Beyond classification: regressing physical merger properties using solid harmonic wavelet bispectrum

arXiv:2611.14630 · astro-ph.IM, astro-ph.GA · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: I'm Vera.

Jocelyn: And I'm Jocelyn, here with Subrahmanyan.

Vera: We're looking at the paper: "Beyond classification: regressing physical merger properties using solid harmonic wavelet bispectrum".

Jocelyn: Here's the gist of the paper. Galaxy mergers play an important role in galaxy evolution; they enhance star formation, transform morphologies, and may drive accretion onto black holes Galaxy mergers play an important role in galaxy evolution; they enhance star formation, transform morphologies, and may drive accretion onto black holes.

Vera: The solid harmonic wavelet bispectrum is introduced as an alternative to Convolutional Neural Networks (CNNs) because it encodes invariances and symmetries in the data which need to be learned by CNNs, as an alternative to CNNs.

Jocelyn: This method is motivated by four requirements for regression on continuous merger properties: (i) Rotation and translation invariance; (ii) Multi-scale sensitivity; (iii) Cross-scale interactions; and (iv) Deformation stability.

Vera: The solid harmonic wavelet scattering bispectrum captures quadratic interactions between morphological structures, while the angular frequency constraint causes rotational phase factors to cancel exactly, preserving invariance without discarding phase information.

Paper summary: Vera: It means we can use mathematical tools that capture intrinsic symmetries in galaxy shapes to extract much richer physical information than simple classification models provide. The authors are pointing toward a method that works well even when we don't have enormous training sets, which is a practical win for survey work.

Jocelyn: I think the implication is that for large-scale surveys like LSST or Euclid, we might be able to use these wavelet representations as powerful features without needing the massive computational power that current deep learning models demand just to train them.

Subrahmanyan: From a theoretical standpoint, if this method holds up on simulated data and shows strong regression performance on mass ratio and time since merger, it suggests that the underlying physical relationships encoded by those solid harmonic wavelets are indeed fundamental to how mergers drive accretion onto black holes in the first place.

Vera: It’s about moving from just knowing *if* a merger happened to understanding *what kind* of merger it was based on its measurable properties, which helps us build better physical models of galaxy growth.

Jocelyn: Ultimately, they are suggesting that for continuous data analysis in cosmology, a method with transparent functional structure and good performance in data-scarce regimes is a strong alternative to black box deep learning approaches.

Conclusion: Vera: So we're wrapping up on these papers about galaxy mergers, and this one is called "Beyond classification: regressing physical merger properties using solid harmonic wavelet bispectrum."

Jocelyn: It basically takes something that's usually just sorting mergers into classes, like classifying them as a one to one hundred mass ratio, and tries to predict the actual numbers—the mass ratio and how long ago the merger happened.

Subrahmanyan: That’s the core idea. They aren't stopping at telling us *what* kind of merger it is; they're trying to get a better estimate of its physical parameters, which are crucial for our structure formation models.

Vera: The authors use this solid harmonic wavelet bispectrum technique as an alternative to something called Convolutional Neural Networks, which have been popular for classification but often need huge datasets.

Jocelyn: And what they found is that their method requires way fewer training samples to get good results, and it’s actually more robust when the data gets noisy.

Subrahmanyan: That robustness is key. If you're using this in a real survey, dealing with noise and limited data is a constant battle for observational astronomy.

Vera: They specifically managed to regress both the merger mass ratio and the time since merger at the same time, with uncertainties around zero point three three dex for mass ratio and about zero point six six gigayears for time since merger.

Jocelyn: That uncertainty level is pretty tight, especially when you think about how hard it is to measure those things from actual telescope images.

Subrahmanyan: For the bigger picture, if this works well on simulated data with these uncertainties, it means the mathematical structure of galaxy shapes—the way they interact across different scales—actually holds real physical meaning for predicting merger outcomes.

Vera: It suggests that instead of just using deep learning to find patterns, we can use these wavelet representations to build more transparent tools for extracting physical properties from images.

Jocelyn: It opens the door for applying this kind of feature extraction to future massive surveys like Euclid or LSST, where they'll be dealing with huge amounts of galaxy data.

