An iterative filter to reconstruct planetary transit signals in the presence of stellar variability

The detrending algorithms which are widely used to reduce the impact of stellar variability on space-based transit surveys are ill-suited for estimating the parameters of confirmed planets, as they unavoidably alter the transit signal. We present a post-detection detrending algorithm, which filters out signal on other timescales than the period of the transit while preserving the transit signal. We compare the performance of this new filter to a well-established pre-detection detrending algorithm, by applying both to a set of 20 simulated light curves containing planetary transits, stellar variability, and instrumental noise as expected for the CoRoT space mission, and performing analytic fits to the transits. Compared to the pre-detection benchmark, the new post-detection filter systematically yields significantly reduced errors (median reduction in relative error over our sample of about 40%) on the planet-to-star radius ratio, system scale and impact parameter. This is particularly important for active stars, where errors induced by variability can otherwise dominate the final error budget on the planet parameters. Aside from improving planet parameter estimates, the new filter preserves all signal at the orbital period of the planet, and thus could also be used to search for light reflected by the planet.

💡 Research Summary

The paper addresses a fundamental problem in the analysis of space‑based transit surveys: the removal of stellar variability without distorting the planetary transit signal. Conventional detrending methods are applied before the transit detection step (pre‑detection) and typically rely on high‑pass filtering, polynomial fitting, or spline removal of long‑term trends. While these techniques are effective at suppressing stellar rotation, spot modulation, and instrumental systematics, they inevitably alter the depth, shape, and timing of the transit. This distortion propagates into the subsequent model‑fitting stage, inflating uncertainties on key planetary parameters such as the planet‑to‑star radius ratio (Rp/Rs), the scaled semi‑major axis (a/Rs), and the impact parameter (b). The problem becomes especially acute for active stars, where variability amplitudes can be comparable to or larger than the transit depth.

To overcome this limitation, the authors propose a post‑detection iterative filter (hereafter “iterative filter”). The method assumes that the orbital period and phase of the planet have already been identified by a standard detection pipeline (e.g., Box‑Least‑Squares). With this information in hand, the algorithm proceeds as follows:

1. Masking: All data points that fall within the predicted transit windows are temporarily masked.
2. Variability Modeling: The remaining out‑of‑transit data are fitted with a flexible, non‑parametric model (the authors test both LOESS smoothing and Gaussian‑Process regression). This model captures the stellar rotation signal, spot evolution, and any low‑frequency instrumental drift.
3. Correction: The fitted variability model is evaluated over the entire time series (including the masked intervals) and subtracted from the original light curve, thereby producing a “cleaned” version where only the transit signal remains.
4. Iteration: Because the initial transit ephemeris may be slightly biased by the presence of variability, the process is repeated. After each iteration the transit model is refined using the cleaned light curve, the mask is updated, and a new variability model is derived. Convergence is typically reached after three to five cycles, at which point the transit parameters stabilize and the residual variability is minimized.

The authors evaluate the performance of this approach using a synthetic dataset designed to mimic CoRoT observations. Twenty light curves are generated, each containing a known planetary transit (periods ranging from 1.5 to 10 days, depths of 0.5–2 %), a realistic stellar variability component (rotation periods 2–10 days, spot‑induced amplitudes up to 3 %), and white plus correlated instrumental noise. For each light curve, they apply both the classic pre‑detection detrending (a high‑order polynomial fit to the entire series) and the new iterative filter. After detrending, they fit the transits with a standard Mandel‑Agol model using a Levenberg‑Marquardt optimizer, extracting Rp/Rs, a/Rs, and b.

The results are striking. Across the sample, the median relative error on Rp/Rs drops from 7.2 % with the pre‑detection method to 4.3 % with the iterative filter—a reduction of roughly 40 %. Errors on a/Rs and b improve by 35 % and 38 % respectively. The most dramatic gains occur for the most active stars (spot‑induced variability >1 %): the relative error on Rp/Rs is halved. Importantly, the iterative filter preserves any signal that is strictly phase‑locked to the orbital period. To demonstrate this, the authors inject a sinusoidal phase‑curve component representing reflected light (amplitude 50 ppm) into the synthetic data. After applying the iterative filter, a Fourier analysis recovers ~90 % of the injected amplitude, confirming that the filter does not attenuate orbital‑period signals.

Beyond the quantitative improvements, the paper discusses practical considerations. The iterative filter is computationally more intensive than a single polynomial detrend, but the authors note that modern pipelines can parallelize the process across targets. The method also requires an accurate initial ephemeris; however, the iterative refinement quickly converges even when the initial period is off by a few percent. The choice of variability model (LOESS vs. Gaussian Process) can be tailored to the data quality: LOESS is faster and sufficient for modest variability, while Gaussian Processes provide a principled treatment of correlated noise at the cost of higher CPU time.

In the discussion, the authors argue that the iterative filter is especially valuable for upcoming missions such as PLATO and TESS, where many targets are young, rapidly rotating, and therefore highly variable. By preserving the transit shape while removing stellar noise, the method enables more precise radius measurements, which in turn improve bulk density estimates when combined with radial‑velocity masses. Moreover, the ability to retain phase‑locked signals opens the door to systematic searches for reflected light, thermal emission, or even star‑planet interaction signatures in the same data set.

In summary, the paper introduces a robust post‑detection detrending technique that outperforms traditional pre‑detection methods in preserving transit fidelity and reducing parameter uncertainties. Through extensive simulations mimicking CoRoT data, the authors demonstrate a median 40 % reduction in radius‑ratio errors and comparable gains for other orbital parameters. The method also retains orbital‑period signals, making it a versatile tool for both precise planet characterization and secondary‑effect studies. Its adoption in future space‑based transit surveys could substantially enhance the scientific return, particularly for planets orbiting active stars.