Publication: Automatic computation of derivatives with the use of the multilevel differentiating technique --- I: Algorithmic basis
Introduction
Applications
Tools
Research Groups
Workshops
Publications
   List Publications
   Advanced Search
   Info
   Add Publications
My Account
About

Automatic computation of derivatives with the use of the multilevel differentiating technique --- I: Algorithmic basis

- Article in a journal -
 

Author(s)
Yu. M. Volin , G. M. Ostrovskii

Published in
Computers and Mathematics with Applications

Year
1985

Abstract
Consider the vector-valued function y = f(x), f : R^m → R. The authors consider calculating the partial derivatives (∂ y) ⁄ (∂ x_i) = (∂ f) ⁄ (∂ x_i) for each 1 ≤ i ≤ m . The first method of calculation, called the forward method , uses the variational form ∂ y = (∂ f) ⁄ (∂ x_i) ∂ x_i . The second method of calculation, called the back method , uses the adjoint variational form ∂ x'_i = ≤ft((∂ f) ⁄ (∂ x_i)ight)^T ∂ y' where .' means adjoint variable and sign T means transposition. The forward method has been discussed elsewhere in the literature, see [Rall81a]. The authors discuss the relative merits of both methods. They provide results which show that the forward method is the more space efficient and the back method is the more time efficient. The authors combine the two methods to produce a multilevel method which is a compromise of the two methods. This has the property that the memory requirements are better than the back method and the time requirements are better than the forward method.

AD Theory and Techniques
General

BibTeX
@ARTICLE{
         Volin1985Aco,
       author = "{Yu.} M. Volin and G. M. Ostrovskii",
       title = "Automatic computation of derivatives with the use of the multilevel differentiating
         technique --- {I}: {A}lgorithmic basis",
       journal = "Computers and Mathematics with Applications",
       volume = "11",
       year = "1985",
       pages = "1099--1114",
       doi = "10.1016/0898-1221(85)90188-9",
       referred = "[Christianson1996SSU], [Irim91a], [Kubo91a], [Tesf91a], [Wexl87a].",
       keywords = "point algorithm; differentiation arithmetic; forward method; back method.",
       abstract = "Consider the vector-valued function $$ y = f(x), \quad f : R^m
         \rightarrow R. $$ The authors consider calculating the partial derivatives $$
         \frac{\partial y}{\partial x_i} = \frac{\partial f}{\partial x_i} $$
         for each $ 1 \le i \le m $. The first method of calculation, called the {\sl forward
         method \/}, uses the variational form $$ \partial y = \frac{\partial
         f}{\partial x_i} \partial x_i $$. The second method of calculation, called the {\sl
         back method \/}, uses the adjoint variational form $$ \partial \overline{x}_i =
         \left(\frac{\partial f} {\partial x_i}ight)^T \partial \overline{y} $$
         where $\overline{.}$ means adjoint variable and sign T means transposition. The forward method
         has been discussed elsewhere in the literature, see [Rall81a]. The authors discuss the relative
         merits of both methods. They provide results which show that the forward method is the more space
         efficient and the back method is the more time efficient. The authors combine the two methods to
         produce a {\sl multilevel method \/} which is a compromise of the two methods. This has
         the property that the memory requirements are better than the back method and the time requirements
         are better than the forward method.",
       ad_theotech = "General"
}


back
  

Contact:
autodiff.org
Username:
Password:
(lost password)