TRANSMISSION SPECTROSCOPY OF EXOPLANET XO-2b OBSERVED WITH HUBBLE SPACE TELESCOPE NICMOS

The instrumental parameters are derived for two reasons: to apply a wavelength-dependent flat-field correction, and to compute a decorrelation from these parameters in the derivation of the white-light transit curve and the planet spectrum. These parameters are represented in Figure 2.

• 1.  
  Position (x, y) of the zeroth order. We derive the (x, y) position of the observed zeroth order by cross-correlation with a reference image. The zeroth-order (x, y) position of the other star, which does not appear in the image, is derived by applying an offset to the zeroth order that is present. This offset is calculated by correlating the first-order spectral traces of both stars, and considering that the same shift applies to the zeroth orders. The best correlation is obtained by applying a constant shift of 118 pixels in the x-direction and 99 pixels in the y-direction for visits 1 and 2 for XO-2 N with respect to XO-2 S, and a shift of 118 pixels in the x-direction and −99 pixels in the y-direction for visit 3 for XO-2 S with respect to XO-2 N.
• 2.  
  Angle and FWHM of the spectral trace. The angle and FWHM of the first-order spectral trace are derived for each star. The spectral trace is divided into boxes of 5 pixels wide in x and 40 pixels wide in y. Each box is averaged in x by a median filter, and a one-dimensional Gaussian fit is computed in the y-direction. A linear fit of the center of all Gaussians yields the angle of the spectral trace. Each Gaussian also provides an FWHM in y, and the median of all values gives the FWHM of the spectral trace.
• 3.  
  Detector temperature and state. The detector temperature is derived from the bias level using a standard procedure for NICMOS (Pirzkal et al. 2009b). In addition, Burke et al. (2010) showed that the detector electronics vary between seven states, resulting in photometric variations approximately at the millimagnitude amplitude, peak to valley (Figure 3). The detector state in a given image is inferred by comparing the bias level in two different quadrants. These seven states are clearly identified in the XO-2 data.
• 4.  
  Filter wheel position. The filter wheel has a constant position during an orbit, but experiences small shifts between orbits due to the blank filter element insertion during Earth occultation of the target. These variations are believed to affect the position of the spectral trace, in particular its angle (Gilliland & Arribas 2003). The filter wheel has two positions.
• 5.  
  HST phase. The HST orbital phase is known to influence NICMOS data. However, this parameter is responsible for variations of other parameters that are more directly linked to the images, such as the FWHM and the detector temperature (Wiklind & Thatte 2009).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The calibration is made from the raw images. The bias calibration is performed by subtracting the first readout from the last (Fixsen et al. 2000). Pixels with a dark current greater than 1e⁻ s⁻¹ are already marked as bad, and lower dark currents are negligible for our purpose.

To correct for bad pixels, we combine the nominal bad pixel mask for the NIC3 detector with the pixel flags provided in each individual image. In addition, we flag as bad pixels those that deviate by more than 3.5σ from their average value in the orbit, or from their value in the adjacent images. We find typically 35 bad pixels out of ∼2100 in the spectral aperture. Each bad pixel's value is then replaced by a spline interpolation of adjacent pixels.

The set of bad pixels for each image defines a mask (Figure 4). In some masks, lines appear along the x-axis at the location of the spectral traces. This behavior appears mostly at the beginning of the orbits, for which the detector temperature is significantly higher and the FWHM significantly smaller than during the rest of the orbit, due to the satellite stabilizing after re-entering the Earth shadow (Figure 2). These lines are thus due to a different flux distribution compared to the rest of the orbit, and not to bad pixels. Typically, the first 20 images of each orbit are affected, plus ∼12 other images in each orbit of the third visit. In total, this represents 37% of the data. Since this fraction is large, we process these images similarly to the rest of the data, without considering those lines as bad pixels. The decorrelation from instrumental parameters will reduce the effects of such variations.

Zoom In Zoom Out Reset image size

Since the response of the pixels is wavelength dependent, a custom-made flat field is required.³ For each image, we calculate the reference position of the star using an image taken without the dispersive element. This reference position and the beam angle are used to calculate the wavelength at each pixel, using the grism's nominal dispersion relation (Pirzkal et al. 2009a). Then, the response of each pixel as a function of wavelength is determined by evaluating a fourth-order polynomial to flat fields taken at 10 different wavelengths. A flat field is made for each image, and an average flat field is built for each visit. The latter one is used for the calibration. Images of a given visit are thus calibrated using the same flat field.

