{"id":"588966c0-e0b6-4bd8-a42f-f89dad51887c","arxiv_id":"2610.37526","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The intrinsic Fisher completion of the full-covariance wrapped Gaussian family is classified by primitive frequency subgroups; a covariance floor restores compactness and near-parametric uniform estimation.","lead":"This paper determines exactly which limiting probability laws complete the wrapped Gaussian model on a torus under the intrinsic Fisher metric, and shows whether those limits can be estimated uniformly. It finds a sharp split: unrestricted covariance creates singular, non-estimable limit points in dimension at least two, while a fixed covariance floor restores compactness and near-parametric estimation rates.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"No flaw found in Proposition 3.5; the load-bearing gap is the abstract's unproved compactness–uniform-consistency equivalence, which a short packing argument can supply.","rationale":"Proposition 3.5 is the technical heart of the classification, and I traced the D/alpha-uniformity carefully. The density tail is controlled by K_B, which is independent of the shift Dk; the connector derivatives are O(e^{-t ell(S)/2}); and the lower bound uses only the lattice minimum of S and a rounding argument, so unbounded cross-covariance blocks do not break the estimate. The reader's weakest assumption therefore does not land as a mathematical objection.\n\nThe real discrepancy is between the abstract and the body. The abstract asserts that compactness of the completed closure is equivalent to uniform consistency, while Section 6.2 explicitly leaves the reverse direction open. This is a load-bearing presentation gap because the paper's headline claim includes that equivalence. However, the missing direction appears to be true by a short metric argument: failure of (6.10) gives a noncompact closure, hence a non-totally-bounded set, hence an infinite epsilon-separated sequence in d_W; the completed density map is 1-Lipschitz, so the same sequence is epsilon-separated in L1; and any family with an infinite epsilon-separated set of densities has minimax L1 risk bounded below by epsilon/2 at every sample size. If added to the paper, this would justify the abstract rather than force its weakening.\n\nFor that reason I do not move the verdict: CONDITIONAL remains appropriate, but the condition should be to add the missing equivalence proof (or explicitly qualify the abstract). I agree with the reader's overall cautious assessment but not with the identification of Proposition 3.5 as the weakest assumption.","tokens_in":34578,"tokens_out":26205,"duration_ms":548052,"concrete_test":"Write out the missing converse of the abstract equivalence. Assume (6.10) fails, so by Proposition 6.5 the d_W-closure of F is noncompact; greedily select P_1, P_2, ... in F with d_W(P_i, P_j) >= epsilon for some epsilon > 0. Since ||F_P - F_Q||_1 <= d_W(P, Q) by Lemma 3.2, these densities are epsilon-separated in L1. Then apply the finite-hypothesis packing bound: for every estimator output f, at most one P_k satisfies ||f - F_{P_k}||_1 < epsilon/2, so sup_k ||f - F_{P_k}||_1 >= epsilon/2, giving a constant minimax lower bound independent of n. If this derivation is valid, replace Section 6.2's open question with the proof; if any step fails, qualify the abstract.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption targets Proposition 3.5, but the estimate survives scrutiny. Endpoint identification only requires connector lengths to tend to zero for fixed S, D, alpha, and the distance bound's independence of D and alpha is supplied by the shift-invariant sums K_B = sup_x sum_l e^{-(l+x)^T B(l+x)/2}, which absorb arbitrary shifts Dk; the lower bound uses rounding of -Dk to a lattice point. No failure mode from unbounded D or alpha is visible in the proof.\n\nThe genuine gap is the abstract's assertion that compactness of the completed closure is equivalent to uniform consistency. The body proves compactness iff uniform moment decay (Proposition 6.5) and compactness implies uniform consistency (Corollary 6.7), but Section 6.2 explicitly says: \"Whether every subfamily violating (6.10) forbids uniform estimation is not settled here.