No mystery.
The training algorithm tries to minimize the error on the training subset. That is why the training error is, most of the time, lower than that of nontraining data.
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!