Space Telescope and Optical Reverberation Mapping Project. VIII. Time Variability of Emission and Absorption in NGC 5548 Based on Modeling the Ultraviolet Spectrum

As described by De Rosa et al. (2015), we used the CalCOS pipeline v2.21 to process our data, but we made a special effort to enhance the calibration files for our particular observations. We developed special flat-field files, enhanced the wavelength calibration through comparison to prior Space Telescope Imaging Spectrograph (STIS) observations of NGC 5548, and tracked the time-dependent sensitivity of COS to achieve higher S/N and better flux reproducibility. For each day, all four exposures were calibrated, aligned in wavelength, and combined into a single spectrum for each grating for each visit. We binned these spectra by 4 pixels (approximately half a resolution element) to reduce residual pattern-noise features and achieve higher S/N. Ultimately, our wavelength scale achieves an rms precision of <6 km s⁻¹, and our flux reproducibility is better than 1.1% for G130M and better than 1.4% for G160M.

To produce a spectral model that can be adjusted to each individual observation in the reverberation campaign, we start with the high-quality mean spectrum produced from the whole campaign data set. This reveals individual weak features that we cannot precisely track in individual observations but that we must include to avoid biasing our results for stronger features of interest. The model of the mean spectrum is described in detail by De Rosa et al. (2015), but we also describe it here once again as a key reference for understanding the components that we vary in the fits to the individual spectra in our campaign's time series.

The basic model starts with the one developed for the XMM-Newton campaign of Kaastra et al. (2014), where the soft X-ray obscuration and broad UV absorption were first discovered. This model used a power-law continuum and multiple Gaussian components for both the emission and the broad absorption features. The high S/N of the mean spectrum from the reverberation mapping campaign requires the addition of more weak emission features, additional weak broad absorption features associated with all permitted transitions in the spectrum, and more components in the bright emission lines. We emphasize that the model is not intended as a physical characterization of the spectrum but rather as an empirical tool that enables us to deblend emission-line components and correct for absorption by using some simple assumptions about the shape of the spectrum.

Starting with the continuum, we use a power law with F_λ(λ) = F_λ(1000 Å) (λ/1000 Å)^−α. Although Kraemer et al. (1998) found a modest amount of internal extinction in the narrow-line region (NLR) of NGC 5548, $E(B-V)\,={0.07}_{-0.06}^{+0.09}$, the continua of Type 1 AGNs in general show little extinction (Hopkins et al. 2004), and none was required in fitting the broadband spectral energy distribution (SED) of NGC 5548 (Mehdipour et al. 2015). Given that our continuum model is simply an empirical characterization of its shape, we assume there is no internal extinction in NGC 5548, and we redden the power law with Galactic foreground extinction of E(B − V) fixed at 0.017 mag (Schlegel et al. 1998; Schlafly & Finkbeiner 2011) using the mean Galactic extinction curve of Cardelli et al. (1989) with R_V = 3.1. Weak, blended Fe ii emission is expected at λ > 1550 Å, which we include as modeled by Wills et al. (1985) and broadened with a Gaussian with FWHM = 4000 km s⁻¹. This model component is essentially a smooth, low-level addition to the continuum at long wavelengths, and it has no predicted or observed spectral features associated with it. Also, as revealed in prior monitoring campaigns on NGC 5548, the Fe ii varies only weakly (Krolik et al. 1991; Vestergaard & Peterson 2005) on timescales of weeks. During AGN STORM, the optical Fe ii emission varied by at most 10% from its mean value (Pei et al. 2017). In the UV model, we therefore keep its intensity fixed, and we normalize its flux using the modeled Fe ii emission from the mean XMM-Newton Optical Monitor grism spectrum of Mehdipour et al. (2015).

For the emission lines, we use multiple Gaussian components. We do not assign any particular physical significance to most of these individual kinematic components, especially since there is no unique way to decompose these line profiles using such nonorthogonal elements. However, the narrow and intermediate-width components of Lyα, N v, C iv, and He ii are discernible as discrete entities in prior observations of NGC 5548 in faint states (Crenshaw et al. 2009). Although narrow Si iv was not present in the 2004 STIS spectrum of Crenshaw et al. (2009), our high-S/N mean spectrum requires it, and we include it in our fit. Similarly, there is a nonvarying intermediate-width component with FWHM ∼ 800 km s⁻¹ in Lyα, N v, Si iv, C iv, and He ii. We call this the intermediate-line region (ILR). In weaker emission lines, such as C iii* λ1176, Si iii λ1260, Si ii + O i λ1304, C ii λ1335, N iv] λ1486, O iii] λ1663, and N iii] λ1750, this intermediate-width component is the only one observed in the spectrum.

Crenshaw et al. (2009) saw little or no evidence for variability of the NLR and ILR components. We allow these to vary freely in determining our best fit to the mean spectrum, but for the bright emission lines (Lyα, N v, Si iv, C iv, and He ii), we keep these components fixed when fitting the individual spectra from the campaign. For the weaker, lower-ionization emission lines listed above, however, we allow their flux, central wavelength, and FWHM to vary, since these lines are not heavily blended with other components.

The strongest emission lines (Lyα, N v, Si iv, C iv, and He ii) all require up to three additional broad components. These have approximate widths of 3000 (broad, or B), 8000 (medium broad, or MB), and 15,000 (very broad, or VB) km s⁻¹. The N v, Si iv, and C iv emission lines are all doublets. We allow for independent narrow, intermediate, B, and MB components for each doublet transition. In each case, we link their wavelengths at the ratio of their vacuum values, assign them the same FWHM, and assume that their relative fluxes have an optically thick 1:1 ratio. For the VB component, however, which is much broader than the doublet separations, we use only a single Gaussian for each ion.

A final set of empirical emission components in our model accounts for weak bumps on the red and blue wings of C iv and the red wing of Lyα. These bumps have an interesting variability pattern that we discuss later, and they are especially visible in the rms spectrum shown by De Rosa et al. (2015) in their Figures 1 and 2. We use Gaussian components for each of these features.

We do not model the narrow, intrinsic absorption lines in NGC 5548 or the foreground interstellar lines, but we do model the variable intrinsic broad absorption associated with the obscurer discovered by Kaastra et al. (2014). The strongest, most easily modeled broad absorption features are on the blue wings of the Lyα, N v, Si iv, and C iv emission lines. For these, like Kaastra et al. (2014), we use an asymmetric Gaussian with negative flux. To specify the asymmetry, we use a larger dispersion on the blue side of the central wavelength than on the red side. We allow the ratio of blue to red dispersion to vary as a free parameter. This parameterization produces a rounded triangular shape with the deepest point in the trough near the red extreme and a blue wing that extends far out along the blue wing of the emission line. (Kaastra et al. 2014 showed these profiles in their Figure 2.) We emphasize that this absorption profile is strictly an empirical characterization of the observed flux in the spectrum that has no independent physical meaning. Deriving physical information requires making further assumptions about which emission components are covered, producing a transmission profile, and integrating that profile to obtain the actual opacity. We discuss these measurements later in Section 3.7.

In addition to these main troughs, additional small depressions appear further out on the blue wings. We model these with additional symmetric Gaussians in negative flux. As with the emission lines, the absorption in N v, Si iv, and C iv is due to doublets. These are unresolved, and we assume they are optically thick. We model each line in the doublet using the same shape and depth. Finally, all of the UV resonance lines in the spectrum have weak, blueshifted absorption troughs at velocities comparable to the main portion of the troughs observed in Lyα, N v, Si iv, and C iv. These cannot be modeled in the detail that we apply to the strongest absorption troughs, and they are only readily apparent in the mean spectrum. For these troughs, we use symmetric Gaussians in negative flux.

The final component of our model is absorption of all model components by damped Lyα from the Milky Way. We fix the column density at N(H i) = 1.45 × 10²⁰ cm⁻² (Wakker et al. 2011).

Figure 1 gives a detailed view of the best-fit model overlaid on the data. We illustrate all individual components of the model, the best-fit model overlaid on the data, and the absorption-corrected model (which is the ultimate goal of our efforts). Note that the C iv absorption is not as deep in the mean spectrum from the reverberation campaign as it was during the deepest phase of the obscuration as observed in the XMM-Newton campaign (see Figure S1 in Kaastra et al. 2014, which shows the individual components of the region surrounding the C iv emission line). For further illustrations of our model of the mean spectrum, see Figures 1 and 2 of De Rosa et al. (2015), which compare the model to the full G130M and G160M spectra.

Zoom In Zoom Out Reset image size

All of the model parameters are listed in Table 1. The model consists of a total of 97 individual components, each with two to four parameters. Although the total number of parameters is 383, many of these are fixed or linked to other parameters. Free parameters in the fit total 143. Although this is a large number, the fit is tightly constrained. The fitted regions in the mean spectrum comprise ∼12,598 points, each approximately half of a resolution element. Thus, each spectrum has ∼6000 spectral elements included in the fit, and that is described by only 143 parameters.

To optimize the parameters of the model and obtain the best fit, we use a combination of minimization algorithms. After determining initial guesses by visual inspection, we start the optimization process using a simplex algorithm (Murty 1983). This works well for problems with many parameters that are not initially well tuned. Once the fit is nearly optimized, this algorithm generally loses efficiency in approaching full convergence. At that point (usually after several tens of iterations), we switch to a Levenberg–Marquardt algorithm (as originally coded by Bevington 1969). When close to an optimum fit, this algorithm converges rapidly, but with the large number of parameters in our fit, it can get stuck in false minima. To escape these pitfalls, we then alternate sets of 5–10 iterations using the simplex algorithm and the Levenberg–Marquardt algorithm until the fit has fully converged. We define full convergence as Δχ² < 0.01 and a change in each parameter value of <1% after a set of iterations.

As we noted above, our Gaussian decomposition of the emission lines is not unique. Therefore, unless one takes care to preserve the overall character of the spectral shape of the model from visit to visit among the individual observations, best fits and parameters can wander far from the character of our fit to the mean spectrum. One could try to avoid this by tailoring initial guesses for fits to each individual spectrum interactively, but this would introduce a unsatisfying degree of subjectivity into our final results. We therefore employed an approach that used the quantitative characteristics of the spectra to tune each individual fit and guide it to an optimum result. We verified the soundness of this approach through multiple trials and experiments before converging on a process that produced consistent results from spectrum to spectrum without drastic changes in parameters that might signify unphysical solutions.

