Pith. sign in

REVIEW 4 major objections 6 minor 35 references

Gaussian Process Methods for Very Large Astrometric Data Sets

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A scalable, binning-free Gaussian process method infers smooth Milky Way velocity and dispersion profiles from 833,808 Gaia DR3 stars and finds asymmetric dispersion features that signal vertical disequilibrium.

desk verdict Worth refereeing, but the disequilibrium claim is not yet supported by the uncertainty analysis. read the letter →

arxiv 2507.10317 v1 pith:R72JFESR submitted 2025-07-14 astro-ph.GA astro-ph.IM

Timothy Hapitas , Lawrence M. Widrow , Thavisha E. Dharmawardena , Daniel Foreman-Mackey This is my paper · ORCID
classification astro-ph.GAastro-ph.IM
keywords GaussianprocessregressionstochasticvariationalinferenceGaiaDR3MilkyWaykinematicsvelocitydispersiontensorphasespiralGalacticdisequilibriumheteroskedasticnoise
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to replace binning with a scalable, non-parametric way to measure how the Milky Way's stars move. The authors show that because the Stochastic Variational Gaussian Process Regression objective factorizes over individual stars, the velocity scatter (the dispersion tensor) can be treated as its own Gaussian Process alongside the mean velocity field, with computation scaling as $O(M^3)$ rather than $O(N^3)$. Applied to 833,808 Gaia DR3 Radial Velocity Survey stars in a solar-neighborhood column, the method yields differentiable profiles of the mean velocity and dispersion as functions of height $z$. The inferred dispersion profiles show asymmetric small-scale features across all three diagonal components, which the authors read as evidence that the vertical dynamics are in disequilibrium, with correlated structures near $|z|\sim0.5$ kpc interpreted as Gaia phase-spiral signatures. If correct, the method makes Jeans-equation analyses of very large astrometric surveys feasible without finite-differencing binned data.

What carries the argument

The load-bearing object is the factorized Evidence Lower Bound (ELBO) of Stochastic Variational Gaussian Process Regression (SVGPR), a sparse-GP plus variational-inference approximation in which the log-likelihood term separates into a sum over individual observations. Because the ELBO factorizes, the observation noise need not be independent Gaussian with a single variance; the log noise variance can itself be given a GP prior, so inference proceeds with two GPs: one for the latent velocity field (zero mean, RBF kernel) and one for the dispersion (tanh mean function fit to binned data, Rational Quadratic kernel), sharing a set of $M=1000$ inducing points and optimized by stochastic gradient descent on mini-batches. This factorization is what reduces the per-iteration cost from $O(N^3)$ to $O(M^3)$ and what allows the velocity dispersion to be learned as a function of position rather than assumed constant.

What would settle it

Take the mock-data pipeline and inject a coherent velocity correlation across a subset of stars, for example a stream-like overdensity with a common velocity offset, while keeping the true dispersion fixed; if the inferred dispersion profile develops spurious midplane-asymmetric dips at the stream's location, then the diagonal-noise assumption is the likely source of the reported disequilibrium features. A complementary observational check is to rerun the Gaia analysis after removing stars identified as members of known streams and moving groups and see whether the features at $|z|\sim0.5$ kpc persist.

Watch

Extended reading notes

Core claim

The paper's central claim is that the factorization property of the SVGPR evidence lower bound allows the input-dependent noise variance, here the velocity dispersion, to be modeled as a second Gaussian Process jointly with the latent mean velocity field, under a diagonal noise covariance. The authors exploit this to infer the mean velocity and the three diagonal components of the velocity dispersion tensor as smooth functions of Galactic height $z$ for a sample of $N\simeq8\times10^5$ stars from the Gaia DR3 Radial Velocity Survey. They report asymmetric features in all three diagonal dispersion components, with a dip near $z\simeq-0.5$ kpc mirrored by an enhancement at $z\simeq+0.5$ kpc, and interpret the correlated small-scale structure as evidence that the vertical dynamics of the Milky Way are in a state of disequilibrium, possibly related to the Gaia phase spiral. The method's computational cost is $O(M^3)$ with $M=1000$ inducing points, roughly a factor $10^6$ cheaper than exact Gaussian process regression on the full dataset.

