CF3: Multidisciplinary Research Insights

• CF3 is a multifaceted term encompassing extragalactic datasets, 3D feature compression in computer vision, and counterfactual learning for financial risk detection.
• In astronomy, CF3 underpins precise reconstructions of galaxy flows and void dynamics using Bayesian inference and MCMC–Gibbs techniques.
• In computer vision and spectroscopy, CF3 enables efficient 3D feature field sparsification and serves as a spectroscopic probe to map local electrostatics.

CF3 denotes several distinct concepts in contemporary research usage. In extragalactic astronomy it most prominently refers to Cosmicflows-3, a galaxy-distance and peculiar-velocity compilation used to reconstruct the nearby density and velocity fields and to map the Local Void (Tully et al., 2019). In computer vision it names “Compact and Fast 3D Feature Fields,” a top-down pipeline for compressing and sparsifying 3D Gaussian feature representations (Lee et al., 7 Aug 2025). In graph machine learning it denotes a counterfactual-and-factual explanation framework for company financial risk detection (Du et al., 9 Mar 2025). The same label also appears as a local identifier for a cold front in Abell 2256 (Ge et al., 2020), as the Snowmass-2013 Cosmic Frontier 3 working group on non-WIMP dark matter (Kusenko et al., 2013), and as the aliphatic –CF$_3$ chemical group in vibrational Stark spectroscopy (Cruz et al., 2024).

1. Cosmicflows-3 as a cosmographic dataset and reconstruction framework

Cosmicflows-3 (CF3) is described as a uniquely rich compilation of galaxy distances and inferred peculiar velocities, comprising approximately $18\,000$ galaxy distances and reaching observed heliocentric redshifts up to $cz\approx 15\,000\ {\rm km\ s^{-1}}$ ($0.05\,c$), with essentially all-sky coverage apart from the Milky-Way zone of avoidance. Its distances are assembled from heterogeneous indicators: the Tully–Fisher relation for spirals, the Fundamental Plane for ellipticals, Cepheid period–luminosity and tip of the red-giant-branch calibrations, and Type Ia supernovae. Cross-calibration uses overlapping samples to establish a consistent zero-point; typical relative distance errors are $\sigma_d/d\simeq 0.15$, and velocity measurement errors are $\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$. A line-of-sight peculiar velocity is obtained to first order from $v_{\rm pec}=cz_{\rm obs}-H_0d$, with a best-fit Hubble constant in the Bayesian/Wiener-filter framework of approximately $75\ {\rm km\ s^{-1}\ Mpc^{-1}}$ (Tully et al., 2019).

The reconstruction assumes $\Lambda$CDM with $\Omega_m=0.3$ and $18\,000$0, linking the peculiar-velocity field $18\,000$1 and density contrast $18\,000$2 in linear theory through $18\,000$3, where $18\,000$4. The Wiener filter is used as the optimal Bayesian linear estimator, combining noisy CF3 constraints with a fiducial CDM power spectrum to produce constrained realizations of $18\,000$5 on a grid with resolution $18\,000$6. An independent but compatible MCMC–Gibbs sampler described by Graziani et al. (2019) alternately samples $18\,000$7, a non-linear dispersion $18\,000$8, the density field, and updated distances. Even behind the Galactic plane, the coherence of the large-scale flow allows Wiener-filter interpolation to remain robust within approximately $18\,000$9.

