StarCAT: A CATALOG OF SPACE TELESCOPE IMAGING SPECTROGRAPH ULTRAVIOLET ECHELLE SPECTRA OF STARS

The pipeline processing is deficient in one important, although subtle, respect: the low order polynomials describing the mapping from pixel coordinates in the raw MAMA frames to wavelengths in the extracted spectra (the "dispersion relations") do not fully capture all the persistent distortions present in the various echelle settings (Ayres 2008). This can be an issue when merging overlapping segments of adjacent orders, because inconsistencies in the assigned wavelengths can lead to imperfect alignment between spectral features in the overlap zones, blurring the resolution.

The "Deep Lamp Project" (Ayres 2008)—which analyzed long exposures of the STIS wavelength calibration sources—provided a solution to this otherwise vexing problem: a post-facto correction that could be applied to the ${\sf x1d}$ pipeline files. As part of the StarCAT effort, empirical distortion models were developed for all 44 supported STIS echelle mode settings, along the lines of the original Deep Lamp study. Figure 2 schematically depicts the resulting maps grouped in the four uber modes. For some of the little used secondary tilts, the pipeline dispersion relations introduced a more-or-less global displacement, reaching as much as 2 km s⁻¹ in equivalent velocity units in one (extreme) example (E230M-2269, where the shift was ∼0.4 pixel).

Zoom In Zoom Out Reset image size

The first step of the ${\sf x1d}$ "initialization" consisted of applying the distortion correction. Like the ${\sf calstis}$ dispersion relations, the two-dimensional corrections were expressed in polynomial expansions of two variables, including cross terms; but utilizing higher terms where appropriate, and fully orthogonal independent variables (see discussion by Ayres 2008). In practice, the distortions were specified as equivalent velocities, υ, as a function of the grating parameter k (≡m × λ) and order number m. The k parameter varies exclusively along the order (image x-axis), whereas m varies exclusively (and trivially) with the orders (image y-axis), thus providing the desirable orthogonality of the independent variables in the bi-polynomial expansions (i.e., containing terms like k, m, k m, k² m, k m², and so forth). This approach was taken—instead of utilizing, say, Legendre polynomials—to maintain heritage with the dispersion relations embedded in the ${\sf calstis}$ pipeline, which in turn can be traced back to the IUE Spectral Imaging Processing System. The number of terms in each model (up to fifth order) was determined by trial and error to minimize the global χ² with the fewest coefficients. The variables k and m were transformed to new variables $\tilde{k}\equiv (m\,\lambda - k_{0})/500$ and $\tilde{m}\equiv (m - m_{0})/10$, where k₀ depends on the uber mode, and m₀ depends on the setting. The transformed variables are order unity over the range of each echelle stripe, to better condition the numerical solutions. The specific models are summarized in Table 3.

Table 3. Wavelength Distortion Correction

Polynomial Model
E140M-1425: $\tilde{m}\equiv \frac{(m - 106)}{10}\,; \tilde{k}\equiv \frac{(m\,\lambda - 147950)}{500}$
$\upsilon =(0.2168845) + (0.5295246)\times \,\tilde{m} + (-0.1133152)\times \,\tilde{m}^2 + (-0.6151966)\times \,\tilde{m}^3 + (0.0123862)\times \,\tilde{m}^4$
$\quad +\, (0.0813750)\times \,\tilde{m}^5 + (0.7533000)\times \,\tilde{k} + (-0.1167010)\times \,\tilde{k}^2 + (-1.1997739)\times \,\tilde{k}^3 + (0.1042297)\times \,\tilde{k}^4$
$\quad+\, (0.2603576)\times \,\tilde{k}^5 + (-0.7276294)\times \,\tilde{m}\,\tilde{k} + (0.3840922)\times \,\tilde{m}^2\,\tilde{k} + (0.1872590)\times \,\tilde{m}^3\,\tilde{k} + (-0.1337153)\times \,\tilde{m}^4\,\tilde{k}$
$\quad+\, (0.0018007)\times \,\tilde{m}^5\,\tilde{k} + (-0.1091477)\times \,\tilde{m}\,\tilde{k}^2 + (-0.2054250)\times \,\tilde{m}^2\,\tilde{k}^2 + (0.1786350)\times \,\tilde{m}^3\,\tilde{k}^2 + (0.0175580)\times \,\tilde{m}^4\,\tilde{k}^2$
$\quad+\, (-0.0321286)\times \,\tilde{m}^5\,\tilde{k}^2 + (0.7644269)\times \,\tilde{m}\,\tilde{k}^3 + (-0.7003058)\times \,\tilde{m}^2\,\tilde{k}^3 + (-0.3217466)\times \,\tilde{m}^3\,\tilde{k}^3 + (0.2486133)\times \,\tilde{m}^4\,\tilde{k}^3$
$\quad+\, (0.0498403)\times \,\tilde{m}^5\,\tilde{k}^3 + (0.1559506)\times \,\tilde{m}\,\tilde{k}^4 + (0.1556453)\times \,\tilde{m}^2\,\tilde{k}^4 + (-0.0563305)\times \,\tilde{m}^3\,\tilde{k}^4 + (-0.0372428)\times \,\tilde{m}^4\,\tilde{k}^4$
$\quad+\, (-0.0015327)\times \,\tilde{m}^5\,\tilde{k}^4 + (-0.1468092)\times \,\tilde{m}\,\tilde{k}^5 + (0.4034192)\times \,\tilde{m}^2\,\tilde{k}^5 + (0.0330954)\times \,\tilde{m}^3\,\tilde{k}^5 + (-0.1312219)\times \,\tilde{m}^4\,\tilde{k}^5$
$\quad+\, (-0.0049503)\times \,\tilde{m}^5\,\tilde{k}^5$
E140H-1234: $\tilde{m}\equiv \frac{(m - 339)}{10}\,; \tilde{k}\equiv \frac{(m\,\lambda - 421000)}{500}$
$\upsilon =(0.0301367) + (0.0382590)\times \,\tilde{m} + (-0.0383733)\times \,\tilde{m}^2 + (-0.0418539)\times \,\tilde{m}^3 + (0.0979899)\times \,\tilde{k}$
$\quad+\, (0.0444978)\times \,\tilde{k}^2 + (-0.1219705)\times \,\tilde{k}^3 + (-0.0260690)\times \,\tilde{m}\,\tilde{k} + (0.0959600)\times \,\tilde{m}^2\,\tilde{k} + (0.0569775)\times \,\tilde{m}^3\,\tilde{k}$
$\quad+\, (0.0352926)\times \,\tilde{m}\,\tilde{k}^2 + (-0.0193594)\times \,\tilde{m}^2\,\tilde{k}^2 + (-0.0123493)\times \,\tilde{m}^3\,\tilde{k}^2 + (-0.0278292)\times \,\tilde{m}\,\tilde{k}^3 + (0.0192519)\times \,\tilde{m}^2\,\tilde{k}^3$
$\quad+\, (0.0079272)\times \,\tilde{m}^3\,\tilde{k}^3$
E230M-2269: $\tilde{m}\equiv \frac{(m - 94)}{10}\,; \tilde{k}\equiv \frac{(m\,\lambda - 204100)}{500}$
$\upsilon =(-1.1279973) + (0.7839361)\times \,\tilde{m} + (-0.3184260)\times \,\tilde{m}^2 + (-0.5095827)\times \,\tilde{m}^3 + (0.3729259)\times \,\tilde{k}$
$\quad+\, (-0.0702246)\times \,\tilde{k}^2 + (-0.0920298)\times \,\tilde{k}^3 + (-0.3107360)\times \,\tilde{m}\,\tilde{k} + (-0.0233978)\times \,\tilde{m}^2\,\tilde{k} + (0.1220877)\times \,\tilde{m}^3\,\tilde{k}$
$\quad+\, (-0.1275635)\times \,\tilde{m}\,\tilde{k}^2 + (0.0957628)\times \,\tilde{m}^2\,\tilde{k}^2 + (0.0868626)\times \,\tilde{m}^3\,\tilde{k}^2 + (0.0405377)\times \,\tilde{m}\,\tilde{k}^3 + (0.0244841)\times \,\tilde{m}^2\,\tilde{k}^3$
$\quad+\, (-0.0092254)\times \,\tilde{m}^3\,\tilde{k}^3$
E230H-2513: $\tilde{m}\equiv \frac{(m - 308)}{10}\,; \tilde{k}\equiv \frac{(m\,\lambda - 772300)}{500}$
$\upsilon =(-0.0738475) + (0.0954277)\times \,\tilde{m} + (-0.0416483)\times \,\tilde{k} + (0.0130481)\times \,\tilde{m}\,\tilde{k}$
etc.

