Spline Methods Draft Tom Lyche and Knut Mørken 24th May 2002 2 Contents 1 Splines and B-splines an introduction 1.1 Convex combinations and convex hulls . . . . . . . . . 1.1.1 Stable computations . . . . . . . . . . . . . . . 1.1.2 The convex hull of a set of points . . . . . . . . 1.2 Some fundamental concepts . . . . . . . . . . . . . . . 1.3 Interpolating polynomial curves . . . . . . . . . . . . . 1.3.1 Quadratic interpolation of three points . . . . . 1.3.2 General polynomial interpolation . . . . . . . . 1.3.3 Interpolation by convex combinations? . . . . . 1.4 Bézier curves . . . . . . . . . . . . . . . . . . . . . . . 1.4.1 Quadratic Bézier curves . . . . . . . . . . . . . 1.4.2 Bézier curves based on four and more points . 1.4.3 Composite Bézier curves . . . . . . . . . . . . . 1.5 A geometric construction of spline curves . . . . . . . 1.5.1 Linear spline curves . . . . . . . . . . . . . . . 1.5.2 Quadratic spline curves . . . . . . . . . . . . . 1.5.3 Spline curves of higher degrees . . . . . . . . . 1.5.4 Smoothness of spline curves . . . . . . . . . . . 1.6 Representing spline curves in terms of basis functions 1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . 2 Basic properties of splines and B-splines 2.1 Some simple consequences of the recurrence relation 2.2 Linear combinations of B-splines . . . . . . . . . . . 2.2.1 Spline functions . . . . . . . . . . . . . . . . . 2.2.2 Spline curves . . . . . . . . . . . . . . . . . . 2.3 A matrix representation of B-splines . . . . . . . . . 2.4 Algorithms for evaluating a spline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Further properties of splines and B-splines 3.1 Linear independence and representation of polynomials . . . . 3.1.1 Some properties of the B-spline matrices . . . . . . . . 3.1.2 Marsden’s identity and representation of polynomials . 3.1.3 Linear independence of B-splines . . . . . . . . . . . . i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 3 4 4 7 8 9 10 14 15 15 17 20 21 21 23 24 28 29 32 . . . . . . 37 37 43 43 46 47 50 . . . . 57 57 57 59 61 ii CONTENTS 3.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 63 66 68 70 4 Knot insertion 4.1 Convergence of the control polygon for spline functions . 4.2 Knot insertion . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Formulas and algorithms for knot insertion . . . 4.3 B-spline coefficients as functions of the knots . . . . . . 4.3.1 The blossom . . . . . . . . . . . . . . . . . . . . 4.3.2 B-spline coefficients as blossoms . . . . . . . . . 4.4 Inserting one knot at a time . . . . . . . . . . . . . . . . 4.5 Bounding the number of sign changes in a spline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 75 78 79 85 85 88 90 92 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 100 100 102 105 105 110 110 112 112 114 117 120 121 122 . . . . . . . . . . 127 . 127 . 127 . 129 . 130 . 131 . 131 . 133 . 134 . 135 . 136 3.3 Differentiation and smoothness of B-splines . . . . . . 3.2.1 Derivatives of B-splines . . . . . . . . . . . . . 3.2.2 Computing derivatives of splines and B-splines 3.2.3 Smoothness of B-splines . . . . . . . . . . . . . B-splines as a basis for piecewise polynomials . . . . . 5 Spline Approximation of Functions and Data 5.1 Local Approximation Methods . . . . . . . . . . . 5.1.1 Piecewise linear interpolation . . . . . . . . 5.1.2 Cubic Hermite interpolation . . . . . . . . . 5.1.3 Estimating the derivatives . . . . . . . . . . 5.2 Cubic Spline Interpolation . . . . . . . . . . . . . . 5.2.1 Interpretations of cubic spline interpolation 5.2.2 Numerical solution and examples . . . . . . 5.3 General Spline Approximation . . . . . . . . . . . 5.3.1 Spline interpolation . . . . . . . . . . . . . 5.3.2 Least squares approximation . . . . . . . . 5.4 The Variation Diminishing Spline Approximation . 5.4.1 Preservation of bounds on a function . . . . 5.4.2 Preservation of monotonicity . . . . . . . . 5.4.3 Preservation of convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Parametric Spline Curves 6.1 Definition of Parametric Curves . . . . . . . . . . . . . . . . . . . . 6.1.1 Regular parametric representations . . . . . . . . . . . . . . 6.1.2 Changes of parameter and parametric curves . . . . . . . . 6.1.3 Arc length parametrization . . . . . . . . . . . . . . . . . . 6.2 Approximation by Parametric Spline Curves . . . . . . . . . . . . . 6.2.1 Definition of parametric spline curves . . . . . . . . . . . . 6.2.2 The parametric variation diminishing spline approximation 6.2.3 Parametric spline interpolation . . . . . . . . . . . . . . . . 6.2.4 Assigning parameter values to discrete data . . . . . . . . . 6.2.5 General parametric spline approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . CONTENTS 1 7 Tensor Product Spline Surfaces 7.1 Explicit tensor product spline surfaces . . . . . . . . . 7.1.1 Definition of the tensor product spline . . . . . 7.1.2 Evaluation of tensor product spline surfaces . . 7.2 Approximation methods for tensor product splines . . 7.2.1 The variation diminishing spline approximation 7.2.2 Tensor Product Spline Interpolation . . . . . . 7.2.3 Least Squares for Gridded Data . . . . . . . . . 7.3 General tensor product methods . . . . . . . . . . . . 7.4 Trivariate Tensor Product Methods . . . . . . . . . . . 7.5 Parametric Surfaces . . . . . . . . . . . . . . . . . . . 7.5.1 Parametric Tensor Product Spline Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 Some Approximation Theory for Splines 8.1 Quasi-interpolants . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 The distance to polynomials . . . . . . . . . . . . . . . . . . . . . 8.3 The distance to splines . . . . . . . . . . . . . . . . . . . . . . . . 8.3.1 The constant and linear cases . . . . . . . . . . . . . . . . 8.3.2 The quadratic case . . . . . . . . . . . . . . . . . . . . . . 8.3.3 The general case . . . . . . . . . . . . . . . . . . . . . . . 8.4 Stability of the B-spline basis . . . . . . . . . . . . . . . . . . . . 8.4.1 A general definition of stability . . . . . . . . . . . . . . . 8.4.2 The condition number of the B-spline basis. Infinity norm 8.4.3 The condition number of the B-spline basis. p-norm . . . 8.5 Polynomial Interpolation with Linear Functionals . . . . . . . . . 8.6 Quasi-Interpolants . . . . . . . . . . . . . . . . . . . . . . . . . . 8.7 Two Quasi-Interpolants Based on Point Functionals . . . . . . . 8.7.1 A Quasi-Interpolant Based on Derivatives . . . . . . . . . 8.7.2 Quasi-Interpolants Based on Evaluation . . . . . . . . . . 9 Shape Preserving Properties of B-splines 9.1 Bounding the number of zeros of a spline 9.2 Uniqueness of spline interpolation . . . . . 9.2.1 Lagrange Interpolation . . . . . . . 9.2.2 Hermite Interpolation . . . . . . . 9.3 Total positivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 . 139 . 139 . 143 . 144 . 144 . 145 . 149 . 153 . 155 . 159 . 160 . . . . . . . . . . . . . . . 163 . 163 . 164 . 166 . 167 . 168 . 170 . 173 . 173 . 175 . 177 . 178 . 182 . 185 . 186 . 187 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 . 193 . 196 . 198 . 199 . 200 A Some Linear Algebra A.1 Matrices . . . . . . . . . . . . . . . . . . . . . . . A.1.1 Nonsingular matrices, and inverses. . . . . A.1.2 Determinants. . . . . . . . . . . . . . . . . A.1.3 Criteria for nonsingularity and singularity. A.2 Vector Norms . . . . . . . . . . . . . . . . . . . . A.3 Vector spaces of functions . . . . . . . . . . . . . A.3.1 Linear independence and bases . . . . . . A.4 Normed Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 205 205 206 206 207 209 210 212 2 CONTENTS CHAPTER 1 Splines and B-splines an introduction In this first chapter we consider the following fundamental problem: Given a set of points in the plane, determine a smooth curve that approximates the points. The algorithm for determining the curve from the points should be well suited for implementation on a computer. That is, it should be efficient and it should not be overly sensitive to roundoff errors in the computations. We only consider methods that involve a relatively small number of elementary arithmetic operations; this ensures that the methods are efficient. The sensitivity of the methods to round-off errors is controlled by insisting that all the operations involved should amount to forming weighted averages of the given points. This has the added advantage that the constructions are geometrical in nature and easy to visualise. In Section 1.1 we discuss affine and convex combinations and the convex hull of a set of points, and relate these concepts to numerical stability (sensitivity to rounding errors), while in Section 1.2 we give a brief and very informal introduction to parametric curves. The first method for curve construction, namely polynomial interpolation, is introduced in Section 1.3. In Section 1.4 we show how to construct Bezier curves, and in Section 1.5 we generalise this construction to spline curves. At the outset, our construction of spline curves is geometrical in nature, but in Section 1.6 we show that spline curves can be written conveniently in terms of certain basis functions, namely B-splines. In the final section, we relate the material in this chapter to the rest of the book. 1.1 Convex combinations and convex hulls An important constraint on our study is that it should result in numerical methods that will ultimately be implemented in floating point arithmetic on a computer. We should therefore make sure that these methods are reasonably insensitive to the primary source of problems, namely round off errors and other numerical uncertainties that occur in numerical computations. This requirement is often referred to by saying that the methods should be numerically stable. 