no. 12 · Tuesday 10 November 2026
The Convergence Rate of Accelerated Proximal Gradient Methods with General Inertial Term for Convex Composite Multiobjective Optimization
Abstract
We study accelerated proximal gradient methods with general stepsizes for convex composite multiobjective optimization, where each objective is a smooth convex term plus a proper closed convex term and the proximal subproblem keeps a separate offset for each objective. For a two-parameter root momentum rule with trajectory-dependent coefficients, we construct a new energy function and establish an energy identity. This yields sharper convergence guarantees: boundedness, weighted square summability of the displacements, a merit-function rate of $o(k^{-2})$, and displacement decay of $o(k^{-1})$ whenever the first parameter converges to a value in $(0,1)$; the second parameter need not converge or have bounded total variation. At the critical value $α=3$, we obtain $u_0(x^N)\le \frac{2\ell R}{N(N+3)},$ together with $O(1/N)$ displacement decay, square summability, and weak Pareto convergence. The remaining little-$o$ question at $α=3$ is reduced to the vanishing of a nonnegative critical energy. For $1<α<3$, we prove the rate $O(N^{-(α-1)})$, and for every $α>0$ we obtain $u_0(x^k)\to0$, $F_i(x^k)\to c_i$, and $\sum_k D_k/k<\infty$. A JOS1 experiment numerically validates the predicted little-$o$ separation and the rate behavior.
- fields
- Optimization and Control