GoKawiilDetails of observations
Our observations, from which we inferred S301, comprise 19 datasets spanning more than 8 years. We first noted S301 in 2023, when pointing directly to Sgr A*. Given the position of the star close to Sgr A* and its large proper motion, it was clear that it might be on a tight orbit, and we followed up with dedicated observing campaigns over the course of the next years. We summarize the data used for this work in Extended Data Table 1.
Based on the orbital coverage from 2023 to 2025, we were able to trace S301 back in time and identified two previously acquired datasets in 2017 and 2021, in which the star should be present. For this, we fit a preliminary orbit to the 2023–2025 data and used samples from a Markov chain to predict where the star should be in 2021. The dataset in 2021 was acquired during an observation that involved multiple other pointings in the GC to create a mosaic of the central about 200 × 200 mas (ref. 22). This particular pointing was affected by bad seeing conditions, however, and we apply stricter cuts and accept as a post-processing step only fringe-tracking ratios of 90% for the individual detector integrations, effectively discarding data with low coherent flux. This procedure led to the inference of S301 in this particular epoch, close to the predicted position (Extended Data Fig. 3, left).
With that position in hand, we repeated the orbit fit and used the updated Markov chain to predict the S301 position in 2017. The 2017 data were acquired as part of the monitoring of the motion of the star S2, an object around five magnitudes brighter than S301. The dataset consists of individual exposures across five consecutive nights in March 2017, partially recorded in split-polarization mode. In the image reconstruction of this dataset, we treat both polarizations in polarized exposures as individual measurements and combine them with the unpolarized data. Again, S301 is found close to the predicted place (Extended Data Fig. 3, right).
Highest-resolution, deep images of the central 800 au
The high-angular-resolution, high-fidelity images of the GC are reconstructed from GRAVITY data, an example of which is shown in Extended Data Fig. 1 with the image reconstruction tool GRAVITY-RESOLVE (GR)22. This code is designed to reconstruct stars in the GC that appear to GRAVITY as unresolved point sources and, in particular, to find faint, yet undiscovered stars. GR is based on a hierarchical forward model that incorporates the instrument response of GRAVITY, including both optical aberrations and the spectral transmission within the beam combiner, as well as a statistical model of the GC. The intrinsic multitude of degrees of freedom of an image is tamed with Bayesian inference, in which the statistical model of the GC provides the previous information. With GR, we exploit supreme imaging resolution of ≃ 1.7 mas of GRAVITY and its phase-referencing abilities, which allow for high contrast images and routine mosaicking as pioneered in radio interferometry (F. Mang et al., manuscript in preparation).
The individual positions of S301 over time, as shown in Fig. 2, were inferred in a first step from imaging. We initialized the model in GR by invoking all known, bright sources within the field of view. Positions and fluxes of the stars and Sgr A* are constrained by previous distributions that reflect current knowledge. We applied GR a total of 10 times to the same dataset with varying initial random seeds. Tentative faint sources that may appear in the image of individual reconstructions are accepted only when they are inferred in at least 5 out of 10 reconstructions. Their corresponding positions in the image grid are then referenced to Sgr A* and subsequently averaged.
Astrometric errors
The uncertainties of the newly inferred sources are derived from the scatter over the different runs, also taking into account the finite resolution of the pixel grid. This is an improvement over22 where the errors were simply approximated by the size of the pixels in the image. If the same faint source is inferred in the same pixel for the 10 individual runs, the corresponding standard deviation is zero, albeit unphysical. Hence, the discretized position space needs to be considered when stating errors on astrometry. Instead of quantifying the error based on the number of same-pixel inferences in a set of detections, we opt for a general approach, which may be conservative, but prevents an underestimation of errors. We derive a discretization error by calculating the root mean square error for a single pixel within the image for both directions, right ascension and declination, independently. Considering a pixel size of 0.8 mas per pixel, this evaluates to about 207 μas and represents the statistical uncertainty in astrometry originating from the pixel grid in GR images. This value acts effectively as a noise floor.
Fitting of GRAVITY data
... continue reading