Alexander I.

Bobenko Editor

Advances in
Discrete Differential
Geometry
Advances in Discrete Differential Geometry
Alexander I. Bobenko
Editor

Advances in Discrete
Differential Geometry
Editor
Alexander I. Bobenko
Institut für Mathematik
Technische Universität Berlin
Berlin
Germany

ISBN 978-3-662-50446-8 ISBN 978-3-662-50447-5 (eBook)

DOI 10.1007/978-3-662-50447-5
Library of Congress Control Number: 2016939574

© The Editor(s) (if applicable) and The Author(s) 2016. This book is published open access.
Open Access This book is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any non-
commercial use, distribution, and reproduction in any medium, provided the original author(s) and source
are credited.
The images or other third party material in this book are included in the work’s Creative Commons
license, unless indicated otherwise in the credit line; if such material is not included in the work’s
Creative Commons license and the respective action is not permitted by statutory regulation, users will
need to obtain permission from the license holder to duplicate, adapt or reproduce the material.
This work is subject to copyright. All commercial rights are reserved by the Publisher, whether the whole
or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publi-
cation does not imply, even in the absence of a specific statement, that such names are exempt from the
relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer-Verlag GmbH Berlin Heidelberg
Preface

In this book we take a closer look at discrete models in differential geometry and
dynamical systems. The curves used are polygonal, surfaces are made from trian-
gles and quadrilaterals, and time runs discretely. Nevertheless, one can hardly see
the difference to the corresponding smooth curves, surfaces, and classical dynam-
ical systems with continuous time. This is the paradigm of structure-preserving
discretizations. The common idea is to find and investigate discrete models that
exhibit properties and structures characteristic of the corresponding smooth geo-
metric objects and dynamical processes. These important and characteristic quali-
tative features should already be captured at the discrete level. The current interest
and advances in this field are to a large extent stimulated by its relevance for
computer graphics, mathematical physics, architectural geometry, etc.
The book focuses on differential geometry and dynamical systems, on smooth
and discrete theories, and on pure mathematics and its practical applications. It
demonstrates this interplay using a range of examples, which include discrete con-
formal mappings, discrete complex analysis, discrete curvatures and special sur-
faces, discrete integrable systems, special texture mappings in computer graphics,
and freeform architecture. It was written by specialists from the DFG Collaborative
Research Center “Discretization in Geometry and Dynamics”. The work involved in
this book and other selected research projects pursued by the Center was recently
documented in the film “The Discrete Charm of Geometry” by Ekaterina Eremenko.
Lastly, the book features a wealth of illustrations, revealing that this new branch
of mathematics is both (literally) beautiful and useful. In particular the cover
illustration shows the discretely conformally parametrized surfaces of the inflated
letters A and B from the recent educational animated film “conform!” by Alexander
Bobenko and Charles Gunn.
At this place, we want to thank the Deutsche Forschungsgesellschaft for its
ongoing support.

Berlin, Germany Alexander I. Bobenko

November 2015

v
Contents

Discrete Conformal Maps: Boundary Value Problems, Circle

Domains, Fuchsian and Schottky Uniformization . . . . . . . . . . . . . . . . . 1
Alexander I. Bobenko, Stefan Sechelmann and Boris Springborn
Discrete Complex Analysis on Planar Quad-Graphs . . . . . . . . . . . . . . . 57
Alexander I. Bobenko and Felix Günther
Approximation of Conformal Mappings Using Conformally
Equivalent Triangular Lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Ulrike Bücking
Numerical Methods for the Discrete Map Za . . . . . . . . . . . . . . . . . . . . 151
Folkmar Bornemann, Alexander Its, Sheehan Olver
and Georg Wechslberger
A Variational Principle for Cyclic Polygons with Prescribed Edge
Lengths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Hana Kouřimská, Lara Skuppin and Boris Springborn
Complex Line Bundles Over Simplicial Complexes
and Their Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
Felix Knöppel and Ulrich Pinkall
Holomorphic Vector Fields and Quadratic Differentials
on Planar Triangular Meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Wai Yeung Lam and Ulrich Pinkall
Vertex Normals and Face Curvatures of Triangle Meshes . . . . . . . . . . 267
Xiang Sun, Caigui Jiang, Johannes Wallner and Helmut Pottmann
S-Conical CMC Surfaces. Towards a Unified Theory of Discrete
Surfaces with Constant Mean Curvature . . . . . . . . . . . . . . . . . . . . . . . 287
Alexander I. Bobenko and Tim Hoffmann

vii
viii Contents

Constructing Solutions to the Björling Problem for Isothermic

Surfaces by Structure Preserving Discretization . . . . . . . . . . . . . . . . . . 309
Ulrike Bücking and Daniel Matthes
On the Lagrangian Structure of Integrable Hierarchies . . . . . . . . . . . . 347
Yuri B. Suris and Mats Vermeeren
On the Variational Interpretation of the Discrete KP Equation . . . . . . . 379
Raphael Boll, Matteo Petrera and Yuri B. Suris
Six Topics on Inscribable Polytopes . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Arnau Padrol and Günter M. Ziegler
DGD Gallery: Storage, Sharing, and Publication of Digital
Research Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
Michael Joswig, Milan Mehner, Stefan Sechelmann, Jan Techter
and Alexander I. Bobenko
Contributors

Alexander I. Bobenko Inst. für Mathematik, Technische Universität Berlin,

Berlin, Germany
Raphael Boll Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Folkmar Bornemann Zentrum Mathematik – M3, Technische Universität
München, Garching bei München, Germany
Ulrike Bücking Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Felix Günther Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Tim Hoffmann Zentrum Mathematik – M10, Technische Universität München,
Garching bei München, Germany
Alexander Its Department of Mathematical Sciences, Indiana University–Purdue
University, Indianapolis, IN, USA
Caigui Jiang King Abdullah Univ. of Science and Technology, Thuwal, Saudi
Arabia
Michael Joswig Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Felix Knöppel Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Hana Kouřimská Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Wai Yeung Lam Technische Universität Berlin, Inst. Für Mathematik, Berlin,
Germany

ix
x Contributors

Daniel Matthes Zentrum Mathematik – M8, Technische Universität München,

Garching bei München, Germany
Milan Mehner Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Sheehan Olver School of Mathematics and Statistics, The University of Sydney,
Sydney, Australia
Arnau Padrol Institut de Mathématiques de Jussieu - Paris Rive Gauche, UPMC
Univ Paris 06, Paris Cedex 05, France
Matteo Petrera Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Ulrich Pinkall Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Helmut Pottmann Vienna University of Technology, Wien, Austria
Stefan Sechelmann Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Lara Skuppin Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Boris Springborn Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Xiang Sun King Abdullah Univ. of Science and Technology, Thuwal, Saudi
Arabia
Yuri B. Suris Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Jan Techter Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Mats Vermeeren Inst. für Mathematik, Technische Universität Berlin, Berlin,
Germany
Johannes Wallner Graz University of Technology, Graz, Austria
Georg Wechslberger Zentrum Mathematik – M3, Technische Universität
München, Garching bei München, Germany
Günter M. Ziegler Inst. für Mathematik, Freie Universität Berlin, Berlin,
Germany
Discrete Conformal Maps: Boundary Value
Problems, Circle Domains, Fuchsian
and Schottky Uniformization

Alexander I. Bobenko, Stefan Sechelmann and Boris Springborn

Abstract We discuss several extensions and applications of the theory of discretely

conformally equivalent triangle meshes (two meshes are considered conformally
equivalent if corresponding edge lengths are related by scale factors attached to
the vertices). We extend the fundamental definitions and variational principles from
triangulations to polyhedral surfaces with cyclic faces. The case of quadrilateral
meshes is equivalent to the cross ratio system, which provides a link to the theory of
integrable systems. The extension to cyclic polygons also brings discrete conformal
maps to circle domains within the scope of the theory. We provide results of numer-
ical experiments suggesting that discrete conformal maps converge to smooth con-
formal maps, with convergence rates depending on the mesh quality. We consider
the Fuchsian uniformization of Riemann surfaces represented in different forms:
as immersed surfaces in R3 , as hyperelliptic curves, and as CP1 modulo a classi-
cal Schottky group, i.e., we convert Schottky to Fuchsian uniformization. Extended
examples also demonstrate a geometric characterization of hyperelliptic surfaces
due to Schmutz Schaller.

1 Introduction

Not one, but several sensible definitions of discrete holomorphic functions and
discrete conformal maps are known today. The oldest approach, which goes back
to the early finite element literature, is to discretize the Cauchy–Riemann equa-

A.I. Bobenko · S. Sechelmann · B. Springborn (B)

Inst. für Mathematik, Technische Universität Berlin, Straße des 17. Juni 136,
10623 Berlin, Germany
e-mail: [email protected]
A.I. Bobenko
e-mail: [email protected]
S. Sechelmann
e-mail: [email protected]

c The Author(s) 2016 1
A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_1
2 A.I. Bobenko et al.

tions [10–14, 27]. This leads to linear theories of discrete complex analysis, which
have recently returned to the focus of attention in connection with conformal models
of statistical physics [8, 9, 22, 23, 29, 40–42], see also [4].
The history of nonlinear theories of discrete conformal maps goes back to
Thurston, who introduced patterns of circles as elementary geometric way to visual-
ize hyperbolic polyhedra [45, Chapter 13]. His conjecture that circle packings could
be used to approximate Riemann mappings was proved by Rodin and Sullivan [35].
This initiated a period of intensive research on circle packings and circle patterns,
which lead to a full-fledged theory of discrete analytic functions and discrete con-
formal maps [44].
A related but different nonlinear theory of discrete conformal maps is based on
a straightforward definition of discrete conformal equivalence for triangulated sur-
faces: Two triangulations are discretely conformally equivalent if the edge lengths
are related by scale factors assigned to the vertices. This also leads to a surprisingly
rich theory [5, 17, 18, 28]. In this article, we investigate different aspects of this
theory (Fig. 1).
We extend the notion of discrete conformal equivalence from triangulated
surfaces to polyhedral surfaces with faces that are inscribed in circles. The basic
definitions and their immediate consequences are discussed in Sect. 2.
In Sect. 3, we generalize a variational principle for discretely conformally equiv-
alent triangulations [5] to the polyhedral setting. This variational principle is the
main tool for all our numerical calculations. It is also the basis for our uniqueness
proof for discrete conformal mapping problems (Theorem 3.9).
Section 4 is concerned with the special case of quadrilateral meshes. We discuss
the emergence of orthogonal circle patterns, a peculiar necessary condition for the
existence of solutions for boundary angle problems, and we extend the method of
constructing discrete Riemann maps from triangulations to quadrangulations.
In Sect. 5, we briefly discuss discrete conformal maps from multiply connected
domains to circle domains, and special cases in which we can map to slit domains.
Section 6 deals with conformal mappings onto the sphere. We generalize the
method for triangulations to quadrangulations, and we explain how the spherical
version of the variational principle can in some cases be used for numerical calcu-
lations although the corresponding functional is not convex.
Section 7 is concerned with the uniformization of tori, i.e., the representation of
Riemann surfaces as a quotient space of the complex plane modulo a period lattice.
We consider Riemann surfaces represented as immersed surfaces in R3 , and as ellip-
tic curves. We conduct numerical experiments to test the conjectured convergence
of discrete conformal maps. We consider the difference between the true modulus
of an elliptic curve (which can be calculated using hypergeometric functions) and
the modulus determined by discrete uniformization, and we estimate the asymptotic
dependence of this error on the number of vertices.
In Sect. 8, we consider the Fuchsian uniformization of Riemann surfaces repre-
sented in different forms. We consider immersed surfaces in R3 (and S 3 ), hyperellip-
tic curves, and Riemann surfaces represented as a quotient of Ĉ modulo a classical
Schottky group. That is, we convert from Schottky uniformization to Fuchsian uni-
formization. The section ends with two extended examples demonstrating, among
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 3

Fig. 1 Uniformization of compact Riemann surfaces. The uniformization of spheres is treated in

Sect. 6. Tori are covered in Sect. 7, and Sect. 8 is concerned with surfaces of higher genus
4 A.I. Bobenko et al.

other things, a remarkable geometric characterization of hyperelliptic surfaces due

to Schmutz Schaller.

2 Discrete Conformal Equivalence of Cyclic Polyhedral

Surfaces

2.1 Cyclic Polyhedral Surfaces

A euclidean polyhedral surface is a surface obtained from gluing euclidean poly-

gons along their edges. (A surface is a connected two-dimensional manifold, pos-
sibly with boundary.) In other words, a euclidean polyhedral surface is a surface
equipped with, first, an intrinsic metric that is flat except at isolated points where it
has cone-like singularities, and, second, the structure of a CW complex with geo-
desic edges. The set of vertices contains all cone-like singularities. If the surface has
a boundary, the boundary is polygonal and the set of vertices contains all corners of
the boundary.
Hyperbolic polyhedral surfaces and spherical polyhedral surfaces are defined
analogously. They are glued from polygons in the hyperbolic and elliptic planes,
respectively. Their metric is locally hyperbolic or spherical, except at cone-like sin-
gularities.
We will only be concerned with polyhedral surfaces whose faces are all cyclic,
i.e., inscribed in circles. We call them cyclic polyhedral surfaces. More precisely,
we require the polygons to be cyclic before they are glued together. It is not required
that the circumcircles persist after gluing; they may be disturbed by cone-like sin-
gularities. A polygon in the hyperbolic plane is considered cyclic if it is inscribed
in a curve of constant curvature. This may be a circle (the locus of points at con-
stant distance from its center), a horocycle, or a curve at constant distance from a
geodesic.
A triangulated surface, or triangulation for short, is a polyhedral surface all of
whose faces are triangles. All triangulations are cyclic.

2.2 Notation

We will denote the sets of vertices, edges, and faces of a CW complex  by V , E  ,

and F , and we will often omit the subscript when there is no danger of confusion.
For notational convenience, we require all CW complexes to be strongly regular.
This means that we require that faces are not glued to themselves along edges or
at vertices, that two faces are not glued together along more than one edge or one
vertex, and that edges have distinct end-points and two edges have at most one
endpoint in common. This allows us to label edges and faces by their vertices. We
will write ij ∈ E for the edge with vertices i, j ∈ V and ijkl ∈ F for the face with
vertices i, j, k, l ∈ V . We will always list the vertices of a face in the correct cyclic
order, so that for example the face ijkl has edges ij, jk, kl, and li. The only reason
for restricting our discussion to strongly regular CW complexes is to be able to use
this simple notation. Everything we discuss applies also to general CW complexes.
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 5

2.3 Discrete Metrics

The discrete metric of a euclidean (or hyperbolic or spherical) cyclic polyhedral sur-
face  is the function  : E  → R>0 that assigns to each edge ij ∈ E  its length ij .
It satisfies the polygon inequalities (one side is shorter than the sum of the others):
⎫
−i1 i2 + i2 i3 + . . . + in−1 in > 0 ⎪
⎪
⎪
i1 i2 − i2 i3 + . . . + in−1 in > 0 ⎪
⎬
.. for all i 1 i 2 . . . i n ∈ F (1)
. ⎪
⎪
⎪
⎪
⎭
i1 i2 + i2 i3 + . . . − in−1 in > 0

In the case of spherical polyhedral surfaces, we also require that

i1 i2 + i2 i3 + . . . + in−1 in < 2π. (2)

The polygon inequalities (1) are necessary and sufficient for the existence of a
unique cyclic euclidean polygon and a unique cyclic hyperbolic polygon with the
given edge lengths. Together with inequality (2) they are necessary and sufficient
for the existence of a unique cyclic spherical polygon. For a new proof of these ele-
mentary geometric facts, see [24]. Thus, a discrete metric determines the geometry
of a cyclic polyhedral surface:

Proposition and Definition 2.1 If  is a surface with the structure of a CW com-

plex and a function  : E  → R>0 satisfies the polygon inequalities (1), then there
is a unique euclidean cyclic polyhedral surface and also a unique hyperbolic cyclic
polyhedral surface with CW complex  and discrete metric . If  also satisfies
the inequalities (2), then there is a unique spherical cyclic polyhedral surface with
CW complex  and discrete metric .
We will denote the euclidean, hyperbolic, and spherical polyhedral surface with
CW complex  and discrete metric  by (, )euc , (, )hyp , and (, )sph , respec-
tively.

2.4 Discrete Conformal Equivalence

We extend the definition of discrete conformal equivalence from triangulations

[5, 28] to cyclic polyhedral surfaces in a straightforward way (Definition 2.2). While
some aspects of the theory carry over to the more general setting (e.g., Möbius
invariance, Proposition 2.5), others do not, like the characterization of discretely
conformally equivalent triangulations in terms of length cross-ratios (Sect. 2.5). We
will discuss similar characterizations for polyhedral surfaces with 2-colorable ver-
tices and the particular case of quadrilateral faces in Sects. 2.7 and 2.8.
6 A.I. Bobenko et al.

We define discrete conformal equivalence only for polyhedral surfaces that are
combinatorially equivalent (see Remark 2.4). Thus, we may assume that the surfaces
˜
share the same CW complex  equipped with different metrics , .
Definition 2.2 Discrete conformal equivalence is an equivalence relation on the set
of cyclic polyhedral surfaces defined as follows:
• Two euclidean cyclic polyhedral surfaces (, )euc and (, ) ˜ euc are discretely
conformally equivalent if there exists a function u : V → R such that

˜ij = e 2 (u i +u j ) ij .
1
(3)

• Two hyperbolic cyclic polyhedral surfaces (, )hyp and (, )˜ hyp are discretely
conformally equivalent if there exists a function u : V → R such that

 ˜   
= e 2 (u i +u j ) sinh
ij 1 ij
sinh . (4)
2 2

• Two spherical cyclic polyhedral surfaces (, )sph and (, )˜ sph are discretely
conformally equivalent if there exists a function u : V → R such that

 ˜   
= e 2 (u i +u j ) sin
ij 1 ij
sin . (5)
2 2
We will also consider mixed versions:
• A euclidean cyclic polyhedral surface (, )euc and a hyperbolic cyclic polyhe-
˜ hyp are discretely conformally equivalent if
dral surface (, )

 ˜ 
= e 2 (u i +u j ) ij .
ij 1
sinh (6)
2
• A euclidean cyclic polyhedral surface (, )euc and a spherical cyclic polyhedral
˜ sph are discretely conformally equivalent if
surface (, )

 ˜ 
= e 2 (u i +u j ) ij .
ij 1
sin (7)
2
• A hyperbolic cyclic polyhedral surface (, )hyp and a spherical cyclic polyhedral
˜ sph are discretely conformally equivalent if
surface (, )

 ˜   
= e 2 (u i +u j ) sinh
ij 1 ij
sin . (8)
2 2

Remark 2.3 Note that relation (5) for spherical edge lengths is equivalent to rela-
tion (3) for the euclidean lengths of the chords in the ambient R3 of the sphere (see
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 7

Fig. 2 Spherical and 

hyperbolic chords

2 sin 2 2 sinh 2

Fig. 2, left). Likewise, relation (4) for hyperbolic edge lengths is equivalent to (3)
for the euclidean lengths of the chords in the ambient R2,1 of the hyperboloid model
of the hyperbolic plane (see Fig. 2, right).

Remark 2.4 For triangulations, the definition of discrete conformal equivalence has
been extended to meshes that are not combinatorially equivalent [5, Definition 5.1.4]
[17, 18]. It is not clear whether or how the following definitions for cyclic polyhe-
dral surfaces can be extended to combinatorially inequivalent CW complexes.

The discrete conformal class of a cyclic polyhedral surface embedded in n-

dimensional euclidean space is invariant under Möbius transformations of the ambi-
ent space:

Proposition 2.5 (Möbius invariance) Suppose P and P̃ are two combinatorially

equivalent euclidean cyclic polyhedral surfaces embedded in Rn (with straight
edges and faces), and suppose there is a Möbius transformation of Rn ∪ {∞} that
maps the vertices of P to the corresponding vertices of P̃. Then P and P̃ are dis-
cretely conformally equivalent.

Note that only vertices are related by the Möbius transformation, not edges and
faces, which remain straight. The simple proof for the case of triangulations [5]
carries over without change.

2.5 Triangulations: Characterization by Length Cross-Ratios

For euclidean triangulations, there is an alternative characterization of conformal

equivalence in terms of length cross-ratios [5]. We review the basic facts in this
section.
For two adjacent triangles ijk ∈ F and jil ∈ F (see Fig. 3), the length cross-ratio
of the common interior edge ij ∈ E is defined as

il jk
lcr ij = . (9)
lj ki

(If the two triangles are embedded in the complex plane, this is just the modulus of
the complex cross-ratio of the four vertices.) This definition of length cross-ratios
8 A.I. Bobenko et al.

Fig. 3 Length cross-ratio j

 jk l j

k lcri j l

ki il

implicitly assumes that an orientation has been chosen on the surface. For non-
orientable surfaces, the length cross-ratio is well-defined on the oriented double
cover.
The product of length cross-ratios around an interior vertex i ∈ V is 1, because
all lengths cancel: 
lcr ij = 1. (10)
iji

Proposition 2.6 Two euclidean triangulations (, )euc and (, )˜ euc are discretely
conformally equivalent if and only if for each interior edge ij ∈ E int , the induced
length cross-ratios agree.
Remark 2.7 Analogous statements hold for spherical and hyperbolic triangulations.
Equation (9) has to be modified by replacing  with sin 2 or sinh 2 , respectively
(compare Remark 2.3).

2.6 Triangulations: Reconstructing Lengths from Length

Cross-Ratios

To deal with Riemann surfaces that are given in terms of Schottky data (Sect. 8.2) we
will need to reconstruct a function  : E  → R>0 satisfying (9) from given length
cross-ratios. (It is not required that the function  satisfies the triangle inequalities.)
To this end, we define auxiliary quantities cjki attached to the angles of the triangu-
lation. The value at vertex i of the triangle ijk ∈ F is defined as

jk
cjki = . (11)
ij ki

Then (9) is equivalent to

cjki
lcr ij = . (12)
clji

Now, given a function lcr : E int → R>0 defined on the set of interior edges E int and
satisfying the product condition (10) around interior vertices, one can find parame-
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 9

ters cjki satisfying (11) by choosing one value at each vertex and then successively
multiplying length cross-ratios. The corresponding function  is then determined by

1 1
ij = = . (13)
j j
cjki cki clji cil

2.7 Bipartite Graphs: Characterization by Length

Multi-Ratios

A different characterization of discrete conformal equivalence in terms of length

multi-ratios holds if the 1-skeleton of the polyhedral surface is bipartite, i.e., if the
vertices can be colored with two colors so that no two neighboring vertices share
the same color.

Proposition 2.8 (i) If two combinatorially equivalent euclidean cyclic polyhedral

˜ euc with discrete metrics  and ˜ are discretely confor-
surfaces (, )euc and (, )
mally equivalent, then the length multi-ratios for even cycles

i 1 i 2 , i 2 i 3 , . . . , i 2n i 1

are equal:
i1 i2 i3 i4 · · · i2n−1 i2n ˜i i ˜i i · · · ˜i2n−1 i2n
= 12 34 . (14)
i2 i3 i4 i5 · · · i2n i1 ˜i2 i3 ˜i4 i5 · · · ˜i2n i1

(ii) If the 1-skeleton of  is bipartite, i.e., if all cycles are even, then this con-
dition is also sufficient: If the length multi-ratios are equal for all cycles, then the
polyhedral surfaces are discretely conformally equivalent.

Proof (i) This is obvious, because all scale factors eu cancel. (ii) It is easy to see that
Eq. (3) can be solved for the scale factors eu/2 if the length multi-ratios are equal.
Note that the scale factors are not uniquely determined: they can be multiplied by λ
and 1/λ on the two vertex color classes, respectively. To find a particular solution,
one can fix the value of eu/2 at one vertex, and find the other values by alternatingly
dividing and multiplying by / ˜ along paths. The equality of length multi-ratios
implies that the obtained values do not depend on the path. 

Remark 2.9 If a polyhedral surface is simply connected, then its 1-skeleton is bipar-
tite if and only if all faces are even polygons. If a polyhedral surface is not simply
connected, then having even faces is only a necessary condition for being bipartite.

A polyhedral surface with bipartite 1-skeleton has even faces. If a polyhedral

surface has even faces and is simply connected, then its 1-skeleton is bipartite, and
the face boundaries generate all cycles. Thus, Proposition 2.8 implies the following
corollary.
10 A.I. Bobenko et al.

Corollary 2.10 Two simply connected combinatorially equivalent euclidean cyclic

polyhedral surfaces with even faces and with discrete metrics  and ˜ are discretely
conformally equivalent if and only if the multi-ratio condition (14) holds for every
face boundary cycle.

Remark 2.11 Analogous statements hold for spherical and hyperbolic cyclic poly-
hedral surfaces. In the multi-ratio condition, one has to replace non-euclidean
lengths  with sin 2 or sinh 2 , respectively (compare Remark 2.3).

2.8 Quadrangulations: The Cross-Ratio System on

Quad-Graphs

The case of cyclic quadrilateral faces is somewhat special (and we will return to it
in Sect. 4), because equal length cross-ratio implies equal complex cross-ratio:

Proposition 2.12 If two euclidean polyhedral surfaces with cyclic quadrilateral

faces are discretely conformally equivalent, then corresponding faces ijkl ∈ F have
the same complex cross-ratio (when embedded in the complex plane):

(z i − z j )(z k − zl ) (z̃ i − z̃ j )(z̃ k − z̃l )

=
(z j − z k )(zl − z i ) (z̃ j − z̃ k )(z̃l − z̃ i )

Proof This follows immediately from Proposition 2.8: The length multi-ratio of a
quadrilateral is the modulus of the complex cross-ratio. If the (embedded) quadri-
laterals are cyclic, then their complex cross-ratios are real and negative, so their
arguments are also equal. 

For planar polyhedral surfaces, i.e., for quadrangulations in the complex plane,
Proposition 2.12 connects discrete conformality with the cross-ratio system on
quad-graphs. A quad-graph in the most general sense is simply an abstract CW
cell decomposition of a surface with quadrilateral faces. Often, more conditions are
added to the definition as needed. Here, we will require that the surface is oriented
and that the vertices are bicolored black and white. For simplicity, we will also
assume that the CW complex is strongly regular (see Sect. 2.2). The cross-ratio sys-
tem on a quad-graph  imposes equations (15) on variables z i that are attached to
the vertices i ∈ V . There is one equation per face ijkl ∈ F :

(z i − z j )(z k − zl )
= Q ijkl , (15)
(z j − z k )(zl − z i )

where we assume that i is a black vertex and the boundary vertices ijkl are listed in
the positive cyclic order. (Here we need the orientation). On the right hand side of
the equation, Q : F → C \ {0, 1} is a given function. In particular, it is required
that the values z i , z j , z k , zl on a face are distinct.
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 11

By Proposition 2.12, two discretely conformally equivalent planar quadrangula-

tions correspond to two solutions of the cross-ratio system on the same quad-graph
with the same cross-ratios Q. The following proposition says that in the simply con-
nected case, one can find complex factors w on the vertices whose absolute values
|w| = eu/2 govern the length change of edges according to (3), and whose arguments
govern the rotation of edges. Note that (3) is obtained from (16) by taking absolute
values.
Proposition 2.13 Let  be a simply connected quad-graph. Two functions z, z̃ :
V → C are solutions of the cross-ratio system on  with the same cross-ratios Q
if and only if there is a function w : V → C such that for all edges ij ∈ E 

z̃ j − z̃ i = wi w j (z j − z i ). (16)

Proof As in the proof of Proposition 2.8, it is easy to see that the system of equa-
tions (16) is solvable for w if and only if the complex multi-ratios for even cycles
are equal. Because  is simply connected, this is the case if and only if the complex
cross-ratios of corresponding faces are equal. 
Remark 2.14 The cross-ratio system on quad-graphs (15) is an integrable system (in
the sense of 3D consistency [6, 7]) if the cross-ratios Q “factor”, i.e., if there exists
a function on the set of edges, a : E  → C, that satisfies the following conditions
for each quadrilateral ijkl ∈ F:
(i) It takes the same value on opposite edges,

aij = akl , ajk = ali . (17)

(ii)
aij
Q ijkl = . (18)
ajk

In Adler et al. classification of integrable equations on quad-graphs [2], the inte-

grable cross-ratio system is called (Q1)δ=0 . It is also known as the discrete
Schwarzian Korteweg–de Vries (dSKdV) equation, especially when it is considered
on the regular square lattice [33] with constant cross-ratios.
If the cross-ratios Q have unit modulus, the cross-ratio system on quad-graphs is
connected with circle patterns with prescribed intersection angles [6, 7].
Remark 2.15 The system of equations (16) is also connected with an integrable
system on quad-graphs. Let bij = z j − z i , so b is a function on the oriented edges
with bij = −bji . Let us also assume that the quad-graph  is simply connected. Then
the system (16) defines a function z : V → C (uniquely up to an additive constant)
if and only if the complex scale factors w : V → C satisfy, for each face ijkl ∈ F
the closure condition

bij wi w j + bjk w j wk + bkl wk wl + bli wl wi = 0. (19)

12 A.I. Bobenko et al.

This system for w is integrable if, for each face ijkl ∈ F,

bij + bkl = 0 and bjk + bli = 0.

In this case, (19) is known as discrete modified Korteweg–de Vries (dmKdV) equa-
tion [33], or as Hirota equation [6, 7].

3 Variational Principles for Discrete Conformal Maps

3.1 Discrete Conformal Mapping Problems

We will consider the following discrete conformal mapping problems. (The notation
(, )g was introduced in Definition 2.1.)

Problem 3.1 (prescribed angle sums) Given

• A euclidean, spherical, or hyperbolic cyclic polyhedral surface (, )g , where
g ∈ {euc, hyp, sph},
• a desired total angle Θi > 0 for each vertex i ∈ V ,
• a choice of geometry g̃ ∈ {euc, hyp, sph},
˜ g̃ of geom-
find a discretely conformally equivalent cyclic polyhedral surface (, )
etry g̃ that has the desired total angles Θ around vertices.

For interior vertices, Θ prescribes a desired cone angle. For boundary vertices,
Θ prescribes a desired interior angle of the polygonal boundary. If Θi = 2π for all
interior vertices i, then Problem 3.1 asks for a flat metric in the discrete conformal
class, with prescribed boundary angles if the surface has a boundary.
More generally, we will consider the following problem, where the logarithmic
scale factors u (see Definition 2.2) are fixed at some vertices and desired angle sums
Θ are prescribed at the other vertices. The problems to find discrete Riemann maps
(Sect. 4.2) and maps onto the sphere (Sect. 6.1) can be reduced to this mapping
problem with some fixed scale factors.

Problem 3.2 (prescribed scale factors and angle sums) Given

• A euclidean, spherical, or hyperbolic cyclic polyhedral surface (, )g , where
g ∈ {euc, hyp, sph},
˙ 1
• a partition V = V0 ∪V
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 13

• a prescribed angle Θi > 0 for each vertex i ∈ V1 ,

• a prescribed logarithmic scale factor u i ∈ R for each vertex i ∈ V0 ,
• a choice of geometry g̃ ∈ {euc, hyp, sph},
˜ g̃ of geom-
find a discretely conformally equivalent cyclic polyhedral surface (, )
etry g̃ that has the desired total angles Θ around vertices in V1 and the fixed scale
factors u at vertices in V0 .

Note that for V0 = ∅, V1 = V , Problem 3.2 reduces to Problem 3.1.

3.2 Analytic Formulation of the Mapping Problems

We rephrase the mapping Problem 3.2 analytically as Problem 3.4. The sides of a
cyclic polygon determine its angles, but practical explicit equations for the angles
as functions of the sides exist only for triangles, e.g., (21). For this reason it makes
sense to triangulate the polyhedral surface. For the angles in a triangulation, we use
the notation shown in Fig. 4. In triangle ijk, we denote the angle at vertex i by αjki .
We denote by βiji the angle between the circumcircle and the edge jk. The angles α
and β are related by
j
αjki + βki + βijk = π,

so betas determine alphas and vice versa:

j
2βjki = π + αjki − αki − αijk , . . . (20)

For euclidean triangles,

j
αjki + αki + αijk = π, βjki = αjki .

Fig. 4 Notation of lengths

and angles in a triangle
ijk ∈ F
14 A.I. Bobenko et al.

The half-angle equation can be used to express the angles as functions of lengths:
⎧ 1
⎪
⎪ (−ij + jk + ki )(ij + jk − ki ) 2
⎪
⎪ (euc)
⎪
⎪ (ij − jk + ki )(ij + jk + ki )

i  ⎪⎪
⎪
     21
αjk ⎨ sinh (ij − jk + ki )/2 sinh (ij + jk − ki )/2
tan =     (hyp)
2 ⎪
⎪ sinh (−ij + jk + ki )/2 sinh (ij + jk + ki )/2
⎪

⎪      21
⎪
⎪
⎪
⎪ sin (ij − jk + ki )/2 sin (ij + jk − ki )/2
⎪
⎩     (sph)
sin (−ij + jk + ki )/2 sin (ij + jk + ki )/2
(21)

Lemma 3.3 (analytic formulation of Problem 3.2) Let

• the polyhedral surface (, )g ,
• ˙ 1,
the partition V0 ∪V
• Θi for i ∈ V1 ,
• u i for i ∈ V0 ,
• the geometry g̃ ∈ {euc, hyp, sph}
be given as in Problem 3.2. Let Δ be an abstract triangulation obtained by adding
non-crossing diagonals to non-triangular faces of . (So V = VΔ , E  ⊆ E Δ , and
the set of added diagonals is E Δ \ E  .) For ij ∈ E  , define λij by
⎧
⎪
⎨2 log ij if g = euc
λij = 2 log sinh 2ij if g = hyp (22)
⎪
⎩ 
2 log sin 2ij if g = sph

Then solving Problem 3.2 is equivalent to solving Problem 3.4 with E 0 = E  and
E1 = EΔ \ E .

Problem 3.4 Given

• an abstract triangulation Δ,
• a partition VΔ = V0 ∪V˙ 1,
• u i ∈ R for i ∈ V0
• Θi ∈ R>0 for i ∈ V1 ,
• ˙ 1,
a partition E Δ = E 0 ∪E
• λij for ij ∈ E 0 ,
• g̃ ∈ {euc, hyp, sph},
find u i ∈ R for i ∈ V1 and λij for ij ∈ E 1 such that

˜ : E Δ → R>0
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 15

defined by
λ̃ij = u i + u j + λij , (23)

and ⎧ 1
⎪ λ̃ij
⎨e 2 if g̃ = euc
˜ij = 2 arsinh e 2 λ̃ij
1
if g̃ = hyp (24)
⎪
⎩
2 arcsin e 2 λ̃ij
1
if g̃ = sph

satisfies for all ijk ∈ FΔ the triangle inequalities

˜ij < ˜jk + ˜ki , ˜jk < ˜ki + ˜ij , ˜ki < ˜ij + ˜jk , (25)

and for g̃ = sph also

˜ij + ˜jk + ˜ki < 2π, (26)

and such that


α̃jki = Θi for all i ∈ V1 , (27)
jk:ijk∈FΔ

β̃ijk + β̃jil = π for all ij ∈ E 1 , (28)

˜ Note
where α̃ and β̃ are defined by (21) and (20) (with α, β,  replaced by α̃, β̃, ).
˜
that for g̃ = sph it is also required that λ̃ < 0 for  to be well-defined.

Proof (of Lemma 3.3) Note that (27) says that the angle sums at vertices in V1 have
˜ g̃ belonging
the prescribed values, and (28) says that neighboring triangles of (Δ, )
to the same face of  share the same circumcircle. So deleting the edges in E Δ \ E  ,
˜ E )g̃ . 
one obtains a cyclic polyhedral surface (, |

3.3 Variational Principles

VΔ
Definition 3.5 For an abstract triangulation Δ and a function Θ ∈ R>0 , define the
three functions
hyp sph
euc
E Δ,Θ , E Δ,Θ , E Δ,Θ : R EΔ × RVΔ −→ R,
g̃
(λ, u) −→ E Δ,Θ (λ, u)

by

g̃
  π  
E Δ,Θ (λ, u) = f g̃ (λ̃ij , λ̃jk , λ̃ki ) − (λ̃jk + λ̃ki + λ̃ij ) + Θi u i , (29)
ijk∈FΔ
2 i∈V Δ
16 A.I. Bobenko et al.

where g̃ ∈ {euc, hyp, sph}, λ̃ is defined as function of λ and u by (23), and the
functions f euc , f hyp , f sph are defined in Sect. 3.4.

We will often omit the subscripts and write simply E euc , E hyp , E sph when this is
unlikely to cause confusion.
g̃
Definition 3.6 We define the feasible regions of the functions E Δ,Θ as the follow-
ing open subsets of their domains:
• The feasible region of E euc and E hyp is the set of all (λ, u) ∈ R EΔ × RVΔ such
that ˜ ∈ R>0
E
defined by (23) and (24) satisfies the triangle inequalities (25)
• The feasible region of E sph is the set of all (λ, u) ∈ R EΔ × RVΔ such that λ̃ defined
˜ which is then well-defined by (24), satisfies the triangle
by (23) is negative, and ,
inequalities (25) and the inequalities (26).

Theorem 3.7 (Variational principles) Every solution (, ) ˜ g̃ of Problem 3.2 cor-
responds via (23) and (24) to a critical point (λ, u) ∈ R × RVΔ of the function
EΔ
g̃
E Δ,Θ under the constraints that λij and u i are fixed for ij ∈ E 0 and i ∈ V0 , respec-
tively. (The triangulation Δ, and E 0 = E  and E 1 = E Δ \ E  are as in Lemma 3.3,
and the given function Θ is extended from V1 to V by arbitrary values on V0 .)
g̃
Conversely, if (λ, u) ∈ R EΔ × RVΔ is a critical point of the function E Δ,Θ under
g̃
the same constraints, and if (λ, u) is contained in the feasible region of E Δ,Θ , then
(, ) ˜ g̃ defined by (23) and (24) is a solution of Problem 3.2.

Proof This follows from the analytic formulation of Problem 3.2 (see Sect. 3.2) and
Proposition 3.8. 

Proposition 3.8 (First derivative of E g̃ ) The partial derivatives of E g̃ are

∂ E g̃ 
(λ, u) = Θi − α̃jki (30)
∂u i ijki

∂ E g̃
(λ, u) = β̃ijk + β̃ijl − π. (31)
∂λij

˜ if (λ, u)
Here α̃, β̃ are defined by (21) and (20) (with α, β,  replaced by α̃, β̃, )
is contained in the feasible region of E g̃ . For (λ, u) not contained in the feasible
region, the definition of α̃, β̃ is extended like in Definition 3.12.

Proof Equations (30) and (31) follow from the definition of E g̃ and Proposition 3.14
on the partial derivatives of f g . 

Theorem 3.9 (Uniqueness for mapping problems) If Problem 3.2 with target
geometry g̃ ∈ {euc, hyp} has a solution, then the solution is unique—except if
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 17

g̃ = euc and V0 = ∅ (the case of Problem 3.1). In this case, the solution is unique
up to scale.
The critical point (λ, u) ∈ R EΔ × RVΔ that corresponds, via (23) and (24), to a
solution (, )˜ g̃ of Problem 3.2 with g̃ ∈ {euc, hyp} is a minimizer of E Δ,Θ
g̃
under
the constraints described in Theorem 3.7. The minimizer is unique except in the
g̃
following cases. If g̃ = euc and V0 = ∅, then E Δ,Θ is constant along all lines in the
“scaling direction” (0, 1VΔ ) ∈ R EΔ × RVΔ . If the 1-skeleton of  is bipartite and
g̃
V0 = ∅, then E Δ,Θ is constant in the direction that is ±1 on the two color classes
of VΔ , respectively, and takes appropriate values on E Δ \ E  so that λ̃ij defined
by (23) remains constant for all ij ∈ E Δ . (In both exceptional cases, one can obtain
a unique minimizer by adding the constraint of fixing u i for some i ∈ VΔ .)

Proof The theorem follows from Theorem 3.7 and the following observations.

(1) If the point (λ, u) ∈ R EΔ × RVΔ corresponds to a solution of Problem 3.2, it is

g̃
contained in the feasible region of E Δ,Θ .
(2) By (29) and Proposition 3.16, the functions E euc and E hyp are convex.
(3) For (λ, u) in the feasible region, the second derivative D 2 E hyp (λ, u) is a posi-
tive definite quadratic form of d λ̃, i.e., D 2 E hyp (λ, u)(λ̇, u̇) ≥ 0 for all (λ̇, u̇) ∈
R EΔ × RVΔ and D 2 E hyp (λ, u)(λ̇, u̇) = 0 if and only if

λ̇ij + u̇ i + u̇ j = 0 for all ij ∈ E Δ .

(4) Similarly, for (λ, u) in the feasible region, the second derivative D 2 E euc (λ, u) is
a positive semidefinite quadratic form with D 2 E euc (λ, u)(λ̇, u̇) = 0 if and only
if
λ̇ij + u̇ i + u̇ j = c for all ij ∈ E Δ , for some c ∈ R.


In the following proposition, we collect explicit formulas for the second deriv-
atives of the functions E g̃ . They are useful for the numerical minimization of E euc
and E hyp , and even for finding critical points of E sph , as explained in Sect. 6.2.

Proposition 3.10 (Second derivative of E g̃ ) The second derivatives of E euc , E hyp ,

and E sph are the quadratic forms

1   k j 
D 2 E g̃ (λ, u) = qij (λ, u) + qjki (λ, u) + qki (λ, u) ,
2 ijk∈F
Δ

where qijk (λ, u) = 0 if ˜ij , ˜jk , ˜ki defined by (23), (24) violate the triangle inequali-
ties (25), or, in the case of g̃ = sph, inequality (26). Otherwise, the quadratic forms
qijk (λ, u) are defined by
18 A.I. Bobenko et al.
⎧
⎪
⎪ cot α̃ijk (dλki − dλjk + du i − du j )2 (euc)
⎪
⎨   ˜  
qijk = cot β̃ijk (dλik − dλkj + du i − du j )2 + tanh2 2ij (dλij + du i + du j )2 (hyp)
⎪
⎪
⎩cot β̃ k (dλ − dλ + du − du )2 − tan2  ˜ij (dλ + du + du )2 
⎪
(sph)
ij ik kj i j 2 ij i j

˜
where α̃, β̃ are defined by (21) and (20) (with α, β,  replaced by α̃, β̃, ).

Proposition 3.10 follows from (29) and Proposition 3.15 about the second deriv-
atives of f g .

3.4 The Triangle Functions

This section is concerned with three real valued functions f euc , f hyp , f sph of three
variables that are the main building blocks for the action functions E euc , E hyp , E sph
of the variational principles. Since we consider single triangles in this section, not
triangulations, we can use simpler notation. For {i, j, k} = {1, 2, 3}, let

λi = λjk , i = jk , αi = αjki , βi = βjki .

The terminology introduced in the following definition makes Definition 3.12 easier
to state.

Definition 3.11 Let the feasible region of f euc and f hyp be the open subset of all
λ ∈ R3 such that  ∈ R3>0 determined by (22) satisfies the triangle inequalities, i.e.,

k < i +  j (32)

for {i, j, k} = {1, 2, 3}.

Let the feasible region of f sph be the open subset of all λ ∈ R3 such that λ < 0,
and such that  ∈ R3>0 , which is then well-defined by (22), satisfies the triangle
inequalities (32) and
1 + 2 + 3 < 2π. (33)

Definition 3.12 We define the three functions

f euc , f hyp , f sph : R3 → R

by

f g (λ1 , λ2 , λ3 ) = β1 λ1 + β2 λ2 + β3 λ3 + L(α1 ) + L(α2 ) + L(α3 )

 
+ L(β1 ) + L(β2 ) + L(β3 ) + L 21 (π − α1 − α2 − α3 ) , (34)

where g ∈ {euc, hyp, sph}, L(x) denotes Milnor’s Lobachevsky function [30]
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 19
 x  
L(x) = − log 2 sin(t) dt, (35)
0

and,
• if λ is in the feasible region of f g , then the angles α, β are defined as the angles
(shown in Fig. 4) in a euclidean, hyperbolic, or spherical triangle (depending on
g) with sides 1 , 2 , 3 determined by (22). That is, α and β are defined by (21)
and (20).
• Otherwise, if g = sph, and if either at least two λs are non-negative or λ < 0 and
inequality (33) is violated, let

αk = αi = α j = π, βk = βi = β j = 0.

• Otherwise, if the triangle inequality (32) is violated, or if g = sph and λk ≥ 0, let

αk = βk = π, αi = α j = βi = β j = 0.

Figure 5 shows a graph of Milnor’s Lobachevsky function. It is continuous, π -

periodic, odd, has zeros at the integer multiples of π/2, and is real analytic except
at integer multiples of π , where the derivative tends to +∞.

Remark 3.13 In the euclidean case, (34) simplifies to

f euc (λ) = αi λi + α j λ j + αk λk + 2L(αi ) + 2L(α j ) + 2L(αk ). (36)

This follows immediately from α1 + α2 + α3 = π , α = β, and L(0) = 0.

Proposition 3.14 (first derivative) The functions f g , g ∈ {euc, hyp, sph}, are con-
tinuously differentiable and
∂f g
= βi . (37)
∂λi

Fig. 5 Graph of Milnor’s

Lobachevsky function,
y = L (x)
20 A.I. Bobenko et al.

Proof Note that the angles α, β are continuous functions of λ on R3 . Hence f g

defined by (34) is also continuous. We will show that f g is continuously differ-
entiable with derivative (37) on an open dense subset of the domain, namely, the
union of (a) the feasible region and (b) the interior of its complement. Since f g is
continuous and df g extends continuously to R3 , the claim follows.
(a) First, suppose λ is contained in the feasible region of f g . By symmetry, it
suffices to consider the derivative with respect to λ1 . From (34) and (35) one obtains

∂f g  
3
 ∂βi
= β1 + λi − log(2 sin βi ) +
∂λ1 i=1
∂λ1
   ∂αi 
− log(2 sin αi ) + 21 log 2 sin( π−α1 −α2 −α3 
) (38)
2
∂λ1

For hyperbolic and spherical triangles, one derives from the respective cosine
rules

i sin βi sin π−α1 −α 2 −α3

sinh2 = 2
(hyperbolic),
2 sin α2 sin α3
i sin βi sin α1 +α2 2+α3 −π
sin2 = (spherical).
2 sin α2 sin α3

In both cases, expand the fraction on the right hand side by four and take logarithms
to find
 
λi = log(2 sin βi ) + log 2 sin π−α1 −α2 −α3 
2
− log(2 sin α j ) − log(2 sin αk ).

Substitute this expression for λi in (38) and use dβi = 21 (dαi − dα j − dαk ) to see
that all terms on the right hand side of (38) cancel, except β1 .
For euclidean triangles, (38) simplifies to

∂f g  3
 ∂αi
= β1 + λi − 2 log(2 sin αi ) ,
∂λ1 i=1
∂λ1

where
i
λi − 2 log(2 sin αi ) = 2 log = 2 log R
2 sin αi

does not depend on i. (R denotes the circumradius.) Equation (37) follows because
the angle sum is constant.
(b) Now suppose λ is contained in the interior of the complement of the feasible
region of f g . Since β1 , β2 , β3 are constant on each connected component of the
complement of the feasible region, and since
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 21

f g (λ1 , λ2 , λ3 ) = β1 λ1 + β2 λ2 + β3 λ3 ,

outside the feasible region, Eq. (37) holds also in this case. This completes the
proof. 

Proposition 3.15 (second derivative) For g ∈ {euc, hyp, sph} the function f g is
twice continuously differentiable on its feasible set and the second derivative is

1
3
D 2 f euc = cot αi (dλ j − dλk )2 , (39)
2 i=1

1    
3
D f2 hyp
= cot βi (dλ j − dλk )2 + tanh2 2i dλi2 , (40)
2 i=1

1    
3
D 2 f sph = cot βi (dλ j − dλk )2 − tan2 2i dλi2 . (41)
2 i=1

On each component of the complement of its feasible set, the function f g is linear
so the second derivative vanishes.

A proof of (39) is contained in [5] (Proposition 4.2.3), see Remark 3.17 below.
Equations (40) and (41) can be derived by lengthy calculations.

Proposition 3.16 (i) The function f euc is convex. On its feasible set, the second
derivative D 2 f euc is positive semidefinite with one-dimensional kernel spanned by
the “scaling direction” (1, 1, 1).
(ii) The function f hyp is convex. On its feasible set, the second derivative D 2 f hyp
is positive definite, so the functions is locally strictly convex.

Part (i) is proved in [5] (Propositions 4.2.4, 4.2.5, note the following remark)
directly from (39). We do not know a similarly straightforward proof of part (ii).
The proof in [5] (Sect. 6.2) is based on a connection with 3-dimensional hyperbolic
geometry: f hyp is the Legendre dual of the volume of an ideal hyperbolic prism
considered as a function of the dihedral angles. This volume function is strictly
concave, as shown by Leibon [26]. His argument uses the decomposition of an ideal
prism into three ideal tetrahedra.

Remark 3.17 The functions f and V̂h defined in [5] (equations (4-3), (6-4)) are
related to the functions f euc and f hyp by

f euc (λ1 , λ2 , λ3 ) = 2 f ( λ21 , λ22 , λ23 ), (42)

f hyp (λ1 , λ2 , λ3 ) = 2 V̂h (λ1 , λ2 , λ3 , 0, 0, 0). (43)
22 A.I. Bobenko et al.

4 Conformal Maps of Cyclic Quadrangulations

Having introduced the mapping problems and variational principles, we return to

conformal maps of cyclic quadrangulations. Some basic facts were already dis-
cussed in Sect. 2.8. Here, in Sect. 4.1, we consider a simple experiment that demon-
strates the somewhat unexpected appearance of orthogonal circle patterns, and also
a necessary condition for the boundary angles. In Sect. 4.2, we discuss a discrete
version of the Riemann mapping problem for quadrangulations.

4.1 Emerging Circle Patterns and a Necessary Condition

Consider the two discrete conformal maps shown in the two rows of Fig. 6. The
domains (shown left) are a square and a rectangle, subdivided into 6 × 6 and 6 × 5
squares, respectively. We solve the mapping Problem 3.1 by minimizing E euc as
explained in Sect. 3.3, prescribing boundary angles to obtain maps to parallelo-
grams: Θ = 50◦ and 130◦ for the corner vertices, Θ = 180◦ for the other bound-
ary vertices, and Θ = 360◦ for interior vertices. The resulting quadrangulations are
shown in the middle.
On first sight, the 6 × 6 example shown in the top row behaves rather like one
would expect from a conformal map. The horizontal and vertical “coordinate lines”
of the domain are mapped to polygonal curves that look more or less like they could
be discretizations of reasonable smooth curves. In the 6 × 5 example shown in the
bottom row, the images of the vertical lines zigzag noticeably.
A closer look at the 6 × 6 example reveals a remarkable phenomenon. Let us
bicolor the vertices black and white so that neighboring vertices have different

Fig. 6 Mapping a rectangle to a parallelogram. Note the orthogonal circle pattern in the top row
and the wiggly vertical lines in the bottom row
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 23

colors, with the corners colored white. Then, in the image quadrangulation, the
edges incident with a black vertex meet at right angles, and the edges incident with
a white vertex have the same length. One can therefore draw a circle around each
white vertex through the neighboring black vertices as shown in Fig. 6 (top right).
At the black vertices, these circles touch and intersect orthogonally. Such circle pat-
terns were studied by Schramm [38] as discrete analogs of conformal maps.
Given such a circle pattern with orthogonally intersecting circles, the quadran-
gulation formed by drawing edges between circle centers and intersection points
consists of quadrilaterals that are right-angled kites. Such kites have complex cross-
ratio −1. Hence, the quadrangulation coming from an orthogonal circle pattern
is discretely conformally equivalent (in our sense) to a combinatorially equivalent
quadrangulation consisting of squares.
The conformal map shown in the top row of Fig. 6 “finds” the orthogonal cir-
cle pattern because that circle pattern exists and the conformal map is unique (by
Theorem 3.9). For the 6 × 5 example shown in the bottom row, a corresponding
orthogonal circle pattern does not exist. No matter which coloring is chosen, there
are two black vertices at which the total angle changes (from 90◦ to 50◦ and 130◦ ,
respectively). The neighbors of a vertex do not lie on a circle. Figure 6 (bottom right)
shows two circles drawn through three out of four neighbors.
If we map an m × n square grid to a parallelogram like in Fig. 6, an orthogonal
circle pattern will appear if m an n are even. No such pattern will appear if one of
the numbers is even and the other is odd. What happens if both m and n are odd?
In this case, the conformal map does not exist. The corners with increasing angle
and the corners with decreasing angle would have different colors. This violates the
necessary condition expressed in the following theorem.

Theorem 4.1 (Necessary condition for the existence of a conformal map) Let  be
an abstract quadrangulation of the closed disk, and let

z, z̃ : V → C

determine two discretely conformally equivalent immersions of  into the complex

plane. Denote their angle sums at boundary vertices v ∈ V by Θv and Θ̃v , respec-
tively. Since the 1-skeleton of  is bipartite, we may assume the vertices are colored
black and white. Let Vb∂ and Vw∂ denote the sets of black and white boundary vertices
of . Then

(Θ̃v − Θv ) ≡ 0 (mod 2π ), (44)
v∈Vb∂

(Θ̃v − Θv ) ≡ 0 (mod 2π ). (45)
v∈Vw∂


(Since Vb∂ ∪Vw∂ (Θ̃v − Θv ) = 0, equations (45) and (44) are equivalent.)
24 A.I. Bobenko et al.

Proof Since z and z̃ are two solutions of the cross-ratio system on  with the same
cross-ratios (see Sect. 2.8), there exists by Proposition 2.13 a function w : V → C
such that (16) holds for all edges ij ∈ E  . Now suppose v0 , . . . , v2n−1 ∈ V are the
boundary vertices in cyclic order (with indices taken modulo 2n). Then

(z̃ vk+1 − z̃ vk )(z vk−1 − z vk ) wv

ei(Θ̃vk −Θvk ) = = k+1 ,
(z̃ vk−1 − z̃ vk )(z vk+1 − z vk ) wvk−1

so

n−1 
n−1
ei(Θ̃v2k −Θv2k ) = ei(Θ̃v2k+1 −Θv2k+1 ) = 1.
k=0 k=0

Equations (44) and (45) follow. 

4.2 Riemann Maps with Cyclic Quadrilaterals

Consider the following discrete version of the Riemann mapping problem: Map a
cyclic polyhedral surface that is topologically a closed disk discretely conformally to
a planar polygonal region with boundary vertices on a circle. An example is shown
in Fig. 7, top row. This type of problem can often be reduced to Problem 3.2. Then,
by the variational principle, if a solution exists, it can be found by minimizing a
convex function. For triangulations, the reduction of the discrete Riemann mapping
problem to Problem 3.2 is explained in [5] (Sect. 3.3). Here, we consider the case
of quadrangulations. (The arguments can be extended to even polygons with more
than four sides. We restrict our attention to quadrilaterals because the combinatorial
restrictions discussed in the following paragraph become even more involved for
surfaces with hexagons, octagons, etc.)
The basic idea is the same as for triangulations: First, map the polyhedral sur-
face to the half plane with one boundary vertex at infinity. Then apply a Möbius
transformation. This leads to a combinatorial restriction: No face may have more
than one edge on the boundary. (The face would degenerate when the boundary is
mapped to a straight line.) For triangulations, this means that no triangle may be
connected to the surface by only one edge. If this condition is violated, cutting off
such “ears” often leads to an admissible triangulation. For quadrangulations, this
fix does not work in typical situations. Instead, if a quadrilateral contains two con-
secutive edges on the boundary, cut off a triangle. The resulting polyhedral surface
will consist mostly of quadrilaterals with some triangles on the boundary, as in the
example shown in Figs. 7, 8.
Suppose (, )euc is a euclidean cyclic polyhedral surface that is homeomorphic
to the closed disk and consists mostly of quadrilaterals. (For the following construc-
tion we really only need a boundary vertex that is incident with quadrilateral faces.)
To map it to a polygonal region inscribed in a circle, proceed as follows (see Fig. 7):
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 25

Fig. 7 Riemann mapping with cyclic quadrilaterals

(1) Choose a vertex k on the boundary of  such that all incident faces are quadri-
laterals.
(2) Apply a discrete conformal change of metric (3) such that all edges incident
with k have the same length. One may choose u = 0 for all vertices except the
neighbors of k. It does not matter if polygon inequalities are violated after this
step.
(3) Let (  ,  )euc be the cyclic polyhedral complex obtained by removing vertex k
and all incident quadrilaterals.
(4) Solve Problem 3.2 for (  ,  )euc with prescribed total angles Θi = 2π for inte-
rior vertices of   , Θi = π for boundary vertices of   that were not neigh-
bors of k in , and fixed logarithmic scale factors u i = 0 for those that were
neighbors of k. The result is a planar polyhedral surface as shown in Fig. 7, bot-
tom. The boundary consists of one straight line segment containing all bound-
ary edges of   that were also boundary edges of , and two or more straight
26 A.I. Bobenko et al.

Fig. 8 Here we show the

face circumcircles of the
solution to the Riemann
mapping problem of Fig. 7.
It looks conspicuously like
an orthogonal circle pattern.
But the face circumcircles
intersect only approximately
but not exactly at right
angles

line segments, each consisting of two edges that were incident with a removed
quadrilateral.
(5) Apply a Möbius transformation (e.g., z → 1/z) to the vertices that maps the
boundary vertices of  to a circle and the other vertices to the inside of this
circle. Reinsert k at the image point of ∞ under this Möbius transformation.
Each face ijmk ∈  incident with k is cyclic because the three vertices i, j, and
m are contained in a line before transformation.
(6) Optionally apply a 2-dimensional version of the Möbius normalization
described in Sect. 6.3.

Proposition 4.2 The result of this procedure is a planar cyclic polyhedral surface
that is discretely conformally equivalent to (, )euc and has its boundary polygon
inscribed in a circle.

Proof That the boundary polygon is inscribed in a circle is obvious from the con-
struction. Using the Möbius invariance of discrete conformal equivalence (Propo-
sition 2.5), it is not difficult to see that the surfaces without quadrilaterals incident
with k are discretely conformally equivalent. To show that the whole surfaces are
equivalent, it suffices to show that corresponding quadrilaterals incident with k have
the same complex cross-ratio.
After step (2), the length cross-ratio of a quadrilateral incident with k is equal to
the simple length ratio of the two edges that are not incident with k.
After step (4), the length cross-ratio of these edges is unchanged due to the fixed
logarithmic scale factors u = 0 on the neighbors of k. Also, these edges are now
collinear because of the prescribed angle Θ = π between them.
After applying the Möbius transformation in step (5), the image of the point at
infinity and the other three vertices of our quadrilateral incident with k form again a
cyclic quadrilateral with the same complex cross-ratio as in the beginning. 
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 27

Fig. 9 Discrete conformal map of a multiply-connected domain (left) to a circle domain (middle).
The images of vertical and horizontal “parameter lines” are shown on the right

5 Multiply Connected Domains

5.1 Circle Domains

Koebe’s generalization of the Riemann mapping theorem says that multiply con-
nected domains are conformally equivalent to domains bounded by circles, and the
uniformizing map to such a circle domain is unique up to Möbius transformations.
A method to construct discrete Riemann maps is described in [5] (Sect. 3.3) for
triangulations and for mostly quadrilateral meshes in the previous Sect. 4.2. Hav-
ing generalized the notion of discrete conformal equivalence from triangulations to
cyclic polyhedral surfaces, it is straightforward to adapt this method to construct
discrete maps to circle domains:
(1) Fill holes by gluing faces to all but one boundary component, so that the result-
ing surface is homeomorphic to a disk.
(2) Construct the discrete Riemann map.
(3) Remove the faces that were added in step (1).
Figure 9 shows an example.

5.2 Special Slit Domains

Any multiply connected domain can be mapped to the complex plain with parallel
slits [32]. In principle, it is possible to construct discrete conformal maps that map
holes to slits by solving Problem 3.1. On each boundary component that should be
mapped to a slit, set the desired total angle Θ = 2π for the two vertices that should
be mapped to the endpoints of the slit, and set Θ = π for all other vertices on that
boundary component. However, this will not work in general. While the resulting
surface will be flat, the developing map to the plane will in general have translational
28 A.I. Bobenko et al.

monodromy for a cycle around the hole. The surface will only close up in the plane
if the vertices that should be mapped to the endpoints of the slit are chosen exactly
right. (This will in general require modifying the original mesh.)
Sometimes, the symmetry of the problem determines the right positions of the
end-vertices, so that discrete conformal maps to slit surfaces can be computed. The
first two rows of Fig. 10 show examples. The bottom row visualizes a discrete con-

Fig. 10 Mapping surfaces with holes to slit surfaces. In all images, the left and right parts of the
boundary are identified by a horizontal translation. Preimages of horizontal lines visualize the flow
of an incompressible inviscid fluid around the hole in a channel with periodic boundary conditions.
Top row A cylinder with a triangular hole is mapped to a cylinder with a slit. One vertex of the
triangle and the midpoint of the opposite side are mapped to the endpoints of the slit. Middle row
An arrow shaped slit is mapped to a straight slit. The two vertices at the arrow’s tip, on either side
of the slit, are mapped to the endpoints of the straight slit. Bottom row Three circular boundary
components are mapped to horizontal slits (The slit surface is not shown.)
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 29

formal map where circular holes are mapped to slits. Here, we use the following
trick: We start with the slit surface and map it to a surface with circular holes as
described in Sect. 5.1.

6 Uniformization of Spheres

This section is concerned with discrete conformal maps of polyhedral surfaces of

genus 0 onto the round sphere. For triangulations, this is described in [5] (Sect. 3.2).
In Sect. 6.1, we adapt this method to quadrangulations. This is similar to the dis-
crete Riemann mapping with quadrilaterals described in Sect. 4.2. Effectively, this
method reduces the problem to minimizing the convex euclidean functional E euc .
The spherical version of the variational principle (Theorem 3.7) involves the non-
convex function E sph . It is not as practical for calculations, because one has to find a
saddle point instead of a minimum. Nevertheless, the spherical functional can often
be used to calculate maps to the sphere. This is explained in Sect. 6.2.

6.1 Uniformizing Quadrangulations of the Sphere

Suppose (, )euc is a cyclic polyhedral surface with quadrilateral faces that is
homeomorphic to the sphere.
(1) Choose a vertex k ∈ V .
(2) Apply a discrete conformal change of metric (3) such that all edges incident
with k have the same length. One may choose u = 0 for all vertices except the
neighbors of k. It does not matter if polygon inequalities are violated after this
step.
(3) Let (  ,  )euc be the complex obtained by removing vertex k and all incident
quadrilaterals.
(4) Solve Problem 3.2 for (  ,  )euc with prescribed total angles Θi = 2π for inte-
rior vertices of   , Θi = π for boundary vertices of   that were not neighbors
of k in , and fixed scale factors u i = 0 for vertices that were neighbors of k
in . The result is a planar polyhedral surface with cyclic quadrilaterals. Con-
secutive boundary edges that belonged to a face incident with vertex k in  are
contained in a straight line.
(5) Map the vertices to the unit sphere by stereographic projection and reinsert the
vertex k at the image point of ∞.
(6) Optionally apply Möbius normalization, see Sect. 6.3.

Proposition 6.1 The result is a cyclic polyhedral surface with vertices on the unit
sphere that is discretely conformally equivalent to (, )euc .
30 A.I. Bobenko et al.

Fig. 11 Discrete conformal map from the cube to the sphere, calculated with the method described
in Sect. 6.1. We apply Möbius normalization (Sect. 6.3) to the polyhedral surface with vertices on
the sphere to achieve rotational symmetry

This can be seen in the same way as the corresponding statement about discrete
Riemann maps with quadrilaterals (Proposition 4.2). Figure 11 shows a discrete con-
formal map calculated by this method.

6.2 Using the Spherical Functional

It is possible to use the spherical functional E sph to calculate maps to the sphere
even though it is not convex. For simplicity, we consider only triangulations, so
all λ variables are fixed and we may consider E sph as function of the logarithmic
scale factors u only (see Sect. 3.3). A numerical method has to find a saddle point
of E sph (u).
Note that the scaling direction 1VΔ ∈ RVΔ is a negative direction of the Hessian at
a critical point: Suppose (Δ, )sph is a spherical triangulation with the desired angle
sph
sum Θi at each vertex i. Then 0 ∈ RVΔ is a critical point of E Δ,Θ (u). If we enlarge
all edge lengths by a common factor eh > 1, then all angles become larger, so every
component (30) of the gradient of E sph becomes negative. Following the negative
gradient would result in even larger lengths.
The following minimax method works in many cases. Define the function Ẽ by
maximizing the functional E sph in the scaling direction,
 
Ẽ(u) = max E sph (u + h1VΔ ) . (46)
h∈R

Minimize functional Ẽ in a hyperplane of RVΔ transverse to the direction 1VΔ .

Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 31

Fig. 12 Mapping conformally to the sphere using the spherical functional. The spherical surfaces
are Möbius-normalized to achieve rotational symmetry

Figure 1 (top) and Fig. 12 show examples of discrete conformal maps to polyhe-
dral surfaces inscribed in a sphere that were calculated using this method.

6.3 Möbius Normalization

The notion of discrete conformal equivalence of euclidean polyhedral surfaces

(, )euc in R3 is Möbius invariant (Proposition 2.5). If all vertices v ∈ V are con-
tained in the unit sphere S 2 ⊂ R3 , then there is a Möbius transformation T of S 2
such that the center of mass of the transformed vertices is the origin [43],

T (v) = 0.
v∈V

The Möbius transformation T is uniquely determined up to post-composition with

a rotation around the origin.
The following method can be used to calculate such a Möbius transformation:
Find the unique minimizer of the function δ defined below. Then choose for T a
Möbius transformation that maps S 2 to itself and the minimizer to the origin. Here,
we only provide explicit formulas for the function δ and its first two derivatives. For
a more detailed account, we refer the reader to [43]. The function δ is defined on the
open unit ball in R3 by

  
−x, v
δ(x) = log √ , (47)
v∈V
−x, x

where
x, y = x1 y1 + x2 y2 + x3 y3 − 1. (48)
32 A.I. Bobenko et al.

The gradient and Hessian matrix of δ are

 v x

grad δ(x) = − , (49)
v∈V
x, v x, x
  xT x vT v

1

Hess δ(x) = 2 − − diag . (50)
v∈V
x, x2 x, v2 x, x

7 Uniformization of Tori

Every Riemann surface R of genus one is conformally equivalent to a flat torus, i.e.,
to a quotient space C/ , where  = Zω1 + Zω2 is some two-dimensional lattice
in C. The biholomorphic map from R to C/ , or from the universal cover of R to
C, is called a uniformizing map. For a polyhedral surface of genus one, construct-
ing a discrete uniformizing map amounts to solving Problem 3.1 with prescribed
total angle Θ = 2π at all vertices. This provides us with a method to calculate
approximate uniformizing maps for Riemann surfaces of genus one given in various
forms. We consider examples of tori immersed in R3 in Sect. 7.1 and elliptic curves
in Sect. 7.2. (We will also consider tori in the form of Schottky uniformization in
Sect. 8.2, as a toy example after treating the higher genus case.)
The belief that discrete conformal maps approximate conformal maps is not
based on a proven theorem but on experimental evidence like the Wente torus exam-
ple of Sect. 7.1 and the numerical experiments of Sect. 7.4.

7.1 Immersed Tori

First we consider a simple example with quadrilateral faces. Figure 13 (left) shows
a coarse discretization of a torus. The faces are isosceles trapezoids, so they are
inscribed in circles. On the right, the figure shows the uniformization obtained by
solving Problem 3.1 with prescribed total angle Θ = 2π at all vertices.
To test the numerical accuracy of our discrete uniformizing maps, we consider
the famous torus of constant mean curvature discovered by Wente [47]. Explicit
doubly periodic conformal immersion formulas (i.e., formulas for the inverse of a
uniformizing map) are known in terms of elliptic functions [1, 3, 46].
Figure 14 (left) shows a triangulated model of the Wente torus constructed by
sampling an explicit immersion formula [3] on a nearly square lattice containing
the period lattice . On the right, the figure shows the discrete uniformization,
which reproduces the regular lattice  to high accuracy. The modulus τ = ω2 /ω1
of the Wente torus has been determined numerically [19] as τ = 0.41300 . . . +
i 0.91073 . . . . The modulus of the discrete uniformization of the discretized sur-
face shown in the figure is τ̃ = 0.41341 . . . + i 0.91061 . . . .
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 33

Fig. 13 Uniformization of an immersed torus with cyclic quadrilateral faces

Fig. 14 Uniformization of the Wente torus

7.2 Elliptic Curves

An algebraic curve of the form


k
μ2 = a (λ − λ j ), (51)
j=1

where the λ j ∈ C are distinct and k = 3 (an elliptic curve) or k = 4 (with the singu-
larity at infinity resolved), represents a Riemann surface of genus one as branched
double cover of the λ-sphere CP1 , which we identify conformally with the unit
sphere S 2 ⊂ R3 . The branch points are λ1 , λ2 , λ3 , ∞ if k = 3 and λ1 , λ2 , λ3 , λ4 if
k = 4. Every Riemann surface of genus one can be represented in this way.
We construct a discrete model for a double cover of S 2 branched at four points
λ1 , . . . , λ4 in the following way. Choose n other points p1 , . . . , pn ∈ S 2 and let P
be the boundary of the convex hull of the points {λ1 , . . . , λ4 , p1 , . . . , pn }. Then
P is a convex polyhedron with n + 4 vertices and with faces inscribed in circles.
(Generically, the faces will be triangles. In Sect. 7.3 we explain the method we used
to obtain “good” triangles.) Find two disjoint simple edge paths γ1 , γ2 joining the
branch points λ j in pairs. Take a second copy P̂ of the polyhedron P. Cut and glue P
and P̂ along the paths γ1 , γ2 to obtain a polyhedral surface of genus 1. Uniformize it
34 A.I. Bobenko et al.

Fig. 15 Discrete uniformization of elliptic curves. Left If the branch points in S 2 are the vertices
of a regular tetrahedron, period lattice is very close to a hexagonal lattice. Middle If the branch
points form a square on the equator, the period lattice is very close to a square lattice. Right an
example with branch points in unsymmetric position

by solving Problem 3.1. One obtains a discrete conformal map to a flat torus, whose
inverse can be seen as a discrete elliptic function. Figure 15 shows examples. We
will treat hyperelliptic curves in a similar fashion in Sect. 8.3.
Remark 7.1 Instead of constructing a doubly covered convex euclidean polyhedron
with vertices on the unit sphere as described above, one could also construct a spher-
ical triangulation of the doubly covered sphere that is invariant under the elliptic
involution (exchanging sheets). These two approaches are in fact equivalent due to
Remark 2.3.
Mapping a flat torus to an elliptic curve. We can also go the opposite way, map-
ping a flat torus to a double cover of S 2 . Start with a triangulated flat torus. The
triangulation should be symmetric with respect to the elliptic involution, i.e., sym-
metric with respect to a half turn around one vertex (which is then also a half turn
around three other vertices). The quotient space of the triangulated torus modulo the
elliptic involution is then a triangulated sphere. Map it to the sphere by the procedure
explained in [5] (Sect. 3.2), see also Sect. 6.1 of the present article.
Figures 16 and 17 show examples where we started with a hexagonal and a square
torus respectively.

√
Fig. 16 Mapping the hexagonal torus C/(Z + τ Z), τ = 21 + i 23 (left) to a double cover of the
sphere (right). Because the regular triangulation of the torus on the left is symmetric with respect
to the elliptic involution, its image projects to a triangulation of the sphere seen on the right
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 35

Fig. 17 Mapping the square torus C/(Z + i Z) (left) to a double cover of the sphere (right). Again,
the triangulation on the left is symmetric with respect to the elliptic involution, so the image on the
right projects to a triangulation of the sphere

7.3 Choosing Points on the Sphere

The uniformization procedure for elliptic curves described in Sect. 7.2 requires
choosing points on the sphere in addition to the four given branch points. For numer-
ical reasons, these points should be chosen so that taking the convex hull leads to
triangles that are close to equilateral. We obtained good triangulations by minimiz-
ing the following energy for n points in R3 while fixing the subset of branch points:
  1
E = n2 (v, v − 1)2 + , (52)
v∈V v,w∈V
w − v, w − v
w=v

where ., . denotes the standard euclidean scalar product of R3 . We do not enforce
the constraint that the points should lie in the unit sphere S 2 . Instead, we simply
project back to S 2 after the optimization.
As initial guess we choose points uniformly distributed in S 2 . To achieve this we
choose points with normally distributed coordinates and project them to S 2 [31].

7.4 Numerical Experiments

Given the branch points of an elliptic curve, the modulus τ can be calculated in
terms of hypergeometric functions. In this section, we compare the theoretical value
of τ with the value τ̂ that we obtain by the discrete uniformization method explained
in Sect. 7.2.
We consider elliptic curves in Weierstrass normal form

μ2 = 4(z − λ1 )(z − λ2 )(z − λ3 )

(53)
= 4z 3 − g2 z − g3 ,
36 A.I. Bobenko et al.

so the branch points λ1 , λ2 , λ3 , ∞ satisfy λ1 + λ2 + λ3 = 0, and

g2 = −4(λ1 λ2 + λ2 λ3 + λ3 λ1 ), g3 = 4λ1 λ2 λ3 . (54)

We calculate the modulus τ with Mathematica using the built-in function Weier-
strassHalfPeriods[{g2 , g3 }]. We normalize τ and the value τ̂ obtained by
discrete uniformization so that they lie in the standard fundamental domain of the
modular group, |τ | > 1 and | Re(τ )| < 21 , and we consider the error |τ − τ̂ |. (We
stay away from the boundary of the fundamental domain.)
Subdivided icosahedron. In this experiment we start with the twelve vertices of
a regular icosahedron and choose the branch points λ1 , . . . , λ4 among them. The
remaining points act the role of p1 , . . . , pn . To study the dependence of |τ − τ̂ | on
the number of points we repeatedly subdivide all triangles into four similar triangles
and project the new vertices to S 2 . The number of vertices grows exponentially
while the triangles remain close to equilateral. Figure 18 shows the result of this
experiment. It suggests the error behaves like

|τ − τ̂ | = O(n α ), α ≈ −0.88. (55)

Dependence on mesh quality. In the second experiment we choose the additional

points randomly to analyze how the quality of the triangulation affects the approxi-
mation error. We use the following quantities to measure the quality of the triangu-
lation based on length cross-ratios for edges:
 
1 1
Q lcr (e) := lcr(e) + − 1,
2 lcr(e)
 
1 1
Q lmr ( f ) := lmr( f ) + − 1,
2 lmr( f )
5e 02
5e 03
^

= 0.8784918
5e 04
5e 05

10 50 100 500 5000 50000

Fig. 18 Left Error for zero to six subdivision steps. The log-log plot shows the error |τ − τ̂ |
against the number of vertices of the subdivided icosahedron (i.e., in one sheet of the doubly cov-
ered sphere). To estimate the asymptotic behavior of the error, we determine the slope α ≈ −0.88
of a line through the last four points by linear regression. Right Result of the discrete uniformiza-
tion after two subdivision steps
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 37

1e 02

1e 02
= 0.6264841
5e 04 2e 03

5e 04 2e 03
^|

^|
|

|
1e 04

1e 04
10 20 50 100 200 500 1000 10 20 50 100 200 500 1000

Fig. 19 Left log-log plot of the error |τ − τ̂ | against the number of vertices for a sample
of optimized random triangulations with no quality constraint. Right Only triangulations with
maxe {Q lmr (e)} < 0.3 are considered. The regression line with slope α ≈ −0.63 is shown in red

where lcr denotes the length cross-ratio (9) of an edge, and lmr denotes the length
multi-ratio defined for faces by lmr( f ) = e∈ f lcr(e). If Q lcr = 0 for all edges, then
the mesh is discretely conformally equivalent to a mesh consisting of equilateral
triangles. So less is better for these quality measures. To get enough “good” trian-
gulations in our samples, we improve random meshes with the procedure described
in Sect. 7.3.
Figure 19 (left) shows a plot of 2600 triangulations ranging from n = 20 to
n = 1500 vertices. No clear convergence rate is discernible. The situation improves
when only samples with a certain minimal mesh quality are considered. For the
plot in Fig. 19 (right) we selected only triangulations with maxe {Q lmr (e)} < 0.3.
(The results are similar when using the quality measures maxe {Q lcr (e)} < x or
meane {Q lcr (e)} < x.)
The results from these two experiments suggest that the error depends on the
number n of vertices asymptotically like n α , where the exponent α < 0 depends on
the mesh.

7.5 Putting a Square Pattern on a Spherical Mesh

We can use a variant of the discrete uniformization of elliptic curves (Sect. 7.2) to
put a square pattern on a surface that is homeomorphic to a sphere. Figure 20 shows
an example.
Pick four vertices of the mesh as ramification points and create a two-sheeted
branched cover of the mesh by gluing two copies along paths connecting the selected
vertices. The resulting surface is a torus. It can be uniformized using the euclidean
functional. The uniformizing group is generated by two translations. This group is
a subgroup of the group generated by rotations around the branch vertices. Hence
we can achieve the same result as follows. Instead of doubling the surface, prescribe
total angles Θ = π at the ramification vertices and Θ = 2π at all other vertices.
The result is a flat surface with four cone-like singularities of cone-angle π . The
38 A.I. Bobenko et al.

Fig. 20 The discrete “Berlin Buddy Bear”, a mascot of the SFB/Transregio 109 “Discretization in
Geometry and Dynamics”. The square pattern is put on a bear model as described in Sect. 7.5. Four
ramification vertices (marked in red) are chosen at the paws. The uniformization of the branched
double cover is shown in the middle. Each fundamental domain covers the bear twice. Fundamental
domains of the group generated by rotations around the branch points are shown on the right. Each
covers the bear once

monodromy of the developing map is generated by half-turns. Avoiding the double

cover is more efficient because one only has to minimize a function of (approxi-
mately) half the number of variables.

8 Uniformization of Surfaces of Higher Genus

As in the case of tori (Sect. 7), we can find uniformizing maps for cyclic polyhe-
dral surfaces of genus g ≥ 2 by solving the hyperbolic version (g̃ = hyp) of Prob-
lem 3.1 with prescribed total angle Θ = 2π at all vertices. (We will only consider

Fig. 21 Discrete uniformization of an embedded triangulated surface of genus 3. A fundamental

polygon with “canonical” edge pairing is shown on the right together with the image mesh. The
edges of the polygon (brown) and the axes of the edge-pairing translations (blue) are pulled back
to the embedded surface shown on the left
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 39

Fig. 22 Left An embedded triangulated surface of genus 5. Right Fundamental polygon with non-
canonical edge-pairing. The axes of the edge pairing translations are shown in blue

triangulations in the following.) This allows us to calculate approximate uniformiza-

tions for Riemann surfaces of genus g ≥ 2 given in various forms, by approximating
them with polyhedral surfaces.
In Sect. 8.1, we briefly discuss how to construct fundamental polygons and group
generators.
Not much needs to be said about the uniformization of immersed surfaces. Exam-
ples are shown in Figs. 1 (bottom), 21, and 22. In Sect. 5.1 we discussed mappings
from multiply connected domains to circle domains. Analogously, one can construct
uniformizations of polyhedral surfaces of genus g ≥ 2 with holes over the hyper-
bolic plane with circular holes. An example is shown in Fig. 23. More precisely, the
holes are bounded by hyperbolic polygons with vertices on a circle.

Fig. 23 Uniformization of a genus 2 surface with three boundary components over the hyperbolic
plane with three circular holes. The three holes are filled with polygons, which are then triangulated
during the calculation, see Sects. 5.1 and 3
40 A.I. Bobenko et al.

We explain how to calculate the Fuchsian uniformization for Riemann surfaces

given in the form of a Schottky uniformization in Sect. 8.2. We discuss the uni-
formization of hyperelliptic curves in Sect. 8.3 and a geometric characterization of
hyperelliptic Riemann surfaces in Sect. 8.4.

8.1 Fundamental Polygons and Group Generators

Basic facts and notation. Every compact Riemann R of genus g ≥ 2 can be rep-
resented as the quotient of the hyperbolic plane H 2 modulo the action of a discrete
group G of hyperbolic translations,

R = H 2 /G. (56)

Presentations of the group G play an important role. We will denote generators by

capital letters and their inverses by primes,

A := A−1 ∈ G. (57)

The uniformization group G can be presented with a finite set of generators

A, B, C, D, . . . ∈ Isom(H 2 )

subject to a single relation r = 1,

G = A, B, C, D, . . . | r = 1, (58)

where r is a product in which all generators and their inverses appear exactly once.
Such presentations are closely related with fundamental polygons: Every fundamen-
tal polygon in which all vertices are identified leads to such a presentation.
A fundamental domain of G is an open connected subset D of the hyperbolic
plane such that the G-orbit of the closure D̄ covers H 2 , and g D ∩ D = ∅ for all
g ∈ G \ {1}. A fundamental polygon of G is a fundamental domain with polygonal
boundary, i.e., the boundary consists of geodesic segments, the edges of the fun-
damental polygon, which are identified in pairs by the action of the group G. For
each edge a, there is exactly one partner edge a  such that there exists a transla-
tion A ∈ G mapping a to a  . These edge-gluing translations form a generating set
of G. If all vertices of the fundamental polygon are identified (i.e., they belong to
the same G-orbit), then the fundamental polygon has 4g edges. In this case there is
only one relation for these generators. The relation can be determined from the edge
labels, which we always list in counterclockwise order. For example, if the edges of
an octagon are labeled “canonically”,

aba  b cdc d  , (59)

Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 41

then the relation for the corresponding edge pairing translations is

DC  D  C B A B  A, (60)

and if opposite edges are identified,

abcda  b c d  , (61)

then the relation is

DC  B A D  C B  A = 1. (62)

Computational aspects. Let (, ) be a closed (euclidean, spherical, or hyperbolic)

triangulated surface of genus g ≥ 2. We solve Problem 3.1 to obtain a combinato-
˜ hyp with angle sum Θ = 2π
rially equivalent hyperbolic triangulated surface (, )
at every vertex. We lay out the triangles in the hyperbolic plane one-by-one, fol-
lowing a breadth-first search of the the 1-skeleton of the dual cell complex of .
(Alternatively, one could use a shortest spanning tree of the 1-skeleton of the dual
complex [15].) The result is a fundamental polygon with many vertices. An example
is shown in Fig. 24a.
We simplify this fundamental polygon by connecting vertices that are identified
with more than one partner by geodesic arcs, as shown in Fig. 24b. The resulting
polygon has in general more than one vertex class.
Now we perform the standard algorithm involving cut-and-glue operations (see,
e.g., [20]) to obtain a fundamental polygon with one vertex class and so-called
canonical edge identification

aba  b cdc d  . . . . (63)

During this process we maintain edge-identification transformations, which we rep-

resent as S O + (2, 1) matrices.
Hyperbolic translations tend to accumulate numerical errors quite fast when
building products. The situation could be ameliorated somewhat by using the
P S L(2, R)-representation of hyperbolic isometries [16], but the fundamental prob-
lem remains. For this reason, it is desirable to perform the cut-and-glue algorithm
in such a way that the number of matrix products required to maintain the gluing
translations is small. We follow the following greedy approach. Repeatedly we have
to choose labels x, y such that the order in the polygon is x · · · y · · · x  · · · y  , and
then perform a cut-and-glue sequence to bring the labels next to each other, x yx  y  .
We always choose a pair x, y, for which this requires the minimal number of matrix
multiplications.
The polygon with canonical edge identifications may not be convex. Following
Keen [21], we can transform this domain into a strictly convex fundamental poly-
gon by choosing a different base vertex for the same group of transformations. Let
aba  b cdc d  . . . be a fundamental polygon and let
42 A.I. Bobenko et al.

(a) (b)

(c) (d)

(e) (f)

Fig. 24 Constructing a fundamental polygon with opposite edges identified. a Laying out hyper-
bolic triangles creates a fundamental polygon with many vertices. b Straighten the edges between
vertices that are identified with more than one partner (shown in red). c Axes of the edge-pairing
translations are shown in blue. d, e Two cut-and-paste operations lead to a fundamental polygon
with one vertex class and opposite edges identified. The axes intersect in one point (see Sect. 8.4).
We move this point to the origin. f Tiling the hyperbolic plane with fundamental polygons
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 43

G = A, B, C, D, . . . ∈ PSL(2, R) | . . . DC  D  C B A B  A = 1 (64)

be the corresponding presentation of the uniformization group, see Fig. 25 (left).

Then the axes of the generators A and B intersect in a point p0 . Choosing p0 as
the base point of a new fundamental polygon as shown in Fig. 25 (right) renders it
convex and uniquely determined for the given group and presentation.
Fundamental polygons with opposite sides identified. When we consider the geo-
metric characterization of hyperelliptic surfaces in Sect. 8.4, we want to transform
fundamental polygons into fundamental polygons with opposite sides identified, i.e.,
polygons with edge labeling

abcd · · · a  b c d  · · · .

Any fundamental polygon can be transformed into a fundamental polygon with

opposite edges identified by cut-and-glue operations: First transform the polygon
to canonical form aba  b cdc d  . . . by the standard algorithm. Playing a sequence
of steps that transforms a polygon with opposite edges identified to canonical form
backwards, transforms the canonical polygon to a polygon with opposite edges iden-
tified.
This algorithm is not optimal with respect to the number of multiplications nec-
essary to maintain the edge-gluing translations. Especially if the original polygon
is already “close” to one with opposite sides identified, the detour via a canonical
polygon is inefficient.
In all examples, we use a heuristic method based on the following idea: Find a
longest sequence of different letters in the edge labeling (ignoring primes), and then
try to move a different letter into this sequence by cutting and gluing.

Fig. 25 The algorithm of Linda Keen to construct strictly convex fundamental polygons.
Start with any canonical fundamental polygon aba  b cdc d  with a corresponding relation
DC  D  C B A B  A = 1 (left). We choose the intersection p0 of the axes of transformations A and
B as base point for the new domain. The new vertices of the fundamental domain are calculated as
p1 = A Bp0 , p2 = A p0 , p3 = Bp0 , and p4 = B A p0 . The other vertices are obtained similarly
from p4 by applying C and D
44 A.I. Bobenko et al.

8.2 From Schottky to Fuchsian Uniformization

In this section, we consider Riemann surfaces presented as quotient spaces of clas-

sical Schottky groups.

Definition 8.1 Let C1 , C1 . . . , C g , C g be circles in Ĉ that bound disjoint disks. A

classical Schottky group is a Kleinian group generated by Möbius transformations
σ1 , . . . , σg , where σ j maps the outside of C j onto the inside of C j .

Each generator σ j has fixed points A j , B j inside C j and C j , respectively. The

limiting set A of G is the union of orbits of the fixed points A j , B j . G acts freely and
properly discontinuously on the domain of discontinuity  = Ĉ \ A. The quotient
space R = /G is a Riemann surface of genus g. The domain outside all of the
circles is a fundamental domain of G. The identified pairs of circles form handles.
We discretize the Riemann surface R = /G determined by a classical Schot-
tky group G as follows. First, construct a triangulation of  whose vertex set and
combinatorics are invariant under the action of G. (Ignore the fact that a Möbius
transformation maps straight edges to circular arcs as in Proposition 2.5 on the
Möbius invariance of conformal classes.) For example, the triangulation may be
the Delaunay triangulation of a G-invariant point set. The following construction
avoids Delaunay triangulations of infinite (but symmetric) point configurations:
If necessary, choose a Möbius normalization for which the fundamental domain
is bounded in C. For each pair of circles C j , C j we construct polygons p1 j , . . . , pn j j
inscribed in C j and p1 j , . . . , pn j j inscribed in C  such that σ j ( pk j ) = pk j . For
example, we may choose a regular n-gon inscribed in C j and map the vertices by σ j
to C j . Triangulate the compact region bounded by these polygons, adding vertices
in the interior as wanted. (For example, use a constrained Delaunay triangulation.)
The images of this triangulation under the action of G (again, considering only
combinatorics and vertex positions) form a G-invariant triangulation Δ̂ of the uni-
versal cover of R, hence a triangulation Δ of R. More precisely, the triangulations Δ̂
and Δ are only defined up to isotopy fixing the vertices. The edge-lengths ˆ (dis-
tances of vertices) do not project from Δ̂ to Δ, but the length cross-ratios l cr calcu-
lated from these edge lengths do, because they are Möbius invariant. The projected
length cross-ratios lcr determine a discrete conformal class for Δ (see Sect. 2.5).
To obtain a Fuchsian uniformization of R, construct edge lengths  from the
length cross-ratios lcr as described in Sect. 2.6. Then solve Problem 3.1 (or rather
the corresponding analytic version, Problem 3.4) for (Δ, )euc with g̃ = hyp and
desired angle sums Θ = 2π at all vertices.
Note that the lengths  calculated from length cross-ratios lcr may not satisfy
all triangle inequalities. This does not matter for the corresponding analytic Prob-
lem 3.4 (with V = V1 , E = E 1 ). If Problem 3.4 has a solution, it is in the discrete
conformal class determined by the length cross-ratios lcr. Also, whether or not Prob-
lem 3.4 has a solution does not depend on the choice of edge lengths  provided they
lead to the same length cross-ratios.
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 45

Fig. 26 Discrete Riemann surface of genus 3 given by Schottky data (left) and its Fuchsian uni-
formization (right). Circles with the same color are identified. The extra points of the triangulation
are chosen so that the triangles are close to equilateral where possible. The shaded region in the
right image corresponds to the fundamental domain of the Schottky group in the left image. Its
boundary consists of curves corresponding to the circles and curves corresponding to lines con-
necting the circles (drawn in gray)

Fig. 27 Left Fundamental domains of Riemann surfaces of genus 1 given by Schottky data C, C  ,
A, B, μ (see (65)). The triangulations use only points on the circles C, C  . We deliberately chose
non-concentric circles, i.e., with centers A = B. Right Representation of the same surfaces as C/ 
for a lattice . Top For real μ = 0.3 we get a rectangular lattice. Bottom μ = 0.08 + 0.01i yields
a parallelogram
46 A.I. Bobenko et al.

Figure 26 shows an example of the Fuchsian uniformization of a genus three

surface presented by its Schottky uniformization.
Tori given by Schottky data. For tori, the Schottky data consist of one generator

σ (z) − A z−A
=μ (65)
σ (z) − B z−B

and one pair of circles. To find a uniformization C/  is elementary. It suffices to

consider the case where A = B = 0 (and C, C  are concentric circles around 0 with
radii i and μ. Figure 27 shows two examples where we apply the discrete method
without adding extra points inside the fundamental domain of the Schottky group.

8.3 Hyperelliptic Curves

A hyperelliptic curve is a complex algebraic curve of the form

μ2 = p(λ), (66)

where p is a polynomial of degree d ≥ 5 with d distinct roots. For d = 2g + 2 or

d = 2g + 1, the hyperelliptic curve becomes a compact Riemann surface of genus
g after singularities at infinity are resolved. For our purposes, a hyperelliptic curve
is just a branched double cover of the λ-sphere with branch points λ1 , . . . , λ2g+1 , ∞
if d = 2g + 1 and branch points λ1 , . . . , λ2g+2 if d = 2g + 2.
We construct a polyhedral approximation of a hyperelliptic curve in the same
way as for elliptic curves (Sect. 7.2). We choose points p1 , . . . , pn in addition to
the 2g + 2 branch points and take the convex hull. We cut the resulting polyhedron
open along edge paths joining pairs of branch points and glue a second copy along
the cuts.
Figure 28 shows uniformizations of the curves

2g  
 ikπ
μ2 = λ λ−e g (67)
k=1

for g = 2, 3, 4 that were obtained this way. The curves are branched at the 2gth
roots of unity and at 0 and ∞.
Mapping a polyhedral surface to a hyperelliptic curve. We can also map a trian-
gulated surface of genus g to a branched double cover of the sphere, provided it is
symmetric with respect to a discrete conformal involution with 2g + 2 fixed points,
which are vertices. In the simplest case, the involution is an isometry. (Compare
Sect. 7.2, where we map flat tori to elliptic curves.) Taking the quotient of the trian-
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 47

Fig. 28 Uniformizations of the hyperelliptic curves (67) with genus 2, 3, and 4. The triangulation
of the surfaces is a regular 1-to-4 subdivision of the convex hull of the branch points. Due to the
symmetries of these curves, the fundamental domains are regular hyperbolic 4g-gons. Since the
triangulation is as symmetric as the curves, and because the solution of the discrete uniformization
problem is unique, the fundamental domains of the polyhedral surfaces are also exactly regular
hyperbolic 4g-gons. Any error in the domains is therefore due to numerics, and not due to the
discretization

Fig. 29 A triangulated genus 2 surface is mapped to a branched cover of Ĉ. The 180◦ rotation
about the horizontal symmetry axis is a discrete conformal involution with 6 fixed points marked
in red, blue, and purple. The texture is a square grid in the plane, pulled back to the doubly covered
sphere by Mercator projection, then pulled back to the surface. Bottom Branched cover of Ĉ, and a
closeup of three branch points

gulation with respect to the involution, we get a triangulated sphere with a discrete
conformal structure, which we map discretely conformally to the sphere. Figure 29
shows an example.
48 A.I. Bobenko et al.

8.4 Geometric Characterization of Hyperelliptic Surfaces

A Riemann surface R of genus g ≥ 2 is called hyperelliptic, if one of the following

equivalent conditions is true (and hence all are):
1. R is conformally equivalent to some hyperelliptic curve.
2. R is conformally equivalent to a branched cover of the sphere with 2g + 2
branch points.
3. There is a conformal involution τ : R → R with exactly 2g + 2 fixed points.
The involution τ is called the hyperelliptic involution of R. By the Riemann-Hurwitz
formula, the quotient surface R/τ is a sphere.
All Riemann surfaces of genus two are hyperelliptic, but for every genus greater
than two, there are Riemann surfaces that are not hyperelliptic. The following
geometric characterization of hyperelliptic Riemann surfaces is due to Schmutz
Schaller [36, 37].
Theorem 8.2 Let R be a closed hyperbolic surface of genus g. Then the following
statements are equivalent:
(i) R is hyperelliptic.
(ii) R has a set of 2g − 2 simple closed geodesics which all intersect in one point
and which intersect in no other point.
(iii) R has a set of 2g simple closed geodesics which all intersect in one point and
which intersect in no other point.
(iv) R has a fundamental polygon that is a 4g-gon with opposite sides identified
and equal opposite angles.
The fundamental polygon of condition (iv) is symmetric with respect to a 180◦
rotation around its center, which corresponds to the hyperelliptic involution on R.
The 2g + 2 fixed points on R are the vertex of the polygon, its center, and the 2g
edge midpoints. The axes of the 2g edge-gluing translations all go through the cen-
ter. They project to 2g simple closed geodesics on R which all intersect in one point
and intersect in no other point.

8.5 Example: Deforming a Hyperelliptic Surface

We uniformize a hyperelliptic surface obtaining a centrally symmetric fundamental

polygon with opposite edges identified as predicted by Theorem 8.2. The axes of
the generators meet in one point. Then we deform the surface slightly to a non-
hyperelliptic surface to see how the fundamental polygon and the axes change. The
result is shown in Fig. 30.
For this example, we construct an elliptic-hyperelliptic triangulated surface with
additional symmetry. A surface is called elliptic-hyperelliptic if it is conformally
equivalent to a two-sheeted branched cover of the torus.
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 49

Fig. 30 Hyperelliptic versus non-hyperelliptic. Left Uniformization of a hyperelliptic surface with

a centrally symmetric fundamental polygon. The axes of the generators meet in a common point.
Right Uniformization of the deformed surface, which is not hyperelliptic. The axes do not meet in
one point

Take two regular tetrahedra (the faces of which are subdivided several times to
obtain a finer mesh), cut them across pairs of opposite edges and glue them together
to obtain a two-sheeted cover of a tetrahedron branched at the four vertices. Now
choose two paths in one of the sheets that connect the centers of the tetrahedron’s
faces in pairs. Cut the surface along these paths, take another copy of this cut surface
and glue corresponding cuts together to form an elliptic-hyperelliptic surface of
genus three that is a four-fold cover of a regular tetrahedron. The surface possesses
six anti-holomorphic involutions corresponding to the six reflectional symmetries of
the tetrahedron, and three holomorphic involutions corresponding to the rotational
symmetries of the tetrahedron of order two. Each of the holomorphic involutions
has eight fixed points covering the midpoints of a pair of opposite edges. Thus, this
elliptic-hyperelliptic surface is also hyperelliptic.
Figure 30 (left) shows a uniformization of the hyperelliptic elliptic-hyperelliptic
surface. Destroying the symmetry by moving all points of the polyhedral surface
in space by a small random offset destroys the hyperellipticity of the surface, see
Fig. 30 (right).
Numerical Data. We list the numerical S O + (2, 1) matrices of the generators of the
group

T1 , T2 , T3 , T4 , T5 , T6 | T6 T5 −1 T4 T3 −1 T2 T1 −1 T6 −1 T5 T4 −1 T3 T2 −1 T1 = 1 (68)

representing the hyperelliptic elliptic-hyperelliptic surface constructed in this section

(see Fig. 31). The matrices satisfy the relation with error ≈ 10−7 .
50 A.I. Bobenko et al.

Fig. 31 Generator labels

⎡ ⎤
2.05443154523212 −4.021591426903446 −4.403849064057392
T1 = ⎣−4.021591427085276 16.338309707059754 16.796236533536394 ⎦
−4.403849064222335 16.796236533484112 17.392741252301292
⎡ ⎤
7.906334736200989 −6.57792280760043 −10.236171033333449
T2 = ⎣−6.5779228079025245 7.265127613618063 9.749417813849163 ⎦
−10.236171033527825 9.749417813638956 14.171462349831586
⎡ ⎤
933.210063638192 509.0929753776527 1063.0407708335915
T3 = ⎣509.09297492442374 279.0228056502974 580.5414569092936 ⎦
1063.0407706165242 580.5414573067374 1211.2328692884857
⎡ ⎤
47.8208492808903 21.282776040302117 −52.33345184418173
T4 = ⎣ 21.28277609643665 10.67424906068982 −23.788571865092973⎦
−52.333451867010325 −23.788571814871467 57.49509834158029
⎡ ⎤
933.2100574645401 509.09297238055706 −1063.040763978619
T5 = ⎣ 509.092972765322 279.02280467565924 −580.5414545474814⎦
−1063.0407641628826 −580.5414542100707 1211.2328621402066
⎡ ⎤
128.62265665383228 90.05086671584104 −157.0093831621644
T6 = ⎣ 90.05086668827934 64.5401174556973 −110.78621463208009⎦
−157.00938314635744 −110.78621465448322 192.16277410952506

8.6 Example: Different Forms of the Same Genus-2 Surface

In this section we present Fuchsian uniformizations of the same Riemann surface

represented in three different ways:
• as hyperelliptic curve μ2 = λ6 − 1 (Fig. 32),
• as Lawson’s genus 2 minimal surface in S 3 [25] (Fig. 33),
• and as a surface glued from six squares (Fig. 34).
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 51

Fig. 32 Uniformization of the hyperelliptic curve μ2 = λ6 − 1. Left Triangulated double cover of

the sphere branched at the 6th roots of unity, with the boundary of the fundamental domain shown
in brown and the axes of generators shown in blue. Right Fuchsian uniformization and fundamental
polygons. Canonical polygon (top), polygon with opposite sides identified (middle), and 12-gon
specially adapted to the six-squares surface (bottom)
52 A.I. Bobenko et al.

Fig. 33 Uniformization of Lawson’s surface. Left Triangulated model [34], with the boundary of
the fundamental domain shown in brown and the axes of the generators shown in blue. Right Fuch-
sian uniformizations and fundamental domains. Canonical domain (top), opposite sides domain
(middle), and 12-gon (bottom)
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 53

Fig. 34 Left A surface glued from six squares. Right Fuchsian uniformization and fundamental
domain

For each representation we choose corresponding fundamental polygons that allow

the comparison of the uniformization:
• an octagon with canonical edge pairing aba  b cdc d  ,
• an octagon with opposite sides identified, abcda  b c d  ,
• a 12-gon that is adapted to the six-squares surface.
All data presented in this section is available on the DGD Gallery webpage [39].
Hyperelliptic curve. We uniformize the hyperelliptic curve μ2 = λ6 − 1 as
described in Sect. 7.2. The results are shown in Fig. 32.
To understand the cuts on the hyperelliptic surface that lead to the 12-gon in the
bottom row, imagine taking the canonical system of loops in the top row, meeting
at the north pole, and deform them until they also meet at the south pole. This
introduces a second vertex class in the fundamental polygon.
Lawson’s surface. Figure 33 shows Fuchsian uniformizations of Lawson’s min-
imal surface in S 3 . The triangulated surface model was kindly provided by Konrad
Polthier [34].
This model of the Lawson surface realizes the hyperelliptic involution as a euclid-
ean rotational symmetry. Its symmetry axis meets the surface in six points. These
fixed points of the hyperelliptic involution correspond to the branch points of the
hyperelliptic curve representation. This allows us to uniformize the model with cor-
responding fundamental domains.
Six-squares surface. Figure 34 (left) shows a surface glued from six squares, which
is conformally equivalent to Lawson’s surface and the hyperelliptic curve. Edges
with the same marking are glued together. We calculate a uniformization using the
triangulation with vertices added in the centers of the squares as shown. An adapted
fundamental domain for this square-tiled translational surface arranges all squares
around a single vertex, see Fig. 34 (right). By comparison with Fig. 32 (bottom) we
54 A.I. Bobenko et al.

see that the vertices in the center of the squares correspond to the branch points of
the hyperelliptic representation of the surface. The black, gray, and white vertices
correspond to the north and south pole of the hyperelliptic representation.

Acknowledgments This research was supported by DFG SFB/TRR 109 “Discretization in Geom-
etry and Dynamics”.

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Abresch, U.: Constant mean curvature tori in terms of elliptic functions. J. Reine Angew.
Math. 374, 169–192 (1987)
2. Adler, V.E., Bobenko, A.I., Suris, Y.B.: Classification of integrable equations on quad-graphs.
The consistency approach. Commun. Math. Phys. 233(3), 513–543 (2003)
3. Bobenko, A.I.: All constant mean curvature tori in R 3 , S 3 , H 3 in terms of theta-functions.
Math. Ann. 290(2), 209–245 (1991)
4. Bobenko, A.I., Günther, F.: Discrete complex analysis on planar quad-graphs. In this volume
5. Bobenko, A.I., Pinkall, U., Springborn, B.: Discrete conformal maps and ideal hyperbolic
polyhedra. Geom. Topology 19(4), 2155–2215 (2015)
6. Bobenko, A.I., Suris, Y.B.: Integrable systems on quad-graphs. Int. Math. Res. Not. 11, 573–
611 (2002)
7. Bobenko, A.I., Suris, Y.B.: Discrete Differential Geometry-Integrable Structure, Graduate
Studies in Mathematics, vol. 98. American Mathematical Society, Providence (2008)
8. Chelkak, D., Smirnov, S.: Discrete complex analysis on isoradial graphs. Adv. Math. 228(3),
1590–1630 (2011)
9. Chelkak, D., Smirnov, S.: Universality in the 2D Ising model and conformal invariance of
fermionic observables. Invent. Math. 189(3), 515–580 (2012)
10. Courant, R., Friedrichs, K., Lewy, H.: Über partielle Differentialgleichungen der mathematis-
chen Physik. Math. Ann. 100, 32–74 (1928)
11. Duffin, R.J.: Discrete potential theory. Duke Math. J. 20, 233–251 (1953)
12. Duffin, R.J.: Basic properties of discrete analytic functions. Duke Math. J. 23, 335–363 (1956)
13. Duffin, R.J.: Distributed and lumped networks. J. Math. Mech. 8, 793–826 (1959)
14. Duffin, R.J.: Potential theory on a rhombic lattice. J. Comb. Theory 5, 258–272 (1968)
15. Erickson, J., Har-Peled, S.: Optimally cutting a surface into a disk. In: Symposium on Com-
putational Geometry, pp. 244–253 (2002)
16. Floyd, W., Weber, B., Weeks, J.: The Achilles’ heel of O(3, 1)? Experiment. Math. 11(1),
91–97 (2002)
17. Gu, X., Luo, F., Sun, J., Wu, T.: A discrete uniformization theorem for ployhedral surfaces.
arXiv:1309.4175 (2013)
Discrete Conformal Maps: Boundary Value Problems, Circle Domains . . . 55

18. Gu, X., Luo, F., Sun, J., Wu, T.: A discrete uniformization theorem for polyhedral surfaces II.
arXiv:1401.4594 (2014)
19. Heil, M.: Numerical tools for the study of finite gap solutions of integrable systems. Ph.D.
thesis, TU-Berlin (1995)
20. Jost, J.: Compact Riemann Surfaces, 3rd edn. Universitext. Springer-Verlag, Berlin (2006)
21. Keen, L.: Canonical polygons for finitely generated Fuchsian groups. Acta Math. 115, 1–16
(1965)
22. Kenyon, R.: Conformal invariance of domino tiling. Ann. Probab. 28(2), 759–795 (2000)
23. Kenyon, R.: The Laplacian and Dirac operators on critical planar graphs. Invent. Math. 150(2),
409–439 (2002)
24. Kouřimská, H., Skuppin, L., Springborn, B.: A variational principle for cyclic polygons with
prescribed edge lengths. In this volume
25. Lawson Jr., H.B.: Complete minimal surfaces in S3 . Ann. Math. 92(3), 335–374 (1970)
26. Leibon, G.: Characterizing the Delaunay decompositions of compact hyperbolic surfaces.
Geom. Topol. 6, 361–391 (2002)
27. Lelong-Ferrand, J.: Représentation conforme et transformations à intégrale de Dirichlet
bornée. Gauthier-Villars, Paris (1955)
28. Luo, F.: Combinatorial Yamabe flow on surfaces. Commun. Contemp. Math. 6(5), 765–780
(2004)
29. Mercat, C.: Discrete Riemann surfaces and the Ising model. Commun. Math. Phys. 218(1),
177–216 (2001)
30. Milnor, J.: Hyperbolic geometry: the first 150 years. Bull. Amer. Math. Soc. (N.S.) 6(1), 9–24
(1982)
31. Muller, M.E.: A note on a method for generating points uniformly on n-dimensional spheres.
Commun. ACM 2(4), 19–20 (1959)
32. Nehari, Z.: Conformal Mapping. McGraw-Hill Book Co., Inc., New York (1952)
33. Nijhoff, F., Capel, H.: The discrete Korteweg-de Vries equation. Acta Appl. Math. 39(1–3),
133–158 (1995). KdV ’95 (Amsterdam, 1995)
34. Oberknapp, B., Polthier, K.: An algorithm for discrete constant mean curvature surfaces. In:
Hege, H.C., Polthier, K. (eds.) Visualization and Mathematics, pp. 141–161. Springer-Verlag,
Berlin (1997)
35. Rodin, B., Sullivan, D.: The convergence of circle packings to the Riemann mapping. J. Differ.
Geom. 26(2), 349–360 (1987)
36. Schmutz Schaller, P.: Geometry of Riemann surfaces based on closed geodesics. Bull. Amer.
Math. Soc. (N.S.) 35(3), 193–214 (1998)
37. Schmutz Schaller, P.: Teichmüller space and fundamental domains of Fuchsian groups.
Enseign. Math. (2) 45(1-2), 169–187 (1999)
38. Schramm, O.: Circle patterns with the combinatorics of the square grid. Duke Math. J. 86,
347–389 (1997)
39. Sechelmann, S., Bobenko, A.I., Springborn, B.: DGD Gallery, Lawson’s surface uniformiza-
tion. https://gallery.discretization.de/models/lawsons_surface_uniformization (2015)
40. Smirnov, S.: Critical percolation in the plane: conformal invariance, Cardy’s formula, scaling
limits. C. R. Acad. Sci. Paris Sér. I Math. 333(3), 239–244 (2001)
41. Smirnov, S.: Conformal invariance in random cluster models. I. Holomorphic fermions in the
Ising model. Ann. Math. (2) 172(2), 1435–1467 (2010)
42. Smirnov, S.: Discrete complex analysis and probability. In: Proceedings of the International
Congress of Mathematicians 2010 (ICM 2010), vol. I: Plenary Lectures and Ceremonies, vols.
II-IV: Invited Lectures, pp. 595–621. Hyderabad, India (2010)
43. Springborn, B.: A unique representation of polyhedral types. Centering via Möbius transfor-
mations. Math. Z. 249(3), 513–517 (2005)
44. Stephenson, K.: Introduction to Circle Packing. The Theory of Discrete Analytic Functions.
Cambridge University Press, Cambridge (2005)
45. Thurston, W.P.: The geometry and topology of three-manifolds. Electronic version 1.1, March
2002, http://www.msri.org/publications/books/gt3m/
56 A.I. Bobenko et al.

46. Walter, R.: Constant mean curvature tori with spherical curvature lines in non-Euclidean
geometry. Manuscripta Math. 63(3), 343–363 (1989)
47. Wente, H.C.: Counterexample to a conjecture of H. Hopf. Pacific J. Math. 121(1), 193–243
(1986)
Discrete Complex Analysis on Planar
Quad-Graphs

Alexander I. Bobenko and Felix Günther

Abstract We develop further a linear theory of discrete complex analysis on

general quad-graphs, extending previous work of Duffin, Mercat, Kenyon, Chelkak
and Smirnov on discrete complex analysis on rhombic quad-graphs. Our approach
based on the medial graph leads to generalizations as well as to new proofs of previ-
ously known discrete analogs of classical theorems. New results include in particular
discretizations of Green’s first identity and Cauchy’s integral formula for the deriv-
ative of a holomorphic function. Another contribution is a discussion on the product
of discrete holomorphic functions that is itself discrete holomorphic in a specific
sense. In this paper, we focus on planar quad-graphs, but many notions and the-
orems can be easily adapted to discrete Riemann surfaces. In the case of planar
parallelogram-graphs with bounded interior angles and bounded ratio of side lengths
explicit formulae for a discrete Green’s function and discrete Cauchy’s kernels are
obtained. This slightly generalizes the previous results on rhombic lattices. When we
further restrict to the integer lattice of a two-dimensional skew coordinate system a
discrete Cauchy’s integral formulae for higher order derivatives is derived.

Keywords Discrete complex analysis · Quad-graph · Green’s function · Cauchy’s

integral formulae · Parallelogram-graph

1 Introduction

Linear theories of discrete complex analysis look back on a long and varied
history. We refer here to the survey of Smirnov [24]. Already Kirchhoff’s circuit laws
describe a discrete harmonicity condition for the potential function whose gradient
describes the current flowing through the electric network. A notable application

A.I. Bobenko · F. Günther (B)

Inst. für Mathematik, Technische Universität Berlin, Straße des 17. Juni 136,
10623 Berlin, Germany
e-mail: [email protected]
A.I. Bobenko
e-mail: [email protected]

© The Author(s) 2016 57

A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_2
58 A.I. Bobenko and F. Günther

of Kirchhoff’s laws in geometry was the article [4] of Brooks, Smith, Stone, and
Tutte, who used coupled discrete harmonic functions (in fact, discrete holomorphic
functions) to construct tilings of rectangles into squares with different integral side
lengths. Discrete harmonic functions on the square lattice were studied by a number
of authors in the 1920s, including Courant, Friedrichs, and Lewy, who showed con-
vergence of solutions of the Dirichlet boundary value problem to their corresponding
continuous counterpart [8].
Discrete holomorphic functions on the square lattice were studied by Isaacs [14].
He proposed two different definitions for holomorphicity. The first one is not symmet-
ric on the square lattice, but it becomes symmetric on the triangular lattice obtained
by inserting all southwest-to-northeast diagonals. Dynnikov and Novikov studied an
equivalent notion in [11]. His second definition was reintroduced by Lelong-Ferrand
in [12]. She developed the theory to a level that allowed her to prove the Riemann
mapping theorem using discrete methods [18]. Duffin also studied discrete complex
analysis on the square grid [9] and was the first who extended the theory to rhombic
lattices [10]. Mercat [19], Kenyon [16], Chelkak and Smirnov [6] resumed the inves-
tigation of discrete complex analysis on rhombic lattices or, equivalently, isoradial
graphs. In these settings, it was natural to split the real and the imaginary part of a
discrete holomorphic function to the two vertex sets of a bipartite decomposition.
Some two-dimensional discrete models in statistical physics exhibit conformally
invariant properties in the thermodynamical limit. Such conformally invariant prop-
erties were established by Smirnov for site percolation on a triangular grid [25] and
for the random cluster model [26], by Chelkak and Smirnov for the Ising model [7],
and by Kenyon for the dimer model on a square grid (domino tiling) [15]. In all cases,
linear theories of discrete analytic functions on regular grids were highly important.
Kenyon [16] as well as Chelkak and Smirnov [6] obtained important analytic results
that were instrumental in the proof that the critical Ising model is universal, i.e.,
that the scaling limit is independent of the shape of the lattice [7]. Already Mercat
related the theory of discrete complex analysis to the Ising model and investigated
criticality [19].
Important non-linear discrete theories of complex analysis involve circle packings
or, more generally, circle patterns. Rodin and Sullivan proved that the Riemann
mapping of a complex domain to the unit disk can be approximated by circle packings
[22]. A similar result for isoradial circle patterns, even with irregular combinatorics,
is due to Bücking [5]. In [2] it was shown that discrete holomorphic functions describe
infinitesimal deformations of circle patterns. Moreover, in the case of parallelogram-
graphs it was discussed that the corresponding theory is integrable by embedding the
quad-graph into Zn .
Our setup in Sect. 2 is a strongly regular cellular decomposition of the com-
plex plane into rectilinear quadrilaterals, called quad-graph. The medial graph of a
quad-graph plays a crucial role in our work. It provides the connection between the
notions of discrete derivatives of Kenyon [16], Mercat [20], Chelkak and Smirnov [6],
extended from rhombic to general quad-graphs, and discrete differential forms and
discrete exterior calculus as suggested by Mercat [19, 21]. Our treatment of discrete
differential forms is close to what Mercat proposed in [21]. However, our version
Discrete Complex Analysis on Planar Quad-Graphs 59

of discrete exterior calculus is slightly more general. Having introduced the basic
notations in the first two sections and the discrete exterior derivative in Sect. 2.3.1,
we come to Proposition 2.13. There, it is described how the medial graph allows to
multiply discrete holomorphic functions to a function that is discrete holomorphic
in a certain sense, in particular that it fulfills (discrete) Morera’s theorem.
The medial graph approach turns out to be quite useful for integration theory. The-
orem 2.16 shows that the discrete exterior derivative is a derivation of the discrete
wedge product. Many further results rely on this result and discrete Stokes’ Theo-
rem 2.9. In particular, this concerns discrete Green’s identities (Theorem 2.23). A dis-
cretization of Green’s second identity was one ingredient in the proof of Skopenkov’s
convergence result in [23]. Before the theorem is proved, we introduce the discrete
wedge product, the discrete Hodge star, and the discrete Laplacian in Sects. 2.3
and 2.4.
Skopenkov’s results on the existence and uniqueness of solutions to the discrete
Dirichlet boundary value problem [23] help us to prove Theorem 2.30. This theorem
states surjectivity of the discrete derivatives and the discrete Laplacian seen as lin-
ear operators. This implies in particular the existence of discrete Green’s functions
and discrete Cauchy’s kernels. Furthermore, discrete Cauchy’s integral formulae for
discrete holomorphic functions are derived in Theorem 2.35 and for the discrete
derivative of a discrete holomorphic function on the vertices of the quad-graph in
Theorem 2.36. Note that discrete Cauchy’s integral formula was used by Chelkak
and Smirnov to derive certain asymptotic estimates in [7].
Section 3 is devoted to discrete complex analysis on planar parallelogram-graphs.
There, explicit formulae for discrete Green’s functions and discrete Cauchy’s kernels
with asymptotics similar to the functions in the rhombic case [5, 6, 16] are given
(Theorems 3.7, 3.8, and 3.10). The general assumption is that the interior angles and
the ratio of side lengths of all parallelograms are bounded. The construction of these
functions is based on the discrete exponential introduced by Kenyon on quasicrys-
tallic rhombic quad-graphs [16] and its extension to quasicrystallic parallelogram-
graphs [2].
In the end, we close with the very special case of the integer lattice of a skew
coordinate system in the complex plane. In this case, discrete Cauchy’s integral
formulae for higher order discrete derivatives of a discrete holomorphic function are
derived in Theorem 3.11.

2 Discrete Complex Analysis on Planar Quad-Graphs

Although we focus on planar quad-graphs in this paper, many of our notions and
theorems generalize to discrete Riemann surfaces. A corresponding linear theory of
discrete Riemann surfaces is discussed in the subsequent paper [1] and can be found
in more detail in the thesis [13].
60 A.I. Bobenko and F. Günther

2.1 Basic Definitions and Notation

The aim of this section is to introduce first planar quad-graphs and some basic notation
in Sect. 2.1.1 and then to discuss the medial graph in Sect. 2.1.2.

2.1.1 Planar Quad-Graphs

Definition A planar quad-graph Λ without boundary is an infinite graph embedded

into the complex plane C such that all edges are straight line segments and such that
all faces are quadrilaterals which may be non-convex. In addition, we assume that Λ
induces a cellular decomposition of the whole complex plane that is locally finite,
i.e., a compact subset of C contains only finitely many quadrilaterals, and strongly
regular, i.e., two different faces are either disjoint or share exactly one vertex or share
exactly one edge (but not two edges).
Let V (Λ) denote the set of vertices, E(Λ) the set of edges, and F(Λ) the set of
faces of Λ.

It is well known that any planar quad-graph is bipartite. We fix one decomposi-
tion of the vertices of Λ into two independent sets and refer to the vertices of this
decomposition as black and white vertices, respectively.

Definition Let Γ and Γ ∗ be the graphs defined on the black and white vertices
where vv is an edge of Γ (or Γ ∗ ) if and only if its two black (or white) endpoints
are vertices of a single face.

Remark The assumption of strong regularity guarantees that any edge of Γ or Γ ∗ is

the (possibly outer) diagonal of exactly one quadrilateral.

