Red and Reddened: Ultraviolet through Near-infrared Observations of Type Ia Supernova 2017erp^*

SN 2017erp was discovered by Itagaki (2017) in images taken on 2017 June 13 15:01:28 (all dates UT). It was classified by Jha et al. (2017) as an extremely young SN Ia.

The host galaxy of SN 2017erp is NGC 5861, classified as SAB(rs)c (de Vaucouleurs et al. 1991), at a redshift of 0.006174 ± 0.000003 (Theureau et al. 2005). The redshift corresponds to a distance modulus of 32.30 ± 0.26 assuming a Hubble constant of 73 km s⁻¹ Mpc⁻¹ and a random velocity uncertainty of 300 km s⁻¹.

The Neil Gehrels Swift Observatory (Gehrels et al. 2004) began observing SN 2017erp on 2017 June 14 04:55:33 with the six photometric filters: uvw2, uvm2, uvw1, u, b, and v of the UVOT (Roming et al. 2005). The lower case filter name convention is typically used in this paper for the photometry or spectrophotometry in the UVOT bands. Since the differences are small, no corrections are made to the standard Johnson B and V filters, and upper case letters are sometimes used to match common usage, such as Δm₁₅(B) or E(B – V). Filter details are available in Roming et al. (2005) and Breeveld et al. (2011). An image of SN 2017erp and its host galaxy is displayed in Figure 1. The photometry was reduced using the pipeline of the Swift Optical/Ultraviolet Supernova Archive (SOUSA; Brown et al. 2014a) using the zero-points of Breeveld et al. (2011) on the Vega system. The Swift/UVOT reduction includes the subtraction of the host galaxy count rate in the UVOT filters using observations taken about 16 months after the SN explosion. The difference between the peak magnitudes with subtraction and no subtraction at all is less than 0.02 mag in the optical using a 3'' aperture. The subtraction of the host galaxy has the strongest impact in the uvm2 filter, equal to 0.25 mag. The final photometric errors include the propogation of the statistical uncertainty of the SN+galaxy and galaxy count rates and a 2% systematic uncertainty on each due to large-scale sensitivity differences across the detector.

Zoom In Zoom Out Reset image size

The light curves are shown in Figure 2 and the magnitudes listed in Table 1. The extremely red early colors prompted us to trigger our HST program once we had confidence that the colors were not solely due to dust reddening and that SN 2017erp would be bright enough in the mid-UV to obtain useful spectra. This assessment was done using color evolution and color–color evolution plots of young SNe Ia previously observed with Swift/UVOT as well as the UV/optical spectral template of SN 2011fe (Pereira et al. 2013) with different amounts of reddening applied to map out the parameter space covered by a reddened NUV-blue SN Ia (Brown et al. 2017).

Zoom In Zoom Out Reset image size

Observations with the HST were triggered as part of the program "Ultraviolet Spectra of a Normal Standard Candle" to obtain UV spectroscopy. Four epochs were obtained with the HST's STIS using the G430L grism and the CCD detector and the G230L grism and the MAMA detector. For these spectra we use the default HST reduction obtained from the Mikulski Archive for Space Telescopes.²² We eliminate bad pixels and cosmic rays, smooth the spectra in 5 Å bins, and combine the MUV G230L and NUV/optical G430L spectra. The four epochs are displayed in Figure 3 with observation times listed in Table 2.

Zoom In Zoom Out Reset image size

Optical light curves and spectra come from the Global Supernova Project, a three-year Key Project to observe light curves and spectroscopy of hundreds of SNe using the Las Cumbres Observatory global network of 21 robotic telescopes. UBVgri images were taken with the Sinistro cameras on the Las Cumbres Observatory network of 1 m telescopes (Brown et al. 2013). Point-spread function (PSF) photometry was extracted using lcogtsnpipe (Valenti et al. 2016), a PyRAF-based photometric reduction pipeline. UBV Vega magnitudes are calibrated to Landolt standard fields taken at the same site and on the same night as observations of the SN. gri AB magnitudes are calibrated to the Sloan Digital Sky Survey.

Broadband BVRI-band photometric observations of SN 2017erp were also obtained with the Tsinghua-NAOC 0.8 m telescope (TNT) at Xinglong Observatory in China (Huang et al. 2012) and AZT-22 1.5 m telescope (hereafter AZT) at Madanak Astronomical Observatory in Uzbekistan, spanning the phases from +2 to +28 days relative to the B-band maximum light. All CCD images were preprocessed using standard routines, which include corrections for bias, flat field, and removal of cosmic rays. As the SN is located relatively far from the galactic center, we did not apply a technique of subtracting the galaxy template from the SN images; instead, the foreground sky was determined locally using a nearby annulus and subtracted. The instrumental magnitudes of both the SN and the reference stars were then measured using the standard PSF. These magnitudes are converted to those of the standard Johnson-Vega system using the APASS catalog.²³

Spectra were taken with the robotic FLOYDS spectrographs on Las Cumbres Observatory's 2 m telescopes on Haleakalā Hawai'i, and in Siding Spring, Australia, and were reduced using the PyRAF-based floydsspec pipeline. Additional spectra were taken with the Southern African Large Telescope (SALT) using the Robert Stobie Spectrograph (Smith et al. 2006) as part of program 2017-1-MLT-002 (PI: Jha). We used the PG0900 grating and 15 longslit, yielding a spectral resolution λ/Δλ ≈ 900 over the wavelength range 3500–9400 Å. The data were reduced with a custom pipeline that incorporates routines from PyRAF and PySALT²⁴ (Crawford et al. 2010). A series of spectra of SN 2017erp also were obtained with the Xinglong 2.16 m telescope (XLT+BFOSC) of NAOC, China, the 2.3 m Australian National University (ANU) telescope (+WiFeS), and the 6.5 m Magellan telescope (+IMACS). All of the spectra were mangled to match the BVgri photometry at that epoch, except for the first SALT spectrum, which predated the photometry. The optical spectra are displayed in Figure 4, and the observation dates and wavelength ranges of all the spectra are given in Table 2.

Zoom In Zoom Out Reset image size

One epoch of NIR spectroscopy was obtained with the FLAMINGOS-2 (Eikenberry et al. 2008) spectrograph at Gemini South on 2017 July 11. The FLAMINGOS-2 data were taken with the JH grism and filter in place, along with a 0.72 arcsec slit width, yielding a wavelength range of 1.0–1.8 μm and R ∼ 1000. The FLAMINGOS-2 longslit data were reduced in a standard way (i.e., image detrending, sky subtraction of the AB pairs, spectral extraction, spectral combination, and wavelength calibration) using the F2 pyraf package provided by the Gemini Observatory. An A0V star was observed near in time and position to the science data in order to make a telluric absorption correction and to flux calibrate the spectra, following the methodology of Vacca et al. (2003). This spectrum is compared to a GNIRS NIR spectrum of SN 2011fe (Hsiao et al. 2013) in Figure 5.

Zoom In Zoom Out Reset image size

We used the SuperNovae in object-oriented Python (SNooPy; Burns et al. 2011) light-curve fitter on the Las Cumbres Observatory BVgri photometry to determine various light-curve parameters. The B band reached a maximum brightness of 13.27 ± 0.01 mag on MJD 57,934.9 (UT 2017 June 30.9). Δm₁₅, a parameterization of the multifilter templates sorted by Δm₁₅(B), is measured to be 1.05 ± 0.06, while Δm₁₅(B) directly measured from the UVOT light curves is 1.11 ± 0.02. SNe Ia with 1.0 < Δm₁₅(B) < 1.4 would be considered normal, with Δm₁₅(B) = 1.1 usually used as the reference value (e.g., Phillips et al. 1999). The color stretch (Burns et al. 2014) measured with SNooPy is s_BV = 0.993 ± 0.03. For comparison, an SN with Δm₁₅(B) = 1.1 would have a s_BV = 0.95. Thus SN 2017erp has a very standard light-curve shape.

