There is actually a good analytical result on how vector similarity can easily fail to recover relevant information https://arxiv.org/pdf/2403.05440 https://arx

There is actually a good analytical result on how vector similarity can easily fail to recover relevant information https://arxiv.org/pdf/2403.05440 https://arxiv.org/pdf/2403.05440 > For some linear models the similarities are not even unique, while for others they are implicitly controlled by the regularization. I am not strong in mathematics but if this paper claims run opposite to each other.