Pith. sign in

REVIEW 3 major objections 4 minor 67 references

A Flag Decomposition for Hierarchical Datasets

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper proves that any matrix with nested column groups of strictly increasing rank admits a hierarchy-preserving flag decomposition D=QRP^T, and gives an algorithm to compute it.

desk verdict Block QR in flag clothing with a real theorem/practice gap on truncated flag types, plus modest experimental gains. read the letter →

arxiv 2502.07782 v2 pith:PMTSPGD3 submitted 2025-02-11 cs.CV

Nathan Mankovich , Ignacio Santamaria , Gustau Camps-Valls , Tolga Birdal This is my paper · ORCID
classification cs.CV MSC 15A2314M1565F25
keywords flagdecompositionmanifoldsStiefelcoordinateshierarchicaldatablockGram-Schmidtfew-shotlearninghyperspectralimagedenoisingsubspacerecovery
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 introduces a matrix factorization that preserves nested hierarchies in data. It proves that whenever a data matrix's columns can be partitioned into nested groups with strictly increasing ranks, the matrix admits an exact decomposition D=QRP^T, where Q is an orthonormal basis for a flag (a nested sequence of subspaces), R is block upper triangular, and P is a permutation. The authors give an algorithm, Flag-BMGS, that constructs this decomposition block by block and a robust variant for data with outlier columns. They show the decomposition improves denoising and clustering of hyperspectral images and, used as class prototypes, improves few-shot classification accuracy on several benchmarks. Standard tools like SVD and QR either ignore the hierarchy or can only recover limited nested structure, so this is a general, hierarchy-preserving alternative.

What carries the argument

The central object is the flag manifold F L(n1,...,nk;n) in Stiefel coordinates: a tall-skinny orthonormal matrix Q=[Q1|...|Qk] whose prefixes span a nested sequence of subspaces. The load-bearing identity is the projection property Π_{Q_i^⊥} ... Π_{Q_1^⊥} B_i = 0, which holds exactly when the nested column groups have strictly increasing rank; it lets each block B_i be reconstructed from its projections onto the flag pieces. Flag-BMGS is a block modified Gram-Schmidt procedure that builds Q block by block, solving a small subspace-recovery problem at each step, and assembles R and P from those projections. The rotational ambiguity of Stiefel coordinates (Prop. 4) means the decomposition is unique as a flag, not as a matrix.

What would settle it

Compute the ranks of [D_{A_i}] for a concrete dataset with a proposed hierarchy; if any two consecutive groups have equal rank, the if-and-only-if theorem says no exact flag decomposition exists, so Flag-BMGS output cannot be hierarchy-preserving. Concretely, duplicating a column across two nested groups makes the rank increments vanish. For the few-shot claim, one can numerically check whether the span of early-layer AlexNet features is contained in the span of the full features on the actual support set; a counterexample would invalidate the flag-prototype distance.

Watch

Extended reading notes

Core claim

The central claim is that flag decomposition is exactly the right factorization for hierarchical data: a matrix D admits a flag decomposition of type (n1,...,nk;n) if and only if the nested column index sets A1⊂...⊂Ak form a column hierarchy, meaning dim([D_{A_{i-1}}]) < dim([D_{A_i}]) for every i. When this holds, one can construct an n×n_k matrix Q with orthonormal columns, partitioned as Q=[Q1|...|Qk], whose prefixes span exactly the same nested subspaces as the column groups of D; together with a block upper triangular R and a permutation P, it satisfies D=QRP^T. The construction is explicit: project each block onto the orthogonal complement of all previous flag pieces, take an orthonormal basis of the result (via SVD, or IRLS-SVD for robustness), and assemble R from the projections. The flag is unique up to block rotations, so the factorization is a property of the hierarchy rather than of a particular basis.

Load-bearing premise

The load-bearing premise is that every nested column group increases the column rank, so that the subspaces spanned by [D_{A_i}] are strictly nested; if a supplied hierarchy violates this (or, in the few-shot application, if early-layer features do not actually lie in the span of the full features), the exact decomposition does not exist.

Editorial extensions

If this is right

  • Any matrix whose nested column groups have strictly increasing ranks can be factored exactly as D=QRP^T with a hierarchy-preserving flag in Stiefel coordinates (Prop. 3).
  • Flag-BMGS computes this decomposition block by block, with cost equal to k SVDs on the hierarchy blocks; the IRLS-SVD variant (RFD) handles outlier columns by using an L1 objective.
  • Truncating the flag type gives a low-rank reconstruction that preserves the column hierarchy, which SVD- and QR-based reconstruction do not; this is what drives the reported gains in hyperspectral image denoising.
  • Flag prototypes for few-shot learning embed all shots of a class into a single flag and score queries by projection residuals; on AlexNet features the flag classifier matches or improves Euclidean and subspace prototypes in the tested 5-way tasks.
  • Chordal distances between flags, computed from the decomposition, give distance matrices and MDS embeddings that cluster noisy hierarchical datasets more cleanly than Euclidean or SVD-based distances.

