessary.}

1165: \end{figure}

1166: 

1167: A simpler way to argue that L's can be rotated is shown in Figure \ref{rotate

1168: L}.  Assume without loss of generality that the initial orientation is the

1169: canonical L.  First we apply induction to the subconfiguration of all coins

1170: except the top-left coin.  Thus all rows except the third row contain at most

1171: one coin each, assuming the L consists of at least three rows.  Now we apply

1172: local operations in $3 \times 3$ or $3 \times 2$ rectangles (similar to Figure

1173: \ref{flip 3L}) to move the top-left coin to the far right.  (We cannot perform

1174: this left-to-right motion in one step using induction, because there may be

1175: only three rows, and hence all coins may be involved in this motion.)  Finally

1176: we flip the L in the top three rows, as described above, thereby obtaining the

1177: desired result.  Again the number of moves and computation time are both

1178: $O(n^2)$.  Note that the same approach of repeated local operations applies

1179: when the L consists of only two rows.

1180: 

1181: \begin{figure}

1182: \centerline{\includegraphics[scale=0.5]{rotate_L.eps}}

1183: \caption{\label{rotate L}

1184:   A general method for rotating an L.  The first step is to apply induction,

1185:   and the remaining steps apply subroutines similar to

1186:   Figure \protect\ref{flip 3L}.}

1187: \end{figure}

1188: 

1189: \subsubsection{Merging L's}

1190: 

1191: %--- Old outline:

1192: % route them close to each other

1193: % if unnecessary coins, done

1194: % else remove an end of an L that matters, apply induction

1195: % how to add a coin back (shift it along using 3-coin subconfigurations)

1196: 

1197: Consider two L's $L_1$ and $L_2$ whose bounding rectangles $R_1$ and

1198: $R_2$ are distance at most two from each other.  Equivalently, consider two L's

1199: such that $\span (L_1 \cup L_2)$ has a single connected component.

1200: This section describes how to merge $L_1$ and $L_2$ into a single L.

1201: This step is the most complicated part of the algorithm, not because it is

1202: difficult in any one case, but because there are many cases involved.

1203: 

1204: First suppose that the rectangles $R_1$ and $R_2$ overlap.

1205: We claim that one of the L's, say $L_1$, can be re-oriented so that one of its

1206: coins is contained in the other L's bounding rectangle, $R_2$.  This coin is

1207: therefore in the span of the $L_2$, and hence redundant, so as described above

1208: we can apply induction and finish the entire canonicalization process.

1209: 

1210: To prove the claim, there are three cases; see

1211: Figure \ref{merge L overlapping}.

1212: If a corner of one of the bounding rectangles, say $R_1$, is in the other

1213: bounding rectangle, $R_2$, then we can re-orient $L_1$ so that one of its end

1214: coins is at that corner of $R_1$ and hence in $R_2$; see

1215: Figure \ref{merge L overlapping}(a)).

1216: Otherwise, we have rectangles that form a kind of ``thick plus sign''

1217: (Figure \ref{merge L overlapping}(b--c));

1218: we distinguish the two rectangles as according to whether they form the

1219: \emph{horizontal stroke} or \emph{vertical stroke} of the plus sign.

1220: If the vertical stroke has width at least two

1221: (Figure \ref{merge L overlapping}(b)),

1222: then that rectangle already contains a coin of the other L,

1223: because that L cannot have two empty columns.

1224: Similarly, if the horizontal stroke has height at least two,

1225: then that rectangle already contains a coin of the other L, because that L

1226: cannot have two empty rows.

1227: Finally, if both strokes are of unit thickness, and there is not already a coin

1228: in their single-position intersection (Figure \ref{merge L overlapping}(c)),

1229: then we can splice and redefine the L's, so that one L is formed by the top

1230: half of the vertical stroke and the left half of the horizontal stroke, and the

1231: other L is formed by the bottom half of the vertical stroke and the right half

1232: of the horizontal stroke, and then we have the first case in which the bounding

1233: rectangles share a corner.

1234: 

1235: \begin{figure}

1236:   \centering

1237:   \includegraphics[scale=0.425]{merge_L_overlapping.eps}

1238:   \caption{\label{merge L overlapping}

1239:     Merging two L's with overlapping bounding rectangles (shaded).

1240:     (a) The corner of one L is contained in the other L's bounding rectangle.