3 4 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION 1.3 1 0.8 c2 0.5 0.3 0 c1 -0.4 Figure 1.1. Some points on the line (1 − λ)c1 + λc2 and the corresponding values of λ. 1.1.1 Stable computations One characteristic of numerical instabilities is that a chain of computations contain numbers of large magnitude even though the numbers that form the input to the computations, and the final result, are not particularly large numbers. A simple way to avoid this is to base the computations on computing weighted averages as in c = (1 − λ)c1 + λc2 . (1.1) Here c1 and c2 are two given numbers and λ a given weight in the range [0, 1]. The result of the computation is the number c which must lie between c1 and c2 as averages always do. A special example is of course computation of the mean between two numbers, c = (c1 + c2 )/2. A computation on the form (1.1) is often referred to as a convex combination, and c is often said to be a convex combination of c1 and c2 . If all our computations are convex combinations, all intermediate results as well as the final result must be within the numerical range of the input data, thereby indicating that the computations are reasonably stable. It is overly optimistic to hope that we can do all our computations by forming convex combinations, but convex combinations will certainly be a guiding principle. 1.1.2 The convex hull of a set of points Convex combinations make sense for vectors as well as for real numbers. If c1 = (x1 , y1 ) and c2 = (x2 , y2 ) then a convex combination of c1 and c2 is an expression on the form c = (1 − λ)c1 + λc2 , (1.2) where the weight λ is some number in the range 0 ≤ λ ≤ 1. This expression is usually implemented on a computer by expressing it in terms of convex combinations of real numbers, (x, y) = (1 − λ)x1 + λx2 , (1 − λ)y1 + λy2 , where (x, y) = c. 1.1. CONVEX COMBINATIONS AND CONVEX HULLS 5 c2 Λ c 1-Μ c Μ c3 1-Λ c1 Figure 1.2. Determining the convex hull of three points. Sometimes combinations on the form (1.1) or (1.2) with λ < 0 or λ > 1 are required. A combination of c1 and c2 as in (1.2) with no restriction on λ other than λ ∈ R is called an affine combination of c1 and c2 . As λ takes on all real numbers, the point c in (1.2) will trace out the whole straight line that passes through c1 and c2 . If we restrict λ to lie in the interval [0, 1], we only get the part of the line that lies between c1 and c2 . This is the convex hull, or the set of all weighted averages, of the two points. Figure 1.1 shows two points c1 and c2 and the line they define, together with some points on the line and their corresponding values of λ. We can form convex and affine combinations in any space dimension, we just let c1 and c2 be points in the appropriate space. If we are working in Rn for instance, then c1 and c2 have n components. In our examples we will mostly use n = 2, as this makes the visualisation simpler. Just as we can take the average of more than two numbers, it is possible to form convex combinations of more than two points. If we have n points (ci )ni=1 , a convex combination of the points is an expression on the form c = λ 1 c 1 + λ2 c 2 + · · · + λn c n P where the n numbers λi sum to one, ni=1 λi = 1, and also satisfy 0 ≤ λi ≤ 1 for i = 1, 2, . . . , n. As for two points, the convex hull of the points (ci )ni=1 is the set of all possible convex combinations of the points. It can be shown that the convex hull of a set of points is the smallest convex set that contain all the points (recall that a set is convex if the straight line connecting any two points in the set is always completely contained in the set). This provides a simple geometric interpretation of the convex hull. As we have already seen, the convex hull of two points can be identified with the straight line segment that connects the points, whereas the convex hull of three points coincides with the triangle spanned by the points, see Figure 1.2. In general, the convex hull of n points is the n-sided polygon with the points as corners. However, if some of the points are contained in the convex hull of the others then the number of edges is reduced correspondingly, see the examples in Figure 1.3. 6 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION (a) Two points. (b) Three points. (c) Four points. (d) Five points. (e) Five points. (f) Five points. Figure 1.3. Examples of convex hulls (shaded area) of points (black dots). 1.2. SOME FUNDAMENTAL CONCEPTS 1.2 7 Some fundamental concepts Our basic challenge in this chapter is to construct a curve from some given points in the plane. The underlying numerical algorithms should be simple and efficient and preferably based on forming repeated convex combinations as in (1.1). To illustrate some fundamental concepts let us consider the case where we are given two points c0 = (x0 , y0 ) and c1 = (x1 , y1 ) (we always denote points and vectors by bold type). The most natural curve to construct from these points is the straight line segment which connects the two points. In Section 1.1.2 we saw that this line segment coincides with the convex hull of the two points and that a point on the line could be represented by a convex combination, see (1.2). More generally we can express this line segment as q(t | c0 , c1 ; t0 , t1 ) = t1 − t t − t0 c0 + c1 t1 − t0 t1 − t0 for t ∈ [t0 , t1 ]. (1.3) Here t0 and t1 are two arbitrary real numbers with t0 < t1 . Note that the two coefficients add to one, t1 − t t − t0 + =1 t1 − t0 t1 − t0 and each of them is nonnegative as long as t is in the interval [t0 , t1 ]. The expression in (1.3) is therefore a convex combination of c0 and c1 . In fact, if we set λ = (t − t0 )/(t1 − t0 ) then (1.3) becomes (1.2). A representation of a line as in (1.3), where we have a function that maps each real number to a point in R2 , is an example of a parametric representation. The line can also be expressed as a linear function y = f (x) = x1 − x x − x0 y0 + y1 x1 − x0 x1 − x0 but here we run into problems if x0 = x1 , i.e., if the line is vertical. Vertical lines can only be expressed as x = c (with each constant c characterising a line) if we insist on using functions. In general, a parametric representation can cross itself or return to its starting point, but this is impossible for a function, which always maps a real number to a real number, see the two examples in Figure 1.4. In this chapter we only work with parametric representations in the plane, and we will refer to these simply as (parametric) curves. All our constructions start with a set of points, from which we generate new points, preferably by forming convex combinations as in (1.2). In our examples the points lie in the plane, but we emphasise again that the constructions will work for curves in any space dimension; just replace the planar points with points with the appropriate number of components. For example, a line in space is obtained by letting c0 and c1 in (1.3) be points in space with three components. In particular, we can construct a function by letting the points be real numbers. In later chapters we will work mainly with functions since the core of the spline theory is independent of the space dimension. The reason for working with planar curves in this chapter is that the constructions are geometric in nature and particularly easy to visualise in the plane. In (1.3) the two parameters t0 and t1 are arbitrary except that we assumed t0 < t1 . Regardless of how we choose the parameters, the resulting curve is always the same. If we consider the variable t to denote time, the parametric representation q(t | c0 , c1 ; t0 , t1 ) 8 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION 4 2 -2 1 -1 2 3 4 -2 -4 (a) (b) Figure 1.4. A function (a) and a parametric curve (b). gives a way to travel from c0 to c1 . The parameter t0 gives the time at which we start at c0 and t1 the time at which we arrive at c1 . With this interpretation, different choices of t0 and t1 correspond to different ways of travelling along the line. The speed