{"id":"ec9adb8a-ec92-4c18-bbc1-768cd8bf000a","arxiv_id":"2508.14369","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A closed-form Hilbert distance on the matrix box {0 ≺ X ≺ I} (the extended Gaussian parameter space) is derived, plus a claimed full isometry classification: orthogonal conjugation and complement X → I−X.","lead":"This paper derives a closed-form distance formula (a Hilbert metric) for the space of symmetric matrices with eigenvalues strictly between 0 and 1, the parameter space for extended Gaussian distributions with possible degenerate covariance or precision. It also characterizes the full isometry group of that metric; both results are new, though the isometry classification has a proof gap.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4's affine-reduction step is unjustified and is false in the 1D analog; without it, the isometry classification (Theorem 5) is unsupported.","rationale":"The reader's weakest assumption is the same as the load-bearing concern: the proof of Theorem 4 asserts that projectivizations of linear cone automorphisms are affine, without proof, and this assertion is false in the one-dimensional analog. The reader's conditional verdict is appropriate: Theorem 1's closed-form Hilbert distance appears correct and independently verifiable via Birkhoff's characterization, and the invariance results are solid; but the classification of all isometries, presented as a main contribution, rests on an unproven and locally contradicted step. I find no reason to move the verdict: the paper should be accepted only after the affine-reduction step in §5.2.1 is either justified for n≥2 or replaced with a correct argument, and the n=1 restriction on Theorem 4 should be stated explicitly if intended. No ad hominem considerations arise; the issue is purely in the mathematical argument.","tokens_in":15280,"tokens_out":12240,"duration_ms":148701,"concrete_test":"Set n=2 and parameterize L(X,t)=(AX+tB, tr(CX)+st). For a candidate with C≠0, compute the induced projectivization f(X)=(AX+B)/(tr(CX)+s) after setting t0=1. Enforce L(C2)=C2 symbolically: f and f^{-1} must send the boundary hypersurfaces det X=0 and det(I−X)=0 to themselves, with a possible swap. Solve the resulting polynomial identities for the entries of A,B,C,s (e.g., with sympy over rationals/generic coefficients). If a solution with nonconstant denominator exists, Theorem 4 is false; if all solutions force C=0 and reduce to P∈O(n) plus X↦I−X, the theorem is true and only the proof's affine step needs repair. A quick partial check: L(X,t)=(2X, tr(X)+t) is not surjective on C2 because (I,1.1) has no preimage in C2, illustrating why the 1D counterexample does not automatically lift.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Location: §5.2.1, proof of Theorem 4, paragraph beginning 'Because L maps the origin...'. The paper claims that a linear automorphism L of Cn 'maps rays through the origin to rays through the origin' and therefore 'for each fixed t0, π∘L(·,t0)∘ιt0 is an affine isomorphism.' The conclusion does not follow: π(L(X,t0)) = (AX+t0B)/(tr(CX)+s t0), which is affine only if C=0, precisely what the proof must establish. The subsequent use of Lemma 3, which classifies affine automorphisms of VPM, therefore applies only to the maps the argument was supposed to characterize. This is not a cosmetic gap: in the n=1 analog, L(x,t)=(2x,x+t) preserves C1={(x,t):0<x<t} but induces x↦2x/(1+x), a non-affine fractional linear map that is a Hilbert isometry of (0,1). The paper supplies no argument excluding such 'matrix Möbius' cone automorphisms for n≥2. Since Theorem 4 is the bridge from 'isometries are collineations' (Corollary 1) to the classification of isometries in Theorem 5, the 'only these isometries' contribution is not proven. Theorem 1's distance formula and the invariance claims in Propositions 7–8 are independent of this step and appear sound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the Hilbert geometry of the open bounded convex set VPM(n) = {X ∈ Sym(n) : 0 ≺ X ≺ I}, the interior of the variance-precision bicone parameterizing extended centered Gaussian families. Theorem 1 gives a closed-form Hilbert distance: dH(A,B)=log( max(λmax(B^{-1}A),λmax((I−B)^{-1}(I−A))) / min(λmin(B^{-1}A),λmin((I−B)^{-1}(I−A))) ). Propositions 