Simulation-Based Inference on EMRI Frequency Tracks: Harmonic Identification and Parameter Constraints from Search Outputs
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: We're looking at the paper: "Simulation-Based Inference on EMRI Frequency Tracks".
Mira: Here's the gist of the paper. Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long, multiharmonic waveforms make their search and parameter estimation computationally expensive and challenging with standard methods Context.
Kai: Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long, multiharmonic waveforms make their search and parameter estimation computationally expensive and challenging with standard methods.
Mira: A promising alternative is to work directly with the output of a time–frequency search pipeline where one or more recovered frequency tracks are identified.
Kai: Aims of the Investigation The investigation aims to determine whether frequency-evolution tracks recovered by a semi-coherent search pipeline can be used to (i) identify which harmonic(s) a recovered track corresponds to; and (ii) constrain the intrinsic EMRI parameters, including an assessment of how these constraints degrade when the tracks are partially masked near plunge.
Title and authors: Kai: So we’re looking at this paper today called "Simulation-Based Inference on EMRI Frequency Tracks: Harmonic Identification and Parameter Constraints from Search Outputs". It deals with how to analyze those long, complicated signals from Extreme Mass Ratio Inspirals, or EMRIs.
Mira: Exactly. These EMRIs are the main targets for LISA, but their waveforms are so long and have so many different frequencies that standard search methods get really slow and messy when trying to figure out what’s going on.
Lev: And the paper suggests a way around that by looking directly at the output of a time-frequency search pipeline, instead of just analyzing the full waveform data itself.
Kai: Right, so they’re taking those recovered frequency tracks—which are essentially these evolving frequencies over time—and trying to use them to do two main things: first, figure out which specific harmonic that track belongs to; and second, constrain the actual physical parameters of the inspiral.
Mira: That makes sense. They want to know if those frequency tracks can actually help us pinpoint the right harmonic or not, and how well they can narrow down the intrinsic parameters when things get messy near plunge.
Lev: And they also have to deal with that problem you mentioned earlier about the signal being spread across multiple harmonics, where each one only carries a small part of the total signal strength depending on things like eccentricity and spin.
Kai: The method they use for harmonic identification involves combining pairs of candidate tracks under different assumptions about which harmonic label they might have, and then using summary statistics like posterior medians and network acceptance rates to sort out the correct labeling.
Mira: That sounds like a clever way to use the outputs from the search pipeline to make a decision about which frequency track is which.
Lev: And for parameter constraints, they are applying this sequential simulation-based inference scheme, or SBI, using a truncated prior approach to progressively shrink the parameter volume very quickly.
Kai: They show that assuming you have a perfect track as your search output, the accuracy and precision of recovering the intrinsic parameters can get arbitrarily close to their true values with this process.
Title and authors: Mira: But they also look at what happens when you introduce noise into those tracks, and how that affects things near the plunge phase.
Lev: In a more realistic scenario with noisy tracks, the sequential pipeline can marginalize over those noise features and shrink the parameter space volume by approximately thirteen orders of magnitude in the best case.
Kai: And then they take that result and apply another detection statistic over that narrowed region, which further shrinks the parameter space volume down to a final size that is about eight orders of magnitude smaller than the starting prior.
Mira: So, the core idea here seems to be that having an efficient time-frequency search pipeline is actually super important because a lot of the physical information content and signal strength lies in that non-linear fraction of the frequency evolution.
Lev: They also found that if you use noisy tracks, the sequential pipeline can marginalize over those noise features, which helps a lot with the parameter constraints.
Kai: It’s interesting because they found that this sequential SBI approach is preferred over running standard MCMC simulations in waveform space because of the multimodal structure of the posterior across parameter space.
Mira: That's significant because MCMC sampling can get stuck in those secondary peaks where it doesn't go to get the best results, but this new method handles that better by shrinking the volume sequentially.
Lev: The paper also mentions some potential improvements for future work, like refining the noise model to be specific to each harmonic frequency track and training a neural network to predict the most powerful set of harmonics given intrinsic parameters.
Kai: I think what they suggest about improving those networks by including phase information and amplitude modulation alongside the fundamental frequency evolutions could make a big difference.
Mira: Yeah, incorporating those extra details might let the AI model get even better at predicting which harmonics are actually going to be most informative for the measurement.
Lev: And that final shrinkage step using detection statistics over the region proposed by sequential SBI sounds like it could take that prior volume down to about eight orders of magnitude smaller than before, which is quite a reduction.
Title and authors: Kai: It really shows how this frequency-track based SBI approach provides a fast, waveform-free way to identify EMRI harmonics and get tight, reliable prior volumes for targeted follow-up parameter estimation.
Mira: That fast inference part is what I think is going to be most useful for people who are already doing MCMC runs and need to narrow down their search space quickly.
Lev: It substantially reduces the burden on those subsequent MCMC analyses, which is something we definitely need in this high-precision regime of EMRI parameter inference.
Kai: So, to wrap up the "Simulation-Based Inference on EMRI Frequency Tracks: Harmonic Identification and Parameter Constraints from Search Outputs" paper: they show that using frequency tracks from a search pipeline allows you to identify the correct harmonic pair with over ninety-five percent of the time across parameter space, and they demonstrate that sequential SBI can shrink the parameter space volume by up to a hundred thirteen orders of magnitude in just a few hours of GPU time.
Mira: That’s a huge reduction, and it really highlights how much physical information is packed into those frequency evolutions that we need to utilize better.
Lev: The caveat they mention is that if the search output is masked, the sequential procedure is less effective providing only about four orders of magnitude shrinkage in the worst case scenario when using detection statistics for further shrinkage.
Kai: It’s a strong result, but it shows that while this method is powerful, its performance does depend on how much information you actually lose due to masking near the plunge.
