If dark matter particles scatter off each other, a galaxy falling through a cluster core should leave part of its dark matter behind. The stars pass through without resistance. The dark halo feels a drag from the cluster's own dark matter and lags. Astronomers measure this lag with gravitational lensing, the bending of background light by foreground mass, and turn it into a limit on the scattering cross-section per unit mass, $\sigma/m$. The measurements are fragile. The same galaxy in Abell 3827 went from a claimed 1.62 kpc offset to 0.54 kpc once new data arrived. The step from offset to cross-section also rests on a guessed infall time and orbit. The field is really asking a worst-case question: could ordinary collisionless gravity, together with every messy astrophysical effect we can bound, have produced this offset? Point estimates and posteriors do not answer that question.
That question is a reachability problem. Control theorists ask the same thing about robots: can bounded disturbances drive the system into a given region of state space? We propose to compute, with guaranteed over-approximation, the set of star–halo offsets that collisionless dynamics can reach under bounded tidal forces, friction, and orbit uncertainty. The result is a certified "offset ceiling." Any observed offset above the ceiling excludes cold dark matter under the stated model. Repeating the computation for each $\sigma/m$ turns an offset measurement into an exclusion interval that stays valid for any unknown orbit or infall time. Standard grid-based methods cannot do this, because their cost grows exponentially with state dimension. The practical content of the work is a model reduction that brings the problem to about 14 dimensions, the range where recent grid-free solvers have been shown to work.
Core Claim
The separation between a galaxy's stars and its dark halo evolves under dynamics that are nearly linear. The two components attract each other with a restoring force that is close to linear for small offsets. Drag and tides enter as forcing terms. This structure fits the conservative linear envelope method of Conservative Linear Envelopes for Nonlinear, High-Dimensional, Hamilton-Jacobi Reachability. That method treats the gap between the nonlinear system and a linear model as an adversarial "error player." It then evaluates the reachable set point by point with the generalized Hopf formula, so it needs no grid. Its authors proved the resulting sets over-approximate the true ones and scaled it to a 15-dimensional pursuit-evasion game.
Our claim is that this machinery, applied to a reduced model of galaxy infall, yields offset ceilings that are both sound (provable over-approximations of what the reduced dynamics can reach) and tight enough to be informative. "Tight enough" here means within a small factor of the largest offset that adversarial trajectory search can actually find. If that holds, the Abell 3827 dispute becomes a clean test: is 0.54 ± 0.23 kpc, or 1.62 ± 0.49 kpc, above what collisionless gravity can reach under bounded confounders?
The reduced model writes the halo-centroid-minus-stellar-centroid displacement $s \in \mathbb{R}^3$ as
$$\ddot{s} = -\Omega^2(r)\, s + a_{\text{drag}}\!\left(\tfrac{\sigma}{m}, \rho(r), v\right) + d(t), \qquad \|d(t)\| \le \delta(r),$$
Here $\Omega(r)$ is the internal restoring frequency of the star–halo pair. It depends on the galaxy's cluster-centric radius $r$ through tidal truncation. $a_{\text{drag}}$ is the drag acceleration, which scales with cluster dark matter density $\rho$, speed squared $v^2$, and $\sigma/m$. $d$ collects everything else we can bound but not model: differential tidal acceleration, differential dynamical friction, pull from neighbouring core galaxies, and centroid shifts from asymmetric stripping, all within a radius-dependent budget $\delta(r)$.
The galaxy's own orbit (position and velocity in the cluster potential, 6 dimensions) drives this 6-dimensional offset subsystem. Adding $\sigma/m$ and infall time as parameters that do not evolve, as in Parameter-Conditioned Reachable Sets for Updating Safety Assurances Online, gives a 14-dimensional problem. The folk estimate behind Massey et al.'s cross-section is the steady state $s^\ast \approx a_{\text{drag}}/\Omega^2$. The reachable set replaces this single number with the full envelope over transients, pericentre passages, and adversarial confounders.
Research Program
The work has three parts.
The first part builds the reduced model and calibrates the disturbance budget. $\Omega(r)$ and $\delta(r)$ come from restricted N-body integrations that we run ourselves. In these runs a live galaxy of stars and halo falls through a static analytic cluster potential with collisionless dark matter. We measure how far the centroids actually separate through tidal stripping and friction alone, and set $\delta(r)$ to cover the largest deviation from the linear model with a stated margin. This calibration is the weakest link, and we treat it as such. The certificate is sound only relative to the reduced model and its budget. We report how the ceiling moves as $\delta$ is inflated.
The second part computes the reachable sets. The initial set covers all bound infall orbits with cluster-centric velocities and radii consistent with the observed line-of-sight velocity and projected position. Infall time ranges from roughly $10^8$ to a few $10^9$ years. The object of interest is the projection of the reachable $s$ onto the sky plane at the present time.
