D RAFT VERSION M ARCH 17, 2023

Typeset using LATEX twocolumn style in AASTeX63

The Empirical Limits of Gyrochronology

L UKE G. B OUMA , 1 , ∗ E LSA K. PALUMBO , 1 AND LYNNE A. H ILLENBRAND 1
1 Department of Astronomy, MC 249-17, California Institute of Technology, Pasadena, CA 91125, USA

(Received 2023 Jan 25; Revised 2023 Mar 3; Accepted 2023 Mar 14)
arXiv:2303.08830v1 [astro-ph.SR] 15 Mar 2023

ABSTRACT
The promise of gyrochronology is that given a star’s rotation period and mass, its age can be inferred. The
reality of gyrochronology is complicated by effects other than ordinary magnetized braking that alter stellar
rotation periods. In this work, we present an interpolation-based gyrochronology framework that reproduces the
time- and mass-dependent spin-down rates implied by the latest open cluster data, while also matching the rate at
which the dispersion in initial stellar rotation periods decreases as stars age. We validate our technique for stars
with temperatures of 3800–6200 K and ages of 0.08–2.6 gigayears (Gyr), and use it to reexamine the empirical
limits of gyrochronology. In line with previous work, we find that the uncertainty floor varies strongly with both
stellar mass and age. For Sun-like stars (≈5800 K), the statistical age uncertainties improve monotonically from
±38% at 0.2 Gyr to ±12% at 2 Gyr, and are caused by the empirical scatter of the cluster rotation sequences
combined with the rate of stellar spin-down. For low-mass K-dwarfs (≈4200 K), the posteriors are highly
asymmetric due to stalled spin-down, and ±1σ age uncertainties vary non-monotonically between 10% and
50% over the first few gigayears. High-mass K-dwarfs (5000 K) older than ≈1.5 Gyr yield the most precise
ages, with limiting uncertainties currently set by possible changes in the spin-down rate (12% systematic), the
calibration of the absolute age scale (8% systematic), and the width of the slow sequence (4% statistical). An
open-source implementation, gyro-interp, is available online at github.com/lgbouma/gyro-interp.

Keywords: Stellar ages (1581), Stellar rotation (1629), Field stars (2103); Bayesian statistics (1900)

1. INTRODUCTION This work aims to clarify the accuracy and precision of gy-
The ages of stars are fundamental for our understanding rochronology for stars on the main-sequence. Our main im-
of planetary, stellar, and galactic evolution. Unfortunately, petus for writing was the realization that available models did
stellar ages are not directly measurable, and so the astronom- not match observations of open cluster rotation periods (e.g.,
ical age scale is tied to a mix of semifundamental, model- Curtis et al. 2019a, 2020). The disagreement was most se-
dependent, and empirical techniques (Soderblom 2010). One vere for K-dwarfs, which have stellar rotation rates that stall
empirical age-dating method is to use a star’s spin-down as a from 0.7 to 1.4 Gyr (Agüeros et al. 2018; Curtis et al. 2020).
clock (Kawaler 1989; Barnes 2003). This gyrochronal tech- While a likely physical explanation centers on the timescale
nique leverages direct measurements of stellar surface rota- for angular momentum exchange between the radiative core
tion periods, typically inferred from photometric modulation and convective envelope (Spada & Lanzafame 2020), accu-
induced by spots or faculae. The clock’s mechanism is mag- racy is paramount because any bias in the rotation models
netized braking that drives rotation periods to increase as the propagates into bias on the inferred ages.
square root of time (Weber & Davis 1967; Skumanich 1972). Regarding precision, previous analytic studies have re-
While data from open clusters have shown the limitations of ported age uncertainties for field FGK dwarfs of 13–20%
this approximation, the idea has been useful, and it has set (Barnes 2007), and have noted that these uncertainties in-
the foundation for many empirical studies of how rotation crease for young stars due to larger empirical scatter in their
period, age, and activity are interrelated (e.g., Noyes et al. rotation sequences (Barnes 2010). The question of how this
1984; Barnes 2007; Mamajek & Hillenbrand 2008; Barnes empirical scatter, often described as “fast” and “slow” se-
2010; Angus et al. 2019; Spada & Lanzafame 2020). quences in the rotation–color plane, limits gyrochronal pre-
cision was analyzed in detail by Epstein & Pinsonneault
(2014). For stars older than 0.5 Gyr, their approach was to
Corresponding author: Luke G. Bouma consider the range of possible ages that a star with fixed ro-
[email protected] tation period and mass might have, and to convert this range
∗
into an age uncertainty. Our work formalizes this idea. If
51 Pegasi b Fellow
an astronomer wishes to infer the age of an individual field
2 B OUMA , PALUMBO , H ILLENBRAND

star, they do not know whether their star is on the fast or slow DR2 GBP − GRP colors by assuming E(GBP − GRP ) = 0.415AV ,
sequence. They simply know the star’s rotation period and similar to Curtis et al. (2020).
mass, and so they must marginalize over the population-level
scatter in order to determine a posterior probability distribu- 2.3. Binarity Filters
tion for the age. Ultimately, Epstein & Pinsonneault (2014) Binarity can affect the locations of stars in rotation–color
emphasized that this type of approach needed empirical guid- space by observationally biasing photometric color measure-
ance in order to mitigate the systematic uncertainties in the ments, and also by physically altering stellar rotation rates
spin-down models; such guidance now exists. through e.g., tidal spin-up or early disk dispersal. To remove
Using the latest available open cluster data (Section 2), we possible binaries from our calibration sample, we applied the
calibrate a new gyrochronal model that interpolates between following filters to each cluster dataset.
the open cluster rotation sequences (Section 3). Given a star’s Photometric binarity—We plotted the Gaia DR3 color–
rotation period, effective temperature, and their uncertainties, absolute magnitude diagrams in MG vs. GBP − GRP , G − GRP ,
our framework returns the implied gyrochronal age poste- and GBP − G, and manually drew loci to remove over or
rior, which is often asymmetric (Section 4). We validate our under-luminous stars in each diagram.
model against both training and test data, and focus our dis- RUWE—We examined diagrams of the Gaia DR3 renor-
cussion and conclusions (Section 5) on the empirical limits malized unit weight error (RUWE) as a function of bright-
of gyrochronal age-dating. An open-source implementation ness, and based on these diagrams required RUWE > 1.2.
is available online at github.com/lgbouma/gyro-interp. Outliers in this space can be caused by astrometric binarity,
or by marginally resolved point-sources fitted with a single-
2. BENCHMARK CLUSTERS source PSF model by the Gaia pipeline.
2.1. Rotation Data Radial velocity scatter—We examined diagrams of Gaia
DR3 “radial velocity error” as a function of G-mag. Since
To calibrate our model, we first collected rotation period this quantity is the standard deviation of the Gaia RV time-
data from open clusters that have been surveyed using pre- series, outliers can imply single-lined spectroscopic binarity.
cise space and ground-based photometers. The clusters that We manually removed such stars.
we examined are listed in Table 1, along with their ages and Crowding—We queried Gaia DR3 to determine how many
V -band extinctions. These clusters were selected based on stars were within 1 instrument pixel distance of each target
the completeness of available rotation period catalogs for F, star (e.g., 400 /px for Kepler). Any stars within ≈20× the
G, K, and early M dwarfs. The Pleiades, Blanco-1, and Psc- brightness of the target star (∆G < 3.25) were noted, and
Eri were concatenated as a 120 megayear (Myr) sequence, the target stars were removed from further consideration. Al-
since their rotation–temperature sequences were visually in- though not all visual companions are binaries, their presence
distinguishable. The upper age anchor, Ruprecht-147, was can complicate rotation period measurements, particularly in
similarly combined with NGC-6819 to make a 2.6 Gyr se- cluster environments.
quence. While older populations have been studied (Barnes Gaia DR3 Non-Single-Stars—Gaia DR3 includes a col-
et al. 2016; Dungee et al. 2022), their rotation–color se- umn to flag known or suspected eclipsing, astrometric, and
quences do not yet have sufficient coverage to be usable in spectroscopic binaries. We directly merged against this col-
our core analysis. Our lower anchor, α Per, was selected umn to remove such sources.
based on its converged rotation–temperature sequence above Final calibration sample—The combination of the filters
0.8 M (Boyle & Bouma 2022). Our model is therefore only described above yields the set of stars that show no evidence
constrained between 80 Myr and 2.6 Gyr. for binarity or crowding. However, some of the rotation pe-
riod analyses in Table 1 include additional relevant quality
2.2. Effective Temperatures flags. For instance, light curves showing multiple photomet-
For our effective temperature scale, we adopted the Cur- ric periods can indicate unresolved binarity. We used all rel-
tis et al. (2020) conversion from dereddened Gaia Data Re- evant filters available from the original authors if they were
lease 2 (DR2) GBP − GRP colors to effective temperatures. designed to select single stars with reliable rotation periods.
This calibration was determined using FGK stars with high- The final combination of these filters with our own flag for
resolution spectra (Brewer et al. 2016), nearby stars with in- possible binarity yields our sample of benchmark rotators.
terferometric radii (Boyajian et al. 2012), and M-dwarfs with
optical and near-infrared spectroscopy (Mann et al. 2015). 2.4. The Single-Star Calibration Sequence
The typical precision in temperature from this relationship is Figure 1 is the result of the data curation process described
50 K for stars near the zero-age main-sequence (ZAMS). We in Sections 2.1 through 2.3. While we have omitted the pos-
explicitly used Gaia DR2 mean photometry to calculate the sible binaries described in Section 2.3 for visual clarity, they
temperatures, since the intrinsic difference between the Gaia are included in the Data behind the Figure. The gray lines are
DR2 and DR3 colors is important at this scale. For all other derived from polynomial fits that we describe in the follow-
Gaia-based quantities in our analysis, we used the DR3 val- ing section. Comparing against the rotation–color sequences
ues. For the extinction corrections, we adopted the reddening in say Godoy-Rivera et al. (2021), it is impressive how sparse
values listed in Table 1. We dereddened the observed Gaia the fast sequence is for hot stars. In the 120 Myr clusters,
 G YROCHRONOLOGY BY I NTERPOLATION 3

