{"id":"5934fb10-2e5a-41f1-b926-86f38a5c5eca","arxiv_id":"2505.23629","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Color image sets are mapped to quaternionic subspaces and classified by a proposed geodesic distance formula, tested on ETH-80 and highway traffic videos.","lead":"The paper proposes a method to represent sets of color images as points on a quaternionic Grassmannian and to classify them by a new distance formula. It is worth a look because it combines quaternion color encoding with manifold geometry, an approach that could improve video and recognition systems that process many images at once.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Theorem 3.4 asserts the key minimization min_j ||2X_j|| = sqrt(sum arg^2(...)) without deriving it, so the claimed shortest-distance formula and the recognition results built on it are not established.","rationale":"The reader's weakest assumption identifies exactly the unproven minimization step in the proof of Theorem 3.4, and this is also the most load-bearing concern I found. Every downstream claim, including the classifier in §4, the ETH-80 and Highway Traffic results in Tables 2 and 3, and the claimed novelty of the paper, depends on formula (3.9) being the shortest distance. The proof of Theorem 3.4 reduces the problem to choosing the minimal-norm generator among all X_j satisfying exp(2X_j) = (I-2Q)(I-2P), but then asserts the minimum equals sqrt(sum a_j^2) without deriving it. This is not a cosmetic omission: the norm of X_j changes when eigenvalues of 2X_j are shifted by 2π i, and the paper does not justify selecting the principal branch, nor does it address the multiplicity of standard eigenvalues for the quaternionic matrix versus its complex representation. The remark that 'similar eigenvalues have the same norm' only shows norm equality for a fixed X_j; it says nothing about the minimum over j. The metric proof in §3.3 is also incomplete, but the failure of the distance formula would already invalidate the experiments. A 2x2 sanity check is the natural way to settle whether the formula is merely unproven or actually wrong. Since the central claim is not established as written, the reader's REJECT verdict remains appropriate; my analysis does not change that verdict.","tokens_in":14778,"tokens_out":15107,"duration_ms":147864,"concrete_test":"Instantiate Gr_{2,1}(H) with P = diag(1,0) and Q = [[cos^2 θ, cosθ sinθ],[cosθ sinθ, sin^2θ]], and compare the right-hand side of (3.9) with the direct minimum of ||X||_H over all skew-Hermitian X satisfying exp(2X)=(I-2Q)(I-2P), for θ = π/4 and θ = π/3. Repeat the check for a few random projection pairs in Gr_{3,1}(H), enumerating all logarithm branches explicitly. Agreement would show the minimization step is a repairable proof gap; disagreement would show the central formula is not the shortest geodesic distance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 3.4 is the mathematical core of the paper: the classifier in §4 and the recognition rates in Tables 2 and 3 directly use formula (3.9). The proof's decisive step is the line \"Since similar eigenvalues have the same norm, we obtain min_j ||2X_j||_H = sqrt(sum_j a_j^2).\" This is not a consequence of similarity of eigenvalues; it is a nontrivial minimization over all geodesic generators X_j satisfying exp(2X_j) = (I-2Q)(I-2P). The eigenvalues of 2X_j are i(a_j + 2π m_j), and the minimal Frobenius norm is obtained only by choosing the principal branch m_j = 0 for every eigenvalue, with special attention to eigenvalues exactly at -1. The proof neither establishes that such a branch choice is compatible with the geodesic connecting P and Q nor handles the multiplicities introduced by passing from the n standard eigenvalues of the quaternionic matrix to the 2n eigenvalues of its complex representation. Since this minimization is the only link between geodesic length and the closed-form eigenvalue expression, the central claim that (3.9) is the shortest distance is unsupported. The metric proof in §3.3 is also incomplete: it invokes an unproved extension of Thompson's theorem to quaternions and uses identities such as (3.17) without demonstrating their validity. The experimental improvements depend on this unproven distance, so the advertised 97.00% and 88.55% results inherit the gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes representing color image sets as points on quaternionic Grassmannians, where each RGB image is encoded as a quaternionic matrix and a set is summarized by a quaternionic subspace projection matrix. The central theoretical contribution is a