1241:     (b) A ``thick plus sign'' in which at least one stroke has thickness

1242:     more than~$1$.  (c) A plus sign in which both strokes have thickness~$1$.}

1243: \end{figure}

1244: 

1245: Now suppose that the rectangles $R_1$ and $R_2$ do not overlap.  Hence, either

1246: they share no $x$ coordinates or they share no $y$ coordinates.  Assume by

1247: symmetry that $R_1$ and $R_2$ share no $x$ coordinates.  Assume again by

1248: symmetry that $R_1$ is to the left of $R_2$.  A \emph{leg} is a horizontal or

1249: vertical segment/edge of an L.  Re-orient $L_1$ so that its vertical leg is on

1250: the right side, and re-orient $L_2$ so that its vertical leg is on the left

1251: side.  Now $L_1$ and $L_2$ have distance at most three from each other; the

1252: distance may be as much as three because of parity.

1253: 

1254: We consider merging $L_1$ with each leg of $L_2$ one at a time.

1255: In other words, we merge $L_1$ with the nearest leg of $L_2$ within distance

1256: three of $L_1$, then we merge the result with the other leg of $L_2$.

1257: The second leg can be treated in the same way as the first leg, by induction.

1258: Thus there are two cases: either the first leg is the horizontal leg of $L_2$,

1259: or it is the vertical leg of $L_2$.

1260: We first show how the latter case reduces to the former case.

1261: 

1262: If the vertical leg of $L_2$ is the first leg, it can have only one coin within

1263: the $y$ range of $R_1$ (by the assumption that extra coins are picked up).  We

1264: can add this coin separately, as if it were a short horizontal leg of its own

1265: L.  This leaves a portion of the vertical leg of $L_2$ outside of the $y$ range

1266: of $R_2$.  Thus what remain of $R_1$ and $R_2$ do not share any $y$

1267: coordinates, so we can rotate the picture $90$ degrees and return to a

1268: horizontal problem.

1269: 

1270: This argument reduces merging two L's to at most three merges between an L and

1271: a horizontal leg.  Still several cases remain, as illustrated in Figure

1272: \ref{merge L I}.  Case 1 is when the horizontal leg is aligned with a coin in

1273: the L.  Case 2 is when they are out of alignment.  Case 3 is a special case

1274: occurring at the corner of the L, where the horizontal legs are aligned but

1275: distances are higher than in Case 1.  The above three cases are subdivided into

1276: subcases (a) and (b), depending on how close the horizontal leg is to the L.

1277: Finally, Case 4 is when the L and horizontal leg do not share $x$ or $y$

1278: coordinates.

1279: 

1280: \begin{figure}

1281: \centerline{\includegraphics[scale=0.37]{merge_L_I.eps}}

1282: \caption{\label{merge L I}

1283:   Merging an L and the horizontal leg of a nearby L.}

1284: \end{figure}

1285: 

1286: By the procedures in Figure \ref{merge L I}, in all cases, the merging can be

1287: done in $O(1)$ flips and rotations of L's, $O(1)$ leapfrogs, and $O(n)$

1288: shuffles.  In total, $O(n^2)$ moves are required to merge an L and a horizontal

1289: leg, or equivalently to merge two L's.

1290: 

1291: %--- Version in original submission: Lacked many details and figures,

1292: %    and best analysis I could formalize was O(n^4) in total.

1293: % Now consider a coin $t$ (in $L_2$) that has maximum $x$

1294: % coordinate, and subject to this restriction is farthest from $R_1$ as possible.

1295: 

1296: % We claim that $\span (L_1 \cup L_2 - t)$ has a single connected component, just

1297: % like $\span (L_1 \cup L_2)$.  There are several cases.  If $L_2$ has only one

1298: % coin, then removing $t$ reduces the span to $\span L_1$, which is connected.

1299: % If $R_2$ has more than one column, then $t$ is the only coin in its column, so

1300: % removal of $t$ simply removes the rightmost column from the span.  So suppose

1301: % $R_2$ has only one column but that $L_2$ has more than one coin.  Thus there

1302: % are several coins in $L_2$ with maximum $x$ coordinate, and we chose $t$ to be

1303: % as far from $R_1$ as possible.  If $t$ shares a $y$ coordinate with $R_1$, then

1304: % removal of $t$ will not affect the span because of other coins in $L_2$.

1305: % Otherwise, removal of $t$ removes either the topmost or bottommost row from

1306: % the span.  In all cases, the span remains connected.

