Cosmological simulations follow dark matter as billions of point particles. Small clumps, called subhalos, orbit inside larger clumps, called host halos, and they often vanish from the catalogs that analysts build from these simulations. Nobody can currently say which of three things happened to a vanished subhalo. It may have been genuinely torn apart by tidal forces, which is real physics. It may have been destroyed because too few particles represented it, which is a flaw of the simulation. Or it may still exist and the software that links clumps between saved time steps simply lost track of it while it passed through the dense center of its host, which is a flaw of the tracking code. This proposal separates the three. It borrows a remedy from video object tracking, where an object hidden behind an obstacle is found again by searching the whole frame rather than only near its last known position. It then judges the result with an internal test that needs no external ground truth: in a simulated universe with no preferred scale, every correct statistic must look identical at every epoch once lengths and masses are rescaled.
The central claim is that "subhalo disruption at low particle number" mixes a reversible tracking failure with an irreversible numerical one, and the two can be measured apart. We will add a memory-guided global re-detection step to a standard catalog pipeline and measure subhalo statistics in scale-free simulations. The drop in the particle count at which those statistics stop being scale-invariant tells us how much of the apparent resolution limit comes from the tracker. The limit that remains is the simulation's real one. Three knobs control the diagnosis. Varying how often snapshots are saved should move only tracking losses. Varying force softening should move only numerical losses. Genuine physical disruption must be unaffected by either.
Background and the gap
Scale-free simulations are the evaluation instrument. They model a matter-only expanding universe whose initial density fluctuations follow a pure power law, so the physics has no characteristic length. Their dimensionless statistics must therefore collapse onto one curve when distances are measured in units of the nonlinear scale $R_{\mathrm{NL}}(a)$ and masses in units of $M_{\mathrm{NL}}(a)$, where $a$ is the cosmic expansion factor. Quantifying resolution in cosmological N-body simulations using self-similarity used this collapse on the two-point correlation function to locate resolution limits at sub-percent precision. It showed that when softening is below about 1/30 of the interparticle spacing, particle mass rather than softening sets the limit, and that the limit shifts in comoving units as $a^{-1/2}$, the signature of artificial two-body scattering. Good and proper confirmed this under softening that is fixed in physical rather than comoving units. Accuracy of power spectra in dissipationless cosmological simulations extended the analysis to the power spectrum. These three papers supply our method. All of them study smooth field statistics and never examine discrete objects.
Two later papers moved the test to halos, and they supply the tension this proposal addresses. Convergence of halo statistics: rockstar vs compaso found that ROCKSTAR, a halo finder that uses both positions and velocities, reaches 1% convergence of the halo mass function at about 100–200 particles. COMPASO needs about 1000 particles, and only after a merger-tree cleaning step; without cleaning it never reaches 1%. Self-similar mass accretion history found ROCKSTAR accretion histories converged to 2% down to about 30 particles, while COMPASO shows an unphysical late-time mass drop. Because a post-hoc tracking fix changes the measured convergence, part of what is called "resolution" is really a tracking effect. Neither paper studies subhalos.
On the subhalo side, Xu et al. treated subhalo identification as unreliable enough to discard. They predicted galaxy counts from host properties instead, which cost them small-scale clustering for red and massive galaxies. This motivates the proposal: the subhalo problem is real, and bypassing it has a price. A deep learning approach to halo merger tree construction generates merger trees with a generative adversarial network trained on existing catalogs, so it inherits whatever tracking errors those catalogs contain.
Visual tracking supplies the mechanism. Local trackers search a small window around the last position. When the target is hidden, they latch onto a distractor, and if they update their appearance model with it, the corruption becomes permanent (Re2EMA studies this drift). Three remedies exist. GlobalTrack searches the whole frame with no memory, and on the TLP long-term benchmark it gained 11.1 points over a local tracker. Siam R-CNN keeps short track segments for the target and for distractors, then picks the most consistent path over the whole history with dynamic programming. RTracker stops updating when the target is lost, stores failed candidates as negative examples, and accepts a re-detection only when the candidate resembles past target appearances more than stored distractors.