Mira: So, for those of us listening who are focused on the big picture, the implication is that we don't necessarily need to run every single waveform simulation if we can get a good frequency track output from a search pipeline.
Lev: And for those of us building hardware, it means that whatever data you feed into your networks has to be very clean and representative of what you expect to measure, because the noise features really matter when you are trying to marginalize over them.
Kai: So this paper gives us a new toolset: a fast way to identify harmonics and a powerful way to constrain parameters, which should definitely make subsequent MCMC analyses much more feasible for LISA data.
The paper's summary: Kai: So this paper shows that instead of just looking at the whole messy waveform to find an EMRI, you can actually use those recovered frequency tracks from a time-frequency search pipeline to do two things: identify which specific harmonic that track is, and then constrain the physical parameters of the inspiral.
Mira: That’s right. They’re suggesting that having these frequency tracks is really useful because a lot of the actual physical information you need about the inspiral, like its mass ratio or spin, is actually hidden in how those frequencies evolve over time in a non-linear way.
Lev: I mean, it's interesting that they focus on what happens when you get noisy tracks. They look at how that noise affects things near the plunge phase and if the pipeline can still do something useful then.
Kai: Yeah, they show that even with noisy tracks, this sequential inference method can marginalize over those noise features and shrink the parameter space volume by a lot in the best case scenario.
Mira: And that final step using a detection statistic over that narrowed region really pushes it to a much tighter constraint, shrinking the volume down to about eight orders of magnitude smaller than what you started with.
Lev: That’s where I see some real value for hardware. If we can get these tracks from our detectors and use this method, it means the subsequent simulations or analyses we have to run are going to be much faster and less computationally expensive for people doing MCMC work.
Kai: Exactly. It moves the heavy lifting of searching and parameter estimation away from needing full waveform generation, which takes a ton of CPU time.
Mira: The authors also point out that they found a way to classify the correct harmonic pair using summary statistics like posterior medians and network acceptance rates, which helps them figure out which label is right.
Lev: And that classification works pretty well across different masking levels, meaning it’s robust even when you lose a lot of data near the plunge.
Kai: So, this whole idea is about taking the output of a search pipeline—those frequency tracks—and turning them into a much more efficient way to narrow down what we actually care about in terms of EMRI parameters.
Mira: It gives us a practical strategy that’s preferred over running standard MCMC runs in the waveform space because the posterior distribution is so complex and multimodal there.
Lev: It's fast, waveform-free inference, which means we can get tight prior volumes for targeted follow-up studies without having to wait for hours of simulation time.
Kai: So this work really underlines that getting a good frequency track from a search pipeline is important because it contains a lot of the physical information content and signal strength that you need.
Mira: And the future work they suggest is pretty cool, like training neural networks to predict which harmonics will be most powerful given the intrinsic parameters.
Lev: If we can get those models trained right, incorporating phase and amplitude modulation into those networks could really help us understand the signal better.
Kai: That sounds like a solid plan for how we can make these frequency-track methods even more accurate down the line.
The paper's improvements: Tom: So we’re looking at how they plan to make this even better than what they just showed us about using frequency tracks for EMRI analysis.
Kai: They are suggesting a few things, like making the noise model more specific to each harmonic frequency track, which means tailoring the way we think about the noise based on what kind of signal we're looking at.
Mira: And they want to use a larger set of harmonic labels for their classification process, trying out more possible assignments to see if that improves how accurately they identify the right harmonic pair.
Lev: I think training a neural network is also big, especially if it can predict which harmonics are going to be most informative based on the intrinsic parameters of the inspiral.
Kai: Right, and they mentioned including phase information and amplitude modulation in those neural networks alongside just looking at the fundamental frequency evolutions.
Mira: That would be a significant step, because phase and amplitude carry a lot of physical detail about the system that you're missing if you only look at one frequency track.
Lev: From my side, I think incorporating those extra details into the network training could lead to better constraints when we try to run this on real hardware later.
Kai: That makes sense, because if the model can see more of the signal structure, it should be able to learn more accurately what’s important.
Mira: And they’re confident that applying those detection statistics over that region proposed by the sequential inference scheme could shrink the parameter space volume down even further, to about eight orders of magnitude smaller than before.
Lev: That final shrinkage step is a big deal because it means we can get really tight prior volumes for follow-up studies without having to run massive amounts of simulation time.
Kai: So they’re showing that this approach isn't just a one-off; it has a clear path forward to refining both the harmonic identification and the parameter constraints.
Mira: It shows that this method is really flexible enough to adapt its assumptions based on the specific signal we are analyzing, which is key for complex systems like EMRIs.
Lev: It’s good to see them thinking about how to handle those noise features more granularly, because in real hardware, those small details matter a lot when you're trying to marginalize over things.
Kai: So this paper lays out a really solid roadmap for improving these frequency-track based methods moving forward.
Mira: It gives us confidence that we can get much tighter constraints on the intrinsic EMRI parameters than we could with the current search outputs alone.
Conclusion: Kai: So to wrap up, this paper on "Simulation-Based Inference on EMRI Frequency Tracks: Harmonic Identification and Parameter Constraints from Search Outputs" shows that using frequency tracks from a search pipeline is a fast way to identify the correct harmonic and get tight constraints on the physical parameters of an EMRI.
Mira: It really highlights how crucial it is to have an efficient time-frequency search pipeline because that non-linear part of the frequency evolution holds a lot of the actual physical information you need.
Lev: I think this approach is going to be a big deal for reducing the computational overhead later, because if we can narrow down the parameter space so much faster, it makes subsequent MCMC analyses much more feasible on real hardware.
