You are on page 1of 32

Chapter 1

Introduction

1.1 Computer Aided Geometric Design


Computer Aided Geometric Design (CAGD) was coined by Barnhill and
Riesenfeld in 1974 and involves the study of geometric algorithms. The prob-
lems addressed in this field are related to curve and surface modeling. Its
primary objective is to represent real-world objects in forms suitable for com-
putations in CAD/CAM systems. Computer-Aided Design (CAD) involves
the use of computer technology for the design of objects, real and virtual.
CAD is used to design curves in two dimensional space, and curves, surfaces,
and solids in three-dimensional space. CAGD has extensive applications in
automotive, shipbuilding and aerospace industries and more recently in com-
puter animation for special effects in movies.
This thesis considers problems in curve approximation which fall within
the scope of CAGD. Being application oriented in nature, several fields of
mathematics come into play in tackling design problems. Tools from ap-
proximation theory, geometric modeling and differential geometry are used to
solve the curve approximation problems in this thesis. The key goal is the
approximation of curves and the asymptotic analysis of the rates of conver-
gence of different approximation schemes with respect to various metrics. To
measure the error, one usually uses the notion of approximation order. In
this thesis we go a step further and look at the complexity of approximation.
The central idea is to approximate a sufficiently smooth parametric curve
: [0, ] Rm , m = 2, 3, by a class of tangent continuous spline curves S,
such that the Hausdorff distance between and a spline curves S S is given
by

H (, S) = max{sup inf d(x, y), sup inf d(x, y)},


x yS yS x
2 1. Introduction

where d is the distance between two points in Rm . Thus the Hausdorff distance
between and S is given by

H (, S) = inf H (, S) = c n + O( n+1 ),
SS

where is the arc length of the curve and n N denotes the approximation
order of by the class of spline curves S and c is a positive constant of approx-
imation. In our case c depends on the differential invariants like curvature
and affine curvature, cf. Chapter 2. The expression of approximation order is
used to compute the complexity of the approximating spline. Complexity de-
scribes the minimum number of components of an optimal spline (with respect
to Hausdorff distance and symmetric difference distance in our case) approx-
imating a given curve and is used in the same sense as by Fejes Toth [42],
McClure and Vitale [82] and Ludwig [74]. If the approximation order is n, the
complexity of the approximating spline is of the form c1/n + O(1), where
is the error (e.g., Hausdorff distance) of approximation. To find the optimal
spline we rely on differential geometric properties of the approximating spline
curve.

1.1.1 Brief history of CAGD


Curves were used in the manufacturing environment since Roman times, for
the purpose of shipbuilding. A ships ribs - wooden planks emanating from
the keel - were produced based on templates which could be reused many
times [41]. The form of the ribs was defined in terms of tangent continuous
circular arcs. Geometric design found application in aeronautics. In his book
Analytical Geometry with Application to Aircraft [69] R. Liming combined
classical drafting methods with computational techniques for the first time.
Liming worked for the NAA (North American Aviation) during World War II,
a company which built the famous Mustang fighter planes. Conics, based on
the constructions of Pascal and Monge, were used in aircraft and shipbuilding
industries. Traditionally these constructions were used by draftsmen as the
basic product information. Liming translated the classical drafting construc-
tions into numerical algorithms. This was advantageous because numbers can
be stored in unambiguous tables.
1.1. Computer Aided Geometric Design 3

Limings conic constructions were an exception but were not widely avail-
able outside the aircraft industry. Thus the problem of communicating the
information stored on blueprints to computers (driving milling machines) re-
mained. Milling machines were used for production of dies and stamps for
sheet metal parts. In France, de Casteljau and Bezier helped designers to
abandon the blueprint process altogether. In the U.S., J. Ferguson at Boe-
ing and S. Coons at M.I.T. provided alternative techniques. General Motors
developed its first CAD/CAM system DAC-I (Design Augmented by com-
puter). It used the fundamental curve and surface techniques developed at
General Motors by researchers such as C. de Boor and W. Gordon. M. Sabin
worked for British Aircraft Corporation and was instrumental in developing
their CAD system, Numerical Master Geometry. All these developments took
place in the sixties. All these methods existed in isolation for a while, until the
seventies started to see a union of different research approaches, culminating
in the creation of a new discipline, CAGD. Without the advent of computers
a discipline such as CAGD would not have emerged.

In order to solve some of the theoretical problems that arose from the
blueprint-to-computer challenge, the French car company Citroen hired de
Casteljau in the year 1959. He developed a system which was primarily aimed
at design of curves and surfaces instead of focusing on the reproduction of
existing blueprints. From the very beginning he adopted the use of Bernstein
polynomials for his curve and surface definitions, together with what is now
known as the de Casteljau algorithm. W. Bohm was the first to give de
Casteljau recognition for his work in the research community. During the
early sixties, Pierre Bezier headed the design department and also realized
the need for computer representations of mechanical parts. Beziers work in
this area was widely published in [7, 8, 9]. The Renault CAD/CAM system
UNISURF was based entirely on Bezier curves and surfaces [11].

Curves have been employed by draftsmen for centuries. The majority


of these curves were circles and some were free-form. Such curves arose
from applications such as ship hull design and architecture. For drawing
purposes carefully designed wooden curves consisting of pieces of spirals and
conics were used. Another mechanical tool, called a spline was also used.
A mechanical spline was a flexible strip of wood that was held in place and
4 1. Introduction

shape by metal weights, known as ducks. A spline bends as little as possible,


resulting in shapes that are both aesthetically pleasing and physically optimal.
The mathematical counterpart to a mechanical spline is a spline curve, one
of the fundamental parametric curve forms. The differential geometry of
curves was well understood since the late 1800s after the work of Frenet-
Serret, Darboux and others. On the other hand the focus of the research in
numerical analysis and approximation theory was entirely on functions. Both
of these areas were brought together when they became important building
blocks of CAGD [41].

Figure 1.1: Spline with ducks (Photo: Westlawn Institute of Marine Technology,
www.westlawn.edu)

1.2 Curve modeling


Geometric modeling of curves provides an invaluable tool for representation
or visualization, analysis and manufacture of any machine part by providing
the basis for the representation of surfaces and solids and thus of real world
1.2. Curve modeling 5

objects. Curve design is fundamentally significant in Computer Aided Design


because it allows the user to manipulate the shape of a curve locally (at a
small region around a point) without altering it significantly. The way in
which active control on curves shape can be sought, is by choosing a set
of data points and requiring to interpolate a best fitting curve through it.
Curve interpolation and curve fitting methods are two of the oldest methods
available in curve design.
A different problem which is closely related to curve interpolation is the
approximation of a parametric or implicit curve by a simpler curve. Various
curve approximation methods exist to approximate a given set of data points,
of a parametric curve or an implicit curve with respect to some error bound.
The motivation behind most curve approximation techniques is to give an
efficiently computable and numerically stable algorithm which preserves the
shape of the given input. The bisection algorithms presented in this thesis
efficiently computes the spline curve with some limitations on shapes. Non-
etheless these algorithms can handle shapes under certain constraints on the
input curve and are numerically stable (cf. Chapters 3 and 4). In this section
we survey some of the curve approximation techniques, without making any
claims of being exhaustive.

