samsara++ LISA Algorithm Explained: New Sampler Untangles Millions of Overlapping Gravitational Waves

Artist's impression of ESA's LISA mission. Image: European Space Agency, CC BY-SA 3.0 IGO (via Wikimedia Commons)

TL;DR: The samsara++ LISA algorithm is a new C++/CUDA sampler that untangles millions of overlapping gravitational waves for ESA's LISA mission. Posted to arXiv on 12 November 2026, it rebuilt the global-fit analysis from empty starting models and, in a stress test, ran 15 million iterations in about 16 hours on just 3 GB of RAM.

Key Highlights

• New paper: "Global Fit Inference for LISA with samsara++" (arXiv:2611.17266), submitted 12 November 2026.
• The problem: LISA will hear ~10 million double white dwarf binaries at once, overlapping in time and frequency.
• Headline number: 16,435 binaries (131,480 parameters), 15 million iterations in ~16 hours, 3 GB of RAM.
• It starts from nothing: the smaller tests began with empty source models — no preliminary search, no hints about the injected population.
• Status: a preprint on simulated data, not yet peer-reviewed.

What is the samsara++ LISA algorithm?

samsara++ is a sampler: software that tests possible explanations of a dataset to find the best fit. Written in C++ with CUDA for parallel GPU computing, it builds on an earlier framework called samsara, rebuilt for the hardest version of the LISA problem: the global fit.

The Laser Interferometer Space Antenna (LISA) is ESA's planned space-based gravitational wave observatory: three spacecraft in a triangle with arms 2.5 million kilometres long, listening in the millihertz band LIGO cannot reach. The Milky Way holds roughly 10 million double white dwarf binaries radiating in LISA's band, and between 1,000 and 10,000 will be loud enough to pick out individually. The rest merge into a background hum, so researchers must perform a "global fit" — modelling every source and the instrument noise at once.

samsara++ combines birth-death processes, which add a source to the model or remove one, with mutation moves that refine each source's parameters; the birth and death rates are driven by the data itself. Shared-memory parallelism, incremental likelihood updates and indexed storage keep costs down; parallel tempering helps it explore multi-peaked posteriors. The team reports the design cuts memory overhead by three orders of magnitude over older approaches.

The stress test in numbers

Tests used simulated one-year LISA data streams of double white dwarf binaries plus Gaussian instrument noise:

Test	Binaries	Parameters	Result
Small population	143	~1,144	Recovered from an empty start
Medium population	866	~6,928	Recovered from an empty start
Stress test	16,435	131,480	15 million iterations in ~16 hours, ~3 GB RAM

The empty start matters: the sampler began with zero sources and had to discover how many binaries were present, purely from the data. It recovered the injected number of sources without prior hints. Caveat: the reconstructed noise showed small upward offsets where overlapping signals were not fully separated.

Artist's conception of J0806, a double white dwarf system about 1,600 light-years away whose stars orbit each other every 321 seconds. Image: NASA / Dana Berry (CXC), public domain (via Wikimedia Commons)

Why the LISA global fit is so hard

Ordinary data analysis assumes you know how many things you are looking for. LISA breaks that assumption: the number of resolvable sources is itself unknown, so it must be treated as a parameter — a "trans-dimensional" problem. Every signal sits on top of every other in time and frequency, the noise must be estimated at the same time, and the parameter space explodes: the stress test alone had 131,480 parameters. Older methods demand heavy memory allocations when sources change, or lean on Python, which limits sampling speed. samsara++ keeps a 15-million-iteration run inside an overnight job on modest memory — the paper's central claim is that birth-death sampling is computationally feasible at galactic scale.

Should you care?

✅ Verdict: If you follow space science, yes — this is the infrastructure work that decides whether a flagship mission delivers science or drowns in its own data. LISA is a decade from launch, so nothing here changes what you can observe today; but every global-fit method proven now is one less risk for the 2030s. Treat it as a preprint result on simulated data, not a finished capability.

What happens next

Next, the team will extend samsara++ beyond white dwarfs to massive black hole binaries, extreme mass ratio inspirals and the unresolved galactic foreground. Others are attacking the same problem differently: an earlier space gravitational-wave algorithm cut phase error by 53% in simulations, and PETRA is being built to turn global-fit results into catalogues. The real test is years away: real LISA data, with gaps, glitches and non-stationary noise, will be far messier than these simulated streams.

FAQs

What is the LISA global fit problem?
LISA will hear millions of overlapping signals at once, so sources cannot be analysed one by one. The global fit models every source and the noise simultaneously, including the unknown number of sources — far harder than fitting a single event.

What is samsara++?
A C++/CUDA sampler for the LISA global fit, posted to arXiv on 12 November 2026. It uses continuous-time birth-death sampling — adding and removing sources as the data demands — with parallel tempering and GPU parallelism.

How many gravitational wave sources will LISA detect?
About 10 million double white dwarfs radiate in LISA's band; 1,000–10,000 should be resolvable individually. LISA will also hear massive black hole mergers and extreme mass ratio inspirals.

When will LISA launch?
LISA is an ESA-led mission adopted in 2024, with launch planned for around 2035. Three spacecraft will fly in a triangle with 2.5-million-kilometre arms, detecting millihertz gravitational waves ground-based observatories cannot reach.

Takeaway

samsara++ shows the hardest part of LISA's data analysis — fitting millions of overlapping signals at once — can run on ordinary computing resources. The mission is a decade away, but its software is being proven now.

Read the paper on arXiv, and see our AIOA explainer on the AI agent that optimised numerical algorithms 22x faster.

Sources

Global Fit Inference for LISA with samsara++ — arXiv:2611.17266 (Valbusa Dall'Armi et al., submitted 12 Nov 2026)
New Algorithm Tackles Millions of Overlapping Space Gravitational Waves — BioScience, 13 Nov 2026
NASA Scientists Show How Gravitational Waves Can Reveal New Worlds — NASA Science
PETRA: From the LISA global fit to a catalog of Galactic binaries — arXiv:2502.14818
samsara: A Continuous-Time Markov Chain Monte Carlo Sampler for Trans-Dimensional Bayesian Analysis — arXiv:2511.07385