Q5. In cross validation, we evaluate the model performance on data that we have used for hyperparameter tuning. Thus, cross validation would underestimate the generalization error.
Q6. Take the linear regression model y=b+w1x1+w2x2 as an example. OLS and LASSO have the same VC dimension (3), but LASSO has lower model complexity.
Q7. I will construct a sequence of random samples from the original sample, with increasing sample size. If the model performance based on each sample improves as the sample size increases, that would justify collecting more data for better model performance.