Within these reconstructions, the Local Void emerges as a complex cavity bounded principally by the Perseus–Pisces filament, the Norma–Pavo–Indus complex, the “Arch,” and the Centaurus–Puppis sheet. Its deepest trough is Lacerta–2.4 with $cz\approx 15\,000\ {\rm km\ s^{-1}}$0, and two other prominent minima are Andromeda–2.3 and Aquila–0.8. At the $cz\approx 15\,000\ {\rm km\ s^{-1}}$1 contour, the Local Void spans approximately $cz\approx 15\,000\ {\rm km\ s^{-1}}$2, implying $cz\approx 15\,000\ {\rm km\ s^{-1}}$3. Beyond it lie the even larger Hercules and Sculptor voids; at $cz\approx 15\,000\ {\rm km\ s^{-1}}$4, these depressions interconnect into a continuous “hypervoid” network. The abstract further notes that the Perseus–Pisces and Norma–Pavo–Indus filaments are separated by $cz\approx 15\,000\ {\rm km\ s^{-1}}$5, that minor filaments course through voids, and that one such chain connects the Virgo and Perseus clusters through the depths of the Local Void.

CF3 is also used to quantify the Local Void’s dynamical effect on the Local Group. A simple spherical-void approximation, confirmed by numerical-action orbits, gives a void “push” of approximately $cz\approx 15\,000\ {\rm km\ s^{-1}}$6–$cz\approx 15\,000\ {\rm km\ s^{-1}}$7, while Virgo’s overdensity produces a comparable “pull.” A Wiener-filter decomposition of the Local Group’s CMB-dipole motion, $cz\approx 15\,000\ {\rm km\ s^{-1}}$8, yields $cz\approx 15\,000\ {\rm km\ s^{-1}}$9, $0.05\,c$0, and a combined $0.05\,c$1, with the remainder attributed to mid- and far-field attractors including Perseus–Pisces, Shapley, and the Dipole Repeller. The same study characterizes the combined perturbations from Local Void repulsion and Virgo attraction as accounting for about $0.05\,c$2 of the Local Group motion in the CMB rest frame.

2. CF3 in local-kinematic inference and bias correction

CF3 also underpins quantitative studies of low-order velocity moments. Qin et al. combine CF3 with the 2MASS Tully-Fisher survey (2MTF) and work in log-distance space with $0.05\,c$3. Galaxies common to the two catalogues yield a zero-point difference of $0.05\,c$4 dex, after which the merged sample contains $0.05\,c$5 galaxies. Using the $0.05\,c$6MLE maximum-likelihood estimator, the study measures at a characteristic depth of $0.05\,c$7 a bulk flow of $0.05\,c$8 toward Galactic coordinates $0.05\,c$9, together with an average shear amplitude of $\sigma_d/d\simeq 0.15$0. Variable-window analyses over $\sigma_d/d\simeq 0.15$1–$\sigma_d/d\simeq 0.15$2 remain consistent with $\sigma_d/d\simeq 0.15$3CDM predictions (Qin et al., 2018).

A separate methodological issue is the lognormal bias induced when Gaussian errors in distance modulus are transformed into distances and peculiar velocities. Hoffman et al. analyze this effect for data designed to statistically follow the grouped CF3 sample and show that naive treatment produces spurious nearby outflow and strong infall at larger distances. Their Bias Gaussianization correction (BGc) exploits the invariance of medians under the lognormal transformation to Gaussianize distances and velocities. Tested on mock CF3 catalogues, BGc eliminates the lognormal bias, yields Wiener-filter reconstructions that recover the underlying density and velocity fields without the spurious outflow/inflow seen in raw data, and gives a residual in the recovered $\sigma_d/d\simeq 0.15$4 of $\sigma_d/d\simeq 0.15$5 relative to simulation. Applied to actual CF3 data, the correction gives $\sigma_d/d\simeq 0.15$6 (Hoffman et al., 2021).

Taken together, these works establish CF3 not only as a catalogue but as an inference substrate for Wiener-filter reconstructions, bulk-flow and shear estimation, and bias-controlled peculiar-velocity analysis. This suggests that the scientific role of CF3 extends beyond data aggregation into the calibration of local-universe kinematics.

