Sampling-Optimisation-Geometry-Bench
OptimisationBench
Algorithms, rates and proof mechanisms.
Public-dev · shared task identities · Dataset manifest
Choose a task
PL descent: the sign of the contraction factorsogb-002 · NL L3 · Lean L3→A quadratic envelope from one midpointsogb-003 · NL L2 · Lean L2→Why a sphere gradient is a projected ambient gradientsogb-005 · NL L3→Reading a PL inequality before using an algorithmsogb-010 · NL L1→Which space contains the sphere gradient?sogb-011 · NL L1→A quadratic gradient through the actual sphere differentialsogb-016 · NL L3 · Lean L3→Why a global Lipschitz function bounds its supporting vectorssogb-019 · NL L3 · Lean L2→From subgradient energy to a finite-run best iteratesogb-020 · NL L3→From a selected-curve chord inequality to its first-order boundsogb-022 · NL L3 · Lean L3→An actual gradient flow: objective decay with the zero-curvature branchsogb-023 · NL L4 · Lean L4→Can the SPHMC smoothing argument survive weak convexity?sogb-026 · NL L4→Chewi Optimization Exercise 3.2: the balanced gradient stepsogb-029 · NL L3→Scope and difficulty
These are views of the same catalog, not new datasets. A cross-listed core keeps its identity, family, split and source binding. NL/Lean levels are provisional, not empirically calibrated.
Topics: Optimisation, Riemannian Optimisation.