In order to make the duality between Γ and Γ ∗ apparent, we consider just for
this paragraph the edges of Γ or Γ ∗ as curves lying totally inside the face they
are a diagonal of. Then, any black edge of Γ corresponds to the white edge of
Γ ∗ that crosses it and vice versa. The black and white vertices are in one-to-one
correspondence to the faces of Γ ∗ and Γ they are contained in.

Definition The complex number assigned to a vertex of Λ is the corresponding

complex value of its embedding in C. To oriented edges of Λ, Γ, Γ ∗ we assign the
complex numbers determined by the difference of the complex numbers assigned to
their two endpoints.

Remark For simplicity, we perform our calculations hereafter directly with the ver-
tices and oriented edges of Λ, Γ, Γ ∗ rather than replacing them with their corre-
sponding complex values.

Definition ♦ := Λ∗ is the dual graph of Λ.

Discrete Complex Analysis on Planar Quad-Graphs 61

In general, we do not specify a planar embedding of the dual graph ♦. We will just
identify vertices or faces of ♦ with their corresponding dual faces and vertices of Λ,
respectively. However, in the particular case that all quadrilaterals are parallelograms,
it makes sense to identify the vertices of ♦ (i.e., faces of the quad-graph Λ) with
the centers of the parallelograms. Here, the center of a parallelogram is the point of
intersection of its two diagonals. Further details will be given in Sects. 2.2.1 and 2.2.3.

Definition If a vertex v ∈ V (Λ) is a vertex of a quadrilateral Q ∈ F(Λ), we write

Q ∼ v or v ∼ Q and say that v and Q are incident to each other.

Throughout our paper, we will denote the vertices of a single quadrilateral Q by

b− , w− , b+ , w+ in counterclockwise order, where b± ∈ V (Γ ) and w± ∈ V (Γ ∗ ).

Definition For a quadrilateral Q ∈ F(Λ) we define

w+ − w−
ρ Q := −i .
b+ − b−

Moreover, let
     
ρQ (b+ − b− )(w+ − w− )
ϕ Q := arccos Re i = arccos Re
|ρ Q | |b+ − b− ||w+ − w− |

be the angle under which the diagonal lines of Q intersect.

Note that 0 < ϕ Q < π . Figure 1 shows a finite bipartite quad-graph together with
the notations we have introduced for a single quadrilateral Q and the notations we
are using later for the star of a vertex v, i.e., the set of all faces incident to v.

v2

v2 v1 v1

b− w+ v
vk

Q
vs−1 Qs
Q vs

w− b+ vs

Fig. 1 Bipartite quad-graph with notations

62 A.I. Bobenko and F. Günther

In addition, we denote by ♦0 a subgraph of ♦ that we always assume to be

connected and by V (♦0 ) ⊆ V (♦) the corresponding subset of faces of the quad-
graph Λ. Through our identification V (♦) ∼ = F(Λ), we can call the elements of
V (♦) quadrilaterals and identify them with the corresponding faces of Λ.

Definition ♦0 ⊆ ♦ is said to form a simply-connected closed region if the union of

all quadrilaterals in V (♦0 ) is a simply-connected closed region in C.

Definition Let Λ0 be the subgraph of Λ whose vertices and edges are exactly the
corners and edges of the quadrilaterals in V (♦0 ). The interior faces of Λ0 are given
by V (♦0 ). Let Γ0 and Γ0∗ denote the subgraphs of Γ and of Γ ∗ whose edges are the
diagonals of quadrilaterals in V (♦0 ) and who do not contain isolated vertices.

Remark Since ♦0 ⊆ ♦ is connected, Γ0 and Γ0∗ are connected as well. Indeed, if

Q, Q  ∈ V (♦0 ) are two quadrilaterals adjacent in ♦0 , then the corresponding sub-
graphs on Γ0 and Γ0∗ consisting of the three black and white vertices of Q and Q 
are paths of two edges each.

Definition Λ0 is called a (planar) quad-graph induced by the subgraph ♦0 ⊆ ♦. Its

boundary ∂Λ0 is the subgraph of Λ0 that consists of all the edges of Λ0 that belong
to both a quadrilateral in V (♦0 ) and one in V (♦)\V (♦0 ).

Remark If ♦0 ⊆ ♦ is finite and forms a simply-connected closed region, then the set
of all interior faces of Λ0 is homeomorphic to a disk and ∂Λ0 is a closed broken line
without self-intersections.

2.1.2 Medial Graph

Definition The medial graph X of Λ is defined as follows. Its vertex set is given
by all the midpoints of the edges of Λ embedded in C, and two vertices x, x  are
adjacent if and only if the corresponding edges belong to the same face Q of Λ and
have a vertex v ∈ V (Λ) in common. We denote this edge by [Q, v]. Taking [Q, v] as
a straight line segment if v is a convex corner of the quadrilateral Q and as a curve
lying inside Q that does not intersect the three other edges [Q, v ] (v ∼ Q, v = v)
inside Q if v is a concave corner, we get an embedding of X into C. Then, the set
F(X ) of faces of X is in bijection with V (Λ) ∪ V (♦): A face Fv of X corresponding
to v ∈ V (Λ) has the midpoints of edges of Λ incident to v as vertices, and a face FQ
of X corresponding to Q ∈ F(Λ) ∼ = V (♦) has the midpoints of the four edges of Λ
belonging to Q as vertices. The vertices of FQ and Fv are colored gray in Fig. 2.

Remark Clearly, any pair Q ∼ v of a face and an incident vertex of Λ corresponds

to an edge [Q, v] of X . Moreover, a face FQ lies inside Q and v is contained in Fv .

Definition As for the vertices and edges of Λ, we assign to a vertex of X the complex
number corresponding to its position in C, and to an oriented edge of X we assign
the difference of the two endpoints.
Discrete Complex Analysis on Planar Quad-Graphs 63

v
Fv

Pv
Q
FQ
PQ

Fig. 2 Bipartite quad-graph (dashed) with medial graph (solid)

Even though not all edges of X might be straight line segments, we actually think
of them as being straight since we assign the vector of its endpoints to it if the edge is
oriented. In this sense, any face FQ , Q ∈ F(Λ), is a parallelogram due to Varignon’s
theorem. Moreover, the complex number assigned to the edge [Q, v0 ], v0 ∼ Q, if
 
oriented from the midpoint of the edge v0 v− to the one of the edge v0 v+ of Λ is
 
just half of e = v+ − v− . We will say that [Q, v0 ] is parallel to e (disregarding the
orientation), as it would be if we considered all edges of X as straight line segments.
Remark If all quadrilaterals of Λ are convex, then the embedding of X given above
consists of straight line segments only. If no Varignon parallelogram of a non-convex
quadrilateral contains another vertex of X apart from its corners, then the correspond-
ing straight line realization gives an embedding equivalent to the one above. In this
case, the face Fv of X corresponding to a vertex v ∈ V (Λ) that is a concave corner
of a quadrilateral does not contain v any longer. However, if such a Varignon paral-
lelogram contains an additional vertex of X , then connecting adjacent vertices of X
by straight line segments does not yield an embedding of X .

Definition For a connected subgraph ♦0 ⊆ ♦, we denote by X 0 ⊆ X the connected

subgraph of X consisting of all edges [Q, v] where Q ∈ V (♦0 ) and v is a corner of
Q. The boundary ∂ X 0 is the subgraph of X 0 whose edges are exactly those [Q, v]
where v ∈ V (∂Λ0 ) and Q ∈ V (♦0 ) is incident to v. We consider the orientation on the
boundary ∂ X 0 that is induced by orienting any of its edges [Q, v] counterclockwise
with regard to Q. For a finite collection F of faces of X 0 , ∂ F denotes the union
of all counterclockwise oriented boundaries of faces in F, where oriented edges in
opposite directions cancel each other out.
Remark In the case that all quadrilaterals in V (♦0 ) are convex, X 0 consists of all
straight edges of X that lie inside the closed region formed by the quadrilaterals in
V (♦0 ). In any case, the medial graph X corresponds to a (strongly regular and locally
finite) cellular decomposition of the plane in a canonical way. In particular, we can
talk about a topological disk D in F(X 0 ) as a finite collection of faces of X 0 that form
64 A.I. Bobenko and F. Günther

a closed region homeomorphic to a disk, and we can consider its counterclockwise

oriented boundary ∂ D as a subgraph of X 0 .

Definition For v ∈ V (Λ) and Q ∈ F(Λ), let Pv and PQ be the closed paths on
X connecting the midpoints of edges of Λ incident to v and Q, respectively, in
counterclockwise direction. In Fig. 2, their vertices are colored gray. We say that Pv
and PQ are discrete elementary cycles.

2.2 Discrete Holomorphicity

To motivate the definition of discrete holomorphicity due to Mercat [21] that was
also used previously in the rhombic setting by Duffin [10] and others, let us have
a short look to the classical theory. There, a real differentiable complex function f
defined on an open subset of the complex plane is holomorphic if and only if in
any point all directional derivatives coincide. Moreover, holomorphic functions with
nowhere-vanishing derivative preserve angles, and at a single point, infinitesimal
lengths are uniformly scaled.

Definition Let Q ∈ V (♦) ∼ = F(Λ) and f be a complex function on the vertices

b− , w− , b+ , w+ of the quadrilateral Q. f is said to be discrete holomorphic at Q if
the discrete Cauchy-Riemann equation is satisfied:

f (b+ ) − f (b− ) f (w+ ) − f (w− )

= .
b+ − b− w+ − w−

Definition Let f : V (Λ0 ) → C. f is said to be discrete holomorphic if f is discrete

holomorphic at all Q ∈ V (♦0 ).

Note that if a discrete holomorphic function f does not have the same value on
both black vertices b− and b+ , then it preserves the angle ϕ Q and f uniformly scales
the lengths of the diagonals of Q. However, the image of Q under f might be a
degenerate or self-intersecting quadrilateral.
We immediately see that for discrete holomorphicity, only the differences at black
and at white vertices matter. Hence, we should not consider constants on V (Λ), but
biconstants [20] determined by each a value on V (Γ ) and V (Γ ∗ ).

Definition A function f : V (Λ0 ) → C that is constant on V (Γ0 ) and constant on

V (Γ0∗ ) is said to be biconstant.
¯ first of functions on V (Λ)
In the following, we will define discrete analogs of ∂, ∂,
in Sect. 2.2.1 and later of functions on V (♦) in Sect. 2.2.3. Before, we introduce
discrete differential forms in Sect. 2.2.2.
Discrete Complex Analysis on Planar Quad-Graphs 65

2.2.1 Discrete Derivatives of Functions on the Vertices of Λ

Definition Let Q ∈ V (♦) = ∼ F(Λ), and let f be a complex function on its vertices
b− , w− , b+ , w+ . The discrete derivatives ∂Λ f , ∂¯Λ f are defined by

f (b+ ) − f (b− ) f (w+ ) − f (w− )

∂Λ f (Q) := λ Q + λ̄ Q ,
b+ − b− w+ − w−
f (b+ ) − f (b− ) f (w+ ) − f (w− )
∂¯Λ f (Q) := λ̄ Q + λQ ,
b+ − b− w+ − w−
  
where 2λ Q := exp −i ϕ Q − π2 / sin(ϕ Q ).

Remark Clearly, biconstant functions have vanishing discrete derivatives.

If the quadrilateral Q is a rhombus, then ϕ Q = π/2 and λ Q = 1/2. Therefore, the

definition above then reduces to the previous one given by Chelkak and Smirnov [6].
The definition of discrete derivatives matches the notion of discrete holomorphicity;
and the discrete derivatives approximate their smooth counterparts correctly up to
order one for general quad-graphs and up to order two for parallelogram-graphs:

Proposition 2.1 Let Q be a face of the quad-graph Λ and f be a complex function

on its vertices b− , w− , b+ , w+ .
(i) f is discrete holomorphic at Q if and only if ∂¯Λ f (Q) = 0.
(ii) For the function f (v) = v, ∂¯Λ f (Q) = 0 and ∂Λ f (Q) = 1.
(iii) If Q is a parallelogram and f (v) = v2 , then ∂¯Λ f (Q) = 0, ∂Λ f (Q) = 2 Q̂.
(iv) If Q is a parallelogram and f (v) = |v|2 , then ∂¯Λ f (Q) = ∂Λ f (Q) = Q̂.
In parts (iii) and (iv), Q̂ denotes the center of the parallelogram Q.

Proof (i) We observe that

     b −b
2 sin(ϕ Q )λ̄ Q exp i ϕ Q − π2 −i exp iϕ Q |b+ −b− |
+ − −i (w+ − w− )
= = = ,
b+ − b− b+ − b− |b+ − b− | |w + − w− ||b+ − b− |
     w −w
2 sin(ϕ Q )λ Q exp i π2 − ϕ Q i exp −iϕ Q |w+ −w− |
+ − i (b+ − b− )
= = = .
w+ − w− w+ − w− |w+ − w− | |w+ − w− ||b+ − b− |

So if we multiply ∂¯Λ f (Q) by 2i|w+ − w− ||b+ − b− | sin(ϕ Q ) = 0, we obtain

(w+ − w− ) ( f (b+ ) − f (b− )) − (b+ − b− ) ( f (w+ ) − f (w− )) ,

which vanishes if and only if the discrete Cauchy-Riemann equation is satisfied.

(ii) Clearly, f (v) = v satisfies the discrete Cauchy-Riemann
  equation.
  By the 
first part, ∂¯Λ f (Q) = 0. Due to 2 sin(ϕ Q ) = exp −i ϕ Q − π2 + exp i ϕ Q − π2 ,
∂Λ f (Q) simplifies to λ Q + λ̄ Q = 1.
66 A.I. Bobenko and F. Günther

(iii) For the function f (v) = v2 , the discrete Cauchy-Riemann equation is equiv-
alent to b+ + b− = w+ + w− . But since Q is a parallelogram, both (b+ + b− )/2 and
(w+ + w− )/2 equal its center Q̂. Thus, f is discrete holomorphic at Q and

∂Λ f (Q) = λ Q (b+ + b− ) + λ̄ Q (w+ + w− ) = 2 Q̂(λ Q + λ̄ Q ) = 2 Q̂.

(iv) Since f is a real function, ∂¯Λ f (Q) = ∂Λ f (Q) follows straight from the defin-
ition. Let z ∈ C be arbitrary. If g(v) := vz̄, then ∂Λ g(Q) = z̄ and ∂Λ ḡ(Q) = 0 by the
second part. So if we define the function h(v) := |v − z|2 = |v|2 − vz̄ − v̄z + |z|2 ,
then ∂Λ h(Q) = ∂Λ f (Q) − z using the second part and observing that constant func-
tions have vanishing derivatives. Hence, the statement is invariant under transla-
tion, and it suffices to consider the case Q̂ = 0. Then, b+ = −b− and w+ = −w−
since Q is a parallelogram. It follows that f (b− ) = f (b+ ) and f (w− ) = f (w+ ), so
∂Λ f (Q) = 0.
Our first discrete analogs of classical theorems are immediate consequences of
the discrete Cauchy-Riemann equation:
Proposition 2.2 Let f : V (Λ0 ) → C be discrete holomorphic.
(i) If f is purely imaginary or purely real, then f is biconstant.
(ii) If ∂Λ f ≡ 0, then f is biconstant.
Proof (i) Let us assume that f is not biconstant. Then, without loss of generality, f
is not constant on Γ0 . Since Γ0 is connected, there are two adjacent vertices b− , b+
of Γ0 such that f (b+ ) = f (b− ). Let b− , w− , b+ , w+ ∈ V (Λ0 ) be the vertices of the
interior face of Λ0 with black diagonal b− b+ . Due to the discrete Cauchy-Riemann
equation,
f (w+ ) − f (w− ) w+ − w−
= .
f (b+ ) − f (b− ) b+ − b−

The left hand side is real and well-defined since f is purely imaginary or purely real
and f (b+ ) = f (b− ). But the right hand side is not, contradicting the assumption that
f is not biconstant.
(ii) Since f is discrete holomorphic,

f (b+ ) − f (b− ) f (w+ ) − f (w− )

= .
b+ − b− w+ − w−

∂Λ f ≡ 0 then yields that both sides of the discrete Cauchy-Riemann equation equal
zero, so f is constant on V (Γ0 ) and on V (Γ0∗ ) since both graphs are connected.

2.2.2 Discrete Differential Forms

In our paper, we mainly consider two type of functions, functions f : V (Λ0 ) → C

and functions h : V (♦0 ) → C. An example for a function on V (♦0 ) is ∂Λ f .
Discrete Complex Analysis on Planar Quad-Graphs 67

Definition A discrete one-form ω is a complex function on the oriented edges of the

medial graph X 0 such that ω(−e) = ω(e) for any oriented edge e of X 0 . Here, −e
denotes the edge e with opposite orientation. 
The evaluation of ω at an oriented edge e of X 0 is denoted by e ω. If P is a
oforiented edges e1 , e2 , . . . , en , then
the discrete
directed path in X 0 consisting
integral along P is P ω = nk=1 ek ω. For closed paths P, we write P ω instead.
If P is the oriented boundary of a topological disk D in F(X 0 ), then we say that the
discrete integral is a discrete contour integral with discrete contour P.
Since we consider the planar case, one-forms dz and d z̄ are globally defined.
 
Definition The discrete one-forms dz and d z̄ are given by e dz = e and e d z̄ = ē
for any oriented edge e of X .
It turns out that discrete one-forms that actually come from discrete one-forms on
Γ and Γ ∗ are of particular interest:
Definition A discrete one-form ω defined on the oriented edges of X 0 is of type ♦ if
for any Q ∈ V (♦0 ) there exist complex numbers p, q such that ω = pdz + qd z̄ on
all edges e = [Q, v], v ∈ V (Λ0 ) incident to Q. ω is of type Λ if for any v ∈ V (Λ0 )
there exist complex numbers p, q such that ω = pdz + qd z̄ on all edges e = [Q, v],
Q ∈ V (♦0 ) incident to v.
Remark Discrete one-forms of type Λ do not play such an important role as discrete
one-forms of type ♦, although they occur as discrete differentials of functions defined
on V (♦0 ). This will become clear in the end of Sect. 2.3.2, one of the reasons being
that discrete one-forms of type Λ are not defined on discrete Riemann surfaces, but
discrete one-forms of type ♦ are.
Definition A discrete two-form Ω is a complex function on the faces of X 0 .
The evaluation of Ω at a face  F of X is

0 denoted by F Ω. If S is a set of
faces F1 , F2 , . . . , Fn of X 0 , then S Ω = nk=1 Fk Ω is the discrete integral of Ω
over S.
As we are mainly interested in functions f : V (Λ0 ) → C and h : V (♦0 ) → C,
discrete two-forms of particular interest are those that vanish on faces of X 0 corre-
sponding to vertices of either ♦ or Λ.
Definition A discrete two-form Ω defined on F(X 0 ) is of type Λ if Ω vanishes on
all faces of X 0 corresponding to V (♦0 ) and of type ♦ if Ω vanishes on all faces of
X corresponding to V (Λ0 ).
Remark These discrete two-forms correspond to functions on V (Λ0 ) or V (♦0 ) by
the discrete Hodge star that will be defined later in Sect. 2.3.3.
Since we did not give a precise embedding of the medial graph into the complex
plane in the general case, we have to specify what the area of a face is. This area
includes a factor of two in order to get the same coefficients in the discrete setup as
in the smooth case.
68 A.I. Bobenko and F. Günther

Definition Let F be a face of the medial graph X . We define ar(F) to be twice the
Euclidean area of the polygon that results from connecting adjacent vertices of F by
straight line segments in the complex plane. In contrast, area(P) will always denote
the Euclidean area of a polygon P.

Remark As we have mentioned before, our main objects either live on the quad-graph
Λ or on its dual ♦. Thus, we have to deal with two different cellular decompositions
at the same time. The medial graph has the crucial property that its faces are in
one-to-one correspondence to vertices of Λ and of ♦, i.e., to faces of ♦ and of Λ.
Furthermore, the Euclidean area of the Varignon parallelogram of Q ∈ F(Λ) is just
half of the area of Q. In some sense, a corresponding statement is true for the cells
of X corresponding to vertices of Λ, i.e., faces of ♦. However, there is not only
no canonical embedding of X , but also no natural embedding of ♦ in the general
setting. But in the particular case of parallelogram-graphs, when we have a canonical
embedding of X with rectilinear edges, we can make the statement precise: If an edge
Q Q  of ♦ is represented by the two line segments that connect the centers of the
parallelograms Q and Q  with the midpoint of their common edge, then the Euclidean
area of the face of X corresponding to a vertex v ∈ V (Λ) ∼ = F(♦) is exactly half of
the area of the face of ♦ corresponding to v.
In summary, the medial graph allows us to deal with just one decomposition of the
complex plane, but we have to count areas twice in order to get the right coefficients
as in the continuous setup.

Definition The discrete two-forms ΩΛ and Ω♦ are defined as being zero on faces
of X corresponding to vertices of ♦ or Λ, respectively, and defined by
 
ΩΛ = −2iar(Fv ) and Ω♦ = −2iar(FQ )
Fv FQ

on faces Fv and FQ corresponding to v ∈ V (Λ) or Q ∈ V (♦). As defined above,

ar(F) is twice the Euclidean area of the straight-line embedding of Fv or FQ .

Remark ΩΛ and Ω♦ are the straightforward discretizations of dz ∧ d z̄ having in

mind that they are essentially defined on faces of ♦ or of Λ, respectively. It turns
out that in local coordinates, we can perform our calculations with ΩΛ and Ω♦ in
the discrete setting exactly as we do with dz ∧ d z̄ in the smooth theory. We will
see in Sect. 2.3.2 that Ω♦ is indeed the discrete wedge product of dz and d z̄ seen
as discrete one-forms of type ♦. The same would be true for ΩΛ if we considered
dz and d z̄ as being of type Λ, but the discrete wedge product is of interest just for
discrete one-forms of type ♦ and we therefore define it just for these forms.
Definition Let f : V (Λ0 ) → C, h : V (♦0 ) → C, ω a discrete one-form defined on
the oriented edges of X 0 , and Ω1 , Ω2 discrete two-forms defined on F(X 0 ) that are
of type Λ and ♦, respectively. For any oriented edge e = [Q, v] and any faces Fv , FQ
of X 0 corresponding to v ∈ V (Λ0 ) or Q ∈ V (♦0 ), we define the products f ω, hω,
f Ω1 , and hΩ2 by
Discrete Complex Analysis on Planar Quad-Graphs 69
    
f ω : = f (v) ω and f Ω1 := f (v) Ω1 , f Ω1 := 0;
e e Fv Fv FQ
    
hω : = h(Q) ω and hΩ2 := 0, hΩ2 := h(Q) Ω2 .
e e Fv FQ FQ

In the following table, we give a quick overview of various discrete differential

forms (most of them will be discussed in Sect. 2.3) and state whether they are essen-
tially functions on V (Λ) (first column) or functions on V (♦) (second column) or
entirely objects on the cellular decomposition X (third column). In the first row we
find functions, in the second discrete one-forms, and in the third discrete two-forms.
So for example, the intersection of the second row with the third column lists discrete
one-forms that are entirely objects on X and cannot be reduced to functions on V (Λ)
or V (♦).

Λ ♦ X 
functions f, g : V (Λ) → C h 1 , h 2 : V (♦) → C f · g = ( f dg + gd f )
∂♦ h, ∂¯♦ h ∂Λ f, ∂¯Λ f
1-forms dh df f dg + gd f
h 1 dz + h 2 d z̄ f dz + gd z̄ f hdz
η of type Λ ω, ω of type ♦ fω
2-forms ΩΛ Ω♦
f h
dω dη d( f hdz)
f dω ω ∧ ω d( f ω)

Remark Although discrete one-forms of type Λ or of type ♦ do not live themselves

on Λ or ♦, they are described by two functions defined on the vertices of Λ or ♦,
respectively.

2.2.3 Discrete Derivatives of Functions on the Faces of Λ

Before we pass on to discrete derivatives of functions on V (♦), we first prove an

alternative formula for the discrete derivatives of functions on V (Λ).

Lemma 2.3 Let Q ∈ V (♦) ∼ = F(Λ) and f be a complex function on the vertices
b− , w− , b+ , w+ of Q. Let PQ be the discrete elementary cycle around Q and F the
face of X corresponding to Q. Then,
 
−1 1
∂Λ f (Q) = f d z̄ and ∂¯Λ f (Q) = f dz.
2iar(F) 2iar(F)
PQ PQ
70 A.I. Bobenko and F. Günther

Proof Since we think of F as a parallelogram (see Sect. 2.1.2), its Euclidean area is
half of the area of Q. So by definition,

1
ar(F) = |b+ − b− ||w+ − w− | sin(ϕ Q ).
2

Furthermore, f (b+ ) and − f (b− ) are multiplied

by the same factor (w+ − w− )/2
when evaluating the discrete contour integral PQ f d z̄. Therefore, the coefficient in
front of f (b+ ) − f (b− ) in the right hand side of the first equation in the lemma is
  
w+ − w− w+ − w− exp −i ϕ Q − π2 λQ
i = −i = =
4ar(F) 2 sin(ϕ Q )|w+ − w− ||b+ − b− | 2 sin(ϕ Q )(b+ − b− ) b+ − b−

(compare with the proof of Proposition 2.1(i)), which is exactly the coefficient appear-
ing in ∂Λ f (Q). Analogously, the coefficients in front of f (w+ ) − f (w− ) are equal.
This shows the first equation. The second one follows from the first, noting that
the coefficients in front of f (b+ ) − f (b− ) and f (w+ ) − f (w− ) on both sides of
the second equation are just complex conjugates of the corresponding coefficients
appearing in the first equation.

Inspired by Lemma 2.3 that is illustrated by Fig. 3a, we can now define the discrete
derivatives for complex functions on V (♦), see Fig. 3b.

Definition Let v ∈ V (Λ) and h be a complex function defined on all quadrilaterals

that are incident to v. Let Pv be the discrete elementary cycle around v and F the face
of X corresponding to v. Then, the discrete derivatives ∂♦ h, ∂¯♦ h at v are defined by
 
−1 1
∂♦ h(v) := hd z̄ and ∂¯♦ h(v) := hdz.
2iar(F) 2iar(F)
Pv Pv

h is said to be discrete holomorphic at v if ∂¯♦ h(v) = 0.

(a) w+ (b)
vs
Pv vs
PQ v Qs

b− b+
vs−1

w−

Fig. 3 Illustrations to the integration formulae for discrete derivatives. a Lemma 2.3 for ∂Λ , ∂¯Λ .
b Definitions of ∂♦ , ∂¯♦
Discrete Complex Analysis on Planar Quad-Graphs 71

Definition h : V (♦0 ) → C is said to be discrete holomorphic if h is discrete holo-

morphic at all v ∈ V (Λ0 )\V (∂Λ0 ).

Note that in the rhombic case, our definition coincides with the one used by
Chelkak and Smirnov in [6]. As an immediate consequence of the definition, we
obtain a discrete Morera’s theorem.

Proposition 2.4 Functions

f : V (Λ0 ) →
C and h : V (♦0 ) → C are discrete holo-
morphic if and only if P f dz = 0 and P hdz = 0 for all discrete contours P on
X 0.

Proof Clearly, Pv f dz = f (v) Pv dz = 0 for any discrete elementary cycle Pv

around a vertex v of V (Λ0 ). Similarly, PQ hdz = 0 for any Q ∈ V (♦0 ). Using

Lemma

2.3 and the
definition of ∂¯♦ , f and h are discrete holomorphic if and only
if P f dz = 0 and P hdz = 0 for all discrete elementary cycles P. To conclude the
proof, we observe that any integration along a discrete contour can be decomposed
into integrations along discrete elementary cycles since by definition, a discrete con-
tour is the boundary of a topological disk in F(X 0 ).

The discrete derivatives of constant functions on V (♦) vanish. As an analog of

Proposition 2.1, we prove that the discrete derivatives ∂♦ , ∂¯♦ locally approximate
their smooth counterparts correctly up to order one if the quadrilaterals in V (♦) are
identified with the midpoints of their black or white edges. In a parallelogram-graph,
these two midpoints coincide for each face, which then gives a global approximation
statement. Note that even for rhombic quad-graphs, the discrete derivatives ∂♦ , ∂¯♦
generally do not coincide with the smooth derivatives in order two.

Proposition 2.5 Let v ∈ V (Λ), and let h be a complex function on all faces incident
to v. As illustrated in Fig. 3b, we counterclockwise enumerate them by Q 1 , . . . , Q k ,
where k is the degree of v in Λ, and their vertices adjacent to v by v1 , v2 , . . . , vk ,

vk+1 
= v1 . Let Q̂ s = (vs−1 + vs )/2. Then, if h(Q s ) = Q̂ s for all s, ∂¯♦ h(v) = 0 and
∂♦ h(v) = 1 hold true.

Proof

4 hdz = 2h(Q s )(vs − vs−1

)= 
(vs−1 + vs )(vs − vs−1

)
Pv Q s ∼v Q s ∼v
 2  2 
= vs − vs−1

= 0,
Q s ∼v
  2 
2  
4 hd z̄ = 
(vs−1 + vs )(vs − vs−1

)= v  − v  − 2i Im v v̄
s s−1 s s−1
Pv Q s ∼v Q s ∼v
  
= −2i Im vs v̄s−1 = −8iar(Fv ).
Q s ∼v
72 A.I. Bobenko and F. Günther

Thus, ∂¯♦ h(v) = 0 and ∂♦ h(v) = 1. Here, we have used that by definition, ar(Fv ) is
half of the Euclidean area of the polygon v1 v2 . . . vk , so ar(F) equals

1 1     1   

area(vvs−1 vs ) = Im vs − v vs−1 −v = Im vs v̄s−1 ,
2 Q ∼v 4 Q ∼v 4 Q ∼v
s s s

    
using that Q s ∼v vv̄s−1 + v̄vs = Q s ∼v vv̄s + v̄vs is real.

In [6], Chelkak and Smirnov used averaging operators to map functions on V (Λ)
to functions on V (♦)
and vice versa. On parallelogram-graphs, the averaging opera-
tor m( f )(Q) := v∼Q f (v)/4 actually maps discrete holomorphic functions f on
V (Λ) to discrete holomorphic functions on V (♦). Our proof will be similar as the
one for rhombic quad-graphs in [6]. Note that discrete holomorphic functions on
V (♦) cannot be averaged to discrete holomorphic functions on V (Λ) in general, so
the averaging operator of Chelkak of Smirnov that mapped functions on V (♦) to
functions on V (Λ) did not preserve discrete holomorphicity.

Proposition 2.6 Let Λ be a parallelogram-graph and f : V (Λ) → C be discrete

holomorphic. Then, m( f ) : V (♦) → C is discrete holomorphic.

Proof Let us consider the star of the vertex v ∈ V (Λ) and use the notation we used in
Proposition 2.5 (illustrated by Fig. 3b). Since f is discrete holomorphic, the discrete
Cauchy-Riemann equation is satisfied on any Q s ∼ v. Therefore, we can express

f (vs ) in terms of f (v), f (vs ) and f (vs−1 ). Plugging this in the definition of the
averaging operator, we obtain
 
vs − v + vs − vs−1  vs − v − vs + vs−1 
4m( f )(Q s ) = 2 f (v) +  f (v ) −  f (vs−1 )
vs − vs−1 s
vs − vs−1
v − v v − v
= 2 f (v) + 2  s  f (vs ) − 2 s−1  f (vs−1 
).
vs − vs−1 vs − vs−1

 
Here, we have used the properties vs − vs−1 = vs − v and vs − vs = vs−1 − v of the
parallelogram Q s . Hence, m( f ) is discrete holomorphic at v by definition due to
 
4 m( f )dz = 2 f (v) dz + (vs − v) f (vs ) − 
(vs−1 
− v) f (vs−1 ) = 0.
Pv Pv Q s ∼v Q s ∼v

Remark As mentioned above, our main interest lies in functions that are defined
either on the vertices or the faces of the quad-graph. Now, extending f : V (Λ) → C
to a complex function on F(X ) by using its average m( f ) on V (♦) seems to be
an option. However, functions on V (Λ) and on V (♦) behave differently. In Corol-
lary 2.11 we will see that ∂Λ f is discrete holomorphic if f is, but ∂♦ m( f ) does not
need to be discrete holomorphic in general. So to make sense of differentiating twice,
we can only consider functions on V (Λ).
Discrete Complex Analysis on Planar Quad-Graphs 73

Definition Let f 1 , f 2 : V (Λ0 ) → C and h 1 , h 2 : V (♦0 ) → C. Their discrete scalar

products are defined as
 
1 1
f 1 , f 2  := − f 1 f¯2 ΩΛ and h 1 , h 2  := − h 1 h̄ 2 Ω♦ ,
2i 2i
F(X 0 ) F(X 0 )

whenever the right hand side converges absolutely.

Proposition 2.7 −∂♦ and −∂¯♦ are the formal adjoints of ∂¯Λ and ∂Λ , respectively.
That is, if f : V (Λ) → C or h : V (♦) → C is compactly supported, then

∂Λ f, h + f, ∂¯♦ h = 0 = ∂¯Λ f, h + f, ∂♦ h.

Proof In Lemma 2.3, we showed how the discrete derivative ∂Λ f (Q) can be
expressed as a contour integration around the face of X associated to Q ∈ V (♦).
Using this, the definitions of ΩΛ and Ω♦ , and ∂♦ h̄ = ∂¯♦ h, we get

∂Λ f, h + f, ∂¯♦ h = ∂Λ f (Q)h̄(Q)ar(FQ ) + f (v)∂¯♦ h(v)ar(Fv )
Q∈V (♦) v∈V (Λ)
 
i i
= h̄(Q) f d z̄ + f (v) h̄d z̄
2 2
Q∈V (♦) PQ v∈V (Λ) Pv

i
= f h̄d z̄ = 0,
2
P

where P is a large contour enclosing all the vertices of Λ and ♦ where f or h do

not vanish. In particular, f h̄ vanishes in a neighborhood of P. In the same way,
∂¯Λ f, h + f, ∂♦ h = 0.

Remark Note that in their work on discrete complex analysis on rhombic quad-
graphs, Kenyon [16] and Mercat [20] did not give explicit formulae for the discrete
derivatives, but defined −∂♦ and −∂¯♦ instead as the formal adjoints of the discrete
derivatives ∂¯Λ and ∂Λ , respectively. In contrast, we derived the formal adjoint property
from our explicit formulae for the discrete derivatives.

In Corollary 2.11, we will prove that ∂Λ f is discrete holomorphic if the function

f : V (Λ) → C is. Conversely, we can find discrete primitives of discrete holomor-
phic functions on subgraphs ♦0 ⊆ ♦ that form a simply-connected closed region,
extending the corresponding result for rhombic quad-graphs given by Chelkak and
Smirnov [6].

Proposition 2.8 Let ♦0 ⊆ ♦ form a simply-connected closed region. Then, for  any
discrete holomorphic function h on V (♦0 ), there is a discrete primitive f := h on
V (Λ0 ), i.e., f is discrete holomorphic and ∂Λ f = h. f is unique up to two additive
constants on Γ0 and Γ0∗ .
74 A.I. Bobenko and F. Günther

Proof Since h is discrete holomorphic, P hdz = 0 for any discrete contour P in

X 0 by discrete Morera’s Theorem 2.23. Therefore, hdz can be integrated to a well-
defined function f X on V (X ) that is unique up to an additive constant. The equation
f X ((v + w) /2) = ( f (v) + f (w)) /2 for any edge (v, w) of Λ defines a function f
on V (Λ). Indeed, since ♦0 forms a simply-connected closed region, it suffices to
check the compatibility of the equations defining f just for one quadrilateral face
Q. Now, the differences of f along the black and white diagonals of Q are given by
integration of hdz along the edges of X that are parallel to these diagonals. Since hdz
is a discrete one-form of type ♦, the two integrations along the edges of X parallel to
one diagonal are the same if they are oriented the same, and compatibility follows.
Given f X , f is unique up to another additive constant.
In summary, f is unique up to two additive constants that can be chosen indepen-
dently on Γ0 and Γ0∗ . By construction, f satisfies

f (b+ ) − f (b− ) f (w+ ) − f (w− )

= h(Q) =
b+ − b− w+ − w−

on any quadrilateral Q ∈ V (♦0 ). So f is discrete holomorphic and ∂Λ f = h.

2.3 Discrete Exterior Calculus

Our treatment of discrete exterior calculus is similar to the approach of Mercat in
[19–21], but differs in some aspects. The main differences are due to our differ-
ent notation of multiplication of functions with discrete one-forms, which allows
us to define a discrete exterior derivative on a larger class of discrete one-forms
in Sect. 2.3.1. It coincides with Mercat’s discrete exterior derivative in the case of
discrete one-forms of type ♦ that Mercat considers. In contrast, our definitions are
based on a coordinate representation and mimic the smooth case. Eventually, they
lead to essentially the same definitions of a discrete wedge product in Sect. 2.3.2 and
a discrete Hodge star in Sect. 2.3.3 as in [21].

2.3.1 Discrete Exterior Derivative

Definition Let f : V (Λ0 ) → C, h : V (♦0 ) → C. We define the discrete exterior
derivatives d f and dh as the discrete one-forms on oriented edges of X 0 given by

d f := ∂Λ f dz + ∂¯Λ f d z̄ and dh := ∂♦ hdz + ∂¯♦ hd z̄.

Definition Let ω be a discrete one-form defined on all boundary edges of a face Fv

of the medial graph X corresponding to v ∈ V (Λ) or on all four boundary edges of a
face FQ of X corresponding to Q ∈ F(Λ). In the first case, we write ω = pdz + qd z̄
with functions p, q defined on all faces of Λ that are incident to v, and in the second
case, we write ω = pdz + qd z̄ with functions p, q defined on all vertices of Λ that
are incident to Q. The discrete exterior derivative dω on Fv or FQ is given by
Discrete Complex Analysis on Planar Quad-Graphs 75
   
dω| Fv := ∂♦ q − ∂¯♦ p ΩΛ and dω| FQ := ∂Λ q − ∂¯Λ p Ω♦ .

The representation of ω as pdz + qd z̄ ( p, q defined on edges of X ) we have used

above may be nonunique. However, dω is well-defined by the following discrete
Stokes’ theorem that also justifies our definition of d f and dω. Note that Mercat
defined the discrete exterior derivative by the discrete Stokes’ theorem [19].
Theorem 2.9 Let f : V (Λ0 ) → C and ω be a discrete one-form defined on oriented
edges of X 0 . Then, for any directed edge e of X 0 starting in the midpoint of the edge
 
vv− and ending in the midpoint of the edge vv+ of Λ0 and for any finite collection of
faces F of X 0 with counterclockwise oriented boundary ∂ F we have:
  ) − f (v )  )  )  
f (v+ − f (v) + f (v+ f (v) + f (v−
df = = − and dω = ω.
2 2 2
e F ∂F

 
Proof Let v− be the other vertex of the quadrilateral Q with vertices v, v− and v+ .
Without loss of generality, let v be white. Since d f = ∂Λ f dz + ∂¯Λ f d z̄, e d f equals

v − v−
 v − v−

∂Λ f + + ∂¯Λ f +
2 2
 
   
1  ) − f (v )) + 1 λ̄ v+ − v− + λ v+ − v− ( f (v) − f (v ))
= (λ Q + λ̄ Q )( f (v+ − Q Q −
2 2 v − v− v − v−
 
 ) − f (v )
f (v+ v − v−  ) − f (v )
f (v+
−
= + Re λ̄ Q + ( f (v) − f (v− )) = −
.
2 v − v− 2

To get to the third line, we used λ Q + λ̄ Q = 1, and for the last step we used
     

v+ − v− π   
arg λ̄ Q = arg ± exp i ϕ Q − exp −iϕ Q = ±π/2.
v − v− 2
 
The sign depends on whether v, v− , v− , v+ denote the corners of Q in clockwise or
counterclockwise order. In either case, the expression inside arg is purely imaginary.
The second identity has to be shown just for one single face of X 0 . Let us write
ω = pdz + qd z̄ on all edges of X 0 that are boundary edges of FQ or Fv , where p, q
are functions defined on the vertices of the quadrilateral Q ∈ V (♦0 ) or on the faces
incident to v ∈ V (Λ0 ). Then, by the representation of ∂Λ , ∂¯Λ as discrete contour
integrals in Lemma 2.3 and the definition of the discrete derivatives ∂♦ , ∂¯♦ ,
  
   
dω = ∂♦ q − ∂¯♦ p Ω♦ = −2iar(FQ ) ∂♦ q − ∂¯♦ p = ( pdz + qd z̄) ,
FQ FQ ∂ FQ
  
   
dω = ∂Λ q − ∂¯Λ p ΩΛ = −2iar(Fv ) ∂Λ q − ∂¯Λ p = ( pdz + qd z̄) .
Fv Fv ∂ Fv
76 A.I. Bobenko and F. Günther

Definition Let ♦0 ⊆ ♦ form a simply-connected closed region. A discrete one-form

ω defined on oriented edges of X 0 is said to be closed if dω ≡ 0.

Note that if ω is a discrete one-form of type ♦, then F dω = 0 for any face F
corresponding to a face of Λ. Examples for closed discrete one-forms are discrete
exterior derivatives of complex functions on V (Λ):
Proposition 2.10 Let f : V (Λ0 ) → C. Then, dd f = 0 on any face Fv of X 0 corre-
sponding to v ∈ V (Λ0 )\V (∂Λ0 ).

Proof By discrete Stokes’ Theorem 2.9, we have to show P d f = 0 for any dis-
crete elementary cycle P in X 0 in order to prove dd f = 0. Since d f is of type
♦, the statement is trivially true if P = PQ for Q ∈ V (♦0 ). So let P = Pv for
v ∈ V (Λ0 )\V (∂Λ0 ). Using discrete Stokes’ Theorem 2.9 again,
 f (vs ) − f (vs−1

)
df = = 0.
Q ∼v
2
Pv s

An immediate corollary of the last proposition is the commutativity of discrete deriv-

atives, generalizing the known result for rhombic quad-graphs in [6].
Corollary 2.11 Let f : V (Λ0 ) → C. Then, ∂♦ ∂¯Λ f (v) = ∂¯♦ ∂Λ f (v) for all vertices
v ∈ V (Λ0 )\V (∂Λ0 ). In particular, ∂Λ f is discrete holomorphic if f is discrete holo-
morphic.
Proof Due to the preceding Proposition
 2.10 and the
 definition of the discrete deriv-
ative, the equation 0 = dd f = ∂♦ ∂¯Λ f − ∂¯♦ ∂Λ f ΩΛ holds on all faces of X 0 cor-
responding to a vertex of Λ0 that is not on the boundary ∂Λ0 . The claim follows
since ΩΛ is nonzero on these faces.
Remark Note that even in the generic rhombic case, ∂Λ ∂¯♦ h does not always equal
∂¯Λ ∂♦ h for a function h : V (♦) → C [6]. Hence, ddh = 0 cannot hold for such
functions h in general.
Corollary 2.12 Let f : V (Λ0 ) → C. Then, f is discrete holomorphic if and only if
d f = pdz for some p : V (♦0 ) → C. In the case that f is discrete holomorphic, p
is discrete holomorphic as well.
Proof Since all quadrilaterals Q ∈ V (♦0 ) are nondegenerate, the representation
of d f |∂ FQ as pdz + qd z̄ is unique (see Lemma 2.14 below). Clearly, we have
d f = ∂Λ f dz + ∂¯Λ f d z̄. It follows that f is discrete holomorphic at Q if and only if
d f |∂ FQ = pdz.
Assuming that d f = pdz for some p : V (♦0 ) → C, dd f = 0 on any face Fv of
X 0 corresponding to v ∈ V (Λ0 )\V (∂Λ0 ) by Proposition 2.10. Thus, ∂¯♦ p(v) = 0 for
any such v and f is discrete holomorphic.
Definition Let ♦0 ⊆ ♦ form a simply-connected closed region. A discrete one-form
ω defined on the oriented edges of X 0 is discrete holomorphic if ω = pdz for some
p : V (♦0 ) → C and dω = 0.
Discrete Complex Analysis on Planar Quad-Graphs 77

Remark This notion recurs in the more general setting of discrete Riemann surfaces
in [1]. By Corollary 2.12, d f is discrete holomorphic if f is, and by Proposition 2.8
on the existence of a discrete primitive for discrete holomorphic functions defined
on the vertices of a subset ♦0 ⊆ ♦ that forms a simply-connected closed region, any
discrete holomorphic one-form ω defined on the oriented edges of X 0 is the discrete
exterior derivative of a discrete holomorphic function on V (Λ0 ).

Due to Chelkak and Smirnov [6], one of the unpleasant facts of all discrete theories
of complex analysis is that (pointwise) multiplication of discrete holomorphic func-
tions does not yield a discrete holomorphic function in general. We define a product
of complex functions on V (Λ) that is defined on V (X ) and a product of complex
functions on V (Λ) with functions on V (♦) that is defined on E(X ). In general, the
product of two discrete holomorphic functions is not discrete holomorphic according
to the classical quad-based definition (on planar quad-graphs different from Λ), but it
will be discrete holomorphic in the sense that a discretization of its exterior derivative
is closed and is of the form pdz, p defined on the edges of the medial graph of the
new quad-graph, or in the sense that it fulfills a discrete Morera’s theorem.

Proposition 2.13 Let f, g : V (Λ) → C and h : V (♦) → C.

(i) f dg + gd f is a closed discrete one-form.
(ii) If f and h are discrete holomorphic, then f hdz is a closed discrete one-form.

Proof (i) Let ω := f dg + gd f . By Proposition 2.10, dd f = 0 and ddg = 0, i.e.,

d f and dg are closed. Thus,
  
ω = f (v) dg + g(v) df = 0
∂ Fv ∂ Fv ∂ Fv

for any face Fv corresponding to v ∈ V (Λ). Using Lemma 2.3 that relates discrete
derivatives with discrete contour integration,
 
 
2iar(FQ ) ω = 2iar(FQ ) f ∂Λ gdz + f ∂¯Λ gd z̄ + g∂Λ f dz + g ∂¯Λ f d z̄
∂ FQ ∂ FQ
 
= ∂¯Λ f ∂Λ g − ∂Λ f ∂¯Λ g + ∂¯Λ g∂Λ f − ∂Λ g ∂¯Λ f (Q) = 0

for any face FQ corresponding to Q ∈ F(Λ). It follows by discrete Stokes’ Theo-

rem 2.9 that dω = 0.
(ii) By discrete Morera’s Theorem 2.23,
discrete holomorphicity

of f and h
implies that f dz and hdz are closed. Thus, ∂ Fv f hdz = f (v) ∂ Fv hdz = 0 as well

as ∂ FQ f hdz = h(Q) ∂ FQ f dz = 0 for any faces Fv and FQ of X corresponding to

v ∈ V (Λ) and Q ∈ F(Λ). Therefore, f hdz is closed.

Remark In particular, for any f, g : V (Λ) → C a product f · g : V (X ) → C can

be defined by integration of f dg + gd f . Note that this product f · g is defined up to
78 A.I. Bobenko and F. Günther

an additive constant. Furthermore, f · h : E(X ) → C can be defined by “pointwise”

multiplication. If f, g, h are discrete holomorphic, then f dg + gd f = pdz is closed,
where p = f · ∂Λ g + g · ∂Λ f : E(X ) → C, and so to say a discrete holomorphic
one-form, meaning that f · g is discrete holomorphic in this sense. Similarly, f hdz is
closed, so f · h is discrete holomorphic in the sense that a discrete Morera’s theorem
holds true. Even though f · g is defined on the vertices of the dual of a planar quad-
graph different from Λ, as well is f · h on the dual of a different planar quad-graph,
these products are generally not discrete holomorphic everywhere according to the
quad-based definition of discrete holomorphicity on the dual of a bipartite quad-graph
given by the definition in Sect. 2.2.3. To define the mentioned planar quad-graphs,
we identify Q ∈ V (♦) with such a point in the interior of the face Q that all line
segments connecting it to the four corners of Q lie inside the quadrilateral.
First, f · g is a complex function on the vertices of X . The medial graph X is
the dual of the bipartite quad-graph with vertex set V (Λ)  V (♦), edges connecting
points Q ∈ V (♦) with all incident vertices v ∈ V (Λ), and faces in one-to-one cor-
respondence to edges of Λ. But even if f and g are discrete holomorphic on V (Λ),
f · g does not need to be a discrete holomorphic function on the faces of the quad-
graph we just defined. For example, consider f (v) = 0 if v is black and f (v) = 1
if v is white and a discrete holomorphic g that is not biconstant. Then, the product
f · g is not discrete holomorphic at all Q ∈ V (♦) (seen as vertices of the quad-graph
described above) where ∂Λ g(Q) = 0.
Second, f · h is a complex function on the edges of X , so it is a function on the
vertices of the medial graph of X . The medial graph of the medial graph of Λ is the
dual of the quad-graph with vertex set (V (Λ) ∪ V (♦))  V (X ), edges connecting
points v ∈ V (Λ) or Q ∈ V (♦) with the midpoints of all incident edges, and each
face being in one-to-one correspondence to an edge of X . Since f hdz is closed,
f · h is discrete holomorphic on the new quad-graph at vertices of Λ or ♦ by discrete
Morera’s Theorem 2.23. But there is no need for f · h to be discrete holomorphic
at vertices of X , even for constant h. For example, take the function f defined by
f (v) = 0 if v is black and f (v) = 1 if v is white.
In summary, we defined products f · g and f · h, where f, g : V (Λ) → C and
h : V (♦) → C are discrete holomorphic, that are local (on each vertex, they depend
just on the values of f and g in a small neighborhood) and discrete holomorphic at
least in the sense that its discrete exterior derivative is closed and of the form pdz or
in the sense that it fulfills a discrete Morera’s theorem.
Somehow missing is a product h · h  , where h, h  : V (♦) → C. In the general
case, we do not know an appropriate product so far. But we want to point out that
Chelkak and Smirnov showed in [7] that  for so-called
 spin holomorphic functions
h, h  , the pointwise product satisfies Re ∂¯♦ h · h  ≡ 0.

2.3.2 Discrete Wedge Product

Following Whitney [27], Mercat defined in [19] a discrete wedge product for discrete
one-forms living on the edges of Λ. Then, the discrete exterior derivative defined by
Discrete Complex Analysis on Planar Quad-Graphs 79

a discretization of Stokes’ theorem is a derivation for the discrete wedge product.

However, a discrete Hodge star cannot be defined on Λ. To circumvent this problem,
Mercat used an averaging map to relate discrete one-forms on the edges of Λ with
discrete one-forms on the edges of Γ and Γ ∗ , i.e., discrete one-forms of type ♦.
Then, he could define a discrete Hodge star; however, the discrete exterior derivative
was not a derivation for the now heterogeneous discrete wedge product.
In this section, a different interpretation of the discrete wedge product is proposed.
Still, the notions of Mercat in [19–21] are recovered. Starting with discrete one-forms
of type ♦ that are defined on the edges of X , a discrete wedge product on (half of)
the faces of X is defined. This definition is different from Whitney’s [27] and has
the advantage that both a discrete wedge product and a discrete Hodge star can be
defined on the same structure. In addition, the discrete exterior derivative is now a
derivation for the discrete wedge product in a well-defined sense, see Theorem 2.16.
It turns out that Theorem 2.16 is a powerful tool leading to discretizations of Green’s
identities in Sect. 2.4.1 and of a Cauchy’s integral formula for the discrete derivative
of a discrete holomorphic function in Sect. 2.6.

Lemma 2.14 Let ω be a discrete one-form of type ♦ defined on the oriented edges of
X 0 . Then, there is a unique representation ω = pdz + qd z̄ with p, q : V (♦0 ) → C.
On a quadrilateral Q ∈ V (♦0 ), p and q are given by
   
ω e∗ ω eω ∗ ω
p(Q) = λ Q e
+ λ̄ Q ∗ and q(Q) = λ̄ Q + λQ e ∗ .
e e ē ē

Here, e is an oriented edge of X 0 parallel to a black edge of Γ0 , and e∗ is parallel to

a white edge of Γ0∗ .

Proof First, we show that a representation ω|∂ FQ = pdz + qd z̄ exists for any face
 Q ∈ V (♦
FQ of X 0 corresponding to a quadrilateral  0 ). Given ω, we have to solve the
system of linear equations e Q ω = p e Q dz + q e Q d z̄ for all four boundary edges
e Q of FQ . Since ω is of type ♦, we just have to consider two equations, namely
one for a boundary edge eb of FQ parallel to a black edge of Γ0 and one equation
for a boundary edge ew parallel to a white edge of Γ0∗ . Since all quadrilaterals are
nondegenerate, the diagonals are not parallel to each other and it follows that the pair
(dz, d z̄) gives different values when integrated over eb and ew . Thus, this system
of two linear equations in two variables is nondegenerate. It follows that p, q are
uniquely defined on V (♦0 ).
Furthermore, we can find for any quadrilateral Q ∈ V (♦0 ) ∼ = F(Λ0 ) a function
f that is defined on the vertices b± , w± of Q such that 2 e ω = f (b+ ) − f (b− ) and
2 e∗ ω = f (w+ ) − f (w− ), where e is one of the two oriented edges of X 0 going
from the midpoint of b− and w± to the midpoint of b+ and w± , and e∗ is one of the
two edges connecting the midpoint of w− and b± with the midpoint of w+ and b± . By
discrete Stokes’ Theorem 2.9, we get ω|∂ FQ = d f = pdz + qd z̄ with p = ∂Λ f (Q)
and q = ∂¯Λ f (Q). Replacing the differences of f in the definition of the discrete
derivative by discrete integrals of ω yields the desired result.
80 A.I. Bobenko and F. Günther

Definition Let ω = pdz + qd z̄ and ω = p  dz + q  d z̄ be two discrete one-forms

of type ♦ defined on the oriented edges of X 0 . Here, p, p  , q, q  : V (♦0 ) → C are
given by the above Lemma 2.14. Then, the discrete wedge product ω ∧ ω is defined
as the discrete two-form of type ♦ defined on F(X 0 ) that equals
 
pq  − qp  Ω♦

on faces of X corresponding to interior faces of the quad-graph Λ0 .

Remark Note that if one considers dz and d z̄ as discrete one-forms of type ♦, then
Ω♦ = dz ∧ d z̄.
Proposition 2.15 Let F be a face of X 0 corresponding to Q ∈ F(Λ0 ), and let e, e∗
be oriented edges of X parallel to the black and white diagonal of the quadrilateral
Q, respectively, such that Im (e∗ /e) > 0. If ω, ω are discrete one-forms of type ♦
defined on the oriented edges of ∂ F, then
    
ω ∧ ω = 2 ω ω − 2 ω ω .
F e e∗ e∗ e

Proof Both sides of the equation are bilinear and antisymmetric in ω, ω . Hence,
it suffices to check the identity for ω = dz, ω = d z̄. On the left hand side, we get
 
F ω ∧ ω = −2iar(F). This equals the right hand side

2eē∗ − 2e∗ ē = 4i Im(eē∗ ) = −i|2e||2e∗ | sin(ϕ Q ) = −2iar(F).

Remark Since the complex numbers e and e∗ are just half of the oriented diagonals,
the above definition of the discrete wedge product is essentially the same as the one
given by Mercat in [19–21].
The discrete exterior derivative is a derivation for the discrete wedge product if
one considers functions on Λ and discrete one-forms of type ♦:
Theorem 2.16 Let f : V (Λ0 ) → C and ω be a discrete one-form of type ♦ defined
on the oriented edges of X 0 . Then, the following identity holds on F(X 0 ):

d( f ω) = d f ∧ ω + f dω.

Proof Let ω = pdz + qd z̄ with p, q : V (♦0 ) → C given by Lemma 2.14. If Fv and

FQ are faces of X 0 corresponding to a vertex v and a face Q of Λ0 , then
     
d( f ω)| Fv = f (v) ∂♦ q (v) − f (v) ∂¯♦ p (v) ΩΛ = f dω| Fv ,
   
d( f ω)| FQ = q(Q) (∂Λ f ) (Q) − p(Q) ∂¯Λ f (Q) Ω♦ = (d f ∧ ω)| FQ .
Discrete Complex Analysis on Planar Quad-Graphs 81

But (d f ∧ ω)| Fv = 0 since Ω♦ | Fv = 0 and f dω| FQ = 0 since ω is of type ♦, so

d( f ω) = d f ∧ ω + f dω.

Remark In [19], Mercat formulated an analog of the above Theorem 2.16 in a setting
where discrete one-forms are defined on edges of Λ. In the setting of discrete one-
forms defined on edges of Γ and Γ ∗ , the claim d( f ω) = d f ∧ ω + f dω could not
be well-defined.

Above, a discrete wedge product just of two discrete one-forms of type ♦ is

defined. Actually, we could define a discrete wedge product of two discrete one-
forms of type Λ in essentially the same way, getting a discrete two-form of type
Λ. Then, the analog of Theorem 2.16 would be true for this kind of discrete wedge
product and functions on V (♦0 ). Also the discrete Hodge star of a discrete one-form
in the next section could be defined not only for those of type ♦. However, there
exist no analogs of Propositions 2.15 and 2.18. These propositions imply that the
discrete wedge product as well as the discrete Hodge star of discrete one-forms of
type ♦ can be defined in a chart-independent way. This enables one to consider these
objects on discrete Riemann surfaces, see [1]. There are no such statements if one
chooses discrete one-forms of type Λ. In fact, a discrete one-form of type Λ cannot be
canonically defined on a discrete Riemann surface as opposed to discrete one-forms
of type ♦. So since our interest lies in the latter, we do not define a discrete wedge
product or a discrete Hodge star for discrete one-forms of type Λ.

2.3.3 Discrete Hodge Star

Definition Let f : F(Λ0 ) → C and h : V (♦0 ) → C, let ω = pdz + qd z̄ be a dis-

crete one-form of type ♦ defined on oriented edges of X 0 with p, q : V (♦0 ) → C,
and let Ω1 , Ω2 : F(X 0 ) → C be discrete two-forms of type Λ and ♦. Then, the
discrete Hodge star is given by

1 1
 f := − f ΩΛ ; h := − hΩ♦ ; ω := −i pdz + iqd z̄;
2i 2i
Ω1 Ω2
Ω1 := −2i ; Ω2 := −2i .
ΩΛ Ω♦

If ω and ω are both discrete one-forms of type ♦ defined on oriented edges of X 0 ,

we define their discrete scalar product

ω, ω  := ω ∧ ω̄ ,
F(X 0 )

whenever the right hand side converges absolutely. Similarly, a discrete scalar product
for discrete two-forms of the same type is defined.
82 A.I. Bobenko and F. Günther

Remark
  in the classical theory, the Hodge star corresponds to a π/2-rotation:
As
ie ω = e ω where ω is a discrete one-form of type ♦, e an oriented edge of X and
ie its (virtual) image under π/2-rotation around the origin.

Corollary 2.17 The following statements are true:

(i) 2 = −Id on discrete one-forms of type ♦ defined on oriented edges of X 0 .
(ii) 2 = Id on complex functions on V (Λ0 ) or V (♦0 ) and discrete two-forms
defined on F(X 0 ) of type Λ or ♦.
(iii) f : V (Λ0 ) → C is discrete holomorphic if and only if d f = −id f .
(iv) f 1 , f 2  = F(X 0 ) f 1  f 2 for functions f 1 , f 2 : V (Λ0 ) → C and

h 1 , h 2  = F(X 0 ) h 1 h 2 for functions h 1 , h 2 : V (♦0 ) → C.
(v) ·, · is a Hermitian scalar product on discrete differential forms (of type Λ or
of type ♦).

Proposition 2.18 Let Q ∈ V (♦), and let e, e∗ be oriented edges of X parallel to

the black and white diagonal of Q, respectively, such that Im (e∗ /e) > 0. If ω is a
discrete one-form of type ♦ defined on the oriented edges of the boundary of the face
of X corresponding to Q, then
  
  |e|
ω = cot ϕ Q ω− ∗   ω,
|e | sin ϕ Q
e e e∗
  
∗
|e |  
ω =   ω − cot ϕ Q ω.
|e| sin ϕ Q
e∗ e e∗

Proof Both sides of any of the two equations are linear and behave the same under
complex conjugation. Thus, it suffices to check the statement for ω = dz. Hence, it
remains to show that
  |e| |e∗ |  
−ie = cot ϕ Q e −   e∗ and e∗ =   e − cot ϕ Q e∗ .
∗
|e | sin ϕ Q |e| sin ϕ Q

Now, both sides of the first equation behave the same under scaling and simultaneous
rotation of e and e∗ , the same statement
  is true for the second equation. Thus, we
may assume e = 1 and e∗ = cos ϕ Q + i sin(ϕ Q ). Multiplying both equations by
sin(ϕ Q ) gives the equivalent statements
     
−i sin(ϕ Q ) = cos ϕ Q − cos ϕ Q + i sin(ϕ Q ) ,
 
−i sin(ϕ Q ) exp(iϕ Q ) = 1 − cos ϕ Q exp(iϕ Q ).
 
Both equations are true, noting that cos ϕ Q − i sin(ϕ Q ) = exp(−iϕ Q ).

Remark Proposition 2.18 shows that our definition of a discrete Hodge star on dis-
crete one-forms coincides with Mercat’s definition given in [21]. But on discrete
Discrete Complex Analysis on Planar Quad-Graphs 83

two-forms and complex functions, our definition of the discrete Hodge star includes
an additional factor of the area of the corresponding face of X .

Proposition 2.19 δ := −  d is the formal adjoint of the discrete exterior

derivative d: Let f : V (Λ) → C, and let ω be a discrete one-form of type ♦ defined
on the oriented edges of X and Ω : F(X ) → C a discrete two-form of type Λ. Assume
that all of them are compactly supported. Then,

d f, ω = f, δω and dω, Ω = ω, δΩ.

Proof By the assumption that all forms are compactly supported, we can take a large
enough finite ♦0 ⊆ ♦ that forms a simply-connected closed region such that f, ω, Ω
vanish outside Λ0 , X 0 , ♦0 and ω is zero on the boundary ∂ X 0 . By discrete Stokes’
Theorem 2.9 and Theorem 2.16 that states that the discrete exterior derivative is a
derivation for the discrete wedge product,
   
0= f  ω̄ = d( f  ω̄) = f d  ω̄ + d f ∧ ω̄ = f, d  ω + d f, ω,
∂ X0 F(X 0 ) F(X 0 ) F(X 0 )
   
0= Ω̄ω = d(Ω̄ω) = Ω̄dω + (d  Ω̄) ∧ ω = dω, Ω − ω, δΩ.
∂ X0 F(X 0 ) F(X 0 ) F(X 0 )

In the last equalities, we have used Corollary 2.17(ii) and (iv) (the basic properties of
the discrete Hodge star) and the observation that complex conjugation commutes with
the discrete Hodge star and the discrete exterior derivative. The latter observation
immediately follows from the definitions that mimic the classical theory.

2.4 Discrete Laplacian

The discrete Laplacian and the discrete Dirichlet energy on general quad-graphs
were introduced by Mercat in [21]. Later, Skopenkov reintroduced these definitions
in [23], taking the same definition in a different notation. In our discussion of the
discrete Laplacian in Sect. 2.4.1, we follow the classical approach of Mercat (up to
sign) and adapt it to our notations. A feature of the medial graph approach is that it
allows to formulate a discrete analog of Green’s first identity from which discrete
Green’s second identity immediately follows.
In Sect. 2.4.2, the discrete Dirichlet energy is investigated. In particular, in
Theorem 2.30 it is shown how uniqueness and existence of solutions to the dis-
crete Dirichlet boundary value problem imply surjectivity of the discrete derivatives
and the discrete Laplacian. We conclude this section with a result concerning the
asymptotics of discrete harmonic functions.
84 A.I. Bobenko and F. Günther

2.4.1 Definition and Basic Properties

Definition The discrete Laplacian on functions f : V (Λ) → C, discrete one-forms

of type ♦, or discrete two-forms of type Λ is defined as the linear operator

 := −δd − dδ = d  d + d  d  .

For a connected subgraph ♦0 ⊆ ♦ and f : V (Λ0 ) → C,  f is still defined by

the formula above as a complex function on V (Λ0 )\V (∂Λ0 ). f is said to be discrete
harmonic at v ∈ V (Λ0 )\V (∂Λ0 ) if  f (v) = 0. f is discrete harmonic if it is discrete
harmonic at all such v.

The following factorization of the discrete Laplacian in terms of discrete deriv-

atives generalizes the corresponding result given by Chelkak and Smirnov in [6] to
general quad-graphs. The local representation of  f at v ∈ V (Λ) is, up to a factor
involving the area of the face Fv of X corresponding to v, the same as Mercat’s [21].

Corollary 2.20 Let f : V (Λ0 ) → C. Then,  f (v) = 4∂♦ ∂¯Λ f (v) = 4∂¯♦ ∂Λ f (v)
for all vertices v ∈ V (Λ0 )\V (∂Λ0 ) and

1 1   
 f (v) = |ρs |2 ( f (vs ) − f (v)) + Im (ρs ) f (vs ) − f (vs−1
 ) .
2ar(Fv ) Re (ρs )
Q s ∼v

Here, ρs := ρ Q s if v is black, and ρs := 1/ρ Q s if v is white.

In particular, Re( f ) ≡  Re( f ) and Im( f ) ≡  Im( f ).

Proof Since the definitions of the discrete Hodge star and the discrete exterior deriv-
ative mimic the classical theory and ∂♦ ∂¯Λ f (v) = ∂¯♦ ∂Λ f (v) by Corollary 2.11,

 f (v) = d  d f (v) = 2∂♦ ∂¯Λ f (v) + 2∂¯♦ ∂Λ f (v) = 4∂♦ ∂¯Λ f (v) = 4∂¯♦ ∂Λ f (v)

holds exactly as in the smooth setting.

For the second statement, let us assume without loss of generality that v ∈ V (Γ0 ).
Then, we have to show that

1  |ρ Q |  

 f (v) = s
( f (vs ) − f (v)) − cot(ϕ Q s ) f (vs ) − f (vs−1
 ) .
2ar(Fv ) sin(ϕ Q s )
Q ∼v
s

The structure is similar to the formula of the discrete Hodge star in Proposition 2.18.
Indeed, if es denotes an edge of X parallel to the black diagonal vvs and es∗ an edge
parallel to the dual diagonal, then
Discrete Complex Analysis on Planar Quad-Graphs 85
 
1 1
 f (v) = d df = d f
ar(Fv ) ar(Fv )
Fv ∂ Fv
⎛ ⎞
⎜  
1 |es∗ | ⎟
= ⎝ d f − cot(ϕ Q s ) df⎠
ar(Fv ) |es | sin(ϕ Q s )
Q s ∼v es es∗
1  |ρ Q |  

= s
( f (vs ) − f (v)) − cot(ϕ Q s ) f (vs ) − f (vs−1

) ,
2ar(Fv ) sin(ϕ Q s )
Q s ∼v

using discrete Stokes’ Theorem 2.9 in the first and third equality, Proposition 2.18
that compares the integration of the discrete Hodge star of a discrete one-form of
type ♦ with the integration of the discrete one-form d f itself in the second equality,
and |ρ Q s | = |es∗ |/|es | for the last step.

Remark In the case when the diagonals of the quadrilaterals are orthogonal to each
other, ρ Q is always a positive real number. In this case, the discrete Laplacian splits
into two separate discrete Laplacians on Γ and Γ ∗ . In this case, it is known and
actually an immediate consequence of the local representation in Corollary 2.20 that
a discrete maximum principle holds true, i.e., a discrete harmonic function can attain
its maximum only at the boundary of a closed region. This is not true for general
quad-graphs, see for example Skopenkov’s paper [23].

Corollary 2.21 Let f : V (Λ0 ) → C.

(i) If f is discrete harmonic, then ∂Λ f is discrete holomorphic.
(ii) If f is discrete holomorphic, then f , Re f , and Im f are discrete harmonic.

Proof By Corollary 2.20,  f ≡ 4∂¯♦ ∂Λ f ≡ 4∂♦ ∂¯Λ f . In particular, ∂¯♦ ∂Λ f ≡ 0 if

 f ≡ 0, which shows (i). Also, f is discrete harmonic if it is discrete holomorphic.
Using Re( f ) ≡  Re( f ) and Im( f ) ≡  Im( f ), Re( f ) and Im( f ) are discrete
harmonic if f is.

Similar to Proposition 2.1 that compares the discrete derivative ∂Λ with the smooth
derivative, the discrete Laplacian coincides with the smooth one up to order one in the
general case and up to order two for parallelogram-graphs. This was already shown
by Skopenkov in [23]. Since this result follows immediately from our previous ones,
we give a proof here as well.

Proposition 2.22 Let f C : C → C and f its restriction to V (Λ).

(i) If f C (z) is a polynomial in Re(z) and Im(z) of degree at most one, then the smooth
and the discrete Laplacian coincide on vertices: C f C (v) =  f (v) = 0.
(ii) Let all faces of Λ be parallelograms. If f C (z) is a polynomial in Re(z) and Im(z)
of degree at most two, then the smooth and the discrete Laplacian coincide on
vertices: C f C (v) =  f (v).
86 A.I. Bobenko and F. Günther

Proof (i) Proposition 2.1(ii) says that the function f (v) = v is discrete holomorphic
and Corollary 2.21(ii) that real and imaginary part of discrete holomorphic functions
are discrete harmonic. Since constants are discrete harmonic, the statement follows.
(ii) In the parallelogram case, let Q̂ denote the center of the parallelogram
Q ∈ F(Λ) ∼ = V (♦). Analogously to (i), f (v) = v2 is discrete harmonic by Propo-
sition 2.1(iii) and Corollary 2.21(ii). Looking at real and imaginary part separately,
 f 12 ≡  f 22 and  ( f 1 f 2 ) ≡ 0 where we consider f 1 (v) = Re(v), f 2 (v) = Im(v).
Finally,
| f |2 ≡ 4∂♦ ∂¯Λ | f |2 ≡ 4∂♦ h = 4

with h(Q) = Q̂ for all Q ∈ V (♦), due to Propositions 2.1(iv) and 2.5 that implied
∂¯Λ | f |2 ≡ h and ∂♦ h ≡ 1. Since any polynomial in Re(z) and Im(z) of monomials
of degree two is a linear combination of f 12 − f 22 , f 12 + f 22 , and f 1 f 2 , and since we
have shown that the discrete Laplacian  and the smooth Laplacian C coincide on
these, we are done.

Remark The second part of the last proposition generalizes the known result for
rhombi given by Chelkak and Smirnov [6]. Note that this is not true for general
quadrilaterals even if one assumes that the diagonals of quadrilaterals are orthogonal
to each other. For this, consider the following (finite) bipartite quad-graph of Fig. 4:
the black vertex 0 is adjacent to the white vertices ±1 and ±i in the quad-graph and
adjacent to the black vertices 2 + 2i, −1 ± i, and 1 − i in the graph on black vertices.
There are no further vertices. Then,  f (0) = 0 for f (v) = v2 . Indeed, we would
get  f (0) = 0 if  v = 2 + 2i by v = 1 + i obtaining a rhombic
 we had replaced
quad-graph; but |ρ Q |2 / Re(ρ Q ) ( f (v) − f (0)) scales by a factor of 2, whereas the
other nonzero summands in the formula for  f (0) remain invariant.
In the case of general quad-graphs, smooth functions fC : C → C, and restrictions
f to V (Λ), Skopenkov compared the integral of C f C over a square domain R and
a sum of  f (v) over black vertices v in R [23]. Moreover, he showed that for
f (v) = |v|2 ,
2 
 f (v) = area(vvs−1 Q̂ s vs )
ar(Fv ) Q ∼v
s

2+2i

−1+i i
Q

−1 0 1

−1−i −i 1−i

Fig. 4  f (0) = 0 for f (v) = v2

Discrete Complex Analysis on Planar Quad-Graphs 87

when Q̂ s is the intersection point of the middle perpendiculars to the diagonals of

the quadrilateral Q s (which equals the intersection point of the diagonals if Q s is a
parallelogram). Note that in general, h(Q) = Q̂ is not discrete holomorphic if Q̂ is
the intersection of these middle perpendiculars.
Definition For a finite connected subset ♦0 ⊂ ♦, let F0 be the set of faces of
X 0 that correspond to a quadrilateral Q ∈ V (♦0 ) that is incident to a vertex in
V (Λ0 )\V (∂Λ0 ). Now, let f 1 : V (Λ0 ) → C and f 2 : V (Λ0 )\V (∂Λ0 ) → C (or vice
versa) be given. Then, we denote by

1
f 1 , f 2  := − f 1 f¯2 ΩΛ
2i
F0

the discrete scalar product of f 1 and f 2 seen as functions on V (Λ0 )\V (∂Λ0 ).
In the rhombic setup, discrete versions of Green’s second identity were already
stated by Mercat [19], whose integrals were not well-defined separately, and Chelkak
and Smirnov [6], whose boundary integral was an explicit sum involving boundary
angles. Skopenkov formulated a discrete Green’s second identity with a vanishing
boundary term [23].
Theorem 2.23 Let ♦0 ⊂ ♦ be finite, and let f, g : V (Λ0 ) → C.

(i) Discrete Green’s first identity: f, g + d f, dg = f  d ḡ.

∂ X0

(ii) Discrete Green’s second identity: f, g −  f, g = ( f  d ḡ − ḡ  d f ) .

∂ X0

Proof (i) Since the discrete exterior derivative is a derivation for the discrete wedge
product by Theorem 2.16,

d ( f  d ḡ) = d f ∧ d ḡ + f  (d  d ḡ) = d f ∧ d ḡ + f  ḡ.

Now, integration over F(X 0 ) yields the desired result together with discrete Stokes’
Theorem 2.9 and the basic properties of the discrete Hodge star given in Corol-
lary 2.17(ii) and (iv).
(ii) Just apply twice discrete Green’s first identity, once with the roles of f and g
interchanged, and subtract the equations from another.
The following discrete Weyl’s lemma is a direct consequence of discrete Green’s
second identity, Theorem 2.23(ii). A version for rhombic quad-graphs was given by
Mercat in [19], proven by an explicit calculation.
Corollary 2.24 f : V (Λ) → C is discrete harmonic if and only if f, g = 0 for
every compactly supported g : V (Λ) → C.
Skopenkov introduced the notion of discrete harmonic conjugates in [23]. We
recover his definitions in our notation, observing that his discrete gradient corre-
sponds to the discrete exterior derivative and his counterclockwise rotation by π/2
to the discrete Hodge star.
88 A.I. Bobenko and F. Günther

Definition Let f be a real (discrete harmonic) function on V (Λ0 ). A real discrete

harmonic function f˜ on V (Λ0 ) is said to be a discrete harmonic conjugate of f if
f + i f˜ is discrete holomorphic at all vertices of ♦0 .
Note that the existence of a real function f˜ such that f + i f˜ is discrete holomor-
phic requires already that f is discrete harmonic at all interior vertices of Λ0 (i.e.,
V (Λ0 )\V (∂Λ0 )) due to Corollary 2.21(ii) implying that the real part of a discrete
holomorphic function is discrete harmonic.
Lemma 2.25 Let f : V (Λ0 ) → R satisfy  f (v) = 0 for all v ∈ V (Λ0 )\V (∂Λ0 ).
(i) The discrete harmonic conjugate f˜ is unique up to two additive real constants
on Γ0 and Γ0∗ .
(ii) If ♦0 forms a simply-connected closed region, then a discrete harmonic
conjugate f˜ exists.
Proof (i) If f˜1 and f˜2 are two real discrete harmonic conjugates, then their difference
f˜1 − f˜2 is real and discrete holomorphic at all vertices of ♦0 . So by Proposition 2.2(ii)
(♦0 is connected), it is biconstant as a real discrete holomorphic function.
(ii) Since f is harmonic, d  d f = 0, i.e., d f is closed and of type ♦. Moreover,
reality of f implies d f = −i∂Λ f dz + i ∂¯Λ f d z̄ = 2 Im (∂Λ f dz). So in the same
manner as in the proof of Proposition 2.8 (existence of a discrete primitive if ♦0
forms a simply-connected closed region), d f can be integrated to a real function f˜
on V (Λ0 ). Since

d( f + i f˜) = d f + i  d f = 2 Re (∂Λ f dz) + 2i Im (∂Λ f dz) = 2∂Λ f dz

is of the form pdz and of type ♦, f + i f˜ is discrete holomorphic by Corollary 2.12.

Note that in the case of quadrilaterals with orthogonal diagonals, such that 
splits into two discrete Laplacians on Γ and Γ ∗ , it follows that a discrete harmonic
conjugate of a discrete harmonic function on V (Γ ) can be defined on V (Γ ∗ ) and
vice versa, as was already noted by Chelkak and Smirnov in [6].
Corollary 2.26 Let f : V (Λ0 ) → C be discrete holomorphic at all vertices of ♦0 .
Then, Im( f ) is uniquely determined by Re( f ) up to two additive constants on Γ0
and Γ0∗ .

2.4.2 Discrete Dirichlet Energy

We follow the classical approach of discretizing the Dirichlet energy introduced by

Mercat in [21]. Note that Skopenkov’s definition in [23] is exactly the same. In par-
ticular, Skopenkov’s results, including an approximation property of the Laplacian,
convergence of the discrete Dirichlet energy to the smooth Dirichlet energy for non-
degenerate uniform sequences of quad-graphs, and further theorems for quad-graphs
with orthogonal diagonals apply as well in our setting. We refer to his work [23] for
details on these results.
Discrete Complex Analysis on Planar Quad-Graphs 89

Definition For a function f : V (Λ0 ) → C, we define the discrete Dirichlet energy

of f on ♦0 as E ♦0 ( f ) := d f, d f  ∈ [0, ∞].
If ♦0 is finite, then the discrete Dirichlet boundary value problem asks for a real
function f on V (Λ0 ) such that f is discrete harmonic at all points of V (Λ0 )\V (∂Λ0 )
and such that f agrees with a preassigned real function f 0 on the boundary V (∂Λ0 ).

Proposition 2.27 Let ♦0 ⊆ ♦ be finite, and let f : V (Λ0 ) → C. Then,

1  
E ♦0 ( f ) =   |ρ Q |2 | f (b+ ) − f (b− )|2 + | f (w+ ) − f (w− )|2
Q∈V (♦0 )
2 Re ρ Q
   
Im ρ Q
+   Re ( f (b+ ) − f (b− )) ( f (w+ ) − f (w− )) .
Q∈V (♦ )
Re ρ Q
0

Proof Since E ♦0 ( f ) is a sum over Q ∈ V (♦0 ), it suffices to check the identity for
just a singular quadrilateral Q. Furthermore, E ♦0 ( f ) = E ♦0 (Re( f )) + E ♦0 (Im( f ))
allows us to restrict to real functions f . Then, E Q ( f ) equals

d f ∧ d f = 4area(Q)∂Λ f (Q)∂¯Λ f (Q)
FQ

= 2|w+ − w− ||b+ − b− | sin(ϕ Q )|∂¯Λ f (Q)|2 .

Here, b− , w− , b+ , w+ are the vertices of Q in counterclockwise order, starting with

a black vertex, and FQ is the face of X corresponding to Q.
In the proof of Proposition 2.1(i), we calculated

(w+ − w− ) ( f (b+ ) − f (b− )) − (b+ − b− ) ( f (w+ ) − f (w− ))

∂¯Λ f (Q) = .
2i|w+ − w− ||b+ − b− | sin(ϕ Q )

It follows that E Q ( f ) equals

|w+ − w− | |b+ − b− |
| f (b+ ) − f (b− )|2 + | f (w+ ) − f (w− )|2
2|b+ − b− | sin(ϕ Q ) 2|w+ − w− | sin(ϕ Q )
 
(w+ − w− )(b+ − b− )
− Re ( f (b+ ) − f (b− )) ( f (w+ ) − f (w− )) .
|w+ − w− ||b+ − b− | sin(ϕ Q )

Remembering ρ Q = −i exp(iϕ Q )|w+ − w− |/|b+ − b− |, the claim follows from

|w+ − w− | |ρ Q |2 |b+ − b− | 1
=  , =  ,
2|b+ − b− | sin(ϕ Q ) 2 Re ρ Q 2|w+ − w− | sin(ϕ Q ) 2 Re ρ Q
 
(w+ − w− )(b+ − b− ) Im(ρ Q )
− Re = .
|w+ − w− ||b+ − b− | sin(ϕ Q ) Re(ρ Q )
90 A.I. Bobenko and F. Günther

The same formula of E ♦0 ( f ) was given by Mercat [21].

In the case of rhombic quad-graphs, Duffin proved in [10] that the discrete Dirich-
let boundary value problem has a unique solution. The same argument applies for
general quad-graphs with the discrete Dirichlet energy defined here. Using a differ-
ent notation, Skopenkov proved existence and uniqueness of solutions of the discrete
Dirichlet boundary value problem as well [23].

Lemma 2.28 Let ♦0 ⊂ ♦ be finite and f 0 : V (∂Λ0 ) → R. We consider the affine

space of real functions f : V (Λ0 ) → R that agree with f 0 on the boundary.
Then, E ♦0 is a strictly convex nonnegative quadratic functional in terms of the
interior values f (v). Furthermore,

∂ E ♦0
− ( f ) = 2ar(Fv ) f (v)
∂ f (v)

for any v ∈ V (Λ0 )\V (∂Λ0 ). In particular, the solution of the discrete Dirichlet
boundary value problem is given by the unique minimizer of E ♦0 .

Proof By construction, E ♦0 is a quadratic form in the vector space of real functions

f : V (Λ0 )\V (∂Λ0 ) → R. In particular, it is convex, nonnegative, and quadratic in
terms of the values f (v). Thus, global minima exist. To prove strict convexity, it
suffices to check that the minimum is unique.
For an interior vertex v0 ∈ V (Λ0 )\V (∂Λ0 ), let φ(v) := δvv0 be the Kronecker
delta function on V (Λ0 ). Then,

∂ E ♦0 d
(f) = E ♦ ( f + tφ)|t=0 = 2 d f, dφ = −2  f, φ = −2ar(Fv0 ) f (v0 )
∂ f (v0 ) dt 0

due to Proposition 2.19 that stated that δ is the formal adjoint of d. To apply the
proposition, we consider φ as a function on V (Λ) and extend f to V (Λ) by setting
it zero on V (Λ)\V (Λ0 ). This changes neither d f, dφ nor 2  f, φ.
It follows that exactly the minima of E ♦0 are discrete harmonic and therefore solve
the discrete Dirichlet boundary value problem. The difference g of two minima is a
discrete harmonic function vanishing on the boundary. Similar to the argument given
in the previous paragraph, E ♦0 (g) = dg, dg = − g, g = 0 by Proposition 2.19
since g is zero on V (∂Λ0 ). But only biconstant functions have zero energy. Thus,
the difference has to vanish everywhere, i.e., minima are unique.

In the following, we apply Lemma 2.28 to show that ∂Λ , ∂¯Λ , ∂♦ , ∂¯♦ ,  are surjec-
tive operators. This implies immediately the existence of discrete Green’s functions
and discrete Cauchy’s kernels, as we will see in Sects. 2.5 and 2.6.

Lemma 2.29 Let ♦0 ⊂ ♦ be finite and assume that it forms a simply-connected

closed region. Then, the discrete derivatives ∂Λ , ∂¯Λ , ∂♦ , ∂¯♦ and the discrete Lapla-
cian  are surjective operators. That means, given any complex functions h 0
on V (♦0 ) and f 0 on V (Λ0 )\V (∂Λ0 ), there exist functions h ∂ , h ∂¯ on V (♦0 ) and
Discrete Complex Analysis on Planar Quad-Graphs 91

f ∂ , f ∂¯ , f  on V (Λ0 ) such that ∂♦ h ∂ = ∂¯♦ h ∂¯ =  f  = f 0 and ∂Λ f ∂ = ∂¯Λ f ∂¯ = h 0 .

If f 0 is real-valued, then f  can be chosen real-valued as well.

Proof Denote by B the number of vertices of ∂Λ0 . By assumption, ∂Λ0 is a simple

closed broken line with B edges.
By the previous Lemma 2.28, the space of real discrete harmonic functions on
V (Λ0 ) has dimension B. Clearly, real and imaginary part of a discrete harmonic
function are itself discrete harmonic. Therefore, the complex dimension of the space
of complex discrete harmonic functions, i.e., of the kernel of , is B as well. Thus,
 : KV (Λ0 ) → KV (Λ0 \∂Λ0 ) is a surjective linear operator with K ∈ {R, C}.
Now,  = 4∂♦ ∂¯Λ = 4∂¯♦ ∂Λ by Corollary 2.20, so ∂♦ , ∂¯♦ : CV (♦0 ) → CV (Λ0 \∂Λ0 )
are surjective as well. The kernel of ∂¯♦ consists of all discrete holomorphic functions
on V (♦0 ). By Proposition 2.8 (♦0 forms a simply-connected closed region), any
such function has a discrete primitive, i.e., the kernel is contained in the image of
∂Λ . Using the surjectivity of , it follows that ∂Λ : CV (Λ0 ) → CV (♦0 ) is surjective.
The same is true for ∂¯Λ .

Theorem 2.30 The discrete derivatives ∂Λ , ∂¯Λ , ∂♦ , ∂¯♦ and the discrete Laplacian
 (defined on complex or real functions) are surjective operators on the vector space
of functions on V (Λ) or V (♦).

Proof Let ♦0 ⊂ ♦1 ⊂ ♦2 ⊂ . . . ⊂ ♦ bea sequence of finite subgraphs forming sim-

ply-connected closed regions such that ∞ k=0 ♦k = ♦. By Λk we denote the subgraph
of Λ whose vertices and edges are the vertices and edges of quadrilaterals in ♦k .
Let us first prove that any h : V (♦) → C has a preimage under the discrete deriv-
atives ∂Λ , ∂¯Λ . By the previous Lemma 2.29, the affine space A(0) k of all complex
functions on V (Λ ) that are mapped to h| by ∂ (or ¯Λ ) is nonempty. Let
∂
 k V (♦ k ) Λ

A(0)
k  denote the affine space of restrictions of these functions to V (Λ j ) ⊆ V (Λk ).
Λj
Clearly,  
(0)  
A(0)
0 ⊇ A 1  ⊇ A(0)
2  ⊇ ...
Λ0 Λ0

Since all affine spaces are finite-dimensional and nonempty, this chain becomes
stationary at some point, giving a function f 0 on V (Λ0 ) mapped to h|V (♦0 ) by ∂Λ (or
∂¯Λ ) that can be extended to a function in A(0) k for any k.
Inductively, assume that f j : V (Λ j ) → C is mapped to h|V (♦ j ) by ∂Λ (or ∂¯Λ ) and
( j) ( j+1)
that f j can be extended to a function in Ak for all k  j. Let Ak , k  j + 1,
be the affine space of all complex functions on V (Λk ) that are mapped to h|V (♦k ) by
∂Λ (or ∂¯Λ ) and whose restriction to V (Λ j ) is equal to f j . By assumption, all these
spaces are nonempty. In the same way as above, there is a function f j+1 extending
f j to V (Λ j+1 ) that is mapped to h|V (♦ j+1 ) by ∂Λ (or ∂¯Λ ) and that can be extended to
( j+1)
a function in Ak for all k  j + 1.
For v ∈ V (Λk ), define f (v) := f k (v). f is a well-defined complex function on
V (Λ) with ∂Λ f = h (or ∂¯Λ f = h). Hence, ∂Λ , ∂¯Λ : CV (Λ) → CV (♦) are surjective.
92 A.I. Bobenko and F. Günther

Replacing V (♦k ) by V (Λk )\V (∂Λk ), we obtain with the same arguments that
 is surjective, regardless whether  is defined on real or complex functions.
Finally, ∂♦ , ∂¯♦ : CV (♦) → CV (Λ) are surjective due to  = 4∂♦ ∂¯Λ = 4∂¯♦ ∂Λ by
Corollary 2.20.

In the case of rhombic quad-graphs with bounded interior angles, Kenyon proved
the existence of a discrete Green’s function and a discrete Cauchy’s kernel with
asymptotic behaviors similar to the classical setting [16]. But in the general case, it
seems to be practically impossible to speak about any asymptotic behavior of certain
discrete functions. For this reason, we will consider functions that discretize Green’s
functions and Cauchy’s kernels apart from their asymptotics in Sects. 2.5 and 2.6. Not
requiring a certain asymptotic behavior leads to non-uniqueness of these functions.
Still, one can expect results concerning the asymptotics of special discrete func-
tions if the interior angles and the side lengths of the quadrilaterals are bounded,
meaning that the quadrilaterals do not degenerate at infinity. And indeed, on such
quad-graphs any discrete harmonic function whose difference functions on V (Γ )
and V (Γ ∗ ) have asymptotics o(v−1/2 ) as |v| → ∞ is biconstant. In the rhombic set-
ting, Chelkak and Smirnov showed that a discrete Liouville’s theorem holds true,
i.e., any bounded discrete harmonic function on V (Λ) vanishes [6].

Theorem 2.31 Assume that there exist constants α0 > 0 and E 1  E 0 > 0 such that
α  α0 and E 1  e  E 0 for all interior angles α and side lengths e of quadrilaterals
−1/2
Q ∈ F(Λ). If f : V (Λ) → C is discrete harmonic and f (v+ ) − f (v− ) = o(v± )
for any two adjacent v± ∈ V (Γ ) or v± ∈ V (Γ ∗ ) as |v± | → ∞, then f is biconstant.

Proof Without loss of generality, we can restrict to real functions f . Assume that
f is not biconstant. Then, d f ∧ d f is nonzero somewhere on a face F of X . In
particular, the discrete Dirichlet energy of f is bounded away from zero if a domain
contains F. Now, the idea of proof is to show that if the domain is large enough but
still compact, the function being zero in the interior and equal to f on the boundary
has a smaller discrete Dirichlet energy than f , contradicting Lemma 2.28 that implies
that f is the unique minimizer of the discrete Dirichlet energy on that domain.
Let us first bound the intersection angles and the lengths of diagonals of the
quadrilaterals. Take Q ∈ F(Λ) and denote its vertices by b− , w− , b+ , w+ in coun-
terclockwise order, starting with a black vertex. Then, there are two opposite interior
angles that are less than π , say α± at vertices b± . Since all interior angles are bounded
by α0 from below, one of α± is less than or equal to π − α0 , say α0  α−  π − α0 .
By triangle inequality, |b+ − b− |, |w+ − w− | < 2E 1 . Twice the area of Q equals

|w− − b− ||w+ − b− | sin(α− ) + |w− − b+ ||w+ − b+ | sin(α+ )  E 02 sin(α0 ),

so |b+ − b− ||w+ − w− | sin(ϕ Q ) = 2area(Q)  E 02 sin(α0 ). It follows that

E 02 sin(α0 ) E 2 sin(α0 )
|b+ − b− |  > 0 =: E 0 .
|w+ − w− | sin(ϕ Q ) 2E 1
Discrete Complex Analysis on Planar Quad-Graphs 93

Similarly, |w+ − w− | > E 0 and sin(ϕ Q ) > E 0 /(2E 1 ). Thus, we can bound

|w+ − w− |   π  |w+ − w− |  
ρQ = exp i ϕ Q − = sin(ϕ Q ) − i cos(ϕ Q )
|b+ − b− | 2 |b+ − b− |
 2
2E 1   E 0
by |ρ Q | <  and Re ρ Q > .
E0 2E 1

For some r > 0, denote by B♦ (0, r ) ⊂ V (♦) the set of quadrilaterals that have a
nonempty intersection with the open ball B(0, r ) around 0 and radius r . Let R > 2E 1 ,
and consider the ball B♦ (0, R) ⊂ V (♦). Since Λ is locally finite, B♦ (0, R) is finite.
Also, if we connect two elements of B♦ (0, R) if they are adjacent in ♦, then we
obtain a connected subgraph of ♦ that we will also denote by B♦ (0, R). To see that
it is connected, we observe that the closed region in the complex plane formed by
the quadrilaterals in B♦ (0, R) is connected, and that if Q ∈ B♦ (0, R), then one of
its corners, say v, has to lie in B(0, R) and so all quadrilaterals incident to v are in
B♦ (0, R). We denote by Λ R the subgraph of Λ that consists of all the vertices and
edges of quadrilaterals in B♦ (0, R)
Since edge lengths are bounded by E 1 , all elements of B♦ (0, R) that are not
completely contained in B(0, R) are contained in B(0, R + 2E 1 )\B(0, R − 2E 1 ).
The area of the latter is 8π R E 1 . Any quadrilateral has area at least E 02 sin(α0 )/2, so
at most 16π R E 1 /(E 02 sin(α0 )) quadrilaterals of B♦ (0, R) do not lie completely in
B(0, R). We call these quadrilaterals for short boundary faces.
Consider the real function f R defined on V (Λ R ) that is equal to f at V (∂Λ R )
and equal to 0 in V (Λ R )\V (∂Λ R ). When computing the discrete Dirichlet energy of
f R on B♦ (0, R), only boundary faces can give nonzero contributions. If we look at
the formula of the discrete Dirichlet energy in Proposition 2.27 and use in addition
that f (v+ ) − f (v− ) = o(R −1/2 ) for vertices of boundary faces, then we see that
any
 contribution
 of a boundary face has asymptotics o(R−1 ).For this, we use that
Re ρ Q )  is bounded from below by a constant and Im ρ Q   ρ Q  < 2E 1 /E  .
0
Using that there are only O(R) faces in the boundary (the constant depending on
E 0 , E 1 , α0 only), the discrete Dirichlet energy E B♦ (0,R) ( f R ), considered as a function
of R, behaves as o(1). So if R is large enough, then

E B♦ (0,R) ( f R ) < d f ∧ d f  E B♦ (0,R) ( f ),
F

contradicting that f minimizes the discrete Dirichlet energy by Lemma 2.28.

2.5 Discrete Green’s Functions

Definition Let v0 ∈ V (Λ). A real function G(·; v0 ) on V (Λ) is a (free) discrete
Green’s function for v0 if
94 A.I. Bobenko and F. Günther

1
G(v0 ; v0 ) = 0 and G(v; v0 ) = δvv for all v ∈ V (Λ).
2ar(Fv0 ) 0

Remark It is important to note that discrete Green’s functions as we defined them

are far from being unique. The contrast to the smooth setting or the rhombic case
investigated by Kenyon [16], Chelkak and Smirnov [6] is that no asymptotics are
required. So it might be more appropriate to call these functions functions of discrete
Green’s function type, but for the sake of convenience, we still call them discrete
Green’s functions.
However, when considering planar parallelogram-graphs with bounded interior
angles and bounded ratio of side lengths in Sect. 3.3, existence of a discrete Green’s
function with asymptotics generalizing the corresponding result for rhombic quad-
graphs is proven.

As a corollary of Theorem 2.30 we get existence of discrete Green’s functions.

Corollary 2.32 A discrete Green’s function exists for any v0 ∈ V (Λ).

Proof By Theorem 2.30,

  issurjective, so there exists a function G : V (Λ) → R
with G(v) = δvv0 / 2ar(Fv0 ) . Since constant functions are discrete harmonic, we
can adjust G to get G(v0 ) = 0.

The following notion of discrete Greens’ functions in a discrete domain follows

the presentation of Chelkak and Smirnov in [6].

Definition Let ♦0 ⊂ ♦ be finite and form a simply-connected closed region. For

a vertex v0 ∈ V (Λ0 )\V (∂Λ0 ), a real function G Λ0 (·; v0 ) on V (Λ0 ) is a discrete
Green’s function in Λ0 for v0 if

G Λ0 (v; v0 ) = 0 for all v ∈ V (∂Λ0 )

1
and G Λ0 (v; v0 ) = δvv for all v ∈ V (Λ0 )\V (∂Λ0 ).
2ar(Fv0 ) 0

An immediate corollary of Lemma 2.29 is now the existence of these functions.

Corollary 2.33 Let ♦0 ⊂ ♦ be finite and form a simply-connected closed region.

Furthermore, let v0 ∈ V (Λ0 )\V (∂Λ0 ). Then, there exists a unique discrete Green’s
function in Λ0 for v0 .

Proof Due to our assumptions on ♦0 , existence follows from Lemma 2.29 stating
surjectivity of  on such domains. Since the difference of two discrete Green’s
functions in Λ0 for v0 is discrete harmonic on V (Λ0 ) and equals zero on the boundary
V (∂Λ0 ), it has to be identically zero by Lemma 2.28 since the zero function is the
unique solution of the corresponding discrete Dirichlet boundary value problem.
Discrete Complex Analysis on Planar Quad-Graphs 95

2.6 Discrete Cauchy’s Integral Formulae

In this section, we first formulate discretizations of the standard Cauchy’s integral

formula, both for discrete holomorphic functions on V (Λ) and V (♦). Later, we
give with Theorem 2.36 a discrete formulation of Cauchy’s integral formula for the
derivative of a holomorphic function. We conclude this part with Sect. 2.6.1, where
we relate our formulation of the discrete Cauchy’s integral formula for discrete
holomorphic functions on V (♦) with Chelkak’s and Smirnov’s notation in [6].

Definition Discrete Cauchy’s kernels with respect to Q 0 ∈ V (♦) and v0 ∈ V (Λ)

are functions K Q 0 : V (Λ) → C and K v0 : V (♦) → C, respectively, that satisfy for
all Q ∈ V (♦), v ∈ V (Λ):
π π
∂¯Λ K Q 0 (Q) = δ Q Q 0 and ∂¯♦ K v0 (v) = δvv0 .
ar(FQ ) ar(Fv )

For fixed Q 0 ∈ V (♦0 ), v0 ∈ V (Λ0 )\V (∂Λ0 ), functions K Q 0 : V (Λ0 ) → C and

K v0 : V (♦0 ) → C satisfying the above equations for all faces Q ∈ V (♦0 ) and ver-
tices v ∈ V (Λ0 )\V (∂Λ0 ) are called discrete Cauchy’s kernels on V (Λ0 ) or V (♦0 ),
respectively.

Clearly, the restrictions of discrete Cauchy’s kernels to V (Λ0 ) or V (♦0 ), respec-

tively, are discrete Cauchy’s kernels on V (Λ0 ) or V (♦0 ), respectively.

Remark It is important to note that exactly as discrete Green’s functions, discrete

Cauchy’s kernels as we defined them are far from being unique. Again, it might be
more appropriate to call these functions functions of discrete Cauchy’s kernel type,
but we still call them discrete Cauchy’s kernels.
But if interior angles and side lengths of quadrilaterals are bounded, then it follows
from Theorem 2.31 that any discrete Cauchy’s kernel with respect to a vertex of ♦
with asymptotics o(v−1/2 ) as |v| → ∞ is necessarily unique, but we cannot prove
existence so far. In Sect. 3.4, explicit formulae for discrete Cauchy’s kernels with
asymptotics similar to the smooth setting are given, generalizing Kenyon’s result
[16] on rhombic quad-graphs to parallelogram-graphs.

The existence of discrete Cauchy’s kernels follows from the surjectivity of discrete
derivatives by Theorem 2.30:

Corollary 2.34 Let Q 0 ∈ V (♦) and v0 ∈ V (Λ) be arbitrary. Then, discrete

Cauchy’s kernels with respect to Q 0 and v0 exist.

Theorem 2.35 Let f and h be discrete holomorphic functions on V (Λ0 ) and V (♦0 ),
respectively. Furthermore, let v0 ∈ V (Λ0 )\V (∂Λ0 ) and Q 0 ∈ V (♦0 ) be given, and
let K v0 : V (♦) → C and K Q 0 : V (Λ) → C be discrete Cauchy’s kernels with respect
to v0 and Q 0 on V (♦0 ) and V (Λ0 ), respectively.
Then, for any discrete contours Cv0 and C Q 0 on X 0 surrounding v0 and Q 0 ,
respectively, once in counterclockwise order, discrete Cauchy’s integral formulae
hold:
96 A.I. Bobenko and F. Günther
 
1 1
f (v0 ) = f K v0 dz and h(Q 0 ) = h K Q 0 dz.
2πi 2πi
C v0 C Q0

Proof Let Pv and PQ be discrete elementary cycles, v being an interior vertex and
Q an interior face of Λ0 . By Lemma 2.3 that relates ∂Λ , ∂¯Λ with discrete contour
integrals and the definition of ∂¯♦ , we get:

1 1
f K v0 dz = ar(Fv ) f (v)∂¯♦ K v0 (v) = δvv0 f (v),
2πi π
Pv

1 1
f K v0 dz = ar(FQ )∂¯Λ f (Q)K v0 (Q) = 0.
2πi π
PQ

Here, we used that the value of the product on [Q  , v ] ∈ E(X 0 ) is f (v )K v0 (Q  ),

so in the first integration we could factor out f (v), in the second one K v0 (Q).
By definition, the discrete contour Cv0 is the oriented boundary of a topological
disk in F(X 0 ), so we can decompose the integration along Cv0 into a couple of
integrations along discrete elementary cycles Pv and PQ as above. Summing up,
only the contribution of Pv0 is nonvanishing, and we get the desired result. The
second formula is shown in an analog fashion.
Remark In the case of rhombic quad-graphs, Mercat formulated a discrete Cauchy’s
integral formula for the average of a discrete holomorphic function on V (Λ) along an
edge of Λ. In [6], Chelkak and Smirnov provided a discrete Cauchy’s integral formula
for discrete holomorphic functions on V (♦) using an integration along cycles on Γ
and Γ ∗ , see Sect. 2.6.1.
Theorem 2.36 Let f : V (Λ0 ) → C be discrete holomorphic, let Q 0 ∈ V (♦0 ), and
let K Q 0 : V (Λ0 ) → C be a discrete Cauchy’s kernel with respect to Q 0 on V (Λ0 ).
Then, for any discrete contour C Q 0 in X 0 surrounding Q 0 once in counterclockwise
order that does not contain any edge [Q 0 , v] of X 0 , v ∼ Q 0 (see Fig. 5), the discrete
Cauchy’s integral formula is true:

1
∂Λ f (Q 0 ) = − f ∂Λ K Q 0 dz.
2πi
C Q0

Q0

Fig. 5 Discrete contour as in Theorem 2.36

Discrete Complex Analysis on Planar Quad-Graphs 97

Proof Let D be the discrete domain in F(X 0 ) bounded by C Q 0 . Since no edge of

C Q 0 passes through Q 0 , the discrete one-form ∂¯Λ K Q 0 d z̄ vanishes on C Q 0 . Therefore,
d K Q 0 = ∂Λ K Q 0 dz on C Q 0 and
   
f ∂Λ K Q 0 dz = f d K Q0 = d( f d K Q 0 ) = d f ∧ d K Q0
C Q0 C Q0 D D

due to discrete Stokes’ Theorem 2.9 in the second equality and Theorem 2.16 that
assures that d( f d K Q 0 ) = d f ∧ d K Q 0 + f dd K Q 0 and Proposition 2.10 that assures
that dd K Q 0 = 0 in the third equality. Now, f is discrete holomorphic, so we obtain
d f ∧ d K Q 0 = ∂Λ f ∂¯Λ K Q 0 Ω♦ . But ∂¯Λ K Q 0 vanishes on all vertices of ♦0 but Q 0 .
Finally,
 
1 1
− f ∂Λ K Q 0 dz = − ∂Λ f ∂¯Λ K Q 0 Ω♦ = ∂Λ f (Q 0 ).
2πi 2πi
C Q0 FQ 0

Remark In general, there exists no analog of the above Theorem 2.36 for the dis-
crete derivative of a discrete holomorphic function on V (♦0 ), because the discrete
derivative itself does not need to be discrete holomorphic. However, in the special
case of integer lattices, any discrete derivative of a discrete holomorphic function is
itself discrete holomorphic. In Sect. 3.5, we will obtain discrete analogs of Cauchy’s
integral formulae for higher derivatives of discrete holomorphic functions.

2.6.1 A Different Notation

Let W be a cycle on the edges of Γ ∗ , having (ordered) white vertices w0 , w1 , . . . , wm ,

wm = w0 . Then, any edge connecting two consecutive vertices wk , wk+1 forms the
diagonal of a quadrilateral face Q(wk , wk+1 ) ∈ V (♦). We denote the set of such faces
together with the induced orientation of their white diagonals by W♦ . For Q ∈ W♦ ,
we denote its white vertices by w− (Q), w+ (Q) such that the corresponding oriented
diagonal goes from w− (Q) to w+ (Q). Its black vertices are denoted by b(Q), b (Q)
in such a way that w− (Q), b(Q), w+ (Q), b (Q) appear in counterclockwise order.
The reason why we do not choose our previous notation of Fig. 1 is that black and
white vertices now play a different role that shall be indicated by the notation.
Now, we construct a cycle B on the edges of Γ having (ordered) black vertices
b0 , b1 , . . . , bn , bn = b0 , in the following way. We start with b0 := b (Q (w0 , w1 )).
In the star of the vertex w1 , there are two simple paths on Γ connecting b0 and
b (Q (w1 , w2 )), and we choose the path that does not go through Q(w0 , w1 ). Note
that it may happen that b (Q (w1 , w2 )) = b0 ; in this case, we do not add any vertices
to B. Also, w2 = w0 is possible, which causes adding the nondirect path connecting
the black vertices b0 and b (Q (w1 , w2 )) = b (Q (w0 , w1 )).
98 A.I. Bobenko and F. Günther

Continuing this procedure till we have connected b (Q (wm−1 , wm )) with b0 , we

end up with a closed path B on Γ . Without loss of generality, any two consec-
utive vertices in B are different. As above, any edge connecting two consecutive
vertices bk , bk+1 forms the diagonal of a face Q(bk , bk+1 ) ∈ V (♦). We denote the
set of such faces together with the induced orientation of their black diagonals by
B♦ . For Q ∈ B♦ , we denote its black vertices by b− (Q), b+ (Q) such that the cor-
responding oriented diagonal goes from b− (Q) to b+ (Q). Finally, its white vertices
are denoted by w(Q), w (Q) in such a way that b− (Q), w (Q), b+ (Q), w(Q) appear
in counterclockwise order.

Definition Let W and B be cycles as above and h a function defined on W♦ ∪ B♦ .

We define the discrete integrals along W and B by

m−1
h(Q)dz := h (Q (wk , wk+1 )) (wk+1 − wk ) ,
W k=0

m−1
h(Q)d z̄ := h (Q (wk , wk+1 )) (wk+1 − wk );
W k=0

n−1
h(Q)dz := h (Q (bk , bk+1 )) (bk+1 − bk ) ,
B k=0

n−1
h(Q)d z̄ := h (Q (bk , bk+1 )) (bk+1 − bk ).
B k=0

In between the closed paths B and W , there is a cycle P on the medial graph
X that comprises exactly all edges [Q, v] with Q ∈ W♦ and v ∈ B incident to Q
and all edges [Q, v] with Q ∈ B♦ and v ∈ W incident to Q. The orientation of
[Q, v] is induced by the orientation of the corresponding parallel white or black
diagonal. Figure 6 gives an example for this construction, where all cycles are oriented
counterclockwise.

P
B

Fig. 6 Cycles W on Γ ∗ , B on Γ , and closed path P on X in between

Discrete Complex Analysis on Planar Quad-Graphs 99

Remark Note that an oriented cycle P on X induces a white cycle W = W (P) and
a black cycle B = B(P) in such a way that W , P, and B are related as above.

Lemma 2.37 Let P be an oriented cycle on X and let W = W (P) and B = B(P)
be the white and black cycles it induces. Let f be a function defined on the vertices
of W and B and h a function defined on W♦ ∪ B♦ . Then,
  
f (b(Q))h(Q)dz + f (w(Q))h(Q)dz = 2 f hdz.
W B P

Proof Any edge e = [Q, b(Q)] (Q ∈ W♦ ) or [Q, w(Q)] (Q ∈ B♦ ) of P is parallel to

either an edge w− (Q)w+ (Q) of W or to an edge b− (Q)b+ (Q) of B, respectively, and
vice versa. Since the complexnumber associated to e is just half of the corresponding
∗
 edge of Γ or Γ , 2 e f h = f (b (Q)) h(Q)(w+ (Q) − w− (Q)) in the first
parallel
and 2 e f h = f (w (Q)) h(Q)(b+ (Q) − b− (Q)) in the second case. Therefore, the
discrete integral along P decomposes into one along B and one along W .

Remark Note that the construction of B and Lemma 2.37 are also valid if W consists
of a single point or of only two edges (being the same, but traversed in both directions).
In both cases, P will be a discrete contour, as well when W is simple and oriented
counterclockwise.

The discrete Cauchy’s integral formula of Chelkak and Smirnov in [6] reads as
 
πi h(Q 0 ) = h(Q)K Q 0 (b(Q))dz + h(Q)K Q 0 (w(Q))dz
W B

if Q 0 ∈ V (♦) is surrounded once by the simple closed curve W on Γ ∗ in counter-

clockwise direction, h is discrete holomorphic on V (♦), and K Q 0 : V (Λ) → C is a
discrete Cauchy’s kernel with respect to Q 0 . The above Lemma 2.37 directly relates
this formulation to the one of the discrete Cauchy’s integral formula in Theorem 2.35.
Finally, we conclude this section with a proposition relying on the decomposition
of a discrete contour into black and white cycles. In Proposition 2.13(i), we have
already seen that f dg + gd f is closed for functions f, g : V (Λ) → C. Actually, a
slightly stronger statement is true:
Proposition 2.38 Let W be a closed cycle on Γ ∗ and B be the corresponding closed
cycle on Γ as above, and let f, g : V (Λ) → C. Then,
 
   
f (b(Q)) ∂Λ g(Q)dz + ∂¯Λ g(Q)d z̄ + g(w(Q)) ∂Λ f (Q)dz + ∂¯Λ f (Q)d z̄ = 0.
W B

Proof We first rewrite the discrete integral along the cycle W

 . By definition, we have
dg = ∂Λ gdz + ∂¯Λ gd z̄. By discrete Stokes’ Theorem 2.9, e dg = g (w+ ) − g (w− )
if e is an edge on X parallel to w− w+ and oriented the same, so
100 A.I. Bobenko and F. Günther

 
f (b(Q)) ∂Λ g(Q)dz + ∂¯Λ g(Q)d z̄ = f (b(Q)) (g (w+ (Q)) − g (w− (Q)))
W Q∈W♦

=− g (w (Q)) ( f (b+ (Q)) − f (b− (Q))) ,
Q∈B♦

where we changed the summation along the path W into a summation along the path
B in the last step. Similar to above, we rewrite the discrete integral along B as

 
g(w(Q)) ∂Λ f (Q)dz + ∂¯Λ f (Q)d z̄ = g(w(Q)) ( f (b+ (Q)) − f (b− (Q))) .
B Q∈B♦

In summary,
 
   
f (b(Q)) ∂Λ g(Q)dz + ∂¯Λ g(Q)d z̄ + g(w(Q)) ∂Λ f (Q)dz + ∂¯Λ f (Q)d z̄ = 0.
W B

3 Discrete Complex Analysis on Planar

Parallelogram-Graphs

3.1 Preliminaries

Definition A planar parallelogram-graph Λ (without boundary) is a planar quad-

graph without boundary such that all its faces are parallelograms. Its dual ♦ is
embedded in the complex plane by placing all Q ∈ V (♦) at the center Q̂ of the
corresponding parallelogram Q and connecting adjacent vertices by straight line
segments.
Remark Remembering the duality V (♦) ∼ = F(Λ), we will omit the hat above Q in
the sequel and identify the vertex Q ∈ V (♦) representing a parallelogram in Λ with
the corresponding point Q̂ in the complex plane. It will be clear from the context
whether Q is meant to be a face of Λ or a point in the complex plane.
In Sect. 3, discrete complex analysis on planar parallelogram-graphs Λ is dis-
cussed. As in Sect. 2, Λ is bipartite, the induced graphs on black and white vertices
are denoted by Γ and Γ ∗ , respectively, and we assume that the cellular decomposi-
tion induced by Λ is locally finite and strongly regular (which already follows from
all faces being parallelograms).
In Propositions 2.1, 2.5, and 2.22, we have already seen that discrete complex
analysis on parallelogram-graphs is closer to the classical theory than on general
quad-graphs. For example, f (v) = v2 is a discrete holomorphic function on V (Λ)
and ∂Λ f (Q) = 2Q; h(Q) = Q is a discrete holomorphic function on V (♦) and
∂♦ h ≡ 1; and the discrete Laplacian  approximates the smooth one correctly up to
order two.
Discrete Complex Analysis on Planar Quad-Graphs 101

In order to concentrate on the calculation of the asymptotics of a certain discrete

Green’s function and discrete Cauchy’s kernels, we postpone the discussion of some
necessary combinatorial and geometric results on planar parallelogram-graphs to the
appendix. Our setup is closely related to the quasicrystallic parallelogram-graphs
discussed in [2]. There, the quad-graph was embedded into Zn and explicit formulae
for a discrete exponential function and a discrete logarithm were introduced. These
results relied on Kenyon’s formulae for a discrete exponential and a discrete Green’s
function in [16].
In the following, the ideas of [2] are adapted to our slightly more general setting in
Sect. 3.2. The discrete exponential is the basic building block for the integral formulae
of a discrete Green’s function in Sect. 3.3 and discrete Cauchy’s kernels in Sect. 3.4.
The corresponding functions can be defined for general planar parallelogram-graphs,
but we need more regularity of the graph to calculate their asymptotics. The two
conditions we use are that all interior angles of the parallelograms are bounded
(the same condition was used by Chelkak and Smirnov in [6]) and that the ratio of
side lengths of the parallelograms is bounded as well. For rhombic quad-graphs, the
second condition is trivially fulfilled; for quasicrystallic graphs, there are only a finite
number of interior angles. Note that instead of using boundedness of the ratio of side
lengths of the parallelograms, we can assume that the side lengths themselves are
bounded. This seems to be a stronger condition at first, but actually, both conditions
are equivalent, see Proposition 4.3 in the appendix.
Remark To our knowledge, it is an open problem to find an explicit formula for a
discrete Green’s function or a discrete Cauchy’s kernel in the case of general quad-
graphs as discussed in Sect. 2.
We conclude this section by a discussion of integer lattices in Sect. 3.5. On these
graphs, discrete holomorphic functions can be discretely differentiated infinitely
many times, and for all higher order discrete derivatives, discrete Cauchy’s integral
formulae with the asymptotics one expects from the asymptotics of the discrete
Cauchy’s integral formula for a discrete holomorphic function hold true.
Lemma 3.1 Let v, v ∈ V (Λ), let k be an odd integer, let e1 , . . . , en be a sequence
of n oriented edges of Λ forming a directed path from v to v on Λ, and let f 1 , . . . , f m
be another such sequence of m oriented edges. Then,

n
m
ekj = f jk .
j=1 j=1

Here, the calculations are performed directly with edges rather than replacing them
with their associated complex numbers (see Sect. 2.1.1).
Proof Consider a path p1 , p2 , p3 , p4 of oriented edges of
Λ going once around a
parallelogram. Since p1 = − p3 and p2 = − p4 , we have 4j=1 p kj = 0. Now, any
closed cycle on the planar graph Λ can be decomposed into elementary oriented
cycles around faces, where edges e, −e with opposite orientation cancel out, and
102 A.I. Bobenko and F. Günther

pairs of oppositely oriented edges (that are not necessarily successive). Using that
ek + (−e)k = 0, the claim now follows from

n
m
ekj + (− f j )k = 0
j=1 j=1

for the cycle e1 , . . . , en , − f 1 , . . . , − f m of oriented edges starting and ending in v.

In the remainder of the paper, we use the following shorthand notation.
Definition Let v, v ∈ V (Λ) and Q, Q  ∈ V (♦).
(i) Choose any directed path of edges e1 , . . . , en on Λ going from v to v. Define

n
J (v, v ) := e−1
j ,
j=1

which does not depend on the choice of path from v to v due to Lemma 3.1.
(ii) Choose any vertex v Q incident to Q and any directed path of edges e1 , . . . , en
on Λ from v to v Q . Moreover, let d1 , d2 be the two oriented edges of Q that
emanate in v Q . We now define

n
1 −1 1 −1
−J (v, Q) = J (Q, v) := e−1
j + d1 + d2 .
j=1
2 2

Note that J (Q, v) does not depend on the choice of path from v to v Q by
Lemma 3.1 nor on the choice of v Q by a similar argument as in the proof of the
above lemma.
Moreover, let τ (v, Q) = τ (Q, v) := 1/(d1 d2 ) if v Q , v are both in V (Γ ) or both
in V (Γ ∗ ) and τ (v, Q) = τ (Q, v) := −1/(d1 d2 ) otherwise. Since Q is a paral-
lelogram, these quantities depend on v and Q only.
(iii) Choose any vertices v Q  incident to Q  and v Q incident to Q and a directed path
of edges e1 , . . . , en on Λ going from v Q  to v Q . Let d1 , d2 be the two oriented
edges of Q  ending in v Q  and f 1 , f 2 the two oriented edges of Q emanating
from v Q . Define

1 −1 1 −1 −1 1 −1 1 −1
n
J (Q, Q  ) := d + d2 + e j + f1 + f2 .
2 1 2 j=1
2 2

J (Q, Q  ) does not depend on the choice of v Q  and v Q or the path from v Q  to v Q .

Furthermore, let τ (Q, Q  ) := 1/(d1 d2 f 1 f 2 ) if v Q , v Q  are both in V (Γ ) or both

in V (Γ ∗ ) and τ (Q, Q  ) := −1/(d1 d2 f 1 f 2 ) otherwise. τ (Q, Q  ) depends on Q
and Q  only, since they are both parallelograms.
Discrete Complex Analysis on Planar Quad-Graphs 103

Remark In the case that all parallelograms are rhombi of side length one, we have
J (x, x  ) = x − x  .

Finally, the notion of the argument of a complex number will become important in
the sequel. In our paper, it will be usually an arbitrary real number and not a number
modulo 2π .

Definition For a complex number λ = 0, a real number φ such that λ/ exp(iφ) is a

positive real number is called argument of λ. On the other hand, arg(λ) denotes the
unique argument φ of λ that is contained in the interval (−π, π ].

