The cosine transform is nowadays widely used in image data compression studies as a result of its observed near optimum performance with respect to variance redistribution andits property of reducing block edge effects which occur at extreme values of compression. The reason for its good performance for a widely used class of sources is now demonstrated theoretically by showing that it can be derived from the optimum (Karhunen-Loève) transform in the limiting case when the adjacent data-element correlation tends to unity.