Evidence for the First Globular Cluster Stellar Stream beyond the Milky Way

This paper presents evidence for the first extragalactic globular cluster stellar stream, identified in deep Hubble Space Telescope imaging of the ultra-diffuse galaxy UGC9050-Dw1. The stream, named Oyashio after a cold Pacific ocean current, is located at a projected distance of approximately 2.5 kpc from the center of UGC9050-Dw1, which is likely associated with the low-surface brightness spiral galaxy UGC 9050 at a distance of 35.2 ± 2.5 Mpc.

The stream candidate was identified independently in both HST and CFHT data, ruling out imaging or data processing artefacts. The signal prominence in the combined HST image is (fstream − f̄background)/σbackground = 7.34, with a measured width of w±σ = 72.3 ± 8.9 pc assuming a Gaussian profile. The amplitude of the fit gives a signal-to-noise ratio of A/√σA = 5.2 for the HST image, with similar fits to CFHT images giving ratios between 2.3 and 3.9 in the g, r, and i bands, while the feature is not detectable in the u and z bands. The length of the stream arm is approximately 2 kpc based on a by-eye estimate.

The width of Oyashio is much smaller than any known stream from a dwarf galaxy progenitor (e.g., the Orphan-Cenab MW stream has a width > 200 pc), pointing towards a globular cluster origin. Milky Way globular cluster streams have widths ranging from a few tens to a few hundred pc. The stellar population of a GC progenitor and its stream should be of the same age and metallicity, following the same isochrone in a color-magnitude diagram. The stream and progenitor cluster candidates have overlapping colors (F555W-F814W = 1.0 ± 0.2 for the stream and 1.1 ± 0.1 for the GC candidate), both falling within the GC candidate selection box used by the previous analysis of this galaxy's GC population. These colors are slightly bluer than the average GC in the Milky Way, consistent with a more recent origin for the clusters in UGC9050-Dw1.

To test whether a GC stream with this morphology is dynamically plausible, the authors apply the X-Stream sampler, which translates stream imaging into constraints on stream progenitors and host dark matter halos. The sampler places an upper limit on the initial progenitor mass of Mprog < 2.5 × 10 6 M⊙ with 95% confidence, consistent with a globular cluster origin. This is less massive than the Milky Way star cluster ωCentauri (M∗ = 3.55 × 10 6 M⊙) and comparable to some GCs in M31.

The surface brightness of Oyashio is F555W = 27.0 ± 0.1 mag/arcsec2 and F814W = 26.1 ± 0.1 mag/arcsec2, which is brighter than the average Milky Way stream, as expected for an extragalactic discovery. Comparing the measured surface brightness to a simulated stellar population based on the archetypal Milky Way GC stream Palomar 5, the authors find that a stream with the same age, metallicity, and level of disruption requires a progenitor 20 times as massive (initial cluster mass of 2 × 10 6 M⊙) to produce a similar surface brightness. A younger progenitor could produce the observed surface brightness with a progenitor mass as low as 1.65 × 10 5 M⊙ for the brightest isochrone.

The X-Stream sampler finds a UGC9050-Dw1 halo scale mass of log10(Mhalo/M⊙) = 11.31+0.67(−0.71) within the 68% confidence limit, corresponding to log10(M200/M⊙) = 11.63+0.71(−0.83) (68%), and finds an inner density slope of γ = 0.92+0.57(−0.58) within the 68% confidence limits. This is the first constraint on a dark matter halo mass and density slope from a stellar stream in an ultra-diffuse galaxy. The results suggest this UDG is slightly less cored than the average low-surface brightness dwarf (γ ∼ 0.2) and than UDG Dragonfly 44 (γ = 0.3), but allows inner slopes comparable to that of UDG AGC 242019 (γ ∼ 0.54). The sampler also places strong constraints on the line-of-sight progenitor position and orbit, while the halo scale radius and outer density slope are unconstrained.

For the model stream with the halo mass at the peak of the posterior distribution, the progenitor's current galactocentric radius is Rgal,today = 2.52 kpc, with an inferred enclosed mass of M(< Rgal,today) = 1.36 × 10 10 M⊙. The total mass (M200 = 1.56 × 10 11 M⊙) is similar to mass estimates of the Large Magellanic Cloud and is within errors of previous halo mass estimates (1.5 ± 0.3 × 10 11 M⊙ or 1.8 ± 0.3 × 10 11 M⊙) that used GC counts. The tidal radius for the fit shown is rt = 133 pc at present day with a minimum of 95 pc, similar to the estimated tidal radius of Pal 5 (≈ 145 pc), confirming that tidal stripping is feasible.

The authors consider alternative explanations for the feature. Merger-induced tidal tails from GC or nuclear star cluster mergers are unlikely since those occur near the host centre, while this feature is >2 kpc from the luminosity centre. Tidal shells formed in radial collisions would have a centre of curvature offset from the host, which is not observed. Gravitational lensing of a background galaxy could create an arc, but no other arcs or plausible lenses are observed. A dusty region would show a reddening trend, which is not found. A stellar stream from a small dwarf galaxy might be misinterpreted as a GC stream, but the observational width measurement and dynamical mass constraints both point to a GC progenitor. Chance alignment of unresolved stars cannot be entirely ruled out, but the similarity in colours, agreement between modelling results, halo mass estimates from GC counts, and mock observations support a GC stream detection.

The authors note that the arm towards positive x-values wraps behind the brighter central part of the UDG itself, which can help explain why only one arm is detected in the HST data. Deeper observations with HST or JWST could distinguish the feature further from background, and spectroscopic studies with e.g. the Keck telescope could compare parts of the stream to the presumed parent cluster.

