{"id":"arxiv/2610.15223","url":"https://arxiv.org/abs/2610.15223","tags":["math.OC","cs.LG"],"score":0,"title":"Non-Euclidean Gradient Descent Has Many Edges of Stability","domain":"math.OC","source":"arxiv","authors":["Peter Zaika","Zakhar Shumaylov"],"reviews":0,"subject":"math.OC","summary":"Recent observations that deep learning models train at the edge of stability have renewed interest in understanding stability properties of classical optimization algorithms. For quadratic objectives, standard gradient descent has a sharp stability threshold: it converges from every initialization when $η 2/L$, and diverges from every nonzero initialization when $η>2/μ$. For non-Euclidean gradient descent, the corresponding known guarantees are weaker: convergence holds for every initialization when $η 2/μ$. The behavior in the intermediate regime $2/S<η<2/μ$ is not fully characterized. We…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-10","reactions":{}}