Frequency-domain identification algorithms are considered. The aim of this paper is to develop a new algorithm that i) converges to a minimum of the objective function, and ii) possesses optimal numerical properties. Hereto, recent results in instrumental variable system identification are exploited. In addition, a new bilinear form is proposed that leads to the novel introduction of bi-orthonormal polynomials in system identification. The combination of these aspects leads to the desired convergence properties in conjunction with optimal numerical conditioning. The results are supported by means of a simulation example.