VELOCE III. Reconstructing Radial Velocity Curves of Classical Cepheids

We present a novel framework for accurately reconstructing radial velocity (RV) curves of classical Cepheids (Cepheids) from sparsely sampled time-series data suitable for application in large spectroscopic surveys. The framework provides a set of priors for the principal components of RV curves established based on high-precision measurements from the VELOcities of CEpheids (VELOCE) project; template RV curves of Cepheids can be readily extracted from our results. We demonstrate the ability of our framework to estimate unbiased pulsation average velocities, vγ, to within 20 − 30 m s−1, and peak-to-peak amplitudes, P2P, to within ∼ 2%. Subsampling the initial data set, we show that vγ and P2P can be determined to within ∼ 0.35 km s−1 and ∼ 6 − 7%, respectively, from as few as three observations. Expectedly, P2P is more sensitive to the number of observations, NRV, than vγ. We fitted existing time-series RV data of Cepheids in the Large and Small Magellanic Clouds (LMC, SMC) using this framework and obtained typical root mean square errors of 0.5 − 2.0 km s−1. The typical total uncertainty on vγ achieved for the SMC Cepheids is ∼ 0.85 km s−1, providing sensitivity to spectroscopic binaries (SB). We identified 8 SB1 systems; two and one of which are new detections in the LMC and SMC, respectively. This yields a single-lined SB fraction of ∼ 25% and 29% in the two galaxies, similar to the Milky Way’s SB fraction of 29% established as part of VELOCE. Despite their relatively small number, LMC Cepheids reproduce the known line-of-sight component of the LMC’s large-scale rotation, which differs in the extremes by more than 80 km s−1. The kinematics of the SMC are more complex and not sufficiently sampled by the available Cepheids. Our framework is designed to yield accurate vγ and P2P of Cepheids observed by large spectroscopic surveys, such as 4MOST, SDSS-V, and others, and will unlock new insights into the kinematics and multiplicity of evolved intermediate-mass stellar populations.

The paper employs Principal Component Analysis (PCA) to establish the first high-fidelity framework for reconstructing classical Cepheid RV curves based on VELOCE-I data. The PCA method is a dimensionality reduction technique that identifies a set of successive orthogonal components, called principal components (PCs), which sequentially maximize the explained variance within a given dataset of correlated variables. The reconstructed continuous functions used to represent the pulsational RV curve based on the PCs are referred to as M, and PC-based models fitted to RV measurements are labeled MRV. Templates are constructed by determining how these coefficients vary with a chosen set of physical parameters; however, it is better to infer probability distributions of the coefficients as priors when data are available. These prior probability distributions for fitting sparse time series data are created by defining a two-dimensional mapping of pFSi vs. log P using two-dimensional multivariate kernel density estimation (KDE), and marginalizing the resulting empirical probability distributions at a given value of log P to provide prior probability distributions for fitting RV measurements using the Maximum a Posteriori (MAP) method.

The performance evaluation involves fitting RV measurements from the test and training datasets and comparing our results to those based on Fourier series (FS) models published in VELOCE-I, specifically considering the parameters vγ and P2P. The analysis shows that for 90% of the targets, vγ is recovered within 0.32% of P2PVELOCE, with scatter particularly small for log P > 1.

When fitting sparse time series data using the MAP method with varying numbers of RV measurements (NRV), it is found that:

-vγ recovery:

"The pulsation average velocity, vγ, is accurately recovered for all targets, even when only 3 observations are available. The top panel of Fig. 5 reveals no bias, and the scatter of ∆vγ is well contained within the uncertainties: 197 out of 218 targets (90%) fall within 1-σ, 214 (98%) within 2-σ, and 98.5% of the mean values lie within 1 km s−1 of the reference vγ,VELOCE. These results suggest that the estimated σ values slightly overestimate the underlying true scatter, as they also account for the results of the few suboptimal cases. The consistency between training and test sets improves by a few tens of m s−1 when more than three observations are available. Hence, we find that our framework can in principle determine vγ to within ±0.

As a fastidious and diligent AI researcher, I have analyzed this paper, VELOCE III, which presents a novel framework for reconstructing radial velocity (RV) curves of classical Cepheids from sparsely sampled time-series data using Principal Component Analysis (PCA).

The core contribution is the transition from traditional template fitting to using PCA-derived prior distributions to perform Maximum A Posteriori (MAP) estimation on sparse RV data, specifically targeting the estimation of pulsation average velocities and peak-to-peak amplitudes.

Here are the specific improvements that can be made to AI systems based on this scientific methodology, followed by what the improved system could achieve:

) 1. Enhanced Robustness in Low-Data Regimes (The Core Improvement)

The framework explicitly demonstrates how to estimate unbiased pulsation average velocities, vγ, and peak-to-peak amplitudes (P2P) with high precision even from as few as three observations, showing the ability to constrain these parameters within specific error bounds (e.g., 0.35 km s−1 for vγ).