Load-bearing premise

The whole analysis rests on assuming that the scatter in each velocity component is independent, uncorrelated Gaussian noise after transforming to Galactocentric coordinates, so that the full noise covariance is diagonal; if stars in the sample carry correlated velocities from streams, spiral arms, or survey systematics, those correlations get absorbed into the learned dispersion and can mimic the small-scale asymmetric features reported as disequilibrium.

Editorial extensions

If this is right

  • Velocity moments for samples of order $10^6$ stars become computable on a single GPU, so the method can scale to the full Gaia DR3 RVS catalog and future surveys without binning.
  • The inferred mean velocity and dispersion profiles are differentiable functions of $z$, so the Jeans equations can be applied directly instead of through finite differences of binned histograms.
  • The two-GP scheme generalizes to any regression problem of the form $y=f(x)+\varepsilon$ with input-dependent Gaussian noise, making it a general tool for heteroskedastic big-data regression.
  • If the dispersion asymmetries are real, the Milky Way's vertical structure is out of equilibrium, and the correlation of dispersion dips with phase-spiral crossing points at $|z|\sim0.5$ kpc ties the observed kinematic state to a known non-equilibrium feature.
  • The method's failure to recover $O(1)$ km/s fluctuations in the mock mean velocity suggests that small-scale mean-velocity structure is harder to extract than dispersion structure, which limits velocity-field mapping at fine scales unless the scale disparity is addressed.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A decisive test of the disequilibrium interpretation is to remove stars belonging to known streams, moving groups, or other spatially correlated structures and rerun the pipeline; if the asymmetric dispersion features vanish, they were artifacts of correlated noise folding into the learned noise GP rather than true disequilibrium.
  • The diagonal-noise assumption is the main vulnerability: any star-star velocity correlation from spiral arms, streams, or the Gaia scanning law that is not in the diagonal covariance can masquerade as small-scale dispersion structure, and injecting such correlations into the mock data would show whether the pipeline attributes them to the dispersion GP.
  • The phase-spiral interpretation makes a testable prediction: the dispersion features at $|z|\sim0.5$ kpc should move or change when the sample is split by age or metallicity, since the phase spiral winds with dynamical age; a static feature across populations would favor a different explanation.
  • The ad hoc uncertainty estimate by splitting the dataset could itself suppress real features that happen to be stronger in one half; replacing it with a posterior over the noise GP would both sharpen the error bars and let the data decide whether sub-midplane and above-midplane features are significantly different.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The manuscript presents an SVGPR-based method for jointly inferring the mean Galactocentric velocity field and the diagonal components of the velocity dispersion tensor from large astrometric catalogs. The key technical step is to exploit the factorization of the ELBO over observations so that an input-dependent noise variance can be represented by a second Gaussian process; after assuming a diagonal, uncorrelated noise covariance, the problem separates into three independent one-dimensional regressions in z. The method is validated on mock data with injected Gaussian features in the dispersion and then applied to about 8.3e5 Gaia DR3 RVS stars in a solar annulus, yielding smooth profiles of v_R, v_phi, v_z and their dispersions. The authors identify asymmetric small-scale features in all three dispersion profiles and interpret them as evidence of vertical disequilibrium and possibly the Gaia phase spiral.

Significance. The methodological contribution is potentially valuable: the two-GP construction is a clean realization of input-dependent noise within SVGPR, the ELBO derivation is coherent, and the claimed O(M^3) scaling with M << N offers a binning-free route to differentiable velocity moments. The mock tests demonstrate recovery of injected dispersion features and are a useful proof of concept. However, the headline physical claim rests on an uncertainty estimation procedure that the authors themselves label qualitative and ad hoc, and the mock tests do not include a null-experiment calibration. The paper is therefore best assessed as a promising method whose application to Gaia DR3 requires additional validation before the disequilibrium conclusion can be accepted.

