Neural-network approximation of functions of several variables

The main result of the work is as follows: in the Chebyshev–Hermite weighted integral metric, it is possible to approximate any function of sufficiently general form by a neural network. The approximating net consists of two layers, where the first uses any predefined sigmoid function of activation and the second uses a linear-threshold function. The Chebyshev–Hermite weight is chosen because it allows one to imitate the distribution of receptors, for example, in the eye of a human or some mammal.     

Local approximation of functions with several variables

This paper presents the problem of local approximation of scalar functions with several variables, including points of non-differentiability. The procedure considers that the mapping displays rates of change of power type with respect to linear changes in the coordinate domain, and the exponents are not necessarily integer. The approach provides a formula describing the local variability of scalar fields which contains and generalizes Taylor’s formula of first order. The functions giving the contact are Müntz polynomials. The knowledge of their coefficients and exponents enables the finding of local extremes including cases of non-smoothness. Sufficient conditions for the existence of global maxima and minima of concave-convex functions are obtained as well.     

Best approximations of functions of several variables by sums of functions of fewer variables

The minimum of the mean-square deviation of approximations of functions of several independent variables by sums of functions of fewer variables in a multidimensional parallelepiped is investigated. The approximating function yielding the minimum mean-square deviation is obtained and this minimum deviation is calculated.Translated from Matematicheskie Zametki, Vol. 5, No. 2, pp. 217–226, February, 1969.The author wishes to thank M. D. Millionshchikov and V. A. Khodakov for their useful suggestions concerning this work.     

Best one-sided approximation of certain classes of functions

This considers the question of the best one-sided approximation of certain classes of continuous periodic functions by means of trigonometric polynomials of order n-1 in the metric L ₂ ^p (1p<). Precise upper bounds are obtained for the best one-sided approximation of classes of 2/n-periodic functions H_,n [having arbitrary prescribed modulus of continuity (t)] in the metric L ₂ ^p , as well as of classes of 2-periodic functions H [having prescribed modulus of continuity (t) with definite limits] in the metric L ₂ ¹ .Translated from Matematicheskie Zametki, Vol. 10, No. 6, pp. 615–626, December, 1971.The author warmly thanks N. P. Korneichuk for his constant attention to this work.     

Piece-wise linear approximation of functions of two variables

The goal of increasing computational efficiency is one of the fundamental challenges of both theoretical and applied research in mathematical modeling. The pursuit of this goal has lead to wide diversity of efforts to transform a specific mathematical problem into one that can be solved efficiently. Recent years have seen the emergence of highly efficient methods and software for solving Mixed Integer Programming Problems, such as those embodied in the packages CPLEX, MINTO, XPRESS-MP. The paper presents a method to develop a piece-wise linear approximation of an any desired accuracy to an arbitrary continuous function of two variables. The approximation generalizes the widely known model for approximating single variable functions, and significantly expands the set of nonlinear problems that can be efficiently solved by reducing them to Mixed Integer Programming Problems. By our development, any nonlinear programming problem, including non-convex ones, with an objective function (and/or constraints) that can be expressed as sums of component nonlinear functions of no more than two variables, can be efficiently approximated by a corresponding Mixed Integer Programming Problem.     

On best approximation of functions of two variables

$\mathop {\lim \sup }\limits_{r \to \infty } \frac{{E_{n_i ,m_i } (f)_L }}{{[E_{n_i ,\infty } (f)_L + E_{\infty ,m_i } (f)_L ]ln\{ 2 + min(n_i ,m_i )\} }}\underset{\raise0.3em\hbox{$\mathop {\lim \sup }\limits_{r \to \infty } \frac{{E_{n_i ,m_i } (f)_L }}{{[E_{n_i ,\infty } (f)_L + E_{\infty ,m_i } (f)_L ]ln\{ 2 + min(n_i ,m_i )\} }}\underset{\raise0.3em\hbox{     

On the best approximation of functions ofn variables

We propose a new approach to the solution of the problem of the best approximation, by a certain subspace for functions ofn variables determined by restrictions imposed on the modulus of, continuity of certain partial derivatives. This approach is based on the duality theorem and on the representation of a function as a countable sum of simple functions. Institute of Mathematics, Ukrainian Academy of Sciences, Kiev. Translated from Ukrainskii Matematicheskii Zhurnal, Vol. 51, No. 10, pp. 1352–1359, October, 1999.