About half of the Sun-like stars in the Galaxy have a companion. I work on binaries at both ends of the range in separation. The widest pairs, thousands of AU apart, are a clean place to test gravity where it is weakest. The closest pairs are too tight to resolve at all, but they still leave two sets of lines in a single spectrum, and large surveys contain tens of thousands of them.
Newton's law of gravity works beautifully until you look at systems where gravitational accelerations drop below about \(10^{-10}\ \mathrm{m\,s^{-2}}\). At that threshold, galaxies stop behaving the way Newton predicts, and the usual fix is to invoke dark matter. But there's another possibility: maybe gravity itself changes at very low accelerations. That idea is called Modified Newtonian Dynamics, or MOND.
The trouble with testing MOND has always been scale. Galaxy rotation curves are suggestive but messy - full of baryonic uncertainties and modeling assumptions. What you really want is a clean, local laboratory. Wide binary stars turn out to be exactly that.
A wide binary is two stars gravitationally bound to each other but separated by thousands of astronomical units, far enough apart that the gravitational acceleration between them drops into the MOND regime. If MOND is right, their relative orbital velocities should be systematically higher than what Newton predicts. If Newton (plus dark matter) is right, they shouldn't.
At Ohio State, I'm building the most precise observational test of this prediction. I combine Gaia astrometry, which gives exquisite proper motions and parallaxes, with high-resolution radial velocities from facilities like Keck, VLT, and Magellan. The radial velocity component is critical: Gaia gives you the sky-plane motion, but you need the line-of-sight velocity to reconstruct the full 3D relative velocity of each pair.
In Saad & Ting (2025), I introduced a technique to measure differential radial velocities of wide binaries from high-resolution, high signal-to-noise spectra, and applied it to 85 pairs from the C3PO survey. The method reaches a precision of 8 to 15 \(\mathrm{m\,s^{-1}}\) per pair, a median improvement of roughly \(24\times\) over Gaia DR3, finally bringing line-of-sight velocities to the accuracy this test demands. Folding 57 gravitationally bound systems into a hierarchical Bayesian model of their 3D orbits, I inferred the global MOND acceleration scale \(a_0\) directly from the data and found tension with MOND: the canonical value is excluded at \(3.1\sigma\) for the simple interpolating function and \(1.9\sigma\) for the standard form. The data and code are on GitHub.
In a companion paper, Saad & Ting (2026), I showed just how sensitive these measurements are to the orbital modeling itself. Chae et al. (2026) had reported a gravitational anomaly, a gravity boost of \(\gamma \approx 1.6\), in a high-quality sample of 36 wide binaries. Reanalyzing the same data with a hierarchical Bayesian model that forward-models the full 3D orbits, I recovered \(\gamma = 1.00 \pm 0.24\), fully consistent with Newtonian gravity. The anomaly reappears only when the semi-major axis is replaced by a geometric de-projection of the observed separation, which pins the discrepancy on the orbit modeling rather than on new physics. The code is on GitHub.
Many stars that look single in a survey are not. When a companion is too close to resolve but bright enough to contribute light, the spectrum carries two sets of absorption lines, shifted against each other by the orbital motion. These double-lined spectroscopic binaries (SB2s) are easy to miss, and a survey pipeline that fits them as one star returns parameters that are quietly wrong.
In Saad & Ting (2026b), I fit every main-sequence dwarf in APOGEE DR19 twice, once as a single star and once as a pair with a tied mass ratio, following El-Badry, Ting & Rix (2018), and keep the binary model when it is clearly preferred. Of 238,205 dwarfs, 41,466 (17.4%) are SB2 candidates, at a false-positive rate of 8.1% measured on held-out stars. The earlier search in DR13 found 2,645 in a sample twelve times smaller. The classifier is trained only on public spectra and labels, and the catalog, model weights and code are on GitHub. The analysis is also packaged as tools that a language-model agent can run, which I describe on the AI in Astronomy page.