3.2 Discrete Exponential Function

Definition Let v0 be a vertex of Λ. Then, the discrete exponentials e(λ, ·; v0 ) and

exp(λ, ·; v0 ) on V (Λ) are the rational functions in the complex variable λ that are
inductively defined by e(λ, v0 ; v0 ) = 1 = exp(λ, v0 ; v0 ) and

e(λ, v ; v0 ) λ + (v − v) exp(λ, v ; v0 ) 1 + λ2 (v − v)

= and =
e(λ, v; v0 ) λ − (v − v) exp(λ, v; v0 ) 1 − λ2 (v − v)

for all vertices v, v ∈ V (Λ) adjacent to each other.

Remark Note that the quotient of e or exp at the vertices of an oriented edge e
is by definition the inverse of the quotient for the edge −e oriented in the opposite
direction. Since all faces of Λ are parallelograms, the complex numbers associated to
opposite edges oriented the same are equal and therefore are corresponding quotients
of the discrete exponentials. Thus, the discrete exponentials are well-defined.

For v ∈ V (Λ), exp(·, v; v0 ) is a rational function on C with poles being the com-
plex numbers associated to the oriented edges of a shortest directed path connecting
v0 with v. It follows from Lemma 4.2 in the appendix that the arguments of all poles
can be chosen to lie in an interval of length less than π . If in addition the interior
angles of parallelograms are bounded from below by α0 , then the arguments of all
poles can be chosen to lie even in an interval of length at most π − α0 .

Remark Note that exp(λ, ·; v0 ) = e(2/λ, ·; v0 ). Hence, e and exp are equivalent up to
reparametrization. On square lattices, the discrete exponential was already considered
by Ferrand [12] and Duffin [9]. The discrete exponential e on rhombic lattices was
used in the papers [2, 5, 16]. To be comparable to these, we use e and not exp
to perform our calculations of the asymptotic behavior. In contrast, Mercat [20],
Chelkak and Smirnov [6] preferred the parametrization of exp that is closer to the
smooth setting. Indeed, Mercat remarked that the discrete exponential exp in the
rhombic setting is a generalization of the formula
104 A.I. Bobenko and F. Günther
 n  
λx
1+ λ3 x 3
exp(λx) = 2n
λx
+O
1− 2n
n2

to the case when the path from the origin to x consists of O(|x|/δ) straight line
segments of length δ of any directions [20].
 
Definition For a face Q ∈ V (♦) with incident vertices v− , v− , v+ , v+ in counter-
clockwise order and v ∈ V (Λ), we define the discrete exponentials as the following
rational functions in the complex variable λ:

e(λ, v; v± )
e(λ, v; Q) :=  
 
,
λ − (v± − v+ ) λ − (v± − v− )
exp(λ, v± ; v)
exp(λ, Q; v) :=   .
1 − λ2 (v+

− v± ) 1 − λ2 (v−

− v± )

Remark For arbitrary Q 0 ∈ V (♦) and v0 ∈ V (Λ), the above definition yields well-
defined rational functions e(·, v0 ; Q 0 ) and exp(·, Q 0 ; v0 ). As long as λ is not a pole,
e(λ, ·; Q 0 ) is a function on V (Λ) and exp(λ, ·; v0 ) is a function on V (♦).
Proposition 3.2 Let v0 ∈ V (Λ), Q ∈ V (♦). Then, for any λ ∈ C that is not a pole
of exp(·, v; v0 ) for any vertex v ∼ Q, exp(λ, ·; v0 ) is discrete holomorphic at Q and

(∂Λ exp (λ, ·; v0 )) (Q) = λ exp(λ, Q; v0 ).

 
Proof Let v− , v− , v+ , v+ be the vertices of Q in counterclockwise order. Let us

denote a := v+ − v− and b := v− 
− v− . Using Lemma 2.3 that describes ∂¯Λ as a
 
discrete contour integration, v+ − v− = a + b, v+ − v− = a − b, and the inductive
 
¯
formula for exp, we get that ∂Λ exp(λ, ·; v0 ) (Q) equals

exp(λ, v− ; v0 ) exp(λ, ·; v0 )
dz
2iar(FQ ) exp(λ, v− ; v0 )
∂ FQ
    
exp(λ, v− ; v0 ) a+b 1 + λ2 b 1 + λ2 a a−b 1 + λ2 b 1 + λ2 a
= · − + · −1
2iar(FQ ) 2 1 − λ2 b 1 − λ2 a 2 1 − λ2 b 1 − λ2 a
exp(λ, v− ; v0 )
=    ((a + b) λ (b − a) + (a − b) λ (a + b)) = 0.
4iar(FQ ) 1 − λ2 a 1 − λ2 b

Here, FQ is the Varignon parallelogram inside Q, and ar(FQ ) = area(Q). Similarly,

we obtain that (∂Λ exp(λ, ·; v0 )) (Q) equals

exp(λ, v− ; v0 )  
 λ
 λ
 −(a + b)λ (b − a) + (b − a)λ (a + b)
4iar(FQ ) 1 − 2 a 1 − 2 b
λ exp(λ, v− ; v0 )
=    2i Im(a b̄) = λ exp(λ, Q; v0 ).
2i|a||b| sin(ϕ Q ) 1 − λ2 a 1 − λ2 b
Discrete Complex Analysis on Planar Quad-Graphs 105

3.3 Asymptotics of the Discrete Green’s Function

Following the presentation in [2], we first define a discrete logarithmic function on

the lift Λ̃v0 of the quad-graph Λ onto the Riemann surface of the logarithmic function
log(· − v0 ), which can be considered as a branched covering of the quad-graph Λ
seen as a cellular decomposition of the complex plane.
Throughout the following paragraphs, fix v0 ∈ V (Λ), and let e1 , e2 , . . . , en be the
directed edges starting in v0 , ordered according to their slopes.

Definition To each of these edges e we assign the angle θe := arg(e) ∈ (−π, π ].

We assume that θe1 < θen . Now, define θa+bn := θa + 2π b, where a ∈ {1, . . . , n} and
b ∈ Z. For m ∈ Z, let em := em mod n , considering the residue classes {1, 2, . . . , n}.

Definition Let e be one of the ek . The sector Ue ⊂ Λ is the subgraph of Λ that

consists of all vertices and edges of directed paths on Λ starting in v0 whose oriented
edges have arguments that can be chosen to lie in [arg(e), arg(e) + π ).

Definition For m ∈ Z, we define the graph Ũm to be the sector Uem with the additional
data that each vertex v of Uem besides v0 is assigned the real number ϑm (v) given by
ϑm (v) ≡ arg(v − v0 ) mod 2π and ϑm (v) ∈ [θm , θm + π ). Then,
∞

Ũ := Ũm
m=−∞

defines a graph Λ̃v0 on the Riemann surface of log(· − v0 ) that projects to the planar
parallelogram graph Λ. Here, vertices v of Uem and v of Uem are equal as vertices
of Ũm and Ũm  if and only if v = v and either v = v = v0 or ϑm (v) = ϑm  (v ).

Remark Apart from the additional data of the vertices, Ũm is composed of all the
vertices of edges of directed paths of edges on Λ starting in v0 whose arguments can
be chosen to lie in [θm , θm + π ). It follows that all Ũm+bn , b ∈ Z and 1  m  n,
cover the same sector Uem , and Ũm ∩ Ũm  contains more than just v0 if and only if
|m − m  | < n. In addition, Lemma 4.2 shows that the union of all Uek , k = 1, . . . , n,
covers the whole quad-graph Λ. It follows that Λ̃v0 is a branched covering of the
cellular decomposition Λ, branched over v0 .

Definition To each vertex ṽ ∈ V (Λ̃v0 ) covering a vertex v = v0 of Λ, let us denote

θṽ := ϑm (v) if ṽ ∈ Ũm .

Remark θṽ increases by 2π when the vertex winds once around v0 in counterclock-
wise order; and if ṽ, ṽ = v0 are adjacent vertices of Λ̃v0 , then |θṽ − θṽ | < π .

Note that by construction, if we connect v0 to some ṽ = v0 by a shortest directed

path of edges of Λ̃v0 , then the arguments of all oriented edges can be chosen to lie
all in (θṽ − π, θṽ + π ).
106 A.I. Bobenko and F. Günther

Definition Let v0 ∈ V (Λ) and let Λ̃v0 be the corresponding branched covering of
Λ. The discrete logarithmic function on V (Λ̃v0 ) is given by log(v0 ; v0 ) := 0 and

1 log(λ)
log(ṽ; v0 ) := e(λ, v; v0 )dλ
2πi 2λ
Cṽ

for ṽ = v0 . Here, Cṽ is a collection of sufficiently small counterclockwise oriented

loops going once around each pole of e(·, v; v0 ), v ∈ V (Λ) being the projection of
ṽ ∈ V (Λ̃v0 ). On each loop around a pole e, we take the branch of logarithm that
satisfies Im(log(e)) ∈ (θṽ − π, θṽ + π ).
Remark Let us suppose that v0 is a black vertex. In the special case of a rhombic qua-
sicrystallic quad-graph, the notion of the discrete logarithm is motivated as follows
[2]: The discrete logarithm is real-valued and does not branch on black vertices; and
it is purely imaginary on white points and increases by 2πi if one goes once around
v0 in counterclockwise order. Therefore, the discrete logarithm models the behavior
of the real and the imaginary part of the smooth logarithm if restricted to black and
white vertices, respectively. As we will see later in the proof of Proposition 3.4, the
values at vertices adjacent to v0 coincide with the smooth logarithm.
Lemma 3.3 Let ṽ, ṽ ∈ V (Λ̃v0 ) be two points that cover the same vertex v ∈ V (Λ)
such that θṽ − θṽ = 2π . Then

log(ṽ ; v0 ) − log(ṽ; v0 ) = 0

if v0 , v are both in V (Γ ) or both in V (Γ ∗ ), and otherwise

log(ṽ ; v0 ) − log(ṽ; v0 ) = 2πi.

Proof By definition,

 1
log(ṽ ; v0 ) − log(ṽ; v0 ) = e(λ, v; v0 )dλ.
2λ
Cṽ

The function that is integrated is meromorphic on C with poles given by the one of
e(·, v; v0 ) and zero. By residue formula, we can replace integration along Cṽ by an
integration along a circle centered at 0 with large radius R (such that all other poles
lie inside the disk) in counterclockwise direction and an integration along a circle
centered at 0 with small radius r (such that all poles lie outside the disk) in clockwise
direction. Now, e(∞, v; v0 ) = 1. If v0 , v are both in V (Γ ) or both in V (Γ ∗ ), then
e(0, v; v0 ) = 1, otherwise e(0, v; v0 ) = −1. Hence, log(ṽ ; v0 ) − log(ṽ; v0 ) = 0 in
the first and log(ṽ ; v0 ) − log(ṽ; v0 ) = 2πi in the latter case.
In particular, the real part of the discrete logarithm log(·; v0 ) is a well-defined
function on V (Λ). Divided by 2π , one actually obtains a discrete Green’s function
Discrete Complex Analysis on Planar Quad-Graphs 107

with respect to v0 . In the rhombic case, it coincides with Kenyon’s discrete logarithm
in [16] as was shown in [2].
Proposition 3.4 Let v0 be a vertex of V (Λ). The function G(·; v0 ) : V (Λ) → R
defined by G(v0 ; v0 ) = 0 and
⎛ ⎞

1 1 log(λ)
G(v; v0 ) = Re ⎝ e(λ, v; v0 )dλ⎠
2π 2πi 2λ
Cv

for each v = v0 is a (free) discrete Green’s function with respect to v0 . Here, Cv is

a collection of sufficiently small counterclockwise oriented loops going once around
each pole of e(·, v; v0 ), and on each loop around a pole e, we take the branch of log
where Im(log(e)) ∈ (arg(v − v0 ) − π, arg(v − v0 ) + π ).
Proof When evaluating the real part of the integral, we can also take the branches of
the logarithm that satisfy Im(log(e)) ∈ (arg(v − v0 ) − π, arg(v − v0 ) + π ) + 2kπ
for all poles e, where k ∈ Z is fixed. Indeed, Lemma 3.3 asserts that under this change,
the discrete logarithm changes by 0 or 2kπi, so the real part does not change.
Consider all faces incident to v0 in Λ and its lift to Λ̃v0 . For any vertex vs ∈ V (Λ)
adjacent to v0 , λ = vs − v0 is the only pole of e(λ, vs ; v0 ). The residue theorem shows
that log(ṽs ; v0 ) and log(vs − v0 ) coincide up to a multiple of 2πi if ṽs covers vs . By
a similar calculation for vertices vs adjacent to v0 in Γ or Γ ∗ , we finally get

1
G(vs ; v0 ) = Re(log(vs − v0 )),
2π  
1       vs − v0
G(vs ; v0 ) = Re log vs − v0 − log vs−1 − v0  ,
2π vs − vs−1


where v0 , vs−1 , vs , vs are the vertices of Q s ∼ v0 in counterclockwise order.
 
As in Corollary   2.20, let ρs := −i(vs − vs−1 )/(vs − v0 ). In addition, we assign
angles θvs ≡ arg vs − v0 mod 2π in such a way that 0 < θvs − θvs−1  < π . Due to
Re (−i/ρs ) = − Im (ρs ) /|ρs | and Im (−i/ρs ) = − Re (ρs ) /|ρs | ,
2 2

 

|ρs |2 (G(vs ; v0 ) − G(v0 ; v0 )) + Im (ρs ) G(vs ; v0 ) − G(vs−1 ; v0 )
Re(ρs )
  
|ρs | Re (−i/ρs ) + Im (ρs )
2  vs − v0  |ρs |2 Im (−i/ρs )  
= 
log   − θvs − θvs−1
2π Re (ρs ) vs−1 − v0  2π Re (ρs )


θvs − θvs−1

= .
2π
It follows from the explicit formula for the discrete Laplacian in Corollary 2.20 that
G(v0 ; v0 ) = 1/(2ar(Fv0 )).
Now, we show that G(·; v0 ) is discrete harmonic away from v0 . For this, we
consider the star of some vertex v = v0 , i.e., all faces of Λ incident to v ∈ V (Λ).
108 A.I. Bobenko and F. Günther

Let us assume that we can find one collection C of loops together with appropriate
branches of log such that for all vertices v of the star, G(v ; v0 ) can be computed
by an integration along C instead of Cv . Then, when we compute G(v; v0 ), we
can exchange the discrete Laplacian not only with the real part, but also with the
integration. Since e(λ, ·; v0 ) is discrete holomorphic by Proposition 3.2, it is also
discrete harmonic by Corollary 2.21. By this, we conclude that G(v; v0 ) = 0.
It remains to show that there exists such a collection C of loops with corresponding
branches of log. We will show that a collection of sufficiently small counterclockwise
oriented loops going once around each pole of e(·, v ; v0 ), v any vertex of the star
of v, does the job, where around a pole e of e(·, v ; v0 ) that branch of logarithm
is taken where Im(log(e)) ∈ (arg(v − v0 ) − π, arg(v − v0 ) + π ). For this, we just
have to show that the branches of the logarithm are well-defined. This is the case if
for two vertices v , v of the star and a common pole e of e(·, v ; v0 ) and e(·, v ; v0 ),
there is an argument of e contained in both (arg(v − v0 ) − π, arg(v − v0 ) + π ) and
(arg(v − v0 ) − π, arg(v − v0 ) + π ).
It easily follows from v = v0 that if v is not adjacent to v , there is a vertex w adja-
cent to v such that all common poles of e(·, v ; v0 ) and e(·, v ; v0 ) are also common
poles of e(·, v ; v0 ) and e(·, w; v0 ). So let us assume without loss of generality that
v and v are adjacent. Clearly, we can also assume that both vertices are different
from v0 since e(·, v0 ; v0 ) ≡ 1.
Let us suppose the converse from our claim, that means suppose that there is
a common pole e of e(·, v ; v0 ) and e(·, v ; v0 ) such that no argument of the edge
e is contained in both the two intervals (arg(v − v0 ) − π, arg(v − v0 ) + π ) and
(arg(v − v0 ) − π, arg(v − v0 ) + π ). This can only happen if the edge v v inter-
sects the ray v0 − te, t  0. But since the edge e is a pole of the discrete exponential,
there is a strip with common parallel e, i.e., an infinite path in the dual graph ♦ with
edges dual to edges of Λ that are parallel to e, that separates v0 from both v and v
in such a way that e is pointing toward the region of v and v (see the first part of the
appendix for more information on a strip). In particular, the edge v v is separated
from the ray v0 − te, t  0, and cannot intersect it, contradiction.

Remark With almost the same arguments as in the proof of Proposition 3.4, we see
that the discrete logarithm is a discrete holomorphic function on the vertices of Λ̃v0 . In
[2], it was shown that the discrete logarithm on rhombic quasicrystallic quad-graphs
is even more than discrete holomorphic, namely isomonodromic.

Before we derive the asymptotics of the discrete Green’s function given in Propo-
sition 3.4, we state and prove some necessary estimations in two separate lemmas
since we will use them later during the corresponding calculations for the discrete
Cauchy’s kernel in Sect. 3.4.

Lemma 3.5 Let E 1  E 0 > 0 be fixed real constants and consider a complex vari-
able λ. Then, for any e ∈ C\ {0} satisfying E 1  |e|  E 0 , the following holds true,
where log denotes the principal branch of the logarithm and constants in O-notation
depend on E 0 and E 1 only:
Discrete Complex Analysis on Planar Quad-Graphs 109

(i) As λ → 0,

λ+e λ λ2
− = 1 + 2 + 2 2 + O(λ3 ),
λ−e e e
   
λ+e λ e λ
log − = 2 + O(λ3 ), and log − = + O(λ2 ).
λ−e e λ−e e

(ii) As |λ| → ∞,

λ+e e e2
= 1 + 2 + 2 2 + O(λ−3 ),
λ−e λ λ
   
λ+e e −3 λ e
log = 2 + O(λ ), and log = + O(λ−2 ).
λ−e λ λ−e λ

Proof (i)
λ   
λ+e 1+ e λ λ λ2 λ λ2
− = λ
= 1+ 1 + + 2 + O(λ ) = 1 + 2 + 2 2 + O(λ3 )
3
λ−e 1− e
e e e e e

shows the first equation and implies the second equation noting that

log(1 + x) = x − x 2 /2 + O(x 3 ) as x → 0.

The series expansion for log also implies the third equation using

−d 1 λ
= λ
=1+ + O(λ2 ).
λ−d 1− d
d

(ii) These equations are shown in a completely analogous way to (i), e/λ taking
the place of λ/e.

Lemma 3.6 Assume that there exist real constants α0 > 0 and E 1  E 0 > 0 such
that α  α0 and E 1  e  E 0 for all interior angles α and side lengths e of paral-
lelograms of Λ. Let v0 ∈ V (Λ) and Q 0 ∈ V (♦) be fixed and consider v ∈ V (Λ) and
Q ∈ V (♦) in the following.
(i) Let k(v) be the combinatorial distance on Λ between v0 and v (or between a
vertex incident to Q 0 and v).
Then, k(v) = Ω(|v − v0 |) (or k(v) = Ω(|v − Q 0 |)) as |v| → ∞.
(ii) J (v, v0 ) = Ω(v − v0 ), J (Q, v0 ) = Ω(Q − v0 ) and J (Q, Q 0 ) = Ω(Q − Q 0 )
as |v|, |Q| → ∞.
(iii) τ (v, Q 0 ) = Ω(1) and τ (Q, Q 0 ) = Ω(1) as |v|, |Q| → ∞.
(iv) Furthermore, assume that |v − v0 |  1 and that the arguments of all oriented
edges of a shortest directed path on Λ from v0 to v can be chosen to lie in
[θ0 , −θ0 ], where θ0 := −(π − α0 )/2.
110 A.I. Bobenko and F. Günther

√ √
Then, for any λ ∈ [−E 1 |v − v0 |, −E 1 / |v − v0 |] :
 √ 
cos(θ0 ) |v − v0 |
|e(λ, v; v0 )|  exp − .
2E 1

Here, constants in the Ω-notation depend on α0 , E 0 , and E 1 only.

Proof Let e1 , e2 , . . . , ek(v) denote the oriented edges of Λ of a shortest directed path
on Λ from v0 (or a vertex incident to Q 0 ) to v. Due to the bound on the interior angles
of parallelograms in Λ, there is a real θ0 such that the arguments of e1 , e2 , . . . , ek(v)
can be chosen to lie all in [θ0 , θ0 + π − α0 ] by Lemma 4.2. All the claims in the
first three parts stay (essentially) the same under rotation of the complex plane, so
we may assume that θ0 = −(π − α0 )/2. The same assumption is used in the fourth
part.
(i) Under the assumptions above, the projections of the ek onto the real axis lie
on the positive axis and are at least E 0 cos(θ0 ) long since edge lengths are bounded
by E 0 from below. It follows that k(v)  Re(v − v0 )/(E 0 cos(θ0 )). Using in addition
that k(v)  |v − v0 |/E 1 , we get k(v) = Ω(|v − v0 |).
(ii) Using 1/|E 0 |  1/|e j |  1/|E 1 | for all j, we get
 
 
 k(v) −1  k(v)
|J (v, v0 )| =  e j   = O(|v − v0 |),
 j=1  E0
⎛ ⎞ ⎛ ⎞

k(v)
ē 1
k(v)
cos(θ0 )|v − v0 |
Re(J (v, v0 )) = Re ⎝
j ⎠
 2 Re ⎝ ej⎠ = .
j=1
|e j | 2 E1 j=1
E 12

Hence, J (v, v0 ) = Ω(|v − v0 |). This also implies that J (Q, v0 ) = Ω(Q − v0 ) since
|J (v, v0 ) − J (Q, v0 )|  1/|E 0 | for any v incident to Q, which follows easily from the
definition and the lower bound for edge lengths. Similarly, J (Q, Q 0 ) = Ω(Q − Q 0 )
follows from the previous statements if we take v0 incident to Q 0 .
(iii) E 0−2  |τ (Q, Q 0 )|  E 1−2 and E 0−4  |τ (Q, Q 0 )|  E 1−4 follow immedi-
ately from the definitions and the boundedness of edge lengths.
(iv) Using that λ < 0 and Re(e) > 0, we get
 
|λ + e|2 4λ Re(e) 4λ Re(e) 4λ Re(e)
= 1 +  1 +  exp .
|λ − e|2 λ2 − 2λ Re(e) + |e|2 (λ − |e|)2 (λ − E 1 )2

Taking the product of such inequalities,

   
2λ Re(v − v0 ) 2λ cos(θ0 )|v − v0 |
|e(λ, v; v0 )|  exp  exp .
(λ − E 1 )2 (λ − E 1 )2

Now, we observe that λ/(λ − E 1 )2 attains its maximum on the boundary. Together
with |v − v0 |  1,
Discrete Complex Analysis on Planar Quad-Graphs 111
√
λ − |v − v0 | −1
  √ 2  √ .
(λ − E 1 )2 E 1 1 + |v − v0 | 4E 1 |v − v0 |

Plugging this into the estimation before gives the desired result.

Theorem 3.7 Assume that there exist real constants α0 > 0 and E 1  E 0 > 0 such
that α  α0 and E 1  e  E 0 for all interior angles α and side lengths e of paral-
lelograms of Λ. Let v0 ∈ V (Λ) be fixed. Then, the following is true:
(i) The discrete Green’s function G(·; v0 ) given in Proposition 3.4 has the following
asymptotic behavior as |v| → ∞:
 
1  v − v0   
G(v; v0 ) = log   + O |v − v0 |−2 if v and v0 are of different color,

4π J (v, v0 )
γEuler + log(2) 1  
G(v; v0 ) = + log |(v − v0 )J (v, v0 )| + O |v − v0 |−2 otherwise.
2π 4π

(ii) There is exactly one discrete Green’s function G : V (Λ) → R for v0 that
behaves for |v| → ∞ as
 
1  v − v0   
G(v) = log   + o |v − v0 |−1/2 if v and v0 are of different color,

4π J (v, v0 )
γEuler + log(2) 1  
G(v) = + log |(v − v0 )J (v, v0 )| + o |v − v0 |−1/2 otherwise.
2π 4π

Here, constants in the O-notation depend on α0 , E 0 , and E 1 only, and γEuler

denotes the Euler-Mascheroni constant.

Remark Note that due to Lemma 3.6(ii), J (v, v0 ) = Ω(v − v0 ) as |v| → ∞.

By Proposition 4.3, we may replace the existence of constants E 1  E 0 > 0 such
that E 1  e  E 0 for all side lengths e of parallelograms by the existence of q0 such
that e/e  q0 for the two side lengths e, e of any parallelogram of Λ since the latter
implies the first assumption. Then, the constants in the O-notation depend instead
of E 0 and E 1 on q0 , e0 , and e1 , where e0 and e1 are the side lengths of an arbitrary
but fixed parallelogram of Λ.

The proof of the first part follows the ideas of Kenyon [16] and Bücking [5].
Both considered just quasicrystallic rhombic quad-graphs. But the main difference
to [16] is that we deform the path of integration into an equivalent one different
from Kenyon’s, since his approach does not generalize to parallelogram-graphs. As
Chelkak and Smirnov did for rhombic quad-graphs with bounded interior angles in
[6], Kenyon used that two points v, v ∈ V (Λ) can be connected by a directed path
of edges such that the angle between each directed edge and v − v is less than π/2
or the angle between the sum of two consecutive edges and v − v is less than π/2.
This is true for rhombic quad-graphs, but not for parallelogram-graphs. Instead, we
use essentially the same deformation of the path of integration as Bücking did.
112 A.I. Bobenko and F. Günther

Proof (i) The poles e1 , . . . , ek(v) of e(·, v; v0 ) correspond to the directed edges of a
path from v0 to v of minimal length k(v). By Lemma 4.2, there is a real θ0 such that
the arguments of all directed edges above can be chosen to lie in [θ0 , θ0 + π − α0 ].
It is easy to check that the claim is invariant under rotation of the complex plane, so
we can assume θ0 = −(π − α0 )/2. By definition,
⎛ ⎞

1 log λ
G(v; v0 ) = Re ⎝ 2 e(λ, v; v0 )dλ⎠ ,
8π i λ
Cv

where Cv is a collection of sufficiently small counterclockwise oriented loops going

once around each e1 , . . . , ek(v) and log is the principal branch of the logarithm since
it satisfies Im(log(e j )) ∈ [θ0 , −θ0 ] for all j.
By residue theorem, we can deform Cv into a new path of integration Cv that goes
first along a circle centered at 0 with large radius R(v) (such that all poles lie inside
this disk) in counterclockwise direction starting and ending in −R(v), then goes
along the line segment [−R(v), −r (v)] followed by the circle centered at 0 with
small radius r (v) (such that all poles lie outside this disk) in clockwise direction,
and finally goes the line segment [−R(v), −r (v)] backwards. Note that the principal
branch of log jumps by 2πi at the negative real axis where the integration along the
two line segments takes place.
By assumption, E 0  |e j |  E 1 for all j. Using |v − v0 |  E 0 , it follows that

E 05 E0 E1
|v − v0 |−4  < |e j | < 2E 1  2 4 |v − v0 |4 .
2 2 E0

In particular, we can choose

E1 E 05
R(v) := 2 |v − v0 | 4
and r (v) = |v − v0 |−4
E 04 2

for all v = v0 . We first look at the contributions of the circles with radii r (v) and
R(v) to G(v; v0 ).
Let λ be on the small circle with radius r (v). Then, λ = Ω(|v − v0 |−4 ) → 0 as
|v| → ∞. In particular, we can apply (−λ + e)/(λ − e) = 1 + 2λ/e + O(λ2 ) by
Lemma 3.5(i) to estimate (−1)k(v) e(λ, v; v0 ). More precisely, the latter is a product
of k(v) = Ω(|v − v0 |) terms (see Lemma 3.6(i)) with e = e j . Multiplying out and
using in addition E 0  |e j |  E 1 easily gives for |v| → ∞ that

(−1)k(v) e(λ, v; v0 ) = 1 + O(|v − v0 |−3 ).

Thus, we get for the integration along the small circle with radius r (v):
Discrete Complex Analysis on Planar Quad-Graphs 113
⎛ ⎞
−π
1 log(r (v)) + iθ  
Re ⎝ 2 (−1)k(v) 1 + O(|v − v0 |−3 ) d (r (v) exp(iθ ))⎠
8π i r (v) exp(iθ )
π
⎛ ⎞
π
1  
= − Re ⎝ 2 (−1)k(v) (log(r (v)) + iθ ) 1 + O(|v − v0 |−3 ) dθ ⎠
8π
−π

(−1)k(v) log(r (v))  

=− 1 + O(|v − v0 |−3 ) .
4π

Let us now consider λ to be on the large circle with radius R(v). Then, we have
|λ| = Ω(|v − v0 |4 ) → ∞ as |v| → ∞. Analogously to above, repeated use of the
−3
first equation in Lemma 3.5(ii) gives e(λ, v;
−3
 v0 ) = 1 + O(|v − v0 | ) as |v| → ∞.
Thus, log(R(v))/(4π ) · 1 + O(|v − v0 | ) is the contribution of the circle of radius
R(v). In total, the asymptotics for the real part of the integration along the two circles
are
1  
log(R(v)) − (−1)k(v) log(r (v)) + O(|v − v0 |−2 ).
4π

The two integrations along [−R(v), −r (v)] can be combined into the integral

 (v)
−r
1 e(λ, v; v0 )
dλ.
4π λ
−R(v)

Since we are interested in the asymptotics for |v| → ∞, we can consider

|v − v0 |  1 large enough and split the integration into the three parts along
  E1 E1
[−R(v), −E 1 |v − v0 |], [−E 1 |v − v0 |, − √ ], and [− √ , −r (v)].
|v − v0 | |v − v0 |

√ √
We first show that the contribution of λ ∈ [−E 1 |v − v0 |, −E 1 / |v − v0 |] can
be neglected. Indeed, it is a consequence of the estimation in Lemma 3.6(iv) that
 √ 
 −E 1 / |v−v0 |   √ 
 1 e(λ, v; v0 )   cos(θ0 ) |v − v0 |

 dλ  E 1 |v − v0 | exp − .
 4π λ  2E 1
 √
−E 1 |v−v0 |


As |v| → ∞, the latter expression

√ decays faster to zero than any power of |v − v0 |.
Now, consider λ ∈ [−E 1 / |v − v0 |, −r (v)]. Then, λ → 0 as |v| → ∞, so using
the second equation in Lemma 3.5(i) k(v) = Ω(|v − v0 |) times gives as |v| → ∞:
 
(−1)k(v) e(λ, v; v0 ) = exp 2λJ (v, v0 ) + O(k(v)λ3 )
 
= exp (2λJ (v, v0 )) 1 + O(|v − v0 |λ3 ) .
114 A.I. Bobenko and F. Günther

Thus, the integral near the origin is equal to

 (v)
−r  
(−1)k(v) exp (2λJ (v, v0 ))
+ exp (2λJ (v, v0 )) O(|v − v0 |λ2 ) dλ.
4π √
λ
−E 1 / |v−v0 |

The expansion of the integral of the second term involves two exponential
√ factors,
one for each bound: exp(−2J (v, v0 )r (v)) and exp(−2E 1 J (v, v0 )/ |v − v0 |). Now,
we will use that J (v, v0 ) = Ω(|v − v0 |) by Lemma 3.6(ii). Since the exponent of
the first factor goes to 0 in speed |v − v0 |−3 , the exponential goes exponentially fast
to 1 as |v| → ∞. For the second factor, we use our assumption that the arguments
of all the poles can be chosen to lie in [−(π − α0 )/2, (π − α0 )/2]. It follows that
Re(J (v, v0 )) is positive and goes to infinity like Ω(|v − v0 |) as |v| → ∞, such that
the second exponential factor goes to zero exponentially fast. Now, it is not hard to
see that the integral of exp (2λJ (v, v0 )) O(|v − v0 |λ3 ) gives O(|v − v0 |−2 ). For the
first term, we get
⎛ ⎞
−1  (v,v0 )
−2r (v)J
(−1)k(v)
⎜ exp(s) exp(s) − 1 ⎟
⎝ ds + ds ⎠
4π √
s s
−2E 1 J (v,v0 )/ |v−v0 | −1

 (v,v0 )
−2r (v)J
(−1)k(v) 1
+ ds
4π s
−1
⎛ ⎞
−1 0
(−1)k(v) ⎝ exp(s) exp(s) − 1 ⎠ (−1)k(v)
= ds + ds + log(2r (v)J (v, v0 ))
4π s s 4π
−∞ −1
⎛ √ ⎞

−2E 1 J (v,v0 )/ |v−v0|  0
(−1)k(v) ⎜ exp(s) exp(s) − 1 ⎟
− ⎝ ds + ds ⎠
4π s s
−∞ −2r (v)J (v,v0 )

(−1)k(v)  
= γEuler + Ω(|v − v0 |−3 ) + log(2r (v)J (v, v0 )) .
4π

To get to the last line, we used that Re(J (v, v0 )) = Ω(|v − v0 |) as |v| → ∞ stays
positive. Indeed, as |v| → ∞, the first integral in the second to last line goes to zero
exponentially fast (to see this, just write the integrand as s exp(s)/s 2 and bound
the absolute value of√the integral from above by s0 exp(s0 ) where s0 denotes the
term −2E 1 J (v, v0 )/ |v − v0 |) and the second integral is of order Ω(|v − v0 |−3 ) as
|v| → ∞ as a Taylor expansion √ of the exponential shows.
Finally, let λ ∈ [−R(v), −E 1 |v − v0 |]. Then, λ → −∞ as |v| → ∞, and repeated
use of the second equation in Lemma 3.5(ii) gives
Discrete Complex Analysis on Planar Quad-Graphs 115
 
2(v − v0 )  
e(λ, v; v0 ) = exp 1 + O(|v − v0 |λ−3 )
λ

as |v| → ∞, so the corresponding contribution of the integral to 4π G(v; v0 ) is

√ √  
−E 1 |v−v0 | −E 1 |v−v0 | 2(v−v0 )
exp(λ, v; v0 ) exp λ
dλ = dλ + O(|v − v0 |−2 )
λ λ
−R(v) −R(v)
√
−1 1 0 |/(2(v−v0 ))
−E 1 |v−v  
exp −1 exp 1s
= s
ds + ds
s s
−R(v)/(2(v−v0 )) −1

−1
1
+ ds + O(|v − v0 |−2 )
s
−R(v)/(2(v−v0 ))
 
R(v)
= γEuler − log + O(|v − v0 |−2 )
2(v − v0 )

by a similar reasoning as above. Summing up the integrals over all four parts of the
contour and taking the real part, we finally get that 4π G(v; v0 ) equals
 
1 + (−1)k(v) (γEuler + log(2)) + log |v − v0 | + (−1)k(v) log |J (v, v0 )| + O(|v − v0 |−2 ).

(ii) We know from Theorem 2.31 that discrete harmonic functions of asymptotics
o(|v − v0 |−1/2 ) as |v| → ∞ are zero. We can apply this result to the discrete harmonic
function G − G(·; v0 ), where G(·; v0 ) from the first part has the desired asymptotics.

Remark Let us compare this result to the case of rhombi of side length one. Assume
that v0 ∈ V (Γ ). Then, the discrete logarithm is purely real and nonbranched on
V (Γ ) and purely imaginary and branched on V (Γ ∗ ). It follows that G(v; v0 ) = 0 if
v ∈ V (Γ ∗ ), well fitting to the fact that  splits into two discrete Laplacians on Γ
and Γ ∗ . Using J (v, v0 ) = v − v0 ,

1
G(v; v0 ) = (γEuler + log(2) + log |v − v0 |) + O(|v − v0 |−2 )
2π

as |v| → ∞ for v ∈ V (Γ ), exactly as in the work of Bücking [5], who slightly

strengthened the error term in Kenyon’s work [16]. In this paper, Kenyon showed
that there is no further discrete Green’s function with asymptotics o(|v − v0 |).
116 A.I. Bobenko and F. Günther

3.4 Asymptotics of Discrete Cauchy’s Kernels

Let v0 ∈ V (Λ) and Q 0 ∈ V (♦) be given. We first give a formula for a discrete
Cauchy’s kernel K v0 with respect to v0 on V (♦) that has asymptotics Ω(|Q − v0 |−1 )
as |Q| → ∞. Remember that the vertex Q ∈ V (♦) ∼ = F(Λ) is placed on the center
of the corresponding parallelogram. Then, we provide a discrete Cauchy’s kernel
K Q 0 with respect to Q 0 on V (Λ) with asymptotics Ω(|v − Q 0 |−1 ) as v → ∞.
In both cases, there are no further discrete Cauchy’s kernels with asymptotics
o(|Q − v0 |−1/2 ) or o(|v − Q 0 |−1/2 ) as |Q|, |v| → ∞. In the end of this section, the
asymptotics of ∂Λ K Q 0 are determined.
Theorem 3.8 Let G(·; v0 ) be a discrete Green’s function on V (♦) for v0 ∈ V (Λ).
(i) K v0 := 8π ∂Λ G(·; v0 ) is a discrete Cauchy’s kernel with respect to v0 .
(ii) Assume additionally that there exist real constants α0 > 0 and E 1  E 0 > 0
such that α  α0 and E 1  e  E 0 for all interior angles α and side lengths e
of parallelograms of Λ. Suppose that G(·; v0 ) is the discrete Green’s function
given in Proposition 3.4 and K v0 the discrete Cauchy’s kernel given in (i). Then,

1 τ (Q, v0 )
K v0 (Q) = + + O(|Q − v0 |−2 )
Q − v0 J (Q, v0 )

as |Q| → ∞. Here, constants in the O-notation depend on α0 , E 0 , and E 1 only.

(iii) Under the conditions of (ii), there is exactly one discrete Cauchy’s kernel with
respect to v0 with asymptotics o(|Q − v0 |−1/2 ) as |Q| → ∞.
Remark Note that due to Lemma 3.6(ii) and (iii),

τ (Q, v0 )  
= Ω (Q − v0 )−1
J (Q, v0 )

as |Q| → ∞. As in Theorem 3.7, we may replace the existence of E 1  E 0 > 0

such that E 1  e  E 0 for all side lengths e of parallelograms by the existence of
q0 such that e/e  q0 for the two side lengths e, e of any parallelogram of Λ since
the latter implies the first assumption by Proposition 4.3. Then, the constants in the
O-notation depend instead of E 0 and E 1 on q0 , e0 , and e1 , where e0 and e1 are the
side lengths of an arbitrary but fixed parallelogram of Λ.
Proof (i) By Corollary 2.20 about the factorization of the discrete Laplacian,
  π
∂¯♦ K v0 (v) = 8π ∂¯♦ ∂Λ G(·; v0 ) (v) = 2π G(v; v0 ) = δvv0 .
ar(Fv0 )

(ii) For the ease of notation, we assume that v0 ∈ V (Γ ), but note that the other
case of a white vertex v0 can be handled in the same manner. Let b− , w− , b+ , w+
denote the vertices of the parallelogram Q in counterclockwise order, starting with
a black vertex. Using the asymptotics of Theorem 3.7, as |Q| → ∞:
Discrete Complex Analysis on Planar Quad-Graphs 117
 
|J (b+ , v0 )||b+ − v0 |
4π G(b+ ; v0 ) − 4π G(b− ; v0 ) = log + O(|Q − v0 |−2 ),
|J (b− , v0 )||b− − v0 |
 
|J (w− , v0 )||w+ − v0 |
4π G(w+ ; v0 ) − 4π G(w− ; v0 ) = log + O(|Q − v0 |−2 ).
|J (w+ , v0 )||w− − v0 |

Let a := w+ − b− and b := w− − b− . Using log |1 + x| = (x + x̄)/2 + O(x 2 )

as x → 0 and J (Q, v0 ) = Ω(|Q − v0 |) as |Q| → ∞ given by Lemma 3.6(ii), we
get as |Q| → ∞:
  
 a+b 
log |b± − v0 | = log (Q − v0 ) 1 ± 
2(Q − v0 ) 
 
a+b
= log |Q − v0 | ± Re + O(|Q − v0 |−2 ),
2(Q − v0 )
 
a−b
log |w± − v0 | = log |Q − v0 | ± Re + O(|Q − v0 |−2 ),
2(Q − v0 )
  
 a −1 + b−1 
log |J (b± , v0 )| = log  J (Q, v0 ) 1 ±
2J (Q, v0 ) 
 −1 
a + b−1
= log |J (Q, v0 )| ± Re + O(|Q − v0 |−2 ),
2J (Q, v0 )
 −1 
a − b−1
log |J (w± , v0 )| = log |J (Q, v0 )| ± Re + O(|Q − v0 |−2 ).
2J (Q, v0 )

Therefore, we get for the discrete derivative of 8π G(·; v0 ) at the face Q:

       
a −1 +b−1 a −1 −b−1
Re a+b
Q−v0
+ Re J (Q,v0 )
Re a−b
Q−v0
− Re J (Q,v0 )
K v0 (Q) = λ Q + λ̄ Q
a+b a−b
−2
+ O(|Q − v0 | )
1 1
= + + O(|Q − v0 |−2 )
Q − v0 ab J (Q, v0 )
  
1 a+b a−b 1 1
+ λQ + λ̄ Q +
2 a+b a−b Q − v0 ab J (Q, v0 )
1 τ (Q, v0 )
= + + O(|Q − v0 |−2 )
Q − v0 J (Q, v0 )
 
as |Q| → ∞. Here, λ Q = exp −i(ϕ Q − π/2) , and we have used the identity

λQ   a+b a−b a+b a−b

− = exp −2iϕ Q = · ⇔ λQ + λ̄ Q = 0.
λ̄ Q a − b a+b a + b a−b
118 A.I. Bobenko and F. Günther

(iii) Let h be the difference of two discrete Cauchy’s kernels with respect to v0
with asymptotics o(|Q − v0 |−1/2 ) as |Q| → ∞. K v0 in the second part is such a
discrete Cauchy’s kernel. Then, h is discrete holomorphic, and by Proposition 2.8, a
discrete primitive f on V (Λ) exists. By construction,

f (b+ ) − f (b− ) f (w+ ) − f (w− )

= = ∂Λ f (Q) = h(Q) = o(|Q − v0 |−1/2 ).
b+ − b− w+ − w−

Since angles and edge lengths of parallelograms are bounded, the conditions of
Theorem 2.31 are fulfilled, implying that f is biconstant, so h vanishes identically.
Since we do not have discrete Green’s functions on V (♦), we have to derive
a discrete Cauchy’s kernels on V (Λ) differently. To do so, we follow the original
approach of Kenyon using the discrete exponential [16] that was reintroduced by
Chelkak and Smirnov in [6].
Proposition 3.9 Let Q 0 ∈ V (♦). The function defined by

 0
1
K Q 0 (v) := log(λ)e(λ, v; Q 0 )dλ = 2 e(λ, v; Q 0 )dλ
πi
Cv −(v−Q 0 )∞

is a discrete Cauchy’s kernel with respect to Q 0 . Here, Cv is a collection of sufficiently

small counterclockwise oriented loops going once around each pole of e(·, v; Q 0 ).
On each loop around a pole e of e(·, v; Q 0 ), we take the branch of logarithm that
satisfies Im(log(e)) ∈ (θv − π, θv + π ), where θv := arg(v − Q 0 ). The second inte-
gration goes along the straight ray that ends in the origin in direction (v − Q 0 ).
Proof Lemma 4.2 assures that the arguments of all poles can be indeed chosen in
such a way that they lie in (θv − π, θv + π ).
By residue theorem, we can replace integration along Cv by an integration along a
circle centered at 0 with large radius R (such that all poles lie inside the disk) in coun-
terclockwise direction, an integration along a circle centered at 0 with small radius
r (such that all poles lie outside the disk) in clockwise direction, and integrations
along the two directions of the line segment [−R, −r ]. Using that

lim λe(λ, v; Q 0 ) log(λ) = 0

|λ|→∞

and taking r arbitrarily small, there is a homotopy between the new integration path
and the path from complex infinity to 0 and back along the ray spanned by −(v − Q 0 )
that does not change the value of the integral. The branch of log we consider jumps
by 2πi on the two sides of the ray, which shows

 0
1
log(λ)e(λ, v; Q 0 )dλ = 2 e(λ, v; Q 0 )dλ.
πi
Cv −(v−Q 0 )∞
Discrete Complex Analysis on Planar Quad-Graphs 119

In particular, θv can be replaced by any θv + 2kπ , k ∈ Z, in the first integration.

Let us consider the parallelogram Q 0 with vertices b− , w− , b+ , w+ in counter-
clockwise direction, starting with a black vertex, and let us denote a := w+ − b− ,
b := w− − b− . Without loss of generality, we rotate the complex plane in such a way
that arg(b+ − Q 0 ) = 0. Then, arg(a) and arg(b) lie between 0 and ±π , and using
residue theorem,

(log |a| + i arg(a) + iπ ) − (log |b| + i arg(b) + iπ )

K Q 0 (b− ) = 2 ,
b−a
(log |b| + i arg(b)) − (log |a| + i arg(a))
K Q 0 (b+ ) = 2 ,
b−a
(log |b| + i arg(b) + 2iπ ) − (log |a| + i arg(a) + iπ )
K Q 0 (w− ) = 2 ,
a+b
(log |a| + i arg(a)) − (log |b| + i arg(b) + iπ )
K Q 0 (w+ ) = 2 .
a+b

Finally, we get by Lemma 2.3 giving an integration formula for ∂¯Λ :

   
(a − b) K Q 0 (b+ ) − K Q 0 (b+ ) + (a + b) K Q 0 (w− ) − K Q 0 (w+ )
∂¯Λ K Q 0 (Q 0 ) =
4iar(FQ 0 )
π
= .
ar(FQ 0 )

By a similar argument as in the proof of Proposition 3.4, we can choose the same
contours of integration for all incident vertices v of a face Q = Q 0 . Then, the discrete
derivative ∂¯Λ commutes with the integration, and ∂¯Λ K Q 0 (Q) = 0 because e(λ, ·; Q 0 )
is discrete holomorphic by Proposition 3.2.

Theorem 3.10 Assume that there are α0 > 0 and E 1  E 0 > 0 such that α  α0
and E 1  e  E 0 for all interior angles α and side lengths e of parallelograms of
Λ. Let Q 0 ∈ V (♦) be fixed.
(i) The discrete Cauchy’s kernel K Q 0 given in Proposition 3.9 has the following
asymptotics as |v| → ∞:

1 τ (v, Q 0 )  
K Q 0 (v) = + + O |v − Q 0 |−3 .
v − Q0 J (v, Q 0 )

(ii) There is no further discrete Cauchy’s kernel with respect to Q 0 that has asymp-
totics o(|v − Q 0 |−1/2 ) as |v| → ∞.
(iii) For the discrete Cauchy’s kernel K Q 0 given in Proposition 3.9, ∂Λ K Q 0 has the
following asymptotics as |v| → ∞:

1 τ (Q, Q 0 )  
∂Λ K Q 0 (Q) = − − + O |Q − Q 0 |−3 .
(Q − Q 0 ) 2 J (Q, Q 0 ) 2
120 A.I. Bobenko and F. Günther

(iv) Up to two additive complex constants on Γ and Γ ∗ , there is no further dis-

crete Cauchy’s kernel with respect to Q 0 such that its discrete derivative has
asymptotics o(|Q − Q 0 |−1/2 ) as |Q| → ∞.
Here, constants in the O-notation depend on α0 , E 0 , and E 1 only.

Remark By Lemma 3.6(ii) and (iii),

τ (v, Q 0 )   τ (Q, Q 0 )  
= Ω (v − Q 0 )−1 and = Ω (Q − Q 0 )−2
J (v, Q 0 ) J (Q, Q 0 ) 2

as |v|, |Q| → ∞. As in the previous Theorem 3.8, we may replace the existence of
constants E 1  E 0 > 0 such that E 1  e  E 0 for all side lengths e of parallelo-
grams by the existence of q0 such that e/e  q0 for the two side lengths e, e of any
parallelogram of Λ since the latter implies the first assumption by Proposition 4.3.
Then, the constants in the O-notation depend instead of E 0 and E 1 on q0 , e0 , and e1 ,
where e0 and e1 are the side lengths of an arbitrary but fixed parallelogram of Λ.

The proof of the first part follows the ideas of Kenyon [16]. Similar to the proof
of Theorem 3.7, the path of integration is deformed into a path different from the
one Kenyon used, (−(v − Q 0 )∞, 0]. For the same reasons as before, his approach
does not generalize to parallelogram-graphs. The second and the fourth part of the
theorem are immediate consequences of Theorem 2.31; the third part is shown by a
direct computation.

Proof (i) Among all the vertices incident to Q 0 , let v0 be the one that is combinatori-
ally closest to v on Λ. Then, the poles d1 , d2 , e1 , . . . , ek(v) of e(·, v; Q 0 ) correspond
to the directed edges of a shortest path from v0 to v of length k(v) and the two directed
edges of Q 0 that point toward v0 . It is easy to deduce from Lemma 4.2 that the argu-
ments of all poles can be chosen to lie in [θ0 , θ0 + π − α0 ]. For more details, we refer
to the appendix in [13]. The claim is invariant under multiplication of the complex
plane, so we can assume that θ0 = −(π − α0 )/2. Then, there are no poles in the left
half-plane, such that we can reduce the calculation to an integration over R:

0 0
K Q 0 (v) = 2 e(λ, v; Q 0 )dλ = 2 e(λ, v; Q 0 )dλ.
−(v−Q 0 )∞ −∞

Since we are interested in the limit |v| → ∞, we take |v − Q 0 |  1 large√ enough

and split
√ the integration into
√ the three parts along
√ the intervals (−∞, −E 1 |v − Q 0 |],
[−E 1 |v − Q 0 |, −E 1 / |v − Q 0 |], [−E 1 / √ |v − Q 0 |, 0]. Lemma
√ 3.6(iv) shows
that the contribution of the range λ ∈ [−E 1 |v − Q 0 |, −E 1 / |v − Q 0 |] decays
faster to zero than any power of v − Q 0 as |v| → ∞. By Lemma 3.6(i) and (ii) we
know that k(v) = Ω(v − Q 0 ) and J (v, Q 0 ) = Ω(v − Q 0 ) as |v| → ∞. Since the
arguments of all the poles can be chosen to lie in [−(π − α0 )/2, (π − α0 )/2], it
follows even that Re(J (v, Q 0 )) = Ω(|v − Q 0 |) as |v| → ∞ stays positive.
Discrete Complex Analysis on Planar Quad-Graphs 121

√
Let λ ∈ [−E 1 / |v − Q 0 |, 0]. Then, λ → 0 as |v| → ∞. Repeated use of the
second and the third equation in Lemma 3.5(i) yields
 
e(λ, v; Q 0 ) = τ (v, Q 0 ) exp (2λJ (v, Q 0 )) 1 + O(λ2 ) + O(|v − Q 0 |λ3 )

as |v| → ∞. Thus, the integral near the origin behaves for |v| → ∞ as

0
τ (v, Q 0 )
2 e(λ, v; Q 0 )dλ = + O(|v − Q 0 |−3 ).
√
J (v, Q 0 )
−E 1 / |v−Q 0 |

√
For this, we used that exp(−2E 1 J (v, Q 0 )/ |v − Q 0 |) goes to zero exponentially
fast as |v| → ∞. √
Now, let λ ∈ (−∞, −E 1 |v − Q 0 |]. Then, λ → −∞ as |v| → ∞. Repeated use
of the second and the third equation in Lemma 3.5(ii) shows that
 
e(λ, v; Q 0 ) = λ−2 exp (2(v − Q 0 )/λ) 1 + O(λ−2 ) + O(|v − Q 0 |λ−3 )

as |v| → ∞. Using the result of above, we get for the integral near minus infinity:
√
−E 1  |v−Q 0 |

2 e(λ, v; Q 0 )dλ
−∞
√
−E 1  |v−Q 0 |
 
=2 λ−2 exp (2(v − Q 0 )/λ) 1 + O(λ−2 ) + O(|v − Q 0 |λ−3 ) dλ
−∞
0
 
=2 exp (2λ(v − Q 0 )) 1 + O(λ2 ) + O(|v − Q 0 |λ3 ) dλ
√
−1/(E 1 |v−Q 0 |)
1
= + O(|v − Q 0 |−3 ).
v − Q0

Summing the contributions of the three ranges up gives the desired result and also
shows the asymptotic behavior required in (ii).
(ii) The difference f of two discrete Cauchy’s kernels with respect to Q 0 of
asymptotics o(|v − Q 0 |−1/2 ) is discrete holomorphic and fulfills the conditions of
Theorem 2.31. Hence, f is biconstant, so f ≡ 0.
(iii) Let b− , w− , b+ , w+ denote the vertices of the parallelogram Q ∈ V (♦) in
counterclockwise order, starting with a black vertex. Let us introduce a := w+ − b−
and b := w− − b− . Using the expansion 1/(1 + x) = 1 − x + O(x 2 ) as x → 0 and
the asymptotics of the first part, we get that K Q 0 (b± ) and K Q 0 (w± ) equal
122 A.I. Bobenko and F. Günther

1 a+b τ (b− , Q 0 ) a −1 + b−1

∓ + ∓ τ (b− , Q 0 ) + O(|Q − Q 0 |−3 ),
Q − Q0 2(Q − Q 0 ) 2 J (Q, Q 0 ) 2J (Q, Q 0 )2
1 a−b τ (w− , Q 0 ) a −1 − b−1
∓ + ∓ τ (w− , Q 0 ) + O(|Q − Q 0 |−3 ),
Q − Q0 2(Q − Q 0 ) 2 J (Q, Q 0 ) 2J (Q, Q 0 )2

respectively. By definition, τ (b− , Q 0 ) = −τ (w− , Q 0 ). Therefore,

−1 −1
+b
a+b
(Q−Q 0 )2
− τ (b− , Q 0 ) Ja(Q,Q
0)
2
∂Λ K Q 0 (Q) = − λ Q
a+b
−1 −1
−b
a−b
(Q−Q 0 )2
− τ (b− , Q 0 ) Ja(Q,Q
0)
2
+ λ̄ Q + O(|Q − Q 0 |−3 )
a−b
1 τ (b− , Q 0 )
=− − + O(|Q − Q 0 |−3 )
(Q − Q 0 )2 ab J (Q, Q 0 )2
1 τ (Q, Q 0 )
=− − + O(|Q − Q 0 |−3 ).
(Q − Q 0 )2 J (Q, Q 0 )2

(iv) Let f be the difference of two discrete Cauchy’s kernels with respect to Q 0
whose discrete derivatives have asymptotics o(|Q − v0 |−1/2 ). Then, f is discrete
holomorphic and

f (b+ ) − f (b− ) f (w+ ) − f (w− )

= = ∂Λ f (Q) = o(|Q − v0 |−1/2 ).
b+ − b− w+ − w−

Since angles and edge lengths are bounded, the conditions of Theorem 2.31 are
fulfilled. Hence, f is biconstant.

Remark Note that Kenyon [16] and Chelkak and Smirnov [6] proved in the rhombic
setting the stronger result that there is a unique discrete Cauchy’s kernel on V (Λ)
with respect to Q 0 with asymptotics o(1) as |v| → ∞.

3.5 Integer Lattice

Let us consider a planar parallelogram-graph Λ such that each vertex has degree
four. With the embedding of ♦ described in Sect. 3.1, this happens if and only if ♦ is
a planar quad-graph or, equivalently, if Λ has the combinatorics of the integer lattice
Z2 . The vertices of ♦ lie at the midpoints of edges of Γ or Γ ∗ . Since any vertex
of Γ or Γ ∗ is enclosed by a quadrilateral of Γ ∗ or Γ , respectively, the faces of ♦
are parallelograms by Varignon’s theorem. Thus, ♦ becomes a planar parallelogram-
graph as well.
Of particular interest is the case that the two notions of discrete holomorphicity
on ♦, the one coming from ♦ being the dual of Λ and the other coming from the quad-
graph ♦ itself, coincide. It is not hard to show that this happens only for the integer
Discrete Complex Analysis on Planar Quad-Graphs 123

lattice of a skew coordinate system, onto which we restrict ourselves in the following.
For more details, see [13]. If e1 , e2 denote two spanning vectors, then ♦ is a parallel
shift of Λ by e1 /2 + e2 /2. Furthermore, the discrete derivatives on ♦ seen as the dual
of Λ coincide with the discrete derivatives on ♦ seen as a parallelogram-graph.
Since corresponding notions coincide and ♦ and Λ are congruent, we can skip all
subscripts Λ and ♦ in the definitions of discrete derivatives. Moreover, the discrete
Laplacian  is now defined for functions on V (Λ) and functions on V (♦) in the
same way. Due to Corollary 2.20, 4∂ ∂¯ =  = 4∂∂ ¯ is now true on both graphs. It
follows that all discrete derivatives ∂ n f of a discrete holomorphic function f are
discrete holomorphic themselves. Conversely, a discrete primitive exists for any
discrete holomorphic function on a simply-connected domain by Proposition 2.8.
Our main interests lie in giving discrete Cauchy’s integral formulae for higher
order derivatives of a discrete holomorphic function and determining the asymptotics
of higher order discrete derivatives of the discrete Cauchy’s kernel given in Sect. 3.4.
Note that due to the uniqueness statements in Theorems 3.8 and 3.10, both formulae
yield the same discrete Cauchy’s kernel.
Without loss of generality, we restrict our attention to functions on V (Λ). For the
ease of notation, we introduce the discrete distance D(·, ·) on V (Λ) ∪ V (♦) that is
induced by the | · |∞ -distance on the integer lattice spanned by e1 /2, e2 /2.

Theorem 3.11 Let Λ be the integer lattice spanned by the pair e1 , e2 of linearly
independent complex vectors. Let v0 ∈ V (Λ), Q 0 := v0 + e1 /2 + e2 /2 ∈ V (♦), let
f be a discrete holomorphic function on V (Λ), and let K v0 and K Q 0 be discrete
Cauchy’s kernels with respect to v0 and Q 0 , respectively. Let n be a nonnegative
integer and define x0 := v0 if n is even and x0 := Q 0 if n is odd. Similarly, let
x ∈ V (Λ) if n is even and x ∈ V (♦) if n is odd.
(i) For any counterclockwise oriented discrete contour C x0 in the medial graph X
enclosing all points x  ∈ V (Λ) ∪ V (♦) with D(x  , x0 )  n/2,

(−1)n
∂ f (x0 ) =
n
f ∂ n K x0 dz.
2πi
C x0

(ii) If K Q 0 is the discrete Cauchy’s kernel given in Proposition 3.9, then

(−1)n n 1 τ  (x, Q 0 )
∂ K Q 0 (x) = + + O(|x − Q 0 |−n−3 )
n! (x − Q 0 )n+1 (J (x, Q 0 )e1 e2 )n+1

as |x| → ∞, where τ  (x, Q 0 ) = 1 if x and Q 0 or (x + e1 /2 + e2 /2) and Q 0 can

be connected by a path on V (♦) of even length and τ  (x, Q 0 ) = −1 otherwise.
Constants in the O-notation depend on the spanning vectors e0 , e1 only.
124 A.I. Bobenko and F. Günther

Remark By Lemma 3.6(ii) and (iii), we have for |x| → ∞:

τ  (x, Q 0 )  
= Ω (x − Q 0 )−(n+1) .
(J (x, Q 0 )e1 e2 ) n+1

Proof (i) Let D be the discrete domain in F(X ) bounded by C x0 . By the assumptions
¯ n−1 K x0 d z̄ vanishes on C x0 . Thus,
on C x0 , the discrete one-form ∂∂
       
f ∂ n K x0 dz = f d ∂ n−1 K x0 = d( f d(∂ n−1 K x0 )) = d f ∧ d ∂ n−1 K x0
C x0 C x0 D D

by discrete Stokes’ Theorem 2.9 in the second equation and Theorem 2.16 and Propo-
sition 2.10 stating that d is a derivation for the discrete wedge product and dd f = 0
in the third
 n−1equation.
 Now, f is discrete holomorphic, meaning that ∂¯ f ≡ 0, so
d f ∧ d ∂ K x0 = ∂ f ∂∂ ¯ n−1 K x0 Ω♦ . But since the discrete derivatives commute
according to Corollary 2.11, ∂∂ ¯ n−1 K x0 = ∂ n−1 ∂¯ K x0 vanishes outside C x0 , so by
repeated use of Proposition 2.7 stating that ∂ is the formal adjoint of −∂, ¯

f ∂ n K x0 dz = −2i ∂ f, ∂¯ n−1 ∂ K̄ x0  = 2i(−1)n ∂ n f, ∂ K̄ x0  = 2πi(−1)n ∂ n f (x0 ).
C x0

(ii) Let us define the discrete exponential e(λ, Q; Q 0 ) for Q ∈ V (♦) in the same
way as a rational function in the complex variable λ as we defined e(λ, v; v0 ). By
inductive use of the formula for the discrete derivative of exp in Proposition 3.2 and
exp(λ, ·; Q 0 ) = e(2/λ, ·; Q 0 ), we get for the discrete exponential e(λ, ·; Q 0 ) that is
defined on V (Λ):

(2λ)n
(∂ n e(λ, ·; Q 0 ))(x) = e(λ, x; Q 0 ).
((λ − e1 )(λ − e2 )(λ + e1 )(λ + e2 ))n/2

To compute ∂ n K Q 0 (x), we need the values of K Q 0 only at vertices v ∈ V (Λ) with

D(v, x)  n/2. For these points,
 0
K Q 0 (v) = 2 e(λ, v; Q 0 )dλ
−(v−Q 0 )∞

by the formula in Proposition 3.9. If D(x, Q 0 ) is sufficiently large, then we can

choose e ∈ {e1 , e2 , −e1 , −e2 } such that D(ke, x) > n/2 for all positive integers k.
So if D(v, x)  n/2 and D(x, Q 0 ) is sufficiently large, then e(·, v; Q 0 ) does not
have the pole e. Since ±e1 , ±e2 are the only possible poles, it follows that there are
no poles in the convex hull of all the rays (−(v − Q 0 )∞, 0] for the vertices v with
D(v, x)  n/2. By residue theorem, we can use the same path of integration for all
relevant values of K Q 0 and get
Discrete Complex Analysis on Planar Quad-Graphs 125

0
∂ K Q 0 (x) = 2
n
(∂ n e(λ, ·; Q 0 ))(x)dλ.
−(x−Q 0 )∞

Now, we follow the proof of Theorem 3.10(i). Again, the claim is invariant under
rotation of the complex plane, so we may assume (x − Q 0 ) > 0.
Let E 1 := max {|e1 |, |e2 |} and E 0 := min {|e1 |, |e2 |}. For |x − Q√0 |  1 large
enough,
√ we split the integration
√ into the three parts
√ along (−∞, −E 1 |x − Q 0 |],
[−E 1 |x − Q 0 |, −E 1 / |x − Q 0 |], and [−E 1 / |x − √ 0 Q |, 0]. √
By Lemma 3.6(iv), the contribution of λ ∈ [−E 1 |x − Q 0 |, −E 1 / |x − Q 0 |]
decays faster to zero than any power of x − Q 0 as |x| → ∞. By Lemma 3.6(i) and (ii),
we know that D(x, Q 0 ) = Ω(x − Q 0 ) and J (x, Q 0 ) = Ω(x − Q 0 ). Furthermore,
the choice of (x − Q 0 ) > 0 implies that Re(J (x, Q 0 )) = Ω(x − Q 0 ) stays positive
as |x| → ∞. √
Let λ ∈ [−E 1 / |x − Q 0 |, 0]. Then, λ → 0 as |x| → ∞, and repeated use of the
second and third equation in Lemma 3.5(i) gives that (∂ n e(λ, ·; Q 0 ))(x) equals

τ  (x, Q 0 )  
(2λ)n exp (2λJ (x, Q 0 )) 1 + O(λ2 ) + O(|x − Q 0 |λ3 ) .
(e1 e2 )n+1

With a similar argument as in the proof of Theorem 3.10(i), the integral near the
origin behaves for |x| → ∞ as

(−1)n n!τ  (x, Q 0 )

+ O(|x − Q 0 |−n−3 ).
(e1 e2 J (x, Q 0 ))n+1
√
Now, let λ ∈ (−∞, −E 1 |x − Q 0 |]. For |x| → ∞, λ → −∞, and repeated use
of the second and third equation in Lemma 3.5(ii) implies that (∂ n e(λ, ·; Q 0 ))(x)
equals
 
n −2n−2 2(x − Q 0 )  
(2λ) λ exp 1 + O(λ−2 ) + O(|x − Q 0 |λ−3 ) .
λ

Using the result of above, we get for the integral near minus infinity for |x| → ∞:
√
−E 1  |x−Q 0 |

(∂ n e(λ, ·; Q 0 ))(x)dλ
−∞
√
−E 1  |x−Q 0 | n  
2 −2 2(x − Q 0 )  
=2 λ exp 1 + O(λ−2 ) + O(|x − Q 0 |λ−3 ) dλ
λ λ
−∞
126 A.I. Bobenko and F. Günther

0
 
=2 (2λ)n exp (2λ(x − Q 0 )) 1 + O(λ2 ) + O(|x − Q 0 |λ3 ) dλ
√
−1/(E 1 |x−Q 0 |)
(−1) n!
n
= + O(|x − Q 0 |−n−3 ).
(x − Q 0 )n+1

Summing up the contributions of the three ranges gives

(−1)n n 1 τ  (x, Q 0 )
∂ K Q 0 (x) = + + O(|x − Q 0 |−n−3 ).
n! (x − Q 0 )n+1 (J (x, Q 0 )e1 e2 )n+1

Acknowledgments The authors thank Richard Kenyon, Christian Mercat, and Mikhail Skopenkov
for fruitful discussions and helpful suggestions. In addition, we are grateful to Mikhail Skopenkov
for his detailed review of previous versions and his numerous and valuable recommendations.
The first author was partially supported by the DFG Collaborative Research Center TRR 109,
“Discretization in Geometry and Dynamics”. The research of the second author was supported
by the Deutsche Telekom Stiftung. Some parts of this paper were written at the IHÉS in Bures-
sur-Yvette, the INIMS in Cambridge, and the ESI in Vienna. The second author thanks the EPDI
program for the opportunity to stay at these institutes. The stay at the Isaac Newton Institute for
Mathematical Sciences was funded through an Engineering and Physical Sciences Research Council
Visiting Fellowship, Grant EP/K032208/1.

Appendix: Planar Parallelogram-Graphs

The aim of this appendix is to discuss some combinatorial and geometric properties
of parallelogram-graphs that were used in Sect. 3. The following notion of a strip is
standard, see for example the book [3].

Definition A strip in a planar quad-graph Λ is a path on its dual ♦ such that two
successive faces share an edge and the strip leaves a face in the opposite edge where
it enters it. Moreover, strips are assumed to have maximal length, i.e., there are no
strips containing it apart from itself.

Note that a strip is uniquely determined by the edges it passes through, meaning
the edges two successive faces share.

Definition For a strip S of a parallelogram-graph Λ, there exists a complex vector

a S such that any (nonoriented) edge through which S passes is ±a S . We call a S a
common parallel.

a S is unique up to sign; the choice of the sign induces an orientation on all edges.
The parallel edges of the strip can be rescaled to length |a S | = 1, without changing the
combinatorics. Hence, rhombic planar quad-graphs and planar parallelogram-graphs
are combinatorially equivalent. Rhombic planar quad-graphs are characterized by the
following proposition of Kenyon and Schlenker [17]:
Discrete Complex Analysis on Planar Quad-Graphs 127

Proposition 4.1 A planar quad-graph Λ admits a combinatorially equivalent

embedding in C with all rhombic faces if and only if the following two conditions
are satisfied:
• No strip crosses itself or is periodic.
• Two distinct strips cross each other at most once.

Remark Let us prove the simpler claim that planar parallelogram-graphs fulfill these
two conditions as was already noted by Kenyon in [16]. The underlying reason is
that any strip S is monotone with respect to the direction ia S : The coordinates of
the endpoints of the edges parallel to a S are strictly increasing or strictly decreasing
if they are projected to ia S . Whether the projections are decreasing or increasing
depends on the direction in which the faces of S are passed through. Without loss
of generality, we assume that the faces of S are passed through in such a way that
the projections of the corresponding coordinates are strictly increasing. For Q ∈ S,
let S Q denote the semi-infinite part of S starting in the quadrilateral Q that passes
through the faces of S in the same order.
As a consequence, no strip crosses itself or is periodic. Furthermore, S divides
the complex plane into two unbounded regions, to one is a S pointing and to the other
−a S . When a distinct strip S  crosses S, it enters a different region determined by S,
say it goes to the one to which a S is pointing. Due to monotonicity, the angle between
ia S  and a S is less than π/2. It follows that S  cannot cross S another time, since it
would then go to the region −a S is pointing to, contradicting that the angle between
ia S  and −a S is greater and not less than π/2.

In order to give explicit formulae for the discrete Green’s function and the discrete
Cauchy’s kernels in Sects. 3.3 and 3.4, we chose a particular directed path connecting
two vertices (or a face and a vertex) by edges of the parallelogram-graph Λ. This
path was monotone in one direction. The existence of such a path follows from the
following lemma, generalizing a result of [2] to general parallelogram-graphs. The
proof bases on the same ideas.

Lemma 4.2 Let Λ be a parallelogram-graph and let v0 ∈ V (Λ) be fixed. For a

directed edge e of Λ starting in v0 , consider the subgraph Ue ⊂ Λ that consists of
all vertices and edges of directed paths on Λ starting in v0 whose oriented edges
have arguments that can be chosen to lie in [arg(e), arg(e) + π ).
Then, the union of all Ue , e a directed edge starting in v0 , covers the whole quad-
graph. If there is a constant α0 > 0 such that α  α0 for all interior angles α of
faces of Λ, then the same statement holds true if [arg(e), arg(e) + π ) is replaced by
[arg(e), arg(e) + π − α0 ].

Proof Let us rescale the edges such that all of them have length one. By this, we
change neither the combinatorics of Λ nor the size of interior angles.
For a directed edge e starting in v0 , let Ue− and Ue+ denote the (directed) paths
on Λ starting in v0 , obtained by choosing the directed edge with the least or largest
128 A.I. Bobenko and F. Günther

argument in [arg(e), arg(e) + π ) (or [arg(e), arg(e) + π − α0 ]) at a vertex, respec-

tively. We first show that all vertices in between Ue− and Ue+ are contained in V (Ue ).
Then, it follows that Ue is the conical sector with boundary Ue− and Ue+ .
Suppose the contrary, i.e., suppose that there is a vertex v between Ue− and Ue+ to
which v0 cannot be connected by a directed path of edges whose arguments lie all in
the interval [arg(e), arg(e) + π ) (or [arg(e), arg(e) + π − α0 ]). Let the combinato-
rial distance between v0 and v be minimal among all such vertices.
In the case that interior angles of rhombi are bounded by α0 from below, they
are bounded from above by π − α0 . Hence, there is a vertex v adjacent to v such
that the argument of the directed edge v v lies in [arg(e), arg(e) + π − α0 ]. Even if
interior angles are not uniformly bounded, v can be chosen in such a way that the
argument of v v lies in [arg(e), arg(e) + π ). By construction, v is not in V (Ue ), but
still between Ue− and Ue+ . Let us look at the strip S passing through v v. Suppose
that the common parallel a S points from v to v.
If S intersects Ue− or Ue+ , then an edge parallel to a S is contained in Ue− or Ue+ ,
respectively. By construction, v0 and v then lie on the same side of the strip S.
If S does neither intersect Ue− nor Ue+ , then it is completely contained in the left
half space determined by the oriented line v0 + te, t ∈ R, as Ue− and Ue+ are. Suppose
S intersects the ray v0 + ta S , t  0. Again, it follows that v0 and v lie on the same
side of S.
It remains the case that S neither intersects Ue− , Ue+ , nor the ray v0 + ta S , t  0.
Consider the quadrilateral area R in between the parallels v0 + ta S , v + ta S and
v0 + te, v + te, t ∈ R. By assumption, the semi-infinite part of S that starts with the
edge v v and then goes into R does not intersect an edge of R incident to v0 , and by
monotonicity, it does not intersect v + ta S again. Now, Λ is locally finite, such that
only finitely many quadrilaterals of S are inside P. Thus, S leaves P on the edge
v + te, t ∈ R, and it follows that S separates v0 and v.
So in any case, S separates v0 from v. Any shortest path P connecting both points
has to pass through S. Let w be the first point of P that lies on the same side of S as
v does. Any strip passing through an edge on the shortest path connecting w and v
on S has to intersect P as well. It follows that replacing all edges of P on the same
side of S as v by the path from w to v does not change its length. But then, v is
combinatorially nearer to v0 than v, contradiction.
Finally, we can cyclically order the directed edges starting in v0 according to their
slopes. Then, the sectors Ue are interlaced, i.e., Ue contains both Ue+− and Ue−+ , where
e− , e, e+ are consecutive according to the cyclic order. As a consequence, the union
of all these Ue covers the whole Λ.

To perform our computations in Sects. 3.3 and 3.4, we needed not only that the
interior angles were bounded, but also that the side lengths were bounded. We can
relax the latter condition to boundedness of the ratio of side lengths.

Proposition 4.3 Let Λ be a parallelogram-graph and assume that there are con-
stants α0 , q0 > 0 such that α  α0 and e/e  q0 for all interior angles α and two
side lengths e, e of any parallelogram of Λ. Then, E 1  e  E 0 for all edge lengths
Discrete Complex Analysis on Planar Quad-Graphs 129

e, where E 0 := e0 q0N and E 1 := e1 q0−N with N := 2π/α0 /2 and e1  e0 being

the side lengths of an arbitrary parallelogram of Λ.

Proof Let Q  ∈ V (♦) be fixed with edge lengths e1  e0 and let Q ∈ V (♦) be
another parallelogram with center x. In the following, we construct a sequence of
n strips such that any two consecutive strips are crossing each other, the first one
contains Q  , the last one contains Q, and n  N . Then, it follows that the side lengths
of Q are bounded by E 0 and E 1 .
Let S0 be a strip containing Q  . If Q ∈ S0 , we are done. Otherwise, we choose
the common parallel a S0 in such a way that x lies in the region −a S0 is pointing to.
Let β0 := arg(a S0 ). Since S0 is monotone in the direction ia S0 and interior angles are
uniformly bounded, the ray x + ta S0 , t > 0, intersects S in exactly one line segment.
Let y0 be the first intersection point and Q 0 a quadrilateral of S containing y0 .
Because Λ is locally finite, the line segment connecting x and y0 intersects only
finitely many parallelograms. Through any such parallelogram at most two strips are
passing. Thus, only a finite number of strips intersect this line segment. Therefore,
we can choose a strip S1 intersecting S0Q 0 in a parallelogram Q 0,1 such that S1 does
not contain Q and does not intersect the line segment connecting x and y0 . Moreover,
Q
we require that Q  ∈ / S0 0,1 . Now, choose the common parallel a S1 of S1 in such a way
that there is an argument β1 of a S1 that satisfies π + β0 > β1 > β0 . By construction,
Q
x lies in the region −a S1 is pointing to. Note that S1 0,1 cannot cross S0 a second time.
Suppose we have already constructed the strip Sk with common parallel a Sk and
argument βk , k > 0, and x lies in the region −a Sk is pointing to. Sk shall not intersect
the line segments connecting x and y0 or connecting x and yk−1 . Moreover, assume
Q
that the semi-infinite part Sk k−1,k starting in the intersection Q k−1,k of Sk with Sk−1
does not cross S0 .
Let yk be the first intersection of the ray x + ta Sk , t > 0, with a quadrilateral Q k
of the strip Sk . By the same arguments as above, there exists a strip Sk+1 intersect-
Q
ing Sk k−1,k ∩ SkQ k that does not contain Q and does not intersect the line segments
connecting x and y0 or x and yk . Choose its common parallel a Sk+1 in such a way
that it has an argument βk+1 that satisfies π + βk > βk+1 > βk . By construction, x
Q k,k+1
lies in the region −a Sk+1 is pointing to. If the semi-infinite part Sk+1 starting in the
intersection Q k,k+1 with Sk does not cross S0 , then we continue this procedure. For
a schematic picture of the proof, see Fig. 7.
Q
After at most l := 2π/α0  steps, we end up with a strip Sl such that Sl l−1,l
intersects S0 . Indeed, let us suppose the contrary, that is, let us suppose that all
Q Q
S2 1,2 , . . . , Sl l−1,l do not cross S0 .
By assumption, βk + π − α0  βk+1  βk + α0 . It follows that the first j such
that β j is greater or equal than β0 + 2π satisfies j  2π/α0 . In addition, we have
β j < β0 + 3π − α0 .
By construction, S j does not intersect the line segment connecting x and y j−1 .
Q Q
Moreover, S j j−1, j cannot cross S j−1 a second time. It follows that S j j−1, j cannot
intersect the ray x + ta S j−1 , t > 0.
130 A.I. Bobenko and F. Günther
S1

aS1 y1
Qk−1,k y
k

Q0,1

x y0

aSk
a S0

Sk S0
aSk+1
Qk,k+1
Sk+1

Fig. 7 Schematic picture of the proof of Proposition 4.3

Q j−1, j
Also, S j does neither cross S0 nor does it intersect the line segment connecting
Q
x and y0 , so it does not intersect the ray x + ta S0 , t > 0. Thus, S j j−1, j is contained in
the cone with tip x spanned by a S j−1 and a S0 (with angle less than π ). This contradicts
Q
the monotonicity of S j j−1, j into the direction ia S j , because the ray x + ta S j , t > 0,
is not contained in the interior of the cone above.
In summary, we found a cycle of m strips S0 , S1 , . . . , Sm−1 surrounding x, where
m  2π/α0  + 1. Actually, m  2π/α0 , because the a Sk are cyclically ordered.
Since only finitely many strips intersect the strip S0 in between Q  and Q 0,1 , we can
Q
assume that Q  is contained in S0 m−1,0 .
These m strips determine a bounded region x is contained in. If Q  = Q m−1,0 ,
then we look at the semi-infinite part of the strip S̃0 different from S0 that passes
through Q  and goes into the interior of the bounded region above. It has to intersect
one of the strips S1 , . . . , Sm−1 , say Sk . Then, S0 , . . . , Sk , S̃0 or S̃0 , Sk , . . . , Sm−1 , S0
determine a bounded region x is contained in (Q may be an element of S̃0 ). Clearly,
they are at most 2π/α0  such strips, and Q  lies on an intersection.
If Q ∈ / S̃0 , then a strip S Q containing Q has to cross two different strips of the
cycle due to local finiteness. In the same way as above, we can find a cycle of at
most m   2π/α0  strips S0 , S1 , . . . , Sm  −1 such that Q lies on one of the strips,
say Sk , and the intersection of S0 and Sm  −1 is Q  . If k  m  /2, then we choose the
sequence of strips S0 , S1 , . . . , Sk ; otherwise, we take Sm  −1 , Sm  −2 , . . . , Sk . Any two
consecutive strips are crossing each other, Q  is on the first strip, Q on the last one,
and there are at most 2π/α0 /2 of them.

Remark In general, the bound 2π/α0 /2 in the proof is optimal. Indeed, consider
n rays emanating from 0 such that the angle between any two neighboring rays is
2π/n. In each of the n segments, choose the quad-graph combinatorially equivalent
to the positive octant of the integer lattice that is spanned by two consecutive rays.
Discrete Complex Analysis on Planar Quad-Graphs 131

For example, if n = 4, we obtain Z2 . Then, any strip passes through exactly two
adjacent segments, and n/2 is the optimal bound.
Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Bobenko, A., Günther, F.: Discrete Riemann surfaces based on quadrilateral cellular decom-
positions (2015). Preprint arXiv:1511.00652
2. Bobenko, A., Mercat, C., Suris, Y.: Linear and nonlinear theories of discrete analytic functions.
Integrable structure and isomonodromic Green’s function. J. Reine Angew. Math. 583, 117–161
(2005)
3. Bobenko, A., Suris, Y.: Discrete differential geometry: integrable structure. In: Graduate Studies
in Mathematics, vol. 98. AMS, Providence (2008)
4. Brooks, R., Smith, C., Stone, A., Tutte, W.: The dissection of rectangles into squares. Duke
Math. J. 7(1), 312–340 (1940)
5. Bücking, U.: Approximation of conformal mappings by circle patterns. Geom. Dedicata 137,
163–197 (2008)
6. Chelkak, D., Smirnov, S.: Discrete complex analysis on isoradial graphs. Adv. Math. 228,
1590–1630 (2011)
7. Chelkak, D., Smirnov, S.: Universality in the 2D Ising model and conformal invariance of
fermionic observables. Invent. Math. 189(3), 515–580 (2012)
8. Courant, R., Friedrichs, K., Lewy, H.: Über diepartiellen Differenzengleichungen der mathe-
matischen Physik. Math. Ann. 100, 32–74 (1928)
9. Duffin, R.: Basic properties of discrete analytic functions. Duke Math. J. 23(2), 335–363 (1956)
10. Duffin, R.: Potential theory on a rhombic lattice. J. Comb. Th. 5, 258–272 (1968)
11. Dynnikov, I., Novikov, S.: Geometry of the triangle equation on two-manifolds. Moscow Math.
J. 3(2), 419–482 (2003)
12. Ferrand, J.: Fonctions préharmoniques et fonctions préholomorphes. Bull. Sci. Math. Ser. 2(68),
152–180 (1944)
13. Günther, F.: Discrete Riemann surfaces and integrable systems. Ph.D. thesis, Technische Uni-
versität Berlin (2014). http://opus4.kobv.de/opus4-tuberlin/files/5659/guenther_felix.pdf
14. Isaacs, R.: A finite difference function theory. Univ. Nac. Tucumán. Rev. A 2, 177–201 (1941)
15. Kenyon, R.: Conformal invariance of domino tiling. Ann. Prob. 28(2), 759–795 (2002)
16. Kenyon, R.: The Laplacian and Dirac operators on critical planar graphs. Invent. Math. 150,
409–439 (2002)
17. Kenyon, R., Schlenker, J.M.: Rhombic embeddings of planar quad-graphs. Trans. Amer. Math.
Soc. 357(9), 3443–3458 (2005)
18. Lelong-Ferrand, J.: Représentation conforme et transformations à intégrale de Dirichlet bornée.
Gauthier-Villars, Paris (1955)
132 A.I. Bobenko and F. Günther

19. Mercat, C.: Discrete Riemann surfaces and the Ising model. Commun. Math. Phys. 218(1),
177–216 (2001)
20. Mercat, C.: Discrete Riemann surfaces. In: Handbook of Teichmüller theory. Vol. I, IRMA Lect.
Math. Theor. Phys., vol. 11, pp. 541–575. Eur. Math. Soc., Zurich (2007)
21. Mercat, C.: Discrete complex structure on surfel surfaces. Proceedings of the 14th IAPR Inter-
national Conference on Discrete Geometry for Computer Imagery. DGCI’08, pp. 153–164.
Springer-Verlag, Berlin, Heidelberg (2008)
22. Rodin, B., Sullivan, D.: The convergence of circle packings to the Riemann mapping. J. Diff.
Geom. 26(2), 349–360 (1987)
23. Skopenkov, M.: The boundary value problem for discrete analytic functions. Adv. Math. 240,
61–87 (2013)
24. Smirnov, S.: Discrete complex analysis and probability. In: Proceedings of the International
Congress of Mathematicians 2010 (ICM 2010), vol. I: Plenary Lectures and Ceremonies, vols.
II-IV: Invited Lectures, pp. 595–621. Hindustan Book Agency, New Delhi, India (2000)
25. Smirnov, S.: Critical percolation in the plane: conformal invariance, Cardy’s formula, scaling
limits. C. R. Math. Acad. Sci. Paris Sér. I 333(3), 239–244 (2001)
26. Smirnov, S.: Conformal invariance in random cluster models. I. Holomorphic fermions in the
Ising model. Ann. Math. 172(2), 1435–1467 (2010)
27. Whitney, H.: On products in a complex. Ann. Math. 39(2), 397–432 (1938)
Approximation of Conformal Mappings
Using Conformally Equivalent Triangular
Lattices

Ulrike Bücking

Abstract Two triangle meshes are conformally equivalent if their edge lengths are
related by scale factors associated to the vertices. Such a pair can be considered
as preimage and image of a discrete conformal map. In this article we study the
approximation of a given smooth conformal map f by such discrete conformal maps
f ε defined on triangular lattices. In particular, let T be an infinite triangulation of
the plane with congruent strictly acute triangles. We scale this triangular lattice by
ε > 0 and approximate a compact subset of the domain of f with a portion of it. For
ε small enough we prove that there exists a conformally equivalent triangle mesh
whose scale factors are given by log | f  | on the boundary. Furthermore we show
that the corresponding discrete conformal (piecewise linear) maps f ε converge to f
uniformly in C 1 with error of order ε.

1 Introduction

Holomorphic functions build the basis and heart of the rich theory of complex analy-
sis. Holomorphic functions with nowhere vanishing derivative, also called conformal
maps, have the property to preserve angles. Thus they may be characterized by the
fact that they are infinitesimal scale-rotations.
In the discrete theory, the idea of characterizing conformal maps as local scale-
rotations may be translated into different concepts. Here we consider the discretiza-
tion coming from a metric viewpoint: Infinitesimally, lengths are scaled by a factor,
i.e. by | f  (z)| for a conformal function f on D ⊂ C. More generally, on a smooth
manifold two Riemannian metrics g and g̃ are conformally equivalent if g̃ = e2u g
for some smooth function u.

U. Bücking (B)
Inst. für Mathematik, Technische Universität Berlin, Straße des 17. Juni 136,
10623 Berlin, Germany
e-mail: [email protected]

© The Author(s) 2016 133

A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_3
134 U. Bücking

(a) (b)

Fig. 1 Lattice triangulation of the plane with congruent triangles. a Example of a triangular lattice.
b Acute angled triangle.

The smooth complex domain (or manifold) is replaced in this discrete setting by
a triangulation of a connected subset of the plane C (or a triangulated piecewise
Euclidean manifold).

1.1 Convergence for Discrete Conformal PL-Maps

on Triangular Lattices

In this article we focus on the case where the triangulation is a (part of a) triangular
lattice. In particular, let T be a lattice triangulation of the whole complex plane C with
congruent triangles, see Fig. 1a. The sets of vertices and edges of T are denoted by V
and E respectively. Edges will often be written as e = [vi , v j ] ∈ E, where vi , v j ∈
V are its incident vertices. For triangular faces we use the notation Δ[vi , v j , vk ]
enumerating the incident vertices with respect to the orientation (counterclockwise)
of C.
On a subcomplex of T we now define a discrete conformal mapping. The main
idea is to change the lengths of the edges of the triangulation according to scale
factors at the vertices. The new triangles are then ‘glued together’ to result in a
piecewise linear map, see Fig. 2 for an illustration. More precisely, we have
Definition 1.1 A discrete conformal PL-mapping g is a continuous and orientation
preserving map of a subcomplex TS of a triangular lattice T to C which is locally
a homeomorphism in a neighborhood of each interior point and whose restriction
to every triangle is a linear map onto the corresponding image triangle, that is the
mapping is piecewise linear. Furthermore, there exists a function u : VS → R on
the vertices, called associated scale factors, such that for all edges e = [v, w] ∈ E S
there holds
|g(v) − g(w)| = |v − w|e(u(v)+u(w))/2 , (1)

where |a| denotes the modulus of a ∈ C.

Approximation of Conformal Mappings … 135

Fig. 2 Example of a discrete conformal PL-map g

Note that Eq. (1) expresses a linear relation for the logarithmic edge lengths, that is

2 log |g(v) − g(w)| = 2 log |v − w| + u(v) + u(w).

In fact, the definition of a discrete conformal PL-map relies on the notion of

discrete conformal triangle meshes. These have been studied by Luo, Gu, Sun, Wu,
Guo [8, 9, 14], Bobenko, Pinkall, and Springborn [1] and others.
As possible application, discrete conformal PL-maps can be used for discrete
uniformization. The simplest case is a discrete Riemann mapping theorem, i.e. the
problem of finding a discrete conformal mapping of a simply connected domain onto
the unit disc. Similarly, we may consider a related Dirichlet problem. Given some
function u ∂ on the boundary of a subcomplex TS , find a discrete conformal PL-map
whose associated scale factors agree on the boundary with u ∂ . For such a Dirichlet
problem (with assumptions on u ∂ and TS ) we will prove existence as part of our
convergence theorem.
In this article we present a first answer to the following problem: Given a smooth
conformal map, find a sequence of discrete conformal PL-maps which approximate
the given map. We study this problem on triangular lattices T with acute angles and
always assume for simplicity that the origin is a vertex. Denote by εT the lattice T
scaled by ε > 0. Using the values of log | f  |, we obtain a discrete conformal PL-map
f ε on a subcomplex of εT from a boundary value problem for the associated scale
factors. More precisely, we prove the following approximation result.

Theorem 1.2 Let f : D → C be a conformal map (i.e. holomorphic with f  = 0).

Let K ⊂ D be a compact set which is the closure of its simply connected interior
int (K ) and assume that 0 ∈ int (K ). Let T be a triangular lattice with strictly acute
angles. For each ε > 0 let TKε be a subcomplex of εT whose support is contained
in K and is homeomorphic to a closed disc. We further assume that 0 is an interior
vertex of TKε . Let e0 = [0, v̂0 ] ∈ E Kε be one of its incident edges.
136 U. Bücking

Then if ε > 0 is small enough (depending on K , f , and T ) there exists a unique

discrete conformal PL-map f ε on TKε which satisfies the following two conditions:
• The associated scale factors u ε : VKε → R satisfy

u ε (v) = log | f  (v)| for all boundary vertices v of VKε . (2)

• The discrete conformal PL-map is normalized according to f ε (0) = f (0) and

arg( f ε (v̂0 ) − f ε (0)) = arg(v̂0 ) + arg( f  ( v̂20 )) (mod 2π ).
Furthermore, the following estimates for u ε and f ε hold for all vertices v ∈ VKε and
points x in the support of TKε respectively with constants C1 , C2 , C3 depending only
on K , f , and T , but not on v or x:
(i) The scale factors u ε approximate log | f  | uniformly with error of order ε2 :
 ε 
u (v) − log | f  (v)|  C1 ε2 . (3)

(ii) The discrete conformal PL-mappings f ε converge to f for ε → 0 uniformly

with error of order ε:
| f ε (x) − f (x)|  C2 ε.

(iii) The derivatives of f ε (in the interior of the triangles) converge to f  uniformly
for ε → 0 with error of order ε:
 
∂z f ε (x) − f  (x)  C3 ε and |∂z̄ f ε (x)|  C3 ε

for all points x in the interior of a triangle Δ of TKε . Here ∂z and ∂z̄ denote the
Wirtinger derivatives applied to the linear maps f ε |Δ .

Note that the subcomplexes TKε may be chosen such that they approximate the
compact set K . Further notice that (3) implies that u ε converges to log | f  | in C 1
with error of order ε, in the sense that also
 ε   
 u (v) − u ε (w) f ((v + w)/2) 
 − Re  C̃ε
 ε f  ((v + w)/2) 

on edges [v, w] uniformly for some constant C̃.

The proof of Theorem 1.2 is given in Sect. 4. The arguments are based on estimates
derived in Sect. 3.
The problem of actually computing the scale factors u for given boundary values
u ∂ such that u gives rise to a discrete conformal PL-map (in case it exists) can be
solved using a variational principle, see [1, 20]. Our proof relies on investigations
using the corresponding convex functional, see Theorem 2.2 in Sect. 2.
Approximation of Conformal Mappings … 137

Remark 1.3 The convergence result of Theorem 1.2 also remains true if linear inter-
polation is replaced with the piecewise projective interpolation schemes described
in [1, 3], i.e., circumcircle preserving, angle bisector preserving and, generally,
exponent-t-center preserving for all t ∈ R. The proof is the same with only small
adaptations. This is due to the fact that the image of the vertices is the same for
all these interpolation schemes and these image points converge uniformly to the
corresponding image points under f with error of order ε. The estimates for the
derivatives similarly follow from Theorem 1.2(i).

1.2 Other Convergence Results for Discrete Conformal Maps

Smooth conformal maps can be characterized in various ways. This leads to different
notions of discrete conformality. Convergence issues have already been studied for
some of these discrete analogs. We only give a very short overview and cite some
results of a growing literature.
In particular, linear definitions can be derived as discrete versions of the Cauchy-
Riemann equations and have a long and still developing history. Connections of such
discrete mappings to smooth conformal functions have been studied for example
in [2, 6, 7, 13, 16, 19, 22].
The idea of characterizing conformal maps as local scale-rotations has lead to the
consideration of circle packings, more precisely to investigations on circle packings
with the same (given) combinatorics of the tangency graph. Thurston [21] first con-
jectured the convergence of circle packings to the Riemann map, which was then
proven by [10, 11, 17].
The theory of circle patterns generalizes the case of circle packings. Also, there
is a link to integrable structures via isoradial circle patterns. The approximation of
conformal maps using circle patterns has been studied in [4, 5, 12, 15, 18].
The approach taken in this article constructs discrete conformal maps from given
boundary values. Our approximation results and some ideas of the proof are therefore
similar to those in [4, 5, 18] for circle patterns which also rely on boundary value
problems.

2 Some Characterizations of Associated Scale Factors

of Discrete Conformal PL-Maps

Consider a subcomplex TS of a triangular lattice T and an arbitrary function u :

VS → R. Assign new lengths to the edges according to (1) by

˜
l([v, w]) = |v − w|e(u(v)+u(w))/2 (4)
138 U. Bücking

In order to obtain new triangles with these lengths (and ultimately a discrete con-
formal PL-map) the triangle inequalities need to hold for the edge lengths l˜ on each
triangle. If we assume this, we can embed the new triangles (respecting orientation)
and immerse sequences of triangles with edge lengths given by l˜ as in (4). In order
to obtain a discrete conformal PL-map, in particular a local homeomorphism, the
interior angles of the triangles need to sum up to 2π at each interior vertex. The angle
at a vertex of a triangle with given side lengths can be calculated. With the notation
of Fig. 1b we have the half-angle formula
 
α 
(−b + a + c)(−c + a + b) 1 − ( b − c )2
tan = =
b ac 2 a . (5)
2 (b + c − a)(a + b + c) (a + a ) − 1

The last expression emphasizes the fact that the angle does not depend on the scaling
of the triangle. Careful considerations of this angle function depending on (scaled)
side lengths of the triangle form the basis for our proof. In particular, we define the
function 
1 − (e−x/2 − e−y/2 )2
θ (x, y) := 2 arctan , (6)
(e −x/2 + e−y/2 )2 − 1

so (5) can be written as

b c
α = θ (x, y) with = e−x/2 and = e−y/2 .
a a
Summing up, we have the following characterization of scale factors associated
to discrete conformal PL-maps.

Proposition 2.1 Let TS be a subcomplex of a triangular lattice T and u : VS → R

a function satisfying the following two conditions.
(i) For every triangle Δ[v1 , v2 , v3 ] of TS the triangle inequalities for l˜ defined by (4)
hold, in particular

|vi − v j |e(u(vi )+u(v j ))/2 < |vi − vk |e(u(vi )+u(vk ))/2 + |v j − vk |e(u(v j )+u(vk ))/2 (7)

for all permutations (i jk) of (123).

(ii) For every interior vertex v0 with neighbors v1 , v2 , . . . , vk , vk+1 = v1 in cyclic
order we have


k
θ (λ(v0 , v j , v j+1 ) + u(v j+1 ) − u(v0 ), λ(v0 , v j+1 , v j ) + u(v j ) − u(v0 )) = 2π,
j=1
(8)
where λ(va , vb , vc ) = 2 log(|vb − vc |/|va − vb |) for a triangle Δ[va , vb , vc ].
Approximation of Conformal Mappings … 139

Then there is a discrete conformal PL-map (unique up to post-composition with

Euclidean motions) such that its associated scale factors are the given function
u : VS → R.
Conversely, given a discrete conformal PL-map on a subcomplex TS of a triangular
lattice T , its associated scale factors u : VS → R satisfy conditions (i) and (ii).

In order to obtain discrete conformal PL-maps from a given smooth conformal

map we will consider a Dirichlet problem for the associated scale factors. Therefore
we will apply a theorem from [1] which characterizes the scale factors u for given
boundary values using a variational principle for a functional E defined in [1, Sect. 4].
Note that we will not need the exact expression for E but only the formula for its
partial derivatives. In fact, the vanishing of these derivatives is equivalent to the
necessary condition (8) for the scale factors to correspond to a discrete conformal
PL-map.

Theorem 2.2 ([1]) Let TS be a subcomplex of a triangular lattice and let u ∂ : V∂ →

R be a function on the boundary vertices V∂ of TS . Then the solution ũ (if it exists)
of Eq. (8) at all interior vertices with ũ|V∂ = u ∂ is the unique argmin of a locally
strictly convex functional E(u) = E TS (u) which is defined for functions u : V → R
satisfying the inequalities (7).
The partial derivative of E with respect to u i = u(vi ) at an interior vertex vi ∈ Vint
with k neighbors vi1 , vi2 , . . . , vik vik+1 = vi1 in cyclic order is

k    
∂E li j+1 ,i j li j+1 ,i j
(u) = 2π − θ (2 log + u i j − u i , 2 log + u i j+1 − u i ),
∂u i j=1
li,i j+1 li,i j
(9)
where l j,k = |v j − vk |.
By Proposition 2.1 such a solution ũ are then scale factors associated to a discrete
conformal PL-map.

Remark 2.3 The functional E can be extended to a convex continuously differen-

tiable function on RV , see [1] for details.

3 Taylor Expansions

We now examine the effect when we take u = log | f  | as ‘scale factors’, i.e. for each
√  we multiply  the length |v − w| of an edge [v, w] by the geometric mean
triangle
| f (v) f (w)| of | f | at the vertices. The proof of Theorem 1.2 is based on the idea
that u = log | f  | almost satisfies the conditions for being the associated scale factors
of an discrete conformal PL-map, that is conditions (i) and (ii) of Proposition 2.1,
and therefore is close to the exact solution u ε .
140 U. Bücking

To be precise, suppose that εT is the equilateral triangulation

√
of the plane. Assume
without loss of generality that the edge lengths equal 2 ε > 0 and edges are parallel
3

to ei jπ/3 for j = 0, 1, . . . , 5. Let the conformal function f , the compact set K , and
the subcomplexes TKε (with vertices VKε and edges E Kε ) be given as in Theorem 1.2.
Let v0 ∈ VKε ,int be an interior vertex. Here and below VKε ,int denotes the set of interior
√ ij π
vertices having six neighbors in VKε . Denote the neighbors of v0 by v j = v0 + ε 3e2 3
and consider the triangle Δ j = Δ[v0 , v j , v j+1 ] for some j ∈ {0, 1, . . . , 5}. Taking
u = log | f  |, we obtain edge lengths of a new triangle Δ̃ j , i.e. satisfying (7), if ε is
small enough. Then the angle in Δ̃ j at the image vertex of v0 is given by
√ ij π √ i( j+1) π
3e 3 3e 3
θ(log | f  (v0 + ε  
)| − log | f (v0 )|, log | f (v0 + ε )| − log | f  (v0 )|)
2 2

according to (6). Summing up these angles—that is inserting log | f  | into (8) instead
of u at an interior vertex v0 ∈ VKε ,int —we obtain the function

Sv0 (ε) =
√ ij π √ i( j+1) π

5
3e 3 3e 3
θ (log | f  (v0 + ε )| − log | f  (v0 )|, log | f  (v0 + ε )| − log | f  (v0 )|)
2 2
j=0

We are interested in the Taylor expansion of Sv0 in ε. The symmetry of the lattice
T implies that Sv0 is an even function, so the expansion contains only even powers
of εn . Using a computer algebra program we arrive at

Sv0 (ε) = 2π + Cv0 ε4 + O(ε6 ). (10)

Here and below, the notation h(ε) = O(εn ) means that there is a constant C , such
that |h(ε)|  C εn holds for all small enough ε > 0. The constant of the ε4 -term is
√    
3 3 f
Cv0 = − Re S( f )(v0 ) (v0 ) ,
32 f

     2
where S( f ) = ff  − 21 ff  is the Schwarzian derivative of f . We will not need
the exact form of this constant, but only the fact that it is bounded on K .
Analogous results to (10) hold for all triangular lattices εT with edge lengths
a ε = ε sin α, bε = ε sin β, cε = ε sin γ , also if the angles are larger than π/2. We
assume without loss of generality the edge directions being parallel to 1, eiα and
ei(α+β) . Arguing as above, we consider the function
Approximation of Conformal Mappings … 141

sin α f  (v0 + ε sin β) sin α f  (v0 + ε sin γ eiα )

Sv0 (ε) = θ (2 log + log | |, 2 log + log | |)
sin γ f  (v0 ) sin β f  (v0 )
sin β f  (v0 + ε sin γ eiα ) sin β f  (v0 + ε sin α ei(α+β) )
+ θ (2 log + log | |, 2 log + log | |)
sin α f  (v0 ) sin γ f  (v0 )
sin γ f  (v0 + ε sin α ei(α+β) ) sin γ f  (v0 − ε sin β)
+ θ (2 log + log |  |, 2 log + log | |)
sin β f (v0 ) sin α f  (v0 )
sin α f  (v0 − ε sin β) sin α f  (v0 − ε sin γ eiα )
+ θ (2 log + log |  |, 2 log + log | )
sin γ f (v0 ) sin β f  (v0 )|
sin β f  (v0 − ε sin γ eiα ) sin β f  (v0 − ε sin α ei(α+β) )
+ θ (2 log + log |  |, 2 log + log | |)
sin α f (v0 ) sin γ f  (v0 )
sin γ f  (v0 − ε sin α ei(α+β) ) sin γ f  (v0 + ε sin β)
+ θ (2 log + log |  |, 2 log + log | |).
sin β f (v0 ) sin α f  (v0 )

Again, Sv0 is an even function. Using a computer algebra program we arrive at

Sv0 (ε) = 2π + Cv0 ε4 + O(ε6 ), (11)

with corresponding constant

   
sin α sin β sin γ f
Cv0 = − Re S( f )(v0 ) (v0 )
4 f
      
1 f  2 f   1 f  
+c(α, β, γ ) − ,
2 f f 3 f

where c(α, β, γ ) = cos β sin3 β + cos γ sin3 γ e2iα + cos α sin3 αe2i(α+β) .
Our key observation is that we can control the sign of the O(ε4 )-term in (10) if we
replace log | f  (x)| by log | f  (x)| + aε2 |x|2 , where a ∈ R is some suitable constant.
In particular, for positive constants M ± , C ± consider the functions

±  ± ± ±ε2 (M ± − C ± |v|2 ) for v ∈ VKε ,int ,
w = log | f | + q with q (v) =
0 for v ∈ ∂ VKε .

Here and below ∂ VKε denotes the set of boundary vertices of VKε . √
Then we obtain for equilateral triangulations with edge length 23 ε the following
Taylor expansion for all interior vertices v0 ∈ VKε ,int whose neighbors are also in
VKε ,int :
√ √

5
3 ij π 3 i( j+1) π
± ± ± 3 ) − w ± (v ))
θ (w (v0 + ε e ) − w (v0 ), w (v0 + ε
3 e 0
j=0
2 2
√
3 3 ± 4
= 2π + (Cv0 ∓ C )ε + O(ε5 ). (12)
2
142 U. Bücking

Fig. 3 Two adjacent

triangles of the triangular
lattice εT and orthogonal
edges e ∈ (εE) (solid) and
e∗ ∈ (εE ∗ ) (dashed)

Again, analogous results hold for all regular triangular lattices, where the corre-
sponding O(ε4 )-term then is

Cv0 ∓ 4 sin α sin β sin γ C ± .

For interior vertices v0 ∈ VKε ,int which are incident to k boundary vertices we
obtain instead of the right-hand side of (12):
√
3 ±
2π ∓ k (M − C ± |v0 |2 )ε2 + O(ε4 ).
4
For general triangular lattices we get for every edge e = [v0 , v j ] which is incident
to a boundary vertex v j ∈ ∂ VKε a term ∓(M ± − C ± |v0 |2 ) cos ϕe sin ϕe ε2 where ϕe is
the angle opposite to the edge e, see Fig. 3.
The following lemma summarizes the main properties of w± which follow from
the definition of w± together with the preceding estimates.

Lemma 3.1 w ± satisfies the boundary condition w ± |∂ VKε = log | f  |∂ V ε .
K
Furthermore, C ± > 0 and M ± > 0 can be chosen such that for all ε small enough
and all v0 ∈ VKε ,int :
(i) q + (v0 ) > 0 and q − (v0 ) < 0
(ii) If v1 , v2 , . . . , v6 , v7 = v1 denote the chain of neighboring vertices of v0 in
cyclic order and λ(va , vb , vc ) = 2 log(|vb − vc |/|va − vb |) for any triangle
Δ[va , vb , vc ], we have

6
θ (λ(v0 , v j+1 , v j ) + w+ (v j ) − w+ (v0 ), λ(v0 , v j , v j+1 ) + w+ (v j+1 ) − w+ (v0 )) < 2π,
j=1


6
θ (λ(v0 , v j+1 , v j ) + w− (v j ) − w− (v0 ), λ(v0 , v j , v j+1 ) + w− (v j+1 ) − w− (v0 )) > 2π
j=1
Approximation of Conformal Mappings … 143

The choices of C ± and M ± only depend on f (and its derivatives), K , and on the
angles of the triangular lattice T .

In analogy to the continuous case we interpret Eq. (8) as a non-linear Laplace

equation for u. In this spirit w+ may be taken as superharmonic function and w − as
subharmonic function.

4 Existence of Discrete Conformal PL-Maps and Estimates

The functions w ± have been introduced in order to ‘catch’ the solution u ε in the
following compact set:

W ε = {u : VKε → R | u(v) = log | f  (v)| for all v ∈ ∂ VKε ,

w − (v)  u(v)  w + (v) for all v ∈ VKε ,int }.

Note that W ε is a n-dimensional interval in Rn for n = |VKε | = number of vertices,

if we identify a function u : VKε → R with the vector of its values u(vi ). Also, for
neighboring vertices vi ∼ v j and u ∈ W ε we have u(v j ) − u(vi ) = O(ε). Therefore,
u ∈ W ε satisfies the triangle inequalities (7) if ε is small enough.
Our aim is to show that for ε small enough there exists a function u ε satisfying
conditions (i) and (ii) of Proposition 2.1 and u ε (v) = log | f  (v)| for all boundary ver-
tices v ∈ ∂ VKε . This function then defines a discrete conformal PL-map f ε (uniquely
if we use the normalization of Theorem 1.2).

Theorem 4.1 Assume that all angles of the triangular lattice T are strictly smaller
than π/2. There is an ε0 > 0 (depending on f , K and the triangulation parameters)
such that for all 0 < ε < ε0 the minimum of the functional E (see Theorem 2.2) with
boundary conditions (2) is attained in W ε .

Corollary 4.2 For all 0 < ε < ε0 there exists a discrete conformal PL-map on TKε
whose associated scale factors satisfy the boundary conditions (2).

The proof of Theorem 4.1 follows from Lemma 4.4 below. It is based on Theo-
rem 2.2 and on monotonicity estimates of the angle function θ (x, y) defined in (6).
It is only here where we need the assumption that all angles of the triangular lattice
T are strictly smaller than π/2.

Lemma 4.3 (Monotonicity lemma) Consider the star of a vertex v0 of a triangular

lattice T and its neighboring vertices v1 , . . . , v6 , v7 = v1 in cyclic order. Denote
λ0,k := 2 log(|vk+1 − vk |/|v0 − vk |). Assume that all triangles Δ(v0 , vk , vk+1 ) are
strictly acute angled, i.e. all angles < π/2.
144 U. Bücking

Then there exists η0 > 0, depending on the λs, such that for all 0  η1 , . . . , η6 ,
η7 = η1 < η0 there holds


6 
6
θ (λ0,k + ηk , λ0,k+1 + ηk+1 )  θ (λ0,k , λ0,k+1 ),
k=1 k=1

and for all 0  η1 , . . . , η6 , η7 = η1 > −η0 we have


6 
6
θ (λ0,k + ηk , λ0,k+1 + ηk+1 )  θ (λ0,k , λ0,k+1 ).
k=1 k=1

Proof First, consider a single acute angled triangle. Observe that with the notation
of Fig. 1b:
∂β 1
= − cot γ .
∂a a
Thus, we easily deduce that

∂ a a  1
θ (2 log( ) + ε, 2 log( )) = cot γ .
∂ε c b ε=0 2

Now the claim follows by Taylor expansion. 

Lemma 4.4 There is an ε0 > 0 such that for all 0 < ε < ε0 the negative gradient
−grad(E) on the boundary of W ε points into the interior of W ε .
Proof For notational simplicity, set u k = u(vk ), wk± = w ± (vk ) for vertices vk ∈ VKε
and λa,b,c = 2 log(|vb − vc |/|va − vb |).
Consider grad(E) on a boundary face Wi+ = {u ∈ W ε : u i = wi+ } of the n-
dimensional interval W ε . Let v1 , . . . , v6 , v7 = v1 denote the neighbors of vi in cyclic
order. Note that w +j − w −j = ε2 (M + + M − − (C + + C − )|v j |2 ) for all vertices v j .
As K is compact we may assume that 0 < ε0 is such that w +j − w −j  ε for 0 < ε < ε0 .
Then using the properties of w+ and u we obtain from Lemmas 4.3 and 3.1

∂E  5
(u) = 2π − θ (λi, j+1, j + u j − u i , λi, j, j+1 + u j+1 − u i )
∂u i  
j=0
=wi+ =wi+
     
w +j −wi+ w +j+1 −wi+


5
 2π − θ (λi, j+1, j + w +j − wi+ , λi, j, j+1 + w +j+1 − wi+ )
j=0

> 0.

An analogous estimate holds for boundary faces Wi− . 

Approximation of Conformal Mappings … 145

We are now ready to deduce our convergence theorem.

Proof (of Theorem 1.2) The existence part follows from Theorem 4.1. The unique-
ness is obvious as the translational and rotational freedom of the image of f ε is fixed
using values of f .
We now deduce the remaining estimates.
Part (i): Together with the definition of w± , Theorem 4.1 implies that for ε > 0 small
enough and all vertices v ∈ VKε

− ε2 (M − − C − |v|2 )
 w − (v) − log | f  (v)|  u ε (v) − log | f  (v)|  w + (v) − log | f  (v)|
 ε2 (M + − C + |v|2 ).

As K is compact, this implies estimate (3).

Part (ii): Given the scale factors u ε associated to the discrete conformal PL-map
f ε on TKε , we can in every image triangle determine the interior angles (using for
example (5)). In particular, we begin by deducing from estimate (3) the change of
these interior angles of the triangles.
Recall that for acute angled triangles the center of the circumcircle lies in the
interior of the triangle. Joining these centers for incident triangles leads to an embed-
ded regular graph εT ∗ = (εV ∗ , εE ∗ ) which is dual to the given triangular lattice
εT . In particular, the vertices εV ∗ are identified with the centers of the circumcir-
cles of the triangles of εT . Furthermore, each edge e∗ ∈ (εE ∗ ) intersects exactly
one edge e ∈ (εE) orthogonally, so e and e∗ are dual, see Fig. 3. Consider an
edge e = [v1 , v2 ] ∈ E Kε with dual edge e∗ = [c1 , c2 ]. Their lengths are related by
|c2 − c1 | = |v2 − v1 | cot ϕe , where ϕe denotes the angle opposite to e in εT . Further-
more we obtain

cot ϕe (log | f  (v2 )| − log | f  (v1 )|) = cot ϕe Re((log f  ) (v1 )(v2 − v1 )) + O(ε2 )
= cot ϕe Im((log f  ) (v1 )i(v2 − v1 )) + O(ε2 )
= Im((log f  ) (v1 )(c2 − c1 )) + O(ε2 )
= 2Im((log f  ) (v1 )(c2 − v1 ))
c2 + c1
+ 2Im((log f  ) (v1 )(v1 − )) + O(ε2 )
2
c2 + c1
= 2 arg f  (c2 ) − 2 arg f  ( ) + O(ε2 )
2
  
v2 +v1
= 2

(13)
v + v
= 2 arg f  ( ) − 2 arg f  (c1 ) + O(ε2 ),
2 1
2
(14)
146 U. Bücking

where we have chosen the notation such that (v2 − v1 )i = (c2 − c1 ) tan ϕe .
Now we estimate the change of the angles in a triangle of TKε compared with
its image triangle under f ε . Assume given a triangle Δ[v0 , v1 , v2 ] and denote e1 =
[v0 , v1 ] and e2 = [v0 , v2 ]. Denote the angle at v0 by θ0 = θ (λ1 , λ2 ), where le j = |v j −
v0 | and λ j = 2 log(|v1 − v2 |/le j+1 ) for j = 1, 2 and e3 = e1 . Consider the Taylor
expansion

cot ϕe1 cot ϕe2

θ (λ1 + x1 ε, λ2 + x2 ε) = θ0 + ε( x1 + x2 ) + O(ε2 ).
2 2
We apply this estimate for the bounded terms

u ε (v j ) − u ε (v0 ) log | f  (v j )| − log | f  (v0 )|

xj = = + O(ε)
ε ε
for j = 1, 2. Denote by δ + θ0 ∈ (0, π ) the angle at the image point of v0 in the
image triangle f ε (Δ[v0 , v1 , v2 ]). Then by (13) and (14) the change of angle δ is
given by
v2 + v0 v0 + v1
δ = arg f  ( ) − arg f  ( ) + O(ε2 ) (15)
2 2
This local change of angles is related to the angle ψ ε (e) by which each edge e of TKε
has to be rotated to obtain the corresponding image edge f ε (e) (or, more precisely, a
parallel edge). The function ψ ε may be defined globally on E Kε such that in the above
notation the change of the angle at v0 is given as δ = ψ ε (e2 ) − ψ ε (e1 ) ∈ (−π, π ).
We fix the value of ψ ε , that is the rotational freedom of the image of TKε under f ε at the
edge e0 according to arg f  , see Theorem 1.2. Then we take shortest simple paths and
deduce from (15) that each edge e = [v j , v j+1 ] ∈ E Kε is rotated counterclockwise by

v j + v j+1
ψ ε (e) = arg f  ( ) + O(ε).
2
This implies together with (3) that for all edges e = [v j , v j+1 ] ∈ E Kε we have uni-
formly
v j + v j+1 u ε (v j ) + u ε (v j+1 )
log f  ( )− − iψ ε (e) = O(ε). (16)
2 2
Therefore the difference of the smooth and discrete conformal maps at vertices
v0 ∈ VKε satisfies uniformly

f (v0 ) − f ε (v0 ) = O(ε)

by suitable integration along shortest simple paths from the reference point as above.
This estimate then also holds for all points in the support of TKε and ε → 0.
Part (iii): As last step we consider the derivatives of f ε restricted to a triangle.
Approximation of Conformal Mappings … 147

Assume given a triangle Δ[v0 , v1 , v2 ] in TKε . As f ε is piecewise linear its restric-

tion to Δ = Δ[v0 , v1 , v2 ] is the restriction of an R-linear map L Δ . This map can be
written for z ∈ C as

L Δ (z) = f ε (v0 ) + a · (z − v0 ) + b · (z − v0 ),

where the constants a, b ∈ C are determined from the conditions L Δ (v j ) = f ε (v j )

for j = 0, 1, 2. Straightforward calculation gives

( f ε (v2 ) − f ε (v0 ))(v1 − v0 ) − ( f ε (v1 ) − f ε (v0 ))(v2 − v0 )

∂z L Δ = a =
(v1 − v0 )(v2 − v0 ) − (v1 − v0 )(v2 − v0 )
( f ε (v2 ) − f ε (v0 ))(v1 − v0 ) − ( f ε (v1 ) − f ε (v0 ))(v2 − v0 )
∂z̄ L Δ = b = .
(v1 − v0 )(v2 − v0 ) − (v1 − v0 )(v2 − v0 )

Note that by definition of f ε and ψ ε we know that

ε
(v j )+u ε (v0 ))/2+iψ ε ([v j ,v0 ])
f ε (v j ) − f ε (v0 ) = (v j − v0 )e(u ,

where we use the rotation function ψ ε on the edges as defined in the previous part (ii)
of the proof. Now (16) together with the above expressions of a and b immediately
implies the desired estimates

∂z f ε |Δ (z) = ∂z L Δ (z) = f  (z) + O(ε) and ∂z̄ f ε |Δ (z) = ∂z̄ L Δ (z) = O(ε).

uniformly on the triangle Δ = Δ[v0 , v1 , v2 ]. Also, the constants in the estimate do

not depend on the choice of the triangle. This finishes the proof. 

Remark 4.5 Theorem 1.2 focuses on a particular way to approximate a given con-
formal map f by a sequence of discrete conformal PL-maps. Namely, we consider
corresponding smooth and discrete Dirichlet boundary value problems and compare
the solutions. There is of course a corresponding problem for Neumann boundary
conditions, i.e. prescribing angle sums of the triangles at boundary vertices using
arg f  . Also, there is a corresponding variational description for conformally equiv-
alent triangle meshes or discrete conformal PL-maps in terms of angles, see [1]. But
unfortunately, the presented methods for a convergence proof seem not to generalize
in a straightforward manner to this case, as the order of the corresponding Taylor
expansion is lower.

Acknowledgments The author would like to thank the anonymous referees for the careful reading
of the initial manuscript and various suggestions for improvement. This research was supported by
the DFG Collaborative Research Center TRR 109 “Discretization in Geometry and Dynamics”.

