Constraint satisfaction modules: a methodology for analog circuit design

This dissertation describes a methodology for solving convex constraint problems using analog circuits. It demonstrates how this methodology can be used to design circuits that solve function-fitting problems through iterated gradient descent. In particular, it shows how to build a small circuit that can model a nonlinearity by observation, and predistort to compensate for this nonlinearity. The system fits into a broader effort to investigate non-traditional approaches to circuit design. First, it breaks the traditional input-output abstraction barrier; all ports are bidirectional. Second, it uses a different methodology for proving system stability with local rather than global properties. Such stability arguments can be scaled to much more complex systems than traditional stability criteria. Thesis Supervisor: Gerald Jay Sussman Title: Panasonic Professor of Electrical Engineering Thesis Supervisor: Thomas F. Knight, Jr Title: Senior Research Scientist