Notes. Complete table available at http://archive.stsci.edu/prepds/starcat.

Download table as:  ASCIITypeset image

The velocity corrections were applied to the individual echelle orders according to

$$\begin{equation} \lambda _{\rm new}= \lambda _{\rm old}\, - \, (\upsilon [k,m]\, /\, c)\times \lambda _{\rm old}. \end{equation} \tag{ 1 }$$

It would be preferable, of course, to have a more accurate dispersion model implemented in ${\sf calstis}$ itself, to avoid the distortion correction step altogether. But, in the interests of producing StarCAT expeditiously in the face of the development effort required to modify the pipeline, it was decided to proceed with the post-facto approach.

Following the distortion correction, the parameters of each ${\sf x1d}$ order were linearly interpolated onto a finer wavelength grid with twice the pixel density delivered by the pipeline, to achieve four points per resol. The (trivial) double-density resampling was helpful in the subsequent concatenation step (and in the various co-addition stages downstream), for which nontrivial interpolations were required to match the wavelengths of one order to those of the next one up: interpolating sub-Nyquist sampled spectra can lead to undesirable smoothing.

At this stage, a slightly modified photometric error was introduced. The pipeline calculates a 1σ value according to the square root of the gross number of counts in the wavelength bin, N (e.g., as reported in the pipeline array "${\sf GROSS}$"⁹), divided by exposure time and sensitivity. However, a careful analysis of Poisson statistics in the low count regime shows that the 1σ error is better represented by $\sqrt{N\,+\,1}$ rather than $\sqrt{N}$.¹⁰ The modified photometric error was calculated according to the $\sqrt{N\,+\,1}$ model.

Ordinarily, one would utilize the ${\sf GROSS}$ array mentioned above. It was found, however, that occasionally, and unaccountably, the pipeline ${\sf GROSS}$ was offset by one entire order, rendering the N values unusable for the local photometric uncertainty calculation. Instead, the appropriate N was derived directly from the pipeline photometric error (which apparently was calculated properly even when there was the Δm = 1 shift of the ${\sf GROSS}$ counts), by multiplying by the sensitivity (s_λ: counts s⁻¹ per unit flux density) and exposure time, then squaring the result. The 1σ limit in flux density was recovered by adding 1 to the derived N (which would be the sum of two derived values in regions of order overlap), taking the square root, and dividing by exposure time and sensitivity (sum of s_λ for overlap zones). In practice, the difference between the modified and pipeline σ_λ was minor, except in regions of very low count rate. A benefit of the modified σ_λ was that the photometric uncertainty always was greater than zero, and thus an S/N calculation would never diverge. In the double density interpolation step, σ_λ was boosted by $\sqrt{2}$ to preserve the average (in quadrature) over the original flux cell.

One consideration to keep in mind, especially when examining the spectral charts presented later, is that the calibrated σ_λ trace can take on a distinct scalloped appearance, which devolves from the rapidly varying sensitivity across each echelle order. For example, if the source f_λ is constant with wavelength, the detected counts will exhibit a bell-shaped distribution across an echelle order, which then will be flattened after application of the inverse sensitivity function, s⁻¹_λ. However, the photometric error, $\sqrt{N\,+\,1}$, would exhibit a shallower profile, especially for small N, which then would take on a "horned" shape after application of s⁻¹_λ: higher at the edges of the order and lower at the center. This can lead to a scalloped appearance of the calibrated σ_λ in an order-concatenated spectrum.

During verification of the modified σ_λ it was noticed that occasionally a pair of sharp spikes would appear at the short-wavelength end of E140M (near 1170 Å, as illustrated later in Figure 8(a)). Often, the associated flux densities would exhibit a strong negative dip, but sometimes a distinct peak. This is the signature of a detector bright spot that most of the time falls in the interorder background, but occasionally inside an order (owing to the slight randomness in the MSM y positioning). Examination of representative E140M raw images, including many wavecal spectra from the previously mentioned Deep Lamp Project, revealed that the bright spot was highly intermittent: sometimes at a level of a few hundred counts, but more often entirely absent. (Coincidentally, the spot was conspicuous in the summed AD Leo E140M exposures of Figure 1(a): it falls in the extreme lower right of the frame, next to the C iii 1175 Å blend, and clearly is a detector artifact because a true spectral feature would be replicated in the next order up.)

A check of E140H raw images, taken in three different wavelength settings, also showed the bright spot at the same location on the FUV MAMA, although again its appearance was highly intermittent. In general, the feature was not flagged by the pipeline, although in a few instances it apparently was assigned a data quality flag of 16, indicating an anomalously high dark current. At worst, the blemish appears as a 5 × 2 pixel patch, with up to a few hundred counts. (In the case of AD Leo, the peak was about 40 counts in each 2.2–2.7 ks exposure). Because the feature is not flagged, it can survive the co-addition process. However, it was decided not to attempt a correction at this time, especially since the issue resides in the pipeline, it affects only a very small number of spectral bins, and is very intermittent in the first place. One should be wary, nevertheless, of the 1170 Å region of E140M spectra: single or paired spikes in the local σ_λ probably signal the presence of the defect. The short-wavelength ends of E140H spectra could be affected similarly. This is another example of a potential area where the pipeline could be improved.

The data quality flag, q_λ, plays a vital role in marking valid fluxes for interorder merging, and in subsequent co-addition stages involving independent exposures. Although the flag should not, strictly speaking, be interpreted as a number, roughly speaking, values larger than 500 indicate more serious problems.¹¹ Of the 1.3 × 10⁸ points in the dearchived ${\sf x1d}$ files, 95% had quality values of 0, 2% were flagged in the range 1–256, and 3% in the range ⩾512 (up to 3620, the top value encountered). The most common flag in the middle range was 16, indicating an anomalously high dark rate for the bin. Examination of representative ${\sf x1d}$ files revealed that spectral points assigned flags in the lower range did not appear to be unusual with respect to their q_λ = 0 neighbors (except for the one possible conspicuous exception of the FUV bright spot described in the previous section), whereas those in the upper range (q_λ>500) were more clearly deviant. Accordingly, q_λ < 500 was chosen as the cutoff for valid fluxes. (The specific value q_λ = 500 was reserved to mark wavelength gaps that sometimes occurred between the ${\sf x1d}$ orders.)

To avoid any edge effects (caused by the detector format or the pipeline modeling of, say, the scattered light correction), a set number of points were clipped at the beginning and end of each order prior to concatenating. For the FUV modes, both trim values were set to 40 points uniformly for all m (the full span of each resampled ${\sf x1d}$ order now was 2048 bins). For the NUV modes, it was found occasionally that the lowest few orders (longest wavelengths) would be affected by unflagged "dropouts" at the short-wavelength edge of the stripe (similar gaps in the higher orders invariably would be flagged correctly). These unmarked dropouts would be treated as valid fluxes and averaged with the good points at the end of the preceding order, leading to false depressions. While it is easy to spot such defects visually, it was not so easy to design a purely automated procedure to recognize them in an arbitrary echellegram (because they occurred without an obvious pattern). For this reason, all of the NUV settings were treated as if they were affected by the dropouts, and additional points were clipped from the short-wavelength edges of the lowest orders (300 points for the lowest m; 240 for the second and third lowest; 160 for the fourth lowest; and the nominal 40 for all the rest). The trim value for the high side of the orders was set to 40 uniformly, same as for FUV.

The expanded dropout exclusion zones were large enough to avoid the worst cases identified in an examination of random NUV exposures, but not so large as to significantly diminish the quality of spectra not affected by unflagged gaps (by eliminating otherwise good fluxes). The strategy resulted in a very substantial decrease in the number of defects in the final spectra, although very occasionally one might still find a small sharp unflagged dropout in one spectrum of an overlapping pair, made obvious during the splicing step described below, where ordinarily there would be excellent point-by-point agreement between the independent flux traces. Better dropout flagging, particularly in NUV, is another example of where the pipeline could be improved.

The conservative order clipping was dictated by the desire to automate the concatenation procedure as much as possible, and avoid the more custom hands-on approach that in principle could be applied to a smaller collection of echellegrams. In any event, less than 5% of the data points were removed in the process. Many of these (at the extremes of the orders) already had been flagged by the pipeline (a fair fraction of the 3% bad points noted earlier) and most represent redundant wavelengths, so information loss was minimal. However, a consequence of the trimming is slightly less overlap than would be expected in the low orders, so there are a few more, and somewhat larger, interorder gaps at the longward ends of some of the tilts. This is particularly true for E140M, where small regularly spaced gaps begin to appear at around 1550 Å, and become progressively wider toward longer wavelengths.

In the process of tracking down sources of the anomalous gaps in some E230 frames, it was noticed that the E230H-2313 setting, and to a lesser extent the neighboring tilt E230H-2363, was affected by a systematic roll-off of the fluxes on the short-wavelength side of each order in the middle range of m's for that setting. The roll-off was not flagged as unusual. Consequently, when the overlap zones were concatenated, the roll-off part would be averaged with the normal fluxes from the end of the previous order, producing a sharp-edged dip with a gradual curve back to the normal flux level on the long-wavelength side. Since this was a ${\sf calstis}$ calibration issue (probably a bad value in one of the echelle ripple parameters) and there was no easy fix, all the E230H-2313 exposures (thankfully, there were only two) were simply excluded from the subsequent processing stages. The E230H-2363 tilt appeared to be mildly affected by this problem, but not so noticeably to justify exclusion of these (more numerous) data sets. Both E230H-2313 and E230H-2363 are little-used secondary settings. The closest prime tilt, E230H-2263, appeared to be perfectly normal.

Following the initialization steps described above, the multiple echelle orders of the ${\sf x1d}$ file were concatenated. This was accomplished in ascending wavelength, by linearly interpolating f_λ and σ_λ at the end of order m onto the wavelength scale of the overlapping zone at the beginning of the next order, m − 1; then averaging, weighting by the relative sensitivity functions, s_λ. The latter was obtained for each order by dividing the trace of net count rate (counts s⁻¹) by the flux density. The sensitivity peaks at the order center, and tapers off to either side. Weighting by (normalized) s_λ is equivalent to adding the net counts in each bin, then dividing by the combined sensitivity and the exposure time. (The s_λ profile also was used in the recalculation of the photometric error, as described earlier.) The data quality flag was "interpolated" according to a nearest neighbor function.¹²

In combining the fluxes from overlapping zones, if both bins were flagged "good" (q_λ < 500, as noted above), the data quality flag of the merged point was set to the higher of the two q_λ. Because most of the points have a zero value for the flag, this strategy should effectively capture any potential issues with merged points devolving from a specific flag of one of the pair ostensibly in the good range, but nevertheless >0. If one or both points was identified as bad (q_λ>500), the value with the lowest q_λ was retained. If a wavelength gap was encountered, additional closely spaced bins with q_λ = 500 and f_λ =  σ_λ = 0 were interjected at both edges to ensure that the blank interval would be properly treated in later interpolations. (The same hierarchy was followed in subsequent co-addition and/or splicing steps.)

The root name of the pipeline data set (e.g., "o61s01010") was retained for the output file, appended with "_n" (n = 1, 2, ...) to identify sub-exposures (multiple repeats within the same observation "wrapper"). For example, the merged save set in the preceding case, a single exposure, would be called "o61s01010_1." The entire collection of ${\sf x1d}$ data sets was processed automatically, and the resulting files then were copied to the appropriate target directory (identical to the consensus object name) for subsequent processing: in this example, as "AD-LEO/o61s01010_1." The full list of STIS exposures can be found in Table 4, explicitly including the sub-exposures.

Table 4. Stage Zero Co-additions: Sub-exposures

Data set	Mode/Setting	Aperture	Band	texp	Modified JD	Qual.	Xcorr Params.
		(001 × 001)	(Å)	(s)	(d)	(%)	(Å or km s−1)
(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
AD-LEO
o61s01010_1	E140M-1425	020 × 020	1140-1729	2200	51613.145	3
o61s01020_1	E140M-1425	020 × 020	1140-1729	2700	51613.205	4
o61s01030_1	E140M-1425	020 × 020	1140-1729	2700	51613.272	4
o61s01040_1	E140M-1425	020 × 020	1140-1729	2700	51613.339	3
o61s01050_1	E140M-1425	020 × 020	1140-1729	2700	51613.406	4
o61s02010_1	E140M-1425	020 × 020	1140-1729	2200	51614.151	3
o61s02020_1	E140M-1425	020 × 020	1140-1729	2700	51614.211	6
o61s02030_1	E140M-1425	020 × 020	1140-1729	2700	51614.278	4
o61s02040_1	E140M-1425	020 × 020	1140-1729	2700	51614.345	4
o61s02050_1	E140M-1425	020 × 020	1140-1729	2700	51614.412	5
o61s03010_1	E140M-1425	020 × 020	1140-1729	2200	51615.090	3
o61s03020_1	E140M-1425	020 × 020	1140-1729	2700	51615.150	4
o61s03030_1	E140M-1425	020 × 020	1140-1729	2700	51615.217	5
o61s03040_1	E140M-1425	020 × 020	1140-1729	2700	51615.284	4
o61s03050_1	E140M-1425	020 × 020	1140-1729	2700	51615.351	4
o61s04010_1	E140M-1425	020 × 020	1140-1729	2200	51616.096	6
o61s04020_1	E140M-1425	020 × 020	1140-1729	2700	51616.156	3
o61s04030_1	E140M-1425	020 × 020	1140-1729	2700	51616.223	3
o61s04040_1	E140M-1425	020 × 020	1140-1729	2700	51616.290	3
o61s04050_1	E140M-1425	020 × 020	1140-1729	2700	51616.357	4
o6jg01010_1	E140M-1425	020 × 020	1140-1709	1750	52426.297	3
o6jg01020_1	E140M-1425	020 × 020	1140-1709	2880	52426.350	4
o6jg01030_1	E140M-1425	020 × 020	1140-1709	2880	52426.416	4
o6jg01040_1	E140M-1425	020 × 020	1140-1709	2880	52426.483	4
o6jg02010_1	E140M-1425	020 × 020	1140-1709	1750	52427.233	3
o6jg02020_1	E140M-1425	020 × 020	1140-1709	2880	52427.285	4
AG-DRA
o6ky01010_1	E140M-1425	020 × 020	1140-1729	2600	52755.815	93
o6ky01020_1	E140M-1425	020 × 020	1140-1729	1837	52755.848	91
o6ky01040_1	E230M-1978	020 × 020	1607-2365	241	52755.881	54
o6ky01050_1	E230M-1978	020 × 020	1607-2365	1872	52755.886	82
o6ky01030_1	E230M-2707	020 × 020	2275-3118	374	52755.873	99
AO-PSC
o50151010_1	E140M-1425	020 × 020	1140-1709	2140	51740.494	90
o50151020_1	E140M-1425	020 × 020	1140-1709	2540	51740.552	92
o50151030_1	E140M-1425	020 × 020	1140-1709	2680	51740.619	91
o50151040_1	E140M-1425	020 × 020	1140-1709	2270	51740.686	90
o50152010_1	E140M-1425	020 × 020	1140-1709	2140	51741.432	90
o50152020_1	E140M-1425	020 × 020	1140-1709	2540	51741.491	91
o50152030_1	E140M-1425	020 × 020	1140-1709	2680	51741.558	91
o50152040_1	E140M-1425	020 × 020	1140-1709	2270	51741.625	91
BG-PHE
o63502010_1	E140M-1425	020 × 020	1140-1709	1440	52041.327	97
BW-SCL
o5b616010_1	E140M-1425	020 × 020	1140-1729	1977	51433.056	2
CI-CAM
o5ci01010_1	E140M-1425	020 × 020	1140-1729	2242	51623.065	11
o5ci01020_1	E140M-1425	020 × 020	1140-1729	1350	51623.120	2	+0.00
o5ci01020_2	E140M-1425	020 × 020	1140-1729	1540	51623.138	3	+0.00
o5ci01020				2890	51623.120	11	...
o5ci01030_1	E140M-1425	020 × 020	1140-1729	1350	51623.187	1	+0.00
o5ci01030_2	E140M-1425	020 × 020	1140-1729	1540	51623.206	2	+0.00
o5ci01030				2890	51623.187	9	...
o5ci01040_1	E230M-2269	020 × 020	1857-2672	1350	51623.254	35	+0.00
o5ci01040_2	E230M-2269	020 × 020	1857-2672	1646	51623.271	37	+0.68
o5ci01040				2996	51623.254	41	2328.48 [0.38]
etc.

Notes. Objects now are listed alphabetically. This is a complete list of STIS echelle exposures incorporated in StarCAT. Multiple sub-exposures, if any, are explicitly identified (by x1d.fits extension number) and grouped together, followed by the name of the co-added spectrum (and its properties). "Modified JD" in Column 6 is JD−2,400,000. Quality factor ("Qual.") in Column 7 is percentage of valid flux points ⩾2.5σ with respect to local photometric error (per resol): fewer than 1% should exceed this significance level by chance (one-sided Gaussian σ). In Column 8, values for the sub-exposures are derived cross-correlation shifts, expressed in equivalent velocities (km s⁻¹): leading spectrum has υ ≡ 0 by definition. For the co-added spectrum, the first value in Column 8, if any, is wavelength (Å) of cross-correlation template feature, followed (parenthetically) by half-width of correlation band (also in Å). If template parameters are not reported, a blind co-addition was performed (all shifts set to zero).(Complete table available at http://archive.stsci.edu/prepds/starcat.)

