Farklı yüksek boyutlu model gösterilim algoritmalarının çok değişkenli interpolasyon uygulamaları

Bu çalışmada N adet bağımsız değişkene bağlı olan birçok değişkenli fonksiyonun değerlerinin bağımsız değişkenlerin sonlu sayıda değer takımı için verildiği ve fonksiyonun analitik yapısının istendiği çok değişkenli interpolasyon problemlerinin daha az değişkenli interpolasyon problemlerine indirgenmesi amaçlanmıştır. Böylelikle, hesaplama karmaşıklığı düşürülecek ve problemin bilgisayar ortamında programlanması da kolaylaşacaktır. Bu amaçla, N değişkenli bir interpolasyon problemi N adet tek bağımsız değişkenli interpolasyon problemi haline getirilmektedir. Belirtilen indirgeme için ilk olarak I.M. Sobol tarafından tasarlanan Yüksek Boyutlu Model Gösterilim (YBMG) yöntemi geliştirilmiştir. Bu yöntem çok değişkenli verinin hiperprizmatik örgünün tüm düğümlerinde verildiği problemlerde veri bölümlemesinde kullanılmaktadır. Bölümleme sonucunda elde edilen tek değişkenli veri kümesinden çok değişkenli fonksiyon için aranılan analitik yapı yaklaşık olarak elde edilebilmektedir. Yöntemin temel felsefesini oluşturan sonlu terimden oluşan açılımın yapısı baskın olarak toplamsal özellikler taşıyan çokdeğişkenli veri kümelerine ait interpolasyon problemlerinde gerçek sonuca yakın gösterilimler elde etmeyi sağlamaktadır. Aranılan analitik yapı, yani verilen çok değişkenli veri kümesinin yapısı, toplamsal özelliklerden uzaklaşıp çarpımsal veya melez özelliklere sahip olmaya başladığında YBMG yönteminin verimi düşmektedir. Bu bağlamda alternatif yöntemlere ihtiyaç duyulmaktadır. Bu amaçla, problemde verilen veri takımının yapısına göre Çarpımsallaştırılmış Yüksek Boyutlu Model Gösterilim (ÇYBMG) ve Melez Yüksek Boyutlu Model Gösterilim (MYBMG) yöntemleri de oluşturulmuştur. Belirtilen bu yöntemler YBMG yöntemi aracılığıyla bölümlenmiş veriyi kullanarak çarpımsal veya melez yapıya sahip fonksiyonlar için daha iyi yaklaşıklık elde eden gösterilimler oluşturmayı hedeflemektedir.

Multivariate interpolation applications of different high dimensional model representations