In order to derive the star and planet parameters, we perform white-light photometry by summing the flux over the spectral trace. We optimize the photometric apertures for the star and sky background. We extract and compare light curves for apertures of different sizes, varying between 6 and 50 pixels in y. The sky background is computed from the same quadrant as the spectral trace, either above or below depending on the position of the star; the width of the sky region varies between 40 and 128 pixels in x, and between 20 and 60 pixels in y. The sky level is calculated with an outlier resistant mean with a threshold at 3.5σ, multiplied by the number of pixels in the star region, and subtracted to the stellar flux. The light-curve rms is calculated from the second and the fifth orbit of each visit, to avoid the first orbit which is commonly affected strongly by systematics, as well as the ingress, egress, and in-transit orbits. Independently for each visit, we determine the optimal height in y as the one resulting in the smallest rms. We find an optimal height in y of 18, 20, and 18 pixels for XO-2 N, and 24, 18, and 16 pixels for XO-2 S, for the first, second, and third visit, respectively.

The rms is calculated after correcting the orbit light curves for the seven detector states (Figure 3), without applying any other decorrelation. Indeed, the detector state clearly affects the photometry: flux measurements corresponding to a given state are systematically offset from the mean flux during the orbit. Following the procedure outlined in Burke et al. (2010), a correction is applied by correcting the mean of the measurements taken in each state to the mean flux of the orbit, after accounting for the transit if present. The seven-state correction lowers the rms by a factor 1.5 on average (Figure 5). Because orbits have a different average flux, the correction cannot be applied for a whole visit at once when extracting these raw light curves. However, in the remainder of this work, the average flux of each orbit is adjusted by optimization processes, and the seven-state correction will be applied for each visit instead of each orbit. Using the larger number of data points in a visit, compared to each orbit, provides better estimates of the seven state-dependent flux offsets and also reduces to insignificance the side effect that fitting these offsets will reduce even random noise simply because each offset is another parameter in the fit.

Zoom In Zoom Out Reset image size

The star and planet parameters are derived through a Markov Chain Monte Carlo (MCMC) minimization process (Burke et al. 2007). We simultaneously fit the light curve with a transit model and an instrumental model. We adopt a quadratic limb-darkened transit model from Mandel & Agol (2002). The model parameters, some of which are degenerate, are the period P, the stellar mass M_⋆, the stellar radius R_⋆, the planet radius R_p, the inclination i, the mid-transit time for each visit t₁, t₂, t₃, and the limb-darkening coefficients u₁, u₂. For an efficient MCMC calculation, we remove degeneracy between parameters by using R_⋆, r = R_p/R_⋆, and the transit duration from first to fourth contact τ₁₄, instead of R_⋆, R_p, and i. For similar reasons, we use a₁ = u₁ + 2u₂ and a₂ = 2u₁ − u₂, instead of u₁ and u₂ (Holman et al. 2006). We fix P and M_⋆, the other parameters are free. We assume zero eccentricity, as indicated by radial velocity measurements (Burke et al. 2007). Initial values are taken from Sing et al. (2011) and Fernandez et al. (2009), except for the limb-darkening coefficients. These are derived in the I, J, H, and K bands from tabulated values (Claret 2000) using the stellar mass, temperature, and metallicity found by Burke et al. (2007), and interpolated at the mean of the NICMOS band, 1.54 μm. We find u₁ = 0.054 and u₂ = 0.417.

The minimization process includes a decorrelation from instrumental parameters. The instrumental model includes a second-order polynomial in the angle α, and a linear combination of x, y, the FWHM w, and the detector temperature T. The three visits are treated separately. In addition, the filter wheel is displaced after HST re-enters the Earth's shadow leading to two distinct positions (see Figure 2). Thus, we divide the data into six groups, with two groups per visit. A multiplicative constant F₀ is used to adjust the flux level between each group. The instrumental model Ψ is computed independently for each group g, whereas the transit model Φ remains the same. The model light curve thus can be written as

$$\begin{eqnarray} F_{{\rm mod}} & = & \Phi (R_\star ,r,\tau _{14},t_1,t_2,t_3,a_1,a_2) \nonumber \\ && \times \lbrace F_{0}\times \Psi (\alpha ^2,\alpha ,x,y,w,T) \rbrace _{g \,\in \, [\![1,6 ]\!]}. \end{eqnarray} \tag{ 1 }$$

