Bisous Process: Cosmic Web Filament Detection

• Bisous process is a stochastic marked point process that models cosmic filaments as chains of small cylindrical segments placed on discrete galaxy distributions.
• The method uses MCMC, generating a filament detection probability field, an orientation field, and filament spines to statistically represent uncertainties.
• It emphasizes local density, geometric connectivity, and alignment over smoothed density fields, distinguishing it from deterministic ridge finders.

The Bisous process—also called the Bisous model or Bisous filament finder—is a stochastic, object-based marked point process with interactions designed to detect and quantify filamentary structure in point distributions, most prominently the cosmic web traced by galaxies. In cosmological applications, it represents filaments as random configurations of small cylindrical segments placed directly on the discrete 3D galaxy distribution, rather than on a precomputed continuous density field. Inference by Markov chain Monte Carlo yields a filament detection probability field, an orientation field, and a catalogue of filament spines, so the method provides both a morphological reconstruction of the network and a statistical characterization of uncertainty (Tempel et al., 2016, Tempel et al., 2013).

1. Conceptual basis and methodological lineage

The Bisous process belongs to the general class of marked point processes with interactions. In this setting, the “points” of the process are not the observed galaxies themselves, but the centres of elementary geometric objects; the “marks” encode object attributes such as orientation, length, and radius. In the cosmological version of the model, those objects are small cylindrical segments, and a filament is interpreted as a connected, aligned chain of such cylinders. This formulation was introduced in the methodological work of Stoica, Gregori & Mateu (2005), further developed by Stoica, Martínez & Saar (2007, 2010), and operationalized for galaxy surveys by Tempel and collaborators, as summarized in later arXiv papers (Tempel et al., 2015).

A defining conceptual feature is that the method does not model galaxies directly, but the structure outlined by galaxy positions. The working hypothesis is that the filamentary network can be approximated by a random configuration of cylinders that preferentially occupy locally overdense, elongated regions and interact so as to form coherent chains. Because neighbouring cylinders are required to be similarly oriented and geometrically connected, the network emerges as a filamentary object process rather than as a thresholded density ridge. This distinguishes Bisous from Hessian-, topology-, or graph-based web finders that begin from a smoothed field or an explicit galaxy-to-galaxy connectivity graph (Tempel et al., 2013).

The model is intrinsically stochastic. A single run produces one realization of a plausible filament network; repeated runs sample a distribution over networks. This stochasticity is not an incidental computational detail but a core part of the formalism, because it is precisely what makes possible a posterior-like description in terms of visit probabilities and directional coherence. A common misconception is therefore that Bisous is simply a deterministic skeletonization algorithm expressed in cylinder language; the published applications consistently treat it instead as a probabilistic model whose primary products are fields and ensembles, not only a single skeleton (Libeskind et al., 2015).

2. Statistical formulation

In its standard formulation, the Bisous model is a Gibbs/Markov marked point process on configurations of cylinders. Let $K \subset \mathbb{R}^3$ denote the observation window and $\mathbf{d}=\{d_1,\dots,d_n\}\subset K$ the observed galaxy positions. A cylinder is specified by a centre $k\in K$, a radius $r$, a length $h$, and an orientation unit vector $\omega$, often written as $s(y)=s(k,r,h,\omega)$. A configuration $\mathbf{y}$ is then a finite set of such cylinders. The model density is written in Gibbs form,

$$p(\mathbf{y}\mid\theta)=\alpha \exp\!\left[-U(\mathbf{y}\mid\theta)\right],$$

with energy decomposed into a data term and an interaction term (Tempel et al., 2016).

The data term rewards cylinders that resemble local filament elements in the galaxy distribution. In the detailed SDSS catalogue formulation, each cylinder is evaluated by hypothesis tests for locally high density relative to a surrounding “shadow” region and for locally uniform spread along its axis, together with a concentration measure quantifying how tightly galaxies cluster around the axis. A representative single-cylinder potential is

$$v(y)= \begin{cases} -\sigma^2(y)+c_{\mathrm{hyp}}\log p_{\mathrm{hyp}}(y), & n\ge n_{\min},\ -\infty, & n<n_{\min}, \end{cases}$$

where $\mathbf{d}=\{d_1,\dots,d_n\}\subset K$0 combines the local-uniformity and local-high-density tests, $\mathbf{d}=\{d_1,\dots,d_n\}\subset K$1 measures transverse concentration, and $\mathbf{d}=\{d_1,\dots,d_n\}\subset K$2 imposes a minimum galaxy count. The data energy is additive,

$\mathbf{d}=\{d_1,\dots,d_n\}\subset K$3

so cylinders with strong overdensity contrast, approximate longitudinal uniformity, and small transverse scatter reduce the total energy and are thus favored (Tempel et al., 2013).