1.2.1 B
ezier curves
Bezier curves are widely used in computer graphics to model smooth curves.
Industrial applications of Bezier techniques are described by Bezier [10] at
Renault, de Casteljau [26] at Citroen, Farin [38] at Daimler-Benz and Hochfeld
and Ahlers [58] at Volkswagen.
The Bernstein basis functions are the building blocks for Bezier curves. A
Bezier curve of degree n is given by

B(t) = ni=0 bi Bin (t), t [0, 1] (1.1)

where Bin (t) = ni (1 t)ni ti , and b0 , , bn are control points. For n = 2




we get a quadratic Bezier curve which is a parabolic segment. The polygon


formed by connecting the Bezier points with line segments, starting with b0
and finishing with bn , is called a control polygon. The convex hull of the
6 1. Introduction

control polygon contains the Bezier curve. De Casteljau devised a recurs-


ive method to evaluate polynomials in Bernstein form or Bezier curves. The
de Casteljaus algorithm can be used to split a single Bezier curve into two
Bezier curves at an arbitrary parameter value. The key idea of de Castel-
jaus algorithm is that the control polygons generated by recursive subdivison
converge to the original Bezier curve. From equation (1.1) the following prop-
erties of Bezier curves are inferred:
- A Bezier curve passes through the control polygon at the endpoints b0 and
bn and hence the curve interpolates the endpoints.
- The Bezier curve generated by an affine map applied to the control poly-
gon is the same as the curve under the same map depending on the original
control polygon. More precisely,

ni=0 (bi ) Bin (t) = (ni=0 bi Bin (t)),

thus we conclude that the curve is invariant under affine maps. We also infer
that Bezier curves are independent of the choice of the origin of the coordinate
system. Other than these two properties relevant in the context of this thesis
the Bezier curves also have the symmetry property, convex hull property and
the variation diminishing property. For more details of these properties we
refer to Farin [40].
Bezier curves provide a powerful tool in curve design, but they have some
limitations. If the curve to be modeled has a complex shape, then its Bezier
representation will have a very high degree (for practical purposes, degrees
exceeding 10 are considered high [40]), in such cases the curves are modeled
using composite Bezier curves. Such piecewise polynomial curves are an ex-
ample of spline curves. In describing a shape using free-form curves, it is
common to use several curve segments which are joined together with some
degree of continuity. There are two types of continuity that can be imposed
on two adjacent Bezier curves: parametric continuity and geometric continu-
ity. In general, two curves which are parametrically continuous to a certain
degree are also geometrically continuous to that same degree, but the converse
is not true. Parametric spline curves are typically constructed so that the first
n parametric derivatives are the same where the curve segments join. This
type of continuity condition is called C n or n-th order parametric continuity.
E.g., C 0 curves are joined, C 1 curves are C 0 and tangent vectors at the join
1.2. Curve modeling 7

are equal. The concept of geometric continuity was an early attempt at de-
scribing, through geometry, the concept of continuity as expressed through a
parametric function. In general Gn continuity exists if a curve can be repara-
metrized to have C n continuity [3]. E.g., G0 continuity implies two curves
meet at a point, G1 implies G0 and that the tangent vectors of the two curves
at their junction point are proportional. In this thesis the concept of G1
continuity, also known as tangent continuity, is used in curve approximation
(cf. Chapters 3 and 4). Modern imaging systems like Postscript, Asymptote
and Metafont use Bezier splines composed of cubic Bezier curves for drawing
curved shapes and allow for curvature (G2 ) continuity.

B-spline curves. A B-spline curve is a generalization of a Bezier curve,


where any B-spline curve is parametrized by spline functions which are linear
combinations of B-splines. B-splines provide a different approach to repres-
enting piecewise polynomial curves. In 1946, I. J. Schoenberg [106] coined
the term B-spline which is short for basis spline. He used B-splines for stat-
istical data smoothing, and his paper started the modern theory of spline
approximation. B-splines were used in industry after de Boor published his
work [24] in the early 1970s. At around the same time Cox published similar
results in [21]. Cox and de Boor found separately a recurrence relation to
compute B-splines and showed that this is a perfect tool for practical and
stable evaluation of B-splines.

1.2.2 Rational curves


Rational techniques and representations were at the root of geometric model-
ing. Curves like conics and circular arcs can be expressed exactly as rational
Bezier curves, cf. Lee [67]. Conic sections and quadric surfaces were initial
building blocks of early CAD systems. In [69, 70] Liming gives many geomet-
ric constructions for aircrafts using conics. Later Coons at Ford introduced
conics into a CAD environment. Independently, conics were used by engineers
at Boeing.
A rational Bezier curve of degree m is a parametric curve which is de-
scribed by control points, bi Rn , n = 2, 3, weights wi and the parameter
t. Without loss of generality we let t vary from 0 to 1. The curve has the
8 1. Introduction

form
m m
i=0 wi bi Bi (t)
c(t) = , (1.2)
m m
i=0 wi Bi (t)

where Bim are the Bernstein Bezier basis functions. If we set all the weights
wi to 1 we obtain a polynomial Bezier curve. The rational Bezier curves
inherit affine invariance, the convex hull property, endpoint interpolation and
the variation diminishing property from their polynomial counterpart. One
important property that rational Bezier curves have but do not share with
polynomial Bezier curves is projective invariance, which means that a rational
Bezier curve is invariant under general projective transformations. A conic
section can be represented as a rational quadratic Bezier curve, and when
w0 , w2 6= 0, we can assume the two weights w0 and w2 to be 1. For a proof
cf. [40, Page 221]. Hence, a conic section is given by

b0 B02 (t) + w1 b1 B12 (t) + b2 B22 (t)


c(t) = ,
B02 (t) + w1 B12 (t) + B22 (t)

with t [0, 1].


Just like in the Bezier curves case we can generalize polynomial B-spline
curves to rational B-spline curves. This generalization defines non-uniform
rational B-splines (NURBS) which are the current industry standard. The
NURBS curves also display affine invariance, the convex hull property, end-
point interpolation and the variation diminishing property. One of the prob-
lems of interpolation with rational curves is that they are very sensitive to
their parametrization as compared to the polynomial spline curves. Approx-
imation with conics is an example of rational Bezier curve approximation.
The most straightforward approach is to fit conics locally through five data
points each and then piece together these conic segments. In general the res-
ulting conic spline curve in this case will only be C 0 continuous. To get a
smooth conic spline curve one prescribes, three data points p0 , p1 , p2 and
two end tangents q0 and q1 . Chapter 3 analyzes the complexity of tangent
continuous conic spline approximation of a planar curve. Osculatory interpol-
ation is another practical interpolation method in which one pieces together
locally fitted arcs with at least G2 continuity. For an example of osculat-
ory interpolation suppose we are given data points pi , unit tangents qi and
1.2. Curve modeling 9

Euclidean curvature values i . We can now fit a rational cubic Bezier curve
segments ci such that ci (0) = pi , ci (1) = pi+1 , the unit tangents c0i (0) = qi ,
c0i (1) = qi+1 and curvatures i . Hollig proved in [60] that such a curve exists
under the restrictions on the geometry induced by the data, where the data
corresponds to a sufficiently smooth curve with non-vanishing curvature and
torsion in space. Furthermore, in [60] he gives a method to compute such a
curvature continuous cubic rational spline. In [104] Shaback shows that, given
an odd number of points in general position in R2 , if they can be interpol-
ated at all by a smooth curve of non-vanishing curvature then this would be a
unique G2 interpolant consisting of pieces of conics under some restrictions on
the data. This method of osculatory interpolation leads us to a more general
method of approximation known as Geometric Hermite Interpolation, which
we describe in Section 1.3.3.

