To map the linear predictor to the range of the response variable.

To ensure the variance is constant.

To define the distribution of the errors.

To allow for non-linear relationships between predictors.