Which of the following is true about the residuals when applying the best subset selection method?

The choice indicating that the residual sum of squares may vary across models is accurate because, in the context of best subset selection, different combinations of predictors will generally lead to different model fits. Each combination uses a subset of the available predictors, which can significantly impact the error in prediction. Consequently, the residuals, which are the differences between observed and predicted values, will change based on which predictors are included in the model.

Since a model with fewer predictors may yield larger residual errors compared to a model with more predictors that captures the relationship more effectively, the residual sum of squares—which quantifies the overall error—will not be constant across different models created with varying numbers and combinations of predictors. Each model's fit is influenced by its unique set of parameters and variables, resulting in differences in residuals and, consequently, the residual sum of squares.

In contrast, the other statements do not hold true because they suggest invariant characteristics of the residuals across all models or overly generalize the behavior of best subset selection, which is a flexible and tailored approach to model selection.