In our first trials, we noticed that most weak emission and absorption features were often too weak to be effectively constrained in a single observation. We therefore produced a series of grouped spectra that improved the S/N for our measurements of these weak spectral features. As described in De Rosa et al. (2015), the observations were done in a cycle of eight, where central wavelength settings and FP-POS positions were changed on a daily basis. Combining spectra according to these natural groups produces better S/N, reduces pattern noise, and provides the full spectral coverage that extends down to the P v region at the short-wavelength end. In the fits to the time series we discuss below, we use the values for the weak features determined from these grouped spectra as our initial guess for starting parameters when fitting the individual spectra in a group.

Another outcome of our trials, perhaps an obvious one, is that good initial choices for parameters led to quicker convergence and less chance of solutions where parameters strayed into unphysical regions of parameter space. Thus, given our exquisite fit to the mean spectrum, its parameters provide the best guess for spectra that are similar. To strengthen this similarity from fit to fit, we tried two different methods. First, we ordered the spectra by the flux level in the 1367 Å continuum window. We then chose the spectrum near the middle of this distribution with the flux most comparable to the flux in the mean spectrum as the first one to fit. The best fit from this spectrum was then used as the initial guesses for parameters in fitting the next spectra in the series. The series of fits followed two parallel paths moving both higher and lower in flux from this initial middle spectrum. Our second method was to keep the spectra ordered by time and then pick a spectrum close to the midpoint of the campaign with a mean continuum close to that of the mean spectrum. Fits then progressed both forward and backward in time from this middle point.

These experiments validated our intentions to develop an objective process for determining the best fit to each spectrum. Both methods achieved good fits for all spectra, and the parameters from each method were close in value to each other (typically within the 1σ uncertainties). Figure 2 compares the values obtained as a function of time for the two different methods for three selected parameters from the model. In Figure 2 and the remainder of the paper, we define times by the Truncated Heliocentric Julian Date (THJD), THJD = HJD−2,400,000.

Zoom In Zoom Out Reset image size

Despite the good agreement in the quality of the fits for the two different methods, our experiments also showed that the second method, "ordered by time," produced better results than "ordered by flux," in the sense that variations in parameter values were smoother, and, as shown in Figure 3, the best-fit χ² was typically slightly less. Despite the significant reduction in χ² we achieved for the fits ordered by time, the differences in the fits are not obvious. We note that for each method, χ² varies systematically with time during the campaign. The variations loosely correspond to the overall variations in brightness of NGC 5548 (as one can see in later figures showing light curves for the continuum and emission line). Our inference for these systematic variations is that the brighter spectra have higher S/N per pixel, and that subtle residual pattern noise in the flat-field properties of the COS detectors degrades the quality of our fits. Figure 4 compares the best-fit models to the data for two extreme cases, our overall lowest χ² solution for visit 38 and our overall highest χ² obtained for visit 60. One is hard-pressed to see the differences between the models fit using either method, or even why χ² is significantly higher for visit 60 compared to visit 38.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Our best inference for why the "ordered by time" sequence produces better results is that the spectrum evolves on timescales of a few days. This is measured in the lags of the emission lines, and, as we will show later in Section 5.2, it is also true for the absorption features. Therefore, although one spectrum might have the same continuum flux as another, if it is separated in time by more than several days, the parameters of major features may differ significantly, making it harder for the minimization algorithm to converge on the best solution.

To fit an individual spectrum, we first determined the best first approximation to the normalization and spectral index of the power-law continuum by fitting only those points identified by De Rosa et al. (2015) as continuum windows. To avoid "contaminating" these windows with broad-line flux, we also set the flux of all of the VB components to zero. In our experiments, we found that if we did not do this, the VB components developed a tendency to grow in width until they formed their own pseudo-continuum across the whole spectrum. In this first pass, only the power-law normalization and index were allowed to vary freely.

After this step, we optimized the strength of the brightest emission lines—Lyα, N v, Si iv, C iv, and He ii—and let the fluxes of their B and MB components and the power-law normalization vary freely, while keeping the power-law index fixed and the VB component turned off. Next, we restored the VB fluxes to their original values, and let the fluxes of all broad components of the above lines vary freely, as well as the power-law normalization.

At this point, the fit formed a remarkably good representation of an individual observation, but it was still far from the best fit. In the next steps, we turned our attention to optimizing the fits for each individual bright emission line. In these separate steps, we kept all parameters not related to the specific spectral region fixed, including the continuum parameters, Fe ii flux, narrow- and intermediate-line components for all lines, and any weak blended lines.

For the Lyα region, we did separate optimization steps in the following order.

• 1.  
  Fit Lyα only (B, MB, VB) using only wavelengths 1150–1245 Å. As before, let the fluxes vary freely first, then both the widths and fluxes. Keep all components of N v fixed throughout this.
• 2.  
  Fit the red wing of N v using only wavelengths 1263–1300 Å. Keep Lyα and all other components fixed. Let the two B components vary freely first, then fix them and let the MB components vary. Finally, fix those and let the VB component vary.
• 3.  
  Fit the N v absorption on its blue wing, keeping the doublet's fluxes tied at a 1:1 ratio, using only wavelengths 1245–1264 Å.
• 4.  
  Free the flux and width of the B, MB, and VB components of Lyα.
• 5.  
  Free the flux and width of the B, MB, and VB components of N v.
• 6.  
  Free the flux of the Lyα B absorber.
• 7.  
  Free the flux, width, and asymmetry of the Lyα B absorber.

We next did the Si iv region. All parameters that we varied freely above for the optimization of the Lyα region were fixed, and we then did the following.

• 1.  
  Free the flux of the B, MB, and VB components of Si iv.
• 2.  
  Free the flux and width of the B, MB, and VB components of Si iv.
• 3.  
  Free the flux of the Si iv B absorber. (Since the Si iv absorption is so weak, we linked the width and asymmetry of the Si iv B absorber to that of C iv.)

We fit C iv and He ii together, since they are tightly blended. For optimizing this region, we fix all of the previous freely varying parameters, then do the following.

• 1.  
  Free the flux and width of the B, MB, and VB components of C iv.
• 2.  
  Free the flux (but not the width) of the B, MB, and VB components of He ii.
• 3.  
  Free the flux of the C iv B absorber.
• 4.  
  Free the flux, width, and asymmetry of the C iv B absorber.

Once the major emission and absorption components are tuned up, we then allow the weaker emission-line features to adjust. We keep the continuum fixed, as well as all of the parameters associated with Lyα, N v, Si iv, C iv, and He ii. The fluxes of all other weak emission features are then allowed to vary.

To complete the optimization, all parameters designated as free to vary in Table 1 are freed, and we iterate the χ² minimization process until it converges. The best-fit parameters for this spectrum are then used as the initial guesses for doing the fit to the next spectrum in the series (with the exception that the initial guesses for the weak features are taken from the best fit to the grouped spectrum corresponding to that spectrum). Final values of all components as a function of wavelength for each spectrum are available as a high-level science product in the Mikulski Archive for Space Telescopes (MAST)¹¹⁴ as the data set identified by 10.17909/t9-ky1s-j932.

In principle, one can obtain 1σ uncertainties on each of the parameters in our model from the best-fit covariance matrix. However, given the 383 parameters, 143 of which are freely varying, this is computationally impractical. An alternative is to assume that parameter space can be approximated by a parabola near the best-fit minimum in χ². Using numerically calculated first and second derivatives, one can then extrapolate from the minimum changes in each parameter to achieve Δχ² = 1, which corresponds to a 1σ uncertainty for a single interesting parameter (Bevington 1969). Unfortunately, owing to the high dimensionality of our parameter space and its poor sampling in our calculations, this method proved inadequate.

All of the main quantities of interest to be extracted from our models are the fluxes of individual features, either in emission or in absorption. We therefore calculate uncertainties for these quantities using the data and associated uncertainties in each original spectrum and scaling them in proportion to the quantities that we integrate from our models. To explain this quantitatively, first we define these quantities:

• f_data,i = flux in pixel i of the original spectrum,
• σ_data,i = 1σ uncertainty for pixel i in the original spectrum,
• f_mod,i = flux in pixel i of the model spectrum,
• σ_mod,i = 1σ uncertainty for pixel i in the model spectrum,
• f_c,j,i = flux in pixel i of the component j (of 97 total), and
• σ_c,j = 1σ uncertainty for component j.

The flux of the model in pixel i can be decomposed into the sum of the contributions of the individual components:

$$\begin{eqnarray}&&{f}_{\mathrm{mod},i}=\displaystyle \sum _{j}{f}_{c,j,i}.\end{eqnarray} \tag{ 1 }$$

The variance of the model in pixel i is then

$$\begin{eqnarray}&&{\sigma }_{\mathrm{mod},i}^{2}=\displaystyle \sum _{j}{\sigma }_{c,j}^{2}.\end{eqnarray} \tag{ 2 }$$

In the limit of very good statistics, the variance predicted by the model should simply be the variance in the data themselves, i.e.,

$$\begin{eqnarray}&&{\sigma }_{\mathrm{mod},i}^{2}\sim {\sigma }_{\mathrm{data},i}^{2},\end{eqnarray} \tag{ 3 }$$

so the variance in the total flux of any individual component is then

$$\begin{eqnarray}&&{\sigma }_{c,j}^{2}=\displaystyle \sum _{i}{\sigma }_{\mathrm{data},i}^{2}\times {f}_{\mathrm{mod},i}^{2}/\left.\left(\displaystyle \sum _{k\ne j}\right.\right){f}_{c,k,i}^{2}.\end{eqnarray} \tag{ 4 }$$

If the quantity of interest is the sum of multiple components (i.e., the several components of an emission line contributing to the flux in a given velocity bin), then the 1σ uncertainty we derive for that quantity is

$$\begin{eqnarray}&&{\sigma }_{\mathrm{tot}}=\sqrt{\displaystyle \sum _{j}{\sigma }_{c,j}^{2}}.\end{eqnarray} \tag{ 5 }$$

Figure 5 shows the consequences of this method for calculating the 1σ uncertainties. When calculated directly from the data, as described above, the uncertainties are more uniform. The numerical instabilities in our method of interpolating in the error matrix of the fit give uncertainties that largely cluster around the calculation based on the data but show large excursions, both higher and lower.

