Dönüşümsel yüksek boyutlu model gösterilimi tabanlı yeni oransal yaklaştıranlar oluşturumu ve LSI sistemlerdeki uygulamaları

Oransal yaklaştırım yöntemlerinin, tek veya çok değişkenli fonksiyonlardaki doğrusal olmama durum- larının veya tekilliklerin üstesinden gelmede kullanılan en etkin ve oluşturulması en kolay yaklaştırımyöntemleri olduğu söylenebilir. Bu yöntemler, çok değişkenli, doğrusal öteleme altında değişmeyensistemlerdeki (ing: Multidimensional Linear Shift Invariant-LSI recursive systems) model indirgemeuygulamalarında, çok düzeyli birleştiricilerdeki (ing: multiplexer models) paket kayıp olasılıklarınınhesaplanmasında ve üç boyutlu şekilleri yeniden oluşturmada (ing: shape reconstruction) kullanılır- lar. Bu çalışmada, çok değişkenli, yeni bir oransal yaklaştırım yöntemi geliştirilmiş, yöntemin özellik- leri, günümüzde oransal yaklaştırım yöntemlerinin en başında gelen Padé yaklaştırım yöntemi ilekarşılaştırmalar yaparak tartışılmıştır. Burada sunulan oransal yaklaştırım çizemi, YBMG’nde (Yük- sek Boyutlu Model Gösterilimi) yapılan kesmelere dayanır. YBMG, analitik yapısı bilinen çok değiş- kenli fonksiyonları daha az değişken içeren bir takım fonksiyonların toplamı olarak yaklaştırmayadayanan ve son 15 yılda geliştirilen bir yaklaştırım yöntemidir. Çalışmada, yaklaştırımın niteliğiniarttırmak amacı ile yöntemin yapısına bir takım esneklikler katılmıştır. YBMG, asıl fonksiyon yerinefonksiyonun birinci derece dönüşüm altındaki görüntüsüne uygulanmıştır. Bu da bize yaklaştırımafazladan esneklik katabilme olanağını vermiştir. YBMG aracılığı ile elde edilen bu yeni yaklaştırımyönteminin fiziksel problemlerdeki davranışını gözlemlemek amacı ile, yöntem doğrusal öteleme al- tında değişmeyen, çok değişkenli sistemlere (LSI) uygulanmıştır. Bunun dışında, tek değişkenli bir ya- pay fonksiyon örneği de, yöntemin tek değişkenli fonksiyonlardaki etkinliğini incelemek için verilmiş- tir. Sonuçlar, tek veya çok değişkenli bir fonksiyonu Padé oransal yaklaştırım yöntemine göre dahageniş bir aralıkta yaklaştırabildiğimizi göstermektedir.

Construction of new rational approximants based on transformational high dimensional model representation and their applications on LSI systems

