Dingo-Pop: GW & H I Population Inference

• Dingo-Pop is a simulation-based framework that provides direct population inference for gravitational-wave catalogs via neural posterior estimation using transformers.
• It scales efficiently to large catalogs (25–1000 events) by bypassing per-event parameter estimation, reducing Monte Carlo noise and accelerating inference.
• In H I studies, Dingo-Pop describes a survey-driven census of neutral hydrogen in dark-matter halos, highlighting distinct observational applications.

Dingo‑Pop denotes a simulation‑based framework for gravitational‑wave population inference that maps catalogs of gravitational‑wave strain data directly to posteriors over population hyperparameters, using a pre‑trained Dingo single‑event encoder, a transformer encoder over event embeddings, and a normalizing flow (Leyde et al., 11 May 2026). In a distinct and informal usage, the term also appears in nearby‑Universe H I studies as a shorthand for a population‑level view of where neutral atomic hydrogen resides in dark‑matter halos, derived from DINGO, GAMA, and WAVES observations (Dev et al., 29 Apr 2026). The term should therefore be interpreted contextually: in gravitational‑wave inference it is the name of a specific end‑to‑end method, whereas in H I halo studies it functions as a descriptive label rather than a separately defined algorithmic system (Leyde et al., 11 May 2026).

1. Terminology and disambiguation

The formal usage of Dingo‑Pop is introduced in "End-to-End Population Inference from Gravitational-Wave Strain using Transformers" (Leyde et al., 11 May 2026). There, Dingo‑Pop is a framework for gravitational‑wave population inference that operates end‑to‑end on gravitational‑wave strain data and is designed to infer properties of the compact‑binary population, including the black‑hole mass spectrum and cosmological parameters such as the Hubble constant $H_0$, without traditional per‑event parameter estimation, explicit per‑event posterior reweighting, or on‑the‑fly numerical integration of selection functions.

A separate paper on the DINGO pilot H I survey states that it provides a population‑level “Dingo‑Pop” view of where H I lives in dark‑matter halos in the nearby Universe and how it is split between centrals and satellites (Dev et al., 29 Apr 2026). In that context, the expression denotes an observational census rather than a standalone software or inference architecture.

A further source of ambiguity is the 2025 diffusion‑LLM paper "DINGO: Constrained Inference for Diffusion LLMs" (Suresh et al., 29 May 2025). That paper explicitly states that “Dingo‑Pop” or “DINGO‑Pop” does not appear anywhere in the paper, and defines only “DINGO” as a dynamic‑programming‑based constrained decoding algorithm for diffusion LLMs. This makes clear that Dingo‑Pop is not a named variant of the diffusion‑LLM decoding method (Suresh et al., 29 May 2025).

Usage	Meaning
Dingo‑Pop	End‑to‑end GW population inference framework from strain (Leyde et al., 11 May 2026)
“Dingo‑Pop” in H I survey context	Population‑level view of halo H I in DINGO/GAMA/WAVES (Dev et al., 29 Apr 2026)
DINGO	Constrained decoding algorithm for diffusion LLMs; not called Dingo‑Pop (Suresh et al., 29 May 2025)

This terminological split is important because the gravitational‑wave usage names a concrete methodological contribution, while the H I usage summarizes an empirical population picture.

2. End‑to‑end gravitational‑wave population inference

In the gravitational‑wave literature, Dingo‑Pop is motivated by the standard hierarchical Bayesian analysis framework. For a population model $p_\text{pop}(\theta\mid\Lambda)$, event parameters $\theta$, and hyperparameters $\Lambda$, the usual population likelihood for $N$ events with data $\{D_i\}$ is written as

$$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$

The numerator marginalizes the single‑event likelihood against the population, and the denominator accounts for selection effects via $p_\mathrm{det}(\theta)$ (Leyde et al., 11 May 2026).

The framework is introduced as a response to the scaling limitations of conventional hierarchical Bayesian analysis. As catalogs grow to hundreds or thousands of events, the Monte Carlo variance in the log‑likelihood can scale as badly as $\sim N^2$, and both numerator and denominator are conventionally evaluated using Monte Carlo methods: reweighting single‑event parameter‑estimation samples for the numerator, and injection campaigns plus importance sampling for the denominator. Dingo‑Pop addresses this by directly learning the mapping

$$\{D_i\}_{i=1}^N \;\longrightarrow\; p(\Lambda \mid \{D_i\})$$

through neural posterior estimation, training on simulated catalogs generated from the forward population model including selection effects (Leyde et al., 11 May 2026).