Subrahmanyan: That’s where we need to keep an eye on it. We need to see if these results hold up when we apply them to real observational data, not just simulations.

Vera: Exactly. The next big question is how this method performs when you start moving from ideal simulated images to the messy, noisy reality of actual telescope observations.

Department of Computer Science, University of Bath · Department of Physics, University of Bath

astro-ph.IM, astro-ph.GA

Submitted: 2026-11-10

Updated: 2026-11-10

Comments: accepted for publication in MNRAS, results for regression of merger properties shown in Figs 4&7

DOI: 10.1093/mnras/stag2160

Code: https://github.com/illustristng/illustris_python

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

Importance score: 83/100

The gist: Galaxy mergers play an important role in galaxy evolution; they enhance star formation, transform morphologies…

Key concepts

Galaxy Mergers
Galaxy mergers are crucial events in galaxy evolution that significantly impact galaxies by enhancing star formation, changing their physical shapes (morphologies), and potentially fueling the accretion of matter onto central black holes.
Solid Harmonic Wavelet Bispectrum (SHWSBic)
This is a mathematical feature derived from solid harmonic wavelet scattering coefficients. It captures quadratic interactions between morphological structures while enforcing rotational invariance, allowing the model to learn physical symmetries directly without needing complex phase information that CNNs might struggle with.
Regression on Physical Properties
Instead of just classifying whether two galaxies are a merger or not, this technique aims to predict continuous numerical values related to the merger, such as the mass ratio (how much more massive one galaxy is than the other) and the time elapsed since they first collided.

Terminology

Summary

Galaxy mergers play an important role in galaxy evolution; they enhance star formation, transform morphologies, and may drive accretion onto black holes Galaxy mergers play an important role in galaxy evolution; they enhance star formation, transform morphologies, and may drive accretion onto black holes. p. 0

Where the paper says it

2 Beyond classification: regressing physical merger properties using solid harmonic wavelet bispectrum Alex Brown1 , Georgios Exarchakis1★, Mathilda Avirett-Mackenzie2 , Carolin Villforth2† 1Department of Computer Science, University of Bath, BA2 7AY, Bath, UK 2Department of Physics, University of Bath, BA2 7AY, Bath, UK Accepted 6th November 2026. Received YYY; in original form ZZZ ABSTRACT Galaxy mergers play an important role in galaxy evolution; they enhance star formation, transform morphologies, and may drive accretion onto black holes.

Methodology Rationale

The solid harmonic wavelet bispectrum is introduced as an alternative to Convolutional Neural Networks (CNNs) because it encodes invariances and symmetries in the data which need to be learned by CNNs, as an alternative to CNNs. This method is motivated by four requirements for regression on continuous merger properties: (i) Rotation and translation invariance; (ii) Multi-scale sensitivity; (iii) Cross-scale interactions; and (iv) Deformation stability. p. 3

Where the paper says it

Rather than learning features from data, we construct a fixed operator — with no trainable parameters — that encodes the relevant physical invariances by design. This is motivated by four requirements for regression on continuous merger properties: (i) Rotation and translation invariance.

The solid harmonic wavelet scattering bispectrum captures quadratic interactions between morphological structures, while the angular frequency constraint causes rotational phase factors to cancel exactly, preserving invariance without discarding phase information. p. 3
Where the paper says it

Taking the bispectrum of scattering responses — the third order of spectral analysis — captures quadratic interactions between morphological structures, while the angular frequency constraint causes rotational phase factors to cancel exactly, preserving invariance without discarding phase information. The resulting operator, the solid harmonic wavelet scattering bispectrum (SHWSB), is entirely determined before any training data is seen.

How the training signal is built

The simulated dataset is created by selecting mergers with a wide range of mass ratios and time since merger from IllustrisTNG following methodology laid out in Avirett-Mackenzie et al. (2024). The data involves idealised images to test algorithms on idealised data, as well as a suite of realistic data to test how the different methods perform on noisy data. p. 1

Where the paper says it

1 describes the selection of the merger sample from IllustrisTNG as well as derivation of merger mass ratios and time since merger used in the analysis. We then download idealised images to test algorithms on idealised data, we also create a suite of realistic data to test how the different methods perform on noisy data (see Section 2.

