Estimating distances from parallaxes. VII. Computing the three-dimensional kinegeometric posterior with one-dimensional sampling
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.
Vera: We're looking at the paper: "Estimating distances from parallaxes. VII. Computing the three-dimensional kinegeometric posterior with one-dimensional sampling".
Jocelyn: Here's the gist of the paper. When parallaxes are of poor quality, proper motions can improve stellar distance estimates using known constraints on stars’ physical velocities.
Vera: The paper introduces the kinegeometric posterior, a three-dimensional (3D) posterior in distance and velocity for given measured parallax and 2D proper motion.
Jocelyn: This method is presented as a faster and more accurate alternative to direct 3D sampling of the joint distance–velocity posterior.
Vera: The kinegeometric posterior formulation The kinegeometric posterior is defined as the product of a covariant Gaussian likelihood in the parallax and proper motion with a distance and velocity prior that can be constructed from a Galaxy model.
Paper summary: Vera: To recap, this paper focuses on estimating stellar distances using both parallax and proper motion when those parallax measurements aren't very good Jocelyn The core idea is to use the kinegeometric posterior which is a three dimensional thing in distance and velocity for a measured parallax and two dimensional proper motion Subrahmanyan They claim this can be done faster than the direct three dimensional sampling method they used before Vera
Jocelyn: The main thesis here is that instead of sampling the whole complicated joint posterior directly, you can reduce it to a one dimensional posterior for distance and then sample the velocity posteriors separately Vera They do this by integrating over velocity first to get that one dimensional distance posterior which they show can be sampled numerically Subrahmanyan
Subrahmanyan: That reduction is really clever because they show that because velocity is linearly proportional to proper motion at a given distance, you can reduce the three dimensional sampling without making any approximations Jocelyn They also take a look at the priors used, showing how you can factorize them into a two dimensional Gaussian prior in velocity conditional on distance and a one dimensional non-Gaussian prior over distance Vera
Jocelyn: So what’s the result? This 1D plus 2D representation is both faster and more accurate to sample compared to the direct three dimensional method Subrahmanyan They also note that by adopting a Gaussian prior on velocities at a given distance, those two dimensional distributions become Gaussian, which makes the velocity posterior a sum of Gaussians Vera <ref:2611.17231#pg1,2D representation is both faster and more accurate to sample>
Vera: That’s what it means for us on the ground. When we get messy parallax data, this method gives us better distance estimates because it properly incorporates that known constraint from the stars' physical velocities Jocelyn It makes computing these estimates for huge datasets much more feasible with this approach Subrahmanyan
Subrahmanyan: I think the implication is that if we can reliably get those 1D distance samples, we can use them to constrain our models of structure formation or even cosmology Vera It links the local stellar kinematics directly to the overall cosmic picture in a way that was harder before Jocelyn <ref:2611.17231#pg1>
Conclusion: Vera: So looking at this paper 'Estimating distances from parallaxes. VII. Computing the three dimensional kinegeometric posterior with one-dimensional sampling' and its authors, what’s the real meaning of all this? Jocelyn It boils down to being able to compute that complicated three dimensional distance velocity posterior in a faster way Subrahmanyan
Subrahmanyan: They showed that you can draw samples from the marginal one dimensional non-Gaussian distance posterior and then use those samples for the conditional velocity posterior which is a two dimensional Gaussian mixture model Vera This is more accurate than trying to sample the full three dimensional thing directly Jocelyn
Jocelyn: The paper really shows that MCMC could be used to sample that one dimensional distribution, but they suggest for these specific posteriors, doing one dimensional sampling with other numerical methods can be much more efficient at higher accuracy Vera It makes sense because the method requires far fewer posterior evaluations in practice Subrahmanyan
Subrahmanyan: I think the practical implication is that we can use this kinegeometric posterior to improve our distance estimates for stars, and since those estimates depend on kinematics, it gives us a new way to constrain how fast the universe is expanding near us Vera It’s a tool that uses all the data available to get a better picture of our local neighborhood Jocelyn
Vera: Right. So we’re talking about using Gaia data and proper motions to get more robust distances, even when the parallax measurements are noisy because this approach is faster and more accurate than what we had before Subrahmanyan It’s a practical tool for observational astronomers everywhere Jocelyn
Coryn A.L. Bailer-Jones
Max Planck Institute for Astronomy
astro-ph.IM
Submitted: 2026-11-12
Updated: 2026-11-12
Comments: For earlier papers in this series see https://bailer-jones.www3.mpia.de/gedr3_distances_velocities.html
Project page: https://statproofbook.github.io/P/mvn-cond.html
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 98/100
The gist: When parallaxes are of poor quality, proper motions can improve stellar distance estimates using known constraints on stars’ physical velocities.