The central conceptual features are stated explicitly. The pipeline is end‑to‑end, from strain to population posterior in a single differentiable pipeline. It requires no per‑event PE at inference time, uses a single network that is size‑agnostic over catalogs with $p_\text{pop}(\theta\mid\Lambda)$0, and after training produces a full population posterior in $p_\text{pop}(\theta\mid\Lambda)$1 s (Leyde et al., 11 May 2026). The target posterior is approximated directly as

$p_\text{pop}(\theta\mid\Lambda)$2

with selection effects represented implicitly by the training distribution rather than explicitly recomputed at inference time.

This design removes per‑analysis Monte Carlo noise from inference itself. The Monte Carlo burden is displaced to the training set construction, where the network is exposed to catalogs drawn from the full hierarchical generative process. A plausible implication is that Dingo‑Pop is best understood not as a faster likelihood evaluator, but as an amortized posterior approximator trained to absorb the full selection‑aware generative structure into a parametric inference map.

3. Architecture, data representation, and selection effects

The Dingo‑Pop inference pipeline has three stages (Leyde et al., 11 May 2026). First, each event’s data $p_\text{pop}(\theta\mid\Lambda)$3—gravitational‑wave strain time series in multiple detectors together with corresponding power spectral densities—are processed by a pre‑trained Dingo encoder: $p_\text{pop}(\theta\mid\Lambda)$4 where $p_\text{pop}(\theta\mid\Lambda)$5. In the reported implementation, the encoder is specialized to detector‑frame component masses and luminosity distance, and is frozen when training Dingo‑Pop.

Second, the set of event embeddings is transformed into catalog tokens. A residual fully connected tokenizer maps each 32‑dimensional embedding to a 1024‑dimensional token, a learnable summary cls‑token is prepended, and a 10‑layer transformer encoder with 8‑head self‑attention and embedding dimension 1024 processes the sequence. No positional encoding is used, so the architecture is permutation‑equivariant to event ordering. The final cls‑token defines a single 1024‑dimensional summary vector,

$p_\text{pop}(\theta\mid\Lambda)$6

which is then mapped by a residual fully connected network with 5 blocks and 1024 hidden units to a 512‑dimensional conditioning vector for the posterior model (Leyde et al., 11 May 2026).

Third, a neural spline flow with 14 steps, rational quadratic splines with 8 bins, and a 9‑dimensional latent space models the posterior over the nine hyperparameters $p_\text{pop}(\theta\mid\Lambda)$7: $p_\text{pop}(\theta\mid\Lambda)$8 At inference time, strain data are encoded eventwise, aggregated by the transformer, and sampled through the conditional flow.

Variable catalog size is handled directly by the transformer. The prior over catalog size during training is uniform,

$p_\text{pop}(\theta\mid\Lambda)$9

and the absence of positional encodings ensures that the final summary is permutation‑invariant with respect to the set of events (Leyde et al., 11 May 2026). This matches the invariance of the population posterior to event ordering.

Selection effects are incorporated during training through two auxiliary neural networks rather than brute‑force simulation of large numbers of undetected events. The first is a detection probability estimator for

$\theta$0

marginalized further to a function of detector‑frame component masses and luminosity distance. It is trained on $\theta$1 simulated events with Gaussian detector noise, with detection defined by a matched‑filter SNR threshold of 12 and optimization by binary cross‑entropy (Leyde et al., 11 May 2026). The second is an embedding emulator, a conditional normalizing flow trained to emulate $\theta$2 for detected events. During training this permits direct sampling of embeddings $\theta$3 without generating raw strain or re‑running the Dingo encoder. In validation and comparison against hierarchical Bayesian analysis, the emulator is not used; full waveforms plus noise are generated and passed through the true encoder.

The training objective is the negative log‑likelihood

$\theta$4

implemented as the average NLL over a batch of populations (Leyde et al., 11 May 2026). Because the training distribution includes explicit simulation of detection and catalog formation, the learned amortized posterior is intended to approximate the exact hierarchical posterior implied by the underlying forward model, including selection effects.

4. Population model, training regime, and empirical performance

The reported population model combines a Power Law + Peak mass spectrum and flat $\theta$5CDM cosmology with fixed $\theta$6 and free $\theta$7 (Leyde et al., 11 May 2026). The nine hyperparameters are

$\theta$8

The prior ranges are given explicitly: $\theta$9 km s$\Lambda$0 Mpc$\Lambda$1, $\Lambda$2, $\Lambda$3, $\Lambda$4, $\Lambda$5, $\Lambda$6, $\Lambda$7, and $\Lambda$8 (Leyde et al., 11 May 2026).

