Pith. sign in

REVIEW 4 major objections 6 minor 40 references

Localized Functional Principal Component Analysis Based on Covariance Structure

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper shows that if a process splits into uncorrelated sub-processes on disjoint intervals, applying FPCA to each piece yields eigenfunctions of the whole process that are automatically localized and orthogonal.

desk verdict A clean, correct L-FPCA method for exactly block-diagonal covariances, but it does not yet face the near-block-diagonal case that real data will present. read the letter →

arxiv 2506.02836 v1 pith:RYVG24HN submitted 2025-06-03 stat.ME

Maria Laura Battagliola , Jan O. Bauer This is my paper · ORCID
classification stat.ME MSC 62R1062H25
keywords functionaldataanalysisprincipalcomponentlocalizedeigenfunctionssparseblockdiagonalcovariancedomainselectionstructuredetection
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

This paper proposes a functional principal component analysis that returns eigenfunctions vanishing outside automatically chosen intervals, without any sparsity penalty. The key claim is that this localization follows directly from the covariance structure: if the process is a sum of uncorrelated sub-processes on disjoint intervals, then the eigenfunctions of each sub-process are already eigenfunctions of the full process on that interval, with the same eigenvalues. Because the supports are disjoint, the localized eigenfunctions are orthonormal, and each eigenvalue directly measures that region's contribution to total variance. The method detects the block-diagonal structure of the sample covariance, runs FPCA on each detected block, and orders the resulting eigenvalues. Simulations and two data examples (pinch force curves and EEG recordings) illustrate that the method recovers supports accurately and preserves explained variance better than penalized sparse FPCA.

What carries the argument

The argument is carried by the block-diagonal covariance structure, the Karhunen-Loève expansion of each sub-process, and BD-SVD, a singular-vector-based algorithm that detects the block-diagonal covariance structure from the noisy sample covariance. Lemma 1 ties zero cross-block covariance to compactly supported sub-process eigenfunctions, Lemma 2 equates sub-process eigenfunctions and eigenvalues with those of the full process on each interval, and Corollary 2 gives orthonormality of the assembled localized eigenfunctions.

What would settle it

Run L-FPCA on simulated curves with a known block-diagonal covariance: if the detected supports are wrong by more than a few grid points, or if the per-block PVE differs from the true block PVE by more than sampling variability, the pipeline fails. A sharper version adds a small constant $\varepsilon$ to off-block covariance entries and checks whether the estimated eigenfunctions leak outside their supports.

Watch

Extended reading notes

Core claim

The central claim is Lemma 2: under the block-diagonal covariance assumption, the eigenfunctions and eigenvalues of each sub-process $X_k$ on its interval $T_k$ coincide with those of the full process $X$ on that interval. Corollary 1 then states that all eigenfunctions of $X$ are compactly supported on single intervals, so the sparse structure of the process is mirrored by the eigenfunctions without any explicit sparsity penalty. Consequently, Procedure 1 — detect blocks via a singular-vector block-detection algorithm, run FPCA on each block, sort the eigenvalues — directly produces orthonormal localized principal components.

Load-bearing premise

The construction assumes the covariance $\Gamma(s,t)$ is exactly zero whenever $s$ and $t$ lie in different intervals of the partition; if small cross-block correlations exist, the true eigenfunctions are not exactly zero outside their blocks and the hard block constraint will bias the estimated eigenfunctions and eigenvalues.

Editorial extensions

If this is right

  • Because the localized eigenfunctions are orthonormal by construction, the explained variance of the full process is additive across sub-processes, so each interval's contribution can be read directly from its eigenvalues.
  • The PVE of a sub-process equals the sum of its eigenvalues divided by the total variance, which follows from Lemma 2; this makes per-region explained variance available even when a region contributes multiple components.
  • Since L-FPCA never penalizes the eigenfunctions, the estimated eigenvalues and PVEs do not suffer the over-regularization shrinkage observed for penalized sparse FPCA in the simulation study.
  • The number of blocks detected by BD-SVD determines both the number of localized components and the partition of the domain, so the method outputs a domain segmentation alongside the principal components.

Reading between the lines

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

  • Beyond the paper: the same block-diagonal reasoning should extend to multidimensional domains $T\subset\mathbb{R}^d$, where blocks become product regions, provided the cross-region covariance vanishes; the orthogonality argument would carry over unchanged.
  • Beyond the paper: the hard zero-covariance assumption invites a natural stress test — simulate from an approximately block-diagonal covariance and measure how support recovery and per-block PVE degrade as off-block correlations grow.
  • Beyond the paper: the EEG result that alcoholic and control groups differ in the number of detected sub-processes implies the block count itself could serve as a classification feature, a use the paper mentions only as a future direction.
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

4 major / 6 minor

Summary. The paper proposes a localized functional principal component analysis (L-FPCA) method built on the assumption that the covariance function of the stochastic process is block-diagonal with respect to a partition of the domain. Under this assumption, the process decomposes into uncorrelated sub-processes on disjoint intervals, and the eigenfunctions of the full process coincide on each interval with the eigenfunctions of the corresponding sub-process. The proposed Procedure 1 detects the block structure via BD-SVD (Bauer 2025), then applies standard FPCA within each detected block, yielding localized eigenfunctions that are orthonormal by construction. The paper provides theoretical lemmas (Lemma 1, Lemma 2, Corollary 1, Corollary 2) establishing these equivalences, a simulation study comparing L-FPCA with thresholded FPCA and SFPCA, and two real-data applications (pinch force curves and EEG data).

Significance. If the block-diagonal assumption holds exactly, the method provides localized eigenfunctions that coincide with the population eigenfunctions on each block, avoiding the over-regularization and non-orthogonality issues of penalized approaches. The theoretical development is elementary but sound under assumption (3), and the simulation study is carefully designed to show the method's support-recovery and variance-explained performance in the exact block-diagonal setting. The paper ships reproducible R code and a detailed supplementary appendix, which is a strength. However, the practical relevance of the method is currently limited by the strength of the block-diagonal assumption and by the absence of any analysis of what happens when the covariance is only approximately block-diagonal, which is the realistic regime for functional data.

major comments (4)
  1. [Section 3, Eq. (3) and Procedure 1] The exactness of Lemma 2 and Corollary 1 relies entirely on the population-level assumption (3) that Γ(s,t)=0 for s and t in different blocks. The last paragraph of Section 3 correctly notes that the empirical covariance perturbs the block structure, but the manuscript gives no analysis of the effect of nonzero off-block entries. When off-block correlations are small but nonzero, the eigenvalues of block submatrices are not eigenvalues of the full covariance operator, and the hard support constraints imposed in Step 1 of Procedure 1 force estimated eigenfunctions to be exactly zero in regions where the true eigenfunctions are merely small. The simulation study in Section 4 generates data from exactly block-diagonal covariances in both designs, so this misspecification regime is never exercised. I request a concrete error analysis (e.g., bounds on eigenfunction and eigenvalue bias as a function of the operator norm of the off-block part of Γ) or, at minimum, a simulation that adds off-block correlations of varying magnitude and reports bias in eigenfunctions, eigenvalues, and support recovery.
  2. [Section 3, last paragraph; Procedure 1, Step 1] The method depends critically on BD-SVD (Bauer 2025) to recover the unknown blocks from finite noisy data, but no consistency theorem for BD-SVD in the functional setting is stated, and no result quantifies how detection errors propagate to the estimated localized eigenfunctions and eigenvalues. Because Step 1 is the only data-dependent step that determines the support of every eigenfunction, an incorrect block partition directly changes the output. Please add either a support-recovery consistency result (e.g., as N and the grid density grow, the estimated blocks converge to the true blocks with probability tending to one) or a sensitivity analysis that varies the signal-to-noise ratio and the separability of blocks in the simulation.
  3. [Section 4.2, Definition of Specificity and Precision] The text states that "TN denotes the number of correctly identified grid points in the support" and "TP the number of correctly identified grid points outside the support," but the formulas immediately below—TN=|S^c_hat∩S^c| and TP=|S_hat∩S|—define TN as correctly identified points outside the support and TP as correctly identified points inside the support. The labeling is reversed and makes the interpretation of Figure 2 ambiguous. Please correct the terminology or the formulas.
  4. [Section 5.2, EEG application] The real-data analysis interprets the intervals detected by BD-SVD as "uncorrelated sub-processes" and draws comparative conclusions about the number of blocks in the alcoholic versus control groups. The paper provides no diagnostic or test supporting the plausibility of the block-diagonal assumption (3) for these data; the observed difference in block counts could partly reflect different noise levels or signal complexity between groups rather than a genuine difference in the number of uncorrelated sub-processes. Please add a diagnostic of off-block correlation strength (e.g., a permutation-based summary of the empirical covariance outside the detected blocks) or soften the causal interpretation of the group comparison.
minor comments (6)
  1. [Section 3, Lemma 1] Lemma 1 is vacuous as stated: the eigenfunctions of the sub-process X_k are compactly supported on T_k by the definition in (2), regardless of whether the cross-block covariance is zero. The substantive statement is Lemma 2, which requires (3). Consider reformulating Lemma 1 as a remark to avoid giving the impression that (3) is needed for the trivial support property.
  2. [Section 4.1 and 4.2, threshold values] The threshold values for FPCA(τ) are inconsistent: Section 4.1 states τ=10^{-3} and τ=10^{-5}, while Section 4.2 and the caption of Figure 2 use 10^{-3} and 10^{-4}. Please align these values.
  3. [Section 4.1, M=10] The simulation uses 8 eigenvalues in the KL representation (8), but the text says "We pick M=10 eigenfunctions in all our scenarios." Clarify whether M=10 refers to the number of estimated components or the truncation level; the relationship between M and the 8 true components is not explained.
  4. [Procedure 1, Step 1] The text refers to "supports T_1,...,T_b" but the partition size is denoted K throughout the paper; this is a typo.
  5. [Section 4.3, equation reference] The text says "computing the percentage of variance explained (PVE) in (4.3)" but the displayed equation is not numbered; add a number or refer to the display directly.
  6. [Introduction, terminology] The phrase "uncorrelated sub-processes" is used as if it follows from the partition, but it is an assumption (3) about the covariance. Please make this explicit early in the paper to avoid confusion.

Circularity Check

1 steps flagged · score 2.0 of 10

No major circularity: the central block-diagonal spectral theorem is genuine, but Lemma 1's localization claim is partly built into the sub-process definition, and the practical block-detection step is delegated to a same-author citation.

  1. self definitional [Section 3, Lemma 1 (with Section 2, Eq. (2))]
    "It holds that Γ(s, t) a.e. = 0 for s ̸∈ Tk and t ∈ Tk if and only if the eigenfunctions {ψk,j(·)}∞ j=1 of sub-process Xk(·) are compactly supported on Tk, namely ψk,j(t) = 0 for t /∈ Tk."

    The sub-process Xk is defined in Eq. (2) to be exactly X(t) on Tk and 0 outside Tk. Therefore its covariance Γk(s,t) is identically zero whenever s or t is outside Tk, and any eigenfunction of Xk automatically vanishes outside Tk (for nonzero eigenvalue). This compact-support conclusion follows from the definition of Xk alone, not from the stated cross-block vanishing condition on Γ. Presenting Lemma 1 as an 'if and only if' consequence of the block structure makes localization look like an output when it is already an input of the decomposition. The genuinely nontrivial part, that these sub-process eigenpairs are also eigenpairs of X, is Lemma 2, which does depend on assumption (3) and is not circular.

full rationale

L-FPCA's central equivalence (Lemma 2 and Corollary 1) is a genuine spectral consequence of the block-diagonal covariance assumption (3): if Γ vanishes across disjoint intervals, the covariance operator splits into a direct sum, so the eigenpairs on each block are eigenpairs of the full operator. This derivation is self-contained and does not reduce to a fitted parameter or to a prediction that is forced by construction. The one step that is self-definitional is Lemma 1, because the sub-process Xk is defined to be zero outside Tk, so its eigenfunctions are automatically compactly supported on Tk regardless of the cross-block condition on Γ; the paper presents this as a derived 'if and only if' property, but the support is an input of the definition. The practical block-detection step is delegated to Bauer (2025), a same-author citation for BD-SVD; this is an external methodological dependency rather than a fitted parameter renamed as a prediction, and it does not make the theoretical claim circular. However, the paper supplies no consistency or misspecification analysis of BD-SVD in the functional setting, and the simulations generate data from exactly the block-diagonal model that BD-SVD is designed to recover, so the misspecified regime is never exercised. These are robustness concerns, not circularity. Overall score 2 reflects the minor self-definitional support claim, not the main spectral result.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method relies on the block-diagonal covariance assumption (3) as its core structural input. The other listed axioms are standard regularity conditions. The practical block detection via BD-SVD is an unproven step in this paper, and the choice of K and boundaries is effectively a free modeling choice.

free parameters (1)
  • Number of sub-processes K and partition boundaries
    Determined by BD-SVD (Bauer, 2025) on the empirical covariance. The paper provides no theoretical analysis of how errors in K and boundaries propagate to the localized eigenfunctions.
assumptions (3)
  • domain assumption The covariance function Γ(s,t) is block diagonal with respect to the partition T1,...,TK, i.e., Γ(s,t)=0 for s in Ti, t in Tj, i≠j (equation (3)).
    This is the central structural assumption of the paper; without it, Lemma 1 and Corollary 1 do not hold and the localized eigenfunctions are not guaranteed to match the true process.
  • standard math The stochastic process X is mean-zero, continuous, and square-integrable on a compact interval T, and its eigenelements have positive eigenvalues.
    These are standard regularity conditions for FPCA and Mercer's theorem, stated in Section 2.
  • ad hoc to paper The finite-dimensional truncation and the empirical covariance estimate preserve the block structure of the true covariance, and BD-SVD recovers this structure correctly.
    The practical implementation relies on BD-SVD to identify the supports; the paper does not prove consistency or error bounds for this detection step in the functional setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Localized Functional Principal Component Analysis Based on Covariance Structure." pith.science (2026). https://pith.science/paper/RYVG24HN

@misc{pith2026250602836,
  author       = {Pith},
  title        = {Pith review of: Localized Functional Principal Component Analysis Based on Covariance Structure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RYVG24HN}},
  note         = {Machine review of arXiv:2506.02836}
}
read the original abstract

Functional principal component analysis (FPCA) is a widely used technique in functional data analysis for identifying the primary sources of variation in a sample of random curves. The eigenfunctions obtained from standard FPCA typically have non-zero support across the entire domain. In applications, however, it is often desirable to analyze eigenfunctions that are non-zero only on specific portions of the original domain-and exhibit zero regions when little is contributed to a specific direction of variability-allowing for easier interpretability. Our method identifies sparse characteristics of the underlying stochastic process and derives localized eigenfunctions by mirroring these characteristics without explicitly enforcing sparsity. Specifically, we decompose the stochastic process into uncorrelated sub-processes, each supported on disjoint intervals. Applying FPCA to these sub-processes yields localized eigenfunctions that are naturally orthogonal. In contrast, approaches that enforce localization through penalization must additionally impose orthogonality. Moreover, these approaches can suffer from over-regularization, resulting in eigenfunctions and eigenvalues that deviate from the inherent structure of their population counterparts, potentially misrepresenting data characteristics. Our approach avoids these issues by preserving the inherent structure of the data. Moreover, since the sub-processes have disjoint supports, the eigenvalues associated to the localized eigenfunctions allow for assessing the importance of each sub-processes in terms of its contribution to the total explained variance. We illustrate the effectiveness of our method through simulations and real data applications. Supplementary material for this article is available online.

Figures

Figures reproduced from arXiv: 2506.02836 by the authors.

Figure 1
Figure 1. Eigenfunctions (dotted black) and localized eigenfunctions (solid gray) of an electroencephalogram dataset ranging from 0 to 256 Hz (3.9-ms epoch). An illustration of the methodology proposed in this work is provided in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Visualization of specificity (left) and precision (right) for L-FPCA, SFPCA with thresholds 10−4 and 10−5 , and FPCA(τ ) with thresholds 10−3 and 10−4 . For both design A (left column) and design B (right column), there are 500 realizations for sample sizes N ∈ {250, 500, 1000}. 4.3 Explained Variance A second motivation for L-FPCA is that the methodology preserves the explained variance of the principal components … view at source ↗
Figure 3
Figure 3. Ratio PVE \k/PVEk of estimated and true PVE for L-FPCA, SFPCA with thresh￾olds 10−4 and 10−5 , and FPCA(τ ) with thresholds 10−3 and 10−4 . The ratio is summarized in a single box-plot summarizing all three sub-processes k ∈ {1, 2, 3} across simulation designs A (left) and B (right) for N ∈ {250, 500, 1000}. In [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Top: Smoothed pinch force curves, with colored segments highlighting the disjoint sub-intervals of the uncorrelated sub-processes identified by our methodology. Bottom: First three localized eigenfunctions obtained using L-FPCA, colored according to the their support. …
Figure 5
Figure 5. Figure 5: Top: Number of identified disjoint sub-intervals of the uncorrelated sub￾processes across all electrodes for the control group (left) and alcoholic group (right). Bottom: Smoothed EEG signals from the 64th electrode for the control (left) and alcoholic group (right). T…
Figure 6
Figure 6. Figure 6: First six localized eigenfunctions obtained via L-FPCA for the EEG curves [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages

  1. [1]

    Bauer, J. O. (2025). High-dimensional block diagonal covariance structure detection using singular vectors. J. Comput. Graph. Stat.\/ 0\/ (0), 1--12

    work page 2025

  2. [2]

    Bauer, J. O. and B. Drabant (2021). Principal loading analysis . J. Multivar. Anal.\/ 184

    work page 2021

  3. [3]

    Beyaztas, U., H. L. Shang, and A.-S. G. Abdel-Salam (2022). Functional linear models for interval-valued data. Commun. Stat. - Simul. Comput.\/ 51\/ (7), 3513--3532

    work page 2022

  4. [4]

    Carrizosa, A

    Blanquero, R., E. Carrizosa, A. Jim \'e nez-Cordero, and B. Mart \' n-Barrag \'a n (2020). Selection of time instants and intervals with support vector regression for multivariate functional data. Comput. Oper. Res\/ 123 , 105050

    work page 2020

  5. [5]

    Carrizosa, C

    Blanquero, R., E. Carrizosa, C. Molero-R \' o, and D. R. Morales (2023). On optimal regression trees to detect critical intervals for multivariate functional data. Comput. Oper. Res\/ 152 , 106152

    work page 2023

  6. [6]

    Chen, K. and J. Lei (2015). Localized functional principal component analysis. J. Am. Stat. Assoc.\/ 110\/ (511), 1266--1275

    work page 2015

  7. [7]

    Crainiceanu, C. M., J. Goldsmith, A. Leroux, and E. Cui (2024). Functional data analysis with R . CRC Press

    work page 2024

  8. [8]

    Delaigle, A., P. Hall, W. Huang, and A. Kneip (2021). Estimating the covariance of fragmented and other related types of functional data. J. Am. Stat. Assoc.\/ 116\/ (535), 1383--1401

    work page 2021

Show all 40 references
  1. [9]

    Hall, and T

    Delaigle, A., P. Hall, and T. Pham (2019). Clustering functional data into groups by using projections. J. R. Stat. Soc. B\/ 81\/ (2), 271--304

    2019

  2. [10]

    Devijver, E. and M. Gallopin (2018). Block-diagonal covariance selection for high-dimensional gaussian graphical models. J. Am. Stat. Assoc.\/ 113\/ (521), 306--314

    2018

  3. [11]

    Di, C.-Z., C. M. Crainiceanu, B. S. Caffo, and N. M. Punjabi (2009). Multilevel functional principal component analysis. Ann. Appl. Stat.\/ 3\/ (1), 458

    2009

  4. [12]

    Fan, J. and R. Li (2001). Variable selection via nonconcave penalized likelihood and its oracle properties. J. Am. Stat. Assoc.\/ 96\/ (456), 1348--1360

    2001

  5. [13]

    Greven, and C

    Goldsmith, J., S. Greven, and C. Crainiceanu (2013). Corrected confidence bands for functional data using principal components. Biometrics\/ 69\/ (1), 41--51

    2013

  6. [14]

    M \"u ller, and J.-L

    Hall, P., H.-G. M \"u ller, and J.-L. Wang (2006). Properties of principal component methods for functional and longitudinal data analysis . Ann. Stat.\/ 34\/ (3), 1493--1517

    2006

  7. [15]

    Ieva, F., A. M. Paganoni, and N. Tarabelloni (2016). Covariance-based clustering in multivariate and functional data analysis. J. Mach. Learn. Res.\/ 17\/ (143), 1--21

    2016

  8. [16]

    James, G. M. and C. A. Sugar (2003). Clustering for sparsely sampled functional data. J. Am. Stat. Assoc.\/ 98\/ (462), 397--408

    2003

  9. [17]

    Xiao, and S

    Li, C., L. Xiao, and S. Luo (2020). Fast covariance estimation for multivariate sparse functional data. Stat\/ 9\/ (1), e245

    2020

  10. [18]

    Wang, and J

    Lin, Z., L. Wang, and J. Cao (2016). Interpretable functional principal component analysis. Biometrics\/ 72\/ (3), 846--854

    2016

  11. [19]

    Nie, Y. and J. Cao (2020). Sparse functional principal component analysis in a new regression framework. Comput. Stat. Data Anal.\/ 152 , 107016

    2020

  12. [20]

    Pini, A. and S. Vantini (2016). The interval testing procedure: a general framework for inference in functional data analysis. Biometrics\/ 72\/ (3), 835--845

    2016

  13. [21]

    Pini, A. and S. Vantini (2017). Interval-wise testing for functional data. J. Nonparametric Stat.\/ 29\/ (2), 407--424

    2017

  14. [22]

    R: A Language and Environment for Statistical Computing

    R Core Team (2022). R: A Language and Environment for Statistical Computing . Vienna, Austria: R Foundation for Statistical Computing

    2022

  15. [23]

    Ramsay, J. (2024). fda: Functional Data Analysis . R package version 6.2.0

    2024

  16. [24]

    Ramsay, J. and B. Silverman (2005). Functional Data Analysis . Springer Series in Statistics. Springer

    2005

  17. [25]

    Rice, J. A. and B. W. Silverman (1991). Estimating the mean and covariance structure nonparametrically when the data are curves. J. R. Stat. Soc. B\/ 53\/ (1), 233--243

    1991

  18. [26]

    Witten, and J

    Saha, A., D. Witten, and J. Bien (2024). Inferring independent sets of gaussian variables after thresholding correlations. J. Am. Stat. Assoc.\/ 0\/ (0), 1--12

    2024

  19. [27]

    S ent \"u rk, S

    Shamshoian, J., D. S ent \"u rk, S. Jeste, and D. Telesca (2022). Bayesian analysis of longitudinal and multidimensional functional data. Biostatistics\/ 23\/ (2), 558--573

    2022

  20. [28]

    Shen, H. and J. Z. Huang (2008). Sparse principal component analysis via regularized low rank matrix approximation. J. Multivar. Anal.\/ 99\/ (6), 1015--1034

    2008

  21. [29]

    Staniswalis, J. G. and J. J. Lee (1998). Nonparametric regression analysis of longitudinal data. J. Am. Stat. Assoc.\/ 93\/ (444), 1403--1418

    1998

  22. [30]

    Tarpey, T. and K. K. Kinateder (2003). Clustering functional data. J. Classif.\/ 20\/ (1)

    2003

  23. [31]

    Li, and C

    Wang, H., B. Li, and C. Leng (2009). Shrinkage tuning parameter selection with a diverging number of parameters. J. R. Stat. Soc. B\/ 71\/ (3), 671--683

    2009

  24. [32]

    Wang, J., R. K. Wong, and X. Zhang (2022). Low-rank covariance function estimation for multidimensional functional data. J. Am. Stat. Assoc.\/ 117\/ (538), 809--822

    2022

  25. [33]

    Chiou, and H.-G

    Wang, J.-L., J.-M. Chiou, and H.-G. M \"u ller (2016). Functional data analysis. Annu. Rev. Stat. Appl.\/ 3\/ (1), 257--295

    2016

  26. [34]

    Wang, S. and G. Cao (2024). Multiclass classification for multidimensional functional data through deep neural networks. Electron. J. Stat.\/ 18\/ (1), 1248--1292

    2024

  27. [35]

    Xiao, L., C. Li, W. Checkley, and C. Crainiceanu (2018). Fast covariance estimation for sparse functional data. Stat. Comput.\/ 28 , 511--522

    2018

  28. [36]

    Zipunnikov, D

    Xiao, L., V. Zipunnikov, D. Ruppert, and C. Crainiceanu (2016). Fast covariance estimation for high-dimensional functional data. Stat. Comput.\/ 26 , 409--421

    2016

  29. [37]

    Yang, and F

    Xue, K., J. Yang, and F. Yao (2024). Optimal linear discriminant analysis for high-dimensional functional data. J. Am. Stat. Assoc.\/ 119\/ (546), 1055--1064

    2024

  30. [38]

    M \"u ller, and J.-L

    Yao, F., H.-G. M \"u ller, and J.-L. Wang (2005). Functional data analysis for sparse longitudinal data. J. Am. Stat. Assoc.\/ 100\/ (470), 577--590

    2005

  31. [39]

    Qing Shi, and W.-S

    Zeng, P., J. Qing Shi, and W.-S. Kim (2019). Simultaneous registration and clustering for multidimensional functional data. J. Comput. Graph. Stat.\/ 28\/ (4), 943--953

    2019

  32. [40]

    Hastie, and R

    Zou, H., T. Hastie, and R. Tibshirani (2006). Sparse principal component analysis. J. Comput. Graph. Stat.\/ 15\/ (2), 265--286

    2006

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.