Download table as:  ASCIITypeset image

When combining the sub-exposures, a new issue came into play. Namely, slight pointing drifts or spacecraft jitter within a visit can shift the apparent radial velocity of each spectrum of a sequence, no matter how well the occasional routine wavecals are able to track the instrumental zero point (which also can change over time, owing to thermal effects). For example, the 02 × 02 aperture ("020 × 020," as described earlier), commonly used to achieve the best photometric accuracy, is about 7 pixels wide for E140M-1425, or around 22 km s⁻¹ in equivalent velocity units. Thus, a slight target drift off-center by only a few tenths of the aperture width could produce a spurious velocity shift of several km s⁻¹ (which incidentally is larger than typical accuracies of bright star radial velocities). Blindly, co-adding a sequence of exposures containing random shifts of this order could lead to undesirable spectral blurring.

To counter that possibility, the individual spectra were aligned by cross-correlation, if there was enough contrast and S/N in a spectral feature, or features, within the grasp of the particular setting. Narrow emission lines proved to be good markers in cool star spectra, as were sharp interstellar absorptions in (the more numerous) hot stars. For late-type chromospheric emission-line objects, atomic lines like O i λ1306 were preferred over highly ionized species such as C ivλ1548, even though the latter might be brighter, because the higher temperature features are known to display more short term stochastic velocity variability (say, due to microflaring) than their lower temperature cousins. Such considerations do not apply to the same degree, of course, to the interstellar absorptions in hot-star spectra.

In practice, the leading subexposure of a sequence was adopted as a reference (if a suitable template feature were available), and all the others were cross-correlated against it to establish a set of relative velocity shifts. The initial exposure of a sequence likely will have the best target centering, because it will be closest in time to an acquisition (and peak-up, if any). The remaining observations therefore should be slaved to the first to preserve "velocity memory."

Because of the excellent internal wavelength precision of the distortion-corrected echellegram, the cross-correlation can be performed on a single spectral feature, as long as S/N is adequate. The rule of thumb for line centroiding is that the precision is roughly the line width divided by the peak S/N (e.g., Lenz & Ayres 1992). The latter often is 20, or better, so precisions of a small fraction of a resol can be achieved under good circumstances. This is more than adequate to prevent spectral blurring. In cases for which suitable narrow features were lacking, or S/N was too low, the subtle effects of spectral blurring would not be recognizable in the first place, so a blind co-addition would be acceptable.

The cross-correlation shifts were reported, and applied, as equivalent velocities, because that is how the echelle spectrum responds to a simple spatial displacement of the target away from the center of the observing aperture (i.e., a set of narrow features at different wavelengths would show the same velocity offset, rather than, say, a constant wavelength shift). Conveniently, this allows the derived shifts to be compared without regard to the specific template wavelength. (The velocity shifts were applied according to Equation (1).)

The co-addition was implemented as a semi-automated procedure, with an "operator" in the loop. The operator first was presented with a view of a 50 Å interval of the leading spectrum of a set. The specific region was chosen autonomously by the "robot" on the basis of a numerical test that identified areas of combined high contrast and high S/N. The operator then would select a specific feature to serve as the cross-correlation template. Next, the other spectra of the sequence were correlated against the template in a narrow wavelength band (also chosen by the operator), and the resulting usually bell-shaped cross-correlation trace was centroided by a Gaussian. Finally, the resulting fits were displayed. If the global cross-correlation was judged a success, the operator would activate the subsequent co-addition step. If the fits appeared to be poor, the operator could choose a new template feature from the same region, move on to a new interval, or, if nothing better seemed to be available, default to a blind co-addition (all relative shifts set to zero).

Following the alignment step, the other sub-exposures were interpolated onto the wavelength grid of the leading one, after applying the velocity corrections derived from the cross-correlation centroids. In principle, this implies a somewhat nonuniform treatment, because the first exposure was not subjected to an interpolation, and consequent potential smoothing. The initial double-density resampling, however, should mitigate this bias. Since the wavelength bins of the sub-exposures were the same, there was no need to modify the photometric error.

Next, the re-gridded spectra were averaged, weighting the flux densities by exposure time and combining the photometric errors in quadrature, $\sigma _{\rm tot}= \sqrt{\sum {\omega _{i}^2\, \sigma _{i}^2}}$, where ω_i ≡ t_i/∑t_j is the normalized exposure time weight. In principle, the weighting factor should be proportional to the number of net counts in each spectrum. For the sub-exposures, however, net counts and exposure time are synonymous. (The camera backgrounds are low and do not depend strongly on circumstances of the observations.)

In the averaging step, only valid points (q_λ < 500) were included. The q_λ for the average flux then was set to the maximum of the set of the q_λ < 500. If all the flux points in the wavelength bin were flagged as bad, the parameters of the point with the smallest q_λ>500 were retained. If all the points were flagged with q_λ = 500 (indicating a hard gap), then the output f_λ and σ_λ were set to zero, and the data quality to 500.

Embedded in Table 4 are summaries of the specific exposure sequences, template wavelengths, cross-correlation bandpass half-widths, and derived velocity shifts (zero, by definition, for the leading exposure of a group). If a template wavelength and band are not reported, a blind co-addition was performed. (Incidentally, Table 4 has a similar structure to the scripts that controlled this and subsequent stages of co-addition and splicing.)

These "o-type" data sets were written to FITS¹³ for later public access. The FITS filename was derived from the original ${\sf x1d}$ root and follows the MAST HLSP convention; e.g., "h_o4o001010_spc.fits." The header of FITS extension zero (EXTEN=0) contains information concerning the target, including coordinates, name assigned by the GO, consensus StarCAT name, and HST proposal identifier; the observation, including start time, exposure duration, STIS echelle mode, setting, and aperture; and the processing configuration, including cross-correlation template parameters and derived velocity shifts (when sub-exposures were involved). Also listed are the wavelength ranges of each echelle order retained in the concatenation process. From these ranges, one can infer the overlap zones and splice points. If the observation consisted of only a single subexposure, then the StarCAT processing flag would be set to "NULL" and the concatenated data set would be stored in the first (and only) extension of the FITS file (EXTEN=1). The parameters are WAVE (wavelength: Å), FLUX (flux density: erg cm⁻² s⁻¹ Å⁻¹), ERROR (photometric error: same units as flux density), and DQ (data quality flag: unitless). If the observation had two or more sub-exposures, then the processing flag would be set to "ZEROTH-STAGE" and the co-added spectrum would be written to EXTEN = 1. The individual sub-exposures 1 through n would be written sequentially to EXTEN = 2 through n + 1. Thus, EXTEN=1 of the o-type FITS files always contains the most refined version of the observation, although all the sub-exposures, if any, would be recorded in the trailing extensions.

Figure 3 illustrates histograms of the derived velocity shifts, separated by uber mode, for Stages 0 and 1. Combining the samples is justified because the distinction between the two types of co-additions is minimal: both consist of sequences of exposures in the same tilt taken close together in time. The Stage 0 processing was done separately mainly to ensure that each individual STIS o-type exposure, say as tabulated in a MAST query, could be associated with a single refined data product.

Zoom In Zoom Out Reset image size

The shifts included in the histograms were taken exclusively from sequences having valid cross-correlation solutions (and ignoring the leading exposures for which υ ≡ 0 by design). The parameters of the distributions are summarized in Table 6. The 1σ standard deviations are slightly more than 1 km s⁻¹ for the M modes, and about half that for the H modes (${\sim} \mbox{$\frac{1}{3}$}$ pixel in both cases). The widths of the profiles—which mainly must reflect spacecraft drifts and/or jitter—are narrow enough that velocity blurring in a blindly co-added sequence should not be troublesome in general.

Although the vast majority of derived cross-correlation shifts at the visit level (Stage 1) were less than a few km s⁻¹, usually much less, there were a few notable exceptions. For HD 36285, the target apparently was "dithered" on the 310 × 005 ND slit, which produced very large, although discrete, velocity shifts between the exposures of that visit (υ = 0 and ±808 km s⁻¹). For HZ-HER, the derived velocity shifts were systematically changing during each visit (owing to rapid orbital motion), with up to tens of km s⁻¹ increments between exposures, and a maximum shift >100 km s⁻¹. In both cases, the anomalous velocities had an external cause, and do not reflect the usual high precision of the HST guiding.