The interaction term encodes the fact that filaments are networks rather than isolated segments. In one explicit implementation, cylinders are counted according to whether they are connected through 0, 1, or 2 endpoints, and whether pairs are repulsive: $\mathbf{d}=\{d_1,\dots,d_n\}\subset K$4 Here $\mathbf{d}=\{d_1,\dots,d_n\}\subset K$5 is the number of repulsive cylinder pairs, $\mathbf{d}=\{d_1,\dots,d_n\}\subset K$6 counts cylinders connected through $\mathbf{d}=\{d_1,\dots,d_n\}\subset K$7 endpoints, and the $\mathbf{d}=\{d_1,\dots,d_n\}\subset K$8-parameters tune the relative preference for chain interiors, endpoints, and forbidden overlaps. In the SDSS filament-catalogue construction, repulsive pairs were forbidden by setting $\mathbf{d}=\{d_1,\dots,d_n\}\subset K$9; isolated cylinders were strongly penalized, 1-connected cylinders mildly penalized, and 2-connected cylinders strongly encouraged (Tempel et al., 2013).

Connectivity itself is defined geometrically. Each cylinder has two endpoints and attraction regions around them; two cylinders are considered connected if they attract each other, do not reject each other by excessive centre overlap, and are sufficiently aligned. Published implementations include criteria such as $k\in K$0 for alignment and $k\in K$1 for orthogonality, together with attraction radii comparable to the cylinder radius. In the SDSS catalogue work, the adopted values included $k\in K$2, $k\in K$3, $k\in K$4, $k\in K$5, and $k\in K$6 (Tempel et al., 2013).

3. Inference, visit maps, and filament spines

Inference is performed by Metropolis–Hastings MCMC on the space of cylinder configurations, typically with birth, death, and change moves. Birth adds a new cylinder; death removes one; change perturbs the centre, orientation, or other mark parameters of an existing cylinder. To improve exploration and convergence, Bisous implementations combine these proposals with simulated annealing, and in some descriptions also simulated tempering. A practical feature of the sampler is that many birth proposals are “connected births,” deliberately proposed near existing cylinders and with compatible orientations so as to accelerate the formation of coherent chains (Tempel et al., 2016).

Because the model is stochastic, the relevant summary is the ensemble of sampled configurations rather than any one configuration. From $k\in K$7 sampled cylinder realizations $k\in K$8, the method constructs a visit map or filament detection probability field

$k\in K$9

which estimates how often a spatial point $r$0 is covered by cylinders across the ensemble. A related weighted density map gives higher weight to cylinders with better data potentials, and an orientation field is obtained by averaging cylinder directions at each location. In the 2016 methodological presentation, the local orientation field is encoded by $r$1, with the dominant direction $r$2 defined by the maximizing orientation and the orientation strength $r$3 quantifying directional coherence (Tempel et al., 2016).

Filament spines are then extracted as continuous curves following high-probability ridges that are also directionally coherent. The published spine-extraction procedure uses thresholds on the visit map and orientation strength, starts from local maxima, and propagates in both directions along the local orientation field while enforcing curvature and directional-consistency conditions. Example thresholds given in the methodological literature are $r$4 and $r$5, together with a curvature cutoff $r$6. The output is a filament catalogue in which each filament is represented by a spine: an ordered set of 3D points that approximates the filament axis (Tempel et al., 2016).

This output structure explains why Bisous is often described as simultaneously morphological and statistical. The spine provides the geometry used in downstream analyses—distances to filaments, local tangent directions, lengths, or line-of-sight projections—while the visit map and orientation strength encode confidence and local ambiguity. A misconception arises when only the final spine catalogue is retained and the probability field is ignored; in the original formulation, the spine is a derived object supported by the probabilistic fields, not the sole primary object.

4. Scale choice, survey construction, and practical deployment

The most consequential free choice in Bisous applications is the cylinder radius, which sets the characteristic filament scale being probed. Different studies adopt different fixed radii according to data density and scientific aim. The SDSS filament catalogue of Tempel et al. searched for filaments with radius about $r$7 (Tempel et al., 2013). Later SDSS and 2MRS analyses adopted a characteristic scale of roughly $r$8, explicitly stating that this corresponds to the scale of galaxy groups or clusters and that the detected filaments are the bridges between them (Libeskind et al., 2015). In EAGLE-based work and in reliability studies using mock catalogues, the cylinder width was taken to be about $r$9, and galaxies were counted as belonging to Bisous filaments if they lay within $h$0 of the spine, often with a visit-map threshold of 0.05 (Veena et al., 2019, Muru et al., 2021).

This scale dependence is substantive rather than cosmetic. The method is not intrinsically multiscale; it is tuned to the width encoded by the cylinders. Published papers therefore note that the detected network is the network at the chosen scale, not the full hierarchy of sub-filaments and super-filaments. This suggests that comparisons across Bisous catalogues must control for the adopted cylinder radius before interpreting differences as physical rather than algorithmic (Tempel et al., 2016).

Practical survey deployment also requires explicit handling of redshift-space distortions, selection functions, and survey edges. In SDSS applications, “fingers of God” were compressed at the group level so that groups became approximately spherical, removing artificial line-of-sight filament-like structures before the Bisous run (Tempel et al., 2015). In the SDSS DR10 pair-alignment study, filaments with radius $h$1 were considered reliably detectable only up to $h$2, and the analysis retained only filaments farther than $h$3 from the survey boundary and longer than $h$4 (Tempel et al., 2015). In the 2MRS–Cosmicflows comparison, the same $h$5 Bisous setup was applied within the local Universe out to $h$6, precisely because that is where the independent velocity reconstruction is reliable (Libeskind et al., 2015).

A notable example of problem-specific adaptation is the SDSS galaxy-pair analysis. There, the aim was to measure the angle between the orientation of a galaxy pair and the orientation of its host filament. To avoid a built-in bias in which the filament finder would inherit the pair axis, each galaxy pair was replaced during filament extraction by its 3D centre point. The resulting filament orientations and pair orientations were therefore measured independently by construction (Tempel et al., 2015).

5. Scientific applications and physical interpretation

Bisous filaments have been used as environmental structures for a broad range of cosmological and galaxy-evolution analyses. A central validation came from the comparison between the P-web—filaments extracted by Bisous from halo positions—and the V-web defined from the velocity shear tensor. In a cosmological simulation, the local directions of Bisous filaments were found to be very strongly aligned with $h$7, the eigenvector associated with the smallest eigenvalue of the shear tensor, with about 80% of detected elements aligned within $h$8. In moderate-density regions the spatial overlap between P-web and V-web filaments reached about 90%, supporting the interpretation that Bisous filaments carry more than purely morphological information even though the method does not use the velocity field as input (Tempel et al., 2013).

An observational analogue was provided by the comparison between Bisous filaments in the 2MRS galaxy distribution and a velocity-shear web reconstructed from Cosmicflows-2. The two methods, though described as radically different in ideology and based on independent data sets, were found to be well aligned in the local Universe, with observational agreement weaker than in simulations but still significant. Bisous spine points preferentially occupied V-web filaments and knots, and the alignment of Bisous directions with the shear-based filament direction $h$9 was strongest in voids, where the velocity reconstruction is most linear and reliable (Libeskind et al., 2015).

The method has also been central to direct environmental alignment measurements. In the SDSS pair-orientation study, the orientation of galaxy pairs was found to correlate strongly with the orientation of their host Bisous filaments. For loose pairs, the excess of aligned pairs was at least 25% relative to a random distribution, and the KS-test probability against uniformity was $\omega$0, corresponding to about 6.5$\omega$1. The observed signal agreed well with the Millennium-simulation analogue, where the same Bisous-based filament geometry was used (Tempel et al., 2015).

In galaxy-spin analyses, Bisous provided both filament axes and, indirectly, sheet orientations. In the SDSS study of spiral and elliptical galaxies, spiral galaxy spins were found to align with the host filament, while elliptical minor axes were preferentially perpendicular to hosting filaments. The same work emphasized that Bisous does not explicitly detect sheets; instead, sheet directions were inferred statistically from anisotropy in the filament detection probability field in the plane perpendicular to the filament spine. That point is important because references to “Bisous sheets” in later literature usually mean sheets inferred from the filament probability geometry, not planar objects fitted directly by the model (Tempel et al., 2013).