claimed closed-form formula (3.9) for the shortest geodesic distance between two points on Gr_{n,k}(H), expressed in terms of the arguments of the standard eigenvalues of (I-2Q)(I-2P). The paper further claims this distance is a metric and uses it in a nearest-centroid classifier evaluated on the ETH-80 dataset and a Highway Traffic video dataset, reporting recognition rates of 97.00% and 88.55%, respectively. The mathematical core is Theorem 3.4 and the metric proof in Section 3.3; the experimental framework in Section 4 and the results in Section 5 depend directly on the unproven distance formula.","tokens_in":15095,"tokens_out":8291,"duration_ms":76388,"significance":"If rigorously established, the quaternionic Grassmannian distance formula would provide a natural color-preserving generalization of complex Grassmannian geodesic distances and a compact representation for image-set recognition that captures inter-channel correlations. The paper has the virtue of shipping MATLAB code and using a simple, reproducible experimental protocol. However, the central proofs contain substantial gaps: the minimization over geodesics in Theorem 3.4 is not derived, the triangle inequality proof in Section 3.3 is circular, and the extension to quaternions in Corollary 3.10 is asserted without proof. Because the experimental results inherit the unproven distance formula, the paper's main claims are not currently supported.","major_comments":[{"comment":"The proof's decisive step, 'Since similar eigenvalues have the same norm, we obtain min_j ||2X_j||_H = sqrt(sum_j a_j^2),' is not a consequence of eigenvalue similarity. The Frobenius norm of a skew-Hermitian matrix is determined by the eigenvalues of that matrix, which are only determined up to branch choices of the form i(a_j + 2π m_j). The proof does not show that the principal branch m_j = 0 yields a valid geodesic connecting P and Q, nor does it handle multiplicities arising from passing from the n standard eigenvalues of the quaternionic matrix to the 2n eigenvalues of its complex representation. Without this minimization, formula (3.9) is not established as the shortest distance.","section":"§3.2, Theorem 3.4, Eq. (3.10)"},{"comment":"The identity e^{X}e^{Y}e^{Z} = e^{-X}e^{-Y}e^{-Z} is asserted without proof. The preceding four-term sinh/cosh expression is algebraically equivalent to this identity, not a consequence of (3.16), so the proof of Theorem 3.9 is circular. Furthermore, the derivation of e^{-2X} = e^{Y}e^{Z}e^{Z}e^{Y} from (3.17) appears to treat exponentials of non-commuting matrices as if they could be freely reordered, which is not justified.","section":"§3.3, Eq. (3.17)"},{"comment":"The extension of the triangle inequality to quaternionic Grassmannians is asserted to follow from Proposition 2.1 and a reference to [29], but no proof is supplied. Thompson's theorem (Theorem 3.8) is stated for complex skew-Hermitian matrices; its quaternionic analogue requires a nontrivial argument concerning eigenvalues, similarity classes, and the branch of the logarithm, none of which is given. Since the metric property of (3.9) is a central claim, this gap is load-bearing.","section":"§3.3, Corollary 3.10"},{"comment":"The proof does not address the ambiguity of arg(-1) (the eigenvalue -1 can be assigned argument π or -π) or the fact that, per Definition 2.2 and Corollary 2.4, the standard eigenvalues of a quaternionic unitary matrix lie in the closed upper half-plane. The proof writes a_j ∈ [-π,π], which is inconsistent with the stated definition of standard eigenvalues, and the choice of branch is directly relevant to the minimization over geodesics.","section":"§3.2, Theorem 3.4"},{"comment":"The Highway Traffic baselines in Table 3 are reported without standard deviations, so the claim that the proposed method 'significantly outperforms' all baselines is not supported; given the reported ±2.46% for the proposed method, overlap with baselines is possible. Additionally, Remark 5.1 shows ETH-80 average rates varying across rounds from 93.0% to 97.25%, so the single reported value 97.00% in Table 2 is not presented as an aggregate over the rounds described, weakening the empirical support.","section":"§5, Tables 2 and 3"}],"minor_comments":[{"comment":"The abstract contains a duplicated word: 'ETH-80 benchmark dataset and and the Highway Traffic video dataset'.","section":"Abstract"},{"comment":"Algorithm 1 line 6 says 'Reduce ... by QSVD' while the surrounding text in §4.1 describes the step as Quaternion Principal Component Analysis (QPCA); please unify the terminology.","section":"§4.1, Algorithm 