Dataset curation involves undersampling majority histogram bins for both merger mass ratio and time since last merger coalescence to create a balanced sample, while keeping the overall sample size large enough for the study. p. 2
Where the paper says it

Below, we show how the level of noise affects the performance of the methodology. 3 Dataset curation From the initial merger sample, we create two further task-specific datasets by undersampling majority histogram bins for both merger mass ratio and time since last merger coalescence.

Where the gains come from

The solid harmonic wavelet bispectrum outperforms standard methodologies by requiring fewer training samples and being robust to noise. For regression on idealised data, the best performance is achieved using SHWSBic features with KNN regression, suggesting that accurate prediction of merger features requires capturing not just global symmetries but also complex relationships between features at varying scales. p. 7

Where the paper says it

Switching from linear regression to KNN regression on SHWS features yields further improvement, and the best performance is achieved using SHWSBic features with KNN regression. This suggests that accurate prediction of merger features requires capturing not just global symmetries but also complex relationships between features at varying scales — precisely what the SHWSBic encodes.

The representation encodes physically meaningful morphological structures such as spiral arms, central bulges, and hierarchical cross-scale features. p. 5
Where the paper says it

1 Convolutional Neural Networks Convolutional Neural Networks (CNNs) have achieved remarkable success in a wide range of tasks by learning hierarchical features through cascaded convolutional layers, nonlinear activations, and pooling operations. In the context of galaxy morphology, early convolutional layers capture low-level structural features such as tidal tails, spiral arms, and central bulges, while deeper layers progressively learn abstract morphological representations relevant to the chosen task.

Performance and Robustness

The methodology demonstrates robustness to noise using a wide range of signal-to-noise conditions, where the bispectral analysis is particularly powerful in identifying non-Gaussianity. The model can regress both time since mergers and mass ratio, expanding on previous work in the literature regressing on time since merger. p. 1

Where the paper says it

Beyond this practical consideration, there is a methodological argument for representations that encode known physical symmetries rather than learning them, since a model whose invariances are structurally guaranteed requires fewer assumptions to be validated numerically, and its predictions can be more directly connected to the physical properties it encodes. In this paper, we aim to expand on the current literature by expanding regression to include both mass ratio and time since merger.

In terms of data efficiency, solid harmonic wavelet bispectrum features with k-nearest-neighbour regression perform extremely well for small sample sizes due to the minimal number of trainable parameters required. p. 1
Where the paper says it

In either case, training data remains scarce relative to the requirements of modern deep learning architectures. Methods with fewer trainable parameters that can perform well on small datasets are therefore not merely desirable but necessary for progress in merger analysis.

The results can be summarized as follows:

To analyse the performance of different approaches on galaxy classification, we select a sample of galaxy mergers from IllustrisTNG, with a wide range of galaxy mass ratios and time since merger.

"We compare several wavelet-based methodologies — Morlet wavelets with linear regression, solid harmonic wavelets with linear regression, and the solid harmonic wavelet bispectrum with both linear regression and KNN regression—against conventional CNNs and Geometric Deep Learning models."

"Although CNNs slightly outperform solid harmonic wavelets in simple binary classification tasks, solid harmonic wavelet bispectrum methodologies allow regression on merger features with uncertainties ∼ 0.33 on log(mu) and ∼ 0. p. 2

Where the paper says it

Example idealised and realistic galaxy images used throughout the paper. Leftmost image: idealised sample from IllustrisTNG simulation cropped to 32 half stellar mass radius, and log normalised to [0, 1].

66 Gyr on Δt on idealized images (Fig. p. 7
Where the paper says it

Images on a log scale over AB magnitude. Model Log 𝜇 Δt (Gyr) MAE RMSE MAE RMSE Random predictor 0.

7)."

We demonstrate the robustness of the proposed method to noise using a wide range of signal-to-noise conditions.

Conclusions

The solid harmonic wavelet bispectrum offers an appealing alternative to conventional deep learning techniques, with better performance especially at small sample sizes and a transparent functional methodology. p. 0

Where the paper says it

66 Gyr uncertainty in time since merger. We have therefore shown that solid harmonic wavelet bispectrum presents an alternative to currently used machine learning techniques.