Table 1. Reference clusters and parameters used for the core gyrochrone calibration.

Name Reference Age Age Provenance AV AV Provenance Instrument Prot Provenance Recovered Age?
α Per 79.0+1.5
−2.3 Myr (1) 0.28 (2† ) TESS (2) 70+8
−8 Myr

Pleiades 127.4+6.3
−10.0 Myr (1) 0.12 (3) K2 (4) 117+6
−6 Myr

Blanco-1 137.1+7.0
−33.0 Myr (1) 0.031 (5) NGTS (5) 134+11
−9 Myr

Psc-Eri stream Pleiades-coeval (6) 0 (6) TESS (6) 137+11

−10 Myr

NGC-3532 300 ± 50 Myr (7) 0.034 (8) Y4KCam (8) 306+11

−9 Myr

Group-X 300 ± 60 Myr (9) 0.016 (9) TESS (9) 260+29

−27 Myr

Praesepe 670 ± 67 Myr (10) 0.035 (3) K2 (11) 738+22

−20 Myr

NGC-6811 1040 ± 70 Myr (12) 0.15 (3) K2 (12) 946+16

−15 Myr

NGC-6819 2.5 ± 0.2 Gyr (13) 0.44 (3) Kepler (14) 2518+32
−33 Myr

Ruprecht-147 2.7 ± 0.2 Gyr (15) 0.30 (3) K2 (3) 2637+51

−52 Myr

N OTE— References: (1) Galindo-Guil et al. (2022); (2) Boyle & Bouma (2022); (3) Curtis et al. (2020); (4) Rebull et al. (2016); (5) Gillen
et al. (2020); (6) Curtis et al. (2019b); (7) Fritzewski et al. (2019); (8) Fritzewski et al. (2021); (9) Messina et al. (2022); (10) Douglas et al.
(2019); (11) Rampalli et al. (2021); (12) Curtis et al. (2019a); (13) Jeffries et al. (2013); (14) Meibom et al. (2015); (15) Torres et al. (2020).
†
The adopted α Per reddening varies across the cluster, per Boyle & Bouma (2022); this table reports the median value. ? See Section 4.1.

both Blanco-1 and Psc-Eri have no apparently single fast ro- By Bayes’ rule, the integrand can be written as
tators hotter than 5000 K. The Pleiades has four. The rapid
rotator sequence is similarly sparse at 300 Myr. The large bi- p(Prot , Teff ,t|D) ∝ p(D|Prot , Teff ,t) · p(Prot , Teff ,t) (2)
nary fraction of fast-sequence stars warrants future analysis,
to understand whether the binary separations and mass ratios where the first term is the likelihood and the latter is the prior.
for these systems are typical of the field binary population.
3.2. Likelihood
3. A GYROCHRONOLOGY MODEL For the likelihood, we assume that the observed rotation
Here we present a model that aims to accurately describe period and temperature have Gaussian uncertainties and are
the evolving rotation period distributions of F7–M0 dwarfs measured independently. In this case,
with ages of 0.08–2.6 Gyr. The goal is to then use this model
to assess the precision with which rotation periods can be p(D|Prot , Teff ,t) = p(P̃rot |Prot , Teff ,t) · p(T̃eff |Prot , Teff ,t), (3)
used to infer ages. To perform this analysis, our model needs
to account for the trends visible in Figure 1: stellar spin- and the temperature and rotation period distributions are
down rates vary with both mass and age; stellar spin-down specified by T̃eff ∼ N (Teff , σT̃2eff ) and P̃rot ∼ N (Prot , σP̃2rot ),
can stall; and higher-mass stars younger than Praesepe tend where N denotes the normal distribution. In other words,
to converge to the slow sequence before lower-mass stars. our likelihood is a product of two normal distributions.
Our approach will ultimately use interpolation, based on the
logic that there are certain regions of Figure 1 in which a hy- 3.3. Prior
pothetical star located between two cluster sequences would The prior is more interesting. By the chain rule,
need to have an age intermediate to those two clusters. A few
formalities are needed to make this idea rigorous. p(Prot , Teff ,t) = p(Prot |Teff ,t) · p(Teff ) · p(t), (4)
3.1. Formalism where we have assumed p(Teff |t) = p(Teff ) because in our
For a given star, we have an observed rotation period model, changes in stellar temperature through time are ig-
P̃rot and stellar effective temperature T̃eff with measurement nored. We assume that age and temperature are uniformly
min max
uncertainties σP̃rot and σT̃eff . Given these data, we want to distributed, t ∼ U(tmin ,tmax ) and Teff ∼ U(Teff , Teff ), where
min max
find the posterior probability distribution for the age t of (tmin ,tmax ), (Teff , Teff ) are the limiting ages and tempera-
the star. We write the corresponding probability density as tures for our model, respectively. We adopt limiting ages of
p(t|D), where D = {P̃rot , T̃eff } are the observed data. We find 0 to 2.6 Gyr, and limiting temperatures of 3800 to 6200 K.
p(t|D) by marginalizing over the joint probability density The upper limit on age is set by the oldest clusters in our
p(t, Prot , Teff |D), where Prot and Teff are the true rotation pe- dataset (Table 1), and the temperature limits are set to in-
riod and temperature of the star. Mathematically, this means clude the regions in which stellar rotation is most correlated
ZZ with age. While one might imagine a prior on temperature
p(t|D) = p(Prot , Teff ,t|D) dProt dTeff . (1) informed by the stellar initial mass function, or a prior on
age informed by the star formation history of the Milky Way,
4 B OUMA , PALUMBO , H ILLENBRAND

