We investigate a sensing system in which a Markov source is observed by a sensor that can communicate noiselessly to a receiver. However, each transmission consumes a fixed amount of power. To conserve energy, at any time instant the sensor may decide not to transmit or to causally encode its observations before transmission. The objective is to choose transmission and estimation policies to minimize a weighted sum of the average transmitted energy and the average distortion. We derive the structure of optimal transmission and estimation policies and use that to derive a dynamic programming decomposition of the system.