Key concepts
- kinegeometric posterior
- This is a three-dimensional probability distribution describing both the distance and velocity of stars, given their measured parallax and proper motion. It combines a likelihood function with prior information derived from galaxy models to constrain the possible physical states of the stars.
- 1D+2D sampling
- This is an efficient computational strategy where the complex 3D problem is broken down. First, a single sample is drawn from the one-dimensional distance posterior, and then for each distance sample, multiple samples are drawn from the resulting two-dimensional conditional velocity posteriors.
- marginal posterior in distance
- This refers to the probability distribution of stellar distances calculated after integrating out all possible velocity components. It is a non-trivial one-dimensional function that can be sampled numerically to provide initial distance estimates for the next step.
- conditional posterior velocity PDF
- This describes the probability density function of a star's velocity given a specific distance and other fixed parameters. This distribution is shown to be a two-dimensional Gaussian, which allows for efficient sampling when combined with the 1D distance samples.
Terminology
Summary
When parallaxes are of poor quality, proper motions can improve stellar distance estimates using known constraints on stars’ physical velocities. p. 0 Bailer-Jones
Max Planck Institute for Astronomy, Heidelberg
https://www. de/homes/bailer-jones
12 November 2026
Abstract
When parallaxes are of poor quality, proper motions can improve stellar distance estimates using known constraints on stars’ physical velocities. de/homes/bailer-jones
12 November 2026
Abstract
When parallaxes are of poor quality, proper motions can improve stellar distance estimates using known constraints on stars’ physical velocities. Formally this may be
achieved via the kinegeometric posterior, a three-dimensional (3D) posterior in distance and velocity for given measured parallax and 2D proper motion. 6 Conclusion
To compute the 3D distance–velocity kinegeometric posterior we can draw Ng samples
from the marginal 1D non-Gaussian distance posterior, followed by the same number
of samples from the conditional velocity posterior, which is a 2D Gaussian mixture
model. This approach is more accurate than directly sampling the 3D non-Gaussian
joint distance–velocity posterior, and in practice requires far fewer posterior evaluations, making it faster.Where the paper says it
Where the paper says it
Where the paper says it
The kinegeometric posterior formulation
The kinegeometric posterior is defined as the product of a covariant Gaussian likelihood in the parallax and proper motion with a distance and velocity prior that can be constructed from a Galaxy model. The authors show that this 3D sampling can be reduced without approximation to a 1D sampling of the distance posterior followed by single samplings of multiple 2D conditional velocity posteriors. p. 1 While doable, this approach is
relatively slow and may also be unreliable if insufficient samples are used or if the
covariances in the data are very large. Here I show that, because velocity is linearly proportional to proper motion at a given
distance, the 3D sampling can be reduced without approximation to a 1D sampling of
the distance posterior followed by single samplings of multiple 2D conditional velocity
posteriors. Here I show that this posterior is equivalent to a 1D posterior in distance combined with a sum of 2D Gaussian
distributions in velocity. This 1D+2D representation is both faster and more accurate
to sample, thereby making the computation of the kinegeometric posterior for a large
set of sources much easier.Where the paper says it
Where the paper says it
Derivation steps for sampling
