SPLIT: Loosely-coherent inference of long-inspiral gravitational-wave sources for robustness against modeling and data outliers

arXiv:2606.13728 · gr-qc, astro-ph.HE, astro-ph.IM · Submitted 2026-06-11 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: Hi, I'm Vera.

Jocelyn: And I'm Jocelyn, here with Subrahmanyan.

Vera: We're looking at the paper: "SPLIT: Loosely-coherent inference of long-inspiral gravitational-wave sources for robustness against modeling and data outliers".

Jocelyn: Here's the gist of the paper. The gist: The SPLIT framework formulates a robust semi-coherent Bayesian inference framework that segments data into independent blocks, analyzes each block coherently, and recombines results incoherently to mitigate systematic biases in long-inspiral gravitational-wave source analysis How it works The core methodology involves segmenting the discrete time-series data into Nblocks contiguous blocks, each analyzed independently using fully-coherent methods and subsequently combined incoherently The joint posterior probability density function is then written as a product over all blocks, where the block-wise likelihoods are evaluated coherently and combined incoherently through the product over all blocks To enhance robustness against transient features, a heavy-tailed Student-t likelihood is constructed for each block This likelihood downweights outliers by choosing a Student-t distribution instead of the standard Gaussian likelihood A.

Vera: General formalism The parameter set is decomposed into static parameters ϑ and evolving parameters φi for the ith block, allowing for a joint semicoherent posterior pdf to be written as a product of terms The likelihood in Eq.

Jocelyn: (2) is rewritten as p(diφi, ϑ) = Y i p(diφi, ϑ) The semi-coherent nature manifests through Eq.

Vera: (4), where the block-wise likelihoods are evaluated coherently and subsequently combined incoherently through the product over all blocks B.

Paper summary: Vera: So we’ve covered the basic idea of SPLIT, which is this robust semi-coherent Bayesian inference framework for long-inspiral gravitational-wave sources Shubham Kejriwal et al. They are looking at EMRIs and IMRIs because they stay in band for months to years, which creates these long integration problems.

Jocelyn: The paper claims that the standard fully phase-coherent analysis in current pipelines has stringent waveform accuracy requirements, and failing to model the signal over such long durations leads to significant systematic biases.

Subrahmanyan: It proposes a solution by segmenting the data into independent blocks, analyzing each block coherently, and then recombining those results incoherently. This prevents the accumulation of phase errors that happen when you try to track everything perfectly over years of observation.

Vera: They formulate this framework using a product over all blocks where the block-wise likelihoods are evaluated coherently and then combined incoherently through that product, which is what they call semi-coherent nature in Eq. (four) <ref:2606.13728#pg3>.

Jocelyn: To make those individual blocks more robust against transient features, they construct a heavy-tailed Student-t likelihood for each block instead of the standard Gaussian likelihood. This helps downweight outliers in any single block that might be noisy or glitchy.

Subrahmanyan: They also link these independent blocks weakly via a Markovian Student-t prior to allow flexibility across blocks in the presence of unmodeled physical effects, but this is designed to penalize excessive discontinuity between adjacent blocks so you don't get bad jumps between segments.

Vera: The computational side is handled by implementing this methodology in a new Python package called SPLIT, which has a highly configurable interface and can parallelize over multiple graphics processing units or GPUs.

Jocelyn: They also propose a sequential Gibbs sampling strategy for efficient posterior recovery, which helps speed up the process of finding the final parameter estimates when you're working with these complex posteriors.

Subrahmanyan: The authors compare their semi-coherent results against fully coherent baselines in Sec. III B, III C, and III D to quantify exactly how much better their method is in terms of bias reduction, for example, showing a maximum bias of four point eight sigma versus less than zero point five sigma in eccentricity when an unmodeled effect is injected.

Vera: So the main thrust here is that by segmenting and using a heavy-tailed likelihood, you get significantly reduced systematic biases without losing too much of the true signal information, even though you have to deal with a slight drop in signal-to-noise ratio per block.

Jocelyn: This whole setup seems designed to make inference on these long signals more reliable than what's possible with standard pipelines right now. It’s about finding a way to keep the analysis stable over very long timescales when the data itself is inherently complex and noisy.

