A large number of equilibrium states or fixed points is in a randomly and symmetrically connected neural network. Recently it has been shown that the maximum number which can be realized depend on the model of the single neuron. Here we show some network properites of the neuronal model dependence which include the maximum number of equibrium states and the activity of these states. Furthermore, the invariant activity in each model is also derived, where the activity does not depend on the statistical parameters designated by the probability distribution of connection weights between neurons and a threshold of neurons.
It is well-known that computing the eigenvalues and eigenvectors of a given matrix A is difficult in case A is not diagonalizable. In this paper, we apply Power methods with Hotelling's, Frazer-Duncan-Collar's and Wielandt's deflations to such matrices and show the numerical properties. Finally, the error analysis is given, particularly for Power method with Hotelling's deflation, when the Jordan canonical form of A has one 2-dimensional Jordan block.
Arithmetic operations and functions of Taylor series can be defined easily by FORTRAN 90 and C++program language. Using this, it is shown that asymptotic expansion of the integral for oscillatory functions over infinite interval: ∫^∞_0f(x)g(x)dx, where f(x) is slowly decaying function, g(x) is sin x, cosx or J_n(x)(the first kind Bessel function of integer order), can be computed easily by partition integration method. Evaluating this expansion gives an effective numerical integration method for this kind of integrals.
This paper presents the study on determining a transformation between two projective planes, i.e., a plane projective transformation, based on quadratic curves. For a plane projective transformation, the parameters representing the corresponding pair of quadratic curves satisfy a linear constraint and its coefficients are expressed as the tensor product of the plane projective transformation and itself. We present a linear algorithm for uniquely determining from these coefficients the plane projective transformation up to a scale factor. Our algorithm makes full use of redundant information to determine the plane projective transformation with only linear computation. We can now handle plane projective transformations based on quadratic curves.
A numerical method for monitoring temperature distribution in which boundary flux and initial state are unknown is presented. Regularizations based on Tikhonov's and Beck's method are employed. And then, regularization parameters are evaluated by L-curve. The method is applied to an actual piping problem in a steam power plant and compared with measured data, and it is also applied to a two-dimensional thermal shock problem.
In our previous paper[17], we naturally generalized the Morse code and we found the associative generalized Fibonacci sequences. Further we studied in[18]the matrix representation of these generalized sequences. In this paper, we introduce a new code which is developed by our preceding studies of the generalized Morse code. Moreover, we examine an efficient algorithm for generating codewords of the new code systematically and show that the number of codeword of equal lengths gives more widely generalized Fibonacci sequences. Subsequently we also introduce the associated widely generalized Lucas numbers and we study the direct representation of these n-th terms of the newly generalized Fibonacci and Lucas sequences by making use of matrices. Furthermore, we study some extended properities concerning these widely generalized sequences.
In recent years, many computer systems are introduced in police station and various scientific crime detection are enforced. However, it is not easy to investigate the scene of crime. Then, we propose the environment for image recognition with some cubic scales in the picture. The aim of this environment is to support analysis and transformation of object data on 3D space. In this systems, we applied a traditional photo survey technique to a problem. Further more, we discussed the on-line dialogue specification on the user oriented investigation systems.
Conjugate Gradient Algorithm and its variants have been widely used as an excellent solution method for linear equations on vector and/or parallel supercomputers. But the effectiveness depends on the ability to parallelize the preconditioning procedure. Here, we first deal with Neumann expansion preconditioning method and extend it to polynomial preconditioning method. Both are very simple and available for a wide class of matrices. The numerical studies are made on vector/parallel supercomputer S-3800 and parallel processors KSR-1 and AP1000 to validate the vactor/parallel effect. The methods are applied to a matrix discretized by Fourier series expansion of a plasma fluid flow equation, as well as usual finite difference and finite element methods.
A new method for approximately calculating of spectral factors having one parameter is proposed. This is based on usual spectral factorization and Hensel construction with respect to the parameter. An application of such approximate spectral factors to optimal control is also presented.