F2 F5 G2 K0 K5 M0 M3
80 Myr Per
120 Myr Pleiades
25 120 Myr Blanco-1
120 Myr Psc-Eri
300 Myr NGC-3532
300 Myr Group-X
20 670 Myr Praesepe
1 Gyr NGC-6811
Rotation Period [days] 2.5 Gyr NGC-6819
2.7 Gyr Rup-147
15

10

0
7000 6000 5000 4000 3000
Effective Temperature [K]
F2 F5 G2 K0 K5 M0 M3
15 80 Myr Per
120 Myr Pleiades
120 Myr Blanco-1
120 Myr Psc-Eri
300 Myr NGC-3532
300 Myr Group-X
670 Myr Praesepe
1 Gyr NGC-6811
Rotation Period [days]

10

0
7000 6000 5000 4000 3000
Effective Temperature [K]
Figure 1. Open cluster data and models. The top panel shows the data that we aim to model, and the bottom panel focuses on the first
gigayear. Gray lines in the top panel show the mean model for the rotation period distribution, and are uniformly spaced at integer multiples
of 100 Myr. They are evaluated using a seventh-order polynomial for each cluster (colored lines, bottom panel), and interpolated piecewise
between those reference loci. The model is defined over temperatures of 3800–6200 K, and ages of 0.08–2.6 Gyr. Data behind the Figure are
available as a machine-readable table.
the star formation rate has been approximately constant over ponent, µslow (Teff ,t), is the rotation period of the star if it
the past 10 Gyr (e.g., Nordström et al. 2004) and incorporat- were exactly on the slow sequence — this is colloquially
ing a stellar mass function would systematically bias already the “mean” gyrochronal model for a star’s rotation period
accurate measurements towards lower temperatures. We do prescribed at any age and temperature. The second compo-
not consider such additions. nent is the residual to that mean model — the probability
The remaining term in Equation 4, p(Prot |Teff ,t), is the distribution for how far the star’s rotation period is from the
core of our model. We propose a functional form for slow sequence at any given age and temperature. This model
p(Prot |Teff ,t) that relies on two components. The first com- parametrization is motivated by how the observed abundance
 G YROCHRONOLOGY BY I NTERPOLATION 5

of rapid rotators changes as a function of both stellar temper- between the slow and fast sequence. While other functional
ature and age. forms are possible, we assumed that at any given time this
function is defined as the temperature of the lowest-mass star
The Mean Model —To parametrize the slow sequence, we fit-
that has just arrived at the main-sequence, since this is the
ted rotation periods in each reference cluster with an N th or- time at which the star’s surface rotation rate is no longer af-
der polynomial over 3800–6200 K. We manually selected the fected by gravitational contraction. We evaluated this quan-
slow sequence stars to perform this fit using the data behind tity through linear interpolation over the solar-metallicity
Figure 1. We investigated the choice of N between 2 and 9, MIST grids (Choi et al. 2016). At 80, 120, and 300 Myr this
and settled on N = 7 as a compromise between overfitting and cut
yielded Teff values of 4620, 4150, and 3440 K, respectively.
accurately capturing the structure of the Praesepe and NGC-
6811 sequences. While lower-order polynomials provide ac- 3.4. Free Parameters
ceptable fits for the 80–300 Myr clusters (e.g., Curtis et al.
The free parameters in the model are as follows. In the
2020, Appendix A), for purposes of homogeneity across all
residual term, there are the amplitudes (a0 , a1 ), the two scal-
clusters we adopted a single polynomial order.
ing parameters (k0 , k1 ), and the slope of the linear amplitude
To model the evolution of the slow sequence, we consid-
decrease g(t) for the “fast sequence” term through time. This
ered a few possible approaches, all based on interpolating
would yield five free parameters, but a0 and a1 are degen-
between the fitted polynomials (see Appendix A). We ulti-
erate, so there are really only four degrees of freedom. We
mately chose at any given temperature to fit 1-D monotonic
fixed the other terms in the model that could in principle be
cubic splines in rotation period as a function of age. This
allowed to vary. These included the polynomial terms in the
guarantees a smooth increase in the slow sequence envelope
slow sequence model µslow , the scatter around the slow se-
while also fitting all available data. Systematic uncertainties
quence σ, and the function specifying the decrease of the ef-
associated with this choice are described in Section 5.2. This cut
fective temperature cutoff through time Teff (t).
procedure yielded the gray lines in Figure 1. At times be-
low 80 Myr, we do not extrapolate; we instead let the “mean 3.5. Fitting the Model
model” µslow (t, Teff ) equal the lowest reference polynomial
rotation period values as set by α Per. This yields poste- To compare the model (Equations 1 through 5) to the data,
rior distributions that are uniformly distributed at ages below we performed the following procedure. For the reference
80 Myr. Possible options regarding extrapolation for older clusters at 120 Myr (N? = 196), 300 Myr (N? = 133), and
stars are discussed in Appendix A. 670 Myr (N? = 100), we divided the data into seven bins,
starting at 3800 K, with uniform bin widths of 350 K. Includ-
The Residual —The top row of Figure 2 shows the residuals ing α Per (80 Myr; N? = 65) as an optional fourth dataset
for the calibration clusters with t ≤ 670 Myr, relative to the yielded similar results, so we omitted it for simplicity. In
polynomial model. Our ansatz is to model this distribution each bin, we counted the number of stars on the slow se-
as a sum of a Gaussian and a uniform distribution, with each quence, and the number of stars on the fast sequence. We
distribution smoothed around a time-dependent transition lo- considered a star to be “slow” if it is within two days of the
cation in effective temperature. This procedure ignores the mean slow sequence model, and “fast” if it is more than two
few positive outliers. days faster than the same model. This cutoff was determined
Mathematically, this means that the rotation period, given based on the uniform scatter of σ ≈ 0.51 days seen around
the age and temperature, is drawn from the slow-sequence for clusters with t ≥ 120 Myr. We then
use the resulting counts to define a “fast fraction,” F, the ra-
Prot ∼ a0 NP (µslow , σ 2 ) ⊗ LT (Teff
cut
(t), k0 ) (5) tio of fast rotating stars to the total number of stars observed
cut in any given temperature bin.


+ a1 g(t) UP (0, µslow ) ⊗ 1 − LT (Teff (t), k1 ) ,
The bottom row of Figure 2 shows this fast fraction as a
where N is a normal distribution, U is a uniform distribu- function of temperature. We calculated the same summary
tion, a0 and a1 are scaling constants, and L(`, k) is the lo- statistic for our model through numerical integration. This
yields a χ2 metric, χ2 = i (Fi − Fi,model )2 /σi2 , where the sum
P
gistic function specified by a location ` and smoothing scale
k. Visual examples are given in the middle row of Figure 2. i is over the three reference sets of open clusters. For the
The subscripts, for instance NP , indicate the dimension over σi , the default Poissonian uncertainties would disfavor the
which the distribution is defined — period (P) or effective small number of stars from 4500–6200 K in Praesepe that
temperature (T ), and ⊗ denotes an outer product. We have are all on the slow sequence. Since auxiliary clusters with
also hidden the dependence of µslow on time and temperature similar ages such as the Hyades (Douglas et al. 2019) and
for simplicity of notation. NGC-6811 (Curtis et al. 2019a) also have fully converged
The first term in Equation 5 parametrizes the slow se- slow sequences, we adopted a prescription for the σi in which
quence using a Gaussian centered on µslow (t, Teff ), with a uni- we set them to be equal to one another at 120 and 300 Myr
versal width σ = 0.51 days set by the observations of clus- and ten times smaller at 670 Myr. This forces the model to
ters at least as old as the Pleiades. The location parameter converge to the fast sequence by the age of Praesepe. The
cut
of the logistic function, Teff (t), is a function that monoton- normalization of the uncertainties was then allowed to float
ically decreases to account for the age-dependent transition in order to yield a reduced χ2 of unity.
6 B OUMA , PALUMBO , H ILLENBRAND

80 Myr 120 Myr 300 Myr 670 Myr

5 5 5 5
0 0 0 0
Rotation Period Data - Model [days]

5 5 5 5
Blanco-1
10 10 Pleiades 10 NGC-3532 10
Per Psc-Eri Group-X Praesepe

6000 5000 4000 6000 5000 4000 6000 5000 4000 6000 5000 4000

Probability
0.01 0.1 1

1.0 Best-fit Model

1.0 Best-fit Model
1.0 Best-fit Model
1.0 Best-fit Model
Samples Samples Samples Samples
0.8 Data 0.8 Data 0.8 Data 0.8 Data
Fast Fraction

0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0.0 0.0 0.0 0.0
6000 5000 4000 6000 5000 4000 6000 5000 4000 6000 5000 4000
Effective Temperature [K]
Figure 2. Data, model, and goodness-of-fit. Top: Cluster rotation periods, minus the corresponding slow-sequence mean model at each
cluster’s age. The lower gray envelope corresponds to a zero-day rotation period. Middle: Probabilistic model for rotation period as a function
of age and temperature (Equation 5), fitted to the 120 Myr, 300 Myr, and 670 Myr clusters. Bottom: Fraction of stars in 350 K bins that rotate
“fast”, as a function of temperature. “Fast” and “slow” stars are squares and circles on the top panel; “very slow” outliers are the crosses. “Slow”
stars show a uniform scatter of σ≈0.51 days around the mean model at t ≥ 120 Myr. The assumed uncertainties for Praesepe are smaller than
the markers (see Section 3.5).
We fitted the model by sampling the posterior probabil- our uncertainties, we performed a cross-validation analysis
ity using emcee (Foreman-Mackey et al. 2013). We sam- in which we randomly dropped 20% of the stars in the refer-
pled over five parameters: a1 /a0 , ln k0 , ln k1 , the slope of ence clusters, without replacement, and then refitted the data.
g(t), and the multiplicative uncertainty normalization. The The resulting parameters all fell within the stated 1σ uncer-
function g(t) was set to unity below 120 Myr, and to de- tainty intervals.
crease linearly to zero while intersecting 300 Myr at a par-
ticular value, yg . The latter value was the free parameter 3.6. Evaluating the Posterior
used to fit the slope of the line. The maximum-likelihood
For any given star, we numerically evaluate Equation 1
values yielded by this exercise were {a1 /a0 , ln k0 , ln k1 , yg } =
using the composite trapezoidal rule. For each age in a re-
{8.26, −4.88, −6.24, 0.67}. To evaluate the posterior, we as-
quested grid, we define linear grids in the dimensions of
sumed a prior on each parameter that was uniformly dis-
temperature and y ≡ P − µslow , each with side length Ngrid .
tributed over a wide boundary.1 We checked convergence
The integration is then performed over dTeff and dy at each
by running the chains out to a factor of 300 times longer
specified age. Runtime scales as O(Ngrid ), and takes under a
than the autocorrelation time. The resulting median pa-
minute on a typical laptop. This runtime estimate however
rameters and their 1σ intervals were {a1 /a0 , ln k0 , ln k1 , yg } =
assumes that the four hyperparameters, a1 /a0 , ln k0 , ln k1 , and
{9.29−+3.62
2.41 , −4.27−1.52 , −6.15−0.25 , 0.63−0.07 }. The lower row of
+2.56 +0.23 +0.03
yg , are fixed. Since these parameters are unknown, the most
Figure 2 shows the best-fit model plotted over 64 samples.
rigorous approach for age inference for any one star requires
Qualitatively, the model fits the fast fraction’s behavior well
sampling from the posterior probability distribution for the
in both temperature and time. To check the accuracy of
hyperparameters. Each sample then yields its own posterior
for the age from Equation 1, from which sub-samples can be
1 a1 /a0 ∼ U (1, 20), ln k0 ∼ U (−10, 0), ln k1 ∼ U (−10, 0), yg ∼ U (0.1, 1). drawn. All the sub-samples can then be combined to numer-
ically yield an average posterior.
 G YROCHRONOLOGY BY I NTERPOLATION 7