Zoom In Zoom Out Reset image size

The resulting best-fit χ² for each spectrum is shown as a time series in Figure 3. The number of points in each spectrum varies slightly, since the spectrum is moved to multiple positions on the detector using different central wavelength settings. Because of the differing central wavelength settings, not all spectra cover the same range in wavelength as the mean spectrum; the settings tend to lose several hundred points on the blue and red ends of each spectrum. On average, individual spectra have ∼11,600 points in each fit. With 143 freely varying parameters, given the χ² values ranging from ∼6800 to ∼9000 in Figure 3, one can see that our uncertainties are too large. This is most likely because in aligning and merging each spectrum, we have resampled the original pixels, introducing correlated errors in adjacent bins after rebinning. Note also that χ² varies systematically with time in the campaign, largely following the light curve of total brightness. A likely explanation for this trend is that it is easier to get a good fit to our complex model when the fluxes are lower and uncertainties are larger.

While the χ² shows that we have good fits overall, we visually examined each individual fit to see if there were any points or regions where there were systematic deviations of the model from the data. It was these inspections in our early experiments that led us to develop the fitting strategies we documented above. The final fits show no gross or systematic residuals. This is illustrated visually by the image in Figure 6 that shows the residuals of Data − Model for each fit stacked into a two-dimensional spectrogram. The only significant features visible are the vertical lines at the positions of interstellar and intrinsic narrow absorption lines, which are not part of our model.

Zoom In Zoom Out Reset image size

Our final set of tests compared light curves of fluxes extracted from our model fits to integrations of the same regions of data used in De Rosa et al. (2015). Again, here we see no systematic deviations, but these figures do illustrate how the continuum regions are slightly contaminated by wings of the broad emission lines. Figure 7 compares light curves for continuum windows integrated from the raw data, as described by De Rosa et al. (2015), to integrations of the same wavelength regions in our models, both for the full model and for just the power-law continuum. The lower half of each panel in the figure shows the differences between the two curves along with uncertainties from the raw data. These uncertainties include the systematic repeatability errors described by De Rosa et al. (2015) that apply to the time-series analysis of fluxes from the campaign. These are δ_P = 1.1% for data with λ < 1425 Å and δ_P = 1.4% for λ > 1425 Å. One can see here that the cleanest continuum window, i.e., the one with the least contamination by surrounding emission lines, is the shortest-wavelength window surrounding 1158 Å. Although this is the cleanest window in terms of total flux, we use the modeled continuum flux at 1367 Å in our subsequent analysis. Since this is deterministically connected to the modeled flux at 1158 Å through the continuum model, there is no difference between using one or the other.

Zoom In Zoom Out Reset image size

One of our main goals for these fits to the emission model of each spectrum is to correct for the effects of intrinsic broad absorption and also to bridge the regions affected by foreground interstellar absorption lines and the narrow intrinsic absorption features in NGC 5548 itself. To correct for the broad absorption, we simply apply the inverse of these model elements to the data. For each pixel corrected in this way, we apply the same scaling to the associated uncertainty, as well as the data themselves. To correct for the narrow absorption features (both foreground and intrinsic), since these are not modeled, we replace the data in the wavelength regions affected by the absorption with the emission model.

More specifically, we first visually examine the fit to the mean spectrum to identify points that were significantly affected by narrow absorption lines. These intervals and their identifications are listed in Table 2. Next, we define the following quantities:

• f_orig = flux in the original spectrum,
• f_mod = flux in the model spectrum (including broad absorption),
• f_abs = model flux (negative) in the broad absorption lines,
• t_G = transmission profile of Galactic damped Lyα, and
• f_cor = flux in the corrected spectrum.

The corrected spectrum is then computed in two steps. First, we replace all pixels in the original spectrum with f_cor if they fall within the wavelength intervals defined in Table 2. We then compute f_cor = (f_orig − f_abs)/t_G.

To calculate the 1σ uncertainties, we scale the original uncertainties in each pixel by the ratio of the corrected flux to the original flux:

$$\begin{eqnarray}&&{\sigma }_{\mathrm{cor}}={\sigma }_{\mathrm{orig}}\times ({f}_{\mathrm{cor}}/{f}_{\mathrm{orig}}).\end{eqnarray} \tag{ 6 }$$

With our model fits to the entire series of spectra from the STORM campaign, we can now extract absorption-corrected spectra for all emission lines across their full, deblended velocity profiles. Our models also allow us to separate the variable and nonvariable components of the strong emission lines, as well as deblend adjacent lines. For example, Crenshaw et al. (2009) were able to use the faint state of NGC 5548 in 2004 to separate and measure the narrow- and intermediate-line width components of the Lyα and C iv emission lines. They demonstrated that these components vary only slightly over timescales of years. Using our model, we are able to exclude these nonvarying components of the emission lines from the overall broad-line profile. Likewise, we can separate the contributions of blended lines from the wings of Lyα and C iv and measure lines such as N v as individual species.

For each individual broad emission line, we construct a model profile at each individual pixel i that includes only the contributions of the relevant broad-line components from our model. As described in Section 3.3, the net flux associated with a given emission feature is

$$\begin{eqnarray}&&{f}_{\mathrm{totem},i}=\displaystyle \sum _{j}{f}_{c,j,i},\end{eqnarray} \tag{ 7 }$$

where the index j runs over all components associated with the desired emission feature. As described in Section 3.3, we calculate the associated 1σ statistical uncertainty as the fraction in quadrature (relative to all model components in that pixel) of the 1σ uncertainty on the data in that pixel,

$$\begin{eqnarray}&&{\sigma }_{\mathrm{totem},i}={\sigma }_{i}\sqrt{\displaystyle \frac{{\displaystyle \sum }_{j}{f}_{c,j,i}^{2}}{{\displaystyle \sum }_{k}{\text{}}{f}_{c,k,i}^{2}}},\end{eqnarray} \tag{ 8 }$$

where the index j runs over all components contributing to the desired emission feature, and the index k runs over all components of the model contributing to the flux in pixel i. As described in detail by De Rosa et al. (2015), systematic repeatability errors affect the data when considering a time-series analysis of these quantities. We therefore add in quadrature the same errors in precision, namely, δ_P = 1.1% for data from grating G130M (λ < 1425 Å) and δ_P = 1.4% for data from grating G160M (λ > 1425 Å). These errors in reproducibility actually dominate the uncertainties for all quantities at λ < 1180 and λ > 1425 Å.

Figure 8 compares light curves for the modeled continuum flux at 1367 Å to the deblended broad emission lines of Lyα, N v, Si iv, C iv, and He ii. Portions of these light curves are tabulated in Table 3, with full tabulations of these quantities and the other continuum windows (1158, 1430, and 1740 Å) available online. The light curves in Figure 8 closely resemble those derived from the original data shown in Figure 3 of De Rosa et al. (2015). In general, Lyα, Si iv, and C iv are brighter due to the corrections for absorption, the additional flux from the wings of the VB emission components, and the elimination of contaminating emission-line flux from the original continuum windows. For He ii, the flux levels are roughly the same; additional flux from the VB component of the emission line and a less contaminated continuum are offset by subtraction of blended emission from C iv. Overall, the error bars are smaller, since the modeled flux for any given component is determined by many more pixels than the limited wavelength range used in the original integrations. The light curve for N v is a new addition enabled by the deblending from Lyα in our model. In some respects, N v differs in character from the other emission lines, especially during the first 75 days of the campaign prior to the BLR holiday. However, starting with the BLR holiday, its behavior is very similar to that of C iv and He ii. We will quantify these similarities and differences in Section 4.1 when we discuss the emission-line lags.

Zoom In Zoom Out Reset image size

Table 3.  Modeled Continuum and Deblended Emission-line Light Curves for NGC 5548

HJDa	${F}_{\lambda }\left(1367\ \mathring{\rm A} \right)$ b	F(Lyα)c	F(N v)c	F(Si iv)c	F(C iv)c	F(He ii)c
56,690.6120	29.71 ± 0.34	66.83 ± 0.76	10.14 ± 0.12	9.03 ± 0.10	63.16 ± 0.90	5.60 ± 0.08
56,691.5416	32.13 ± 0.36	65.48 ± 0.73	9.80 ± 0.11	8.66 ± 0.10	62.46 ± 0.88	6.13 ± 0.09
56,692.3940	34.50 ± 0.39	69.99 ± 0.78	7.89 ± 0.09	8.16 ± 0.09	61.50 ± 0.87	5.82 ± 0.08
56,693.3237	34.69 ± 0.39	65.72 ± 0.73	9.00 ± 0.11	8.37 ± 0.09	62.78 ± 0.88	5.79 ± 0.08
56,695.2701	36.78 ± 0.41	66.80 ± 0.74	10.09 ± 0.11	8.92 ± 0.10	62.69 ± 0.89	6.67 ± 0.10
56,696.2459	40.76 ± 0.46	66.78 ± 0.74	9.72 ± 0.11	7.73 ± 0.09	61.19 ± 0.87	6.12 ± 0.09
56,697.3080	42.63 ± 0.48	69.09 ± 0.77	8.74 ± 0.11	7.56 ± 0.08	63.59 ± 0.89	5.52 ± 0.08
56,698.3041	44.49 ± 0.50	68.66 ± 0.76	9.78 ± 0.11	7.69 ± 0.09	64.00 ± 0.90	6.30 ± 0.09
56,699.2338	43.14 ± 0.48	70.28 ± 0.78	9.73 ± 0.11	7.99 ± 0.09	64.20 ± 0.91	6.75 ± 0.10
56,700.2299	42.82 ± 0.48	71.07 ± 0.79	8.47 ± 0.10	7.61 ± 0.09	64.28 ± 0.91	5.80 ± 0.08

Notes. Modeled light curves are in the observed frame. Flux uncertainties include both statistical and systematic errors.

^aMidpoint of the observation (HJD − 2,400,000). ^bUnits of 10⁻¹⁵ erg s⁻¹ cm⁻² Å⁻¹. ^cUnits of 10⁻¹³ erg s⁻¹ cm⁻².

Only a portion of this table is shown here to demonstrate its form and content. A machine-readable version of the full table is available.

