BRYAN CHEN Tutorial: Add Annealed Sinkhorn tutorial notebook

[BryanBradfo]

This PR adds a tutorial notebook exploring the Annealed Sinkhorn algorithm, based on the paper:

• Annealed Sinkhorn for Optimal Transport: convergence, regularization path and debiasing by Lénaïc Chizat (2024).

The tutorial (docs/tutorials/linear/Annealed_Sinkhorn.ipynb) covers:

1. Setup of an OT problem and calculation of the exact cost.
2. Analysis of the speed-accuracy trade-off for standard Sinkhorn.
3. Implementation and convergence analysis of standard Sinkhorn (LSE mode).
4. Implementation and convergence analysis of Annealed Sinkhorn (LSE mode, kappa=0.5).
5. Implementation and convergence analysis of Debiased Annealed Sinkhorn (LSE mode, kappa=2/3).
6. Comparison of the convergence trajectories of the three methods.
7. Visualization of the final transport plans obtained by each method.

The experiments visually confirm the findings of the paper, notably the improved convergence rate of the Debiased Annealed Sinkhorn compared to the standard Annealed version, often approaching or improving upon the Pareto front of standard Sinkhorn. Manual LSE implementations were used to track per-iteration suboptimality during annealing. The symmetric version is not covered as noted in the notebook.

The main file added is docs/tutorials/linear/Annealed_Sinkhorn.ipynb.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[marcocuturi]

super nice work! thorough and clear.