This methodology therefore represents an alternative to CNNs, which better performance for small datasets (Fig. p. 1
Where the paper says it

In either case, training data remains scarce relative to the requirements of modern deep learning architectures. Methods with fewer trainable parameters that can perform well on small datasets are therefore not merely desirable but necessary for progress in merger analysis.

5) and able to perform regression on merger properties, with performance exceeding traditionally used methods. p. 1
Where the paper says it

2026) further increased the number of time bins when classifying mergers, showing that a more detailed analysis of merger properties is feasible, showing they can classify post-mergers into four time bins achieving accuracy exceeding 70%. The model used combines traditional CNNs with Vision Transformers.

In the future, such models can be applied to new large datasets (e. p. 1
Where the paper says it

2025), achieving strong performance but requiring substantially larger training sets due to their increased parameter counts. Fine-tuning foundation models, for specific tasks, may reduce the data requirements but results in an operator that is more opaque as it is trained on different datasets with a different objective.

g. LSST, Euclid) to better understand the impact of mergers on galaxy evolution. p. 0
Where the paper says it

2017; Kaviraj 2025, and references therein). Given their complex impact on galaxy evolution, detecting mergers has therefore been of significant interest.

Key Findings Summary

We show for the first time regression on both mass ratio and time since mergers, expanding on previous enhanced classification of mergers (Pearson et al. 2024; Ferreira et al. p. 1

Where the paper says it

2025; Ferreira et al. 2024; Huertas-Company et al.

2026).

"Our results can be summarized as follows: • To analyse the performance of different approaches on galaxy classification, we select a sample of galaxy mergers from IllustrisTNG, with a wide range of galaxy mass ratios and time since merger (Section 2) • We compare several wavelet-based methodologies — Morlet wavelets with linear regression, solid harmonic wavelets with linear regression, and the solid harmonic wavelet bispectrum with both linear regression and KNN regression—against conventional CNNs and Geometric Deep Learning models (Section 4.1)."

Although CNNs slightly outperform solid harmonic wavelets in simple binary classification tasks (Section 4.1), solid harmonic wavelet scattering features substantially linearise the classification problem. p. 7

Where the paper says it

15M Hybrid (Morlet) 78. 01 172k Table 1 shows that wavelet scattering features substantially linearise the classification problem.

p. 7
Where the paper says it

15M Hybrid (Morlet) 78. 01 172k Table 1 shows that wavelet scattering features substantially linearise the classification problem.

p. 7
Where the paper says it

15M Hybrid (Morlet) 78. 01 172k Table 1 shows that wavelet scattering features substantially linearise the classification problem.

The main result we wish to pursue is regression on merger mass ratio and time since merger, where the interpretable representations provided by solid harmonic wavelets become essential.

We demonstrate that the two labels are effectively independent, so the two regression tasks can be treated as separate.

The representation is fixed, so we can remove a component, retrain the regressor and keep every other part of the analysis unchanged.

This paper is 2611.14630.

--- Page 1 ---

Galaxy mergers play an important role in galaxy evolution; they enhance star formation, transform morphologies, and may drive accretion onto black holes

Pruning redundant reasoning tokens

Galaxy mergers play an important role in galaxy evolution; they enhance star formation, transform morphologies, and may drive accretion onto black holes.

To create a merger sample, we select mergers with a wide range of mass ratios and time since merger from IllustrisTNG following methodology laid out in Avirett-Mackenzie et al. (2024). The selected merger sample has a mass ratio (mu) range of 1:1 to 1:100 and time since merger (Δt) up to 4 Gyr. p. 2

Where the paper says it

We select only completed mergers, rather than pre-mergers. The selected merger sample has a mass ratio (𝜇) range of 1:1 to 1:100 and time since merger (Δt) up to 4 Gyr, distributions are shown in Figure 2.

