Pith. sign in

REVIEW 5 minor 21 references

Nonstandard likelihood-ratio limits under semidefinite rank constraints

T0 review · 0 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper establishes that likelihood-ratio tests for the hypothesis that a positive-semidefinite matrix has rank at most r are governed, after nuisance profiling, by a single reduced Gaussian experiment: the limit is a difference of squar

desk verdict A careful, genuinely useful paper on stratified LRT calibration under PSD rank constraints; the core reduction and isotropic dominance are new and the paper is honest about its anisotropic gap. read the letter →

arxiv 2607.13761 v1 pith:UD2I3QJP submitted 2026-07-15 math.ST stat.TH

Didier Concordet This is my paper
classification math.STstat.TH MSC 62F0562E2062H1590C22
keywords likelihood-ratiotestrankconstraintpositivesemidefinitechi-bar-squarestratifiednullconetransitionleastfavourabledistribution
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper studies likelihood-ratio tests of the hypothesis that a positive-semidefinite matrix has rank at most r. Because the null set is a union of rank strata, the usual chi-bar-square calibration—correct at matrices of rank exactly r—does not by itself control the test at lower-rank matrices or along sequences where the rank changes at the n^{-1/2} scale. The paper shows that after profiling all regular nuisance parameters, every fixed null rank and every local null rank transition is described by one reduced Gaussian experiment: the limit is the difference of squared Frobenius distances to a rank-constrained semidefinite cone and to the semidefinite cone. On the top stratum this recovers the chi-bar-square law; on lower strata it is generally a nonconvex projection law. The main calibration result is that, under isotropy, the top-stratum law dominates all fixed strata and all admissible transitions, so its critical value is least favourable for the composite null; for anisotropic information the same dominance is proved only when the active corank is one.

What carries the argument

The load-bearing object is the active compression of the true matrix's kernel: for a null matrix of rank s, with kernel dimension k_s=q−s, the active block is A_{P_s}(H)=U_s^T H U_s, and the reduced covariance is S_{P_s}=A_{P_s} I_eff^{-1} A_{P_s}^*. Profiling the nuisance parameters and all off-kernel matrix directions reduces the Gaussian limit to a difference of squared distances between a whitened random matrix Y and the cones C_{P_s}=S^{-1/2}(S_+^k) and D_{P_s,m_s}=S^{-1/2}(K_m^k), where K_m^k is the nonconvex cone of PSD matrices of rank ≤m. The tangent-cone identity T={H:A_{P_s}(H)∈K_m^k} is what makes the reduction exact; it converts a semidefinite-rank problem into a finite-dimensio

What would settle it

Take an anisotropic reduced covariance with active space of dimension three and rank budget m=1 (active corank two), for instance the elliptic-cone family with S=diag(1,1,γ) in adapted coordinates; compute the 95th percentile of the fixed-stratum and transition laws Δ_{0,1}(C;Y) for several drifts C and compare with the top-stratum chi-bar-square 95th percentile. If any of those quantiles exceeds the top value, the isotropic dominance theorem does not extend to this anisotropic case.

Watch

Extended reading notes

Core claim

The central claim is Theorem 3.3: at a null matrix of rank s≤r, after nuisance profiling the likelihood-ratio statistic converges to Δ_{s,r}(P_s)=dist²_F(Y,D_{P_s,m_s})−dist²_F(Y,C_{P_s}), where Y is standard Gaussian in the active space, C_{P_s} is the whitened PSD cone, and D_{P_s,m_s} is the whitened set of PSD matrices of rank at most m_s=r−s. This formula holds for every fixed rank stratum, and with a deterministic Gaussian shift it also holds for null paths that cross a rank interface at the local scale. On the top stratum s=r the rank constraint is inactive, D reduces to {0}, and the limit becomes the classical chi-bar-square law. On lower strata the limit is a nonconvex distance-diff

Load-bearing premise

The least-favourable calibration claim assumes the reduced active covariance is proportional to the identity on every stratum; when this isotropy fails and the active corank q−r is at least two, the paper does not prove that the top-stratum law dominates lower strata and rank transitions.

