We discuss typical `potential' performance of sparse signal recovery in the case where iterative recovery algorithms fail to converge. We especially focus on sparse signal recovery from a linear measurement whose measurement matrix has large coherence and evaluate the mean square error of the estimate in such a case by using the statistical mechanical method. It can be considered that this kind of analysis gives a theoretical limit for iterative sparse recovery algorithms. When coherence of the measurement matrix is large, the mean square error between an original signal to be estimated and its estimate tends not to depend on the compression rate.

We challenge the human identification problem from the perspective of gait and body shape. Conventional methods depend on the camera viewing direction, and since they are based on matching image silhouettes or features their identification accuracy is low when there is a big difference between the camera viewing direction of the test and training data. Thus, if a person is walking in an arbitrary direction, they may not be accurately identified. In this paper, we propose a novel method that does not depend on the camera viewing direction. We develop a state space model called a "cyclic motion model" whose state variables are not only the phase of the motions but also the camera viewing direction. We learn model parameters for each candidate person, and represent their walking with the cyclic motion model. To identify a person from the observed image sequence, we first compute the model likelihoods for the sequence using a particle filter that represents a probability distribution by a set of weighted samples, We then identify the person from model likelihoods.