The critical gradient extrapolation conjecture

Consider gradient descent run for a fixed number NN of steps with constant stepsize h(0,1]h\in(0,1], starting from x0x_0 with x0x=D\lVert x_0-x_\star\rVert=D. For simple extrapolation by a factor cc, let xσx_\sigma denote the extrapolated point and let LL be the smoothness constant. Critical gradient extrapolation conjecture. There exists ccrit,grad=1+Θ(1/N)c_{\mathrm{crit,grad}}=1+\Theta(1/N) such that, for every c[1,ccrit,grad]c\in[1,c_{\mathrm{crit,grad}}], simple extrapolation has the tight worst-case bound

f(xσ)LDNhc+1.\lVert\nabla f(x_\sigma)\rVert\leq\frac{LD}{Nhc+1}.

The conjecture formalizes the observed 1+O(1/N)1+O(1/N) shrinkage of the critical extrapolation factor for gradient-norm performance and predicts that the simple bound remains tight throughout this interval.

Sources & referencesView supporting material

Primary source

Alan Luner and Benjamin Grimmer, “On Averaging and Extrapolation for Gradient Descent”, arXiv:2402.12493 (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.