Apresentamos um novo algoritmo, baseado no algoritmo de inversão de matrizes de Leverrier-Fadeev, para extrair os autovalores e os coeficientes do polinômio característico da matriz (si+ A), não-simétrica, que surge em conexão com o método LTSN - o qual utiliza a transformada de Laplace para a solução da equação de ordenadas discretas S N. O algoritmo baseia-se em propriedades exibidas pela matriz, cuja estrutura e valores dos elementos fazem com que todos os seus autovalores sejam reais e simétricos em relação a zero. Evidências experimentais demonstram que, os autovalores do bloco superior esquerdo da matriz, de dimensão N /2, entrelaçam os autovalores negativos de -A. O algoritmo foi implementado em FORTRAN 77, usando algumas rotinas do BLAS e do LAPACK, e estruturado de forma a explorar a estrutura da matriz, permitindo efetuar os cálculos necessários em um menor tempo e com um menor gasto de menória. No entanto, apesar de ganhos obtidos em comparação com o algoritmo usualmente utilizado, proposto por Barichello, nossos experimentos demonstram a instabilidade numérica do algoritmo de Leverrier-Fadeev. / We present a new algorithm to compute the eigenvalues and the coefficients o f the characteristic polynomial o f a nonsymmetric matrix o f the form (sI+ A), which arises in connection with the LTSN method for the solution of thc discrete ordinates equations S N. Our algorithm is a modifi.cation of the matrix inversion Leverrier-Fadeev algorithm, exploiting the pattern existent in the matrix -A and some properties exhibited by its eigenvalues, which have been determined experimentally. More specifi.cally, its eigenvalues alllie on the real axis and are symmetrically distributed around zero. Also, -A has a block structure and the eigenvalues of the left-hand superior block interleave the negative eigenvalues of the matrix. The algorithm was designed to exploit these characteristics, computing only the nega:tive eigenvalues of -A (due to their symmetrical distribution) by means of the well-know bisection method to obtain the zeros of thc characteristic polynomial. Since the eigenvalues of the left-hand superior block of A interleave those of the matrix, it is possible to use intervals made of pairs of those eigenvalues which contain just a single eigenvalue of - A. Also, the structure of -A was used to develop optimized sections of code of thc algorithm to reduce the number of operations required. The whole algorithm was implementcd in FORTRAN 77, making use of some of the BLAS and LAPACK routines. The results obtained although presenting a better performance than that used currently, due to Barichello, show that the algorithm is susceptible to the ill-conditioning of the matrix.

Apresentamos um novo algoritmo, baseado no algoritmo de inversão de matrizes de Leverrier-Fadeev, para extrair os autovalores e os coeficientes do polinômio característico da matriz (si+ A), não-simétrica, que surge em conexão com o método LTSN - o qual utiliza a transformada de Laplace para a solução da equação de ordenadas discretas S N. O algoritmo baseia-se em propriedades exibidas pela matriz, cuja estrutura e valores dos elementos fazem com que todos os seus autovalores sejam reais e simétricos em relação a zero. Evidências experimentais demonstram que, os autovalores do bloco superior esquerdo da matriz, de dimensão N /2, entrelaçam os autovalores negativos de -A. O algoritmo foi implementado em FORTRAN 77, usando algumas rotinas do BLAS e do LAPACK, e estruturado de forma a explorar a estrutura da matriz, permitindo efetuar os cálculos necessários em um menor tempo e com um menor gasto de menória. No entanto, apesar de ganhos obtidos em comparação com o algoritmo usualmente utilizado, proposto por Barichello, nossos experimentos demonstram a instabilidade numérica do algoritmo de Leverrier-Fadeev. / We present a new algorithm to compute the eigenvalues and the coefficients o f the characteristic polynomial o f a nonsymmetric matrix o f the form (sI+ A), which arises in connection with the LTSN method for the solution of thc discrete ordinates equations S N. Our algorithm is a modifi.cation of the matrix inversion Leverrier-Fadeev algorithm, exploiting the pattern existent in the matrix -A and some properties exhibited by its eigenvalues, which have been determined experimentally. More specifi.cally, its eigenvalues alllie on the real axis and are symmetrically distributed around zero. Also, -A has a block structure and the eigenvalues of the left-hand superior block interleave the negative eigenvalues of the matrix. The algorithm was designed to exploit these characteristics, computing only the nega:tive eigenvalues of -A (due to their symmetrical distribution) by means of the well-know bisection method to obtain the zeros of thc characteristic polynomial. Since the eigenvalues of the left-hand superior block of A interleave those of the matrix, it is possible to use intervals made of pairs of those eigenvalues which contain just a single eigenvalue of - A. Also, the structure of -A was used to develop optimized sections of code of thc algorithm to reduce the number of operations required. The whole algorithm was implementcd in FORTRAN 77, making use of some of the BLAS and LAPACK routines. The results obtained although presenting a better performance than that used currently, due to Barichello, show that the algorithm is susceptible to the ill-conditioning of the matrix.

Apresentamos um novo algoritmo, baseado no algoritmo de inversão de matrizes de Leverrier-Fadeev, para extrair os autovalores e os coeficientes do polinômio característico da matriz (si+ A), não-simétrica, que surge em conexão com o método LTSN - o qual utiliza a transformada de Laplace para a solução da equação de ordenadas discretas S N. O algoritmo baseia-se em propriedades exibidas pela matriz, cuja estrutura e valores dos elementos fazem com que todos os seus autovalores sejam reais e simétricos em relação a zero. Evidências experimentais demonstram que, os autovalores do bloco superior esquerdo da matriz, de dimensão N /2, entrelaçam os autovalores negativos de -A. O algoritmo foi implementado em FORTRAN 77, usando algumas rotinas do BLAS e do LAPACK, e estruturado de forma a explorar a estrutura da matriz, permitindo efetuar os cálculos necessários em um menor tempo e com um menor gasto de menória. No entanto, apesar de ganhos obtidos em comparação com o algoritmo usualmente utilizado, proposto por Barichello, nossos experimentos demonstram a instabilidade numérica do algoritmo de Leverrier-Fadeev. / We present a new algorithm to compute the eigenvalues and the coefficients o f the characteristic polynomial o f a nonsymmetric matrix o f the form (sI+ A), which arises in connection with the LTSN method for the solution of thc discrete ordinates equations S N. Our algorithm is a modifi.cation of the matrix inversion Leverrier-Fadeev algorithm, exploiting the pattern existent in the matrix -A and some properties exhibited by its eigenvalues, which have been determined experimentally. More specifi.cally, its eigenvalues alllie on the real axis and are symmetrically distributed around zero. Also, -A has a block structure and the eigenvalues of the left-hand superior block interleave the negative eigenvalues of the matrix. The algorithm was designed to exploit these characteristics, computing only the nega:tive eigenvalues of -A (due to their symmetrical distribution) by means of the well-know bisection method to obtain the zeros of thc characteristic polynomial. Since the eigenvalues of the left-hand superior block of A interleave those of the matrix, it is possible to use intervals made of pairs of those eigenvalues which contain just a single eigenvalue of - A. Also, the structure of -A was used to develop optimized sections of code of thc algorithm to reduce the number of operations required. The whole algorithm was implementcd in FORTRAN 77, making use of some of the BLAS and LAPACK routines. The results obtained although presenting a better performance than that used currently, due to Barichello, show that the algorithm is susceptible to the ill-conditioning of the matrix.