New Algorithm Tackles Millions of Overlapping Space Gravitational Waves

A high-performance C++ computing tool successfully processes over 130,000 source parameters to untangle crowded signals for the upcoming LISA observatory.

By Karan Das

Published: November 13, 2026

New Algorithm Tackles Millions of Overlapping Space Gravitational Waves — Norm of the deterministic signal Fourier transform, |d~(f)||\tilde{d}(f)| (top row), and PSD (bottom row) for the AA (left), EE (middle), and TT (right) channels; the PSD panels share the same axes. The blue line is the median obtained with samsara++, while the black line represents the injected value. The dark and light shaded regions mark the 16–84% and 5–95% credible intervals, respectively, and the orange line is the PSD of the injected data. In the top row, the bands are the credible intervals of |d~||\tilde{d}| itself, computed frequency by frequency in each channel. The deterministic component is identically zero outside the frequency windows around each source; these bins are omitted on the logarithmic axis. The insets in the top row show |d~||\tilde{d}| on a linear scale, bin by bin, around the injected frequency f0f_{0} (red dashed line) of two sources. The left inset shows a source at f0=1.916mHzf_{0}=1.916\,\mathrm{mHz} with SNR 8.48.4, whose signal overlaps with those of Lorenzo Valbusa Dall'Armi et al. / arXiv

A new high-performance computing algorithm could help scientists untangle millions of overlapping gravitational wave signals expected to flood the upcoming Laser Interferometer Space Antenna (LISA) space mission. The tool, called samsara++, is designed to simultaneously solve the complex data jigsaw puzzle presented by thousands of compact binary star systems in our Milky Way.

How Will LISA Hear the Milky Way White Dwarf Population?

Set to launch in the 2030s, the Laser Interferometer Space Antenna will operate as a space-based gravitational wave observatory with arms stretching 2.5 million kilometers across interplanetary space. Unlike ground-based detectors such as LIGO, which primarily catch fleeting transient signals from colliding black holes and neutron stars, LISA will face a perpetually crowded sky. The authors report the new findings are available on the arXiv preprint server.

Population synthesis studies indicate that the most numerous sources in LISA’s frequency band will be roughly 10 million double white dwarf binary systems scattered throughout our galaxy. Between 1,000 and 10,000 of these individual pairs will emit continuous gravitational waves loud enough to be resolved directly, while the rest blend together into an incoherent background roar below a few millihertz. Because these signals overlap continuously in both time and frequency, researchers must perform a massive global fit analysis to separate individual sources, stochastic foregrounds, and instrumental noise.

Figure 2: Inference on the single frequency bin [1.1728,1.2324]mHz[1.1728,1.2324]\,\rm mHz. Top left: trace of the number of DWDs as a function of the generation, for PT runs with 8, 16, 32, and 64 injected sources (dashed horizontal lines), starting from an empty state. Top right: posterior on the number of DWDs (grey) for 143 injections (dark grey dashed line), and number of sources in the PETRA catalog with pdetp_{\rm det} greater than 0.1, 0.5, and 0.9 (orange dot-dashed lines, from light to dark). Bottom left: number of sources in 23 equally spaced frequency sub-bins; the blue line and the light blue band are the median and the 16–84% credible interval over the societies explored by samsara++, while the dark grey dashed line is the injected population. Bottom right: posterior predictive of the modulus of the AA-channel data, |d~A(f)||\tilde{d}_{A}(f)|, over the same frequency range as the bottom-left panel. The blue line is the median, the dark and light shaded regions mark the 50% and 90 — Inference on the single frequency bin [1.1728,1.2324]mHz[1.1728,1.2324]\,\rm mHz. Top left: trace of the number of DWDs as a function of the generation, for PT runs with 8, 16, 32, and 64 injected sources (dashed horizontal lines), starting from an empty state. Top right: posterior on the number of DWDs (grey) for 143 injections (dark grey dashed line), and number of sources in the PETRA catalog with pdetp_{\rm det} greater than 0.1, 0.5, and 0.9 (orange dot-dashed lines, from light to dark). Bottom left: number of sources in 23 equally spaced frequency sub-bins; the blue line and the light blue band are the median and the 16–84% credible interval over the societies explored by samsara++, while the dark grey dashed line is the injected population. Bottom right: posterior predictive of the modulus of the AA-channel data, |d~A(f)||\tilde{d}_{A}(f)|, over the same frequency range as the bottom-left panel. The blue line is the median, the dark and light shaded regions mark the 50% and 90 Credit: Lorenzo Valbusa Dall'Armi et al. / arXiv

Why Is Global Fit Inference So Computationally Demanding?

Traditional Bayesian data analysis struggles with trans-dimensional problems where the exact number of sources is unknown and must be treated as an active parameter. The research team, consisting of Lorenzo Valbusa Dall’Armi, Gabriele Astorino, Riccardo Buscicchio, Angelo Ricciardone, and Walter Del Pozzo, built their new sampler to replace older methods that require extensive memory allocations or rely heavily on Python limitations.

Also ReadBlood Moon Eclipse to Light Up UK Skies This Sunday

By writing the framework in C++ and CUDA, the team utilized graphics processing units to execute batched parallel calculations and shared-memory threading. They introduced specialized dynamic array structures that eliminate redundant memory allocations whenever stellar sources are added or removed during Markov chain Monte Carlo iterations. This design cuts memory overhead by three orders of magnitude and reduces the computational complexity of removing a binary source from an operations-heavy step down to a near-instantaneous index swap.

What Did the Stress Tests Reveal About the Algorithm?

The investigators tested their sampler using simulated one-year LISA data streams containing thousands of synthetic double white dwarf binaries alongside stationary Gaussian instrumental noise. In a targeted benchmark involving 16,435 injected binaries, which created a staggering parameter space of 131,480 dimensions at the true population size, the system completed 15 million iterations in approximately 16 hours using only 3 gigabytes of RAM.

Figure 3: Left panel: histogram of the absolute difference between the number of injected sources and the posterior mode of the number of sources in each frequency bin, for all the 1000 bins (blue) and for the 281 bins containing at least one injected source (orange, dashed). Middle panel: trace of the number of sources as a function of the generation for the 19th frequency bin; the dashed line marks the injected number. Right panel: median and 16–84% and 5–95% bands, across the 281 non-empty bins, of the difference between the number of sources and the injected one as a function of the generation. — Left panel: histogram of the absolute difference between the number of injected sources and the posterior mode of the number of sources in each frequency bin, for all the 1000 bins (blue) and for the 281 bins containing at least one injected source (orange, dashed). Middle panel: trace of the number of sources as a function of the generation for the 19th frequency bin; the dashed line marks the injected number. Right panel: median and 16–84% and 5–95% bands, across the 281 non-empty bins, of the difference between the number of sources and the injected one as a function of the generation. Credit: Lorenzo Valbusa Dall'Armi et al. / arXiv

The benchmarking results demonstrate that continuous-time birth-death sampling is computationally feasible for massive galactic populations. The authors emphasize that their algorithm successfully recovered the injected number of sources starting from completely empty initial states without needing prior astrophysical hints. While minor upward offsets appeared in the reconstructed instrument noise spectral density due to incomplete separation of overlapping signals, the overall signal reconstruction matched the simulated data within close credible intervals.

What Steps Come Next for Space-Based Gravitational Wave Astronomy?

The research group aims to expand their sampler to tackle an even wider array of astrophysical phenomena in future simulations. Planned expansions involve incorporating massive black hole binaries, extreme mass ratio inspirals, and the total unresolved galactic foreground into a unified global fitting architecture for upcoming data challenges.

Also ReadWatch asteroid 152637 (1997 NC1) zip past Earth at its closest in 400 years this June

The study was posted to the arXiv preprint server on November 12, 2026.

Fact Checked

This article has been fact checked for accuracy, with information verified against reputable sources. Learn more about us and our editorial process.

Last reviewed on November 13, 2026.

Article history

November 13, 2026 Latest version

Reference(s)

Dall'Armi, Lorenzo. “Global Fit Inference for LISA with samsara++.” arXiv.org <http://arxiv.org/abs/2611.17266v1>.

Cite this page:

Das, Karan. “New Algorithm Tackles Millions of Overlapping Space Gravitational Waves.” BioScience. BioScience ISSN 2521-5760, 13 November 2026. <https://www.bioscience.com.pk/en/subject/astronomy/global-fit-inference-for-lisa-with-samsara>. Das, K. (2026, November 13). “New Algorithm Tackles Millions of Overlapping Space Gravitational Waves.” BioScience. ISSN 2521-5760. Retrieved November 13, 2026 from https://www.bioscience.com.pk/en/subject/astronomy/global-fit-inference-for-lisa-with-samsara Das, Karan. “New Algorithm Tackles Millions of Overlapping Space Gravitational Waves.” BioScience. ISSN 2521-5760. https://www.bioscience.com.pk/en/subject/astronomy/global-fit-inference-for-lisa-with-samsara (accessed November 13, 2026).

Posted by Karan Das

Follow us on social media

Google News Facebook Twitter Instagram LinkedIn Telegram YouTube

End of the article