Exploring the Limitations of Structured Orthogonal Dictionary Learning

Aditya Siripuram; Anirudh Dash

REVIEW 2 major objections 6 minor 17 references

Exploring the Limitations of Structured Orthogonal Dictionary Learning

T0 review · 2 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper shows that two binary-coded columns of $Y=HX$ uniquely determine a Householder reflection matrix $H$ and the coefficient matrix $X$.

desk verdict The approximation half is a real contribution; the two-sample identifiability theorem is false as stated, so the paper needs major surgery. read the letter →

arxiv 2501.15094 v2 pith:JQ7DIXJM submitted 2025-01-25 eess.SP

Anirudh Dash , Aditya Siripuram This is my paper · ORCID

classification eess.SP MSC 15A2315A1865F25

keywords Householderreflectionsstructureddictionarylearningorthogonalmatrixfactorizationsamplecomplexitybinarycoefficientfasttransformsrecoveryeigendecomposition

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 addresses two questions about dictionaries built from Householder reflections, matrices of the form $H = I - 2uu^T$ with $u$ a unit vector. For approximation, it gives an algorithm that, given an orthogonal matrix $V$ and a target number $m$ of reflections, returns a product of $m$ Householders whenever $V$ is exactly such a product, finds the smallest such $m$, and, when truncated, comes with an explicit Frobenius-norm error bound. For learning, it claims that if observed data obey $Y = HX$ with $H$ a single Householder matrix and $X$ a binary ($0/1$) coefficient matrix, then two columns of $Y$ determine both $H$ and $X$ uniquely. This matters because storing and applying a product of $m$ Householders costs $O(mn)$ rather than $O(n^2)$, and two-column identifiability is far below what classical dictionary learning requires. The paper also proves the binary assumption is load-bearing: without it, no finite number of columns makes the recovery unique.

What carries the argument

The central objects are Householder reflections $H = I - 2uu^T$ for unit vectors $u$, and their symmetric parts. For approximation, the carrier of the argument is the eigenspace $E^1_V$ of eigenvalue $1$: a product of $m$ Householders has an eigenvalue-$1$ eigenspace of dimension at least $n-m$, and Algorithm 1 chooses $u$ as the eigenvector of the smallest eigenvalue of $(V_k + V_k^T)/2$, forms $H_k = I - 2uu^T$, and updates $V_{k+1} = H_k V_k$. Lemma 3 is the key mechanism: each iteration raises the dimension of the eigenvalue-$1$ eigenspace by exactly one, forcing termination at the minimal factor count. For the dictionary-learning theorem, the load-bearing structure is the binary support pattern of a column of $X$. The proof partitions rows into four sets $P_1,\ldots,P_4$ according to which of two candidate solutions have zero coefficients, introduces $\delta_1$ and $\delta_2$ as the sums of the candidate Householder vectors over zero-support positions, and Lemma 4 expresses one candidate vector entrywise as an affine function of the other; Lemma 5 then derives $|P_2| = |P_4|$, $\delta_2/\delta_1 = c_1/c_2$, and ratio identities that leave no room for a second consistent vector when two distinct binary columns are used.

What would settle it

Take $u=(1/\sqrt{2},-1/\sqrt{2},0,0)$ with binary columns $x_1=(1,1,0,0)^T$ and $x_2=(0,0,1,1)^T$; then $H=I-2uu^T$ fixes both columns, so $Y=X$, and any other Householder reflection whose vector is orthogonal to the two columns also fixes them, contradicting the unqualified two-column uniqueness claim unless the zero-support-sum condition is added.

