New blog unpacks removing the square root in Adam: what second-order methods reveal

burny_tech · x · 2026-09-04

Researcher zhanpengzhou published his fifth blog on optimization, examining why adaptive optimizers like Adam(W) use a square root: it guarantees loss-scaling invariance, and removing it (to get closer to Newton's method) requires first resolving the scaling mismatch.

The post builds on the arXiv paper Can We Remove the Square-Root in Adaptive Gradient Methods? A Second-Order Perspective (Wu Lin et al.). Key findings: square-root-free adaptive methods close the generalization gap to SGD on convolutional architectures while matching root-based performance on transformers, and they work well with half-precision since they avoid numerically unstable matrix root decompositions — opening a practical path to non-diagonal preconditioner methods.

Original post →

More from Research

Research channel →