The top panels of Figure 3 show the results of this sam- mative in that it provides an upper limit on the star’s age.
pling procedure in dotted lines, plotted underneath an alter- As the star ages, the age posterior becomes two-sided, with
native: simply adopting the best-fit model (solid lines). The a best-case statistical precision of ±12% at 2 Gyr. For a
results are similar, although there are differences for most low-mass K-dwarf, the evolution of the age posterior is more
rapidly rotating stars. While the sampling procedure is rela- complicated. These stars only converge to the slow sequence
tively simple to parallelize, it is a factor of ≈103 times more by the age of Praesepe. Their spin-down is then observed to
expensive than using the best-fit model; for most practition- stall, which leads to highly asymmetric posteriors between
ers, the rigor is unlikely to justify the runtime cost. As we ages of 0.5–1.3 Gyr. For instance, a 4000 K star on the slow-
will discuss in Section 5.2, this model has other systematic sequence at 200 Myr has a +1σ uncertainty of 88%, and a
uncertainties that are more important. −1σ uncertainty of 13%. Nonetheless, statistical age preci-
sions for such stars are predicted to improve after the era of
4. RESULTS stalled spin-down, reaching ±9% by 2 Gyr. The implication
4.1. Model Validation is that the rotation periods of such stars can be predictive of
age, but only at certain times.
As a validation test, we calculated gyrochronal age pos-
teriors for all 3800–6200 K stars in Figure 1. To infer the 5. DISCUSSION & CONCLUSIONS
implied age for each cluster, we “stack” the posteriors us- 5.1. The Gyrochronal Precision Floor
ing PosteriorStacker2 , which considers two hierarchi-
cal Bayesian models for the intrinsic age distribution of each A key simplifying factor in our analysis is that we as-
cluster: a Gaussian, and a non-parametric histogram. After sumed the scatter of rotation periods around the slow se-
omitting a few extreme outliers3 , the two approaches give quence, σ ≈ 0.51 days, is fixed in time. Based on the data,
similar results, and so in the “recovered age” column of Ta- σ appears to be constant between 120 Myr and 1 Gyr (Fig-
ble 1 we report the median and uncertainty of the mean clus- ure 2, top panel). In α Per, the scatter is larger (0.85 days),
ter age assuming that the individual stellar ages in each clus- likely because the stars are only just converging to the slow
ter are drawn from a Gaussian. The resulting ages agree with sequence. The scatter is also larger in the Ruprecht-147 data,
the literature ages for every cluster to within 2σ, as we would but this is likely due to observational uncertainty in the period
expect for a sample of ten clusters. measurements. This empirical ≈0.51-day scatter could come
As an additional test, we repeated the exercise, but us- from a number of sources, including differential rotation on
ing data for two open clusters outside of our training data: stellar surfaces (Epstein & Pinsonneault 2014), uncertainties
M34 (≈240 Myr; Meibom et al. 2011) and M37 (≈500 Myr; in the effective temperature scale, or differing wind strengths
Hartman et al. 2009). For M34, fitting the data after apply- between stars of the same mass and age.
ing the binarity filters described in Section 2.3 yielded an Regardless of the scatter’s origin, it sets the floor for
age of 222 ± 20 Myr. For M37, the same procedure yielded gyrochronal precision, in tandem with the intrinsic spin-
463 ± 18 Myr. The latter estimate agrees with the isochronal down rates. In line with previous results (Barnes 2007),
age found by Hartman et al. (2008) without convective over- gyrochronal ages for F-dwarfs are less precise than for G-
shoot (485 ± 28 Myr), and is 2.5σ below their isochronal age dwarfs, because F-dwarfs spin down more slowly. However
that included convective overshoot (550 ± 30 Myr). in detail, Figure 3 shows that such statements depend on both
mass and age. More broadly, Figure 3 also implies that ac-
4.2. Precision of Gyrochronology counting for the evolving dispersion of the rotation period
distributions is a required ingredient for producing accurate
Having demonstrated that our method can recover the ages age uncertainties.
of known cluster stars, here we examine its statistical limits
for individual field stars. The bottom panel of Figure 3 shows 5.2. Systematic Uncertainties
the ±1σ uncertainties, normalized by the median of the gy- The uncertainties described thus far have been statistical,
rochronal age posteriors, over a grid of rotation periods and rather than systematic. Key systematic uncertainties include
temperatures. Broadly speaking, the regions in which rota- the time-varying nature of the spin-down rate, the accuracy
tion periods evolve the least, such as the hottest stars and of the absolute age scale, and stellar binarity.
stalled ≈1 Gyr K-dwarfs, have the worst inferred precisions. Regarding the spin-down rate, our interpolation approach
The top panel of Figure 3 visualizes vertical slices of the guarantees accuracy near any given reference cluster. How-
bottom panel for a few canonical cases. For a Sun-like star ever, far from the reference clusters, the choice of interpo-
(≈5800 K) in its early life, the rotation period is only infor- lation method can affect the inferred ages. We estimated the
associated systematic uncertainties by evaluating grids of Prot
2
vs. Teff analogous to Figure 3, but assuming i) piecewise lin-
See github.com/JohannesBuchner/PosteriorStacker, and Appendix A of
Baronchelli et al. (2020). This approach is viable because gyro-interp ear interpolation, and ii) piecewise cubic hermite interpolat-
adopts a uniform prior over age, and so the hierarchical likelihood simpli- ing polynomials calibrated only on the 0.08–2.6 Gyr data (see
fies to a product of the likelihoods for each star. Appendix A). The difference in the medians of the age pos-
3 TIC 44647574 in Psc-Eri; EPIC 212008710 in Praesepe; KIC 5026583 and
teriors relative to our default interpolation method is an indi-
KIC 5024122 in NGC-6819; and EPIC 219774323 in Ruprecht-147 cator of the systematic uncertainty. This procedure showed a
8 B OUMA , PALUMBO , H ILLENBRAND

5800K 5000K 4200K

2.0d 4.0d
8 10 9.0d
Probability (10 3 Myr 1)

Probability (10 3 Myr 1)

8 3.5d 5.5d

Probability (10 3 Myr 1)

10.5d
5.0d 7.0d 12.0d
6.5d 8.5d 8 13.5d
6 8.0d 6 10.0d 15.0d
9.5d 11.5d 16.5d
11.0d 13.0d 6
4 4
4
2 2 2
0 0 0
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
Age [Myr] Age [Myr] Age [Myr]
G2 K0 K5 M0 G2 K0 K5 M0
+1 t/t 2.5 -1 t/t 2.5

20 1
2.0 2.0
Rotation Period [days]

15

±1 t/median(t)
1.5 1.5
0.3
1.0 1.0
10
0.5 0.5 0.1
5

0 6000 5000 4000 6000 5000 4000 0.03

Effective Temperature [K] Effective Temperature [K]
Figure 3. Precision of gyrochronal ages from our method. Top: Age posteriors across rotation–temperature space. In each subplot, each
line represents a pair of Prot and Teff , and assumes a precision of 50 K in effective temperature, and 1% in rotation period. The solid lines come
from the best-fit model in Figure 2. The under-plotted dotted lines come from a more rigorous approach that samples over the population-level
hyperparameters discussed in Section 3.6. Bottom: +1σ (left) and −1σ uncertainty (right) of the age posterior, normalized by the median value.
For instance “±1σt /median(t) = 0.3” corresponds to 30% relative precision. Thick gray lines are at integer multiples of 500 Myr, and dotted
lines are spaced every 100 Myr; labels are in units of gigayears.
<1% bias in the inferred ages for 5000–6200 K stars younger Regarding the absolute age scale, Table 1 reports age pre-
than 1 Gyr, due to the dense sampling of the calibration clus- cisions for the calibration clusters of 3–20%, with the largest
ters. For cooler stars however, a linear spin-down rate would uncertainties for the 300 ± 60 Myr NGC-3532 and Group-X.
yield differences of up to ±15% in the median age due to To assess how shifts in this scale might affect our gyrochronal
the rapid spin-down from 0.1–0.3 Gyr (see Figure 1). For ages, we again calculated grids of Prot vs. Teff , but in this case
older stars between 1 and 2.6 Gyr, the cubic interpolation we shifted all the reference cluster ages by either +1σ or −1σ.
yielded ±100 Myr differences, while the linear interpolation The results showed what one would naively expect: if all
yielded −200 to +50 Myr differences, with the largest differ- the clusters are ±1σ older than their reference ages, then the
ences again for stars cooler than 4500 K. The summary is changes in the inferred ages match however much freedom
that from 1–2.6 Gyr, there is a 6–12% systematic uncertainty, there is in the local age scale. For example, for a 5800 K star
with the maximum uncertainty at 1.8 Gyr, half-way between with a 5.1 day rotation period, our method statistically yields
the two reference clusters. t = 308+70
−81 Myr, roughly the age of NGC-3532. However, the
 G YROCHRONOLOGY BY I NTERPOLATION 9