Reading between the lines

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

  • The exactness condition suggests a diagnostic for downstream use: run a rank-increment check before applying Flag-BMGS, so that invalid hierarchies are rejected instead of silently producing approximations.
  • The decomposition points toward a hierarchy-aware analogue of PCA in which nested subspaces, rather than principal directions, are the preserved quantities; a statistical model for flag-type estimation under noise would be a natural next step.
  • The few-shot distance of Eq. (15) is differentiable, so fine-tuning the feature extractor with a flag-classifier loss is a direct extension the paper leaves open.
  • Because the flag is determined only up to block rotations, distances between flag representations are invariant to the particular Stiefel coordinates chosen, making the representation well suited to clustering and averaging on the flag manifold.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces the Flag Decomposition (FD), a factorization D = Q R P^T in which Q is a Stiefel-coordinate representative of a flag associated with a user-specified hierarchy on the columns of D, R is block upper triangular, and P is a permutation that orders columns by hierarchy blocks. The authors prove existence for the exact flag type n_i = rank(D_{A_i}) (Props. 1-3 / 5-7), state a block-rotational ambiguity (Prop. 4/8), propose the Flag-BMGS algorithm and a robust IRLS variant to compute the decomposition, and apply the method to denoising, clustering, and few-shot learning. Experiments on simulated data, AVIRIS hyperspectral images, and ImageNet-feature few-shot benchmarks report improvements over SVD and QR baselines.

Significance. The exact existence result is correct, and the paper ships code, which makes the experiments reproducible. The geometric framing (Stiefel coordinates for flags, chordal distances on flag manifolds) is appropriate, and the few-shot prototype idea is original. However, the advertised 'hierarchy-preserving' claim is proved only when the flag type equals the empirical ranks of the hierarchy blocks. All applications use truncated flag types, so a central theoretical support for those results is missing. The paper is nevertheless a meaningful step, and the gap is addressable with an explicit approximation analysis or with a flag-type selection step.

major comments (3)
  1. [Sec. 3, Prop. 3 (and App. B, Prop. 7)] As stated, Proposition 3 asserts the equivalence between existence of an FD of type (n1,...,nk;n) and A1⊂...⊂Ak being a column hierarchy. This is missing the condition n_i = rank(D_{A_i}) for all i. Definition 5 requires [D_{A_i}] = [Q_1,...,Q_i]; since Q has orthonormal columns, dim([Q_1,...,Q_i]) = n_i, so equality forces n_i = dim([D_{A_i}]). The proof of Prop. 7 itself uses this dimension matching when it derives Eq. (32). Consequently, if the user specifies any truncated type with n_i < rank(D_{A_i}) — precisely the regime recommended in Sec. 3.2 ('we choose a flag type where n_k < rank(˜D)') — no exact hierarchy-preserving FD exists. The statement should either include the rank constraint or define an approximate FD and bound how well [Q_1,...,Q_i] approximates [D_{A_i}].
  2. [Secs. 5.3 and 5.5; Sec. 3.2] The experimental sections use truncated flag types: HSI denoising uses (8,9,10;2500) without reporting the ranks of the band blocks, and few-shot uses (s−1,2(s−1)) for data matrices in R^{4096×2s}. In both cases the rank of D_{A_i} is almost certainly larger than n_i, so Definition 5 is violated exactly. The claim in the abstract that FD 'decomposes arbitrary hierarchical real-valued data into a hierarchy-preserving flag representation' is therefore not supported for the configurations that are actually evaluated. Please add a formal statement for the truncated case (e.g., an approximation bound in chordal distance or a guarantee on the nested subspace residuals) and verify the rank condition for the feature hierarchy in Example 2.3.
  3. [Sec. 2, Example 2.3] The feature hierarchy assumes [f^(1)_Θ(x_1),...,f^(1)_Θ(x_s)] ⊆ [f_Θ(x_1),...,f_Θ(x_s)] ('information flows from f^(1)_Θ to f^(2)_Θ'). This inclusion is not tested on the AlexNet features used in Sec. 5.5. If it fails, then A1 ⊂ A2 is not a column hierarchy for D, and even the full-rank version of Prop. 3 does not apply. The few-shot experiments should either verify this inclusion empirically or state the hierarchy as an assumption of the method.
minor comments (4)
  1. [Table 1] The caption spells 'Stiefel' as 'Steifel'; please correct the typo.
  2. [Fig. 6] The vertical axis is labeled 'LSRE' but the metric defined in Table 3 is 'LRSE'; please make the labels consistent.
  3. [Sec. 4.3, Eq. (15)] The sentence following Eq. (15) says the distance is 'proportional to the squared chordal distance on FL(1,2;n)' when the query matrix is in Stiefel coordinates. The constant of proportionality and the exact condition on the query matrix should be stated explicitly.
  4. [App. D, Algorithm 2] The output line states 'D=QRP^T' without qualification. For truncated flag types the equality is only approximate; please use '≈' or include a residual term so the algorithm's output matches the theory.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FD is a constructive linear-algebra decomposition; self-citations are background and no fitted parameter is relabeled as a prediction.