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
148 U. Bücking

The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Bobenko, A.I., Pinkall, U., Springborn, B.: Discrete conformal maps and ideal hyperbolic
polyhedra. Geom. Topol. 19, 2155–2215 (2015)
2. Bobenko, A.I., Skopenkov, M.: Discrete Riemann surfaces: linear discretization and its con-
vergence. To appear in J. Reine Angew, Math (2014)
3. Born, S., Bücking, U., Springborn, B.: Quasiconformal distortion of projective transformations,
with an application to discrete conformal maps. arXiv:1505.01341 [math.CV]
4. Bücking, U.: Approximation of conformal mappings by circle patterns and discrete mini-
mal surfaces. Ph.D. thesis, Technische Universität Berlin (2007). http://opus.kobv.de/tuberlin/
volltexte/2008/1764/
5. Bücking, U.: Approximation of conformal mapping by circle patterns. Geom. Dedicata 137,
163–197 (2008)
6. Chelkak, D., Smirnov, S.: Universality in the 2D Ising model and conformal invariance of
fermionic observables. Invent. math. 189, 515–580 (2012)
7. Courant, R., Friedrichs, K., Lewy, H.: Über die partiellen Differenzengleichungen der math-
ematischen Physik. Math. Ann. 100, 32–74 (1928). English translation: IBM Journal (1967),
215–234
8. Gu, X., Guo, R., Luo, F., Sun, J., Wu, T.: A discrete uniformization theorem for polyhedral
surfaces II. arXiv:1401.4594 [math.GT]
9. Gu, X., Luo, F., Sun, J., Wu, T.: A discrete uniformization theorem for polyhedral surfaces.
arXiv:1309.4175 [math.GT]
10. He, Z.X., Schramm, O.: On the convergence of circle packings to the Riemann map. Invent.
Math. 125, 285–305 (1996)
11. He, Z.X., Schramm, O.: The C ∞ -convergence of hexagonal disk packings to the Riemann map.
Acta Math. 180, 219–245 (1998)
12. Lan, S.Y., Dai, D.Q.: The C ∞ -convergence of SG circle patterns to the Riemann mapping. J.
Math. Anal. Appl. 332, 1351–1364 (2007)
13. Lelong-Ferrand, J.: Représentation conforme et transformations à intégrale de Dirichlet bornée.
Gauthier-Villars, Paris (1955)
14. Luo, F.: Combinatorial Yamabe flow on surfaces. Commun. Contemp. Math. 6(5), 765–780
(2004)
15. Matthes, D.: Convergence in discrete Cauchy problems and applications to circle patterns.
Conform. Geom. Dyn. 9, 1–23 (2005)
16. Mercat, C.: Discrete Riemann Surfaces. In: Papadopoulos, A. (ed.) Handbook of Teichmüller
theory, vol. I, pp. 541–575. Eur. Math. Soc., Zürich (Ed.) (2007)
17. Rodin, B., Sullivan, D.: The convergence of circle packings to the Riemann mapping. J. Diff.
Geom. 26, 349–360 (1987)
18. Schramm, O.: Circle patterns with the combinatorics of the square grid. Duke Math. J. 86,
347–389 (1997)
19. Skopenkov, M.: The boundary value problem for discrete analytic functions. Adv. Math. 240,
61–87 (2013)
Approximation of Conformal Mappings … 149

20. Springborn, B., Schröder, P., Pinkall, U.: Conformal equivalence of triangle meshes. ACM
Trans. Graph. 27(3) (2008)
21. Thurston, B.: The finite Riemann mapping theorem. Invited address at the International Sym-
posioum in Celebration of the proof of the Bieberbach Conjecture, Purdue University (1985)
22. Werness, B.M.: Discrete analytic functions on non-uniform lattices without global geometric
control (2014). Preprint
Numerical Methods for the Discrete Map Z a

Folkmar Bornemann, Alexander Its, Sheehan Olver

and Georg Wechslberger

Abstract As a basic example in nonlinear theories of discrete complex analysis,

we explore various numerical methods for the accurate evaluation of the discrete
map Z a introduced by Agafonov and Bobenko. The methods are based either on a
discrete Painlevé equation or on the Riemann–Hilbert method. In the latter case, the
underlying structure of a triangular Riemann–Hilbert problem with a non-triangular
solution requires special care in the numerical approach. Complexity and numerical
stability are discussed, the results are illustrated by numerical examples.

1 Introduction

Following the famous ideas of Thurston’s for a nonlinear theory of discrete com-
plex analysis based on circle packings, Bobenko and Pinkall [5] defined a discrete
conformal map as a complex valued function f : Z2 ⊂ C → C satisfying

( f n,m − f n+1,m )( f n+1,m+1 − f n,m+1 )

= −1. (1)
( f n+1,m − f n+1,m+1 )( f n,m+1 − f n,m )

F. Bornemann (B) · G. Wechslberger

Zentrum Mathematik – M3, Technische Universität München,
Boltzmannstr. 3, 85747 Garching bei München, Germany
e-mail: [email protected]
G. Wechslberger
e-mail: [email protected]
A. Its
Department of Mathematical Sciences, Indiana University–Purdue University,
402 N. Blackford, Indianapolis, IN 46202, USA
e-mail: [email protected]
S. Olver
School of Mathematics and Statistics, The University of Sydney,
Sydney, NSW 2006, Australia
e-mail: [email protected]

© The Author(s) 2016 151

A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_4
152 F. Bornemann et al.

That is, the cross ratio on each elementary quadrilateral (fundamental cell) of the lat-
tice Z2 is −1; infinitesimally, this property characterizes conformal maps among the
smooth ones. A discrete conformal map f n,m is called an immersion if the interiors
of adjacent elementary quadrilaterals are disjoint.
A central problem in discrete complex analysis is to find discrete conformal ana-
logues of classical holomorphic functions that are immersions; simply evolving just
the boundary values of the classical function by (1) would not work [6]. To solve this
problem, Bobenko [3] suggested to augment (1) by another equation: using meth-
ods from the theory of integrable systems it can be shown that the non-autonomous
system of constraints

( f n+1,m − f n,m )( f n,m − f n−1,m ) ( f n,m+1 − f n,m )( f n,m − f n,m−1 )

a f n,m = 2n + 2m ,
( f n+1,m − f n−1,m ) ( f n,m+1 − f n,m−1 )
(2)
obtained as an integrable discretization of the differential equation

a f = x fx + y f y = z fz

that would define f (z) = z a up to scaling, is compatible with (1). Agafonov and
Bobenko [1] proved that, for 0 < a < 2, the system (1) and (2) of recursions,
applied to the three initial values

f 0,0 = 0, f 1,0 = 1, f 0,1 = eiaπ/2 , (3)

defines a unique discrete conformal map Z n,m a

= f n,m that is an immersion
[1, Theorem 1].
Moreover, they showed [1, Sect. 3] that this discrete conformal map Z a deter-
mines a circle pattern of Schramm type, i.e., an orthogonal circle pattern with the

Fig. 1 Left Red dots are the discrete Z 2/3 for 0  n, m  19; blue circles are the asymptotics
given by (4). Right The Schramm circle pattern of the discrete Z 2/3 [courtesy of J. Richter-Gebert]
Numerical Methods for the Discrete Map Z a 153

2/3
Fig. 2 Numerical discrete Z n,m (0  n, m  49): recursing from the initial values (3) by a
straightforward application of the system (1) and (2) quickly develops numerical instabilities. The
color cycles with the coordinate m

combinatorics of the square grid, see Fig. 1. They conjectured, recently proved by
Bobenko and Its [4] using the Riemann–Hilbert method, that asymptotically
 a   
n + im 1
a
Z n,m = ca 1+O (n 2 + m 2 → ∞) (4a)
2 n2 + m 2
 
with the constant Γ 1 − a2
ca =  . (4b)
Γ 1 + a2

For 0 < a  1, as exemplified in Fig. 1, this asymptotics is already accurate to

plotting accuracy for all but the very smallest values of n and m. If a → 2, however,
it requires increasingly larger values of n and m to become accurate.
In this work we study the stable and accurate numerical calculation of Z a ; to
the best of our knowledge for the first time in the literature. This is an interesting
mathematical problem in itself, but the underlying methods should be applicable to
a large set of similar discrete integrable systems. Now, the basic difficulty is that
the evolution of the discrete dynamical system (1) and (2), starting from the initial
values (3), is numerically highly unstable, see Fig. 2.1

1 All numerical calculations are done in hardware arithmetic using double precision.
154 F. Bornemann et al.

The support of the stencils of (1) and (2) has the form of a square and a five-point
cross in the lattice Z2 , that is,

f n,m+1
f n,m+1 f n+1,m+1
and f n−1,m f n,m f n+1,m ,
f n,m f n+1,m
f n,m−1

with the latter reducing to be dimensional along the boundary of Z2+ , namely

f 0,m+1
f 0,m , resp. f n−1,0 f n,0 f n+1,0 ,
f 0,m−1

if n = 0 or m = 0. Thus, the forward evolution can be organized as follows. If f n,m

is known for 0  n, m  N the upper index N is increased to N + 1 according to:
(i) use (2) to compute the boundary values f N +1,0 and f 0,N +1 ;
(ii) use (1) to compute the row f N +1,m , 1  m  N ;
(iii) use (1) to compute the column f n,N +1 , 1  n  N ;
(iv) use (1) to compute the diagonal value f N +1,N +1 .
It is this algorithm that gives the unstable calculation shown in Fig. 2. Alterna-
tively, one could use (2) to calculate the row values f N +1,m , 1  m  N − 1, and
column values f n,N +1 , 1  n  N − 1, up to the first sub- and superdiagonal (note
that these calculations do not depend on order within the rows and columns). The
missing values are then completed by using (1). However, this alternative forward
evolution gives a result that is visually indistinguishable from Fig. 2.
As can be seen from Fig. 2, the numerical instability starts spreading from the
diagonal elements f n,n . In fact, there is an initial exponential growth of numerical
errors to be found in the diagonal entries, see Fig. 3. Such a numerical instability of

Fig. 3 Numerical error of

the diagonal values f n,n 10 0
from Fig. 2 (blue), of the xn
as in (5) and computed by
forward evolution of the 10 −4
discrete Painlevé II equation
(red), and of the
error

corresponding invariant
10 −8
|xn | = 1 (yellow); a = 2/3.
They share the same rate of
initial exponential growth
10 −12

10 −16
0 10 20 30 40 50
n
Numerical Methods for the Discrete Map Z a 155

an evolution is the direct consequence of the instability of the underlying dynamical

system, that is, of positive Lyapunov exponents.
a
As a remedy we suggest two different approaches to calculating Z n,m . In Sect. 2
we stabilize the calculation of the diagonal values by solving a boundary value prob-
lem for an underlying discrete Painlevé II equation and in Sects. 3–8 we explore
numerical methods based on the Riemann–Hilbert method. The latter reveals an
interesting structure (Sect. 4): the Riemann–Hilbert problem has triangular data but
a non-triangular solution; the operator equation can thus be written as a uniquely
solvable block triangular system where the infinite-dimensional diagonal operators
are not invertible. We discuss two different ways to prevent this particular struc-
ture from hurting finite-dimensional numerical schemes: a coefficient-based spectral
method with infinite-dimensional linear algebra in Sect. 6 and a modified Nyström
method based on least squares in Sect. 8.

2 Discrete Painlevé II Separatrix as a Boundary

Value Problem

Since the source of the numerical instability of the direct evolution of the discrete
dynamical system (1) and (2) is found in the diagonal elements f n,n , we first express
the f n,n directly in terms of a one-dimensional three-term recursion and then study
its stable numerical evaluation. To begin with, Agafonov and Bobenko [1, Proposi-
tion 3] proved that the geometric quantities

f n,n+1 − f n,n
xn2 = , arg xn ∈ (0, π/2), (5)
f n+1,n − f n,n

have invariant magnitude |xn | = 1 (see the circle packing in Fig. 1) and that they
satisfy the following form of the discrete Painlevé II equation
   
xn+1 − i xn xn−1 + i xn
(n + 1)(xn2 − 1) − n(xn2 + 1) = axn , (6)
i + xn xn+1 i + xn−1 xn

with initial value x0 = eiaπ/4 . Note that for n = 0 this nonlinear three-term recur-
rence degenerates and gives the missing second initial value, namely

x0 (x02 + a − 1)
x1 = . (7)
i((a − 1)x02 + 1)

Reversely, given the solution xn of this equation, the diagonal elements f n,n can be
calculated according to the simple recursion [1, p. 176]
156 F. Bornemann et al.

rn
un = , rn+1 = u n · Imxn , gn+1 = gn + u n , f n+1,n+1 = gn+1 eiaπ/4 ,
Rexn
(8)
with inital values g0 = 0, r0 = 1 (note that u n , rn , gn are all positive); the sub- and
superdiagonal elements f n+1,n and f n,n+1 are obtained from (1) and (5) by

(xn2 − 1) f n,n + (xn2 + 1) f n+1,n+1

f n+1,n = , (9a)
2xn2
(1 − xn2 ) f n,n + (1 + xn2 ) f n+1,n+1
f n,n+1 = . (9b)
2
However, given that xn is a separatrix solution of the discrete Painlevé II equation
[1, p. 167], we expect that a forward evolution of (6), starting with the initial values
x0 and x1 , suffers from exactly the same instability as the calculation of the diagonal
values f n,n by evolving (1) and (2). Figure 3 shows that this is indeed the case,
exhibiting the same initial exponential growth rate; it also shows that the deviation
of the calculated values of |xn | from its invariant value 1 can serve as an explicitly
computable error indicator.
In the continuous case of the Hastings–McLeod solution of Painlevé II, which
also constitutes a separatrix, Bornemann [8, Sect. 3.2] suggested to address such
problems by solving an asymptotic two-point boundary value problem instead of
the originally given evolution problem. To this end, one has to solve the connection
problem first, that is, one has to establish the asymptotics of xn as n → ∞. By
a
inserting the known asymptotics (4) of Z n,m into the defining Eq. (5), we obtain

xn = eiπ/4 (1 + O(n −1 )) (n → ∞).

Since in actual numerical calculations we need accurate approximations already for

moderately large n, we match the coefficients of an expansion in terms of n −1 to the
discrete Painlevé II equation (6) and get, as n → ∞,
  
iπ i(a − 1) −a 2 + (2 − 2i)a − (1 − 2i) i a 3 − (3 − 2i)a 2 − (1 + 4i)a + (3 + 2i)
xn = e 4 1+ + −
2n 8n 2 16n 3
3a 4 − (12 − 12i)a 3 − (2 + 36i)a 2 + (28 + 4i)a − (17 − 20i)
+
128n 4
 
i 3a 5 − (15 − 12i)a 4 − (30 + 48i)a 3 + (150 + 24i)a 2 − (5 − 48i)a − (103 + 36i)
+ 5
256n

+ O(n −6 ) .

We denote the r.h.s. of this asymptotic formula, without the O(n −6 ) term, by xn,6 .
Next, using Newton’s method, we solve the nonlinear system of N + 1 equations
in N + 1 unknowns x0 , . . . , x N given by the discrete Painlevé equation (6) for 1 
n  N − 1 and the two boundary conditions
Numerical Methods for the Discrete Map Z a 157

x0 (x02 + a − 1)
x1 = , x N = x N ,6 .
i((a − 1)x02 + 1)

Note that the value x0 = eiaπ/4 is not explicitly used and must be obtained as output
of the Newton solve, that is, it can be used as an measure of success. We choose N
.
large enough that |x N ,6 | = 1 up to machine precision (about N ≈ 300 uniformly in
a). Then, using the excellent initial guesses (for the accuracy of the asymptotics cf.
the left panel of Fig. 1)

x0(0) = eiaπ/4 , xn(0) = xn,6 (1  n  N ),

Newton’s method will converge in about just 10 iterations to machine precision

yielding a numerical solution that satisfies the invariant |xn | = 1 also up to machine
precision. Since the Jacobian of a nonlinear system stemming from a three-term
recurrence is tridiagonal, each Newton step has an operation count of order O(N ).
Hence, the overall complexity of accurately calculating the values xn , 0  n  N ,
is of optimal order O(N ).
Finally, having accurate values of xn at hand, and therefore by (8) and (9) also
those of f n,n , f n+1,n and f n,n+1 , one can calculate the missing values of f n,m row-
and column-wise, starting from the second sup- and superdiagonal and evolving
to the boundary, either by evolving the cross-ratio relations (1) or by evolving the
discrete differential equation (2). It turns out that the first option develops numerical
instabilities spreading from the boundary, see Fig. 4, whereas the second option is,

2/3
Fig. 4 Numerical discrete Z n,m (0  n, m  49): evolving from accurate values of f n,n , f n+1,n ,
f n,n+1 close to the diagonal back to the boundary by using the cross-ratio relations (1) develops
numerical instabilities. The color cycles with the coordinate m
158 F. Bornemann et al.

2/3
Fig. 5 Numerical discrete Z n,m (0  n, m  49): recursing from accurate values of f n,n , f n+1,n ,
f n,n+1 close to the diagonal back to the boundary by using the discrete differential equation (2) is
perfectly stable. The color cycles with the coordinate m

for a wide range of the parameter a, numerically observed to be perfectly stable,

see Fig. 5. Note that this stable algorithm only differs from the alternative direct
evolution discussed in the introduction in how the values close to the diagonal, that
is f n,n , f n+1,n and f n,n+1 , are computed.
The total complexity of this stable numerical calculation of the array f n,m with
0  n, m  N is of optimal order O(N 2 ).

3 The Riemann–Hilbert Method

Based on the integrability of the system (1) and (2), by identifying (1) as the com-
patibility condition of a Lax pair of linear difference equations [5] and by using
isomonodromy, Bobenko and Its [4, p. 15] expressed the Z a map in terms of the fol-
lowing Riemann–Hilbert problem (which is a slightly transformed and transposed
version of the X -RHP by these authors): Let Γ1 be the oriented contour built of two
non-intersecting circles in the complex plane centered at z = ±1 (see Fig. 6 left),
the holomorphic function X : C \ Γ1 → GL(2) satisfies the jump condition

X + (ζ ) = G 1 (ζ )X − (ζ ) (ζ ∈ Γ1 ) (10a)
Numerical Methods for the Discrete Map Z a 159

Fig. 6 Contours for the X -RHP [4, p. 15]. Left Two non-intersecting circles Γ1 centered at ±1
(black); right additional circle Γ2 centered at 0 (red) for standard normalization at ∞

with the jump matrix2

 
1 0
G 1 (ζ ) = iaπ/2 −a/2 (10b)
e ζ (ζ − 1)−m (ζ + 1)−n 1

subject to the following normalization

 m+n 
z 2 0
X (z) = (I + O(z −1 )) (z → ∞). (10c)
0 z− 2
m+n

Here, we restrict ourselves to values of n and m having the same parity such that
(m + n)/2 is an integer. The discrete Z a map is now given by the values f n,m
extracted from an LU -decomposition at z = 0, namely,
  
1 0 • •
X (0) = ,
(−1)m+1 f n,m 1 0 •

that is,
X 21 (0)
f n,m = (−1)m+1 .
X 11 (0)

Subsequently, using the Deift–Zhou nonlinear steepest decent method, Bobenko and
Its [4] transform this X -RHP to a series of Riemann–Hilbert problems that are more
suitable for asymptotic analysis. The last one of this series before introducing a

2 Tomake G 1 holomorphic in the vicinity of Γ1 we place the branch-cut of ζ −a/2 at the negative
imaginary axis, that is, we take, using the principal branch Log of the logarithm,
a
eiaπ/2 ζ −a/2 = eiaπ/4 e− 2 Log(ζ /i) .

.
160 F. Bornemann et al.

Fig. 7 Left Contour of the S-RHP [4, pp.24–27] centered at z = 0. Right Modified contour after
normalizing the RHP at z = 0 and z → ∞ (analogously to Fig. 6); the relative size of the inner and
outer circles is chosen depending on n and m and has a major influence on the condition number
of the spectral collocation method. Proper choices steer the condition number into a regime, which
corresponds to a loss of about three to eight digits, see [24, Sect. 5.4]

global parametrix,3 the S-RHP [4, pp. 24–27], is based on the contour shown in the
left part of Fig. 7. This rather elaborate S-RHP is, after normalizing at z = 0 and
z → ∞ appropriately, amenable to the spectral collocation method of Olver [18];
we skip the details which can be found in the thesis of the fourth author G.W. [24,
Sect. 5.4] that extends previous work on automatic contour deformation by Borne-
mann and Wechslberger [10, 25]. Here, the relative size of the inner and outer cir-
cles shaping the contour system shown in the right part of Fig. 7 have to be carefully
adjusted to the parameters n and m to keep the condition number at a reasonable
size. The complexity of computing f n,m for fixed n and m is then basically indepen-
dent of m and n.
In the rest of this work we explore to what extent the analytic transformation
from the X -RHP to the S-RHP is a necessary preparatory step also numerically, or
whether one can use the originally given X -RHP as the basis for numerical calcula-
tions. To this end, we replace the normalization (10c) by the standard one, that is,

X (z) = I + O(z −1 ) (z → ∞), (10d)

and introduce a further circle Γ2 as shown in the right part of Fig. 6 with the jump
condition
 m+n 
ζ 2 0
X + (ζ ) = G 2 (ζ )X − (ζ ) (ζ ∈ Γ2 ), G 2 (ζ ) = . (10e)
0 ζ− 2
m+n

3 Though the parametrix leads to a near-identity RHP, the actually computation of the parametrix
would require solving a problem that is, numerically, of similar difficulty as the S-RHP itself.
Numerical Methods for the Discrete Map Z a 161

We define Γ = Γ1 ∪ Γ2 and put G(ζ ) = G j (ζ ) for ζ ∈ Γ j ( j = 1, 2). That way,

the Riemann-Hilbert problem is given in the standard form

X + (ζ ) = G(ζ )X − (ζ ) (ζ ∈ Γ ), X (z) = I + O(z −1 ) (z → ∞). (11)

Because of det G = 1, the solution X ∈ C ω (C \ Γ, GL(2)) is unique, see

[12, p. 104].

4 Lower Triangular Jump Matrices and Indices

We note that the jump matrix G defined in (10b) and (10e) is lower triangular. How-
ever, even though the non-singular lower triangular matrices form a multiplicative
group and the normalization at z → ∞ is also lower triangular, the solution X turns
out to not be lower triangular. Arguably the most natural source of RHPs exhibiting
this structure are connected to orthogonal polynomials. By renormalizing at z → ∞
the standard RHP for the system of orthogonal polynomials on the unit circle with
complex weight e z , we are led to consider the following model problem (m ∈ N)4 :
 m 
ζ 0
Y+ (ζ ) = Y (ζ ) (|ζ | = 1), Y (z) = I + O(z −1 ) (z → ∞).
eζ ζ −m −
(12)
Though one could perform a set of transformations to this problem that are stan-
dard in the RHP approach to the asymptotics of orthogonal polynomials on the
circle, basically resulting in an analogue of the S-RHP of [4], our point here is to
understand the issues of a direct numerical approach to the X -RHP (11) in a simple
model case. It is straightforward to check that the unique solution of (12) is given
explicitly by
⎧

⎪
⎪ 1 −z −m em (−z)
⎪
⎪ (|z| > 1),
⎪
⎪
⎨ 0 1
Y (z) =
⎪
⎪

⎪
⎪ zm −em (−z)
⎪
⎪ (|z| < 1),
⎩ z −m
e z (1 − e z em (−z))

4 The standard form, see [2, p. 1124], of that orthogonal polynomial RHP would be
   m 
1 0 z 0
X + (ζ ) = ζ −m X − (ζ ) (|ζ | = 1), X (z) = −m (I + O(z −1 )) (z → ∞).
e ζ 1 0 z

The model problem (12) is obtained by putting the diagonal scaling at z → ∞ into the jump matrix.
162 F. Bornemann et al.

with

z2 z k−1 Γ (k, z)
ek (z) = 1 + z + + ··· + = ez , (13)
2! (k − 1)! Γ (k)

where Γ (z) and Γ (k, z) denote the Gamma function and the incomplete Gamma
function. In particular, we observe that Y12 (0) = −1 = 0.
The nontrivial 12-component v of a Riemann–Hilbert problem with lower trian-
gular jump matrices, such as (11) or (12), can be expressed independently of the
other components, it satisfies a homogeneous scalar Riemann–Hilbert problem of
its own. Namely, denoting the 11-component of G by g, we get

v+ (ζ ) = g(ζ )v− (ζ ) (ζ ∈ Γ ), v(z) = O(z −1 ) (z → ∞). (14)

If the contour is a cycle as in (11), or as in the model problem above, the gen-
eral theory [17, Sect. 127] of Riemann–Hilbert problems with Hölder continuous
boundary regularity states that the Noether index5 κ of (14) is given by the winding
number

κ = indΓ g.

More precisely, the nullity is the sum of the positive partial indices and the defi-
ciency is the sum of the magnitudes of the negative partial indices, see
[17, Eq. (127.30)]. Since there is just one partial index in the scalar case, the nullity
of (14) is κ if κ > 0, and the deficiency is −κ if κ < 0.
Thus, in the case of the RHP (11), the nullity of the scalar sub-RHP for the 12-
component is
n+m
indΓ g = indΓ1 1 + indΓ2 ζ (n+m)/2 = ,
2
in the case of the model RHP (12) the corresponding nullity is m. In both cases, the
unique non-zero solution of (14) that is induced by the solution of the defining 2 × 2
RHP is precisely selected by the compatibility conditions set up by the remaining
linear relations of that RHP: the homogeneous part of these relations must then have
Noether index −κ.

Impact on Numerical Methods

This particular substructure of a Riemann–Hilbert problem with lower triangular

jump matrices G is a major challenge for numerical methods. If a discretization of
the 2 × 2 RHP induces a discretization of the scalar subproblem (14) that results in

5 Here, we identify a RHP with an equivalent linear operator equation T u = · · · , see, e.g., (15) in

the next section. We recall that λ = dim ker T is called the nullity, μ = dim coker T the deficiency
and κ = λ − μ the Noether index of a linear operator T with closed range.
Numerical Methods for the Discrete Map Z a 163

a homogeneous linear system with a square matrix S N (that is, the same number of
equations and unknowns), there are just two (non exclusive) options:
• S N is non-singular, which results in a 12-component v N = 0 that does not con-
verge;
• the full system is singular and therefore numerically of not much use (ill-
conditioning and convergence issues will abound).
Such methods compute fake lower triangular solutions, are ill-conditioned, or both.
To understand this claim, let us denote the 12-component of the 2 × 2 discrete
solution matrix by v N and the vector of the three other components by w N . By inher-
iting the subproblem structure such as (14) for the 12-component, the discretization
results then in a linear system of the block matrix form
    
SN 0 vN 0
=
• TN wN •
  
=A N

Because of det(A N ) = det(S N ) det(TN ) a non-singular discretization matrix A N

implies a non-singular S N and, thus, a non-convergent trivial component v N = 0.
Such a non-convergent zero 12-component is what one gets, for example, if one
applies the spectral collocation method of [18] (with square contours replacing the
circles) to the Riemann–Hilbert problems (11) or to the model problem (14). As a
hint of failure, the resulting discrete system is ill-conditioned; details which can be
found in the thesis of the fourth author G.W. [24, Sect. 5.4].
The deeper structural reason for this problem can be seen in the fact that the
Noether index of finite-dimensional square matrices is always zero, whereas the
index of the infinite-dimensional subproblem (14) is strictly positive.
We suggest two approaches to deal with this problem: first, an infinite-
dimensional discretization using sequence spaces, that is, without truncation, and
using infinite-dimensional numerical linear algebra, and second, using underdeter-
mined discretizations with rectangular linear systems that are complemented by a
set of explicit compatibility conditions.

5 RHPs as Integral Equations with Singular Kernels

In this section, we recall a way to express the RHP (11), with standard normalization
at infinity, as a particular system of singular integral equations, cf. [11, 16, 18]. We
introduce the Cauchy transform

1 f (ζ )
C f (z) = dζ (z ∈
/ Γ)
2πi Γ ζ −z
164 F. Bornemann et al.

and their directional limits C± when approaching the left or right of the oriented
contour Γ , defined by

1 f (ζ )
C± f (η) = lim dζ (η ∈ Γ ).
z→η± 2πi Γ ζ − z

Note that C± can be extended as bounded linear operators mapping L 2 (Γ ) (or

spaces of Hölder continuous functions) into itself, and C (suitably extended) maps
such functions into functions that are holomorphic on C \ Γ , see [12, p. 100]. By
using the decomposition C+ − C− = id, the ansatz (by letting C act component-
wise on the matrix-valued function u)

X (z) = I + Cu(z), u ∈ L 2 (Γ, C2×2 ), (15a)

establishes the equivalence of a RHP of the form (11) and the system of singular
integral equations

(id −(G − I )C− )u = G − I (15b)

As the following theorem shows, singular integral operators of the form

TG = id −(G − I )C− : L 2 (Γ, C2×2 ) → L 2 (Γ, C2×2 )

can be preconditioned by operators of exactly the same form.

Theorem 1 Let Γ be a smooth, bounded, and non-self intersecting6 contour system

and G : Γ → GL(2) a system of jump matrices which continues analytically to a
vicinity of Γ . Then, TG −1 is a Fredholm regulator of TG , that is, TG −1 TG = id +K
with a compact operator K : L 2 (Γ, C2×2 ) → L 2 (Γ, C2×2 ) that can be represented
as a regular integral operator.

Proof The Sokhotski–Plemelj formula [17, Eq. (17.2)] gives that 2C− = − id +H ,
where H denotes a variant of the Hilbert transform (normalized as in [17]),

1 f (η)
H f (ζ ) = dη (ζ ∈ Γ ),
πi Γ η−ζ

with the integral understood in the sense of principle values. This way, we have

6 Points of self intersection are allowed if certain cyclic conditions are satisfied [13]: at such a point

the product of the corresponding parts of the jump matrix should be the identity matrix. These
conditions guarantee smoothness in the sense of [26], where the analog of Theorem 1 is proved for
the general smooth Riemann–Hilbert data.
Numerical Methods for the Discrete Map Z a 165

1 1
TG = A1 id +B1 H, A1 = (I + G), B1 = (I − G),
2 2
1 1
TG −1 = A2 id +B2 H, A2 = (I + G ), B2 = (I − G −1 ).
−1
2 2
By a product formula of Muskhelishvili [17, Eq. (130.15)], which directly follows
from the Poincaré–Betrand formula [17, Eq. (23.8)], one has

TG −1 TG = A id +B H + K ,

where K represents a regular integral operator and the coefficient matrices A and B
are given by the expressions

A = A2 A1 + B2 B1 , B = A2 B1 + B2 A1 .

Here, we thus obtain A = I and B = 0, which finally proves the assertion. 

This theorem implies that the operator TG is Fredholm, that is, its nullity and
deficiency are finite. In fact, since in our examples det G ≡ 1, we have that the
Noether index of TG is zero. The possibility to use the Fredholm theory is extremely
important in studying RHPs: it allows one to use, when proving the solvability of
Riemann-Hilbert problems, the “vanishing lemma” [26], see also [12, Chap. 5]. For
the use of Fredholm regulators in iterative methods applied to solving singular inte-
gral equations, see [23].

6 A Well-Conditioned Spectral Method

for Closed Contours

We follow the ideas of Olver and Townsend [20] on spectral methods for differential
equations, recently extended by Olver and Slevinsky [19] to singular integral equa-
tions. First, the solution u and the data G − I of the singular integral equation (15b)
are expanded7 in the Laurent bases of the circles that built up the cycle Γ . Next, the
resulting linear system is solved using the framework of infinite-dimensional linear
algebra [14, 21], built out of the adaptive QR factorization introduced in [20].
To be specific, we describe the details for the model RHP (12), where the cycle
Γ is just the unit circle. Here, we have the expansions
∞
 ∞

u(ζ ) = Uk ζ k , G(ζ ) − I = Ak ζ k (ζ ∈ Γ ),
k=−∞ k=−∞

7 It
is actually implemented this way in SingularIntegralEquations.jl, a J ULIA soft-
ware package described in [19].
166 F. Bornemann et al.

both rapidly decaying with 2 × 2 coefficient matrices Uk and Ak . In the Laurent

basis, the operator C− acts diagonally in the simple form

0 k  0,
C− ζ =
k
−ζ k k < 0.

which gives
∞

C− u(ζ ) = − U−k ζ −k (ζ ∈ Γ ).
k=1

Note that −C− acts as a projection to the subspace spanned by the basis elements
with negative index. This way, the system (15b) of singular integral equations is
transformed to8
∞

Uk + [k − j < 0] A j Uk− j = Ak (k ∈ Z). (16)
j=−∞

Up to a given accuracy, we may assume that the data is given as a finite sum,


n1
G(ζ ) − I ≈ Ak ζ k ,
k=−n 1

likewise for G −1 (ζ ) − I with a truncation at n 2 . Thus, writing the discrete system

(16) in matrix-vector form, the corresponding double-infinite matrix has a band-
width of order O(n 1 ). Preconditioning this system, following Theorem 1, by the
multiplication with the double-infinite matrix belonging to G −1 instead of G, results
in a double-infinite matrix that has a bandwidth of order O(n 1 + n 2 ). Since the right
hand side of (16) is truncated at indices of magnitude O(n 1 ), application of the
adaptive QR factorization [20], after re-ordering the double-infinite coefficients as
U0 , U−1 , U1 , . . . in order to be singly infinite, will result in an algorithm that has a
complexity of order O((n 1 + n 2 )2 n 3 ), where n 3 is the number of coefficients needed
to resolve u, dictated by a specified tolerance.
Remark 1 The extension to systems Γ of closed contours built from several circles
is straightforward. The jump data and the solution, restricted to a circle centered at
a are expanded in the Laurent basis (z − a)k , k ∈ Z. When instead evaluated at a
circle centered at b, a change of basis is straightforwardly computed using

∞  
j
(z − a) =j
(b − a) j−k (z − b)k ( j ∈ Z),
k=0
k

8 We use the Iverson bracket of a condition: [P ] = 1 if the predicate P is true, [P ] = 0 otherwise.

 Numerical Methods for the Discrete Map Z a 167

valid for |z − b| < |b − a|. Because of a geometric decay, one can truncate those
series at k = O(1) as long as |z − b| < θ |b − a| with 0 < θ < 1 small enough. The
adaptive QR factorization can then be applied by interlacing the Laurent coefficients
on each circle to obtain a singly infinite unknown vector of coefficients.

Numerical Example 1: Model problem

Because of the entries ζ m and ζ −m in the jump matrix of the model problem (12),
we have that n 1 , n 2 , n 3 = O(m) in order to resolve the data and the solution; hence
the computational complexity of the method scales as O(m 3 ). Using the J ULIA
software package SingularIntegralEquations.jl9 (v0.0.1) the problem
is numerically solved by the following short code showing that the user has to do
little more than just providing the data and entering the singular integral equation
(15b) as a mathematical expression:

1 using ApproxFun, SingularIntegralEquations

2

3 m = 100
4 Γ = Circle(0.0,1.0)
5 G = Fun(z -> [z^m 0; exp(z) 1/z^m],Γ )
6 C = Cauchy(-1)
7 @time u = (I-(G-I)*C)\(G-I)
8 Y = z ->I+cauchy(u,z)
9 err = norm(Y(0)-[0 -1; 1 (-1)^m*exp(-lfact(m))],2)

The run time10 is 2.8 seconds, the error of Y (0) is 4.22 · 10−15 (spectral norm),
which corresponds to a loss of one digit in absolute error.

Numerical Example 2: Riemann–Hilbert Problem for the Discrete Z2/3

Now, we apply the method to the Riemann–Hilbert problem (11) encoding the dis-
crete Z a map. Here, because of the exponents −m, −n and ±(n + m)/2 in (10), we
have n 1 , n 2 , n 3 = O(n + m) in order to resolve the data and the solution, see Fig. 8;
hence the computational complexity scales as O((n + m)3 ). Note that this is far
from optimal, using the stabilized recursion of Sect. 2 to compute a table including
a
Z n,m would give a complexity of order O((n + m)2 ). Once more, however, the code
requires little more than typing the mathematical equations of the RHP.

1 using ApproxFun, SingularIntegralEquations

2

3 a = 2/3
4 n = 6; m = 8; # n+m must be even
5 pow = z -> exp(1im*a*pi/4)*exp(-a/2*log(z/1im))
6 Γ = Circle(-1.0,0.3) ∪ Circle(+1.0,0.3) ∪ Circle(0.0,3.0)

9 https://github.com/ApproxFun/SingularIntegralEquations.jl, cf. [19].

10 Using a MacBook Pro with a 3.0 GHz Intel Core i7-4578U processor and 16 GB of RAM.
 168 F. Bornemann et al.

2·10 4 1·10 6

1·10 4 5·10 5

u21 0 u 21 0

−1·10 4 −5·10 5

−2·10 4 −1·10 6
0 0.5 1.0 0 0.5 1.0
t t

Fig. 8 Left u 21 (ζ ) on the circle ζ = −1 + 0.3e2πit ; right u 21 (ζ ) on ζ = +1 + 0.3e2πit . The real

parts are shown in blue, the imaginary part in yellow. Note that there are m = 6 oscillations on the
left and n = 8 oscillations on the right; the maximum amplitude is about 1.1 · 104 on the left and
7.5 · 105 on the right

7 G = Fun(z -> in(z,Γ [3])?[z^((m+n)/2) 0; 0 1/z^((n+m)/2)]:[1 0;

pow(z)/(z-1)^m/(z+1)^n 1],Γ )
8 C = Cauchy(-1)
9 @time u = (I-(G-I)*C)\(G-I)
10 X = z -> I+cauchy(u,z);
11 X0 = X(0)
12 Za0 = (-1)^(m+1)*X0[2,1]/X0[1,1]
13 Za1 = 3.610326860525178 + 2.568086087959661im # exact solution
from the recursion as in Section 1.2 using bigfloats
14 err = abs(Za0 - Za1)

The run time is 2.7 s, the absolute error of Z 6,8 is 3.38 · 10−8 , which corresponds
2/3

to a loss of about 7 digits. This loss of accuracy can be explained by comparing the
magnitude of the 21-component of u as shown in Fig. 8, along the two black circles
of Fig. 6, with that of the corresponding component of the solution matrix at z = 0,
namely,
 
−3.38121 −12.2073 + 8.68324i
X (0) ≈ .
12.2073 + 8.68324i 66.0758

We observe that during the evaluation of the Cauchy transform (15a), which maps
u
→ X (0) by means of an integral, at least 5 digits must have been lost by
cancellation—a loss, which structurally cannot be avoided for oscillatory integrands
with large amplitudes. (Note that this is not an issue of frequency: just one oscilla-
tion with a large amplitude suffices to get such a severe cancellation.)
Since the amplitudes of u 21 grow exponentially with n and m, the algorithm
a
for computing Z n,m based on the numerical evaluation of (15) applied to the RHP
(11) is numerically unstable. Even though the initial step, the spectral method in
coefficient space applied to (15b) is perfectly stable, stability is destructed by the
bad conditioning of the post-processing step, that is, the evaluation of the integral in
Numerical Methods for the Discrete Map Z a 169

(15a). We refer to [7] for an analysis that algorithms with a badly conditioned post
processing of intermediate solutions are generally prone to numerical instability.

7 RHPs as Integral Equations with Nonsingular Kernels

By reversing the orientation of the two small circles in the RHP (11), and by simul-
taneously replacing the jump matrix G 1 by G̃ 1 = G −1 1 , the RHP is transformed to
an equivalent one with a contour system
that satisfies the following properties,
see Fig. 9: it is a union of non-self intersecting smooth curves, that bound a domain
+ to its left. By − we will denote the (generally not connected) region which is
the complement of + ∪ Γ . Note that the model problem (12) falls into that class
of contours without any further transformation.
We drop the tilde from the jump matrices and consider RHPs of the form

Φ+ (ζ ) = G(ζ )Φ− (ζ ) (ζ ∈
), Φ(z) = I + O(z −1 ) (z → ∞), (17)

on such contours systems

. It will either represent the aforementioned transforma-
tion of (11) or the model problem (12). In particular, G is lower triangular and can
be analytically continued to a vicinity of
.
The classical theory developed by Plemelj (see [17, Sect. 126]) for such prob-
lems teaches the following: the directed boundary values Φ− of the unique analytic
solution Φ : C \
→ GL(2) of the RHP (17) satisfy a Fredholm integral equation
[17, Eq. (126.5)] of the second kind on
, namely

1 G −1 (ζ )G(η) − I
Φ− (ζ ) − Φ− (η) dη = I (ζ ∈
), (18)
2πi
η−ζ

Fig. 9 A modified, but equivalent, contour system

for the for the RHP (11) obtained by revers-
ing the orientation and by simultaneously replacing the jump matrix G 1 by G −1
1 . Now, there is a
bounded domain + (marked in green) to the left of
, cf. the original system shown in Fig. 6
170 F. Bornemann et al.

understood here as an integral equation in L 2 (

, C2×2 ). The matrix kernel of this
equation, that is,

G −1 (ζ )G(η) − I G(η) − G(ζ )

K (ζ, η) = = G −1 (ζ ) ,
η−ζ η−ζ

is smooth on
×
, since it extends as an analytic function and since the singular-
ity at ζ = η is removable. Integral equations of the form (18) with a smooth kernel
are, in principle, amenable to fast quadrature based methods, see the next section.
We note that, given the boundary values Φ− (ζ ) for ζ ∈
, the solution of the
RHP (17) can be reconstructed by
⎧ 
⎪ 1 Φ− (ζ )
⎪
⎪ I − dζ z ∈ − ,
⎪
⎨ 2πi
ζ − z
Φ(z) =  (19)
⎪
⎪
⎪
⎪ 1 G(ζ )Φ− (ζ )
⎩ dζ z ∈ + .
2πi
ζ −z

In general, however, the Fredholm equation (18) is not equivalent to the RHP, see
[17, p. 387]: the Fredholm equation has but a kernel of the same dimension as the
kernel of the associated homogeneous RHP, defined as

Ψ+ (ζ ) = G −1 (ζ )Ψ− (ζ ) (ζ ∈
), Ψ (z) = O(z −1 ) (z → ∞). (20)

As we will show now, the kernel of the associated RHP is nontrivial in the examples
studied in this work.
First, we observe, by the lower triangular form of G, that the 11- and the 12-
components of Ψ both satisfy a scalar RHP of the form (14) with a jump function g
that has a winding number which is

n+m
ind
g = −
2
for the discrete map Z a , and which is ind
g = −m for the model problem (12).
Note that this winding number has the sign opposite to the results of Sect. 4 since
the underlying 2 × 2 RHP is based on G −1 instead of G. Hence, the nullity of the
scalar RHPs for the 11- and the 12-components of Ψ is zero and the deficiency is
(n + m)/2 (m in case of the model problem). As a consequence, the 11- and the
12-components of Ψ must both be identically zero.
Next, since we now know that Ψ has a zero first row, also the 21- and 22-
components of Ψ satisfy a scalar RHP of the form (14) each, but with a jump
function g that has the positive winding number

n+m
ind
g =
2
Numerical Methods for the Discrete Map Z a 171

for the discrete map Z a , and ind

g = m for the model problem (12), just as dis-
cussed in Sect. 4. Hence, the deficiency of the scalar RHPs for the 21- and the 22-
components of Ψ is zero and the nullity is (n + m)/2 (m in case of the model
problem). Since both components are linearly independent from of each other, we
have thus proven the following lemma.

Lemma 1 The nullity of the associated homogeneous RHP (20), and hence, that of
the Fredholm integral equation (18) is n + m in the case of the discrete map Z a and
2m in the case of the model problem (12).

Example 1 For the model RHP (12) the smooth kernel of the Fredholm integral
equation (18) can be constructed explicitly. Here we have

(η/ζ )m −1

G −1 (ζ )G(η) − I η−ζ
0
K (ζ, η) = = eη ζ m −eζ ηm (ζ /η)m −1 .
η−ζ η−ζ η−ζ

A column of a matrix belonging to the kernel of (18) satisfies the equation

    
u − (ζ ) 1 u (η)
= K (ζ, η) − dη (ζ ∈
), (21)
w− (ζ ) 2πi
w− (η)

where
is the positively oriented unit circle. We will construct solutions that extend
analytically as u − (z) and w− (z) for z = 0, such that
   
u − (z) u − (η)
= resη=0 K (z, η) (z = 0).
w− (z) w− (η)

By recalling the notation introduced in (13) we observe, for k = 0, . . . , m − 1, that

resη=0 K 11 (z, η)ηk−m = z k−m ,

resη=0 K 21 (z, η)ηk−m = −z k em−k (z),
resη=0 K 22 (z, η)ηk = −z k .

Using the coefficients a (m)

jk that induce a change of polynomial basis by


m−1
zk = ak(m)
j z em− j (z)
j
(k = 0, . . . , m − 1),
k=0

we define the polynomials


m−1
pk(m) (z) = ak(m)
j z
j
(k = 0, . . . , m − 1),
j=0
172 F. Bornemann et al.

each of which has degree at most m − 1. Then, the m linear independent vectors
   
u − (ζ ) −2ζ −m pk(m) (ζ )
= (k = 0, . . . , m − 1) (22)
w− (ζ ) ζk

are solutions of (21) each. Thus, since its dimension is 2m by Lemma 1, the kernel
of the integral equation (18) is spanned by the 2 × 2 matrices whose columns are
linear combinations of these vectors. 
The unique solution Φ− of the RHP (17) can be picked among the solutions
of (18) by imposing additional linear conditions, namely n + m independent such
conditions in the case of the discrete map Z a and 2m in the case of the model
problem. Specifically, for the model problem (12), we obtain such conditions as
follows. First, since Φ− (z) continues analytically to |z| > 1 and since Φ− (z) = I +
O(z −1 ) as z → ∞, we get by Cauchy’s formula for the Laurent coefficients at z =
∞ that

1 dζ
Φ− (ζ ) k = [k = 1] · I (k = 1, 2, . . .).
2πi
ζ

Second, by restricting this relation to the second row of the matrix Φ− for k =
1, . . . , m, we get the conditions

1   dζ  
0 1 · Φ− (ζ ) k = 0 [k = 1] (k = 1, . . . , m). (23)
2πi
ζ

In fact, these conditions force all the components of the columns (22) that would
span an offset from the kernel of (18) to be zero.
For the Z a -RHP, similar arguments prove that the kernel of (17) is spanned by
matrices whose second row extends to polynomials of degree smaller than (n +
m)/2 to the outside of the outer circle in Fig. 9. Thus, the same form of conditions
as in (23) can be applied for picking the proper solution Φ− (ζ ), except that one
would have to replace
by that outer circle and the upper index m by (n + m)/2.

8 A Modified Nyström Method

Fredholm integral equations of the second kind with smooth kernels defined on a
system of circular contours are best discretized by the classical Nyström method
[15, Sect. 12.2]. Here, one uses the composite trapezoidal rule as the underlying
quadrature formula, that is,
 N −1
1 r   
f (z) dz ≈ f z 0 + r e2πi j/N e2πi j/N .
2πi ∂ Br (z 0 ) N j=0
Numerical Methods for the Discrete Map Z a 173

For integrands that extend analytically to a vicinity of the contour, this quadrature
formula is spectrally accurate, see, e.g., [9, §2] or [22, §2].
Since the Fredholm integral equation (18) has a positive nullity, applying the
Nyström method to it will yield, for N large enough, a numerically singular linear
system. However, the theory of the last section suggests a simple modification of
the Nyström method: we use the conditions (23) (after approximating them by the
same quadrature formula as for the Nyström method) as additional equations and
solve the resulting overdetermined linear system by the least squares method.

Numerical Example 1: Model problem

We apply the modified Nyström method to the Fredholm integral equation repre-
senting the model problem (12). By the sampling condition, see, e.g., [9, §2], the
number N of quadrature points will scale as N = O(m), hence the computational
complexity scales as O(m 3 ). To check the accuracy we compare with
 
1 dζ 1 dζ
Y (0) = G(ζ )Φ− (ζ ) , I = Φ− (ζ ) ,
2πi Γ ζ 2πi Γ ζ

evaluated by the same quadrature formula as for the Nyström method. For the par-
ticular parameters m = 100 and N = 140 we get, within a run-time of 0.49 s for
a straightforward Matlab implementation, a maximum error of these two quanti-
ties, measured in 2-norm, of 1.33 · 10−14 . The condition number of the least squares
matrix grows just moderately with m: it is about 23 for m = 1 and about 650 for
m = 1000.

Numerical Example 2: Discrete Z 2/3

Now, we apply the modified Nyström method to the Fredholm integral equation
representing the RHP (11) subject to a transformation to the form (17). Here, the
sampling condition requires N = O(n + m), hence the computational complexity
2/3
scales as O((n + m)3 ). For Z 6,8 , the modified Nyström method yields the conver-
gence plot shown in Fig. 10: it exhibits exponential (i.e., spectral) convergence until
a noise level of about 10−9 is reached, which corresponds to a loss of about 6 digits.
The reason for this loss is that this method for approximating the discrete Z a suffers
the same issue with a bad conditioning of the post-processing step, that is, of

1 dζ
Φ− (·)
→ X (0) = G(ζ )Φ− (ζ ) ,
2πi Γ ζ

as the spectral method for the singular integral equation discussed in Sect. 6. Here,
the amplitude of the real and imaginary part of Φ− (ζ ) along the two inner circles is
of the order 104 which causes a cancellation of at least 4 significant digits.
174 F. Bornemann et al.

10 0

10−3

error
10−6

10−9

10−12
10 20 30 40 50
N0
2/3
Fig. 10 Absolute error of the approximation of Z 6,8 by the modified Nyström method vs. the
number of quadrature points N0 on each of the three circles in Fig. 9 (the radii are 1/2 for the
inner circles, 3 for the outer one); the total number of quadrature points is then N = 3 × N0 . One
observes, after a threshold caused by a sampling condition, exponential (i.e., spectral) convergence
that saturates at a level of numerical noise at an error of about 10−9 . Run time of a Mathematica
implementation with N0 = 42 is about 0.15 s

9 Conclusion

To summarize, there are two fundamental options for the stable numerical evaluation
a
of the discrete map Z n,m .
• Computing all the values of the array 1  n, m  N at once by, first, computing
the diagonal using a boundary value solve for the discrete Painlevé II equation (5)
and, then, by recursing from the diagonal to the boundary using the discrete dif-
ferential equation (2). This approach has optimal complexity O(N 2 ).
• Computing just a single value for a given index pair (n, m) by using the RHP (11)
and one of the methods discussed in Sect. 6 or 8. Since both methods suffer from
an instability caused by a post-processing quadrature for larger values of n and
m, one would rather mix this approach with the asymptotics (4). For instance,
using the numerical schemes for n, m  10, and the asymptotics otherwise, gives
a uniform precision of about 5 digits for a = 2/3. Higher accuracy would require
the calculation of the next order terms of the asymptotics as in Sect. 2. This mixed
numerical-asymptotic method has optimal complexity O(1).

Acknowledgments The research of F.B., A.I., and G.W. was supported by the DFG-Collaborative
Research Center, TRR 109, “Discretization in Geometry and Dynamics.”

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
Numerical Methods for the Discrete Map Z a 175

The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Agafonov, S.I., Bobenko, A.: Discrete Z γ and Painlevé equations. Int. Math. Res. Notices
2000(4), 165–193 (2000)
2. Baik, J., Deift, P., Johansson, K.: On the distribution of the length of the longest increasing
subsequence of random permutations. J. Am. Math. Soc. 12(4), 1119–1178 (1999)
3. Bobenko, A.: Discrete conformal maps and surfaces. Symmetries and Integrability of Differ-
ence Equations (Canterbury. 1996), London Mathematical Society Lecture Note Series, vol.
255, pp. 97–108. Cambridge University Press, Cambridge (1999)
4. Bobenko, A., Its, A.: The asymptotic behaviour of the discrete holomorphic map Z a via the
Riemann–Hilbert method (2014). arXiv:1409.2667
5. Bobenko, A., Pinkall, U.: Discrete isothermic surfaces. J. Reine Angew. Math. 475, 187–208
(1996)
6. Bobenko, A., Pinkall, U.: Discretization of surfaces and integrable systems. Discrete Inte-
grable Geometry and Physics (Vienna, 1996). Oxford Lecture Series in Mathematics and Its
Applications, vol. 16, pp. 3–58. Oxford University Press, New York (1999)
7. Bornemann, F.: A model for understanding numerical stability. IMA J. Numer. Anal. 27(2),
219–231 (2007)
8. Bornemann, F.: On the numerical evaluation of distributions in random matrix theory: a
review. Markov Process. Rel. Fields 16(4), 803–866 (2010)
9. Bornemann, F.: Accuracy and stability of computing high-order derivatives of analytic func-
tions by Cauchy integrals. Found. Comput. Math. 11(1), 1–63 (2011)
10. Bornemann, F., Wechslberger, G.: Optimal contours for high-order derivatives. IMA J. Numer.
Anal. 33(2), 403–412 (2013)
11. Deift, P., Zhou, X.: A priori L p -estimates for solutions of Riemann-Hilbert problems. Int.
Math. Res. Not. 40, 2121–2154 (2002)
12. Fokas, A.S., Its, A.R., Kapaev, A.A., Novokshenov, V.Y.: Painlevé Transcendents-The
Riemann-Hilbert Approach. American Mathematical Society, Providence (2006)
13. Fokas, A.S., Zhou, X.: On the solvability of Painlevé II and IV. Comm. Math. Phys. 144(3),
601–622 (1992)
14. Hansen, A.C.: Infinite-dimensional numerical linear algebra: theory and applications. Proc. R.
Soc. Lond. Ser. A Math. Phys. Eng. Sci. 466(2124), 3539–3559 (2010)
15. Kress, R.: Linear Integral Equations, vol. 82, 3rd edn. Springer, New York (2014)
16. Litvinchuk, G.S., Spitkovskii, I.M.: Factorization of Measurable Matrix Functions. Birkhäuser
Verlag, Basel (1987)
17. Muskhelishvili, N.I.: Singular Integral Equations. Wolters-Noordhoff Publ, Groningen (1972)
18. Olver, S.: A general framework for solving Riemann-Hilbert problems numerically. Numer.
Math. 122(2), 305–340 (2012)
19. Olver, S., Slevinsky, R.M.: A fast and well-conditioned spectral method for singular integral
equations (2015). arXiv:1507.00596
20. Olver, S., Townsend, A.: A fast and well-conditioned spectral method. SIAM Rev. 55(3),
462–489 (2013)
176 F. Bornemann et al.

21. Olver, S., Townsend, A.: A practical framework for infinite-dimensional linear algebra. In:
Proceedings of the First Workshop for High Performance Technical Computing in Dynamic
Languages, pp. 57–62 (2014)
22. Trefethen, L.N., Weideman, J.A.C.: The exponentially convergent trapezoidal rule. SIAM Rev.
56(3), 385–458 (2014)
23. Trogdon, T.: On the application of GMRES to oscillatory singular integral equations. BIT
55(2), 591–620 (2015)
24. Wechslberger, G.: Automatic Contour Deformation of Riemann–Hilbert Problems. Ph.D. the-
sis, Technische Universität München (2015)
25. Wechslberger, G., Bornemann, F.: Automatic deformation of Riemann-Hilbert problems with
applications to the Painlevé II transcendents. Constr. Approx. 39(1), 151–171 (2014)
26. Zhou, X.: The Riemann-Hilbert problem and inverse scattering. SIAM J. Math. Anal. 20(4),
966–986 (1989)
A Variational Principle for Cyclic Polygons
with Prescribed Edge Lengths

Hana Kouřimská, Lara Skuppin and Boris Springborn

Abstract We provide a new proof of the elementary geometric theorem on the exis-
tence and uniqueness of cyclic polygons with prescribed side lengths. The proof
is based on a variational principle involving the central angles of the polygon as
variables. The uniqueness follows from the concavity of the target function. The
existence proof relies on a fundamental inequality of information theory. We also
provide proofs for the corresponding theorems of spherical and hyperbolic geom-
etry (and, as a byproduct, in 1 + 1 spacetime). The spherical theorem is reduced
to the Euclidean one. The proof of the hyperbolic theorem treats three cases sepa-
rately: Only the case of polygons inscribed in compact circles can be reduced to the
Euclidean theorem. For the other two cases, polygons inscribed in horocycles and
hypercycles, we provide separate arguments. The hypercycle case also proves the
theorem for “cyclic” polygons in 1 + 1 spacetime.

1 Introduction

This article is concerned with cyclic polygons, i.e., convex polygons inscribed in a
circle. We will provide a new proof of the following elementary theorem in Sect. 2.

Theorem 1.1 There exists a Euclidean cyclic polygon with n ≥ 3 sides of lengths
1 , . . . , n ∈ R>0 if and only if they satisfy the polygon inequalities

H. Kouřimská · L. Skuppin · B. Springborn (B)

Inst. für Mathematik, Technische Universität Berlin,
Straße des 17. Juni 136, 10623 Berlin, Germany
e-mail: [email protected]
H. Kouřimská
e-mail: [email protected]
L. Skuppin
e-mail: [email protected]


c The Author(s) 2016 177
A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_5
178 H. Kouřimská et al.


n
k < i , (1)
i=1
i=k

and this cyclic polygon is unique.

Our proof involves a variational principle with the central angles as variables.
The variational principle has a geometric interpretation in terms of volume in
3-dimensional hyperbolic space (see Remark 2.6). Another striking feature of
our proof is the use of a fundamental inequality of information theory:

Theorem (Information Inequality) Let p = ( p1 , . . . , pm ) and q = (q1 , . . . , qm ) be

discrete probability distributions, then


m
pk
pk log ≥ 0, (2)
k=1
qk

and equality holds if and only if p = q.

The left hand side of inequality (2) is called the Kullback–Leibler divergence
or information gain of q from p, also the relative entropy of p with respect to q.
The inequality follows from the strict concavity of the logarithm function (see, e.g.,
Cover and Thomas [3]).
In Sects. 3 and 4 we provide proofs for non-Euclidean versions of Theorem 1.1.
The spherical version requires an extra inequality:

Theorem 1.2 There exists a spherical cyclic polygon with n ≥ 3 sides of lengths
1 , . . . , n ∈ R>0 if and only if they satisfy the polygon inequalities (1) and


n
i < 2π, (3)
i=1

and this cyclic spherical polygon is unique.

Inequality (3) is necessary because the perimeter of a circle in the unit sphere
cannot be greater than 2π , and the perimeter of the inscribed polygon is a lower
bound. We require strict inequality to exclude polygons that degenerate to great
circles (with all interior angles equal to π ).
In Sect. 3, we prove Theorem 1.2 by a straightforward reduction to Theorem 1.1:
connecting the vertices of a spherical cyclic polygon by straight line segments in the
ambient Euclidean R3 , one obtains a Euclidean cyclic polygon.
In the case of hyperbolic geometry, the notion of “cyclic polygon” requires addi-
tional explanation. We call a convex hyperbolic polygon cyclic if its vertices lie on
a curve of constant non-zero curvature. Such a curve is either
A Variational Principle for Cyclic Polygons with Prescribed Edge Lengths 179

• a hyperbolic circle if the curvature is greater than 1,

• a horocycle if the curvature is equal to 1,
• a hypercycle, i.e., a curve at constant distance from a geodesic if the curvature is
strictly between 0 and 1.

Theorem 1.3 There exists a hyperbolic cyclic polygon with n ≥ 3 sides of lengths
1 , . . . , n ∈ R if and only if they satisfy the polygon inequalities (1), and this cyclic
hyperbolic polygon is unique.

We prove this theorem in Sect. 4. The case of hyperbolic polygons inscribed in

circles can be reduced to Theorem 1.1 by considering the hyperboloid model of the
hyperbolic plane: Connecting the vertices of a hyperbolic polygon inscribed in a
circle by straight line segments in the ambient R2,1 , one obtains a Euclidean cyclic
polygon.
The cases of polygons inscribed in horocycles and hypercycles cannot be reduced
to the Euclidean case because the intrinsic geometry of the affine plane of the poly-
gon is not Euclidean: In the horocycle case, the scalar product is degenerate with a
1-dimensional kernel. Hence, this case reduces to the case of degenerate polygons
inscribed in a straight line. It is easy to deal with. In the hypercycle case, the scalar
product is indefinite. This case reduces to polygons inscribed in hyperbolas in flat
1 + 1 spacetime. The variational principle of Sect. 2 can be adapted for this case (see
Sect. 5), but the corresponding target function fails to be concave or convex. It may
be possible to base a proof of existence and uniqueness on this variational principle,
perhaps using a min-max-argument, but we do not pursue this route in this arti-
cle. Instead, we deal with polygons inscribed in hypercycles using a straightforward
analytic argument.
Some history, from ancient to recent. Theorems 1.1–1.3 belong to the circle of
results connected with the classical isoperimetric problem. As the subject is ancient
and the body of literature is vast, we can only attempt to provide a rough histor-
ical perspective and ask for leniency regarding any essential work that we fail to
mention.
The early history of the relevant results about polygons is briefly discussed by
Steinitz [13, Sect. 16]. Steinitz goes on to discuss analogous results for polyhedra, a
topic into which we will not go. A more recent and comprehensive survey of proofs
of the isoperimetric property of the circle was given by Blåsjö [2].
It was known to Pappus that the regular n-gon had the largest area among n-gons
with the same perimeter, and that the area grew with the number of sides. This was
used to argue for the isoperimetric property of the circle:

Theorem 1.4 (Isoperimetric Theorem) Among all closed planar curves with given
length, only the circle encloses the largest area.

It is not clear who first stated the following theorem about polygons:

Theorem 1.5 (Secant Polygon) Among all n-gons with given side lengths, only the
one inscribed in a circle has the largest area.
180 H. Kouřimská et al.

This was proved by Moula [8], by L’Huilier [5] (who cites Moula), and by
Steiner [12] (who cites L’Huilier). L’Huilier also proved the following theorem:

Theorem 1.6 (Tangent Polygon) Among all convex n-gons with given angles, only
the one circumscribed to a circle has the largest area when the perimeter is fixed
and and smallest perimeter when the area is fixed.

Steiner also proves versions of Theorems 1.5 and 1.6 for spherical polygons.
None of these authors deemed it necessary to prove the existence of a maximizer,
an issue that became generally recognized only after Weierstrass [14]. For polygons,
the existence of a maximizer follows by a standard compactness argument.
Blaschke [1, Sect. 12], notes that the quadrilateral case (n = 4) of Theorem 1.5
can easily be deduced from the Isoperimetric Theorem 1.4 using Steiner’s four-
hinge method. Conversely, one can similarly deduce Theorem 1.4 and the general
Theorem 1.5 from the quadrilateral case of Theorem 1.5. He remarks that the quadri-
lateral case of Theorem 1.5 can be proved directly by deriving the following equa-
tion for the area A of a quadrilateral with sides k :

A2 = (s − 1 )(s − 2 )(s − 3 )(s − 4 ) − 1 2 3 4 cos2 θ, (4)

where s = (1 + 2 + 3 + 4 )/2 is half the perimeter, and θ is the arithmetic mean
of two opposite angles.
Neither Blaschke, nor Steiner, L’Huilier, or Moula provide an argument for the
uniqueness of the maximizer in Theorem 1.5 or 1.6. It seems that even after Weier-
strass, the fact that the sides determine a cyclic polygon uniquely was considered
too obvious to deserve a proof.
Penner [9, Theorem 6.2] gives a complete proof of Theorem 1.1. He proceeds
by showing that there is one and only one circumcircle radius that allows the con-
struction of a Euclidean cyclic polygon with given sides (provided they satisfy the
polygon inequalities).
Schlenker [11] proves Theorems 1.2 and 1.3, and also the isoperimetric prop-
erty of non-Euclidean cyclic polygons, i.e., the spherical and hyperbolic versions of
Theorem 1.5. His proofs of the isoperimetric property are based on the remarkable
equation 
α̇i vi = 0 (5)

characterizing the change of angles αi of a spherical or hyperbolic polygon under

infinitesimal deformations with fixed side lengths. Here, vi ∈ R3 are the position
vectors of the polygon’s vertices in the sphere or in the hyperboloid, respectively.
To prove the uniqueness of spherical and hyperbolic cyclic polygons with given
sides he uses separate arguments similar to Penner’s.
A Variational Principle for Cyclic Polygons with Prescribed Edge Lengths 181

2 Euclidean Polygons. Proof of Theorem 1.1

To construct an inscribed polygon with given side lengths  = (1 , . . . , n ) ∈ Rn>0

(see Fig. 1) is equivalent to finding a point (α1 , . . . αn ) in the set
 
 
n
Dn = α ∈ Rn>0  αk = 2π ⊂ Rn (6)
k=1

satisfying, for some R ∈ R and for all k ∈ {1, . . . , n},

k αk
= R sin . (7)
2 2
This problem admits the following variational formulation. Define the function
f  : Rn → R by

n
 
f  (α) = Cl2 (αk ) + log(k ) αk (8)
k=1

where Cl2 denotes Clausen’s integral [4]:

 
x
 t 
Cl2 (x) = − log 2 sin  dt. (9)
0 2

Clausen’s integral is closely related to Milnor’s Lobachevsky function [6]:

1
L(x) = Cl2 (2x).
2

Fig. 1 Euclidean polygon

inscribed in a circle
182 H. Kouřimská et al.

Fig. 2 Graph of Clausen’s

integral Cl2 (x)

The function Cl2 : R → R is continuous, 2π -periodic, and odd. It is differentiable

except at integer multiples of 2π where the graph has vertical tangents (see Fig. 2).

Proposition 2.1 (Variational Principle) A point α ∈ Dn is a critical point of f 

restricted to Dn if and only if there exists an R ∈ R satisfying equations (7).

Proof A point α ∈ Dn is a critical point of f  restricted to Dn if and only if there

log R such that ∇ f  (α) = (log R)∇g(α) for the con-
exists a Lagrange multiplier
straint function g(α) = αk , i.e.,
⎛   ⎞ ⎡ ⎤
− log 2 sin α1 
2
+ log 1 1
⎜ .. ⎟ ⎢ .. ⎥
⎝ .  ⎠ = log R ⎣ . ⎦.

− log 2 sin αn  + log n
2
1

Since 0 < αk < 2π we may omit the absolute value signs, obtaining equations
(7). 

Thus, to prove Theorem 1.1, we need to show that f  has a critical point in Dn
if and only if the polygon inequalities (1) are satisfied, and that this critical point
is then unique. The following proposition and corollary deal with the uniqueness
claim.

Proposition 2.2 The function f  is strictly concave on Dn .

Corollary 2.3 If f  has a critical

point in Dn , it is the unique maximizer of f  in
the closure D̄n = {α ∈ Rn≥0 | αk = 2π }.

This proves the uniqueness claim of Theorem 1.1.

Proof (of Proposition 2.2) We will show that


n
Vn (α) = Cl2 (αk ) (10)
k=1
A Variational Principle for Cyclic Polygons with Prescribed Edge Lengths 183

is strictly concave on Dn . Since Vn differs from f  by a linear function, this is

equivalent to the claim.
Rivin [10, Theorem 2.1] showed that V3 is strictly concave on D3 . For n > 3 we
proceed by induction on n by “cutting off a triangle”: first, note the obvious identity

Vn (α1 , . . . , αn ) = Vn−1 (α1 , . . . , αn−1 + αn ) − Cl2 (αn−1 + αn )

+ Cl2 (αn−1 ) + Cl2 (αn ).

Since Clausen’s integral is 2π -periodic and odd,

 n−2 

−Cl2 (αn−1 + αn ) = Cl2 (2π − αn−1 − αn ) = Cl2 αk ,
k=1

so
 n−2 

Vn (α1 , . . . , αn ) = Vn−1 (α1 , . . . , αn−1 + αn ) + V3 αk , αn−1 , αn .
k=1

Hence, if Vn−1 and V3 are strictly concave on Dn−1 and D3 , respectively, the claim
for Vn follows. 

Since f  attains its maximum on the compact set D̄n , it remains to show that the
maximum is attained in Dn if and only if the polygon inequalities (1) are satisfied.
This is achieved by the following Propositions 2.4 and 2.5.
Note that D̄n is an (n − 1)-dimensional simplex in Rn . Its vertices are the points
2π e1 , . . . , 2π en , where ek are the canonical basis vectors of Rn . The relative bound-
ary of the simplex D̄n is

∂ D̄n = {α ∈ D̄n | αk = 0 for at least one k}. (11)

Proposition 2.4 If the function f  attains its maximum on the simplex D̄n at a
boundary point α ∈ ∂ D̄n , then α is a vertex.

Proof Suppose α ∈ ∂ D̄n is not a vertex. We need to show that f  does not attain
its maximum at α. This follows from the fact that the derivative of f  in a direction
pointing towards Dn is +∞.
Indeed, suppose v ∈ Rn≥0 , k vk = 0 and vk > 0 if αk = 0. Then α + tv ∈ Dn
for small enough t > 0, and because lim x→0 Cl
2 (x) = +∞,

d
lim f  (α + tv) = +∞. (12)
t→0 dt

Hence f  (α + tv) > f  (α) for small enough t > 0. 

184 H. Kouřimská et al.

Proposition 2.5 The function f  attains its maximum on D̄n at a vertex 2π ek if and
only if
 n
k ≥ i . (13)
i=1
i=k

Proof By symmetry, it is enough to consider the case k = n, i.e., to show that the
function f  attains its maximum on D̄n at the vertex (0, . . . , 0, 2π ) if and only
n−1
if n ≥ k=1 k . To this end, we will calculate the directional derivative of f  in
directions v ∈ Rn pointing inside Dn , i.e., satisfying


n−1
vk ≥ 0 for k ∈ {1, . . . , n − 1}, vn = − vk < 0.
k=1

Since we are only interested in the sign, we may assume v to be scaled so that


n−1
vk = 1, vn = −1.
k=1

Clausen’s integral has the asymptotic behavior

Cl2 (x) = −x log |x| + x + o(x) as x → 0. (14)

This can be seen by considering

x   x
Cl2 (x) = − 0 log (2 sin 2t )/t  dt − 0 log |t| dt.

Using (14) and the 2π -periodicity of Clausen’s integral, one obtains


n
 
f  (2π en + tv) − f  (2π en ) = − tvk log |vk | + tvk log k + o(t)
k=1


n−1
vk
=− tvk log − t log n + o(t),
k=1
n

and hence
d  
n−1
vk
 f (2π en + tv) = − vk log − log n .
dt t=0 k=1
k

Now we invoke the information inequality (2) for the discrete probability distri-
n−1
butions (v1 , . . . , vn−1 ) and (1 , . . . , n−1 )/ k=1 k . Thus,
A Variational Principle for Cyclic Polygons with Prescribed Edge Lengths 185
   
d  
n−1 n−1
vk k=1 k
 f (2π en + tv) = − vk log n−1 + log .
dt t=0 k=1 n / m=1 m n
  
≤0

n−1
If n ≥ k , then
d 
k=1

 f (2π en + tv) ≤ 0.
dt t=0
With the concavity of f  (Proposition 2.2), this implies that f  attains its maximum
on D̄n at (0, . . . , 0, 2π ). n−1 n−1
If, on the other hand, n < k=1 k , then we obtain, for vk = k / m=1 m ,

d 
 f (2π en + tv) > 0.
dt t=0

This implies that f  does not attain its maximum at (0, . . . , 0, 2π ). 

This completes the proof of Theorem 1.1.

Remark 2.6 The function Vn has the following interpretation in terms of hyper-
bolic volume [6]. Consider a Euclidean cyclic n-gon with central angles α1 , . . . , αn .
Imagine the Euclidean plane of the polygon to be the ideal boundary of hyperbolic
3-space in the Poincaré upper half-space model. Then the vertical planes through
the edges of the polygon and the hemisphere above its circumcircle bound a hyper-
bolic pyramid with vertices at infinity. Its volume is 21 Vn (α1 , . . . , αn ). Together with
Schläfli’s differential volume equation (rather, Milnor’s generalization that allows
for ideal vertices [7]), this provides another way to prove Proposition 2.1.

3 Spherical Polygons. Proof of Theorem 1.2

The polygon inequalities (1) are clearly necessary for the existence of a spheri-
cal cyclic polygon because every side is a shortest geodesic. That inequality (3) is
also necessary was already noted in the introduction. It remains to show that these
inequalities are also sufficient, and that the polygon is unique.
We reduce the spherical case to the Euclidean one as shown in Fig. 3. Connecting
the vertices of a spherical cyclic polygon with line segments in the ambient Euclid-
ean space, one obtains a Euclidean cyclic polygon whose circumradius is smaller
than 1. Conversely, every Euclidean polygon inscribed in a circle of radius less than
1 corresponds to a unique spherical cyclic polygon. The spherical side lengths  are
related to the Euclidean lengths ¯ by


¯ = 2 sin . (15)
2
186 H. Kouřimská et al.

Fig. 3 Spherical and

Euclidean polygons

It remains to show the following two propositions:

Proposition 3.1 If the spherical lengths  ∈ Rn>0 satisfy the inequalities (1) and (3),
then the Euclidean lengths ¯ defined by (15) satisfy the inequalities (1) as well. By
Theorem 1.1 there is then a unique Euclidean cyclic polygon P¯ with side lengths . ¯
Proposition 3.2 The circumradius R̄ of the polygon P¯ of Proposition 3.1 is strictly
less than 1.
We will use the following estimate in the proof of Proposition 3.1:

Lemma 3.3 (Sum of Sines Estimate) If β1 , . . . , βn ∈ R≥0 satisfy nk=1 βk ≤ π ,
then   
n n
sin βk ≤ sin βk . (16)
k=1 k=1

Proof (of Lemma 3.3) By induction on n, the base case n = 1 being trivial. For the
inductive step, use the addition theorem,

n+1  
n  
n 
sin βk = sin βk cos βn+1 + cos βk sin βn+1 ,
k=1 k=1 k=1

and note that the cosines are ≤ 1. 

Remark 3.4 The statement of Lemma 3.3 can be strengthened. Equality holds
in (16) if and only if at most one βk is greater than zero. This is easy to see, but
we do not need this stronger statement in the following proof.
A Variational Principle for Cyclic Polygons with Prescribed Edge Lengths 187

Proof (of Proposition 3.1) Suppose 1 , . . . , n ∈ R>0 satisfy the polygon inequali-
ties (1) and (3). We need to show that ¯1 , . . . , ¯n defined by (15) satisfy

¯k < ¯i . (17)
i=k

To this end, we will show that

 
k i
sin < sin , (18)
2 i=k
2

from which inequality (17) follows by Lemma 3.3. To prove inequality (18), we
consider two cases separately.

• i=k i ≤ π . Inequality (18) simply follows from the polygon inequality k <
i=k i and the monotonicity closed interval [0, π2 ].
of the sine function on the
• i=k i ≥ π . Note that 2π > i i implies 2π − k > i=k i , and hence

2π > 2π − k > i ≥ π. (19)
i=k

Inequality (18) follows from sin 2k = sin(π − k

2
) and the monotonicity of the
sine function on the closed interval [ π2 , π ].
This completes the proof of (18) and hence the proof of Proposition 3.1. 
Proof (of Proposition 3.2) Let αk be the central angles of the Euclidean cyclic poly-
gon P¯. Then
k ¯k αk
sin = = R̄ sin , (20)
2 2 2
by (7) and (15). Note that αk are the central angles of both the Euclidean and the
spherical polygon (provided it exists). We consider
two cases separately.

First, suppose that αk ≤ π for all k. Since k k < 2π = k αk , there is some
Then sin 2k < sin α2k , and equation (20) implies that R̄ < 1.
k such that k < αk .
Otherwise, since k αk = 2π , there is exactly one i such that αi > π , and αk <
π for all k = i. By symmetry, it is enough to consider the case

α1 > π, αk < π for k ∈ {2, . . . , n}.

For future reference, we note that α1 > π implies that ¯1 is the longest side of P¯.
(Use (20) and the monotonicity of the sine function.)
We will show R̄ < 1 by induction on n. First, assume n = 3. Then (18) says

1 2 + 3
sin < sin .
2 2
188 H. Kouřimská et al.

By (20) and using 2π − α1 = α2 + α3 , we have

1 α2 α3 α2 α3
sin = R̄ sin cos + R̄ cos sin , (21)
2 2 2 2 2
and
2 + 3 2 3 2 3
sin = sin cos + cos sin
2 2 2 2 2 (22)
α2 3 2 α3
= R̄ sin cos + R̄ cos sin .
2 2 2 2

For at least one k ∈ {2, 3}, cos α2k < cos 2k and hence sin α2k > sin 2k . Equation (20)
implies R̄ < 1.
Now assume that R̄ < 1 has already been shown if P¯ has at most n sides. Sup-
pose P¯ has n + 1 sides. The idea of the following argument is to cut off a triangle
with sides ¯n , ¯n+1 , and λ̄ = 2 R̄ sin αn +α
2
n+1
. Since λ̄ ≤ ¯1 (the longest side), and
¯1 ≤ 2 by (15), we may define λ = 2 arcsin λ̄2 . Now assume R̄ ≥ 1. Then, by the
inductive hypothesis, the polygon inequalities (1) or (3) are violated for the cut-off
triangle and the remaining n-gon. Inequality (3) cannot be violated because it was
assumed to hold for 1 , . . . , n+1 . Hence,

1 ≥ 2 + · · · + n−1 + λ and λ ≥ n + n+1 .

This implies 1 ≥ 2 + · · · + n+1 . Conversely, if (1) and (3) hold, then R̄ < 1. This
completes the proof of Proposition 3.2. 

4 Hyperbolic Polygons. Proof of Theorem 1.3

The polygon inequalities (1) are clearly necessary for the existence of a hyperbolic
cyclic polygon, because every side is a shortest geodesic. It remains to show that
they are also sufficient, and that the polygon is unique, i.e., Proposition 4.2. First,
we review some basic facts from hyperbolic geometry.
As in the spherical case (Sect. 3), we will connect vertices by straight line seg-
ments in the ambient vector space. But instead of the sphere, we consider the hyper-
bolic plane in the hyperboloid model,

H2 = {x ∈ R2,1 | x, x = −1, x3 > 0},

where R2,1 denotes the vector space R3 equipped with the scalar product

x, y = x1 y1 + x2 y2 − x3 y3 ,
A Variational Principle for Cyclic Polygons with Prescribed Edge Lengths 189

and lengths and angles in H2 are measured using the Riemannian metric induced by
this scalar product.
Straight lines (i.e., geodesics) in H2 are the intersections of H2 with
2-dimensional subspaces of R3 . The length  of the geodesic segment connecting
points p, q ∈ H2 is determined by

cosh  = − p, q .

The length of the straight line segment connecting points p, q ∈ H2 in the ambient
R2,1 is 
¯ =  p − q, p − q .

This chordal length ¯ and the hyperbolic distance  are related by

¯ 
= sinh . (23)
2 2

An affine plane in R2,1 is called spacelike, lightlike, or timelike, if the restric-

tion of the scalar product ·, · to (the tangent space of) the affine plane is positive
definite, positive semidefinite, or indefinite, respectively. In terms of the standard
Euclidean metric on R3 , a plane is spacelike, lightlike, or timelike if its slope is less
than, equal to, or greater than 45◦ .
A curve of intersection of H2 with an affine plane in R2,1 that does not contain 0
is a hyperbolic circle, a horocycle, or a hypercycle, depending on whether the plane
is spacelike, lightlike, or timelike.
Thus, connecting the vertices of a hyperbolic cyclic polygon by straight line seg-
ments in the ambient R2,1 , one obtains a planar polygon in R2,1 , but the intrinsic
geometry of the plane will only be Euclidean if the hyperbolic polygon is inscribed
in a circle (see Fig. 4). If the polygon is inscribed in a horocycle or hypercycle,
then the geometry of the plane will be degenerate with signature (+, 0) or a 1 + 1-
spacetime with signature (+, −), respectively.

Proposition 4.1 Let P be a hyperbolic cyclic polygon with side lengths 1 , . . . ,

n ∈ R>0 , and let ¯k be the chordal lengths (23). If P is inscribed in
(i) a circle then

n
¯k < ¯i for all k. (24)
i=1
i=k

(ii) a horocycle then


n
¯k = ¯i for one k. (25)
i=1
i=k
190 H. Kouřimská et al.

Fig. 4 Hyperbolic polygon

inscribed in a circle, shown
in the hyperboloid model

(iii) a hypercycle then


n
¯k > ¯i for one k. (26)
i=1
i=k

Proof (i) If P is inscribed in a circle, then the chordal polygon obtained by con-
necting the vertices of P by straight line segments in R2,1 is a Euclidean polygon.
Hence, its side lengths ¯ satisfy (24).
(ii) If P is inscribed in a horocycle, then the chordal length ¯k of a side is equal
to the length of the arc of the horocycle between its vertices (see Fig. 5). Since one
horocycle arc comprises all others, this implies (25).
(iii) If P is inscribed in a hypercycle at distance R from a geodesic g, then the
chordal lengths ¯k , the hypercycle “radius” R, and the distances ak between the foot
points of the perpendiculars from the vertices to g (see Fig. 6) are related by

¯k ak
= cosh(R) sinh . (27)
2 2

¯
5 Polygon inscribed in a horocycle, shown in the Poincaré half-plane model. Here, n =
Fig.
¯
i=n i is the largest chordal length. Since all horocycles are congruent, we may without loss of
generality assume that the polygon is inscribed in the horocycle y = 1
A Variational Principle for Cyclic Polygons with Prescribed Edge Lengths 191

Fig. 6 Polygon inscribed in

a hypercycle, shown in the
Poincaré disk model. Here,
n and ¯n are the largest side
length and chordal length,
respectively

Since one of the segments of g comprises all others,


ak = ai for one k.
i=k

With (27) this implies (26):

¯k  ai !  ai  ¯i
= cosh(R) sinh > cosh(R) sinh = ,
2 i=k
2 i=k
2 i=k
2

where we have used the inequality sinh(x + y) > sinh(x) + sinh(y), which holds
for positive x, y. This follows immediately from the addition theorem for the hyper-
bolic sine function.
This completes the proof of Proposition 4.1. 
Proposition 4.2 If  ∈ Rn>0 satisfies the polygon inequalities (1), then there exists
a unique hyperbolic cyclic polygon with these side lengths.
Proof Suppose  ∈ Rn>0 satisfies the polygon inequalities (1). Let ¯ be the corre-
sponding chordal lengths (23). We will treat each case of Proposition 4.1 separately.
In each case, we will tacitly use Proposition 4.1 and its proof. Our treatment of case
(iii) is analogous to Penner’s proof [9] of Theorem 1.1 (his Theorem 6.2).
(i) If the chordal lengths ¯ satisfy condition (24), then the existence and unique-
ness of a hyperbolic cyclic polygon with side lengths  follows from the existence
and uniqueness of a Euclidean cyclic polygon with side lengths , ¯ i.e., from Theo-
rem 1.1 (see Fig. 4).
192 H. Kouřimská et al.

(ii) If the chordal lengths ¯ satisfy condition (25), then the corresponding hyper-
bolic cyclic polygon can be constructed by marking off the lengths ¯i for i = k
along a horocycle (see Fig. 5). To see the uniqueness claim, note that all horocycles
are congruent.
(iii) It remains to consider the case that the chordal lengths ¯ satisfy condi-
tion (26). For simplicity, we will assume that n is the largest side length. Then
¯n is the largest chordal length and condition (26) says


n−1
¯n > ¯k . (28)
k=1

Now suppose P is a hyperbolic polygon with side lengths  that is inscribed in

a hypercycle at distance R from its geodesic g, and let

R̄ = cosh R. (29)

Then the distances ak between the foot points (see Fig. 6) satisfy


n−1
an = ak ,
k=1

Using (27), one obtains

¯n !  ¯k !
n−1
arsinh = arsinh . (30)
2 R̄ k=1
2 R̄

Conversely, if, for given , a number R̄ > 1 satisfies (30) then R defined by (29)
is the correct hypercycle distance. More precisely, one can then construct a hyper-
bolic cyclic polygon with side lengths  by marking off the distances a1 , . . . , an−1
determined by (27) along a geodesic and intersect the perpendiculars in the marked
points with a hypercycle at distance R (see Fig. 6).
It remains to show that there is exactly one R̄ > 1 satisfying (30). To this end,
consider the function

¯n !  ¯k !
n−1
Φ(x) = arsinh − arsinh . (31)
2x k=1
2x

We need to show that Φ has exactly one zero in the interval (1, ∞). Using (23), we
see
1 
n−1 !
Φ(1) = n − k < 0. (32)
2 k=1
A Variational Principle for Cyclic Polygons with Prescribed Edge Lengths 193

For x → ∞,
1 ¯ 
n−1 ! 1!
Φ(x) = n − ¯k + o , (33)
2x k=1
x

so
Φ(x) > 0 for large x.

By continuity, Φ has at least one zero in the interval (1, ∞).

Finally, we will show that the derivative
 ¯n 
n−1 ¯k 
1
Φ
(x) = −" 2x
 ¯n 2 + " 2x
 ¯k 2 , (34)
x
1+ 2x
k=1 1+ 2x

is positive at the positive zeroes of Φ. This implies that Φ has at most one zero in
R>0 . Let us define
¯k !
ak (x) = arsinh ,
2x
so

n−1
Φ(x) = an (x) − ak (x),
k=1

and
1  !
n−1
Φ
(x) = − tanh an (x) + tanh ak (x) .
x k=1

The claim follows from the inequality


m 
m
tanh aj < tanh a j ,
j=1 j=1

which holds for m ≥ 2 and positive numbers a j . (Use induction on m and the addi-
tion theorem for tanh for the base case m = 2.)
This concludes the proof of Proposition 4.2. 

5 Concluding Remarks on 1 + 1 Spacetime

The scalar product of R1,1 is

x, y 1,1 = x1 y1 − x2 y2 ,
194 H. Kouřimská et al.

and the length of a spacelike vector x is  = x, x 1,1 . The proof of Theorem 1.3
for polygons inscribed in hypercycles (Sect. 4) also proves the following theorem
about “cyclic” polygons in 1 + 1 spacetime.

Theorem 5.1 There exists a polygon in R1,1 with n ≥ 3 spacelike sides with lengths
1 , . . . , n > 0 that is inscribed in one branch of a hyperbola x, x 1,1 = −R 2 if and
only if
n
k > i for one k, (35)
i=1
i=k

and this polygon is unique.

Without loss of generality, we will assume that the nth side is the longest, i.e., k =
n in (35). Like in the Euclidean case (Sect. 2), the construction of such an inscribed
polygon in R1,1 is equivalent to the following analytic problem: Find a point a ∈
Rn>0 satisfying

n−1
an = ai (36)
i=1

and
k ak
= R sinh (37)
2 2
for some R ∈ R and all k ∈ {1, . . . , n}.
This problem admits the following variational formulation. Define the function
ϕ : Rn → R by


n−1
   
ϕ (a) = Clh2 (ak ) + log(k ) ak − Clh2 (an ) + log(n ) an ,
k=1

where Clh2 denotes the “hyperbolic version” of Clausen’s integral,

 
x
 t 
Clh2 (x) = − log 2 sinh  dt.
0 2

(This notation is not standard.) The function Clh2 (x) can be expressed in terms of
the real part of the dilogarithm function:

x2 π2
Clh2 (x) = Re Li2 (e x ) + − .
4 6
Like in the Euclidean case, one sees that a ∈ Rn>0 is a critical point of ϕ under
the constraint (36) if and only if there is an R satisfying equations (37). However,
the function ϕ is neither concave nor convex on the subspace (36), so any proof
A Variational Principle for Cyclic Polygons with Prescribed Edge Lengths 195

of Theorem 5.1 (or the hypercycle case of Theorem 1.3) based on this variational
principle would have to be more involved.

Acknowledgments This research was supported by DFG SFB/TRR 109 “Discretization in Geom-
etry and Dynamics”.

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Blaschke, W.: Kreis und Kugel. Veit u. Co., Leipzig (1916)

2. Blåsjö, V.: The isoperimetric problem. Am. Math. Monthly 112(6), 526–566 (2005)
3. Cover, T.M., Thomas, J.A.: Elements of Information Theory. Wiley, New York (1991)
4. Lewin, L.: Polylogarithms and Associated Functions. North-Holland Publishing Co., New
York-Amsterdam (1981)
5. L’Huilier, S.: De relatione mutua capacitatis et terminorum figurarum, geometrice considerata:
seu de maximis et minimis. Michael Gröll, Warsaw (1782). http://dx.doi.org/10.3931/e-rara-
4049
6. Milnor, J.: Hyperbolic geometry: the first 150 years. Bull. Am. Math. Soc. (N.S.) 6(1), 9–24
(1982)
7. Milnor, J.: The Schläfli differential equality. In: Collected Papers, vol. 1, pp. 281–295. Publish
or Perish Inc., Houston (1994)
8. Moula, F.: De maximis in figuris rectilineis. In: Commentarii Academiae Scientiarum Impe-
rialis Petropolitanae, vol. 9, pp. 138–159. Typis Academiae, Petropolis (1737). http://www.
archive.org/details/commentariiacade09impe
9. Penner, R.C.: The decorated Teichmüller space of punctured surfaces. Comm. Math. Phys.
113(2), 299–339 (1987)
10. Rivin, I.: Euclidean structures on simplicial surfaces and hyperbolic volume. Ann. Math. (2)
139(3), 553–580 (1994)
11. Schlenker, J.M.: Small deformations of polygons and polyhedra. Trans. Am. Math. Soc.
359(5), 2155–2189 (2007)
12. Steiner, J.: Sur le maximum et le minimum des figures dans le plan, sur la sphère, et dans
l’espace en général. J. Reine Angew. Math. 24, 93–152, 190–250 (1842). http://dx.doi.org/10.
1515/crll.1842.24.93
13. Steinitz, E.: Polyeder und Raumeinteilungen. In: Enzyklopädie der mathematischen Wis-
senschaften mit Einschluss ihrer Anwendungen, vol. 3-1-2, chap. IIIAB12. B. G. Teubner,
Leipzig (1914–1931)
14. Weierstrass, K.: Mathematische Werke. VII. Vorlesungen über Variationsrechnung. Georg
Olms Verlagsbuchhandlung, Hildesheim; Johnson Reprint Corp., New York (1967)
Complex Line Bundles Over Simplicial
Complexes and Their Applications

Felix Knöppel and Ulrich Pinkall

Abstract Discrete vector bundles are important in Physics and recently found
remarkable applications in Computer Graphics. This article approaches discrete
bundles from the viewpoint of Discrete Differential Geometry, including a complete
classification of discrete vector bundles over finite simplicial complexes. In particu-
lar, we obtain a discrete analogue of a theorem of André Weil on the classification of
hermitian line bundles. Moreover, we associate to each discrete hermitian line bun-
dle with curvature a unique piecewise-smooth hermitian line bundle of piecewise-
constant curvature. This is then used to define a discrete Dirichlet energy which
generalizes the well-known cotangent Laplace operator to discrete hermitian line
bundles over Euclidean simplicial manifolds of arbitrary dimension.

1 Introduction

Vector bundles are fundamental objects in Differential Geometry and play an impor-
tant role in Physics [2]. The Physics literature is also the main place where discrete
versions of vector bundles were studied: First, there is a whole field called Lat-
tice Gauge Theory where numerical experiments concerning connections in bundles
over discrete spaces (lattices or simplicial complexes) are the main focus. Some of
the work that has been done in this context is quite close to the kind of problems we
are going to investigate here [3, 4, 6].
Vector bundles make their most fundamental appearance in Physics in the form
of the complex line bundle whose sections are the wave functions of a charged par-
ticle in a magnetic field. Here the bundle comes with a connection whose curvature
is given by the magnetic field [2]. There are situations where the problem itself sug-
gests a natural discretization: The charged particle (electron) may be bound to a

F. Knöppel (B) · U. Pinkall

Inst. für Mathematik, Technische Universität Berlin,
Straße des 17. Juni 136, 10623 Berlin, Germany
e-mail: [email protected]
U. Pinkall
e-mail: [email protected]


c The Author(s) 2016 197
A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_6
198 F. Knöppel and U. Pinkall

Fig. 1 A smooth
triangulation of a manifold

certain arrangement of atoms. Modelling this situation in such a way that the elec-
tron can only occupy a discrete set of locations then leads to the “tight binding
approximation” [1, 12, 15].
Recently vector bundles over discrete spaces also have found striking applica-
tions in Geometry Processing and Computer Graphics. We will describe these in
detail in Sect. 2.
In order to motivate the basic definitions concerning vector bundles over sim-
plicial complexes let us consider a smooth manifold M̃ that comes with smooth
triangulation (Fig. 1).
Let Ẽ be a smooth vector bundle over M̃ of rank K. Then we can define a discrete
version E of Ẽ by restricting Ẽ to the vertex set V of the triangulation. Thus E assigns
to each vertex i ∈ V the K-dimensional real vector space Ei := Ẽi . This is the way
vector bundles over simplicial complexes are defined in general: Such a bundle E
assigns to each vertex i a K-dimensional real vector space Ei in such a way that
Ei ∩ E j = ∅ for i = j.
So far the notion of a discrete vector bundle is completely uninteresting mathe-
matically: The obvious definition of an isomorphism between two such bundles E
and Ê just would require a vector space isomorphism f i : Ei → Êi for each vertex
i. Thus, unless we put more structure on our bundles, any two vector bundles of the
same rank over a simplicial complex are isomorphic.
Suppose now that Ẽ comes with a connection ∇. Then we can use the parallel
transport along edges ij of the triangulation to define vector space isomorphisms

ηij : Ẽi → Ẽ j
Complex Line Bundles Over Simplicial Complexes . . . 199

This leads to the standard definition of a connection on a vector bundle over a

simplicial complex: Such a connection is given by a collection of isomorphisms
ηij : Ei → E j defined for each edge ij such that

η ji = ηij−1 .

Now the classification problem becomes non-trivial because for an isomorphism f

between two bundles E and Ê with connection we have to require compatibility with
the transport maps ηij :

f j ◦ ηij = η̂ij ◦ f i .

Given a connection η and a closed edge path γ = e · · · e1 (compare Sect. 4) of

the simplicial complex we can define the monodromy Pγ ∈ Aut(Ei ) around γ as

Pγ = ηe ◦ . . . ◦ ηe1 .

In particular the monodromies around triangular faces of the simplicial complex

provide an analog for the smooth curvature in the discrete setting. In Sect. 4 we will
classify vector bundles with connection in terms of their monodromies.
Let us look at the special case of a rank 2 bundle E that is oriented and comes
with a Euclidean scalar product. Then the 90◦ -rotation in each fiber makes it into
1-dimensional complex vector space, so we effectively are dealing with a hermitian
complex line bundle. If ijk is an oriented face of our simplicial complex, the mon-
odromy P∂ ijk : Ei → Ei around the triangle ijk is multiplication by a complex num-
ber h ijk of norm one. Writing h ijk = eıαijk with −π < αijk ≤ π we see that this mon-
odromy can also be interpreted as a real curvature αijk ∈ (−π, π]. It thus becomes
apparent that the information provided by the connection η cannot encode any cur-
vature that integrated over a single face is larger than ±π. This can be a serious
restriction for applications: We effectively see a cutoff for the curvature that can be
contained in a single face.
Remember however our starting point: We asked for structure that can be nat-
urally transferred from the smooth setting to the discrete one. If we think again
about a triangulated smooth manifold it is clear that we can associate to each two-
dimensional face ijk the integral Ωijk of the curvature 2-form over this face. This
is just a discrete 2-form in the sense of discrete exterior calculus [5]. Including this
discrete curvature 2-form with the parallel transport η brings discrete complex line
bundles much closer to their smooth counterparts:

Definition A hermitian line bundle with curvature over a simplicial complex X is

a triple (E, η, Ω). Here E is complex hermitian line bundle over X, for each edge
ij the maps ηij : Ei → E j are unitary and the closed real-valued 2-form Ω on each
face ijk satisfies

ηki ◦ η jk ◦ ηij = eıΩijk idEi .

200 F. Knöppel and U. Pinkall

In Sect. 7 we will prove for hermitian line bundles with curvature the discrete
analog of a well-known theorem by André Weil on the classification of hermitian
line bundles.
In Sect. 8 we will define for hermitian line bundles with curvature a degree
(which can be an arbitrary integer) and we will prove a discrete version of the
Poincaré-Hopf index theorem concerning the number of zeros of a section (counted
with sign and multiplicity).
Finally we will construct in Sect. 10 for each hermitian line bundle with curva-
ture a piecewise-smooth bundle with a curvature 2-form that is constant on each
face. Sections of the discrete bundle can be canonically extended to sections of the
piecewise-smooth bundle. This construction will provide us with finite elements for
bundle sections and thus will allow us to compute the Dirichlet energy on the space
of sections.

2 Applications of Vector Bundles in Geometry Processing

Several important tasks in Geometry Processing (see the examples below) lead to
the problem of coming up with an optimal normalized section φ of some Euclid-
ean vector bundle E over a compact manifold with boundary M. Here “normalized
section” means that φ is defined away from a certain singular set and where defined
it satisfies |φ| = 1.
In all the mentioned situations E comes with a natural metric connection ∇ and
it turns out that the following method for finding φ yields surprisingly
 good results:
Among all sections ψ of E find one which minimizes M |∇ψ| 2
under the con-
straint M |ψ|2 = 1. Then away from the zero set of ψ use φ = ψ/|ψ|.
The term “optimal” suggests that there is a variational functional which is mini-
mized by φ and this is in fact the case. Moreover, in each of the applications there
are heuristic arguments indicating that φ is indeed a good choice for the problem
at hand. For the details we refer to the original papers. Here we are only concerned
with the Discrete Differential Geometry involved in the discretization of the above
variational problem.

2.1 Direction Fields on Surfaces

Here M is a surface with a Riemannian metric, E = TM is the tangent bundle and

∇ is the Levi-Civita connection. Figure 2 shows the resulting unit vector field φ. If
we consider TM as a complex line bundle, normalized sections of the tensor square
L = TM ⊗ TM describe unoriented direction fields on M. Similarly, “higher order
direction fields” like cross fields are related to higher tensor powers of TM. Higher
order direction fields also have important applications in Computer Graphics.
Complex Line Bundles Over Simplicial Complexes . . . 201

Fig. 2 An optimal direction field on a surface

2.2 Stripe Patterns on Surfaces

A stripe pattern on a surface M is a map which away from a certain singular set
assigns to each point p ∈ M an element φ( p) ∈ S = {z ∈ C||z| = 1}. Such a map φ
can be used to color M in a periodic fashion according to a color map that assigns a
color to each point on the unit circle S. Suppose we are given a 1-form ω on M that
specifies a desired direction and spacing of the stripes, which means that ideally we
would wish for something like φ = eiα with dα = ω. Then the algorithm in [9] says
that we should use a φ that comes from taking E as the trivial bundle E = M × C and
∇ψ = dψ − iωψ. Sometimes the original data come from an unoriented direction
field and (in order to obtain the 1-form ω) we first have to move from M to a double
branched cover M̃ of M. This is for example the case in Fig. 3.

2.3 Decomposing Velocity Fields into Fields Generated

by Vortex Filaments

The velocity fields that arise in fluid simulations quite often can be understood as
a superposition of interacting vortex rings. It is therefore desirable to have an algo-
rithm that reconstructs the underlying vortex filaments from a given velocity field.
Let the velocity field v on a domain M ⊂ R3 be given as a 1-form ω = v, ·. Then
the algorithm proposed in [19] uses the function φ : M → C that results from taking
the trivial bundle E = M × C endowed with the connection ∇ψ = dψ − iωψ. Note
that so far this is just a three-dimensional version of the situation in Sect. 2.2. This
time however we even forget φ in the end and only retain the zero set of ψ as the
filament configuration we are looking for (Fig. 4).
202 F. Knöppel and U. Pinkall

Fig. 3 An optimal stripe pattern aligned to an unoriented direction field

Fig. 4 A knotted vortex

filament defined as the zero
set of a complex valued
function ψ. It is shown as the
intersection of the zero set of
Re ψ with the zero set of  ψ

2.4 Close-To-Conformal Deformations of Volumes

Here the data are a domain M ⊂ R3 and a function u : M → R. The task is to find a
map f : M → R3 which is approximately conformal with conformal factor eu , i.e.
for all tangent vectors X ∈ TM we want

|d f (X )| ≈ eu |X |.

The only exact solutions of this equations are the Möbius transformations. For these
we find

d f (X ) = eu ψ X ψ
Complex Line Bundles Over Simplicial Complexes . . . 203

Fig. 5 Close-to-conformal deformation of a sphere based on a desired conformal factor specified

as the potential of a collection of point charges

for some map ψ : M → H with |ψ| = 1 which in addition satisfies

dψ(X ) = − 21 (grad u × X ) ψ.

Note that here we have identified R3 with the space of purely imaginary quaternions.
Let us define a connection ∇ on the trivial rank 4 vector bundle M × H by

∇ X ψ := dψ(X ) + 21 (grad u × X )ψ.

Then we can apply the usual method and find a section φ : M → H with |φ| = 1. In
general there will not be any f : M → R3 that satisfies

d f (X ) = eu φX φ (1)

exactly but we can always look for an f that satisfies (1) in the least squares sense.
See Fig. 5 for an example.

3 Discrete Vector Bundles with Connection

An (abstract) simplicial complex is a collection X of finite non-empty sets such that

if σ is an element of X so is every non-empty subset of σ [14].
An element of a simplicial complex X is called a simplex and each non-empty
subset of a simplex σ is called a face of σ. The elements of a simplex are called
vertices and the union of all vertices V = ∪σ∈X σ is called the vertex set of X. The
dimension of a simplex is defined to be one less than the number of its vertices:
204 F. Knöppel and U. Pinkall

dim σ := |σ| − 1. A simplex of dimension k is also called a k-simplex. The dimen-

sion of a simplicial complex is defined as the maximal dimension of its simplices.
To avoid technical difficulties, we will restrict our attention to finite simplicial
complexes only. The main concepts are already present in the finite case. Though,
the definitions carry over verbatim to infinite simplicial complexes.
Definition 3.1 Let F be a field and let X be a simplicial complex with vertex set V.
A discrete F-vector bundle of rank K ∈ N over X is a map π : E → V such that for
each vertex i ∈ V the fiber over i

Ei := π −1 ({i})

has the structure of a K-dimensional F-vector space.

Most of the time, we slightly abuse notation and denote a discrete vector bundle
over a simplicial complex schematically by E → X.
The usual vector space constructions carry fiberwise over to discrete vector bun-
dles. So we can speak of tensor products or dual bundles. Moreover, the fibers can be
equipped with additional structures. In particular, a real vector bundle whose fibers
are Euclidean vector spaces is called a discrete Euclidean vector bundle. Similarly, a
complex vector bundle whose fibers are hermitian vector spaces is called a discrete
hermitian vector bundle.
So far discrete vector bundles are completely uninteresting from the mathemat-
ical viewpoint—the obvious definition of an isomorphism f between two discrete
vector bundles E and Ẽ just requires isomorphisms between the fibers f i : Ei → Ẽi .
Thus any two bundles of rank K are isomorphic. This changes if we connect the
fibers along the edges by isomorphisms.
Let σ = {i 0 , . . . , i k } be a k-simplex. We define two orderings of its vertices to be
equivalent if they differ by an even permutation. Such an equivalence class is then
called an orientation of σ and a simplex together with an orientation is called an
oriented simplex. We will denote the oriented k-simplex just by the word i 0 · · · i k .
Further, an oriented 1-simplex is called an edge.
Definition 3.2 Let E → X be a discrete vector bundle over a simplicial complex. A
discrete connection on E is a map η which assigns to each edge ij an isomorphism
ηij : Ei → E j of vector spaces such that

η ji = ηij−1 .

Remark 3.3 Here and in the following a morphism of vector spaces is a linear map
that also preserves all additional structures—if any present. E.g., if we are dealing
with hermitian vector spaces, then a morphism is a complex-linear map that pre-
serves the hermitian metric, i.e. it is a complex linear isometric immersion.
Definition 3.4 A morphism of discrete vector bundles with connection is a map
f : E → F between discrete vector bundles E → X and F → X with connections η
and θ (resp.) such that
Complex Line Bundles Over Simplicial Complexes . . . 205

(i) for each vertex i we have that f (Ei ) ⊂ Fi and the map f i = f |Ei : Ei → Fi is
a morphism of vector spaces,
(ii) for each edge ij the following diagram commutes:

i.e. θij ◦ f i = f j ◦ ηij .

An isomorphism is a morphism which has an inverse map, which is also a morphism.

Two discrete vector bundles with connection are called isomorphic, if there exists
an isomorphism between them.
Again let V denote the vertex set of X. A discrete vector bundle E → X with
connection η is called trivial, if it is isomorphic to the product bundle

FK := V × FK

over X equipped with the connection which assigns to each edge the identity idFK .
It is a natural question to ask how many non-isomorphic discrete vector bundles
with connection exist on a given simplicial complex X.

Remark 3.5 For k ∈ N, the k-skeleton Xk of a simplicial complex X is the subcom-

plex that consists of all simplices σ ∈ X of dimension ≤ k,
 
Xk := σ ∈ X | dim σ ≤ k .

The classification of vector bundles over X only involves its 1-skeleton X1 and could
be equally done just for discrete vector bundles over 1-dimensional simplicial com-
plexes, i.e. graphs. Later on, when we consider discrete hermitian line bundles with
curvature in Sect. 7, the 2- and 3-skeleton come into play and finally, in Sect. 11, we
will use the whole simplicial complex.

4 Monodromy—A Discrete Analogue of Kobayashi’s

Theorem

Let X be a simplicial complex. Each edge ij of X has a start vertex s(ij) := i

and a target vertex t (ij) := j. A edge path γ is a sequence of successive edges
(e1 , . . . , e ), i.e. s(ek+1 ) = t (ek ) for all k = 1, . . . ,  − 1, and will be denoted by
the word:
γ = e · · · e1 .
206 F. Knöppel and U. Pinkall

If i = s(e1 ), we say that γ starts at i, and if j = t (e ) that γ ends at j. The complex
X is called connected, if any two of its vertices can be joined by an edge path. From
now on we will only consider connected simplicial complexes.
Now, let E → X be a discrete vector bundle with connection η. To each edge
path γ = e · · · e1 from i to j, we define the parallel transport Pγ : Ei → E j along
γ by

Pγ := ηe ◦ · · · ◦ ηe1 .

To each edge path γ = e · · · e1 we can assign an inverse path γ −1 . If γ̃ = em · · ·

e+1 starts where γ ends, we can build the concatenation γ̃γ: With the notation
ij −1 := ji, we have

γ −1 := e1−1 · · · e−1 , γ̃γ := em · · · e+1 e · · · e1 .

Whenever γ̃γ is defined,

Pγ̃γ = Pγ̃ ◦ Pγ , Pγ −1 = Pγ−1 . (2)

The elements of the fundamental group are identified with equivalence classes of
edge loops, i.e. edge paths starting and ending a given base vertex i of X, where two
such loops are identified if they differ by a sequence of elementary moves [16]:

e · · · ek+1 e−1 e ek · · · e1 ←→ e · · · ek+1 ek · · · e1 .

Now, by Eq. (2), we see that the parallel transport descends to a representation of
the fundamental group π1 (X1 , i). We encapsulate this in the following

Proposition 4.1 Let E → X be a discrete vector bundle with connection over a

connected simplicial complex. The parallel transport descends to a representation
of the fundamental group π1 (X1 , i):

M : π1 (X1 , i) → Aut(Ei ), [γ] → Pγ .

The representation M will be called the monodromy of the discrete vector bundle E.

If we change the base vertex this leads to an isomorphic representation—an iso-

morphism is just given by the parallel transport Pγ along an edge path joining the
two base vertices. Moreover, if f : E → Ẽ is an isomorphism of discrete vector bun-
dles with connection over the simplicial complex X, for any edge path γ = e · · · e1
from a vertex i to a vertex j the following equality holds:

P̃γ = f j ◦ Pγ ◦ f i−1 .
Complex Line Bundles Over Simplicial Complexes . . . 207

Here P and P̃ denote the parallel transports of E and Ẽ. Thus we obtain:
Proposition 4.2 Isomorphic discrete vector bundles with connection have isomor-
phic monodromies.
In fact, the monodromy completely determines a discrete vector bundle with
connection up to isomorphism, which provides a complete classification of dis-
crete vector bundles with connection: Let X be a connected simplicial complex.
Let E → X be a discrete F-vector bundle of rank K with connection and let
M : π1 (X1 , i) → Aut(Ei ) denote its monodromy.Any choice of a basis of the fiber
Ei determines a group homomorphism ρ ∈ Hom π1 (X1 , i), GL(K, F) . Any other
choice of basis determines a group homomorphism ρ̃ which is related to ρ by con-
jugation, i.e. there is S ∈ GL(K, F) such that

ρ̃([γ]) = S · ρ([γ]) · S −1 for all [γ] ∈ π1 (X1 , i).

Hence the monodromy M determines a well-defined conjugacy class of group

homomorphisms from π1 (X1 , i) to GL(K, F), which we will denote by [M]. The
group GL(K, F) will be referred to as the structure group of E.
Let VK F (X) denote the set of isomorphism
 classes of F-vector
 bundles of rank
K with connection over X and let Hom π1 (X1 , i), GL(K, F) /∼ denote the set of
conjugacy classes of group homomorphisms from the fundamental group π1 (X1 , i)
into the structure group GL(K, F).
 
Theorem 4.3 F : VK F (X) → Hom π1 (X , i), GL(K, F) /∼ , [E]  → [M] is bijec-
1

tive.

Proof By Proposition 4.2, F is well-defined. First we show injectivity. Consider two

discrete vector bundles E, Ẽ over X with connections η, η̃ and let M, M̃ denote their
monodromies. Suppose that [M] = [M̃]. If we choose bases {X i1 , . . . , X iK } of Ei
and { X̃ i1 , . . ., X̃ iK } of Ẽi , then M and M̃ are represented by group homomorphisms
ρ, ρ̃ ∈ Hom π1 (X1 , i), GL(K, F) which are related by conjugation. Without loss of
generality, we can assume that ρ = ρ̃. Now, let T be a spanning tree of X with root i.
Then for each vertex j of X there is an edge path γ i, j from the root i to the vertex j
entirely contained in T. Since T contains no loops, the path γ i, j is essentially unique,
i.e. any two such paths differ by a sequence of elementary moves. Thus we can use
the parallel transport to obtain bases {X 1j , . . . , X Kj } ⊂ E j and { X̃ 1j , . . . , X̃ Kj } ⊂ Ẽ j
at every vertex j of X. With respect to these bases the connections η and η̃ are
represented by elements of GL(K, F). By construction, for each edge e in T the
connection is represented by the identity matrix. Moreover, to each edge e = jk
not contained in T there corresponds a unique [γe ] ∈ π1 (X1 , i). With the notation
−1
above, it is given by γe = γi,k e γ i, j . In particular, on the edge e both connections
are represented by the same matrix ρ([γe ]) = ρ̃([γe ]). Thus, if we define f : E → Ẽ
by f (X mj ) := X̃ mj for m = 1, . . . , K, we obtain an isomorphism, i.e. E ∼ = Ẽ. Hence
F is injective.
208 F. Knöppel and U. Pinkall

K
To see that F is surjective we use T to equip  E := V × F
 the1 product bundle
with a particular connection η. Let ρ ∈ Hom π1 (X , i), GL(K, F) . If e lies in T
we set ηe = id else we set ηe := ρ([γe ]). By construction, F([E]) = [ρ]. Thus F is
surjective. 

Remark 4.4 Note that Theorem 4.3 can be regarded as a discrete analogue of a the-
orem of S. Kobayashi [10, 13], which states that the equivalence classes of connec-
tions on principal G-bundles over a manifold M are in one-to-one correspondence
with the conjugacy classes of continuous homomorphisms from the path group
(M) to the structure group G. In fact, the fundamental group of the 1-skeleton
is a discrete analogue of (M).

5 Discrete Line Bundles—The Abelian Case

In this section we want to focus on discrete line bundles, i.e. discrete vector bundle
of rank 1. Here the monodromy descends to a group homomorphism from the closed
1-chains to the multiplicative group F∗ := F \ {0} of the underlying field. This leads
to a description by discrete differential forms (Sect. 6).
Let L → X be discrete F-line bundle over a connected simplicial complex. In
this case the structure group is just F∗ , which is abelian. Thus we obtain
   
Hom π1 (X1 , i), F∗ ) /∼ = Hom π1 (X1 , i), F∗ .
 
Hom π1 (X1 , i), F∗ carries a natural group structure. Moreover, the isomorphism
classes of discrete line bundles over X form an abelian group. The group structure
is given by the tensor product: For [L], [L̃] ∈ V1F (X), we have

[L][L̃] = [L ⊗ L̃], [L]−1 = [L∗ ].

The identity element is given by the trivial bundle. In the following we will denote
 of 1F-line bundles
the group of isomorphism classes  over X by LFX .
F
The map F : LX → Hom π1 (X , i), F∗ , [L] → [M] is a group homomor-
phism. By Theorem 4.3, F is then an isomorphism.  
Now, since F∗ is abelian, each homomorphism ρ ∈ Hom π1 (X1 , i), F∗ factors
through the abelianization

π1 (X1 , i)ab = π1 (X1 , i)/[π1 (X1 , i), π1 (X1 , i)],

   
i.e. for each ρ ∈ Hom π1 (X1 , i), F∗ there is a unique ρab ∈ Hom π1 (X1 , i)ab , F∗
such that

ρ = ρab ◦ πab .
Complex Line Bundles Over Simplicial Complexes . . . 209

Here πab : π1 (X1 , i) → π1 (X

 , i)ab
1
denotes
 the canonical
 projection.This yields an
isomorphism between Hom π1 (X1 , i), F∗ and Hom π1 (X1 , i)ab , F∗ . In particular,
 
LFX ∼
= Hom π1 (X1 , i)ab , F∗ .

As we will see below, the abelianization π1 (X1 , i)ab is naturally isomorphic to the
group of closed 1-chains.
The group of k-chains Ck (X, Z) is defined as the free abelian group which is
generated by the k-simplices of X. More precisely, let Xork denote the set of oriented
k-simplices of X. Clearly, for k > 0, each k-simplex has two orientations. Inter-
changing these orientations yields a fixed-point-free involution ρk : Xor
k → Xk . The
or

group of k-chains is then explicitly given as follows:

 
Ck (X, Z) := c : Xor
k → Z | c ◦ ρk = −c .

0 = X0 . Thus,
Since simplices of dimension zero have only one orientation, Xor
 
C0 (X, Z) := c : Xor
k →Z .

It is common to identify an oriented k-simplex σ with its elementary k-chain, i.e. the
chain which is 1 for σ, −1 for the oppositely oriented simplex and zero else. With
this identification a k-chain c can be written as a formal sum of oriented k-simplices
with integer coefficients:


m
c= n i σi , n i ∈ Z, σi ∈ Xor
k .
i=1

The boundary operator ∂k : Ck (X, Z) → Ck−1 (X, Z) is then the homomorphism

which is uniquely determined by


k
∂k i 0 · · · i k = (−1) j i 0 · · · ij · · · i k .
j=0

It well-known that ∂k ◦ ∂k+1 ≡ 0. Thus we get a chain complex

∂0 ∂1 ∂2 ∂k ∂k+1
0←
− C0 (X, Z) ←
− C1 (X, Z) ←
− ··· ←
− Ck (X, Z) ←−− · · · .

The simplicial Homology groups Hk (X, Z) may be regarded as a measure for the
deviation of exactness:

Hk (X, Z) := ker ∂k /im ∂k+1 .

The elements of ker ∂k are called k-cycles, those of im ∂k+1 are called k-boundaries.
210 F. Knöppel and U. Pinkall

It is a well-known fact that the abelianization of the first fundamental group is the
first homology group (see e.g. [7]). Now, since the first homology of the 1-skeleton
consists exactly of all closed chains of X, we obtain

π1 (X1 , i)ab ∼
= ker ∂1 .

The isomorphism is induced by the map π1 (X1 , i) → ker ∂1 given by [γ] → j e j ,
where γ = e · · · e1 . We summarize the above discussion in the following theorem.

Theorem 5.1 The group of isomorphism classes of line bundles LFX is naturally
isomorphic to the group Hom(ker ∂1 , F∗ ):

LFX ∼
= Hom(ker ∂1 , F∗ ).

The isomorphism of Theorem 5.1 can be made explicit using discrete F∗ -valued
1-forms associated to the connection of a discrete line bundle.

6 Discrete Connection Forms

Throughout this section X denotes a connected simplicial complex.

Definition 6.1 Let G be an abelian group. The group of G-valued discrete k-forms
is defined as follows:
 
Ω k (X, G) := ω : Ck (X) → G | ω group homomorphism .

The discrete exterior derivative dk is then defined to be the adjoint of ∂k+1 , i.e.

dk : Ω k (X, G) → Ω k+1 (X, G), dk ω := ω ◦ ∂k+1 .

By construction, we immediately get that dk+1 ◦ dk ≡ 0. The corresponding

cochain complex is called the discrete de Rahm complex with coefficients in G:

d0 d1 dk−1 dk
0 → Ω 0 (X, G) −
→ Ω 1 (X, G) −
→ · · · −−→ Ω k (X, G) −
→ ··· .

In analogy to the construction of the homology groups, the k-th de Rahm Cohomol-
ogy group Hk (X, G) with coefficients in G is defined as the quotient group

Hk (X, G) := ker dk /im dk−1 .

The discrete k-forms in ker dk are called closed, those in im dk−1 are called exact.
Complex Line Bundles Over Simplicial Complexes . . . 211

Now, let CL denote the space of connections on the discrete F-line bundle
L → X:
 
CL := η | η connection on L .

Any two connections η, θ ∈ CL differ by a unique discrete 1-form ω ∈ Ω 1 (X, F∗ ):

θ = ωη.

Hence the group Ω 1 (X, F∗ ) acts simply transitively on the space of connections CL .
In particular, each choice of a base connection β ∈ CL establishes an identification

CL  η = ωβ ←→ ω ∈ Ω 1 (X, F∗ ).

Remark 6.2 Note that each discrete vector bundle admits a trivial connection. To
see this choose for each vertex a basis of the corresponding fiber. The correspond-
ing coordinates establish an identification with the product bundle. Then there is a
unique connection that makes the diagrams over all edges commute.

Definition 6.3 Let η ∈ CL . A connection form representing the connection η is a

1-form ω ∈ Ω 1 (X, F∗ ) such that η = ωβ for some trivial base connection β.

Clearly, there are many connection forms representing a connection. We want to

see how two such forms are related.
More generally, two connections η and θ in CL lead to isomorphic discrete line
bundles if and only if for each fiber there is a vector space isomorphism f i : Li →
Li , such that for each edge ij:

θij ◦ f i = f j ◦ ηij .

Since ηe and θe are linear, this boils down to discrete F∗ -valued functions and the
relation characterizing an isomorphism becomes
 
θij = g j gi−1 ηij = (dg)ij ηij ,

i.e. η and θ differ by an exact discrete F∗ -valued 1-form. In particular, the difference
of two connection forms representing the same connection η is exact.
Thus we obtain a well-defined map sending a discrete line bundle L with con-
nection to the corresponding equivalence class of connection forms

[ω] ∈ Ω 1 (X, F∗ )/dΩ 0 (X, F∗ ).

