1. Introduction
Many of the spline based solutions to multivariate estimation problems involve tensor product spaces that by necessity depend on the choice of coordinate system (see, for example, Friedman 1991; Gu 1993; Kooperberg, Bose and Stone 1995; Kooperberg, Stone and Truong 1995; Stone, Hansen, Kooperberg and Truong 1995). Since these procedures are also highly adaptive, the estimates can change significantly when the data are rotated, a property that is clearly undesirable when the given coordinate system is arbitrary, as is the case with spatial or compositional data. Moreover, the tensor product structure of these spaces implicitly defines the domain of an unknown function to be a hyper-rectangle and can restrict the resulting estimators from capturing major features in the data that are not oriented along one of the major axes.In the last 15 years, very few applications of multivariate splines other than tensor product spaces have appeared in statistical journals. During the same period, however, approximation theorists, numerical analysts and computer scientists have amassed a considerable body of literature on constructing and representing smooth, piecewise polynomial surfaces over meshes in many variables. When the underlying partitions consist of triangles or high-dimensional simplices, the Bernstein-Bezier (or B -net) representation has been extremely useful in characterizing multivariate spline bases (Chui 1988; de Boor 1987; Farin 1986). In this paper we develop a procedure for bivariate function estimation using a rotation invariant spline basis built from barycentric coordinate functions, a construction that is at the core of the B -net representation.
Our estimates, named Triograms, are continuous, piecewise linear functions defined over adaptively selected triangulations in the plane. The fitting is done via maximum likelihood, and the methodology can be applied to any of the so-called extended linear models, including density and conditional density estimation, generalized linear models, polychotomous regression, and hazard regression (Stone et al. 1995). In a process similar in spirit to stepwise knot addition and deletion in a univariate spline space, the underlying triangulation is constructed adaptively by adding and deleting vertices. These computations are made efficient through the use of the Rao (score) statistic for addition and the Wald statistic for deletion.
While the focus of our paper is on estimating bivariate functions, Triograms can be applied much more generally. For example, let denote an unknown function of three variables. Following the ANOVA framework developed by Hastie and Tibshirani (1990) and Stone (1994) for multivariate function estimation, subject to certain identifiability constraints, we can write
By ignoring higher order interaction terms in this expansion, the convergence rate of the remaining problem is governed by the largest of the dimensions of the terms that are estimated, taming the ``curse of dimensionality." Polynomial splines can be used to model the main effect spaces and tensor products of univariate spline spaces can be used to estimate the various interactions in the ANOVA decomposition. Alternately, we can use the Triogram method presented in our paper to model any of the two-dimensional components, or two-factor interactions. While we ignore this possibility in most of the Triogram paper, this is an important consideration in the theoretical framework for extended linear models that was developed by Hansen (1994) and is summarized in Section 5 of the Triogram paper (PostScript version).
In this online manuscript, you will find a somewhat abreviated discussion of the Triogram methodology in Section 2. In the remaining three sections of this document we discuss some applications of the Triogram methodology, focusing our attention on regression and density estimation. We have found that static displays of our Triogram fits can be somewhat unsatisfying, and have designed this site for interested researchers to learn more about Triograms. We have included a number of interactive Java applets to make this learning process easier. We have decided not to put the entire paper online in this interactive form, but invite you to pick up a PostScript version of our paper either from Charles Kooperberg's home page in Seattle or from Mark Hansen's publication list .
The PostScript version of our paper is organized as follows. In Section 2 we define barycentric coordinate functions and derive some basic facts to motivate their usefulness in representing polynomials over triangles in the plane. Triogram models are first introduced in Section 3 together with general stepwise algorithms that are used to adaptively refine the underlying spline spaces. The barycentric coordinate functions again prove convenient for computing the various statistics required to perform this adaptation. In Section 4 we present a number of examples in which Triograms are used to construct bivariate regression and bivariate log-density estimates. As mentioned above, we reserve Section 5 for more technical remarks concerning theoretical rates of convergence for Triograms in the context of an extended linear model. Finally, in Section 6 we give some concluding remarks.
If you have any comments on either this (eeee)lectronic manuscript or on our PostScript paper, please send e-mail to either Mark Hansen or Charles Kooperberg . The interactive plotting routines like the one on the previous page began with a piece of Java code distributed by James Gosling at Sun Microsystems for viewing wire frame models. Many moons of Java programming later, we have the interactive Triogram displays you will encounter as you read through this document. Rather than use Gif versions of LaTeX equations, our equations are also done through Java using the WebEQ applets distributed by Robert Miner at the Geometry Center.
You are currently reading Section 1, the introduction.
- Introduction and acknowledgements
- A short description of Triogram modeling
- A regression example with simulated data
- A bivariate density estimation example
- A regression surface with a ridge
- References
- A PostScript version of Triogram Models
- Send us comments !
You can also bail out and head back to my home page .