1"},{"comment":"The sentence 'Let X ∈ UH(n)' appears to be a typo: the generator of a geodesic on the Grassmannian should be a skew-Hermitian matrix in sH(n), not a unitary matrix.","section":"§3.2, before Theorem 3.4"},{"comment":"There is an unresolved citation '[ ?]' in the sentence 'particularly in [ ?] and [27]'.","section":"§3.3"},{"comment":"The notation \\(\\widetilde{X}_{PQ}\\) is used without a formal definition, and the placement of the tilde is inconsistent across the proof of Theorem 3.9.","section":"§3.3, Proposition 3.7"},{"comment":"It is unclear whether the baseline methods in Table 2 were re-run under the same random splits as the proposed method or taken from their original papers; please clarify the comparison protocol.","section":"§5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is potentially useful, and the code release is a strength. However, the theoretical gaps are substantial and the experimental comparisons lack some necessary detail. If the authors can supply rigorous proofs for the distance formula and the metric property, the paper could become acceptable. The current version is not ready for publication in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea—encode color image sets as points on a quaternionic Grassmannian and use a principal-angle-like distance—is genuinely new for the computer-vision side, and the paper ships a working MATLAB implementation. That is real and worth acknowledging.\n\nBut the mathematical core does not hold up as written. The claimed shortest-distance formula (3.9) is very likely correct; for complex Grassmannians it is the standard principal-angle distance, scaled by 1/2. The problem is that Theorem 3.4 does not prove it. The line 'Since similar eigenvalues have the same norm, we obtain min_j ||2X_j||_H = sqrt(sum a_j^2)' is not a consequence of eigenvalue similarity; it is a minimization over logarithm branches. One has to show that the principal branch, with eigenvalues in [-π,π], corresponds to a valid geodesic connecting P and Q. That is exactly the missing step. The proof also does not handle eigenvalue multiplicities when passing to the complex representation.\n\nThe triangle-inequality proof is in worse shape. It invokes an unproved quaternionic version of Thompson's theorem and then uses identity (3.17), e^A e^B e^C = e^{-A} e^{-B} e^{-C}, which appears to be false for skew-Hermitian matrices. If that identity does not hold, the metric argument collapses.\n\nThe experimental section is honest in some ways—the authors disclose variability and limitations—but the comparisons are not fully fair. On ETH-80 the gain over GGPLCR is 97.00 vs 96.75, well within the reported standard deviations. On Highway Traffic the baselines are borrowed from a different paper without re-implementing them under the same splits, so the 88.55% jump is not substantiated.\n\nWho should read this? People working on Grassmannian methods for image sets will find the quaternionic encoding idea attractive, and the GitHub code is a starting point. But the paper currently overclaims: it says 'direct formula' and 'shortest distance' without a sound proof.\n\nMy recommendation: send it to review, but expect major revision. The referee should insist on a complete proof of Theorem 3.4, a correct triangle-inequality argument, and re-implemented baselines. If the authors can close the gap, the paper becomes useful; in its current form it presents an interesting idea with unproven foundations.","headline":"A fresh application idea with a working implementation, but the central distance formula is not proven and the metric proof contains an apparent false identity.","tokens_in":15614,"tokens_out":4274,"would_cite":false,"duration_ms":40549,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a closed-form shortest-geodesic distance on quaternionic Grassmannians, combined with a nearest-centroid rule, recognizes color image sets with high accuracy.","keywords":["quaternionic Grassmannian","color image set recognition","shortest geodesic distance","standard eigenvalues","quaternion PCA","nearest-centroid classification","ETH-80","Highway Traffic dataset"],"falsifier":"Take two points $P,Q$ in $\\mathrm{Gr}_{n,k}(\\mathbb{H})$ such that $(I-2Q)(I-2P)$ has $-1$ as an eigenvalue and compute the formula's value under both choices $\\arg(-1)=\\pm\\pi$; if the two values differ while a numerical geodesic-distance optimizer gives a single number, the formula as stated is not well-defined. Alternatively, randomly sample triples with eigenvalues near $-1$ and check the triangle inequality numerically; a violation