Normal multi-scale transforms for curves (Q656816)

The \textit{normal multi-scale transform} gives a recursive prediction-correction method of representing a curve \({\mathcal C}\) in the plane. Starting with a set of samples \({\mathbf v}^0\) of \({\mathcal C}\), the \(j\)-th level of recursion is obtained from the previous level by applying a fixed subdivision scheme \(S\) to \({\mathbf v}^{j-1}\) to obtain prediction points \(S({\mathbf v}^{j-1})\) for \({\mathcal C}\). For each prediction point \(\bigl(S({\mathbf v}^{j-1})\bigr)_i\), choose a normal direction \(\hat{\mathbf n}_i^j\), and find the intersection point of \({\mathcal C}\) and the line passing through the prediction point in this direction: \({\mathbf v}_i^j=\bigl(S({\mathbf v}^{j-1})\bigr)_i+d_i^j\hat{\mathbf n}_i^j\). The normal multi-scale transform of \({\mathcal C}\) then consists of the initial sample points \({\mathbf v}^0\) and the \textit{detail coefficients} \(d^j\) from each level \(j\geq 1\). Since there is not necessarily a unique intersection point between the line and the curve \({\mathcal C}\), the above construction is not well-defined in general. The authors of the article show that the normal multi-scale transform is well-defined provided that \({\mathcal C}\) is a regular \(C^k\) curve with Lipschitz order \(\alpha\leq 1\), the spacing between the initial samples is sufficiently small, and the normals are reasonably chosen. A bound on the size of the detail coefficients is also obtained. In addition, the specific case when \(S\) is the Chaikin corner-cutting subdivision scheme is discussed in depth, and the results of numerical experiments using higher order \(B\)-spline subdivision schemes are given. The article is intended for a general audience familiar with numerical analysis and subdivision schemes. The article is well-written, although the proof of the main result is lengthy and intricate.