Seeking Spectroscopic Binaries with Data-Driven Models

Problem Statement
Data-driven stellar classification has become a standard tool for deriving stellar properties (labels) such as effective temperature ($T_{\text{eff}}$), radius ($R_*$), metallicity ([Fe/H]), and projected rotational velocity ($v \sin i$) from large spectroscopic surveys. While these models excel at label transfer, their application to detecting unresolved stellar binaries remains challenging. Identifying binaries requires an accurate model of the single-star spectrum to distinguish subtle deviations caused by a fainter secondary star. Traditional searches are often limited to close binaries with distinct spectral features (SB2s), missing wider systems where spectral features overlap. The authors investigate whether data-driven models, specifically trained on high-resolution Keck HIRES spectra, can achieve the per-pixel flux accuracy necessary to detect these unresolved "spectral binaries" in the California-Kepler Survey (CKS) sample.

Methodology
The authors employed The Cannon, a data-driven framework that models stellar flux as a function of stellar labels, to create a spectral emulator for HIRES data.

1. Training Data: The model was trained on the SpecMatch-Emp library, a set of 404 well-characterized stars ($T_{\text{eff}} \sim 3000\text{--}6600$ K). The sample was rigorously cleaned of known binaries using SIMBAD, Gaia metrics, and literature catalogs. To address the lack of $v \sin i$ values for slow rotators in the cool star sample, the authors included artificially broadened copies of these spectra.
2. Data Preprocessing:
   • Blaze Removal & Registration: Spectra were shifted to the rest frame and corrected for the instrumental blaze function.
   • Wavelet Filtering: A novel step was introduced using discrete wavelet decomposition (specifically the "ym5" mode) to remove low-frequency, non-astrophysical variations (e.g., residual blaze function mismatches and night-to-night systematics) while preserving astrophysical line information.
   • Pixel Masking: Telluric absorption lines were masked. Additionally, a "binary signature" pixel mask was created using ab initio simulations to select pixels most sensitive to secondary star contributions.
3. Model Architecture:
   • Single-Star Model: A piecewise Cannon model was trained on hot ($T_{\text{eff}} > 5200$ K) and cool ($T_{\text{eff}} < 5300$ K) stars separately to maintain the accuracy of the quadratic flux-label approximation. The model predicts flux based on labels ($T_{\text{eff}}$, $R_*$, [Fe/H], $v \sin i$) and the instrumental Point Spread Function (PSF).
   • Binary Model: A composite binary model was constructed as a linear combination of two Doppler-corrected single-star spectra, weighted by their flux ratio. The model allows for independent velocities for the primary and secondary to account for imperfect rest-frame shifting and small orbital offsets.
4. Validation: The model's performance was tested via a leave-20%-out validation for label transfer and by comparing single-star vs. binary fits on a validation set of 102 known single stars and 97 known unresolved binaries from the CKS sample. Goodness-of-fit was evaluated using the Bayesian Information Criterion (BIC) and an improvement fraction ($f_{\text{imp}}$).

Key Contributions

• Wavelet-Based Filtering: The paper introduces and validates a wavelet-filtering technique that significantly improves the accuracy of data-driven label predictions, particularly for moderate Signal-to-Noise (S/N) spectra, by removing low-frequency instrumental systematics without flattening the continuum in a way that loses astrophysical information.
• HIRES Spectral Emulator: The authors successfully trained a data-driven model on HIRES spectra that achieves competitive precision in predicting stellar labels ($T_{\text{eff}}$, $R_*$, [Fe/H], $v \sin i$, PSF) compared to existing models.
• Binary Detection Limits: The study provides a critical assessment of the limits of current data-driven models for binary detection, demonstrating that while label transfer is robust, per-pixel flux emulation is not yet sufficient for confident binary identification in this dataset.

Results

• Label Transfer: The wavelet-filtered model showed superior agreement with ground-truth labels compared to the unfiltered model. For the SpecMatch-Emp training set, the filtered model achieved precisions of 27 K in $T_{\text{eff}}$, 0.10 $R_\odot$ in radius, 0.09 dex in [Fe/H], and 0.87 km s$^{-1}$ in $v \sin i$. This improvement was even more pronounced for the moderate S/N ($\sim 45$) CKS survey spectra.
• Binary Detection Failure: Despite accurate label predictions, the model failed to confidently distinguish between single stars and unresolved binaries in the validation set.
  • The distributions of $\Delta$BIC (difference in BIC between binary and single-star fits) and $f_{\text{imp}}$ for known binaries and single stars showed significant overlap.
  • The model residuals for both single stars and binaries were comparable, typically ranging from 2% to 5%.
  • The authors attribute this failure to the model's per-pixel flux prediction accuracy being limited to $\sim$3%. Since the flux contribution of a secondary star in many unresolved binaries is often $<3\%$, the model's intrinsic scatter obscures the binary signature.
• Synthetic Data Context: Simulations using ab initio (PHOENIX) spectra convolved to HIRES resolution suggested that binaries should be detectable (with $\Delta$BIC strongly favoring the binary model) if the underlying spectral model were perfect. This confirms that the limitation lies in the data-driven model's fidelity rather than the physical detectability of the signatures.

Significance and Claims
The paper concludes that while data-driven models are highly effective for label transfer (inferring stellar parameters), they currently face significant hurdles in spectral emulation (predicting the exact flux at every pixel). The authors state that the $\sim$2%–5% per-pixel flux uncertainty in their model is the primary bottleneck preventing the detection of unresolved binaries in the CKS sample.

The significance of the work lies in:

1. Demonstrating the Gap: Highlighting that accurate stellar labels do not guarantee accurate spectral reconstruction, a crucial distinction for applications requiring precise flux modeling (like binary detection).
2. Methodological Improvement: Validating wavelet filtering as a superior preprocessing step for removing systematics in high-resolution spectroscopy compared to traditional continuum normalization.
3. Setting Expectations: Providing a realistic assessment that current data-driven approaches, even with large training sets, may not yet be sensitive enough to detect the subtle signatures of wide-separation binaries in high-resolution ground-based data without further improvements in model complexity or training data homogeneity. The authors note that the latent dimensionality of HIRES data (due to PSF instability and non-uniform resolution) presents a greater challenge than in lower-resolution surveys like APOGEE.