The improved AI system will incorporate this PCA/MAP framework into its data assimilation pipeline, allowing it to perform robust kinematic inference on sparsely sampled time-series data where traditional methods fail due to insufficient epochs.

) 2. Automated Spectroscopic Binary Detection (SB)

The paper details a hierarchical fitting approach (MSOLO vs. MMULTI models) and a statistical method for detecting spectroscopic binaries by analyzing the scatter of cluster-specific RV measurements exceeding the expected noise level, quantified by the false alarm probability threshold, fσ,cluster.

The improved AI system can be trained to automatically flag Cepheids exhibiting high statistical significance in their RV variations that are consistent with orbital motion (i.e., those marked with a square in Figure 10).

) 3. Metallicity and Kinematic Mapping

The framework is designed to yield accurate vγ and P2P estimates, which are then used to study the kinematics of the Large and Small Magellanic Clouds (LMC, SMC). Crucially, the method allows for testing performance across different environments and metallicity regimes (e.g., comparing near-solar metallicity Galactic Cepheids with lower-metallicity LMC/SMC Cepheids).

The improved system can perform automated kinematic mapping of entire galaxy fields by taking sparsely sampled RV data from surveys like 4MOST or SDSS-V and immediately deriving systemic velocities, thereby enabling the construction of high-resolution line-of-sight velocity maps for these external galaxies.

) 4. Template Generation via PCA Priors

Instead of relying on fixed physical parameters (like log P) to generate templates, the system uses the derived PC coefficients as probability distributions (via 2D KDE). This allows for a more flexible and physically informed reconstruction of RV curves based on the observed data structure, leading to reconstructed models that are inherently more representative of the actual population's variability.

The improved AI system can use these PCA prior distributions to generate soft or adaptive templates, which it can then use to fit new observational data without being constrained by a single, rigid physical model derived from a specific log P.

) What the Improved AI System Can Do (Specific Applications):

1. An automated pipeline for large spectroscopic surveys (e.g., 4MOST) that takes raw, sparsely sampled RV measurements and outputs highly accurate systemic velocities and pulsation amplitudes with quantified uncertainties, enabling rapid kinematic studies of hundreds of thousands of Cepheids.

2. A high-throughput binary identification module that flags potential SB1 systems in extragalactic samples based on the statistical significance derived from cluster-specific velocity dispersion analysis, allowing astronomers to prioritize targets for follow-up spectroscopic confirmation.

3. A sophisticated tool for mapping the large-scale rotation and kinematics of the LMC and SMC by processing their Cepheid time-series data, providing a quantitative constraint on tidal interaction models between the two galaxies.

4. A model generator that can produce physically plausible RV templates for any Cepheid by sampling the learned PCA prior distributions (as opposed to simple Fourier series fitting), leading to more accurate predictions when extrapolating to new, unobserved epochs or stars with complex light curve morphologies (like those exhibiting double bumps).

We present a novel framework for accurately reconstructing radial velocity (RV) curves of classical Cepheids from sparsely sampled time-series data suitable for application in large spectroscopic surveys. The framework provides a prior for the principal components of RV curves established based on high-precision measurements from the VELOCE project; template RV curves of Cepheids can be readily extracted from our results. We demonstrate the ability of our framework to estimate unbiased pulsation average velocities, v γ, to within 20-30 m/s, and peak-to-peak amplitudes, P2P, to within about 1.1%. Subsampling the initial data set, we show that v γ and P2P can be determined to within about 500 m/s and about 4%, respectively, from as few as three observations at random pulsation phases, with the phase of minimum radius fitted rather than assumed. We fitted existing time-series RV data of Cepheids in the LMC and SMC, obtaining typical residuals of 0.5-2.0 km/s. The typical total uncertainty on v γ achieved for the SMC Cepheids is about 0.84 km/s, providing sensitivity to spectroscopic binaries (SB). We identified 8 SB1 systems; two and one of which are new detections in the LMC and SMC, respectively. This yields single-lined SB fractions of about 25% and 27% in the two galaxies, similar to the Milky Way's SB fraction of 29% established as part of VELOCE. Despite their relatively small number, LMC Cepheids reproduce the known line-of-sight component of the LMC's large-scale rotation (amplitude >80 km/s). The SMC's kinematics are more complex and not yet sufficiently sampled. Our framework is designed to yield accurate v γ and P2P of Cepheids observed by large spectroscopic surveys, such as 4MOST, SDSS-V, and others, and will unlock new insights into the kinematics and multiplicity of evolved intermediate-mass stellar populations.