One-dimensional worst-case conjecture for gradient descent on composed operator classes

At least 2 years old · documented by

Let Cμg0\mathcal{C}_{\mu_g}^{0} and DμgμM\mathcal{D}_{\mu_g}^{\mu_M} be the operator-composed function classes defined in the source, and let Cμg0,(1)\mathcal{C}_{\mu_g}^{0,(1)} and DμgμM,(1)\mathcal{D}_{\mu_g}^{\mu_M,(1)} denote their subclasses with one-dimensional gg and scalar linear operator MM. One-dimensional worst-case conjecture. The worst-case performances are attained in these one-dimensional subclasses:

w(Cμg0;h)=w(Cμg0,(1);h),w(DμgμM;h)=w(DμgμM,(1);h).w(\mathcal{C}_{\mu_g}^{0};h)=w(\mathcal{C}_{\mu_g}^{0,(1)};h),\qquad w(\mathcal{D}_{\mu_g}^{\mu_M};h)=w(\mathcal{D}_{\mu_g}^{\mu_M,(1)};h).

The conjecture is motivated by the authors' numerical experiments, but the supplied text gives no proof or resolution.

References

Primary source

Nizar Bousselmi, Julien M. Hendrickx and François Glineur, “Interpolation Conditions for Linear Operators and Applications to Performance Estimation Problems”, arXiv:2302.08781 (2024).

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.