1.2.3 Implicit curve modeling


An overview of various curve modeling methods would be incomplete without
considering modeling of implicit curves. Implicit objects have gained increas-
ing importance in geometric modeling, visualization, animation and computer
graphics due to their nice geometric properties which give some advantages
over traditional modeling methods. The bisection algorithms presented in this
thesis give good approximation of a given parametric curve where, the theoret-
ical and algorithmic results on the number of optimal patches, match exactly
in the parabolic and conic case. But what these algorithms do not consider is a
topologically correct approximation of a given curve. The primary challenge in
implicit curve modeling is to develop rigorous, robust and tractable methods
for guaranteeing that computational approximations of smooth curves and
surfaces preserve critical topological characteristics. Implicit planar curves
are defined as the zero set of a function F : R2 R. The approxima-
tion of implicit planar curves by line segments is a very classical problem.
The approximation schemes for implicit curves can be roughly divided in
continuation methods and adaptive enumeration techniques. Plantinga and
Vegter [96] present an algorithm which uses interval arithmetic to determ-
ine the global properties of the implicit function. They use these properties
to generate a piecewise linear approximation of an implicit curve or surface
10 1. Introduction

which is isotopic to the object itself. For curves and surfaces this paper gives
one of the first practical algorithms that guarantees correct topology. Vacav-
ant, Coeurjolly and Tougne [111] use tools from discrete geometry to build an
original geometrical and topological representation of the implicit curve. We
refer to the thesis of Plantinga [95] for algorithms on isotopic approximation
of implicit curves and surfaces and for detailed survey of various methods in
implicit curve and surface approximation we refer to [52].

1.3 Approximation Theory


The theory of spline functions and their applications is an area of active
research in Approximation Theory since the 1900s. The rapid development of
spline functions is due primarily to their great usefulness in applications. Since
they are easy to store, evaluate and manipulate on a computer, a plethora
of applications have been found. These include, for example, data fitting,
function approximation, numerical quadrature and the numerical solution of
operator equations, integral equations, optimal control problems, to name a
few. As a result of these innumerable applications spline theory has become
a firmly entrenched part of Approximation Theory and Numerical Analysis.
For a detailed review of approximation theory including the theory of splines
we refer to [23, 99, 108, 107].

Approximation problem. One of the basic problems in Approximation


Theory can be stated in generic form as follows: Given some element g in a
metric space X (with metric d), find a best approximation m to g from some
given subset M of X, i.e., find

m M s.t. d(g, m ) = inf d(g, m) =: dist(g, M ) (1.3)


mM

and we consider the set of all such elements m . The basic problems that arise
are existence, uniqueness and characterization of a best approximation m .
The best approximation is defined implicitly by (1.3). A characterization is an
explicit description, that can be turned into an algorithm for the computation
of this best approximations. See Chapters 3, 4 and 5 in this thesis, where we
1.3. Approximation Theory 11

solve this approximation problem for curves in the plane with sets of biarcs,
parabolic arcs and conic arcs and for curves in space with bihelical arcs.

Figure 1.2: General spline element(blue) of a given curve (red)

Figure 1.2 illustrates the approximation of a curve by a spline element.


The important general steps (taken in the conic case) for approximating a
curve with a spline element are given below. In the conic case, X is the space
of all arcs of planar curves with monotone, non-vanishing Euclidean curvature,
M is the space of all bitangent conic arcs. Furthermore, d in this case is
the Hausdorff distance, between two curves. Since the curves are compact
subsets of R2 , the Hausdorff distance in this case is a metric. We also consider
the case when the distance function is the symmetric difference distance, for
the definition of symmetric difference distance, we refer to Section 3.1 in
Chapter 3.
An offset curve to a planar parametric curve , defined over interval
I, with arc length parameter s, cf. Chapter 2, is given by

(s) = (s) + f (s)N (s).

Here N (s) denotes the unit normal to at (s) (cf. Chapter 2).

The distance function f is used to compute the Hausdorff distance


H (, ) = ||f || = maxsI |f (s)|.

If and have n + 1 intersections, such that (si ) = (si ), where


s0 s1 sn , then the distance function is expressed in terms of
a divided-difference formula, cf. Chapter 3

f (s) = (s s0 ) (s sn )[s0 , , sn , s]f,


12 1. Introduction

where [s0 , , sn ]f is the nth divided difference and is defined as the


coefficient of sn in the polynomial of degree n that interpolates f at
s0 , , sn . For more information we refer to Appendix A and the ref-
erences in it.

Furthermore, denoting sn s0 by and using the Hermite-Genocchi


identity as explained in Appendix A, the function f has the following
form
(n+1)
f (s) = 1
(n+1)! f (s0 )(s s0 ) (s sn ) + O( n+2 ).

One of the goals of this thesis is to express f (n+1) (s0 ) in terms of differen-
tial invariants of the curve, like (affine) arc length, (affine) curvature and its
derivatives.

1.3.1 Approximation order and complexity


The approximation order is characterized by the properties of the approxim-
ation scheme, the metric of approximation as well as by a definite property
of the approximated object. In numerical analysis, the approximation order
of a numerical method having error O(hn ), where h is the step-size,( i.e., the
size h in which an interval is subdivided) is the exponent n. Thus, e.g., given
a real valued function f over an interval [a, b], a function g is an n-th order
approximation of f if and only if the interval [a, b] can be subdivided into
intervals of size h, such that as h 0,

sup |f (x) g(x)| C hn ,


axb

where C is a positive constant. We then say that g approximates f with an


error O(hn ) and n is the approximation order. Note that n also depends on
the approximation scheme. In this thesis we make explicit the constant C of
approximation.
An associated notion is that of complexity. In this thesis complexity refers
to space complexity, i.e., the number of elements in a spline of a certain type
the approximates the curve to within distance . We express this space com-
plexity in terms of and differential invariants of the curve. Another notion
1.3. Approximation Theory 13

of complexity is that of time complexity. The paper by Drysdale, Rote and


Sturm [34] is an example where curve approximation algorithms with respect
to time complexity are addressed. They consider approximation of polygonal
curve with a minimum number of circular arcs and biarcs. Their algorithm
takes O(n2 log n) time to compute a series of circular arcs to approximate a
polygonal line with n vertices. We discuss the relevance of their and similar
results in greater detail in applications of our work in Section 1.3.3. In gen-
eral, an algorithm with better time complexity implies a faster algorithm, but
apart from this its space complexity is also important, this is essentially the
number of memory cells which an algorithm needs. A good algorithm keeps
this number as small as possible. There is often a time-space-tradeoff involved
in a problem, that is, it cannot be solved with less computing time and low
memory consumption. One then has to make a compromise and to exchange
computing time for memory consumption or vice-versa, depending on which
algorithm one chooses. Although, in this thesis we focus on algorithms which
look at complexity in terms of space, a possible extension of our work is to
use the theoretical results on complexity to determine the time complexity
of algorithms. We discuss this probable application of our work in greater
detail in the context of geometric Hermite interpolation and refer reader to
Section 1.3.3.

1.3.2 Curve interpolation


Curve interpolation, as already discussed earlier is the constructive (possibly
approximate) recovery of a curve of a certain class by its known values, or
by known values of its derivatives, at given points. In this section we briefly
describe some of the existing curve interpolation methods. In Section 1.3.3
a more general interpolation scheme, Geometric Hermite interpolation is dis-
cussed in some detail as the curve approximation problems of this thesis lie
within the scope of geometric Hermite interpolation. Interpolation of curves
are described by points which either come from some mathematical construc-
tion, or are obtained as data using some mechanical device such as a digitizer.

Curve interpolation using polynomials. The curve interpolation prob-


lem, using polynomials of degree n can be solved uniquely, if we are given n+1
14 1. Introduction

pairwise distinct points Pi R3 , i = 0, 1, , n associated with appropriately


selected parameter values ti . Therefore depending on the basis of Pn , where
Pn is the space of polynomials of degree at most n, one gets,

interpolation using monomials. In which case the polynomial basis


is given by {1, t, , tn }.

Interpolation with Lagrange polynomials. In this case the poly-


nomial basis is given by {L0 , L1 , , Ln }, where
(
1, if i = k,
Li (tk ) = ik :=
0, if i 6= k

where ik is the Kronecker delta. The condition assumes that Li has


value 1 at t = ti and vanishes at other parameter values.

Interpolation with Newton polynomials. The Newton polyno-


mial basis is given by {N0 , N1 , , Nn }, where Ni (t) = (t t0 ) (t
t1 ) (t ti1 ), and N0 (t) = 1.

1.3.3 Geometric Hermite interpolation


Geometric Hermite Interpolation (GHI) is a generalization of different approx-
imation schemes that have been developed in the past for parametric curves
to be approximated by polynomial or rational curves. The central idea is to
consider the curve independent of its actual parametrization. In general a
curve in this situation is not approximated by a single polynomial or rational
curve, but by a geometric spline.
There are two analogous problems that one considers in this case.

The first problem involves the computation of complexity of the best


approximating spline curve, given an error bound w.r.t. a metric
and a parametric curve. Furthermore, a part of this problem is to give
an algorithm to compute this spline and compare the theoretical and
algorithmic results. This is the problem of curve approximation that is
considered in this thesis.
1.3. Approximation Theory 15

The second analogous problem is to distribute knots over a given para-


metric curve, given a knot count n such that the approximating spline
is the optimal one, i.e., one would like to distribute n knots such that
the error is minimized.
A knot is a point on the parametric curve where two adjoining spline elements
meet with prescribed conditions of continuity. Although the second part of
the problem is not considered in this thesis, we can nonetheless reduce it to
the first problem by using our theoretical results. Given knot count n we
can compute an asymptotically optimal error bound using our theoretical
results and hence generate the spline with n knots.
In fact work on approximation of convex discs with polytopes and its
asymptotic error analysis was done in late 1940s. Fejes Toth [42] gave asymp-
totic formulae for the distance between a smooth convex disc and its best
approximating inscribed or circumscribed polygons with at most n vertices
( as n ), where the distance is in the sense of the Hausdorff metric.
In our case, given a regular curve : [a, b] R2 , let d denote the distance
function between and a conic C which is tangent to at (a) and (b).
The distance function d is given by

C(s) = (s) + d(s) N (s),

where s [a, b] and N is the unit normal vector to the curve . Thus C(s) is a
point on the conic C which is the intersection of the normal line of at (s),
and d(s) denotes the distance of the point C(s) from (s) along the normal
N (s). The Hausdorff distance between and C is thus given by the maximum
of the function d attained in the interval [a, b] and is denoted by H (, C). In
fact it can be easily verified that the Hausdorff distance is attained at a point
s [a, b] where the tangent to at (s) and tangent to conic C at C(s) are
parallel. Let Pni denote the set of all inscribed polygons of the plane convex
curve C and let Pnc denote the set of all circumscribed polygons of C, then
the error with respect to the Hausdorff distance is given by
Z l
1
i c
H (C, Pn ) = H (C, Pn ) = 8 ( 1/2 (s) ds)2 2 + O( n14 ),
1
0 n
here is the Euclidean curvature of the curve C and s its arc length parameter
and l is the arc length. Furthermore in [42], Fejes Toth computes the error
16 1. Introduction

of approximation with respect to the symmetric difference distance :


1
R l 1/3 3
S (C, Pni ) = 12 0 (t) dt n12 + O( n14 )
R l 1/3 3 (1.4)
1
S (C, Pnc ) = 24 0 (t) dt n12 + O( n14 )

Ludwig in [74] extends the result by Fejes Toth in [42] for approximation
of plane curves with polygons by including second terms of the asymptotic
expansions in (1.4).

1 4
Z
i 1
S (C, Pn ) = k(s) ds n14 + O( n14 ),
12 n2 2 5! 0
where is affine arc length and k is affine curvature, cf. Chapter 2. McClure
and Vitale give sharp estimates of the order of convergence of best approx-
imations of a convex curve by circumscribed and inscribed polygons with n
edges. They compute these results with respect to Hausdorff distance, area
and perimeter metrics. They first obtain the asymptotic characterizations
of best approximations and use this to give a construction of a distribution
function which then is utilized in distributing n knots optimally on a given
convex curve.
The case of asymptotic error analysis of approximation of convex bodies in
d-space by inscribed and circumscribed polytopes with respect to the symmet-
ric difference metric is considered by Gruber in [53, 54, 55]. Given a convex
body C, Pni = Pni (C), (n = d + 1, d + 2, ) denotes the family of convex
polytopes P having at most n vertices inscribed into C, then the asymptotic
error expansion with respect to symmetric difference distance metric is given
by
c
S (C, Pni ) = + O(1) as n .
n2/(d1)
The error with respect to symmetric difference metric for n vertices is thus
given by
Z
i 1
S (C, Pn (C)) = 43 ( K 1/4 ds)2 n1 + O(1),
C

as n , where C denotes the boundary of C and K is the Gaussian


curvature. In 3-d, ( C K 1/4 ds)2 12V (C), where V (.) denotes the volume.
R
1.3. Approximation Theory 17