Download table as:  DataTypeset image

The intrinsic narrow absorption lines comprise six discrete velocity components. We adopt the nomenclature of Mathur et al. (1999), numbering each component in order starting at the highest blueshifted velocity. To illustrate this kinematic structure, Figure 9 shows normalized absorption profiles for the most prominent intrinsic absorption lines as a function of velocity relative to the systemic velocity of the host galaxy NGC 5548. For this, we adopt the H i 21 cm redshift of z = 0.017175 (de Vaucouleurs et al. 1991).

Zoom In Zoom Out Reset image size

Measuring the strengths of the intrinsic narrow absorption lines is straightforward. Using the complete model of each spectrum (including the broad absorption components, since they help define the local continuum surrounding the narrow intrinsic absorption lines), we measure the equivalent widths (EWs) by integrating across each absorption-line profile in the normalized spectrum. These integrations are performed as discrete sums over pixels lying within the wavelength regions defined for each feature in Table 4:

$$\begin{eqnarray}&&\mathrm{EW}=\displaystyle \sum _{i}({f}_{\mathrm{orig},i}-{f}_{\mathrm{mod},i})/{f}_{\mathrm{mod},i}\times {\rm{\Delta }}\lambda .\end{eqnarray} \tag{ 9 }$$

The 1σ uncertainty for EW is obtained by simply propagating the uncertainty associated with each data point in the original spectrum used in the sum:

$$\begin{eqnarray}&&{\sigma }_{\mathrm{EW}}=\sqrt{\displaystyle \sum _{i}({\sigma }_{\mathrm{orig},i}^{2}{/{f}_{\mathrm{orig},i}}^{2})\times {\rm{\Delta }}{\lambda }^{2}}.\end{eqnarray} \tag{ 10 }$$

Table 4.  Properties of Narrow Intrinsic Absorption Lines in NGC 5548

Feature	λoa	λ1b	λ2c	EWd	ve	cff	log Niong
	(Å)	(Å)	(Å)	(km s−1)	(Å)		(cm−2)
P vb 1	1117.98	1132.30	1133.05	−0.167 ± 0.015	−1207	1.00	${13.57}_{-0.20}^{+0.42}$
Fe iii 1	1122.52	1136.99	1137.60	−0.037 ± 0.004	−1205	1.00	${13.63}_{-0.27}^{+0.30}$
Fe iii* 1	1124.87	1139.31	1140.01	−0.031 ± 0.004	−1199	1.00	${13.73}_{-0.38}^{+0.20}$
P vr 1h	1126.72	1142.49	1143.33	−0.200 ± 0.004	−828	1.00	${13.57}_{-0.20}^{+0.42}$
C iii* 1	1175.26	1190.54	1190.98	−0.122 ± 0.002	−1195	0.65	${13.86}_{-0.24}^{+0.17}$
C iii* 1	1175.71	1190.98	1191.59	−0.183 ± 0.002	−1189	0.65	${14.23}_{-0.28}^{+0.29}$
C iii* 1	1176.37	1191.59	1192.35	−0.130 ± 0.002	−1190	0.65	${14.49}_{-0.24}^{+0.17}$
S iii 1i	1190.20	1205.50	1207.06	−0.571 ± 0.003	−1093	0.65	${14.37}_{-0.38}^{+0.10}$
Si ii 1	1193.29	1208.55	1209.38	−0.023 ± 0.002	−1196	0.34	${12.99}_{-0.49}^{+0.25}$
S iii* 1	1194.06	1209.53	1210.05	−0.030 ± 0.002	−1187	0.34	${14.37}_{-0.27}^{+0.32}$
Si ii* 1	1194.50	1210.05	1210.61	−0.034 ± 0.002	−1194	0.34	${13.25}_{-0.25}^{+0.25}$
Si ii* 1	1197.39	1212.95	1213.49	−0.026 ± 0.003	−1166	0.34	${13.00}_{-1.0}^{+0.25}$
Si iii 1	1206.50	1221.76	1223.08	−0.124 ± 0.003	−1176	0.30	${13.45}_{-0.62}^{+0.06}$
Si iii 3	1206.50	1224.20	1225.07	−0.048 ± 0.002	−638	0.30	${12.91}_{-0.61}^{+0.09}$
Si iii 5	1206.50	1225.07	1225.84	−0.028 ± 0.002	−454	0.30	${12.67}_{-0.61}^{+0.15}$
Lyα 1	1215.67	0.017175	1232.98	−0.845 ± 0.002	−1156	0.70	>14.60
Lyα 2	1215.67	0.017175	1233.54	−0.301 ± 0.001	−785	0.90	>14.07
Lyα 3	1215.67	0.017175	1234.21	−0.515 ± 0.001	−659	0.90	>14.37
Lyα 4	1215.67	0.017175	1235.19	−0.791 ± 0.001	−470	0.95	>14.56
Lyα 5	1215.67	0.017175	1236.10	−0.182 ± 0.002	−299	0.70	>13.97
Lyα 6	1215.67	0.017175	1236.73	−0.105 ± 0.002	−22	0.70	${13.00}_{-1.0}^{+0.3}$
N vb 1	1238.82	1254.21	1256.46	−0.776 ± 0.003	−1147	0.80	>14.94
N vb 2	1238.82	1256.46	1256.95	−0.115 ± 0.001	−791	0.70	>14.14
N vb 3	1238.82	1256.95	1257.82	−0.637 ± 0.001	−648	0.90	>14.87
N vb 4	1238.82	1257.82	1258.57	−0.555 ± 0.001	−478	0.95	>14.78
N vr 2	1242.80	1260.70	1260.98	−0.051 ± 0.001	−789	0.70	>13.93
N vr 3	1242.80	1260.98	1261.88	−0.536 ± 0.001	−645	0.90	>14.98
N vr 4	1242.80	1261.88	1262.63	−0.482 ± 0.001	−475	0.95	>14.96
N vr 5	1242.80	1262.63	1263.56	−0.230 ± 0.002	−302	0.70	>14.74
N vr 6	1242.80	1263.67	1264.34	−0.095 ± 0.002	−35	0.70	>14.25
Si ii 1	1260.42	1276.54	1277.50	−0.046 ± 0.003	−1193	0.28	${13.07}_{-0.57}^{+0.17}$
Si ii* 1	1264.74	1280.93	1282.20	−0.062 ± 0.003	−1166	0.28	${13.08}_{-0.58}^{+0.45}$
Si iii* 1	1296.73	1313.40	1314.60	−0.001 ± 0.004	−1600	0.90	${13.00}_{-1.0}^{+0.3}$
Si iii* 1	1298.96	1315.66	1316.56	−0.026 ± 0.003	−1185	0.90	${13.00}_{-1.0}^{+0.30}$
Si iii* 1	1303.32	1320.15	1320.80	−0.008 ± 0.003	−1195	0.90	${12.70}_{-0.7}^{+0.30}$
Si ii 1	1304.37	1321.10	1322.00	−0.015 ± 0.003	−1175	0.92	${13.09}_{-0.59}^{+0.15}$
Si ii* 1	1309.28	1326.18	1327.14	−0.020 ± 0.003	−1179	0.92	${13.09}_{-0.59}^{+0.15}$
C ii 1	1334.53	1351.56	1352.75	−0.078 ± 0.003	−1185	1.00	${13.61}_{-0.11}^{+0.38}$
C ii* 1	1335.71	1352.75	1353.90	−0.100 ± 0.003	−1194	1.00	${13.61}_{-0.11}^{+0.38}$
P iii* 1	1344.33	1361.31	1362.61	−0.009 ± 0.004	−1270	1.00	${13.20}_{-0.21}^{+0.76}$
Si ivb 1	1393.76	1411.59	1412.93	−0.260 ± 0.003	−1181	0.50	${14.11}_{-0.06}^{+0.11}$
Si ivb 3	1393.76	1413.78	1415.20	−0.178 ± 0.003	−644	1.00	${13.32}_{-0.05}^{+0.05}$
Si ivb 4	1393.76	1415.20	1416.04	−0.071 ± 0.003	−455	1.00	${12.90}_{-0.09}^{+0.10}$
Si ivb 5	1393.76	1416.04	1416.47	−0.040 ± 0.002	−305	1.00	${12.66}_{-0.07}^{+0.07}$
Si ivb 6	1393.76	1416.47	1416.89	−0.021 ± 0.002	−228	1.00	${12.37}_{-0.15}^{+0.11}$
Si ivr 1	1402.77	1420.68	1422.05	−0.210 ± 0.003	−1179	0.50	${14.18}_{-0.13}^{+0.04}$
Si ivr 3	1402.77	1423.22	1424.39	−0.069 ± 0.003	−631	1.00	${13.18}_{-0.12}^{+0.11}$
Si ivr 4	1402.77	1424.39	1425.24	−0.021 ± 0.003	−451	1.00	${12.67}_{-0.38}^{+0.20}$
Si ivr 5	1402.77	1425.24	1425.67	−0.016 ± 0.002	−295	1.00	${12.54}_{-0.21}^{+0.15}$
Si ivr 6	1402.77	1425.67	1426.09	−0.011 ± 0.002	−216	1.00	${12.38}_{-0.25}^{+0.20}$
Si ii 1	1526.71	1546.53	1547.21	−0.023 ± 0.003	−1190	0.70	${13.08}_{-0.58}^{+0.16}$
Si ii* 1	1533.45	1553.26	1554.18	−0.039 ± 0.003	−1193	0.70	${13.30}_{-0.16}^{+0.14}$
C ivb 1	1548.19	1567.55	1570.26	−0.724 ± 0.004	−1162	0.70	>14.62
C ivb 2	1548.19	1570.26	1570.84	−0.169 ± 0.002	−793	0.70	>14.05
C ivb 4	1548.19	1571.87	1572.91	−0.704 ± 0.001	−478	0.95	>14.54
C ivr 4	1550.77	1574.40	1575.51	−0.551 ± 0.002	−478	0.95	>14.62
C ivr 5	1550.77	1575.51	1576.73	−0.131 ± 0.003	−322	0.50	>14.32
C ivr 6	1550.77	1576.84	1577.51	−0.005 ± 0.003	0	0.50	${12.50}_{-0.5}^{+0.5}$

Notes.

^aRest wavelength. ^bStarting wavelength for EW integration. ^cEnding wavelength for EW integration. ^dThe EW of the absorption feature. ^eVelocity of the feature relative to the systemic redshift of the host galaxy, z = 0.017175. ^fCovering factor of the absorption feature. ^gColumn density of the absorption feature. ^hBlended with Galactic Fe ii λ1143. ⁱBlended with Galactic Si iii λ1206.