1307: 

1308: % \xxx{The above probably needs a figure to convince the reader.}

1309: 

1310: % Now we can apply the induction hypothesis to $L_1 \cup L_2 - t$ and re-arrange

1311: % it into an L.  By re-orienting the L, we can arrange for $t$ to be within

1312: % distance two from the L.  Using local operations each involving at most three

1313: % coins, we can move $t$ to the end of the L, thereby forming a slightly

1314: % longer chain.  Finally, this chain can be re-arranged into an L using

1315: % more local operations.

1316: 

1317: % \xxx{Also draw a figure for the above.}

1318: 

1319: % Because we may have to re-orient L's, each step in this induction may cost as

1320: % much as $\Theta(n^2)$, for a total cost of $O(n^3)$ per merge.

1321: 

1322: % % We first apply the re-orientation method of the previous section so that the

1323: % % two L's have distance at most three, i.e., there are two coins of different L's

1324: % % with distance at most three.  If $R_1$ and $R_2$ do not share any $x$ or $y$

1325: % % coordinates, we can orient the L's to have two ends within distance two of each

1326: % % other.  If $R_1$ and $R_2$ overlap (have distance two), then we can orient the

1327: % % L's so that they ``overlap'' (have distance at most two).  If $R_1$ and $R_2$

1328: % % are distance one or two from each other, say along horizontal edges, then we

1329: % % can orient the L's against these horizontal edges.  Because of parity we may

1330: % % end up with every two coins from different L's having distance at least three.

1331: % %

1332: % % [figure abut_L]

1333: 

1334: \subsection{Final Sweep}

1335: \label{Final Sweep}

1336: 

1337: Thus far we have shown how to reversibly re-arrange a configuration ($A$ or

1338: $B$) into the canonical configuration, using two extra coins.  However, during

1339: this process, we may have picked up extra coins, and now need to drop them

1340: appropriately.  In reality, these coins sit in arbitrary locations on the

1341: board.  For re-arranging the source configuration $A$ into the canonical

1342: configuration, the moves need not be reversible, so we can simply drop the

1343: extra coins in the canonical order, as in Figure \ref{canonical}.  For

1344: re-arranging the destination configuration $B$ into the canonical

1345: configuration, we need to effectively drop these coins by a sequence of reverse

1346: moves.

1347: 

1348: More directly, starting from the canonical configuration, we need to show how

1349: to distribute the extra coins to arbitrary locations on the board.  We

1350: can achieve this effect by making a complete sweep over the board.  More

1351: precisely, we flip the L as in Section \ref{Re-orienting L's}, which has the

1352: effect of passing over every position on the board with the operations shown in

1353: Figure \ref{flip 3L}.  During this process, we will pass over the extra coins;

1354: at this point we treat them as if they were picked up, applying the emulation

1355: in Section \ref{Picking Up and Dropping Tokens}.

1356: Then we flip the L back to its original orientation.

1357: On the way back, whenever we apply an operation in Figure \ref{flip 3L} and

1358: pass over the desired destination $d$ for one of the extra coins, we move the

1359: extra coin to $d$ while there are at least two adjacent coins from the L.

1360: By monotonicity of the flipping process, this extra coin will not be passed

1361: over later by the flip, so once an extra coin is placed in its desired

1362: location, it remains there.

1363: 

1364: \subsection{Reducing Span}

1365: 

1366: Now that we know any configuration can be brought to the corresponding

1367: canonical configuration with a sequence of (mostly) reversible moves, it

1368: follows immediately that any configuration can be re-arranged into any

1369: configuration with the same span.  More generally, if we are given

1370: configurations $A$ and $B$ satisfying $\span A \supseteq \span B$, we can first

1371: pick up all coins in $A \setminus \span B$, then reversibly re-arrange both

1372: configurations into the same canonical configuration.  Putting these two

1373: sequence of moves together, we obtain a re-arrangement from $A \setminus \span

1374: B$ to $B$ with some coins missing.  Then we simply drop the previously picked

1375: up coins in the appropriate positions to create $B$.

1376: 

1377: Note that these moves need not be reversible, because we are only concerned

1378: with the direction from $A$ to $B$.  Indeed, the moves cannot be made

1379: reversible, because the span cannot increase (Lemma \ref{span monotone}).

1380: 

1381: This concludes the proof of Theorem \ref{square grid}.

