Seams and Tears in the Type Ia Supernova Map

Astronomers sort exploding white dwarfs (Type Ia supernovae) into subtypes: "normal," dim and fast-fading "91bg-like," and hot, bright "91T-like." These labels matter for cosmology and for identifying which stellar systems explode. Recent work argues that the subtypes are points on one continuous family, not separate kinds of explosion. The evidence so far, however, cannot tell two situations apart. In the first, the population is continuous but thin in places: transitional objects exist and are merely rare. In the second, there is a real gap, meaning two physically distinct channels whose outskirts happen to look alike. A clustering method that finds two clumps cannot distinguish these, and a two-dimensional picture of a deep-learning embedding that "looks unbroken" cannot either.

We propose to borrow a structural idea from how the brain lays out the body surface. Somatotopic maps, the patches of cortex where neighboring neurons respond to neighboring patches of skin, are built from local patches stitched together. Where the skin is continuous, the patches join smoothly, forming a seam. Where the sensory surface is truly disconnected, the map has a border, a tear. The rodent whisker cortex makes this concrete: it has one discrete column ("barrel") per whisker, with continuous structure inside each column. We build a generative model of supernova spectra and light curves out of such stitched patches ("charts"), and we ask of each adjacent pair whether gluing them costs predictive accuracy on held-out supernovae. Because the question is about joining patches rather than counting clumps, a thin but connected bridge and a genuine gap give different answers. We calibrate that answer on simulated populations where the truth is known.

Core Claim

Current evidence mixes up two properties: continuity of the support (the region where supernovae can exist at all) and unimodality of the density (whether they pile up in one clump or several). Density-based tools such as K-means and mixture-model comparison answer the second question. Visual embeddings answer neither rigorously. A model built as an atlas of local charts, with a learned pattern of which charts touch and an explicit penalty for mismatch along their shared edges, can test support continuity directly. We expect the test to be well calibrated in two cases where density methods fail. It should report "continuous" for populations that are bimodal but connected. It should report "torn" for populations with a real gap, even when observational noise smears the gap.

Applied to the Zwicky Transient Facility (ZTF) SN Ia second data release, the test gives a quantitative, uncertainty-aware verdict on two open disputes. The first is the normal–91T boundary, which 1991T-Like Type Ia Supernovae as an Extension of the Normal Population argues is a smooth physical gradient. The second is the normal–91bg boundary, which SN 2012ij argues is bridged by at least about 2% of all SNe Ia, while noting that no explosion model fits every spectral feature.

Where the Pieces Come From

The motivating evidence comes from the supernova side, and each piece has a specific limitation. Exploring the spectroscopic diversity of type Ia supernovae with Deep Learning and Unsupervised Clustering compresses spectra into four latent dimensions. It shows the subtypes lying along an unbroken manifold, but only by K-means clustering and Isomap visualization, using roughly 150 maximum-light spectra. The O'Brien et al. paper fits a neural emulator of a radiative-transfer code to 40 supernovae and finds 91T-like events at the edge of the normal parameter range. That result rests on a small sample of single-epoch, one-dimensional models. SN 2012ij supplies a single bridging object and a rate estimate. Normal or transitional? is a constraint on our design: it shows that light-curve shape and spectral line strength can disagree about the same object. A continuity test using only photometry could therefore invent or hide tears, so our model encodes light curves and spectra jointly.

The method comes from the topographic side, from two papers that each supply one piece. Localist Topographic Expert Routing builds a barrel-cortex network with one module per sensor, weights shared across modules, and communication restricted to the four nearest neighboring modules. Sharing cut parameters by 97% without losing accuracy. We reuse this pattern of shared-weight modules with sparse neighbor links. We change what the modules are: they are charts over the space of physical explosion properties, not sensor channels, and the neighbor links become the seams we test. Topographic VAEs learn Equivariant Capsules places latent variables on a grid whose neighbors share variance. It shows that a known continuous transformation can become a shift along that grid, but its grid is fixed in advance, and fixing it assumes the answer to our question. We instead learn which charts are adjacent and let the data decide whether each seam holds.

Evaluation of Representational Similarity Scores Across Human Visual Cortex compares ways of matching a representation's geometry to physical layout and finds first-order shape metrics the most stable across subjects. We use such a metric only for evaluation: to check whether the learned atlas preserves the geometry of true simulated physical parameters.

Research Program

The model has $K$ charts. Each chart $k$ carries a low-dimensional coordinate $u \in [0,1]^d$ with $d \le 4$, following the four-dimension finding of Ishida et al. One shared decoder maps a chart index, a coordinate, and the observation phase $t$ (days from peak) to a spectrum and multi-band fluxes, with a noise model for each instrument.

Phase enters the decoder as an input. This separates time evolution from physical identity, and it lets every epoch of the same supernova constrain one shared coordinate. The model therefore uses all 5,138 ZTF spectra and 3,591 light curves rather than a maximum-light snapshot. An encoder amortizes inference, so that each object gets a posterior over which chart it belongs to and where it sits within that chart. Objects with only photometry receive correspondingly wider posteriors.

Adjacency between charts is learned, with each chart linked to at most a few neighbors. Along each shared edge, a seam penalty with strength $\lambda_{kl}$ forces the two charts' decoded outputs to agree. For each adjacent pair we fit two versions: glued ($\lambda_{kl}$ large) and torn ($\lambda_{kl}=0$). The seam statistic is

$$\Delta_{kl} = \frac{1}{N_{kl}} \sum_{i \in \text{held-out}} \left[ \log p_{\text{glued}}(x_i) - \log p_{\text{torn}}(x_i) \right],$$

where $x_i$ is the full data for held-out supernova $i$ and $N_{kl}$ is the number of held-out objects whose posterior mass lies mostly in charts $k$ or $l$. Positive $\Delta_{kl}$ means gluing helps. That happens when bridge objects are well predicted by interpolating across the seam. Strongly negative $\Delta_{kl}$ means a forced join puts probability where no supernovae live. We obtain a null distribution for $\Delta_{kl}$ by parametric bootstrap: we generate data from the fitted glued model at ZTF cadence and noise, then refit.

The program has three stages. First, build and calibrate the test on simulations with known ground truth. Second, apply it to ZTF DR2 and to the 40-object WISeREP sample from O'Brien et al. Third, relate chart coordinates to physical parameters.

Baselines

The naive reference point takes the standard two-number light-curve summaries from SALT2 fits already provided by ZTF DR2: stretch $x_1$ and color $c$. It fits one- versus two-component Gaussian mixtures by the Bayesian information criterion. This is what a working cosmologist would do. It sees only photometry, and the ZTF overview reports that an uncorrected detector nonlinearity shifts $x_1$ by an average of −0.14. Beating it means correct verdicts on simulated populations where it fails.

The strongest existing approach to the continuity question is the Ishida et al. pipeline: an autoencoder, then K-means and Isomap. We retrain it on the same ZTF spectra and add a quantitative verdict it lacks, the connectivity of a nearest-neighbor graph in its latent space, so that it is scored rather than eyeballed. No reported result found on Lacuna for any continuity-versus-gap test on supernova data, so there is no published number to beat, and the comparison is entirely on our calibrated simulations.

For reconstruction quality, the strongest generative model is DiTSNe-Ia from Variational diffusion transformers for conditional sampling of supernovae spectra. It reports a mean squared error of 0.108 on the Goldstein & Kasen grid, against 0.508 for SALT3, with after-peak MSE of 0.0191 and 90% interval coverage of 0.845. We do not need to beat this. The atlas must stay within a modest factor of it on the same held-out split, which shows that the seam test is not running on a badly underfitting model.

The most dangerous competitor is a phase-conditioned variational autoencoder, a standard compression network, with one unstructured latent space. It uses the same encoder, decoder, data, and tuning budget as our model. It would be paired with mixture-model comparison and with persistent-homology or nearest-neighbor connectivity tests on its latent codes.

Two ablations isolate the key ingredient. The first removes all seams, which yields a plain mixture of local factor models and is equivalent to asking only about density. The second fixes chart adjacency on a grid in the style of the Topographic VAE instead of learning it.

All models get the same data splits, the same injected noise, and equal hyperparameter-search budgets measured in GPU-hours. Beating a baseline means a higher area under the curve for detecting true gaps at matched false-tear rate, above all on the bimodal-but-connected control.

Why not just…?

Why not just compare a one-component and a two-component mixture model? This answers whether the density has two clumps, not whether the support is torn. A population with a 2% bridge, the rate SN 2012ij implies for normal–91bg, will usually be declared two classes. The bimodal-but-connected simulation below is designed to show this failure directly. If mixture comparison unexpectedly handles that control well, our claim is wrong.

Why not just run a topology or connectivity test on an ordinary VAE latent space? This is the real threat. Nearest-neighbor connectivity and persistent homology do target support continuity, and they need no new model. Their weaknesses are that they work on noisy point estimates of latent codes, they depend on a chosen neighborhood scale, and they cannot tell a sparsely sampled bridge from noise that merely scatters points into a gap. Our test puts the noise model inside the likelihood and compares predictions on held-out objects. The calibration experiment scores both methods on identical simulated populations across gap widths and bridge fractions. If the latent-space topology test matches the seam test, we will report that. The contribution would then shrink to the phase-aware, multi-epoch joint encoder plus the calibration suite, and we would recommend the simpler estimator.

Why not just run full physical inference, emulator plus nested sampling, on everything? The O'Brien et al. approach gives physically interpretable answers, but it has been applied to 40 objects with one-dimensional single-epoch models. The SN 2012ij paper shows that the underlying explosion models themselves fail on some features, so a continuity verdict inside a model family inherits that model's blind spots. We use its 40-object results as an external check: the atlas should call the normal–91T seam glued, as O'Brien et al. found. The atlas asks the continuity question in a data-driven space that does not depend on any single explosion model.

Proposed Experiments

Calibration uses the Goldstein & Kasen grid of 4,500 simulated SNe Ia, whose physical parameters are known. We build three kinds of synthetic population, each resampled at ZTF cadence and noise with host redshifts perturbed.

Continuous populations draw parameters from a connected region. Gapped populations remove a slab of nickel-56 mass or ejecta mass of controlled width, and optionally leave a controlled fraction of bridge objects between 0.5% and 5%. The decisive control, bimodal-but-connected populations, draws parameters from a two-peaked distribution with no gap at all.

We report the false-tear rate on continuous and bimodal populations, and the true-tear area under the curve as a function of gap width, bridge fraction, and signal-to-noise. The seam test is compared against every baseline above. Success means that seams are held on the bimodal control, where mixture comparison declares two classes, and that the seam test beats latent-space topology at small gap widths.

Two further checks run on the same simulations. Chart coordinates are compared with the true physical parameters using a first-order shape metric, with a raw-latent VAE as reference. Reconstruction MSE is reported against DiTSNe-Ia's 0.108.

On real data, the model is trained on ZTF DR2 using the 2,667 objects that pass quality cuts plus all their spectra, and we report the seam statistics between the charts that absorb normal, 91T-like, and 91bg-like objects. Two external checks follow. The seam near 91T-like objects should be glued, consistent with the WISeREP analysis. Objects like SN 2012ij and SN 2020ue should sit near seams with broad chart posteriors. From the fraction of posterior mass in seam regions we estimate the transitional rate and compare it with the lower limit of about 2% from the SN 2012ij paper.

The result that would change the field's picture is a seam with a significantly negative $\Delta$ that survives the bootstrap null and the mixture-of-factor-models ablation. That would be data-driven evidence for a distinct channel. The most likely location is the 91bg side, where the SN 2012ij paper finds neither Chandrasekhar-mass nor sub-Chandrasekhar-mass models fully adequate. A uniformly glued map, calibrated to detect gaps of a stated width, would instead be a quantitative version of the continuum claim, with detection limits attached.