betaFeedback

WorksPhysics and Astronomy

New optimisations for exact, scalable 2D Gaussian process inference, with applications to astronomical time-series

Abstract

Astronomical datasets often consist of a grid of datapoints in two or more dimensions, such as images, spectroscopic time-series, and imaging time-series. Gaussian processes (GPs) are a powerful tool for dealing with correlated noise in these data, arising from both astrophysical and instrumental systematics. Despite this, popular GP optimisations such as celerite are largely restricted to 1D datasets. Extensions to 2D data typically come with limited flexibility or poor scaling. As a motivating example, this limitation is significant for transiting exoplanet spectroscopy with JWST, as current analyses typically assume each spectroscopic light curve is statistically independent, despite many potential sources of wavelength-correlation, including detector readout noise and stellar granulation. This introduces a potential bias in atmospheric retrievals which is typically ignored. To solve this problem, we present new, exact GP optimisations implemented in the JAX-based package luas, which extend 1D GP optimisations to multidimensional data using Kronecker product algebra. These GP optimisations are highly flexible, permit efficient analysis of 2D datasets containing millions of datapoints, and in many cases have a cheaper cost than transit light curve modelling. These optimisations are key to pushing the limits of exoplanet spectroscopy to smaller targets around bright host stars, as we increasingly hit instrumental and astrophysical noise floors. We suggest further applications to the analysis of ground-based high resolution spectroscopic time-series, extreme precision radial velocities, and high-contrast imaging, drawing on the connection that certain 2D GP models represent a probabilistic generalisation of PCA. Applications may even extend to machine learning and geostatistics, as we present novel optimisations for the Linear Mixing Model and the Linear Model of Coregionalisation.

Introduction

Gaussian processes (GPs) are a non-parametric fitting tool, which makes them very useful at modelling signals we cannot describe explicitly. They have therefore seen widespread use across many areas of astronomy, such as modelling stellar activity (Aigrain et al. [3]; Rajpaul et al. [97]), asteroseismology and stellar granulation (Brewer and Stello 2009; Foreman-Mackey et al. [42]), brown dwarf variability (Vos et al. [124]), gravitational lensing (Press et al. [94]), pulsar timing delays (Haasteren and Vallisneri [120]; Susobhanan and van Haasteren [111]), atmospheric retrievals (Rotman et al. [99]), supernova light curves (Lochner et al. [80]), and many more (see Aigrain and Foreman-Mackey [2] for a review). Outside of astronomy, they have also seen very diverse applications, including their original application for geostatistics (Krige [77]), and later uses within epidemiology (Diggle et al. [33]), ecology (Dovers et al. [35]), genetics (Rakitsch et al. [98]; Velten et al. [121]), and neuroimaging (Cai et al. [20]).

This work has mainly been motivated by their applications for modelling correlated noise in exoplanet transit time-series, originally introduced in Gibson et al. [48]. Noise which is correlated in time, often appearing as ‘bumps’ in the light curves, can be referred to as either correlated noise, red noise, or systematics, while noise which is independent in time (e.g. due to photon noise) is referred to as ‘white noise’. Correlated noise may arise from a variety of sources, typically split between instrumental sources (e.g. pointing instabilities, thermal breathing), and astrophysical sources (e.g. stellar granulation, asteroseismology and stellar activity). The key reason why we may want to joint-fit transit observations with both a deterministic transit model and a GP for modelling systematics, is that the GP can marginalise over an infinite space of possible systematics models and weight towards the best-fitting ones, which propagates the uncertainty introduced by complex noise sources into our measured transit depth uncertainties. Gibson [47] demonstrated using simulated data that GPs can produce robust results despite contamination from a wide variety of systematics models, and are often more robust than common alternatives, such as marginalising over a set of detrending vectors.

While Gaussian processes are typically applied to one-dimensional datasets (e.g. a time-series of flux values), there are many cases where it is useful to apply them to 2D datasets, modelling correlations across two input dimensions, such as wavelength and time. Despite being a standard approach within exoplanet atmospheres, fitting transit light curves one wavelength band at a time ignores the presence of wavelength-correlated noise, which can arise from effects such as detector readout noise or stellar granulation (e.g. Schlawin et al. [105]; Gordon et al. [53]). Fortune et al. [44] demonstrated using simulated data that simultaneously fitting full spectroscopic time-series with a 2D Gaussian process can produce more robust transmission spectra in the presence of both wavelength- and time-correlated systematics. This could enable us to reliably analyse spectroscopic time-series at higher spectroscopic-resolution, while accounting for wavelength correlation due to instrumental effects such as interpixel capacitance, and 1/f readout noise (e.g. Schlawin et al. [105]; Ahrer et al. [1]). In the case of exoplanet detection, multi-band transit fitting (Gordon et al. [53]), as well as joint-fitting radial velocities with multiple stellar activity indicators (Rajpaul et al. [97]), have been shown to more reliably recover exoplanet signals in the presence of stellar variability. Additional uses of GPs applied to multi-dimensional datasets include joint-fitting pixel light curves for eclipse photometry (Fortune et al. [45]), modelling pulsar profile variations (Keith et al. [71]), and fitting multi-band supernova light curves (Villar et al. [122]).

Unfortunately, GPs can be expensive to scale to large datasets, especially data with multiple dimensions. In the most general case, GP log-likelihood evaluations scale as O(N3)\mathcal{O}(N^{3}) in runtime and O(N2)\mathcal{O}(N^{2}) in memory for NN datapoints. Fortunately, plenty of optimisations exist with significantly improved scaling in both runtime and memory, which either approximate log-likelihood calculations, or use a restricted set of kernel functions which have special mathematical structure. A popular example of the latter is CELERITE, a GP optimisation developed to fit one-dimensional inputs (Foreman-Mackey et al. [42]), which has been used extensively for modelling systematics in transit light curves (e.g. Espinoza et al. [40]; Barros et al. [10]; Alderson et al. [4]), as well as for modelling stellar activity, granulation and oscillations (e.g. Angus et al. [8]; Morris et al. [89]). In fact, CELERITE is a special case of a broader class of GP optimisations which scale linearly on 1D (sortable) inputs. Quasiseparable kernels, as implemented in TINYGP (Foreman-Mackey et al. [43]), S+LEAF matrices from the package SPLEAF (Delisle et al. [30, 31]), and state-space models (Hartikainen and Särkkä [59]; Jordán et al. [68]) implemented in SMOLGP (Rubenzahl et al. [101]), can all be viewed as generalisations of CELERITE which can implement all CELERITE kernels, plus exact implementations of Matérn half-integer kernels, GP derivatives, and banded covariance matrices.

Various GP optimisations have been introduced for 2D ‘rectangular’ datasets, such as a spectroscopic time-series, images, parallel time-series of stellar activity indicators, etc. The various 2D GP optimisations introduced include extensions of CELERITE (Gordon et al. [53]; Keith et al. [71]), extensions of state-space models (Särkkä et al. [104]), and various approximations including inducing point formulations (e.g. Snelson and Ghahramani [108]; Wilson and Nickisch [130]) and HODLR (Ambikasaran et al. [7]), not all of which require a complete rectangular dataset. Some very flexible approaches work by exploiting Kronecker product algebra, which use clever linear transformations to split up covariance matrix decomposition into separate covariance matrices in each dimension, such as in work by Saatchi [102], Rakitsch et al. [98], and Brunisma et al. (2020). These algorithms do tend to be restricted to complete rectangular data without missing values, but non-uniform spacing in each dimension is allowed. In this work, we expand on such optimisations to include a much broader range of kernel functions, as well as identifying which algorithms can be combined with common 1D GP optimisations, including CELERITE, to achieve excellent scaling in one or multiple dimensions, while log-likelihood calculations remain exact.

These optimisations could be relevant to a wide range of exoplanet applications, including transit spectroscopy at both low- and high-resolution, as well as extreme precision radial velocities, and high-contrast imaging. In the discussion section, we demonstrate the close connections between these methods and popular data-driven algorithms like PCA and SysRem, but implemented as a likelihood-based approach which marginalises over components, can include smoothing priors on components, and features a principled treatment of noise. Of relevance outside of astronomy, where 2D GPs may instead be referred to as spatiotemporal GPs, multi-output GPs, or multi-task GPs, we provide a new, exact optimisation for the Linear Mixing Model, which combined with Gordon et al. [53] and quasiseparable kernels, extends the Orthogonal Instantaneous Linear Mixing Model (OILMM) from Bruinsma et al. [19] to drop the orthogonality requirement for the mixing matrix, while maintaining a linear runtime scaling in the number of datapoints.

Each optimisation identified has been implemented and validated in a new update to our open-source Python package LUAS, with tutorials and extensive documentation available1.

Methods

Introduction to Gaussian processes

Gaussian processes provide a very versatile method for calculating the mean and uncertainty of deterministic model parameters, given a dataset containing complex noise sources. As an example, suppose that we have a single transit light curve with time-correlated noise, where the data can be described as:

y=μ(θ,t)+fsys+ε(1)\mathbf{y} = \boldsymbol{\mu}(\boldsymbol{\theta},\mathbf{t}) + \mathbf{f}_{\mathrm{sys}} + \boldsymbol{\varepsilon} \tag*{(1)}

where μ(θ,t)\boldsymbol{\mu}(\boldsymbol{\theta},\mathbf{t}) is an analytical model of the transit light curve observed at times t\mathbf{t}, and parameterised by model parameters θ\boldsymbol{\theta}, fsys\mathbf{f}_{\mathrm{sys}} is an unknown flux time-series due to some form of systematics (e.g. stellar variability), and ε\boldsymbol{\varepsilon} is the observed white noise arising from Poisson processes like photon counting statistics, with ε∼N(0,Σ)\boldsymbol{\varepsilon} \sim\mathcal{N}(\mathbf{0},\boldsymbol{\Sigma}), where the covariance matrix describing the white noise Σ\boldsymbol{\Sigma} is diagonal. μ(θ,t)\boldsymbol{\mu}(\boldsymbol{\theta},\mathbf{t}) is typically referred to as the ‘mean function’, and for transits is often modelled following Mandel and Agol [85]. For any fixed mean function and systematics model, the likelihood function (the probability of the data given the model) is simply the probability of generating ε\boldsymbol{\varepsilon} from Σ\boldsymbol{\Sigma}, which is given by the Gaussian probability density function:

p(D∣θ,t,fsys)=1(2π)N∣Σ∣exp⁡(−12εTΣ−1ε)(2)p(\mathbf{D} \mid\boldsymbol{\theta},\mathbf{t},\mathbf{f}_{\mathrm{sys}}) = \frac{1}{\sqrt{(2\pi)^{N} \lvert\boldsymbol{\Sigma} \rvert}} \exp\left(-\frac{1}{2}\boldsymbol{\varepsilon}^{T}\boldsymbol{\Sigma}^{-1}\boldsymbol{\varepsilon}\right) \tag*{(2)}

which after taking the logarithm (useful for numerical stability) yields

log⁡p(D∣θ,t,fsys)=−12εTΣ−1ε−12log⁡∣Σ∣−N2log⁡2π.(3)\log p(\mathbf{D} \mid\boldsymbol{\theta},\mathbf{t},\mathbf{f}_{\mathrm{sys}}) = -\frac{1}{2}\boldsymbol{\varepsilon}^{T}\boldsymbol{\Sigma}^{-1}\boldsymbol{\varepsilon} - \frac{1}{2}\log\lvert\boldsymbol{\Sigma} \rvert- \frac{N}{2}\log2\pi. \tag*{(3)}

The systematics model may in reality describe the host star flux varying in time due to chaotic processes such as stellar granulation, it could describe variations in the throughput of the telescope, or it could be a combination of multiple processes varying over time which we may have no way to analytically model. To span this full space of models, we could fit for every element of fsysf_{\mathrm{sys}}, introducing NN free parameters for NN datapoints, but doing this without any constraints would completely overfit our data. So we add a prior constraint on fsysf_{\mathrm{sys}} which penalises complexity, normally we impose that it should vary with some amplitude hh, and it should have a correlation timescale it changes over, with nearby points in time having similar values, and widely separated points in time being almost fully independent. We can implement this by choosing a multivariate Gaussian prior fsys∼N(0,Ksys)f_{\mathrm{sys}} \sim\mathcal{N}(0, K_{\mathrm{sys}}), where the covariance matrix KsysK_{\mathrm{sys}} is calculated using a kernel function which describes the covariance between each pair of points. For example, a common choice of kernel function is the exponential kernel

kℓ(t,t′)=h2exp⁡(−∣t−t′∣ℓt),(4)k_{\ell}(t,t') = h^2 \exp\left(-\frac{|t-t'|}{\ell_t}\right), \tag*{(4)}

where the covariance between two points decays exponentially with their separation in time, parameterised by the timescale ℓt\ell_t. The kernel function can be thought of as a ‘smoothness prior’ on our systematics model fsysf_{\mathrm{sys}}, i.e. whatever process is the cause of these systematics, we assume that it is continuous, varying over some characteristic time scale ℓt\ell_t, and with an average amplitude hh.

Now that we have this smoothing prior on fsysf_{\mathrm{sys}}, we can freely fit NN parameters to NN datapoints. However, if we optimised fsysf_{\mathrm{sys}} to its best-fit values and then kept it fixed, we would essentially be assuming that we have perfectly modelled out the systematics without any error, but the presence of white noise introduces uncertainty in the exact systematics model. Instead, we marginalise over all of the parameter space fsys∈RNf_{\mathrm{sys}} \in\mathbb{R}^{N}, weighting towards the best-fitting models. This could be performed by marginalising all NN parameters of fsysf_{\mathrm{sys}} in an MCMC, while simultaneously inferring our model parameters θ\boldsymbol{\theta}. Thankfully, GPs offer a much more efficient approach, since fsysf_{\mathrm{sys}} consists of linear parameters with the multivariate Gaussian prior p(fsys)=N(0,Ksys)p(f_{\mathrm{sys}}) = \mathcal{N}(0, K_{\mathrm{sys}}), we can exactly marginalise over it analytically (e.g. Luger et al. 2017). The result of integrating Equation (3) over all values of fsysf_{\mathrm{sys}} results in the marginal log-likelihood

log⁡p(D∣θ,t)=−12rTK−1r−12log⁡∣K∣−N2log⁡(2π),(5)\log p(D|\boldsymbol{\theta},t) = -\frac{1}{2}\boldsymbol{r}^{T}K^{-1}\boldsymbol{r} - \frac{1}{2}\log|K| - \frac{N}{2}\log(2\pi), \tag*{(5)}

where we have written our GP covariance matrix as K=Ksys+ΣK = K_{\mathrm{sys}} + \Sigma, and our data residuals as r=y−μ(θ,t)\boldsymbol{r} = \boldsymbol{y} - \boldsymbol{\mu}(\boldsymbol{\theta},t). Combining this marginalised log-likelihood with Bayesian inference methods (e.g. MCMC, Nested Sampling) allows us to propagate the uncertainty introduced by the presence of unknown systematics into our transit model parameters θ\boldsymbol{\theta}, within a statistically rigorous and interpretable framework, and without sampling a huge number of free parameters.

If the constraint on the systematics model fsysf_{\mathrm{sys}} is desired, it is described by the GP predictive mean and predictive covariance fsys∼N(μfsys,Kfsys)f_{\mathrm{sys}} \sim\mathcal{N}(\boldsymbol{\mu}_{f_{\mathrm{sys}}}, K_{f_{\mathrm{sys}}}), as given by the equations:

μfsys=KsysK−1r(6)\boldsymbol{\mu}_{f_{\mathrm{sys}}} = K_{\mathrm{sys}}K^{-1}\boldsymbol{r} \tag*{(6)}
Kfsys=Ksys−KsysTK−1Ksys.(7)K_{f_{\mathrm{sys}}} = K_{\mathrm{sys}} - K_{\mathrm{sys}}^{T}K^{-1}K_{\mathrm{sys}}. \tag*{(7)}

This can be used to visualise the most strongly weighted systematics models the GP is fitting, and in fact the distribution of systematics can even be predicted for unseen observations using similar formulae (see Appendix § 13). The reason why we can get a constraint on unseen data is because the correlation described by the kernel function restricts how much the systematics can deviate for predicted locations close to the observed data, which should give a robust predictive distribution if our kernel functions are an appropriate model of the systematics. The predictive mean and covariance on unseen data can be used to perform cross-validation tests, which can help to validate the accuracy of a GP model.

Implementing GP Inference

Performing inference with GPs typically involves calculating the log-likelihood function in Equation (5) many times (∼105−106\sim10^{5} - 10^{6} times for MCMC inference), which requires calculating both rTK−1r\boldsymbol{r}^{T}K^{-1}\boldsymbol{r}, and the log-determinant log⁡∣K∣\log|K| each time. There are various approaches to do this depending on the problem, and methods like Cholesky decomposition can avoid explicitly calculating K−1K^{-1}.

Using Cholesky Decomposition

A common approach is to calculate the Cholesky decomposition K=LLTK = LL^{T}, where LL is a lower triangular matrix. Since K−1=L−TL−1K^{-1} = L^{-T}L^{-1}, we can calculate:

rTK−1r=rT(L−1)TL−1r,(8)\boldsymbol{r}^{T}K^{-1}\boldsymbol{r} = \boldsymbol{r}^{T}(L^{-1})^{T}L^{-1}\boldsymbol{r}, \tag*{(8)}
=r~Tr~,(9)= \tilde{\boldsymbol{r}}^{T}\tilde{\boldsymbol{r}}, \tag*{(9)}

where r~=L−1r\tilde{\boldsymbol{r}} = L^{-1}\boldsymbol{r}. We can use the fact that LL is lower triangular to efficiently solve the linear system Lr~=rL\tilde{\boldsymbol{r}} = \boldsymbol{r} using an algorithm known as forward-substitution, which scales as O(N2)\mathcal{O}(N^{2}) for a dense triangular solve. The log-determinant is simply calculated as log⁡∣K∣=∑i2log⁡(Lii)\log|K| = \sum_i 2\log(L_{ii}).

Using Eigendecomposition

Cholesky decomposition is not the only matrix decomposition that can be used for GP log-likelihood calculations. We could instead calculate the eigendecomposition K=QΛQTK = Q\Lambda Q^{T}, where QQ is a matrix of eigenvectors of KK, and Λ\Lambda is a diagonal matrix formed using the corresponding eigenvalues of those eigenvectors. While typically more expensive than Cholesky decomposition (see benchmarking in Figure 1), eigendecomposition can be used to calculate the log-likelihood in a similar way

rTK−1r=rT(Λ−12QT)T(Λ−12QT)r,(10)\boldsymbol{r}^{T}K^{-1}\boldsymbol{r} = \boldsymbol{r}^{T}(\Lambda^{-\frac{1}{2}}Q^{T})^{T}(\Lambda^{-\frac{1}{2}}Q^{T})\boldsymbol{r}, \tag*{(10)}
=r~Tr~,(11)= \tilde{\boldsymbol{r}}^{T}\tilde{\boldsymbol{r}}, \tag*{(11)}

now with r~=Λ−12QTr\tilde{\boldsymbol{r}} = \Lambda^{-\frac{1}{2}}Q^{T}\boldsymbol{r}, and the log-determinant is simply log⁡∣K∣=∑ilog⁡Λii\log|K| = \sum_i \log\Lambda_{ii}.

Using Scalable 1D Approaches

Despite their performance differences, Cholesky decomposition and eigendecomposition both scale as O(N3)\mathcal{O}(N^{3}) in runtime and O(N2)\mathcal{O}(N^{2}) in memory for a general covariance matrix. These approaches can always be used for any valid covariance matrix, regardless of kernel function choice or the dimensions of the dataset. However, if we build the covariance matrix using CELERITE terms, the Cholesky decomposition can be computed in O(NJ2)\mathcal{O}(NJ^{2}) time, forward substitution can be applied in O(NJ)\mathcal{O}(NJ) time, and the memory cost also scales as O(NJ)\mathcal{O}(NJ) (Foreman-Mackey et al. 2017), where JJ refers to the number of CELERITE terms in the kernel function, which can be expressed as

k(t,t′)=∑j=1Je−cj∣t−t′∣[ajcos⁡(dj∣t−t′∣)+bjsin⁡(dj∣t−t′∣)].(12)k(t,t') = \sum_{j=1}^{J} e^{-c_j|t-t'|}\left[a_j\cos(d_j|t-t'|) + b_j\sin(d_j|t-t'|)\right]. \tag*{(12)}

CELERITE can implement the exponential kernel and the simple harmonic oscillator kernel, which can be very useful for modelling stellar activity [42, 8, 89].

Runtime to decompose matrices using an M4 MacBook Air and an NVIDIA A100 GPU

Figure 1. Runtime to decompose matrices using an M4 MacBook Air and an NVIDIA A100 GPU for eigendecomposition (using JAX.NUMPY.LINALG.EIGH), general Cholesky decomposition (using JAX.SCIPY.LINALG.CHOLESKY), and quasiseparable Cholesky decomposition of an exponential kernel (using TINYGP). The GPU is typically slower for small matrices, but outperforms the CPU for larger matrices by exploiting more parallel workers.

For our work, we use the package TINYGP, which implements a generalisation of CELERITE [43], using the algebra of quasiseparable matrices [38], and state-space models [e.g. 59, 68]. It can implement all CELERITE kernels with the same runtime and memory scaling, but it also features numerically stable, exact implementations of the Matérn-3/2 and Matérn-5/2 kernels, and banded covariance matrices [43]. It can also calculate GP derivatives following the framework of state-space models, similar to Rubenzahl et al. [101]. While similar extensions have also been implemented in the package S+LEAF [30, 31], the quasiseparable kernels in TINYGP are implemented entirely within JAX, which combined with our JAX-based implementation of LUAS, allows for efficient auto-differentiation and GPU hardware acceleration [16]. This is useful for scaling to problems involving large datasets, and for performing high-dimensional parameter inference using efficient gradient-based methods. For this work, we will use JJ to refer to the quasiseparable rank, where the exponential kernel is quasiseparable rank 1, the Matérn-3/2 kernel is rank 2, and the kernel in Equation (12) actually has rank 2J2J. Summing kernels will sum their quasiseparable ranks, while multiplying kernels multiplies their ranks. We note that some quasiseparable kernels theoretically follow a O(NJ3)\mathcal{O}(NJ^3) runtime scaling and O(NJ2)\mathcal{O}(NJ^2) memory scaling, such as high-order Matérn kernels and wide banded matrices, but these are not typically used in practice, and sums of arbitrary quasiseparable kernels will still tend towards CELERITE scalings by exploiting block diagonal transition matrices [43], so we use the scalings associated with CELERITE from Foreman-Mackey et al. [42] for calculations in this work.

Unfortunately, while all these methods have excellent scaling, they depend on the sort order of their inputs, and cannot add or

multiply kernels whose inputs have a different sort order. In practice, this makes them challenging to extend to 2D datasets, for example, if wavelength and time are kernel inputs on a spectroscopic time-series, it is not possible to sort all datapoints into chronological order, while never decreasing in wavelength. Creative approaches do exist to avoid this problem [e.g. 104, 53, 71], but they can come at significant computational cost for large datasets.

Using Any Matrix Square Root

Both the Cholesky factor L\mathbf{L}, and the term QΛ12\mathbf{Q}\mathbf{\Lambda}^{\frac{1}{2}}, are distinct examples of matrix square roots K12\mathbf{K}^{\frac{1}{2}} of K\mathbf{K}, where we will use the notation that a matrix square root satisfies

References

  1. [1]Ahrer, E.-M., et al. 2023, Nature, 614, 653
  2. [2]Aigrain, S., & Foreman-Mackey, D. 2023, ARA&A, 61, 329
  3. [3]Aigrain, S., Pont, F., & Zucker, S. 2012, MNRAS, 419, 3147
  4. [4]Alderson, L., et al. 2023, Nature, 614, 664
  5. [5]Alie, R., Stephens, D. A., & Schmidt, A. M. 2024, arXiv:2402.08877arxiv.org/abs/2402.08877
  6. [7]Ambikasaran, S., et al. 2015, IEEE Transactions on Pattern Analysis and Machine Intelligence, 38, 252
  7. [8]Angus, R., et al. 2018, MNRAS, 474, 2094
  8. [9]Barragán, O., Aigrain, S., Rajpaul, V. M., & Zicher, N. 2022, MNRAS, 509, 866
  9. [10]Barros, S. C. C., et al. 2020, A&A, 634, A75
  10. [12]Bell, T. J., et al. 2024, NatAs, 8, 879
  11. [13]Belouchrani, A., Abed-Meraim, K., Cardoso, J.-F., & Moulines, E. 1997, IEEE Transactions on Signal Processing, 45, 434
  12. [14]Birkby, J. L., et al. 2013, MNRAS, 436, L35
  13. [15]Bonilla, E., Chai, K., & Williams, C. 2007, Advances in Neural Information Processing Systems, 20
  14. [16]Bradbury, J., et al. 2018, JAX: composable transformations of Python+NumPy programs, 0.3.13, http://github.com/google/jax
  15. [18]Brogi, M., & Line, M. R. 2019, AJ, 157, 14
  16. [19]Bruinsma, W. P., et al. 2020, Proceedings of the 37th International Conference on Machine Learning
  17. [20]Cai, M. B., et al. 2020, arXiv e-prints, arXiv:2005.04879DOI
  18. [21]Cheverall, C. J., & Madhusudhan, N. 2024, AJ, 167, 272
  19. [22]Claringbold, A. B., et al. 2026, MNRAS, 546, stag143
  20. [23]Connors, N. J., Monaghan, C., Benneke, B., & Dang, L. 2025, ApJ, 989, L11
  21. [24]Costes, J. C., et al. 2026, MNRAS, 550, stag1173
  22. [25]Cui, T., Pagendam, D., & Gilfedder, M. 2021, Environmental Modelling & Software, 144, 105170
  23. [27]Das, S., et al. 2026, arXiv e-prints, arXiv:2607.21431arxiv.org/abs/2607.21431
  24. [28]Davey, J. J., Yip, K. H., Al-Reffaie, A. F., & Waldmann, I. P. 2025, MNRAS, 536, 2618
  25. [29]de Kok, R. J., et al. 2013, A&A, 554, A82
  26. [30]Delisle, J. B., Hara, N., & Ségransan, D. 2020, A&A, 638, A95
  27. [31]Delisle, J. B., Unger, N., Hara, N. C., & Ségransan, D. 2022, A&A, 659, A182
  28. [32]Deming, D., et al. 2015, ApJ, 805, 132
  29. [33]Diggle, P., Moraga, P., Rowlingson, B., & Taylor, B. 2013, Statistical Science, 28
  30. [34]Donnelly, J., Daneshkhah, A., & Abolfathi, S. 2024, Engineering Applications of Artificial Intelligence, 128, 107536
  31. [35]Dovers, E., Popovic, G. C., & Warton, D. I. 2024, Methods in Ecology and Evolution, 15, 191
  32. [36]Dyrek, A., et al. 2026, arXiv e-prints, arXiv:2608.04110DOI
  33. [37]Edenhofer, G., et al. 2024, JOSS, 9, 6593
  34. [38]Eidelman, Y., & Gohberg, I. 1999, Integral Equations and Operator Theory, 34, 293
  35. [39]Enßlin, T. 2013, AIP Conference Proceedings, 1553, 184
  36. [40]Espinoza, N., et al. 2019, MNRAS, 482, 2065
  37. [41]Event Horizon Telescope Collaboration, et al. 2019, ApJ, 875, L4
  38. [42]Foreman-Mackey, D., Agol, E., Ambikasaran, S., & Angus, R. 2017, AJ, 154, 220
  39. [43]Foreman-Mackey, D., et al. 2024, dfm/tinygp: The tiniest of Gaussian Process libraries, v0.3.0, Zenodo, doi:10.5281/zenodo.10463641
  40. [44]Fortune, M., et al. 2024, A&A, 686, A89
  41. [45]Fortune, M., et al. 2025, A&A, 701, A25
  42. [46]Gardner, J. R., et al. 2018, arXiv e-prints, arXiv:1809.11165arxiv.org/abs/1809.11165
  43. [47]Gibson, N. P. 2014, MNRAS, 455, 3401
  44. [48]Gibson, N. P., et al. 2012, MNRAS, 419, 2683
  45. [49]Gibson, N. P., et al. 2020, MNRAS, 493, 2215
  46. [51]Gilbertson, C., Ford, E. B., Jones, D. E., & Stenning, D. C. 2020, ApJ, 905, 155
  47. [52]Golub, G. H., & Van Loan, C. F. 2013, Matrix Computations (4th edn.; Baltimore, MD: Johns Hopkins University Press), https://epubs.siam.org/doi/book/10.1137/1.9781421407944
  48. [53]Gordon, T. A., Agol, E., & Foreman-Mackey, D. 2020, AJ, 160, 240
  49. [54]Gordon, T. A., et al. 2026, AJ, 171, 178
  50. [55]Greco, J. P., & Brandt, T. D. 2016, ApJ, 833, 134
  51. [57]Hara, N. C., & Delisle, J.-B. 2025, A&A, 696, A141
  52. [58]Harris, C. R., et al. 2020, Nature, 585, 357
  53. [59]Hartikainen, J., & Särkkä, S. 2010, in 2010 IEEE International Workshop on Machine Learning for Signal Processing, 379, doi:10.1109/MLSP.2010.5589113
  54. [60]Hattori, S., et al. 2022, AJ, 163, 284
  55. [61]Hattori, S., et al. 2024, exoplanet-dev/jaxoplanet: Astronomical time series analysis with JAX, v0.0.2, Zenodo, doi:10.5281/zenodo.10369396DOI
  56. [63]Hoffman, M. D., & Gelman, A. 2014, Journal of Machine Learning Research, 15, 1593
  57. [64]Holmberg, M., & Madhusudhan, N. 2023, MNRAS, 524, 377
  58. [65]Hong, Y. P., Horn, R. A., & Johnson, C. R. 1986, Linear Algebra and its Applications, 73, 213
  59. [66]Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90
  60. [68]Jordán, A., Eyheramendy, S., & Buchner, J. 2021, RNAAS, 5, 107
  61. [69]Jutten, C., & Herault, J. 1991, Signal Processing, 24, 1
  62. [70]Kallinger, T., et al. 2014, A&A, 570, A41
  63. [71]Keith, M. J., et al. 2025, MNRAS, 540, 2486
  64. [72]Kirk, J., et al. 2024, AJ, 167, 90
  65. [73]Kirk, J., et al. 2025, MNRAS, 537, 3027
  66. [74]Kisi, O., et al. 2025, Scientific Reports, 15
  67. [75]Knollmüller, J., & Enßlin, T. A. 2019, arXiv e-prints, arXiv:1901.11033arxiv.org/abs/1901.11033
  68. [76]Kreidberg, L. 2015, PASP, 127, 1161
  69. [77]Krige, D. 1951, Journal of the Southern African Institute of Mining and Metallurgy, 52, 119
  70. [78]Langford, Z., et al. 2025, PASP, 137, 114503
  71. [79]Liao, H.-J., Liu, J.-G., Wang, L., & Xiang, T. 2019, PhRvX, 9, 031041
  72. [80]Lochner, M., et al. 2016, ApJS, 225, 31
  73. [82]Luger, R., et al. 2016, AJ, 152, 100
  74. [83]Lund, M. N. 2019, MNRAS, 489, 1072
  75. [84]Maguire, C., et al. 2024, A&A, 692, A8
  76. [85]Mandel, K., & Agol, E. 2002, ApJ, 580, L171
  77. [86]McWilliam, N., et al. 2026, AJ, 171, 233
  78. [87]Meech, A., Aigrain, S., Brogi, M., & Birkby, J. L. 2022, MNRAS, 512, 2604
  79. [88]Morello, G. 2015, ApJ, 808, 56
  80. [89]Morris, B. M., et al. 2020, MNRAS, 493, 5489
  81. [90]Nasedkin, E., et al. 2023, A&A, 678, A41
  82. [91]Pérez, F., & Granger, B. E. 2007, Computing in Science and Engineering, 9, 21
  83. [92]Pleiss, G., Gardner, J., Weinberger, K., & Wilson, A. G. 2018, Proceedings of the 35th International Conference on Machine Learning, 80, 4114
  84. [93]Pope, B. J. S., et al. 2019, ApJS, 245, 8
  85. [94]Press, W. H., Rybicki, G. B., & Hewitt, J. N. 1992, ApJ, 385, 416
  86. [95]Qiu, B., Lu, T., Liu, S., & Luo, A.-L. 2026, AJ, 171, 314
  87. [96]Radica, M., et al. 2026, AJ, 171, 314
  88. [97]Rajpaul, V., et al. 2015, MNRAS, 452, 2269
  89. [98]Rakitsch, B., Lippert, C., Borgwardt, K., & Stegle, O. 2013, Advances in Neural Information Processing Systems, 26
  90. [99]Rotman, Y., et al. 2025, ApJ, 989, 201
  91. [100]Roy-Perez, J., Pérez-Hoyos, S., Barrado-Izagirre, N., & Chen-Chen, H. 2026, A&A, 707, A297
  92. [101]Rubenzahl, R. A., et al. 2026, AJ, 171, 259
  93. [102]Saatchi, Y. 2011, PhD thesis, University of Cambridge
  94. [103]Sarkka, S., & Hartikainen, J. 2012, Proceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics, 22, 993
  95. [104]Särkkä, S., Solin, A., & Hartikainen, J. 2013, IEEE Signal Processing Magazine, 30, 51
  96. [105]Schlawin, E., et al. 2020, AJ, 160, 231
  97. [106]Schmidt, S. P., et al. 2025, AJ, 170, 298
  98. [108]Snelson, E., & Ghahramani, Z. 2005, Advances in Neural Information Processing Systems, 18
  99. [110]Stevenson, K. B., et al. 2012, ApJ, 754, 136
  100. [111]Susobhanan, A., & van Haasteren, R. 2025, MNRAS, 542, 2892
  101. [112]Svendsen, D. H., et al. 2020, ISPRS Journal of Photogrammetry and Remote Sensing, 166, 68
  102. [113]Tamuz, O., Mazeh, T., & Zucker, S. 2005, MNRAS, 356, 1466
  103. [114]Teh, Y. W., Seeger, M., & Jordan, M. I. 2005, Proceedings of the Tenth International Workshop on Artificial Intelligence and Statistics, R5, 333
  104. [116]Tipping, M. E., & Bishop, C. M. 1999, Journal of the Royal Statistical Society Series B: Statistical Methodology, 61, 611
  105. [118]Truffinet, O., Ammar, K., Argaud, J.-P., & Bouriquet, B. 2026, Neurocomputing, 697, 134156
  106. [120]van Haasteren, R., & Vallisneri, M. 2014, PhRvD, 90, 104012
  107. [121]Velten, B., et al. 2022, Nature Methods, 19
  108. [122]Villar, V. A., et al. 2020, ApJ, 905, 94
  109. [123]Virtanen, P., et al. 2020, Nature Methods, 17, 261
  110. [124]Vos, J. M., et al. 2022, ApJ, 924, 68
  111. [125]Waalkes, W. C., et al. 2026, arXiv e-prints, arXiv:2610.24550arxiv.org/abs/2610.24550
  112. [126]Waldmann, I. P. 2012, ApJ, 747, 12
  113. [127]Wang, D., Hogg, D. W., Foreman-Mackey, D., & Schölkopf, B. 2016, PASP, 128, 094503
  114. [129]Weeks, A., et al. 2026, arXiv e-prints, arXiv:2610.30594arxiv.org/abs/2610.30594
  115. [130]Wilson, A. G., & Nickisch, H. 2015, arXiv e-prints, arXiv:1503.01057arxiv.org/abs/1503.01057
  116. [131]Wilson, J., et al. 2020, Proceedings of the 37th International Conference on Machine Learning, 119, 10292
  117. [133]Yu, B. M., et al. 2008, 21
  118. [134]Yu, H., et al. 2024, MNRAS, 535, 634
  119. [135]Zhang, C., Li, W., & Travis, D. 2007, International Journal of Remote Sensing, 28, 5103

Paper details

vlnc:W_28642u57
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

coding assistants

View statement
Looking for related papers…
Contents