Download table as:  ASCIITypeset images: 1 2

We use caution in performing these integrations to avoid features blended with Galactic absorption lines or other transitions. For example, the close velocity spacings of the C iv and N v doublets (498 and 964 km s⁻¹, respectively) cause components 1, 2, 3, and 5 in C iv to overlap and components 1 and 5 in N v to overlap, as shown in Figure 10. The red transition of component 1 in the N v doublet is blended with Galactic S ii λ1259, and the blue transition of component 6 is blended with Si ii λ1260. Therefore, Table 4 gives measurements only for clean, unblended features.

Zoom In Zoom Out Reset image size

The EW of each broad absorption feature is calculated from the normalized modeled spectrum,

$$\begin{eqnarray}&&{f}_{\mathrm{norm},i}={f}_{\mathrm{mod},i}/({f}_{\mathrm{mod},i}-{f}_{\mathrm{abs},i}),\end{eqnarray} \tag{ 11 }$$

as

$$\begin{eqnarray}&&\mathrm{EW}=\displaystyle \sum _{i}(1-{f}_{\mathrm{norm},i})\times ({\lambda }_{i+1}-{\lambda }_{i}),\end{eqnarray} \tag{ 12 }$$

where λ_i is the wavelength of pixel i. Since our spectra are linearized, Δλ = λ_i+1 − λ_i is actually a constant. The corresponding 1σ uncertainty is

$$\begin{eqnarray}&&{\sigma }_{\mathrm{EW}}=({f}_{\mathrm{totabs}}/\mathrm{EW})\times {\sigma }_{\mathrm{totabs}},\end{eqnarray} \tag{ 13 }$$

where f_totabs and σ_totabs are defined below.

The broad UV absorption troughs associated with the obscurer in NGC 5548 are shown in Figure 2 of Kaastra et al. (2014). These broad troughs are asymmetric, and they extend from near zero velocity in the systemic frame of the host galaxy to ∼−5500 km s⁻¹. The time-varying strengths of the intrinsic broad absorption lines in NGC 5548 that are associated with the obscurer are part of the models we have fit to all of the spectra. These all have one main component, but there are also weaker components on the high-velocity blue wing of the absorption profile. These individual weak components are often not well constrained by the model fits. We therefore calculate the total absorption, f_totabs, for the sum of all components associated with a given spectral transition. The total flux is then

$$\begin{eqnarray}&&{f}_{\mathrm{totabs}}=\displaystyle \sum _{j}{f}_{c,j},\end{eqnarray} \tag{ 14 }$$

where the index j runs over all components associated with the desired line, and the associated uncertainty is calculated as the quadrature sum of the 1σ uncertainties of each component j:

$$\begin{eqnarray}&&{\sigma }_{\mathrm{totabs}}=\sqrt{\displaystyle \sum _{j}{\sigma }_{c,j}^{2}}.\end{eqnarray} \tag{ 15 }$$

Table 5 shows sample portions of the light curves for the broad absorption in C iv and the intrinsic narrow absorption associated with C ii λ1334. Full light curves for all features listed in Table 4 are published in the online version of this paper. Figure 11 shows sample light curves for the EWs of the narrow absorption features associated with component 1 for C ii λ1334, Si iii λ1206, and Si iv λ1393, plus component 3 for Si iv λ1393, all compared to the UV continuum flux at 1367 Å. Light curves for the broad absorption in C iv, N v, Lyα, and Si iv are shown in Figure 12.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Our absorption-corrected, deblended emission-line profiles described in Section 3.6 allow us to remove the uncertainties in emission-line lags that may have been introduced by the variable intrinsic absorption in NGC 5548, as well as to separate the behaviors of adjacent blended lines. In addition, for the brightest two lines, Lyα and C iv, we can determine velocity-binned lags for each line uncontaminated by absorption or blended contributions from other lines.

Following De Rosa et al. (2015), we measured the emission-line lags for the species tabulated in Table 3 by cross-correlating the time series with the continuum light curve using the interpolated cross-correlation function (ICCF) as implemented by Peterson et al. (2004). The procedure and resulting associated uncertainties are described in detail by De Rosa et al. (2015). Briefly, the technique uses a Monte Carlo method of "flux randomization and random subset selection" to generate a large set of realizations of the light curves. For each realization, we determine the cross-correlation function, its maximum correlation coefficient r_max, and associated peak lag τ_peak. We also use the region surrounding the peak with r(τ) > 0.8 to calculate the centroid of the cross-correlation function, τ_cent. A few thousand realizations of each cross-correlation function then gives distribution functions for τ_peak and τ_cent from which we measure the median values to give the lags for each emission line as presented in Table 6. The associated uncertainties represent the 68% confidence intervals of each Monte Carlo distribution.

Table 6.  Emission-line Lags from the Modeled NGC 5548 Spectra

Emission Line	Lyα	N v	Si iv	C iv	He ii
Whole Campaign–Modeled Data
τcenta	5.1 ± 0.3	7 ± 8	8.1 ± 0.7	5.8 ± 0.5	2.2 ± 0.3
τpeakb	5.1 ± 0.6	6 ± 7	8.1 ± 1.0	5.7 ± 0.6	1.9 ± 0.4
rpeakc	0.71 ± 0.03	0.17 ± 0.06	0.16 ± 0.04	0.39 ± 0.04	0.56 ± 0.03
Pre-BLR Holiday, THJD = 56,691–56,765
τcenta	4.8 ± 0.3	⋯	8.0 ± 0.5	4.4 ± 0.3	2.4 ± 0.4
τpeakb	4.8 ± 0.4	⋯	8.1 ± 0.6	4.5 ± 0.5	2.2 ± 0.4
rpeakc	0.94 ± 0.01	⋯	0.7 ± 0.04	0.91 ± 0.02	0.91 ± 0.02
BLR Holiday, THJD = 56,766–56,829
τcenta	5 ± 1	4.5 ± 0.4	7.3 ± 1	7.1 ± 0.6	2.1 ± 0.4
τpeakb	5 ± 1	4.4 ± 0.6	7.3 ± 1	7.3 ± 0.7	2.1 ± 0.5
rpeakc	0.74 ± 0.06	0.79 ± 0.04	0.66 ± 0.06	0.73 ± 0.07	0.75 ± 0.05
Post-BLR Holiday, THJD = 56,830–56,866
τcenta	6 ± 1	5 ± 3	10 ± 2	8 ± 1	7 ± 5
τpeakb	7 ± 1	${2}_{-4}^{+1}$	10 ± 3	8.5 ± 2.2	${10}_{-8}^{+1}$
rpeakc	0.85 ± 0.04	0.72 ± 0.06	0.77 ± 0.19	0.80 ± 0.13	0.63 ± 0.09
Whole Campaign–Original Data
τcenta	6.2 ± 0.3	⋯	5.3 ± 0.7	5.3 ± 0.5	2.5 ± 0.3
τpeakb	6.1 ± 0.4	⋯	5.4 ± 1.1	5.2 ± 0.7	2.4 ± 0.6
rpeakc	0.77 ± 0.02	⋯	0.46 ± 0.06	0.36 ± 0.04	0.65 ± 0.03
Pre-BLR Holiday, THJD = 56,691–56,765
τcenta	5.8 ± 0.3	⋯	5.2 ± 0.8	4.3 ± 0.3	2.4 ± 0.4
τpeakb	5.9 ± 0.4	⋯	5 ± 1	4.4 ± 0.5	${1.4}_{-0.1}^{+0.9}$
rpeakc	0.94 ± 0.01	⋯	0.7 ± 0.04	0.93 ± 0.02	0.90 ± 0.02
BLR Holiday, THJD = 56,766–56,829
τcenta	5.2 ± 0.6	⋯	6 ± 2	6.9 ± 0.9	3.3 ± 0.6
τpeakb	${5}_{-0.4}^{+1}$	⋯	6 ± 2	7 ± 1	3.3 ± 0.8
rpeakc	0.84 ± 0.04	⋯	0.52 ± 0.09	0.64 ± 0.09	0.74 ± 0.06
Post-BLR Holiday, THJD = 56,830–56,866
τcenta	7 ± 1	⋯	7 ± 1	8 ± 1	3 ± 1
τpeakb	8 ± 2	⋯	7 ± 2	8 ± 2	3 ± 1
rpeakc	0.83 ± 0.08	⋯	0.75 ± 0.11	0.82 ± 0.12	0.81 ± 0.05

Notes. Delays measured in days in the rest frame of NGC 5548.

^aCentroid of the ICCF lag distribution for r > 0.8r_peak. ^bPeak lag. ^cPeak correlation coefficient.

Download table as:  ASCIITypeset image

While De Rosa et al. (2015) arbitrarily split the data set in two midway through the campaign, we now know that a more logical breaking point for examining any changes is at day 75 in the campaign, which is the beginning of the period when the broad emission-line fluxes become decorrelated from the continuum variations (Goad et al. 2016), also known as the BLR holiday. We therefore quote lags not only for the full campaign but also for the first 75 days, the preholiday period, when the emission-line and continuum fluxes correlated normally; for the holiday period, days 76–129; and for the postholiday period at the end of the campaign. Comparing the lags in Table 6 to De Rosa et al. (2015), we see that Lyα is slightly shorter, Si iv is longer, and C iv and He ii are about the same. Within the error bars, the lags for the UV model data set are consistent with the prior results using the original data.

Comparing results for the different time intervals within the campaign reveals an interesting evolution in the emission-line lags. As expected, during the preholiday period, when the emission-line fluxes correlate well with the continuum fluctuations, correlation coefficients are high, exceeding r = 0.9 for Lyα, C iv, and He ii. For N v, however, the correlation is so poor that we cannot determine a lag in the preholiday period, as expected from the lack of any strong features in this region of the light curve in Figure 8. During the holiday period, correlation coefficients are lower, but they are still very good, with r > 0.65 for all lines. Lags for all lines during the holiday are about the same as for the preholiday period, except for C iv. The C iv lag during the holiday is significantly longer than for the preholiday period by almost 3 days. In the postholiday period, correlation coefficients are again very good, and C iv shows significantly longer lags compared to the preholiday period. Other lines hint at such a difference, but not significantly. These changing lags with time explain why the correlation coefficients for the overall campaign are low despite the much longer data set. Together with changes in the velocity-resolved lags discussed below, this may indicate that significant changes in the structure (or at least our viewpoint), the illumination, or both of the BLR are occurring on the timescale of our campaign. Given that the orbital timescale at a radius of 1 lt-day in NGC 5548 is 115 days, such changes seem plausible.

