The Masses of Supernova Remnant Progenitors in M83

We took the locations of SNRs and strong SNR candidates in M83 from a compilation of catalogs obtained by several surveys. We provide the full combined catalog in Table 1. The survey where most of the SNRs were discovered was from the Magellan 6.5 m telescope at Las Campanas (Blair et al. 2012, B12 in the table). Those SNR candidates found first using HST are from Dopita et al. (2010) and Blair et al. (2014), which are D10 and B14 in the table. We note that the Dopita et al. (2010) catalog covered the complex starburst nucleus where confusion effects prevent resolved photometry even for HST, so many of these objects are excluded from the photometric analysis. There are a handful of SNR candidates in both the Magellan data set and the HST data that have not been reported previously but are included here for completeness. For the SNRs that have identified X-ray counterparts from Long et al. (2014, L14 in the table), a cross reference is also provided. Spectroscopic confirmations of a significant subsample of these SNRs were reported by Winkler et al. (2017). Furthermore, there are six historical SNe whose locations are well-determined, which we also include in the sample. An additional young SNR that may represent an unobserved SN in the last century, SNR 238 (B12-174a), is also included (Blair et al. 2015). Table 1 represents the most complete listing of SNRs and good SNR candidates in M83 to date.

Table 1 also includes a quality flag for each SNR candidate, based on the amount of supporting evidence used to confirm that the object is an SNR. The quality flag values indicate the following categories:

1a: Optical imaging candidate confirmed by spectroscopy and having X-ray emission.

1b: Optical imaging candidate strongly confirmed by optical spectroscopy (e.g., spectroscopic [S ii]:Hα > 0.5) but no X-ray detection.

1c: Optical imaging candidate with photometric [S ii]:Hα > 0.5 but no optical spectral confirmation available and no X-ray detection.

1d: Optical imaging candidate with X-ray but no spectra available.

1e: Optical imaging candidate with X-ray but marginal spectral ratio (likely due to H ii contamination).

1f: Solid Imaging candidate, no X-ray, but with clear H ii contamination in spectrum.

1g: Strong HST [Fe ii] 1.64 μm candidate.

1h: HST Nuclear candidates (mostly Dopita et al. 2010) with no additional supporting evidence.

1z: Special cases.

2: Optical candidate with imaging or spectroscopic ratio 0.25 < [S ii]:Hα < 0.5. Many of these are likely SNRs, but H ii contamination causes uncertainty. Some have detectable [Fe ii], which strengthens their case.

h: Historical supernova (event was observed).

Finally, the catalog has a use flag column, which indicates how each SNR was used or not used in the analysis. This flag has the following categories:

• 1.  
  p: Progenitor mass constrained from measured local young population.
• 2.  
  n: No young population detected in the fit (Type Ia candidate).
• 3.  
  o: Outside of the HST coverage; no measurement performed.
• 4.  
  c: Too close to the center of M83; no measurement performed.
• 5.  
  i: Insufficient photometry for a measurement; fitting failed.

In Figure 1, we show the locations of all of the SNRs for which we were able to obtain sufficient photometry for a measurement (use flags "p" and "n"). The circles mark remnants of use flag "p" and are color-coded by the most likely progenitor mass from our technique. Yellow squares mark the locations of remnants with "n" use flags.

Zoom In Zoom Out Reset image size

The SNR catalog has 307 historical SNe, SNRs, and SNR candidates. There are 271 of these within the seven-field HST/UVIS footprint for which we have resolved stellar photometry. However, 24 of those locations are too close to the complex galaxy nucleus to provide reliable resolved stellar photometry. Of the remaining 247 SNRs, 10 more had samples that proved insufficient to allow the fitting to converge. Our final measured sample therefore consisted of 237 SNR locations (see Figure 1) and a 100 random control sample positions.

For our program, we first ran photometry on all of the resolved stars in the HST imaging data. Then, to quantify the errors and completeness of the photometry as a function of star color, magnitude, and location, we ran over a million artificial star tests (ASTs). As in previous population-based progenitor studies, we then fit the photometry data within 50 pc of each SNR with stellar evolution models to constrain the age distribution of the young stars near each SNR. We now discuss the production and analysis of our resolved stellar photometry.

We measured resolved stellar photometry from the F336W, F438W, and F814W imaging with the point-spread function (PSF) fitting photometry pipeline used for the Panchromatic Hubble Andromeda Treasury (PHAT; Williams et al. 2014a). In short, this pipeline uses an astrodrizzle PyRAF routine to flag cosmic-ray affected pixels, and the DOLPHOT (Dolphin 2000, 2016) PSF-fitting photometry package with the TinyTim-generated PSFs (Krist et al. 2011) to find and measure point-source photometry for all sources in a set of HST images. It starts by reading all of the flc images into memory to search for stars using the full image stack. Then the photometry is performed on the individual flc images, where a PSF is fitted to the location of all star locations in every image. For the multiple exposures in a given filter, the measurements are combined to reduce the photometric uncertainty and increase the signal-to-noise of the measurements.

The pipeline has been updated based on some of the detailed tests performed on the PHAT photometry. In particular, we used the TinyTim PSFs (Krist et al. 2011) and we used the CTE-corrected flc images for photometry instead of post-photometry CTE corrections. Both of these updates have resulted in more precise photometry with smaller systematic uncertainties.

From the DOLPHOT measurements, we first produced full-field color–magnitude diagrams (CMDs) of our photometry. We show examples of these in Figure 2, where it is clear that we have detected the upper main-sequence of blue stars. At the distance of M83, the depth of these CMDs is to m_F435W ∼ 26; M_F435W ∼ −2.5, which corresponds to a main-sequence turnoff age of 70 Myr, which is the lifetime of a 6.3 M_⊙ star in the Padova models. Therefore we are sensitive to all populations relevant for CCSNe progenitor measurements.

Zoom In Zoom Out Reset image size

After this check, we focused our further analysis on the regions within 50 pc of each SNR location, and on a control set of photometry taken from random positions of the same physical size. These control regions were distributed in proportion to the real SNRs in each of the seven M83 UVIS fields, and their photometry samples were run through the same analysis routines as those of the SNR regions. The results from the control samples allow us to quantitatively check that the measurements from the SNR locations differ from the field. Examples of the relevant regions around several SNRs are shown in Figure 3 to demonstrate how well-resolved the individual stars are and to show the diversity of the dust and stellar density properties of the regions.

Zoom In Zoom Out Reset image size

Fitting CMDs with models requires convolving models with the errors, bias, and completeness of the photometry at each color and magnitude. These quantities are best determined through ASTs, whereby a star of known location, color, and brightness is added to the images, and then the photometry of that section of the images is rerun. This operation is run hundreds of thousands to millions of times to cover location, color, and brightness space well enough to properly model the photometric quality of each CMD that we fit.

We ran ASTs covering the range of colors and magnitudes seen in the observed stellar catalogs. The differences between the input star properties and the output measurements allow us to calculate the bias, uncertainty, and completeness of the observed stellar catalog as a function of star color, magnitude, and environment. These quality metrics are then used by the fitting software to convolve model CMDs so that they have the same photometric quality as the observed CMDs. Then we can determine the model CMD that best matches the observed CMD.

ASTs require significant computational resources as they require running the photometric measurements many times. Therefore, to increase the efficiency of ours, we assume regions of similar stellar density and exposure time will have resolved star samples of the same photometric quality. This assumption allows us to apply one set of ASTs to several SNRs. A similar technique was used by the PHAT team to drastically decrease the number of ASTs necessary to measure SFHs as a function of position over the large area covered by the PHAT project (Williams et al. 2017).

To optimize the efficiency of our ASTs, we first measured the relative stellar density as a function of position in the images. To avoid biases due to completeness differences we limited the magnitude range over which we counted stars to where the completeness was high (>80%) over the entire survey (F438W > 26.2). At each image location, we counted stars above this brightness limit within 50 pc and divided by the area to determine the stellar density.

