Comparison Guide

How dynaris compares to other state-space / Kalman filter libraries.

Feature	dynaris	dynamax	statsmodels	filterpy	pykalman
Backend	JAX	JAX	NumPy/Cython	NumPy	NumPy
GPU/TPU	Yes	Yes	No	No	No
JIT compilation	Yes	Yes	No	No	No
Autodiff (gradients)	Yes (full)	Yes	Limited	No	No
Kalman filter	Yes	Yes	Yes	Yes	Yes
RTS smoother	Yes	Yes	Yes	Yes	Yes
Extended KF	Yes	Yes	No	Yes	No
Unscented KF	Yes	Yes	No	Yes	No
Particle filter	Yes	Yes (via Blackjax)	No	Yes	No
Hamilton filter	Yes	Yes	Yes	No	No
DLM components	Yes (6 types)	No	Yes (via UnobservedComponents)	No	No
Model composition (+)	Yes	No	No	No	No
Dynamic Factor Models	Yes	No	Yes	No	No
Bayesian (MCMC)	Yes (NumPyro)	No (manual)	No	No	No
MLE estimation	Yes (JAX grad)	Yes	Yes (scipy)	No	Yes (EM only)
EM estimation	Yes	Yes	No	No	Yes
Batch (vmap)	Yes	Yes	No	No	No
Missing data	Yes (NaN)	Yes	Yes	Manual	No
Forecast	Yes	Yes	Yes	Manual	No
Plotting	Yes (matplotlib)	No	Yes	No	No
pandas integration	Yes	No	Yes	No	No

When to use dynaris

• You want composable DLM components (trend + seasonal + regression via +)

• You need gradient-based optimization through the filter (MLE, Bayesian)

• You want GPU/TPU acceleration for large-scale or batch inference

• You need nonlinear filters (EKF, UKF, particle) alongside linear DLMs

• You want regime-switching models with the Hamilton filter

• You need Bayesian estimation with NUTS/HMC via NumPyro

When to use alternatives

• dynamax: If you want a general-purpose JAX state-space library with more model types (HMMs, LDS) and variational inference.

• statsmodels: If you need a mature, well-tested library with extensive econometric functionality and don’t need GPU/autodiff.

• filterpy: If you need a lightweight, NumPy-only library for real-time tracking applications (robotics, navigation).

• pykalman: If you need a simple EM-based Kalman filter with minimal dependencies and don’t need nonlinear filters or GPU.