v0.1.4

This release fixes a minor error with the PSD normalization, bringing the code into agreement with the convention used in both Solin & Sarkka (2014) and Rubenzahl et al. (2026) by removing an erroneous factor of 2pi.

v0.1.3

This release includes functionality for generalized multivariate output data y with shape (NxD), with corresponding measurement uncertainties R with shape (NxDxD). The standard case of a 1-D time series is auto-casted into the multidimensional shape, and D>1 datasets can simply be input into the noise attribute of the GaussianProcess object and the y argument for conditioning the GP.

A new tutorial (Multivariate) is also added, which uses the FF' model as an example of a parallel time series (a value and its derivative) using the new functionality with D=2.

v0.1.2

This release drops the requirement of installing tinygp from its Github source. Version 0.3.1 of tinygp has been released (https://github.com/dfm/tinygp/releases/tag/v0.3.1) which includes the Block matrix object definitions used in smolgp. This version and future versions of smolgp now require tinygp>=0.3.1.

v0.1.1

Some minor changes and improvements, including updated documentation, tutorials, and moving the dense integrated SHO kernel from testing into smolgp.kernels.dense so it can be imported.

v0.1.0

smolgp v0.1.0 (Jan 29, 2026)

This is the initial release of smolgp which includes general-use functionality, tests, and initial documentation.

Highlights

With v0.1.0 of smolgp, users can:

• Build Gaussian Process models from kernels or sums/products of kernels
• Condition GPs on data (which may or may not include integrated measurements)
• Utilize parallel solvers on GPU
• Make predictions at arbitrary times
• Decompose multicomponent models into individual predictive components.

Planned upcoming features

These are in-development for a v1.0 release to match all functionality of tinygp:

• Sampling procedure using the state space definition.
• Nonzero mean functions (see Fittng a Mean Function).
• More expressive noise models (see tinygp.noise).
• Full conditioned covariance matrix as a product of the GP object

Other upcoming features

These are planned features that smolgp is uniquely capable of incorporating:

• Predictions at arbitrary times with arbitrary exposure times
• Tutorials for building models with multivariate inputs/outputs, and spatiotemporal GPs.