Legible motion --- motion that communicates its intent to a human observer --- is crucial for enabling seamless human-robot collaboration. In this paper, we propose a functional gradient optimization technique for autonomously generating legible motion. Our algorithm optimizes a legibility metric inspired by the psychology of action interpretation in humans, resulting in motion trajectories that purposefully deviate from what an observer would expect in order to better convey intent. A trust region constraint on the optimization ensures that the motion does not become too surprising or unpredictable to the observer. Our studies with novice users that evaluate the resulting trajectories support the applicability of our method and of such a trust region. They show that within the region, legibility as measured in practice does significantly increase. Outside of it, however, the trajectory becomes confusing and the users' confidence in knowing the robot's intent significantly decreases.