Global Fit Inference for LISA with samsara++
summary
The gist
The paper presents samsara++, a C++/CUDA sampler designed to address the trans-dimensional Global Fit (GF) problem for LISA by combining birth-death processes with mutation moves within a…
In short
samsara++ is a C++/CUDA sampler designed for LISA data to solve trans-dimensional Global Fit problems. It uses birth-death processes within a continuous-time Markov chain to jointly infer unknown source counts, parameters, and instrumental noise. This method offers an efficient way to explore complex parameter spaces by proposing changes in the number of sources. The results show successful recovery of injected source numbers even without prior knowledge of the count.
Key concepts
- Global Fit (GF)
- A technique needed for LISA data where signals from many different gravitational wave sources overlap in both time and frequency. GF aims to jointly model these overlapping signals, including resolvable sources, stochastic foregrounds, and instrumental noise.
- Birth-Death Process (BDP)
- This is a mathematical process used in samsara++ to propose new states in the parameter space. It models the evolution of the number of sources by proposing 'birth' (adding a source), 'death' (removing a source), or 'mutation' (changing parameters) moves.
- Parallel Tempering (PT)
- A method used within samsara++ to improve exploration of complex parameter spaces. It runs multiple MCMC chains at different temperatures simultaneously, allowing the algorithm to jump out of local optima and explore more probable configurations effectively.
Terminology used across episodes
This episode discusses
- Global Fit Inference for LISA with samsara++ · Paper Radio
- Observation of Gravitational Waves from a Binary Black Hole Merger
- GWTC-1: A Gravitational-Wave Transient Catalog of Compact Binary Mergers Observed by LIGO and Virgo during the First and Second Observing Runs
- Population Properties of Compact Objects from the Second LIGO-Virgo Gravitational-Wave Transient Catalog
- GWTC-3: Compact Binary Coalescences Observed by LIGO and Virgo During the Second Part of the Third Observing Run
- GWTC-4.0: Updating the Gravitational-Wave Transient Catalog with Observations from the First Part of the Fourth LIGO-Virgo-KAGRA Observing Run
- The NANOGrav 15-year Data Set: Evidence for a Gravitational-Wave Background
- The second data release from the European Pulsar Timing Array III. Search for gravitational wave signals
- The Parkes Pulsar Timing Array Third Data Release
- Searching for the nano-Hertz stochastic gravitational wave background with the Chinese Pulsar Timing Array Data Release I
- Laser Interferometer Space Antenna
- LISA Definition Study Report
- Astrophysics with the Laser Interferometer Space Antenna
- Cosmology with the Laser Interferometer Space Antenna
- New Horizons for Fundamental Physics with LISA
- Report on the second Mock LISA Data Challenge
- The Mock LISA Data Challenges: from Challenge 3 to Challenge 4
- A LISA Data-Analysis Primer
- Science with the space-based interferometer eLISA. I: Supermassive black hole binaries
- Science with the space-based interferometer LISA. V: Extreme mass-ratio inspirals
- The promise of multi-band gravitational wave astronomy
The paper
Global Fit Inference for LISA with samsara++ · Read on arXiv
Dipartimento di Fisica “E. Fermi”, University of Pisa · INFN, Sezione di Pisa · Gran Sasso Science Institute (GSSI) · INFN, Laboratori Nazionali del Gran Sasso · Dipartimento di Fisica “G. Occhialini”, University of Milano-Bicocca · Institute for Gravitational Wave Astronomy & School of Physics and Astronomy, University of Birmingham
Transcript
Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.
Vera: I'm Vera.
Jocelyn: And I'm Jocelyn, here with Subrahmanyan.
Vera: We're looking at the paper: "Global Fit Inference for LISA with samsara++".
Jocelyn: Here's the gist of the paper. The paper presents samsara++, a C++/CUDA sampler designed to address the trans-dimensional Global Fit (GF) problem for LISA by combining birth-death processes with mutation moves within a continuous-time Markov chain framework This algorithm is crucial because it allows for the joint inference of an unknown number of sources, parameters, and instrumental noise in LISA data, which overlap in both time and frequency.
Vera: The core methodology The approach utilizes a birth-death process (BDP) within a continuous-time (CT) MCMC framework to explore the trans-dimensional parameter space by proposing new states with a larger (birth), smaller (death), or unchanged (mutation) number of sources.
