Supernova Refsdal: Lensed Transient & Cosmography

• Supernova Refsdal is a strongly lensed Type II supernova with multiple, resolved images that validate theoretical time-delay predictions.
• The event showcases a multi-scale lensing configuration, where both cluster-scale and galaxy-scale lenses work together to produce distinctive image patterns.
• Time-delay measurements from its images have been used to tightly constrain the Hubble constant, illustrating the potential of lensing for cosmology.

Supernova Refsdal is the first known strongly lensed supernova with multiple resolved images, discovered in the field of the galaxy cluster MACS J1149.5+2223. It exploded in a spiral host galaxy at redshift $z=1.49$, while the principal lensing structures lie at $z \approx 0.54$. The system realized Sjur Refsdal’s 1964 proposal that time delays between multiply imaged transients can probe both the lensing mass distribution and the cosmic expansion rate. Its importance derives from an unusual conjunction of observables: an Einstein-cross quartet around a cluster member galaxy, a later reappearance at a different host-galaxy image, rich host-galaxy morphology for lens reconstruction, and sufficiently precise light curves to support time-delay cosmography (Kelly et al., 2014).

1. Discovery and lensing configuration

SN Refsdal was discovered as four point sources, labeled S1–S4, in an Einstein-cross configuration around an early-type cluster galaxy in MACS J1149.5+2223. The host galaxy is a face-on spiral at $z = 1.491$, strongly lensed by the cluster, while the local four-image pattern is generated by secondary lensing from a nearby cluster member galaxy, denoted G3 in the revised lens analysis. The host itself is lensed into three full images and one partial image, and the supernova occurs on a spiral-arm segment that is particularly sensitive to the local lens potential (Sharon et al., 2014).

The geometry is therefore intrinsically multi-scale. Cluster-scale lensing creates the widely separated host-galaxy images, and galaxy-scale lensing creates the local Einstein cross. In later cosmographic treatments, the system is described as having six images of the same supernova: S1–S4, the later image SX, and a predicted but unobserved SY. The time-delay baseline between the earliest and latest detected image is of order one year, which is unusually favorable for direct transient-based time-delay measurements in a cluster lens (Liu et al., 2024).

This configuration made SN Refsdal distinct from earlier lensed quasars and galaxies. The source is transient, so relative arrival times are measurable directly from light curves, while the host is morphologically resolved, so the local and global lens potentials can be constrained not only by image centroids but also by internal structure.

2. Supernova classification and intrinsic properties

Photometric and spectroscopic follow-up classified SN Refsdal as a luminous and blue SN 1987A-like Type II supernova. Its Hubble Space Telescope light curve showed a slow rise over $\sim 150$ days to a broad peak, and the spectra provided strong evidence for a broad H$\alpha$ P-Cygni profile at the host-galaxy redshift. The measured H$\alpha$ expansion velocities were $v_{\mathrm{H}\alpha} = -8356 \pm 1105$ km s$^{-1}$ from the HST grism at phase $-47$ days and $v_{\mathrm{H}\alpha} = -6465 \pm 2918$ km s$z \approx 0.54$0 from X-shooter at phase $z \approx 0.54$1 days. The same analysis found a subsolar host-galaxy metallicity near the explosion site, with $z \approx 0.54$2 from MOSFIRE and $z \approx 0.54$3 from X-shooter, and noted that if the event is modeled as a scaled version of SN 1987A, the ejecta mass is $z \approx 0.54$4 solar masses (Kelly et al., 2015).

A later radiation-hydrodynamics analysis pushed the physical interpretation further and argued that the progenitor was likely a more massive and energetic analog of SN 1987A, specifically a blue supergiant. The best-fit parameters were

$z \approx 0.54$5

$z \approx 0.54$6

That work also used physically modeled light curves, rather than only empirical templates, to refine relative time delays and magnification ratios for S2–S4 and to derive an estimate of $z \approx 0.54$7 from the lens-model predictions (Baklanov et al., 2020).

Taken together, these analyses place SN Refsdal within the SN 1987A-like family but at the luminous and blue end of that class. A plausible implication is that the combination of low metallicity, strong magnification, and exceptionally dense monitoring made it possible to characterize a high-redshift blue-supergiant explosion in substantially greater detail than is usually possible at $z \approx 0.54$8.

3. Lens modeling and predictive success

The first major modeling refinement after discovery replaced sparse knot-based constraints with a denser set of emission knots distributed along the spiral arms of the host galaxy and incorporated the precise positions of the four SN images. The model increased flexibility by freeing the parameters of the cluster galaxies nearest images 1.1 and 1.2, including ellipticity, orientation, core and cut radii, and velocity dispersion. It also moved to galaxy catalogs directly based on archival HST imaging. In that framework, the image-plane scatter for host-galaxy knots improved from $z \approx 0.54$9 in the earlier model to $z = 1.491$0 for most individual features, and the source-plane reconstruction placed the SN on a spiral arm (Sharon et al., 2014).

In parallel, a free-form analysis based on WSLAP+ used HII regions in the spiral host as localized constraints, together with other reliable lensed galaxies, to predict the future appearance of the supernova. That model predicted a reappearance at

$z = 1.491$1

and on November $z = 1.491$2 2015, with an estimated error of $z = 1.491$3 days, corresponding to an accuracy of $z = 1.491$4 (Diego et al., 2015).

A coordinated comparison of seven lens models from five independent methods framed the event explicitly as a blind test of extragalactic prediction. Those models agreed reasonably well with the measured time delays and magnification ratios between the known images even though those quantities were not used as inputs, and they predicted that a future image would reach peak in the first half of 2016 while another image had appeared between 1994 and 2004 (Treu et al., 2015).

The subsequent detection of SX made SN Refsdal the first case in which the appearance of a supernova at a particular time and location in the sky was successfully predicted in advance. This predictive success did not eliminate model uncertainty, but it established that cluster-scale strong-lens models could be confronted with a genuinely blind transient test on a human timescale.

4. Time delays, magnifications, and the reappearance SX

The reappearance was confirmed in HST observations taken on 11 December 2015, which showed a new source at the predicted position approximately $z = 1.491$5 from S1–S4. That detection supplied the decisive long time delay for cosmography and the first blind test of both time-delay and magnification predictions for a lensed supernova (Kelly et al., 2015).

Before SX, the first-year HST campaign measured relative delays and magnification ratios among S1–S4 using both SN 1987A-like templates and more flexible polynomial fits. The polynomial fits yielded $z = 1.491$6 days for S2–S1, $z = 1.491$7 days for S3–S1, and $z = 1.491$8 days for S4–S1, with magnification ratios $z = 1.491$9, $\sim 150$0, and $\sim 150$1, respectively. The template-based fits were consistent but somewhat broader (Rodney et al., 2015).

A later blinded analysis of all five observed images introduced a new photometry pipeline based on DOLPHOT and measured fluxes from difference images. It combined four independent light-curve methods and propagated uncertainties from microlensing and millilensing. The resulting measurements were (Kelly et al., 2023):

Image pair	Time delay (days)	Magnification ratio
S2–S1	$\sim 150$2	$\sim 150$3
S3–S1	$\sim 150$4	$\sim 150$5
S4–S1	$\sim 150$6	$\sim 150$7
SX–S1	$\sim 150$8	$\sim 150$9

The $\alpha$0-day SX–S1 delay corresponds to a $\alpha$1 precision on the time delay and $\alpha$2 precision for the magnification ratios in that analysis. The same work emphasized that the uncertainties include millilensing and microlensing, and that the latter is relatively mild because the stellar surface-mass fraction along the relevant lines of sight is low (Kelly et al., 2023).

The long delay is the key observable for cosmography, but the short delays remain astrophysically informative because they probe the detailed mass profile of the galaxy-scale perturber embedded in the cluster potential. Differences among predicted and measured short delays have therefore been used less as failures of the method than as diagnostics of local lens-model fidelity.

5. Time-delay cosmography and the Hubble constant

The cosmographic use of SN Refsdal rests on the standard lensing relation

$\alpha$3

so that, for a fixed mass model, $\alpha$4 (Grillo et al., 2018). In one formulation used for the SX analysis,

$\alpha$5

which makes the leverage of the SX–S1 delay on $\alpha$6 explicit (Kelly et al., 2023).

Early cluster-lens cosmography with SN Refsdal already showed that the longest measurable delay, SX–S1, dominates the $\alpha$7 constraint. A blind-systematics study found an approximately linear relation between that delay and $\alpha$8, and concluded that a $\alpha$9 uncertainty on the SX–S1 delay translates into approximately $\alpha$0 and $\alpha$1 uncertainties for $\alpha$2 and $\alpha$3, respectively, with statistical uncertainties dominating the total error budget over the specific systematics tested (Grillo et al., 2020).

Selected $\alpha$4 determinations illustrate how the inferred value depends on the adopted lens-model ensemble and inference strategy:

Methodology	$\alpha$5 result	Source
Re-scaling time delays predicted by different lens models to match radiation-hydrodynamics delays	$\alpha$6 km s$\alpha$7 Mpc$\alpha$8	(Baklanov et al., 2020)
Blinded analysis using eight cluster lens models	$\alpha$9 km / s / Mpc	(Kelly et al., 2023)
Cosmological-model-independent Gaussian-process method with eight mass models	$v_{\mathrm{H}\alpha} = -8356 \pm 1105$0 km s$v_{\mathrm{H}\alpha} = -8356 \pm 1105$1 Mpc$v_{\mathrm{H}\alpha} = -8356 \pm 1105$2	(Li et al., 2024)
Combination of 23 lens mass models with equal weighting	$v_{\mathrm{H}\alpha} = -8356 \pm 1105$3 km/s/Mpc	(Liu et al., 2024)

These results do not collapse to a single consensus value, but neither do they imply catastrophic instability. A 2024 systematic study using 23 lens mass models concluded that the dependence on the choice of lens mass models is not significantly large and found clear correlations between best-fitting $v_{\mathrm{H}\alpha} = -8356 \pm 1105$4 values and both the radial density profile of the lensing cluster and the magnification factors of the supernova images. That study argued that additional multiple images at varied radii would tighten the cluster density profile and further improve the cosmographic constraint (Liu et al., 2024).

A common misconception is that SN Refsdal by itself has already reduced cluster-lens cosmography to a systematics-free problem. The literature instead supports a more precise statement: in current analyses, the longest time delay is measured extremely well, and the dominant residual limitation has shifted toward lens-model structure and its controlled comparison across parametric, hybrid, and extended-image approaches.

6. Environment, data products, and later developments

SN Refsdal also became a probe of its local interstellar environment. MUSE observations detected MgII emission at all observed and model-predicted SN positions, with weak FeII* emission at two positions. The measured [OII]/MgII ratios of $v_{\mathrm{H}\alpha} = -8356 \pm 1105$5–$v_{\mathrm{H}\alpha} = -8356 \pm 1105$6 indicate a high degree of ionization with low metallicity. Because the same ionization state is present at all positions, including those corresponding to phases before and after the explosion, the study concluded that the high ionization was produced by previous supernovae or a young and hot stellar population rather than by the current SN. No variability of the [OII] line was detected over 57 days (Karman et al., 2015).

MUSE spectroscopy simultaneously strengthened the lensing analysis. An early program measured 117 secure redshifts in the cluster core, including 68 cluster members and 18 multiple images from 7 lensed sources, and used these data with HST imaging to build six strong-lensing models. The best of those reproduced observed multiple-image positions with an rms offset of $v_{\mathrm{H}\alpha} = -8356 \pm 1105$7 and predicted the next emerging image to peak between March and June 2016 (Grillo et al., 2015). A later northern-field MUSE campaign expanded the spectroscopic catalog by 162 secure redshifts, confirmed 22 previously photometric cluster members, added ten new ones, and identified 17 new spectroscopic multiple images from 6 background sources. The resulting improved mass model used 106 multiple images from 34 families over $v_{\mathrm{H}\alpha} = -8356 \pm 1105$8–$v_{\mathrm{H}\alpha} = -8356 \pm 1105$9, included 308 total mass components for member galaxies and four additional mass profiles, and achieved a multiple-image position rms of $^{-1}$0 (Schuldt et al., 2024).

The most consequential recent development is the shift from point-like constraints to full extended-image modeling of the host galaxy. A 2026 analysis incorporated 77,000 HST pixels from the SN host in addition to 106 point-like multiple images. With the SN host’s extended image included, the statistical uncertainties of all 34 free model parameters were reduced by factors ranging from one to two orders of magnitude compared to the point-like-only model, and the statistical uncertainty in predicted time delays fell below even the small systematic uncertainty assessed by comparing different approaches. The same work presented the delensed host galaxy in multiple HST bands and argued that extended-image models can measure the cluster mass distribution, cosmological parameters, and source properties with unparalleled precision (Schuldt et al., 12 Feb 2026).

The practical legacy of SN Refsdal is therefore twofold. Observationally, it established the feasibility of transient strong-lens prediction and verification. Methodologically, it drove cluster-lens modeling from image-centroid fitting toward hybrid and pixel-level inference. Lens-model products from the revised MACS J1149 analysis, including maps of magnification, arrival-time surfaces, model parameters, and critical curves, were made available through MAST, facilitating reuse by the broader community (Sharon et al., 2014).