{"id":"e62cb002-2f0d-408e-82c8-1c4d6e10dc35","arxiv_id":"2412.08934","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A review-style cheat sheet that compiles existing orientation distributions (von Mises, vMF, Kent, Bingham, ACG, Matrix Fisher, tangent-space Gaussians) with fitting and sampling recipes, plus a Python library and two applications.","lead":"This paper assembles a practical cheat sheet of probability distributions for angles, directions, quaternions, and rotation matrices, with density formulas, fitting routines, and sampling algorithms. It also ships a Python library and two small robotics examples comparing four orientation models.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (38) uses Tr(D) as the normalizing constant of the tangent-space Gaussian; the correct determinant is det(D), so the cheat sheet's density formula is wrong and users will compute mis-scaled densities.","rationale":"The paper's central claim is to be a reliable 'cheat sheet' giving engineers the strictly necessary density, fitting, and sampling tools for orientation distributions. I checked the main formulas against standard directional-statistics results; most are standard and the accompanying library is a useful contribution. The reader's weakest_assumption targets the Section 8.2 comparison of concentration parameters across families without uncertainty quantification; I agree those experimental conclusions are over-claimed. However, the single most load-bearing concern for the paper's stated purpose is in Section 5.3: Eq. (38) writes the normalizing constant of the tangent-space Gaussian as sqrt((2π)^(d−1) Tr(D)) instead of sqrt((2π)^(d−1) det(D)). This is not a matter of convention: with the paper's own diagonal covariance in Table 1, the two constants differ by about a factor of 55 in the density value. The problem is internal and checkable, it affects a core model rather than just the experiments, and it directly undermines the reliability promise of the cheat sheet. Since the reader's verdict is already CONDITIONAL and this finding strengthens the need for a corrected revision, I do not propose moving the verdict; I would keep it CONDITIONAL with a mandatory fix of Eq. (38).","tokens_in":25310,"tokens_out":8707,"duration_ms":91157,"concrete_test":"Analytically verify Eq. (38) against the standard MVN density in tangent coordinates: take d = 4 and D = diag(0.227, 0.033, 0.012), and evaluate ∫_{R^3} f(y) dy for y = B log_μ(x). The printed normalizer with Tr(D) gives ∫ f ≠ 1 (approximately sqrt(det(D)/Tr(D)) ≈ 0.018), while replacing Tr(D) by det(D) integrates to 1. If the printed version fails this integral check, Eq. (38) must be corrected to det(D).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.3 defines the tangent-space Gaussian density (Eq. (38)) as f_TG,μ(x | B,D) = 1/sqrt((2π)^(d−1) Tr(D)) exp(−1/2 log_μ(x)^T B^T D^{−1} B log_μ(x)), with D = diag(λ1,…,λ_{d−1}). For y = B log_μ(x) ∈ R^{d−1}, a zero-mean MVN with covariance D has normalizer (2π)^(−(d−1)/2) (det D)^(−1/2). Replacing det(D) by Tr(D) is not a legitimate normalization choice. In the paper's own experiment 1, D = diag(0.227, 0.033, 0.012): Tr(D) = 0.272, det(D) ≈ 8.99×10^−5, so the two normalizers differ by sqrt(Tr(D)/det(D)) ≈ 55. Density values computed from Eq. (38) would be off by this factor. This matters because the paper promises a reliable density/fitting/sampling reference, and Eq. (38) is one of the core models used in the tangent-space rows of Tables 1–2. The caveat that the density is not normalized over D_μ only explains why ∫_{D_μ} f ≠ 1; it does not justify changing det(D) to Tr(D) inside the Gaussian constant. This is an internally checkable mathematical error, not a modeling preference.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript is a review/tutorial on probability distributions for orientational data, covering 1-DOF (wrapped normal, von Mises), 2-DOF (vMF, Kent, ESAG), and 3-DOF (Bingham, ACG, tangent-space Gaussians, matrix Fisher, Lie-algebra Gaussians) models. For each model, it presents density formulas, parameter estimation and sampling methods, and it is accompanied by a Python library (RotStats) and two real-data application experiments (a pouring task and a camera-calibration comparison). The paper concludes that the angular central Gaussian (ACG) distribution is the most convenient model for orientation data.","tokens_in":25662,"tokens_out":13483,"duration_ms":121271,"significance":"If