Pith.

Statistics Theory · arXiv:2610.37526 v1 · submitted 27 Oct 2026 · 38 references

Fisher Completions and Statistical Regularization of Wrapped Gaussian Families

Gunhee Cho, Hyun Chul Jang

The paper says

“We determine the intrinsic Fisher completion of the full-covariance wrapped Gaussian family on a fixed torus and relate its geometry to uniform estimation. The completion is homeomorphic to the total-variation closure.”

more of the abstract

Every added law is, in suitable integer torus coordinates, a product of a uniform law and a lower-dimensional wrapped Gaussian. All these densities are positive and real analytic, but in dimension at least two no added point has a compact neighborhood. Small balls remain contractible, and explicit minimizing geodesics pass through added points. In dimension at least two, every fixed Fisher neighborhood of an added point has a positive minimax risk at every sample size. For subfamilies with a common positive lower bound on the variance in every nonzero integer direction, we prove that compactness of the completed closure is equivalent to uniform consistency in either L1 or completed-Fisher loss. Equivalently, the Fourier moments must decay uniformly over the subfamily. A fixed covariance floor guarantees this condition without an upper covariance bound and gives L1 risk at most Cε,dn−1/2{log(n + 1)}d/4. A common-phase example illustrates how partially uniform limits can retain relative-phase information while remaining uniformly estimable.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 28 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. 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. This line is wrong

1 major objection · 4 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 38 references say.

The paper says, at Abstract; Section 6.2

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.67):

“The logarithmic gap is not resolved here. 6.2.The exact compactness criterion.The spectral floor is one way to exclude the shear mechanism.”

Pith says: Objection 1 of 1 · major

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.'

more of objection 1

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).

4 minor objections

The paper says, at Proposition 3.5, proof

Pith says: minor objection 1 of 4

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.

The paper says, at Corollary 6.1, proof

Pith says: minor objection 2 of 4

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.

show 2 more

The paper says, at Theorem 5.2, residual-rank case

Pith says: minor objection 3 of 4

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.

The paper says, at Section 6.3

Pith says: minor objection 4 of 4

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.

What would settle it

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.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 28 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 9 min 8 s · 34,578 tokens read, 26,205 written
  3. Editor wrote the verdict above 3 min 4 s · 35,465 tokens read, 26,325 written
  4. Referee major revision: 1 major objection, 4 minor 7 min 3 s · 34,954 tokens read, 23,579 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Read the whole thing; the body is better than the abstract.

What 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.

The 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 KB absorbing the Dk shifts, with the rounding argument for the lower bound. I do not see the failure mode the reader worried about.

The 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 ≤ dW, so dW-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.

Who 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.

The strongest objection

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.

The 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.

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.

The 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.

For 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.

The referee's report

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.

Pith's reading of the paper (a machine's summary, not the authors')

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.

What carries the argument. 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 claim. 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.

The premise it rests on. 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.

If the paper is right (Pith's inferences, not the authors')

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.

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.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

We determine the intrinsic Fisher completion of the full-covariance wrapped Gaussian family on a fixed torus and relate its geometry to uniform estimation. The completion is homeomorphic to the total-variation closure. Every added law is, in suitable integer torus coordinates, a product of a uniform law and a lower-dimensional wrapped Gaussian. All these densities are positive and real analytic, but in dimension at least two no added point has a compact neighborhood. Small balls remain contractible, and explicit minimizing geodesics pass through added points. In dimension at least two, every fixed Fisher neighborhood of an added point has a positive minimax risk at every sample size. For subfamilies with a common positive lower bound on the variance in every nonzero integer direction, we prove that compactness of the completed closure is equivalent to uniform consistency in either L1 or completed-Fisher loss. Equivalently, the Fourier moments must decay uniformly over the subfamily. A fixed covariance floor guarantees this condition without an upper covariance bound and gives L1 risk at most Cε,dn−1/2{log(n + 1)}d/4. A common-phase example illustrates how partially uniform limits can retain relative-phase information while remaining uniformly estimable.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-28