Schneider in [105] shows that for a convex C 3 hypersurface in Rd the minimal


number of vertices of an inscribed, -approximating polytope with respect to
Hausdorff distance is
Z
(1d)/2
Nmin () = Cd KdF + O(1),
F

where Cd is a constant which depends only on dimension, and K is the Gaus-


sian curvature of F . For a definition of Gaussian curvature cf. [30]. Note
that for the case, d = 2, the above result reduces to the case of polygonal
approximation of plane curves and coincides with Fejes Toths result in [42]
Geometric design saw the advent of the theory of GHI after some ele-
mentary attempts with quadratic or cubic curves - with the famous paper of
de Boor, H olling and Sabin [25] in 1987. In [25] the scheme of parametric
cubic spline interpolation for planar curves is considered where in addition to
position and tangent, the curvature is prescribed at each knot. This ensures
that the resulting interpolating piecewise cubic curve is C 2 with respect to arc
length and can be constructed locally. Moreover, under appropriate assump-
tions, the interpolant preserves convexity and the order of approximation is 6.
Meek and Walton in [84, 83] present an arbitrarily close arc spline approxim-
ation of a smooth curves in the geometric Hermite sense. They show that, for
a particular choice of a biarc, the asymptotic error between biarc and spiral
1
is given by 324 0 h3 + O(h4 ), where 0 is derivative with respect to arc length
of the Euclidean curvature of the given spiral curve and h is the maximum
length of the parameter intervals. The error of approximation in this case is
of order 3. A biarc is a pair of circular arcs joined with a continuous unit
tangent. A biarc can join two points and match the unit tangents at those
two points. However, Meek and Walton do not prove the optimal error of
approximation of planar curves with biarcs. In Chapter 4 we consider the
approximation of spiral arcs with biarcs and give a proof for the optimal er-
ror of approximation of such an arc with biarcs. Our expression for the error
matches the expression given by Meek and Walton in [83]. A spiral curve in
the plane is a curve which has monotonically increasing/decreasing Euclidean
curvature. In [84] Meek and Walton give new method for construction of
planar osculating arc splines, they are G1 arcs called triarcs. A triarc is a set
of three circular arcs joined with a continuous unit tangent. A triarc can join
18 1. Introduction

a pair of points, matching unit tangents and curvatures at the two points. The
triarc spline is a more controlled curve, than arc splines
produced by biarcs
1 2 33 3/2 0 3
h + O(h4 ).

and the asymptotic error of approximation is 24 3
The paper by Degen [27] deals with approximation of parametric curves. As
discussed earlier on in this section, in [25] deBoor, Hollig and Sabin consider
approximation of planar curves with curvature continuous cubic polynomial
spline curve. Degen in [27] generalizes their work to consider rational cubic
spline approximation of plane curves and imposes third order contact at the
end points, he shows that this raises the approximation order to 8.
In 1995 H ollig and Koch [59] introduced a general conjecture for GHI
of smooth curves in Rn . They conjecture that, under suitable assumptions,
splines of degree d can interpolate points on a smooth curve in Rn with
approximation order
d1
m = d + 1 + b n1 c.

A lot of work has been done on this conjecture, but only special cases of
it have been solved till date. A. Rababah was one of the first authors to
compare GHI approximants with Taylor polynomials in [100]. He showed, for
an arbitrary degree n, there exists a polynomial one-point approximant having
approximation order b4n/3c. Other special case proofs of the conjecture by
Hollig and Koch can be found in the works of M. Floater. Floater [43, 44]
obtained the optimal approximation order O(h2n ) approximating conics with
polynomial curves of degree n. The approximation of circles with polynomial
curves of degree n was studied earlier by Dokken, et al. in [32]. In a recent
paper [45] in 2006, Floater gave a construction of a polynomial curve of degree
at most n + k 2 for rational curves of degree k and he showed that the order
of approximation was 2n. A. Lin and M. Walker treat Hermite interpolation
of space curves in [71]. They give a construction of a degree 5 Bezier curve
which interpolates given positional, tangent, curvature and torsion data at the
endpoints and show that the order of approximation is O(h8 ). Note that A.
Lin and M. Walkers result satisfies the conjecture of Hollig and Koch in [59]
for R3 . In fact the Hollig-Koch conjecture is also satisfied by parabolic arcs
and proved in Chapter 3. Parabolic arcs are polynomial curves of degree 2 in
R2 and approximate a planar curve with approximation order 4. Furthermore,
the cubic spline approximation of plane curves, which approximates a plane
1.3. Approximation Theory 19

curve with approximation order 6 as shown in [25] is also a special case of


Hollig-Koch conjecture.

Open Problems. A recent paper by Degen [28], published in 2005 provides


an interesting survey of results in geometric Hermite interpolation using tools
from differential geometry. Moreover, this paper is a source of current open
problems in this area. Besides the special case of the Hollig-Koch conjecture,
which states that in planar case there exists a polynomial approximant of
degree n having approximation order 2n, there are few more accessible open
problems in this field.

1. Find explicit error bounds for instance for the method of deBoor, Hollig
and Sabin [25]. In particular, if the method admits several solutions:
derive a criterion for the one which yields the lowest error.

2. Are GHI approximants shape preserving, i.e., having the same signs
of curvature, not more inflection points and not more vertices than the
curve to be approximated? - In case they are not, derive additional
conditions to guarantee that property.

3. Compare results of time and space complexity algorithms for curve ap-
proximations. Furthermore, investigate the question whether it is better
to use low degree together with a finer subdivision of the whole curve or
to take higher degrees of the approximant and less segments, taking into
account existence problems, stability of algorithms etc. and comparing
cost of computing under the same tolerance requirements.

Applications for curve approximation. Among the various applica-


tions of the complexity results of geometric Hermite interpolation of different
schemes some are briefly outlined here. In [34] Drysdale, Rote and Sturm
present an algorithm for approximating a polygonal curve with G0 -spline
consisting of a minimum number of circular arcs within a tolerance region
. They also present a similar algorithm to approximate a polygonal curve
with a series of tangent continuous biarcs given points and tangent directions,
with a runtime of O(n log n). Similarly, in [57] Held and Eibl present an al-
gorithm for approximating a simple planar polygon by tangent continuous
20 1. Introduction