Subrahmanyan: For cosmology, this is relevant because it shows how modeling uncertainties—like unmodeled physical effects—can be managed systematically rather than just hoping they don't ruin your result in a single massive fit.

Vera: It’s a way of managing the complexity inherent in long-duration signals by breaking them down into manageable, coherent pieces and then combining them intelligently.

Jocelyn: And the next thing we need to talk about is what this all actually means for the people who are building these detectors like LISA.

Conclusion: Vera: So we’ve gone through the paper "SPLIT: Loosely-coherent inference of long-inspiral gravitational-wave sources for robustness against modeling and data outliers". It seems like they've developed a method that is much more robust for analyzing EMRIs and IMRIs.

Jocelyn: The title itself points right to the core idea: loosely coherent inference, which means you analyze parts of the signal coherently but stitch them together incoherently. This is a way to achieve robustness against modeling issues and data outliers in these long inspirals.

Subrahmanyan: What this means practically is that for future detectors like LISA, you get much more trustworthy parameter estimates from their signals than if they relied on the standard fully coherent methods when dealing with these very long sources.

Vera: It’s about moving away from a single, potentially biased estimate toward a collection of accurate estimates derived from segmented analysis, which is a major improvement in reliability for observational astronomy.

Jocelyn: The authors show that the method successfully reduces systematic biases down to less than zero point five sigma in eccentricity when compared to the truth, even when an unmodeled astrophysical effect is deliberately injected into the data.

Subrahmanyan: That level of precision improvement is what makes this work interesting from a theoretical side because it shows that this type of segmentation isn't just a trick; it’s a valid way to handle signal complexity in these regimes.

Vera: It seems like the main implication is that for EMRIs and IMRIs, we can expect much higher confidence in the parameters we extract from the data, which is crucial for testing theories about black holes and neutron stars.

Jocelyn: And this whole SPLIT package gives researchers a practical tool to actually try out this methodology with their own data injections or real observational data.

Subrahmanyan: So, in short, the paper suggests that when dealing with signals that span years, breaking the analysis into blocks and using semi-coherent techniques leads to much more reliable parameter extraction than trying to do one massive coherent fit.

Shubham Kejriwal

National University of Singapore

gr-qc, astro-ph.HE, astro-ph.IM

Submitted: 2026-06-11

Updated: 2026-10-18

Comments: before proofs accepted; 28 pages, 8 figures

Code: https://github.com/perturber/SPLIT

License: http://creativecommons.org/licenses/by-sa/4.0/

Importance score: 88/100

The gist: The gist: The SPLIT framework formulates a robust semi-coherent Bayesian inference framework that segments data into independent blocks, analyzes each block coherently, and recombines results

Key concepts

Segmented Inference: The method divides the continuous gravitational-wave data into discrete, contiguous blocks. Each block is analyzed independently using a fully coherent technique, and the results are then combined incoherently. This segmentation limits the duration over which phase errors can accumulate, improving robustness for long inspirals.
Student-t Likelihood: Instead of a standard Gaussian likelihood, this framework uses a Student-t distribution for each data block. This heavy-tailed distribution is chosen because it effectively downweights outliers in individual blocks, making the analysis more robust against transient features or data anomalies that might otherwise skew results.
Semi-coherent Posterior: The joint posterior probability density function is constructed as a product of terms, where block-wise likelihoods are evaluated coherently but combined incoherently. This structure allows for the exploitation of local coherence within blocks while maintaining overall robustness against long-term phase errors.
Fractional SNR Loss: Because the signal is analyzed in smaller blocks rather than one continuous piece, the optimal signal-to-noise ratio (SNR) is distributed across these blocks. This means each individual block's SNR will be about $\sqrt{N_{blocks}}$ lower than the full optimal SNR, but this trade-off is accepted for vastly improved parameter accuracy.

