3.2. A Key Subtle Twist

  If the background solutions from BCKGRD faithfully respond to changes in the background surface, then any changes in the background surface by themselves should not cause the spurious time-dependence at Lyman $\alpha$ shown in Figure 1. To investigate the time-dependence of the line depths further, we examined the prototype results for all 23 images in Table 2 and compared the interpolated Pass 1 background solutions at the Lyman $\alpha$ line with the gross flux in the Lyman $\alpha$ core flux within m = 113 itself. Invariably we found that Pass 1 solutions for the late-epoch (1992+) images were 7th-degree polynomials. This means that they had not failed any pathology tests at the conclusion of this pass. The same solutions also matched the true zero-flux level as determined by the Lyman $\alpha$ core. In contrast, solutions for early-epoch (<1982) images were usually 5th degree (degraded) polynomials, so these solutions had twice failed the solution-pathology tests. Moreover, they fell some 7% higher, relative to the local order height, (or 28 FN for the example of SWP04262) than the Lyman $\alpha$ core. Next we ran these data through BCKGRD ``with intervention," that is by forcing BCKGRD to accept the first, 7th-degree trial solution of Pass 1 without modification from the pathology tests. These solutions are depicted in Figure Figure 7a  & b for a pair of early- and late-epoch images, respectively. For late-epoch images notice that the solution runs nicely through a level corresponding to the flux of the Lyman $\alpha$ core. Indeed, the solutions for these images falls close to the line-core flux level. Panel a shows how the circumstances responsible for the solutions of early- and mid-epoch images are far more complicated. This time the pipeline-processing solution, as produced for the final IUE archive, falls well below the flux level of the Lyman line core. By ''tricking" BCKGRD into accepting the 7th-degree solution, we could force the solution in order m = 113 through almost exactly the flux level of the Lyman $\alpha$ core. However, this close match at this wavelength is coincidental because this same solution was unacceptable at long-wavelength orders; note the mismatch of the dashed line in the center of Figure 7a. Similarly, the forced solutions failed in the Lyman $\alpha$ region and for even shorter wavelength orders where negligible target flux falls as well as in the long-wavelength regions dominated by halation-caused overlap (lines 400-600). These experiments are instructive for several reasons. First, they show that BCKGRD works generally rather well as a general-purpose estimator of the global background surface. The algorithm usually did a good job of fitting within an error which remains roughly constant over the image (about 1 rms of adjacent background pixel-to-pixel fluctuations; typically 3% of the continuum flux). Note, first, that this error represents a larger percentage error of the continuum flux at short wavelengths. Second, one finds that even if BCKGRD ``knew" what the correct background is in Pass 1, e.g., for a hypothetical image containing numerous saturated lines, a 7th-order polynomial does not provide enough the curvature needed to fit the background function it would ``know" to be correct. In order to fit the steep slope of the null flux between m = 113 and 125 (lines 195 and 128, respectively), a polynomial of somewhat higher degree than 7 would be needed for early-epochs images. However, as already mentioned, extensive testing during the development stages of BCKGRD showed that permitting steep curvatures would in turn result in spurious ringing in attempting to match problems in other images. The tendency toward ringing is a potential flaw in a pipeline (automated) processing algorithm because it leads to unpredictable behavior in unchecked solutions, i.e., to ``cures" that are worse than inaccurate fits. BCKGRD generally catches the problem in Pass 1 and responds with a lower degree fit which does not ring. The expense of this procedure is that the resulting solution sometimes can may not follow the curvature that is needed, as shown in the example above. In such cases the solutions can be improved by human intervention, though with a loss of reproducibility. This fact may be important when comparing different spectra of the same target.

Figure 7: (Panel a): A comparison of the Pass 1 extraction in the spatial direction with the Chebyshev solution for the early-epoch image SWP04262. Three solutions are shown: the first, 7th-degree solution before a pathology check (dashed line), the 5th-degree solution actually accepted by the BCKGRD pipeline version (solid), and a ``customized" solution obtained by the author using the IRAF task CONTINUUM by weighting only the lower envelope of interorder fluxes (dot-dashed line). Note that none of these solutions is a satisfactory fit to the true background level in all regions. The location of Lyman $\alpha$ is shown by an arrow.

(Panel b): A comparison of the Pass 1 extraction in the spatial direction with the Chebyshev solution for the late-epoch image SWP55997 of $\tau$ Sco. Note how well the solution passes through the flux at the core of Lyman $\alpha$ (arrow).