biarc spline. Their experimental results demonstrate that they compute close
to a minimum number of biarcs of an n vertex polygon is roughly O(n log n)
time. This kind of algorithms find applications in fields like robot motion
planning and computer aided manufacturing environments. To assess the
quality of this kind of algorithms theoretical bounds are necessary and hence
the complexity results described previously in this section and the ones ob-
tained in this thesis are useful. Furthermore, the expression for the Hausdorff
distance can be used to develop algorithms which are useful in distributing
n points optimally on a curve. This kind of algorithms gives time complex-
ity results which can be helpful in comparing the time complexity results of
algorithms obtained in [34, 57].
The complexity results of this thesis may be of use in applications like
determining the intersection of high-degree curves and the building of ar-
rangements of algebraic curves. For example, in [5] Berberich et al. give an
exact geometry kernel for conic arcs and present a sweep line algorithm for
computing arrangements of curved arcs and in [36, 37] Emiris et al. investigate
the Voronoi diagrams of ellipses. A possible application of our work in such
algorithms is to compute conic arc approximations, of a given arrangements of
curves. Although certified approximation of curves in the plane by conic arcs
is still an open problem, by using the methods suggested in this thesis it is
possible to find sharper bounds between a curve and its approximating conic.
This in turn might be useful in robust computation of geometric structures of
curved objects, like its Voronoi diagram, or its medial axis. In this approach
the number of elements of a conic spline approximation would be orders of
magnitude smaller than the number of line segments needed to approximate
a curved object with the same accuracy. Whether this feature outweighs the
added complexity of the geometric primitives in the computation of Voronoi
diagrams would have to be the goal of extensive experiments.

1.4 Classifications in singularity theory


Classifying objects in mathematics is a fundamental activity. Each branch
has its own notion of equivalence and it is equally natural to list the objects
in respective branches up to the equivalence under study. The final part of
1.4. Classifications in singularity theory 21

this thesis lies in the area of classifications of singularities of secant maps.


In this section the background and motivation for studying classifications in
singularity theory are provided, following which we give an overview of the
classification problem considered in this thesis. Singularity theory provides
some useful tools for the study of the local geometry of curves and surfaces, i.e.,
study of local properties of a single point or of a suitably small neighborhood of
a point. It is a rigorous body of mathematics which arose from the interactions
between topology, algebraic geometry and differential geometry. Both these
fields owe their origins to theory of smooth functions of one variable. For
example by using methods from differential calculus we can recognize the
shape of graphs of smooth functions. The following algorithm helps us to
recognize the shape of the graph.
- Given a function f : I R, compute f 0 (x) and find a point x0 such that
f 0 (x0 ) = 0.
- Calculate the second derivative f 00 (x), then the point x0 is a minimum if
f 00 (x0 ) > 0 and is a maximum if f 00 (x0 ) < 0.
Then x0 is called a critical point or a singular point of f if f 0 (x0 ) = 0. The
shape of the graph y = f (x) changes drastically around the singular point.
Therefore the singular points of f (x) gives crucial information on the shape of
the graph y = f (x) (locally, i.e., close to x = x0 ). Moreover x0 is a degenerate
singular point if f 0 (x0 ) = f 00 (x0 ) = 0. Thus the nondegenerate critical points
are the maximal or minimal points. We say that x0 is an Ak -type singular
point if f 0 (x0 ) = f 00 (x0 ) = = f (k) (x0 ) = 0 and f (k+1) (x0 ) 6= 0. A major
mathematical source of singularity theory is the classification of the types of
critical points or singularities.

1.4.1 Background
Let f : Rn R be a smooth function. A point u Rn is a critical point of f
if
f f
x1 |u = = xn |u = 0.

The value f (u) at a critical point u is called a critical value of f . Critical


points occur when the graph of f has a horizontal tangent. We have already
mentioned briefly the cases for n = 1, for n = 2 there are more possibilities.
22 1. Introduction

The most common are (local) maxima, minima and saddles. However, there
are a wide variety of more complicated types e.g., the monkey-saddle and fold.
An important distinction between the different critical points is the following.

Saddle, nondegenerate Pig Trough, degenerate Monkey Saddle, degener-


corank 1 ate corank 2

Figure 1.3: Some singularities in R2 .

We say that f has a nondegenerate critical point at u if Df |u=0 = 0 and if


D2 f |u is a nondegenerate quadratic form i.e., the Hessian matrix
2f 
Hf (u) = (u)
xi xj
is non-singular. Note that saddle (given by x2 y 2 ) has a nondegenerate
critical point, the pig trough (given by x2 + y 2 ) is a degenerate critical point
of corank 1, whereas the monkey saddle (given by x3 3xy 2 ) has a degenerate
critical point of corank 2. Thus we conclude that for a real valued function f a
singular point x0 Rn is nondegenerate if det Hf (x0 ) 6= 0. The Morse lemma
(cf. Milnor [87]) is useful in classifying the nondegenerate critical points of
a real valued function. Before stating the lemma formally, we introduce the
notion of germs for convenience, it is a term signifying a localization of various
mathematical objects e.g., germs of functions, germs of mappings, germs of
analytic sets, etc.
Defintion 1.4.1. Given a point x of a topological space X, and two maps
f, g : X Y , then f and g define the same germ at x if there is a neighborhood
U of x such that restricted to U , f and g are equal i.e., f |U = g|U . This defines
an equivalence relation and a germ is an equivalence class.
1.4. Classifications in singularity theory 23

Then we have the following result

Lemma 1.4.2 (Morse lemma). Let 0 Rn be a nondegenerate singular point


of f (x1 , xn ). Then there exists a local diffeomorphism (germ) : (Rn , 0)
(Rn , 0) such that

f (x1 , , xn ) = x21 x2n + f (0).

By the Morse lemma we can recognize the shape of the graph y =


f (x1 , xn ) around a nondegenerate singular point very well. On the other
hand, if the Hessian Hf of f is singular then we get a degenerate singular
point, the natural question to ask is how does one study degenerate singu-
lar points. We can measure how critical a degenerate singular point is by
computing its corank

corank (f)(x0 ) = n rank Hf(x0 ).

Therefore, a singular point x0 of f is nondegenerate if and only if


corank (f)(x0 ) = 0. The Morse lemma is a classification theorem of smooth
functions near a corank zero singular point. Another version of the Morse
lemma allows us to tidy up somewhat, a degenerate critical point, by split-
ting the function into a Morse piece over one set of variables and a degenerate
piece on a different set, whose number is equal to the corank.

Lemma 1.4.3 (Thoms splitting lemma (cf. [78, 97])). Let 0 Rn be a


singular point of f : (Rn , 0) R with corank (f)(0) = r. Then there exists
a local diffeomorphism : (Rn , 0) (Rn , 0) and a smooth local function
g : (Rr , 0) R at 0 such that

f (x1 , , xn ) = g(x1 , , xr ) x2r+1 x2n + f (0)

and rank H(g)(0) = 0.

We call g a residual singularity of f . By the splitting lemma the behavior


of a function near a degenerate singular point can be found by studying its
residual singularity.
24 1. Introduction

Stability. The notion of stability of different singular points is considered,


where the effect of a small deformation on a critical point is studied. A
deformation p is considered to be small if the function and all its partial
derivatives are small for all x in a neighborhood of 0. Assuming that the
derivative of p, Dp|0 = 0, suppose f is Morse, i.e., the critical point at 0 is
nondegenerate, then for a small enough (here C 2 small is sufficient) p, f + p
is also Morse,

If det Hf |0 6= 0 then det H(f + p)|0 6= 0.