Jocelyn: The transition rates are chosen to satisfy detailed balance, analogous to the Metropolis–Hastings acceptance ratio.
Vera: The sampler is designed to address the challenge of combining high dimensionality with uncertainty in the number of sources, which requires methods capable of exploring models of different dimensions.
Title and authors: Vera: So, to wrap up what we just talked about, this paper introduces samsara++, which is this new C++ CUDA sampler for LISA data that tackles that whole global fit problem—getting the number of sources, their parameters, and the noise all at once.
Jocelyn: Yeah, basically it uses a continuous-time Markov chain with birth-death processes to propose different numbers of sources, like adding one more or removing one less. It's a way to explore that huge parameter space by following a path that’s driven by the posterior distribution itself.
Subrahmanyan: From my side, what I find interesting is how they handle the high dimensionality; they use specific optimizations, like indexed storage in C++ and CUDA to save memory when things are changing. That’s a big engineering win for running this kind of thing on real hardware.
Vera: And not just that, they’re using parallel tempering to help the sampler explore different regions of the result space better. It’s designed to help the chain escape getting stuck in local modes where there might be too many redundant sources.
Jocelyn: Right, and they tested it on mock data and showed it can recover the number of sources injected without needing any prior knowledge about how many there should be. That’s a strong result for figuring out if we’re missing something in our source counts.
Subrahmanyan: It means this method could be the foundation for doing a proper global fit, which is what LISA really needs to do—jointly modeling everything from Galactic binaries to extreme mass ratio inspirals.
Vera: So, they're moving away from the old methods like RJ-MCMC by following a trajectory through all those dimensions simultaneously. It’s about making the inference process itself smarter and more efficient than just running separate fits for each thing.
Jocelyn: And I think what this shows is that we can tackle problems where the model structure itself, like the number of sources, is unknown, instead of having to guess it ahead of time.
Subrahmanyan: Exactly. This opens up possibilities for combining different types of signals—like foregrounds and astrophysical sources—into one coherent picture instead of treating them separately.
Vera: It’s a powerful tool for testing how well we can actually fit the data when the complexity is this high, which is exactly what LISA will be dealing with.
Jocelyn: We need to see if this efficiency holds up when we start feeding it all that real noise and all those different source populations into the model.
The paper's summary: Vera: So, to recap, samsara++ is this new sampler for LISA data that uses birth-death processes to handle the problem of figuring out unknown source numbers and noise all together in a global fit analysis.
Jocelyn: Yeah, it’s using that continuous-time Markov chain idea to propose different numbers of sources at each step, like birth or death moves. It’s designed to explore those complicated parameter spaces without having to guess the number of sources first.
Subrahmanyan: From a theoretical side, what I like is how they model the signals—they use mildly chirping sinusoids for Galactic binaries and account for Doppler modulation in the waveform parameters like frequency and amplitude. That’s a physically motivated way to build the signal model.
Vera: And they’re using heterodyning to speed up those time-domain calculations, which cuts down on the number of operations needed to calculate likelihoods by a factor related to the frequency points. That makes it computationally feasible for real data sets.
Jocelyn: Right, and they’ve also made some big computational improvements, like using C++ CUDA for parallel computation and an indexed storage scheme so memory doesn't get wasted when the population sizes change. That’s crucial for handling high-dimensional problems.
Subrahmanyan: I’m interested in their proposal to use a trans-dimensional parallel tempering scheme to improve mixing. If the sampler can explore different temperature chains well, it means we have a better chance of finding the true posterior mode without getting stuck in weird local configurations.
Vera: So they aren't just presenting one method, they are suggesting ways to make this sampler even more robust and efficient by adding these specific computational tricks and sampling techniques.
Jocelyn: And there’s the idea that they could extend this further by using parallel evolution for different species of sources. If you have many non-overlapping frequency bands, you can run those populations in parallel, which is a great speedup.
Subrahmanyan: That sounds promising for scaling up the model; if we can evolve those different source groups simultaneously, it gets closer to modeling the full complexity of LISA’s expected input.
Vera: So what this suggests is that we aren't just looking at one fix, but a whole pipeline of improvements—from the core sampler design to how they handle the signal modeling and how they run the code efficiently.
Jocelyn: And it points toward a future where we can apply this framework not just to DWDs, but to include all those other sources we’ve been discussing, like MBHBHs and EMRIs, in that single global fit.