Single‑event simulations use IMRPhenomXPHM waveforms in stationary Gaussian noise with O3 sensitivity for Advanced LIGO Hanford and Livingston in a two‑detector HL configuration, with network matched‑filter SNR $\Lambda$9 as the detection threshold (Leyde et al., 11 May 2026). The Dingo encoder is trained over detector‑frame component masses $N$0, luminosity distance $N$1 Mpc, spin magnitudes in $N$2 with isotropic tilts, and isotropic sky positions and orientations. The population priors are chosen to remain within this encoder domain.

Training of the main Dingo‑Pop network uses AdamW with initial learning rate $N$3, weight decay 0.01, cosine annealing from $N$4 to 0 over 800 epochs, 128 populations per batch, and 50,000 populations per epoch. Total training is 800 epochs, corresponding to approximately $N$5 population draws, on a single NVIDIA A100 GPU, with wall‑time of about 11 days (Leyde et al., 11 May 2026). Inference cost is reported as $N$6 s for generating 5000 posterior samples for a single catalog on GPU.

Calibration is assessed through probability–probability plots on 2500 catalogs with $N$7 and 5000 posterior samples per catalog. The P–P curves for all nine hyperparameters lie within 1–3$N$8 bands around the diagonal, with combined Kolmogorov–Smirnov p‑value $N$9 (Leyde et al., 11 May 2026). The supplementary results show slight degradation for some hyperparameters, such as $\{D_i\}$0, near $\{D_i\}$1, which the authors connect to the upper edge of the training range and the tightening of posteriors.

Comparison against conventional hierarchical Bayesian analysis is performed on two simulated populations of 500 events each. The reference pipeline uses Dingo and Bilby for single‑event parameter estimation, icarogw for population analysis, and injection‑based selection estimation with $\{D_i\}$2 and $\{D_i\}$3 detected injections in two runs (Leyde et al., 11 May 2026). The reported medians and 90% credible intervals for each hyperparameter show close agreement between “SBI” and “HBA”. For Population 2, for example, the true value $\{D_i\}$4 yields SBI $\{D_i\}$5 and HBA $\{D_i\}$6, while true $\{D_i\}$7 yields SBI $\{D_i\}$8 and HBA $\{D_i\}$9 (Leyde et al., 11 May 2026).

The comparison also shows that discrepancies between two HBA runs with different injection sets are comparable to the differences between HBA and Dingo‑Pop, indicating that Monte Carlo uncertainty in the selection function dominates many of the observed differences. Some Dingo‑Pop posteriors are described as slightly broader, especially in some shape parameters, consistent with potential information loss in the 32‑dimensional embedding and/or conservative learning (Leyde et al., 11 May 2026).

5. Spectral‑siren cosmology, advantages, and limitations

A principal application of Dingo‑Pop is spectral‑siren inference of the Hubble constant. In this setting, cosmology enters only through $$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$0 in flat $$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$1CDM with fixed $$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$2, while the observed detector‑frame mass distribution is stretched by $$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$3 and therefore carries information about redshift and $$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$4 (Leyde et al., 11 May 2026). Dingo‑Pop returns the joint posterior over both mass‑spectrum hyperparameters and $$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$5 directly from the event catalog.

The paper reports a scaling study in which 128 independent populations are generated, events are added incrementally from 1 up to 1000, and the relative uncertainty on $$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$6 is measured as the 2‑$$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$7 posterior width divided by the posterior median. The median relative uncertainty is approximately $$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$8 at $$p(\{D_i\}_{i=1}^{N}\mid\Lambda) = \prod_{i=1}^{N} \frac{\int p(D_i\mid \theta) p_\text{pop}(\theta \mid \Lambda)\,\mathrm{d}\theta}{\int p_\text{det}(\theta)\,p_\text{pop}(\theta\mid \Lambda)\, \mathrm{d}\theta}.$$9 and approximately $p_\mathrm{det}(\theta)$0 at $p_\mathrm{det}(\theta)$1, with median scaling between 500 and 1000 events of

$p_\mathrm{det}(\theta)$2

This is described as somewhat shallower than $p_\mathrm{det}(\theta)$3, with selection biases, the role of other population parameters, and degeneracies cited as the expected reasons (Leyde et al., 11 May 2026).

The main advantages emphasized for Dingo‑Pop are the elimination of per‑event Monte Carlo noise at inference time, amortized inference across catalog sizes from 25 to 1000 events, and end‑to‑end speed of about a second once the encoder and population model are trained (Leyde et al., 11 May 2026). The framework is also presented as providing a direct route to training on astrophysical simulations rather than exclusively analytic parametric forms of $p_\mathrm{det}(\theta)$4.

