Hi! Id like to propose adding the SSBroyden and SSBFGS optimizers to optax.contrib.
These are full-matrix quasi-Newton methods from Urbán et al. (2025) that use a self-scaled inverse Hessian update. The paper demonstrates strong results on PINNs and similar small-to-medium scale problems.
I've ported the PyTorch implementation from SciMBA to JAX/optax following the existing patterns (scale_by_lbfgs / lbfgs). The implementation includes:
ssbroyden() and ssbfgs() as separate public optimizers- A shared
scale_by_ss_quasi_newton() core transformation
- Integration with
scale_by_zoom_linesearch (same pattern as optax.lbfgs)
- Tests on quadratic and Rosenbrock functions
I have a working implementation ready. Would you be open to a PR?
Reference:
Urbán, J. F., Stefanou, P., & Pons, J. A. (2025). Unveiling the optimization process of physics informed neural networks: How accurate and competitive can PINNs be? Journal of Computational Physics, 523, 113656. DOI
Hi! Id like to propose adding the SSBroyden and SSBFGS optimizers to
optax.contrib.These are full-matrix quasi-Newton methods from Urbán et al. (2025) that use a self-scaled inverse Hessian update. The paper demonstrates strong results on PINNs and similar small-to-medium scale problems.
I've ported the PyTorch implementation from SciMBA to JAX/optax following the existing patterns (
scale_by_lbfgs/lbfgs). The implementation includes:ssbroyden()andssbfgs()as separate public optimizersscale_by_ss_quasi_newton()core transformationscale_by_zoom_linesearch(same pattern asoptax.lbfgs)I have a working implementation ready. Would you be open to a PR?
Reference:
Urbán, J. F., Stefanou, P., & Pons, J. A. (2025). Unveiling the optimization process of physics informed neural networks: How accurate and competitive can PINNs be? Journal of Computational Physics, 523, 113656. DOI