7 and 8 record the isometries X↦I−X and X↦U^T XU, and Theorem 5 asserts that these generate the full isometry group for n>1. The derivation of Theorem 1 via Birkhoff's characterization is essentially correct, and the invariance proofs are correct. The isometry classification, however, depends on Theorem 4, whose proof contains an unjustified and in general false affine-reduction step; the 1D analogue provides a concrete non-affine projective automorphism. Thus the 'only these isometries' contribution is not proven as written.","tokens_in":15451,"tokens_out":13258,"duration_ms":140864,"significance":"If the missing step can be supplied, the paper is a useful contribution: it provides an explicit Hilbert metric on the variance-precision manifold that depends only on extreme eigenvalues, exhibits two natural invariance groups, and gives a plausible classification of isometries. The Birkhoff-cone derivation is reproducible from cited standard results and involves no fitted parameters or numerical tuning. The weaknesses are confined to the classification part, but that part is advertised as a main result, so the manuscript requires a substantial repair before it can be accepted.","major_comments":[{"comment":"The paragraph beginning 'Because L maps the origin...' asserts that a linear automorphism L of Cn induces an affine isomorphism on each projective slice. This is false. In the 1D analogue, the linear map L(x,t)=(2x,x+t) preserves the cone C1={(x,t):0<x<t}, but the induced projectivization on the chart t=1 is x↦2x/(1+x), a non-affine fractional linear map that is nonetheless a Hilbert isometry of (0,1). Therefore the application of Lemma 3 to the slice maps is not justified, and the rest of the proof—in particular the conclusion C=0—does not follow. Since Theorem 4 is the bridge from Corollary 1 to Theorem 5, the claim that the only isometries are O(n)-conjugation and complement is unsupported. The authors need to exclude 'matrix Möbius' automorphisms for n≥2 by a genuinely different argument (e.g., by showing that a projective automorphism of VPM(n) must preserve the boundary hypersurfac","section":"§5.2.1, proof of Theorem 4"},{"comment":"The slice restriction is also not set up correctly. The chart identifying VPM(n) with the slice t=t0 of Cn is X↦(t0X,t0), not X↦(X,t0). With the latter, π∘L(·,t0)∘ιt0 need not map VPM(n) into itself, so Lemma 3 cannot be applied even if that map happened to be affine. The proof should fix the chart t=1 and work with the homogeneous scaling consistently.","section":"§5.2.1, proof of Theorem 4 (slice chart)"},{"comment":"The proof of Lemma 3 does not establish the claimed dichotomy. It shows that no interior point can be mapped to 0, but that does not imply L(0) and L(I) must be the two corners. The later step 'substituting Y=0 ... and Y=I' uses points in the closure, not in the open set VPM(n), and requires a limiting argument that is not supplied. This can likely be repaired by working with the closure, but as written the lemma is incomplete.","section":"Lemma 3"}],"minor_comments":[{"comment":"In the statement, 'minimal and maximal eigenvalues of the A^{-1}B matrix' should read B^{-1}A (the proof uses B^{-1}A).","section":"Theorem 1"},{"comment":"In the computation of m, the sentence 'taking the infimum' should be 'taking the supremum', since m is defined by a supremum.","section":"Theorem 1 proof"},{"comment":"VPM(n) is used for both the open manifold (Definition 7) and the closed model from [16] (Remark 2). Please adopt a consistent notation such as VPM and overline.","section":"Definition 7 / Remark 2"},{"comment":"'L is onto VPM(n)' should read 'onto VPM(n)' (closure); the argument with Y=0 and Y=I is a limiting argument.","section":"Lemma 3 proof"},{"comment":"The phrase 'the infimum is achieved by letting Z→...' is imprecise, since the extremal Z lies in the closure of VPM(n); write 'attained in the limit'.","section":"Proposition 10"},{"comment":"The final sentence 'We leave these extensions ... for the next revision of this manuscript' is informal; replace with a standard future-work statement.","section":"§6"},{"comment":"The map X↦I−X is called 'inversion'; consider using 'complement' to avoid confusion with matrix inversion.","section":"Theorem 