3. CF3 as “Compact and Fast 3D Feature Fields”

In computer vision, CF3 refers to “Compact and Fast 3D Feature Fields,” a method built on 3D Gaussian Splatting (3DGS). The scene is represented as a set of anisotropic Gaussians $\sigma_d/d\simeq 0.15$7, where each Gaussian $\sigma_d/d\simeq 0.15$8 has a center, covariance, opacity, and feature vector. The stated goal is to lift high-dimensional 2D features from multiple views into a 3D Gaussian field and then compress and sparsify that field so that the rendered 3D feature field faithfully reproduces multi-view 2D features while drastically reducing the total number of Gaussians. Rather than direct joint optimization, CF3 uses a top-down pipeline with three stages: fast weighted fusion of multi-view 2D features with pre-trained Gaussians, a per-Gaussian autoencoder, and adaptive sparsification (Lee et al., 7 Aug 2025).

The feature-lifting stage computes a fused reference feature for each Gaussian and also estimates a per-Gaussian variance; the top $\sigma_d/d\simeq 0.15$9 of Gaussians with the largest variance norm are discarded because they typically lie on boundaries or in inaccurate geometry. The autoencoder is a small MLP with five layers, for example encoder dimensions $\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$0 and a symmetric decoder, compressing each lifted feature into a 3D latent code that can be stored in the Gaussian’s RGB channels. Its loss combines mean-squared error, cosine loss, and a structural term that preserves pairwise cosine relations in the encoded space.

Adaptive sparsification then jointly optimizes Gaussian attributes while pruning and merging redundancies. Pruning removes Gaussians with low global contribution $\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$1. Merging is triggered for surviving Gaussians with small gradient norm; it uses $\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$2-nearest neighbors, cosine similarity of codes, and a Mahalanobis-distance criterion, followed by moment matching for the merged Gaussian’s center, opacity, covariance, and code. The training loop optimizes encoded RGB, structure, and opacity with Adam while periodically invoking prune and merge operations.

The reported complexity reduction is substantial. Before sparsification, the field contains approximately $\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$3K Gaussians, requires approximately $\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$4 GB of storage, and renders at approximately $\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$5 FPS. After CF3 sparsification, the representation uses approximately $\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$6–$\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$7K Gaussians, occupies approximately $\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$8–$\sigma_{cz}\simeq 50\ {\rm km\ s^{-1}}$9 MB, renders at approximately $v_{\rm pec}=cz_{\rm obs}-H_0d$0–$v_{\rm pec}=cz_{\rm obs}-H_0d$1 FPS, and preserves feature quality within $v_{\rm pec}=cz_{\rm obs}-H_0d$2–$v_{\rm pec}=cz_{\rm obs}-H_0d$3 mIoU of the uncompressed baseline. On Replica with LSeg features, the paper reports that Feature-3DGS uses $v_{\rm pec}=cz_{\rm obs}-H_0d$4K Gaussians, $v_{\rm pec}=cz_{\rm obs}-H_0d$5 GB, $v_{\rm pec}=cz_{\rm obs}-H_0d$6 FPS, and $v_{\rm pec}=cz_{\rm obs}-H_0d$7, whereas CF3 uses $v_{\rm pec}=cz_{\rm obs}-H_0d$8K Gaussians, $v_{\rm pec}=cz_{\rm obs}-H_0d$9 MB, $75\ {\rm km\ s^{-1}\ Mpc^{-1}}$0 FPS, and $75\ {\rm km\ s^{-1}\ Mpc^{-1}}$1. The authors state that the approach can achieve a competitive 3D feature field using as little as $75\ {\rm km\ s^{-1}\ Mpc^{-1}}$2 of the Gaussians compared to Feature-3DGS.

4. CF$75\ {\rm km\ s^{-1}\ Mpc^{-1}}$3 as counterfactual-and-factual reasoning for financial risk detection

In graph machine learning, CF$75\ {\rm km\ s^{-1}\ Mpc^{-1}}$4 denotes “Counterfactual and Factual learning for company Financial risk detection.” It is defined on a heterogeneous company knowledge graph $75\ {\rm km\ s^{-1}\ Mpc^{-1}}$5 with node-feature matrix $75\ {\rm km\ s^{-1}\ Mpc^{-1}}$6, and for a target node $75\ {\rm km\ s^{-1}\ Mpc^{-1}}$7 it operates on an $75\ {\rm km\ s^{-1}\ Mpc^{-1}}$8-hop computation subgraph $75\ {\rm km\ s^{-1}\ Mpc^{-1}}$9. The framework consists of three trainable modules together with a fixed pre-trained target GNN $\Lambda$0: meta-path attribution, an edge-type-aware graph generator, and a layer-based feature masker. At inference time, meta-path attribution first constructs an attribution subgraph $\Lambda$1; the generator then produces a soft adjacency $\Lambda$2, and the feature masker produces a masked feature matrix $\Lambda$3; factual and counterfactual tests on the fixed predictor are used to encourage $\Lambda$4 and $\Lambda$5 to be jointly necessary and sufficient for the original prediction (Du et al., 9 Mar 2025).

The attribution stage uses Granger causality. For each predefined meta-path $\Lambda$6, the method compares the target GNN’s loss on the original computation subgraph and on the graph obtained by deleting all edges participating in $\Lambda$7, defining $\Lambda$8. Meta-paths are ranked by $\Lambda$9, and the top-$\Omega_m=0.3$0 paths induce the attribution subgraph. The edge-type-aware generator then runs a 2-layer GCN separately on each edge type to produce per-type node embeddings and decodes them to soft relation-wise adjacencies by inner product. The feature masker learns layer-specific mask matrices $\Omega_m=0.3$1, normalizes them into layer masks, and fuses them into a global mask $\Omega_m=0.3$2 that emphasizes feature dimensions important for prediction.

The learning objective combines factual sufficiency, attribution matching, and counterfactual necessity: $\Omega_m=0.3$3 Here the factual terms require predictions under masked structure and features to preserve the original label, while the counterfactual term requires removal of important edges and features to flip the label.

The reported experiments use three real-world datasets: SMEsD, SME, and CSI300E. Evaluation metrics are fidelity$\Omega_m=0.3$4, fidelity$\Omega_m=0.3$5, characterization, GEF, Recall-on-Risk, and Fleiss’ $\Omega_m=0.3$6. Across all three datasets and two target GNNs, CF$\Omega_m=0.3$7 attains the best or second-best characterization, fidelity$\Omega_m=0.3$8, fidelity$\Omega_m=0.3$9, GEF, RoR, and $18\,000$00. The paper gives specific examples: on SMEsD with a ComRisk target, CF$18\,000$01 achieves characterization $18\,000$02 versus the best baseline GMT at $18\,000$03; on SME with an HAT target, it reports $18\,000$04 versus $18\,000$05 for GMT and $18\,000$06 versus $18\,000$07; on CSI300E with a DANSMP target, it reports characterization $18\,000$08 versus $18\,000$09 and $18\,000$10 versus $18\,000$11. The method induces explanations in $18\,000$12 per instance and, after amortized training, is reported as $18\,000$13–$18\,000$14 faster than PGExplainer.

5. Other astrophysical and particle-physics usages

In the X-ray study of Abell 2256, CF3 is the label assigned to one of three cold fronts detected near the cluster center. It is the westernmost of the three contact discontinuities and lies immediately west of the primary-cluster X-ray peak, at a projected distance of roughly $18\,000$15–$18\,000$16 ($18\,000$17–$18\,000$18 kpc) from the primary cluster X-ray centroid. The surface-brightness edge is sharp and arc-shaped, convex toward the hotter ambient gas, and subtends an azimuthal angle of approximately $18\,000$19. A broken power-law fit to the surface-brightness profile gives a density jump $18\,000$20, while spectral fits give $18\,000$21 keV inside the front and $18\,000$22 keV outside, implying $18\,000$23. The pressure ratio $18\,000$24 is consistent with unity, confirming a contact discontinuity rather than a shock. The edge is unresolved at Chandra’s $18\,000$25 PSF, implying $18\,000$26 kpc and requiring suppression of perpendicular thermal conduction by factors $18\,000$27 relative to the Spitzer value. In the proposed merger scenario, CF3 is one of a pair of sloshing cold fronts induced by an earlier passage of a small galaxy group through the core (Ge et al., 2020).

A different use appears in particle-physics community planning. The Snowmass-2013 Cosmic Frontier 3 (CF3) Working Group, convened by Alexander Kusenko and Leslie J. Rosenberg, was tasked with surveying dark-matter candidates outside the canonical WIMP paradigm. Its remit included defining the theoretical and phenomenological landscape of non-WIMP dark matter, classifying candidates such as axions, asymmetric dark matter, sterile neutrinos, primordial black holes, mirror and self-interacting dark sectors, superheavy relics, and SUSY Q-balls, reviewing direct, indirect, collider, and astrophysical probes, and identifying open questions and future priorities. The report emphasizes that no single experimental approach can cover all non-WIMP candidates and therefore argues for a diversified program spanning mass scales from sub-$18\,000$28eV axions to $18\,000$29 GeV superWIMPs (Kusenko et al., 2013).

These usages show that CF3 can function either as a programmatic acronym or as a local object label within a larger astrophysical system. The referent is determined by disciplinary context rather than by a shared formal definition.

6. CF$18\,000$30 as the aliphatic trifluoromethyl group in vibrational Stark spectroscopy

In chemistry, CF$18\,000$31 denotes the trifluorinated aliphatic group considered alongside CF and CF$18\,000$32 as a vibrational Stark effect reporter. Cruz et al. examine the C–F stretching modes of these fluorination patterns and determine magnitudes and orientations of the difference-dipole vectors by density functional theory and MD-assisted solvatochromic analysis. The study reports Stark tuning rates of $18\,000$33–$18\,000$34, with the smallest and largest electric-field sensitivities for CF$18\,000$35 and CF$18\,000$36, respectively. Average electric fields of solvation are oriented along the main symmetry axis of the CFn group, while the Stark tuning-rate vectors can be tilted by up to $18\,000$37, which the authors note may enable mapping electrostatics in multiple dimensions (Cruz et al., 2024).

For 1,1,1-trifluorohexane, the two CF$18\,000$38-based normal modes are a symmetric stretch with CCC deformation at approximately $18\,000$39 and an antisymmetric stretch with CCC deformation at approximately $18\,000$40. Because the $18\,000$41 feature is heavily overlapped and its $18\,000$42 is nearly perpendicular to the solvent-field direction, the $18\,000$43 antisymmetric band is identified as the more practical solvatochromic probe. For the all-trans conformer, the antisymmetric mode has $18\,000$44 and a tilt of $18\,000$45 from the $18\,000$46 axis, whereas the symmetric mode has $18\,000$47 and a tilt of $18\,000$48. In the 1-gauche rotamer, the symmetric and antisymmetric modes have $18\,000$49 and $18\,000$50, with tilts of $18\,000$51 and $18\,000$52, respectively.

The study further identifies conformational heterogeneity and electrostatic modeling as central practical issues. DFT indicates that mode frequencies shift only weakly with rotamer, but $18\,000$53 magnitude and orientation vary; low rotational barriers imply time averaging on approximately $18\,000$54 ps scales. Fixed-charge GAFF/AMBER reproduces average electric-field magnitudes only qualitatively, whereas AMOEBA, with atomic multipoles and polarizability, yields electrostatic potentials within $18\,000$55–$18\,000$56 of the MP2 reference. The authors therefore recommend multipolar and/or polarizable MD force fields for accurate extraction of $18\,000$57. In this usage, CF$18\,000$58 is not an acronym but a chemical group whose spectroscopic response is used to probe local electrostatics in fluorinated molecules, peptides, proteins, and pharmaceuticals.