Polyhedral Patterns

Caigui Jiang* Chengcheng Tang* Amir Vaxman Peter Wonka† Helmut Pottmann†
King Abdullah University of Science and Technology (KAUST) TU Wien / Utrecht University KAUST TU Wien / KAUST

Figure 1: Polyhedral patterns on a knot. Top: three polyhedral patterns tiling a knot and optimized by our framework. All examples
are combinatorially equivalent to a semi-regular pattern (3, 4, 6, 4). Bottom: each of the three solutions is induced by a choice of strip
decomposition and corresponding affine symmetries. For each model, we show the strip decomposition (left) with the pattern in the plane
colored by yellow and blue strips. We show the deformed pattern upon mapping to a cylinder, suggesting the feasible symmetries (right). The
different colors encode different choices of symmetries. For instance, blue faces are symmetric with respect to the barycenter.

1 Introduction
Architects and engineers are constantly pushing design boundaries
Abstract by exploring new building shapes and modeling their appearances.
Advances in architectural geometry have made it possible for many
We study the design and optimization of polyhedral patterns, which buildings to be shaped as freeform surfaces. To conform to construc-
are patterns of planar polygonal faces on freeform surfaces. Working tion constraints, such designs are often rationalized with meshes
with polyhedral patterns is desirable in architectural geometry and that have planar faces. These faces are then realized with common
industrial design. However, the classical tiling patterns on the plane materials, such as wood (see Fig. 3) or glass.
must take on various shapes in order to faithfully and feasibly ap- Symmetric tessellation patterns have often been used in art, architec-
proximate curved surfaces. We define and analyze the deformations ture, and product design for their aesthetic merits. However, the use
these tiles must undertake to account for curvature, and discover of these patterns was restricted to planar surfaces, such as windows,
the symmetries that remain invariant under such deformations. We walls, or floors. Notable examples are Arabesques, stained-glass pat-
propose a novel method to regularize polyhedral patterns while main- terns, and mosaics [Abas et al. 1995; Lu and Steinhardt 2007]. Here,
taining these symmetries into a plethora of aesthetic and feasible we seek to enrich architectural design by meshing freeform surfaces
patterns. with tessellation patterns. Examples of those are in Figures 1 and 2.
There are several key challenges in designing polyhedral patterns
CR Categories: I.3.5 [Computer Graphics]: Computational Ge- on curved freeform surfaces. Simply placing a given pattern on
ometry and Object Modeling—Curve, surface, solid, and object such a surface (e.g., by a parametrization) and optimizing for tile
representations planarity without any regularization of tile shapes is bound to fail;
planarity alone is severly underconstrained, and the process is apt
Keywords: Architectural geometry, discrete differential geometry, to degenerate to solutions with zero-length edges or foldovers (see
polyhedral meshes, wallpaper tiling groups Figure 4). Moreover, an arbitrary choice of tile-shape regularization
with commonly used measures, such as face or angle distortion,
Laplacian smoothness, or edge-length preservation, might clash with
* Joint first authors. the planarity constraint, resulting in over-constrained optimization.
† Joint last authors. As a consequence, either the mesh is not planarized, or the desired
regularity is not satisfied. (Figures 27 and 28 show examples.)
Our solution to the problem is to study explicit constructions of poly-
hedral patterns that approximate surfaces with varying Gaussian cur-
vature. We observe curvature-invariant regularities, namely different
types of symmetries. We introduce a theoretical study of polyhedral
patterns that explains our choice of regularizers, and which leads
to an objective function that is neither over- nor under-constrained.
 Figure 4: Under- and over-constrained optimization illustrated on
a quad mesh. The initial mesh is aligned with the parameter lines
on a bilinear surface. The task is to planarize it. Left: Using mesh
polyline smoothness as a regularizer results in an over-constrained
problem, and acceptable planarity is not achieved (red color). Mid-
dle: Dropping the regularizer leads to an under-constrained problem,
Figure 2: Cladding an interior space with a polyhedral pattern where the faces are perfectly planar, yet their appearance is chaotic
using wooden panels. The pattern transforms smoothly from positive and unaesthetic. Right: Using our regularizer, based on affine sym-
to negative curvature regions. metries (with respect to edge midpoints), yields an aesthetically
pleasing pattern with planar quads.

triangle meshes [Zimmer 2014; Li et al. 2014] or remeshing triangle

meshes by parametrization and deformation [Vaxman and Ben-Chen
2015]. They are problematic in several aspects: the face shapes
have to change considerably and even become concave in negatively
curved areas. In addition, they have to transition smoothly between
regions of negative and positive Gaussian curvature. We present a
systematic way to regularize patterns between different curvature
Figure 3: A planar-hexagonal pavilion constructed with wooden regions, including PH meshes as a special case.
panels [Krieg et al. 2014].
Special polyhedral patterns appear as by-products with circle-
packing meshes [Schiftner et al. 2009] and special hexagonal support
structures [Jiang et al. 2014]. However, these papers do not consider
We show a set of results that demonstrate, for the first time, the polyhedral patterns in a systematic way; aesthetics and regularity
computation of such patterns on surfaces that satisfy both planarity largely come from the structures the patterns have been derived
and regularity constraints. We focus on semi-regular patterns (see from.
Figure 5), which are patterns comprising regular polygons.
Numerical optimization schemes for computing polyhedral surfaces
Our contributions are: include the alternating least-squares approach of [Poranne et al.
• An analysis of surface approximation with polyhedral patterns 2013], the local-global projection method of [Bouaziz et al. 2012],
by describing tile shape deformations, in order to accommodate the augmented Lagrangian algorithm of [Deng et al. 2015], and the
for curvature. Consequently, we show how different strip guided projection method of [Tang et al. 2014]. Our computations
decompositions result in a variety of patterns. are based on the latter approach, but are distinct from all the other
aforementioned works by the novel use of local affine symmetries
• Affine symmetries that are curvature-invariant. We construct a as regularizers, which are able to adapt to freeform geometry, and
family of regularizers encoding such symmetries: e.g., sym- by the study of the curvature-dependent appearance of polyhedral
metries with respect to axes passing through vertices, edge patterns.
midpoints, or face barycenters, and reflective symmetries with
respect to planes. Triangle meshes are trivial polyhedral meshes. Regularizing them
for face and edge repetitivity is the aim of [Singh and Schaefer 2010]
• A variety of polyhedral patterns that have not been demon- and [Huard et al. 2014]. These meshes also appear in triangle-based
strated before. point folding structures [Zimmer et al. 2012]. Our symmetry-based
regularizers can also be used for triangle mesh optimization.
2 Previous Work
3 Geometry of Polyhedral Patterns
Approximation with polyhedral meshes can be achieved with varia-
tional shape approximation [Cohen-Steiner et al. 2004]. However, Creating polyhedral patterns first and foremost poses a theoretical
the resulting unstructured mesh is built to satisfy required approxi- challenge, since we do not possess the knowledge of how such
mation accuracy and does not follow a prescribed pattern. patterns behave in different curvature regions. We have an under-
Most previous works that focused on polyhedral mesh creation tar- standing of planar-quad meshes as given by [Liu et al. 2006]. If the
geted planar quad (PQ) meshes. PQ meshes play a central role in dis- network of polylines that is characteristic of quad meshes follows
crete differential geometry [Sauer 1970; Bobenko and Suris 2008], conjugate directions, it is possible to achieve a mesh with smooth
and have attracted considerable interest in recent years, cf. [Liu et al. polylines. Unfortunately, this is not possible for any orientation
2006; Liu et al. 2011; Zadravec et al. 2010], as their design is a core of quads (see auxiliary material for a theoretical proof), and there
problem in architectural geometry. has been no suggestion for what could be done in this case. An
analysis is provided for feasible planar hexagonal tile shapes by [Li
Planar hexagonal (PH) meshes have also been studied, but to a lesser et al. 2014]. However, the description is particular for hexagons in
extent. The simplest way to produce them is by taking the dual of principal directions, and the generalization to semi-regular patterns
 (a) (34 , 6) (b) (3, 4, 6, 4) (c) (3, 6, 3, 6) (d) (3, 12, 12) (e) (4, 6, 12) (f) (4, 8, 8)