of travel along the curve is given by the tangent vector or derivative q 0 (t | c0 , c1 ; t0 , t1 ) = c1 − c0 , t1 − t0 while the scalar speed or velocity is given by the length of the tangent vector p 2 2 0 q (t | c0 , c1 ; t0 , t1 ) = |c1 − c0 | = (x1 − x0 ) + (y1 − y0 ) . t1 − t0 t1 − t0 If t1 − t0 is small (compared to |c1 − c0 |), then we have to travel quickly to reach c1 at time t1 whereas if t1 − t0 is large then we have to move slowly to arrive at c1 exactly at time t1 . Note that regardless of our choice of t0 and t1 , the speed along the curve is independent of t and therefore constant. This reflects the fact that all the representations of the line given by (1.3) are linear in t. This discussion shows how we must differentiate between the geometric curve in question (a straight line in our case) and the parametric representation of the curve. Loosely speaking, a curve is defined as the collection of all the different parametric representations of the curve. In practise a curve is usually given by a particular parametric representation, and we will be sloppy and often refer to a parametric representation as a curve. The distinction between a curve and a particular parametric representation is not only of theoretical significance. When only the geometric shape is significant we are discussing curves and their properties. Some examples are the outlines of the characters in a font and the level curves on a map. When it is also significant how we travel along the curve (how it is represented) then we are talking about a particular parametric representation of the underlying geometric curve, which in mathematical terms is simply a vector valued function. An example is the path of a camera in a computer based system for animation. 1.3 Interpolating polynomial curves A natural way to construct a curve from a set of given points is to force the curve to pass through the points, or interpolate the points, and the simplest example of this is the 1.3. INTERPOLATING POLYNOMIAL CURVES 9 (a) t = (0, 1, 2). (b) t = (0, 0.5, 2). (c) t = (0, 1, 2). (d) t = (0, 0.5, 2). Figure 1.5. Some examples of quadratic interpolation. straight line between the points. In this section we show how to construct curves that interpolate any number of points. 1.3.1 Quadratic interpolation of three points How can we construct a curve that interpolates three points? In addition to the three given interpolation points c0 , c1 and c2 we also need three parameters (ti )2i=0 . We first construct the two straight lines q 0,1 (t) = q(t | c0 , c1 ; t0 , t1 ) and q 1,1 (t) = q(t | c1 , c2 ; t1 , t2 ). If we now form the weighted average q 0,2 (t) = q(t | c0 , c1 , c2 ; t0 , t1 , t2 ) = t2 − t t − t0 q 0,1 (t) + q (t), t2 − t0 t2 − t0 1,1 we obtain a curve that is quadratic in t, and it is easy to check that it passes through the given points as required, q 0,2 (t0 ) = q 0,1 (t0 ) = c0 , t1 − t0 t2 − t1 t1 − t0 t2 − t1 q 0,2 (t1 ) = q 0,1 (t1 ) + q 1,1 (t1 ) = c1 + c1 = c1 , t2 − t0 t2 − t0 t 2 − t0 t2 − t0 q 0,2 (t2 ) = q 1,1 (t2 ) = c2 . Four examples are shown in Figure 1.5, with the interpolation points (ci )2i=0 given as black dots and the values of the three parameters t = (ti )2i=0 shown below each plot. 10 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION The tangent vector at the end of the curve (at t = t2 ) is also displayed in each case. Note that the interpolation points are the same in plots (a) and (b), and also in plots (c) and (d). When we only had two points, the linear interpolant between the points was independent of the values of the parameters t0 and t1 ; in the case of three points and quadratic interpolation the result is clearly highly dependent on the choice of parameters. It is possible to give qualitative explanations of the results if we view q 0,2 (t) as the position at time t of someone travelling along the curve. In the first two plots the given points are quite uniformly spaced and the uniform distribution of parameters in plot (a) seems to connect the points with a ’nice’ curve. In plot (b) the value of t1 has been lowered, leaving more ‘time’ for travelling from c1 to c2 than from c0 to c1 with the effect that the curve bulges out between c1 and c2 . This makes the journey between these points longer and someone travelling along the curve can therefore spend the extra time allocated to this part of the ‘journey’. The curves in Figure 1.5 (c) and (d) can be explained similarly. The interpolation points are the same in both cases, but now they are not uniformly distributed. In plot (a) the parameters are uniform which means that we must travel much faster between c1 and c2 (which are far apart) than between c0 and c1 (which are close together). The result is a curve that is almost a straight line between the last two points and bulges out between the first two points. In plot (d) the parameters have been chosen so as to reflect better the geometric spacing between the points, and this gives a more uniformly rounded curve. 1.3.2 General polynomial interpolation To construct a cubic curve that interpolates four points we follow the same strategy that was used to construct the quadratic interpolant. If the given points are (ci )3i=0 we first choose four parameters t = (ti )3i=0 . We then form the two quadratic interpolants q 0,2 (t) = q(t | c0 , c1 , c2 ; t0 , t1 , t2 ), q 1,2 (t) = q(t | c1 , c2 , c3 ; t1 , t2 , t3 ), and combine these to obtain the cubic interpolant q 0,3 (t), q 0,3 (t) = t − t0 t3 − t q (t) + q (t). t3 − t0 0,2 t3 − t0 1,2 At t0 this interpolant agrees with q 0,2 (t0 ) = c0 and at t3 it agrees with q 1,2 (t3 ) = c3 . At an interior point ti it is a convex combination of q 0,1 (ti ) and q 1,1 (ti ) which both interpolate ci at ti . Hence we also have q 0,3 (ti ) = ci for i = 1 and i = 2 so q 0,3 interpolates the four points (ci )3i=0 as it should. Some examples of cubic interpolants are shown in Figure 1.6, and the same interpolation points are used in (a) and (b), and (c) and (d) respectively. The qualitative comments that we made about the quadratic interpolants also apply here. The pleasing shape of the curve in Figure 1.6 (a) is quite natural since both the interpolation points and parameters are quite uniformly spaced. However, by adjusting the parameters, quite strange behaviour can occur, even with these ‘nice’ interpolation points. In (b) there is so much time to ‘waste’ between c1 and c2 that the curve makes a complete loop. In (c) and (d) we see two different approaches to jumping from one level in the data to another. In (c) there is too much time to be spent between c0 and c1 , and between c2 and c3 , the result 1.3. INTERPOLATING POLYNOMIAL CURVES 11 (a) t = (0, 1, 2, 3). (b) t = (0, 0.3, 2.7, 3). (c) t = (0, 0.75, 2.25, 3). (d) t = (0, 0.3, 2.8, 3). Figure 1.6. Some examples of cubic interpolation. 12 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION being bulges between these points. In Figure 1.6 (d) there is too much time between c1 and c2 leading to the two big wiggles and almost straight lines between c0 and c1 , and c2 and c3 respectively. The general strategy for constructing interpolating curves should now be clear. Given d+1 points (ci )di=0 and parameters (ti )di=0 , the curve q 0,d of degree d that satisfies q 0,d (tj ) = cj for j = 0, . . . , d is constructed by forming a convex combination between the two curves d of degree d − 1 that interpolate (ci )d−1 i=0 and (ci )i=1 , q 0,d (t) = td − t t − t0 q 0,d−1 (t) + q (t). td − t0 td − t0 1,d−1 (1.4) If we expand out this equation we find that q 0,d (t) can be written q 0,d (t) = c0 `0,d (t) + c1 `1,d (t) + · · · + cd `d,d (t), (1.5) where the functions {`i,d }di=0 are the Lagrange polynomials of degree d given by `i,d (t) = Y (t − tj ) . ti − tj (1.6) 0≤j≤d j6=i It is easy to check that these polynomials satisfy the condition ( 1, if k = i, `i,d (tk ) = 0, otherwise, which is necessary since q 0,d (tk ) = ck . The complete computations involved in computing q 0,d (t) is summarised in the following algorithm. Algorithm 1.1 (Neville/Aitken method). Let d be a positive integer and let the d + 1 points (ci )di=0 be given together with d + 1 strictly increasing parameter values t = (ti )di=0 . There is a polynomial curve q 0,d of degree d that satisfies the conditions q 0,d (ti ) = ci for i = 0, 1, . . . , d, and for any real number t the following algorithm computes the point q 0,d (t). First set q i,0 (t) = ci for i = 0, 1, . . . , d and then compute q i,r (t) = ti+r − t t − ti q i,r−1 (t) + q (t) ti+r − ti ti+r − ti i+1,r−1 for i = 0, 1, . . . , d − r and r = 1, 2, . . . , d. The computations involved in determining a cubic interpolating curve are shown in the triangular table in Figure 1.7. The computations start from the right and proceed to the left and at any point a quantity q i,r is computed by combining, in an affine combination, the two quantities at the beginning of the two arrows meeting at q i,r . The expression between the two arrows is the denominator of the weights in the affine combination while the two numerators are written along the respective arrows. 