The first orbit of each visit is strongly affected by HST settling; those data are discarded. The time for each observation is the midpoint of the exposure converted to barycentric Julian date. The priors for the parameters are similar to those defined in Burke et al. (2007, 2010): equivalent to uniform on R_⋆, R_p, and i for R_⋆, r, and τ, and uniform for the other parameters. A correction for the seven readout states is applied during the process, using the known detector state at each image as calculated in Section 3.1. The flux offset generated by each state is calculated using the whole visit. The likelihood is modeled as independent Gaussian residuals with uncertainty adjusted from two trial runs to enforce a final reduced χ² of 1. The value σ = 380 ppm is adopted, and remains fixed during the MCMC run. This uncertainty is treated as white noise. We adopt the median as the best estimate of the posterior probability, and the 1σ uncertainties of the parameter estimates are determined such that they demarcate 68.3% of the MCMC samples. The first 1000 iterations are discarded as burn-in period of the MCMC procedure.

Our choice of the parameters for the instrumental model is guided by the data. In many trial parameter sets, we examine sets including F₀, x, y, α, w, T, the HST orbital phase ϕ, and quadratic and cubic powers of x, y, and α. We perform short runs of 10⁴ iterations for each set of parameter combinations. In particular, we evaluate the importance of each parameter in a simple iterative process: one parameter is added each time, and the one that yields the lower rms is retained, if the rms is lowered by at least 10 ppm (this arbitrary choice constitutes a limitation of the method). Examples of combinations are given in Table 1. This empirical approach is useful to understand the NICMOS systematics. We find that F₀ is the most important parameter. The average flux varies by up to 0.6% between groups. These variations are also present in previous studies of NICMOS data (Burke et al. 2010; Gibson et al. 2011b), and their origin is not understood. Incidentally, F₀ also accounts for stellar variability between visits. The second-most-important parameter is the angle of the spectral trace, α. In decreasing order of effect, we then find x and y, then w and T. Among higher-order terms, only α² yields a significant improvement. Including the HST orbital phase yields a negligible improvement and is not included. As seen in Figure 2, the HST orbital phase is in fact responsible for variations of the actual instrumental parameters w and T, which are already accounted for. After examining a large number of possible combinations, we adopt the parameters expressed in Equation (1).

To derive the star and planet parameters, we run five MCMC chains of 10⁵ iterations and merge their posterior probability distributions. The results are reported in Table 2. The best-fit model yields a point-to-point precision of 366 ppm (Figure 6). The Poisson noise limit is 203 ppm per exposure. The minimum χ² is 716 for 719 degrees of freedom, indicating uncertainties are well estimated.

Zoom In Zoom Out Reset image size

To identify the main sources of systematics affecting the spectrum, we first build the XO-2 Stellar spectrum in a simple way: on each image, the flux is summed over each column of the spectral trace, after subtracting the sky background in that column. The bounds in the y-direction of the spectral trace and the sky region are the same as in the white-light analysis in Section 4.1. These spectra are affected by strong variations due to the grism response in wavelength: for example, 15% between 1.54 and 1.58 μm (Figure 7). We estimate the influence of the sky background. The median sky background is 27e⁻ pixel⁻¹, which is 5.4 × 10⁻⁴ lower than the median flux pixels inside the stellar region. A determination of sky background in the whole image rather than in each column leads to no significant changes.

Zoom In Zoom Out Reset image size

A common method to extract the planetary spectrum is to divide an in-transit by an out-of-transit stellar spectrum. To empirically infer the uncertainties of this method for our data, we use all possible pairs of orbits for each star and each visit including pairs of out-of-transit orbits, but as before, discarding the first orbit of each visit. We do not use the ingress and egress. The resulting spectra have strong variations, comprised between 0.3 and 3% peak to peak (Figure 8). For comparison, the precision of the Poisson noise limited spectrum would be 370 ppm. Evidently, the large variations are due to systematic effects.

Zoom In Zoom Out Reset image size

Similar to our analysis of the white-light curve, we investigated correlations of the spectrum with instrumental parameters. We calculate the variation of a given parameter between two orbits by comparing its mean value in these orbits. We find a clear correlation with α: a larger difference in α yields a larger rms variation in the ratio of two spectra (Figure 9). The 0.6% at Δα = 06 is consistent with 10% intrapixel variations for in-focus G141 spectra (R. Bohlin 2005, private communication) diluted by the defocus of 5 pixels we intentionally introduced specifically to reduce the effect of intrapixel sensitivity variations (Lauer 1999). Angle variations and more generally position variations combined to intrapixel variations apparently are a significant source of NICMOS systematics.

