This is apparently the first paper which can verify that the classification does not change under non-trivial combinations of meaningful geometric transformations like rotations, translations, scaling, shear etc. I think this is an important step in the right direction in the area of provable robustness guarantees for neural networks. For the final version I recommend to - include test errors of all the models - train models with data augmentation and adversarial data augmentation and show how that affects the robustness guarantees. This will increase the practical impact of this paper.