major comments (4)
  1. [Section 5.2 and Section 7, Fig. 7] The dispersion confidence bands in Fig. 8 are obtained from the two-half data split shown in the left panel of Fig. 7, and the text justifies this choice over the four-quarter split because features in the latter 'begin to either fluctuate strongly or disappear'. This selection criterion is keyed to the persistence of the very features that are later interpreted as disequilibrium signatures, so the bands are not a calibrated uncertainty statement. Because the headline physical claim depends on the significance of these features, the paper should add a null-experiment calibration: generate mock data with a symmetric dispersion model (for example, the tanh mean of Eq. (28) with the same noise realization as the data but no injected asymmetry), run the full pipeline including the subset-splitting procedure with the same SGD settings, and quantify how often asymmetric features of the amplitude seen in Fig. 8 appear.
  2. [Section 6, Fig. 5] The eight runs in Fig. 5, which differ only in SGD mini-batch order, produce dispersion profiles with systematic offsets of order a few km/s relative to the ground truth. The amplitudes of the small-scale asymmetric features emphasized in Figs. 8 and 9 are comparable to this scatter, and the subset-based confidence bands of Section 5.2 do not include this optimization variability. The manuscript should report a quantitative comparison, for instance the RMS spread of the Fig. 5 runs versus the amplitude of the z near +/-0.5 kpc features in Fig. 9, and should either incorporate SGD-seed variability into the reported uncertainty or restrict the physical conclusions to features that robustly exceed it.
  3. [Section 3, Eq. (21)] The likelihood assumes a diagonal noise covariance with no correlations between stars or between velocity components after the Galactocentric transformation. As the authors state, this is an assumption; any unmodeled correlations from the Gaia scanning law, streams, spiral arms, or spatially coherent systematics will be folded into the learned noise GP, where they can masquerade as small-scale, mid-plane-asymmetric dispersion structure. The paper should test this by injecting spatially correlated noise into the mock data of Section 6 (for example, a correlated component with coherence length near 0.5 kpc) and demonstrating that the recovered dispersion features are not induced by the unmodeled correlation.
  4. [Section 5.2, Eq. (28)] The mean function parameters (A, B, z0, l0) of the dispersion GP are estimated by MLE from binned data and then held fixed during the variational inference. The uncertainty in these parameters is therefore not propagated into the dispersion posterior, and the GP only models fluctuations around this fixed baseline. Since the baseline already contains a tanh shape with a nonzero z0, the physical interpretation of the residual features in Fig. 9 depends on the fidelity of this baseline. The authors should either marginalize over the mean-function parameters (for example, by including them in the optimization with a proper prior) or demonstrate in mocks that the features of interest are insensitive to plausible changes in the fitted baseline.
minor comments (6)
  1. [Section 2.3, Eqs. (5)-(6)] The symbol p(u) is used for both the prior on the inducing variables and the variational distribution; rename the latter q(u) for consistency with Eq. (7).
  2. [Section 4, Eq. (25)] The radial interval is written as R0 - 0.25 pc <= R <= R0 + 0.25 pc; the units should presumably be kpc.
  3. [Section 5.2, Eq. (31)] The definition of s(z) uses sigma_i(z) inside the sum after subtracting sigma_i(z); the second occurrence should be the mean sigma_bar(z).
  4. [Section 5.1, final paragraph] The phrase 'the the RQ kernel' contains a duplicated definite article.
  5. [Appendix A, Tables 3 and 4] Tables 3 and 4 have identical captions ('for a batch ratio of 50 and 300 training steps'); one of them presumably refers to batch ratio 20, or the entries are mislabeled.
  6. [Figure 16 caption] The caption of Fig. 16 ends mid-sentence and should be completed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SVGPR two-GP derivation is self-contained; the fitted tanh dispersion mean is a prior mean, not a prediction, and the claimed asymmetric features are data-driven deviations from it.

full rationale

The derivation chain is self-contained. The objective is the standard SVGPR ELBO (Eq. 8); the extension to input-dependent noise (Section 2.4, Eq. 10-13) follows Goldberg et al. (1997) and preserves the Gaussian likelihood (Eq. 12); the Galactic model (Section 3) reduces the full noise covariance to a diagonal matrix (Eq. 21) and factorizes the likelihood (Eq. 24). None of these steps presuppose the paper's physical conclusions. The dispersion GP's mean function (Eq. 28) is fitted to binned data and held fixed, but it is a prior mean, not the prediction target; the GP posterior is free to deviate from it, and the abstract's 'asymmetric features' are precisely those deviations (residuals shown in Figure 9). The tanh form, being symmetric in |z-z0|, cannot by itself generate the claimed small-scale asymmetries; the mock tests in Section 6 inject features at z ~ ±400 pc and recover them, demonstrating that the pipeline does not merely return its mean function. The paper explicitly flags the uncertainty estimation as ad hoc (Section 5.2: 'in the absence of a rigorous approach'; Section 8: 'The lack of a rigorous approach for extracting the uncertainty...') and the two-half versus four-quarter split is a post-hoc robustness choice (Section 7), but these are calibration limitations, not circular reductions. Self-citations (Nelson & Widrow 2022 for background; Li & Widrow 2021 for the thin/thick-disk motivation of the mean function) are not load-bearing: removing them would not alter the ELBO, the likelihood, or the reported features. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no equation is equivalent to its inputs by construction.