The derivation proceeds through several steps to arrive at an efficient sampling scheme. Step 4 involves integrating the complete kinegeometric posterior over velocity to obtain the marginal posterior in distance, which is a non-trivial 1D function of distance that can be sampled numerically. p. 4 (12b)
This is a product of the geometric posterior with Q. All terms depend on distance, so
this is a non-trivial 1D function of distance that we can sample numerically. Step 3: The 3D joint distance–velocity posterior. We can now write the complete kinegeometric posterior (expression 3a) as
P
∗
(r, v | ϖ, µ, Σ3, p) = Nv(mv, Σv)
| {z }
f(v,r)
Q(r, ϖ, µ) P(ϖ | r, Σϖϖ)P(r | p)
| {z }
g(r)
(11a)
= P
∗
(v | r) P
∗
(r) (11b)
which defines f(v, r) and g(r): that these are equal to the conditional velocity posterior
P
∗
(v | r) and the marginal distance posterior P
∗
(r) respectively will be shown in the
next two steps.Where the paper says it
Where the paper says it
The sampling procedure
The summary of the procedure to sample the kinegeometric posterior involves drawing samples from two main steps. First, one must sample the 1D marginal distance posterior P∗(r ϖ, p) to provide a set of Ng samples. p. 6 4 Summary of the procedure to sample the kinegeometric posterior
1. Sample the 1D marginal distance posterior P
∗
(r | ϖ, p) (≡ g(r)) in equation 12b to
provide a set of Ng samples {ri}. Sample the 1D marginal distance posterior P
∗
(r | ϖ, p) (≡ g(r)) in equation 12b to
provide a set of Ng samples {ri}. For each sample ri
, the conditional posterior velocity PDF P
∗
(v | ri
, ϖ, µ, p) (≡
f(v, ri)) in equation 14 is a 2D Gaussian in velocity with mean and covariance
given by equation 10 (using also equations 5 and 2). For each sample ri
, the conditional posterior velocity PDF P
∗
(v | ri
, ϖ, µ, p) (≡
f(v, ri)) in equation 14 is a 2D Gaussian in velocity with mean and covariance
given by equation 10 (using also equations 5 and 2). The average of these Ng
Gaussians (equation 13c) gives the marginal 2D velocity posterior.Where the paper says it
Where the paper says it
Where the paper says it
Performance and comparison
The 1D+2D sampling method is demonstrated to be more accurate than directly sampling the 3D non-Gaussian joint distance–velocity posterior. Furthermore, this method was found to be 15 times faster to compute than the 3D MCMC method. p. 8 This demonstrates that the 1D+2D sampling method derived in this article is an accurate representation of the original 3D posterior. Furthermore, the 1D+2D method
was 15 times faster to compute than the 3D MCMC method. They are quite close to this, as also quantified by the median
and standard deviation written at the top of each panel. Deviations from the ideal delta
functions are presumably due to the finite number of samples used in both methods,
plus imperfect convergence and residual correlation in the MCMC.Where the paper says it
Where the paper says it
Conclusion
To compute the 3D distance–velocity kinegeometric posterior, one can draw Ng samples from the marginal 1D non-Gaussian distance posterior, followed by the same number of samples from the conditional velocity posterior, which is a 2D Gaussian mixture model. This approach is more accurate than directly sampling the 3D non-Gaussian joint distance–velocity posterior, and in practice requires far fewer posterior evaluations, making it faster. p. 8 6 Conclusion
To compute the 3D distance–velocity kinegeometric posterior we can draw Ng samples
from the marginal 1D non-Gaussian distance posterior, followed by the same number
of samples from the conditional velocity posterior, which is a 2D Gaussian mixture
model. This approach is more accurate than directly sampling the 3D non-Gaussian
joint distance–velocity posterior, and in practice requires far fewer posterior evaluations, making it faster. This approach is more accurate than directly sampling the 3D non-Gaussian
joint distance–velocity posterior, and in practice requires far fewer posterior evaluations, making it faster. MCMC could be used to sample the 1D distribution, but for