Beyond orientation studies, Bisous spines have been treated as one-dimensional backbones for point-process analyses along filaments. In the “pearl necklace” study, galaxies and groups projected onto Bisous spines were found not to be uniformly distributed along the filaments. The characteristic regularity scale was around $\omega$2, with a secondary scale around $\omega$3, revealed by both the 1D two-point correlation function and the Rayleigh $\omega$4 statistic (Tempel et al., 2014).

Hydrodynamical simulations extended the Bisous programme from morphology to baryon localization. In EAGLE, Bisous filaments extracted from the galaxy distribution were compared to NEXUS+ filaments derived from the mass distribution. The two methods traced similar dominant structures, with a median mutual misalignment of about $\omega$5 at the positions of galaxies common to both filament sets, and 4277 galaxies lying in filaments according to both methods (Veena et al., 2019). In the missing-baryon analysis, Bisous filaments occupied only about 5% of the full simulation volume but contained 23%–25% of the total baryon budget in diffuse hot intergalactic gas, corresponding to 79%–87% of all the hot WHIM outside halo virial radii. Selecting Bisous filaments with high galaxy luminosity density yielded an “optimal” sample whose missing-baryon mass fraction was about 82%, and the derived radial gas profiles indicated that the hot WHIM is strongly concentrated toward the filament axes (Tuominen et al., 2020).

6. Reliability, limitations, and current extensions

A substantial part of the recent Bisous literature has focused on reliability under sparse sampling and on extensions beyond purely spectroscopic redshift catalogues. In a controlled MultiDark-Galaxies study, the method was applied to subsamples with varying galaxy number density and to 200 repeated runs on identical input. The principal result was that Bisous becomes less complete as the galaxy density decreases, but the filaments that are found are largely persistent. When a galaxy was identified as lying in a filament in a lower-density sample, there was a 97% chance that the same galaxy would also lie in a filament when more complete input data were used, and about 85% of filaments were persistent under that increase in completeness. The Pearson correlation coefficient between visit maps from 200 runs on the same input was 0.98, indicating very low run-to-run stochastic variation at the field level (Muru et al., 2021).

These results clarify a recurrent misconception. Sparse sampling does not primarily cause Bisous to hallucinate large numbers of spurious filaments; rather, it causes the recovered network to become incomplete. In the same study, as the input density decreased, the filament volume-filling fraction dropped from about 7% to nearly 0%, and the fraction of galaxies in filaments fell from about 80% to about 15%. The reliability issue is therefore chiefly one of missed structure, not of catastrophic instability (Muru et al., 2021).

The main structural limitations are consistent across methodological and application papers. Bisous is scale dependent, because the cylinder radius must be specified. It approximates geometry by piecewise cylinders, so very sharply curved or highly branching structures are only represented through locally aligned segments. It is computationally expensive because it relies on large MCMC ensembles, and it requires sufficient tracer density to satisfy the cylinder-level data tests. In observational catalogues, residual redshift-space effects, magnitude limits, survey masks, and edge cuts remain significant practical concerns (Tempel et al., 2016, Libeskind et al., 2015).

An active development path is the incorporation of photometric redshift information. In a proof-of-concept study, the Bisous model itself was left unchanged, but galaxies with photometric redshifts were assigned different line-of-sight positions in different runs by sampling from their uncertainty distributions. Pure photometric samples showed strong line-of-sight alignment bias and a recall that could drop close to 20% as the uncertainty increased, whereas mixed spectroscopic–photometric samples improved the number of detected filaments, mitigated the alignment bias, and kept recall between 40% and 80%. In every tested sample the false discovery rate stayed below 5%. Mixed samples consistently outperformed corresponding photometric-only or spectroscopic-only samples with the same uncertainty scale or spectroscopic fraction (Muru et al., 2023).

Taken together, these studies place the Bisous process in a distinctive methodological position. It is neither a purely geometric ridge finder nor a purely dynamical classifier, but a probabilistic object process on interacting cylinders that has repeatedly been shown to recover structures consistent with velocity-based web definitions, galaxy-alignment signals, and hydrodynamical baryon distributions. Its present limitations are explicit and well documented—fixed scale, dependence on tracer density, MCMC cost, and indirect sheet inference—but so are its strengths: direct use of discrete galaxy positions, explicit encoding of connectivity and alignment, and native probabilistic outputs that make uncertainty part of the representation rather than an afterthought.