We propose a new appearance-based feature for real-time gesture recognition from motion images. The feature is the shape of the trajectory caused by gesture, in pattern space defined by inner-product between patterns on frame images. It has three merits, 1) invariant for the target human's position, size and lie, 2) gesture recognition without interpreting frame image contents, 3) no costly statistical calculation. And it gives us a theoretical guarantee about the effectiveness of several time-sequence matching methods using shapes in the eigenspace or results of position tracking. In this paper, we describe the properties of the gesture trajectory feature, and some experimental results in order to show its applicability to gesture recognition with a theoretical consideration.