age of that cluster is uncertain at the 20% level, and so the to complement our own is the BAFFLES code (Stanford-
median age from our estimate for this worst-case scenario Moore et al. 2020), which returns age posterior probabilities
could be systematically shifted either up or down by ±20% based on a star’s surface lithium content. Other age-dating
to match the true age of the reference cluster. From the un- tools, including activity (Ca HK, Ca IRT, x-ray, UV excess),
certainties quoted in Table 1, and from comparable studies in isochrones, and asteroseismology, can similarly be combined
the literature (e.g., Dahm 2015), the age scale itself seems to with our gyrochronal posteriors to verify the accuracy of our
currently be defined at a ∼10% level of accuracy, at best. rotation-based ages, and to improve on their precision.
Finally, regarding binarity, the presence of even a wide Angus et al. (2019) presented an important step in
binary during the pre-main-sequence can prompt fast disk this vein, through a method that simultaneously fitted an
clearing, which could alter a star’s rotation period by halting isochronal and gyrochronal model to determine a star’s age.
disk-locking (Meibom et al. 2007). This mechanism might Their statistical framework could certainly encompass the
explain the abundance of fast rotators in ≈120 Myr open model developed in this manuscript. The main advantages
clusters (Bouma et al. 2021). A separate concern with bina- of our particular gyrochronology model however are i) im-
ries is photometric blending of the rotation signal. Because proved accuracy for stalling K-dwarfs, ii) improved accuracy
of these issues, our framework is only strictly applicable to in treating the growth of the slow sequence and decay of the
stars that are apparently single. Section 2.3 summarizes some fast sequence over the first gigayear, and iii) incorporation of
of the information that can be used to determine whether a the astrophysical width of the slow sequence for FGK stars.
given field star meets this designation. Appendix B discusses The main disadvantage is that our model is not applicable
the potential impact of ignoring binarity entirely. beyond 2.6 Gyr, though we caution that this is because the
calibration data are more sparse in this regime, and so the
5.3. Future Directions ages have larger systematic uncertainties.
The need for intermediate-age calibrators —The region of Fig- Physics-based models —A separate issue with our model is
ure 1 with the largest gap, near 1.8 Gyr, has the largest sys- that it is empirical, and so it does not yield physical under-
tematic uncertainties in our model. These uncertainties could standing. Physics-based gyrochronology models have pro-
be addressed by measuring rotation periods in a cluster at vided crucial insight into what gives the data in Figure 1 their
this age. Considering clusters from Cantat-Gaudin et al. structure. The relevant physics likely includes decoupling be-
(2020) older than 1 Gyr, within 1 kpc, and with more than 100 tween the radiative core and convective envelope (Gallet &
members yields eight objects. Sorted near to far, they are: Bouvier 2013), angular momentum transport to recouple the
Ruprecht-147, NGC-752, IC-4756, NGC-6991, NGC-2682, core and envelope (Gallet & Bouvier 2015; Spada & Lan-
NGC-7762, NGC-2423, and IC-4561. The closest two have zafame 2020), and spin-down rates that vary depending on
been studied by Curtis et al. (2020) and Agüeros et al. (2018), whether the magnetic dynamo is saturated (e.g., Sills et al.
though rotation periods in NGC-752 (1.34 ± 0.06 Gyr, d ∼ 2000; Matt et al. 2015). At older ages, additional physics
440 pc) could be worth revisiting using data from the Tran- may well be needed to explain the lethargic spin-down of
siting Exoplanet Survey Satellite and the Zwicky Transient stars with Rossby numbers comparable to the Sun (Brown
Facility. IC-4756 and NGC-6991 could similarly merit fur- 2014; van Saders et al. 2016; David et al. 2022). A separate
ther study, though it would be wise to confirm their ages be- issue that also merits attention is the exact role of binarity on
fore delving in a rotation period analysis. stellar rotation. Our filtering process (Section 2.3) removed
potential binaries based on a gamut of tracers, because ob-
Going older —M67 (4 Gyr) will likely be the next rung in the servations have shown that rapid rotators are often binaries
gyrochronology ladder: the analyses by Barnes et al. (2016) (Meibom et al. 2007; Stauffer et al. 2016; Gillen et al. 2020).
and Dungee et al. (2022) have nearly completed its rotation– The exact properties of these binaries, for instance their sep-
color sequence. As described in Appendix A, we used their arations and masses, would help in clarifying the physical
data on M67 to calibrate the rate of spin-down between 1 origin of this correlation. The issue of whether binarity leads
and 2.6 Gyr. This choice is connected to a generic issue with to early disk dispersal seems likely to be related, and also
interpolation-based methods: the systematic uncertainty in deserves attention (Cieza et al. 2009).
the model increases near the boundaries of the interpolation
domain. By this logic, incorporating the 4 Gyr data in the
most reliable way would require an even older population of
ACKNOWLEDGMENTS
stars. Clusters such as NGC 6791 (8 Gyr; Chaboyer et al.
1999), or else a precise set of asteroseismic calibrators (e.g., This work was supported by the Heising-Simons 51 Pe-
van Saders et al. 2016) might be the most plausible paths gasi b Fellowship (LGB) and the Arthur R. Adams SURF
toward this goal, though the complicating effects of stellar Fellowship (EKP).
evolution bear consideration.
Facilities: Gaia (Gaia Collaboration et al. 2022), Ke-
Precision age-dating of field stars —The best way to demon- pler (Borucki et al. 2010), TESS (Ricker et al. 2015), NGTS
strate the reliability of a star’s age is to measure it using (Wheatley et al. 2018)
independent techniques. One framework that we expect
10 B OUMA , PALUMBO , H ILLENBRAND