Zoom In Zoom Out Reset image size

To investigate the variations in the spectra, we analyze in detail the light curves of each column. We find that the noise in the light curves is ∼1.5 times the Poisson noise within a single orbit. In contrast, global variations appear between different orbits. As an example, Figure 10 illustrates the light curves of columns 70 and 69 during the second and fifth orbits of the first visit. Those columns correspond respectively to the left and right circled points in Figure 8. A systematic flux difference between the two orbits appears clearly for column 70, and not for 69. Because this variation differs for each column, it is not due to a white-light flux variations. Moreover, in our example, the white-light flux differs by 650 ppm between both orbits, whereas the mean of the column 70 light curve differs between orbits by ∼15 times this value. This variation is thus due to more complicated systematic effects. The same behavior appears when extracting the light curves directly from the raw images, and is therefore present in the original data.

Zoom In Zoom Out Reset image size

Images do not show any particular feature that could result in a systematic flux difference. Only the level of each individual pixel differs slightly between orbits. Considering, for example, the 30th image of both orbits, this results in a flux variation of 900 ppm for column 69, and in a variation 10 times greater for column 70. These differences at the pixel level can typically be caused by intrapixel variations combined with motion of the spectral trace between orbits. We conclude that this is a main limitation of NICMOS for the extraction of planetary spectra.

Correlations with other parameters are also identified in the column light curves, as shown Figure 11. A correlation with the x position appears for most of the columns, which may be caused by the variations of the grism response in wavelength. Indeed, the observed ∼1% variations in the light curves is consistent with grism response variations up to 3% pixel⁻¹ and a measured Δx = 0.3 pixels between the first and second visits. Intrapixel variations may also play a role. We also find a frequent correlation with the FWHM. Similar correlations appear for XO-2 S. Note that correlations such as shown in Figure 11 may be the result of multi-dimensional correlations projected onto a single-parameter axis. Removing one-dimensional correlations by polynomial fit reduces the noise in the spectra to ∼0.1% (1000 ppm). Another factor of ∼10 is necessary to extract the planetary signal. A more sophisticated decorrelation method is required, as described in the next section.

Zoom In Zoom Out Reset image size

The apparent nonlinear variations may be the result of a multi-parameter dependence projected on a single-parameter axis, and we keep the parameters defined in Section 4.2 for the decorrelation in Section 5.2.

A Monte Carlo analysis is performed to derive the spectrum of XO-2b. The columns are binned by 5 (the size of the FWHM), giving 22 independent channels along the spectral trace. The corresponding spectral resolution is R = 37. To aid comparison with other NICMOS analysis (Swain et al. 2009a; Tinetti et al. 2010), each bin is then shifted by 0, 1, 2, 3, and 4 pixels, leading to as many points as the number of columns; the analysis is performed independently for each of the 22 × 5 channels. An alternative approach to MCMC would be to use Gaussian processes to account for systematics, as done by Gibson et al. (2012a, 2012b).

A light curve is extracted for each channel using a box of 5 pixel wide in x and bounds in y that were found optimal in Section 4.1. We exclude the first orbit of each visit and individual points departing by more than 3σ from the mean flux of the orbit. These light curves are then fitted with a Monte Carlo procedure similarly to Section 4.2. The parameters P, M_⋆, R_⋆, i, and t_i are fixed to their value in Table 2. The limb-darkening coefficients u₁ and u₂ are fixed for each channel at values interpolated from broadband I, J, H, and K coefficients appropriate for XO-2 (Claret 2000). Among physical parameters, only the planetary radius R_p is a variable; its initial value is that of Table 2. Similarly to Section 4.2, the instrumental model is a linear combination of α², α, x, y, w, and T, and F₀. Each detector column and therefore each spectral channel has its own dependence to instrumental parameters (Figure 11). The instrumental model is thus computed independently for each channel, as well as for the six groups of data:

$$\begin{equation} F_{{\rm mod}}=\Phi (r)\times \lbrace F_0\times \Psi (\alpha ^2,\alpha ,x,y,w,T) \rbrace _{g \,\in \, [\![1,6]\!] .} \end{equation} \tag{ 2 }$$

