cWB-space: A time-frequency unmodelled transient-search pipeline for LISA

arXiv:2609.12986 · gr-qc, astro-ph.IM · Submitted 2026-09-11 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: General Relativity and Quantum Cosmology. The day in general relativity and quantum cosmology.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "cWB-space: A time-frequency unmodelled transient-search pipeline for LISA".

Mira: Here's the gist of the paper. Gravitational-wave observations can reveal sources whose signals are not known in advance The central physical difficulty is that a feature in the data may come either from a gravitational wave or from a disturbance within the instrument.

Kai: This work presents cWB-space, which identifies candidate signals by following how their power is distributed in time and frequency, and then reconstructs parts of their waveforms without imposing a source model.

Mira: Pipeline Overview cWB-space is organized as a sequence of modular functions that includes a response stage, an instrument stage, a TDI-construction stage, and a search stage.

Kai: The simulator starts by generating GW source time series and propagating them through the LISA response into the six one-way links.

Title and authors: Kai: So, let’s talk about the title of this paper, cWB-space: A time-frequency unmodelled transient-search pipeline for LISA. It immediately tells you what this is about—it's a pipeline for searching for gravitational wave transients that doesn't require a predicted waveform.

Mira: I agree. The authors are Shubhanshu Tiwari, Yumeng Xu, and Giovanni A Prodi from the University of Trento, along with collaborators from ETH Zurich and IAC in Spain. It’s a team pulling together expertise from different physics areas to tackle this specific LISA challenge.

Lev: From my side, it’s interesting that they are focusing on LISA specifically. The challenge there is dealing with the overlapping astrophysical sources and the instrumental disturbances simultaneously, which makes any search strategy really tough.

Kai: It seems like they’re addressing that head-on by focusing on how power is distributed across time and frequency instead of relying solely on a predicted waveform to look for signals. That's a big shift in approach.

Mira: They are essentially using the structure of the data itself—the time-frequency distribution—as the primary search tool, rather than trying to match it directly against a known waveform family.

Lev: And from my perspective as someone building real hardware, this means they’re looking for statistical signatures in the measured strain that could point toward a signal even if we don't have a perfect model of what that signal should look like. It’s about finding anomalies in the noise structure.

Kai: So it’s less about matching shapes and more about finding patterns in how energy moves through the detector over time and frequency. It’s an unmodelled approach, which is what makes it so appealing for discovering novel signals.

The paper's summary: Mira: Looking at the summary, cWB-space is organized as a sequence of modular functions that includes a response stage, an instrument stage, a TDI-construction stage, and finally a search stage. It’s not one big black box; it’s built in stages.

Kai: That modular structure is smart because it allows you to test each piece individually. They start by generating gravitational wave source time series and running them through the LISA response into those six one-way links.

Lev: And then they generate realistic noise by first evolving the LISA Instrument simulation for a duration of zero point one second, and then evaluating those TDI observables at that same cadence to get the realistic noisy links.

Mira: The core idea is that this pipeline groups concentrations of power in two combined LISA data streams and ranks them using three things: their strength, their relative strength in the two streams, and their frequency evolution.

Kai: And they use this ranking statistic called rho event which we talked about earlier, which combines the total power in a cluster with its balance between A and E plus the evolution of its centroid frequency. That gives them a way to prioritize candidates based on physical characteristics like chirp-like tracks.

Lev: It’s interesting how they then use the WDM transform to compute a coefficient map where each pixel is a time and frequency pair, and then they do event clustering based on that map.

Mira: So the whole process flows from simulating sources and noise, through whitening the WDM map, clustering bright pixels into candidate events, ranking them by rho event, and then using glitch classification to filter out instrumental disturbances.

Kai: And it’s important to note that this entire pipeline is designed to work on both simulated data branches and real-data branches. This means you can test how it handles the actual noisy links versus perfectly simulated ones.

The paper's improvements: Lev: What they suggest as improvements is focused on making the identification of candidate signals more robust by integrating detector response comparisons. They want to evaluate where a similar response can arise from an instrumental origin.

Kai: That directly improves instrumental disturbance discrimination. It means the method isn't just looking at one observable; it’s checking how a feature affects different TDI observables to see if that pattern is consistent with a GW or if it looks like an instrument glitch.

Mira: They also improve the ranking statistics by defining rho event more precisely to explicitly incorporate the total power in a cluster and its balance between A and E plus the evolution of its centroid frequency.

Lev: That refinement should give us a better way to prioritize signals based on those physical characteristics we discussed, like tracking chirp-like tracks, without relying solely on the original definition.

Kai: And they add an event-level consistency test using Sagnac and zeta TDI combinations to evaluate how different link patterns and strengths affect the response statistic. That’s a new way to check for instrumental transients when A/E alone isn't sufficient.