1.3. INTERPOLATING POLYNOMIAL CURVES 13 c0 t t1 - q0,1 t2 t1 -t0 -t t- t 0 q0,2 t3 c1 t2 -t0 -t t2 -t t- t 0 q0,3 q1,1 t3 -t0 t3 t2 -t1 -t t- t t- t 0 1 q1,2 c2 t3 -t1 t3 - t t- t 1 q3,\ 1 t3 -t2 t- t 2 c3 Figure 1.7. Computing a point on a cubic interpolating curve. (a) t = (0, 1, 2, 3, 4, 5). (b) t = (0, 0.5, 2, 3, 4.5, 5). Figure 1.8. Two examples of interpolation with polynomial curves of degree five. 14 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION Two examples of curves of degree five are shown in Figure 1.8, both interpolating the same points. The wiggles in (a) indicate that t1 − t0 and t6 − t5 should be made smaller and the result in (b) confirms this. It should be emphasised that choosing the correct parameter values is a complex problem. Our simple analogy with travelling along a road may seem to explain some of the behaviour we have observed, but to formalise these observations into a foolproof algorithm for choosing parameter values is a completely different matter. As we shall see later, selection of parameter values is also an issue when working with spline curves. The challenge of determining good parameter values is not the only problem with polynomial interpolation. A more serious limitation is the fact that the polynomial degree is only one less than the number of interpolation points. In a practical situation we may be given several thousand points which would require a polynomial curve of an impossibly high degree. To compute a point on a curve of degree d requires a number of multiplications and additions that are at best proportional to d (using the Newton form of the interpolating polynomial); the algorithm we have presented here requires roughly d2 additions and multiplications. If for example d = 1000, computer manipulations like plotting and interactive editing of the curve would be much too slow to be practical, even on today’s fast computers. More importantly, it is well known that round-off errors in the computer makes numerical manipulations of high degree polynomials increasingly (with the degree) inaccurate. We therefore need alternative ways to approximate a set of points by a smooth curve. 1.3.3 Interpolation by convex combinations? In the interpolation algorithm for polynomials of degree d, Algorithm 1.1, the last step is to form a convex combination between two polynomials of degree d − 1, q 0,d (t) = t − t0 td − t q 0,d−1 (t) + q (t). td − t0 td − t0 1,d−1 More precisely, the combination is convex as long as t lies in the interval [t0 , td ]. But if the algorithm is based on forming convex combinations, any point on the final curve should be within the convex hull of the given interpolation points. By merely looking at the figures it is clear that this is not true, except in the case where we only have two points and the interpolant is the straight line that connects the points. To see what is going on, let us consider the quadratic case in detail. Given the points (ci )2i=0 and the parameters (ti )2i=0 , we first form the two straight lines t1 − t c0 + t1 − t0 t2 − t q 1,1 (t) = c1 + t2 − t1 q 0,1 (t) = t − t0 c1 , t1 − t0 t − t1 c2 , t2 − t1 (1.7) (1.8) and from these the quadratic segment q 0,2 (t) = t2 − t t − t0 q 0,1 (t) + q (t). t2 − t0 t2 − t0 1,1 (1.9) The combination in (1.7) is convex as long as t is in [t0 , t1 ], the combination in (1.8) is convex when t lies within [t1 , t2 ], and the combination in (1.9) is convex when t is restricted 1.4. BÉZIER CURVES 15 (a) Two points on the curve. (b) Thirty points on the curve. Figure 1.9. The geometry of quadratic interpolation. to [t0 , t2 ]. But in computing q 0,2 (t) we also have to compute q 0,1 (t) and q 1,1 (t), and one of these latter combinations will not be convex when t is in [t0 , t2 ] (except when t = t1 ). The problem lies in the fact that the two line segments are defined over different intervals, namely [t0 , t1 ] and [t1 , t2 ] that only has t1 in common, so t cannot be in both intervals simultaneously. The situation is illustrated in Figure 1.9. In the next section we shall see how we can construct polynomial curves from points in the plane by only forming convex combinations. The resulting curve will then lie within the convex hull of the given points, but will not interpolate the points. 1.4 Bézier curves The curve construction method that we consider in this section is an alternative to polynomial interpolation and produces what we call Bézier curves. Bézier curves are also polynomial curves and for that reason not very practical, but they avoid the problem of wiggles and bulges because all computations are true convex combinations. It also turns out that segments of Bézier curves can easily be joined smoothly together to form more complex shapes. This avoids the problem of using curves of high polynomial degree when many points are approximated. Bézier curves are a special case of the spline curves that we will construct in Section 1.5. 1.4.1 Quadratic Bézier curves We have three points in the plane c0 , c1 and c2 , and based on these points we want to construct a smooth curve, by forming convex combinations of the given points. With polynomial interpolation this did not work because the two line segments (1.7) and (1.8) are defined over different intervals. The natural solution is to start by defining the two line segments over the same interval, say [0, 1] for simplicity, p1,1 (t) = p(t | c0 , c1 ) = (1 − t)c0 + tc1 , (1.10) p2,1 (t) = p(t | c1 , c2 ) = (1 − t)c1 + tc2 . (1.11) (The curves we construct in this section and the next are related and will be denoted by p to distinguish them from the interpolating curves of Section 1.3.) Now we have no 16 CHAPTER 1. SPLINES AND B-SPLINES (a) AN INTRODUCTION (b) Figure 1.10. A Bézier curve based on three points. (a) (b) Figure 1.11. Two examples of quadratic Bézier curves. problem forming a true convex combination, p2,2 (t) = p(t | c0 , c1 , c2 ) = (1 − t)p1,1 (t) + tp2,1 (t). (1.12) The construction is illustrated in Figure 1.10 (a). In Figure 1.10 (b), where we have repeated the construction for 15 uniformly spaced values of t, the underlying curve is clearly visible. If we insert the explicit expressions for the two lines in (1.10) and (1.11) in (1.12) we find p2,2 (t) = (1 − t)2 c0 + 2t(1 − t)c1 + t2 c2 = b0,2 (t)c0 + b1,2 (t)c1 + b2,2 (t)c2 . (1.13) This is called a quadratic Bézier curve; the points (ci )2i=0 are called the control points of the curve and the piecewise linear curve connecting the control points is called the control polygon of the curve. Two examples of quadratic Bézier curves with their control points and control polygons are shown in Figure 1.11 (the two sets of interpolation points in Figure 1.5 have been used as control points). Some striking geometric features are clearly visible in Figures 1.10 and 1.11. We note that the curve interpolates c0 at t = 0 and c2 at t = 1. This can be verified algebraically 1.4. BÉZIER CURVES 17 (a) (b) Figure 1.12. Constructing a Bézier curve from four points. by observing that b0,2 (0) = 1 and b1,2 (0) = b2,2 (0) = 0, and similarly b2,2 (1) = 1 while b0,2 (1) = b1,2 (1) = 0. The line from c0 to c1 coincides with the direction of the tangent to the curve at t = 0 while the line from c1 to c2 coincides with the direction of the tangent at t = 1. This observation can be confirmed by differentiating equation (1.13). We find p02,2 (0) = 2(c1 − c0 ), p02,2 (1) = 2(c2 − c1 ). The three polynomials in (1.13) add up to 1, (1 − t)2 + 2t(1 − t) + t2 = (1 − t + t)2 = 1, and since t varies in the interval [0, 1], we also have 0 ≤ bi,2 (t) ≤ 1 for i = 0, 1, 2. This confirms that p2,2 (t) is a convex combination of the three points (ci )2i=0 . The geometric interpretation of this is that the curve lies entirely within the triangle formed by the three given points, the convex hull of c0 , c1 and c2 . 