Theorem 6.4 The map F : LFX → Ω 1 (X, F∗ )/dΩ 0 (X, F∗ ), [L] → [ω], where ω is
a connection form of L, is an isomorphism of groups.
212 F. Knöppel and U. Pinkall

Proof Clearly, F is well-defined. Let L and L̃ be two discrete complex line bundle
with connections η and θ, respectively. If β ∈ CL and β̃ ∈ CL̃ are trivial, so is β ⊗
β̃ ∈ CL⊗L̃ . Hence, with η = ωβ and η̃ = ω̃ β̃, we get

F([L ⊗ L̃]) = [ω ω̃] = [ω][ω̃] = F([L])F([L̃]).

By the preceding discussion, F is injective. Surjectivity is also easily checked. 

Next we want to prove that the map given by

Ω 1 (X, F∗ )/dΩ 0 (X, F∗ )  [ω] → ω|ker ∂1 ∈ Hom(ker ∂1 , F∗ )

is a group isomorphism. Clearly, it is a well-defined group homomorphism. We

show its bijectivity in two steps. First, the surjectivity is provided by the following

Lemma 6.5 Let X be a simplicial complex and G be an abelian group. Then the
restriction map  : Ω k (X, G) → Hom(ker ∂k , G), ω → ω|ker ∂k is surjective.

Proof If we choose an orientation for each simplex in X, then ∂k is given by

an integer matrix. Now, there is a unimodular matrix U such that ∂k U = (0|H )
has Hermite normal form. Write U = (A|B), where ∂k A = 0 and ∂k B = H and
let ai denote the columns of A, i.e. A = (a1 , . . . , a ). Clearly, ai ∈ ker ∂k . More-
over, if c ∈ ker ∂k , then 0 = ∂k c = (0|H )U −1 c. Hence U −1 c = (q, 0) , q ∈ Z ,
and thus c = Aq. Therefore {ai | i = 1, . . . , } is a basis of ker ∂k . Now, let μ ∈
Hom(ker ∂k , Z). A homomorphism is completely determined by its values on a
basis. We define ω = (μ(a1 ), . . . , μ(a ), 0 . . . , 0)U −1 . Then ω ∈ Ω k (X, Z) and
ω A = (μ(a1 ), . . . , μ(a )). Hence (ω) = μ and  is surjective for forms with coef-
ficients in Z. Now, let G be an arbitrary abelian group. And μ ∈ Hom(ker ∂k , G).
Now, if a1 , .., a is an arbitrary basis of ker ∂k , then there are forms ω1 , . . . , ω ∈
Ω k (X, Z) such that ωi (a j ) = δij . Since Z acts on G, we can multiply ωi with ele-

ments g ∈ G to obtain forms with coefficients in G. Now, set ω = i=1 ωi · μ(ai ).
Then ω ∈ Ω k (X, G) and ω(ai ) = μ(ai ) for i = 1, . . . , . Thus (ω) = μ. Hence 
is surjective for forms with coefficients in arbitrary abelian groups. 

The injectivity is actually easy to see: If ω|ker ∂1 = 0, we define an F∗ -valued

function f by integration along paths: Fix some vertex i. Then


f ( j) := ω := ω(e),
γ e∈γ

where γ is some path joining i to j. Since ω|ker ∂1 = 0, the value f ( j) does not
depend on the choice of the path γ. Moreover, d f = ω. Together with Lemma 6.5,
this yields the following theorem.

Theorem 6.6 The map F : Ω 1 (X, F∗ )/dΩ 0 (X, F∗ ) → Hom(ker ∂1 , F∗ ) given by

[ω] → ω|ker ∂1 is an isomorphism of groups.
Complex Line Bundles Over Simplicial Complexes . . . 213

Now, let us relate this to Theorem 5.1. Let L → X be a line bundle with connec-
tion η and ω be a connection form representing η, i.e. η = ωβ for some trivial base
connection β. Let [γ] ∈ π1 (X1 , i), where γ = e · · · e1 . By linearity and since trivial
connections have vanishing monodromy, we obtain

M([γ]) = ηe ◦ · · · ◦ ηe1 = ωe · · · ωe1 · βe ◦ · · · ◦ βe1 = ω(πab ([γ])) · id|Li .

Hence, by the uniqueness of [M]ab , we obtain the following theorem that brings
everything nicely together.
Theorem 6.7 Let L → X be a line bundle with connection η. Let M denote its
monodromy and let ω be some connection form representing η. Then, with the iden-
tifications above,

[M]ab = [ω].

7 Curvature—A Discrete Analogue of Weil’s Theorem

In this section we describe complex and hermitian line bundles by their curvature.
For the first time we use more than the 1-skeleton.
Let X be a connected simplicial complex and G an abelian group. Since d 2 = 0,
the exterior derivative descends to a well-defined map on Ω k (X, G)/dΩ k−1 (X, G),
which again will be denoted by d. Explicitly,

d : Ω k (X, G)/dΩ k−1 (X, G) → Ω k+1 (X, G), [ω] → dω.

Definition 7.1 The F∗ -curvature of a discrete F-line bundle L → X is the discrete

2-form Ω ∈ Ω 2 (X, F∗ ) given by

Ω = d[ω],

where [ω] ∈ Ω 1 (X, F∗ )/dΩ 0 (X, F∗ ) represents the isomorphism class [L].

Remark 7.2 Note that Ω just encodes the parallel transport along the boundary of
the oriented 2-simplices of X—the “local monodromy”.

From the definition it is obvious that the F∗ -curvature is invariant under isomor-
phisms. Thus, given a prescribed 2-form Ω ∈ Ω 2 (X, F∗ ), it is a natural question to
ask how many non-isomorphic line bundles have curvature Ω.
Actually, this question is answered easily: If d[ω] = Ω = d[ω̃], then the differ-
ence of ω and ω̃ is closed. Factoring out the exact 1-forms we see that the space
of non-isomorphic line bundles with curvature Ω can be parameterized by the first
214 F. Knöppel and U. Pinkall

cohomology group H1 (X, F∗ ). Furthermore, the existence of a line bundle with cur-
vature Ω ∈ Ω 2 (X, F∗ ) is equivalent to the exactness of Ω.
But when is a k-form Ω exact? Certainly it must be closed. Even more, it must
vanish on every closed k-chain: If Ω = im d and S is a closed k-chain, then

Ω(S) = dω(S) = ω(∂ S) = 0.

For k = 1, as we have seen, this criterion is sufficient for exactness. For k > 1 this
is not true with coefficients in arbitrary groups.

Example 7.3 Consider a triangulation X of the real projective plane RP2 . The zero-
chain is the only closed 2-chain and hence each Z2 -valued 2-form vanishes on every
closed 2-chain. But H2 (X, Z2 ) = Z2 and hence there exists a non-exact 2-form.

In the following we will see that this cannot happen for fields of characteristic
zero or, more generally, for groups that arise as the image of such fields.
Clearly, there is a natural pairing of Z-modules between Ω k (X, G) and Ck (X, Z):

., . : Ω k (X, G) × Ck (X, Z) → G, (ω, c) → ω(c).

This pairing is degenerate if and only if all elements of G have bounded order. In
particular, if G is a field F of characteristic zero, ., . yields a group homomorphism

Fk : Ck (X, Z) → HomF (Ω k (X, F), F) = (Ω k (X, F))∗ .

A basis of Ck (X, Z) is mapped under Fk to a basis of (Ω k (X, F))∗ and hence

Ck (X, Z) appears as an n k -dimensional lattice in (Ω k (X, F))∗ .
Let dk∗ denote the adjoint of the discrete exterior derivative dk with respect to the
natural pairing between Ω k (X, F) and (Ω k (X, F))∗ . Clearly,

dk∗ ◦ Fk = Fk ◦ ∂k+1 .

Now, since the simplicial complex is finite, we can choose bases of Ck (X, Z) for all
k. This in turn yields bases of (Ω k (X, F))∗ and hence, by duality, bases of Ω k (X, F).
With respect to these bases we have

Ck (X, Z) = Zn k ⊂ Fn k = (Ω k (X, F))∗ = Ω k (X, F), (3)

where n k denotes the number of k-simplices. Moreover, the pairing is represented

∗
by the standard product. The operator dk−1 = ∂k is then just an integer matrix and


∂k = dk−1 .
Complex Line Bundles Over Simplicial Complexes . . . 215

Fig. 6 With the

identifications (3), the space
of k-forms becomes a direct
sum of the image of dk−1
and the kernel of its adjoint
∗ , the latter of which
dk−1
contains the closed k-chains
as a lattice

∗
We have im dk−1 ⊥ ker dk−1 . Moreover, by the rank-nullity theorem,
∗ ∗ ∗
n k = dim im dk−1 + dim ker dk−1 = dim im dk−1 + dim ker dk−1 .

∗
Hence, under the identifications above, we have that Fn k = im dk−1 ⊕⊥ ker dk−1 (see
∗
Fig. 6). Moreover, ker ∂k contains a basis of ker dk−1 . From this we conclude imme-
diately the following lemma.

Lemma 7.4 Let ω ∈ Ω k (X, F), where F is a field of characteristic zero. Then

ω ∈ im dk−1 ⇐⇒ ω, c = 0 for all c ∈ ker ∂k .

Remark 7.5 Note, that for boundary cycles the condition is nothing but the closed-
ness of the form ω. Thus Lemma 7.4 states that a closed form ω ∈ Ω k (X, F) is exact
if and only if the integral over all homology classes [c] ∈ Hk (X, Z) vanishes.

Let G be an abelian group. The sequence below will be referred to as the k-th
fundamental sequence of forms with coefficients in G:

dk−1 k
Ω k−1 (X, G) −−→ Ω k (X, G) −→ Hom(ker ∂k , G) → 0,

where k denotes the restriction to the kernel of ∂k , i.e. k (ω) := ω|ker ∂k .

Combining Lemmas 6.5 and 7.4 we obtain that the fundamental sequence with
coefficients in a field F of characteristic zero is exact for all k > 1. This serves as an
anchor point. The exactness propagates under surjective group homomorphisms.
f
Lemma 7.6 Let A − → B → 0 be a an exact sequence of abelian groups. Then, if
the k-th fundamental sequence of forms is exact with coefficients in A, so it is with
coefficients in B.
216 F. Knöppel and U. Pinkall

Proof By Lemma 6.5 the restriction map k is surjective for every abelian group.
It is left to check that ker k = im dk−1 with coefficients in B. Let Ω ∈ Ω k (X, B)
such that k (Ω) = 0. Since f : A → B is surjective, there is a form  ∈ Ω k (X, A)
such that Ω = f ◦ . Since 0 = k (Ω) = f ◦ k (), we obtain that k () takes
its values in ker f . Since k is surjective for arbitrary groups, there is  ∈ Ω k (X,
ker f ) such that k () = k (). Hence k ( − ) = 0. Thus there is a form ξ ∈
Ω k−1 (X, A) such that dk−1 ξ =  − . Now, let ω := f ◦ ξ ∈ Ω k−1 (X, B). Then

dk−1 ω = dk−1 f ◦ ξ = f ◦ dk−1 ξ = f ◦ ( − ) = f ◦  = Ω.

Hence ker k = im dk−1 and the sequence (with coefficients in B) is exact. 

Remark 7.7 The map f : C → C, z → exp(2πı z) provides a surjective group

homomorphism from C onto C∗ , and similarly from R onto S. Hence the k-th fun-
damental sequence of forms is exact for coefficients in C∗ and in the unit circle S.

Remark 7.8 The k-th fundamental sequence with coefficients in an abelian group
G is exact if and only if Ω k (X, G)/dΩ k−1 (X, G) ∼
= Hom(ker ∂k , G). The isomor-
phism is induced by the restriction map k .

The following corollary is a consequence of Remark 7.7. It nicely displays the

fibration of the complex line bundles by their C∗ -curvature.

Corollary 7.9 For G = S, C∗ the following sequence is exact:

d
1 → H1 (X, G) → Ω 1 (X, G)/dΩ 0 (X, G) −
→ Ω 2 (X, G) → Hom(ker ∂2 , G) → 1.

Definition 7.10 Let Ω ∗ ∈ Ω k (X, S). A real-valued form Ω ∈ Ω 2 (X, R) is called

compatible with Ω ∗ if Ω ∗ = exp ıΩ . A discrete hermitian line bundle with cur-
vature is a discrete hermitian line bundle L with connection equipped with a closed
2-form compatible with the S-curvature of L.

For real-valued forms it is common to denote the natural pairing with the k-chains
by an integral sign, i.e. for ω ∈ Ω k (X, R) and c ∈ Ck (X, Z) we write

ω := ω, c = ω(c).
c

Theorem 7.11 Let L be a discrete hermitian line bundle with curvature Ω. Then
Ω is integral, i.e.

Ω ∈ 2π Z, for all c ∈ ker ∂2 .

c
Complex Line Bundles Over Simplicial Complexes . . . 217
 
Proof By definition the curvature form Ω satisfies exp ıΩ = dω for some con-
nection form ω ∈ Ω 1 (X, S). Thus, if c ∈ ker ∂2 ,

 
exp ı Ω = exp(ıΩ), c = dω, c = ω, ∂c = 1.
c

This proves the claim. 

Conversely, Corollary 7.9 yields a discrete version of a theorem of André Weil

[11, 18], which states that any closed smooth integral 2-form on a manifold M can
be realized as the curvature of a hermitian line bundle. This plays a prominent role
in the process of prequantization [17].

Theorem 7.12 If Ω ∈ Ω 2 (X, R) is integral, then there exists a hermitian line bun-
dle with curvature Ω.

Proof Consider Ω ∗ := exp(iΩ). Since Ω is integral, Ω ∗ , c = 1 for all c ∈ ker ∂2 .

By Corollary 7.9, there exists r ∈ Ω 1 (X, S) such that dr = Ω ∗ . This in turn defines
a hermitian line bundle with curvature Ω. 

Remark 7.13 Moreover, Corollary 7.9 shows that the connections of two such bun-
dles differ by an element of H1 (X, S). Thus the space of discrete hermitian line
bundles with fixed curvature Ω can be parameterized by H1 (X, S).

8 The Index Formula for Hermitian Line Bundles

Before we define the degree of a discrete hermitian line bundle with curvature or
the index form of a section, let us first recall the situation in the smooth setting.
Therefore, let L → M be a smooth hermitian line bundle with connection. Since
the curvature tensor R ∇ of ∇ is a 2-form taking values in the skew-symmetric
endomorphisms of L, it is completely described by a closed real-valued 2-form
Ω ∈ Ω 2 (M, R),

R ∇ = −ıΩ.

The following theorem shows an interesting relation between the index sum of a
section ψ ∈ Γ (L), the curvature 2-form Ω, and the rotation form ξ ψ of ψ. This
form is defined as follows:
∇ψ, ıψ
ξ ψ := .
ψ, ψ
218 F. Knöppel and U. Pinkall

Theorem 8.1 Let L → M be a smooth hermitian line bundle with connection and
Ω its curvature 2-form. Let ψ ∈ Γ (L) be a section with a discrete zero set Z . Then,
if C is a finite smooth 2-chain such that ∂C ∩ Z = ∅,


2π indψp = ξψ + Ω.
p∈C ∩ Z ∂C C

Proof We can assume that C is a single smooth triangle. Then we can express ψ on
C in terms of a complex-valued function z and a pointwise-normalized local section
φ, i.e. ψ = z φ. Since Im( dzz ) = d arg(z), we obtain

1 dz
ξψ = dz φ + z ∇φ, ı z φ = φ, ıφ + ∇φ, ıφ = d arg(z) + ∇φ, ıφ.
|z|2 z

Moreover, away from zeros, we have

d ∇φ, ıφ = R ∇ φ, ıφ + ∇φ ∧ ı∇φ = R ∇ φ, ıφ = −Ω.

Hence we obtain





ξψ = d arg(z) + ∇φ, ıφ = 2π index p (ψ) − Ω.

∂C ∂C ∂C p∈C ∩ Z C

This proves the claim. 

In the case that L is a hermitian line bundle

 with connection over a closed ori-
ented surface M, Theorem 8.1 tells us that M Ω ∈ 2πZ. This yields a well-known
topological invariant—the degree of L:

  1
deg L := Ω.
2π M

From Theorem 8.1 we immediately obtain the famous Poincaré-Hopf index theo-
rem.

Theorem 8.2 Let L → M be a smooth hermitian line bundle over a closed oriented
surface. Then, if ψ ∈ Γ (L) is a section with isolated zeros,
  
deg L = indψp .
p∈M

Now, let us consider the discrete case. In general, a section of a discrete vec-
tor bundle E → X with vertex set V is a map ψ : V → E such that the following
diagram commutes
Complex Line Bundles Over Simplicial Complexes . . . 219

i.e. π ◦ ψ = id. As in the smooth case, the space of sections of E is denoted by

Γ (E).
Now, let L → X be a discrete hermitian line bundle with curvature Ω and let
ψ ∈ Γ (L) be a nowhere-vanishing section such that

ηij (ψi ) = −ψ j (4)

for each edge ij of X. Here η denotes the connection of L as usual. The rotation
form ξ ψ of ψ is then defined as follows:

ψ  ψj 
ξij := arg ∈ (−π, π).
ηij (ψi )

Remark 8.3 Equation (4) can be interpreted as the condition that no zero lies in the
1-skeleton of X (compare Sect. 11). Actually, given a consistent choice of the argu-
ment on each oriented edge, we could drop this condition. Figuratively speaking, if
a section has a zero in the 1-skeleton, then we decide whether we push it to the left
or the right face of the edge.

Now we can define the index form of a discrete section:

Definition 8.4 Let L → X be a discrete hermitian line bundle with curvature Ω.

For ψ ∈ Γ (L), we define the index form of ψ by

1  ψ 
indψ := dξ + Ω .
2π
Theorem 8.5 The index form of a nowhere-vanishing discrete section is Z-valued.

Proof Let L be a discrete hermitian line bundle with curvature and let η be its con-
nection. Let ψ ∈ Γ (L) be a nowhere-vanishing section. Now, choose a connection
form ω, i.e. η = ωβ, where β is a trivial connection on L. Then we can write ψ with
respect to a non-vanishing parallel section φ of β, i.e. there is a C-valued function z
ψ  z 
such that ψ = zφ. Then ξij = arg ωijjzi and thus

 ψ    zi   zj   z k  1
exp 2πı dξijk = exp ı arg + ı arg + ı arg = .
ωki z k ωij z i ω jk z j dωijk
220 F. Knöppel and U. Pinkall

Thus  
 ψ  exp ıΩijk
exp 2πı indijk = = 1.
dωijk

This proves the claim. 

If L is a discrete hermitian line bundle with curvature Ω over a closed oriented

surface X, then we define the degree of L just as in the smooth case:

  1
deg L := Ω.
2π X

Here we have identified X by the corresponding closed 2-chain. From Theorem 7.11
we obtain the following corollary.

Corollary 8.6 The degree of a discrete hermitian line bundle with curvature is an
integer:
 
deg L ∈ Z.

The discrete Poincaré-Hopf index theorem follows easily from the definitions:

Theorem 8.7 Let L → X be a discrete hermitian line bundle with curvature Ω over
an oriented simplicial surface. If ψ ∈ Γ (L) is a non-vanishing discrete section, then
   ψ
deg L = indijk .
ijk∈X

Proof Since the integral of an exact form over a closed oriented surface vanishes,



  ψ
2π deg L = Ω= dξ ψ + Ω = 2π indijk ,
X X ijk∈X

as was claimed. 

9 Piecewise-Smooth Vector Bundles over Simplicial

Complexes

It is well-known that each abstract simplicial complex X has a geometric realization

which is unique up to simplicial isomorphism. In particular, each abstract simplex
is realized as an affine simplex. Moreover, each face σ of a simplex σ ∈ X comes
with an affine embedding
ισ σ : σ → σ.
Complex Line Bundles Over Simplicial Complexes . . . 221

In the following, we will not distinguish between the abstract simplicial complex
and its geometric realization.

Definition 9.1 A piecewise-smooth vector bundle E over a simplicial complex X is

a topological vector bundle π : E → X such that
(a) for each σ ∈ X the restriction Eσ := E|σ is a smooth vector bundle over σ,
(b) for each face σ of σ ∈ X, the inclusion Eσ → Eσ is a smooth embedding.

As a simplicial complex, X has no tangent bundle. Nonetheless, differential

forms survive as collections of smooth differential forms defined on the simplices
which are compatible in the sense that they agree on common faces:

Definition 9.2 Let E be a piecewise-smooth vector bundle over X. An E-valued

differential k-form is a collection ω = {ωσ ∈ Ω k (σ, Eσ )}σ∈X such that for each face
σ of a simplex σ ∈ X the following relation holds:

ι∗σ σ ωσ = ωσ ,

where ισ σ : σ → σ denotes the inclusion. The space of E-valued differential k-

forms is denoted by Ω kps (X, E).

Remark 9.3 Note that a 0-form defines a continuous map on the simplicial complex.
Hence the definition includes functions and, more generally, sections: A piecewise-
smooth section of E is a continuous section ψ : X → E such that for each simplex
σ ∈ X the restriction ψσ := ψ|σ : σ → Eσ is smooth, i.e.
 
Γ ps (E) := ψ : X → E | ψσ ∈ Γ (Eσ ) for all σ ∈ X .

Since the pullback commutes with the wedge-product ∧ and the exterior deriva-
tive d of real-valued forms, we can define the wedge product and the exterior deriv-
ative of piecewise-smooth differential forms by applying it componentwise.

Definition 9.4 For ω = {ωσ }σ∈X ∈ Ω kps (X, R), η = {ησ }σ∈X ∈ Ω ps (X, R),

ω ∧ η := {ωσ ∧ ησ }σ∈X , dω := {dωσ }σ∈X .

All the standard properties of ∧ and d also hold in the piecewise-smooth case.

Definition 9.5 A connection on a piecewise-smooth vector bundle E over X is a

linear map ∇ : Γ ps (E) → Ω 1ps (X, E) such that

∇( f ψ) = d f ψ + f ∇ψ, for all f ∈ Ω 0ps (X, R), ψ ∈ Γ ps (E).

Once we have a connection on a smooth vector bundle we obtain a corresponding

exterior derivative d ∇ on E-valued forms.
222 F. Knöppel and U. Pinkall

Theorem 9.6 Let E be a piecewise-smooth vector bundle over X. Then there is

a unique linear map d ∇ : Ω kps (X, E) → Ω ps
k+1
(X, E) such that d ∇ ψ = ∇ψ for all
ψ ∈ Γ ps (E) and

d ∇ (ω ∧ η) = dω ∧ η + (−1)k ω ∧ d ∇ η

for all ω ∈ Ω kps (X, R) and η ∈ Ω ps (X, E).

The curvature tensor survives as a piecewise-smooth End(E)-valued 2-form:

Definition 9.7 Let E → X be a piecewise-smooth vector bundle. The endomor-

phism-valued curvature 2-form of a connection ∇ on E is defined as follows:

d ∇ ◦ d ∇ ∈ Ω 2ps (X, End(E)).

10 The Associated Piecewise-Smooth Hermitian

Line Bundle

Let L̃ → X be a piecewise-smooth hermitian line bundle with connection ∇ over a

simplicial complex. Just as in the smooth case the endomorphism-valued curvature
2-form takes values in the skew-adjoint endomorphisms and hence is given by a
piecewise-smooth real-valued 2-form Ω̃:

d ∇ ◦ d ∇ = −ı Ω̃.

Since each simplex of X has an affine structure, we can speak of piecewise-constant

forms.
The goal of this section will be to construct for each discrete hermitian line
bundle with curvature a piecewise-smooth hermitian line bundle with piecewise-
constant curvature which in a certain sense naturally contains the discrete bundle.
We first prove two lemmata.

Lemma 10.1 To each closed discrete real-valued k-form ω there corresponds a

unique piecewise-constant piecewise-smooth k-form ω̃ such that

ω(c) = ω̃, for all c ∈ Ck (X, Z).

c

The form ω̃ will be called the piecewise-smooth form associated to ω.

Proof It is enough to consider just a single n-simplex σ. We denote the space of

piecewise-constant piecewise-smooth k-forms on σ by Ωck and the space of discrete
Complex Line Bundles Over Simplicial Complexes . . . 223

k-forms on σ by Ωdk . Consider the linear map F : Ωck → Ωdk that assigns to ω̃ ∈ Ωck
the discrete k-form given by

F(ω̃)σ := ω̃.
σ

Clearly, F is injective. Moreover, since each piecewise-constant piecewise-smooth

form is closed, we have im F ⊂ ker dk , where dk denotes the discrete exterior deriv-
  on σ is of
ative. Hence it is enough to show that the space of closed discrete k-forms
dimension nk . This we do by induction. Clearly, dim ker d0 = 1 = n0 . Now sup-
 n 
pose that dim ker dk−1 = k−1 . By Lemma 7.4, we have ker dk = im dk−1 . Hence
n+1   n 
dim ker dk = dim im dk−1 = dim Ωdk − dim ker dk−1 = k
− n
k−1
= k
.

Therefore, for each closed discrete k-form we obtain a unique piecewise-constant

piecewise-smooth k-form which has the desired integrals on the k-simplices. 
It is a classical result that on star-shaped domains U ⊂ R N each closed form is
exact: If Ω ∈ Ω k (U, R) is closed, then there exists a form ω ∈ Ω k−1 (U, R) such
that Ω = dω. Moreover, the potential can be constructed explicitly by the map K :
Ω k (U, R) → Ω k−1 (U, R) given by

 
k 
1 
K (Ω) = (−1)α−1 t k−1 Ωi 1 ···i k (t x)dt xi α d xi 1 ∧ . . . ∧ d
xi α ∧ . . . ∧ d xi k ,
i 1 <···<i k α=1 0

where Ω = i 1 <···<i k Ωi1 ···ik d xi1 ∧ . . . ∧ d xik . One directly can check that

K (dΩ) + d K (Ω) = Ω.

Hence, if dΩ = 0, we get Ω = d K (Ω). The same construction works for

piecewise-smooth forms defined on the star of a simplex. This yields the follow-
ing piecewise-smooth version of the Poincaré-Lemma.

Lemma 10.2 On the star of a simplex each closed piecewise-smooth form is exact.

Now we are ready to prove the main result of this section.

Theorem 10.3 Let L → X be a discrete hermitian line bundle with curvature Ω

over a simplicial complex and let Ω̃ be the piecewise-smooth 2-form associated to
Ω. Then there is a piecewise-smooth hermitian line bundle L̃ → X with connection
∇˜ of curvature Ω̃ such that L̃i = Li for each vertex i and the parallel transports
coincide along each edge path. The bundle L̃ is unique up to isomorphism.

Proof First we construct the piecewise-smooth hermitian line bundle. Let L → X

be a discrete hermitian line bundle with curvature Ω and let η denote its connection.
Let V be the vertex set of X and let Si denote the open vertex star of the vertex i.
224 F. Knöppel and U. Pinkall

Further, since Ω is closed, by Lemma 10.1, there is a piecewise-constant piecewise-

smooth form Ω̃ associated to Ω. Now, consider the set

L̂ := ! Si × Li .
i∈V

Note, that Si ∩ S j = ∅ if and only if ij is an edge of X or i = j. Thus, if we set

ηii := id|Li , we can define an equivalence relation on L̂ as follows:

 
(i, p, u) ∼ ( j, q, v) :⇐⇒ p = q and v = exp −ı Ω̃ ηij (u),
p
ij

p
where ij denotes the oriented triangle spanned by the point i, j and p. Note here
p
that ij is completely contained in some simplex of X. Let us check shortly that
this really defines an equivalence relation. Here the only non-trivial property is tran-
sitivity. Therefore, let (i, p, u) ∼ ( j, q, v) and ( j, q, v) ∼ (k, r, w). Thus we have
p = q = r and p lies in a simplex which contains the oriented triangle ijk. Clearly,
p p p
the 2-chain ij + jk + ki is homologous to ijk and since piecewise-constant
forms are closed we get

Ω̃ = − Ω̃ + Ω̃ = Ω̃ + Ωijk .
p p p p
ij + jk ki ijk ik

Hence we obtain

     
w = exp −ı Ω̃ η jk exp −ı Ω̃ ηij (u)
p p
jk ij

 
= exp −ı Ω̃ η jk ◦ ηij (u)
p p
ij + jk

 
= exp −ı Ω̃ − ıΩijk η jk ◦ ηij (u)
p
ik

 
= exp −ı Ω̃ ηik (u),
p
ik

and thus (i, p, u) ∼ (k, r, w). Hence ∼ defines an equivalence relation. One can
check now that the quotient L̃ := L̂/∼ is a piecewise-smooth line bundle over X. The
local trivializations are then basically given by the inclusions Si × Li → L̃ sending
a point to the corresponding equivalence class. Moreover, all transition maps are
unitary so that the hermitian
 metric of L extends to L̃ and turns L̃ into a hermitian

line bundle. Clearly, L̃ = L.
V
Next, we need to construct the connection. Therefore we will use an explicit
system of local sections: Choose for each vertex i ∈ V a unit vector X i ∈ Li and
define φi ( p) := [i, p, X i ]. This yields for each vertex i a piecewise-smooth section
Complex Line Bundles Over Simplicial Complexes . . . 225

φi define on the star Si . For each non-empty intersection Si ∩ S j = ∅ we then obtain

a function gij : Si ∩ S j → S. By the above construction, we find that, if ηij (X i ) =
rij X j ,

 
gij ( p) = rij exp −ı Ω̃ . (5)
p
ij

Since Ω̃ is closed, Lemma 10.2 tells us that Ω̃| Si is exact. Hence there is a piecewise-
smooth 1-form ωi defined on Si such that dωi = Ω̃| Si . In general, the form ωi is only
unique up to addition of an exact 1-form, but among those there is a unique form ωi
which is zero along the radial directions originating from i. To see this, just choose
some potential ω̃i of Ω| Si and
 define a function f : Si → R as follows:
p
For p ∈ Si , let f ( p) := γ p ω̃i , where γi denote the linear path from the vertex
i
i to the point p. Then ωi := ω̃i − d f is a piecewise-smooth potential of Ω| Si and
vanishes on radial directions. For the uniqueness, let ω̂i be another such potential.
Then, the difference ωi − ω̂i is closed and hence exact on Si , i.e. there is f : Si → R
such that d f = ωi − ω̂i . Since d f vanishes on radial directions f is constant on
radial lines starting at i and hence constant on Si . Thus ωi = ω̂i .
Suppose that for each edge ij the forms ωi and ω j are compatible, i.e., wherever
both are defined,

ıω j = ıωi + d log gij .

 
Then we can define a connection ∇ as follows: Let ψ ∈ Γ L̃ and let X ∈ T p σ
for some simplex σ of X, then there is some Si  p. On Si we can express ψ with
respect to φi , i.e. ψ = z φi for some piecewise-smooth function z : Si → C. Then
we define
 
∇ X ψ := dz(X ) − ıωi (X )z φi . (6)

In general there are several stars that contain the point p. From compatibility easily
follows that the definition does not depend on the choice of the vertex. Hence we
have constructed a piecewise smooth connection ∇. One easily checks that ∇ is
unitary and since dωi = Ω̃| Si we get d ∇ ◦ d ∇ = −ı Ω̃ as desired.
So it is left to check the compatibility of the forms ωij constructed above. Let ij
be some edge and let p0 be a point in  its interior. Since ωi − ω j is closed, we can
define ϕ : Si ∩ S j → R by ϕ( p) := γ p ωi − ω j , where γ p is some path in Si ∩ S j
from the point p0 to the point p. Then, for p ∈ Si ∩ S j ,

Ω= ωj = ωj = − ωj = ωi − ω j = ϕ( p),
p p p p
p ∂ p ij+γ j −γi γi γi

p p
where as above γi denotes the linear path from i to p and, similarly, γ j denotes
the linear path from j to the point p. From this we obtain
226 F. Knöppel and U. Pinkall

ωi − ω j = dϕ = d Ω
p

and in particular ıω j = ıωi + d log gij . This shows the existence.

Now suppose there are two such piecewise-smooth bundles L̃ and L̂ with con-
nection ∇˜ and ∇, ˆ respectively. We want to construct an isomorphism between L̃ and
L̂. Therefore we again use local systems. Explicitly, we choose a discrete direction
field X ∈ L. This yields for each vertex i a vector X i ∈ L̃i = L̂i which extends by
parallel transport along rays starting at i to a local sections φ̃i of L̃ and, similarly, to
a local section φ̂i of L̂ defined on Si .
Now we define F : L̃ → L̂ to be unique map which is linear on the fibers and
satisfies F(φ̃i ) = φ̂i on Si . To see that F is well-defined, we need to check that
it is compatible with the transition maps. But by construction both systems have
equal transition maps, namely the the functions gij from Eq. (5) with rij given by
ηij (X i ) = rij X j . Now, if z i φ̃i = z j φ̃ j , then z i = z j gij and hence

F(z i φ̃i ) = z i φ̂i = z i gij φ̂ j = z j φ̂ j = F(z j φ̃ j ).

Using Eq. (6) one similarly shows that F ◦ ∇˜ = ∇ˆ ◦ F. Thus L̃ ∼

= L̂. 

11 Finite Elements for Hermitian Line Bundles

with Curvature

In this section we want to present a specific finite element space on the associated
piecewise-smooth hermitian line bundle of a discrete hermitian line with curvature.
They are constructed from the local systems that played such a prominent role in
the proof of Theorem 10.3 and the usual piecewise-linear hat function.
Let L̃ be the associated piecewise-smooth bundle of a discrete hermitian line
bundle L → X and let xi : X → R denote the barycentric coordinate of the vertex
i ∈ V, i.e. the unique piecewise-linear function such that xi ( j) = δij , where δ is the
Kronecker delta. Clearly,

Γ (L) = Li .
i∈V

To each X ∈ Li we now construct a piecewise-smooth section ψ̃ as follows: First,

we extend X to the vertex star Si of the vertex i using the parallel transport along
rays starting at i. To get a global section ψ̃ ∈ Γ ps (L) we use xi to scale φ̃ down to
zero on ∂ Si and extend it by zero to X, i.e.

xi ( p)φ̃ p for p ∈ Si ,
ψ̃ p :=
0 else.
Complex Line Bundles Over Simplicial Complexes . . . 227

The above construction yields a linear map ι : Γ (L) → Γ ps (L̃). Clearly, ι is

injective—a left-inverse is just given by the restriction map

Γ ps (L̃)  ψ̃ → ψ̃ V ∈ Γ (L).

Definition 11.1 The space of piecewise-linear sections is given by Γ pl (L̃) := im ι.

Thus we identified each section of a discrete hermitian line bundle with curva-
ture with a piecewise-linear section of the associated piecewise-smooth bundle. This
allows to define a discrete hermitian inner product and a discrete Dirichlet energy on
Γ (L), which is a generalization of the well-known cotangent Laplace operator for
discrete functions on triangulated surfaces. Before we come to the Dirichlet energy,
we define Euclidean simplicial complexes.
Similarly to piecewise-smooth forms we can define piecewise-smooth (con-
travariant) k-tensors as collections of compatible k-tensors: A piecewise-smooth k-
tensor is a collection T = {Tσ }σ∈X of smooth contravariant k-tensors Tσ on σ such
that

ι∗σ σ Tσ = Tσ ,

whenever σ is a face of σ. A Riemannian simplicial complex is then a simplicial

complex X equipped with a piecewise-smooth Riemannian metric, i.e. a piecewise-
smooth positive-definite symmetric 2-tensor g on X.
The following lemma tells us that the space of constant piecewise-smooth sym-
metric tensors is isomorphic to functions on 1-simplices.
Lemma 11.2 Let X be a simplicial complex and let E denote the set of its 1-
simplices. For each function f : E → R there exists a unique constant piecewise-
smooth symmetric 2-tensor S such that for each 1-simplex e = {i, j}

Se ( j − i, j − i) = f (e).

Proof It is enough to consider a single affine n-simplex σ = {i 0 , . . . , i n } with vector

space V . Consider the map F that sends a symmetric 2-tensor S on V to the function
given by

F(S)(e) := S(i k − i j , i k − i j ), e = {i j , i k } ⊂ σ.

Clearly, F is linear. Moreover, if Q denotes the quadratic form corresponding to S,

i.e. Q(X ) := S(X, X ), then
 
S(X, Y ) = 1
2
Q(X ) + Q(Y ) − Q(X − Y ) .

Hence, from F(S) = 0 follows S = 0. Thus F is injective. Clearly, the space of

symmetric bilinear forms is of dimension n(n + 1)/2, which equals the number of
1-simplices. Thus F is an isomorphism. This proves the claim. 
228 F. Knöppel and U. Pinkall

It is also easy to write down the corresponding symmetric tensor in coordinates:

Let σ = {i 0 , . . . , i n } be a simplex. The vectors e j := i j − i 0 , j = 1, . . . , n, then
yield a basis of the corresponding vector space. Let f be a function defined on the
unoriented edges of σ and let xi j denote the barycentric coordinates of its vertices
i j , then the corresponding symmetric bilinear form Sσf is given by
   
Sσf = f i0 i j d xi j ⊗ d xi j + 1
2
f i0 i j + f i0 ik − f i j ik d xi j ⊗ d xik . (7)
1≤ j≤n 1≤ j,k≤n, j=k

Thus starting with a positive function f , by Sylvester’s criterion, it has to satisfy on

each n-simplex n − 1 inequalities to determine a positive-definite form. If the cor-
responding piecewise-smooth form is positive-definite, we call f a discrete metric.

Definition 11.3 A Euclidean simplicial complex is a simplicial complex X equipped

with a discrete metric, i.e. a map  that assigns to each 1-simplex e a length e > 0
such that for each simplex σ the symmetric tensor Sσ is positive-definite.

Now, let X be a Euclidean simplicial manifold of dimension n and denote by Xn

the set of its top-dimensional simplices. Since each simplex of X is equipped with
a scalar product it comes with a corresponding density and hence we know how to
integrate functions over the simplices of X. Now, we define the integral over X as
follows:



f := f σ , f ∈ Ω 0ps (X, R).

X σ∈Xn σ

Moreover, given a piecewise-smooth hermitian line bundle L̃ → X with curvature,

then there is a canonical hermitian product ., . on Γ ps (L̃): If ψ̃, φ̃ ∈ Γ ps (L̃), then

ψ̃, φ̃ = ψ̃, φ̃.

X

In particular, if L̃ is the associated piecewise-smooth bundle of a discrete her-

mitian line bundle L with curvature Ω, then we can use ι to pull ., . back to
Γ (L). Since ι is injective this yields a hermitian product on Γ (L).
Now we want to compute this metric explicitly in terms of given discrete data.

Definition 11.4 A piecewise-linear section ψ̃ ∈ Γ pl (L̃) is called concentrated at a

vertex i, if it is of the form ψ̃ = ι(ψi ) for some vector ψi ∈ Li .

It is basically enough to compute the product of two such concentrated sections.

Therefore, let ψi ∈ Li and ψ j ∈ L j and let ψ̃ i and ψ̃ j denote the corresponding
piecewise-linear concentrated sections.
Complex Line Bundles Over Simplicial Complexes . . . 229

Now consider their product ψ̃ i , ψ̃ j . Clearly, this product has support Si ∩ S j .

For simplicity, we extend the discrete connection η to arbitrary pairs ij in such way
that ηii = id and ηij : Li → L j is zero whenever {i, j} ∈ / X. With this convention,
Eq. (5) yields

 
ψ̃ , ψ̃  = ψ j , ηij (ψi ) xi x j exp −ı
j i
Ω̃ , (8)
p
ij

where Ω̃ denotes the constant piecewise-smooth curvature form associated to Ω.

p
Now, let us express the integral over ij on a given n-simplex. Therefore consider
an n-simplex σ = {i 0 , . . . , i n }. The hat functions xi1 , . . . , xin yield affine coordinates
on σ and we can express any 2-form with respect to the basis forms d xi j ∧ d xik . One
can show that


± 21 for σ = ±i j i k i  ,
d xi j ∧ d xik =
σ 0 else.

Thus we obtain

Ω̃ = 2 Ωi0 i j ik d xi j ∧ d xik .
1≤ j<k≤n

p
Now we want to compute the integral over the triangle i0 i1 ⊂ σ. By Stokes
theorem,


i1
p
i0
d xi j ∧ d xik = xi j d xik + xi j d xik + xi j d xik ,
p
i i0 i1 p
0 i1

where the integrals are computed along straight lines. A small computation shows

1 
d xi j ∧ d xik = δ1j xik ( p) − δ1k xi j ( p) ,
p
i i 2
0 1


Thus, for j < k, we get i
p d xi j ∧ d xik = 21 δ1j xik ( p) and hence
0 i1



Ω̃ = 2 Ωi0 i j ik d xi j ∧ d xik = Ωi0 i1 i j xi j ( p),
p p
i i 1≤ j<k≤n i i j
0 1 0 1

where we have used the convention that Ω vanishes on all triples not representing
an oriented 2-simplex of X. With this convention Eq. (8) becomes
  
ψ̃ j , ψ̃ i  = ψ j , ηij (ψi ) xi x j exp −ı Ωi jk xk . (9)
k
230 F. Knöppel and U. Pinkall

Fig. 7 The graph of the norm of a piecewise-linear section of a bundle over a torus consisting of
two triangles. Its two smooth parts fit continuously together along the diagonal. In this example the
curvature of the bundle over each triangles is equal to 4π. Note that the section has 4 zeros—just
as predicted

In particular, using Eq. (9), we can compute the norm of a piecewise-linear

section ψ̃ on a given triangle ijk. Therefore we distinguish one of its vertices, say
i, and write ψ̃ with respect to a section which is radially parallel with respect to i.
Now, one checks that

|ψ̃| = |ci + x j (c j eiΩijk xk − ci ) + xk (ck e−iΩijk x j − ci )|,

where ci , c j , ck ∈ C are constants depending on the explicit form of ψ̃. An example

of the norm of a piecewise-linear section is shown in Fig. 7.
As the next proposition shows, the identification of discrete and piecewise-linear
sections perfectly fits together with the definitions in Sect. 8.
Proposition 11.5 Let ψ ∈ Γ (L) be a discrete section and let ψ̃ ∈ Γ pl (L̃) be the
corresponding piecewise-linear section, i.e. ψ̃ = ι(ψ). Then, if ψ̃ has no zeros on
edges, the discrete rotation form ξ ψ and the piecewise-smooth rotation form ξ ψ̃ are
related as follows: For each oriented edge ij,

ψ
ξij = ξ ψ̃ .
ij

Proof The claim follows easily by expressing ψ̃ with respect to some non-vanishing
parallel section along the edge ij. 
In particular, by Theorem 8.1, the index form of a non-vanishing section of a
discrete hermitian line bundle with curvature counts the number of (signed) zeros
of the corresponding piecewise-linear section of the associated piecewise-smooth
bundle.
Complex Line Bundles Over Simplicial Complexes . . . 231

Let us continue with the computation of the metric on Γ (L). To write down the
formula we give the following definition.
Definition 11.6 Let X be an n-dimensional simplicial manifold and let Ω ∈ Ω 2
(X, R). To an n-simplex σ and vertices i, j, k, l of X we assign the value

1   
Ω
σ,i, j (k, l) := xk xl exp −ı Ωijm xm ,
vol (σ) σ m

where we have chosen for integration an arbitrary discrete metric on X.

Remark 11.7 Note that the functions Ωσ,i, j are indeed well-defined. On a simplex,
any two such measures induced by a discrete metric differ just by a constant.
With Definition 11.6 and Eq. (9) we obtain the following form of the metric:
Theorem 11.8 (Product of Discrete Sections) Let L be a discrete hermitian line
bundle with curvature Ω over an n-dimensional Euclidean simplicial manifold X,
then the product on Γ (L) induced by the associated piecewise-smooth
hermitian

line bundle is given as follows: Given two discrete sections ψ = i ψi , φ = i φi ,
 
Ω
ij ij
ψ, φ = μΩ ψ j , ηij (φi ), where μΩ = σ,i, j (i, j) vol (σ).
i, j {i, j}⊃σ∈Xn

Note that Ω
ij
σ,i, j (k, l), and hence μΩ , can be computed explicitly using Fubini’s
theorem and the following small lemma, which can be shown by induction.
Lemma 11.9 Let c ∈ C∗ , n ∈ N and [a, b] ⊂ R be an interval. Then

b
n! 
n
(cx)n−k   b
x n exp(cx)d x = (−1)k exp cx  .
a c n+1
k=0
(n − k)! a

Next, we would like to compute the Dirichlet energy of a section ψ̃ ∈ Γ pl (L̃), i.e.

 2
E D (ψ̃) = ∇ ψ̃  .
X

Note, that the Dirichlet energy comes with a corresponding positive-semidefinite

hermitian form ., .D —called the Dirichlet product. Clearly, like the metric, the
Dirichlet product is completely determined by the values it takes on concentrated
sections.
In general, if ψ̃ = 0 is piecewise-linear section concentrated at i, it is given on
the vertex star Si as a product ψ̃ = xi φ̃, where xi denotes the barycentric coordinate
of the vertex i and φ̃ is a local section radially parallel with respect to i. Clearly,

∇ ψ̃ = d xi φ̃ + ı xi ωi φ̃,
232 F. Knöppel and U. Pinkall

where ωi denotes the rotation form of φ̃, i.e. ∇ φ̃ = ıωi φ̃. Note here that ωi does not
depend on the actual value of ψ̃ at i, but is the same for all non-vanishing piecewise-
linear sections concentrated at i.
To compute the rotation form ωi at a given point p0 ∈ Si , we use a local section
ζ which is radially parallel with respect to p0 such that ζ p0 = φ̃ p0 . Then we can
express φ̃ in terms of ζ, i.e.

φ̃ = z ζ,

for some piecewise-smooth C∗ -valued function z defined locally at p0 . Clearly, |z|

is constant, and hence


ıωi | p0 φ̃ p0 = ∇ φ̃ = dz| p0 ζ p0 + z( p0 ) ∇ζ| p0 = d log z| p0 φ̃ p0 = ıd arg z| p0 φ̃ p0 .
p0

The clue is that we can now use the relation of parallel transport and curvature to
obtain an explicit formula for z. If p is sufficiently close to p0 , then the three points
p, i and p0 determine an oriented triangle p which is contained in a simplex of X.
Its boundary curve γ p consists of three line segments γ1 , γ2 , γ3 connecting p to i,
i to p0 and p0 back to p. Hence on each of these segments either φ̃ of ζ is parallel
and

 
ζ p = Pγ p (φ̃ p ) = exp ı Ω̃ φ̃ p .
p

  
Thus we obtain that z( p) = exp −ı p Ω̃ and hence


  
ωi  = −d Ω̃  .
p0 p p0

Now, if p is contained in a simplex σ = {i 0 , . . . , i n }, one verifies that

1 
d xi j ∧ d xik = xi ( p0 )xik ( p) − xik ( p0 )xi j ( p) .
p 2 j

Thus,

    
d Ω̃  = 2 Ωi0 i j ik d d xi j ∧ d xik 
p p0 p p0
1≤ j<k≤n
  
= Ωi0 i j ik xi j d xik − xik d xi j  ,
p0
1≤ j<k≤n
   

= Ωi0 i j ik xik d xi j 
p0
1≤ j≤n k= j
Complex Line Bundles Over Simplicial Complexes . . . 233

and, using the convention on Ω from above, we find the following simple formula:
  

ωi = Ωi jk xk d x j  , (10)
Si
j k

where we sum over the whole vertex set of X.

Now, given this local form expressions, we can finally return to the computa-
tion of the products which we are actually interested in. Therefore we consider two
piecewise-linear sections concentrated at the vertices i and j:

ψ̃ i := ι(ψi ), ψ̃ j := ι(ψ j ),

for some ψi ∈ Li and ψ j ∈ L j . On their common support Si ∩ S j both section can be

expressed, just as above, as products of a real-valued piecewise-linear hat functions
xi and x j and radially parallel local sections φ̃i and φ̃ j :

ψ̃ i = xi φ̃i , ψ̃ j = x j φ̃ j .

Clearly,

ψ̃ j , ψ̃ i D = d x j φ̃ j + ı x j ω j φ̃ j , d xi φ̃i + ı xi ωi φ̃i 

Si ∩ S j

= d x j + ı x j ω j , d xi + ı xi ωi  φ̃ j , φ̃i .
Si ∩ S j

With Eq. (9) we see that

  
φ̃ j , φ̃i  = ψ j , ηij (ψi ) exp −ı Ωi jm xm .
m

Moreover, by Eq. (10),

  
d x j + ı x j ω j , d xi + ı xi ωi  = d x j , d xi  + Ωik l Ω jk l x j xi xl xl d xk , d xk 
k ,k ,l ,l
 
+ı (Ωik l xi xl d x j , d xk  − Ω jk l x j xl d xk , d xi ) .
k ,l

The constants d xk , d xl  are basically provided by the following lemma.

Lemma 11.10 Let σ = {v0 , . . . , vn } be a Euclidean simplex of dimension n > 0
and let xi denote its barycentric coordinate functions. Then

1
gradxi = − Ni ,
hi
234 F. Knöppel and U. Pinkall

where h i denotes the distance between vi and σi = σ \ {vi } and Ni denotes the
outward-pointing unit normal of σi .

Proof This immediately follows from two basic facts: First, d xi (v j − v0 ) = δij for
i, j > 0. Second, h i = v0 − vi , Ni . 

Lemma 11.10 yields almost immediately a higher dimensional analogue of the

well-known cotangent formula for surfaces.

Theorem 11.11 (Cotangent Formula) Let σ be a simplex of a Euclidean simplicial

complex X and let dim σ > 1. If i = j,

  
− n(n−1)
1
cot αijσ vol σ \ {i, j} , if {i, j} ⊂ σ,
cσij := d xi , d x j  =
σ 0 else.

Here αijσ denotes the angle between the faces σ \ {i} and σ \ { j}. Moreover,

  
1
vol σ \ {i} , if i ∈ σ,
cσii := |d xi | =
2 n hi
σ 0 else,

where h i denotes the distance between the vertex i and the face σ \ {i}.

Proof Clearly, if {i, j} ⊂ σ, then σ d xi , d x j  = 0. Now, let {i, j} ⊂ σ, i = j. With
the notation of Lemma 11.10, we have

Ni , N j 
d xi , d x j  = gradxi , gradx j  vol σ = vol σ.
σ hi h j
 
Furthermore, cos αijσ = − Ni , N j  and n! vol σ=(n − 2)! h i h j sin αijσ vol σ \ {i, j} .
This yields the first part of the theorem. Similarly, n vol σ = h i vol σ \ {i} . Setting
then i = j yields the second part. 

Definition 11.12 Let X be an n-dimensional simplicial manifold and let Ω ∈

Ω 2 (X, R). Let σ be an n-simplex and i, j, k, l be vertices of X. Then, let

1   
ΛΩ
σ,i, j := exp −ı Ωijm xm ,
vol (σ) σ m

1   
Ω
σ,i, j (k, l) := xi x j xk xl exp −ı Ωijm xm ,
vol (σ) σ m

where we choose for the integration an arbitrary discrete metric on X.

Remark 11.13 Just like the functions Ω Ω

σ,i, j , the values Λσ,i, j and the functions
Ω
σ,i, j and are well-defined (compare Remark 11.7).
Complex Line Bundles Over Simplicial Complexes . . . 235

Now, with these definitions, we can summarize the above discussion by the fol-
lowing theorem.

Theorem 11.14 (Discrete Dirichlet Energy) Let L be a discrete hermitian line bun-
dle with curvature Ω over an n-dimensional Euclidean simplicial manifold X, then
the Dirichlet product on Γ (L) induced by the associated piecewise-smooth
her-
mitian line bundle is given as follows: If φ = i φi and ψ = i ψi are two discrete
sections,
 ij 
ij Ω
φ, ψD = wΩ φ j , ηij (ψi ), wΩ = Wσ,i, j,
i, j {i, j}⊃σ∈Xn

where
  
Ω Ω Ω
Wσ,i, j = cσ
ij
Λσ,i, j + Ωik l Ω jk l cσ
kk
 σ,i, j (l , l ) (11)
k ,k ,l ,l
 
+ı Ωik l cσjk Ω Ω
σ,i, j (i, l ) − Ω jk l cσ σ,i, j ( j, l ) .
ik

k ,l

12 Discrete Energies on Surfaces—An Example

While the computation of the Dirichlet product ., .D and the metric ., . of dis-
crete sections is quite complicated and tedious for higher dimensional simplicial
manifolds, it is manageable for the 2-dimensional case. We are going to compute it
explicitly.
Throughout this section let L denote a discrete hermitian line bundle with cur-
vature Ω over a Euclidean simplicial surface X and let σ = {i, j, k} be one of its
triangles.
The metric ., . is easily obtained. We basically just need to compute the values
Ω Ω
σ,i,i (i, i) and σ,i, j (i, j), which can be done over the standard triangle. We get

1 exp(−ıΩijk ) − 1 + ıΩijk + 21 Ωijk

2
− ı 16 Ωijk
3
Ω
σ,i,i (i, i) = , Ω
σ,i, j (i, j) = 2 .
6 Ωijk
4

(12)
Now, we compute the Dirichlet product ., .D on X. For n = 2, the expressions
Ω Ω
Wσ,i,i and Wσ,i, j simplify drastically. First, we look at the diagonal terms. We have


cσk k Ωik l Ωik l Ω
σ,i,i (l , l )
k ,k ,l ,l
 
= cσj j Ω
σ,i,i (k, k) − 2cσ
jk Ω
 σ,i,i ( j, k) + cσ
kk Ω
 σ,i,i ( j, j) Ωijk
2
,
236 F. Knöppel and U. Pinkall

and with
1 1
Λσ,i,i = 1, σ,i,i ( j, j) = = σ,i,i (k, k), σ,i,i ( j, k) =
90 180
we get the following formula:

Ω cσj j − cσjk + cσkk 2

Wσ,i,i = cσii + Ωijk .
90
Now we would like to obtain a similar formula for the off-diagonal terms. Since
d xi + d x j = −d xk , we have cσjk + cσki = −cσkk . Hence,

cσk k Ωik l Ω jk l Ω
σ,i, j (l , l )
k ,k ,l ,l
  Ω  2
= − cσij Ω Ω
σ,i, j (k, k) + cσ σ,i, j (i, j) + σ,i, j ( j, k) Ωijk .
kk

This time the expressions become more complicated. We get

2     
Ω
σ,i, j (k, k) = 20 − 12ıΩijk − 3Ωijk
2
+ 13 ıΩijk
3
+ −20 − 8ıΩijk + Ωijk
2
exp −ıΩijk ,
Ωijk
6

2     
Ω Ω
σ,i, j (i, j) + σ,i, j ( j, k) = −6 + 4ıΩijk + Ωijk
2
+ 12 Ωijk
1 4
− 1 5
30 ıΩijk + 6 + 2ıΩijk exp −ıΩijk .
Ωijk
6

Thus,

cσk k Ωik l Ω jk l Ω
σ,i, j (l , l ) =
k ,k ,l ,l
2  kk ij   ij   ij  2 ij kk
3 − cσ Ω 4
4
6cσ − 20cσ + 12cσ − 4cσkk ıΩijk + 3cσ − cσkk Ωijk − c3σ ıΩijk 12 ijk
Ωijk
cσkk 5 + 20cij − 6ckk +8cij − 2ckk ıΩ ij 2   
+ 30 ıΩijk σ σ σ σ ijk − cσ Ωijk exp −ıΩijk .

Now, let us look at the second sum in Eq. (11). We have

 
ı Ωik l cσjk Ω Ω
σ,i, j (i, l ) − Ω jk l cσ σ,i, j ( j, l )
ik

k ,l
 
= cσii Ω jj Ω kk Ω
σ,i, j ( j, k) + cσ σ,i, j (k, i) + cσ σ,i, j (i, j) ıΩijk .

The formula for Ω

σ,i, j (i, j) is already given in Eq. (12). Further, we have

2  
2 + −3 + ıΩ  exp−ıΩ  = Ω (k, i).
Ω
σ,i, j ( j, k) = 4
3 − 2ıΩijk − 21 Ωijk ijk ijk σ,i, j
Ωijk
Complex Line Bundles Over Simplicial Complexes . . . 237

Thus we get
 
Ωik l cσ Ω ik Ω
jk
ı σ,i, j (i, l ) − Ω jk l cσ σ,i, j ( j, l ) =
k ,l
2  jj   jj  2  jj
4
3(cσii + cσ ) − cσkk ıΩijk + 2(cσii + cσ ) − cσkk Ωijk + 21 cσkk − cσii − cσ ıΩijk
3
Ωijk
ckk4 +  kk jj   jj 2   
+ 6σ Ωijk cσ − 3(cσii + cσ ) ıΩijk + cσii + cσ Ωijk exp −ıΩijk .

Hence, with

2  2  
ΛΩ
σ,i, j = Ωijk − ıΩ 3
ijk − Ω 2
ijk exp −ıΩijk ,
Ωijk
4

Equation (11) becomes

Ω 2  kk      2
Wσ,i, j = 6σ − 20cσij + 12cσij + 3(cσii + cσj j ) − 5cσkk ıΩijk + 4cσij + 2(cσii + cσj j − cσkk ) Ωijk
Ωijk
4
  3
+ 61 3(cσkk − cσii − cσj j ) − 8cσij ıΩijk + 12 cσ Ωijk + 30
1 kk 4
cσ Ωijk
1 kk 4

 ij      2   
+ 20cσ − 6cσ + 8cσ − 3(cσ + cσ ) − cσ ıΩijk + cσii − 2cσij + cσj j ] Ωijk
kk ij ii jj kk
exp −ıΩijk .

Since n = 2, the weights cσij are just given as follows:

cot αijσ ij

cσij = − , cσkk = ,
2 2h k

where ij denotes the edge length. We would like to express them explicitly in terms
of the Euclidean metric g of σ. In fact, we can distinguish the vertex k as origin and
use the hat functions xi and x j as coordinates on σ. With respect to these coordi-
nates, the metric is given by a matrix:
 
g11 g12
g= .
g21 g22

In terms of g the cotangent weights are given as follows:

g12 g11 − g12 g22 − g12

cσij = − √ , cσjk = − √ , cσki = − √ ,
2 det g 2 det g 2 det g
g11 − 2g12 + g22 g22 g11
cσkk = √ , cσii = √ , cσj j = √ ,
2 det g 2 det g 2 det g
238 F. Knöppel and U. Pinkall

and we have rederived the formulas in [8]:

1    
Ω
j = 3g11 + 4g12 + 3g22 − g11 + g12 + g22 ıΩijk + g612 ıΩijk 3
Wσ,i,
vol (σ)Ωijk 4
 
+ g11 −2g2412 +g22 Ωijk
4
+ g11 −2g6012 +g22 Ωijk
4
− 3g11 + 4g12 + 3g22
    2   
+ 2g11 + 3g12 + 2g22 ıΩijk − 21 g11 + 2g12 + g22 Ωijk exp −ıΩijk .

Acknowledgments This research was supported by the DFG Collaborative Research Center TRR
109, “Discretization in Geometry and Dynamics”.

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Avron, J., Osadchy, D., Seiler, R.: A topological look at the quantum Hall effect. Phys. Today
56, 38–42 (2003)
2. Bott, R.: On some recent interactions between mathematics and physics. Canad. Math. Bull.
28, 129–164 (1985)
3. Christiansen, S., Halvorsen, T.: A gauge invariant discretization on simplicial grids of the
Schrödinger eigenvalue problem in an electromagnetic field. SIAM J. Numer. Anal. 49, 331–
345 (2011)
4. Christiansen, S., Halvorsen, T.: A simplicial gauge theory. J. Math. Phys. 53 (2012)
5. Desbrun, M., Kanso, E., Tong, Y.: Discrete differential forms for computational modeling. In:
Discrete Differential Geometry, pp. 287–324. Birkhäuser Basel (2008)
6. Halvorsen, T., Sørensen, T.: Simplicial gauge theory and quantum gauge theory simulation.
Nucl. Phys. B 854, 166–183 (2012)
7. Hatcher, A.: Algebraic Topology. Cambridge University Press (2002)
8. Knöppel, F., Crane, K., Pinkall, U., Schröder, P.: Globally optimal direction fields. ACM
Trans. Graph. 32 (2013)
9. Knöppel, F., Crane, K., Pinkall, U., Schröder, P.: Stripe patterns on surfaces. ACM Trans.
Graph. 34 (2015)
10. Kobayashi, S.: La connexion des variétés fibrés I and II. Comptes Rendus de l’Académie
Sciences, Paris 54(318–319), 443–444 (1954)
11. Kostant, B.: Quantization and unitary representations. In: Lectures in Modern Analysis and
Applications III, vol. 170, pp. 87–208. Springer (1970)
12. Kreft, C., Seiler, R.: Models of the Hofstadter type. J. Math. Phys. 37, 5207–5243 (1996)
Complex Line Bundles Over Simplicial Complexes . . . 239

13. Morrison, K.: Yang-Mills connections on surfaces and representations of the path group. Proc.
Am. Math. Soc. 112, 1101–1106 (1991)
14. Munkres, J.R.: Elements of algebraic topology. In: Advanced Book Classics. Perseus Books
(1984)
15. Sasaki, K., Kawazoe, Y., Saito, R.: Aharonov-Bohm effect in higher genus materials. Physica
A 321, 369–375 (2004)
16. Seifert, H., Threlfall, W.: Lehrbuch der Topologie: mit 132 Figuren. Chelsea Scientific Books,
Chelsea (1934)
17. Simms, D.J., Woodhouse, N.M.J.: Lectures on geometric quantization. In: Lecture Notes in
Physics. Springer, Berlin (1976)
18. Weil, A.: Introduction a l’etude des varietes Kähleriennes. Actualités Scientifiques et Indus-
trielles. Hermann (1958)
19. Weißmann, S., Pinkall, U., Schröder, P.: Smoke rings from smoke. ACM Trans. Graph. 33
(2014)
Holomorphic Vector Fields and Quadratic
Differentials on Planar Triangular Meshes

Wai Yeung Lam and Ulrich Pinkall

Abstract Given a triangulated region in the complex plane, a discrete vector field
Y assigns a vector Yi ∈ C to every vertex. We call such a vector field holomorphic
if it defines an infinitesimal deformation of the triangulation that preserves length
cross ratios. We show that each holomorphic vector field can be constructed based
on a discrete harmonic function in the sense of the cotan Laplacian. Moreover, to
each holomorphic vector field we associate in a Möbius invariant fashion a certain
holomorphic quadratic differential. Here a quadratic differential is defined as an
object that assigns a purely imaginary number to each interior edge. Then we derive
a Weierstrass representation formula, which shows how a holomorphic quadratic
differential can be used to construct a discrete minimal surface with prescribed Gauß
map and prescribed Hopf differential.

1 Introduction

Consider an open subset U in the complex plane C ∼

= R2 with coordinates z =
x + i y together with a holomorphic vector field

∂
Y = f .
∂x

Here Y is a real vector field. It assigns to each p ∈ R2 the vector f ( p) ∈ C ∼

= R2 .
∂
We do not consider objects like ∂z which are sections of the complexified tangent
bundle (T R2 )C .

W.Y. Lam (B) · U. Pinkall

Technische Universität Berlin, Inst. Für Mathematik,
Straße des 17. Juni 136, 10623 Berlin, Germany
e-mail: [email protected]
U. Pinkall
e-mail: [email protected]

© The Author(s) 2016 241

A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_7
242 W.Y. Lam and U. Pinkall

Note f : U → C is a holomorphic function, i.e.

 
1 ∂f ∂f
0 = f z̄ = +i .
2 ∂x ∂y

Let t → gt denote the local flow of Y (defined for small t on open subsets of U with
compact closure in U ). Then the euclidean metric pulled back under gt is confor-
mally equivalently to the original metric:

gt∗  ,  = e2u  , 

for some real-valued function u. The infinitesimal change in scale u̇ is given by

1
u̇ = div Y = Re ( f z ) .
2
Note that u̇ is a harmonic function:

u̇ z z̄ = 0.

On the other hand, differentiating u̇ twice with respect to z yields one half the third
derivative of f :
1
u̇ zz = f zzz .
2
It is well-known that the vector field Y corresponds to an infinitesimal Möbius trans-
formation of the extended complex plane C if and only if f is a quadratic polyno-
mial. In this sense f zzz measures the infinitesimal “change in Möbius structure”
under Y (Möbius structures are sometimes also called “complex projective struc-
tures” [6]). Moreover, the holomorphic quadratic differential

q := f zzz dz 2

is invariant under Möbius transformations Φ. This is equivalent to saying that q is

unchanged under a change of variable Φ(z) = w = ξ + iη whenever Φ is a Möbius
transformation. This is easy to see if Φ(z) = az + b is an affine transformation. In
this case

dw = a dz
d 1 d
=
dw a dz

and therefore
∂
Y = f˜
∂ξ
Holomorphic Vector Fields and Quadratic Differentials . . . 243

with

f˜ = a f.

Thus we indeed have

f˜www dw 2 = f zzz dz 2 .

A similar argument applies to Φ(z) = 1z and therefore to all Möbius transforma-

tions.
For realizations from an open subset U of the Riemann sphere CP1 the vanishing
of the Schwarzian derivative characterizes Möbius transformations. The quadratic
differential q plays a similar role for vector fields. We call q the Möbius derivative
of Y .
An important geometric context where holomorphic quadratic differentials arise
comes from the theory of minimal surfaces: Given a simply connected Riemann
surface M together with a holomorphic immersion g : M → S 2 ⊂ R3 and a holo-
morphic quadratic differential q on M, there is a minimal surface F : M → R3
(unique up to translations) whose Gauß map is g and whose second fundamental
form is Re q.
In this paper we will provide a discrete version for all details of the above story.
Instead of smooth surfaces we will work with triangulated surfaces of arbitrary
combinatorics. The notion of conformality will be that of conformal equivalence
as explained in [3]. Holomorphic vector fields will be defined as infinitesimal con-
formal deformations.
There is also a completely parallel discrete story where conformal equivalence of
planar triangulations is replaced by preserving intersection angles of circumcircles.
To some extent we also tell this parallel story that belongs to the world of circle
patterns.
The results on planar triangular meshes in this paper are closely related to isother-
mic triangulated surfaces in Euclidean space [8].

2 Discrete Conformality

In this section, we review two notions of discrete conformality for planar triangular
meshes. We first start with some notations of triangular meshes.
Definition 2.1 A triangular mesh M is a simplicial complex whose underlying
topological space is a connected 2-manifold (with boundary). The set of vertices
(0-cells), edges (1-cells) and triangles (2-cells) are denoted as V , E and F.
We denote E int the set of interior edges and Vint the set of interior vertices.
Without further notice we will assume that all triangular meshes under consideration
are oriented.
244 W.Y. Lam and U. Pinkall

Fig. 1 Two neighboring and

oriented triangles

Definition 2.2 A realization z : V → C of a triangular mesh M in the extended

complex plane assigns to each vertex i ∈ V a point z i ∈ C in such a way that
for each triangle {ijk} ∈ F the points corresponding to its three vertices are not
collinear.
Given two complex numbers z 1 , z 2 ∈ C we write

z 1 , z 2  := Re(z̄ 1 z 2 ).

We are looking for suitable definitions of conformal structure of a realization z.

In particular, we want z to be conformally equivalent to g ◦ z whenever g : C →
C is a Möbius transformations. This requirement will certainly be met if we base
our definitions on complex cross ratios: Given a triangular mesh z : V → C, we
associate a complex number to each interior edge {ij} ∈ E int , namely the cross ratio
of the corresponding four vertices (See Fig. 1)

(z j − z k )(z i − zl )
cr z,ij = .
(z k − z i )(zl − z j )

Notice that cr z,ij = cr z, ji and hence cr z : E int → C is well defined. It is easy to

see that two realizations differ only by a Möbius transformation if and only if their
corresponding cross ratios are the same. In order to arrive at a more flexible notion
of conformality we need to relax the condition that demands the equality of all cross
ratios. Two natural ways to do this is to only require equality of either the norm or
alternatively the argument of the cross ratios. This leads to two different notions
of discrete conformality: conformal equivalence theory [9, 13] and circle pattern
theory [11].
Note that for the sake of simplicity of exposition we are ignoring here realizations
in C where one of the vertices is mapped to infinity.

2.1 Conformal Equivalence

The edge lengths of a triangular mesh realized in the complex plane provide a dis-
crete counterpart for the induced Euclidean metric in the smooth theory. A notion
Holomorphic Vector Fields and Quadratic Differentials . . . 245

of conformal equivalence based on edge lengths was proposed by Luo [9]. Later
Bobenko et al. [3] stated this notion in the following form:
Definition 2.3 Two realizations of a triangular mesh z, w : V → C are conformally
equivalent if the norm of the corresponding cross ratios are equal:

| cr z | ≡ | cr w |,

i.e. for each interior edge {ij}

|(z j − z k )||(z i − zl )| |(wj − wk )||(wi − wl )|

= .
|(z k − z i )||(zl − z j )| |(wk − wi )||(wl − wj )|

This definition can be restated in an equivalent form that closely mirrors the
notion of conformal equivalence of Riemannian metrics:
Theorem 2.4 Two realizations of a triangular mesh z, w : V → C are conformally
equivalent if and only if there exists u : V → R such that
u i +u j
|wj − wi | = e 2 |z j − z i |.

Proof It is easy to see that the existence of u implies conformal equivalence. Con-
versely, for two conformally equivalent realizations z, w, we define a function
σ : E → R by
|wj − wi | = eσij |z j − z i |.

Since z, w are conformally equivalent σ satisfies for each interior edge {ij}

σ jk − σki + σil − σlj = 0.

For any vertex i and any triangle {ijk} containing it we then define

eu i := eσki +σij −σjk .

Note the vertex star of i is a triangulated disk if i is interior, or is a fan if i is

a boundary vertex. Hence the value u i defined in this way is independent of the
chosen triangle. 

2.2 Circle Patterns

Given a triangular mesh realized in the complex plane we consider the circum-
scribed circles of its triangles. These circles inherit an orientation from their tri-
angles. The intersection angles of these circles from neighboring triangles (Fig. 2)
246 W.Y. Lam and U. Pinkall

Fig. 2 The intersection

angle of two neighboring
circumscribed circles

define a function φ : E int → [0, 2π ) which is related to the argument of the corre-
sponding cross ratio via

φi j = Arg(cr z,i j ). (1)

Based on these angles we obtain another notion of discrete conformality which

reflects the angle-preserving property that we have in the smooth theory.
Definition 2.5 Two realizations of a triangular mesh z, w : V → C have the same
pattern structure if the corresponding intersection angles of neighboring circum-
scribed circles are equal:

Arg(cr z,ij ) = Arg(cr w,ij ),

i.e. for each interior edge {ij}

(z j − z k )(z i − zl ) (wj − wk )(wi − wl )

Arg = Arg .
(z k − z i )(zl − z j ) (wk − wi )(wl − wj )

Just as conformal equivalence was related to scale factors u at vertices, having

the same pattern structure is related to the existence of certain angular velocities α
located at vertices:
Theorem 2.6 Two realizations of a triangular mesh z, w : V → C have the same
pattern structure if and only if there exists α : V → [0, 2π ) such that

wj − wi αi +αj z − z
j i
= ei 2 .
|wj − wi | |z j − z i |

Proof The argument is very similar to the one for Theorem 2.4. In particular, the
existence of the function α easily implies equality of the pattern structures. Con-
versely, assuming identical pattern structures we take any ω : E → R that satisfies

wj − wi zj − zi
= eiωij .
|wj − wi | |z j − z i |
Holomorphic Vector Fields and Quadratic Differentials . . . 247

For any vertex i and any triangle {ijk} containing it we define αi ∈ [0, 2π ) such that

eiαi = ei(ωki +ωij −ω jk ) .

Note the vertex star of i is a triangulated disk if i is interior, or is a fan if i is a

boundary vertex. Hence having the same pattern structure implies that the value αi
is independent of the chosen triangle. 

3 Infinitesimal Deformations and Linear Conformal

Theory

We will linearize both of the above notions of discrete conformality by considering

infinitesimal deformations. This will allow us to relate them to linear discrete com-
plex analysis, based on a discrete analogue of the Cauchy Riemann equations [4, 5,
10] (See the survey [12]).
Definition 3.1 An infinitesimal conformal deformation of a realization z : V → C
of a triangular mesh is a map ż : V → C such that there exists u : V → R satisfying

ż j − ż i ż j − ż i , z j − z i  ui + uj
Re = = .
zj − zi |z j − z i |2 2

We call u the scale change at vertices.

Definition 3.2 An infinitesimal pattern deformation of a realization z : V → C of

a triangular mesh is a map ż : V → C such that there exists α : V → R satisfying

ż j − ż i ż j − ż i , i(z j − z i ) αi + αj
Im = = .
zj − zi |z j − z i |2 2

We call α the angular velocities at vertices.

Example 3.3 The infinitesimal deformations ż := az 2 + bz + c, where a, b, c ∈ C

are constants, are both conformal and pattern deformations since

ż j − ż i
= (az i + b/2) + (az j + b/2).
zj − zi

Infinitesimal conformal deformations and infinitesimal pattern deformations are

closely related:

Theorem 3.4 Suppose z : V → C is a realization of a triangular mesh. Then an

infinitesimal deformation ż : V → C is conformal if and only if i ż is a pattern defor-
mation.
248 W.Y. Lam and U. Pinkall

Proof Notice

ż j − ż i , z j − z i  i ż j − i ż i , i(z j − z i )
= .
|z j − z i |2 |z j − z i |2

and the claim follows from Definition 3.1 and 3.2. 

3.1 Infinitesimal Deformations of a Triangle

Let z : V → C be a realization of a triangulated mesh and ż an infinitesimal defor-

mation. Up to an infinitesimal translation ż is completely determined by the infini-
tesimal scalings and rotations that it induces on each edge. These infinitesimal scal-
ings and rotations of edges satisfy certain compatibility conditions on each triangle.
These conditions involve the cotangent coefficients well known from the theory of
discrete Laplacians. As we will see in Sect. 3.2, for conformal deformations (as
well as for pattern deformations) the infinitesimal scalings and rotations of edges
are indeed discrete harmonic functions.
Consider three pairwise distinct points z 1 , z 2 , z 3 ∈ C that do not lie on a line. In
the following i, j, k denotes any cyclic permutation of the indexes 1, 2, 3. The trian-
gle angle at the vertex i is denoted by βi . We adopt the convention that all β1 , β2 , β3
have positive sign if the triangle z 1 , z 2 , z 3 is positively oriented and a negative sign
otherwise. Suppose we have an infinitesimal deformation of this triangle. Then there
exists σij , ωij ∈ R such that

ż j − ż i = (σij + iωij )(z j − z i ). (2)

The scalars σij and ωij describe the infinitesimal scalings and rotations of the edges.
They satisfy the following compatibility conditions:

Lemma 3.5 Given σij , ωij ∈ R the following statements are equivalent:
(a) There exist ż i such that (2) holds.
(b) We have

0 = (σ12 + iω12 )(z 2 − z 1 ) + (σ23 + iω23 )(z 3 − z 2 ) + (σ31 + iω31 )(z 1 − z 3 ). (3)

(c) There exists ω ∈ R such that

iω = iω23 + i cot β1 (σ31 − σ12 )

= iω31 + i cot β2 (σ12 − σ23 )
= iω12 + i cot β3 (σ23 − σ31 ).
Holomorphic Vector Fields and Quadratic Differentials . . . 249

(d) There exist σ ∈ R such that

σ = σ23 + i cot β1 (iω31 − iω12 )

= σ31 + i cot β2 (iω12 − iω23 )
= σ12 + i cot β3 (iω23 − iω31 ).

Proof The relation between (a) and (b) is obvious. We show the equivalence
between (b) and (c). With A denoting the signed triangle area we have the following
identities:

0 = i(z j − z i ), z j − z i ,
2 A = i(z j − z i ), z k − z j ,
i(z j − z i ), i(z j − z i ) = z j − z i , z j − z i .

Using these identities and z 3 − z 2 ∈ spanR {i(z 1 − z 3 ), i(z 2 − z 1 )} we obtain

z 3 − z 2 = cot(β3 )i(z 2 − z 1 ) − cot(β2 )i(z 1 − z 3 ). (4)

Cyclic permutation yields

z 1 − z 3 = cot(β1 )i(z 3 − z 2 ) − cot(β3 )i(z 2 − z 1 ),

z 2 − z 1 = cot(β2 )i(z 1 − z 3 ) − cot(β1 )i(z 3 − z 2 ).

Substituting these identities into Equation (3) we obtain

 
0= σ1 cot(β3 )i(z 2 − z 1 ) − cot(β2 )i(z 1 − z 3 ) + ω23 i(z 3 − z 2 )
 
+ σ2 cot(β1 )i(z 3 − z 2 ) − cot(β3 )i(z 2 − z 1 ) + ω31 i(z 1 − z 3 )
 
+ σ3 cot(β2 )i(z 1 − z 3 ) − cot(β1 )i(z 3 − z 2 ) + ω12 i(z 2 − z 1 )
 
= ω1 + cot β1 (σ2 − σ3 ) i(z 3 − z 2 )
 
+ ω2 + cot β2 (σ3 − σ1 ) i(z 1 − z 3 )
 
+ ω3 + cot β3 (σ1 − σ2 ) i(z 2 − z 1 ).

Now we use that λ1 , λ2 , λ3 ∈ C satisfy

λ1 i(z 3 − z 2 ) + λ2 i(z 1 − z 3 ) + λ3 i(z 2 − z 1 ) = 0,

if and only if λ1 = λ2 = λ3 . This establishes the equivalence of (b) and (c). The
equivalence of (b) and (d) is seen in a similar fashion by eliminating i(z j − z i ) in
(3) instead of (z j − z i ). 
250 W.Y. Lam and U. Pinkall

The quantity ω above describes the average rotation speed of the triangle. Simi-
larly, it can be verified that the above σ satisfies

Ṙ
σ =
R
where R denotes the circumradius of the triangle. Thus σ signifies an average scal-
ing of the triangle.

3.2 Harmonic Functions with Respect to the Cotangent

Laplacian

In smooth complex analysis conformal maps are closely related to harmonic func-
tions. If a conformal map preserves orientation it is holomorphic and satisfies the
Cauchy Riemann equations. In particular, its real part and the imaginary part are
conjugate harmonic functions. Conversely, given a harmonic function on a simply
connected surface then it is the real part of some conformal map.
A similar relationship manifests between discrete harmonic functions (in the
sense of the cotangent Laplacian) and infinitesimal deformations of triangular
meshes. Discrete harmonic functions can be regarded as the real part of holomorphic
functions which satisfies a discrete analogue of the Cauchy Riemann equations. In
particular, a relation between discrete harmonic functions and infinitesimal pattern
deformations was found by Bobenko, Mercat and Suris [2]. Integrable systems were
involved in this context. We extend their result to include the case of infinitesimal
conformal deformations.

Theorem 3.6 Let z : V → C be a simply connected triangular mesh realized in the

complex plane and h : V → R be a function. Then the following are equivalent:
(a) h is a harmonic function with respect to the cotangent Laplacian, i.e. using
the notation of Fig. 1, for all interior vertices i ∈ Vint we have

(cot βijk + cot β lji )(h j − h i ) = 0. (5)
j

(b) There exists an infinitesimal conformal deformation ż : V → C with scale

factors given by h. It is unique up to infinitesimal rotations and translations.
(c) There exists an infinitesimal pattern deformation i ż : V → C with h as angu-
lar velocities. It is unique up to infinitesimal scalings and translations.

Proof We show the equivalence of the first two statements. The equivalence of the
first and the third follows similarly.
Holomorphic Vector Fields and Quadratic Differentials . . . 251

Suppose h is a harmonic function. Since the triangular mesh is simply connected,

equation (5) implies the existence of a function ω̃ : F → R such that for all interior
edges {ij} we have

i ω̃ijk − i ω̃ jil = i(cot βijk + cot β lji )(h j − h i ).

Here ω̃ is unique up to an additive constant and called the conjugate harmonic func-
tion of h. Using ω̃ we define a function ω : E → R via

iωij = i ω̃ijk − i cot βijk (h j − h i ).

Lemma 3.5 now implies that there exists ż : V → C such that

 
hi + hj
(ż j − ż i ) = + iωij (z j − z i ).
2

This gives us the desired infinitesimal conformal deformation of z with h as scale

factors.
To show uniqueness, suppose ż, ż  are infinitesimal conformal deformations with
the same scale factors. Then ż − ż  preserves all the edge lengths of the triangular
mesh and hence is induced from an Euclidean transformation.
Conversely, given an infinitesimal conformal deformation ż with scale factors h.
We write
 
hi + hj
ż j − ż i = + iωij (z j − z i )
2

for some ω : E → R. Lemma 3.5 implies that there is a function ω̃ : F → R such

that

i ω̃ijk = iωij + i cot βijk (h j − h i ).

We have

i ω̃ijk − i ω̃ jil = i(cot βijk + cot β lji )(h j − h i )

and

(cot βijk + cot β lji )(h j − h i ) = 0.
j

Therefore h is harmonic. 
252 W.Y. Lam and U. Pinkall

4 Holomorphic Quadratic Differentials

In this section, we introduce a discrete analogue of holomorphic quadratic differen-

tials. We illustrate their correspondence to discrete harmonic functions. It reflects
the property in the smooth theory that holomorphic quadratic differentials parame-
trize Möbius structures on Riemann surfaces ([6, Chap. 9]).
To simplify the notation, we make use of discrete differential forms. We denote E
the set of oriented edges and E int the set of oriented interior edges. Given an oriented
triangular mesh M, a complex-valued function η : E → C is called a discrete 1-
form if

η(eij ) = −η(eji )
∀eij ∈ E.

It is closed if for every face {ijk}

η(eij ) + η(ejk ) + η(eki ) = 0.

It is exact if there exists a function f : V → C such that

η(eij ) = d f (eij ) := f j − f i .

Similarly, we can consider discrete 1-forms on the dual graph M ∗ of M and these are
called dual 1-forms. Given an oriented edge e, we denote e∗ its dual edge oriented
from the right face of e to its left face. The set of oriented dual edges is denoted
by E ∗ .

Definition 4.1 Given a triangular mesh z : V → C realized on the complex plane, a

function q : E int → iR defined on interior edges is a discrete holomorphic quadratic
differential if it satisfies for every interior vertex i ∈ Vint

qij = 0,
j

qij /dz(eij ) = 0.
j

Theorem 4.2 Let q : E int → iR be a holomorphic quadratic differential on a real-

ization z : V → C of a triangular mesh. Suppose Φ : C → C is a Möbius transfor-
mation which does not map any vertex to infinity. Then q is again a holomorphic
quadratic differential on w := Φ ◦ z.

Proof Since Möbius transformations are generated by Euclidean transformations

and inversions, it suffices to consider the inversion in the unit circle at the origin

w := Φ(z) = 1/z.
Holomorphic Vector Fields and Quadratic Differentials . . . 253

We have
   
qij /dw(eij ) = −z i z j qij /dz(eij ) = −z i qij − z i2 qij /dz(eij ) = 0.
j j j j

Hence the claims follow. 

We are going to show that on a simply connected triangular mesh, there is a

correspondence between discrete holomorphic quadratic differentials and discrete
harmonic functions.
We first show how to construct a discrete holomorphic quadratic differential from
a harmonic function. Given a function u : V → R on a realization of z : V → C of
a triangular mesh M. If we interpolate it piecewise-linearly over each triangular
face, its gradient is constant on each face and we have grad z u : F → C given by

u i dz(e jk ) + u j dz(eki ) + u k dz(eij )

grad z u ijk = i .
2 Aijk

Note that we ignore here the non-generic case (which leads to the vanishing of the
area) where the triangle degenerates in the sense that its circumcircle passes through
the point at infinity. Also note that for a non-degenerate triangle that is mapped by z
in C in an orientation reversing fashion the area Aijk is considered to have a negative
sign. Granted this, one can verify that the gradient of u satisfies

grad z u ijk , dz(eij ) = u j − u i ∀{ij} ⊂ {ijk} ∈ F.

We define u z : F → C by
1
u z := grad z u.
2

and the dual 1-form du z : E int

∗
→ C on M by

du z (eij∗ ) := (u z )ijk − (u z ) jil

where {ijk} is the left face and { jil} is the right face of the oriented edge eij .

Lemma 4.3 Given a function u : V → R on a realization of a triangular mesh

z : V → C, we have

du z (eij∗ )dz(eij )
−i  j j 
= cot β ijk (u k − u j ) + cot βki (u k − u i ) + cot βil (u l − u i ) + cot βlij (u l − u j )
2
which is purely imaginary (Fig. 1).
254 W.Y. Lam and U. Pinkall

Proof Since

grad z u i jk , dz(eij ) = u j − u i = grad z u jkl , dz(ei j ),

we have

Re(du z (ei∗j )dz(ei j )) = 0.

On the other hand, using equation (4) we get

Re(du z (eij∗ )idz(eij ))

= Re(((u z )ijk − (u z )jil )idz(eij ))
j
= (grad z u ijk , cot βjk
i
dz(ejk ) − cot βki dz(eki )
j
+ grad z u jil , cot βil dz(eil ) − cot βlij dz(elj ))/2
1 j j 
= cot βjk
i
(u k − u j ) + cot βki (u k − u i ) + cot βil (u l − u i ) + cot βlji (u l − u j ) .
2
Hence the claim follows. 

Lemma 4.4 Given a realization z : V → C of a triangular mesh. A function u :

V → R is harmonic if and only if the function q : E int → iR defined by

qij := du z (eij∗ )dz(eij )

is a holomorphic quadratic differential.

Proof Note q is well defined since

qij = du z (eij∗ )dz(eij ) = du z (e∗ji )dz(e ji ) = q ji .

It holds for general functions u : V → R that

Re(q) ≡ 0
 
qij /dz(eij ) = du z (eij∗ ) = 0 ∀i ∈ Vint .
j j

We know from Lemma 4.3 that for every interior vertex i ∈ Vint
  i 
qij = du z (eij∗ )dz(eij ) = (cot βijk + cot β lji )(u j − u i ).
j j
2 j

Hence, u is harmonic if and only if q is a holomorphic quadratic differential. 

Holomorphic Vector Fields and Quadratic Differentials . . . 255

Lemma 4.5 Let z : V → C be a realization of a simply connected triangular mesh.

Given a function q : E int → iR such that for every interior vertex i ∈ Vint

qij /dz(eij ) = 0,
j

there exists a function u : V → R such that for every interior edge {ij}

qij = du z (eij∗ )dz(eij ).

Proof We consider a dual 1-form τ on M defined by

τ (eij∗ ) = qij /dz(eij ).

Since M is simply connected and

 
τ (eij∗ ) = qij /dz(eij ) = 0,
j j

there exists a function h : F → C such that

dh(eij∗ ) := h ijk − h jil = τ (eij∗ ).

It implies we have Re(dh(e∗ )dz(e)) = Re(q) ≡ 0 and

ω(eij ) := 2h̄ ijk , dz(eij ) = 2h̄ jil , dz(eij )

is a well-defined R-valued 1-form. Since the triangular mesh is simply connected

and for every face {ijk}

ω(eij ) + ω(ejk ) + ω(eki ) = 0,

there exists a function u : V → R such that for every oriented edge eij

du(eij ) = u j − u i = ω(eij ).

It can be verified that

1
h= grad z u = u z .
2
Hence we obtain

qij = τ (eij∗ )dz(eij ) = dh(eij∗ )dz(eij ) = du z (eij∗ )dz(eij )

for every interior edge {ij}. 

256 W.Y. Lam and U. Pinkall

Theorem 4.6 Suppose z : V → C is a realization of a simply connected triangular

mesh. Then any holomorphic quadratic differential q : E int → iR is of the form

qij = du z (eij∗ )dz(eij ) ∀eij ∈ E int

for some harmonic function u : V → R.

Furthermore, the space of holomorphic quadratic differentials is a vector space
isomorphic to the space of discrete harmonic functions module linear functions.

Proof The first part of the statement follows from Lemmas 4.4 and 4.5. In order to
show the second part, it suffices to observe that

du z ≡ 0 ⇐⇒ grad u ≡ a ⇐⇒ du = a, dz ⇐⇒ u = a, z + b

for some a, b ∈ C. 

In previous sections, we showed that every harmonic function corresponds to

an infinitesimal conformal deformation. The following shows that discrete holo-
morphic quadratic differentials are the change in the intersection angles of circum-
scribed circles.

Theorem 4.7 Let z : V → C be a realization of a simply connected triangular

mesh. Suppose u : V → R is a discrete harmonic function and ż is an infinitesi-
mal conformal deformation with u as scale factors. Then we have

˙z
1 cr i
du z dz = − = − φ̇
2 cr z 2

where φ̇ : E int → R denotes the change in the intersection angles of neighboring

circumscribed circles.
h i +h j
Proof We write (ż j − ż i ) = ( 2
+ iωij )(z j − z i ). Applying Lemma 4.3 we have

˙ z,ij / cr z,ij = iω jk − iωki + iωil − iωlj

cr
 j j 
= i cot β ijk (u k − u j ) + cot βki (u k − u i ) + cot βil (u l − u i ) + cot βlji (u l − u j )
= −2du z (eij∗ )dz(eij ).

The equality
˙z
cr
= i φ̇
cr z

follows from Equation (1). 

Holomorphic Vector Fields and Quadratic Differentials . . . 257

5 Conformal Deformations in Terms of End(C2 )

In this section we show how an infinitesimal conformal deformation gives rise to

a discrete analogue of a holomorphic null curve in C3 . Later we will see that the
real parts of such a “holomorphic null curve” can be regarded as the Weierstrass
representation of a discrete minimal surface.
Up to now we have mostly treated the Riemann sphere CP1 as the extended
complex plane C = C ∪ {∞}. In this section we will take a more explicitly Möbius
geometric approach: We will represent fractional linear transformations of C by
linear transformations of C2 with determinant one. Actually, the group of Möbius
transformations is

= PSL (2, C) ∼
Möb(C) ∼ = SL (2, C)/(±I ). (6)

However, since we are mainly interested in infinitesimal deformations and any

map into PSL(2, C) whose values stay close to the identity admits a canonical lift
to SL(2, C), we can safely ignore the difference between PSL(2, C) and SL(2, C).
Given a realization z : V → C of a triangular mesh we consider its lift
ψ : V → C2
 
z
ψ :=
1

and regard the realization as a map Ψ : V → CP1 where

 
z
Ψ := C = [ψ].
1

The action of a Möbius transformation on the Riemann sphere is given by a

matrix A ∈ SL(2, C), which is unique up to sign:

[ϕ] → [Aϕ].

Before we investigate infinitesimal deformations we first consider finite defor-

mations of a triangular mesh Ψ : V → CP1 . Given such a finite deformation, the
change in the positions of the three vertices of a triangle {ijk} can be described by
a Möbius transformation, which is represented by G ijk ∈ SL(2, C). They satisfy a
compatibility condition on each interior edge {ij} (see Fig. 1):

[G ijk ψi ] = [G jil ψi ],
[G ijk ψj ] = [G jil ψj ].
258 W.Y. Lam and U. Pinkall

Suppose now that the mesh is simply connected. Then up to a global Möbius trans-
formation the map G : F → SL(2, C) can be uniquely reconstructed from the mul-
tiplicative dual 1-form defined as
−1
G(eij∗ ) := G jil G ijk .

G(eij∗ ) is defined whenever {ij} is an interior edge and we have

G(eij∗ ) = G(eji∗ )−1 .

Moreover, for every interior vertex i we have


G(eij∗ ) = I.
j

The compatibility conditions imply that for interior each edge {ij} there exist
λij,i , λij, j ∈ C\{0} such that

G(eij∗ )ψi = λij,i ψi

G(eij∗ )ψj = λij,j ψj .

Since λij,i λij, j = det(G(eij∗ )) = 1, we have

λij := λij,i = 1/λij,j .

Because of G(eij∗ ) = G(eij∗ )−1 we know

λij = λij,i = 1/λji,i = λji .

Hence λ defines a complex-valued function on the set E int of interior edges.

We now show that for each interior edge λij determines the change in the cross
ratio of the four points of the two adjacent triangles. Note that the cross ratio of four
points in CP1 can expressed as

det(ψk , ψj ) det(ψl , ψi )
cr([ψj ], [ψk ], [ψi ], [ψl ]) = .
det(ψi , ψk ) det(ψj , ψl )

Lemma 5.1 Suppose we are given four points [ψi ], [ψj ], [ψk ], [ψl ] ∈ CP1 and G ∈
SL(2, C) with

Gψi = λ−1 ψi
Gψj = λψj
Holomorphic Vector Fields and Quadratic Differentials . . . 259

for some λ ∈ C\{0}. Then the cross ratio of the four transformed points

[ψ̃i ] = [Gψi ], [ψ̃j ] = [Gψj ], [ψ̃k ] = [Gψk ], [ψ̃l ] = [ψl ]

is given by

cr([ψ̃j ], [ψ̃k ], [ψ̃i ], [ψ̃l ]) = cr([ψj ], [ψk ], [ψi ], [ψl ])/λ2 .

Proof

det(Gψk , Gψj ) det(ψl , Gψi )

cr([ψ̃j ], [ψ̃k ], [ψ̃i ], [ψ̃l ]) =
det(Gψi , Gψk ) det(Gψj , ψl )
= cr([ψj ], [ψk ], [ψi ], [ψl ])/λ2 .

We now can summarize the information about finite deformations of a realization

as follows:

Theorem 5.2 Let Ψ : V → CP1 be a realization of a simply connected triangu-

lar mesh. Then there is a bijection between finite deformations of Ψ in CP1 mod-
ulo global Möbius transformations and multiplicative dual 1 forms G : E int∗
→
SL(2, C) satisfying for every interior vertex i

G(eij∗ ) = I
j

and for every interior edge

G(eij∗ ) = G(e∗ji )−1

G(eij∗ )ψi = λij−1 ψi
G(eij∗ )ψj = λij ψj .

Here λ : E int → C\{0}. We denote cr : E int → C the cross ratios of Ψ and cr :
E int → C the cross ratios of a new realization described by G. Then

cr = cr /λ2 .

In particular,

|λ| ≡ 1 =⇒ the deformation is conformal.

Arg(λ) ≡ 0 =⇒ the deformation is a pattern deformation.
260 W.Y. Lam and U. Pinkall

Suppose we have a family of deformations described by dual 1-forms G t :

E int → SL(2, C) with G 0 ≡ I . By considering η := dtd |t=0 G t we obtain the fol-
lowing description of infinitesimal deformations:

Corollary 5.3 Let Ψ : V → CP1 be a realization of a simply connected triangular

mesh. Then there is a bijection between infinitesimal deformations of Ψ in CP1
modulo infinitesimal Möbius transformations and dual 1 forms η : E int → sl(2, C)
satisfying for every interior vertex i

η(eij∗ ) = 0 (7)
j

and for every interior edge

η(eij∗ ) = −η(e∗ji )
η(eij∗ )ψi = −μij ψi
η(eij∗ )ψj = μij ψj .

Here μ : E int → C. We denote cr : E int → C the cross ratios of Ψ and cr

˙ : E int →
C the rate of change in cross ratios induced by the infinitesimal deformation
described by η. Then
˙
1 cr
μ=− .
2 cr
In particular,

Re(μ) ≡ 0 =⇒ the infinitesimal deformation is conformal,

Im(μ) ≡ 0 =⇒ the infinitesimal deformation is a pattern deformation.

Note that given a mesh, the 1-form η is uniquely determined by the eigenfunction
μ. We now investigate the constraints on μ implied by the closedness condition (7)
of η.
Consider the symmetric bilinear form ( , ) : C2 × C2 → sl(2, C)

(φ, ϕ)v := det(φ, v)ϕ + det(ϕ, v)φ.

For ψi = ψj ∈ C2 we define

1
m ij := (ψj , ψi ) ∈ sl(2, C).
det(ψi , ψj )
Holomorphic Vector Fields and Quadratic Differentials . . . 261

The matrix m ij is independent of the representatives of [ψi ], [ψj ] ∈ CP1 and we

have
m ij = −m ji
m ij ψi = −ψi
m ij ψj = ψj .
 
zi
Using the representatives ψi = we obtain
1

μij
η(eij∗ ) = (ψi , ψj )
det(ψj , ψi )
 
μij z i + z j −2z i z j
= .
zj − zi 2 −z i − z j

Hence
  
η(eij∗ ) = 0 ⇐⇒ μij = 0 and μij /(z j − z i ) = 0. (8)
j j j

We consider the Pauli matrices

     
01 0 i 1 0
σ1 = , σ2 = , σ3 =
10 −i 0 0 −1

which form a basis of sl(2, C). Then

μij
η(eij∗ ) = ((1 − z i z j )σ1 + i(1 + z i z j )σ2 + (z i + z j )σ3 ).
zj − zi

If we now identify sl(2, C) with C3 via

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
1 0 0
σi → ⎝ 0 ⎠ , σ2 → ⎝ 1 ⎠ , σ3 → ⎝ 0 ⎠ ,
0 0 1

we obtain
⎛ ⎞
1 − zi zj
μij ⎝
η(eij∗ ) = i(1 + z i z j ) ⎠ . (9)
zj − zi zi + zj

Thus to every infinitesimal deformation of a realized triangular mesh we can asso-

ciate a closed sl(2, C)-valued dual 1-form. In the special case of an infinitesimal
conformal deformation (i.e. μ is real-valued) we will see that this yields a discrete
analogue of the Weierstrass representation for minimal surfaces.
262 W.Y. Lam and U. Pinkall

6 Weierstrass Representation of Discrete Minimal Surfaces

The Weierstrass representation for minimal surfaces in R3 is the most classical

example for applications of complex analysis:

Theorem 6.1 Given two meromorphic functions g, h : U ⊂ C → C such that g 2 h

is holomorphic. Then f : U → R3 defined by
⎛⎛ ⎞ ⎞ ⎛⎛ ⎞ ⎞
1 − g2 1 − g2
q
d f = Re ⎝⎝ i(1 + g 2 ) ⎠ h(z)dz ⎠ = Re ⎝⎝ i(1 + g 2 ) ⎠ ⎠
2g 2g dg

is a minimal surface. Its Gauß map n is the stereographic projection of g

⎛ ⎞
2Reg
1 ⎝ 2Img ⎠ .
n=
|g|2 + 1 |g|2 − 1

The holomorphic quadratic differential q := hgz dz 2 is called the Hopf differential

of f and encodes its second fundamental form: The direction defined by a nonzero
tangent vector W is

an asymptotic direction ⇐⇒ q(W ) ∈ iR.

a principal curvature direction ⇐⇒ q(W ) ∈ R.

Locally, every minimal surface can be written in this form.

We now develop a discrete version of this theorem for arbitrary triangular meshes
realized in the complex plane. A similar formula for quadrilateral meshes with fac-
torized real cross ratios was established by Bobenko and Pinkall [1]. Here we will
use the definition of a discrete minimal surface f with Gauß map n given in [8]:

Definition 6.2 Let n : V → S2 be a realization of a triangular mesh on the unit

sphere in R3 . Then a map f : F → R3 defined on the set F of faces is called a
discrete minimal surface with Gauß map n if for all oriented interior edges eij we
have

(n j − n i ) × ( f ijk − f jil ) = 0.

Here {ijk} and { jil} denote the left and the right faces of eij .

This definition mirrors the fact from the smooth theory that minimal surfaces are
Christoffel duals of their Gauß maps (Fig. 3). The correspondence between discrete
harmonic functions and discrete minimal surfaces was observed in [8]. Here is a
Weierstrass representation for discrete minimal surfaces in terms of their Gauß map
and their Hopf differential:
Holomorphic Vector Fields and Quadratic Differentials . . . 263

Fig. 3 Left A triangulated surface n : V → S2 with vertices on the unit sphere. Right A discrete
minimal surface f : F → R3 satisfying Definition 6.2

Theorem 6.3 Let z : V → C be a realization of a simply connected triangular

mesh and q : E int → iR a holomorphic quadratic differential. Then there exists
f : F → R3 such that for every interior edge {ij}
⎛ ⎛ ⎞⎞
1 − zi zj
qij
d f (eij∗ ) = Re ⎝ ⎝ i(1 + z i z j ) ⎠⎠ . (10)
i(z j − z i ) zi + zj

Moreover f is a discrete minimal surface with Gauß map

⎛ ⎞
2Rez
1 ⎝ 2Imz ⎠ .
n= 2
|z| + 1 |z|2 − 1

Locally, every discrete minimal surface can be written in this form.

Proof Suppose q : E int → iR is a holomorphic quadratic differential. Then by (8)
and (9) the dual 1-form η defined as
⎛ ⎞
1 − zi zj
qij ⎝ i(1 + z i z j ) ⎠
η(eij∗ ) :=
i(z j − z i ) zi + zj

satisfies

η(eij∗ ) = 0
j

for all interior vertices i. Therefore, since the triangular mesh is simply connected,
there exists F : F → C3 such that for any interior edge e we have

dF(e∗ ) = η(e∗ ).
264 W.Y. Lam and U. Pinkall

Thus the map f : F → R3 defined by f := ReF satisfies Eq. (10). To show that f
is a discrete minimal surface we define a function k : E int → R by

kij := −i qij /|z j − z i |2 .

Then by direct computation we obtain

kij (1 + |z i |2 )(1 + |z j |2 )
d f (eij∗ ) = (n j − n i ). (11)
2
This shows that f is a discrete minimal surface with Gauß map n. The converse is
straightforward: Given a discrete minimal surface f with Gauß map n we define
k : E int → R via (11). Then it can be shown that the function

qij := i kij |z j − z i |2

is a holomorphic quadratic differential. 

Remark 6.4 The discrete minimal surfaces given by (10) are trivalent meshes with
planar vertex stars for purely imaginary q. It is closely related to discrete asymptotic
nets. The factor i in front of z j − z i appears since the integration is taken over a dual
mesh while in the smooth theory ∗dz = idz.

Note that we could also consider the periodic one-parameter family of maps f α :
F → R3 defined for α ∈ R by

f α := Re(eiα F).

This family of discrete surfaces can be regarded as an associate family of minimal

surfaces and is investigated in [7].

Acknowledgments This research was supported by DFG SFB/TRR 109 “Discretization in Geom-
etry and Dynamics”

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.
Holomorphic Vector Fields and Quadratic Differentials . . . 265

References

1. Bobenko, A., Pinkall, U.: Discrete isothermic surfaces. J. Reine Angew. Math. 475, 187–208
(1996)
2. Bobenko, A.I., Mercat, C., Suris, Y.B.: Linear and nonlinear theories of discrete analytic func-
tions. Integrable structure and isomonodromic Green’s function. J. Reine Angew. Math. 583,
117–161 (2005)
3. Bobenko, A.I., Pinkall, U., Springborn, B.A.: Discrete conformal maps and ideal hyperbolic
polyhedra. Geom. Topol. 19(4), 2155–2215 (2015)
4. Duffin, R.J.: Basic properties of discrete analytic functions. Duke Math. J. 23, 335–363 (1956)
5. Ferrand, J.: Fonctions préharmoniques et fonctions préholomorphes. Bull. Sci. Math. 2(68),
152–180 (1944)
6. Gunning, R.C.: Lectures on Riemann surfaces. Princeton Mathematical Notes. Princeton Uni-
versity Press, Princeton (1966)
7. Lam, W.Y.: Discrete minimal surfaces: critical points of the area functional from integrable
systems (2015). arXiv:1510.08788
8. Lam, W.Y., Pinkall, U.: Isothermic triangulated surfaces (2015). arXiv:1501.02587
9. Luo, F.: Combinatorial Yamabe flow on surfaces. Commun. Contemp. Math. 6(5), 765–780
(2004)
10. Mercat, C.: Discrete Riemann surfaces and the Ising model. Comm. Math. Phys. 218(1), 177–
216 (2001)
11. Schramm, O.: Circle patterns with the combinatorics of the square grid. Duke Math. J. 86(2),
347–389 (1997)
12. Smirnov, S.: Discrete complex analysis and probability. Proceedings of the International
Congress of Mathematicians. Volume I, pp. 595–621. Hindustan Book Agency, New Delhi
(2010)
13. Springborn, B., Schröder, P., Pinkall, U.: Conformal equivalence of triangle meshes. ACM
Trans. Graph. 27(3), 77:1–77:11 (2008)
Vertex Normals and Face Curvatures
of Triangle Meshes

Xiang Sun, Caigui Jiang, Johannes Wallner and Helmut Pottmann

Abstract This study contributes to the discrete differential geometry of triangle

meshes, in combination with discrete line congruences associated with such meshes.
In particular we discuss when a congruence defined by linear interpolation of vertex
normals deserves to be called a ‘normal’ congruence. Our main results are a dis-
cussion of various definitions of normality, a detailed study of the geometry of such
congruences, and a concept of curvatures and shape operators associated with the
faces of a triangle mesh. These curvatures are compatible with both normal congru-
ences and the Steiner formula.

1 Introduction

The system of lines orthogonal to a surface (called the normal congruence of that
surface) has close relations to the surface’s curvatures and is a well studied object
of classical differential geometry, see e.g. [14]. It is quite surprising that this natural
correspondence has not been extensively exploited in discrete differential geometry:
most notions of discrete curvature are constructed in a way not involving normals, or
involving normals only implicitly. There are however applications such as support
structures and shading/lighting systems in architectural geometry where line con-
gruences, and in particular normal congruences, come into play [21]. We continue

X. Sun · C. Jiang
King Abdullah Univ. of Science and Technology, Thuwal 23955, Saudi Arabia
e-mail: [email protected]
C. Jiang
e-mail: [email protected]
J. Wallner (B)
Graz University of Technology, 8010 Graz, Austria
e-mail: [email protected]
H. Pottmann
Vienna University of Technology, 1040 Wien, Austria
e-mail: [email protected]

© The Author(s) 2016 267

A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_8
268 X. Sun et al.

this study, elaborate on discrete normal congruences in more depth and present a
novel discrete curvature theory for triangle meshes which is based on discrete line
congruences.
Contributions and overview. We organize our presentation as as follows. Section 2
summarizes properties of smooth congruences and elaborates on an important exam-
ple arising in the context of linear interpolation of surface normals.
Section 3 first recalls discrete congruences following the work of Wang et al. [21]
and then focuses on the interesting geometry of a new version of discrete normal
congruences (defined over triangle meshes). We shed new light onto the behavior
of linearly interpolated surface normals and discuss the problem of choosing vertex
normals.
In Sect. 4, discrete normal congruences lead to a curvature theory for triangle
meshes which has many analogies to the classical smooth setting. Unlike most other
concepts of discrete curvature, it assigns values of the curvatures (principal, mean,
Gaussian) to the faces of a triangle mesh. We discuss internal consistency of this
theory and show by examples (Sect. 5), that it is well suited for curvature estimation
and other applications.
Previous work. Smooth line congruences represent a classical subject. An intro-
duction may be found in the monograph by Pottmann and Wallner [16]. Discrete
congruences have appeared both in discrete differential geometry and geometry
processing. Let us first mention contributions which study congruences based on
triangle meshes: A computational framework for normal congruences and for esti-
mating focal surfaces of meshes with known or estimated normals has been presented
by Yu et al. [22]. The paper by Wang et al. [21] is described in more detail below.
Congruences associated with quad meshes are discrete versions of parametrized
congruences associated with parametrized surfaces. In particular, the so-called torsal
parametrizations are discussed from the integrable systems perspective by Bobenko
and Suris [3]. An earlier contribution in this direction is due to Doliwa et al. [6]. These
special parametrizations also occur as node axes in torsion-free support structures in
architectural geometry [12, 15, 17].
Curvatures of triangle meshes are a well studied subject. One may distinguish
between numerical approximation schemes (such as the jet fitting approach [4] or
integral invariants [18]) on the one hand, and extensive studies from the discrete
differential geometry perspective on the other hand. Without going into any detail
we mention that these include discrete exterior calculus [5], the geometry of offset-
like sets and distance functions [13], or various ways of defining shape operators
[8, 9]. Naturally, also Yu et al. [22] address this topic when studying discrete normal
congruences and focal surfaces. We present here yet another definition of curvatures
for triangle meshes which is based on discrete normal congruences, and which is at
the same time motivated by the Steiner formula (which also plays an important role
in [2, 13, 15]).
Vertex Normals and Face Curvatures of Triangle Meshes 269

2 Smooth Line Congruences

The introduction into line congruences in this section follows the paper by Wang et al.
[21]. A line congruence L is a smooth 2D manifold of lines described locally by lines
L(u, v) which connect corresponding points a(u, v) and b(u, v) of two surfaces. With
e(u, v) = b(u, v) − a(u, v) indicating the direction of the line L(u, v) (see Fig. 1),
we employ the volumetric parametrization

x(u, v, λ) = a(u, v) + λe(u, v) = (1 − λ)a(u, v) + λb(u, v).

Any 1-parameter family R (t) = L(u(t), v(t)) of lines results in a ruled surface
r(t, λ) = x(u(t), v(t), λ) contained in the congruence. We are particularly interested
in developable ruled surfaces: The developability condition reads

u 2t [eu , au , e] + u t vt ([eu , av , e] + [ev , au , e]) + vt2 [ev , av , e]

  
[eu , au , e] [eu , av , e] + [ev , au , e] vt
= (u t , vt ) = 0, (1)
symm. [ev , av , e] ut

if we use subscripts to indicate differentiation and square brackets for the determinant.
Equation (1) tells us that for any (u, v) there are up to two so-called torsal directions
u t : vt which belong to developable surfaces. This behaviour is quite analogous to the
fact that for any point in a smooth surface there are two principal tangent directions
which belong to principal curvature lines. By integrating the torsal directions one
creates ruled surfaces which are developable, which is analogous to finding principal
curvature lines by integrating principal directions.

(a) (b)

Fig. 1 (a) A line congruence L is described by a surface a(u, v), and direction vectors e(u, v).
(b) Developables R1 , R2 contained in L . The set of all regression curves ci of these developables
makes up the focal sheets F1 , F2 of the congruence (here only F1 is shown). The tangent planes of
R1 , R2 along the common line are the torsal planes or focal planes of that line. These images are
taken from [21]
270 X. Sun et al.

Normal Congruences.
The normals of a surface constitute the normal congruence of that surface. For such
congruences the analogy between torsal directions and principal directions men-
tioned above is actually an equality: The surface normals along a curve form a
developable surface if and only if that curve is a principal curvature line [14].
The reference surface a(u, v) might be the base surface the lines of L are orthog-
onal to, but this does not have to be the case. The congruence does not change if
the reference surface is changed to a∗ (u, v) = a(u, v) + λ(u, v)e(u, v), so deciding
whether or not L is a normal congruence depends on existence of an alternative
reference surface a∗ orthogonal to the lines of L , i.e., e, au∗  = e, av∗  = 0. Assum-
ing without loss of generality that e(u, v) = 1 and using e, eu  = e, ev  = 0 the
orthogonality condition reduces to λu = −au , e, λv = −av , e. This PDE for the
function λ has a solution if and only if the integrability condition λuv = λvu holds.
It is easy to see that this is equivalent to

au , ev  = av , eu . (2)

It is not difficult to see that (2) is equivalent to the condition that developables
contained in L intersect at right angles.
Focal surfaces and focal planes.
Loosely speaking, an intersection point of a line in L with an infinitesimally neigh-
bouring line produces a focal point of the congruence L . The rigorous definition of
focal point is a point x(u, v, λ) where the derivatives of x are not linearly independent:
One gets the condition
 
[xu , xv , xλ ] = [eu , ev , e]λ2 + [au , ev , e] + [eu , av , e] λ + [au , av , e] = 0, (3)

i.e., up to two focal points per line. It is not difficult to see that such singularities are
exactly the singularities of developables contained in L , see Fig. 1b. For this reason,
the tangent planes of developables contained in L are called focal planes as well as
torsal planes. Such a focal plane/torsal plane is spanned by a line L(u, v) together
with a torsal direction.
For normal congruences, the focal points are precisely the principal centers of
curvature; they exist always unless one of the principal curvatures is zero. In each
point of the surface, the focal plane (i.e., torsal plane) is spanned by the surface
normal and a principal tangent.
Example: Congruences defined by linear interpolation.
Congruences of the special form
   
x(u, v, λ) = (1 − λ) a0 + a10 u + a20 v + λ b0 + b10 u + b20 v
   
= a0 + a10 u + a20 v + λ e0 + e10 u + e20 v (4)
Vertex Normals and Face Curvatures of Triangle Meshes 271

play an important role, both for us and in other places: for example, the set of lines
described by such a congruence is the one generated by Phong shading, when one
linearly interpolates vertex normals in a triangle.
We consider the planes “Pα ” which are defined as the set of all points x(u, v, α),
and we study the affine mappings

φαβ : Pα → Pβ , x(u, v, α) → x(u, v, β).

The lines L(u, v) of the congruence are precisely the lines which connect points
x(u, v, α) ∈ Pα and x(u, v, β) ∈ Pβ . These congruences are studied e.g. in [16, Ex.
7.1.2]. Let us summarize some of their properties, which are illustrated by Fig. 2.

(i) Each intersection line L = Pα ∩ Pβ of two planes in the family Pλ is contained

in the congruence L . This follows from the fact that L is spanned by the points
−1
X = φαβ (L) ∩ L and φαβ (X ) = L ∩ φαβ (L).
(ii) The lines Pα ∩ Pβ with α fixed, constitute a developable surface Rα ⊂ L
which is planar and contained in Pα (in general, it is the tangent surface of
a parabola rα ).
(iii) For properties of the focal surface, see Fig. 2.

(a) (b)

Fig. 2 Congruences defined by a “linear” volumetric parametrization x(u, v, λ) turn out to be useful
for linear interpolation of triangle meshes, but they have counter-intuitive properties. (a) Planes Pλ
defined by λ = const are visualized as triangles. Interestingly, all of these triangles contain a planar
developable Rλ ⊂ L with a parabola rλ as curve of regression. In particular the red triangle Pλ1
represents a torsal plane for the blue line L( 13 , 13 ) which connects the barycenters of triangles Pλ .
The image further shows many lines Pλ1 ∩ Pβ , of the planar developable R λ1 . (b) The focal surface
F of L agrees with the envelope of the family of planes Pλ . It is in general the tangent surface of a
cubic polynomial curve r. We show in red and yellow the two sheets of this tangent surface F which
are separated by the regression curve r. We also indicate the point of tangency Tλ where Pλ touches
r. The hyperbolic congruence lines (those which are contained in two focal planes) are bitangents
of the focal surface, i.e., they touch F in two points. The regression parabolas rλ are contained in
F and are obtained by intersecting F with one of its tangent planes Pλ
272 X. Sun et al.

3 Discrete Normal Congruences

Wang et al. [21] define discrete congruences by means of correspondences between

combinatorially equivalent triangle meshes A, B with vertices {ai } and {bi }. Each
pair of corresponding triangles ai a j ak and bi b j bk defines, via linear interpolation, a
piece of a smooth line congruence of the kind described by Eq. (4):

x(u, v, λ) = a(u, v) + λe(u, v),

a(u, v) = ai + ua ji + vaki , e(u, v) = ei + ue ji + veki , where
ei = bi − ai , ai j = ai − a j , ei j = ei − e j . (5)

If the domain is restricted to u ≥ 0, v ≥ 0, u + v ≤ 1, then the correspondence

x(u, v, 0) −→ x(u, v, 1) is precisely the affine mapping of triangle ai a j ak to triangle
bi b j bk . Equations (1) and (3) serve to compute torsal directions and focal points of
this congruence, and also to trace the developables contained in this congruence (see
Fig. 3).
Discrete normal congruences—Version 1.
It is not straightforward to define which correspondence between triangle meshes
defines a normal congruence. Firstly this is because congruences of the form (4)
are never normal except for degenerate cases. Secondly such a normal congruence
would automatically lead to a good definition of constant-distance offset mesh of a
triangle mesh which is lacking so far.
We discuss two suitable definitions of “normal congruence” and start with a
version already published. Wang et al. [21] require normality to hold only in the
barycenters of faces (i.e., they require that Eq. (2) holds for barycenters of faces),
see Fig. 4. Figure 5 shows an example demonstrating the efficiency of this definition.
Proposition 3.1 below gives an equivalent analytic condition.

(a) (b)

Fig. 3 A piecewise-linear correspondence between meshes A and B defines a piecewise-smooth

congruence L . (a) Integrating torsal directions yields corresponding polylines in meshes A and B.
(b) Connecting corresponding points of those two polylines yields a piecewise-flat developable
R ⊂ L . These images are taken from [21]
Vertex Normals and Face Curvatures of Triangle Meshes 273

Fig. 4 Congruences defined by the piecewise-affine correspondence of meshes A, B can be called

discrete-normal, if the normality condition is fulfilled for barycenters of faces. This figure also
illustrates the auxiliary projection used by Eq. (7). This normality condition is called ‘version 1
normality’ here (image taken from [21])

Fig. 5 We demonstrate that Eq. (7) is a working definition of normality: Given a triangle mesh
{ai } (white), we find unit vectors ei by optimizing for shading effects according to Wang et al. [21]
under the normality constraint (7). Subsequently we check if a triangle mesh {ai∗ } orthogonal to
the congruence can be found. We let ai∗ = ai + λi ei and solve for λi such that the faces of the new
mesh are orthogonal to the congruence in their barycenters. The result of this computation yields
a mesh {ai∗ } (yellow) where face normals and congruence lines (in face barycenters) differ by an
angle β, which assumes a maximum of 4.1◦ , a mean of 0.9◦ , and a median of 0.8◦ . Instead of the
mesh computed here, any constant-distance offset would have been a solution as well. We chose
one which lies at a small distance from the original mesh

Proposition 3.1 Consider two combinatorially equivalent triangle meshes and the
line congruence L defined by the piecewise-linear correspondence of faces. For
each pair a1 a2 a3 , b1 b2 b3 of corresponding faces perform orthogonal projection in
direction of the line which connects their respective barycenters, yielding triangles
ā1 ā2 ā3 , b̄1 b̄2 b̄3 . Then L is normal in the barycenters of the two faces if and only if
the following analogue of (2) holds:
274 X. Sun et al.

