1: \begin{abstract}

2: % \signSGD is a simple modification of SGD, where the gradient is compressed into a binary-like vector. In recent years, \signSGD has garnered interest from community folk-lore as an effective intermediary for studying Adam.

3: % Although prior studies have have shown that \signSGD can have guaranteed convergence, the dynamical behavior of \signSGD's binary-like gradient across the loss-landscape are still unknown. In this work, we derive a deterministic ODE that is able to capture the evolution of \signSGD  in the high-dimensional limit. Using this ODE we show that \signSGD is able to symmetrize  any noise distribution, effectively balancing it out. Moreover, we derive an explicit preconditioning effect \signSGD has on the covariance matrix of the data.

4: % \end{abstract}

5: