Characterisation of the TOI-421 planetary system using CHEOPS, TESS, and archival radial velocity data ^††thanks: Based on data from CHEOPS Guaranteed Time Observations, collected under Programme ID CH_PR100024. The raw and detrended photometric time series data are available in electronic form at the CDS via anonymous ftp to cdsarc.cds.unistra.fr (130.79.128.5) or via https://cdsarc.cds.unistra.fr/cgi-bin/qcat?J/A+A/

CHEOPS performed a total of six observations (visits) of TOI-421 between November 2021 and December 2022 (see Table 4). Each visit contains a single transit event and also includes data being acquired both before and after the transit. Each individual visit comprises several CHEOPS orbits and a single orbit covers roughly 100 minutes. However, the target cannot be observed throughout an entire CHEOPS orbit, due to Earth occultations and South Atlantic Anomaly (SAA) crossings. This leads to gaps in the observations with a width that varies from visit to visit depending on target coordinates and observation time. The duration and the observing efficiency (i.e. the relative amount of time in which the target is visible to CHEOPS throughout a visit) of each visit are also listed in Table 4.

The CHEOPS observations are available as sub-array data products (Benz et al. 2021) at a cadence equal to the exposure time of 60 s. The sub-arrays contain a circular region around the target with a radius of 100 pixels. We processed the data with PSF photometry utilising PIPE⁵⁵5http://github.com/alphapsa/PIPE(see also Morris et al. 2021; Szabó et al. 2021; Brandeker et al. 2022). In general, CHEOPS observations are affected by instrumental noise such as stray light from the Earth and Moon (Moon glint), smearing effects, or spacecraft jitter. The flux measurements usually show a particularly strong correlation with the spacecraft roll angle (see also Lendl et al. 2020; Bonfanti et al. 2021a). The spacecraft is designed to rotate around itself exactly once every orbit around the Earth. Therefore, the roll angle parameter is directly linked to the orbital position of the spacecraft. Instrumental noise must be accounted for during data analysis in order to identify and measure the transit signals of the planets (see Section 4). Prior to performing data analysis, we removed all points that were flagged by PIPE, which for example include those contaminated by cosmic rays. We performed a sigma clip** and removed all points with the median absolute deviation (MAD) higher than five to discard outliers. We also removed points with peculiar high backgrounds by removing any point with a background larger than four times the median background value, as well as points with a peculiar high pointing offset by removing all points with a centroid-offset of more than one pixel.

TOI-421 was observed by TESS in Sectors 5, 6 and 32 at 2 min cadence. The TESS photometric baseline, therefore, spans from July 2018 to December 2020 (see Table 3 for further details). In our analysis, we used the Pre-search Data Conditioning Simple Aperture Photometry (PDCSAP) flux, provided by the Science Processing Operations Center (SPOC) pipeline (Smith et al. 2012; Stumpe et al. 2014; Jenkins et al. 2016). This light curve type is subject to more treatment than the Simple Aperture Photometry (SAP) flux and is specifically intended for detecting exoplanets. The pipeline attempts to remove systematic artefacts, while kee** planetary transits intact. It also already accounts for dilution caused by nearby contaminating stars. The data were downloaded from the Mikulski Archive for Space Telescopes⁶⁶6See https://mast.stsci.edu/portal/Mashup/Clients/Mast/Portal.html. (MAST). We did not apply outlier removal in the TESS dataset. To save computation time and simplify the treatment of long-term trends caused by stellar activity in the dataset we only used data points, which we expected to be within a 0.25 day window centred around the expected transit mid-point. We used the following transit timing parameters to define this window after having performed a preliminary analysis of the data to retrieve these values: $T_{\rm 0,b}=2459189.7336$ BJD, $P_{\rm b}=5.1975736$ days, and $T_{\rm 0,c}=2459195.30753$ BJD, $P_{\rm c}=16.0675425$ days.

Carleo et al. (2020) published a total of 156 radial velocity (RV) measurements of TOI-421 obtained between February 2019 and March 2020 by four different instruments: The FIbre-fed Échellé Spectrograph at La Palma, Spain (FIES; Frandsen & Lindberg 1999; Telting et al. 2014); the High Accurarcy Radial velocity Planet Searcher at La Silla, Chile (HARPS; Mayor et al. 2003); the HIgh Resolution Échelle Spectrometer on Hawaii, USA (HIRES; Vogt et al. 1994); and the Planet Finder Spectograph at Las Campanas, Chile (PFS; Crane et al. 2006, 2008, 2010). An overview of the observation dates and the number of acquired RV points is listed in Table 4. A detailed description of the data can be found in Carleo et al. (2020).

In the case of the FIES, HIRES, and PFS data we used the reduced RV data published by Carleo et al. (2020) and made available at the Centre Données astronomiques de Strasbourg (CDS) astronomical data centre⁷⁷7https://cdsarc.cds.unistra.fr/viz-bin/cat/J/AJ/160/114. The corresponding data products provide only the time stamp of the observation and the measured RV value including its uncertainty. The HARPS data reduced using the HARPS data reduction software (DRS) are also available at the CDS and accompanied with line profile asymmetry indicators, namely the FWHM, the bisector inverse slope of the cross-correlation function (CCF) and the CaII H and K lines activity indicator $\log{R^{\prime}_{HK}}$.

We did not use the published HARPS radial velocities reduced using the DRS, but employed a novel reduction approach, using the CCFs of the published HARPS spectra, along with their $\log{R^{\prime}_{HK}}$ activity indicators. Namely, we performed a Skew Normal fit onto the CCFs (Simola et al. 2019), and for each epoch of observation we derived the radial velocity, the Skew-Normal full width half maximum (FHWM_SN), the contrast ($A$), and the asymmetry parameter ($\gamma$) along with their uncertainties. Details of this reduction process are described in Bonfanti et al. (2023) and Fridlund et al. (2023, accepted).

We used the PyCHEOPS python package (Maxted et al. 2021) to correct the CHEOPS data for instrumental effects and long-term temporal trends. To this end we excluded all data points that were observed during a transit event and then fitted coefficients of detrending models on the out-of-transit data only. The detrending models were chosen independently for each visit by individually adding detrending basis vectors one by one and retaining an additional vector only when supported by a higher Bayes factor computed from the Bayesian information criterion (BIC; Schwarz 1978). In this way we used up to second-order harmonics of the roll angle, up to second-order polynomials in PSF centroid position and time, and a first-order linear model on background. Additionally, a 18-segment spline was fit on top of these models. The spline fits small-scale flux changes as a function of roll angle to remove glint effects. After fitting the coefficients of the detrending models on the out-of-transit data, the models were also applied to the in-transit-data and the uncertainties of the individual data-points were inflated to account for the uncertainties of the detrending coefficients. In our joint analysis we then used the detrended datasets only and no additional detrending models.

Carleo et al. (2020) have shown that the radial velocity data are affected by stellar activity. Specifically, they performed an in-depth frequency analysis of the HARPS data and found a significant peak in the RV residuals following the subtraction of the planetary signals corresponding to a period of $\sim 42$ days and an RV semi-amplitude of $\sim 2.4$ m/s. They also recovered this peak in the periodogram of the $\log{R^{\prime}_{HK}}$ and FWHM activity indicators in the HARPS DRS data. They attribute this peak to a plage-dominated activity signal after carefully analysing a contour map of the HARPS CCF residuals and also searching for rotational modulation in the available WASP-South and TESS photometry. In their subsequent global analysis, Carleo et al. (2020) account for this activity induced signal by fitting a third Keplerian to the RV data, resulting in an additional sinusoidal signal at a period of $43.24^{+0.57}_{-0.55}$ days and with a semi-amplitude of $2.4\pm 0.3$ m/s. The epoch of this signal is however fairly unconstrained, with an uncertainty of $\sim 5$ days, as reported by Carleo et al. (2020). Furthermore, stellar activity in general is not strictly sinusoidal (e.g. Lanza et al. 2001) and adding sinusoidal baselines has been shown to sometimes introduce spurious harmonics that bias the final result (Pont et al. 2011; Tuomi et al. 2014; Rajpaul et al. 2015). Because the epoch is unconstrained and the known problems of adding a simple sinusoidal model, we aim to provide alternative baseline models. To compare different approaches, we analysed the HARPS data, which is the only RV dataset that also provides activity indicators, obtained through two different extractions methods and using two different approaches, as explained below. To help constraining both the ephemeris and the eccentricity of the orbits, we also included all photometric data in this particular analysis. We used the stellar parameters derived in Section 2. To show that the corresponding results are not an artefact of the data reduction process, we performed the analysis once using the DRS-reduced data published in Carleo et al. (2020) and once using the data reduced with the novel approach of performing a Skew Normal fit to the CCFs (see Section 3.3). The later reduction approach will be denoted as SNF.

The first baseline model we used is a model fitting polynomial functions to the activity indicators acompanying both the HARPS DRS and HARPS SNF data. The analysis was performed using the MCMCI code (Bonfanti & Gillon 2020). As previously done with the CHEOPS data, we choose the detrending models independently for both HARPS datasets by individually adding detrending basis vectors one by one and retaining an additional vector only when supported by a higher Bayes factor computed from the BIC. In this way, we ended up using a second-order polynomial in FWHM and a first-order polynomial in $\log{R^{\prime}_{HK}}$ for the DRS data and first-order polynomials in FWHM_SN, contrast, and $\log{R^{\prime}_{HK}}$ for the SNF data. We did also perform a breakpoint analysis (Simola et al. 2022; Bonfanti et al. 2023) to determine possible parts of the RV time series that would require different detrending basis vectors, but did not find any breakpoints indicating that the whole time series can be adequately detrended with the same detrending model. The 14 individual transit events present in the TESS data were treated as 14 individual light curves. For each of them a second-order polynomial in time was used as a baseline model.

The second baseline model we used is a Gaussian process (GP; Rasmussen & Williams 2005) using a critically damped Simple-Harmonic-Oscillator (SHO) kernel (Foreman-Mackey et al. 2017). The kernel ($\kappa$) is defined by an amplitude of variations $S_{0}$ and an angular frequency of the variations $\omega_{0}$, which can be translated to a period of variations by $P=2\pi/\omega_{0}$:

We further constrained the period of the variations by fitting a Gaussian function to the strongest peak in the periodogram of the FWHM_SN activity indicator in the HARPS SNF data, which resulted in a prior on the period of variations at $41.2\pm 2.1$ days. We also observed a similar peak in the periodogram of the $\log{R^{\prime}_{HK}}$ activity indicator. The analysis was performed using the allesfitter Python package (Günther & Daylan 2021). The hybrid spline baseline option available in allesfitter was applied to the TESS photometry.

Table 8 shows the retrieved planetary parameters for all combinations of RV baseline models and datasets, as well as the values reported in Carleo et al. (2020). We find that as expected the retrieved planetary radii are not dependent on the chosen RV baseline model. We do not find significant differences in the retrieved mass and eccentricity of planet b, with all results agreeing with each other well within their $1\sigma$ confidence intervals. For the mass of planet c, we find that our new analyses result in a smaller planetary mass compared to the value reported in Carleo et al. (2020). The masses retrieved when using the SHO-GP baseline are smaller than those obtained using the polynomials baseline, but all four new model approaches are consistent with each other within their $1\sigma$ confidence intervals. In the case of the eccentricity of planet c, we find that all of the new analyses find a higher eccentricity than reported in Carleo et al. (2020). Again, we also find that all four new approaches agree with each other well within their $1\sigma$ confidence intervals.

Regardless of the adopted technique to extract the RV time series and the indicators of stellar activity (either DRS or SNF), we obtain consistent results in terms of planetary masses and eccentricities after applying the polynomial detrending baseline based on the activity indicators. This proves that the SNF-based technique enables us to infer reliable RV-related parameters, and thus we adopt the SNF-based HARPS RVs also in the following joint analysis (Sec. 4.3).

Because two completely independent approaches (i.e. the polynomial-based and GP-based detrending) retrieve consistent results independently on how the data were reduced, we believe that both approaches are justified. To choose a model for our final global analysis, we also looked at the periodogram of the residuals when applying the different baseline models (see Figure 1). We find that the approach using an SHO-GP baseline performs best at removing the activity-induced signal. We therefore chose this approach for our global analysis using all of the available radial velocity datasets.

We performed a joint analysis of all photometric and RV datasets using allesfitter. For each of the planets we fitted for the radius-ratio $R_{\rm p}/R_{\star}$, the cosine of the inclination $i$, the term $(R_{\rm p}+R_{\star})/a$, the transit mid-time, the orbital period, the RV semi-amplitude, and the parameters $\sqrt{e}\,\rm cos(\omega)$ and $\sqrt{e}\,\rm sin(\omega)$, which determine the eccentricity $e$ and the argument of periapsis $\omega$.

While using uniform priors for the planetary parameters, we did apply Gaussian priors on the stellar radius and mass according to the values presented in Section 2. This resulted in a prior on the mean stellar density, which in turn constrains the transit model. We also applied a quadratic limb-darkening (LD) law using the parametrisation proposed by Kip** (2013) and used Gaussian priors on the LD coefficients according to theoretically computed values including their uncertainties using the LDCU Python package (see Table 9). LDCU¹⁰¹⁰10https://github.com/delinea/LDCU is a modified version of the python routine implemented by Espinoza & Jordán (2015) that computes the LD coefficients and their corresponding uncertainties using a set of stellar intensity profiles accounting for the uncertainties on the stellar parameters. The stellar intensity profiles are generated based on two libraries of synthetic stellar spectra: ATLAS (Kurucz 1979) and PHOENIX (Husser et al. 2013). To check the validity of this additional prior, we performed the same analysis presented here also without constraining the LD coefficients except for their physical boundaries. Although in this case the LD coefficients are only poorly constrained, all of the retrieved planetary parameters are consistent with those presented in this work within their $1\sigma$ confidence intervals.

As discussed in Carleo et al. (2020), the TESS light curves are contaminated by a M dwarf companion (GAIA-ID 2984582227215748224) at an angular distance of $29.4^{\prime\prime}$. The contamination was computed by Carleo et al. (2020) to be $1.8\pm 0.4\%$. The TESS PDSCAP light curves already account for the dilution of the light curve caused by possible contaminants via the contamination estimate in the TESS Input Catalog (TIC). We note that the estimated contamination ratio of 0.0246 is consistent with the value computed by Carleo et al. (2020). Because we used PSF and not aperture photometry for the extraction of the CHEOPS photometry, the CHEOPS data are not affected by contamination.

For the photometric observations, we did fit for an additional white-noise term per instrument. For the RV data, we also added a further jitter-term and a constant offset for each instrument. In terms of baseline models, we used the already detrended CHEOPS data (see Section 4.1) and applied the hybrid spline baseline option in allesfitter to the TESS photometry. As discussed in Section 4.2, for the baselines of the RV data we used a GP with an SHO kernel. While fitting for individual amplitudes of variations $S_{0}$ for each instrument, we jointly fit for a single angular frequency (and therefore period) of variations. We again imposed a prior on this parameter defined by the strongest peak in the periodogram of the FWHM_SN activity indicator in the HARPS SNF data.

The global analysis was performed using the dynamic nested sampling algorithm (e.g. Feroz & Hobson 2008; Feroz et al. 2019) implemented in allesfitter via the dynesty Python package (Speagle 2020). The results of our fit and the successively computed planetary parameters are listed in Table 9. We note that both planetary radii and masses are consistent with the results presented in Carleo et al. (2020) at the $1\sigma$ level. The uncertainties on the radius of both planets decreased by a factor of $\sim 2.5$. Despite the more conservative uncertainty on the stellar mass, the uncertainty on the mass of planet b also decreased. Noteworthy is also the difference in the median values of the transit timing. There is a difference of 74 seconds when comparing the period of planet b as derived by us and as derived by Carleo et al. (2020). This adds up to a difference of 87 minutes ($\sim 1.5$ hours) per year. In the case of planet c, the two periods differ by 56 seconds, which amounts to 21 minutes per year. Considering the prospects of future follow-up spectroscopic observations (see Section 8), these differences of several hours since the publication of the original ephemeris in Carleo et al. (2020) are especially relevant. Finally, we note that we have also been able to significantly detect the eccentricity of both planets, which is especially important in the context of planet formation models and possible observations of the planetary eclipses for atmospheric studies.

The TOI-421 system is composed by two close-in low mass planets ($M_{b}\approx 6.7\,M_{\oplus}$, $M_{c}\approx 14.1\,M_{\oplus}$) in moderate eccentric orbits ($e_{b}\approx 0.13$, $e_{c}\approx 0.19$) and near a 3:1 mean motion resonance ($P_{c}/P_{b}\approx 3.091$, Table 9). Because the masses of the planets are in the super-Earth regime, the stability should be assured (eg. Elser et al. 2013). Nevertheless, the significant eccentricity of the planets can introduce many higher order mean motion resonances that can disturb the system.

In order to get a reliable and comprehensive view of the stability of the system, we performed a global frequency analysis (Laskar 1990, 1993) in the same way as achieved for other planetary systems (e.g. Correia et al. 2005, 2010). The system was integrated on a regular 2D mesh of initial conditions in the vicinity of the best fit (Table 9). We used the symplectic integrator SABAC4 (Laskar & Robutel 2001), with a step size of $5\times 10^{-4}$ yr and general relativity corrections. Each initial condition was integrated for 5000 yr, and a stability indicator, $\Delta=|1-n_{b}^{\prime}/n_{b}|$, was computed. Here, $n_{b}$ and $n_{b}^{\prime}$ are the main frequency of the mean longitude of the planet over 2500 yr and 5000 yr, respectively, calculated via the frequency analysis (Laskar 1993). In Figure 4, the results are reported in colour, where orange and red represent strongly chaotic trajectories with $\Delta>10^{-2}$, while extremely stable systems for which $\Delta<10^{-8}$ are shown in cyan and blue. Yellow indicates the transition between the two, with $\Delta\sim 10^{-4}$.

In a first experiment, we explored the stability of the system by varying the orbital period and the eccentricity of the inner planet (Fig. 4, left). We confirm that the best fit solution is stable, despite the existence of some higher order mean motion resonances for eccentricities $e_{b}>0.3$. We also note the presence of a large unstable region at low eccentricity on the right hand side of the figure, corresponding to the 3:1 mean motion resonance. The inner planet is close enough to the star to undergo strong tidal interactions that drive the period ratio to a value above the exact resonance, as it was already reported for many other near resonant systems comprising low-mass planets (e.g. Lissauer et al. 2011; Delisle & Laskar 2014).

As the eccentricity plays an important role in the stability of the system, in a second experiment, we varied the eccentricities of both planets (Fig. 4, right). Though the eccentricities are not very well constrained in the best fit solution (Table 9), we observe that the system is stable as long as $e_{b}\lesssim 0.2$ and $e_{c}\lesssim 0.3$, that is, the system remains stable even if one takes the $3\sigma$ maximum for both eccentricities. We conclude that the TOI-421 planetary system as presented in Table 9 is reliable and resilient to the uncertainties in the determination of the orbital parameters.