The line-of-sight dust extinction through the Milky Way (MW) is estimated to be A_V = 0.296 based on the Schlegel et al. (1998) dust maps recalibrated by Schlafly & Finkbeiner (2011), corresponding to E(B − V) = 0.095 mag, assuming R_V = 3.1. The optical spectra show absorption by interstellar Na I D from both the Milky Way and the host galaxy NGC 5861 with comparable strengths. Our best optical spectrum (in terms of signal-to-noise and resolution in that region), the SALT spectrum taken 6 days after maximum light, is used to measure the EW of both components. Using a variety of places on the spectrum to set the continuum level and estimate the uncertainty, we measure EWs of 0.65 ± 0.05 and 0.77 ± 0.06. This corresponds to E(B − V) = 0.09 for the MW component, consistent with the Schlafly & Finkbeiner (2011) determination, using the correlation of Poznanski et al. (2011). The host component corresponds to E(B − V) = 0.11 ± 0.03. It is actually the extinction A_V that should correlate with the Na ID EW Phillips et al. (2013), but we are making an implicit assumption that the relevant host galaxy properties (dust to gas ratio, R_V, etc.) are similar to the MW.

For other estimates of the total reddening, we compare the B_peak–V_peak pseudocolor to Equation (7) of Phillips et al. (1999), yielding a host E(B − V) of 0.13 mag. By the term pseudocolor we refer to the subtraction of the nonsimultaneous peak magnitudes in different filters, while the peak color would refer to the subtraction of magnitudes at the time of maximum light in the B-band light curve. An alternative estimate of the reddening comes by comparing B − V colors between 30 and 90 days after the time of V-band maximum to the Lira relation (Phillips et al. 1999), which is a linear fit to the fairly uniform color decay of SNe Ia. The host E(B − V) color excess measured for individual observations compared to the Lira relation after correction for the MW component has an mean of 0.16 mag and a standard deviation of 0.03 mag.

Light-curve fitters such as MLCS2k2 (Jha et al. 2007), SNooPy (Burns et al. 2011), and SALT2 (Guy et al. 2010) use more of the light curve in determining the color differences, and thus reddening, between the SN Ia being fit and the training sample.

The MLCS2k2 (Jha et al. 2007) fit to the optical BVri light-curve data with Milky Way E(B – V) = 0.095 yields Δ = −0.19 ± 0.02 and host A_V = 0.39 ± 0.04 for fixed host R_V = 1.9. This corresponds to a host E(B – V) = 0.21 ± 0.02 and thus a total E(B − V) = 0.31. The MLCS2k2 distance modulus is 32.22 ± 0.09 (on an H₀ = 72 km s⁻¹ Mpc⁻¹ scale), marginally consistent with the Hubble flow distance modulus of 32.30 mag. A fit that included U and g resulted in poor fits in those filters and pulled the fit to a higher host E(B – V) = 0.29 ± 0.02 and a closer distance modulus of 32.07 ± 0.09. This could be due to differences between the filter transmission curves used in the observations and those used for training similar filters or the strong Ca H&K feature being interpreted as reddening.

The color model of SNooPy estimates the E(B − V)_host = 0.179 ± 0.005 (statistical) ±0.060 (systematic) mag with R_V = 2.80 ± 0.51. Fitting similar to Prieto et al. (2006) yields E(B − V)_host = 0.097 ± 0.005 (statistical) ±0.060 (systematic) mag, assuming R_V = 3.1.

The optical light curves were also fit with the SALT2 model (Guy et al. 2010) using the SNCosmo package (Barbary et al. 2016). The best-fit parameters are m_B = 13.28, x₁ = 0.46, and c = 0.05, where m_B is the peak magnitude at rest frame B band, x₁ is the shape parameter and c is the color parameter. Using the Tripp relation (Tripp 1998): μ = m_B + αx₁ − βc − M and setting α = 0.14, β = 3.1, and M = −19.1, we determine the distance modulus as 32.29 ± 0.12, where the uncertainty has been estimated based on the typical scatter for SN Ia. The SALT2 color term does not translate directly into a host reddening E(B – V) because it empirically combines intrinsic color as well as the dust reddening. The distance modulus is consistent with the MLCS2k2 distance modulus. Vinkó et al. (2012, 2018) found that ≤0.2 mag differences are possible between the distance modulus determinations for well-observed SNe as determined by different fitting methods and using the same method on separate subsets of BVRI and g'r'i'z' data.

The various reddening estimates are summarized in Table 3. Clearly there are differences larger than the claimed uncertainties. Since the difference would be due to systematic uncertainties in the fitting techniques or the filters used, instead of averaging or otherwise combining the values we will instead present the results for different assumptions in the host galaxy reddening, utilizing either all of the values or two values representing the clusters of values, namely host E(B – V) = 0.10 and 0.18, with R_V = 3.1 and 1.9, respectively.

The observed color evolution and peak maximum light colors of SN 2017erp are consistent with the NUV-red SNe Ia as defined by Milne et al. (2013). Figure 6 shows the uvw1 – v color evolution compared to shaded regions representing the NUV-blue and NUV-red groups (Brown et al. 2017). This definition utilized the observed colors and did not take into account dust reddening.

Zoom In Zoom Out Reset image size

As shown in Brown et al. (2017), however, some SNe Ia with observed colors similar to NUV-red SNe could be NUV-blue SNe reddened by dust. Determining the intrinsic colors is complicated by the possibility of different intrinsic optical colors (Milne et al. 2013), uncertainty in the appropriate dust reddening law, and how much the intrinsic spectral differences would affect the broadband filter extinction coefficients determined from a NUV-blue spectral template like SN 2011fe. The latter uncertainty is removed by having a new spectral sequence of a NUV-red SN Ia. In Figure 6 we also plot the spectrophotometric color evolution from the HST observations with different amounts of reddening applied. With the larger host E(B – V) = 0.18 inferred from the multiband data, SN 2017erp crosses into the color evolution space occupied by NUV-blue SNe Ia. The B − V evolution in such a case becomes bluer than the unreddened SN 2011fe template.

In Figure 7 we focus on the peak color of SN 2017erp and compare with those of the Brown et al. (2017) sample with 1.0 < Δm₁₅(B) < 1.4. Lines show the reddening vectors for SN 2011fe and SN 2017erp corresponding to different dust reddening laws. The dust reddening laws include the Cardelli et al. (1989) law parameterized with R_V = 3.1 (like the MW) and R_V = 1.7 (similar to that found for SNe Ia; e.g., Kessler et al. 2009), an extinction law measured for the SMC (Prevot et al. 1984), and a Large Magellenic Cloud extinction law modified by circumstellar scattering (Goobar 2008; Brown et al. 2010). We also show symbols corresponding to dereddening the SN 2017erp spectrum for MW reddening with E(B − V) = 0.095 mag and R_V = 3.1 and after correcting for the MW reddening and the SNooPy estimated host reddening of E(B − V) = 0.097 mag (with R_V = 3.1).

Zoom In Zoom Out Reset image size

For a host reddening of E(B – V) = 0.1, the dereddened colors of SN 2017erp are about 0.5 mag redder in the uvw1 – v color than the low reddened SN 2011fe, with comparable colors to several NUV-red SNe Ia. For the larger host reddening of E(B – V) = 0.18, the uvw1 – v color becomes comparable to the NUV-blue SN2011by. The dereddened B − V color, however, would be significantly bluer than the observed colors of the other SNe. This is understandable as the SNooPy fitting seeks to correct the colors to match the bluest in the CSP training sample. In this case, the B − V color at peak ends up about one σ bluer than the mean for objects of similar light-curve stretch (Burns et al. 2014). It would be the bluest SN Ia of the Swift sample in both colors. The reddening estimated from the peak B − V colors is smaller than that from the whole light-curve fitters or the late-time Lira relation, so the larger estimates push the reddening-corrected peak color bluer.

