Optimal estimation of atmospheric temperature and composition from limb sounding observations is extended to the direct retrieval of line-of-sight atmospheric structure that can be obtained in certain limb viewing geometries. The approach is to divide the dataset into slightly overlapping chunks of several atmospheric profiles and retrieve estimates for all profiles concurrently. The method is made efficient due to the sparse nature of the matrices involved. In the case where the number of radiance measurements is significantly larger than the length of the state vector, the computational effort scales linearly with the number of profiles in the chunk. Prototype simulations, done for the EOS MLS experiment, show that application of this method can give significant improvements in accuracy. ¿ 2000 American Geophysical Union