Identification and Mitigation of a Vibrational Telescope Systematic with Application to Spitzer

Exoplanet science has pushed the Spitzer Space Telescope (Werner et al., 2004) far beyond its initial design. Transiting and eclipsing exoplanet signals are on the order of 1% and 0.1% of their host star, respectively, far below expected performance of the InfraRed Array Camera (IRAC, Fazio et al., 2004). Soon, the James Webb Space Telescope (JWST) will provide an unprecedented combination of spectral resolution, wavelength coverage, collecting area, and stability for exoplanet science (Deming & Seager, 2017). Currently, the field is limited to 1D characterization of most planets (e.g., Kreidberg et al., 2018), 2D thermal mapping of very hot targets (de Wit et al., 2012; Majeau et al., 2012), and 3D characterization with Hubble Space Telescope data of the hottest known planets (e.g., Stevenson et al., 2014, 2017). JWST will significantly improve planetary characterization possiblities, but small and cold planets will still be a challenge. Hotter terrestrial targets, like TRAPPIST-1b (Gillon et al., 2016, 2017), will require ${\sim}10$ secondary eclipses for a confident detection (Morley et al., 2017), but temperate Earth-like targets will be difficult, if not impossible (Rauer et al., 2011; Rugheimer et al., 2015; Batalha et al., 2018; Beichman & Greene, 2018). We must take advantage of every technique available if we hope to characterize these planets.

Spitzer IRAC suffers from two primary systematic effects: an easily-removed “ramp” that causes measured flux to vary with time, and an intrapixel gain variation that creates correlations between flux and target position on the detector at a subpixel level (e.g., Charbonneau et al., 2005). These effects are present in all 3.6 and 4.5 µm observations, although the ramp is sometimes weak enough to be ignored. Several independent modeling techniques, such as Gaussian-weighted maps (Ballard et al., 2010; Lewis et al., 2013), BiLinearly Interpolated Subpixel Sensitivity (BLISS, Stevenson et al., 2012), Pixel-Level Decorrelation (PLD, Deming et al., 2015), and Independent Component Analysis (Morello, 2015) have successfully removed the position-correlated noise, enabling transiting exoplanet observations with uncertainties $<$100 ppm and retrieving accurate planetary parameters (Ingalls et al., 2016).

A third, much less common effect creates light-curve features that resemble transiting exoplanets. This effect has been linked to activity in the “noise pixel” parameter (Mighell, 2005; Lewis et al., 2013), a measurement of the number of pixels that contribute to centering and photometry, or the size of the point-response function (PRF). Spikes in the noise pixels are known to correlate with transit-like signals, and are likely caused by high-frequency telescope oscillations of unknown origin (see irachpp.spitzer.caltech.edu/page/np_spikes). Previous studies show that PRF Gaussian-width dependent models can account for time-evolving point-source shapes (e.g., Lanotte et al., 2014; Mendonça et al., 2018; Mansfield et al., 2020), which may be related to the noise-pixel effect.

We observed Proxima Centauri (hereafter Proxima) in 2016 and 2017 with Spitzer IRAC to search for transits of the planet Proxima b (Anglada-Escudé et al., 2016). Jenkins et al. (2019), hereafter J19, presented the results from the first observation. When following standard data reduction procedures and dividing out best-fitting systematic models (with only a flat astrophysical model), these observations contain three transit-like events (see Figure 1) that resemble the asymmetric shapes created by transits of evaporating planets (e.g., Rappaport et al., 2014; Sanchis-Ojeda et al., 2015; Vanderburg et al., 2015) or comets (e.g., Rappaport et al., 2018). We now know, conclusively, that these events are localized systematic effects due to high-frequency telescopic vibration of unknown origin. When the telescope vibrates, the PRF smears along the direction of the vibration. During the vibration, fixed-radius photometry apertures spill light, resulting in lower measured flux with larger vibrational amplitudes.

In this work, we present evidence that the systematic is due to vibration, several new methods to identify when this vibrational systematic occurs, and a new aperture photometry method to correct it. The paper is organized as follows: in Section 2 we describe our observations, in Section 5 we discuss how to identify this systematic, in Section 6 we present our elliptical photometry method, in Section 8 we interpret our findings, and in Section 9 we lay out our conclusions.