The paper's improvements: Vera: So we've covered how samsara++ uses birth-death processes to do that trans-dimensional global fit for LISA data, figuring out source counts, parameters, and noise all at once.
Jocelyn: Yeah, it shows a way to explore those huge parameter spaces by proposing moves that add or subtract sources from the current model state. It's about following the posterior distribution without knowing exactly how many sources there are beforehand.
Subrahmanyan: From my perspective, the modeling of those signals as mildly chirping sinusoids is pretty standard, but I appreciate how they’ve tied their spectral width to things like observation time and Doppler modulation. That keeps the physics grounded in what we actually see.
Vera: And that whole process is made feasible because of the computational efficiency gains they made, using parallelization and those clever indexing schemes to keep memory management clean as the chain runs.
Jocelyn: I think it really opens up possibilities for combining all these different signal types—the Galactic binaries, the foregrounds, and the exotic sources like MBHBHs—into one coherent analysis. It moves us closer to a full global fit model.
Subrahmanyan: If this method can scale up to handle all those source species without losing its efficiency, it becomes a real tool for testing cosmological models where we need that level of joint inference.
Vera: It’s a solid piece of work because they've shown that the sampler can recover the number of injections even when starting from an empty state, which is exactly what we need to test our analysis pipelines.
Jocelyn: And it gives us a clear path forward for LISA data analysis, showing how to handle this kind of complexity in a practical way.
Subrahmanyan: Ultimately, the success of this Global Fit Inference for LISA with samsara++ really depends on whether we can successfully extend it to include all those different source populations. That’s the next big challenge for structure formation studies.
Conclusion: Vera: So we're wrapping up on "Global Fit Inference for LISA with samsara++," which is this new C++ CUDA sampler that tackles figuring out source counts, parameters, and noise all at once for LISA data.
Jocelyn: Yeah, it shows a way to explore those huge parameter spaces by proposing moves that add or subtract sources from the current model state. It's about following the posterior distribution without knowing exactly how many sources there are beforehand.
Subrahmanyan: From my side, I think what they did well is how they handle the high dimensionality; using indexed storage in C++ and CUDA to save memory when population sizes change is a big engineering win for running this kind of thing on real hardware.
Vera: And that parallel tempering setup they introduced really helps them sample better by letting the algorithm explore more probable configurations without getting stuck in weird local spots.
Jocelyn: Right, and the results with those mock analyses showed that the sampler can recover injected source numbers even when starting from an empty state, which is a strong result for figuring out if we're missing something in our source counts.
Subrahmanyan: It means this method could be the foundation for doing a proper global fit, which is what LISA really needs to do—jointly modeling everything from Galactic binaries to extreme mass ratio inspirals.
Vera: They’re moving away from those older RJ-MCMC methods by following a trajectory through all those dimensions at once, which makes the inference process itself much more efficient.
Jocelyn: And I think what this shows is that we can tackle problems where the model structure, like the number of sources, is unknown, instead of having to guess it ahead of time.
Subrahmanyan: Exactly. This opens up possibilities for combining different signal types—like foregrounds and astrophysical sources—into one coherent picture instead of treating them separately.
Vera: It’s a solid piece of work because they've shown that the sampler can recover the number of injections even when starting from an empty state, which is exactly what we need to test our analysis pipelines.
Jocelyn: And it gives us a clear path forward for LISA data analysis, showing how to handle this kind of complexity in a practical way.
Subrahmanyan: Ultimately, the success of "Global Fit Inference for LISA with samsara++" really depends on whether we can successfully extend it to include all those different source populations. That’s the next big challenge for structure formation studies.
More episodes
- 2607.17929-gevolution 2.0: GPU-accelerated relativistic N-body simulations for cosmology
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2509.02657-On the synergetic use of Ariel and JWST for exoplanet atmospheric science
- 2609.10042-Modelling Fourier filtering wavefront sensors for PSD-based methods: the AOPERA tool
- 2511.05639-Performance of the AstroPix Prototype Module for the Barrel Imaging Calorimeter at the ePIC Detector and in Space-Based Payloads
- 2610.19909-RPCA: a Resolved PSF reconstruction method for undersampled images
- 2512.03440-An Analysis of LIGO Glitches Using t-SNE During the First Part of the Fourth LIGO-Virgo-KAGRA Observing Run
- 2608.10284-Lunar Reflective Interferometry
- 2610.16633-Institutional broadening of the Japanese astronomy and astrophysics research community during 1996-2025