Conclusion: Mira: So, wrapping up the paper cWB-space, they provide template free transient search, waveform reconstruction without imposing a source model, and additional information carried by the LISA detector response via glitch classification.

Kai: It’s a really useful combination of tools here. They can find things they don't know about before they have a good model for them, and then reconstruct some waveform info afterwards.

Lev: For me, the main implication is that this method gives us a way to investigate signals whose waveforms are uncertain or haven't been anticipated yet, which could help us discover unexpected dynamics or physical processes outside of established source models.

Kai: That leads into how we might use it in global analysis. This capability also motivates a role within the global analysis of LISA data because it allows for this kind of discovery without needing to assign those waveforms to a particular astrophysical population right away.

Mira: It’s about following the power distribution, not fitting a model, and that’s where it opens up new avenues for exploring what might be happening in space. We're still waiting to see what kinds of interesting things this pipeline can actually find when run on the real data sets.

Lev: I just think this framework is solid for testing how we can use these complex time-frequency maps to pull out physical information from noisy detector output. It’s a strong step toward separating signal from noise in a way that respects the physics of the measurement itself.

Shubhanshu Tiwari, Yumeng Xu, Giovanni A Prodi

Department of Earth and Planetary Sciences, ETH Zurich · Departament de Fósica, Universitat de les Illes Balears · Department of Physics, University of Trento and INFN-TIFPA

gr-qc, astro-ph.IM

Submitted: 2026-09-11

Updated: 2026-10-30

Comments: 19 Pages, 10 figures,

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

Importance score: 90/100

The gist: Gravitational-wave observations can reveal sources whose signals are not known in advance The central physical difficulty is that a feature in the data may come either from a gravitational wave or…

Key concepts

WDM Transform
This transform creates a map showing how signal power is distributed across both time and frequency simultaneously. Each spot on this map represents a specific time and frequency pair, allowing the system to visualize the overall structure of the data.
Detection Statistic (rhoevent)
This formula ranks candidate signals by combining several factors: their total power, how balanced they are between two data streams (A and E), and how their central frequency changes over time. This helps prioritize potential events based on physical characteristics like a chirp.
Template-free Search
Instead of looking for signals matching a known waveform, this method searches for any pattern of power distribution that looks like a transient event. This is crucial for finding unexpected or novel gravitational wave signals whose exact shape is not yet predicted by current models.
TDI Observables
These are specific data products derived from the raw LISA detector measurements. The pipeline uses these observables to compare how power concentrates in different channels, which helps distinguish between a true gravitational wave and noise originating from the instrument itself.

Terminology

Summary

Gravitational-wave observations can reveal sources whose signals are not known in advance The central physical difficulty is that a feature in the data may come either from a gravitational wave or from a disturbance within the instrument. p. 0

Where the paper says it

A transient search could identify intervals deserving attention before a complete description of all those sources is available. The central physical difficulty is that a feature in the data may come either from a gravitational wave or from a disturbance within the instrument.

This work presents cWB-space, which identifies candidate signals by following how their power is distributed in time and frequency, and then reconstructs parts of their waveforms without imposing a source model. p. 0
Where the paper says it

For the Laser Interferometer Space Antenna (LISA), the challenge is to find such signals among overlapping astrophysical sources and disturbances in the instrument. We present cWB-space, which identifies candidate signals by following how their power is distributed in time and frequency, and then reconstructs parts of their waveforms without imposing a source model.

Pipeline Overview

cWB-space is organized as a sequence of modular functions that includes a response stage, an instrument stage, a TDI-construction stage, and a search stage. The simulator starts by generating GW source time series and propagating them through the LISA response into the six one-way links. p. 1

Where the paper says it

In case of real LISA data the downstream pipeline from common links is carried forward. Simulator The simulator starts by generating GW source time series and propagating them through the LISA response into the six one-way links.

The realistic noise is generated by first evolving the LISA Instrument simulation at ∆t = 0. p. 1
Where the paper says it

The response uses LISA Orbits and LISA GW Response, and the noisy-link simulation uses LISA Instrument [19–22]. The realistic-noise is generated by first evolving the LISA Instrument simulation at ∆t = 0.

1 s and evaluating the TDI observables at the same cadence. p. 1
Where the paper says it

The realistic-noise is generated by first evolving the LISA Instrument simulation at ∆t = 0. 1 s and evaluating the TDI observables at the same cadence.

The pipeline groups concentrations of power in two combined LISA data streams and ranks them using their strength, relative strength in the two streams, and frequency evolution. p. 1
Where the paper says it

We present a pipeline connecting candidate finding, detector-response comparisons, and waveform reconstruction. It groups concentrations of power in two combined LISA data streams and ranks them using their strength, relative strength in the two streams, and frequency evolution.