Kai: Exactly. And they prove that even with noisy tracks, their sequential inference method can marginalize over those noise features and shrink the volume by about thirteen orders of magnitude in the best case scenario.
Mira: That final step using detection statistics over that narrowed region is what really pushes it to a final volume about eight orders of magnitude smaller than the starting point.
Lev: It’s a strong result, but I do want to mention that they flag that if the search output is masked, the sequential procedure isn't as effective, giving only about four orders of magnitude shrinkage in the worst case.
Kai: So it shows that while this method is powerful, its performance does depend on how much information you actually lose due to masking near the plunge.
Mira: It’s a solid tool for getting tight prior volumes quickly without needing to do massive waveform simulations every single time we want to estimate parameters.
Lev: I think moving away from full waveform simulation for those initial steps is something we need, especially as we build out our error correction codes, because the overhead of running those simulations is just too much.
Kai: So this work really gives us a new toolset: a fast way to identify harmonics and a powerful way to constrain parameters that should make subsequent analyses much more feasible for LISA data.
Mira: It's definitely something we need to keep an eye on as we try to build the next generation of EMRI analysis tools.
Lev: Next time, I think we should look at how these constraints translate into actual limits on the black hole parameters themselves.
F. Fantoccoli, L. Speri, B. Depoorter, J. R. Gair
Max Planck Institute for Gravitational Physics (Albert Einstein Institute) · European Space Agency (ESTEC) · Leiden Observatory, Leiden University · Department of Physics and Astronomy, KU Leuven · Leuven Gravity Institute, KU Leuven · Royal Observatory of Belgium
gr-qc
Submitted: 2026-11-10
Updated: 2026-11-10
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 91/100
The gist: Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long…
Key concepts
- Extreme Mass Ratio Inspirals (EMRIs)
- EMRIs are binary systems where a compact object, like a black hole, orbits a much more massive one. These systems are key targets for LISA because their long, complex signals allow for very precise measurements of the objects' physical properties over many orbital cycles.
- Time-Frequency Search Pipeline
- This is an initial search method that analyzes gravitational wave data to recover frequency tracks, which show how the signal's frequency evolves over time. The paper uses these recovered tracks as input instead of analyzing the full waveform directly.
- Simulation-Based Inference (SBI)
- SBI is a machine learning technique used to approximate complex posterior probability distributions, like those describing EMRI parameters. It works by training a neural network on simulated data generated from different parameter sets to learn how to map observed data back to the most likely physical parameters.
Terminology
Summary
Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long, multiharmonic waveforms make their search and parameter estimation computationally expensive and challenging with standard methods Context. p. 0 Box 9513, 2300 RA Leiden, The Netherlands
4 Department of Physics and Astronomy , KU Leuven, Celestijnenlaan 200D, B-3001 Leuven, Belgium
5 Leuven Gravity Institute, KU Leuven, Celestijnenlaan 200D box 2415, 3001 Leuven, Belgium
6 Royal Observatory of Belgium, Avenue Circulaire, 3, 1180 Uccle, Belgium
Received —, accepted —
ABSTRACT
Context. Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long, multiharmonic waveforms make their search and parameter estimation computationally expensive and challenging with standard methods. p. 0 Box 9513, 2300 RA Leiden, The Netherlands
4 Department of Physics and Astronomy , KU Leuven, Celestijnenlaan 200D, B-3001 Leuven, Belgium
5 Leuven Gravity Institute, KU Leuven, Celestijnenlaan 200D box 2415, 3001 Leuven, Belgium
6 Royal Observatory of Belgium, Avenue Circulaire, 3, 1180 Uccle, Belgium
Received —, accepted —
ABSTRACT
Context. Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long, multiharmonic waveforms make their search and parameter estimation computationally expensive and challenging with standard methods. Box 9513, 2300 RA Leiden, The Netherlands
4 Department of Physics and Astronomy , KU Leuven, Celestijnenlaan 200D, B-3001 Leuven, Belgium
5 Leuven Gravity Institute, KU Leuven, Celestijnenlaan 200D box 2415, 3001 Leuven, Belgium
6 Royal Observatory of Belgium, Avenue Circulaire, 3, 1180 Uccle, Belgium
Received —, accepted —
ABSTRACT
Context. Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long, multiharmonic waveforms make their search and parameter estimation computationally expensive and challenging with standard methods. Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long, multiharmonic waveforms make their search and parameter estimation computationally expensive and challenging with standard methods. A promising alternative is to work directly with the output of a time–frequency search pipeline where one or more recovered frequency
tracks are identified. Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long, multiharmonic waveforms make their search and parameter estimation computationally expensive and challenging with standard methods. A promising alternative is to work directly with the output of a time–frequency search pipeline where one or more recovered frequency
tracks are identified.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Aims of the Investigation
The investigation aims to determine whether frequency-evolution tracks recovered by a semi-coherent search pipeline can be used to (i) identify which harmonic(s) a recovered track corresponds to; and (ii) constrain the intrinsic EMRI parameters, including an assessment of how these constraints degrade when the tracks are partially masked near plunge. The main goal is to use the output of an EMRI search providing one or more frequency tracks to address three specific questions. p. 1 For
this reason, we aim to understand whether the frequency evolution alone can provide useful constraints on the source parameters and thereby inform subsequent, computationally expensive
waveform-based analyses. Our main goal is to use the output of an EMRI search providing one or more frequency tracks to address the following