To ensure that this apparent increase in the emission-line lags over the course of the campaign is not an artifact of our modeling of the data, we reanalyzed the original data of De Rosa et al. (2015) by splitting it into the same time intervals. The results are given in the bottom half of Table 6. The lags for the whole campaign replicate the original results of De Rosa et al. (2015), and we see the same lengthening of lags toward the end of the campaign with similar values to those found using the modeled data.

Also interesting are the velocity-binned results for Lyα and C iv. Absorption in NGC 5548 largely obscured the inner few thousand km s⁻¹ of the blue side of the profile of each emission line, and N v or He ii emission contaminated the far red wings. Following De Rosa et al. (2015), we use bins of 500 km s⁻¹ spanning each profile. Figure 13 compares the mean spectrum for the modeled broad component of Lyα, its corresponding rms spectrum, and, finally, the velocity-binned profile to the original data from De Rosa et al. (2015). All data are for the full campaign. Our modeled profile provides full velocity coverage across the Lyα emission line. The most noticeable characteristic of the lag profile is its distinct "M" shape, with a local minimum in the lag near zero velocity and maxima on the red and blue sides at ±2500 km s⁻¹. A prominent feature in the rms spectrum is the "red bump" on the Lyα emission-line profile at +1500 km s⁻¹, which loosely corresponds to the local peak in the velocity-dependent lag profile on the red wing of Lyα.

Zoom In Zoom Out Reset image size

Figure 14 shows the corresponding set of results for the C iv emission line. For C iv, there are emission bumps on both the red and blue wings of the profile in the rms spectrum. These bumps are at higher velocity than the red bump in Lyα, at roughly ±5000 km s⁻¹, and they appear to correspond to local minima in the lag profile, as opposed to the maxima seen in Lyα. As with Lyα, C iv shows a slight hint of an "M" shape to its profile, with a shorter lag near the center and local peaks at ±2500 km s⁻¹. However, the contrast is not as distinctive as in Lyα. The central dip in C iv has a confidence level of only ∼90%.

Zoom In Zoom Out Reset image size

Examining the velocity-binned profiles for the separate, distinct time intervals of the campaign reveals even more complex behavior. Figure 15 compares the rms spectra and velocity-dependent lags for C iv and Lyα from the full campaign to the first 75 days (the preholiday period), the period of the BLR holiday, and the postholiday period concluding the campaign. The emission bumps on the red and blue wings of C iv and the red wing of Lyα are most prominent early in the campaign and diminish in flux (or disappear, in the case of C iv red) by the end of the campaign. We show light curves for these features in Figure 16. The red emission bump in C iv also has associated features in the lag profiles that evolve from a local minimum on the red side of the bump during the preholiday period to a local maximum in the lag on the blue side of the bump. These more detailed changes in the emissivity and lag profiles again suggest that we are seeing changes in the structure of the BLR over the course of the campaign. This may be due to the presence of some outflowing components, as discussed in Section 5, but this is speculative and more easily investigated with two-dimensional reverberation maps and models.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The very high S/N of the mean spectrum makes accurate measures of weak features possible. These are particularly useful since they are often unsaturated and can therefore provide better diagnostic information on physical conditions in the absorbing gas. The last four columns of Table 4 give the EW, the velocity relative to the systemic velocity of the host galaxy, the covering factor, and the inferred column density for the narrow absorption lines in the mean spectrum. For Galactic interstellar medium (ISM) features, the inferred column densities are at best lower limits, since the lines are saturated and the profiles have not been corrected for the COS line-spread function. For absorption lines intrinsic to NGC 5548, the line widths are broad enough (FWHM typically >80 km s⁻¹) that the COS line-spread function has little effect. To measure column densities, we integrate the apparent optical depth across the absorption-line profile between the wavelength limits given in Table 4, assuming a uniform covering factor, as given in the next-to-last column of the table (see Appendix A of Arav et al. 2015, for a description of the technique). For doublets and absorption lines with multiple transitions (e.g., P v or Si ii), we can determine the covering factor such that integration of each line profile gives consistent column densities. For other absorption lines, particularly the deep, heavily blended features associated with Lyα, N v, and C iv, we use a covering factor determined by the deepest point of the absorption-line profile and assume the line is saturated. Like Arav et al. (2015), this gives a lower limit on the column density.

Examining the covering factors in Table 4 is instructive. Lines far from the centers of the bright emission lines (e.g., P v, Si ii λ1302, C ii λ1335) have covering factors near unity, indicating that the absorbing gas fully covers, or nearly fully covers, the continuum emission region. These absorption lines are also multiplets, so their covering factors are well determined. Other absorption lines embedded in the profiles of bright emission lines, such as C iii* λ1176, Si ii λ1193, and Si ii λ1260, have well-determined covering factors that are significantly less than unity and vary depending on their distance from the center of the emission line. The sense of this variation is such that covering factors are lower for lines in the brighter portions of the emission-line profile. From this we conclude that, at least for component 1, the absorbing gas nearly fully covers the continuum-emitting region but only covers less than half of the BLR.

The column densities in Table 4 for the STORM campaign mean that the spectra are factors of several lower than those observed by Arav et al. (2015) during the XMM-Newton campaign. This is consistent with the higher brightness of NGC 5548 during the STORM campaign—$\langle {F}_{\lambda 1367}\rangle =4.30\times {10}^{-14}\,\mathrm{erg}\,{\mathrm{cm}}^{-2}\,{{\rm{s}}}^{-1}\,{\mathring{\rm A} }^{-1}$ versus $\langle {F}_{\lambda 1367}\rangle \,=3.11\times {10}^{-14}\,\mathrm{erg}\,{\mathrm{cm}}^{-2}\,{{\rm{s}}}^{-1}\,{\mathring{\rm A} }^{-1}$. Simple scaling of the continuum would imply an increase in the ionization parameter of Δlog U of 0.14. Although this seems small, it is sufficient to account for the observed differences, since the weak, low-ionization species in component 1 are formed in a thin hydrogen ionization front, and their column densities are highly nonlinear with changes in ionizing flux.

We tabulate the properties of the broad absorption features separately, since these are most likely associated with the soft X-ray obscurer discovered by Kaastra et al. (2014). To obtain the EW, mean transmission-weighted velocity, transmission-weighted velocity dispersion, and column density of each broad absorption trough, we use the normalized mean spectrum of NGC 5548. For each trough, we use the highest blueshifted velocity at which the trough drops by more than 1σ below the normalized spectrum. All troughs are integrated up to zero velocity. Since these broad troughs are well resolved, we use the apparent optical depth method of Savage & Sembach (1991) to calculate the column densities of each trough. The deepest absorption troughs in N v, Si iv, and C iv appear to be saturated, since they have similar depths at the velocities of the red and blue members of their respective doublets. Since Lyα is of similar depth, we also assume it is saturated. We therefore measure a covering factor C_f at the deepest point of the trough and use this in our apparent optical depth calculation (Arav et al. 2002). This column density is only a lower limit to the actual column density. The shallower absorption troughs are significantly less deep. If they were saturated and had similar covering factors, they would likely have similar depths to those of the stronger ions. We therefore assume that they lie on the linear portion of the curve of growth and use their apparent optical depths to obtain a direct measure of the column density assuming covering factors of unity. Table 7 summarizes the broad absorption trough properties in detail.

We can now use these measures of ionic column densities in the UV, together with the X-ray opacity measurements from the XMM-Newton spectra of Kaastra et al. (2014), to determine the ionization state of the obscurer more accurately. Because of the low X-ray flux resulting from the heavy X-ray absorption, the XMM-Newton spectra have no detectable spectral features that we can use to determine the ionization properties of the obscurer. However, the X-ray spectrum does provide a good measure of the total column density. As in the study of the obscurer in NGC 3783 (Kriss et al. 2019), we examine joint photoionization models that use the UV ionic column densities along with the total column density determined from the X-ray observations to more precisely determine the physical properties of the obscurer. Since the obscuring gas is illuminated by the bare active nucleus in NGC 5548, we use the unobscured SED shown in Figure 2 of Arav et al. (2015) for our photoionization models. Using Cloudy v17.00 (Ferland et al. 2017), we run a grid of models covering a range of −1.5 to 2.0 in ionization parameter log ξ and total column densities from log N_H = 21.0 to 23.5.¹¹⁵ Figure 17 shows the allowed space of the photoionization solutions, which are at the two points where the X-ray column densities of the two obscurer components intersect the measured column densities of C ii, C iii*, Si ii, and P v in the grid of photoionization models. Note that no single solution fits all measured ions, but some of this incommensurability could be due to the unknown covering fraction of the weak, low-ionization species. In addition, all of these low-ionization species are produced in very narrow ionization fronts, so there are likely systematic errors in the photoionization modeling (see Mehdipour et al. 2016a) that are larger than the statistical uncertainties we show in the figure. Kaastra et al. (2014) used only the XMM-Newton X-ray spectra to determine the ionization parameter and column density of the obscurer. Since there are no spectral features to constrain the X-ray models, a broad range of ionization parameters produces acceptable fits. As shown by Mehdipour et al. (2017), Kriss et al. (2019), and Longinotti et al. (2019), including the UV absorption as an additional constraint indicates that the obscuring gas likely has higher ionization, even in the observations of the original XMM-Newton campaign. From our new analysis that includes the broad UV absorption as part of the solution, we see that the obscurer is much more highly ionized than originally thought. Component 1, with log N_H cm⁻² = 22.08, actually has an ionization parameter in the range log ξ₁ = 0.8–0.95, and component 2 is even more highly ionized at log ξ₂ = 1.5–1.6. These differ sufficiently from the original fits of Kaastra et al. (2014; log ξ₁ = −0.8 and log ξ₂ = −4.5) that another look at the X-ray spectral analysis is warranted.