Though SN 2017erp could be dereddened enough to be close in uvw1 – v color to those observed for some NUV-blue SNe Ia, its dereddening tracks in the color–color diagram are most consistent with the NUV-red SNe. A clear understanding of the intrinsic colors would require a correct dust reddening correction for the sample—a correction that might depend on the intrinisic spectral shape. Brown et al. (2017) considered the perpendicular offset from SN 2011fe's reddening vector as a more appropriate measure of the intrinsic UV differences than simply the NUV-optical color used by Milne et al. (2013). However, an intrinsic difference in B – V colors correlated with the NUV colors (Milne et al. 2013; possibly due to metallicity Höflich et al. 1998; Walker et al. 2012) is also a possible complication. The histogram of colors offset from the SN2011fe with MW dust reddening vector in Brown et al. (2017) showed less of a dichotomy in the NUV colors. In that paradigm, the colors range from 0 (SN2011fe) to 0.8 mag, with SN 2017erp near 0.5. Though not at the extreme red end, this makes the SN 2017erp HST spectral series complementary to those of the NUV-blue SNe 2011fe (Mazzali et al. 2014) and 2011by (Foley & Kirshner 2013) and the reddened NUV-blue SN 2015F (Foley et al. 2016). A diversity in spectral templates to match the observations is important regardless of whether the SNe Ia at the NUV-blue and NUV-red ends represent distinct groups or a continuously varying color difference.

3.3.1. Optical Spectra

The full optical spectral evolution of SN 2017erp is shown in Figure 4, while Figure 8 focuses on the region of the Si ii 6355 Å line in the early spectra. The velocity of the Si ii 6355 Å absorption begins at a high velocity of 21.5 Mm s⁻¹ (Jha et al. 2017) but drops quickly to a value of 10.4 Mm s⁻¹ at maximum light. The early velocity is actually that of a high-velocity component. The asymmetric Si ii profile 11 days before maximum light shows the transition between the feature being dominated by the high-velocity component to being dominated by the photospheric component. SN 2009ig showed similar behavior, and Marion et al. (2013) measured the velocity evolution of the separate components. Figure 9 shows the velocity evolution of the Si ii 6355 line (considering only the dominant component) and demonstrates SN 2017erp is not significantly different from the "gold standards" SNe 2011fe (Pereira et al. 2013; Zhang et al. 2016) and 2005cf (Garavini et al. 2007; Pastorello et al. 2007; Wang et al. 2009b) in terms of photospheric velocity evolution. At shorter wavelengths, the Ca ii H&K feature also begins by being very broad and stretched to blue wavelengths. By day 15 (though probably earlier where our spectra are noisier) the Ca ii H&K feature is split into two distinct features with a high-velocity feature at a similar velocity as near maximum light.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Milne et al. (2013) found that SNe Ia with high velocities (HV; Si ii ejecta velocities greater than about 12 Mm s⁻¹ Wang et al. 2009a) are almost exclusively NUV-red, while normal-velocity SNe Ia can be either NUV-red or NUV-blue. Meanwhile, the NUV-blue SNe Ia all have velocities below 12 Mm s⁻¹. Brown et al. (2018) confirmed this, showing that SNe Ia with normal velocities have a range of UV colors with no correlation with the velocities. Thus the photospheric velocity does not differentiate NUV-red and NUV-blue SNe Ia, consistent with the optical similarities seen here. However, the similarity in the high-velocity features of SNe 2017erp and 2005cf suggest a possibility that high-velocity features may be related to the similar NUV-red colors. This would be consistent with the high-velocity features and the UV continuum both arising from the outer layers of the SN ejecta. Brown et al. (2018) found that the asymmetric models of Kasen & Plewa (2007) did not reproduce the observed UV colors, but further work modeling asymmetry (e.g., Maeda et al. 2010; Maund et al. 2010) and its effect on the UV (e.g., Kromer & Sim 2009; Brown et al. 2014b will better constrain how much of an effect asymmetry could have on the UV colors.