We then tested the assumption that ASTs taken from regions of similar stellar density are interchangeable. For this test, we ran ASTs at the location of an observed stellar sample, as well as in another location in the image with similar stellar density. We then fit the observed CMD of one of the locations twice: once with each set of AST results. We found that ASTs from locations with stellar densities within a factor of two of one another resulted in equivalent age measurements. Thus, we performed 42 sets of ASTs obtained from regions spanning the full range of stellar densities in our sample (0.4–20 arcsec⁻²), and stay well within this factor of two variation. We then applied this library of ASTs during model fitting to our photometry. Namely, when fitting each SNR and control sample, we convolved the models with the photometric errors, bias, and completeness by applying the ASTs measured from a region of appropriate stellar density.

We employ the CMD fitting package MATCH (Dolphin 2002, 2013), which fits the CMDs from resolved stellar photometry to determine the best SFH of each stellar sample near an SNR and control region. The SFH is the distribution of stellar ages and metallicities present in the sample, as well as the uncertainties on that distribution. This package has been used by many investigators to measure the star formation rate as a function of lookback time for a number of nearby galaxies (e.g., Williams et al. 2009; McQuinn et al. 2010; Weisz et al. 2011; Skillman et al. 2017, and many others), as well as to measure star cluster ages (e.g., Senchyna et al. 2015), and the ages and masses of supernova progenitors (e.g., Murphy et al. 2011; Jennings et al. 2012, J14).

We applied time bins starting at 6.6–8.0 log years in steps of 0.05 dex and then from 8.0 to 10.1 log years in steps of 0.1 dex using Padova isochrones (Girardi et al. 2002; Marigo et al. 2008; Girardi et al. 2010). Furthermore, we constrain the present day metallicity from −0.6 ≤ [Fe/H] ≤ 0.1, as appropriate for M83 (Bresolin et al. 2009). We also applied reddening (A_V) and differential reddening (dA_V) to our model fits (see Section 2.5). In addition, the MATCH package allows the inclusion of a contamination CMD to look for populations that are present in the sampled region at higher density than the surrounding field. We used the entire WFC3 162'' × 162'' UVIS M83 field surrounding the SNR, scaled to the area of the extraction region area, for this contamination CMD.

M83 is characterized by an abundance of dust embedded in the disk. For extinction, it is valuable to consider multiple colors. Therefore, we consider both the F336W–F438W and F438W–F814W CMDs when finding the extinction properties of each location. These extinction properties are both the overall extinction affecting the entire location, A_V (with the minimum bounded by the foreground Galactic extinction of 0.2; Schlafly & Finkbeiner 2011), and extinction spread due to the distribution of stars along the line of sight (dA_V). Each of these parameters affect the CMD differently, and are therefore they applied to the models independently. First, A_V is applied to all of the stars in each model CMD. Then dA_V is applied to the model CMD to spread the stars along the reddening line as they would be if they were randomly distributed within a uniform dust layer. In essence, A_V affects the location of features in the CMD (e.g., the main sequence), while dA_V affects the sharpness of the features in the CMD. For example, stars embedded within dusty regions of a galaxy (high dA_V) will produce CMDs with features smeared along the reddening line, even if they are behind little foreground A_V, while stars not embedded in a dusty region (low dA_V) will produce CMDs with sharp features, even if all of the stars are behind a large foreground A_V.

We determine the values for A_V and dA_V by finding the best maximum likelihood fit value reported by MATCH (Dolphin 2012) for a grid of possibilities when fitting (see Figures 4 and 5). We then fix the A_V and dA_V values to those determined from the multiple colors when fitting the deepest single CMD for the final age distribution and uncertainties out of these, as described below.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

View figure set (199 panels)

Tarball of figure set images

Zoom In Zoom Out Reset image size

Our uncertainties are mainly due to photometric error, incompleteness, and the number of stars in our sample available to constrain the age distribution of the young component of the population. All of these vary significantly for the SNR sample. For example, the number of stars ranges from 10s to 100s of stars per SNR. Thus we measure the uncertainties separately for each SNR sample.

We characterize these uncertainties using a hybrid Monte Carlo (MC) algorithm within MATCH using the task hybridMC (Dolphin 2013). This task accepts or rejects thousands of potential SFHs surrounding the best-fitting one, based on likelihood. It then applies the distribution of accepted SFHs to determine the 68% uncertainty regions surrounding the best fit. The hybridMC uncertainty determination applies statistics that are valid only for independent CMDs. Applying these statistics to two CMDs that share a filter would produce incorrect uncertainties because the CMDs are correlated. Therefore, we must choose one of our CMDs when fitting for the final SFHs and uncertainties. The M83 data contain imaging in F336W, F438W, and F814W.

To determine which CMD fit to use for our final SFH and uncertainties, we fit both the F336W versus F336W–F438W and the F438W versus F438W–F814W CMD independently, fixing A_V and dA_V measured values (see Section 2.5). We then compared the best-fit SFH to the best-fit SFH from the simultaneous fit of both CMDs. We found overall consistency between the techniques, but the single CMD fits had reliable uncertainties. The best constraints were almost always from the fits to the F336W versus F336W–F438W CMDs. This result was expected because the F814W photometry is the shallowest and least sensitive to the young populations. However, there were a handful of cases, mostly those with the highest reddening, where the F336W–F438W CMD alone did not produce a result consistent with the two CMD fit from the A_V, dA_V runs (185, 278, 059, 197, 090, 117, 150, 227, 246, SN 1950B). In these cases we found that the fit to the F438W versus F438W–F814W CMD produced a result consistent with the two CMD fit. We report as our final measurement of the SFH and uncertainties, the results from the F336W versus F336W–F438W CMD, except for these exceptions.

Our method for constraining the progenitor mass is similar to that of previous work (e.g., Jennings et al. 2012; J14; Williams et al. 2018). The method begins by finding the most likely age of the progenitor star. In short, we measure the SFH from broadband photometry of the resolved stars within 50 pc of the location of each SNR. This region size assumes that most young stars are coeval within our desired extracted region, as suggested by the work of Bastian & Goodwin (2006). Indeed, the validity of this assumption has been strengthened by the works of Gogarten et al. (2009) and Murphy et al. (2011).

We restrict the ages in the SFH that we apply to our progenitor analysis. The ages of SNR progenitors measured with deeper photometry in more nearby galaxies have established that the maximum progenitor age for CCSNe is <50 Myr (young enough for a ∼7 M_⊙ star to still exist). While older progenitors are allowed according to binary evolution models (e.g., Xiao et al. 2019), a maximum of 50 Myr is empirically measured for CCSNe progenitors by J14 and confirmed by Díaz-Rodríguez et al. (2018) in M31 and M33, even when they include older ages in their analysis. Furthermore Xiao et al. (2019) find a similar empirical age constraint using an independent emission line fitting technique that uses the latest binary evolution models to infer ages of H ii regions that have hosted CCSNe.

We take advantage of these empirical results to limit the part of the SFH that we apply to determine the progenitor mass. We assume only the part of the SFH <50 Myr is relevant for determining progenitor ages. When we infer the most likely progenitor age of each SNR, we look for the most prominent peak in the age distribution, only considering ages <50 Myr. This assumption means that we are not able to use these data to probe the high-age (low-mass) end of the progenitor mass function, but it allows us to avoid including the higher age bins, which have large uncertainties at the photometric depth of the observations used here, into our age determinations.

We derive the mass fraction for each time bin over the past 50 Myr, taking advantage of the empirical constraints so that we do not need to include older ages that have main-sequence turnoffs near our completeness limit, which lead to large SFH uncertainties. Furthermore, these older bins are more likely to contain stars unrelated to the SNR progenitor because stars of older ages have had more time to migrate away from their siblings. Thus, limiting the ages to those known to be relevant to CCSNe progenitors helps to minimize contamination from unrelated populations.

Figures 4 and 5 summarize the progenitor mass measurement technique for two SNR locations (SNR 295 and SNR 057, respectively). We include figures of this format for all 199 SNRs with progenitor constraints in the supplemental materials. The effect of the assumed differential extinction is shown in the upper left panel, where the resulting age distribution from each assumed differential extinction amount is plotted with a different color. In Figure 4 the general pattern in ages is relatively insensitive to the assumed dA_V. The lower left panels of the figure show the star formation history (SFH) along with uncertainties for the best dA_V result. Peaks in the rate and stellar mass fraction correspond to the most likely progenitor age.

The full SFHs show how well the median age represents the full age distribution of the population. An age distribution may correspond to several epochs of SF or may be dominated by one epoch. In the case of the former, the median age is less likely to represent the correct age of the progenitor. The full probability distribution for each event is a more precise characterization of the complex uncertainties.

Based on these measurements, we use the age where 50% of the stars (as determined from the bottom left panel) have formed as the most likely progenitor age. We then take our uncertainties to include all ages that may contain the median age. This is shown with the tan vertical shaded region in the bottom left panel of the mass measurement figures (e.g., Figures 4 and 5). These limits define our 1σ progenitor age range for the progenitor mass. From this age range, we infer the zero age main-sequence (ZAMS) mass range using the Girardi et al. (2010) isochrone library as a lookup table for the most massive star present at any given age. We set a floor for the uncertainty to be at least 0.1 M_⊙. In the cases where we find there to be zero SF in the last 50 Myr, we assume these to be the remnants of SNe Ia or runaway massive stars, and we provide no measurement of M_ZAMS.

The derived progenitor masses are plotted on an image of M83 in Figure 1 to show the spatial distribution as a function of derived progenitor mass. The symbol colors denote derived mass with masses above 20 M_⊙ shown as red, and lower-mass ranges progressing from orange to green to blue. Yellow squares designate those SNRs for which no young stellar population was identified, and these are the candidates for SNIa progenitors. The higher-mass progenitors generally follow the spiral arms, but we note that the most massive progenitor masses appear to be most closely associated with arms and/or star-forming regions. This result is consistent with findings in clustering studies that show trends where stars of higher masses are more clustered (e.g., Kang et al. 2009; Bianchi et al. 2012; Kim et al. 2012). Lower-mass progenitors, while also associated with the arms and star-forming regions, seem to show a broader spatial distribution. We also see that the Type Ia candidates tend to be even more generally distributed, including the interarm regions.

Our sample of 199 SNe and SNRs with inferred progenitor masses is the largest ever produced for any single galaxy. Thus the distribution has the potential to yield the tightest constraints on the shape of the progenitor mass distribution to date. We plot the distribution of all the progenitor masses in Figures 6 and 7. Figure 6 shows a histogram of the progenitor mass distribution for M83 along with the combined M31+M33 (J14) distribution for comparison. While the number of progenitors in the lowest mass bin is comparable in Figure 6, the fraction of the total in that lowest bin is less, with significantly more progenitors in the 10–15 M_⊙ range and near 20 M_⊙. The M83 distribution has a significantly higher proportion of higher-mass progenitors as well.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

In Figure 7, the data points mark the masses at the median ages as described in Section 2.6, along with their uncertainties. They are plotted in order of mass to show the cumulative mass distribution. Lines mark fits from various power-law mass distributions. The left panel is the best-fit to the data, with an index of $-{3.0}_{-0.7}^{+0.2}$ (χ²/dof < 1.1). The middle panel has lines drawn from a Kroupa IMF (index of 2.3), and the right panel has lines drawn from the index 4.4 that fits the M31+M33 SNR distribution from J14. Note that this plot excludes the assumed SNe Ia, or objects where no young stellar population was detected.

In addition, we can study only the spectroscopically confirmed SNR sample (quality flag 1a in the Table 1), along with the observed SNe, which comprises a subsample of 40. The mass distribution of this subsample has a best-fitting power-law index of 2.8, making it slightly closer to a standard IMF. Moreover, a more recent reanalysis of the M31+M33 progenitor distribution that uses Bayesian inference to better account for the uncertainties, finds a power-law index for that distribution of $-{3.0}_{-0.5}^{+0.3}$ (Díaz-Rodríguez et al. 2018), which is consistent with this M83 result.

We also compare the SNR distribution to the control sample. The first difference is that the random positions have double the fraction of locations with no young population. In addition, the distribution of inferred masses for the random positions is significantly steeper, with a best-fit power-law index of −3.7. The SNR locations appear to be providing a distribution different from the general field. The facts that the random field is steeper and that the spectroscopically confirmed sample is less steep than the full sample suggest that contaminants in the full sample may be pushing the best-fitting power-law index higher.

We apply a Kolmogorov–Smirnov test against the mass distribution of M31+M33 J14, and obtain a P-value of 0.02, suggesting only a 2% chance of them having the same parent distribution. Indeed, visually, there is a noticeable difference particularly around ∼10 M_⊙ and with more massive progenitors. Thus, the M83 SNR sample really does appear to have more high-mass progenitors than those in M31 and M33.

While Figure 6 shows that the M83 distribution is more top-heavy than the M31 and M33 samples, it is still not consistent with a standard IMF (Kroupa 2001); there are too few high-mass progenitors for the number at the lower-mass end. This result may be due to biases in our technique against the youngest ages dominating the local stellar mass. For example, SNe that occur within very young associations, superbubbles, or highly photoionized regions may not leave visible SNRs, thus biasing our survey based on optical SNRs. However, there is also a possibility that some fraction of the most massive stars do not explode as SNe. This possibility seems to be corroborated by the discovery of a vanishing red supergiant in NGC 6946 with no observed SN (Adams et al. 2017b; Murphy et al. 2018).

It does appear to be the case that at least some stars >30 M_⊙ explode as SNe. One such high-mass progenitor is that of the SNR 057 (see Figure 5). This case is the highest progenitor mass constraint to date from population fitting techniques. The surrounding population in this case is limited almost entirely to the youngest ages for which we have models, making the progenitor highly likely to have been >20 M_⊙. This single measurement provides some of our strongest observational evidence to date that such massive stars produce SNe, suggesting that any shortage of progenitor masses >20 M_⊙ is more likely to be attributable to a bias against young ages or some massive stars failing to explode, rather than to a hard high end cutoff in the progenitor mass distribution. Furthermore, there has been some evidence from stellar population analysis that such high-mass progenitors may be more likely to produce stripped envelope SNe (Maund 2018), making SNR 057 a candidate remnant from a stripped envelope event.

The M83 sample includes six historical supernovae, where the actual explosion was observed. These include SN 1923A, SN 1945B, SN 1950B, SN 1957D, SN 1968L, and SN 1983N. For these, the types are known to be II, unknown, unknown, unknown, II, and Ib, respectively. Thus, the three known types were all core-collapse SNe, which is not surprising given the high star formation rate of M83. Of these, SN 1968L occurred deep in the complex nuclear starburst region where its parent population is not resolved even at HST resolution, but the high star formation rate in this region favors a core-collapse event. For SN 1923A, we find a likely massive progenitor (${53}_{-44}^{+13}$ M_⊙), as with SN 1983N (${20}_{-5}^{+6}$ M_⊙). Of the three unknowns, SN 1945B did not have sufficient local stellar photometry for the fit to converge.

SN 1950B and SN 1957D are candidates for CCSNe progenitors, with masses of ${7.6}_{-0.3}^{+3.3}$ M_⊙ and ${7.6}_{-0.3}^{+8.2}$ M_⊙ respectively. This measurement for SN 1957D is lower than that of Long et al. (2012), who found a lower-mass limit of >17 M_⊙, also looking at the local stellar population. Our analysis detects the presence of two populations at this location: one at an age consistent with the analysis of Long et al. (2012), and an underlying one at ∼40–50 Myr, which could have produced a lower-mass progenitor. Thus, this SNR is a good example of why it is important to consider the full probability distribution, as the single value with uncertainties does not always capture multiple peaks in the age distribution. Interestingly, a similar situation occurs with the SNR 238, which Blair et al. (2015) found to be hosted by a population of very young stars, suggesting a progenitor with a mass >17 M_⊙. We also detect this young population; however, we also detect a population with an age of 40–50 Myr, making our mass measurement ${8.1}_{-0.4}^{+12.5}$ M_⊙. Again, the double-peaked age distribution is important to consider.

Thus, it appears that five to six of the six historical SNe are core-collapse (with SN 1968L very likely core-collapse and SN 1945B still unknown), which is similar to the fraction of the total SNR population and roughly consistent with expectations for an Sbc galaxy like M83 (∼20%, Li et al. 2011). Furthermore, the four historical SNe for which we have measured ages appear to cover the full range of progenitor masses to which we are sensitive.