Time-Frequency Representation

The WDM transform computes a coefficient map where a time-frequency pixel denotes one indexed pair (ti, fj) on the map. The default cWB-space whitening is the self-A/E median normalization, using the analyzed data segment itself to estimate this background. p. 3

Where the paper says it

Whitening We divide channel power by a frequency-dependent empirical noise-floor to compare excesses within the event map. The default cWB-space whitening is the self-A/E median normalization, using the analyzed data segment itself to estimate this background.

The event-clustering stage groups bright pixels in the whitened WDM map into candidate events based on a chosen time window and frequency band. p. 4
Where the paper says it

Bright-pixel event clustering The clustering stage groups bright pixels in the whitened WDM map into candidate events. It operates on the combined search power PAE in a chosen time window and frequency band.

Detection Statistic

The ranking statistic is defined as ρevent = p EAE p BAE 1 + mchirp squared. This statistic combines the total power in a cluster, its balance between A and E, and the evolution of its centroid frequency. p. 4

Where the paper says it

The choices made here for the sparseness and the connectivity of the pixels are based on empirical tests done for a small set of MBHB injections, and can be improved and targeted to other signal classes. Pipeline detection statistic The ranking combines the total power in a cluster, its balance between A and E, and the evolution of its centroid frequency.

The chirp term is a preference in the ranking, not a strict requirement on candidate morphology. p. 4
Where the paper says it

(14) The three factors are basically the weighting the normalized power with channel balance, and approximately non-decreasing centroid frequency. The chirp term is a preference in the ranking, not a strict requirement on candidate morphology: its multiplicative factor lies between 1/2 and 1, so a nonchirping event is not rejected solely for lacking an increasing frequency.

Validation Study

The validation study addresses three connected questions: which candidates are found, what their channel responses reveal, and what waveform information can be recovered. The MBHB-plus-glitch validation uses a continuously generated two-day instrumental-noise data and prescribed sources to test candidate finding and TDI response comparisons. p. 6

Where the paper says it

OpenAI Codex (GPT-6 Astra) assisted in the development and check of scripts for some of the validation, benchmarking, and plotting studies under the authors directions. MBHB-plus-glitch validation This study use a continuously generated two-day instrumental-noise data, sampled initially at 0.

The Sangria simulation tests the association of the leading candidates with the unblinded content of MBHB. p. 6
Where the paper says it

The MBHB and generic-burst data examine candidate finding and TDI response comparisons for prescribed gravitational-wave and link-glitch injections. The Sangria analysis tests the association of the leading candidates with the unblinded content of MBHB.

Waveform Recovery

The reconstruction studies assess waveform extraction separately with different settings from candidate identification. For example, the reconstruction-to-source norm ratios are within approximately 1. p. 9

Where the paper says it

1151, and 0. 1309; the reconstruction-to-source norm ratios are 1.

3% of the reference values, indicating close agreement in overall amplitude. p. 3
Where the paper says it

The reference search uses K = 32, while the reconstruction examples use K = 256. These are empirical working configurations for the examples studied here, not the outcome of a systematic optimization of K or uniquely preferred values.

The results demonstrate accurate local waveform recovery for all six events. p. 7
Where the paper says it

VI F; Figs. No instrument-monitor comparison; reference combines all black-hole signals; local recovery is not an isolatedsource full-waveform test.

Conclusions

cWB-space provides template-free transient search, waveform reconstruction, and additional information carried by the LISA detector response via glitch classification. The analysis conducted in this paper follows gravitational-wave and link-glitch injections through the LISA response to second-generation TDI observables in noisy LISA data. p. 2

Where the paper says it

3 Simulator input strain waveform  or generic burst phenomxpy LISA response six GW links LISA Orbits GW Response Instrument model realistic noisy links LISA Instrument Common link data (12, 23, 31, 13, 32, 31) Real LISA input Calibrated one-way link data TDI observables X, Y, Z; A, E, T; α, β, γ, ζ PyTDI A/E WDM maps self-median whitening wdm-wavelet Bright-pixel event clustering Sagnac/zeta WDM positive-excess ratio Event catalog detection statistic and glitch classifier wdm-wavelet FIG. Schematic overview of the cWB-space analysis chain.

The generic-burst examples illustrate the breadth of the search without assigning those waveforms to a particular astrophysical population. p. 0
Where the paper says it

For the Laser Interferometer Space Antenna (LISA), the challenge is to find such signals among overlapping astrophysical sources and disturbances in the instrument. We present cWB-space, which identifies candidate signals by following how their power is distributed in time and frequency, and then reconstructs parts of their waveforms without imposing a source model.

This capability also motivates a role within the global analysis of LISA data. p. 0
Where the paper says it

Finding a feature and testing its physical origin are therefore connected parts of the analysis. This is also relevant to LISA’s global fit: the effort to disentangle many sources and the noise together.

Acknowledgments

The authors thank Johan Robertsson for fruitful discussions and continued support. The analysis uses PyTDI, LISA Orbits, LISA GW Response, LISA Instrument, wdm-wavelet, phenomxpy, and lisaglitch. p. 2

Where the paper says it

3 Simulator input strain waveform  or generic burst phenomxpy LISA response six GW links LISA Orbits GW Response Instrument model realistic noisy links LISA Instrument Common link data (12, 23, 31, 13, 32, 31) Real LISA input Calibrated one-way link data TDI observables X, Y, Z; A, E, T; α, β, γ, ζ PyTDI A/E WDM maps self-median whitening wdm-wavelet Bright-pixel event clustering Sagnac/zeta WDM positive-excess ratio Event catalog detection statistic and glitch classifier wdm-wavelet FIG. Schematic overview of the cWB-space analysis chain.

OpenAI Codex (GPT-6 Astra) assisted with the literature and citation checks under the authors’ direction. p. 6
Where the paper says it

Reconstruction studies assess waveform extraction separately with different settings from candidate identification. OpenAI Codex (GPT-6 Astra) assisted in the development and check of scripts for some of the validation, benchmarking, and plotting studies under the authors directions.

The data sets are available from the LISA Data Challenge release. p. 2
Where the paper says it

data sets are generated in overlapped chunks and stitched only after this removal, filtering, and downsampling procedure has been applied. The noise and SNR conventions are checked against the standard LISA sensitivity normalization of Babak et al.

The algorithm and supporting data for the present study will be made publicly available. p. 0
Where the paper says it

LISA will observe the millihertz gravitational-wave sky, with many sources present at the same time [4]. A transient search could identify intervals deserving attention before a complete description of all those sources is available.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Automated transient candidate identification based on time-frequency power distribution enables systems to identify candidate signals by following how their power is distributed in time and frequency. This allows AI to search for unexpected signals without needing a predicted waveform from a specific source class.

  2. Automated waveform reconstruction capability allows systems to reconstruct parts of their waveforms without imposing a source model after identifying candidate transients. This enables the system to generate reconstructed channel responses by using the inverse transforms provided by wdm-wavelet.

  3. Instrumental disturbance discrimination is improved by integrating detector response comparisons, as the method can evaluate where a similar response can arise from an instrumental origin. This allows AI to distinguish between gravitational wave signals and disturbances based on how they affect different TDI observables.

  4. Improved ranking statistics allow systems to rank candidates using the total power in a cluster, its balance between A and E, and the evolution of its centroid frequency by calculating the statistic rhoevent. This enables better prioritization of candidates based on physical characteristics like chirp-like tracks.

  5. Automated classification of instrumental glitches is enhanced by adding an event-level consistency test using Sagnac and zeta TDI combinations to evaluate how different link patterns and strengths affect the response statistic. This helps in distinguishing instrumental transients from gravitational wave signals when the A/E statistic alone is insufficient.

  6. Template-free search capability allows systems to investigate signals whose waveforms are uncertain or have not yet been anticipated, which can help identify unexpected dynamics or physical processes outside established source models. This supports the discovery of novel signal morphologies and helps guide iterative source modeling in global fits.

Abstract

The Laser Interferometer Space Antenna (LISA) will observe massive-black-hole mergers and other transient phenomena alongside many long-lived gravitational-wave signals. Some transients may be incompletely modeled or have waveforms that have not been anticipated. We present cWB-space, a pipeline that combines an unmodeled transient search, waveform reconstruction, and a time-frequency glitch classifier. The search identifies concentrations of signal power in time and frequency and reconstructs candidate waveforms without requiring a source template. The glitch classifier compares how candidates appear in different combinations of LISA's laser measurements, with the background estimated from the observed data. We test the pipeline with simulated black-hole binaries with aligned and precessing spins, burst signals with different shapes, and instrumental glitches. The response comparisons help assess a candidate's astrophysical or instrumental origin, while also revealing correlated glitches that can resemble gravitational waves. In the Sangria simulated LISA data, the search recovers all six injected massive-black-hole binaries as its six highest-ranked candidates. Their local reconstructions recover the main features of the injected waveforms over the intervals studied. These results demonstrate how an unmodeled search can connect transient identification with investigation of a candidate's origin and waveform recovery.

Sources

Related papers