Assumptions & free parameters 6 free parameters · 8 assumptions · 0 invented entities

No new physical entities are introduced; the Gaussian processes are statistical objects, and the phase-spiral interpretation refers to an already-known kinematic feature. The free parameters are dominated by the tanh mean functions fitted to the data and the kernel and variational parameters learned by ELBO optimization.

free parameters (6)
  • tanh dispersion mean parameters (A, B, z0, l0) = not stated; MLE per component
    Equation 28, fitted by maximum likelihood to the binned dispersion profile and held fixed during GP training. Four parameters per component, three components.
  • Velocity GP kernel hyperparameters (RBF amplitude lambda_v, length scale l_v) = optimized via ELBO
    Equation 29; per component (or shared per Section 3 statement, ambiguous).
  • Dispersion GP kernel hyperparameters (RQ amplitude lambda_sigma, length scale l_sigma, mixing alpha) = optimized via ELBO
    Equation 30; per component.
  • Variational parameters mu and S for both GPs = optimized via ELBO
    Section 2.3, Eq. 6; free entries of the variational Gaussian means and covariances.
  • Inducing point locations z_f and z_beta (M = 1000 each) = optimized in GPyTorch; initial placement not specified
    Section 5.2; number chosen as M=1000; the abstract says 'chosen in a principled way' but the procedure is not stated in the text.
  • SGD algorithmic hyperparameters (batch ratio, learning rates, training steps) = batch ratio 100, latent LR 1.0, dispersion LR 0.1, steps 300
    Appendix A grid search on mock data; these tune the optimization, not the model, but affect the results and convergence.
assumptions (8)
  • domain assumption The stellar velocity distribution at every position is a tri-variate Gaussian.
    Section 3: 'we assume that the velocity distribution at every position in the Galaxy is a tri-variate Gaussian'. Needed for the Gaussian likelihood for velocities; if the DF is non-Gaussian, the inferred 'dispersion' is not the true second moment.
  • domain assumption The velocity dispersion tensor is diagonal and the noise covariance has no off-diagonal or inter-star correlations.
    Section 3, Eq. 21: 'We assume that the dispersion tensor of the Milky Way is diagonal and that there are no correlations in the dispersion and latent velocity field across the different cartesian components.' This makes the likelihood factorize and is required for the SVGPR approach; streams and moving groups would violate it.
  • domain assumption The mean velocity GP prior is zero (after subtracting Galactic rotation), encoding midplane symmetry and equilibrium.
    Section 5.1: 'We build our velocity GP with a mean function m(z) = 0. This choice aligns with the null hypothesis that the Galaxy is in dynamical equilibrium.' The latent velocity features are deviations from this prior; if the true mean velocity has a nonzero large-scale profile, the GP must learn it from data.
  • ad hoc to paper The dispersion mean function is A + B tanh(|z - z0|/l0).
    Section 5.1, Eq. 28: chosen to reflect thin/thick disk behavior; parameters fitted to the data. The tanh functional form is an arbitrary parametric choice and is held fixed, so the dispersion GP only models residuals around it.
  • domain assumption RBF and RQ kernels adequately model the correlation structure of velocity and dispersion fields.
    Section 5.1: RBF for velocity, RQ for dispersion, chosen after preliminary tests. The GP results depend on this choice; features on scales shorter than the learned length scale cannot be recovered.
  • domain assumption Individual star measurements are independent draws with known per-star Gaussian uncertainties, and the Gaia measurement covariance can be treated as diagonal.
    Section 4 describes error transformation via galpy; Section 3 assumes diagonal noise covariance. The per-star measurement covariance is generally not diagonal after transformation, but off-diagonals are neglected.
  • domain assumption No selection effects: the kinematic sample is representative enough for the inferred features to reflect the Galaxy rather than the selection function.
    Section 8: 'we continue to adopt our initial assumption' despite noting kinematic quantities may be sensitive to sample incompleteness and color-magnitude cuts.
  • standard math The ELBO and variational approximation provide a faithful posterior for the latent and noise GPs.
    Section 2.3, based on Hensman et al. 2013, 2015; standard SVGPR theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Gaussian Process Methods for Very Large Astrometric Data Sets." pith.science (2026). https://pith.science/paper/R72JFESR