We observed Proxima with the Spitzer InfraRed Array Camera, with both the 3.6 and 4.5 µm filters, for a total of $>80$ hours (Table 1). The 48-hour stare bracketed the predicted transit time of Proxima b, and shorter observations occurred at times when further transits should occur, if the feature in the stare was caused by Proxima b (Anglada-Escudé et al., 2016). All observations were in $32\times 32$ subarray mode and centered on the IRAC “sweet spot”, at (-0.352″, 0.064″) and (-0.511″, 0.039″) for 3.6 µm and 4.5 µm, respectively. We bracketed each science observation with an initial 30-minute observation to minimize telescope pointing settling and a final 10-minute observation, for those who wish to generate their own dark frames, per Spitzer Science Center recommendations.

Notably, due to the brightness of the target, our observations utilized the shortest frame time, 0.02 seconds, which allows temporal resolution of high-frequency effects. To handle the large data volume from this cadence, our observations have data gaps. The 48 hour 4.5 µm stare has 17-second gaps between 64-frame subarray chunks, 24-second gaps between Astronomical Observation Requests (AORs), and ${\sim}4$ minute gaps every 16 hours for data downlink and target reacquisition. Both 3.6 µm observations have 6 second gaps between subarray chunks and 14 second gaps between AORs. The shortest 4.5 µm observation has 2 – 2.5-second gaps between subarray chunks, and only one AOR. The November 2017 4.5 µm  observation has the same gaps as the 48-hour stare, without the downlink and target reacquisition.

The telescope’s heater, which introduces motion on the detector in $\sim$40-minute cycles, was turned off for the duration of all five observations, following then-current Spitzer procedures for exoplanet observations. This minimizes the impact of the intrapixel systematic, allowing closer study of other effects.

We use the Photometry for Orbits, Eclipses, and Transits code (POET, Stevenson et al., 2012; Blecic et al., 2013; Cubillos et al., 2013; Blecic et al., 2014; Cubillos et al., 2014; Hardy et al., 2017, e.g.,) for all analyses herein. The steps in producing light curves are bad pixel identification, determining the location of the target (centering), and measuring its brightness (photometry). We identify bad pixels by performing a twice-iterated 4$\sigma$ rejection on 64-frame chunks, and discard these pixels from further analysis. This work relies heavily on different centering and photometry methods, so we describe our implementations in detail.

We modeled our light curves with both PLD and BLISS to correct the intrapixel systematic and to assess each model’s ability to address the vibrational systematic. BLISS maps correct for intrapixel sensitivity variations by gridding the detector into fine subpixels. We assign each frame to a subpixel based on the target position from centering, compute the sensitivity of each subpixel based on the average flux of all frames associated with them, once all other models (astrophysical or otherwise) have been removed, and bilinearly interpolate the sensitivity grid to find a correction factor for each frame. The generic full model formula is

PLD removes the same effect by treating the data as a weighted sum of normalized pixel values, where the weights are free parameters of the model. The model is

Figure 1 shows the fixed-aperture light curves, modeled with BLISS, to highlight the vibrational systematic. Aperture radii are 3.00 pixels (November 2016 4.5 µm), 2.25 pixels (May 2017 4.5 µm), 3.25 pixels (June 2017 3.6 µm), 4.50 pixels (July 2017 4.5 µm), and 2.75 pixels (November 2017 4.5 µm). The systematic is present in the November 2016 4.5 µm, June 2017 3.6 µm, and July 2017 3.6 µm light curves, so we focus on these observations going forward. While the vibration appears to be more common at 3.6 µm (2 – 3 occurrences) than at 4.5 µm (1 occurrence), possibly due to a narrower PRF, our sample size is only large enough to confirm that the vibration is not limited to one bandpass.

Past works used the noise pixel measurement (Equation 2) to identify activity in the PRF, and correct for it with variable-aperture photometry (e.g., Lewis et al., 2013; Deming et al., 2015; Garhart et al., 2018; Jenkins et al., 2019). Effectively, noise pixels measure the number of pixels above the background (contributing to centering and photometry). A wider (narrower) PRF should result in a larger (smaller) noise-pixel value. Since noise pixels measure an area, the radius of the photometry aperture required for the PRF is the root of the noise pixels, commonly with additional multiplicative and/or additive scaling (see Section 3.2). Thus, as the PRF size varies throughout the observation, so does the photometry aperture radius.

