When performing feature selection for the Wisconsin Breast Cancer Data set, there appear to be covariates that are inherently related by their definition. Given the approximate circular nature of breast cancer tumours, the area and radius of tumours will be related quadratically as A_tumour \approx pi * r_tumour^2. Pearson correlation only captures linear correlations between covariates. As these features are basically expressing the same information, would it be optimal here to make one of these features redundant to improve the classifiers metrics?
You can eliminate features that are correlated to each other. This will help build the model with fewer parameters and train efficiently.