The solid harmonic wavelet bispectrum captures quadratic interactions between morphological structures, while the angular frequency constraint causes rotational phase factors to cancel exactly, preserving invariance without discarding phase information. The resulting operator, the solid harmonic wavelet scattering bispectrum (SHWSB), is entirely determined before any training data is seen. p. 3

Where the paper says it

Taking the bispectrum of scattering responses — the third order of spectral analysis — captures quadratic interactions between morphological structures, while the angular frequency constraint causes rotational phase factors to cancel exactly, preserving invariance without discarding phase information. The resulting operator, the solid harmonic wavelet scattering bispectrum (SHWSB), is entirely determined before any training data is seen.

The SHWSBic features with k-nearest-neighbour regression perform extremely well for small sample sizes, due to the minimal number of trainable parameters required. The model reaches a test MAE of 0. p. 8

Where the paper says it

Noise resistance of solid harmonic wavelet bispectrum. Mean and standard deviation of the mean absolute error (MAE), averaged across ten train-test splits, as a function of exposure time 15 × 2 𝑖 s for 𝑖 = 0, .

414 ± 0.006 in log µ and 0. p. 2
Where the paper says it

Example idealised and realistic galaxy images used throughout the paper. Leftmost image: idealised sample from IllustrisTNG simulation cropped to 32 half stellar mass radius, and log normalised to [0, 1].

842 ± 0.023 Gyr in Δt on idealised images. p. 7
Where the paper says it

Images on a log scale over AB magnitude. Model Log 𝜇 Δt (Gyr) MAE RMSE MAE RMSE Random predictor 0.

In this work, we performed regression on continuous merger features (both time since merger and mass ratio) from morphological image data, with applications in large-scale astronomical surveys using the solid harmonic wavelet bispectrum. We demonstrate the strengths of these methodologies in data-scarce regimes. p. 6

Where the paper says it

Comparatively, for smaller training set sizes, a much simpler architecture with low angular frequency is required to prevent overfitting. H-Nets serve as a good benchmark for equivariant deep learning architectures, which leverage known symmetries to more efficiently learn patterns in the data compared to conventional CNNs, and thus typically perform better in smaller data regimes.

p. 6
Where the paper says it

Comparatively, for smaller training set sizes, a much simpler architecture with low angular frequency is required to prevent overfitting. H-Nets serve as a good benchmark for equivariant deep learning architectures, which leverage known symmetries to more efficiently learn patterns in the data compared to conventional CNNs, and thus typically perform better in smaller data regimes.

The solid harmonic wavelet bispectrum offers an appealing alternative to conventional deep learning techniques, with better performance especially at small sample sizes and a transparent functional methodology.

To analyse the performance of different approaches on galaxy classification, we select a sample of galaxy mergers from IllustrisTNG, with a wide range of galaxy mass ratios and time since merger (Section 2).

"We compare several wavelet-based methodologies — Morlet wavelets with linear regression, solid harmonic wavelets with linear regression, and the solid harmonic wavelet bispectrum with both linear regression and KNN regression—against conventional CNNs and Geometric Deep Learning models (Section 4.1)."

This paper is 2611.14630.

--- Page 2 ---

Brown et al.

To create a merger sample, we select mergers with a wide range of mass ratios and time since merger from IllustrisTNG following methodology laid out in Avirett-Mackenzie et al. (2024).

The solid harmonic wavelet bispectrum captures quadratic interactions between morphological structures, while the angular frequency constraint causes rotational phase factors to cancel exactly, preserving invariance without discarding phase information.

The SHWSBic features with k-nearest-neighbour regression perform extremely well for small sample sizes, due to the minimal number of trainable parameters required.

In this work, we performed regression on continuous merger features (both time since merger and mass ratio) from morphological image data, with applications in large-scale astronomical surveys using the solid harmonic wavelet bispectrum. We demonstrate the strengths of these methodologies in data-scarce regimes.

This paper is 2611.14630.

--- Page 3 ---

Regressing physical merger properties 3

--- Page 4 ---

Brown et al.

