Performance limits for distributed estimation over LMS adaptive networks