Milne et al. (2013) also found that all of their NUV-blue SNe Ia had carbon detected in their spectra. The majority of NUV-red SNe Ia had no detections, with SN 2005cf as the lone exception. SN 2017erp would be another exception, as it shows strong C ii absorption in the earliest spectrum. This is shown in Figure 8. However, the C ii is gone by the fifth spectrum, taken 12 days before maximum light. If the earlier spectra had not been taken, the spectrum 12 days before maximum would have counted as a nondetection. C ii was detected in SN 2005cf eleven days (Silverman & Filippenko 2012) and 9.4 days maximum light (Thomas et al. 2011) and then not detected two days later. These two exceptions might be due to how early the observations of SNe 2005cf and 2017erp began. Other NUV-red SNe may have had detectable carbon had they been observed at similarly early epochs. A better understanding of the relationship between carbon and UV colors would require both to be treated as continuous parameters (which each change with time) rather than individual classes.

3.3.2. Near-infrared Spectrum

3.3.3. Ultraviolet Spectra

Now we examine how the intrinsic differences seen here could be interpreted as peculiar dust extinction. In Figure 10 we compare the difference between SNe 2017erp and 2011fe to the various reddening laws discussed previously and the color law used in the SALT2 light-curve fitter (Guy et al. 2010; Betoule et al. 2014). Because we are interested in the shape of the wavelength dependence, the laws have been scaled to give the same B − V differences as between the SNe and then shifted to zero in the V band. The difference between SNe 2011fe and 2017erp matches broadly the shape of the SALT2 color law (which itself is made from a polynomial fit to broadband observations of higher redshift SNe Ia between 2800–7000 Å and extrapolated to wavelengths above and below that; Guy et al. 2010).

Zoom In Zoom Out Reset image size

We suggest that in the UV, the steep intrinsic differences could dominate the derivation and use of the SALT2 color law over dust reddening. The SALT2 color law is derived empirically and combines the potential effects of dust reddening and intrinsic color variations. This is difficult to disentangle in the optical, where both dust reddening and intrinsic red color seem to correlate with lower luminosity (e.g., Scolnic et al. 2014). However, the NUV differences do not appear to correlate with luminosity (Maguire et al. 2012; Brown et al. 2017). The use of UV colors to correct for reddening could lead to biased distances especially if the UV colors may change with redshift (Milne et al. 2015; but see also Cinabro et al. 2017).

We have focused on the colors and relative flux levels, as a well-measured distance to the host of SN 2017erp is not yet available. Observations of NGC 5861 with the HST have been made with the goal to measure the periods of Cepheid variables (PI: Riess) and calculate a distant in a consistent manner with SNe 2011by, 2011fe, and others (Riess et al. 2016; Foley et al. 2018). This will also shed light on the differences in dust extinction and inferred distance modulus estimated for SN 2017erp by the different light-curve fitters.

Understanding the origin of the UV differences is important to better characterizing the progenitors and explosion mechanisms of SNe Ia and how they might change with redshift. Following comparisons from Smitka (2016) that the NUV-blue/red spectral difference between 2700 and 3300 Å (see also Milne et al. 2013, 2015) could be caused by metallicity, we next compare the Walker et al. (2012) models to SNe 2011fe and 2017erp. The Walker models are shown in the left panel of Figure 11. We use a maximum light spectrum of SN 2011fe from Mazzali et al. (2014) and our maximum light spectrum of SN 2017erp dereddened by an MW extinction law with R_V = 3.1 and E(B − V) = 0.2 mag (for host and MW reddening) and also the spectrum dereddened by E(B – V) = 0.1 from the MW and also E(B – V) = 0.18 from the host (the latter corrected with an extinction law with R_V = 1.9). The spectra are normalized in the optical between 4000 and 4800 Å, so that we can compare the relative flux levels between the optical and the near-UV. The main features in the near-UV are the peaks identified as λ₁ and λ₂ in Ellis et al. (2008) and found by Walker et al. (2012) to be caused by reverse fluorescence. These comparisons are shown in the right panel of Figure 11.

Zoom In Zoom Out Reset image size

We find a reasonable match between SN 2011fe and the model with one-fifth the metallicity of the baseline model, while SN 2017erp matches the models with a factor of 2–5 increase in metallicity. The comparison plot is similar to that made by Smitka (2016) using Swift/UVOT grism spectra of SN 2005df for the NUV-red SN near peak. If the NUV differences are solely due to metallicity and the magnitude of the Walker et al. (2012) models is correct, that would imply a factor of 10 difference in the metallicity between SNe 2011fe and 2017erp. This value should be used with caution, as Mazzali et al. (2014) remark that the Walker et al. (2012) models had a larger fraction of unburned material and thus a stronger dependence on metallicity. Walker et al. (2012) explicitly note that their method of altering the model metallicity cannot distinguish between primordial metallicity of the progenitor and upmixing of products of explosive nucleosynthesis. Distinguishing the two is significant because of the effect of progenitor metallicity on the radioactive output of an SN Ia and thus luminosity (Timmes et al. 2003; Foley et al. 2018).

The uncertainty in the reddening correction does not greatly effect the continuum levels in the near-UV after the normalization in the B band. However, there is also an inconsistency in correcting SNe Ia to have similar optical colors and then comparing them to models with differences in B – V colors. In the Walker et al. (2012) models the SN Ia models with the largest metallicities are 0.1 mag redder than the low metallicity models. The effect of metallicity on SN Ia cosmology is not just a shift in the luminosities (Timmes et al. 2003; Foley et al. 2018) but also the intrinsic colors (Höflich et al. 2000; Podsiadlowski et al. 2006), which affect the extinction corrections and thus the luminosity distances. For this reason it is important to constrain the metallicity effects in SNe Ia for precision cosmology.

Continuing blueward, the mid-UV flux of SN 2017erp falls well below that of SN 2011fe and the best-fitting Walker et al. (2012) model. Foley & Kirshner (2013) have demonstrated using SNe 2011by and 2011fe that SNe with nearly identical optical and near-UV spectra can have very different continuum levels in the mid-UV. They attribute the spectral difference in the mid-UV to be caused by metallicity using the Lentz et al. (2000) models. We do not use the Lentz et al. (2000) models due to the large color difference from actual observations (Brown et al. 2015). Regardless of the cause, the differences between SNe 2011fe and 2011by and SNe 2011fe and 2017erp show that a separate mechanism must affect the mid-UV independent from the near-UV. The Walker et al. (2012) models were not first principle models, nor based on either of these objects, but abundances and density gradients were adjusted to match UV/optical spectra of SN 2005cf near maximum light. Other differences in the spectra are thus not accounted for, but we seek general trends causing the largest differences between SNe 2011fe and 2017erp. Similar to Foley & Kirshner (2013) we have compared two well-observed examples, but larger samples (Smitka 2016; Pan et al. 2018) of SNe Ia might help disentangle multiple effects.

There is a connection between our proposed cause of the NUV-blue/red dispersion and the general trend seen by Thomas et al. (2011) and Milne et al. (2013) that the NUV-blue SNe Ia frequently have detections of C ii in their spectra while NUV-red SNe Ia generally do not. Heringer et al. (2017) found that emission from iron can hide the absorption from carbon, implying that SNe Ia with carbon signatures have a lower metal content in their outer layers. As described above, SN 2017erp would be another exception to this general trend (as was SN 2005cf; Silverman et al. 2012), as there are dips in the early optical spectra redward of Si ii 6355 Å that are most likely C ii. However, the presence of C ii and its detectability is not a binary property, and Parrent et al. (2011) discussed a variety of causes (such as S/N, abundance, and velocity differences), which may affect if, and at what epochs, C ii is detectable. If metallicity is related to both the NUV dispersion and the strength of C ii features, each would be expected to have a continuous distribution. A larger sample of SNe Ia with measurements of UV color and the strengths of C ii (corrected to some common epoch or otherwise accounting for the time-variation of the C ii strength) is needed to understand this further, treating both parameters as continuous variables.

Foley & Kirshner (2013) and Graham et al. (2015a) conjecture that metallicity is the cause of the mid-UV flux differences of the otherwise nearly identical SNe 2011fe and 2011by based on comparisons with models from Lentz et al. (2000). Multiple theoretical models also show near-UV differences (Höflich et al. 1998; Sauer et al. 2008; Walker et al. 2012; Wang et al. 2012; Baron et al. 2015; Miles et al. 2016), including the near-peak and later models of Lentz et al. (2000). The similarity of the SN 2017erp spectra in the near-UV region to the higher metallicity models of Walker et al. (2012) strengthen our confidence that at least some of the differences in the near-UV are caused by metallicity. The differences between the models, however, and the other factors affecting the UV make us cautious about quantitative statements regarding metallicity using the models. Further work is needed to resolve to what extent the differences between SNe 2011fe and 2011by and between SNe 2011fe and 2017erp are due to metallicity and how much is due to other causes. What we do think is clear, however, is that there are multiple regimes of UV dispersion—SNe Ia with similar optical colors and light-curve shapes can have different near-UV spectra, as shown here, as well as different near-UV photometric colors (Milne et al. 2013; Brown et al. 2017), and SNe Ia with similar optical and near-UV colors and light-curve shapes can have different mid-UV flux (Foley & Kirshner 2013; Graham et al. 2015b).

Foley et al. (2018) further tie the metallicity differences they infer between SNe 2011fe and 2011by to the significant luminosity differences between them without affecting the optical colors or spectra. Recent modeling by Miles et al. (2016) predicts that a change in progenitor metallicity does cause a change in the bolometric luminosity in the same direction as predicted by Timmes et al. (2003) and assumed by Foley et al. (2018). The models also, however, predict a change in the Δm₁₅(B) of up to 0.4 mag, depending on the model, and changes in the temperature and thus spectral shape as well as numerous features in the UV and the optical. This makes it difficult to isolate an effect like metallicity.

In scientific experiments, one changes a single physical characteristic at a time (the independent variable) and measures the difference (if any) on one or more other characteristics (dependent variables). In an observational science, we have a multitude of observables that may or may not be dependent on each other. We may attempt to limit the number of differences by using subsamples where one observable is similar. In this case we compare SNe Ia with similar velocities and Δm₁₅(B) and seek the cause of a difference in the UV flux continuum and features. However, if the cause of the UV flux differences also causes differences in other observables such as intrinsic B – V color, velocity, or Δm₁₅(B), then such sample cuts may hide the relevant correlations.

The multiplicity of models predicting different UV effects from metallicity (Höflich et al. 1998; Lentz et al. 2000; Sauer et al. 2008; Walker et al. 2012; Miles et al. 2016) as well as the other effects that could be going on (asymmetry, density gradients, etc., see, e.g., Brown et al. 2014b) make it seem premature to conclude metallicity is the only cause of either the near-UV or mid-UV differences at this time. Multiple physical differences likely exist between SNe 2011fe and 2017erp, which are not matched by the change of a single parameter like metallicity. What is needed are models that accurately predict the observed UV-optical-NIR spectra and light curves for a subset of SNe and grids of parameter variations. The large sample of UV, optical, and NIR spectra that now exist should be able to constrain the physical parameters that serve as inputs to the models.