Optimal first-order oracle complexity for heterogeneous Hölder-smooth sums

Let J\mathcal{J} be a finite index set, and for each jJj\in\mathcal{J} let fjf_j be a convex (Mj,vj)(M_j,v_j)-Hölder smooth function. Let x0x_0 be the initial point, xx^* an optimizer of jJfj(x)\sum_{j\in\mathcal{J}}f_j(x), and suppose

12x0x2R.\frac{1}{2}\|x_0-x^*\|^2\leq R.

For a target accuracy ϵ>0\epsilon>0, the optimal first-order oracle complexity conjecture. The optimal first-order oracle complexity for minimizing jJfj(x)\sum_{j\in\mathcal{J}}f_j(x) is

jJ[cˉjJ13vj1+3vj(Mjϵ)21+3vjR1+vj1+3vj],\sum_{j\in\mathcal{J}} \left[\bar c_j|\mathcal{J}|^{\frac{1-3v_j}{1+3v_j}}\left(\frac{M_j}{\epsilon}\right)^{\frac{2}{1+3v_j}}R^{\frac{1+v_j}{1+3v_j}} \right],

where cˉj\bar c_j depends only on universal constants and vjv_j. This conjecture proposes the optimal dependence on the number of summands, improving the match between the upper and lower complexity bounds; its resolution would characterize the optimal performance of first-order methods for heterogeneous sums.

Sources & referencesView supporting material

Primary source

Benjamin Grimmer, “On Optimal Universal First-Order Methods for Minimizing Heterogeneous Sums”, arXiv:2208.08549 (2023).

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.