21,864 Unresolved, Low-mass Binaries Identified via their Overluminosity in \textit{Gaia} DR3 and a Catalog of 347,440 Systems within 100 pc of the Sun

Abstract

The fundamental parameters of a low-mass star can potentially be determined from its photometry and astrometry. This is complicated by the fact that 10-20 percent of low-mass stars are predicted to be equal-mass binaries. These unresolved systems appear more luminous compared to single stars with the same fundamental parameters. We present a method to differentiate binary stars from single-star main sequence K and M dwarfs using their \textit{Gaia} DR3 XP spectra. We assemble a training set of stars which have pristine astrometry and photometry, are located within 100pc of the Sun, and exclude stars with \textit{Gaia} DR3 flags suggesting they may be unequal mass systems, thereby leaving stars that are predominantly either single- or equal-mass binaries. We then iteratively train Random Forest Regression (RFR) models to predict absolute magnitude and color given the RP spectral coefficients of a star. After each model, we remove the stars that have absolute magnitudes significantly brighter than their predicted values. This method converges on a model trained only on single stars. We then use this model to identify the ``overluminous'' K and M stars in \textit{Gaia} DR3 within 100 parsecs, with some quality cuts. We find that $\sim13\%$ of the sample is significantly overluminous and assume these to be unresolved binaries. We aggregate several multiplicity surveys across different projected separations and incorporate our overluminous binaries to create a general \textit{Catalog of Systems} within 100 pc. We use this \textit{Catalog} to provide lower limits on the multiplicity fraction for stars between $0.1$ and $0.7~M_{\odot}$.

Figures (16)

• Figure 1: The strong relationship between the optical photometry and spectra for low-mass stars. Left panel: the dereddened, Gaia color-magnitude diagram for our stellar sample defined in §\ref{['sec:data']}. Two examples, a metal-rich dwarf (PM J08202+0532) and a metal-poor subdwarf (G 192-52), are pulled from this distribution. Right panels: the low-resolution, XP spectrum (red) and a medium-resolution, LAMOST DR9 (gray) spectra for these examples. Their location on the CMD is shown and arrows point to their spectra on the right panels. Labels above the spectra identify the strongest molecular features in the SED showing that there are drastic changes for different metallicities. This change is strong enough to effect the photometry.
• Figure 2: The results of different RUWE cuts for low-mass stars from a simulated set of 200,000 binaries. The predicted RUWE parameters were generated using the GaiaUnlimitedempirical_GUsubsample_GU package. A linear interpolation was used to relate mass and absolute magnitude from the_bible. Left panels: the top panel displays the mass ratio distribution of our simulation, as found by elbadry_twin. The resulting mass ration distribution from two RUWE limits, $1.4$ in orange and $1.2$ in green, are also shown. The bottom panel shows the fraction of stars removed at a given mass ratio for different distance limits using the constraint RUWE$<1.2$ which we adopt in §\ref{['sec:data']}. Right panels: the distribution (top panel) and fraction of stars removed (bottom panel) for a given primary flux fraction for the full sample and the stars below the RUWE thresholds. The primary flux fraction is defined as the amount of flux emitted from the system which comes from the brigther star. The RUWE cuts are less effective at removing the equal-mass systems ($\sim0.5$) and effectively single stars ($\sim1.0$).
• Figure 3: The winnowing from all stars within 100 pc to the Regression Sample in four parts. Top left: the CMD of all low-mass stars in Gaia with a parallax $\varpi>10~mas$. Top right: the resulting sample given our ADQL query described in §\ref{['sec:data']}. It can be seen that there are still issues with the photometry, especially in the lowest-mass stars. Bottom left: the distribution of the corrected phot_bp_rp_excess_factor, which we represent as $C^*$riello_photometry. The red line at $log_{10}|C^*| = -1.4$ represents our cut for stars with poor photometry. Bottom right: the Gaia CMD with the high $C^*$ sources removed, representing the final Regression Sample of 127,843 stars.
• Figure 4: A comparison between the predictions from our first HistGradientBoostingRegressor model and the measured values from Gaia. Left panel: the precision of the predicted color compared to the measured values. The low scatter shows that we recover this feature well. Middle panel: the precision of the absolute magnitude prediction. Unlike the color, there is a cloud of stars hovering above the central distribution. These systems have a lower absolute magnitude and are therefore brighter than their prediction: the overluminous binary systems. We identify these systems with the dashed red line which lies at 3 standard deviations above zero. Right panel: the overluminous sources are highlighted in red on a Gaia CMD.
• Figure 5: Progress of our iterative method shown in two aspects. Left panel: number of stars flagged as "overluminous" and removed at each iteration. Right panel: absolute error in the predicted absolute magnitude after each iteration. The error for each iteration's test sample is shown in black and the error for the full sample (including the overluminous systems) is shown in orange. We can see that there are progressive gains but with diminishing return. We note the elbow in the total sample's error, where we identify that each iteration is no longer removing binaries. We mark the 40th iteration as the breaking point by eye and use this as our terminal model.