questions:
1.Where the paper says it
Methods for Harmonic Identification
For harmonic identification, the authors combine pairs of candidate tracks under different harmonic-label assumptions, using the resulting posterior medians and network acceptance rates as summary statistics to classify the correct labelling. The pipeline recovers the correct harmonic pair among the top-2 ranked candidates with probability greater than 0. p. 0 For parameter identification, we apply a sequential (truncated-prior) SBI scheme to progressively shrink the
parameter volume, and further refine the results by cross-checking posterior samples against a time–frequency detection statistic. The harmonic-identification pipeline recovers the correct harmonic pair among the top-2 ranked candidates with probability
> 0. The harmonic-identification pipeline recovers the correct harmonic pair among the top-2 ranked candidates with probability
> 0. 99 across masking levels, and correctly identifies it outright in 98% of cases when restricted to physically plausible candidates.Where the paper says it
Where the paper says it
Methods for Parameter Constraints
The sequential SBI parameter-identification pipeline applies a truncated-prior scheme to progressively shrink the parameter volume, achieving a reduction of up to 10 13 within a few hours of GPU time. This is further refined by cross-checking posterior samples against a time–frequency detection statistic, which reduces the constrained volume by up to four additional orders of magnitude, approaching precision achievable in the least-masked cases even for heavily masked tracks. p. 0 The sequential SBI parameter-identification pipeline shrinks the intrinsic-parameter prior volume by up to 1013 within a few hours of
GPU time. Combining the SBI posterior with a detection-statistic-based refinement step further reduces the constrained volume by
up to four additional orders of magnitude, approaching the precision achievable in the least-masked cases even for heavily masked
tracks.Where the paper says it
The Training Signal Construction
The training set is built by randomly sampling 5 × 10 5 parameter sets from uniform distributions over the ranges given in Table 1, and then rolling each two-years track to the sampled time to plunge, applying a mask by drawing the masked fraction uniformly within the interval reported in Table 1. The noisy version of the fundamental frequency tracks is produced through fnoisy(t) = fclean(t) 1 + δ f(t), where δ f(t) is of the order of 10-3. p. 3 The resulting residual band
is shown in the bottom panel of Fig. We use this band to map
realistic search pipeline uncertainties onto a different set of generated tracks as:
fnoisy(t) = fclean(t)
1 + δ f(t)
, (6)
where fclean is the track generated with fewtrax masked and
rolled as explained above, meanwhile fnoisy is our model of the
realistic output of a noisy search.Where the paper says it
Results and Inference
The sequential SBI process demonstrates that assuming a perfect track as search output, the accuracy and precision of the recovery through this process can be brought arbitrarily close to the true values of the intrinsic parameters. In a more realistic scenario with noisy tracks, the sequential pipeline is able to marginalize over noise features and shrink the parameter space volume by approximately 13 orders of magnitude in the best case scenario. p. 9 We found that assuming a perfect track as search output, the
accuracy and precision of the recovery through the sequential
SBI process can be brought arbitrarily close to the true values
of the intrinsic parameters. In a more realistic scenario, where the search pipeline recovers noisy tracks, we find that the sequential pipeline is able to
marginalize over the noise features and shrink the parameter
space volume by ∼ 13 orders of magnitude with respect to
the prior in the best case scenario. 3 for more details), the sequential
procedure is less effective providing ∼ 4 orders of magnitude
shrinkage in the worst case scenario. However, we demonstrated that a subsequent application of the same detection
statistic used in the search over the region of parameter space
proposed by the sequential SBI shrinks the parameter space
volume down to a final volume that is ∼ 8 orders of magnitude smaller than the prior.Where the paper says it
Where the paper says it
Discussion and Conclusion
The results underline that having an efficient time–frequency search pipeline able to recover the complete frequency track is important because a large part of the physical information content and SNR lies in the non-linear fraction of the frequency evolution. The work provides a first exploration of EMRI time–frequency tracks, demonstrating that given two frequency tracks, the pipeline recovers the correct harmonic pair associated with both tracks more than 95% of the time across parameter space. p. 9 We
summarize our main results, which address the questions raised
in Sec. We demonstrated that given two frequency tracks our
pipeline is able to recover the correct harmonic pair associated to both tracks more than 95% of the time, across parameter space. A different prior in further rounds can be adopted, but
the final posterior distributions must be corrected accordingly. We choose the sequential SBI approach instead of a MCMC run
in waveform space because of the multimodal structure of the
posterior across parameter space: Chua & Cutler (2022) showed
that it is possible to explore the multimodality through a series of
MCMC runs in the identified volume of parameter space.Where the paper says it
Where the paper says it
Future Improvements
Possible improvements include refining the noise model to be specific to each harmonic frequency track, using a larger set of harmonic labels, and training a neural network to predict the most powerful set of harmonics given intrinsic parameters. Significant improvements could involve including phase information and amplitude modulation into the neural network training alongside the fundamental frequency evolutions. p. 1 1, we work with single harmonic frequency evolution tracks, focusing on the frequency evolution
alone and neglecting the information carried by the amplitude
of the signal in its time–frequency representation. The intrinsic
parameters of an EMRI shape its frequency evolution changing its frequency span (primary mass dependence), its derivative (mostly determined by primary and secondary mass) and