J19 found that, using common techniques, centering and photometry selection criteria selected against variable photometry apertures. We have since improved the variable-aperture photometry by calculating the aperture radii after background subtraction, which reduces uncertainty introduced by unimportant pixels. Figure 2 shows a comparison of aperture radii ($\sqrt{\beta}$ in Equation 1) in the July 3.6 µm observation when $\beta$ is computed before and after background subtraction. In this case, the standard deviation of $\sqrt{\beta}$ decreases by $\sim 14$%. With this improvement, variable-aperture radii are preferred over fixed-aperture radii for these observations, although they still introduce noise to the light curve due to the additional noise-pixel parameter. It is unclear if Lewis et al. (2013), who introduced variable photometry apertures, calculated noise pixels and aperture radii before or after background subtraction.

Oscillations in the telescope, if higher frequency than the exposure time, could be hidden from centering, but they would be evident in a widening of the PRF in the direction of the vibration. By fitting a Gaussian to the PRF, we determine 1$\sigma$ widths in $x$ and $y$ (see Figure 3, second and third rows) and notice a prominent widening in the PRF at the time of the systematic. This widening is even more evident in a measure of the 3$\sigma$ area of the Gaussian, which we compute as an ellipse with axes along the $x$ and $y$ directions (see Figure 3, fourth row). We also measure the variance in this elliptical area, on a 64-frame basis, to look for PRF activity (see Figure 3, fifth row).

Our short exposures (0.02 seconds) allow temporal resolution of high-frequency effects. Figure 4 shows the 3$\sigma$ elliptical area of a single set of 64 frames during the peak of the systematic in the July 2017 3.6 µm observation. We find a clear sinusoidal pattern with a period of 0.45 seconds, evidence for telescope oscillation. Since the shape of the PRF is changing, and the photometric effect is a net loss in flux, integrating exposures by a multiple of the oscillation timescale will not correct the effect.

The periodicity is localized in time, so we apply a continuous Morlet wavelet transform, using the pywavelets (Lee et al., 2019) Python package (see Figure 5). Wavelet transforms assume a uniform sampling, but our observations are sets of 64 short-cadence frames separated by relatively long gaps, to work around data storage limits. This results in spurious periodicity in the wavelet transforms. Despite this limitation, a wavelet transform reveals periodic activity in the elliptical area of the PRF at the time that the systematic occurred, near frames 40,000 – 50,000. In particular, there is a cluster of stronger amplitudes at $\sim 2$ Hz, damping out to lower frequencies over the course of several thousand frames.

Lomb-Scargle periodograms are well-suited to finding periodicity in non-uniformly sampled data, but unlike wavelet transforms, they provide no temporal resolution of localized activity. To gain some insight into local periodicity, we use a windowed Lomb-Scargle periodogram (see Figure 6). The periodogram shows a strong peak at $\sim$ 2 Hz (as well as several weaker resonances), which matches the periodic behavior seen in Figure 4.

Until now, we calculated elliptical area from the $x$ and $y$ widths of the PRF. However, this measure of area is only accurate if vibrations are oriented along those axes. To more accurately measure the shape of the PRF, we rotate a Gaussian clockwise from the $x$ axis. This detaches the $x$ and $y$ widths from their respective axes, instead measuring the semimajor and semiminor axes of the ellipse. A rotated 2D Gaussian is described by

We applied this rotated-Gaussian centering method to the observations affected by the systematic. The results are displayed in the first seven rows of Figure 8. They match the non-rotated Gaussian fits in elliptical area and elliptical area variance. These systematic identification methods perform nearly equivalently when using the non-rotated Gaussian. However, the rotated Gaussian has implications for elliptical photometry, which is discussed in Section 6.

There are bimodalities in the fitted $y$ position, the axes lengths, and rotation of the ellipse when the center of the PRF passes below the center of a pixel. This behavior may be due to the asymmetry of the IRAC PRF, which has a roughly-triangular shape (e.g., the second panel of Figure 7). The ellipse is swapping between the asymmetric edges of the triangle (Figure 9). We see this behavior in the Proxima images and the synthetic images created with IRACSIM for the Spitzer data challenge (Ingalls et al., 2016), but not with simple synthetic Gaussians (Figure 7), suggesting it is a real effect of the complex PRF.

Past works have removed this vibrational systematic prior to modeling with variable, circular apertures (e.g., Lewis et al., 2013; Deming et al., 2015; Garhart et al., 2018; Jenkins et al., 2019). These apertures adjust to avoid spilling light. However, due to their circular shape, they must either spill flux from the aperture or overcompensate in size for the elliptically-smeared PRF to capture all the important pixels; thus, they include unnecessary background noise.

Instead, we use elliptical photometry, where we use an elliptical aperture described by the fitted parameters from the non-rotated Gaussian or rotated-Gaussian centering methods described in Section 5. With rotated Gaussian centering, we apply the rotation to the elliptical aperture. Similar to using variable-aperture photometry, elliptical apertures attempt to remove the effects of PRF activity prior to modeling, but only including the most important pixels, resulting in less noise. Several elliptical photometry packages exist (e.g., Laher et al., 2012; Barbary, 2016; Merlin et al., 2019), although application has been limited to correcting atmospheric effects in ground-based observations (Bowman & Holdsworth, 2019), measuring the radial surface brightness profiles of physically elliptical galaxies (e.g., Davis et al., 1985; Djorgovski, 1985; Cornell, 1989; Ryder, 1992; McNamara & O’Connell, 1992; Hayes et al., 2005), and measuring photometry of comets that move significantly during each exposure (Miles, 2009). To our knowledge, none have used elliptical apertures to correct for vibrational effects.

Qualitatively, we find that elliptical photometry almost entirely removes the vibrational systematic from the light curve, with the non-rotated ellipses outperforming the rotated ones (see Figures 3 and 8, last rows). To assess performance quantitatively, we fit BLISS and PLD models to the three observations which include the vibrational systematic. PLD performs poorly when applied to observations longer than typical eclipses and transits (Deming et al., 2015), so, for the 48-hour observation, we only consider the final 16 hours (after the final data downlink). Many PLD implementations also bin the data (e.g., Deming et al., 2015; Wong et al., 2015; Buhler et al., 2016), which can reduce short-period correlated noise. We choose not to bin to isolate each model’s ability to address correlated noise. Table 3 lists the results: $\chi^{2}_{\textrm{b}in}$-minimized photometry aperture sizes for each combination of centering and photometry methods, as well as the $\chi^{2}_{\textrm{b}in}$ (lowest in bold) and SDNR for each combination, both for BLISS and PLD fits. Figure 10 shows the $\chi^{2}_{\textrm{b}in}$-minimized BLISS light curves compared with the fixed-radius aperture light curves in Figure 1. We reduced the strength of the systematic from 0.16% to 0.06% in the November 2016 4.5 µm observation, from 0.14% to 0.06% in the June 2017 3.6 µm observation, and from 0.30% to 0.03% in July 2017 3.6 µm observation, measured by the minimum of the binned photometry presented in Figure 10.

Some analyses of Spitzer IRAC data found that light-curve correlated noise could be significantly reduced by combining a BLISS map with a model dependent on the Gaussian widths of the PRF (e.g., Lanotte et al., 2014; Mendonça et al., 2018; Mansfield et al., 2020). Here we compare that approach with elliptical-aperture photometry, and test the results of combining the two methods.

For the Gaussian-width detrending, we use the following generic quadratic model:

We repeat the process used in Section 6 of optimizing for the minimum $\chi^{2}_{\textrm{bin}}$ in each photometry method, although we are restricted to Gaussian centering, as the Gaussian widths are an input to the light-curve model. A sample of the results is shown in Figure 11. Unsurprisingly, models that include $W$ lead to better fits (lower $\chi^{2}$) due to increased flexibility. In most cases, the noise reduction is clearly significant when applied to fixed-radius aperture photometry. However, the improvement is more marginal when the model is applied to non- and rotated elliptical apertures. For example, for the July 2017 3.6 µm observation, $\chi^{2}_{\textrm{bin}}$ for non-rotated elliptical apertures improves from 4.9 to 4.8 and for rotated elliptical apertures improves from 23.7 to 19.5. In some cases including $W$ actually leads to more correlated noise because model-fitting seeks to minimize $\chi^{2}$, not $\chi^{2}_{\textrm{bin}}$. For all three observations, elliptical apertures led to a lower $\chi^{2}_{\textrm{bin}}$ than fixed-radius circular apertures with a Gaussian-width model.

We cannot use a statistic like the Bayesian Information Criterion (Raftery, 1995) to decide if $W$ is worth including. The $\chi^{2}_{\textrm{bin}}$-optimized photometry while including the $W$ model is different than the $\chi^{2}_{\textrm{bin}}$-optimized photometry without $W$; we would be comparing model fits to different data sets.

We draw several conclusions from the results in Table 3. First, we find elliptical photometry superior or equivalent to variable, circular apertures when using BLISS maps. The vibrational systematic is not correlated with position, especially if the vibration occurs at a period shorter than the exposure time, and thus cannot be corrected by a BLISS map. By removing the vibrational systematic with elliptical photometry, the accuracy of the BLISS map improves for the entire observation.

The PLD model is more flexible in its noise removal. It assumes that flux variations are tied to fluctuations in the pixel brightnesses. As the target moves on the detector, pixels brighten and dim. Likewise, if the PRF is smeared, pixels near the center of the target dim and pixels along the vibration axis brighten. Thus, the PLD model is able to correct for the vibrational systematic without explicit knowledge of the vibration, minimizing the advantage gained by using elliptical photometry. This is convenient, but we achieve much lower correlated noise in the BLISS models where the systematics are corrected with a physical description of their effects (see $\chi^{2}_{\textrm{bin}}$ values in Table 3). We do not use binning in our application of PLD, which would reduce correlated noise, but again without explicit knowledge of the vibration. For example, when testing bin sizes of 1, 2, 4, 8, 18, 32, 64, 128, and 256 on the optimal PLD light curves from Table 3, we improve the $\chi^{2}_{\textrm{bin}}$ from 34.7 to 34.3 for the November 2016 4.5 µm observation, from 28.8 to 12.0 for the June 2017 3.6 µm observation, and from 36.9 to 5.2 for the July 2017 3.6 microns observation. This suggests that temporal binning provides a significant portion of PLD’s correlated-noise correction capabilities.

The rotated ellipse is never preferred over the non-rotated case. As mentioned above, the Spitzer PRF is highly asymmetric, and slightly triangular in shape (see Figure 7, right panel), which creates a challenge when fitting a rotated ellipse. The vibration-induced elliptical shape is less prominent than the already-present asymmetry in the PRF, as evidenced by the bimodal distribution in rotation (Figure 8). We suspect that the rotated elliptical Gaussian is fitting to the sides of the triangular PRF, which creates additional noise in the resulting light curve. In particular, the bimodality creates problems for BLISS maps in two ways:

1. 1.

   BLISS maps rely on multiple frames at the same location to accurately calculate subpixel sensitivity. The bimodality reduces the number of frames in each sensitivity map grid cell, leading to greater uncertainty.

2. 2.

   The bimodality increases the size of the grid cells, which we calculate as the RMS of the point-to-point target position variation in $x$ and $y$. This reduces the flexibility of the model, resulting in worse fits.

Since the Spitzer PRF is a complex shape, ideally we would determine flux by directly fitting the PRF, but that has proven challenging. The Spitzer PRF is underresolved, especially at shorter wavelengths, and the true PSF is not known at a high resolution, only as a map of a point source at a 5$\times$5 grid of positions within a central pixel. Hence, we recommend overresolved PRFs for high-precision point-source instruments like exoplanet telescopes, or a high-resolution lab-measured PRF tested in comparison to real data with a routine to accurately bin to the native pixel level. One could also fit a shape more representative of the PRF, like a tri-lobed Gaussian with a radial scale, rotation, stretching factor, and stretching axis. However, that is beyond the scope of this work.

In general, we find that PLD is agnostic to the centering method used. In two observations, we prefer center-of-light centering, and in the third there is no strong preference for any of the methods. This would suggest that, when using PLD models, it is acceptable to only apply center-of-light centering, although we recommend always applying all methods available.

BLISS maps, on the other hand, are extremely sensitive to the centering method because 1) target position is an input to the model, and 2) we use BLISS map $x$ and $y$ grid sizes equal to the RMS of the point-to-point $x$ and $y$ target position motion, respectively. Thus, higher precision centering methods result in maps with finer structure. Compared to Gaussian and least-asymmetry centering, center-of-light centering results in high RMS of point-to-point $x$ and $y$ target position motion and, thus, poor maps, at least for 3.6 and 4.5 µm observations (Table 3). Therefore, center-of-light can be ignored with BLISS maps, although applying all analysis methods will ensure the best is chosen.