1.4.2 Bézier curves based on four and more points The construction of quadratic Bézier curves generalises naturally to any number of points and any polynomial degree. If we have four points (ci )3i=0 we can form the cubic Bézier curve p3,3 (t) = p(t | c0 , c1 , c2 , c3 ) by taking a weighted average of two quadratic curves, p3,3 (t) = (1 − t)p2,2 (t) + tp3,2 (t). If we insert the explicit expressions for p2,2 (t) and p3,2 (t), we find p3,3 (t) = (1 − t)3 c0 + 3t(1 − t)2 c1 + 3t2 (1 − t)c2 + t3 c3 . The construction is illustrated in Figure 1.12. Figure (a) shows the construction for a given value of t, and in Figure (b) the cubic and the two quadratic curves are shown together with the lines connecting corresponding points on the two quadratics (every point on the cubic lies on such a line). The data points are the same as those used in Figure 1.6 (a) and (b). Two further examples are shown in Figure 1.13, together with the control points and control polygons which are defined just as in the quadratic case. The data points in Figure 1.13 are the same as those used in Figure 1.6 (c) and (d). In 18 CHAPTER 1. SPLINES AND B-SPLINES (a) AN INTRODUCTION (b) Figure 1.13. Two examples of cubic Bézier curves. Figure 1.13 (b) the control polygon crosses itself with the result that the underlying Bézier curve does the same. To construct Bézier curves of degree d, we start with d + 1 control points (ci )di=0 , and form a curve pd,d (t) = p(t | c0 , . . . , cd ) based on these points by taking a convex combination of the two Bézier curves pd−1,d−1 and pd,d−1 of degree d − 1 which are based d on the control points (ci )d−1 i=0 and (ci )i=1 respectively, pd,d (t) = (1 − t)pd−1,d−1 (t) + tpd,d−1 (t). If we expand out we find by an inductive argument that pd,d (t) = b0,d (t)c0 + · · · + bd,d (t)cd , (1.14) where d i bi,d (t) = t (1 − t)d−i . i As in the quadratic case we have b0,d (t) + b1,d (t) + · · · + bd,d (t) = (1 − t + t)d = 1 and 0 ≤ bi,d (t) ≤ 1 for any t in [0, 1] and 0 ≤ i ≤ d. For any t in [0, 1] the point pd,d (t) therefore lies in the convex hull of the points (ci )di=0 . The curve interpolates the first and last control points and the tangent at t = 0 points in the direction from c0 to c1 and the tangent at t = 1 points in the direction from cd−1 to cd , p0d,d (0) = d(c1 − c0 ), p0d,d (1) = d(cd − cd−1 ). (1.15) As in the quadratic and cubic cases the piecewise linear curve with the control points as vertices is called the control polygon of the curve. The complete computations involved in computing a point on a Bézier curve are given in Algorithm 1.2 and depicted graphically in the triangular table in Figure 1.14. Algorithm 1.2. Let d be a positive integer and let the d + 1 points (ci )di=0 be given. The point pd,d (t) on the Bézier curve p0,d of degree d can be determined by the following computations. First set pi,0 (t) = ci for i = 0, 1, . . . , d and then compute pd,d (t) by pi,r (t) = (1 − t)pi−1,r−1 (t) + tpi,r−1 (t) 1.4. BÉZIER CURVES 19 c0 1- t p1, 1 1- t t p2, 2 c1 1- t 1- t t p3,3 p2,1 1- t t t p3, 2 c2 1- t t p3,\ 1 t c3 Figure 1.14. Computing a point on a cubic Bézier curve. for i = r, . . . , d and r = 1, 2, . . . , d. Two examples of Bézier curves of degree five are shown in Figure 1.15. The curve in Figure (a) uses the interpolation points of the two curves in Figure 1.8 as control points. We have defined Bézier curves on the interval [0, 1], but any nonempty interval would work. If the interval is [a, b] we just have to use convex combinations on the form c= b−t t−a c0 + c1 b−a b−a instead. Equivalently, we can use a linear change of parameter; if pd,d (t) is a Bézier curve on [0, 1] then p̃d,d (s) = pd,d (t − a)/(b − a) (a) (b) Figure 1.15. Two Bézier curves of degree five. 20 CHAPTER 1. SPLINES AND B-SPLINES (a) AN INTRODUCTION (b) Figure 1.16. Different forms of continuity between two segments of a cubic Bézier curve. is a Bézier curve on [a, b]. 1.4.3 Composite Bézier curves By using Bézier curves of sufficiently high degree we can represent a variety of shapes. However, Bézier curves of high degree suffer from the same shortcomings as interpolating polynomial curves: 1. As the degree increases, the complexity and therefore the processing time increases. 2. Because of the increased complexity, curves of high degree are more sensitive to round-off errors. 3. The relation between the given data points (ci )di=0 and the curve itself becomes less intuitive when the degree is large. Because of these shortcomings it is common to form complex shapes by joining together several Bézier curves, most commonly of degree two or three. Such composite Bézier curves are also referred to as Bézier curves. A Bézier curve of degree d consisting of n segments is given by n sets of control points i (c0 , . . . , cid )ni=1 . It is common to let each segment be defined over [0, 1], but it is also possible to form a curve defined over the interval [0, n] with segment i defined on the interval [i − 1, i]. By adjusting the control points appropriately it is possible to ‘glue’ together the segments with varying degrees of continuity. The minimal form of continuity = ci0 which ensures that segments i − 1 and i join together continuously as is to let ci−1 d in Figure 1.16 (a). We obtain a smoother join by also letting the tangents be continuous at the join. From (1.15) we see that the tangent at the join between segments i − 1 and i will be continuous if i i ci−1 − ci−1 d d−1 = c1 − c0 . An example is shown in Figure 1.16 (b). Quadratic Bézier curves form the basis for the TrueType font technology, while cubic Bézier curves lie at the heart of PostScript and a number of draw programs like Adobe Illustrator. Figure 1.17 shows one example of a complex Bézier curve. It is the letter S in the Postscript font Times Roman, shown with its control polygon and control points. 1.5. A GEOMETRIC CONSTRUCTION OF SPLINE CURVES 21 This is essentially a cubic Bézier curve, interspersed with a few straight line segments. Each cubic curve segment can be identified by the two control points on the curve giving the ends of the segment and the two intermediate control points that lie off the curve. 1.5 A geometric construction of spline curves The disadvantage of Bézier curves is that the smoothness between neighbouring polynomial pieces can only be controlled by choosing the control points appropriately. It turns out that by adjusting the construction of Bézier curves slightly, we can produce pieces of polynomial curves that automatically tie together smoothly. These piecewise polynomial curves are called spline curves. 1.5.1 Linear spline curves The construction of spline curves is also based on repeated averaging, but we need a slight generalisation of the Bézier curves, reminiscent of the construction of the interpolating polynomials in Section 1.3. In Section 1.3 we introduced the general representation (1.3) for a straight line connecting two points. In this section we use the same general representation, but with a different labelling of the points and parameters. If we have two points c1 and c2 we now represent the straight line between them by p(t | c1 , c2 ; t2 , t3 ) = t − t2 t3 − t c1 + c2 , t 3 − t2 t3 − t2 t ∈ [t2 , t3 ], (1.16) provided t2 < t3 . By setting t2 = 0 and t3 = 1 we get back to the linear Bézier curve. The construction of a piecewise linear curve based on some given points (ci )ni=1 is quite obvious; we just connect each pair of neighbouring points by a straight line. More specifically, we choose n numbers (ti )n+1 i=2 with ti < ti+1 for i = 2, 3, . . . , n, and define the curve f by t ∈ [t2 , t3 ), p(t | c1 , c2 ; t2 , t3 ), p(t | c2 , c3 ; t3 , t4 ), t ∈ [t3 , t4 ), f (t) = (1.17) . .. .. . p(t | cn−1 , cn ; tn , tn+1 ), t ∈ [tn , tn+1 ]. The points (ci )ni=1 are called the control points of the curve, while the parameters t = (ti )n+1 i=2 , which give the value of t at the control points, are referred to as the knots, or knot vector, of the curve. If we introduce the piecewise constant functions Bi,0 (t) defined by ( 1, ti ≤ t < ti+1 , Bi,0 (t) = (1.18) 0, otherwise, and set pi,1 (t) = p(t | ci−1 , ci ; ti , ti+1 ), we can write f (t) more succinctly as f (t) = n X pi,1 (t)Bi,0 (t). (1.19) i=2 This construction can be generalised to produce smooth, piecewise polynomial curves of higher degrees. 22 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION Figure 1.17. The letter S in the Postscript font Times Roman. 1.5. A GEOMETRIC CONSTRUCTION OF SPLINE CURVES 23 Figure 1.18. Construction of a segment of a quadratic spline curve. 