Watch

Extended reading notes

Core claim

On the approximation side, the paper's central claim is that membership in $\mathcal{H}_m$ — the set of orthogonal matrices expressible as a product of at most $m$ Householder reflections — can be tested by repeatedly peeling off the reflection whose vector is the unit eigenvector of the smallest eigenvalue of the current residual's symmetric part $(V+V^T)/2$. Algorithm 1 performs this peeling; Theorem 1 states that it terminates at exactly the minimal $m$ with $V \in \mathcal{H}_m$, and Theorem 2 gives the error bound $\|V - \hat{V}_m\|_F \leq \sqrt{2(n - \operatorname{tr}(V) - 2\lfloor m/2 \rfloor + \sum_{i=1}^m \lambda_i)}$, with $\lambda_i$ the sorted eigenvalues of $V_{\mathrm{sym}}$. On the learning side, the paper's central claim is Theorem 3: for $Y = HX$ with $H = I - 2uu^T$ and $X$ an arbitrary binary matrix, two columns in $Y$ uniquely recover $H$ and $X$. The proof enumerates the $2^n$ binary vectors for the first column, solves the $n$ linear equations $Y_{ij} = \sum_k (\delta_{ik} - 2u_i u_k)X_{kj}$ for $u$, and shows through Lemmas 4 and 5 that two distinct columns cannot both admit the same spurious reflection vector unless the columns are identical, so only the true $u$ survives. Theorem 4 complements this by exhibiting two genuinely different real non-binary pairs $(H,X)$ with the same product, so the binary restriction is what makes two-sample recovery possible.

Load-bearing premise

The uniqueness proof for the two-column recovery divides by $\delta_1$ and $\delta_2$, the sums of the true Householder vector entries over the zero positions of the two binary columns, so it assumes both support sums are nonzero and that the two columns are distinct; if a column's zero-support sum vanishes, the same data can be consistent with many different Householder dictionaries.

Editorial extensions

If this is right

An orthogonal matrix known to be a product of $m$ Householders can be certified and factored in $O(n^3 m)$ time, yielding $O(mn)$ storage and $O(mn)$ per matrix-vector product instead of $O(n^2)$.
Truncating Algorithm 1 at $m$ steps gives a guaranteed approximation bound in terms of the spectrum of $V_{\mathrm{sym}}$, so the same procedure doubles as a fast-transforms approximation scheme.
For structured dictionary learning with binary coefficients, two distinct data columns are information-theoretically sufficient, so the sample complexity is exactly $2$ rather than growing with $n$.
Without a binary or similarly discrete coefficient constraint, the dictionary recovery problem is unidentifiable no matter how many samples are provided, so any polynomial-time method must exploit additional structure.
The constructive proof enumerates binary vectors and is exponential-time, so the two-sample guarantee does not promise a polynomial-time algorithm; the paper notes that meaningful lower bounds for efficient algorithms remain open.

Reading between the lines

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

The proof as written divides by $\delta_1$ and $\delta_2$ but the theorem statement does not state the nonzero-support-sum condition; counterexamples such as $u=(1/\sqrt{2},-1/\sqrt{2},0,0)$ with columns $(1,1,0,0)$ and $(0,0,1,1)$ suggest the statement needs that hypothesis made explicit.
Because arbitrary real coefficients defeat recovery even with unlimited samples, the discrete-versus-continuous gap suggests that identifiability of orthogonal dictionaries is governed by the coefficient model at least as much as by sample count.
The eigenvector-peeling strategy may extend to the generalized reflectors $z_1 I - z_2 uu^T$ floated in the discussion, where a richer class of orthogonal matrices could be approximated with fewer building blocks; testing this would require re-deriving the one-dimensional eigenspace growth step.
A natural next question is whether any polynomial-time algorithm can achieve the two-sample bound; the paper's exponential enumeration suggests a lower bound for efficient algorithms may exist, and proving one would sharpen the sample-complexity picture.

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.

Desk Editor's Note

read the letter

Two contributions, one good, one broken. The eigenspace-based algorithm for factoring an orthogonal matrix into a product of m Householder matrices, with the smallest-m guarantee and the truncation error bound (Theorems 1 and 2), is a plausible and useful variant of the known construction in [16]. The error bound in terms of the eigenvalues of the symmetric part is new as far as I know, and the numerical experiments support the claim. I did not find an obvious gap in that half.