We transfer the control logic of these trackers, not their neural networks. The logic has four parts: pause updates while the target is lost, search globally, verify a candidate against both the target's history and known distractors, and choose the path over the full history. A host halo during a subhalo's closest approach to its center, called pericenter, is exactly an occluder and a distractor. Particles are labeled, so no learned appearance model is needed.
Research Program
We will build a recovery layer that sits on top of ROCKSTAR and COMPASO catalogs from the ABACUS scale-free runs. The runs cover spectral indices $n=-1.5$ and $-2.0$, sizes of $1024^3$ and $4096^3$ particles, and softenings of 1/15, 1/30, 1/60 and proper.
Losing a subhalo. When a subhalo's branch ends without a clear merger, the layer does not delete it. It marks the subhalo "lost" and freezes its memory. The positive memory holds the IDs of its most-bound particles over its last several snapshots and its recent bound-mass trajectory. The negative memory holds the host's central particles and any neighboring subhalos that captured its particles.
Searching for it. In each later snapshot, over a window of a few host dynamical times, the layer locates the remembered particles wherever they are inside the host, with no local search box. It then runs gravitational unbinding on the particles clustered around them. A candidate bound remnant is accepted only if its overlap with the positive memory exceeds its overlap with the negative memory by a margin. This is RTracker's relative rule, replacing the fixed "50% of most-bound particles" threshold that tree builders use.
Choosing the path. Accepted candidates form short track segments. Following Siam R-CNN, a dynamic-programming pass over each host's full history chooses the most consistent assignment of segments to lost subhalos. This prevents one remnant from being claimed twice and prevents a host fluctuation from being mistaken for a survivor.
Measuring the outcome. We measure four statistics in self-similar variables:
- subhalo mass function in $m_{\mathrm{sub}}/M_{\mathrm{host}}$, at fixed $M_{\mathrm{host}}/M_{\mathrm{NL}}$;
- radial distribution in units of host radius;
- survival probability as a function of time since infall, measured in $\ln a$;
- the rate at which subhalos lose mass.
Time since infall in $\ln a$ is a natural variable because in this universe every host's dynamical time scales with cosmic time. For each statistic we define the resolved particle count $N_{\mathrm{res}}$ as the smallest count at which it stays self-similar within a tolerance, using the sliding-window criterion from the accretion-history paper.
Decomposing the losses. The difference in $N_{\mathrm{res}}$ with and without recovery measures tracking-induced loss. The residual is attributed to the simulation. Each part is then checked against its predicted signature. Tracking loss should grow when snapshots are thinned and should not depend on softening. Numerical loss should depend on softening and particle count, and should follow the two-body-scattering trend of the correlation-function studies. Physical disruption, the part that survives at the highest resolution, must depend on neither knob.
Baselines
Standard catalogs, no recovery. The naive reference is the standard catalogs as produced: ROCKSTAR with its usual tree linking, and COMPASO with and without its cleaning step. These run on identical snapshots and particle sets. The relevant published numbers are all for host halos:
- ROCKSTAR mass function converged to 1% at about 100–200 particles, per Convergence of halo statistics;
- ROCKSTAR accretion history converged to 2% at about 30 particles, per Self-similar mass accretion history.
For subhalo statistics in scale-free simulations there is no reported result on Lacuna, so our first deliverable is measuring this baseline ourselves.
Orphan tracking, the strongest simple competitor. A lost subhalo's most-bound particle is followed forever and counted as a survivor until a fixed merger clock expires. It is the strongest competitor because it recovers every temporary loss by construction and costs almost nothing. We will give it the same particle memory and the same search window. We will tune its merger clock on the $1024^3$ run with the same budget we spend tuning our acceptance margin. Beating it means matching its low-$N$ recovery while not overcounting at high resolution. Concretely, our layer should achieve a lower $N_{\mathrm{res}}$, and in the $4096^3$ run it should flag fewer false survivors, meaning subhalos with no counterpart there.
Ablations of the key ingredients. Four ablations isolate what matters:
- recovery with a local search box around the extrapolated orbit instead of global search;
- an absolute overlap threshold instead of the relative positive-versus-negative rule;
- no negative memory;
- no dynamic-programming pass.