REFERENCES
Agüeros, M. A., Bowsher, E. C., Bochanski, J. J., et al. 2018, ApJ, Galindo-Guil, F. J., Barrado, D., Bouy, H., et al. 2022, A&A, 664,
862, 33 A70
Angus, R., Morton, T. D., Foreman-Mackey, D., et al. 2019, AJ, Gallet, F., & Bouvier, J. 2013, A&A, 556, A36
158, 173 —. 2015, A&A, 577, A98
Barnes, S. A. 2003, ApJ, 586, 464 Gillen, E., Briegal, J. T., Hodgkin, S. T., et al. 2020, MNRAS, 492,
—. 2007, ApJ, 669, 1167 1008
—. 2010, ApJ, 722, 222 Godoy-Rivera, D., Pinsonneault, M. H., & Rebull, L. M. 2021,
Barnes, S. A., Weingrill, J., Fritzewski, D., Strassmeier, K. G., & ApJS, 257, 46
Platais, I. 2016, ApJ, 823, 16 Hartman, J. D., Gaudi, B. S., Holman, M. J., et al. 2008, ApJ, 675,
Baronchelli, L., Nandra, K., & Buchner, J. 2020, MNRAS, 498, 1233
5284 Hartman, J. D., Gaudi, B. S., Pinsonneault, M. H., et al. 2009, ApJ,
Borucki, W. J., Koch, D., Basri, G., et al. 2010, Science, 327, 977 691, 342
Bouma, L. G., Curtis, J. L., Hartman, J. D., Winn, J. N., & Bakos, Jeffries, Mark W., J., Sandquist, E. L., Mathieu, R. D., et al. 2013,
G. Á. 2021, AJ, 162, 197 AJ, 146, 58
Boyajian, T. S., von Braun, K., van Belle, G., et al. 2012, ApJ, 757, Kawaler, S. D. 1989, ApJL, 343, L65
112 Mamajek, E. E., & Hillenbrand, L. A. 2008, ApJ, 687, 1264
Boyle, A. W., & Bouma, L. G. 2022, arXiv e-prints, Mann, A. W., Feiden, G. A., Gaidos, E., Boyajian, T., & von
arXiv:2211.09822 Braun, K. 2015, ApJ, 804, 64
Brewer, J. M., Fischer, D. A., Valenti, J. A., & Piskunov, N. 2016, Matt, S. P., Brun, A. S., Baraffe, I., Bouvier, J., & Chabrier, G.
ApJS, 225, 32 2015, ApJL, 799, L23
Meibom, S., Barnes, S. A., Platais, I., et al. 2015, Nature, 517, 589
Brown, T. M. 2014, ApJ, 789, 101
Meibom, S., Mathieu, R. D., & Stassun, K. G. 2007, ApJL, 665,
Cantat-Gaudin, T., Anders, F., Castro-Ginard, A., et al. 2020,
L155
A&A, 640, A1
Meibom, S., Mathieu, R. D., Stassun, K. G., Liebesny, P., & Saar,
Chaboyer, B., Green, E. M., & Liebert, J. 1999, AJ, 117, 1360
S. H. 2011, ApJ, 733, 115
Choi, J., Dotter, A., Conroy, C., et al. 2016, ApJ, 823, 102
Messina, S., Nardiello, D., Desidera, S., et al. 2022, A&A, 657, L3
Cieza, L. A., Padgett, D. L., Allen, L. E., et al. 2009, ApJL, 696,
Nordström, B., Mayor, M., Andersen, J., et al. 2004, A&A, 418,
L84
989
Curtis, J. L., Agüeros, M. A., Douglas, S. T., & Meibom, S. 2019a,
Noyes, R. W., Hartmann, L. W., Baliunas, S. L., Duncan, D. K., &
ApJ, 879, 49
Vaughan, A. H. 1984, ApJ, 279, 763
Curtis, J. L., Agüeros, M. A., Mamajek, E. E., Wright, J. T., &
Rampalli, R., Agüeros, M. A., Curtis, J. L., et al. 2021, ApJ, 921,
Cummings, J. D. 2019b, AJ, 158, 77
167
Curtis, J. L., Agüeros, M. A., Matt, S. P., et al. 2020, ApJ, 904, 140
Rebull, L. M., Stauffer, J. R., Bouvier, J., et al. 2016, AJ, 152, 114
Dahm, S. E. 2015, ApJ, 813, 108
Ricker, G. R., Winn, J. N., Vanderspek, R., et al. 2015, JATIS, 1,
David, T. J., Angus, R., Curtis, J. L., et al. 2022, ApJ, 933, 114
014003
Douglas, S. T., Curtis, J. L., Agüeros, M. A., et al. 2019, ApJ, 879,
Sills, A., Pinsonneault, M. H., & Terndrup, D. M. 2000, ApJ, 534,
100
335
Dungee, R., van Saders, J., Gaidos, E., et al. 2022, ApJ, 938, 118 Skumanich, A. 1972, ApJ, 171, 565
Epstein, C. R., & Pinsonneault, M. H. 2014, ApJ, 780, 159 Soderblom, D. R. 2010, ARA&A, 48, 581
Foreman-Mackey, D., Hogg, D. W., Lang, D., & Goodman, J. Spada, F., & Lanzafame, A. C. 2020, A&A, 636, A76
2013, PASP, 125, 306 Stanford-Moore, S. A., Nielsen, E. L., De Rosa, R. J., Macintosh,
Fritsch, F. N., & Butland, J. 1984, SIAM Journal on Scientific and B., & Czekala, I. 2020, ApJ, 898, 27
Statistical Computing, 5, 300 Stauffer, J., Rebull, L., Bouvier, J., et al. 2016, AJ, 152, 115
Fritzewski, D. J., Barnes, S. A., James, D. J., et al. 2019, A&A, Torres, G., Vanderburg, A., Curtis, J. L., et al. 2020, ApJ, 896, 162
622, A110 van Saders, J. L., Ceillier, T., Metcalfe, T. S., et al. 2016, Nature,
Fritzewski, D. J., Barnes, S. A., James, D. J., & Strassmeier, K. G. 529, 181
2021, A&A, 652, A60 Weber, E. J., & Davis, Leverett, J. 1967, ApJ, 148, 217
Gaia Collaboration, Vallenari, A., Brown, A. G. A., et al. 2022, Wheatley, P. J., West, R. G., Goad, M. R., et al. 2018, MNRAS,
arXiv e-prints, arXiv:2208.00211 475, 4476
 G YROCHRONOLOGY BY I NTERPOLATION 11

APPENDIX

A. INTERPOLATION METHODS & LITERATURE COMPARISON

How does the slow sequence evolve between each reference cluster? In other words, what is the functional form of µslow (t, Teff ),
the rotation period of a star evolving exactly along the slow sequence? Figure 4 summarizes a few possible answers, evaluated
at 5800 K, 5000 K, and 4200 K. Data from Barnes et al. (2016) and Dungee et al. (2022) have been included as the 4 Gyr M67
data points, to assess how well the interpolation methods succeed at extrapolating beyond 2.6 Gyr. As a qualitative note, the stars
of interest in this work (0.5–1.2 M ; 3800–6200 K) have temperature changes of .2.5% between 80 Myr and 2.6 Gyr, according
to the MIST grids (Choi et al. 2016). Past ≈3 Gyr, the most massive 1.2 M (≈ 6200 K) stars begin to turn onto the subgiant
branch, and temperature becomes a more ambiguous parameter when modeling stellar spin-down.
The simplest plausible model would be if the slow sequence’s evolution followed a power-law, with a flexible color or tem-
perature calibration similar to that suggested by many authors (e.g., Skumanich 1972; Noyes et al. 1984; Barnes 2003). In this
approach, for every temperature, we would set Prot ∝ t n , where canonically n = 1/2. We would then scale based on some fiducial
rotation period, say at 120 Myr. Figure 4 shows how well this type of scaling works, letting n float in order to match the data as
well as possible. For Sun-like stars (≈5800 K), this type of scaling works surprisingly well, yielding agreement with the cluster
data at the <15% level for n = 0.47 out to 4 Gyr. The agreement is significantly worse for lower mass stars, due to their stalled
spin-down at intermediate ages.
An alternative approach would be to directly interpolate between the cluster sequences, ignoring our expectation for any kind of
power-law spin-down. The resulting linear and quadratic interpolation cases are shown as the dot and dot-dash lines in Figure 4.
While these approaches tautologically fit the data, they suffer from sharp transitions in the spin-down rate at every reference
cluster. Quadratic interpolation is also not guaranteed to be monotonic, which is probably a desirable property for a stellar spin-
down. A final concern is that interpolating in this way is not guaranteed to be predictive; leaving the M67 data out, extrapolating
based on the 1–2.6 Gyr data will generally over or under-estimate the rotation periods in the 2.6–4 Gyr interval.
An approach closer to interpolation that still incorporates a form of power-law scaling is as follows. For a point (Ti , Pi )
intermediate between the loci of two clusters with ages t0 and t1 and rotation periods P0 and P1 at the same temperature Ti , set
 n

ti log P1 /P0
Pi = P0 , for n =
. (A1)
t0 log t1 /t0

In other words, given the full set of reference loci {µ0 , µ1 , . . . , µk }, their ratios {µ1 /µ0 , . . . , µk /µk−1 } can be used to define power-
law scalings that are accurate at a piecewise level. While this tautologically fits the data, there is a concern that for cool stars older
than 1 Gyr, it may over-estimate the rotation periods. This concern is primarily based on the sharp transition visible in Figure 4
in the spin-down rate at 1 Gyr for the 4200 K case.
A final approach is based on PCHIP interpolation (Piecewise Cubic Hermite Interpolating Polynomials; Fritsch & Butland
1984). This approach is monotonic, and continuous in the first derivatives at each reference cluster. While it is interpolation-
based, and therefore not predictive outside of its training bounds, we can include the M67 data in order to define the most accurate
possible slow sequence evolution over the 1–2.6 Gyr interval. The results are shown with the black line in Figure 4 in the method
labeled “pchip_m67,” which we adopt as our default. This approach leaves the slope of Prot vs. t even less constrained in the
2.6–4 Gyr interval, which is why we do not advocate using our model for stars older than 2.6 Gyr.
Finally, the models from Mamajek & Hillenbrand (2008) (MH08), Angus et al. (2019) (A19), and Spada & Lanzafame (2020)
(SL20) are also shown in Figure 4 for comparison. The MH08 model is defined over 0.5 < (B − V )0 < 0.9, or roughly 5050–
6250 K. The Teff = 5000 K case is therefore a mild over-extrapolation, but we nonetheless show the result for illustrative purposes.
Of the three cases, the Spada & Lanzafame (2020) model generally provides the best match to the data.