Zoom In Zoom Out Reset image size

The narrow intrinsic absorption features in NGC 5548 lie at distances of ∼3 to hundreds of pc (Arav et al. 2015). They vary on timescales of days (as seen above in Figure 11) to years (Arav et al. 2015) and appear to be associated with the X-ray warm absorber in NGC 5548 (Mathur et al. 1995; Crenshaw et al. 2009; Kaastra et al. 2014; Arav et al. 2015). Using density-sensitive transitions in the metastable excited states of C iii and Si iii, Arav et al. (2015) determined the density of the gas producing the absorption features associated with component 1 as log n cm⁻³ = 5.8 ± 0.3, placing it at a distance of 3.5 ± 1.0 pc, consistent with the density and the 1–3 pc location of the emission-line gas in the NLR (Peterson et al. 2013). Using the time variability of these absorption features, we can independently measure the density by measuring the recombination time in the gas. Krongold et al. (2007) used the variability of the aggregate soft X-ray absorption as it responded to continuum flux changes in NGC 4051 to estimate recombination times, but here we have the opportunity to make such measurements with distinct, resolved absorption lines in the UV.

For gas in photoionization equilibrium, the population density n_i in a state i depends on the balance between the ionizing photon flux causing ionizations to more highly ionized states and recombinations from those states. Following Krolik & Kriss (1995), this can be expressed as

$$\begin{eqnarray}\begin{array}{rcl}{{dn}}_{i}/{dt} & = & -({F}_{\mathrm{ion},i}{\sigma }_{\mathrm{ion},i}+{n}_{e}{\alpha }_{\mathrm{rec},i-1}){n}_{i}\\ & & +\,{n}_{e}{n}_{i+1}{\alpha }_{\mathrm{rec},i}+{F}_{\mathrm{ion},i-1}{\sigma }_{\mathrm{ion},i-1}{n}_{i-1}.\end{array}\end{eqnarray} \tag{ 16 }$$

For the ions we are measuring, generally, ${n}_{i-1}\ll {n}_{i}\ll {n}_{i+1}$, so we can simplify to

$$\begin{eqnarray}&&{{dn}}_{i}/{dt}=-{F}_{\mathrm{ion},i}{\sigma }_{\mathrm{ion},i}{n}_{i}+{n}_{e}{n}_{i+1}{\alpha }_{\mathrm{rec},i}.\end{eqnarray} \tag{ 17 }$$

In general, as long as there is a copious increase in the ionizing flux, the $-{F}_{\mathrm{ion},i}{\sigma }_{\mathrm{ion},i}{n}_{i}$ term dominates, and ions n_i are destroyed instantly. Conversely, when the flux decreases abruptly, ${n}_{e}{n}_{i+1}{\alpha }_{\mathrm{rec},i}$ dominates, and n_i reappears more slowly, on the recombination timescale

$$\begin{eqnarray}&&{\tau }_{\mathrm{rec}}=({n}_{i}/{n}_{i+1})/({n}_{e}{\alpha }_{\mathrm{rec},i}).\end{eqnarray} \tag{ 18 }$$

Figure 18 beautifully illustrates this simple behavior. We can see the absolute magnitude of the EW of the C ii λ1334 absorption features decrease in absolute value immediately when the continuum flux increases. Conversely, when the continuum flux decreases, there is a noticeable delay in the C ii response as it takes a measurable amount of time for recombinations to repopulate the C ii ionization state.

Zoom In Zoom Out Reset image size

Our high data quality and good sampling enable us to measure recombination delays directly from our light curves. To obtain an objective, empirical measure of the recombination time, we cross-correlated the continuum light curve with the absorption-line light curves (using the ICCF of Peterson et al. 2004) but restricted our cross-correlation to time intervals when the continuum flux had reached a peak and then fell to a minimum. We obtained good cross-correlations only for a select number of absorption lines in components 1 and 3. We tabulate the measured recombination times in Table 8. To convert these times into densities, we use Cloudy 17.00 (Ferland et al. 2017) to obtain the recombination rates for each of the ions in Table 8, assuming a fiducial density of log n = 4.8 cm⁻³. The best-fit photoionization solutions from Arav et al. (2015) give log U = −1.5 and log N_H = 21.5 cm⁻² for component 1 and log U = −1.3 and log N_H = 21.2 cm⁻² for component 3. We scale up the ionization parameters by the ratio of the continuum fluxes at 1367 Å for the mean spectrum from the XMM-Newton campaign (3.11 × 10⁻¹⁴ erg cm⁻² s⁻¹ Å⁻¹) to the value from the mean spectrum of the STORM campaign (4.30 × 10⁻¹⁴ erg cm⁻² s⁻¹ Å⁻¹). The print ionization rates command in Cloudy then gives total recombination rates for the relevant ionic states, and we scale the fiducial density of log n = 4.8 cm⁻³ by the ratio of the inferred recombination time from Cloudy to our measurements in Table 8 to obtain the tabulated inferred densities.

These density measurements are reassuring both for their internal consistency and for their agreement with the independent determination of log n_e = 4.8 ± 0.3 obtained by Arav et al. (2015) using density-sensitive absorption-line ratios. Despite our simplifying assumptions, it is gratifying that the atomic physics of this gas produces such consistent results. C. Silva et al. (2019, in preparation) discussed the limitations of our simplifying assumptions in more detail and produced a more rigorous time-dependent photoionization model of the various light curves that verify these empirical results with greater precision.

This classic ionization response describes the light curves of the low-ionization transitions visible in component 1. However, this idealized behavior only holds true for approximately the first third of the campaign. At later times, the absorption lines do not follow the continuum so closely, during either the ionization or recombination phases. The correlation plot comparing EW(C ii) to the UV continuum flux F_λ(1367 Å) in Figure 19 shows a good linear correlation (linear correlation coefficient r = 0.82) for the first 75 days of the campaign, or THJD < 56,766, but much more scatter at later times (r = 0.39).

Zoom In Zoom Out Reset image size

A partial explanation for the inconsistent correlation of the absorption-line strength with the observed UV continuum is that the observed continuum is only one determinant of the actual ionizing flux beyond the Lyman limit. Historically, the extreme UV (EUV) continuum varies with greater amplitude than the FUV (Marshall et al. 1997). We also know that the soft X-ray continuum is obscured by optically thick gas that only partially covers the continuum source (Kaastra et al. 2014). This opaque partial coverage also shadows the ionizing UV, as shown by the photoionization analysis of the UV absorption lines (Arav et al. 2015). The obscuration is also variable, with the variability predominantly explained by variations in the covering fraction (Di Gesu et al. 2015; Cappi et al. 2016; Mehdipour et al. 2016b). If the covering fraction of the obscurer is varying, then one would expect variations in the ionizing flux that are independent of the strength of the observed UV continuum. To test this hypothesis, we did a joint correlation analysis of the variations in EW(C ii) with the UV continuum and the hardness ratio (HR) as measured with Swift. The HR is defined as HR = (HX − SX)/(HX + SX), where SX is the soft X-ray count rate in the 0.3–0.8 keV band, and HX is the count rate in the hard X-ray band, 0.8–10.0 keV (Edelson et al. 2015). The HR is almost a direct measure of the covering fraction of the obscurer (Mehdipour et al. 2016b). A linear fit of the EW(C ii) to the UV continuum flux, F_λ(1367 Å), yields the red line shown in Figure 20, with χ² = 224.2 for 171 points and 2 degrees of freedom. This is a good correlation, but the fit is not a statistically acceptable predictor of the strength of the EW(C ii). If we include the HR as an additional independent variable, the goodness of fit improves dramatically to χ² = 156.1, which is statistically acceptable. This strongly bolsters the interpretation that the ionizing continuum for C ii is determined by both the observed intensity of the UV continuum and the fraction of the ionizing continuum that is covered by the X-ray obscurer.

Zoom In Zoom Out Reset image size

The bottom panel of Figure 20 shows how including HR as a predictor of the EW(C ii) improves the fit. The red line is the prediction based solely on the UV continuum flux. The observed EWs have a large scatter about this line, as seen by the black points. When the HR for each observation is taken into account, the predicted EWs then vary from the simple linear fit in a manner consistent with variations in the covering fraction. For a given UV flux, high HRs, indicative of high covering fractions, mean that more of the ionizing continuum is obscured, so predicted EWs are greater in magnitude (more negative). For low HRs, covering fractions are lower, more ionizing flux leaks past the obscurer, and EWs decrease in magnitude as more C ii is ionized.

An alternative possibility is that the shape of the ionizing continuum is varying, and that the soft X-ray flux might be a good measure of this when combined with the UV continuum flux. We therefore tested a joint correlation of the EW(C ii) with the UV continuum flux and the soft X-ray count rate, SX. This also gives an improved fit, χ² = 180.7, which is statistically acceptable but not as good as the fit using HR. The improvement can readily be explained as a consequence of SX being determined partially by intrinsic flux variations but mostly by variations in the covering fraction of the obscurer, as argued by Mehdipour et al. (2016b).

The lack of a direct, exclusive correlation with the observed UV continuum flux is even more striking for all of the high-ionization absorption lines, Si iv, C iv, and N v. Figure 21 illustrates these effects in the blue component of C iv λ1548 associated with the absorbing gas in component 1. This transition has the highest blueshift of any narrow intrinsic absorption feature for C iv, and therefore it is not blended with any other components. In Figure 21, one can see that it tracks the UV continuum variations very closely up to THJD = 56,766 during both increases in flux and decreases in flux. For the first 75 days of the STORM campaign, before the BLR holiday, C iv has a linear correlation coefficient of r = 0.86. For the remainder of the campaign, however, the correlation drops significantly, with r = 0.06. In contrast to C ii, C iv does not show any recombination delay when continuum flux levels fall, even during the first 75 days of the campaign; the absorption increases in strength almost immediately when the flux levels rise (for THJD < 56,766). The other high-ionization ions, Si iv and N v, also show this instantaneous response. At later times, however, C iv becomes almost completely decorrelated. The correlation plot comparing the EW(C iv) to the UV continuum flux F_λ(1367 Å) in Figure 22 shows a tight correlation with the UV continuum during the first 75 days, just like for C ii, but a large degree of scatter at later times. During the period of the BLR holiday, after THJD = 56,776, there is no coherent correlation between EW(C iv) and the UV continuum flux. The particularly discordant points in the upper right corner of Figure 22 correspond to the FUV continuum flux peak near the end of the holiday period at THJD = 56,820.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

As for C ii, we investigated whether other variables might also have a strong correlation with the ionizing flux that is controlling the population of C iv ions. In Figure 23, we compare EW(C iv) to the soft X-ray flux as measured by Swift (Edelson et al. 2015). Although not perfect, the correlation improves significantly, suggesting that the soft X-ray flux plays a more dominant role in controlling the C iv ionic population than the FUV continuum.

