Occupying Halos by Their Place in the Cosmic Web's Connectivity

Galaxy surveys are interpreted through a simple empirical rule called the halo occupation distribution, or HOD. Dark matter clumps into "halos," and the HOD is a probability recipe for how many galaxies a halo of a given mass hosts. The recipe works well for galaxy clustering. It fails a known consistency test, though: once tuned to how galaxies cluster, it overpredicts how strongly those galaxies bend background light through gravitational lensing, a problem known as "lensing is low." Letting occupation depend on the local density around each halo closes roughly a third of the gap (AbacusHOD). The rest is unexplained.

This note proposes giving each halo a second label that local density cannot express. The label records how the halo is connected into the filaments, loops and voids of the cosmic web. It is computed with persistent homology, a method from topology that tracks when connected clumps, loops and cavities appear and disappear as one lowers a density threshold. We then test whether occupation conditioned on this label accounts for part of the remaining lensing discrepancy. It must do so at a fixed number of free parameters, and the label must survive checks that it is not an artifact of how the density field was approximated. That last requirement borrows a hard lesson from electronic-structure physics, where approximate calculations routinely produce topological labels that turn out to be false.

Core Claim

At fixed halo mass and fixed local density, a halo's position in the density filtration carries information about how many galaxies it hosts. The filtration is the sequence of excursion sets obtained as the density threshold is lowered. The most useful quantity is expected to be the merger density: the threshold at which the halo's isolated clump first joins a larger connected structure. A halo that merges into the web only at very low density is isolated. One that merges at high density sits on a thick filament or inside a node. Spherical density averages mix these two situations up, because a sphere of radius 5–8 Mpc/h cannot tell a halo on a filament from a halo beside it in a void wall at the same mean density.

The claim is falsifiable in three ways. It fails if the topology-conditioned model adds no clustering or lensing information beyond a multi-scale spherical density baseline with the same number of parameters. It fails if the fitted topology dependence changes sign or size when the smoothing scale or particle subsampling changes. It fails if the gain disappears once environment is described by a flexible non-topological summary.

Why This Combination

The cosmology side supplies the application and the tension. Pranav et al. computed persistent Betti numbers of cosmic density fields: $\beta_0$ counts isolated clumps, $\beta_1$ counts loops and tunnels, and $\beta_2$ counts enclosed voids. They showed these numbers separate cluster-dominated from sponge-like structure. The older Euler characteristic, the alternating sum $\beta_0-\beta_1+\beta_2$, cannot make this separation, because the terms cancel. That work motivates the proposal, but it uses topology only as a global diagnostic of the field. It never assigns a topological quantity to individual halos, and it never feeds such a quantity into $P(N|M)$, the probability that a halo of mass $M$ hosts $N$ galaxies. We do both. AbacusHOD supplies the method. It is a fast, particle-based HOD code on the AbacusSummit simulations that already includes environment-based secondary bias parameters $B_{\text{cent}}$ and $B_{\text{sat}}$ (detected at $>3\sigma$ in BOSS CMASS) and concentration-based assembly bias parameters (about $2\sigma$ for $A_{\text{cent}}$). Our topological label enters the model in exactly the slot $B$ occupies, so the comparison is direct.

The band-topology and density-functional-theory (DFT) directions serve as a methodological constraint rather than a mechanism. DFT is the standard approximate method for computing electron densities and energies in solids. In high-throughput topological materials screening, the crystal's topological label is read from approximate DFT band structures (Catalogue of topological electronic materials). Vidal et al. showed this is dangerous. The standard approximation pushes bands together, which creates spurious "band inversions," the reordering of bands that signals a topological insulator. The standard approximation gives InAs an inversion energy of −0.6 eV, falsely labeling it topological, while a more accurate quasiparticle method gives +0.65 eV. A popular corrected method, the HSE06 hybrid functional, errs the other way on CuTlS₂. The general lesson is that a topological label computed from an approximated field is only trustworthy if it is stable when the approximation is improved. For the cosmic web, the approximations are the density estimator, the smoothing scale, and the particle subsampling. Loosening any of them can create or destroy loops just as band-gap errors create or destroy inversions.