While including a Gaussian-width model can significantly reduce noise in fixed-aperture photometry, the addition of this model only slightly improves elliptical-aperture photometry. We recommend elliptical apertures over a Gaussian-width model for two reasons: 1) to reduce model complexity, and 2) to correct for correlated noise in image processing, rather than creating it with circular apertures and then modeling it out.

Finally, in nearly all cases, non-rotated elliptical photometry results in the lowest SDNR. With BLISS, elliptical photometry improves SDNR by up to 11.2% over fixed circular apertures and up to 6.0% over variable, circular apertures. With PLD, we see up to 9.4% improvement over fixed apertures and up to 6.3% improvement over variable apertures. These statistics are for the entire modeled light curve; the improvement is even more pronounced if we only consider data when the systematic is present. This suggests SDNR can be useful when optimizing photometric extraction, with the caveat that the SDNR values listed in Table 3 are for the $\chi^{2}_{\textrm{bin}}$-optimized light curves, not SDNR-optimized light curves. In our case, if we optimized solely for SDNR, we would be left with fixed-radius apertures for the November 2016 4.5 µm observation, which is clearly unsatisfactory (see Figure 1, top panel).

The optimal light curves presented here are available, in machine- and human-readable formats, in a compendium archive available at https://doi.org/10.5281/zenodo.3759914 (catalog https://doi.org/10.5281/zenodo.3759914). The compendium also includes best-fit models and correlated noise diagnostics.

We have identified a vibrational systematic in Spitzer photometry that mimics planetary or cometary transits. With our short exposure times, we were able to resolve this vibration in the size and shape of the PRF, both on sub-second timescales and with periodograms. We caution against false positive detections of planets, and recommend applying the techniques described here to identify and correct the systematic.

“Noise pixels” can occasionally identify this systematic, but they can be misleading, as noise pixel activity does not always correspond with the systematic, and can frequently be hidden in the baseline activity. Several other metrics are better suited to identifying this vibration:

1. 1.

   $x$ and $y$ widths from Gaussian centering, both rotated and non-rotated.

2. 2.

   Elliptical area of Gaussian centering, both rotated and non-rotated.

3. 3.

   Variance of noise pixels.

4. 4.

   Variance of elliptical area.

5. 5.

   Wavelet amplitude over a variety of frequencies.

6. 6.

   Lomb-Scargle periodograms of elliptical area.

To correct this vibrational systematic, we developed an adaptive elliptical-photometry technique. We fit an asymmetric Gaussian to the PRF to determine target position and PRF shape, and use this parameterization to create an elliptical aperture that adapts its shape to the PRF as it changes with time. We applied elliptical photometry to three observations known to include the vibrational systematic, with both BLISS and PLD models to assess relative performance. With BLISS models, elliptical photometry results in reduced correlated noise in two of our three observations, and reduced SDNR in all observations. PLD prefers variable, circular apertures over elliptical apertures, but, without binning, is less capable of removing correlated noise compared to BLISS. We also used a rotated elliptical aperture, but found that the complex shape of the Spitzer PRF created bimodalities in the orientation of the ellipse and noise in the resulting light curve. Other shapes, like a tri-lobed Gaussian, are an area of potential future study. Finally, we found that elliptical apertures outperformed traditional fixed-radius circular apertures with a Gaussian-width model.

We cannot determine the source of the vibration, though we speculate that it could be micrometeorite impacts or wear-and-tear on the telescope, such as a defect in the gyroscopes. If the source is micrometeorites, this systematic should be present in many past observations, at roughly the same rate as in our observations (four instances in 80 hours). Reanalyses with our techniques may be able to rescue data sets deemed unsalvageable, or at least improve the uncertainties on measured planetary transmission, emission, and phase curve variation. If wear-and-tear is the source of the systematic, then older observations may be unaffected, but more recent observations would still be affected. Spitzer produced high-profile exoplanet science for 16 years (e.g., Gillon et al., 2017; Kreidberg et al., 2019), much of which is done at the limit of detection. Elliptical photometry could make the difference between speculation and discovery.

Elliptical photometry is not limited to Spitzer. TESS and Kepler (and K2) are purely photometric observatories that may suffer from the same systematic. JWST also has photometric modes which will surely be used to push transiting exoplanet photometry to the smallest and coldest objects possible. Optimistically assuming that we reach the noise floor, we will need large amounts of JWST time to study these planets (e.g., Morley et al., 2017), and require the absolute best data reduction and noise removal techniques.