Its limitations are equally explicit. Dingo‑Pop depends on the training distribution: waveform model IMRPhenomXPHM, Gaussian noise, SNR$p_\mathrm{det}(\theta)$5 selection, and the parametric PL+Peak mass model with flat‑$p_\mathrm{det}(\theta)$6CDM cosmology. If the true population or detection process differs substantially, the amortized posterior may be biased through model misspecification (Leyde et al., 11 May 2026). The Dingo encoder also has a limited domain in masses, distances, detectors, and data conditioning; out‑of‑domain application would require retraining or extension. The current detection probability estimator is trained on synthetic SNR thresholding rather than real LVK pipeline injections, and the framework inherits waveform systematics and noise assumptions from the simulations.

The authors also discuss robustness to out‑of‑distribution structure by manually introducing a gap in the primary mass distribution. In that test, Dingo‑Pop posteriors broadly track those of HBA, though with heavier tails in some parameters (Leyde et al., 11 May 2026). This suggests partial robustness rather than formal misspecification immunity. Future directions named explicitly include training on real LVK selection functions, extending the encoder, including more complex population models such as spins, eccentricities, redshift evolution, formation channels, and non‑parametric mass spectra, handling correlated events and multi‑messenger data, and replacing full attention with sparse or linear alternatives for the $p_\mathrm{det}(\theta)$7–$p_\mathrm{det}(\theta)$8 event catalogs expected from third‑generation detectors (Leyde et al., 11 May 2026).

6. Alternate usage in H I halo studies

In observational extragalactic astronomy, the phrase “Dingo‑Pop” is used in the DINGO/GAMA/WAVES study as a label for a population‑level view of the H I content of halos in the nearby Universe (Dev et al., 29 Apr 2026). The analysis combines ASKAP DINGO pilot 100‑hour H I data, GAMA spectroscopy, and WAVES photometric data to measure the H I–halo mass relation over

$p_\mathrm{det}(\theta)$9

using direct detections and spectral stacking.

That study finds that the H I–halo mass relation exhibits a double power‑law form with turnover near $\sim N^2$0, that central galaxies dominate the halo H I budget below $\sim N^2$1, and that satellites dominate at higher halo masses (Dev et al., 29 Apr 2026). Including WAVES photometric members increases the measured H I content in halos above $\sim N^2$2 by a factor of 1.5–3, which the paper attributes to gas‑rich satellites that fall below the spectroscopic completeness limit of GAMA. The comparison with previous group‑stacking studies indicates that low‑surface‑brightness galaxies and intra‑group H I structures contribute only a minor fraction to the total halo H I mass (Dev et al., 29 Apr 2026).

Here Dingo‑Pop is not a neural architecture or an amortized inference system. It denotes, instead, a survey‑driven population census of where cold gas resides across halo mass and between centrals and satellites. A plausible implication is that the term has acquired a broader descriptive life around the DINGO survey name, but only the gravitational‑wave usage defines Dingo‑Pop as a formal method.

7. Position in the literature

Dingo‑Pop in the gravitational‑wave sense is situated at the intersection of neural posterior estimation for gravitational waves, simulation‑based inference for populations, and transformer architectures for set‑valued data (Leyde et al., 11 May 2026). The framework builds on Dingo for single‑event neural posterior estimation from strain, on earlier work using flows for parameter estimation and hierarchical analysis, and on DeepSets and set transformers for permutation‑invariant processing of unordered collections.

The novelty claims are threefold. First, it is presented as the first end‑to‑end gravitational‑wave population inference method directly from strain using transformers, without intermediate per‑event posteriors at inference time. Second, it is amortized over variable catalog sizes from 25 to 1000 events in a single network. Third, it eliminates per‑analysis Monte Carlo noise in both numerator and denominator of the hierarchical likelihood by learning the posterior over the full simulated training measure (Leyde et al., 11 May 2026).

This positions Dingo‑Pop as a methodological response to the computational and statistical bottlenecks of classical hierarchical Bayesian analysis for large catalogs. At the same time, the explicit caveats about model misspecification, encoder domain, and selection realism indicate that its claims are bounded by the fidelity of the simulator and training distribution rather than by likelihood exactness in the conventional sense (Leyde et al., 11 May 2026). In the H I literature, by contrast, the phrase marks a population‑level observational synthesis rather than a new inferential formalism (Dev et al., 29 Apr 2026).

Taken together, these usages show that Dingo‑Pop has become a context‑dependent term spanning two astrophysical domains. In gravitational‑wave astronomy it names a specific end‑to‑end amortized posterior estimator from strain; in H I halo studies it denotes a population census of neutral gas in halos. The two are connected only by nomenclature, not by method or scientific target (Leyde et al., 11 May 2026).