would refute the metric claim.","tokens_in":14561,"feed_emoji":"🎨","tokens_out":6202,"duration_ms":60252,"temperature":0.7,"pith_summary":"This paper claims that color image sets can be recognized by lifting each set to a point on a quaternionic Grassmannian and using a newly derived closed-form formula for the shortest geodesic distance between two such points. The central contribution is Theorem 3.4, which expresses that distance as half the root-sum-square of the arguments of the standard eigenvalues of the quaternionic unitary matrix $(I-2Q)(I-2P)$. The authors argue this quantity is a genuine metric and build a nearest-centroid classifier on it. On the ETH-80 benchmark the classifier reaches 97.00% average accuracy, and on the Highway Traffic video dataset it reaches 88.55%, both above the baselines compared in the paper.","feed_headline":"Quaternionic Grassmannian distance hits 97% on color image sets","feed_subtitle":"A closed-form shortest distance between color-image subspaces lets one classifier use RGB structure directly.","key_machinery":"The load-bearing object is the quaternionic Grassmannian $\\mathrm{Gr}_{n,k}(\\mathbb{H})$, realized as the set of $n\\times n$ quaternionic orthogonal projection matrices of rank $k$. Geodesics are written $\\gamma(t)=\\exp(tX)P\\exp(-tX)$ with $X$ skew-Hermitian quaternionic, and the endpoint condition $\\exp(2X)=(I-2Q)(I-2P)$ links the geodesic generator to the two endpoints. The distance formula then reads off the generator's size from the arguments of the standard eigenvalues of that unitary product. Color enters by encoding each RGB image as a pure quaternionic matrix, reducing the set with quaternionic PCA, orthonormalizing, and forming the projection $XX^\\ast$.","core_discovery":"The paper's central claim is that for any two points $P$ and $Q$ on the quaternionic Grassmannian $\\mathrm{Gr}_{n,k}(\\mathbb{H})$, the shortest geodesic distance is $$ \\hat d(P,Q)=\\frac12\\sqrt{\\sum_j \\$arg^{2}$\\big(\\hat\\lambda_j((I-2Q)(I-2P))\\big)}, $$ where $\\hat\\lambda_j$ runs over the standard eigenvalues of the quaternionic unitary matrix. The authors state this formula is compatible with real and complex Grassmannians, that it satisfies non-negativity, symmetry, and the triangle inequality, and that it supplies the distance used in their classification framework. The proof routes through the geodesic equation $\\gamma(t)=e^{tX}Pe^{-tX}$ with boundary condition $e^{2X}=(I-2Q)(I-2P)$, the quaternionic Schatten-2 norm, and the complex representation of quaternionic matrices.","pith_inferences":["If the distance formula is genuinely a metric, a natural extension is to feed it into discriminative learning on quaternionic Grassmannians; the paper itself only uses a nearest-centroid rule.","The branch ambiguity at eigenvalue $-1$ is the spot where the argument of the formula could break; a numerical experiment comparing the formula against a direct geodesic optimization on pairs with an eigenvalue exactly at $-1$ would settle whether the claimed shortest distance is correct.","The same eigenvalue-argument construction may transfer to other symmetric spaces of unitary type, giving closed-form distances in settings the paper does not discuss.","The reported standard deviations (up to about $\\pm3.5\\%$ across rounds on ETH-80) suggest accuracy depends on the random split, so quantifying how much of the advantage comes from the metric versus the quaternionic representation would clarify the method's real margin."],"forward_implications":["On the paper's terms, the formula supplies a metric on $\\mathrm{Gr}_{n,k}(\\mathbb{H})$, so nearest-centroid and nearest-neighbor classifiers can be run directly on color image sets without splitting RGB channels.","The framework extends to color video by treating each video as a set of frames and mapping the set to one quaternionic Grassmannian point.","Because the formula is claimed compatible with real and complex Grassmannians, the same distance expression can serve subspace comparisons in settings beyond color.","The reported ETH-80 accuracy of 97.00% and Highway Traffic accuracy of 88.55% indicate the representation preserves enough chromatic and geometric structure to beat the compared subspace baselines.","The computation of standard eigenvalues of $(I-2Q)(I-2P)$ is the main bottleneck, so faster eigenvalue algorithms would make the method practical for large image sets."],"supporting_citations":[{"why":"Supplies the geodesic boundary condition $\\exp(2X)=(I-2Q)(I-2P)$ and the tangent-vector properties used to express the distance.","marker":"[27]"},{"why":"Provides the quaternionic singular value decomposition and spectral facts underlying the Schatten-2 norm and standard eigenvalues.","marker":"[21]"},{"why":"Gives the complex representation of quaternionic matrices and the eigenvalue symmetry used to justify standard eigenvalues.","marker":"[19]"},{"why":"Supplies the Grassmannian tangent-space parametrization that the proof adapts to the quaternionic setting.","marker":"[25]"},{"why":"Supports the quaternionic version of the exponential factorization identity used in the triangle-inequality proof.","marker":"[29]"},{"why":"Provides quaternion PCA for reducing color image sets before forming the Grassmannian representation.","marker":"[12]"},{"why":"Supplies the ETH-80 dataset used in the recognition benchmark.","marker":"[32]"},{"why":"Supplies the Highway Traffic video dataset used in the second recognition benchmark.","marker":"[33]"}],"fun_headline_variants":["Quaternionic Grassmannian distance unlocks color image sets","Closed-form distance on quaternionic Grassmannian for color sets","Quaternion geometry gives direct distance for image set recognition","Quaternionic Grassmannian formula improves color image set classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that the shortest geodesic is obtained by counting each principal angle exactly once, i.e., that the norm of the optimal generator equals the square root of the sum of squared arguments of the standard eigenvalues, with no extra multiplicities and no inconsistency in choosing the argument of $-1$; if that step is false, the formula is only a partial or scaled distance.","fun_headline_variants_meta":{"raw":{"variants":["Quaternionic Grassmannian distance unlocks color image sets","Closed-form distance on quaternionic Grassmannian for color sets","Quaternion geometry gives direct distance for image set recognition","Quaternionic Grassmannian formula improves color image set classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000903,"raw_usage":{"total_tokens":3831,"prompt_tokens":834,"completion_tokens":2997,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":2928}},"tokens_in":450,"tokens_out":2997,"duration_ms":21472,"temperature":1.0,"reasoning_tokens":2928,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:43:20.455203+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two points $P,Q$ in $\\mathrm{Gr}_{n,k}(\\mathbb{H})$ such that $(I-2Q)(I-2P)$ has $-1$ as an eigenvalue and compute the formula's value under both choices $\\arg(-1)=\\pm\\pi$; if the two values differ while a numerical geodesic-distance optimizer gives a single number, the formula as stated is not well-defined. Alternatively, randomly sample triples with eigenvalues near $-1$ and check the triangle inequality numerically; a violation would refute the metric claim.","supporting_citations":[{"cited_title":"Linear Algebra Application 466, 83–101 (2015)","cited_arxiv_id":null,"evidence_quote":"Supplies the geodesic boundary condition $\\exp(2X)=(I-2Q)(I-2P)$ and the tangent-vector properties used to express the distance."},{"cited_title":"Linear algebra and its applications 251, 21–57 (1997)","cited_arxiv_id":null,"evidence_quote":"Provides the quaternionic singular value decomposition and spectral facts underlying the Schatten-2 norm and standard eigenvalues."},{"cited_title":"In: Proceedings of the Royal Irish Academy","cited_arxiv_id":null,"evidence_quote":"Gives the complex representation of quaternionic matrices and the eigenvalue symmetry used to justify standard eigenvalues."},{"cited_title":"Advances in Computational Mathematics 50, 6 (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the Grassmannian tangent-space parametrization that the proof adapts to the quaternionic setting."},{"cited_title":"Princeton Series in Applied Mathematics","cited_arxiv_id":null,"evidence_quote":"Supports the quaternionic version of the exponential factorization identity used in the triangle-inequality proof."},{"cited_title":"An Empirical Analysis of Privacy in the Lightning Network","cited_arxiv_id":"2003.12470","evidence_quote":"Provides quaternion PCA for reducing color image sets before forming the Grassmannian representation."},{"cited_title":"Interference as an information-theoretic game","cited_arxiv_id":"2003.12114","evidence_quote":"Supplies the ETH-80 dataset used in the recognition benchmark."},{"cited_title":"In: 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition 1, 846–851 (2005)","cited_arxiv_id":null,"evidence_quote":"Supplies the Highway Traffic video dataset used in the second recognition benchmark."}],"review_version":1}