Mirror descent selects among the minimizers the one closest to initialization for the associated Bregman divergence (its "implicit bias"). Euclidean descent computes an orthogonal projection on the minimizers.

Using the language of this equivalently reads as the fact that the mirror flow (i.e. when tau->0) has conservation laws.

The difference between this and our paper is that these conservation laws of implicit bias depend on the data (via the matrix A) whereas we studied conservation laws which are independent of the data (e.g. depend only on the neural network architecture).

It sounds be the growth rate of a curves with the definition of the derivatives

-e=(imc)²