A straight line trying to fit a curved relationship is a classic case. Remedies include a larger or more expressive model, better features, more training or less regularization.
The goal is the balance between underfitting and overfitting, where a model generalizes well to new data.