Then f is stable and for all sufficiently small smooth functions p, the critical
points of f and f + p have the same type. For degenerate critical points
the situation is completely different. In fact every degenerate critical point is
unstable [97].

Families of functions. Two functions f, g : Rn R are right equivalent


around 0 if there is a diffeomorphism germ y : Rn Rn and a constant term
, such that g(x) = f (y(x)) + . Similarly, there is a notion of equivalence
for families of functions f, g : Rn Rr R which is given by the following
definition.

Defintion 1.4.4. If there exists a diffeomorphism e : Rr Rr , a smooth


map y : Rn Rr Rn , such that for each s Rr the map ys : Rn Rn ,
ys (x) = y(x, s) is a diffeomorphism and a smooth map : Rr R, such
that g(x, s) = f (ys (x), e(s)) + (s) for all (x, s) Rn Rr , then f and g are
equivalent.

As there is the notion of stability of functions, similarly there is the notion


of stability for families of functions in a natural way. If f : Rn Rr
R is equivalent in the above sense to any family f + p : Rn Rr R,
where p is a sufficiently small perturbation family p : Rn Rr R, then
f is stable. This brings us to a much celebrated result due to Thom, which
states that a generic k-parameter family, k 4, of functions is stable and
in the neighborhood of a critical point it behaves, up to sign and change of
variable, like one of the seven cases as( cf. Table 1.4.1) The term codimension
(codim) serves as a measure of complexity of a critical point. Any small
1.4. Classifications in singularity theory 25

Notation Codim Germ Name


A2 1 x3 + y 2 Fold
A3 2 x4 + y 2 Cusp
A4 3 x5 + y 2 Swallow-tail
D4 3 x3 + xy 2 Hyperbolic umbilic
D4+ 3 x3 xy 2 Elliptic umbilic
A5 4 x6 + y 2 Butterfly
D5 4 x4 + xy 2 Parabolic umbilic

Table 1.1: Seven classes of singularities due to Thom

perturbation of a function f of codim r leads to a function with at most r


critical points. Apart from an algebraic definition, a germ has codim k if it is
the smallest k such that the germ occurs in a generic k-parameter families of
local functions. Thom considered the classification of degenerate singularities.
However, in many cases, one is interested not in an individual object, but
in a collection of them, depending on some parameters. Towards the end
of the 1960s Mather closed the gaps in Thoms work by generalizing the
classification of stable map-germs, he devised a method of producing classes
of stable map-germs for any source and target dimension, making crucial use
of the concept of contact equivalence. For a collection of Mathers works we
refer to cf. [77, 79, 78, 81, 80] There is a rather huge collection of literature
of work on classification of singularities in various kinds of cases. For our
purpose we will focus on Z2 -equivariant maps in the following section.

1.4.2 Singularities of secant maps


In Chapter 6 our goal is to study the local structure of the secant map of
an immersion. We do this by classifying the singularities of such maps. The
principal contribution is the classification of the germs of secant maps of
generically immersed surfaces in Rn , for n 3. In this section, we elaborate
the concept of secant maps and illustrate the classification problem with an
example from Bruce [16]. The classification problem considered in Chapter 6
is an extension of the classification problem of space curves considered by
Bruce.
26 1. Introduction

Secant maps of space curves. Suppose : R R3 ,is a regular, smooth


embedding, the projectivized secant map S : R R P2 is defined by
1 , t2 ) = p((t1 ) (t2 )), if t1 6= t2
S(t
t)
S(t, = p( 0 (t)),

where P2 is the real projective two space and p : R3 \ {0} P2 , is the


natural projection onto a real projective space P2 , assigning to each point the
unoriented line through that point. Thus, the points on the diagonal =
{(t, t) | t R}, get mapped to the corresponding tangent line by the secant
map S. We consider the local behavior of the secant map for points on the
diagonal. Observe that the secant map has the following property, S(t 1 , t2 ) =
2 , t1 ), so we conclude that S has Z2 symmetry along the diagonal. Thus
S(t
to obtain local models of the secant map S, we work in the space of Z2
equivariant maps, where a function h : R2 , 0 R2 is said to be Z2 equivariant
if h(x) = h(x), for all x R2 .
The parametrized form of a generic regular space curve , upto a local
change of coordinates in R and R3 , is given by,
N
X N
X
: (R, 0) R3 , (t) = t, a2j tj + O(tN +1 ), a3j tj + O(tN +1 ) .


j=2 j=2

Thus the secant germ is given by S : (R2 , 0) R3 , such that


j
tj2 ) + O(tN
1 , t2 ) = PN +1
S(t t1 t2 , j=2 a2j (t1 1 tN
2
+1
),

j
tj2 ) + O(tN
PN +1
tN +1

j=2 a3j (t1 1 2 ) .

Since t1 6= t2 , S(t 1 , t2 ) spans the same line as 1, P a2j j +


2 (t1 , t2 ), a3j j + 3 (t1 , t2 ) , where j (t1 , t2 ) = (tj1 tj2 )(t1 t2 ) and
P 

j (t1 , t2 ) = O(tN 1
+1
tN
2
+1
). Hence the projectivized secant germ S has
a representation as the last two coordinates of S. Now since S is also a Z2
equivariant germ, composing it with the symmetry preserving map such that
x := 21 (t1 + t2 ) and y := 12 (t1 t2 ), we obtain a smooth map S : R2 , 0 R2 , 0,
with S(x, y) = S(x, y).

Proposition 1.4.5. Let be a space curve with nowhere vanishing curvature.


1.4. Classifications in singularity theory 27

1. If the torsion of the curve at t = 0 , (0) 6= 0, then S : R R R3


has a local representation at (0, 0), of the form S(x, y) = (x, y 2 ), which
is a fold 1.4.

2. If (0) = 0, but (0), 3(0) (0) 6= 0, then S has a local


(0) (0)
representation at (0, 0), of the form S(x, y) = (x, xy 2 +y 4 ), see figure 1.4.

Fold Pitch Fork

Figure 1.4: Singularities of secant maps of space curves

Proof. We give a proof sketch for the first case for Proposition 1.4.5. Thus
note that 2 = 2x, 3 = 3x2 +y 2 , 4 = 4(x3 +xy 2 ) and 5 = 5x4 +10x2 y 2 +y 4 .
The 2 jet of S at (0, 0) is

(2a22 x + a23 (3x2 + y 2 ), 2a32 x + a33 (3x2 + y 2 )).

This jet is Z2 equivalent to (x, 3x2 +y 2 ) and hence (x, y 2 ) (under Z2 equivari-
ant coordinate changes in the source and arbitrary coordinate changes in the
target) if and only if a22 a33 a23 a32 6= 0. Furthermore, the torsion of the
2-jet of the curve at t = 0 is given by
1

(0) = a222 +a232
a22 a33 a23 a32 ,

thus using Malgrange preparation theorem we conclude that the secant germ
(x, y 2 ) is equivalent to the condition that that (0) 6= 0.