There are plenty of methods to approximate a multi- variate function arising from different concepts.Some of them are based on the truncation of infiniteseries like Taylor series or orthogonal expansions.These do not explicitly reflect the singularities of thefunction to be approximated. One of the infinite se- ries explicitly showing singularities is the Laurentseries, which include finite or infinite simple frac- tions to represent polar singularities. Since any ra- tional function can be uniquely represented by aLaurent series, it is common practice to use rationalapproximants for approximating the functions toexplicitly reflect all polar singularities. Padé approximants are perhaps the most frequentlyused ones in this category. Functions having strongsingularities like branch points or essential singu- larities need special attention and can be approxi- mated by using more specific techniques. Rationalapproximation techniques are perhaps the leadingand easiest ones used to handle nonlinearities orsingularities in the approximation of univariate ormultivariate functions in science and engineering.They can be used in model-reduction of multidimen- sional Linear Shift Invariant (LSI) recursive systems,computing packet loss probabilities in multiplexermodels, and shape reconstruction. In this work, we construct a new multivariate ra- tional approximation method and discuss its ad- vantages and disadvantages by comparing with thePadé technique, which is at the moment the leadingrational approximation technique. The rational ap- proximation scheme proposed here is formulated byusing HDMR (High Dimensional Model Representa- tion) truncations. HDMR has been developed in thelast 15 years to approximate multivariate functionsby a sum of less variate functions. Here, we insertcertain flexibilities into the approximation, in orderto improve its quality. The most general class of thistype of HDMR is called “Transformational HighDimensional Model Representation” (THDMR). InTHDMR, the image of the original multivariatefunction is approximated by a sum of less variatefunctions. The coefficients of the transformation areconsidered as optimization parameters, which makethe error of the method as small as possible. In thiswork, a first degree polynomial structure is used asTHDMR’s operator and the coefficients of this firstdegree transformation are assumed to be varyingwith independent variables. This is for giving flexi- bilities to the transformation and fixing them to ap- proximate the HDMR expansion optimally. AlthoughTHDMR has a more general structure than we usedhere, we chose to work with the first degree polyno- mial structure for unique invertibility. Once an ap- proximation for this first order polynomial structureconstructed by HDMR method, the approximationfor the original function can be obtained by invert- ing the polynomial structure. The rational structureof the method comes from this special choice ofTHDMR operator. The univariate and bivariate results are quite im- pressive and imply that we can approximate a multi- variate function over a quite larger interval whencompared with the Padé method. This is caused bythe fact that these two methods involve differentstrategies of extracting information from the origi- nal function. In THDMR method, the integral of theoriginal function is computed over a specified inter- val to extract information from it, while Padé ap- proximation method uses the derivatives of the func- tion in a specific point. So THDMR method approx- imates the function more uniformly compared withPadé approximation technique which exhibits a lo- cal character. Moreover, the quality of the approxi- mation can be improved by changing the integrationintervals in THDMR. An important defect of the rational approximationtechniques is that they can produce some poles overthe focused interval, although the original functionhas no singularity over the same interval at all. Thesingularity problem exists also in THDMR method.However, in THDMR, in order to obtain the rationalapproximant, certain integrals are needed to becomputed and the location of the possible poles ofthe approximant can be altered by changing the in- terval of the integration. This facility of changing thelocations of the poles with the integration intervalmakes the method attractive and efficient. In orderto see the behavior of the method on physical prob- lems, we have applied it on LSI systems to reducethe system. The transfer function of the LSI systemhas been approximated by THDMR and Padé meth- ods. The results show that, the newly developedTHDMR method can be used on LSI systems withhigh level approximation quality.

___

  • Abouir, J. ve Cuyt, A.A.M., (2007). Stable multi- dimensional model reduction and IIR filter de- sign, International Journal of Computing Science and Mathamatics, 1, 16-27.
  • Baker, G.E., (2006). Padé approximants, Cambridge University Press.
  • Cuyt, A.A.M., Ogawa, S. ve Verdonk, B., (1992). Model reduction of multidimensional linear shift- ınvariant recursive systems using padé tech- niques, Multidimensional Systems and Signal Processing, 3, 309-322.
  • Demiralp, M., (2003). High dimensional model rep- resentation and its applications, Tools for Mathe- matical Modelling, 9, 146-159.
  • Powell, M.J.D., (1996). Approximation theory and methods, Cambridge University Press.
  • Sobol, I.M., (1993). Sensitivity estimates for nonlin- ear mathematical models, MMCE (Mathematical Modelling and Computational Experiment), 1-4, 407-419.
  • Tunga, M.A. ve Demiralp M., (2005). A factorized high dimensional model representation on the nodes of a finite hyperprismatic regular grid, Ap- plied Mathematics and Computation, 164, 865- 883.
  • Yaman, İ. ve Demiralp M., (2004). High dimension- al model representation approximation of an evo- lution operator with a first order partial differen- tial operator argument, ANACM (Numerical Anal- ysis and Computational Methods), 1, 280-289.