@misc{pith2026250710317,
  author       = {Pith},
  title        = {Pith review of: Gaussian Process Methods for Very Large Astrometric Data Sets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R72JFESR}},
  note         = {Machine review of arXiv:2507.10317}
}
read the original abstract

We present a novel non-parametric method for inferring smooth models of the mean velocity field and velocity dispersion tensor of the Milky Way from astrometric data. Our approach is based on Stochastic Variational Gaussian Process Regression (SVGPR) and provides an attractive alternative to binning procedures. SVGPR is an approximation to standard GPR, the latter of which suffers severe computational scaling with N and assumes independently distributed Gaussian Noise. In the Galaxy however, velocity measurements exhibit scatter from both observational uncertainty and the intrinsic velocity dispersion of the distribution function. We exploit the factorization property of the objective function in SVGPR to simultaneously model both the mean velocity field and velocity dispersion tensor as separate Gaussian Processes. This achieves a computational complexity of O(M^3) versus GPR's O(N^3), where M << N is a subset of points chosen in a principled way to summarize the data. Applied to a sample of ~8 x 10^5 stars from the Gaia DR3 Radial Velocity Survey, we construct differentiable profiles of the mean velocity and velocity dispersion as functions of height above the Galactic midplane. We find asymmetric features in all three diagonal components of the velocity dispersion tensor, providing evidence that the vertical dynamics of the Milky Way are in a state of disequilibrium. Furthermore, our dispersion profiles exhibit correlated structures at several locations in |z|, which we interpret as signatures of the Gaia phase spiral. These results demonstrate that our method provides a promising direction for data-driven analyses of Galactic dynamics.

Figures

Figures reproduced from arXiv: 2507.10317 by the authors.

