Is That Dark Matter Clump Real? Testing Free-Form Lens Substructure Against the Choice of Mass Basis

When a galaxy cluster bends light from galaxies behind it, astronomers see the same background galaxy several times. Free-form lens models reconstruct the cluster's mass from those repeated images without assuming that mass sits where the light is. Their best-known products are "light-unaffiliated" dark matter clumps: concentrations of mass with no galaxy on top. These clumps matter because they test dark matter physics. They are also disputed. In the cluster MACS J0416, a free-form model finds a clump of about $5.7\times10^{11}\,M_\odot$ (called M2). A parametric model, which builds the cluster from a few smooth analytic profiles, fits newer JWST data without it and drives its mass to zero. Neither side can currently tell whether the clump is required by the data or created by its own modeling choices.

Image registration, the problem of computing the spatial warp that aligns two images, has already faced this question in another form. When most of an image is featureless, the warp in the empty regions is set entirely by how the method spreads information outward from a few landmarks, not by the data. We propose to treat cluster lensing the same way. Multiple images are sparse landmarks that cover a few percent of the field. Mass between them is filled in by the basis, meaning the family of building-block mass shapes the model is allowed to sum. We will build one differentiable free-form solver with interchangeable bases, plus a map of where the data actually constrain the mass. We will then ask a sharp question: does a candidate clump survive when the basis is swapped but the fit to the images stays equally good? The method is calibrated on simulated clusters with planted clumps and then applied to M2.

Core Claim

A light-unaffiliated clump should count as a data-driven detection only if it meets two conditions. It must persist across mass bases that are structurally different but fit the images equally well. It must also sit in a region where the multiple images carry independent information about the mass. Current free-form practice estimates uncertainty by rerunning one basis family with different random seeds. The Abell 2744 robustness study (Free-form Grale reconstruction of Abell 2744) examines how stable results are when the input image catalog changes. Reruns of that kind cannot reveal an artifact that the basis itself induces, because every rerun shares the same basis. Our new contribution is a test of how much each feature depends on the basis, together with a data-support map, both calibrated on simulated clusters with known truth.

Where the Two Directions Meet

On the lensing side, the main free-form code is GRALE (Liesenborgs et al., Liesenborgs et al.). It represents mass as many Plummer spheres, simple round blobs of mass, placed on a grid that subdivides where lensing information is dense (De Rijcke et al.). A genetic algorithm then searches for blob weights such that the multiple images, traced back to the source, land on top of one another. This code produced the MACS J0416 model with lens-plane root-mean-square position error (RMS) of 0.191″ and clumps M1 and M2 (BUFFALO wild wings). It also produced the mass-to-light offsets reported in Abell 370 (Further support for a trio of mass-to-light deviations in Abell 370). These papers motivate the proposal: their claims are exactly what we want to test.

The JWST reanalysis (Testing light-unaffiliated mass clumps in MACS 0416) constrains the proposal. It shows that a parametric LENSTOOL model reaches 0.57″ RMS on 303 images without cluster-scale clumps, compared with 0.53″ for a model that includes them. When M2 is added to LENSTOOL, its mass goes to zero. In the M2 region, the parametric and free-form surface densities differ by 5–6%. Parametric fits tie mass to light, so they cannot settle a light-unaffiliated question on their own; Testing light-traces-mass in MACS-J0416 explains why this reasoning is circular. Free-form fits, for their part, have many local degrees of freedom that could invent structure. Exact degeneracies sharpen the worry. The mass-sheet transformation adds a uniform sheet of mass and rescales the rest, and other transformations of the projected potential change it far from the images (Extrapolating the projected potential of gravitational lens models). Under these transformations, different mass maps reproduce identical images, so any feature that lies along a degenerate direction is not measured at all.

On the registration side, Gaussian Primitive Optimization supplies the method we reuse (Gaussian Primitive Optimized Deformable Retinal Image Registration). In retinal images, vessels cover less than 15% of the area, so gradient-based warps get almost no signal from the rest; the paper calls this gradient dilution. The method anchors Gaussian blobs, each with a position, a displacement, and an adjustable radius, at vessel junctions. It then spreads their displacements into the empty background by blending each point's nearest anchors. This reaches 2.35 px target registration error on the FIRE retinal benchmark, against 2.77 px for RetinaRegNet. We borrow two things from it. The first is a second basis family that is genuinely different: anchored at the data, with adaptive radii and nearest-neighbor blending, unlike a refined grid. The second is the lesson that in data-poor regions the interpolator decides the answer.

The disentanglement work in registration (Bayesian Unsupervised Disentanglement of Anatomy and Geometry, LADDA) and the gradient-based multimodal similarity measure of MMRegNet also map onto lensing. We considered building on them and set them aside. Their natural lensing use is to compare a source's appearance across wavelength bands. A simpler alternative, reconstructing a separate source for each band, likely does that job as well, and it does not address the clump dispute.

Research Program

We will write a differentiable lens solver (in JAX) whose convergence $\kappa$ (the surface mass density in units of the critical density) is a sum of non-negative primitives drawn from a chosen basis $\mathcal{B}$. Each primitive has a closed-form deflection. All bases are fitted with the same objective and the same gradient-based optimizer, so they differ only in the basis. The objective has three terms. The first is GRALE's source-coincidence criterion, the spread of back-traced images of each source in the source plane, weighted by local magnification so that it approximates image-plane error. The second penalizes predicted extra images where none are observed. The third is a weak positivity and smoothness term that is the same for every basis.

We will compare three bases. The first follows GRALE: Plummer spheres on an adaptively refined grid. The second follows the retinal method: Gaussians initialized at observed image positions and at the midpoints between images, with free radii, plus a few broad cluster-scale components. The third is a pixel grid with a penalty on large second derivatives. Running GRALE itself alongside the first basis lets us confirm that our solver reproduces published behavior.

The data-support map measures how strongly the fit responds to a small local change in mass at each position $\theta$. Changes that the known degeneracies leave invisible are removed first:

$$S(\theta) = \big\| P_{\perp}\, \partial r / \partial \kappa(\theta) \big\|^2 .$$

Here $r$ is the vector of position residuals divided by their measurement errors. $\partial r/\partial\kappa(\theta)$ is the change in those residuals when a compact unit of mass is added at $\theta$. $P_\perp$ projects out the mass-sheet direction and the potential transformations that preserve the observables. A clump earns the label data-required under two conditions. Deleting it and re-optimizing every other component must worsen the fit by more than the noise allows. That must hold for all three bases. A clump that only one basis needs, at matched fit quality, is classified as basis-induced.

Baselines

The naive reference point is a smooth model: our solver restricted to broad cluster-scale components with no small-scale freedom. It shows how much of the image fit any substructure can buy. For MACS J0416, this smooth model gives the RMS that a substructure-free free-form solution reaches on the same catalog, the free-form counterpart of the LENSTOOL test.

The strongest existing free-form competitor is GRALE itself, run with its published configuration and an ensemble of seeds. On the 237-image catalog it reports $\Delta_{\rm RMS}=0.191''$ (BUFFALO wild wings). The strongest parametric competitor is LENSTOOL, which reports 0.57″ RMS on 303 JWST images without cluster-scale clumps and 0.53″ with them (Testing light-unaffiliated mass clumps in MACS 0416).

For fairness, every method receives the same image catalogs and positional errors. Each new basis gets an optimization budget matched to GRALE's ensemble in wall-clock time and number of restarts. Each basis also gets the same number of hyperparameter trials, selected on the simulated clusters rather than on MACS J0416. Our goal is not a lower RMS than 0.191″. On MACS J0416 we count it a success if the alternative bases fit within GRALE's RMS plus one measurement-noise unit, because only then is the basis swap a valid test.

Two ablations isolate the key ingredients. One keeps the solver fixed and swaps only the basis, comparing the grid with the image-anchored primitives; this isolates the effect of the basis. The other removes $P_\perp$ from the support map; this tests whether accounting for degeneracies changes which clumps are judged supported.

On the simulated clusters, the comparison project reports that the best methods recover radial profiles within 10% and magnification to about 10% at $\mu=3$, degrading to about 30% at $\mu=10$ (The Frontier Fields lens modelling comparison project). We found no reported result on Lacuna for clump false-positive rates on these mocks. Our test therefore beats existing practice if it gives a lower false-detection rate than seed-ensemble uncertainty at an equal detection rate for planted clumps.

Why not just…?

Why not just rerun GRALE with many seeds, as is already done? Seed ensembles measure how much the optimizer wanders within one basis family. A structure favored by Plummer spheres on a refined grid can appear in every run and look highly significant. The planted-clump experiment tests this directly. If seed ensembles already achieve the same false-positive rate as our cross-basis test, the basis swap adds nothing and we will say so.

Why not just settle it with parametric model comparison, as the JWST paper did? LENSTOOL's null result for M2 is informative. However, a parametric model ties its small-scale mass to galaxies, so it can absorb a light-unaffiliated clump into nearby galaxy halos or the smooth halo (Testing light-traces-mass in MACS-J0416). The comparison project also found that the parametric advantage narrows on the irregular simulated cluster Hera. We will run the same planted-clump test on LENSTOOL-style fits to measure how often they miss clumps that are really present.

Why not just use a learned solver such as the recurrent inference machine? The recurrent-inference-machine paper reconstructs pixelated mass with a network trained on 400,000 simulated galaxy-scale lenses. It reports a systematics floor: at high signal-to-noise, the learned prior limits fidelity regardless of data quality. A learned prior is just another basis whose behavior in empty regions is fixed by training, which makes it a candidate fourth basis for our test rather than a replacement for it.

Why not just wait for more images? Scaling to 1,000 images (Free-form grale lens inversion of galaxy clusters with up to 1000 multiple images) will shrink the empty regions, but it will not remove them. The support map predicts which proposed clumps more images would actually resolve, which is useful for planning observations.

Proposed Experiments

First, as a correctness check on the reused method, we will reimplement the anchored-Gaussian blending and confirm that it comes close to the reported 2.35 px on FIRE. We will not tune it further.

Second, we will use the Ares and Hera simulated clusters from the comparison project. We will plant clumps of $10^{11}$–$10^{12}\,M_\odot$, with and without accompanying light, at varying distances from the nearest image. Then we fit each basis and GRALE to the simulated image catalogs and measure three things: detection rate against planted mass and distance, false-positive clumps per field, and whether our support map $S(\theta)$ predicts where false positives occur.

Third, we will apply the calibrated test to MACS J0416, using both the 237-image catalog and the 303-image JWST catalog. We will check whether M2 is data-required, basis-induced, or located in a low-support region. The same test will be run on the three Abell 370 offsets.

The decisive check is falsifiable in both directions. If M2 persists across all bases at matched RMS and sits at high support, while the same procedure rejects spurious clumps in Hera, the free-form detection gains a calibrated standing that LENSTOOL's null result cannot override. If M2 disappears under the anchored basis with no loss of fit, it was produced by the basis rather than required by the data.

References

  1. Free-form Grale reconstruction of Abell 2744: robustness of uncertainties against changes in lensing data
  2. A genetic algorithm for the non-parametric inversion of strong lensing systems
  3. The role of multiple images and model priors in measuringH0 from supernova Refsdal in galaxy cluster MACS J1149.5+2223
  4. Non-parametric strong lens inversion of Cl 0024+1654: illustrating the monopole degeneracy
  5. BUFFALO wild wings: a high-precision free-form lens model of MACSJ0416 with constraints on dark matter from substructure and highly magnified arcs
  6. Further support for a trio of mass-to-light deviations in Abell 370: free-form <scp>grale</scp> lens inversion using BUFFALO strong lensing data
  7. Testing light-unaffiliated mass clumps in MACS 0416 on galaxy and galaxy-cluster scales using the JWST
  8. Testing light-traces-mass in Hubble Frontier Fields Cluster MACS-J0416.1-2403
  9. Extrapolating the projected potential of gravitational lens models: property-preserving degeneracies
  10. Gaussian Primitive Optimized Deformable Retinal Image Registration
  11. Bayesian Unsupervised Disentanglement of Anatomy and Geometry for Deep Groupwise Image Registration
  12. LADDA: Latent Diffusion-based Domain-adaptive Feature Disentangling for Unsupervised Multi-modal Medical Image Registration
  13. Unsupervised Multi-modality Registration Network Based on Spatially Encoded Gradient Information
  14. The Frontier Fields lens modelling comparison project
  15. Pixelated Reconstruction of Foreground Density and Background Surface Brightness in Gravitational Lensing Systems Using Recurrent Inference Machines
  16. Free-form grale lens inversion of galaxy clusters with up to 1000 multiple images