This paper presents a new method for estimating a given latent trait variable by the least-squares approach. The beta weights are obtained recursively with the help of Fourier series and expressed as functions of item parameters of response curves. The values of the latent trait variable estimated by this method and by maximum likelihood method were compared using real data: A 48-item test in matrix algebra administered by computer, in an adaptive testing situation. The results were very close, and yet the values of the latent trait variable estimated by the multiple regression method were always obtainable, even for a small number of examinees and a small number of items. The maximum likelihood method, on the other hand, often fails to converge. (Author)