Editorial extensions

If this is right

  • At a rank-r null point the limiting law is the classical chi-bar-square law, so existing top-stratum critical values remain valid there and only there.
  • At lower-rank null points the limit is a nonconvex rank-constrained projection law; using the top-stratum critical value is conservative under isotropy, while using the central lower-stratum critical value can under-cover along rank transitions.
  • Null sequences whose rank changes at the n^{-1/2} scale are covered by the same reduced experiment with a deterministic drift, so no separate asymptotic construction is needed.
  • Plug-in critical values based on a consistent estimator of the reduced covariance converge by continuity of the stratified law.
  • On the top stratum the critical value is differentiable at positive levels under the paper's regularity conditions, giving a formula for sensitivity to orientation and nuisance information.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the isotropy assumption can be relaxed, the practical payoff would be large: a single top-stratum chi-bar-square quantile would calibrate the whole composite bounded-rank test in covariance models without knowing the true rank.
  • The unresolved anisotropic active-corank ≥2 case is the natural place to look for a counterexample; a numerical search over elliptic cones with k=3,m=1 could settle whether the dominance holds more generally or fails.
  • Because lower-stratum projections are nonunique at eigenvalue ties, extending sensitivity to lower strata would require set-valued derivatives; one testable consequence is that coverage and critical values may jump at ties.
  • The reduced covariance S_{P_s} functions as a normal-form summary of the model; estimating it directly offers a plug-in calibration route for real-data applications.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

0 major / 5 minor

Summary. The paper studies likelihood-ratio tests for the positive-semidefinite rank hypothesis H0: rank(Σ) ≤ r against the unconstrained PSD alternative. Under uniform LAN and local set convergence, it develops a stratified active-block reduction: after profiling regular nuisance parameters, the limiting statistic depends only on the kernel block of the true matrix. The main reduction (Theorem 3.3) represents the limit as a distance-difference between projections onto S^{-1/2}(S^k_+) and S^{-1/2}(K^k_m). On the top stratum this recovers the classical chi-bar-square law. Under isotropy, interlacing and compression establish that the top-stratum law stochastically dominates every fixed lower stratum and every admissible local null rank transition (Theorem 3.10); for arbitrary anisotropy the same transition dominance is proved when the active corank is one (Proposition 3.12). A conditional positive-level Hadamard shape derivative for the top-stratum distribution and quantile is given in Theorem 4.3, with an explicit verification family. Numerical experiments cover nuisance profiling, lower-stratum rank transitions, anisotropic lower-stratum laws, and derivative/ascent checks. The paper is explicit about the open anisotropic corank-at-least-two interface problem and about the conditional nature of Assumption 4.2.

Significance. Assuming the results hold, this is a significant contribution to constrained likelihood-ratio inference. It provides a unified treatment of the stratified PSD rank null, identifies precisely where chi-bar-square calibration fails, and gives a least-favourable calibration result under isotropy and in the corank-one anisotropic case. The geometric proofs — tangent-cone structure, local Hausdorff convergence, interlacing/compression dominance — are careful and self-contained. The assumptions are stated openly, and the numerical validation is reproducible, with exact finite-sample likelihood calculations and no parameters fitted to the target claims. The honest scoping of the unresolved anisotropic interface and the conditional sensitivity theorem is a strength rather than a defect: the main least-favourable theorem is explicitly conditional on isotropy, and the sensitivity theorem is explicitly conditional on Assumption 4.2.

minor comments (5)
  1. [§3.6] The symbol C is used both for the cone C_{P_s} in (3.4) and for an admissible drift C in Proposition 3.9 and Theorem 3.10. This is locally clear but can be confusing; consider using a different letter (e.g., M) for the drift.
  2. [§5.4] The transition from the q=3 top-stratum example to the q=2 lower-stratum experiment is abrupt. A sentence stating explicitly that the same elliptic cone family Cγ is being reused for a different active-space problem (q=2, r=1, s=0) would help the reader.
  3. [§5.3] The sentence 'The same active-dimension-two law is the γ=1 member of the analytic chi-bar-square family below' is potentially confusing because the preceding sentence refers to the rank-one top stratum. Clarify that the top-stratum law, not the lower-stratum law, is the γ=1 member.
  4. [§5.5] In the paragraph before Table 7, 'T wo runs satisfy' should read 'Two runs satisfy the gradient tolerance directly.'
  5. [§4.2] In Assumption 4.2(2), the notation 'whenever (R̃,t)→(R,c)' is fine, but it may be worth writing 'where R and c are fixed' to avoid confusion with the evaluation point on the right-hand side of (4.5).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity

full rationale

The paper's central chain is a mathematical derivation from explicit assumptions, not a fit or a self-citation loop. Theorem 3.3 derives the stratified active-block limit from uniform LAN (Assumption 2.1), local set convergence (Assumption 3.2), and a self-contained tangent-cone proof (Theorem 3.1); the affine profiling identity (3.3) is proven by completing the square, and the Gaussian limit follows by change of variable. The top-stratum chi-bar-square law is obtained from Moreau decomposition and standard conic Steiner theory used as independent mathematical inputs. The isotropic least-favourable Theorem 3.10 is proved by Cauchy interlacing and a compression inequality, not assumed; Proposition 3.12 gives an independent supporting-hyperplane proof in the corank-one anisotropic case. The sensitivity Theorem 4.3 is explicitly conditional on Assumption 4.2, and the paper states that Assumption 4.2 is not asserted for all transformed PSD cones; this is honest scoping, not circularity. Numerical validations compare independent finite-sample likelihood ratios or Monte Carlo laws against analytically derived limits and analytic chi-bar-square weights; the Monte Carlo top-stratum value is explicitly retained only as a reproducibility diagnostic while the analytic value is used for decisions. There are no fitted parameters renamed as predictions, no load-bearing self-citations (the cited tangent-cone and variational references are external and the paper also provides its own proof), and no ansatz smuggled in via citation. The explicit open problems—anisotropic active corank at least two and lower-stratum set-valued sensitivity—are stated as unresolved, further confirming that the derived results are not being assumed through the back door.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper introduces no fitted free parameters and no invented physical or mathematical entities. Its results rest on standard LAN and tangent-cone assumptions, the explicitly stated isotropy condition for the least-favourable theorem, and the conditional level-set regularity assumption for sensitivity.

assumptions (5)
  • domain assumption Uniform LAN: sup over compact h of |ℓ_n(θ0+n^{-1/2}h)-ℓ_n(θ0)-⟨h,Z_n⟩+1/2⟨h,Ih⟩| converges to 0 in probability.
    Assumption 2.1; this is the standard local asymptotic normality required for all distributional limits in the paper.
  • domain assumption Local set convergence and stochastic localisation of constrained maximisers.
    Assumption 3.2; verified for the direct PSD rank constraint in Prop. S2.3, and retained as an explicit condition for smooth model embeddings.
  • domain assumption Reduced active covariance S_Ps is proportional to the identity on every stratum.
    This isotropy condition is the premise of Theorem 3.10, the main least-favourable calibration theorem.
  • domain assumption Assumption 4.2: uniform finiteness and continuity of level-set surface integrals and Gaussian-tail tightness.
    Needed for the Hadamard shape derivative in Theorem 4.3; the paper explicitly says it is not verified for every linearly transformed semidefinite cone.
  • domain assumption Gaussian covariance model and standard distributional assumptions for numerical experiments.
    Used in Section 5 for finite-sample simulations; these are model specifications, not claims of general validity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nonstandard likelihood-ratio limits under semidefinite rank constraints." pith.science (2026). https://pith.science/paper/UD2I3QJP

@misc{pith2026260713761,
  author       = {Pith},
  title        = {Pith review of: Nonstandard likelihood-ratio limits under semidefinite rank constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UD2I3QJP}},
  note         = {Machine review of arXiv:2607.13761}
}
abstract

We study likelihood-ratio tests for the hypothesis that a positive-semidefinite matrix has rank at most a prescribed value. The null hypothesis is stratified: points of maximal allowed rank lie on a regular boundary stratum, whereas lower-rank points are singular. Consequently, the usual chi-bar-square calibration on the top stratum does not by itself describe the whole composite null, especially along sequences whose rank changes at the local $n^{-1/2}$ scale. After profiling regular nuisance parameters, we derive a common reduced Gaussian experiment for every fixed null rank and for all admissible local rank transitions. On the top stratum, the classical chi-bar-square law is recovered. At lower ranks, the limit generally involves projection onto a nonconvex rank-constrained semidefinite set. Our main calibration result shows that, under isotropy, the top-stratum law is least favourable over all fixed null strata and all local null rank transitions. We also prove the corresponding transition dominance under arbitrary anisotropy when the active corank is one. Finally, on the top stratum, we obtain a conditional shape derivative for the limiting distribution and its critical value. Gaussian covariance models and finite-sample experiments illustrate nuisance profiling, rank transitions, anisotropy, and orientation sensitivity.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 2 linked inside Pith

  1. [1]

    barticle [author] Baey , C. C. , Courn \`e de , P. H. P. H. Kuhn , E. E. ( 2019 ). Asymptotic distribution of likelihood ratio test statistics for variance components in nonlinear mixed effects models . Comput. Statist. Data Anal. 135 107--122 . barticle

    2019

  2. [2]

    bbook [author] Bonnans , J. F. J. F. Shapiro , A. A. ( 2000 ). Perturbation Analysis of Optimization Problems . Springer , New York . bbook