A spectrum is built for each visit (Figure 12). We restrict the range to 1.2–1.8 μm. In this range, we obtain a standard deviation σ of 430, 510, and 1000 ppm for each spectrum for the first, second, and third visits, respectively. The approximately twice higher noise for the third visit than for the other two visits can be explained by larger instrumental parameter variations within orbits, in particular for y and α (Figure 2). As a consequence, we ignore the third visit for the remainder of this analysis. We focus on the first and second visits. The mean absolute difference between the first and second visit individual spectra is 680 ppm. A joint analysis of these two visits is then performed. We run the MCMC using the data of both visits, with a single set of transit parameters. Again, the instrumental parameters are optimized for each group of data. This yields a spectrum with σ = 290 ppm (Figure 12). For comparison, the expected Poisson noise per 5 column channel is 101 ppm for the combination of all observations. The corresponding channel light curves for XO-2 N before and after the decorrelation process are shown in Figure13.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

For context, we perform a similar analysis using XO-2 S, a nearly identical star. We extract the channel light curves, and numerically inject a simulated transit with a depth constant at all wavelengths. We use the same transit parameters as for XO-2 N. The extracted spectra are shown in Figure 14. Whereas the extracted simulated planetary spectrum should be flat, we find similar variations as we did for XO-2 N: σ = 370, 500, and 1480 ppm for the first, second, and third visits, respectively, σ = 250 ppm for the joint analysis of the first and second visits, and a mean absolute difference of 610 ppm between the first and second visit spectra. Therefore, these levels represent the level of remaining NICMOS systematics for these data.

Zoom In Zoom Out Reset image size

Uniquely for XO-2, we can derive a planetary spectrum for XO-2b using XO-2 S as a reference star for XO-2 N. We divide the spectrum of XO-2 N by that of XO-2 S using the joint analysis of the first and second visits. The result is a normalized spectrum shown in Figure 15. The corresponding transit depths are reported in Table 3. The standard deviation of the normalized spectrum is 3.74%, equivalent to σ = 400 ppm in transit depth. The expected Poisson noise is 142 ppm. Again, using the first, the second, or both visits leads to planetary spectra with features that change significantly more than that expected for random noise. Apparently, some instrumental systematics of NICMOS are not common mode and therefore do not cancel. Otherwise, the 1σ uncertainty of the spectrum in Table 3 would have more closely matched the Poisson noise.

Zoom In Zoom Out Reset image size

Although systematics are clearly remaining, we compare for context the normalized XO-2b spectrum (obtained by dividing the spectrum of XO-2 N by that of XO-2 S after numerical injection of a flat transit, using the joint analysis of the first and second visits) to a normalized synthetic spectrum of the irradiated giant planet HD209458b, from Burrows et al. (2010). In that model, spectral features in the range 1.2–1.8 μm are largely due to water vapor absorption (see also, e.g., Brown 2001, for the identification of spectral features). In the following, we designate the Burrows et al. (2010) model as water model.

In the range 1.2–1.8 μm, the mean absolute difference between the data and the water model is 2.70%, corresponding to 290 ppm in transit depth. The comparison of the data to the water model yields χ² = 10.4 for N = 15 degrees of freedom. A comparison of the data to a flat spectrum model (with a constant level adjusted to yield the minimum χ²) yields χ² = 13.5 for N = 14 degrees of freedom. Both models are therefore acceptable fits to the data. A broad feature is identified around 1.4 μm, corresponding to an expected water feature. To infer the amplitude of the observed feature, we fit the water model to the data over the range 1.2–1.8 μm varying the model's amplitude and reference level. The amplitude is calculated as the difference between the minimum and maximum model averaged points, at 1.245 and 1.405 μm, respectively, which corresponds to the amplitude at 1.4 μm. A χ² minimization yields a best-fit amplitude of 580 ± 330 ppm. This level is in good agreement with the theoretical value of 650 ppm albeit with a large uncertainty, and is comparable to those reported for HD 189733b (Swain et al. 2008) and XO-1b (Tinetti et al. 2010). However, this is also comparable to the level of remaining systematics in our data, in particular to the mean absolute difference between the first and second visit individual spectra. Therefore, this feature may simply be due to systematics. We do not attempt to fit models with additional molecular species. Interpreted as a 3σ upper limit on the presence of water vapor absorption in the atmosphere of XO-2b, excess absorption at 1.4 μm is less than 1570 ppm. Note this simple χ² analysis is limited by the fact that residuals may not be white noise.

