A recent study published on arXiv investigates the adversarial robustness of deep neural networks in classification tasks by comparing them to optimal classifiers. The research focuses on identifying the smallest perturbations capable of changing a classifier's output, highlighting the networks' vulnerability to such minimal changes. Through a matrix-theoretic approach, the paper provides a novel perspective on why these neural networks exhibit fragility when subjected to adversarial inputs. Central to the findings is the identification of feature compression as the root cause of this adversarial fragility. This insight advances the understanding of how deep learning models process information and why they may fail under adversarial conditions. The study contributes to ongoing discussions about improving the robustness of AI systems against subtle manipulations. It also aligns with broader research efforts aimed at enhancing the security and reliability of neural network classifiers.