🔭 astrophysics

A Linearized Approach to Radial-Velocity Extraction. II: Shot-Noise-Limited Precision via Spectral Factorization

This paper presents a generalized, shot-noise-limited method for extracting radial velocities from spectroscopic time series by factorizing unknown spectral components via singular value decomposition, achieving ~30 cm/s precision and enabling the detection of Earth-like planets around solar-type stars.

Original authors: Sahar Shahaf, Barak Zackay

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sahar Shahaf, Barak Zackay

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Cosmic Wobble" Problem

Imagine you are trying to find a tiny, invisible mouse (an Earth-like planet) hiding in a massive, noisy library (a star). You can't see the mouse, but you know that if it's there, it will make the library shelves wobble slightly as it orbits.

In astronomy, we look for these "wobbles" by measuring the Radial Velocity of a star. As a planet orbits, it pulls the star toward us and then pushes it away. This changes the color of the starlight slightly (the Doppler effect), just like the pitch of a siren changes as an ambulance drives past you.

The Problem: Stars aren't quiet libraries. They are noisy, chaotic places. They have sunspots, magnetic storms, and surface bubbles (convection) that change the shape of their light. These changes can look exactly like a planet wobble, or they can hide a real planet signal. It's like trying to hear a whisper in a library where the librarian is constantly shouting, rearranging books, and slamming doors.

The Solution: The "Spectral Factorization" Method

The authors (Shahaf and Zackay) have developed a new mathematical "magic trick" to separate the planet's whisper from the librarian's shouting. They call it Spectral Factorization.

Here is how it works, using a few analogies:

1. The "Chopped-Up Song" (STFT)

Imagine the star's light spectrum is a long, complex song. To understand it, the scientists don't listen to the whole song at once. Instead, they use a technique called the Short-Time Fourier Transform (STFT).

Think of this like taking a long movie and slicing it into thousands of tiny, overlapping frames. They look at small chunks of the light spectrum at a time. This allows them to see how the "song" changes moment by moment.

2. The "Deconstructing the Orchestra" (SVD)

Once they have these tiny slices, they use a mathematical tool called Singular Value Decomposition (SVD).

Imagine an orchestra playing a piece of music.

The Old Way: You try to guess which instrument is playing which note by listening to the whole mix. It's messy.
The New Way (This Paper): The scientists act like a super-smart audio engineer. They take the recording and mathematically "un-mix" it. They separate the recording into two parts:
1. The Instruments (Principal Spectra): The unique "sound" of the star's features (like a sunspot or a quiet patch).
2. The Conductor's Baton (Kernels): The timing and intensity of when those features appear.

The brilliant part of this paper is that they don't need to know what the instruments sound like beforehand. In previous methods, you had to know exactly what a sunspot looks like to remove it. Here, the math figures out what the "instruments" sound like just by looking at the changes in the music over time.

3. Finding the Wobble (Phase Differences)

Once they have separated the "noise" (stellar activity) from the "signal," they look for the planet.

In the mathematical world of this paper, a planet's movement doesn't change the shape of the light; it just shifts the timing (phase) of the wave.

Analogy: Imagine two identical runners. One is running on a track. The other is running on the same track but starts 1 second later. If you look at their footprints, they are identical, but the timing is shifted.
The scientists measure this tiny timing shift. Because they have already mathematically removed the "noise" (the sunspots and storms), the only thing left causing a timing shift is the planet pulling the star.

The Results: Hearing the Whisper

The team tested this method in two ways:

Fake Data: They created a computer simulation of a star with a fake planet and fake sunspots. Their method successfully found the fake planet and ignored the fake sunspots, reaching a precision of about 30 centimeters per second. That is incredibly precise—like measuring the speed of a snail with a stopwatch accurate to a fraction of a second.
Real Data: They applied this to real data from the EXPRES telescope looking at two stars: HD 34411 and Tau Ceti.
- They found that the method could reach the "instrument limit," meaning the telescope itself was the only thing stopping them from being even more precise.
- They successfully filtered out the "p-mode" noise (the star's natural pulsations, like a heartbeat) and recovered coherent signals.