B. WHAT IF WE IGNORED BINARITY?

In this work we argued that omitting binaries from gyrochronology analyses is important due to the observational and astro-
physical biases that they can otherwise induce on rotation periods. In Section 2.3, we described the set of quality filters that
we used to expunge binaries from our calibration data, to guarantee that we were considering only apparently single stars with
reliable rotation period measurements. For generic field stars, not all of these conditions are necessarily applicable. For instance,
outliers in color–absolute magnitude diagrams might be challenging to identify due to the lack of an immediately obvious ref-
erence sequence (although the locus of the main-sequence is itself well-known, and Gaia for instance can now be used to query
local spatial volumes around arbitrary field stars to construct well-defined reference samples).
In general, we strongly recommend applying our method only to stars that are thought to be single and on the main sequence.
For instance, spectroscopic surface gravity estimates should be used, if available, to expunge evolved stars since they are not in
 12 B OUMA , PALUMBO , H ILLENBRAND

5800 K 5000 K 4200 K

skumanich_fix_n_0.47 skumanich_fix_n_0.33
50 skumanich_fix_n_0.25
25 skumanich_vary_n 40 skumanich_vary_n skumanich_vary_n
1d_linear
1d_quadratic
1d_linear
1d_quadratic 40 1d_linear
1d_quadratic
20 1d_pchip
30
1d_pchip 1d_pchip
Prot [days]

Prot [days]

Prot [days]
pchip_m67 pchip_m67 pchip_m67
MH08 MH08 30 A19
15 A19 A19 SL20
SL20
20 SL20
20
10
5 10 10

30 30 30
20 20 20
Residual [%]

Residual [%]

Residual [%]
10 10 10
0 0 0
10 10 10
20 20 20
30 30 30
100 1000 100 1000 100 1000
Age [Myr] Age [Myr] Age [Myr]
5800 K 5000 K 4200 K
skumanich_fix_n_0.47 skumanich_fix_n_0.33
50 skumanich_fix_n_0.25
25 skumanich_vary_n 40 skumanich_vary_n skumanich_vary_n
1d_linear
1d_quadratic
1d_linear
1d_quadratic 40 1d_linear
1d_quadratic
20 1d_pchip
30
1d_pchip 1d_pchip
Prot [days]

Prot [days]

pchip_m67
MH08
pchip_m67
MH08 Prot [days] 30 pchip_m67
A19
15 A19 A19 SL20
SL20
20 SL20
20
10
5 10 10

30 30 30
20 20 20
Residual [%]

Residual [%]

Residual [%]

10 10 10
0 0 0
10 10 10
20 20 20
30 30 30
0 2000 4000 0 2000 4000 0 2000 4000
Age [Myr] Age [Myr] Age [Myr]
Figure 4. Different approaches for interpolating between reference clusters. Prot denotes the rotation period of the star if it were evolving
exactly along the slow sequence. The top two and bottom two rows show identical data, but are scaled logarithmically and linearly in time. The
“residual” is defined versus the pchip_m67 interpolation method, calculated for each model i as (Prot,i − Prot,pchip_m67 )/Prot,i . The “+” data points
are evaluated from polynomial fits to the data in Figure 1. The fixed power laws (“skumanich_fix_n_0.XX”) are extrapolated based on
the rotation period at 120 Myr. MH08: Mamajek & Hillenbrand (2008). A19: Angus et al. (2019). SL20: Spada & Lanzafame (2020).
 G YROCHRONOLOGY BY I NTERPOLATION 13

80 Myr 120 Myr 300 Myr 670 Myr

5 5 5 5
0 0 0 0
Rotation Period Data - Model [days]

5 5 5 5
Blanco-1
10 10 Pleiades 10 NGC-3532 10
Per Psc-Eri Group-X Praesepe

6000 5000 4000 6000 5000 4000 6000 5000 4000 6000 5000 4000

Probability
0.01 0.1 1

1.0 Best-fit Model

1.0 Best-fit Model
1.0 Best-fit Model
1.0 Best-fit Model
Samples Samples Samples Samples
0.8 Data 0.8 Data 0.8 Data 0.8 Data
Fast Fraction

0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0.0 0.0 0.0 0.0
6000 5000 4000 6000 5000 4000 6000 5000 4000 6000 5000 4000
Effective Temperature [K]
Figure 5. What if we loosened the quality cuts? This plot is the same as Figure 2, but systems that are known or suspected to be visual,
photometric, astrometric, and spectroscopic binaries are now displayed along with the single stars. The model is the same as in Figure 2, as is
the panel ordering.

our calibration data. Nonetheless, it is interesting to consider how well our method translates for samples that are messier, and
that have binarity rates in line with field populations. Figure 5 shows the result of dropping all of the quality cuts described in
Section 2.3, using the data included behind Figure 1.
The first noticeable effect is that without any quality cuts, there are more stars. The star count in α Per jumps from 65 to
128; in the 120 Myr clusters from 196 to 364, in the 300 Myr clusters from 133 to 301; and in Praesepe from 100 to 250. In
addition, without quality cuts, the width of the slow sequence increases. The mean residual width for the t ≥ 120 Myr stars
within 2 days of the slow sequence is 0.72 days, a 40% increase from σ = 0.51 days observed in the cleaned sample. This scatter
term is proportional to the statistical age uncertainty at late times, in the regime of very precise rotation period and temperature
measurements (Barnes 2007). This suggests that if one wished to apply our gyrochronology model to a population with a mixture
of single and binary stars, the model would need to be refit to account for the wider intrinsic scatter in such a population.
Finally, we can ask to what degree the ratio between fast and slow rotators changes when we omit all quality cuts. The results
are shown in the bottom row of Figure 5, and compared against the original best-fit model (trained on the cleaned data) from
Figure 2. While the visual agreement remains good at t ≥ 120 Myr, the hot stars in the raw α Per sample have a larger fast fraction
than in the cleaned sample, and so the model provides a worse match to those stars. A second qualitatively important difference is
present in Praesepe: the raw data show around a dozen rapid outliers, none of which are present in the cleaned dataset (Figure 2).
If any of these stars were single and rapidly rotating, we might construe them as motivation to lengthen our model’s timescale
for the decay of the fast sequence. However, since they are most likely binaries, and the Hyades similarly shows no evidence for
rapidly rotating single stars hotter than 3800 K (Douglas et al. 2019). The NGC-6811 data at 1 Gyr similarly have no reported
rapid rotators (Curtis et al. 2019a). We therefore simply note that these outlying stars do exist at 0.7 Gyr, and that practitioners
aiming to perform gyrochronology analyses on populations of stars that include binaries should consider them.