The materials literature also offers a second, structural parallel. Cheap local diagnostics can miss topology that only a global calculation reveals. Symmetry indicators, which read topology from wavefunctions at a few special momentum points, miss the "hidden" topology that Wilson loops expose; Wilson loops track the phase a wavefunction picks up around a closed path in momentum space (Catalogue; thallium nitride). Local spherical density plays the role of the cheap local diagnostic in our setting, and filtration-based connectivity plays the role of the global one. This parallel motivates the design but is not evidence for it.

Research Program

We build a per-halo topological catalog for AbacusSummit boxes. The density field is reconstructed from particle subsamples using the Delaunay tessellation estimator (DTFE) used by Pranav et al. We then run a superlevel-set filtration, which adds regions from the densest downward, and compute three quantities for each halo. The first is its merger density $\rho_{\text{merge}}$, the death threshold of the $\beta_0$ component the halo belongs to. The second is the persistence of the most persistent $\beta_1$ loop passing within the halo's splashback radius. The third is the local $\beta_2$ count within a fixed window. Each descriptor is rank-normalized within narrow mass bins. This removes its trivial correlation with mass, and a second rank-normalization within bins of local density $\delta_R$ does the same for density. The resulting descriptor $\hat\tau$ is precomputed once, like AbacusHOD's environment column, so it does not slow the fitting loop.

Occupation follows the AbacusHOD form. Centrals are a smoothed step in mass and satellites are Poisson with a power-law mean. The topology dependence shifts the characteristic mass for centrals,

$$ \log_{10} M_{\text{cut}}^{\text{eff}} = \log_{10} M_{\text{cut}} + B_{\text{cent}}\,\hat\delta + T_{\text{cent}}\,\hat\tau , $$

where $M_{\text{cut}}$ is the central threshold mass and $\hat\delta$ is the rank-normalized local density already used in AbacusHOD. $T_{\text{cent}}$ is the new coefficient, and an analogous $T_{\text{sat}}$ enters the satellite mass scale. Because $\hat\tau$ is orthogonalized against $\hat\delta$ within mass bins, a nonzero $T$ cannot be absorbed by density.

The robustness protocol follows directly from the Vidal lesson. We recompute $\hat\tau$ under three changes: two particle subsampling rates, DTFE versus Gaussian-smoothed fields at several scales, and two halo finders. We report the fraction of halos whose $\hat\tau$ quartile changes. We also report the posterior on $T$ under each variant. A descriptor whose sign flips under these changes is treated as a false positive and discarded, however well it fits.

Baselines

The naive reference point is the mass-only five-parameter HOD run through the same AbacusHOD pipeline on the same boxes and data vector. Every other model is scored by its improvement over this one in clustering $\chi^2$ and in the ratio of predicted to observed lensing amplitude.

The strongest existing method is AbacusHOD with environment-based secondary bias (AbacusHOD). Its reported result is that $B_{\text{cent}}, B_{\text{sat}}$ are detected at $>3\sigma$ in BOSS CMASS and reduce predicted galaxy–galaxy lensing by 10–15%, about one third of the "lensing is low" discrepancy. We rerun it rather than quote it. It uses the same CMASS projected and redshift-space correlation functions, the same AbacusSummit cosmology and boxes, the same sampler, and the same number of likelihood evaluations as our model. Adding $T_{\text{cent}}, T_{\text{sat}}$ on top of $B$ adds two parameters. The fair comparison therefore also includes AbacusHOD with $B$ plus the concentration-based $A_{\text{cent}}, A_{\text{sat}}$, which is likewise $+2$. Beating this baseline means one of two things, with comparable clustering fit and a penalized information criterion that does not worsen. Either the topological model reduces the predicted lensing by clearly more than the 10–15% the environment model achieves, or $T$ is detected at a significance at least comparable to $B$'s while the concentration parameters stay near the reported $\sim 2\sigma$.

The multi-scale density ablation isolates the key ingredient. It replaces $\hat\tau$ with local density at a second and third radius, for example 2 and 15 Mpc/h, also orthogonalized, with the same parameter count. If this ablation matches the topological model, the gain came from scale information rather than connectivity.

A fourth baseline swaps $\hat\tau$ for a per-halo local Euler characteristic in the same window. It tests whether persistence, rather than any topological count, matters. According to Pranav et al., cancellation among the Betti numbers should weaken it.

Why not just…?

Why not just add more smoothing scales to the environment variable? This is the most serious threat. It is exactly the multi-scale density ablation above, and it would plausibly capture much of the connectivity signal, since filament thickness correlates with density at small radii. The deciding measurement is the conditional one. We take halos matched in mass and in density at all three radii, split them by $\rho_{\text{merge}}$, and check whether their mock occupation and lensing differ. If they do not, the proposal reduces to multi-scale density and we report that.

Why not just use the Euler characteristic or genus, which cosmologists already compute? Pranav et al. show that the Euler characteristic can stay constant while loops and voids grow together, so it loses exactly the distinction we care about. We do not assume this carries over to per-halo descriptors; the Euler baseline measures it directly.

Why not just fit a flexible learned model of occupation on the raw local density field? A convolutional or graph model on a density patch could in principle learn connectivity without topology. Such a model has far more parameters than two, so it cannot be compared at matched complexity, and its dependence on smoothing is harder to audit. We therefore treat it as an upper bound. We fit a small network to predict occupation in a hydrodynamic mock and compare its explained variance with the $\hat\tau$ model. The file lists only THESAN-HR, which covers the reionization epoch, so a low-redshift hydrodynamic catalog would have to be sourced separately; no reported result found on Lacuna for this comparison. If the network captures substantially more variance, topology is an incomplete summary. We would say so and use the network to search for the missing descriptor.

Why not just build a "cosmological density functional theory" that derives galaxy density from matter density, as the research file's second gap suggests? DFT rests on the Hohenberg–Kohn theorem for interacting electrons. The file contains no evidence that an analogous functional exists for collisionless gravitating matter, and building one would be a separate theoretical program with no clear test. We take from DFT only its demonstrated failure mode, false topology from approximate fields, and turn it into a robustness requirement.

Why not just attribute the remaining lensing gap to baryonic feedback or cosmology rather than occupation? That may well be the truth, and this proposal cannot rule it out. What it can do is bound the occupation-driven share. If even the topological model with robust descriptors leaves the gap near two thirds, occupation-based explanations lose ground.

Proposed Experiments

The first experiment is a controlled recovery test that runs before any real data. We populate AbacusSummit halos using a mock in which occupation truly depends on $\rho_{\text{merge}}$ at a chosen strength. We check that the full model recovers $T$ and that the environment-only and multi-scale models show biased $B$ or residual lensing errors. A null mock with no topology dependence checks that we do not detect spurious $T$.

The main experiment fits BOSS CMASS projected and redshift-space clustering. We report the posterior on $T_{\text{cent}}, T_{\text{sat}}$, the predicted lensing amplitude, and the fraction of the discrepancy closed, set against the rerun environment baseline's 10–15% reduction.

The transfer test applies the same descriptors to DESI One-Percent emission-line galaxies (DESI One-Percent ELG HOD). These galaxies sit in halos of mean mass $\log_{10}\langle M_h\rangle \approx 11.9$ at $0.8<z<1.6$. Their satellites needed an ad hoc outskirts component on 0.1–1 Mpc/h scales, and about 0.5% of satellites reside in halo outskirts (companion paper). The question is whether halos with high loop persistence host these outskirt satellites, which would replace an empirical profile fix with an environmental cause.

The decisive checks are three. Topology must beat matched multi-scale density on the conditional split. The sign and magnitude of $T$ must be stable across density estimators, subsampling rates and halo finders. The lensing reduction must exceed what the environment model achieves on the same data and budget. Failing the first check means the idea reduces to density at multiple scales. Failing the second means the label is an approximation artifact, the cosmic-web version of the InAs false positive.