Each is run with the same compute and the same windows.
Why not just…?
Why not just run at higher resolution and compare? Resolution doubling can show apparent agreement when both runs share the same numerical error, which is the central warning of Joyce and Eisenstein. Doubling also cannot tell tracking losses from numerical ones, because both shrink as particle counts rise. We do use the $1024^3$ and $4096^3$ pair, but only to check false recoveries. It is not the convergence criterion.
Why not just use orphan tracking? This is a real threat. If orphans reproduce the self-similar subhalo statistics as well as our layer, the tracking-inspired machinery is unnecessary and we will report that. Our expectation that it falls short rests on one mechanism: an orphan never checks whether anything is still bound to it. Near the resolution limit it should therefore keep numerically destroyed subhalos alive, and that would show up as a violation of self-similarity above the true $N_{\mathrm{res}}$ and as unmatched survivors in the high-resolution run. The orphan-baseline comparison settles this.
Why not just drop subhalos as Xu et al. did? Their own results show a deficit in small-scale clustering for red and massive galaxies and an underprediction for halos hosting many galaxies (Xu et al.). Dropping subhalos also answers nothing about whether the simulation or the tracker was at fault.
Why not just train a neural re-detector like Siam R-CNN? Particle IDs make appearance matching exact, so a learned matcher adds cost without adding information. We keep only the control logic. If the ablations show the relative rule and the dynamic-programming pass add nothing beyond global search, the method collapses to the simpler GlobalTrack-style variant, and we will report that simpler variant.
Proposed Experiments
Decisive check 1: does recovery move the limit? On the $n=-1.5$, $4096^3$ run with 1/30 softening, compare $N_{\mathrm{res}}$ for each subhalo statistic with and without recovery. A substantial downward shift would support the claim that tracking losses are a large share of subhalo "disruption." No shift would show that the simulation's own losses dominate. Both outcomes are publishable.
Decisive check 2: snapshot cadence. Rerun all pipelines on every second and every fourth snapshot. Losses without recovery should grow as cadence thins. Losses with recovery should stay roughly flat. Any piece that still grows with sparser cadence is unrepaired tracking loss.
Decisive check 3: softening. Across the 1/15, 1/30, 1/60 and proper-softening runs, the residual limit after recovery should vary with softening and with $N$ in the way two-body scattering predicts. The recovered fraction should not.
Decisive check 4: false recoveries. Match subhalos between the $1024^3$ and $4096^3$ runs of the same $n$. The fraction of low-resolution recoveries with no high-resolution counterpart is our false-positive rate, compared directly against the orphan baseline.
Decisive check 5: generality. Repeat everything at $n=-2.0$. This tests whether the decomposition depends on the spectral index. Box-size effects are known to dominate at late times for redder spectra (Maleubre et al.), so late epochs at $n=-2.0$ will be excluded using their published range of reliable wavenumbers.
Risk and cost. The main practical risk is that full particle snapshots at fine cadence may not be stored for the largest run. If so, we fall back to the $1024^3$ runs, which weakens only the high-resolution false-positive check. Global search costs one ID lookup and one local unbinding per lost subhalo per snapshot, which avoids the frame-rate cost that burdens full-frame visual re-detection.
References
- Quantifying resolution in cosmological N-body simulations using self-similarity
- Good and proper: self-similarity of N-body simulations with proper force softening
- Accuracy of power spectra in dissipationless cosmological simulations
- Convergence of halo statistics: code comparison between <scp>rockstar</scp> and <scp>compaso</scp> using scale-free simulations
- Self-similar mass accretion history in scale-free simulations
- A FIRST LOOK AT CREATING MOCK CATALOGS WITH MACHINE LEARNING TECHNIQUES
- A deep learning approach to halo merger tree construction
- Re2EMA: Regularized and Reinitialized Exponential Moving Average for Target Model Update in Object Tracking
- GlobalTrack: A Simple and Strong Baseline for Long-Term Tracking
- Siam R-CNN: Visual Tracking by Re-Detection
- RTracker: Recoverable Tracking via PN Tree Structured Memory