1382: 

1383: \subsection{Lower Bound}

1384: \label{Lower Bound}

1385: 

1386: The bound on the number of moves in Theorem \ref{square grid} is in fact tight:

1387: 

1388: \begin{theorem}

1389: The ``V to diagonal'' puzzle in Figure \ref{lower bound sqr}

1390: requires $\Theta(n^3)$ moves to solve.

1391: \end{theorem}

1392: 

1393: \begin{figure}

1394: \centerline{\includegraphics[scale=\coinscale]{lower_bound_sqr.eps}}

1395: \caption{\label{lower bound sqr}

1396:   Re-arranging the V-shape in the upper left into the diagonal in the

1397:   lower right requires repeated rotations

1398:   of diagonals as in Figure \protect\ref{diagonal} (or repeated rotations of

1399:   L's).}

1400: \end{figure}

1401: 

1402: \begin{proof}

1403: We claim that re-arranging the V shape into a diagonal effectively requires

1404: repeated ``diagonal flipping.''  At any time, only one component of coins can

1405: be actively manipulated (drawn with dotted lines in the figure); all other

1406: coins are isolated from movement.  Thus we must repeatedly re-arrange the

1407: active component so that it can reach the nearest isolated coin.  More

1408: specifically, we must re-arrange the active component into a chain starting at

1409: the corner of the bounding rectangle that is near the isolated coin, and ending

1410: at the opposite corner of the bounding rectangle of the active component.

1411: These two corners alternate for each isolated coin we pick up, and that is the

1412: sense in which we must ``flip a diagonal.''  It is fairly easy to see that

1413: each diagonal flipping of a chain with $k$ coins takes $\Omega(k^2)$ time.

1414: In total, the puzzle requires $\Theta(\sum_{k=1}^n k^2) = \Theta(n^3)$ moves.

1415: \end{proof}

1416: 

1417: This theorem is the motivation for the puzzle in Figure \ref{vee}.

1418: 

1419: \subsection{Labeled Coins}

1420: \label{Labeled Coins}

1421: 

1422: We conjecture that Theorem \ref{square grid} holds even when coins are labeled,

1423: subject to a few constraints.  The idea is that permutation of the coins is

1424: relatively easy once we reach the canonical configuration.  Examples of methods

1425: for swapping coins within one L are shown in Figure \ref{swap L}.  The top

1426: figure shows how to swap a pair of coins when the canonical configuration is

1427: nothing more than a canonical L.  The middle figure shows how to perform the

1428: same swap when there are four additional coins.  Note that swapping the corner

1429: coin 3 works in exactly the same way; indeed, this method works whenever the

1430: coins to be swapped have two other coins adjacent to them, and there is another

1431: valid destination.  The bottom figure shows how to swap one of the end coins,

1432: which is more difficult.  This last method begins with moving the bend of the L

1433: toward the end coin, and then works locally on the coins $1,2,3$.

1434: 

1435: \begin{figure}

1436: \centerline{\includegraphics[scale=0.6]{swap_L.eps}}

1437: \caption{\label{swap L}

1438:   Three cases of swapping coins in the canonical configuration.

1439:   The coins to be swapped have a thick outline.}

1440: \end{figure}

1441: 

1442: One obvious constraint for these methods is that if there are no valid moves,

1443: permutation is impossible.  Also, if the bounding rectangle of an L has width

1444: or height $1$, then the two end coins of the L cannot be moved.  Subject to

1445: these constraints, Figure \ref{swap L} proves that the coins in a single

1446: connected component of the span, other than the extra coins $e_1$ and $e_2$,

1447: can be permuted arbitrarily.

1448: 

1449: It only remains to show that a coin can be swapped with $e_1$ or $e_2$, which

1450: implies that coins between different connected components of the span can be

1451: swapped.  We have not proved this in general yet, but one illustrating example

1452: is the puzzle in Figure \ref{swap5}, whose solution is shown in Figure

1453: \ref{swap5 solution}.  The idea is that coins 2 and 4 are $e_1$ and $e_2$, and

1454: so we succeed in swapping $e_2$ with coin 3.  A slight generalization of this

1455: approach may complete a solution to the labeled coins.

