1: %% Flipping Cubical Meshes, Bern, Eppstein, and Erickson

2: 

3: \documentclass[11pt]{llncs}

4: \usepackage{graphicx}

5: \usepackage{cite}

6: \usepackage{url}

7: \urlstyle{same}

8: 

9: \usepackage{times}

10: \usepackage{mathfont}

11: \def\log{\mathop{{\rm log}}}

12: \def\min{\mathop{\rm min}}

13: \def\max{\mathop{\rm max}}

14: \def\sin{\mathop{\rm sin}}

15: \def\cos{\mathop{\rm cos}}

16: 

17: \newtheorem{open}{Open Problem}

18: 

19: \DeclareSymbolFont{AMSb}{U}{msb}{m}{n}

20: \DeclareSymbolFontAlphabet{\Bbb}{AMSb}

21: \def\Real{\ensuremath{\Bbb R}}

22: 

23: \let\sevsfb\small

24: 

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

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

27: \setlength{\evensidemargin}{0in}

28: \setlength{\oddsidemargin}{0in}

29: \setlength{\topmargin}{-.5in}

30: 

31: % magic to make big-O notation use script font

32: \mathcode`O="724F

33: 

34: % fix broken caption in llncs

35: \makeatletter

36: \def\hb@xt@{\hbox to }

37: \makeatother

38: 

39: % MARK ENDS OF PROOFS!

40: \let\oldendproof\endproof

41: \def\endproof{\qed\oldendproof}

42: 

43: \newcommand{\OCfigure}[3]{

44: \begin{figure}[t]

45:  \centering\includegraphics[width=#3]{#1}

46:  \caption{#2}

47:  \label{#1}

48: \end{figure}}

49: 

50: \let\SCfigure\OCfigure

51: %\renewcommand\topfraction{.5}

52: %\renewcommand\floatpagefraction{.7}

53: %\renewcommand\textfraction{.3}

54: \renewcommand\dbltopfraction{.7}

55: 

56: \pagestyle{plain}

57: 

58: \begin{document}

59: 

60: \title{Flipping Cubical Meshes} 

61: 

62: \author{Marshall Bern\inst{1},

63: David Eppstein\inst{2}, and

64: Jeff Erickson\inst{3}}

65: 

66: \institute{Xerox Palo Alto Research Center\\

67: 3333 Coyote Hill Rd., Palo Alto, CA 94304\\

68: \email{bern@parc.xerox.com}

69: \medskip

70: \and

71: University of California, Irvine\\

72: Dept. of Information \& Computer Science,

73: Irvine, CA 92697\\

74: \email{eppstein@ics.uci.edu}

75: \medskip

76: \and

77: Univ. of Illinois, Urbana-Champaign\\

78: Dept. of Computer Science,

79: 1304 W. Springfield Ave.,

80: Urbana, IL 61801\\

81: \email{jeffe@cs.uiuc.edu}}

82: 

83: \maketitle   

84: 

85: \begin{abstract}We define and examine flip operations

86: for quadrilateral and hexahedral meshes, similar to the flipping

87: transformations previously used in triangular and tetrahedral mesh

88: generation.

89: 

90: \medskip

91: {\bf Keywords:} quadrilateral meshing, hexahedral meshing, flip graph

92: connectivity, mesh improvement.

93: \end{abstract}

94: 

95: \section{Introduction}

96: 

97: In this paper we propose flipping transformations for cubical meshes,

98: such as quad and hex meshes. 

99: These flipping transformations are precise analogues of the 

100: well-known flipping transformations for simplicial meshes.  

101: We did not invent these transformations. 

102: They have already appeared---although they were not 

103: explicitly enumerated---in the mathematical

104: polytope literature~\cite{BilStu-AM-92,Zie-95}. 

105: We have not found any previous references 

106: in the meshing literature. The closest work seems

107: to be the quad refinement operations of Canann,

108: Muthukrishnan, and Phillips~\cite{CanMutPhi-EC-98}

109: and the hex reconnection primitives of 

110: Knupp and Mitchell~\cite{KnuMit-TR-99}.

111: 

112: As in the simplicial case, quad and hex flipping could be useful for mesh

113: generation, improvement, and refinement/derefinement. We also use

114: flipping to study existence questions for hexahedral

115: meshes.  Up until this present paper, the theoretical results on

116: quad-to-hex extension have been positive, suggesting that  (under

117: simple necessary conditions) it should always be possible.

118: Mitchell~\cite{Mit-STACS-96} and Thurston~\cite{Thu-93} independently

119: showed that any quad surface mesh with an even number of quadrilaterals,

120: and topologically equivalent to a sphere, can be extended to a {\it

121: topological\/} hex volume mesh, which allows elements to have warped,

122: nonplanar sides.  (An even number of surface quadrilaterals is a

123: necessary condition, because hex elements each have six sides,  and each

124: interior face will account for two hex sides.) 

125: Eppstein~\cite{Epp-CGTA-99} gave another proof of this result and

126: improved the complexity bound from $O(n^2)$ to $O(n)$ elements. Each of

127: the two proofs generalize to certain more topologically  complicated

128: inputs as well.  However, it has not been clear under what circumstances

129: these topological meshes can be realized using flat-faced cuboids. In

130: this context we discuss a simple 10-element boundary complex, the {\em

131: bicuboid}, which we have been unable to mesh, and use flipping to show

132: that many possible topological meshes for this complex are unrealizable.

133: 

134: This is primarily a theoretical paper---realizability

135: is not of great practical importance since the

136: isoparametric trilinear elements typically used with unstructured

137: hexahedral meshes allow the cells to have curved boundaries.

138: However, the existence or nonexistence of flat-faced meshes may

139: shed light on the question of how many curved elements are needed to achieve satisfactory mesh quality.  In addition,

140: our study of flip graph connectivity relates to mesh

141: improvement algorithms: by characterizing the connected components of the

142: quadrilateral flip graph for simply-connected regions

143: (Theorem~\ref{thm:qfc}) we show that any local improvement operation

144: that remeshes bounded regions of a quadrilateral mesh can be simulated

145: by a sequence of simpler flip and parity-change operations.

146: 

147: \section{Definitions}\label{def-sec}

148: 

149: A {\it polyhedron\/} is a set of closed planar

150: polygonal faces in $\Real^3$, that intersect

151: only at shared vertices and edges, and such that each edge is shared

152: by exactly two faces and each vertex is shared by exactly one

153: cycle of faces.\footnote{ This definition of polyhedron allows  

154: multiple connected components (two cubes can be one polyhedron!),

155: as well as voids and tunnels.

156: It disallows non-manifold boundaries, such as those used to 

157: model objects made from two different materials.}

158: A polyhedron divides $\Real^3$ into a bounded

159: {\it interior\/} and an unbounded {\it exterior\/}, each of which

160: may contain more than one connected component.

161: 

162: A {\it cuboid\/} in $\Real^d$ is a convex polytope, 

163: combinatorially equivalent to the unit cube $[0,1]^d$.  

164: {\it Combinatorial equivalence\/} means that the faces 

165: (vertices, edges, and so forth) of the cuboid are

166: in one-to-one correspondence with the faces of the unit cube,

167: and that this correspondence preserves intersections. 

168: In $\Real^2$ a cuboid is simply a convex quadrilateral,

169: called a {\it quad\/} for short. 

170: In $\Real^3$ a cuboid is a convex polyhedron called a {\it hex\/}. 

171: 

172: A {\it quad surface mesh\/} is a polyhedron

173: whose faces are all quads.

174: A {\it hex mesh\/} is a quad surface mesh, along with interior quads 

175: that subdivide the polyhedron interior into hexes.  

176: The intersection of any pair of hexes is either empty, a single

177: vertex, a single edge, or a single quad. 

178: Quad surface meshes and hex meshes are both examples

179: of {\it cubical complexes\/} \cite{BliBli-DM-98,JosZie-DCG-00,math.GT/0204085}, the analogue of simplicial complexes in which each $k$-dimensional face is a $k$-dimensional cuboid.

180: 

181: We also consider {\em topological hex meshes}: cubical complexes

182: homeomorphic to the input domain, with a boundary mesh combinatorially equivalent to that of the input.  In this context,

183: we call the quad and hex meshes defined earlier {\em geometric meshes}

184: to emphasize the geometric requirements on their cell shapes. Topological

185: meshes can be {\em realized} by specifying locations in the domain for

186: each internal vertex of the complex; the realization is a hex mesh if

187: each cell's eight vertex locations have a convex hull that is a cuboid,

188: and no two of these cuboids have intersecting interiors. A realization is

189: {\em self-intersecting} if each four vertices of a quadrilateral in the

190: mesh have locations that form a convex quadrilateral in the domain,

191: without respect for whether these quadrilaterals form the boundaries of

192: disjoint cuboids; a self-intersecting mesh can be viewed as a

193: generalized geometric mesh in which some cells are allowed to be

194: inverted.

195: 

196: \section{Flipping}\label{flipping-sec}

197: 

198: \OCfigure{triflips}{The 2-2 and 1-3 flips for triangular meshes

199: correspond to switching upper and lower facets of a tetrahedron.}

200: {4.2in}

201: 

202: \OCfigure{jeffqf}{Flips for quad

203: meshes correspond to switching upper and lower facets of a

204: three-dimensional cuboid.  The bold cuboid edges mark the

205: division between upper and lower facets.} {6.5in}

206: 

207: Consider the well-known {\it flip\/} transformation

208: for a triangular mesh, which

209: switches one diagonal for the other in a convex quadrilateral

210: $Q$ formed by two triangular elements sharing a side. 

211: We can think of this 2-2 flip (two elements in and

212: two elements out) as exchanging the upper and lower

213: facets of a three-dimensional simplex $P$ that projects to $Q$,

214: as shown on the left in Figure~\ref{triflips}.

215: This view leads us to consider the 

216: refinement transformation that subdivides a triangle into

217: three smaller ones by adding an interior vertex

218: as a 1-3 flip (Figure~\ref{triflips} right side). 

219: The reverse of the 1-3 flip is a 3-1 flip, which

220: derefines the triangulation by removing a vertex.

221: 

222: The almost-as-well-known flips for tetrahedral meshes

223: can be defined analogously by the exchange of the upper and lower

224: facets of a four-dimensional simplex.  

225: The flips include the 2-3 flip, which exchanges two tetrahedra

226: sharing a face for three sharing an edge, and the 1-4 flip

227: which subdivides a tetrahedron by adding an interior vertex,

228: along with their reverse flips.

229: 

230: This view of flipping extends to quad and hex meshes.

231: The flips for quad meshes are induced by the combinatorially

232: distinct exchanges of upper and lower facets of a three-dimensional cuboid,

233: as shown in Figure~\ref{jeffqf}, and the flips for hexahedral meshes are induced by the combinatorially distinct exchanges of upper and lower facets of a four-dimensional hypercube, as shown in Figure~\ref{hextriangle}.

234: 

235: The facets of a $d$-dimensional cube or hypercube can be grouped into $d$ opposite pairs.

236: In order to discover all possible flips, we describe any set of facets of a cube or hypercube by how it relates to these pairs.  To any set $S$ of cube or hypercube facets, we assign

237: a pair of integers $(X,Y)$, where $X$ denotes the number of pairs of facets eXcluded from $S$, and $Y$ denotes the number of pairs of facets Yncluded in $S$.

238: The remaining $d-X-Y$ pairs of facets have one of their two facets included in $S$, and the other facet excluded from $S$.  The symmetries of the hypercube include all permutations of opposite pairs of facets,

239: and all swaps of the two facets in each pair, so any two sets $S$ described by the same pair $(X,Y)$ are combinatorially equivalent.

240: 

241: The meshes in Figures \ref{jeffqf}

242: and~\ref{hextriangle} are labeled by these pairs of numbers.  If a set $S$ is represented by a pair $(X,Y)$, the complementary set is represented by pair $(Y,X)$.  So, a flip can be described as the replacement of a set of mesh cells combinatorially equivalent to the set $(X,Y)$ by a different set of cells combinatorially equivalent to $(Y,X)$.  In order to find all possible flips, it remains only to determine which pairs $(X,Y)$ form topological disks.

243: 

244: \begin{lemma}

245: Let $S$ be a set of facets of a $d$-hypercube described by pair $(X,Y)$.

246: Then $S$ forms a topological disk if and only if $X+Y<d$.

247: \end{lemma}

248: 

249: \begin{proof}

250: If $X+Y<d$, there is some facet $f\in S$ such that the opposite facet is not contained in $S$; then $f$ is adjacent to every facet in $S$.  We can retract all facets to $f$ by an affine transformation that shrinks the coordinate axis perpendicular to $f$; therefore $S$ has a retraction to a $(d-1)$-dimensional disk and is itself a disk.

251: 

252: On the other hand, suppose $X+Y=d$.  We show by induction on $X$ that $S$ has a retraction to a $Y$-sphere, and therefore cannot be a disk.  More specifically, we can retract $S$ in a facet-preserving way to the boundary of a $Y$-hypercube.

253: As a base case, if $X=0$ and $Y=d$, $S$ is the whole boundary of the $d$-hypercube.

254: Otherwise, $S$ is formed by removing a pair of facets from a set $S^+$ that can be retracted to a $(Y+1)$-hypercube $C^+$.  Removing the final two facets from $S^+$ corresponds, in its retraction, to removing two opposite facets from the boundary of $C^+$, and the remaining facets of $C^+$ can be contracted to a $Y$-hypercube by again using an affine map that contracts the coordinate axis perpendicular to the removed facets.

255: \end{proof}

256: 

257: We cannot allow sets $S$ that do not form disks to count as flips, because exchanging such a set for its complement would change the topology of the domain.

258: The possible flips in a $d$-dimensional cubical mesh correspond to subsets of a $(d+1)$-hypercube, and are therefore given by the pairs $(X,Y)$--$(Y,X)$ where $X+Y\le d$.

259: 

260: \begin{corollary}

261: The $d$-dimensional cubical meshes have

262: $(d+1)(d+2)/2$ combinatorially distinct flippable submeshes $(X,Y)$

263: which can be grouped into $\lfloor (d+2)^2/4 \rfloor$ flip pairs $(X,Y)$--$(Y,X)$.

264: \end{corollary}

265: 

266: The possible quadrilateral mesh flips are therefore

267: $(2,0)$--$(0,2)$ (changing one quadrilateral for five),

268: $(1,0)$--$(0,1)$ (changing two quadrilaterals for four),

269: and two changes of three for three quads:

270: $(0,0)$--$(0,0)$ and $(1,1)$--$(1,1)$.

271: along with their reverses. We can picture the $(2,0)$ mesh as corresponding to the unique lowest face of a cuboid, and the flipped $(0,2)$ mesh as the unique highest cuboid face together with the four neighboring side faces. The $(1,0)$--$(0,1)$ flip can similarly be pictured with

272: unique lowest and highest edges, and the $(0,0)$--$(0,0)$ flip can be pictured with unique lowest and highest vertices.  The remaining

273: $(1,1)$--$(1,1)$ flip transforms the top, front, and back faces of a cube into the bottom, left, and right faces.

274: 

275: \OCfigure{hextriangle}{The flips for hex

276: meshes correspond to switching upper and lower facets of a

277: four-dimensional cube.}{6.5in}

278: 

279: Analogously, the flips for hex meshes are induced by all possible

280: exchanges of upper and lower facets of a four-dimensional cuboid, as

281: shown in Figure~\ref{hextriangle}. The $(3,0)$--$(0,3)$ flip converts one hex to seven, and

282: can be pictured with unique lowest and highest facets of a hypercuboid in

283: $\Real^4$. The $(2,0)$--$(0,2)$ converts two hexes to six, and can be pictured with unique highest and lowest two-dimensional

284: faces of a hypercuboid. There are two different flips from three to five hexes: $(1,0)$--$(0,1)$ and

285: $(2,1)$--$(1,2)$.  In the $(2,1)$ three-hex mesh, the hexes are connected in a path. This mesh  can be pictured most symmetrically with a unique lowest hypercuboid facet

286: bordered by two other mutually disjoint almost-lowest facets, and the complementary $(1,2)$ mesh can be pictured with a unique highest facet (a sort of hex

287: tunnel in Figure~\ref{hextriangle}) surrounded by four other high facets.

288: The other three-hex mesh in our set of flips, $(1,0)$, has three hexes connected in a cycle,

289: and can be pictured  with a unique lowest hypercuboid

290: edge surrounded by three facets; its complementary five-hex mesh, $(0,1)$, has a

291: unique highest edge, whose endpoints do not appear on the lower convex

292: hull, surrounded by three facets.  Finally, there are two different flips involving sets of four hexes.

293: The $(1,1)$ flip corresponds to a lower-convex-hull facet connectivity of a clique minus an edge, and can be pictured with a unique lowest two-dimensional hypercuboid face; the $(0,0)$ flip has four mutually adjacent hexes, and can be pictured with a unique lowest hypercuboid vertex.

294: 

295: The regions of $\Real^3$ occupied by the cells of a flip are all combinatorially inequivalent:

296: a cuboid for the $(3,0)$--$(0,3)$ flip,

297: a bicuboid for the $(2,0)$--$(0,2)$ flip,

298: a twisted rhombic dodecahedron for the $(1,0)$--$(0,1)$ flip,

299: a tricuboid for the $(2,1)$--$(1,2)$ flip,

300: a fourteen-faced shape resembling a square orthobicupola for the $(1,1)$--$(1,1)$ flip, and a rhombic dodecahedron for the $(0,0)$--$(0,0)$ flip.

301: These shapes can be determined from the two-dimensional quadrilateral flips:

302: if one forms a polyhedron from the convex hull of two parallel copies of

303: an $(X,Y)$ quadrilateral mesh, the result is the region of an $(X,Y+1)$--$(Y+1,X)$ flip

304: (for instance, the $(0,2)$ bicuboid is formed by the convex hull of two parallel two-quadrilateral $(0,1)$ meshes), and if one instead forms the convex hull of an $(X,Y)$ mesh and a parallel $(Y,X)$ mesh, the resulting three-dimensional shape is the region for an $(X,Y)$--$(Y,X)$ flip.

305: 

306: Three of the hexahedral flips are listed by Knupp and

307: Mitchell~\cite{KnuMit-TR-99}: the $(3,0)$--$(0,3)$ flip is given as an example of

308: their ``pillowing'' transformation, the $(2,0)$--$(0,2)$ flip is given as an example

309: of their ``inflating hex ring'' transformation, and the $(0,0)$--$(0,0)$ flip

310: is called the ``rotate three hexes primitive'' despite its symmetric

311: action on four hexes.  The $(3,0)$--$(0,3)$ and $(2,0)$--$(0,2)$ flips are also used by

312: Mar\'echal~\cite{Mar-IMR-01} to reduce the complexity of some case

313: analysis in his conformal mesh refinement algorithm.

314: 

315: \section{Flippability}

316: 

317: \OCfigure{frustum}{The $(2,0)$--$(0,2)$ quad flip and $(3,0)$--$(0,3)$ hex flip can be performed

318: geometrically by projecting the truncation of a higher dimensional

319: pyramid.}{5.5in}

320: 

321: We say that a set of cells in a mesh is {\em flippable} if it can be

322: replaced by one of the flips described above.  For topological meshes,

323: a set of cells is flippable whenever it has the appropriate connectivity

324: pattern, but for geometric meshes additional conditions will likely

325: be required.  For instance, in triangular meshes, a pair of triangles is

326: flippable if and only if it forms a convex quadrilateral.

327: 

328: We say that a hex mesh is {\em generic} if no three vertices are

329: collinear and no two quadrilaterals are coplanar. We say that a type of flip is {\em automatically flippable}

330: if, whenever the configuration of cells on one side of the flip occurs

331: in a generic geometric or self-intersecting mesh, the flip can be

332: performed resulting in another generic (possibly self-intersecting)

333: mesh.  That is, for these types of flips, flippability of geometric

334: meshes may involve convexity constraints disallowing inverted elements,

335: but does not involve additional constraints on the flatness of cell

336: boundaries. In triangular, quadrilateral, or tetrahedral

337: meshes, all flips are automatically flippable, because all

338: cell boundaries are always flat.  The requirement that each

339: quadrilateral face be flat leads to more interesting theory in the

340: hexahedral case.

341: 

342: \begin{theorem}\label{automatic}

343: The $(3,0)$--$(0,3)$, $(0,3)$--$(3,0)$, $(2,0)$--$(0,2)$, $(0,2)$--$(2,0)$, and $(0,0)$--$(0,0)$ hex mesh flips are automatically flippable.

344: \end{theorem}

345: 

346: \begin{proof}

347: The result is trivial for the $(0,3)$--$(3,0)$ flip, since it just involves removal

348: of some faces from the $(0,3)$ mesh.  We can flip the $(3,0)$ mesh by viewing the

349: initial cuboid as the base of a pyramid in four dimensions, forming a

350: four-dimensional cuboid by truncating the pyramid's apex by a

351: hyperplane, and replacing the base by the orthogonal

352: projection of the truncated pyramid's remaining faces

353: (Figure~\ref{frustum}).  There are enough degrees of freedom in this

354: construction to preserve genericity.

355: 

356: To understand the $(2,0)$--$(0,2)$ flips, let's analyze the conditions under

357: which the six-cuboid $(0,2)$ mesh is possible. This mesh (shown in the upper

358: right corner of Figure~\ref{hextriangle}) breaks a {\em

359: bicuboid} into a {\em ring} of four cuboids connected to the bicuboid

360: sides, and a {\em core} of two cuboids connected to the top and bottom.

361: 

362: Let's first look at the ring. The planes $P_i$ ($i=0,1,2,3$) where its

363: cuboids connect to each other are determined by triples of

364: points on the bicuboid.  Each such plane contains four ring vertices,

365: three of which are already determined.  Fixing the position of the

366: fourth vertex $v_i$ on plane $P_i$ causes the planarity constraints of

367: two inner quadrilaterals to combine to fix the position of $v_{i+1}$ on

368: plane $P_{i+1}$, etc: $v_{i+1} = f_i(v_i)$.  In order for the ring to be

369: formed correctly, these functions must return back to the starting

370: position: $v_0=f_3(f_2(f_1(f_0(v_0))))$.  But each $f_i$ is an affine

371: transformation, and we already know three noncollinear points of $P_0$

372: that are returned to themselves by the composition

373: $f_3(f_2(f_1(f_0(x))))$: the three original bicuboid vertices.  So the

374: composition must be the identity map, and {\em any} point of $P_0$ is

375: mapped by the functions $f_i$ to four points determining a geometrically

376: embedded ring.

377: 

378: The only remaining question is, when are these four points coplanar?

379: We say that a point $x$ in $P_0$ is {\em plane-generating}

380: when its quadruple $x$, $f_0(x)$, $f_1(f_0(x))$, $f_2(f_1(f_0(x)))$ forms a planar quadrilateral.

381: Then the plane-generating points form a linear subspace of $P_0$: if you have any two such

382: points $p$ and $q$, and a third point $r$ on line $pq$, then each point in the quadruple of $r$

383: is just the same linear combination of the images of $p$ and $q$,

384: so the quadruple of $r$ lies on a plane that is the same linear combination

385: of the planes for $p$ and $q$.  Note also that we already know two points

386: in $P_0$ that determine a coplanar quadrilateral: the vertices of the top

387: and bottom faces of the bicuboid.  So if there were a third affinely

388: independent point determining a coplanar quadrilateral then {\em all}

389: points of $P_0$ would determine a coplanar quadrilateral.

390: So the existence of a flat middle face for the $(2,0)$ mesh determines the

391: existence of a flat middle face for the $(0,2)$ hex and vice versa.

392: Since there are two degrees of freedom for placing the points of the

393: middle face, we can do so in such a way that the mesh remains generic.

394: 

395: Finally, consider the $(0,0)$--$(0,0)$ flip. The hexes of a $(0,0)$ mesh form a rhombic dodecahedron.

396: Color the degree-4 vertices of the 

397: dodecahedron gray, and color the degree-3 vertices alternately white 

398: and black, so that the cuboids in the given $(0,0)$ mesh have

399: one white vertex each.  To show that it can be 

400: subdivided into black cuboids, lift the vertices of the mesh into 

401: $\Real^4$ by assigning $w$-coordinates as follows.  Place the interior 

402: point $a$ (the apex) on the plane $w=1$ and the gray vertices 

403: $g_1,g_2,g_3,g_4,g_5,g_6$ on the plane $w=0$.  Each quadruple 

404: $ag_ig_jg_k$ determines a hyperplane, and these hyperplanes determine 

405: the $w$-coordinates of the black and white vertices.  The lifted 

406: cuboid mesh is the upper surface of a hypercuboid.  The single 

407: remaining hypercuboid vertex~$z$ (the zenith) is located at the 

408: intersection of four hyperplanes, each determined by a black point and 

409: its three gray neighbors.  Projecting the lower facets of the 

410: resulting hypercuboid back to $\Real^3$ gives us the flipped cuboid 

411: mesh.  In particular, the projection of $z$ lies in the interior of 

412: the original rhombic dodecahedron.

413: \end{proof}

414: 

415: The same argument shows that, for geometric hex meshes, regardless of

416: genericity, the $(3,0)$--$(0,3)$, $(0,3)$--$(3,0)$, $(2,0)$--$(0,2)$, $(0,2)$--$(2,0)$,

417: and $(0,0)$--$(0,0)$ flips are possible if and only if

418: the set of cuboids to be flipped forms a strictly convex subset of the

419: domain.

420: 

421: \SCfigure{unflippable}{Unflippable $(2,1)$ configuration.} {2in}

422: 

423: \SCfigure{unflip11}{Unflippable $(1,1)$ configuration.} {2in}

424: 

425: \SCfigure{quartz}{Vertex labels of $(0,1)$ configuration for proof of $(1,0)$ unflippability.}{2in}

426: 

427: However, not all flips are automatically flippable:

428: 

429: \begin{theorem}

430: \label{stickshift}

431: The $(1,1)$--$(1,1)$, $(2,1)$--$(1,2)$ and $(1,2)$--$(2,1)$ flips are not automatically flippable for geometric or self-intersecting meshes.  The $(1,0)$--$(0,1)$ flip is not automatically flippable for geometric meshes.

432: \end{theorem}

433: 

434:  \begin{proof}

435: One

436: can form a set of vertices in the pattern of a $(2,1)$ mesh

437: by placing four axis-aligned squares in four parallel planes: two

438: large ones above each other in the two inner planes, and two small ones

439: inside the projections of the large squares in the remaining planes

440: (Figure~\ref{unflippable}).  With this placement, the three-hex $(2,1)$

441: mesh has all faces flat.  However, the small squares can be

442: translated arbitrarily, causing the diagonal interior faces of the

443: five-hex $(1,2)$ mesh (such as the face formed by the four marked vertices in

444: the figure) to be warped, so that the $(2,1)$ mesh is not flippable.

445: 

446: Similarly, by centering these four squares on the same axis, but then shifting a pair of corresponding vertices in the two central squares along the edges connecting these two vertices to their counterparts in the outer two squares, one can create an unflippable $(1,2)$ mesh.

447: 

448: Next, choose parameters $r>0$ and $0<\theta<\pi/4$, and let $p(i,j)$ denote the point

449: $(r\cos i\theta,r\sin i\theta,j)\in\Real^3$.  Then the eight points

450: $p(\pm 1,0)$, $p(\pm 2,1)$, $p(\pm 3,1)$, $p(\pm 4,0)$ form the vertices

451: of a cuboid in which all facets are trapezoids (shown as the top eight vertices

452: of Figure~\ref{unflip11}).

453: If we make another copy of this set of eight vertices, translated downwards, and connect each bottom face of the cuboid to each bottom face of the copy, we get a $(1,1)$ mesh.

454: However, if the copies of $p(-2,1)$ and $p(-3,1)$ (the black marked points in the figure) are shifted less far than the copies of $p(2,1)$ and $p(3,1)$ (the grey marked points), then these four points will form a skew quadrilateral, preventing this mesh from being flipped.

455: 

456: Finally, consider the $(1,0)$ mesh, and label the $14$ vertices of the corresponding

457: $(0,1)$ mesh as illustrated in

458: Figure~\ref{quartz}.  Suppose also that the segments $p_iq_i$, $az$, and $by$ are all parallel.

459: We have the following equations for the two interior vertices $b$ and $y$:

460: \begin{eqnarray*}

461: b &= p_6p_1p_2 \cap p_2p_3p_4 \cap p_4p_5p_6 \\

462: y &= q_6q_1q_2 \cap q_2q_3q_4 \cap q_4q_5q_6

463: \end{eqnarray*}

464: 

465: Here, each triple denotes a plane through the three points.

466: Now consider shrinking the mesh by simultaneously contracting the parallel edges

467: $p_iq_i$ and  $az$.  As long as these edges have non-zero

468: length, the $(0,1)$ mesh is still valid.  But if we make them short

469: enough, $by$ is inverted, so the top and bottom cubes in the $(1,0)$ mesh

470: intersect. Thus, the $(0,1)$--$(1,0)$

471: flip is not automatically flippable.

472: \end{proof}

473: 

474: \begin{open}

475: Is the $(1,0)$ flip automatically flippable for self-intersecting meshes?  Is the $(0,1)$ flip automatically flippable?

476: \end{open}

477: 

478: Our results on automatic flippability are related to scene analysis

479: questions studied by Whiteley and Sugihara (see~\cite{Whi-DCG-89}). In

480: its most basic form, scene analysis asks whether a given planar drawing

481: of vertices connected by faces is the projection  of flat faces in

482: $\Real^3$. Here we are interested in whether a given set of hexes is the

483: projection of a cuboid in $\Real^4$, however our three-dimensional

484: scenes are not generic (due to the flatness constraints) so Whiteley's

485: formula for the number of degrees of freedom does not seem to apply. It

486: would be interesting to apply the  matroid techniques used in scene

487: analysis and rigidity theory to hex meshing.

488: 

489: \section{Bicuboid}\label{bicuboid-sec}

490: 

491: \SCfigure{bicuboid-fig}{The bicuboid.}{2.5in}

492: 

493: \OCfigure{3x3}{3-by-3 subdivision of boundary can be

494: simulated by covering quads with four cuboids.}{5in}

495: 

496: \OCfigure{ngwbc}{Non-generic warped bicuboid (left)

497: with a self-intersecting mesh (right).}{4.5in}

498: 

499: In this section, we discuss a simple but important solid, called

500: a {\it bicuboid\/}. 

501: A bicuboid has the same topology as two hex elements sharing a face, as 

502: shown in Figure~\ref{bicuboid-fig}.  

503: We call a bicuboid {\it warped\/} if its central four vertices,

504: shown dotted in Figure~\ref{bicuboid-fig}, are not coplanar.

505: The bicuboid is important because, as we now show, the question of

506: whether topologically meshable domains have polyhedral hex meshes can be

507: reduced to this case, improving a reduction to a more complicated but

508: finite set of cases from our previous work~\cite{Epp-CGTA-99}.

509: 

510: \begin{theorem}\label{thm:bicuboid}

511: If every bicuboid has a geometric hex mesh, then any quad surface

512: mesh, with an even number of quad elements

513: and topologically equivalent to a sphere,

514: can be extended to a geometric hex volume mesh.

515: \end{theorem}

516: 

517: \begin{proof}

518: We start by finding a topological hex volume mesh

519: as in~\cite{Epp-CGTA-99,Mit-STACS-96,Thu-93}.

520: In such a mesh, each element is a topological cuboid,

521: meaning a cuboid with curved edges and faces.

522: We then subdivide all topological cuboids into 

523: $m \times m \times m$ grids of small topological cuboids

524: that are ``geometrically close'' to true cuboids. 

525: For example, if we require each face

526: of a small topological cuboid to fit within a thin slab

527: defined by parallel planes, we could pick a sufficiently large $m$

528: in order to make this possible.

529: 

530: Each small topological cuboid is then subdivided into seven pieces as

531: in a $(3,0)$--$(0,3)$ flip: a true 

532: cuboid in the center, and six ``almost-cuboidal'' warped hexes, each

533: with five planar faces and one nonplanar face.

534: The almost-cuboidal hexes of the subdivided mesh

535: match up in pairs to form bicuboids.  Meshing each bicuboid and

536: combining the results gives a mesh for the overall domain.

537: 

538: The only remaining problem is that the $m\times m\times m$ subdivision

539: also subdivides the boundary of the domain, which is not allowed.

540: To get around this problem, we let $m$ be a power of three,

541: and simulate the $3\times 3$ subdivision of a boundary quadrilateral by

542: placing four flat cuboids along it, as shown in Figure~\ref{3x3}, so that

543: the cuboids from adjacent boundary quadrilaterals meet face-to-face. The

544: number of degrees of freedom for the new cuboid vertices exceeds the

545: number of coplanarity constraints, allowing this construction to be

546: performed geometrically.

547: \end{proof}

548: 

549: Theorem~\ref{thm:bicuboid} shows that a bicuboid is

550: the crucial test case for the question of whether every

551: geometric quad surface mesh can be extended to a geometric hex mesh.

552: Theorem~\ref{automatic}, however, leads us to suspect that 

553: the answer is no, since it shows that a warped bicuboid cannot be

554: meshed by any generic mesh that can be reached from the two-hex mesh by

555: 

556: automatically flippable flips---reversing those flips would give a

557: geometric two-hex mesh, which does not exist.

558: 

559: \begin{open}

560: Does every warped bicuboid have a geometric hex mesh?

561: Does every warped cuboid have a generic

562: self-intersecting hex mesh?

563: \end{open}

564: 

565: The question is not open for non-generic self-intersecting meshes: a

566: warped cuboid formed by placing the four vertices of a warped

567: quadrilateral around the equator of a unit cube has a non-generic

568: self-intersecting six-hex mesh in which the inner quadrilateral's

569: vertices also lie on the surface of the cube (Figure~\ref{ngwbc}).

570: Less degenerately, if one has a warped bicuboid such that the convex hull of its top and bottom quadrilaterals forms a cuboid, then the toroidal difference between the bicuboid and this central cuboid can be partitioned into four triangular prisms; if we then slice the central cuboid in two by a plane, this plane divides a face of each prism in two, forming degenerate cuboids each with a single $180^\circ$ dihedral.

571: 

572: \section{Parity}

573: \label{parsec}

574: 

575: \OCfigure{quad2to3}{Parity-changing transformation of a quadrilateral

576: mesh.}

577: {3.25in}

578: 

579: \OCfigure{hexdual}{Cuboid divided into octants by three dual surfaces

580: (left), and the set of dual surfaces (right) for a more complicated

581: four-hex mesh (center).}{5in}

582: 

583: Each of the flip transformations preserves

584: the parity (odd or even) of the number of hex elements in the mesh.

585: In two dimensions, any meshable domain has meshes of both parities, as

586: can easily be seen by performing a two-to-three quad replacement (not a

587: flip), as shown in Figure~\ref{quad2to3}, so not all quad

588: meshes are obtainable by flipping from a given starting mesh.

589: A similar result holds for hex meshes, but the construction is more

590: complicated: 

591: 

592: \begin{theorem}\label{parity}

593: If a given quad surface mesh can be extended to a topological hex mesh,

594: it has topological hex meshes with both even and odd numbers of

595: hexes. 

596: \end{theorem}

597: 

598: \begin{proof}

599: Suppose we have some hex mesh, and we wish to find an alternative mesh

600: with the opposite parity.  Following Mitchell~\cite{Mit-STACS-96} and

601: Thurston~\cite{Thu-93}, we form three quadrilateral surface patches

602: within each hex, medial between each of the three pairs of opposite

603: faces; these patches join up to form an arrangement of surfaces, with

604: one triple intersection per cuboid (Figure~\ref{hexdual}).  The

605: boundaries of these surfaces form a graph on the domain boundary that is

606: the dual of the original quad surface mesh.

607: 

608: Conversely, if we have an arrangement of surfaces that meets the domain

609: boundary in the dual of the quad mesh, where at most three surfaces meet

610: in any point and we have no pinch points or other topological anomalies,

611: we can form a graph with one vertex per triple intersection and with

612: edges connecting the vertices along the curves where two surfaces meet. 

613: As long as this graph is connected and has no self-loops or multiple

614: adjacencies, we can form a valid topological hex mesh whose cuboids and

615: quadrilaterals correspond to the vertices and edges of this

616: graph.  If the graph does have self-loops or multiple adjacencies,

617: Mitchell~\cite{Mit-STACS-96} shows that they can be removed by the

618: ``pillowing'' ($(3,0)$--$(0,3)$ flip) operation.

619: 

620: In order to change the parity of the number of hexes in the mesh, then,

621: we can introduce a new surface in such a way that the number of triple

622: intersections increases by an odd number.  A suitable candidate is Boy's

623: surface~\cite{Boy-MA-03,Fra-87}, a projective plane embedded in

624: $\Real^3$ in such a way that there is one triple intersection point,

625: connected to itself by double-intersection curves in a pattern of three

626: self-loops.  We place a copy of Boy's surface within a

627: simply-connected region of the dual arrangement, in such a way that the

628: augmented arrangement remains connected and continues to avoid multiple

629: adjacencies.  Then each of the three loops of Boy's surface will cross

630: the other arrangement surfaces an even number of times, and the

631: double intersection curves of those other surfaces will cross the new

632: Boy's surface an even number of times, so that the only change in parity

633: will be from the single triple intersection of Boy's surface with

634: itself.  The resulting graph may have some self-loops or multiple

635: adjacencies, but we can remove these by pillowing without further

636: changes of parity.  The hex mesh formed by this graph will therefore

637: have the opposite parity from the mesh we started with.

638: \end{proof}

639: 

640: \SCfigure{Boy}{Boy's surface.}{2.5in}

641: 

642: \begin{open}

643: Are there domains with geometric hex meshes of both parities?

644: \end{open}

645: 

646: For quadrilateral meshes, one can allow parity-changing operations by

647: adding Figure~\ref{quad2to3} to the repertoire of flips.

648: 

649: \begin{open}

650: Is there a simpler parity-changing operation for hexahedral meshes?

651: What is the minimum possible number of hexahedra involved in such an operation?

652: \end{open}

653: 

654: For non-simply-connected domains, parity changing may be easier: for instance, a torus with a hexagonal cross section and an odd number of segments may be filled with either two or three hexes per segment, as in the planar parity-changing operation.  For instance, with three segments, this leads to an even mesh with six hexes and an odd mesh with nine.  However such a torus cannot be part of a mesh of a simply-connected domain because the complement of the torus would have a contractible odd cycle.

655: 

656: \section{Bubble-wrapping and Shelling}

657: 

658: \SCfigure{quaddual}{Curve arrangement dual to a quadrilateral

659: mesh.}{3.5in}

660: 

661: \OCfigure{dualjqf}{Actions of quadrilateral flips on dual curves.}{6.5in}

662: 

663: In order to prove our results in the next section on flip graph

664: connectivity, we need a technical strengthening of the Mitchell-Thurston

665: result on topological mesh existence: every ball-shaped domain with an

666: even number of quadrilateral faces has a {\em shellable} hexahedral mesh,

667: where a {\em shelling} of a complex~\cite{Zie-95} is an ordering of its

668: cells such that the union of the cells in any suffix of the ordering

669: forms a simply-connected set, and such that each cell meets the union of

670: the set of cells coming after it in the ordering in a simply-connected

671: subset.  In this section we prove a simpler version of this

672: strengthening: define a {\em pseudo-shelling} to be an ordering of the

673: cells such that the union of the cells in any suffix of the ordering is

674: simply connected, without the requirement on how each cell meets this

675: union.  That is, we want to remove elements one by one from our mesh in

676: such a way that the remaining elements always form a ball-shaped

677: domain.  Then we prove here that every simply-connected domain with an

678: even number of quadrilateral faces has a pseudo-shelling.

679: 

680: To do this, we take a dual view similar to that in

681: Theorem~\ref{parity} and in~\cite{Mue-EC-99}. For any topological quadrilateral

682: mesh on a topological-sphere surface, define a curve arrangement by

683: connecting each opposite pair of edge midpoints in each quadrilateral by

684: a segment of a curve (Figure~\ref{quaddual}).  This is a {\em simple}

685: arrangement, meaning that at most two curves intersect at any point, and

686: whenever two curves intersect they cross each other at a single point.

687: Because the surface has no boundary, the curves have no endpoints; each

688: curve crossing corresponds to a mesh element.  Each element removal in a

689: pseudo-shelling corresponds to a flip in the boundary quadrilateral

690: mesh, and we consider these flips in terms of their effect on the dual

691: curve arrangement (Figure~\ref{dualjqf}).  By finding a sequence of flips that can be used to

692: transform any such arrangement into the arrangement dual to a single

693: cube's boundary, we will show that a pseudo-shelling exists.

694: 

695: \SCfigure{bubblewrap}{Bubble-wrapping a curve arrangement}{4in}

696: 

697: \begin{lemma}

698: A simple curve arrangement is dual to a quadrilateral mesh if and only if

699: the graph formed by its crossing vertices and edges is 3-connected.

700: \end{lemma}

701: 

702: \begin{proof}

703: Any simple curve arrangement can be dualized by forming a quadrilateral

704: for each crossing, however the resulting collection of cells must meet

705: face-to-face and vertex-to-vertex to be a cell complex.

706:  It is not hard to see that, if the graph has

707: one or two vertices that separate it into nontrivial components, then

708: it must have a degeneracy that prevents it from being a cell complex;

709: for instance, a single articulation point corresponds to a

710: quadrilateral that is adjacent to itself along a pair of edges.

711: Conversely, if an arrangement forms a 3-connected planar graph, then by

712: Steinitz' theorem~\cite{Zie-95} it can be represented as the vertices and edges of a

713: convex polyhedron, and the polar polyhedron's facets form the desired

714: dual mesh.

715: \end{proof}

716: 

717: Since not all arrangements are 3-connected, it is convenient to work

718: with a modification of the mesh that is guaranteed to be 3-connected.

719: For a given arrangement $A$, we define the {\em bubble-wrapped

720: arrangement} $B(A)$ by adding curves to $A$ as follows: first, form a

721: small circular curve around each vertex of $A$.  Second, repeat this

722: step around each new crossing point formed by the new circles and the

723: edges of $A$.  Each edge has two circles added in this second step, and

724: these circles should be drawn so that they cross each other, and so that

725: the edge passes through their intersection.  Figure~\ref{bubblewrap}

726: depicts these bubble-wrapping steps.

727: 

728: \SCfigure{arrops}{Three operations on arrangement: pulling apart or

729: pushing together (left); inverting a triangle (middle), and switching

730: (right).}{3.5in}

731: 

732: \begin{lemma}

733: If $A$ is a connected arrangement, $B(A)$ is 3-connected.

734: \end{lemma}

735: 

736: \begin{lemma}\label{wrappable}

737: If $A$ is the dual to a quadrilateral mesh, $B(A)$ can be formed by

738: flips from $A$.

739: \end{lemma}

740: 

741: \begin{proof}

742: We create all the circles used in the bubble-wrapping process by $(2,0)$--$(0,2)$

743: flips, and cross the two circles on each edge by a $(1,0)$--$(0,1)$ flip.

744: It is not hard to see that each step in this process preserves

745: 3-connectivity of the arrangement.

746: \end{proof}

747: 

748: We now define three operations that modify the topology of a

749: curve arrangement in simpler ways than the modifications performed by the flip duals; in some ways these operations are similar to the well-known Reidemeister moves for knots in $\Real^3$, or the Whitney moves for curves in $\Real^2$ \cite{MehYap-SJC-91,Whi-CM-37}.  If the arrangement has a pair of crossings that are

750: connected by two edges, without other curves between them, we can {\em

751: pull apart} the crossings and form an arrangement with two fewer

752: vertices; conversely we can {\em push together} two nearby curves to add

753: two vertices to an arrangement (Figure~\ref{arrops}, left).

754: If the arrangement has a face formed by a triangle of three crossings

755: and three curves, we can {\em invert} the triangle by passing one curve

756: across the crossing of the other two (Figure~\ref{arrops}, middle).

757: Finally, if the arrangement contains two nearby connected and

758: non-crossing segments of curves, we can {\em switch} these segments by

759: removing them from the arrangement and reconnecting the curves by two

760: other non-crossing segments (Figure~\ref{arrops}, right).

761: 

762: \begin{lemma}\label{wrapped-ops}

763: Suppose two connected arrangements $A$ and $A'$ are related to each

764: other by one of the three operations described above.  Then $B(A)$ can

765: be transformed into $B(A')$ by a sequence of flips.

766: \end{lemma}

767: 

768: \SCfigure{invbub}{Steps in inversion of a triangle in a bubble-wrapped

769: arrangement.}{4.5in}

770: 

771: \begin{proof}

772: We begin by describing the simplest case, inversion of a triangle.

773: If $A$ were not bubble-wrapped, we could perform this operation by a

774: single $(0,0)$--$(0,0)$ flip, however the difficulty is in rearranging the

775: bubble-wrap curves to allow this flip without violating 3-connectivity.

776: Suppose

777: $A$ contains a triangle that can be inverted; its bubble-wrapped version

778: $B(A)$ is depicted in Figure~\ref{invbub} (left). We perform a $(2,0)$--$(0,2)$ flip

779: at one of the crossings of $B(A)$ in this triangle, adding a new curve

780: to the arrangement, and perform a sequence of $(1,0)$--$(0,1)$ and $(0,0)$--$(0,0)$ flips to

781: extend the region interior to this new curve until it covers the whole

782: triangle without extending outside the bubble-wrapping of the triangle

783: (Figure~\ref{invbub}, middle).  Next, we remove the nine curves forming

784: the bubble-wrapping around the triangle, by using $(0,0)$--$(0,0)$ and $(0,1)$--$(1,0)$ flips

785: to reduce the size of the region each one contains until it can be

786: removed by a $(0,2)$--$(2,0)$ flip (Figure~\ref{invbub}, right).  At this point the

787: triangle is exposed in the interior of the newly added curve and can be

788: inverted by a $(0,0)$--$(0,0)$ flip, after which we reverse the above steps to

789: restore the bubble-wrapping around the triangle.  It is not difficult to

790: verify that each step in this process can be performed while preserving

791: 3-connectivity of the overall arrangement.

792: 

793: Similarly, pulling apart and pushing together can be performed by

794: modifying the bubble-wrapping to remove the eight curves protecting the crossings and edges in the pulled-apart region, and replace them by a single curve surrounding an adjacent face of the arrangement,

795: in a way that allows the change to $A$ to

796: be made by a single $(1,0)$--$(0,1)$ or $(0,1)$--$(1,0)$ flip.

797: Switching can be performed by adding curves around the face of the arrangement between the switched edges, so that one curve crosses both switched edges and another curve passes between them, in a way that allows the change to $A$

798: to be made by a single $(1,1)$--$(1,1)$ flip.  The details are more complicated

799: than inversion, and we omit them.

800: \end{proof}

801: 

802: \SCfigure{switchcross}{Switching near a crossing.}{2.5in}

803: 

804: \SCfigure{switchloop}{Switching near a loop.}{3.75in}

805: 

806: \begin{lemma}

807: Any simply-connected domain, with a boundary consisting of an even

808: number of quadrilaterals, has a mesh with a pseudo-shelling.

809: \end{lemma}

810: 

811: \begin{proof}

812: As stated at the start of the section, we describe the pseudo-shelling

813: in terms of the effect each removal of a hexahedral element

814: causes to the curve arrangement $A$ dual to the domain boundary. By

815: Lemma~\ref{wrappable}, we can form an initial sequence of flips to

816: bubble-wrap $A$.  Next, we apply Lemma~\ref{wrapped-ops} to perform a

817: sequence of switching operations near each crossing of $A$.  Each such

818: operation creates a loop connecting the crossing to itself (Figure~\ref{switchcross}), and reduces

819: the number of crossings without such loops; it is always possible to

820: choose a way to perform the switching operation that preserves the

821: connectivity of the overall arrangement.

822: 

823: \SCfigure{loopcancel}{Removing two adjacent loops by a push,

824: invert, and two pull operations (the Whitney trick).}{5in}

825: 

826: After this step, the transformed arrangement must consist of a single

827: simple closed curve, decorated with a sequence of loops.  Since $A$

828: initially had an even number of crossings, and each flip preserves

829: parity, there are an even number of loops along the curve.  Note that

830: a switching operation near the crossing of a loop can result in a new loop

831: on the other side of the simple closed curve (Figure~\ref{switchloop}).  We perform additional

832: switching operations near the loops' crossings if necessary so that the

833: loops can be grouped into adjacent pairs, with one member of each pair

834: interior to the simple closed curve and one member exterior.  Each pair

835: can be removed by a sequence of a push, triangle inversion, and two pull

836: operations (Figure~\ref{loopcancel}); this sequence of operations is known to topologists as the {\em Whitney trick}.  After repeated Whitney tricks, the arrangement is transformed into a simple closed curve protected by bubbles; we remove bubbles until only two are left, forming

837: the dual arrangement of a cube, the desired final state of a

838: pseudo-shelling.

839: \end{proof}

840: 

841: The existence of a mesh with a shelling, rather than just a

842: pseudo-shelling, can be deduced via the methods used to prove

843: Theorem~\ref{thm:qfc} below.

844: 

845: \section{Connectivity}

846: 

847: The results of Section~\ref{parsec} naturally raise the question

848: whether parity is the only obstacle to reachability by flips, or whether

849: there might be some more subtle property of a mesh that prevents it being

850: formed by flipping from some other mesh.  Phrased another way, can every

851: mesh transformation that replaces some bounded submesh by another be

852: simulated by a sequence of flips?  Figure~\ref{hardtoflip} depicts

853: a difficult example, closely related to a hexahedral meshing problem

854: posed by Schneiders~\cite{Sch-Open}.  We leave finding a flip sequence

855: for this example as a puzzle for the reader.

856: 

857: For any domain with boundary mesh, and a type of mesh to use for

858: that domain, define the {\em flip graph}~\cite{San-JAMS-00} to be a graph with (infinitely

859: many) vertices corresponding to possible meshes of the

860: domain, and an edge connecting two vertices whenever the corresponding

861: two meshes can be transformed into each other by a single flip.  In this

862: framework, the question above can be phrased as asking for a description

863: of the connected components of the flip graph.

864: 

865: \begin{theorem}\label{thm:qfc}

866: The flip graph for topological quad meshes of any simply-connected

867: domain has exactly two connected components.

868: \end{theorem}

869: 

870: \begin{proof}

871: Due to parity, there must be at least two components. It remains to show

872: that any two meshes $M_1$ and $M_2$ of the same parity can be flipped

873: into each other. Consider a three-dimensional ball, the upper hemisphere

874: of which is partitioned into mesh $M_1$ and the lower hemisphere of

875: which is partitioned into mesh $M_2$, the two meshes meeting along the

876: equator of the ball at corresponding boundary edges.  Because $M_1$ and

877: $M_2$ have the same parity, the ball has an even number of quadrilateral

878: faces and so has a topological hex mesh $H$.

879: 

880: The possible existence of

881: pairs of quadrilaterals meeting in more than one edge (along the ball's

882: equator) does not lead to difficulties with degenerate meshes,

883: as we can remove these degeneracies by performing $(2,0)$--$(0,2)$ and $(0,2)$--$(2,0)$ flips in

884: $M_1$ and $M_2$ respectively.

885: 

886: We now wish to remove the cuboids from $H$ one by one,

887: reducing the initial ball to successively smaller shapes.  At each step

888: in the removal process, we will maintain a mesh $M$ describing 

889: describing the upper boundary of the remaining shape; initially

890: $M=M_1$.  As long as each removed cuboid touches the upper boundary of $M$ in a simply

891: connected set of faces, the change to $M$ caused by its removal is a

892: flip.  If we can remove all the cuboids in this way, we will have

893: flipped $M_1$ into $M_2$.

894: 

895: The requirement of simply-connected incidence with $M$ for each removed

896: cuboid is very similar to the definition of a {\em shelling} of a

897: complex, as described in the previous section; it differs from a shelling only in the requirement that cuboids be removed from the upper boundary.

898: As shown in the previous section, we can at least choose $H$ to

899: be pseudo-shellable.  We then partition the pseudo-shelling sequence into

900: {\em layers} of independent sets of hexahedral elements, such that no two elements in the

901: same layer share a facet.  Let $B_i$ denote the ball formed by the union

902: of layer $i$ and all successive layers. The boundary of each $B_i$ is a

903: quadrilateral mesh; we thicken each quadrilateral of this mesh into a

904: hexahedron, forming another layer of hexahedra separating layer $i$ from

905: layer $i-1$.  In terms of the dual surface arrangement this corresponds

906: to adding a topological-sphere surface surrounding $B_i$.

907: After performing this thickening process, we have a mesh with

908: alternating nested layers of cells: half of the layers consist of cells

909: connecting two nested and combinatorially equivalent quad meshes, while

910: the other half of the layers consist of cells not sharing any facets

911: with each other.

912: 

913: We order the cells of this mesh as follows: start by choosing some cell

914: adjacent to

915: $M_1$, then some cell in the next layer adjacent to the starting cell,

916: and continue ``drilling down'' by removing at most one cell from each

917: layer until reaching the center of the set of concentric spheres.  Then,

918: remove the layers from the inside out, removing all cells from a given

919: layer before starting the next outer layer.  The existence of an

920: appropriate ordering within each layer follows from the existence of

921: shellings for all planar complexes.

922: \end{proof}

923: 

924: \SCfigure{hardtoflip}{Transformation of one four-quad

925: mesh into another. Can it be simulated by a sequence of flips?}{3in}

926: 

927: \OCfigure{simflips}{Simulation of flips by parity-change operation:

928: half of $(1,1)$--$(1,1)$ flip (top), $(0,0)$--$(0,0)$ flip (bottom, first and third positions), and

929: $(1,0)$--$(0,1)$ flip (bottom,

930: second and fourth positions).}{5.25in}

931: 

932: \OCfigure{unflipq}{Transformation of a non-simply-connected quad mesh

933: that cannot be achieved by flips. The three edges indicated by hash

934: marks form a contractible triangle in the region

935: bounded above and below by the left and right meshes.}{5in}

936: 

937: As a consequence, the set of flips together with the parity-changing

938: operation depicted in Figure~\ref{quad2to3} form a complete set of local

939: mesh transformations, in that they are able to simulate any other

940: transformation operation.  In fact, since the parity-changing operation

941: can simulate many of the possible flips (Figure~\ref{simflips}),

942: it and the $(2,0)$--$(0,2)$ flip together form a complete set.

943: 

944: Combining the proof method with our previous linear-complexity

945: hex meshing techniques~\cite{Epp-CGTA-99} shows that any two equal-parity

946: quad meshes $M_1$ and $M_2$ of a simply-connected region can be flipped

947: one to the other (as topological meshes) in a number of steps linear in

948: the total number of cells in the two meshes.

949: 

950: Conversely to the techniques of this proof, any sequence of flips from

951: one quad mesh to another can be seen as forming a hex mesh of a three

952: dimensional region bounded above and below by the two meshes. The

953: assumption of simple connectivity is necessary for the result, as can be

954: seen from Figure~\ref{unflipq}: the region bounded by the two

955: meshes has odd cycles on its boundary (triangles formed by

956: one diagonal edge of the four-quad mesh and two edges of the eight-quad

957: mesh) that can be contracted to points in the interior of the region;

958: Mitchell~\cite{Mit-STACS-96} has shown that such contractible odd cycles

959: make hex meshing an impossibility. Therefore there is no way to

960: flip one mesh to the other, despite their having equal

961: parity.

962: 

963: \begin{open}

964: Can we characterize the flip graphs for geometric quad meshes of non-simply-connected regions?  Is the number of components of the flip graph determined by the number of topological holes in the region?

965: \end{open}

966: 

967: For the analogous problem of geometric triangle mesh flip graph

968: connectivity, a single connected component is known to exist due to the

969: ability to flip to the Delaunay triangulation, and the ability to

970: add or remove vertices by changing the weights of a regular

971: triangulation.  However, higher dimensional flip graphs can be disconnected~\cite{San-JAMS-00}.  There has also been some work on finding efficient

972: sequences of

973: flips~\cite{HanOttSch-JUCS-96,HurNoyUrr-DCG-99,SleTarThu-JAMS-88}

974: although finding the shortest such sequence even for triangulations of

975: convex polygons remains a major open problem.

976: 

977: \begin{open}

978: Can the flip graph for topological or geometric hex meshes of a simply

979: connected region have more than two components?

980: \end{open}

981: 

982: Answering this question seems to require a four dimensional

983: generalization of Mitchell and Thurston's result.  Even the tetrahedral

984: mesh version of this problem seems difficult: topological tetrahedron

985: mesh flip graphs are connected~\cite{Lic-KF-99},

986: and flipping can be used to form Delaunay

987: triangulations~\cite{EdeSha-Algo-96,Joe-CAGD-91}, but this does not seem

988: to imply geometric tetrahedral flip graph connectivity since the

989: flipping-based Delaunay construction algorithms

990: do not allow the initial triangulation or sequence of flips to be chosen

991: arbitrarily.

992: 

993: \section{Conclusions}\label{disc-sec}

994: 

995: In this paper, we have introduced (or at least popularized)

996: flipping moves for quad and hex meshes.  We have also considered the

997: conditions under which flips are realizable, tested the connectivity of

998: the flip graph, and used flipping to show some evidence that not all quad

999: surface meshes can be extended to  hex volume meshes. 

1000: 

1001: \section*{Acknowledgements}

1002: 

1003: A preliminary version of this paper appeared at the 10th International

1004: Meshing Roundtable.  Work of Eppstein was supported in part by NSF grants

1005: CCR-9258355 and CCR-9912338

1006: and by matching funds from Xerox Corp., and performed in

1007: part while at Xerox PARC.

1008: Work of Erickson was supported in part by a Sloan Fellowship and by NSF grants CCR-0093348 and DMR-0121695.

1009: We thank Scott Mitchell, Scott Canann, and the

1010: Roundtable referees for many helpful comments.

1011: 

1012: \raggedright

1013: \bibliographystyle{abuser}

1014: \bibliography{hexmesh}

1015: 

1016: \end{document}

1017: