Pith. sign in

REVIEW 2 major objections 4 minor 17 references

Overlapping clusters in functional data are recoverable from covariance operators alone, up to label switching, when each latent cluster has two observable anchor variables.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

funFMC recovers overlapping cluster memberships in multivariate functional data with provable identifiability, consistency, and a central limit theorem for inference.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Solid identifiability and clustering theory, but Theorem 5's covariance formula is asymmetric as written; the CLT needs a fix before the inferential claims can be trusted. the 2 major comments →

arxiv 2607.13197 v1 pith:INLGSUIE submitted 2026-07-14 stat.ME math.STstat.TH

funFMC: Overlapping Clustering for Functional Data

Abhiti Mishra (1) , Kean Ming Tan (1) , Tailen Hsing (1) ((1) Department of Statistics , University of Michigan) This is my paper
classification stat.ME math.STstat.TH MSC 62H3062R1062F1262H25
keywords overlapping clusteringfunctional datalatent factor modelidentifiabilitypure variablesoperator normHilbert-Schmidtcentral limit theorem
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper claims that overlapping cluster memberships in multivariate functional data can be recovered from the covariance structure of the observed curves, and that the recovery comes with frequentist guarantees. The authors propose a latent factor model in which a real-valued loading matrix encodes which of several latent functional factors each observed curve depends on, so one curve can belong to several clusters at once. They prove that, provided each factor is anchored by at least two 'pure' variables that load only on that factor, the loading matrix is identifiable up to signed permutation, meaning the cluster structure is recoverable up to label switching. Their two-step estimation procedure first finds pure variables by comparing operator norms of cross-covariances, then solves an operator regression using Hilbert-Schmidt inner products to estimate the overlap weights of the remaining variables. The paper further establishes a central limit theorem for those overlap weights, enabling hypothesis tests on cluster membership; simulations and an fMRI application support the method. If true, this is the first frequentist inferential framework for overlapping functional clustering.

Core claim

The central discovery is that the infinite-dimensional covariance operator of a multivariate functional variable carries enough information to identify both the number of overlapping clusters and the membership weights, despite the lack of a natural ordering of cross-covariance operators. Identifiability holds up to a signed permutation matrix under Assumptions 2.1–2.3, with pure variables acting as observable proxies for latent factors. Estimation proceeds by locating pure variables from row maxima of cross-covariance operator norms (Theorem 2), and then recovering non-pure loadings by solving W^{-1}A_I^T Σ_IJ = C A_J^T, which is re-expressed as a finite-dimensional least-squares problem us

What carries the argument

The central objects are the loading matrix A ∈ R^{p×K}, whose entries are cluster membership weights, and the covariance operator Σ of the p-variate functional variable, viewed as a block operator of cross-covariances. Pure variables — variables with |A_ja|=1 and zero elsewhere — act as anchors: their cross-covariances have operator norm exactly ||C_aa||, the maximum in the row, which lets Algorithm 1 recover them from Σ. The non-pure loadings are then obtained from the operator regression W^{-1}A_I^T Σ_IJ = C A_J^T; the key trick is to take Hilbert-Schmidt inner products of both sides with the covariance blocks C_{k,s}, converting the infinite-dimensional problem into inversion of a real K×

Load-bearing premise

Each latent cluster must contain at least two 'pure' variables that load exclusively on that cluster; without these anchors the loading matrix is not identifiable and the algorithm has no starting point.

What would settle it

Find two loading matrices A and A' that are not related by a signed permutation, both satisfying Assumptions 2.1–2.3, that produce the same covariance operator Σ = A C A^T + Γ = A' C' A'^T + Γ'; such a pair would disprove Theorem 1. A more targeted check: simulate Model (2) with a cluster containing only one pure variable and see whether two different ground-truth loadings yield indistinguishable covariance operators.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Overlapping cluster memberships in multivariate functional data are identifiable from the covariance operator up to label switching and sign, provided each cluster has two pure variables.
  • The number of clusters is estimated consistently with high probability, and pure-variable recovery is characterized up to quasi-pure contamination (Theorem 4).
  • The estimated overlap weights for non-pure variables are asymptotically normal with an explicit covariance, so practitioners can test H0: A_jk=0 and run multiple-testing corrections.
  • The method applies to non-overlapping clusters as a special case, where it competes with functional k-means and model-based clustering without needing K in advance.
  • In the fMRI application, the estimated overlapping clusters correspond to interpretable networks (visual, attentional, default-mode), consistent with overlapping brain parcellations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper assumes fully observed functions, but all experiments discretize on m points; the bias from this discretization and from B-spline smoothing is not part of the theory, so practitioners should expect the finite-sample guarantees to degrade as m shrinks.
  • The pure-variable anchor assumption could be relaxed in future work by using other identifiability routes, e.g., higher-order cumulants or non-Gaussianity, though the paper leaves this open.
  • The CLT's covariance V^{-1}TV^{-1} enables not only pointwise tests but also simultaneous confidence bands on membership profiles; such bands are not constructed in the paper but follow directly from the stated convergence.
  • The method's reliance on uncorrelated errors and independent observations suggests a natural extension to temporally or spatially dependent functional data; until then, applications to longitudinal or gridded spatial curves need caution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The manuscript develops a latent-factor model for multivariate functional data, X=AZ+E with X in H^p and Z in H^K, where the real loading matrix A encodes possibly overlapping cluster memberships. Under Assumptions 2.1–2.3 (row scaling, at least two pure variables per factor, separated and independent latent-factor covariance operators), Theorem 1 establishes identifiability of A and C from the covariance operator Sigma up to signed permutation. Estimation proceeds by extracting pure variables from row maxima of cross-covariance operator norms (Algorithm 1), constructing C from pure variables, and solving a Hilbert-Schmidt regression (Algorithm 2). The paper proves concentration of the sample covariance operator (Proposition 1), high-probability recovery of K and I with a quasi-pure characterization (Theorem 4), and asymptotic normality of the estimated non-pure loadings (Theorem 5), enabling tests on membership entries. Experiments validate the method on simulated overlapping and non-overlapping data and apply it to fMRI parcellation time series.

Significance. If correct, this is a substantial contribution: it extends the overlapping-clustering factor model of Bing et al. (2020) to functional data and provides frequentist inferential guarantees—identifiability, consistency, and a CLT—for soft functional clustering, which is new to this literature. The paper is notable for its clean operator/Hilbert-Schmidt formalism, explicit and honest assumption statements, and a Monte Carlo validation of the claimed asymptotic covariance. The authors also provide a code repository and are careful to state limitations of the theory (fully observed functions, pure-variable anchor, independence assumptions). The central concerns below are localized to the covariance formula in Theorem 5 and to one algebraic step in the proof of Theorem 4(c), both of which appear fixable without changing the overall framework.

major comments (2)
  1. [Section 4.2, Eq. (20), and Supplement C] The displayed formula for T_{s1j1,s2j2} is not the covariance of G(Z) as claimed. Terms IV and V contain (A_J)_{j2s} but no corresponding (A_J)_{j1s}; for the entry (s1j1,s2j2) one needs both Cov(G1,G3A_J^T) and Cov(G3A_J^T,G1), and similarly for G2/G3. Supplement C computes only the first orientation of each cross-covariance. These two scalar covariances differ when j1≠j2, so the resulting T is generally not symmetric and cannot be the stated positive semidefinite covariance. Diagonal entries, which drive single-coefficient z-tests, survive, but the full covariance matrix and any covariance-aware multiple-testing procedure based on V^{-1}TV^{-1} are unsupported as written. The proof should add the reverse cross-terms and the numerical check in Section 5.4 should be rerun against the corrected formula.
  2. [Supplement B.3, eq. (S.23)] The identities for ||Sigma_ii + Sigma_ij|| and ||Sigma_ii - Sigma_ij|| are algebraically incorrect for quasi-pure variables. For i in I_a and j in I_a ∪ J_a^1 with sign(A_{iπ(a)}) = sign(A_{jπ(a)}), the coefficient of A_{ib}C_{π(a)b} in the first sum of Sigma_ii+Sigma_ij is 2+|A_{jπ(a)}|, not 3, and the corresponding coefficient in Sigma_ii-Sigma_ij is 2-|A_{jπ(a)}|, not 1. Since quasi-pure variables need not have |A_{jπ(a)}|=1, the displayed equalities do not hold; the subsequent lower bound on ||Sigma_ii+Sigma_ij||-||Sigma_ii-Sigma_ij|| uses these expressions as exact identities. The statement of Theorem 4(c) may be true, but the current proof needs reworking, or the theorem needs an additional margin assumption that controls this coefficient error.
minor comments (4)
  1. [Section 5 and Section 7] The theory assumes fully observed functions, while all experiments use m-point discretization and B-spline basis approximation. This gap is acknowledged in Section 7, but it should also be stated when the discrete norms are introduced in Section 5, since the reported Riemann-sum approximations introduce a bias not covered by the theoretical guarantees.
  2. [Section 5.4] The validation of Theorem 5 is performed only for H=R. This is stated, but it would be helpful to note explicitly that the functional case is not empirically checked and that the scalar case may not exercise all features of the operator-valued formula.
  3. [Proposition 1] The symbol 'W' in the definitions of delta_a and delta_b is not defined; it appears to denote max (vee). Please define it or use standard notation.
  4. [Tables 1-3] The numbers in parentheses are presumably standard deviations across simulation replicates; the captions should state this explicitly.

Circularity Check

0 steps flagged

No significant circularity: the identifiability, clustering, and CLT results are derived from the model and validated against Monte Carlo, with no fitted quantity disguised as a prediction.

full rationale

The paper's derivation chain is self-contained. Theorem 1 is a mathematical identifiability result proved from Model (2) and Assumptions 2.1-2.3; Assumption 2.2 is a stated existence condition (pure variables), not the conclusion of the theorem. Theorem 2 derives a data-dependent criterion (row maxima of operator norms) for recovering pure variables; the proof uses the model and lemmas establishing that pure variables are exactly the rows/columns where the norm maximum is attained. Algorithm 1 computes the estimated pure set and number of clusters from the estimated covariance; Theorem 4 quantifies false discoveries in terms of the sampling error delta and separation nu, so the recovery claim is not assumed. Theorem 3 is an application of the standard Moore-Penrose least-squares solution in Hilbert-Schmidt spaces (cited to Hsing & Eubank 2015, a textbook theorem; despite one coauthor, it is a standard external result, not a novel self-citation). Theorem 5 derives asymptotic normality of sqrt(n)(bA_J^T - A_J^T) by a delta method from the known CLT for covariance operators; the limiting covariance is given explicitly and is validated in Section 5.4 and Supplement E.2 by Monte Carlo using the true S and C, so it is a genuine prediction check rather than a fit to the same data. The acknowledged limitations in Section 7 (two pure variables per factor, uncorrelated errors, fully observed functions vs. discrete grids) are scope restrictions, not circularity. The skeptic's concern that the displayed covariance in Theorem 5 is asymmetric is a mathematical correctness issue about omitted/reversed cross-covariance terms; it does not make the theorem circular, because the formula is not obtained by assuming the target conclusion.

Axiom & Free-Parameter Ledger

3 free parameters · 8 axioms · 1 invented entities

The honest price of admission is the factor model plus Assumptions 2.1–2.3. The pure-variable anchor (2.2) and the operator-separation condition (2.3) are the load-bearing postulates; both are flagged in Section 7. δ, α, and the basis size are practical tuning choices, not fitted targets.

free parameters (3)
  • δ (Algorithm 1 tuning radius) = chosen by cross-validation over grid δ_q = c_q sqrt(log p / n)
    Controls pure-variable detection in Algorithm 1; Proposition 1 bounds δ only up to an unspecified constant c_0, so in practice δ is tuned by CV (Section 5).
  • membership threshold α = 0.10
    Loadings with |Â_ik| ≤ 0.10 are zeroed when forming clusters in the overlapping-cluster simulations (§5.3) and the fMRI analysis (§6); the paper notes results depend on α.
  • B-spline basis = order 4, 10 basis functions
    Computational representation of functional samples (Section 5); impacts accuracy of the operator-norm and HS-norm approximations used throughout estimation.
axioms (8)
  • domain assumption Model (2): X = AZ + E with uncorrelated noise E and functional latent factors Z
    The entire clustering structure (clusters G_a = {i: A_ia ≠ 0}) is defined by this decomposition; no alternative data-generating mechanism is considered. Invoked in §2.2, eq. (2).
  • domain assumption Assumption 2.1: row-wise ℓ1 scaling Σ_a |A_ja| ≤ 1
    Needed in Lemma 1 to ensure pure variables achieve the row-maximum operator norm; a scaling convention that excludes heavy multi-membership rows.
  • domain assumption Assumption 2.2: at least two pure variables per cluster
    Load-bearing: Theorem 2, Lemma 2 and Algorithm 1 identify clusters by locating pure variables. Acknowledged as restrictive in Section 7.
  • domain assumption Assumption 2.3: minimum separation Δ(C) = min(‖C_aa‖∧‖C_bb‖−‖C_ab‖) > 0 and C has independent columns
    Guarantees no two latent factors are norm-equivalent and that the regression W^{-1}A_I^T Σ_IJ = C A_J^T has a unique solution. Needed for Theorems 1, 4, 5.
  • domain assumption X is centered and subgaussian (for concentration, Proposition 1)
    The operator-norm concentration of bΣ (Prop. 1) inherits subgaussianity from Theorem 9 of Koltchinskii & Lounici (2017).
  • standard math E‖X‖^4 < ∞ (fourth moment) for the CLT
    Standard moment condition for sample covariance CLT in Hilbert spaces (Hsing & Eubank 2015, Ch. 8), used in Theorem 5.
  • standard math Background operator theorems from Hsing & Eubank (2015): Moore-Penrose least-squares solution (Thm 3.5.10), cross-covariance inner-product identity (Thm 7.2.9), sample-covariance CLT (Ch. 8)
    Used in the proofs of Theorem 3 and Theorem 5. Standard textbook results, not the paper's target claims.
  • domain assumption Functions are fully observed; in practice they are observed on m-point grids with B-spline approximation
    Section 5 (m equidistant points, 10 B-splines) and Section 7 acknowledge the unmodeled approximation bias in operator-norm/HS-norm estimates (eq. S.51, E.1).
invented entities (1)
  • latent factors Z ∈ H^K (one per cluster) independent evidence
    purpose: Postulated shared sources of dependence; cluster G_a is defined by non-zero loadings on Z_a
    The model provides falsifiable handles outside itself: loading estimates, CLT-based tests of A_jk = 0, and the pure-variable structure are all checkable from data (Sections 4.2, 5.4).

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of funFMC: Overlapping Clustering for Functional Data." pith.science (2026). https://pith.science/paper/INLGSUIE

@misc{pith2026260713197,
  author       = {Pith},
  title        = {Pith review of: funFMC: Overlapping Clustering for Functional Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/INLGSUIE}},
  note         = {Machine review of arXiv:2607.13197}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In applications such as neuroscience and environmental science, data are naturally modeled as multivariate functional data and often exhibit overlapping cluster structure. Existing clustering methods for functional data typically impose mutually exclusive memberships and therefore fail to capture such structure. We propose a latent factor model based approach with functional factors and a real-valued loading matrix that encodes potentially overlapping cluster memberships. Under mild conditions, we establish identifiability of the loading matrix up to permutation, ensuring that the overlapping cluster structure is recoverable up to label switching. We develop a procedure for estimating both the number of clusters and the associated cluster memberships. This involves solving an infinite-dimensional regression problem in operators, whose solution is characterized using the inner product on the space of Hilbert-Schmidt operators and expressed in terms of real-valued matrices. This formulation enables rigorous asymptotic analysis, and we establish a central limit theorem to facilitate statistical inference on overlapping cluster memberships. We demonstrate the performance of our method using numerical studies and an application to functional magnetic resonance imaging data.

Figures

Figures reproduced from arXiv: 2607.13197 by Abhiti Mishra (1), Kean Ming Tan (1), Tailen Hsing (1) ((1) Department of Statistics, University of Michigan).

Figure 1
Figure 1. Figure 1: Empirical covariances (computed using Monte Carlo) versus theoretical covariances. [PITH_FULL_IMAGE:figures/full_fig_p019_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dependence of parcellations on the estimated clusters [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean cluster loadings by Yeo-17 network and hemisphere [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references

  1. [1]

    & Molinari, N

    Abraham, C., Cornillon, P., Matzner-Løber, E. & Molinari, N. (2003), ‘Unsupervised curve clustering usingB-splines’,Scandinavian Journal of Statistics30, 581 –

    2003

  2. [4]

    In the case that sign(A iπ(a))̸= sign(A jπ(a) ), we observe that the expressions for∥Σ ii + Σij∥ and∥Σ ii −Σ ij∥in (S.23) are reversed

    Then ∥Σii + Σij∥ − ∥Σii −Σ ij∥ ≥ ∥Cπ(a)π(a)∥ −2∥Γii∥+ 12δ >12δ.(S.27) We have proved that in the case that sign(A iπ(a)) = sign(A jπ(a) ),∥Σ ii + Σij∥ − ∥Σii −Σ ij∥> 12δ >0. In the case that sign(A iπ(a))̸= sign(A jπ(a) ), we observe that the expressions for∥Σ ii + Σij∥ and∥Σ ii −Σ ij∥in (S.23) are reversed. That is, ∥Σii −Σ ij∥= 2Cπ(a)π(a) +A iπ(a) X b̸=...

    2015

  3. [5]

    Recall that in Section 5.4, we used Monte-Carlo simulations to estimate bAJ under the assumption thatA I is known. Under the same simulation setting as Section 5.4, we check the distribution of elements of √n( bAJ −A J ), and check coverage of the confidence intervals computed using the asymptotic distribution. We perform all three diagnostic checks for a...

    2020

  4. [14]

    The (a, b)th element of the matrices in (S.3) 27 yields the equality|I a∥Ib|Cab = P i1,i2∈I Ai1aΣi1i2Ai2b.The elements{A ia :i∈I}are non-zero if and only ifi∈I a. Therefore, Cab = 1 |Ia∥Ib| X i1∈Ia,i2∈Ib Ai1aAi2bΣi1i2.(S.4) Recall the expression for Σ IJ from (S.1) and pre-multiply both sides byW −1A⊤ I to derive W −1A⊤ I ΣIJ =CA ⊤ J ,(S.5) where all term...

    2015

  5. [15]

    We define the event Eij(δij) ={∥ bΣij −Σ ij∥ ≤δij}

    The computations above simplify to ∥bΣij −Σ ij∥ ≤ ∥bΨij −Ψ ij∥. We define the event Eij(δij) ={∥ bΣij −Σ ij∥ ≤δij}. An application of (S.14) shows that there exists a constantc ij such that for allt≥1, P[Eij(δij)]≥1−e −t,(S.16) whereδ ij =c ij∥Ψij∥ q r(Ψij ) n W r(Ψij ) n W q t n W t n . 31 For eventE(δ) in (15) to occur, all eventsE ij(δij) wherei, j∈[p]...

    2011

  6. [85]

    & Guo, C

    Sonkusare, S., Breakspear, M. & Guo, C. (2019), ‘Naturalistic stimuli in neuroscience: Critically acclaimed’,Trends in Cognitive Sciences23(8), 699–714. Tarpey, T. & Kinateder, K. (2003), ‘Clustering functional data’,Journal of Classification 20, 093–114. Tokushige, S., Yadohisa, H. & Inada, K. (2007), ‘Crisp and fuzzyk-means clustering algorithms for mul...

    2019

  7. [131]

    & Kokoszka, P

    Horv´ ath, L. & Kokoszka, P. (2012),Inference for Functional Data with Applications, Springer Series in Statistics, Springer, New York, NY. Hsing, T. & Eubank, R. (2015),Theoretical Foundations of Functional Data Analysis, with an Introduction to Linear Operators, John Wiley and Sons, Ltd. Ieva, F., Paganoni, A. M., Pigoli, D. & Vitelli, V. (2013), ‘Multi...

    2012

  8. [192]

    & M¨ uller, H.-G

    Chiou, J.-M. & M¨ uller, H.-G. (2013), ‘Linear manifold modelling of multivariate functional data’,Journal of the Royal Statistical Society Series B: Statistical Methodology76(3), 605–

    2013

  9. [304]

    & Sansonnet, L

    Denis, C., Lebarbier, ´E., L´ evy-Leduc, C., Martin, O. & Sansonnet, L. (2020), ‘A novel regu- larized approach for functional data clustering: An application to milking kinetics in dairy goats’,Journal of the Royal Statistical Society: Series C (Applied Statistics)69(3), 623–640. Dubin, J. A. & M¨ uller, H.-G. (2005), ‘Dynamical correlation for multivari...

    2020

  10. [512]

    & Gelfand, A

    Nguyen, X. & Gelfand, A. E. (2011), ‘The Dirichlet labeling process for clustering functional data’,Statistica Sinica21, 1249–1289. Pollard, D. (1981), ‘Strong consistency ofk-means clustering’,The Annals of Statistics 9(1), 135–140. Ramos-Carre˜ no, C., Torrecilla, J. L., Carbajo-Berrocal, M., Marcos, P. & Su´ arez, A. (2024), ‘scikit-fda: A python packa...

    2011

  11. [595]

    Anderson, T. W. & Rubin, H. (1956), ‘Statistical inference in factor analysis’,Proceedings of the Third Berkeley Symposium on Mathematical Statistics and ProbabilityV 111-150, 1954–

    1956

  12. [626]

    & Shulman, G

    Corbetta, M., Patel, G. & Shulman, G. L. (2008), ‘The reorienting system of the human brain: From environment to theory of mind’,Neuron58(3), 306–324. Corbetta, M. & Shulman, G. L. (2002), ‘Control of goal-directed and stimulus-driven attention in the brain’,Nature Reviews Neuroscience3(3), 201–215. Delaigle, A., Hall, P. & Pham, T. (2019), ‘Clustering fu...

    2008

  13. [988]

    & Matricardi, P

    22 Giordani, P., Perna, S., Bianchi, A., Pizzulli, A., Tripodi, S. & Matricardi, P. M. (2020), ‘A study of longitudinal mobile health data through fuzzy clustering methods for functional data: The case of allergic rhinoconjunctivitis in childhood’,PLoS ONE15(11), e0242197. Golovkine, S., Klutchnikoff, N. & Patilea, V. (2022), ‘Clustering multivariate func...

    2020

  14. [1131]

    Serinko, R. J. & Babu, G. J. (1992), ‘Weak limit theorems for univariatek-mean clustering under a nonregular condition’,Journal of Multivariate Analysis41(2), 273–296. Snoek, L., van der Miesen, M. M., Beemsterboer, T., van der Leij, A., Eigenhuis, A. & Scholte, H. S. (2021), ‘The Amsterdam Open MRI Collection, a set of multimodal MRI datasets for individ...

    1992

  15. [1760]

    & Jacques, J

    Bouveyron, C. & Jacques, J. (2011), ‘Model-based clustering of time series in group-specific functional subspaces’,Advances in Data Analysis and Classification5, 281–300. Bouveyron, C., Jacques, J., Schmutz, A., Sim˜ oes, F. & Bottini, S. (2022), ‘Co-clustering of multivariate functional data for the analysis of air pollution in the South of France’,The A...

    2011

  16. [1955]

    & Wegkamp, M

    Bing, X., Bunea, F., Ning, Y. & Wegkamp, M. (2020), ‘Adaptive estimation in structured factor models with applications to overlapping clustering’,The Annals of Statistics48, 2055–2081. Bouveyron, C., Cˆ ome, E. & Jacques, J. (2015), ‘The discriminative functional mixture model for a comparative analysis of bike sharing systems’,The Annals of Applied Stati...

    2020

  17. [7751]

    & Lounici, K

    Koltchinskii, V. & Lounici, K. (2017), ‘Concentration inequalities and moment bounds for sample covariance operators’,Bernoulli23(1), 110–133. Koopmans, T. C. & Reiersol, O. (1950), ‘The identification of structural characteristics’,Annals of Mathematical Statitics21, 165–181. Kringelbach, M. L., Perl, Y. S., Tagliazucchi, E. & Deco, G. (2023), ‘Toward na...

    2017

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.