cartan

Optimisation Benchmarks

Wall-clock timing for Riemannian optimisation algorithms. cartan is compared against geoopt (RiemannianSGD) for gradient descent and against geomstats for the Frechet mean estimator. All times are median over 20 repetitions.

Riemannian Gradient Descent

RGD wall-clock time: cartan vs geoopt across dimensions on the sphere
Wall-clock time for 200 iterations of Riemannian gradient descent on the sphere, minimising the squared geodesic distance to a target point. cartan's RCG (dashed) shows the convergence advantage of conjugate gradient.

cartan provides three Riemannian optimisers:

Optimisercartangeoopt equivalent
RGD (gradient descent)minimize_rgdRiemannianSGD
RCG (conjugate gradient)minimize_rcgn/a
RTR (trust region)minimize_rtrn/a

RCG and RTR are cartan-only; geoopt has no second-order Riemannian optimiser.

Frechet Mean

Frechet mean wall-clock time: cartan vs geomstats across sample sizes
Wall-clock time for computing the Fréchet mean of random points on . cartan uses an iterative gradient scheme; geomstats uses its built-in FrechetMean estimator.