\" This leaves the abstract's equivalence unsupported. It is not false: if (6.10) fails, Proposition 6.5 gives a noncompact d_W-closure, hence an infinite epsilon-separated subset in d_W; since the completed density map is 1-Lipschitz (Lemma 3.2), the same subset is epsilon-separated in L1, and the standard packing test gives a minimax lower bound of order epsilon for every sample size. Thus the missing direction is true and elementary, but it must be written into the paper before the abstract claim is justified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper studies the intrinsic Fisher metric completion of the full-covariance wrapped Gaussian family on a fixed torus. It characterizes the completion as the total-variation closure of the finite family, with points indexed by a primitive subgroup L of integer frequencies, a character χ, and a positive definite quadratic form Q (Theorems 1.1 and 3.7). It further proves that every added point in dimension d≥2 is non-locally compact, that every small completed-Fisher ball around an added point has minimax L1 risk of order the ball radius for every sample size (Corollary 6.1), and that a fixed covariance floor Σ⪰εI makes the closure compact and yields L1 risk of order n^{-1/2}{log(n+1)}^{d/4} with model-valued estimators uniformly consistent in Fisher distance (Theorem 6.2). The paper also establishes contractible small balls, explicit minimizing geodesics through added points, and a compactness criterion for subfamilies in terms of uniform Fourier-moment decay (Proposition 6.5).","tokens_in":34954,"tokens_out":23579,"duration_ms":422678,"significance":"This is a substantial contribution if the proofs are correct. The main load-bearing estimates are treated in detail: Proposition 3.5 supplies length and distance bounds for covariance-divergent paths with arbitrary cross-covariance, and Theorem 3.7 converts moment convergence into intrinsic endpoint convergence. I checked the point flagged in the stress-test note and do not find a gap there: the shifted lattice sums K_B and K_B^{(1)} absorb arbitrary shifts, the connector arguments are fixed-(S,D,α) and need no uniformity, and the lower bound in (3.6) uses rounding of -Dk. The minimax obstruction is not just packing: equation (6.3) gives an explicit chi-squared identity independent of the shear index, which is what yields the every-n lower bound. Strengths include the exact classification, the explicit finite-cutoff risk bound (6.9), and the ancillary verification scripts. My main concern is an overclaim in the abstract about equivalence between compactness and uniform consistency; the body leaves that converse open.","major_comments":[{"comment":"The abstract states that compactness of the completed closure is equivalent to uniform consistency in L1 or completed-Fisher loss, and equivalently to uniform decay of the Fourier moments. The body proves (6.10) as equivalent to compact closure (Proposition 6.5) and compact closure as implying uniform consistency (Corollary 6.7), but Section 6.2 explicitly says: 'Whether every subfamily violating (6.10) forbids uniform estimation is not settled here.' The missing converse is also not a direct packing consequence: the paper itself notes immediately after Corollary 6.1 that infinite packing alone does not give the quantitative, sample-size-uniform obstruction and that the likelihood-ratio control in (6.3) is needed. Thus the abstract overstates the result. Please prove the missing implication or weaken the abstract's wording to match Section 6.2, for example by stating compactness as a sufficient condition for uniform consistency and reserving the equivalence for the covariance-floor class M_d(ε).","section":"Abstract; Section 6.2","authors_acknowledge":"The logarithmic gap is not resolved here. 6.2.The exact compactness criterion.The spectral floor is one way to exclude the shear mechanism.","authors_acknowledge_p":0.67}],"minor_comments":[{"comment":"The symbol 'KB2' in the derivative bound is ambiguous; it should be defined explicitly, for example as a polynomial-weighted variant of K_B or as K_B multiplied by a universal constant.","section":"Proposition 3.5, proof","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"It would help the reader if the proof stated explicitly that the uniform-in-k control of the likelihood ratios in (6.3) is what allows J to be sent to infinity at fixed R and n; this makes the role of the chi-squared identity clear.","section":"Corollary 6.1, proof","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"In the residual-rank construction, the constant B_11 should be explicitly identified as the first diagonal entry of the residual covariance in the adapted coordinate system used for the shear x_1 - k y_1.","section":"Theorem 5.2, residual-rank case","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The Monte Carlo illustration is clearly labelled as an illustration with one estimator and finite subfamilies; it would be useful to add one sentence stating that the fitted slope is not used as evidence for the analytic rate.","section":"Section 6.3","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Read the whole thing; the body is better than the abstract.\n\nWhat is actually new: the intrinsic Fisher completion of the full wrapped Gaussian family is classified by primitive frequency subgroups, characters, and positive quadratic forms, with the extended density map a homeomorphism onto the total-variation closure (Theorems 1.1 and 3.7). That does not reduce to Csiszár–Mátus or to the Fisher–Rao geometry papers. The integer-shear packing (Theorem 5.2, Corollary 5.3) yields local noncompactness and infinite Assouad dimension at every added point, and the exact χ² identity (6.3) converts that into a minimax lower bound that holds at every sample size; the likelihood-ratio control is doing real work there. The spectral-floor regularization (Theorem 6.2), the compactness criterion (Proposition 6.5), the sharp ray asymptotics (Proposition 3.6), and the geodesic nonattainment (Theorem 4.8) are detailed and, as far as I can tell, correct. The citation pattern is standard and appropriate. The ancillary scripts carry an honest caveat: they are not interval certificates, and no proof depends on them.\n\nThe reader's flagged weak spot, Proposition 3.5, survives. Endpoint identification only requires connector lengths to vanish for fixed S, D, α, and the uniformity in D and α comes from the shift-invariant sums K_B absorbing the Dk shifts, with the rounding argument for the lower bound. I do not see the failure mode the reader worried about.\n\nThe real problem is the abstract's equivalence between compactness of the closure and uniform consistency. The body proves compactness iff uniform moment decay (Proposition 6.5) and compactness implies uniform consistency (Corollary 6.7), but Section 6.2 states in so many words that the reverse direction is not settled. The stress-test note's proposed one-paragraph fix does not close the gap. Its transfer step has the Lipschitz direction wrong: Lemma 3.2 gives L1 ≤ d_W, so d_W-separation does not imply L1-separation (the homeomorphism does give an L1-separated set, but that is beside the point). More importantly, a minimax lower bound valid for every n needs uniform control of the likelihood ratios against a common reference, which is exactly what (6.3) supplies for the shears and what noncompactness alone does not. I suspect the reverse direction is false: a thin family such as {(0, σ²I) : σ → 0} violates (6.10) (h(n) = 0 for every n) but admits a uniformly consistent plug-in estimator, since the scaled sample variance has relative error O(1/√n) uniformly in σ. If that is right, the abstract states as proved an equivalence that is false, and the hedged sentence in Section 6.2 is the correct position.\n\nWho this is for: directional statistics, information geometry, and statistical decision theory. It deserves a serious referee. The required fix is localized: qualify the abstract and the framing of Theorem 1.2(2) so they assert what the body proves, and leave the reverse direction open unless someone settles it.","headline":"The completion classification and the every-n minimax obstruction are real, well-proved contributions; the abstract's compactness–consistency equivalence is unsupported, and the proposed quick fix for it is invalid.","tokens_in":35465,"tokens_out":26325,"duration_ms":184113},"pith_extraction":{"msc":["53B12","62B11","62H11","62C20","53C23","11H06","60E05"],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Wrapped Gaussian laws on a torus are completed by limits indexed by integer frequency subgroups, and a covariance floor is what makes those limits uniformly estimable.","keywords":["Fisher information","wrapped normal distribution","directional statistics","metric completion","covariance regularization","local compactness","Assouad dimension","minimax risk"],"falsifier":"A direct numerical check of Proposition 3.5 would settle the classification: take the covariance block form of equation (3.5) with fixed $S, B$ and a sequence of cross-covariance blocks $D_j$ of growing norm, and measure the intrinsic Fisher distance $d_W(P_j, E(\\beta, B))$; the proposition requires it to stay at order $e^{-\\ell(S)/2}$ uniformly in the $D_j$, so a ratio growing with $\\|D_j\\|$ would falsify the completion theorem. A second check targets Corollary 6.1: on a fixed Fisher ball around an added law in dimension two, run any estimator on the shear family $\\Sigma_k = a(1+k^2, -k; -k, 1)$; the prediction is positive minimax risk proportional to the ball radius at every sample size, so an estimator with error tending to zero on that fixed ball would disprove the obstruction.","tokens_in":36045,"feed_emoji":"📐","tokens_out":15625,"duration_ms":106879,"temperature":0.7,"pith_summary":"The paper asks which probability laws complete the wrapped Gaussian family on a fixed torus when distance is measured by the intrinsic Fisher metric, and what that completion means for estimation. It establishes that the metric completion is homeomorphic to the total-variation closure of the finite family: every added law is a product of a uniform law and a lower-dimensional wrapped Gaussian, indexed by a primitive subgroup $L$ of integer frequencies, a character $\\chi$ for their limiting phases, and a positive definite quadratic form $Q$ for their limiting variances. In dimension at least two, every Fisher ball around every added law contains infinitely many laws separated by a fixed relative scale, so no added point has a compact neighborhood and the local minimax $\\mathrm{L}^1$ risk is at least a constant times the ball radius at every sample size. A fixed lower bound on covariance, with no upper bound, restores compactness and uniform estimation at rate $n^{-1/2}\\{\\log(n+1)\\}^{d/4}$. The mechanism is concrete: the obstruction comes from Fourier mass of fixed size migrating to arbitrarily high integer frequencies, which a spectral floor forbids through a common Gaussian envelope.","feed_headline":"Covariance floor restores uniform estimation on the torus","feed_subtitle":"Without one, every Fisher ball around an added law has minimax risk proportional to its radius at any sample size.","key_machinery":"The argument turns on four objects. (1) The survivor triple $(L, \\chi, Q)$: a primitive subgroup $L$ of integer frequencies that stay nonuniform in the limit, a character $\\chi$ giving their limiting phases, and a positive definite quadratic form $Q$ giving their limiting variances; the limiting density is a smoothed sum of the moments $\\chi(n)e^{-Q(n)/2}$ on $L$. (2) The lattice minimum $\\ell(S)=\\min_{k\\neq 0} k^\\top S k$ of the relevant Schur complement: $\\ell(S_j)\\to\\infty$ is exactly the convergence criterion, distances to partial-uniform endpoints are comparable to $e^{-\\ell(S)/2}$, and along fixed covariance rays the distance has sharp leading coefficient $\\sqrt{N(S)}\\,e^{-t\\ell(S)/2}$ determined by the shortest lattice shell. (3) Integer shears $\\Sigma_k = a\\, T_k^{-1} T_k^{-\\top}$ with $T_k = I + k\\, e_1 e_2^\\top$: these are Fisher isometries of the torus fixing the endpoint, so every shear law sits at the same distance from uniformity while the frequency $(1,k)$ separates it from all others by a fixed amount, which is the engine of local noncompactness, infinite Assouad dimension, and the minimax obstruction. (4) The uniform length estimates of Proposition 3.5: the divergence curve with covariance $\\mathrm{diag}(tS, B)$ reaches its endpoint with Fisher length and distance $O(e^{-\\ell(S)/2})$, uniformly in the cross-covariance block and the mean, which is what turns total-variation limits into genuine metric endpoints. Heat contraction $\\Phi_t(\\mu, \\Sigma) = (\\mu, \\Sigma + tI)$ then supplies contractible small balls and a global deformation retraction onto the uniform law.","core_discovery":"The central claim is that the intrinsic Fisher completion of the full-covariance wrapped Gaussian family equals its total-variation closure, point for point. Each completion point is classified by a triple $(L, \\chi, Q)$: $L$ is the primitive subgroup of integer frequencies $n$ whose angles $n\\cdot Y$ remain nonuniform in the limit, $\\chi(n)$ is the limiting phase of that angle, and $Q(n)$ its limiting variance; the limiting density has moments $\\chi(n)e^{-Q(n)/2}$ on $L$ and zero off $L$, which in adapted integer coordinates is exactly $U^{d-r} \\otimes \\mathrm{WN}^r(\\beta, B)$, a uniform factor times a lower-dimensional wrapped Gaussian. All these densities are positive and real analytic, and a finite-parameter sequence converges to such a point precisely when the quadratic forms $n^\\top \\Sigma_j n$ converge to $Q(n)$ on $L$ and diverge uniformly off $L$. The statistical payload is a dichotomy. For $d \\geq 2$, every fixed Fisher ball around an added law has minimax $\\mathrm{L}^1$ risk of order the ball radius at every sample size, while on the spectral-floor family $\\Sigma \\succeq \\varepsilon I$, with no upper covariance bound, the completed closure is compact and uniformly estimable at rate $n^{-1/2}\\{\\log(n+1)\\}^{d/4}$. Compactness of a subfamily's closure is equivalent to uniform decay of its Fourier moments.","pith_inferences":["Editorial extension: the compactness criterion is a template for other non-closed parametric families: uniform consistency over a family should be diagnosed by uniform decay of its Fourier or moment coefficients, and the integer-shear construction shows the precise way a family can evade that while every individual parameter stays estimable.","Editorial extension: the logarithmic factor in the spectral-floor rate is left open by the paper, and the natural refinement is an estimator whose Fourier cutoff adapts to the $\\varepsilon$-envelope of the frequency lattice; if such an estimator attains order $n^{-1/2}$ with constants depending only on $\\varepsilon$ and $d$, the gap in Theorem 6.2 would close.","Editorial extension: quotienting the model by integer torus automorphisms would identify the shear laws and remove the source of the noncompactness, since the obstruction is tied to a fixed coordinate frame; the paper notes that quotient as a different problem, and its completion is a natural target for the same machinery."],"forward_implications":["In dimension $d \\geq 2$, no estimator, even one allowed to output arbitrary densities, is uniformly consistent on any fixed Fisher ball around an added law; the minimax $\\mathrm{L}^1$ risk on such a ball is at least a fixed fraction of its radius for every sample size (Corollary 6.1).","A spectral floor $\\Sigma \\succeq \\varepsilon I$ with no upper bound on covariance makes the completed family compact and uniformly estimable, with $\\mathrm{L}^1$ risk at most $C_{\\varepsilon,d}\\, n^{-1/2}\\{\\log(n+1)\\}^{d/4}$ and model-valued estimators consistent in expected Fisher distance (Theorem 6.2).","Across all subfamilies, compact closure in the completed Fisher metric is equivalent to uniform decay of Fourier moments; the spectral floor is sufficient but not necessary, since some families whose smallest eigenvalue tends to zero still satisfy the criterion (Proposition 6.5, Example 3.1).","The completion admits explicit minimizing geodesics through every added law, and at residual rank $d-1$ these connect finite-parameter laws whose Fisher distance is not attained by any curve inside the finite family; compatible transverse pieces meet at right angles asymptotically (Theorem 4.8, Corollary 4.9).","In dimension one the completion adds only the uniform law and the Fisher metric extends smoothly and positively across it, so the statistical obstruction is intrinsically higher-dimensional."],"supporting_citations":[{"why":"Supplies the lattice Gaussian-sum estimates behind Lemma 3.4 and the uniform length bounds of Proposition 3.5.","marker":"[MR07]"},{"why":"Provides the orthogonal-series bias-variance argument used for the spectral-floor estimator and the rate in Theorem 6.2.","marker":"[Tsy09]"},{"why":"Supplies the minimum-distance density selection used in Lemma 6.6 to convert compact closure into uniform consistency.","marker":"[Yat85]"},{"why":"The classical non-existence result that frames the local minimax obstruction of Corollary 6.1.","marker":"[BS56]"},{"why":"Provides the convex normal neighborhood theorem used in Lemma 4.7 for the circular-disk geodesic statement.","marker":"[Lee18]"},{"why":"Gives the strong data-processing inequality behind the strict heat contraction of Theorem 4.1.","marker":"[PW17]"},{"why":"Supplies the square-root Fisher-Rao chord used as the lower bound in the sharp distance asymptotics of Proposition 3.6.","marker":"[BM19]"}],"fun_headline_variants":["Torus wrapped Gaussian completion equals total-variation closure","A covariance floor enables uniform estimation on the torus","Every added torus law is a uniform law times a wrapped Gaussian","Without a covariance floor, Fisher balls are not uniformly estimable","Fourier moment decay controls compactness of wrapped Gaussian closures"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The load-bearing premise is the uniform length estimate of Proposition 3.5: every covariance-divergence curve with blocks $\\mathrm{diag}(tS, B)$ reaches its partial-uniform endpoint at Fisher distance $O(e^{-\\ell(S)/2})$, uniformly in the cross-covariance block and the mean, and if that uniformity failed for unbounded cross-covariance, some total-variation limits would not be genuine metric endpoints and the classification would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Torus wrapped Gaussian completion equals total-variation closure","A covariance floor enables uniform estimation on the torus","Every added torus law is a uniform law times a wrapped Gaussian","Without a covariance floor, Fisher balls are not uniformly estimable","Fourier moment decay controls compactness of wrapped Gaussian closures"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"cost":0.0001812356,"is_byok":false,"cost_details":{"upstream_inference_cost":0.0001812356,"upstream_inference_prompt_cost":2.98116e-05,"upstream_inference_completions_cost":0.000151424},"total_tokens":2002,"prompt_tokens":1170,"completion_tokens":832,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":1053,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":761}},"tokens_in":1170,"tokens_out":832,"cost_source":"openrouter_usage","duration_ms":10494,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":761,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-28T03:48:07.144750+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"A direct numerical check of Proposition 3.5 would settle the classification: take the covariance block form of equation (3.5) with fixed $S, B$ and a sequence of cross-covariance blocks $D_j$ of growing norm, and measure the intrinsic Fisher distance $d_W(P_j, E(\\beta, B))$; the proposition requires it to stay at order $e^{-\\ell(S)/2}$ uniformly in the $D_j$, so a ratio growing with $\\|D_j\\|$ would falsify the completion theorem. A second check targets Corollary 6.1: on a fixed Fisher ball around an added law in dimension two, run any estimator on the shear family $\\Sigma_k = a(1+k^2, -k; -k, 1)$; the prediction is positive minimax risk proportional to the ball radius at every sample size, so an estimator with error tending to zero on that fixed ball would disprove the obstruction.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}