(g) (34 , 6)∗ (h) (3, 4, 6, 4)∗ (i) (3, 6, 3, 6)∗ (j) (3, 12, 12)∗ (k) (4, 6, 12)∗ (l) (4, 8, 8)∗

Figure 5: Several patterns used in this paper: selected semi-regular patterns (top row; labels correspond to the valences of faces around a
vertex) and their duals (bottom row). In our results, patterns (a) - (e) and (g) - (j) are derived from a hex-mesh, patterns (f) and (l) from a quad
mesh, and pattern (k) from a triangle mesh.

is not obvious. respective principal directions. The paraboloids are characterized

as either elliptic (both curvatures are nonzero and have the same
In the following, we provide an analysis of feasible planar tile shapes sign), hyperbolic (different nonzero signs), or cylindrical (one of
in different curvature regions, for the general case of semi-regular the curvatures is zero). In case that κ1 = κ2 = 0, the osculating
tilings. In Section 6, we utilize the insights gained from this analysis, paraboloid is a plane, and we do not need to deal further with this
to establish a set of symmetries that remain invariant in each curva- trivial case.
ture region. We consequently use these symmetries as regularizers
in our planarization algorithm. We consider discretizations of paraboloids by polyhedral patterns
characterized by two properties: First, they are inscribed, which
We base our geometric constructions on semi-regular patterns, means the vertices of the pattern lie exactly on the paraboloid. Sec-
which are tilings that can be derived in the plane by altering ei- ond, we have normal adherence. Assuming that the supporting
ther of the three regular tilings: triangle, square, or hexagonal grids. plane to the inscribed tile encloses a well-defined small patch of
Semi-regular tilings are characterized by several properties: First, the paraboloid, then there must be a point within the patch whose
the neighborhood of any vertex is perfectly similar to the neighbor- tangent plane is parallel to the supporting plane. Both properties can
hood of any other. Second, such tiles constitute an orthogonal circle be generally relaxed, but it is cogent to study the pattern symmetry
pattern: every tile has a circumcircle, and the dual segments between and regularity emerging from these most restrictive requirements.
neighboring tile (circle) centers are orthogonal to the primal edge
they share. This property is important when we discuss construction
by lifting. We depict a range of semi-regular tilings that we employ Lifting The analysis we give for tiling surfaces relies on lift-
in Figure 5. We denote tilings using vertex configuration shorthand ing planar tiles onto paraboloids (see Figure 7). We then ex-
by numbering the degree of faces around each vertex and using pow- plore the tile shapes and topologies for which the lifting pro-
ers for multiples, e.g., a pure hexagonal pattern is 63 . We use the duces polyhedral patterns (preserving tile planarity). Given the
term “tile” to indicate a single face of the tiling, and either “tiling” paraboloid S2 , the vertices are

lifted (bijectively) by the function
or “pattern” to indicate the entire set. We often refer to the dual (x, y) → x, y, κ1 x2 + κ2 y 2 . The intersection of the supporting
pattern as the pattern that is made by connecting dual face centers plane of the tile and the paraboloid is a conic, related to the Dupin
of adjacent primal tiles. Tilings that are not semi-regular exhibit indicatrix. The projection of the conic down to the plane is again a
several of these properties, and our consequent optimization results planar conic, endowed with required properties that we next detail.
in interesting polyhedral patterns as well. Moreover, the conic on the paraboloid is an affine image of the planar
tile. See Figure 6 for an example.
3.1 Discretization of osculating paraboloids

We construct some explicit embeddings of polyhedral patterns on

surfaces to study necessary deformations in tile shapes. Such defor-
Figure 6: Lifting. Top: a circle lifted to an
mations are the result of fitting planar patterns with given connec-
ellipse in a rotational paraboloid. Bottom: a
tivities onto curved surfaces, while constraining each tile to remain
hyperbola lifted to a hyperbolic paraboloid.
planar. Our purpose is to derive the invariants of the required defor-
mations, focusing on symmetries they fix. An understanding of such
invariant symmetries serves as a guide to predicting the resulting
pattern shapes expected within our optimization process.
Consistent tilings Polyhedral patterns on paraboloids are syn-
We locally approximate the original surface, S, to a second-order onymous with consistent structures that are pivotal to our framework
in a point, p ∈ S, with an osculating paraboloid, S2 . Assuming and that govern the deformations induced on tiles for consistent ap-
the z direction is parametrized to be in the direction of the normal, proximation of paraboloids. Assume that there are two neighboring
the formula defining the paraboloid is 2z = κ1 x2 + κ2 y 2 , where tiles, i, j. Their centers are defined by looking at the centers of the
κ1 , κ2 are the principal curvatures, and the x and y axes are the conics on the paraboloid and projecting them down on the tiling
 pp
ppiiiii ccccjjjjj ppiiiii
pp ccccjjjjj pppiiiii ccccjjjjj pppiiiii ccccjjjjj
ccciiiiii pp cccciiiiii pp cccciiiiii pp cccciiiiii pjjjjjj
pp
pjjjjjj pjjjjjj pjjjjjj

Figure 7: Lifting consistent tilings to paraboloids. Under the in-

duced metric in the plane, the dual edge cij and the primal edge
pij are conjugate. Both vectors are orthogonal in the rotational
paraboloid case (left).
Figure 8: Deforming and lifting. Top: the paraboloid tilings. Bot-
tom: top view of the tiling. Left to right: original (fit to the canonical
plane, producing ci , cj . The intersection points between them are paraboloid), anisotropic elliptic, cylindrical, and hyperbolic.
the projected common vertices, pi , pj . We define the primal vector
pij = pj − pi , and the dual vector cij similarly. Next, we consider
the metric induced by the paraboloid: ha, bi := κ1 xa xb + κ2 ya yb .
The lift of a planar tiling into a paraboloid is polyhedral if and only conic centers of the same strips are denoted as dual rulings, and the
if: sequence of intersection edges bounding two strips are denoted as
primal rulings. Next, fix all the tile (dual) centers and deform all
1. Conjugacy: the primal and dual vectors are conjugate. i.e., (primal) vertices orthogonally to the ruling until they are on a parallel
they are orthogonal with respect to the metric (hpij , cij i = 0). primal ruling. The actual position of the primal ruling is set as the
2. Bisection: The dual edge bisects the primal edge. ruling that is closest to the original primal vertices (see Figure 8
for a depiction of this process). Denote the original position of any
We call a tiling that obeys both properties consistent. For complete- primal vertex as pi , and the deformation vector for a cylindrical
ness, we include a proof in the Appendix. Consistency brings about pattern as uc,i . Then, ŷ · uc,i = 0.
several key consequences:
• If the paraboloid is a rotational-symmetric canonical Tiling elliptic and hyperbolic paraboloids To unify our defor-
paraboloid 2z = x2 + y 2 , then the duals and the primals mation setting and make it fit all types of paraboloids, we rephrase
are in fact also Euclidean orthogonal, as we demand from the our construction for a cylindrical paraboloid in a local and continu-
original tilings. ous manner: we deform the primal edges of an initial tiling so that
the (constant) dual and the primal are conjugate according to the
• If the paraboloid is cylindrical, then either the dual or the induced metric, and the deformation is done according to the choice
primal must be in the direction of the ruling (the direction of of strip decomposition, as in the cylindrical case. Suppose again that
zero curvature). That means that the tiling has to comprise the chosen direction of the dual strip is ŷ. We then need to find the
strips of faces that are parallel to the ruling direction. deformation vector ui of vertex pi so that for each primal edge pij
• If the paraboloid is hyperbolic, the dual and the primal can and dual edge cij we get hpj + uj − pi − ui , cij i = 0, according
be identical (in asymptotic directions). We can potentially to the metric. Furthermore, we constrain ui = −uj for symme-
produce degenerate and concave tiles in this manner. try. Since ui = −uj are orthogonal to ŷ, it is straightforward to
compute the actual deformation.
3.2 Fitting tiles to paraboloids
Discussion Our deformation process is obviously invariant to
Given a planar tiling and a paraboloid of any shape, we next wish scale. More accurately, it only pertains to the ratio of κ1 and κ2
to deform the tiling on the plane, such that the lifting produces a and not their sizes. In addition, it is evident that tiles with more
polyhedral pattern. This should be done while keeping the shapes of than 4 vertices must become non-convex in order to be inscribed in
the tiles as symmetric and as regular as possible and with minimal negatively curved regions; this is expected, since the conic in which
deformation. It is worth noting at this point that we use this construc- the tile is inscribed is a hyperbola.
tion for a theoretical, rather than a direct algorithmic, purpose; it is a
general and arbitrary way to observe which symmetries are invariant The deformation orthogonal to the chosen strip direction is the
within the deformation, and to motivate our use of such symmetries only one possible for tilings such as 63 . However, by constraining
in pattern optimization for general surfaces, locally approximated as the consistency, and by conforming to the aesthetic request that
paraboloids. We initiate this analysis by considering the canonical the repeating faces of the same type must stay congruent, some
rotational paraboloid Sr : 2z = x2 + y 2 . We take any semi-regular tilings may in fact allow more degrees of freedom in deformation
tiling on the plane. Without any deformation, the circular faces of the possibilities. This works only to our advantage.
semi-regular tiling are then projected into ellipses in the paraboloid,
which are planar by definition. We thus obtain a natural polyhedral
Violating consistency The deformations we defined maintain
pattern embedded in Sr for every semi-regular pattern.
conjugacy and bisection for most semi-regular patterns, but not for
some. For example, consider the (4, 6, 12) example in Figure 9: by
Tiling cylindrical paraboloids Without loss of generality, we fixing all the dual vertices, our deformation would violate bisection
assume that our cylinder is defined by Sc : z = κ2 y 2 , κ2 > 0. By between some of the quads and the 12-sided faces. This is caused
the rule of conjugacy, the tiling must comprise strips that are parallel by the special structure of the (4, 6, 12) pattern, for which the dual
to the ruling direction, ŷ, to consistently tile the cylinder. In light of centers of the hexes and the oblique quads cannot conform to straight
this, and opting to deform the tiling as little and as symmetrically rulings should they stay fixed. Our correction is simple: allow the
as possible, we do the following: decompose the tiling into strips centers of the hexes to deform as well, so that they line up with the
of faces that are parallel to the rulings. The lines between dual (fixed) centers of the oblique quads.
Figure 11: Different strip decompositions increase the available types of polyhedral patterns. With the three strip decompositions shown in
Figure 10, we obtain hexagon patterns approximating a cyclide with different appearances.

