This paper presents an initial rectifying learning method for trajectory tracking of linear time-invariant systems with rank-defective Markov parameters. The initial shift problem is addressed through introduction of the initial rectifying action. The role of the rectifying action is examined in case of systems with row and column rank-defective Markov parameters, respectively. Sufficient conditions for convergence of the proposed learning algorithms are derived, by which the learning gains can be chosen. It is shown that the output trajectory converges to the desired one with a smooth transition. The merging of the transition to the desired trajectory occurs at a pre-specified time instant.