This work extends the reach of globular cluster stream analysis to external galaxies, providing a new tool to study ultra-diffuse galaxies. Cold dark matter models predict the formation of low-mass subhalos devoid of stars, and extending the sample of GC stellar streams to extragalactic hosts allows study of host galaxies with fewer baryonic perturbers than the Milky Way, increasing the chances of unambiguous subhalo detection. With the capabilities of the Euclid and Roman Space Telescopes, the authors expect to detect many more GC streams, with this discovery serving as a precursor to future detections and representing an independent way to probe the dark matter mass and density profiles of ultra-diffuse galaxies.

Improvements to AI Systems Based on This Paper:

1. AI for Automated Stream Detection in Low-Surface-Brightness Imaging

• Train a deep-learning model (e.g., convolutional neural network) on synthetic and real HST/CFHT images of ultra-diffuse galaxies to automatically identify faint, narrow, arc-like stellar streams (like Oyashio) with signal-to-noise ratios as low as 5.

• The improved system can flag candidate streams across large survey datasets (Euclid, Roman) without human bias, including those hidden behind brighter galactic regions, by learning to separate real astrophysical features from imaging artifacts, dust lanes, and gravitational arcs.

2. AI for Multi-Band Photometric Consistency Checks

• Develop a Bayesian classifier that combines multi-band surface brightness measurements (e.g., F555W, F814W, g, r, i, u, z) to validate stream candidates by requiring consistent colors and non-detection in bands where the model predicts low signal (as seen with Oyashio’s absence in u and z).

• The improved system can automatically reject false positives (e.g., dusty regions, lensed arcs) by cross-correlating color trends and spatial morphology, reducing manual follow-up time.

3. AI for Dynamical Stream Modeling and Halo Parameter Inference

• Integrate the X-Stream sampler’s outputs into a neural density estimator (e.g., normalizing flows) to accelerate posterior sampling of progenitor mass, halo scale mass, inner density slope, and orbital parameters from stream morphology.

• The improved system can, in real time, constrain dark matter halo properties (e.g., log10(Mhalo/M⊙) = 11.31, γ = 0.92) for any newly discovered extragalactic stream, enabling rapid population studies of ultra-diffuse galaxies without running expensive MCMC chains.

4. AI for Progenitor Classification (Globular Cluster vs. Dwarf Galaxy)

• Train a random forest or gradient-boosting model on synthetic streams with known progenitor masses, widths, and surface brightness profiles to classify observed streams as globular-cluster-origin (mass < 2.5 × 10 6 M⊙, width < 100 pc) versus dwarf-galaxy-origin (wider, more massive).

• The improved system can automatically estimate the probability that a stream originates from a GC, using width, amplitude, and color-matching to candidate clusters, as done for Oyashio (width 72 pc, color match to GC candidate).

5. AI for Mock Observation and Survey Strategy Optimization

• Use generative adversarial networks (GANs) to create realistic mock HST/JWST/Euclid images of GC streams around UDGs, varying host halo mass, stream age, metallicity, and distance.

• The improved system can predict detectability limits (e.g., surface brightness thresholds, required exposure times) and recommend optimal filter sets and observation strategies for future telescopes, maximizing the discovery rate of extragalactic streams.

6. AI for Anomaly Detection in Deep Imaging Data

• Implement an unsupervised anomaly detection algorithm (e.g., autoencoder) on pixel-level data from deep surveys to find low-surface-brightness features that deviate from smooth galaxy models, without prior assumptions about stream shape.

• The improved system can discover new classes of tidal features (streams, shells, tails) in UDGs and other faint galaxies, potentially revealing more subhalo-induced perturbations than current by-eye searches.

7. AI for Cross-Instrument Data Fusion

• Build a multi-modal AI that jointly analyzes HST and ground-based (CFHT) images, accounting for different point-spread functions, depths, and filter responses, to confirm faint features like Oyashio with higher confidence.

• The improved system can automatically co-register and combine datasets, outputting a unified significance map (e.g., combined S/N = 7.34) and reducing false positives from single-instrument artifacts.

8. AI for Tidal Radius and Disruption Level Prediction

• Train a regression model on N-body simulations of GC disruption to predict tidal radius and mass-loss rate from stream width, surface brightness, and host halo profile (as inferred for Oyashio: rt = 133 pc).

• The improved system can quickly estimate whether a stream is currently being tidally stripped and predict its future evolution, aiding in selecting targets for spectroscopic follow-up (e.g., Keck) to measure radial velocities.

The dark matter content of ultra-diffuse galaxies is the subject of considerable debate. Stellar streams, which form when a host galaxy tidally strips stars from an orbiting stellar system, provide a powerful technique to constrain the dark matter content of external galaxies. The stripped stars form long, thin leading and trailing tidal arms that persist for billions of years. Stellar streams from globular clusters are particularly sensitive probes of dark matter halos and substructure. Globular cluster streams are expected to exist in a variety of host galaxy types, but so far, they have only been observed in the Milky Way. We present evidence for the first extragalactic globular cluster stellar stream, identified in deep Hubble Space Telescope imaging of the ultra-diffuse galaxy, UGC9050-Dw1. The stream's morphology, colour, and apparent association with a compact source support the globular cluster progenitor interpretation observationally, and we reproduce the observed surface brightness with simulated globular cluster stellar populations. We use generative stream modelling, which fits dynamical models directly to the stream morphology, to constrain the mass of the progenitor and present the first stream-based halo constraint for an ultra-diffuse galaxy. The stream models point to a globular cluster origin and suggest a massive dark matter host halo. By extending the reach of globular cluster stream analysis to external galaxies, this work opens a new chapter in dark matter science.