betaFeedback

WorksPhysics and Astronomy

Revisiting the Photometric Modulation Amplitudes of Kepler Planet Hosts: A Confounder-Adjusted Test of Stellar Spin-Orbit Alignment

Abstract

T. Mazeh and coworkers pioneered a method using rotational photometric modulation as a proxy to infer stellar obliquities in Kepler transiting planetary systems, based on the expected positive correlation between spot-induced modulation amplitudes and stellar inclinations. They found that planet hosts cooler than the Sun showed larger modulation amplitudes than comparison stars, consistent with low stellar obliquities. For hotter hosts, however, they reported the opposite - a negative amplitude difference (hosts minus comparison stars) - which is difficult to interpret in terms of viewing geometry. We show that this comparison is affected by confounding variables not controlled for in the sample selection. One such confounder is apparent magnitude, whose distributions differ markedly between the two samples for hotter stars and affect the observed modulation amplitudes. We develop resampling and likelihood-based approaches that enable comparison of the amplitude distributions at fixed apparent magnitude, stellar radius, and rotation period. The previously reported negative amplitude difference for hot hosts is no longer present in either analysis. The likelihood-based analysis finds a positive amplitude difference up to ∼6100 K\sim 6100\,\mathrm{K} that becomes consistent with zero at higher temperatures, while the resampling analysis shows a similar but less pronounced trend. This result is qualitatively consistent with independent vsin⁡iv\sin i analyses suggesting low obliquities for cooler super-Earth and sub-Neptune hosts and broader obliquity distributions for hotter ones. This may also indicate a gradual change in the obliquity distribution with stellar temperature, although changes in spot distributions or selection effects could also contribute to the photometric trends in hot stars.

Introduction

Rossiter–McLaughlin (RM) measurements of hot Jupiters have shown that close-in giant planets span a wide range of spin–orbit configurations (e.g., [64, 3]. These measurements also revealed a clear dependence on hot-star temperature: hot Jupiters around stars cooler than the Kraft break tend to have low obliquities, whereas those around hotter stars show a broader distribution ([67, 59]).

One interpretation of this temperature trend is that the low obliquities of cool stars are due to tidal re-alignment of the stellar spin ([67, 2]), in which case the present-day alignment of cool hosts need not be primordial and hot Jupiters in general could initially have a broad obliquity distribution. Such a distribution could arise naturally from high-eccentricity migration, rather than from smooth migration within an initially aligned disk (see review by R. I. Dawson & J. A. Johnson [17]). Alternatively, the trend may arise from properties of the proto-planetary disks rather than from the planets themselves. For example, disks may frequently form misaligned with their host stars and/or subsequently undergo realignment in a manner that depends on stellar temperature or mass (e.g., [5, 36, 21, 62, 7]). In this case, even hot Jupiters that migrate smoothly within their disks could end up on misaligned orbits. These possibilities are not mutually exclusive and may operate simultaneously. In any case, the tem perature dependence provides an important clue to how stellar spins and planetary orbital planes are established and subsequently evolve.

Obliquity measurements for close-in super-Earths and sub-Neptunes with radii ≲4R⊕\lesssim4 R_{\oplus},1 which have been discovered in large numbers by transit surveys such as Kepler (W. J. Borucki et al. [8]) and TESS (G. R. Ricker et al. [58]), help distinguish between these interpretations. Their typical masses are roughly an order of magnitude or more lower than those of hot Jupiters, resulting in much weaker tidal interactions. The stellar obliquities in these systems are therefore expected to retain more primordial information. Many of these systems also contain multiple transiting planets, suggesting that nearly coplanar configurations are common (J. J. Lissauer et al. [38]; S. Tremaine & S. Dong [63]; D. C. Fabrycky et al. [20]); the intrinsic fraction of such systems may be even higher after accounting for transit geometry and detection incompleteness (W. Zhu et al. [70]; M. Y. He et al. 2019). Such near-coplanar architectures also suggest that many of these systems have preserved orbital configurations inherited from their protoplanetary disks.

Despite the observational challenges compared with hot Jupiters, the number of RM measurements for such small transiting planets has been growing in recent years (Figure 1). Although a population analysis based on a well-defined selection function has yet to be performed, the existing measurements suggest that low obliquities might be common among small-planet systems around cool stars (top panel). At least for systems with multiple transiting planets, statistically significant evidence for preferentially low obliquities has been reported (S. Albrecht et al. [1]; E. Knudstrup et al. [34]; A. S. Polanski et al. [54]; L. B. Handley et al. [27]). However, the current measurements are restricted to stars cooler than ∼6200\sim6200 K.

Sky-projected stellar obliquity as a function of host-star effective temperature for planets with different radii

Figure 1. Sky-projected stellar obliquity λ\lambda as a function of host-star effective temperature for planets with orbital periods P<100P < 100 days. From top to bottom, the panels show systems with Rp<4R⊕R_{\mathrm{p}} < 4 R_{\oplus}, 4R⊕<Rp<8R⊕4 R_{\oplus} < R_{\mathrm{p}} < 8 R_{\oplus}, and Rp>8R⊕R_{\mathrm{p}} > 8 R_{\oplus}. The obliquities, stellar effective temperatures, orbital periods, and published planetary radii are compiled from TEP-Cat (J. Southworth [61]); the horizontal and vertical error bars show the cataloged uncertainties in TeffT_{\mathrm{eff}} and λ\lambda, respectively, where available. This is a heterogeneous compilation rather than a uniformly selected sample, and the measurements differ in precision and reliability. Accordingly, this figure is intended to illustrate the current state of the available measurements rather than to support a statistical inference about the underlying obliquity distributions. The giant-planet sample spans a wide range of obliquities but also shows a pronounced concentration of aligned systems near λ=0∘\lambda= 0^{\circ}, extending up to Teff∼6500T_{\mathrm{eff}} \sim6500 K (X.-Y. Wang et al. [65]).

Utilizing a large sample of transiting planets for which orbits are known to be edge-on, population-level studies focusing on spin-axis inclinations have also been performed (e.g., K. C. Schlaufman [59]; T. Hirano et al. [31]; T. L. Campante et al. [11]; J. N. Winn et al. [68]; D. J. Muñoz & H. B. Perets [47]; E. M. Louden et al. [39]). In particular, a comparison of the projected rotation velocity vsin⁡iv \sin i with the equatorial rotation velocity vv inferred from the stellar rotation period and radius suggests low obliquities for small planets around cool stars ([68, 47]). Using a vsin⁡iv\sin i-only comparison likely less sensitive to rotation-period selection, E. M. Louden et al. [39] further supported this result, finding larger mean vsin⁡iv\sin i for cool transiting-planet hosts (5950–6250 K) than for matched control stars. Although the sample of hotter stars in E. M. Louden et al. [39] was smaller, both that study and a subsequent larger-sample analysis (E. M. Louden et al. [40]) reported evidence for a broader obliquity distribution at higher temperatures. This apparent similarity to the hot-Jupiter trend suggests that the temperature dependence may not arise solely from tidal realignment induced by hot Jupiters.

Here we focus on another statistical approach based on the amplitudes of rotational photometric modulation, which has the potential to probe stellar obliquities of small planets across a broad range of stellar temperatures. The method was pioneered by T. Mazeh et al. [44], hereafter M15, who compared the rotational modulation amplitudes of Kepler stars with detected transiting planets (Kepler Objects of Interest; KOIs) with those of a reference sample of Kepler stars (KIC stars). This comparison probes the correlation between the stellar spin and planetary orbit: if they are aligned, stars with transiting planets tend to be viewed edge-on, resulting in larger modulation amplitudes if spots tend to occur at low latitudes, as on the Sun. For cool stars, they found that planet hosts showed stronger modulation, consistent with low obliquities. For hotter stars, however, they found the opposite trend: planet hosts showed smaller modulation amplitudes than those in the comparison sample. Although this trend may be partly affected by transit-detectability bias, the hot-star result is not straightforward to interpret in terms of viewing geometry alone ([44] (G. Li & J. N. Winn 2016)).

We revisit this problem because comparisons of photometric amplitudes may be affected by confounding variables that were not explicitly controlled for in M15. Here, confounding variables refer to stellar properties correlated with both the modulation amplitude and whether a star belongs to the planet-host or control sample. If the amplitude depends on these properties and their distributions differ between the two samples, the observed amplitude distributions may differ even if the underlying obliquity distributions are identical. In § 2, we explain this concept and identify the additional variables that should be controlled. § 3 then gives a unified mathematical treatment of these confounding variables and the relevant selection effects. In § 4, we present a resampling-based analysis that reweights and resamples the control stars to match the stellar-parameter distribution of the planet-host sample, thereby mitigating the effects of confounding variables. In § 5, we present a likelihood-based analysis that explicitly models the joint distribution of variability amplitude and stellar properties, allowing the amplitude distributions to be compared at fixed stellar properties. In § 6, we repeat the analysis using a smaller subset of stars with effective temperatures from LAMOST low-resolution spectroscopy and find no evidence that the results are sensitive to the source of the effective temperatures. Both analyses show that controlling for confounding variables removes the previously reported negative amplitude offset for hotter KOIs. The likelihood-based analysis yields a positive median shift of approximately 0.07–0.09 dex below ∼5,500\sim5{,}500 K; the shift decreases with increasing temperature, remains positive up to ∼6,100\sim6{,}100 K, and is consistent with zero at higher temperatures. The resampling analysis yields a qualitatively consistent but less pronounced trend. We interpret these results and discuss limitations of this type of analysis in § 7, and summarize our main findings in § 8.

Summary of the Symbols

Table 1 summarizes the variables and conditions (selection criteria) discussed in this paper. For conditions, the notation AB\mathcal{A}\mathcal{B} denotes the logical conjunction of A\mathcal{A} and B\mathcal{B}. The photometric modulation amplitude, rotation period, and stellar radius are expressed as base-10 logarithms.

SymbolMeaning
R\mathcal{R}KIC star with rotational modulation detected in T. Mazeh et al. [44] or in A. McQuillan et al. [45]
T\mathcal{T}star with a transiting planet
D\mathcal{D}star with a detected transiting planet (always conditioned on T\mathcal{T})
KOI=DTR\mathrm{KOI}=\mathcal{D}\mathcal{T}\mathcal{R}KOI with detected rotational modulation
r=log⁡Rr=\log Rlogarithmic amplitude of rotational modulation (ppm)
TeffT_{\mathrm{eff}}effective temperature of a star (K)
mKepm_{\mathrm{Kep}}Kepler magnitude of a star
log⁡Prot\log P_{\mathrm{rot}}logarithmic rotation period of a star (day)
log⁡R⋆\log R_{\star}logarithmic radius of a star (R⊙R_{\odot})

Table 1. Summary of the Symbols

Why Matching in Effective Temperature Is Not Enough

M15 compared the photometric modulation amplitudes rr of KOIs with those measured for KIC stars by A. McQuillan et al. [45] using the same method, to investigate the correlation between stellar spins and planetary orbits. Because transiting planets have nearly edge-on orbits, such a correlation would produce different stellar inclination distributions for KOIs and KIC stars, potentially leading to different amplitude distributions. M15 performed this comparison as a function of effective temperature TeffT_{\rm eff}. This is a natural comparison given the temperature dependence of stellar obliquities observed in hot-Jupiter systems. Comparing stars at fixed TeffT_{\rm eff} also helps control for differences in the temperature distributions of the two samples, which can affect their amplitude distributions because rr is correlated with TeffT_{\rm eff}.

From this latter perspective, however, fixing TeffT_{\rm eff} alone is not sufficient. Stars have other properties, some of which are correlated with rr. In the comparison performed by M15 at fixed TeffT_{\rm eff}, these properties are marginalized over. If such variables have different distributions among KIC and KOI stars, the amplitude distributions will differ because of these variables, not because of inclinations. For this to occur, the correlation between a “confounder” and the amplitude need not be causal or intrinsic; a statistical correlation introduced by, e.g., selection criteria is sufficient. M15 did examine one such effect: because rotation period is correlated with modulation amplitude, they tried controlling for the rotation period log⁡Prot\log P_{\rm rot} within relatively broad (∼500\sim500 K) TeffT_{\rm eff} bins, and found that this did not substantially change their result.

One important variable that was not controlled in M15 is the apparent Kepler magnitude mKepm_{\rm Kep}. Although it is not obvious whether photometric amplitudes should intrinsically depend on mKepm_{\rm Kep}, it certainly does when we consider stars with detected rotational modulation in A. McQuillan et al. [45] or in M15, as the weakest modulations are missed among the faintest stars (K. Masuda [42]). As shown in Figure 2, the mKepm_{\rm Kep} distributions in their KIC and KOI samples differ, especially for hotter stars, where KOIs are on average brighter. Thus, weaker modulation can be detected among the brighter KOIs than among the fainter KIC stars, and the rr distribution of KOIs is biased toward smaller values when marginalized over mKepm_{\rm Kep}. This aligns with the sign of the KOI–KIC amplitude difference reported by M15 for hot stars. See also § 4 for the differences in the other variables.

Kepler magnitude distributions of KIC and KOI stars in the M15 sample

Figure 2. The Kepler magnitude distributions of KIC (black solid) and KOI (orange dotted) stars in the M15 sample.

Given these considerations, at least two more variables are worth controlling for. As mentioned above, log⁡Prot\log P_{\rm rot} remains important given its correlation with rr. Its distribution may differ between KICs and KOIs, as log⁡Prot\log P_{\rm rot} is related to stellar activity and hence the noise level in the Kepler light curve that affects planet detectability. Another good candidate is the stellar radius log⁡R⋆\log R_\star, which was also not considered in M15. For stars somewhat more massive than the Sun, the radius at a given TeffT_{\rm eff} reflects the evolutionary phase and hence the convective-zone depth, which is related to rr at a given log⁡Prot\log P_{\rm rot} through the associated convective turnover time and hence the Rossby number (e.g., R. W. Noyes et al. [50]; V. See et al. [60]; E. Corsaro et al. [13]). The log⁡R⋆\log R_\star distributions of KOIs and KIC stars can differ for multiple reasons, given that log⁡R⋆\log R_\star affects both the geometric transit probability and the transit depth.

Thus, in this work, we control for z=(mKep,log⁡R⋆,log⁡Prot)z = (m_{\rm Kep}, \log R_\star, \log P_{\rm rot}) in addition to TeffT_{\rm eff}. We do so jointly because their correlations with one another and with rr may also matter. It is interesting to note that the distributions of mKepm_{\rm Kep} and log⁡R⋆\log R_\star are broader for stars hotter than the Sun, which span different evolutionary phases and hence a range of sizes and intrinsic brightnesses, so the impact of this correction is likely to be larger for them. In addition to these confounding issues, we need to account for the fact that the observed KOI sample is shaped by planet detectability, as was also discussed by T. Mazeh et al. [44], and that both samples are conditioned on the detection of rotational modulation. The next section gives a unified mathematical formulation of these effects.

Formal Treatment of Confounding Variables and Selection Effects

We now formalize the discussion in § 2 in probabilistic terms, starting with the comparison performed by M15. We then examine how confounding variables and selection effects enter this comparison, providing the basis for the analyses in § 4 and § 5.

Revisiting the 2015ApJ...801....3M Comparison

T. Mazeh et al. [44] compared the photometric modulation amplitudes of KIC and KOI stars as a function of TeffT_{\rm eff} to investigate stellar spin–orbit alignment. Their main comparison used 250 K temperature bins, with the median value of rr evaluated separately for each sample in each bin.

The binning can be viewed as a way to examine the conditional amplitude distribution p(r∣Teff)p(r|T_{\rm eff}) using samples drawn from the joint distribution p(r,Teff)p(r,T_{\rm eff}), with this interpretation becoming exact in the narrow-bin limit.

When we select stars in a bin centered on Teff=τT_{\rm eff}=\tau, we assign weights w(Teff−τ)w(T_{\rm eff}-\tau) to the observed pairs (ri,Teff,i)(r_i,T_{{\rm eff},i}) in each sample, with a constant weight inside the bin and zero outside it. The resulting amplitude distribution of the selected stars is given by weighting the joint distribution and normalizing over rr:

∫p(r,Teff)w(Teff−τ) dTeff∫p(Teff)w(Teff−τ) dTeff\frac{\int p(r,T_{\rm eff})w(T_{\rm eff}-\tau)\,\mathrm{d}T_{\rm eff}}{\int p(T_{\rm eff})w(T_{\rm eff}-\tau)\,\mathrm{d}T_{\rm eff}}

In the narrow-bin limit, the normalized weight function approaches δ(Teff−τ)\delta(T_{\rm eff}-\tau), so this reduces to

p(r,Teff)p(Teff)∣Teff=τ=p(r∣Teff)p(Teff)p(Teff)∣Teff=τ=p(r∣τ).\left.\frac{p(r,T_{\rm eff})}{p(T_{\rm eff})}\right|_{T_{\rm eff}=\tau} = \left.\frac{p(r|T_{\rm eff})p(T_{\rm eff})}{p(T_{\rm eff})}\right|_{T_{\rm eff}=\tau} = p(r|\tau).

In practice, M15 used the median to summarize this conditional amplitude distribution.

Applying this procedure separately to the KIC and KOI samples, we can therefore describe the comparison of M15 as a comparison of (the medians of)

p(r∣Teff,X)andp(r∣Teff,XT),p(r|T_{\rm eff},\mathcal{X}) \qquad\text{and} \qquad p(r|T_{\rm eff},\mathcal{X}\mathcal{T}),

where X\mathcal{X} denotes the common selection criteria and T\mathcal{T} denotes the presence of a transiting planet. This comparison examines how the amplitude distribution differs with T\mathcal{T} at a given TeffT_{\rm eff}. For now, this description omits the additional detection requirements of the observed samples; we return to them below.

As discussed in § 2, examining this difference as a function of TeffT_{\rm eff} is not only physically motivated but also helps distinguish a difference associated with T\mathcal{T} from one caused by different temperature distributions. To illustrate this second point, consider what would happen if the two samples were compared without conditioning on TeffT_{\rm eff}. The corresponding marginal distributions are

p(r∣X)=∫p(r,Teff∣X) dTeff=∫p(r∣Teff,X)p(Teff∣X) dTeff,p(r|\mathcal{X})=\int p(r,T_{\rm eff}|\mathcal{X})\,\mathrm{d}T_{\rm eff} =\int p(r|T_{\rm eff},\mathcal{X})p(T_{\rm eff}|\mathcal{X})\,\mathrm{d}T_{\rm eff},
p(r∣XT)=∫p(r,Teff∣XT) dTeff=∫p(r∣Teff,XT)p(Teff∣XT) dTeff.p(r|\mathcal{X}\mathcal{T})=\int p(r,T_{\rm eff}|\mathcal{X}\mathcal{T})\,\mathrm{d}T_{\rm eff} =\int p(r|T_{\rm eff},\mathcal{X}\mathcal{T})p(T_{\rm eff}|\mathcal{X}\mathcal{T})\,\mathrm{d}T_{\rm eff}.

These expressions average the conditional amplitude distributions, p(r∣Teff,X)p(r|T_{\rm eff},\mathcal{X}) and p(r∣Teff,XT)p(r|T_{\rm eff},\mathcal{X}\mathcal{T}), over the temperature distribution of each sample, p(Teff∣X)p(T_{\rm eff}|\mathcal{X}) and p(Teff∣XT)p(T_{\rm eff}|\mathcal{X}\mathcal{T}). If rr depends on TeffT_{\rm eff}, differences between p(Teff∣X)p(T_{\rm eff}|\mathcal{X}) and p(Teff∣XT)p(T_{\rm eff}|\mathcal{X}\mathcal{T}) can produce different marginal amplitude distributions even when the conditional distributions at each TeffT_{\rm eff} are identical. Comparing the samples at fixed TeffT_{\rm eff} avoids this source of difference.

In summary, M15 used binning to compare rr distributions conditioned on TeffT_{\rm eff} and interpreted the KOI–KIC difference as an effect of conditioning on T\mathcal{T}. Below, we clarify two limitations of this comparison: conditioning only on TeffT_{\rm eff} and not explicitly accounting for the detection requirements of the observed samples.

Importance of Controlling Other Variables

We first consider how other stellar properties can affect the comparison, as discussed in § 2. Each star has properties other than rr and TeffT_{\mathrm{eff}}, which we collectively denote as zz. When comparing p(r∣Teff)p(r|T_{\mathrm{eff}}), these variables are implicitly marginalized over their joint distribution specific to each sample:

p(r∣Teff,X)=∫p(r,z∣Teff,X) dz=∫p(r∣z,Teff,X)p(z∣Teff,X) dz,\begin{aligned} p(r|T_{\mathrm{eff}},\mathcal{X}) = \int p(r,z|T_{\mathrm{eff}},\mathcal{X})\,dz \\ &= \int p(r|z,T_{\mathrm{eff}},\mathcal{X})p(z|T_{\mathrm{eff}},\mathcal{X})\,dz, \end{aligned}
p(r∣Teff,XT)=∫p(r∣z,Teff,XT)p(z∣Teff,XT) dz.p(r|T_{\mathrm{eff}},\mathcal{X}^{T}) = \int p(r|z,T_{\mathrm{eff}},\mathcal{X}^{T})p(z|T_{\mathrm{eff}},\mathcal{X}^{T})\,dz.

This is the same issue illustrated above for TeffT_{\mathrm{eff}}: if rr depends on zz when conditioned on TeffT_{\mathrm{eff}} and X\mathcal{X} or XT\mathcal{X}^{T}, the difference between p(z∣Teff,X)p(z|T_{\mathrm{eff}},\mathcal{X}) and p(z∣Teff,XT)p(z|T_{\mathrm{eff}},\mathcal{X}^{T}) results in a difference between p(r∣Teff,X)p(r|T_{\mathrm{eff}},\mathcal{X}) and p(r∣Teff,XT)p(r|T_{\mathrm{eff}},\mathcal{X}^{T}). We should therefore compare p(r∣z,Teff,X)p(r|z,T_{\mathrm{eff}},\mathcal{X}) and p(r∣z,Teff,XT)p(r|z,T_{\mathrm{eff}},\mathcal{X}^{T}), and any such stellar parameter needs to be included in zz on the right side of the bar. For the present analysis, we adopt z=(mKep,log⁡R⋆,log⁡Prot)z=(m_{\mathrm{Kep}},\log R_{\star},\log P_{\mathrm{rot}}), as discussed in § 2, and compare the amplitude distributions at fixed (z,Teff)(z,T_{\mathrm{eff}}) using the resampling and likelihood-based approaches described below.

Impacts of Conditioning on Detected Transiting Planets and Rotational Modulation

Another important consideration is that both the KIC and KOI samples require detected rotational modulation, while the KOI sample additionally requires a detected transiting planet. We therefore need to account for how these detection requirements affect the comparison. Writing the rotational-modulation and planet-detection conditions as R\mathcal{R} and D\mathcal{D}, respectively, the observed comparison is between XR\mathcal{X}\mathcal{R} and XTRD\mathcal{X}^{T}\mathcal{R}\mathcal{D}, whereas the comparison of interest is between X\mathcal{X} and XT\mathcal{X}^{T}. For clarity, we suppress X\mathcal{X} from the notation below, with all distributions understood to be conditioned on it.

We first consider the impact of D\mathcal{D}. Since we have

p(r,z,Teff∣DTR)=p(D∣r,z,Teff,TR)p(r,z,Teff∣TR)p(D∣TR)=πD∣TR(r,z,Teff)p(r,z,Teff∣TR)p(D∣TR),(1)\begin{aligned} p(r,z,T_{\mathrm{eff}}|\mathcal{D}\mathcal{T}\mathcal{R}) = \frac{p(\mathcal{D}|r,z,T_{\mathrm{eff}},\mathcal{T}\mathcal{R})p(r,z,T_{\mathrm{eff}}|\mathcal{T}\mathcal{R})}{p(\mathcal{D}|\mathcal{T}\mathcal{R})} \\ &= \frac{\pi_{\mathcal{D}|\mathcal{T}\mathcal{R}}(r,z,T_{\mathrm{eff}})p(r,z,T_{\mathrm{eff}}|\mathcal{T}\mathcal{R})}{p(\mathcal{D}|\mathcal{T}\mathcal{R})}, \tag*{(1)} \end{aligned}

one can “correct for” the D\mathcal{D} effect to obtain p(r,z,Teff∣TR)p(r,z,T_{\mathrm{eff}}|\mathcal{T}\mathcal{R}) essentially by reweighting the samples from DTR\mathcal{D}\mathcal{T}\mathcal{R} by the inverse2 of the detectability of a transiting planet πD∣TR(r,z,Teff)≡p(D∣r,z,Teff,TR)\pi_{\mathcal{D}|\mathcal{T}\mathcal{R}}(r,z,T_{\mathrm{eff}}) \equiv p(\mathcal{D}|r,z,T_{\mathrm{eff}},\mathcal{T}\mathcal{R}) defined in the last equation. We assume that this term can be estimated from the available information about the Kepler planet-detection pipeline, and discuss the implementation of this correction in § 4.2.

Such a correction essentially allows us to compare the TR\mathcal{T}\mathcal{R} sample against the R\mathcal{R} sample. As we discussed above, we wish to perform this comparison in a manner conditioned on (z,Teff)(z,T_{\mathrm{eff}}):

p(r∣z,Teff,TR)p(r∣z,Teff,R)=p(R∣r,z,Teff,T)p(r∣z,Teff,T)p(R∣z,Teff,T)(p(R∣r,z,Teff)p(r∣z,Teff)p(R∣z,Teff))−1\begin{aligned} \frac{p(r|z,T_{\mathrm{eff}},\mathcal{T}\mathcal{R})}{p(r|z,T_{\mathrm{eff}},\mathcal{R})} \\ &= \frac{p(\mathcal{R}|r,z,T_{\mathrm{eff}},\mathcal{T})p(r|z,T_{\mathrm{eff}},\mathcal{T})}{p(\mathcal{R}|z,T_{\mathrm{eff}},\mathcal{T})} \left(\frac{p(\mathcal{R}|r,z,T_{\mathrm{eff}})p(r|z,T_{\mathrm{eff}})}{p(\mathcal{R}|z,T_{\mathrm{eff}})}\right)^{-1} \end{aligned}
∝p(R∣r,z,Teff,T)p(R∣r,z,Teff)p(r∣z,Teff,T)p(r∣z,Teff)\propto\frac{p(\mathcal{R}|r,z,T_{\mathrm{eff}},\mathcal{T})}{p(\mathcal{R}|r,z,T_{\mathrm{eff}})} \frac{p(r|z,T_{\mathrm{eff}},\mathcal{T})}{p(r|z,T_{\mathrm{eff}})}

focusing only on the rr dependence. If we assume that the detectability of rotational modulation R\mathcal{R} is not affected by the presence of transiting planets at fixed (r,z,Teff)(r,z,T_{\mathrm{eff}}), i.e., p(R∣r,z,Teff,T)=p(R∣r,z,Teff)p(\mathcal{R}|r,z,T_{\mathrm{eff}},\mathcal{T})=p(\mathcal{R}|r,z,T_{\mathrm{eff}}), then the first factor in the above equation cancels. The ratio of the two amplitude distributions conditioned on R\mathcal{R} therefore has the same rr dependence as the corresponding ratio without conditioning on R\mathcal{R}. In particular, conditioning on R\mathcal{R} does not by itself generate an amplitude difference under the null hypothesis p(r∣z,Teff,T)=p(r∣z,Teff)p(r|z,T_{\mathrm{eff}},\mathcal{T})=p(r|z,T_{\mathrm{eff}}), which forms the basis of the comparison below. We assume that this is the case in the samples of M15 and A. McQuillan et al. [45], given that the same pipeline was applied to the KOI sample after masking transits and that R\mathcal{R} is conditioned not only on rr but also on the main relevant variables in zz, particularly apparent magnitude (see K. Masuda [42]).3

We reiterate that these arguments rely on (z,Teff)(z,T_{\mathrm{eff}}) adequately accounting for the relevant confounding variables, as discussed in § 3.2. Any unaccounted-for variable that affects the amplitude distribution and is distributed differently between stars with and without transiting planets could still bias the comparison.

Detectability-Weighted Resampling

The considerations in § 3 show that the effects of the identified confounding variables and planet detectability can be accounted for by conditioning the comparison on (TeffT_{\mathrm{eff}}, zz) and subjecting the KIC sample to the same planet-detection selection as the KOI sample through detectability weighting. Here we implement this comparison by resampling the detectability-weighted KIC sample.4 This procedure extends the weighted-sample approach of M15, which conditioned the comparison on TeffT_{\mathrm{eff}} alone, to incorporate planet detectability and the additional conditioning variables zz.

Principle of the Approach

Given the above considerations, we propose to compare the following two samples to examine the difference between the rr distributions conditioned on zz and TeffT_{\mathrm{eff}} (at fixed zz and TeffT_{\mathrm{eff}}):

  1. The KOI sample (DTR\mathcal{D}\mathcal{T}\mathcal{R}) weighted by the narrow w(Teff−τ)w(T_{\mathrm{eff}}-\tau) around each τ\tau and marginalized over the other variables; this is essentially the same as discussed in M15.

  1. The KIC sample (R\mathcal{R}) first weighted by the narrow w(Teff−τ)w(T_{\mathrm{eff}}-\tau) and the KOI detectability πD∣TR\pi_{\mathrm{D}|\mathrm{TR}}, and then resampled so that the zz distribution of the detectability-weighted sample at each given τ\tau aligns with that of the KOI sample above.

The latter sample is designed to subject the KIC sample to the same selection on D\mathcal{D} while controlling for variables that could affect the rr distribution.

The following equations show how these samples meet our requirements. First, the KOI sample defined above in item 1 follows

p1(r∣τ)∝∫p(r,z,Teff∣DTR)δ(Teff−τ) dTeffdz∝∫p(r,z∣τ,DTR) dz=∫p(r∣z,τ,DTR)p(z∣τ,DTR) dz=∫πD∣TR(r,z,τ)p(r∣z,τ,TR)CKOI(z,τ)p(z∣τ,DTR) dz,(2)\begin{aligned} p_1(r|\tau) &\propto\int p(r,z,T_{\mathrm{eff}}|\mathcal{D}\mathcal{T}\mathcal{R})\delta(T_{\mathrm{eff}}-\tau)\,\mathrm{d}T_{\mathrm{eff}}\mathrm{d}z \\ &\propto\int p(r,z|\tau,\mathcal{D}\mathcal{T}\mathcal{R})\,\mathrm{d}z \\ &= \int p(r|z,\tau,\mathcal{D}\mathcal{T}\mathcal{R})p(z|\tau,\mathcal{D}\mathcal{T}\mathcal{R})\,\mathrm{d}z \\ &= \int\frac{\pi_{\mathrm{D}|\mathrm{TR}}(r,z,\tau)p(r|z,\tau,\mathcal{T}\mathcal{R})}{C_{\mathrm{KOI}}(z,\tau)}p(z|\tau,\mathcal{D}\mathcal{T}\mathcal{R})\,\mathrm{d}z, \tag*{(2)} \end{aligned}
CKOI(z,τ)≡∫πD∣TR(r,z,τ)p(r∣z,τ,TR) dr.C_{\mathrm{KOI}}(z,\tau)\equiv\int\pi_{\mathrm{D}|\mathrm{TR}}(r,z,\tau)p(r|z,\tau,\mathcal{T}\mathcal{R})\,\mathrm{d}r.

On the other hand, the detectability-weighted KIC samples in item 2 follow pKIC∗(r,z,Teff)∝πD∣TR(r,z,Teff)p(r,z,Teff∣R)p_{\mathrm{KIC}}^*(r,z,T_{\mathrm{eff}})\propto\pi_{\mathrm{D}|\mathrm{TR}}(r,z,T_{\mathrm{eff}})p(r,z,T_{\mathrm{eff}}|\mathcal{R}), which gives

pKIC∗(r∣z,Teff)=πD∣TR(r,z,Teff)p(r∣z,Teff,R)CKIC(z,Teff),p_{\mathrm{KIC}}^*(r|z,T_{\mathrm{eff}})=\frac{\pi_{\mathrm{D}|\mathrm{TR}}(r,z,T_{\mathrm{eff}})p(r|z,T_{\mathrm{eff}},\mathcal{R})}{C_{\mathrm{KIC}}(z,T_{\mathrm{eff}})},
CKIC(z,Teff)≡∫πD∣TR(r,z,Teff)p(r∣z,Teff,R) drC_{\mathrm{KIC}}(z,T_{\mathrm{eff}})\equiv\int\pi_{\mathrm{D}|\mathrm{TR}}(r,z,T_{\mathrm{eff}})p(r|z,T_{\mathrm{eff}},\mathcal{R})\,\mathrm{d}r

at fixed (zz, TeffT_{\mathrm{eff}}). The zz matching is then applied to this weighted distribution to obtain the reweighted and resampled KIC distribution as:

p2(r∣τ)∝∫pKIC∗(r,z,Teff)δ(Teff−τ)p(z∣Teff,DTR)pKIC∗(z∣Teff) dTeffdz∝∫pKIC∗(r,z∣τ)p(z∣τ,DTR)pKIC∗(z∣τ) dz=∫pKIC∗(r∣z,τ)p(z∣τ,DTR) dz=∫πD∣TR(r,z,τ)p(r∣z,τ,R)CKIC(z,τ)p(z∣τ,DTR) dz.(3)\begin{aligned} p_2(r|\tau)&\propto\int p_{\mathrm{KIC}}^*(r,z,T_{\mathrm{eff}})\delta(T_{\mathrm{eff}}-\tau)\frac{p(z|T_{\mathrm{eff}},\mathcal{D}\mathcal{T}\mathcal{R})}{p_{\mathrm{KIC}}^*(z|T_{\mathrm{eff}})}\,\mathrm{d}T_{\mathrm{eff}}\mathrm{d}z\\ &\propto\int p_{\mathrm{KIC}}^*(r,z|\tau)\frac{p(z|\tau,\mathcal{D}\mathcal{T}\mathcal{R})}{p_{\mathrm{KIC}}^*(z|\tau)}\,\mathrm{d}z\\ &=\int p_{\mathrm{KIC}}^*(r|z,\tau)p(z|\tau,\mathcal{D}\mathcal{T}\mathcal{R})\,\mathrm{d}z\\ &=\int\frac{\pi_{\mathrm{D}|\mathrm{TR}}(r,z,\tau)p(r|z,\tau,\mathcal{R})}{C_{\mathrm{KIC}}(z,\tau)}p(z|\tau,\mathcal{D}\mathcal{T}\mathcal{R})\,\mathrm{d}z. \tag*{(3)} \end{aligned}

Equations 2 and 3 show that the two distributions differ only through p(r∣z,τ,TR)p(r|z,\tau,\mathcal{T}\mathcal{R}) and p(r∣z,τ,R)p(r|z,\tau,\mathcal{R}), including their corresponding selection normalizations CC: under the null hypothesis that these conditional distributions are equal, CKIC(z,τ)=CKOI(z,τ)C_{\mathrm{KIC}}(z,\tau)=C_{\mathrm{KOI}}(z,\tau), and hence p1(r∣τ)=p2(r∣τ)p_1(r|\tau)=p_2(r|\tau). Thus, subject to the assumptions discussed above, a difference between p1p_1 and p2p_2 can be attributed to the condition T\mathcal{T}.

Implementation of the Resampling

We start with the same KIC and KOI samples analyzed by M15, for which TeffT_{\mathrm{eff}}, log⁡Prot\log P_{\mathrm{rot}}, and mKepm_{\mathrm{Kep}} are available. For the stellar radii, we adopted the values from T. A. Berger et al. [6]. Cross-matching with this catalog and requiring CDPP measurements for the KIC stars (see below) resulted in final samples of 29700 KIC stars and 880 KOIs.

We require an estimate of πD∣TR\pi_{\mathrm{D}|\mathrm{TR}}, the detectability of a transiting planet conditional on transit and marginalized over planet parameters. We estimated this for the KIC stars with detected rotational modulation using a synthetic physical catalog generated within the SysSim framework (M. Y. He et al. 2019, 2020). We selected planets satisfying ∣b∣<1|b|<1 and drew a set of 1,000 period-radius pairs from this transiting population. For each KIC star and simulated planet, we evaluated the expected transit signal-to-noise ratio as

S/N=106(RpR⋆)2(TobsfdutyP)1/21CDPP(Tdur),\mathrm{S/N}=10^6\left(\frac{R_{\mathrm{p}}}{R_\star}\right)^2\left(\frac{T_{\mathrm{obs}}f_{\mathrm{duty}}}{P}\right)^{1/2}\frac{1}{\mathrm{CDPP}(T_{\mathrm{dur}})},
Tdur=13 hr(P365.25 day)1/3(R⋆R⊙)(M⋆M⊙)−1/3π4.T_{\mathrm{dur}}=13\ \mathrm{hr}\left(\frac{P}{365.25\ \mathrm{day}}\right)^{1/3}\left(\frac{R_\star}{R_\odot}\right)\left(\frac{M_\star}{M_\odot}\right)^{-1/3}\frac{\pi}{4}.

The factor π/4\pi/4 averages the chord-length factor 1−b2\sqrt{1-b^2} over the uniform impact-parameter distribution conditional on transit (∣b∣<1|b| < 1). Here TobsT_{\mathrm{obs}} and fdutyf_{\mathrm{duty}} are the data span and duty cycle. We evaluated the CDPP at TdurT_{\mathrm{dur}} by interpolating the tabulated values in log–log space; durations outside the tabulated range were assigned the value at the nearest endpoint. We then converted S/N into a detection probability using the gamma-function completeness curve adopted in B. J. Fulton et al. [23], based on the Kepler injection–recovery calibration of J. L. Christiansen et al. [12]. The value of πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} for each star was obtained by averaging this probability over the 1,000 simulated planets. Note that this is the detectability conditioned on T\mathcal{T}: that is, we evaluate the probability that a planet would be detected given that it transits. The dependence on the geometric transit probability itself is treated separately by controlling for the stellar radius.

We then repeated the following procedure 200 times:

  1. Sample 29700 stars with replacement from KIC, with probabilities proportional to the value of πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} assigned to each star.

  1. In each 100 K bin over 3500 K≤Teff≤6500 K3500\ \mathrm{K} \le T_{\mathrm{eff}} \le6500\ \mathrm{K}, match the z=(mKep,log⁡R⋆,log⁡Prot)\boldsymbol{z} = (m_{\mathrm{Kep}}, \log R_{\star}, \log P_{\mathrm{rot}}) distribution to that of KOI. For each KOI, we identify the k=32k = 32 nearest stars in the sample from step 1 using the scaled distance

d2=(ΔmKep0.1)2+(Δlog⁡R⋆0.05)2+(Δlog⁡Prot0.05)2,d^2 = \left(\frac{\Delta m_{\mathrm{Kep}}}{0.1}\right)^2 + \left(\frac{\Delta\log R_{\star}}{0.05}\right)^2 + \left(\frac{\Delta\log P_{\mathrm{rot}}}{0.05}\right)^2,

and draw one of them with probability proportional to exp⁡(−d2/2)\exp(-d^2/2). This sampling is performed with replacement and yields the same number of stars as in the KOI sample in each temperature bin.

  1. Compute the median rr in each 150 K bin of TeffT_{\mathrm{eff}} in the sample from step 2.

This yields 200 realizations of the rr–TeffT_{\mathrm{eff}} relation corresponding to p2p_2 in Equation (3).

Results

Figure 3 shows the median and 16th–84th percentile interval (KIC resamples; blue circles), together with the KOI sample (orange squares). We find that the resampling lowers the control-sample rr above ∼5500 K\sim5500\ \mathrm{K} (blue vs gray), where the distributions of z\boldsymbol{z}, the transit probability, and the transit detectability vary more strongly than for cooler stars. As a result, the KOI sample no longer lies clearly below the KIC sample; on average, it is higher up to ∼6000\sim6000 K. For hotter stars, the two distributions become indistinguishable within the uncertainties. Figure 4 shows that the resampling approximately matches the relevant stellar-parameter distributions. The marginal and joint distributions of TeffT_{\mathrm{eff}}, mKepm_{\mathrm{Kep}}, log⁡Prot\log P_{\mathrm{rot}}, and log⁡R⋆\log R_{\star} in the resampled KIC sample closely follow those of the KOI sample, whereas the original KIC sample shows noticeable differences, particularly in the pairwise distributions involving mKepm_{\mathrm{Kep}} and log⁡R⋆\log R_{\star}. Figure 5 compares the empirical cumulative dis-

Median photometric modulation amplitude as a function of effective temperature

Figure 3. Median photometric modulation amplitude as a function of effective temperature for the original KIC sample (gray circles), the KIC resamples (blue circles), and the KOI sample (orange squares). Small gray and orange points show the individual KIC and KOI stars, respectively. The blue shaded region shows the 16th–84th percentile interval among the 200 resampling realizations. Medians are evaluated in 150 K bins.

Marginal and pairwise joint distributions of $T_{\mathrm{eff}}$, $r$, $m_{\mathrm{Kep}}$, $\log P_{\mathrm{rot}}$, and $\log R_{\star}$

Figure 4. Marginal and pairwise joint distributions of TeffT_{\mathrm{eff}}, rr, mKepm_{\mathrm{Kep}}, log⁡Prot\log P_{\mathrm{rot}}, and log⁡R⋆\log R_{\star} for the original KIC sample (black), the resampled KIC sample (blue), and the KOI sample (orange). The resampled distribution is represented by 50,000 stars randomly selected from the pooled 200 realizations. The resampling matches TeffT_{\mathrm{eff}}, mKepm_{\mathrm{Kep}}, log⁡Prot\log P_{\mathrm{rot}}, and log⁡R⋆\log R_{\star} to the KOI sample; rr is shown for comparison but is not used in the matching. The contours enclose approximately 12%, 39%, 68%, and 86% of each distribution, corresponding to the default 0.5, 1, 1.5, and 2σ2\sigma levels for a two-dimensional Gaussian distribution.

Empirical cumulative distributions of $r$ in 500 K bins

Figure 5. Empirical cumulative distributions of rr in 500 K bins for the original KIC sample (black solid lines), the KIC sample pooled over the 200 resampling realizations (blue dashed lines), and the KOI sample (orange dotted lines).

tributions of rr in 500 K bins. Below 6000 K, the KOI distributions are generally shifted toward larger amplitudes relative to the resampled KIC distributions. In the 6000–6500 K bin, the original KIC distribution is shifted toward larger amplitudes than the KOI distribution, while the resampled KIC and KOI distributions nearly overlap.

Likelihood-based Approach

In § 4, we compared p(r∣z,Teff,TR)p(r\mid z,T_{\mathrm{eff}},T\mathcal{R}) and p(r∣z,Teff,R)p(r\mid z,T_{\mathrm{eff}},\mathcal{R}) after marginalizing both over the same zz distribution. Although this controls for differences in the distribution of zz, binning in effective temperature and summarizing the resulting marginalized distributions by their medians may reduce sensitivity to differences in the full conditional amplitude distributions. We therefore complement the resampling analysis with a likelihood-based comparison at each star’s observed y=(z,Teff)y=(z,T_{\mathrm{eff}}), evaluating the required conditional likelihood through a model of the joint distribution of (r,y)(r,y).

Conditional Formulation Using a Joint Reference Density

We test whether the distribution of rKOIr_{\mathrm{KOI}} at fixed yy is consistent with that of the KIC sample after accounting for planet detectability. The null hypothesis is

p(r∣y,TR)=p(r∣y,R)≡pKIC(r∣y),p(r\mid y,\mathcal{T}\mathcal{R})=p(r\mid y,\mathcal{R})\equiv p_{\mathrm{KIC}}(r\mid y),

that is, presence of a transiting planet does not affect the rr distribution at fixed yy. Using Equation (1), the conditional amplitude distribution expected for KOIs under this null hypothesis is

p(r∣y,DTR)=πD∣TR(r,y)p(r∣y,TR)∫πD∣TR(r′,y)p(r′∣y,TR) dr′=πD∣TR(r,y)pKIC(r∣y)∫πD∣TR(r′,y)pKIC(r′∣y) dr′≡pKIC∗(r∣y).\begin{aligned} p(r\mid y,\mathcal{D}\mathcal{T}\mathcal{R}) &= \frac{\pi_{\mathcal{D}\mid\mathcal{T}\mathcal{R}}(r,y)p(r\mid y,\mathcal{T}\mathcal{R})} {\int\pi_{\mathcal{D}\mid\mathcal{T}\mathcal{R}}(r',y)p(r'\mid y,\mathcal{T}\mathcal{R})\,\mathrm{d}r'} \\ &= \frac{\pi_{\mathcal{D}\mid\mathcal{T}\mathcal{R}}(r,y)p_{\mathrm{KIC}}(r\mid y)} {\int\pi_{\mathcal{D}\mid\mathcal{T}\mathcal{R}}(r',y)p_{\mathrm{KIC}}(r'\mid y)\,\mathrm{d}r'} \\ &\equiv p_{\mathrm{KIC}}^{*}(r\mid y). \end{aligned}

To quantify the departure of the KOI amplitude distribution from this conditional null, we introduce a KOI-minus-KIC shift α\alpha in logarithmic amplitude and assume

p(r∣y,DTR)=pKIC∗(r−α∣y).(4)p(r\mid y,\mathcal{D}\mathcal{T}\mathcal{R})=p_{\mathrm{KIC}}^{*}(r-\alpha\mid y). \tag*{(4)}

Given the observed pairs (rKOI,i,yKOI,i)(r_{\mathrm{KOI},i},y_{\mathrm{KOI},i}) and treating yKOI,iy_{\mathrm{KOI},i} as fixed, the likelihood function for α\alpha is then

L(α)=∏ipKIC∗(rKOI,i−α∣yKOI,i)=∏ipKIC∗(rKOI,i−α,yKOI,i)pKIC∗(yKOI,i)∝∏ipKIC∗(rKOI,i−α,yKOI,i).(5)\begin{aligned} \mathcal{L}(\alpha) &= \prod_i p_{\mathrm{KIC}}^{*}(r_{\mathrm{KOI},i}-\alpha\mid y_{\mathrm{KOI},i}) \\ &= \prod_i \frac{p_{\mathrm{KIC}}^{*}(r_{\mathrm{KOI},i}-\alpha,y_{\mathrm{KOI},i})}{p_{\mathrm{KIC}}^{*}(y_{\mathrm{KOI},i})} \\ &\propto\prod_i p_{\mathrm{KIC}}^{*}(r_{\mathrm{KOI},i}-\alpha,y_{\mathrm{KOI},i}). \tag*{(5)} \end{aligned}

where

pKIC∗(r,y)≡πD∣TR(r,y)pKIC(r,y)∫πD∣TR(r′,y′)pKIC(r′,y′) dr′dy′.(6)p_{\mathrm{KIC}}^{*}(r,y)\equiv \frac{\pi_{\mathcal{D}\mid\mathcal{T}\mathcal{R}}(r,y)p_{\mathrm{KIC}}(r,y)} {\int\pi_{\mathcal{D}\mid\mathcal{T}\mathcal{R}}(r',y')p_{\mathrm{KIC}}(r',y')\,\mathrm{d}r'\mathrm{d}y'}. \tag*{(6)}

The last proportionality in Equation (5) holds because pKIC∗(yKOI,i)p_{\mathrm{KIC}}^{*}(y_{\mathrm{KOI},i}) is independent of α\alpha. Thus the inference requires the values of the joint reference density only up to an α\alpha-independent factor.

We have samples from pKIC(r,y)p_{\mathrm{KIC}}(r,y), rather than its probability density at arbitrary (r,y)(r,y). In principle, one could construct a multidimensional histogram and evaluate its height to estimate the density. In practice, this becomes difficult because yy contains several parameters, so we instead use the normalizing-flow model described below.

Since rr is a logarithmic amplitude, α\alpha describes a multiplicative shift in the amplitude distribution. This location-shift model is a simplification, although the empirical distributions in Figure 5 suggest that it provides a reasonable description. The resulting α\alpha is only a descriptive parameter for a departure from the detectability-matched, post-selection null distribution pKIC∗p^{*}_{\mathrm{KIC}}. A nonzero value implies that the conditional amplitude distributions of the KOI and KIC samples differ after accounting for planet detectability. However, the value of α\alpha does not necessarily equal an intrinsic shift αint\alpha_{\mathrm{int}} in the underlying pre-selection amplitude distribution, for example one defined by p(r∣y,T)=p(r−αint∣y)p(r|y,\mathcal{T}) = p(r-\alpha_{\mathrm{int}}|y). Inferring αint\alpha_{\mathrm{int}} would require both the rotational-modulation detection probability p(R∣r,y)p(\mathcal{R}|r,y), which is not well quantified, and an explicit evaluation of the αint\alpha_{\mathrm{int}}-dependent selection normalization. Such an analysis is beyond the scope of this work.

Normalizing Flow Model for KIC Stars

A normalizing flow represents a multidimensional probability density as an invertible transformation of a simple base distribution, such as a multivariate normal distribution (D. Jimenez Rezende & S. Mohamed 2015; I. Kobyzev et al. 2021). The change-of-variables formula then permits direct evaluation of the probability density at any point. A normalizing flow thus provides a smooth and flexible analogue of a multidimensional histogram. See also J. F. Crenshaw et al. [14] for an astronomical application and accessible introduction to normalizing flows.

For the main analysis, we modeled the five-dimensional vector θ=(Teff,r,log⁡Prot,log⁡R⋆,mKep)\boldsymbol{\theta}=(T_{\mathrm{eff}},r,\log P_{\mathrm{rot}},\log R_{\star},m_{\mathrm{Kep}}). We standardized these variables and constructed a masked autoregressive flow (G. Papamakarios et al. [51]) with 16-knot rational-quadratic spline transforms (C. Durkan et al. [19]) using FlowJAX version 17.2.1 (D. Ward [66]). The model was trained on 29700 KIC stars by maximizing the πD∣TR\pi_{\mathrm{D|TR}}-weighted log likelihood,

ϕ^=arg⁡max⁡ϕ ∑nπD∣TR(θn)log⁡qϕ(θn)∑nπD∣TR(θn)\hat{\phi}=\underset{\phi}{\arg\max}\ \frac{\sum_n \pi_{\mathrm{D|TR}}(\theta_n)\log q_{\phi}(\theta_n)}{\sum_n \pi_{\mathrm{D|TR}}(\theta_n)}

where qϕq_{\phi} denotes the flow density. We used a learning rate of 3×10−43\times10^{-4}, a batch size of 1,024, and at most 500 training epochs, with early stopping after 50 epochs without improvement. The fitted density qϕ^q_{\hat{\phi}} approximates the detectability-weighted joint reference density pKIC⋆(r,y)p_{\mathrm{KIC}}^{\star}(r,y) in Equation (6). It can therefore be evaluated directly at each shifted KOI data point in Equation (5). Figure 6 shows that samples from the fitted flow reproduce the one- and two-dimensional marginal distributions of the weighted KIC sample reasonably well.

One- and two-dimensional marginal distributions of $T_{\mathrm{eff}}$, $r$, $m_{\mathrm{Kep}}$, $\log P_{\mathrm{rot}}$, and $\log R_{\star}$

Figure 6. One- and two-dimensional marginal distributions of (TeffT_{\mathrm{eff}}, rr, mKepm_{\mathrm{Kep}}, log⁡Prot\log P_{\mathrm{rot}}, log⁡R⋆\log R_{\star}) for the KIC sample weighted by πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} (blue) and for 60,000 samples drawn from the fitted normalizing-flow model (orange). The diagonal panels show the marginal distributions, and the off-diagonal panels show the joint contours. The contours enclose approximately 12%, 39%, 68%, and 86% of each distribution, corresponding to the default 0.5, 1, 1.5, and 2σ2\sigma levels for a two-dimensional Gaussian distribution.

With a finite KIC sample, the fitted five-dimensional density need not perfectly represent the underlying reference distribution. As a simple check of the adopted flow model, we randomly excluded 2,914 KIC stars (approximately 10%) and trained the same flow model on the remaining stars using the same procedure. We then tested whether it could predict the amplitude distribution of the excluded stars. For each excluded star, we fixed yy at its observed value and obtained the conditional density p(r∣y)p(r|y) by numerically normalizing the fitted joint density over rr, then integrated this density to obtain the cumulative distribution. Figure 7 compares the predicted and observed cumulative distributions, weighted by πD∣TR\pi_{\mathrm{D|TR}}, for these stars. They agree reasonably well within the six temperature bins shown. While this agreement supports the use of the fitted flow as the reference density for the following analysis, this check does not quantify uncertainty in the estimated reference density or its sensitivity to alternative flow architectures or training procedures. The posterior intervals for α\alpha reported below are therefore conditional on the fitted reference density. This constitutes a limitation of the present likelihood-based analysis.

Cumulative distributions of $r$ for excluded KIC stars

Figure 7. Validation of the flow model using the cumulative distributions of rr for the 2,914 KIC stars excluded from training. Black solid lines show the empirical cumulative distributions, with each star weighted by πD∣TR\pi_{\mathrm{D|TR}}. Gray dashed lines show the conditional cumulative distributions predicted at each star’s observed (Teff,mKep,log⁡R⋆,log⁡ProtT_{\mathrm{eff}},m_{\mathrm{Kep}},\log R_{\star},\log P_{\mathrm{rot}}) and averaged over the stars in the panel using the same weights.

The flow fitted to the full KIC sample also provides a model-based counterpart to the resampling comparison in § 4. We evaluate the conditional distribution pKIC⋆(r∣y)p_{\mathrm{KIC}}^{\star}(r|y) at each KOI’s observed yy without binning in TeffT_{\mathrm{eff}}, then average the corresponding cumulative distributions within the same 500 K bins as Figure 5 for display. The resulting predictions under α=0\alpha= 0, shown in Figure 8, essentially correspond to p2p_2 in Equation (3) but averaged over the KOI temperature distribution within each bin. They agree reasonably well with the resampled KIC distributions, providing a consistency check between the two approaches. As in the resampling comparison, the observed KOI distributions are generally shifted toward larger amplitudes below 6000 K. Below, we quantify departures from this conditional reference model through α(Teff)\alpha(T_{\mathrm{eff}}) using the individual conditional densities.

Cumulative distributions of $r$ in 500 K temperature bins

Figure 8. Cumulative distributions of rr in the same 500 K bins as Figure 5, comparing the KIC sample pooled over 200 resampling realizations (blue dashed lines), the flow predictions under α=0\alpha= 0 (purple solid lines), and the KOI sample (orange dotted lines). The flow predictions average the conditional cumulative distributions evaluated at each KOI’s observed yy, with equal weights. The purple bands show the 16th–84th percentile range at each value of rr among 5,000 simulated KOI samples, generated by independently drawing one amplitude per KOI while holding its yy fixed. These bands describe the variation expected for the observed number of KOIs at fixed flow density and do not include uncertainty in its estimation.

Inference

We infer α\alpha as a function of TeffT_{\mathrm{eff}} using the 880 KOIs with all five variables available. We place 20 equally spaced grid points over the temperature range spanned by these KOIs, from Tmin⁡=3,500T_{\min}=3{,}500 K to Tmax⁡=6,500T_{\max}=6{,}500 K, and infer the value αj\alpha_j at each grid point. Linear interpolation between the grid points defines the function fα(Teff,α)f_\alpha(T_{\mathrm{eff}},\alpha). The likelihood is

p(D∣α)∝∏ipKIC∗(rKOI,i−fα(Teff,i,α),yKOI,i),p(D \mid\alpha) \propto\prod_i p_{\mathrm{KIC}}^*\left(r_{\mathrm{KOI},i}-f_\alpha(T_{\mathrm{eff},i},\alpha),y_{\mathrm{KOI},i}\right),

which we evaluate using the fitted normalizing-flow density qϕ^q_{\hat{\phi}}. Thus only the amplitude coordinate of each KOI is shifted; its observed TeffT_{\mathrm{eff}}, mKepm_{\mathrm{Kep}}, log⁡R⋆\log R_\star, and log⁡Prot\log P_{\mathrm{rot}} are held fixed.

We place a Gaussian-process prior on α\alpha to regularize the variation between neighboring grid points. We define the scaled temperature coordinate

T~≡Teff−Tmin⁡Tmax⁡−Tmin⁡\widetilde{T} \equiv\frac{T_{\mathrm{eff}}-T_{\min}}{T_{\max}-T_{\min}}

and adopt

p(α∣μα,σα,ℓ)=N(α;μα1,Σα),p(\alpha\mid\mu_\alpha,\sigma_\alpha,\ell)=\mathcal{N}(\alpha;\mu_\alpha\mathbf{1},\Sigma_\alpha),
(Σα)ij=σα2(1+3∣T~i−T~j∣ℓ)exp⁡(−3∣T~i−T~j∣ℓ),(\Sigma_\alpha)_{ij}=\sigma_\alpha^2\left(1+\frac{\sqrt{3}\lvert\widetilde{T}_i-\widetilde{T}_j\rvert}{\ell}\right)\exp\left(-\frac{\sqrt{3}\lvert\widetilde{T}_i-\widetilde{T}_j\rvert}{\ell}\right),

which is the Matérn-3/2 kernel [55]. Here μα\mu_\alpha is a constant mean, σα\sigma_\alpha is the marginal standard deviation, and ℓ\ell is a dimensionless correlation length on the scaled temperature coordinate. We restrict each αj\alpha_j to [−0.2,0.2][-0.2,0.2] and adopt a uniform prior μα∼U(−0.2,0.2)\mu_\alpha\sim\mathcal{U}(-0.2,0.2), along with independent log-uniform priors over 0.005<σα<0.50.005<\sigma_\alpha<0.5 and 0.05<ℓ<50.05<\ell<5.

The joint posterior is therefore

p(α,μα,σα,ℓ∣D)∝p(D∣α) p(α∣μα,σα,ℓ) p(μα) p(σα) p(ℓ).(7)p(\alpha,\mu_\alpha,\sigma_\alpha,\ell\mid D) \propto p(D\mid\alpha)\,p(\alpha\mid\mu_\alpha,\sigma_\alpha,\ell)\,p(\mu_\alpha)\,p(\sigma_\alpha)\,p(\ell). \tag*{(7)}

We initialize the chains at a variational maximum-aposteriori solution and sample Equation (7) using the No-U-Turn sampler (M. D. Hoffman & A. Gelman [32]) implemented in NumPyro (D. Phan et al. [53]). We use four chains, each with 1,000 warm-up steps and 1,000 retained samples, and marginalize over μα\mu_\alpha, σα\sigma_\alpha, and ℓ\ell. We confirmed that split-rhat (A. Gelman et al. [24]) was less than 1.03 for all parameters.

Results

Figure 9 shows the inferred α(Teff)\alpha(T_{\rm eff}) in orange. The median shift is approximately 0.07–0.09 over most of the range below ∼5,500\sim5{,}500 K, decreases toward higher temperatures, remains positive up to ∼6,100\sim6{,}100 K, and becomes consistent with zero for hotter stars. For comparison, the gray curve in Figure 9 shows the same detectability-weighted analysis using only (Teff,r)(T_{\rm eff},r) in the joint density, without conditioning on zz. This case thus corresponds to the analysis of M15 after accounting for detectability of transiting planets, but not for confounding variables. This curve crosses zero near the solar temperature and becomes negative at higher temperatures, reproducing the qualitative trend reported by M15.

Inferred amplitude shift as a function of effective temperature

Figure 9. The inferred amplitude shift α(Teff)\alpha(T_{\rm eff}), defined as KOI minus KIC in logarithmic amplitude. The curves and shaded regions show the posterior medians and 16th–84th percentile intervals. The orange curve conditions on (Teff,mKep,log⁡R⋆,log⁡Prot)(T_{\rm eff},m_{\rm Kep},\log R_{\star},\log P_{\rm rot}), whereas the gray curve conditions only on TeffT_{\rm eff}. Both analyses use KIC reference densities weighted by the detectability of transiting planets, πD∣TR\pi_{\mathrm{D}|\mathrm{TR}}. Positive values indicate larger amplitudes for KOIs.

Because both curves include the same πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} correction, their difference isolates the effect of conditioning on (mKep,log⁡R⋆,log⁡Prot)(m_{\rm Kep},\log R_{\star},\log P_{\rm rot}). Including these variables removes the apparent negative shift for hotter stars and extends the positive KOI–KIC difference to nearly 6,100 K.

This result is qualitatively consistent with the resampling analysis in § 4, but the KOI–KIC separation is more pronounced in the likelihood-based analysis. The two approaches need not yield identical offsets because they summarize the conditional difference differently: the resampling analysis compares medians after approximate matching and marginalization over a common zz distribution, whereas the likelihood-based analysis fits a translational shift to the full conditional density at the observed zz of each KOI. The difference may reflect departures from a pure location shift, or the finite-bin and nearest-neighbor approximations used in the resampling procedure. In addition, as noted in § 5.2, the posterior intervals in the likelihood-based analysis do not include uncertainty associated with estimation of the joint density. We therefore place greater emphasis on the qualitative features shared by the two approaches than on the precise magnitude or detailed temperature dependence inferred from the flow-based analysis alone.

Motivated by the period dependence discussed by G. Li & J. N. Winn (2016) and more recently by E. M. Louden et al. [40], we also cross-matched the 880 KOIs with the candidate-only KOI catalog used to assign the orbital period of the innermost transiting planet PinnerP_{\mathrm{inner}} to each star. Of these, 810 had a match; the remaining 70 were excluded from the period-split analysis. We divided the matched KOIs at Pinner=8.6P_{\mathrm{inner}} = 8.6 days, with 405 stars in each subsample. For each subsample, we recomputed πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} by marginalizing over simulated planets in the corresponding period range and trained a separate weighted KIC flow. Figure 10 shows that both subsamples follow the same broad temperature dependence as the full sample. The long-period curve declines more gradually, whereas the short-period curve is systematically higher at the sampled points with Teff≳5,500T_{\mathrm{eff}} \gtrsim5{,}500 K. This difference is qualitatively consistent with the trend discussed by E. M. Louden et al. [40], in which systems with shorter-period planets tend to be more closely aligned. However, the separation is modest, and the result may be sensitive to the accuracy of πD∣TR\pi_{\mathrm{D}|\mathrm{TR}}. We therefore regard the evidence for period dependence as tentative.

Inferred $\alpha(T_{\mathrm{eff}})$ for all KOIs and period-split subsamples

Figure 10. The inferred α(Teff)\alpha(T_{\mathrm{eff}}) for all KOIs (orange dotted curve) and for two equal-sized subsamples divided by the orbital period of the innermost planet candidate: Pinner<8.6P_{\mathrm{inner}} < 8.6 days (salmon stars) and Pinner>8.6P_{\mathrm{inner}} > 8.6 days (blue squares). The shaded regions show the 16th–84th percentile intervals for the two subsamples. For each subsample, the KIC reference density was trained using πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} marginalized over the corresponding period range of the simulated planet population.

Sensitivity to the Adopted Effective Temperatures

The effective temperatures adopted by M15 and used in the analyses above are primarily based on the KIC (T. M. Brown et al. [10]); we refer to them below as the M15 temperatures. Here we test the sensitivity of our results to the adopted effective-temperature scale by repeating the analysis in § 5 using homogeneous spectroscopic values from the LAMOST DR11 low-resolution AFGK stellar-parameter catalog (X.-Q. Cui et al. [15]; A. Ren et al. [57]).

We constructed LAMOST-matched samples starting from the full catalogs underlying our analysis, which contain 33,614 KIC stars and 993 KOIs. We first associated each KIC identifier with a Gaia DR3 source identifier using the KIC–Gaia correspondence table in D. Godoy-Rivera et al. [26], and then performed an exact match in Gaia DR3 source identifier to the LAMOST DR11 catalog (A.-L. Luo et al. [41]). For a star with multiple LAMOST spectra, we adopted the spectrum having the highest gg-band signal-to-noise ratio. This procedure yielded finite LAMOST temperatures for 10,779 KIC stars and 338 KOIs.5 For the flow-based analysis, we retained KIC stars with 3500 K<Teff,LAMOST<6800 K3500\ \mathrm{K} < T_{\mathrm{eff,LAMOST}} < 6800\ \mathrm{K} for density estimation and KOI stars with 3500 K<Teff,LAMOST<6500 K3500\ \mathrm{K} < T_{\mathrm{eff,LAMOST}} < 6500\ \mathrm{K}. The wider training interval for the KIC stars provides support beyond the upper boundary of the inference interval. The resulting sample contains 9,401 KIC stars and 287 KOIs. The two temperature scales broadly agree within ∼\sim 100 K (Figure 11, left). The medians and 16th–84th percentile ranges of Teff,LAMOST−Teff,M15T_{\mathrm{eff,LAMOST}}-T_{\mathrm{eff,M15}} are +88−129+169+88^{+169}_{-129} K for the KIC stars and +54−123+127+54^{+127}_{-123} K for the KOIs, respectively. These differences provide a rough indication of the systematic uncertainty associated with the adopted temperature scale.

Sensitivity to the adopted effective temperatures

Figure 11. Sensitivity to the adopted effective temperatures. Left: Teff,LAMOST−Teff,M15T_{\mathrm{eff,LAMOST}} - T_{\mathrm{eff,M15}} for the final LAMOST-matched KIC and KOI samples, summarized in 250 K bins of Teff,M15T_{\mathrm{eff,M15}}. The curves show the bin medians and the shaded regions enclose the 16th–84th percentiles of the individual temperature differences. Right: the inferred amplitude shift for the same 9,401 KIC stars and 287 KOIs, using either the M15 effective temperatures (green triangles) or the LAMOST DR11 effective temperatures (orange circles) consistently for both the KIC reference density and the KOI likelihood. The curves and shaded regions show the posterior medians and 16th–84th percentile intervals. All inputs and analysis settings other than the effective-temperature measurements are held fixed; only the common, well-supported interval 4000–6400 K is shown in the right panel.

To further quantify the impact of the adopted temperature scale on the amplitude difference, we repeated the five-dimensional, all-period analysis in § 5 using this sample. We performed two analyses, one with the LAMOST temperatures and the other with the M15 temperatures, while keeping the sample membership fixed to isolate the impact of the temperature measurements from that of selecting the smaller sample. The other four flow coordinates, the πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} weights, the flow architecture, and the inference model were identical. The results are shown in the right panel of Figure 11. The inferred curves agree within their posterior uncertainties, suggesting that the analysis is not sensitive to the source of the effective temperatures. It is worth noting, however, that α\alpha is estimated to be smaller than in the full-sample analysis at the highest temperatures, while remaining consistent with zero within the uncertainties. The same behavior is seen when the M15 temperatures are used, indicating that the difference arises from restricting the analysis to the smaller sub-set of stars with LAMOST temperatures. This suggests that the sample definition may be more consequential than the choice of temperature catalog.

Discussion

Interpretation of the Trend

We find that stars with transiting planets and Teff≲6100T_{\mathrm{eff}} \lesssim6100 K have larger rr than stars in the comparison sample. This difference is consistent with a tendency toward low stellar obliquities. The inferred shift of ∼0.08\sim0.08 dex is also numerically similar to expectations from models for solar-like distributions of magnetic features. N.-E. Némec et al. [48], for example, found that the mean rotational variability of an isotropically oriented ensemble of otherwise solar-like stars is about 15% (0.07 dex) lower than that seen equator-on, although we note that the two quantities are not directly equivalent because our α\alpha is a descriptive shift of the post-selection distribution and need not equal the intrinsic amplitude shift (as mentioned at the end of § 5.1). This interpretation is qualitatively similar to that in T. Maze h et al. (2015), but after controlling for confounding variables, the photometric evidence consistent with spin-orbit alignment extends to higher temperatures. Our result provides further support for the low obliquities of small-planet hosts around cool stars inferred from previous stellar-inclination analyses and recent RM measurements (§ 1).

For hotter stars, we find that the rr distributions of stars with and without transiting planets are consistent with each other. This contrasts with M15, who found smaller amplitudes among planet hosts, a trend that is not straightforward to interpret geometrically. Thus, the modulation amplitudes no longer provide evidence for an obliquity distribution that would preferentially suppress the amplitudes of planet hosts, such as a predominantly polar distribution (e.g., S. H. Albrecht et al. [4]; J. Dong & D. Foreman-Mackey [18]). There are at least three possible explanations for the lack of difference that are not mutually exclusive:

  1. The two samples have similar inclination distributions because the correlation between stellar spins and planetary orbits in hot-star systems is weak (consistent with a broad obliquity distribution).

  1. The distribution of magnetic features differs between hotter and cooler stars, making rr less sensitive to the viewing angle for hotter stars.

  1. Conditioning on the detection of rotational modulation (R\mathcal{R}) can introduce inclination-dependent selection effects and attenuate the observed star–planet correlation. For example, if rotational modulation in hot stars is detectable mainly for nearly equator-on stars, both samples may be restricted to similar inclination ranges regardless of their underlying obliquity distributions.

Scenarios 2 and 3 imply that modulation amplitudes provide little leverage on the obliquities of these hot stars. One motivation for considering scenarios 2 and 3 is that the KOI–KIC difference in rr begins to diminish over the same TeffT_{\mathrm{eff}} range in which rr declines in the KIC population as a whole (e.g., Figure 3). This coincidence raises the possibility that the properties of magnetic features, such as their characteristic sizes or spatial distribution, change across this temperature range, potentially reducing the sensitivity of rr to viewing angle. The decline in rr could also make the detection of rotational modulation more susceptible to selection effects. The extent to which these mechanisms operate remains unclear. Related effects have been demonstrated in specific models (T. Reinhold et al. [56]; N.-E. Némec et al. [49]), but neither study establishes that they occur in the hot-star population considered here. Although these scenarios are difficult to distinguish using photometric modulation alone, independent vsin⁡iv \sin i analyses, which do not rely on spot distributions or the detection of rotational modulation, are consistent with a broad obliquity distribution among hot small-planet hosts and therefore lend support to scenario 1 for these stars [39, 40]. The tentative period dependence in our results may also be qualitatively consistent with that reported by E. M. Louden et al. [40]. Taken together, these results are consistent with a temperature-dependent picture in which small-planet hosts tend to have lower obliquities at lower temperatures and a broader obliquity distribution at higher temperatures. The similarity of this trend to that observed for hot Jupiters suggests that the temperature dependence may not arise solely from tidal realignment by hot Jupiters.

The possible similarity in the temperature dependence does not necessarily imply the same transition shape. Our rKOIr_{\mathrm{KOI}}–rKICr_{\mathrm{KIC}} comparison may instead suggest a relatively smooth change over 500 K, although its inferred smoothness may also be influenced by the Gaussian-process regularization. Such a smooth transition is not obviously inconsistent with the vsin⁡iv \sin i results of E. M. Louden et al. [39, 40], who did not attempt to constrain the detailed shape of the temperature dependence. For hot Jupiters, the relatively sharp transition has motivated explanations involving tidal realignment [67, 2, 16]. The absence of a comparably sharp transition for small planets, even if real, would not be surprising, given their much weaker tidal interactions with their host stars. However, scenarios 2 and 3 could also contribute to the gradual decline, so the smooth temperature dependence may not be entirely due to a change in the obliquity distribution. Similar vsin⁡iv \sin i comparisons spanning this temperature range could help distinguish these possibilities.

Considerations for More Accurate Comparisons

Although our analysis accounts for the main identified confounding variables, several aspects could be improved. Because the methodology presented here can also be applied to other analyses, we summarize caveats relevant to future applications.

Have we controlled all the relevant variables?—Our adopted set of zz approximately controls for stellar evolutionary state and includes variables related to detectability, so it should capture the most important effects. Nevertheless, it is generally difficult to establish that the choice includes all relevant confounding variables. In the present problem, one remaining candidate is metallicity, which affects the thickness of the convective envelope and hence activity and rr, and whose distribution may differ between the KIC and KOI samples through its potential correlation with planet occurrence. If activity amplitude and planet occurrence both increase with metallicity, controlling for it could reduce the inferred KOI–KIC amplitude difference, although the current uncertainties for hot stars are already large. Unresolved binary companions may provide another confounder because they can dilute the photometric signal and may also correlate with planet occurrence, particularly for close companions (e.g., [46]). The flow-based framework presented in § 5 may be useful when zz includes many variables.

Models of the Planet Population and Its Detectability—Our comparison avoids the need to model the detection function of rotational modulation, for which we do not have a quantitative model. We nevertheless account empirically for planet detectability using πD∣TR\pi_{\mathrm{D}|\mathrm{TR}}. Although πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} is estimated from injection–recovery simulations performed on actual stellar light curves, it still depends on the unknown underlying planet population over which it is marginalized. Thus, limitations of the adopted population model can propagate into πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} and potentially affect our results. We adopted a population inferred from the Kepler data, but, for example, did not account for its dependence on TeffT_{\mathrm{eff}} or for planet multiplicity (see review by W. Zhu & S. Dong [69]).

As a simple sensitivity check, we instead estimated πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} using the observed KOI population as a function of TeffT_{\mathrm{eff}} and found that the inferred trends were not significantly affected. Although the observed KOI population is itself affected by detection biases and therefore does not represent the true planet population, this agreement suggests that the population assumptions tested here do not strongly affect the results. This likely occurs because the values of πD∣TR\pi_{\mathrm{D}|\mathrm{TR}} for individual stars are correlated between the two assumptions, so their relative weights change little. A more accurate model of the planet population would nevertheless be preferable.

Beyond the Temperature Dependence—Throughout this paper, we modeled the KOI–KIC amplitude shift as a function of TeffT_{\mathrm{eff}}. This choice was motivated by the temperature dependence reported for hot Jupiters and by the analysis of M15, which this work builds on. However, the amplitude shift may depend on other stellar or planetary properties as well. One can instead model α\alpha as a function of multiple variables and estimate parameters that quantify the strength of each dependence. This approach would allow possible dependencies to be explored in a data-driven manner. The normalizing flow framework can naturally be extended in this direction.

Model for the Amplitude Difference—Our model treats α\alpha as a translational shift of the KOI rr distribution relative to the KIC distribution at fixed y=(z,Teff)y=(z,T_{\mathrm{eff}}), and therefore assumes that the two distributions have the same shape. This approximation is sufficient for capturing the amplitude offset as considered by T. Mazeh et al. [44]. If the difference is driven by stellar inclination, however, the KOI distribution need not be a simple translation of the KIC distribution: a change in the inclination distribution may instead alter its skewness or the relative weight of its high-amplitude tail. A more flexible model could allow both the location and shape of the distribution to vary, or could introduce the stellar inclination as a latent variable and marginalize over its distribution.

A separate limitation concerns the interpretation of this shift. As noted below Equation (4), the inferred α\alpha is a descriptive shift away from the null distribution pKIC∗p_{\mathrm{KIC}}^*, rather than an intrinsic shift between the pre-selection amplitude distributions of planet hosts and comparison stars. A more precise treatment would therefore require a full forward model of the selection process. In addition to a more accurate model of the planet population and its detectability, as discussed above, such a model would require a quantitative description of the detectability of rotational modulation (R\mathcal{R}). Such an analysis is possible in principle, but we do not pursue it here because a more detailed statistical model alone would not provide a unique obliquity interpretation: even an intrinsic difference in amplitude distributions could also arise from changes in the distribution of magnetic features on the stellar surface. Nevertheless, full forward modeling would be a possible future extension once the relevant selection functions and surface-variability models can be calibrated.

Summary and Conclusions

We revisited the comparison of rotational-modulation amplitudes between Kepler stars with and without detected transiting planets. We showed that a direct comparison is sensitive to differences in stellar properties that are correlated with both sample membership and modulation amplitude, particularly apparent magnitude, stellar radius, and rotation period, as well as to planet detectability. We addressed these effects using two complementary approaches: a resampling analysis that matches the distributions of the relevant stellar properties, and a likelihood-based analysis that enables comparisons at fixed stellar properties by modeling their joint distribution with the amplitude using a normalizing flow. The agreement between the two approaches indicates that controlling for these effects is particularly important for stars hotter than the Sun.

After these corrections, we find no evidence that stars with transiting planets have smaller modulation amplitudes at any of the temperatures considered. The likelihood-based analysis instead yields a positive shift of approximately 0.07–0.09 dex below ∼5,500\sim5{,}500 K; the shift decreases toward higher temperatures, remains positive up to ∼6,100\sim6{,}100 K, and is consistent with zero for hotter stars. The resampling analysis gives a qualitatively consistent, although less pronounced, result. Dividing the sample by the period of the innermost planet also hints at a larger shift for shorter-period systems at Teff≳5,500T_{\mathrm{eff}} \gtrsim5{,}500 K, but the evidence for this dependence remains tentative.

Provided that modulation amplitude is, on average, positively correlated with stellar inclination, the positive shift for cooler stars can be interpreted as evidence for a tendency toward low stellar obliquities. For hotter stars, the lack of a detectable shift is compatible with the broader obliquity distribution inferred from vsin⁡iv\sin i analyses and hence with more frequent misalignment. The tentative tendency toward larger shifts for shorter-period systems may further suggest that the period dependence of alignment reported by E. M. Louden et al. [40] extends toward cooler stars. Taken together, these results are consistent with a possible unified picture in which cool small-planet hosts tend to be aligned and hot hosts are more frequently misaligned. The qualitative similarity of the temperature dependence to that seen for hot Jupiters suggests that this pattern may not be produced solely by tidal realignment of hot-Jupiter hosts. Nevertheless, photometric modulation alone may not provide a unique obliquity diagnostic, because the observed trends could also be affected by the distribution of magnetic features on the stellar surface and by selection effects associated with detecting rotational modulation. Thus, the obliquity interpretation rests on the combined evidence from photometric modulation and vsin⁡iv\sin i. Future vsin⁡iv\sin i measurements and more complete forward models of the relevant selection functions will be important for testing whether the observed temperature dependence reflects a genuine change in obliquity.

More generally, this work demonstrates that comparisons between detected populations require careful treatment of confounding variables and selection effects. The resampling and conditional-density frameworks developed here provide general approaches for such comparisons.

Acknowledgments

We acknowledge the work of T. Mazeh et al. [44], which motivated the present study. We are grateful to the reviewer for their careful reading of the manuscript and constructive suggestions, which helped improve the paper. We also thank Takato Tokuno for helpful discussions. Work by KM was supported by JSPS KAK-ENHI grant Nos. 21H04998, 25K07387, 26K22393, and 26H02074. The authors used Codex (OpenAI) to assist with language editing and improving the clarity of the manuscript. The authors reviewed all changes and take full responsibility for the content of the manuscript.

Software: JAX (J. Bradbury et al. [9]), FlowJAX (D. Ward [66]), NumPyro (D. Phan et al. [53]), corner.py (D. Foreman-Mackey [22]), SysSim (M. Y. He [28])

References

  1. [1]Albrecht, S., Winn, J. N., Marcy, G. W., et al. 2013, ApJ, 771, 11, doi: 10.1088/0004-637X/771/1/11DOI
  2. [2]Albrecht, S., Winn, J. N., Johnson, J. A., et al. 2012, ApJ, 757, 18, doi: 10.1088/0004-637X/757/1/18DOI
  3. [3]Albrecht, S. H., Dawson, R. I., & Winn, J. N. 2022, PASP, 134, 082001, doi: 10.1088/1538-3873/ac6c09DOI
  4. [4]Albrecht, S. H., Marcussen, M. L., Winn, J. N., Dawson, R. I., & Knudstrup, E. 2021, ApJL, 916, L1, doi: 10.3847/2041-8213/ac0f03DOI
  5. [5]Batygin, K. 2012, Nature, 491, 418, doi: 10.1038/nature11560DOI
  6. [6]Berger, T. A., Huber, D., Gaidos, E., & van Saders, J. L. 2018, ApJ, 866, 99, doi: 10.3847/1538-4357/aada83DOI
  7. [7]Biddle, L. I., Bowler, B. P., Morgan, M., Tran, Q. H., & Wu, Y.-L. 2025, Nature, 644, 356, doi: 10.1038/s41586-025-09324-0DOI
  8. [8]Borucki, W. J., Koch, D., Basri, G., et al. 2010, Science, 327, 977, doi: 10.1126/science.1185402DOI
  9. [9]Bradbury, J., Frostig, R., Hawkins, P., et al. 2018, JAX: composable transformations of Python+NumPy programs, 0.2.5 http://github.com/google/jax
  10. [10]Brown, T. M., Latham, D. W., Everett, M. E., & Esquerdo, G. A. 2011, AJ, 142, 112, doi: 10.1088/0004-6256/142/4/112DOI
  11. [11]Campante, T. L., Lund, M. N., Kuszlewicz, J. S., et al. 2016, ApJ, 819, 85, doi: 10.3847/0004-637X/819/1/85DOI
  12. [12]Christiansen, J. L., Clarke, B. D., Burke, C. J., et al. 2016, ApJ, 828, 99, doi: 10.3847/0004-637X/828/2/99DOI
  13. [13]Corsaro, E., Bonanno, A., Mathur, S., et al. 2021, A&A, 652, L2, doi: 10.1051/0004-6361/202141395DOI
  14. [14]Crenshaw, J. F., Kalmbach, J. B., Gagliano, A., et al. 2024, AJ, 168, 80, doi: 10.3847/1538-3881/ad54bfDOI
  15. [15]Cui, X.-Q., Zhao, Y.-H., Chu, Y.-Q., et al. 2012, Research in Astronomy and Astrophysics, 12, 1197, doi: 10.1088/1674-4527/12/9/003DOI
  16. [16]Dawson, R. I. 2014, ApJL, 790, L31, doi: 10.1088/2041-8205/790/2/L31DOI
  17. [17]Dawson, R. I., & Johnson, J. A. 2018, ARA&A, 56, 175, doi: 10.1146/annurev-astro-081817-051853DOI
  18. [18]Dong, J., & Foreman-Mackey, D. 2023, AJ, 166, 112, doi: 10.3847/1538-3881/ace105DOI
  19. [19]Durkan, C., Bekasov, A., Murray, I., & Papamakar ios, G. 2019, arXiv e-prints, arXiv:1906.04032, doi: 10.48550/arXiv.1906.04032DOI
  20. [20]Fabrycky, D. C., Lissauer, J. J., Ragozzine, D., et al. 2014, ApJ, 790, 146, doi: 10.1088/0004-637X/790/2/146DOI
  21. [21]Fielding, D. B., McKee, C. F., Socrates, A., Cunningham, A. J., & Klein, R. I. 2015, MNRAS, 450, 3306, doi: 10.1093/mnras/stv836DOI
  22. [22]Foreman-Mackey, D. 2016, The Journal of Open Source Software, 24, doi: 10.21105/joss.00024
  23. [23]Fulton, B. J., Petigura, E. A., Howard, A. W., et al. 2017, AJ, 154, 109, doi: 10.3847/1538-3881/aa80ebDOI
  24. [24]Gelman, A., Carlin, J. B., Stern, H. S., et al. 2014, Bayesian data analysis, 3rd edn., Texts in statistical science (CRC Press). https://ci.nii.ac.jp/ncid/BB13945229
  25. [25]Gilbert, G. J., Petigura, E. A., & Entrican, P. M. 2025, Proceedings of the National Academy of Science, 122, e2405295122, doi: 10.1073/pnas.2405295122DOI
  26. [26]Godoy-Rivera, D., Mathur, S., García, R. A., et al. 2025, A&A, 696, A243, doi: 10.1051/0004-6361/2023487375
  27. [27]Handley, L. B., Howard, A. W., Dai, F., et al. 2026, arXiv e-prints, arXiv:2603.23713, doi: 10.48550/arXiv.2603.23713DOI
  28. [28]He, M. Y. 2025, SysSimPyPlots: Functions for plotting galleries of systems., Astrophysics Source Code Library, record ascl:2507.004 http://ascl.net/2507.004
  29. [31]Hirano, T., Sanchis-Ojeda, R., Takeda, Y., et al. 2012, ApJ, 756, 66, doi: 10.1088/0004-637X/756/1/66DOI
  30. [32]Hoffman, M. D., & Gelman, A. 2011, arXiv e-prints, arXiv:1111.4246. https://arxiv.org/abs/1111.4246
  31. [34]Knudstrup, E., Albrecht, S. H., Winn, J. N., et al. 2024, A&A, 690, A379, doi: 10.1051/0004-6361/202450627DOI
  32. [36]Lai, D. 2014, MNRAS, 440, 3532, doi: 10.1093/mnras/stu485DOI
  33. [38]Lissauer, J. J., Ragozzine, D., Fabrycky, D. C., et al. 2011, ApJS, 197, 8, doi: 10.1088/0067-0049/197/1/8DOI
  34. [39]Louden, E. M., Winn, J. N., Petigura, E. A., et al. 2021, AJ, 161, 68, doi: 10.3847/1538-3881/abcebdDOI
  35. [40]Louden, E. M., Wang, S., Winn, J. N., et al. 2024, ApJL, 968, L2, doi: 10.3847/2041-8213/ad4b1bDOI
  36. [41]Luo, A.-L., Zhao, Y.-H., Zhao, G., et al. 2026, VizieR Online Data Catalog: LAMOST DR11 catalogs (Luo+, 2026)., VizieR On-line Data Catalog: V/162. Originally published in: 2026RAA...in.prep..L
  37. [42]Masuda, K. 2022a, ApJ, 933, 195, doi: 10.3847/1538-4357/ac7527DOI
  38. [44]Mazeh, T., Perets, H. B., McQuillan, A., & Goldstein, E. S. 2015, ApJ, 801, 3, doi: 10.1088/0004-637X/801/1/3DOI
  39. [45]McQuillan, A., Mazeh, T., & Aigrain, S. 2014, ApJS, 211, 24, doi: 10.1088/0067-0049/211/2/24DOI
  40. [46]Moe, M., & Kratter, K. M. 2021, MNRAS, 507, 3593, doi: 10.1093/mnras/stab2328DOI
  41. [47]Muñoz, D. J., & Perets, H. B. 2018, AJ, 156, 253, doi: 10.3847/1538-3881/aae7d0DOI
  42. [48]Němec, N.-E., Işık, E., Shapiro, A. I., et al. 2020, A&A, 638, A56, doi: 10.1051/0004-6361/202038054DOI
  43. [49]Němec, N.-E., Shapiro, A. I., Işık, E., Solanki, S. K., & Reinhold, T. 2023, A&A, 672, A138, doi: 10.1051/0004-6361/202244412DOI
  44. [50]Noyes, R. W., Hartmann, L. W., Baliunas, S. L., Duncan, D. K., & Vaughan, A. H. 1984, ApJ, 279, 763, doi: 10.1086/161945DOI
  45. [51]Papamakarios, G., Pavlakou, T., & Murray, I. 2017, arXiv e-prints, arXiv:1705.07057, doi: 10.48550/arXiv.1705.07057
  46. [52]Petigura, E. A., Marcy, G. W., Winn, J. N., et al. 2018, AJ, 155, 89, doi: 10.3847/1538-3881/aaa54cDOI
  47. [53]Phan, D., Pradhan, N., & Jankowiak, M. 2019, arXiv preprint arXiv:1912.11554
  48. [54]Polanski, A. S., Crossfield, I. J. M., Seifahrt, A., et al. 2025, AJ, 170, 182, doi: 10.3847/1538-3881/adf070DOI
  49. [55]Rasmussen, C. E., & Williams, C. K. I. 2006, Gaussian Processes for Machine LearningDOI
  50. [56]Reinhold, T., Shapiro, A. I., Witzke, V., et al. 2021, ApJL, 908, L21, doi: 10.3847/2041-8213/abde46DOI
  51. [57]Ren, A., Fu, J., De Cat, P., et al. 2016, ApJS, 225, 28, doi: 10.3847/0067-0049/225/2/28DOI
  52. [58]Ricker, G. R., Winn, J. N., Vander spek, R., et al. 2015, Journal of Astronomical Telescopes, Instruments, and Systems, 1, 014003, doi: 10.1117/1.JATIS.1.1.014003DOI
  53. [59]Schlaufman, K. C. 2010, ApJ, 719, 602, doi: 10.1088/0004-637X/719/1/602DOI
  54. [60]See, V., Roquette, J., Amard, L., & Matt, S. P. 2021, ApJ, 912, 127, doi: 10.3847/1538-4357/abed47DOI
  55. [61]Southworth, J. 2026, Universe, 12, 62, doi: 10.3390/universe12030062DOI
  56. [62]Spalding, C., & Batygrin, K. 2015, ApJ, 811, 82, doi: 10.1088/0004-637X/811/2/82DOI
  57. [63]Tremaine, S., & Dong, S. 2012, AJ, 143, 94, doi: 10.1088/0004-6256/143/4/94DOI
  58. [64]Triaud, A. H. M. J. 2018, The Rossiter-McLaughlin Effect in Exoplanet Research, 2, doi: 10.1007/978-3-319-53333-7_2
  59. [65]Wang, X.-Y., Wang, S., & Ong, J. M. J. 2026, ApJL, 996, L7, doi: 10.3847/2041-8213/ae21c5DOI
  60. [66]Ward, D. 2025, FlowJAX: Distributions and Normalizing Flows in Jax, 17.2.1 doi: 10.5281/zenodo.10402073
  61. [67]Winn, J. N., Fabrycky, D., Albrecht, S., & Johnson, J. A. 2010, ApJL, 718, L145, doi: 10.1088/2041-8205/718/2/L145DOI
  62. [68]Winn, J. N., Petigura, E. A., Morton, T. D., et al. 2017, AJ, 154, 270, doi: 10.3847/1538-3881/aa93e3DOI
  63. [69]Zhu, W., & Dong, S. 2021, ARA&A, 59, 291, doi: 10.1146/annurev-astro-112420-020055DOI
  64. [70]Zhu, W., Petrovich, C., Wu, Y., Dong, S., & Xie, J. 2018, ApJ, 860, 101, doi: 10.3847/1538-4357/aacd6d

Paper details

vlnc:W_cxshnp48
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Codex

View statement
Looking for related papers…
Contents