The primary solver is the Hopf-based envelope. The error-player bound absorbs the nonlinearity in the orbit-dependent coefficients $\Omega(r)$ and $\rho(r)$. We use a second route to get a whole $\sigma/m$ family in one solve: a DeepReach network (a neural network with sinusoidal activations trained to satisfy the Hamilton-Jacobi equation) conditioned on $\sigma/m$ as a virtual state, following Parameter-Conditioned Reachable Sets. This route is not sound by itself. Its role is to interpolate between Hopf evaluations and to show where the linear envelope is loose.
The third part is inference. Lensing measurement error enters as a target tube around the observed offset. We cannot use distribution-free calibration of that error in the style of Conformal Reachability for Safe Control in Unknown Environments, because it needs a stream of exchangeable residuals that a single cluster does not provide. So we use the published 68% intervals and state their Gaussian assumption openly. A value of $\sigma/m$ is excluded if the tube does not intersect the reachable offset set at that $\sigma/m$ for any admissible orbit and infall time.
Prior work shapes each part in a specific way. The Abell 3827 papers, The behaviour of dark matter associated with four bright cluster galaxies in the 10 kpc core of Abell 3827 and Dark matter dynamics in Abell 3827: new data consistent with standard cold dark matter, motivate the work. They show an offset claim collapsing under better data and a cross-section inference that hinges on an unknown infall time. The first paper explicitly calls for dynamical modelling to separate self-interaction from astrophysical effects. Mass-galaxy offsets in Abell 3827, 2218 and 1689: intrinsic properties or line-of-sight substructures? supplies the multi-cluster test set. It reports intrinsic kpc-scale offsets in Abell 2218, none in the centre of Abell 1689, and uncertainties from ensembles that the authors note can be underestimated by a factor of two. That last caveat is why we inflate lensing error bars in a sensitivity run.
Two papers constrain the design. Learning Hybrid Control Barrier Functions from Data shows that learned safety certificates need sample density exponential in dimension, so we do not make a learned certificate the source of soundness. The linear-envelope paper itself warns that strongly nonlinear dynamics can make envelopes empty or vacuous. This is why the reduction isolates the nearly linear offset subsystem instead of wrapping full multi-galaxy orbital mechanics. We deliberately do not attempt the 24-dimensional coupled four-galaxy core. The other core galaxies enter only through $\delta$.
Baselines
The naive reference point is the steady-state estimate $s^\ast = a_{\text{drag}}/\Omega^2$ evaluated at nominal orbit parameters. It is the reasoning behind the cross-section in Massey et al. 2015, which reported $\sigma/m \sim (1.7 \pm 0.7)\times 10^{-4}\,(t_{\text{infall}}/10^9\,\text{yr})^{-2}$ cm²/g from a 1.62 ± 0.49 kpc offset. We compute it with the same $\Omega$ and $\rho$ calibration as our method, so any difference comes from handling transients and confounders, not from inputs. Our method beats this baseline if the ceiling at $\sigma/m = 0$ is significantly non-zero: collisionless dynamics can then reach offsets that the steady-state formula would misread as drag. It also beats the baseline if our exclusion intervals differ from the formula's by more than the formula's quoted uncertainty.
The strongest competitor is adversarial Monte Carlo search. We sample orbits, infall times, and disturbance time series inside the same bounds, integrate the reduced model, and add gradient-based optimization of the disturbance sequence to push the offset outward. It gets the same model, the same bounds, and the same GPU-hour budget as the Hopf solve. This baseline gives a lower bound on the true ceiling. Our method gives an upper bound. We beat it in usefulness if the certified ceiling is within a factor of about 1.5 of the best adversarial offset. If the gap is larger, the certificate is sound but too loose to matter. No reported result was found on Lacuna for offset ceilings from any method, since the problem has not been posed. The relevant existing numbers are the observed offsets themselves.
The observational baseline is the published free-form and parametric lensing results taken at face value. For Abell 3827, the 2018 LENSTOOL fit reached a reduced $\chi^2$ of 31.7 with 29 degrees of freedom and offset 0.54 ± 0.23 kpc (Massey et al. 2018). For Abell 1689, the free-form reconstruction in An excursion into the core of the cluster lens Abell 1689 gives $\sigma/m \lesssim 1$ cm²/g from the survival of clump M3. We do not try to beat these as lens models. We test whether our dynamical exclusion is consistent with them and whether it sharpens them.
Two ablations isolate the key ingredients. The first sets $\delta = 0$, which removes the confounder budget and shows how much of the ceiling comes from astrophysical confounders. The second replaces Hopf evaluation with the DeepReach network alone, which shows what formal soundness costs in tightness and runtime. On reachability benchmarks the file reports, for context, SEEV exact verification of an 8-dimensional system in about 22 seconds against 3-hour baseline timeouts (SEEV), and about 75 minutes of offline training for the 4-dimensional Air3D game (Parameter-Conditioned Reachable Sets). No runtime is reported for the 15-dimensional Hopf benchmark.
Why not just…?
