GPU-aware data-center scheduling, flat-minima stepsizes, and an LLM-written solver benchmark
The day in optimization
Optimization theory led the day. New results covered adaptive stepsizes, sampling-based control, and quasi-Newton methods. Verification and automation were the second theme: a Rust SDP solver checked in Verus, a Lean autoformalization pipeline, and a benchmark of LLM agents designing OR algorithms. Applied papers on grid-aware AI data centers and GPU-accelerated facility location rounded out the day.
Theory
Stepsizes and first-order methods. Critically damped Polyak stepsizes multiply the classical Polyak rule by 4 and a mild damping factor. The authors say the factor 4 makes the iterates equioscillate around the set of minimizers. They report the first uniform-in-time guarantees on both sharpness and loss reduction for a broad class of locally quadratic objectives. Sharpness here means how steeply the loss curves, so the rule biases gradient descent toward flatter minima. A related paper, Safeguarded Learning-Assisted Nonlinear Conjugate Gradient Methods, keeps the standard two-term NCG structure. It adds small online-trained models that predict a coefficient only after a restart or a stall, so the classical safeguards stay in place. Dimension-Free Buffer Bounds for Compressed-Momentum Error Feedback shows that the error buffer need not carry the usual compressor contraction constant. The trick is to re-inject the compressed momentum into the accumulator before compressing. The bound is ‖E_t‖_F ≤ (1−β₁)‖G_t‖F + γ‖E{t−1}‖_F, which says each step's buffer is a small multiple of the new gradient plus a decayed copy of the old buffer.
Newton-type and manifold methods. R2N is a modified quasi-Newton method that approximately minimizes a quadratic model plus adaptive regularization. The paper treats convergence, worst-case complexity and implementation, and the objective is only assumed to be C¹. Preconditioning for quadratic matrix optimization via constraint reshaping takes a different route for ill-conditioned problems over orthogonality or symplectic constraints. Instead of choosing a Riemannian metric, it redesigns the problem to get the narrowest possible Hessian spectrum at the minimizer.
Sampling-based control. Semi-global convergence and finite-sample bounds for MPPI studies Model Predictive Path Integral control in the low-temperature setting. For objectives with a unique global minimum, allowing non-convexity and non-smoothness, the population-level iteration converges semi-globally to that minimum. With high probability the finite-sample iteration comes close to it. A second control paper, Computing Optimal Non-stationary Policies for Probabilistic Reach-Avoid Problems, shows that stationary policies can be suboptimal for maximizing the probability of reaching a goal while avoiding bad states. It develops an algorithm for the non-stationary optimum.
Verification and automation. Formally Verified Semidefinite Programming describes a hand-written Rust first-order SDP solver. Claude generated its full verification in Verus. Paper2Lean converts research papers into Lean projects through a dependency graph and specialized agents. The FrontierOR Challenge asks whether LLM agents can design OR algorithms that scale. Its abstract says that in one-shot evaluation the strongest model matches Gurobi to within 1% on some instances. That is a benchmark result, not a general claim.
Hard instances and negative results. Integer Programming Strategies for the Snake-in-the-Box Problem targets the longest induced path in a hypercube. Optimality is proven only up to dimension 8, and the smallest open instance has about 512 binary variables. A non-zero gap for the planar Mumford-Shah convex lifting shows the calibration-based convex lifting is not sharp in 2D. A lifted competitor beats the crack-tip subgraph, so a strict gap separates the original and lifted problems.
In practice
Flexible AI data centers. The day's paper Optimal Scheduling of Multi-Site AI Data Centers in Day-Ahead Markets models GPU power caps, deferrable jobs and job placement across sites against electricity prices. It fits a wider push to make AI load grid-friendly. NVIDIA's blog describes power-flexible AI factories that adjust power use during peak demand while protecting priority workloads, developed with Emerald AI, EPRI, National Grid and Nebius. A search also surfaced a Phoenix trial under EPRI's DCFlex initiative, in which the utility set a 25% reduction target for an AI cluster. Both are vendor-side accounts rather than independent audits. A Boston University study estimates that AI data centers could supply regulation reserves of roughly 18% to 55% of average power.
Planning at an energy retailer. A Timefold and AWS case study of e-switch reports cost savings above $100 million and over 10 million kg less CO₂ from planning optimization. These are the vendor's claims, and the case study is not from this week.
Worth watching
- Covering Location Optimization at GPU Speed: a GPU multi-trajectory search for maximal covering location using exact incremental one-swap gains.
- PWLNet: a Python framework that embeds neural surrogates in Pyomo through exact piecewise-linear or MILP reformulations.
- Compact Slack-Free Penalty Functions for HUBO: exact penalties for inequality constraints on binary-encoded integers, with no slack variables.
- A Tutorial on Complementarity-Based Contact Models for Robotics: a self-contained introduction to cone complementarity formulations in simulation.
Sources: - Timefold and AWS: e-switch - NVIDIA: power-flexible AI factories - Boston University FlexDC paper
Try ProvenDecisions free: type, photograph or say an optimization problem; every answer is checked by a solver.