the formulas and code are correct, this is a genuinely useful single reference for engineers and computer scientists who need to work with orientation data. The paper's strengths are its breadth, the explicit treatment of density/fitting/sampling for each model, the inclusion of a companion library, and the visualizations on S^2 and SO(3). However, the presence of several mathematical errors in core equations (the tangent-space Gaussian normalizer, the vMF concentration approximation, and the ESAG density) undermines its reliability as a 'cheat sheet'. These errors are fixable and do not invalidate the overall approach, but they must be corrected before the paper can serve as a dependable reference.","major_comments":[{"comment":"The normalizing constant of the tangent-space Gaussian is written with Tr(D), but the correct constant for a zero-mean MVN with covariance D in R^{d-1} is (2π)^(d-1)/2 (det D)^(1/2). For the dataset in Table 1, D = diag(0.227, 0.033, 0.012), Tr(D) = 0.272 and det(D) ≈ 8.99×10^-5, so densities computed with Eq. (38) are off by a factor of sqrt(Tr(D)/det(D)) ≈ 55. Since the paper promises accurate density formulas, this is a load-bearing error that must be fixed.","section":"Eq. (38), Section 5.3"},{"comment":"The large-ρ approximation for the vMF concentration parameter is stated as κ ≈ (d−1)/(d(1−ρ)). For a vMF on S^{d-1}, the score equation ρ = I_{d/2}(κ)/I_{d/2−1}(κ) has the large-κ expansion ρ ≈ 1 − (d−1)/(2κ), so the correct leading approximation is κ ≈ (d−1)/(2(1−ρ)). The factor d in the denominator is incorrect; for d=3, the manuscript gives 2/(3(1−ρ)) instead of 1/(1−ρ), which will materially bias the estimated concentration.","section":"Eq. (13), Section 3.2"},{"comment":"The ESAG density has a denominator 2π(x^T V^{-1}x), but the correct expression (Paine et al., Statistics and Computing, 2018) has denominator 2π sqrt(x^T V^{-1}x). As written, the density is not the ESAG density and will yield incorrect likelihoods in the fitting procedure of Eq. (27).","section":"Eq. (22), Section 4.3"},{"comment":"Eq. (18) defines jK = kK × jK, which is self-referential and cannot be evaluated; it should be jK = kK × iK (or an equivalent right-handed construction). In Eq. (24), the text introduces a 'frame K' but then labels its axes {iE, jE, kE}, and the line 'Since kE = γ1' contradicts the previous definition kE = ξ3. These typos will confuse any reader implementing the reparametrizations and should be corrected.","section":"Eq. (18) and Eq. (24), Sections 4.2 and 4.3"},{"comment":"The conclusion that the CAD calibration is more precise than HEC is based on direct comparison of concentration parameter values across different model families (ACG eigenvalue ratios, Bingham κ_i, Matrix Fisher sums ψ_j+ψ_k, tangent-space eigenvalues) without error bars, confidence intervals, or significance tests. Because these parameters have different meanings and scales, the claim that 'all four models agree' is not statistically established. This does not affect the central cheat-sheet content, but it should be reframed as qualitative or supplemented with uncertainty quantification.","section":"Section 8.2, Table 2"}],"minor_comments":[{"comment":"The text says 1-DOF orientations form 'the manifold of a circle S^1 and the group R^1'; the group should be SO(2) (or S^1), not R^1.","section":"Section 2.1"},{"comment":"The word 'Condiser' is a typo and should be 'Consider'.","section":"Section 5.2"},{"comment":"The word 'distribuitons' is a typo and should be 'distributions'.","section":"Section 8.1"},{"comment":"The phrase 'This defers from its use' should be 'This differs from its use'.","section":"Section 2"},{"comment":"The log map with the antipodal metric (37) is ambiguous for points with μ^T x < 0; please clarify how x is represented when projecting to the disk D_μ, since the formula uses x directly.","section":"Eq. (36), Section 5.