Zoom In Zoom Out Reset image size

Again, we investigated whether a multivariate correlation would be better than either continuum measure on its own. First, as expected, a simple linear fit of EW(C iv) versus F_λ(1367 Å) is statistically unacceptable with χ² = 6271 for 171 points and two free parameters. We then included the HR in the correlation, leading to a dramatic improvement, but the fit is still poor with χ² = 4314. Finally, in contrast to C ii, using the soft X-ray flux in tandem with F_λ(1367 Å) gives the best result, χ² = 3608. This is still statistically unacceptable, but it does show the stronger influence that the soft X-ray continuum is having on the C iv ionization than the FUV continuum. As a final trial, assuming that the HR might serve as an additional measure of the covering fraction of the obscurer, we carried out a trivariate analysis using SX, F_λ(1367 Å), and HR. This provided no additional improvement in χ², with Δχ² < 1. Figure 24 shows the correlation of the EW(C iv) with F_λ(1367 Å) and how including SX as an additional independent variable improves the fit. Much of the scatter about the linear fit is explained by variations in the soft X-ray flux. The low EW points in the upper right corner of the top panel of Figure 24 coincide with some of the lowest soft X-ray fluxes observed, as shown in the bimodal fit presented in the bottom panel. The fact that it is the soft X-ray flux directly rather than just a covering factor effect (as for C ii) may be telling us that the EUV and soft X-ray continuum shape is also varying, since it is this portion of the continuum that is responsible for ionizing C iv. This is not the whole story, however. As shown in Figure 24, there is a large set of points (shown in green) where variations in the EW(C iv) are not explained by either UV or soft X-ray continuum variations. These green points encompass times from THJD = 56,776 to 56,827. This evidence, plus the complete decoupling from the continuum variations in the light curve after THJD ∼ 56,750, appears to be related to the BLR holiday (Goad et al. 2016), which sets in at the same time. The green points in Figure 24 lie in the deepest portion of the BLR holiday. In Paper VII (Mathur et al. 2017), we speculated that the BLR holiday was related to changes in the EUV/soft X-ray continuum shape, and that these changes in the higher-energy range of the SED were not reflected in the visible FUV continuum. The responses of the high-ionization intrinsic absorption lines enable us to directly measure these changes in the SED.

Zoom In Zoom Out Reset image size

While the responses of the broad emission lines also give insight into the behavior of the continuum, their large-scale spatial distribution and differing views of the continuum source complicate such an analysis. Since the narrow intrinsic absorption lines in NGC 5548 are directly along our line of sight, they are exposed to the same continuum that we see. This allows us to take advantage of their unique point of view and to obtain a direct measure of the continuum strength at a variety of energies. To a good first approximation, the continuum flux at an ion's ionization potential is the ionizing flux driving the population of that ion. To use the strength of the absorption lines as a measure of this flux, however, we must be certain that the absorption-line variations are reflecting changes in column density in response to changes in the ionizing flux, rather than being caused by variations in covering factor, transverse motion, or both.

The best evidence in favor of a dominant photoionization response is the good correlation we see between the strength of the individual absorption lines and either the FUV continuum (for low-ionization lines like C ii) or the soft X-ray flux (for high-ionization lines like C iv). Our biggest concern is for C iv, because it was heavily saturated during the XMM-Newton campaign. The photoionization models of Arav et al. (2015) show that it has an optical depth of 200 during the observations in 2014. Figure 25 shows that the bottom of the absorption trough in C iv λ1548 component 1 remained relatively constant in 2014, in agreement with heavy saturation. Likewise, Figure 23 shows that during the XMM-Newton campaign, the strength of the absorption in C iv λ1548 component 1 is independent of the soft X-ray flux and much stronger than during the AGN STORM campaign. In contrast, Figure 23 shows a strong correlation with the soft X-ray flux during the AGN STORM campaign, indicative of a photoionization response. In Figure 26, we show that the absorption-line profiles change as expected if the column density in C iv component 1 was varying in response to changes in the ionizing flux.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Assuming that the absorption lines in component 1 are indeed tracing a response to change in the ionizing flux, we can use this behavior as a diagnostic of the ionizing continuum. For the first 75 days of the campaign (dates prior to THJD = 56,766), both the absorption lines and the emission lines show a good correlation in their variations with the FUV continuum flux at 1367 Å. We use this proportionality to derive a linear relationship between a line's EW and the continuum flux of

$$\begin{eqnarray}&&\mathrm{EW}={a}_{0}+{a}_{1}\times {F}_{\lambda }(1367).\end{eqnarray} \tag{ 19 }$$

During this early part of the campaign, when everything correlates well, we assume that the continuum shape is that shown in Figure 4 of Mehdipour et al. (2015). We designate points lying on this fiducial SED as F₀(λ), where λ may be either the observed wavelength 1367 Å or the wavelength at the ionization potential of the relevant ion, λ_IP. We can then use variations in the EW of a given line to measure changes in the flux at the ionization potential relative to this fiducial SED shape of

$$\begin{eqnarray}&&\displaystyle \frac{F({\lambda }_{\mathrm{IP}})}{{F}_{0}({\lambda }_{\mathrm{IP}})}=\displaystyle \frac{(\mathrm{EW}-{a}_{0})/{a}_{1}}{{F}_{0}(1367)}.\end{eqnarray} \tag{ 20 }$$

Following this methodology, we derive relative ionizing continuum light curves for each absorption line representative of the time-varying relative fluxes at their ionization potentials. Figure 27 shows these relative light curves for the blue components of the Si iv, C iv, and N v doublets of the intrinsic absorption line for component 1 in NGC 5548. The figure also compares these relative light curves to the deficiency in the broad C iv emission-line flux during the BLR holiday (Goad et al. 2016), defined as the percentage diminution in the C iv emission-line flux during the holiday compared to the expected flux based on the correlation of C iv emission-line flux with the continuum flux observed during the first 75 days of the AGN STORM campaign. Note that the relative decrease in broad C iv emission-line flux occurs during the same time interval (56,776 < THJD < 56,827) as the inferred decrease in EUV continuum flux at the ionization potentials of Si iv (45.1 eV), C iv (64.5 eV), and N v (97.9 eV) deduced from the behavior of the intrinsic absorption line. All three inferred continua show significant correlations with the deficit in the C iv emission. However, there is a delay between the C iv deficit and the inferred continua. Using ICCF, we find delays and peak correlation coefficients of 11.6 ± 1 days and r = 0.66 for Si iv, 10.5 ± 1 days and r = 0.70 for C iv, and 12.3 ± 1 days and r = 0.64 for N v. These delays are actually with respect to the FUV continuum at 1158 Å, since Goad et al. (2016) shifted the C iv light curve by −5 days to align it with the continuum variations, so the inferred EUV continua variations are delayed by 5–7 days relative to the C iv deficit. These delays are puzzling, since one might expect the C iv deficit to lag the continuum deficits rather than vice versa if they were the direct cause of the BLR holiday. Dehghanian et al. (2019) investigated the complex responses of the intrinsic narrow absorption lines to changes in the ionizing continuum and the intervening obscurer in greater detail, although they also did not explain this puzzling behavior.

Zoom In Zoom Out Reset image size

As discussed in Section 5.2, the broad absorption features that appear as extensive blueshifted troughs on all of the resonant UV lines in the spectrum of NGC 5548 are the UV counterparts to the soft X-ray obscurer discovered by Kaastra et al. (2014). These broad UV absorption features also vary with time. In fact, in the observations from that campaign, there was a significant anticorrelation between the strength of the absorption troughs and the soft X-ray flux as measured by Swift (linear correlation coefficient r = −0.37). As the soft X-ray intensity increased, the broad UV absorption strength decreased. This is not due to a photoionization response, however. The broad UV absorption lines are saturated, so their strength is also governed more by the covering factor than by the ionization state or column density of the absorbing gas.

The more extensive monitoring of the time variability of the broad absorption features enabled by the STORM campaign reveals an even more complex picture. As shown in Figure 28, the anticorrelation of the EW of the broad absorption is not as clean as it was in the six points available from the XMM-Newton campaign.

Zoom In Zoom Out Reset image size

While there is a rough anticorrelation (r = −0.41) early in the campaign (THJD < 56,766), as with the behavior of the narrow absorption lines, the correlation disappears after THJD = 56,766 (r = +0.02), approximately the time of the onset of the BLR holiday (Goad et al. 2016). Figure 29 shows this even more dramatically. Light curves for the EWs of the broad absorption features are eerily similar to the time variation of the anomalous deficit in the response of the flux of the broad emission lines. The connection between these two behaviors is puzzling. On the one hand, the lack of response in the BLR during the holiday appears to be due to a change in the SED, in which the soft X-ray excess decouples from the visible UV continuum. A decrease in the BLR flux during the holiday implies a decrease in the ionizing UV driving the line emission. On the other hand, a decrease in the broad absorption-line EW implies a decrease in the covering factor of the obscurer, because these broad troughs appear to be saturated (Section 4.2), and we see no dramatic changes in their velocity widths. This is also seen in the Swift monitoring of NGC 5548 (Mehdipour et al. 2016b), and it implies that more of the EUV and soft X-ray flux is being transmitted past the obscurer. Plus, since the BLR response represents a global view of the continuum source, and our view of the obscurer is restricted purely to our line of sight, this suggests that our view of the obscurer may not be a special geometrical arrangement, such as a single cloud crossing our line of sight, but rather a more axisymmetric arrangement affecting all sight lines that illuminate the BLR. Dehghanian et al. (2019) presented further photoionization analysis of both the narrow and the broad absorbers that support this interpretation.

Zoom In Zoom Out Reset image size

An approximate axisymmetric arrangement for the obscurer is even more natural when one considers the timescales covered by our observations. From the beginning of the XMM-Newton campaign in 2013 June to the end of the reverberation mapping campaign in 2014 July, the elapsed time was 415 days. For a black hole mass of 5.2 × 10⁷ M_☉ (Bentz & Katz 2015), clouds at a radius of 1 lt-day at the innermost edge of the BLR would have an orbital timescale of 115 days. Thus, the obscurer and the innermost part of the BLR could have completed more than one full revolution over the course of our observing programs.