Why This Matters

This paper is a major step toward finding Earth-like planets.

Before: We were limited by how well we could model a star's "mood swings" (activity). If the star was too active, we couldn't find small planets.
Now: This method is "data-driven." It learns the star's behavior directly from the data, rather than relying on imperfect guesses. It acts like a noise-canceling headphone for starlight, allowing us to hear the faint "wobble" of an Earth twin.

Summary in One Sentence

The authors invented a smart mathematical way to "un-mix" a star's chaotic light, separating the noise of stellar storms from the tiny, rhythmic wobble caused by a hidden planet, allowing us to detect Earth-like worlds with unprecedented precision.

1. Problem Statement

Extracting precise stellar radial velocities (RV) is a primary method for detecting exoplanets, but it faces a fundamental limit: stellar spectral variability. While modern spectrographs (like EXPRES) have achieved instrumental stability allowing for cm/s precision, stellar phenomena such as magnetic activity, convection, and surface inhomogeneities (spots/faculae) induce spectral distortions that mimic or obscure genuine Doppler shifts.

Previous methods often rely on:

Spectral compression: Reducing thousands of lines to a few metrics.
Per-line analysis: Analyzing lines independently.
Data-driven decompositions: Projecting time series onto learned variability components.

The authors' previous work (Paper I) introduced a linearized factorization approach assuming the "principal spectra" (chromatic responses to variability) were known. However, in many practical cases, these spectral responses are unknown. This paper addresses the challenge of simultaneously recovering unknown principal spectra and radial velocity shifts from a spectroscopic time series without prior knowledge of the stellar variability model.

2. Methodology

The core of the method is a generalization of the Short-Time Fourier Transform (STFT) combined with Singular Value Decomposition (SVD).

A. Mathematical Formulation

The stellar spectrum $S(\lambda)$ is modeled as a sum of convolutions between principal spectra ( $f^{(k)}$ , representing chromatic components) and principal kernels ( $U^{(k)}$ , representing time-dependent weights).
In the Fourier domain (conjugate to inverse wavelength $\zeta$ ), a spectral segment centered at $\Lambda$ is expressed as a bilinear form:
$\tilde{S}_{t\Lambda}(\zeta) = \sum_k \tilde{U}^{(k)}_t \tilde{f}^{(k)}_\Lambda(\zeta)$
Here, $\tilde{U}^{(k)}_t$ are the time-varying kernels, and $\tilde{f}^{(k)}_\Lambda$ are the fixed principal spectra for that segment.

B. Factorization via SVD

For a given inverse wavelength $\zeta$ , the time series of spectral segments forms a matrix $\tilde{\mathbf{S}}_\zeta$ (rows = time, columns = segments). This matrix is factorized using SVD:
$\tilde{\mathbf{S}}_\zeta = \mathbf{U}_\zeta \mathbf{\Sigma}_\zeta \mathbf{V}^\dagger_\zeta$

$\mathbf{U}_\zeta$ (Left Singular Vectors): Correspond to the time-dependent principal kernels.
$\mathbf{V}^\dagger_\zeta$ (Right Singular Vectors): Correspond to the principal spectra.
$\mathbf{\Sigma}_\zeta$ (Singular Values): Indicate the significance of each component.

In the high Signal-to-Noise Ratio (SNR) regime, the zeroth-order component ( $k=0$ ) dominates and captures the coherent Doppler shift, while higher-order components capture stellar variability (activity, p-modes).

C. Radial Velocity Extraction

The Doppler shift manifests as a complex phase in the Fourier domain. The radial velocity difference between two epochs ( $t_i$ and $t_j$ ) is derived from the phase difference of the zeroth-order kernels:
$2\pi \zeta v = -\arg\left( \frac{\tilde{U}^{(0)}_{t_i}}{\tilde{U}^{(0)}_{t_j}} \right)$
The method assumes white Gaussian noise to derive a maximum likelihood estimator for velocity. The final velocity estimate is a weighted average of estimates across all inverse wavelengths, where weights are determined by the Fisher information (scaling with $\zeta^2 \times \text{SNR}^2$ ).

D. Handling Stellar Activity (Cross-Talk)

A critical challenge is "cross-talk," where stellar activity leaks into the zeroth-order kernel, contaminating the velocity measurement.

The authors model this leakage as a perturbation from the first-order kernel ( $k=1$ ).
They propose a linear correction scheme: Extract velocities from both the zeroth ( $v^{(0)}$ ) and first ( $v^{(1)}$ ) order kernels.
Since $v^{(1)}$ is dominated by activity, a model (e.g., harmonic fit or Gaussian Process) is fitted to $v^{(1)}$ and subtracted from $v^{(0)}$ to isolate the true Doppler signal.

3. Key Contributions

Generalization of STFT-SVD: Extends the linearized RV extraction framework to cases where principal spectra are unknown, deriving them empirically from the data itself.
Shot-Noise-Limited Precision: Demonstrates that the method can reach the theoretical instrumental precision limit (shot-noise limit) in the presence of spectral variability, provided the variability is linearizable into a few components.
Activity Mitigation: Introduces a robust, data-driven linear correction technique to remove activity-induced signals by utilizing higher-order SVD components, avoiding the need for external activity models.
Phase Retrieval Framework: Formulates RV extraction explicitly as a phase retrieval problem in the Fourier domain, clarifying the statistical properties of the estimator.

4. Results

A. Synthetic and SOAP Simulations

White Noise: Validated on synthetic Sun-like star data with injected RV signals (amplitude ~~4 m/s). The method recovered velocities with residuals consistent with the estimated shot-noise error (~~58 cm/s).
Stellar Activity (SOAP): Using SOAP-GPU simulations with spots and faculae, the method successfully separated the 10-day rotation signal (activity) from a 17-day planetary signal (30 cm/s amplitude).
- The zeroth-order velocity showed contamination from the rotation period.
- The first-order velocity captured the rotation signal.
- Linear subtraction of the fitted activity model from the first-order data cleaned the zeroth-order velocity, recovering the injected planetary signal with high fidelity.

B. Real Data: HD 34411 and $\tau$ Ceti

The method was applied to EXPRES (Extreme Precision Spectrograph) observations.

HD 34411:
- Injected a synthetic 75-day signal with 50 cm/s semi-amplitude.
- Recovered signal: $47.3 \pm 6.8$ cm/s.
- Residuals showed no significant long-term correlations.
- Achieved a precision floor of ~30 cm/s, matching the instrumental shot-noise limit.
- Analyzed p-mode modulation: Showed that RMS velocity scatter decreases with exposure time, following the expected $\sin(x)/x$ trend for p-mode averaging, reaching ~30 cm/s for 8-minute exposures.
$\tau$ Ceti (HD 10700):
- Analyzed 174 measurements over 15.5 months.
- Overall RMS: ~100 cm/s.
- High-cadence nights revealed p-mode oscillations with periods of 5–5.5 minutes.
- No zero-point drift was detected, confirming the stability of the underlying velocity series.
- The method successfully isolated stochastic p-mode variability from the stable velocity baseline.

5. Significance and Future Outlook

Exoplanet Detection: This framework represents a significant step toward detecting Earth-like planets (requiring ~10 cm/s precision) around solar-type stars by effectively separating planetary signals from stellar noise.
Data-Driven Approach: Unlike methods requiring physical models of stellar surfaces, this approach learns the variability directly from the data, making it applicable to a wide range of stars without prior characterization.
Astrophysical Applications: Beyond RVs, the decomposition of principal spectra and kernels offers a pathway for Doppler imaging, tomography, and measuring stellar obliquity during transits without assuming specific rotation profiles.
Limitations: The method assumes spectral variability can be linearized into a small number of components. It requires sufficient phase coverage of the variability and high SNR. Future work will address highly active stars, correlated noise, and optimized observing strategies to mitigate p-mode modulation.

In conclusion, Shahaf & Zackay present a powerful, linearized, and statistically rigorous method that pushes radial velocity precision to the shot-noise limit, effectively decoupling instrumental and stellar noise from genuine planetary signals.