ā j − āi , b̄k − b̄i  = āk − āi , b̄ j − b̄i , or equivalently, (6)

ā j − āi , ēk − ēi  = āk − āi , ē j − ēi , where ei = bi − ai . (7)

It is sufficient that these conditions hold for at least one choice of indices i, j, k ∈
{1, 2, 3}, i = j = k.

Discrete normal congruences—Version 2.

There is an obvious analogy between conditions (2) and (7): they express normality in
the smooth and discrete cases respectively. However Eq. (7) is not entirely satisfying
as a definition since it involves a projection operator. It is therefore natural to define
discrete-normality by the following two equations which replace Eqs. (6), (7):

a j − ai , bk − bi  = ak − ai , b j − bi  or, equivalently, (6∗ )

a j − ai , ek − ei  = ak − ai , e j − ei . (7∗ )

We will show that theses conditions are suitable to define normality of discrete
congruences defined by a correspondence of triangle meshes. Besides numerical
experiments (see later), we show geometric properties of congruences which fulfill
these conditions. The first property is a discrete version of the following two facts (i)
A normal congruence L has a 1-parameter family of surfaces orthogonal to it, and
(ii) for any point in such a surface there are 3 mutually orthogonal planes spanned
by the normal and the two principal directions. We show that in the discrete-normal
case, there are analogous principal trihedra:
Proposition 3.2 Consider two combinatorially equivalent triangle meshes and the
line congruence L defined by the piecewise-affine correspondence of faces, and
consider in particular one such pair a1 a2 a3 , b1 b2 b3 of corresponding faces. In the
generic case, the normality condition (6∗ ) implies the following property:
For each plane Pλ spanned by the vertices (1 − λ)ai + λbi there is a congruence
line Nλ = L(u λ , vλ ) such that the two focal planes of that line together with Pλ form
a trihedron of mutually orthogonal planes.
The meaning of “generic” is discussed in the proof.
Proof Generically, vectors ei = bi − ai are linearly independent, so we can express
a normal
3 vector n of the triangle  a1 a2 a3 (which spans P0 ) as a linear combination
n = i=1  αi ei . Generically, αi = 0, so by multiplying n with a factor we can
achieve αi = 1 and by relabeling the coefficients αi we get n = (1 − u − v)e1 +
ue2 + ve3 . Then Equation (5) shows that the line L(u, v) is orthogonal to P0 .
Consider the affine correspondence of triangles a1 a2 a3 and b1 b2 b3 followed by
orthogonal projection onto P0 . A vertex ai is mapped to b̄i = bi + λi n. There is a
linear mapping α with α(ai − a j ) = b̄i − b̄ j . It is clear from Fig. 3 that the eigenvec-
tors of α indicate the directions of torsal planes through the line L(u, v). Conditions
(6∗ ), (7∗ ) imply
Vertex Normals and Face Curvatures of Triangle Meshes 275

a j − ai , b̄k − b̄i  − ak − ai , b̄ j − b̄i 

= a j − ai , bk + λk n − bi − λi n − ak − ai , b j + λ j n − bi − λi n
= a j − ai , bk − bi  − ak − ai , b j − bi  = 0,

i.e., symmetry of α and orthogonality of eigenvectors of α. This shows orthogonality

of torsal planes and verifies the statement for the case λ = 0. The case λ = 1 is
analogous, since condition (6∗ ) is invariant if we replace ai by bi and vice versa.
For all other values λ = 1 we note that replacing vertices bi by vertices ai + λei
inflicts the change ei → λei without changing ai , which does not affect the normality
condition (7∗ ).

As illustrated by Fig. 2, congruences defined by the affine correspondence of trian-

gles have counter-intuitive properties: The planes Pλ generated by linear interpolation
of the defining triangles at the same time are the focal planes of L (and vice versa)
since any Pλ carries the developable surface generated by the lines {Pλ ∩ Pα }α∈R .
The torsal planes Pλ are tangent to the focal surface F of L . It is known that F is
the tangent surface of a cubic polynomial curve, cf. [16, Ex. 7.1.2]. Proposition 3.2
now tells us that this curve has infinitely many triples of mutually orthogonal tan-
gent planes. Translating these planes (the principal trihedra) through the origin, they
become tangent planes of the directing cone of F, which is a quadratic cone. This
cone is quadratic and must likewise have infinitely many orthogonal circumscribed
trihedra. It is therefore a so-called Monge cone, see Fig. 6.
There is a phenomenon in geometry, called porism, cf. [7]. It refers to situa-
tions where existence of one object of a certain kind implies existence of an entire
1-parameter family of such objects. Monge cones are an instance of a porism: If a
quadratic cone has one circumscribed orthogonal trihedron, then one can move this
trihedron around the cone while it remains tangential. This fact is classical knowledge
in projective geometry, see e.e. [1, pp. 33–34].

Fig. 6 The “principal” trihedra mentioned in Proposition 3.2, when moved to the origin, lie tangent
to a so-called Monge cone. Since these planes rotate about an entire cone as the interpolation
parameter λ varies, one cannot without restrictions interpret these principal trihedra as tangent
planes plus principal planes of an offset family of surfaces. Such an interpretation is valid only for
small λ
276 X. Sun et al.

The same porism is hidden in the proof of Proposition 3.2: The normality condition
(6∗ ) was equivalent to existence of the principal trihedron associated with P0 , but it
also implied existence of the trihedron for all Pλ .
Details on principal trihedra in discrete-normal congruences.
We wish to interpret the three mutually orthogonal planes referred to by Proposition
3.2 as the tangent plane and principal planes of a surface. In particular the normal
vector nλ of Pλ shall be the normal vector, and the line Nλ shall be the surface
normal, while the torsal planes should represent the principal directions. In order
to understand better the behaviour of the objects involved, we study the volumetric
parametrization according to Eq. (4) in an adapted coordinate system: the plane P0
is the x y plane, and the two torsal planes associated with it shall be the x z and zy
planes. Since the affine correspondence between planes P0 , P1 may be defined by
any pair of corresponding triangles, we choose a1 = o, a2 = (1, 0, 0), a3 = (0, 1, 0).
We may still change the plane P1 without changing the congruence, so we choose
b1 = (0, 0, 1). The vertices b2 , b3 must lie in the x y and x z planes because of our
assumption on the torsal planes. Thus we get
⎛ ⎞ ⎛ ⎞
u −κ1 u
x(u, v, λ) = ⎝ v ⎠ + λ ⎝ −κ2 v ⎠
0 au + bv + 1
⎛ ⎞
aλ(κ2 λ − 1)
∂x ∂x ⎝ ⎠.
=⇒ nλ = × = bλ(κ1 λ − 1) (8)
∂u ∂v (κ λ − 1)(κ λ − 1)
1 2

We will later interpret κ1 , κ2 as principal curvatures and vectors (1, 0, 0) and (0, 1, 0)
as principal directions. Obviously, they are eigenvectors of the linear map α which
occurs in the proof of Proposition 3.2. The plane Pλ is given as

n 1,λ x1 + n 2,λ x2 + n 3,λ x3 − n 0,λ = aλ(κ2 λ − 1)x1 + bλ(κ1 λ − 1)x2

+ (κ1 λ − 1)(κ2 λ − 1)x3 − λ(κ1 λ − 1)(κ2 λ − 1) = 0.

This is a cubic family of planes. Translating them through the origin yields the
planes n 1,λ x1 + n 2,λ x2 + n 3,λ x3 = 0, which are tangent planes of the tangent cone
illustrated in Fig. 6. Since the plane coefficients satisfy the quadratic equation (κ1 −
κ2 )n 1 n 2 − an 2 n 3 + bn 1 n 3 = 0, it is indeed a quadratic cone.1
We now look for a line L(u λ , vλ ) orthogonal to Pλ . The direction of L(u, v) can
be read off (8), so the condition L(u λ , vλ )  nλ reads

1 The vector of coefficients (n 1 , n 2 , n 3 ) of the equation of a plane is a normal vector of that plane.
This shows that the orthogonal polar cone of the Monge cone fulfills the equation (κ1 − κ2 )x1 x2 −
ax2 x3 + bx1 x3 = 0. Since the Monge cone had many circumscribed orthogonal trihedra, its polar
cone has many inscribed orthogonal frames. These frames are generated by translating the frames
seen in Fig. 7b through the origin.
Vertex Normals and Face Curvatures of Triangle Meshes 277

(a) (b)

Fig. 7 Behaviour of the principal trihedron and the normal Nλ of planes Pλ in a congruence defined
by the affine correspondence between two triangles. (a) The normals Nλ (green) intersect the plane
P0 in the points c(u λ , vλ , 0) of a conic (red). (b) As λ changes, the apex cλ = x(u λ , vλ , λ) of the
principal trihedron (yellow) moves along a straight straight line (blue). The ruled surfaces traced
out by the edges of the trihedron are shown; their union forms one algebraic ruled surface of degree
four

κ1 u λ a(κ2 λ − 1) κ1 u λ aλ
= , =
κ2 vλ b(κ1 λ − 1) au λ + bvλ + 1 1 − κ1 λ
aλκ2 (1 − κ2 λ) bλκ1 (1 − κ1 λ)
=⇒ u λ = , vλ = ,
νλ νλ
where νλ = κ1 κ2 (κ1 λ − 1)(κ2 λ − 1) + a 2 κ2 λ(κ2 λ − 1) + b2 κ1 λ(κ1 λ − 1). (9)

In particular we see that the curve x(u λ , vλ , 0), consisting of all points Nλ ∩ P0 , is
a conic. In fact, for every α, the curve {Nλ ∩ Pα }λ∈R is a conic it corresponds to the
curve Nλ ∩ P0 under the affine mapping φ0α : x(u, v, 0) → (u, v, α), see Fig. 7a.
The surface of all Nλ ’s is then algebraic of 4◦ .
Let us now compute the “apex” cλ = Nλ ∩ Pλ = x(u λ , vλ , λ) of the principal
trihedron: From
⎛ ⎞
κ a
λ(1 − κ1 λ)(1 − κ2 λ) ⎝ 2 ⎠
cλ = κ1 b (10)
νλ κκ 1 2

we see that cλ moves on a straight line, but the parametrization of this line is cubic.
Since the planes Pλ and the torsal planes stem from the same 1-parameter family
of planes, any torsal plane will play the role of Pλ for another value λ ; in total
each orthogonal trihedron will occur three times, and each of the three edges of the
trihedron will play the role of Nλ three times (see Fig. 7b). We summarize:
Proposition 3.3 If a congruence is defined by the affine correspondence between
two triangles a1 a2 a3 and b1 b2 b3 and satisfies the normality condition (6∗ ), then
its focal surface has a 1-parameter family of circumscribed ‘principal’ orthogonal
trihedra whose apex moves on a straight line and whose edges form an algebraic
surface of degree 4 which contains that line as a triple line.
The complicated geometry of these congruences reflects the difficulties in defining
offset pairs of triangle meshes.
278 X. Sun et al.

Discrete normal congruences — Version 3.

An elementary computation shows that either of the two conditions (6∗ ), (7∗ ) is
implied by the stronger condition

a j − ai , e j + ei  = 0, (11)

when imposed on all three edges of a triangle. This third version of normality is a
more direct expression of the orthogonality between triangle mesh and congruence:
the edges ai a j of the mesh are required to be orthogonal to the arithmetic mean of
normal vectors ei , e j at either endpoint of the edge.
Comparison of definitions.
The various definitions of discrete normal congruences have different advantages.
When one wants to design a normal congruence (as in Wang et al. [21]), version 1
may be better because it ensures orthogonality of focal planes in the part of the line
congruence which is actually realized. Using version 2, orthogonal focal planes may
occur outside the realized part. On the other hand, when using the normal congruence
of a given surface, version 2 has the advantage that one plane of a principal frame
contains the base mesh triangle; moreover discrete principal directions are orthogonal
and lie in the plane of the triangle. Version 3 normality is not used here except for
Fig. 8 where we show that imposing version 3 normality leads to results comparable
to version 2. Since the weaker condition of version 2 is sufficient to achieve the same
results, it is not necessary to impose version 3 normality.

Fig. 8 Optimization of normal congruences. For a given mesh with vertices ai , a discrete-normal
congruence, defined by unit vectors ei , has been found by global optimization such that one of the
normality conditions considered here is fulfilled. Each of these conditions is linear, so optimization
was done by least squares. It turns out that there is no substantial difference between Eqs. (6∗ )
and (11). Faces are colored according to the angle β enclosed between the congruence line at the
barycenter and the face’s normal there. We also give statistics on β for each figure
Vertex Normals and Face Curvatures of Triangle Meshes 279

4 Curvatures of Faces of Triangle Meshes

Recall that a smooth normal congruence L possesses a surface A orthogonal to the

lines of L . Then automatically all offsets At also lie orthogonal to L . We assume
labeling of offsets such that surfaces At , As are at constant distance |t − s| from
each other. Then corresponding infinitesimal surface area elements “dAt (u, v)” obey
Steiner’s formula

dAt (u, v)
= 1 − 2t H (u, v) + t 2 K (u, v), (12)
dA0 (u, v)

where H and K denote mean and Gaussian curvature of the surface A0 , respectively.
The sign of H depends on the unit normal vector field; in our case the unit normal
vector field points from A0 to the surfaces At with t > 0.
We now return to a discrete congruence L defined by the piecewise-linear cor-
respondence between triangle meshes A, B. Assuming A, B approximate an offset
pair of surfaces at distance 1, we consider corresponding faces a1 a2 a3 and b1 b2 b3 .
We write bi = ai + ei , where the vectors ei approximate unit normal vectors of the
mesh A. An offset mesh at distance approximately t then has vertices and faces

ait = ai + tei Δt = a1t a2t a3t .

We further assume that the congruence L is a normal congruence (which we have

defined in two different ways).
• If L is normal in the sense of Eqs. (6) and (7), then we apply the projection
mentioned in Proposition 3.1, resulting in vertices ā1 ā2 ā3 , b̄1 b̄2 b̄3 . The projection
is in the direction of a certain unit vector n.
• As an alternative, the congruence may be normal in the sense of Eqs. (6∗ ), (7∗ ).
Here we consider orthogonal projection onto the plane P0 which contains a1 a2 a3 .
This projection results in vertices āi = ai and b̄i . The projection is in direction of
the unit normal vector n = n0 of the plane P0 .

We now study the behaviour of the area of the face Δt as t changes. We do not measure
the actual area, but apply the projection just mentioned. The area of projected triangles
is measured via a determinant in the plane:

1 1 1
p-area(x1 x2 x3 ) = [x̄2 − x̄1, x̄3 − x̄1 ] = [n, x̄2 − x̄1, x̄3 − x̄1 ] = [n, x2 − x1, x3 − x1 ]
2 2 2

With the notation āi j = āi − ā j , b̄i j = b̄i − b̄ j , ēi j = b̄i j − āi j we get

2 [ā12 + t ē12, ā13 + t ē13 ]

1
p-area(t ) [ā12, ē13 ] + [ē12, ā13 ] [ē12, ē13 ]
= =1+t + t2 .
p-area(0 ) 2 [ā12 , ā13 ]
1 [ā12, ā13 ] [ā12, ā13 ]
280 X. Sun et al.

Discrete curvatures and shape operator.

The obvious similarity of this relation with (12) immediately leads to a definition of
the mean curvature H and the Gauss curvature K of the face a1 a2 a3 under consid-
eration:
[n, e12, e13 ] [n, a12, e13 ] + [e12, a13 ]
K = , 2H = − . (13)
[n, a12, a13 ] [n, a12, a13 ]

Principal curvatures κ1 , κ2 are defined by the relations

κ1 + κ2 = 2H, κ1 κ2 = K .

Completing the analogy with the smooth case, we define a shape operator Λ as the
linear mapping which maps

Λ
āi − ā j −→ −(ēi − ē j ), for all i, j ∈ {1, 2, 3}.

Recall that the bar indicates projection (which in turn depends on which version
of “normality” we employ). In analogy to the smooth case, principal directions are
given by the focal planes of the congruence L . All these notions fit together:

Proposition 4.1 The eigenvalues of the shape operator Λ are the principal curva-
tures κ1 , κ2 , and its trace and determinant are given by 2H and K , respectively.
Eigenvectors of Λ indicate the principal directions.

Proof We first show the statement for ‘version 2’ normality. Recall the linear map-
α
ping α in the proof of Proposition 3.2 which maps āi − ā j −→ (āi + ēi ) − (ā j + ē j ).
Since by construction, Λ = id − α, Λ has the same eigenvectors as α, i.e., the
torsal directions. The statement about tr Λ and det Λ follows from the relations
det Λ = det(Λ(x),Λ(y))
det(x,y)
and tr Λ = det(Λ(x),y)+det(x,Λ(y))
det(x,y)
which generally hold for linear
mappings of R . The statement about eigenvalues follows immediately.
2

For version 1 normality the proof is the same, only the bars have a different
meaning. The mapping α is also referred to in the proof of Proposition 3.1 in [21].

Since we have defined principal curvatures κ1 , κ2 implicitly via mean curvature

H and Gauss curvature K , their relation to focal geometry is still unclear. In the
smooth case, points at distance 1/κi from the surface are focal points of the normal
congruence. This property holds in the discrete case too, if we use version 2 normality:

Proposition 4.2 Consider a congruence with parametric representation x(u, v, λ)

which is defined by the correspondence of two triangles a1 a2 a3 and b1 b2 b3 . Assume
that it is normal in the sense of Eq. (6∗ ), and consider (in the notation of Proposition
3.2) the plane P0 which contains a1 a2 a3 and the corresponding normal L(u 0 , v0 ).
Then the focal points of that line lie at distance 1/κ1 , 1/κ2 from the plane P0 , with κi as
the principal curvatures, i.e., the focal points are precisely the points x(u 0 , v0 , 1/κi ).
Vertex Normals and Face Curvatures of Triangle Meshes 281

Proof We consider the parametrization (8) which is with respect to an adapted coordi-
nate system, so that u 0 = 0 and v0 = 0. It is easy to see that the values κ1 , κ2 occurring
there are indeed the principal curvatures. A simple computation shows that for the
special case u = v = 0, the determinant of partial derivatives of x(u, v, λ) specializes
to [xu , xv , xλ ] = (1 − λκ1 )(1 − λκ2 ). Thus we have a singularity if λ = 1/κi .

Special cases.
An umbilic point is characterized by equality of principal curvatures, i.e., κ1 = κ2 =
κ. In this case some of the geometric objects discussed above simplify. E.g. the above-
mentioned cubic family of planes becomes the set of tangent planes of a quadratic
cone with vertex (0, 0, 1/κ). Such an umbilic occurs every time two corresponding
triangles a1 a2 a3 and b1 b2 b3 are in homothetic position, but the converse is not true.
A parabolic point is characterized by one principal curvature, say κ1 , being zero. In
this case, Eq. (8) immediately shows that the congruence vectors e1 , e2 , e3 associated
with vertices a1 , a2 , a3 are not linearly independent, so Proposition 3.2 does not
apply. Along the x axis, the lines of the congruence are parallel to each other, which
is in accordance with the fact that the focal point (0, 0, 1/κ1 ) has moved to infinity.
The above-mentioned cubic family of planes is quadratic (in fact, it is the family of
tangent planes of a parabolic cylinder).

Remark 4.3 We should mention that the approach to curvatures presented here car-
ries over to relative differential geometry where the image of the Gauss map is not
a sphere but a general convex body [19]. Another straightforward extension is to
curvatures at vertices, which however does not lead to a shape operator in such a
natural manner.

5 Results and Discussion

Numerical examples. Vertex normals of a mesh can be estimated (e.g. as area-

weighted averages of face normals). Any such collection of sensible normals is not
far away from being a “normal” congruence in our sense. By applying optimization,
we can make it as normal as possible, meaning that (6) is fulfilled in the least-squares
sense. Numerical experiments show that this improves the quality of the normal field
(even if there are not enough d.o.f. to satisfy (6) fully if the vertices of the mesh
are kept fixed). Since curvatures and the distribution of normals are inseparable, it
makes sense to study curvatures not only as quantities derived from a mesh, but as
quantities which arise naturally from the the result of the optimization procedure just
mentioned. In this way the natural sensitivity of curvatures with respect to noise is
moderated.
The basic task is, of course, the computation of a normal congruence for a given
mesh. This is done via a standard optimization procedure, which is initialized from
estimated vertex normals. We express the validity of the normality condition in terms
282 X. Sun et al.

of least squares, and minimize subject to the constraints that (in the terminology
of previous sections), vectors ei are of unit length. Figure 8 shows an example. In
particular one can see that normality according to Eq. (6) (“version 1”) behaves
differently from normality according to Eq. (6∗ ) (“version 2”), while there is hardly
any difference between conditions (6∗ ) and (11).
Degrees of freedom and topology. When optimizing a normal congruence of a mesh
with v vertices, e edges and f faces, we count 3v variables for the normals and
f + v constraints. If a number b of boundary vertices is present, we fix the normals
at the boundary, resulting in 3(v − b) variables and f + (v − b) constraints, i.e.,
2v − f − 2b d.o.f. Elementary manipulations show that

d.o.f. = 2χ − b,

with χ = f + v − e as the Euler characteristic. We see for meshes of sphere topology

we can expect a unique solution, but topological features diminish the available
degrees of freedom. If boundary normals are kept fixed, long boundaries diminish
this freedom even more. By allowing vertices to move during optimization, we can
achieve zero residual again, but of course a compromise has to be found between
the quality of the normal congruence and the deviation of the mesh from its previous
shape. Table 1 shows some numerical experiments.
Computing Curvatures. Once a normal congruence is available, we can compute
curvatures (see Fig. 9) and we can integrate the field of principal curvature directions
as well as the field of asymptotic directions (see Fig. 10 for an example). It must
be said, however, that we do not want to compete with the many other methods for
computing curvatures, and we do not regard the ability to compute curvatures a main
result of this study.
Robustness by using normals. Fig. 11 demonstrates that considering a mesh and its
normal congruence together allows us to handle optimization/smoothing in a stable
way. After a mesh and its normals have been perturbed (Fig. 11b), an optimization
procedure attempts to restore both. We use a target functional composed of a sum
of least squares expressing condition (6∗ ) and also the property of vectors ei having

Table 1 Comparison of residuals regarding normalcy of the congruence (“c”) and unit vectors
being normalized (“n”) when optimizing congruences
Sphere Torus Disk w/holes, see Fig. 10
Fixed vertices Fixed vertices Moving vertices Fixed vertices Moving vertices
c n c n c n c n c n
v. 1 7.8 × 10−3 0 7.7 × 100 0 – – 1.5 × 10 0 0 – –
v. 2 9.7 × 10−5 0 9.6 × 10−1 0 6.9 × 10−5 8.1 × 10−7 1.9 × 10−2 0 4.0 × 10−5 2.2 × 10−9
v. 3 9.0 × 10−2 0 1.3 × 10−1 0 6.9 × 10−4 6.3 × 10−4 2.4 × 10−1 0 9.6 × 10−5 9.5 × 10−10
All meshes are normalized for unit average edge length, and a zero means a zero up to machine
precision. The rows in this table correspond to versions 1, 2, 3 of the normalcy condition for
congruences. One can see that zero residual happens only for sphere topology
Vertex Normals and Face Curvatures of Triangle Meshes 283

Fig. 9 Computing mean curvature H and Gaussian curvature K by means of normal congruences:
“version 1” and “version 2” refer to normality defined by Eqs. (6) and (6∗ ), respectively. Estimated
normals are optimized so as to become a normal congruence which allows us to compute curvatures
in faces. For comparison, curvatures computed by a 3rd order jet fit have been used, cf. [4]. The
color scale is the same for each kind of curvature and each model, throughout the 3 methods of
computation. One can hardly see any difference. For each mesh, normal congruences have been
computed in the way employed for Fig. 8

Fig. 10 We compute asymptotic lines and principal curvature lines of meshs by various means.
For the figures of the first column, we have used the 3rd order jet fit method of [4]. For the second
column, we used the method of normal cycles (see e.g. [20]). The 3rd and 4th column are computed
using our the shape operators, where version 1 and version 2 refer to normality w.r.t. Eqs. (6),
(6∗ ), respectively. In both cases the normal congruence needed for defining the shape operator was
obtained in the same way as for Fig. 8
284 X. Sun et al.

Fig. 11 Computing normals and principal curvature lines for noisy data. Subfigure a shows a
triangulated cylinder and some of its principal curvature lines. In b the jet fit method has been
used to obtain principal curves for data where noise has been added to both vertex coordinates and
normals. Subfigure c shows the result of optimization applied to (b), which results in a smooth mesh
equipped with a normal congruence. For c we again show the principal curvature lines computed
by our method

length 1 (weight 1), proximity to the input data (weight 1/4), Laplacian fairing for
the mesh (weight 10−6 ), Laplacian fairing for the normal vectors (weight 10−4 ) and
compatibility between normal and mesh by penalizing deviation from orthogonality
between congruence lines in mesh barycenters and face (weight 10−4 ). Figure 11c
shows the repaired mesh.
Relevance for discrete differential geometry.
The idea of employing the Steiner formula for defining curvatures has proved very
helpful in bringing together various different notions of curvature, and indeed, various
different notions of discrete surfaces (like discrete minimal surfaces and discrete cmc
surfaces) which were defined in a way not involving curvature directly but by other
means like Christoffel duality. We refer to [2, 3] for more details. The theory presented
in [2] is restricted to offset-like pairs of polyhedral surfaces where corresponding
edges and faces are parallel. There are ongoing efforts to extend this theory to more
general situations (we point to recent work on quad meshes [10] and on isothermic
triangle meshes of constant mean curvature [11]). It is therefore remarkable that at
least for the situation described here, triangle meshes allow an approach to curvatures
and even a shape operator which is likewise guided by the Steiner formula, but without
the rather restrictive property of parallelity (which for triangle meshes would be even
more restrictive).
Future Research.
As to discrete differential geometry, it is still unclear how known constructions of
special discrete surfaces relate to the curvatures defined here: For instance, it seems
difficult to gain nice geometric properties from the condition vanishing mean cur-
vature. Nevertheless one of the known constructions of discrete minimal surfaces
might be equipped with a canonical normal congruence such that, when our theory
is applied, mean curvature vanishes.
Further applications of line congruences have been discussed by Wang et al. [21],
but there might be other examples of geometry processing tasks where the notion of
line congruence, or even normal congruence, becomes relevant.
Vertex Normals and Face Curvatures of Triangle Meshes 285

Acknowledgments The authors are grateful to Alexander Bobenko for fruitful discussions and to
the anonymous reviewers for their suggestions. This research was supported by the DFG Collabo-
rative Research Center, TRR 109, “Discretization in Geometry and Dynamics” through grants I705
and I706 of the Austrian Science Fund (FWF).

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Baker, H.F.: Principles of Geometry, vol. II, 2nd edn. Cambridge University Press (1930)
2. Bobenko, A., Pottmann, H., Wallner, J.: A curvature theory for discrete surfaces based on mesh
parallelity. Math. Annalen 348, 1–24 (2010)
3. Bobenko, A., Suris, Yu.: Discrete Differential Geometry: Integrable Structure. American Math-
ematical Society (2009)
4. Cazals, F., Pouget, M.: Estimating differential quantities using polynomial fitting of osculating
jets. In: Kobbelt, L., Schröder, P., Hoppe, H. (eds.) Proceedings of Symposium on Geometry
Processing, pp. 177–187. Eurographics Association (2003)
5. Desbrun, M., Meyer, M., Schröder, P., Barr, A.: Discrete differential-geometry operators for
triangulated 2-manifolds. In: Visualization & Mathematics, vol. 3, pp. 35–57. Springer (2003)
6. Doliwa, A., Santini, P., Mañas, M.: Transformations of quadrilateral lattices. J. Math. Phys.
41, 944–990 (2000)
7. Dragović, V., Radnović, M.: Poncelet porisms and beyond. Birkhäuser, Basel (2011)
8. Grinspun, E., Gingold, Y., Reisman, J., Zorin, D.: Computing discrete shape operators on
general meshes. Comput. Graph. Forum 25(3), 547–556 (2006). Proc. Eurographics
9. Hildebrandt, K., Polthier, K.: Generalized shape operators on polyhedral surfaces. Comput.
Aided Geom. Des. 28, 321–343 (2011)
10. Hoffmann, T., Sageman-Furnas, A., Wardetzky, M.: A discrete parametrized surface theory in
R3 . Int. Math. Res. Not. (2016), to appear
11. Lam, W.Y., Pinkall, U.: Isothermic triangulated surfaces. Math. Annalen (2016), to appear
12. Liu, Y., Pottmann, H., Wallner, J., Yang, Y.L., Wang, W.: Geometric modeling with conical
meshes and developable surfaces. ACM Trans. Graphics 25(3), 681–689 (2006). Proc. SIG-
GRAPH
13. Morvan, J.M.: Generalized Curvatures. Springer (2008)
14. Porteous, I.R.: Geometric Differentiation for the Intelligence of Curves and Surfaces. Cam-
bridge University Press (1994)
15. Pottmann, H., Liu, Y., Wallner, J., Bobenko, A., Wang, W.: Geometry of multi-layer freeform
structures for architecture. ACM Trans. Graphics 26(3), # 65,1–11 (2007). Proc. SIGGRAPH
16. Pottmann, H., Wallner, J.: Computational Line Geometry. Springer (2001)
17. Pottmann, H., Wallner, J.: The focal geometry of circular and conical meshes. Adv. Comp.
Math 29, 249–268 (2008)
18. Pottmann, H., Wallner, J., Huang, Q., Yang, Y.L.: Integral invariants for robust geometry
processing. Comput. Aided Geom. Des. 26, 37–60 (2009)
286 X. Sun et al.

19. Simon, U., Schwenck-Schellschmidt, A., Viesel, H.: Introduction to the affine differential
geometry of hypersurfaces. TU Berlin & Science Univ, Tokyo (1992)
20. Sun, X., Morvan, J.M.: Curvature measures, normal cycles and asymptotic cones. Actes des
recontres du C.I.R.M. 3(1), 3–10 (2013). Proc. “Courbure discrète: théorie et applications”
21. Wang, J., Jiang, C., Bompas, P., Wallner, J., Pottmann, H.: Discrete line congruences for shading
and lighting. Comput. Graph. Forum 32(5), 53–62. Proc. Symp, Geometry Processing (2013)
22. Yu, J., Yin, X., Gu, X., McMillan, L., Gortler, S.: Focal surfaces of discrete geometry. In:
Belyaev, A., Garland, M. (eds.) Proceedings of Symposium on Geometry Processing, pp. 23–
32. Eurographics Association (2007)
S-Conical CMC Surfaces. Towards
a Unified Theory of Discrete Surfaces
with Constant Mean Curvature

Alexander I. Bobenko and Tim Hoffmann

Abstract We introduce a novel class of s-conical nets and, in particular, study s-

conical nets with constant mean curvature. Moreover we give a unified description
of nets of various types: circular, conical and s-isothermic. The later turn out to be
interpolating between the circular net discretization and the s-conical one.

1 Introduction

A variety of approaches have been pursued to obtain a notion of discrete constant

mean curvature (cmc) surfaces. Two different starting points arise from the interpre-
tation of cmc surfaces as critical points of an area functional [9, 16], and on the other
hand from an integrable systems point of view [4, 10]. One principal difference of
the two approaches is that, in the first approach, the underlying combinatorial struc-
ture is naturally that of a simplicial surface, whereas the integrable systems approach
demands a quadrilateral structure of the discrete surface, and one reads about discrete
parametrized surfaces. Recently these two approaches were partially merged as soon
as a curvature theory for general polyhedral surfaces based on the notions of parallel
surfaces and mixed area was developed [5].
Restricting to quadrilateral meshes, and, in particular, to planar quadrilaterals
(discrete conjugate nets), may at first sight appear to be a severe restriction. How-
ever, it turns out that every surface can be approximated by discrete nets with these

A.I. Bobenko (B)

Inst. für Mathematik, Technische Universität Berlin,
Straße des 17. Juni 136, 10623 Berlin, Germany
e-mail: [email protected]
T. Hoffmann
Zentrum Mathematik – M10, Technische Universität München,
Boltzmannstr. 3, 85747 Garching bei München, Germany
e-mail: [email protected]

© The Author(s) 2016 287

A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_9
288 A.I. Bobenko and T. Hoffmann

properties [3, 6], reflecting the existence of the corresponding parametrizations in

the smooth case. In fact these integrable discretizations are structure preserving,
i.e. they preserve the key aspects of the geometry. A characteristic feature of sur-
faces described by integrable systems is the existence of special transformations of
Bäcklung and Darboux type that preserve the class of surfaces. In the discrete setup
these transformations lead to consistent multi-dimensional nets. The last property
has established itself as the integrable structure in discrete differential geometry [6].
There exists a rather developed theory of discrete cmc surfaces from an integrable
point of view, reflected in numerous publications. They include Q-nets (nets with
planar quadrilaterals) [5], circular nets (nets with circular quadrilaterals) [4, 6, 8,
10, 11, 14, 18, 19], semi-discrete nets [15], s-isothermic nets [12]. The latter can be
characterized geometrically [6] by having a sphere at every vertex of the net such
that the intersection angles of the spheres along opposite edges in every quadrilateral
are the same and the four spheres either have a common orthogonal circle, share a
pair of points, or (as a limiting case) all meet in exactly one point.
The key observation of the present paper is that the last subclass of s-isothermic
nets is in fact conical. Although the class of conical nets introduced in [13] is very
important for applications and belongs to integrable discrete differential geometry [6,
17], investigation of discrete conical surfaces with constant curvature has just started
[2]. In this paper we introduce a novel class of s-conical nets and, in particular, study
s-conical nets that are cmc. The identification of s-conical surfaces with special
s-isothermic surfaces mentioned above leads not only to a description of conical
cmc nets but also to a unified description of discrete cmc nets. One can think of
s-isothermic nets as “interpolating” between the circular net discretization and the
s-conical one.

2 Conical Nets

Here we consider Q-nets, which are discrete surfaces with planar quadrilateral faces.
Since we are mostly developing a local theory, for simplicity we consider surfaces
with the combinatorics of the square grid f : Z2 → R3 . Some parts of the theory
can be generalized to a more general combinatorics f : G → R3 , where G is a
quad-graph. The latter is a strongly regular cell decomposition of a two-dimensional
manifold with all faces being quadrilaterals. Moreover in the developed theory of
discrete CMC surfaces the quad-graph should be edge-bipartite, i.e. there is a black
and white edge coloring such that for each quadrilateral opposite edges are of the
same color.
Through this paper we will use a notation that indicates shifts in the various
directions by subscript. For a net f : Z2 → R3 we will denote a generic point f (k, l)
simply by f . Then it is understood that f 1 = f (k + 1, l), f 2 = f (k, l + 1), f 12 =
f (k + 1, l + 1), f 1̄ = f (k − 1, l) and so forth. This is of particular use in case of Zn
lattices but also as long as only one or two neighboring quadrilaterals of a quad-graph
are concerned it is a useful shorthand. The following definition first appeared in [13].
S-Conical CMC Surfaces. Towards a Unified Theory . . . 289

Definition 2.1 A Q-net f : Z2 → R3 is said to be conical if for each vertex all

incident faces touch a common sphere σ .
The touching spheres in the Definition 2.1 of conical nets are not unique. In general
there is a 1-parameter family of spheres touching all incident faces for each vertex.
These spheres are inscribed in a cone, and their centers lie on the cone axis. This
furnishes a canonical normal direction (or line) in each vertex of f .
The normal lines of neighboring vertices are coplanar. Due to this fact conical
nets have parallel nets in normal direction. The following theorem as well as all
properties of conical nets mentioned in this section are due to [13] (see also [5, 6]).

Theorem 2.2 Let f : Z2 → R3 be a conical net. There is a 1-parameter family of

conical nets f t : Z2 → R3 such that ( f t − f ) lies in normal direction to f , f and
f t have parallel edges, and the distance between corresponding faces is constant.

Proof Given a conical net f : Z2 → R3 one chooses f t at a vertex f such that f t − f

is in normal direction. Then there is a touching sphere σ with center at f t . Let r
be its radius. Now one can find touching spheres σ of radius r at all other vertices.
Set f t to be their centers. Then ( f f t ) is in normal direction everywhere. Consider
a quadrilateral ( f, f 1 , f 12 , f 2 ). Since the edge ( f f 1 ) lies in the intersection of two
common tangent planes of σ1 and σ it is parallel to the difference of their centers
f 1t − f t as well (the spheres are of equal radius). Thus the corresponding edges (and
hence also the faces) of the quadrilaterals ( f, f 1 , f 12 , f 2 ) and ( f t , f 1t , f 12
t
, f 2t ) are
parallel. Moreover since all the spheres have the same radius r , the distance between
the planes of these quadrilaterals is r . Since the faces of f and f t are parallel f t is
conical as well.

Definition 2.3 Let f : Z2 → R3 be a conical net. n : Z2 → R3 is called a Gauss map

of f if n points in normal direction and the parallel net f 1 = f + n has constant
face offset 1, i.e. the distance between the planes of the corresponding faces of f
and f 1 is equal to 1.

Proposition 2.4 Let f : Z2 → R3 be a conical net and n : Z2 → R3 its Gauss map.

Then the edges of n are parallel to the edges of f and the faces of n touch the unit
sphere S 2 .

Proof Since f + n is parallel to f , n is parallel to f as well. Moreover since the

distance between the faces of f + n and f is equal to 1, the projection of n to the
face normals N of f (and n) is 1. Thus the faces of n touch the unit sphere.

The parallel nets in Theorem 2.2 are then given by

f t = f + tn. (1)

We will refer to these nets as (parallel) nets in constant distance.

We can use the notion of parallel nets to give yet another characterization for
conical nets:
290 A.I. Bobenko and T. Hoffmann

Proposition 2.5 A Q-net f : Z2 → R3 is a conical net if and only if f has parallel

nets with constant face offset. If f is generic the constant face offset nets lie in normal
direction.
Proof Theorem 2.2 shows that conical nets have parallel nets with constant face
offset. Now if f is a net with parallel net f˜ in constant face distance d, then fˆ =
f˜ − f is parallel to f as well, and the faces of fˆ touch a sphere of radius d. Thus
fˆ is conical. This also ensures conicality of all its parallel nets, in particular of f
and f˜. Finally, fˆ = f˜ − f is the cone axes for fˆ and therefore ( f f˜) is the normal
direction for f˜ and f as well.
The following angle characterization of conical nets was proven in [20].
Proposition 2.6 A Q-net f : Z2 → R3 is a conical net if and only if at each vertex
the sums of the opposite angles (of quadrilaterals) are equal.
Example 2.7 Take any planar polygon γ (k) = (γ 1 (k), γ 2 (k)) with non vanishing
edges and an angle 0 < φ < π. Then the discrete rotational net

f (k, l) = (γ 1 (k), cos(lφ)γ 2 (k), sin(lφ)γ 2 (k))

is a conical net.

3 Curvatures of Conical Nets via Steiner’s Formula

The classical Steiner formula couples the areas of a surface f and a parallel offset
surface f t with the mean and Gauss curvature of f . If f is an infinitesimal surface
patch and f t the parallel one in distance t in normal direction, then Steiner’s formula
gives:
A( f t ) = A( f )(1 − 2H t + K t 2 ), (2)

where A is the area, and H and K are the mean and the Gauss curvatures of f .
A discrete analogue of this formula was used in [5] to define curvatures for Q-nets
with a given Gauss map (see also [18] where this formula first appeared for Q-nets
with circular quadrilaterals). Let Q = (q, q1 , q12 , q2 ) be a planar quadrilateral and
N a unit normal of the plane of Q. Denoting its diagonals with d1 = q12 − q and
d2 = q2 − q1 the area A(Q) can be computed as

1
A(Q) = det(d1 , d2 , N ).
2
If P is another quadrilateral with edges parallel to the edges of Q and with diagonals
c1 and c2 the area of P + t Q is equal to

1
A(P + t Q) = A(P) + 2t (det(d1 , c2 , N ) + det(c1 , d2 , N )) + t 2 A(Q). (3)
4
S-Conical CMC Surfaces. Towards a Unified Theory . . . 291

The space of all planar quadrilaterals with edges parallel to a given one Q is a
four dimensional vector space. Moding out the translations leaves a two dimensional
one. On this space the area is a quadratic form A(P) and the mixed area is the
corresponding symmetric form A(P1 , P2 ).
Since a conical mesh f and its Gauss map n have parallel edges, the area of the
quadrilaterals of the offset net f t = f + tn is quadratic in the distance t. A discrete
version of Steiner’s formula suggested in [5] (see also [6]) reads as follows.

Definition 3.1 Let f : Z2 → R3 be a conical net with the Gauss map n : G → R3 .

Then the discrete Steiner formula

A( f + tn) = A( f )(1 − 2H t + K t 2 ) (4)

defines a discrete mean curvature H and a discrete Gauss curvature K on the faces
of the net f :
A( f, n) A(n)
H =− , K = . (5)
A( f ) A( f )

Here A( f ) and A(n) are the areas of the quadrilaterals ( f, f 1 , f 12 , f 2 ) and (n, n 1 ,
n 12 , n 2 ) respectively, and A( f, n) is their mixed area.

4 Dual Quadrilaterals and Koenigs Nets

As we mentioned already, the mixed area A(P, Q) is a symmetric bilinear form on

a two-dimensional vector space of quadrilaterals with parallel edges. Quadrilaterals
P and Q with A(P, Q) = 0 are “orthogonal” with respect to this form. For any non-
vanishing planar quadrilateral Q there is a P with A(P, Q) = 0 and P is unique
up to scaling.

Definition 4.1 Two planar quadrilaterals P and Q with parallel edges are called
dual to each other if their mixed area vanishes:

A(P, Q) = 0.

Whenever scaling is unimportant we will simply talk about the dual quadrilateral
P ∗ , satisfying A(P, P ∗ ) = 0.
Let P = ( p, p1 , p12 , p2 ) and Q = (q, q1 , q12 , q2 ) be two planar quadrilaterals
with parallel edges and let N be their common normal then (3) implies the following
formula for their mixed area
1
A(P, Q) = (det( p12 − p, q2 − q1 , N ) + det(q12 − q, p1 − p2 , N )) . (6)
4
The duality can be described in terms of the diagonals [5, 6].
292 A.I. Bobenko and T. Hoffmann

Proposition 4.2 The mixed area A(P, Q) of two planar quadrilaterals with parallel
edges P = ( p, p1 , p12 , p2 ) and Q(q, q1 , q12 , q2 ) vanishes if and only if their non-
corresponding diagonals are parallel, i.e. p12 − p  q2 − q1 and q12 − q  p2 − p1 .

Let us present also useful explicit formulas for the dual quadrilateral. Let o be
the intersection point of the diagonals of a planar quadrilateral P = ( p, p1 , p12 , p2 ),
−p
e1 =  pp12
12 − p
, e2 =  pp22 − p1
− p1 
be the unit vectors of diagonals and define α, β, γ , δ
as the oriented lengths of the connection intervals of the intersection point o to
the vertices: p − o = αe1 , p12 − o = γ e1 , p1 − o = βe2 , p2 − o = δe2 . Then the
quadrilateral P ∗ = ( p ∗ , p1∗ , p12 ∗
, p2∗ ) determined by

1 1 1 1
p ∗ − o∗ = − e2 , p12
∗
− o∗ = − e2 , p1∗ − o∗ = e1 , p2∗ − o∗ = e1 (7)
α γ β δ

with some o∗ is dual to P.

Every planar quadrilateral has a dual one. However generic Q-nets are not dual-
izable as a whole. Dualizable Q-nets were introduced in [7] (see also [6]).

Definition 4.3 A Q-net f : Z2 → R3 is called a Koenigs net if there is a Q-net

f ∗ : Z2 → R3 such that the corresponding quadrilaterals of f and f ∗ are dual to
each other; f ∗ is called (Christoffel) dual of f .

Koenigs nets can be characterized [6, 7] in terms of the intersection points of the
diagonals.

Proposition 4.4 A Q-net f : Z2 → R3 in general position (each vertex and its

neighbours are not co-planar) is a Koenigs net if and only if the intersection points
of the diagonals of four quadrilaterals around each vertex are coplanar.

Since the planes of a planar quadrilateral and its dual are parallel, the conicality
conditions for f ∗ and f are satisfied simultaneously.

Proposition 4.5 If f : Z2 → R3 is conical and Koenigs, its Christoffel dual f ∗ :

Z2 → R3 is conical as well.

5 Conical Nets with Constant Mean Curvature

and S-Conical Nets

Definition 5.1 A conical net f : Z2 → R3 is called a net with constant mean cur-
vature (cmc net) if its mean curvature H defined by (5) is constant. Conical nets with
vanishing mean curvature are called minimal.

Theorem 5.2 A conical net f : Z2 → R3 is a cmc net with mean curvature H = 0

if and only if it is a Koenigs net with its Christoffel dual in constant distance H1 :
S-Conical CMC Surfaces. Towards a Unified Theory . . . 293

1
H = const ⇔ f ∗ = f − n. (8)
H
The dual net f ∗ is a conical cmc net with mean curvature H as well.

Indeed, we have

1
H A( f, f ) = −A( f, n) ⇔ 0 = A( f, f − n).
H
Thus f ∗ exists and is given by (8).
There is a tight connection (see [6, 17]) of conical nets and circular nets, i.e.
Q-nets with circular faces. Given a conical net one can choose a point on one of
the face’s planes arbitrarily. Mirroring this point at the planes spanned by the face’s
edges and their incident normals gives a circular net. Each vertex of this circular net
corresponds to a face of the conical net and vice versa. This way each conical net
gives rise to a 2-parameter family of circular nets. The circle centers of the circular
net lie on the cone axes of the conical one.
On the other hand, given a circular net, we can choose an initial unit vertex normal
arbitrarily. Mirroring that normal at the edge-bisecting planes gives rise to a consistent
set of vertex normals such that neighboring normals intersect. Their orthogonal planes
passing through the vertices of the circular net give rise to a conical net. This way
each circular Q-net gives rise to a 2-parameter family of conical nets. We will call
these conical and circular nets corresponding to each other.
We complete this section with an introduction of particular conical nets.

Definition 5.3 A discrete conical net f : Z2 → R3 is called s-conical if for every

vertex the intersection points of the diagonals of the quadrilaterals sharing the vertex
lie on a circle, and the axis of this circle is the vertex normal (the cone axis at the
vertex).

Fig. 1 To definition of
s-conical nets. The cone
touches all four neighboring
quadrilaterals. The
intersection points of the
diagonals are circular, and
the axis of the circle
coincides with the axis of the
cone
294 A.I. Bobenko and T. Hoffmann

It is easy to see that an s-conical net and its circular net formed by the intersection
points of the diagonals are corresponding (see Fig. 1). Indeed two triangles on two
neighboring faces sharing an edge built by this edge and the intersection points of the
diagonals are congruent. Therefore they are symmetric with respect to the reflection
in the symmetry plane of the planes of the neighboring faces.

Theorem 5.4 (i) S-conical nets are Koenigs nets.

(ii) The dual net of an s-conical net is also s-conical.

Proof (i) follows from Proposition 4.4 since the intersection points of the diagonals
are circular.
(ii) Consider two (congruent) triangles Δ1 , Δ2 on two neighboring faces of f with a
common edge built by this edge and the intersection points of the diagonals. As
it was explained above they are symmetric with respect to the reflection in the
symmetry plane P of the planes of the faces. The dualization formula (7) implies
that the corresponding triangles Δ∗1 , Δ∗2 of the dual net f ∗ are also symmetric
with respect to P. This implies that f ∗ is s-conical.

In Sect. 7 we give an alternative definition of this class and investigate it.

6 S-Isothermic Nets

For an s-conical net the intersection points o of the diagonals are circular with the
circle centers lying on the corresponding cone axes. Let f be a vertex of an s-conical
net and o1̄2̄ , o2̄ , o, o1̄ the diagonal intersection points of the quadrilaterals incident
to f . The points o1̄2̄ , o2̄ , o, o1̄ lie on a sphere s with center f . This furnishes a map
s : Z2 → {spheres in R3 } such that each sphere s is centered at the vertex f and
in each diagonals intersection point o four spheres s, s1 , s12 , s2 meet. Furthermore,
opposite spheres s and s12 as well as s1 and s2 touch at o.
In order to proceed further we recall some basic facts about s-isothermic nets and
their Möbius geometric description (see [6] for details).
Let R4,1 be the five dimensional Minkowski space with the standard Lorenz inner
product
4
x, y = −x0 y0 + xk yk .
k=1

There is a bijection between oriented spheres and planes s in R3 and unit vectors
ŝ ∈ R4,1 (here we consider planes as degenerate spheres), as well as points p in
R3 ∪ ∞ and isotropic vectors p̂ ∈ L4 ⊂ R4,1 . For points p one sets

1 +  p2 1 −  p2
p̂ = ( , p, ) (9)
2 2
S-Conical CMC Surfaces. Towards a Unified Theory . . . 295

with  p̂, p̂ = 0. The point at infinity ∞ is then given by (1, 0, 0, 0, −1). An oriented
sphere s with center c and radius r in R3 is represented as
 
1 1 + (c2 − r 2 ) 1 − (c2 − r 2 )
ŝ = , c, . (10)
r 2 2

A plane with the normal form v, n|R3 = d, n = 1 is is given by

ŝ = (d, n, −d).

In both cases ŝ, ŝ = 1. Changing the orientation of the sphere or plane corresponds
to the transformation ŝ → −ŝ. A point p lies on a sphere s if and only if

p ∈ s ⇔  p̂, ŝ = 0. (11)

The intersection angle α between two spheres s1 and s2 can be calculated as

cos α = ŝ1 , ŝ2 . (12)

In particular touching spheres satisfy

ŝ1 , ŝ2  = −1.

From here on we will use the same notation p, s for points and spheres in R3 and
their representatives (10) in R4,1 . The meaning will be clear from the context.

Definition 6.1 A net s : Z2 → R4,1 of space-like unit vectors solving the discrete
Moutard equation
s + s12 = λ(s1 + s2 ), λ = 0, (13)

is called s-isothermic.

Proposition 6.2 Let s : Z2 → R4,1 be an s-isothermic net. Then

s12 , s1  = s2 , s and s12 , s2  = s1 , s. (14)

Proof The unit length condition implies

1 = s12 , s12  = λ2 s1 + s2 , s1 + s2  − 2λs1 + s2 , s + 1,

and finally
λs1 + s2 , s1 + s2  − 2s1 + s2 , s = 0. (15)
296 A.I. Bobenko and T. Hoffmann

If s1 + s2 , s1 + s2  = 0 (this is the touching spheres case s1 , s2  = −1) we get from
(15) s1 , s = −s2 , s. This implies s12 , s1  = λ(s1 + s2 ) − s, s1  = −s, s1  =
s, s2 . In the non-touching case s1 + s2 , s1 + s2  = 0 substituting λ given by (15)
into (13) we obtain (14).

S-isothermic nets have the following geometric properties formulated in terms of

the centers c and the radii r of the corresponding spheres (10) (see [6, 7] for the
proof).
Theorem 6.3 (i) The centers c ∈ R3 of an s-isothermic net build Koenigs nets.
(ii) Let s be an s-isothermic net of spheres with the centers c : Z2 → R3 and (signed)
radii r : Z2 → R. The spheres s ∗ with the centers c∗ : Z2 → R3 and the radii
r ∗ : Z2 → R given by

c1 − c ∗ c2 − c ∗ 1
c1∗ − c∗ = , c2 − c∗ = − ,r = , (16)
r1 r r2 r r

build an s-isothermic net (called Christoffel dual net).

S-isothermic nets belong to integrable structures of discrete differential geometry
[6], i.e. they can be extended consistently to the Z N lattice so that all two-dimensional
coordinate subnets are s-isothermic. This property can be interpreted as a transfor-
mation of two-dimensional s-isothermic nets called Darboux transformation [12].

Definition 6.4 Consider two s-isothermic nets s : Z2 → R4,1 and ŝ : Z2 → R4,1 ,

and their spheres with the same indexes as corresponding. Thus, for example, the
spheres s, s1 , s12 , s2 , ŝ, ŝ1 , ŝ12 , ŝ2 build a combinatorial cube shown in Fig. 2, and we
treat the faces of s and ŝ as bottom and top faces of the cube. The s-isothermic nets s
and ŝ are said to be Darboux transforms of each other if the spheres corresponding to
the side faces of the cube also satisfy the discrete Moutard equations (with different
signs corresponding to two different pairs of opposite faces):

ŝ1 + s = a(ŝ + s1 ), ŝ12 + s2 = a2 (sˆ2 + s12 ),

ŝ2 − s = b(ŝ − s2 ), ŝ12 − s1 = b1 (sˆ1 − s12 ).

Fig. 2 A Darboux cube

S-Conical CMC Surfaces. Towards a Unified Theory . . . 297

Remark 6.5 Given s its Darboux transform ŝ is uniquely determined by one vertex
of ŝ which can be chosen arbitrary.

Remark 6.6 Note that the sides of a Darboux cube that correspond to the solutions
of the Moutard equation with plus sign have embedded quadrilaterals while the ones
with minus sign give rise to non-embedded quadrilaterals.

Remark 6.7 To pass to a multidimensional consistent picture mentioned above one

should change the orientations of the spheres for every second line si j → (−1)i si j .

7 S-Conical Nets as S-Isothermic Nets

The Moutard equation implies that the four spheres of each quadrilateral are linearly
dependent vectors in R4,1 . Thus an s-isothermic net s is in particular a Q-net. We
obtain three types of s-isothermic surfaces characterized by the fact that the four
spheres s, s1 , s2 , s12 :
share a common orthogonal circle (Type 1),
intersect in a pair of points (Type 2),
intersect in exactly one point (Type 3).
These three cases are distinguished by the signature of the Lorentz metric restricted
to the subspace of the corresponding spheres s, s1 , s12 , s2 which span a 3-dimensional
subspace U ⊂ R4,1 . Denote its orthogonal complement by U ⊥ . If the inner product
on U ⊥ is positive definite the unit vectors therein form a 1-parameter family of
spheres orthogonal to s, s1 , s12 , and s2 . This family shares a circle (given by the
isotropic vectors of U ) which thus is perpendicular to s, s1 , s12 , and s2 .
If the inner product on U ⊥ is indefinite and non degenerate then it contains two
isotropic directions p̂1 and p̂2 which give rise to two points p1 and p2 that are
contained in all four spheres.
Finally if the inner product is degenerate on U ⊥ the subspace touches the light-
cone of R4,1 and that direction gives rise to one common point of s, s1 , s12 , and
s2 .
The first type has a particularly nice special case when all the inner products are
1. In this case the four spheres for each quadrilateral touch cyclically [1, 6]. The
orthogonal circle then must pass through the four touching points.
S-isothermic surfaces of type 3 are of particular interest for us. It is the case where
the orthogonal circle (or the pair of common points) collapses into a point (Fig. 3).
298 A.I. Bobenko and T. Hoffmann

Fig. 3 An s-conical
quadrilateral as an
s-isothermic net of type 3

Theorem 7.1 For a map s : Z2 → R4,1 with s, s = 1 the following statements are
equivalent:
(i) s is s-isothermic of type 3.
(ii) s is a solution to the Moutard equation (13) and the intersection angles of the
spheres are complimentary: s, s1  = −s, s2  and |s, s1 | ≤ 1.
(iii) s + s12 and s1 + s2 are parallel isotropic vectors.
(iv) The centers c of s form an s-conical net and the intersection points o of the
diagonals lie on the corresponding spheres s, s1 , s12 , s2 .
Proof (i) ⇒ (ii). The centers of the spheres s, s1 , s12 , s2 are coplanar and the
opposite spheres touch in a common point. Thus this must be the intersection point
of the diagonals. The intersection angle of the diagonals α and its complimentary
angle π − α are exactly the intersection angles of the corresponding neighboring
spheres. Then the claim follows from (12).
(ii) ⇒ (iii). For s, s1  = −s, s2  Eq. (15) implies that s1 + s2 is isotropic (note
that λ = 0). Due to (13) s12 + s is parallel and therefore isotropic as well.
(iii) ⇒ (iv). The isotropy s1 + s2 , s1 + s2  = 0 implies the touching condition
s1 , s2  = −1. Moreover s1 + s2 and s + s12 are equivalent projective representa-
tions of the common touching point o. The centers c of the spheres form a planar
quadrilateral for which the diagonals (c1 , c2 ) and (c, c12 ) intersect at the touching
points o. Since o1̄2̄ , o2̄ , o, o1̄ are the points in which the spheres s1̄ , s2̄ , s1 , s2 touch
cyclically, o1̄2̄ , o2̄ , o, o1̄ lie on a circle,1 which we denote by C. This implies that
the net c is a Koenigs net with a circular net of points o. It remains to show that
it is conical as well.
Since the quadrilateral formed by say c, o, c2 , and o1̄ is a folded kite (two pairs
of non-opposing edges are equal in length), the planes spanned by c, o, c2 and
c, c2 , o1̄ are symmetric with respect to the plane spanned by c, c2 , and the axis
of the circle C. The same holds for the other edges and thus c is a conical net
corresponding to the circular net o.
(iv) ⇒ (i). Let s have a central net c that is s-conical. The intersection points
o, o1̄ , o1̄2̄ , o2̄ lie on the sphere s. Since the connection line through the centers

1 This is a simple fact from Möbius geometry that four spheres touching cyclically always have a
circle through their touching points.
S-Conical CMC Surfaces. Towards a Unified Theory . . . 299

c, c12 passes through the common point o of the spheres s, s12 , these must touch
at o. The isotropic vector of o is projectively equivalently represented by s12 + s
or s1 + s2 . This implies that these vectors are parallel, thus s solves the Moutard
equation. Since the four spheres share a single point the solution is of the third
type.
We have shown that s-conical nets and s-isothermic nets of type 3 can be canon-
ically identified: the centers of the spheres of an s-isothermic net of type 3 are the
vertices of the corresponding s-conical net, and their intersection (touching) points
o are the intersection points of the diagonals of the s-conical net.
Christoffel dualizations defined for both these classes also coincide. The following
theorem follows from the dualization formulas (7) and (16).
Proposition 7.2 The Christoffel dual net of an s-isothermic net of type 3 is an
s-isothermic net of type 3. Moreover the centers c and c∗ of the corresponding
spheres build Christoffel dual s-conical nets in the sense of Theorem 5.4.
Finally a Darboux transform of s-isothermic surfaces preserves nets of type 3 and
therefore is well defined for s-conical surfaces.
Theorem 7.3 Let c : Z2 → R3 be s-conical with the corresponding s-isothermic net
s : Z2 → R4,1 and let ŝ : Z2 → R4,1 be a Darboux transform of s with central net
ĉ : Z2 → R3 . Then ĉ is s-conical as well.
Moreover every Darboux cube possesses a (Ribaucour) sphere R ∈ R4,1 which
is orthogonal to all spheres s, s1 , . . . , ŝ12 of the Darboux cube,

R, s = R, s1  = . . . = R, ŝ12  = 0,

and passes through the intersection points o, ô of diagonals of c and ĉ.

Proof Let us use the characterization (ii) of s-isothermic nets of type 3 from Theo-
rem 7.1 that is s, s1  = −s, s2 . Due to the Moutard equation the scalar products
of the spheres at the opposite edges of any face of the Darboux cube are equal (see
Proposition 6.2), for example, s, s1  = s ∗ , s1∗ . This implies s ∗ , s1∗  = −s ∗ , s2∗ .
To prove the second claim, consider the one-parameter family of the spheres
touching the quadrilateral (c, c1 , c12 , c2 ) at the point o. They all are orthogonal to
s, s1 , s12 , s2 . Take the sphere R from this family that is orthogonal to ŝ. Due to
Moutard equations it is orthogonal to ŝ1 , ŝ2 , ŝ12 as well. Thus it touches the quadri-
lateral (ĉ, ĉ1 , ĉ2 , ĉ12 ) at ô.

8 S-Conical Nets with Constant Mean Curvature

In this section we consider s-conical nets with constant mean curvature H = 0.

S-conical minimal surfaces, H = 0, build an interesting subclass of s-conical sur-
faces with a rich theory (associated family, Weierstrass representation, variational
principle). This is a subject of a separate publication [2].
300 A.I. Bobenko and T. Hoffmann

Lemma 8.1 Let f, f ∗ : Z2 → R3 be two s-conical nets. Then any two of the fol-
lowing conditions imply the third:
(i) f ∗ is a Christoffel dual of f .
(ii) f ∗ is a Darboux transform of f .
(iii) f ∗ and f are in constant face offset, and it is equal to the distance o∗ − o
between the intersection points of diagonals of the corresponding faces of f
and f ∗ .
Proof 1. and 2. ⇒ 3.:
Let s and s ∗ be the s-isothermic nets corresponding to f and f ∗ respectively. Consider
the Darboux cube formed by the spheres s, s1 , s2 , s12 and their dual s ∗ , s1∗ , s2∗ , s12 ∗
.
In the following f, f 1 etc. will denote the sphere’s centers, which were previously
denoted by c, c1 etc. Since the quadrilaterals ( f, f 1 , f 12 , f 2 ) and ( f ∗ , f 1∗ , f 12
∗
, f 2∗ )
are dual they are parallel. Due to Theorem 7.3 they touch the Ribaucour sphere R at
o and o∗ respectively. This implies that the line trough o and o∗ intersects these two
quadrilaterals orthogonally (Fig. 4).
Next, consider two neighboring quadrilaterals ( f, f 1 , f 12 , f 2 ) and ( f 1 , f 11 , f 112 ,
f 12 ) with their duals. The common edge f 12 f 1 of the quadrilaterals is the axis of the
circle C = s1 ∩ s12 . Obviously o, o1 ∈ C. Moreover because of the orthogonality the
lines (oo∗ ) and (o1 o1∗ ) are tangents to C. The same argument implies that they are
also tangent to the circle C ∗ = s1∗ ∩ s12 ∗
.
∗
The circles C and C are co-planar and two co-planar circles have two sets of
common tangents distinguished by whether they intersect the edge between the circle
centers or not. Since the Darboux cubes are “flipped over” this happens for one of
the lattice directions and not for the other (see Remark 6.6). Therefore o∗ − o =
o1 − o1∗ . Since those lines are orthogonal to the quads, the face distance for the
nets is constant as well.
2. and 3. ⇒ 1.:
Again consider the Darboux cube. All sides are planar quadrilaterals and since f and
f ∗ are in constant face distance, they have parallel edges. Moreover all the diagonal
intersection points for a Darboux cube are collinear. Let us assume that the quadri-
lateral ( f, f 1 , f 1∗ , f ∗ ) is the one of the embedded sides of the Darboux cube (see
Remark 6.6). Its diagonal intersection point o and the points o, o∗ are collinear. The
triangles ( f, f 1 , o ) and ( f 1∗ , f ∗ , o ) are obviously similar. The triangles ( f, f 1 , o)
and ( f 1∗ , f ∗ , o∗ ) are their orthogonal projections and are similar as well. Thus f − o
and f 1∗ − o∗ are parallel, i.e. the diagonals for f and f ∗ satisfy the duality criteria
of Proposition 4.2.
3. and 1. ⇒ 2.:
Again, the planes of corresponding quadrilaterals for f and f ∗ are parallel since f ∗
is dual to f . Since ( f, f 2 , o) and ( f 2∗ , f ∗ , o∗ ) are similar triangles, the arguments of
the previous item in the proof imply that the intersection points of the diagonals of
all side-quadrilaterals lie on the line connecting o and o∗ . In particular this is the case
with the intersection point õ of the diagonals of the quadrilateral ( f, f 2 , f 2∗ , f ∗ ). We
will prove the claim in case that this quadrilateral is non-embedded, for embedded
quadrilaterals the proof differs just by a minor change of signs ±.
S-Conical CMC Surfaces. Towards a Unified Theory . . . 301

Fig. 4 A geometric cube in

Lemma 8.1. The
quadrilaterals ( f f 1 f 12 f 2 ),
( f ∗ f 1∗ f 12
∗ f ∗ ), ( f f f ∗ f ∗ )
2 1 1
are embedded, the
quadrilateral ( f f 2 f 2∗ f ∗ ) is
not. The diagonal
intersection points
o, o∗ , o , õ lie on a common
line orthogonal to the planes
of the quadrilaterals
( f f 1 f 12 f 2 ) and
( f ∗ f 1∗ f 12
∗ f ∗)
2

We can assume that õ lies at the origin. The triangles (õ, f ∗ , o∗ ) and (õ, f 2 , o)
are similar and o f 2  = r2 , o∗ f ∗  = r ∗ = r1 . This implies for f ∗ and similarly for
f 2∗ :
1 1
f∗ = f 2 , f 2∗ = f.
rr2 rr2

Then the middle three coordinates (10) of the spheres s, s2 , s2∗ and s ∗ are equal
f f∗ ∗

r
, rf22 , r 2∗ = rf , and rf ∗ = rf22 respectively. To prove the Moutard identity
2

r2
s − s2∗ = μ(s ∗ − s2 ), μ = −
r
1  f 2 −r 2
it remains to verify it for the terms of the form r
and r
in (10). The identity
302 A.I. Bobenko and T. Hoffmann

1 1 1 1
− ∗ = μ( ∗ − )
r r2 r r2

is obvious. For the terms of the form  f r −r the computation is slightly more
2 2

involved. The claim finally follows from

 f 2 − r 2 =  f 2 2 − r22 = õ − o2 .

Thus the cube is Darboux.

Remark 8.2 Considering degenerated examples it is not difficult to convince yourself
that the extra condition that the face distance equals the distance between the points
o and o∗ (or equivalently that the line (oo∗ ) is perpendicular to the quadrilaterals) in
Lemma 8.1 cannot be dropped.
Let us recall that a parallel net in constant distance in the sense of (1) to a conical
net lies in a constant distance face offset. We formulate the obtained properties of
s-conical cmc nets in the following
Theorem 8.3 (i) An s-conical net f is cmc if and only if it has a Christoffel dual
f ∗ in constant distance. The lines (oo∗ ) connecting the points of intersection of
diagonals of the corresponding quadrilaterals of f and f ∗ intersect the planes
of these quadrilaterals at constant angle θ . The distance oo∗  is constant for
the whole net.
(ii) An s-conical net f has a Darboux transform f D in constant distance if and
only if f is cmc with θ = π/2. Then the Darboux transform is Chistoffel dual
f D = f ∗.
Proof The first claim in (i) is Theorem 5.2. To prove the claim about the angles,
note that the the points o and o∗ corresponding to four quadrilaterals with a common
vertex f lie on two circles with the common axis ( f f ∗ ). Moreover the corresponding
intervals oo∗ are symmetric with respect to the reflections in the symmetry planes that
identify the planes of the faces. This implies that the distance oo∗  is constant for the
whole net, and the lines (oo∗ ) intersect the planes of the corresponding quadrilaterals
at constant angle.
A Darboux transform f D in constant distance (thus in constant distance face
offset) implies that the line through the touching points of the Ribaucour sphere
(oo D ) is orthogonal to the faces of the quadrilaterals, and the distance oo D  is
constant. Here we have denoted by o D the diagonal intersection points of f D . Then
the Darboux transform is Chistoffel dual due to Lemma 8.1, which implies that f is
cmc. The reverse statement also follows from this Lemma. Indeed, if f is cmc and
f ∗ its dual with orthogonal (oo∗ ) then f ∗ is a Darboux transform of f .
Remark 8.4 Geometric characterizations of s-conical cmc surfaces turn out to be
quite similar to the ones of circular cmc surfaces, that also can be characterized
as possessing a Christoffel dual, which is simultaneously a Darboux transform, at
constant distance [10].
S-Conical CMC Surfaces. Towards a Unified Theory . . . 303

9 Delaunay Nets

In this section we will construct s-conical Delaunay nets that are s-conical analogues
of rotationally symmetric cmc surfaces.
A planar polygon c : Z → R2 with non-vanishing edges, gives rise to a discrete
surface of revolution. Denoting the components of a vertex ck by ck = (xk , yk ) and
choosing an angle φ one can form

f (k,l) = (xk , cos(lφ)yk , sin(lφ)yk ).

Since all the polygons f (·, l0 ) are planar we will not distinguish between the gen-
erating polygon c and its resulting rotational symmetric net f in what follows. This
abuses the notation to some extend but it is understood that shifts in the first direction
are along the polygon and whenever a shift in the second direction occurs it refers to
the rotational direction of the net.
By symmetry c is s-isothermic. In order for it to be of type 1 with touching spheres
(s1 , s = s2 , s = 1) the polygon must satisfy in addition

φ
c1 − c = sin (y1 + y).
2
Instead of the polygon c one can look at the midpoint connectors in axial direction.
They form a planar polygon as well (see Fig. 5). Calling this polygon p = (x, v) one
finds that above condition translates to
φ
 p1 − p2 = 4 tan2 v1 v.
2
Thus in order to allow for the net to be s-isothermic with touching spheres one
needs
 p1 − p2
= const. (17)
vv1

Fig. 5 Rotationally
symmetric nets
304 A.I. Bobenko and T. Hoffmann

Fig. 6 Unrolling the billiard

in an ellipse: consecutively
place the edges of the billiard
on the axis and mark the
position of one of the foci.
The second polygon is
generated by the other focus
and mirroring on the axis

This holds true in particular for a polygon p that arises when unrolling the billiard
in an ellipse.
The construction is as follows [11] (see also [5]): Starting with a polygon given
by a billiard in an ellipse, unroll it by placing each edge on the axis consecutively
and marking the location of one of its foci in this process. This gives rise to a new
polygon p (see Fig. 6). The same can be done with the other focus and after mirroring
that second polygon along the axis the two polygons p and p ∗ are known to give rise
to a pair of discrete cmc nets in the discrete isothermic sense. In particular pairs of
edges p1 p and p1∗ p ∗ form trapezoids (see Fig. 7) with constant distances and diag-
onals  p ∗ − p = d =  p1∗ − p1  and  p ∗ − p1  = l =  p1∗ − p|. Corresponding
quadrilaterals from the two nets are dual to each other.
To see that this polygon satisfies Eq. (17) one needs to know that tan α tan β, with
the angles α, β in Fig. 7, is an integral of motion for the billiard in an ellipse, i.e.

α β α1 β1
tan tan = tan tan . (18)
2 2 2 2
We have
γ = π − α − β,
 p1 − p = 2 p − o sin γ2 ,
v =  p − o sin α,
v1 =  p − o sin β.

This implies

 p1 − p2 α β α β
= cot cot + tan tan − 2 = const.
v1 v 2 2 2 2

Summarizing, the billiard in an ellipse gives rise to a pair of polygons p and

p ∗ that generate a discrete isothermic cmc surface of revolution and its dual. The
same polygons p = (x, v), p ∗ = (x ∗ , v ∗ ) can be read as midpoint connectors in two
rotational symmetric s-isothermic nets that are generated by polygons c, c∗ where
S-Conical CMC Surfaces. Towards a Unified Theory . . . 305

Fig. 7 The polygons from

unrolling an ellipse

v v∗
c = (x, y) = (x, φ
), c∗ = (x ∗ , y ∗ ) = (x ∗ , ).
cos 2
cos φ2

Note that the corresponding scaling (x, y) → (x, y/ cos φ2 ) is an affine mapping
and thus preserves parallelity. Due to Proposition 4.2 the Christoffel duality can be
formulated in terms of parallel edges and diagonals, therefore dual quadrilaterals
stay dual after such scaling. So, the two resulting s-isothermic nets c and c∗ are dual
to each other as well.
Recall that the s-isothermic nets of type 1 have an orthogonal circle for each
quadrilateral of spheres and in the case of touching spheres the circle passes through
the touching points. For the s-isothermic nets c and c∗ one can think of the edges p1 p
as diameters in the inscribed circles and we will denote the touching points along
the edges c1 c with p̃ (see Fig. 8).
Since p, p1 , p1∗ , and p ∗ form a symmetric trapezoid, the circles for the quadri-
laterals (c, c1 , c12 , c2 ) and its dual (c∗ , c1∗ , c12
∗
, c2∗ ) are co-axial. Moreover, since the
∗ ∗
lines (c1 c) and (c1 c ) are parallel and tangent (in p̃ and p̃ ∗ ) to co-axial circles, their
distance is  p̃ − p̃ ∗  = l. Likewise (c2 c) and (c2∗ c∗ ) are parallel and touch the circles
in p and p ∗ thus their distance is  p ∗ − p = d. Together we see that the two dual
s-isothermic nets are in constant edge distance (but with two different distances for
the two lattice directions). This is a way to define s-isothermic cmc nets (see [12]).
Now define a third pair of polygons

y v y∗ v∗
q = (x, u) = (x, φ
) = (x, ), q ∗ = (x, u ∗ ) = (x, ) = (x, ).
cos 2 cos2 φ2 cos φ
2 cos2 φ
2
306 A.I. Bobenko and T. Hoffmann

Fig. 8 The net c and its dual c∗ with the inscribed circles

Fig. 9 An s-conical
Delaunay net

Rotating these polygons one gets nets for which one can think of c and c∗ as the
polygons of midpoint connectors. The resulting faces for q and q ∗ thus will have
constant face offset. Moreover they will be dual by the same argument as above.
The sphere s centered at c contains p̃. Since (cq) and (c p̃) are orthogonal a sphere
t centered at q that meets p̃ also contains p̃1̄ . By symmetry the quadrilaterals formed
by the p̃ are circular. Finally one finds that the points p̃ are the diagonal intersection
points of the quadrilaterals of the net q. One can see this for example by noting that
the radii for the s-isothermic spheres are r = v tan φ2 , so  p̃ − c = r and half an
edge in rotational direction is g = c − q = y tan φ2 . Thus gr = vy which is constant
by construction.
S-Conical CMC Surfaces. Towards a Unified Theory . . . 307

Fig. 10 Nested circular,

s-isothermic, and s-conical
Delaunay nets

We see that the third net formed by q is in fact an s-conical net with the dual
net q ∗ in constant face offset equal the distance between the corresponding diagonal
intersection points. Due to Theorem 8.3 this an s-conical cmc net. Figure 9 shows
an example of such an s-conical Delaunay net. It is worth mentioning that the three
types of discrete Delaunay nets (discrete isothermic, s-isothermic in the touching
case, and s-conical) can be arranged in such a way that the second touches the first
and the third touches the second as shown in Fig. 10. This is a direct consequence of
the construction.

Acknowledgments This research was supported by the DFG Collaborative Research Center, TRR
109 “Discretization in Geometry and Dynamics”.

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Bobenko, A., Hoffmann, T., Springborn, B.: Minimal surfaces from circle patterns: geometry
from combinatorics. Ann. Math. 164(1), 231–264 (2006)
2. Bobenko, A.I., Hoffmann, T., König, B., Sechelmann, S.: S–conical minimal surfaces. Towards
a unified theory of discrete minimal surfaces. (2015). Preprint
3. Bobenko, A.I., Matthes, D., Suris, Y.B.: Discrete and smooth orthogonal systems: C ∞ -
approximation. Int. Math. Res. Not. 2003(45), 2415–2459 (2003)
4. Bobenko, A.I., Pinkall, U.: Discretization of surfaces and integrable systems. In: Bobenko,
A.I., Seiler, R. (eds.) Discrete Integrable Geometry and Physics, pp. 3–58. Oxford University
Press, New York (1999)
308 A.I. Bobenko and T. Hoffmann

5. Bobenko, A.I., Pottmann, H., Wallner, J.: A curvature theory for discrete surfaces based on
mesh parallelity. Math. Ann. 348, 1–24 (2010)
6. Bobenko, A.I., Suris, Y.B.: Discrete differential geometry. Integrable structure. Graduate Stud-
ies in Mathematics, vol. 98. American Mathematical Society, Providence (2008)
7. Bobenko, A.I., Suris, Y.B.: Discrete Koenigs nets and discrete isothermic surfaces. Int. Math.
Res. Not. 2009(11), 1976–2012 (2009)
8. Burstall, F., Hertrich-Jeromin, U., Rossman, W., Santos, S.: Discrete special isothermic sur-
faces. Geom. Dedicata 174, 1–11 (2015)
9. Grosse-Brauckmann, K., Polthier, K.: Numerical examples of compact surfaces of constant
mean curvature. In: Elliptic and Parabolic Methods in Geometry, pp. 23–46. A K Peters,
Wellesley (1996)
10. Hertrich-Jeromin, U., Hoffmann, T., Pinkall, U.: A discrete version of the Darboux transfor-
mation for isothermic surfaces. In: Bobenko, A., Seiler, R. (eds.) Discrete Integrable Geometry
and Physics, pp. 59–81. Oxford University Press (1999)
11. Hoffmann, T.: Discrete rotational cmc surfaces and the elliptic billiard. In: Hege, H.C., Polthier,
K. (eds.) Mathematical Visualisation, pp. 117–124. Springer (1998)
12. Hoffmann, T.: A Darboux transformation for discrete s-isothermic surfaces. J. Math-For-
Industry (2010)
13. Liu, Y., Pottmann, H., Wallner, J., Wang, W., Yang, Y.: Geometric modeling with conical
meshes and developable surfaces. ACM Trans. Graphics 25(3), 681–689 (2006)
14. Müller, C.: On discrete constant mean curvature surfaces. Discrete Comput. Geom. 51(3),
516–538 (2014)
15. Müller, C.: Semi-discrete constant mean curvature surfaces. Math. Z. 279(1–2), 459–478 (2015)
16. Pan, H., Choi, Y.K., Liu, Y., Hu, W., Du, Q., Polthier, K., Zhang, C., Wang, W.: Robust modeling
of constant mean curvature surfaces. ACM Trans. Graphics 31(4), Article No. 85 (2012)
17. Pottmann, H., Wallner, J.: The focal geometry of circular and conical meshes. Adv. Comput.
Math. 29(3), 249–268 (2008)
18. Schief, W.K.: On the unification of classical and novel integrable surfaces. II. Difference
geometry. R. Soc. Lond. Proc. Ser. A Math. Phys. Eng. Sci. 459, 373–391 (2003)
19. Schief, W.K.: On a maximum principle for minimal surfaces and their integrable discrete
counterparts. J. Geom. Phys. 56(9), 1484–1495 (2006)
20. Wang, W., Wallner, J., Y., L.: An angle criterion for conical mesh vertices. J. Geom. Graphics
11(2), 199–208 (2007)
Constructing Solutions to the Björling
Problem for Isothermic Surfaces
by Structure Preserving Discretization

Ulrike Bücking and Daniel Matthes

Abstract In this article, we study an analog of the Björling problem for isothermic
surfaces (that are a generalization of minimal surfaces): given a regular curve γ in
R3 and a unit normal vector field n along γ, find an isothermic surface that contains
γ, is normal to n there, and is such that the tangent vector γ  bisects the principal
directions of curvature. First, we prove that this problem is uniquely solvable locally
around each point of γ, provided that γ and n are real analytic. The main result is that
the solution can be obtained by constructing a family of discrete isothermic surfaces
(in the sense of Bobenko and Pinkall) from data that is read off from γ, and then
passing to the limit of vanishing mesh size. The proof relies on a rephrasing of the
Gauss-Codazzi-system as analytic Cauchy problem and an in-depth-analysis of its
discretization which is induced from the geometry of discrete isothermic surfaces.
The discrete-to-continuous limit is carried out for the Christoffel and the Darboux
transformations as well.

1 Introduction

Isothermic surfaces are among the most classical objects in differential geometry:
these are surfaces that admit a conformal parametrization along curvature lines,
see Definition 1. Like various particular geometries—special coordinate systems,
minimal surfaces, surfaces of constant curvature—they have been introduced and
intensively studied in the second half of the 19th century [9, 24]. Also, like the
many of these classical objects, they have been “rediscovered” in the 1990s, both
in connection with integrable systems and in the context of discrete differential

U. Bücking
Inst. für Mathematik, Technische Universität Berlin,
Straße des 17. Juni 136, 10623 Berlin, Germany
e-mail: [email protected]
D. Matthes (B)
Zentrum Mathematik – M8, Technische Universität München,
Boltzmannstr. 3, D-85747 Garching bei München, Germany
e-mail: [email protected]

© The Author(s) 2016 309

A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_10
310 U. Bücking and D. Matthes

Fig. 1 Initial zig-zag on a discrete isothermic surface

geometry. The first description of isothermic surfaces as soliton surfaces is found

in [11]. The first definition for discrete isothermic surfaces was made shortly after
in [3]. In the most simple case, these are immersions of Z2 into R3 such that the
vertices of each elementary quadrilateral are conformally equivalent to the corners
of a planar square, see Definition 3.
This discrete surface class, its transformations and invariances has been studied
e.g. in [7, 8, 20]. A systematic presentation of the theory of isothermic surfaces in
the context of Möbius geometry can be found in [15]. Finally, we refer to [4] for
a detailed overview on discrete isothermic surfaces as part of discrete differential
geometry, including historical remarks.
Despite the manifold results on (classical) isothermic surfaces and the related
equations, the fundamental question about their construction from suitably chosen
data has apparently been left open. On the one hand, the machinery of integrable
systems enables one to construct a rich variety of “solitonic” isothermic surfaces [11].
But on the other hand, nothing seems to be known about the well-posedness of an
initial or boundary value problem for the Gauss-Codazzi-equations in general. The
latter form a PDE system (cf. Eq. (5)) which contains both elliptic and hyperbolic
equations.1 The appearance of an elliptic equation suggests that data for the surface
boundary should be prescribed, as it is done for minimal surfaces for example. The
hyperbolic equations, on the other hand, suggest to provide data for two curvature
lines instead, like in the case of level surfaces in triply orthogonal systems [1]. Neither
of the two approaches seems promising for the coupled system.
In contrast, there is a canonical way to pose an initial value problem for a discrete
isothermic surface. One prescribes the vertices in R3 for a “zig-zag”-curve in para-
meter space as indicated in Fig. 1. For vertices in general position, these data can be
extended to a discrete isothermic surface in a unique way. In fact, all vertices on the
discrete surface are easily obtained inductively from the prescribed data.
In this paper, we formulate and prove solvability of a Björling problem for real
analytic isothermic surfaces. And we prove that the solution can be obtained as the
continuous limit of discrete isothermic surfaces.

1 The system of Gauss-Codazzi-equations can be simplified to Calapso’s equation [8], which is a

single scalar fourth order PDE, but unfortunately of indefinite type.
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 311

The classical Björling problem is to find a minimal surface that touches a given
curve in R3 along prescribed tangent planes. This problem has been solved in general,
see [13]. An extension of the Björling problem to surfaces of constant mean curvature
has been posed (and solved) in [5]. A natural formulation of the Björling problem in
the yet more general class of isothermic surfaces reads as follows.
Problem 1 Given a regular curve γ in R3 , and two mutually orthogonal unit vector
fields v, w along γ, neither of which is tangent to γ at any point. Find an isothermic
surface S containing γ such that v and w are the principal directions of curvature at
each point of γ.
We believe that this problem is solvable locally provided that γ, v and w are all real
analytic. Indeed, the non-tangency of the vector fields allows to give a reformulation
in terms of a non-characteristic Cauchy problem. However, we do not address the
Björling problem in this full generality here, but stick to the following restricted
setting, where we do not prescribe two tangent vector fields v and w individually, but
only a two-dimensional tangent plane:
Problem 2 Given a regular curve γ in R3 , and a unit vector field n that is orthogonal
to the tangent vectors γ  at each point. Find an isothermic surface S containing γ
such that at each point of γ, the vector n is normal to S, and each of the two directions
of principal curvature encloses an angle π/4 with γ  .
As a corollary of the results presented here, it follows that this problem is uniquely
solvable for real analytic γ and n, at least locally around each point of γ. Existence
and uniqueness of a real analytic isothermic surface S for given data is the minor
result of this paper, see Theorem 1. The main result is that the real analytic data
can be “sampled” with a mesh width  > 0 in a suitable way such that the discrete
isothermic surfaces S  constructed from the discrete data converge in C 1 to S. The
precise formulation is given in Theorem 2.
It is remarkable that naive numerical experiments suggest that such an approxi-
mation result might not be true. It was already noted in [3] that discrete isothermic
surfaces depend very sensitively on their initial data. The limit  → 0 is delicate,
and inappropriate choices of the initial zig-zag cause the sequence S  to diverge
rapidly. In fact, even the possibility to construct any sequence of discrete isothermic
surfaces that approximates a given smooth one is not obvious. Discrete isothermic
surfaces are one of many examples of a discretized geometric structure for which the
passage back to the original continuous structure needs a highly non-trivial approx-
imation result, the proof of which is analysis-based and goes far beyond elementary
geometric considerations. Further such non-trivial convergence results are available,
for instance, for discrete surfaces of constant negative Gaussian curvature [2], for
discrete triply orthogonal systems [1], and, most importantly, for circle patterns
[6, 21, 22] as approximations to conformal maps.
The core of our convergence proof is a stability analysis of the discrete Gauss-
Codazzi system that we derive for discrete isothermic surfaces. We show that the
solution to the discrete Gauss-Codazzi equations with sampled data as initial condi-
tion remains close to the solution of the classical Gauss-Codazzi system for the same
312 U. Bücking and D. Matthes

(continuous) initial data. In a second step, this implies proximity of the respective
discrete and continuous surfaces. We are able to quantify the approximation error in
terms of the supremum-distance between analytic functions on complex domains: it
is linear in the mesh size. In fact, we conjecture that this result is sub-optimal, and
second-order approximation should be provable, using a more refined analysis and
a more careful approximation of the data.
The techniques used in the proof are similar to those employed by one of the
authors [17] to prove convergence of circle patterns to conformal maps. The geo-
metric situation for isothermic surfaces, however, is much more complicated, and
the structure of the Gauss-Codazzi system is much more complex than the Cauchy-
Riemann equations. The proof of stability relies on estimates for the solution of
analytic Cauchy problems in scales of Banach spaces. These estimates have been
developed—in the classical, non-discretized setting—in Nagumo’s famous article
[18] as part of the existence proof for analytic Cauchy problems. Here, we shall
rather use Nirenberg’s [19] version of these estimates. For an overview over the his-
tory of analytic Cauchy problems and the related estimates, see the beautiful article
of Walter [23].
Note that the convergence proof here is more direct than the one in [17]. While
the latter was based on purely discrete considerations, the current proof uses semi-
discrete techniques: a–somewhat artificial–extension of the discrete functions to con-
tinuous domains allows to formulate estimates more easily. The main simplification,
however, is that we separate the proofs for existence of a classical solution and its
approximation by discrete solutions.

The paper is organized as follows. In Sect. 2 we formulate the Gauss-Codazzi-

system for smooth isothermic surfaces in the framework of analytic Cauchy prob-
lems and prove unique local solvability of the Björling problem by the Cauchy-
Kowalevskaya theorem. In Sect. 3 we derive an analogous system of difference
equations for discrete isothermic surfaces. For appropriate initial conditions, the
convergence of the discrete solutions to the corresponding smooth ones is proven in
Sect. 4. Then, in Sect. 5 we explain how to discretize the Björling initial data appropri-
ately, and prove convergence of the discrete surfaces to the respective smooth one.
Finally, in Sect. 6, the convergence result is extended to Christoffel and Darboux
transformations.

2 Smooth Isothermic Surfaces

We start by summarizing basic properties of smooth isothermic surfaces and proving

our first result on the local solvability of the Björling problem.
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 313

Fig. 2 Relation between the

coordinates (x, y) and (ξ, η)

2.1 Coordinates and Domains

For concise statements and proofs, we need to work with two different coordinate
systems (ξ, η) and (x, y) on R2 simultaneously. These coordinates are related to each
other by
x−y x+y
ξ= , η= ⇔ x = η + ξ, y = η − ξ, (1)
2 2
see Fig. 2. Accordingly, the partial derivatives transform as follows:

∂ξ = ∂ x − ∂ y , ∂ η = ∂ x + ∂ y .

Observe in particular that

∂ξ2 + ∂η2 = 2(∂x2 + ∂ y2 ). (2)

It will be convenient to consider (ξ, η) as the “basic” coordinates and (x, y) as the
auxiliary ones. More precisely: in the rare cases that we need to specify explicitly the
arguments of a function g : Ω → R defined on a domain Ω ⊂ R2 , then we shall write
g(ξ, η) for the value of g at the point with coordinates x = η + ξ and y = η − ξ.
For further reference, define for r ≥ h > 0 the domains
 
Ω(r |h) = (ξ, η) ∈ R2 ; |ξ| + |η| ≤ r, −h < η ≤ h .

In the (x, y)-coordinates, Ω(r |h) is a axes-parallel square of side length 2r , centered
at the origin, that is cut off at the top-right and bottom-left corners.

2.2 Definition and Equations

By abuse of notation, we use the term “(parametrized) surface” for a smooth and
non-degenerate map F : Ω(r |h) → R3 . Here non-degeneracy means that the vector
314 U. Bücking and D. Matthes

fields Fx and Fy are linearly independent. Every such surface comes with a smooth
normal map N : Ω(r |h) → S2 , given by

Fx × Fy
N= .
Fx × Fy

Definition 1 F : Ω(r |h) → R3 is a (parametrized) isothermic surface, if

(1) F is conformal, i.e., there exists a conformal factor u : Ω(r |h) → R such that

Fx 2 = Fy 2 = e2u ,
Fx , Fy  = 0, (3)

(2) F parametrizes along curvature lines, i.e., the normal map N : Ω(r |h) → S2
satisfies

Fx y , N  = 0. (4)

The quantities k, l : Ω(r |h) → R in

−
N x , Fx  = eu k, −
N y , Fy  = eu l,

are the (scaled) principal curvatures.

Remark 1 The genuine principal curvature functions are given by e−u k and e−u l.
The quantities k and l are better suited for the calculations below.
The next result is classical.
Lemma 1 Assume that an isothermic surface F : Ω(r |h) → R3 is given. Then the
conformal factor u and the scaled curvatures k, l satisfy the Gauss-Codazzi equations

− (u x x + u yy ) = kl, lx = ku x , k y = lu y . (5)

Conversely, if functions u, k, l : Ω(r |h) → R satisfy the system (5), then there exists
an isothermic surface F : Ω(r |h) → R that has u as its conformal factor and has
scaled curvatures k, l. Moreover, F is uniquely determined up to Euclidean motions.
We briefly recall the proof, since we shall need some of the calculations later.
Proof (Sketch) For a given isothermic surface F : Ω(r |h) → R3 , introduce the
adapted frame
 
Ψ := e−u Fx , e−u Fy , N : Ω(r |h) → SO(3)

and define the transition matrices U, V : Ω(r |h) → so(3) implicitly by

Ψx = Ψ U, Ψ y = Ψ V. (6)
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 315

Using the defining properties of the isothermic parametrization, one easily obtains
the following explicit expressions for U and V :
⎛ ⎞ ⎛ ⎞
0 u y −k 0 −u x 0
U = ⎝−u y 0 0 ⎠ , V = ⎝u x 0 −l⎠ (7)
k 0 0 0 l 0

The compatibility condition U y − Vx = U V − V U implies the equations in (5).

Conversely, if u, k, l satisfy (5), then the matrix functions U, V : Ω(r |h) → so(3)
defined by (7) satisfy compatibility condition U y − Vx = U V − V U . Consequently,
one can define a further matrix function Ψ = (Ψ1 , Ψ2 , Ψ3 ) : Ω(r |h) → SO(3) as
solution to the system (6). Clearly, the solution Ψ is uniquely determined by its
value Ψ (0) ∈ SO(3) at (x, y) = 0. The particular form of U and V imply that

∂x (eu Ψ2 ) = ∂ y (eu Ψ1 ),

which further implies the existence of a map F : Ω(r ) → R3 such that

∂x F = eu Ψ1 and ∂ y F = eu Ψ2 . (8)

The map F is non-degenerate, and it is uniquely determined by its value F(0) ∈ R3

at (x, y) = 0. Clearly Ψ is an adapted frame for the surface defined by F, whose
normal vector field is given by Ψ3 . It follows directly from (8) that F is conformal (3).
The property (4) is a further direct consequence of (6) and the special form of U and
V from (7). 
A different form of the Gauss-Codazzi equations (5) is needed in the following.
Remind the relation between the coordinates (x, y) and (ξ, η) by x = η + ξ and
y = η − ξ in Sect. 2.1. Introduce auxiliary functions v, w : Ω(r |h) → R by

1 1
v= uξ , w = uη .
2 2
Further recall (2). Then the Gauss-Codazzi system (5) attains the form

vη = wξ , (9)
wη = −vξ − kl, (10)
k y = l(w − v), (11)
lx = k(w + v). (12)
316 U. Bücking and D. Matthes

2.3 Local Solution of the Björling Problem

The following result implies local solvability of (the restricted version of) the Björling
problem for isothermic surfaces, with real analytic data. To see the equivalence to
Problem 2 stated in the introduction, observe that the conformal parametrization
F of an isothermic surface and our coordinates in (1) are such that the images of
{x = const} and of {y = const} are mapped to curvature lines under F, whereas the
tangent to each curve ξ → F(ξ, η) is always at an angle of π/4 to both curvature
directions.
Theorem 1 Let an analytic and regular curve f : (−r, r ) → R3 and an analytic
normal unit vector field n : (−r, r ) → S2 be given, that is
f  , n ≡ 0. Then, for some
h > 0 with h ≤ r , there exists a unique analytic isothermic surface F : Ω(r |h) → R3
such that F and its normal vector field N satisfy

F(ξ, 0) = f (ξ), N (ξ, 0) = n(ξ) for all ξ ∈ (−r, r ). (13)

Remark 2 The original Björling problem consists in finding a minimal surface in R3

that touches a given curve along prescribed tangent planes. See [5] for an extension
to constant mean curvature surfaces. Our problem is a bit different since (13) implies
in addition that the tangential vector to the data curve is everywhere at angle π/4
with the directions of principal curvature, see (14). Such additional restrictions are
expected to guarantee unique solvability of the Björling problem in the much larger
class of isothermic surfaces.
Proof (of Theorem 1) If there exists an isothermic surface F : Ω(r |h) → R3 with
the properties (13), then

f  (ξ) = Fξ (ξ, 0) = Fx (ξ, 0) − Fy (ξ, 0) (14)

at every ξ ∈ (−r, r ), and in particular

f  (ξ) 2 = Fx (ξ, 0) 2 + Fy (ξ, 0) 2 −

Fx (ξ, 0), Fy (ξ, 0) = 2e2u(ξ,0) .

It follows that f and n determine both the conformal factor u and the adapted frame
Ψ = (e−u Fx , e−u Fy , N ) : Ω(r |h) → SO(3) uniquely on η = 0; denote the corre-
sponding functions by u 0 : (−r, r ) → R and Ψ 0 : (−r, r ) → SO(3), respectively.
Next, introduce functions v0 , w0 , k0 , l0 : (−r, r ) → R by
⎛ ⎞
0 2w0 −k0
(u 0 ) = 2v0 u 0 , (Ψ 0 ) = Ψ 0 ⎝−2w0 0 l0 ⎠. (15)
k0 −l0 0

The line {η = 0} = {x + y = 0} is obviously non-characteristic for the system of

equations (9)–(12). Hence, the Cauchy-Kowalevskaya theorem applies in this sit-
uation. For some sufficiently small h > 0, there exists a unique analytic solution
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 317

v, w, k, l : Ω(r |h) → R to (9)–(12) with the initial conditions v0 , w0 , k0 , l0 at η = 0.

Since (9) is a compatibility condition for the linear system

u ξ = 2v, u η = 2w,

there exists a unique analytic solution u : Ω(r |h) → R with u = u 0 for η = 0. The
triple (u, k, l) satisfies (5). Lemma 1 guarantees the existence of a unique isothermic
surface F : Ω(r |h) → R3 with u as conformal factor, with scaled principle curva-
tures k and l, and with the normalizations

F(0) = f (0), N (0) = n(0), Fx (0) − Fy (0) = f  (0). (16)

Analyticity of F is clear from its construction in the proof. To see that F attains
the initial data (13), first observe that an adapted frame Ψ necessarily satisfies Ψξ =
Ψx − Ψ y = Ψ (U − V ), and so Ψ = Ψ 0 on η = 0, thanks to (7) and (15), (16). In
particular, we have that Ψ3 (ξ, 0) = N (ξ). And further, Fξ = Ψ1 − Ψ2 = Ψ10 − Ψ20
implies F = f on η = 0.
Concerning uniqueness: f and Ψ 0 determine the initial data (v0 , w0 , k0 , l0 ) for (9)–
(12)—and hence also its solution (v, w, k, l)—uniquely. Invoking again Lemma 1, it
follows that F with the normalization (16) is unique as well. 

3 Discrete Isothermic Surfaces

Throughout this section, we assume that some (small) parameter  > 0 is given, which
quantifies the average mesh width of the considered discrete isothermic surfaces. We
introduce the abbreviation

z ∗ = 1 − 2 z 2 (17)

for arbitrary quantities z, assuming that |z| < 1.

3.1 Coordinates and Domains

Recall that we are working with the two coordinate systems from (1) simultaneously,
(ξ, η) being the “basic” coordinates and (x, y) being the “auxiliary” ones. Introduce
the associated shift-operators T x , T y , T ξ , T η by

  
T x (ξ, η) = (ξ + , η + ), T ξ (ξ, η) = (ξ + , η)
4 4 2
  
T y (ξ, η) = (ξ − , η + ), T η (ξ, η) = (ξ, η + ).
4 4 2
318 U. Bücking and D. Matthes

By slight abuse of notation, we shall use the same symbols for the associated contra-
variant shifts of functions f : Ω(r |h) → R, i.e., T x f := f ◦ T x etc. The associated
central difference quotient operators are defined by

1 1
δx f = (T x f − T −1
x f ), δξ f = (T ξ f − T −1
ξ f)
 
1 1
δ y f = (T y f − T −1
y f ), δη f = (T η f − T −1
η f ).
 
It is a notorious inconvenience in discrete differential geometry that the various quan-
tities which are derived from discrete geometric objects are associated to different
natural domains of definition. To account for that, we need to single out specific
subdomains inside our basic domain Ω(r |h): let

Ω [x] (r |h) = Ω(r |h) ∩ T x Ω(r |h) ∩ T −1

x Ω(r |h),

be the natural domain of definition for δx f , when f is defined on Ω(r |h). Likewise,
we define Ω [y] (r |h). The domain
 
Ω [x y] (r |h) = Ω(r − |h − )
2 2
is such that the mixed difference quotient δx δ y f is well-defined there; notice that δξ f
and δη f are well-defined on Ω [x y] (r |h). In the same spirit, we introduce Ω [x x y] (r |h)
as domain for δx2 δ y f etc. For each point ζ ∈ Ω [x y] (r |h), we say that the four points
−1 −1
T ξ ζ, T η ζ, T ξ ζ and T η ζ form an elementary -square.

3.2 Definition of Discrete Isothermic Surfaces

In this section, we give a variant of the definition for discrete isothermic surfaces
from [3], which is well-suited for the passage to the continuum limit. First, we need
auxiliary notation.
Definition 2 Four points p1 , . . . , p4 ∈ R3 form a (non-degenerate) conformal
square iff they lie on a circle, but no three of them are on a line, they are cycli-
cally ordered,2 and their mutual distances are related by

p1 − p2 · p3 − p4 = p1 − p4 · p2 − p3 . (18)

Remark 3 The name refers to the fact that p1 , . . . , p4 form a (non-degenerate) con-
formal square if and only if there is a Möbius transformation of R3 which takes these
points to the corners of the unit square, (0, 0, 0), (1, 0, 0), (1, 1, 0) and (0, 1, 0),

2 Cyclicordering means that walking around the circle either clockwise or anti-clockwise, one
passes p1 , p2 , p3 and p4 in that order, see Fig. 3 (left).
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 319

Fig. 3 Conformal squares

and the association of
quantities to lattice points

respectively. Notice that the non-degeneracy condition is important for the equiva-
lence, since certain point configurations on a straight line can be Möbius transformed
into the unit square as well.
Alternatively, one could define conformal squares by saying that p1 to p4 have
cross-ratio equal to minus one, either in the sense of quaternions, see for example [15],
or after identification of these points with complex numbers in their common plane.
Again, non-degeneracy is important for equivalence of the definitions.

The following is an easy exercise in elementary geometry.

Lemma 2 For any given three points p1 , p2 , p3 ∈ R3 (with ordering) that are not
collinear (and in particular pairwise distinct), there exists precisely one fourth
point p4 ∈ R3 that completes the conformal square. Moreover, the coordinates of
p4 depend analytically on those of p1 , p2 and p3 .
We are now going to state the main definition, namely the one for discrete isother-
mic surfaces. Originally [3], discrete surfaces have been introduced as particular
immersed lattices in R3 . Having the continuous limit in mind, we give a slightly
different definition, which describes a continuous immersion in R3 , corresponding
to a two-parameter family of lattices.
Definition 3 A map F  : Ω(r |h) → R3 is called (the parametrization of) a -
discrete isothermic surface, if elementary -squares are mapped to conformal squares
in R3 .

Remark 4 Since no continuity is required for F  : Ω(r |h) → R3 , one can think of
it—at this point—for instance as the piecewise constant extension of a map F̃  :
Λ (r |h) → R3 that is only defined on a suitable lattice Λ (r |h) ⊂ Ω(r |h), e.g. on
 
ξ η
Λ (r |h) = (ξ, η) ∈ Ω(r |h) ; + ∈ Z .
 

Alternatively, one can say that F  : Ω(r |h) → R3 is a discrete isothermic surface,
if and only if the four vectors
  −1  −1 
T y δx F , T x δy F , T y δx F , T x δ y F
320 U. Bücking and D. Matthes

always lie in one common plane and satisfy

T y δx F  T −1 δx F  = T x δ y F  T −1 δ y F  . (19)
y x

Note that this identity is a discrete replacement for the relation Fx 2 = Fy 2 on

smooth conformally parametrized surfaces.

3.3 The Discrete Björling Problem

We introduce the analog of the Björling problem for -discrete isothermic surfaces.
In contrast to its continuous counterpart, its solution is immediate. First, we need
some more notation to formulate conditions on the data.
Definition 4 A function f  : Ω(r |h) → R3 is said to be non-degenerate if neither
any of the point triples
 −1

Tξ f  , T −1  
η f , T ξ f (ξ, η),

nor any of the point triples

 −1

Tξ f  , T η f  , T ξ f  (ξ  , η  )

are collinear, where (ξ, η), (ξ  , η  ) ∈ Ω(r |h) are arbitrary points such that these
respective values of f  are defined. If collinearities occur, then f  is called degen-
erate.

Definition 5 We call a function f  : Ω(r | 2 ) → R3 Björling data for the construc-

tion of an -discrete isothermic surface if it is non-degenerate.

Proposition 1 Let h̄ and  > 0 with r > h̄ > 2 and Björling data f  be given. Then,
there exists some maximal h ∈ ( 2 , h̄] and a unique -discrete isothermic surface F  :
Ω(r |h) → R3 such that F  = f  on Ω(r | 2 ). Here maximal has to be understood
as follows: either h = h̄, or the restriction of F  to Ω(r |h − 2 ) is degenerate.

Proof The proof is a direct application of Lemma 2: from the data f  given on
Ω(r | 2 ), one directly calculates the values of F  on Ω(r |). These are then extended
to Ω(r |3 2 ) in the next step, and so on. The procedure works as long as no degeneracies
occur. 
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 321

3.4 Discrete Quantities and Basic Relations

Let some discrete isothermic surface F  : Ω(r |h) → R3 be given. Below, we intro-
duce quantities that play an analogous role for F  as u, k, l etc. do for F. Figure 3
(right) indicates, on which lattices these respective quantities live.
Define the discrete conformal factors û : Ω [x] (r |h) → R and ǔ : Ω [y] (r |h) →
R, respectively, by

eû = δx F  , eǔ = δ y F  .

Thanks to the property (19) of discrete isothermic surfaces, these seemingly different
quantities are related to each other by the identity

T x ǔ + T −1 −1
x ǔ = T y û + T y û

that holds on Ω [x y] (r |h). We may thus unambiguously define the discrete derivatives
v, w : Ω [x y] (r |h) → R of the conformal factor by
−1
T x ǔ − T y û T y û − T −1
x ǔ T x ǔ − T −1
y û T y û − T −1
x ǔ
v= = , w= = . (20)
   

Next, define the discrete unit tangent vectors a : Ω [x] (r |h) → S2 and b : Ω [y] (r |h)
→ S2 , respectively, by

a = e−û δx F  , b = e−ǔ δ y F  .

Since conformal squares are planar, there is a natural notion of normal field N :
Ω [x y] (r |h) → S2 , namely

T y δx F × T x δy F 
N= .
T y δx F  × T x δ y F 

With the help of the discrete orthonormal frame (a, b, N ), we introduce the discrete
scaled principal curvatures k : Ω [x x y] (r |h) → R and l : Ω [x yy] (r |h) → R, respec-
tively, by
k = −
T −1 −1
x N × T x N , b, l =
T y N × T y N , a. (21)

Note that k and l are equal to sin ∠(T −1 −1

x N , T x N ) and to sin ∠(T y N , T y N ), respec-
tively, with the signs chosen to maintain consistency with the continuous quantities.
Finally, to facilitate the calculations below, we need two more discrete functions
ṽ, w̃ : Ω [x y] (r |h) → R, given by
322 U. Bücking and D. Matthes

T y δx F  , T −1 
x δy F 

T −1  
y δx F , T x δ y F 
ṽ = =− ,
T y δx F  T −1 
x δy F T −1  
y δx F T x δ y F
(22)

T y δx F  , T x δ y F  
T −1  −1 
y δx F , T x δ y F 
w̃ = = − .
T y δx F  T x δ y F  T −1  −1
y δx F T x δ y F


The equalities follow since opposite angles in a conformal square sum up to π. The
two pairs (v, w) and (ṽ, w̃) are just different representations of the same geometric
information.
Lemma 3 There is a one-to-one correspondence between the pairs (v, w) and (ṽ, w̃)
of functions. Specifically, recalling the ∗ -notation introduced in (17),

ṽw̃∗ w̃ṽ∗
sinh(v) =  ∗
and sinh(w) =  ∗ . (23)
ṽ w̃

Moreover, the pair (v, w̃) uniquely determines the pair (ṽ, w), and vice versa.

Proof This is a general statement about four geometric quantities defined for con-
formal squares. It thus suffices to consider a single conformal square with vertices
−1 
p1 = T −1   
η F , p2 = T ξ F , p3 = T η F , p4 = T ξ F .

The respective four real numbers v, w, ṽ, w̃ are given by

p2 − p1 p3 − p2 p3 − p2 p3 − p4
ev = = , ew = = ,
p1 − p4 p4 − p3 p2 − p1 p4 − p1
ṽ = cos(∠ p1 p2 p3 ) = − cos(∠ p3 p4 p1 ), w̃ = cos(∠ p2 p3 p4 ) = − cos(∠ p4 p1 p2 ).

Observe that

p3 − p2 2 + p1 − p2 2 − 2
p3 − p2 , p1 − p2  = p3 − p4 2
+ p1 − p4 2 − 2
p3 − p4 , p1 − p4 

since both expressions are equal to p3 − p1 2 . Divide by p3 − p2 2 and use the

definitions of v, w, ṽ, w̃ to obtain, after simplification, that

1 + e−2w − 2ṽe−w = e−2v (1 + e−2w + 2ṽe−w ). (24)

The analogous considerations with p4 − p2 2 in place of p3 − p1 2 give (24) with

w̃ in place of ṽ, and with the roles of w and v exchanged. Clearly, these equations are
uniquely solvable for (ṽ, w̃) in terms of (v, w):

ṽ = tanh(v) cosh(w), w̃ = tanh(w) cosh(v). (25)

Constructing Solutions to the Björling Problem for Isothermic Surfaces … 323

Note that in particular

2 ṽw̃ = sinh(v) sinh(w). (26)

To derive (23) from here, take the square of the equations in (25), and express cosh2
and tanh2 in terms of sinh2 only. Then use (26) to eliminate sinh2 (w) from the first
equation and sinh2 (v) from the second one. This yields
 ṽw̃∗ 2  w̃ṽ∗ 2
sinh2 (v) =  ∗ , sinh2 (w) =  ∗ .
ṽ w̃

Now take the square root, bearing in mind that v, ṽ have the same sign, and w, w̃
have the same sign by (25).
Finally, to calculate ṽ from a given (v, w̃) using the first relation in (23), it suffices
to invert the (strictly increasing) function ṽ → ṽ/ṽ∗ . Then, knowing ṽ and w̃, the
value of w can be obtained from the second relation in (23). 

Recall that all discrete quantities defined above depend on the parameter . To stress
this fact, we will in the following use the superscript .
For later reference, we draw some first consequences of the definitions above.
Specifically, we summarize the relations between the geometric quantities (a  , b ,
û  , ǔ  ), and, of course, to F  itself, to the more abstract quantities (v , w , k , l ) that
satisfy the Gauss-Codazzi system (31)–(34). These relations can be seen as a discrete
analog of the frame equations (6) and (7).
Lemma 4 On Ω [x x yy] (r |h), one has

δ y F  = exp(û  )a  , δx F  = exp(ǔ  )b , (27)

δ y û  = w − v , δx ǔ  = w + v , (28)
  ∗   
 (ṽ )   −1  1 (ṽ )∗
δy a = w̃ − ṽ T x b + − 1 T −1 
y a , (29)
(w̃ )∗  (w̃ )∗
  ∗   
 (ṽ )   −1  1 (ṽ )∗
δx b = w̃ − ṽ T y a + − 1 T −1 
x b . (30)
(w̃ )∗  (w̃ )∗

Proof The two equations in (28) are obtained by rearranging the identities in (20).
For the derivation of (29), one makes the ansatz

T ya = μa T −1  −1 
y a + μb T x b .

Such a representation of T y a  must exist since elementary squares are mapped to

(flat) quadrilaterals by F  . The coefficients μa and μb can be determined by solving
the system of equations
324 U. Bücking and D. Matthes

1 = T y a  2 = μa2 + μ2b − 2μa μb w̃ , ṽ =

T y a  , T −1  
x b  = −μa w̃ + μb .

The analogous ansatz—with the roles of a  and b interchanged—leads to (30). 

3.5 Discrete Gauss-Codazzi System

This section is devoted to derive a discrete version of the Gauss-Codazzi equa-

tions (9)–(12). The following definition is needed to classify the difference between
the continuous and the discrete system.
Definition 6 A family (h  )>0 of real functions on respective domains D ⊂ Rn
is called asymptotically analytic on Cn if the following is true. For every M > 0,
there is an (M) > 0 such that each h  with 0 <  < (M) extends from D to a
complex-analytic function h̃  : DnM → C on the n-dimensional complex multi-disc
  
DnM = z = (z 1 , . . . , z n ) ∈ Cn  |z j | < M for each j = 1, . . . , n .

And the extensions h̃  are bounded on DnM , uniformly in 0 <  < (M).
The prototypical example for a family (h  )>0 that is asymptotically analytic on
C is given by h  (z) = 1/z ∗ = (1 − 2 z 2 )−1/2 . It is further easily seen that also the
functions g = −2 (h  − 1) form such a family; this is a very strong way of saying
that h  = 1 + O(2 ).
Proposition 2 There are four families (h 1, )>0 , . . . , (h 4, )>0 of asymptotically
analytic functions on C8 for which the following is true: let any -discrete isothermic
surface F  : Ω(r |h) → R3 be given, and define the functions v , w , k , l accord-
ingly. Then the following system of discrete equations is satisfied on Ω [x y] (r |h):

δη v  = δξ w  , (31)

δη w = δξ v − (T −1
  −1   
y k )(T x l ) + h 2 (T θ ), (32)

δy k = (T −1  −1   2  
x l )(T η w − T ξ v ) +  h 3 (T θ ) (33)
 −1 
δx l = (T −1  −1  
y k )(T η w + T ξ v ) + h 4 (T θ ),

(34)

where the h j are evaluated on

 
Tθ

= Tξv

, T ξ w , T −1  −1  −1  −1  −1  −1 
ξ v , Tξ w , Tη v , Tη w , T y k , Tx l .

Remark 5 Equations (31)–(34) are explicit in η-direction in the sense that they
express the “unknown” quantities T η v , T η w , T y k and T x l in terms of the “given”
eight quantities summarized in T θ .
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 325

Fig. 4 Four elementary

squares with discrete
quantities for the Cauchy
problem

The rest of this section is devoted to the proof of Proposition 2. Since  > 0 is
fixed in the derivation of (31)–(34), we shall omit the superscript  on the occurring
quantities.
For the derivation of (31)–(34), one can obviously work locally: it suffices to fix
some point in Ω [x x yy] (r |h) and to consider the eight values of v, w on the midpoints
of the four elementary squares incident to that vertex, and the four values of k, l on
the respective connecting edges.
The setup is visualized in Fig. 4. The “unknown” quantities v+ , w+ and k+ , l+ are
marked by ◦, the “given” quantities v0 , w0 , v L , w L , v R , w R and k0 , l0 are marked by •.
To facilitate the calculations, we also assume that values for a0 , b0 , û 0 , ǔ 0 , N0 , N L ,
N R are given; and then obtain the values of a+ , b+ , û + , ǔ + , N+ , see Fig. 4 right.
Naturally, the final formulas for v+ , w+ and k+ , l+ will be independent of these
quantities.

3.5.1 Derivation of Equation (31)

Compare the following two alternative ways to calculate û + , the logarithmic length
of the edge separating the right and the top plaquettes, from û 0 , the logarithmic length
of the edge between the plaquettes at bottom and left:

eû 0 ev0 ew R = eû+ = eû 0 ewL ev+

holds by property (20) of the functions v and w. Take the logarithm to obtain (31).

3.5.2 Derivation of Equation (32)

First recall that by Lemma 3, there is a one-to-one correspondence between (v, w)

and (ṽ, w̃), so we can assume that values for (ṽ0 , w̃0 ), (ṽ L , w̃ L ), (ṽ R , w̃ R ) are given
as well. Using that N R is the normalized cross product a+ × b0 , it is elementary to
derive the following representation of a+ :
326 U. Bücking and D. Matthes

a+ = ṽ R b0 + ṽ∗R (b0 × N R ). (35)

Taking the scalar product with N0 , one obtains

a+ , N0  = ṽ∗R
b0 , N R × N0  = ṽ∗R k0 .

Hence a+ can be expanded in the basis a0 , b0 and N0 as follows:

a+ = μa a0 + μb b0 + ṽ∗R k0 N0 (36)

with some real coefficients μa and μb to be determined. Calculating the square norm
on both sides gives

1 = μa2 + μ2b + 2w̃0 μa μb + 2 (ṽ∗R )2 k20 , (37)

and the scalar product with b0 yields

ṽ R = w̃0 μa + μb . (38)

Use (38) to eliminate μb from (37), then solve for μa . This gives

ṽ∗R k∗0 ṽ∗R k∗0

μa = , μb = ṽ R −  w̃0 . (39)
w̃0∗ ṽ0∗

On the other hand, starting from

b+ = λb b0 + λa a0 + ṽ∗L l0 N0 (40)

instead of (36), one obtains by analogous calculations that

ṽ∗L l∗0 ṽ∗L l∗0

λb = , λa = −ṽ L −  w̃0 .
w̃0∗ w̃0∗

Since w̃+ = −
a+ , b+ , it eventually follows that

ṽ∗L ṽ∗R k∗0 l∗0 ṽ∗L l∗0 ṽ∗R k∗0

w̃+ = w̃0 − ṽ R + ṽ L − ṽ∗R ṽ∗L k0 l0
(w̃0∗ )2 w̃0∗ w̃0∗
 
ṽ∗R ṽ L k∗0 − ṽ R ṽ∗L l∗0 ṽ∗R ṽ∗L k∗0 l∗0 2
+  w̃0 −ṽ R ṽ L +
2
w̃0 + w̃ . (41)
w̃0∗ (w̃0∗ )2 0

Next, recall that one may consider (v, w) as a function of (ṽ, w̃). More precisely,
by (23), one has that v and w approximate ṽ and w̃, respectively, to order 2 , in the
sense that the family of functions
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 327
 
(ṽ, w̃) → −2 (v − ṽ), −2 (w − w̃)
   ṽw̃∗     w̃ṽ∗  
1 −1 1 −1
= 2  arsinh  ∗ − ṽ , 2  arsinh  ∗ − w̃
 ṽ  w̃

is asymptotically analytic on C2 , see Definition 6. Observe further that −2 (1 − ṽ∗ )

etc. are asymptotically analytic as well. With this, it is straight-forward to conclude
(32) from (41).

3.5.3 Derivation of Equation (33)

In analogy to (35), one obtains by elementary considerations the following represen-

tation of a+ :
∗
a+ = −w̃+ b+ + w̃+ (b+ × N+ ).

Using the definition (21) of k, it then follows that

∗ ∗

a+ , N L  = w̃+
b+ , N+ × N L  = w̃+ k+ .

On the other hand, the computation (36)–(39) implies that

a+ , N L  = μb
b0 , N L  + ṽ∗R k0
N0 , N L 
μb
= ∗
b0 , a0 × b+  + ṽ∗R l∗0 k0
ṽ L
μb w̃∗
= − ∗ 0
b+ , N0  + ṽ∗R l∗0 k0
ṽ L
= −2 (w̃0∗ ṽ R − ṽ∗R k∗0 w̃0 )l0 + ṽ∗R l∗0 k0 .

In combination, this yields

∗
w̃+ k+ = ṽ∗R l∗0 k0 + (ṽ∗R k∗0 w̃0 − w̃0∗ ṽ R )l0 .

We can now substitute (41) to express the unknown w̃+ in terms of the known
quantities only. Using once again that −2 (1 − w̃+
∗
) etc. are asymptotically analytic
according to Definition 6, we arrive at (33).
The derivation of Eq. (34) is analogous.

4 The Abstract Convergence Result

In this section, we analyze the convergence of solutions to the classical Gauss-

Codazzi system (9)–(12) by solutions to the discrete system (31)–(34). This is the
core part of the convergence proof, from which our main result will be easily deduced
in the next section.
328 U. Bücking and D. Matthes

4.1 Domains

A key concept in the proof is to work with analytic extensions of the quantities v, w, k
and l defined in Sect. 3.4. The analytic setting forces us to introduce yet another class
of domains, and corresponding spaces of real analytic functions. In the following, we
assume that r > 0 and ρ̄ > 0 are fixed parameters (which will be frequently omitted
in notations), while h ∈ (0, ρ̄) and  > 0 may vary, with the restriction that  < h.
For each domain Ω(r |h), introduce its analytic fattening Ω ρ̄ (r |h) as follows:
 
ρ̄ (r |h) = (ξ, η) ∈ C × R ; ∃(ξ  , η) ∈ Ω(r |h) s.t. |ξ − ξ  |/ρ̄ + |η|/ h < 1 .
Ω

On these domains, we introduce the function class

   
 |h) := f : Ω
C ω Ω(r ρ̄ (r |h) → C ; f (·, η) is real analytic, for each η .

Notice that we require analyticity with respect to ξ, but not even continuity
 
 |h) ,
with respect to η. Next, introduce semi-norms |·|η,ρ for functions f ∈ C ω Ω(r
depending on parameters η ∈ [−h, h] and ρ ∈ [0, ρ̄] with ρ/ρ̄ + |η|/ h < 1 as fol-
lows:
 
| f |η,ρ = sup | f (ξ, η)| ; ξ ∈ C s.t. ∃(ξ  , η) ∈ Ω(r |h) with |ξ − ξ  | < ρ .

These semi-norms are perfectly suited to apply Cauchy estimates; indeed, one easily
proves with the Cauchy integral formula that
  1
∂ξ f  ≤ | f |η,ρ , (42)
η,ρ ρ − ρ

 that ρ > ρ. The semi-norms are now combined into a genuine norm · h
provided
ω 
on C Ω(r |h) as follows:
 
|η| ρ
f h = sup Λ(η, ρ) | f |η,ρ ; + <1 , (43)
h ρ̄

where the positive weight Λ is given by

|η|/ h
Λ(η, ρ) = 1 − . (44)
1 − ρ/ρ̄
 
This norm makes C ω Ω(r |h) a Banach space.
 
There is another semi-norm · h,δ that will be of importance below: for each
δ ∈ [0, 1], let
 
  |η| ρ
f h,δ = sup | f |η,ρ ; + ≤1−δ .
h ρ̄
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 329

By definition (44) of the weight Λ, the following estimate is immediate:

 
f h,δ ≤ δ −1 f h , (45)

provided that δ > 0.

Replacing Ω(r |h) by Ω [x y] (r |h) above yields  definitions for analytically fattened
domains Ω ρ̄[x y] (r |h) with respective spaces C ω Ω
[x y] (r |h) , semi-norms |·|[x y]
η,ρ and
 [x y] [x y]
· h,δ , and norms · h etc.

4.2 Statement of the Approximation Result

Recall that r > 0 and ρ̄ > 0 are fixed parameters.

Definition 7 An analytic solution θ = (v, w, k, l) of the classical Gauss-Codazzi
system on Ω ρ̄ (r |h) consists of four functions v, w, k, l ∈ C ω Ω(r  |h) that are globally

bounded on Ωρ̄ (r |h), are continuously differentiable with respect to η, and satisfy
Eqs. (9)–(12) on Ω ρ̄ (r |h).
An analytic solution θ = (v , w , k , l ) of the -discrete  [xGauss-Codazzi

system on Ω ρ̄ (r |h) consists of four functions v , w ∈ C ω Ω  y] (r |h) , k ∈
   
Cω Ω [x x y] (r |h) , l ∈ C ω Ω ρ̄[x x yy] (r |h).
[x yy] (r |h) that satisfy Eqs. (31)–(34) on Ω

A suitable norm to measure the deviation of an -discrete solution θ to a classical

ρ̄ (r |h) is given by the norms of the differences of
solution θ on the same domain Ω
the four components,
 
[x y] [x y] [x x y] [x yy]
|||θ − θ|||h = max v − v h , w − w h , k − k h , l − l h .

Proposition 3 Let an analytic solution θ to the Gauss-Codazzi system on Ω ρ̄ (r |h̄)


be given, and consider a family (θ )>0 of (a priori not necessarily analytic) solutions
θ = (v , w , k , l ) to the -discrete Gauss-Codazzi equations on Ω(r |h  ). Then there
are numbers A, B > 0 and ¯ > 0 such that the following is true for all  ∈ (0, ¯): if
θ possesses sufficient regularity to admit ξ-analytic complex extensions for η near
zero such that

|||θ − θ ||| < A, (46)

then θ as a whole extends to an analytic solution θ of the -discrete Gauss-Codazzi

system on Ω ρ̄ (r |h) , and

|||θ − θ |||h  ≤ B. (47)

330 U. Bücking and D. Matthes

Remark 6 The formulation of the proposition suggests that the height h of the domain
on which convergence takes place is small. However, this is misleading in general.
As it turns out in the proof, the limitation for h is mostly determined by the value
of ρ̄. In many examples of interest, ρ̄ is large compared to the region of interest
(determined by h̄ and r ), and consequently, one has h  = h̄ above, i.e., convergence
takes place on the entire domain of definition of θ.

The rest of this section is devoted to the proof of Proposition 3.

4.3 Consistency

We start with an evaluation of the difference between the classical and the -discrete
Gauss-Codazzi equations. Here, we need yet another measure for the deviation of θ
from θ:
    [x y]  [x y]  [x x y]  [x yy] 
{θ − θ} h,δ = max v − v h,δ , w − w h,δ , k − k h,δ , l − l h,δ .

This semi-norm is similar to |||θ − θ|||h . For further reference, we note that

   1
{θ − θ} h,δ ≤ |||θ − θ|||h , (48)
δ
thanks to (45), provided that δ > 0. Furthermore, we denote for abbreviation the
difference between corresponding discrete and continuous quantities by Δ, i.e. Δv =
v − v etc.
Lemma 5 Let an analytic solution θ to the classical Gauss-Codazzi system and
an analytic solution θ to the -discrete Gauss-Codazzi  [x x yy] system
 be given, both on
ρ̄ (r |h). Define the residuals g̃1 , . . . , g̃4 ∈ C ω Ω
Ω  (r |h) by

δη Δv = δξ Δw + g̃1 (49)


δη Δw = −δξ Δv + T −1
  −1  −1  −1 
y k T x Δl + T y Δk T x l +  g̃2 (50)
δ y Δk = (T −1  −1   −1  −1 
x l ) (T η Δw − T ξ Δv ) + (T x Δl ) (T η w − T ξ v) +  g̃3 (51)
 −1 −1
δx Δl = (T −1  −1   −1  −1 
y k ) (T η Δw + T ξ Δv ) + (T y Δk ) (T η w + T ξ v) +  g̃4 . (52)

Then the g̃ j are uniformly bounded with respect to  < ¯ on their respective domains:

ρ̄[x x yy] (r |h), for each j = 1, . . . , 4,

|g̃ j | ≤ G on Ω (53)
 
with a suitable constant G that depends on θ, and θ only via {θ − θ} h,0 , but is
independent of .
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 331

Proof By analyticity of θ it is clear that the central difference quotients obey

 
δξ v = ∂ξ v + gv,ξ , δη v = ∂η v + gv,η etc.
 [x x yy] 

with functions gv,ξ 
, gv,η 
, . . . ∈ Cω Ω (r |h) that are bounded uniformly w.r.t.
. The classical Gauss-Codazzi system (9)–(12) thus implies that

δη v = δξ w + g1 ,
δη w = −δξ v − (T −1 −1 
y k)(T x l) + g2 ,

δ y k = (T −1 −1 
x l)(T η w − T ξ v) + g3 ,
−1
δx l = (T −1 −1 
y k)(T η w + T ξ v) + g4 ,

where each of the functions g j is bounded on Ω [x x yy] (r |h), with an -independent

bound. Taking the difference between each equation of this system and the respective
equation of the -discrete Gauss-Codazzi equation (31)–(34) yields (49)–(52), with

g̃ j = h j (T θ ) − g j .

Since the h j are asymptotically analytic on C8 , it follows that the modulus of h j (T θ )

is uniformly controlled on Ω [x x yy] (r |h) by the supremum of the modulus of (θ )’s
components. 

4.4 Stability

Stability is shown inductively. More precisely, we prove for each n = 1, 2, . . . with

n 2 ≤ h that

|||θ − θ|||n 2 < B. (54)

In fact, there is nothing to show for n = 1. For n = 2, the claim (54) is a consequence
of estimate (46) on the initial data. Now assume that (54) has been shown for some
n ≥ 2. We are going to extend the estimate to n + 1.
Estimate on Δv . We begin by proving the estimate for the v-component of Δθ .
Since v is defined on Ωρ̄[x y] (r |h), the step n → n + 1 requires to estimate the values
of Δv (·, η ) for η ∈ ((n − 1) 2 , n 2 ]. Choose such an η ∗ , and define accordingly 
 ∗ ∗

such that η0∗ := η ∗ −  ∈ (− 2 , 2 ]; in fact, 2 = n if n is even, and n = 2 + 1 if n

is odd. For 0 ≤ k ≤ 2, introduce

ηk∗ = η ∗ − (2 − k) ; (55)
2
332 U. Bücking and D. Matthes

non-integer values of k are admitted. (55) is consistent with the definition of η0∗ , and
∗
moreover, η ∗ = η2 . Using the evolution equation (49), we obtain


  
Δv (·, η ∗ ) = Δv (·, η0∗ ) + T η Δv

− T −1  ∗
η Δv (·, η2k−1 )
k=1

 


= Δv (·, η0∗ ) + δξ Δw  ∗
(·, η2k−1 ) +
2
g̃1 (·, η2k−1
∗
). (56)
k=1 k=1

Next, pick a ρ∗ > 0 such that

ρ∗ η∗
+ < 1. (57)
ρ̄ h

We estimate:

 

    [x x yy]
|Δv |[x y]  [x y] δξ Δw [x∗yξ] ∗ + 2 g̃  ∗
η ∗ ,ρ∗ ≤ |Δv |η ∗ ,ρ∗ + 
0 η ,ρ 2k−1
1 η ,ρ∗ 2k−1
k=1 k=1
=: (I) + (II) + (III). (58)

We consider the terms (I)−(III) separately. First, thanks to our hypothesis (46) on
the initial conditions, we find that
[x y]
(I) = |Δv |η∗ ,ρ∗ ≤ Δv [x

y]
≤ A.
0

[x y]
Second, recalling the definition of · h , and using a Cauchy estimate (42), we
∗ ∗
obtain for given ρ2k−1 > ρ —yet to be determined—

   |Δw | ∗ [x y]
       η2k−1 ,ρ∗2k−1
(II) =  δξ Δw [x∗yξ] ∗ ≤  ∂ξ Δw [x∗y] ∗ ≤ 
η2k−1 ,ρ η2k−1 ,ρ ∗
k=1 k=1 k=1
ρ2k−1 − ρ∗
  
 1
≤ ∗ ∗ )Λ(η ∗ ∗
Δw [x y]
n 2 .
k=1
(ρ 2k−1 − ρ 2k−1 , ρ2k−1 )

We make the particular choice

 ∗ 
ρ̄ η2k−1 ρ∗
ρ∗2k−1 := 1− + ,
2 h ρ̄
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 333

which yields that

 ∗ 
ρ̄ η2k−1 ρ∗
ρ∗2k−1 −ρ = ∗
1− − ,
2 h ρ̄
∗
η2k−1 ρ∗2k−1 ∗
η2k−1 ρ∗
∗
1− − ρ̄
1− − ρ̄
, ρ∗2k−1 )
h h
Λ(η2k−1 = ρ∗2k−1
= ∗
η2k−1
.
ρ∗
1− ρ̄
1+ h
− ρ̄

And so we obtain
      
2 η∗ ρ∗ ∗
η2k−1 ρ∗ −2
(II) ≤ 1+ −  1− − Δw [x y]
n 2 .
ρ̄ h ρ̄ k=1
h ρ̄

To estimate the sum above, define ϕ : (η0∗ , η ∗ ) → R by

 
η ρ∗ −2
ϕ(η) = 1 − − .
h ρ̄

Since ϕ is a convex function, Jensen’s inequality implies that

 ∗
  ∗

η2k η2k
∗ ∗ 1 ∗
ϕ(η) dη ≥ (η2k − η2k−2 )ϕ ∗ ∗ η dη = ϕ(η2k−1 ).
∗
η2k−2 η2k − η2k−2 ∗
η2k−2

Hence the sum is bounded from above by the respective integral,

    η2∗  
η∗ ρ∗ −2 η ρ∗ −2
 1 − 2k−1 − ≤ 1− − dη
k=1
h ρ̄ η0∗ h ρ̄
 
η∗ ρ∗ −1 h
≤h 1− − = .
h ρ̄ Λ(η ∗ , ρ∗ ) (1 − ρ∗ /ρ̄)

The last term (III) is estimated with the help of the bound (53).  However,
 there is a
subtlety: a priori, the constant G there is controlled in terms of {θ − θ} n  ,0 , but the
2
induction estimate (54) is not sufficient to provide such a uniform bound, due to the
weight Λ. Fortunately, a close inspection of the terms in (III) reveals in combination
with (57) that we only need bounds on |g̃ j |η,ρ where ρ/ρ̄ < 1 − η ∗ / h − 2 / h. It
 
is easily deduced from Lemma 5 that an -uniform estimate on {θ − θ} n  ,δ with
2
δ := ρ̄/ h 2 > 0 suffices in this case, and the latter is obtained by combining (54)
with (48). Enlarging G if necessary, we arrive at



(III) ≤ 2 G = ()G ≤ Gh.
k=1
334 U. Bücking and D. Matthes

After multiplication of (58) by Λ(η ∗ , ρ∗ ) ≤ 1, we arrive at

[x y] 2h 1 + η ∗ / h − ρ∗ /ρ̄
Λ(η ∗ , ρ∗ ) |Δv |η∗ ,ρ∗ ≤ A + Δw [x y]
n 2 + Gh
ρ̄ 1 − ρ∗ /ρ̄
 
4h
≤ A+ B + Gh , (59)
ρ̄

[x y]
where we have used the induction hypothesis (54) for estimation of Δw n  , and
2
the relation (57) for estimation of the quotient. We have just proven inequality (59)
for every η ∗ ∈ ((n − 1) 2 , n 2 ], and for every ρ∗ ≥ 0 that satisfies (57). Taking the
supremum with respect to these quantities yields
 
4h
Δv [x y]
(n+1)  ≤ A+ B + Gh . (60)
2 ρ̄

Estimate on Δw . For estimation of the w-component, let η ∗ ∈ ((n − 1) 2 , n 2 ] be

given as before, and define ηk∗ as in (55). In analogy to (56), we have


 

Δw (·, η ∗ ) = Δw (·, η0∗ ) +  δξ Δv (·, η2k−1
∗
) + 2 g̃2 (·, η2k−1
∗
)
k=1 k=1

 

 −1  −1 
 ∗
 −1  −1
 ∗
+ T y k T x Δl (·, η2k−1 )+ T y Δk T x l (·, η2k−1 ).
k=1 k=1

[x y]
Taking the |·|η∗ ,ρ∗ -norm on both sides, multiplying by Λ(η ∗ , ρ∗ ) < 1, and estimating
the first couple of terms as above, we find that
 
 [x y]  4h
Λ(η ∗ , ρ∗ ) Δw η∗ ,ρ∗ ≤ A+ B + Gh 
ρ̄

 
   [x x y]   [x yy] [x yy]   [x x y]
+   ∗ ∗   ∗ ∗  
k η∗ ,ρ∗ Λ(η , ρ ) Δl η∗ ,ρ∗ + |l|η∗ ,ρ∗ Λ(η , ρ ) Δk η∗ ,ρ∗
k=1 2k− 23 2k− 23 2k− 23 2k− 23

 
  
4h    [x x y]  [x x y] [x yy]  [x x y]
≤ A+ B + Gh  +   
k η∗ ,ρ∗ Δl n 
+ |l|η∗ ,ρ∗ Δk n  .
ρ̄ 3 2 2k− 2 3 2 2k− 2
k=1
(61)

ρ̄ (r |h̄), and so

On the one hand, the analytic solution θ is bounded on Ω

k n[xx y] ≤ Θ := sup |θ|. (62)

2
ρ̄ (r |h̄)
Ω
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 335

∗ ∗
On the other hand, since η2k− 3 ≤ η − 4 , and because of (57), we have that
3
2

∗ ∗
1 − η2k− 3 / h − ρ /ρ̄ 3
∗ ∗
Λ(η2k− 3,ρ ) = 2
≥ ,
2 1 − ρ∗ /ρ̄ 4h

and therefore, using the induction hypothesis (54),

Δk n[xx y]

|k |η[x∗x y],ρ∗ ≤ |k|η[x∗x y],ρ∗ + |Δk |η[x∗x y],ρ∗ ≤ sup |θ| + ∗ ∗
2

2k− 23 2k− 23 2k− 23 ρ̄ (r |h̄)

Ω Λ(η2k− 3,ρ )
2

B 4
≤Θ+ = Θ + Bh. (63)
(3)/(4h) 3
[x yy] [x yy]
The remaining terms Δk n  and Δl n  in (61) can be estimated directly
2 2
by (54). Substitution of these partial estimates into (61), and recalling that  ≤ h,
leads to
   
[x y] 4 4
Λ(η ∗ , ρ∗ ) |Δv |η∗ ,ρ∗ ≤ A + + 2Θ + Bh Bh + Gh . (64)
ρ̄ 3

Estimate on Δk . Finally, let us estimate Δk (·, η ∗ ) at some η ∗ ∈ ((n − 23 ) 2 , (n −

1 
) ]. For the estimates below, let in addition a ξ ∗ ∈ C be given such that (ξ ∗ , η ∗ ) ∈
2 2
ρ̄[x x y] (r |h). We need to use a slightly different normalization for the ηk∗ in (55): write
Ω
∗  ∗  
η ∗ = η− 1 + m 2 for suitable η 1 ∈ (− 4 , 4 ] and a (uniquely determined) m ∈ N.
−2
2
Now define
1  1 
ξk∗ = ξ ∗ + (m − k + ) , ηk∗ = η ∗ − (m − k + ) ;
2 2 2 2
∗ ∗
note that ξ ∗ = ξm−1/2 and η ∗ = ηm− 1 . With these notations:
2


m−1
  ∗ ∗
Δk (ξ ∗ , η ∗ ) = Δk (ξ−∗ 1 , η−
∗
1) + T y Δk

− T −1 
y Δk (ξk , ηk )
2 2
k=0


m−1
 
= Δk (ξ−∗ 1 , η−
∗
1) +  l (ξk+
∗
1,η
∗
k− 1
) Δw (ξk∗ , ηk−1
∗
) − Δv (ξk+1
∗
, ηk )
2 2 2 2
k=0


m−1
  
m−1
+ Δl (ξk+
∗
1,η
∗
k−
∗ ∗ ∗ ∗
1 ) w(ξk , ηk−1 ) − v(ξk+1 , ηk ) + 
2
g̃3 (ξk∗ , ηk∗ ).
2 2
k=0 k=0

It is straight-forward to verify that all the terms on the right-hand side are well-defined
for the given arguments. For a given ρ∗ that satisfies (57), we apply the semi-norm
|·|η[x∗x,ρy]
∗ to both sides and estimate further, using the triangle inequality:
336 U. Bücking and D. Matthes


m−1  
|Δk |η[x∗x,ρy]  [x x y]
∗ ≤ |Δk | ∗
η ,ρ∗ +  |l |η[x∗x y],ρ∗ |Δw |[x y]  [x y]
ηk−1 ,ρ∗ + |Δv |ηk ,ρ∗
− 21 k− 21
k=0


m−1     [x y]
m−1
+ |Δl |η[x∗x y],ρ∗ |w|[x y]
∗ + |v| [x y]
∗ + 2 g̃   ∗ ∗ . (65)
ηk−1 ,ρ η ,ρ ∗ 3 η ,ρ
k− 21 k k
k=0 k=0

On the one hand, we have that

|w|[x y] [x y]

ηk−1 ,ρ∗ + |v|ηk ,ρ∗ ≤ sup |w| + sup |v| ≤ 2Θ,
ρ̄ (r |h̄)
Ω ρ̄ (r |h̄)
Ω

with the bound Θ from (62). And on the other hand, arguing like in (63) on grounds
∗ ∗
of ηk− 1 ≤ η − 4  for all k = 0, . . . , m − 1, we have the estimate
3
2

4
|l |η[x∗x y],ρ∗ ≤ |l|η[x∗x y],ρ∗ + |Δl |η[x∗x y],ρ∗ ≤ Θ + Bh.
k− 21 k− 21 k− 21 3

Substitute this into (65) and multiply by Λ(η ∗ , ρ∗ ) to obtain

   
[x x y] 1
Δk (n+1)  ≤ A + 4 Θ + Bh Bh + Gh . (66)
2 3

Estimate on Δl . This is completely analogous to the estimate for Δk above.
Summarizing the results in (60), (64) and (66), we obtain (54) with n + 1 in
place of n, for an arbitrary choice of B > A, and any corresponding h > 0 that is
sufficiently small to make the coefficients in front of  in (60), (64) and (66) smaller
than B. Notice that the implied smallness condition on h is independent of .

5 The Continuous Limit of Discrete Isothermic Surfaces

We are finally in the position to formulate and prove our main approximation result.

5.1 From Björling Data to Cauchy Data and Back

Given analytic Björling data ( f, n) in the sense of Theorem 1, first compute the
associated frame Ψ 0 , the conformal factor u 0 , and the derived quantities v0 , w0 , k0 , l0
as functions on (−r, r ) as detailed in the proof there. We claim that, for any sufficiently
small  > 0, associated Björling data f  : Ω(r | 2 ) → R3 for construction of an -
discrete isothermic surface can be prescribed such that the following are true:
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 337

(1) The initial surface piece and its tangent vectors are approximated to first order
in ,

f  (ξ, η) = f (ξ) + O(),

δx f  (ξ, η) = exp(u 0 (ξ))Ψ10 (ξ) + O(),
δ y f  (ξ, η) = exp(u 0 (ξ))Ψ20 (ξ) + O(), (67)

where the O() indicate -smallness that is uniform in (ξ, η) on the domains
Ω(r | 2 ) for f  , and Ω [x] (r | 2 ) for δx f  , and Ω [y] (r | 2 ) for δ y f  , respectively.
(2) The derived quantities (v , w , k , l ) satisfy

v (ξ, η) = v0 (ξ), w̃ (ξ, η) = w0 (ξ), k (ξ, η) = k0 (ξ), l (ξ, η) = l0 (ξ),

(68)

at each point (ξ, η) in Ω [x y] (r |) for v , w̃ , in Ω [x x y] (r |) for k , and in
Ω [x yy] (r |) for l , respectively.
Notice that the data (v , w , k , l ) are ξ-analytic quantities; ironically, one cannot
even expect continuity of the respective data f  in general.
For later reference, we briefly sketch one possible construction of such data f  . We
start by defining f  on point triples in the strip −3 4 < ξ ≤ 3 4 : let (ξ, η) ∈ Ω(r | 2 )
be a point with − 4 < ξ ≤ 4 . We distinguish two cases. If 0 < η ≤ 2 , then we define
f  (ξ, η − 2 ) = f (0), and there is a unique way to assign data f  at the two points
(ξ − 2 , η) and (ξ + 2 , η) such that for the vectors

1     1    
a= f (ξ + , η) − f  (ξ, η − ) , b = f (ξ − , η) − f  (ξ, η − ) ,
 2 2  2 2
the following is true:
(1) a is parallel to Ψ10 (0), and b is orthogonal to n(0),

a, b
(2) w0 (ξ) = ,
a b
 0   
(3) a = exp u (0) + 2 v0 (ξ) and b = exp u 0 (0) − 2 v0 (ξ) .
If instead − 2 < η ≤ 0, then we define f  (ξ, η + 2 ) = f (0), and we assign data f 
at (ξ + 2 , η) and at (ξ − 2 , η) with the respective adaptations for the conditions on
the vectors.
Up to here, there has been a certain degree of freedom in the choice of the f  . From
now on, there is a unique way to extend the already prescribed f  to all of Ω(r | 2 ) such
that (68)—and, incidentally, also (67)—holds. We briefly indicate how to proceed
in the next step; the further steps are then made inductively in the same way. Let
(ξ, η) be a point with 3 4 < ξ ≤ 5 4 , and with − 2 < η ≤ 0. Note that f  is already
defined at the following points: (ξ − , η), (ξ − 2 , η + 2 ) and (ξ − 3 2 , η + 2 ). Let
us introduce the vectors
1     1    
a= f (ξ − , η + ) − f  (ξ − , η) , b = f (ξ − 3 , η + ) − f  (ξ − , η) .
 2 2  2 2
338 U. Bücking and D. Matthes

Then, there is a unique choice for f  (ξ, η) such that the new vector

1    
c= f (ξ − , η + ) − f  (ξ, η)
 2 2
satisfies the following conditions:
(1) the sin-value of the angle between the planes spanned by (a, b) and by (b, c),
respectively, equals to k0 (ξ − 3 4 ),

a, c 
(2) = w0 (ξ − ),
a c 2
(3) c = b exp( 2 v0 (ξ − 2 )).
By continuing this construction in an inductive manner, we enlarge the domain of
definition with respect to ξ by 2 in both directions in each step, until f  is defined
on all of Ω(r | 2 ). It is obvious from the construction that (68) holds. The verification
of (67) is a tedious but straight-forward exercise in elementary geometry that we leave
to the interested reader. An important point is that the aforementioned construction
only uses data that can be obtained very directly from the Björling data ( f, n). Indeed,
the calculation of u 0 , Ψ 0 and (v0 , w0 , k0 , l0 ) from ( f, n) only involves differentiation
and inversion of matrices. In particular, all operations are local.
Definition 8 Assume that analytic Björling data ( f, n) and discrete data f  : Ω(r | 2 )
are given such that (67) and (68) are satisfied. The maximal -discrete isothermic
surface F  : Ω(r |h  ) that is obtained from f  as Björling data—see Proposition 1—is
referred to as grown from ( f, n).

5.2 Main Result

The central approximation result is the following.

Theorem 2 Let analytic Björling data ( f, n) on (−r, r ) be given, and let F :
Ω(r |h̄) → R3 be the corresponding real-analytic isothermic surface. Further, let
F  : Ω(r |h  ) → R3 be the family of -discrete isothermic surfaces that are grown
from ( f, n).
Then, there are some h > 0 and C > 0, such that for all sufficiently small  > 0,
we have that h  ≥ h, and

F  (ξ, η) − F(ξ, η) ≤ C, δx F  (ξ,η) − Fx (ξ, η) ≤ C,

δ y F  (ξ, η) − Fy (ξ, η) ≤ C, (69)

for all (ξ, η) ∈ Ω [x y] (r |h).

Theorem 2 above gives an answer to the question of how to approximate the unique
isothermic surface F that is determined by given Björling data by a family of -
isothermic surfaces F  . Our construction of the surface F  is completely explicit,
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 339

Fig. 5 A part of a sphere in different degrees of discretization

Fig. 6 Examples of discrete isothermic surfaces: torus (left), hyperbolic paraboloid (right)

and it requires no a priori knowledge about F. The plots in Figs. 5 and 6 illustrate
that our construction can be used to generate pictures of the surfaces F  with just a
few lines of code.

Remark 7 Note that if the discrete isothermic surfaces F  converge to a smooth

isothermic surface F, then also the discrete Christoffel and Darboux transforms of
F  converge to the corresponding smooth Christoffel and Darboux transforms of F.
This will be proven in Sect. 6.

Proof (of Theorem 2) Since F is real-analytic on Ω(r |h̄), the derived quantities
 w and k, l are real-analytic there as well, and can be extended to functions in
u, v,
 |h̄) , for a suitable choice of the “fattening parameter” ρ̄ > 0, after dimin-
C ω Ω(r
ishing h̄ > 0 if necessary. The extensions satisfy the Gauss-Codazzi system (9)–(12)
on the complexified domain.
Next, consider the -discrete isothermic surfaces F  : Ω(r |h  ) → R that are
grown from the Björling data ( f, n). Define the associated quantities v , w , k , l .
Thanks to (68), these are real analytic functions on Ω [x y] (r |), and they extend
complex-analytically w.r.t. ξ to Ω ρ̄[x y] (r |). Since the quadruple (v , w , k , l ) sat-
isfies the discrete Gauss-Codazzi equations (31)–(34), the ξ-analyticity is prop-
 
 from the initial strip to the maximal domain of existence, i.e., v , w ∈
agated
ω [x y]
C Ω (r |h  ) etc.
Moreover, again thanks to (68), the differences Δv , Δk and Δl are of order
O() on the initial strip:

Δv [x

y]
≤ A, Δk [x x y] ≤ A, Δl [x

yy]
≤ A,
340 U. Bücking and D. Matthes

with a suitable -independent constant A. For the remaining differences Δw , it

follows via Lemma 3 on the equivalence of (v , w ) and (ṽ , w̃ ) that they satisfy the
same estimate (enlarging A if necessary):

Δw [x

y]
≤ A.

We are thus in the situation to apply Proposition 3. From the estimate (47), it follows
in particular that

v = v + O(), w = w + O(), k = k + O(), l = l + O(). (70)

Here and below, we use the slightly ambiguous notation O() to express that the
discrete quantities approximate the associated continuous ones uniformly on their
respective (real) domains Ω [x y] (r |h  ) or Ω [x x y] (r |h  ), Ω [x yy] (r |h  ), with a maximal
error of order .
Next, we conclude from (70) that also

a  = a + O(), b = b + O(), û  = u + O(), ǔ  = u + O(). (71)

Indeed, it follows directly from the definition of these quantities that (71) implies
  
δx F  = exp(û  )a  = exp u + O() a + O() = exp(u)a + O() = ∂x F + O(),

and, likewise, δ y F  = ∂ y F + O(), which, eventually, implies further that also

F  = F + O(), thanks to F  (ξ, η) = F(0) for − 4 < ξ < 4 and − 2 < η ≤ 2 by
construction. Therefore, our claim (69) is a consequence of (71).
We only sketch the proof of (71), that is little more than a repeated application of
the Gronwall lemma. For further technical details, we refer the reader to [2], where
the relevant estimates have been carried out in a very similar situation, see the proof
of Theorem 5.4 therein.
First of all, the claim (71) holds on the strip Ω [x y] (r |) thanks to (67). Now
compare the frame equations (6) & (7) with their discrete analogs from (28)–(30):

∂x u = w + v and δx ǔ  = w + v ,
∂ y u = w − v and δ y û  = w − v ,
∂x b = (w − v)(T −1    −1  −1 
y a + O()) and δx b = (w − v + O())T y a + O()T x b ,
∂ y a = (w − v)(T −1    −1  −1 
x b + O()) and δ y a = (w − v + O())T x b + O()T y a .

Subtract the respective equations, recall (70), and use a standard Gronwall argument
to conclude that the validity of (71) extends from the “initial strip” to the entire
domain Ω [x x yy] (r |h).
A posteriori, we conclude that h  ↑ h as  ↓ 0. where h > 0 is the constant
obtained in the proof of Proposition 3. Indeed, thanks to the uniform closeness of
the discrete tangent vectors δx F  , δ y F  to their continuous counterparts ∂x F, ∂ y F—
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 341

which are orthogonal with non-vanishing length—it easily follows that there cannot
occur any degeneracies in F  at any h  < h. 

6 Transformations

Isothermic surfaces have an exceptionally rich transformation theory. For the defi-
nition of discrete isothermic surfaces used in this paper this transformation theory
carries over to the discrete setup.
We consider two important transformations, namely the Christoffel transformation
and Darboux transformation. Their analogs for discrete isothermic surfaces may for
example be found in [3, 4, 14–16]. It is a natural question whether the convergence
results of Theorem 2 can be generalized to imply the convergence of the transformed
surfaces.

6.1 Christoffel Transformation

We briefly remind the classical definition of the Christoffel transformation. The

included existence claim was first proved by Christoffel [10].
Definition 9 Let F : Ω(r |h) → R3 be an isothermic surface. Then the R3 -valued
one-form d F  defined by

Fx Fy
Fx = , Fy = − ,
Fx 2 Fy 2

is closed. The surface F  : Ω(r |h) → R3 , defined (up to translation) by integration

of this one-form, is isothermic and is called dual to the surface F or Christoffel
transform of the surface F.
Given any isothermic surface F and its dual F  , straightforward calculation leads
to the following relations between corresponding quantities.

Fx = e−2u Fx , Fy = −e−2u Fy , N  = −N ,

u  = −u, v = −v, w = −w, k = −k, l = l.

The discrete case is nearly the same, see for example [3].
342 U. Bücking and D. Matthes

Definition 10 Let F  : Ω(r |h) → R3 be a discrete isothermic surface. Then the

R3 -valued discrete one-form δ(F  ) defined by

δx F  δy F 
δx (F  ) = , δ y (F  ) = − ,
δx F  2 δ y F  2

is closed. The surface (F  ) : Ω(r |h) → R3 , defined (up to translation) by integration

of this discrete one-form, is a discrete isothermic surface and is called dual to F  or
Christoffel transform of F  .
Corollary 1 Under the assumptions of Theorem 2 not only the discrete isothermic
surface itself converges to the corresponding smooth isothermic surface, but also the
discrete Christoffel transforms converge to the corresponding Christoffel transforms
of the smooth isothermic surface.
Proof Recall the definitions of the discrete quantities in Sect. 3.4. We immediately
deduce the following relations for a discrete isothermic surface F  and its dual (F  ) .
For better reading we omit the superscript .

a  = a, b = −b, û  = −û, ǔ  = −ǔ, N  = −N ,

ṽ = −ṽ, w̃ = −w̃, v = −v, w = −w, k = −k, l = l.

Now the proof follows directly from the corresponding proofs in Sects. 4 and 5. 

6.2 Darboux Transformation

The Darboux transformation for isothermic surfaces was introduced by Darboux [12].
It is a special case of a Ribaucour transformation and is closely connected to Möbius
geometry as well as to the integrable system approach to isothermic surfaces, see for
example [15].
Definition 11 Let F : Ω(r |h) → R3 be an isothermic surface. Then the isothermic
surface F + : Ω(r |h) → R3 is called a Darboux transform of F if
 
F + − F 2 F+ − F F+ − F
Fx+ =− Fx − 2
Fx ,  ,
C Fx 2 F + − F F + − F
 
F + − F 2 F+ − F F+ − F
Fy+ = Fy − 2
F y ,  ,
C Fy 2 F + − F F + − F

where C ∈ R, C = 0, is a constant which is called parameter of the Darboux trans-

formation.
In the discrete case the definition of the Darboux transformation may be interpreted as
“discrete Ribaucour transformation” using intersecting instead of touching spheres.
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 343

In particular, recall the definition of the cross-ratio q( p1 , p2 , p3 , p4 ) of four coplanar

points p1 , . . . , p4 . After identification of their common plane with the complex plane
C the cross-ratio may be calculated by the formula

q( p1 , p2 , p3 , p4 ) := ( p1 − p2 )( p2 − p3 )−1 ( p3 − p4 )( p4 − p1 )−1

The following definition first appeared in [16], see also [4, 15].

Definition 12 Let F  : Ω(r |h) → R3 be a discrete isothermic surface. Then the dis-
crete isothermic surface (F + ) : Ω(r |h) → R3 is called a discrete Darboux trans-
form of F  if the following conditions are satisfied.
(i) The four points T −1   −1 +  + 
x F , T x F , T x (F ) , T x (F ) lie in a common plane and
the same is true for T y F , T y F , T y (F ) , T y (F + ) .
−1   −1 + 

1
(ii) q(T −1   +  −1 + 
x F , T x F , T x (F ) , T x (F ) ) = and
γ
1
q(T −1   +  −1 + 
y F , T y F , T y (F ) , T y (F ) ) = − , where γ ∈ R, γ  = 0 is a constant
γ
which is called parameter of the Darboux transformation.

Note that given any discrete isothermic surface, a discrete Darboux transform may
be obtained by prescribing the value of (F + ) at one point and using the conditions
of the definition to successively build a new surface which is also discrete isothermic
(as long as the surface does not degenerate).
In order to obtain convergence of the discrete Darboux transform to the corre-
sponding continuous one, we choose γ = C/2 .

Corollary 2 Under the assumptions of Theorem 2 not only the discrete isothermic
surface itself converges to the corresponding smooth isothermic surface, but also
the discrete Darboux transforms (with γ = C/2 ) converge to the corresponding
Darboux transforms (with parameter C) of the smooth isothermic surface.

Proof Assume that the discrete isothermic surface itself converges to the corre-
sponding smooth isothermic surface with errors of order O() as in the proofs of
Theorem 2. Now start with (F + ) (0, 0) = F + (0, 0) and build the discrete Darboux
transform successively using the above definition. Denote the distance between cor-
responding points by d  = (F + ) − F  .
In order to relate corresponding discrete and smooth quantities, we first use the
simple equivalence

( p2 − p1 )( p4 − p3 ) 1 ( p4 − p1 ) − ( p2 − p1 )
= ⇐⇒ p3 − p2 = . (72)
( p3 − p2 )( p1 − p4 ) q 1 − q (( pp24 − p1 )
− p1 )
344 U. Bücking and D. Matthes

Then we use the fact that in our case ( p2 − p1 ) = O() and q = C/2 . Thus we
obtain that
 
( p4 − p1 ) −  ( p2 − p1 ) ( p4 − p1 )2 ( p2 − p1 )
p3 − p2 = = ( p4 − p1 ) +  − + O (2 ).
1− 2 ( p4 − p1 )
C( p2 − p1 )
C ( p2 − p1 ) 

Now we identify
( p2 − p1 )
( p3 − p2 ) = T x d  , ( p4 − p1 ) = T −1 
x d , = δx F 

and easily deduce by straightforward identifications of complex numbers and vectors
that
T x d − T −1
x d
 1
= (− T −1  2  −1  −1   
x d δx F + 2 T x d
T x d , δx F ) − δx F + O ()
 C δx F  2
= Fx+ − Fx + O ().

Analogously, we obtain
T yd − T −1
y d

1
= ( T −1  2  −1  −1   
y d δ y F − 2 T y d
T y d , δ y F ) − δ y F + O ()
 C δ y F  2
= Fy+ − Fy + O ().

Thus starting with (F + ) (0, 0) = F + (0, 0) and building the discrete Darboux
transform successively using the above definition, in each step we add an error of
order O(2 ) to the difference d = F + − F of the Darboux pair. Therefore we obtain
(F + ) = F + + O() as claimed. 
Remark 8 Using the definitions of continuous and discrete Darboux transformations,
corresponding formulas for a + , b+ , û + , ǔ + , N + , ṽ+ , w̃+ , v+ , w+ , k+ , l+ may be
deduced, which also converge under the assumptions of Theorem 2.

Acknowledgments The authors would like to thank Sepp Dorfmeister and Fran Burstall for stim-
ulating and helpful discussions. We further thank the anonymous referee for the careful reading of
the initial manuscript and various suggestions for improvement. This research was supported by
the DFG Collaborative Research Center TRR 109 “Discretization in Geometry and Dynamics”.
Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.
Constructing Solutions to the Björling Problem for Isothermic Surfaces … 345

References

1. Bobenko, A.I., Matthes, D., Suris, Y.B.: Discrete and smooth orthogonal systems: C ∞ -
approximation. Int. Math. Res. Not. 45, 2415–2459 (2003)
2. Bobenko, A.I., Matthes, D., Suris, Y.B.: Nonlinear hyperbolic equations in surface theory:
integrable discretizations and approximation results. St. Petersburg Math. J. 17, 39–61 (2006)
3. Bobenko, A.I., Pinkall, U.: Discrete isothermic surfaces. J. reine angew. Math. 475, 187–208
(1996)
4. Bobenko, A.I., Suris, Y.B.: Discrete differential geometry. Integrable structure. In: Graduate
Studies in Mathematics, vol. 98. AMS (2008)
5. Brander, D., Dorfmeister, J.F.: The Björling problem for non-minimal constant mean curvature
surfaces. Commun. Anal. Geom. 18(1), 171–194 (2010)
6. Bücking, U.: Approximation of conformal mapping by circle patterns. Geom. Dedicata 137,
163–197 (2008)
7. Burstall, F.: Isothermic surfaces: conformal geometry, Clifford algebras and integrable systems.
In: Terng, C.L. (ed.) Integrable Systems, Geometry and Topology, AMS/IP Studies in Advanced
Math., vol. 36, pp. 1–82. American Mathematical Society (2006)
8. Burstall, F., Hertrich-Jeromin, U., Pedit, F., Pinkall, U.: Curved flats and isothermic surfaces.
Math. Z. 225(2), 199–209 (1997)
9. Cayley, A.: On the determination of the surfaces divisible into squares by means of their curves
of curvature. Proc. London Math. Soc. 4, 8–9 (1872)
10. Christoffel, E.: Ueber einige allgemeine Eigenschaften der Minimumsflächen. Crelles J. 67,
218–228 (1867)
11. Cieśliński, J., Goldstein, P., Sym, A.: Isothermic surfaces in E3 as soliton surfaces. Phys. Lett.
A 205(1), 37–43 (1995)
12. Darboux, G.: Sur les surfaces isothermiques. Comptes Rendus 128(1299–1305), 1538 (1899)
13. Dierkes, U., Hildebrandt, S., Küster, A., Wohlrab, O.: Minimal surfaces. I, Grundlehren der
Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 295.
Springer-Verlag, Berlin (1992)
14. Hertrich-Jeromin, U.: Transformations of discrete isothermic nets and discrete cmc-1 surfaces
in hyperbolic space. Manuscr. Math. 102, 465–486 (2000)
15. Hertrich-Jeromin, U.: Introduction to Möbius differential geometry. In: London Mathematical
Society Lecture Note Series, vol. 300. Cambridge University Press, Cambridge (2003)
16. Hertrich-Jeromin, U., Hoffmann, T., Pinkall, U.: A discrete version of the Darboux transform
for isothermic surfaces. In: Bobenko, A.I., Seiler, R. (eds.) Discrete Integrable Geometry and
Physics. Oxford Lecture Ser. Math. Appl, vol. 16, pp. 59–81. Clarendon Press, Oxford (1999)
17. Matthes, D.: Convergence in discrete Cauchy problems and applications to circle patterns.
Conform. Geom. Dyn. 9, 1–23 (2005)
18. Nagumo, M.: Über das Anfangswertproblem partieller Differentialgleichungen. Jap. J. Math.
18, 41–47 (1942)
19. Nirenberg, L.: An abstract form of the nonlinear Cauchy-Kowalewski theorem. J. Differ. Geom.
6, 561–576 (1972)
20. Schief, W.K.: Isothermic surfaces in spaces of arbitrary dimension: integrability, discretization,
and Bäcklund transformations - a discrete Calapso equation. Stud. Appl. Math. 106(1), 85–137
(2001)
21. Schramm, O.: Circle patterns with the combinatorics of the square grid. Duke Math. J. 86(2),
347–389 (1997)
22. Stephenson, K.: Introduction to circle packing: the theory of discrete analytic functions. Cam-
bridge University Press, New York (2005)
23. Walter, W.: An elementary proof of the Cauchy-Kowalevsky theorem. Amer. Math. Monthly
92(2), 115–126 (1985)
24. Weingarten, J.: Über die Differentialgleichung der Oberflächen, welche durch ihre Krüm-
mungslinien in unendlich kleine Quadrate geteilt werden können. Sitzungsber. königl. preuß.
Akad. Wiss. pp. 1163–1166 (1883)
On the Lagrangian Structure of Integrable
Hierarchies

Yuri B. Suris and Mats Vermeeren

Abstract We develop the concept of pluri-Lagrangian structures for integrable hier-

archies. This is a continuous counterpart of the pluri-Lagrangian (or Lagrangian mul-
tiform) theory of integrable lattice systems. We derive the multi-time Euler Lagrange
equations in their full generality for hierarchies of two-dimensional systems, and con-
struct a pluri-Lagrangian formulation of the potential Korteweg-de Vries hierarchy.

1 Introduction

In this paper, our departure point are two developments which have taken place in
the field of discrete integrable systems in recent years.
• Firstly, multi-dimensional consistency of lattice systems has been proposed as a
notion of integrability [8, 15]. In retrospect, this notion can be seen as a discrete
counterpart of the well-known fact that integrable systems never appear alone but
are organized into integrable hierarchies. Based on the notion of multi-dimensional
consistency, a classification of two-dimensional integrable lattice systems (the so
called ABS list) was given in [1]. Moreover, for all equations of the ABS list,
considered as equations on Z2 , a variational interpretation was found in [1].
• Secondly, the idea of the multi-dimensional consistency was blended with the vari-
ational formulation in [13], where it was shown that solutions of any ABS equation

on any quad surface Σ in Z N are critical points of a certain action functional Σ L
obtained by integration of a suitable discrete Lagrangian two-form L . Moreover,
it was observed in [13] that the critical value of the action remains invariant under
local changes of the underlying quad-surface, or, in other words, that the 2-form
L is closed on solutions of quad-equations, and it was suggested to consider this
as a defining feature of integrability. However, later research [10] revealed that

Y.B. Suris (B) · M. Vermeeren

Inst. für Mathematik, Technische Universität Berlin, Straße des 17. Juni 136,
10623 Berlin, Germany
e-mail: [email protected]
M. Vermeeren
e-mail: [email protected]
© The Author(s) 2016 347
A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_11
348 Y.B. Suris and M. Vermeeren

L is closed not only on solutions of (non-variational) quad-equations, but also

on general solutions of the corresponding Euler-Lagrange equations. Therefore, at
least for discrete systems, the closedness condition is implicitly contained in the
variational formulation.
A general theory of multi-time one-dimensional Lagrangian systems, both discrete
and continuous, has been developed in [20]. A first attempt to formulate the theory for
continuous two-dimensional systems wasmade in [21]. For such systems, a solution
is a critical point of the action functional S L on any two-dimensional surface S in
R N , where L is a suitable differential two-form. The treatment in [21] was restricted
to second order Lagrangians, i.e. to two-forms L that only depend on the second jet
bundle. In the present work we will extend this to Lagrangians of any order.
As argued in [10], the unconventional idea to consider the action on arbitrary
two-dimensional surfaces in the multi-dimensional space of independent variables
has significant precursors. These include:
• Theory of pluriharmonic functions and, more generally, of pluriharmonic maps
[11, 17, 19]. By definition, a pluriharmonic function of several
 complex variables
f : C N → R minimizes the Dirichlet functional E Γ = Γ |( f ◦ Γ )z |2 dz ∧ d z̄
along any holomorphic curve in its domain Γ : C → C N . Differential equations
governing pluriharmonic functions,

∂2 f
=0 for all i, j = 1, . . . , N ,
∂z i ∂ z̄ j

are heavily overdetermined. Therefore it is not surprising that pluriharmonic func-

tions (and maps) belong to the theory of integrable systems.
This motivates the term pluri-Lagrangian systems, which was proposed in [9, 10].
• Baxter’s Z-invariance of solvable models of statistical mechanics [3, 4]. This
concept is based on invariance of the partition functions of solvable models under
elementary local transformations of the underlying planar graphs. It is well known
(see, e.g., [7]) that one can identify planar graphs underlying these models with
quad-surfaces in Z N . On the other hand, the classical mechanical analogue of
the partition function is the action functional. This suggests the relation of Z-
invariance to the concept of closedness of the Lagrangian 2-form, at least at the
heuristic level. This relation has been made mathematically precise for a number
of models, through the quasiclassical limit [5, 6].
• The classical notion of variational symmetry, going back to the seminal work of
E. Noether [16], has been shown to be directly related to the closedness of the
Lagrangian form in the multi-time [21].
The main goal of this paper is two-fold: to derive the Euler Lagrange equations
for two-dimensional pluri-Lagrangian problems of arbitrary order, and to state the
(potential) KdV hierarchy as a pluri-Lagrangian system. We will also discuss the
closedness of the Lagrangian two-form, which turns out to be related to the Hamil-
tonian theory of integrable hierarchies.
On the Lagrangian Structure of Integrable Hierarchies 349

Note that the influential monograph [12], according to the foreword, is “about
hierarchies of integrable equations rather than about individual equations”. However,
its Lagrangian part (Chaps. 19, 20) only deals with individual equations. The reason
for this is apparently the absence of the concept of pluri-Lagrangian systems. We hope
that this paper opens up the way for a variational approach to integrable hierarchies.

2 Pluri-Lagrangian Systems

2.1 Definition

We place our discussion in the formalism of the variational bicomplex as presented

in [12, Chap. 19] (and summarized, for the reader’s convenience, in Appendix A).
Slightly different versions of this theory can be found in [18] and in [2].
Consider a vector bundle X : R N → R and its nth jet bundle J n X . Let L ∈
(0,d)
A (J n X ) be a smooth horizontal d-form. In other words, L is a d-form on R N
whose coefficients depend on a function u : R N → R and its partial derivatives up
to order n. We call R N the multi-time, u the field, and L the Lagrangian d-form.
We will use coordinates (t1 , . . . , t N ) on R N . Recall that in the standard calculus of
variations the Lagrangian is a volume form, so that d = N .

 u solves the pluri-Lagrangian problem for L if

Definition 2.1 We say that the field
u is a critical point of the action S L simultaneously for all d-dimensional surfaces
S in R N . The equations describing this condition are called the multi-time Euler-
Lagrange equations. We say that they form a pluri-Lagrangian system and that L
is a pluri-Lagrangian structure for these equations.

To discuss critical points of a pluri-Lagrangian problem, consider the vertical

derivative δL of the (0,d)-form L in the variational bicomplex, and a variation
V . Note that we consider variations V as vertical vector fields; such a restriction is
justified by our interest, in the present paper, in autonomous systems only. Besides,
in the context of discrete systems only vertical vectorfields seem to possess a natural
analogs. The criticality condition of the action, δ S L = 0, is described by the
equation 
ιpr V δL = 0, (1)
S

which has to be satisfied for any variation V on S that vanishes at the boundary
∂ S. Recall that pr V is the nth jet prolongation of the vertical vector field V , and
that ι stands for the contraction. One fundamental property of critical points can be
established right from the outset.

Proposition 2.2 The exterior derivative dL of the Lagrangian is constant on crit-

ical points u.
350 Y.B. Suris and M. Vermeeren

Proof Consider a critical point u and a small (d + 1)-dimensional ball B. Because

S := ∂ B has no boundary, Eq. (1) is satisfied for any variation V . Using Stokes’
theorem and the properties that δd + dδ = 0 and ιpr V d + d ιpr V = 0 (Propositions
A.1 and A.4 in Appendix A), we find that
   
0= ιpr V δL = d(ιpr V δL ) = − ιpr V d(δL ) = ιpr V δ(dL ).
∂B B B B

Since this holds for any ball B it follows that ιpr V δ(dL ) = 0 for any variation V
of a critical point u. Therefore, δ(dL ) = 0, so that dL is constant on critical points
u. Note that here we silently assume that the space of critical points is connected. It
would be difficult to justify this property in any generality, but it is usually clear in
applications, where the critical points are solutions of certain well-posed systems of
partial differential equations. 


We will take a closer look at the property dL = const in Sect. 6, when we discuss
the link with Hamiltonian theory. It will be shown that vanishing of this constant,
i.e., closedness of L on critical points, is related to integrability of the multi-time
Euler-Lagrange equations.

2.2 Approximation by Stepped Surfaces

For computations, we will use the multi-index notation for partial derivatives. For
any multi-index I = (i 1 , . . . , i N ) we set

∂ |I | u
uI = ,
(∂t1 )i1 . . . (∂t N )i N

where |I | = i 1 + . . . + i N . The notations I k and I k α will represent the multi-indices

(i 1 , . . . , i k + 1, . . . i N ) and (i 1 , . . . , i k + α, . . . i N ) respectively. When convenient we
will also use the notations I tk and I tkα for these multi-indices. We will write k ∈ / I
if i k = 0 and k ∈ I if i k > 0. We will denote by Di or Dti the total derivative with
respect to coordinate direction ti ,
 ∂
Di := Dti := uIi
I
∂u I

and by D I := Dit11 . . . Dit NN the corresponding higher order derivatives.

Our main general result is the derivation of the multi-time Euler-Lagrange equa-
tions for two-dimensional surfaces (d = 2). That will allow us to study the KdV
hierarchy as a pluri-Lagrangian system. However, it is instructive to first derive the
multi-time Euler-Lagrange equations for curves (d = 1).
On the Lagrangian Structure of Integrable Hierarchies 351

The key technical result used to derive multi-time Euler-Lagrange equations is

the observation that it suffices to consider a very specific type of surface.

Definition 2.3 A stepped d-surface is a d-surface that is a finite union of coordinate

d-surfaces. A coordinate d-surface of the direction (i 1 , . . . , i d ) is a d-surface lying
in an affine d-plane {(t1 , . . . , t N ) | t j = c j for j = i 1 , . . . , i d }.

Lemma 2.4 If the action is stationary on any stepped surface, then it is stationary
on any smooth surface.

The proof of this Lemma can be found in appendix B.

2.3 Multi-time Euler-Lagrange Equations for Curves

N
Theorem 2.5 Consider a Lagrangian 1-form L = i=1 L i dti . The multi-time
Euler-Lagrange equations for curves are:

δi L i
= 0 ∀I 
i, (2)
δu I
δi L i δj L j
= ∀I, (3)
δu I i δu I j

where i and j are distinct, and the following notation is used for the variational
derivative corresponding to the coordinate direction i:

δi L i  ∂ Li ∂ Li ∂ Li ∂ Li
= (−1)α Diα = − Di + Di2 − ....
δu I α≥0
∂u I i α ∂u I ∂u I i ∂u I i 2

Remark 2.6 In the special case that L only depends on the first jet bundle, system
(2)–(3) reduces to the equations found in [20]:

δi L i ∂ Li ∂ Li
=0 ⇔ − Di = 0,
δu ∂u ∂u i
δi L i ∂ Li
=0 ⇔ = 0 for i = j,
δu j ∂u j
δi L i δj L j ∂ Li ∂L j
= ⇔ = for i = j.
δu i δu j ∂u i ∂u j

Proof (of Theorem 2.5) According to Lemma 2.4, it is sufficient to look at a general
L-shaped curve S = Si ∪ S j , where Si is a line segment of the coordinate direction
i and S j is a line segment of the coordinate direction j. Denote the cusp by p :=
Si ∩ S j . We orient the curve such that Si induces the positive orientation on the
point p and S j the negative orientation. There are four cases, depending on how
352 Y.B. Suris and M. Vermeeren

tj tj tj tj
p p
Si Si
Sj Sj Sj Sj
Si Si
p p
ti ti ti ti
(+1, +1) (+1, −1) (−1, +1) (−1, −1)

Fig. 1 The four L-shaped curves with their values of (εi , ε j )

the L-shape is rotated. They are depicted in Fig. 1. To each case we associate a pair
(εi , ε j ) ∈ {−1, +1}2 , where the positive value is taken if the respective piece of curve
is oriented in the coordinate direction, and negative if it is oriented opposite to the
coordinate direction.
The variation of the action is
  
ιpr V δL = εi (ιpr V δL i ) dti + ε j (ιpr V δL j ) dt j
S Si Sj
   
∂ Li ∂L j
= εi δu I (V ) dti + ε j δu I (V ) dt j .
Si I
∂u I Sj I
∂u I

Note that these sums are actually finite. Indeed, since L depends on the nth jet
bundle all terms with |I | := i 1 + . . . + i N > n vanish.
Now we expand the sum in the first of the integrals and perform integration by
parts.

εi (ιpr V δL i ) dti
Si
  
∂ Li ∂ Li ∂ Li
= εi δu I (V ) + δu I i (V ) + δu I i 2 (V ) + . . . dti
Si I 
i ∂u I ∂u I i ∂u I i 2
  
∂ Li ∂ Li ∂ Li ∂ Li
= εi − Di + Di2 − Di3 + . . . δu I (V ) dti
Si I 
i ∂u I ∂u I i ∂u I i 2 ∂u I i 3
  ∂ Li ∂ Li ∂ Li
+ δu I (V ) + δu I i (V ) − Di δu I (V )
∂u I i ∂u I i 2 ∂u I i 2
I 
i

∂ Li ∂ Li ∂ Li 
+ δu I i 2 (V ) − Di (V )δu I i (V ) + Di2 δu I (V ) + . . .  .
∂u I i 3 ∂u I i 3 ∂u I i 3 p

Using the language of variational derivatives, this reads

  
δi L i
εi (ιpr V δL i ) dti = εi δu I (V ) dti
Si Si I 
i δu I
On the Lagrangian Structure of Integrable Hierarchies 353

  δi L i δi L i


+ δu I (V ) + δu I i (V ) + . . . 
I 
i
δu I i δu I i 2 p
    
δi L i δi L i 
= εi δu I (V ) dti + δu I (V )  .
Si I 
i δu I I
δu I i p

The other piece, S j , contributes

  
δj L j  δj L j 

εj ιpr V δL j dt j = ε j δu I (V ) dt j − δu I (V )  ,
Sj S j I 
j δu I I
δu Ij p

where the minus sign comes from the fact that S j induces negative orientation on the
point p. Summing the two contributions, we find
    
δi L i δj L j
ιpr V δL = εi δu I (V ) dti + ε j δu I (V ) dt j
S Si I 
i δu I S j I 
j δu I

  δi L i δj L j


+ δu I (V ) − δu I (V )  . (4)
I
δu I i δu I j p

Now require that the variation (4) of the action is zero for any variation V . If we
consider variations that vanish on S j , then we find for every multi-index I which
does not contain i that
δi L i
= 0.
δu I

Given this equation, and its analogue for the index j, only the last term remains in
the right hand side of Eq. (4). Considering variations around the cusp p we find for
every multi-index I that
δi L i δj L j
= .
δu I i δu I j

It is clear these equations combined are also sufficient for the action to be critical. 


2.4 Multi-time Euler-Lagrange Equations

for Two-Dimensional Surfaces

The two-dimensional case (d = 2) covers many known integrable hierarchies, includ-

ing the potential KdV hierarchy
which we will discuss in detail later on. We consider
a Lagrangian two-form L = i< j L i j dti ∧ dt j and we will use the notational con-
vention L ji = −L i j .
354 Y.B. Suris and M. Vermeeren

Theorem 2.7 The multi-time Euler-Lagrange equations for two-dimensional sur-

faces are

δi j L i j
= 0, ∀I 
i, j, (5)
δu I
δi j L i j δik L ik
= ∀I 
i, (6)
δu I j δu I k
δi j L i j δ jk L jk δki L ki
+ + =0 ∀I, (7)
δu I i j δu I jk δu I ki

where i, j and k are distinct, and the following notation is used for the variational
derivative corresponding to the coordinate directions i, j:

δi j L i j  β ∂ Li j
:= (−1)α+β Diα D j .
δu I α,β≥0
∂u I i α j β

Remark 2.8 In the special case that L only depends on the second jet bundle, this
system reduces to the equations stated in [21].
Before proceeding with the proof of Theorem 2.7, we introduce some terminol-
ogy and prove a lemma. A two-dimensional stepped surface consisting of q flat
pieces intersecting at some point p is called a q-flower around p, the flat pieces are
called its petals. If the action is stationary on every q-flower, it is stationary on any
stepped surface. By Lemma 2.4 the action will then be stationary on any surface.
The following Lemma shows that it is sufficient to consider 3-flowers.
Lemma 2.9 If the action is stationary on every 3-flower, then it is stationary on
every q-flower for any q > 3.
Proof Let F be a q-flower. Denote its petals corresponding to coordinate directions
(ti1 , ti2 ), (ti2 , ti3 ), . . . , (tiq , ti1 ) by S12 , S23 , . . . , Sq1 respectively. Consider the 3-flower
F123 = S12 ∪ S23 ∪ S31 , where S31 is a petal in the coordinate direction (ti3 , ti1 ) such
that F123 is a flower around the same point as F. Similarly, define F134 , . . . , F1 q−1 q .
Then (for any integrand)
  
+ + ... +
F123 F134 F1 q−1 q
        
= + + + + + +... + + + .
S12 S23 S31 S13 S34 S41 S1 q−1 Sq−1 q Sq1

Here, S21 , S32 , … are the petals S12 , S23 , … but with opposite orientation (see Fig. 2).
Therefore all terms where the index of S contains 1 cancel, except for the first and
last, leaving
       
+... + = + + +... + + = .
F123 F1 q−1 q S12 S23 S34 Sq−1 q Sq1 F
On the Lagrangian Structure of Integrable Hierarchies 355

Fig. 2 Two 3-flowers tj

composed to form a 4-flower.
The common petal does not
contribute to the integral
because it occurs twice with
opposite orientation
ti

tk

tl

By assumption the action is stationary on every 3-flower, so

  
ιpr V δL = ιpr V δL + . . . + ιpr V δL = 0.
F F123 F1 q−1 q




Proof (of Theorem 2.7) Consider a 3-flower S = Si j ∪ S jk ∪ Ski around the point
p = Si j ∩ S jk ∩ Ski . Denote its interior edges by

∂ Si := Si j ∩ Ski , ∂ S j := S jk ∩ Si j , ∂ Sk := Ski ∩ S jk .

On ∂ Si , ∂ S j and ∂ Sk we choose the orientations that induce negative orientation

on p. We consider the case where these orientations correspond to the coordinate
directions, as in Fig. 3. The cases where one or more of these orientations are opposite
to the corresponding coordinate direction (see Fig. 4) can be treated analogously and
yield the same result.

Fig. 3 A 3-flower. Different tj

petals induce the opposite
orientation on the common Sij
boundary
∂ Sj
Sjk
ti
p ∂Si
∂Sk

Ski
tk
356 Y.B. Suris and M. Vermeeren

tj

tj
tj
tk ti ti
ti tk

tk

Fig. 4 Three of the other 3-flowers. The orientations of the interior edges do not all correspond to
the coordinate direction

We choose the orientation on the petals in such a way that the orientations of ∂ Si ,
∂ S j and ∂ Sk are induced by Si j , S jk and Ski respectively. Then the orientations of
∂ Si , ∂ S j and ∂ Sk are the opposite of those induced by Ski , Si j and S jk respectively
(see Fig. 3).
We will calculate
   
ιpr V δL = ιpr V δL + ιpr V δL + ιpr V δL (8)
S Si j S jk Ski

and require it to be zero for any variation V which vanishes on the (outer) boundary
of S. This will give us the multi-time Euler-Lagrange equations.
For the first term of Eq. (8) we find
   ∂ Li j
ιpr V δL = δu I (V ) dti ∧ dt j
Si j Si j I
∂u I
   ∂ Li j
= δu I i λ j μ (V ) dti ∧ dt j .
Si j I 
i, j λ,μ≥0
∂u I i λ j μ

First we perform integration by parts with respect to ti as many times as possible.

    ∂ Li j
ιpr V δL = (−1)λ Diλ δu I j μ (V ) dti ∧ dt j
Si j Si j I 
i, j λ,μ≥0 ∂u I i λ j μ
    λ−1
∂ Li j
− (−1)π Diπ δu I i λ−π−1 j μ (V ) dt j .
∂ Sj I 
i, j λ,μ≥0 π=0
∂u I i λ j μ

Next integrate by parts with respect to t j as many times as possible.

    μ ∂ Li j
ιpr V δL = (−1)λ+μ Diλ D j δu I (V ) dti ∧ dt j (9)
Si j Si j I 
i, j λ,μ≥0 ∂u I i λ j μ
On the Lagrangian Structure of Integrable Hierarchies 357

    λ−1
∂ Li j
− (−1)π Diπ δu I i λ−π−1 j μ (V ) dt j (10)
∂ Sj I 
i, j λ,μ≥0 π=0
∂u Iiλ jμ

    μ−1
ρ ∂ Li j
− (−1)λ+ρ Diλ D j δu I j μ−ρ−1 (V ) dti . (11)
∂ Si I 
i, j λ,μ≥0 ρ=0
∂u I i λ jμ

The signs of (10) and (11) are due to the choice of orientations (see Fig. 3). We can
rewrite the integral (9) as
  δi j L i j
δu I (V ) dti ∧ dt j .
Si j I 
i, j
δu I

The last integral (11) takes a similar form if we replace the index μ by β = μ − ρ − 1.

    μ−1
ρ ∂ Li j
− (−1)λ+ρ Diλ D j δu I j μ−ρ−1 (V ) dti
∂ Si I 
i, j λ,μ≥0 ρ=0
∂u I i λ j μ
  
ρ ∂ Li j
=− (−1)λ+ρ Diλ D j δu I j β (V ) dti
∂ Si I 
i, j β,λ,ρ≥0 ∂u I i λ j β+ρ+1
 
δi j L i j
=− δu I j β (V ) dti .
∂ Si I 
i, j β≥0 δu I j β+1

To write the other boundary integral (10) in this form we first perform integration by
parts.
    λ−1
∂ Li j
− (−1)π Diπ δu I i λ−π−1 j μ (V ) dt j
∂ Sj I 
i, j λ,μ≥0 π=0
∂u I i λ j μ
 λ−1
   μ ∂ Li j
=− (−1)π+μ Diπ D j δu I i λ−π−1 (V ) dt j
∂ S j I 
i, j λ,μ≥0 π=0 ∂u I i λ j μ

   μ−1
λ−1   
π ρ ∂ Li j

+ (−1) π+ρ
Di D j δu I i λ−π−1 j μ−ρ−1 (V )  .
I 
i, j λ,μ≥0 π=0 ρ=0
∂u I i λ j μ p

Then we replace λ by α = λ − π − 1 and in the last term μ by β = μ − ρ − 1.

 λ−1
  ∂ Li j
− (−1)π Diπ δu I i λ−π−1 j μ (V ) dt j
∂ S j I 
i, j λ,≥0 π=0 ∂u I i λ j μ
   μ ∂ Li j
=− (−1)π+μ Diπ D j δu I i α (V ) dt j
∂ S j I 
i, j α,μ,π≥0 ∂u I i α+π+1 j μ
358 Y.B. Suris and M. Vermeeren

   
∂ Li j 
δu I i α j β (V ) 
ρ
+ (−1)π+ρ Diπ D j
I 
i, j α,β,π,ρ≥0
∂u I i α+π+1 j β+ρ+1
p
   δi j L i j    δi j L i j 

=− δu I i (V ) dt j +
α δu I i j (V )  .
α β
∂ Sj I 
i, j α≥0
δu I i α+1 I 
i, j α,β≥0
δu I i α+1 j β+1 p

Putting everything together we find

   δi j L i j  
δi j L i j
ιpr V δL = δu I (V ) dti ∧ dt j − δu I (V )dti
Si j Si j I 
i, j δu I ∂ Si I 
i δu I j
   
δi j L i j δi j L i j 
− δu I (V ) dt j + δu I (V )  .
∂ S j I 
j δu I i I
δu I i j p

Expressions for the integrals over S jk and Ski are found by cyclic permutation of the
indices. Finally we obtain
   δi j L i j
ιpr V δL = δu I (V ) dti ∧ dt j
S Si j I 
i, j δu I
   δki L ki 
δi j L i j
− δu I (V ) + δu I (V ) dti
∂ Si I 
i
δu I j I 
i
δu I k

 δi j L i j 

+ δu I (V )  + cyclic permutations in i, j, k.
I
δu I i j p
(12)

From this we can read off the multi-time Euler-Lagrange equations. 



3 Pluri-Lagrangian Structure of the Sine-Gordon Equation

We borrow our first example of a pluri-Lagrangian system from [21].

The Sine-Gordon equation u x y = sin u is the Euler-Lagrange equation for

1
L= u x u y − cos u.
2
∂
Consider the vector field ϕ ∂u with

1
ϕ = u x x x + u 3x
2
On the Lagrangian Structure of Integrable Hierarchies 359


and its prolongation Dϕ := I ϕ I ∂u∂ I . It is known that Dϕ is a variational symmetry
for the sine-Gordon equation [18, p. 336]. In particular, we have that

Dϕ L = Dx N + D y M (13)

with
1 1 1
M= ϕu x − u 4x + u 2x x ,
2 8 2
1 1 2
N = ϕu y − u x cos u − u x x (u x y − sin u).
2 2
Now we introduce a new independent variable z corresponding to the “flow”
of the generalized vector field Dϕ , i.e. u z = ϕ. Consider simultaneous solutions of
the Euler-Lagrange equation δδuL = 0 and of the flow u z = ϕ as functions of three
independent variables x, y, z. Then Eq. (13) expresses the closedness of the two-form

L = L d x ∧ dy − M dz ∧ d x − N dy ∧ dz.

The fact that dL = 0 on solutions is consistent with Proposition 2.2. Hence L is a

reasonable candidate for a Lagrangian two-form.

Theorem 3.1 The multi-time Euler-Lagrange equations for the Lagrangian two-
form
L = L 12 d x ∧ dy + L 13 d x ∧ dz + L 23 dy ∧ dz

with the components

1
L 12 = u x u y − cos u, (14)
2
1 1 1
L 13 = u x u z − u 4x + u 2x x , (15)
2 8 2
1 1 2
L 23 = − u y u z + u x cos u + u x x (u x y − sin u), (16)
2 2
consist of the sine-Gordon equation

u x y = sin u,

the modified KdV equation

1
u z = u x x x + u 3x ,
2

and corollaries thereof. On solutions of either of these equations the two-form L is

closed.
360 Y.B. Suris and M. Vermeeren

Proof Let us calculate the multi-time Euler-Lagrange Eqs. (5)–(7) one by one:
δ12 L 12
• The equation = 0 yields ..........................................u x y = sin u.
δu
δ12 L 12
For any α > 0 the equation = 0 yields 0 = 0.
δu z α
δ13 L 13
• The equation = 0 yields .............................u x z = 23 u 2x u x x + u x x x x .
δu
δ13 L 13
For any α > 0 the equation = 0 yields 0 = 0.
δu y α
δ23 L 23
• The equation = 0 yields ...................... u yz = 21 u 2x sin u + u x x cos u.
δu
δ23 L 23
The equation = 0 yields .................................... u yx x = u x cos u.
δu x
δ23 L 23
The equation = 0 yields ..........................................u x y = sin u.
δu x x
δ23 L 23
For any α > 2, the equation = 0 yields 0 = 0.
δu x α
δ13 L 13 δ23 L 23
• The equation = yields ............................ u z = u x x x + 21 u 3x .
δu x δu y
δ13 L 13 δ23 L 23
The equation = yields u x x = u x x .
δu x x δu x y
δ13 L 13 δ23 L 23
For any other I the equation = yields 0 = 0.
δu I x δu I y
δ12 L 12 δ13 L 13
• The equation = yields 21 u x = 21 u x .
δu y δu z
δ12 L 12 δ13 L 13
For any nonempty I , the equation = yields 0 = 0.
δu I y δu I z
δ12 L 12 δ23 L 32
• The equation = yields 21 u y = 21 u y .
δu x δu z
δ12 L 12 δ23 L 32
For any nonempty I , the equation = yields 0 = 0.
δu I x δu I z
δ12 L 12 δ23 L 23 δ13 L 31
• For any I the equation + + = 0 yields 0 = 0.
δu I x y δu I yz δu I zx
It remains to notice that all nontrivial equations in this list are corollaries of the
equations u x y = sin u and u z = u x x x + 21 u 3x , derived by differentiation.
The closedness of L can be verified by direct calculation:

1
Dz L 12 − D y L 13 + Dx L 23 = (u yz u x + u x z u y ) + u z sin u
2
1 1 1
− u yz u x − u z u x y + u 3x u x y − u x x u x x y
2 2 2
On the Lagrangian Structure of Integrable Hierarchies 361

1 1 1
− u x z u y − u z u x y + u x u x x cos u − u 3x sin u
2 2 2
+ u x x x (u x y − sin u) + u x x (u x x y − u x cos u)
 
1
= − u z − u 3x − u x x x (u x y − sin u).
2




Remark 3.2 The Sine-Gordon equation and the modified KdV equation are the sim-
plest equations of their respective hierarchies. Furthermore, those hierarchies can be
seen as the positive and negative parts of one single hierarchy that is infinite in both
directions [14, Sect. 3c and 5k]. It seems likely that this whole hierarchy possesses
a pluri-Lagrangian structure.

4 The KdV Hierarchy

Our second and the main example of a pluri-Lagrangian system will be the (potential)
KdV hierarchy. This section gives an overview of the relevant known facts about
KdV, mainly following Dickey [12, Sect. 3.7]. The next section will present its pluri-
Lagrangian structure.
One way to introduce the Korteweg-de Vries (KdV) hierarchy is to consider a
formal power series
∞
R= rk z −2k−1 ,
k=0

with the coefficients rk = rk [u] being polynomials of u and its partial derivatives
with respect to x, satisfying the equation

Rx x x + 4u Rx + 2u x R − z 2 Rx = 0. (17)

Multiplying this equation by R and integrating with respect to x we find

 
1 2 1
R Rx x − Rx + 2 u − z 2 R 2 = C(z), (18)
2 4

where C(z) = ∞ k=0 ck z
−2k
is a formal power series in z −2 , with coefficients ck being
constants. Different choices of C(z) correspond to different normalizations of the
KdV hierarchy. We take C(z) = 18 , i.e. c0 = 18 and ck = 0 for k > 0. The first few
coefficients of the power series R = r0 z −1 + r1 z −3 + r2 z −5 + . . . are

1
r0 = , r1 = u, r2 = u x x + 3u 2 , r3 = u x x x x + 10uu x x + 5u 2x + 10u 3 .
2
The Korteweg-de Vries hierarchy is defined as follows.
362 Y.B. Suris and M. Vermeeren

Definition 4.1 • The KdV hierarchy is the family of equations

u tk = (rk [u])x .

• Write gk [v] := rk [vx ]. The potential KdV (PKdV) hierarchy is the family of equa-
tions
vtk = gk [v].

• The differentiated potential KdV (DPKdV) hierarchy is the family of equations

vxtk = (gk [v])x .

The right-hand sides of first few PKdV equations are

g1 = vx , g2 = vx x x + 3v2x , g3 = vx x x x x + 10vx vx x x + 5v2x x + 10v3x .

Remark 4.2 The first KdV and PKdV equations, u t1 = u x , resp. vt1 = vx , allow us
to identify x with t1 .

Proposition 4.3 The differential polynomials rk [u] satisfy

δrk
= (4k − 2) rk−1 ,
δu
δ δ1
where δu
is shorthand notation for δu
.

A proof of this statement can be found in [12, 3.7.11–3.7.14].

Corollary 4.4 Set h k [v] := 4k+2

1
gk+1 [v], then the differential polynomials gk and h k
satisfy
δgk δh k
= (4k − 2) gk−1 and = gk .
δvx δvx

Before we proceed, let us formulate a simple Lemma.

Lemma 4.5 For any multi-index I and for any differential polynomial f [v] we have:
 
δf ∂f δf
Dx = − .
δv I x ∂v I δv I

Proof By direct calculation:

   
δf ∂f ∂f ∂f
Dx = Dx − Dx + D2x − ...
δv I x ∂v I x ∂v I x 2 ∂v I x 3
On the Lagrangian Structure of Integrable Hierarchies 363

∂f ∂f ∂f ∂f δf
= Dx − D2x + D2x − ... = − .
∂v I x ∂v I x 2 ∂v I x 3 ∂v I δv I




We can now find Lagrangians for the the DPKdV equations.

Proposition 4.6 The DPKdV equations are Lagrangian, with the Lagrange func-
tions
1
L k [v] = vx vtk − h k [v].
2

Proof Since h k = 4k+2

1
gk+1 does not depend on v directly, it follows from Lemma
4.5 and Corollary 4.4 that

δ Lk δh k δh k
= −vtk x − = −vtk x + Dx = −vtk x + (gk )x .
δv δv δvx



5 Pluri-Lagrangian Structure of PKdV Hierarchy

Since the individual KdV and PKdV equations are evolutionary (not variational), it
seems not very plausible that they could have a pluri-Lagrangian structure. However,
it turns out that the PKdV hierarchy as a whole is pluri-Lagrangian. Let us stress that
this structure is only visible if one considers several PKdV equations simultaneously
and not individually. We consider a finite-dimensional multi-time R N parametrized
by t1 , t2 , . . . , t N supporting the first N flows of the PKdV hierarchy. Recall that the
first PKdV equation reads vt1 = vx , which allows us to identify t1 with x.
The formulation of the main result involves certain differential polynomials intro-
duced in the following statement.
Lemma 5.1 • There exist differential polynomials bi j [v] depending on v and vx α ,
α > 0, such that
Dx (gi )g j = Dx (bi j ). (19)

• These polynomials satisfy

bi j + b ji = gi g j . (20)

• The differential polynomials ai j [v] (depending on vx α and vx α t j , α ≥ 0) defined by

δ1 h i δ1 h i δ1 h i
ai j := vt j + vxt j + vx xt j + ... (21)
δvx δvx x δvx x x
364 Y.B. Suris and M. Vermeeren

satisfy
D j (h i ) + Dx (gi )vt j = Dx (ai j ). (22)

Proof The existence of polynomials bi j is shown in [12, 3.7.9]. Since

Dx (bi j + b ji ) = Dx (gi )g j + gi Dx (g j ) = Dx (gi g j ),

and since neither bi j + b ji nor gi g j contain constant terms, Eq. (20) follows. The last
claim is a straightforward calculation using Lemma 4.5:
 
δ1 h i δ1 h i δ1 h i
Dx (ai j ) = Dx vt j + vxt j + vx xt j + ...
δvx δvx x δvx x x
δ1 h i δ1 h i δ1 h i
= vxt j + vx xt j + vx x xt j + ...
δvx δvx x δvx x x
     
δ1 h i δ1 h i δ1 h i
+ vt j Dx + vxt j Dx + vx xt j Dx + ...
δvx δvx x δvx x x
δ1 h i δ1 h i δ1 h i
= vxt j + vx xt j + vx x xt j + ...
δvx δvx x δvx x x
δ1 h i ∂h i δ1 h i ∂h i δ1 h i ∂h i
− vt j + vt j − vxt j + vxt j − vx xt j + vx xt j − ...
δv ∂v δvx ∂vx δvx x ∂vx x
δ1 h i
= D j h i − vt j = D j h i + Dx (gi )vt j .
δv



Now we are in a position to give a pluri-Lagrangian formulation of the PKdV

hierarchy.
Theorem 5.2
 The multi-time Euler-Lagrange equations for the Lagrangian two-
form L = i< j L i j dti ∧ dt j , with coefficients given by

1
L 1i := L i = vx vti − h i (23)
2
and
1 1
L i j := (vti g j − vt j gi ) + (ai j − a ji ) − (bi j − b ji ) for j > i > 1 (24)
2 2
are the first N − 1 nontrivial PKdV equations

vt2 = g2 , vt3 = g3 , . . . vt N = g N ,

and equations that follow from these by differentiation.

On the Lagrangian Structure of Integrable Hierarchies 365

5.1 Variational Symmetries and the Pluri-Lagrangian Form

Before proving Theorem 5.2, let us give an heuristic derivation of expression (24) for
L i j . The ansatz is that different flows of the PKdV hierarchy should be variational
symmetries of each other. (We are grateful to V. Adler who proposed this derivation
to us in a private communication.)
Fix two distinct integers i, j ∈ {2, 3, . . . , N }. Consider the the ith DPKdV equa-
tion, which is nothing but the conventional two-dimensional variational system gen-
erated in the (x, ti )-plane by the Lagrange function

1
L 1i [v] = vx vti − h i [v].
2
Consider the evolutionary equation vt j = g j [v], i.e., the jth PKdV equation, and the
corresponding generalized vector field
 ∂
Dg j := (D I g j ) .
I 
j
∂v I

We want to show that Dg j is a variational symmetry of L 1i . For this end, we look for
L i j such that

(g )
Dg j (L 1i ) − Di L 1 j j + Dx (L i j ) = 0. (25)

(g )
Here, L 1 j j is the Lagrangian defined by (23) but with vt j replaced by g j :

(g ) 1
L 1 j j := vx g j − h j .
2
We have:

1 1
(g )
Di L 1 j j = vti x g j + vx (g j )ti − Di (h j ),
2 2
1 1
Dg j (L 1i ) = (g j )x vti + vx (g j )ti − Dg j (h i ).
2 2
Upon using (22) and (19), and introducing the polynomial

(g ) δ1 h i δ1 h i δ1 h i
ai j j := g j + (g j )x + (g j )x x + ...
δvx δvx x δvx x x

obtained from ai j through the replacement of vt j by g j , we find:

(g )
1 1
Di L 1 j j − Dg j (L 1i ) = vti x g j − (g j )x vti − Di (h j ) + Dg j (h i )
2 2
366 Y.B. Suris and M. Vermeeren

1 1 (g )

= vti x g j − (g j )x vti − (a ji )x + (g j )x vti + ai j j − (gi )x g j

2 2 x
1 1 (g )

= vti x g j + (g j )x vti − a ji − ai j j − (gi )x g j

2 2 x
1 (g )

= (vti g j )x + ai j j − a ji − (bi j )x .
2 x

We denote the antiderivative with respect to x of this quantity by

(g )
L i(i)j := vti g j + ai j j − a ji − bi j .
2
The analogous calculation with coordinates x and t j yields

1

(g ) (g )
Dgi (L 1 j ) − D j L 1i i = − (vt j gi )x + ai j − a ji i + (b ji )x .
2 x

We denote its antiderivative by

( j) (g )
L i j := − vt j gi + ai j − a ji i + b ji .
2
Now we look for a differential polynomial L i j [v] depending on the partial deriv-
( j)
atives of v with respect to x, ti and t j that reduces to L i(i)j and to L i j after the substi-
tutions vt j = g j and vti = gi , respectively. It turns out that there is a one-parameter
family of such functions, given by
   
1 1 1
L i j = cvti vt j + (ai j − a ji ) + − c vti g j − + c vt j gi + (b ji − bi j ) + cgi g j
2 2 2

for c ∈ R. Checking this is a straightforward calculation using Eq. (20). Our theory
does not depend in any essential way on the choice of L i j within this family. For
aesthetic reasons we chose c = 0, which gives us Eq. (24).

Remark 5.3 We could  also take L to be the c-linear part of the form we have just
obtained, i.e. L = 1<i< j (vti − gi )(vt j − g j ) dti ∧ dt j . One can think of this as
choosing c = ∞. Such a two-form L can be considered for any family of evolu-
tionary equations vti = gi [v]. However, due to the vanishing components L 1i , this
form L has no relation to the classical variational formulation of the individual
differential equations vxti = (gi )x .

Eventually, Eq. (25) leads to the following closedness property.


Proposition 5.4 The two-form L = i< j L i j dti ∧ dt j , with coefficients given by
(23) and (24), is closed as soon as v solves all but one of the PKdV equations
vt2 = g2 , . . . , vt N = g N .
On the Lagrangian Structure of Integrable Hierarchies 367

Proof We use the notation


dL = Mi jk dti ∧ dt j ∧ dtk , Mi jk = Dk L i j − D j L ik + Di L jk (26)
i< j<k

We start by showing that M1 jk = Dk L 1 j − D j L 1k + Dx L jk vanishes as soon as

either vt j = g j or vtk = gk is satisfied. Indeed, we have:

M1 jk = Dk L 1 j − D j L 1k + Dx L jk
1 1 1 1
= vt j tk vx + vt j vxtk − Dk h j − vt j tk vx − vtk vxt j + D j h k
2 2 2 2
1 
+ vxt j gk + vt j Dx gk − vxtk g j − vtk Dx g j
2
1
+ Dk h j + vtk Dx g j − D j h k − vt j Dx gk − (gk Dx g j − g j Dx gk )
2
1
= vt vxt − vtk vxt j + vxt j gk − vt j Dx gk
2 j k

− vxtk g j + vtk Dx g j − gk Dx g j + g j Dx gk
1    1   
= vt j − g j Dx vtk − gk − vtk − gk Dx vt j − g j . (27)
2 2
For the case i, j, k > 1, we assume without loss of generality that vti = gi and vt j =
g j are satisfied. We do not assume that vtk = gk holds, and correspondingly we do
not make any identification involving vtk , vxtk , …. Using Eq. (27), we find:
 
Dx Mi jk = Dx Dk L i j − D j L ik + Di L jk
 
 
= Dk Di L 1 j − D j L 1i − D j Di L 1k − Dk L 1i + Di D j L 1k − Dk L 1 j
= 0.

Since these polynomials do not contain constant terms, it follows that

Dk L i j − D j L ik + Di L jk = 0.




Remark 5.5 Assuming that the statement of Theorem 5.2 holds true, one can easily
prove a somewhat weaker claim than Proposition 5.4, namely that the two-form L is
closed on simultaneous solutions of all the PKdV equations. Indeed, by Proposition
2.2, dL is constant on solutions of the multi-time Euler-Lagrange equations vti = gi .
Vanishing of this constant follows from the fact that dL = 0 on the trivial solution
v ≡ 0.
368 Y.B. Suris and M. Vermeeren

5.2 The Multi-time Euler-Lagrange Equations

Proof (of Theorem 5.2) We check all multi-time Euler-Lagrange Eqs. (5)–(7) indi-
vidually. If N > 3, we fix k > j > i > 1. If N = 3, we take j = 3, i = 2, and in the
following ignore all equations containing k. We use the convention L ji = −L i j , etc.

Equations (7)

• The equations
δ1i L 1i δi j L i j δ1 j L j1
+ + =0
δv I xti δv I ti t j δv I t j x

and
δi j L i j δ jk L jk δki L ki
+ + =0
δv I ti t j δv I t j tk δv I tk ti

are trivial because all terms vanish.

Equations (6)
• The equation
δ1i L 1i δi j L ji
=
δvx δvt j

yields

1 δ1i h i 1 δi j ai j
vti − = gi −
2 δvx 2 δvt j
 
1 δi j δ1 h i δ1 h i δ1 h i
= gi − vt j + vt j x + vt j x x + ...
2 δvt j δvx δvx x δvx x x
1 δ1 h i
= gi − .
2 δvx

This simplifies to the PKdV equation

vti = gi . (28)

• For α > 0, the equation

δ1i L 1i δi j L ji
=
δvx α+1 δvt j x α

yields
 
δ1i h i δi j δ1 h i δ1 h i δ1 h i
− =− vt j + vt j x + vt j x x + ...
δvx α+1 δvt j x α δvx δvx x δvx x x
On the Lagrangian Structure of Integrable Hierarchies 369

δ1 h i
=− ,
δvx α+1

which is trivial.
• Similarly, the equation
δ1 j L 1 j δi j L i j
=
δvx δvti

yields PKdV equation

vt j = g j , (29)

and for α > 0, the equation

δ1 j L 1 j δi j L i j
=
δvx α+1 δvti x α

is trivial.
• All equations of the form

δ1i L 1i δi j L ji δ1 j L 1 j δi j L i j
= (ti ∈
/ I) and = (t j ∈
/ I)
δvx I δvt j I δvx I δvti I

where I contains any tl (l > 1) are trivial because each term is zero.
• The equations
δ1i L 1i δ1 j L 1 j
= (x ∈
/ I)
δv I ti δv I t j

and
δi j L i j δik L ik
= (ti ∈
/ I)
δv I t j δv I tk

are easily seen to be trivial as well.

Equations (5)
δ1i L 1i
• By construction, the equations = 0 for i > 1 are the equations
δv
vxti = Dx gi . (30)

δ1i L 1i
For I containing any tl , l > 1, l = i, the equations = 0 are trivial.
δvt I
• The last family of equations we discuss as a lemma because its calculation is far
from trivial.
δi j L i j
Lemma 5.6 The equations = 0 are corollaries of the PKdV equations.
δvx α
370 Y.B. Suris and M. Vermeeren

Proof (of Lemma 5.6) From Eq. (24) we see that the variational derivative of L i j
contains only three nonzero terms,
   
δi j L i j ∂ Li j ∂ Li j ∂ Li j
= − Di − Dj . (31)
δvx α ∂vx α ∂vx α ti ∂vx α t j

δi j L i j
In particular, the equation = 0 yields Di g j − D j gi = 0, that is, the com-
δv
patibility condition of the flows vti = gi and vt j = g j . To determine the first term on
the right hand side of Eq. (31) for an arbitrary α > 0, we use an indirect method.
Assume that the dimension of multi-time N is at least 4 and fix k > 1 distinct from
i and j. Let v be a solution of all PKdV equations except vtk = gk . By Proposition
5.4 we have
 ∂ Li j
v I t = Dk L i j = D j L ik − Di L jk . (32)
I
∂v I k

∂ Li j
Since ∂v I
does not contain any derivatives with respect to tk , we can determine
∂ Li j
∂vx α
by looking at the terms in the right hand side of Eq. (32) containing vx α tk . These
are
 
1 δ1 h i δ1 h i
D j − gi vtk + vtk + vxtk + ...
2 δvx δvx x
 
1 δ1 h j δ1 h j
− Di − g j vtk + vtk + vxtk + ... .
2 δvx δvx x

Now we expand the brackets. By again throwing out all terms that do not contain
any vx α tk , and those that cancel modulo vti = gi or vt j = g j , we get
     
δ1 h i δ1 h i δ1 h i
− vtk D j + vxtk D j + vx xtk D j + ...
δvx δvx x δvx x x
     
δ1 h j δ1 h j δ1 h j
+ vtk Di − vxtk Di − vx xtk Di − ....
δvx δvx x δvx x x

Comparing this to Eq. (32), we find that

   
∂ Li j δ1 h j δ1 h i
= − Di + Dj .
∂vx α δvx α+1 δvx α+1

On the other hand we have

       
∂ Li j ∂ Li j δ1 h j δ1 h i
− Di − Dj = Di − Dj ,
∂vx α ti ∂vx α t j δvx α+1 δvx α+1
On the Lagrangian Structure of Integrable Hierarchies 371

δi j L i j
so Equation (31) implies that = 0 for any α.
δvx α
δ23 L 23
Since = 0 does not depend on the dimension N  3, the result for N  4
δvx α
implies the claim for N = 3.  

This concludes the proof of Theorem 5.2. 



6 Relation to Hamiltonian Formalism

In this last section, we briefly discuss the connection between the closedness of L
and the involutivity of the corresponding Hamiltonians.
In Proposition 2.2 we saw that dL is constant on solutions. For the one–
dimensional case (d = 1) with L depending on the first jet bundle only, it has
been shown in [20] that this is equivalent to the commutativity of the corresponding
Hamiltonian flows. If the constant is zero then the Hamiltonians are in involution.
Now we will prove a similar result for the two-dimensional case.
We will use a Poisson bracket on formal integrals, i.e. equivalence classes of 
functions modulo x-derivatives [12, Chap. 1–2]. In this section, the integral sign
will always denote an equivalence class, not an integration operator. The Poisson
bracket due to Gardner-Zakharov-Faddeev is defined by
  
   δ1 F δ1 G
F, G = Dx .
δu δu

Using integration by parts, we see that this bracket is anti-symmetric. Less obvious
is the fact that it satisfies the Jacobi identity [18, Chap. 7]. As we did when studying
the KdV hierarchy, we introduce a potential v that satisfies vx = u, and we identify
the space-coordinate x with the first coordinate t1 of multi-time. We can now re-write
the Poisson bracket as
   
   δ 1 F δ1 G δ1 F δ1 G
F, G = Dx =− , (33)
δvx δvx δv δvx

for functions F and G that depend on the x-derivatives of v but not on v itself.
Assume that the coefficients L 1 j of the Lagrangian two-from L are given by

1
L1 j = vx vt j − h j ,
2
where h j is a differential polynomial in vx , vx x , . . .. This is the case for the PKdV
hierarchy. The L 1 j are Lagrangians of the equations

vxt j = Dx g j or u t j = Dx g j ,
372 Y.B. Suris and M. Vermeeren

δ1 h j δ h 
where g j := δv x
, hence 1δv j = − Dx g j . It turns out that the formal integral h j
is the Hamilton functional for the equation u t j = Dx g j with respect to the Poisson
bracket (33). Formally:

     δ1 h j
h j , u(y) = h j , u δ(· − y) = − δ(x − y) = Dx g j (y),
δv

where δ denotes the Dirac delta.

Theorem 6.1 If dL = 0 on solutions, then the Hamiltonians are in involution,

  
h i , h j = 0.

Proof Recall notation (26). We have


  
M1 jk = Dx L jk − D j L 1k + Dk L 1 j

 
= − D j L 1k + Dk L 1 j
  
1 1 1 1
= − vxt j vtk − vx vtk t j + D j h k + vxtk vt j + vx vt j tk − Dk L 1 j
2 2 2 2
  
1 
= vxtk vt j − vxt j vtk − Dk L 1 j + D j h k
2

Using Eq. (21) (which, as opposed to Eq. (19), is independent of the form of h i and
gi ), the evolution equations vt j = g j , and integration by parts, we find that
  
 1 
M1 jk = vxtk vt j − vxt j vtk − Dx a jk + vtk Dx g j + Dx ak j − vt j Dx gk
2
  
1 
= − g j Dx gk − gk Dx g j − Dx a jk + Dx ak j
2

= gk Dx g j

δ1 h j δ1 h k
=−
δv δv
   x
= h j , hk .

Hence if dL = 0 on solutions of the evolution equations vt j = g j , then the Hamilton

functionals are in involution. 

On the Lagrangian Structure of Integrable Hierarchies 373

7 Conclusion

We have formulated the pluri-Lagrangian theory of integrable hierarchies, and pro-

pose it as a definition of integrability. The motivation for this definition comes from
the discrete case [10, 13, 20] and the fact that we have established a relation with the
Hamiltonian side of the theory. For the Hamiltonians to be in involution, we need the
additional fact that the Lagrangian two-form is closed. However, we believe that the
essential part of the theory is inherently contained in the pluri-Lagrangian formalism.
Since the KdV hierarchy is one of the most important examples of an integrable
hierarchy, our construction of a pluri-Lagrangian structure for the PKdV hierarchy
is an additional indication that the existence of a pluri-Lagrangian structure is a
reasonable definition of integrability.
It is remarkable that multi-time Euler-Lagrange equations are capable of produc-
ing evolutionary equations. This is a striking difference from the discrete case, where
the evolution equations (quad equations) imply the multi-time Euler–Lagrange equa-
tions (corner equations), but are themselves not variational [10].

Acknowledgments This research is supported by the Berlin Mathematical School and the DFG
Collaborative Research Center TRR 109 “Discretization in Geometry and Dynamics”.

A. A very short introduction to the variational bicomplex

Here we introduce the variational bicomplex and derive the basic results that we
use in the text. We follow Dickey, who provides a more complete discussion in
[12, Chap. 19]. Another good source on a (subtly different) variational bicomplex is
Anderson’s unfinished manuscript [2]. For ease of notation we restrict to real fields
u : R N → R, rather than vector-valued fields.
The space of ( p, q)-forms A ( p,q) consists of all formal sums

ω p,q = f δu I1 ∧ . . . ∧ δu I p ∧ dt j1 ∧ . . . ∧ dt jq ,

where f is a polynomial in u and partial derivatives of u of arbitrary order with

respect to any coordinates. The vertical one-forms δu I are dual to the vector fields
∂
∂u I
. The action of the derivative Di on ω p,q is

Di ω p,q = (Di f ) δu I1 ∧ . . . ∧ δu I p ∧ dt j1 ∧ . . . ∧ dt jq
+ f δu I1 i ∧ . . . ∧ δu I p ∧ dt j1 ∧ . . . ∧ dt jq
+ . . . + f δu I1 ∧ . . . ∧ δu I p i ∧ dt j1 ∧ . . . ∧ dt jq .

The integral of ω p,q over an q-dimensional manifold is the ( p, 0)-form defined by

374 Y.B. Suris and M. Vermeeren
   
ω p,q = f dt j1 ∧ . . . ∧ dt jq δu I1 ∧ . . . ∧ δu I p .

We call (0, q)-forms horizontal and ( p, 0)-forms vertical. The horizontal exterior
derivative d : A ( p,q) → A ( p,q+1) and the vertical exterior derivative δ : A ( p,q) →
A ( p+1,q) are defined by the anti-derivation property
 p ,q p ,q  p ,q p ,q p ,q p ,q
(a) d ω1 1 1 ∧ ω2 2 2 = dω1 1 1 ∧ ω2 2 2 + (−1) p1 +q1 ω1 1 1 ∧ dω2 2 2 ,
 p ,q p ,q  p ,q p ,q p ,q p ,q
δ ω1 1 1 ∧ ω2 2 2 = δω1 1 1 ∧ ω2 2 2 + (−1) p1 +q1 ω1 1 1 ∧ δω2 2 2 ,
and by the way they act on (0, 0)-, (1, 0)-, and (0, 1)-forms:
 ∂f  ∂f   ∂f
(b) d f = D j f dt j = + u I j dt j , δ f = δu I ,
∂t j ∂u I ∂u I
j
 j I I

(c) d(δu I ) = − δu I j ∧ dt j , δ(δu I ) = 0,

j

(d) d(dt j ) = 0, δ(dt j ) = 0, δ(du I ) = −d(δu I ) = δu I j ∧ dt j .
j

Properties (a)–(d) determine the action of d and δ on any form. The corresponding
mapping diagram is known as the variational bicomplex.

.. .. .. ..
. . . .
↑δ ↑δ ↑δ ↑δ
d d d d
A (1,0) −
→ A (1,1) − → A (1,n−1) −
→ ... − → A (1,n)
↑δ ↑δ ↑δ ↑δ
d d d d
A (0,0) −
→ A (0,1) − → A (0,n−1) −
→ ... − → A (0,n)

The following claims follow immediately from the definitions.

Proposition A.1 We have d2 = δ 2 = 0 and dδ + δd = 0.
k
Remark A.2 This implies that d + δ : A k → A k+1 , where A k := i=0 A (i,k−i) , is
an exterior derivative as well.
Proposition A.3 We have Di δ = δ Di .
Proposition A.4 For a differential
polynomial h, define the corresponding vertical
generalized vector field by ∂h := I h I ∂u∂ I . We have d ι∂h + ι∂h d = 0.
Proof It suffices to show this for (0,0)-forms (polynomials f in u and partial deriva-
tives of u), for (0,1)-forms dt j , and for (1,0)-forms δu I . For (0,0)-forms, both terms
of the claimed identity are zero:

 
d ι∂h f = 0, ι∂h (d f ) = ι∂h D j f dt j = 0.
j
On the Lagrangian Structure of Integrable Hierarchies 375

Likewise for (0,1)-forms:

d ι∂h dt j = 0, ι∂h (ddt j ) = 0.

For (1,0)-forms we find:

   

ι∂h (dδu I ) = ι∂h − δu I j ∧ dt j =− h I j dt j = −dh I = −d ι∂h δu I .

j j




B. Proof of Lemma 2.4

Assume that the action is stationary on all d-dimensional stepped surfaces in R N . Let
S be a smooth d-dimensional surface in R N . Partition the space R N into hypercubes
Ci of edge length ε. We can choose this partitioning in such a way that the surface S
does not contain the center of any of the hypercubes. Denote SiN := S ∩ Ci .
We give each hypercube its own coordinate system [−1, 1] N → Ci and identify
the hypercube with its coordinates. In each punctured hypercube [−1, 1] N \ {0} we
define a family of balloon maps
⎧
⎨ αx if xmax < α
BαN : [−1, 1] N \ {0} → [−1, 1] N \ {0} : x → xmax
⎩x if xmax  α

for α ∈ [0, 1]. Here, xmax := max(|x1 |, . . . |x N |) denotes the maximum norm with
respect to the local coordinates. The idea is that from the center of each hypercube,
we inflate a square balloon which pushes the curve away from the center, until it lies
on the boundary of the hypercube.
Indeed, the deformed surface SiN −1 := B1N (SiN ) = B1N (S ∩ Ci ) lies on the bound-
ary of the hypercube, i.e. within the (N − 1)-faces of the hypercube. We want it to lie
within the d-faces of the hypercube, which would imply that it is a stepped surface.
To achieve this, we introduce a balloon map
⎧
⎨ αx if xmax < α
BαN −1, j : [−1, 1] N −1
\ {0} → [−1, 1] N −1
\ {0} : x → xmax
⎩x if xmax  α

j
in each of the (N − 1)-faces Ci of the hypercube Ci , which pushes the surface into
the (N − 2)-faces. We denote the surface we obtain this way by SiN −2 . If the surface
happens to contain the center of a (N − 1)-face, we can slightly perturb the surface
k, j
without affecting the argument. By iterating this procedure, using balloon maps Bα
376 Y.B. Suris and M. Vermeeren

j
in each k-face Ci (N  k  d + 1), we obtain a surface Sid that lies in the d-faces
(Figs. 5 and 6).
Consider the (d + 1)-dimensional surface


N   j
Mi := Bαk, j (Sik ∩ Ci )
k=d+1 j: C j is a α∈[0,1]
i
k -face of Ci

that is swept out by the consecutive application of the balloon maps to SiN := S ∩ Ci .
Assuming that ε is small  compared to the curvature of S, the (d + 1)-dimensional
k, j j
volume of each of the α∈[0,1] Bα (Sik ∩ Ci ) is of the order εd+1 . The number of
such volumes making up Mi only depends on the dimensions N and d, not on ε, so
the (d + 1)-dimensional volume |Mi | of Mi is of the order |Mi | = O(εd+1 ).
Now consider a variation  V with compact support and restrict the surface S to
this support. Denote by  S := i Sid the stepped surface obtained  from S by repeated
application of balloon maps in all the hypercubes, and by M := i Mi the (d + 1)-
dimensional surface swept out by these balloon maps. The bounary of M consists
of S, S, and a small strip of area O(ε) connecting the boundaries of S and  S (the
dotted line in Fig. 5). The number of hypercubes intersecting S is of order ε−d , so
|M| = O(ε−d )O(εd+1 ) = O(ε). It follows that
    
   
 ιpr V δL − ιpr V δL  =  ιpr V δL  + O(ε)
   
S S
∂ M 
 

=  d(ιpr V δL ) + O(ε) → 0
M

as ε → 0. By assumption, S ιpr V
 δL = 0 for all ε, so the action on S will be
stationary as well. 


Fig. 5 Balloon maps in nine adjacent squares deforming a curve in R2 . From left to right: α = 0.2,
α = 0.7 and α = 1
On the Lagrangian Structure of Integrable Hierarchies 377

Fig. 6 The second and last iteration for a curve in R3 . From left to right: α = 0.1, α = 0.6 and
α=1

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Adler, V., Bobenko, A., Suris, Y.: Classification of integrable equations on quad-graphs. the
consistency approach. Commun. Math. Phys. 233, 513–543 (2003)
2. Anderson, I.: The variational bicomplex. Preprint (1989)
3. Baxter, R.: Solvable eight-vertex model on an arbitrary planar lattice. Philos. Trans. R. Soc.
Lond. 289, 315–346 (1978)
4. Baxter, R.: Free-fermion, checkerboard and Z-invariant lattice models in statistical mechanics.
Proc. R. Soc. Lond. A 404, 1–33 (1986)
5. Bazhanov, V., Mangazeev, V., Sergeev, S.: A master solution of the quantum Yang-Baxter
equation and classical discrete integrable equations. Adv. Theor. Math. Phys. 16, 65–95 (2012)
6. Bazhanov, V., Mangazeev, V., Sergeev, S.: Faddeev-Volkov solution of the Yang-Baxter equa-
tion and discrete conformal symmetry. Nucl. Phys. B 784, 234–258 (2007)
7. Bobenko, A., Mercat, C., Suris, Y.: Linear and nonlinear theories of discrete analytic functions.
Intergable structure and isomonodromic Green’s function. J. Reine Angew. Math. 583, 117–161
(2005)
8. Bobenko, A., Suris., Y.: Integrable systems on quad-graphs. Intern. Math. Res. Not. 2002(11),
573–611 (2002)
9. Bobenko, A., Suris, Y.: Discrete pluriharmonic functions as solutions of linear pluri-Lagrangian
systems. Commun. Math. Phys. 336, 199–215 (2015)
10. Boll, R., Petrera, M., Suris, Y.: What is integrability of discrete variational systems? Proc. R.
Soc. A 470, 20130, 550 (2014)
378 Y.B. Suris and M. Vermeeren

11. Burstall, F., Ferus, D., Pedit, F., Pinkall, U.: Harmonic tori in symmetric spaces and commuting
Hamiltonian systems on loop algebras. Ann. Math. 138, 173–212 (1993)
12. Dickey, L.: Soliton equations and Hamiltonian systems, 2nd edn. World Scientific (2003)
13. Lobb, S., Nijhoff., F.: Lagrangian multiforms and multidimensional consistency. J. Phys. A:
Math. Theor. 42, 454,013 (2009)
14. Newell, A.: Solitons in mathematics and physics. SIAM (1985)
15. Nijhoff, F.: Lax pair for the Adler (lattice Krichever-Novikov) system. Phys. Lett. A 297, 49–58
(2002)
16. Noether, E.: Invariante Variationsprobleme. Nachrichten von der Gesellschaft der Wis-
senschaften zu Göttingen, Math.-Phys. Kl. 235–257 (1918)
17. Ohnita, Y., Valli, G.: Pluriharmonic maps into compact Lie groups and factorization into uni-
tons. Proc. Lond. Math. Soc. 61, 546–570 (1990)
18. Olver, P.: Applications of Lie groups to differential equations, 2nd edn. Springer (1993)
19. Rudin, W.: Function theory in polydiscs. Benjamin (1969)
20. Suris, Y.: Variational formulation of commuting Hamiltonian flows: multi-time Lagrangian
1-forms. J. Geom. Mech. 5, 365–379 (2013)
21. Suris, Y.: Variational symmetries and pluri-Lagrangian systems. In: Dynamical Systems, Num-
ber Theory and Applications. A Festschrift in Honor of Armin Leutbecher’s 80th Birthday,
Hagen, Th., Rupp, F., Scheurle, J. (eds.) World Scientific, pp. 255–266 (2016)
On the Variational Interpretation
of the Discrete KP Equation

Raphael Boll, Matteo Petrera and Yuri B. Suris

Abstract We study the variational structure of the discrete Kadomtsev-Petviashvili

(dKP) equation by means of its pluri-Lagrangian formulation. We consider the dKP
equation and its variational formulation on the cubic lattice Z N as well as on the root
lattice Q(A N ). We prove that, on a lattice of dimension at least four, the corresponding
Euler-Lagrange equations are equivalent to the dKP equation.

1 Introduction

We developed the theory of pluri-Lagrangian problems (integrable systems of varia-

tional origin) in recent papers [2–6, 15, 16], influenced by the fundamental insight of
[11–13, 17]. In the present paper, we consider the pluri-Lagrangian formulation of the
discrete bilinear Kadomtsev-Petviashvili (dKP) equation on three-dimensional lat-
tices and its consistent extension to higher dimensional lattices. This equation belongs
to integrable octahedron-type equations which were classified in [1]. A Lagrangian
formulation of this equation was given in [13]. There, the authors consider a discrete
3-form on the lattice Z3 together with the corresponding Euler-Lagrange equations
which are shown to be satisfied on solutions of the dKP equation. They also show
that this 3-form is closed on solutions of the dKP equation, namely, the so-called
4D closure relation is satisfied. The main goal of the present paper is to provide a
more precise understanding of the findings in that paper. More concretely:
• In the framework of the pluri-Lagrangian formulation, we construct the elementary
building blocks of Euler-Lagrange equations, which, in the present situation, are
the so-called 4D corner equations.

R. Boll · M. Petrera · Y.B. Suris (B)

Inst. für Mathematik, Technische Universität Berlin,
Straße des 17. Juni 136, 10623 Berlin, Germany
e-mail: [email protected]
R. Boll
e-mail: [email protected]
M. Petrera
e-mail: [email protected]
© The Author(s) 2016 379
A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_12
380 R. Boll et al.

• In the two-dimensional case, as noticed in [4], the corresponding 3D corner equa-

tions build a consistent system. Its solutions are more general then the solutions
of the underlying hyperbolic system of quad-equations. On the contrary, in the
present three-dimensional situation, the system of 4D corner equations is not con-
sistent in the usual sense (i.e., it does not allow to determine general solutions
with the maximal number of initial data). However, this system turns out to be
equivalent, in a sense which we are going to explain later, to the corresponding
hyperbolic system, namely the dKP equation.
• We provide a rigorous consideration of the branches of the logarithm functions
involved in the Euler-Lagrange equations. This leads to the following more precise
result: the system of 4D corner equations is equivalent, and thus provides a varia-
tional formulation, to two different hyperbolic equations, namely the dKP equation
itself and its version obtained under inversion x → x −1 of all fields which will be
denoted by dKP− .
One can consider the dKP equation on the cubic lattice Z3 and its higher dimen-
sional analogues Z N , but, as discussed in [1, 8, 9] another natural setting the dKP
equation (and related octahedron-type equations) is the three-dimensional root lattice

Q(A3 ) := {(n i , n j , n k , n  ) : n i + n j + n k + n  = 0}.

Also in this setting, the dKP equation can be extended in a consistent way to the
higher dimensional lattices Q(A N ) with N > 3.
Both lattices have their advantages and disadvantages. The cubic lattice Z N , on the
one hand, is more manageable and easier to visualize. Its cell structure is very simple:
for every dimension N , all N -dimensional elementary cells are N -dimensional cubes.
On the other hand, it is less natural to consider dKP on the lattice Z3 , because this
equation depends on the variables assigned to six out of eight vertices of a (three-
dimensional) cube.
The root lattice Q(A N ), in contrast, has a more complicated cell structure, because
the number of different N -dimensional elementary cells increases with the dimen-
sion N . For instance, for N = 3 there are two types of elementary cells octahedra
and tetrahedra. Moreover, especially in higher dimensions, a visualization of the ele-
mentary cells is difficult, if not impossible. However, this lattice is more natural for
the consideration of dKP from the combinatorial point of view, because this equation
depends on variables which can be assigned to the six vertices of an octahedron, one
of the elementary cells of the lattice. Furthermore, the four-dimensional elementary
cells are combinatorially smaller (they contain only 10 vertices, as compared with
16 vertices of a four-dimensional cube) and possess higher symmetry than the cubic
ones. Since they support the equations which serve as variational analogue of the
dKP equation, this leads to a simpler situation.
We will see that a four-dimensional cube is combinatorially equivalent to the sum
of four elementary cells of the root lattice Q(A4 ). Therefore, several results in the
cubic case can be seen as direct consequences of results of the more fundamental
Q(A N )-case.
On the Variational Interpretation of the Discrete KP Equation 381

Let us start with some concrete definitions valid for an arbitrary N -dimensional
lattice X .
Definition 1.1 (Discrete 3-form) A discrete 3-form on X is a real-valued function
L of oriented 3-cells σ depending on some field x : X → R, such that L changes
the sign by changing the orientation of σ .
For instance, in Q(A N ), the 3-cells are tetrahedra and octahedra, and, in Z N , the
3-cells are 3D cubes.
Definition 1.2 (3-dimensional pluri-Lagrangian problem) Let L be a discrete 3-
form on X depending on x : X → R.
• To an arbitrary 3-manifold Σ ⊂ X , i.e., a union of oriented 3-cells which forms
an oriented three-dimensional topological manifold, there corresponds the action
functional, which assigns to x|V (Σ) , i.e., to the fields in the set of the vertices V (Σ)
of Σ, the number

SΣ := L (σ ).
σ ∈Σ

• We say that the field x : V (Σ) → R is a critical point of SΣ , if at any interior

point n ∈ V (Σ), we have
∂ SΣ
= 0. (1)
∂ x(n)

Equation (1) are called discrete Euler-Lagrange equations for the action SΣ .
• We say that the field x : X → R solves the pluri-Lagrangian problem for the
Lagrangian 3-form L if, for any 3-manifold Σ ⊂ X , the restriction x|V (Σ) is a
critical point of the corresponding action SΣ .
In the present paper, we focus on the variational formulation of the dKP equation on
Q(A N ) and Z N . Let us formulate the main results of the paper.
On the lattice Q(A N ), we consider discrete 3-forms vanishing on all tetrahe-
dra. One can show (see Corollary 2.5) that, for an arbitrary interior vertex of any
3-manifold in Q(A N ), the Euler-Lagrange equations follow from certain elemen-
tary building blocks. These so-called 4D corner equations are the Euler-Lagrange
equations for elementary 4-cells of Q(A N ) different from 4-simplices, so-called
4-ambo-simplices. Such a 4-ambo-simplex has ten vertices. Therefore, the crucial
issue is the study of the system consisting of the corresponding ten corner equations.
In our case, each corner equation depends on all ten fields at the vertices of the 4-
ambo-simplex. Therefore, one could call this system consistent if any two equations
are functionally dependent. It turns out that this is not the case. We will prove the
following statement:
Theorem 1.3 Every solution of the system of ten corner equations for a 4-ambo-
simplex in Q(A N ) satisfies either the system of five dKP equations or the system of
five dKP− equations on the five octahedral facets of the 4-ambo-simplex.
382 R. Boll et al.

Thus, one can prescribe arbitrary initial values at seven vertices of a 4-ambo-simplex.
We will also prove the following theorem:
Theorem 1.4 The discrete 3-form L is closed on any solution of the system of
corner equations.
In [4, 15], it was shown that in dimensions 1 and 2 the analogues of the property
formulated in Theorem 1.4 are related to more traditional integrability attributes.
For the case of the cubic lattice Z N , the situation is similar: one can show (see
Corollary 4.2) that, for an arbitrary interior vertex of any 3-manifold in Z3 , the Euler-
Lagrange equations follow from certain elementary building blocks. These so-called
4D corner equations are the Euler-Lagrange equations for elementary 4D cubes in
Z N . A 4D cube has sixteen vertices, but in our case the action on a 4D cube turns out
to be independent of the fields on two of the vertices. Therefore, the crucial issue is the
study of the system consisting of the corresponding fourteen corner equations. Six of
the fourteen corner equations depend each on thirteen of the fourteen fields. There do
not exist pairs of such equations which are independent of one and the same field. All
other equations depend each on ten of the fourteen fields. Therefore, one could call
this system consistent if it would have the minimal possible rank 2 (assign twelve
fields arbitrarily and use two of the six corner equations—depending on thirteen
fields—to determine the remaining two fields, then all twelve remaining equations
should be satisfied automatically). It turns out that the system of the fourteen corner
equations is not consistent in this sense. We will prove the following analogue of
Theorem 1.3:
Theorem 1.5 Every solution of the system of fourteen corner equations for a
4D cube in Z N satisfies either the system of eight dKP equations or the system
of eight dKP− equations on the eight cubic facets of the 4D cube.
Thus, one can prescribe arbitrary initial values at nine vertices of a 4D cube. Corre-
spondingly, we will also prove the following statement:
Theorem 1.6 The discrete 3-form L is closed on any solution of the system of
corner equations.
The paper is organized as follows: we start with the root lattice Q(A N ), thus con-
sidering the combinatorial issues and some general properties of pluri-Lagrangian
systems. Then we introduce the dKP equation and its pluri-Lagrangian structure. In
the second part of the paper the present similar considerations for the cubic lattice Z N .

2 The Root Lattice Q( A N )

We consider the root lattice

Q(A N ) := {n := (n 0 , n 1 , . . . , n N ) ∈ Z N +1 : n 0 + n 1 + . . . + n N = 0},
On the Variational Interpretation of the Discrete KP Equation 383

where N ≥ 3. The three-dimensional sub-lattices Q(A3 ) are given by

Q(A3 ) := {(n i , n j , n k , n  ) : n i + n j + n k + n  = const}.

We consider fields x : Q(A N ) → R, and use the shorthand notations

xı̄ = x(n − ei ), x = x(n), and xi = x(n + ei ),

where ei is the unit vector in the ith coordinate direction. Furthermore, the shift
functions Ti and Tı̄ are defined by

Ti xα := xiα and Tı̄ xα := xı̄α

for a multiindex α. For simplicity, we sometimes abuse notations by identifying

lattice points n with the corresponding fields x(n).
We now give a very brief introduction to the Delaunay cell structure of the n-
dimensional root lattice Q(A N ) [7, 14]. Here, we restrict ourselves to a very elemen-
tary description which is appropriate to our purposes and follow the considerations
in [1]. For each N there are N sorts of N -cells of Q(A N ) denoted by P(k, N ) with
k = 1, . . . , N :
• Two sorts of 2-cells:
P(1, 2) : black trianglesi jk := {xi , x j , xk };
P(2, 2) : white triangles i jk
:= {xi j , xik , x jk };
• Three sorts of 3-cells:
P(1, 3) : black tetrahedrai jk := {xi , x j , xk , x };
P(2, 3) : octahedra[i jk] := {xi j , xik , xi , x jk , x j , xk };
P(3, 3) : white tetrahedra i jk
:= {xi jk , xi j , xik , x jk };
• Four sorts of 4-cells:
P(1, 4) : black 4-simplicesi jkm := {xi , x j , xk , x , xm };
P(2, 4) : black 4-ambo-simplicesi jkm := {xαβ : α, β ∈ {i, j, k, , m}, α = β};
P(3, 4) : white 4-ambo-simplices i jkm
:= {xαβγ : α, β, γ ∈ {i, j, k, , m},
α = β = γ = α};
P(4, 4) : white 4-simplices i jkm

:= {xi jk , xi jkm , xi jm , xikm , x jkm }.

The facets of 3-cells and 4-cells can be found in Appendix 1.
In the present paper we will consider objects on oriented manifolds. We say that
a black triangle i jk and white triangle i jk
are positively oriented if i < j < k
(see Fig. 1). Any permutation of two indices changes the orientation to the opposite
one.
When we use the bracket notation, we always write the letters in brackets in
increasing order, so, e.g., in writing i jk we assume that i < j < k and avoid the
notation  jik or ik j for the negatively oriented triangle −i jk.
There is a simple recipe to derive the orientation of facets of an N -cell: On every
index in the brackets we put alternately a “+” or a “−” starting with a “+” on the
384 R. Boll et al.

