Conditionally optimal algorithms and estimation of reduced order models (Q1112002)

The paper presents optimality results for algorithms used in the field of system identification in the light of the information-based complexity theory. A class of restricted algorithms is defined by means of a restriction in the space of solution elements, and a corresponding conditional worst case error is introduced. We define conditional central algorithms and show their optimality. A conditional central algorithm is then constructed by modifying a projection algorithm and obtaining in this way a conditional projection algorithm. This algorithm is shown to enjoy local optimality properties with respect to the problem element space within the class of conditionally correct algorithms. Finally, it is shown how the obtained results can be used to handle the problem of reduced order model estimation.