1456: 

1457: \begin{figure}

1458: \centerline{\includegraphics[scale=\coinscale]{swap5_solution.eps}}

1459: \caption{\label{swap5 solution}

1460:   Solution to the puzzle in Figure \protect\ref{swap5}.}

1461: \end{figure}

1462: 

1463: \subsection{Fewer Extra Coins}

1464: \label{Fewer Extra Coins}

1465: 

1466: We have shown that the configuration space is essentially strongly connected

1467: provided there is a pair of extra coins, i.e., the removal of these two coins

1468: does not reduce the span.  This section summarizes what we know about

1469: configurations without this property.

1470: 

1471: If we have a span-minimal configuration with no extra coins, Lemma

1472: \ref{span-minimal fragile} tells us that every move decreases the span.  With

1473: an overhead of a factor of $n^2$, we can simply try all possible moves, in each

1474: case obtaining a configuration with smaller span, which furthermore must have

1475: an extra coin (the moved coin).  Now we only need to recursively check these

1476: configurations.

1477: 

1478: Unfortunately, the situation is trickier with one extra coin.  The key

1479: difficulty is that multiple coins could individually be considered extra,

1480: but no pair of coins is extra.  In other words, there may be two coins such

1481: that removing either one does not reduce the span, but removing both of them

1482: reduces the span.  Two simple examples are shown in Figure \ref{two extra}.

1483: 

1484: \begin{figure}

1485: \centerline{\includegraphics[scale=\coinscale]{two_extra.eps}}

1486: \caption{\label{two extra}

1487:   The shaded coins are individually extra, but do not form a pair of

1488:   extra coins suitable for Theorem \protect\ref{square grid}.}

1489: \end{figure}

1490: 

1491: This difficulty makes ``one'' extra coin surprisingly powerful.  For example,

1492: using one extra coin, an L with odd parity can be flipped, although it cannot

1493: be rotated, and an L with even parity cannot be flipped or rotated.

1494: In Figure \ref{one extra} we exploit this property to make an interesting

1495: solvable puzzle initially with no pair of extra coins; it takes significant

1496: work before a pair of extra coins appears.

1497: 

1498: \begin{figure}

1499: \centerline{\includegraphics[scale=\coinscale]{one_extra.eps}}

1500: \caption{\label{one extra}

1501:   A puzzle on the square grid with no initial pair of extra coins.}

1502: \end{figure}

1503: 

1504: \section{Conclusion}

1505: \label{Conclusion}

1506: 

1507: We have begun the study of deciding solvability of coin-moving puzzles and more

1508: generally token-moving puzzles.  We gave an exact characterization of solvable

1509: puzzles with labeled coins on the equilateral-triangle grid.  By introducing

1510: the notion of a constant number of extra coins, we have given a tight theorem

1511: characterizing solvable puzzles on the square grid.  Specifically, we have

1512: shown that any configuration can be re-arranged into any configuration with the

1513: same or smaller span using two extra coins, and that this is best possible in

1514: general.  The number of moves is also best possible in the worst case.

1515: 

1516: Several open questions remain:

1517: 

1518:   \begin{enumerate}

1519:   \item What is the complexity of solving a puzzle using the fewest moves?

1520:         %--- This is flat-out wrong:

1521:         %Even the exact combinatorics of turning a general pyramid upside-down

1522:         %as in Figure \ref{upside-down} are open \cite{Gardner-1975-penny}.

1523:   \item How do our results change if moves are forced to be \emph{slides}

1524:         that avoid other coins?  We conjecture that Theorem \ref{triangular

1525:         grid} still holds for unlabeled coins.

1526:   %\item What puzzles on the square grid with zero or one extra coins are

1527:   %      solvable?  Can this be decided in polynomial time?

1528:   %\item Can our results about square grids be generalized to labeled

1529:   %      coins?  We outline some approaches to solving this in

1530:   %      Section \ref{Labeled Coins}.

1531:   \item Can we extend our results on the square grid to the hypercube lattice

1532:         in any dimension?

1533:   \item Can we combine Theorems \ref{triangular grid} and \ref{square grid}

1534:         to deal with a mix of the square and equilateral-triangle lattice,

1535:         like the second puzzle in Figure \ref{HOH}?

1536:   \item Can we prove similar results for general graphs?

1537:   \end{enumerate}

1538: 

1539: 

1540: \section*{Acknowledgments}

1541: 

1542: We thank J.\ P.\ Grossman for writing a program to find optimal solutions to

1543: the puzzles in Figures \ref{diagonal} and \ref{spindle}.

1544: 

1545: \bibliography{coinsliding}

1546: \bibliographystyle{plain}

1547: 

1548: \end{document}

1549: