Overview of cd-dynamax

The primary goal of this codebase is to extend dynamax to a continuous-discrete (CD) state-space-modeling setting, that is, to problems where

• the underlying dynamics are continuous in time,
• and measurements can arise at arbitrary (i.e., non-regular) discrete times.

To address these gaps, cd-dynamax modifies dynamax to accept irregularly sampled data and implements classical algorithms for continuous-discrete filtering and smoothing.

Mathematical Framework: continuous-discrete state-space models

In this repository, we build an expanded toolkit for filtering, forecasting and learning dynamical systems that underpin real-world messy time-series data.

We move towards this goal by working with the following flexible mathematical setting:

• We assume there exists a (possibly unknown) stochastic dynamical system of form

$$dx(t) = f(x(t), u(t), t)dt + L(x(t), u(t), t) dw(t)$$

where $x \in \mathbb{R}^{d_x}$, $x(0) \sim p(x_0; \varphi_{x_0})$, $f$ is a possibly time-dependent drift function, $L$ is a possibly state and/or time-dependent diffusion coefficient, $u(t)$ denotes optional input covariates, and $dw$ is the derivative of a $d_x$-dimensional Brownian motion with covariance $Q$.

• We assume data are available at arbitrary times $\{t_k\}_{k=1}^K$ and observed via a measurement process dictated by

$$p(y(t_k) \mid x(t_k), u(t_k), t_k; \varphi_y)$$

The library provides the Gaussian model classes ContDiscreteLinearGaussianSSM and ContDiscreteNonlinearGaussianSSM, together with a ContDiscreteNonlinearSSM class for generic initial conditions and observation distributions.

We denote the collection of all parameters as $\theta = \{f,\ L,\ \varphi_{x_0},\ Q,\ \varphi_y \}$.

Note:

• The Gaussian model classes ContDiscreteLinearGaussianSSM and ContDiscreteNonlinearGaussianSSM use Gaussian observation noise.

  • At a high level, $\varphi_y$ collects the corresponding emission mean and covariance parameters.
  • These models remain in the standard continuous (dynamics) - discrete (observation) setting with conditionally independent observation noise across observation times.

• ContDiscreteNonlinearSSM supports generic initial conditions and generic observation distributions for nonlinear CD-SSMs.

  • These observation distributions can depend on state, inputs, and time.
  • This includes non-Gaussian emissions such as Poisson observations.

• Other extensions of the overall framework include categorical state spaces and additional non-Gaussian observation models.

  • These can fit into our broader code framework, and some related cases are already covered in dynamax, but they have not been our main focus here.

cd-dynamax goals and approach

For a given set of observations $Y_K = [y(t_1),\ \dots ,\ y(t_K)]$, we wish to:

• Filter: estimate $x(t_K) \ | \ Y_K, \ \theta$
• Smooth: estimate $\{x(t)\}_t \ | \ Y_K, \ \theta$
• Predict: estimate $x(t > t_K)\ |\ Y_K, \ \theta$
• Infer parameters: estimate $\theta \ |\ Y_K$

All of these problems are deeply interconnected.

• In cd-dynamax, we enable filtering, smoothing, and parameter inference for a single system under multiple trajectory observations ($[Y^{(1)}, \ \dots \, \ Y^{(N)}]$).

  • In these cases, we assume that each trajectory represents an independent realization of the same dynamics-data model, which we may be interested in learning, filtering, smoothing, or predicting.
    • In the future, we would like to have options to perform hierarchical inference, where we assume that each trajectory came from a different, yet similar set of system-defining parameters $\theta^{(n)}$.

• We implement such filtering/smoothing algorithms in an efficient, autodifferentiable framework.

  • We enable usage of modern general-purpose tools for parameter inference (e.g., stochastic gradient descent, Hamiltonian Monte Carlo).

• In cd-dynamax, we take onto the parameter inference case by relying on marginalizing out unobserved states $\{x(t)\}_t$

  • this is a design choice of ours, other alternatives are possible.
  • This marginalization is performed (approximately, in cases of non-linear dynamics) via filtering/smoothing algorithms.

