Strong ironing property for rolling-ball optimizers

Let f,g:RdRf,g:\mathbb{R}^d\to\mathbb{R} be two continuous functions, and let Γ,Γ\Gamma,\Gamma^\prime be their respective graphs. Write dH(Γ,Γ)\operatorname{d_{\mathrm{H}}}(\Gamma^\prime,\Gamma) for the Hausdorff distance between the graphs, and let S(Γ,ρ)\operatorname{S}(\Gamma,\rho) denote the rolling-ball optimizer associated with Γ\Gamma at scale ρ\rho. Two spaces are ε\varepsilon-almost isometric if there is a map between them whose distance distortion is at most ε\varepsilon. Strong ironing property. If

dH(Γ,Γ)<+,\operatorname{d_{\mathrm{H}}}(\Gamma^\prime,\Gamma)<+\infty,

then S(Γ,ρ)\operatorname{S}(\Gamma,\rho) and S(Γ,ρ)\operatorname{S}(\Gamma^\prime,\rho) are ερ\varepsilon_\rho-almost isometric, with ερ0\varepsilon_\rho\to0 as ρ+\rho\to+\infty. The claim would extend the observed ironing phenomenon from affine and other controlled perturbations to arbitrary continuous functions; the supplied context states that it remains unproved and is deferred to future work.

Sources & referencesView supporting material

Primary source

Mohammed Djameleddine Belgoumri, Mohamed Reda Bouadjenek, Hakim Hacid, Imran Razzak and Sunil Aryal, “Rolling Ball Optimizer: Learning by ironing out loss landscape wrinkles”, arXiv:2505.19527 (2025).

Progress summary

Never refreshed

Nothing recorded yet. Refresh searches the literature and the public web for attempts on this problem, and writes the first summary here.

Solutions 0

No solutions have been posted yet.