The second half, however, is not correct. Theorem 3 claims that two columns of Y = HX with binary X always determine H and X uniquely. The proof only works under an unstated assumption that you never divide by zero; the line "We assume that we do not divide by 0 in any case" appears in the proof of Lemma 4, but the theorem statement does not carry any such condition. When the assumption fails, the claim is simply false. Take n=4, u=(1/sqrt(2), -1/sqrt(2), 0, 0), and the two distinct binary columns x1=(1,1,0,0)^T and x2=(0,0,1,1)^T. Both columns are orthogonal to u, so H = I - 2uu^T fixes them, Y = X. For any unit vector v in the two-dimensional subspace {v1+v2=0, v3+v4=0}, the Householder matrix I - 2vv^T also fixes both columns, so infinitely many distinct Householder dictionaries give exactly the same Y. There is also a gap in the proof after Lemma 5: the equality |P2| = |P4| does not imply that the two solutions differ only by permutation. That step is asserted, not derived.

The non-uniqueness result for arbitrary X (Theorem 4) is true but unsurprising; it does not compensate. The discussion's concluding remark about polynomial-time restrictions is sensible but underdeveloped.

Net: the advertised sample-complexity result is load-bearing and false. The approximation half may be salvageable as a standalone contribution, and a corrected identifiability theorem with explicit nondegeneracy assumptions might be recoverable. As written, I would not accept the paper. I would still send it to peer review rather than desk-reject, because the approximation algorithm is a real, defensible contribution and the counterexample should be put on record. With a major revision focused on the dictionary-learning section, this could become a useful paper.

Referee Report

2 major / 6 minor

Summary. The paper studies two problems. The first is approximation of a real orthogonal matrix by a product of a small number of Householder reflections. The authors propose an eigenspace-based greedy algorithm (Algorithm 1), prove that it terminates exactly when the input lies in H_m (Theorem 1), and give a Frobenius-norm error bound for arbitrary inputs (Theorem 2). The second problem is identifiability in the structured dictionary-learning model Y = HX, where H is an unknown Householder matrix and X is an unknown binary coefficient matrix. Theorem 3 claims that two columns of Y suffice to recover (H,X) uniquely, and Theorem 4 claims that without the binary restriction recovery is impossible. The paper also contains simulations for the approximation algorithm and a small illustrative table for the two-sample claim.

Significance. If Theorem 3 were correct, it would be a striking result: only two samples would be needed to identify a Householder dictionary from binary coefficients, much smaller than the sample sizes required in prior dictionary-learning work. The approximation algorithm and its error bound are of moderate interest, and the explicit non-uniqueness construction for arbitrary X (Theorem 4) is a useful sanity check. However, Theorem 3 is false as stated, and the proof contains a non-rigorous key step. Since the two-sample identifiability claim is the paper's main advertised contribution, the current version cannot be accepted. The first part of the paper might be salvageable separately, but the central claim of Section 3 is invalid.

major comments (2)

[Section 3.2, Theorem 3 and Lemma 4] Theorem 3 is false as stated. Let n=4, u=(1/sqrt(2), -1/sqrt(2), 0, 0), H=I-2uu^T, and let the binary coefficient matrix be X=[x1 x2] with x1=(1,1,0,0)^T and x2=(0,0,1,1)^T. Since u^T x1 = u^T x2 = 0, we have HX=X, so Y=X. For every unit vector v in the two-dimensional subspace {v in R^4 : v1+v2=0, v3+v4=0}, the Householder matrix H_v=I-2vv^T also fixes x1 and x2, hence H_v X = Y. This is a continuum of distinct Householder dictionaries producing the same data, so (H,X) is not uniquely recoverable from two distinct binary columns. In the proof of Lemma 4 the sentence "We assume that we do not divide by 0 in any case" hides the assumptions delta1 != 0 and delta2 != 0, which fail in this example (delta1=delta2=0). These nondegeneracy assumptions are absent from the theorem statement, so the theorem and the abstract's claim that two samples suffice are false.
[Section 3.2, Lemma 5 and following paragraph] The conclusion "According to result 1 in Lemma 5, the two matrices are only different due to permutation" is not justified. Lemma 5(1) states only |P2|=|P4|, an equality of cardinalities of two index sets; it says nothing about the coefficient matrices or Householder vectors being permutations of one another. This non sequitur is used as a step toward uniqueness. Even apart from the division-by-zero issue, the proof of uniqueness for the second column relies on informal assertions such as "there is a one-to-one relationship" and does not constitute a rigorous derivation. Thus the proof of Theorem 3 is incomplete in multiple places.

