Kevin W. Aiton,Tobin A. Driscoll
Abstract:Spectral polynomial approximation of smooth functions allows real-time manipulation of and computation with them, as in the Chebfun system. Extension of the technique to two-dimensional and three-dimensional functions on hyperrectangles has mainly focused on low-rank approximation. While this method is very effective for some functions, it is highly anisotropic and unacceptably slow for many functions of potential interest. A method based on automatic recursive domain splitting, with a partition of unity to define the global approximation, is easy to construct and manipulate. Experiments show it to be as fast as existing software for many low-rank functions, and much faster on other examples, even in serial computation. It is also much less sensitive to alignment with coordinate axes. Some steps are also taken toward approximation of functions on nonrectangular domains, by using least-squares polynomial approximations in a manner similar to Fourier extension methods, with promising results.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is to perform efficient and accurate approximation of complex functions in high - dimensional spaces, especially in view of the anisotropy problems encountered by existing methods such as Chebfun2 and Chebfun3 when dealing with certain functions. Specifically:
1. **Anisotropy problem**: Existing low - rank approximation methods (such as Chebfun2 and Chebfun3) exhibit high anisotropy when dealing with some functions, that is, the approximation effect of the function depends heavily on the direction of the coordinate axes. For example, rotating the coordinate axes may make a originally low - rank function become high - rank, thereby significantly increasing the computation time and complexity.
2. **Function approximation on non - rectangular domains**: For functions defined on non - rectangular domains, existing low - rank or sparse - grid approximation methods lack an effective global generalization method. This limits their use in more extensive application scenarios.
To solve these problems, the author proposes a new technique based on the adaptive partition of unity method. This method recursively divides the hyper - rectangular domain into multiple sub - domains and uses tensor - product Chebyshev polynomial interpolation to approximate the function on each sub - domain. Then, these local approximations are combined into a global approximation through a partition of unity function. This method has the following advantages:
- **Reducing anisotropy**: Since the division of sub - domains is adaptive and can have a smooth transition through the overlapping area, the new method is insensitive to changes in the direction of the coordinate axes.
- **Improving computational efficiency**: Experiments show that this method is faster than existing low - rank approximation methods in many cases, especially when dealing with high - rank functions.
- **Extension to non - rectangular domains**: By introducing the least - squares polynomial approximation, this method can effectively handle the problem of function approximation on non - rectangular domains.
In conclusion, this paper aims to provide a new, more flexible and efficient method for approximating multivariate functions to overcome the limitations of existing methods in anisotropy and non - rectangular domain approximation.