The same-setting, same-visit co-addition was described as, e.g., "E140H-1271_020X009_51773." The first part is the echelle mode designator and scan setting; the middle part is an aperture code (height times width, in units of 10 mas, as before); and the final part is the last five digits of the Julian start time, in integer days, of the leading exposure of the sequence (i.e., JD – 2,400,000: "modified Julian Date" or MJD).

There were a few cases in which separate same-setting co-adds were from clearly delineated exposure groups, but close enough in time that the integer MJD flag was the same. In these instances, "a" and "b" would be appended to the aperture designator, for example, if there were two such identical file names. There are other examples where "F" or "N" appear at the end of the aperture flag. The former refers to observations taken through the FP-split slits mentioned earlier, and the latter to one of the ND filtered options. A few instances of "F"- and "N"-suffix exposures also contained independent sequences close enough in time to yield identical file names. In such cases, the first co-add would replace suffix "N" with "a;" the second sequence, with "b," and so forth.

The Stage 1 data sets were written to FITS, with the same general structure as the o-type files. Again, the EXTEN = 0 header contains target information and digests of the Stage 1 processing: constituent exposures, cross-correlation template (central wavelength and half width), and derived velocity shifts. In the exposure list, an o-type spectrum with suffix "_1" designates a single exposure, while an unadorned o-type indicates a Stage 0 co-add. (This convention is maintained in the Stage 2 and 3 FITS headers described later.) The first (and only) data extension (EXTEN = 1) contains the spectral parameters (WAVE, FLUX, ERROR, and DQ). The HLSP FITS file would be called something like, "h_ad-leo_e140m-1425_020x020_51613_spc.fits." (The StarCAT target name was incorporated to minimize potential naming degeneracies.)

For the mixed epoch and/or aperture sequences, the cross-correlation and averaging procedures were modified in several important ways compared with the single epoch case.

First, the derived velocity shifts—as benchmarked against the template spectrum—subsequently were normalized to an average over all the shifts. Each individual observation (which might be a Stage 0 or 1 co-add) can be viewed as an independent realization of the target centering process (which includes acquisition, peak-up, and guiding), which certainly must suffer some degree of random error. Accordingly, all the contributing spectra should be treated equally. In principle, subtracting the average shift should improve the velocity zero point by something like the square root of the number of independent contributors.

The second refinement was that because all the exposures were treated on an equal basis, any one of them could serve as the cross-correlation reference. The data set with the highest "quality factor" (QF: percentage of flux points exceeding 2.5σ with respect to the local photometric error (per resol), a measure of relative S/N) was selected for the purpose. This differs from the single epoch case where the leading spectrum of a sequence always was designated as the benchmark, because in principle its velocity zero point should be the most accurate of the set. Unlike at Stages 0 and 1—where the constituent exposures usually were similar in duration—now one or more of the spectra might be a multi-repeat co-add, which then would be the preferred reference from an S/N standpoint.

The third difference for the mixed epoch/aperture co-adds was that occasionally an echelle order or two would be dropped in one set or the other, because of the specific positioning of the MSM in each independent visit. The consequence was that the extreme echelle orders might not fully benefit from the co-addition, which then would be reflected in a higher σ_λ for the affected orders.

The fourth difference is that the overall weighting in the co-addition was by σ⁻²_λ instead of t_exp. This is equivalent to weighting by (S/N)², which is nearly the same as weighting by net counts per bin.¹⁷ This was done primarily to account for the mixed aperture cases: while the assigned flux density levels might be the same in two observations of similar duration, the underlying count rates could be quite different owing to transmission losses by, say, a narrow slit compared with the broader photometric aperture.

The result of a Stage 2 co-addition would be designated, for example, "E140M-1425_020X020_XXXXX-YYYYY" for the average of a set of E140M-1425_020X020 exposures taken on different dates spanning the range of integer MJDs XXXXX–YYYYY. If two different apertures were involved, both would be appended and the name would be, e.g., "E140M-1425_020X020_020X060_XXXXX," if the observations were in the same epoch, and "..._XXXXX-YYYYY" otherwise. Any additional suffixes from Stage 0 or 1 (e.g., "a" or "N") were not carried over to the new filename. Table 7 outlines the co-addition scheme for these types of mixed exposure sequences. Again, the co-additions were restricted to objects displaying minimal variability over their observational time lines. Figure 4 illustrates histograms of relative velocity shifts derived from any cross-correlations performed within this set. Although the samples are smaller, the distributions are similar to those of Stage 0 + 1, perhaps slightly narrower for the H modes.

Zoom In Zoom Out Reset image size

Table 7. Stage Two Co-additions: Same Setting, Different Epochs and/or Apertures

Data Set	Mode/Setting	Aperture	Band	texp	Modified JD	Qual.	Xcorr Params.
		(001 × 001)	(Å)	(s)	(d)	(%)	(Å or km s−1)
(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
AD-LEO
E140M-1425_020X020_51614			1140-1729	13000	51614.151	12	+0.37
E140M-1425_020X020_51615			1140-1729	13000	51615.090	12	+0.50
E140M-1425_020X020_51616			1140-1729	13000	51616.096	12	−0.95
E140M-1425_020X020_51613			1140-1729	13000	51613.145	10	+0.34
E140M-1425_020X020_52426			1140-1709	10390	52426.297	8	+0.30
E140M-1425_020X020_52427			1140-1709	4630	52427.233	5	−0.56
E140M-1425_020X020_51613-52427			1140-1729	67024	51613.145	31	1306.08 [0.12]
AG-DRA
E140M-1425_020X020_52755			1140-1729	4437	52755.815	94
E230M-1978_020X020_52755			1607-2365	2113	52755.881	83
o6ky01030	E230M-2707	020 × 020	2275-3118	374	52755.873	99
AO-PSC
E140M-1425_020X020_51741			1140-1709	9630	51741.432	95	+0.49
E140M-1425_020X020_51740			1140-1709	9630	51740.494	95	−0.49
E140M-1425_020X020_51740-51741			1140-1709	19261	51740.494	96	1334.49 [0.14]
BG-PHE
o63502010	E140M-1425	020 × 020	1140-1709	1440	52041.327	97
BW-SCL
o5b616010	E140M-1425	020 × 020	1140-1729	1977	51433.056	2
CI-CAM
E140M-1425_020X020_51623			1140-1729	8022	51623.065	47
o5ci01040	E230M-2269	020 × 020	1857-2672	2996	51623.254	41
CY-TAU
E140M-1425_020X020_51884			1140-1729	5178	51884.739	3
o5cf03010	E230M-1978	020 × 020	1607-2365	1260	51884.671	1
E230M-2707_020X020_51884			2275-3118	1020	51884.690	7
DG-TAU
E140M-1425_020X006_51839			1140-1729	12295	51839.408	0
E230M-2707_600X020_52974			2275-3118	4800	52974.980	3
E230M-2707_020X020_51960			2274-3117	11114	51960.193	86
DN-LEO
o66v04010	E140M-1425	020 × 020	1140-1729	144	51951.170	90
E230M-2269_020X006_52653			1841-2673	4721	52653.932	99
DR-TAU
E140M-1425_020X020_51785			1140-1729	5794	51785.668	8
o5cf02010	E230M-1978	020 × 020	1607-2365	1080	51785.609	13
E230M-2707_020X020_51949			2274-3117	5207	51949.284	91	−0.72
o5cf02020	E230M-2707	020 × 020	2276-3119	916	51785.625	85	+0.72
E230M-2707_020X020_51785-51949			2274-3117	6123	51785.625	92	2796.55 [0.48]
etc.

Notes. Objects are listed alphabetically. Same-setting, different-epoch, and/or different-aperture data sets are grouped together (if multiple observations were obtained and judged to be suitable for combining), followed by the name (and properties) of the co-added spectrum. Note that data sets in a grouping now are ordered in descending quality factor, so that the template spectrum (leading one) will be optimum from an S/N standpoint. Columns are same as for Table 4. (Complete table available at http://archive.stsci.edu/prepds/starcat.)

Download table as:  ASCIITypeset image

The Stage 2 data sets were written to FITS with the same structure as the Stage 1 files. The resulting MAST HLSP file name would be something like "h_ad-leo_e140m-1425_020x020_51613-52427_spc.fits."

When a target had been observed in multiple wavelength regions, a scheme was applied to knit these segments together. The procedure was trivial for those observations lacking wavelengths in common, for example, an E140M-1425 (1150–1700 Å) plus an E230H-2762 (2623–2900 Å). Then, the two spectra (which might themselves be co-adds of one type or another) were simply butted together in wavelength, with some points trimmed off the end of the first spectrum and the beginning of the second, especially if flagged as "bad." For the nontrivial cases, involving overlapping wavelengths, a more elaborate splicing procedure was developed (which will be described shortly).

Table 8 summarizes the specific groupings that were spliced. In one extreme case, the eclipsing binary VV Cep (HD 208816: M2 Iab + B6 II), there were 21 separate spliced spectra covering the 21 independent epochs. At the other extreme is WD0501+527 (calibration target G191-B2B, a DA white dwarf). Here, the extensive collections of M and H settings were spliced separately, owing to the complete coverage at each resolution, and regardless of epoch because the source is very stable. There were seven mixed epoch co-adds and single exposures in the "M" splice (with a maximum monochromatic exposure,¹⁸ t_max = 15.6 ks), and a remarkable 36 independent spectra in the "H" splice (t_max = 22.6 ks). Also worth mentioning is the single-setting multi-epoch co-add of the dM flare star AD Leo "E140M-1425_020X020_51613-52427," which encompassed six independent visits and a total of 26 exposures, with t_max = 67 ks. Although several "flares" were captured in the individual pointings, most were relatively small, and it was judged that the super-co-add would have value as the deepest FUV exposure of a late-type star obtained by STIS. (See, also, Figure 1(a), which was based on a sum of the first 20 exposures of the sequence, taken within a few days of one another.)

Table 8. Stage Three Co-additions: Wavelength Splicing

Data set	Mode/Setting	Aperture	Band	texp	Modified JD	Qual.	Fscl	Vshft	Xcorr Params.
		(001 × 001)	(Å)	(s)	(d)	(%)		(km s−1)	(Å)
(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
AD-LEO
E140M-1425_020X020_51613-52427			1140.0-1729.0	67024	51613.145	31
AG-DRA
E140M-1425_020X020_52755			1150.0-1705.0	4437	52755.815	94	1.025	+1.23	...
E230M-1978_020X020_52755			1610.1-2347.3	2113	52755.881	83	1.025	−0.29	1639.61 [0.37]
o6ky01030	E230M-2707	020 × 020	2280.0-3118.3	374	52755.873	99	1.000	−0.94	2305.69 [0.42]
UVSUM_1M_52755			1150.0-3118.3	6550	52755
AO-PSC
E140M-1425_020X020_51740-51741			1140.0-1709.0	19261	51740.494	96
BG-PHE
o63502010	E140M-1425	020 × 020	1140.0-1709.0	1440	52041.327	97
BW-SCL
o5b616010	E140M-1425	020 × 020	1140.0-1729.0	1977	51433.056	2
CI-CAM
E140M-1425_020X020_51623			1150.0-1708.7	8022	51623.065	47	1.000	+0.00	...
o5ci01040	E230M-2269	020 × 020	1861.3-2672.6	2996	51623.254	41	1.000	+0.00	...
UVSUM_1M_51623			1150.0-2672.6	8022	51623
CY-TAU
E140M-1425_020X020_51884			1150.0-1705.1	5178	51884.739	3	1.000	−0.04	...
o5cf03010	E230M-1978	020 × 020	1610.1-2342.6	1260	51884.671	1	1.000	−0.04	...
E230M-2707_020X020_51884			2279.9-3118.2	1020	51884.690	7	1.000	+0.08	2326.21 [0.38]
UVSUM_1M_51884			1150.0-3118.2	6438	51884
DG-TAU
E140M-1425_020X006_51839			1150.0-1699.9	12295	51839.408	0	1.000	+0.00	...
E230M-2707_020X020_51960			2289.7-3117.7	11114	51960.193	86	1.000	+0.00	...
UVSUM_1M_51839-51960			1150.0-3117.7	12295	51839
etc.

Notes. Data sets that were spliced are grouped together (if multiple observations were obtained and judged to be suitable for combining), followed by the name (and properties) of the spliced spectrum. Columns are same as for Table 7, except with addition of the flux scale factor ("Fscl" (Column 8)), and now the derived velocity shift and cross-correlation template parameters are listed on each line (Columns 9 and 10). Also note that the entry in Column 5 for the spliced spectrum is the maximum monochromatic exposure, t_max. Second exposure of a sequence is assigned template parameters referring to overlap with the first spectrum; third exposure (if any) will have parameters for spectrum 2 and 3 overlap; and so forth. Velocity shifts ("Vshft") initially were benchmarked against first exposure, then average for the whole group was subtracted from each entry. If Vshfts and template parameters are blank, the adjacent spectra either were simply concatenated, if there was no overlap between them; or "blindly spliced" in cases of overlapping wavelengths, but insufficient S/N for a reliable Vshft determination or simply lack of a suitable template feature. Similarly, an Fscl of 1.000 in a list of generally >1 values would typically be the spectrum with highest apparent throughput, against which the other wavelength regions were normalized. Unit Fscl also will occur as default if segments of an exposure group were not overlapping, including cases of insufficient S/N in one, or both, components to reliably measure a flux ratio. (Complete table available at http://archive.stsci.edu/prepds/starcat.)

Download table as:  ASCIITypeset image

As for the splicing procedure itself, it is worthwhile reviewing how this was handled in the earlier CoolCAT project. There, the individual FUV and NUV segments of a target were draped over a broadband UV energy distribution abstracted from the extensive collection of IUE low resolution (R ∼ 300) and echelle (R ∼ 10⁴) spectra, covering the period 1978–1995. However, this procedure was viewed as less practical for StarCAT owing to the much larger sample of objects (more than ten times greater) in the face of the labor intensive nature of building a consensus IUE SED. Furthermore, the HST radiometric calibration is superior to that of the earlier observatory, and should be preferred. Finally, there is the issue of variability, since the IUE spectra were obtained years to decades prior to the STIS pointings, and the spectral energy distribution (SED) of the earlier period might not be appropriate to the later HST era.

Thus, it was decided to bypass the external calibration step, and rely instead on overlapping spectral intervals to serve as an internal radiometric scaffold. Consequently there are a few examples when a target had been observed with non-overlapping echelle settings, and the segments do not appear to line up particularly well. In some instances, the disjoint appearance might be misleading due to, say, strong interstellar absorption between well-separated FUV and NUV segments (e.g., the 2200 Å "bump"). In other cases, it might represent a legitimate throughput issue: a target imperfectly centered in a narrow aperture.

For the more favorable situation of overlapping spectral intervals, flux ratios of wavelength bands in common to adjacent segments were measured (if enough signal was present in both members of a pair). Such ratios (or a default of unity in the absence of any overlap or insufficient signal) were propagated through the full sequence to determine a flux scaling factor for each spectrum relative to the first of the group (which was assigned an initial scale factor of 1). Specifically, the scale factor for spectrum 2 would be the overlap-2/overlap-1 flux ratio; the scale factor for spectrum 3 would be that of 2 multiplied by the overlap-3/overlap-2 ratio; and so on. The maximum value of the set pointed to the spectrum that presumably achieved the best throughput. That value then was divided by the others to obtain flux correction factors ("Fscl," always ⩾1) to pass to the splicing procedure. At worst, the "fluxing" strategy would accurately trace the shape of the target SED, but at best capture the absolute radiometric level as well.

As in the previous Stages, wavelength registration was accomplished by cross-correlating contrasty features in the overlap regions of adjacent segments. Like the multi-epoch co-adds, the velocity shifts for each overlap zone (or zero if there was a gap between neighboring segments) were referenced to the initial segment (shortest wavelength); then the average of the whole set was subtracted for the same reason of not wishing to bias the global velocity zero point, treating each measurable overlap region as an independent realization of the target centering process. Of course, many of the overlap zones were contributed by multi-epoch co-adds for which the internal velocity zero in principle already had been refined, to the further benefit of the global accuracy.

Following alignment in flux and wavelength, the overlapping segments were averaged. The quantities of the shorter wavelength interval ("1") were linearly interpolated onto the wavelength scale of the longer one ("2"), except for data quality, which was treated in a nearest neighbor fashion. Prior to interpolation, the photometric error of segment 1 was multiplied by $\sqrt{\Delta \lambda _{1}/\Delta \lambda _{2}}$ to compensate for the, in general, different binning (Δλ: Å per bin, variable with λ) of the two intervals.

Like the Stage 2 procedure, the weighting factor was σ⁻²_λ. This was crucial to the success of the splicing, because of the typically much different monochromatic sensitivities of the overlapping modes, and because it often was the case that at the extreme end of a segment, the sensitivity would be varying rapidly with wavelength through the overlap zone. The monochromatic photometric error conveniently captures such differences and variations. The hierarchy for assigning flux values when one or both points were flagged as bad was the same as in the initial ${\sf x1d}$ concatenation step.

Commonly, one of the contributing spectra dominated in the overlap zone, because each exposure originally was set to optimally record some feature typically at the center of its grasp, rather than at the peripheries, so the resulting exposure depths often were quite different. Here, the role of the weaker spectrum would be to set the relative flux scale, and the velocity offset if a suitable marker was present, rather than contributing directly to improving S/N. There were cases, however, where one spectrum dominated, but had flagged gaps in the overlap zone. These gaps then were filled by the weaker spectrum, and the local photometric error consequently would take on a periodic picket-fence appearance. This effect could be minimized by choosing the long-wavelength splice point short enough to exclude as many of the gaps as possible, although at the cost of eliminating a few tens of Å of higher S/N spectrum between the gaps. Such choices were made on a case by case basis.

The philosophy that guided the specific splicing sequences was to include at each wavelength the highest spectral resolution material available, and to minimize, to the extent possible, averaging unlike resolutions (i.e., M and H). Within this strategy, M and H sequences were spliced into separate spectra only if there was complete, or nearly complete, wavelength coverage in each set. The alternative would have been to routinely splice M and H sequences separately. However, the mixed resolution philosophy was dictated by the calibration strategy (comparing flux ratios in overlapping regions): incorporating all the material available for each target maximized the overlaps for that purpose. Maximizing overlap opportunities also was important for ensuring accurate wavelength continuity throughout the spliced spectrum (by cross-correlating features in overlap zones). Otherwise, there might be small velocity discontinuities between the different segments.

For the cases of overlap between M and H tilts, the co-addition zone was chosen to be as small as possible, but sufficiently wide to provide an accurate flux ratio, if feasible; and (ideally) a high contrast cross-correlation feature, to ensure accurate velocity alignment. Co-addition of M and H overlaps was performed—instead of, say, simply introducing a sharp discontinuity—mainly to achieve the higher S/N and ensure a smoother spectral transition between the unlike modes. Because of the averaging, however, narrow features not resolved at M resolution would acquire hybrid line shapes. Broad features, especially the continuum, would be less affected, and more fully benefit from the increased signal (if M and H had similar S/N). In any event, all the constituent exposures reside in StarCAT, so specific cases where the M + H averaging was viewed as undesirable could be examined in whichever of the original modes was preferred.

In some instances, one finds a single H setting fully contained within the range of an M tilt. The splicing sequence was as follows: (1) retain the M values below the shortward edge of H, with minimal overlap between M and H at the shortward side of the latter as described above; (2) include the full range of H up to its long-wavelength edge; and (3) splice back the remainder of the M region, again with minimal overlap at the longward edge of H. In Table 8, one would find the M tilt repeated twice, sandwiching the H setting, with different slicing parameters reflecting the specific choices of overlap zones at the boundaries of H.

Included in Table 8 are flux scale correction factors, parameters of the cross-correlation templates, if any, and derived velocity shifts. Figure 5 illustrates a histogram of the latter (restricted to cases of successful cross-correlation solutions). Figure 6 is a histogram of the relative flux corrections determined from overlap zones jointly containing sufficient S/N to yield an accurate flux ratio. Because the corrections are inverses of ratios referenced to the largest observed value in a set, they always are ⩾1. For the histogram, only exposures with a valid cross-correlation solution were considered (specifically, the second of a spliced pair). If both members of a spliced pair had sufficient S/N to allow a cross-correlation solution, they would be assured of having sufficient S/N for a valid flux ratio measurement. Within this sample, only values ⩾1.001 (the next significant figure in Table 8 above ≡1) were included, to avoid the artificially sharp peak contributed by the reference (normalizing) exposures. The cutoff excludes a few cases where the derived correction was very close to 1, and reported as such owing to roundoff of significant digits in the table, but the main purpose of the figure is illustrative. Most of the scale factors are close to unity, the desirable outcome if the aperture calibrations and overall broadband radiometry are accurate.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The few cases of large corrections were examined more closely. Observations for which Fscl > 1.5 were isolated from the full list of correction factors, regardless of cross-correlation success. Of the 38 single exposures or co-adds identified in this way, almost all were H settings (more NUV than FUV). Fully one-third had been taken through the tiny Jenkins slit, for which centering errors can dominate over knowledge of the transmission factors. Another quarter of the exposures had utilized one of the ND filtered apertures, and about a sixth had used the (also nonstandard) 010 × 020 slot. Only 10% of the observations had been through the 020 × 020 photometric aperture. The preference of the anomalous Fscl exposures for very narrow or nonstandard apertures is consistent with the idea that the deviations arise from centering errors on the one hand or less well-calibrated transmission factors on the other.

A successful splicing was designated something like, "UVSUM_1M_XXXXX-YYYYY." The "UVSUM" part indicates that multiple wavelength regions were joined; the middle numeral points to a particular grouping of data sets, which might represent a single epoch of a variable source, or multiple visits of a less variable object; the letter appended to the group index encodes whether the constituent spectra all were medium resolution ("M"), all high resolution ("H"), or mixed ("X"); and the final part summarizes the range of modified Julian start times representing the specific sequence. The MJD designator was composed of the extreme values of the set, considering that one, or more, of the constituent spectra might itself be a mixed epoch co-add. If only a single epoch was represented in the spliced group, then just the one integer MJD would appear.

The Stage 3 UVSUM data sets were written to FITS files of similar structure to the E-type. The EXTEN = 0 header contains target information and splicing parameters; EXTEN = 1 holds the spectral data values. The HLSP file name would be, e.g., "h_ag-dra_uvsum_1m_52755_spc.fits."

Table 9 lists the set of fully distilled spectra for each target, constituting the final catalog. At worst, this would be just a single exposure. At best, there would be a fully spliced spectrum covering the entire UV range without gaps. In the middle are single-setting co-adds, possibly multi-epoch, or multiple epochs of partially or fully spliced spectra. Figure 7 illustrates a low resolution tracing of the final spectrum (or spectra) of representative objects, together with a schematic time line depicting the individual exposures as a function of mode, setting, and epoch. In some instances, like VV Cep, the spectra are highly varied, following the carefully orchestrated multi-Cycle GO program (to catch critical orbital events such as ingress, egress, total secondary eclipse, quadrature, and so forth). In other cases, like G191-B2B, the spectra are time independent, but combine to produce composites that achieve outstanding S/N throughout the full UV grasp of STIS. (The top-level previews residing in StarCAT are equivalent to a complete version of Figure 7 for the whole catalog.)

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Table 9. StarCAT Final Catalog

Object Name	Data set	Mode/Setting	Aperture	Band	tmax	Modified JD	Qual.
			(001 × 001)	(Å)	(s)	(d)	(%)
(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
AD-LEO	E140M-1425_020X020_51613-52427			1140-1729	67024	51613.145	31
AG-DRA	UVSUM_1M_52755			1150-3118	6550	52755	96
AO-PSC	E140M-1425_020X020_51740-51741			1140-1709	19261	51740.494	96
BG-PHE	o63502010	E140M-1425	020 × 020	1140-1709	1440	52041.327	97
BW-SCL	o5b616010	E140M-1425	020 × 020	1140-1729	1977	51433.056	2
CI-CAM	UVSUM_1M_51623			1150-2672	8022	51623	45
CY-TAU	UVSUM_1M_51884			1150-3118	6438	51884	3
DG-TAU	UVSUM_1M_51839-51960			1150-3117	12295	51839	29
DN-LEO	UVSUM_1M_51951-52653			1150-2673	4721	51951	95
DR-TAU	UVSUM_1M_51785-51949			1150-3117	7203	51785	31
DS-TAU	UVSUM_1M_51780			1150-3119	6986	51780	33
EF-PEG	E140M-1425_020X020_51713			1140-1709	6883	51713.080	0
EK-TRA	o5b612010	E140M-1425	020 × 020	1140-1709	4302	51384.772	39
EV-LAC	E140M-1425_020X020_52172			1140-1729	10920	52172.698	4
EX-HYA	E140M-1425_020X020_51682-51694			1140-1709	15201	51682.894	97
FQ-AQR	E230M-2269_020X006_52527			1840-2672	7620	52527.226	99
FU-ORI	E230M-2707_020X020_51962			2274-3117	5217	51962.936	72
HD100340	o63557010	E140M-1425	020 × 020	1140-1729	1440	51999.496	97
HD100546	UVSUM_1H_51747			1150-2887	3134	51747	95
	UVSUM_2H_51852			1494-2887	1970	51852	99
HD101131	o6lz48010	E140H-1489	020 × 020	1390-1586	300	52740.936	99
HD101190	o6lz49010	E140H-1489	020 × 020	1390-1586	300	52577.357	99
HD101436	o6lz51010	E140H-1489	020 × 020	1390-1586	600	52786.196	99
HD102065	UVSUM_1H_50900-50901			1150-2888	8879	50900	97
HD102634	o4ao06010	E230M-2707	020 × 006	2275-3118	1440	51128.343	100
HD103095	UVSUM_1X_52839			1150-3158	11097	52839	53
HD103779	UVSUM_1H_51284-51959			1160-1901	1466	51284	97
HD104705	UVSUM_1H_51171			1160-1586	6220	51171	97
HD106343	UVSUM_1H_51284			1163-1901	1486	51284	96
HD106516	UVSUM_1M_51163-51202			1150-3118	1728	51163	34
HD106943	UVSUM_1H_52786			1160-1551	2220	52786	95
HD10700	E140M-1425_020X020_51757			1140-1709	13450	51757.124	18
HD107113	UVSUM_1M_50736			1607-3118	3481	50736	88
HD107213	UVSUM_1M_51378			1607-3116	3539	51378	76
HD107969	E140M-1425_020X020_52281			1140-1729	5197	52281.514	97
HD108	UVSUM_1H_51677			1163-1901	3988	51677	95
etc.