minor comments (6)

[Section 2.5] The text "we find matrcies H3, H4" contains a typo; the surrounding algebra is dense and would benefit from a cleaner derivation or a reference to a supplementary computation.
[Equation (3)] The phrase "Here delta_ik is the standard indicator for j = k" should read "for i = k"; the Kronecker delta in the formula compares row indices i and k.
[Algorithm 2] Algorithm 2 is not a fully specified deterministic algorithm: it says "Set the first column of X as a random n length binary vector" and "Find u" without characterizing the solution set for a given guess or giving a formal termination condition. The exponential complexity is acknowledged, but the correctness statement is a direct consequence of the (false) Theorem 3 and therefore cannot be accepted.
[Proof of Theorem 4] The remark that binary constraints "ensure that up to permutation, only O(n) such possibilities existed" is inaccurate: there are 2^n binary vectors of length n, not O(n). This does not affect the example itself, but the motivating statement should be corrected.
[Section 4] The conclusion that "meaningful lower bounds on sample complexity ... can only be obtained by restricting the dictionary learning algorithm to be polynomial time" is not supported by the preceding theorems, which concern information-theoretic identifiability rather than algorithmic lower bounds; the discussion should be clarified.
[Throughout] There are numerous typos and formatting issues: "consoderation", "his will hopefully", missing spaces in the product notation, and garbled table entries in the Section 3.3 illustration. A careful editorial pass is needed before resubmission.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the derivation chain is self-contained; the unstated division-by-zero assumption in Lemma 4 is a correctness gap, not a circular dependency.

full rationale

Walking the derivation chain, I find no circular step. Section 2's approximation algorithm is self-contained: given V, it iteratively computes eigenvectors of the symmetric part and applies Householder updates. Theorem 1's proof shows that if the residual V_{k+1}=I then V=H_1...H_k by construction, and the converse uses only the standard external fact [16] that every orthogonal matrix is a product of at most n Householder matrices. Theorem 2's error bound follows from the trace recursion in Lemma 3; no parameter is fitted to the quantity being bounded. Section 3's identifiability result is likewise non-circular: Theorem 3 is proved by enumerating binary support sets and solving the data equations Y_{ij} = sum_k (delta_{ik} - 2 u_i u_k) X_{kj} for u, and the uniqueness argument in Lemmas 4-5 derives algebraic consequences of assuming two solutions and rules out collisions. The recovered u is an output, not an input, of the proof. The only flagged passage is in the proof of Lemma 4: "We assume that we do not divide by 0 in any case" (Section 3.2) — an unstated nondegeneracy condition not carried into Theorem 3's statement. When that condition fails, the theorem is false as stated, but this is a correctness gap, not circularity. No claim reduces to its own inputs, no fitted parameter is renamed as a prediction, and the only self-citation ([15]) is contextual rather than load-bearing. Hence score 0.

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

No numerical parameters are fitted to data. The central proofs rely on standard spectral facts about real orthogonal matrices, on the cited theorem that every orthogonal matrix is a product of at most n Householder reflections, and on an unstated nonzero-support condition for the two data columns in Theorem 3.

assumptions (3)

standard math Every real orthogonal n by n matrix can be written as a product of at most n Householder matrices
Used in Lemma 3 and the proof of Theorem 1, cited to reference [16]. This is a prior theorem, not derived here.
standard math Real orthogonal matrices are normal and admit a unitary eigendecomposition with conjugate-pair eigenvalues
Lemma 1 uses this to relate eigenspaces of V and of its symmetric part Vsym.
ad hoc to paper The two data columns have nonzero support sums delta1 and delta2, meaning neither column is orthogonal to the true Householder vector u, and the two columns are distinct
The proof of Lemma 4 explicitly says 'We assume that we do not divide by 0 in any case' and the uniqueness argument assumes the columns differ; the theorem statement does not include these conditions, and without them the two-sample claim fails.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exploring the Limitations of Structured Orthogonal Dictionary Learning." pith.science (2026). https://pith.science/paper/JQ7DIXJM

