Least-Squares Deconvolution in Stellar Spectroscopy

• Least-Squares Deconvolution (LSD) is a multiline deconvolution technique that extracts an average high-S/N line profile from stellar spectra.
• It models observed spectra as a convolution of a mean line profile with a pre-defined mask of spectral lines to retrieve radial velocities and diagnose stellar activity.
• Advanced implementations incorporate regularization and joint continuum fitting to reduce noise and enhance the detection of exoplanet signals and stellar phenomena.

Least-Squares Deconvolution (LSD) is a multiline deconvolution technique developed to extract high-fidelity, high-S/N average line profiles from stellar spectra, particularly in the presence of substantial line blending and noise. Its core principle is that the observed spectrum can be modeled as a convolution of an average (“mean”) intrinsic line profile with a known pattern of spectral lines (“mask”), allowing for precise retrieval of profile shapes, radial velocities (RVs), and diagnostics of stellar activity, rotation, and planetary transits. LSD is foundational in contemporary stellar spectroscopy and RV exoplanet searches, especially when applied to high-resolution échelle spectra and spectropolarimetric data.

1. Mathematical Foundation of LSD

LSD assumes the normalized observed spectrum $r(\lambda)$ can be approximated as a sum (or convolution) of scaled, velocity-shifted copies of an unknown mean line profile $z(v)$, distributed according to the positions and strengths of spectral lines. In matrix notation, this is

$r_j = 1 + \sum_k \alpha_{jk} z_k,$

where $r_j$ is the normalized flux at pixel $j$, $z_k$ is the mean line profile at velocity bin $k$, and $\alpha_{jk}$ is the design/mask matrix, encoding the properties (e.g., rest wavelength $\lambda_i$, normalized depth $d_i$) of each line and its contribution to each pixel (Dolan et al., 2024, Reeth et al., 2013, Kochukhov et al., 2010).

Least-squares minimization seeks the $z_k$ that minimizes

$$\chi^2 = \sum_j \left[ \frac{ (r_j - 1) - \sum_k \alpha_{jk} z_k }{ \sigma_j } \right]^2,$$

where $\sigma_j$ is the noise per pixel. The solution and covariance are

$$\mathbf{z} = (\alpha^T S^2 \alpha)^{-1} \alpha^T S^2 \mathbf{R}, \quad \mathbf{R} = r-1, \quad S^2 = \mathrm{diag}(1/\sigma_j^2).$$

This linear inversion framework admits efficient matrix solutions and error propagation.

2. Construction and Properties of Line Masks

The LSD mask encodes all chosen absorption lines, each with parameters (rest wavelength, predicted central depth, Landé factor for spectropolarimetry). The selection and weighting ($w_i$) of lines are critical:

• Weights are typically $w_i = d_i$ (intensity), $w_i = d_i \lambda_i g_i$ (Stokes V), or $w_i = d_i \lambda_i^2 G_i$ (Stokes Q/U) (Kochukhov et al., 2010).
• Only lines deeper than a threshold (e.g., $d_i > 0.1$ or $d_{\mathrm{min}} = 1/(3\,\mathrm{S/N})$) are included (Dolan et al., 2024).
• Exclusion of broad lines, telluric-contaminated regions, and anomalous blends is essential to prevent ill-conditioning and non-physical profile shapes.

Parametric and randomized line selection algorithms can be employed to optimize the mask for specific science goals, such as jitter minimization in RVs or sensitivity to magnetic features. Randomized selection has achieved >50% reduction in activity-induced RV RMS for active M dwarfs, far outperforming simple parametric thresholds, and enables recovery of planetary signals otherwise lost in activity jitter (Bellotti et al., 2021).

3. Extensions and Regularization Techniques

The classical LSD algorithm is linear, but refinements address key deficiencies:

• Continuum normalization: Accurate profile retrieval depends on robust continuum fitting. The A.C.I.D. method introduces simultaneous fitting of both the spectral continuum (modeled as a low-order polynomial) and the line profile, optimizing both in a joint likelihood/MCMC framework (Dolan et al., 2024).
• Optical-depth deconvolution: Working in 'effective optical depth' ($\tau = -\ln(r)$, with $d_i^\tau = -\ln(1-d_i)$) rather than flux space reduces artificial enhancements of deep blends and yields more physical residuals (Dolan et al., 2024).
• Regularization: Tikhonov penalties or Gaussian process (GP) priors are implemented to control noise-amplification and enforce profile smoothness, respectively (Ramos et al., 2015, Kochukhov et al., 2010).
• Multiprofile LSD: Extended to decompose composite spectra (e.g., binaries or chemical groups), yielding separate mean profiles for multiple stellar components or line groups (Tkachenko et al., 2013, Kochukhov et al., 2010).
• Bayesian inference: Full Bayesian LSD with GP priors delivers credible intervals on $z_k$ and enables robust modeling of noise properties and signal detection thresholds (Ramos et al., 2015).

4. Implementation and Algorithmic Workflow

A typical LSD reduction pipeline (see (Dolan et al., 2024, Reeth et al., 2013, Lienhard et al., 2022)) comprises:

1. Data ingestion: Read normalized, blaze-corrected spectra with per-pixel uncertainties.
2. Mask construction: Assemble a depth/parameter-filtered line list, optionally removing blended, telluric, or variable lines.
3. Design matrix assembly: Calculate $\alpha_{jk}$ or equivalent, typically using fast interpolation (e.g., triangular kernel).
4. Matrix inversion: Solve the normal equations for $z_k$ and compute its formal errors.
5. Validation and regularization: Optionally, refine via iterative line-depth correction, employ regularization, or run joint continuum/profile fits (e.g., as in A.C.I.D.).
6. Combining orders/nights: Co-add profiles from different spectral orders and epochs by S/N-weighted averaging.
7. Extraction of observables: Fit centroids (RVs), equivalent widths, or higher moments of $\mathbf{z}$.

The computational cost scales linearly with the number of pixels and lines; masking and selection steps are critical for stability and performance in large datasets (Lienhard et al., 2022).

5. Applications to Stellar Physics and Exoplanetary Science

LSD is widely employed in:

• Radial velocity planet searches: LSD profiles provide RVs with higher S/N and stability than single-line or CCF approaches. MM-LSD achieves ~12% lower RV RMS than CCFs in HARPS-N data for FGK stars (Lienhard et al., 2022).
• Rossiter-McLaughlin effect: Extraction of fine time-dependent distortions in the mean profile during planetary transits, supporting aligned/oblique spin-orbit angle measurements (Dolan et al., 2024, Strachan et al., 2017, Tkachenko et al., 2022).
• Stellar activity studies: Activity-induced line shape modulation and jitter are mitigated by careful mask construction and profile analysis; randomized line selection enhances sensitivity to rotational modulation and planetary signals (Bellotti et al., 2021).
• Spectropolarimetry: LSD cross-correlation of Stokes profiles enables the detection of weak magnetic fields in cool stars, with GP-regularized Bayesian methods allowing error-controlled inference even in the low-S/N regime (Ramos et al., 2015).
• Atmospheric parameter determination: High-S/N LSD model spectra facilitate recovery of $T_\mathrm{eff}$, $\log g$, metallicity, rotation rates, and mode identification in pulsating and binary stars, even for faint or rapid rotators (Reeth et al., 2013, Tkachenko et al., 2013, Kochukhov et al., 2010).
• Double-lined binaries: Simultaneous recovery of multiple stellar profiles, flux ratios, and eclipse-distorted Rossiter-McLaughlin signatures (LSDBinary) (Tkachenko et al., 2022, Tkachenko et al., 2013).

6. Limitations, Validation, and Best Practices

LSD's accuracy relies on:

• Correctness of the line similarity assumption: The extraction of a single mean profile is valid only for weak, self-similar, non-blended lines. Strong-line saturation and blends yield systematic deviations ("LSD profile does not behave as a real line" for Stokes $I$ and $Q$) (Kochukhov et al., 2010).
• Robustness to continuum errors: Profile depth estimation is highly sensitive to continuum normalization. Joint fitting (as in A.C.I.D.) recovers true depths and continuum to within noise for synthetic and real data; even small continuum errors yield significant biases (Dolan et al., 2024).
• Mask and parameter selection: Exclusion of non-physical, telluric, or anomalous regions and optimal thresholding is vital for outlier rejection and signal fidelity (Dolan et al., 2024, Bellotti et al., 2021, Lienhard et al., 2022).
• Line blending and resolution: LSD presumes quasi-linear blend addition; at low spectral resolving power ($R \lesssim 10,000$) or for densely blended spectra, deconvolved profiles become unreliable (Tkachenko et al., 2013).
• Interpretation of moments: Equating LSD profiles to real spectral lines is problematic outside the weak-line regime; direct modeling and forward application of LSD to synthetic spectra is preferred for quantitative physical inference (Kochukhov et al., 2010).

Validation with synthetic data and controlled injection/recovery experiments are essential to establish retrieval biases and performance metrics for specific regimes (Dolan et al., 2024, Reeth et al., 2013, Tkachenko et al., 2013).

7. Recent Innovations and Comparative Performance

Recent advances include:

• A.C.I.D. incorporates joint MCMC optimization of continuum and line profile, and optical-depth deconvolution, yielding Voigt-shaped mean profiles that more realistically encode thermal, pressure, and instrumental broadening. In HARPS transit observations of HD189733b, ACID profiles exhibit ∼0.01 continuum depression, reduced RMS compared to CCFs (>5% improvement), and enhanced profile uniformity (σ(depth) $3 \times 10^{-4}$ for A.C.I.D., $1.2 \times 10^{-3}$ for CCF) (Dolan et al., 2024).
• Multi-Mask LSD: Simultaneous exploration of multiple mask configurations (hyperparameter grid) and averaging over the best-performing RV series yields robust ∼12% RV RMS reductions in solar and FGK-star samples (Lienhard et al., 2022).
• Randomised mask optimization: Systematic jitter reduction exceeding 50% in M dwarf RVs and improved planetary signal recovery by eliminating lines most contaminated by activity via randomized search strategies (Bellotti et al., 2021).
• Bayesian/GP-regularized LSD: Delivers principled error bars per velocity bin and automatic smoothness, facilitating detection limits and signal significance quantification in spectropolarimetric surveys (Ramos et al., 2015).

A.C.I.D., MM-LSD, LSDBinary, Bayesian LSD, and dLSD (differential LSD) each address specific systematic limitations of classical linear LSD, and their comparative adoption is tailored to the target system's characteristics and the science objectives.

---

For further implementation details and software, the open-source A.C.I.D. codebase is available at https://github.com/ldolan05/ACID (Dolan et al., 2024).