Deformations and symmetries The practical meaning of choos-

ing strips is to contrast a chosen strip direction with the principal
directions of the paraboloid. Choosing a strip means choosing a
dual axis for every consequent pair of faces, and forcing the primal
vertices to move in directions that are orthogonal to the strip axis. In
other words, we constrain a plane of symmetry that is orthogonal to
the tile, and passes through its center.
However, as we explain above, the principal directions may mandate
(e.g., if they are rulings), that the surface forms lines of primal ver-
tices along such rulings. At any rate, the tile must also be inscribed
to a conic of the same nature. The canonical choice of strips is where
Figure 9: The dual centers of the (4, 6, 12) pattern do not corre- the dual axis is aligned with the rulings (see Figure 10 left). However,
spond to possible rulings if fixed upon deformation. Allowing the other choices may produce interesting patterns due to the mismatch
hex (pink) centers to deform fixes this problem. between the constrained symmetry and the rulings (see Figure 12),
and they may potentially form invalid configurations. Such configu-
rations arise when the strips are along the asymptotic directions of
the surface, i.e., where the dual direction is self conjugate.
3.3 Strip decompositions
General patterns Not all patterns can be decomposed to strips.
Our explicit construction provides a canonical way to approximate For instance, the tri-hex pattern (3, 6, 3, 6) cannot be decomposed.
paraboloids, by relying on a single possible choice for defining strips. Such patterns cannot therefore comply to the normal adherence
In the following, we explore other possible constructed solutions, by and cannot be made consistent by deformation. However, as our
choosing different alignments, corresponding to decomposing the algorithm requires consistency only for the theoretical analysis, we
patterns into different strips. still utilize these patterns in practice, just without any guarantees.
Figures 17 and 23 provide examples.
Decomposable patterns Essentially, strip decomposition is a 3.4 Regularizers Motivated by Symmetries
combinatorial refinement of the original pattern. A feasible strip
decomposition is a collection of disjoint dual strings (trees with Ideally, we would like to achieve the described symmetric and pla-
2-valence nodes; see Figure 10 for examples). Since a strip decom- nar tile shapes for meshes initially tiled with semi-regular convex
position assigns primal vertices to dual vertices, it is actually a strip patterns. However, general meshes are not paraboloids, and they
decomposition of the dual pattern as well. Regular quads, hexagons, have a variety of strip decompositions and varying curvature regions.
(34 , 6), (3, 4, 6, 4), (4, 6, 12), (4, 8, 8) and their dual patterns have Not wanting to be particular for every pattern, we instead opt for
infinitely many strip decompositions. However, patterns such as the most general way to make any type of semi-regular pattern de-
(3, 6, 3, 6), (3, 12, 12) and their duals cannot be decomposed to form properly. Our point is to utilize what remains invariant under
strips by definition. curvature-based tile deformations, rather then what deforms. There-

Figure 10: Different strip decompositions for regular hexagons. Figure 12: Transformation of the regular hexagon pattern from a
The three decompositions from the left correspond to the ones shown rotational paraboloid (left) via a parabolic cylinder (middle) to a
in Figure 11. In addition, the transformation corresponding to the hyperbolic paraboloid (right) with the strip decomposition shown in
decomposition second from the left is shown in Figure 12. Figure 10, second from left.
 (a)
(a)
(a) (b)
(b)
(b) (c)
(c)
(c) (d)
(d)
(d)
Figure 13: Framework overview: a) for an initial triangle, quad, or hex mesh, we can generate a pattern mesh using simple geometric rules.
b) The initial pattern mesh might already be aesthetically pleasing, but the faces are typically not planar. c) A regularizer can be configured by
specifying symmetries that should be preserved in the pattern. In this case, face symmetries are chosen. Corresponding vertex pairs are shown
using the same number and the symmetry centers are shown in blue and red. d) Finally, the optimization generates a mesh with planar faces.
The most interesting aspect of polyhedral patterns is that most of them have to transform so that they look different in regions of positive, zero,
and negative Gaussian curvature (see insets).

fore, we identify invariant symmetries of tiles and then regularize the Symmetry optimization Our algorithm optimizes the pattern for
mesh in our planarization process to maintain them. The symmetries planarity and aesthetics (using the regularizer configured in the
that we identify include reflection through axes and through planes previous step) through non-linear optimization. The details for the
as well as reflections through the centers of tiles or edges. It is optimization framework are presented in Section 5, and those for the
straightforward to check that such symmetries are general enough symmetry regularizers are given in Section 6.
to contain the deformations we describe here. The symmetries are
described in greater detail in Section 6.
5 Optimization Framework
4 Overview and User Interaction We next describe the regularity-based planarity optimization frame-
work that our work builds upon. The inputs are a reference surface,
Our framework comprises four stages, shown in Figure 13. S, given as a triangle mesh, and an initial polygonal mesh with
vertices, vi , that approximates the reference surface. The goal is to
optimize the initial polygonal mesh, M = (V, E, F ), according to
Pre-processing Initial meshes are generated using triangular-, three terms: the planarity of the faces, the closeness to the reference
quad-, or hex-based remeshing techniques in a separate program, ac- surface, and the regularity of the mesh. We rely on existing methods
cording to the desired pattern (see Figure 5 for a description). Many (described in this section) to formulate planarity and closeness terms.
patterns are initialized using the hex-based remeshing approach The regularity terms are our contribution.
proposed by Vaxman and Ben-Chen [2015] with their planarity opti-
mization omitted. We therefore have two input meshes in our system:
Variables We denote the vertex coordinates of M as vi , i ∈ V ,
A finely-tessellated triangle mesh to define the reference surface and
and the unit face normals as nk , k ∈ F . Vertices are not constrained
a coarser (non-planar) remeshed triangle, quad, or hex mesh.
to lie on the reference surface, S, exactly. The closest point on S for
a vertex vi is vi∗ with corresponding normal n∗i (see Fig. 14).
Pattern generation The user can transform the initial coarse mesh n
nkkkkkk
n
into a pattern mesh by selecting from a list of pre-defined patterns. n∗i∗i∗i∗i∗i∗i
n
n
The transformation is implemented using a sequence of geometric v
viiiiii
v
v
rules, e.g., subdivision rules. The implementation of such rules
is fairly straightforward and follows the framework proposed by v
v v
vjjjjjj
v vi∗ii∗∗ii∗∗i∗
v
v
viiiiii v
Akleman et al. [2005].

Figure 14: Notation: faces (left), closest point projection (right)

Symmetry configuration The user can then specify a desired
strip decomposition and configure the regularizer by assigning sym-
metries. We offer axial symmetries with respect to an axis passing
through a vertex, an edge midpoint, or a face barycenter. We also Problem formulation The objective function we minimize is
offer reflective symmetries with respect to a plane, e.g., a plane pass- X j
ing through an edge. The user specifies the symmetry assignment E(vi ) = λ1 Eplan + λ2 Eclose + µj Ereg . (1)
for one or more elements, and the system propagates the assign- j
ment over the whole mesh according to the strip decomposition. To
guide the user in his/her selection, we provide a list of suggested Following the reasoning of [Tang et al. 2014], we set up a system
symmetries. The suggestions are generated by mapping each strip with energies that are at most quartic, which entails soft constraints
decomposition of each pattern to a cylinder and then observing what that are at most quadratic, since this formulation is easy to optimize
symmetries are feasible (see Sec. 3). using a standard regularized Gauss-Newton algorithm.
Planarity The planarity constraint is necessary for all non- ❶ ❸ ❷
triangular faces. We adapt the formulation of [Tang et al. 2014]
and express Eplan as ⓿ ⓿ ⓿ ❶ ❹ ⓿ ❶ ❺ ⓿ ⓿

X X X ❶ ❶ ⓿ ❺ ❶ ⓿ ❹ ❶ ❶
Eplan = ((vi −vj )·nk )2 + (nk ·nk −1)2 , (2)
❷ ❸
k∈F (i,j)∈E(fk ) k

which is zero if all face edges are orthogonal to a unit length normal. Figure 16: Left to right: symmetry with respect to a vertex, an edge
midpoint, a face barycenter, and an edge.
Closeness The closeness constraint of a vertex, vi , to a reference
surface is modeled by requiring vi to move only on the tangent
plane associated with its closest point, vi∗ , on the reference surface, Axial symmetries An affine reflection in an axis, Ak , requires the
S: additional prescription of a reference plane, Tk (not parallel to Ak ;
see Fig. 15, left). Then, a pair of vertices vi and vj is symmetric
X
Eclose = ((vi − vi∗ ) · n∗i )2 . (3)
vi ∈V
with respect to Ak if the midpoint between vi and vj lies on Ak and
the vector vi − vj is parallel to Tk . Let Ak be defined by a direction
As shown in Figure 14, n∗i
is the normal of the tangent plane at vi∗ , vector ak and a point ck , and let n∗k be a normal vector of Tk . Then,
1
and it is kept constant in every iteration. Alternatively, for coarse the axial symmetry regularizer Ereg is encoded as follows:
and inconsistent tilings, we may use closeness of face barycenters X
instead to relax this constraint (see Figures 18 and 29, left). ((vi + vj )/2 − (ck + λkl ak ))2 + ((vi − vj ) · n∗k )2 (5)
(i,j,k)

Previous iteration We add a term in each iteration that dampens

the optimization for stability by closeness to the previous iteration: The triplets (i, j, k) are chosen according to the user-assigned sym-
metries. That is, vi and vj are selected to have affine symmetry with
vi − vim−1 2 + β
X  m  X  m
ni − nm−1
2 respect to Ak . Furthermore, ak and λkl are considered as additional
Eprev = β i
 , (4)
variables in our optimization. The axis point ck is a vertex, an edge
vi ∈V ni ∈F
midpoint, or a face barycenter, and thus it is a linear combination
of vertex coordinates. The n∗k variables can be either considered as
where m denotes the number of an iteration and vim is the value of additional variables, or approximated at the beginning of each itera-
vi at iteration m. We use β = 10−6 in all our examples. tion as the normal of S at the closest point to ck . We do the latter.
Symmetry is applied to local neighborhoods as illustrated in Fig. 16.
6 Regularization with Affine Symmetries Note that the symmetry of a planar face with respect to its barycenter
does not require an axis. n∗k = ak models a Euclidean reflection in
We next define the invariant symmetries of deforming patterns (see Ak and is therefore suitable to enforce Euclidean symmetries.
Section 3) and how we utilize them in practice to regularize pat-
terns undergoing deformations through the planarization process. Plane-reflective symmetries Fig. 15 (right) presents an affine
Generally speaking, there are several ways to represent feasible reg- reflection with respect to a plane, Pk (through point bk and with
ularities, such as enforcing specific angles, polyline smoothness of normal vector n̄k ), in the direction rk . Plane-reflective symmetry
selected sequences of non-adjacent vertices, ratios of edge lengths, of vi and vj requires their midpoint to be located on Pk , and the
and more. We choose to use local affine symmetries as described in vector vi − vj to be parallel to rk . We encode this requirement in
the following, because they are simple, sparse, local and linear (e.g., 2
the regularizer, Ereg , as follows:
compared with angle-based formulations), and this is important for
computational efficiency. X
(((vi + vj )/2 − bk ) · n̄k )2 + ((vi − vj ) − λkl rk )2 . (6)
We describe the practical implementation of various symmetry regu- (i,j,k)
larizers and adapt them to the discrete surface by two approaches:
affine symmetries in space and in a tangential projection. Each We use rk for each reflection plane and the scale variable λkl as
approach has different merits and shortcomings. additional variables. The plane Pk is commonly the bisection plane
of two adjacent faces in a polyhedral mesh, and so bk can be the
midpoint of the common edge of two adjacent faces. The normal,
6.1 Affine symmetries
n̄k , is then pre-estimated in each iteration. With rk = n̄k , we obtain
a Euclidean reflection.
Affine symmetries can be defined with respect to either an axis or a
plane. We can distinguish four different generators of symmetries:
vertices, faces, edge midpoints, and edges. Vertices, faces, and edge Symmetry centers In Fig. 16, as exemplified on a quad mesh, the
midpoints generate symmetries with respect to an axis and edges three leftmost images show point symmetries in 2D, equivalent to
generate symmetries with respect to a plane. 3D axial symmetries. The blue dot represents the symmetry center.
Each pair of symmetric points is labelled the same (orange dots).
Ak Their symmetry centers are located at a vertex, an edge midpoint or a
vi vj vj face barycenter, and their symmetries are denoted accordingly. The
rightmost edge symmetry is equivalent to a 3D planar symmetry.
n∗k ak vi rk
bk
n̄k
Tk ck Pk 6.2 Symmetry in a tangential projection

A relaxed version of affine axial symmetry is symmetry in a projec-

Figure 15: Affine reflection in an axis (left) and in a plane (right) tion parallel to a certain direction (the image plane of the projection
does not matter). To achieve it, we simply discard the second part of
Figure 17: Semi-regular patterns on a Dupin cyclide. Left: A (3, 4, 6, 4) pattern using face symmetries. Middle: A (3, 6, 3, 6)∗ pattern using
symmetries with respect to an edge. Right: A (4, 8, 8) pattern using face symmetries.

Equation 5. We use the normal at the closest point, c∗k , to ck as the

projection direction, and thus enforce a symmetry that is relative to
the tangent plane of S at c∗k . This yields the regularizer Ereg
3
:
X
((vi + vj )/2 − (ck + λkl ak ))2 . (7)
(i,j,k)

By approximating ak with n∗k , this expression can be simplified to

the following equivalent formulation:
X
(((vi + vj )/2 − ck ) · tkl )2 . (8)
(i,j,k,l)

In addition, we need to sum over two orthogonal directions, tk1 and

α = 34.4◦ α = 56.3◦
tk2 . Both directions are orthogonal to n∗k , and they are estimated at γ β = 68.7◦ β = 27.2◦
the beginning of each iteration. γ
α β γ = 76.9◦ α β γ = 96.5◦

6.3 Avoiding self-intersection

Figure 18: Triangle meshes constructed by a single type of triangle
The proposed symmetry regularizers cannot prevent self- for the (4, 8, 8)∗ pattern (left) and the (4, 6, 12)∗ pattern (right)
intersections within the pattern. To counter that, we introduce an using symmetries with respect to an edge.
4
additional regularizer, Ereg , assuming that a line segment connect-
ing the face barycenter to a vertex is within the corresponding face:
X 7 Results
((v̂ki − ck ) × (v̂ki+1 − ck ) · nk − νki
2 2
) . (9)
k∈F We present several results and discuss parameters, planarity, running
0≤i<|fk |
time, failure cases, and comparison to related work.
We assume that the faces are consistently oriented with vertices
|f |−1
(v̂k0 , v̂k1 , . . . , v̂k k ), where indices in the sum are taken modulo Models and patterns We generate results for a set of selected
the face valence |fk |. The face normals, nk , and barycenters, ck , are surfaces showing multiple patterns per surface (see Figures 17, 23,
evaluated prior to each iteration, and considered as constants. We in- and 24) where all patterns use the simplest strip decomposition.
troduce νki as slack variables to encode the inequality requirements Throughout the paper, we select the surfaces to highlight the be-
that the vectorial areas of ck v̂ki v̂ki+1 are aligned with nk . havior of our regularizer in different situations. We select models
to include regions of positive and negative Gaussian curvature, as
6.4 Edge length regularization well as interesting transition regions between them. We exemplify
patterns on both open and closed surfaces, as well as surfaces with
To avoid short edges, we regularize selected length differences of topological holes. We show a wood construction of an interior
adjacent edges. For example, if ei and ej are the edge vectors of cladding of an architectural model in Figure 2. Finally, our method
two neighboring edges, we regulate the ratio of their lengths into a is used to generate rough patterns consisting of identical triangles
5
given interval, and Ereg is defined as: (see Figure 18) and a (3, 4, 6, 4) pattern on a non-architectural model
X (see Figure 19). Note that our algorithm typically targets architec-
(kei k − r kej k − µ2ij )2 + (kej k2 − r2 kei k2 − µ2ji )2 , (10)
2 2 2
tural models with moderate curvature variations and a sparse set of
(i,j) singularities. This model is therefore mainly depicted to demonstrate
where the summation is over the pairs of edges chosen based on the robustness.
used strip decomposition and r is set to 0.8 in our implementation.
6
We can require a lower bound, lmin , on edge lengths by Ereg : Strip decompositions We demonstrate a pattern on a surface
X 2 2 2 2 with different strip decompositions (see Figures 1 and 30). Different
(kei k − lmin − γi ) . (11) symmetries are used to accommodate each decomposition. Com-
i∈E
bining different strips and symmetries leads to a large variety of
µij and γi are slack variables for the inequality requirements. aesthetic results from a single basic pattern (see Figure 29).
 Figure 19: We show a
(3, 4, 6, 4) pattern on the
Moomoo model with 57 sin-
gularities. This is the most
complicated model shown
in the paper due to the
high number of singulari-
ties (hexagons replaced by
septagons and pentagons) Figure 21: Failure case on a monkey saddle: due to the initialization
and the high curvature vari- used on the left, the pattern (3, 4, 6, 4) degenerates as some of the
ations. quads collapse to lines. With a different initialization that is better
aligned with the principal curvature directions, the pattern can be
mapped correctly (right).

Planarity All our models are successfully planarized. For assess-

ment, we measure face planarity as the maximum distance between
a vertex and a regression plane computed using PCA, normalized by
dividing by the average edge length in the model. Our tolerance for
this measure is under 10−2 . We illustrate one example in Figure 13
where we show how the Soumaya museum model is planarized.
Before the optimization, many faces of the model are considerably
non-planar. The optimization is nevertheless successful.

Parameters The main parameters stem from the configuration of

the regularizer, i.e., the selected symmetries and strip decomposi-
tions. The weights for the different terms in the optimization vary
slightly per pattern. In the optimization, we use the default parame-
ters of 1.0 for face planarity, 0.1 for the closeness to the reference Figure 22: A failure case due to improper initialization. With a hex-
surface, and 0.01 for the symmetry-based regularizers. This choice dominant mesh dualized from of an arbitrary triangle mesh (left), the
is good for most examples. For example, the (34 , 6) pattern in Fig- computed result (right) cannot achieve regularity or planarity, be-
ure 24, right and the (3, 4, 6, 4) pattern in Figure 17, left use these cause there are too many singularities, leading to a over-fragmented
parameters without change. strip decomposition.
Alternatively, the user can adjust the parameters to trade off reg-
ularity for stricter planarity or closeness to the reference surface;
see Figure 20 for example. For the (4, 6, 12) patterns in Figure 23 432 vertices took 0.53 seconds to optimize.
and 24, we set the planarity to 5.0 and 10.0, respectively, while
setting the closeness and symmetry parameters to the default ones. Failure cases and limitations Our framework is sensitive to the
For the remaining terms, we advise that the edge-length parameter triangle, quad or hex mesh that is used as input. A poor initialization
be equated with the symmetry regularization parameter, and the leads to poor results. In Figure 21, we contrast the results of a poorly
self-intersection avoidance parameter with the planarity parameter. initialized optimization with a good initialization on the monkey
saddle. In Figure 22, we show a poor initialization of the Soumaya
model with many singularities in comparison with the example
Implementation details and running times We implemented
demonstrated above in Figure 13. In addition, we depend on the
our framework in C++ using OpenMesh [Botsch et al. 2002] for the
quality of existing code for the creation of the initial hex or quad
mesh data structures and TAUCS [Krieg et al. 2003] as the library
patterns on the surface.
for sparse linear solvers. The running times for our examples are
typically under one minute with an Intel Xeon X5550 2.67GHz
processor. For example, the Soumaya model in Figure 13 has 7034 Coarse meshes Our algorithm can create adequate results on
vertices and took 43 seconds to optimize. Smaller examples are coarse meshes as well. However, we observe that coarse meshes can
much faster. For example, a (4, 6, 12) pattern on an HP surface with typically only be planarized and regularized when sacrificing the
closeness to the reference surface, as demonstrated in Figure 20.

Comparison to PQ meshing Our regularizers provide more de-

grees of freedom than does the traditional regularizer based on
polyline fairness [Tang et al. 2014]. Therefore, our planarization
is less sensitive to PQ meshes that are not initialized according to
Figure 20: A planarized cyclide model. Hotter vertex colors indi- conjugate directions. We show examples in Figures 4, 27, and 29.
cate greater distances to the reference surface (left). A coarse input A detailed analysis of the difference is provided in the additional
mesh can be successfully planarized and regularized at the cost of materials.
lower fidelity to the reference surface (second from left). Forcing
closeness to the reference surface sacrifices regularity (second from Comparison to PH meshing Li et al. [2014] proposed a regular-
right) and may cause edge degeneracies (right). izer for planar hex meshes, meshing the positively and negatively
Figure 23: Left: A (3, 4, 6, 4) pattern using face symmetries. Middle: A (3, 6, 3, 6) pattern using vertex symmetries. Right: A (4, 6, 12)
pattern using face symmetries.

Figure 24: Semi-regular patterns on an architectural six shape. Left: A (4, 6, 12) pattern using face symmetries. A (34 , 6)∗ pattern using
vertex symmetries. Note how prominent feature lines form automatically due to the regularization.

curved regions separately. However, they do not propose a specific

solution for the transition region and cannot automatically assign
which regularizer to use. This may lead to artifacts in the transition
region (see Figure 25) and to possible failures in the planarization
(Figure 26). Our solution can produce significantly better results.
It also caters to non-canonical strip decomposition. However, we
note that Li et al. [2014] propose a complete framework for PH Figure 26: Comparison of Phex mesh planarity. Left: [Li et al.
remeshing, while our paper focuses only on the regularizer used 2014]. Right: our method achieves better planarity.
after the generation of an initial mesh layout.

Comparison to ad-hoc regularizers A large variety of regular-

izers has been proposed in other contexts. When using ad-hoc
regularizers, typical problems occur, depending on how the regular-
izer is weighted. On the one hand, using a high weight leads to a
mesh that is visually pleasing, but not polyhedral. On the other hand,
using a low weight leads to a planar mesh that is highly irregular,
including degeneracies like self-intersections. There is no effective
weight that can achieve both planarity and regularity simultaneously. Figure 29: Different patterns generated by different symmetries.
Figures 27 and 28 present examples. The pattern on the left uses face symmetries; the one on the right is
based on edge midpoint symmetries.

8 Conclusions

We consider the design and optimization of polyhedral patterns,

i.e. patterns of planar polygonal faces, on freeform surfaces. Our
contributions are the description of a novel class of regularizers
based on affine symmetries and a theoretical analysis of polyhedral
Figure 25: Comparison of Phex mesh aesthetics: with the same patterns. In future work, we plan to study mixed patterns and their
initialization, the regularizer of [Li et al. 2014] generates the left transition regions, volumetric patterns such as frame structures for
mesh, while our approach leads to more natural transitions (right). support in architectural applications, folding patterns, and time-
varying polyhedral patterns for shading systems.
 Figure 27: Mesh planarity. From left to right: initialization, polyline fairness, Laplacian, edge length, angles, face area, ours.

Figure 28: Mesh planarity. From left to right: initialization, Laplacian, edge length, angles, face area, ours.

Acknowledgments K RIEG , O. D., S CHWINN , T., M ENGES , A., L I , J.-M., K NIPPERS ,

J., S CHMITT, A., AND S CHWIEGER , V. 2014. Biomimetic
We thank the anonymous reviewers for their insightful comments and lightweight timber plate shells: Computational integration of
suggestions for improving the paper. This research was supported robotic fabrication, architectural geometry and structural design.
by the Visual Computing Center (VCC) at KAUST and by the In Advances in Architectural Geometry. Springer, 109–125.
Austrian Science Fund (FWF) via grants P23735-N13 and I706-
N26 (DFG-Collaborative Research Center, TRR 109, Discretization L I , Y., L IU , Y., AND WANG , W. 2014. Planar hexagonal meshing
in Geometry and Dynamics). Caigui Jiang and Chengcheng Tang for architecture. IEEE Trans. Vis. Comp. Graphics 21, 95–106.
were supported by KAUST baseline funding. L IU , Y., P OTTMANN , H., WALLNER , J., YANG , Y.-L., AND
WANG , W. 2006. Geometric modeling with conical meshes
References and developable surfaces. ACM Trans. Graph. 25, 3, 681–689.
L IU , Y., X U , W., WANG , J., Z HU , L., G UO , B., C HEN , F., AND
A BAS , S. J., S ALMAN , A. S., M OUSTAFA , A., AND ATIYAH , M. WANG , G. 2011. General planar quadrilateral mesh design using
1995. Symmetries of Islamic geometrical patterns, vol. 3. World conjugate direction field. ACM Trans. Graph. 30, #140, 1–10.
Scientific.
L U , P. J., AND S TEINHARDT, P. J. 2007. Decagonal and quasi-
A KLEMAN , E., S RINIVASAN , V., AND M ANDAL , E. 2005. crystalline tilings in medieval islamic architecture. Science 315,
Remeshing schemes for semi-regular tilings. In Proc. Inter. Conf. 5815, 1106–1110.
on Shape Modeling and Applications, 44–50.
P ORANNE , R., OVREIU , E., AND G OTSMAN , C. 2013. Interactive
B OBENKO , A., AND S URIS , Y U . 2008. Discrete differential geom- planarization and optimization of 3D meshes. Comput. Graph.
etry: Integrable Structure. American Math. Soc. Forum 32, 1, 152–163.
B OTSCH , M., S TEINBERG , S., B ISCHOFF , S., AND KOBBELT, L., S AUER , R. 1970. Differenzengeometrie. Springer.
2002. Openmesh - a generic and efficient polygon mesh data
structure. S CHIFTNER , A., H ÖBINGER , M., WALLNER , J., AND P OTTMANN ,
H. 2009. Packing circles and spheres on surfaces. ACM Trans.
B OUAZIZ , S., S CHWARTZBURG , Y., W EISE , T., AND PAULY, M. Graph. 28, 5, #139,1–8.
2012. Shaping discrete geometry with projections. Computer
Grapics Forum 31, 1657–1667. S INGH , M., AND S CHAEFER , S. 2010. Triangle surfaces with
discrete equivalence classes. ACM Trans. Graph. 29, #46,1–7.
C OHEN -S TEINER , D., A LLIEZ , P., AND D ESBRUN , M. 2004.
Variational shape approximation. ACM Trans. Graphics 23, 3, TANG , C., S UN , X., G OMES , A., WALLNER , J., AND P OTTMANN ,
905–914. H. 2014. Form-finding with polyhedral meshes made simple.
ACM Trans. Graphics 33, 4.
D ENG , B., B OUAZIZ , S., D EUSS , M., K ASPAR , A., S CHWARTZ -
BURG , Y., AND PAULY, M. 2015. Interactive design exploration VAXMAN , A., AND B EN -C HEN , M. 2015. Dupin meshing: A
for constrained meshes. Computer-Aided Design 61, 13–23. parameterization approach to planar hex-dominant meshing. Tech.
Rep. CS-2015-01 (CS series), Technion.
H UARD , M., B OMPAS , P., AND E IGENSATZ , M. 2014. Planar
panelization with extreme repetition. In Advances in Architectural Z ADRAVEC , M., S CHIFTNER , A., AND WALLNER , J. 2010. De-
Geometry 2014, P. Block et al., Eds. Springer. signing quad-dominant meshes with planar faces. Comput. Graph.
Forum 29, 5, 1671–1679.
J IANG , C., WANG , J., WALLNER , J., AND P OTTMANN , H. 2014.
Freeform honeycomb structures. Comput. Graph. Forum 33, 5, Z IMMER , H., C AMPEN , M., B OMMES , D., AND KOBBELT, L.
185–194. 2012. Rationalization of triangle-based point-folding structures.
Comput. Graph. Forum 31, 611–620.
K RIEG , O. D., S CHWINN , T., M ENGES , A., L I , J.-M., K NIPPERS ,
J., S CHMITT, A., AND S CHWIEGER , V. 2003. Taucs. In A Z IMMER , H. 2014. Optimization of 3D models for fabrication. PhD
Library of Sparse Linear Solvers. TAU. thesis, RWTH Aachen.
 Figure 30: Three strip decompositions on the knot model for the (4, 8, 8) pattern. Figure 1 presents an explanation of the color coding.