    2000

  3. [3]

    barticle [author] Chan , Z. X. Z. X. Sun , D. D. ( 2008 ). Constraint nondegeneracy, strong regularity, and nonsingularity in semidefinite programming . SIAM J. Optim. 19 370--396 . barticle

    2008

  4. [4]

    barticle [author] Chen , Q. Q. Fang , Z. Z. ( 2019 ). Improved inference on the rank of a matrix . Quant. Econ. 10 1787--1824 . barticle

    2019

  5. [5]

    barticle [author] Chernoff , H. H. ( 1954 ). On the distribution of the likelihood ratio . Ann. Math. Statist. 25 573--578 . barticle

    1954

  6. [6]

    barticle [author] Drton , M. M. ( 2009 ). Likelihood ratio tests and singularities . Ann. Statist. 37 979--1012 . barticle

    2009

  7. [7]

    barticle [author] Geyer , C. J. C. J. ( 1994 ). On the asymptotics of constrained M -estimation . Ann. Statist. 22 1993--2010 . barticle

    1994

  8. [8]

    bmisc [author] Hantoute , A. A. , Henrion , R. R. P\'erez-Aros , P. P. ( 2017 ). Subdifferential characterization of probability functions under Gaussian distribution . arXiv:1705.10160 . bmisc

    arXiv 2017

Show all 21 references
  1. [9]

    barticle [author] McCoy , M. B. M. B. Tropp , J. A. J. A. ( 2014 ). From Steiner formulas for cones to concentration of intrinsic volumes . Discrete Comput. Geom. 51 926--963 . barticle

    2014

  2. [10]

    bmisc [author] Mitchell , J. D. J. D. , Allman , E. S. E. S. Rhodes , J. A. J. A. ( 2018 ). Hypothesis testing near singularities and boundaries . arXiv:1806.08458 . bmisc

    2018 arXiv

  3. [11]

    barticle [author] Moreau , J. J. J. J. ( 1962 ). D \'e composition orthogonale d'un espace hilbertien selon deux c \^o nes mutuellement polaires . C. R. Acad. Sci. Paris 255 238--240 . barticle

    1962

  4. [12]

    barticle [author] Olikier , G. G. , Mlinari \'c , P. P. , Absil , P. A. P. A. Uschmajew , A. A. ( 2026 ). The tangent cone to the real determinantal variety: Various expressions and a proof . Set-Valued Var. Anal. 34 . Article 8 . barticle

    2026

  5. [13]

    barticle [author] Robinson , S. M. S. M. ( 1980 ). Strongly regular generalized equations . Math. Oper. Res. 5 43--62 . barticle

    1980

  6. [14]

    barticle [author] Schneider , R. R. Uschmajew , A. A. ( 2015 ). Convergence results for projected line-search methods on varieties of low-rank matrices via ojasiewicz inequality . SIAM J. Optim. 25 622--646 . barticle

    2015

  7. [15]

    barticle [author] Self , S. G. S. G. Liang , K. Y. K. Y. ( 1987 ). Asymptotic properties of maximum likelihood estimators and likelihood ratio tests under nonstandard conditions . J. Amer. Statist. Assoc. 82 605--610 . barticle

    1987

  8. [16]

    barticle [author] Shapiro , A. A. ( 1985 ). Asymptotic distribution of test statistics in the analysis of moment structures under inequality constraints . Biometrika 72 133--144 . barticle

    1985

  9. [17]

    barticle [author] Shapiro , A. A. ( 2016 ). Differentiability properties of metric projections onto convex sets . J. Optim. Theory Appl. 169 953--964 . barticle

    2016

  10. [18]

    barticle [author] Shapiro , A. A. ( 2019 ). Statistical inference of semidefinite programming . Math. Program. 174 77--97 . barticle

    2019

  11. [19]

    bbook [author] Silvapulle , M. J. M. J. Sen , P. K. P. K. ( 2005 ). Constrained Statistical Inference: Inequality, Order, and Shape Restrictions . Wiley , Hoboken, NJ . bbook

    2005

  12. [20]

    barticle [author] Uryasev , S. S. ( 1994 ). Derivatives of probability functions and integrals over sets given by inequalities . J. Comput. Appl. Math. 56 197--223 . barticle

    1994

  13. [21]

    bmisc [author] Yang , Y. Y. , Gao , B. B. Yuan , Y.-x. Y.-x. ( 2025 ). Variational analysis of determinantal varieties . arXiv:2511.22613 . bmisc

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.