Fig. 1 Orientation of (a) xk (b) x jk xik

triangles: a the black triangle
i jk; b the white triangle
i jk

xi xj xij

last index. Then we get each of its facets by deleting one index and putting the
corresponding sign in front of the bracket. For instance, the black 4-ambo-simplex

+−+−+
 i j k  m

has the five octahedral facets [i jk], −[i jkm], [i jm], −[ikm], and [ jkm].
The following two definitions are valid for arbitrary N -dimensional lattices X .
Definition 2.1 (Adjacent N -cell) Given an N -cell σ , another N -cell σ̄ is called
adjacent to σ if σ and σ̄ share a common (N − 1)-cell. The orientation of this
(N − 1)-cell in σ must be opposite to its orientation in σ̄ .
The latter property guarantees that the orientations of the adjacent N -cells agree.
Definition 2.2 (Flower) A 3-manifold in X with exactly one interior vertex x is
called a flower with center x. The flower at an interior vertex x of a given 3-manifold
is the flower with center x which lies completely in the 3-manifold.
As a consequence, in Q(A N ), in each flower every tetrahedron has exactly three
adjacent 3-cells and every octahedron has exactly four adjacent 3-cells.
Examples for open 3-manifolds in Q(A N ) are the three-dimensional sub-lattices
Q(A3 ). Here, the flower at an interior vertex consists of eight tetrahedra (four black
and four white ones) and six octahedra.
Examples of closed 3-manifolds in Q(A N ) are the set of facets of a 4-ambo-
simplex (consisting of five tetrahedra) and the set of facets of a 4-ambo-simplex
(consisting of five tetrahedra and five octahedra).
The elementary building blocks of 3-manifolds are so-called 4D corners:
Definition 2.3 (4D corner) A 4D corner with center x is a 3-manifold consisting
of all facets of a 4-cell adjacent to x.
In Q(A N ), there are two different types of 4D corners: a corner on a 4-simplex
(consisting of a four tetrahedra) and a corner on a 4-ambo-simplex (consisting of
two tetrahedra and three octahedra), see Appendix 2 for details.
The following combinatorial statement will be proven in Appendix 3:
Theorem 2.4 The flower at any interior vertex of any 3-manifold in Q(A N ) can be
represented as a sum of 4D corners in Q(A N +2 ).
On the Variational Interpretation of the Discrete KP Equation 385

Let L be a discrete 3-form on Q(A N ). The exterior derivative dL is a discrete

4-form whose value at any 4-cell in Q(A N ) is the action functional of L on the 3-
manifold consisting of the facets of the 4-cell. For our purposes, we consider discrete
3-forms L vanishing on all tetrahedra. In particular, we have

dL (i jkm) ≡ 0 and dL ( i jkm

) ≡ 0

since a 4-simplices only contain tetrahedra. The exterior derivative on a black 4-

ambo-simplex i jkm is given by

S i jkm := dL (i jkm)

(2)
= L ([i jk]) + L (−[i jkm]) + L ([i jm]) + L (−[ikm]) + L ([ jkm]).

The exterior derivative on a white 4-ambo-simplex i jkm

is given by

S̄ i jkm := dL ( i jkm
)
= L (Tm [i jk]) + L (−T [i jkm]) + L (Tk [i jm]) + L (−T j [ikm])
+ L (Ti [ jkm]).
(3)
Accordingly, the Euler-Lagrange equations on black 4-ambo-simplices i jkm
are

∂ S i jkm ∂ S i jkm ∂ S i jkm ∂ S i jkm ∂ S i jkm

= 0, = 0, = 0, = 0, = 0,
∂ xi j ∂ xik ∂ xi ∂ xim ∂ x jk
∂ S i jkm ∂ S i jkm ∂ S i jkm ∂ S i jkm
∂ S i jkm
= 0, = 0, = 0, = 0,
= 0.
∂ x j ∂ x jm ∂ xk ∂ xkm
∂ xm
(4)
and the Euler-Lagrange equations on white 4-ambo-simplices i jkm
are

∂ S̄ i jkm ∂ S̄ i jkm ∂ S̄ i jkm ∂ S̄ i jkm ∂ S̄ i jkm

= 0, = 0, = 0, = 0, = 0,
∂ xi jk ∂ xi j ∂ xi jm ∂ xik ∂ xikm
∂ S̄ i jkm ∂ S̄ i jkm ∂ S̄ i jkm ∂ S̄ i jkm ∂ S̄ i jkm
= 0, = 0, = 0, = 0, = 0.
∂ xim ∂ x jk ∂ x jkm ∂ x jm ∂ xkm
(5)
The last two systems are called corner equations.
The following statement is an immediate consequence of Theorem 2.4:
Theorem 2.5 For discrete every 3-form on Q(A N ) and every 3-manifold in Q(A N )
all corresponding Euler-Lagrange equations can be written as a sum of corner
equations.
386 R. Boll et al.

3 The dKP Equation on Q( A N )

We will now introduce the dKP equation on the root lattice Q(A3 ). Every oriented
octahedron [i jk] (i < j < k < ) in Q(A3 ) supports the equation

xi j xk − xik x j + xi x jk = 0. (6)

We can extend this system in a consistent way (see [1]) to the four-dimensional
root lattice Q(A4 ) and higher-dimensional analogues, such that the five octahedral
facets [i jk], [ jkm], −[ikm], [i jm], and −[i jkm] of the black 4-ambo-simplex
i jkm support the equations

xi j xk − xik x j + xi x jk = 0,

x jk xm − x j xkm + x jm xk = 0,
xk xim − xkm xi + xik xm = 0, (7)
xm xi j − xi x jm + x j xim = 0,
xim x jk − x jm xik + xkm xi j = 0

and the five octahedral facets Tm [i jk], Ti [ jkm], −T j [ikm], Tk [i jm], and
−T [i jkm] of the white 4-ambo-simplex i jkm
support the equations

xi jm xkm − xikm x jm + xim x jkm = 0,

xi jk xim − xi j xikm + xi jm xik = 0,
x jk xi jm − x jkm xi j + xi jk x jm = 0, (8)
xkm xi jk − xik x jkm + x jk xikm = 0,
xim x jk − x jm xik + xkm xi j = 0.

In both systems one can derive one equation from another by cyclic permutations of
indices (i jkm).
We propose the following discrete 3-form L defined on oriented octahedra [i jk]:
      
1 xi j xk xik x j xi x jk
L ([i jk]) := Λ +Λ +Λ − , (9)
2 xik x j xi x jk xi j xk

where
   z
1 log |1 − x|
Λ(z) := λ(z) − λ and λ(z) := − d x. (10)
z 0 x

The discrete 3-form (9) has its motivation in [13]. Indeed, in [13], the authors consider
a similar discrete 3-form on the cubic lattice Z N . One can also consider our 3-form
on the cubic lattice Z N . Then one would assign to each 3D cube the 3-form at its
On the Variational Interpretation of the Discrete KP Equation 387

inscribed octahedron. This 3-form differs from their one by an additive constant and
a slightly different definition of the function λ(z): they use the function
 z
log(1 − x)
Li2 (z) := − dx (11)
0 x

instead of λ(z). Our choice of λ(z) allows us for a more precise consideration of the
branches of the occurring logarithm.
Observe that the expression (9) only changes its sign under the cyclic permutation
of indices (i jkm). This follows from Λ(z) = −Λ(z −1 ). As a consequence, the
exterior derivatives S i jkm and S̄ i jkm defined in (2) and (3), respectively, are invariant
under the cyclic permutation of indices (i jkm). Therefore, one can obtain all corner
equations in (4) and (5) by (iterated) cyclic permutation (i jkm) from

∂ S i jkm ∂ S i jkm ∂ S̄ i jkm ∂ S̄ i jkm

= 0, = 0, and = 0, = 0.
∂ xi j ∂ xik ∂ xi jk ∂ xi j

Let us study separately the corner equations on black and white 4-ambo-simplices.
The corner equations which live on the black 4-ambo-simplex i jkm are given by

∂ S i jkm ∂L ([i jk]) ∂L (−[i jkm]) ∂L ([i jm])

= + + =0
∂ xi j ∂ xi j ∂ xi j ∂ xi j

and

∂ S i jkm ∂L ([i jk]) ∂L (−[i jkm]) ∂L (−[ikm])

= + + = 0.
∂ xik ∂ xik ∂ xik ∂ xik

Explicitly, they read

1 1
log |E i j | = 0 and log |E ik | = 0, (12)
xi j xik

where
xi j xk + xi x jk xi j xkm − xik x jm xi j xm + xim x j
E i j := · ·
xi j xk − xik x j xi j xkm + xim x jk xi j xm − xi x jm

and
xik x j − xi j xk xik x jm − xim x jk xik xm − xi xkm
E ik := · · .
xik x j − xi x jk xik x jm − xi j xkm xik xm + xim xk

For every corner equation (12) there are two classes of solutions, because any solution
can either solve E i j = −1 or E i j = 1. Hereafter, we only consider solutions, where
all fields xi j are non-zero (we call such solutions non-singular).
388 R. Boll et al.

Theorem 3.1 Every solution of the system (4) solves either the system

E i j = −1, E ik = −1, E i = −1, E im = −1, E jk = −1,

(13)
E j = −1, E jm = −1, E k = −1, E km = −1, E m = −1

or the system

E i j = 1, E ik = 1, E i = 1, E im = 1, E jk = 1,
(14)
E j = 1, E jm = 1, E k = 1, E km = 1, E m = 1.

Furthermore, the system (13) is equivalent to the system (7) (that is dKP on the
corresponding black 4-ambo-simplex). The system (14) is equivalent to the system

xik xi x jk x j − xi j xi x jk xk + xi j xik x j xk = 0,

x j x jm xk xkm − x jk x jm xk xm + x jk x j xkm xm = 0,
xkm xik xm xi − xk xik xm xim + xk xkm xi xim = 0, (15)
xi x j xim x jm − xm x j xim xi j + xm xi x jm xi j = 0,
x jm xkm xi j xik − xim xkm xi j x jk + xim x jm xik x jk = 0,

which is the system (7) after the transformation x → x −1 of fields (that is dKP− on
the corresponding black 4-ambo-simplex).

Proof Consider a solution x of (4) that solves E i j = −1 and E jk = −1. We set

ai j := xm xi j − xi x jm + x j xim , (16)

aik := xk xim − xkm xi + xik xm , (17)

and

a jk := x jk xm − x j xkm + x jm xk , (18)

and use these equations to substitute xi j , xik and x jk in E i j = −1 and E jk = −1.

Writing down the result in polynomial form, we get
2
xm (ai j + xi x jm − xim x j )ei j = 0

and
2
xm (a jk + x j xkm − x jm xk )e jk = 0,

where ei j and e jk are certain polynomials. Since for every solutions of (4) all fields
are non-zero this leads us to ei j = 0 and e jk = 0. Computing the difference of the
latter two equations we get
On the Variational Interpretation of the Discrete KP Equation 389

ai j xk xkm (ai j + xi x jm − xim x j ) − a jk xi xim (a jk + x j xkm − x jm xk ) = 0

and, with the use of (16) and (18),

xm (ai j xi j xk xkm − a jk x jk xi xim ) = 0,

which depends on seven independent fields, i.e., no subset of six fields belong to one
octahedron. Then comparing coefficients leads to ai j = a jk = 0. Substituting

xi x jm − xim x j x j xkm − x jm xk

xi j = and x jk =
xm xm

into E i j = −1 and solving the resulting equation with respect to xik , we get

xi xkm − xim xk

xik = .
xm

Substituting xi j , xik and x jk in E ik by using the last three equations, we get E ik = −1.
Analogously, one can prove that, for a solution x of (4) which solves E i j = −1
and E ik = −1, we have E jk = −1, and for a solution x of (4) which solves E ik = −1
and E i = −1, we have E k = −1. Therefore, for every solution x of (4) and for
every white triangle {xα , xβ , xγ } on the black 4-ambo-simplex i jkm we proved
the following: if E α = −1 and E β = −1 then E γ = −1, too.
On the other hand, one can easily see that x solves E i j = 1 or E jk = 1 if and only
if x −1 solves E i j = −1 or E ik = −1, respectively. Therefore, we also know that, if
E α = 1 and E β = 1 then E γ = 1, too.
Summarizing, we proved that every solution x of (4) solves either (13) and then
also (7) or (14) and then also (15).
Consider a non-singular solution x of the system (7). Then

xi j xk + xi x jk xi j xkm − xik x jm xi j xm + xim x j

Ei j = · ·
xi j xk − xik x j xi j xkm + xim x jk xi j xm − xi x jm
xik x j −xim x jk xi x jm
= · · = −1
−xi x jk xik x jm −xim x j

and
xik x j − xi j xk xik x jm − xim x jk xik xm − xi xkm
E ik = · ·
xik x j − xi x jk xik x jm − xi j xkm xik xm + xim xk
xi x jk xi j xkm xim xk
= · · = −1.
xi j xk xim x jk −xi xkm

This proves the equivalence of (13) and (7) and also the equivalence of (14)
and (15) since x solves E i j = −1 or (7) if and only if x −1 solves E i j = 1 or (15),
respectively. 
390 R. Boll et al.

We will present the closure relation which can be seen as a criterion of integrability:
Theorem 3.2 (Closure relation) There holds:

π2
S i jkm ± =0
4
on all solutions of (13) and (14), respectively. Therefore, one can redefine the 3-form
L as
π2
L˜ ([i jk]) := L ([i jk]) ±
4

in order to get S i jkm = 0 on all solutions of (13) and (14), respectively.

Proof The set of solutions S + of (13), as well as the set of solutions S − (14),
is a connected seven-dimensional algebraic manifold which can be parametrized
by the set of variables {xi j , xik , xi , xim , x jk , x j , x jm }. We want to show that the
directional derivatives of S i jkm along tangent vectors of S ± vanish. It is easy to see
that the stronger property gradS i jkm = 0 on S ± , where we S i jkm is considered as a
function of ten variables xi j , is a consequence of (13), respectively (14). Therefore,
the function S i jkm is constant on S ± .
To determine the value of S i jkm on solutions of (13), we consider the constant
solution of (7)

xi j = x jk = xk = xm = xim = a,

(19)
xik = x j = xkm = xi = x jm = −1,

where
√
1 5
a := − .
2 2

(Indeed, for this point every equation from (7) looks like a 2 − 1 − a = 0.) Therefore,
this point satisfies (13), because (7) and (13) are equivalent.
Consider the dilogarithm as defined in (11) and suppose that z > 1. According
to [10], we derive:
1 π2
Li2 (z) = −Li2 (z −1 ) − log2 z + − iπ log z
2 3
and

1 z log(1 − 2x cos 0 + x 2 )
Re Li2 (z) =Re Li2 (ze ) = −
i0
dx
2 0 x
 z 
1 log(1 − x)2 z
log |1 − x|
=− dx = − d x = λ(z),
2 0 x 0 x
On the Variational Interpretation of the Discrete KP Equation 391

where λ(z) is the same function as in (9). Therefore, we have

⎧
⎨Li2 (z), z ≤ 1,
λ(z) =
⎩−Li2 (z −1 ) − 1 log2 z + π , z > 1.
2

2 3
By using the following special values [10]

Li2 (a 2 ) = π15 − log2 (−a), Li2 (−a) = π10 − log2 (−a),

2 2

Li2 (a) = − π15 + 21 log2 (−a), Li2 (a −1 ) = − π10 − log2 (−a).

2 2

a straightforward computation gives

L ([i jk]) =L (−[i jkm]) = L ([i jm]) = L (−[ikm]) = L ([ jkm])

1 π2
= (Λ(a 2 ) + Λ(−a −1 ) + Λ(a −1 )) = −
2 20
and

S i jkm =L ([i jk]) + L (−[i jkm]) + L ([i jm]) + L (−[ikm]) + L ([ jkm])

π2
=− .
4
This is, because the expression for L ([i jk]) (see (9)) changes the sign under the
cyclic permutation of indices (i jk) and the solution is invariant under cyclic per-
mutation of indices (i jkm).
Let us now consider the second branch of solutions: one can easily see that

xi j = x jk = xk = xm = xim = a −1 ,

(20)
xik = x j = xkm = xi = x jm = −1

with
√
1 5
a= −
2 2
is a solution of (14) and (15), because (19) is a solution of (13) and (7). Therefore,
on the solution (20) as well as on all other solutions of (14), we have

π2
S i jkm = ,
4

where we used Λ(z) = λ(z) − λ(z −1 ), and, therefore, Λ(z −1 ) = −Λ(z). 

392 R. Boll et al.

Analogously, we get similar results for the white 4-ambo-simplex i jkm

. Here,
the corner equations are:

∂ S̄ i jkm ∂L (Tk [i jm]) ∂L (−T j [ikm]) ∂L (Ti [ jkm])

= + + =0
∂ xi jk ∂ xi jk ∂ xi jk ∂ xi jk

and

∂ S̄ i jkm ∂L (−T [i jkm]) ∂L (−T j [ikm]) ∂L (Ti [ jkm])

= + + = 0.
∂ xi j ∂ xi j ∂ xi j ∂ xi j

Explicitly, they read

1 1
log |E i jk | = 0 and log |E i j | = 0, (21)
xi jk xi j

where
xi jk xkm + xikm x jk xi jk x jm − xi j x jkm xi jk xim + xi jm xik
E i jk := · ·
xi jk xkm − xik x jkm xi jk x jm + xi jm x jk xi jk xim − xi j xikm

and
xi j xkm − xik x jm xi j x jkm − xi jm x jk xi j xikm − xi jk xim
E i j := · · .
xi j xkm + xim x jk xi j x jkm − xi jk x jm xi j xikm − xi jm xik

The analogue of Theorem 3.1 reads:

Theorem 3.3 Every solution of the system (5) solves either the system

E i jk = −1, E i j = −1, E i jm = −1, E ik = −1, E ikm = −1,

(22)
E im = −1, E jk = −1, E jkm = −1, E jm = −1, E km = −1

or the system

E i jk = 1, E i j = 1, E i jm = 1, E ik = 1, E ikm = 1,
(23)
E im = 1, E jk = 1, E jkm = 1, E jm = 1, E km = 1.

Furthermore the system (22) is equivalent to the system (8) (that is dKP on the
corresponding white 4-ambo-simplex). The system (23) is equivalent to the system
On the Variational Interpretation of the Discrete KP Equation 393

xikm xim x jkm x jm − xi jm xim x jkm xkm + xi jm xikm x jm xkm = 0,
xi j xi jm xik xikm − xi jk xi jm xik xim + xi jk xi j xik xim = 0,
x jkm xi jk x jm xi j − x jk xi jk x jm xi jm + x jk x jkm x jm xi jm = 0, (24)
xik x jk xikm x jkm − xkm x jk xikm xi jk + xkm xik xikm xi jk = 0,
x jm xkm xi j xik − xim xkm xi j x jk + xim x jm xi j x jk = 0,

which is the system (8) after the transformation x → x −1 of fields (that is dKP− on
the corresponding white 4-ambo-simplex).
The analogue of Theorem 3.2 reads:
Theorem 3.4 (Closure relation) There holds:

π2
S̄ i jkm ± =0
4
on all solutions of (22) and (23), respectively. Therefore, one can redefine the 3-form
L as

π2
L˜ ([i jk]) := L ([i jk]) ±
4

in order to get S̄ i jkm = 0 on all solutions of (22) and (23), respectively.

4 The Cubic Lattice Z N

We will now consider the relation between the elementary cells of the root lattice
Q(A N ) and the cubic lattice Z N . The points of Q(A N ) and of Z N are in a one-to-one
correspondence via

Pi : Q(A N ) → Z N , x(n 0 , . . . , n i−1 , n i , n i+1 , . . . , n N ) → x(n 0 , . . . , n i−1 , n i+1 , . . . , n N ).

In the present paper, we will always apply Pi with i < j, k, , . . .

We denote by

{ jk} := {x, x j , xk , x , x jk , x j , xk , x jk }

the oriented 3D cubes of Z N . We say that the 3D cube { jk} is positively oriented if
j < k < . Any permutation of two indices changes the orientation to the opposite
one. Also in this case, we always write the letters in the brackets in increasing order,
so, e.g., in writing { jk} we assume that j < k <  and avoid the notation {k j} or
{ jk} for the negatively oriented 3D cube −{ jk}.
394 R. Boll et al.

xk
(a) (b) (c)
xi xi x j xk x ı̄jk

x j

xik
xik x jk

xii xij xij x jk

(d) xk xı̄jk  (e) xk x jk

xi x
x j x j

xik xk
x jk x jk

xii xij x xj

Fig. 2 Three adjacent 3-cells of the lattice Q(A N ): a black tetrahedron −Ti i jk, b octahedron
[i jk], c white tetrahedron −Tı̄ i jk
. The sum d of these 3-cells corresponds to a 3D cube e

The object in Q(A N ) which corresponds to the 3D cube { jk} is the sum of three
adjacent 3-cells, namely
• the black tetrahedron −Ti i jk (see Fig. 2a),
• the octahedron [i jk] (see Fig. 2b),
• and the white tetrahedron −Tı̄ i jk
(see Fig. 2c).
It contains sixteen triangles and to every quadrilateral face of { jkl} there corresponds
a pair of these triangles containing one black and one white triangle. Here, the map
Pi reads as follows:

xii → x, xi j → x j , x jk → x jk , and xı̄ jk → x jk .

As a four-dimensional elementary cell of Z N , we consider an oriented 4D cube

{ jkm} := {x, x j , xk , x , xm , x jk , x j , x jm , xk , xkm , xm , x jk , x jkm , x jm , xkm , x jkm }.

The 4D cube { jkm} corresponds to the sum of four 4-cells in Q(A N ):

• the black 4-simplex −Ti i jkm,
• the black 4-ambo-simplex i jkm,
On the Variational Interpretation of the Discrete KP Equation 395

xı̄km xı̄ı̄jkm

xm x ı̄jm

xk
x ı̄jk
xi x j

xik x jk
xii
xij

xkm xı̄jkm

xim x jm

Fig. 3 The sum of the black 4-simplex −Ti i jkm, the adjacent black 4-ambo-simplex i jkm,
the adjacent white 4-ambo-simplex −Tı̄ i jkm
, and the adjacent white 4-simplex Tı̄ Tı̄ i jkm

corresponds to the 4D cube { jkm}

• the white 4-ambo-simplex −Tı̄ i jkm

, and
• the white 4-simplex Tı̄ Tı̄ i jkm

(see Fig. 3). It contains sixteen tetrahedra (eight black and eight white ones) and eight
octahedra. Here, the map Pi reads as follows:

xii → x, xi j → x j , x jk → x jk , xı̄ jk → x jk , and xı̄ ı̄ jkm → x jkm .

Also in the cubic case there is an easy recipe to obtain the orientation of the facets
of an (oriented) 4D cube: on every index between the brackets we put alternately
a “+” and a “−” starting with a “+” on the last index. Then we get each facet by
deleting one index and putting the corresponding sign in front of the bracket. For
instance., the 4D cube
−+−+
{ j k  m}

has the eight 3D facets: { jk}, −{ jkm}, { jm}, −{km} and the opposite ones
−Tm { jk}, T { jkm}, −Tk { jm}, and T j {km}.
As a consequence of Definition 2.2, in each flower in Z N , every 3D cube has
exactly four adjacent 3D cubes.
We will now prove the analogue of Theorem 2.5. This proof is easier than the one
for Q(A N ), because of the simpler combinatorial structure.
396 R. Boll et al.

Theorem 4.1 The flower at any interior vertex of any 3-manifold in Z N can be
represented as a sum of 4D corners in Z N +1 .
Proof Set M := N + 1 and consider the flower of an interior vertex x of an arbitrary
3-manifold in Z N . Over each 3D corner { jk} (petal) of the flower, we can build
a 4D corner adjacent to x on the 4D cube { jkM}. Then the vertical 3D cubes
coming from two successive petals of the flower carry opposite orientations, so that
all vertical squares cancel away from the sum of the 4D corners. 
Let L be a discrete 3-form on Z N . The exterior derivative dL is a discrete 4-form
whose value at any 4D cube in Z N is the action functional of L on the 3-manifold
consisting of the facets of the 4D cube:

S jkm :=dL({ jkm}) = L({ jk}) + L(−{ jkm}) + L({ jm}) + L(−{km})
+ L(−Tm { jk}) + L(T { jkm}) + L(−Tk { jm}) + L(T j {km}).

Accordingly, the Euler-Lagrange equations on the 4D cube { jkm} are given by

∂ S jkm
= 0,
∂x
∂ S jkm ∂ S jkm ∂ S jkm ∂ S jkm
= 0, = 0, = 0, = 0,
∂x j ∂ xk ∂ x ∂ xm
∂ S jkm ∂ S jkm ∂ S jkm ∂ S jkm ∂ S jkm ∂ S jkm
= 0, = 0, = 0, = 0, = 0, = 0,
∂ x jk ∂ x j ∂ x jm ∂ xk ∂ xkm ∂ xm
∂ S jkm ∂ S jkm ∂ S jkm ∂ S jkm
= 0, = 0, = 0, = 0,
∂ x jk ∂ x jkm ∂ x jm ∂ xkm
∂ S jkm
= 0. (25)
∂ x jkm

They are called corner equations.

The following statement is an immediate consequence of Theorem 4.1:
Theorem 4.2 For every discrete 3-form on Z N and every 3-manifold in Z N all cor-
responding Euler-Lagrange equations can be written as a sum of corner equations.

5 The dKP Equation on Z N

On the 3D cube { jk} in Z3 ( j < k < ) we put the equation

x j xk − xk x j + x x jk = 0. (26)

We can extend this system in a consistent way (see [1]) to the four-dimensional cubic
lattice Z4 and its higher-dimensional analogues, such that the eight facets { jk},
On the Variational Interpretation of the Discrete KP Equation 397

−{ jkm}, { jm}, −{km}, −Tm { jk}, T { jkm}, −Tk { jm}, T j {km} of a 4D cube
{ jkm} carry the equations

x j xk − xk x j + x x jk = 0, x jm xkm − xkm x jm + xm x jkm = 0,

x j xkm − xk x jm + xm x jk = 0, x jk xkm − xk x jkm + xkm x jk = 0,
(27)
x j xm − x x jm + xm x j = 0, x j xkm − xk x jm + xm x jk = 0,
xk xm − x xkm + xm xk = 0, x jk x jm − x j x jkm + x jm x jk = 0.

Note that, in the four equations in the left column, the fields with one index always
appear with increasing order of indices. The equations in the right column are shifted
copies of the ones in the left column. One can derive the system (27) from the system
of dKP equations (7) on the black 4-ambo-simplex i jkm and the system of dKP
equations (8) on the white 4-ambo-simplex Tı̄ i jkm
, by removing the equations
on the octahedra [ jkm] and [ jkm], respectively, from both systems and applying
the transformation Pi to the fields in the remaining eight equations.
We propose the discrete 3-form L defined as

L := (Pi ) L ,

where L is the discrete 3-form on the root lattice Q(A N ) (see (9)). Therefore, L
evaluated at the 3D cube { jk} reads as

L({ jk}) = ((Pi ) L )(Pi (−Ti i jk + [i jk] − Tı̄ i jk

))
= (Pi ) (L (−Ti i jk) +L ([i jk]) − L (−Tı̄ i jk
)) = (Pi ) L ([i jk]).

 
 
=0 =0
For this discrete 3-form, there are no corner equations on the 4D cube { jkm}
centered at x and x jkm since S jkm does not depend on these two variables. The
remaining corner equations from (25) are given by

∂ S jkm ∂L({ jk}) ∂L(−{ jkm}) ∂L({ jm}) ∂L(T j {km})

= + + +
∂x j ∂x j ∂x j ∂x j ∂x j

 
≡0
 
∂L ([i jk]) ∂L (−[i jkm]) ∂L ([i jm]) 1
=(Pi ) + + = log |E j | = 0,
∂ xi j ∂ xi j ∂ xi j xj
∂ S jkm ∂L({ jk}) ∂L(−{ jkm}) ∂L(−Tk { jm}) ∂L(T j {km})
= + + +
∂ x jk ∂ x jk ∂ x jk ∂ x jk ∂ x jk

∂L ([i jk]) ∂L (−[i jkm]) ∂L (−Tı̄ Tk [i jm])
=(Pi ) + +
∂ x jk ∂ x jk ∂ x jk

∂L (Tı̄ T j [ikm])
+
∂ x jk

E
1 jk
= log = 0,
x jk E¯jk
398 R. Boll et al.

∂ S jkm ∂L({ jk}) ∂L(T { jkm}) ∂L(−Tk { jm}) ∂L(T j {km})

= + + +
∂ x jk ∂ x jk ∂ x jk ∂ x jk ∂ x jk

 
≡0
 
∂L (Tı̄ T [i jkm]) ∂L (−Tı̄ Tk [i jm]) ∂L (Tı̄ T j [ikm])
=(Pi ) + +
∂ xı̄ jk ∂ xı̄ jk ∂ xı̄ jk

1 1
= log = 0, (28)
x jk E jk

where

E j := (Pi ) E i j , E jk := (Pi ) E jk , E¯jk := (Pi ) E i jk , and E jk := (Pi ) E jk .

Hereafter, we only consider solutions, where all fields are non-zero (we call these
solutions non-singular). As in the case of the root lattice Q(A N ) every corner equation
has two classes of solutions.
Theorem 5.1 Every solution of the system (25) solves either the system

E j = −1, Ek = −1, E = −1, Em = −1,

E jk = −1, E j = −1, E jm = −1, E k = −1, E km = −1, E m = −1,
E¯jk = −1, E¯j = −1, E¯jm = −1, E¯k = −1, E¯km = −1, E¯m = −1,
E jk = −1, E jkm = −1, E jm = −1, Ekm = −1 (29)

or the system

E j = 1, Ek = 1, E = 1, Em = 1,
E jk = 1, E j = 1, E jm = 1, E k = 1, E km = 1, E m = 1,
E¯jk = 1, E¯j = 1, E¯jm = 1, E¯k = 1, E¯km = 1, E¯m = 1,
E jk = 1, E jkm = 1, E jm = 1, Ekm = 1. (30)

Furthermore the system (29) is equivalent to the system (27) (this is dKP on the
corresponding 4D cube). The system (30) is equivalent to the system

xk x x jk x j − x j x x jk xk + x j xk x j xk = 0,
xk xm x jk x jm − x j xm x jk xkm + x j xk x jm xkm = 0,
x xm x j x jm − x j xm x j xm + x j x x jm xm = 0,
x xm xk xkm − xk xm xk xm + xk x xkm xm = 0,
(31)
xkm xm x jkm x jm − x jm xm x jkm xkm + x jm xkm x jm xkm = 0,
xk xm x jk x jm − x j xm x jk xkm + x j xk x jm xkm = 0,
xk xkm x jk x jkm − x jk xkm x jk xkm + x jk xk x jkm xkm = 0,
x j x jm x jk x jkm − x jk x jm x jk x jm + x jk x j x jkm x jm = 0,
On the Variational Interpretation of the Discrete KP Equation 399

which is the system (27) after the transformation x → x −1 of fields (this is dKP− on
the corresponding 4D cube).

Proof Let x be a solution of the system (25) such that E j = −1 and Ek = −1. Then
we know from the proof of Theorem 3.1 that

E j = −1, Ek = −1, E = −1, Em = −1,

E jk = −1, E j = −1, E jm = −1, E k = −1, E km = −1, E m = −1

and that the latter system is equivalent to

x j xk − xk x j + x x jk = 0,
x j xkm − xk x jm + xm x jk = 0,
x j xm − x x jm + xm x j = 0,
xk xm − x xkm + xm xk = 0,
x jk xm − x j xkm + x jm xk = 0.

On the other hand, if we consider a solution x of (25) such that E j = 1 and Ek = 1,

we know from the proof of Theorem 3.1 that

E j = 1, Ek = 1, E = 1, Em = 1,
E jk = 1, E j = 1, E jm = 1, E k = 1, E km = 1, E m = 1

and that the latter system is equivalent to

xk x x jk x j − x j x x jk xk + x j xk x j xk = 0,
xk xm x jk x jm − x j xm x jk xkm + x j xk x jm xkm = 0,
x xm x j x jm − x j xm x j xm + x j x x jm xm = 0,
x xm xk xkm − xk xm xk xm + xk x xkm xm = 0,
x j x jm xk xkm − x jk x jm xk xm + x jk x j xkm xm = 0.

Now, let x be a solution of the system (25) such that E jk = −1 and E jkm = −1.
Then we know from the proof of Theorem 3.3 that

E¯jk = 1, E¯j = 1, E¯jm = 1, E¯k = 1, E¯km = 1, E¯m = 1,

E jk = 1, E jkm = 1, E jm = 1, Ekm = 1

and that the latter system is equivalent to

400 R. Boll et al.

xm x jkm − xkm x jm + x jm xkm = 0,

xkm x jk − xk x jkm + x jk xkm = 0,
xm x jk − xk x jm + x j xkm = 0,
x jm x jk − x j x jkm + x jk x jm = 0,
x jk xm − x j xkm + x jm xk = 0.

On the other hand, if we consider a solution x of (25) such that E j = 1 and Ek = 1,

we know from the proof of Theorem 3.3 that

E¯jk = 1, E¯j = 1, E¯jm = 1, E¯k = 1, E¯km = 1, E¯m = 1,

E jk = 1, E jkm = 1, E jm = 1, Ekm = 1

and that the latter system is equivalent to

xkm xm x jkm x jm − x jm xm x jkm xkm + x jm xkm x jm xkm = 0,
xk xm x jk x jm − x j xm x jk xkm + x j xk x jm xkm = 0,
xk xkm x jk x jkm − x jk xkm x jk xkm + x jk xk x jkm xkm = 0,
x j x jm x jk x jkm − x jk x jm x jk x jm + x jk x j x jkm x jm = 0,
x j x jm xk xkm − x jk x jm xk xm + x jk x j xkm xm = 0.

Since a solution x of (25) cannot solve

x jk xm − x j xkm + x jm xk = 0

and
x j x jm xk xkm − x jk x jm xk xm + x jk x j xkm xm = 0

at the same time, this proves the theorem. 

Theorem 5.2 (Closure relation) There holds S jkm = 0 on all solutions of (25).

Proof Let x be a solution of (29) or (30). Then

S jkm = dL({ jk}) = (Pi ) (dL (i jkm) + dL (− i jkm

)) = S i jkm − S̄ i jkm
π2 π2
=± ∓ =0
4 4
due to Theorems 3.2 and 3.4 since every solution of (29) solves (13) and (22) after
the transformation Pi of variables and every solution of (30) solves (14) and (23)
after the transformation Pi of variables. 
On the Variational Interpretation of the Discrete KP Equation 401

6 Conclusion

The fact that the three-dimensional (hyperbolic) dKP equation is, in a sense, equiva-
lent to the Euler-Lagrange equations of the corresponding action is rather surprising
since for the two-dimensional (hyperbolic) quad-equations an analogous statement
is not true (see [4, 6] for more details). On the other hand, in the continuous situa-
tion there is an example of a 2-form whose Euler-Lagrange equations are equivalent
to the set of equations consisting of the (hyperbolic) sine-Gordon equation and the
(evolutionary) modified Korteweg-de Vries equation (see [16] for more details). So,
the general picture remains unclear.
In particular, the variational formulation for the other equations of octahedron
type in the classification of [1] is still an open problem.

Acknowledgments This research was supported by the DFG Collaborative Research Center TRR
109 “Discretization in Geometry and Dynamics”.

Appendix 1: Facets of N-Cells of the Root Lattice Q( A N )

Facets of 3-cells:

Black tetrahedrai jk: four black trianglesi jk, −i j, ik, and −  jk;
Octahedra[i jk]: four black trianglesT i jk, −Tk i j, T j ik,
and − Ti  jk,
four white triangles i jk
, − i j
, ik
, and − jk
;
White tetrahedra i jk
: four white trianglesT i jk
, −Tk i j
, T j ik
,
and − Ti jk
;
Facets of 4-cells:

Black 4-simplicesi jkm: f iveblacktetrahedrai jk, −i jkm, i jm,

−ikm, and jkm;
Black 4-ambo-simplicesi jkm: five black tetrahedraTm i jk, −T i jkm,
Tk i jm, −T j ikm, andTi  jkm,
and five octahedra[i jk], −[i jkm], [i jm],
−[ikm], and[ jkm];
White 4-ambo-simplices i jkm
: five octahedraTm [i jk], −T [i jkm], Tk [i jm],
−T j [ikm], andTi [ jkm],
and five white tetrahedra i jk
, − i jkm
,
i jm
, − ikm
, and jkm
;
White 4-simplices i jkm

: five white tetrahedraTm i jk

, −T i jkm
,
Tk i jm
, −T j ikm
, andTi jkm
.
402 R. Boll et al.

Appendix 2: 4D Corners on 4-Cells of the Root Lattice

Q( A N )

Black 4-simplex i jkm:

The 4D corner with center vertex xi contains
• the four black tetrahedra i jk, −i jkm, i jm, and −ikm;
Black 4-ambo-simplex i jkm:
The 4D corner with center vertex xi j contains
• the two black tetrahedra −T j ikm, and Ti  jkm,
• and the three octahedra [i jk], −[i jkm], and [i jm];
White 4-ambo-simplex i jkm
:
The 4D corner with center vertex xi jk contains
• the three octahedra Tk [i jm], −T j [ikm], and Ti [ jkm],
• and the two white tetrahedra i jk
, and − i jkm
;
White 4-simplex i jkm

:
The 4D corner with center vertex xi jk contains
• the four white tetrahedra −T i jkm
, Tk i jm
, −T j ikm
, and Ti jkm
.

Appendix 3: Proof of Theorem 2.4

Set M := N + 1 and L := N + 2. Then, for the construction of the sum Σ of 4D cor-

ners representing the flower σ centered in X , we use the following algorithm:
(i) For every black tetrahedron ±i jk ∈ σ at the interior vertex X we add the
4D corner with center vertex X on the black 4-simplex ±i jkM to Σ.
(ii) For every octahedron ±[i jk] ∈ σ we add the 4D corner with center vertex X
on the black 4-ambo-simplex ±i jkM to Σ.
(iii) For every white tetrahedron ± i jk
∈ σ we add the 4D corner with center
vertex X on the white 4-ambo-simplex ± i jkM
to Σ.
(iv) For every white tetrahedron ± i jk M
∈ Σ \ σ which appeared in Σ during the
previous step we add the 4D corner with center vertex X on the white 4-simplex
∓TL̄ i jk M L

to Σ.
Therefore, we have to prove that Σ = σ .
Assume that X = xi . Then for each black tetrahedron ±i jk ∈ σ we added
the three black tetrahedra ∓i jk M, ±i jM, and ∓ikM to Σ which do not
belong to σ . Moreover, ±i jk has three black triangular facets adjacent to xi ,
namely ±i jk, which is the common triangle with ∓i jk M (up to orientation),
∓i j, which is the common triangle with ±i jM, and ±ik, which is the
common triangle with ∓ikM. Therefore, each of these black tetrahedra has to
On the Variational Interpretation of the Discrete KP Equation 403

cancel away with the corresponding black tetrahedra from the 4D corner which is
coming from the 3-cell adjacent to ±i jk via the corresponding black triangle.
Assume that X = xi j . Then for each octahedron ±[i jk] ∈ σ we added the two
black tetrahedra ∓T j ikM and ±Ti  jkM as well as the two octahedra ∓[i jk M]
and ±[i jM] to Σ which do not belong to σ . Moreover, ±[i jk] has two black
tetrahedral facets adjacent to xi j , namely ±i jk, which is the common triangle with
∓T j ikM, and ∓i j, which is the common triangle with ±Ti  jkM, as well as
two white tetrahedral facets adjacent to xi j , namely ±T j ik
, which is the common
triangle with ∓[i jk M] and ±[i jM], and ∓ jk
, which is the common triangle
with ±[i jM]. Therefore, each of the black tetrahedra ∓T j ikM and ±Ti  jkM
has to cancel away with the corresponding black tetrahedron from the 4D corner
which is coming from the 3-cell adjacent to±[i jk] via the corresponding black
triangle, and each of the octahedra ∓[i jk M] and ±[i jM] has to cancel away with
the corresponding octahedron coming the 4D corner which is coming from the 3-cell
adjacent to ±[i jk] via the corresponding white triangle.
Assume that X = xi jk . Then for each white tetrahedron ± i jk
∈ σ we added
the three octahedra ±Tk [i jM], ∓T j [ikM], and ±Ti [ jkM] as well as the white
tetrahedron ∓ i jk M
to Σ which do not belong to σ . Moreover, ± i jk
has three
white triangular facets adjacent to xi jk , namely ∓Tk i j
, which is the common
triangle with ±Tk [i jM], ±T j ik
, which is the common triangle with ∓T j [ikM],
and ∓Ti jk
, which is the common triangle with ±Ti [ jkM]. Therefore, each
of these octahedra has to cancel away with the corresponding octahedron from the
4D corner which is coming from the 3-cell adjacent to ± i jk
via the corresponding
white triangle.
Consider two 3-cells , ¯ ∈ σ adjacent via the black triangle i jk, say i jk
belongs to and −i jk belongs to . ¯ Then the 4D corner corresponding to con-
tributes the black tetrahedron −i jk M to Σ, whereas the 4D corner corresponding
to ¯ contributes the black tetrahedron i jk M to Σ. Therefore, the latter two black
tetrahedra cancel out.
Consider two 3-cells , ¯ ∈ σ adjacent via the white triangle i jk
, say i jk

belongs to and − i jk
belongs to . ¯ Then the 4D corner corresponding to
contributes the octahedron −[i jk M] to Σ, whereas the 4D corner corresponding to
¯ contributes the octahedron [i jk M] to Σ. Therefore, the latter two octahedra cancel
out.
Up to now we proved that all black tetrahedra and all octahedra in Σ \ σ cancel
out. We will now consider with the white tetrahedra in Σ \ σ .

Lemma 6.1 The white tetrahedra i jk M arising in the third step of the algorithm
build flowers which only contain white tetrahedra.

Proof We have two prove that each of these white tetrahedra has exactly three adja-
cent white tetrahedra in the flowers, one via each white triangle adjacent to X . They
are not adjacent to the 3-cells in σ , but each of them has three common neigh-
bors adjacent to X with the corresponding white tetrahedron in σ . These common
neighbors are octahedra which cancel out in the previous steps of the algorithm.
404 R. Boll et al.

Consider now two white tetrahedra T, T̄ ∈ Σ \ σ after the third step of the algo-
rithm, where the corresponding white tetrahedra in σ are adjacent, i.e., there is a pair
of octahedra with the same set of points and different orientation, one adjacent to
T and the other adjacent to T̄ . Therefore, T and T̄ share a common white triangle
(up to orientation), i.e., they are adjacent.
Consider the 4D corner which we add to Σ for an octahedron. Its two octahedra
which do not belong to σ share a common white triangle (up to orientation) which
does no lie in σ . Furthermore, consider a sequence of adjacent octahedra in σ , where
the common triangles are all white triangles. Then the octahedra in the corresponding
4D corners which are not in σ all share a common white triangle (up to orientation).
Consider now two white tetrahedra T, T̄ ∈ Σ \ σ after the third step of the algo-
rithm, where the corresponding tetrahedra in σ are connected by a sequence of
octahedra adjacent via white triangles. Then, there is a pair of octahedra, one of
them adjacent to T and the other one adjacent to T̄ , which share a common white
triangle (up to orientation). This triangle does not belong to any 3-cell in σ and,
therefore, is a common triangle of T and T̄ , i.e., T and T̄ are adjacent. 
Now we continue with the proof of Theorem 2.4. We already proved that a flower
containing only black tetrahedra can be written as a sum of 3D corners on black 4-
simplices (see proof of step 1). Analogously, one can write every flower containing
only white tetrahedra as a sum of 3D corners on white 4-simplices. So we write for
each of the flowers of white tetrahedra in Σ \ σ after the third step of the algorithm
the flower of opposite orientation as a sum of 3D corners on white 4-simplices and
add this sums to Σ. Then Σ = σ .
Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Adler, V.E., Bobenko, A.I., Suris, Y.B.: Classification of integrable discrete equations of octa-
hedron type. Intern. Math. Res. Not. 2012(8), 1822–1889 (2012)
2. Bobenko, A.I., Suris, Y.B.: Discrete pluriharmonic functions as solutions of linear pluri-
Lagrangian systems. Commun. Math. Phys. 336(1), 199–215 (2015)
3. Boll, R., Petrera, M., Suris, Y.B.: Multi-time Lagrangian 1-forms for families of Bäcklund
transformations. Toda-type systems. J. Phys. A: Math. Theor. 46(275204) (2013)
4. Boll, R., Petrera, M., Suris, Y.B.: What is integrability of discrete variational systems? Proc.
R. Soc. A 470(20130550) (2014)
On the Variational Interpretation of the Discrete KP Equation 405

5. Boll, R., Petrera, M., Suris, Y.B.: Multi-time Lagrangian 1-forms for families of Bäcklund
transformations. Relativistic Toda-type systems. J. Phys. A: Math. Theor. 48(085203) (2015)
6. Boll, R., Petrera, M., Suris, Y.B.: On integrability of discrete variational systems. Octahedron
relations. Intern. Math. Research Notices rnv140 (2015)
7. Conway, J.H., Sloane, N.J.A.: The cell structures of certain lattices. In: Hilton, P., Hirzebruch,
F., Remmert, R. (eds.) Miscellanea Mathematica, pp. 71–107. Springer, Berlin (1991)
8. Doliwa, A.: Desargues maps and the Hirota-Miwa equation. Proc. R. Soc. A 466, 1177–1200
(2010)
9. Doliwa, A.: The affine Weyl group symmetry of Desargues maps and of the non-commutative
Hirota-Miwa system. Phys. Lett. A 375, 1219–1224 (2011)
10. Lewin, L.: Polylogarithms and Associated Functions. Elsevier North Holland, New York (1981)
11. Lobb, S., Nijhoff, F.W.: Lagrangian multiforms and multidimensional consistency. J. Phys. A:
Math. Theor. 42(454013) (2009)
12. Lobb, S., Nijhoff, F.W.: Lagrangian multiform structure for the lattice Gel’fand-Dikii hierarchy.
J. Phys. A: Math. Theor. 43(072003) (2010)
13. Lobb, S., Nijhoff, F.W., Quispel, G.R.W.: Lagrangian multiform structure for the lattice KP
system. J. Phys. A: Math. Theor. 42(472002) (2009)
14. Moody, R.V., Patera, J.: Voronoi and Delaunay cells of root lattices: classification of their facets
by Coxeter-Dynkin diagrams. J. Phys. A: Math. Gen. 25(5089) (1992)
15. Suris, Y.B.: Variational formulation of commuting Hamiltonian flows: multi-Lagrangian 1-
forms. J. Geom. Mech. 5, 365–379 (2013)
16. Suris, Y.B.: Variational symmetries and pluri-Lagrangian systems. arXiv:1307.2639 [math-ph]
(2013)
17. Yoo-Kong, S., Lobb, S., Nijhoff, F.W.: Discrete-time Calogero-Moser system and Lagrangian
1-form structure. J. Phys. A: Math. Theor. 44(365203) (2011)
Six Topics on Inscribable Polytopes

Arnau Padrol and Günter M. Ziegler

Abstract We discuss six topics related to inscribable polytopes, both in dimension

3 (where the topic was started with a problem posed by Steiner in 1832) and in
higher dimensions.
Jakob Steiner ended his 1832 geometry book Systematische Entwicklung der Abhän-
gigkeit geometrischer Gestalten von einander [41] with a list of 85 open problems.
Problem 77 reads as follows:
77) Wenn irgend ein convexes Polyeder gegeben ist läßt sich dann immer (oder in welchen
Fällen nur) irgend ein anderes, welches mit ihm in Hinsicht der Art und der Zusammenset-
zung der Grenzflächen übereinstimmt (oder von gleicher Gattung ist), in oder um eine
Kugelfläche, oder in oder um irgend eine andere Fläche zweiten Grades beschreiben (d.h.
daß seine Ecken alle in dieser Fläche liegen oder seine Grenzflächen alle diese Fläche
berühren)?

It asks whether every (3-dimensional) polytope is inscribable; that is, whether for
every 3-polytope there is a combinatorially equivalent polytope with all the vertices
on the sphere. And if not, which are the cases of 3-polytopes that do have such a
realization? He also asks the same question for circumscribable polytopes, those
that have a realization with all the facets tangent to the sphere, as well as for other
surfaces of degree 2.
There was no progress on this question until 1928, when Ernst Steinitz showed
that inscribability and circumscribability are polar concepts and presented the first

A. Padrol (B)
Institut de Mathématiques de Jussieu - Paris Rive Gauche (UMR 7586),
Sorbonne Universités, UPMC Univ Paris 06, 4 place Jussieu, Case 247,
75252 Paris Cedex 05, France
e-mail: [email protected]
G.M. Ziegler
Institut für Mathematik, Freie Universität Berlin,
Arnimallee 14, 14195 Berlin, Germany
e-mail: [email protected]

© The Author(s) 2016 407

A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_13
408 A. Padrol and G.M. Ziegler

examples of polytopes that cannot be inscribed/circumscribed. Since then, the inter-

est on inscribability of polytopes has soared, partially because of tight relations with
Delaunay subdivisions and hyperbolic geometry. It was in the context of the latter
that, more than 50 years after Steinitz’s results, Igor Rivin finally found a charac-
terization of 3-dimensional inscribable polytopes [34]: A 3-connected planar graph
describes an inscribable polytope if and only if a certain inequality system has a
solution.
What about other quadric surfaces? First of all, since these are not necessarily
convex, one has to decide to either consider realizations with the vertices on the
surface, or realizations whose intersection with the surface are only the vertices. The
weakly inscribable spherical polytopes considered in [8] belong to the first category.
For the second version of the definition, Danciger et al. have very recently extended
Rivin’s results to arbitrary quadrics in R3 [9]: A 3-polytope is inscribable in the
hyperboloid or the cylinder if and only if it is inscribable in the sphere and its graph
is Hamiltonian.
So, is the inscribability problem completely solved? We do not think so: Many
fundamental questions in this area remain still wide open. In particular, very little
is known about inscribability for higher dimensional polytopes. Here we present
some intriguing open questions and problems motivated by some recent (and not so
recent) results on inscribable polytopes.

1 Inscribability of 3-Polytopes

Incribability and circumscribability are polar concepts: A polytope is inscribable if

and only if its polar is circumscribable. Steinitz [42] (cf. [17, Thm. 13.5.2]) con-
structed non-circumscribable polytopes using the following simple fact. Paint some
of the facets of a 3-polytope P black in such a way that there are no two neighboring
black facets (just like on a soccer ball). If you can paint more than half of the facets
black, then P is not circumscribable (unlike the soccer ball).
His argument was the following. Observe that each facet F has a point of contact
with the sphere p F . We associate to each edge e of F the angle with which we see
e from p F . A reflection about the plane spanned by e and the center of the sphere
shows that this angle is the same for the two facets incident to e. Now, if we add all
these angles for the edges incident to black facets, we get 2π for each black facet.
No two of these share an edge, so if we count the contributions for the white facets
we should get at least the same value. However, we get at most 2π times the number
of white facets, which is smaller than the number of black facets by hypothesis. 
The same argument also works if exactly half of the facets are black as long
as there is at least one edge between two white facets. This provides us with our
first example of a non-circumscribable polytope: Take a simplex and truncate all its
vertices. The facets arising from the truncation do not share an edge, so painting
them black shows non-circumscribability.
The polar argument says that if the graph of a 3-polytope has an independent set
(a subset of vertices no two of which are connected by an edge) of more than half of
Six Topics on Inscribable Polytopes 409

the vertices (or exactly half of them but not incident to every edge), then the polytope
is not inscribable. For example, the triakis tetrahedron, a convex polytope obtained
by stacking a tetrahedron onto each facet of a tetrahedron, is not inscribable.
Steinitz’s condition was later subsumed by results by Dillencourt [10] and by Dil-
lencourt together with Smith [11]. In [10] it is shown that the graph of any inscribed
polytope is 1-tough, which means that for any k, removing k vertices splits the graph
into at most k connected components.
This proves non-inscribability for polytopes with an independent set that contains
more than half of the vertices, by removing the other vertices. For independent sets
that collect exactly half the vertices, we need a slight improvement from [11]: the
graph of an inscribable 3-polytope is either bipartite (with both sides of the same
size) or 1-supertough, which means that for any k  2, the removal of k vertices
splits the graph into less than k components.
Toughness is a necessary combinatorial condition for inscribability that is easy
to check, but it is not sufficient.
Besides the mentioned necessary conditions, Dillencourt and Smith also found
some sufficient combinatorial conditions [12], which they summarize as: “If a poly-
hedron has a sufficiently rich collection of Hamiltonian subgraphs, then it is of
inscribable type.” For example, this implies that 3-polytopes whose graphs are
4-connected, or where all the vertex degrees are between 4 and 6, are always inscrib-
able.
So, how can we decide whether a given polytope is incribable? Fundamental
results on hyperbolic polyhedra by Rivin [34] provided an easy way to decide
inscribability/circumscribability of 3-polytopes by linear programming [19] (see
also [32, 33, 35]). If one identifies the ball with the Klein model of the hyperbolic
space, then an inscribed polytope is an ideal hyperbolic polyhedron. Rivin showed
that the dihedral angles at the edges completely characterize these polyhedra.

Theorem 1.1 ([19, Thm. 1]) A 3-polytope P is circumscribable if and only if there
exist numbers ω(e) associated to the edges e of P such that:
• 0< ω(e) < π,
• e∈F ω(e) = 2π for each facet F of P, and
• e∈C ω(e) > 2π for each simple circuit C that does not bound a facet.

Using this result, inscribability and circumscribability can be efficiently checked.

Yet, the characterization depends on a linear programming type feasibility computa-
tion. As Dillencourt and Smith pointed out in [11], it is an outstanding open problem
to find a graph-theoretical characterization of inscribable 3-polytopes. For simple
polytopes, they found such a characterization [11]: A simple 3-polytope is inscrib-
able if and only if its graph is either bipartite and has a 4-connected dual or it is
1-supertough.

Question 1.2 (Dillencourt and Smith [11]) Is there a purely combinatorial charac-
terization of the graphs of inscribable 3-polytopes?
410 A. Padrol and G.M. Ziegler

2 A Characterization in Higher Dimensions

It is not likely that there is a characterization as nice and simple as Rivin’s for higher-
dimensional inscribable polytopes. To start with, 4-dimensional polytopes already
present universality in the sense of Mnëv: This is a very strong statement whose his-
tory took off with groundbreaking results in Nikolai Mnëv’s Ph.D. thesis [25, 26],
and that we discuss a little further in Sect. 5 (see also [30]). It has several implica-
tions, among them that it is already very hard to decide whether a given face lattice
corresponds to a 4-polytope. And as we will see later, higher-dimensional inscrib-
able polytopes also present universality features, so one should not expect to be able
to decide inscribability in higher dimensions easily or quickly.
A more realistic goal would hence be to look for strong necessary conditions
for inscribability in higher dimensions as well as for good (that is, weak) sufficient
conditions. A set of necessary conditions is available, since Steinitz’s proof carries
over directly to higher dimensions, as Grünbaum and Jucovič have observed already
in 1974 [18]. Let’s see how.

Theorem 2.1 Let P be a d-polytope with graph G. If G has an independent set

that contains more than half of all the vertices, or exactly half the vertices when G
is not bipartite, then P is not inscribable.

Proof Again, we prove the polar statement. Assume that P ◦ is circumscribed. Each
facet F touches the ball in a single point p F . To each of the facets of F, which are
ridges of P ◦ , we can intersect the cone with apex p F spanned by the ridge with a
small ball centered at p F . The (normalized) solid angle associated to the ridge is the
ratio of the volume of this intersection to the volume of the ball. Again, a reflection
shows that the solid angle associated to each ridge does not depend on which of the
two incident facets we take the apex from.
Now assume that the facets of P ◦ are painted in black and white, in such a way
that there are no two neighboring black facets. If we add the contributions of the
angles of the ridges associated to black facets, it should be at most (or less than
if the dual graph is not bipartite) the sum of the contributions for the white facets.
Since the sum along each facet is 1, the number of black facets cannot exceed the
number of white facets, and they also cannot be the same if there are two adjacent
white facets. 

The more general necessary conditions by Dillencourt, such as 1-toughness, have

however not yet been generalized to higher dimensions, as far as we know; neither
have the sufficient conditions of Dillencourt and Smith. Is it true that if the graph
of a d-polytope has a rich enough structure of Hamiltonian subgraphs then the cor-
responding polytope is inscribable? This could explain the phenomena observed in
the next section.

Question 2.2 Find strong necessary and sufficient conditions for inscribability of
higher-dimensional polytopes.
Six Topics on Inscribable Polytopes 411

Firsching recently achieved a complete enumeration of the simplicial 4-polytopes

with 10 vertices [13]: There are exactly 162 004 combinatorial types. Firsching also
managed to decide inscribability for all but 13 of these polytopes: The remaining
161 991 are divided into 161 978 inscribable and 13 non-inscribable. So, most sim-
plicial 4-polytopes with 10 vertices are inscribable. Moreover, the very few that
are not inscribable have very few facets and edges. More precisely, every non-
inscribable simplicial 4-polytope with 10 vertices has less than 27 facets and less
than 37 edges, but 146 104 out of the 162 004 simplicial 4-polytopes with 10 ver-
tices do have at least 27 facets and 37 edges. This seems indicate that there might
be combinatorial sufficient conditions for inscribability based only on the f -vector.

Question 2.3 Is it true that most simplicial 4-polytopes on n vertices are inscribable,
for n → ∞?

In 1991 Smith [40] proved that although there are exponentially many inscribable
and circumscribable 3-polytopes with n vertices (because there are many that are
4-connected), most simplicial 3-polytopes with n vertices are neither inscribable
nor circumscribable. The proof consists in showing that the probability of finding a
fixed non-inscribable subgraph in a random 3-connected planar triangulations tends
to 1 as n → ∞. The same argument can be used on random 3-connected planar
graphs, see [6, Thm. 2 and Cor. 1], which shows that most combinatorial types of
3-polytopes on a given large number of edges are not inscribable/circumscribable.
This contrasts with what happens with simplicial 3-polytopes with few vertices.
In the same paper [40], Smith classified these according to their inscribability and
circumscribility, and most turned out to be both inscribable and circumscribable.
The referee for this paper suggested to use the strategy for 3-polytopes in higher
dimension: To show that a large random simplicial d-polytope is likely to contain a
fixed non-inscribable subcomplex.

3 Neighborly Polytopes

What is the maximal number of faces that a d-dimensional polytope with n vertices
can have? The answer to this fundamental question in the combinatorial theory of
polytopes was not established until 1970, by McMullen [22], although Motzkin had
already guessed the answer, as we know from a 1957 abstract [27]. And the answer
is that the cyclic d-polytopes with n vertices have the maximal number of k-faces
among all d-polytopes with n vertices (for all k!). This is the polytope one obtains
when taking the convex hull of n points on the moment curve given by γ(t) :=
(t, t 2 , t 3 , . . . , t d ).
The cyclic polytopes owe their name to Gale [14], one of several (re-) inven-
tors (cf. [17, Sec. 7.4]), although they were essentially already known to Constantin
Carathéodory in 1911 [7], who had studied the convex hull of the trigonometric
moment curve (sin t, cos t, sin 2t, cos 2t, . . . , sin kt, cos kt). The representation on
412 A. Padrol and G.M. Ziegler

the trigonometric moment curve, which is projectively equivalent to the monomial

one [45, pp. 75/76], shows that in even dimensions, the cyclic polytopes are inscrib-
able. Since then, several inscribed realizations of the cyclic polytope have been
found (also in odd dimensions), see [17, p. 67], [39, p. 521] and [16, Prop. 17].
Thus the upper bound theorem is established for inscribed polytopes too.
The next natural question asks for all the polytopes that have this many facets.
McMullen also provided the answer to this: This characterizes the simplicial
  neigh-
borly polytopes. These are the simplicial polytopes with a complete d2 -skeleton.
(All even-dimensional neighborly polytopes are simplicial, but not in general the
odd-dimensional ones. In particular, not all odd-dimensional neighborly polytopes
have the maximal number of facets, only those that are simplicial.) Even if Motzkin
claimed that the cyclic polytopes are the only neighborly polytopes (in the same
1957 abstract [27] mentioned above), there are actually plenty. The number of neigh-
borly d-polytopes with n vertices grows at least as n d/2n(1−o(1)) for fixed d [29].
Compare this to the number of d-polytopes with n vertices, which is not larger than
2
n d n(1+o(1)) [3]. As it turns out, each one of the neighborly polytopes used to provide
this lower bound is also inscribable [15] (for d  4). And even more, as we discuss
in the next section, they are inscribable in any strictly convex body! This surprising
behavior let Gonska and Padrol [15] to ask:

Question 3.1 (Gonska and Padrol [15]) Is every (even-dimensional) neighborly

polytope inscribable?

Firsching [13] undertook a quest to find a counterexample. He successfully used

non-linear optimization techniques to find polytope realizations. In particular, he
tried to inscribe the neighborly polytopes enumerated in [24]. The results are sur-
prising: Every neighborly 4-polytope with n  11 vertices, every simplicial neigh-
borly 5-polytope with n  10 vertices, every neighborly 6-polytope with n  11
vertices, and every simplicial neighborly 7-polytope with n  11 vertices is inscrib-
able! Even more, every simplicial 2-neighborly 6-polytope with n  10 vertices is
also inscribable. This lead Firsching to ask whether the even stronger statement that
all 2-neighborly polytopes are inscribable might be true [13, Conj. 1]. This may be a
very brave conjecture, for which not much evidence is available yet. However, recall
that the graph of a 2-neighborly polytope is complete, and hence has the richest pos-
sible structure of Hamiltonian subgraphs . . .

Question 3.2 (Firsching [13]) Is every (simplicial) 2-neighborly polytope inscrib-

able?

We can also ask the polar question: What about circumscribability of neighborly
polytopes? The results are completely opposite. Chen and Padrol proved that, for
any d  4, no cyclic d-polytope on sufficiently many vertices is circumscribable [8].
(The proof will be sketched below in the last section.) They even conjecture that
this holds in more generality for neighborly polytopes [8, Conj. 7.4]. Notice that,
since neighborliness can be read on the f -vector, this would imply that there is an
f -vector of a convex polytope that does not belong to any inscribable polytope,
Six Topics on Inscribable Polytopes 413

namely that of the polar of a neighborly 4-polytope with sufficiently many ver-
tices. No example of such an f -vector has been established so far. Actually, what is
known points into the other direction: Every f -vector of a 3-dimensional polytope
is inscribable [16]. However, f -vectors that are not k-scribable are known for d  4
and 1  k  d − 2 [8]. (A polytope is k-scribable if it has a realization with all its
k-faces tangent to the sphere.)
Question 3.3 (Gonska and Ziegler [16]) Is there an f -vector that is not inscribable?

4 Universally Inscribable

As we just mentioned, the proof of inscribability for cyclic and many more neigh-
borly polytopes given in [15] still works when we replace the unit ball by any other
smooth strictly convex body. (This is what Oded Schramm called an egg in his cel-
ebrated paper “How to cage an egg?” [37].) Here smoothness is not very important,
but strict convexity is. For example, the pigeonhole principle tells us that no simpli-
cial d-polytope with more than d(d + 1) vertices can be inscribed on the boundary
of a d-simplex.
Hence many neighborly polytopes (among them all cyclic polytopes) are univer-
sally inscribable: They can be inscribed into any egg. Other examples of such poly-
topes are the stacked d-polytopes arising as a join of a path with a (d − 2)-simplex;
and also Lawrence polytopes [15].
Question 4.1 (Gonska and Padrol [15]) Which polytopes are inscribable into the
boundary of every (smooth) strictly convex body?
An observation of Karim Adiprasito [15] shows that being inscribable on the
sphere is not sufficient for being universally inscribable. (The proof uses projec-
tively unique polytopes [2].) We thank the anonymous reviewer for suggesting the
following opposite question. (The reviewer’s conjectured answer is yes.)
Question 4.2 Are there polytopes that are inscribable in every egg other than the
ellipsoid?
The celebrated Koebe–Andreev–Thurston Theorem states that every 3-polytope
has a realization with all its edges tangent to the sphere. This amazing result has a
long history. It seems that it was first proved by Paul Koebe, but only for simple and
simplicial polytopes [21]. Thurston later realized [44] that it followed from results
of Andreev on hyperbolic polyhedra [4, 5]. Since then, several proofs have been
found (see [46, Sect. 1.3] and references therein). Schramm went even further and
proved that every 3-polytope has a realization with all its edges tangent to any given
egg [37].
Theorem 4.3 (Schramm [37]) For every 3-polytope P, and every smooth strictly
convex body K , there is a realization Q of P such that each edge of Q is tangent
to K .
414 A. Padrol and G.M. Ziegler

There is also the other side of universal inscribability, which would be to focus
on the convex bodies. The mathoverflow user called Samsa asked [36]:
Is there a convex body in Rd such that every combinatorial type of a d-dimensional convex
polytope can be realized with vertices on its surface?

We reproduce the beautiful affirmative answer by Ivanov [20]:

Yes, there is such a body. Actually there is one very close to the standard unit ball and
containing disjoint representatives of each combinatorial type (but these representatives are
very small).
Indeed, every combinatorial type of a d-polytope has a realization which looks as follows:
there is a “large” (d − 1)-dimensional facet and the remaining surface is a graph over this
facet. To construct such a realization, choose a (d − 1)-facet, pick a hyperplane parallel to
it and very close to it (but not intersecting the polytope), and apply a projective map which
sends this hyperplane to infinity.
We can further “flatten” this realization so that it is very close to its large face. Choose a
very small ε > 0, apply a homothety such that the diameter of the polytope becomes less
than ε, and place the resulting tiny polytope so that it touches the sphere by a point on its
“large” face. Then consider the convex hull of the sphere and the polytope. All vertices will
be on the boundary of this convex hull if the polytope is sufficiently √ “flattened.” And the
convex hull diverges from the ball only in a neighborhood of size ∼ ε.
Now pick another combinatorial type of a polytope and repeat the procedure with a much
smaller ε and a location on the sphere chosen so that the neighborhood affected by the sec-
ond polytope does not interfere with the first one. And so on. Since there are only countably
many combinatorial types, they all can be packed into the sphere, provided that ε goes to 0
sufficiently fast.

5 Universality

We move on to a slightly different topic, and a different notion of “universality.”

Take an inscribable polytope P, and consider the set of all inscribed realizations
of P. How does this set look like? To start off, it can be parametrized by the ver-
tex coordinates. Moreover, since any Möbius transformation that preserves the unit
sphere as well as its interior sends an inscribed polytope onto an inscribed polytope,
we can safely mod out the action of this group. This defines Rins (P), the realization
space of an inscribed polytope P (with n vertices in Rd ):
 
d−1 n 

Rins (P) = A ∈ (S ) conv(A)
P Möb(Sd−1 ). (1)

This concept is inspired by an analogue definition for general polytopes. In that

setting, one considers Rpol (P), the set of all realizations of P, up to affine trans-
formation. From Steinitz’s proofs of Steinitz’s Theorem [43] we know that these
sets are relatively nice when P is of dimension at most 3: They are contractible,
contain rational points, etc.; see [30, 31]. For higher-dimensional polytopes, how-
ever, the behavior of realization spaces is much wilder. This was first observed by
Six Topics on Inscribable Polytopes 415

Mnëv, with his celebrated Universality Theorem [25, 26] for polytopes and ori-
ented matroids. The polytopal version reads: For every primary basic [open] semi-
algebraic set defined over Z there is a [simplicial] polytope whose realization space
is stably equivalent to it. This has many consequences, among them:
• topological: Rpol (P) can have the homotopy type of any arbitrary finite simplicial
complex,
• algebraic: there are polytopes that cannot be realized with rational coordinates,
• algorithmic: it is ETR-hard to decide if a lattice is the face lattice of a polytope.
Mnëv’s proof provided polytopes with the desired realization spaces, but could not
say anything about their dimensions. Another major step was done later by Jür-
gen Richter-Gebert, who proved that there is universality already for 4-dimensional
polytopes [30].
The inscribed picture is similar. Up to dimension 3, inscribed realization spaces
are reasonable. This follows from the results of Rivin that we have already men-
tioned [33], which imply that for a 3-polytope P, Rins (P) is homeomorphic to the
polyhedron of angle structures (and hence are contractible). In contrast, results of
Adiprasito and Padrol with Louis Theran show that, in arbitrarily high dimensions,
there is again universality [1]. We have not found yet an appropriate notion of sta-
ble equivalence for this context, while a universality theorem for general polytopes
without stable equivalence is neither available nor in sight. This forces us to separate
the topological, algebraic and algorithmic statements:

Theorem 5.1 (Adiprasito et al. [1])

• For every primary basic semi-algebraic set there is an inscribed polytope whose
realization space is homotopy equivalent to it.
• For every finite field extension F/Q of the rationals, there is an inscribed polytope
that cannot be realized with coordinates in F.
• The problem of deciding if a poset is the face lattice of an inscribed [simplicial]
polytope is polynomially equivalent to the existential theory of the reals (ETR).
In particular, it is NP-hard.

In the last point we can even ask the polytopes to be simplicial. This follows from
a weak universality theorem for inscribed simplicial polytopes, which also appears
in [1]. In this case, we can only find polytopes whose realization space retracts onto
the semi-algebraic set, instead of having homotopy equivalence as in the general
case.
The inscribed analogue to Richter-Gebert’s result for 4-polytopes is still missing.

Question 5.2 (Adiprasito et al. [1]) Is there universality for inscribed polytopes in
bounded dimension, say for 4-polytopes inscribed into S3 ?

The proof of Theorem 5.1 strongly relies on the results of Mnëv. The strategy
is to start with certain polytopes with intricate realization spaces, and then to show
that their inscribed realization spaces are equally involved. In particular, it does not
416 A. Padrol and G.M. Ziegler

prove that it is hard to decide inscribability once we already know that the face lat-
tice corresponds to a polytope. However, inscribability is itself a complex condition,
and hence one can expect that it increases the complexity of the corresponding real-
ization spaces. This could lead to a proof of universality that is intrinsic to inscribed
polytopes, and hopefully to advances in the previous question. A first step in this
direction could be to find a polytope P such that Rins (P) is disconnected while
Rpol (P) is not.

Question 5.3 Is there universality for the realization spaces Rins (P) for a class of
inscribable polytopes P whose general realization spaces Rpol (P) are trivial (in
particular, contractible)?

6 (i, j)-Scribability

We have already referred to the Koebe–Andreev–Thurston Theorem. One is tempted

to ask if similar behaviors might also appear in higher dimensions. The answer is no.
Egon Schulte used an inductive argument over non-inscribable/non-circumscribable
3-polytopes to show that, for every d  4, and every 0  k  d − 1, there are d-
polytopes that cannot be realized with all their k-faces tangent to the sphere [38].
Of course, this opens the door to ask for a characterization of k-scribable poly-
topes. There are almost no results in this direction, which is definitely an interesting
research topic.
We will, however, consider a different problem. If we take an edge-scribed real-
ization of a 3-polytope, it has also the following property: All the vertices are outside
the ball while all the facets cut the ball. This kind of realizations are studied in [8].
Here a polytope is said to be (i, j)-scribed if all its i-faces “avoid” the sphere while
all the j-faces “cut” it. The definitions for cutting and avoiding that seem to work
better say that a face F cuts the ball if there is a point of the unit ball in its rela-
tive interior, and that it avoids the ball if there is a hyperplane H , supporting for F,
which completely contains P and the ball in one of the closed halfspaces it defines.
(This somehow involved definition is needed if we want a self-dual concept that
reduces to classical k-scribability when i = j = k.)
The weakest condition occurs when i = 0 and j = d − 1:
Question 6.1 Does every polytope have a realization where every vertex avoids the
ball and every facet cuts the ball?
The answer is most probably no, but so far no proof has been given (although
Karim Adiprasito has suggested that one should be able to obtain counterexamples
from gluing some of the large projectively unique polytopes from [2]). There are
some results in the opposite direction: For d = 3, the edge-scribed realization is
also (0, d − 1)-scribed. Every inscribable polytope has directly a (0, d − 1)-scribed
realization too. So, in particular, cyclic polytopes and many (all?) neighborly poly-
topes have such realizations. And obviously, circumscribable polytopes also have
Six Topics on Inscribable Polytopes 417

such a realization. This includes stacked polytopes, which are always circumscrib-
able [8].
What about other values of i and j? The best offenders found in [8] are even-
dimensional cyclic polytopes (and their duals):

Theorem 6.2 (Chen and Padrol [8]) If an even-dimensional cyclic polytope has
sufficiently many vertices, then it is not (1, d − 1)-scribable, which in particular
implies that it is not circumscribable.

Proof (Sketch) The first step is to associate to each vertex of a (1, d − 1)-scribed
cyclic polytope the spherical cap consisting of the points of Sd−1 visible from it.
This yields a configuration of spherical caps on Sd−1 . These are said to form a k-ply
system if no point of Sd−1 is contained in the interior of more than k caps.
Now the Sphere Separator Theorem of Miller et al. [23] states that the intersec-
tion graph of a k-ply system on Sd−1 has a separator of size O(k 1/(d−1) n 1−1/(d−1) ).
The proof is astonishingly simple and beautiful (cf. [28, Thm. 8.5]): With a Möbius
transformation, we can assume that the origin is a center-point of the centers of the
caps. The next step is to compute the probability that a random linear hyperplane
intersects a cap, which depends only on the area covered by the cap. Since our sys-
tem is k-ply, we can estimate the sum of these volumes because we know the surface
area of the sphere. This is used to show that a random linear hyperplane hits very
few caps, whose removal separates the graph.
So how do we tie this in with cyclic polytopes? The key observation is that a
set of points induce a k-ply system if and only if the convex hull of every k-set
intersects the sphere. A k-set is a subset of k points that can be separated from
the others with a hyperplane. Even-dimensional cyclic polytopes have a lot of nice
properties, among them oriented matroid rigidity. This allows us to show that every
k-set with k  23 d − 1 contains a facet. If the realization was (0, d − 1)-scribed,
this facet would intersect the sphere, and hence the cyclic polytope induces a k-ply
set. But in this case, the intersection graph consists of the edges of the polytope
avoiding the sphere. If all the edges avoided the sphere, this would be a complete
graph, which obviously does not have a small separator. 

Acknowledgments We are grateful to Karim Adiprasito, Hao Chen, Igor Rivin, Juanjo Rué, Fran-
cisco Santos, and the anonymous referee for very valuable discussions, suggestions, and references;
and to Sergei Ivanov for letting us reproduce his mathoverflow answer.

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.
418 A. Padrol and G.M. Ziegler

References

1. Adiprasito, K., Padrol, A., Theran, L.: Universality theorems for inscribed polytopes and
Delaunay triangulations. Discrete Comput. Geom. 54, 412–431 (2015)
2. Adiprasito, K., Ziegler, G.M.: Many projectively unique polytopes. Inventiones Math. 199,
581–652 (2015)
3. Alon, N.: The number of polytopes, configurations and real matroids. Mathematika 33(1),
62–71 (1986)
4. Andreev, E.M.: On convex polyhedra in Lobačevskıı̆ spaces. Math. USSR, Sb. 10, 413–440
(1971)
5. Andreev, E.M.: On convex polyhedra of finite volume in Lobačevskıı̆ space. Math. USSR, Sb.
12, 255–259 (1971)
6. Bender, E.A., Gao, Z.C., Richmond, L.B.: Submaps of maps. I. general 0–1 laws. J. Combinat.
Theory, Ser. B 55(1), 104–117 (1992)
7. Carathéodory, C.: Über den Variabilitätsbereich der Fourier’schen Konstanten von positiven
harmonischen Funktionen. Rendiconto del Circolo Matematico di Palermo 32, 193–217
(1911)
8. Chen, H., Padrol, A.: Scribability problems for polytopes. Preprint, August 2015, 24 pp.,
http://arxiv.org/abs/1508.03537
9. Danciger, J., Maloni, S., Schlenker, J.M.: Polyhedra inscribed in a quadric. Preprint, Octo-
ber 2014, 42 pp., http://arxiv.org/abs/1410.3774
10. Dillencourt, M.B.: Toughness and Delaunay triangulations. Discrete Comput. Geom. 5(6),
575–601 (1990)
11. Dillencourt, M.B., Smith, W.D.: A linear-time algorithm for testing the inscribability of triva-
lent polyhedra. Internat. J. Comput. Geom. Appl. 5(1-2), 21–36 (1995). Eighth Annual ACM
Symposium on Computational Geometry (Berlin, 1992)
12. Dillencourt, M.B., Smith, W.D.: Graph-theoretical conditions for inscribability and Delaunay
realizability. Discrete Math. 161(1–3), 63–77 (1996)
13. Firsching, M.: Realizability and inscribability for some simplicial spheres and matroid poly-
topes. Preprint, October 2015, 25 pp., http://arxiv.org/abs/1508.02531
14. Gale, D.: Neighborly and cyclic polytopes. In: Proceedings of Symposia in Pure Mathematics,
vol. VII, pp. 225–232. American Mathematical Society, Providence (1963)
15. Gonska, B., Padrol, A.: Neighborly inscribed polytopes and Delaunay triangulations. Adv.
Geom. (in press). Preprint http://arxiv.org/abs/1308.5798v2
16. Gonska, B., Ziegler, G.M.: Inscribable stacked polytopes. Adv. Geom. 13(4), 723–740 (2013)
17. Grünbaum, B.: Convex polytopes. Graduate Texts in Mathematics, vol. 221, 2nd edn.
Springer-Verlag, New York (2003)
18. Grünbaum, B., Jucovič, E.: On non-inscribable polytopes. Czechoslovak Math. J. 24(99), 424–
429 (1974)
19. Hodgson, C.D., Rivin, I., Smith, W.D.: A characterization of convex hyperbolic polyhedra and
of convex polyhedra inscribed in the sphere. Bull. Amer. Math. Soc. 27(2), 246–251 (1992)
20. Ivanov, S.: Can all convex polytopes be realized with vertices on surface of convex body?
MathOverflow, http://mathoverflow.net/q/107113, September 2012
21. Koebe, P.: Kontaktprobleme der konformen Abbildung. Berichte Verh. Sächs. Akademie der
Wissenschaften Leipzig, Math.-Phys. Klasse 88, 141–164 (1936)
22. McMullen, P.: The maximum numbers of faces of a convex polytope. Mathematika 17, 179–
184 (1970)
23. Miller, G.L., Teng, S.H., Thurston, W., Vavasis, S.A.: Separators for sphere-packings and near-
est neighbor graphs. J. ACM 44(1), 1–29 (1997)
24. Miyata, H., Padrol, A.: Enumerating of neighborly polytopes and oriented matroids. Exp.
Math. 24(4), 489–505 (2015)
25. Mnëv, N.E.: The topology of configuration varieties and convex polytopes varieties. Ph.D.
thesis, St. Petersburg State University, St. Petersburg, RU (1986). 116 pp., http://www.pdmi.
ras.ru/~mnev/mnev_phd1.pdf
Six Topics on Inscribable Polytopes 419

26. Mnëv, N.E.: The universality theorems on the classification problem of configuration varieties
and convex polytopes varieties. In: Topology and Geometry—Rohlin Seminar. Lecture Notes
in Mathematics, vol. 1346, pp. 527–544. Springer-Verlag, Berlin (1988)
27. Motzkin, T.S.: Comonotone curves and polyhedra (Abstract). Bull. Am. Math. Soc. 63, 35
(1957)
28. Pach, J., Agarwal, P.K.: Combinatorial Geometry. Wiley-Interscience Series in Discrete Math-
ematics and Optimization. Wiley, New York (1995)
29. Padrol, A.: Many neighborly polytopes and oriented matroids. Discrete Comput. Geom. 50(4),
865–902 (2013)
30. Richter-Gebert, J.: Realization spaces of polytopes. Lecture Notes in Mathematics, vol. 1643.
Springer-Verlag, Berlin (1996)
31. Richter-Gebert, J., Ziegler, G.M.: Realization spaces of 4-polytopes are universal. Bull. Am.
Math. Soc. 32, 403–412 (1995)
32. Rivin, I.: On geometry of convex ideal polyhedra in hyperbolic 3-space. Topology 32(1), 87–
92 (1993)
33. Rivin, I.: Euclidean structures on simplicial surfaces and hyperbolic volume. Ann. Math. (2)
139(3), 553–580 (1994)
34. Rivin, I.: A characterization of ideal polyhedra in hyperbolic 3-space. Ann. Math. (2) 143(1),
51–70 (1996)
35. Rivin, I.: Combinatorial optimization in geometry. Adv. Appl. Math. 31(1), 242–271 (2003)
36. Samsa, G.: Can all convex polytopes be realized with vertices on surface of convex body?
MathOverflow. http://mathoverflow.net/q/107096, September 2012
37. Schramm, O.: How to cage an egg. Inventiones Math. 107(3), 543–560 (1992)
38. Schulte, E.: Analogues of Steinitz’s theorem about non-inscribable polytopes. In: Proceedings
of “Intuitive Geometry” (Siófok, 1985). Colloq. Math. Soc. János Bolyai, vol. 48, pp. 503–
516. North-Holland, Amsterdam (1987)
39. Seidel, R.: Exact upper bounds for the number of faces in d-dimensional Voronoı̆ diagrams. In:
Applied Geometry and Discrete Mathematics. DIMACS Ser. Discrete Math. Theoret. Comput.
Sci., vol. 4, pp. 517–529. American Mathematical Society, Providence (1991)
40. Smith, W.D.: On the enumeration of inscribable graphs. Manuscript, 7 pp., NEC Research
Institute, 1991; http://citeseer.ist.psu.edu/viewdoc/summary?doi=10.1.1.29.4543
41. Steiner, J.: Systematische Entwicklung der Abhängigkeit geometrischer Gestalten von einan-
der. Fincke, Berlin (1832). Also in: Gesammelte Werke, vol. 1, Reimer, Berlin 1881, pp. 229–
458
42. Steinitz, E.: Über isoperimetrische Probleme bei konvexen Polyedern. J. Reine Angew. Math.
159, 133–143 (1928)
43. Steinitz, E., Rademacher, H.: Vorlesungen über die Theorie der Polyeder. Springer-Verlag,
Berlin (1934). Reprint, Springer-Verlag (1976)
44. Thurston, W.P.: Geometry and topology of 3-manifolds. In: Lecture Notes. Princeton Univer-
sity, Princeton (1977–1978); http://library.msri.org/books/gt3m/
45. Ziegler, G.M.: Lectures on polytopes. Graduate Texts in Mathematics, vol. 152. Springer, New
York (1995)
46. Ziegler, G.M.: Convex polytopes: extremal constructions and f -vector shapes. In: Geomet-
ric Combinatorics. IAS/Park City Math. Ser., vol. 13, pp. 617–691. American Mathematical
Society, Providence (2007)
DGD Gallery: Storage, Sharing,
and Publication of Digital Research Data

Michael Joswig, Milan Mehner, Stefan Sechelmann,

Jan Techter and Alexander I. Bobenko

Abstract We describe a project, called the DGD Gallery, whose goal is to store
geometric data and to make it publicly available. The DGD Gallery offers an online
web service for the storage, sharing, and publication of digital research data.

1 Introduction

Software produces data. Mathematical software produces scientific data, and this is
often worth keeping. One reason for this can be the vast amount of CPU time spent
on a specific experiment. Another reason can be that the output is obtained only via a
complex interaction process between the software and its user. That latter situation is
typical in mathematical visualization, where producing a satisfying or even beautiful
picture of a geometric object is a form of art. The purpose of this text is to describe a
new project, called the “Discretization in Geometry and Dynamics Gallery”, or DGD
Gallery for short, whose goal is to store geometric data and to make it publicly
available. The URL of the web-site is
http://gallery.discretization.de

M. Joswig (B) · M. Mehner · S. Sechelmann · J. Techter · A.I. Bobenko

Inst. für Mathematik, Technische Universität Berlin,
Straße des 17. Juni 136, 10623 Berlin, Germany
e-mail: [email protected]
M. Mehner
e-mail: [email protected]
S. Sechelmann
e-mail: [email protected]
J. Techter
e-mail: [email protected]
A.I. Bobenko
e-mail: [email protected]

© The Author(s) 2016 421

A.I. Bobenko (ed.), Advances in Discrete Differential Geometry,
DOI 10.1007/978-3-662-50447-5_14
422 M. Joswig et al.

Today it is safe to say that finally mathematical software has reached every branch
of mathematics. While computers have played a role in mathematical applications
for a long time, it took considerably longer for software to be appreciated fully
in parts of mathematics traditionally considered as “pure”. The array of tools at our
fingertips now includes solvers for linear programs and partial differential equations,
but also software for dealing with real algebraic sets [1] or delicate constructions in
sheaf theory [2]. To get an idea of how rich the mathematical software landscape has
become, see, e.g., [3]. The success of each single software system raises the question
of how the respective data produced should be stored. With an increasing number of
relevant mathematical results relying on non-trivial computations in an essential way
(see, e.g., the Flyspeck project [4]) it becomes more and more crucial to publish
such results in a way such that they can be scrutinized (and used) by the mathematical
community.
The mathematical data we have in mind for the DGD Gallery are the geo-
metric objects that occur naturally on the border between differential geometry and
geometric combinatorics. This includes various classes of surfaces (embedded or
immersed) in 3-space, convex polytopes and polyhedral fans of various dimensions,
circle patterns, and many more. Yet, we believe that several of our design decisions
and architecture ingredients will be useful for other collections of mathematical data.
Key features include the following:
• structured storage of research data,
• review process for increased reliability,
• migration process for sustainability,
• licensing scheme.
To further stress the relevance of our endeavor, it is worth noting that scientific
funding agencies have begun to add requirements concerning the preservation of
scientific data to their regulations. For instance, in a recent announcement [5] of
Deutsche Forschungsgemeinschaft (DFG) says1 :
The documentation of research data according to standards depending on the subject and
their long-term archival are relevant for controlling the quality of scientific work. Further,
these data are the basic requirements for the subsequent use of research results.

The DGD Gallery evolved as a project within the DFG Collaborative Research
Center SFB/TRR 109 “Discretization in Geometry and Dynamics”. Its usage is cur-
rently restricted to the members of the center. However, it is intended that future
versions allow other researchers to contribute their work, too.
The paper is organized as follows. First we compare our design to existing collec-
tions of geometric data (Sect. 2). Then, in Sect. 3, we exhibit some examples already
published on the gallery. This should give a good idea of what kind of collection
we have in mind. At the same time this also shows some of the technical features
and capabilities. The core is Sect. 4, where we elaborate on the architecture and the
design decisions. The key concept is the model, which is our technical realization of
a geometric object. Some aspects of the implementation are covered in Sect. 5. For

1 Translated from German.

DGD Gallery: Storage, Sharing, … 423

instance, we explain how we use the XML document database BaseX [6] and meet
current standards of web technology.

2 Comparison with Previous Work

To store geometric data digitally and make it accessible through a web-site is clearly
not a new idea. On the contrary, since the early days of the Internet people have
set up numerous web-sites with all kinds of information on geometric objects, e.g.:
The Geometry Center’s “Geometry Reference Archive” [7], “The Scientific Graph-
ics Project”of MSRI [8], or David Eppstein’s “Geometry Junkyard” [9], to name a
few prominent examples. Clearly, all of the above still contain lots of interesting
information. However, there are some shortcomings. In the case of the archive of
The Geometry Center we have a static collection of data that will not see any updates
or additions. Yet there is the advantage that all data is available from one source, and
so it cannot degenerate over time (except for eventually outdated file formats). Not
so with the “Geometry Junkyard”. This is a collection of links to other interesting
resources on the web. Many of the links are dead already. This is mainly due to a
discontinued provider service or simply a change of position of the person who pro-
vided the data. The “Scientific Graphics Project” is mainly a collection of surfaces
and differential geometry related publications. The DGD Gallery wants to cover
geometric objects from a much wider collection.
A more recent project is the “GeometrieWerkstatt” [10] maintained by a group
of geometers at Tübingen University. It contains visualizations of mostly smooth
constant mean curvature surfaces. Surfaces are visualized using videos, images, and
interactive 3D viewers. The main difference to the DGD Gallery is that there is
no geometric data that can be accessed via the web-site. On the other hand, how to
provide the data for smooth surfaces is far from obvious and cannot be separated
from the mathematical methods. For the DGD Gallery we propose to include a
discretization of a smooth surface in a reasonable resolution.
“IMAGINARY—open mathematics”is a platform [11] which has a strong edu-
cational focus. It features images and mathematical software for a broad audience
such as exhibitions, high school education, and museums. As an essential feature,
IMAGINARY is open for the public to contribute material by cross-linking to other
web-sites. In this way it works like a collection of collections.
The focus of the SymbolicData project [12] is on developing concepts and tools
for profiling, testing and benchmarking Computer Algebra Software. This includes
storing scientific data from various sources, but visualization does not play a role.
The project that is most similar in spirit to our DGD Gallery is “Electronic
Geometry Models” [13, 14], which is a refereed online journal for digital geome-
try models on the web. It features XML file formats, visualization separated from
descriptions, and a reviewing process. All of these are also implemented in the DGD
Gallery.
However, our technical realization substantially differs from “Electronic Geom-
etry Models”. The DGD Gallery employs modern web technology for the user
interface and a standard data base implementation for storing. One advantage of this
424 M. Joswig et al.

is the possibility to work in teams. Each team member can contribute to a model if
he/she is a registered user with the suitable permissions. Permissions can be granted
by owners of content; for details see Fig. 7 below. Moreover, the entire work flow
from the submission, through reviewing and revising, to the final publication has one
consistent setup through a common front end. Most importantly, the overall design
is highly modularized. For instance, the DGD Gallery features a variety of media
renderers with different visualization strengths to accommodate for heterogeneous
hard- and software environments at the users’ end. This is also relevant for being
able to preserve the data over a long period.
Another difference to “Electronic Geometry Models” is that the DGD Gallery
aims at a broader outreach and therefore seeks to include more models of purely
educational value. This results in a different set of criteria for accepting a model
for publication. Moreover, the DGD Gallery allows for changes to a model after
publication.

3 Examples
In this section we present some selected models from the early contributions to the
DGD Gallery. They are intended as guidelines and inspiration for future models
to be submitted.

3.1 Discrete S–Conical Catenoid and Helicoid

Authors: Alexander Bobenko, Tim Hoffmann, Benno König, and Stefan Sechelmann
(Fig. 1)
http://gallery.discretization.de/models/sc-catenoid

Fig. 1 Screenshot of two media objects contained in the “Discrete S-Conical Catenoid and Heli-
coid” model as presented by a modern web browser. Left Discrete s-conical catenoid. Right Asso-
ciated family animation between s-conical catenoid and its conjugate, the discrete helicoid
DGD Gallery: Storage, Sharing, … 425

This model shows discrete s-conical versions of the catenoid and the helicoid, which
are classical minimal surfaces [15]. The smooth versions are among the first classical
minimal surfaces ever investigated. Their s-conical counterparts are quadrilateral
polyhedral surfaces with the property that at each vertex the adjacent faces are tangent
to a cone of revolution. The theory of these discrete minimal surfaces is closely related
to the theory of orthogonal circle patterns and Koebe polyhedra; see Sect. 3.3 below.
Its features and constructions are similar to the theory of s-isothermic surfaces. A
minimal surface is (Christoffel) dual to its Gauss map. This property is preserved
in the discrete setup, and so discrete minimal surfaces are constructed from Koebe
polyhedra. The associate family of minimal surfaces is contained in the discrete
theory as well.
The model features images of the catenoid and helicoid using representations
with discrete curvature line parameterizations as well as discrete asymptotic line
parameterizations in the associate family. It contains a video with an animation of
the associate family animating the angle parameter. Geometric data is given as OBJ
files and corresponding preview images.

3.2 z a Circle Pattern

Authors: Jan Techter and Jürgen Richter-Gebert (Fig. 2)

http://gallery.discretization.de/models/zalpha_circle_pattern

Fig. 2 Screenshot of the interactive element of the “z a circle pattern” model. A user can adjust the
number of circles in a row as well as the overall scale of the drawing. The angle α is entered by
moving the axes with the mouse
426 M. Joswig et al.

The representation of discrete holomorphic functions by circle patterns with square-

grid combinatorics was first studied by Schramm [16].
This model shows the Schramm type circle pattern corresponding to the holomor-
phic map z → z a for 0 < a < 2 in the first quadrant of the complex plane. Taking
the centers and intersections of the circles as complex fields on the first quadrant of
Z2 , the discrete map was introduced in [17] as a special isomonodromic solution of
the cross-ratio equation (cross-ratio equal to −1 on each elementary quadrilateral).
The numerics of these discrete maps is studied in [18].
The model features an interactive Cinderella [19] application where the user can
adjust the exponent a and the number of circles, see also Sect. 5.3.

3.3 Koebe Polyhedra

Author: Stefan Sechelmann (Fig. 3)

http://gallery.discretization.de/models/koebe_polyhedra

A Koebe polyhedron is a 3-dimensional convex polytope whose edges are tangent to

the unit sphere. Koebe polyhedra have a strong connection to the theory of circle pat-
terns, see [20]. The theory of discrete minimal surfaces of s-isothermic and s-conical
type is based on Koebe polyhedra. Each combinatorial type of 3-polytope admits a
representation as Koebe polyhedron, which is unique up to Möbius transformation.
The first step for the construction of a Koebe polyhedron is to create an orthogonal
circle pattern corresponding to the desired polytopal cell decomposition of the sphere.
This is generally done by finding critical points of a functional expressed in the the
variables ρi = log tan r2i given by the spherical radii ri . Once the radii are known

Fig. 3 Screenshot of two media objects of the “Koebe Polyhedra” model. The two images show
the two corresponding Koebe polyhedra for a given circle pattern
DGD Gallery: Storage, Sharing, … 427

the circles can be layed out. The still remaining freedom of applying a Möbius
transformation can be fixed (up to a simple rotation) by requiring the center of mass
to be at the sphere center. The vertices of the circumscribed Koebe polyhedron that
corresponds to the circle pattern can now easily be found by inverting the euclidean
centers of the circles in S2 (the cone tips are the points polar to the planes containing
the circles). Here we have the freedom to choose one of the two orthogonal families of
circles to become vertices of the Koebe polyhedron, and the other family to become
faces.
The online model features a selection of Koebe polyhedra. Each one with an OBJ
geometry file and a PNG image file.

3.4 Lawson’s Surface Uniformization

Authors: Stefan Sechelmann, Alexander Bobenko, and Boris Springborn (Fig. 4)

http://gallery.discretization.de/models/lawsons_surface_uniformization

Fuchsian uniformizations of the Riemann surface of Lawson’s genus 2 minimal sur-

face in S3 [21] are presented in this model. The results were created in [22] using
the discrete uniformization theory. Three different conformally equivalent represen-
tations of the surface and of the corresponding hyperbolic tilings are presented.
Lawson’s minimal surface in S3 is conformally equivalent to the hyperelliptic
curve μ2 = λ6 − 1. The branch points λ1 , . . . , λ6 are the 6th roots of unity.
An embedding of Lawson’s surface in R3 , see Fig. 4, is obtained via stereographic
projection from S3 [23]. For this surface the hyperelliptic involution of the Riemann

Fig. 4 Left The Lawson surface in R3 , the boundary curves of the fundamental domain of the
uniformizing group in the right picture are shown in red. Blue curves correspond to simple closed
geodesics corresponding to the axes of generators of the group. Right The uniformization of the
Lawson surface in the Poincaré model of hyperbolic space with a canonical fundamental domain
(red) and axes of the hyperbolic generators of the uniformizing group
428 M. Joswig et al.

surface is realized as a rotation by 180◦ . The axis meets the surface in six points,
which are the branch points of the hyperelliptic curve.
The third realization of the Riemann surface is made of squares identified along
suitable edges. The fundamental domain is identified with the two others.
The model features the data of the discrete uniformizations in XML format. It
contains the combinatorial data, the coordinates of the points, and the uniformizing
groups data. PDF vector graphics and PNG images provide 2D renderings of objects
in 3D space.

3.5 Tropical Grassmannian TropGr(2,6)

Authors: Michael Joswig and Benjamin Schröter (Fig. 5)

http://gallery.discretization.de/models/tropical_grassmannian_gr26

Tropical geometry studies piecewise linear images of classical algebraic varieties.

Many interesting properties remain visible in the tropicalization. Additionally, this
method reveals relations between geometry and optimization. One outcome are com-
binatorial algorithms for dealing with classical objects.
The tropical Grassmannian TropGr(d, n) is the tropicalization of the classical
Grassmannian Gr(d, n), defined over some field. It parameterizes the tropical d-
planes in the tropical (n − 1)-torus; see [24, §4.3]. For d = 2 the tropical Grass-
mannian coincides with the corresponding Dressian, which arises as the subfan
of the secondary fan of the hypersimplex (d, n) corresponding to those regular
decompositions whose cells are matroid polytopes [25].

Fig. 5 Screenshots of two images contained in the “Tropical Grassmannian TropGr(2,6)” model
DGD Gallery: Storage, Sharing, … 429

How to properly visualize TropGr(2, 6) is far from obvious, since (modulo its
lineality space and intersected with the corresponding unit sphere) this is a 2-
dimensional spherical simplicial complex naturally embedded in the 8-sphere. It has
25 vertices, 105 edges and 105 triangles. The approach here employs a fixed copy
of smaller tropical Grassmannian TropGr(2, 5), obtained by deletion, as a frame of
reference and uses projections. The deletion of a matroid as a smaller matroid which
is induced on fewer elements, and this notion carries over matroid decompositions.
The media objects associated with this model are a polymake [26] description
and pictures of various projections, in PNG format.

4 Architecture

In this section we describe the structure of a model and the organization of data
within the DGD Gallery. It is also explained how users create, edit, and interact
with models using model permissions. Finally, we give details on the submission
system and the review process.

4.1 What Is a Model?

The architecture of the DGD Gallery is built around the definition of the Model,
see Fig. 6. The teletype font is used to indicate that a word is the name of an
abstract data type, one of its attributes, or an admissible value. From a high level
perspective a model is a collection of files together with a description. The description
contains fields for the title, authors, a description text, keywords, literature references,
and the creation date. The data files associated with a model are bundled into media
objects. A media object is a set of files together with a title and a description text.
These files may be images, videos or data for specific software systems. While some
file formats are more common (and more reasonable) than others, conceptually we
allow for any file format to become part of a media object. In this way our design is
very flexible and thus could be applied in other contexts.
The data type Model has a key, a version number, a status field, and an
edited-by username. The model key is a unique identifier that is used, e.g., to
assemble the permanent link of the model on the web. The version number is assigned
automatically for keeping track of a model’s history. Throughout the following the
word “model” refers both to a specific version and to the entire history of a model. The
standard representative of a model is given by its latest version. The edited-by
field contains the username of the author of a particular model version. Hence, in a
database a model can be uniquely identified by its key, a model version is identified
by its key and version number.
430 M. Joswig et al.

Fig. 6 The structure of a model. A model is a collection of media objects together with a description

The status of a model can take the values edit, pending, rejected, or
approved. See Sect. 4.5 for a detailed description of the model status and the
submission process for models in the gallery.
The model description is a collection of the following information that is provided
by the editor of a model. While this somehow resembles the structure of a traditional
research paper, there are some notable differences.

• The Title of the model.

• A sorted list of Authors. Since this is a frequent source of misunderstanding, it
is worth explaining. The authors of the model are those who create the content
that is presented online. Like for a research paper all the scientific work that leads
to the model must be properly acknowledged in the references. However, clearly,
the set of authors cannot comprise all the authors who contributed something to
the entire history of a mathematical idea. For instance, suppose that Alice first
describes a new type of surface in a traditional research paper, and Bob afterwards
produces a model from Alices description (without Alice’s help). Then Bob is the
only author, who must cite Alice’s paper.
DGD Gallery: Storage, Sharing, … 431

• The description Text, which can contain any valid LATEX source code that can
be compiled using the MathJax library, see [27]. References to the literature or
to media objects can be cited via the \cite{.} command. Previews of media
objects can be included with the \media{.} command.
• A set of keywords can be assigned to the model. The keywords are used on the
web-site to, e.g., improve search features.
• A set of references each of which consitsts of a reference key that is to be used
in \cite commands and a set of key value pairs. The user interface of the
gallery maps BibTE X entries to model references. Conversely a model reference
is rendered and referenced using common BibTE X styles.
• The date field of the description contains the creation date of the particular version
of the model.

A model contains a number of media objects for visualization and use in other
software systems. This concept will be explained below.

4.2 Media Objects and Data Files

A media object is a collection of files that describe the same set of data associated
with the model. For instance, several media objects might correspond to various
views of the same model; e.g., see the tropical Grassmanian in Example in Sect. 3.5.
A different use case for several media objects for the same model is displayed for
the discrete catenoid and helicoid model in Example in Sect. 3.1. One media object
shows a catenoid, whereas the other media object contains a dynamic rendering of
the transformation from the catenoid to the helicoid.
The various file formats for one media object are meant to display one view of the
model on several backends. For instance, the discrete catenoid media object comes
with a PNG file to be displayed in a standard web browser and with an OBJ file which
allows 3-dimensional interactive visualization with a suitable viewer software. The
data files comprising the media objects are stored in the file system separately from
the model database. The media objects of a model contain links to those data files,
see Sect. 4.1.
In principle, we do not restrict the file formats for data files of any media objects.
This makes the DGD Gallery very flexible, but this also creates potential trouble
with file formats that are uncommon. We support the direct visualization of a few
well chosen standard file formats. So far these include the following: PNG and JPG
(for raster image data), SVG and PDF (for vector graphics), OBJ (for 3-dimensional
geometric data), MOV, MP4, and OGV (for video content), POLY (for polymake
data), and others. Interactive content is not excluded, see Sect. 5.3, but the danger
of a particularly low stability over time should be well considered. We rely on the
review process for a sound selection.
432 M. Joswig et al.

4.3 Versioning

The DGD Gallery tracks the history of each model via the version attribute,
see Fig. 6. Editing a model amounts to adding a new version with modified content.
If a model is deleted, all versions of the model and the data files linked are deleted
from the database. A data file is kept in the file system as long as there exists a link
to it from some version of a model.
The version system is particularly useful for models with several authors who can
collaborate through our front end.
It is worth noting that we also allow published models to be edited further and
resubmitted. Upon acceptance this new version will appear as the current version of
the model on the web page. The previously published versions remain visible and can
be compared. This way authors can keep their models up to date; see also Sect. 5.2
below which describes our migration process.

4.4 Users

The users of the DGD Gallery are represented by their usernames, i.e., their
login names on the web-site. The access is password restricted, see also Sect. 5.1.
In addition to the username and password we store the name and email address of
the person that is associated with the user.
A user has a global user-role that can take the values admin, reviewer, or
author. In addition to the global user-role we store model-roles for each model
associated with a user. A user can be the owner or an editor of a model. The read
and write access to models is restricted such that it is based on a combination of the
global user-role, the model-role and the state of the model, Fig. 7. This implementa-
tion allows reviewers to act as model authors but prevents them from approving their
own models.
A notable design decision is that a reviewer can modify a submitted model to
correct obvious typos and other minor changes before approval. Each owner of a
model can invite other users to become either owners or editors of that model.

4.5 Submission Process

The DGD Gallery uses a submission system to publish models on the web-site.
The idea is that a board of reviewers approves, sends back for revision, or rejects a
submitted model. The review process should concern the quality of the content and
address technical issues with the digital data. The review board has to work out and
agree on some quality criteria for a model.
DGD Gallery: Storage, Sharing, … 433

Fig. 7 Read and write permissions during the life-cycle of a model. A user with global admin
privileges can read and write on the model at any state (first row). The author of the model can
edit his model if it is in edit state (second row). A reviewer can edit a model if it has been
submitted (pending state, third row)

During its life-cycle a model has assigned a status value. A newly created model
starts its life in edit state. It can be previewed and edited by the owners of the
model, typically the creator of the model, and any additional user with the editor
model-role, see Sect. 4.4.
A model can be submitted by a user with the owner model-role. The status of
the model changes to pending. A model with pending status is read-only for the
owner and all editors.
Reviewers can preview and edit pending models. A reviewer edits a model to
resolve small issues such as typos. If the quality of a model is sufficiently high then
a reviewer can accept a model. The status of the model is changed to approved.
If the content has flaws or technical issues that can be resolved by the creator of the
model, the reviewer sends the model back to edit state. Any action by a reviewer
is accompanied by a review text, which is presented to the authors of the model.
If a model is sent back for revision, the authors can edit the model according to
the review text and resubmit. If the model is rejected it can neither be edited nor
resubmitted. A model will be rejected if it contains major flaws or its content is not
appropriate for publication in the DGD Gallery.
Approved models become publicly available on the DGD Gallery web-site (see
Sect. 4.6). To further improve public models, e.g. by correcting errors or replacing
outdated file formats, a new version of an approved model can be created, which is
back in edit state. To publish the new version it has to be submitted und undergo
the revision process again.
434 M. Joswig et al.

The model submission system dispatches messages to the users of the DGD
Gallery on every model status change. Reviewers are notified about submitted
models. Model owners/editors are notified upon acceptance, rejection, or call for
revision.
In principle any reviewer can accept, send back, or reject a model. We rely on
a reasonable communication between the reviewers to organize the review process.
Accepting a model is based on formal correctness, technical soundness, mathematical
content and visualization quality.

4.6 Publication and Licensing

Content that has been approved by the board of reviewers is published on the DGD
Gallery web page. The presentation of the content on this page is equivalent to the
preview during edit state of the model. The key defines the permanent absolute
URL of a model:
https://gallery.discretization.de/model/

The content of the DGD Gallery is published under the Creative Commons
Attribution-ShareAlike 4.0 International license, short CC BY-SA 4.0, see [28]. This
means in particular that we allow for our data to be used commercially, enabling
newspapers or commercial web blogs to include content from the gallery without
further complications. Appropriate credit must be given if any content is reproduced
or used, and this includes a link to the DGD Gallery.

5 Implementation

In this section we elaborate on the technical decisions that we made in order to imple-
ment the DGD gallery. It should give an impression of the system architecture,
libraries, frameworks, and languages in use and their respective purposes.
We imposed some a priori constraints on the implementation mainly to ensure
reusability and persistence of the data over time.
• Human readable data format (with enough structure to allow for easy validation
and transformation): We chose XML for storage on the server and as the web server
API data format. It fits the tree-like structure of our data and can be transformed to
anything else, e.g. using XSLT. This allows for easy migrations which can range
from changing the structure of the models to getting rid of XML itself (replacing
it with some more sophisticated data format in the future). To ensure that all
stored data, and in particular data entered by users of the system, agrees with our
specifications we use the XML Schema concept [30]. This allows to validate all
data on insert and during migration, see Sect. 5.2.
DGD Gallery: Storage, Sharing, … 435

Fig. 8 System architecture of the DGD Gallery. On the file system level we store XML model
data and data files. A BaseX server manages the read/write access to the XML documents and data
files. It maintains an instance of a document database to optimize access to the XML data. At the
same time a BaseX servlet provides a REST API [29] to connect the HTML/Javascript web front end
of the gallery. It runs inside an Apache Tomcat servlet container executed within an Apache HTTP
web server. The front end uses AJAX techniques and XSLT to create an interactive application
using the API provided by the application server
 436 M. Joswig et al.

• Database framework agnostic storage (while still using a database): The XML
and any binary data are stored and handled using the XML document database
BaseX [6], see Sect. 5.1. This gives us low access times (for data cached in main
memory) and a transparent mechanism for permanent storage on the server’s file
system. The binary data files of the model’s media objects are stored next to the
XML data and linked appropriately, see the file system section in Fig. 8.
• Separation of data and presentation: We separate our application into a back end
(on the server for database management only) and a front end (creating a HTML
representation on the client machine). The BaseX database already provides the
means for a complete implementation of the back end via XQuery. This includes
the specification of a REST API [29], which is a standard way to define the
communication interface between server and client in the internet. The API returns
XML or binary data in response to specified HTTP requests from the front end.
XML is already close to HTML, while still not carrying explicit information on
the visualization. This allows for the easy generation of multiple presentations
from just one XML. In Sect. 5.3 we elaborate on the front end, which is based an
XSLT, JavaScript and AJAX [31].

5.1 XML Based Backend and the XML Document

Database BaseX

We use the established XML document database BaseX for storing our data. This
automatically provides us with permissions, versioning, and life-cycle management
for the models. BaseX runs on any Java application server. We use Apache Tomcat 7,
see [32].
BaseX allows for the implementation of (web) applications using the XML query
language XQuery, see [33]. It combines the database access and application server
logic implementation into one language. Additionally BaseX can be used to imple-
ment a REST API via RESTXQ, which is a set of XQuery annotations for handling
HTTP requests and generating HTTP responses [34], see Listing 1.
1 (: ~
2 : REST API f u n c t i o n to c r e a t e a new model .
3 :
4 : @ p a r a m $ t i t l e m a p p e d to P O S T p a r a m e t e r title , the title of the
new model
5 : @ p a r a m $user o p t i o n a l user name if no s e s s i o n can be i n f e r r e d by
the s e r v e r
6 : @param $pass optional password
7 :)
8 d e c l a r e % rest : POST
9 % rest : path ( " / c r e a t e m o d e l " )
10 % rest : form - param ( " title " , " { $ t i t l e } " )
11 % rest : form - param ( " user " , " { $user } " , " d u m m y _ i d " )
12 % rest : form - param ( " pass " , " { $pass } " , " " )
13 % o u t p u t : m e t h o d ( " text " )
14 % u p d a t i n g f u n c t i o n api : c r e a t e M o d e l (
15 $ t i t l e as xs : string ,
16 $user as xs : ID ,
 DGD Gallery: Storage, Sharing, … 437

17 $pass as xs : s t r i n g
18 ) {
19 let $user := user : c h e c k U s e r ( $user , $pass )
20 r e t u r n model : c r e a t e M o d e l ( $user , $ t i t l e )
21 };
Listing 1 XQuery with RESTXQ annotations. The function api:createModel defines the web API
function to create a model for a specified user and title string. The RESTXQ annotations, lines 8–13,
define the REST interface of the server. User permissions are checked, line 19, and the corresponding
database function to create a new model is called, line 20. User credentials ($user, $pass) are
optional parameters and are transmitted using HTTPS API calls. Once logged in we use session
cookies to authenticate users.

Generally, all API calls to our back end have to be authenticated. Either a username
and password pair, or a session-cookie has to be provided along each request. The
front end implementation uses session-cookies, which are obtained by an authenti-
cated call to the login API function. A user’s password is stored in the form of a salted
bcrypt hash to provide protection against password recovery through an attacker in
case of a server breach, see [35].

5.2 A Fail-Safe Release and Migration Process

While a project like the DGD Gallery evolves the precise technical requirements
for the database are likely to change. This means that old versions will have to be
migrated into new ones. We implemented a release process for new versions of the
web application and its data using Apache ant [36]. We use XSLT 2.0 and XML
Schema to define and validate database migrations [37].
In principle this allows for more general migrations than just XML to XML
conversions between different schema versions of the database. We can envision
scenarios in the future where XML may turn into a legacy format and will be replaced
by a more general versatile format. With XSLT we can also convert our XML data
into arbitrary text based formats allowing for a final conversion into file formats
entirely different from XML.

5.3 A JavaScript Web Front End

The standard way to enter a new model into the DGD Gallery is through our web
front end. This part of the application is completely separated from the back end,
relying only on the REST API to BaseX for communication.
We use the AJAX scheme of web application development. The application, with
its HTML, XSLT, and JavaScript components, is initially loaded from the server. The
access to the database is organized as HTTP connections via JavaScript. Once XML
model data has arrived from the server we process it with XSL Transformations [37]
438 M. Joswig et al.

to provide dynamic HTML and JavaScript for each client. We use the SaxonCE
XSLT JavaScript framework to execute XSLT 2.0 in the browser, see [38].
Media renderers are provided for several common media formats, see Sect. 4.2. In
the case of images and videos we are relying on the standards built into HTML5. We
have support for the web capabilities of Cinderella to allow for interactive content [19,
39]. For the web browser in particular we use CindyJS [40], an open source JavaScript
variant of Cinderella that aims to be compatible with Cinderella.

Acknowledgments This research was supported by DFG SFB/TRR 109 “Discretization in Geom-
etry and Dynamics”.

Open Access This chapter is distributed under the terms of the Creative Commons Attribution-
Noncommercial 2.5 License (http://creativecommons.org/licenses/by-nc/2.5/) which permits any
noncommercial use, distribution, and reproduction in any medium, provided the original author(s)
and source are credited.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.

References

1. Bates, D.J., Hauenstein, J.D., Sommese, A.J., Wampler, C.W.: Numerically solving polynomial
systems with Bertini, Software, Environments, and Tools, vol. 25. Society for Industrial and
Applied Mathematics (SIAM), Philadelphia, PA (2013)
2. Barakat, M., Lange-Hegermann, M.: Gabriel morphisms and the computability of Serre quo-
tients with applications to coherent sheaves (2014). Preprint arXiv:1409.2028
3. Hong, H., Yap, C. (eds.): Mathematical software–ICMS 2014. Lecture Notes in Computer
Science, vol. 8592. Springer, Heidelberg (2014)
4. Hales, T., Adams, M., Bauer, G., Dang, D.T., Harrison, J., Hoang, T.L., Kaliszyk, C., Magron,
V., McLaughlin, S., Nguyen, T.T., Nguyen, T.Q., Nipkow, T., Obua, S., Pleso, J., Rute, J.,
Solovyev, A., Ta, A.H.T., Tran, T.N., Trieu, D.T., Urban, J., Vu, K.K., Zumkeller, R.: A formal
proof of the Kepler conjecture (2015). Preprint arXiv:1501.02155
5. DFG verabschiedet Leitlinien zum Umgang mit Forschungsdaten. http://www.dfg.de/
foerderung/info_wissenschaft/2015/info_wissenschaft_15_66/
6. BaseX Team: BaseX. The XML database (2014). http://basex.org
7. The Geometry Center—Geometry Reference Archive (2000). http://www.geom.uiuc.edu/docs/
reference/
8. Hoffman, D., Hoffman, J., Weber, M.: The Scientific Graphics Project (1998–2004). http://
www.msri.org/publications/sgp/SGP/
9. Eppstein, D.: The Geometry Junkyard. https://www.ics.uci.edu/~eppstein/junkyard/
10. Bohle, C., Loose, F., Schmitt, N., Heller, S.: GeometrieWerkstatt. https://www.math.uni-
tuebingen.de/ab/GeometrieWerkstatt/
11. Imaginary | open mathematics. http://imaginary.org/
12. The SymbolicData project. http://wiki.symbolicdata.org/
13. EG-Models. http://www.eg-models.de
DGD Gallery: Storage, Sharing, … 439

14. Joswig, M., Polthier, K.: EG-Models—a New Journal for Digital Geometry Models. In: Bor-
wein J., Morales M., Polthier K., Rodrigues J. (eds.) Multimedia Tools for Communicating
Mathematics, pp. 165–190. Springer (2002)
15. Bobenko, A.I., Hoffmann, T., König, B., Sechelmann, S.: S-conical minimal surfaces. Towards
a unifying theory of discrete minimal surfaces (2015)
16. Schramm, O.: Circle patterns with the combinatorics of the square grid. Duke Math. J. 86,
347–389 (1997)
17. Bobenko, A.I.: Discrete conformal maps and surfaces. In: Clarkson P., Nijhof F. (eds.) Sym-
metries and Integrability of Difference Equations, Proceedings of the SIDE II Conference,
Canterbury, July 1–5, 1996, pp. 97–108. Cambridge University Press (1999)
18. Bornemann, F., Its, A., Olver, P., Wechslberger, G.: Numerical methods for the discrete map
z a (in this volume)
19. Cinderella (2013). http://www.cinderella.de
20. Bobenko, A.I., Springborn, B.A.: Variational principles for circle patterns and Koebe’s theorem.
Trans. Amer. Math. Soc 356, 659–689 (2004)
21. Lawson, H.B.J.: Complete minimal surfaces in S3 . Annals of Mathematics 92(3), 335–374
(1970)
22. Bobenko, A.I., Sechelmann, S., Springborn, B.: Discrete conformal maps: Boundary value
problems, circle domains, Fuchsian and Schottky uniformization (In this volume)
23. Oberknapp, B., Polthier, K.: An algorithm for discrete constant mean curvature surfaces. In:
Hege, H.C., Polthier, K. (eds.) Visualization and Mathematics, pp. 141–161. Springer, Berlin
Heidelberg (1997)
24. Maclagan, D., Sturmfels, B.: Introduction to Tropical Geometry, Graduate Studies in Mathe-
matics, vol. 161. American Mathematical Society, Providence, RI (2015)
25. Herrmann, S., Jensen, A., Joswig, M., Sturmfels, B.: How to draw tropical planes. Electronic
J. Combin. 16(2), R6 (2009–2010). http://www.combinatorics.org/ojs/index.php/eljc/article/
view/v16i2r6
26. Gawrilow, E., Joswig, M.: polymake: a framework for analyzing convex polytopes. In:
Polytopes—combinatorics and computation (Oberwolfach, 1997), DMV Sem., vol. 29, pp.
43–73. Birkhäuser, Basel (2000)
27. The MathJax Consortium: Mathjax. https://www.mathjax.org
28. creativecommons.org: CC BY-SA 4.0. http://creativecommons.org/licenses/by-sa/4.0/
29. REST—Representational State Transfer. https://en.wikipedia.org/wiki/Representational_
state_transfer
30. W3C: XML Schema (2004). http://www.w3.org/TR/xmlschema-0/
31. Ajax programming. http://wikipedia.org/wiki/Ajax_(programming)
32. Apache: Apache Tomcat (2015). https://tomcat.apache.org
33. W3C: Xquery 1.0: An xml query language (2010). http://www.w3.org/TR/xquery/
34. BaseX Team: RESTXQ documentation web page. http://docs.basex.org/wiki/RESTXQ
35. Provos, N., Mazières, D.: A future-adaptive password scheme. Proceedings of the Annual
Conference on USENIX Annual Technical Conference. ATEC ’99, pp. 32–32. USENIX Asso-
ciation, Berkeley, CA, USA (1999)
36. Apache: Apache ant (2012). http://ant.apache.org/
37. W3C: XSL Transformations (XSLT) version 2.0 (2007). http://www.w3.org/TR/xslt20/
38. Saxonica: Saxon-CE (Client Edition) (2015). http://www.saxonica.com/ce/index.xml
39. Richter-Gebert, J., Kortenkamp, U.: The Cinderella.2 Manual: Working with The Interactive
Geometry Software. Springer (2012)
40. CindyJS—A JavaScript framework for interactive (mathematical) content. https://github.com/
CindyJS