Terminology

Accuracy — 2× in this paper · explained in 1 episode
Bayesian Inference — 1× in this paper · explained in 6 episodes
Laser Interferometer Space Antenna — 1× in this paper
Robustness — 1× in this paper · explained in 5 episodes
Systematic Bias — 1× in this paper · explained in 3 episodes

Summary

The gist: The SPLIT framework formulates a robust semi-coherent Bayesian inference framework that segments data into independent blocks, analyzes each block coherently, and recombines results incoherently to mitigate systematic biases in long-inspiral gravitational-wave source analysis

How it works

The core methodology involves segmenting the discrete time-series data into Nblocks contiguous blocks, each analyzed independently using fully-coherent methods and subsequently combined incoherently The joint posterior probability density function is then written as a product over all blocks, where the block-wise likelihoods are evaluated coherently and combined incoherently through the product over all blocks To enhance robustness against transient features, a heavy-tailed Student-t likelihood is constructed for each block This likelihood downweights outliers by choosing a Student-t distribution instead of the standard Gaussian likelihood

A. General formalism

The parameter set is decomposed into static parameters ϑ and evolving parameters φi for the ith block, allowing for a joint semicoherent posterior pdf to be written as a product of terms The likelihood in Eq. (2) is rewritten as p(diφi, ϑ) = Y i p(diφi, ϑ) The semi-coherent nature manifests through Eq. (4), where the block-wise likelihoods are evaluated coherently and subsequently combined incoherently through the product over all blocks

B. A Student-t likelihood for robustness against transient features

The Student-t likelihood is chosen to downweight outliers by choosing a distribution with a heavier tail compared to the Gaussian likelihood The degrees-of-freedom parameter νk is defined such that νk > 0 recovers the Gaussian likelihood when νk → ∞ The likelihood in Eq. (11) is independent of all other blocks j ≠ i conditioned on ϑ⃗ such that Eq. (4) holds

C. A Markovian Student-t prior for robust phase tracking

To break degeneracies between orbital separation and phase parameters, a first-order discrete Markov chain prior is proposed This transition probability can be modeled as another multivariate Student-t distribution in the realistic case to account for modeling uncertainties from secular perturbative effects

D. Fractional loss of signal-to-noise ratio

The total optimal SNR is distributed among Nblocks, such that the SNR in each block ρ i is a factor ≈ √Nblocks lower than the optimal SNR of the full-duration signal ρopt The ratio between posterior widths in the ith block (σ 2 i) and in the fully-coherent case (σ 2 FC) can be written as σ 2 i / σ 2 FC ≈ Nblocks

IV. DISCUSSION

The semi-coherent pipeline is significantly more robust when an unmodeled astrophysical effect was injected, as the fully-coherent analysis incurred a maximum bias ≈ 4.8σ in eccentricity and a full-dimensional Mahalanobis distance of ≈ 1200 In comparison, the maximum bias in the semi-coherent case was ≈ 0.5σ, also in eccentricity, and a Mahalanobis distance of ≲ 0.9 across all blocks

B. Future directions

Future iterations of the SPLIT pipeline can benefit from the integration of simulation-based inference (SBI) methods, particularly conditional normalizing flows (CNFs), to mitigate the computational cost associated with inferring the large Dinferdimensional posterior A primary advantage of the Student-t likelihood constructed in Sec. II B is its heavy-tailed nature, which offers robustness against significant data outliers within any single block Future studies should validate whether the Student-t likelihood (11) can help mitigate such biases and maintain consistency with the injected values

ACKNOWLEDGEMENTS

I am grateful to Alvin J. K. Chua, Jonathan Gair, and Ilya Mandel for useful comments during the preparation of this manuscript I also acknowledge the support of the NUS Research Scholarship and the computational resources provided by the NUS IT Research and Computing Group

REFERENCES

[1] M. Colpi, K. Danzmann, M. Hewitson, K. HolleyBockelmann, P. Jetzer, G. Nelemans, A. Petiteau, D. Shoemaker, C. Sopuerta, R. Stebbins, N. Tanvir, H. Ward, W. J. Weber (2017), arXiv:2603.17072 [astro-ph IM].

[16] C. Cutler and M. Vallisneri (2007), LISA detections of massive black hole inspirals: Parameter extraction errors due to inaccurate template waveforms, Phys. Rev. D 76, 104018 (2007), arXiv:0707.2982 [gr-qc].

[35] S. Kejriwal, Split - semi-coherent posteriors for longinspiral templates (2026).

