Skip to main content
Usami Laboratory

An arXiv preprint on invertible graph aggregation and classifier width is now available

Our preprint “Invertible Graph Aggregation Can Increase ReLU Readout Width” is now available on arXiv.

The study

For every even N≥4, a tree with 2N vertices and maximum degree three needs one hidden unit on raw scalar features and exactly N units after one lazy neighborhood mean. The paper also studies scalar perturbations and a specified noise model, and reports controlled fits and comparisons on public graphs.

Paper and release details

Paper
Invertible Graph Aggregation Can Increase ReLU Readout Width
Authors
Hiroyasu Usami, Masanori Tsukamoto, Kohei Hashiguchi
Public version
arXiv:2611.02274 [cs.LG] (preprint; not peer reviewed)
arXiv v1 submitted
2026-11-01 (UTC)