Appendix: conjugacy of patterns We next prove how planar lifting leads to consistency.

We prove that a primal pattern can be vertically lifted to a Lemma 3. The vertical projection of the intersection curve between
paraboloid, S, while every face remains planar, if and only if S and a plane P := z = ax + by + e onto the (x, y)-plane is a
every primal edge in the pattern is conjugated to and bisected conic of the form |p − c|2 = γ, γ ∈ R.
by the corresponding dual edge (i.e., the edge between the conic
centers of the neighboring faces). The paraboloid, S, is defined Proof. The surface obtained by the P − S subtraction is a vertical
as z = κ1 x2 + κ2 y 2 . We currently assume κ1 κ2 6= 0 and refer cylinder passing through the intersection curve, which intersects the
to this assumption later. We consider the induced quadratic form (x, y)-plane with a conic of the form κ1 (xp −xc )2 +κ2 (yp −yc )2 =
ha, bi := κ1 xa xb + κ2 ya yb on the (x, y)-plane and denote the |p − c|2 = γ, where xc = 2κa1 , yc = 2κb 2 , and γ ∈ R.
squared norm |a|2 = ha, ai accordingly. The conjugacy relation is
thus ha, bi = 0. We prove our claim using the following: Lemma 4. On the (x, y)-plane, if two similar conics ci :
|p − ci |2 = γi and cj : |p − cj |2 = γj intersect at pi and pj ,
Lemma 1. A polygon with vertices (p1 , p2 , . . . , pn ) on the x, y- then pj − pi is conjugate to and bisected by cj − ci .
plane, inscribed to a conic of the form |p − c|2 = γ, γ ∈ R,
remains planar when lifted to S. Proof. The subtraction of the two conics, cj − ci , produces a line
l defined by hp, cj − ci i = const. As both pi and pj are on l,
Proof. The lifted points of pi on the paraboloid are on the in- hpj − pi , cj − ci i = 0.
tersection of S of the vertical cylinder C extruding the conic:
To show that bisection holds as well, we apply a shearing transforma-
κ1 (x − xc )2 + κ2 (y − yc )2 = γ. Thus, they also lie on any
tion so that c̄j − c̄i and p̄j − p̄i become orthogonal and thus aligned
linear combination of C and S. The C − S surface is a plane, as the
with the axes. As c̄i and c̄j are both reflectively symmetric with
quadratic terms cancel out.
respect to c̄j −c̄i , so are their intersections. The midpoint of pj −pi
Lemma 2. Let pj − pi be a primal edge conjugate to the dual edge therefore lies on cj − ci before the shearing transformation.
cj − ci and bisected by cj − ci at point d. Then, |pi − ci |2 = Corollary 2. Planar lifting to consistency: Consider a pattern on
|pj − ci |2 and |pi − cj |2 = |pj − cj |2 . the (x, y)-plane which is the vertical projection of a polyhedral
pattern inscribed on S. Then, there exists a dual pattern with each
Proof. The conjugacy of pj −pi and cj −ci implies hpj −pi , cj − edge conjugate to and bisecting the corresponding primal edge.
ci i = 0, and the bisecting condition implies |pi − d|2 = |pj − d|2 .
Thus: Proof. Due to Lemma 3, the faces of the primal pattern are all
inscribed to conics with the form |p − ci |2 = γi , γi ∈ R. Consider
|pi − ci |2 = |d − ci + pi − d|2 the dual pattern formed by connecting the adjacent face circumconic
= hd − ci , d − ci i + hpj − d, pj − di centers. Then, the primal edges pj −pi are conjugate to and bisected
by the dual edges cj − ci due to Lemma 4.
= hpj − ci , pj − ci i.

Similarly, hpi − cj , pi − cj i = hpj − cj , pj − cj i. Cylindrical paraboloids When κ1 κ2 = 0, κ1 6= κ2 , Corollary 1

still holds, but Corollary 2 is no longer generally exact, as the proof
Corollary 1. Consistency to planar lifting: A primal pattern on of Lemma 3 requires κ1 6= 0 and κ2 6= 0. The reason for this is that
the (x, y)-plane with every edge conjugate to and bisected by the the conics made by this condition are pairs of straight lines without
dual edge is liftable to S while the faces remain planar. any center, and thus the duality is not well defined. However, by
identifying centers, this is still true in the limit of our deformation
Proof. Due to Lemma 2, for a given polygon {pi } and center c, we scheme, as we continuously deform regular tilings from elliptic
have hpi − c, pi − ci = γ for some constant γ ∈ R, and thus the to hyperbolic conics with well-defined centers. Thus, we have a
vertices are on a planar conic and can be lifted to S while keeping well-defined dual center obeying these conditions.
face planarity, due to Lemma 1.