The minimum of the empirical risk function on a training dataset can occur at a completely different location than the minimum of the true risk function. This illustrates why simply minimizing the training error via optimization may not necessarily lead to the lowest possible generalization error.