In this work, the main purpose is to reduce the multivariate interpolation problems to the less-variate interpolation problems in which the values of a multivariate function having N number of independent variables are given for a finite number of data and it is asked to determine an analytical structure for this function. As a result, the computational complexity of the problem will decrease and it will become easier to write programs for the computer-based applications. For this purpose, a package of N number of univariate interpolation problems is constructed from a N dimensional interpolation problem. High Dimensional Model Representation (HDMR) method is developed for the mentioned reduction process of the interpolation problem to determine approximate representation for the analytical structure of the sought function. HDMR is a divide – conquer method and was first proposed by I.M. Sobol, then generalized by H. Rabitz. HDMR has an expansion for a given multivariate function such that its components are ordered starting from a constant component (zeroth order multivariance) and continuing in ascending multivariance, that is, univariate, bivariate, trivariate components and so on. Components of this representation are determined by using an imposition of vanishing integrals. Since the main purpose of this work is to partition the given multivariate data into lower variate data, HDMR algorithm is reconstructed for data partitioning. This new method can be used for partitioning the data of multivariate interpolation problems in which the values of the sought function are given at all nodes of the hyperprismatic grid. Using these partitioned data the analytical structure for the sought function is obtained through Lagrange interpolation formula. When the nature of the HDMR expansion given below and the numerical implementations are examined it is seen that new methods are needed to obtain better approximate representations when the sought function does not have purely or dominantly additive nature. Hence, it can be said that the nature of the sought multivariate function and the features of the given data set have characteristic roles on the development of these methods. The sought function may have a multiplicative or an intermediate nature. Certain other methods are developed for interpolation problems having these types of structures. Factorized form of the HDMR method is called Factorized High Dimensional Model Representation (FHDMR). This method has a multiplicative expansion and the components of FHDMR expansion are evaluated by making comparisons between the HDMR and the FHDMR expansions of the multivariate function. To construct a unique comparison procedure certain idempotent operators are inserted into the HDMR expansion. After inserting these mentioned operators and expanding the FHDMR expansion into an additive expression, relations for FHDMR components of the multivariate function can be obtained in terms of the components of the data partitioning technique. In most cases the nature of the given multivariate data and the sought multivariate function have neither a purely additive nor a purely multiplicative nature. They have a hybrid nature. So, a new method is developed to obtain better results and it is called Hybrid High Dimensional Model Representation (HHDMR). This new method has an expansion including both the HDMR and the FHDMR expansions of the multivariate function through a hybridity parameter. The main problem in this method is to determine the best value for this parameter to obtain the best representation in the given interpolation problem. A cost functional is defined to obtain this mentioned value for the hybridity parameter. Another cost functional is defined to find the best representation obtained through three methods that were mentioned; HDMR, FHDMR and HHDMR for the sought multivariate function. Several numerical implementations are also given in this paper to test the efficiency of all these three methods. Various test functions are selected to examine the performance of the given methods. When the norm values, defined for finding the best representation, obtained for each representation method in the implementations are examined the best representation for the purely or dominantly additive functions are obtained through HDMR method. If the sought function has a purely or dominantly multiplicative nature, FHDMR method gives the best representation. On the other hand when the sought function has an intermediate nature then HHDMR method is needed to determine a better representation.

___

  • 1.Sobol, I. M., (1993). Sensitivity estimates for nonlinear mathematical models, Mathematical Modeling and Computational Experiments, 1, 407-414.
  • 2.Sobol, I. M., (2003). Theorems and examples on high dimensional model representation, Reliability Engineering and System Safety, 79, 187-193.
  • 3.Rabitz, H. ve Alış, Ö. F., (1999). General foundations of high dimensional model representations, Journal of Mathematical Chemistry, 25, 197-233.
  • 4.Alış, Ö. F. ve Rabitz, H., (2001). Efficient implementation of high dimensional model representations, Journal of Mathematical Chemistry, 29, 127-142.
  • 5.Li, G., Rosenthal, C. ve Rabitz, H., (2001). High dimensional model representations, Journal of Physical Chemistry A, 105, 7765-7777.
  • 6.Li, G., Wang, S.-W. ve Rabitz, H., (2002). Practical approaches to construct RS-HDMR component functions, Journal of Physical Chemistry A, 106, 8721-8733.
  • 7.Demiralp, M., (2003). High dimensional model representation and its application varieties, Mathematical Research, 9, 146-159.
  • 8.Baykara, N.A. ve Demiralp, M., (2003). Hyperspherical or hyperellipsoidal coordinates in the evaluation of high dimensional model representation approximants, Mathematical Research, 9, 48-62.
  • 9.Demiralp M. ve Tunga, M.A., (2001). High dimensional model representation of multivariate interpolation via hypergrids, the sixteenth international symposium on computer and information sciences, November 5-7, 416-423, Antalya, Turkey.
  • 10.Tunga, M.A. ve Demiralp, M., (2004). A factorized high dimensional model representation on the nodes of a finite hyperprismatic regular grid, Applied Mathematics and Computation, (Kabul edildi).
  • 11.Tunga, B. ve Demiralp, M., (2003). Hybrid high dimensional model representation approximants and their utilization in applications, Mathematical Research, 9, 438-446.
  • 12.Zemanian, A. H., (1987). Distribution theory and transform analysis, an ıntroduction to generalized functions with applications, Dover Publications Inc., New York.
  • 13.Oevel, W., Postel, F., Wehmeier, S. ve Gerhard, J., (2000). The Mupad Tutorial, Springer.
  • 14.Deitel, H. M., Deitel, P. J., Nieto, T. R. ve McPhie, D. C., (2001). Perl How to Program, Prentice Hall.