full rationale

The paper's central claim, that a matrix D admits a hierarchy-preserving flag decomposition when the nested column groups have strictly increasing column-space dimensions, is a constructive existence theorem. The proof builds Q by successive orthogonalization of the column blocks (Prop. 5), then defines R and P directly from Q and the input hierarchy (Prop. 6), so no quantity is fitted to any evaluation metric or derived from the target outcome. The flag type and column hierarchy are explicit user inputs, not learned parameters, and the experimental comparisons report classification/denoising accuracy on fixed benchmarks without tuning the decomposition to those numbers. Self-citations to the authors' earlier flag-manifold work appear in related-work and distance/background contexts but are not load-bearing: the Stiefel-coordinate model is attributed to Ye et al. [60] and the block Gram-Schmidt machinery to Barlow [3] and Jalby-Philippe [19], both independent sources. The most serious issue in the paper is not circularity: Prop. 3 is stated for an arbitrary flag type, but the proof constructs the full-rank type n_i = rank(D_Ai), while the applications use truncated types (e.g., n_k < rank(D_tilde)), so the exact hierarchy-preservation guarantee is silently absent for those experiments. Likewise, the few-shot feature hierarchy of Example 2.3 is an untested assumption, not a constructed equivalence. These are correctness/scope gaps, not reductions of a prediction to its own input, and therefore do not raise the circularity score.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities; its central claim rests on standard linear algebra (block QR/SVD, projections) and on user-supplied inputs (hierarchy and flag type). The main domain-specific assumption is that deep feature hierarchies yield nested column spaces, which is plausible but unverified for AlexNet.

free parameters (4)
  • flag type (n1,...,nk;n) = e.g., (8,9,10;2500) for KSC, (1,8;176) for KSC patches, (s-1,2(s-1);4096) for few-shot
    User-specified input to Flag-BMGS; controls the rank of each nested subspace and is chosen by hand per experiment, not learned or validated.
  • column hierarchy A1 subset ... subset Ak = e.g., first 40, first 100, all 176 bands for KSC; center pixel vs 3x3 patch; f1 vs f features
    Must be provided a priori and satisfy the rank condition; the paper selects it from domain knowledge (spectral ranges, patch geometry, network layers).
  • robust cost parameters (q,r) = q=1, r=2 for RFD; r=q=2 for FD
    Algorithm hyperparameters; q=1 gives L1 robustness, q=2 gives SVD.
  • IRLS weight floor = 10^-8
    Stabilizes IRLS-SVD; not justified by convergence analysis.
assumptions (5)
  • standard math Standard facts about orthogonal projections, SVD, QR, and orthonormal bases.
    Used throughout the existence proofs and Algorithm 1.
  • standard math Flag manifold is diffeomorphic to St(n_k,n)/(O(m1)x...xO(mk)).
    Invoked in Definition 4 and Prop. 4; from Ye et al. [60].
  • domain assumption The column hierarchy satisfies dim([D_{A_{i-1}}]) < dim([D_{A_i}]) for each i.
    Definition 1; Prop. 3 shows this is necessary and sufficient for an exact FD; if violated, the claimed decomposition does not exist.
  • domain assumption In the few-shot setup, features from the earlier layer f^(1) span a subspace of the full features f.
    Example 2.3 and Section 5.5 rely on this to build a meaningful flag from AlexNet features; not empirically verified.
  • domain assumption The pre-trained AlexNet can be written as f = f^(2) composed with f^(1), both outputs in R^4096.
    Used to construct the feature hierarchy in few-shot experiments; the decomposition of the network into two blocks is standard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Flag Decomposition for Hierarchical Datasets." pith.science (2026). https://pith.science/paper/PMTSPGD3

@misc{pith2026250207782,
  author       = {Pith},
  title        = {Pith review of: A Flag Decomposition for Hierarchical Datasets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PMTSPGD3}},
  note         = {Machine review of arXiv:2502.07782}
}
read the original abstract

Flag manifolds encode nested sequences of subspaces and serve as powerful structures for various computer vision and machine learning applications. Despite their utility in tasks such as dimensionality reduction, motion averaging, and subspace clustering, current applications are often restricted to extracting flags using common matrix decomposition methods like the singular value decomposition. Here, we address the need for a general algorithm to factorize and work with hierarchical datasets. In particular, we propose a novel, flag-based method that decomposes arbitrary hierarchical real-valued data into a hierarchy-preserving flag representation in Stiefel coordinates. Our work harnesses the potential of flag manifolds in applications including denoising, clustering, and few-shot learning.

Figures

Figures reproduced from arXiv: 2502.07782 by the authors.

Figure 1
Figure 1. A flag decomposition (center) is used for a hierarchy [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. We recover a flag from D with hierarchy A1 ⊂ A2. Columns of D are plotted as points with A1 in blue and A2 \ A1 in orange. FD is the only method that recovers the flag (line inside plane). SVD correctly recovers the plane but not the line whereas QR only recovers the line and the plane misses the orange points. Definition 6 (Flag Decomposition (FD)). Let D ∈ R n×p be data with the hierarchically nested sequence of c… view at source ↗
Figure 3
Figure 3. Images from the YFB [5] are flattened and horizontally stacked into D. We use the hierarchy with the images of the first subject as A1 and all images as A2. We run FD (flag type (1, 2)) and baselines (rank 2). FD is the only method to correctly recon￾struct the subjects. We plot the basis vectors (eigenfaces) on the right and find FD extracts basis elements that most closely resem￾ble the subjects. 4. Applications B… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: To perform few shot-learning, we embed all [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: FD & RFD improve flag recovery while maintaining ac [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: (Top row) distance matrices using Euclidean distance (Euclidean) and chordal distance between flags (SVD and FD). (Bottom row) 2D representation of the data colored by cluster via MDS applied to the distance matrix [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: FD improves hyperspectral image denoising over SVD [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 66 canonical work pages

  1. [1]

    Gen- eralized Weiszfeld algorithms for lq optimization.IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(4):728–745, 2014

    Khurrum Aftab, Richard Hartley, and Jochen Trumpf. Gen- eralized Weiszfeld algorithms for lq optimization.IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(4):728–745, 2014. 4

    work page 2014

  2. [2]

    Bandos, Lorenzo Bruzzone, and Gustavo Camps- Valls

    Tatyana V . Bandos, Lorenzo Bruzzone, and Gustavo Camps- Valls. Classification of hyperspectral images with regular- ized linear discriminant analysis.IEEE Transactions on Geo- science and Remote Sensing, 47(3):862–873, 2009. 7

    work page 2009

  3. [3]

    Block modified Gram–Schmidt algorithms and their analysis.SIAM Journal on Matrix Analysis and Applications, 40(4):1257–1290, 2019

    Jesse L Barlow. Block modified Gram–Schmidt algorithms and their analysis.SIAM Journal on Matrix Analysis and Applications, 40(4):1257–1290, 2019. 4

    work page 2019

  4. [4]

    Weiszfeld’s method: Old and new results.Journal of Optimization Theory and Appli- cations, 164:1–40, 2015

    Amir Beck and Shoham Sabach. Weiszfeld’s method: Old and new results.Journal of Optimization Theory and Appli- cations, 164:1–40, 2015. 4

    work page 2015

  5. [5]

    Belhumeur, Joao P Hespanha, and David J

    Peter N. Belhumeur, Joao P Hespanha, and David J. Krieg- man. Eigenfaces vs. fisherfaces: Recognition using class specific linear projection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(7):711–720, 1997. 5

    work page 1997

  6. [6]

    Numerical methods for computing angles between linear subspaces.Mathematics of computation, 27(123):579–594, 1973

    Ake Bjorck and Gene H Golub. Numerical methods for computing angles between linear subspaces.Mathematics of computation, 27(123):579–594, 1973. 3

    work page 1973

  7. [7]

    Introduction to hierarchical matrices with applications.En- gineering analysis with boundary elements, 27(5):405–422,

    Steffen B ¨orm, Lars Grasedyck, and Wolfgang Hackbusch. Introduction to hierarchical matrices with applications.En- gineering analysis with boundary elements, 27(5):405–422,

    work page

  8. [8]

    Manopt, a Matlab toolbox for optimization on manifolds.JMLR, 15(1):1455–1459, 2014

    Nicolas Boumal, Bamdev Mishra, P-A Absil, and Rodolphe Sepulchre. Manopt, a Matlab toolbox for optimization on manifolds.JMLR, 15(1):1455–1459, 2014. 6

    work page 2014

Show all 67 references
  1. [9]

    Shape spaces of nonlinear flags

    Ioana Ciuclea, Alice Barbora Tumpach, and Cornelia Viz- man. Shape spaces of nonlinear flags. InInternational Con- ference on Geometric Science of Information, pages 41–50. Springer, 2023. 1

    2023

  2. [10]

    SIAM, 1997

    James W Demmel.Applied numerical linear algebra. SIAM, 1997. 4

    1997

  3. [11]

    A flag representation for finite collec- tions of subspaces of mixed dimensions.Linear Algebra and its Applications, 451:15–32, 2014

    Bruce Draper, Michael Kirby, Justin Marks, Tim Marrinan, and Chris Peterson. A flag representation for finite collec- tions of subspaces of mixed dimensions.Linear Algebra and its Applications, 451:15–32, 2014. 1, 6

    2014

  4. [12]

    The ge- ometry of algorithms with orthogonality constraints.SIAM Journal on Matrix Analysis and Applications, 20(2):303– 353, 1998

    Alan Edelman, Tom ´as A Arias, and Steven T Smith. The ge- ometry of algorithms with orthogonality constraints.SIAM Journal on Matrix Analysis and Applications, 20(2):303– 353, 1998. 3

    1998

  5. [13]

    A review on the selection criteria for the truncated SVD in Data Science applications.Journal of Computational Mathematics and Data Science, 5, 2022

    Antonella Falini. A review on the selection criteria for the truncated SVD in Data Science applications.Journal of Computational Mathematics and Data Science, 5, 2022. 5

    2022

  6. [14]

    The geometric median on riemannian manifolds with application to robust atlas estimation.NeuroImage, 45(1): S143–S152, 2009

    P Thomas Fletcher, Suresh Venkatasubramanian, and Sarang Joshi. The geometric median on riemannian manifolds with application to robust atlas estimation.NeuroImage, 45(1): S143–S152, 2009. 6

    2009

  7. [15]

    Subspace averaging and order determina- tion for source enumeration.IEEE Transactions on Signal Processing, 67(11):3028–3041, 2019

    Vaibhav Garg, Ignacio Santamaria, David Ramirez, and Louis L Scharf. Subspace averaging and order determina- tion for source enumeration.IEEE Transactions on Signal Processing, 67(11):3028–3041, 2019. 4

    2019

  8. [16]

    Matan Gavish and David L. Donoho. The optimal hard threshold for singular values is 4√ 3 .IEEE Transactions on Information Theory, 60(8):5040–5053, 2014. 5

    2014

  9. [17]

    A block Householder– based algorithm for the QR decomposition of hierarchical matrices.SIAM Journal on Matrix Analysis and Applica- tions, 45(2):847–874, 2024

    Vincent Griem and Sabine Le Borne. A block Householder– based algorithm for the QR decomposition of hierarchical matrices.SIAM Journal on Matrix Analysis and Applica- tions, 45(2):847–874, 2024. 5

    2024

  10. [18]

    Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217–2226, 2019. 8

    2019

  11. [19]

    Stability analysis and improvement of the block Gram–Schmidt algorithm.SIAM journal on scientific and statistical computing, 12(5):1058– 1073, 1991

    William Jalby and Bernard Philippe. Stability analysis and improvement of the block Gram–Schmidt algorithm.SIAM journal on scientific and statistical computing, 12(5):1058– 1073, 1991. 4, 15

    1991

  12. [20]

    Learning multiple layers of features from tiny images.Technical report, 2009

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.Technical report, 2009. 8

    2009

  13. [21]

    Imagenet classification with deep convolutional neural net- works.Advances in neural information processing systems, 25, 2012

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works.Advances in neural information processing systems, 25, 2012. 8

    2012

  14. [22]

    Sage, 1978

    Joseph B Kruskal and Myron Wish.Multidimensional scal- ing. Sage, 1978. 5

    1978

  15. [23]

    A scalable second order method for ill-conditioned matrix completion from few samples

    Christian K ¨ummerle and Claudio M Verdun. A scalable second order method for ill-conditioned matrix completion from few samples. InInternational Conference on Machine Learning, pages 5872–5883. PMLR, 2021. 4

    2021

  16. [24]

    Iteratively reweighted least squares for basis pursuit with global linear convergence rate.Advances in Neural In- formation Processing Systems, 34:2873–2886, 2021

    Christian K ¨ummerle, Claudio Mayrink Verdun, and Dominik St¨oger. Iteratively reweighted least squares for basis pursuit with global linear convergence rate.Advances in Neural In- formation Processing Systems, 34:2873–2886, 2021. 4

    2021

  17. [25]

    Hierarchy, complexity, society

    David Lane. Hierarchy, complexity, society. InHierarchy in natural and social sciences, pages 81–119. Springer, 2006. 1

    2006

  18. [26]

    Hypersdfusion: Bridging hierarchical structures in language and geometry for enhanced 3d text2shape genera- tion

    Zhiying Leng, Tolga Birdal, Xiaohui Liang, and Federico Tombari. Hypersdfusion: Bridging hierarchical structures in language and geometry for enhanced 3d text2shape genera- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 19691–19700,

  19. [27]

    Fast, robust and non-convex subspace recovery.Information and Inference: A Journal of the IMA, 7(2):277–336, 2018

    Gilad Lerman and Tyler Maunu. Fast, robust and non-convex subspace recovery.Information and Inference: A Journal of the IMA, 7(2):277–336, 2018. 4

    2018

  20. [28]

    An overview of robust sub- space recovery.Proceedings of the IEEE, 106(8), 2018

    Gilad Lerman and Tyler Maunu. An overview of robust sub- space recovery.Proceedings of the IEEE, 106(8), 2018

    2018

  21. [29]

    Robust computation of linear models by convex re- laxation.Foundations of Computational Mathematics, 15: 363–410, 2015

    Gilad Lerman, Michael B McCoy, Joel A Tropp, and Teng Zhang. Robust computation of linear models by convex re- laxation.Foundations of Computational Mathematics, 15: 363–410, 2015. 4

    2015

  22. [30]

    Robust recovery of subspace structures by low-rank representation.IEEE transactions on pattern anal- ysis and machine intelligence, 35(1):171–184, 2012

    Guangcan Liu, Zhouchen Lin, Shuicheng Yan, Ju Sun, Yong Yu, and Yi Ma. Robust recovery of subspace structures by low-rank representation.IEEE transactions on pattern anal- ysis and machine intelligence, 35(1):171–184, 2012. 4

    2012

  23. [31]

    Mouton, 1966

    Robert E Longacre.Hierarchy in language. Mouton, 1966. 1

    1966

  24. [32]

    Self- organizing mappings on the flag manifold with applications to hyper-spectral image data analysis.Neural Computing and Applications, 34(1):39–49, 2022

    Xiaofeng Ma, Michael Kirby, and Chris Peterson. Self- organizing mappings on the flag manifold with applications to hyper-spectral image data analysis.Neural Computing and Applications, 34(1):39–49, 2022. 1

    2022

  25. [33]

    Chordal averaging on flag manifolds and its applications

    Nathan Mankovich and Tolga Birdal. Chordal averaging on flag manifolds and its applications. InICCV, pages 3881– 3890, 2023. 1, 6

    2023

  26. [34]

    The flag median and FlagIRLS

    Nathan Mankovich, Emily J King, Chris Peterson, and Michael Kirby. The flag median and FlagIRLS. InCVPR, pages 10339–10347, 2022. 4, 6

    2022

  27. [35]

    Module representatives for refining gene co-expression modules.Physical Biology, 20 (4):045001, 2023

    Nathan Mankovich, Helene Andrews-Polymenis, David Threadgill, and Michael Kirby. Module representatives for refining gene co-expression modules.Physical Biology, 20 (4):045001, 2023. 1

    2023

  28. [36]

    Fun with Flags: Robust principal directions via flag mani- folds

    Nathan Mankovich, Gustau Camps-Valls, and Tolga Birdal. Fun with Flags: Robust principal directions via flag mani- folds. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2024. 1

    2024

  29. [37]

    PhD thesis, Colorado State University, 2023

    Nathan J Mankovich.Subspace and Network Averaging for Computer Vision and Bioinformatics. PhD thesis, Colorado State University, 2023. 6

    2023

  30. [38]

    Finding the subspace mean or median to fit your need

    Tim Marrinan, J Ross Beveridge, Bruce Draper, Michael Kirby, and Chris Peterson. Finding the subspace mean or median to fit your need. InCVPR, 2014. 1

    2014

  31. [39]

    On a minimum enclosing ball of a collection of linear subspaces.Linear Algebra and its Applications, 625:248–278, 2021

    Tim Marrinan, P-A Absil, and Nicolas Gillis. On a minimum enclosing ball of a collection of linear subspaces.Linear Algebra and its Applications, 625:248–278, 2021. 5

    2021

  32. [40]

    Automated flower classification over a large number of classes

    Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In2008 Sixth Indian conference on computer vision, graphics & im- age processing, pages 722–729. IEEE, 2008. 8

    2008

  33. [41]

    Riemannian optimization method on generalized flag mani- folds for complex and subspace ICA

    Yasunori Nishimori, Shotaro Akaho, and Mark D Plumbley. Riemannian optimization method on generalized flag mani- folds for complex and subspace ICA. InAIP Conference Pro- ceedings, pages 89–96. American Institute of Physics, 2006. 1

    2006

  34. [42]

    Riemannian optimization method on generalized flag man- ifolds for complex and subspace ICA

    Yasunori Nishimori, Shotaro Akaho, and Mark D Plumbley. Riemannian optimization method on generalized flag man- ifolds for complex and subspace ICA. InAIP Conference, 2006

    2006

  35. [43]

    Riemannian optimization method on the flag manifold for independent subspace analysis

    Yasunori Nishimori, Shotaro Akaho, and Mark D Plumbley. Riemannian optimization method on the flag manifold for independent subspace analysis. InInternational conference on independent component analysis and signal separation, pages 295–302. Springer, 2006

    2006

  36. [44]

    Flag manifolds for subspace ICA prob- lems

    Yasunori Nishimori, Shotaro Akaho, Samer Abdallah, and Mark D Plumbley. Flag manifolds for subspace ICA prob- lems. InICASSP, pages IV–1417. IEEE, 2007

    2007

  37. [45]

    Natural conjugate gradient on complex flag manifolds for complex independent subspace analysis

    Yasunori Nishimori, Shotaro Akaho, and Mark D Plumb- ley. Natural conjugate gradient on complex flag manifolds for complex independent subspace analysis. InInternational Conference on Artificial Neural Networks. Springer, 2008. 1

    2008

  38. [46]

    Beyond low rank+ sparse: Multiscale low rank matrix decomposition.IEEE journal of selected topics in signal processing, 10(4), 2016

    Frank Ong and Michael Lustig. Beyond low rank+ sparse: Multiscale low rank matrix decomposition.IEEE journal of selected topics in signal processing, 10(4), 2016. 5, 15

    2016

  39. [47]

    Barycentric subspace analysis on manifolds

    Xavier Pennec. Barycentric subspace analysis on manifolds. Annals of Statistics, 46(6A), 2018. 1

    2018

  40. [48]

    Flag orbit codes and their expansion to Stiefel codes

    Renaud-Alexandre Pitaval and Olav Tirkkonen. Flag orbit codes and their expansion to Stiefel codes. InIEEE Informa- tion Theory Workshop, pages 1–5. IEEE, 2013. 3

    2013

  41. [49]

    Columbia University Press, 1985

    Stanley N Salthe.Evolving hierarchical systems: their struc- ture and representation. Columbia University Press, 1985. 1

    1985

  42. [50]

    Adaptive subspaces for few-shot learn- ing

    Christian Simon, Piotr Koniusz, Richard Nock, and Mehrtash Harandi. Adaptive subspaces for few-shot learn- ing. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 4136–4145,

  43. [51]

    Prototypical networks for few-shot learning.Advances in neural informa- tion processing systems, 30, 2017

    Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning.Advances in neural informa- tion processing systems, 30, 2017. 8, 17

    2017

  44. [52]

    Rethinking the Rieman- nian logarithm on flag manifolds as an orthogonal alignment problem

    Tom Szwagier and Xavier Pennec. Rethinking the Rieman- nian logarithm on flag manifolds as an orthogonal alignment problem. InInternational Conference on Geometric Science of Information, pages 375–383. Springer, 2023. 2

    2023

  45. [53]

    The curse of isotropy: from principal components to principal subspaces, 2024

    Tom Szwagier and Xavier Pennec. The curse of isotropy: from principal components to principal subspaces, 2024. 1, 5, 6

    2024

  46. [54]

    Nested subspace learning with flags, 2025

    Tom Szwagier and Xavier Pennec. Nested subspace learning with flags, 2025. 1

    2025

  47. [55]

    Representing part-whole hierarchies in foundation models by learning localizability composability and decomposability from anatomy via self supervision

    Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Representing part-whole hierarchies in foundation models by learning localizability composability and decomposability from anatomy via self supervision. In Proceedings of the IEEE/CVF Conference on Compute...

    2024

  48. [56]

    Pymanopt: A Python toolbox for optimization on man- ifolds using automatic differentiation.arXiv preprint arXiv:1603.03236, 2016

    James Townsend, Niklas Koep, and Sebastian Weichwald. Pymanopt: A Python toolbox for optimization on man- ifolds using automatic differentiation.arXiv preprint arXiv:1603.03236, 2016. 6

    2016 arXiv

  49. [57]

    Dual principal component pursuit.Journal of Machine Learning Research, pages 1—- 50, 2018

    Manolis Tsakiris and Ren ´e Vidal. Dual principal component pursuit.Journal of Machine Learning Research, pages 1—- 50, 2018. 4

    2018

  50. [58]

    Fast, blind, and accurate: Tuning-free sparse re- gression with global linear convergence

    Claudio Mayrink Verdun, Oleh Melnyk, Felix Krahmer, and Peter Jung. Fast, blind, and accurate: Tuning-free sparse re- gression with global linear convergence. InThe Thirty Sev- enth Annual Conference on Learning Theory, pages 3823–

  51. [59]

    Hd- cnn: hierarchical deep convolutional neural networks for large scale visual recognition

    Zhicheng Yan, Hao Zhang, Robinson Piramuthu, Vignesh Jagadeesh, Dennis DeCoste, Wei Di, and Yizhou Yu. Hd- cnn: hierarchical deep convolutional neural networks for large scale visual recognition. InProceedings of the IEEE international conference on computer vision, 2015. 1

    2015

  52. [60]

    Optimiza- tion on flag manifolds.Mathematical Programming, 194(1): 621–660, 2022

    Ke Ye, Ken Sze-Wai Wong, and Lek-Heng Lim. Optimiza- tion on flag manifolds.Mathematical Programming, 194(1): 621–660, 2022. 2, 3

    2022

  53. [61]

    A novel m-estimator for ro- bust PCA.The Journal of Machine Learning Research, 15 (1):749–808, 2014

    Teng Zhang and Gilad Lerman. A novel m-estimator for ro- bust PCA.The Journal of Machine Learning Research, 15 (1):749–808, 2014. 4 Appendices We provide alternative methods for flag recovery in Sec. A, proofs of each proposition in Sec. B, a discussion of block matrix decompo...

    2014

  54. [62]

    By way of contradiction, assumeC j =0

    Non-zero. By way of contradiction, assumeC j =0. ThenΠ Q⊥ :j−1 Bj =0. This means each column ofB j is in the column space ofQ :j−1. In terms of subspaces, this implies [Bj]⊆[Q 1,Q 2, . . . ,Qj−1] = [B1,B 2, . . . ,Bj−1](20) where the second equality follows from the induction ...

  55. [63]

    It suffices to show Q⊤ j Q:j−1 = [Q⊤ j Q1|Q⊤ j Q2| · · · |Q⊤ j Qj−1] =0(23) which is equivalent to showing[Q j]is orthogonal to [Q1,Q 2,

    Coordinates. It suffices to show Q⊤ j Q:j−1 = [Q⊤ j Q1|Q⊤ j Q2| · · · |Q⊤ j Qj−1] =0(23) which is equivalent to showing[Q j]is orthogonal to [Q1,Q 2, . . . ,Qj−1]. By construction, [Qj] = [Cj] = [ΠQ⊥ :j−1 Bj](24) which is orthogonal to[Q 1, . . . ,Qj−1]

  56. [64]

    Hierarchy. UsingQ ⊤ j Q:j−1 =0, we have ΠQ⊥ :j =I−Q :jQ⊤ :j =I− jX ℓ=1 QℓQ⊤ ℓ =I−Q jQ⊤ j −Q :j−1Q⊤ :j−1 =I−Q jQ⊤ j −Q :j−1Q⊤ :j−1 +Q j Q⊤ j Q:j−1 | {z } 0 Q⊤ :j−1, = (I−Q jQ⊤ j )(I−Q :j−1Q⊤ :j−1), =Π Q⊥ j ΠQ⊥ :j−1 . (25) By Eq. (25) and the construction[Q j] = [Π Q⊥ :j−1 Bj], ...

  57. [65]

    Using Eq

    Projection property. Using Eq. (25) and the induction hypothesis (4) thatΠ Q⊥ :j−1 =Π Q⊥ j−1 · · ·ΠQ⊥ 1 , we have ΠQ⊥ :j =Π Q⊥ j ΠQ⊥ :j−1 , =Π Q⊥ j ΠQ⊥ j−1 · · ·ΠQ⊥ 1 . (26) By construction[Q j] = [ΠQ⊥ :j−1 Bj]. Thus ΠQ⊥ j ΠQ⊥ :j−1 Bj =0. Using Eq. (26), we have ΠQ⊥ j ΠQ⊥ j−1 ...

  58. [66]

    By the induction hypothesis (5),Q i ∈ St(mi, n)fori= 1,2,

    Dimensions. By the induction hypothesis (5),Q i ∈ St(mi, n)fori= 1,2, . . . , j−1. SoQ:j−1 ∈St(n j−1, n) withn j−1 = Pj−1 i=1 mi. Let nj = rank(DAj ), = dim([B1,B 2, . . . ,Bj]), = dim([Q1,Q 2, . . . ,Qj]). ThusQ :j ∈R n×nj andQ j ∈R n×mj withm j =n j − nj−1.Q j has orthonorma...

  59. [67]

    for alli= 1,2, . . . , k. Specifically, given a column hi- erarchy onD, we have found coordinates for a hierarchy- preserving flag[ [Q] ]∈ F L(n1, n2, . . . , nk;n)that satisfies the projection property. Although we can writeR i,j =Q ⊤ i Bj forj≥i, an equivalent definition is ...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.