Codebase description and status

The cd-dynamax codebase extends the dynamax library to support continuous-discrete state space models, where observations are made at specified discrete times rather than at regular intervals.

• We leverage dynamax code

  • Currently, based on a local directory with Dynamax release 0.1.5

• We have implemented the cd-dynamax codebase to deal with continuous-discrete linear, nonlinear Gaussian, and generic nonlinear models, along with several filtering and smoothing algorithms.

• The codebase is organized into several key directories:

.
├── cd_dynamax/                  # Source code for cd-dynamax library
│   ├── src/                     # Core source code
│   │   ├── continuous_discrete_linear_gaussian_ssm/  # CD-LGSSM models and algorithms
│   │   ├── continuous_discrete_nonlinear_gaussian_ssm/ # CD-NLGSSM models and algorithms
│   │   ├── continuous_discrete_nonlinear_ssm/ # CD-NLSSM models with generic initial/emission distributions
│   │   ├── ssm_temissions.py    # Modified SSM class for discrete emissions
│   │   └── utils/               # Utility functions and example models
│   └── dynamax/                 # Original dynamax library (as a submodule)
├── demos/                       # Python demos showcasing cd-dynamax functionality
│   ├── python/scripts/          # Python scripts for running demos
│   ├── python/notebooks/        # Jupyter notebooks for interactive demos
│   └── python/configs/          # Configuration files for demos
└── tests/                       # Tests for cd-dynamax functionality

Demos

We provide a set of demos that showcase key functionality of cd-dynamax.

These scripts and notebooks illustrate how to learn components of continuous-discrete SDEs from data.

For instance:

• Filtering-based likelihood tutorial to filtering-based likelihood computation for continuous-discrete SDEs.

• SGD-based model fitting tutorial to SGD-based fitting of continuous-discrete SDE model to data.

• MCMC-based model fitting tutorial to MCMC-based fitting of continuous-discrete SDE model to data.

Tests

• Several tests to establish cd-dynamax general functionality, as well as linear and non-linear filters/smoothers tests: e.g., checks that non-linear algorithms applied to linear problems return similar results as linear algorithms.

Makefile

• We provide a Makefile to automate common tasks, such as running tests and demos.

• To run all tests, simply execute:

make test

• For linting, we use ruff:

make lint

• We can also format files using ruff:

make clean

• The docs can be built using mkdocs as:

make build_docs

Installation

Install from PyPI (recommended), from source in editable mode, or with a Conda-managed environment.

---

Option 1: Install from PyPI (recommended)

# Create and activate a virtual environment
python -m venv .venv        # or `uv venv`
source .venv/bin/activate   # on macOS/Linux
.venv\Scripts\activate      # on Windows

# Upgrade pip
pip install --upgrade pip

# Install latest release from PyPI
pip install cd-dynamax

cd-dynamax is currently not available on Conda Forge.

---

Option 2: Install from source (editable)

# Create and activate a virtual environment
python -m venv .venv        # or `uv venv`
source .venv/bin/activate   # on macOS/Linux
.venv\Scripts\activate      # on Windows

# Upgrade pip
pip install --upgrade pip

# Install in editable mode for local development
pip install -e .[dev]

---

Option 3: Conda environment + pip install

# Create and activate a Conda environment with Python 3.11
conda create -n cd_dynamax python=3.11
conda activate cd_dynamax

# Install latest release from PyPI
pip install cd-dynamax

GPU support

If you want GPU acceleration with JAX, you must install a CUDA-enabled jaxlib wheel.

Check the JAX installation docs for the exact commands for your system.

---

Notes

• pip install -e . puts the repo in editable mode, so changes to source code are immediately available without reinstalling.

• If you plan to use plotting features that rely on graphviz, make sure the system binary is installed:

  • macOS: brew install graphviz
  • Ubuntu/Debian: sudo apt install graphviz
  • Windows (conda): conda install graphviz

• The [dev] extra installs additional developer tools (like pytest).

  • Once your environment is installed, you can run automated tests:

  pytest