1.5.2 Quadratic spline curves In the definition of the quadratic Bézier curve, a point on p2,2 (t) is determined by taking three averages, all with weights 1 − t and t since both the two line segments (1.10) and (1.11), and the quadratic curve itself (1.12), are defined with respect to the interval [0, 1]. The construction of spline functions is a hybrid between the interpolating polynomials of Section 1.3 and the Bézier curve of Section 1.4 in that we retain the convex combinations, but use more general weighted averages of the type in (1.16). To construct a spline curve based on the three control points c1 , c2 , and c3 , we introduce four knots (ti )5i=2 , with the assumption that t2 ≤ t3 < t4 ≤ t5 . We represent the line connecting c1 and c2 by p(t | c1 , c2 ; t2 , t4 ) for t ∈ [t2 , t4 ], and the line connecting c2 and c3 by p(t | c2 , c3 ; t3 , t5 ) for t ∈ [t3 , t5 ]. The reason for picking every other knot in the representation of the line segments is that then the interval [t3 , t4 ] is within the domain of both segments. This ensures that the two line segments can be combined in a convex combination to form a quadratic curve, p(t | c1 , c2 , c3 ; t2 , t3 , t4 , t5 ) = t4 − t t − t3 p(t | c1 , c2 ; t2 , t4 ) + p(t | c2 , c3 ; t3 , t5 ) (1.20) t4 − t3 t4 − t3 with t varying in [t3 , t4 ]. Of course we are free to vary t throughout the real line R since p is a polynomial in t, but then the three combinations involved are no longer all convex. The construction is illustrated in Figure 1.18. Note that if t2 = t3 = 0 and t4 = t5 = 1 we are back in the Bézier setting. The added flexibility provided by the knots t2 , t3 , t4 and t5 turns out to be exactly what we need to produce smooth, piecewise quadratic curves, and by including sufficiently many control points and knots we can construct curves of almost any shape. Suppose we have n control points (ci )ni=1 and a sequence of knots (ti )n+2 i=2 that are assumed to be increasing except that we allow t2 = t3 and tn+1 = tn+2 . We define the quadratic spline curve f (t) by p(t | c1 , c2 , c3 ; t2 , t3 , t4 , t5 ), t 3 ≤ t ≤ t4 , p(t | c2 , c3 , c4 ; t3 , t4 , t5 , t6 ), t 4 ≤ t ≤ t5 , f (t) = (1.21) . .. .. . p(t | cn−2 , cn−1 , cn ; tn−1 , tn , tn+1 , tn+2 ), tn ≤ t ≤ tn+1 . An example with n = 4 is shown in Figure 1.19. Part (a) of the figure shows a quadratic curve defined on [t3 , t4 ] and part (b) a curve defined on the adjacent interval [t4 , t5 ]. In 24 CHAPTER 1. SPLINES AND B-SPLINES (a) AN INTRODUCTION (b) (c) Figure 1.19. A quadratic spline curve (c) and its two polynomial segments (a) and (b). part (c) the two curves in (a) and (b) have been superimposed in the same plot, and, quite strikingly, it appears that the curves meet smoothly at t4 . The precise smoothness properties of splines will be proved in Section 3.2.3 of Chapter 3; see also exercise 6. By making use of the piecewise constant functions {Bi,0 }ni=2 defined in (1.18) and the abbreviation pi,2 (t) = p(t | ci−2 , ci−1 , ci ; ti−1 , ti , ti+1 , ti+2 ), we can write f (t) as f (t) = n X pi,2 (t)Bi,0 (t). (1.22) i=2 Two examples of quadratic spline curves are shown in Figure 1.20. The control points are the same as those in Figure 1.13. We observe that the curves behave like Bézier curves at the two ends. 1.5.3 Spline curves of higher degrees The construction of spline curves can be generalised to arbitrary polynomial degrees by forming more averages. A cubic spline segment requires four control points ci−3 , ci−2 , ci−1 , ci , and six knots (tj )i+3 j=i−2 which must form a nondecreasing sequence of numbers 1.5. A GEOMETRIC CONSTRUCTION OF SPLINE CURVES (a) 25 (b) Figure 1.20. Two quadratic spline curves, both with knots t = (0, 0, 0, 1, 2, 2, 2). with ti < ti+1 . The curve is the average of two quadratic segments, p(t | ci−3 , ci−2 , ci−1 , ci ; ti−2 , ti−1 , ti , ti+1 , ti+2 , ti+3 ) = ti+1 − t p(t | ci−3 , ci−2 , ci−1 ; ti−2 , ti−1 , ti+1 , ti+2 )+ ti+1 − ti t − ti p(t | ci−2 , ci−1 , ci ; ti−1 , ti , ti+2 , ti+3 ), (1.23) ti+1 − ti with t varying in [ti , ti+1 ]. The two quadratic segments are given by convex combinations of linear segments on the two intervals [ti−1 , ti+1 ] and [ti , ti+2 ], as in (1.20). The three line segments are in turn given by convex combinations of the given points on the intervals [ti−2 , ti+1 ], [ti−1 , ti+2 ] and [ti , ti+3 ]. Note that all these intervals contain [ti , ti+1 ] so that when t varies in [ti , ti+1 ] all the combinations involved in the construction of the cubic curve will be convex. This also shows that we can never get division by zero since we have assumed that ti < ti+1 . The explicit notation in (1.23) is too cumbersome, especially when we consider spline curves of even higher degrees, so we generalise the notation in (1.19) and (1.22) and set pi,k,s (t) = p(t | ci−k , . . . , ci , ti−k+1 , . . . , ti , ti+s , . . . , ti+k+s−1 ), (1.24) for positive s and r, assuming that the control points and knots in question are given. The first subscript i in pi,k,s indicates which control points and knots are involved (in general we work with many spline segments and therefore long arrays of control points and knots), the second subscript k gives the polynomial degree, and the last subscript s, gives the gap between the knots in the computation of the weight (t − ti )/(ti+s − ti ). With the abbreviation (1.24), equation (1.23) becomes pi,3,1 (t) = ti+1 − t t − ti pi−1,2,2 (t) + p (t). ti+1 − ti ti+1 − ti i,2,2 Note that on both sides of this equation, the last two subscripts sum to four. Similarly, if the construction of quadratic splines given by (1.20) is expressed with the abbreviation given in (1.24), the last two subscripts add to three. The general pattern is that in the 26 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION ci-3 t i+1 pi-2,1 t i+1 t ti+1 -ti-2 t t- t i- 2 pi-1,2 t i+1 ci-2 tt+1 -ti-1 t t i+2 t- t t i- 1 pi,3 pi-1,1 ti+1 -ti t i+ t- t ti+2 -ti-1 t 2- t- t i- 1 i pi,2 ci-1 ti+2 -ti t i+ t- t t 3- i pi,1 ti+3 -ti t- t i ci Figure 1.21. Computing a point on a cubic spline curve. recursive formulation of spline curves of degree d, the last two subscripts always add to d + 1. Therefore, when the degree of the spline curves under construction is fixed we can drop the third subscript and write pi,k,s = pi,k . The complete computations involved in computing a point on the cubic segment pi,3 (t) can be arranged in the triangular array shown in Figure 1.21 (all arguments to the pi,k have been omitted to conserve space). The labels should be interpreted as in Figure 1.7. A segment of a general spline curve of degree d requires d + 1 control points (cj )ij=i−d and 2d knots (tj )i+d j=i−d+1 that form a nondecreasing sequence with ti < ti+1 . The curve is a weighted average of two curves of degree d − 1, pi,d (t) = t − ti ti+1 − t pi−1,d−1 (t) + p (t). ti+1 − ti ti+1 − ti i,d−1 (1.25) Because of the assumption ti < ti+1 we never get division by zero in (1.25). The two curves of degree d − 1 are obtained by forming similar convex combinations of curves of degree d − 2. For example, pi,d−1 (t) = ti+2 − t t − ti pi−1,d−2 (t) + p (t), ti+2 − ti ti+2 − ti i,d−2 and again the condition ti < ti+1 saves us from dividing by zero. At the lowest level we have d line segments that are determined directly from the control points, pj,1 (t) = tj+d − t t − tj cj−1 + cj tj+d − tj tj+d − tj for j = i − d + 1, . . . , i. The denominators in this case are ti+1 − ti−d+1 , . . . , ti+d − ti , all of which are positive since the knots are nondecreasing with ti < ti+1 . As long as 1.5. A GEOMETRIC CONSTRUCTION OF SPLINE CURVES (a) 27 (b) Figure 1.22. Two cubic spline curves, both with knots t = (0, 0, 0, 0, 1, 2, 3, 3, 3, 3). t is restricted to the interval [ti , ti+1 ], all the operations involved in computing pi,d (t) are convex combinations. The complete computations are summarised in the following algorithm. Algorithm 1.3. Let d be a positive integer and let the d + 1 points (cj )dj=i−d be given together with the 2d knots t = (tj )i+d j=i−d+1 . The point pi,d (t) on the spline curve pi,d of degree d is determined by the following computations. First set pj,0 (t) = cj for j = i − d, i − d + 1, . . . , i and then compute pj,r (t) = tj+d−r+1 − t t − tj pj−1,r−1 (t) + p (t) tj+d−r+1 − tj tj+d−r+1 − tj j,r−1 (1.26) for j = i − d + r, . . . , i and r = 1, 2, . . . , d. A spline curve of degree d with n control points (ci )ni=1 and knots (ti )n+d i=2 is given by pd+1,d (t) t ∈ [td+1 , td+2 ], p d+2,d (t), t ∈ [td+2 , td+3 ]; f (t) = .. .. . . pn,d (t), t ∈ [tn , tn+1 ], where as before it is assumed that the knots are nondecreasing and in addition that ti < ti+1 for i = d + 1, . . . , n. Again we can express f in terms of the piecewise constant functions given by (1.18), n X f (t) = pi,d (t)Bi,0 (t). (1.27) i=d+1 It turns out that spline curves of degree d have continuous derivatives up to order d − 1, see Section 3.2.3 in Chapter 3. Figure 1.22 shows two examples of cubic spline curves with control points taken from the two Bézier curves of degree five in Figure 1.15. Again we note that the curves behave like Bézier curves at the ends because there are four identical knots at each end. 28 CHAPTER 1. SPLINES AND B-SPLINES (a) AN INTRODUCTION (b) Figure 1.23. A quadratic spline with a double knot at the circled point (a) and a cubic spline with a double knot at the circled point (b). 