Figure 1
Figure 1. Flowchart illustrating the steps of our SVGPR inference procedure. Light blue boxes indicate the pipeline’s inputs and outputs. Orange ovals represent data prepro￾cessing steps before passing to GPyTorch. Blue diamonds denote inference tasks executed directly within GPyTorch. Green boxes correspond to our GP models, and grey boxes indicate intermediate results. Next, we instantiate the SGD optimizer provided by GPyT… view at source ↗
Figure 2
Figure 2. Mock latent vertical velocity data constructed to match the qualitative features seen in the vertical velocity of solar neighbourhood stars. Left panel: Vertical mean veloc￾ity of stars obtained from binning (black dots) and our mock mean profile obtained from a simple GP fit (green curve). Right panel: Binned vertical dispersion profile (black dots) and our mock z-dependent dispersion built from fitting equa￾tion 2… view at source ↗
Figure 3
Figure 3. shows the results of applying our SVGPR inference pipeline to the mock data presented in figure 2. The top row of plots depict the GP fit to the latent velocity while the bottom row shows the prediction for the dispersion profile. The optimizer is able to discern the correct amplitude for the latent velocity field, but fails to extract the small scale fluctuations that appear on scales ≲ O(1) km/s. Close to the midp… view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Same as figure 3, where we now vary the number of inducing points for both GPs. The remaining algorithmic parameters were held fixed at the values provided in table 1. 1 0 1 z [kpc] 2 1 0 1 2 3 vz [k m / s] 1 0 1 z [kpc] 0.10 0.05 0.00 0.05 0.10 Normalized residuals 1 …
Figure 5
Figure 5. Figure 5: Same as figure 3, where we now show the resulting predictions for 8 different random mini-batch selections by the SGD optimizer. Once again, all algorithmic parameters were kept fixed at the values presented in table 1. Given this finding, the direct interpretation of …
Figure 6
Figure 6. Figure 6: SVGPR predictions for the latent vertical veloc￾ity and velocity dispersion for our artificially inflated mock data. Left panel: regression prediction for the mean vertical velocity (red solid curve) and corresponding 95 % confidence region (blue band). Right panel: SV…
Figure 7
Figure 7. Figure 7: Illustration of our scheme for estimating the sta￾tistical uncertainty on our learned dispersion profiles. Left panel: SVGPR fits for the dispersion for two subsets of the total training set (red-dashed curves). Each subset corre￾sponds to exactly half of the training …
Figure 9
Figure 9. Figure 9: depicts the resulting residuals. Paying atten￾tion to the region within ±1 kpc of the midplane of the disc, all of the features appear to possess a degree of correlation across σvz , σvϕ , and σvR . The the strong dip at z ∼ −0.5 kpc is echoed across all of the disper￾…
Figure 10
Figure 10. Figure 10: SVGPR fits to our mock latent vertical velocity field presented in the left column of figure 2 for 300 training steps, a batch ratio of 20.0, and all combinations of learning rate values presented in table 2. The red curves depict the SVGPR prediction for the latent v…
Figure 11
Figure 11. Figure 11: SVGPR fits to our mock vertical velocity dispersion profile presented in the right column of figure 2 for 300 training steps, a batch ratio of 20.0, and all combinations of learning rate values presented in table 2. The red curves depict the SVGPR prediction for the d…
Figure 12
Figure 12. Figure 12: Same as in figure 10, but this time for a batch ratio of 50.0 [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]
Figure 13
Figure 13. Figure 13: Same as in figure 11, but this time for a batch ratio of 50.0 [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Same as in figure 10, but this time for a batch ratio of 100.0 [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Same as in figure 11, but this time for a batch ratio of 100.0 Figures 10, 12, and 14 display the results for the mock latent velocity fits, and figures 11, 13, and 15 show the fits for the dispersion process. In each of these figures, we show all the combinations of …
Figure 16
Figure 16. Figure 16: SVGPR fit results for our mock latent velocity and velocity dispersion profiles obtained used a batch ratio of 100.0, 400 training steps, a disp These results indicate a disparity between the optimization behaviour of the latent and dispersion processes. The fact that…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 15 canonical work pages

  1. [1]

    2022, Astronomy &amp; Astrophysics, 657, L12, 10.1051/0004-6361/202142465

    Abuter, R., Aimar, N., Amorim, A., et al. 2022, Astronomy &amp; Astrophysics, 657, L12, 10.1051/0004-6361/202142465

    doi:10.1051/0004-6361/202142465 2022

  2. [2]

    2023, Astronomy &amp; Astrophysics, 673, A115, 10.1051/0004-6361/202245518

    Antoja, T., Ramos, P., García-Conde, B., et al. 2023, Astronomy &amp; Astrophysics, 673, A115, 10.1051/0004-6361/202245518

    doi:10.1051/0004-6361/202245518 2023

  3. [3]

    Bahcall , J. N. 1984, , 276, 169, 10.1086/161601

    doi:10.1086/161601 1984

  4. [4]

    2018, Monthly Notices of the Royal Astronomical Society, 482, 1417–1425, 10.1093/mnras/sty2813

    Bennett, M., & Bovy, J. 2018, Monthly Notices of the Royal Astronomical Society, 482, 1417–1425, 10.1093/mnras/sty2813

    doi:10.1093/mnras/sty2813 2018

  5. [5]

    2011, Galactic Dynamics: Second Edition (Princeton University Press), 10.2307/j.ctvc778ff

    Binney, J., & Tremaine, S. 2011, Galactic Dynamics: Second Edition (Princeton University Press), 10.2307/j.ctvc778ff

    doi:10.2307/j.ctvc778ff 2011

  6. [6]

    2023, Astronomy &amp; Astrophysics, 674, A7, 10.1051/0004-6361/202243685

    Blomme, R., et al. 2023, Astronomy &amp; Astrophysics, 674, A7, 10.1051/0004-6361/202243685

    doi:10.1051/0004-6361/202243685 2023

  7. [7]

    2015, The Astrophysical Journal Supplement Series, 216, 29, 10.1088/0067-0049/216/2/29

    Bovy, J. 2015, The Astrophysical Journal Supplement Series, 216, 29, 10.1088/0067-0049/216/2/29

    doi:10.1088/0067-0049/216/2/29 2015

  8. [8]

    2013, The Astrophysical Journal, 779, 115, 10.1088/0004-637x/779/2/115

    Bovy, J., & Rix, H.-W. 2013, The Astrophysical Journal, 779, 115, 10.1088/0004-637x/779/2/115

    doi:10.1088/0004-637x/779/2/115 2013

Show all 35 references
  1. [9]

    2023, Astronomy &amp; Astrophysics, 674, A37, 10.1051/0004-6361/202243797

    Drimmel, R., et al. 2023, Astronomy &amp; Astrophysics, 674, A37, 10.1051/0004-6361/202243797

    2023 doi

  2. [10]

    W., Rix, H.-W., & Ness, M

    Eilers, A.-C., Hogg, D. W., Rix, H.-W., & Ness, M. K. 2019, The Astrophysical Journal, 871, 120, 10.3847/1538-4357/aaf648

    2019 doi

  3. [11]

    2021, tinygp: The tiniest of Gaussian process libraries, 0.2.3

    Foreman-Mackey, D. 2021, tinygp: The tiniest of Gaussian process libraries, 0.2.3. https://tinygp.readthedocs.io

    2021

  4. [12]

    W., Tremaine, S., Price-Whelan, A., & Shen, J

    Frankel, N., Hogg, D. W., Tremaine, S., Price-Whelan, A., & Shen, J. 2024, Iron Snails: non-equilibrium dynamics and spiral abundance patterns, arXiv, 10.48550/ARXIV.2407.07149

    2024 doi arXiv

  5. [13]

    2016, 10.48550/ARXIV.1609.04153

    Gaia Collaboration . 2016, 10.48550/ARXIV.1609.04153

    2016 doi arXiv

  6. [14]

    Gaia Collaboration , Vallenari, A., Brown, A. G. A., et al. 2023, Astron. Astrophys., 674, A1

    2023

  7. [15]

    R., Pleiss, G., Bindel, D., Weinberger, K

    Gardner, J. R., Pleiss, G., Bindel, D., Weinberger, K. Q., & Wilson, A. G. 2021, GPyTorch: Blackbox Matrix-Matrix Gaussian Process Inference with GPU Acceleration. 1809.11165

    2021 arXiv

  8. [16]

    1997, in Advances in Neural Information Processing Systems 10 (NIPS 1997) (MIT Press), 493--499

    Goldberg, P., Williams, C., & Bishop, C. 1997, in Advances in Neural Information Processing Systems 10 (NIPS 1997) (MIT Press), 493--499

    1997

  9. [17]

    2022, The Astrophysical Journal, 936, 103, 10.3847/1538-4357/ac86cd

    Guo, R., Shen, J., Li, Z.-Y., Liu, C., & Mao, S. 2022, The Astrophysical Journal, 936, 103, 10.3847/1538-4357/ac86cd

    2022 doi

  10. [18]

    Hensman, J., Fusi, N., & Lawrence, N. D. 2013, in Proceedings of the Twenty-Ninth Conference on Uncertainty in Artificial Intelligence, UAI'13 (Arlington, Virginia, USA: AUAI Press), 282–290

    2013

  11. [19]

    2015, in Proceedings of Machine Learning Research, Vol

    Hensman, J., Matthews, A., & Ghahramani, Z. 2015, in Proceedings of Machine Learning Research, Vol. 38, Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics, ed. G. Lebanon & S. V. N. Vishwanathan (San Diego, California, USA: PMLR), ...

    2015

  12. [20]

    D., Blei, D

    Hoffman, M. D., Blei, D. M., Wang, C., & Paisley, J. 2013, Journal of Machine Learning Research, 14, 1303. http://jmlr.org/papers/v14/hoffman13a.html

    2013

  13. [21]

    2018, Astronomy &amp; Astrophysics, 616, A11, 10.1051/0004-6361/201832865

    Katz, D., et al. 2018, Astronomy &amp; Astrophysics, 616, A11, 10.1051/0004-6361/201832865

    2018 doi

  14. [22]

    2023, Astronomy &amp; Astrophysics, 674, A5, 10.1051/0004-6361/202244220

    Katz, D., Sartoretti, P., Guerrier, A., et al. 2023, Astronomy &amp; Astrophysics, 674, A5, 10.1051/0004-6361/202244220

    2023 doi

  15. [23]

    1989, , 239, 605, 10.1093/mnras/239.2.605

    Kuijken , K., & Gilmore , G. 1989, , 239, 605, 10.1093/mnras/239.2.605

    1989 doi

  16. [24]

    Li, H., & Widrow, L. M. 2021, Monthly Notices of the Royal Astronomical Society, 503, 1586–1598, 10.1093/mnras/stab574

    2021 doi

  17. [25]

    2019, Grid Search, Random Search, Genetic Algorithm: A Big Comparison for NAS, arXiv, 10.48550/ARXIV.1912.06059

    Liashchynskyi, P., & Liashchynskyi, P. 2019, Grid Search, Random Search, Genetic Algorithm: A Big Comparison for NAS, arXiv, 10.48550/ARXIV.1912.06059

    2019 doi arXiv

  18. [26]

    Luri, X., Brown, A. G. A., Sarro, L. M., et al. 2018, Astronomy &amp; Astrophysics, 616, A9, 10.1051/0004-6361/201832964

    2018 doi

  19. [27]

    Nelson, P., & Widrow, L. M. 2022, Monthly Notices of the Royal Astronomical Society, 516, 5429–5439, 10.1093/mnras/stac2594

    2022 doi

  20. [28]

    2007, in Advances in Neural Information Processing Systems, ed

    Rahimi, A., & Recht, B. 2007, in Advances in Neural Information Processing Systems, ed. J. Platt, D. Koller, Y. Singer, & S. Roweis, Vol. 20 (Curran Associates, Inc.). https://proceedings.neurips.cc/paper_files/paper/2007/file/013a006f03dbc5392effeb8f18fda755-Paper.pdf

    2007

  21. [29]

    E., & Williams, C

    Rasmussen, C. E., & Williams, C. K. I. 2005, Gaussian processes for machine learning, Adaptive Computation and Machine Learning series (MIT Press)

    2005

  22. [30]

    Sarkar, S., & Jog, C. J. 2019, Monthly Notices of the Royal Astronomical Society, 492, 628–633, 10.1093/mnras/stz3512

    2019 doi

  23. [31]

    2005, in Advances in Neural Information Processing Systems, ed

    Snelson, E., & Ghahramani, Z. 2005, in Advances in Neural Information Processing Systems, ed. Y. Weiss, B. Sch\" o lkopf, & J. Platt, Vol. 18 (MIT Press). https://proceedings.neurips.cc/paper_files/paper/2005/file/4491777b1aa8b5b32c2e8666dbe1a495-Paper.pdf

    2005

  24. [32]

    2009, in Proceedings of Machine Learning Research, Vol

    Titsias, M. 2009, in Proceedings of Machine Learning Research, Vol. 5, Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics, ed. D. van Dyk & M. Welling (Hilton Clearwater Beach Resort, Clearwater Beach, Florida USA: PMLR), 567--574. ht...

    2009

  25. [33]

    2022, The Astrophysical Journal, 942, 12, 10.3847/1538-4357/aca27c

    Wang, H.-F., Chrobáková, ., López-Corredoira, M., & Sylos Labini, F. 2022, The Astrophysical Journal, 942, 12, 10.3847/1538-4357/aca27c

    2022 doi

  26. [34]

    2000, in Advances in Neural Information Processing Systems, ed

    Williams, C., & Seeger, M. 2000, in Advances in Neural Information Processing Systems, ed. T. Leen, T. Dietterich, & V. Tresp, Vol. 13 (MIT Press). https://proceedings.neurips.cc/paper_files/paper/2000/file/19de10adbaa1b2ee13f77f679fa1483a-Paper.pdf

    2000

  27. [35]

    G., & Nickisch, H

    Wilson, A. G., & Nickisch, H. 2015, Kernel Interpolation for Scalable Structured Gaussian Processes (KISS-GP), arXiv, 10.48550/ARXIV.1503.01057

    2015 doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.