[36] L. S. Finn (1992), Detection, measurement, and gravitational radiation, Phys. Rev. D 46, 5236 (1992), arXiv:grqc/9209010 [gr-qc].

[37] C. Cutler and É. E. Flanagan (1994), Gravitational waves from merging compact binaries: How accurately can one extract the binary’s parameters from the inspiral waveform?, Phys. Rev. D 49, 2658 (1994), arXiv:gr-qc/9402014 [grqc].

[38] C. Röver, R. Meyer, and N. Christensen (2011), Modelling coloured residual noise in gravitational-wave signal processing, Classical and Quantum Gravity 28, 015010 (2011), arXiv:0804.3853 [stat ME].

[47] D. Foreman-Mackey, D. W. Hogg, D. Lang, and J. Goodman (2013), emcee: The MCMC Hammer, PASP 125, 306 (2013), arXiv:1202.3665 [astro-ph IM].

[48] M. Maggiore et al. (2020), Science case for the Einstein telescope, J. Cosmology Astropart. Phys. 2020, 050 (2020), arXiv:1912.02622 [astro-ph CO].

[51] A. Gelman and D. B. Rubin (1992), Inference from Iterative Simulation Using Multiple Sequences, Statistical Science 7, 457 (1992).

[53] S. Isoyama et al.

Improvements for AI systems

No longer suffer from significant systematic biases in long-inspiral gravitational-wave source parameter estimation by implementing a robust semi-coherent Bayesian inference framework that segments the data into independent blocks, analyzes each block coherently, and recombines the results incoherently.
The improved AI system can accurately recover parameters for Extreme and Intermediate Mass Ratio Inspirals (EMRIs/IMRIs) over long durations (months to years) by restricting phase-tracking to shorter block durations, which prevents significant accumulation of phase errors and restricts systematic biases to ≲ 0.5σ compared to a fully-coherent analysis's maximum 1D systematic bias of ≈ 4.8σ.
The system can robustly handle data outliers and instrumental artifacts by employing a heavy-tailed Student-t likelihood within each block, which makes the likelihood more robust by ensuring that data segments with large outliers do not dominate the total likelihood product (4).
The AI system can model unmodeled secular perturbations in astrophysical environments more flexibly by using a Markovian Student-t prior, allowing it to account for deviations from vacuum-GR trajectories via a transition probability that can be modeled as another multivariate Student-t distribution.
The system can efficiently traverse the high dimensionality of the joint posterior space by utilizing a sequential Gibbs sampling strategy that avoids full-dimensional jumps, allowing it to perform inference even when the total dimensionality scales linearly with the number of data blocks, which is crucial for LISA analysis where Dinfer = Nblocks × Devol + Dstatic.
The system can provide a more reliable estimate of the true parameter region by using the derived statistical measures, such as calculating the 1D z-scores for each modelled parameter θi = (ϑ, φi) as zi = θtruth ij − µij /σij and the Mahalanobis distance DMaha,i, which reveal that the semi-coherent analysis yields a significantly smaller bias compared to the fully-coherent case.

Abstract

The Laser Interferometer Space Antenna (LISA) will detect gravitational waves (GWs) from dozens of extreme- and intermediate-mass-ratio inspirals (EMRIs/IMRIs). These sources will stay in-band for months to years, offering extraordinary scientific potential. However, their fully phase-coherent analysis standard in current pipelines imposes stringent waveform accuracy requirements; failing to model the signal over such long durations can result in significant systematic biases. To address this, we formulate a robust Bayesian inference framework that segments the data into blocks, infers each block coherently, and recombines the results using loosely-coherent phase information. By restricting full phase-tracking to much shorter block durations, this approach prevents significant accumulation of phase errors. We implement this methodology in SPLIT (Segmented Posteriors for Long-Inspiral Templates), a GPU-accelerated Python package. Applying SPLIT to an environment-rich injection, we demonstrate that while a fully-coherent vacuum-GR analysis incurs a maximum 1D systematic bias of about 4.8σ from the truth, the shorter integration window of our loosely-coherent approach restricts such biases to 0.5σ. Overall, despite a fractional loss of optimal signal-to-noise ratio, the substantial improvement in parameter accuracy offered by our method presents a highly advantageous trade-off for LISA and other future GW detectors.