Benchmarks / Solvers
Solvers — head-to-head
basin's five general optimizers — gradient descent, Nelder–Mead, BFGS,
L-BFGS, and CMA-ES — from several seeded starting points sampled
uniformly in each problem's domain. Each run is capped on a fixed 20.00 ms wall-clock budget and stopped early on reaching suboptimality 1e−10, so a line that ends at the right
edge of a panel never made it within the time given. Lines within a
panel share the same f(x₀); lower and
further left is better.
Rosenbrock
n = 10, starts sampled uniformly in [−2, 2]10.
Seed 1 f(x₀) = 700
Seed 2 f(x₀) = 5836
Seed 3 f(x₀) = 3435
Seed 4 f(x₀) = 7689
Seed 5 f(x₀) = 2849
Seed 6 f(x₀) = 5014
wall-clock time
Measured 2026-05-28 on AMD Ryzen 9 7900 12-Core Processor
(linux/x64). Every solver runs on the Vec<f64> backend, capped at
20.00 ms per (solver, seed) run. Each
per-iteration timestamp is the median over 11 repetitions of the same
deterministic run; absolute times are machine-specific — compare curves
within a panel, not across machines. Some seeds land in the basin of
Rosenbrock's spurious local minimum near (−1, 1, …, 1) (which appears for n ≥ 4); a line that flattens around f ≈ 4 is a solver caught in that trap.
For the basin-versus-other-libraries view, see the competitors axis; for backend cost on the same solvers, see the backends axis. To watch the same solvers converge interactively, try the visualizer.