2025 , Volume 30, № 2, p.73-86

Purpose. Developing a tool with precision tracking for solving ill-conditioned quadratic unconditional optimization problems of enlarged dimension.

Methodology. The standard conjugate gradient (CG) method was chosen as the basis of the algorithm for finding the solution. The work uses the ball arithmetic (van der Hoeven, Johansson), which provides convenient tools for operations with arbitrary precision numbers, as well as vectors and matrices based on them. The precision indicators of ball numbers and composite numeric objects in bits and decimal digits are described. A centering procedure is proposed to speed up calculations in iterations of the CG algorithm. Computational experiments have been carried out to illustrate the precision degradation process in CG iterations and to find out mantissa length of ball numbers required to successfully find a solution to a number of ill-conditioned quadratic unconditional optimization problems. The effect of centering on the total solution time has been examined experimentally.

Findings. The pattern of the precision degradation in the computing process is different and significantly depends on the type of quadratic problem. According to the results of the computational experiment, the centering significantly reduces the time for finding solution to the optimization problems under consideration. A hybrid CG algorithm with a feedback mechanism using centering in a precision depletion situation of intermediate data is proposed.

Value. The proposed approach along with the hybrid CG algorithm allows obtaining stable and time-acceptable solutions for several types of poorly conditioned quadratic unconditional optimization problems of enlarged dimension.