Adaptive and local regularization for data fitting by tensor-product spline surfaces (Q6171718)

When approximating noisy data with a spline surface, it is often necessary to add a regularization term to the objective function. Changing its weight, allows for a user selected trade-off between surface quality and approximation error. This approach works well for data with a regular distribution of noise over the parameter domain but has shortcomings in the presence of irregularities like holes or sharp features. This article proposes to replace the globally constant weight by a non-constant regularization weight function (RWF) \(\lambda\) that makes possible a local control of regularization. The authors consider the case of \(\lambda\) being a classical spline function but also the discrete limiting case where \(\lambda\) is a sum of Dirac delta functions. One of the benefits of regularization is to ensure uniqueness of the optimal solution. Some mild assumptions on \(\lambda\) in the classical smooth case guarantee this also in the setting of this paper. Also an error estimate and optimal convergence results are available in the smooth setting. On the downside, the condition number of the approximation problem increases with smaller mesh size. The resulting numerical problems could be alleviated by more advanced regularization techniques or by using a lower degree spline function, thus sacrificing optimal convergence. The important case for industrial applications is that of a discrete RWF for which above results no longer apply. (Obtaining comparable results also for the discrete setting is mentioned as possible topic of future research.) Two methods for the construction of suitable RWFs are proposed. The first one is adapted to data with holes. The second one involves a more sophisticated iterative procedure and is suitable for data with varying sampling density, for example sharp features. Both approaches are discussed and compared at hand of artificial and practical examples. The bottom line is that they do what they were designed for. In the discussed scenarios, they outperform traditional regularization with constant weight, provided some heuristic advise on setting parameters is followed.