1.5.4 Smoothness of spline curves The geometric construction of one segment of a spline curve, however elegant and numerically stable it may be, would hardly be of much practical interest was it not for the fact that it is possible to smoothly join together neighbouring segments. We will study this in much more detail in Chapter 3, but will take the time to state the exact smoothness properties of spline curves here. Theorem 1.4. Suppose that the number ti+1 occurs m times among the knots (tj )m+d j=i−d , with m some integer bounded by 1 ≤ m ≤ d + 1, i.e., ti < ti+1 = · · · = ti+m < ti+m+1 . Then the spline function f (t) = pi,d,1 (t)Bi,0 (t) + pi+m,d,1 (t)Bi+m,0 (t) has continuous derivatives up to order d − m at the join ti+1 . This theorem introduces a generalisation of our construction of spline curves by permitting ti+1 , . . . , ti+m to coalesce, but if we assume that m = 1 the situation corresponds to the construction above. Theorem 1.4 tells us that in this standard case the spline curve f will have d continuous derivatives at the join ti+1 : namely f , f 0 , . . . , f d−1 will all be continuous at ti+1 . This means that if the knots are all distinct, then a linear spline will be continuous, a quadratic spline will also have a continuous first derivative, while for a cubic spline even the second derivative will be continuous. Examples of spline curves with this maximum smoothness can be found above. What happens when m > 1? Theorem 1.4 tells us that each time we add a knot at ti+1 the number of continuous derivatives is reduced by one. So a quadratic spline will in general only be continuous at a double knot, whereas a cubic spline will be continuous and have a continuous derivative at a double knot. This ability to control the smoothness of a spline by varying the multiplicity of the knots is important in practical applications. For example it is often necessary to represent curves with a sharp corner (discontinuous derivative). With a spline curve of degree d this can be done by letting the appropriate knot occur d times. We will see many examples of how the multiplicity of the knots influence the smoothness of a spline in later chapters. Two examples of spline curves with reduced smoothness are shown in Figure 1.23. 1.6. REPRESENTING SPLINE CURVES IN TERMS OF BASIS FUNCTIONS 29 Figure (a) shows a quadratic spline with a double knot and a discontinuous derivative at the encircled point, while Figure (b) shows a cubic spline with a double knot and a discontinuous second derivative at the encircled point. 1.6 Representing spline curves in terms of basis functions In Section 1.4 we saw that a Bézier curve g of degree d with control points (ci )di=0 can be written as a linear combination of the Bernstein polynomials {bi,d }di=0 with the control points as coefficients, see (1.14). In this section we want to develop a similar representation for spline curves. n+d If we have n control points (ci )ni=1 and the n + d − 1 knots t = (ti )i=2 for splines of degree d; we have seen that a typical spline can be written f (t) = n X pi,d (t)Bi,0 (t), t ∈ [td+1 , tn+1 ], (1.28) i=d+1 where {Bi,0 }ni=d+1 are given by (1.18). When this representation was introduced at the end of Section 1.5.3 we assumed that td+1 < td+2 < · · · < tn+1 (although the end knots were allowed to coincide). To accommodate more general forms of continuity, we know from Theorem 1.4 that we must allow some of the interior knots to coincide as well. If for example ti = ti+1 for some i with d + 1 < i < n + 1, then the corresponding segment pi,d is completely redundant and (1.25) does not make sense since we get division by zero. This is in fact already built into the representation in (1.28), since Bi,0 (t) is identically zero in this case, see (1.18). A more explicit definition of Bi,0 makes this even clearer, 1, ti ≤ t < ti+1 , Bi,0 (t) = 0, t < ti or t ≥ ti+1 , (1.29) 0, ti = ti+1 . The representation (1.28) is therefore valid even if some of the knots occur several times. The only complication is that we must be careful when we expand out pi,d according to (1.25) as this will give division by zero if ti = ti+1 . One might argue that there should be no need to apply (1.25) if ti = ti+1 since the result is zero anyway. However, in theoretical developments it is convenient to be able to treat all the terms in (1.28) similarly, and this may then lead to division by zero. It turns out though that this problem can be circumvented quite easily by giving an appropriate definition of ‘division by zero’ in this context, see below. Let us now see how f can be written more directly in terms of the control points. By making use of (1.25) we obtain f (t) = n X i=d+1 = n−1 X i=d+1 ti+1 − t t − ti pi,d−1 (t)Bi,0 (t) + p (t)Bi,0 (t) ti+1 − ti ti+1 − ti i−1,d−1 t−t ti+2 − ti i Bi,0 (t) + Bi+1,0 (t) pi,d−1 (t)+ ti+1 − ti ti+2 − ti+1 td+2 − t t − tn Bd+1,0 (t)pd,d−1 (t) + Bn,0 (t)pn,d−1 (t). td+2 − td+1 tn+1 − tn (1.30) 30 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION This is a typical situation where we face the problem of division by zero if ti = ti+1 for some i. The solution is to declare that ‘anything divided by zero is zero’ since we know that if ti = ti+1 the answer should be zero anyway. In (1.30) we have two ‘boundary terms’ that complicate the expression. But since t is assumed to lie in the interval [td+1 , tn+1 ] we may add the expression tn+2 − t t − td Bd,0 (t)pd,d−1 (t) + Bn+1,0 (t)pn,d−1 (t) td+1 − td tn+1 − tn+1 which is identically zero as long as t is within this [td+1 , tn+1 ]. By introducing the functions Bi,1 (t) = t − ti ti+2 − t Bi,0 (t) + Bi+1,0 (t) ti+1 − ti ti+2 − ti+1 (1.31) for i = d, . . . , n, we can then write f as f (t) = n X pi,d−1 (t)Bi,1 (t). i=d This illustrates the general strategy: Successively apply the relations in (1.26) in turn and rearrange the sums until we have an expression where the control points appear explicitly. The functions that emerge are generalisations of Bi,1 and can be defined recursively by Bi,r (t) = ti+r+1 − ti t − ti Bi,r−1 (t) + Bi+1,r−1 (t), ti+r − ti ti+r+1 − ti (1.32) for r = 1, 2, . . . , d, starting with Bi,0 as defined in (1.18). Again we use the convention that ‘anything divided by zero is zero’. It follows by induction that Bi,r (t) is identically zero if ti = ti+r+1 and Bi,r (t) = 0 if t < ti or t > ti+r+1 , see exercise 7. To prove by induction that the functions defined by the recurrence (1.32) appear in the process of unwrapping all the averaging in (1.26), we consider a general step. Suppose that after r − 1 applications of (1.26) we have f (t) = n X pi,d−r+1 (t)Bi,r−1 (t). i=d+2−r One more application yields f (t) = n X i=d+2−r = n−1 X i=d+2−r ti+r − t t − ti pi−1,d−r (t)Bi,r−1 (t) + p (t)Bi,r−1 (t) ti+r − ti ti+r − ti i,d−r t−t ti+r+1 − t i Bi,r−1 (t) + Bi+1,r−1 (t) pi,d−r (t)+ ti+r − ti ti+r+1 − ti+1 td+2 − t t − tn Bd+2−r,r−1 (t)pd+1−r,d−r (t) + Bn,r−1 (t)pn,d−r (t). td+2 − td+2−r tn+r − tn Just as above we can include the boundary terms in the sum by adding t − td+1−r tn+r+1 − t Bd+1−r,r−1 (t)pd+1−r,d−r (t) + Bn+1,r−1 (t)pn,d−r (t) td+1 − td+1−r tn+r+1 − tn+1 1.6. REPRESENTING SPLINE CURVES IN TERMS OF BASIS FUNCTIONS 31 which is zero since Bi,r−1 (t) is zero when t < ti or t > ti+r . The result is that n X f (t) = pi,d−r (t)Bi,r (t). i=d+1−r P After d − 1 steps we have f (t) = ni=2 pi,1,d−1 (t)Bi,d−1 (t). In the last application of (1.26) we recall that pj,0 (t) = cj for j = i − d, . . . , i. After rearranging the sum and adding zero terms as before we obtain n X f (t) = ci Bi,d (t). i=1 But note that in this final step we need two extra knots, namely t1 and tn+d+1 which are used by B1,d−1 and Bn+1,d−1 , and therefore also by B1,d and Bn,d . The value of the spline in the interval [td+1 , tn+1 ] is independent of these knots, but it is customary to demand that t1 ≤ t2 and tn+d+1 ≥ tn+d to ensure that the complete knot vector t = (ti )n+d+1 is a i=1 nondecreasing sequence of real numbers. The above discussion can be summarised in the following theorem. Theorem 1.5. Let (ci )ni=1 be a set of control points for a spline curve f of degree d, with nondecreasing knots (ti )n+d+1 , i=1 f (t) = n X pi,d (t)Bi,0 (t) i=d+1 where pi,d is given recursively by pi,d−r+1 (t) = t − ti ti+r − t pi−1,d−r (t) + p (t) ti+r − ti ti+r − ti i,d−r (1.33) for i = d − r + 1, . . . , n, and r = d, d − 1, . . . , 1, while pi,0 (t) = ci for i = 1, . . . , n. The functions {Bi,0 }ni=d+1 are given by ( 1, ti ≤ t < ti+1 , Bi,0 (t) = (1.34) 0, otherwise. The spline f can also be written f (t) = n X ci Bi,d (t) (1.35) ti+1+d − ti+1 t − ti Bi,d−1 (t) + Bi+1,d−1 (t). ti+d − ti ti+1+d − ti+1 (1.36) i=1 where Bi,d is given by the recurrence relation Bi,d (t) = In both (1.33) and (1.36) possible divisions by zero are resolved by the convention that ‘anything divided by zero is zero’. The function Bi,d = Bi,d,t is called a B-spline of degree d (with knots t). B-splines have many interesting and useful properties and in the next chapter we will study these functions in detail. 