Design Rationale

User-facing design documentation for normix — the why behind the API and algorithms. For class and method reference see the API Reference.

Topics

• Exponential Family Core
• Mixture Architecture
• EM Framework
• Solvers and Bessel Functions
• Why not gradient descent

Quick lookups

Question	Where to read
Why two classes for each mixture distribution?	Mixture Architecture § 1
Why three classmethod tiers for the log-partition?	Exponential Family Core § 2
What does 'det_sigma_x' regularisation do?	EM Framework § 5
Why CPU backend for Bessel and GIG solve?	Solvers and Bessel Functions § 3–4
Is jax.grad(log_kv) the same as log_kv_moments.d_arg?	Solvers and Bessel Functions § 3.1; Bessel functions and log_kv
Why not Adam / L-BFGS on the observed NLL?	Why not gradient descent

See also

• Mathematical Background — formal derivations

• API Reference — module and class reference

• Quickstart — install and first fit