Limited-memory first-order methods, a Douglas–Rachford generalisation, and a new Reinhardt proof
The day in optimization
First-order theory led the day. Several papers ask how much freedom an algorithm has in its memory, step-sizes or assumptions, and what that freedom buys. Convex-geometry and MIP papers filled out the rest of the list. Fei Xue also posted a claimed proof of the Reinhardt conjecture, the lattice-packing problem for centrally symmetric convex disks. The question is a famous open one, so the claim needs community checking.
Theory
Limited memory and optimized methods. A general first-order method may add a combination of all past gradients at every step. Many good methods need only a few stored vectors. Vernimmen, Glineur, Goujaud et al. define "limited-memory" methods through the lower-triangular coefficient matrix that describes a fixed-coefficient method. The abstract title also promises identification, reconstruction and an "H-duality" for optimized methods. The excerpt we have is truncated, so we cannot say what the results are.
Operator splitting. Briceño-Arias, Roldán and Taylor relax the symmetry of Douglas–Rachford, a classical method for finding a zero of the sum of two monotone operators. Standard Douglas–Rachford reflects both operators the same way and uses one step-size. Their version uses relaxed resolvents, so each operator gets its own step-size. They report that this extra freedom is available precisely when the operators are cocoercive. Related work from Alcantara treats degenerate preconditioned proximal point methods for nonmonotone inclusions. It writes them as forward methods under "semimonotonicity" assumptions, which allow indefinite weights.
Complexity under non-standard growth. Kovalev and Malinovsky study convex problems where the subgradient-norm bound depends on the objective value. They prove a lower bound on the iterations any first-order method needs. They also give two algorithms that match it up to logarithmic factors. One is projected gradient descent with a warm-up stepsize schedule. A lower bound limits every method. The matching upper bound is what shows the rate is achievable. Rudomanov, Fomin, Gasnikov et al. do the analogous job for monotone variational inequalities. There the Jacobian norm is bounded by an affine function of the operator norm, ‖DF(z)‖ ≤ L₀ + L₁‖F(z)‖ (a "generalised smoothness" condition that lets curvature grow with the size of F). Weraarchakul and Zheng propose a distance-certified Polyak-type stepsize for subgradient methods under Hölder growth. It targets the usual Polyak stepsize's need to know the optimal value.
Conic and constrained methods. Jiang, Shao and Vandenberghe reformulate conic LPs as exact-penalty problems over truncated cones, using self-dual embeddings. Bregman distances then give simpler subproblems than Euclidean projections for non-polyhedral cones. Xin He builds Nesterov-accelerated augmented Lagrangian methods for convex problems with nonlinear inequality constraints. Zuo, Lin, Liu et al. give a first-order smoothed exact-penalty method (SEPi-Prox) for nonconvex regularized problems with equality constraints. Hoteit et al. present COSMIC, a continuous-time solver for inequality constraints built on sliding-mode control. They prove its trajectories reach a manifold in finite time.
Convex relaxations for energy storage. Morales addresses a known flaw of convex storage models. Dropping the ban on simultaneous charging and discharging lets the relaxation cheat in set-point tracking, because the error depends only on net power. He shows that the convex hull of the single-period error epigraph under complementarity is the epigraph of a parabolic cylinder. This gives a second-order cone inequality.
Discrete optimization. Fiorini, Kober and McFarland give a polynomial-time algorithm for maximum-weight independent set on graphs embedded in a bounded-genus surface with a bounded number of vertex-disjoint odd faces. Abiad, Golak and Schulz combine branch-and-cut, clique subproblems and spectral bounds for maximum distance-independent sets. Yarkony, van Hoeve et al. shrink arc-flow vehicle-routing formulations by dual-guided aggregation. They report recovering the LP bound of the original networks.
In practice
Operators use mixed-integer programming for real-time power system dispatch. A CIGRE 2026 session abstract describes a framework developed for Colombia's system operator. It models ramping limits, minimum up/down times, hydro cascades and combined-cycle plants. The abstract reports fewer manual interventions, better alignment of scheduled and actual generation, and easier integration of renewables. These are simulation results, and the abstract gives no figures.
In routing, a Timefold and AWS write-up on e-switch credits optimization with large annual savings. The vendor-side claim is over $100 million a year, and the page does not say how it was measured, so treat it with caution. The same search turned up vendor numbers such as a custom routing platform reportedly cutting cost per stop by 19%. These are not independently verified. Neither is dated to this week.
Worth watching
- Qu and Lee show a gap in the integer decomposition property for disjunctions of three staircase simplices, while two simplices always have it.
- Fischetti's note asks whether the LP in rank-one "anti-loop" edits of a small reasoning model is needed. A simple sort-based substitute replaces it.
- Kovalev, Stonyakin and Gasnikov study nonsmooth convex optimization from noisy pairwise comparisons only.
- Wang and Ning analyse long-time behaviour of continuous-time AdamW, including bias correction and weight decay.
- Zurek and Chen report that plug-in RL with a generative model needs O(SA(1−γ)⁻³ε⁻² log(SA/δ)) samples, where S, A, γ and ε are state count, action count, discount and accuracy.
Sources: - Limited-Memory First-Order Methods - Douglas–Rachford with Relaxed Resolvents - Colombian real-time optimization framework (CIGRE) - e-switch with Timefold and AWS
Try ProvenDecisions free: type, photograph or say an optimization problem; every answer is checked by a solver.