For a detailed discussion on Z2 equivariant mappings from (R2 , 0)


(R2 , 0) we refer to Bruce [16, Appendix].
28 1. Introduction

1.5 Overview and main results


In this section we give a brief overview of chapters in this thesis and highlight
the main contributions.
Chapter 2 provides a review of the basic concepts of differential geometry
of plane and space curves. In this chapter the affine differential geometry of
plane curves is treated in some detail as it is used extensively in approximation
with conics in Chapter 3. In particular, we introduce affine arc length and
affine curvature, which are invariant under equiaffine transformations. Conic
arcs are the only curves in the plane having constant affine curvature, which
explains the relevance of these notions from affine differential geometry for
our work. Moreover, in this chapter we review the differential geometry of
generalized circular helices in three space. A derivation of another normal
frame for a space curve is given other than the Frenet-Serret frame, this
normal frame contains the normalized Darboux vector. The conditions for a
space curve to be a generalized helix is also reviewed.
In Chapter 3, we study the approximation of curves in the plane by arcs
of conics. We show that the complexity of a parabolic or conic spline approx-
imating a sufficiently smooth curve with non-vanishing curvature to within
Hausdorff distance is c1 1/4 + O(1), if the spline consists of parabolic arcs,
and c2 1/5 + O(1), if it is composed of general coinc arcs of varying type.
The constants c1 and c2 are expressed in the Euclidean and affine curvature
of the curve,
1
RL 1/4 (s)5/12 ds

c1 = 0 |k(s)|

4
128
1
RL 0
|k (s)|1/5 (s)2/5 ds ,

c2 = 5 0
2000 5

where L is the length of the given arc , k denotes its affine curvature, the
Euclidean curvature and s is the arc length parameter of . We also show that
the Hausdorff distance between a curve and an optimal conic arc tangent at its
endpoints is increasing with its arc length, provided the affine curvature along
the arc is monotone. This property yields a simple bisection algorithm for the
computation of an optimal parabolic or conic spline. Furthermore, we also
derive the complexity result for approximation of plane curves by parabolic
and conic arcs with respect to the symmetric difference distance. We show
1.5. Overview and main results 29

that the complexity of an optimal parabolic or conic spline approximating


a smooth curve with non-vanishing curvature to within symmetric difference
distance is c3 1/4 + O(1) for parabolic arcs and c4 1/5 + O(1), if the spline
is composed of general conic arcs, where
R% 5/4
c3 = (240)1/4 0 |k(r)|1/5 dr
R% 6/5
c4 = (7680)1/5 0 |k 0 (r)|1/6 dr ,

r denotes the affine arc length parameter. We define an equisymmetric conic


arc tangent to a curve at its endpoints, to be the (unique) conic such that the
areas of the two moons formed by this conic and the given curve are equal,
and show that its complexity is asymptotically equal to the complexity of an
optimal conic spline. We also show that the symmetric difference distance
between a curve and an equisymmetric conic arc tangent at its endpoints is
increasing with affine arc length, provided the affine curvature along the arc is
monotone. This property yields a simple and an efficient bisection algorithm
for the computation of an optimal parabolic or equisymmetric conic spline.
Chapter 4 deals with approximation of curves in space with bihelical
splines. In this chapter we show that the complexity of bihelical spline
approximating a sufficiently smooth space curve with non-vanishing, mono-
tonically increasing/decreasing curvature to within Hausdorff distance is
c1/3 + O(1), such that
Z L
1
|0 (s)|1/2 ds ,

c =
3
324
0

and is the Euclidean curvature, L length of the arc of the space curve and s
is the arc length parameter. A bihelical spline is composed of bihelical arcs. A
bihelical arc is composed of two helical arcs which are tangent to each other
at the junction point and tangent to the curve at the endpoints. We also
conclude that the expression for complexity of a bihelical spline is the same
as that of biarcs in the planar case. Furthermore, given points p0 , p1 and
tangents T0 , T1 , respectively we show that there exists a unique helical arc
through these points with the given tangents if and only if

hp0 p1 , T0 T1 i = 0.
30 1. Introduction

Given two points pL , pR and two tangents TL , TR , if we fix a direction


T0 S2 \ {TL , TR } and set it to be the junction tangent, then the locus
of points p0 , such that there is a bihelical arc through pL , pR and p0 with
tangents TL , TR and T0 is a line denoted by l(T0 ). Now as T0 traces a circle in
S2 , we get a family of lines l(T0 ), we show that the union of all such lines l(T0 )
is a cylinder and we call it the junction cylinder. Thus as T0 varies in S2 we
get a one parameter family of junction cylinder. From our local computations
on bihelices we get an expression of the optimal junction tangent, we use this
expression to define a near optimal junction tangent Tnopt , we use this expres-
sion for the near optimal junction tangent to obtain the corresponding line of
junction points l(Tnopt ). From this line of junction points we choose the one
that is closest to the given curve and define the bihelical arc associated with
it to be near optimal. Our experimental results suggest that the monotonicity
property is true and thus we implement a bisection algorithm for computing
a near optimal bihelical spline approximation of a given space curve. The
complexity results of our algorithm are promising, the experimental and the-
oretical results of complexity are almost the same for the curves that were
tested. Moreover, in the planar case of approximation with biarcs, we show
that in case of approximation of a planar spiral with biarcs, there is a unique
biarc minimizing the Hausdorff distance. Fixing two points pL and pR and
the tangents at these points TL and TR , there exists a one parameter family
of biarcs. The locus of the junction points of this one parameter family is a
circle and we call it the junction circle.

Finally in Chapter 6 we study the singularities of secant maps of immer-


sions of surfaces in Rn for n 3. We define the projectivized secant map of an
immersion X : Rn Rm+1 , and relate the secant map to n parameter family
of inner projections of X(Rn ) from centers in X(Rn ). We show that the sec-
ant map at a point outside the diagonal is a versal n parameter deformation
of a germ Rn Rm for a residual set of embeddings. Moreover, under some
restrictions on the pair of dimensions we show that the secant map germ along
the diagonal is AZ2 equivalent to some Z2 -stable germ for a residual set of
immersions X : Rn Rm , m > n. We also give the classification of germs of
secant maps at the diagonal of surfaces generically immersed in Rn , for n 3.
Also we give the classification of Z2 stable germs from R4 Rn , for n 2 ,
1.5. Overview and main results 31

this classification is relevant for the secant maps germs of generic immersions
X : R2 Rn+1 .

Publications. Parts of Chapter 2 and Chapter 3 are published as Ap-


proximations by conic splines in the journal Mathematics in Computer Sci-
ence [47]. Parts of Chapter 3 appeared as Minimizing the symmetric difference
distance in conic spline approximation in the proceedings of MACIS [50] and
as extended abstracts in [94, 51]. Chapter 4 is a preprint Approximation of
Space Curves by Helix Splines [49].
Chapter 6 is published as Singularities of secant maps of immersed surfaces
in journal Geometriae Dedicata [48].

You might also like