Why not just run a grid of SIDM N-body simulations and match offsets? Forward simulation answers "what offset does this orbit produce," not "what is the largest offset any admissible orbit and confounder history can produce." Random sampling of a roughly 14-dimensional space systematically underestimates the maximum. For exclusion that is the dangerous direction, because it rules out cold dark matter too eagerly. We still run simulations, but for a different purpose: to calibrate $\delta$ and to check that no simulated trajectory escapes the certified envelope. This alternative becomes a real threat if the adversarial Monte Carlo baseline turns out to match the certified ceiling closely. In that case sampling is good enough in practice, and the certificate adds only rigor. We would report that outcome plainly.
Why not just use the steady-state formula with generous error bars? It is the obvious sanity check, and in quiet, non-merging systems it may be adequate. It ignores the transient overshoot of the star–halo pair after a pericentre passage, when $\Omega$ drops and tidal forcing peaks. That transient is exactly the regime of a merging core like Abell 3827. The $\delta = 0$ ablation, compared with the full ceiling on orbits reaching pericentre within the last few hundred million years, measures directly how much the formula misses. If the difference is negligible, the formula is vindicated and our contribution shrinks to a rigorous justification of it.
Why not just free the halo centres in LENSTOOL, or use simulation-based posterior inference? These improve the measurement of the offset, not its interpretation. The 2018 Abell 3827 analysis still left degeneracies among galaxies N2–N4 even with ALMA and MUSE data (Massey et al. 2018). Posterior methods such as TMNRE, a neural method that learns likelihood ratios from simulations, give probabilities under a prior on orbits and confounders (Brehmer et al.; Anau Montel et al.). When that prior is poorly known, as the infall-time degeneracy shows it is, a prior-free worst-case statement is the complementary answer. The two are not rivals. Our method consumes whatever offset tube the best lens model delivers.
Why not just use grid-based level sets or Taylor-model zonotopes? A grid at 100 points per axis in 14 dimensions is $10^{28}$ cells, and exponential grid cost is documented as limiting exact methods to about 4–5 dimensions (Tomlin et al.). On the 15-dimensional benchmark, Taylor-model zonotopes lost to the Hopf envelope because their over-approximation compounds over long horizons (Conservative Linear Envelopes). Our horizon covers several orbital times, so we expect the same failure. We include a zonotope run on a 6-dimensional fixed-orbit slice to confirm this.
Proposed Experiments
The first experiment is a soundness and tightness study on synthetic clusters. We use the restricted N-body setup with known $\sigma/m$, including zero, and hold out orbits that were not used to calibrate $\delta$. Two quantities are reported. The first is the fraction of held-out trajectories whose projected offset exceeds the certified ceiling, which should be zero up to the stated calibration margin. The second is the ratio of the certified ceiling to the adversarial Monte Carlo maximum, as a function of $\delta$ and time since pericentre.
The decisive check has a clear failure condition. If the ratio exceeds about 2 over most of the orbit space, the linear envelope is too conservative for this physics, and the project pivots. The fallback is to present the adversarial lower bound with DeepReach as an approximate map, and to drop the claim of certified exclusion.
The second experiment applies the method to data. We compute cold dark matter offset ceilings and $\sigma/m$ exclusion intervals for galaxy N1 of Abell 3827 under both the 2015 and 2018 offset measurements, for the central galaxies of Abell 2218 using the reported few-kpc offsets, and for the aligned centre of Abell 1689. The result is interesting if the 1.62 kpc measurement sits above the collisionless ceiling while 0.54 kpc sits below it. That would say the dispute between the two papers is exactly a dispute about whether self-interaction is required. It is equally interesting if both sit below it: then even the original claim never required self-interaction once confounders are bounded. A sensitivity run doubles the lensing uncertainties, following the underestimation caveat in Mohammed et al., and reports whether any conclusion flips.
The third experiment checks cross-cluster consistency. Every cluster must admit a common $\sigma/m$. We intersect the per-cluster intervals and check them against the $\lesssim 1$ cm²/g bound from Abell 1689's surviving clump. An empty intersection would flag either a mis-set disturbance budget or a genuine tension. Either would be a useful finding.
References
- Conservative Linear Envelopes for Nonlinear, High-Dimensional, Hamilton-Jacobi Reachability
- Parameter-Conditioned Reachable Sets for Updating Safety Assurances Online
- Conformal Reachability for Safe Control in Unknown Environments
- The behaviour of dark matter associated with four bright cluster galaxies in the 10 kpc core of Abell 3827
- Dark matter dynamics in Abell 3827: new data consistent with standard cold dark matter
- Mass-galaxy offsets in Abell 3827, 2218 and 1689: intrinsic properties or line-of-sight substructures?
- Learning Hybrid Control Barrier Functions from Data
- An excursion into the core of the cluster lens Abell 1689
- SEEV: Synthesis with Efficient Exact Verification for ReLU Neural Barrier Functions
- Mining for Dark Matter Substructure: Inferring Subhalo Population Properties from Strong Lenses with Machine Learning
- The effect of the perturber population on subhalo measurements in strong gravitational lenses
- Reducing Conservativeness in Safety Guarantees by Learning Disturbances Online: Iterated Guaranteed Safe Online Learning