1. Six subjects performed fast, "single-joint" flexions at either the elbow or shoulder over three angular distances in a sagittal plane. Movement endpoints were located to require flexion of only a single, "focal" joint, without any external, mechanical constraint on the other, "nonfocal" joint. Three subjects performed another series of movements between two targets while moving along different paths and in which both joints were flexed. 2. We compared the torque patterns that were produced at the two joints. For single-joint movements, they were both biphasic pulses that accelerated and then decelerated the limb. 3. The torque at the nonfocal joint of a single joint movement was very close to linearly proportional to that at the focal joint throughout the movement. Elbow and shoulder torques differed by a linear scaling constant and went through extrema and zero crossings almost simultaneously. 4. In contrast, during movements in which subjects were explicitly instructed to use a hand path they would not naturally, use the linear interjoint torque scaling rule did not apply. This demonstrated that when we wish to move along a path between two targets that is not produced by linear torque covariation, we are able to modify that rule at will. 5. We speculate that linear, dynamic covariation of the torque patterns across two joints may be an important principle for reducing the number of degrees of freedom that the nervous system must independently control in performing unconstrained limb movements over naturally chosen paths.