@misc{pith2026250115094,
  author       = {Pith},
  title        = {Pith review of: Exploring the Limitations of Structured Orthogonal Dictionary Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JQ7DIXJM}},
  note         = {Machine review of arXiv:2501.15094}
}

abstract

This work is motivated by recent applications of structured dictionary learning, in particular when the dictionary is assumed to be the product of a few Householder atoms. We investigate the following two problems: 1) How do we approximate an orthogonal matrix $\mathbf{V}$ with a product of a specified number of Householder matrices, and 2) How many samples are required to learn a structured (Householder) dictionary from data? For 1) we discuss an algorithm that decomposes $\mathbf{V}$ as a product of a specified number of Householder matrices. We see that the algorithm outputs the decomposition when it exists, and give bounds on the approximation error of the algorithm when such a decomposition does not exist. For 2) given data $\mathbf{Y}=\mathbf{HX}$, we show that when assuming a binary coefficient matrix $\mathbf{X}$, the structured (Householder) dictionary learning problem can be solved with just $2$ samples (columns) in $\mathbf{Y}$.

Figures

Figures reproduced from arXiv: 2501.15094 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 16 canonical work pages

[16]

Constructive ways for generating (generalized) real orthogonal matrices as products of (gen- eralized) symmetries,

F. Uhlig, “Constructive ways for generating (generalized) real orthogonal matrices as products of (gen- eralized) symmetries,” Linear Algebra and its Applications , vol. 332, pp. 459–467, 2001

work page 2001
[1]

Sparse coding with an overcomplete basis set: A strategy employed by V1?,

B. A. Olshausen and D. J. Field, “Sparse coding with an overcomplete basis set: A strategy employed by V1?,” Vision Research, vol. 37, no. 23, pp. 3311–3325, 1997

work page 1997
[2]

Method of optimal directions for frame design,

K. Engan, S. O. Aase, and J. H. Husoy, “Method of optimal directions for frame design,” in 1999 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings. ICASSP99 (Cat. No. 99CH36258) , vol. 5, pp. 2443–2446, 1999

work page 1999
[3]

K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation,

M. Aharon, M. Elad, and A. Bruckstein, “K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation,” IEEE Transactions on Signal Processing , vol. 54, no. 11, pp. 4311–4322, 2006

work page 2006
[4]

Online dictionary learning for sparse coding,

J. Mairal, F. Bach, J. Ponce, and G. Sapiro, “Online dictionary learning for sparse coding,” in Proceed- ings of the 26th Annual International Conference on Machine Learning , pp. 689–696, 2009

work page 2009
[5]

Complete dictionary recovery over the sphere,

J. Sun, Q. Qu, and J. Wright, “Complete dictionary recovery over the sphere,” in 2015 International Conference on Sampling Theory and Applications (SampTA) , pp. 407–410, 2015

work page 2015
[6]

Matrix factorization techniques in machine learning, signal processing, and statistics,

K.-L. Du, M. N. S. Swamy, Z.-Q. Wang, and W. H. Mow, “Matrix factorization techniques in machine learning, signal processing, and statistics,” Mathematics, vol. 11, no. 12, p. 2674, 2023

work page 2023
[7]

Efficient dictionary learning with gradient descent,

Dar Gilboa, Sam Buchanan, and John Wright, “Efficient dictionary learning with gradient descent,” in International Conference on Machine Learning . PMLR, 2019, pp. 2252–2259

work page 2019

Show all 17 references

[8]

Learning unions of orthonormal bases with thresholded singular value decomposition,

S. Lesage, R. Gribonval, F. Bimbot, and L. Benaroya, “Learning unions of orthonormal bases with thresholded singular value decomposition,” inProceedings (ICASSP’05). IEEE International Conference on Acoustics, Speech, and Signal Processing, 2005 , vol. 5, pp. v–293, 2005

2005
[9]

Fast sparsity-based orthogonal dictionary learning for image restoration,

C. Bao, J.-F. Cai, and H. Ji, “Fast sparsity-based orthogonal dictionary learning for image restoration,” in Proceedings of the IEEE International Conference on Computer Vision , pp. 3384–3391, 2013

2013
[10]

Simple alternating minimization provably solves complete dictionary learning,

G. Liang, G. Zhang, S. Fattahi, and R. Y. Zhang, “Simple alternating minimization provably solves complete dictionary learning,” arXiv preprint arXiv:2210.12816 , 2022

2022 arXiv
[11]

Complete dictionary learning via l4-norm maxi- mization over the orthogonal group,

Y. Zhai, Z. Yang, Z. Liao, J. Wright, and Y. Ma, “Complete dictionary learning via l4-norm maxi- mization over the orthogonal group,” Journal of Machine Learning Research, vol. 21, no. 165, pp. 1–68, 2020

2020
[12]

Improved online dictionary learning for sparse signal representation,

F. Yeganli, A. K. Fletcher, and A. O. Hero, “Improved online dictionary learning for sparse signal representation,” in 2014 22nd Signal Processing and Communications Applications Conference (SIU) . IEEE, 2014, pp. 1702–1705

2014
[13]

Fast orthonormal sparsifying transforms based on householder reflectors,

C. Rusu, N. Gonz´ alez-Prelcic, and R. W. Heath, “Fast orthonormal sparsifying transforms based on householder reflectors,” IEEE Transactions on Signal Processing , vol. 64, no. 24, pp. 6589–6599, 2016

2016
[14]

Learning fast sparsifying transforms,

C. Rusu and J. Thompson, “Learning fast sparsifying transforms,” IEEE Transactions on Signal Pro- cessing, vol. 65, no. 16, pp. 4367–4378, 2017

2017
[15]

Fast structured orthogonal dictionary learning using householder reflections,

Anirudh Dash and Aditya Siripuram, “Fast structured orthogonal dictionary learning using householder reflections,” arXiv preprint arXiv:2409.09138 , 2024. 13

2024 arXiv
[17]

G. H. Golub and C. F. Van Loan, Matrix computations. JHU Press, 2013. 14

2013

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [16]

Constructive ways for generating (generalized) real orthogonal matrices as products of (gen- eralized) symmetries,

F. Uhlig, “Constructive ways for generating (generalized) real orthogonal matrices as products of (gen- eralized) symmetries,” Linear Algebra and its Applications , vol. 332, pp. 459–467, 2001

work page 2001

[2] [1]

Sparse coding with an overcomplete basis set: A strategy employed by V1?,

B. A. Olshausen and D. J. Field, “Sparse coding with an overcomplete basis set: A strategy employed by V1?,” Vision Research, vol. 37, no. 23, pp. 3311–3325, 1997

work page 1997

[3] [2]

Method of optimal directions for frame design,

K. Engan, S. O. Aase, and J. H. Husoy, “Method of optimal directions for frame design,” in 1999 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings. ICASSP99 (Cat. No. 99CH36258) , vol. 5, pp. 2443–2446, 1999

work page 1999

[4] [3]

K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation,

M. Aharon, M. Elad, and A. Bruckstein, “K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation,” IEEE Transactions on Signal Processing , vol. 54, no. 11, pp. 4311–4322, 2006

work page 2006

[5] [4]

Online dictionary learning for sparse coding,

J. Mairal, F. Bach, J. Ponce, and G. Sapiro, “Online dictionary learning for sparse coding,” in Proceed- ings of the 26th Annual International Conference on Machine Learning , pp. 689–696, 2009

work page 2009

[6] [5]

Complete dictionary recovery over the sphere,

J. Sun, Q. Qu, and J. Wright, “Complete dictionary recovery over the sphere,” in 2015 International Conference on Sampling Theory and Applications (SampTA) , pp. 407–410, 2015

work page 2015

[7] [6]

Matrix factorization techniques in machine learning, signal processing, and statistics,

K.-L. Du, M. N. S. Swamy, Z.-Q. Wang, and W. H. Mow, “Matrix factorization techniques in machine learning, signal processing, and statistics,” Mathematics, vol. 11, no. 12, p. 2674, 2023

work page 2023

[8] [7]

Efficient dictionary learning with gradient descent,

Dar Gilboa, Sam Buchanan, and John Wright, “Efficient dictionary learning with gradient descent,” in International Conference on Machine Learning . PMLR, 2019, pp. 2252–2259

work page 2019