5"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the main gap is real and is confirmed by the 1D counterexample. The Hilbert distance formula and invariance results appear sound, but the isometry classification is not proven as written. I would not reject outright, because the missing argument may be repairable for n≥2, but the authors must either supply a correct proof of Theorem 4 or retract the uniqueness claim in Theorem 5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the closed-form Hilbert metric on VPM(n) in Theorem 1 checks out. Recomputing from the Birkhoff cone construction gives the same symmetric, positive, non-degenerate formula, and it matches the known interval and diagonal cases. That is a genuinely new and useful result, and the extreme-eigenvalue cost is nice. The invariance under X ↦ I−X and orthogonal conjugation are also correct. The motivation from extended Gaussians is sensible; the paper engages the relevant Hilbert-geometry literature.\n\nThe soft spot is Theorem 4, which is the bridge to the isometry classification. The proof asserts that a linear cone automorphism induces, on each projective slice, an affine isomorphism of Sym(n). That does not follow from mapping rays to rays. The 1D example L(x,t)=(2x,x+t) preserves the cone over (0,1) but induces x↦2x/(1+x) on the interval, which is fractional linear, not affine. The paper does assume n≥2 in Section 5, so the 1D example is not a counterexample to the theorem statement, but it is a counterexample to the inference, and no n≥2 argument replaces it. Since Theorem 5 uses Theorem 4 to conclude that only O(n)-conjugation and complement are isometries, that classification is not proven as written. I don't see a way to patch it in a sentence; it needs either a real argument showing C=0 in the linear automorphism for n≥2, or a separate treatment of possible matrix Möbius maps.\n\nThere are also smaller issues: the quoted Walsh theorem (Theorem 2) should be checked against the source, because its statement as written seems to conflict with the inversion isometry on the projectivized PSD cone for n≥3. The paper only uses the non-symmetric branch, so this may be harmless, but it's worth a look. The internal numbering is off: §1.3 cites 'Theorem 4' for the distance result that is Theorem 1, and there are A^{-1}B vs B^{-1}A slips in Theorem 1's text.\n\nBottom line: Theorem 1 is solid and worth having; the isometry classification is an unfinished proof, not a false result as far as I can tell. The paper deserves a serious referee. I'd recommend major revision (or conditional acceptance) with the isometry part either fixed or clearly labelled as open. I'd cite Theorem 1 in my own work if the need comes up.","headline":"The paper gets the Hilbert distance formula right, but the isometry classification rests on an unproven affine-reduction step.","tokens_in":16109,"tokens_out":8423,"would_cite":true,"duration_ms":88104,"reading_group":"yes","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 proves that the extended Gaussian parameter space, viewed as a symmetric positive-definite bicone, carries an explicit Hilbert metric whose distances are read from four extreme eigenvalues, and shows its isometries are exactly ort","keywords":["Hilbert geometry","extended Gaussian distributions","variance-precision model","symmetric positive semi-definite bicone","Birkhoff projective distance","matrix Möbius transformations","isometry classification","convex cones"],"falsifier":"Compute the Hilbert distance in VPM(2) directly from boundary intersections for a pair of noncommuting matrices and compare with the four-eigenvalue formula; a mismatch would refute Theorem 1. For the isometry claim, test the one-dimensional analog: on (0,1), x ↦ 2x/(1+x) preserves the cone over (0,1) and is a Hilbert isometry but is not affine; seek an analogous non-affine projective map preserving VPM(n) for n = 2. Existence of such a map in Isom(VPM(2)) outside the group generated by orthogonal conjugation and X ↦ I−X would refute Theorem 5.","tokens_in":14992,"feed_emoji":"📐","tokens_out":7405,"duration_ms":77438,"temperature":0.7,"pith_summary":"The paper studies the open bounded convex set VPM(n) = {0 ≺ X ≺ I} of symmetric matrices, which parameterizes the extended Gaussian family: ordinary Gaussians with covariance Σ correspond to the interior point X = Σ(I+Σ)^{-1}, and degenerate covariance or precision matrices appear on the boundary. The central result is an explicit Hilbert distance formula: for A, B in VPM(n), dH(A,B) is the logarithm of the ratio of the larger of two maximal eigenvalues to the smaller of two minimal eigenvalues. Because the formula uses only extreme eigenvalues, distances are computable in closed form. The paper also proves that for n > 1 the only distance-preserving transformations are the complement map X ↦ I-X and conjugation by orthogonal matrices X ↦ UᵀXU.","feed_headline":"Four eigenvalues fix the distance on extended Gaussians","feed_subtitle":"A closed-form metric on extended Gaussians; its only isometries are orthogonal conjugation and the complement map.","key_machinery":"The load-bearing object is the cone C_n = {(tX, t) : X ∈ VPM(n), t > 0} over the bicone. Hilbert distance on VPM is Birkhoff distance between rays of this cone; the paper computes the Birkhoff constants by solving Loewner-order inequalities, which collapse to ratios of extreme eigenvalues. For isometries, the paper shows C_n is not symmetric for n > 1, invokes the result that Hilbert isometries coincide with collineations in that case, and uses projective geometry to classify those collineations as linear cone automorphisms; a classification of symmetric-positive-definite-cone automorphisms then forces the orthogonal form.","core_discovery":"On the open bicone VPM(n) = {X ∈ Sym(n) : 0 ≺ X ≺ I}, the Hilbert distance between A and B is log( max( λmax(B^{-1}A), λmax((I−B)^{-1}(I−A)) ) / min( λmin(B^{-1}A), λmin((I−B)^{-1}(I−A)) ) ). The authors derive this by viewing Hilbert distance as Birkhoff projective distance on rays of the cone over VPM, where the required bounds reduce to Loewner-order inequalities A ⪯ λB and I−A ⪯ λ(I−B). They further prove that for n > 1 the isometry group is generated by X ↦ UᵀXU with U orthogonal and X ↦ I−X. Degenerate covariance and precision Gaussians lie on the boundary; the paper introduces an enlarged bicone VPM_epsilon that makes distances between such boundary points finite.","pith_inferences":["Because only extreme eigenvalues are needed, the Hilbert VPM distance may be evaluated with a single extreme-eigenvalue routine per ratio matrix; benchmarking it against AIRM in high-dimensional Gaussian inference would be a natural next test.","The complement map X ↦ I−X swaps the two boundary cones, so the metric treats degenerate covariance and degenerate precision symmetrically; this may yield a natural loss for parameter estimation when both kinds of singular Gaussians are present.","The paper's enlarged bicone VPM_epsilon could define a practical 'nearly degenerate' metric with a well-posed optimization landscape over the full closed extended Gaussian family, though the paper leaves that application open."],"forward_implications":["The Hilbert VPM distance needs only four extreme eigenvalues, so it is cheaper to compute than the affine-invariant Riemannian metric, which requires the full spectrum.","Boundary points corresponding to degenerate covariance or precision matrices have infinite Hilbert distance, but the enlarged bicone VPM_epsilon assigns them finite distances and lower-bounds the original distance.","Because straight lines are geodesics in Hilbert geometry, computational geometry primitives such as smallest enclosing balls and Voronoi diagrams transfer directly to the extended Gaussian parameter space.","The isometry invariance under orthogonal conjugation matches the invariance used in diffusion-tensor style processing, while the complement map I−X mirrors the covariance-precision duality.","The paper shows the Hilbert VPM distance is invariant under the two named transformations and, for n > 1, that these generate all its isometries."],"supporting_citations":[{"why":"Defines the variance-information manifold and the reparameterization L = Σ(I+Σ)^{-1} that turns the extended Gaussian family into the bicone VPM(n).","marker":"[16]"},{"why":"Supplies the Birkhoff characterization of Hilbert distance through cone rays, used in the proof of Theorem 1 to convert the distance into a ratio of Loewner-order bounds.","marker":"[17]"},{"why":"Provides the theorem that Hilbert isometries coincide with collineations for symmetric non-Lorentzian cones, the key step in the isometry classification.","marker":"[35]"},{"why":"Classifies linear automorphisms of the symmetric positive-definite cone as conjugations, used to force the orthogonal form in Theorem 4.","marker":"[14]"},{"why":"Gives the local-to-global projective extension lemma that lets the paper move from collineations on VPM(n) to linear automorphisms of the cone C_n.","marker":"[29]"},{"why":"Provides the basic implication that collineations of an open bounded convex domain are Hilbert isometries, used at the start of the isometry argument.","marker":"[13]"}],"fun_headline_variants":["Closed-form Hilbert distance for extended Gaussians","Four eigenvalue ratios fix the distance","Isometry group of extended Gaussian distances: conjugation and complement","Hilbert metric on degenerate Gaussians: explicit formula","Extended Gaussian distances from four eigenvalues"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The proof that there are no other isometries rests on the unstated claim that every projective automorphism of the cone over VPM induces an affine isomorphism on each horizontal matrix slice; if non-affine fractional-linear maps occur for n ≥ 2, the isometry classification could be missing a family, as happens already in dimension one.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form Hilbert distance for extended Gaussians","Four eigenvalue ratios fix the distance","Isometry group of extended Gaussian distances: conjugation and complement","Hilbert metric on degenerate Gaussians: explicit formula","Extended Gaussian distances from four eigenvalues"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000714,"raw_usage":{"total_tokens":3025,"prompt_tokens":698,"completion_tokens":2327,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":2259}},"tokens_in":442,"tokens_out":2327,"duration_ms":18493,"temperature":1.0,"reasoning_tokens":2259,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:44:41.469010+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the Hilbert distance in VPM(2) directly from boundary intersections for a pair of noncommuting matrices and compare with the four-eigenvalue formula; a mismatch would refute Theorem 1. For the isometry claim, test the one-dimensional analog: on (0,1), x ↦ 2x/(1+x) preserves the cone over (0,1) and is a Hilbert isometry but is not affine; seek an analogous non-affine projective map preserving VPM(n) for n = 2. Existence of such a map in Isom(VPM(2)) outside the group generated by orthogonal conjugation and X ↦ I−X would refute Theorem 5.","supporting_citations":[{"cited_title":"The variance information manifold and the functions on it","cited_arxiv_id":null,"evidence_quote":"Defines the variance-information manifold and the reparameterization L = Σ(I+Σ)^{-1} that turns the extended Gaussian family into the bicone VPM(n)."},{"cited_title":"Birkhoff’s version of Hilbert’s metric and its applications in analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the Birkhoff characterization of Hilbert distance through cone rays, used in the proof of Theorem 1 to convert the distance into a ratio of Loewner-order bounds."},{"cited_title":"Gauge-reversing maps on cones, and Hilbert and Thompson isometries","cited_arxiv_id":null,"evidence_quote":"Provides the theorem that Hilbert isometries coincide with collineations for symmetric non-Lorentzian cones, the key step in the isometry classification."},{"cited_title":"Seetharama Gowda, Roman Sznajder, and Jiyuan Tao","cited_arxiv_id":null,"evidence_quote":"Classifies linear automorphisms of the symmetric positive-definite cone as conjugations, used to force the orthogonal form in Theorem 4."},{"cited_title":"Synthetic projective geometry and Poincar´ e’s theorem on automorphisms of the ball","cited_arxiv_id":null,"evidence_quote":"Gives the local-to-global projective extension lemma that lets the paper move from collineations on VPM(n) to linear automorphisms of the cone C_n."},{"cited_title":"Geometric structures on manifolds, volume 227","cited_arxiv_id":null,"evidence_quote":"Provides the basic implication that collineations of an open bounded convex domain are Hilbert isometries, used at the start of the isometry argument."}],"review_version":1}