these posteriors, 1D sampling can be done more efficiently at higher accuracy with
other numerical methods.Where the paper says it
Where the paper says it
REFERENCE
Bailer-Jones C.A.L., 2023. Estimating distances from parallaxes. p. 0 Estimating distances from parallaxes. Computing the three-dimensional kinegeometric posterior with
one-dimensional sampling
Coryn A.Where the paper says it
Derflinger G., Hormann W., Leydold J., 2010. ¨ Random variate generation by numerical inversion when only the density is known
Foreman-Mackey D., Hogg D.W., Lang D., Goodman J., 2013. emcee: The MCMC Hammer
Gaia Collaboration, 2021. Gaia Early Data Release 3. p. 0 2018) and third (BailerJones et al. 2021, hereafter paper V) Gaia data releases (Gaia Collaboration 2018, 2021).Where the paper says it
Lindegren L., Hernandez J., Bombrun A., Klioner S., Bastian U., et al. 2018 ´ Gaia Data Release 2. p. 0 2018) and third (BailerJones et al. 2021, hereafter paper V) Gaia data releases (Gaia Collaboration 2018, 2021). 2018) and third (BailerJones et al. 2021, hereafter paper V) Gaia data releases (Gaia Collaboration 2018, 2021).Where the paper says it
Where the paper says it
Luri X., Brown A.G.A., Sarro L. M., Arenou F., Bailer-Jones C. p. 0 Computing the three-dimensional kinegeometric posterior with
one-dimensional sampling
Coryn A. Bailer-Jones
Max Planck Institute for Astronomy, Heidelberg
https://www.Where the paper says it
Queiroz A B A., Anders F., Santiago B X., Chiappini C., Steinmetz M., et al., 2018.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved sampling efficiency for kinegeometric posterior computation allows for faster inference on large datasets by reducing dimensionality from a three-dimensional joint distance-velocity space to a one-dimensional distance posterior followed by single samplings of multiple two-dimensional conditional velocity posteriors.
-
The system can compute the marginal 2D velocity posterior by drawing one sample from each of the Ng conditional velocity functions at representative distances, rather than attempting direct sampling of the full three-dimensional non-Gaussian joint distance-velocity posterior.
-
The AI system can characterize 3D posteriors by computing summary statistics such as means and variances, 1D quantiles on the distance and velocities, as well as correlation coefficients derived from the samples obtained via the 1D+2D method.
Abstract
When parallaxes are of poor quality, proper motions can improve stellar distance estimates using known constraints on stars' physical velocities. Formally this may be achieved via the kinegeometric posterior, a three-dimensional (3D) posterior in distance and velocity for given measured parallax and 2D proper motion. This posterior is the product of a covariant Gaussian likelihood in the parallax and proper motion with a distance and velocity prior that can be constructed from a Galaxy model. In earlier work, I showed how to infer this posterior for stars in Gaia DR3 by explicitly sampling it with a Markov Chain Monte Carlo method. Here I show that this posterior is equivalent to a 1D posterior in distance combined with a sum of 2D Gaussian distributions in velocity. This 1D+2D representation is both faster and more accurate to sample, thereby making the computation of the kinegeometric posterior for a large set of sources much easier.
Related papers
- On-sky demonstration of a vector Zernike wavefront sensor in a cascaded adaptive optics system
- Euclid Data Release 1 (DR1): Calibration strategy and observations
- The General Antiparticle Spectrometer (GAPS) Antarctic Balloon Payload
- Detection and excision of radio-frequency interference in radio arrays
- Unified matched-filter and excess-power pipeline for gravitational-wave detector characterization
- Ushering in sub-arcsecond angular resolution observatories in the far-infrared: The case for a double-Fourier interferometer