Notes. Refer to Tables 4, 5, 7, and 8 for specific groupings of exposures in each final data set. In a few cases for which complete spectral coverage was available in both M and H sequences, these were separately spliced. In most situations, however, the final UVSUM spectrum is either entirely M, or entirely H, and only occasionally a mixture of the two ("X-type"). (Complete table available at http://archive.stsci.edu/prepds/starcat.)

Download table as:  ASCIITypeset image

Figure 8 illustrates a few examples of the StarCAT spectra at higher resolution (objects of Figure 1), showing the remarkable detail that often is present in the co-added and spliced STIS echellegrams. All the intermediate data products are accessible through StarCAT, so an investigator can examine the material for a given object at any level of detail desired.

Zoom In Zoom Out Reset image size

The StarCAT effort was especially careful with the wavelength scales. The distortion correction to the ${\sf x1d}$ files improved the internal consistency of the relative wavelengths, and removed small systematic errors in the assigned velocity zero points in some instances (especially for the less heavily used secondary tilts). Taking the Deep Lamp project as a guide, the internal precision of the wavelength scale of an arbitrary corrected STIS exposure should be better than the 600 m s⁻¹ for M modes and 300 m s⁻¹ for H measured in uncorrected wavecals (roughly $\frac{1}{10}$ resol, or $\frac{1}{5}$ pixel, in both cases). Such precision could be realized in spectra with S/N > 10. The cited wavelength precision is similar to that quoted in Table 16.2 of the STIS Instrument Handbook, noting that the values there are 2σ.

The accuracy of the velocity zero point of an arbitrary STIS echellogram is trickier to assess. In principle, it depends mainly on the target centering, and thus, in practice, on the time interval from the last peak-up, and the success of the centroiding at that peak-up. One could carry out, say, a careful evaluation of the apparent UV radial velocities of StarCAT objects for which accurate ground-based υ_r have been measured. Such a comparison was made for the earlier CoolCAT effort. There, sharp chromospheric emission lines deviated from the known stellar (photospheric) radial velocities by only +0.1 ± 1.4 km s⁻¹ (Ayres 2002; considering just narrow-line dwarfs and giants, excluding dMe dwarfs; from Tables 1 and 2 of that article, 13 stars in all). Although the analogous StarCAT comparison has not yet been done, the CoolCAT empirical standard deviation (1.4 km s⁻¹) probably is a reasonable estimate of the true absolute accuracy (especially since the empirical σ probably is boosted somewhat by uncertainties in the υ_r themselves). This estimate, based exclusively on E140M exposures, lies between the $\mbox{$\frac{1}{2}$}$–1 pixel absolute accuracy (2σ) quoted by the Instrument Handbook.

What can be gleaned straightforwardly from the StarCAT processing are the relative velocity shifts between observations in a sequence. These show a standard deviation somewhat over 1 km s⁻¹ for M modes and about half that for H modes (${\sim} \mbox{$\frac{1}{3}$}$ pixel in both cases). This is a measure of pointing jitter and/or drifts (the latter can be a problem if only a single guide star was available), and would be the velocity blurring expected in a typical sequence of exposures that was blindly co-added. Such blurring would be inconsequential except perhaps in the highest S/N studies of intrinsically narrow spectral features with the H modes. However, in a high-S/N spectrum containing conspicuous narrow features, the cross-correlation registration normally would succeed, thereby minimizing any profile blurring. (The "centroiding error" of a high quality cross-correlation measurement—templates with peak S/N > 20—would be at a level of $\frac{1}{20}$ resol ($\frac{1}{10}$ pixel), which already is twice as good as the internal wavelength precisions cited earlier.)

In spliced spectra, the precision of the wavelength scales would be comparable to those of single-setting exposures, if there was good overlap between the spliced segments, and if the overlap regions all contained suitable cross-correlation features. Otherwise, particularly when intra-tilt gaps were present, one segment of the spliced spectrum might deviate from the velocity scale of the next segment by the absolute accuracy factor, although the relative precision within each segment still would be high. One could consult the Stage 3 splicing sequence for a particular object to judge whether the spectral continuity was such that the internal velocity scale likely would achieve the single-setting precision over the full range of the splice, or whether some of the segments were isolated from others, and thus the velocity scales might be less well connected.

For the multi-epoch co-adds and spliced spectra, where average velocities were subtracted from the group cross-correlation shifts, the accuracy of the absolute velocity in principle would be improved by $\sqrt{n}$, where n is the number of υ involved. This would be exclusive, however, of any systematic errors in the assignment of the velocity zero point (including corrections for the spacecraft orbital motion, the telluric component, and so forth). The CoolCAT analysis suggests that any systematic zero-point errors must be small, because the average deviation (over the thirteen stars considered here) was close to zero.

There are two main sources of uncertainty for the StarCAT flux scales. One is the absolute accuracy of the STIS radiometric calibration, and its repeatability, which includes knowledge of the aperture transmission factors for each of the settings. Again, for an isolated exposure, the applicability of the default transmission factors will depend on the quality of the target centering. The absolute radiometric calibration of the STIS echelles—as determined from UV standard stars—is quoted as 3% (1σ) for a well-centered target in the 020 × 020 aperture, and the repeatability also is 3%, leading to a total error budget of about 4% (in quadrature; see Bohlin 1998).

The second major source of uncertainty is the fluxing procedure that joined overlapping spectral segments. As with the velocity scale, if the overlap zones in a group of spectra all have high S/N, so that accurate flux ratios can be derived for each individual pair, then the spliced spectrum should have high relative flux precision throughout the full range, and the absolute accuracy should approach that associated with whatever single exposure or co-add was selected by the splicing procedure as the reference spectrum (for normalizing the Fscl factors). Because each flux ratio is determined by averaging over many wavelength bins, the precision of the local measurement can be better than 1%, even when the monochromatic S/N is only 10 or so. Thus, the relative flux precision of a well-spliced spectrum should be much better than the absolute accuracy (and limited mainly by knowledge of the SEDs of UV standard stars). Like the velocity case, however, if the splice included gaps or low-S/N overlaps, the affected segments could not be empirically aligned, and then might suffer significant relative flux scale errors, especially if one or both sets of underlying observations had a poorly centered target.

Some idea of the photometric repeatability can be gleaned from the Fscl factors. Considering only the high-S/N sample (successful cross-correlation solutions), one finds that the average deviation for all the values (including those equal to 1, since the exposure assigned unit Fscl either was itself the peak-throughput spectrum of a group, or had unit flux ratio with it), is 1.06 (with standard deviation 0.14) for the 020 × 020 aperture (154 cases), and 1.07 (with standard deviation 0.18) when either the 020 × 009 or 020 × 006 aperture was used (78 cases). It is somewhat surprising that the narrower slits showed essentially the same average deviation as the photometric aperture. A simplistic interpretation would say that the true repeatability of the STIS photometry for echelle spectroscopy probably is closer to 7% when supported apertures were used. This would imply an overall flux accuracy of about 8%, including the standard star contribution, for an arbitrary exposure taken with one of the supported apertures. A spliced spectrum, however, could approach the 4% accuracy cited earlier, if it consisted of many constituent spectra, because there would be a better statistical chance that the selected benchmark spectrum (highest relative flux ratio) actually achieved ideal throughput (as assumed in the ${\sf calstis}$ processing).