--- Page 5 ---

Regressing physical merger properties 5

The Solid Harmonic Wavelet Scattering (SHWS) transform cascades these wavelet transforms with modulus nonlinearities, analogously to Morlet scattering (Bruna & Mallat 2012) but with built-in rotational structure. The bispectrum of solid harmonic wavelet scattering coefficients is therefore defined as the triple-product of intermediate scattering responses with respect to angular frequency. p. 4

Where the paper says it

While SHWS produces roto-translation invariant features and linearises deformations, it does so by discarding complex phase information which can carry discriminative structural information — which SHWSB retains. The bispectrum of solid harmonic wavelet scattering coefficients is therefore defined as the triple-product of intermediate scattering responses with respect to angular frequency.

The SHWSBic is derived by normalising with the product of the scattering magnitudes, which isolates phase alignment from amplitude, and SHWSBic inherits the same [0, 1] range and covariance properties.

SHWSBic is less sensitive to contrast variations and better suited for discriminative tasks because it quantifies the alignment between rotational symmetries in galaxy morphology. The SHWSBic inherits the covariance properties of the bispectrum. p. 4

Where the paper says it

As with the non-scattering case, the bicoherence inherits the covariance properties of the bispectrum. In our regression experiments, we use SHWSBic as it empirically yields better performance.

In our regression experiments, we use SHWSBic as it empirically yields better performance. The results can be summarized as follows:

"Table 2 shows that models trained on our representations effectively regress merger features in idealised samples without observational noise or sky addition with a typical error of ∼ 0.3 dex over 2 orders of magnitude in merger mass ratio. p. 7

Where the paper says it

Table 2 shows that models trained on our representations effectively regress merger features in idealised samples without observational noise or sky addition with a typical error of ∼ 0. 3 dex over 2 orders of magnitude in merger mass ratio.

"

The solid harmonic wavelet bispectrum offers an appealing alternative to conventional deep learning techniques, with better performance especially at small sample sizes and a transparent functional methodology.

--- Page 6 ---

Brown et al.

The bispectrum of solid harmonic wavelet scattering coefficients is therefore defined as the triple-product of intermediate scattering responses with respect to angular frequency.

We extract features from Equations 13–16 by applying Lp pooling on the spatial coordinates of the representations, which we denote with the parameter p as: SHWSBic[x, j1, l2] = (15).

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. Reduced parameter count for feature extraction allows for deployment in data-scarce regimes where conventional deep learning architectures struggle with limited training samples. The solid harmonic wavelet bispectrum features require only 2520 trained parameters to perform regression, outperforming a CNN with 8,294,849 parameters in the same task.

  2. Regression models can predict both merger mass ratio and time since merger simultaneously, providing physical merger properties rather than just binary classification labels. This capability is demonstrated by showing that the method can regress on both merger mass ratio and time since merger simultaneously.

  3. Improved robustness to noise enables reliable performance on realistic astronomical data, even with limited exposure times. The bispectral analysis is particularly powerful in identifying non-Gaussianity and therefore this noise is essentially nullified by the bispectrum over wavelet scattering coefficients.

  4. Enhanced interpretability allows for the decomposition of predictions into contributions from specific morphological structures at different scales and angular frequencies. The model's prediction decomposes exactly into contributions of these named components, making it possible to test which of them carry the information.

Abstract

Galaxy mergers play an important role in galaxy evolution; they enhance star formation, transform morphologies, and may drive accretion onto black holes. Due to their complex morphological features, identification of mergers poses a significant challenge. Convolutional neural networks (CNNs) have been successful at identifying mergers with good accuracy. However, similar to other deep learning techniques such as Vision Transformers, they require large training samples. Current methods are also limited to classification, rather than regression. Here, we apply the solid harmonic wavelet scattering bispectrum, a method that encodes invariances and symmetries in the data which need to be learned by CNNs, as an alternative to CNNs. We apply our methodology to a simulated dataset of galaxies with a wide range of merger properties. We show that this method outperforms standard methodologies, while requiring fewer training samples and being robust to noise. Most importantly, we show that for the first time this method is able to regress on both merger mass ratio and time since merger simultaneously, with uncertainties 0.33 dex uncertainty in merger mass ratio and 0.66 Gyr uncertainty in time since merger. We have therefore shown that solid harmonic wavelet bispectrum presents an alternative to currently used machine learning techniques.

Sources

Related papers