The authors' main claim, as the classifier picked it out 0.64

“We determine the intrinsic Fisher completion of the full-covariance wrapped Gaussian family on a fixed torus and relate its geometry to uniform estimation.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.77 · human participants 0.02 · controlled comparison 0.02 · randomized assignment 0.01 · states quantitative result 0.96 · compares with prior results 0.09 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
We determine the intrinsic Fisher completion of the full-covariance wrapped Gaussian family on a fixed torus and relate its geometry to uniform estimation. contribution 0.99 0.95 0.48 asserted 1.00
The completion is homeomorphic to the total-variation closure. result 1.00 0.85 0.16 asserted 1.00
Every added law is, in suitable integer torus coordinates, a product of a uniform law and a lower-dimensional wrapped Gaussian. result 1.00 0.88 0.22 asserted 1.00
All these densities are positive and real analytic, but in dimension at least two no added point has a compact neighborhood. result 0.99 0.86 0.11 asserted 1.00
Small balls remain contractible, and explicit minimizing geodesics pass through added points. result 1.00 0.86 0.06 asserted 1.00
In dimension at least two, every fixed Fisher neighborhood of an added point has a positive minimax risk at every sample size. result 0.99 0.84 0.15 asserted 1.00
For subfamilies with a common positive lower bound on the variance in every nonzero integer direction, we prove that compactness of the completed closure is equivalent to uniform consistency in either $L^1$ or completed-Fisher loss. result 0.95 0.91 0.26 asserted 1.00
Equivalently, the Fourier moments must decay uniformly over the subfamily. result 0.84 0.80 0.11 asserted 1.00
A fixed covariance floor guarantees this condition without an upper covariance bound and gives $L^1$ risk at most $C_{\varepsilon,d}n^{-1/2}\{\log(n+1)\}^{d/4}$. result 0.98 0.87 0.14 asserted 1.00
A common-phase example illustrates how partially uniform limits can retain relative-phase information while remaining uniformly estimable. interpretation 0.48 0.84 0.06 asserted 0.98

What kind of paper the classifier reports this is: theoretical result 1.00 · evidence from proof or derivation 1.00

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-28: 1 request, 7,482 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (5)
Figure 1 of the paper
Figure 1. The covariance floor and the shear mechanism. For Bk =  1+k 2 −k −k 1  , compare Σk = 8Bk with akBk, where ak = max{8, 0.5/λmin(Bk)}. The matched frequency (1, k) has moment magnitude e −ak/2 . Panel (b) divides this magnitude by its initial value e −4 . Keeping the smallest covariance eigenvalue above 0.5 forces this Fourier amplitude to vanish. The plotted values are computed from these matrices. specific instances of the added-point distance and geodesic questions without asserting existence for every pair in the nonproper completion. Context and contribution. Wrapped Ga
Figure 2 of the paper
Figure 2. One observation dimension: lengths of the coordinate circles {s}×T for s ≥ s0 = 1.2, plotted against the intrinsic distance from the circle at s0 in each model. The wrapped circle closes at the uniform law, at distance D∗ ≈ 0.7697, with limiting slope −2π; the quotient circle closes only at infinite distance. The wrapped curve is computed numerically from the Fisher coefficients and is an illustration only. 3. The full intrinsic Fisher completion Write Mcd,W for the metric completion and use dW also for its completed distance. All density norms below use normalized Haar measure ν.…
Figure 3 of the paper
Figure 3. Surviving frequencies for d = 2. Each disk is an integer frequency n = (n1, n2), shaded by |mn| = e −n ⊤Σn/2 at t = 6, b = 1 (last panel: j = 6). Rings mark the frequencies that survive in the limit t → ∞ or j → ∞; the shading is a single snapshot. The surviving set is the subgroup L of Theorem 1.1. In the last panel the smallest covariance eigenvalue tends to zero, yet every nonzero frequency disappears. 3.1. The statistical closure. Lemma 3.2 (Lipschitz observables). For a bounded complex observable φ on the torus, P 7→ EP φ is ∥φ∥∞- Lipschitz for dW. In particular, (3.2) |mn(P) −
Figure 4 of the paper
Figure 4. Diagonal and unrestricted covariance in two observation dimensions. (a) For comparison: with its own intrinsic distance, the diagonal subfamily is the Riemannian prod￾uct of two circular families, since independence makes Fisher speeds add, and each factor completes to a disk; the upward-closing surfaces are schematic, not isometric embeddings. (b) The full family contains the integer-shear sequence Pk at a common distance from U2, with the separation bound (5.3). The vertical scale is symbolic. (c) Relative densities of P0, . . . , P3 on [−π, π] 2 with opposite edges identified.…
Figure 5 of the paper
Figure 5. Worst-case Monte Carlo L 1 error of the truncated Fourier estimator, over the shear family at a common distance from U2 and over a spectral-floor family Σ ⪰ 0.5I2, for d = 2. Twelve repetitions per law and sample size; the dotted line has slope −1/2. Computed by the ancillary script simulate estimation.py. Every added law lies on a minimizing geodesic; at residual rank d−1 this gives finite endpoints for which the original manifold has no distance-attaining curve. Compatible transverse pieces have a right-angle distance asymptotic and a shorter bridge than the path through their…

Figures are the authors', reproduced from the paper.

References (38)
  1. [1] Amari, S.-I. and Nagaoka, H. , title =. 2000 , doi = on Pith
  2. [2] Bauer, M. and Le Brigant, A. and Lu, Y. and Maor, C. , title =. Calc. Var. Partial Differential Equations , volume =. 2024 , pages = on Pith
  3. [3] Bruveris, M. and Michor, P. W. , title =. Math. Nachr. , volume =. 2019 , pages = on Pith
  4. [4] Coles, S. , title =. Stat. Comput. , volume =. 1998 , pages = on Pith
  5. [5] Costa, S. I. R. and Santos, S. A. and Strapasson, J. E. , title =. Discrete Appl. Math. , volume =. 2015 , pages = on Pith
  6. [6] Csisz. Closures of exponential families , journal =. 2005 , pages = on Pith
  7. [7] Globke, W. and Quiroga-Barranco, R. , title =. Inf. Geom. , volume =. 2021 , pages = on Pith
  8. [8] Huang, Z. and Lami, L. and Wilde, M. M. , title =. PRX Quantum , volume =. 2024 , pages = on Pith
  9. [9] Ishiki, Y. , title =. J. Fractal Geom. , volume =. 2021 , pages = on Pith
  10. [10] Kurz, G. and Gilitschenski, I. and Hanebeck, U. D. , title =. Sensor Data Fusion: Trends, Solutions, Applications (SDF) , publisher =. 2014 , pages = on Pith
  11. [11] Le Brigant, A. and Preston, S. C. and Puechmorel, S. , title =. Differential Geom. Appl. , volume =. 2021 , pages = on Pith
  12. [12] Louis, T. A. , title =. J. Roy. Statist. Soc. Ser. B , volume =. 1982 , pages = on Pith
  13. [13] Mardia, K. V. and Jupp, P. E. , title =. 1999 , doi = on Pith
  14. [14] Micciancio, D. and Regev, O. , title =. SIAM J. Comput. , volume =. 2007 , pages = on Pith
  15. [15] Nodehi, A. and Golalizadeh, M. and Maadooliat, M. and Agostinelli, C. , title =. Comput. Statist. , volume =. 2021 , pages = on Pith
  16. [16] Saraceno, G. and Agostinelli, C. and Greco, L. , title =. METRON , volume =. 2021 , pages = on Pith
  17. [17] Stein, E. M. and Shakarchi, R. , title = on Pith
  18. [18] Tsybakov, A. B. , title =. 2009 , doi = on Pith
  19. [19] Ay, Nihat and Jost, J. Information Geometry , series =. 2017 , doi = on Pith
  20. [20] Jammalamadaka, S. Rao and SenGupta, Ashis , title =. 2001 , doi = on Pith
show 18 more
  1. [21] Ley, Christophe and Verdebout, Thomas , title =. 2017 , doi = on Pith
  2. [22] Conway, John H. and Sloane, Neil J. A. , title =. 1999 , doi = on Pith
  3. [23] Banaszczyk, Wojciech , title =. Mathematische Annalen , volume =. 1993 , doi = on Pith
  4. [24] Le Cam, Lucien , title =. The Annals of Statistics , volume =. 1973 , doi = on Pith
  5. [25] Bahadur, R. R. and Savage, Leonard J. , title =. The Annals of Mathematical Statistics , volume =. 1956 , doi = on Pith
  6. [26] Yu, Bin , title =. Festschrift for Lucien Le Cam: Research Papers in Probability and Statistics , editor = on Pith
  7. [27] Raginsky, Maxim , title =. IEEE Transactions on Information Theory , volume =. 2016 , doi = on Pith
  8. [28] Polyanskiy, Yury and Wu, Yihong , title =. Convexity and Concentration , editor =. 2017 , doi = on Pith
  9. [29] McCullagh, Peter , title =. The Annals of Statistics , volume =. 1996 , doi = on Pith
  10. [30] Dempster, A. P. and Laird, N. M. and Rubin, D. B. , title =. Journal of the Royal Statistical Society. Series B (Methodological) , volume = on Pith
  11. [31] Orchard, Terence and Woodbury, Max A. , title =. Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Theory of Statistics , editor = on Pith
  12. [32] Heinonen, Juha , title =. 2001 , doi = on Pith
  13. [33] Fraser, Jonathan M. , title =. 2020 , doi = on Pith
  14. [34] Mumford, David , title =. 1983 , doi = on Pith
  15. [35] Jona-Lasinio, Giovanna and Gelfand, Alan and Jona-Lasinio, Mattia , title =. The Annals of Applied Statistics , volume =. 2012 , doi = on Pith
  16. [36] Lee, John M. , title=. 2018 , doi= on Pith
  17. [37] Yatracos, Yannis G. , title =. The Annals of Statistics , volume =. 1985 , doi = on Pith
  18. [38] Devroye, Luc and Lugosi, G. Combinatorial Methods in Density Estimation , series =. 2001 , doi = on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

  1. Reader · 2026-10-30

    It’s very….pithy.

Sign in to comment