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a single-author review with a useful scope and a companion library, but it contains several errors in core formulas (Eqs. 38, 13, 22) that would mislead practitioners. The author should be asked to correct these and to carefully re-verify all other equations, as the density formulas and estimation procedures are the main deliverable. The experimental comparison in Section 8.2 is suggestive but lacks statistical rigor; the conclusions about CAD vs HEC and the general preference for ACG should be softened or properly quantified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this as a useful cheat sheet with one load-bearing typo. The compilation is genuinely handy: density, fitting, sampling for vM, vMF, Kent, ESAG, Bingham, ACG, Matrix Fisher, tangent-space Gaussians, plus a Python library and two real-data examples. Most formulas check out and are cited to Mardia & Jupp, Kent, Tyler, Lee, etc. The library is a real deliverable; I'd use it after the errors below are fixed.\n\nThe stress-test is right and it is not minor. Eq. (38) defines the tangent-space Gaussian with 1/sqrt((2π)^(d−1) Tr(D)) instead of 1/sqrt((2π)^(d−1) det(D)). That mis-scales every density from that model by sqrt(Tr(D)/det(D)) — about 55 in their own Experiment 1. The caveat that the density isn't normalized over D_μ is true but does not justify replacing det with Tr. A cheat sheet whose selling point is reliability cannot ship a wrong normalizer in one of the core models used in Tables 1–2. This must be fixed before the paper is useful.\n\nOther soft spots: Eq. (18) defines jK = kK × jK, self-referential; should be kK × iK. Section 4.3 reuses the same frame letters and then says \"Since kE = γ1\" when γ1 isn't defined there; that's a naming slip. Experiments: N=20 and N=30, no error bars, and the conclusion that CAD calibration is more precise than HEC depends on comparing concentration parameters across different families. That comparison is suggestive at best. The broader claim that ACG is \"most convenient\" follows from model properties (closed-form normalizer, easy sampling), not from the experiments, and the paper mostly frames it that way in the conclusions.\n\nWhat's good: the scope is appropriate, notation compromise is sensible, the library appears to cover what the paper promises, and the visualizations (axis marginalization on S^2) are a nice touch. Citation practice is fine; self-citations are to the author's own relevant preprint and library, not padding.\n\nWho this is for: engineers and roboticists who need a practical orientation-distribution reference. It deserves peer review — not because it breaks new statistical ground, but because a serious referee will catch exactly the kind of error above before engineers copy formulas into code. With the normalizer fixed and the experimental claims softened, I'd happily point students to it.","headline":"Useful practical cheat sheet with a load-bearing normalization error in Eq. (38); fix that and the experimental overreach, and it earns its place.","tokens_in":26135,"tokens_out":3283,"would_cite":false,"duration_ms":34019,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H11","62F10"],"pacs":[],"model":"deepseek-v4-flash","headline":"One reference now gives density, fitting, and sampling recipes for every common orientation distribution.","keywords":["orientational data","directional statistics","probability distributions on manifolds","quaternions","SO(3)","angular central Gaussian distribution","Bingham distribution","acceptance-rejection sampling"],"falsifier":"Fit all four 3-DOF models (ACG, Bingham, matrix Fisher, tangent-space Gaussian) to synthetic orientation data drawn from a known distribution, then compare held-out log-likelihood and sampling cost at matched concentration; if ACG is not competitive with Bingham or matrix Fisher, the paper's convenience claim is wrong. On the calibration experiment, bootstrap confidence intervals for the eigenvalue ratio a1/a2 of the ACG fits would show whether the CAD advantage is statistically real.","tokens_in":25109,"feed_emoji":"🧭","tokens_out":7415,"duration_ms":76035,"temperature":0.7,"pith_summary":"Orientational data—angles, unit vectors, quaternions, and rotation matrices—live on curved manifolds where ordinary multivariate normal statistics break down. This paper collects the probability distributions that respect those manifolds and, for each, gives the density, a parameter-fitting procedure, and a sampling algorithm, with a companion Python library. Its working claim is that a single practical reference can close the gap between directional statistics and engineering practice. From the theory and two small real-data experiments, the paper concludes that the angular central Gaussian distribution is the most convenient model for 3-DOF orientations, because its normalising constant is closed-form, fitting is a simple iteration, and sampling reduces to normalising normal draws.","feed_headline":"Orientation statistics: density, fitting, and sampling in one place","feed_subtitle":"A practical reference for circular, spherical, quaternion, and rotation-matrix data, with code to match.","key_machinery":"The central object is the manifold of the data: S, S2, S3/Z2, and SO(3). The carrying mechanisms are the scatter matrix and mean resultant length for descriptive statistics; acceptance-rejection sampling using enveloping distributions, notably an ACG envelope for the Bingham distribution; and the exponential and logarithmic maps used for tangent-space and Lie-algebra Gaussians. The key identity that makes ACG convenient is that it is the distribution of the direction of a zero-mean multivariate normal: y/∥y∥ with y ∼ N(0, Λ), so sampling is trivial and the normalising constant is a closed-form gamma expression instead of an infinite sum.","core_discovery":"The paper's central claim is that the practical toolkit for modelling orientational data reduces to a small set of recipes, one per manifold: wrapped normal and von Mises on the circle, von Mises-Fisher, Kent, and ESAG on the sphere, Bingham and angular central Gaussian on the quaternion sphere with antipodal symmetry, and matrix Fisher and tangent-space Gaussian on rotation matrices. For each model the paper supplies a density, an estimator, and a sampler, avoiding expensive normalising constants where possible, and the accompanying library implements them. The survey argues that the angular central Gaussian is the most convenient model for 3-DOF orientations: its density has a closed-form normalising constant, its maximum-likelihood fit is a fixed-point iteration, and sampling is simply drawing from a zero-mean multivariate normal and normalising to unit length; Bingham and matrix Fisher sampling are themselves done through an ACG envelope. The two applications—a pouring-demonstration dataset and an RGBD camera calibration comparison—are offered as evidence that the models work on real data and that the ACG-based comparison favours CAD calibration over hand-eye calibration.","pith_inferences":["The paper stops short of a statistical comparison of models; a principled benchmark using held-out log-likelihood on synthetic ground truth would test its convenience claim directly.","The CAD-versus-HEC result would be stronger with resampling confidence intervals; the agreement of four models is suggestive but not a significance test.","The cheat-sheet format points toward a natural extension: coupled models for full SE(3) poses, pairing a position Gaussian with these orientation models; the paper notes the bi-invariant-metric caveat but does not develop it.","Because ACG sampling is essentially free, replacing Bingham-mixture components with ACG components in grasping and robot-learning pipelines is a cheap, testable engineering experiment."],"forward_implications":["An engineer can now treat density, fitting, and sampling for 1-DOF, 2-DOF, and 3-DOF orientations as solved recipes rather than literature searches.","The angular central Gaussian should become the default first choice for 3-DOF orientation data, ahead of Bingham and matrix Fisher, because its normaliser is closed-form and sampling is just normalising normal draws.","The companion Python library lets any practitioner reproduce the pouring-task and camera-calibration results and apply the same models to their own orientation data.","For concentrated data, tangent-space or Lie-algebra Gaussians are acceptable and simple, but only the spherical-normal variant carries a proper normalising constant.","The diffusion-equation section connects heat-kernel densities on S, S2, and SO(3), giving a route to non-parametric density estimation on orientations."],"supporting_citations":[{"why":"Supplies the wrapped-normal and von Mises densities, the sample mean and mean resultant length, and the acceptance-rejection scheme for von Mises.","marker":"[39]"},{"why":"Defines the Bingham distribution for antipodally symmetric data on the sphere and its normalising constant.","marker":"[7]"},{"why":"Defines the Kent (FB5) distribution with elliptical contours and its moment-based parameter estimation.","marker":"[24]"},{"why":"Defines the elliptically symmetric angular Gaussian (ESAG) distribution with a closed-form normalising constant.","marker":"[42]"},{"why":"Defines the angular central Gaussian distribution and proves the fixed-point maximum-likelihood iteration.","marker":"[58]"},{"why":"Presents the matrix Fisher distribution on SO(3), its normalising constant integrals, and the Bingham-based sampling scheme.","marker":"[33]"},{"why":"Provides the acceptance-rejection algorithm that samples the Bingham distribution using an ACG envelope.","marker":"[25]"},{"why":"Introduces the spherical normal distribution in the tangent space with a correct normalising constant.","marker":"[21]"},{"why":"Gives the efficient sampling algorithm used for the von Mises-Fisher distribution.","marker":"[63]"},{"why":"Brings the Riemannian Gaussian model for quaternions into robot learning, the setting the paper targets.","marker":"[64]"}],"fun_headline_variants":["Orientation distributions cheat sheet: density, fit, sample, Python","Modeling orientations: a cheat sheet with code","Orientation stats made practical: from circles to rotations","Density, fitting, sampling for orientation data—in one sheet","Angular central Gaussian and friends: orientation stats cheat sheet"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion that ACG is the most convenient model and that CAD calibration is more precise rests on directly comparing concentration parameters across different model families using N=20 and N=30 datasets without error bars or statistical tests.","fun_headline_variants_meta":{"raw":{"variants":["Orientation distributions cheat sheet: density, fit, sample, Python","Modeling orientations: a cheat sheet with code","Orientation stats made practical: from circles to rotations","Density, fitting, sampling for orientation data—in one sheet","Angular central Gaussian and friends: orientation stats cheat sheet"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000274,"raw_usage":{"total_tokens":1626,"prompt_tokens":915,"completion_tokens":711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":629}},"tokens_in":531,"tokens_out":711,"duration_ms":7820,"temperature":1.0,"reasoning_tokens":629,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:21:48.970433+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit all four 3-DOF models (ACG, Bingham, matrix Fisher, tangent-space Gaussian) to synthetic orientation data drawn from a known distribution, then compare held-out log-likelihood and sampling cost at matched concentration; if ACG is not competitive with Bingham or matrix Fisher, the paper's convenience claim is wrong. On the calibration experiment, bootstrap confidence intervals for the eigenvalue ratio a1/a2 of the ACG fits would show whether the CAD advantage is statistically real.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the wrapped-normal and von Mises densities, the sample mean and mean resultant length, and the acceptance-rejection scheme for von Mises."},{"cited_title":"B INGHAM , An antipodally symmetric distribution on the sphere, The Annals of Statistics, (1974), pp","cited_arxiv_id":null,"evidence_quote":"Defines the Bingham distribution for antipodally symmetric data on the sphere and its normalising constant."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Kent (FB5) distribution with elliptical contours and its moment-based parameter estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the elliptically symmetric angular Gaussian (ESAG) distribution with a closed-form normalising constant."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the angular central Gaussian distribution and proves the fixed-point maximum-likelihood iteration."},{"cited_title":"LEE, Bayesian attitude estimation with the matrix Fisher distribution on SO(3), IEEE Transactions on Automatic Control, 63 (2018), pp","cited_arxiv_id":null,"evidence_quote":"Presents the matrix Fisher distribution on SO(3), its normalising constant integrals, and the Bingham-based sampling scheme."},{"cited_title":"H AUBERG , Directional statistics with the spherical normal distribution, in 2018 21st International Conference on Information Fusion (FUSION), IEEE, 2018, pp","cited_arxiv_id":null,"evidence_quote":"Introduces the spherical normal distribution in the tangent space with a correct normalising constant."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the efficient sampling algorithm used for the von Mises-Fisher distribution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Brings the Riemannian Gaussian model for quaternions into robot learning, the setting the paper targets."}],"review_version":1}