SPLIT (Segmented Posteriors for Long-Inspiral Templates)

SPLIT is a Python package designed for Loosely-Coherent EMRI (Extreme-Mass-Ratio Inspiral) Parameter Estimation on Multi-GPUs.

SPLIT exploits the physical hierarchy of the system by decomposing the parameter space into static and evolving parameters to construct a natural loosely-coherent inference framework. The primary motivation for this hierarchical approach is to ensure robustness against non-stationarities in the detector data and potential waveform modeling inaccuracies. In purely coherent inference, these effects accumulate over the long inspiral duration, leading to severe signal dephasing and strong parameter biases. SPLIT mitigates this by slicing the data into Nblocks blocks and inferring the joint parameter space loosely-coherently with a more forgiving joint prior and likelihood across all blocks.

Full documentation is available at semi-coherent-posteriors-for-long-inspiral-templates.readthedocs.io.

If you use any part of this repository in your work, please cite the paper and the repository. See the Citation section below for the preferred citation.

Features

  • Multi-GPU Orchestration: Utilizes CuPy and the multiprocessing library to distribute parallel likelihood evaluations across dedicated GPUs.

  • Eryn MCMC Sampler: Hosts the core inference pipeline, natively utilizing Eryn’s “branches” (enabling static vs. evolving parameter decomposition) and “leaves” (handling multiple independent blocks), which significantly simplifies the loosely-coherent inference architecture.

  • Custom MCMC Moves: Implements customized block updating moves like SequentialAdaptiveBlockedGibbsGaussianMove and SequentialBlockedGibbsStretchMove to efficiently search the “static” and “evolving” parameter branches.

  • Markovian Student-t Prior: A heavy-tailed Student-t transition probability between consecutive blocks, penalizing excessive deviations from theoretical vacuum-GR trajectories, but still allowing flexibility to account for genuine waveform inaccuracies.

  • Student-t Block Likelihood: A block-independent heavy-tailed Student-t likelihood for robustness against non-stationarities in the data such as noise glitches.

  • Automated Diagnostics: Tracks Gelman-Rubin convergence and autocorrelation times natively, and automatically produces comprehensive diagnostic plots for the block-level evolving parameters, static parameters, and the joint parameter set projected at t=0.

Compatibility

Note: SPLIT is currently GPU-only. It is implemented for extreme-mass-ratio inspirals (EMRIs) in the LISA band using the FastEMRIWaveforms (FEW) package and requires NVIDIA GPUs with CUDA 12.x drivers. CPU-only evaluation of N blocks is not supported and would be prohibitively expensive in any case.