the growth in the lead up to the plunge phase (shaped by the dimensionless spin parameter and eccentricity). 3 for more details), the sequential
procedure is less effective providing ∼ 4 orders of magnitude
shrinkage in the worst case scenario. However, we demonstrated that a subsequent application of the same detection
statistic used in the search over the region of parameter space
proposed by the sequential SBI shrinks the parameter space
volume down to a final volume that is ∼ 8 orders of magnitude smaller than the prior.Where the paper says it
Where the paper says it
Acknowledgments
The authors thank Stephen Green, Diganta Bandopadhyay, Xue-Ting Zhang, Ludovica Carbone, Philippa Cole and Christian Chapman-Bird for useful discussions, comments and suggestions received during the course of this work. The paper is manuscript no. p. 0 Astronomy & Astrophysics manuscript no. aanda ©ESO 2026
November 11, 2026
Simulation-Based Inference on EMRI Frequency Tracks: Harmonic
Identification and Parameter Constraints from Search Outputs
F.Where the paper says it
Key Words
Methods: data analysis, Gravitational waves Methods include data analysis and gravitational waves. A&A proofs manuscript no. p. 1 A&A proofs: manuscript no. aanda
ler 2022) in parameter space, which make MCMC sampling
extremely inefficient and potentially biased: at these degeneracies the posterior surface develops secondary peaks, where the
MCMC walkers tend to linger. Frequency-track-based SBI provides a fast, waveform-free method to identify EMRI harmonics and to obtain tight,
reliable prior volumes to enable targeted follow-up parameter estimation, substantially reducing the burden on subsequent MCMCbased analyses and offering a practical strategy for the multimodal, high-precision regime characteristic of EMRI parameter inference. Methods: data analysis, Gravitational waves
1. Frequency-track-based SBI provides a fast, waveform-free method to identify EMRI harmonics and to obtain tight,
reliable prior volumes to enable targeted follow-up parameter estimation, substantially reducing the burden on subsequent MCMCbased analyses and offering a practical strategy for the multimodal, high-precision regime characteristic of EMRI parameter inference. Methods: data analysis, Gravitational waves
1.Where the paper says it
Where the paper says it
Where the paper says it
References
Babak, S., Gair, J., Sesana, A., et al. 2017 Physical Review D, 95 Babak, S. p. 0 The gravitational wave (GW)
emission from EMRIs spans the frequency range 10−4–10−1 Hz,
making such systems a primary target for the LISA mission
(Babak et al. 2017; Berry et al. In particular, sequential simulation-based inference (SBI) techniques, such as Truncated Marginal Neural Ratio Estimation, have shown potential for efficiently constraining
the high-dimensional EMRI parameter space starting from broad
prior distributions (Cole et al. A distinct class of methods
operates in the time–frequency domain. EMRIs are binary systems in
which a compact object (CO), most likely a black hole (BH) with
mass m2 ∈ [10, 100] M⊙, orbits a massive black hole (MBH)
with typical mass m1 ∈ [105
, 107
] M⊙, corresponding to mass
ratios q in the range 10−3–10−6
. The gravitational wave (GW)
emission from EMRIs spans the frequency range 10−4–10−1 Hz,
making such systems a primary target for the LISA mission
(Babak et al. 2025; Babak et al. 2009b; Cornish
2011). The gravitational wave (GW)
emission from EMRIs spans the frequency range 10−4–10−1 Hz,
making such systems a primary target for the LISA mission
(Babak et al. 2017; Berry et al. Box 9513, 2300 RA Leiden, The Netherlands
4 Department of Physics and Astronomy , KU Leuven, Celestijnenlaan 200D, B-3001 Leuven, Belgium
5 Leuven Gravity Institute, KU Leuven, Celestijnenlaan 200D box 2415, 3001 Leuven, Belgium
6 Royal Observatory of Belgium, Avenue Circulaire, 3, 1180 Uccle, Belgium
Received —, accepted —
ABSTRACT
Context. Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long, multiharmonic waveforms make their search and parameter estimation computationally expensive and challenging with standard methods. The use of veto likelihoods has also been proposed as
a mechanism to suppress spurious local maxima and improve
convergence towards the true signal parameters (Chua 2022). Machine-learning techniques have emerged as promising tools
for accelerating both detection and inference. These systems pose a number of data analysis challenges. First, the length and complexity of the GW signal make waveform generation and likelihood evaluation expensive (∼ 100 ms
for a 4 year long, q = 10−5
signal (Chapman-Bird et al. 2017; Berry et al. 2019; Speri et al. 2017; Berry et al. 2019; Speri et al. 2017; Berry et al. 2019; Speri et al. 2008b; Babak et al. 2009a; Chua & CutArticle number, page 1 of 15
arXiv:2611. A different prior in further rounds can be adopted, but
the final posterior distributions must be corrected accordingly. We choose the sequential SBI approach instead of a MCMC run
in waveform space because of the multimodal structure of the
posterior across parameter space: Chua & Cutler (2022) showed
that it is possible to explore the multimodality through a series of
MCMC runs in the identified volume of parameter space. 2009b; Cornish
2011). The use of veto likelihoods has also been proposed as
a mechanism to suppress spurious local maxima and improve
convergence towards the true signal parameters (Chua 2022). 2025b)),
a cost that could become prohibitive when repeated over the
many iterations required to search for and then infer the EMRI
parameters. Second, because of the precision with which the
parameters can be measured, the ratio between the posterior
and prior volumes is 10−5–10−6 per parameter, which makes
fully-coherent matched filtering searches unfeasible (Gair et al. The gravitational wave (GW)
emission from EMRIs spans the frequency range 10−4–10−1 Hz,
making such systems a primary target for the LISA mission
(Babak et al. 2017; Berry et al. 2025), and perform rapid parameter estimation. In particular, sequential simulation-based inference (SBI) techniques, such as Truncated Marginal Neural Ratio Estimation, have shown potential for efficiently constraining
the high-dimensional EMRI parameter space starting from broad
prior distributions (Cole et al. Methods: data analysis, Gravitational waves
1. Introduction
The detection and characterisation of Extreme Mass Ratio Inspirals (EMRIs) represents one of the most interesting challenges
in the context of Laser Interferometer Space Antenna (LISA)
(Colpi et al. Introduction
The detection and characterisation of Extreme Mass Ratio Inspirals (EMRIs) represents one of the most interesting challenges
in the context of Laser Interferometer Space Antenna (LISA)
(Colpi et al. 2024) data analysis. In particular, sequential simulation-based inference (SBI) techniques, such as Truncated Marginal Neural Ratio Estimation, have shown potential for efficiently constraining
the high-dimensional EMRI parameter space starting from broad
prior distributions (Cole et al. A distinct class of methods
operates in the time–frequency domain. Simulation-based inference
We introduce our notation starting from Bayes’ theorem,
p(θ|d) =
p(d|θ)p(θ)
p(d)
, (2)
where θ is the N-dimensional array of parameters and d the
dataset, p(d|θ) is the likelihood function, p(θ) the prior on the
parameters, p(d) the evidence and p(θ|d) the posterior distribution. The machine learning strategy at the centre of this
work is Simulation-Based Inference (SBI) (Cranmer et al. We provide some references to
applications of SBI in a gravitational-wave context: Green et al. (2020); Green & Gair (2021); Dax et al. (3) and train the NF approximation of the posterior qϕ(θ|d). We provide some references to
applications of SBI in a gravitational-wave context: Green et al. The machine learning strategy at the centre of this
work is Simulation-Based Inference (SBI) (Cranmer et al. 2020;
Deistler et al. 2025), and perform rapid parameter estimation. In particular, sequential simulation-based inference (SBI) techniques, such as Truncated Marginal Neural Ratio Estimation, have shown potential for efficiently constraining
the high-dimensional EMRI parameter space starting from broad
prior distributions (Cole et al. Gair1
1 Max Planck Institute for Gravitational Physics (Albert Einstein Institute), Am Mühlenberg 1, D-14476 Potsdam, Germany e-mail:
federico. fantoccoli@aei. p. 0 Gair1
1 Max Planck Institute for Gravitational Physics (Albert Einstein Institute), Am Mühlenberg 1, D-14476 Potsdam, Germany e-mail:
federico. fantoccoli@aei.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
Astronomy & Astrophysics manuscript no. aanda ©ESO 2026 November 11, 2026 Simulation-Based Inference on EMRI Frequency Tracks: Harmonic Identification and Parameter Constraints from Search Outputs F. p. 0 Astronomy & Astrophysics manuscript no. aanda ©ESO 2026
November 11, 2026
Simulation-Based Inference on EMRI Frequency Tracks: Harmonic
Identification and Parameter Constraints from Search Outputs
F. Gair1
1 Max Planck Institute for Gravitational Physics (Albert Einstein Institute), Am Mühlenberg 1, D-14476 Potsdam, Germany e-mail:
federico. fantoccoli@aei. The size of the bank depends on
the application, and for this work we use 5 × 105 parameter sets. We use the fewtrax python package (in preparation Depoorter
et al. Depoorter4, 5, 6
, and J. Gair1
1 Max Planck Institute for Gravitational Physics (Albert Einstein Institute), Am Mühlenberg 1, D-14476 Potsdam, Germany e-mail:
federico. fantoccoli@aei. de
2 European Space Agency (ESA), European Space Research and Technology Centre (ESTEC), Keplerlaan 1, 2201 AZ Noordwijk,
The Netherlands
3 Leiden Observatory, Leiden University, P. de
2 European Space Agency (ESA), European Space Research and Technology Centre (ESTEC), Keplerlaan 1, 2201 AZ Noordwijk,
The Netherlands
3 Leiden Observatory, Leiden University, P. Box 9513, 2300 RA Leiden, The Netherlands
4 Department of Physics and Astronomy , KU Leuven, Celestijnenlaan 200D, B-3001 Leuven, Belgium
5 Leuven Gravity Institute, KU Leuven, Celestijnenlaan 200D box 2415, 3001 Leuven, Belgium
6 Royal Observatory of Belgium, Avenue Circulaire, 3, 1180 Uccle, Belgium
Received —, accepted —
ABSTRACT
Context. A promising alternative is to work directly with the output of a time–frequency search pipeline where one or more recovered frequency
tracks are identified. We investigate whether frequency-evolution tracks recovered by a semi-coherent search pipeline can, on their own, be used
to (i) identify which harmonic(s) a recovered track corresponds to; and (ii) constrain the intrinsic EMRI parameters, including an
assessment of how these constraints degrade when the tracks are partially masked near plunge. We investigate whether frequency-evolution tracks recovered by a semi-coherent search pipeline can, on their own, be used
to (i) identify which harmonic(s) a recovered track corresponds to; and (ii) constrain the intrinsic EMRI parameters, including an
assessment of how these constraints degrade when the tracks are partially masked near plunge. We generate large banks of azimuthal and radial fundamental-frequency tracks and train Simulation-Based Inference (SBI)
networks on single-frequency and combined inputs. We generate large banks of azimuthal and radial fundamental-frequency tracks and train Simulation-Based Inference (SBI)
networks on single-frequency and combined inputs. For harmonic identification, we combine pairs of candidate tracks under different
harmonic-label assumptions, use the resulting posterior medians and network acceptance rates as summary statistics, and classify
the correct labelling. For harmonic identification, we combine pairs of candidate tracks under different
harmonic-label assumptions, use the resulting posterior medians and network acceptance rates as summary statistics, and classify
the correct labelling. For parameter identification, we apply a sequential (truncated-prior) SBI scheme to progressively shrink the
parameter volume, and further refine the results by cross-checking posterior samples against a time–frequency detection statistic.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 2 ---
A&A proofs: manuscript no. aanda ler 2022
Introduction
The detection and characterisation of Extreme Mass Ratio Inspirals (EMRIs) represents one of the most interesting challenges in the context of Laser Interferometer Space Antenna (LISA) data analysis. EMRIs are binary systems in which a compact object (CO), most likely a black hole (BH) with mass m2 ∈ [10, 100] M⊙, orbits a massive black hole (MBH) with typical mass m1 ∈ [10 5, 10 7] M⊙, corresponding to mass ratios q in the range 10-3–10-6. p. 0 2024) data analysis. EMRIs are binary systems in
which a compact object (CO), most likely a black hole (BH) with
mass m2 ∈ [10, 100] M⊙, orbits a massive black hole (MBH)
with typical mass m1 ∈ [105
, 107
] M⊙, corresponding to mass
ratios q in the range 10−3–10−6
. 2019; Speri et al. Owing to their slow orbital evolution, a typical EMRI remains in the
LISA band for several years, during which it undergoes many
orbital cycles, allowing extremely precise measurements of its
physical properties. These systems pose a number of data analysis challenges. First, the length and complexity of the GW signal make waveform generation and likelihood evaluation expensive (∼ 100 ms
for a 4 year long, q = 10−5
signal (Chapman-Bird et al. 2025b)),
a cost that could become prohibitive when repeated over the
many iterations required to search for and then infer the EMRI
parameters. Second, because of the precision with which the
parameters can be measured, the ratio between the posterior
and prior volumes is 10−5–10−6 per parameter, which makes
fully-coherent matched filtering searches unfeasible (Gair et al. Second, because of the precision with which the
parameters can be measured, the ratio between the posterior
and prior volumes is 10−5–10−6 per parameter, which makes
fully-coherent matched filtering searches unfeasible (Gair et al. Moreover, the complex trajectories followed by the CO
in the gravitational field of the more massive body imply that the
GW signal is a superposition of many harmonics, constructed
from combinations of the orbital frequencies along the radial
(Ωr), azimuthal (Ωϕ) and polar (Ωθ) directions.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Methods for Frequency Evolution
The intrinsic parameters are split into intrinsic and extrinsic groups, with intrinsic parameters including m1, m2, a, ef, Tpl, Φϕ0, Φr0, and Φθ0. The authors focus on equatorial prograde EMRIs where the orbit of the secondary lies in the equatorial plane throughout the evolution with its orbital angular momentum aligned with the spin of the primary. p. 1 In this work
we are mainly interested in constraining the intrinsic parameters. Furthermore, we restrict ourselves to equatorial, prograde
EMRIs, in which the orbit of the secondary lies in the equatorial
plane throughout the evolution, with its orbital angular momentum aligned with the spin of the primary. 2017), which is less accurate. A generic equatorial
EMRI frequency track fm,n can be decomposed as a function of
the fundamental frequencies in Kerr spacetime as follow:
fm,n(t) =
mΩϕ(t) + nΩr(t)
2π
, (1)
where the pair (m, n) identifies the harmonic and Ωϕ(t), Ωr(t)
are the fundamental azimuthal and radial frequencies in Kerr
spacetime (Fujita & Hikida 2009; Schmidt 2002).Where the paper says it
Where the paper says it
Simulation-Based Inference Framework
The machine learning strategy centers on Simulation-Based Inference (SBI), which approximates the posterior distribution using a Normalizing Flow (NF) network trained by minimizing the Kullback–Leibler (KL) divergence between the true underlying posterior and the NF approximation. In practice, samples are drawn from the prior distribution p(θ), used to produce simulated datasets d (likelihood draws), and these draws are used to minimize the KL divergence in Eq. p. 2 (3)
In practice, we draw samples from the prior distribution p(θ) of
the physical parameters θ and use them to produce the simulated
datasets d (i. , the likelihood draws), then we use such draws to
minimize the KL divergence in Eq. , the likelihood draws), then we use such draws to
minimize the KL divergence in Eq. (3) and train the NF approximation of the posterior qϕ(θ|d).Where the paper says it
Where the paper says it
Harmonic Labeling Procedure
To identify harmonics, the authors combine pairs of candidate tracks under different harmonic-label assumptions, using posterior medians and network acceptance rates as summary statistics to classify the correct labelling. They loop over plausible harmonic label assignments and use single frequency networks and a composite network to obtain three sets of posterior samples for each assignment. p. 5 Given a pair of single harmonic frequency evolutions, we
loop over the plausible harmonic label assignments and, for each
of them, pass the reconstructed fundamental frequencies through
both single frequency networks and the composite network, obtaining three sets of posterior samples in the physical parameters. For every assignment we then collect the medians of the
three posteriors for each parameter, together with the acceptance
rate of the composite network; we refer to these quantities as the
summary set. Given a pair of single harmonic frequency evolutions, we
loop over the plausible harmonic label assignments and, for each
of them, pass the reconstructed fundamental frequencies through
both single frequency networks and the composite network, obtaining three sets of posterior samples in the physical parameters. For every assignment we then collect the medians of the
three posteriors for each parameter, together with the acceptance
rate of the composite network; we refer to these quantities as the
summary set.Where the paper says it
Where the paper says it
Parameter Volume Shrinkage
The sequential SBI approach starts from a round-0 prior, where samples are drawn, an SBI network is trained to predict the posterior, and this information is used to select a smaller range of parameter space as the new prior for the next round. The final posterior distributions can be shrunk to obtain an arbitrarily accurate point estimate of intrinsic parameters in the absence of noise and provided a sufficiently dense training dataset. p. 6 We repeat the process starting from the new
round-1 prior in such a way that at every iteration we increase
the density of the sampling space around the region of interest. In the absence of noise in the observation (and in training
data) and provided a sufficiently dense training dataset, we find
that the posterior from a sequential-SBI can be shrank to obtain
an arbitrarily accurate point estimate of the intrinsic parameters. For the purpose of this work, we
define Vprior = Πθi
|θR,i −θR,i
|, where θL,i and θR,i are the left-hand
side and right-hand side of the prior edge for each parameter,
indexed by i. The volume shrinking is extremely important in
the context of EMRIs, where the posterior of intrinsic parameters is extremely sharp, with a posterior-to-prior one dimensional
volume ratio that can be of the order of 10−4 − 10−6
, and the
overall volume compression required for the whole posterior is
(10−4
)
5 ∼ 10−20 with respect to the prior volume (the needlein-a-haystack problem (Gair et al.Where the paper says it
Where the paper says it
Results on Constraints
The sequential SBI algorithm shows that the larger the masking fraction, the broader the per-round posterior after round-0, as information for O(1)% precision is contained in the plunging fraction of the inspiral. The detection statistic score is a reliable tool to shrink the interesting region of parameter space further than what is obtained through SBI, reducing the prior volume ratio by four orders of magnitude to a final level of Vbox/Vprior approximately 3. p. 9 The results are shown in Fig. As observed in the figure, the detection statistic score is a
reliable tool to shrink the interesting region of parameter space
further than what is obtained through SBI, reducing the prior
volume ratio by four orders of magnitude to a final level of
Vbox/Vprior ≃ 3. If the search output is a
masked track (see Sec. 3 for more details), the sequential
procedure is less effective providing ∼ 4 orders of magnitude
shrinkage in the worst case scenario.Where the paper says it
Where the paper says it
Conclusion
Frequency-track-based SBI provides a fast, waveform-free method to identify EMRI harmonics and to obtain tight, reliable prior volumes to enable targeted follow-up parameter estimation. This strategy substantially reduces the burden on subsequent MCMC-based analyses and offers a practical strategy for the multimodal, high-precision regime characteristic of EMRI parameter inference. p. 0 Conclusions. Frequency-track-based SBI provides a fast, waveform-free method to identify EMRI harmonics and to obtain tight,
reliable prior volumes to enable targeted follow-up parameter estimation, substantially reducing the burden on subsequent MCMCbased analyses and offering a practical strategy for the multimodal, high-precision regime characteristic of EMRI parameter inference.Where the paper says it
Methods: data analysis, Gravitational waves Methods include data analysis and gravitational waves. The key words are methods data analysis and gravitational waves.
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
-
Accurate harmonic identification: The system can recover correct harmonic pairs among top-2 candidates with probability greater than 0.99 across masking levels, and correctly identify it outright in 98% of cases when restricted to physically plausible candidates.
-
Parameter constraint refinement: A sequential simulation-based inference scheme can shrink the intrinsic-parameter prior volume by up to 1013 within a few hours of GPU time, and combining this with a detection-statistic-based refinement step further reduces the constrained volume by up to four additional orders of magnitude.
-
Fast, waveform-free inference: The frequency-track based SBI approach provides a fast, waveform-free method to identify EMRI harmonics and obtain tight, reliable prior volumes to enable targeted follow-up parameter estimation, substantially reducing the burden on subsequent MCMCbased analyses.
-
Robustness to noise and masking: The sequential pipeline is able to marginalize over noise features and shrink the parameter space volume by up to 13 orders of magnitude in the best case scenario, with a final volume comparable to the fmask = 0.15 case when using detection statistics for further shrinkage.
-
Harmonic misclassification mitigation: A classifier can be trained on summary statistics like posterior medians and acceptance rates to select the most probable harmonic pairs, outperforming simpler filtering strategies across all tested harmonic pairs.
Abstract
Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna, but their long, multi-harmonic waveforms make search and parameter estimation computationally expensive. A promising alternative is to work directly with frequency tracks recovered by a time-frequency search pipeline. We investigate whether such tracks can be used to (i) identify which harmonic(s) a recovered track corresponds to, and (ii) constrain the intrinsic EMRI parameters, including when tracks are partially masked near plunge. We first train Simulation-Based Inference (SBI) networks on large banks of azimuthal and radial fundamental-frequency tracks. For harmonic identification, we combine pairs of candidate tracks under different harmonic-label assumptions and use the resulting posterior medians and network acceptance rates to classify the correct labelling. For parameter identification, we apply a sequential (truncated-prior) SBI scheme and refine the results by cross-checking posterior samples against a time-frequency detection statistic. The correct harmonic pair is among the top-2 ranked candidates with probability >0.99 across masking levels, and is identified outright in 98% of cases when restricted to physically plausible candidates. Sequential SBI shrinks the intrinsic-parameter prior volume by up to 10 13 within a few hours of GPU time, and the detection-statistic refinement reduces it by up to four further orders of magnitude, even for heavily masked tracks. Frequency-track-based SBI thus provides a fast, waveform-free way to identify EMRI harmonics and obtain tight prior volumes for targeted follow-up, substantially reducing the burden on subsequent MCMC-based analyses.
Sources
- Sequential simulation-based inference for extreme mass ratio inspirals
- Unlocking New Paths for Science with Extreme-Mass-Ratio Inspirals: Machine Learning-Enhanced MCMC for Accurate Parameter Inversion
- Fast and Fourier: Extreme Mass Ratio Inspiral Waveforms in the Frequency Domain
- Searching for extreme mass ratio inspirals in LISA: from identification to parameter estimation
Related papers
- Accurate waveforms for generic planar-orbit binary black holes: The multipolar effective-one-body model SEOBNRv6EHM
- Analytic Marginalization Over Calibration Uncertainties in Gravitational Wave Inference and Application to Black Hole Ringdown
- A catalog of long, eccentric and precessing binary black hole simulations from ICTS-TIFR
- Pole Skipping, Avoided Crossing, and Resonant Excitation in Kerr Quasinormal Modes near Algebraically Special Frequencies
- Gravitational Waves from Green's Function Decomposition for a Kerr black hole: I. Equatorial ISCO Plunge
- A New Spin on Dissipative Tides: First-Post-Newtonian Effects in Compact Binary Inspirals