32 CHAPTER 1. SPLINES AND B-SPLINES 1.7 AN INTRODUCTION Conclusion Our starting point in this chapter was the need for efficient and numerically stable methods for determining smooth curves from a set of points. We considered three possibilities, namely polynomial interpolation, Bézier curves and spline curves. In their simplest forms, all three methods produce polynomial curves that can be expressed as g(t) = d X ai Fi (t), i=0 where d is the polynomial degree, (ai )di=0 are the coefficients and {Fi }di=0 are the basis polynomials. The difference between the three methods lie in the choice of basis polynomials, or equivalently, how the given points relate to the final curve. In the case of interpolation the coefficients are points on the curve with the Lagrange polynomials as basis polynomials. For Bézier and spline curves the coefficients are control points with the property that the curve itself lies inside the convex hull of the control points, while the basis polynomials are the Bernstein polynomials and (one segment of) B-splines respectively. Although all three methods are capable of generating any polynomial curve, their differences mean that they lead to different representations of polynomials. For our purposes Bézier and spline curves are preferable since they can be constructed by forming repeated convex combinations. As we argued in Section 1.1, this should ensure that the curves are relatively insensitive to round off errors. The use of convex combinations also means that the constructions have simple geometric interpretations. This has the advantage that a Bézier curve or spline curve can conveniently be manipulated interactively by manipulating the curve’s control points, and as we saw in Section 1.4.3 it also makes it quite simple to link several Bézier curves smoothly together. The advantage of spline curves over Bézier curves is that smoothness between neighbouring polynomial pieces is built into the basis functions (B-splines) instead of being controlled by constraining control points according to specific rules. In the coming chapters we are going to study various aspects of splines, primarily by uncovering properties of B-splines. This means that our point of view will be shifted somewhat, from spline curves to spline functions (each control point is a real number), since B-splines are functions. However, virtually all the properties we obtain for spline functions also make sense for spline curves, and even tensor product spline surfaces, see Chapters 6 and 7. We were led to splines and B-splines in our search for approximation methods based on convex combinations. The method which uses given points (ci )ni=1 as control points for a spline as in n X f (t) = ci Bi,d (t) (1.37) i=1 is often referred to as Schoenberg’s variation diminishing spline approximation. This is a widely used approximation method that we will study in detail in Section 5.4, and because of the intuitive relation between the spline and its control points the method is often used in interactive design of spline curves. However, there are many other spline approximation methods. For example, we may approximate certain given points (bi )m i=1 by a spline curve that passes through these points, or we may decide that we want a spline curve that 1.7. CONCLUSION 33 approximates these points in such a way that some measure of the error is as small as possible. To solve these kinds of problems, we are faced with three challenges: we must pick a suitable polynomial degree and an appropriate set of knots, and then determine control points so that the resulting spline curve satisfies our chosen criteria. Once this is accomplished we can compute points on the curve by Algorithm 1.3 and store it by storing the degree, the knots and the control points. We are going to study various spline approximation methods of this kind in Chapter 5. But before turning to approximation with splines, we need to answer some basic questions: Exactly what functions can be represented as linear combinations of B-splines as in (1.37)? Is a representation in terms of B-splines unique, or are there several choices of control points that result in the same spline curve? These and many other questions will be answered in the next two chapters. Exercises for Chapter 1 1.1 Recall that a subset A of Rn is said to be convex if whenever we pick two points in A, the line connecting the two points is also in A. In this exercise we are going to prove that the convex hull of a finite set of points is the smallest convex set that contains the points. This is obviously true if we only have one or two points. To gain some insight we will first show that it is also true in the case of three points before we proceed to the general case. We will use the notation CH(c1 , . . . , cn ) to denote the convex hull of the points c1 , . . . , cn . a) Suppose we have three points c1 , c2 and c3 . We know that the convex hull of c1 and c2 is the straight line segment that connects the points. Let c̃ be a point on this line, i.e., c̃ = (1 − λ)c1 + λc2 (1.38) for some λ with 0 ≤ λ ≤ 1. Show that any convex combination of c̃ and c3 is a convex combination of c1 , c2 and c3 . Explain why this proves that CH(c1 , c2 , c3 ) contains the triangle with the three points at its vertices. The situation is depicted graphically in Figure 1.2. b) It could be that CH(c1 , c2 , c3 ) is larger than the triangle formed by the three points since the convex combination that we considered above was rather special. We will now show that this is not the case. Show that any convex combination of c1 , c2 and c3 gives rise to a convex combination on the form (1.38). Hint: Show that if c is a convex combination of the three points, then we can write c = λ 1 c 1 + λ2 c2 + λ 3 c 3 = (1 − λ3 )c̃ + λ3 c3 , where c̃ is some convex combination of c1 and c2 . Why does this prove that the convex hull of three points coincides with the triangle formed by the points? Explain why this shows that if B is a convex set that contains c1 , c2 and c3 then B must also contain the convex hull of the three points which allows us to conclude that the convex hull of three points is the smallest convex set that contains the points. 34 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION c) The general proof that the convex hull of n points is the smallest convex set that contains the points is by induction on n. We know that this is true for n = 2 and n = 3 so we assume that n ≥ 4. Let B be a convex set that contains c1 , . . . , cn . Use the induction hypothesis and show that B contains any point on a straight line that connects cn and an arbitrary point in CH(c1 , . . . , cn−1 ). d) From what we have found in (c) it is not absolutely clear that any convex set B that contains c1 , . . . , cn also contains all convex combinations of the points. To settle this show that any point c in CH(c1 , . . . , cn ) can be written c = λc̃ + (1 − λ)cn for some λ in [0, 1] and some point c̃ in CH(c1 , . . . , cn−1 ). Hint: Use a trick similar to that in (b). Explain why this lets us conclude that CH(c1 , . . . , cn ) is the smallest convex set that contains c1 , . . . , cn . 1.2 Show that the interpolatory polynomial curve q 0,d (t) given by (1.4) can be written as in (1.5) with `i,d given by (1.6). 1.3 Implement Algorithm 1.1 in a programming language of your choice. Test the code by interpolating points on a semicircle and plot the results. Perform four tests, with 3, 7, 11 and 15 uniformly sampled points. Experiment with the choice of parameter values (ti ) and try to find both some good and some bad approximations. 1.4 Implement Algorithm 1.2 in your favourite programming language. Test the program on the same data as in exercise 3. 1.5 In this exercise we are going to write a program for evaluating spline functions. Use whatever programming language you prefer. a) Implement Algorithm 1.3 in a procedure that takes as input an integer d (the degree), d + 1 control points in the plane, 2d knots and a parameter value t. P n+d+1 b) If we have a complete spline curve f = ni=1 ci Bi,d with knots t = (ti )i=1 that we want to evaluate at t we must make sure that the correct control points and knots are passed to the routine in (a). If tµ ≤ t < tµ+1 (1.39) then (ci )µi=µ−d and (ti )µ+d i=µ−d+1 are the control points and knots needed in (a). Write a procedure which takes as input all the knots and a value t and gives as output the integer µ such that (1.39) holds. c) Write a program that plots a spline function by calling the two routines from (a) and (b). Test your program by picking control points from the upper half of the unit circle and plotting the resulting spline curve. Use cubic splines and try with n = 4, n = 8 and n = 16 control points. Use the knots t = (0, 0, 0, 0, 1, 1, 1, 1) when n = 4 and add the appropriate number of knots between 0 and 1 when n is increased. Experiment with the choice of interior knots when n = 8 and n = 16. Is the resulting curve very dependent on the knots? 1.6 Show that a quadratic spline is continuous and has a continuous derivative at a single knot. 1.7. CONCLUSION 35 1.7 Show by induction that Bi,d depends only on the knots ti , ti+1 , . . . , ti+d+1 . Show also that Bi,d (t) = 0 if t < ti or t > ti+d+1 . 36 CHAPTER 1. SPLINES AND B-SPLINES AN INTRODUCTION

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement