Mapping interstellar dust with Gaussian processes

Andrew C. Miller, Lauren Anderson, Boris Leistedt, John P. Cunningham, David W. Hogg, David M. BleiThe Annals of Applied Statistics 2022 · arXiv

Abstract

Interstellar dust corrupts nearly every stellar observation, and accounting for it is crucial to measuring physical properties of stars. We model the dust distribution as a spatially varying latent field with a Gaussian process (GP) and develop a likelihood model and inference method that scales to millions of astronomical observations. Modeling interstellar dust is complicated by two factors. The first is integrated observations. The data come from a vantage point on Earth and each observation is an integral of the unobserved function along our line of sight, resulting in a complex likelihood and a more difficult inference problem than in classical GP inference. The second complication is scale; stellar catalogs have millions of observations. To address these challenges we develop ziggy, a scalable approach to GP inference with integrated observations based on stochastic variational inference. We study ziggy on synthetic data and the Ananke dataset, a high-fidelity mechanistic model of the Milky Way with millions of stars. ziggy reliably infers the spatial dust map with well-calibrated posterior uncertainties.

Introduction

Interstellar dust acts as a cosmic fog, dimming and reddening starlight—a phenomenon known as extinction. This effect complicates the measurement of a star’s true distance and intrinsic brightness. To accurately study the Milky Way, astronomers require high-fidelity 3D maps of this dust distribution.

However, constructing such maps presents two significant challenges. First, we cannot measure dust density at a specific point in space. Instead, we observe the cumulative extinction along the line of sight from Earth to a star. This means our data consists of integrals of the latent dust density field, rather than direct samples. Second, modern stellar catalogs, such as Gaia, contain millions to billions of stars. Standard Gaussian Process (GP) regression, which scales cubically ($O(N^3)$) with the number of observations, is computationally infeasible at this scale.

Comparison of pointwise and integrated observations. (a) Standard GP regression uses pointwise samples. (b) In dust mapping, observations are integrals along the line of sight from Earth (red dot) to stars (grey dots).

The ziggy Framework

To overcome these hurdles, we introduce ziggy, a scalable approach based on Stochastic Variational Gaussian Processes (SVGP). ziggy incorporates several key innovations to handle the unique nature of astronomical data:

Inducing Points and Scalability

Rather than processing every star simultaneously, ziggy utilizes a set of "inducing points" to represent the global dust field. This reduces the computational complexity from $O(N^3)$ to $O(M^3)$, where $M$ is the number of inducing points and $M \ll N$. This allows the model to scale to millions of observations.

Monte Carlo Semi-Integrated Kernels

The line-of-sight integrals often lack closed-form solutions. ziggy employs a Monte Carlo estimator that samples points along the ray between Earth and each star. This approximates the covariance between the integrated observation and the latent dust field, enabling efficient inference without requiring spatial discretization into voxels.

Fast Variance Interpolation

Calculating the variance of an integrated observation—a "doubly-integrated" kernel—is typically a bottleneck. By exploiting the rotational invariance of isotropic kernels, we pre-calculate this variance as a 1D function of distance. This allows for rapid linear interpolation during training and inference.

Whitened Parameterization

Optimizing a dust map alongside GP hyperparameters (like length scales) is notoriously difficult. ziggy uses a "whitened" coordinate system to decouple the geometry of the dust map from the GP's hyperparameters. This stabilization significantly improves the convergence and speed of the optimization process.

Experimental Evaluation

We evaluated ziggy using two primary benchmarks:

  1. Synthetic 2D Data: A controlled environment to validate the mathematical framework and compare the performance of different kernels (e.g., Matérn vs. Squared Exponential).
  2. The Ananke Dataset: A high-fidelity simulation of the Milky Way based on FIRE cosmological simulations. This provides a "ground truth" for 1 million stars, mimicking the complexity of real-world Gaia data.

Key Results

Our experiments demonstrate that ziggy effectively leverages large datasets. As the number of stars increased from $10^3$ to $10^6$, the accuracy of the dust map improved significantly. On the Ananke dataset, the Root Mean Squared Error (RMSE) for predicted extinction decreased by a factor of approximately 2.6 when scaling from 100,000 to 1,000,000 stars.

Furthermore, ziggy produces well-calibrated posterior uncertainties, ensuring that the model's confidence levels accurately reflect its predictive performance. Unlike previous voxel-based methods, ziggy treats dust as a continuous field, avoiding blocky artifacts and better representing small-scale structures.

Limitations and Future Directions

While ziggy represents a significant step forward, several challenges remain:

  • Distance Uncertainty: The current model assumes stellar distances are known exactly. Future work will involve jointly modeling distance and dust extinction.
  • Non-Gaussian Noise: Real-world data may contain heavy-tailed outliers that deviate from the assumed Gaussian error model.
  • Resolution vs. Complexity: Resolving extremely fine dust features requires increasing the number of inducing points ($M$), which eventually hits the $O(M^3)$ computational ceiling.

The ultimate objective is to scale ziggy to the full Gaia dataset, encompassing over 1.3 billion stars, to provide the most detailed 3D map of our galaxy's dust to date.

Summary updated

Abstract

Interstellar dust corrupts nearly every stellar observation and accounting for it is crucial to measuring physical properties of stars. We model the dust distribution as a spatially varying latent field with a Gaussian process (GP) and develop a likelihood model and inference method that scales to millions of astronomical observations. Modeling interstellar dust is complicated by two factors. The first is integrated observations. The data come from a vantage point on Earth, and each observation is an integral of the unobserved function along our line of sight, resulting in a complex likelihood and a more difficult inference problem than in classical GP inference. The second complication is scale; stellar catalogs have millions of observations. To address these challenges, we develop ziggy, a scalable approach to GP inference with integrated observations based on stochastic variational inference. We study ziggy on synthetic data and the Ananke dataset, a high-fidelity mechanistic model of the Milky Way with millions of stars. ziggy reliably infers the spatial dust map with well-calibrated posterior uncertainties.
BibTeX
@article{miller2022mapping,
  author = {Andrew C. Miller and Lauren Anderson and Boris Leistedt and John P. Cunningham and David W. Hogg and David M. Blei},
  title = {{Mapping interstellar dust with Gaussian processes}},
  journal = {The Annals of Applied Statistics},
  year = {2022},
}
Figure