SlotFlow-GB: A Neural Framework for Trans-Dimensional Inference of Overlapping Galactic Binaries

Niklas Houba, Giovanni Giarda, Davide PirasarXiv 2026 · arXiv

The upcoming Laser Interferometer Space Antenna (LISA) mission faces a daunting "big data" problem: its millihertz gravitational-wave sky will be crowded with millions of Galactic binaries (GBs). Tens of thousands of these will be individually resolvable, but their signals will overlap significantly in both time and frequency.

In a new paper, researchers introduce SlotFlow-GB, a neural framework designed to perform "trans-dimensional inference"—simultaneously determining how many binaries are in a data segment and what their physical parameters are—in a fraction of a second.

TL;DR

SlotFlow-GB uses a novel combination of slot attention and normalizing flows to perform near-instantaneous inference on overlapping gravitational-wave signals. It accurately estimates source counts and parameter distributions for up to ten overlapping binaries in just 41 milliseconds, providing a powerful "fast-forward" button for LISA data analysis.

What Problem Does This Paper Address?

Traditional gravitational-wave analysis often treats signals as isolated events. However, LISA will observe a "signal-dominated" regime where the primary challenge isn't just detecting a source against noise, but disentangling it from thousands of other overlapping sources.

This creates a trans-dimensional inference problem: because the number of sources ($K$) is unknown, the total number of parameters to be inferred changes. Standard techniques like Reversible-Jump Markov Chain Monte Carlo (RJMCMC) can handle this but are computationally grueling, often requiring millions of likelihood evaluations to "jump" between different dimensional spaces. SlotFlow-GB aims to replace this expensive search with a single, amortized neural network evaluation.

Methods

The architecture of SlotFlow-GB (illustrated in Figure 2) consists of several key stages:

  1. Preprocessing: Year-long data from LISA’s Time-Delay Interferometry (TDI) channels are segmented and Fourier-transformed to capture time-dependent responses.
  2. Convolutional Encoder: A CNN processes the spectra to extract high-level features.
  3. Competing Source Slots: The model uses a "Slot Attention" mechanism with a fixed pool of 13 slots. Through iterative attention, these slots "compete" for signal information, encouraging the model to assign distinct binaries to distinct slots.
  4. Normalizing Flows: Each slot predicts an existence probability ($p_i$) and conditions a shared Rational-Quadratic Spline Flow. This flow represents an 8-dimensional probability density for that specific binary's parameters (e.g., frequency, sky position, amplitude).
  5. Trans-Dimensional Output: The source count $K$ is derived as a Poisson binomial distribution over the slots.
FIG. 2. Architecture of SlotFlow-GB. The TDI observation and nominal frequency pointer (1) undergo fixed preprocessing, including time segmentation, spectral selection, and summary-feature extraction (2). A learned encoder combines convolutional features from the A and E channels with projected summary features from all three TDI channels (3). Competing source slots (4) predict existence probabilities and condition a shared normalizing flow for the source parameters (5). The resulting outputs comprise a source-count distribution q(K | d) and the parameter densities qi(θi | d) of the selected sources (6).

Datasets

The authors trained and evaluated the model on simulated year-long LISA observations:

  • Training Set: 2 million simulated "scenes," each containing between 0 and 10 binaries within a 2 µHz frequency band.
  • Evaluation Sets: As detailed in Table I (Figure 4), the model was tested on a general population set, a "two-source" set to probe resolution limits, and a "multiplicity grid" to see how performance scales with crowding.

Key Insights and Ideas

  • Permutation Invariance: By using slot attention and Hungarian matching during training, the model learns that the order of sources doesn't matter—a fundamental requirement for set-based inference.
  • Frequency Anchoring: To achieve the extreme precision required for gravitational waves, the model doesn't predict absolute frequency directly. Instead, it identifies a spectral peak to use as an "anchor" and predicts a small offset, significantly easing the learning task.
  • Amortized Initialization: Perhaps the most practical insight is that SlotFlow-GB doesn't have to be the final word. Its output can be used to "seed" traditional likelihood-based samplers, allowing them to skip the expensive global search phase and jump straight to refining the solution.

Key Results

  • Speed: A complete evaluation takes just 41 ms on an NVIDIA GH200 GPU. For comparison, traditional global fits can take hours or days for similar tasks.
  • Counting Accuracy: The model identifies the exact number of sources in 61% of cases and is correct to within $\pm 1$ source in 94% of cases (see Figure 6).
  • Resolution: SlotFlow-GB can distinguish individual binaries even when their frequencies differ by only ~73 nHz—roughly the width of a single Fourier bin in the input data.
  • Refinement Boost: When used to initialize an MCMC sampler, SlotFlow-GB allowed the sampler to reach the true likelihood mode significantly faster than starting from a blind prior.
A demonstration of SlotFlow-GB resolving seven overlapping binaries, showing the measured band, the individual source densities, and the inferred source count.

Limitations and Future Work

While a successful proof-of-concept, the authors note several areas for improvement:

  • High-SNR Focus: This study focused on relatively "loud" sources to isolate the challenge of overlap. Real LISA data will include a "confusion noise" background of millions of weak, unresolvable sources.
  • Factorized Posterior: The current model assumes that the parameter densities of different slots are independent. In reality, the parameters of overlapping sources are often highly correlated.
  • Calibration: The source-count probabilities are currently somewhat over-conservative, meaning the model is more uncertain about the number of sources than its actual error rate suggests.
  • Real-world Noise: Future iterations will need to handle non-stationary noise, data gaps, and "glitches" common in space-borne detectors.

Summary updated

Abstract

The Laser Interferometer Space Antenna (LISA) will observe the millihertz gravitational-wave sky, including millions of long-lived Galactic binaries (GBs), tens of thousands of which are expected to be individually resolvable throughout the mission's lifetime. Extracting these sources requires jointly determining their number and physical parameters from strongly overlapping signals, posing a computationally demanding trans-dimensional inference problem. We introduce SlotFlow-GB, a neural framework that constructs a learned approximation to the trans-dimensional posterior, producing a probability distribution over source multiplicity and parameter densities for individual binaries in a single evaluation, without requiring their number to be specified in advance. We demonstrate this framework using year-long LISA simulations containing up to ten binaries within a 2 $μ$Hz frequency interval, focusing on a high-SNR regime in which source overlap, rather than individual-source detectability, is the primary challenge. The inferred source count differs from the true count by at most one in 94% of cases, while the model identifies individual binaries even when their signals overlap strongly in frequency. A complete evaluation, including parameter sampling, takes just 41 ms per band on one NVIDIA GH200 GPU, or 530 ms on eight laptop CPU threads. We further demonstrate that these distributions can be used to initialize subsequent likelihood-based inference, connecting fast amortized predictions with joint parameter estimation.
BibTeX
@misc{houba2026slotflow,
  author = {Niklas Houba and Giovanni Giarda and Davide Piras},
  title = {{SlotFlow-GB: A Neural Framework for Trans-Dimensional Inference of Overlapping Galactic Binaries}},
  year = {2026},
  eprint = {2611.26057},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26057},
}
Figure