Models predict water vapor in the atmosphere of hot Jupiters as a major absorber in the near-infrared (Fortney et al. 2008, 2010; Burrows et al. 2010). In this domain, pM and pL class planets exhibit the same spectral features, with a larger amplitude for the pM class due to a greater atmospheric scale height. Fortney et al. (2010) calculated models of transmission spectrum as a function of the planet temperature for two surface gravities. With surface gravity g = 14.71 m s⁻² (Fernandez et al. 2009) and dayside temperature T = 1500 K (Machalek et al. 2009), these models predict a water vapor absorption feature at 1.4 μm that would increase the apparent planetary radius by 3%, i.e., the transit depth by 600 ppm. A similar feature is predicted by Burrows et al. (2010), which we show in Figure 15. Our XO-2b spectrum shows an increase in absorption of 580 ± 330 ppm, in good agreement with these predictions, albeit with a large uncertainty. Given the level of remaining systematics in our data, this feature may however be spurious. Our 3σ upper limit on the 1.4 μm molecular absorption, 1570 ppm, is consistent with these models.

Transit spectroscopy is at the limit of NICMOS capabilities. With the simplest data analysis, i.e., without a decorrelation process or alternative approach, the derived planetary spectrum of XO-2b (Figure 8) exhibits noise-like features that are approximately one order of magnitude greater than signals anticipated from a planetary atmosphere. The dominant noise source is due to the combination of the different positioning of the spectral trace on the detector, notably the angle α of the spectral trace, combined with intrapixel variations. Previous NICMOS measurements presumably are affected similarly as well.

In this work, we decorrelated the NICMOS data for XO-2b with the MCMC technique using a quadratic function for the spectral tilt angle α and linear functions for the other parameters. Previous studies (e.g., Tinetti et al. 2010) have performed similar parametric analyses, and others have used alternative methods such as Gaussian processes (e.g., Gibson et al. 2012a, 2012b). For the NICMOS data of XO-2b, the availability of a nearly identical companion star (XO-2 S) observed simultaneously provides a unique opportunity for data analysis and calibration. The presence of this bright, nearly identical reference star could be used by other investigators as a benchmark to test their NICMOS light-curve and spectral extraction methods.

To use the companion star XO-2 S in our analysis, we numerically injected into the data of XO-2 S a transit with a depth constant in wavelength, and then derived a spectrum with the same procedure as for the planet host star, XO-2 N. The derived simulated spectrum (of XO-2 S) has residuals comparable to those evident in the planetary spectrum derived for XO-2 N, but which can only be due to instrumental effects or effects of the analysis because we injected a flat, featureless spectrum. In addition, the residuals exhibited in the synthetic spectrum (of XO-2 S) differ for each visit: they have a typical scale of 0.05 μm in wavelength and 500 ppm in amplitude. Therefore, features of that scale derived by our technique cannot be confidently ascribed to molecular signatures in XO-2b. We note that the spectrum derived for XO-1b in the Appendix exhibits similar residuals, which also differ for the two visits. In both cases, XO-2b and XO-1b, a broader 0.2 μm wide feature appears at 1.4 μm, where extant models predict extra absorption due to water vapor. However, the observed feature has an amplitude comparable to the level of systematics and is thus of weak statistical significance.

Our MCMC decorrelation process may not correct for all systematics. In particular, Pont et al. (2009) and Gibson et al. (2011b) suggest that the in-transit instrumental parameters should be completely sampled during the out-of-transit orbits. This is unfortunately not the case for the XO-2 data, in particular for the angle of the spectral trace. Whereas the need for sophisticated analysis techniques is clear, different analysis of previous NICMOS data lead to different results (Swain et al. 2008; Tinetti et al. 2010; Gibson et al. 2011b). This supports our remark that extracting planetary spectra is at the limit of NICMOS capabilities.

Whether the above conclusions apply only to those data that we analyzed (XO-1 and XO-2), and/or only for our analysis technique, or apply more generally to all NICMOS spectra remains an open question, one that is beyond the scope of this analysis. A program to analyze all relevant NICMOS data has been approved under HST archival program 12844, "Exoplanetary Spectroscopy with NICMOS Revisited" (PI: Deming).

Planetary spectra have been observed with HST's new Wide Field Camera 3 (WFC3), which has much smaller, indeed as yet unmeasurable, intrapixel sensitivity variations than NICMOS, for which they are an important effect (Pavlovsky et al. 2011). Using WFC3, Berta et al. (2012) reported a result for the planet orbiting the H = 9.1 magnitude M dwarf GJ1214, and showed that the results are nearly Poisson-noise limited after a few corrections are carefully applied to the WFC3 data.